xref: /freebsd/usr.sbin/bhyve/pci_nvme.c (revision 2008043f386721d58158e37e0d7e50df8095942d)
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 	((cc) >> NVME_CC_REG_EN_SHIFT & NVME_CC_REG_EN_MASK)
373 #define	NVME_CC_GET_CSS(cc) \
374 	((cc) >> NVME_CC_REG_CSS_SHIFT & NVME_CC_REG_CSS_MASK)
375 #define	NVME_CC_GET_SHN(cc) \
376 	((cc) >> NVME_CC_REG_SHN_SHIFT & NVME_CC_REG_SHN_MASK)
377 #define	NVME_CC_GET_IOSQES(cc) \
378 	((cc) >> NVME_CC_REG_IOSQES_SHIFT & NVME_CC_REG_IOSQES_MASK)
379 #define	NVME_CC_GET_IOCQES(cc) \
380 	((cc) >> NVME_CC_REG_IOCQES_SHIFT & NVME_CC_REG_IOCQES_MASK)
381 
382 #define	NVME_CC_WRITE_MASK \
383 	((NVME_CC_REG_EN_MASK << NVME_CC_REG_EN_SHIFT) | \
384 	 (NVME_CC_REG_IOSQES_MASK << NVME_CC_REG_IOSQES_SHIFT) | \
385 	 (NVME_CC_REG_IOCQES_MASK << NVME_CC_REG_IOCQES_SHIFT))
386 
387 #define	NVME_CC_NEN_WRITE_MASK \
388 	((NVME_CC_REG_CSS_MASK << NVME_CC_REG_CSS_SHIFT) | \
389 	 (NVME_CC_REG_MPS_MASK << NVME_CC_REG_MPS_SHIFT) | \
390 	 (NVME_CC_REG_AMS_MASK << NVME_CC_REG_AMS_SHIFT))
391 
392 /* Controller Status utils */
393 #define	NVME_CSTS_GET_RDY(sts) \
394 	((sts) >> NVME_CSTS_REG_RDY_SHIFT & NVME_CSTS_REG_RDY_MASK)
395 
396 #define	NVME_CSTS_RDY	(1 << NVME_CSTS_REG_RDY_SHIFT)
397 #define	NVME_CSTS_CFS	(1 << NVME_CSTS_REG_CFS_SHIFT)
398 
399 /* Completion Queue status word utils */
400 #define	NVME_STATUS_P	(1 << NVME_STATUS_P_SHIFT)
401 #define	NVME_STATUS_MASK \
402 	((NVME_STATUS_SCT_MASK << NVME_STATUS_SCT_SHIFT) |\
403 	 (NVME_STATUS_SC_MASK << NVME_STATUS_SC_SHIFT))
404 
405 #define NVME_ONCS_DSM	(NVME_CTRLR_DATA_ONCS_DSM_MASK << \
406 	NVME_CTRLR_DATA_ONCS_DSM_SHIFT)
407 
408 static void nvme_feature_invalid_cb(struct pci_nvme_softc *,
409     struct nvme_feature_obj *,
410     struct nvme_command *,
411     struct nvme_completion *);
412 static void nvme_feature_temperature(struct pci_nvme_softc *,
413     struct nvme_feature_obj *,
414     struct nvme_command *,
415     struct nvme_completion *);
416 static void nvme_feature_num_queues(struct pci_nvme_softc *,
417     struct nvme_feature_obj *,
418     struct nvme_command *,
419     struct nvme_completion *);
420 static void nvme_feature_iv_config(struct pci_nvme_softc *,
421     struct nvme_feature_obj *,
422     struct nvme_command *,
423     struct nvme_completion *);
424 static void nvme_feature_async_event(struct pci_nvme_softc *,
425     struct nvme_feature_obj *,
426     struct nvme_command *,
427     struct nvme_completion *);
428 
429 static void *aen_thr(void *arg);
430 
431 static __inline void
432 cpywithpad(char *dst, size_t dst_size, const char *src, char pad)
433 {
434 	size_t len;
435 
436 	len = strnlen(src, dst_size);
437 	memset(dst, pad, dst_size);
438 	memcpy(dst, src, len);
439 }
440 
441 static __inline void
442 pci_nvme_status_tc(uint16_t *status, uint16_t type, uint16_t code)
443 {
444 
445 	*status &= ~NVME_STATUS_MASK;
446 	*status |= (type & NVME_STATUS_SCT_MASK) << NVME_STATUS_SCT_SHIFT |
447 		(code & NVME_STATUS_SC_MASK) << NVME_STATUS_SC_SHIFT;
448 }
449 
450 static __inline void
451 pci_nvme_status_genc(uint16_t *status, uint16_t code)
452 {
453 
454 	pci_nvme_status_tc(status, NVME_SCT_GENERIC, code);
455 }
456 
457 /*
458  * Initialize the requested number or IO Submission and Completion Queues.
459  * Admin queues are allocated implicitly.
460  */
461 static void
462 pci_nvme_init_queues(struct pci_nvme_softc *sc, uint32_t nsq, uint32_t ncq)
463 {
464 	uint32_t i;
465 
466 	/*
467 	 * Allocate and initialize the Submission Queues
468 	 */
469 	if (nsq > NVME_QUEUES) {
470 		WPRINTF("%s: clamping number of SQ from %u to %u",
471 					__func__, nsq, NVME_QUEUES);
472 		nsq = NVME_QUEUES;
473 	}
474 
475 	sc->num_squeues = nsq;
476 
477 	sc->submit_queues = calloc(sc->num_squeues + 1,
478 				sizeof(struct nvme_submission_queue));
479 	if (sc->submit_queues == NULL) {
480 		WPRINTF("%s: SQ allocation failed", __func__);
481 		sc->num_squeues = 0;
482 	} else {
483 		struct nvme_submission_queue *sq = sc->submit_queues;
484 
485 		for (i = 0; i < sc->num_squeues + 1; i++)
486 			pthread_mutex_init(&sq[i].mtx, NULL);
487 	}
488 
489 	/*
490 	 * Allocate and initialize the Completion Queues
491 	 */
492 	if (ncq > NVME_QUEUES) {
493 		WPRINTF("%s: clamping number of CQ from %u to %u",
494 					__func__, ncq, NVME_QUEUES);
495 		ncq = NVME_QUEUES;
496 	}
497 
498 	sc->num_cqueues = ncq;
499 
500 	sc->compl_queues = calloc(sc->num_cqueues + 1,
501 				sizeof(struct nvme_completion_queue));
502 	if (sc->compl_queues == NULL) {
503 		WPRINTF("%s: CQ allocation failed", __func__);
504 		sc->num_cqueues = 0;
505 	} else {
506 		struct nvme_completion_queue *cq = sc->compl_queues;
507 
508 		for (i = 0; i < sc->num_cqueues + 1; i++)
509 			pthread_mutex_init(&cq[i].mtx, NULL);
510 	}
511 }
512 
513 static void
514 pci_nvme_init_ctrldata(struct pci_nvme_softc *sc)
515 {
516 	struct nvme_controller_data *cd = &sc->ctrldata;
517 	int ret;
518 
519 	cd->vid = 0xFB5D;
520 	cd->ssvid = 0x0000;
521 
522 	cpywithpad((char *)cd->mn, sizeof(cd->mn), "bhyve-NVMe", ' ');
523 	cpywithpad((char *)cd->fr, sizeof(cd->fr), "1.0", ' ');
524 
525 	/* Num of submission commands that we can handle at a time (2^rab) */
526 	cd->rab   = 4;
527 
528 	/* FreeBSD OUI */
529 	cd->ieee[0] = 0xfc;
530 	cd->ieee[1] = 0x9c;
531 	cd->ieee[2] = 0x58;
532 
533 	cd->mic = 0;
534 
535 	cd->mdts = NVME_MDTS;	/* max data transfer size (2^mdts * CAP.MPSMIN) */
536 
537 	cd->ver = NVME_REV(1,4);
538 
539 	cd->cntrltype = NVME_CNTRLTYPE_IO;
540 	cd->oacs = 1 << NVME_CTRLR_DATA_OACS_FORMAT_SHIFT;
541 	cd->oaes = NVMEB(NVME_CTRLR_DATA_OAES_NS_ATTR);
542 	cd->acl = 2;
543 	cd->aerl = 4;
544 
545 	/* Advertise 1, Read-only firmware slot */
546 	cd->frmw = NVMEB(NVME_CTRLR_DATA_FRMW_SLOT1_RO) |
547 	    (1 << NVME_CTRLR_DATA_FRMW_NUM_SLOTS_SHIFT);
548 	cd->lpa = 0;	/* TODO: support some simple things like SMART */
549 	cd->elpe = 0;	/* max error log page entries */
550 	/*
551 	 * Report a single power state (zero-based value)
552 	 * power_state[] values are left as zero to indicate "Not reported"
553 	 */
554 	cd->npss = 0;
555 
556 	/* Warning Composite Temperature Threshold */
557 	cd->wctemp = 0x0157;
558 	cd->cctemp = 0x0157;
559 
560 	/* SANICAP must not be 0 for Revision 1.4 and later NVMe Controllers */
561 	cd->sanicap = (NVME_CTRLR_DATA_SANICAP_NODMMAS_NO <<
562 			NVME_CTRLR_DATA_SANICAP_NODMMAS_SHIFT);
563 
564 	cd->sqes = (6 << NVME_CTRLR_DATA_SQES_MAX_SHIFT) |
565 	    (6 << NVME_CTRLR_DATA_SQES_MIN_SHIFT);
566 	cd->cqes = (4 << NVME_CTRLR_DATA_CQES_MAX_SHIFT) |
567 	    (4 << NVME_CTRLR_DATA_CQES_MIN_SHIFT);
568 	cd->nn = 1;	/* number of namespaces */
569 
570 	cd->oncs = 0;
571 	switch (sc->dataset_management) {
572 	case NVME_DATASET_MANAGEMENT_AUTO:
573 		if (sc->nvstore.deallocate)
574 			cd->oncs |= NVME_ONCS_DSM;
575 		break;
576 	case NVME_DATASET_MANAGEMENT_ENABLE:
577 		cd->oncs |= NVME_ONCS_DSM;
578 		break;
579 	default:
580 		break;
581 	}
582 
583 	cd->fna = NVME_CTRLR_DATA_FNA_FORMAT_ALL_MASK <<
584 	    NVME_CTRLR_DATA_FNA_FORMAT_ALL_SHIFT;
585 
586 	cd->vwc = NVME_CTRLR_DATA_VWC_ALL_NO << NVME_CTRLR_DATA_VWC_ALL_SHIFT;
587 
588 	ret = snprintf(cd->subnqn, sizeof(cd->subnqn),
589 	    "nqn.2013-12.org.freebsd:bhyve-%s-%u-%u-%u",
590 	    get_config_value("name"), sc->nsc_pi->pi_bus,
591 	    sc->nsc_pi->pi_slot, sc->nsc_pi->pi_func);
592 	if ((ret < 0) || ((unsigned)ret > sizeof(cd->subnqn)))
593 		EPRINTLN("%s: error setting subnqn (%d)", __func__, ret);
594 }
595 
596 static void
597 pci_nvme_init_nsdata_size(struct pci_nvme_blockstore *nvstore,
598     struct nvme_namespace_data *nd)
599 {
600 
601 	/* Get capacity and block size information from backing store */
602 	nd->nsze = nvstore->size / nvstore->sectsz;
603 	nd->ncap = nd->nsze;
604 	nd->nuse = nd->nsze;
605 }
606 
607 static void
608 pci_nvme_init_nsdata(struct pci_nvme_softc *sc,
609     struct nvme_namespace_data *nd, uint32_t nsid,
610     struct pci_nvme_blockstore *nvstore)
611 {
612 
613 	pci_nvme_init_nsdata_size(nvstore, nd);
614 
615 	if (nvstore->type == NVME_STOR_BLOCKIF)
616 		nvstore->deallocate = blockif_candelete(nvstore->ctx);
617 
618 	nd->nlbaf = 0; /* NLBAF is a 0's based value (i.e. 1 LBA Format) */
619 	nd->flbas = 0;
620 
621 	/* Create an EUI-64 if user did not provide one */
622 	if (nvstore->eui64 == 0) {
623 		char *data = NULL;
624 		uint64_t eui64 = nvstore->eui64;
625 
626 		asprintf(&data, "%s%u%u%u", get_config_value("name"),
627 		    sc->nsc_pi->pi_bus, sc->nsc_pi->pi_slot,
628 		    sc->nsc_pi->pi_func);
629 
630 		if (data != NULL) {
631 			eui64 = OUI_FREEBSD_NVME_LOW | crc16(0, data, strlen(data));
632 			free(data);
633 		}
634 		nvstore->eui64 = (eui64 << 16) | (nsid & 0xffff);
635 	}
636 	be64enc(nd->eui64, nvstore->eui64);
637 
638 	/* LBA data-sz = 2^lbads */
639 	nd->lbaf[0] = nvstore->sectsz_bits << NVME_NS_DATA_LBAF_LBADS_SHIFT;
640 }
641 
642 static void
643 pci_nvme_init_logpages(struct pci_nvme_softc *sc)
644 {
645 	__uint128_t power_cycles = 1;
646 
647 	memset(&sc->err_log, 0, sizeof(sc->err_log));
648 	memset(&sc->health_log, 0, sizeof(sc->health_log));
649 	memset(&sc->fw_log, 0, sizeof(sc->fw_log));
650 	memset(&sc->ns_log, 0, sizeof(sc->ns_log));
651 
652 	/* Set read/write remainder to round up according to spec */
653 	sc->read_dunits_remainder = 999;
654 	sc->write_dunits_remainder = 999;
655 
656 	/* Set nominal Health values checked by implementations */
657 	sc->health_log.temperature = NVME_TEMPERATURE;
658 	sc->health_log.available_spare = 100;
659 	sc->health_log.available_spare_threshold = 10;
660 
661 	/* Set Active Firmware Info to slot 1 */
662 	sc->fw_log.afi = (1 << NVME_FIRMWARE_PAGE_AFI_SLOT_SHIFT);
663 	memcpy(&sc->fw_log.revision[0], sc->ctrldata.fr,
664 	    sizeof(sc->fw_log.revision[0]));
665 
666 	memcpy(&sc->health_log.power_cycles, &power_cycles,
667 	    sizeof(sc->health_log.power_cycles));
668 }
669 
670 static void
671 pci_nvme_init_features(struct pci_nvme_softc *sc)
672 {
673 	enum nvme_feature	fid;
674 
675 	for (fid = 0; fid < NVME_FID_MAX; fid++) {
676 		switch (fid) {
677 		case NVME_FEAT_ARBITRATION:
678 		case NVME_FEAT_POWER_MANAGEMENT:
679 		case NVME_FEAT_INTERRUPT_COALESCING: //XXX
680 		case NVME_FEAT_WRITE_ATOMICITY:
681 			/* Mandatory but no special handling required */
682 		//XXX hang - case NVME_FEAT_PREDICTABLE_LATENCY_MODE_CONFIG:
683 		//XXX hang - case NVME_FEAT_HOST_BEHAVIOR_SUPPORT:
684 		//		  this returns a data buffer
685 			break;
686 		case NVME_FEAT_TEMPERATURE_THRESHOLD:
687 			sc->feat[fid].set = nvme_feature_temperature;
688 			break;
689 		case NVME_FEAT_ERROR_RECOVERY:
690 			sc->feat[fid].namespace_specific = true;
691 			break;
692 		case NVME_FEAT_NUMBER_OF_QUEUES:
693 			sc->feat[fid].set = nvme_feature_num_queues;
694 			break;
695 		case NVME_FEAT_INTERRUPT_VECTOR_CONFIGURATION:
696 			sc->feat[fid].set = nvme_feature_iv_config;
697 			break;
698 		case NVME_FEAT_ASYNC_EVENT_CONFIGURATION:
699 			sc->feat[fid].set = nvme_feature_async_event;
700 			/* Enable all AENs by default */
701 			sc->feat[fid].cdw11 = PCI_NVME_AEN_DEFAULT_MASK;
702 			break;
703 		default:
704 			sc->feat[fid].set = nvme_feature_invalid_cb;
705 			sc->feat[fid].get = nvme_feature_invalid_cb;
706 		}
707 	}
708 }
709 
710 static void
711 pci_nvme_aer_reset(struct pci_nvme_softc *sc)
712 {
713 
714 	STAILQ_INIT(&sc->aer_list);
715 	sc->aer_count = 0;
716 }
717 
718 static void
719 pci_nvme_aer_init(struct pci_nvme_softc *sc)
720 {
721 
722 	pthread_mutex_init(&sc->aer_mtx, NULL);
723 	pci_nvme_aer_reset(sc);
724 }
725 
726 static void
727 pci_nvme_aer_destroy(struct pci_nvme_softc *sc)
728 {
729 	struct pci_nvme_aer *aer = NULL;
730 
731 	pthread_mutex_lock(&sc->aer_mtx);
732 	while (!STAILQ_EMPTY(&sc->aer_list)) {
733 		aer = STAILQ_FIRST(&sc->aer_list);
734 		STAILQ_REMOVE_HEAD(&sc->aer_list, link);
735 		free(aer);
736 	}
737 	pthread_mutex_unlock(&sc->aer_mtx);
738 
739 	pci_nvme_aer_reset(sc);
740 }
741 
742 static bool
743 pci_nvme_aer_available(struct pci_nvme_softc *sc)
744 {
745 
746 	return (sc->aer_count != 0);
747 }
748 
749 static bool
750 pci_nvme_aer_limit_reached(struct pci_nvme_softc *sc)
751 {
752 	struct nvme_controller_data *cd = &sc->ctrldata;
753 
754 	/* AERL is a zero based value while aer_count is one's based */
755 	return (sc->aer_count == (cd->aerl + 1U));
756 }
757 
758 /*
759  * Add an Async Event Request
760  *
761  * Stores an AER to be returned later if the Controller needs to notify the
762  * host of an event.
763  * Note that while the NVMe spec doesn't require Controllers to return AER's
764  * in order, this implementation does preserve the order.
765  */
766 static int
767 pci_nvme_aer_add(struct pci_nvme_softc *sc, uint16_t cid)
768 {
769 	struct pci_nvme_aer *aer = NULL;
770 
771 	aer = calloc(1, sizeof(struct pci_nvme_aer));
772 	if (aer == NULL)
773 		return (-1);
774 
775 	/* Save the Command ID for use in the completion message */
776 	aer->cid = cid;
777 
778 	pthread_mutex_lock(&sc->aer_mtx);
779 	sc->aer_count++;
780 	STAILQ_INSERT_TAIL(&sc->aer_list, aer, link);
781 	pthread_mutex_unlock(&sc->aer_mtx);
782 
783 	return (0);
784 }
785 
786 /*
787  * Get an Async Event Request structure
788  *
789  * Returns a pointer to an AER previously submitted by the host or NULL if
790  * no AER's exist. Caller is responsible for freeing the returned struct.
791  */
792 static struct pci_nvme_aer *
793 pci_nvme_aer_get(struct pci_nvme_softc *sc)
794 {
795 	struct pci_nvme_aer *aer = NULL;
796 
797 	pthread_mutex_lock(&sc->aer_mtx);
798 	aer = STAILQ_FIRST(&sc->aer_list);
799 	if (aer != NULL) {
800 		STAILQ_REMOVE_HEAD(&sc->aer_list, link);
801 		sc->aer_count--;
802 	}
803 	pthread_mutex_unlock(&sc->aer_mtx);
804 
805 	return (aer);
806 }
807 
808 static void
809 pci_nvme_aen_reset(struct pci_nvme_softc *sc)
810 {
811 	uint32_t	atype;
812 
813 	memset(sc->aen, 0, PCI_NVME_AE_TYPE_MAX * sizeof(struct pci_nvme_aen));
814 
815 	for (atype = 0; atype < PCI_NVME_AE_TYPE_MAX; atype++) {
816 		sc->aen[atype].atype = atype;
817 	}
818 }
819 
820 static void
821 pci_nvme_aen_init(struct pci_nvme_softc *sc)
822 {
823 	char nstr[80];
824 
825 	pci_nvme_aen_reset(sc);
826 
827 	pthread_mutex_init(&sc->aen_mtx, NULL);
828 	pthread_create(&sc->aen_tid, NULL, aen_thr, sc);
829 	snprintf(nstr, sizeof(nstr), "nvme-aen-%d:%d", sc->nsc_pi->pi_slot,
830 	    sc->nsc_pi->pi_func);
831 	pthread_set_name_np(sc->aen_tid, nstr);
832 }
833 
834 static void
835 pci_nvme_aen_destroy(struct pci_nvme_softc *sc)
836 {
837 
838 	pci_nvme_aen_reset(sc);
839 }
840 
841 /* Notify the AEN thread of pending work */
842 static void
843 pci_nvme_aen_notify(struct pci_nvme_softc *sc)
844 {
845 
846 	pthread_cond_signal(&sc->aen_cond);
847 }
848 
849 /*
850  * Post an Asynchronous Event Notification
851  */
852 static int32_t
853 pci_nvme_aen_post(struct pci_nvme_softc *sc, pci_nvme_async_type atype,
854 		uint32_t event_data)
855 {
856 	struct pci_nvme_aen *aen;
857 
858 	if (atype >= PCI_NVME_AE_TYPE_MAX) {
859 		return(EINVAL);
860 	}
861 
862 	pthread_mutex_lock(&sc->aen_mtx);
863 	aen = &sc->aen[atype];
864 
865 	/* Has the controller already posted an event of this type? */
866 	if (aen->posted) {
867 		pthread_mutex_unlock(&sc->aen_mtx);
868 		return(EALREADY);
869 	}
870 
871 	aen->event_data = event_data;
872 	aen->posted = true;
873 	pthread_mutex_unlock(&sc->aen_mtx);
874 
875 	pci_nvme_aen_notify(sc);
876 
877 	return(0);
878 }
879 
880 static void
881 pci_nvme_aen_process(struct pci_nvme_softc *sc)
882 {
883 	struct pci_nvme_aer *aer;
884 	struct pci_nvme_aen *aen;
885 	pci_nvme_async_type atype;
886 	uint32_t mask;
887 	uint16_t status;
888 	uint8_t lid;
889 
890 	assert(pthread_mutex_isowned_np(&sc->aen_mtx));
891 	for (atype = 0; atype < PCI_NVME_AE_TYPE_MAX; atype++) {
892 		aen = &sc->aen[atype];
893 		/* Previous iterations may have depleted the available AER's */
894 		if (!pci_nvme_aer_available(sc)) {
895 			DPRINTF("%s: no AER", __func__);
896 			break;
897 		}
898 
899 		if (!aen->posted) {
900 			DPRINTF("%s: no AEN posted for atype=%#x", __func__, atype);
901 			continue;
902 		}
903 
904 		status = NVME_SC_SUCCESS;
905 
906 		/* Is the event masked? */
907 		mask =
908 		    sc->feat[NVME_FEAT_ASYNC_EVENT_CONFIGURATION].cdw11;
909 
910 		DPRINTF("%s: atype=%#x mask=%#x event_data=%#x", __func__, atype, mask, aen->event_data);
911 		switch (atype) {
912 		case PCI_NVME_AE_TYPE_ERROR:
913 			lid = NVME_LOG_ERROR;
914 			break;
915 		case PCI_NVME_AE_TYPE_SMART:
916 			mask &= 0xff;
917 			if ((mask & aen->event_data) == 0)
918 				continue;
919 			lid = NVME_LOG_HEALTH_INFORMATION;
920 			break;
921 		case PCI_NVME_AE_TYPE_NOTICE:
922 			if (aen->event_data >= PCI_NVME_AEI_NOTICE_MAX) {
923 				EPRINTLN("%s unknown AEN notice type %u",
924 				    __func__, aen->event_data);
925 				status = NVME_SC_INTERNAL_DEVICE_ERROR;
926 				lid = 0;
927 				break;
928 			}
929 			if ((PCI_NVME_AEI_NOTICE_MASK(aen->event_data) & mask) == 0)
930 				continue;
931 			switch (aen->event_data) {
932 			case PCI_NVME_AEI_NOTICE_NS_ATTR_CHANGED:
933 				lid = NVME_LOG_CHANGED_NAMESPACE;
934 				break;
935 			case PCI_NVME_AEI_NOTICE_FW_ACTIVATION:
936 				lid = NVME_LOG_FIRMWARE_SLOT;
937 				break;
938 			case PCI_NVME_AEI_NOTICE_TELEMETRY_CHANGE:
939 				lid = NVME_LOG_TELEMETRY_CONTROLLER_INITIATED;
940 				break;
941 			case PCI_NVME_AEI_NOTICE_ANA_CHANGE:
942 				lid = NVME_LOG_ASYMMETRIC_NAMESPACE_ACCESS;
943 				break;
944 			case PCI_NVME_AEI_NOTICE_PREDICT_LATENCY_CHANGE:
945 				lid = NVME_LOG_PREDICTABLE_LATENCY_EVENT_AGGREGATE;
946 				break;
947 			case PCI_NVME_AEI_NOTICE_LBA_STATUS_ALERT:
948 				lid = NVME_LOG_LBA_STATUS_INFORMATION;
949 				break;
950 			case PCI_NVME_AEI_NOTICE_ENDURANCE_GROUP_CHANGE:
951 				lid = NVME_LOG_ENDURANCE_GROUP_EVENT_AGGREGATE;
952 				break;
953 			default:
954 				lid = 0;
955 			}
956 			break;
957 		default:
958 			/* bad type?!? */
959 			EPRINTLN("%s unknown AEN type %u", __func__, atype);
960 			status = NVME_SC_INTERNAL_DEVICE_ERROR;
961 			lid = 0;
962 			break;
963 		}
964 
965 		aer = pci_nvme_aer_get(sc);
966 		assert(aer != NULL);
967 
968 		DPRINTF("%s: CID=%#x CDW0=%#x", __func__, aer->cid, (lid << 16) | (aen->event_data << 8) | atype);
969 		pci_nvme_cq_update(sc, &sc->compl_queues[0],
970 		    (lid << 16) | (aen->event_data << 8) | atype, /* cdw0 */
971 		    aer->cid,
972 		    0,		/* SQID */
973 		    status);
974 
975 		aen->event_data = 0;
976 		aen->posted = false;
977 
978 		pci_generate_msix(sc->nsc_pi, 0);
979 	}
980 }
981 
982 static void *
983 aen_thr(void *arg)
984 {
985 	struct pci_nvme_softc *sc;
986 
987 	sc = arg;
988 
989 	pthread_mutex_lock(&sc->aen_mtx);
990 	for (;;) {
991 		pci_nvme_aen_process(sc);
992 		pthread_cond_wait(&sc->aen_cond, &sc->aen_mtx);
993 	}
994 	pthread_mutex_unlock(&sc->aen_mtx);
995 
996 	pthread_exit(NULL);
997 	return (NULL);
998 }
999 
1000 static void
1001 pci_nvme_reset_locked(struct pci_nvme_softc *sc)
1002 {
1003 	uint32_t i;
1004 
1005 	DPRINTF("%s", __func__);
1006 
1007 	sc->regs.cap_lo = (ZERO_BASED(sc->max_qentries) & NVME_CAP_LO_REG_MQES_MASK) |
1008 	    (1 << NVME_CAP_LO_REG_CQR_SHIFT) |
1009 	    (60 << NVME_CAP_LO_REG_TO_SHIFT);
1010 
1011 	sc->regs.cap_hi = 1 << NVME_CAP_HI_REG_CSS_NVM_SHIFT;
1012 
1013 	sc->regs.vs = NVME_REV(1,4);	/* NVMe v1.4 */
1014 
1015 	sc->regs.cc = 0;
1016 
1017 	assert(sc->submit_queues != NULL);
1018 
1019 	for (i = 0; i < sc->num_squeues + 1; i++) {
1020 		sc->submit_queues[i].qbase = NULL;
1021 		sc->submit_queues[i].size = 0;
1022 		sc->submit_queues[i].cqid = 0;
1023 		sc->submit_queues[i].tail = 0;
1024 		sc->submit_queues[i].head = 0;
1025 	}
1026 
1027 	assert(sc->compl_queues != NULL);
1028 
1029 	for (i = 0; i < sc->num_cqueues + 1; i++) {
1030 		sc->compl_queues[i].qbase = NULL;
1031 		sc->compl_queues[i].size = 0;
1032 		sc->compl_queues[i].tail = 0;
1033 		sc->compl_queues[i].head = 0;
1034 	}
1035 
1036 	sc->num_q_is_set = false;
1037 
1038 	pci_nvme_aer_destroy(sc);
1039 	pci_nvme_aen_destroy(sc);
1040 
1041 	/*
1042 	 * Clear CSTS.RDY last to prevent the host from enabling Controller
1043 	 * before cleanup completes
1044 	 */
1045 	sc->regs.csts = 0;
1046 }
1047 
1048 static void
1049 pci_nvme_reset(struct pci_nvme_softc *sc)
1050 {
1051 	pthread_mutex_lock(&sc->mtx);
1052 	pci_nvme_reset_locked(sc);
1053 	pthread_mutex_unlock(&sc->mtx);
1054 }
1055 
1056 static int
1057 pci_nvme_init_controller(struct pci_nvme_softc *sc)
1058 {
1059 	uint16_t acqs, asqs;
1060 
1061 	DPRINTF("%s", __func__);
1062 
1063 	/*
1064 	 * NVMe 2.0 states that "enabling a controller while this field is
1065 	 * cleared to 0h produces undefined results" for both ACQS and
1066 	 * ASQS. If zero, set CFS and do not become ready.
1067 	 */
1068 	asqs = ONE_BASED(sc->regs.aqa & NVME_AQA_REG_ASQS_MASK);
1069 	if (asqs < 2) {
1070 		EPRINTLN("%s: illegal ASQS value %#x (aqa=%#x)", __func__,
1071 		    asqs - 1, sc->regs.aqa);
1072 		sc->regs.csts |= NVME_CSTS_CFS;
1073 		return (-1);
1074 	}
1075 	sc->submit_queues[0].size = asqs;
1076 	sc->submit_queues[0].qbase = vm_map_gpa(sc->nsc_pi->pi_vmctx,
1077 	    sc->regs.asq, sizeof(struct nvme_command) * asqs);
1078 	if (sc->submit_queues[0].qbase == NULL) {
1079 		EPRINTLN("%s: ASQ vm_map_gpa(%lx) failed", __func__,
1080 		    sc->regs.asq);
1081 		sc->regs.csts |= NVME_CSTS_CFS;
1082 		return (-1);
1083 	}
1084 
1085 	DPRINTF("%s mapping Admin-SQ guest 0x%lx, host: %p",
1086 	        __func__, sc->regs.asq, sc->submit_queues[0].qbase);
1087 
1088 	acqs = ONE_BASED((sc->regs.aqa >> NVME_AQA_REG_ACQS_SHIFT) &
1089 	    NVME_AQA_REG_ACQS_MASK);
1090 	if (acqs < 2) {
1091 		EPRINTLN("%s: illegal ACQS value %#x (aqa=%#x)", __func__,
1092 		    acqs - 1, sc->regs.aqa);
1093 		sc->regs.csts |= NVME_CSTS_CFS;
1094 		return (-1);
1095 	}
1096 	sc->compl_queues[0].size = acqs;
1097 	sc->compl_queues[0].qbase = vm_map_gpa(sc->nsc_pi->pi_vmctx,
1098 	    sc->regs.acq, sizeof(struct nvme_completion) * acqs);
1099 	if (sc->compl_queues[0].qbase == NULL) {
1100 		EPRINTLN("%s: ACQ vm_map_gpa(%lx) failed", __func__,
1101 		    sc->regs.acq);
1102 		sc->regs.csts |= NVME_CSTS_CFS;
1103 		return (-1);
1104 	}
1105 	sc->compl_queues[0].intr_en = NVME_CQ_INTEN;
1106 
1107 	DPRINTF("%s mapping Admin-CQ guest 0x%lx, host: %p",
1108 	        __func__, sc->regs.acq, sc->compl_queues[0].qbase);
1109 
1110 	return (0);
1111 }
1112 
1113 static int
1114 nvme_prp_memcpy(struct vmctx *ctx, uint64_t prp1, uint64_t prp2, uint8_t *b,
1115 	size_t len, enum nvme_copy_dir dir)
1116 {
1117 	uint8_t *p;
1118 	size_t bytes;
1119 
1120 	if (len > (8 * 1024)) {
1121 		return (-1);
1122 	}
1123 
1124 	/* Copy from the start of prp1 to the end of the physical page */
1125 	bytes = PAGE_SIZE - (prp1 & PAGE_MASK);
1126 	bytes = MIN(bytes, len);
1127 
1128 	p = vm_map_gpa(ctx, prp1, bytes);
1129 	if (p == NULL) {
1130 		return (-1);
1131 	}
1132 
1133 	if (dir == NVME_COPY_TO_PRP)
1134 		memcpy(p, b, bytes);
1135 	else
1136 		memcpy(b, p, bytes);
1137 
1138 	b += bytes;
1139 
1140 	len -= bytes;
1141 	if (len == 0) {
1142 		return (0);
1143 	}
1144 
1145 	len = MIN(len, PAGE_SIZE);
1146 
1147 	p = vm_map_gpa(ctx, prp2, len);
1148 	if (p == NULL) {
1149 		return (-1);
1150 	}
1151 
1152 	if (dir == NVME_COPY_TO_PRP)
1153 		memcpy(p, b, len);
1154 	else
1155 		memcpy(b, p, len);
1156 
1157 	return (0);
1158 }
1159 
1160 /*
1161  * Write a Completion Queue Entry update
1162  *
1163  * Write the completion and update the doorbell value
1164  */
1165 static void
1166 pci_nvme_cq_update(struct pci_nvme_softc *sc,
1167 		struct nvme_completion_queue *cq,
1168 		uint32_t cdw0,
1169 		uint16_t cid,
1170 		uint16_t sqid,
1171 		uint16_t status)
1172 {
1173 	struct nvme_submission_queue *sq = &sc->submit_queues[sqid];
1174 	struct nvme_completion *cqe;
1175 
1176 	assert(cq->qbase != NULL);
1177 
1178 	pthread_mutex_lock(&cq->mtx);
1179 
1180 	cqe = &cq->qbase[cq->tail];
1181 
1182 	/* Flip the phase bit */
1183 	status |= (cqe->status ^ NVME_STATUS_P) & NVME_STATUS_P_MASK;
1184 
1185 	cqe->cdw0 = cdw0;
1186 	cqe->sqhd = sq->head;
1187 	cqe->sqid = sqid;
1188 	cqe->cid = cid;
1189 	cqe->status = status;
1190 
1191 	cq->tail++;
1192 	if (cq->tail >= cq->size) {
1193 		cq->tail = 0;
1194 	}
1195 
1196 	pthread_mutex_unlock(&cq->mtx);
1197 }
1198 
1199 static int
1200 nvme_opc_delete_io_sq(struct pci_nvme_softc* sc, struct nvme_command* command,
1201 	struct nvme_completion* compl)
1202 {
1203 	uint16_t qid = command->cdw10 & 0xffff;
1204 
1205 	DPRINTF("%s DELETE_IO_SQ %u", __func__, qid);
1206 	if (qid == 0 || qid > sc->num_squeues ||
1207 	    (sc->submit_queues[qid].qbase == NULL)) {
1208 		WPRINTF("%s NOT PERMITTED queue id %u / num_squeues %u",
1209 		        __func__, qid, sc->num_squeues);
1210 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1211 		    NVME_SC_INVALID_QUEUE_IDENTIFIER);
1212 		return (1);
1213 	}
1214 
1215 	sc->submit_queues[qid].qbase = NULL;
1216 	sc->submit_queues[qid].cqid = 0;
1217 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1218 	return (1);
1219 }
1220 
1221 static int
1222 nvme_opc_create_io_sq(struct pci_nvme_softc* sc, struct nvme_command* command,
1223 	struct nvme_completion* compl)
1224 {
1225 	if (command->cdw11 & NVME_CMD_CDW11_PC) {
1226 		uint16_t qid = command->cdw10 & 0xffff;
1227 		struct nvme_submission_queue *nsq;
1228 
1229 		if ((qid == 0) || (qid > sc->num_squeues) ||
1230 		    (sc->submit_queues[qid].qbase != NULL)) {
1231 			WPRINTF("%s queue index %u > num_squeues %u",
1232 			        __func__, qid, sc->num_squeues);
1233 			pci_nvme_status_tc(&compl->status,
1234 			    NVME_SCT_COMMAND_SPECIFIC,
1235 			    NVME_SC_INVALID_QUEUE_IDENTIFIER);
1236 			return (1);
1237 		}
1238 
1239 		nsq = &sc->submit_queues[qid];
1240 		nsq->size = ONE_BASED((command->cdw10 >> 16) & 0xffff);
1241 		DPRINTF("%s size=%u (max=%u)", __func__, nsq->size, sc->max_qentries);
1242 		if ((nsq->size < 2) || (nsq->size > sc->max_qentries)) {
1243 			/*
1244 			 * Queues must specify at least two entries
1245 			 * NOTE: "MAXIMUM QUEUE SIZE EXCEEDED" was renamed to
1246 			 * "INVALID QUEUE SIZE" in the NVM Express 1.3 Spec
1247 			 */
1248 			pci_nvme_status_tc(&compl->status,
1249 			    NVME_SCT_COMMAND_SPECIFIC,
1250 			    NVME_SC_MAXIMUM_QUEUE_SIZE_EXCEEDED);
1251 			return (1);
1252 		}
1253 		nsq->head = nsq->tail = 0;
1254 
1255 		nsq->cqid = (command->cdw11 >> 16) & 0xffff;
1256 		if ((nsq->cqid == 0) || (nsq->cqid > sc->num_cqueues)) {
1257 			pci_nvme_status_tc(&compl->status,
1258 			    NVME_SCT_COMMAND_SPECIFIC,
1259 			    NVME_SC_INVALID_QUEUE_IDENTIFIER);
1260 			return (1);
1261 		}
1262 
1263 		if (sc->compl_queues[nsq->cqid].qbase == NULL) {
1264 			pci_nvme_status_tc(&compl->status,
1265 			    NVME_SCT_COMMAND_SPECIFIC,
1266 			    NVME_SC_COMPLETION_QUEUE_INVALID);
1267 			return (1);
1268 		}
1269 
1270 		nsq->qpriority = (command->cdw11 >> 1) & 0x03;
1271 
1272 		nsq->qbase = vm_map_gpa(sc->nsc_pi->pi_vmctx, command->prp1,
1273 		              sizeof(struct nvme_command) * (size_t)nsq->size);
1274 
1275 		DPRINTF("%s sq %u size %u gaddr %p cqid %u", __func__,
1276 		        qid, nsq->size, nsq->qbase, nsq->cqid);
1277 
1278 		pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1279 
1280 		DPRINTF("%s completed creating IOSQ qid %u",
1281 		         __func__, qid);
1282 	} else {
1283 		/*
1284 		 * Guest sent non-cont submission queue request.
1285 		 * This setting is unsupported by this emulation.
1286 		 */
1287 		WPRINTF("%s unsupported non-contig (list-based) "
1288 		         "create i/o submission queue", __func__);
1289 
1290 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1291 	}
1292 	return (1);
1293 }
1294 
1295 static int
1296 nvme_opc_delete_io_cq(struct pci_nvme_softc* sc, struct nvme_command* command,
1297 	struct nvme_completion* compl)
1298 {
1299 	uint16_t qid = command->cdw10 & 0xffff;
1300 	uint16_t sqid;
1301 
1302 	DPRINTF("%s DELETE_IO_CQ %u", __func__, qid);
1303 	if (qid == 0 || qid > sc->num_cqueues ||
1304 	    (sc->compl_queues[qid].qbase == NULL)) {
1305 		WPRINTF("%s queue index %u / num_cqueues %u",
1306 		        __func__, qid, sc->num_cqueues);
1307 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1308 		    NVME_SC_INVALID_QUEUE_IDENTIFIER);
1309 		return (1);
1310 	}
1311 
1312 	/* Deleting an Active CQ is an error */
1313 	for (sqid = 1; sqid < sc->num_squeues + 1; sqid++)
1314 		if (sc->submit_queues[sqid].cqid == qid) {
1315 			pci_nvme_status_tc(&compl->status,
1316 			    NVME_SCT_COMMAND_SPECIFIC,
1317 			    NVME_SC_INVALID_QUEUE_DELETION);
1318 			return (1);
1319 		}
1320 
1321 	sc->compl_queues[qid].qbase = NULL;
1322 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1323 	return (1);
1324 }
1325 
1326 static int
1327 nvme_opc_create_io_cq(struct pci_nvme_softc* sc, struct nvme_command* command,
1328 	struct nvme_completion* compl)
1329 {
1330 	struct nvme_completion_queue *ncq;
1331 	uint16_t qid = command->cdw10 & 0xffff;
1332 
1333 	/* Only support Physically Contiguous queues */
1334 	if ((command->cdw11 & NVME_CMD_CDW11_PC) == 0) {
1335 		WPRINTF("%s unsupported non-contig (list-based) "
1336 		         "create i/o completion queue",
1337 		         __func__);
1338 
1339 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1340 		return (1);
1341 	}
1342 
1343 	if ((qid == 0) || (qid > sc->num_cqueues) ||
1344 	    (sc->compl_queues[qid].qbase != NULL)) {
1345 		WPRINTF("%s queue index %u > num_cqueues %u",
1346 			__func__, qid, sc->num_cqueues);
1347 		pci_nvme_status_tc(&compl->status,
1348 		    NVME_SCT_COMMAND_SPECIFIC,
1349 		    NVME_SC_INVALID_QUEUE_IDENTIFIER);
1350 		return (1);
1351  	}
1352 
1353 	ncq = &sc->compl_queues[qid];
1354 	ncq->intr_en = (command->cdw11 & NVME_CMD_CDW11_IEN) >> 1;
1355 	ncq->intr_vec = (command->cdw11 >> 16) & 0xffff;
1356 	if (ncq->intr_vec > (sc->max_queues + 1)) {
1357 		pci_nvme_status_tc(&compl->status,
1358 		    NVME_SCT_COMMAND_SPECIFIC,
1359 		    NVME_SC_INVALID_INTERRUPT_VECTOR);
1360 		return (1);
1361 	}
1362 
1363 	ncq->size = ONE_BASED((command->cdw10 >> 16) & 0xffff);
1364 	if ((ncq->size < 2) || (ncq->size > sc->max_qentries))  {
1365 		/*
1366 		 * Queues must specify at least two entries
1367 		 * NOTE: "MAXIMUM QUEUE SIZE EXCEEDED" was renamed to
1368 		 * "INVALID QUEUE SIZE" in the NVM Express 1.3 Spec
1369 		 */
1370 		pci_nvme_status_tc(&compl->status,
1371 		    NVME_SCT_COMMAND_SPECIFIC,
1372 		    NVME_SC_MAXIMUM_QUEUE_SIZE_EXCEEDED);
1373 		return (1);
1374 	}
1375 	ncq->head = ncq->tail = 0;
1376 	ncq->qbase = vm_map_gpa(sc->nsc_pi->pi_vmctx,
1377 		     command->prp1,
1378 		     sizeof(struct nvme_command) * (size_t)ncq->size);
1379 
1380 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1381 
1382 
1383 	return (1);
1384 }
1385 
1386 static int
1387 nvme_opc_get_log_page(struct pci_nvme_softc* sc, struct nvme_command* command,
1388 	struct nvme_completion* compl)
1389 {
1390 	uint64_t logoff;
1391 	uint32_t logsize;
1392 	uint8_t logpage;
1393 
1394 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1395 
1396 	/*
1397 	 * Command specifies the number of dwords to return in fields NUMDU
1398 	 * and NUMDL. This is a zero-based value.
1399 	 */
1400 	logpage = command->cdw10 & 0xFF;
1401 	logsize = ((command->cdw11 << 16) | (command->cdw10 >> 16)) + 1;
1402 	logsize *= sizeof(uint32_t);
1403 	logoff  = ((uint64_t)(command->cdw13) << 32) | command->cdw12;
1404 
1405 	DPRINTF("%s log page %u len %u", __func__, logpage, logsize);
1406 
1407 	switch (logpage) {
1408 	case NVME_LOG_ERROR:
1409 		if (logoff >= sizeof(sc->err_log)) {
1410 			pci_nvme_status_genc(&compl->status,
1411 			    NVME_SC_INVALID_FIELD);
1412 			break;
1413 		}
1414 
1415 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
1416 		    command->prp2, (uint8_t *)&sc->err_log + logoff,
1417 		    MIN(logsize - logoff, sizeof(sc->err_log)),
1418 		    NVME_COPY_TO_PRP);
1419 		break;
1420 	case NVME_LOG_HEALTH_INFORMATION:
1421 		if (logoff >= sizeof(sc->health_log)) {
1422 			pci_nvme_status_genc(&compl->status,
1423 			    NVME_SC_INVALID_FIELD);
1424 			break;
1425 		}
1426 
1427 		pthread_mutex_lock(&sc->mtx);
1428 		memcpy(&sc->health_log.data_units_read, &sc->read_data_units,
1429 		    sizeof(sc->health_log.data_units_read));
1430 		memcpy(&sc->health_log.data_units_written, &sc->write_data_units,
1431 		    sizeof(sc->health_log.data_units_written));
1432 		memcpy(&sc->health_log.host_read_commands, &sc->read_commands,
1433 		    sizeof(sc->health_log.host_read_commands));
1434 		memcpy(&sc->health_log.host_write_commands, &sc->write_commands,
1435 		    sizeof(sc->health_log.host_write_commands));
1436 		pthread_mutex_unlock(&sc->mtx);
1437 
1438 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
1439 		    command->prp2, (uint8_t *)&sc->health_log + logoff,
1440 		    MIN(logsize - logoff, sizeof(sc->health_log)),
1441 		    NVME_COPY_TO_PRP);
1442 		break;
1443 	case NVME_LOG_FIRMWARE_SLOT:
1444 		if (logoff >= sizeof(sc->fw_log)) {
1445 			pci_nvme_status_genc(&compl->status,
1446 			    NVME_SC_INVALID_FIELD);
1447 			break;
1448 		}
1449 
1450 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
1451 		    command->prp2, (uint8_t *)&sc->fw_log + logoff,
1452 		    MIN(logsize - logoff, sizeof(sc->fw_log)),
1453 		    NVME_COPY_TO_PRP);
1454 		break;
1455 	case NVME_LOG_CHANGED_NAMESPACE:
1456 		if (logoff >= sizeof(sc->ns_log)) {
1457 			pci_nvme_status_genc(&compl->status,
1458 			    NVME_SC_INVALID_FIELD);
1459 			break;
1460 		}
1461 
1462 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
1463 		    command->prp2, (uint8_t *)&sc->ns_log + logoff,
1464 		    MIN(logsize - logoff, sizeof(sc->ns_log)),
1465 		    NVME_COPY_TO_PRP);
1466 		memset(&sc->ns_log, 0, sizeof(sc->ns_log));
1467 		break;
1468 	default:
1469 		DPRINTF("%s get log page %x command not supported",
1470 		        __func__, logpage);
1471 
1472 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1473 		    NVME_SC_INVALID_LOG_PAGE);
1474 	}
1475 
1476 	return (1);
1477 }
1478 
1479 static int
1480 nvme_opc_identify(struct pci_nvme_softc* sc, struct nvme_command* command,
1481 	struct nvme_completion* compl)
1482 {
1483 	void *dest;
1484 	uint16_t status;
1485 
1486 	DPRINTF("%s identify 0x%x nsid 0x%x", __func__,
1487 	        command->cdw10 & 0xFF, command->nsid);
1488 
1489 	status = 0;
1490 	pci_nvme_status_genc(&status, NVME_SC_SUCCESS);
1491 
1492 	switch (command->cdw10 & 0xFF) {
1493 	case 0x00: /* return Identify Namespace data structure */
1494 		/* Global NS only valid with NS Management */
1495 		if (command->nsid == NVME_GLOBAL_NAMESPACE_TAG) {
1496 			pci_nvme_status_genc(&status,
1497 			    NVME_SC_INVALID_NAMESPACE_OR_FORMAT);
1498 			break;
1499 		}
1500 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
1501 		    command->prp2, (uint8_t *)&sc->nsdata, sizeof(sc->nsdata),
1502 		    NVME_COPY_TO_PRP);
1503 		break;
1504 	case 0x01: /* return Identify Controller data structure */
1505 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
1506 		    command->prp2, (uint8_t *)&sc->ctrldata,
1507 		    sizeof(sc->ctrldata),
1508 		    NVME_COPY_TO_PRP);
1509 		break;
1510 	case 0x02: /* list of 1024 active NSIDs > CDW1.NSID */
1511 		dest = vm_map_gpa(sc->nsc_pi->pi_vmctx, command->prp1,
1512 		                  sizeof(uint32_t) * 1024);
1513 		/* All unused entries shall be zero */
1514 		memset(dest, 0, sizeof(uint32_t) * 1024);
1515 		((uint32_t *)dest)[0] = 1;
1516 		break;
1517 	case 0x03: /* list of NSID structures in CDW1.NSID, 4096 bytes */
1518 		if (command->nsid != 1) {
1519 			pci_nvme_status_genc(&status,
1520 			    NVME_SC_INVALID_NAMESPACE_OR_FORMAT);
1521 			break;
1522 		}
1523 		dest = vm_map_gpa(sc->nsc_pi->pi_vmctx, command->prp1,
1524 		                  sizeof(uint32_t) * 1024);
1525 		/* All bytes after the descriptor shall be zero */
1526 		memset(dest, 0, sizeof(uint32_t) * 1024);
1527 
1528 		/* Return NIDT=1 (i.e. EUI64) descriptor */
1529 		((uint8_t *)dest)[0] = 1;
1530 		((uint8_t *)dest)[1] = sizeof(uint64_t);
1531 		memcpy(((uint8_t *)dest) + 4, sc->nsdata.eui64, sizeof(uint64_t));
1532 		break;
1533 	case 0x13:
1534 		/*
1535 		 * Controller list is optional but used by UNH tests. Return
1536 		 * a valid but empty list.
1537 		 */
1538 		dest = vm_map_gpa(sc->nsc_pi->pi_vmctx, command->prp1,
1539 		                  sizeof(uint16_t) * 2048);
1540 		memset(dest, 0, sizeof(uint16_t) * 2048);
1541 		break;
1542 	default:
1543 		DPRINTF("%s unsupported identify command requested 0x%x",
1544 		         __func__, command->cdw10 & 0xFF);
1545 		pci_nvme_status_genc(&status, NVME_SC_INVALID_FIELD);
1546 		break;
1547 	}
1548 
1549 	compl->status = status;
1550 	return (1);
1551 }
1552 
1553 static const char *
1554 nvme_fid_to_name(uint8_t fid)
1555 {
1556 	const char *name;
1557 
1558 	switch (fid) {
1559 	case NVME_FEAT_ARBITRATION:
1560 		name = "Arbitration";
1561 		break;
1562 	case NVME_FEAT_POWER_MANAGEMENT:
1563 		name = "Power Management";
1564 		break;
1565 	case NVME_FEAT_LBA_RANGE_TYPE:
1566 		name = "LBA Range Type";
1567 		break;
1568 	case NVME_FEAT_TEMPERATURE_THRESHOLD:
1569 		name = "Temperature Threshold";
1570 		break;
1571 	case NVME_FEAT_ERROR_RECOVERY:
1572 		name = "Error Recovery";
1573 		break;
1574 	case NVME_FEAT_VOLATILE_WRITE_CACHE:
1575 		name = "Volatile Write Cache";
1576 		break;
1577 	case NVME_FEAT_NUMBER_OF_QUEUES:
1578 		name = "Number of Queues";
1579 		break;
1580 	case NVME_FEAT_INTERRUPT_COALESCING:
1581 		name = "Interrupt Coalescing";
1582 		break;
1583 	case NVME_FEAT_INTERRUPT_VECTOR_CONFIGURATION:
1584 		name = "Interrupt Vector Configuration";
1585 		break;
1586 	case NVME_FEAT_WRITE_ATOMICITY:
1587 		name = "Write Atomicity Normal";
1588 		break;
1589 	case NVME_FEAT_ASYNC_EVENT_CONFIGURATION:
1590 		name = "Asynchronous Event Configuration";
1591 		break;
1592 	case NVME_FEAT_AUTONOMOUS_POWER_STATE_TRANSITION:
1593 		name = "Autonomous Power State Transition";
1594 		break;
1595 	case NVME_FEAT_HOST_MEMORY_BUFFER:
1596 		name = "Host Memory Buffer";
1597 		break;
1598 	case NVME_FEAT_TIMESTAMP:
1599 		name = "Timestamp";
1600 		break;
1601 	case NVME_FEAT_KEEP_ALIVE_TIMER:
1602 		name = "Keep Alive Timer";
1603 		break;
1604 	case NVME_FEAT_HOST_CONTROLLED_THERMAL_MGMT:
1605 		name = "Host Controlled Thermal Management";
1606 		break;
1607 	case NVME_FEAT_NON_OP_POWER_STATE_CONFIG:
1608 		name = "Non-Operation Power State Config";
1609 		break;
1610 	case NVME_FEAT_READ_RECOVERY_LEVEL_CONFIG:
1611 		name = "Read Recovery Level Config";
1612 		break;
1613 	case NVME_FEAT_PREDICTABLE_LATENCY_MODE_CONFIG:
1614 		name = "Predictable Latency Mode Config";
1615 		break;
1616 	case NVME_FEAT_PREDICTABLE_LATENCY_MODE_WINDOW:
1617 		name = "Predictable Latency Mode Window";
1618 		break;
1619 	case NVME_FEAT_LBA_STATUS_INFORMATION_ATTRIBUTES:
1620 		name = "LBA Status Information Report Interval";
1621 		break;
1622 	case NVME_FEAT_HOST_BEHAVIOR_SUPPORT:
1623 		name = "Host Behavior Support";
1624 		break;
1625 	case NVME_FEAT_SANITIZE_CONFIG:
1626 		name = "Sanitize Config";
1627 		break;
1628 	case NVME_FEAT_ENDURANCE_GROUP_EVENT_CONFIGURATION:
1629 		name = "Endurance Group Event Configuration";
1630 		break;
1631 	case NVME_FEAT_SOFTWARE_PROGRESS_MARKER:
1632 		name = "Software Progress Marker";
1633 		break;
1634 	case NVME_FEAT_HOST_IDENTIFIER:
1635 		name = "Host Identifier";
1636 		break;
1637 	case NVME_FEAT_RESERVATION_NOTIFICATION_MASK:
1638 		name = "Reservation Notification Mask";
1639 		break;
1640 	case NVME_FEAT_RESERVATION_PERSISTENCE:
1641 		name = "Reservation Persistence";
1642 		break;
1643 	case NVME_FEAT_NAMESPACE_WRITE_PROTECTION_CONFIG:
1644 		name = "Namespace Write Protection Config";
1645 		break;
1646 	default:
1647 		name = "Unknown";
1648 		break;
1649 	}
1650 
1651 	return (name);
1652 }
1653 
1654 static void
1655 nvme_feature_invalid_cb(struct pci_nvme_softc *sc __unused,
1656     struct nvme_feature_obj *feat __unused,
1657     struct nvme_command *command __unused,
1658     struct nvme_completion *compl)
1659 {
1660 	pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1661 }
1662 
1663 static void
1664 nvme_feature_iv_config(struct pci_nvme_softc *sc,
1665     struct nvme_feature_obj *feat __unused,
1666     struct nvme_command *command,
1667     struct nvme_completion *compl)
1668 {
1669 	uint32_t i;
1670 	uint32_t cdw11 = command->cdw11;
1671 	uint16_t iv;
1672 	bool cd;
1673 
1674 	pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1675 
1676 	iv = cdw11 & 0xffff;
1677 	cd = cdw11 & (1 << 16);
1678 
1679 	if (iv > (sc->max_queues + 1)) {
1680 		return;
1681 	}
1682 
1683 	/* No Interrupt Coalescing (i.e. not Coalescing Disable) for Admin Q */
1684 	if ((iv == 0) && !cd)
1685 		return;
1686 
1687 	/* Requested Interrupt Vector must be used by a CQ */
1688 	for (i = 0; i < sc->num_cqueues + 1; i++) {
1689 		if (sc->compl_queues[i].intr_vec == iv) {
1690 			pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1691 		}
1692 	}
1693 }
1694 
1695 #define NVME_ASYNC_EVENT_ENDURANCE_GROUP		(0x4000)
1696 static void
1697 nvme_feature_async_event(struct pci_nvme_softc *sc __unused,
1698     struct nvme_feature_obj *feat __unused,
1699     struct nvme_command *command,
1700     struct nvme_completion *compl)
1701 {
1702 	if (command->cdw11 & NVME_ASYNC_EVENT_ENDURANCE_GROUP)
1703 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1704 }
1705 
1706 #define NVME_TEMP_THRESH_OVER	0
1707 #define NVME_TEMP_THRESH_UNDER	1
1708 static void
1709 nvme_feature_temperature(struct pci_nvme_softc *sc,
1710     struct nvme_feature_obj *feat __unused,
1711     struct nvme_command *command,
1712     struct nvme_completion *compl)
1713 {
1714 	uint16_t	tmpth;	/* Temperature Threshold */
1715 	uint8_t		tmpsel; /* Threshold Temperature Select */
1716 	uint8_t		thsel;  /* Threshold Type Select */
1717 	bool		set_crit = false;
1718 	bool		report_crit;
1719 
1720 	tmpth  = command->cdw11 & 0xffff;
1721 	tmpsel = (command->cdw11 >> 16) & 0xf;
1722 	thsel  = (command->cdw11 >> 20) & 0x3;
1723 
1724 	DPRINTF("%s: tmpth=%#x tmpsel=%#x thsel=%#x", __func__, tmpth, tmpsel, thsel);
1725 
1726 	/* Check for unsupported values */
1727 	if (((tmpsel != 0) && (tmpsel != 0xf)) ||
1728 	    (thsel > NVME_TEMP_THRESH_UNDER)) {
1729 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1730 		return;
1731 	}
1732 
1733 	if (((thsel == NVME_TEMP_THRESH_OVER)  && (NVME_TEMPERATURE >= tmpth)) ||
1734 	    ((thsel == NVME_TEMP_THRESH_UNDER) && (NVME_TEMPERATURE <= tmpth)))
1735 		set_crit = true;
1736 
1737 	pthread_mutex_lock(&sc->mtx);
1738 	if (set_crit)
1739 		sc->health_log.critical_warning |=
1740 		    NVME_CRIT_WARN_ST_TEMPERATURE;
1741 	else
1742 		sc->health_log.critical_warning &=
1743 		    ~NVME_CRIT_WARN_ST_TEMPERATURE;
1744 	pthread_mutex_unlock(&sc->mtx);
1745 
1746 	report_crit = sc->feat[NVME_FEAT_ASYNC_EVENT_CONFIGURATION].cdw11 &
1747 	    NVME_CRIT_WARN_ST_TEMPERATURE;
1748 
1749 	if (set_crit && report_crit)
1750 		pci_nvme_aen_post(sc, PCI_NVME_AE_TYPE_SMART,
1751 		    sc->health_log.critical_warning);
1752 
1753 	DPRINTF("%s: set_crit=%c critical_warning=%#x status=%#x", __func__, set_crit ? 'T':'F', sc->health_log.critical_warning, compl->status);
1754 }
1755 
1756 static void
1757 nvme_feature_num_queues(struct pci_nvme_softc *sc,
1758     struct nvme_feature_obj *feat __unused,
1759     struct nvme_command *command,
1760     struct nvme_completion *compl)
1761 {
1762 	uint16_t nqr;	/* Number of Queues Requested */
1763 
1764 	if (sc->num_q_is_set) {
1765 		WPRINTF("%s: Number of Queues already set", __func__);
1766 		pci_nvme_status_genc(&compl->status,
1767 		    NVME_SC_COMMAND_SEQUENCE_ERROR);
1768 		return;
1769 	}
1770 
1771 	nqr = command->cdw11 & 0xFFFF;
1772 	if (nqr == 0xffff) {
1773 		WPRINTF("%s: Illegal NSQR value %#x", __func__, nqr);
1774 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1775 		return;
1776 	}
1777 
1778 	sc->num_squeues = ONE_BASED(nqr);
1779 	if (sc->num_squeues > sc->max_queues) {
1780 		DPRINTF("NSQR=%u is greater than max %u", sc->num_squeues,
1781 					sc->max_queues);
1782 		sc->num_squeues = sc->max_queues;
1783 	}
1784 
1785 	nqr = (command->cdw11 >> 16) & 0xFFFF;
1786 	if (nqr == 0xffff) {
1787 		WPRINTF("%s: Illegal NCQR value %#x", __func__, nqr);
1788 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1789 		return;
1790 	}
1791 
1792 	sc->num_cqueues = ONE_BASED(nqr);
1793 	if (sc->num_cqueues > sc->max_queues) {
1794 		DPRINTF("NCQR=%u is greater than max %u", sc->num_cqueues,
1795 					sc->max_queues);
1796 		sc->num_cqueues = sc->max_queues;
1797 	}
1798 
1799 	/* Patch the command value which will be saved on callback's return */
1800 	command->cdw11 = NVME_FEATURE_NUM_QUEUES(sc);
1801 	compl->cdw0 = NVME_FEATURE_NUM_QUEUES(sc);
1802 
1803 	sc->num_q_is_set = true;
1804 }
1805 
1806 static int
1807 nvme_opc_set_features(struct pci_nvme_softc *sc, struct nvme_command *command,
1808 	struct nvme_completion *compl)
1809 {
1810 	struct nvme_feature_obj *feat;
1811 	uint32_t nsid = command->nsid;
1812 	uint8_t fid = NVMEV(NVME_FEAT_SET_FID, command->cdw10);
1813 	bool sv = NVMEV(NVME_FEAT_SET_SV, command->cdw10);
1814 
1815 	DPRINTF("%s: Feature ID 0x%x (%s)", __func__, fid, nvme_fid_to_name(fid));
1816 
1817 	if (fid >= NVME_FID_MAX) {
1818 		DPRINTF("%s invalid feature 0x%x", __func__, fid);
1819 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1820 		return (1);
1821 	}
1822 
1823 	if (sv) {
1824 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1825 		    NVME_SC_FEATURE_NOT_SAVEABLE);
1826 		return (1);
1827 	}
1828 
1829 	feat = &sc->feat[fid];
1830 
1831 	if (feat->namespace_specific && (nsid == NVME_GLOBAL_NAMESPACE_TAG)) {
1832 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1833 		return (1);
1834 	}
1835 
1836 	if (!feat->namespace_specific &&
1837 	    !((nsid == 0) || (nsid == NVME_GLOBAL_NAMESPACE_TAG))) {
1838 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1839 		    NVME_SC_FEATURE_NOT_NS_SPECIFIC);
1840 		return (1);
1841 	}
1842 
1843 	compl->cdw0 = 0;
1844 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1845 
1846 	if (feat->set)
1847 		feat->set(sc, feat, command, compl);
1848 	else {
1849 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1850 		    NVME_SC_FEATURE_NOT_CHANGEABLE);
1851 		return (1);
1852 	}
1853 
1854 	DPRINTF("%s: status=%#x cdw11=%#x", __func__, compl->status, command->cdw11);
1855 	if (compl->status == NVME_SC_SUCCESS) {
1856 		feat->cdw11 = command->cdw11;
1857 		if ((fid == NVME_FEAT_ASYNC_EVENT_CONFIGURATION) &&
1858 		    (command->cdw11 != 0))
1859 			pci_nvme_aen_notify(sc);
1860 	}
1861 
1862 	return (0);
1863 }
1864 
1865 #define NVME_FEATURES_SEL_SUPPORTED	0x3
1866 #define NVME_FEATURES_NS_SPECIFIC	(1 << 1)
1867 
1868 static int
1869 nvme_opc_get_features(struct pci_nvme_softc* sc, struct nvme_command* command,
1870 	struct nvme_completion* compl)
1871 {
1872 	struct nvme_feature_obj *feat;
1873 	uint8_t fid = command->cdw10 & 0xFF;
1874 	uint8_t sel = (command->cdw10 >> 8) & 0x7;
1875 
1876 	DPRINTF("%s: Feature ID 0x%x (%s)", __func__, fid, nvme_fid_to_name(fid));
1877 
1878 	if (fid >= NVME_FID_MAX) {
1879 		DPRINTF("%s invalid feature 0x%x", __func__, fid);
1880 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1881 		return (1);
1882 	}
1883 
1884 	compl->cdw0 = 0;
1885 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1886 
1887 	feat = &sc->feat[fid];
1888 	if (feat->get) {
1889 		feat->get(sc, feat, command, compl);
1890 	}
1891 
1892 	if (compl->status == NVME_SC_SUCCESS) {
1893 		if ((sel == NVME_FEATURES_SEL_SUPPORTED) && feat->namespace_specific)
1894 			compl->cdw0 = NVME_FEATURES_NS_SPECIFIC;
1895 		else
1896 			compl->cdw0 = feat->cdw11;
1897 	}
1898 
1899 	return (0);
1900 }
1901 
1902 static int
1903 nvme_opc_format_nvm(struct pci_nvme_softc* sc, struct nvme_command* command,
1904 	struct nvme_completion* compl)
1905 {
1906 	uint8_t	ses, lbaf, pi;
1907 
1908 	/* Only supports Secure Erase Setting - User Data Erase */
1909 	ses = (command->cdw10 >> 9) & 0x7;
1910 	if (ses > 0x1) {
1911 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1912 		return (1);
1913 	}
1914 
1915 	/* Only supports a single LBA Format */
1916 	lbaf = command->cdw10 & 0xf;
1917 	if (lbaf != 0) {
1918 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1919 		    NVME_SC_INVALID_FORMAT);
1920 		return (1);
1921 	}
1922 
1923 	/* Doesn't support Protection Information */
1924 	pi = (command->cdw10 >> 5) & 0x7;
1925 	if (pi != 0) {
1926 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1927 		return (1);
1928 	}
1929 
1930 	if (sc->nvstore.type == NVME_STOR_RAM) {
1931 		if (sc->nvstore.ctx)
1932 			free(sc->nvstore.ctx);
1933 		sc->nvstore.ctx = calloc(1, sc->nvstore.size);
1934 		pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1935 	} else {
1936 		struct pci_nvme_ioreq *req;
1937 		int err;
1938 
1939 		req = pci_nvme_get_ioreq(sc);
1940 		if (req == NULL) {
1941 			pci_nvme_status_genc(&compl->status,
1942 			    NVME_SC_INTERNAL_DEVICE_ERROR);
1943 			WPRINTF("%s: unable to allocate IO req", __func__);
1944 			return (1);
1945 		}
1946 		req->nvme_sq = &sc->submit_queues[0];
1947 		req->sqid = 0;
1948 		req->opc = command->opc;
1949 		req->cid = command->cid;
1950 		req->nsid = command->nsid;
1951 
1952 		req->io_req.br_offset = 0;
1953 		req->io_req.br_resid = sc->nvstore.size;
1954 		req->io_req.br_callback = pci_nvme_io_done;
1955 
1956 		err = blockif_delete(sc->nvstore.ctx, &req->io_req);
1957 		if (err) {
1958 			pci_nvme_status_genc(&compl->status,
1959 			    NVME_SC_INTERNAL_DEVICE_ERROR);
1960 			pci_nvme_release_ioreq(sc, req);
1961 		} else
1962 			compl->status = NVME_NO_STATUS;
1963 	}
1964 
1965 	return (1);
1966 }
1967 
1968 static int
1969 nvme_opc_abort(struct pci_nvme_softc *sc __unused, struct nvme_command *command,
1970     struct nvme_completion *compl)
1971 {
1972 	DPRINTF("%s submission queue %u, command ID 0x%x", __func__,
1973 	        command->cdw10 & 0xFFFF, (command->cdw10 >> 16) & 0xFFFF);
1974 
1975 	/* TODO: search for the command ID and abort it */
1976 
1977 	compl->cdw0 = 1;
1978 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1979 	return (1);
1980 }
1981 
1982 static int
1983 nvme_opc_async_event_req(struct pci_nvme_softc* sc,
1984 	struct nvme_command* command, struct nvme_completion* compl)
1985 {
1986 	DPRINTF("%s async event request count=%u aerl=%u cid=%#x", __func__,
1987 	    sc->aer_count, sc->ctrldata.aerl, command->cid);
1988 
1989 	/* Don't exceed the Async Event Request Limit (AERL). */
1990 	if (pci_nvme_aer_limit_reached(sc)) {
1991 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1992 				NVME_SC_ASYNC_EVENT_REQUEST_LIMIT_EXCEEDED);
1993 		return (1);
1994 	}
1995 
1996 	if (pci_nvme_aer_add(sc, command->cid)) {
1997 		pci_nvme_status_tc(&compl->status, NVME_SCT_GENERIC,
1998 				NVME_SC_INTERNAL_DEVICE_ERROR);
1999 		return (1);
2000 	}
2001 
2002 	/*
2003 	 * Raise events when they happen based on the Set Features cmd.
2004 	 * These events happen async, so only set completion successful if
2005 	 * there is an event reflective of the request to get event.
2006 	 */
2007 	compl->status = NVME_NO_STATUS;
2008 	pci_nvme_aen_notify(sc);
2009 
2010 	return (0);
2011 }
2012 
2013 static void
2014 pci_nvme_handle_admin_cmd(struct pci_nvme_softc* sc, uint64_t value)
2015 {
2016 	struct nvme_completion compl;
2017 	struct nvme_command *cmd;
2018 	struct nvme_submission_queue *sq;
2019 	struct nvme_completion_queue *cq;
2020 	uint16_t sqhead;
2021 
2022 	DPRINTF("%s index %u", __func__, (uint32_t)value);
2023 
2024 	sq = &sc->submit_queues[0];
2025 	cq = &sc->compl_queues[0];
2026 
2027 	pthread_mutex_lock(&sq->mtx);
2028 
2029 	sqhead = sq->head;
2030 	DPRINTF("sqhead %u, tail %u", sqhead, sq->tail);
2031 
2032 	while (sqhead != atomic_load_acq_short(&sq->tail)) {
2033 		cmd = &(sq->qbase)[sqhead];
2034 		compl.cdw0 = 0;
2035 		compl.status = 0;
2036 
2037 		switch (cmd->opc) {
2038 		case NVME_OPC_DELETE_IO_SQ:
2039 			DPRINTF("%s command DELETE_IO_SQ", __func__);
2040 			nvme_opc_delete_io_sq(sc, cmd, &compl);
2041 			break;
2042 		case NVME_OPC_CREATE_IO_SQ:
2043 			DPRINTF("%s command CREATE_IO_SQ", __func__);
2044 			nvme_opc_create_io_sq(sc, cmd, &compl);
2045 			break;
2046 		case NVME_OPC_DELETE_IO_CQ:
2047 			DPRINTF("%s command DELETE_IO_CQ", __func__);
2048 			nvme_opc_delete_io_cq(sc, cmd, &compl);
2049 			break;
2050 		case NVME_OPC_CREATE_IO_CQ:
2051 			DPRINTF("%s command CREATE_IO_CQ", __func__);
2052 			nvme_opc_create_io_cq(sc, cmd, &compl);
2053 			break;
2054 		case NVME_OPC_GET_LOG_PAGE:
2055 			DPRINTF("%s command GET_LOG_PAGE", __func__);
2056 			nvme_opc_get_log_page(sc, cmd, &compl);
2057 			break;
2058 		case NVME_OPC_IDENTIFY:
2059 			DPRINTF("%s command IDENTIFY", __func__);
2060 			nvme_opc_identify(sc, cmd, &compl);
2061 			break;
2062 		case NVME_OPC_ABORT:
2063 			DPRINTF("%s command ABORT", __func__);
2064 			nvme_opc_abort(sc, cmd, &compl);
2065 			break;
2066 		case NVME_OPC_SET_FEATURES:
2067 			DPRINTF("%s command SET_FEATURES", __func__);
2068 			nvme_opc_set_features(sc, cmd, &compl);
2069 			break;
2070 		case NVME_OPC_GET_FEATURES:
2071 			DPRINTF("%s command GET_FEATURES", __func__);
2072 			nvme_opc_get_features(sc, cmd, &compl);
2073 			break;
2074 		case NVME_OPC_FIRMWARE_ACTIVATE:
2075 			DPRINTF("%s command FIRMWARE_ACTIVATE", __func__);
2076 			pci_nvme_status_tc(&compl.status,
2077 			    NVME_SCT_COMMAND_SPECIFIC,
2078 			    NVME_SC_INVALID_FIRMWARE_SLOT);
2079 			break;
2080 		case NVME_OPC_ASYNC_EVENT_REQUEST:
2081 			DPRINTF("%s command ASYNC_EVENT_REQ", __func__);
2082 			nvme_opc_async_event_req(sc, cmd, &compl);
2083 			break;
2084 		case NVME_OPC_FORMAT_NVM:
2085 			DPRINTF("%s command FORMAT_NVM", __func__);
2086 			if ((sc->ctrldata.oacs &
2087 			    (1 << NVME_CTRLR_DATA_OACS_FORMAT_SHIFT)) == 0) {
2088 				pci_nvme_status_genc(&compl.status, NVME_SC_INVALID_OPCODE);
2089 				break;
2090 			}
2091 			nvme_opc_format_nvm(sc, cmd, &compl);
2092 			break;
2093 		case NVME_OPC_SECURITY_SEND:
2094 		case NVME_OPC_SECURITY_RECEIVE:
2095 		case NVME_OPC_SANITIZE:
2096 		case NVME_OPC_GET_LBA_STATUS:
2097 			DPRINTF("%s command OPC=%#x (unsupported)", __func__,
2098 			    cmd->opc);
2099 			/* Valid but unsupported opcodes */
2100 			pci_nvme_status_genc(&compl.status, NVME_SC_INVALID_FIELD);
2101 			break;
2102 		default:
2103 			DPRINTF("%s command OPC=%#X (not implemented)",
2104 			    __func__,
2105 			    cmd->opc);
2106 			pci_nvme_status_genc(&compl.status, NVME_SC_INVALID_OPCODE);
2107 		}
2108 		sqhead = (sqhead + 1) % sq->size;
2109 
2110 		if (NVME_COMPLETION_VALID(compl)) {
2111 			pci_nvme_cq_update(sc, &sc->compl_queues[0],
2112 			    compl.cdw0,
2113 			    cmd->cid,
2114 			    0,		/* SQID */
2115 			    compl.status);
2116 		}
2117 	}
2118 
2119 	DPRINTF("setting sqhead %u", sqhead);
2120 	sq->head = sqhead;
2121 
2122 	if (cq->head != cq->tail)
2123 		pci_generate_msix(sc->nsc_pi, 0);
2124 
2125 	pthread_mutex_unlock(&sq->mtx);
2126 }
2127 
2128 /*
2129  * Update the Write and Read statistics reported in SMART data
2130  *
2131  * NVMe defines "data unit" as thousand's of 512 byte blocks and is rounded up.
2132  * E.g. 1 data unit is 1 - 1,000 512 byte blocks. 3 data units are 2,001 - 3,000
2133  * 512 byte blocks. Rounding up is achieved by initializing the remainder to 999.
2134  */
2135 static void
2136 pci_nvme_stats_write_read_update(struct pci_nvme_softc *sc, uint8_t opc,
2137     size_t bytes, uint16_t status)
2138 {
2139 
2140 	pthread_mutex_lock(&sc->mtx);
2141 	switch (opc) {
2142 	case NVME_OPC_WRITE:
2143 		sc->write_commands++;
2144 		if (status != NVME_SC_SUCCESS)
2145 			break;
2146 		sc->write_dunits_remainder += (bytes / 512);
2147 		while (sc->write_dunits_remainder >= 1000) {
2148 			sc->write_data_units++;
2149 			sc->write_dunits_remainder -= 1000;
2150 		}
2151 		break;
2152 	case NVME_OPC_READ:
2153 		sc->read_commands++;
2154 		if (status != NVME_SC_SUCCESS)
2155 			break;
2156 		sc->read_dunits_remainder += (bytes / 512);
2157 		while (sc->read_dunits_remainder >= 1000) {
2158 			sc->read_data_units++;
2159 			sc->read_dunits_remainder -= 1000;
2160 		}
2161 		break;
2162 	default:
2163 		DPRINTF("%s: Invalid OPC 0x%02x for stats", __func__, opc);
2164 		break;
2165 	}
2166 	pthread_mutex_unlock(&sc->mtx);
2167 }
2168 
2169 /*
2170  * Check if the combination of Starting LBA (slba) and number of blocks
2171  * exceeds the range of the underlying storage.
2172  *
2173  * Because NVMe specifies the SLBA in blocks as a uint64_t and blockif stores
2174  * the capacity in bytes as a uint64_t, care must be taken to avoid integer
2175  * overflow.
2176  */
2177 static bool
2178 pci_nvme_out_of_range(struct pci_nvme_blockstore *nvstore, uint64_t slba,
2179     uint32_t nblocks)
2180 {
2181 	size_t	offset, bytes;
2182 
2183 	/* Overflow check of multiplying Starting LBA by the sector size */
2184 	if (slba >> (64 - nvstore->sectsz_bits))
2185 		return (true);
2186 
2187 	offset = slba << nvstore->sectsz_bits;
2188 	bytes = nblocks << nvstore->sectsz_bits;
2189 
2190 	/* Overflow check of Number of Logical Blocks */
2191 	if ((nvstore->size <= offset) || ((nvstore->size - offset) < bytes))
2192 		return (true);
2193 
2194 	return (false);
2195 }
2196 
2197 static int
2198 pci_nvme_append_iov_req(struct pci_nvme_softc *sc __unused,
2199     struct pci_nvme_ioreq *req, uint64_t gpaddr, size_t size, uint64_t offset)
2200 {
2201 	int iovidx;
2202 	bool range_is_contiguous;
2203 
2204 	if (req == NULL)
2205 		return (-1);
2206 
2207 	if (req->io_req.br_iovcnt == NVME_MAX_IOVEC) {
2208 		return (-1);
2209 	}
2210 
2211 	/*
2212 	 * Minimize the number of IOVs by concatenating contiguous address
2213 	 * ranges. If the IOV count is zero, there is no previous range to
2214 	 * concatenate.
2215 	 */
2216 	if (req->io_req.br_iovcnt == 0)
2217 		range_is_contiguous = false;
2218 	else
2219 		range_is_contiguous = (req->prev_gpaddr + req->prev_size) == gpaddr;
2220 
2221 	if (range_is_contiguous) {
2222 		iovidx = req->io_req.br_iovcnt - 1;
2223 
2224 		req->io_req.br_iov[iovidx].iov_base =
2225 		    paddr_guest2host(req->sc->nsc_pi->pi_vmctx,
2226 				     req->prev_gpaddr, size);
2227 		if (req->io_req.br_iov[iovidx].iov_base == NULL)
2228 			return (-1);
2229 
2230 		req->prev_size += size;
2231 		req->io_req.br_resid += size;
2232 
2233 		req->io_req.br_iov[iovidx].iov_len = req->prev_size;
2234 	} else {
2235 		iovidx = req->io_req.br_iovcnt;
2236 		if (iovidx == 0) {
2237 			req->io_req.br_offset = offset;
2238 			req->io_req.br_resid = 0;
2239 			req->io_req.br_param = req;
2240 		}
2241 
2242 		req->io_req.br_iov[iovidx].iov_base =
2243 		    paddr_guest2host(req->sc->nsc_pi->pi_vmctx,
2244 				     gpaddr, size);
2245 		if (req->io_req.br_iov[iovidx].iov_base == NULL)
2246 			return (-1);
2247 
2248 		req->io_req.br_iov[iovidx].iov_len = size;
2249 
2250 		req->prev_gpaddr = gpaddr;
2251 		req->prev_size = size;
2252 		req->io_req.br_resid += size;
2253 
2254 		req->io_req.br_iovcnt++;
2255 	}
2256 
2257 	return (0);
2258 }
2259 
2260 static void
2261 pci_nvme_set_completion(struct pci_nvme_softc *sc,
2262     struct nvme_submission_queue *sq, int sqid, uint16_t cid, uint16_t status)
2263 {
2264 	struct nvme_completion_queue *cq = &sc->compl_queues[sq->cqid];
2265 
2266 	DPRINTF("%s sqid %d cqid %u cid %u status: 0x%x 0x%x",
2267 		 __func__, sqid, sq->cqid, cid, NVME_STATUS_GET_SCT(status),
2268 		 NVME_STATUS_GET_SC(status));
2269 
2270 	pci_nvme_cq_update(sc, cq, 0, cid, sqid, status);
2271 
2272 	if (cq->head != cq->tail) {
2273 		if (cq->intr_en & NVME_CQ_INTEN) {
2274 			pci_generate_msix(sc->nsc_pi, cq->intr_vec);
2275 		} else {
2276 			DPRINTF("%s: CQ%u interrupt disabled",
2277 						__func__, sq->cqid);
2278 		}
2279 	}
2280 }
2281 
2282 static void
2283 pci_nvme_release_ioreq(struct pci_nvme_softc *sc, struct pci_nvme_ioreq *req)
2284 {
2285 	req->sc = NULL;
2286 	req->nvme_sq = NULL;
2287 	req->sqid = 0;
2288 
2289 	pthread_mutex_lock(&sc->mtx);
2290 
2291 	STAILQ_INSERT_TAIL(&sc->ioreqs_free, req, link);
2292 	sc->pending_ios--;
2293 
2294 	/* when no more IO pending, can set to ready if device reset/enabled */
2295 	if (sc->pending_ios == 0 &&
2296 	    NVME_CC_GET_EN(sc->regs.cc) && !(NVME_CSTS_GET_RDY(sc->regs.csts)))
2297 		sc->regs.csts |= NVME_CSTS_RDY;
2298 
2299 	pthread_mutex_unlock(&sc->mtx);
2300 
2301 	sem_post(&sc->iosemlock);
2302 }
2303 
2304 static struct pci_nvme_ioreq *
2305 pci_nvme_get_ioreq(struct pci_nvme_softc *sc)
2306 {
2307 	struct pci_nvme_ioreq *req = NULL;
2308 
2309 	sem_wait(&sc->iosemlock);
2310 	pthread_mutex_lock(&sc->mtx);
2311 
2312 	req = STAILQ_FIRST(&sc->ioreqs_free);
2313 	assert(req != NULL);
2314 	STAILQ_REMOVE_HEAD(&sc->ioreqs_free, link);
2315 
2316 	req->sc = sc;
2317 
2318 	sc->pending_ios++;
2319 
2320 	pthread_mutex_unlock(&sc->mtx);
2321 
2322 	req->io_req.br_iovcnt = 0;
2323 	req->io_req.br_offset = 0;
2324 	req->io_req.br_resid = 0;
2325 	req->io_req.br_param = req;
2326 	req->prev_gpaddr = 0;
2327 	req->prev_size = 0;
2328 
2329 	return req;
2330 }
2331 
2332 static void
2333 pci_nvme_io_done(struct blockif_req *br, int err)
2334 {
2335 	struct pci_nvme_ioreq *req = br->br_param;
2336 	struct nvme_submission_queue *sq = req->nvme_sq;
2337 	uint16_t code, status;
2338 
2339 	DPRINTF("%s error %d %s", __func__, err, strerror(err));
2340 
2341 	/* TODO return correct error */
2342 	code = err ? NVME_SC_DATA_TRANSFER_ERROR : NVME_SC_SUCCESS;
2343 	status = 0;
2344 	pci_nvme_status_genc(&status, code);
2345 
2346 	pci_nvme_set_completion(req->sc, sq, req->sqid, req->cid, status);
2347 	pci_nvme_stats_write_read_update(req->sc, req->opc,
2348 	    req->bytes, status);
2349 	pci_nvme_release_ioreq(req->sc, req);
2350 }
2351 
2352 /*
2353  * Implements the Flush command. The specification states:
2354  *    If a volatile write cache is not present, Flush commands complete
2355  *    successfully and have no effect
2356  * in the description of the Volatile Write Cache (VWC) field of the Identify
2357  * Controller data. Therefore, set status to Success if the command is
2358  * not supported (i.e. RAM or as indicated by the blockif).
2359  */
2360 static bool
2361 nvme_opc_flush(struct pci_nvme_softc *sc __unused,
2362     struct nvme_command *cmd __unused,
2363     struct pci_nvme_blockstore *nvstore,
2364     struct pci_nvme_ioreq *req,
2365     uint16_t *status)
2366 {
2367 	bool pending = false;
2368 
2369 	if (nvstore->type == NVME_STOR_RAM) {
2370 		pci_nvme_status_genc(status, NVME_SC_SUCCESS);
2371 	} else {
2372 		int err;
2373 
2374 		req->io_req.br_callback = pci_nvme_io_done;
2375 
2376 		err = blockif_flush(nvstore->ctx, &req->io_req);
2377 		switch (err) {
2378 		case 0:
2379 			pending = true;
2380 			break;
2381 		case EOPNOTSUPP:
2382 			pci_nvme_status_genc(status, NVME_SC_SUCCESS);
2383 			break;
2384 		default:
2385 			pci_nvme_status_genc(status, NVME_SC_INTERNAL_DEVICE_ERROR);
2386 		}
2387 	}
2388 
2389 	return (pending);
2390 }
2391 
2392 static uint16_t
2393 nvme_write_read_ram(struct pci_nvme_softc *sc,
2394     struct pci_nvme_blockstore *nvstore,
2395     uint64_t prp1, uint64_t prp2,
2396     size_t offset, uint64_t bytes,
2397     bool is_write)
2398 {
2399 	uint8_t *buf = nvstore->ctx;
2400 	enum nvme_copy_dir dir;
2401 	uint16_t status;
2402 
2403 	if (is_write)
2404 		dir = NVME_COPY_TO_PRP;
2405 	else
2406 		dir = NVME_COPY_FROM_PRP;
2407 
2408 	status = 0;
2409 	if (nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, prp1, prp2,
2410 	    buf + offset, bytes, dir))
2411 		pci_nvme_status_genc(&status,
2412 		    NVME_SC_DATA_TRANSFER_ERROR);
2413 	else
2414 		pci_nvme_status_genc(&status, NVME_SC_SUCCESS);
2415 
2416 	return (status);
2417 }
2418 
2419 static uint16_t
2420 nvme_write_read_blockif(struct pci_nvme_softc *sc,
2421     struct pci_nvme_blockstore *nvstore,
2422     struct pci_nvme_ioreq *req,
2423     uint64_t prp1, uint64_t prp2,
2424     size_t offset, uint64_t bytes,
2425     bool is_write)
2426 {
2427 	uint64_t size;
2428 	int err;
2429 	uint16_t status = NVME_NO_STATUS;
2430 
2431 	size = MIN(PAGE_SIZE - (prp1 % PAGE_SIZE), bytes);
2432 	if (pci_nvme_append_iov_req(sc, req, prp1, size, offset)) {
2433 		err = -1;
2434 		goto out;
2435 	}
2436 
2437 	offset += size;
2438 	bytes  -= size;
2439 
2440 	if (bytes == 0) {
2441 		;
2442 	} else if (bytes <= PAGE_SIZE) {
2443 		size = bytes;
2444 		if (pci_nvme_append_iov_req(sc, req, prp2, size, offset)) {
2445 			err = -1;
2446 			goto out;
2447 		}
2448 	} else {
2449 		void *vmctx = sc->nsc_pi->pi_vmctx;
2450 		uint64_t *prp_list = &prp2;
2451 		uint64_t *last = prp_list;
2452 
2453 		/* PRP2 is pointer to a physical region page list */
2454 		while (bytes) {
2455 			/* Last entry in list points to the next list */
2456 			if ((prp_list == last) && (bytes > PAGE_SIZE)) {
2457 				uint64_t prp = *prp_list;
2458 
2459 				prp_list = paddr_guest2host(vmctx, prp,
2460 				    PAGE_SIZE - (prp % PAGE_SIZE));
2461 				if (prp_list == NULL) {
2462 					err = -1;
2463 					goto out;
2464 				}
2465 				last = prp_list + (NVME_PRP2_ITEMS - 1);
2466 			}
2467 
2468 			size = MIN(bytes, PAGE_SIZE);
2469 
2470 			if (pci_nvme_append_iov_req(sc, req, *prp_list, size,
2471 			    offset)) {
2472 				err = -1;
2473 				goto out;
2474 			}
2475 
2476 			offset += size;
2477 			bytes  -= size;
2478 
2479 			prp_list++;
2480 		}
2481 	}
2482 	req->io_req.br_callback = pci_nvme_io_done;
2483 	if (is_write)
2484 		err = blockif_write(nvstore->ctx, &req->io_req);
2485 	else
2486 		err = blockif_read(nvstore->ctx, &req->io_req);
2487 out:
2488 	if (err)
2489 		pci_nvme_status_genc(&status, NVME_SC_DATA_TRANSFER_ERROR);
2490 
2491 	return (status);
2492 }
2493 
2494 static bool
2495 nvme_opc_write_read(struct pci_nvme_softc *sc,
2496     struct nvme_command *cmd,
2497     struct pci_nvme_blockstore *nvstore,
2498     struct pci_nvme_ioreq *req,
2499     uint16_t *status)
2500 {
2501 	uint64_t lba, nblocks, bytes;
2502 	size_t offset;
2503 	bool is_write = cmd->opc == NVME_OPC_WRITE;
2504 	bool pending = false;
2505 
2506 	lba = ((uint64_t)cmd->cdw11 << 32) | cmd->cdw10;
2507 	nblocks = (cmd->cdw12 & 0xFFFF) + 1;
2508 	bytes = nblocks << nvstore->sectsz_bits;
2509 	if (bytes > NVME_MAX_DATA_SIZE) {
2510 		WPRINTF("%s command would exceed MDTS", __func__);
2511 		pci_nvme_status_genc(status, NVME_SC_INVALID_FIELD);
2512 		goto out;
2513 	}
2514 
2515 	if (pci_nvme_out_of_range(nvstore, lba, nblocks)) {
2516 		WPRINTF("%s command would exceed LBA range(slba=%#lx nblocks=%#lx)",
2517 		    __func__, lba, nblocks);
2518 		pci_nvme_status_genc(status, NVME_SC_LBA_OUT_OF_RANGE);
2519 		goto out;
2520 	}
2521 
2522 	offset = lba << nvstore->sectsz_bits;
2523 
2524 	req->bytes = bytes;
2525 	req->io_req.br_offset = lba;
2526 
2527 	/* PRP bits 1:0 must be zero */
2528 	cmd->prp1 &= ~0x3UL;
2529 	cmd->prp2 &= ~0x3UL;
2530 
2531 	if (nvstore->type == NVME_STOR_RAM) {
2532 		*status = nvme_write_read_ram(sc, nvstore, cmd->prp1,
2533 		    cmd->prp2, offset, bytes, is_write);
2534 	} else {
2535 		*status = nvme_write_read_blockif(sc, nvstore, req,
2536 		    cmd->prp1, cmd->prp2, offset, bytes, is_write);
2537 
2538 		if (*status == NVME_NO_STATUS)
2539 			pending = true;
2540 	}
2541 out:
2542 	if (!pending)
2543 		pci_nvme_stats_write_read_update(sc, cmd->opc, bytes, *status);
2544 
2545 	return (pending);
2546 }
2547 
2548 static void
2549 pci_nvme_dealloc_sm(struct blockif_req *br, int err)
2550 {
2551 	struct pci_nvme_ioreq *req = br->br_param;
2552 	struct pci_nvme_softc *sc = req->sc;
2553 	bool done = true;
2554 	uint16_t status;
2555 
2556 	status = 0;
2557 	if (err) {
2558 		pci_nvme_status_genc(&status, NVME_SC_INTERNAL_DEVICE_ERROR);
2559 	} else if ((req->prev_gpaddr + 1) == (req->prev_size)) {
2560 		pci_nvme_status_genc(&status, NVME_SC_SUCCESS);
2561 	} else {
2562 		struct iovec *iov = req->io_req.br_iov;
2563 
2564 		req->prev_gpaddr++;
2565 		iov += req->prev_gpaddr;
2566 
2567 		/* The iov_* values already include the sector size */
2568 		req->io_req.br_offset = (off_t)iov->iov_base;
2569 		req->io_req.br_resid = iov->iov_len;
2570 		if (blockif_delete(sc->nvstore.ctx, &req->io_req)) {
2571 			pci_nvme_status_genc(&status,
2572 			    NVME_SC_INTERNAL_DEVICE_ERROR);
2573 		} else
2574 			done = false;
2575 	}
2576 
2577 	if (done) {
2578 		pci_nvme_set_completion(sc, req->nvme_sq, req->sqid, req->cid,
2579 		    status);
2580 		pci_nvme_release_ioreq(sc, req);
2581 	}
2582 }
2583 
2584 static bool
2585 nvme_opc_dataset_mgmt(struct pci_nvme_softc *sc,
2586     struct nvme_command *cmd,
2587     struct pci_nvme_blockstore *nvstore,
2588     struct pci_nvme_ioreq *req,
2589     uint16_t *status)
2590 {
2591 	struct nvme_dsm_range *range = NULL;
2592 	uint32_t nr, r, non_zero, dr;
2593 	int err;
2594 	bool pending = false;
2595 
2596 	if ((sc->ctrldata.oncs & NVME_ONCS_DSM) == 0) {
2597 		pci_nvme_status_genc(status, NVME_SC_INVALID_OPCODE);
2598 		goto out;
2599 	}
2600 
2601 	nr = cmd->cdw10 & 0xff;
2602 
2603 	/* copy locally because a range entry could straddle PRPs */
2604 	range = calloc(1, NVME_MAX_DSM_TRIM);
2605 	if (range == NULL) {
2606 		pci_nvme_status_genc(status, NVME_SC_INTERNAL_DEVICE_ERROR);
2607 		goto out;
2608 	}
2609 	nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, cmd->prp1, cmd->prp2,
2610 	    (uint8_t *)range, NVME_MAX_DSM_TRIM, NVME_COPY_FROM_PRP);
2611 
2612 	/* Check for invalid ranges and the number of non-zero lengths */
2613 	non_zero = 0;
2614 	for (r = 0; r <= nr; r++) {
2615 		if (pci_nvme_out_of_range(nvstore,
2616 		    range[r].starting_lba, range[r].length)) {
2617 			pci_nvme_status_genc(status, NVME_SC_LBA_OUT_OF_RANGE);
2618 			goto out;
2619 		}
2620 		if (range[r].length != 0)
2621 			non_zero++;
2622 	}
2623 
2624 	if (cmd->cdw11 & NVME_DSM_ATTR_DEALLOCATE) {
2625 		size_t offset, bytes;
2626 		int sectsz_bits = sc->nvstore.sectsz_bits;
2627 
2628 		/*
2629 		 * DSM calls are advisory only, and compliant controllers
2630 		 * may choose to take no actions (i.e. return Success).
2631 		 */
2632 		if (!nvstore->deallocate) {
2633 			pci_nvme_status_genc(status, NVME_SC_SUCCESS);
2634 			goto out;
2635 		}
2636 
2637 		/* If all ranges have a zero length, return Success */
2638 		if (non_zero == 0) {
2639 			pci_nvme_status_genc(status, NVME_SC_SUCCESS);
2640 			goto out;
2641 		}
2642 
2643 		if (req == NULL) {
2644 			pci_nvme_status_genc(status, NVME_SC_INTERNAL_DEVICE_ERROR);
2645 			goto out;
2646 		}
2647 
2648 		offset = range[0].starting_lba << sectsz_bits;
2649 		bytes = range[0].length << sectsz_bits;
2650 
2651 		/*
2652 		 * If the request is for more than a single range, store
2653 		 * the ranges in the br_iov. Optimize for the common case
2654 		 * of a single range.
2655 		 *
2656 		 * Note that NVMe Number of Ranges is a zero based value
2657 		 */
2658 		req->io_req.br_iovcnt = 0;
2659 		req->io_req.br_offset = offset;
2660 		req->io_req.br_resid = bytes;
2661 
2662 		if (nr == 0) {
2663 			req->io_req.br_callback = pci_nvme_io_done;
2664 		} else {
2665 			struct iovec *iov = req->io_req.br_iov;
2666 
2667 			for (r = 0, dr = 0; r <= nr; r++) {
2668 				offset = range[r].starting_lba << sectsz_bits;
2669 				bytes = range[r].length << sectsz_bits;
2670 				if (bytes == 0)
2671 					continue;
2672 
2673 				if ((nvstore->size - offset) < bytes) {
2674 					pci_nvme_status_genc(status,
2675 					    NVME_SC_LBA_OUT_OF_RANGE);
2676 					goto out;
2677 				}
2678 				iov[dr].iov_base = (void *)offset;
2679 				iov[dr].iov_len = bytes;
2680 				dr++;
2681 			}
2682 			req->io_req.br_callback = pci_nvme_dealloc_sm;
2683 
2684 			/*
2685 			 * Use prev_gpaddr to track the current entry and
2686 			 * prev_size to track the number of entries
2687 			 */
2688 			req->prev_gpaddr = 0;
2689 			req->prev_size = dr;
2690 		}
2691 
2692 		err = blockif_delete(nvstore->ctx, &req->io_req);
2693 		if (err)
2694 			pci_nvme_status_genc(status, NVME_SC_INTERNAL_DEVICE_ERROR);
2695 		else
2696 			pending = true;
2697 	}
2698 out:
2699 	free(range);
2700 	return (pending);
2701 }
2702 
2703 static void
2704 pci_nvme_handle_io_cmd(struct pci_nvme_softc* sc, uint16_t idx)
2705 {
2706 	struct nvme_submission_queue *sq;
2707 	uint16_t status;
2708 	uint16_t sqhead;
2709 
2710 	/* handle all submissions up to sq->tail index */
2711 	sq = &sc->submit_queues[idx];
2712 
2713 	pthread_mutex_lock(&sq->mtx);
2714 
2715 	sqhead = sq->head;
2716 	DPRINTF("nvme_handle_io qid %u head %u tail %u cmdlist %p",
2717 	         idx, sqhead, sq->tail, sq->qbase);
2718 
2719 	while (sqhead != atomic_load_acq_short(&sq->tail)) {
2720 		struct nvme_command *cmd;
2721 		struct pci_nvme_ioreq *req;
2722 		uint32_t nsid;
2723 		bool pending;
2724 
2725 		pending = false;
2726 		req = NULL;
2727 		status = 0;
2728 
2729 		cmd = &sq->qbase[sqhead];
2730 		sqhead = (sqhead + 1) % sq->size;
2731 
2732 		nsid = le32toh(cmd->nsid);
2733 		if ((nsid == 0) || (nsid > sc->ctrldata.nn)) {
2734 			pci_nvme_status_genc(&status,
2735 			    NVME_SC_INVALID_NAMESPACE_OR_FORMAT);
2736 			status |=
2737 			    NVME_STATUS_DNR_MASK << NVME_STATUS_DNR_SHIFT;
2738 			goto complete;
2739  		}
2740 
2741 		req = pci_nvme_get_ioreq(sc);
2742 		if (req == NULL) {
2743 			pci_nvme_status_genc(&status,
2744 			    NVME_SC_INTERNAL_DEVICE_ERROR);
2745 			WPRINTF("%s: unable to allocate IO req", __func__);
2746 			goto complete;
2747 		}
2748 		req->nvme_sq = sq;
2749 		req->sqid = idx;
2750 		req->opc = cmd->opc;
2751 		req->cid = cmd->cid;
2752 		req->nsid = cmd->nsid;
2753 
2754 		switch (cmd->opc) {
2755 		case NVME_OPC_FLUSH:
2756 			pending = nvme_opc_flush(sc, cmd, &sc->nvstore,
2757 			    req, &status);
2758  			break;
2759 		case NVME_OPC_WRITE:
2760 		case NVME_OPC_READ:
2761 			pending = nvme_opc_write_read(sc, cmd, &sc->nvstore,
2762 			    req, &status);
2763 			break;
2764 		case NVME_OPC_WRITE_ZEROES:
2765 			/* TODO: write zeroes
2766 			WPRINTF("%s write zeroes lba 0x%lx blocks %u",
2767 			        __func__, lba, cmd->cdw12 & 0xFFFF); */
2768 			pci_nvme_status_genc(&status, NVME_SC_SUCCESS);
2769 			break;
2770 		case NVME_OPC_DATASET_MANAGEMENT:
2771  			pending = nvme_opc_dataset_mgmt(sc, cmd, &sc->nvstore,
2772 			    req, &status);
2773 			break;
2774  		default:
2775  			WPRINTF("%s unhandled io command 0x%x",
2776 			    __func__, cmd->opc);
2777 			pci_nvme_status_genc(&status, NVME_SC_INVALID_OPCODE);
2778 		}
2779 complete:
2780 		if (!pending) {
2781 			pci_nvme_set_completion(sc, sq, idx, cmd->cid, status);
2782 			if (req != NULL)
2783 				pci_nvme_release_ioreq(sc, req);
2784 		}
2785 	}
2786 
2787 	sq->head = sqhead;
2788 
2789 	pthread_mutex_unlock(&sq->mtx);
2790 }
2791 
2792 static void
2793 pci_nvme_handle_doorbell(struct pci_nvme_softc* sc,
2794 	uint64_t idx, int is_sq, uint64_t value)
2795 {
2796 	DPRINTF("nvme doorbell %lu, %s, val 0x%lx",
2797 	        idx, is_sq ? "SQ" : "CQ", value & 0xFFFF);
2798 
2799 	if (is_sq) {
2800 		if (idx > sc->num_squeues) {
2801 			WPRINTF("%s queue index %lu overflow from "
2802 			         "guest (max %u)",
2803 			         __func__, idx, sc->num_squeues);
2804 			return;
2805 		}
2806 
2807 		atomic_store_short(&sc->submit_queues[idx].tail,
2808 		                   (uint16_t)value);
2809 
2810 		if (idx == 0) {
2811 			pci_nvme_handle_admin_cmd(sc, value);
2812 		} else {
2813 			/* submission queue; handle new entries in SQ */
2814 			if (idx > sc->num_squeues) {
2815 				WPRINTF("%s SQ index %lu overflow from "
2816 				         "guest (max %u)",
2817 				         __func__, idx, sc->num_squeues);
2818 				return;
2819 			}
2820 			pci_nvme_handle_io_cmd(sc, (uint16_t)idx);
2821 		}
2822 	} else {
2823 		if (idx > sc->num_cqueues) {
2824 			WPRINTF("%s queue index %lu overflow from "
2825 			         "guest (max %u)",
2826 			         __func__, idx, sc->num_cqueues);
2827 			return;
2828 		}
2829 
2830 		atomic_store_short(&sc->compl_queues[idx].head,
2831 				(uint16_t)value);
2832 	}
2833 }
2834 
2835 static void
2836 pci_nvme_bar0_reg_dumps(const char *func, uint64_t offset, int iswrite)
2837 {
2838 	const char *s = iswrite ? "WRITE" : "READ";
2839 
2840 	switch (offset) {
2841 	case NVME_CR_CAP_LOW:
2842 		DPRINTF("%s %s NVME_CR_CAP_LOW", func, s);
2843 		break;
2844 	case NVME_CR_CAP_HI:
2845 		DPRINTF("%s %s NVME_CR_CAP_HI", func, s);
2846 		break;
2847 	case NVME_CR_VS:
2848 		DPRINTF("%s %s NVME_CR_VS", func, s);
2849 		break;
2850 	case NVME_CR_INTMS:
2851 		DPRINTF("%s %s NVME_CR_INTMS", func, s);
2852 		break;
2853 	case NVME_CR_INTMC:
2854 		DPRINTF("%s %s NVME_CR_INTMC", func, s);
2855 		break;
2856 	case NVME_CR_CC:
2857 		DPRINTF("%s %s NVME_CR_CC", func, s);
2858 		break;
2859 	case NVME_CR_CSTS:
2860 		DPRINTF("%s %s NVME_CR_CSTS", func, s);
2861 		break;
2862 	case NVME_CR_NSSR:
2863 		DPRINTF("%s %s NVME_CR_NSSR", func, s);
2864 		break;
2865 	case NVME_CR_AQA:
2866 		DPRINTF("%s %s NVME_CR_AQA", func, s);
2867 		break;
2868 	case NVME_CR_ASQ_LOW:
2869 		DPRINTF("%s %s NVME_CR_ASQ_LOW", func, s);
2870 		break;
2871 	case NVME_CR_ASQ_HI:
2872 		DPRINTF("%s %s NVME_CR_ASQ_HI", func, s);
2873 		break;
2874 	case NVME_CR_ACQ_LOW:
2875 		DPRINTF("%s %s NVME_CR_ACQ_LOW", func, s);
2876 		break;
2877 	case NVME_CR_ACQ_HI:
2878 		DPRINTF("%s %s NVME_CR_ACQ_HI", func, s);
2879 		break;
2880 	default:
2881 		DPRINTF("unknown nvme bar-0 offset 0x%lx", offset);
2882 	}
2883 
2884 }
2885 
2886 static void
2887 pci_nvme_write_bar_0(struct pci_nvme_softc *sc, uint64_t offset, int size,
2888     uint64_t value)
2889 {
2890 	uint32_t ccreg;
2891 
2892 	if (offset >= NVME_DOORBELL_OFFSET) {
2893 		uint64_t belloffset = offset - NVME_DOORBELL_OFFSET;
2894 		uint64_t idx = belloffset / 8; /* door bell size = 2*int */
2895 		int is_sq = (belloffset % 8) < 4;
2896 
2897 		if ((sc->regs.csts & NVME_CSTS_RDY) == 0) {
2898 			WPRINTF("doorbell write prior to RDY (offset=%#lx)\n",
2899 			    offset);
2900 			return;
2901 		}
2902 
2903 		if (belloffset > ((sc->max_queues+1) * 8 - 4)) {
2904 			WPRINTF("guest attempted an overflow write offset "
2905 			         "0x%lx, val 0x%lx in %s",
2906 			         offset, value, __func__);
2907 			return;
2908 		}
2909 
2910 		if (is_sq) {
2911 			if (sc->submit_queues[idx].qbase == NULL)
2912 				return;
2913 		} else if (sc->compl_queues[idx].qbase == NULL)
2914 			return;
2915 
2916 		pci_nvme_handle_doorbell(sc, idx, is_sq, value);
2917 		return;
2918 	}
2919 
2920 	DPRINTF("nvme-write offset 0x%lx, size %d, value 0x%lx",
2921 	        offset, size, value);
2922 
2923 	if (size != 4) {
2924 		WPRINTF("guest wrote invalid size %d (offset 0x%lx, "
2925 		         "val 0x%lx) to bar0 in %s",
2926 		         size, offset, value, __func__);
2927 		/* TODO: shutdown device */
2928 		return;
2929 	}
2930 
2931 	pci_nvme_bar0_reg_dumps(__func__, offset, 1);
2932 
2933 	pthread_mutex_lock(&sc->mtx);
2934 
2935 	switch (offset) {
2936 	case NVME_CR_CAP_LOW:
2937 	case NVME_CR_CAP_HI:
2938 		/* readonly */
2939 		break;
2940 	case NVME_CR_VS:
2941 		/* readonly */
2942 		break;
2943 	case NVME_CR_INTMS:
2944 		/* MSI-X, so ignore */
2945 		break;
2946 	case NVME_CR_INTMC:
2947 		/* MSI-X, so ignore */
2948 		break;
2949 	case NVME_CR_CC:
2950 		ccreg = (uint32_t)value;
2951 
2952 		DPRINTF("%s NVME_CR_CC en %x css %x shn %x iosqes %u "
2953 		         "iocqes %u",
2954 		        __func__,
2955 			 NVME_CC_GET_EN(ccreg), NVME_CC_GET_CSS(ccreg),
2956 			 NVME_CC_GET_SHN(ccreg), NVME_CC_GET_IOSQES(ccreg),
2957 			 NVME_CC_GET_IOCQES(ccreg));
2958 
2959 		if (NVME_CC_GET_SHN(ccreg)) {
2960 			/* perform shutdown - flush out data to backend */
2961 			sc->regs.csts &= ~(NVME_CSTS_REG_SHST_MASK <<
2962 			    NVME_CSTS_REG_SHST_SHIFT);
2963 			sc->regs.csts |= NVME_SHST_COMPLETE <<
2964 			    NVME_CSTS_REG_SHST_SHIFT;
2965 		}
2966 		if (NVME_CC_GET_EN(ccreg) != NVME_CC_GET_EN(sc->regs.cc)) {
2967 			if (NVME_CC_GET_EN(ccreg) == 0)
2968 				/* transition 1-> causes controller reset */
2969 				pci_nvme_reset_locked(sc);
2970 			else
2971 				pci_nvme_init_controller(sc);
2972 		}
2973 
2974 		/* Insert the iocqes, iosqes and en bits from the write */
2975 		sc->regs.cc &= ~NVME_CC_WRITE_MASK;
2976 		sc->regs.cc |= ccreg & NVME_CC_WRITE_MASK;
2977 		if (NVME_CC_GET_EN(ccreg) == 0) {
2978 			/* Insert the ams, mps and css bit fields */
2979 			sc->regs.cc &= ~NVME_CC_NEN_WRITE_MASK;
2980 			sc->regs.cc |= ccreg & NVME_CC_NEN_WRITE_MASK;
2981 			sc->regs.csts &= ~NVME_CSTS_RDY;
2982 		} else if ((sc->pending_ios == 0) &&
2983 		    !(sc->regs.csts & NVME_CSTS_CFS)) {
2984 			sc->regs.csts |= NVME_CSTS_RDY;
2985 		}
2986 		break;
2987 	case NVME_CR_CSTS:
2988 		break;
2989 	case NVME_CR_NSSR:
2990 		/* ignore writes; don't support subsystem reset */
2991 		break;
2992 	case NVME_CR_AQA:
2993 		sc->regs.aqa = (uint32_t)value;
2994 		break;
2995 	case NVME_CR_ASQ_LOW:
2996 		sc->regs.asq = (sc->regs.asq & (0xFFFFFFFF00000000)) |
2997 		               (0xFFFFF000 & value);
2998 		break;
2999 	case NVME_CR_ASQ_HI:
3000 		sc->regs.asq = (sc->regs.asq & (0x00000000FFFFFFFF)) |
3001 		               (value << 32);
3002 		break;
3003 	case NVME_CR_ACQ_LOW:
3004 		sc->regs.acq = (sc->regs.acq & (0xFFFFFFFF00000000)) |
3005 		               (0xFFFFF000 & value);
3006 		break;
3007 	case NVME_CR_ACQ_HI:
3008 		sc->regs.acq = (sc->regs.acq & (0x00000000FFFFFFFF)) |
3009 		               (value << 32);
3010 		break;
3011 	default:
3012 		DPRINTF("%s unknown offset 0x%lx, value 0x%lx size %d",
3013 		         __func__, offset, value, size);
3014 	}
3015 	pthread_mutex_unlock(&sc->mtx);
3016 }
3017 
3018 static void
3019 pci_nvme_write(struct pci_devinst *pi, int baridx, uint64_t offset, int size,
3020     uint64_t value)
3021 {
3022 	struct pci_nvme_softc* sc = pi->pi_arg;
3023 
3024 	if (baridx == pci_msix_table_bar(pi) ||
3025 	    baridx == pci_msix_pba_bar(pi)) {
3026 		DPRINTF("nvme-write baridx %d, msix: off 0x%lx, size %d, "
3027 		         " value 0x%lx", baridx, offset, size, value);
3028 
3029 		pci_emul_msix_twrite(pi, offset, size, value);
3030 		return;
3031 	}
3032 
3033 	switch (baridx) {
3034 	case 0:
3035 		pci_nvme_write_bar_0(sc, offset, size, value);
3036 		break;
3037 
3038 	default:
3039 		DPRINTF("%s unknown baridx %d, val 0x%lx",
3040 		         __func__, baridx, value);
3041 	}
3042 }
3043 
3044 static uint64_t pci_nvme_read_bar_0(struct pci_nvme_softc* sc,
3045 	uint64_t offset, int size)
3046 {
3047 	uint64_t value;
3048 
3049 	pci_nvme_bar0_reg_dumps(__func__, offset, 0);
3050 
3051 	if (offset < NVME_DOORBELL_OFFSET) {
3052 		void *p = &(sc->regs);
3053 		pthread_mutex_lock(&sc->mtx);
3054 		memcpy(&value, (void *)((uintptr_t)p + offset), size);
3055 		pthread_mutex_unlock(&sc->mtx);
3056 	} else {
3057 		value = 0;
3058                 WPRINTF("pci_nvme: read invalid offset %ld", offset);
3059 	}
3060 
3061 	switch (size) {
3062 	case 1:
3063 		value &= 0xFF;
3064 		break;
3065 	case 2:
3066 		value &= 0xFFFF;
3067 		break;
3068 	case 4:
3069 		value &= 0xFFFFFFFF;
3070 		break;
3071 	}
3072 
3073 	DPRINTF("   nvme-read offset 0x%lx, size %d -> value 0x%x",
3074 	         offset, size, (uint32_t)value);
3075 
3076 	return (value);
3077 }
3078 
3079 
3080 
3081 static uint64_t
3082 pci_nvme_read(struct pci_devinst *pi, int baridx, uint64_t offset, int size)
3083 {
3084 	struct pci_nvme_softc* sc = pi->pi_arg;
3085 
3086 	if (baridx == pci_msix_table_bar(pi) ||
3087 	    baridx == pci_msix_pba_bar(pi)) {
3088 		DPRINTF("nvme-read bar: %d, msix: regoff 0x%lx, size %d",
3089 		        baridx, offset, size);
3090 
3091 		return pci_emul_msix_tread(pi, offset, size);
3092 	}
3093 
3094 	switch (baridx) {
3095 	case 0:
3096        		return pci_nvme_read_bar_0(sc, offset, size);
3097 
3098 	default:
3099 		DPRINTF("unknown bar %d, 0x%lx", baridx, offset);
3100 	}
3101 
3102 	return (0);
3103 }
3104 
3105 static int
3106 pci_nvme_parse_config(struct pci_nvme_softc *sc, nvlist_t *nvl)
3107 {
3108 	char bident[sizeof("XXX:XXX")];
3109 	const char *value;
3110 	uint32_t sectsz;
3111 
3112 	sc->max_queues = NVME_QUEUES;
3113 	sc->max_qentries = NVME_MAX_QENTRIES;
3114 	sc->ioslots = NVME_IOSLOTS;
3115 	sc->num_squeues = sc->max_queues;
3116 	sc->num_cqueues = sc->max_queues;
3117 	sc->dataset_management = NVME_DATASET_MANAGEMENT_AUTO;
3118 	sectsz = 0;
3119 	snprintf(sc->ctrldata.sn, sizeof(sc->ctrldata.sn),
3120 	         "NVME-%d-%d", sc->nsc_pi->pi_slot, sc->nsc_pi->pi_func);
3121 
3122 	value = get_config_value_node(nvl, "maxq");
3123 	if (value != NULL)
3124 		sc->max_queues = atoi(value);
3125 	value = get_config_value_node(nvl, "qsz");
3126 	if (value != NULL) {
3127 		sc->max_qentries = atoi(value);
3128 		if (sc->max_qentries <= 0) {
3129 			EPRINTLN("nvme: Invalid qsz option %d",
3130 			    sc->max_qentries);
3131 			return (-1);
3132 		}
3133 	}
3134 	value = get_config_value_node(nvl, "ioslots");
3135 	if (value != NULL) {
3136 		sc->ioslots = atoi(value);
3137 		if (sc->ioslots <= 0) {
3138 			EPRINTLN("Invalid ioslots option %d", sc->ioslots);
3139 			return (-1);
3140 		}
3141 	}
3142 	value = get_config_value_node(nvl, "sectsz");
3143 	if (value != NULL)
3144 		sectsz = atoi(value);
3145 	value = get_config_value_node(nvl, "ser");
3146 	if (value != NULL) {
3147 		/*
3148 		 * This field indicates the Product Serial Number in
3149 		 * 7-bit ASCII, unused bytes should be space characters.
3150 		 * Ref: NVMe v1.3c.
3151 		 */
3152 		cpywithpad((char *)sc->ctrldata.sn,
3153 		    sizeof(sc->ctrldata.sn), value, ' ');
3154 	}
3155 	value = get_config_value_node(nvl, "eui64");
3156 	if (value != NULL)
3157 		sc->nvstore.eui64 = htobe64(strtoull(value, NULL, 0));
3158 	value = get_config_value_node(nvl, "dsm");
3159 	if (value != NULL) {
3160 		if (strcmp(value, "auto") == 0)
3161 			sc->dataset_management = NVME_DATASET_MANAGEMENT_AUTO;
3162 		else if (strcmp(value, "enable") == 0)
3163 			sc->dataset_management = NVME_DATASET_MANAGEMENT_ENABLE;
3164 		else if (strcmp(value, "disable") == 0)
3165 			sc->dataset_management = NVME_DATASET_MANAGEMENT_DISABLE;
3166 	}
3167 
3168 	value = get_config_value_node(nvl, "bootindex");
3169 	if (value != NULL) {
3170 		if (pci_emul_add_boot_device(sc->nsc_pi, atoi(value))) {
3171 			EPRINTLN("Invalid bootindex %d", atoi(value));
3172 			return (-1);
3173 		}
3174 	}
3175 
3176 	value = get_config_value_node(nvl, "ram");
3177 	if (value != NULL) {
3178 		uint64_t sz = strtoull(value, NULL, 10);
3179 
3180 		sc->nvstore.type = NVME_STOR_RAM;
3181 		sc->nvstore.size = sz * 1024 * 1024;
3182 		sc->nvstore.ctx = calloc(1, sc->nvstore.size);
3183 		sc->nvstore.sectsz = 4096;
3184 		sc->nvstore.sectsz_bits = 12;
3185 		if (sc->nvstore.ctx == NULL) {
3186 			EPRINTLN("nvme: Unable to allocate RAM");
3187 			return (-1);
3188 		}
3189 	} else {
3190 		snprintf(bident, sizeof(bident), "%u:%u",
3191 		    sc->nsc_pi->pi_slot, sc->nsc_pi->pi_func);
3192 		sc->nvstore.ctx = blockif_open(nvl, bident);
3193 		if (sc->nvstore.ctx == NULL) {
3194 			EPRINTLN("nvme: Could not open backing file: %s",
3195 			    strerror(errno));
3196 			return (-1);
3197 		}
3198 		sc->nvstore.type = NVME_STOR_BLOCKIF;
3199 		sc->nvstore.size = blockif_size(sc->nvstore.ctx);
3200 	}
3201 
3202 	if (sectsz == 512 || sectsz == 4096 || sectsz == 8192)
3203 		sc->nvstore.sectsz = sectsz;
3204 	else if (sc->nvstore.type != NVME_STOR_RAM)
3205 		sc->nvstore.sectsz = blockif_sectsz(sc->nvstore.ctx);
3206 	for (sc->nvstore.sectsz_bits = 9;
3207 	     (1U << sc->nvstore.sectsz_bits) < sc->nvstore.sectsz;
3208 	     sc->nvstore.sectsz_bits++);
3209 
3210 	if (sc->max_queues <= 0 || sc->max_queues > NVME_QUEUES)
3211 		sc->max_queues = NVME_QUEUES;
3212 
3213 	return (0);
3214 }
3215 
3216 static void
3217 pci_nvme_resized(struct blockif_ctxt *bctxt __unused, void *arg,
3218     size_t new_size)
3219 {
3220 	struct pci_nvme_softc *sc;
3221 	struct pci_nvme_blockstore *nvstore;
3222 	struct nvme_namespace_data *nd;
3223 
3224 	sc = arg;
3225 	nvstore = &sc->nvstore;
3226 	nd = &sc->nsdata;
3227 
3228 	nvstore->size = new_size;
3229 	pci_nvme_init_nsdata_size(nvstore, nd);
3230 
3231 	/* Add changed NSID to list */
3232 	sc->ns_log.ns[0] = 1;
3233 	sc->ns_log.ns[1] = 0;
3234 
3235 	pci_nvme_aen_post(sc, PCI_NVME_AE_TYPE_NOTICE,
3236 	    PCI_NVME_AEI_NOTICE_NS_ATTR_CHANGED);
3237 }
3238 
3239 static int
3240 pci_nvme_init(struct pci_devinst *pi, nvlist_t *nvl)
3241 {
3242 	struct pci_nvme_softc *sc;
3243 	uint32_t pci_membar_sz;
3244 	int	error;
3245 
3246 	error = 0;
3247 
3248 	sc = calloc(1, sizeof(struct pci_nvme_softc));
3249 	pi->pi_arg = sc;
3250 	sc->nsc_pi = pi;
3251 
3252 	error = pci_nvme_parse_config(sc, nvl);
3253 	if (error < 0)
3254 		goto done;
3255 	else
3256 		error = 0;
3257 
3258 	STAILQ_INIT(&sc->ioreqs_free);
3259 	sc->ioreqs = calloc(sc->ioslots, sizeof(struct pci_nvme_ioreq));
3260 	for (uint32_t i = 0; i < sc->ioslots; i++) {
3261 		STAILQ_INSERT_TAIL(&sc->ioreqs_free, &sc->ioreqs[i], link);
3262 	}
3263 
3264 	pci_set_cfgdata16(pi, PCIR_DEVICE, 0x0A0A);
3265 	pci_set_cfgdata16(pi, PCIR_VENDOR, 0xFB5D);
3266 	pci_set_cfgdata8(pi, PCIR_CLASS, PCIC_STORAGE);
3267 	pci_set_cfgdata8(pi, PCIR_SUBCLASS, PCIS_STORAGE_NVM);
3268 	pci_set_cfgdata8(pi, PCIR_PROGIF,
3269 	                 PCIP_STORAGE_NVM_ENTERPRISE_NVMHCI_1_0);
3270 
3271 	/*
3272 	 * Allocate size of NVMe registers + doorbell space for all queues.
3273 	 *
3274 	 * The specification requires a minimum memory I/O window size of 16K.
3275 	 * The Windows driver will refuse to start a device with a smaller
3276 	 * window.
3277 	 */
3278 	pci_membar_sz = sizeof(struct nvme_registers) +
3279 	    2 * sizeof(uint32_t) * (sc->max_queues + 1);
3280 	pci_membar_sz = MAX(pci_membar_sz, NVME_MMIO_SPACE_MIN);
3281 
3282 	DPRINTF("nvme membar size: %u", pci_membar_sz);
3283 
3284 	error = pci_emul_alloc_bar(pi, 0, PCIBAR_MEM64, pci_membar_sz);
3285 	if (error) {
3286 		WPRINTF("%s pci alloc mem bar failed", __func__);
3287 		goto done;
3288 	}
3289 
3290 	error = pci_emul_add_msixcap(pi, sc->max_queues + 1, NVME_MSIX_BAR);
3291 	if (error) {
3292 		WPRINTF("%s pci add msixcap failed", __func__);
3293 		goto done;
3294 	}
3295 
3296 	error = pci_emul_add_pciecap(pi, PCIEM_TYPE_ROOT_INT_EP);
3297 	if (error) {
3298 		WPRINTF("%s pci add Express capability failed", __func__);
3299 		goto done;
3300 	}
3301 
3302 	pthread_mutex_init(&sc->mtx, NULL);
3303 	sem_init(&sc->iosemlock, 0, sc->ioslots);
3304 	blockif_register_resize_callback(sc->nvstore.ctx, pci_nvme_resized, sc);
3305 
3306 	pci_nvme_init_queues(sc, sc->max_queues, sc->max_queues);
3307 	/*
3308 	 * Controller data depends on Namespace data so initialize Namespace
3309 	 * data first.
3310 	 */
3311 	pci_nvme_init_nsdata(sc, &sc->nsdata, 1, &sc->nvstore);
3312 	pci_nvme_init_ctrldata(sc);
3313 	pci_nvme_init_logpages(sc);
3314 	pci_nvme_init_features(sc);
3315 
3316 	pci_nvme_aer_init(sc);
3317 	pci_nvme_aen_init(sc);
3318 
3319 	pci_nvme_reset(sc);
3320 done:
3321 	return (error);
3322 }
3323 
3324 static int
3325 pci_nvme_legacy_config(nvlist_t *nvl, const char *opts)
3326 {
3327 	char *cp, *ram;
3328 
3329 	if (opts == NULL)
3330 		return (0);
3331 
3332 	if (strncmp(opts, "ram=", 4) == 0) {
3333 		cp = strchr(opts, ',');
3334 		if (cp == NULL) {
3335 			set_config_value_node(nvl, "ram", opts + 4);
3336 			return (0);
3337 		}
3338 		ram = strndup(opts + 4, cp - opts - 4);
3339 		set_config_value_node(nvl, "ram", ram);
3340 		free(ram);
3341 		return (pci_parse_legacy_config(nvl, cp + 1));
3342 	} else
3343 		return (blockif_legacy_config(nvl, opts));
3344 }
3345 
3346 static const struct pci_devemu pci_de_nvme = {
3347 	.pe_emu =	"nvme",
3348 	.pe_init =	pci_nvme_init,
3349 	.pe_legacy_config = pci_nvme_legacy_config,
3350 	.pe_barwrite =	pci_nvme_write,
3351 	.pe_barread =	pci_nvme_read
3352 };
3353 PCI_EMUL_SET(pci_de_nvme);
3354