xref: /freebsd/usr.sbin/bhyve/pci_nvme.c (revision 22cf89c938886d14f5796fc49f9f020c23ea8eaf)
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 
518 	cd->vid = 0xFB5D;
519 	cd->ssvid = 0x0000;
520 
521 	cpywithpad((char *)cd->mn, sizeof(cd->mn), "bhyve-NVMe", ' ');
522 	cpywithpad((char *)cd->fr, sizeof(cd->fr), "1.0", ' ');
523 
524 	/* Num of submission commands that we can handle at a time (2^rab) */
525 	cd->rab   = 4;
526 
527 	/* FreeBSD OUI */
528 	cd->ieee[0] = 0xfc;
529 	cd->ieee[1] = 0x9c;
530 	cd->ieee[2] = 0x58;
531 
532 	cd->mic = 0;
533 
534 	cd->mdts = NVME_MDTS;	/* max data transfer size (2^mdts * CAP.MPSMIN) */
535 
536 	cd->ver = NVME_REV(1,4);
537 
538 	cd->cntrltype = NVME_CNTRLTYPE_IO;
539 	cd->oacs = 1 << NVME_CTRLR_DATA_OACS_FORMAT_SHIFT;
540 	cd->oaes = NVMEB(NVME_CTRLR_DATA_OAES_NS_ATTR);
541 	cd->acl = 2;
542 	cd->aerl = 4;
543 
544 	/* Advertise 1, Read-only firmware slot */
545 	cd->frmw = NVMEB(NVME_CTRLR_DATA_FRMW_SLOT1_RO) |
546 	    (1 << NVME_CTRLR_DATA_FRMW_NUM_SLOTS_SHIFT);
547 	cd->lpa = 0;	/* TODO: support some simple things like SMART */
548 	cd->elpe = 0;	/* max error log page entries */
549 	/*
550 	 * Report a single power state (zero-based value)
551 	 * power_state[] values are left as zero to indicate "Not reported"
552 	 */
553 	cd->npss = 0;
554 
555 	/* Warning Composite Temperature Threshold */
556 	cd->wctemp = 0x0157;
557 	cd->cctemp = 0x0157;
558 
559 	/* SANICAP must not be 0 for Revision 1.4 and later NVMe Controllers */
560 	cd->sanicap = (NVME_CTRLR_DATA_SANICAP_NODMMAS_NO <<
561 			NVME_CTRLR_DATA_SANICAP_NODMMAS_SHIFT);
562 
563 	cd->sqes = (6 << NVME_CTRLR_DATA_SQES_MAX_SHIFT) |
564 	    (6 << NVME_CTRLR_DATA_SQES_MIN_SHIFT);
565 	cd->cqes = (4 << NVME_CTRLR_DATA_CQES_MAX_SHIFT) |
566 	    (4 << NVME_CTRLR_DATA_CQES_MIN_SHIFT);
567 	cd->nn = 1;	/* number of namespaces */
568 
569 	cd->oncs = 0;
570 	switch (sc->dataset_management) {
571 	case NVME_DATASET_MANAGEMENT_AUTO:
572 		if (sc->nvstore.deallocate)
573 			cd->oncs |= NVME_ONCS_DSM;
574 		break;
575 	case NVME_DATASET_MANAGEMENT_ENABLE:
576 		cd->oncs |= NVME_ONCS_DSM;
577 		break;
578 	default:
579 		break;
580 	}
581 
582 	cd->fna = NVME_CTRLR_DATA_FNA_FORMAT_ALL_MASK <<
583 	    NVME_CTRLR_DATA_FNA_FORMAT_ALL_SHIFT;
584 
585 	cd->vwc = NVME_CTRLR_DATA_VWC_ALL_NO << NVME_CTRLR_DATA_VWC_ALL_SHIFT;
586 }
587 
588 static void
589 pci_nvme_init_nsdata_size(struct pci_nvme_blockstore *nvstore,
590     struct nvme_namespace_data *nd)
591 {
592 
593 	/* Get capacity and block size information from backing store */
594 	nd->nsze = nvstore->size / nvstore->sectsz;
595 	nd->ncap = nd->nsze;
596 	nd->nuse = nd->nsze;
597 }
598 
599 static void
600 pci_nvme_init_nsdata(struct pci_nvme_softc *sc,
601     struct nvme_namespace_data *nd, uint32_t nsid,
602     struct pci_nvme_blockstore *nvstore)
603 {
604 
605 	pci_nvme_init_nsdata_size(nvstore, nd);
606 
607 	if (nvstore->type == NVME_STOR_BLOCKIF)
608 		nvstore->deallocate = blockif_candelete(nvstore->ctx);
609 
610 	nd->nlbaf = 0; /* NLBAF is a 0's based value (i.e. 1 LBA Format) */
611 	nd->flbas = 0;
612 
613 	/* Create an EUI-64 if user did not provide one */
614 	if (nvstore->eui64 == 0) {
615 		char *data = NULL;
616 		uint64_t eui64 = nvstore->eui64;
617 
618 		asprintf(&data, "%s%u%u%u", get_config_value("name"),
619 		    sc->nsc_pi->pi_bus, sc->nsc_pi->pi_slot,
620 		    sc->nsc_pi->pi_func);
621 
622 		if (data != NULL) {
623 			eui64 = OUI_FREEBSD_NVME_LOW | crc16(0, data, strlen(data));
624 			free(data);
625 		}
626 		nvstore->eui64 = (eui64 << 16) | (nsid & 0xffff);
627 	}
628 	be64enc(nd->eui64, nvstore->eui64);
629 
630 	/* LBA data-sz = 2^lbads */
631 	nd->lbaf[0] = nvstore->sectsz_bits << NVME_NS_DATA_LBAF_LBADS_SHIFT;
632 }
633 
634 static void
635 pci_nvme_init_logpages(struct pci_nvme_softc *sc)
636 {
637 	__uint128_t power_cycles = 1;
638 
639 	memset(&sc->err_log, 0, sizeof(sc->err_log));
640 	memset(&sc->health_log, 0, sizeof(sc->health_log));
641 	memset(&sc->fw_log, 0, sizeof(sc->fw_log));
642 	memset(&sc->ns_log, 0, sizeof(sc->ns_log));
643 
644 	/* Set read/write remainder to round up according to spec */
645 	sc->read_dunits_remainder = 999;
646 	sc->write_dunits_remainder = 999;
647 
648 	/* Set nominal Health values checked by implementations */
649 	sc->health_log.temperature = NVME_TEMPERATURE;
650 	sc->health_log.available_spare = 100;
651 	sc->health_log.available_spare_threshold = 10;
652 
653 	/* Set Active Firmware Info to slot 1 */
654 	sc->fw_log.afi = (1 << NVME_FIRMWARE_PAGE_AFI_SLOT_SHIFT);
655 	memcpy(&sc->fw_log.revision[0], sc->ctrldata.fr,
656 	    sizeof(sc->fw_log.revision[0]));
657 
658 	memcpy(&sc->health_log.power_cycles, &power_cycles,
659 	    sizeof(sc->health_log.power_cycles));
660 }
661 
662 static void
663 pci_nvme_init_features(struct pci_nvme_softc *sc)
664 {
665 	enum nvme_feature	fid;
666 
667 	for (fid = 0; fid < NVME_FID_MAX; fid++) {
668 		switch (fid) {
669 		case NVME_FEAT_ARBITRATION:
670 		case NVME_FEAT_POWER_MANAGEMENT:
671 		case NVME_FEAT_INTERRUPT_COALESCING: //XXX
672 		case NVME_FEAT_WRITE_ATOMICITY:
673 			/* Mandatory but no special handling required */
674 		//XXX hang - case NVME_FEAT_PREDICTABLE_LATENCY_MODE_CONFIG:
675 		//XXX hang - case NVME_FEAT_HOST_BEHAVIOR_SUPPORT:
676 		//		  this returns a data buffer
677 			break;
678 		case NVME_FEAT_TEMPERATURE_THRESHOLD:
679 			sc->feat[fid].set = nvme_feature_temperature;
680 			break;
681 		case NVME_FEAT_ERROR_RECOVERY:
682 			sc->feat[fid].namespace_specific = true;
683 			break;
684 		case NVME_FEAT_NUMBER_OF_QUEUES:
685 			sc->feat[fid].set = nvme_feature_num_queues;
686 			break;
687 		case NVME_FEAT_INTERRUPT_VECTOR_CONFIGURATION:
688 			sc->feat[fid].set = nvme_feature_iv_config;
689 			break;
690 		case NVME_FEAT_ASYNC_EVENT_CONFIGURATION:
691 			sc->feat[fid].set = nvme_feature_async_event;
692 			/* Enable all AENs by default */
693 			sc->feat[fid].cdw11 = PCI_NVME_AEN_DEFAULT_MASK;
694 			break;
695 		default:
696 			sc->feat[fid].set = nvme_feature_invalid_cb;
697 			sc->feat[fid].get = nvme_feature_invalid_cb;
698 		}
699 	}
700 }
701 
702 static void
703 pci_nvme_aer_reset(struct pci_nvme_softc *sc)
704 {
705 
706 	STAILQ_INIT(&sc->aer_list);
707 	sc->aer_count = 0;
708 }
709 
710 static void
711 pci_nvme_aer_init(struct pci_nvme_softc *sc)
712 {
713 
714 	pthread_mutex_init(&sc->aer_mtx, NULL);
715 	pci_nvme_aer_reset(sc);
716 }
717 
718 static void
719 pci_nvme_aer_destroy(struct pci_nvme_softc *sc)
720 {
721 	struct pci_nvme_aer *aer = NULL;
722 
723 	pthread_mutex_lock(&sc->aer_mtx);
724 	while (!STAILQ_EMPTY(&sc->aer_list)) {
725 		aer = STAILQ_FIRST(&sc->aer_list);
726 		STAILQ_REMOVE_HEAD(&sc->aer_list, link);
727 		free(aer);
728 	}
729 	pthread_mutex_unlock(&sc->aer_mtx);
730 
731 	pci_nvme_aer_reset(sc);
732 }
733 
734 static bool
735 pci_nvme_aer_available(struct pci_nvme_softc *sc)
736 {
737 
738 	return (sc->aer_count != 0);
739 }
740 
741 static bool
742 pci_nvme_aer_limit_reached(struct pci_nvme_softc *sc)
743 {
744 	struct nvme_controller_data *cd = &sc->ctrldata;
745 
746 	/* AERL is a zero based value while aer_count is one's based */
747 	return (sc->aer_count == (cd->aerl + 1U));
748 }
749 
750 /*
751  * Add an Async Event Request
752  *
753  * Stores an AER to be returned later if the Controller needs to notify the
754  * host of an event.
755  * Note that while the NVMe spec doesn't require Controllers to return AER's
756  * in order, this implementation does preserve the order.
757  */
758 static int
759 pci_nvme_aer_add(struct pci_nvme_softc *sc, uint16_t cid)
760 {
761 	struct pci_nvme_aer *aer = NULL;
762 
763 	aer = calloc(1, sizeof(struct pci_nvme_aer));
764 	if (aer == NULL)
765 		return (-1);
766 
767 	/* Save the Command ID for use in the completion message */
768 	aer->cid = cid;
769 
770 	pthread_mutex_lock(&sc->aer_mtx);
771 	sc->aer_count++;
772 	STAILQ_INSERT_TAIL(&sc->aer_list, aer, link);
773 	pthread_mutex_unlock(&sc->aer_mtx);
774 
775 	return (0);
776 }
777 
778 /*
779  * Get an Async Event Request structure
780  *
781  * Returns a pointer to an AER previously submitted by the host or NULL if
782  * no AER's exist. Caller is responsible for freeing the returned struct.
783  */
784 static struct pci_nvme_aer *
785 pci_nvme_aer_get(struct pci_nvme_softc *sc)
786 {
787 	struct pci_nvme_aer *aer = NULL;
788 
789 	pthread_mutex_lock(&sc->aer_mtx);
790 	aer = STAILQ_FIRST(&sc->aer_list);
791 	if (aer != NULL) {
792 		STAILQ_REMOVE_HEAD(&sc->aer_list, link);
793 		sc->aer_count--;
794 	}
795 	pthread_mutex_unlock(&sc->aer_mtx);
796 
797 	return (aer);
798 }
799 
800 static void
801 pci_nvme_aen_reset(struct pci_nvme_softc *sc)
802 {
803 	uint32_t	atype;
804 
805 	memset(sc->aen, 0, PCI_NVME_AE_TYPE_MAX * sizeof(struct pci_nvme_aen));
806 
807 	for (atype = 0; atype < PCI_NVME_AE_TYPE_MAX; atype++) {
808 		sc->aen[atype].atype = atype;
809 	}
810 }
811 
812 static void
813 pci_nvme_aen_init(struct pci_nvme_softc *sc)
814 {
815 	char nstr[80];
816 
817 	pci_nvme_aen_reset(sc);
818 
819 	pthread_mutex_init(&sc->aen_mtx, NULL);
820 	pthread_create(&sc->aen_tid, NULL, aen_thr, sc);
821 	snprintf(nstr, sizeof(nstr), "nvme-aen-%d:%d", sc->nsc_pi->pi_slot,
822 	    sc->nsc_pi->pi_func);
823 	pthread_set_name_np(sc->aen_tid, nstr);
824 }
825 
826 static void
827 pci_nvme_aen_destroy(struct pci_nvme_softc *sc)
828 {
829 
830 	pci_nvme_aen_reset(sc);
831 }
832 
833 /* Notify the AEN thread of pending work */
834 static void
835 pci_nvme_aen_notify(struct pci_nvme_softc *sc)
836 {
837 
838 	pthread_cond_signal(&sc->aen_cond);
839 }
840 
841 /*
842  * Post an Asynchronous Event Notification
843  */
844 static int32_t
845 pci_nvme_aen_post(struct pci_nvme_softc *sc, pci_nvme_async_type atype,
846 		uint32_t event_data)
847 {
848 	struct pci_nvme_aen *aen;
849 
850 	if (atype >= PCI_NVME_AE_TYPE_MAX) {
851 		return(EINVAL);
852 	}
853 
854 	pthread_mutex_lock(&sc->aen_mtx);
855 	aen = &sc->aen[atype];
856 
857 	/* Has the controller already posted an event of this type? */
858 	if (aen->posted) {
859 		pthread_mutex_unlock(&sc->aen_mtx);
860 		return(EALREADY);
861 	}
862 
863 	aen->event_data = event_data;
864 	aen->posted = true;
865 	pthread_mutex_unlock(&sc->aen_mtx);
866 
867 	pci_nvme_aen_notify(sc);
868 
869 	return(0);
870 }
871 
872 static void
873 pci_nvme_aen_process(struct pci_nvme_softc *sc)
874 {
875 	struct pci_nvme_aer *aer;
876 	struct pci_nvme_aen *aen;
877 	pci_nvme_async_type atype;
878 	uint32_t mask;
879 	uint16_t status;
880 	uint8_t lid;
881 
882 	assert(pthread_mutex_isowned_np(&sc->aen_mtx));
883 	for (atype = 0; atype < PCI_NVME_AE_TYPE_MAX; atype++) {
884 		aen = &sc->aen[atype];
885 		/* Previous iterations may have depleted the available AER's */
886 		if (!pci_nvme_aer_available(sc)) {
887 			DPRINTF("%s: no AER", __func__);
888 			break;
889 		}
890 
891 		if (!aen->posted) {
892 			DPRINTF("%s: no AEN posted for atype=%#x", __func__, atype);
893 			continue;
894 		}
895 
896 		status = NVME_SC_SUCCESS;
897 
898 		/* Is the event masked? */
899 		mask =
900 		    sc->feat[NVME_FEAT_ASYNC_EVENT_CONFIGURATION].cdw11;
901 
902 		DPRINTF("%s: atype=%#x mask=%#x event_data=%#x", __func__, atype, mask, aen->event_data);
903 		switch (atype) {
904 		case PCI_NVME_AE_TYPE_ERROR:
905 			lid = NVME_LOG_ERROR;
906 			break;
907 		case PCI_NVME_AE_TYPE_SMART:
908 			mask &= 0xff;
909 			if ((mask & aen->event_data) == 0)
910 				continue;
911 			lid = NVME_LOG_HEALTH_INFORMATION;
912 			break;
913 		case PCI_NVME_AE_TYPE_NOTICE:
914 			if (aen->event_data >= PCI_NVME_AEI_NOTICE_MAX) {
915 				EPRINTLN("%s unknown AEN notice type %u",
916 				    __func__, aen->event_data);
917 				status = NVME_SC_INTERNAL_DEVICE_ERROR;
918 				lid = 0;
919 				break;
920 			}
921 			if ((PCI_NVME_AEI_NOTICE_MASK(aen->event_data) & mask) == 0)
922 				continue;
923 			switch (aen->event_data) {
924 			case PCI_NVME_AEI_NOTICE_NS_ATTR_CHANGED:
925 				lid = NVME_LOG_CHANGED_NAMESPACE;
926 				break;
927 			case PCI_NVME_AEI_NOTICE_FW_ACTIVATION:
928 				lid = NVME_LOG_FIRMWARE_SLOT;
929 				break;
930 			case PCI_NVME_AEI_NOTICE_TELEMETRY_CHANGE:
931 				lid = NVME_LOG_TELEMETRY_CONTROLLER_INITIATED;
932 				break;
933 			case PCI_NVME_AEI_NOTICE_ANA_CHANGE:
934 				lid = NVME_LOG_ASYMMETRIC_NAMESPACE_ACCESS;
935 				break;
936 			case PCI_NVME_AEI_NOTICE_PREDICT_LATENCY_CHANGE:
937 				lid = NVME_LOG_PREDICTABLE_LATENCY_EVENT_AGGREGATE;
938 				break;
939 			case PCI_NVME_AEI_NOTICE_LBA_STATUS_ALERT:
940 				lid = NVME_LOG_LBA_STATUS_INFORMATION;
941 				break;
942 			case PCI_NVME_AEI_NOTICE_ENDURANCE_GROUP_CHANGE:
943 				lid = NVME_LOG_ENDURANCE_GROUP_EVENT_AGGREGATE;
944 				break;
945 			default:
946 				lid = 0;
947 			}
948 			break;
949 		default:
950 			/* bad type?!? */
951 			EPRINTLN("%s unknown AEN type %u", __func__, atype);
952 			status = NVME_SC_INTERNAL_DEVICE_ERROR;
953 			lid = 0;
954 			break;
955 		}
956 
957 		aer = pci_nvme_aer_get(sc);
958 		assert(aer != NULL);
959 
960 		DPRINTF("%s: CID=%#x CDW0=%#x", __func__, aer->cid, (lid << 16) | (aen->event_data << 8) | atype);
961 		pci_nvme_cq_update(sc, &sc->compl_queues[0],
962 		    (lid << 16) | (aen->event_data << 8) | atype, /* cdw0 */
963 		    aer->cid,
964 		    0,		/* SQID */
965 		    status);
966 
967 		aen->event_data = 0;
968 		aen->posted = false;
969 
970 		pci_generate_msix(sc->nsc_pi, 0);
971 	}
972 }
973 
974 static void *
975 aen_thr(void *arg)
976 {
977 	struct pci_nvme_softc *sc;
978 
979 	sc = arg;
980 
981 	pthread_mutex_lock(&sc->aen_mtx);
982 	for (;;) {
983 		pci_nvme_aen_process(sc);
984 		pthread_cond_wait(&sc->aen_cond, &sc->aen_mtx);
985 	}
986 	pthread_mutex_unlock(&sc->aen_mtx);
987 
988 	pthread_exit(NULL);
989 	return (NULL);
990 }
991 
992 static void
993 pci_nvme_reset_locked(struct pci_nvme_softc *sc)
994 {
995 	uint32_t i;
996 
997 	DPRINTF("%s", __func__);
998 
999 	sc->regs.cap_lo = (ZERO_BASED(sc->max_qentries) & NVME_CAP_LO_REG_MQES_MASK) |
1000 	    (1 << NVME_CAP_LO_REG_CQR_SHIFT) |
1001 	    (60 << NVME_CAP_LO_REG_TO_SHIFT);
1002 
1003 	sc->regs.cap_hi = 1 << NVME_CAP_HI_REG_CSS_NVM_SHIFT;
1004 
1005 	sc->regs.vs = NVME_REV(1,4);	/* NVMe v1.4 */
1006 
1007 	sc->regs.cc = 0;
1008 
1009 	assert(sc->submit_queues != NULL);
1010 
1011 	for (i = 0; i < sc->num_squeues + 1; i++) {
1012 		sc->submit_queues[i].qbase = NULL;
1013 		sc->submit_queues[i].size = 0;
1014 		sc->submit_queues[i].cqid = 0;
1015 		sc->submit_queues[i].tail = 0;
1016 		sc->submit_queues[i].head = 0;
1017 	}
1018 
1019 	assert(sc->compl_queues != NULL);
1020 
1021 	for (i = 0; i < sc->num_cqueues + 1; i++) {
1022 		sc->compl_queues[i].qbase = NULL;
1023 		sc->compl_queues[i].size = 0;
1024 		sc->compl_queues[i].tail = 0;
1025 		sc->compl_queues[i].head = 0;
1026 	}
1027 
1028 	sc->num_q_is_set = false;
1029 
1030 	pci_nvme_aer_destroy(sc);
1031 	pci_nvme_aen_destroy(sc);
1032 
1033 	/*
1034 	 * Clear CSTS.RDY last to prevent the host from enabling Controller
1035 	 * before cleanup completes
1036 	 */
1037 	sc->regs.csts = 0;
1038 }
1039 
1040 static void
1041 pci_nvme_reset(struct pci_nvme_softc *sc)
1042 {
1043 	pthread_mutex_lock(&sc->mtx);
1044 	pci_nvme_reset_locked(sc);
1045 	pthread_mutex_unlock(&sc->mtx);
1046 }
1047 
1048 static int
1049 pci_nvme_init_controller(struct pci_nvme_softc *sc)
1050 {
1051 	uint16_t acqs, asqs;
1052 
1053 	DPRINTF("%s", __func__);
1054 
1055 	/*
1056 	 * NVMe 2.0 states that "enabling a controller while this field is
1057 	 * cleared to 0h produces undefined results" for both ACQS and
1058 	 * ASQS. If zero, set CFS and do not become ready.
1059 	 */
1060 	asqs = ONE_BASED(sc->regs.aqa & NVME_AQA_REG_ASQS_MASK);
1061 	if (asqs < 2) {
1062 		EPRINTLN("%s: illegal ASQS value %#x (aqa=%#x)", __func__,
1063 		    asqs - 1, sc->regs.aqa);
1064 		sc->regs.csts |= NVME_CSTS_CFS;
1065 		return (-1);
1066 	}
1067 	sc->submit_queues[0].size = asqs;
1068 	sc->submit_queues[0].qbase = vm_map_gpa(sc->nsc_pi->pi_vmctx,
1069 	    sc->regs.asq, sizeof(struct nvme_command) * asqs);
1070 	if (sc->submit_queues[0].qbase == NULL) {
1071 		EPRINTLN("%s: ASQ vm_map_gpa(%lx) failed", __func__,
1072 		    sc->regs.asq);
1073 		sc->regs.csts |= NVME_CSTS_CFS;
1074 		return (-1);
1075 	}
1076 
1077 	DPRINTF("%s mapping Admin-SQ guest 0x%lx, host: %p",
1078 	        __func__, sc->regs.asq, sc->submit_queues[0].qbase);
1079 
1080 	acqs = ONE_BASED((sc->regs.aqa >> NVME_AQA_REG_ACQS_SHIFT) &
1081 	    NVME_AQA_REG_ACQS_MASK);
1082 	if (acqs < 2) {
1083 		EPRINTLN("%s: illegal ACQS value %#x (aqa=%#x)", __func__,
1084 		    acqs - 1, sc->regs.aqa);
1085 		sc->regs.csts |= NVME_CSTS_CFS;
1086 		return (-1);
1087 	}
1088 	sc->compl_queues[0].size = acqs;
1089 	sc->compl_queues[0].qbase = vm_map_gpa(sc->nsc_pi->pi_vmctx,
1090 	    sc->regs.acq, sizeof(struct nvme_completion) * acqs);
1091 	if (sc->compl_queues[0].qbase == NULL) {
1092 		EPRINTLN("%s: ACQ vm_map_gpa(%lx) failed", __func__,
1093 		    sc->regs.acq);
1094 		sc->regs.csts |= NVME_CSTS_CFS;
1095 		return (-1);
1096 	}
1097 	sc->compl_queues[0].intr_en = NVME_CQ_INTEN;
1098 
1099 	DPRINTF("%s mapping Admin-CQ guest 0x%lx, host: %p",
1100 	        __func__, sc->regs.acq, sc->compl_queues[0].qbase);
1101 
1102 	return (0);
1103 }
1104 
1105 static int
1106 nvme_prp_memcpy(struct vmctx *ctx, uint64_t prp1, uint64_t prp2, uint8_t *b,
1107 	size_t len, enum nvme_copy_dir dir)
1108 {
1109 	uint8_t *p;
1110 	size_t bytes;
1111 
1112 	if (len > (8 * 1024)) {
1113 		return (-1);
1114 	}
1115 
1116 	/* Copy from the start of prp1 to the end of the physical page */
1117 	bytes = PAGE_SIZE - (prp1 & PAGE_MASK);
1118 	bytes = MIN(bytes, len);
1119 
1120 	p = vm_map_gpa(ctx, prp1, bytes);
1121 	if (p == NULL) {
1122 		return (-1);
1123 	}
1124 
1125 	if (dir == NVME_COPY_TO_PRP)
1126 		memcpy(p, b, bytes);
1127 	else
1128 		memcpy(b, p, bytes);
1129 
1130 	b += bytes;
1131 
1132 	len -= bytes;
1133 	if (len == 0) {
1134 		return (0);
1135 	}
1136 
1137 	len = MIN(len, PAGE_SIZE);
1138 
1139 	p = vm_map_gpa(ctx, prp2, len);
1140 	if (p == NULL) {
1141 		return (-1);
1142 	}
1143 
1144 	if (dir == NVME_COPY_TO_PRP)
1145 		memcpy(p, b, len);
1146 	else
1147 		memcpy(b, p, len);
1148 
1149 	return (0);
1150 }
1151 
1152 /*
1153  * Write a Completion Queue Entry update
1154  *
1155  * Write the completion and update the doorbell value
1156  */
1157 static void
1158 pci_nvme_cq_update(struct pci_nvme_softc *sc,
1159 		struct nvme_completion_queue *cq,
1160 		uint32_t cdw0,
1161 		uint16_t cid,
1162 		uint16_t sqid,
1163 		uint16_t status)
1164 {
1165 	struct nvme_submission_queue *sq = &sc->submit_queues[sqid];
1166 	struct nvme_completion *cqe;
1167 
1168 	assert(cq->qbase != NULL);
1169 
1170 	pthread_mutex_lock(&cq->mtx);
1171 
1172 	cqe = &cq->qbase[cq->tail];
1173 
1174 	/* Flip the phase bit */
1175 	status |= (cqe->status ^ NVME_STATUS_P) & NVME_STATUS_P_MASK;
1176 
1177 	cqe->cdw0 = cdw0;
1178 	cqe->sqhd = sq->head;
1179 	cqe->sqid = sqid;
1180 	cqe->cid = cid;
1181 	cqe->status = status;
1182 
1183 	cq->tail++;
1184 	if (cq->tail >= cq->size) {
1185 		cq->tail = 0;
1186 	}
1187 
1188 	pthread_mutex_unlock(&cq->mtx);
1189 }
1190 
1191 static int
1192 nvme_opc_delete_io_sq(struct pci_nvme_softc* sc, struct nvme_command* command,
1193 	struct nvme_completion* compl)
1194 {
1195 	uint16_t qid = command->cdw10 & 0xffff;
1196 
1197 	DPRINTF("%s DELETE_IO_SQ %u", __func__, qid);
1198 	if (qid == 0 || qid > sc->num_squeues ||
1199 	    (sc->submit_queues[qid].qbase == NULL)) {
1200 		WPRINTF("%s NOT PERMITTED queue id %u / num_squeues %u",
1201 		        __func__, qid, sc->num_squeues);
1202 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1203 		    NVME_SC_INVALID_QUEUE_IDENTIFIER);
1204 		return (1);
1205 	}
1206 
1207 	sc->submit_queues[qid].qbase = NULL;
1208 	sc->submit_queues[qid].cqid = 0;
1209 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1210 	return (1);
1211 }
1212 
1213 static int
1214 nvme_opc_create_io_sq(struct pci_nvme_softc* sc, struct nvme_command* command,
1215 	struct nvme_completion* compl)
1216 {
1217 	if (command->cdw11 & NVME_CMD_CDW11_PC) {
1218 		uint16_t qid = command->cdw10 & 0xffff;
1219 		struct nvme_submission_queue *nsq;
1220 
1221 		if ((qid == 0) || (qid > sc->num_squeues) ||
1222 		    (sc->submit_queues[qid].qbase != NULL)) {
1223 			WPRINTF("%s queue index %u > num_squeues %u",
1224 			        __func__, qid, sc->num_squeues);
1225 			pci_nvme_status_tc(&compl->status,
1226 			    NVME_SCT_COMMAND_SPECIFIC,
1227 			    NVME_SC_INVALID_QUEUE_IDENTIFIER);
1228 			return (1);
1229 		}
1230 
1231 		nsq = &sc->submit_queues[qid];
1232 		nsq->size = ONE_BASED((command->cdw10 >> 16) & 0xffff);
1233 		DPRINTF("%s size=%u (max=%u)", __func__, nsq->size, sc->max_qentries);
1234 		if ((nsq->size < 2) || (nsq->size > sc->max_qentries)) {
1235 			/*
1236 			 * Queues must specify at least two entries
1237 			 * NOTE: "MAXIMUM QUEUE SIZE EXCEEDED" was renamed to
1238 			 * "INVALID QUEUE SIZE" in the NVM Express 1.3 Spec
1239 			 */
1240 			pci_nvme_status_tc(&compl->status,
1241 			    NVME_SCT_COMMAND_SPECIFIC,
1242 			    NVME_SC_MAXIMUM_QUEUE_SIZE_EXCEEDED);
1243 			return (1);
1244 		}
1245 		nsq->head = nsq->tail = 0;
1246 
1247 		nsq->cqid = (command->cdw11 >> 16) & 0xffff;
1248 		if ((nsq->cqid == 0) || (nsq->cqid > sc->num_cqueues)) {
1249 			pci_nvme_status_tc(&compl->status,
1250 			    NVME_SCT_COMMAND_SPECIFIC,
1251 			    NVME_SC_INVALID_QUEUE_IDENTIFIER);
1252 			return (1);
1253 		}
1254 
1255 		if (sc->compl_queues[nsq->cqid].qbase == NULL) {
1256 			pci_nvme_status_tc(&compl->status,
1257 			    NVME_SCT_COMMAND_SPECIFIC,
1258 			    NVME_SC_COMPLETION_QUEUE_INVALID);
1259 			return (1);
1260 		}
1261 
1262 		nsq->qpriority = (command->cdw11 >> 1) & 0x03;
1263 
1264 		nsq->qbase = vm_map_gpa(sc->nsc_pi->pi_vmctx, command->prp1,
1265 		              sizeof(struct nvme_command) * (size_t)nsq->size);
1266 
1267 		DPRINTF("%s sq %u size %u gaddr %p cqid %u", __func__,
1268 		        qid, nsq->size, nsq->qbase, nsq->cqid);
1269 
1270 		pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1271 
1272 		DPRINTF("%s completed creating IOSQ qid %u",
1273 		         __func__, qid);
1274 	} else {
1275 		/*
1276 		 * Guest sent non-cont submission queue request.
1277 		 * This setting is unsupported by this emulation.
1278 		 */
1279 		WPRINTF("%s unsupported non-contig (list-based) "
1280 		         "create i/o submission queue", __func__);
1281 
1282 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1283 	}
1284 	return (1);
1285 }
1286 
1287 static int
1288 nvme_opc_delete_io_cq(struct pci_nvme_softc* sc, struct nvme_command* command,
1289 	struct nvme_completion* compl)
1290 {
1291 	uint16_t qid = command->cdw10 & 0xffff;
1292 	uint16_t sqid;
1293 
1294 	DPRINTF("%s DELETE_IO_CQ %u", __func__, qid);
1295 	if (qid == 0 || qid > sc->num_cqueues ||
1296 	    (sc->compl_queues[qid].qbase == NULL)) {
1297 		WPRINTF("%s queue index %u / num_cqueues %u",
1298 		        __func__, qid, sc->num_cqueues);
1299 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1300 		    NVME_SC_INVALID_QUEUE_IDENTIFIER);
1301 		return (1);
1302 	}
1303 
1304 	/* Deleting an Active CQ is an error */
1305 	for (sqid = 1; sqid < sc->num_squeues + 1; sqid++)
1306 		if (sc->submit_queues[sqid].cqid == qid) {
1307 			pci_nvme_status_tc(&compl->status,
1308 			    NVME_SCT_COMMAND_SPECIFIC,
1309 			    NVME_SC_INVALID_QUEUE_DELETION);
1310 			return (1);
1311 		}
1312 
1313 	sc->compl_queues[qid].qbase = NULL;
1314 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1315 	return (1);
1316 }
1317 
1318 static int
1319 nvme_opc_create_io_cq(struct pci_nvme_softc* sc, struct nvme_command* command,
1320 	struct nvme_completion* compl)
1321 {
1322 	struct nvme_completion_queue *ncq;
1323 	uint16_t qid = command->cdw10 & 0xffff;
1324 
1325 	/* Only support Physically Contiguous queues */
1326 	if ((command->cdw11 & NVME_CMD_CDW11_PC) == 0) {
1327 		WPRINTF("%s unsupported non-contig (list-based) "
1328 		         "create i/o completion queue",
1329 		         __func__);
1330 
1331 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1332 		return (1);
1333 	}
1334 
1335 	if ((qid == 0) || (qid > sc->num_cqueues) ||
1336 	    (sc->compl_queues[qid].qbase != NULL)) {
1337 		WPRINTF("%s queue index %u > num_cqueues %u",
1338 			__func__, qid, sc->num_cqueues);
1339 		pci_nvme_status_tc(&compl->status,
1340 		    NVME_SCT_COMMAND_SPECIFIC,
1341 		    NVME_SC_INVALID_QUEUE_IDENTIFIER);
1342 		return (1);
1343  	}
1344 
1345 	ncq = &sc->compl_queues[qid];
1346 	ncq->intr_en = (command->cdw11 & NVME_CMD_CDW11_IEN) >> 1;
1347 	ncq->intr_vec = (command->cdw11 >> 16) & 0xffff;
1348 	if (ncq->intr_vec > (sc->max_queues + 1)) {
1349 		pci_nvme_status_tc(&compl->status,
1350 		    NVME_SCT_COMMAND_SPECIFIC,
1351 		    NVME_SC_INVALID_INTERRUPT_VECTOR);
1352 		return (1);
1353 	}
1354 
1355 	ncq->size = ONE_BASED((command->cdw10 >> 16) & 0xffff);
1356 	if ((ncq->size < 2) || (ncq->size > sc->max_qentries))  {
1357 		/*
1358 		 * Queues must specify at least two entries
1359 		 * NOTE: "MAXIMUM QUEUE SIZE EXCEEDED" was renamed to
1360 		 * "INVALID QUEUE SIZE" in the NVM Express 1.3 Spec
1361 		 */
1362 		pci_nvme_status_tc(&compl->status,
1363 		    NVME_SCT_COMMAND_SPECIFIC,
1364 		    NVME_SC_MAXIMUM_QUEUE_SIZE_EXCEEDED);
1365 		return (1);
1366 	}
1367 	ncq->head = ncq->tail = 0;
1368 	ncq->qbase = vm_map_gpa(sc->nsc_pi->pi_vmctx,
1369 		     command->prp1,
1370 		     sizeof(struct nvme_command) * (size_t)ncq->size);
1371 
1372 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1373 
1374 
1375 	return (1);
1376 }
1377 
1378 static int
1379 nvme_opc_get_log_page(struct pci_nvme_softc* sc, struct nvme_command* command,
1380 	struct nvme_completion* compl)
1381 {
1382 	uint64_t logoff;
1383 	uint32_t logsize;
1384 	uint8_t logpage;
1385 
1386 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1387 
1388 	/*
1389 	 * Command specifies the number of dwords to return in fields NUMDU
1390 	 * and NUMDL. This is a zero-based value.
1391 	 */
1392 	logpage = command->cdw10 & 0xFF;
1393 	logsize = ((command->cdw11 << 16) | (command->cdw10 >> 16)) + 1;
1394 	logsize *= sizeof(uint32_t);
1395 	logoff  = ((uint64_t)(command->cdw13) << 32) | command->cdw12;
1396 
1397 	DPRINTF("%s log page %u len %u", __func__, logpage, logsize);
1398 
1399 	switch (logpage) {
1400 	case NVME_LOG_ERROR:
1401 		if (logoff >= sizeof(sc->err_log)) {
1402 			pci_nvme_status_genc(&compl->status,
1403 			    NVME_SC_INVALID_FIELD);
1404 			break;
1405 		}
1406 
1407 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
1408 		    command->prp2, (uint8_t *)&sc->err_log + logoff,
1409 		    MIN(logsize - logoff, sizeof(sc->err_log)),
1410 		    NVME_COPY_TO_PRP);
1411 		break;
1412 	case NVME_LOG_HEALTH_INFORMATION:
1413 		if (logoff >= sizeof(sc->health_log)) {
1414 			pci_nvme_status_genc(&compl->status,
1415 			    NVME_SC_INVALID_FIELD);
1416 			break;
1417 		}
1418 
1419 		pthread_mutex_lock(&sc->mtx);
1420 		memcpy(&sc->health_log.data_units_read, &sc->read_data_units,
1421 		    sizeof(sc->health_log.data_units_read));
1422 		memcpy(&sc->health_log.data_units_written, &sc->write_data_units,
1423 		    sizeof(sc->health_log.data_units_written));
1424 		memcpy(&sc->health_log.host_read_commands, &sc->read_commands,
1425 		    sizeof(sc->health_log.host_read_commands));
1426 		memcpy(&sc->health_log.host_write_commands, &sc->write_commands,
1427 		    sizeof(sc->health_log.host_write_commands));
1428 		pthread_mutex_unlock(&sc->mtx);
1429 
1430 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
1431 		    command->prp2, (uint8_t *)&sc->health_log + logoff,
1432 		    MIN(logsize - logoff, sizeof(sc->health_log)),
1433 		    NVME_COPY_TO_PRP);
1434 		break;
1435 	case NVME_LOG_FIRMWARE_SLOT:
1436 		if (logoff >= sizeof(sc->fw_log)) {
1437 			pci_nvme_status_genc(&compl->status,
1438 			    NVME_SC_INVALID_FIELD);
1439 			break;
1440 		}
1441 
1442 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
1443 		    command->prp2, (uint8_t *)&sc->fw_log + logoff,
1444 		    MIN(logsize - logoff, sizeof(sc->fw_log)),
1445 		    NVME_COPY_TO_PRP);
1446 		break;
1447 	case NVME_LOG_CHANGED_NAMESPACE:
1448 		if (logoff >= sizeof(sc->ns_log)) {
1449 			pci_nvme_status_genc(&compl->status,
1450 			    NVME_SC_INVALID_FIELD);
1451 			break;
1452 		}
1453 
1454 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
1455 		    command->prp2, (uint8_t *)&sc->ns_log + logoff,
1456 		    MIN(logsize - logoff, sizeof(sc->ns_log)),
1457 		    NVME_COPY_TO_PRP);
1458 		memset(&sc->ns_log, 0, sizeof(sc->ns_log));
1459 		break;
1460 	default:
1461 		DPRINTF("%s get log page %x command not supported",
1462 		        __func__, logpage);
1463 
1464 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1465 		    NVME_SC_INVALID_LOG_PAGE);
1466 	}
1467 
1468 	return (1);
1469 }
1470 
1471 static int
1472 nvme_opc_identify(struct pci_nvme_softc* sc, struct nvme_command* command,
1473 	struct nvme_completion* compl)
1474 {
1475 	void *dest;
1476 	uint16_t status;
1477 
1478 	DPRINTF("%s identify 0x%x nsid 0x%x", __func__,
1479 	        command->cdw10 & 0xFF, command->nsid);
1480 
1481 	status = 0;
1482 	pci_nvme_status_genc(&status, NVME_SC_SUCCESS);
1483 
1484 	switch (command->cdw10 & 0xFF) {
1485 	case 0x00: /* return Identify Namespace data structure */
1486 		/* Global NS only valid with NS Management */
1487 		if (command->nsid == NVME_GLOBAL_NAMESPACE_TAG) {
1488 			pci_nvme_status_genc(&status,
1489 			    NVME_SC_INVALID_NAMESPACE_OR_FORMAT);
1490 			break;
1491 		}
1492 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
1493 		    command->prp2, (uint8_t *)&sc->nsdata, sizeof(sc->nsdata),
1494 		    NVME_COPY_TO_PRP);
1495 		break;
1496 	case 0x01: /* return Identify Controller data structure */
1497 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
1498 		    command->prp2, (uint8_t *)&sc->ctrldata,
1499 		    sizeof(sc->ctrldata),
1500 		    NVME_COPY_TO_PRP);
1501 		break;
1502 	case 0x02: /* list of 1024 active NSIDs > CDW1.NSID */
1503 		dest = vm_map_gpa(sc->nsc_pi->pi_vmctx, command->prp1,
1504 		                  sizeof(uint32_t) * 1024);
1505 		/* All unused entries shall be zero */
1506 		memset(dest, 0, sizeof(uint32_t) * 1024);
1507 		((uint32_t *)dest)[0] = 1;
1508 		break;
1509 	case 0x03: /* list of NSID structures in CDW1.NSID, 4096 bytes */
1510 		if (command->nsid != 1) {
1511 			pci_nvme_status_genc(&status,
1512 			    NVME_SC_INVALID_NAMESPACE_OR_FORMAT);
1513 			break;
1514 		}
1515 		dest = vm_map_gpa(sc->nsc_pi->pi_vmctx, command->prp1,
1516 		                  sizeof(uint32_t) * 1024);
1517 		/* All bytes after the descriptor shall be zero */
1518 		memset(dest, 0, sizeof(uint32_t) * 1024);
1519 
1520 		/* Return NIDT=1 (i.e. EUI64) descriptor */
1521 		((uint8_t *)dest)[0] = 1;
1522 		((uint8_t *)dest)[1] = sizeof(uint64_t);
1523 		memcpy(((uint8_t *)dest) + 4, sc->nsdata.eui64, sizeof(uint64_t));
1524 		break;
1525 	case 0x13:
1526 		/*
1527 		 * Controller list is optional but used by UNH tests. Return
1528 		 * a valid but empty list.
1529 		 */
1530 		dest = vm_map_gpa(sc->nsc_pi->pi_vmctx, command->prp1,
1531 		                  sizeof(uint16_t) * 2048);
1532 		memset(dest, 0, sizeof(uint16_t) * 2048);
1533 		break;
1534 	default:
1535 		DPRINTF("%s unsupported identify command requested 0x%x",
1536 		         __func__, command->cdw10 & 0xFF);
1537 		pci_nvme_status_genc(&status, NVME_SC_INVALID_FIELD);
1538 		break;
1539 	}
1540 
1541 	compl->status = status;
1542 	return (1);
1543 }
1544 
1545 static const char *
1546 nvme_fid_to_name(uint8_t fid)
1547 {
1548 	const char *name;
1549 
1550 	switch (fid) {
1551 	case NVME_FEAT_ARBITRATION:
1552 		name = "Arbitration";
1553 		break;
1554 	case NVME_FEAT_POWER_MANAGEMENT:
1555 		name = "Power Management";
1556 		break;
1557 	case NVME_FEAT_LBA_RANGE_TYPE:
1558 		name = "LBA Range Type";
1559 		break;
1560 	case NVME_FEAT_TEMPERATURE_THRESHOLD:
1561 		name = "Temperature Threshold";
1562 		break;
1563 	case NVME_FEAT_ERROR_RECOVERY:
1564 		name = "Error Recovery";
1565 		break;
1566 	case NVME_FEAT_VOLATILE_WRITE_CACHE:
1567 		name = "Volatile Write Cache";
1568 		break;
1569 	case NVME_FEAT_NUMBER_OF_QUEUES:
1570 		name = "Number of Queues";
1571 		break;
1572 	case NVME_FEAT_INTERRUPT_COALESCING:
1573 		name = "Interrupt Coalescing";
1574 		break;
1575 	case NVME_FEAT_INTERRUPT_VECTOR_CONFIGURATION:
1576 		name = "Interrupt Vector Configuration";
1577 		break;
1578 	case NVME_FEAT_WRITE_ATOMICITY:
1579 		name = "Write Atomicity Normal";
1580 		break;
1581 	case NVME_FEAT_ASYNC_EVENT_CONFIGURATION:
1582 		name = "Asynchronous Event Configuration";
1583 		break;
1584 	case NVME_FEAT_AUTONOMOUS_POWER_STATE_TRANSITION:
1585 		name = "Autonomous Power State Transition";
1586 		break;
1587 	case NVME_FEAT_HOST_MEMORY_BUFFER:
1588 		name = "Host Memory Buffer";
1589 		break;
1590 	case NVME_FEAT_TIMESTAMP:
1591 		name = "Timestamp";
1592 		break;
1593 	case NVME_FEAT_KEEP_ALIVE_TIMER:
1594 		name = "Keep Alive Timer";
1595 		break;
1596 	case NVME_FEAT_HOST_CONTROLLED_THERMAL_MGMT:
1597 		name = "Host Controlled Thermal Management";
1598 		break;
1599 	case NVME_FEAT_NON_OP_POWER_STATE_CONFIG:
1600 		name = "Non-Operation Power State Config";
1601 		break;
1602 	case NVME_FEAT_READ_RECOVERY_LEVEL_CONFIG:
1603 		name = "Read Recovery Level Config";
1604 		break;
1605 	case NVME_FEAT_PREDICTABLE_LATENCY_MODE_CONFIG:
1606 		name = "Predictable Latency Mode Config";
1607 		break;
1608 	case NVME_FEAT_PREDICTABLE_LATENCY_MODE_WINDOW:
1609 		name = "Predictable Latency Mode Window";
1610 		break;
1611 	case NVME_FEAT_LBA_STATUS_INFORMATION_ATTRIBUTES:
1612 		name = "LBA Status Information Report Interval";
1613 		break;
1614 	case NVME_FEAT_HOST_BEHAVIOR_SUPPORT:
1615 		name = "Host Behavior Support";
1616 		break;
1617 	case NVME_FEAT_SANITIZE_CONFIG:
1618 		name = "Sanitize Config";
1619 		break;
1620 	case NVME_FEAT_ENDURANCE_GROUP_EVENT_CONFIGURATION:
1621 		name = "Endurance Group Event Configuration";
1622 		break;
1623 	case NVME_FEAT_SOFTWARE_PROGRESS_MARKER:
1624 		name = "Software Progress Marker";
1625 		break;
1626 	case NVME_FEAT_HOST_IDENTIFIER:
1627 		name = "Host Identifier";
1628 		break;
1629 	case NVME_FEAT_RESERVATION_NOTIFICATION_MASK:
1630 		name = "Reservation Notification Mask";
1631 		break;
1632 	case NVME_FEAT_RESERVATION_PERSISTENCE:
1633 		name = "Reservation Persistence";
1634 		break;
1635 	case NVME_FEAT_NAMESPACE_WRITE_PROTECTION_CONFIG:
1636 		name = "Namespace Write Protection Config";
1637 		break;
1638 	default:
1639 		name = "Unknown";
1640 		break;
1641 	}
1642 
1643 	return (name);
1644 }
1645 
1646 static void
1647 nvme_feature_invalid_cb(struct pci_nvme_softc *sc __unused,
1648     struct nvme_feature_obj *feat __unused,
1649     struct nvme_command *command __unused,
1650     struct nvme_completion *compl)
1651 {
1652 	pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1653 }
1654 
1655 static void
1656 nvme_feature_iv_config(struct pci_nvme_softc *sc,
1657     struct nvme_feature_obj *feat __unused,
1658     struct nvme_command *command,
1659     struct nvme_completion *compl)
1660 {
1661 	uint32_t i;
1662 	uint32_t cdw11 = command->cdw11;
1663 	uint16_t iv;
1664 	bool cd;
1665 
1666 	pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1667 
1668 	iv = cdw11 & 0xffff;
1669 	cd = cdw11 & (1 << 16);
1670 
1671 	if (iv > (sc->max_queues + 1)) {
1672 		return;
1673 	}
1674 
1675 	/* No Interrupt Coalescing (i.e. not Coalescing Disable) for Admin Q */
1676 	if ((iv == 0) && !cd)
1677 		return;
1678 
1679 	/* Requested Interrupt Vector must be used by a CQ */
1680 	for (i = 0; i < sc->num_cqueues + 1; i++) {
1681 		if (sc->compl_queues[i].intr_vec == iv) {
1682 			pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1683 		}
1684 	}
1685 }
1686 
1687 #define NVME_ASYNC_EVENT_ENDURANCE_GROUP		(0x4000)
1688 static void
1689 nvme_feature_async_event(struct pci_nvme_softc *sc __unused,
1690     struct nvme_feature_obj *feat __unused,
1691     struct nvme_command *command,
1692     struct nvme_completion *compl)
1693 {
1694 	if (command->cdw11 & NVME_ASYNC_EVENT_ENDURANCE_GROUP)
1695 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1696 }
1697 
1698 #define NVME_TEMP_THRESH_OVER	0
1699 #define NVME_TEMP_THRESH_UNDER	1
1700 static void
1701 nvme_feature_temperature(struct pci_nvme_softc *sc,
1702     struct nvme_feature_obj *feat __unused,
1703     struct nvme_command *command,
1704     struct nvme_completion *compl)
1705 {
1706 	uint16_t	tmpth;	/* Temperature Threshold */
1707 	uint8_t		tmpsel; /* Threshold Temperature Select */
1708 	uint8_t		thsel;  /* Threshold Type Select */
1709 	bool		set_crit = false;
1710 	bool		report_crit;
1711 
1712 	tmpth  = command->cdw11 & 0xffff;
1713 	tmpsel = (command->cdw11 >> 16) & 0xf;
1714 	thsel  = (command->cdw11 >> 20) & 0x3;
1715 
1716 	DPRINTF("%s: tmpth=%#x tmpsel=%#x thsel=%#x", __func__, tmpth, tmpsel, thsel);
1717 
1718 	/* Check for unsupported values */
1719 	if (((tmpsel != 0) && (tmpsel != 0xf)) ||
1720 	    (thsel > NVME_TEMP_THRESH_UNDER)) {
1721 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1722 		return;
1723 	}
1724 
1725 	if (((thsel == NVME_TEMP_THRESH_OVER)  && (NVME_TEMPERATURE >= tmpth)) ||
1726 	    ((thsel == NVME_TEMP_THRESH_UNDER) && (NVME_TEMPERATURE <= tmpth)))
1727 		set_crit = true;
1728 
1729 	pthread_mutex_lock(&sc->mtx);
1730 	if (set_crit)
1731 		sc->health_log.critical_warning |=
1732 		    NVME_CRIT_WARN_ST_TEMPERATURE;
1733 	else
1734 		sc->health_log.critical_warning &=
1735 		    ~NVME_CRIT_WARN_ST_TEMPERATURE;
1736 	pthread_mutex_unlock(&sc->mtx);
1737 
1738 	report_crit = sc->feat[NVME_FEAT_ASYNC_EVENT_CONFIGURATION].cdw11 &
1739 	    NVME_CRIT_WARN_ST_TEMPERATURE;
1740 
1741 	if (set_crit && report_crit)
1742 		pci_nvme_aen_post(sc, PCI_NVME_AE_TYPE_SMART,
1743 		    sc->health_log.critical_warning);
1744 
1745 	DPRINTF("%s: set_crit=%c critical_warning=%#x status=%#x", __func__, set_crit ? 'T':'F', sc->health_log.critical_warning, compl->status);
1746 }
1747 
1748 static void
1749 nvme_feature_num_queues(struct pci_nvme_softc *sc,
1750     struct nvme_feature_obj *feat __unused,
1751     struct nvme_command *command,
1752     struct nvme_completion *compl)
1753 {
1754 	uint16_t nqr;	/* Number of Queues Requested */
1755 
1756 	if (sc->num_q_is_set) {
1757 		WPRINTF("%s: Number of Queues already set", __func__);
1758 		pci_nvme_status_genc(&compl->status,
1759 		    NVME_SC_COMMAND_SEQUENCE_ERROR);
1760 		return;
1761 	}
1762 
1763 	nqr = command->cdw11 & 0xFFFF;
1764 	if (nqr == 0xffff) {
1765 		WPRINTF("%s: Illegal NSQR value %#x", __func__, nqr);
1766 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1767 		return;
1768 	}
1769 
1770 	sc->num_squeues = ONE_BASED(nqr);
1771 	if (sc->num_squeues > sc->max_queues) {
1772 		DPRINTF("NSQR=%u is greater than max %u", sc->num_squeues,
1773 					sc->max_queues);
1774 		sc->num_squeues = sc->max_queues;
1775 	}
1776 
1777 	nqr = (command->cdw11 >> 16) & 0xFFFF;
1778 	if (nqr == 0xffff) {
1779 		WPRINTF("%s: Illegal NCQR value %#x", __func__, nqr);
1780 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1781 		return;
1782 	}
1783 
1784 	sc->num_cqueues = ONE_BASED(nqr);
1785 	if (sc->num_cqueues > sc->max_queues) {
1786 		DPRINTF("NCQR=%u is greater than max %u", sc->num_cqueues,
1787 					sc->max_queues);
1788 		sc->num_cqueues = sc->max_queues;
1789 	}
1790 
1791 	/* Patch the command value which will be saved on callback's return */
1792 	command->cdw11 = NVME_FEATURE_NUM_QUEUES(sc);
1793 	compl->cdw0 = NVME_FEATURE_NUM_QUEUES(sc);
1794 
1795 	sc->num_q_is_set = true;
1796 }
1797 
1798 static int
1799 nvme_opc_set_features(struct pci_nvme_softc *sc, struct nvme_command *command,
1800 	struct nvme_completion *compl)
1801 {
1802 	struct nvme_feature_obj *feat;
1803 	uint32_t nsid = command->nsid;
1804 	uint8_t fid = NVMEV(NVME_FEAT_SET_FID, command->cdw10);
1805 	bool sv = NVMEV(NVME_FEAT_SET_SV, command->cdw10);
1806 
1807 	DPRINTF("%s: Feature ID 0x%x (%s)", __func__, fid, nvme_fid_to_name(fid));
1808 
1809 	if (fid >= NVME_FID_MAX) {
1810 		DPRINTF("%s invalid feature 0x%x", __func__, fid);
1811 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1812 		return (1);
1813 	}
1814 
1815 	if (sv) {
1816 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1817 		    NVME_SC_FEATURE_NOT_SAVEABLE);
1818 		return (1);
1819 	}
1820 
1821 	feat = &sc->feat[fid];
1822 
1823 	if (feat->namespace_specific && (nsid == NVME_GLOBAL_NAMESPACE_TAG)) {
1824 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1825 		return (1);
1826 	}
1827 
1828 	if (!feat->namespace_specific &&
1829 	    !((nsid == 0) || (nsid == NVME_GLOBAL_NAMESPACE_TAG))) {
1830 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1831 		    NVME_SC_FEATURE_NOT_NS_SPECIFIC);
1832 		return (1);
1833 	}
1834 
1835 	compl->cdw0 = 0;
1836 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1837 
1838 	if (feat->set)
1839 		feat->set(sc, feat, command, compl);
1840 	else {
1841 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1842 		    NVME_SC_FEATURE_NOT_CHANGEABLE);
1843 		return (1);
1844 	}
1845 
1846 	DPRINTF("%s: status=%#x cdw11=%#x", __func__, compl->status, command->cdw11);
1847 	if (compl->status == NVME_SC_SUCCESS) {
1848 		feat->cdw11 = command->cdw11;
1849 		if ((fid == NVME_FEAT_ASYNC_EVENT_CONFIGURATION) &&
1850 		    (command->cdw11 != 0))
1851 			pci_nvme_aen_notify(sc);
1852 	}
1853 
1854 	return (0);
1855 }
1856 
1857 #define NVME_FEATURES_SEL_SUPPORTED	0x3
1858 #define NVME_FEATURES_NS_SPECIFIC	(1 << 1)
1859 
1860 static int
1861 nvme_opc_get_features(struct pci_nvme_softc* sc, struct nvme_command* command,
1862 	struct nvme_completion* compl)
1863 {
1864 	struct nvme_feature_obj *feat;
1865 	uint8_t fid = command->cdw10 & 0xFF;
1866 	uint8_t sel = (command->cdw10 >> 8) & 0x7;
1867 
1868 	DPRINTF("%s: Feature ID 0x%x (%s)", __func__, fid, nvme_fid_to_name(fid));
1869 
1870 	if (fid >= NVME_FID_MAX) {
1871 		DPRINTF("%s invalid feature 0x%x", __func__, fid);
1872 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1873 		return (1);
1874 	}
1875 
1876 	compl->cdw0 = 0;
1877 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1878 
1879 	feat = &sc->feat[fid];
1880 	if (feat->get) {
1881 		feat->get(sc, feat, command, compl);
1882 	}
1883 
1884 	if (compl->status == NVME_SC_SUCCESS) {
1885 		if ((sel == NVME_FEATURES_SEL_SUPPORTED) && feat->namespace_specific)
1886 			compl->cdw0 = NVME_FEATURES_NS_SPECIFIC;
1887 		else
1888 			compl->cdw0 = feat->cdw11;
1889 	}
1890 
1891 	return (0);
1892 }
1893 
1894 static int
1895 nvme_opc_format_nvm(struct pci_nvme_softc* sc, struct nvme_command* command,
1896 	struct nvme_completion* compl)
1897 {
1898 	uint8_t	ses, lbaf, pi;
1899 
1900 	/* Only supports Secure Erase Setting - User Data Erase */
1901 	ses = (command->cdw10 >> 9) & 0x7;
1902 	if (ses > 0x1) {
1903 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1904 		return (1);
1905 	}
1906 
1907 	/* Only supports a single LBA Format */
1908 	lbaf = command->cdw10 & 0xf;
1909 	if (lbaf != 0) {
1910 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1911 		    NVME_SC_INVALID_FORMAT);
1912 		return (1);
1913 	}
1914 
1915 	/* Doesn't support Protection Information */
1916 	pi = (command->cdw10 >> 5) & 0x7;
1917 	if (pi != 0) {
1918 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
1919 		return (1);
1920 	}
1921 
1922 	if (sc->nvstore.type == NVME_STOR_RAM) {
1923 		if (sc->nvstore.ctx)
1924 			free(sc->nvstore.ctx);
1925 		sc->nvstore.ctx = calloc(1, sc->nvstore.size);
1926 		pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1927 	} else {
1928 		struct pci_nvme_ioreq *req;
1929 		int err;
1930 
1931 		req = pci_nvme_get_ioreq(sc);
1932 		if (req == NULL) {
1933 			pci_nvme_status_genc(&compl->status,
1934 			    NVME_SC_INTERNAL_DEVICE_ERROR);
1935 			WPRINTF("%s: unable to allocate IO req", __func__);
1936 			return (1);
1937 		}
1938 		req->nvme_sq = &sc->submit_queues[0];
1939 		req->sqid = 0;
1940 		req->opc = command->opc;
1941 		req->cid = command->cid;
1942 		req->nsid = command->nsid;
1943 
1944 		req->io_req.br_offset = 0;
1945 		req->io_req.br_resid = sc->nvstore.size;
1946 		req->io_req.br_callback = pci_nvme_io_done;
1947 
1948 		err = blockif_delete(sc->nvstore.ctx, &req->io_req);
1949 		if (err) {
1950 			pci_nvme_status_genc(&compl->status,
1951 			    NVME_SC_INTERNAL_DEVICE_ERROR);
1952 			pci_nvme_release_ioreq(sc, req);
1953 		} else
1954 			compl->status = NVME_NO_STATUS;
1955 	}
1956 
1957 	return (1);
1958 }
1959 
1960 static int
1961 nvme_opc_abort(struct pci_nvme_softc *sc __unused, struct nvme_command *command,
1962     struct nvme_completion *compl)
1963 {
1964 	DPRINTF("%s submission queue %u, command ID 0x%x", __func__,
1965 	        command->cdw10 & 0xFFFF, (command->cdw10 >> 16) & 0xFFFF);
1966 
1967 	/* TODO: search for the command ID and abort it */
1968 
1969 	compl->cdw0 = 1;
1970 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
1971 	return (1);
1972 }
1973 
1974 static int
1975 nvme_opc_async_event_req(struct pci_nvme_softc* sc,
1976 	struct nvme_command* command, struct nvme_completion* compl)
1977 {
1978 	DPRINTF("%s async event request count=%u aerl=%u cid=%#x", __func__,
1979 	    sc->aer_count, sc->ctrldata.aerl, command->cid);
1980 
1981 	/* Don't exceed the Async Event Request Limit (AERL). */
1982 	if (pci_nvme_aer_limit_reached(sc)) {
1983 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
1984 				NVME_SC_ASYNC_EVENT_REQUEST_LIMIT_EXCEEDED);
1985 		return (1);
1986 	}
1987 
1988 	if (pci_nvme_aer_add(sc, command->cid)) {
1989 		pci_nvme_status_tc(&compl->status, NVME_SCT_GENERIC,
1990 				NVME_SC_INTERNAL_DEVICE_ERROR);
1991 		return (1);
1992 	}
1993 
1994 	/*
1995 	 * Raise events when they happen based on the Set Features cmd.
1996 	 * These events happen async, so only set completion successful if
1997 	 * there is an event reflective of the request to get event.
1998 	 */
1999 	compl->status = NVME_NO_STATUS;
2000 	pci_nvme_aen_notify(sc);
2001 
2002 	return (0);
2003 }
2004 
2005 static void
2006 pci_nvme_handle_admin_cmd(struct pci_nvme_softc* sc, uint64_t value)
2007 {
2008 	struct nvme_completion compl;
2009 	struct nvme_command *cmd;
2010 	struct nvme_submission_queue *sq;
2011 	struct nvme_completion_queue *cq;
2012 	uint16_t sqhead;
2013 
2014 	DPRINTF("%s index %u", __func__, (uint32_t)value);
2015 
2016 	sq = &sc->submit_queues[0];
2017 	cq = &sc->compl_queues[0];
2018 
2019 	pthread_mutex_lock(&sq->mtx);
2020 
2021 	sqhead = sq->head;
2022 	DPRINTF("sqhead %u, tail %u", sqhead, sq->tail);
2023 
2024 	while (sqhead != atomic_load_acq_short(&sq->tail)) {
2025 		cmd = &(sq->qbase)[sqhead];
2026 		compl.cdw0 = 0;
2027 		compl.status = 0;
2028 
2029 		switch (cmd->opc) {
2030 		case NVME_OPC_DELETE_IO_SQ:
2031 			DPRINTF("%s command DELETE_IO_SQ", __func__);
2032 			nvme_opc_delete_io_sq(sc, cmd, &compl);
2033 			break;
2034 		case NVME_OPC_CREATE_IO_SQ:
2035 			DPRINTF("%s command CREATE_IO_SQ", __func__);
2036 			nvme_opc_create_io_sq(sc, cmd, &compl);
2037 			break;
2038 		case NVME_OPC_DELETE_IO_CQ:
2039 			DPRINTF("%s command DELETE_IO_CQ", __func__);
2040 			nvme_opc_delete_io_cq(sc, cmd, &compl);
2041 			break;
2042 		case NVME_OPC_CREATE_IO_CQ:
2043 			DPRINTF("%s command CREATE_IO_CQ", __func__);
2044 			nvme_opc_create_io_cq(sc, cmd, &compl);
2045 			break;
2046 		case NVME_OPC_GET_LOG_PAGE:
2047 			DPRINTF("%s command GET_LOG_PAGE", __func__);
2048 			nvme_opc_get_log_page(sc, cmd, &compl);
2049 			break;
2050 		case NVME_OPC_IDENTIFY:
2051 			DPRINTF("%s command IDENTIFY", __func__);
2052 			nvme_opc_identify(sc, cmd, &compl);
2053 			break;
2054 		case NVME_OPC_ABORT:
2055 			DPRINTF("%s command ABORT", __func__);
2056 			nvme_opc_abort(sc, cmd, &compl);
2057 			break;
2058 		case NVME_OPC_SET_FEATURES:
2059 			DPRINTF("%s command SET_FEATURES", __func__);
2060 			nvme_opc_set_features(sc, cmd, &compl);
2061 			break;
2062 		case NVME_OPC_GET_FEATURES:
2063 			DPRINTF("%s command GET_FEATURES", __func__);
2064 			nvme_opc_get_features(sc, cmd, &compl);
2065 			break;
2066 		case NVME_OPC_FIRMWARE_ACTIVATE:
2067 			DPRINTF("%s command FIRMWARE_ACTIVATE", __func__);
2068 			pci_nvme_status_tc(&compl.status,
2069 			    NVME_SCT_COMMAND_SPECIFIC,
2070 			    NVME_SC_INVALID_FIRMWARE_SLOT);
2071 			break;
2072 		case NVME_OPC_ASYNC_EVENT_REQUEST:
2073 			DPRINTF("%s command ASYNC_EVENT_REQ", __func__);
2074 			nvme_opc_async_event_req(sc, cmd, &compl);
2075 			break;
2076 		case NVME_OPC_FORMAT_NVM:
2077 			DPRINTF("%s command FORMAT_NVM", __func__);
2078 			if ((sc->ctrldata.oacs &
2079 			    (1 << NVME_CTRLR_DATA_OACS_FORMAT_SHIFT)) == 0) {
2080 				pci_nvme_status_genc(&compl.status, NVME_SC_INVALID_OPCODE);
2081 				break;
2082 			}
2083 			nvme_opc_format_nvm(sc, cmd, &compl);
2084 			break;
2085 		case NVME_OPC_SECURITY_SEND:
2086 		case NVME_OPC_SECURITY_RECEIVE:
2087 		case NVME_OPC_SANITIZE:
2088 		case NVME_OPC_GET_LBA_STATUS:
2089 			DPRINTF("%s command OPC=%#x (unsupported)", __func__,
2090 			    cmd->opc);
2091 			/* Valid but unsupported opcodes */
2092 			pci_nvme_status_genc(&compl.status, NVME_SC_INVALID_FIELD);
2093 			break;
2094 		default:
2095 			DPRINTF("%s command OPC=%#X (not implemented)",
2096 			    __func__,
2097 			    cmd->opc);
2098 			pci_nvme_status_genc(&compl.status, NVME_SC_INVALID_OPCODE);
2099 		}
2100 		sqhead = (sqhead + 1) % sq->size;
2101 
2102 		if (NVME_COMPLETION_VALID(compl)) {
2103 			pci_nvme_cq_update(sc, &sc->compl_queues[0],
2104 			    compl.cdw0,
2105 			    cmd->cid,
2106 			    0,		/* SQID */
2107 			    compl.status);
2108 		}
2109 	}
2110 
2111 	DPRINTF("setting sqhead %u", sqhead);
2112 	sq->head = sqhead;
2113 
2114 	if (cq->head != cq->tail)
2115 		pci_generate_msix(sc->nsc_pi, 0);
2116 
2117 	pthread_mutex_unlock(&sq->mtx);
2118 }
2119 
2120 /*
2121  * Update the Write and Read statistics reported in SMART data
2122  *
2123  * NVMe defines "data unit" as thousand's of 512 byte blocks and is rounded up.
2124  * E.g. 1 data unit is 1 - 1,000 512 byte blocks. 3 data units are 2,001 - 3,000
2125  * 512 byte blocks. Rounding up is achieved by initializing the remainder to 999.
2126  */
2127 static void
2128 pci_nvme_stats_write_read_update(struct pci_nvme_softc *sc, uint8_t opc,
2129     size_t bytes, uint16_t status)
2130 {
2131 
2132 	pthread_mutex_lock(&sc->mtx);
2133 	switch (opc) {
2134 	case NVME_OPC_WRITE:
2135 		sc->write_commands++;
2136 		if (status != NVME_SC_SUCCESS)
2137 			break;
2138 		sc->write_dunits_remainder += (bytes / 512);
2139 		while (sc->write_dunits_remainder >= 1000) {
2140 			sc->write_data_units++;
2141 			sc->write_dunits_remainder -= 1000;
2142 		}
2143 		break;
2144 	case NVME_OPC_READ:
2145 		sc->read_commands++;
2146 		if (status != NVME_SC_SUCCESS)
2147 			break;
2148 		sc->read_dunits_remainder += (bytes / 512);
2149 		while (sc->read_dunits_remainder >= 1000) {
2150 			sc->read_data_units++;
2151 			sc->read_dunits_remainder -= 1000;
2152 		}
2153 		break;
2154 	default:
2155 		DPRINTF("%s: Invalid OPC 0x%02x for stats", __func__, opc);
2156 		break;
2157 	}
2158 	pthread_mutex_unlock(&sc->mtx);
2159 }
2160 
2161 /*
2162  * Check if the combination of Starting LBA (slba) and number of blocks
2163  * exceeds the range of the underlying storage.
2164  *
2165  * Because NVMe specifies the SLBA in blocks as a uint64_t and blockif stores
2166  * the capacity in bytes as a uint64_t, care must be taken to avoid integer
2167  * overflow.
2168  */
2169 static bool
2170 pci_nvme_out_of_range(struct pci_nvme_blockstore *nvstore, uint64_t slba,
2171     uint32_t nblocks)
2172 {
2173 	size_t	offset, bytes;
2174 
2175 	/* Overflow check of multiplying Starting LBA by the sector size */
2176 	if (slba >> (64 - nvstore->sectsz_bits))
2177 		return (true);
2178 
2179 	offset = slba << nvstore->sectsz_bits;
2180 	bytes = nblocks << nvstore->sectsz_bits;
2181 
2182 	/* Overflow check of Number of Logical Blocks */
2183 	if ((nvstore->size <= offset) || ((nvstore->size - offset) < bytes))
2184 		return (true);
2185 
2186 	return (false);
2187 }
2188 
2189 static int
2190 pci_nvme_append_iov_req(struct pci_nvme_softc *sc __unused,
2191     struct pci_nvme_ioreq *req, uint64_t gpaddr, size_t size, uint64_t offset)
2192 {
2193 	int iovidx;
2194 	bool range_is_contiguous;
2195 
2196 	if (req == NULL)
2197 		return (-1);
2198 
2199 	if (req->io_req.br_iovcnt == NVME_MAX_IOVEC) {
2200 		return (-1);
2201 	}
2202 
2203 	/*
2204 	 * Minimize the number of IOVs by concatenating contiguous address
2205 	 * ranges. If the IOV count is zero, there is no previous range to
2206 	 * concatenate.
2207 	 */
2208 	if (req->io_req.br_iovcnt == 0)
2209 		range_is_contiguous = false;
2210 	else
2211 		range_is_contiguous = (req->prev_gpaddr + req->prev_size) == gpaddr;
2212 
2213 	if (range_is_contiguous) {
2214 		iovidx = req->io_req.br_iovcnt - 1;
2215 
2216 		req->io_req.br_iov[iovidx].iov_base =
2217 		    paddr_guest2host(req->sc->nsc_pi->pi_vmctx,
2218 				     req->prev_gpaddr, size);
2219 		if (req->io_req.br_iov[iovidx].iov_base == NULL)
2220 			return (-1);
2221 
2222 		req->prev_size += size;
2223 		req->io_req.br_resid += size;
2224 
2225 		req->io_req.br_iov[iovidx].iov_len = req->prev_size;
2226 	} else {
2227 		iovidx = req->io_req.br_iovcnt;
2228 		if (iovidx == 0) {
2229 			req->io_req.br_offset = offset;
2230 			req->io_req.br_resid = 0;
2231 			req->io_req.br_param = req;
2232 		}
2233 
2234 		req->io_req.br_iov[iovidx].iov_base =
2235 		    paddr_guest2host(req->sc->nsc_pi->pi_vmctx,
2236 				     gpaddr, size);
2237 		if (req->io_req.br_iov[iovidx].iov_base == NULL)
2238 			return (-1);
2239 
2240 		req->io_req.br_iov[iovidx].iov_len = size;
2241 
2242 		req->prev_gpaddr = gpaddr;
2243 		req->prev_size = size;
2244 		req->io_req.br_resid += size;
2245 
2246 		req->io_req.br_iovcnt++;
2247 	}
2248 
2249 	return (0);
2250 }
2251 
2252 static void
2253 pci_nvme_set_completion(struct pci_nvme_softc *sc,
2254     struct nvme_submission_queue *sq, int sqid, uint16_t cid, uint16_t status)
2255 {
2256 	struct nvme_completion_queue *cq = &sc->compl_queues[sq->cqid];
2257 
2258 	DPRINTF("%s sqid %d cqid %u cid %u status: 0x%x 0x%x",
2259 		 __func__, sqid, sq->cqid, cid, NVME_STATUS_GET_SCT(status),
2260 		 NVME_STATUS_GET_SC(status));
2261 
2262 	pci_nvme_cq_update(sc, cq, 0, cid, sqid, status);
2263 
2264 	if (cq->head != cq->tail) {
2265 		if (cq->intr_en & NVME_CQ_INTEN) {
2266 			pci_generate_msix(sc->nsc_pi, cq->intr_vec);
2267 		} else {
2268 			DPRINTF("%s: CQ%u interrupt disabled",
2269 						__func__, sq->cqid);
2270 		}
2271 	}
2272 }
2273 
2274 static void
2275 pci_nvme_release_ioreq(struct pci_nvme_softc *sc, struct pci_nvme_ioreq *req)
2276 {
2277 	req->sc = NULL;
2278 	req->nvme_sq = NULL;
2279 	req->sqid = 0;
2280 
2281 	pthread_mutex_lock(&sc->mtx);
2282 
2283 	STAILQ_INSERT_TAIL(&sc->ioreqs_free, req, link);
2284 	sc->pending_ios--;
2285 
2286 	/* when no more IO pending, can set to ready if device reset/enabled */
2287 	if (sc->pending_ios == 0 &&
2288 	    NVME_CC_GET_EN(sc->regs.cc) && !(NVME_CSTS_GET_RDY(sc->regs.csts)))
2289 		sc->regs.csts |= NVME_CSTS_RDY;
2290 
2291 	pthread_mutex_unlock(&sc->mtx);
2292 
2293 	sem_post(&sc->iosemlock);
2294 }
2295 
2296 static struct pci_nvme_ioreq *
2297 pci_nvme_get_ioreq(struct pci_nvme_softc *sc)
2298 {
2299 	struct pci_nvme_ioreq *req = NULL;
2300 
2301 	sem_wait(&sc->iosemlock);
2302 	pthread_mutex_lock(&sc->mtx);
2303 
2304 	req = STAILQ_FIRST(&sc->ioreqs_free);
2305 	assert(req != NULL);
2306 	STAILQ_REMOVE_HEAD(&sc->ioreqs_free, link);
2307 
2308 	req->sc = sc;
2309 
2310 	sc->pending_ios++;
2311 
2312 	pthread_mutex_unlock(&sc->mtx);
2313 
2314 	req->io_req.br_iovcnt = 0;
2315 	req->io_req.br_offset = 0;
2316 	req->io_req.br_resid = 0;
2317 	req->io_req.br_param = req;
2318 	req->prev_gpaddr = 0;
2319 	req->prev_size = 0;
2320 
2321 	return req;
2322 }
2323 
2324 static void
2325 pci_nvme_io_done(struct blockif_req *br, int err)
2326 {
2327 	struct pci_nvme_ioreq *req = br->br_param;
2328 	struct nvme_submission_queue *sq = req->nvme_sq;
2329 	uint16_t code, status;
2330 
2331 	DPRINTF("%s error %d %s", __func__, err, strerror(err));
2332 
2333 	/* TODO return correct error */
2334 	code = err ? NVME_SC_DATA_TRANSFER_ERROR : NVME_SC_SUCCESS;
2335 	status = 0;
2336 	pci_nvme_status_genc(&status, code);
2337 
2338 	pci_nvme_set_completion(req->sc, sq, req->sqid, req->cid, status);
2339 	pci_nvme_stats_write_read_update(req->sc, req->opc,
2340 	    req->bytes, status);
2341 	pci_nvme_release_ioreq(req->sc, req);
2342 }
2343 
2344 /*
2345  * Implements the Flush command. The specification states:
2346  *    If a volatile write cache is not present, Flush commands complete
2347  *    successfully and have no effect
2348  * in the description of the Volatile Write Cache (VWC) field of the Identify
2349  * Controller data. Therefore, set status to Success if the command is
2350  * not supported (i.e. RAM or as indicated by the blockif).
2351  */
2352 static bool
2353 nvme_opc_flush(struct pci_nvme_softc *sc __unused,
2354     struct nvme_command *cmd __unused,
2355     struct pci_nvme_blockstore *nvstore,
2356     struct pci_nvme_ioreq *req,
2357     uint16_t *status)
2358 {
2359 	bool pending = false;
2360 
2361 	if (nvstore->type == NVME_STOR_RAM) {
2362 		pci_nvme_status_genc(status, NVME_SC_SUCCESS);
2363 	} else {
2364 		int err;
2365 
2366 		req->io_req.br_callback = pci_nvme_io_done;
2367 
2368 		err = blockif_flush(nvstore->ctx, &req->io_req);
2369 		switch (err) {
2370 		case 0:
2371 			pending = true;
2372 			break;
2373 		case EOPNOTSUPP:
2374 			pci_nvme_status_genc(status, NVME_SC_SUCCESS);
2375 			break;
2376 		default:
2377 			pci_nvme_status_genc(status, NVME_SC_INTERNAL_DEVICE_ERROR);
2378 		}
2379 	}
2380 
2381 	return (pending);
2382 }
2383 
2384 static uint16_t
2385 nvme_write_read_ram(struct pci_nvme_softc *sc,
2386     struct pci_nvme_blockstore *nvstore,
2387     uint64_t prp1, uint64_t prp2,
2388     size_t offset, uint64_t bytes,
2389     bool is_write)
2390 {
2391 	uint8_t *buf = nvstore->ctx;
2392 	enum nvme_copy_dir dir;
2393 	uint16_t status;
2394 
2395 	if (is_write)
2396 		dir = NVME_COPY_TO_PRP;
2397 	else
2398 		dir = NVME_COPY_FROM_PRP;
2399 
2400 	status = 0;
2401 	if (nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, prp1, prp2,
2402 	    buf + offset, bytes, dir))
2403 		pci_nvme_status_genc(&status,
2404 		    NVME_SC_DATA_TRANSFER_ERROR);
2405 	else
2406 		pci_nvme_status_genc(&status, NVME_SC_SUCCESS);
2407 
2408 	return (status);
2409 }
2410 
2411 static uint16_t
2412 nvme_write_read_blockif(struct pci_nvme_softc *sc,
2413     struct pci_nvme_blockstore *nvstore,
2414     struct pci_nvme_ioreq *req,
2415     uint64_t prp1, uint64_t prp2,
2416     size_t offset, uint64_t bytes,
2417     bool is_write)
2418 {
2419 	uint64_t size;
2420 	int err;
2421 	uint16_t status = NVME_NO_STATUS;
2422 
2423 	size = MIN(PAGE_SIZE - (prp1 % PAGE_SIZE), bytes);
2424 	if (pci_nvme_append_iov_req(sc, req, prp1, size, offset)) {
2425 		err = -1;
2426 		goto out;
2427 	}
2428 
2429 	offset += size;
2430 	bytes  -= size;
2431 
2432 	if (bytes == 0) {
2433 		;
2434 	} else if (bytes <= PAGE_SIZE) {
2435 		size = bytes;
2436 		if (pci_nvme_append_iov_req(sc, req, prp2, size, offset)) {
2437 			err = -1;
2438 			goto out;
2439 		}
2440 	} else {
2441 		void *vmctx = sc->nsc_pi->pi_vmctx;
2442 		uint64_t *prp_list = &prp2;
2443 		uint64_t *last = prp_list;
2444 
2445 		/* PRP2 is pointer to a physical region page list */
2446 		while (bytes) {
2447 			/* Last entry in list points to the next list */
2448 			if ((prp_list == last) && (bytes > PAGE_SIZE)) {
2449 				uint64_t prp = *prp_list;
2450 
2451 				prp_list = paddr_guest2host(vmctx, prp,
2452 				    PAGE_SIZE - (prp % PAGE_SIZE));
2453 				if (prp_list == NULL) {
2454 					err = -1;
2455 					goto out;
2456 				}
2457 				last = prp_list + (NVME_PRP2_ITEMS - 1);
2458 			}
2459 
2460 			size = MIN(bytes, PAGE_SIZE);
2461 
2462 			if (pci_nvme_append_iov_req(sc, req, *prp_list, size,
2463 			    offset)) {
2464 				err = -1;
2465 				goto out;
2466 			}
2467 
2468 			offset += size;
2469 			bytes  -= size;
2470 
2471 			prp_list++;
2472 		}
2473 	}
2474 	req->io_req.br_callback = pci_nvme_io_done;
2475 	if (is_write)
2476 		err = blockif_write(nvstore->ctx, &req->io_req);
2477 	else
2478 		err = blockif_read(nvstore->ctx, &req->io_req);
2479 out:
2480 	if (err)
2481 		pci_nvme_status_genc(&status, NVME_SC_DATA_TRANSFER_ERROR);
2482 
2483 	return (status);
2484 }
2485 
2486 static bool
2487 nvme_opc_write_read(struct pci_nvme_softc *sc,
2488     struct nvme_command *cmd,
2489     struct pci_nvme_blockstore *nvstore,
2490     struct pci_nvme_ioreq *req,
2491     uint16_t *status)
2492 {
2493 	uint64_t lba, nblocks, bytes;
2494 	size_t offset;
2495 	bool is_write = cmd->opc == NVME_OPC_WRITE;
2496 	bool pending = false;
2497 
2498 	lba = ((uint64_t)cmd->cdw11 << 32) | cmd->cdw10;
2499 	nblocks = (cmd->cdw12 & 0xFFFF) + 1;
2500 	bytes = nblocks << nvstore->sectsz_bits;
2501 	if (bytes > NVME_MAX_DATA_SIZE) {
2502 		WPRINTF("%s command would exceed MDTS", __func__);
2503 		pci_nvme_status_genc(status, NVME_SC_INVALID_FIELD);
2504 		goto out;
2505 	}
2506 
2507 	if (pci_nvme_out_of_range(nvstore, lba, nblocks)) {
2508 		WPRINTF("%s command would exceed LBA range(slba=%#lx nblocks=%#lx)",
2509 		    __func__, lba, nblocks);
2510 		pci_nvme_status_genc(status, NVME_SC_LBA_OUT_OF_RANGE);
2511 		goto out;
2512 	}
2513 
2514 	offset = lba << nvstore->sectsz_bits;
2515 
2516 	req->bytes = bytes;
2517 	req->io_req.br_offset = lba;
2518 
2519 	/* PRP bits 1:0 must be zero */
2520 	cmd->prp1 &= ~0x3UL;
2521 	cmd->prp2 &= ~0x3UL;
2522 
2523 	if (nvstore->type == NVME_STOR_RAM) {
2524 		*status = nvme_write_read_ram(sc, nvstore, cmd->prp1,
2525 		    cmd->prp2, offset, bytes, is_write);
2526 	} else {
2527 		*status = nvme_write_read_blockif(sc, nvstore, req,
2528 		    cmd->prp1, cmd->prp2, offset, bytes, is_write);
2529 
2530 		if (*status == NVME_NO_STATUS)
2531 			pending = true;
2532 	}
2533 out:
2534 	if (!pending)
2535 		pci_nvme_stats_write_read_update(sc, cmd->opc, bytes, *status);
2536 
2537 	return (pending);
2538 }
2539 
2540 static void
2541 pci_nvme_dealloc_sm(struct blockif_req *br, int err)
2542 {
2543 	struct pci_nvme_ioreq *req = br->br_param;
2544 	struct pci_nvme_softc *sc = req->sc;
2545 	bool done = true;
2546 	uint16_t status;
2547 
2548 	status = 0;
2549 	if (err) {
2550 		pci_nvme_status_genc(&status, NVME_SC_INTERNAL_DEVICE_ERROR);
2551 	} else if ((req->prev_gpaddr + 1) == (req->prev_size)) {
2552 		pci_nvme_status_genc(&status, NVME_SC_SUCCESS);
2553 	} else {
2554 		struct iovec *iov = req->io_req.br_iov;
2555 
2556 		req->prev_gpaddr++;
2557 		iov += req->prev_gpaddr;
2558 
2559 		/* The iov_* values already include the sector size */
2560 		req->io_req.br_offset = (off_t)iov->iov_base;
2561 		req->io_req.br_resid = iov->iov_len;
2562 		if (blockif_delete(sc->nvstore.ctx, &req->io_req)) {
2563 			pci_nvme_status_genc(&status,
2564 			    NVME_SC_INTERNAL_DEVICE_ERROR);
2565 		} else
2566 			done = false;
2567 	}
2568 
2569 	if (done) {
2570 		pci_nvme_set_completion(sc, req->nvme_sq, req->sqid, req->cid,
2571 		    status);
2572 		pci_nvme_release_ioreq(sc, req);
2573 	}
2574 }
2575 
2576 static bool
2577 nvme_opc_dataset_mgmt(struct pci_nvme_softc *sc,
2578     struct nvme_command *cmd,
2579     struct pci_nvme_blockstore *nvstore,
2580     struct pci_nvme_ioreq *req,
2581     uint16_t *status)
2582 {
2583 	struct nvme_dsm_range *range = NULL;
2584 	uint32_t nr, r, non_zero, dr;
2585 	int err;
2586 	bool pending = false;
2587 
2588 	if ((sc->ctrldata.oncs & NVME_ONCS_DSM) == 0) {
2589 		pci_nvme_status_genc(status, NVME_SC_INVALID_OPCODE);
2590 		goto out;
2591 	}
2592 
2593 	nr = cmd->cdw10 & 0xff;
2594 
2595 	/* copy locally because a range entry could straddle PRPs */
2596 	range = calloc(1, NVME_MAX_DSM_TRIM);
2597 	if (range == NULL) {
2598 		pci_nvme_status_genc(status, NVME_SC_INTERNAL_DEVICE_ERROR);
2599 		goto out;
2600 	}
2601 	nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, cmd->prp1, cmd->prp2,
2602 	    (uint8_t *)range, NVME_MAX_DSM_TRIM, NVME_COPY_FROM_PRP);
2603 
2604 	/* Check for invalid ranges and the number of non-zero lengths */
2605 	non_zero = 0;
2606 	for (r = 0; r <= nr; r++) {
2607 		if (pci_nvme_out_of_range(nvstore,
2608 		    range[r].starting_lba, range[r].length)) {
2609 			pci_nvme_status_genc(status, NVME_SC_LBA_OUT_OF_RANGE);
2610 			goto out;
2611 		}
2612 		if (range[r].length != 0)
2613 			non_zero++;
2614 	}
2615 
2616 	if (cmd->cdw11 & NVME_DSM_ATTR_DEALLOCATE) {
2617 		size_t offset, bytes;
2618 		int sectsz_bits = sc->nvstore.sectsz_bits;
2619 
2620 		/*
2621 		 * DSM calls are advisory only, and compliant controllers
2622 		 * may choose to take no actions (i.e. return Success).
2623 		 */
2624 		if (!nvstore->deallocate) {
2625 			pci_nvme_status_genc(status, NVME_SC_SUCCESS);
2626 			goto out;
2627 		}
2628 
2629 		/* If all ranges have a zero length, return Success */
2630 		if (non_zero == 0) {
2631 			pci_nvme_status_genc(status, NVME_SC_SUCCESS);
2632 			goto out;
2633 		}
2634 
2635 		if (req == NULL) {
2636 			pci_nvme_status_genc(status, NVME_SC_INTERNAL_DEVICE_ERROR);
2637 			goto out;
2638 		}
2639 
2640 		offset = range[0].starting_lba << sectsz_bits;
2641 		bytes = range[0].length << sectsz_bits;
2642 
2643 		/*
2644 		 * If the request is for more than a single range, store
2645 		 * the ranges in the br_iov. Optimize for the common case
2646 		 * of a single range.
2647 		 *
2648 		 * Note that NVMe Number of Ranges is a zero based value
2649 		 */
2650 		req->io_req.br_iovcnt = 0;
2651 		req->io_req.br_offset = offset;
2652 		req->io_req.br_resid = bytes;
2653 
2654 		if (nr == 0) {
2655 			req->io_req.br_callback = pci_nvme_io_done;
2656 		} else {
2657 			struct iovec *iov = req->io_req.br_iov;
2658 
2659 			for (r = 0, dr = 0; r <= nr; r++) {
2660 				offset = range[r].starting_lba << sectsz_bits;
2661 				bytes = range[r].length << sectsz_bits;
2662 				if (bytes == 0)
2663 					continue;
2664 
2665 				if ((nvstore->size - offset) < bytes) {
2666 					pci_nvme_status_genc(status,
2667 					    NVME_SC_LBA_OUT_OF_RANGE);
2668 					goto out;
2669 				}
2670 				iov[dr].iov_base = (void *)offset;
2671 				iov[dr].iov_len = bytes;
2672 				dr++;
2673 			}
2674 			req->io_req.br_callback = pci_nvme_dealloc_sm;
2675 
2676 			/*
2677 			 * Use prev_gpaddr to track the current entry and
2678 			 * prev_size to track the number of entries
2679 			 */
2680 			req->prev_gpaddr = 0;
2681 			req->prev_size = dr;
2682 		}
2683 
2684 		err = blockif_delete(nvstore->ctx, &req->io_req);
2685 		if (err)
2686 			pci_nvme_status_genc(status, NVME_SC_INTERNAL_DEVICE_ERROR);
2687 		else
2688 			pending = true;
2689 	}
2690 out:
2691 	free(range);
2692 	return (pending);
2693 }
2694 
2695 static void
2696 pci_nvme_handle_io_cmd(struct pci_nvme_softc* sc, uint16_t idx)
2697 {
2698 	struct nvme_submission_queue *sq;
2699 	uint16_t status;
2700 	uint16_t sqhead;
2701 
2702 	/* handle all submissions up to sq->tail index */
2703 	sq = &sc->submit_queues[idx];
2704 
2705 	pthread_mutex_lock(&sq->mtx);
2706 
2707 	sqhead = sq->head;
2708 	DPRINTF("nvme_handle_io qid %u head %u tail %u cmdlist %p",
2709 	         idx, sqhead, sq->tail, sq->qbase);
2710 
2711 	while (sqhead != atomic_load_acq_short(&sq->tail)) {
2712 		struct nvme_command *cmd;
2713 		struct pci_nvme_ioreq *req;
2714 		uint32_t nsid;
2715 		bool pending;
2716 
2717 		pending = false;
2718 		req = NULL;
2719 		status = 0;
2720 
2721 		cmd = &sq->qbase[sqhead];
2722 		sqhead = (sqhead + 1) % sq->size;
2723 
2724 		nsid = le32toh(cmd->nsid);
2725 		if ((nsid == 0) || (nsid > sc->ctrldata.nn)) {
2726 			pci_nvme_status_genc(&status,
2727 			    NVME_SC_INVALID_NAMESPACE_OR_FORMAT);
2728 			status |=
2729 			    NVME_STATUS_DNR_MASK << NVME_STATUS_DNR_SHIFT;
2730 			goto complete;
2731  		}
2732 
2733 		req = pci_nvme_get_ioreq(sc);
2734 		if (req == NULL) {
2735 			pci_nvme_status_genc(&status,
2736 			    NVME_SC_INTERNAL_DEVICE_ERROR);
2737 			WPRINTF("%s: unable to allocate IO req", __func__);
2738 			goto complete;
2739 		}
2740 		req->nvme_sq = sq;
2741 		req->sqid = idx;
2742 		req->opc = cmd->opc;
2743 		req->cid = cmd->cid;
2744 		req->nsid = cmd->nsid;
2745 
2746 		switch (cmd->opc) {
2747 		case NVME_OPC_FLUSH:
2748 			pending = nvme_opc_flush(sc, cmd, &sc->nvstore,
2749 			    req, &status);
2750  			break;
2751 		case NVME_OPC_WRITE:
2752 		case NVME_OPC_READ:
2753 			pending = nvme_opc_write_read(sc, cmd, &sc->nvstore,
2754 			    req, &status);
2755 			break;
2756 		case NVME_OPC_WRITE_ZEROES:
2757 			/* TODO: write zeroes
2758 			WPRINTF("%s write zeroes lba 0x%lx blocks %u",
2759 			        __func__, lba, cmd->cdw12 & 0xFFFF); */
2760 			pci_nvme_status_genc(&status, NVME_SC_SUCCESS);
2761 			break;
2762 		case NVME_OPC_DATASET_MANAGEMENT:
2763  			pending = nvme_opc_dataset_mgmt(sc, cmd, &sc->nvstore,
2764 			    req, &status);
2765 			break;
2766  		default:
2767  			WPRINTF("%s unhandled io command 0x%x",
2768 			    __func__, cmd->opc);
2769 			pci_nvme_status_genc(&status, NVME_SC_INVALID_OPCODE);
2770 		}
2771 complete:
2772 		if (!pending) {
2773 			pci_nvme_set_completion(sc, sq, idx, cmd->cid, status);
2774 			if (req != NULL)
2775 				pci_nvme_release_ioreq(sc, req);
2776 		}
2777 	}
2778 
2779 	sq->head = sqhead;
2780 
2781 	pthread_mutex_unlock(&sq->mtx);
2782 }
2783 
2784 static void
2785 pci_nvme_handle_doorbell(struct pci_nvme_softc* sc,
2786 	uint64_t idx, int is_sq, uint64_t value)
2787 {
2788 	DPRINTF("nvme doorbell %lu, %s, val 0x%lx",
2789 	        idx, is_sq ? "SQ" : "CQ", value & 0xFFFF);
2790 
2791 	if (is_sq) {
2792 		if (idx > sc->num_squeues) {
2793 			WPRINTF("%s queue index %lu overflow from "
2794 			         "guest (max %u)",
2795 			         __func__, idx, sc->num_squeues);
2796 			return;
2797 		}
2798 
2799 		atomic_store_short(&sc->submit_queues[idx].tail,
2800 		                   (uint16_t)value);
2801 
2802 		if (idx == 0) {
2803 			pci_nvme_handle_admin_cmd(sc, value);
2804 		} else {
2805 			/* submission queue; handle new entries in SQ */
2806 			if (idx > sc->num_squeues) {
2807 				WPRINTF("%s SQ index %lu overflow from "
2808 				         "guest (max %u)",
2809 				         __func__, idx, sc->num_squeues);
2810 				return;
2811 			}
2812 			pci_nvme_handle_io_cmd(sc, (uint16_t)idx);
2813 		}
2814 	} else {
2815 		if (idx > sc->num_cqueues) {
2816 			WPRINTF("%s queue index %lu overflow from "
2817 			         "guest (max %u)",
2818 			         __func__, idx, sc->num_cqueues);
2819 			return;
2820 		}
2821 
2822 		atomic_store_short(&sc->compl_queues[idx].head,
2823 				(uint16_t)value);
2824 	}
2825 }
2826 
2827 static void
2828 pci_nvme_bar0_reg_dumps(const char *func, uint64_t offset, int iswrite)
2829 {
2830 	const char *s = iswrite ? "WRITE" : "READ";
2831 
2832 	switch (offset) {
2833 	case NVME_CR_CAP_LOW:
2834 		DPRINTF("%s %s NVME_CR_CAP_LOW", func, s);
2835 		break;
2836 	case NVME_CR_CAP_HI:
2837 		DPRINTF("%s %s NVME_CR_CAP_HI", func, s);
2838 		break;
2839 	case NVME_CR_VS:
2840 		DPRINTF("%s %s NVME_CR_VS", func, s);
2841 		break;
2842 	case NVME_CR_INTMS:
2843 		DPRINTF("%s %s NVME_CR_INTMS", func, s);
2844 		break;
2845 	case NVME_CR_INTMC:
2846 		DPRINTF("%s %s NVME_CR_INTMC", func, s);
2847 		break;
2848 	case NVME_CR_CC:
2849 		DPRINTF("%s %s NVME_CR_CC", func, s);
2850 		break;
2851 	case NVME_CR_CSTS:
2852 		DPRINTF("%s %s NVME_CR_CSTS", func, s);
2853 		break;
2854 	case NVME_CR_NSSR:
2855 		DPRINTF("%s %s NVME_CR_NSSR", func, s);
2856 		break;
2857 	case NVME_CR_AQA:
2858 		DPRINTF("%s %s NVME_CR_AQA", func, s);
2859 		break;
2860 	case NVME_CR_ASQ_LOW:
2861 		DPRINTF("%s %s NVME_CR_ASQ_LOW", func, s);
2862 		break;
2863 	case NVME_CR_ASQ_HI:
2864 		DPRINTF("%s %s NVME_CR_ASQ_HI", func, s);
2865 		break;
2866 	case NVME_CR_ACQ_LOW:
2867 		DPRINTF("%s %s NVME_CR_ACQ_LOW", func, s);
2868 		break;
2869 	case NVME_CR_ACQ_HI:
2870 		DPRINTF("%s %s NVME_CR_ACQ_HI", func, s);
2871 		break;
2872 	default:
2873 		DPRINTF("unknown nvme bar-0 offset 0x%lx", offset);
2874 	}
2875 
2876 }
2877 
2878 static void
2879 pci_nvme_write_bar_0(struct pci_nvme_softc *sc, uint64_t offset, int size,
2880     uint64_t value)
2881 {
2882 	uint32_t ccreg;
2883 
2884 	if (offset >= NVME_DOORBELL_OFFSET) {
2885 		uint64_t belloffset = offset - NVME_DOORBELL_OFFSET;
2886 		uint64_t idx = belloffset / 8; /* door bell size = 2*int */
2887 		int is_sq = (belloffset % 8) < 4;
2888 
2889 		if ((sc->regs.csts & NVME_CSTS_RDY) == 0) {
2890 			WPRINTF("doorbell write prior to RDY (offset=%#lx)\n",
2891 			    offset);
2892 			return;
2893 		}
2894 
2895 		if (belloffset > ((sc->max_queues+1) * 8 - 4)) {
2896 			WPRINTF("guest attempted an overflow write offset "
2897 			         "0x%lx, val 0x%lx in %s",
2898 			         offset, value, __func__);
2899 			return;
2900 		}
2901 
2902 		if (is_sq) {
2903 			if (sc->submit_queues[idx].qbase == NULL)
2904 				return;
2905 		} else if (sc->compl_queues[idx].qbase == NULL)
2906 			return;
2907 
2908 		pci_nvme_handle_doorbell(sc, idx, is_sq, value);
2909 		return;
2910 	}
2911 
2912 	DPRINTF("nvme-write offset 0x%lx, size %d, value 0x%lx",
2913 	        offset, size, value);
2914 
2915 	if (size != 4) {
2916 		WPRINTF("guest wrote invalid size %d (offset 0x%lx, "
2917 		         "val 0x%lx) to bar0 in %s",
2918 		         size, offset, value, __func__);
2919 		/* TODO: shutdown device */
2920 		return;
2921 	}
2922 
2923 	pci_nvme_bar0_reg_dumps(__func__, offset, 1);
2924 
2925 	pthread_mutex_lock(&sc->mtx);
2926 
2927 	switch (offset) {
2928 	case NVME_CR_CAP_LOW:
2929 	case NVME_CR_CAP_HI:
2930 		/* readonly */
2931 		break;
2932 	case NVME_CR_VS:
2933 		/* readonly */
2934 		break;
2935 	case NVME_CR_INTMS:
2936 		/* MSI-X, so ignore */
2937 		break;
2938 	case NVME_CR_INTMC:
2939 		/* MSI-X, so ignore */
2940 		break;
2941 	case NVME_CR_CC:
2942 		ccreg = (uint32_t)value;
2943 
2944 		DPRINTF("%s NVME_CR_CC en %x css %x shn %x iosqes %u "
2945 		         "iocqes %u",
2946 		        __func__,
2947 			 NVME_CC_GET_EN(ccreg), NVME_CC_GET_CSS(ccreg),
2948 			 NVME_CC_GET_SHN(ccreg), NVME_CC_GET_IOSQES(ccreg),
2949 			 NVME_CC_GET_IOCQES(ccreg));
2950 
2951 		if (NVME_CC_GET_SHN(ccreg)) {
2952 			/* perform shutdown - flush out data to backend */
2953 			sc->regs.csts &= ~(NVME_CSTS_REG_SHST_MASK <<
2954 			    NVME_CSTS_REG_SHST_SHIFT);
2955 			sc->regs.csts |= NVME_SHST_COMPLETE <<
2956 			    NVME_CSTS_REG_SHST_SHIFT;
2957 		}
2958 		if (NVME_CC_GET_EN(ccreg) != NVME_CC_GET_EN(sc->regs.cc)) {
2959 			if (NVME_CC_GET_EN(ccreg) == 0)
2960 				/* transition 1-> causes controller reset */
2961 				pci_nvme_reset_locked(sc);
2962 			else
2963 				pci_nvme_init_controller(sc);
2964 		}
2965 
2966 		/* Insert the iocqes, iosqes and en bits from the write */
2967 		sc->regs.cc &= ~NVME_CC_WRITE_MASK;
2968 		sc->regs.cc |= ccreg & NVME_CC_WRITE_MASK;
2969 		if (NVME_CC_GET_EN(ccreg) == 0) {
2970 			/* Insert the ams, mps and css bit fields */
2971 			sc->regs.cc &= ~NVME_CC_NEN_WRITE_MASK;
2972 			sc->regs.cc |= ccreg & NVME_CC_NEN_WRITE_MASK;
2973 			sc->regs.csts &= ~NVME_CSTS_RDY;
2974 		} else if ((sc->pending_ios == 0) &&
2975 		    !(sc->regs.csts & NVME_CSTS_CFS)) {
2976 			sc->regs.csts |= NVME_CSTS_RDY;
2977 		}
2978 		break;
2979 	case NVME_CR_CSTS:
2980 		break;
2981 	case NVME_CR_NSSR:
2982 		/* ignore writes; don't support subsystem reset */
2983 		break;
2984 	case NVME_CR_AQA:
2985 		sc->regs.aqa = (uint32_t)value;
2986 		break;
2987 	case NVME_CR_ASQ_LOW:
2988 		sc->regs.asq = (sc->regs.asq & (0xFFFFFFFF00000000)) |
2989 		               (0xFFFFF000 & value);
2990 		break;
2991 	case NVME_CR_ASQ_HI:
2992 		sc->regs.asq = (sc->regs.asq & (0x00000000FFFFFFFF)) |
2993 		               (value << 32);
2994 		break;
2995 	case NVME_CR_ACQ_LOW:
2996 		sc->regs.acq = (sc->regs.acq & (0xFFFFFFFF00000000)) |
2997 		               (0xFFFFF000 & value);
2998 		break;
2999 	case NVME_CR_ACQ_HI:
3000 		sc->regs.acq = (sc->regs.acq & (0x00000000FFFFFFFF)) |
3001 		               (value << 32);
3002 		break;
3003 	default:
3004 		DPRINTF("%s unknown offset 0x%lx, value 0x%lx size %d",
3005 		         __func__, offset, value, size);
3006 	}
3007 	pthread_mutex_unlock(&sc->mtx);
3008 }
3009 
3010 static void
3011 pci_nvme_write(struct pci_devinst *pi, int baridx, uint64_t offset, int size,
3012     uint64_t value)
3013 {
3014 	struct pci_nvme_softc* sc = pi->pi_arg;
3015 
3016 	if (baridx == pci_msix_table_bar(pi) ||
3017 	    baridx == pci_msix_pba_bar(pi)) {
3018 		DPRINTF("nvme-write baridx %d, msix: off 0x%lx, size %d, "
3019 		         " value 0x%lx", baridx, offset, size, value);
3020 
3021 		pci_emul_msix_twrite(pi, offset, size, value);
3022 		return;
3023 	}
3024 
3025 	switch (baridx) {
3026 	case 0:
3027 		pci_nvme_write_bar_0(sc, offset, size, value);
3028 		break;
3029 
3030 	default:
3031 		DPRINTF("%s unknown baridx %d, val 0x%lx",
3032 		         __func__, baridx, value);
3033 	}
3034 }
3035 
3036 static uint64_t pci_nvme_read_bar_0(struct pci_nvme_softc* sc,
3037 	uint64_t offset, int size)
3038 {
3039 	uint64_t value;
3040 
3041 	pci_nvme_bar0_reg_dumps(__func__, offset, 0);
3042 
3043 	if (offset < NVME_DOORBELL_OFFSET) {
3044 		void *p = &(sc->regs);
3045 		pthread_mutex_lock(&sc->mtx);
3046 		memcpy(&value, (void *)((uintptr_t)p + offset), size);
3047 		pthread_mutex_unlock(&sc->mtx);
3048 	} else {
3049 		value = 0;
3050                 WPRINTF("pci_nvme: read invalid offset %ld", offset);
3051 	}
3052 
3053 	switch (size) {
3054 	case 1:
3055 		value &= 0xFF;
3056 		break;
3057 	case 2:
3058 		value &= 0xFFFF;
3059 		break;
3060 	case 4:
3061 		value &= 0xFFFFFFFF;
3062 		break;
3063 	}
3064 
3065 	DPRINTF("   nvme-read offset 0x%lx, size %d -> value 0x%x",
3066 	         offset, size, (uint32_t)value);
3067 
3068 	return (value);
3069 }
3070 
3071 
3072 
3073 static uint64_t
3074 pci_nvme_read(struct pci_devinst *pi, int baridx, uint64_t offset, int size)
3075 {
3076 	struct pci_nvme_softc* sc = pi->pi_arg;
3077 
3078 	if (baridx == pci_msix_table_bar(pi) ||
3079 	    baridx == pci_msix_pba_bar(pi)) {
3080 		DPRINTF("nvme-read bar: %d, msix: regoff 0x%lx, size %d",
3081 		        baridx, offset, size);
3082 
3083 		return pci_emul_msix_tread(pi, offset, size);
3084 	}
3085 
3086 	switch (baridx) {
3087 	case 0:
3088        		return pci_nvme_read_bar_0(sc, offset, size);
3089 
3090 	default:
3091 		DPRINTF("unknown bar %d, 0x%lx", baridx, offset);
3092 	}
3093 
3094 	return (0);
3095 }
3096 
3097 static int
3098 pci_nvme_parse_config(struct pci_nvme_softc *sc, nvlist_t *nvl)
3099 {
3100 	char bident[sizeof("XXX:XXX")];
3101 	const char *value;
3102 	uint32_t sectsz;
3103 
3104 	sc->max_queues = NVME_QUEUES;
3105 	sc->max_qentries = NVME_MAX_QENTRIES;
3106 	sc->ioslots = NVME_IOSLOTS;
3107 	sc->num_squeues = sc->max_queues;
3108 	sc->num_cqueues = sc->max_queues;
3109 	sc->dataset_management = NVME_DATASET_MANAGEMENT_AUTO;
3110 	sectsz = 0;
3111 	snprintf(sc->ctrldata.sn, sizeof(sc->ctrldata.sn),
3112 	         "NVME-%d-%d", sc->nsc_pi->pi_slot, sc->nsc_pi->pi_func);
3113 
3114 	value = get_config_value_node(nvl, "maxq");
3115 	if (value != NULL)
3116 		sc->max_queues = atoi(value);
3117 	value = get_config_value_node(nvl, "qsz");
3118 	if (value != NULL) {
3119 		sc->max_qentries = atoi(value);
3120 		if (sc->max_qentries <= 0) {
3121 			EPRINTLN("nvme: Invalid qsz option %d",
3122 			    sc->max_qentries);
3123 			return (-1);
3124 		}
3125 	}
3126 	value = get_config_value_node(nvl, "ioslots");
3127 	if (value != NULL) {
3128 		sc->ioslots = atoi(value);
3129 		if (sc->ioslots <= 0) {
3130 			EPRINTLN("Invalid ioslots option %d", sc->ioslots);
3131 			return (-1);
3132 		}
3133 	}
3134 	value = get_config_value_node(nvl, "sectsz");
3135 	if (value != NULL)
3136 		sectsz = atoi(value);
3137 	value = get_config_value_node(nvl, "ser");
3138 	if (value != NULL) {
3139 		/*
3140 		 * This field indicates the Product Serial Number in
3141 		 * 7-bit ASCII, unused bytes should be space characters.
3142 		 * Ref: NVMe v1.3c.
3143 		 */
3144 		cpywithpad((char *)sc->ctrldata.sn,
3145 		    sizeof(sc->ctrldata.sn), value, ' ');
3146 	}
3147 	value = get_config_value_node(nvl, "eui64");
3148 	if (value != NULL)
3149 		sc->nvstore.eui64 = htobe64(strtoull(value, NULL, 0));
3150 	value = get_config_value_node(nvl, "dsm");
3151 	if (value != NULL) {
3152 		if (strcmp(value, "auto") == 0)
3153 			sc->dataset_management = NVME_DATASET_MANAGEMENT_AUTO;
3154 		else if (strcmp(value, "enable") == 0)
3155 			sc->dataset_management = NVME_DATASET_MANAGEMENT_ENABLE;
3156 		else if (strcmp(value, "disable") == 0)
3157 			sc->dataset_management = NVME_DATASET_MANAGEMENT_DISABLE;
3158 	}
3159 
3160 	value = get_config_value_node(nvl, "bootindex");
3161 	if (value != NULL) {
3162 		if (pci_emul_add_boot_device(sc->nsc_pi, atoi(value))) {
3163 			EPRINTLN("Invalid bootindex %d", atoi(value));
3164 			return (-1);
3165 		}
3166 	}
3167 
3168 	value = get_config_value_node(nvl, "ram");
3169 	if (value != NULL) {
3170 		uint64_t sz = strtoull(value, NULL, 10);
3171 
3172 		sc->nvstore.type = NVME_STOR_RAM;
3173 		sc->nvstore.size = sz * 1024 * 1024;
3174 		sc->nvstore.ctx = calloc(1, sc->nvstore.size);
3175 		sc->nvstore.sectsz = 4096;
3176 		sc->nvstore.sectsz_bits = 12;
3177 		if (sc->nvstore.ctx == NULL) {
3178 			EPRINTLN("nvme: Unable to allocate RAM");
3179 			return (-1);
3180 		}
3181 	} else {
3182 		snprintf(bident, sizeof(bident), "%u:%u",
3183 		    sc->nsc_pi->pi_slot, sc->nsc_pi->pi_func);
3184 		sc->nvstore.ctx = blockif_open(nvl, bident);
3185 		if (sc->nvstore.ctx == NULL) {
3186 			EPRINTLN("nvme: Could not open backing file: %s",
3187 			    strerror(errno));
3188 			return (-1);
3189 		}
3190 		sc->nvstore.type = NVME_STOR_BLOCKIF;
3191 		sc->nvstore.size = blockif_size(sc->nvstore.ctx);
3192 	}
3193 
3194 	if (sectsz == 512 || sectsz == 4096 || sectsz == 8192)
3195 		sc->nvstore.sectsz = sectsz;
3196 	else if (sc->nvstore.type != NVME_STOR_RAM)
3197 		sc->nvstore.sectsz = blockif_sectsz(sc->nvstore.ctx);
3198 	for (sc->nvstore.sectsz_bits = 9;
3199 	     (1U << sc->nvstore.sectsz_bits) < sc->nvstore.sectsz;
3200 	     sc->nvstore.sectsz_bits++);
3201 
3202 	if (sc->max_queues <= 0 || sc->max_queues > NVME_QUEUES)
3203 		sc->max_queues = NVME_QUEUES;
3204 
3205 	return (0);
3206 }
3207 
3208 static void
3209 pci_nvme_resized(struct blockif_ctxt *bctxt __unused, void *arg,
3210     size_t new_size)
3211 {
3212 	struct pci_nvme_softc *sc;
3213 	struct pci_nvme_blockstore *nvstore;
3214 	struct nvme_namespace_data *nd;
3215 
3216 	sc = arg;
3217 	nvstore = &sc->nvstore;
3218 	nd = &sc->nsdata;
3219 
3220 	nvstore->size = new_size;
3221 	pci_nvme_init_nsdata_size(nvstore, nd);
3222 
3223 	/* Add changed NSID to list */
3224 	sc->ns_log.ns[0] = 1;
3225 	sc->ns_log.ns[1] = 0;
3226 
3227 	pci_nvme_aen_post(sc, PCI_NVME_AE_TYPE_NOTICE,
3228 	    PCI_NVME_AEI_NOTICE_NS_ATTR_CHANGED);
3229 }
3230 
3231 static int
3232 pci_nvme_init(struct pci_devinst *pi, nvlist_t *nvl)
3233 {
3234 	struct pci_nvme_softc *sc;
3235 	uint32_t pci_membar_sz;
3236 	int	error;
3237 
3238 	error = 0;
3239 
3240 	sc = calloc(1, sizeof(struct pci_nvme_softc));
3241 	pi->pi_arg = sc;
3242 	sc->nsc_pi = pi;
3243 
3244 	error = pci_nvme_parse_config(sc, nvl);
3245 	if (error < 0)
3246 		goto done;
3247 	else
3248 		error = 0;
3249 
3250 	STAILQ_INIT(&sc->ioreqs_free);
3251 	sc->ioreqs = calloc(sc->ioslots, sizeof(struct pci_nvme_ioreq));
3252 	for (uint32_t i = 0; i < sc->ioslots; i++) {
3253 		STAILQ_INSERT_TAIL(&sc->ioreqs_free, &sc->ioreqs[i], link);
3254 	}
3255 
3256 	pci_set_cfgdata16(pi, PCIR_DEVICE, 0x0A0A);
3257 	pci_set_cfgdata16(pi, PCIR_VENDOR, 0xFB5D);
3258 	pci_set_cfgdata8(pi, PCIR_CLASS, PCIC_STORAGE);
3259 	pci_set_cfgdata8(pi, PCIR_SUBCLASS, PCIS_STORAGE_NVM);
3260 	pci_set_cfgdata8(pi, PCIR_PROGIF,
3261 	                 PCIP_STORAGE_NVM_ENTERPRISE_NVMHCI_1_0);
3262 
3263 	/*
3264 	 * Allocate size of NVMe registers + doorbell space for all queues.
3265 	 *
3266 	 * The specification requires a minimum memory I/O window size of 16K.
3267 	 * The Windows driver will refuse to start a device with a smaller
3268 	 * window.
3269 	 */
3270 	pci_membar_sz = sizeof(struct nvme_registers) +
3271 	    2 * sizeof(uint32_t) * (sc->max_queues + 1);
3272 	pci_membar_sz = MAX(pci_membar_sz, NVME_MMIO_SPACE_MIN);
3273 
3274 	DPRINTF("nvme membar size: %u", pci_membar_sz);
3275 
3276 	error = pci_emul_alloc_bar(pi, 0, PCIBAR_MEM64, pci_membar_sz);
3277 	if (error) {
3278 		WPRINTF("%s pci alloc mem bar failed", __func__);
3279 		goto done;
3280 	}
3281 
3282 	error = pci_emul_add_msixcap(pi, sc->max_queues + 1, NVME_MSIX_BAR);
3283 	if (error) {
3284 		WPRINTF("%s pci add msixcap failed", __func__);
3285 		goto done;
3286 	}
3287 
3288 	error = pci_emul_add_pciecap(pi, PCIEM_TYPE_ROOT_INT_EP);
3289 	if (error) {
3290 		WPRINTF("%s pci add Express capability failed", __func__);
3291 		goto done;
3292 	}
3293 
3294 	pthread_mutex_init(&sc->mtx, NULL);
3295 	sem_init(&sc->iosemlock, 0, sc->ioslots);
3296 	blockif_register_resize_callback(sc->nvstore.ctx, pci_nvme_resized, sc);
3297 
3298 	pci_nvme_init_queues(sc, sc->max_queues, sc->max_queues);
3299 	/*
3300 	 * Controller data depends on Namespace data so initialize Namespace
3301 	 * data first.
3302 	 */
3303 	pci_nvme_init_nsdata(sc, &sc->nsdata, 1, &sc->nvstore);
3304 	pci_nvme_init_ctrldata(sc);
3305 	pci_nvme_init_logpages(sc);
3306 	pci_nvme_init_features(sc);
3307 
3308 	pci_nvme_aer_init(sc);
3309 	pci_nvme_aen_init(sc);
3310 
3311 	pci_nvme_reset(sc);
3312 
3313 done:
3314 	return (error);
3315 }
3316 
3317 static int
3318 pci_nvme_legacy_config(nvlist_t *nvl, const char *opts)
3319 {
3320 	char *cp, *ram;
3321 
3322 	if (opts == NULL)
3323 		return (0);
3324 
3325 	if (strncmp(opts, "ram=", 4) == 0) {
3326 		cp = strchr(opts, ',');
3327 		if (cp == NULL) {
3328 			set_config_value_node(nvl, "ram", opts + 4);
3329 			return (0);
3330 		}
3331 		ram = strndup(opts + 4, cp - opts - 4);
3332 		set_config_value_node(nvl, "ram", ram);
3333 		free(ram);
3334 		return (pci_parse_legacy_config(nvl, cp + 1));
3335 	} else
3336 		return (blockif_legacy_config(nvl, opts));
3337 }
3338 
3339 static const struct pci_devemu pci_de_nvme = {
3340 	.pe_emu =	"nvme",
3341 	.pe_init =	pci_nvme_init,
3342 	.pe_legacy_config = pci_nvme_legacy_config,
3343 	.pe_barwrite =	pci_nvme_write,
3344 	.pe_barread =	pci_nvme_read
3345 };
3346 PCI_EMUL_SET(pci_de_nvme);
3347