xref: /freebsd/usr.sbin/bhyve/pci_nvme.c (revision 731d06abf2105cc0873fa84e972178f9f37ca760)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2017 Shunsuke Mie
5  * Copyright (c) 2018 Leon Dang
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 /*
30  * bhyve PCIe-NVMe device emulation.
31  *
32  * options:
33  *  -s <n>,nvme,devpath,maxq=#,qsz=#,ioslots=#,sectsz=#,ser=A-Z
34  *
35  *  accepted devpath:
36  *    /dev/blockdev
37  *    /path/to/image
38  *    ram=size_in_MiB
39  *
40  *  maxq    = max number of queues
41  *  qsz     = max elements in each queue
42  *  ioslots = max number of concurrent io requests
43  *  sectsz  = sector size (defaults to blockif sector size)
44  *  ser     = serial number (20-chars max)
45  *
46  */
47 
48 /* TODO:
49     - create async event for smart and log
50     - intr coalesce
51  */
52 
53 #include <sys/cdefs.h>
54 __FBSDID("$FreeBSD$");
55 
56 #include <sys/types.h>
57 
58 #include <assert.h>
59 #include <pthread.h>
60 #include <semaphore.h>
61 #include <stdbool.h>
62 #include <stddef.h>
63 #include <stdint.h>
64 #include <stdio.h>
65 #include <stdlib.h>
66 #include <string.h>
67 
68 #include <machine/atomic.h>
69 #include <machine/vmm.h>
70 #include <vmmapi.h>
71 
72 #include <dev/nvme/nvme.h>
73 
74 #include "bhyverun.h"
75 #include "block_if.h"
76 #include "pci_emul.h"
77 
78 
79 static int nvme_debug = 0;
80 #define	DPRINTF(params) if (nvme_debug) printf params
81 #define	WPRINTF(params) printf params
82 
83 /* defaults; can be overridden */
84 #define	NVME_MSIX_BAR		4
85 
86 #define	NVME_IOSLOTS		8
87 
88 /* The NVMe spec defines bits 13:4 in BAR0 as reserved */
89 #define NVME_MMIO_SPACE_MIN	(1 << 14)
90 
91 #define	NVME_QUEUES		16
92 #define	NVME_MAX_QENTRIES	2048
93 
94 #define	NVME_PRP2_ITEMS		(PAGE_SIZE/sizeof(uint64_t))
95 #define	NVME_MAX_BLOCKIOVS	512
96 
97 /* helpers */
98 
99 /* Convert a zero-based value into a one-based value */
100 #define ONE_BASED(zero)		((zero) + 1)
101 /* Convert a one-based value into a zero-based value */
102 #define ZERO_BASED(one)		((one)  - 1)
103 
104 /* Encode number of SQ's and CQ's for Set/Get Features */
105 #define NVME_FEATURE_NUM_QUEUES(sc) \
106 	(ZERO_BASED((sc)->num_squeues) & 0xffff) | \
107 	(ZERO_BASED((sc)->num_cqueues) & 0xffff) << 16;
108 
109 #define	NVME_DOORBELL_OFFSET	offsetof(struct nvme_registers, doorbell)
110 
111 enum nvme_controller_register_offsets {
112 	NVME_CR_CAP_LOW = 0x00,
113 	NVME_CR_CAP_HI  = 0x04,
114 	NVME_CR_VS      = 0x08,
115 	NVME_CR_INTMS   = 0x0c,
116 	NVME_CR_INTMC   = 0x10,
117 	NVME_CR_CC      = 0x14,
118 	NVME_CR_CSTS    = 0x1c,
119 	NVME_CR_NSSR    = 0x20,
120 	NVME_CR_AQA     = 0x24,
121 	NVME_CR_ASQ_LOW = 0x28,
122 	NVME_CR_ASQ_HI  = 0x2c,
123 	NVME_CR_ACQ_LOW = 0x30,
124 	NVME_CR_ACQ_HI  = 0x34,
125 };
126 
127 enum nvme_cmd_cdw11 {
128 	NVME_CMD_CDW11_PC  = 0x0001,
129 	NVME_CMD_CDW11_IEN = 0x0002,
130 	NVME_CMD_CDW11_IV  = 0xFFFF0000,
131 };
132 
133 #define	NVME_CQ_INTEN	0x01
134 #define	NVME_CQ_INTCOAL	0x02
135 
136 struct nvme_completion_queue {
137 	struct nvme_completion *qbase;
138 	uint32_t	size;
139 	uint16_t	tail; /* nvme progress */
140 	uint16_t	head; /* guest progress */
141 	uint16_t	intr_vec;
142 	uint32_t	intr_en;
143 	pthread_mutex_t	mtx;
144 };
145 
146 struct nvme_submission_queue {
147 	struct nvme_command *qbase;
148 	uint32_t	size;
149 	uint16_t	head; /* nvme progress */
150 	uint16_t	tail; /* guest progress */
151 	uint16_t	cqid; /* completion queue id */
152 	int		busy; /* queue is being processed */
153 	int		qpriority;
154 };
155 
156 enum nvme_storage_type {
157 	NVME_STOR_BLOCKIF = 0,
158 	NVME_STOR_RAM = 1,
159 };
160 
161 struct pci_nvme_blockstore {
162 	enum nvme_storage_type type;
163 	void		*ctx;
164 	uint64_t	size;
165 	uint32_t	sectsz;
166 	uint32_t	sectsz_bits;
167 };
168 
169 struct pci_nvme_ioreq {
170 	struct pci_nvme_softc *sc;
171 	struct pci_nvme_ioreq *next;
172 	struct nvme_submission_queue *nvme_sq;
173 	uint16_t	sqid;
174 
175 	/* command information */
176 	uint16_t	opc;
177 	uint16_t	cid;
178 	uint32_t	nsid;
179 
180 	uint64_t	prev_gpaddr;
181 	size_t		prev_size;
182 
183 	/*
184 	 * lock if all iovs consumed (big IO);
185 	 * complete transaction before continuing
186 	 */
187 	pthread_mutex_t	mtx;
188 	pthread_cond_t	cv;
189 
190 	struct blockif_req io_req;
191 
192 	/* pad to fit up to 512 page descriptors from guest IO request */
193 	struct iovec	iovpadding[NVME_MAX_BLOCKIOVS-BLOCKIF_IOV_MAX];
194 };
195 
196 struct pci_nvme_softc {
197 	struct pci_devinst *nsc_pi;
198 
199 	pthread_mutex_t	mtx;
200 
201 	struct nvme_registers regs;
202 
203 	struct nvme_namespace_data  nsdata;
204 	struct nvme_controller_data ctrldata;
205 	struct nvme_error_information_entry err_log;
206 	struct nvme_health_information_page health_log;
207 	struct nvme_firmware_page fw_log;
208 
209 	struct pci_nvme_blockstore nvstore;
210 
211 	uint16_t	max_qentries;	/* max entries per queue */
212 	uint32_t	max_queues;	/* max number of IO SQ's or CQ's */
213 	uint32_t	num_cqueues;
214 	uint32_t	num_squeues;
215 
216 	struct pci_nvme_ioreq *ioreqs;
217 	struct pci_nvme_ioreq *ioreqs_free; /* free list of ioreqs */
218 	uint32_t	pending_ios;
219 	uint32_t	ioslots;
220 	sem_t		iosemlock;
221 
222 	/*
223 	 * Memory mapped Submission and Completion queues
224 	 * Each array includes both Admin and IO queues
225 	 */
226 	struct nvme_completion_queue *compl_queues;
227 	struct nvme_submission_queue *submit_queues;
228 
229 	/* controller features */
230 	uint32_t	intr_coales_aggr_time;   /* 0x08: uS to delay intr */
231 	uint32_t	intr_coales_aggr_thresh; /* 0x08: compl-Q entries */
232 	uint32_t	async_ev_config;         /* 0x0B: async event config */
233 };
234 
235 
236 static void pci_nvme_io_partial(struct blockif_req *br, int err);
237 
238 /* Controller Configuration utils */
239 #define	NVME_CC_GET_EN(cc) \
240 	((cc) >> NVME_CC_REG_EN_SHIFT & NVME_CC_REG_EN_MASK)
241 #define	NVME_CC_GET_CSS(cc) \
242 	((cc) >> NVME_CC_REG_CSS_SHIFT & NVME_CC_REG_CSS_MASK)
243 #define	NVME_CC_GET_SHN(cc) \
244 	((cc) >> NVME_CC_REG_SHN_SHIFT & NVME_CC_REG_SHN_MASK)
245 #define	NVME_CC_GET_IOSQES(cc) \
246 	((cc) >> NVME_CC_REG_IOSQES_SHIFT & NVME_CC_REG_IOSQES_MASK)
247 #define	NVME_CC_GET_IOCQES(cc) \
248 	((cc) >> NVME_CC_REG_IOCQES_SHIFT & NVME_CC_REG_IOCQES_MASK)
249 
250 #define	NVME_CC_WRITE_MASK \
251 	((NVME_CC_REG_EN_MASK << NVME_CC_REG_EN_SHIFT) | \
252 	 (NVME_CC_REG_IOSQES_MASK << NVME_CC_REG_IOSQES_SHIFT) | \
253 	 (NVME_CC_REG_IOCQES_MASK << NVME_CC_REG_IOCQES_SHIFT))
254 
255 #define	NVME_CC_NEN_WRITE_MASK \
256 	((NVME_CC_REG_CSS_MASK << NVME_CC_REG_CSS_SHIFT) | \
257 	 (NVME_CC_REG_MPS_MASK << NVME_CC_REG_MPS_SHIFT) | \
258 	 (NVME_CC_REG_AMS_MASK << NVME_CC_REG_AMS_SHIFT))
259 
260 /* Controller Status utils */
261 #define	NVME_CSTS_GET_RDY(sts) \
262 	((sts) >> NVME_CSTS_REG_RDY_SHIFT & NVME_CSTS_REG_RDY_MASK)
263 
264 #define	NVME_CSTS_RDY	(1 << NVME_CSTS_REG_RDY_SHIFT)
265 
266 /* Completion Queue status word utils */
267 #define	NVME_STATUS_P	(1 << NVME_STATUS_P_SHIFT)
268 #define	NVME_STATUS_MASK \
269 	((NVME_STATUS_SCT_MASK << NVME_STATUS_SCT_SHIFT) |\
270 	 (NVME_STATUS_SC_MASK << NVME_STATUS_SC_SHIFT))
271 
272 static __inline void
273 cpywithpad(char *dst, size_t dst_size, const char *src, char pad)
274 {
275 	size_t len;
276 
277 	len = strnlen(src, dst_size);
278 	memset(dst, pad, dst_size);
279 	memcpy(dst, src, len);
280 }
281 
282 static __inline void
283 pci_nvme_status_tc(uint16_t *status, uint16_t type, uint16_t code)
284 {
285 
286 	*status &= ~NVME_STATUS_MASK;
287 	*status |= (type & NVME_STATUS_SCT_MASK) << NVME_STATUS_SCT_SHIFT |
288 		(code & NVME_STATUS_SC_MASK) << NVME_STATUS_SC_SHIFT;
289 }
290 
291 static __inline void
292 pci_nvme_status_genc(uint16_t *status, uint16_t code)
293 {
294 
295 	pci_nvme_status_tc(status, NVME_SCT_GENERIC, code);
296 }
297 
298 static __inline void
299 pci_nvme_toggle_phase(uint16_t *status, int prev)
300 {
301 
302 	if (prev)
303 		*status &= ~NVME_STATUS_P;
304 	else
305 		*status |= NVME_STATUS_P;
306 }
307 
308 static void
309 pci_nvme_init_ctrldata(struct pci_nvme_softc *sc)
310 {
311 	struct nvme_controller_data *cd = &sc->ctrldata;
312 
313 	cd->vid = 0xFB5D;
314 	cd->ssvid = 0x0000;
315 
316 	cpywithpad((char *)cd->mn, sizeof(cd->mn), "bhyve-NVMe", ' ');
317 	cpywithpad((char *)cd->fr, sizeof(cd->fr), "1.0", ' ');
318 
319 	/* Num of submission commands that we can handle at a time (2^rab) */
320 	cd->rab   = 4;
321 
322 	/* FreeBSD OUI */
323 	cd->ieee[0] = 0x58;
324 	cd->ieee[1] = 0x9c;
325 	cd->ieee[2] = 0xfc;
326 
327 	cd->mic = 0;
328 
329 	cd->mdts = 9;	/* max data transfer size (2^mdts * CAP.MPSMIN) */
330 
331 	cd->ver = 0x00010300;
332 
333 	cd->oacs = 1 << NVME_CTRLR_DATA_OACS_FORMAT_SHIFT;
334 	cd->acl = 2;
335 	cd->aerl = 4;
336 
337 	cd->lpa = 0;	/* TODO: support some simple things like SMART */
338 	cd->elpe = 0;	/* max error log page entries */
339 	cd->npss = 1;	/* number of power states support */
340 
341 	/* Warning Composite Temperature Threshold */
342 	cd->wctemp = 0x0157;
343 
344 	cd->sqes = (6 << NVME_CTRLR_DATA_SQES_MAX_SHIFT) |
345 	    (6 << NVME_CTRLR_DATA_SQES_MIN_SHIFT);
346 	cd->cqes = (4 << NVME_CTRLR_DATA_CQES_MAX_SHIFT) |
347 	    (4 << NVME_CTRLR_DATA_CQES_MIN_SHIFT);
348 	cd->nn = 1;	/* number of namespaces */
349 
350 	cd->fna = 0x03;
351 
352 	cd->power_state[0].mp = 10;
353 }
354 
355 static void
356 pci_nvme_init_nsdata(struct pci_nvme_softc *sc)
357 {
358 	struct nvme_namespace_data *nd;
359 
360 	nd = &sc->nsdata;
361 
362 	nd->nsze = sc->nvstore.size / sc->nvstore.sectsz;
363 	nd->ncap = nd->nsze;
364 	nd->nuse = nd->nsze;
365 
366 	/* Get LBA and backstore information from backing store */
367 	nd->nlbaf = 0; /* NLBAF is a 0's based value (i.e. 1 LBA Format) */
368 	/* LBA data-sz = 2^lbads */
369 	nd->lbaf[0] = sc->nvstore.sectsz_bits << NVME_NS_DATA_LBAF_LBADS_SHIFT;
370 
371 	nd->flbas = 0;
372 }
373 
374 static void
375 pci_nvme_init_logpages(struct pci_nvme_softc *sc)
376 {
377 
378 	memset(&sc->err_log, 0, sizeof(sc->err_log));
379 	memset(&sc->health_log, 0, sizeof(sc->health_log));
380 	memset(&sc->fw_log, 0, sizeof(sc->fw_log));
381 }
382 
383 static void
384 pci_nvme_reset_locked(struct pci_nvme_softc *sc)
385 {
386 	DPRINTF(("%s\r\n", __func__));
387 
388 	sc->regs.cap_lo = (ZERO_BASED(sc->max_qentries) & NVME_CAP_LO_REG_MQES_MASK) |
389 	    (1 << NVME_CAP_LO_REG_CQR_SHIFT) |
390 	    (60 << NVME_CAP_LO_REG_TO_SHIFT);
391 
392 	sc->regs.cap_hi = 1 << NVME_CAP_HI_REG_CSS_NVM_SHIFT;
393 
394 	sc->regs.vs = 0x00010300;	/* NVMe v1.3 */
395 
396 	sc->regs.cc = 0;
397 	sc->regs.csts = 0;
398 
399 	sc->num_cqueues = sc->num_squeues = sc->max_queues;
400 	if (sc->submit_queues != NULL) {
401 		for (int i = 0; i < sc->num_squeues + 1; i++) {
402 			/*
403 			 * The Admin Submission Queue is at index 0.
404 			 * It must not be changed at reset otherwise the
405 			 * emulation will be out of sync with the guest.
406 			 */
407 			if (i != 0) {
408 				sc->submit_queues[i].qbase = NULL;
409 				sc->submit_queues[i].size = 0;
410 				sc->submit_queues[i].cqid = 0;
411 			}
412 			sc->submit_queues[i].tail = 0;
413 			sc->submit_queues[i].head = 0;
414 			sc->submit_queues[i].busy = 0;
415 		}
416 	} else
417 		sc->submit_queues = calloc(sc->num_squeues + 1,
418 		                        sizeof(struct nvme_submission_queue));
419 
420 	if (sc->compl_queues != NULL) {
421 		for (int i = 0; i < sc->num_cqueues + 1; i++) {
422 			/* See Admin Submission Queue note above */
423 			if (i != 0) {
424 				sc->compl_queues[i].qbase = NULL;
425 				sc->compl_queues[i].size = 0;
426 			}
427 
428 			sc->compl_queues[i].tail = 0;
429 			sc->compl_queues[i].head = 0;
430 		}
431 	} else {
432 		sc->compl_queues = calloc(sc->num_cqueues + 1,
433 		                        sizeof(struct nvme_completion_queue));
434 
435 		for (int i = 0; i < sc->num_cqueues + 1; i++)
436 			pthread_mutex_init(&sc->compl_queues[i].mtx, NULL);
437 	}
438 }
439 
440 static void
441 pci_nvme_reset(struct pci_nvme_softc *sc)
442 {
443 	pthread_mutex_lock(&sc->mtx);
444 	pci_nvme_reset_locked(sc);
445 	pthread_mutex_unlock(&sc->mtx);
446 }
447 
448 static void
449 pci_nvme_init_controller(struct vmctx *ctx, struct pci_nvme_softc *sc)
450 {
451 	uint16_t acqs, asqs;
452 
453 	DPRINTF(("%s\r\n", __func__));
454 
455 	asqs = (sc->regs.aqa & NVME_AQA_REG_ASQS_MASK) + 1;
456 	sc->submit_queues[0].size = asqs;
457 	sc->submit_queues[0].qbase = vm_map_gpa(ctx, sc->regs.asq,
458 	            sizeof(struct nvme_command) * asqs);
459 
460 	DPRINTF(("%s mapping Admin-SQ guest 0x%lx, host: %p\r\n",
461 	        __func__, sc->regs.asq, sc->submit_queues[0].qbase));
462 
463 	acqs = ((sc->regs.aqa >> NVME_AQA_REG_ACQS_SHIFT) &
464 	    NVME_AQA_REG_ACQS_MASK) + 1;
465 	sc->compl_queues[0].size = acqs;
466 	sc->compl_queues[0].qbase = vm_map_gpa(ctx, sc->regs.acq,
467 	         sizeof(struct nvme_completion) * acqs);
468 	DPRINTF(("%s mapping Admin-CQ guest 0x%lx, host: %p\r\n",
469 	        __func__, sc->regs.acq, sc->compl_queues[0].qbase));
470 }
471 
472 static int
473 nvme_prp_memcpy(struct vmctx *ctx, uint64_t prp1, uint64_t prp2, uint8_t *src,
474 	size_t len)
475 {
476 	uint8_t *dst;
477 	size_t bytes;
478 
479 	if (len > (8 * 1024)) {
480 		return (-1);
481 	}
482 
483 	/* Copy from the start of prp1 to the end of the physical page */
484 	bytes = PAGE_SIZE - (prp1 & PAGE_MASK);
485 	bytes = MIN(bytes, len);
486 
487 	dst = vm_map_gpa(ctx, prp1, bytes);
488 	if (dst == NULL) {
489 		return (-1);
490 	}
491 
492 	memcpy(dst, src, bytes);
493 
494 	src += bytes;
495 
496 	len -= bytes;
497 	if (len == 0) {
498 		return (0);
499 	}
500 
501 	len = MIN(len, PAGE_SIZE);
502 
503 	dst = vm_map_gpa(ctx, prp2, len);
504 	if (dst == NULL) {
505 		return (-1);
506 	}
507 
508 	memcpy(dst, src, len);
509 
510 	return (0);
511 }
512 
513 static int
514 nvme_opc_delete_io_sq(struct pci_nvme_softc* sc, struct nvme_command* command,
515 	struct nvme_completion* compl)
516 {
517 	uint16_t qid = command->cdw10 & 0xffff;
518 
519 	DPRINTF(("%s DELETE_IO_SQ %u\r\n", __func__, qid));
520 	if (qid == 0 || qid > sc->num_squeues) {
521 		WPRINTF(("%s NOT PERMITTED queue id %u / num_squeues %u\r\n",
522 		        __func__, qid, sc->num_squeues));
523 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
524 		    NVME_SC_INVALID_QUEUE_IDENTIFIER);
525 		return (1);
526 	}
527 
528 	sc->submit_queues[qid].qbase = NULL;
529 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
530 	return (1);
531 }
532 
533 static int
534 nvme_opc_create_io_sq(struct pci_nvme_softc* sc, struct nvme_command* command,
535 	struct nvme_completion* compl)
536 {
537 	if (command->cdw11 & NVME_CMD_CDW11_PC) {
538 		uint16_t qid = command->cdw10 & 0xffff;
539 		struct nvme_submission_queue *nsq;
540 
541 		if ((qid == 0) || (qid > sc->num_squeues)) {
542 			WPRINTF(("%s queue index %u > num_squeues %u\r\n",
543 			        __func__, qid, sc->num_squeues));
544 			pci_nvme_status_tc(&compl->status,
545 			    NVME_SCT_COMMAND_SPECIFIC,
546 			    NVME_SC_INVALID_QUEUE_IDENTIFIER);
547 			return (1);
548 		}
549 
550 		nsq = &sc->submit_queues[qid];
551 		nsq->size = ONE_BASED((command->cdw10 >> 16) & 0xffff);
552 
553 		nsq->qbase = vm_map_gpa(sc->nsc_pi->pi_vmctx, command->prp1,
554 		              sizeof(struct nvme_command) * (size_t)nsq->size);
555 		nsq->cqid = (command->cdw11 >> 16) & 0xffff;
556 		nsq->qpriority = (command->cdw11 >> 1) & 0x03;
557 
558 		DPRINTF(("%s sq %u size %u gaddr %p cqid %u\r\n", __func__,
559 		        qid, nsq->size, nsq->qbase, nsq->cqid));
560 
561 		pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
562 
563 		DPRINTF(("%s completed creating IOSQ qid %u\r\n",
564 		         __func__, qid));
565 	} else {
566 		/*
567 		 * Guest sent non-cont submission queue request.
568 		 * This setting is unsupported by this emulation.
569 		 */
570 		WPRINTF(("%s unsupported non-contig (list-based) "
571 		         "create i/o submission queue\r\n", __func__));
572 
573 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
574 	}
575 	return (1);
576 }
577 
578 static int
579 nvme_opc_delete_io_cq(struct pci_nvme_softc* sc, struct nvme_command* command,
580 	struct nvme_completion* compl)
581 {
582 	uint16_t qid = command->cdw10 & 0xffff;
583 
584 	DPRINTF(("%s DELETE_IO_CQ %u\r\n", __func__, qid));
585 	if (qid == 0 || qid > sc->num_cqueues) {
586 		WPRINTF(("%s queue index %u / num_cqueues %u\r\n",
587 		        __func__, qid, sc->num_cqueues));
588 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
589 		    NVME_SC_INVALID_QUEUE_IDENTIFIER);
590 		return (1);
591 	}
592 
593 	sc->compl_queues[qid].qbase = NULL;
594 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
595 	return (1);
596 }
597 
598 static int
599 nvme_opc_create_io_cq(struct pci_nvme_softc* sc, struct nvme_command* command,
600 	struct nvme_completion* compl)
601 {
602 	if (command->cdw11 & NVME_CMD_CDW11_PC) {
603 		uint16_t qid = command->cdw10 & 0xffff;
604 		struct nvme_completion_queue *ncq;
605 
606 		if ((qid == 0) || (qid > sc->num_cqueues)) {
607 			WPRINTF(("%s queue index %u > num_cqueues %u\r\n",
608 			        __func__, qid, sc->num_cqueues));
609 			pci_nvme_status_tc(&compl->status,
610 			    NVME_SCT_COMMAND_SPECIFIC,
611 			    NVME_SC_INVALID_QUEUE_IDENTIFIER);
612 			return (1);
613 		}
614 
615 		ncq = &sc->compl_queues[qid];
616 		ncq->intr_en = (command->cdw11 & NVME_CMD_CDW11_IEN) >> 1;
617 		ncq->intr_vec = (command->cdw11 >> 16) & 0xffff;
618 		ncq->size = ONE_BASED((command->cdw10 >> 16) & 0xffff);
619 
620 		ncq->qbase = vm_map_gpa(sc->nsc_pi->pi_vmctx,
621 		             command->prp1,
622 		             sizeof(struct nvme_command) * (size_t)ncq->size);
623 
624 		pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
625 	} else {
626 		/*
627 		 * Non-contig completion queue unsupported.
628 		 */
629 		WPRINTF(("%s unsupported non-contig (list-based) "
630 		         "create i/o completion queue\r\n",
631 		         __func__));
632 
633 		/* 0x12 = Invalid Use of Controller Memory Buffer */
634 		pci_nvme_status_genc(&compl->status, 0x12);
635 	}
636 
637 	return (1);
638 }
639 
640 static int
641 nvme_opc_get_log_page(struct pci_nvme_softc* sc, struct nvme_command* command,
642 	struct nvme_completion* compl)
643 {
644 	uint32_t logsize = (1 + ((command->cdw10 >> 16) & 0xFFF)) * 2;
645 	uint8_t logpage = command->cdw10 & 0xFF;
646 
647 	DPRINTF(("%s log page %u len %u\r\n", __func__, logpage, logsize));
648 
649 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
650 
651 	switch (logpage) {
652 	case NVME_LOG_ERROR:
653 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
654 		    command->prp2, (uint8_t *)&sc->err_log, logsize);
655 		break;
656 	case NVME_LOG_HEALTH_INFORMATION:
657 		/* TODO: present some smart info */
658 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
659 		    command->prp2, (uint8_t *)&sc->health_log, logsize);
660 		break;
661 	case NVME_LOG_FIRMWARE_SLOT:
662 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
663 		    command->prp2, (uint8_t *)&sc->fw_log, logsize);
664 		break;
665 	default:
666 		WPRINTF(("%s get log page %x command not supported\r\n",
667 		        __func__, logpage));
668 
669 		pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
670 		    NVME_SC_INVALID_LOG_PAGE);
671 	}
672 
673 	return (1);
674 }
675 
676 static int
677 nvme_opc_identify(struct pci_nvme_softc* sc, struct nvme_command* command,
678 	struct nvme_completion* compl)
679 {
680 	void *dest;
681 
682 	DPRINTF(("%s identify 0x%x nsid 0x%x\r\n", __func__,
683 	        command->cdw10 & 0xFF, command->nsid));
684 
685 	switch (command->cdw10 & 0xFF) {
686 	case 0x00: /* return Identify Namespace data structure */
687 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
688 		    command->prp2, (uint8_t *)&sc->nsdata, sizeof(sc->nsdata));
689 		break;
690 	case 0x01: /* return Identify Controller data structure */
691 		nvme_prp_memcpy(sc->nsc_pi->pi_vmctx, command->prp1,
692 		    command->prp2, (uint8_t *)&sc->ctrldata,
693 		    sizeof(sc->ctrldata));
694 		break;
695 	case 0x02: /* list of 1024 active NSIDs > CDW1.NSID */
696 		dest = vm_map_gpa(sc->nsc_pi->pi_vmctx, command->prp1,
697 		                  sizeof(uint32_t) * 1024);
698 		((uint32_t *)dest)[0] = 1;
699 		((uint32_t *)dest)[1] = 0;
700 		break;
701 	case 0x11:
702 		pci_nvme_status_genc(&compl->status,
703 		    NVME_SC_INVALID_NAMESPACE_OR_FORMAT);
704 		return (1);
705 	case 0x03: /* list of NSID structures in CDW1.NSID, 4096 bytes */
706 	case 0x10:
707 	case 0x12:
708 	case 0x13:
709 	case 0x14:
710 	case 0x15:
711 	default:
712 		DPRINTF(("%s unsupported identify command requested 0x%x\r\n",
713 		         __func__, command->cdw10 & 0xFF));
714 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
715 		return (1);
716 	}
717 
718 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
719 	return (1);
720 }
721 
722 static int
723 nvme_set_feature_queues(struct pci_nvme_softc* sc, struct nvme_command* command,
724 	struct nvme_completion* compl)
725 {
726 	uint16_t nqr;	/* Number of Queues Requested */
727 
728 	nqr = command->cdw11 & 0xFFFF;
729 	if (nqr == 0xffff) {
730 		WPRINTF(("%s: Illegal NSQR value %#x\n", __func__, nqr));
731 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
732 		return (-1);
733 	}
734 
735 	sc->num_squeues = ONE_BASED(nqr);
736 	if (sc->num_squeues > sc->max_queues) {
737 		DPRINTF(("NSQR=%u is greater than max %u\n", sc->num_squeues,
738 					sc->max_queues));
739 		sc->num_squeues = sc->max_queues;
740 	}
741 
742 	nqr = (command->cdw11 >> 16) & 0xFFFF;
743 	if (nqr == 0xffff) {
744 		WPRINTF(("%s: Illegal NCQR value %#x\n", __func__, nqr));
745 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
746 		return (-1);
747 	}
748 
749 	sc->num_cqueues = ONE_BASED(nqr);
750 	if (sc->num_cqueues > sc->max_queues) {
751 		DPRINTF(("NCQR=%u is greater than max %u\n", sc->num_cqueues,
752 					sc->max_queues));
753 		sc->num_cqueues = sc->max_queues;
754 	}
755 
756 	compl->cdw0 = NVME_FEATURE_NUM_QUEUES(sc);
757 
758 	return (0);
759 }
760 
761 static int
762 nvme_opc_set_features(struct pci_nvme_softc* sc, struct nvme_command* command,
763 	struct nvme_completion* compl)
764 {
765 	int feature = command->cdw10 & 0xFF;
766 	uint32_t iv;
767 
768 	DPRINTF(("%s feature 0x%x\r\n", __func__, feature));
769 	compl->cdw0 = 0;
770 
771 	switch (feature) {
772 	case NVME_FEAT_ARBITRATION:
773 		DPRINTF(("  arbitration 0x%x\r\n", command->cdw11));
774 		break;
775 	case NVME_FEAT_POWER_MANAGEMENT:
776 		DPRINTF(("  power management 0x%x\r\n", command->cdw11));
777 		break;
778 	case NVME_FEAT_LBA_RANGE_TYPE:
779 		DPRINTF(("  lba range 0x%x\r\n", command->cdw11));
780 		break;
781 	case NVME_FEAT_TEMPERATURE_THRESHOLD:
782 		DPRINTF(("  temperature threshold 0x%x\r\n", command->cdw11));
783 		break;
784 	case NVME_FEAT_ERROR_RECOVERY:
785 		DPRINTF(("  error recovery 0x%x\r\n", command->cdw11));
786 		break;
787 	case NVME_FEAT_VOLATILE_WRITE_CACHE:
788 		DPRINTF(("  volatile write cache 0x%x\r\n", command->cdw11));
789 		break;
790 	case NVME_FEAT_NUMBER_OF_QUEUES:
791 		nvme_set_feature_queues(sc, command, compl);
792 		break;
793 	case NVME_FEAT_INTERRUPT_COALESCING:
794 		DPRINTF(("  interrupt coalescing 0x%x\r\n", command->cdw11));
795 
796 		/* in uS */
797 		sc->intr_coales_aggr_time = ((command->cdw11 >> 8) & 0xFF)*100;
798 
799 		sc->intr_coales_aggr_thresh = command->cdw11 & 0xFF;
800 		break;
801 	case NVME_FEAT_INTERRUPT_VECTOR_CONFIGURATION:
802 		iv = command->cdw11 & 0xFFFF;
803 
804 		DPRINTF(("  interrupt vector configuration 0x%x\r\n",
805 		        command->cdw11));
806 
807 		for (uint32_t i = 0; i < sc->num_cqueues + 1; i++) {
808 			if (sc->compl_queues[i].intr_vec == iv) {
809 				if (command->cdw11 & (1 << 16))
810 					sc->compl_queues[i].intr_en |=
811 					                      NVME_CQ_INTCOAL;
812 				else
813 					sc->compl_queues[i].intr_en &=
814 					                     ~NVME_CQ_INTCOAL;
815 			}
816 		}
817 		break;
818 	case NVME_FEAT_WRITE_ATOMICITY:
819 		DPRINTF(("  write atomicity 0x%x\r\n", command->cdw11));
820 		break;
821 	case NVME_FEAT_ASYNC_EVENT_CONFIGURATION:
822 		DPRINTF(("  async event configuration 0x%x\r\n",
823 		        command->cdw11));
824 		sc->async_ev_config = command->cdw11;
825 		break;
826 	case NVME_FEAT_SOFTWARE_PROGRESS_MARKER:
827 		DPRINTF(("  software progress marker 0x%x\r\n",
828 		        command->cdw11));
829 		break;
830 	case 0x0C:
831 		DPRINTF(("  autonomous power state transition 0x%x\r\n",
832 		        command->cdw11));
833 		break;
834 	default:
835 		WPRINTF(("%s invalid feature\r\n", __func__));
836 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
837 		return (1);
838 	}
839 
840 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
841 	return (1);
842 }
843 
844 static int
845 nvme_opc_get_features(struct pci_nvme_softc* sc, struct nvme_command* command,
846 	struct nvme_completion* compl)
847 {
848 	int feature = command->cdw10 & 0xFF;
849 
850 	DPRINTF(("%s feature 0x%x\r\n", __func__, feature));
851 
852 	compl->cdw0 = 0;
853 
854 	switch (feature) {
855 	case NVME_FEAT_ARBITRATION:
856 		DPRINTF(("  arbitration\r\n"));
857 		break;
858 	case NVME_FEAT_POWER_MANAGEMENT:
859 		DPRINTF(("  power management\r\n"));
860 		break;
861 	case NVME_FEAT_LBA_RANGE_TYPE:
862 		DPRINTF(("  lba range\r\n"));
863 		break;
864 	case NVME_FEAT_TEMPERATURE_THRESHOLD:
865 		DPRINTF(("  temperature threshold\r\n"));
866 		switch ((command->cdw11 >> 20) & 0x3) {
867 		case 0:
868 			/* Over temp threshold */
869 			compl->cdw0 = 0xFFFF;
870 			break;
871 		case 1:
872 			/* Under temp threshold */
873 			compl->cdw0 = 0;
874 			break;
875 		default:
876 			WPRINTF(("  invalid threshold type select\r\n"));
877 			pci_nvme_status_genc(&compl->status,
878 			    NVME_SC_INVALID_FIELD);
879 			return (1);
880 		}
881 		break;
882 	case NVME_FEAT_ERROR_RECOVERY:
883 		DPRINTF(("  error recovery\r\n"));
884 		break;
885 	case NVME_FEAT_VOLATILE_WRITE_CACHE:
886 		DPRINTF(("  volatile write cache\r\n"));
887 		break;
888 	case NVME_FEAT_NUMBER_OF_QUEUES:
889 		compl->cdw0 = NVME_FEATURE_NUM_QUEUES(sc);
890 
891 		DPRINTF(("  number of queues (submit %u, completion %u)\r\n",
892 		        compl->cdw0 & 0xFFFF,
893 		        (compl->cdw0 >> 16) & 0xFFFF));
894 
895 		break;
896 	case NVME_FEAT_INTERRUPT_COALESCING:
897 		DPRINTF(("  interrupt coalescing\r\n"));
898 		break;
899 	case NVME_FEAT_INTERRUPT_VECTOR_CONFIGURATION:
900 		DPRINTF(("  interrupt vector configuration\r\n"));
901 		break;
902 	case NVME_FEAT_WRITE_ATOMICITY:
903 		DPRINTF(("  write atomicity\r\n"));
904 		break;
905 	case NVME_FEAT_ASYNC_EVENT_CONFIGURATION:
906 		DPRINTF(("  async event configuration\r\n"));
907 		sc->async_ev_config = command->cdw11;
908 		break;
909 	case NVME_FEAT_SOFTWARE_PROGRESS_MARKER:
910 		DPRINTF(("  software progress marker\r\n"));
911 		break;
912 	case 0x0C:
913 		DPRINTF(("  autonomous power state transition\r\n"));
914 		break;
915 	default:
916 		WPRINTF(("%s invalid feature 0x%x\r\n", __func__, feature));
917 		pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
918 		return (1);
919 	}
920 
921 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
922 	return (1);
923 }
924 
925 static int
926 nvme_opc_abort(struct pci_nvme_softc* sc, struct nvme_command* command,
927 	struct nvme_completion* compl)
928 {
929 	DPRINTF(("%s submission queue %u, command ID 0x%x\r\n", __func__,
930 	        command->cdw10 & 0xFFFF, (command->cdw10 >> 16) & 0xFFFF));
931 
932 	/* TODO: search for the command ID and abort it */
933 
934 	compl->cdw0 = 1;
935 	pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
936 	return (1);
937 }
938 
939 static int
940 nvme_opc_async_event_req(struct pci_nvme_softc* sc,
941 	struct nvme_command* command, struct nvme_completion* compl)
942 {
943 	DPRINTF(("%s async event request 0x%x\r\n", __func__, command->cdw11));
944 
945 	/*
946 	 * TODO: raise events when they happen based on the Set Features cmd.
947 	 * These events happen async, so only set completion successful if
948 	 * there is an event reflective of the request to get event.
949 	 */
950 	pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
951 	    NVME_SC_ASYNC_EVENT_REQUEST_LIMIT_EXCEEDED);
952 	return (0);
953 }
954 
955 static void
956 pci_nvme_handle_admin_cmd(struct pci_nvme_softc* sc, uint64_t value)
957 {
958 	struct nvme_completion compl;
959 	struct nvme_command *cmd;
960 	struct nvme_submission_queue *sq;
961 	struct nvme_completion_queue *cq;
962 	int do_intr = 0;
963 	uint16_t sqhead;
964 
965 	DPRINTF(("%s index %u\r\n", __func__, (uint32_t)value));
966 
967 	sq = &sc->submit_queues[0];
968 
969 	sqhead = atomic_load_acq_short(&sq->head);
970 
971 	if (atomic_testandset_int(&sq->busy, 1)) {
972 		DPRINTF(("%s SQ busy, head %u, tail %u\r\n",
973 		        __func__, sqhead, sq->tail));
974 		return;
975 	}
976 
977 	DPRINTF(("sqhead %u, tail %u\r\n", sqhead, sq->tail));
978 
979 	while (sqhead != atomic_load_acq_short(&sq->tail)) {
980 		cmd = &(sq->qbase)[sqhead];
981 		compl.status = 0;
982 
983 		switch (cmd->opc) {
984 		case NVME_OPC_DELETE_IO_SQ:
985 			DPRINTF(("%s command DELETE_IO_SQ\r\n", __func__));
986 			do_intr |= nvme_opc_delete_io_sq(sc, cmd, &compl);
987 			break;
988 		case NVME_OPC_CREATE_IO_SQ:
989 			DPRINTF(("%s command CREATE_IO_SQ\r\n", __func__));
990 			do_intr |= nvme_opc_create_io_sq(sc, cmd, &compl);
991 			break;
992 		case NVME_OPC_DELETE_IO_CQ:
993 			DPRINTF(("%s command DELETE_IO_CQ\r\n", __func__));
994 			do_intr |= nvme_opc_delete_io_cq(sc, cmd, &compl);
995 			break;
996 		case NVME_OPC_CREATE_IO_CQ:
997 			DPRINTF(("%s command CREATE_IO_CQ\r\n", __func__));
998 			do_intr |= nvme_opc_create_io_cq(sc, cmd, &compl);
999 			break;
1000 		case NVME_OPC_GET_LOG_PAGE:
1001 			DPRINTF(("%s command GET_LOG_PAGE\r\n", __func__));
1002 			do_intr |= nvme_opc_get_log_page(sc, cmd, &compl);
1003 			break;
1004 		case NVME_OPC_IDENTIFY:
1005 			DPRINTF(("%s command IDENTIFY\r\n", __func__));
1006 			do_intr |= nvme_opc_identify(sc, cmd, &compl);
1007 			break;
1008 		case NVME_OPC_ABORT:
1009 			DPRINTF(("%s command ABORT\r\n", __func__));
1010 			do_intr |= nvme_opc_abort(sc, cmd, &compl);
1011 			break;
1012 		case NVME_OPC_SET_FEATURES:
1013 			DPRINTF(("%s command SET_FEATURES\r\n", __func__));
1014 			do_intr |= nvme_opc_set_features(sc, cmd, &compl);
1015 			break;
1016 		case NVME_OPC_GET_FEATURES:
1017 			DPRINTF(("%s command GET_FEATURES\r\n", __func__));
1018 			do_intr |= nvme_opc_get_features(sc, cmd, &compl);
1019 			break;
1020 		case NVME_OPC_ASYNC_EVENT_REQUEST:
1021 			DPRINTF(("%s command ASYNC_EVENT_REQ\r\n", __func__));
1022 			/* XXX dont care, unhandled for now
1023 			do_intr |= nvme_opc_async_event_req(sc, cmd, &compl);
1024 			*/
1025 			break;
1026 		default:
1027 			WPRINTF(("0x%x command is not implemented\r\n",
1028 			    cmd->opc));
1029 		}
1030 
1031 		/* for now skip async event generation */
1032 		if (cmd->opc != NVME_OPC_ASYNC_EVENT_REQUEST) {
1033 			struct nvme_completion *cp;
1034 			int phase;
1035 
1036 			cq = &sc->compl_queues[0];
1037 
1038 			cp = &(cq->qbase)[cq->tail];
1039 			cp->cdw0 = compl.cdw0;
1040 			cp->sqid = 0;
1041 			cp->sqhd = sqhead;
1042 			cp->cid = cmd->cid;
1043 
1044 			phase = NVME_STATUS_GET_P(cp->status);
1045 			cp->status = compl.status;
1046 			pci_nvme_toggle_phase(&cp->status, phase);
1047 
1048 			cq->tail = (cq->tail + 1) % cq->size;
1049 		}
1050 		sqhead = (sqhead + 1) % sq->size;
1051 	}
1052 
1053 	DPRINTF(("setting sqhead %u\r\n", sqhead));
1054 	atomic_store_short(&sq->head, sqhead);
1055 	atomic_store_int(&sq->busy, 0);
1056 
1057 	if (do_intr)
1058 		pci_generate_msix(sc->nsc_pi, 0);
1059 
1060 }
1061 
1062 static int
1063 pci_nvme_append_iov_req(struct pci_nvme_softc *sc, struct pci_nvme_ioreq *req,
1064 	uint64_t gpaddr, size_t size, int do_write, uint64_t lba)
1065 {
1066 	int iovidx;
1067 
1068 	if (req != NULL) {
1069 		/* concatenate contig block-iovs to minimize number of iovs */
1070 		if ((req->prev_gpaddr + req->prev_size) == gpaddr) {
1071 			iovidx = req->io_req.br_iovcnt - 1;
1072 
1073 			req->io_req.br_iov[iovidx].iov_base =
1074 			    paddr_guest2host(req->sc->nsc_pi->pi_vmctx,
1075 			                     req->prev_gpaddr, size);
1076 
1077 			req->prev_size += size;
1078 			req->io_req.br_resid += size;
1079 
1080 			req->io_req.br_iov[iovidx].iov_len = req->prev_size;
1081 		} else {
1082 			pthread_mutex_lock(&req->mtx);
1083 
1084 			iovidx = req->io_req.br_iovcnt;
1085 			if (iovidx == NVME_MAX_BLOCKIOVS) {
1086 				int err = 0;
1087 
1088 				DPRINTF(("large I/O, doing partial req\r\n"));
1089 
1090 				iovidx = 0;
1091 				req->io_req.br_iovcnt = 0;
1092 
1093 				req->io_req.br_callback = pci_nvme_io_partial;
1094 
1095 				if (!do_write)
1096 					err = blockif_read(sc->nvstore.ctx,
1097 					                   &req->io_req);
1098 				else
1099 					err = blockif_write(sc->nvstore.ctx,
1100 					                    &req->io_req);
1101 
1102 				/* wait until req completes before cont */
1103 				if (err == 0)
1104 					pthread_cond_wait(&req->cv, &req->mtx);
1105 			}
1106 			if (iovidx == 0) {
1107 				req->io_req.br_offset = lba;
1108 				req->io_req.br_resid = 0;
1109 				req->io_req.br_param = req;
1110 			}
1111 
1112 			req->io_req.br_iov[iovidx].iov_base =
1113 			    paddr_guest2host(req->sc->nsc_pi->pi_vmctx,
1114 			                     gpaddr, size);
1115 
1116 			req->io_req.br_iov[iovidx].iov_len = size;
1117 
1118 			req->prev_gpaddr = gpaddr;
1119 			req->prev_size = size;
1120 			req->io_req.br_resid += size;
1121 
1122 			req->io_req.br_iovcnt++;
1123 
1124 			pthread_mutex_unlock(&req->mtx);
1125 		}
1126 	} else {
1127 		/* RAM buffer: read/write directly */
1128 		void *p = sc->nvstore.ctx;
1129 		void *gptr;
1130 
1131 		if ((lba + size) > sc->nvstore.size) {
1132 			WPRINTF(("%s write would overflow RAM\r\n", __func__));
1133 			return (-1);
1134 		}
1135 
1136 		p = (void *)((uintptr_t)p + (uintptr_t)lba);
1137 		gptr = paddr_guest2host(sc->nsc_pi->pi_vmctx, gpaddr, size);
1138 		if (do_write)
1139 			memcpy(p, gptr, size);
1140 		else
1141 			memcpy(gptr, p, size);
1142 	}
1143 	return (0);
1144 }
1145 
1146 static void
1147 pci_nvme_set_completion(struct pci_nvme_softc *sc,
1148 	struct nvme_submission_queue *sq, int sqid, uint16_t cid,
1149 	uint32_t cdw0, uint16_t status, int ignore_busy)
1150 {
1151 	struct nvme_completion_queue *cq = &sc->compl_queues[sq->cqid];
1152 	struct nvme_completion *compl;
1153 	int do_intr = 0;
1154 	int phase;
1155 
1156 	DPRINTF(("%s sqid %d cqid %u cid %u status: 0x%x 0x%x\r\n",
1157 		 __func__, sqid, sq->cqid, cid, NVME_STATUS_GET_SCT(status),
1158 		 NVME_STATUS_GET_SC(status)));
1159 
1160 	pthread_mutex_lock(&cq->mtx);
1161 
1162 	assert(cq->qbase != NULL);
1163 
1164 	compl = &cq->qbase[cq->tail];
1165 
1166 	compl->sqhd = atomic_load_acq_short(&sq->head);
1167 	compl->sqid = sqid;
1168 	compl->cid = cid;
1169 
1170 	// toggle phase
1171 	phase = NVME_STATUS_GET_P(compl->status);
1172 	compl->status = status;
1173 	pci_nvme_toggle_phase(&compl->status, phase);
1174 
1175 	cq->tail = (cq->tail + 1) % cq->size;
1176 
1177 	if (cq->intr_en & NVME_CQ_INTEN)
1178 		do_intr = 1;
1179 
1180 	pthread_mutex_unlock(&cq->mtx);
1181 
1182 	if (ignore_busy || !atomic_load_acq_int(&sq->busy))
1183 		if (do_intr)
1184 			pci_generate_msix(sc->nsc_pi, cq->intr_vec);
1185 }
1186 
1187 static void
1188 pci_nvme_release_ioreq(struct pci_nvme_softc *sc, struct pci_nvme_ioreq *req)
1189 {
1190 	req->sc = NULL;
1191 	req->nvme_sq = NULL;
1192 	req->sqid = 0;
1193 
1194 	pthread_mutex_lock(&sc->mtx);
1195 
1196 	req->next = sc->ioreqs_free;
1197 	sc->ioreqs_free = req;
1198 	sc->pending_ios--;
1199 
1200 	/* when no more IO pending, can set to ready if device reset/enabled */
1201 	if (sc->pending_ios == 0 &&
1202 	    NVME_CC_GET_EN(sc->regs.cc) && !(NVME_CSTS_GET_RDY(sc->regs.csts)))
1203 		sc->regs.csts |= NVME_CSTS_RDY;
1204 
1205 	pthread_mutex_unlock(&sc->mtx);
1206 
1207 	sem_post(&sc->iosemlock);
1208 }
1209 
1210 static struct pci_nvme_ioreq *
1211 pci_nvme_get_ioreq(struct pci_nvme_softc *sc)
1212 {
1213 	struct pci_nvme_ioreq *req = NULL;;
1214 
1215 	sem_wait(&sc->iosemlock);
1216 	pthread_mutex_lock(&sc->mtx);
1217 
1218 	req = sc->ioreqs_free;
1219 	assert(req != NULL);
1220 
1221 	sc->ioreqs_free = req->next;
1222 
1223 	req->next = NULL;
1224 	req->sc = sc;
1225 
1226 	sc->pending_ios++;
1227 
1228 	pthread_mutex_unlock(&sc->mtx);
1229 
1230 	req->io_req.br_iovcnt = 0;
1231 	req->io_req.br_offset = 0;
1232 	req->io_req.br_resid = 0;
1233 	req->io_req.br_param = req;
1234 	req->prev_gpaddr = 0;
1235 	req->prev_size = 0;
1236 
1237 	return req;
1238 }
1239 
1240 static void
1241 pci_nvme_io_done(struct blockif_req *br, int err)
1242 {
1243 	struct pci_nvme_ioreq *req = br->br_param;
1244 	struct nvme_submission_queue *sq = req->nvme_sq;
1245 	uint16_t code, status;
1246 
1247 	DPRINTF(("%s error %d %s\r\n", __func__, err, strerror(err)));
1248 
1249 	/* TODO return correct error */
1250 	code = err ? NVME_SC_DATA_TRANSFER_ERROR : NVME_SC_SUCCESS;
1251 	pci_nvme_status_genc(&status, code);
1252 
1253 	pci_nvme_set_completion(req->sc, sq, req->sqid, req->cid, 0, status, 0);
1254 	pci_nvme_release_ioreq(req->sc, req);
1255 }
1256 
1257 static void
1258 pci_nvme_io_partial(struct blockif_req *br, int err)
1259 {
1260 	struct pci_nvme_ioreq *req = br->br_param;
1261 
1262 	DPRINTF(("%s error %d %s\r\n", __func__, err, strerror(err)));
1263 
1264 	pthread_cond_signal(&req->cv);
1265 }
1266 
1267 
1268 static void
1269 pci_nvme_handle_io_cmd(struct pci_nvme_softc* sc, uint16_t idx)
1270 {
1271 	struct nvme_submission_queue *sq;
1272 	uint16_t status;
1273 	uint16_t sqhead;
1274 	int err;
1275 
1276 	/* handle all submissions up to sq->tail index */
1277 	sq = &sc->submit_queues[idx];
1278 
1279 	if (atomic_testandset_int(&sq->busy, 1)) {
1280 		DPRINTF(("%s sqid %u busy\r\n", __func__, idx));
1281 		return;
1282 	}
1283 
1284 	sqhead = atomic_load_acq_short(&sq->head);
1285 
1286 	DPRINTF(("nvme_handle_io qid %u head %u tail %u cmdlist %p\r\n",
1287 	         idx, sqhead, sq->tail, sq->qbase));
1288 
1289 	while (sqhead != atomic_load_acq_short(&sq->tail)) {
1290 		struct nvme_command *cmd;
1291 		struct pci_nvme_ioreq *req = NULL;
1292 		uint64_t lba;
1293 		uint64_t nblocks, bytes, size, cpsz;
1294 
1295 		/* TODO: support scatter gather list handling */
1296 
1297 		cmd = &sq->qbase[sqhead];
1298 		sqhead = (sqhead + 1) % sq->size;
1299 
1300 		lba = ((uint64_t)cmd->cdw11 << 32) | cmd->cdw10;
1301 
1302 		if (cmd->opc == NVME_OPC_FLUSH) {
1303 			pci_nvme_status_genc(&status, NVME_SC_SUCCESS);
1304 			pci_nvme_set_completion(sc, sq, idx, cmd->cid, 0,
1305 			                        status, 1);
1306 
1307 			continue;
1308 		} else if (cmd->opc == 0x08) {
1309 			/* TODO: write zeroes */
1310 			WPRINTF(("%s write zeroes lba 0x%lx blocks %u\r\n",
1311 			        __func__, lba, cmd->cdw12 & 0xFFFF));
1312 			pci_nvme_status_genc(&status, NVME_SC_SUCCESS);
1313 			pci_nvme_set_completion(sc, sq, idx, cmd->cid, 0,
1314 			                        status, 1);
1315 
1316 			continue;
1317 		}
1318 
1319 		nblocks = (cmd->cdw12 & 0xFFFF) + 1;
1320 
1321 		bytes = nblocks * sc->nvstore.sectsz;
1322 
1323 		if (sc->nvstore.type == NVME_STOR_BLOCKIF) {
1324 			req = pci_nvme_get_ioreq(sc);
1325 			req->nvme_sq = sq;
1326 			req->sqid = idx;
1327 		}
1328 
1329 		/*
1330 		 * If data starts mid-page and flows into the next page, then
1331 		 * increase page count
1332 		 */
1333 
1334 		DPRINTF(("[h%u:t%u:n%u] %s starting LBA 0x%lx blocks %lu "
1335 		         "(%lu-bytes)\r\n",
1336 		         sqhead==0 ? sq->size-1 : sqhead-1, sq->tail, sq->size,
1337 		         cmd->opc == NVME_OPC_WRITE ?
1338 			     "WRITE" : "READ",
1339 		         lba, nblocks, bytes));
1340 
1341 		cmd->prp1 &= ~(0x03UL);
1342 		cmd->prp2 &= ~(0x03UL);
1343 
1344 		DPRINTF((" prp1 0x%lx prp2 0x%lx\r\n", cmd->prp1, cmd->prp2));
1345 
1346 		size = bytes;
1347 		lba *= sc->nvstore.sectsz;
1348 
1349 		cpsz = PAGE_SIZE - (cmd->prp1 % PAGE_SIZE);
1350 
1351 		if (cpsz > bytes)
1352 			cpsz = bytes;
1353 
1354 		if (req != NULL) {
1355 			req->io_req.br_offset = ((uint64_t)cmd->cdw11 << 32) |
1356 			                        cmd->cdw10;
1357 			req->opc = cmd->opc;
1358 			req->cid = cmd->cid;
1359 			req->nsid = cmd->nsid;
1360 		}
1361 
1362 		err = pci_nvme_append_iov_req(sc, req, cmd->prp1, cpsz,
1363 		    cmd->opc == NVME_OPC_WRITE, lba);
1364 		lba += cpsz;
1365 		size -= cpsz;
1366 
1367 		if (size == 0)
1368 			goto iodone;
1369 
1370 		if (size <= PAGE_SIZE) {
1371 			/* prp2 is second (and final) page in transfer */
1372 
1373 			err = pci_nvme_append_iov_req(sc, req, cmd->prp2,
1374 			    size,
1375 			    cmd->opc == NVME_OPC_WRITE,
1376 			    lba);
1377 		} else {
1378 			uint64_t *prp_list;
1379 			int i;
1380 
1381 			/* prp2 is pointer to a physical region page list */
1382 			prp_list = paddr_guest2host(sc->nsc_pi->pi_vmctx,
1383 			                            cmd->prp2, PAGE_SIZE);
1384 
1385 			i = 0;
1386 			while (size != 0) {
1387 				cpsz = MIN(size, PAGE_SIZE);
1388 
1389 				/*
1390 				 * Move to linked physical region page list
1391 				 * in last item.
1392 				 */
1393 				if (i == (NVME_PRP2_ITEMS-1) &&
1394 				    size > PAGE_SIZE) {
1395 					assert((prp_list[i] & (PAGE_SIZE-1)) == 0);
1396 					prp_list = paddr_guest2host(
1397 					              sc->nsc_pi->pi_vmctx,
1398 					              prp_list[i], PAGE_SIZE);
1399 					i = 0;
1400 				}
1401 				if (prp_list[i] == 0) {
1402 					WPRINTF(("PRP2[%d] = 0 !!!\r\n", i));
1403 					err = 1;
1404 					break;
1405 				}
1406 
1407 				err = pci_nvme_append_iov_req(sc, req,
1408 				    prp_list[i], cpsz,
1409 				    cmd->opc == NVME_OPC_WRITE, lba);
1410 				if (err)
1411 					break;
1412 
1413 				lba += cpsz;
1414 				size -= cpsz;
1415 				i++;
1416 			}
1417 		}
1418 
1419 iodone:
1420 		if (sc->nvstore.type == NVME_STOR_RAM) {
1421 			uint16_t code, status;
1422 
1423 			code = err ? NVME_SC_LBA_OUT_OF_RANGE :
1424 			    NVME_SC_SUCCESS;
1425 			pci_nvme_status_genc(&status, code);
1426 
1427 			pci_nvme_set_completion(sc, sq, idx, cmd->cid, 0,
1428 			                        status, 1);
1429 
1430 			continue;
1431 		}
1432 
1433 
1434 		if (err)
1435 			goto do_error;
1436 
1437 		req->io_req.br_callback = pci_nvme_io_done;
1438 
1439 		err = 0;
1440 		switch (cmd->opc) {
1441 		case NVME_OPC_READ:
1442 			err = blockif_read(sc->nvstore.ctx, &req->io_req);
1443 			break;
1444 		case NVME_OPC_WRITE:
1445 			err = blockif_write(sc->nvstore.ctx, &req->io_req);
1446 			break;
1447 		default:
1448 			WPRINTF(("%s unhandled io command 0x%x\r\n",
1449 				 __func__, cmd->opc));
1450 			err = 1;
1451 		}
1452 
1453 do_error:
1454 		if (err) {
1455 			uint16_t status;
1456 
1457 			pci_nvme_status_genc(&status,
1458 			    NVME_SC_DATA_TRANSFER_ERROR);
1459 
1460 			pci_nvme_set_completion(sc, sq, idx, cmd->cid, 0,
1461 			                        status, 1);
1462 			pci_nvme_release_ioreq(sc, req);
1463 		}
1464 	}
1465 
1466 	atomic_store_short(&sq->head, sqhead);
1467 	atomic_store_int(&sq->busy, 0);
1468 }
1469 
1470 static void
1471 pci_nvme_handle_doorbell(struct vmctx *ctx, struct pci_nvme_softc* sc,
1472 	uint64_t idx, int is_sq, uint64_t value)
1473 {
1474 	DPRINTF(("nvme doorbell %lu, %s, val 0x%lx\r\n",
1475 	        idx, is_sq ? "SQ" : "CQ", value & 0xFFFF));
1476 
1477 	if (is_sq) {
1478 		atomic_store_short(&sc->submit_queues[idx].tail,
1479 		                   (uint16_t)value);
1480 
1481 		if (idx == 0) {
1482 			pci_nvme_handle_admin_cmd(sc, value);
1483 		} else {
1484 			/* submission queue; handle new entries in SQ */
1485 			if (idx > sc->num_squeues) {
1486 				WPRINTF(("%s SQ index %lu overflow from "
1487 				         "guest (max %u)\r\n",
1488 				         __func__, idx, sc->num_squeues));
1489 				return;
1490 			}
1491 			pci_nvme_handle_io_cmd(sc, (uint16_t)idx);
1492 		}
1493 	} else {
1494 		if (idx > sc->num_cqueues) {
1495 			WPRINTF(("%s queue index %lu overflow from "
1496 			         "guest (max %u)\r\n",
1497 			         __func__, idx, sc->num_cqueues));
1498 			return;
1499 		}
1500 
1501 		sc->compl_queues[idx].head = (uint16_t)value;
1502 	}
1503 }
1504 
1505 static void
1506 pci_nvme_bar0_reg_dumps(const char *func, uint64_t offset, int iswrite)
1507 {
1508 	const char *s = iswrite ? "WRITE" : "READ";
1509 
1510 	switch (offset) {
1511 	case NVME_CR_CAP_LOW:
1512 		DPRINTF(("%s %s NVME_CR_CAP_LOW\r\n", func, s));
1513 		break;
1514 	case NVME_CR_CAP_HI:
1515 		DPRINTF(("%s %s NVME_CR_CAP_HI\r\n", func, s));
1516 		break;
1517 	case NVME_CR_VS:
1518 		DPRINTF(("%s %s NVME_CR_VS\r\n", func, s));
1519 		break;
1520 	case NVME_CR_INTMS:
1521 		DPRINTF(("%s %s NVME_CR_INTMS\r\n", func, s));
1522 		break;
1523 	case NVME_CR_INTMC:
1524 		DPRINTF(("%s %s NVME_CR_INTMC\r\n", func, s));
1525 		break;
1526 	case NVME_CR_CC:
1527 		DPRINTF(("%s %s NVME_CR_CC\r\n", func, s));
1528 		break;
1529 	case NVME_CR_CSTS:
1530 		DPRINTF(("%s %s NVME_CR_CSTS\r\n", func, s));
1531 		break;
1532 	case NVME_CR_NSSR:
1533 		DPRINTF(("%s %s NVME_CR_NSSR\r\n", func, s));
1534 		break;
1535 	case NVME_CR_AQA:
1536 		DPRINTF(("%s %s NVME_CR_AQA\r\n", func, s));
1537 		break;
1538 	case NVME_CR_ASQ_LOW:
1539 		DPRINTF(("%s %s NVME_CR_ASQ_LOW\r\n", func, s));
1540 		break;
1541 	case NVME_CR_ASQ_HI:
1542 		DPRINTF(("%s %s NVME_CR_ASQ_HI\r\n", func, s));
1543 		break;
1544 	case NVME_CR_ACQ_LOW:
1545 		DPRINTF(("%s %s NVME_CR_ACQ_LOW\r\n", func, s));
1546 		break;
1547 	case NVME_CR_ACQ_HI:
1548 		DPRINTF(("%s %s NVME_CR_ACQ_HI\r\n", func, s));
1549 		break;
1550 	default:
1551 		DPRINTF(("unknown nvme bar-0 offset 0x%lx\r\n", offset));
1552 	}
1553 
1554 }
1555 
1556 static void
1557 pci_nvme_write_bar_0(struct vmctx *ctx, struct pci_nvme_softc* sc,
1558 	uint64_t offset, int size, uint64_t value)
1559 {
1560 	uint32_t ccreg;
1561 
1562 	if (offset >= NVME_DOORBELL_OFFSET) {
1563 		uint64_t belloffset = offset - NVME_DOORBELL_OFFSET;
1564 		uint64_t idx = belloffset / 8; /* door bell size = 2*int */
1565 		int is_sq = (belloffset % 8) < 4;
1566 
1567 		if (belloffset > ((sc->max_queues+1) * 8 - 4)) {
1568 			WPRINTF(("guest attempted an overflow write offset "
1569 			         "0x%lx, val 0x%lx in %s",
1570 			         offset, value, __func__));
1571 			return;
1572 		}
1573 
1574 		pci_nvme_handle_doorbell(ctx, sc, idx, is_sq, value);
1575 		return;
1576 	}
1577 
1578 	DPRINTF(("nvme-write offset 0x%lx, size %d, value 0x%lx\r\n",
1579 	        offset, size, value));
1580 
1581 	if (size != 4) {
1582 		WPRINTF(("guest wrote invalid size %d (offset 0x%lx, "
1583 		         "val 0x%lx) to bar0 in %s",
1584 		         size, offset, value, __func__));
1585 		/* TODO: shutdown device */
1586 		return;
1587 	}
1588 
1589 	pci_nvme_bar0_reg_dumps(__func__, offset, 1);
1590 
1591 	pthread_mutex_lock(&sc->mtx);
1592 
1593 	switch (offset) {
1594 	case NVME_CR_CAP_LOW:
1595 	case NVME_CR_CAP_HI:
1596 		/* readonly */
1597 		break;
1598 	case NVME_CR_VS:
1599 		/* readonly */
1600 		break;
1601 	case NVME_CR_INTMS:
1602 		/* MSI-X, so ignore */
1603 		break;
1604 	case NVME_CR_INTMC:
1605 		/* MSI-X, so ignore */
1606 		break;
1607 	case NVME_CR_CC:
1608 		ccreg = (uint32_t)value;
1609 
1610 		DPRINTF(("%s NVME_CR_CC en %x css %x shn %x iosqes %u "
1611 		         "iocqes %u\r\n",
1612 		        __func__,
1613 			 NVME_CC_GET_EN(ccreg), NVME_CC_GET_CSS(ccreg),
1614 			 NVME_CC_GET_SHN(ccreg), NVME_CC_GET_IOSQES(ccreg),
1615 			 NVME_CC_GET_IOCQES(ccreg)));
1616 
1617 		if (NVME_CC_GET_SHN(ccreg)) {
1618 			/* perform shutdown - flush out data to backend */
1619 			sc->regs.csts &= ~(NVME_CSTS_REG_SHST_MASK <<
1620 			    NVME_CSTS_REG_SHST_SHIFT);
1621 			sc->regs.csts |= NVME_SHST_COMPLETE <<
1622 			    NVME_CSTS_REG_SHST_SHIFT;
1623 		}
1624 		if (NVME_CC_GET_EN(ccreg) != NVME_CC_GET_EN(sc->regs.cc)) {
1625 			if (NVME_CC_GET_EN(ccreg) == 0)
1626 				/* transition 1-> causes controller reset */
1627 				pci_nvme_reset_locked(sc);
1628 			else
1629 				pci_nvme_init_controller(ctx, sc);
1630 		}
1631 
1632 		/* Insert the iocqes, iosqes and en bits from the write */
1633 		sc->regs.cc &= ~NVME_CC_WRITE_MASK;
1634 		sc->regs.cc |= ccreg & NVME_CC_WRITE_MASK;
1635 		if (NVME_CC_GET_EN(ccreg) == 0) {
1636 			/* Insert the ams, mps and css bit fields */
1637 			sc->regs.cc &= ~NVME_CC_NEN_WRITE_MASK;
1638 			sc->regs.cc |= ccreg & NVME_CC_NEN_WRITE_MASK;
1639 			sc->regs.csts &= ~NVME_CSTS_RDY;
1640 		} else if (sc->pending_ios == 0) {
1641 			sc->regs.csts |= NVME_CSTS_RDY;
1642 		}
1643 		break;
1644 	case NVME_CR_CSTS:
1645 		break;
1646 	case NVME_CR_NSSR:
1647 		/* ignore writes; don't support subsystem reset */
1648 		break;
1649 	case NVME_CR_AQA:
1650 		sc->regs.aqa = (uint32_t)value;
1651 		break;
1652 	case NVME_CR_ASQ_LOW:
1653 		sc->regs.asq = (sc->regs.asq & (0xFFFFFFFF00000000)) |
1654 		               (0xFFFFF000 & value);
1655 		break;
1656 	case NVME_CR_ASQ_HI:
1657 		sc->regs.asq = (sc->regs.asq & (0x00000000FFFFFFFF)) |
1658 		               (value << 32);
1659 		break;
1660 	case NVME_CR_ACQ_LOW:
1661 		sc->regs.acq = (sc->regs.acq & (0xFFFFFFFF00000000)) |
1662 		               (0xFFFFF000 & value);
1663 		break;
1664 	case NVME_CR_ACQ_HI:
1665 		sc->regs.acq = (sc->regs.acq & (0x00000000FFFFFFFF)) |
1666 		               (value << 32);
1667 		break;
1668 	default:
1669 		DPRINTF(("%s unknown offset 0x%lx, value 0x%lx size %d\r\n",
1670 		         __func__, offset, value, size));
1671 	}
1672 	pthread_mutex_unlock(&sc->mtx);
1673 }
1674 
1675 static void
1676 pci_nvme_write(struct vmctx *ctx, int vcpu, struct pci_devinst *pi,
1677                 int baridx, uint64_t offset, int size, uint64_t value)
1678 {
1679 	struct pci_nvme_softc* sc = pi->pi_arg;
1680 
1681 	if (baridx == pci_msix_table_bar(pi) ||
1682 	    baridx == pci_msix_pba_bar(pi)) {
1683 		DPRINTF(("nvme-write baridx %d, msix: off 0x%lx, size %d, "
1684 		         " value 0x%lx\r\n", baridx, offset, size, value));
1685 
1686 		pci_emul_msix_twrite(pi, offset, size, value);
1687 		return;
1688 	}
1689 
1690 	switch (baridx) {
1691 	case 0:
1692 		pci_nvme_write_bar_0(ctx, sc, offset, size, value);
1693 		break;
1694 
1695 	default:
1696 		DPRINTF(("%s unknown baridx %d, val 0x%lx\r\n",
1697 		         __func__, baridx, value));
1698 	}
1699 }
1700 
1701 static uint64_t pci_nvme_read_bar_0(struct pci_nvme_softc* sc,
1702 	uint64_t offset, int size)
1703 {
1704 	uint64_t value;
1705 
1706 	pci_nvme_bar0_reg_dumps(__func__, offset, 0);
1707 
1708 	if (offset < NVME_DOORBELL_OFFSET) {
1709 		void *p = &(sc->regs);
1710 		pthread_mutex_lock(&sc->mtx);
1711 		memcpy(&value, (void *)((uintptr_t)p + offset), size);
1712 		pthread_mutex_unlock(&sc->mtx);
1713 	} else {
1714 		value = 0;
1715                 WPRINTF(("pci_nvme: read invalid offset %ld\r\n", offset));
1716 	}
1717 
1718 	switch (size) {
1719 	case 1:
1720 		value &= 0xFF;
1721 		break;
1722 	case 2:
1723 		value &= 0xFFFF;
1724 		break;
1725 	case 4:
1726 		value &= 0xFFFFFFFF;
1727 		break;
1728 	}
1729 
1730 	DPRINTF(("   nvme-read offset 0x%lx, size %d -> value 0x%x\r\n",
1731 	         offset, size, (uint32_t)value));
1732 
1733 	return (value);
1734 }
1735 
1736 
1737 
1738 static uint64_t
1739 pci_nvme_read(struct vmctx *ctx, int vcpu, struct pci_devinst *pi, int baridx,
1740     uint64_t offset, int size)
1741 {
1742 	struct pci_nvme_softc* sc = pi->pi_arg;
1743 
1744 	if (baridx == pci_msix_table_bar(pi) ||
1745 	    baridx == pci_msix_pba_bar(pi)) {
1746 		DPRINTF(("nvme-read bar: %d, msix: regoff 0x%lx, size %d\r\n",
1747 		        baridx, offset, size));
1748 
1749 		return pci_emul_msix_tread(pi, offset, size);
1750 	}
1751 
1752 	switch (baridx) {
1753 	case 0:
1754        		return pci_nvme_read_bar_0(sc, offset, size);
1755 
1756 	default:
1757 		DPRINTF(("unknown bar %d, 0x%lx\r\n", baridx, offset));
1758 	}
1759 
1760 	return (0);
1761 }
1762 
1763 
1764 static int
1765 pci_nvme_parse_opts(struct pci_nvme_softc *sc, char *opts)
1766 {
1767 	char bident[sizeof("XX:X:X")];
1768 	char	*uopt, *xopts, *config;
1769 	uint32_t sectsz;
1770 	int optidx;
1771 
1772 	sc->max_queues = NVME_QUEUES;
1773 	sc->max_qentries = NVME_MAX_QENTRIES;
1774 	sc->ioslots = NVME_IOSLOTS;
1775 	sc->num_squeues = sc->max_queues;
1776 	sc->num_cqueues = sc->max_queues;
1777 	sectsz = 0;
1778 
1779 	uopt = strdup(opts);
1780 	optidx = 0;
1781 	snprintf(sc->ctrldata.sn, sizeof(sc->ctrldata.sn),
1782 	         "NVME-%d-%d", sc->nsc_pi->pi_slot, sc->nsc_pi->pi_func);
1783 	for (xopts = strtok(uopt, ",");
1784 	     xopts != NULL;
1785 	     xopts = strtok(NULL, ",")) {
1786 
1787 		if ((config = strchr(xopts, '=')) != NULL)
1788 			*config++ = '\0';
1789 
1790 		if (!strcmp("maxq", xopts)) {
1791 			sc->max_queues = atoi(config);
1792 		} else if (!strcmp("qsz", xopts)) {
1793 			sc->max_qentries = atoi(config);
1794 		} else if (!strcmp("ioslots", xopts)) {
1795 			sc->ioslots = atoi(config);
1796 		} else if (!strcmp("sectsz", xopts)) {
1797 			sectsz = atoi(config);
1798 		} else if (!strcmp("ser", xopts)) {
1799 			/*
1800 			 * This field indicates the Product Serial Number in
1801 			 * 7-bit ASCII, unused bytes should be space characters.
1802 			 * Ref: NVMe v1.3c.
1803 			 */
1804 			cpywithpad((char *)sc->ctrldata.sn,
1805 			           sizeof(sc->ctrldata.sn), config, ' ');
1806 		} else if (!strcmp("ram", xopts)) {
1807 			uint64_t sz = strtoull(&xopts[4], NULL, 10);
1808 
1809 			sc->nvstore.type = NVME_STOR_RAM;
1810 			sc->nvstore.size = sz * 1024 * 1024;
1811 			sc->nvstore.ctx = calloc(1, sc->nvstore.size);
1812 			sc->nvstore.sectsz = 4096;
1813 			sc->nvstore.sectsz_bits = 12;
1814 			if (sc->nvstore.ctx == NULL) {
1815 				perror("Unable to allocate RAM");
1816 				free(uopt);
1817 				return (-1);
1818 			}
1819 		} else if (optidx == 0) {
1820 			snprintf(bident, sizeof(bident), "%d:%d",
1821 			         sc->nsc_pi->pi_slot, sc->nsc_pi->pi_func);
1822 			sc->nvstore.ctx = blockif_open(xopts, bident);
1823 			if (sc->nvstore.ctx == NULL) {
1824 				perror("Could not open backing file");
1825 				free(uopt);
1826 				return (-1);
1827 			}
1828 			sc->nvstore.type = NVME_STOR_BLOCKIF;
1829 			sc->nvstore.size = blockif_size(sc->nvstore.ctx);
1830 		} else {
1831 			fprintf(stderr, "Invalid option %s\n", xopts);
1832 			free(uopt);
1833 			return (-1);
1834 		}
1835 
1836 		optidx++;
1837 	}
1838 	free(uopt);
1839 
1840 	if (sc->nvstore.ctx == NULL || sc->nvstore.size == 0) {
1841 		fprintf(stderr, "backing store not specified\n");
1842 		return (-1);
1843 	}
1844 	if (sectsz == 512 || sectsz == 4096 || sectsz == 8192)
1845 		sc->nvstore.sectsz = sectsz;
1846 	else if (sc->nvstore.type != NVME_STOR_RAM)
1847 		sc->nvstore.sectsz = blockif_sectsz(sc->nvstore.ctx);
1848 	for (sc->nvstore.sectsz_bits = 9;
1849 	     (1 << sc->nvstore.sectsz_bits) < sc->nvstore.sectsz;
1850 	     sc->nvstore.sectsz_bits++);
1851 
1852 	if (sc->max_queues <= 0 || sc->max_queues > NVME_QUEUES)
1853 		sc->max_queues = NVME_QUEUES;
1854 
1855 	if (sc->max_qentries <= 0) {
1856 		fprintf(stderr, "Invalid qsz option\n");
1857 		return (-1);
1858 	}
1859 	if (sc->ioslots <= 0) {
1860 		fprintf(stderr, "Invalid ioslots option\n");
1861 		return (-1);
1862 	}
1863 
1864 	return (0);
1865 }
1866 
1867 static int
1868 pci_nvme_init(struct vmctx *ctx, struct pci_devinst *pi, char *opts)
1869 {
1870 	struct pci_nvme_softc *sc;
1871 	uint32_t pci_membar_sz;
1872 	int	error;
1873 
1874 	error = 0;
1875 
1876 	sc = calloc(1, sizeof(struct pci_nvme_softc));
1877 	pi->pi_arg = sc;
1878 	sc->nsc_pi = pi;
1879 
1880 	error = pci_nvme_parse_opts(sc, opts);
1881 	if (error < 0)
1882 		goto done;
1883 	else
1884 		error = 0;
1885 
1886 	sc->ioreqs = calloc(sc->ioslots, sizeof(struct pci_nvme_ioreq));
1887 	for (int i = 0; i < sc->ioslots; i++) {
1888 		if (i < (sc->ioslots-1))
1889 			sc->ioreqs[i].next = &sc->ioreqs[i+1];
1890 		pthread_mutex_init(&sc->ioreqs[i].mtx, NULL);
1891 		pthread_cond_init(&sc->ioreqs[i].cv, NULL);
1892 	}
1893 	sc->ioreqs_free = sc->ioreqs;
1894 	sc->intr_coales_aggr_thresh = 1;
1895 
1896 	pci_set_cfgdata16(pi, PCIR_DEVICE, 0x0A0A);
1897 	pci_set_cfgdata16(pi, PCIR_VENDOR, 0xFB5D);
1898 	pci_set_cfgdata8(pi, PCIR_CLASS, PCIC_STORAGE);
1899 	pci_set_cfgdata8(pi, PCIR_SUBCLASS, PCIS_STORAGE_NVM);
1900 	pci_set_cfgdata8(pi, PCIR_PROGIF,
1901 	                 PCIP_STORAGE_NVM_ENTERPRISE_NVMHCI_1_0);
1902 
1903 	/*
1904 	 * Allocate size of NVMe registers + doorbell space for all queues.
1905 	 *
1906 	 * The specification requires a minimum memory I/O window size of 16K.
1907 	 * The Windows driver will refuse to start a device with a smaller
1908 	 * window.
1909 	 */
1910 	pci_membar_sz = sizeof(struct nvme_registers) +
1911 	    2 * sizeof(uint32_t) * (sc->max_queues + 1);
1912 	pci_membar_sz = MAX(pci_membar_sz, NVME_MMIO_SPACE_MIN);
1913 
1914 	DPRINTF(("nvme membar size: %u\r\n", pci_membar_sz));
1915 
1916 	error = pci_emul_alloc_bar(pi, 0, PCIBAR_MEM64, pci_membar_sz);
1917 	if (error) {
1918 		WPRINTF(("%s pci alloc mem bar failed\r\n", __func__));
1919 		goto done;
1920 	}
1921 
1922 	error = pci_emul_add_msixcap(pi, sc->max_queues + 1, NVME_MSIX_BAR);
1923 	if (error) {
1924 		WPRINTF(("%s pci add msixcap failed\r\n", __func__));
1925 		goto done;
1926 	}
1927 
1928 	pthread_mutex_init(&sc->mtx, NULL);
1929 	sem_init(&sc->iosemlock, 0, sc->ioslots);
1930 
1931 	pci_nvme_reset(sc);
1932 	pci_nvme_init_ctrldata(sc);
1933 	pci_nvme_init_nsdata(sc);
1934 	pci_nvme_init_logpages(sc);
1935 
1936 	pci_lintr_request(pi);
1937 
1938 done:
1939 	return (error);
1940 }
1941 
1942 
1943 struct pci_devemu pci_de_nvme = {
1944 	.pe_emu =	"nvme",
1945 	.pe_init =	pci_nvme_init,
1946 	.pe_barwrite =	pci_nvme_write,
1947 	.pe_barread =	pci_nvme_read
1948 };
1949 PCI_EMUL_SET(pci_de_nvme);
1950