xref: /illumos-gate/usr/src/uts/common/io/nvme/nvme.c (revision 7931524763ef94dc16989451dddd206563d03bb4)
1 /*
2  * This file and its contents are supplied under the terms of the
3  * Common Development and Distribution License ("CDDL"), version 1.0.
4  * You may only use this file in accordance with the terms of version
5  * 1.0 of the CDDL.
6  *
7  * A full copy of the text of the CDDL should have accompanied this
8  * source.  A copy of the CDDL is also available via the Internet at
9  * http://www.illumos.org/license/CDDL.
10  */
11 
12 /*
13  * Copyright 2018 Nexenta Systems, Inc.
14  * Copyright 2016 Tegile Systems, Inc. All rights reserved.
15  * Copyright (c) 2016 The MathWorks, Inc.  All rights reserved.
16  * Copyright 2018 Joyent, Inc.
17  */
18 
19 /*
20  * blkdev driver for NVMe compliant storage devices
21  *
22  * This driver was written to conform to version 1.2.1 of the NVMe
23  * specification.  It may work with newer versions, but that is completely
24  * untested and disabled by default.
25  *
26  * The driver has only been tested on x86 systems and will not work on big-
27  * endian systems without changes to the code accessing registers and data
28  * structures used by the hardware.
29  *
30  *
31  * Interrupt Usage:
32  *
33  * The driver will use a single interrupt while configuring the device as the
34  * specification requires, but contrary to the specification it will try to use
35  * a single-message MSI(-X) or FIXED interrupt. Later in the attach process it
36  * will switch to multiple-message MSI(-X) if supported. The driver wants to
37  * have one interrupt vector per CPU, but it will work correctly if less are
38  * available. Interrupts can be shared by queues, the interrupt handler will
39  * iterate through the I/O queue array by steps of n_intr_cnt. Usually only
40  * the admin queue will share an interrupt with one I/O queue. The interrupt
41  * handler will retrieve completed commands from all queues sharing an interrupt
42  * vector and will post them to a taskq for completion processing.
43  *
44  *
45  * Command Processing:
46  *
47  * NVMe devices can have up to 65535 I/O queue pairs, with each queue holding up
48  * to 65536 I/O commands. The driver will configure one I/O queue pair per
49  * available interrupt vector, with the queue length usually much smaller than
50  * the maximum of 65536. If the hardware doesn't provide enough queues, fewer
51  * interrupt vectors will be used.
52  *
53  * Additionally the hardware provides a single special admin queue pair that can
54  * hold up to 4096 admin commands.
55  *
56  * From the hardware perspective both queues of a queue pair are independent,
57  * but they share some driver state: the command array (holding pointers to
58  * commands currently being processed by the hardware) and the active command
59  * counter. Access to a queue pair and the shared state is protected by
60  * nq_mutex.
61  *
62  * When a command is submitted to a queue pair the active command counter is
63  * incremented and a pointer to the command is stored in the command array. The
64  * array index is used as command identifier (CID) in the submission queue
65  * entry. Some commands may take a very long time to complete, and if the queue
66  * wraps around in that time a submission may find the next array slot to still
67  * be used by a long-running command. In this case the array is sequentially
68  * searched for the next free slot. The length of the command array is the same
69  * as the configured queue length. Queue overrun is prevented by the semaphore,
70  * so a command submission may block if the queue is full.
71  *
72  *
73  * Polled I/O Support:
74  *
75  * For kernel core dump support the driver can do polled I/O. As interrupts are
76  * turned off while dumping the driver will just submit a command in the regular
77  * way, and then repeatedly attempt a command retrieval until it gets the
78  * command back.
79  *
80  *
81  * Namespace Support:
82  *
83  * NVMe devices can have multiple namespaces, each being a independent data
84  * store. The driver supports multiple namespaces and creates a blkdev interface
85  * for each namespace found. Namespaces can have various attributes to support
86  * protection information. This driver does not support any of this and ignores
87  * namespaces that have these attributes.
88  *
89  * As of NVMe 1.1 namespaces can have an 64bit Extended Unique Identifier
90  * (EUI64). This driver uses the EUI64 if present to generate the devid and
91  * passes it to blkdev to use it in the device node names. As this is currently
92  * untested namespaces with EUI64 are ignored by default.
93  *
94  * We currently support only (2 << NVME_MINOR_INST_SHIFT) - 2 namespaces in a
95  * single controller. This is an artificial limit imposed by the driver to be
96  * able to address a reasonable number of controllers and namespaces using a
97  * 32bit minor node number.
98  *
99  *
100  * Minor nodes:
101  *
102  * For each NVMe device the driver exposes one minor node for the controller and
103  * one minor node for each namespace. The only operations supported by those
104  * minor nodes are open(9E), close(9E), and ioctl(9E). This serves as the
105  * interface for the nvmeadm(1M) utility.
106  *
107  *
108  * Blkdev Interface:
109  *
110  * This driver uses blkdev to do all the heavy lifting involved with presenting
111  * a disk device to the system. As a result, the processing of I/O requests is
112  * relatively simple as blkdev takes care of partitioning, boundary checks, DMA
113  * setup, and splitting of transfers into manageable chunks.
114  *
115  * I/O requests coming in from blkdev are turned into NVM commands and posted to
116  * an I/O queue. The queue is selected by taking the CPU id modulo the number of
117  * queues. There is currently no timeout handling of I/O commands.
118  *
119  * Blkdev also supports querying device/media information and generating a
120  * devid. The driver reports the best block size as determined by the namespace
121  * format back to blkdev as physical block size to support partition and block
122  * alignment. The devid is either based on the namespace EUI64, if present, or
123  * composed using the device vendor ID, model number, serial number, and the
124  * namespace ID.
125  *
126  *
127  * Error Handling:
128  *
129  * Error handling is currently limited to detecting fatal hardware errors,
130  * either by asynchronous events, or synchronously through command status or
131  * admin command timeouts. In case of severe errors the device is fenced off,
132  * all further requests will return EIO. FMA is then called to fault the device.
133  *
134  * The hardware has a limit for outstanding asynchronous event requests. Before
135  * this limit is known the driver assumes it is at least 1 and posts a single
136  * asynchronous request. Later when the limit is known more asynchronous event
137  * requests are posted to allow quicker reception of error information. When an
138  * asynchronous event is posted by the hardware the driver will parse the error
139  * status fields and log information or fault the device, depending on the
140  * severity of the asynchronous event. The asynchronous event request is then
141  * reused and posted to the admin queue again.
142  *
143  * On command completion the command status is checked for errors. In case of
144  * errors indicating a driver bug the driver panics. Almost all other error
145  * status values just cause EIO to be returned.
146  *
147  * Command timeouts are currently detected for all admin commands except
148  * asynchronous event requests. If a command times out and the hardware appears
149  * to be healthy the driver attempts to abort the command. The original command
150  * timeout is also applied to the abort command. If the abort times out too the
151  * driver assumes the device to be dead, fences it off, and calls FMA to retire
152  * it. In all other cases the aborted command should return immediately with a
153  * status indicating it was aborted, and the driver will wait indefinitely for
154  * that to happen. No timeout handling of normal I/O commands is presently done.
155  *
156  * Any command that times out due to the controller dropping dead will be put on
157  * nvme_lost_cmds list if it references DMA memory. This will prevent the DMA
158  * memory being reused by the system and later be written to by a "dead" NVMe
159  * controller.
160  *
161  *
162  * Locking:
163  *
164  * Each queue pair has its own nq_mutex, which must be held when accessing the
165  * associated queue registers or the shared state of the queue pair. Callers of
166  * nvme_unqueue_cmd() must make sure that nq_mutex is held, while
167  * nvme_submit_{admin,io}_cmd() and nvme_retrieve_cmd() take care of this
168  * themselves.
169  *
170  * Each command also has its own nc_mutex, which is associated with the
171  * condition variable nc_cv. It is only used on admin commands which are run
172  * synchronously. In that case it must be held across calls to
173  * nvme_submit_{admin,io}_cmd() and nvme_wait_cmd(), which is taken care of by
174  * nvme_admin_cmd(). It must also be held whenever the completion state of the
175  * command is changed or while a admin command timeout is handled.
176  *
177  * If both nc_mutex and nq_mutex must be held, nc_mutex must be acquired first.
178  * More than one nc_mutex may only be held when aborting commands. In this case,
179  * the nc_mutex of the command to be aborted must be held across the call to
180  * nvme_abort_cmd() to prevent the command from completing while the abort is in
181  * progress.
182  *
183  * Each minor node has its own nm_mutex, which protects the open count nm_ocnt
184  * and exclusive-open flag nm_oexcl.
185  *
186  *
187  * Quiesce / Fast Reboot:
188  *
189  * The driver currently does not support fast reboot. A quiesce(9E) entry point
190  * is still provided which is used to send a shutdown notification to the
191  * device.
192  *
193  *
194  * Driver Configuration:
195  *
196  * The following driver properties can be changed to control some aspects of the
197  * drivers operation:
198  * - strict-version: can be set to 0 to allow devices conforming to newer
199  *   major versions to be used
200  * - ignore-unknown-vendor-status: can be set to 1 to not handle any vendor
201  *   specific command status as a fatal error leading device faulting
202  * - admin-queue-len: the maximum length of the admin queue (16-4096)
203  * - io-queue-len: the maximum length of the I/O queues (16-65536)
204  * - async-event-limit: the maximum number of asynchronous event requests to be
205  *   posted by the driver
206  * - volatile-write-cache-enable: can be set to 0 to disable the volatile write
207  *   cache
208  * - min-phys-block-size: the minimum physical block size to report to blkdev,
209  *   which is among other things the basis for ZFS vdev ashift
210  *
211  *
212  * TODO:
213  * - figure out sane default for I/O queue depth reported to blkdev
214  * - FMA handling of media errors
215  * - support for devices supporting very large I/O requests using chained PRPs
216  * - support for configuring hardware parameters like interrupt coalescing
217  * - support for media formatting and hard partitioning into namespaces
218  * - support for big-endian systems
219  * - support for fast reboot
220  * - support for firmware updates
221  * - support for NVMe Subsystem Reset (1.1)
222  * - support for Scatter/Gather lists (1.1)
223  * - support for Reservations (1.1)
224  * - support for power management
225  */
226 
227 #include <sys/byteorder.h>
228 #ifdef _BIG_ENDIAN
229 #error nvme driver needs porting for big-endian platforms
230 #endif
231 
232 #include <sys/modctl.h>
233 #include <sys/conf.h>
234 #include <sys/devops.h>
235 #include <sys/ddi.h>
236 #include <sys/sunddi.h>
237 #include <sys/sunndi.h>
238 #include <sys/bitmap.h>
239 #include <sys/sysmacros.h>
240 #include <sys/param.h>
241 #include <sys/varargs.h>
242 #include <sys/cpuvar.h>
243 #include <sys/disp.h>
244 #include <sys/blkdev.h>
245 #include <sys/atomic.h>
246 #include <sys/archsystm.h>
247 #include <sys/sata/sata_hba.h>
248 #include <sys/stat.h>
249 #include <sys/policy.h>
250 #include <sys/list.h>
251 
252 #include <sys/nvme.h>
253 
254 #ifdef __x86
255 #include <sys/x86_archext.h>
256 #endif
257 
258 #include "nvme_reg.h"
259 #include "nvme_var.h"
260 
261 /*
262  * Assertions to make sure that we've properly captured various aspects of the
263  * packed structures and haven't broken them during updates.
264  */
265 CTASSERT(sizeof (nvme_identify_ctrl_t) == 0x1000);
266 CTASSERT(offsetof(nvme_identify_ctrl_t, id_oacs) == 256);
267 CTASSERT(offsetof(nvme_identify_ctrl_t, id_sqes) == 512);
268 CTASSERT(offsetof(nvme_identify_ctrl_t, id_subnqn) == 768);
269 CTASSERT(offsetof(nvme_identify_ctrl_t, id_nvmof) == 1792);
270 CTASSERT(offsetof(nvme_identify_ctrl_t, id_psd) == 2048);
271 CTASSERT(offsetof(nvme_identify_ctrl_t, id_vs) == 3072);
272 
273 CTASSERT(sizeof (nvme_identify_nsid_t) == 0x1000);
274 CTASSERT(offsetof(nvme_identify_nsid_t, id_fpi) == 32);
275 CTASSERT(offsetof(nvme_identify_nsid_t, id_nguid) == 104);
276 CTASSERT(offsetof(nvme_identify_nsid_t, id_lbaf) == 128);
277 CTASSERT(offsetof(nvme_identify_nsid_t, id_vs) == 384);
278 
279 CTASSERT(sizeof (nvme_identify_primary_caps_t) == 0x1000);
280 CTASSERT(offsetof(nvme_identify_primary_caps_t, nipc_vqfrt) == 32);
281 CTASSERT(offsetof(nvme_identify_primary_caps_t, nipc_vifrt) == 64);
282 
283 
284 /* NVMe spec version supported */
285 static const int nvme_version_major = 1;
286 
287 /* tunable for admin command timeout in seconds, default is 1s */
288 int nvme_admin_cmd_timeout = 1;
289 
290 /* tunable for FORMAT NVM command timeout in seconds, default is 600s */
291 int nvme_format_cmd_timeout = 600;
292 
293 static int nvme_attach(dev_info_t *, ddi_attach_cmd_t);
294 static int nvme_detach(dev_info_t *, ddi_detach_cmd_t);
295 static int nvme_quiesce(dev_info_t *);
296 static int nvme_fm_errcb(dev_info_t *, ddi_fm_error_t *, const void *);
297 static int nvme_setup_interrupts(nvme_t *, int, int);
298 static void nvme_release_interrupts(nvme_t *);
299 static uint_t nvme_intr(caddr_t, caddr_t);
300 
301 static void nvme_shutdown(nvme_t *, int, boolean_t);
302 static boolean_t nvme_reset(nvme_t *, boolean_t);
303 static int nvme_init(nvme_t *);
304 static nvme_cmd_t *nvme_alloc_cmd(nvme_t *, int);
305 static void nvme_free_cmd(nvme_cmd_t *);
306 static nvme_cmd_t *nvme_create_nvm_cmd(nvme_namespace_t *, uint8_t,
307     bd_xfer_t *);
308 static void nvme_admin_cmd(nvme_cmd_t *, int);
309 static void nvme_submit_admin_cmd(nvme_qpair_t *, nvme_cmd_t *);
310 static int nvme_submit_io_cmd(nvme_qpair_t *, nvme_cmd_t *);
311 static void nvme_submit_cmd_common(nvme_qpair_t *, nvme_cmd_t *);
312 static nvme_cmd_t *nvme_unqueue_cmd(nvme_t *, nvme_qpair_t *, int);
313 static nvme_cmd_t *nvme_retrieve_cmd(nvme_t *, nvme_qpair_t *);
314 static void nvme_wait_cmd(nvme_cmd_t *, uint_t);
315 static void nvme_wakeup_cmd(void *);
316 static void nvme_async_event_task(void *);
317 
318 static int nvme_check_unknown_cmd_status(nvme_cmd_t *);
319 static int nvme_check_vendor_cmd_status(nvme_cmd_t *);
320 static int nvme_check_integrity_cmd_status(nvme_cmd_t *);
321 static int nvme_check_specific_cmd_status(nvme_cmd_t *);
322 static int nvme_check_generic_cmd_status(nvme_cmd_t *);
323 static inline int nvme_check_cmd_status(nvme_cmd_t *);
324 
325 static int nvme_abort_cmd(nvme_cmd_t *, uint_t);
326 static void nvme_async_event(nvme_t *);
327 static int nvme_format_nvm(nvme_t *, uint32_t, uint8_t, boolean_t, uint8_t,
328     boolean_t, uint8_t);
329 static int nvme_get_logpage(nvme_t *, void **, size_t *, uint8_t, ...);
330 static int nvme_identify(nvme_t *, uint32_t, void **);
331 static int nvme_set_features(nvme_t *, uint32_t, uint8_t, uint32_t,
332     uint32_t *);
333 static int nvme_get_features(nvme_t *, uint32_t, uint8_t, uint32_t *,
334     void **, size_t *);
335 static int nvme_write_cache_set(nvme_t *, boolean_t);
336 static int nvme_set_nqueues(nvme_t *, uint16_t *);
337 
338 static void nvme_free_dma(nvme_dma_t *);
339 static int nvme_zalloc_dma(nvme_t *, size_t, uint_t, ddi_dma_attr_t *,
340     nvme_dma_t **);
341 static int nvme_zalloc_queue_dma(nvme_t *, uint32_t, uint16_t, uint_t,
342     nvme_dma_t **);
343 static void nvme_free_qpair(nvme_qpair_t *);
344 static int nvme_alloc_qpair(nvme_t *, uint32_t, nvme_qpair_t **, int);
345 static int nvme_create_io_qpair(nvme_t *, nvme_qpair_t *, uint16_t);
346 
347 static inline void nvme_put64(nvme_t *, uintptr_t, uint64_t);
348 static inline void nvme_put32(nvme_t *, uintptr_t, uint32_t);
349 static inline uint64_t nvme_get64(nvme_t *, uintptr_t);
350 static inline uint32_t nvme_get32(nvme_t *, uintptr_t);
351 
352 static boolean_t nvme_check_regs_hdl(nvme_t *);
353 static boolean_t nvme_check_dma_hdl(nvme_dma_t *);
354 
355 static int nvme_fill_prp(nvme_cmd_t *, bd_xfer_t *);
356 
357 static void nvme_bd_xfer_done(void *);
358 static void nvme_bd_driveinfo(void *, bd_drive_t *);
359 static int nvme_bd_mediainfo(void *, bd_media_t *);
360 static int nvme_bd_cmd(nvme_namespace_t *, bd_xfer_t *, uint8_t);
361 static int nvme_bd_read(void *, bd_xfer_t *);
362 static int nvme_bd_write(void *, bd_xfer_t *);
363 static int nvme_bd_sync(void *, bd_xfer_t *);
364 static int nvme_bd_devid(void *, dev_info_t *, ddi_devid_t *);
365 
366 static int nvme_prp_dma_constructor(void *, void *, int);
367 static void nvme_prp_dma_destructor(void *, void *);
368 
369 static void nvme_prepare_devid(nvme_t *, uint32_t);
370 
371 static int nvme_open(dev_t *, int, int, cred_t *);
372 static int nvme_close(dev_t, int, int, cred_t *);
373 static int nvme_ioctl(dev_t, int, intptr_t, int, cred_t *, int *);
374 
375 #define	NVME_MINOR_INST_SHIFT	9
376 #define	NVME_MINOR(inst, nsid)	(((inst) << NVME_MINOR_INST_SHIFT) | (nsid))
377 #define	NVME_MINOR_INST(minor)	((minor) >> NVME_MINOR_INST_SHIFT)
378 #define	NVME_MINOR_NSID(minor)	((minor) & ((1 << NVME_MINOR_INST_SHIFT) - 1))
379 #define	NVME_MINOR_MAX		(NVME_MINOR(1, 0) - 2)
380 
381 static void *nvme_state;
382 static kmem_cache_t *nvme_cmd_cache;
383 
384 /*
385  * DMA attributes for queue DMA memory
386  *
387  * Queue DMA memory must be page aligned. The maximum length of a queue is
388  * 65536 entries, and an entry can be 64 bytes long.
389  */
390 static ddi_dma_attr_t nvme_queue_dma_attr = {
391 	.dma_attr_version	= DMA_ATTR_V0,
392 	.dma_attr_addr_lo	= 0,
393 	.dma_attr_addr_hi	= 0xffffffffffffffffULL,
394 	.dma_attr_count_max	= (UINT16_MAX + 1) * sizeof (nvme_sqe_t) - 1,
395 	.dma_attr_align		= 0x1000,
396 	.dma_attr_burstsizes	= 0x7ff,
397 	.dma_attr_minxfer	= 0x1000,
398 	.dma_attr_maxxfer	= (UINT16_MAX + 1) * sizeof (nvme_sqe_t),
399 	.dma_attr_seg		= 0xffffffffffffffffULL,
400 	.dma_attr_sgllen	= 1,
401 	.dma_attr_granular	= 1,
402 	.dma_attr_flags		= 0,
403 };
404 
405 /*
406  * DMA attributes for transfers using Physical Region Page (PRP) entries
407  *
408  * A PRP entry describes one page of DMA memory using the page size specified
409  * in the controller configuration's memory page size register (CC.MPS). It uses
410  * a 64bit base address aligned to this page size. There is no limitation on
411  * chaining PRPs together for arbitrarily large DMA transfers.
412  */
413 static ddi_dma_attr_t nvme_prp_dma_attr = {
414 	.dma_attr_version	= DMA_ATTR_V0,
415 	.dma_attr_addr_lo	= 0,
416 	.dma_attr_addr_hi	= 0xffffffffffffffffULL,
417 	.dma_attr_count_max	= 0xfff,
418 	.dma_attr_align		= 0x1000,
419 	.dma_attr_burstsizes	= 0x7ff,
420 	.dma_attr_minxfer	= 0x1000,
421 	.dma_attr_maxxfer	= 0x1000,
422 	.dma_attr_seg		= 0xfff,
423 	.dma_attr_sgllen	= -1,
424 	.dma_attr_granular	= 1,
425 	.dma_attr_flags		= 0,
426 };
427 
428 /*
429  * DMA attributes for transfers using scatter/gather lists
430  *
431  * A SGL entry describes a chunk of DMA memory using a 64bit base address and a
432  * 32bit length field. SGL Segment and SGL Last Segment entries require the
433  * length to be a multiple of 16 bytes.
434  */
435 static ddi_dma_attr_t nvme_sgl_dma_attr = {
436 	.dma_attr_version	= DMA_ATTR_V0,
437 	.dma_attr_addr_lo	= 0,
438 	.dma_attr_addr_hi	= 0xffffffffffffffffULL,
439 	.dma_attr_count_max	= 0xffffffffUL,
440 	.dma_attr_align		= 1,
441 	.dma_attr_burstsizes	= 0x7ff,
442 	.dma_attr_minxfer	= 0x10,
443 	.dma_attr_maxxfer	= 0xfffffffffULL,
444 	.dma_attr_seg		= 0xffffffffffffffffULL,
445 	.dma_attr_sgllen	= -1,
446 	.dma_attr_granular	= 0x10,
447 	.dma_attr_flags		= 0
448 };
449 
450 static ddi_device_acc_attr_t nvme_reg_acc_attr = {
451 	.devacc_attr_version	= DDI_DEVICE_ATTR_V0,
452 	.devacc_attr_endian_flags = DDI_STRUCTURE_LE_ACC,
453 	.devacc_attr_dataorder	= DDI_STRICTORDER_ACC
454 };
455 
456 static struct cb_ops nvme_cb_ops = {
457 	.cb_open	= nvme_open,
458 	.cb_close	= nvme_close,
459 	.cb_strategy	= nodev,
460 	.cb_print	= nodev,
461 	.cb_dump	= nodev,
462 	.cb_read	= nodev,
463 	.cb_write	= nodev,
464 	.cb_ioctl	= nvme_ioctl,
465 	.cb_devmap	= nodev,
466 	.cb_mmap	= nodev,
467 	.cb_segmap	= nodev,
468 	.cb_chpoll	= nochpoll,
469 	.cb_prop_op	= ddi_prop_op,
470 	.cb_str		= 0,
471 	.cb_flag	= D_NEW | D_MP,
472 	.cb_rev		= CB_REV,
473 	.cb_aread	= nodev,
474 	.cb_awrite	= nodev
475 };
476 
477 static struct dev_ops nvme_dev_ops = {
478 	.devo_rev	= DEVO_REV,
479 	.devo_refcnt	= 0,
480 	.devo_getinfo	= ddi_no_info,
481 	.devo_identify	= nulldev,
482 	.devo_probe	= nulldev,
483 	.devo_attach	= nvme_attach,
484 	.devo_detach	= nvme_detach,
485 	.devo_reset	= nodev,
486 	.devo_cb_ops	= &nvme_cb_ops,
487 	.devo_bus_ops	= NULL,
488 	.devo_power	= NULL,
489 	.devo_quiesce	= nvme_quiesce,
490 };
491 
492 static struct modldrv nvme_modldrv = {
493 	.drv_modops	= &mod_driverops,
494 	.drv_linkinfo	= "NVMe v1.1b",
495 	.drv_dev_ops	= &nvme_dev_ops
496 };
497 
498 static struct modlinkage nvme_modlinkage = {
499 	.ml_rev		= MODREV_1,
500 	.ml_linkage	= { &nvme_modldrv, NULL }
501 };
502 
503 static bd_ops_t nvme_bd_ops = {
504 	.o_version	= BD_OPS_VERSION_0,
505 	.o_drive_info	= nvme_bd_driveinfo,
506 	.o_media_info	= nvme_bd_mediainfo,
507 	.o_devid_init	= nvme_bd_devid,
508 	.o_sync_cache	= nvme_bd_sync,
509 	.o_read		= nvme_bd_read,
510 	.o_write	= nvme_bd_write,
511 };
512 
513 /*
514  * This list will hold commands that have timed out and couldn't be aborted.
515  * As we don't know what the hardware may still do with the DMA memory we can't
516  * free them, so we'll keep them forever on this list where we can easily look
517  * at them with mdb.
518  */
519 static struct list nvme_lost_cmds;
520 static kmutex_t nvme_lc_mutex;
521 
522 int
523 _init(void)
524 {
525 	int error;
526 
527 	error = ddi_soft_state_init(&nvme_state, sizeof (nvme_t), 1);
528 	if (error != DDI_SUCCESS)
529 		return (error);
530 
531 	nvme_cmd_cache = kmem_cache_create("nvme_cmd_cache",
532 	    sizeof (nvme_cmd_t), 64, NULL, NULL, NULL, NULL, NULL, 0);
533 
534 	mutex_init(&nvme_lc_mutex, NULL, MUTEX_DRIVER, NULL);
535 	list_create(&nvme_lost_cmds, sizeof (nvme_cmd_t),
536 	    offsetof(nvme_cmd_t, nc_list));
537 
538 	bd_mod_init(&nvme_dev_ops);
539 
540 	error = mod_install(&nvme_modlinkage);
541 	if (error != DDI_SUCCESS) {
542 		ddi_soft_state_fini(&nvme_state);
543 		mutex_destroy(&nvme_lc_mutex);
544 		list_destroy(&nvme_lost_cmds);
545 		bd_mod_fini(&nvme_dev_ops);
546 	}
547 
548 	return (error);
549 }
550 
551 int
552 _fini(void)
553 {
554 	int error;
555 
556 	if (!list_is_empty(&nvme_lost_cmds))
557 		return (DDI_FAILURE);
558 
559 	error = mod_remove(&nvme_modlinkage);
560 	if (error == DDI_SUCCESS) {
561 		ddi_soft_state_fini(&nvme_state);
562 		kmem_cache_destroy(nvme_cmd_cache);
563 		mutex_destroy(&nvme_lc_mutex);
564 		list_destroy(&nvme_lost_cmds);
565 		bd_mod_fini(&nvme_dev_ops);
566 	}
567 
568 	return (error);
569 }
570 
571 int
572 _info(struct modinfo *modinfop)
573 {
574 	return (mod_info(&nvme_modlinkage, modinfop));
575 }
576 
577 static inline void
578 nvme_put64(nvme_t *nvme, uintptr_t reg, uint64_t val)
579 {
580 	ASSERT(((uintptr_t)(nvme->n_regs + reg) & 0x7) == 0);
581 
582 	/*LINTED: E_BAD_PTR_CAST_ALIGN*/
583 	ddi_put64(nvme->n_regh, (uint64_t *)(nvme->n_regs + reg), val);
584 }
585 
586 static inline void
587 nvme_put32(nvme_t *nvme, uintptr_t reg, uint32_t val)
588 {
589 	ASSERT(((uintptr_t)(nvme->n_regs + reg) & 0x3) == 0);
590 
591 	/*LINTED: E_BAD_PTR_CAST_ALIGN*/
592 	ddi_put32(nvme->n_regh, (uint32_t *)(nvme->n_regs + reg), val);
593 }
594 
595 static inline uint64_t
596 nvme_get64(nvme_t *nvme, uintptr_t reg)
597 {
598 	uint64_t val;
599 
600 	ASSERT(((uintptr_t)(nvme->n_regs + reg) & 0x7) == 0);
601 
602 	/*LINTED: E_BAD_PTR_CAST_ALIGN*/
603 	val = ddi_get64(nvme->n_regh, (uint64_t *)(nvme->n_regs + reg));
604 
605 	return (val);
606 }
607 
608 static inline uint32_t
609 nvme_get32(nvme_t *nvme, uintptr_t reg)
610 {
611 	uint32_t val;
612 
613 	ASSERT(((uintptr_t)(nvme->n_regs + reg) & 0x3) == 0);
614 
615 	/*LINTED: E_BAD_PTR_CAST_ALIGN*/
616 	val = ddi_get32(nvme->n_regh, (uint32_t *)(nvme->n_regs + reg));
617 
618 	return (val);
619 }
620 
621 static boolean_t
622 nvme_check_regs_hdl(nvme_t *nvme)
623 {
624 	ddi_fm_error_t error;
625 
626 	ddi_fm_acc_err_get(nvme->n_regh, &error, DDI_FME_VERSION);
627 
628 	if (error.fme_status != DDI_FM_OK)
629 		return (B_TRUE);
630 
631 	return (B_FALSE);
632 }
633 
634 static boolean_t
635 nvme_check_dma_hdl(nvme_dma_t *dma)
636 {
637 	ddi_fm_error_t error;
638 
639 	if (dma == NULL)
640 		return (B_FALSE);
641 
642 	ddi_fm_dma_err_get(dma->nd_dmah, &error, DDI_FME_VERSION);
643 
644 	if (error.fme_status != DDI_FM_OK)
645 		return (B_TRUE);
646 
647 	return (B_FALSE);
648 }
649 
650 static void
651 nvme_free_dma_common(nvme_dma_t *dma)
652 {
653 	if (dma->nd_dmah != NULL)
654 		(void) ddi_dma_unbind_handle(dma->nd_dmah);
655 	if (dma->nd_acch != NULL)
656 		ddi_dma_mem_free(&dma->nd_acch);
657 	if (dma->nd_dmah != NULL)
658 		ddi_dma_free_handle(&dma->nd_dmah);
659 }
660 
661 static void
662 nvme_free_dma(nvme_dma_t *dma)
663 {
664 	nvme_free_dma_common(dma);
665 	kmem_free(dma, sizeof (*dma));
666 }
667 
668 /* ARGSUSED */
669 static void
670 nvme_prp_dma_destructor(void *buf, void *private)
671 {
672 	nvme_dma_t *dma = (nvme_dma_t *)buf;
673 
674 	nvme_free_dma_common(dma);
675 }
676 
677 static int
678 nvme_alloc_dma_common(nvme_t *nvme, nvme_dma_t *dma,
679     size_t len, uint_t flags, ddi_dma_attr_t *dma_attr)
680 {
681 	if (ddi_dma_alloc_handle(nvme->n_dip, dma_attr, DDI_DMA_SLEEP, NULL,
682 	    &dma->nd_dmah) != DDI_SUCCESS) {
683 		/*
684 		 * Due to DDI_DMA_SLEEP this can't be DDI_DMA_NORESOURCES, and
685 		 * the only other possible error is DDI_DMA_BADATTR which
686 		 * indicates a driver bug which should cause a panic.
687 		 */
688 		dev_err(nvme->n_dip, CE_PANIC,
689 		    "!failed to get DMA handle, check DMA attributes");
690 		return (DDI_FAILURE);
691 	}
692 
693 	/*
694 	 * ddi_dma_mem_alloc() can only fail when DDI_DMA_NOSLEEP is specified
695 	 * or the flags are conflicting, which isn't the case here.
696 	 */
697 	(void) ddi_dma_mem_alloc(dma->nd_dmah, len, &nvme->n_reg_acc_attr,
698 	    DDI_DMA_CONSISTENT, DDI_DMA_SLEEP, NULL, &dma->nd_memp,
699 	    &dma->nd_len, &dma->nd_acch);
700 
701 	if (ddi_dma_addr_bind_handle(dma->nd_dmah, NULL, dma->nd_memp,
702 	    dma->nd_len, flags | DDI_DMA_CONSISTENT, DDI_DMA_SLEEP, NULL,
703 	    &dma->nd_cookie, &dma->nd_ncookie) != DDI_DMA_MAPPED) {
704 		dev_err(nvme->n_dip, CE_WARN,
705 		    "!failed to bind DMA memory");
706 		atomic_inc_32(&nvme->n_dma_bind_err);
707 		nvme_free_dma_common(dma);
708 		return (DDI_FAILURE);
709 	}
710 
711 	return (DDI_SUCCESS);
712 }
713 
714 static int
715 nvme_zalloc_dma(nvme_t *nvme, size_t len, uint_t flags,
716     ddi_dma_attr_t *dma_attr, nvme_dma_t **ret)
717 {
718 	nvme_dma_t *dma = kmem_zalloc(sizeof (nvme_dma_t), KM_SLEEP);
719 
720 	if (nvme_alloc_dma_common(nvme, dma, len, flags, dma_attr) !=
721 	    DDI_SUCCESS) {
722 		*ret = NULL;
723 		kmem_free(dma, sizeof (nvme_dma_t));
724 		return (DDI_FAILURE);
725 	}
726 
727 	bzero(dma->nd_memp, dma->nd_len);
728 
729 	*ret = dma;
730 	return (DDI_SUCCESS);
731 }
732 
733 /* ARGSUSED */
734 static int
735 nvme_prp_dma_constructor(void *buf, void *private, int flags)
736 {
737 	nvme_dma_t *dma = (nvme_dma_t *)buf;
738 	nvme_t *nvme = (nvme_t *)private;
739 
740 	dma->nd_dmah = NULL;
741 	dma->nd_acch = NULL;
742 
743 	if (nvme_alloc_dma_common(nvme, dma, nvme->n_pagesize,
744 	    DDI_DMA_READ, &nvme->n_prp_dma_attr) != DDI_SUCCESS) {
745 		return (-1);
746 	}
747 
748 	ASSERT(dma->nd_ncookie == 1);
749 
750 	dma->nd_cached = B_TRUE;
751 
752 	return (0);
753 }
754 
755 static int
756 nvme_zalloc_queue_dma(nvme_t *nvme, uint32_t nentry, uint16_t qe_len,
757     uint_t flags, nvme_dma_t **dma)
758 {
759 	uint32_t len = nentry * qe_len;
760 	ddi_dma_attr_t q_dma_attr = nvme->n_queue_dma_attr;
761 
762 	len = roundup(len, nvme->n_pagesize);
763 
764 	q_dma_attr.dma_attr_minxfer = len;
765 
766 	if (nvme_zalloc_dma(nvme, len, flags, &q_dma_attr, dma)
767 	    != DDI_SUCCESS) {
768 		dev_err(nvme->n_dip, CE_WARN,
769 		    "!failed to get DMA memory for queue");
770 		goto fail;
771 	}
772 
773 	if ((*dma)->nd_ncookie != 1) {
774 		dev_err(nvme->n_dip, CE_WARN,
775 		    "!got too many cookies for queue DMA");
776 		goto fail;
777 	}
778 
779 	return (DDI_SUCCESS);
780 
781 fail:
782 	if (*dma) {
783 		nvme_free_dma(*dma);
784 		*dma = NULL;
785 	}
786 
787 	return (DDI_FAILURE);
788 }
789 
790 static void
791 nvme_free_qpair(nvme_qpair_t *qp)
792 {
793 	int i;
794 
795 	mutex_destroy(&qp->nq_mutex);
796 	sema_destroy(&qp->nq_sema);
797 
798 	if (qp->nq_sqdma != NULL)
799 		nvme_free_dma(qp->nq_sqdma);
800 	if (qp->nq_cqdma != NULL)
801 		nvme_free_dma(qp->nq_cqdma);
802 
803 	if (qp->nq_active_cmds > 0)
804 		for (i = 0; i != qp->nq_nentry; i++)
805 			if (qp->nq_cmd[i] != NULL)
806 				nvme_free_cmd(qp->nq_cmd[i]);
807 
808 	if (qp->nq_cmd != NULL)
809 		kmem_free(qp->nq_cmd, sizeof (nvme_cmd_t *) * qp->nq_nentry);
810 
811 	kmem_free(qp, sizeof (nvme_qpair_t));
812 }
813 
814 static int
815 nvme_alloc_qpair(nvme_t *nvme, uint32_t nentry, nvme_qpair_t **nqp,
816     int idx)
817 {
818 	nvme_qpair_t *qp = kmem_zalloc(sizeof (*qp), KM_SLEEP);
819 
820 	mutex_init(&qp->nq_mutex, NULL, MUTEX_DRIVER,
821 	    DDI_INTR_PRI(nvme->n_intr_pri));
822 	sema_init(&qp->nq_sema, nentry, NULL, SEMA_DRIVER, NULL);
823 
824 	if (nvme_zalloc_queue_dma(nvme, nentry, sizeof (nvme_sqe_t),
825 	    DDI_DMA_WRITE, &qp->nq_sqdma) != DDI_SUCCESS)
826 		goto fail;
827 
828 	if (nvme_zalloc_queue_dma(nvme, nentry, sizeof (nvme_cqe_t),
829 	    DDI_DMA_READ, &qp->nq_cqdma) != DDI_SUCCESS)
830 		goto fail;
831 
832 	qp->nq_sq = (nvme_sqe_t *)qp->nq_sqdma->nd_memp;
833 	qp->nq_cq = (nvme_cqe_t *)qp->nq_cqdma->nd_memp;
834 	qp->nq_nentry = nentry;
835 
836 	qp->nq_sqtdbl = NVME_REG_SQTDBL(nvme, idx);
837 	qp->nq_cqhdbl = NVME_REG_CQHDBL(nvme, idx);
838 
839 	qp->nq_cmd = kmem_zalloc(sizeof (nvme_cmd_t *) * nentry, KM_SLEEP);
840 	qp->nq_next_cmd = 0;
841 
842 	*nqp = qp;
843 	return (DDI_SUCCESS);
844 
845 fail:
846 	nvme_free_qpair(qp);
847 	*nqp = NULL;
848 
849 	return (DDI_FAILURE);
850 }
851 
852 static nvme_cmd_t *
853 nvme_alloc_cmd(nvme_t *nvme, int kmflag)
854 {
855 	nvme_cmd_t *cmd = kmem_cache_alloc(nvme_cmd_cache, kmflag);
856 
857 	if (cmd == NULL)
858 		return (cmd);
859 
860 	bzero(cmd, sizeof (nvme_cmd_t));
861 
862 	cmd->nc_nvme = nvme;
863 
864 	mutex_init(&cmd->nc_mutex, NULL, MUTEX_DRIVER,
865 	    DDI_INTR_PRI(nvme->n_intr_pri));
866 	cv_init(&cmd->nc_cv, NULL, CV_DRIVER, NULL);
867 
868 	return (cmd);
869 }
870 
871 static void
872 nvme_free_cmd(nvme_cmd_t *cmd)
873 {
874 	/* Don't free commands on the lost commands list. */
875 	if (list_link_active(&cmd->nc_list))
876 		return;
877 
878 	if (cmd->nc_dma) {
879 		if (cmd->nc_dma->nd_cached)
880 			kmem_cache_free(cmd->nc_nvme->n_prp_cache,
881 			    cmd->nc_dma);
882 		else
883 			nvme_free_dma(cmd->nc_dma);
884 		cmd->nc_dma = NULL;
885 	}
886 
887 	cv_destroy(&cmd->nc_cv);
888 	mutex_destroy(&cmd->nc_mutex);
889 
890 	kmem_cache_free(nvme_cmd_cache, cmd);
891 }
892 
893 static void
894 nvme_submit_admin_cmd(nvme_qpair_t *qp, nvme_cmd_t *cmd)
895 {
896 	sema_p(&qp->nq_sema);
897 	nvme_submit_cmd_common(qp, cmd);
898 }
899 
900 static int
901 nvme_submit_io_cmd(nvme_qpair_t *qp, nvme_cmd_t *cmd)
902 {
903 	if (sema_tryp(&qp->nq_sema) == 0)
904 		return (EAGAIN);
905 
906 	nvme_submit_cmd_common(qp, cmd);
907 	return (0);
908 }
909 
910 static void
911 nvme_submit_cmd_common(nvme_qpair_t *qp, nvme_cmd_t *cmd)
912 {
913 	nvme_reg_sqtdbl_t tail = { 0 };
914 
915 	mutex_enter(&qp->nq_mutex);
916 	cmd->nc_completed = B_FALSE;
917 
918 	/*
919 	 * Try to insert the cmd into the active cmd array at the nq_next_cmd
920 	 * slot. If the slot is already occupied advance to the next slot and
921 	 * try again. This can happen for long running commands like async event
922 	 * requests.
923 	 */
924 	while (qp->nq_cmd[qp->nq_next_cmd] != NULL)
925 		qp->nq_next_cmd = (qp->nq_next_cmd + 1) % qp->nq_nentry;
926 	qp->nq_cmd[qp->nq_next_cmd] = cmd;
927 
928 	qp->nq_active_cmds++;
929 
930 	cmd->nc_sqe.sqe_cid = qp->nq_next_cmd;
931 	bcopy(&cmd->nc_sqe, &qp->nq_sq[qp->nq_sqtail], sizeof (nvme_sqe_t));
932 	(void) ddi_dma_sync(qp->nq_sqdma->nd_dmah,
933 	    sizeof (nvme_sqe_t) * qp->nq_sqtail,
934 	    sizeof (nvme_sqe_t), DDI_DMA_SYNC_FORDEV);
935 	qp->nq_next_cmd = (qp->nq_next_cmd + 1) % qp->nq_nentry;
936 
937 	tail.b.sqtdbl_sqt = qp->nq_sqtail = (qp->nq_sqtail + 1) % qp->nq_nentry;
938 	nvme_put32(cmd->nc_nvme, qp->nq_sqtdbl, tail.r);
939 
940 	mutex_exit(&qp->nq_mutex);
941 }
942 
943 static nvme_cmd_t *
944 nvme_unqueue_cmd(nvme_t *nvme, nvme_qpair_t *qp, int cid)
945 {
946 	nvme_cmd_t *cmd;
947 
948 	ASSERT(mutex_owned(&qp->nq_mutex));
949 	ASSERT3S(cid, <, qp->nq_nentry);
950 
951 	cmd = qp->nq_cmd[cid];
952 	qp->nq_cmd[cid] = NULL;
953 	ASSERT3U(qp->nq_active_cmds, >, 0);
954 	qp->nq_active_cmds--;
955 	sema_v(&qp->nq_sema);
956 
957 	ASSERT3P(cmd, !=, NULL);
958 	ASSERT3P(cmd->nc_nvme, ==, nvme);
959 	ASSERT3S(cmd->nc_sqe.sqe_cid, ==, cid);
960 
961 	return (cmd);
962 }
963 
964 static nvme_cmd_t *
965 nvme_retrieve_cmd(nvme_t *nvme, nvme_qpair_t *qp)
966 {
967 	nvme_reg_cqhdbl_t head = { 0 };
968 
969 	nvme_cqe_t *cqe;
970 	nvme_cmd_t *cmd;
971 
972 	(void) ddi_dma_sync(qp->nq_cqdma->nd_dmah, 0,
973 	    sizeof (nvme_cqe_t) * qp->nq_nentry, DDI_DMA_SYNC_FORKERNEL);
974 
975 	mutex_enter(&qp->nq_mutex);
976 	cqe = &qp->nq_cq[qp->nq_cqhead];
977 
978 	/* Check phase tag of CQE. Hardware inverts it for new entries. */
979 	if (cqe->cqe_sf.sf_p == qp->nq_phase) {
980 		mutex_exit(&qp->nq_mutex);
981 		return (NULL);
982 	}
983 
984 	ASSERT(nvme->n_ioq[cqe->cqe_sqid] == qp);
985 
986 	cmd = nvme_unqueue_cmd(nvme, qp, cqe->cqe_cid);
987 
988 	ASSERT(cmd->nc_sqid == cqe->cqe_sqid);
989 	bcopy(cqe, &cmd->nc_cqe, sizeof (nvme_cqe_t));
990 
991 	qp->nq_sqhead = cqe->cqe_sqhd;
992 
993 	head.b.cqhdbl_cqh = qp->nq_cqhead = (qp->nq_cqhead + 1) % qp->nq_nentry;
994 
995 	/* Toggle phase on wrap-around. */
996 	if (qp->nq_cqhead == 0)
997 		qp->nq_phase = qp->nq_phase ? 0 : 1;
998 
999 	nvme_put32(cmd->nc_nvme, qp->nq_cqhdbl, head.r);
1000 	mutex_exit(&qp->nq_mutex);
1001 
1002 	return (cmd);
1003 }
1004 
1005 static int
1006 nvme_check_unknown_cmd_status(nvme_cmd_t *cmd)
1007 {
1008 	nvme_cqe_t *cqe = &cmd->nc_cqe;
1009 
1010 	dev_err(cmd->nc_nvme->n_dip, CE_WARN,
1011 	    "!unknown command status received: opc = %x, sqid = %d, cid = %d, "
1012 	    "sc = %x, sct = %x, dnr = %d, m = %d", cmd->nc_sqe.sqe_opc,
1013 	    cqe->cqe_sqid, cqe->cqe_cid, cqe->cqe_sf.sf_sc, cqe->cqe_sf.sf_sct,
1014 	    cqe->cqe_sf.sf_dnr, cqe->cqe_sf.sf_m);
1015 
1016 	if (cmd->nc_xfer != NULL)
1017 		bd_error(cmd->nc_xfer, BD_ERR_ILLRQ);
1018 
1019 	if (cmd->nc_nvme->n_strict_version) {
1020 		cmd->nc_nvme->n_dead = B_TRUE;
1021 		ddi_fm_service_impact(cmd->nc_nvme->n_dip, DDI_SERVICE_LOST);
1022 	}
1023 
1024 	return (EIO);
1025 }
1026 
1027 static int
1028 nvme_check_vendor_cmd_status(nvme_cmd_t *cmd)
1029 {
1030 	nvme_cqe_t *cqe = &cmd->nc_cqe;
1031 
1032 	dev_err(cmd->nc_nvme->n_dip, CE_WARN,
1033 	    "!unknown command status received: opc = %x, sqid = %d, cid = %d, "
1034 	    "sc = %x, sct = %x, dnr = %d, m = %d", cmd->nc_sqe.sqe_opc,
1035 	    cqe->cqe_sqid, cqe->cqe_cid, cqe->cqe_sf.sf_sc, cqe->cqe_sf.sf_sct,
1036 	    cqe->cqe_sf.sf_dnr, cqe->cqe_sf.sf_m);
1037 	if (!cmd->nc_nvme->n_ignore_unknown_vendor_status) {
1038 		cmd->nc_nvme->n_dead = B_TRUE;
1039 		ddi_fm_service_impact(cmd->nc_nvme->n_dip, DDI_SERVICE_LOST);
1040 	}
1041 
1042 	return (EIO);
1043 }
1044 
1045 static int
1046 nvme_check_integrity_cmd_status(nvme_cmd_t *cmd)
1047 {
1048 	nvme_cqe_t *cqe = &cmd->nc_cqe;
1049 
1050 	switch (cqe->cqe_sf.sf_sc) {
1051 	case NVME_CQE_SC_INT_NVM_WRITE:
1052 		/* write fail */
1053 		/* TODO: post ereport */
1054 		if (cmd->nc_xfer != NULL)
1055 			bd_error(cmd->nc_xfer, BD_ERR_MEDIA);
1056 		return (EIO);
1057 
1058 	case NVME_CQE_SC_INT_NVM_READ:
1059 		/* read fail */
1060 		/* TODO: post ereport */
1061 		if (cmd->nc_xfer != NULL)
1062 			bd_error(cmd->nc_xfer, BD_ERR_MEDIA);
1063 		return (EIO);
1064 
1065 	default:
1066 		return (nvme_check_unknown_cmd_status(cmd));
1067 	}
1068 }
1069 
1070 static int
1071 nvme_check_generic_cmd_status(nvme_cmd_t *cmd)
1072 {
1073 	nvme_cqe_t *cqe = &cmd->nc_cqe;
1074 
1075 	switch (cqe->cqe_sf.sf_sc) {
1076 	case NVME_CQE_SC_GEN_SUCCESS:
1077 		return (0);
1078 
1079 	/*
1080 	 * Errors indicating a bug in the driver should cause a panic.
1081 	 */
1082 	case NVME_CQE_SC_GEN_INV_OPC:
1083 		/* Invalid Command Opcode */
1084 		if (!cmd->nc_dontpanic)
1085 			dev_err(cmd->nc_nvme->n_dip, CE_PANIC,
1086 			    "programming error: invalid opcode in cmd %p",
1087 			    (void *)cmd);
1088 		return (EINVAL);
1089 
1090 	case NVME_CQE_SC_GEN_INV_FLD:
1091 		/* Invalid Field in Command */
1092 		if (!cmd->nc_dontpanic)
1093 			dev_err(cmd->nc_nvme->n_dip, CE_PANIC,
1094 			    "programming error: invalid field in cmd %p",
1095 			    (void *)cmd);
1096 		return (EIO);
1097 
1098 	case NVME_CQE_SC_GEN_ID_CNFL:
1099 		/* Command ID Conflict */
1100 		dev_err(cmd->nc_nvme->n_dip, CE_PANIC, "programming error: "
1101 		    "cmd ID conflict in cmd %p", (void *)cmd);
1102 		return (0);
1103 
1104 	case NVME_CQE_SC_GEN_INV_NS:
1105 		/* Invalid Namespace or Format */
1106 		if (!cmd->nc_dontpanic)
1107 			dev_err(cmd->nc_nvme->n_dip, CE_PANIC,
1108 			    "programming error: invalid NS/format in cmd %p",
1109 			    (void *)cmd);
1110 		return (EINVAL);
1111 
1112 	case NVME_CQE_SC_GEN_NVM_LBA_RANGE:
1113 		/* LBA Out Of Range */
1114 		dev_err(cmd->nc_nvme->n_dip, CE_PANIC, "programming error: "
1115 		    "LBA out of range in cmd %p", (void *)cmd);
1116 		return (0);
1117 
1118 	/*
1119 	 * Non-fatal errors, handle gracefully.
1120 	 */
1121 	case NVME_CQE_SC_GEN_DATA_XFR_ERR:
1122 		/* Data Transfer Error (DMA) */
1123 		/* TODO: post ereport */
1124 		atomic_inc_32(&cmd->nc_nvme->n_data_xfr_err);
1125 		if (cmd->nc_xfer != NULL)
1126 			bd_error(cmd->nc_xfer, BD_ERR_NTRDY);
1127 		return (EIO);
1128 
1129 	case NVME_CQE_SC_GEN_INTERNAL_ERR:
1130 		/*
1131 		 * Internal Error. The spec (v1.0, section 4.5.1.2) says
1132 		 * detailed error information is returned as async event,
1133 		 * so we pretty much ignore the error here and handle it
1134 		 * in the async event handler.
1135 		 */
1136 		atomic_inc_32(&cmd->nc_nvme->n_internal_err);
1137 		if (cmd->nc_xfer != NULL)
1138 			bd_error(cmd->nc_xfer, BD_ERR_NTRDY);
1139 		return (EIO);
1140 
1141 	case NVME_CQE_SC_GEN_ABORT_REQUEST:
1142 		/*
1143 		 * Command Abort Requested. This normally happens only when a
1144 		 * command times out.
1145 		 */
1146 		/* TODO: post ereport or change blkdev to handle this? */
1147 		atomic_inc_32(&cmd->nc_nvme->n_abort_rq_err);
1148 		return (ECANCELED);
1149 
1150 	case NVME_CQE_SC_GEN_ABORT_PWRLOSS:
1151 		/* Command Aborted due to Power Loss Notification */
1152 		ddi_fm_service_impact(cmd->nc_nvme->n_dip, DDI_SERVICE_LOST);
1153 		cmd->nc_nvme->n_dead = B_TRUE;
1154 		return (EIO);
1155 
1156 	case NVME_CQE_SC_GEN_ABORT_SQ_DEL:
1157 		/* Command Aborted due to SQ Deletion */
1158 		atomic_inc_32(&cmd->nc_nvme->n_abort_sq_del);
1159 		return (EIO);
1160 
1161 	case NVME_CQE_SC_GEN_NVM_CAP_EXC:
1162 		/* Capacity Exceeded */
1163 		atomic_inc_32(&cmd->nc_nvme->n_nvm_cap_exc);
1164 		if (cmd->nc_xfer != NULL)
1165 			bd_error(cmd->nc_xfer, BD_ERR_MEDIA);
1166 		return (EIO);
1167 
1168 	case NVME_CQE_SC_GEN_NVM_NS_NOTRDY:
1169 		/* Namespace Not Ready */
1170 		atomic_inc_32(&cmd->nc_nvme->n_nvm_ns_notrdy);
1171 		if (cmd->nc_xfer != NULL)
1172 			bd_error(cmd->nc_xfer, BD_ERR_NTRDY);
1173 		return (EIO);
1174 
1175 	default:
1176 		return (nvme_check_unknown_cmd_status(cmd));
1177 	}
1178 }
1179 
1180 static int
1181 nvme_check_specific_cmd_status(nvme_cmd_t *cmd)
1182 {
1183 	nvme_cqe_t *cqe = &cmd->nc_cqe;
1184 
1185 	switch (cqe->cqe_sf.sf_sc) {
1186 	case NVME_CQE_SC_SPC_INV_CQ:
1187 		/* Completion Queue Invalid */
1188 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_CREATE_SQUEUE);
1189 		atomic_inc_32(&cmd->nc_nvme->n_inv_cq_err);
1190 		return (EINVAL);
1191 
1192 	case NVME_CQE_SC_SPC_INV_QID:
1193 		/* Invalid Queue Identifier */
1194 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_CREATE_SQUEUE ||
1195 		    cmd->nc_sqe.sqe_opc == NVME_OPC_DELETE_SQUEUE ||
1196 		    cmd->nc_sqe.sqe_opc == NVME_OPC_CREATE_CQUEUE ||
1197 		    cmd->nc_sqe.sqe_opc == NVME_OPC_DELETE_CQUEUE);
1198 		atomic_inc_32(&cmd->nc_nvme->n_inv_qid_err);
1199 		return (EINVAL);
1200 
1201 	case NVME_CQE_SC_SPC_MAX_QSZ_EXC:
1202 		/* Max Queue Size Exceeded */
1203 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_CREATE_SQUEUE ||
1204 		    cmd->nc_sqe.sqe_opc == NVME_OPC_CREATE_CQUEUE);
1205 		atomic_inc_32(&cmd->nc_nvme->n_max_qsz_exc);
1206 		return (EINVAL);
1207 
1208 	case NVME_CQE_SC_SPC_ABRT_CMD_EXC:
1209 		/* Abort Command Limit Exceeded */
1210 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_ABORT);
1211 		dev_err(cmd->nc_nvme->n_dip, CE_PANIC, "programming error: "
1212 		    "abort command limit exceeded in cmd %p", (void *)cmd);
1213 		return (0);
1214 
1215 	case NVME_CQE_SC_SPC_ASYNC_EVREQ_EXC:
1216 		/* Async Event Request Limit Exceeded */
1217 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_ASYNC_EVENT);
1218 		dev_err(cmd->nc_nvme->n_dip, CE_PANIC, "programming error: "
1219 		    "async event request limit exceeded in cmd %p",
1220 		    (void *)cmd);
1221 		return (0);
1222 
1223 	case NVME_CQE_SC_SPC_INV_INT_VECT:
1224 		/* Invalid Interrupt Vector */
1225 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_CREATE_CQUEUE);
1226 		atomic_inc_32(&cmd->nc_nvme->n_inv_int_vect);
1227 		return (EINVAL);
1228 
1229 	case NVME_CQE_SC_SPC_INV_LOG_PAGE:
1230 		/* Invalid Log Page */
1231 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_GET_LOG_PAGE);
1232 		atomic_inc_32(&cmd->nc_nvme->n_inv_log_page);
1233 		return (EINVAL);
1234 
1235 	case NVME_CQE_SC_SPC_INV_FORMAT:
1236 		/* Invalid Format */
1237 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_NVM_FORMAT);
1238 		atomic_inc_32(&cmd->nc_nvme->n_inv_format);
1239 		if (cmd->nc_xfer != NULL)
1240 			bd_error(cmd->nc_xfer, BD_ERR_ILLRQ);
1241 		return (EINVAL);
1242 
1243 	case NVME_CQE_SC_SPC_INV_Q_DEL:
1244 		/* Invalid Queue Deletion */
1245 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_DELETE_CQUEUE);
1246 		atomic_inc_32(&cmd->nc_nvme->n_inv_q_del);
1247 		return (EINVAL);
1248 
1249 	case NVME_CQE_SC_SPC_NVM_CNFL_ATTR:
1250 		/* Conflicting Attributes */
1251 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_NVM_DSET_MGMT ||
1252 		    cmd->nc_sqe.sqe_opc == NVME_OPC_NVM_READ ||
1253 		    cmd->nc_sqe.sqe_opc == NVME_OPC_NVM_WRITE);
1254 		atomic_inc_32(&cmd->nc_nvme->n_cnfl_attr);
1255 		if (cmd->nc_xfer != NULL)
1256 			bd_error(cmd->nc_xfer, BD_ERR_ILLRQ);
1257 		return (EINVAL);
1258 
1259 	case NVME_CQE_SC_SPC_NVM_INV_PROT:
1260 		/* Invalid Protection Information */
1261 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_NVM_COMPARE ||
1262 		    cmd->nc_sqe.sqe_opc == NVME_OPC_NVM_READ ||
1263 		    cmd->nc_sqe.sqe_opc == NVME_OPC_NVM_WRITE);
1264 		atomic_inc_32(&cmd->nc_nvme->n_inv_prot);
1265 		if (cmd->nc_xfer != NULL)
1266 			bd_error(cmd->nc_xfer, BD_ERR_ILLRQ);
1267 		return (EINVAL);
1268 
1269 	case NVME_CQE_SC_SPC_NVM_READONLY:
1270 		/* Write to Read Only Range */
1271 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_NVM_WRITE);
1272 		atomic_inc_32(&cmd->nc_nvme->n_readonly);
1273 		if (cmd->nc_xfer != NULL)
1274 			bd_error(cmd->nc_xfer, BD_ERR_ILLRQ);
1275 		return (EROFS);
1276 
1277 	default:
1278 		return (nvme_check_unknown_cmd_status(cmd));
1279 	}
1280 }
1281 
1282 static inline int
1283 nvme_check_cmd_status(nvme_cmd_t *cmd)
1284 {
1285 	nvme_cqe_t *cqe = &cmd->nc_cqe;
1286 
1287 	/*
1288 	 * Take a shortcut if the controller is dead, or if
1289 	 * command status indicates no error.
1290 	 */
1291 	if (cmd->nc_nvme->n_dead)
1292 		return (EIO);
1293 
1294 	if (cqe->cqe_sf.sf_sct == NVME_CQE_SCT_GENERIC &&
1295 	    cqe->cqe_sf.sf_sc == NVME_CQE_SC_GEN_SUCCESS)
1296 		return (0);
1297 
1298 	if (cqe->cqe_sf.sf_sct == NVME_CQE_SCT_GENERIC)
1299 		return (nvme_check_generic_cmd_status(cmd));
1300 	else if (cqe->cqe_sf.sf_sct == NVME_CQE_SCT_SPECIFIC)
1301 		return (nvme_check_specific_cmd_status(cmd));
1302 	else if (cqe->cqe_sf.sf_sct == NVME_CQE_SCT_INTEGRITY)
1303 		return (nvme_check_integrity_cmd_status(cmd));
1304 	else if (cqe->cqe_sf.sf_sct == NVME_CQE_SCT_VENDOR)
1305 		return (nvme_check_vendor_cmd_status(cmd));
1306 
1307 	return (nvme_check_unknown_cmd_status(cmd));
1308 }
1309 
1310 static int
1311 nvme_abort_cmd(nvme_cmd_t *abort_cmd, uint_t sec)
1312 {
1313 	nvme_t *nvme = abort_cmd->nc_nvme;
1314 	nvme_cmd_t *cmd = nvme_alloc_cmd(nvme, KM_SLEEP);
1315 	nvme_abort_cmd_t ac = { 0 };
1316 	int ret = 0;
1317 
1318 	sema_p(&nvme->n_abort_sema);
1319 
1320 	ac.b.ac_cid = abort_cmd->nc_sqe.sqe_cid;
1321 	ac.b.ac_sqid = abort_cmd->nc_sqid;
1322 
1323 	cmd->nc_sqid = 0;
1324 	cmd->nc_sqe.sqe_opc = NVME_OPC_ABORT;
1325 	cmd->nc_callback = nvme_wakeup_cmd;
1326 	cmd->nc_sqe.sqe_cdw10 = ac.r;
1327 
1328 	/*
1329 	 * Send the ABORT to the hardware. The ABORT command will return _after_
1330 	 * the aborted command has completed (aborted or otherwise), but since
1331 	 * we still hold the aborted command's mutex its callback hasn't been
1332 	 * processed yet.
1333 	 */
1334 	nvme_admin_cmd(cmd, sec);
1335 	sema_v(&nvme->n_abort_sema);
1336 
1337 	if ((ret = nvme_check_cmd_status(cmd)) != 0) {
1338 		dev_err(nvme->n_dip, CE_WARN,
1339 		    "!ABORT failed with sct = %x, sc = %x",
1340 		    cmd->nc_cqe.cqe_sf.sf_sct, cmd->nc_cqe.cqe_sf.sf_sc);
1341 		atomic_inc_32(&nvme->n_abort_failed);
1342 	} else {
1343 		dev_err(nvme->n_dip, CE_WARN,
1344 		    "!ABORT of command %d/%d %ssuccessful",
1345 		    abort_cmd->nc_sqe.sqe_cid, abort_cmd->nc_sqid,
1346 		    cmd->nc_cqe.cqe_dw0 & 1 ? "un" : "");
1347 		if ((cmd->nc_cqe.cqe_dw0 & 1) == 0)
1348 			atomic_inc_32(&nvme->n_cmd_aborted);
1349 	}
1350 
1351 	nvme_free_cmd(cmd);
1352 	return (ret);
1353 }
1354 
1355 /*
1356  * nvme_wait_cmd -- wait for command completion or timeout
1357  *
1358  * In case of a serious error or a timeout of the abort command the hardware
1359  * will be declared dead and FMA will be notified.
1360  */
1361 static void
1362 nvme_wait_cmd(nvme_cmd_t *cmd, uint_t sec)
1363 {
1364 	clock_t timeout = ddi_get_lbolt() + drv_usectohz(sec * MICROSEC);
1365 	nvme_t *nvme = cmd->nc_nvme;
1366 	nvme_reg_csts_t csts;
1367 	nvme_qpair_t *qp;
1368 
1369 	ASSERT(mutex_owned(&cmd->nc_mutex));
1370 
1371 	while (!cmd->nc_completed) {
1372 		if (cv_timedwait(&cmd->nc_cv, &cmd->nc_mutex, timeout) == -1)
1373 			break;
1374 	}
1375 
1376 	if (cmd->nc_completed)
1377 		return;
1378 
1379 	/*
1380 	 * The command timed out.
1381 	 *
1382 	 * Check controller for fatal status, any errors associated with the
1383 	 * register or DMA handle, or for a double timeout (abort command timed
1384 	 * out). If necessary log a warning and call FMA.
1385 	 */
1386 	csts.r = nvme_get32(nvme, NVME_REG_CSTS);
1387 	dev_err(nvme->n_dip, CE_WARN, "!command %d/%d timeout, "
1388 	    "OPC = %x, CFS = %d", cmd->nc_sqe.sqe_cid, cmd->nc_sqid,
1389 	    cmd->nc_sqe.sqe_opc, csts.b.csts_cfs);
1390 	atomic_inc_32(&nvme->n_cmd_timeout);
1391 
1392 	if (csts.b.csts_cfs ||
1393 	    nvme_check_regs_hdl(nvme) ||
1394 	    nvme_check_dma_hdl(cmd->nc_dma) ||
1395 	    cmd->nc_sqe.sqe_opc == NVME_OPC_ABORT) {
1396 		ddi_fm_service_impact(nvme->n_dip, DDI_SERVICE_LOST);
1397 		nvme->n_dead = B_TRUE;
1398 	} else if (nvme_abort_cmd(cmd, sec) == 0) {
1399 		/*
1400 		 * If the abort succeeded the command should complete
1401 		 * immediately with an appropriate status.
1402 		 */
1403 		while (!cmd->nc_completed)
1404 			cv_wait(&cmd->nc_cv, &cmd->nc_mutex);
1405 
1406 		return;
1407 	}
1408 
1409 	qp = nvme->n_ioq[cmd->nc_sqid];
1410 
1411 	mutex_enter(&qp->nq_mutex);
1412 	(void) nvme_unqueue_cmd(nvme, qp, cmd->nc_sqe.sqe_cid);
1413 	mutex_exit(&qp->nq_mutex);
1414 
1415 	/*
1416 	 * As we don't know what the presumed dead hardware might still do with
1417 	 * the DMA memory, we'll put the command on the lost commands list if it
1418 	 * has any DMA memory.
1419 	 */
1420 	if (cmd->nc_dma != NULL) {
1421 		mutex_enter(&nvme_lc_mutex);
1422 		list_insert_head(&nvme_lost_cmds, cmd);
1423 		mutex_exit(&nvme_lc_mutex);
1424 	}
1425 }
1426 
1427 static void
1428 nvme_wakeup_cmd(void *arg)
1429 {
1430 	nvme_cmd_t *cmd = arg;
1431 
1432 	mutex_enter(&cmd->nc_mutex);
1433 	cmd->nc_completed = B_TRUE;
1434 	cv_signal(&cmd->nc_cv);
1435 	mutex_exit(&cmd->nc_mutex);
1436 }
1437 
1438 static void
1439 nvme_async_event_task(void *arg)
1440 {
1441 	nvme_cmd_t *cmd = arg;
1442 	nvme_t *nvme = cmd->nc_nvme;
1443 	nvme_error_log_entry_t *error_log = NULL;
1444 	nvme_health_log_t *health_log = NULL;
1445 	size_t logsize = 0;
1446 	nvme_async_event_t event;
1447 
1448 	/*
1449 	 * Check for errors associated with the async request itself. The only
1450 	 * command-specific error is "async event limit exceeded", which
1451 	 * indicates a programming error in the driver and causes a panic in
1452 	 * nvme_check_cmd_status().
1453 	 *
1454 	 * Other possible errors are various scenarios where the async request
1455 	 * was aborted, or internal errors in the device. Internal errors are
1456 	 * reported to FMA, the command aborts need no special handling here.
1457 	 *
1458 	 * And finally, at least qemu nvme does not support async events,
1459 	 * and will return NVME_CQE_SC_GEN_INV_OPC | DNR. If so, we
1460 	 * will avoid posting async events.
1461 	 */
1462 
1463 	if (nvme_check_cmd_status(cmd) != 0) {
1464 		dev_err(cmd->nc_nvme->n_dip, CE_WARN,
1465 		    "!async event request returned failure, sct = %x, "
1466 		    "sc = %x, dnr = %d, m = %d", cmd->nc_cqe.cqe_sf.sf_sct,
1467 		    cmd->nc_cqe.cqe_sf.sf_sc, cmd->nc_cqe.cqe_sf.sf_dnr,
1468 		    cmd->nc_cqe.cqe_sf.sf_m);
1469 
1470 		if (cmd->nc_cqe.cqe_sf.sf_sct == NVME_CQE_SCT_GENERIC &&
1471 		    cmd->nc_cqe.cqe_sf.sf_sc == NVME_CQE_SC_GEN_INTERNAL_ERR) {
1472 			cmd->nc_nvme->n_dead = B_TRUE;
1473 			ddi_fm_service_impact(cmd->nc_nvme->n_dip,
1474 			    DDI_SERVICE_LOST);
1475 		}
1476 
1477 		if (cmd->nc_cqe.cqe_sf.sf_sct == NVME_CQE_SCT_GENERIC &&
1478 		    cmd->nc_cqe.cqe_sf.sf_sc == NVME_CQE_SC_GEN_INV_OPC &&
1479 		    cmd->nc_cqe.cqe_sf.sf_dnr == 1) {
1480 			nvme->n_async_event_supported = B_FALSE;
1481 		}
1482 
1483 		nvme_free_cmd(cmd);
1484 		return;
1485 	}
1486 
1487 
1488 	event.r = cmd->nc_cqe.cqe_dw0;
1489 
1490 	/* Clear CQE and re-submit the async request. */
1491 	bzero(&cmd->nc_cqe, sizeof (nvme_cqe_t));
1492 	nvme_submit_admin_cmd(nvme->n_adminq, cmd);
1493 
1494 	switch (event.b.ae_type) {
1495 	case NVME_ASYNC_TYPE_ERROR:
1496 		if (event.b.ae_logpage == NVME_LOGPAGE_ERROR) {
1497 			(void) nvme_get_logpage(nvme, (void **)&error_log,
1498 			    &logsize, event.b.ae_logpage);
1499 		} else {
1500 			dev_err(nvme->n_dip, CE_WARN, "!wrong logpage in "
1501 			    "async event reply: %d", event.b.ae_logpage);
1502 			atomic_inc_32(&nvme->n_wrong_logpage);
1503 		}
1504 
1505 		switch (event.b.ae_info) {
1506 		case NVME_ASYNC_ERROR_INV_SQ:
1507 			dev_err(nvme->n_dip, CE_PANIC, "programming error: "
1508 			    "invalid submission queue");
1509 			return;
1510 
1511 		case NVME_ASYNC_ERROR_INV_DBL:
1512 			dev_err(nvme->n_dip, CE_PANIC, "programming error: "
1513 			    "invalid doorbell write value");
1514 			return;
1515 
1516 		case NVME_ASYNC_ERROR_DIAGFAIL:
1517 			dev_err(nvme->n_dip, CE_WARN, "!diagnostic failure");
1518 			ddi_fm_service_impact(nvme->n_dip, DDI_SERVICE_LOST);
1519 			nvme->n_dead = B_TRUE;
1520 			atomic_inc_32(&nvme->n_diagfail_event);
1521 			break;
1522 
1523 		case NVME_ASYNC_ERROR_PERSISTENT:
1524 			dev_err(nvme->n_dip, CE_WARN, "!persistent internal "
1525 			    "device error");
1526 			ddi_fm_service_impact(nvme->n_dip, DDI_SERVICE_LOST);
1527 			nvme->n_dead = B_TRUE;
1528 			atomic_inc_32(&nvme->n_persistent_event);
1529 			break;
1530 
1531 		case NVME_ASYNC_ERROR_TRANSIENT:
1532 			dev_err(nvme->n_dip, CE_WARN, "!transient internal "
1533 			    "device error");
1534 			/* TODO: send ereport */
1535 			atomic_inc_32(&nvme->n_transient_event);
1536 			break;
1537 
1538 		case NVME_ASYNC_ERROR_FW_LOAD:
1539 			dev_err(nvme->n_dip, CE_WARN,
1540 			    "!firmware image load error");
1541 			atomic_inc_32(&nvme->n_fw_load_event);
1542 			break;
1543 		}
1544 		break;
1545 
1546 	case NVME_ASYNC_TYPE_HEALTH:
1547 		if (event.b.ae_logpage == NVME_LOGPAGE_HEALTH) {
1548 			(void) nvme_get_logpage(nvme, (void **)&health_log,
1549 			    &logsize, event.b.ae_logpage, -1);
1550 		} else {
1551 			dev_err(nvme->n_dip, CE_WARN, "!wrong logpage in "
1552 			    "async event reply: %d", event.b.ae_logpage);
1553 			atomic_inc_32(&nvme->n_wrong_logpage);
1554 		}
1555 
1556 		switch (event.b.ae_info) {
1557 		case NVME_ASYNC_HEALTH_RELIABILITY:
1558 			dev_err(nvme->n_dip, CE_WARN,
1559 			    "!device reliability compromised");
1560 			/* TODO: send ereport */
1561 			atomic_inc_32(&nvme->n_reliability_event);
1562 			break;
1563 
1564 		case NVME_ASYNC_HEALTH_TEMPERATURE:
1565 			dev_err(nvme->n_dip, CE_WARN,
1566 			    "!temperature above threshold");
1567 			/* TODO: send ereport */
1568 			atomic_inc_32(&nvme->n_temperature_event);
1569 			break;
1570 
1571 		case NVME_ASYNC_HEALTH_SPARE:
1572 			dev_err(nvme->n_dip, CE_WARN,
1573 			    "!spare space below threshold");
1574 			/* TODO: send ereport */
1575 			atomic_inc_32(&nvme->n_spare_event);
1576 			break;
1577 		}
1578 		break;
1579 
1580 	case NVME_ASYNC_TYPE_VENDOR:
1581 		dev_err(nvme->n_dip, CE_WARN, "!vendor specific async event "
1582 		    "received, info = %x, logpage = %x", event.b.ae_info,
1583 		    event.b.ae_logpage);
1584 		atomic_inc_32(&nvme->n_vendor_event);
1585 		break;
1586 
1587 	default:
1588 		dev_err(nvme->n_dip, CE_WARN, "!unknown async event received, "
1589 		    "type = %x, info = %x, logpage = %x", event.b.ae_type,
1590 		    event.b.ae_info, event.b.ae_logpage);
1591 		atomic_inc_32(&nvme->n_unknown_event);
1592 		break;
1593 	}
1594 
1595 	if (error_log)
1596 		kmem_free(error_log, logsize);
1597 
1598 	if (health_log)
1599 		kmem_free(health_log, logsize);
1600 }
1601 
1602 static void
1603 nvme_admin_cmd(nvme_cmd_t *cmd, int sec)
1604 {
1605 	mutex_enter(&cmd->nc_mutex);
1606 	nvme_submit_admin_cmd(cmd->nc_nvme->n_adminq, cmd);
1607 	nvme_wait_cmd(cmd, sec);
1608 	mutex_exit(&cmd->nc_mutex);
1609 }
1610 
1611 static void
1612 nvme_async_event(nvme_t *nvme)
1613 {
1614 	nvme_cmd_t *cmd;
1615 
1616 	cmd = nvme_alloc_cmd(nvme, KM_SLEEP);
1617 	cmd->nc_sqid = 0;
1618 	cmd->nc_sqe.sqe_opc = NVME_OPC_ASYNC_EVENT;
1619 	cmd->nc_callback = nvme_async_event_task;
1620 	cmd->nc_dontpanic = B_TRUE;
1621 
1622 	nvme_submit_admin_cmd(nvme->n_adminq, cmd);
1623 }
1624 
1625 static int
1626 nvme_format_nvm(nvme_t *nvme, uint32_t nsid, uint8_t lbaf, boolean_t ms,
1627     uint8_t pi, boolean_t pil, uint8_t ses)
1628 {
1629 	nvme_cmd_t *cmd = nvme_alloc_cmd(nvme, KM_SLEEP);
1630 	nvme_format_nvm_t format_nvm = { 0 };
1631 	int ret;
1632 
1633 	format_nvm.b.fm_lbaf = lbaf & 0xf;
1634 	format_nvm.b.fm_ms = ms ? 1 : 0;
1635 	format_nvm.b.fm_pi = pi & 0x7;
1636 	format_nvm.b.fm_pil = pil ? 1 : 0;
1637 	format_nvm.b.fm_ses = ses & 0x7;
1638 
1639 	cmd->nc_sqid = 0;
1640 	cmd->nc_callback = nvme_wakeup_cmd;
1641 	cmd->nc_sqe.sqe_nsid = nsid;
1642 	cmd->nc_sqe.sqe_opc = NVME_OPC_NVM_FORMAT;
1643 	cmd->nc_sqe.sqe_cdw10 = format_nvm.r;
1644 
1645 	/*
1646 	 * Some devices like Samsung SM951 don't allow formatting of all
1647 	 * namespaces in one command. Handle that gracefully.
1648 	 */
1649 	if (nsid == (uint32_t)-1)
1650 		cmd->nc_dontpanic = B_TRUE;
1651 
1652 	nvme_admin_cmd(cmd, nvme_format_cmd_timeout);
1653 
1654 	if ((ret = nvme_check_cmd_status(cmd)) != 0) {
1655 		dev_err(nvme->n_dip, CE_WARN,
1656 		    "!FORMAT failed with sct = %x, sc = %x",
1657 		    cmd->nc_cqe.cqe_sf.sf_sct, cmd->nc_cqe.cqe_sf.sf_sc);
1658 	}
1659 
1660 	nvme_free_cmd(cmd);
1661 	return (ret);
1662 }
1663 
1664 static int
1665 nvme_get_logpage(nvme_t *nvme, void **buf, size_t *bufsize, uint8_t logpage,
1666     ...)
1667 {
1668 	nvme_cmd_t *cmd = nvme_alloc_cmd(nvme, KM_SLEEP);
1669 	nvme_getlogpage_t getlogpage = { 0 };
1670 	va_list ap;
1671 	int ret;
1672 
1673 	va_start(ap, logpage);
1674 
1675 	cmd->nc_sqid = 0;
1676 	cmd->nc_callback = nvme_wakeup_cmd;
1677 	cmd->nc_sqe.sqe_opc = NVME_OPC_GET_LOG_PAGE;
1678 
1679 	getlogpage.b.lp_lid = logpage;
1680 
1681 	switch (logpage) {
1682 	case NVME_LOGPAGE_ERROR:
1683 		cmd->nc_sqe.sqe_nsid = (uint32_t)-1;
1684 		/*
1685 		 * The GET LOG PAGE command can use at most 2 pages to return
1686 		 * data, PRP lists are not supported.
1687 		 */
1688 		*bufsize = MIN(2 * nvme->n_pagesize,
1689 		    nvme->n_error_log_len * sizeof (nvme_error_log_entry_t));
1690 		break;
1691 
1692 	case NVME_LOGPAGE_HEALTH:
1693 		cmd->nc_sqe.sqe_nsid = va_arg(ap, uint32_t);
1694 		*bufsize = sizeof (nvme_health_log_t);
1695 		break;
1696 
1697 	case NVME_LOGPAGE_FWSLOT:
1698 		cmd->nc_sqe.sqe_nsid = (uint32_t)-1;
1699 		*bufsize = sizeof (nvme_fwslot_log_t);
1700 		break;
1701 
1702 	default:
1703 		dev_err(nvme->n_dip, CE_WARN, "!unknown log page requested: %d",
1704 		    logpage);
1705 		atomic_inc_32(&nvme->n_unknown_logpage);
1706 		ret = EINVAL;
1707 		goto fail;
1708 	}
1709 
1710 	va_end(ap);
1711 
1712 	getlogpage.b.lp_numd = *bufsize / sizeof (uint32_t) - 1;
1713 
1714 	cmd->nc_sqe.sqe_cdw10 = getlogpage.r;
1715 
1716 	if (nvme_zalloc_dma(nvme, getlogpage.b.lp_numd * sizeof (uint32_t),
1717 	    DDI_DMA_READ, &nvme->n_prp_dma_attr, &cmd->nc_dma) != DDI_SUCCESS) {
1718 		dev_err(nvme->n_dip, CE_WARN,
1719 		    "!nvme_zalloc_dma failed for GET LOG PAGE");
1720 		ret = ENOMEM;
1721 		goto fail;
1722 	}
1723 
1724 	if (cmd->nc_dma->nd_ncookie > 2) {
1725 		dev_err(nvme->n_dip, CE_WARN,
1726 		    "!too many DMA cookies for GET LOG PAGE");
1727 		atomic_inc_32(&nvme->n_too_many_cookies);
1728 		ret = ENOMEM;
1729 		goto fail;
1730 	}
1731 
1732 	cmd->nc_sqe.sqe_dptr.d_prp[0] = cmd->nc_dma->nd_cookie.dmac_laddress;
1733 	if (cmd->nc_dma->nd_ncookie > 1) {
1734 		ddi_dma_nextcookie(cmd->nc_dma->nd_dmah,
1735 		    &cmd->nc_dma->nd_cookie);
1736 		cmd->nc_sqe.sqe_dptr.d_prp[1] =
1737 		    cmd->nc_dma->nd_cookie.dmac_laddress;
1738 	}
1739 
1740 	nvme_admin_cmd(cmd, nvme_admin_cmd_timeout);
1741 
1742 	if ((ret = nvme_check_cmd_status(cmd)) != 0) {
1743 		dev_err(nvme->n_dip, CE_WARN,
1744 		    "!GET LOG PAGE failed with sct = %x, sc = %x",
1745 		    cmd->nc_cqe.cqe_sf.sf_sct, cmd->nc_cqe.cqe_sf.sf_sc);
1746 		goto fail;
1747 	}
1748 
1749 	*buf = kmem_alloc(*bufsize, KM_SLEEP);
1750 	bcopy(cmd->nc_dma->nd_memp, *buf, *bufsize);
1751 
1752 fail:
1753 	nvme_free_cmd(cmd);
1754 
1755 	return (ret);
1756 }
1757 
1758 static int
1759 nvme_identify(nvme_t *nvme, uint32_t nsid, void **buf)
1760 {
1761 	nvme_cmd_t *cmd = nvme_alloc_cmd(nvme, KM_SLEEP);
1762 	int ret;
1763 
1764 	if (buf == NULL)
1765 		return (EINVAL);
1766 
1767 	cmd->nc_sqid = 0;
1768 	cmd->nc_callback = nvme_wakeup_cmd;
1769 	cmd->nc_sqe.sqe_opc = NVME_OPC_IDENTIFY;
1770 	cmd->nc_sqe.sqe_nsid = nsid;
1771 	cmd->nc_sqe.sqe_cdw10 = nsid ? NVME_IDENTIFY_NSID : NVME_IDENTIFY_CTRL;
1772 
1773 	if (nvme_zalloc_dma(nvme, NVME_IDENTIFY_BUFSIZE, DDI_DMA_READ,
1774 	    &nvme->n_prp_dma_attr, &cmd->nc_dma) != DDI_SUCCESS) {
1775 		dev_err(nvme->n_dip, CE_WARN,
1776 		    "!nvme_zalloc_dma failed for IDENTIFY");
1777 		ret = ENOMEM;
1778 		goto fail;
1779 	}
1780 
1781 	if (cmd->nc_dma->nd_ncookie > 2) {
1782 		dev_err(nvme->n_dip, CE_WARN,
1783 		    "!too many DMA cookies for IDENTIFY");
1784 		atomic_inc_32(&nvme->n_too_many_cookies);
1785 		ret = ENOMEM;
1786 		goto fail;
1787 	}
1788 
1789 	cmd->nc_sqe.sqe_dptr.d_prp[0] = cmd->nc_dma->nd_cookie.dmac_laddress;
1790 	if (cmd->nc_dma->nd_ncookie > 1) {
1791 		ddi_dma_nextcookie(cmd->nc_dma->nd_dmah,
1792 		    &cmd->nc_dma->nd_cookie);
1793 		cmd->nc_sqe.sqe_dptr.d_prp[1] =
1794 		    cmd->nc_dma->nd_cookie.dmac_laddress;
1795 	}
1796 
1797 	nvme_admin_cmd(cmd, nvme_admin_cmd_timeout);
1798 
1799 	if ((ret = nvme_check_cmd_status(cmd)) != 0) {
1800 		dev_err(nvme->n_dip, CE_WARN,
1801 		    "!IDENTIFY failed with sct = %x, sc = %x",
1802 		    cmd->nc_cqe.cqe_sf.sf_sct, cmd->nc_cqe.cqe_sf.sf_sc);
1803 		goto fail;
1804 	}
1805 
1806 	*buf = kmem_alloc(NVME_IDENTIFY_BUFSIZE, KM_SLEEP);
1807 	bcopy(cmd->nc_dma->nd_memp, *buf, NVME_IDENTIFY_BUFSIZE);
1808 
1809 fail:
1810 	nvme_free_cmd(cmd);
1811 
1812 	return (ret);
1813 }
1814 
1815 static int
1816 nvme_set_features(nvme_t *nvme, uint32_t nsid, uint8_t feature, uint32_t val,
1817     uint32_t *res)
1818 {
1819 	_NOTE(ARGUNUSED(nsid));
1820 	nvme_cmd_t *cmd = nvme_alloc_cmd(nvme, KM_SLEEP);
1821 	int ret = EINVAL;
1822 
1823 	ASSERT(res != NULL);
1824 
1825 	cmd->nc_sqid = 0;
1826 	cmd->nc_callback = nvme_wakeup_cmd;
1827 	cmd->nc_sqe.sqe_opc = NVME_OPC_SET_FEATURES;
1828 	cmd->nc_sqe.sqe_cdw10 = feature;
1829 	cmd->nc_sqe.sqe_cdw11 = val;
1830 
1831 	switch (feature) {
1832 	case NVME_FEAT_WRITE_CACHE:
1833 		if (!nvme->n_write_cache_present)
1834 			goto fail;
1835 		break;
1836 
1837 	case NVME_FEAT_NQUEUES:
1838 		break;
1839 
1840 	default:
1841 		goto fail;
1842 	}
1843 
1844 	nvme_admin_cmd(cmd, nvme_admin_cmd_timeout);
1845 
1846 	if ((ret = nvme_check_cmd_status(cmd)) != 0) {
1847 		dev_err(nvme->n_dip, CE_WARN,
1848 		    "!SET FEATURES %d failed with sct = %x, sc = %x",
1849 		    feature, cmd->nc_cqe.cqe_sf.sf_sct,
1850 		    cmd->nc_cqe.cqe_sf.sf_sc);
1851 		goto fail;
1852 	}
1853 
1854 	*res = cmd->nc_cqe.cqe_dw0;
1855 
1856 fail:
1857 	nvme_free_cmd(cmd);
1858 	return (ret);
1859 }
1860 
1861 static int
1862 nvme_get_features(nvme_t *nvme, uint32_t nsid, uint8_t feature, uint32_t *res,
1863     void **buf, size_t *bufsize)
1864 {
1865 	nvme_cmd_t *cmd = nvme_alloc_cmd(nvme, KM_SLEEP);
1866 	int ret = EINVAL;
1867 
1868 	ASSERT(res != NULL);
1869 
1870 	if (bufsize != NULL)
1871 		*bufsize = 0;
1872 
1873 	cmd->nc_sqid = 0;
1874 	cmd->nc_callback = nvme_wakeup_cmd;
1875 	cmd->nc_sqe.sqe_opc = NVME_OPC_GET_FEATURES;
1876 	cmd->nc_sqe.sqe_cdw10 = feature;
1877 	cmd->nc_sqe.sqe_cdw11 = *res;
1878 
1879 	/*
1880 	 * For some of the optional features there doesn't seem to be a method
1881 	 * of detecting whether it is supported other than using it.  This will
1882 	 * cause "Invalid Field in Command" error, which is normally considered
1883 	 * a programming error.  Set the nc_dontpanic flag to override the panic
1884 	 * in nvme_check_generic_cmd_status().
1885 	 */
1886 	switch (feature) {
1887 	case NVME_FEAT_ARBITRATION:
1888 	case NVME_FEAT_POWER_MGMT:
1889 	case NVME_FEAT_TEMPERATURE:
1890 	case NVME_FEAT_ERROR:
1891 	case NVME_FEAT_NQUEUES:
1892 	case NVME_FEAT_INTR_COAL:
1893 	case NVME_FEAT_INTR_VECT:
1894 	case NVME_FEAT_WRITE_ATOM:
1895 	case NVME_FEAT_ASYNC_EVENT:
1896 		break;
1897 
1898 	case NVME_FEAT_WRITE_CACHE:
1899 		if (!nvme->n_write_cache_present)
1900 			goto fail;
1901 		break;
1902 
1903 	case NVME_FEAT_LBA_RANGE:
1904 		if (!nvme->n_lba_range_supported)
1905 			goto fail;
1906 
1907 		cmd->nc_dontpanic = B_TRUE;
1908 		cmd->nc_sqe.sqe_nsid = nsid;
1909 		ASSERT(bufsize != NULL);
1910 		*bufsize = NVME_LBA_RANGE_BUFSIZE;
1911 		break;
1912 
1913 	case NVME_FEAT_AUTO_PST:
1914 		if (!nvme->n_auto_pst_supported)
1915 			goto fail;
1916 
1917 		ASSERT(bufsize != NULL);
1918 		*bufsize = NVME_AUTO_PST_BUFSIZE;
1919 		break;
1920 
1921 	case NVME_FEAT_PROGRESS:
1922 		if (!nvme->n_progress_supported)
1923 			goto fail;
1924 
1925 		cmd->nc_dontpanic = B_TRUE;
1926 		break;
1927 
1928 	default:
1929 		goto fail;
1930 	}
1931 
1932 	if (bufsize != NULL && *bufsize != 0) {
1933 		if (nvme_zalloc_dma(nvme, *bufsize, DDI_DMA_READ,
1934 		    &nvme->n_prp_dma_attr, &cmd->nc_dma) != DDI_SUCCESS) {
1935 			dev_err(nvme->n_dip, CE_WARN,
1936 			    "!nvme_zalloc_dma failed for GET FEATURES");
1937 			ret = ENOMEM;
1938 			goto fail;
1939 		}
1940 
1941 		if (cmd->nc_dma->nd_ncookie > 2) {
1942 			dev_err(nvme->n_dip, CE_WARN,
1943 			    "!too many DMA cookies for GET FEATURES");
1944 			atomic_inc_32(&nvme->n_too_many_cookies);
1945 			ret = ENOMEM;
1946 			goto fail;
1947 		}
1948 
1949 		cmd->nc_sqe.sqe_dptr.d_prp[0] =
1950 		    cmd->nc_dma->nd_cookie.dmac_laddress;
1951 		if (cmd->nc_dma->nd_ncookie > 1) {
1952 			ddi_dma_nextcookie(cmd->nc_dma->nd_dmah,
1953 			    &cmd->nc_dma->nd_cookie);
1954 			cmd->nc_sqe.sqe_dptr.d_prp[1] =
1955 			    cmd->nc_dma->nd_cookie.dmac_laddress;
1956 		}
1957 	}
1958 
1959 	nvme_admin_cmd(cmd, nvme_admin_cmd_timeout);
1960 
1961 	if ((ret = nvme_check_cmd_status(cmd)) != 0) {
1962 		boolean_t known = B_TRUE;
1963 
1964 		/* Check if this is unsupported optional feature */
1965 		if (cmd->nc_cqe.cqe_sf.sf_sct == NVME_CQE_SCT_GENERIC &&
1966 		    cmd->nc_cqe.cqe_sf.sf_sc == NVME_CQE_SC_GEN_INV_FLD) {
1967 			switch (feature) {
1968 			case NVME_FEAT_LBA_RANGE:
1969 				nvme->n_lba_range_supported = B_FALSE;
1970 				break;
1971 			case NVME_FEAT_PROGRESS:
1972 				nvme->n_progress_supported = B_FALSE;
1973 				break;
1974 			default:
1975 				known = B_FALSE;
1976 				break;
1977 			}
1978 		} else {
1979 			known = B_FALSE;
1980 		}
1981 
1982 		/* Report the error otherwise */
1983 		if (!known) {
1984 			dev_err(nvme->n_dip, CE_WARN,
1985 			    "!GET FEATURES %d failed with sct = %x, sc = %x",
1986 			    feature, cmd->nc_cqe.cqe_sf.sf_sct,
1987 			    cmd->nc_cqe.cqe_sf.sf_sc);
1988 		}
1989 
1990 		goto fail;
1991 	}
1992 
1993 	if (bufsize != NULL && *bufsize != 0) {
1994 		ASSERT(buf != NULL);
1995 		*buf = kmem_alloc(*bufsize, KM_SLEEP);
1996 		bcopy(cmd->nc_dma->nd_memp, *buf, *bufsize);
1997 	}
1998 
1999 	*res = cmd->nc_cqe.cqe_dw0;
2000 
2001 fail:
2002 	nvme_free_cmd(cmd);
2003 	return (ret);
2004 }
2005 
2006 static int
2007 nvme_write_cache_set(nvme_t *nvme, boolean_t enable)
2008 {
2009 	nvme_write_cache_t nwc = { 0 };
2010 
2011 	if (enable)
2012 		nwc.b.wc_wce = 1;
2013 
2014 	return (nvme_set_features(nvme, 0, NVME_FEAT_WRITE_CACHE, nwc.r,
2015 	    &nwc.r));
2016 }
2017 
2018 static int
2019 nvme_set_nqueues(nvme_t *nvme, uint16_t *nqueues)
2020 {
2021 	nvme_nqueues_t nq = { 0 };
2022 	int ret;
2023 
2024 	nq.b.nq_nsq = nq.b.nq_ncq = *nqueues - 1;
2025 
2026 	ret = nvme_set_features(nvme, 0, NVME_FEAT_NQUEUES, nq.r, &nq.r);
2027 
2028 	if (ret == 0) {
2029 		/*
2030 		 * Always use the same number of submission and completion
2031 		 * queues, and never use more than the requested number of
2032 		 * queues.
2033 		 */
2034 		*nqueues = MIN(*nqueues, MIN(nq.b.nq_nsq, nq.b.nq_ncq) + 1);
2035 	}
2036 
2037 	return (ret);
2038 }
2039 
2040 static int
2041 nvme_create_io_qpair(nvme_t *nvme, nvme_qpair_t *qp, uint16_t idx)
2042 {
2043 	nvme_cmd_t *cmd = nvme_alloc_cmd(nvme, KM_SLEEP);
2044 	nvme_create_queue_dw10_t dw10 = { 0 };
2045 	nvme_create_cq_dw11_t c_dw11 = { 0 };
2046 	nvme_create_sq_dw11_t s_dw11 = { 0 };
2047 	int ret;
2048 
2049 	dw10.b.q_qid = idx;
2050 	dw10.b.q_qsize = qp->nq_nentry - 1;
2051 
2052 	c_dw11.b.cq_pc = 1;
2053 	c_dw11.b.cq_ien = 1;
2054 	c_dw11.b.cq_iv = idx % nvme->n_intr_cnt;
2055 
2056 	cmd->nc_sqid = 0;
2057 	cmd->nc_callback = nvme_wakeup_cmd;
2058 	cmd->nc_sqe.sqe_opc = NVME_OPC_CREATE_CQUEUE;
2059 	cmd->nc_sqe.sqe_cdw10 = dw10.r;
2060 	cmd->nc_sqe.sqe_cdw11 = c_dw11.r;
2061 	cmd->nc_sqe.sqe_dptr.d_prp[0] = qp->nq_cqdma->nd_cookie.dmac_laddress;
2062 
2063 	nvme_admin_cmd(cmd, nvme_admin_cmd_timeout);
2064 
2065 	if ((ret = nvme_check_cmd_status(cmd)) != 0) {
2066 		dev_err(nvme->n_dip, CE_WARN,
2067 		    "!CREATE CQUEUE failed with sct = %x, sc = %x",
2068 		    cmd->nc_cqe.cqe_sf.sf_sct, cmd->nc_cqe.cqe_sf.sf_sc);
2069 		goto fail;
2070 	}
2071 
2072 	nvme_free_cmd(cmd);
2073 
2074 	s_dw11.b.sq_pc = 1;
2075 	s_dw11.b.sq_cqid = idx;
2076 
2077 	cmd = nvme_alloc_cmd(nvme, KM_SLEEP);
2078 	cmd->nc_sqid = 0;
2079 	cmd->nc_callback = nvme_wakeup_cmd;
2080 	cmd->nc_sqe.sqe_opc = NVME_OPC_CREATE_SQUEUE;
2081 	cmd->nc_sqe.sqe_cdw10 = dw10.r;
2082 	cmd->nc_sqe.sqe_cdw11 = s_dw11.r;
2083 	cmd->nc_sqe.sqe_dptr.d_prp[0] = qp->nq_sqdma->nd_cookie.dmac_laddress;
2084 
2085 	nvme_admin_cmd(cmd, nvme_admin_cmd_timeout);
2086 
2087 	if ((ret = nvme_check_cmd_status(cmd)) != 0) {
2088 		dev_err(nvme->n_dip, CE_WARN,
2089 		    "!CREATE SQUEUE failed with sct = %x, sc = %x",
2090 		    cmd->nc_cqe.cqe_sf.sf_sct, cmd->nc_cqe.cqe_sf.sf_sc);
2091 		goto fail;
2092 	}
2093 
2094 fail:
2095 	nvme_free_cmd(cmd);
2096 
2097 	return (ret);
2098 }
2099 
2100 static boolean_t
2101 nvme_reset(nvme_t *nvme, boolean_t quiesce)
2102 {
2103 	nvme_reg_csts_t csts;
2104 	int i;
2105 
2106 	nvme_put32(nvme, NVME_REG_CC, 0);
2107 
2108 	csts.r = nvme_get32(nvme, NVME_REG_CSTS);
2109 	if (csts.b.csts_rdy == 1) {
2110 		nvme_put32(nvme, NVME_REG_CC, 0);
2111 		for (i = 0; i != nvme->n_timeout * 10; i++) {
2112 			csts.r = nvme_get32(nvme, NVME_REG_CSTS);
2113 			if (csts.b.csts_rdy == 0)
2114 				break;
2115 
2116 			if (quiesce)
2117 				drv_usecwait(50000);
2118 			else
2119 				delay(drv_usectohz(50000));
2120 		}
2121 	}
2122 
2123 	nvme_put32(nvme, NVME_REG_AQA, 0);
2124 	nvme_put32(nvme, NVME_REG_ASQ, 0);
2125 	nvme_put32(nvme, NVME_REG_ACQ, 0);
2126 
2127 	csts.r = nvme_get32(nvme, NVME_REG_CSTS);
2128 	return (csts.b.csts_rdy == 0 ? B_TRUE : B_FALSE);
2129 }
2130 
2131 static void
2132 nvme_shutdown(nvme_t *nvme, int mode, boolean_t quiesce)
2133 {
2134 	nvme_reg_cc_t cc;
2135 	nvme_reg_csts_t csts;
2136 	int i;
2137 
2138 	ASSERT(mode == NVME_CC_SHN_NORMAL || mode == NVME_CC_SHN_ABRUPT);
2139 
2140 	cc.r = nvme_get32(nvme, NVME_REG_CC);
2141 	cc.b.cc_shn = mode & 0x3;
2142 	nvme_put32(nvme, NVME_REG_CC, cc.r);
2143 
2144 	for (i = 0; i != 10; i++) {
2145 		csts.r = nvme_get32(nvme, NVME_REG_CSTS);
2146 		if (csts.b.csts_shst == NVME_CSTS_SHN_COMPLETE)
2147 			break;
2148 
2149 		if (quiesce)
2150 			drv_usecwait(100000);
2151 		else
2152 			delay(drv_usectohz(100000));
2153 	}
2154 }
2155 
2156 
2157 static void
2158 nvme_prepare_devid(nvme_t *nvme, uint32_t nsid)
2159 {
2160 	/*
2161 	 * Section 7.7 of the spec describes how to get a unique ID for
2162 	 * the controller: the vendor ID, the model name and the serial
2163 	 * number shall be unique when combined.
2164 	 *
2165 	 * If a namespace has no EUI64 we use the above and add the hex
2166 	 * namespace ID to get a unique ID for the namespace.
2167 	 */
2168 	char model[sizeof (nvme->n_idctl->id_model) + 1];
2169 	char serial[sizeof (nvme->n_idctl->id_serial) + 1];
2170 
2171 	bcopy(nvme->n_idctl->id_model, model, sizeof (nvme->n_idctl->id_model));
2172 	bcopy(nvme->n_idctl->id_serial, serial,
2173 	    sizeof (nvme->n_idctl->id_serial));
2174 
2175 	model[sizeof (nvme->n_idctl->id_model)] = '\0';
2176 	serial[sizeof (nvme->n_idctl->id_serial)] = '\0';
2177 
2178 	nvme->n_ns[nsid - 1].ns_devid = kmem_asprintf("%4X-%s-%s-%X",
2179 	    nvme->n_idctl->id_vid, model, serial, nsid);
2180 }
2181 
2182 static int
2183 nvme_init_ns(nvme_t *nvme, int nsid)
2184 {
2185 	nvme_namespace_t *ns = &nvme->n_ns[nsid - 1];
2186 	nvme_identify_nsid_t *idns;
2187 	int last_rp;
2188 
2189 	ns->ns_nvme = nvme;
2190 
2191 	if (nvme_identify(nvme, nsid, (void **)&idns) != 0) {
2192 		dev_err(nvme->n_dip, CE_WARN,
2193 		    "!failed to identify namespace %d", nsid);
2194 		return (DDI_FAILURE);
2195 	}
2196 
2197 	ns->ns_idns = idns;
2198 	ns->ns_id = nsid;
2199 	ns->ns_block_count = idns->id_nsize;
2200 	ns->ns_block_size =
2201 	    1 << idns->id_lbaf[idns->id_flbas.lba_format].lbaf_lbads;
2202 	ns->ns_best_block_size = ns->ns_block_size;
2203 
2204 	/*
2205 	 * Get the EUI64 if present. Use it for devid and device node names.
2206 	 */
2207 	if (NVME_VERSION_ATLEAST(&nvme->n_version, 1, 1))
2208 		bcopy(idns->id_eui64, ns->ns_eui64, sizeof (ns->ns_eui64));
2209 
2210 	/*LINTED: E_BAD_PTR_CAST_ALIGN*/
2211 	if (*(uint64_t *)ns->ns_eui64 != 0) {
2212 		uint8_t *eui64 = ns->ns_eui64;
2213 
2214 		(void) snprintf(ns->ns_name, sizeof (ns->ns_name),
2215 		    "%02x%02x%02x%02x%02x%02x%02x%02x",
2216 		    eui64[0], eui64[1], eui64[2], eui64[3],
2217 		    eui64[4], eui64[5], eui64[6], eui64[7]);
2218 	} else {
2219 		(void) snprintf(ns->ns_name, sizeof (ns->ns_name), "%d",
2220 		    ns->ns_id);
2221 
2222 		nvme_prepare_devid(nvme, ns->ns_id);
2223 	}
2224 
2225 	/*
2226 	 * Find the LBA format with no metadata and the best relative
2227 	 * performance. A value of 3 means "degraded", 0 is best.
2228 	 */
2229 	last_rp = 3;
2230 	for (int j = 0; j <= idns->id_nlbaf; j++) {
2231 		if (idns->id_lbaf[j].lbaf_lbads == 0)
2232 			break;
2233 		if (idns->id_lbaf[j].lbaf_ms != 0)
2234 			continue;
2235 		if (idns->id_lbaf[j].lbaf_rp >= last_rp)
2236 			continue;
2237 		last_rp = idns->id_lbaf[j].lbaf_rp;
2238 		ns->ns_best_block_size =
2239 		    1 << idns->id_lbaf[j].lbaf_lbads;
2240 	}
2241 
2242 	if (ns->ns_best_block_size < nvme->n_min_block_size)
2243 		ns->ns_best_block_size = nvme->n_min_block_size;
2244 
2245 	/*
2246 	 * We currently don't support namespaces that use either:
2247 	 * - protection information
2248 	 * - illegal block size (< 512)
2249 	 */
2250 	if (idns->id_dps.dp_pinfo) {
2251 		dev_err(nvme->n_dip, CE_WARN,
2252 		    "!ignoring namespace %d, unsupported feature: "
2253 		    "pinfo = %d", nsid, idns->id_dps.dp_pinfo);
2254 		ns->ns_ignore = B_TRUE;
2255 	} else if (ns->ns_block_size < 512) {
2256 		dev_err(nvme->n_dip, CE_WARN,
2257 		    "!ignoring namespace %d, unsupported block size %"PRIu64,
2258 		    nsid, (uint64_t)ns->ns_block_size);
2259 		ns->ns_ignore = B_TRUE;
2260 	} else {
2261 		ns->ns_ignore = B_FALSE;
2262 	}
2263 
2264 	return (DDI_SUCCESS);
2265 }
2266 
2267 static int
2268 nvme_init(nvme_t *nvme)
2269 {
2270 	nvme_reg_cc_t cc = { 0 };
2271 	nvme_reg_aqa_t aqa = { 0 };
2272 	nvme_reg_asq_t asq = { 0 };
2273 	nvme_reg_acq_t acq = { 0 };
2274 	nvme_reg_cap_t cap;
2275 	nvme_reg_vs_t vs;
2276 	nvme_reg_csts_t csts;
2277 	int i = 0;
2278 	uint16_t nqueues;
2279 	char model[sizeof (nvme->n_idctl->id_model) + 1];
2280 	char *vendor, *product;
2281 
2282 	/* Check controller version */
2283 	vs.r = nvme_get32(nvme, NVME_REG_VS);
2284 	nvme->n_version.v_major = vs.b.vs_mjr;
2285 	nvme->n_version.v_minor = vs.b.vs_mnr;
2286 	dev_err(nvme->n_dip, CE_CONT, "?NVMe spec version %d.%d",
2287 	    nvme->n_version.v_major, nvme->n_version.v_minor);
2288 
2289 	if (nvme->n_version.v_major > nvme_version_major) {
2290 		dev_err(nvme->n_dip, CE_WARN, "!no support for version > %d.x",
2291 		    nvme_version_major);
2292 		if (nvme->n_strict_version)
2293 			goto fail;
2294 	}
2295 
2296 	/* retrieve controller configuration */
2297 	cap.r = nvme_get64(nvme, NVME_REG_CAP);
2298 
2299 	if ((cap.b.cap_css & NVME_CAP_CSS_NVM) == 0) {
2300 		dev_err(nvme->n_dip, CE_WARN,
2301 		    "!NVM command set not supported by hardware");
2302 		goto fail;
2303 	}
2304 
2305 	nvme->n_nssr_supported = cap.b.cap_nssrs;
2306 	nvme->n_doorbell_stride = 4 << cap.b.cap_dstrd;
2307 	nvme->n_timeout = cap.b.cap_to;
2308 	nvme->n_arbitration_mechanisms = cap.b.cap_ams;
2309 	nvme->n_cont_queues_reqd = cap.b.cap_cqr;
2310 	nvme->n_max_queue_entries = cap.b.cap_mqes + 1;
2311 
2312 	/*
2313 	 * The MPSMIN and MPSMAX fields in the CAP register use 0 to specify
2314 	 * the base page size of 4k (1<<12), so add 12 here to get the real
2315 	 * page size value.
2316 	 */
2317 	nvme->n_pageshift = MIN(MAX(cap.b.cap_mpsmin + 12, PAGESHIFT),
2318 	    cap.b.cap_mpsmax + 12);
2319 	nvme->n_pagesize = 1UL << (nvme->n_pageshift);
2320 
2321 	/*
2322 	 * Set up Queue DMA to transfer at least 1 page-aligned page at a time.
2323 	 */
2324 	nvme->n_queue_dma_attr.dma_attr_align = nvme->n_pagesize;
2325 	nvme->n_queue_dma_attr.dma_attr_minxfer = nvme->n_pagesize;
2326 
2327 	/*
2328 	 * Set up PRP DMA to transfer 1 page-aligned page at a time.
2329 	 * Maxxfer may be increased after we identified the controller limits.
2330 	 */
2331 	nvme->n_prp_dma_attr.dma_attr_maxxfer = nvme->n_pagesize;
2332 	nvme->n_prp_dma_attr.dma_attr_minxfer = nvme->n_pagesize;
2333 	nvme->n_prp_dma_attr.dma_attr_align = nvme->n_pagesize;
2334 	nvme->n_prp_dma_attr.dma_attr_seg = nvme->n_pagesize - 1;
2335 
2336 	/*
2337 	 * Reset controller if it's still in ready state.
2338 	 */
2339 	if (nvme_reset(nvme, B_FALSE) == B_FALSE) {
2340 		dev_err(nvme->n_dip, CE_WARN, "!unable to reset controller");
2341 		ddi_fm_service_impact(nvme->n_dip, DDI_SERVICE_LOST);
2342 		nvme->n_dead = B_TRUE;
2343 		goto fail;
2344 	}
2345 
2346 	/*
2347 	 * Create the admin queue pair.
2348 	 */
2349 	if (nvme_alloc_qpair(nvme, nvme->n_admin_queue_len, &nvme->n_adminq, 0)
2350 	    != DDI_SUCCESS) {
2351 		dev_err(nvme->n_dip, CE_WARN,
2352 		    "!unable to allocate admin qpair");
2353 		goto fail;
2354 	}
2355 	nvme->n_ioq = kmem_alloc(sizeof (nvme_qpair_t *), KM_SLEEP);
2356 	nvme->n_ioq[0] = nvme->n_adminq;
2357 
2358 	nvme->n_progress |= NVME_ADMIN_QUEUE;
2359 
2360 	(void) ddi_prop_update_int(DDI_DEV_T_NONE, nvme->n_dip,
2361 	    "admin-queue-len", nvme->n_admin_queue_len);
2362 
2363 	aqa.b.aqa_asqs = aqa.b.aqa_acqs = nvme->n_admin_queue_len - 1;
2364 	asq = nvme->n_adminq->nq_sqdma->nd_cookie.dmac_laddress;
2365 	acq = nvme->n_adminq->nq_cqdma->nd_cookie.dmac_laddress;
2366 
2367 	ASSERT((asq & (nvme->n_pagesize - 1)) == 0);
2368 	ASSERT((acq & (nvme->n_pagesize - 1)) == 0);
2369 
2370 	nvme_put32(nvme, NVME_REG_AQA, aqa.r);
2371 	nvme_put64(nvme, NVME_REG_ASQ, asq);
2372 	nvme_put64(nvme, NVME_REG_ACQ, acq);
2373 
2374 	cc.b.cc_ams = 0;	/* use Round-Robin arbitration */
2375 	cc.b.cc_css = 0;	/* use NVM command set */
2376 	cc.b.cc_mps = nvme->n_pageshift - 12;
2377 	cc.b.cc_shn = 0;	/* no shutdown in progress */
2378 	cc.b.cc_en = 1;		/* enable controller */
2379 	cc.b.cc_iosqes = 6;	/* submission queue entry is 2^6 bytes long */
2380 	cc.b.cc_iocqes = 4;	/* completion queue entry is 2^4 bytes long */
2381 
2382 	nvme_put32(nvme, NVME_REG_CC, cc.r);
2383 
2384 	/*
2385 	 * Wait for the controller to become ready.
2386 	 */
2387 	csts.r = nvme_get32(nvme, NVME_REG_CSTS);
2388 	if (csts.b.csts_rdy == 0) {
2389 		for (i = 0; i != nvme->n_timeout * 10; i++) {
2390 			delay(drv_usectohz(50000));
2391 			csts.r = nvme_get32(nvme, NVME_REG_CSTS);
2392 
2393 			if (csts.b.csts_cfs == 1) {
2394 				dev_err(nvme->n_dip, CE_WARN,
2395 				    "!controller fatal status at init");
2396 				ddi_fm_service_impact(nvme->n_dip,
2397 				    DDI_SERVICE_LOST);
2398 				nvme->n_dead = B_TRUE;
2399 				goto fail;
2400 			}
2401 
2402 			if (csts.b.csts_rdy == 1)
2403 				break;
2404 		}
2405 	}
2406 
2407 	if (csts.b.csts_rdy == 0) {
2408 		dev_err(nvme->n_dip, CE_WARN, "!controller not ready");
2409 		ddi_fm_service_impact(nvme->n_dip, DDI_SERVICE_LOST);
2410 		nvme->n_dead = B_TRUE;
2411 		goto fail;
2412 	}
2413 
2414 	/*
2415 	 * Assume an abort command limit of 1. We'll destroy and re-init
2416 	 * that later when we know the true abort command limit.
2417 	 */
2418 	sema_init(&nvme->n_abort_sema, 1, NULL, SEMA_DRIVER, NULL);
2419 
2420 	/*
2421 	 * Setup initial interrupt for admin queue.
2422 	 */
2423 	if ((nvme_setup_interrupts(nvme, DDI_INTR_TYPE_MSIX, 1)
2424 	    != DDI_SUCCESS) &&
2425 	    (nvme_setup_interrupts(nvme, DDI_INTR_TYPE_MSI, 1)
2426 	    != DDI_SUCCESS) &&
2427 	    (nvme_setup_interrupts(nvme, DDI_INTR_TYPE_FIXED, 1)
2428 	    != DDI_SUCCESS)) {
2429 		dev_err(nvme->n_dip, CE_WARN,
2430 		    "!failed to setup initial interrupt");
2431 		goto fail;
2432 	}
2433 
2434 	/*
2435 	 * Post an asynchronous event command to catch errors.
2436 	 * We assume the asynchronous events are supported as required by
2437 	 * specification (Figure 40 in section 5 of NVMe 1.2).
2438 	 * However, since at least qemu does not follow the specification,
2439 	 * we need a mechanism to protect ourselves.
2440 	 */
2441 	nvme->n_async_event_supported = B_TRUE;
2442 	nvme_async_event(nvme);
2443 
2444 	/*
2445 	 * Identify Controller
2446 	 */
2447 	if (nvme_identify(nvme, 0, (void **)&nvme->n_idctl) != 0) {
2448 		dev_err(nvme->n_dip, CE_WARN,
2449 		    "!failed to identify controller");
2450 		goto fail;
2451 	}
2452 
2453 	/*
2454 	 * Get Vendor & Product ID
2455 	 */
2456 	bcopy(nvme->n_idctl->id_model, model, sizeof (nvme->n_idctl->id_model));
2457 	model[sizeof (nvme->n_idctl->id_model)] = '\0';
2458 	sata_split_model(model, &vendor, &product);
2459 
2460 	if (vendor == NULL)
2461 		nvme->n_vendor = strdup("NVMe");
2462 	else
2463 		nvme->n_vendor = strdup(vendor);
2464 
2465 	nvme->n_product = strdup(product);
2466 
2467 	/*
2468 	 * Get controller limits.
2469 	 */
2470 	nvme->n_async_event_limit = MAX(NVME_MIN_ASYNC_EVENT_LIMIT,
2471 	    MIN(nvme->n_admin_queue_len / 10,
2472 	    MIN(nvme->n_idctl->id_aerl + 1, nvme->n_async_event_limit)));
2473 
2474 	(void) ddi_prop_update_int(DDI_DEV_T_NONE, nvme->n_dip,
2475 	    "async-event-limit", nvme->n_async_event_limit);
2476 
2477 	nvme->n_abort_command_limit = nvme->n_idctl->id_acl + 1;
2478 
2479 	/*
2480 	 * Reinitialize the semaphore with the true abort command limit
2481 	 * supported by the hardware. It's not necessary to disable interrupts
2482 	 * as only command aborts use the semaphore, and no commands are
2483 	 * executed or aborted while we're here.
2484 	 */
2485 	sema_destroy(&nvme->n_abort_sema);
2486 	sema_init(&nvme->n_abort_sema, nvme->n_abort_command_limit - 1, NULL,
2487 	    SEMA_DRIVER, NULL);
2488 
2489 	nvme->n_progress |= NVME_CTRL_LIMITS;
2490 
2491 	if (nvme->n_idctl->id_mdts == 0)
2492 		nvme->n_max_data_transfer_size = nvme->n_pagesize * 65536;
2493 	else
2494 		nvme->n_max_data_transfer_size =
2495 		    1ull << (nvme->n_pageshift + nvme->n_idctl->id_mdts);
2496 
2497 	nvme->n_error_log_len = nvme->n_idctl->id_elpe + 1;
2498 
2499 	/*
2500 	 * Limit n_max_data_transfer_size to what we can handle in one PRP.
2501 	 * Chained PRPs are currently unsupported.
2502 	 *
2503 	 * This is a no-op on hardware which doesn't support a transfer size
2504 	 * big enough to require chained PRPs.
2505 	 */
2506 	nvme->n_max_data_transfer_size = MIN(nvme->n_max_data_transfer_size,
2507 	    (nvme->n_pagesize / sizeof (uint64_t) * nvme->n_pagesize));
2508 
2509 	nvme->n_prp_dma_attr.dma_attr_maxxfer = nvme->n_max_data_transfer_size;
2510 
2511 	/*
2512 	 * Make sure the minimum/maximum queue entry sizes are not
2513 	 * larger/smaller than the default.
2514 	 */
2515 
2516 	if (((1 << nvme->n_idctl->id_sqes.qes_min) > sizeof (nvme_sqe_t)) ||
2517 	    ((1 << nvme->n_idctl->id_sqes.qes_max) < sizeof (nvme_sqe_t)) ||
2518 	    ((1 << nvme->n_idctl->id_cqes.qes_min) > sizeof (nvme_cqe_t)) ||
2519 	    ((1 << nvme->n_idctl->id_cqes.qes_max) < sizeof (nvme_cqe_t)))
2520 		goto fail;
2521 
2522 	/*
2523 	 * Check for the presence of a Volatile Write Cache. If present,
2524 	 * enable or disable based on the value of the property
2525 	 * volatile-write-cache-enable (default is enabled).
2526 	 */
2527 	nvme->n_write_cache_present =
2528 	    nvme->n_idctl->id_vwc.vwc_present == 0 ? B_FALSE : B_TRUE;
2529 
2530 	(void) ddi_prop_update_int(DDI_DEV_T_NONE, nvme->n_dip,
2531 	    "volatile-write-cache-present",
2532 	    nvme->n_write_cache_present ? 1 : 0);
2533 
2534 	if (!nvme->n_write_cache_present) {
2535 		nvme->n_write_cache_enabled = B_FALSE;
2536 	} else if (nvme_write_cache_set(nvme, nvme->n_write_cache_enabled)
2537 	    != 0) {
2538 		dev_err(nvme->n_dip, CE_WARN,
2539 		    "!failed to %sable volatile write cache",
2540 		    nvme->n_write_cache_enabled ? "en" : "dis");
2541 		/*
2542 		 * Assume the cache is (still) enabled.
2543 		 */
2544 		nvme->n_write_cache_enabled = B_TRUE;
2545 	}
2546 
2547 	(void) ddi_prop_update_int(DDI_DEV_T_NONE, nvme->n_dip,
2548 	    "volatile-write-cache-enable",
2549 	    nvme->n_write_cache_enabled ? 1 : 0);
2550 
2551 	/*
2552 	 * Assume LBA Range Type feature is supported. If it isn't this
2553 	 * will be set to B_FALSE by nvme_get_features().
2554 	 */
2555 	nvme->n_lba_range_supported = B_TRUE;
2556 
2557 	/*
2558 	 * Check support for Autonomous Power State Transition.
2559 	 */
2560 	if (NVME_VERSION_ATLEAST(&nvme->n_version, 1, 1))
2561 		nvme->n_auto_pst_supported =
2562 		    nvme->n_idctl->id_apsta.ap_sup == 0 ? B_FALSE : B_TRUE;
2563 
2564 	/*
2565 	 * Assume Software Progress Marker feature is supported.  If it isn't
2566 	 * this will be set to B_FALSE by nvme_get_features().
2567 	 */
2568 	nvme->n_progress_supported = B_TRUE;
2569 
2570 	/*
2571 	 * Identify Namespaces
2572 	 */
2573 	nvme->n_namespace_count = nvme->n_idctl->id_nn;
2574 
2575 	if (nvme->n_namespace_count == 0) {
2576 		dev_err(nvme->n_dip, CE_WARN,
2577 		    "!controllers without namespaces are not supported");
2578 		goto fail;
2579 	}
2580 
2581 	if (nvme->n_namespace_count > NVME_MINOR_MAX) {
2582 		dev_err(nvme->n_dip, CE_WARN,
2583 		    "!too many namespaces: %d, limiting to %d\n",
2584 		    nvme->n_namespace_count, NVME_MINOR_MAX);
2585 		nvme->n_namespace_count = NVME_MINOR_MAX;
2586 	}
2587 
2588 	nvme->n_ns = kmem_zalloc(sizeof (nvme_namespace_t) *
2589 	    nvme->n_namespace_count, KM_SLEEP);
2590 
2591 	for (i = 0; i != nvme->n_namespace_count; i++) {
2592 		mutex_init(&nvme->n_ns[i].ns_minor.nm_mutex, NULL, MUTEX_DRIVER,
2593 		    NULL);
2594 		if (nvme_init_ns(nvme, i + 1) != DDI_SUCCESS)
2595 			goto fail;
2596 	}
2597 
2598 	/*
2599 	 * Try to set up MSI/MSI-X interrupts.
2600 	 */
2601 	if ((nvme->n_intr_types & (DDI_INTR_TYPE_MSI | DDI_INTR_TYPE_MSIX))
2602 	    != 0) {
2603 		nvme_release_interrupts(nvme);
2604 
2605 		nqueues = MIN(UINT16_MAX, ncpus);
2606 
2607 		if ((nvme_setup_interrupts(nvme, DDI_INTR_TYPE_MSIX,
2608 		    nqueues) != DDI_SUCCESS) &&
2609 		    (nvme_setup_interrupts(nvme, DDI_INTR_TYPE_MSI,
2610 		    nqueues) != DDI_SUCCESS)) {
2611 			dev_err(nvme->n_dip, CE_WARN,
2612 			    "!failed to setup MSI/MSI-X interrupts");
2613 			goto fail;
2614 		}
2615 	}
2616 
2617 	nqueues = nvme->n_intr_cnt;
2618 
2619 	/*
2620 	 * Create I/O queue pairs.
2621 	 */
2622 
2623 	if (nvme_set_nqueues(nvme, &nqueues) != 0) {
2624 		dev_err(nvme->n_dip, CE_WARN,
2625 		    "!failed to set number of I/O queues to %d",
2626 		    nvme->n_intr_cnt);
2627 		goto fail;
2628 	}
2629 
2630 	/*
2631 	 * Reallocate I/O queue array
2632 	 */
2633 	kmem_free(nvme->n_ioq, sizeof (nvme_qpair_t *));
2634 	nvme->n_ioq = kmem_zalloc(sizeof (nvme_qpair_t *) *
2635 	    (nqueues + 1), KM_SLEEP);
2636 	nvme->n_ioq[0] = nvme->n_adminq;
2637 
2638 	nvme->n_ioq_count = nqueues;
2639 
2640 	/*
2641 	 * If we got less queues than we asked for we might as well give
2642 	 * some of the interrupt vectors back to the system.
2643 	 */
2644 	if (nvme->n_ioq_count < nvme->n_intr_cnt) {
2645 		nvme_release_interrupts(nvme);
2646 
2647 		if (nvme_setup_interrupts(nvme, nvme->n_intr_type,
2648 		    nvme->n_ioq_count) != DDI_SUCCESS) {
2649 			dev_err(nvme->n_dip, CE_WARN,
2650 			    "!failed to reduce number of interrupts");
2651 			goto fail;
2652 		}
2653 	}
2654 
2655 	/*
2656 	 * Alloc & register I/O queue pairs
2657 	 */
2658 	nvme->n_io_queue_len =
2659 	    MIN(nvme->n_io_queue_len, nvme->n_max_queue_entries);
2660 	(void) ddi_prop_update_int(DDI_DEV_T_NONE, nvme->n_dip, "io-queue-len",
2661 	    nvme->n_io_queue_len);
2662 
2663 	for (i = 1; i != nvme->n_ioq_count + 1; i++) {
2664 		if (nvme_alloc_qpair(nvme, nvme->n_io_queue_len,
2665 		    &nvme->n_ioq[i], i) != DDI_SUCCESS) {
2666 			dev_err(nvme->n_dip, CE_WARN,
2667 			    "!unable to allocate I/O qpair %d", i);
2668 			goto fail;
2669 		}
2670 
2671 		if (nvme_create_io_qpair(nvme, nvme->n_ioq[i], i) != 0) {
2672 			dev_err(nvme->n_dip, CE_WARN,
2673 			    "!unable to create I/O qpair %d", i);
2674 			goto fail;
2675 		}
2676 	}
2677 
2678 	/*
2679 	 * Post more asynchronous events commands to reduce event reporting
2680 	 * latency as suggested by the spec.
2681 	 */
2682 	if (nvme->n_async_event_supported) {
2683 		for (i = 1; i != nvme->n_async_event_limit; i++)
2684 			nvme_async_event(nvme);
2685 	}
2686 
2687 	return (DDI_SUCCESS);
2688 
2689 fail:
2690 	(void) nvme_reset(nvme, B_FALSE);
2691 	return (DDI_FAILURE);
2692 }
2693 
2694 static uint_t
2695 nvme_intr(caddr_t arg1, caddr_t arg2)
2696 {
2697 	/*LINTED: E_PTR_BAD_CAST_ALIGN*/
2698 	nvme_t *nvme = (nvme_t *)arg1;
2699 	int inum = (int)(uintptr_t)arg2;
2700 	int ccnt = 0;
2701 	int qnum;
2702 	nvme_cmd_t *cmd;
2703 
2704 	if (inum >= nvme->n_intr_cnt)
2705 		return (DDI_INTR_UNCLAIMED);
2706 
2707 	if (nvme->n_dead)
2708 		return (nvme->n_intr_type == DDI_INTR_TYPE_FIXED ?
2709 		    DDI_INTR_UNCLAIMED : DDI_INTR_CLAIMED);
2710 
2711 	/*
2712 	 * The interrupt vector a queue uses is calculated as queue_idx %
2713 	 * intr_cnt in nvme_create_io_qpair(). Iterate through the queue array
2714 	 * in steps of n_intr_cnt to process all queues using this vector.
2715 	 */
2716 	for (qnum = inum;
2717 	    qnum < nvme->n_ioq_count + 1 && nvme->n_ioq[qnum] != NULL;
2718 	    qnum += nvme->n_intr_cnt) {
2719 		while ((cmd = nvme_retrieve_cmd(nvme, nvme->n_ioq[qnum]))) {
2720 			taskq_dispatch_ent((taskq_t *)cmd->nc_nvme->n_cmd_taskq,
2721 			    cmd->nc_callback, cmd, TQ_NOSLEEP, &cmd->nc_tqent);
2722 			ccnt++;
2723 		}
2724 	}
2725 
2726 	return (ccnt > 0 ? DDI_INTR_CLAIMED : DDI_INTR_UNCLAIMED);
2727 }
2728 
2729 static void
2730 nvme_release_interrupts(nvme_t *nvme)
2731 {
2732 	int i;
2733 
2734 	for (i = 0; i < nvme->n_intr_cnt; i++) {
2735 		if (nvme->n_inth[i] == NULL)
2736 			break;
2737 
2738 		if (nvme->n_intr_cap & DDI_INTR_FLAG_BLOCK)
2739 			(void) ddi_intr_block_disable(&nvme->n_inth[i], 1);
2740 		else
2741 			(void) ddi_intr_disable(nvme->n_inth[i]);
2742 
2743 		(void) ddi_intr_remove_handler(nvme->n_inth[i]);
2744 		(void) ddi_intr_free(nvme->n_inth[i]);
2745 	}
2746 
2747 	kmem_free(nvme->n_inth, nvme->n_inth_sz);
2748 	nvme->n_inth = NULL;
2749 	nvme->n_inth_sz = 0;
2750 
2751 	nvme->n_progress &= ~NVME_INTERRUPTS;
2752 }
2753 
2754 static int
2755 nvme_setup_interrupts(nvme_t *nvme, int intr_type, int nqpairs)
2756 {
2757 	int nintrs, navail, count;
2758 	int ret;
2759 	int i;
2760 
2761 	if (nvme->n_intr_types == 0) {
2762 		ret = ddi_intr_get_supported_types(nvme->n_dip,
2763 		    &nvme->n_intr_types);
2764 		if (ret != DDI_SUCCESS) {
2765 			dev_err(nvme->n_dip, CE_WARN,
2766 			    "!%s: ddi_intr_get_supported types failed",
2767 			    __func__);
2768 			return (ret);
2769 		}
2770 #ifdef __x86
2771 		if (get_hwenv() == HW_VMWARE)
2772 			nvme->n_intr_types &= ~DDI_INTR_TYPE_MSIX;
2773 #endif
2774 	}
2775 
2776 	if ((nvme->n_intr_types & intr_type) == 0)
2777 		return (DDI_FAILURE);
2778 
2779 	ret = ddi_intr_get_nintrs(nvme->n_dip, intr_type, &nintrs);
2780 	if (ret != DDI_SUCCESS) {
2781 		dev_err(nvme->n_dip, CE_WARN, "!%s: ddi_intr_get_nintrs failed",
2782 		    __func__);
2783 		return (ret);
2784 	}
2785 
2786 	ret = ddi_intr_get_navail(nvme->n_dip, intr_type, &navail);
2787 	if (ret != DDI_SUCCESS) {
2788 		dev_err(nvme->n_dip, CE_WARN, "!%s: ddi_intr_get_navail failed",
2789 		    __func__);
2790 		return (ret);
2791 	}
2792 
2793 	/* We want at most one interrupt per queue pair. */
2794 	if (navail > nqpairs)
2795 		navail = nqpairs;
2796 
2797 	nvme->n_inth_sz = sizeof (ddi_intr_handle_t) * navail;
2798 	nvme->n_inth = kmem_zalloc(nvme->n_inth_sz, KM_SLEEP);
2799 
2800 	ret = ddi_intr_alloc(nvme->n_dip, nvme->n_inth, intr_type, 0, navail,
2801 	    &count, 0);
2802 	if (ret != DDI_SUCCESS) {
2803 		dev_err(nvme->n_dip, CE_WARN, "!%s: ddi_intr_alloc failed",
2804 		    __func__);
2805 		goto fail;
2806 	}
2807 
2808 	nvme->n_intr_cnt = count;
2809 
2810 	ret = ddi_intr_get_pri(nvme->n_inth[0], &nvme->n_intr_pri);
2811 	if (ret != DDI_SUCCESS) {
2812 		dev_err(nvme->n_dip, CE_WARN, "!%s: ddi_intr_get_pri failed",
2813 		    __func__);
2814 		goto fail;
2815 	}
2816 
2817 	for (i = 0; i < count; i++) {
2818 		ret = ddi_intr_add_handler(nvme->n_inth[i], nvme_intr,
2819 		    (void *)nvme, (void *)(uintptr_t)i);
2820 		if (ret != DDI_SUCCESS) {
2821 			dev_err(nvme->n_dip, CE_WARN,
2822 			    "!%s: ddi_intr_add_handler failed", __func__);
2823 			goto fail;
2824 		}
2825 	}
2826 
2827 	(void) ddi_intr_get_cap(nvme->n_inth[0], &nvme->n_intr_cap);
2828 
2829 	for (i = 0; i < count; i++) {
2830 		if (nvme->n_intr_cap & DDI_INTR_FLAG_BLOCK)
2831 			ret = ddi_intr_block_enable(&nvme->n_inth[i], 1);
2832 		else
2833 			ret = ddi_intr_enable(nvme->n_inth[i]);
2834 
2835 		if (ret != DDI_SUCCESS) {
2836 			dev_err(nvme->n_dip, CE_WARN,
2837 			    "!%s: enabling interrupt %d failed", __func__, i);
2838 			goto fail;
2839 		}
2840 	}
2841 
2842 	nvme->n_intr_type = intr_type;
2843 
2844 	nvme->n_progress |= NVME_INTERRUPTS;
2845 
2846 	return (DDI_SUCCESS);
2847 
2848 fail:
2849 	nvme_release_interrupts(nvme);
2850 
2851 	return (ret);
2852 }
2853 
2854 static int
2855 nvme_fm_errcb(dev_info_t *dip, ddi_fm_error_t *fm_error, const void *arg)
2856 {
2857 	_NOTE(ARGUNUSED(arg));
2858 
2859 	pci_ereport_post(dip, fm_error, NULL);
2860 	return (fm_error->fme_status);
2861 }
2862 
2863 static int
2864 nvme_attach(dev_info_t *dip, ddi_attach_cmd_t cmd)
2865 {
2866 	nvme_t *nvme;
2867 	int instance;
2868 	int nregs;
2869 	off_t regsize;
2870 	int i;
2871 	char name[32];
2872 
2873 	if (cmd != DDI_ATTACH)
2874 		return (DDI_FAILURE);
2875 
2876 	instance = ddi_get_instance(dip);
2877 
2878 	if (ddi_soft_state_zalloc(nvme_state, instance) != DDI_SUCCESS)
2879 		return (DDI_FAILURE);
2880 
2881 	nvme = ddi_get_soft_state(nvme_state, instance);
2882 	ddi_set_driver_private(dip, nvme);
2883 	nvme->n_dip = dip;
2884 
2885 	mutex_init(&nvme->n_minor.nm_mutex, NULL, MUTEX_DRIVER, NULL);
2886 
2887 	nvme->n_strict_version = ddi_prop_get_int(DDI_DEV_T_ANY, dip,
2888 	    DDI_PROP_DONTPASS, "strict-version", 1) == 1 ? B_TRUE : B_FALSE;
2889 	nvme->n_ignore_unknown_vendor_status = ddi_prop_get_int(DDI_DEV_T_ANY,
2890 	    dip, DDI_PROP_DONTPASS, "ignore-unknown-vendor-status", 0) == 1 ?
2891 	    B_TRUE : B_FALSE;
2892 	nvme->n_admin_queue_len = ddi_prop_get_int(DDI_DEV_T_ANY, dip,
2893 	    DDI_PROP_DONTPASS, "admin-queue-len", NVME_DEFAULT_ADMIN_QUEUE_LEN);
2894 	nvme->n_io_queue_len = ddi_prop_get_int(DDI_DEV_T_ANY, dip,
2895 	    DDI_PROP_DONTPASS, "io-queue-len", NVME_DEFAULT_IO_QUEUE_LEN);
2896 	nvme->n_async_event_limit = ddi_prop_get_int(DDI_DEV_T_ANY, dip,
2897 	    DDI_PROP_DONTPASS, "async-event-limit",
2898 	    NVME_DEFAULT_ASYNC_EVENT_LIMIT);
2899 	nvme->n_write_cache_enabled = ddi_prop_get_int(DDI_DEV_T_ANY, dip,
2900 	    DDI_PROP_DONTPASS, "volatile-write-cache-enable", 1) != 0 ?
2901 	    B_TRUE : B_FALSE;
2902 	nvme->n_min_block_size = ddi_prop_get_int(DDI_DEV_T_ANY, dip,
2903 	    DDI_PROP_DONTPASS, "min-phys-block-size",
2904 	    NVME_DEFAULT_MIN_BLOCK_SIZE);
2905 
2906 	if (!ISP2(nvme->n_min_block_size) ||
2907 	    (nvme->n_min_block_size < NVME_DEFAULT_MIN_BLOCK_SIZE)) {
2908 		dev_err(dip, CE_WARN, "!min-phys-block-size %s, "
2909 		    "using default %d", ISP2(nvme->n_min_block_size) ?
2910 		    "too low" : "not a power of 2",
2911 		    NVME_DEFAULT_MIN_BLOCK_SIZE);
2912 		nvme->n_min_block_size = NVME_DEFAULT_MIN_BLOCK_SIZE;
2913 	}
2914 
2915 	if (nvme->n_admin_queue_len < NVME_MIN_ADMIN_QUEUE_LEN)
2916 		nvme->n_admin_queue_len = NVME_MIN_ADMIN_QUEUE_LEN;
2917 	else if (nvme->n_admin_queue_len > NVME_MAX_ADMIN_QUEUE_LEN)
2918 		nvme->n_admin_queue_len = NVME_MAX_ADMIN_QUEUE_LEN;
2919 
2920 	if (nvme->n_io_queue_len < NVME_MIN_IO_QUEUE_LEN)
2921 		nvme->n_io_queue_len = NVME_MIN_IO_QUEUE_LEN;
2922 
2923 	if (nvme->n_async_event_limit < 1)
2924 		nvme->n_async_event_limit = NVME_DEFAULT_ASYNC_EVENT_LIMIT;
2925 
2926 	nvme->n_reg_acc_attr = nvme_reg_acc_attr;
2927 	nvme->n_queue_dma_attr = nvme_queue_dma_attr;
2928 	nvme->n_prp_dma_attr = nvme_prp_dma_attr;
2929 	nvme->n_sgl_dma_attr = nvme_sgl_dma_attr;
2930 
2931 	/*
2932 	 * Setup FMA support.
2933 	 */
2934 	nvme->n_fm_cap = ddi_getprop(DDI_DEV_T_ANY, dip,
2935 	    DDI_PROP_CANSLEEP | DDI_PROP_DONTPASS, "fm-capable",
2936 	    DDI_FM_EREPORT_CAPABLE | DDI_FM_ACCCHK_CAPABLE |
2937 	    DDI_FM_DMACHK_CAPABLE | DDI_FM_ERRCB_CAPABLE);
2938 
2939 	ddi_fm_init(dip, &nvme->n_fm_cap, &nvme->n_fm_ibc);
2940 
2941 	if (nvme->n_fm_cap) {
2942 		if (nvme->n_fm_cap & DDI_FM_ACCCHK_CAPABLE)
2943 			nvme->n_reg_acc_attr.devacc_attr_access =
2944 			    DDI_FLAGERR_ACC;
2945 
2946 		if (nvme->n_fm_cap & DDI_FM_DMACHK_CAPABLE) {
2947 			nvme->n_prp_dma_attr.dma_attr_flags |= DDI_DMA_FLAGERR;
2948 			nvme->n_sgl_dma_attr.dma_attr_flags |= DDI_DMA_FLAGERR;
2949 		}
2950 
2951 		if (DDI_FM_EREPORT_CAP(nvme->n_fm_cap) ||
2952 		    DDI_FM_ERRCB_CAP(nvme->n_fm_cap))
2953 			pci_ereport_setup(dip);
2954 
2955 		if (DDI_FM_ERRCB_CAP(nvme->n_fm_cap))
2956 			ddi_fm_handler_register(dip, nvme_fm_errcb,
2957 			    (void *)nvme);
2958 	}
2959 
2960 	nvme->n_progress |= NVME_FMA_INIT;
2961 
2962 	/*
2963 	 * The spec defines several register sets. Only the controller
2964 	 * registers (set 1) are currently used.
2965 	 */
2966 	if (ddi_dev_nregs(dip, &nregs) == DDI_FAILURE ||
2967 	    nregs < 2 ||
2968 	    ddi_dev_regsize(dip, 1, &regsize) == DDI_FAILURE)
2969 		goto fail;
2970 
2971 	if (ddi_regs_map_setup(dip, 1, &nvme->n_regs, 0, regsize,
2972 	    &nvme->n_reg_acc_attr, &nvme->n_regh) != DDI_SUCCESS) {
2973 		dev_err(dip, CE_WARN, "!failed to map regset 1");
2974 		goto fail;
2975 	}
2976 
2977 	nvme->n_progress |= NVME_REGS_MAPPED;
2978 
2979 	/*
2980 	 * Create taskq for command completion.
2981 	 */
2982 	(void) snprintf(name, sizeof (name), "%s%d_cmd_taskq",
2983 	    ddi_driver_name(dip), ddi_get_instance(dip));
2984 	nvme->n_cmd_taskq = ddi_taskq_create(dip, name, MIN(UINT16_MAX, ncpus),
2985 	    TASKQ_DEFAULTPRI, 0);
2986 	if (nvme->n_cmd_taskq == NULL) {
2987 		dev_err(dip, CE_WARN, "!failed to create cmd taskq");
2988 		goto fail;
2989 	}
2990 
2991 	/*
2992 	 * Create PRP DMA cache
2993 	 */
2994 	(void) snprintf(name, sizeof (name), "%s%d_prp_cache",
2995 	    ddi_driver_name(dip), ddi_get_instance(dip));
2996 	nvme->n_prp_cache = kmem_cache_create(name, sizeof (nvme_dma_t),
2997 	    0, nvme_prp_dma_constructor, nvme_prp_dma_destructor,
2998 	    NULL, (void *)nvme, NULL, 0);
2999 
3000 	if (nvme_init(nvme) != DDI_SUCCESS)
3001 		goto fail;
3002 
3003 	/*
3004 	 * Attach the blkdev driver for each namespace.
3005 	 */
3006 	for (i = 0; i != nvme->n_namespace_count; i++) {
3007 		if (ddi_create_minor_node(nvme->n_dip, nvme->n_ns[i].ns_name,
3008 		    S_IFCHR, NVME_MINOR(ddi_get_instance(nvme->n_dip), i + 1),
3009 		    DDI_NT_NVME_ATTACHMENT_POINT, 0) != DDI_SUCCESS) {
3010 			dev_err(dip, CE_WARN,
3011 			    "!failed to create minor node for namespace %d", i);
3012 			goto fail;
3013 		}
3014 
3015 		if (nvme->n_ns[i].ns_ignore)
3016 			continue;
3017 
3018 		nvme->n_ns[i].ns_bd_hdl = bd_alloc_handle(&nvme->n_ns[i],
3019 		    &nvme_bd_ops, &nvme->n_prp_dma_attr, KM_SLEEP);
3020 
3021 		if (nvme->n_ns[i].ns_bd_hdl == NULL) {
3022 			dev_err(dip, CE_WARN,
3023 			    "!failed to get blkdev handle for namespace %d", i);
3024 			goto fail;
3025 		}
3026 
3027 		if (bd_attach_handle(dip, nvme->n_ns[i].ns_bd_hdl)
3028 		    != DDI_SUCCESS) {
3029 			dev_err(dip, CE_WARN,
3030 			    "!failed to attach blkdev handle for namespace %d",
3031 			    i);
3032 			goto fail;
3033 		}
3034 	}
3035 
3036 	if (ddi_create_minor_node(dip, "devctl", S_IFCHR,
3037 	    NVME_MINOR(ddi_get_instance(dip), 0), DDI_NT_NVME_NEXUS, 0)
3038 	    != DDI_SUCCESS) {
3039 		dev_err(dip, CE_WARN, "nvme_attach: "
3040 		    "cannot create devctl minor node");
3041 		goto fail;
3042 	}
3043 
3044 	return (DDI_SUCCESS);
3045 
3046 fail:
3047 	/* attach successful anyway so that FMA can retire the device */
3048 	if (nvme->n_dead)
3049 		return (DDI_SUCCESS);
3050 
3051 	(void) nvme_detach(dip, DDI_DETACH);
3052 
3053 	return (DDI_FAILURE);
3054 }
3055 
3056 static int
3057 nvme_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
3058 {
3059 	int instance, i;
3060 	nvme_t *nvme;
3061 
3062 	if (cmd != DDI_DETACH)
3063 		return (DDI_FAILURE);
3064 
3065 	instance = ddi_get_instance(dip);
3066 
3067 	nvme = ddi_get_soft_state(nvme_state, instance);
3068 
3069 	if (nvme == NULL)
3070 		return (DDI_FAILURE);
3071 
3072 	ddi_remove_minor_node(dip, "devctl");
3073 	mutex_destroy(&nvme->n_minor.nm_mutex);
3074 
3075 	if (nvme->n_ns) {
3076 		for (i = 0; i != nvme->n_namespace_count; i++) {
3077 			ddi_remove_minor_node(dip, nvme->n_ns[i].ns_name);
3078 			mutex_destroy(&nvme->n_ns[i].ns_minor.nm_mutex);
3079 
3080 			if (nvme->n_ns[i].ns_bd_hdl) {
3081 				(void) bd_detach_handle(
3082 				    nvme->n_ns[i].ns_bd_hdl);
3083 				bd_free_handle(nvme->n_ns[i].ns_bd_hdl);
3084 			}
3085 
3086 			if (nvme->n_ns[i].ns_idns)
3087 				kmem_free(nvme->n_ns[i].ns_idns,
3088 				    sizeof (nvme_identify_nsid_t));
3089 			if (nvme->n_ns[i].ns_devid)
3090 				strfree(nvme->n_ns[i].ns_devid);
3091 		}
3092 
3093 		kmem_free(nvme->n_ns, sizeof (nvme_namespace_t) *
3094 		    nvme->n_namespace_count);
3095 	}
3096 
3097 	if (nvme->n_progress & NVME_INTERRUPTS)
3098 		nvme_release_interrupts(nvme);
3099 
3100 	if (nvme->n_cmd_taskq)
3101 		ddi_taskq_wait(nvme->n_cmd_taskq);
3102 
3103 	if (nvme->n_ioq_count > 0) {
3104 		for (i = 1; i != nvme->n_ioq_count + 1; i++) {
3105 			if (nvme->n_ioq[i] != NULL) {
3106 				/* TODO: send destroy queue commands */
3107 				nvme_free_qpair(nvme->n_ioq[i]);
3108 			}
3109 		}
3110 
3111 		kmem_free(nvme->n_ioq, sizeof (nvme_qpair_t *) *
3112 		    (nvme->n_ioq_count + 1));
3113 	}
3114 
3115 	if (nvme->n_prp_cache != NULL) {
3116 		kmem_cache_destroy(nvme->n_prp_cache);
3117 	}
3118 
3119 	if (nvme->n_progress & NVME_REGS_MAPPED) {
3120 		nvme_shutdown(nvme, NVME_CC_SHN_NORMAL, B_FALSE);
3121 		(void) nvme_reset(nvme, B_FALSE);
3122 	}
3123 
3124 	if (nvme->n_cmd_taskq)
3125 		ddi_taskq_destroy(nvme->n_cmd_taskq);
3126 
3127 	if (nvme->n_progress & NVME_CTRL_LIMITS)
3128 		sema_destroy(&nvme->n_abort_sema);
3129 
3130 	if (nvme->n_progress & NVME_ADMIN_QUEUE)
3131 		nvme_free_qpair(nvme->n_adminq);
3132 
3133 	if (nvme->n_idctl)
3134 		kmem_free(nvme->n_idctl, NVME_IDENTIFY_BUFSIZE);
3135 
3136 	if (nvme->n_progress & NVME_REGS_MAPPED)
3137 		ddi_regs_map_free(&nvme->n_regh);
3138 
3139 	if (nvme->n_progress & NVME_FMA_INIT) {
3140 		if (DDI_FM_ERRCB_CAP(nvme->n_fm_cap))
3141 			ddi_fm_handler_unregister(nvme->n_dip);
3142 
3143 		if (DDI_FM_EREPORT_CAP(nvme->n_fm_cap) ||
3144 		    DDI_FM_ERRCB_CAP(nvme->n_fm_cap))
3145 			pci_ereport_teardown(nvme->n_dip);
3146 
3147 		ddi_fm_fini(nvme->n_dip);
3148 	}
3149 
3150 	if (nvme->n_vendor != NULL)
3151 		strfree(nvme->n_vendor);
3152 
3153 	if (nvme->n_product != NULL)
3154 		strfree(nvme->n_product);
3155 
3156 	ddi_soft_state_free(nvme_state, instance);
3157 
3158 	return (DDI_SUCCESS);
3159 }
3160 
3161 static int
3162 nvme_quiesce(dev_info_t *dip)
3163 {
3164 	int instance;
3165 	nvme_t *nvme;
3166 
3167 	instance = ddi_get_instance(dip);
3168 
3169 	nvme = ddi_get_soft_state(nvme_state, instance);
3170 
3171 	if (nvme == NULL)
3172 		return (DDI_FAILURE);
3173 
3174 	nvme_shutdown(nvme, NVME_CC_SHN_ABRUPT, B_TRUE);
3175 
3176 	(void) nvme_reset(nvme, B_TRUE);
3177 
3178 	return (DDI_FAILURE);
3179 }
3180 
3181 static int
3182 nvme_fill_prp(nvme_cmd_t *cmd, bd_xfer_t *xfer)
3183 {
3184 	nvme_t *nvme = cmd->nc_nvme;
3185 	int nprp_page, nprp;
3186 	uint64_t *prp;
3187 
3188 	if (xfer->x_ndmac == 0)
3189 		return (DDI_FAILURE);
3190 
3191 	cmd->nc_sqe.sqe_dptr.d_prp[0] = xfer->x_dmac.dmac_laddress;
3192 	ddi_dma_nextcookie(xfer->x_dmah, &xfer->x_dmac);
3193 
3194 	if (xfer->x_ndmac == 1) {
3195 		cmd->nc_sqe.sqe_dptr.d_prp[1] = 0;
3196 		return (DDI_SUCCESS);
3197 	} else if (xfer->x_ndmac == 2) {
3198 		cmd->nc_sqe.sqe_dptr.d_prp[1] = xfer->x_dmac.dmac_laddress;
3199 		return (DDI_SUCCESS);
3200 	}
3201 
3202 	xfer->x_ndmac--;
3203 
3204 	nprp_page = nvme->n_pagesize / sizeof (uint64_t) - 1;
3205 	ASSERT(nprp_page > 0);
3206 	nprp = (xfer->x_ndmac + nprp_page - 1) / nprp_page;
3207 
3208 	/*
3209 	 * We currently don't support chained PRPs and set up our DMA
3210 	 * attributes to reflect that. If we still get an I/O request
3211 	 * that needs a chained PRP something is very wrong.
3212 	 */
3213 	VERIFY(nprp == 1);
3214 
3215 	cmd->nc_dma = kmem_cache_alloc(nvme->n_prp_cache, KM_SLEEP);
3216 	bzero(cmd->nc_dma->nd_memp, cmd->nc_dma->nd_len);
3217 
3218 	cmd->nc_sqe.sqe_dptr.d_prp[1] = cmd->nc_dma->nd_cookie.dmac_laddress;
3219 
3220 	/*LINTED: E_PTR_BAD_CAST_ALIGN*/
3221 	for (prp = (uint64_t *)cmd->nc_dma->nd_memp;
3222 	    xfer->x_ndmac > 0;
3223 	    prp++, xfer->x_ndmac--) {
3224 		*prp = xfer->x_dmac.dmac_laddress;
3225 		ddi_dma_nextcookie(xfer->x_dmah, &xfer->x_dmac);
3226 	}
3227 
3228 	(void) ddi_dma_sync(cmd->nc_dma->nd_dmah, 0, cmd->nc_dma->nd_len,
3229 	    DDI_DMA_SYNC_FORDEV);
3230 	return (DDI_SUCCESS);
3231 }
3232 
3233 static nvme_cmd_t *
3234 nvme_create_nvm_cmd(nvme_namespace_t *ns, uint8_t opc, bd_xfer_t *xfer)
3235 {
3236 	nvme_t *nvme = ns->ns_nvme;
3237 	nvme_cmd_t *cmd;
3238 
3239 	/*
3240 	 * Blkdev only sets BD_XFER_POLL when dumping, so don't sleep.
3241 	 */
3242 	cmd = nvme_alloc_cmd(nvme, (xfer->x_flags & BD_XFER_POLL) ?
3243 	    KM_NOSLEEP : KM_SLEEP);
3244 
3245 	if (cmd == NULL)
3246 		return (NULL);
3247 
3248 	cmd->nc_sqe.sqe_opc = opc;
3249 	cmd->nc_callback = nvme_bd_xfer_done;
3250 	cmd->nc_xfer = xfer;
3251 
3252 	switch (opc) {
3253 	case NVME_OPC_NVM_WRITE:
3254 	case NVME_OPC_NVM_READ:
3255 		VERIFY(xfer->x_nblks <= 0x10000);
3256 
3257 		cmd->nc_sqe.sqe_nsid = ns->ns_id;
3258 
3259 		cmd->nc_sqe.sqe_cdw10 = xfer->x_blkno & 0xffffffffu;
3260 		cmd->nc_sqe.sqe_cdw11 = (xfer->x_blkno >> 32);
3261 		cmd->nc_sqe.sqe_cdw12 = (uint16_t)(xfer->x_nblks - 1);
3262 
3263 		if (nvme_fill_prp(cmd, xfer) != DDI_SUCCESS)
3264 			goto fail;
3265 		break;
3266 
3267 	case NVME_OPC_NVM_FLUSH:
3268 		cmd->nc_sqe.sqe_nsid = ns->ns_id;
3269 		break;
3270 
3271 	default:
3272 		goto fail;
3273 	}
3274 
3275 	return (cmd);
3276 
3277 fail:
3278 	nvme_free_cmd(cmd);
3279 	return (NULL);
3280 }
3281 
3282 static void
3283 nvme_bd_xfer_done(void *arg)
3284 {
3285 	nvme_cmd_t *cmd = arg;
3286 	bd_xfer_t *xfer = cmd->nc_xfer;
3287 	int error = 0;
3288 
3289 	error = nvme_check_cmd_status(cmd);
3290 	nvme_free_cmd(cmd);
3291 
3292 	bd_xfer_done(xfer, error);
3293 }
3294 
3295 static void
3296 nvme_bd_driveinfo(void *arg, bd_drive_t *drive)
3297 {
3298 	nvme_namespace_t *ns = arg;
3299 	nvme_t *nvme = ns->ns_nvme;
3300 
3301 	/*
3302 	 * blkdev maintains one queue size per instance (namespace),
3303 	 * but all namespace share the I/O queues.
3304 	 * TODO: need to figure out a sane default, or use per-NS I/O queues,
3305 	 * or change blkdev to handle EAGAIN
3306 	 */
3307 	drive->d_qsize = nvme->n_ioq_count * nvme->n_io_queue_len
3308 	    / nvme->n_namespace_count;
3309 
3310 	/*
3311 	 * d_maxxfer is not set, which means the value is taken from the DMA
3312 	 * attributes specified to bd_alloc_handle.
3313 	 */
3314 
3315 	drive->d_removable = B_FALSE;
3316 	drive->d_hotpluggable = B_FALSE;
3317 
3318 	bcopy(ns->ns_eui64, drive->d_eui64, sizeof (drive->d_eui64));
3319 	drive->d_target = ns->ns_id;
3320 	drive->d_lun = 0;
3321 
3322 	drive->d_model = nvme->n_idctl->id_model;
3323 	drive->d_model_len = sizeof (nvme->n_idctl->id_model);
3324 	drive->d_vendor = nvme->n_vendor;
3325 	drive->d_vendor_len = strlen(nvme->n_vendor);
3326 	drive->d_product = nvme->n_product;
3327 	drive->d_product_len = strlen(nvme->n_product);
3328 	drive->d_serial = nvme->n_idctl->id_serial;
3329 	drive->d_serial_len = sizeof (nvme->n_idctl->id_serial);
3330 	drive->d_revision = nvme->n_idctl->id_fwrev;
3331 	drive->d_revision_len = sizeof (nvme->n_idctl->id_fwrev);
3332 }
3333 
3334 static int
3335 nvme_bd_mediainfo(void *arg, bd_media_t *media)
3336 {
3337 	nvme_namespace_t *ns = arg;
3338 
3339 	media->m_nblks = ns->ns_block_count;
3340 	media->m_blksize = ns->ns_block_size;
3341 	media->m_readonly = B_FALSE;
3342 	media->m_solidstate = B_TRUE;
3343 
3344 	media->m_pblksize = ns->ns_best_block_size;
3345 
3346 	return (0);
3347 }
3348 
3349 static int
3350 nvme_bd_cmd(nvme_namespace_t *ns, bd_xfer_t *xfer, uint8_t opc)
3351 {
3352 	nvme_t *nvme = ns->ns_nvme;
3353 	nvme_cmd_t *cmd;
3354 	nvme_qpair_t *ioq;
3355 	boolean_t poll;
3356 	int ret;
3357 
3358 	if (nvme->n_dead)
3359 		return (EIO);
3360 
3361 	cmd = nvme_create_nvm_cmd(ns, opc, xfer);
3362 	if (cmd == NULL)
3363 		return (ENOMEM);
3364 
3365 	cmd->nc_sqid = (CPU->cpu_id % nvme->n_ioq_count) + 1;
3366 	ASSERT(cmd->nc_sqid <= nvme->n_ioq_count);
3367 	ioq = nvme->n_ioq[cmd->nc_sqid];
3368 
3369 	/*
3370 	 * Get the polling flag before submitting the command. The command may
3371 	 * complete immediately after it was submitted, which means we must
3372 	 * treat both cmd and xfer as if they have been freed already.
3373 	 */
3374 	poll = (xfer->x_flags & BD_XFER_POLL) != 0;
3375 
3376 	ret = nvme_submit_io_cmd(ioq, cmd);
3377 
3378 	if (ret != 0)
3379 		return (ret);
3380 
3381 	if (!poll)
3382 		return (0);
3383 
3384 	do {
3385 		cmd = nvme_retrieve_cmd(nvme, ioq);
3386 		if (cmd != NULL)
3387 			nvme_bd_xfer_done(cmd);
3388 		else
3389 			drv_usecwait(10);
3390 	} while (ioq->nq_active_cmds != 0);
3391 
3392 	return (0);
3393 }
3394 
3395 static int
3396 nvme_bd_read(void *arg, bd_xfer_t *xfer)
3397 {
3398 	nvme_namespace_t *ns = arg;
3399 
3400 	return (nvme_bd_cmd(ns, xfer, NVME_OPC_NVM_READ));
3401 }
3402 
3403 static int
3404 nvme_bd_write(void *arg, bd_xfer_t *xfer)
3405 {
3406 	nvme_namespace_t *ns = arg;
3407 
3408 	return (nvme_bd_cmd(ns, xfer, NVME_OPC_NVM_WRITE));
3409 }
3410 
3411 static int
3412 nvme_bd_sync(void *arg, bd_xfer_t *xfer)
3413 {
3414 	nvme_namespace_t *ns = arg;
3415 
3416 	if (ns->ns_nvme->n_dead)
3417 		return (EIO);
3418 
3419 	/*
3420 	 * If the volatile write cache is not present or not enabled the FLUSH
3421 	 * command is a no-op, so we can take a shortcut here.
3422 	 */
3423 	if (!ns->ns_nvme->n_write_cache_present) {
3424 		bd_xfer_done(xfer, ENOTSUP);
3425 		return (0);
3426 	}
3427 
3428 	if (!ns->ns_nvme->n_write_cache_enabled) {
3429 		bd_xfer_done(xfer, 0);
3430 		return (0);
3431 	}
3432 
3433 	return (nvme_bd_cmd(ns, xfer, NVME_OPC_NVM_FLUSH));
3434 }
3435 
3436 static int
3437 nvme_bd_devid(void *arg, dev_info_t *devinfo, ddi_devid_t *devid)
3438 {
3439 	nvme_namespace_t *ns = arg;
3440 
3441 	/*LINTED: E_BAD_PTR_CAST_ALIGN*/
3442 	if (*(uint64_t *)ns->ns_eui64 != 0) {
3443 		return (ddi_devid_init(devinfo, DEVID_SCSI3_WWN,
3444 		    sizeof (ns->ns_eui64), ns->ns_eui64, devid));
3445 	} else {
3446 		return (ddi_devid_init(devinfo, DEVID_ENCAP,
3447 		    strlen(ns->ns_devid), ns->ns_devid, devid));
3448 	}
3449 }
3450 
3451 static int
3452 nvme_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
3453 {
3454 #ifndef __lock_lint
3455 	_NOTE(ARGUNUSED(cred_p));
3456 #endif
3457 	minor_t minor = getminor(*devp);
3458 	nvme_t *nvme = ddi_get_soft_state(nvme_state, NVME_MINOR_INST(minor));
3459 	int nsid = NVME_MINOR_NSID(minor);
3460 	nvme_minor_state_t *nm;
3461 	int rv = 0;
3462 
3463 	if (otyp != OTYP_CHR)
3464 		return (EINVAL);
3465 
3466 	if (nvme == NULL)
3467 		return (ENXIO);
3468 
3469 	if (nsid > nvme->n_namespace_count)
3470 		return (ENXIO);
3471 
3472 	if (nvme->n_dead)
3473 		return (EIO);
3474 
3475 	nm = nsid == 0 ? &nvme->n_minor : &nvme->n_ns[nsid - 1].ns_minor;
3476 
3477 	mutex_enter(&nm->nm_mutex);
3478 	if (nm->nm_oexcl) {
3479 		rv = EBUSY;
3480 		goto out;
3481 	}
3482 
3483 	if (flag & FEXCL) {
3484 		if (nm->nm_ocnt != 0) {
3485 			rv = EBUSY;
3486 			goto out;
3487 		}
3488 		nm->nm_oexcl = B_TRUE;
3489 	}
3490 
3491 	nm->nm_ocnt++;
3492 
3493 out:
3494 	mutex_exit(&nm->nm_mutex);
3495 	return (rv);
3496 
3497 }
3498 
3499 static int
3500 nvme_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
3501 {
3502 #ifndef __lock_lint
3503 	_NOTE(ARGUNUSED(cred_p));
3504 	_NOTE(ARGUNUSED(flag));
3505 #endif
3506 	minor_t minor = getminor(dev);
3507 	nvme_t *nvme = ddi_get_soft_state(nvme_state, NVME_MINOR_INST(minor));
3508 	int nsid = NVME_MINOR_NSID(minor);
3509 	nvme_minor_state_t *nm;
3510 
3511 	if (otyp != OTYP_CHR)
3512 		return (ENXIO);
3513 
3514 	if (nvme == NULL)
3515 		return (ENXIO);
3516 
3517 	if (nsid > nvme->n_namespace_count)
3518 		return (ENXIO);
3519 
3520 	nm = nsid == 0 ? &nvme->n_minor : &nvme->n_ns[nsid - 1].ns_minor;
3521 
3522 	mutex_enter(&nm->nm_mutex);
3523 	if (nm->nm_oexcl)
3524 		nm->nm_oexcl = B_FALSE;
3525 
3526 	ASSERT(nm->nm_ocnt > 0);
3527 	nm->nm_ocnt--;
3528 	mutex_exit(&nm->nm_mutex);
3529 
3530 	return (0);
3531 }
3532 
3533 static int
3534 nvme_ioctl_identify(nvme_t *nvme, int nsid, nvme_ioctl_t *nioc, int mode,
3535     cred_t *cred_p)
3536 {
3537 	_NOTE(ARGUNUSED(cred_p));
3538 	int rv = 0;
3539 	void *idctl;
3540 
3541 	if ((mode & FREAD) == 0)
3542 		return (EPERM);
3543 
3544 	if (nioc->n_len < NVME_IDENTIFY_BUFSIZE)
3545 		return (EINVAL);
3546 
3547 	if ((rv = nvme_identify(nvme, nsid, (void **)&idctl)) != 0)
3548 		return (rv);
3549 
3550 	if (ddi_copyout(idctl, (void *)nioc->n_buf, NVME_IDENTIFY_BUFSIZE, mode)
3551 	    != 0)
3552 		rv = EFAULT;
3553 
3554 	kmem_free(idctl, NVME_IDENTIFY_BUFSIZE);
3555 
3556 	return (rv);
3557 }
3558 
3559 static int
3560 nvme_ioctl_capabilities(nvme_t *nvme, int nsid, nvme_ioctl_t *nioc,
3561     int mode, cred_t *cred_p)
3562 {
3563 	_NOTE(ARGUNUSED(nsid, cred_p));
3564 	int rv = 0;
3565 	nvme_reg_cap_t cap = { 0 };
3566 	nvme_capabilities_t nc;
3567 
3568 	if ((mode & FREAD) == 0)
3569 		return (EPERM);
3570 
3571 	if (nioc->n_len < sizeof (nc))
3572 		return (EINVAL);
3573 
3574 	cap.r = nvme_get64(nvme, NVME_REG_CAP);
3575 
3576 	/*
3577 	 * The MPSMIN and MPSMAX fields in the CAP register use 0 to
3578 	 * specify the base page size of 4k (1<<12), so add 12 here to
3579 	 * get the real page size value.
3580 	 */
3581 	nc.mpsmax = 1 << (12 + cap.b.cap_mpsmax);
3582 	nc.mpsmin = 1 << (12 + cap.b.cap_mpsmin);
3583 
3584 	if (ddi_copyout(&nc, (void *)nioc->n_buf, sizeof (nc), mode) != 0)
3585 		rv = EFAULT;
3586 
3587 	return (rv);
3588 }
3589 
3590 static int
3591 nvme_ioctl_get_logpage(nvme_t *nvme, int nsid, nvme_ioctl_t *nioc,
3592     int mode, cred_t *cred_p)
3593 {
3594 	_NOTE(ARGUNUSED(cred_p));
3595 	void *log = NULL;
3596 	size_t bufsize = 0;
3597 	int rv = 0;
3598 
3599 	if ((mode & FREAD) == 0)
3600 		return (EPERM);
3601 
3602 	switch (nioc->n_arg) {
3603 	case NVME_LOGPAGE_ERROR:
3604 		if (nsid != 0)
3605 			return (EINVAL);
3606 		break;
3607 	case NVME_LOGPAGE_HEALTH:
3608 		if (nsid != 0 && nvme->n_idctl->id_lpa.lp_smart == 0)
3609 			return (EINVAL);
3610 
3611 		if (nsid == 0)
3612 			nsid = (uint32_t)-1;
3613 
3614 		break;
3615 	case NVME_LOGPAGE_FWSLOT:
3616 		if (nsid != 0)
3617 			return (EINVAL);
3618 		break;
3619 	default:
3620 		return (EINVAL);
3621 	}
3622 
3623 	if (nvme_get_logpage(nvme, &log, &bufsize, nioc->n_arg, nsid)
3624 	    != DDI_SUCCESS)
3625 		return (EIO);
3626 
3627 	if (nioc->n_len < bufsize) {
3628 		kmem_free(log, bufsize);
3629 		return (EINVAL);
3630 	}
3631 
3632 	if (ddi_copyout(log, (void *)nioc->n_buf, bufsize, mode) != 0)
3633 		rv = EFAULT;
3634 
3635 	nioc->n_len = bufsize;
3636 	kmem_free(log, bufsize);
3637 
3638 	return (rv);
3639 }
3640 
3641 static int
3642 nvme_ioctl_get_features(nvme_t *nvme, int nsid, nvme_ioctl_t *nioc,
3643     int mode, cred_t *cred_p)
3644 {
3645 	_NOTE(ARGUNUSED(cred_p));
3646 	void *buf = NULL;
3647 	size_t bufsize = 0;
3648 	uint32_t res = 0;
3649 	uint8_t feature;
3650 	int rv = 0;
3651 
3652 	if ((mode & FREAD) == 0)
3653 		return (EPERM);
3654 
3655 	if ((nioc->n_arg >> 32) > 0xff)
3656 		return (EINVAL);
3657 
3658 	feature = (uint8_t)(nioc->n_arg >> 32);
3659 
3660 	switch (feature) {
3661 	case NVME_FEAT_ARBITRATION:
3662 	case NVME_FEAT_POWER_MGMT:
3663 	case NVME_FEAT_TEMPERATURE:
3664 	case NVME_FEAT_ERROR:
3665 	case NVME_FEAT_NQUEUES:
3666 	case NVME_FEAT_INTR_COAL:
3667 	case NVME_FEAT_WRITE_ATOM:
3668 	case NVME_FEAT_ASYNC_EVENT:
3669 	case NVME_FEAT_PROGRESS:
3670 		if (nsid != 0)
3671 			return (EINVAL);
3672 		break;
3673 
3674 	case NVME_FEAT_INTR_VECT:
3675 		if (nsid != 0)
3676 			return (EINVAL);
3677 
3678 		res = nioc->n_arg & 0xffffffffUL;
3679 		if (res >= nvme->n_intr_cnt)
3680 			return (EINVAL);
3681 		break;
3682 
3683 	case NVME_FEAT_LBA_RANGE:
3684 		if (nvme->n_lba_range_supported == B_FALSE)
3685 			return (EINVAL);
3686 
3687 		if (nsid == 0 ||
3688 		    nsid > nvme->n_namespace_count)
3689 			return (EINVAL);
3690 
3691 		break;
3692 
3693 	case NVME_FEAT_WRITE_CACHE:
3694 		if (nsid != 0)
3695 			return (EINVAL);
3696 
3697 		if (!nvme->n_write_cache_present)
3698 			return (EINVAL);
3699 
3700 		break;
3701 
3702 	case NVME_FEAT_AUTO_PST:
3703 		if (nsid != 0)
3704 			return (EINVAL);
3705 
3706 		if (!nvme->n_auto_pst_supported)
3707 			return (EINVAL);
3708 
3709 		break;
3710 
3711 	default:
3712 		return (EINVAL);
3713 	}
3714 
3715 	rv = nvme_get_features(nvme, nsid, feature, &res, &buf, &bufsize);
3716 	if (rv != 0)
3717 		return (rv);
3718 
3719 	if (nioc->n_len < bufsize) {
3720 		kmem_free(buf, bufsize);
3721 		return (EINVAL);
3722 	}
3723 
3724 	if (buf && ddi_copyout(buf, (void*)nioc->n_buf, bufsize, mode) != 0)
3725 		rv = EFAULT;
3726 
3727 	kmem_free(buf, bufsize);
3728 	nioc->n_arg = res;
3729 	nioc->n_len = bufsize;
3730 
3731 	return (rv);
3732 }
3733 
3734 static int
3735 nvme_ioctl_intr_cnt(nvme_t *nvme, int nsid, nvme_ioctl_t *nioc, int mode,
3736     cred_t *cred_p)
3737 {
3738 	_NOTE(ARGUNUSED(nsid, mode, cred_p));
3739 
3740 	if ((mode & FREAD) == 0)
3741 		return (EPERM);
3742 
3743 	nioc->n_arg = nvme->n_intr_cnt;
3744 	return (0);
3745 }
3746 
3747 static int
3748 nvme_ioctl_version(nvme_t *nvme, int nsid, nvme_ioctl_t *nioc, int mode,
3749     cred_t *cred_p)
3750 {
3751 	_NOTE(ARGUNUSED(nsid, cred_p));
3752 	int rv = 0;
3753 
3754 	if ((mode & FREAD) == 0)
3755 		return (EPERM);
3756 
3757 	if (nioc->n_len < sizeof (nvme->n_version))
3758 		return (ENOMEM);
3759 
3760 	if (ddi_copyout(&nvme->n_version, (void *)nioc->n_buf,
3761 	    sizeof (nvme->n_version), mode) != 0)
3762 		rv = EFAULT;
3763 
3764 	return (rv);
3765 }
3766 
3767 static int
3768 nvme_ioctl_format(nvme_t *nvme, int nsid, nvme_ioctl_t *nioc, int mode,
3769     cred_t *cred_p)
3770 {
3771 	_NOTE(ARGUNUSED(mode));
3772 	nvme_format_nvm_t frmt = { 0 };
3773 	int c_nsid = nsid != 0 ? nsid - 1 : 0;
3774 
3775 	if ((mode & FWRITE) == 0 || secpolicy_sys_config(cred_p, B_FALSE) != 0)
3776 		return (EPERM);
3777 
3778 	frmt.r = nioc->n_arg & 0xffffffff;
3779 
3780 	/*
3781 	 * Check whether the FORMAT NVM command is supported.
3782 	 */
3783 	if (nvme->n_idctl->id_oacs.oa_format == 0)
3784 		return (EINVAL);
3785 
3786 	/*
3787 	 * Don't allow format or secure erase of individual namespace if that
3788 	 * would cause a format or secure erase of all namespaces.
3789 	 */
3790 	if (nsid != 0 && nvme->n_idctl->id_fna.fn_format != 0)
3791 		return (EINVAL);
3792 
3793 	if (nsid != 0 && frmt.b.fm_ses != NVME_FRMT_SES_NONE &&
3794 	    nvme->n_idctl->id_fna.fn_sec_erase != 0)
3795 		return (EINVAL);
3796 
3797 	/*
3798 	 * Don't allow formatting with Protection Information.
3799 	 */
3800 	if (frmt.b.fm_pi != 0 || frmt.b.fm_pil != 0 || frmt.b.fm_ms != 0)
3801 		return (EINVAL);
3802 
3803 	/*
3804 	 * Don't allow formatting using an illegal LBA format, or any LBA format
3805 	 * that uses metadata.
3806 	 */
3807 	if (frmt.b.fm_lbaf > nvme->n_ns[c_nsid].ns_idns->id_nlbaf ||
3808 	    nvme->n_ns[c_nsid].ns_idns->id_lbaf[frmt.b.fm_lbaf].lbaf_ms != 0)
3809 		return (EINVAL);
3810 
3811 	/*
3812 	 * Don't allow formatting using an illegal Secure Erase setting.
3813 	 */
3814 	if (frmt.b.fm_ses > NVME_FRMT_MAX_SES ||
3815 	    (frmt.b.fm_ses == NVME_FRMT_SES_CRYPTO &&
3816 	    nvme->n_idctl->id_fna.fn_crypt_erase == 0))
3817 		return (EINVAL);
3818 
3819 	if (nsid == 0)
3820 		nsid = (uint32_t)-1;
3821 
3822 	return (nvme_format_nvm(nvme, nsid, frmt.b.fm_lbaf, B_FALSE, 0, B_FALSE,
3823 	    frmt.b.fm_ses));
3824 }
3825 
3826 static int
3827 nvme_ioctl_detach(nvme_t *nvme, int nsid, nvme_ioctl_t *nioc, int mode,
3828     cred_t *cred_p)
3829 {
3830 	_NOTE(ARGUNUSED(nioc, mode));
3831 	int rv = 0;
3832 
3833 	if ((mode & FWRITE) == 0 || secpolicy_sys_config(cred_p, B_FALSE) != 0)
3834 		return (EPERM);
3835 
3836 	if (nsid == 0)
3837 		return (EINVAL);
3838 
3839 	rv = bd_detach_handle(nvme->n_ns[nsid - 1].ns_bd_hdl);
3840 	if (rv != DDI_SUCCESS)
3841 		rv = EBUSY;
3842 
3843 	return (rv);
3844 }
3845 
3846 static int
3847 nvme_ioctl_attach(nvme_t *nvme, int nsid, nvme_ioctl_t *nioc, int mode,
3848     cred_t *cred_p)
3849 {
3850 	_NOTE(ARGUNUSED(nioc, mode));
3851 	nvme_identify_nsid_t *idns;
3852 	int rv = 0;
3853 
3854 	if ((mode & FWRITE) == 0 || secpolicy_sys_config(cred_p, B_FALSE) != 0)
3855 		return (EPERM);
3856 
3857 	if (nsid == 0)
3858 		return (EINVAL);
3859 
3860 	/*
3861 	 * Identify namespace again, free old identify data.
3862 	 */
3863 	idns = nvme->n_ns[nsid - 1].ns_idns;
3864 	if (nvme_init_ns(nvme, nsid) != DDI_SUCCESS)
3865 		return (EIO);
3866 
3867 	kmem_free(idns, sizeof (nvme_identify_nsid_t));
3868 
3869 	rv = bd_attach_handle(nvme->n_dip, nvme->n_ns[nsid - 1].ns_bd_hdl);
3870 	if (rv != DDI_SUCCESS)
3871 		rv = EBUSY;
3872 
3873 	return (rv);
3874 }
3875 
3876 static int
3877 nvme_ioctl(dev_t dev, int cmd, intptr_t arg, int mode, cred_t *cred_p,
3878     int *rval_p)
3879 {
3880 #ifndef __lock_lint
3881 	_NOTE(ARGUNUSED(rval_p));
3882 #endif
3883 	minor_t minor = getminor(dev);
3884 	nvme_t *nvme = ddi_get_soft_state(nvme_state, NVME_MINOR_INST(minor));
3885 	int nsid = NVME_MINOR_NSID(minor);
3886 	int rv = 0;
3887 	nvme_ioctl_t nioc;
3888 
3889 	int (*nvme_ioctl[])(nvme_t *, int, nvme_ioctl_t *, int, cred_t *) = {
3890 		NULL,
3891 		nvme_ioctl_identify,
3892 		nvme_ioctl_identify,
3893 		nvme_ioctl_capabilities,
3894 		nvme_ioctl_get_logpage,
3895 		nvme_ioctl_get_features,
3896 		nvme_ioctl_intr_cnt,
3897 		nvme_ioctl_version,
3898 		nvme_ioctl_format,
3899 		nvme_ioctl_detach,
3900 		nvme_ioctl_attach
3901 	};
3902 
3903 	if (nvme == NULL)
3904 		return (ENXIO);
3905 
3906 	if (nsid > nvme->n_namespace_count)
3907 		return (ENXIO);
3908 
3909 	if (IS_DEVCTL(cmd))
3910 		return (ndi_devctl_ioctl(nvme->n_dip, cmd, arg, mode, 0));
3911 
3912 #ifdef _MULTI_DATAMODEL
3913 	switch (ddi_model_convert_from(mode & FMODELS)) {
3914 	case DDI_MODEL_ILP32: {
3915 		nvme_ioctl32_t nioc32;
3916 		if (ddi_copyin((void*)arg, &nioc32, sizeof (nvme_ioctl32_t),
3917 		    mode) != 0)
3918 			return (EFAULT);
3919 		nioc.n_len = nioc32.n_len;
3920 		nioc.n_buf = nioc32.n_buf;
3921 		nioc.n_arg = nioc32.n_arg;
3922 		break;
3923 	}
3924 	case DDI_MODEL_NONE:
3925 #endif
3926 		if (ddi_copyin((void*)arg, &nioc, sizeof (nvme_ioctl_t), mode)
3927 		    != 0)
3928 			return (EFAULT);
3929 #ifdef _MULTI_DATAMODEL
3930 		break;
3931 	}
3932 #endif
3933 
3934 	if (nvme->n_dead && cmd != NVME_IOC_DETACH)
3935 		return (EIO);
3936 
3937 
3938 	if (cmd == NVME_IOC_IDENTIFY_CTRL) {
3939 		/*
3940 		 * This makes NVME_IOC_IDENTIFY_CTRL work the same on devctl and
3941 		 * attachment point nodes.
3942 		 */
3943 		nsid = 0;
3944 	} else if (cmd == NVME_IOC_IDENTIFY_NSID && nsid == 0) {
3945 		/*
3946 		 * This makes NVME_IOC_IDENTIFY_NSID work on a devctl node, it
3947 		 * will always return identify data for namespace 1.
3948 		 */
3949 		nsid = 1;
3950 	}
3951 
3952 	if (IS_NVME_IOC(cmd) && nvme_ioctl[NVME_IOC_CMD(cmd)] != NULL)
3953 		rv = nvme_ioctl[NVME_IOC_CMD(cmd)](nvme, nsid, &nioc, mode,
3954 		    cred_p);
3955 	else
3956 		rv = EINVAL;
3957 
3958 #ifdef _MULTI_DATAMODEL
3959 	switch (ddi_model_convert_from(mode & FMODELS)) {
3960 	case DDI_MODEL_ILP32: {
3961 		nvme_ioctl32_t nioc32;
3962 
3963 		nioc32.n_len = (size32_t)nioc.n_len;
3964 		nioc32.n_buf = (uintptr32_t)nioc.n_buf;
3965 		nioc32.n_arg = nioc.n_arg;
3966 
3967 		if (ddi_copyout(&nioc32, (void *)arg, sizeof (nvme_ioctl32_t),
3968 		    mode) != 0)
3969 			return (EFAULT);
3970 		break;
3971 	}
3972 	case DDI_MODEL_NONE:
3973 #endif
3974 		if (ddi_copyout(&nioc, (void *)arg, sizeof (nvme_ioctl_t), mode)
3975 		    != 0)
3976 			return (EFAULT);
3977 #ifdef _MULTI_DATAMODEL
3978 		break;
3979 	}
3980 #endif
3981 
3982 	return (rv);
3983 }
3984