xref: /illumos-gate/usr/src/uts/common/io/nvme/nvme.c (revision cb41b9c565d4eec9e1f06e24d429696f59f2f07d)
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 2017 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  *   versions or namespaces with EUI64 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 /* NVMe spec version supported */
263 static const int nvme_version_major = 1;
264 static const int nvme_version_minor = 2;
265 
266 /* tunable for admin command timeout in seconds, default is 1s */
267 int nvme_admin_cmd_timeout = 1;
268 
269 /* tunable for FORMAT NVM command timeout in seconds, default is 600s */
270 int nvme_format_cmd_timeout = 600;
271 
272 static int nvme_attach(dev_info_t *, ddi_attach_cmd_t);
273 static int nvme_detach(dev_info_t *, ddi_detach_cmd_t);
274 static int nvme_quiesce(dev_info_t *);
275 static int nvme_fm_errcb(dev_info_t *, ddi_fm_error_t *, const void *);
276 static int nvme_setup_interrupts(nvme_t *, int, int);
277 static void nvme_release_interrupts(nvme_t *);
278 static uint_t nvme_intr(caddr_t, caddr_t);
279 
280 static void nvme_shutdown(nvme_t *, int, boolean_t);
281 static boolean_t nvme_reset(nvme_t *, boolean_t);
282 static int nvme_init(nvme_t *);
283 static nvme_cmd_t *nvme_alloc_cmd(nvme_t *, int);
284 static void nvme_free_cmd(nvme_cmd_t *);
285 static nvme_cmd_t *nvme_create_nvm_cmd(nvme_namespace_t *, uint8_t,
286     bd_xfer_t *);
287 static void nvme_admin_cmd(nvme_cmd_t *, int);
288 static void nvme_submit_admin_cmd(nvme_qpair_t *, nvme_cmd_t *);
289 static int nvme_submit_io_cmd(nvme_qpair_t *, nvme_cmd_t *);
290 static void nvme_submit_cmd_common(nvme_qpair_t *, nvme_cmd_t *);
291 static nvme_cmd_t *nvme_unqueue_cmd(nvme_t *, nvme_qpair_t *, int);
292 static nvme_cmd_t *nvme_retrieve_cmd(nvme_t *, nvme_qpair_t *);
293 static void nvme_wait_cmd(nvme_cmd_t *, uint_t);
294 static void nvme_wakeup_cmd(void *);
295 static void nvme_async_event_task(void *);
296 
297 static int nvme_check_unknown_cmd_status(nvme_cmd_t *);
298 static int nvme_check_vendor_cmd_status(nvme_cmd_t *);
299 static int nvme_check_integrity_cmd_status(nvme_cmd_t *);
300 static int nvme_check_specific_cmd_status(nvme_cmd_t *);
301 static int nvme_check_generic_cmd_status(nvme_cmd_t *);
302 static inline int nvme_check_cmd_status(nvme_cmd_t *);
303 
304 static int nvme_abort_cmd(nvme_cmd_t *, uint_t);
305 static void nvme_async_event(nvme_t *);
306 static int nvme_format_nvm(nvme_t *, uint32_t, uint8_t, boolean_t, uint8_t,
307     boolean_t, uint8_t);
308 static int nvme_get_logpage(nvme_t *, void **, size_t *, uint8_t, ...);
309 static int nvme_identify(nvme_t *, uint32_t, void **);
310 static int nvme_set_features(nvme_t *, uint32_t, uint8_t, uint32_t,
311     uint32_t *);
312 static int nvme_get_features(nvme_t *, uint32_t, uint8_t, uint32_t *,
313     void **, size_t *);
314 static int nvme_write_cache_set(nvme_t *, boolean_t);
315 static int nvme_set_nqueues(nvme_t *, uint16_t *);
316 
317 static void nvme_free_dma(nvme_dma_t *);
318 static int nvme_zalloc_dma(nvme_t *, size_t, uint_t, ddi_dma_attr_t *,
319     nvme_dma_t **);
320 static int nvme_zalloc_queue_dma(nvme_t *, uint32_t, uint16_t, uint_t,
321     nvme_dma_t **);
322 static void nvme_free_qpair(nvme_qpair_t *);
323 static int nvme_alloc_qpair(nvme_t *, uint32_t, nvme_qpair_t **, int);
324 static int nvme_create_io_qpair(nvme_t *, nvme_qpair_t *, uint16_t);
325 
326 static inline void nvme_put64(nvme_t *, uintptr_t, uint64_t);
327 static inline void nvme_put32(nvme_t *, uintptr_t, uint32_t);
328 static inline uint64_t nvme_get64(nvme_t *, uintptr_t);
329 static inline uint32_t nvme_get32(nvme_t *, uintptr_t);
330 
331 static boolean_t nvme_check_regs_hdl(nvme_t *);
332 static boolean_t nvme_check_dma_hdl(nvme_dma_t *);
333 
334 static int nvme_fill_prp(nvme_cmd_t *, bd_xfer_t *);
335 
336 static void nvme_bd_xfer_done(void *);
337 static void nvme_bd_driveinfo(void *, bd_drive_t *);
338 static int nvme_bd_mediainfo(void *, bd_media_t *);
339 static int nvme_bd_cmd(nvme_namespace_t *, bd_xfer_t *, uint8_t);
340 static int nvme_bd_read(void *, bd_xfer_t *);
341 static int nvme_bd_write(void *, bd_xfer_t *);
342 static int nvme_bd_sync(void *, bd_xfer_t *);
343 static int nvme_bd_devid(void *, dev_info_t *, ddi_devid_t *);
344 
345 static int nvme_prp_dma_constructor(void *, void *, int);
346 static void nvme_prp_dma_destructor(void *, void *);
347 
348 static void nvme_prepare_devid(nvme_t *, uint32_t);
349 
350 static int nvme_open(dev_t *, int, int, cred_t *);
351 static int nvme_close(dev_t, int, int, cred_t *);
352 static int nvme_ioctl(dev_t, int, intptr_t, int, cred_t *, int *);
353 
354 #define	NVME_MINOR_INST_SHIFT	9
355 #define	NVME_MINOR(inst, nsid)	(((inst) << NVME_MINOR_INST_SHIFT) | (nsid))
356 #define	NVME_MINOR_INST(minor)	((minor) >> NVME_MINOR_INST_SHIFT)
357 #define	NVME_MINOR_NSID(minor)	((minor) & ((1 << NVME_MINOR_INST_SHIFT) - 1))
358 #define	NVME_MINOR_MAX		(NVME_MINOR(1, 0) - 2)
359 
360 static void *nvme_state;
361 static kmem_cache_t *nvme_cmd_cache;
362 
363 /*
364  * DMA attributes for queue DMA memory
365  *
366  * Queue DMA memory must be page aligned. The maximum length of a queue is
367  * 65536 entries, and an entry can be 64 bytes long.
368  */
369 static ddi_dma_attr_t nvme_queue_dma_attr = {
370 	.dma_attr_version	= DMA_ATTR_V0,
371 	.dma_attr_addr_lo	= 0,
372 	.dma_attr_addr_hi	= 0xffffffffffffffffULL,
373 	.dma_attr_count_max	= (UINT16_MAX + 1) * sizeof (nvme_sqe_t) - 1,
374 	.dma_attr_align		= 0x1000,
375 	.dma_attr_burstsizes	= 0x7ff,
376 	.dma_attr_minxfer	= 0x1000,
377 	.dma_attr_maxxfer	= (UINT16_MAX + 1) * sizeof (nvme_sqe_t),
378 	.dma_attr_seg		= 0xffffffffffffffffULL,
379 	.dma_attr_sgllen	= 1,
380 	.dma_attr_granular	= 1,
381 	.dma_attr_flags		= 0,
382 };
383 
384 /*
385  * DMA attributes for transfers using Physical Region Page (PRP) entries
386  *
387  * A PRP entry describes one page of DMA memory using the page size specified
388  * in the controller configuration's memory page size register (CC.MPS). It uses
389  * a 64bit base address aligned to this page size. There is no limitation on
390  * chaining PRPs together for arbitrarily large DMA transfers.
391  */
392 static ddi_dma_attr_t nvme_prp_dma_attr = {
393 	.dma_attr_version	= DMA_ATTR_V0,
394 	.dma_attr_addr_lo	= 0,
395 	.dma_attr_addr_hi	= 0xffffffffffffffffULL,
396 	.dma_attr_count_max	= 0xfff,
397 	.dma_attr_align		= 0x1000,
398 	.dma_attr_burstsizes	= 0x7ff,
399 	.dma_attr_minxfer	= 0x1000,
400 	.dma_attr_maxxfer	= 0x1000,
401 	.dma_attr_seg		= 0xfff,
402 	.dma_attr_sgllen	= -1,
403 	.dma_attr_granular	= 1,
404 	.dma_attr_flags		= 0,
405 };
406 
407 /*
408  * DMA attributes for transfers using scatter/gather lists
409  *
410  * A SGL entry describes a chunk of DMA memory using a 64bit base address and a
411  * 32bit length field. SGL Segment and SGL Last Segment entries require the
412  * length to be a multiple of 16 bytes.
413  */
414 static ddi_dma_attr_t nvme_sgl_dma_attr = {
415 	.dma_attr_version	= DMA_ATTR_V0,
416 	.dma_attr_addr_lo	= 0,
417 	.dma_attr_addr_hi	= 0xffffffffffffffffULL,
418 	.dma_attr_count_max	= 0xffffffffUL,
419 	.dma_attr_align		= 1,
420 	.dma_attr_burstsizes	= 0x7ff,
421 	.dma_attr_minxfer	= 0x10,
422 	.dma_attr_maxxfer	= 0xfffffffffULL,
423 	.dma_attr_seg		= 0xffffffffffffffffULL,
424 	.dma_attr_sgllen	= -1,
425 	.dma_attr_granular	= 0x10,
426 	.dma_attr_flags		= 0
427 };
428 
429 static ddi_device_acc_attr_t nvme_reg_acc_attr = {
430 	.devacc_attr_version	= DDI_DEVICE_ATTR_V0,
431 	.devacc_attr_endian_flags = DDI_STRUCTURE_LE_ACC,
432 	.devacc_attr_dataorder	= DDI_STRICTORDER_ACC
433 };
434 
435 static struct cb_ops nvme_cb_ops = {
436 	.cb_open	= nvme_open,
437 	.cb_close	= nvme_close,
438 	.cb_strategy	= nodev,
439 	.cb_print	= nodev,
440 	.cb_dump	= nodev,
441 	.cb_read	= nodev,
442 	.cb_write	= nodev,
443 	.cb_ioctl	= nvme_ioctl,
444 	.cb_devmap	= nodev,
445 	.cb_mmap	= nodev,
446 	.cb_segmap	= nodev,
447 	.cb_chpoll	= nochpoll,
448 	.cb_prop_op	= ddi_prop_op,
449 	.cb_str		= 0,
450 	.cb_flag	= D_NEW | D_MP,
451 	.cb_rev		= CB_REV,
452 	.cb_aread	= nodev,
453 	.cb_awrite	= nodev
454 };
455 
456 static struct dev_ops nvme_dev_ops = {
457 	.devo_rev	= DEVO_REV,
458 	.devo_refcnt	= 0,
459 	.devo_getinfo	= ddi_no_info,
460 	.devo_identify	= nulldev,
461 	.devo_probe	= nulldev,
462 	.devo_attach	= nvme_attach,
463 	.devo_detach	= nvme_detach,
464 	.devo_reset	= nodev,
465 	.devo_cb_ops	= &nvme_cb_ops,
466 	.devo_bus_ops	= NULL,
467 	.devo_power	= NULL,
468 	.devo_quiesce	= nvme_quiesce,
469 };
470 
471 static struct modldrv nvme_modldrv = {
472 	.drv_modops	= &mod_driverops,
473 	.drv_linkinfo	= "NVMe v1.1b",
474 	.drv_dev_ops	= &nvme_dev_ops
475 };
476 
477 static struct modlinkage nvme_modlinkage = {
478 	.ml_rev		= MODREV_1,
479 	.ml_linkage	= { &nvme_modldrv, NULL }
480 };
481 
482 static bd_ops_t nvme_bd_ops = {
483 	.o_version	= BD_OPS_VERSION_0,
484 	.o_drive_info	= nvme_bd_driveinfo,
485 	.o_media_info	= nvme_bd_mediainfo,
486 	.o_devid_init	= nvme_bd_devid,
487 	.o_sync_cache	= nvme_bd_sync,
488 	.o_read		= nvme_bd_read,
489 	.o_write	= nvme_bd_write,
490 };
491 
492 /*
493  * This list will hold commands that have timed out and couldn't be aborted.
494  * As we don't know what the hardware may still do with the DMA memory we can't
495  * free them, so we'll keep them forever on this list where we can easily look
496  * at them with mdb.
497  */
498 static struct list nvme_lost_cmds;
499 static kmutex_t nvme_lc_mutex;
500 
501 int
502 _init(void)
503 {
504 	int error;
505 
506 	error = ddi_soft_state_init(&nvme_state, sizeof (nvme_t), 1);
507 	if (error != DDI_SUCCESS)
508 		return (error);
509 
510 	nvme_cmd_cache = kmem_cache_create("nvme_cmd_cache",
511 	    sizeof (nvme_cmd_t), 64, NULL, NULL, NULL, NULL, NULL, 0);
512 
513 	mutex_init(&nvme_lc_mutex, NULL, MUTEX_DRIVER, NULL);
514 	list_create(&nvme_lost_cmds, sizeof (nvme_cmd_t),
515 	    offsetof(nvme_cmd_t, nc_list));
516 
517 	bd_mod_init(&nvme_dev_ops);
518 
519 	error = mod_install(&nvme_modlinkage);
520 	if (error != DDI_SUCCESS) {
521 		ddi_soft_state_fini(&nvme_state);
522 		mutex_destroy(&nvme_lc_mutex);
523 		list_destroy(&nvme_lost_cmds);
524 		bd_mod_fini(&nvme_dev_ops);
525 	}
526 
527 	return (error);
528 }
529 
530 int
531 _fini(void)
532 {
533 	int error;
534 
535 	if (!list_is_empty(&nvme_lost_cmds))
536 		return (DDI_FAILURE);
537 
538 	error = mod_remove(&nvme_modlinkage);
539 	if (error == DDI_SUCCESS) {
540 		ddi_soft_state_fini(&nvme_state);
541 		kmem_cache_destroy(nvme_cmd_cache);
542 		mutex_destroy(&nvme_lc_mutex);
543 		list_destroy(&nvme_lost_cmds);
544 		bd_mod_fini(&nvme_dev_ops);
545 	}
546 
547 	return (error);
548 }
549 
550 int
551 _info(struct modinfo *modinfop)
552 {
553 	return (mod_info(&nvme_modlinkage, modinfop));
554 }
555 
556 static inline void
557 nvme_put64(nvme_t *nvme, uintptr_t reg, uint64_t val)
558 {
559 	ASSERT(((uintptr_t)(nvme->n_regs + reg) & 0x7) == 0);
560 
561 	/*LINTED: E_BAD_PTR_CAST_ALIGN*/
562 	ddi_put64(nvme->n_regh, (uint64_t *)(nvme->n_regs + reg), val);
563 }
564 
565 static inline void
566 nvme_put32(nvme_t *nvme, uintptr_t reg, uint32_t val)
567 {
568 	ASSERT(((uintptr_t)(nvme->n_regs + reg) & 0x3) == 0);
569 
570 	/*LINTED: E_BAD_PTR_CAST_ALIGN*/
571 	ddi_put32(nvme->n_regh, (uint32_t *)(nvme->n_regs + reg), val);
572 }
573 
574 static inline uint64_t
575 nvme_get64(nvme_t *nvme, uintptr_t reg)
576 {
577 	uint64_t val;
578 
579 	ASSERT(((uintptr_t)(nvme->n_regs + reg) & 0x7) == 0);
580 
581 	/*LINTED: E_BAD_PTR_CAST_ALIGN*/
582 	val = ddi_get64(nvme->n_regh, (uint64_t *)(nvme->n_regs + reg));
583 
584 	return (val);
585 }
586 
587 static inline uint32_t
588 nvme_get32(nvme_t *nvme, uintptr_t reg)
589 {
590 	uint32_t val;
591 
592 	ASSERT(((uintptr_t)(nvme->n_regs + reg) & 0x3) == 0);
593 
594 	/*LINTED: E_BAD_PTR_CAST_ALIGN*/
595 	val = ddi_get32(nvme->n_regh, (uint32_t *)(nvme->n_regs + reg));
596 
597 	return (val);
598 }
599 
600 static boolean_t
601 nvme_check_regs_hdl(nvme_t *nvme)
602 {
603 	ddi_fm_error_t error;
604 
605 	ddi_fm_acc_err_get(nvme->n_regh, &error, DDI_FME_VERSION);
606 
607 	if (error.fme_status != DDI_FM_OK)
608 		return (B_TRUE);
609 
610 	return (B_FALSE);
611 }
612 
613 static boolean_t
614 nvme_check_dma_hdl(nvme_dma_t *dma)
615 {
616 	ddi_fm_error_t error;
617 
618 	if (dma == NULL)
619 		return (B_FALSE);
620 
621 	ddi_fm_dma_err_get(dma->nd_dmah, &error, DDI_FME_VERSION);
622 
623 	if (error.fme_status != DDI_FM_OK)
624 		return (B_TRUE);
625 
626 	return (B_FALSE);
627 }
628 
629 static void
630 nvme_free_dma_common(nvme_dma_t *dma)
631 {
632 	if (dma->nd_dmah != NULL)
633 		(void) ddi_dma_unbind_handle(dma->nd_dmah);
634 	if (dma->nd_acch != NULL)
635 		ddi_dma_mem_free(&dma->nd_acch);
636 	if (dma->nd_dmah != NULL)
637 		ddi_dma_free_handle(&dma->nd_dmah);
638 }
639 
640 static void
641 nvme_free_dma(nvme_dma_t *dma)
642 {
643 	nvme_free_dma_common(dma);
644 	kmem_free(dma, sizeof (*dma));
645 }
646 
647 /* ARGSUSED */
648 static void
649 nvme_prp_dma_destructor(void *buf, void *private)
650 {
651 	nvme_dma_t *dma = (nvme_dma_t *)buf;
652 
653 	nvme_free_dma_common(dma);
654 }
655 
656 static int
657 nvme_alloc_dma_common(nvme_t *nvme, nvme_dma_t *dma,
658     size_t len, uint_t flags, ddi_dma_attr_t *dma_attr)
659 {
660 	if (ddi_dma_alloc_handle(nvme->n_dip, dma_attr, DDI_DMA_SLEEP, NULL,
661 	    &dma->nd_dmah) != DDI_SUCCESS) {
662 		/*
663 		 * Due to DDI_DMA_SLEEP this can't be DDI_DMA_NORESOURCES, and
664 		 * the only other possible error is DDI_DMA_BADATTR which
665 		 * indicates a driver bug which should cause a panic.
666 		 */
667 		dev_err(nvme->n_dip, CE_PANIC,
668 		    "!failed to get DMA handle, check DMA attributes");
669 		return (DDI_FAILURE);
670 	}
671 
672 	/*
673 	 * ddi_dma_mem_alloc() can only fail when DDI_DMA_NOSLEEP is specified
674 	 * or the flags are conflicting, which isn't the case here.
675 	 */
676 	(void) ddi_dma_mem_alloc(dma->nd_dmah, len, &nvme->n_reg_acc_attr,
677 	    DDI_DMA_CONSISTENT, DDI_DMA_SLEEP, NULL, &dma->nd_memp,
678 	    &dma->nd_len, &dma->nd_acch);
679 
680 	if (ddi_dma_addr_bind_handle(dma->nd_dmah, NULL, dma->nd_memp,
681 	    dma->nd_len, flags | DDI_DMA_CONSISTENT, DDI_DMA_SLEEP, NULL,
682 	    &dma->nd_cookie, &dma->nd_ncookie) != DDI_DMA_MAPPED) {
683 		dev_err(nvme->n_dip, CE_WARN,
684 		    "!failed to bind DMA memory");
685 		atomic_inc_32(&nvme->n_dma_bind_err);
686 		nvme_free_dma_common(dma);
687 		return (DDI_FAILURE);
688 	}
689 
690 	return (DDI_SUCCESS);
691 }
692 
693 static int
694 nvme_zalloc_dma(nvme_t *nvme, size_t len, uint_t flags,
695     ddi_dma_attr_t *dma_attr, nvme_dma_t **ret)
696 {
697 	nvme_dma_t *dma = kmem_zalloc(sizeof (nvme_dma_t), KM_SLEEP);
698 
699 	if (nvme_alloc_dma_common(nvme, dma, len, flags, dma_attr) !=
700 	    DDI_SUCCESS) {
701 		*ret = NULL;
702 		kmem_free(dma, sizeof (nvme_dma_t));
703 		return (DDI_FAILURE);
704 	}
705 
706 	bzero(dma->nd_memp, dma->nd_len);
707 
708 	*ret = dma;
709 	return (DDI_SUCCESS);
710 }
711 
712 /* ARGSUSED */
713 static int
714 nvme_prp_dma_constructor(void *buf, void *private, int flags)
715 {
716 	nvme_dma_t *dma = (nvme_dma_t *)buf;
717 	nvme_t *nvme = (nvme_t *)private;
718 
719 	dma->nd_dmah = NULL;
720 	dma->nd_acch = NULL;
721 
722 	if (nvme_alloc_dma_common(nvme, dma, nvme->n_pagesize,
723 	    DDI_DMA_READ, &nvme->n_prp_dma_attr) != DDI_SUCCESS) {
724 		return (-1);
725 	}
726 
727 	ASSERT(dma->nd_ncookie == 1);
728 
729 	dma->nd_cached = B_TRUE;
730 
731 	return (0);
732 }
733 
734 static int
735 nvme_zalloc_queue_dma(nvme_t *nvme, uint32_t nentry, uint16_t qe_len,
736     uint_t flags, nvme_dma_t **dma)
737 {
738 	uint32_t len = nentry * qe_len;
739 	ddi_dma_attr_t q_dma_attr = nvme->n_queue_dma_attr;
740 
741 	len = roundup(len, nvme->n_pagesize);
742 
743 	q_dma_attr.dma_attr_minxfer = len;
744 
745 	if (nvme_zalloc_dma(nvme, len, flags, &q_dma_attr, dma)
746 	    != DDI_SUCCESS) {
747 		dev_err(nvme->n_dip, CE_WARN,
748 		    "!failed to get DMA memory for queue");
749 		goto fail;
750 	}
751 
752 	if ((*dma)->nd_ncookie != 1) {
753 		dev_err(nvme->n_dip, CE_WARN,
754 		    "!got too many cookies for queue DMA");
755 		goto fail;
756 	}
757 
758 	return (DDI_SUCCESS);
759 
760 fail:
761 	if (*dma) {
762 		nvme_free_dma(*dma);
763 		*dma = NULL;
764 	}
765 
766 	return (DDI_FAILURE);
767 }
768 
769 static void
770 nvme_free_qpair(nvme_qpair_t *qp)
771 {
772 	int i;
773 
774 	mutex_destroy(&qp->nq_mutex);
775 	sema_destroy(&qp->nq_sema);
776 
777 	if (qp->nq_sqdma != NULL)
778 		nvme_free_dma(qp->nq_sqdma);
779 	if (qp->nq_cqdma != NULL)
780 		nvme_free_dma(qp->nq_cqdma);
781 
782 	if (qp->nq_active_cmds > 0)
783 		for (i = 0; i != qp->nq_nentry; i++)
784 			if (qp->nq_cmd[i] != NULL)
785 				nvme_free_cmd(qp->nq_cmd[i]);
786 
787 	if (qp->nq_cmd != NULL)
788 		kmem_free(qp->nq_cmd, sizeof (nvme_cmd_t *) * qp->nq_nentry);
789 
790 	kmem_free(qp, sizeof (nvme_qpair_t));
791 }
792 
793 static int
794 nvme_alloc_qpair(nvme_t *nvme, uint32_t nentry, nvme_qpair_t **nqp,
795     int idx)
796 {
797 	nvme_qpair_t *qp = kmem_zalloc(sizeof (*qp), KM_SLEEP);
798 
799 	mutex_init(&qp->nq_mutex, NULL, MUTEX_DRIVER,
800 	    DDI_INTR_PRI(nvme->n_intr_pri));
801 	sema_init(&qp->nq_sema, nentry, NULL, SEMA_DRIVER, NULL);
802 
803 	if (nvme_zalloc_queue_dma(nvme, nentry, sizeof (nvme_sqe_t),
804 	    DDI_DMA_WRITE, &qp->nq_sqdma) != DDI_SUCCESS)
805 		goto fail;
806 
807 	if (nvme_zalloc_queue_dma(nvme, nentry, sizeof (nvme_cqe_t),
808 	    DDI_DMA_READ, &qp->nq_cqdma) != DDI_SUCCESS)
809 		goto fail;
810 
811 	qp->nq_sq = (nvme_sqe_t *)qp->nq_sqdma->nd_memp;
812 	qp->nq_cq = (nvme_cqe_t *)qp->nq_cqdma->nd_memp;
813 	qp->nq_nentry = nentry;
814 
815 	qp->nq_sqtdbl = NVME_REG_SQTDBL(nvme, idx);
816 	qp->nq_cqhdbl = NVME_REG_CQHDBL(nvme, idx);
817 
818 	qp->nq_cmd = kmem_zalloc(sizeof (nvme_cmd_t *) * nentry, KM_SLEEP);
819 	qp->nq_next_cmd = 0;
820 
821 	*nqp = qp;
822 	return (DDI_SUCCESS);
823 
824 fail:
825 	nvme_free_qpair(qp);
826 	*nqp = NULL;
827 
828 	return (DDI_FAILURE);
829 }
830 
831 static nvme_cmd_t *
832 nvme_alloc_cmd(nvme_t *nvme, int kmflag)
833 {
834 	nvme_cmd_t *cmd = kmem_cache_alloc(nvme_cmd_cache, kmflag);
835 
836 	if (cmd == NULL)
837 		return (cmd);
838 
839 	bzero(cmd, sizeof (nvme_cmd_t));
840 
841 	cmd->nc_nvme = nvme;
842 
843 	mutex_init(&cmd->nc_mutex, NULL, MUTEX_DRIVER,
844 	    DDI_INTR_PRI(nvme->n_intr_pri));
845 	cv_init(&cmd->nc_cv, NULL, CV_DRIVER, NULL);
846 
847 	return (cmd);
848 }
849 
850 static void
851 nvme_free_cmd(nvme_cmd_t *cmd)
852 {
853 	/* Don't free commands on the lost commands list. */
854 	if (list_link_active(&cmd->nc_list))
855 		return;
856 
857 	if (cmd->nc_dma) {
858 		if (cmd->nc_dma->nd_cached)
859 			kmem_cache_free(cmd->nc_nvme->n_prp_cache,
860 			    cmd->nc_dma);
861 		else
862 			nvme_free_dma(cmd->nc_dma);
863 		cmd->nc_dma = NULL;
864 	}
865 
866 	cv_destroy(&cmd->nc_cv);
867 	mutex_destroy(&cmd->nc_mutex);
868 
869 	kmem_cache_free(nvme_cmd_cache, cmd);
870 }
871 
872 static void
873 nvme_submit_admin_cmd(nvme_qpair_t *qp, nvme_cmd_t *cmd)
874 {
875 	sema_p(&qp->nq_sema);
876 	nvme_submit_cmd_common(qp, cmd);
877 }
878 
879 static int
880 nvme_submit_io_cmd(nvme_qpair_t *qp, nvme_cmd_t *cmd)
881 {
882 	if (sema_tryp(&qp->nq_sema) == 0)
883 		return (EAGAIN);
884 
885 	nvme_submit_cmd_common(qp, cmd);
886 	return (0);
887 }
888 
889 static void
890 nvme_submit_cmd_common(nvme_qpair_t *qp, nvme_cmd_t *cmd)
891 {
892 	nvme_reg_sqtdbl_t tail = { 0 };
893 
894 	mutex_enter(&qp->nq_mutex);
895 	cmd->nc_completed = B_FALSE;
896 
897 	/*
898 	 * Try to insert the cmd into the active cmd array at the nq_next_cmd
899 	 * slot. If the slot is already occupied advance to the next slot and
900 	 * try again. This can happen for long running commands like async event
901 	 * requests.
902 	 */
903 	while (qp->nq_cmd[qp->nq_next_cmd] != NULL)
904 		qp->nq_next_cmd = (qp->nq_next_cmd + 1) % qp->nq_nentry;
905 	qp->nq_cmd[qp->nq_next_cmd] = cmd;
906 
907 	qp->nq_active_cmds++;
908 
909 	cmd->nc_sqe.sqe_cid = qp->nq_next_cmd;
910 	bcopy(&cmd->nc_sqe, &qp->nq_sq[qp->nq_sqtail], sizeof (nvme_sqe_t));
911 	(void) ddi_dma_sync(qp->nq_sqdma->nd_dmah,
912 	    sizeof (nvme_sqe_t) * qp->nq_sqtail,
913 	    sizeof (nvme_sqe_t), DDI_DMA_SYNC_FORDEV);
914 	qp->nq_next_cmd = (qp->nq_next_cmd + 1) % qp->nq_nentry;
915 
916 	tail.b.sqtdbl_sqt = qp->nq_sqtail = (qp->nq_sqtail + 1) % qp->nq_nentry;
917 	nvme_put32(cmd->nc_nvme, qp->nq_sqtdbl, tail.r);
918 
919 	mutex_exit(&qp->nq_mutex);
920 }
921 
922 static nvme_cmd_t *
923 nvme_unqueue_cmd(nvme_t *nvme, nvme_qpair_t *qp, int cid)
924 {
925 	nvme_cmd_t *cmd;
926 
927 	ASSERT(mutex_owned(&qp->nq_mutex));
928 	ASSERT3S(cid, <, qp->nq_nentry);
929 
930 	cmd = qp->nq_cmd[cid];
931 	qp->nq_cmd[cid] = NULL;
932 	ASSERT3U(qp->nq_active_cmds, >, 0);
933 	qp->nq_active_cmds--;
934 	sema_v(&qp->nq_sema);
935 
936 	ASSERT3P(cmd, !=, NULL);
937 	ASSERT3P(cmd->nc_nvme, ==, nvme);
938 	ASSERT3S(cmd->nc_sqe.sqe_cid, ==, cid);
939 
940 	return (cmd);
941 }
942 
943 static nvme_cmd_t *
944 nvme_retrieve_cmd(nvme_t *nvme, nvme_qpair_t *qp)
945 {
946 	nvme_reg_cqhdbl_t head = { 0 };
947 
948 	nvme_cqe_t *cqe;
949 	nvme_cmd_t *cmd;
950 
951 	(void) ddi_dma_sync(qp->nq_cqdma->nd_dmah, 0,
952 	    sizeof (nvme_cqe_t) * qp->nq_nentry, DDI_DMA_SYNC_FORKERNEL);
953 
954 	mutex_enter(&qp->nq_mutex);
955 	cqe = &qp->nq_cq[qp->nq_cqhead];
956 
957 	/* Check phase tag of CQE. Hardware inverts it for new entries. */
958 	if (cqe->cqe_sf.sf_p == qp->nq_phase) {
959 		mutex_exit(&qp->nq_mutex);
960 		return (NULL);
961 	}
962 
963 	ASSERT(nvme->n_ioq[cqe->cqe_sqid] == qp);
964 
965 	cmd = nvme_unqueue_cmd(nvme, qp, cqe->cqe_cid);
966 
967 	ASSERT(cmd->nc_sqid == cqe->cqe_sqid);
968 	bcopy(cqe, &cmd->nc_cqe, sizeof (nvme_cqe_t));
969 
970 	qp->nq_sqhead = cqe->cqe_sqhd;
971 
972 	head.b.cqhdbl_cqh = qp->nq_cqhead = (qp->nq_cqhead + 1) % qp->nq_nentry;
973 
974 	/* Toggle phase on wrap-around. */
975 	if (qp->nq_cqhead == 0)
976 		qp->nq_phase = qp->nq_phase ? 0 : 1;
977 
978 	nvme_put32(cmd->nc_nvme, qp->nq_cqhdbl, head.r);
979 	mutex_exit(&qp->nq_mutex);
980 
981 	return (cmd);
982 }
983 
984 static int
985 nvme_check_unknown_cmd_status(nvme_cmd_t *cmd)
986 {
987 	nvme_cqe_t *cqe = &cmd->nc_cqe;
988 
989 	dev_err(cmd->nc_nvme->n_dip, CE_WARN,
990 	    "!unknown command status received: opc = %x, sqid = %d, cid = %d, "
991 	    "sc = %x, sct = %x, dnr = %d, m = %d", cmd->nc_sqe.sqe_opc,
992 	    cqe->cqe_sqid, cqe->cqe_cid, cqe->cqe_sf.sf_sc, cqe->cqe_sf.sf_sct,
993 	    cqe->cqe_sf.sf_dnr, cqe->cqe_sf.sf_m);
994 
995 	if (cmd->nc_xfer != NULL)
996 		bd_error(cmd->nc_xfer, BD_ERR_ILLRQ);
997 
998 	if (cmd->nc_nvme->n_strict_version) {
999 		cmd->nc_nvme->n_dead = B_TRUE;
1000 		ddi_fm_service_impact(cmd->nc_nvme->n_dip, DDI_SERVICE_LOST);
1001 	}
1002 
1003 	return (EIO);
1004 }
1005 
1006 static int
1007 nvme_check_vendor_cmd_status(nvme_cmd_t *cmd)
1008 {
1009 	nvme_cqe_t *cqe = &cmd->nc_cqe;
1010 
1011 	dev_err(cmd->nc_nvme->n_dip, CE_WARN,
1012 	    "!unknown command status received: opc = %x, sqid = %d, cid = %d, "
1013 	    "sc = %x, sct = %x, dnr = %d, m = %d", cmd->nc_sqe.sqe_opc,
1014 	    cqe->cqe_sqid, cqe->cqe_cid, cqe->cqe_sf.sf_sc, cqe->cqe_sf.sf_sct,
1015 	    cqe->cqe_sf.sf_dnr, cqe->cqe_sf.sf_m);
1016 	if (!cmd->nc_nvme->n_ignore_unknown_vendor_status) {
1017 		cmd->nc_nvme->n_dead = B_TRUE;
1018 		ddi_fm_service_impact(cmd->nc_nvme->n_dip, DDI_SERVICE_LOST);
1019 	}
1020 
1021 	return (EIO);
1022 }
1023 
1024 static int
1025 nvme_check_integrity_cmd_status(nvme_cmd_t *cmd)
1026 {
1027 	nvme_cqe_t *cqe = &cmd->nc_cqe;
1028 
1029 	switch (cqe->cqe_sf.sf_sc) {
1030 	case NVME_CQE_SC_INT_NVM_WRITE:
1031 		/* write fail */
1032 		/* TODO: post ereport */
1033 		if (cmd->nc_xfer != NULL)
1034 			bd_error(cmd->nc_xfer, BD_ERR_MEDIA);
1035 		return (EIO);
1036 
1037 	case NVME_CQE_SC_INT_NVM_READ:
1038 		/* read fail */
1039 		/* TODO: post ereport */
1040 		if (cmd->nc_xfer != NULL)
1041 			bd_error(cmd->nc_xfer, BD_ERR_MEDIA);
1042 		return (EIO);
1043 
1044 	default:
1045 		return (nvme_check_unknown_cmd_status(cmd));
1046 	}
1047 }
1048 
1049 static int
1050 nvme_check_generic_cmd_status(nvme_cmd_t *cmd)
1051 {
1052 	nvme_cqe_t *cqe = &cmd->nc_cqe;
1053 
1054 	switch (cqe->cqe_sf.sf_sc) {
1055 	case NVME_CQE_SC_GEN_SUCCESS:
1056 		return (0);
1057 
1058 	/*
1059 	 * Errors indicating a bug in the driver should cause a panic.
1060 	 */
1061 	case NVME_CQE_SC_GEN_INV_OPC:
1062 		/* Invalid Command Opcode */
1063 		if (!cmd->nc_dontpanic)
1064 			dev_err(cmd->nc_nvme->n_dip, CE_PANIC,
1065 			    "programming error: invalid opcode in cmd %p",
1066 			    (void *)cmd);
1067 		return (EINVAL);
1068 
1069 	case NVME_CQE_SC_GEN_INV_FLD:
1070 		/* Invalid Field in Command */
1071 		if (!cmd->nc_dontpanic)
1072 			dev_err(cmd->nc_nvme->n_dip, CE_PANIC,
1073 			    "programming error: invalid field in cmd %p",
1074 			    (void *)cmd);
1075 		return (EIO);
1076 
1077 	case NVME_CQE_SC_GEN_ID_CNFL:
1078 		/* Command ID Conflict */
1079 		dev_err(cmd->nc_nvme->n_dip, CE_PANIC, "programming error: "
1080 		    "cmd ID conflict in cmd %p", (void *)cmd);
1081 		return (0);
1082 
1083 	case NVME_CQE_SC_GEN_INV_NS:
1084 		/* Invalid Namespace or Format */
1085 		if (!cmd->nc_dontpanic)
1086 			dev_err(cmd->nc_nvme->n_dip, CE_PANIC,
1087 			    "programming error: invalid NS/format in cmd %p",
1088 			    (void *)cmd);
1089 		return (EINVAL);
1090 
1091 	case NVME_CQE_SC_GEN_NVM_LBA_RANGE:
1092 		/* LBA Out Of Range */
1093 		dev_err(cmd->nc_nvme->n_dip, CE_PANIC, "programming error: "
1094 		    "LBA out of range in cmd %p", (void *)cmd);
1095 		return (0);
1096 
1097 	/*
1098 	 * Non-fatal errors, handle gracefully.
1099 	 */
1100 	case NVME_CQE_SC_GEN_DATA_XFR_ERR:
1101 		/* Data Transfer Error (DMA) */
1102 		/* TODO: post ereport */
1103 		atomic_inc_32(&cmd->nc_nvme->n_data_xfr_err);
1104 		if (cmd->nc_xfer != NULL)
1105 			bd_error(cmd->nc_xfer, BD_ERR_NTRDY);
1106 		return (EIO);
1107 
1108 	case NVME_CQE_SC_GEN_INTERNAL_ERR:
1109 		/*
1110 		 * Internal Error. The spec (v1.0, section 4.5.1.2) says
1111 		 * detailed error information is returned as async event,
1112 		 * so we pretty much ignore the error here and handle it
1113 		 * in the async event handler.
1114 		 */
1115 		atomic_inc_32(&cmd->nc_nvme->n_internal_err);
1116 		if (cmd->nc_xfer != NULL)
1117 			bd_error(cmd->nc_xfer, BD_ERR_NTRDY);
1118 		return (EIO);
1119 
1120 	case NVME_CQE_SC_GEN_ABORT_REQUEST:
1121 		/*
1122 		 * Command Abort Requested. This normally happens only when a
1123 		 * command times out.
1124 		 */
1125 		/* TODO: post ereport or change blkdev to handle this? */
1126 		atomic_inc_32(&cmd->nc_nvme->n_abort_rq_err);
1127 		return (ECANCELED);
1128 
1129 	case NVME_CQE_SC_GEN_ABORT_PWRLOSS:
1130 		/* Command Aborted due to Power Loss Notification */
1131 		ddi_fm_service_impact(cmd->nc_nvme->n_dip, DDI_SERVICE_LOST);
1132 		cmd->nc_nvme->n_dead = B_TRUE;
1133 		return (EIO);
1134 
1135 	case NVME_CQE_SC_GEN_ABORT_SQ_DEL:
1136 		/* Command Aborted due to SQ Deletion */
1137 		atomic_inc_32(&cmd->nc_nvme->n_abort_sq_del);
1138 		return (EIO);
1139 
1140 	case NVME_CQE_SC_GEN_NVM_CAP_EXC:
1141 		/* Capacity Exceeded */
1142 		atomic_inc_32(&cmd->nc_nvme->n_nvm_cap_exc);
1143 		if (cmd->nc_xfer != NULL)
1144 			bd_error(cmd->nc_xfer, BD_ERR_MEDIA);
1145 		return (EIO);
1146 
1147 	case NVME_CQE_SC_GEN_NVM_NS_NOTRDY:
1148 		/* Namespace Not Ready */
1149 		atomic_inc_32(&cmd->nc_nvme->n_nvm_ns_notrdy);
1150 		if (cmd->nc_xfer != NULL)
1151 			bd_error(cmd->nc_xfer, BD_ERR_NTRDY);
1152 		return (EIO);
1153 
1154 	default:
1155 		return (nvme_check_unknown_cmd_status(cmd));
1156 	}
1157 }
1158 
1159 static int
1160 nvme_check_specific_cmd_status(nvme_cmd_t *cmd)
1161 {
1162 	nvme_cqe_t *cqe = &cmd->nc_cqe;
1163 
1164 	switch (cqe->cqe_sf.sf_sc) {
1165 	case NVME_CQE_SC_SPC_INV_CQ:
1166 		/* Completion Queue Invalid */
1167 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_CREATE_SQUEUE);
1168 		atomic_inc_32(&cmd->nc_nvme->n_inv_cq_err);
1169 		return (EINVAL);
1170 
1171 	case NVME_CQE_SC_SPC_INV_QID:
1172 		/* Invalid Queue Identifier */
1173 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_CREATE_SQUEUE ||
1174 		    cmd->nc_sqe.sqe_opc == NVME_OPC_DELETE_SQUEUE ||
1175 		    cmd->nc_sqe.sqe_opc == NVME_OPC_CREATE_CQUEUE ||
1176 		    cmd->nc_sqe.sqe_opc == NVME_OPC_DELETE_CQUEUE);
1177 		atomic_inc_32(&cmd->nc_nvme->n_inv_qid_err);
1178 		return (EINVAL);
1179 
1180 	case NVME_CQE_SC_SPC_MAX_QSZ_EXC:
1181 		/* Max Queue Size Exceeded */
1182 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_CREATE_SQUEUE ||
1183 		    cmd->nc_sqe.sqe_opc == NVME_OPC_CREATE_CQUEUE);
1184 		atomic_inc_32(&cmd->nc_nvme->n_max_qsz_exc);
1185 		return (EINVAL);
1186 
1187 	case NVME_CQE_SC_SPC_ABRT_CMD_EXC:
1188 		/* Abort Command Limit Exceeded */
1189 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_ABORT);
1190 		dev_err(cmd->nc_nvme->n_dip, CE_PANIC, "programming error: "
1191 		    "abort command limit exceeded in cmd %p", (void *)cmd);
1192 		return (0);
1193 
1194 	case NVME_CQE_SC_SPC_ASYNC_EVREQ_EXC:
1195 		/* Async Event Request Limit Exceeded */
1196 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_ASYNC_EVENT);
1197 		dev_err(cmd->nc_nvme->n_dip, CE_PANIC, "programming error: "
1198 		    "async event request limit exceeded in cmd %p",
1199 		    (void *)cmd);
1200 		return (0);
1201 
1202 	case NVME_CQE_SC_SPC_INV_INT_VECT:
1203 		/* Invalid Interrupt Vector */
1204 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_CREATE_CQUEUE);
1205 		atomic_inc_32(&cmd->nc_nvme->n_inv_int_vect);
1206 		return (EINVAL);
1207 
1208 	case NVME_CQE_SC_SPC_INV_LOG_PAGE:
1209 		/* Invalid Log Page */
1210 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_GET_LOG_PAGE);
1211 		atomic_inc_32(&cmd->nc_nvme->n_inv_log_page);
1212 		return (EINVAL);
1213 
1214 	case NVME_CQE_SC_SPC_INV_FORMAT:
1215 		/* Invalid Format */
1216 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_NVM_FORMAT);
1217 		atomic_inc_32(&cmd->nc_nvme->n_inv_format);
1218 		if (cmd->nc_xfer != NULL)
1219 			bd_error(cmd->nc_xfer, BD_ERR_ILLRQ);
1220 		return (EINVAL);
1221 
1222 	case NVME_CQE_SC_SPC_INV_Q_DEL:
1223 		/* Invalid Queue Deletion */
1224 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_DELETE_CQUEUE);
1225 		atomic_inc_32(&cmd->nc_nvme->n_inv_q_del);
1226 		return (EINVAL);
1227 
1228 	case NVME_CQE_SC_SPC_NVM_CNFL_ATTR:
1229 		/* Conflicting Attributes */
1230 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_NVM_DSET_MGMT ||
1231 		    cmd->nc_sqe.sqe_opc == NVME_OPC_NVM_READ ||
1232 		    cmd->nc_sqe.sqe_opc == NVME_OPC_NVM_WRITE);
1233 		atomic_inc_32(&cmd->nc_nvme->n_cnfl_attr);
1234 		if (cmd->nc_xfer != NULL)
1235 			bd_error(cmd->nc_xfer, BD_ERR_ILLRQ);
1236 		return (EINVAL);
1237 
1238 	case NVME_CQE_SC_SPC_NVM_INV_PROT:
1239 		/* Invalid Protection Information */
1240 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_NVM_COMPARE ||
1241 		    cmd->nc_sqe.sqe_opc == NVME_OPC_NVM_READ ||
1242 		    cmd->nc_sqe.sqe_opc == NVME_OPC_NVM_WRITE);
1243 		atomic_inc_32(&cmd->nc_nvme->n_inv_prot);
1244 		if (cmd->nc_xfer != NULL)
1245 			bd_error(cmd->nc_xfer, BD_ERR_ILLRQ);
1246 		return (EINVAL);
1247 
1248 	case NVME_CQE_SC_SPC_NVM_READONLY:
1249 		/* Write to Read Only Range */
1250 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_NVM_WRITE);
1251 		atomic_inc_32(&cmd->nc_nvme->n_readonly);
1252 		if (cmd->nc_xfer != NULL)
1253 			bd_error(cmd->nc_xfer, BD_ERR_ILLRQ);
1254 		return (EROFS);
1255 
1256 	default:
1257 		return (nvme_check_unknown_cmd_status(cmd));
1258 	}
1259 }
1260 
1261 static inline int
1262 nvme_check_cmd_status(nvme_cmd_t *cmd)
1263 {
1264 	nvme_cqe_t *cqe = &cmd->nc_cqe;
1265 
1266 	/*
1267 	 * Take a shortcut if the controller is dead, or if
1268 	 * command status indicates no error.
1269 	 */
1270 	if (cmd->nc_nvme->n_dead)
1271 		return (EIO);
1272 
1273 	if (cqe->cqe_sf.sf_sct == NVME_CQE_SCT_GENERIC &&
1274 	    cqe->cqe_sf.sf_sc == NVME_CQE_SC_GEN_SUCCESS)
1275 		return (0);
1276 
1277 	if (cqe->cqe_sf.sf_sct == NVME_CQE_SCT_GENERIC)
1278 		return (nvme_check_generic_cmd_status(cmd));
1279 	else if (cqe->cqe_sf.sf_sct == NVME_CQE_SCT_SPECIFIC)
1280 		return (nvme_check_specific_cmd_status(cmd));
1281 	else if (cqe->cqe_sf.sf_sct == NVME_CQE_SCT_INTEGRITY)
1282 		return (nvme_check_integrity_cmd_status(cmd));
1283 	else if (cqe->cqe_sf.sf_sct == NVME_CQE_SCT_VENDOR)
1284 		return (nvme_check_vendor_cmd_status(cmd));
1285 
1286 	return (nvme_check_unknown_cmd_status(cmd));
1287 }
1288 
1289 static int
1290 nvme_abort_cmd(nvme_cmd_t *abort_cmd, uint_t sec)
1291 {
1292 	nvme_t *nvme = abort_cmd->nc_nvme;
1293 	nvme_cmd_t *cmd = nvme_alloc_cmd(nvme, KM_SLEEP);
1294 	nvme_abort_cmd_t ac = { 0 };
1295 	int ret = 0;
1296 
1297 	sema_p(&nvme->n_abort_sema);
1298 
1299 	ac.b.ac_cid = abort_cmd->nc_sqe.sqe_cid;
1300 	ac.b.ac_sqid = abort_cmd->nc_sqid;
1301 
1302 	cmd->nc_sqid = 0;
1303 	cmd->nc_sqe.sqe_opc = NVME_OPC_ABORT;
1304 	cmd->nc_callback = nvme_wakeup_cmd;
1305 	cmd->nc_sqe.sqe_cdw10 = ac.r;
1306 
1307 	/*
1308 	 * Send the ABORT to the hardware. The ABORT command will return _after_
1309 	 * the aborted command has completed (aborted or otherwise), but since
1310 	 * we still hold the aborted command's mutex its callback hasn't been
1311 	 * processed yet.
1312 	 */
1313 	nvme_admin_cmd(cmd, sec);
1314 	sema_v(&nvme->n_abort_sema);
1315 
1316 	if ((ret = nvme_check_cmd_status(cmd)) != 0) {
1317 		dev_err(nvme->n_dip, CE_WARN,
1318 		    "!ABORT failed with sct = %x, sc = %x",
1319 		    cmd->nc_cqe.cqe_sf.sf_sct, cmd->nc_cqe.cqe_sf.sf_sc);
1320 		atomic_inc_32(&nvme->n_abort_failed);
1321 	} else {
1322 		dev_err(nvme->n_dip, CE_WARN,
1323 		    "!ABORT of command %d/%d %ssuccessful",
1324 		    abort_cmd->nc_sqe.sqe_cid, abort_cmd->nc_sqid,
1325 		    cmd->nc_cqe.cqe_dw0 & 1 ? "un" : "");
1326 		if ((cmd->nc_cqe.cqe_dw0 & 1) == 0)
1327 			atomic_inc_32(&nvme->n_cmd_aborted);
1328 	}
1329 
1330 	nvme_free_cmd(cmd);
1331 	return (ret);
1332 }
1333 
1334 /*
1335  * nvme_wait_cmd -- wait for command completion or timeout
1336  *
1337  * In case of a serious error or a timeout of the abort command the hardware
1338  * will be declared dead and FMA will be notified.
1339  */
1340 static void
1341 nvme_wait_cmd(nvme_cmd_t *cmd, uint_t sec)
1342 {
1343 	clock_t timeout = ddi_get_lbolt() + drv_usectohz(sec * MICROSEC);
1344 	nvme_t *nvme = cmd->nc_nvme;
1345 	nvme_reg_csts_t csts;
1346 	nvme_qpair_t *qp;
1347 
1348 	ASSERT(mutex_owned(&cmd->nc_mutex));
1349 
1350 	while (!cmd->nc_completed) {
1351 		if (cv_timedwait(&cmd->nc_cv, &cmd->nc_mutex, timeout) == -1)
1352 			break;
1353 	}
1354 
1355 	if (cmd->nc_completed)
1356 		return;
1357 
1358 	/*
1359 	 * The command timed out.
1360 	 *
1361 	 * Check controller for fatal status, any errors associated with the
1362 	 * register or DMA handle, or for a double timeout (abort command timed
1363 	 * out). If necessary log a warning and call FMA.
1364 	 */
1365 	csts.r = nvme_get32(nvme, NVME_REG_CSTS);
1366 	dev_err(nvme->n_dip, CE_WARN, "!command %d/%d timeout, "
1367 	    "OPC = %x, CFS = %d", cmd->nc_sqe.sqe_cid, cmd->nc_sqid,
1368 	    cmd->nc_sqe.sqe_opc, csts.b.csts_cfs);
1369 	atomic_inc_32(&nvme->n_cmd_timeout);
1370 
1371 	if (csts.b.csts_cfs ||
1372 	    nvme_check_regs_hdl(nvme) ||
1373 	    nvme_check_dma_hdl(cmd->nc_dma) ||
1374 	    cmd->nc_sqe.sqe_opc == NVME_OPC_ABORT) {
1375 		ddi_fm_service_impact(nvme->n_dip, DDI_SERVICE_LOST);
1376 		nvme->n_dead = B_TRUE;
1377 	} else if (nvme_abort_cmd(cmd, sec) == 0) {
1378 		/*
1379 		 * If the abort succeeded the command should complete
1380 		 * immediately with an appropriate status.
1381 		 */
1382 		while (!cmd->nc_completed)
1383 			cv_wait(&cmd->nc_cv, &cmd->nc_mutex);
1384 
1385 		return;
1386 	}
1387 
1388 	qp = nvme->n_ioq[cmd->nc_sqid];
1389 
1390 	mutex_enter(&qp->nq_mutex);
1391 	(void) nvme_unqueue_cmd(nvme, qp, cmd->nc_sqe.sqe_cid);
1392 	mutex_exit(&qp->nq_mutex);
1393 
1394 	/*
1395 	 * As we don't know what the presumed dead hardware might still do with
1396 	 * the DMA memory, we'll put the command on the lost commands list if it
1397 	 * has any DMA memory.
1398 	 */
1399 	if (cmd->nc_dma != NULL) {
1400 		mutex_enter(&nvme_lc_mutex);
1401 		list_insert_head(&nvme_lost_cmds, cmd);
1402 		mutex_exit(&nvme_lc_mutex);
1403 	}
1404 }
1405 
1406 static void
1407 nvme_wakeup_cmd(void *arg)
1408 {
1409 	nvme_cmd_t *cmd = arg;
1410 
1411 	mutex_enter(&cmd->nc_mutex);
1412 	cmd->nc_completed = B_TRUE;
1413 	cv_signal(&cmd->nc_cv);
1414 	mutex_exit(&cmd->nc_mutex);
1415 }
1416 
1417 static void
1418 nvme_async_event_task(void *arg)
1419 {
1420 	nvme_cmd_t *cmd = arg;
1421 	nvme_t *nvme = cmd->nc_nvme;
1422 	nvme_error_log_entry_t *error_log = NULL;
1423 	nvme_health_log_t *health_log = NULL;
1424 	size_t logsize = 0;
1425 	nvme_async_event_t event;
1426 
1427 	/*
1428 	 * Check for errors associated with the async request itself. The only
1429 	 * command-specific error is "async event limit exceeded", which
1430 	 * indicates a programming error in the driver and causes a panic in
1431 	 * nvme_check_cmd_status().
1432 	 *
1433 	 * Other possible errors are various scenarios where the async request
1434 	 * was aborted, or internal errors in the device. Internal errors are
1435 	 * reported to FMA, the command aborts need no special handling here.
1436 	 *
1437 	 * And finally, at least qemu nvme does not support async events,
1438 	 * and will return NVME_CQE_SC_GEN_INV_OPC | DNR. If so, we
1439 	 * will avoid posting async events.
1440 	 */
1441 
1442 	if (nvme_check_cmd_status(cmd) != 0) {
1443 		dev_err(cmd->nc_nvme->n_dip, CE_WARN,
1444 		    "!async event request returned failure, sct = %x, "
1445 		    "sc = %x, dnr = %d, m = %d", cmd->nc_cqe.cqe_sf.sf_sct,
1446 		    cmd->nc_cqe.cqe_sf.sf_sc, cmd->nc_cqe.cqe_sf.sf_dnr,
1447 		    cmd->nc_cqe.cqe_sf.sf_m);
1448 
1449 		if (cmd->nc_cqe.cqe_sf.sf_sct == NVME_CQE_SCT_GENERIC &&
1450 		    cmd->nc_cqe.cqe_sf.sf_sc == NVME_CQE_SC_GEN_INTERNAL_ERR) {
1451 			cmd->nc_nvme->n_dead = B_TRUE;
1452 			ddi_fm_service_impact(cmd->nc_nvme->n_dip,
1453 			    DDI_SERVICE_LOST);
1454 		}
1455 
1456 		if (cmd->nc_cqe.cqe_sf.sf_sct == NVME_CQE_SCT_GENERIC &&
1457 		    cmd->nc_cqe.cqe_sf.sf_sc == NVME_CQE_SC_GEN_INV_OPC &&
1458 		    cmd->nc_cqe.cqe_sf.sf_dnr == 1) {
1459 			nvme->n_async_event_supported = B_FALSE;
1460 		}
1461 
1462 		nvme_free_cmd(cmd);
1463 		return;
1464 	}
1465 
1466 
1467 	event.r = cmd->nc_cqe.cqe_dw0;
1468 
1469 	/* Clear CQE and re-submit the async request. */
1470 	bzero(&cmd->nc_cqe, sizeof (nvme_cqe_t));
1471 	nvme_submit_admin_cmd(nvme->n_adminq, cmd);
1472 
1473 	switch (event.b.ae_type) {
1474 	case NVME_ASYNC_TYPE_ERROR:
1475 		if (event.b.ae_logpage == NVME_LOGPAGE_ERROR) {
1476 			(void) nvme_get_logpage(nvme, (void **)&error_log,
1477 			    &logsize, event.b.ae_logpage);
1478 		} else {
1479 			dev_err(nvme->n_dip, CE_WARN, "!wrong logpage in "
1480 			    "async event reply: %d", event.b.ae_logpage);
1481 			atomic_inc_32(&nvme->n_wrong_logpage);
1482 		}
1483 
1484 		switch (event.b.ae_info) {
1485 		case NVME_ASYNC_ERROR_INV_SQ:
1486 			dev_err(nvme->n_dip, CE_PANIC, "programming error: "
1487 			    "invalid submission queue");
1488 			return;
1489 
1490 		case NVME_ASYNC_ERROR_INV_DBL:
1491 			dev_err(nvme->n_dip, CE_PANIC, "programming error: "
1492 			    "invalid doorbell write value");
1493 			return;
1494 
1495 		case NVME_ASYNC_ERROR_DIAGFAIL:
1496 			dev_err(nvme->n_dip, CE_WARN, "!diagnostic failure");
1497 			ddi_fm_service_impact(nvme->n_dip, DDI_SERVICE_LOST);
1498 			nvme->n_dead = B_TRUE;
1499 			atomic_inc_32(&nvme->n_diagfail_event);
1500 			break;
1501 
1502 		case NVME_ASYNC_ERROR_PERSISTENT:
1503 			dev_err(nvme->n_dip, CE_WARN, "!persistent internal "
1504 			    "device error");
1505 			ddi_fm_service_impact(nvme->n_dip, DDI_SERVICE_LOST);
1506 			nvme->n_dead = B_TRUE;
1507 			atomic_inc_32(&nvme->n_persistent_event);
1508 			break;
1509 
1510 		case NVME_ASYNC_ERROR_TRANSIENT:
1511 			dev_err(nvme->n_dip, CE_WARN, "!transient internal "
1512 			    "device error");
1513 			/* TODO: send ereport */
1514 			atomic_inc_32(&nvme->n_transient_event);
1515 			break;
1516 
1517 		case NVME_ASYNC_ERROR_FW_LOAD:
1518 			dev_err(nvme->n_dip, CE_WARN,
1519 			    "!firmware image load error");
1520 			atomic_inc_32(&nvme->n_fw_load_event);
1521 			break;
1522 		}
1523 		break;
1524 
1525 	case NVME_ASYNC_TYPE_HEALTH:
1526 		if (event.b.ae_logpage == NVME_LOGPAGE_HEALTH) {
1527 			(void) nvme_get_logpage(nvme, (void **)&health_log,
1528 			    &logsize, event.b.ae_logpage, -1);
1529 		} else {
1530 			dev_err(nvme->n_dip, CE_WARN, "!wrong logpage in "
1531 			    "async event reply: %d", event.b.ae_logpage);
1532 			atomic_inc_32(&nvme->n_wrong_logpage);
1533 		}
1534 
1535 		switch (event.b.ae_info) {
1536 		case NVME_ASYNC_HEALTH_RELIABILITY:
1537 			dev_err(nvme->n_dip, CE_WARN,
1538 			    "!device reliability compromised");
1539 			/* TODO: send ereport */
1540 			atomic_inc_32(&nvme->n_reliability_event);
1541 			break;
1542 
1543 		case NVME_ASYNC_HEALTH_TEMPERATURE:
1544 			dev_err(nvme->n_dip, CE_WARN,
1545 			    "!temperature above threshold");
1546 			/* TODO: send ereport */
1547 			atomic_inc_32(&nvme->n_temperature_event);
1548 			break;
1549 
1550 		case NVME_ASYNC_HEALTH_SPARE:
1551 			dev_err(nvme->n_dip, CE_WARN,
1552 			    "!spare space below threshold");
1553 			/* TODO: send ereport */
1554 			atomic_inc_32(&nvme->n_spare_event);
1555 			break;
1556 		}
1557 		break;
1558 
1559 	case NVME_ASYNC_TYPE_VENDOR:
1560 		dev_err(nvme->n_dip, CE_WARN, "!vendor specific async event "
1561 		    "received, info = %x, logpage = %x", event.b.ae_info,
1562 		    event.b.ae_logpage);
1563 		atomic_inc_32(&nvme->n_vendor_event);
1564 		break;
1565 
1566 	default:
1567 		dev_err(nvme->n_dip, CE_WARN, "!unknown async event received, "
1568 		    "type = %x, info = %x, logpage = %x", event.b.ae_type,
1569 		    event.b.ae_info, event.b.ae_logpage);
1570 		atomic_inc_32(&nvme->n_unknown_event);
1571 		break;
1572 	}
1573 
1574 	if (error_log)
1575 		kmem_free(error_log, logsize);
1576 
1577 	if (health_log)
1578 		kmem_free(health_log, logsize);
1579 }
1580 
1581 static void
1582 nvme_admin_cmd(nvme_cmd_t *cmd, int sec)
1583 {
1584 	mutex_enter(&cmd->nc_mutex);
1585 	nvme_submit_admin_cmd(cmd->nc_nvme->n_adminq, cmd);
1586 	nvme_wait_cmd(cmd, sec);
1587 	mutex_exit(&cmd->nc_mutex);
1588 }
1589 
1590 static void
1591 nvme_async_event(nvme_t *nvme)
1592 {
1593 	nvme_cmd_t *cmd;
1594 
1595 	cmd = nvme_alloc_cmd(nvme, KM_SLEEP);
1596 	cmd->nc_sqid = 0;
1597 	cmd->nc_sqe.sqe_opc = NVME_OPC_ASYNC_EVENT;
1598 	cmd->nc_callback = nvme_async_event_task;
1599 	cmd->nc_dontpanic = B_TRUE;
1600 
1601 	nvme_submit_admin_cmd(nvme->n_adminq, cmd);
1602 }
1603 
1604 static int
1605 nvme_format_nvm(nvme_t *nvme, uint32_t nsid, uint8_t lbaf, boolean_t ms,
1606     uint8_t pi, boolean_t pil, uint8_t ses)
1607 {
1608 	nvme_cmd_t *cmd = nvme_alloc_cmd(nvme, KM_SLEEP);
1609 	nvme_format_nvm_t format_nvm = { 0 };
1610 	int ret;
1611 
1612 	format_nvm.b.fm_lbaf = lbaf & 0xf;
1613 	format_nvm.b.fm_ms = ms ? 1 : 0;
1614 	format_nvm.b.fm_pi = pi & 0x7;
1615 	format_nvm.b.fm_pil = pil ? 1 : 0;
1616 	format_nvm.b.fm_ses = ses & 0x7;
1617 
1618 	cmd->nc_sqid = 0;
1619 	cmd->nc_callback = nvme_wakeup_cmd;
1620 	cmd->nc_sqe.sqe_nsid = nsid;
1621 	cmd->nc_sqe.sqe_opc = NVME_OPC_NVM_FORMAT;
1622 	cmd->nc_sqe.sqe_cdw10 = format_nvm.r;
1623 
1624 	/*
1625 	 * Some devices like Samsung SM951 don't allow formatting of all
1626 	 * namespaces in one command. Handle that gracefully.
1627 	 */
1628 	if (nsid == (uint32_t)-1)
1629 		cmd->nc_dontpanic = B_TRUE;
1630 
1631 	nvme_admin_cmd(cmd, nvme_format_cmd_timeout);
1632 
1633 	if ((ret = nvme_check_cmd_status(cmd)) != 0) {
1634 		dev_err(nvme->n_dip, CE_WARN,
1635 		    "!FORMAT failed with sct = %x, sc = %x",
1636 		    cmd->nc_cqe.cqe_sf.sf_sct, cmd->nc_cqe.cqe_sf.sf_sc);
1637 	}
1638 
1639 	nvme_free_cmd(cmd);
1640 	return (ret);
1641 }
1642 
1643 static int
1644 nvme_get_logpage(nvme_t *nvme, void **buf, size_t *bufsize, uint8_t logpage,
1645     ...)
1646 {
1647 	nvme_cmd_t *cmd = nvme_alloc_cmd(nvme, KM_SLEEP);
1648 	nvme_getlogpage_t getlogpage = { 0 };
1649 	va_list ap;
1650 	int ret;
1651 
1652 	va_start(ap, logpage);
1653 
1654 	cmd->nc_sqid = 0;
1655 	cmd->nc_callback = nvme_wakeup_cmd;
1656 	cmd->nc_sqe.sqe_opc = NVME_OPC_GET_LOG_PAGE;
1657 
1658 	getlogpage.b.lp_lid = logpage;
1659 
1660 	switch (logpage) {
1661 	case NVME_LOGPAGE_ERROR:
1662 		cmd->nc_sqe.sqe_nsid = (uint32_t)-1;
1663 		/*
1664 		 * The GET LOG PAGE command can use at most 2 pages to return
1665 		 * data, PRP lists are not supported.
1666 		 */
1667 		*bufsize = MIN(2 * nvme->n_pagesize,
1668 		    nvme->n_error_log_len * sizeof (nvme_error_log_entry_t));
1669 		break;
1670 
1671 	case NVME_LOGPAGE_HEALTH:
1672 		cmd->nc_sqe.sqe_nsid = va_arg(ap, uint32_t);
1673 		*bufsize = sizeof (nvme_health_log_t);
1674 		break;
1675 
1676 	case NVME_LOGPAGE_FWSLOT:
1677 		cmd->nc_sqe.sqe_nsid = (uint32_t)-1;
1678 		*bufsize = sizeof (nvme_fwslot_log_t);
1679 		break;
1680 
1681 	default:
1682 		dev_err(nvme->n_dip, CE_WARN, "!unknown log page requested: %d",
1683 		    logpage);
1684 		atomic_inc_32(&nvme->n_unknown_logpage);
1685 		ret = EINVAL;
1686 		goto fail;
1687 	}
1688 
1689 	va_end(ap);
1690 
1691 	getlogpage.b.lp_numd = *bufsize / sizeof (uint32_t) - 1;
1692 
1693 	cmd->nc_sqe.sqe_cdw10 = getlogpage.r;
1694 
1695 	if (nvme_zalloc_dma(nvme, getlogpage.b.lp_numd * sizeof (uint32_t),
1696 	    DDI_DMA_READ, &nvme->n_prp_dma_attr, &cmd->nc_dma) != DDI_SUCCESS) {
1697 		dev_err(nvme->n_dip, CE_WARN,
1698 		    "!nvme_zalloc_dma failed for GET LOG PAGE");
1699 		ret = ENOMEM;
1700 		goto fail;
1701 	}
1702 
1703 	if (cmd->nc_dma->nd_ncookie > 2) {
1704 		dev_err(nvme->n_dip, CE_WARN,
1705 		    "!too many DMA cookies for GET LOG PAGE");
1706 		atomic_inc_32(&nvme->n_too_many_cookies);
1707 		ret = ENOMEM;
1708 		goto fail;
1709 	}
1710 
1711 	cmd->nc_sqe.sqe_dptr.d_prp[0] = cmd->nc_dma->nd_cookie.dmac_laddress;
1712 	if (cmd->nc_dma->nd_ncookie > 1) {
1713 		ddi_dma_nextcookie(cmd->nc_dma->nd_dmah,
1714 		    &cmd->nc_dma->nd_cookie);
1715 		cmd->nc_sqe.sqe_dptr.d_prp[1] =
1716 		    cmd->nc_dma->nd_cookie.dmac_laddress;
1717 	}
1718 
1719 	nvme_admin_cmd(cmd, nvme_admin_cmd_timeout);
1720 
1721 	if ((ret = nvme_check_cmd_status(cmd)) != 0) {
1722 		dev_err(nvme->n_dip, CE_WARN,
1723 		    "!GET LOG PAGE failed with sct = %x, sc = %x",
1724 		    cmd->nc_cqe.cqe_sf.sf_sct, cmd->nc_cqe.cqe_sf.sf_sc);
1725 		goto fail;
1726 	}
1727 
1728 	*buf = kmem_alloc(*bufsize, KM_SLEEP);
1729 	bcopy(cmd->nc_dma->nd_memp, *buf, *bufsize);
1730 
1731 fail:
1732 	nvme_free_cmd(cmd);
1733 
1734 	return (ret);
1735 }
1736 
1737 static int
1738 nvme_identify(nvme_t *nvme, uint32_t nsid, void **buf)
1739 {
1740 	nvme_cmd_t *cmd = nvme_alloc_cmd(nvme, KM_SLEEP);
1741 	int ret;
1742 
1743 	if (buf == NULL)
1744 		return (EINVAL);
1745 
1746 	cmd->nc_sqid = 0;
1747 	cmd->nc_callback = nvme_wakeup_cmd;
1748 	cmd->nc_sqe.sqe_opc = NVME_OPC_IDENTIFY;
1749 	cmd->nc_sqe.sqe_nsid = nsid;
1750 	cmd->nc_sqe.sqe_cdw10 = nsid ? NVME_IDENTIFY_NSID : NVME_IDENTIFY_CTRL;
1751 
1752 	if (nvme_zalloc_dma(nvme, NVME_IDENTIFY_BUFSIZE, DDI_DMA_READ,
1753 	    &nvme->n_prp_dma_attr, &cmd->nc_dma) != DDI_SUCCESS) {
1754 		dev_err(nvme->n_dip, CE_WARN,
1755 		    "!nvme_zalloc_dma failed for IDENTIFY");
1756 		ret = ENOMEM;
1757 		goto fail;
1758 	}
1759 
1760 	if (cmd->nc_dma->nd_ncookie > 2) {
1761 		dev_err(nvme->n_dip, CE_WARN,
1762 		    "!too many DMA cookies for IDENTIFY");
1763 		atomic_inc_32(&nvme->n_too_many_cookies);
1764 		ret = ENOMEM;
1765 		goto fail;
1766 	}
1767 
1768 	cmd->nc_sqe.sqe_dptr.d_prp[0] = cmd->nc_dma->nd_cookie.dmac_laddress;
1769 	if (cmd->nc_dma->nd_ncookie > 1) {
1770 		ddi_dma_nextcookie(cmd->nc_dma->nd_dmah,
1771 		    &cmd->nc_dma->nd_cookie);
1772 		cmd->nc_sqe.sqe_dptr.d_prp[1] =
1773 		    cmd->nc_dma->nd_cookie.dmac_laddress;
1774 	}
1775 
1776 	nvme_admin_cmd(cmd, nvme_admin_cmd_timeout);
1777 
1778 	if ((ret = nvme_check_cmd_status(cmd)) != 0) {
1779 		dev_err(nvme->n_dip, CE_WARN,
1780 		    "!IDENTIFY failed with sct = %x, sc = %x",
1781 		    cmd->nc_cqe.cqe_sf.sf_sct, cmd->nc_cqe.cqe_sf.sf_sc);
1782 		goto fail;
1783 	}
1784 
1785 	*buf = kmem_alloc(NVME_IDENTIFY_BUFSIZE, KM_SLEEP);
1786 	bcopy(cmd->nc_dma->nd_memp, *buf, NVME_IDENTIFY_BUFSIZE);
1787 
1788 fail:
1789 	nvme_free_cmd(cmd);
1790 
1791 	return (ret);
1792 }
1793 
1794 static int
1795 nvme_set_features(nvme_t *nvme, uint32_t nsid, uint8_t feature, uint32_t val,
1796     uint32_t *res)
1797 {
1798 	_NOTE(ARGUNUSED(nsid));
1799 	nvme_cmd_t *cmd = nvme_alloc_cmd(nvme, KM_SLEEP);
1800 	int ret = EINVAL;
1801 
1802 	ASSERT(res != NULL);
1803 
1804 	cmd->nc_sqid = 0;
1805 	cmd->nc_callback = nvme_wakeup_cmd;
1806 	cmd->nc_sqe.sqe_opc = NVME_OPC_SET_FEATURES;
1807 	cmd->nc_sqe.sqe_cdw10 = feature;
1808 	cmd->nc_sqe.sqe_cdw11 = val;
1809 
1810 	switch (feature) {
1811 	case NVME_FEAT_WRITE_CACHE:
1812 		if (!nvme->n_write_cache_present)
1813 			goto fail;
1814 		break;
1815 
1816 	case NVME_FEAT_NQUEUES:
1817 		break;
1818 
1819 	default:
1820 		goto fail;
1821 	}
1822 
1823 	nvme_admin_cmd(cmd, nvme_admin_cmd_timeout);
1824 
1825 	if ((ret = nvme_check_cmd_status(cmd)) != 0) {
1826 		dev_err(nvme->n_dip, CE_WARN,
1827 		    "!SET FEATURES %d failed with sct = %x, sc = %x",
1828 		    feature, cmd->nc_cqe.cqe_sf.sf_sct,
1829 		    cmd->nc_cqe.cqe_sf.sf_sc);
1830 		goto fail;
1831 	}
1832 
1833 	*res = cmd->nc_cqe.cqe_dw0;
1834 
1835 fail:
1836 	nvme_free_cmd(cmd);
1837 	return (ret);
1838 }
1839 
1840 static int
1841 nvme_get_features(nvme_t *nvme, uint32_t nsid, uint8_t feature, uint32_t *res,
1842     void **buf, size_t *bufsize)
1843 {
1844 	nvme_cmd_t *cmd = nvme_alloc_cmd(nvme, KM_SLEEP);
1845 	int ret = EINVAL;
1846 
1847 	ASSERT(res != NULL);
1848 
1849 	if (bufsize != NULL)
1850 		*bufsize = 0;
1851 
1852 	cmd->nc_sqid = 0;
1853 	cmd->nc_callback = nvme_wakeup_cmd;
1854 	cmd->nc_sqe.sqe_opc = NVME_OPC_GET_FEATURES;
1855 	cmd->nc_sqe.sqe_cdw10 = feature;
1856 	cmd->nc_sqe.sqe_cdw11 = *res;
1857 
1858 	/*
1859 	 * For some of the optional features there doesn't seem to be a method
1860 	 * of detecting whether it is supported other than using it.  This will
1861 	 * cause "Invalid Field in Command" error, which is normally considered
1862 	 * a programming error.  Set the nc_dontpanic flag to override the panic
1863 	 * in nvme_check_generic_cmd_status().
1864 	 */
1865 	switch (feature) {
1866 	case NVME_FEAT_ARBITRATION:
1867 	case NVME_FEAT_POWER_MGMT:
1868 	case NVME_FEAT_TEMPERATURE:
1869 	case NVME_FEAT_ERROR:
1870 	case NVME_FEAT_NQUEUES:
1871 	case NVME_FEAT_INTR_COAL:
1872 	case NVME_FEAT_INTR_VECT:
1873 	case NVME_FEAT_WRITE_ATOM:
1874 	case NVME_FEAT_ASYNC_EVENT:
1875 		break;
1876 
1877 	case NVME_FEAT_WRITE_CACHE:
1878 		if (!nvme->n_write_cache_present)
1879 			goto fail;
1880 		break;
1881 
1882 	case NVME_FEAT_LBA_RANGE:
1883 		if (!nvme->n_lba_range_supported)
1884 			goto fail;
1885 
1886 		cmd->nc_dontpanic = B_TRUE;
1887 		cmd->nc_sqe.sqe_nsid = nsid;
1888 		ASSERT(bufsize != NULL);
1889 		*bufsize = NVME_LBA_RANGE_BUFSIZE;
1890 		break;
1891 
1892 	case NVME_FEAT_AUTO_PST:
1893 		if (!nvme->n_auto_pst_supported)
1894 			goto fail;
1895 
1896 		ASSERT(bufsize != NULL);
1897 		*bufsize = NVME_AUTO_PST_BUFSIZE;
1898 		break;
1899 
1900 	case NVME_FEAT_PROGRESS:
1901 		if (!nvme->n_progress_supported)
1902 			goto fail;
1903 
1904 		cmd->nc_dontpanic = B_TRUE;
1905 		break;
1906 
1907 	default:
1908 		goto fail;
1909 	}
1910 
1911 	if (bufsize != NULL && *bufsize != 0) {
1912 		if (nvme_zalloc_dma(nvme, *bufsize, DDI_DMA_READ,
1913 		    &nvme->n_prp_dma_attr, &cmd->nc_dma) != DDI_SUCCESS) {
1914 			dev_err(nvme->n_dip, CE_WARN,
1915 			    "!nvme_zalloc_dma failed for GET FEATURES");
1916 			ret = ENOMEM;
1917 			goto fail;
1918 		}
1919 
1920 		if (cmd->nc_dma->nd_ncookie > 2) {
1921 			dev_err(nvme->n_dip, CE_WARN,
1922 			    "!too many DMA cookies for GET FEATURES");
1923 			atomic_inc_32(&nvme->n_too_many_cookies);
1924 			ret = ENOMEM;
1925 			goto fail;
1926 		}
1927 
1928 		cmd->nc_sqe.sqe_dptr.d_prp[0] =
1929 		    cmd->nc_dma->nd_cookie.dmac_laddress;
1930 		if (cmd->nc_dma->nd_ncookie > 1) {
1931 			ddi_dma_nextcookie(cmd->nc_dma->nd_dmah,
1932 			    &cmd->nc_dma->nd_cookie);
1933 			cmd->nc_sqe.sqe_dptr.d_prp[1] =
1934 			    cmd->nc_dma->nd_cookie.dmac_laddress;
1935 		}
1936 	}
1937 
1938 	nvme_admin_cmd(cmd, nvme_admin_cmd_timeout);
1939 
1940 	if ((ret = nvme_check_cmd_status(cmd)) != 0) {
1941 		boolean_t known = B_TRUE;
1942 
1943 		/* Check if this is unsupported optional feature */
1944 		if (cmd->nc_cqe.cqe_sf.sf_sct == NVME_CQE_SCT_GENERIC &&
1945 		    cmd->nc_cqe.cqe_sf.sf_sc == NVME_CQE_SC_GEN_INV_FLD) {
1946 			switch (feature) {
1947 			case NVME_FEAT_LBA_RANGE:
1948 				nvme->n_lba_range_supported = B_FALSE;
1949 				break;
1950 			case NVME_FEAT_PROGRESS:
1951 				nvme->n_progress_supported = B_FALSE;
1952 				break;
1953 			default:
1954 				known = B_FALSE;
1955 				break;
1956 			}
1957 		} else {
1958 			known = B_FALSE;
1959 		}
1960 
1961 		/* Report the error otherwise */
1962 		if (!known) {
1963 			dev_err(nvme->n_dip, CE_WARN,
1964 			    "!GET FEATURES %d failed with sct = %x, sc = %x",
1965 			    feature, cmd->nc_cqe.cqe_sf.sf_sct,
1966 			    cmd->nc_cqe.cqe_sf.sf_sc);
1967 		}
1968 
1969 		goto fail;
1970 	}
1971 
1972 	if (bufsize != NULL && *bufsize != 0) {
1973 		ASSERT(buf != NULL);
1974 		*buf = kmem_alloc(*bufsize, KM_SLEEP);
1975 		bcopy(cmd->nc_dma->nd_memp, *buf, *bufsize);
1976 	}
1977 
1978 	*res = cmd->nc_cqe.cqe_dw0;
1979 
1980 fail:
1981 	nvme_free_cmd(cmd);
1982 	return (ret);
1983 }
1984 
1985 static int
1986 nvme_write_cache_set(nvme_t *nvme, boolean_t enable)
1987 {
1988 	nvme_write_cache_t nwc = { 0 };
1989 
1990 	if (enable)
1991 		nwc.b.wc_wce = 1;
1992 
1993 	return (nvme_set_features(nvme, 0, NVME_FEAT_WRITE_CACHE, nwc.r,
1994 	    &nwc.r));
1995 }
1996 
1997 static int
1998 nvme_set_nqueues(nvme_t *nvme, uint16_t *nqueues)
1999 {
2000 	nvme_nqueues_t nq = { 0 };
2001 	int ret;
2002 
2003 	nq.b.nq_nsq = nq.b.nq_ncq = *nqueues - 1;
2004 
2005 	ret = nvme_set_features(nvme, 0, NVME_FEAT_NQUEUES, nq.r, &nq.r);
2006 
2007 	if (ret == 0) {
2008 		/*
2009 		 * Always use the same number of submission and completion
2010 		 * queues, and never use more than the requested number of
2011 		 * queues.
2012 		 */
2013 		*nqueues = MIN(*nqueues, MIN(nq.b.nq_nsq, nq.b.nq_ncq) + 1);
2014 	}
2015 
2016 	return (ret);
2017 }
2018 
2019 static int
2020 nvme_create_io_qpair(nvme_t *nvme, nvme_qpair_t *qp, uint16_t idx)
2021 {
2022 	nvme_cmd_t *cmd = nvme_alloc_cmd(nvme, KM_SLEEP);
2023 	nvme_create_queue_dw10_t dw10 = { 0 };
2024 	nvme_create_cq_dw11_t c_dw11 = { 0 };
2025 	nvme_create_sq_dw11_t s_dw11 = { 0 };
2026 	int ret;
2027 
2028 	dw10.b.q_qid = idx;
2029 	dw10.b.q_qsize = qp->nq_nentry - 1;
2030 
2031 	c_dw11.b.cq_pc = 1;
2032 	c_dw11.b.cq_ien = 1;
2033 	c_dw11.b.cq_iv = idx % nvme->n_intr_cnt;
2034 
2035 	cmd->nc_sqid = 0;
2036 	cmd->nc_callback = nvme_wakeup_cmd;
2037 	cmd->nc_sqe.sqe_opc = NVME_OPC_CREATE_CQUEUE;
2038 	cmd->nc_sqe.sqe_cdw10 = dw10.r;
2039 	cmd->nc_sqe.sqe_cdw11 = c_dw11.r;
2040 	cmd->nc_sqe.sqe_dptr.d_prp[0] = qp->nq_cqdma->nd_cookie.dmac_laddress;
2041 
2042 	nvme_admin_cmd(cmd, nvme_admin_cmd_timeout);
2043 
2044 	if ((ret = nvme_check_cmd_status(cmd)) != 0) {
2045 		dev_err(nvme->n_dip, CE_WARN,
2046 		    "!CREATE CQUEUE failed with sct = %x, sc = %x",
2047 		    cmd->nc_cqe.cqe_sf.sf_sct, cmd->nc_cqe.cqe_sf.sf_sc);
2048 		goto fail;
2049 	}
2050 
2051 	nvme_free_cmd(cmd);
2052 
2053 	s_dw11.b.sq_pc = 1;
2054 	s_dw11.b.sq_cqid = idx;
2055 
2056 	cmd = nvme_alloc_cmd(nvme, KM_SLEEP);
2057 	cmd->nc_sqid = 0;
2058 	cmd->nc_callback = nvme_wakeup_cmd;
2059 	cmd->nc_sqe.sqe_opc = NVME_OPC_CREATE_SQUEUE;
2060 	cmd->nc_sqe.sqe_cdw10 = dw10.r;
2061 	cmd->nc_sqe.sqe_cdw11 = s_dw11.r;
2062 	cmd->nc_sqe.sqe_dptr.d_prp[0] = qp->nq_sqdma->nd_cookie.dmac_laddress;
2063 
2064 	nvme_admin_cmd(cmd, nvme_admin_cmd_timeout);
2065 
2066 	if ((ret = nvme_check_cmd_status(cmd)) != 0) {
2067 		dev_err(nvme->n_dip, CE_WARN,
2068 		    "!CREATE SQUEUE failed with sct = %x, sc = %x",
2069 		    cmd->nc_cqe.cqe_sf.sf_sct, cmd->nc_cqe.cqe_sf.sf_sc);
2070 		goto fail;
2071 	}
2072 
2073 fail:
2074 	nvme_free_cmd(cmd);
2075 
2076 	return (ret);
2077 }
2078 
2079 static boolean_t
2080 nvme_reset(nvme_t *nvme, boolean_t quiesce)
2081 {
2082 	nvme_reg_csts_t csts;
2083 	int i;
2084 
2085 	nvme_put32(nvme, NVME_REG_CC, 0);
2086 
2087 	csts.r = nvme_get32(nvme, NVME_REG_CSTS);
2088 	if (csts.b.csts_rdy == 1) {
2089 		nvme_put32(nvme, NVME_REG_CC, 0);
2090 		for (i = 0; i != nvme->n_timeout * 10; i++) {
2091 			csts.r = nvme_get32(nvme, NVME_REG_CSTS);
2092 			if (csts.b.csts_rdy == 0)
2093 				break;
2094 
2095 			if (quiesce)
2096 				drv_usecwait(50000);
2097 			else
2098 				delay(drv_usectohz(50000));
2099 		}
2100 	}
2101 
2102 	nvme_put32(nvme, NVME_REG_AQA, 0);
2103 	nvme_put32(nvme, NVME_REG_ASQ, 0);
2104 	nvme_put32(nvme, NVME_REG_ACQ, 0);
2105 
2106 	csts.r = nvme_get32(nvme, NVME_REG_CSTS);
2107 	return (csts.b.csts_rdy == 0 ? B_TRUE : B_FALSE);
2108 }
2109 
2110 static void
2111 nvme_shutdown(nvme_t *nvme, int mode, boolean_t quiesce)
2112 {
2113 	nvme_reg_cc_t cc;
2114 	nvme_reg_csts_t csts;
2115 	int i;
2116 
2117 	ASSERT(mode == NVME_CC_SHN_NORMAL || mode == NVME_CC_SHN_ABRUPT);
2118 
2119 	cc.r = nvme_get32(nvme, NVME_REG_CC);
2120 	cc.b.cc_shn = mode & 0x3;
2121 	nvme_put32(nvme, NVME_REG_CC, cc.r);
2122 
2123 	for (i = 0; i != 10; i++) {
2124 		csts.r = nvme_get32(nvme, NVME_REG_CSTS);
2125 		if (csts.b.csts_shst == NVME_CSTS_SHN_COMPLETE)
2126 			break;
2127 
2128 		if (quiesce)
2129 			drv_usecwait(100000);
2130 		else
2131 			delay(drv_usectohz(100000));
2132 	}
2133 }
2134 
2135 
2136 static void
2137 nvme_prepare_devid(nvme_t *nvme, uint32_t nsid)
2138 {
2139 	/*
2140 	 * Section 7.7 of the spec describes how to get a unique ID for
2141 	 * the controller: the vendor ID, the model name and the serial
2142 	 * number shall be unique when combined.
2143 	 *
2144 	 * If a namespace has no EUI64 we use the above and add the hex
2145 	 * namespace ID to get a unique ID for the namespace.
2146 	 */
2147 	char model[sizeof (nvme->n_idctl->id_model) + 1];
2148 	char serial[sizeof (nvme->n_idctl->id_serial) + 1];
2149 
2150 	bcopy(nvme->n_idctl->id_model, model, sizeof (nvme->n_idctl->id_model));
2151 	bcopy(nvme->n_idctl->id_serial, serial,
2152 	    sizeof (nvme->n_idctl->id_serial));
2153 
2154 	model[sizeof (nvme->n_idctl->id_model)] = '\0';
2155 	serial[sizeof (nvme->n_idctl->id_serial)] = '\0';
2156 
2157 	nvme->n_ns[nsid - 1].ns_devid = kmem_asprintf("%4X-%s-%s-%X",
2158 	    nvme->n_idctl->id_vid, model, serial, nsid);
2159 }
2160 
2161 static int
2162 nvme_init_ns(nvme_t *nvme, int nsid)
2163 {
2164 	nvme_namespace_t *ns = &nvme->n_ns[nsid - 1];
2165 	nvme_identify_nsid_t *idns;
2166 	int last_rp;
2167 
2168 	ns->ns_nvme = nvme;
2169 
2170 	if (nvme_identify(nvme, nsid, (void **)&idns) != 0) {
2171 		dev_err(nvme->n_dip, CE_WARN,
2172 		    "!failed to identify namespace %d", nsid);
2173 		return (DDI_FAILURE);
2174 	}
2175 
2176 	ns->ns_idns = idns;
2177 	ns->ns_id = nsid;
2178 	ns->ns_block_count = idns->id_nsize;
2179 	ns->ns_block_size =
2180 	    1 << idns->id_lbaf[idns->id_flbas.lba_format].lbaf_lbads;
2181 	ns->ns_best_block_size = ns->ns_block_size;
2182 
2183 	/*
2184 	 * Get the EUI64 if present. Use it for devid and device node names.
2185 	 */
2186 	if (NVME_VERSION_ATLEAST(&nvme->n_version, 1, 1))
2187 		bcopy(idns->id_eui64, ns->ns_eui64, sizeof (ns->ns_eui64));
2188 
2189 	/*LINTED: E_BAD_PTR_CAST_ALIGN*/
2190 	if (*(uint64_t *)ns->ns_eui64 != 0) {
2191 		uint8_t *eui64 = ns->ns_eui64;
2192 
2193 		(void) snprintf(ns->ns_name, sizeof (ns->ns_name),
2194 		    "%02x%02x%02x%02x%02x%02x%02x%02x",
2195 		    eui64[0], eui64[1], eui64[2], eui64[3],
2196 		    eui64[4], eui64[5], eui64[6], eui64[7]);
2197 	} else {
2198 		(void) snprintf(ns->ns_name, sizeof (ns->ns_name), "%d",
2199 		    ns->ns_id);
2200 
2201 		nvme_prepare_devid(nvme, ns->ns_id);
2202 	}
2203 
2204 	/*
2205 	 * Find the LBA format with no metadata and the best relative
2206 	 * performance. A value of 3 means "degraded", 0 is best.
2207 	 */
2208 	last_rp = 3;
2209 	for (int j = 0; j <= idns->id_nlbaf; j++) {
2210 		if (idns->id_lbaf[j].lbaf_lbads == 0)
2211 			break;
2212 		if (idns->id_lbaf[j].lbaf_ms != 0)
2213 			continue;
2214 		if (idns->id_lbaf[j].lbaf_rp >= last_rp)
2215 			continue;
2216 		last_rp = idns->id_lbaf[j].lbaf_rp;
2217 		ns->ns_best_block_size =
2218 		    1 << idns->id_lbaf[j].lbaf_lbads;
2219 	}
2220 
2221 	if (ns->ns_best_block_size < nvme->n_min_block_size)
2222 		ns->ns_best_block_size = nvme->n_min_block_size;
2223 
2224 	/*
2225 	 * We currently don't support namespaces that use either:
2226 	 * - protection information
2227 	 * - illegal block size (< 512)
2228 	 */
2229 	if (idns->id_dps.dp_pinfo) {
2230 		dev_err(nvme->n_dip, CE_WARN,
2231 		    "!ignoring namespace %d, unsupported feature: "
2232 		    "pinfo = %d", nsid, idns->id_dps.dp_pinfo);
2233 		ns->ns_ignore = B_TRUE;
2234 	} else if (ns->ns_block_size < 512) {
2235 		dev_err(nvme->n_dip, CE_WARN,
2236 		    "!ignoring namespace %d, unsupported block size %"PRIu64,
2237 		    nsid, (uint64_t)ns->ns_block_size);
2238 		ns->ns_ignore = B_TRUE;
2239 	} else {
2240 		ns->ns_ignore = B_FALSE;
2241 	}
2242 
2243 	return (DDI_SUCCESS);
2244 }
2245 
2246 static int
2247 nvme_init(nvme_t *nvme)
2248 {
2249 	nvme_reg_cc_t cc = { 0 };
2250 	nvme_reg_aqa_t aqa = { 0 };
2251 	nvme_reg_asq_t asq = { 0 };
2252 	nvme_reg_acq_t acq = { 0 };
2253 	nvme_reg_cap_t cap;
2254 	nvme_reg_vs_t vs;
2255 	nvme_reg_csts_t csts;
2256 	int i = 0;
2257 	uint16_t nqueues;
2258 	char model[sizeof (nvme->n_idctl->id_model) + 1];
2259 	char *vendor, *product;
2260 
2261 	/* Check controller version */
2262 	vs.r = nvme_get32(nvme, NVME_REG_VS);
2263 	nvme->n_version.v_major = vs.b.vs_mjr;
2264 	nvme->n_version.v_minor = vs.b.vs_mnr;
2265 	dev_err(nvme->n_dip, CE_CONT, "?NVMe spec version %d.%d",
2266 	    nvme->n_version.v_major, nvme->n_version.v_minor);
2267 
2268 	if (NVME_VERSION_HIGHER(&nvme->n_version,
2269 	    nvme_version_major, nvme_version_minor)) {
2270 		dev_err(nvme->n_dip, CE_WARN, "!no support for version > %d.%d",
2271 		    nvme_version_major, nvme_version_minor);
2272 		if (nvme->n_strict_version)
2273 			goto fail;
2274 	}
2275 
2276 	/* retrieve controller configuration */
2277 	cap.r = nvme_get64(nvme, NVME_REG_CAP);
2278 
2279 	if ((cap.b.cap_css & NVME_CAP_CSS_NVM) == 0) {
2280 		dev_err(nvme->n_dip, CE_WARN,
2281 		    "!NVM command set not supported by hardware");
2282 		goto fail;
2283 	}
2284 
2285 	nvme->n_nssr_supported = cap.b.cap_nssrs;
2286 	nvme->n_doorbell_stride = 4 << cap.b.cap_dstrd;
2287 	nvme->n_timeout = cap.b.cap_to;
2288 	nvme->n_arbitration_mechanisms = cap.b.cap_ams;
2289 	nvme->n_cont_queues_reqd = cap.b.cap_cqr;
2290 	nvme->n_max_queue_entries = cap.b.cap_mqes + 1;
2291 
2292 	/*
2293 	 * The MPSMIN and MPSMAX fields in the CAP register use 0 to specify
2294 	 * the base page size of 4k (1<<12), so add 12 here to get the real
2295 	 * page size value.
2296 	 */
2297 	nvme->n_pageshift = MIN(MAX(cap.b.cap_mpsmin + 12, PAGESHIFT),
2298 	    cap.b.cap_mpsmax + 12);
2299 	nvme->n_pagesize = 1UL << (nvme->n_pageshift);
2300 
2301 	/*
2302 	 * Set up Queue DMA to transfer at least 1 page-aligned page at a time.
2303 	 */
2304 	nvme->n_queue_dma_attr.dma_attr_align = nvme->n_pagesize;
2305 	nvme->n_queue_dma_attr.dma_attr_minxfer = nvme->n_pagesize;
2306 
2307 	/*
2308 	 * Set up PRP DMA to transfer 1 page-aligned page at a time.
2309 	 * Maxxfer may be increased after we identified the controller limits.
2310 	 */
2311 	nvme->n_prp_dma_attr.dma_attr_maxxfer = nvme->n_pagesize;
2312 	nvme->n_prp_dma_attr.dma_attr_minxfer = nvme->n_pagesize;
2313 	nvme->n_prp_dma_attr.dma_attr_align = nvme->n_pagesize;
2314 	nvme->n_prp_dma_attr.dma_attr_seg = nvme->n_pagesize - 1;
2315 
2316 	/*
2317 	 * Reset controller if it's still in ready state.
2318 	 */
2319 	if (nvme_reset(nvme, B_FALSE) == B_FALSE) {
2320 		dev_err(nvme->n_dip, CE_WARN, "!unable to reset controller");
2321 		ddi_fm_service_impact(nvme->n_dip, DDI_SERVICE_LOST);
2322 		nvme->n_dead = B_TRUE;
2323 		goto fail;
2324 	}
2325 
2326 	/*
2327 	 * Create the admin queue pair.
2328 	 */
2329 	if (nvme_alloc_qpair(nvme, nvme->n_admin_queue_len, &nvme->n_adminq, 0)
2330 	    != DDI_SUCCESS) {
2331 		dev_err(nvme->n_dip, CE_WARN,
2332 		    "!unable to allocate admin qpair");
2333 		goto fail;
2334 	}
2335 	nvme->n_ioq = kmem_alloc(sizeof (nvme_qpair_t *), KM_SLEEP);
2336 	nvme->n_ioq[0] = nvme->n_adminq;
2337 
2338 	nvme->n_progress |= NVME_ADMIN_QUEUE;
2339 
2340 	(void) ddi_prop_update_int(DDI_DEV_T_NONE, nvme->n_dip,
2341 	    "admin-queue-len", nvme->n_admin_queue_len);
2342 
2343 	aqa.b.aqa_asqs = aqa.b.aqa_acqs = nvme->n_admin_queue_len - 1;
2344 	asq = nvme->n_adminq->nq_sqdma->nd_cookie.dmac_laddress;
2345 	acq = nvme->n_adminq->nq_cqdma->nd_cookie.dmac_laddress;
2346 
2347 	ASSERT((asq & (nvme->n_pagesize - 1)) == 0);
2348 	ASSERT((acq & (nvme->n_pagesize - 1)) == 0);
2349 
2350 	nvme_put32(nvme, NVME_REG_AQA, aqa.r);
2351 	nvme_put64(nvme, NVME_REG_ASQ, asq);
2352 	nvme_put64(nvme, NVME_REG_ACQ, acq);
2353 
2354 	cc.b.cc_ams = 0;	/* use Round-Robin arbitration */
2355 	cc.b.cc_css = 0;	/* use NVM command set */
2356 	cc.b.cc_mps = nvme->n_pageshift - 12;
2357 	cc.b.cc_shn = 0;	/* no shutdown in progress */
2358 	cc.b.cc_en = 1;		/* enable controller */
2359 	cc.b.cc_iosqes = 6;	/* submission queue entry is 2^6 bytes long */
2360 	cc.b.cc_iocqes = 4;	/* completion queue entry is 2^4 bytes long */
2361 
2362 	nvme_put32(nvme, NVME_REG_CC, cc.r);
2363 
2364 	/*
2365 	 * Wait for the controller to become ready.
2366 	 */
2367 	csts.r = nvme_get32(nvme, NVME_REG_CSTS);
2368 	if (csts.b.csts_rdy == 0) {
2369 		for (i = 0; i != nvme->n_timeout * 10; i++) {
2370 			delay(drv_usectohz(50000));
2371 			csts.r = nvme_get32(nvme, NVME_REG_CSTS);
2372 
2373 			if (csts.b.csts_cfs == 1) {
2374 				dev_err(nvme->n_dip, CE_WARN,
2375 				    "!controller fatal status at init");
2376 				ddi_fm_service_impact(nvme->n_dip,
2377 				    DDI_SERVICE_LOST);
2378 				nvme->n_dead = B_TRUE;
2379 				goto fail;
2380 			}
2381 
2382 			if (csts.b.csts_rdy == 1)
2383 				break;
2384 		}
2385 	}
2386 
2387 	if (csts.b.csts_rdy == 0) {
2388 		dev_err(nvme->n_dip, CE_WARN, "!controller not ready");
2389 		ddi_fm_service_impact(nvme->n_dip, DDI_SERVICE_LOST);
2390 		nvme->n_dead = B_TRUE;
2391 		goto fail;
2392 	}
2393 
2394 	/*
2395 	 * Assume an abort command limit of 1. We'll destroy and re-init
2396 	 * that later when we know the true abort command limit.
2397 	 */
2398 	sema_init(&nvme->n_abort_sema, 1, NULL, SEMA_DRIVER, NULL);
2399 
2400 	/*
2401 	 * Setup initial interrupt for admin queue.
2402 	 */
2403 	if ((nvme_setup_interrupts(nvme, DDI_INTR_TYPE_MSIX, 1)
2404 	    != DDI_SUCCESS) &&
2405 	    (nvme_setup_interrupts(nvme, DDI_INTR_TYPE_MSI, 1)
2406 	    != DDI_SUCCESS) &&
2407 	    (nvme_setup_interrupts(nvme, DDI_INTR_TYPE_FIXED, 1)
2408 	    != DDI_SUCCESS)) {
2409 		dev_err(nvme->n_dip, CE_WARN,
2410 		    "!failed to setup initial interrupt");
2411 		goto fail;
2412 	}
2413 
2414 	/*
2415 	 * Post an asynchronous event command to catch errors.
2416 	 * We assume the asynchronous events are supported as required by
2417 	 * specification (Figure 40 in section 5 of NVMe 1.2).
2418 	 * However, since at least qemu does not follow the specification,
2419 	 * we need a mechanism to protect ourselves.
2420 	 */
2421 	nvme->n_async_event_supported = B_TRUE;
2422 	nvme_async_event(nvme);
2423 
2424 	/*
2425 	 * Identify Controller
2426 	 */
2427 	if (nvme_identify(nvme, 0, (void **)&nvme->n_idctl) != 0) {
2428 		dev_err(nvme->n_dip, CE_WARN,
2429 		    "!failed to identify controller");
2430 		goto fail;
2431 	}
2432 
2433 	/*
2434 	 * Get Vendor & Product ID
2435 	 */
2436 	bcopy(nvme->n_idctl->id_model, model, sizeof (nvme->n_idctl->id_model));
2437 	model[sizeof (nvme->n_idctl->id_model)] = '\0';
2438 	sata_split_model(model, &vendor, &product);
2439 
2440 	if (vendor == NULL)
2441 		nvme->n_vendor = strdup("NVMe");
2442 	else
2443 		nvme->n_vendor = strdup(vendor);
2444 
2445 	nvme->n_product = strdup(product);
2446 
2447 	/*
2448 	 * Get controller limits.
2449 	 */
2450 	nvme->n_async_event_limit = MAX(NVME_MIN_ASYNC_EVENT_LIMIT,
2451 	    MIN(nvme->n_admin_queue_len / 10,
2452 	    MIN(nvme->n_idctl->id_aerl + 1, nvme->n_async_event_limit)));
2453 
2454 	(void) ddi_prop_update_int(DDI_DEV_T_NONE, nvme->n_dip,
2455 	    "async-event-limit", nvme->n_async_event_limit);
2456 
2457 	nvme->n_abort_command_limit = nvme->n_idctl->id_acl + 1;
2458 
2459 	/*
2460 	 * Reinitialize the semaphore with the true abort command limit
2461 	 * supported by the hardware. It's not necessary to disable interrupts
2462 	 * as only command aborts use the semaphore, and no commands are
2463 	 * executed or aborted while we're here.
2464 	 */
2465 	sema_destroy(&nvme->n_abort_sema);
2466 	sema_init(&nvme->n_abort_sema, nvme->n_abort_command_limit - 1, NULL,
2467 	    SEMA_DRIVER, NULL);
2468 
2469 	nvme->n_progress |= NVME_CTRL_LIMITS;
2470 
2471 	if (nvme->n_idctl->id_mdts == 0)
2472 		nvme->n_max_data_transfer_size = nvme->n_pagesize * 65536;
2473 	else
2474 		nvme->n_max_data_transfer_size =
2475 		    1ull << (nvme->n_pageshift + nvme->n_idctl->id_mdts);
2476 
2477 	nvme->n_error_log_len = nvme->n_idctl->id_elpe + 1;
2478 
2479 	/*
2480 	 * Limit n_max_data_transfer_size to what we can handle in one PRP.
2481 	 * Chained PRPs are currently unsupported.
2482 	 *
2483 	 * This is a no-op on hardware which doesn't support a transfer size
2484 	 * big enough to require chained PRPs.
2485 	 */
2486 	nvme->n_max_data_transfer_size = MIN(nvme->n_max_data_transfer_size,
2487 	    (nvme->n_pagesize / sizeof (uint64_t) * nvme->n_pagesize));
2488 
2489 	nvme->n_prp_dma_attr.dma_attr_maxxfer = nvme->n_max_data_transfer_size;
2490 
2491 	/*
2492 	 * Make sure the minimum/maximum queue entry sizes are not
2493 	 * larger/smaller than the default.
2494 	 */
2495 
2496 	if (((1 << nvme->n_idctl->id_sqes.qes_min) > sizeof (nvme_sqe_t)) ||
2497 	    ((1 << nvme->n_idctl->id_sqes.qes_max) < sizeof (nvme_sqe_t)) ||
2498 	    ((1 << nvme->n_idctl->id_cqes.qes_min) > sizeof (nvme_cqe_t)) ||
2499 	    ((1 << nvme->n_idctl->id_cqes.qes_max) < sizeof (nvme_cqe_t)))
2500 		goto fail;
2501 
2502 	/*
2503 	 * Check for the presence of a Volatile Write Cache. If present,
2504 	 * enable or disable based on the value of the property
2505 	 * volatile-write-cache-enable (default is enabled).
2506 	 */
2507 	nvme->n_write_cache_present =
2508 	    nvme->n_idctl->id_vwc.vwc_present == 0 ? B_FALSE : B_TRUE;
2509 
2510 	(void) ddi_prop_update_int(DDI_DEV_T_NONE, nvme->n_dip,
2511 	    "volatile-write-cache-present",
2512 	    nvme->n_write_cache_present ? 1 : 0);
2513 
2514 	if (!nvme->n_write_cache_present) {
2515 		nvme->n_write_cache_enabled = B_FALSE;
2516 	} else if (nvme_write_cache_set(nvme, nvme->n_write_cache_enabled)
2517 	    != 0) {
2518 		dev_err(nvme->n_dip, CE_WARN,
2519 		    "!failed to %sable volatile write cache",
2520 		    nvme->n_write_cache_enabled ? "en" : "dis");
2521 		/*
2522 		 * Assume the cache is (still) enabled.
2523 		 */
2524 		nvme->n_write_cache_enabled = B_TRUE;
2525 	}
2526 
2527 	(void) ddi_prop_update_int(DDI_DEV_T_NONE, nvme->n_dip,
2528 	    "volatile-write-cache-enable",
2529 	    nvme->n_write_cache_enabled ? 1 : 0);
2530 
2531 	/*
2532 	 * Assume LBA Range Type feature is supported. If it isn't this
2533 	 * will be set to B_FALSE by nvme_get_features().
2534 	 */
2535 	nvme->n_lba_range_supported = B_TRUE;
2536 
2537 	/*
2538 	 * Check support for Autonomous Power State Transition.
2539 	 */
2540 	if (NVME_VERSION_ATLEAST(&nvme->n_version, 1, 1))
2541 		nvme->n_auto_pst_supported =
2542 		    nvme->n_idctl->id_apsta.ap_sup == 0 ? B_FALSE : B_TRUE;
2543 
2544 	/*
2545 	 * Assume Software Progress Marker feature is supported.  If it isn't
2546 	 * this will be set to B_FALSE by nvme_get_features().
2547 	 */
2548 	nvme->n_progress_supported = B_TRUE;
2549 
2550 	/*
2551 	 * Identify Namespaces
2552 	 */
2553 	nvme->n_namespace_count = nvme->n_idctl->id_nn;
2554 
2555 	if (nvme->n_namespace_count == 0) {
2556 		dev_err(nvme->n_dip, CE_WARN,
2557 		    "!controllers without namespaces are not supported");
2558 		goto fail;
2559 	}
2560 
2561 	if (nvme->n_namespace_count > NVME_MINOR_MAX) {
2562 		dev_err(nvme->n_dip, CE_WARN,
2563 		    "!too many namespaces: %d, limiting to %d\n",
2564 		    nvme->n_namespace_count, NVME_MINOR_MAX);
2565 		nvme->n_namespace_count = NVME_MINOR_MAX;
2566 	}
2567 
2568 	nvme->n_ns = kmem_zalloc(sizeof (nvme_namespace_t) *
2569 	    nvme->n_namespace_count, KM_SLEEP);
2570 
2571 	for (i = 0; i != nvme->n_namespace_count; i++) {
2572 		mutex_init(&nvme->n_ns[i].ns_minor.nm_mutex, NULL, MUTEX_DRIVER,
2573 		    NULL);
2574 		if (nvme_init_ns(nvme, i + 1) != DDI_SUCCESS)
2575 			goto fail;
2576 	}
2577 
2578 	/*
2579 	 * Try to set up MSI/MSI-X interrupts.
2580 	 */
2581 	if ((nvme->n_intr_types & (DDI_INTR_TYPE_MSI | DDI_INTR_TYPE_MSIX))
2582 	    != 0) {
2583 		nvme_release_interrupts(nvme);
2584 
2585 		nqueues = MIN(UINT16_MAX, ncpus);
2586 
2587 		if ((nvme_setup_interrupts(nvme, DDI_INTR_TYPE_MSIX,
2588 		    nqueues) != DDI_SUCCESS) &&
2589 		    (nvme_setup_interrupts(nvme, DDI_INTR_TYPE_MSI,
2590 		    nqueues) != DDI_SUCCESS)) {
2591 			dev_err(nvme->n_dip, CE_WARN,
2592 			    "!failed to setup MSI/MSI-X interrupts");
2593 			goto fail;
2594 		}
2595 	}
2596 
2597 	nqueues = nvme->n_intr_cnt;
2598 
2599 	/*
2600 	 * Create I/O queue pairs.
2601 	 */
2602 
2603 	if (nvme_set_nqueues(nvme, &nqueues) != 0) {
2604 		dev_err(nvme->n_dip, CE_WARN,
2605 		    "!failed to set number of I/O queues to %d",
2606 		    nvme->n_intr_cnt);
2607 		goto fail;
2608 	}
2609 
2610 	/*
2611 	 * Reallocate I/O queue array
2612 	 */
2613 	kmem_free(nvme->n_ioq, sizeof (nvme_qpair_t *));
2614 	nvme->n_ioq = kmem_zalloc(sizeof (nvme_qpair_t *) *
2615 	    (nqueues + 1), KM_SLEEP);
2616 	nvme->n_ioq[0] = nvme->n_adminq;
2617 
2618 	nvme->n_ioq_count = nqueues;
2619 
2620 	/*
2621 	 * If we got less queues than we asked for we might as well give
2622 	 * some of the interrupt vectors back to the system.
2623 	 */
2624 	if (nvme->n_ioq_count < nvme->n_intr_cnt) {
2625 		nvme_release_interrupts(nvme);
2626 
2627 		if (nvme_setup_interrupts(nvme, nvme->n_intr_type,
2628 		    nvme->n_ioq_count) != DDI_SUCCESS) {
2629 			dev_err(nvme->n_dip, CE_WARN,
2630 			    "!failed to reduce number of interrupts");
2631 			goto fail;
2632 		}
2633 	}
2634 
2635 	/*
2636 	 * Alloc & register I/O queue pairs
2637 	 */
2638 	nvme->n_io_queue_len =
2639 	    MIN(nvme->n_io_queue_len, nvme->n_max_queue_entries);
2640 	(void) ddi_prop_update_int(DDI_DEV_T_NONE, nvme->n_dip, "io-queue-len",
2641 	    nvme->n_io_queue_len);
2642 
2643 	for (i = 1; i != nvme->n_ioq_count + 1; i++) {
2644 		if (nvme_alloc_qpair(nvme, nvme->n_io_queue_len,
2645 		    &nvme->n_ioq[i], i) != DDI_SUCCESS) {
2646 			dev_err(nvme->n_dip, CE_WARN,
2647 			    "!unable to allocate I/O qpair %d", i);
2648 			goto fail;
2649 		}
2650 
2651 		if (nvme_create_io_qpair(nvme, nvme->n_ioq[i], i) != 0) {
2652 			dev_err(nvme->n_dip, CE_WARN,
2653 			    "!unable to create I/O qpair %d", i);
2654 			goto fail;
2655 		}
2656 	}
2657 
2658 	/*
2659 	 * Post more asynchronous events commands to reduce event reporting
2660 	 * latency as suggested by the spec.
2661 	 */
2662 	if (nvme->n_async_event_supported) {
2663 		for (i = 1; i != nvme->n_async_event_limit; i++)
2664 			nvme_async_event(nvme);
2665 	}
2666 
2667 	return (DDI_SUCCESS);
2668 
2669 fail:
2670 	(void) nvme_reset(nvme, B_FALSE);
2671 	return (DDI_FAILURE);
2672 }
2673 
2674 static uint_t
2675 nvme_intr(caddr_t arg1, caddr_t arg2)
2676 {
2677 	/*LINTED: E_PTR_BAD_CAST_ALIGN*/
2678 	nvme_t *nvme = (nvme_t *)arg1;
2679 	int inum = (int)(uintptr_t)arg2;
2680 	int ccnt = 0;
2681 	int qnum;
2682 	nvme_cmd_t *cmd;
2683 
2684 	if (inum >= nvme->n_intr_cnt)
2685 		return (DDI_INTR_UNCLAIMED);
2686 
2687 	if (nvme->n_dead)
2688 		return (nvme->n_intr_type == DDI_INTR_TYPE_FIXED ?
2689 		    DDI_INTR_UNCLAIMED : DDI_INTR_CLAIMED);
2690 
2691 	/*
2692 	 * The interrupt vector a queue uses is calculated as queue_idx %
2693 	 * intr_cnt in nvme_create_io_qpair(). Iterate through the queue array
2694 	 * in steps of n_intr_cnt to process all queues using this vector.
2695 	 */
2696 	for (qnum = inum;
2697 	    qnum < nvme->n_ioq_count + 1 && nvme->n_ioq[qnum] != NULL;
2698 	    qnum += nvme->n_intr_cnt) {
2699 		while ((cmd = nvme_retrieve_cmd(nvme, nvme->n_ioq[qnum]))) {
2700 			taskq_dispatch_ent((taskq_t *)cmd->nc_nvme->n_cmd_taskq,
2701 			    cmd->nc_callback, cmd, TQ_NOSLEEP, &cmd->nc_tqent);
2702 			ccnt++;
2703 		}
2704 	}
2705 
2706 	return (ccnt > 0 ? DDI_INTR_CLAIMED : DDI_INTR_UNCLAIMED);
2707 }
2708 
2709 static void
2710 nvme_release_interrupts(nvme_t *nvme)
2711 {
2712 	int i;
2713 
2714 	for (i = 0; i < nvme->n_intr_cnt; i++) {
2715 		if (nvme->n_inth[i] == NULL)
2716 			break;
2717 
2718 		if (nvme->n_intr_cap & DDI_INTR_FLAG_BLOCK)
2719 			(void) ddi_intr_block_disable(&nvme->n_inth[i], 1);
2720 		else
2721 			(void) ddi_intr_disable(nvme->n_inth[i]);
2722 
2723 		(void) ddi_intr_remove_handler(nvme->n_inth[i]);
2724 		(void) ddi_intr_free(nvme->n_inth[i]);
2725 	}
2726 
2727 	kmem_free(nvme->n_inth, nvme->n_inth_sz);
2728 	nvme->n_inth = NULL;
2729 	nvme->n_inth_sz = 0;
2730 
2731 	nvme->n_progress &= ~NVME_INTERRUPTS;
2732 }
2733 
2734 static int
2735 nvme_setup_interrupts(nvme_t *nvme, int intr_type, int nqpairs)
2736 {
2737 	int nintrs, navail, count;
2738 	int ret;
2739 	int i;
2740 
2741 	if (nvme->n_intr_types == 0) {
2742 		ret = ddi_intr_get_supported_types(nvme->n_dip,
2743 		    &nvme->n_intr_types);
2744 		if (ret != DDI_SUCCESS) {
2745 			dev_err(nvme->n_dip, CE_WARN,
2746 			    "!%s: ddi_intr_get_supported types failed",
2747 			    __func__);
2748 			return (ret);
2749 		}
2750 #ifdef __x86
2751 		if (get_hwenv() == HW_VMWARE)
2752 			nvme->n_intr_types &= ~DDI_INTR_TYPE_MSIX;
2753 #endif
2754 	}
2755 
2756 	if ((nvme->n_intr_types & intr_type) == 0)
2757 		return (DDI_FAILURE);
2758 
2759 	ret = ddi_intr_get_nintrs(nvme->n_dip, intr_type, &nintrs);
2760 	if (ret != DDI_SUCCESS) {
2761 		dev_err(nvme->n_dip, CE_WARN, "!%s: ddi_intr_get_nintrs failed",
2762 		    __func__);
2763 		return (ret);
2764 	}
2765 
2766 	ret = ddi_intr_get_navail(nvme->n_dip, intr_type, &navail);
2767 	if (ret != DDI_SUCCESS) {
2768 		dev_err(nvme->n_dip, CE_WARN, "!%s: ddi_intr_get_navail failed",
2769 		    __func__);
2770 		return (ret);
2771 	}
2772 
2773 	/* We want at most one interrupt per queue pair. */
2774 	if (navail > nqpairs)
2775 		navail = nqpairs;
2776 
2777 	nvme->n_inth_sz = sizeof (ddi_intr_handle_t) * navail;
2778 	nvme->n_inth = kmem_zalloc(nvme->n_inth_sz, KM_SLEEP);
2779 
2780 	ret = ddi_intr_alloc(nvme->n_dip, nvme->n_inth, intr_type, 0, navail,
2781 	    &count, 0);
2782 	if (ret != DDI_SUCCESS) {
2783 		dev_err(nvme->n_dip, CE_WARN, "!%s: ddi_intr_alloc failed",
2784 		    __func__);
2785 		goto fail;
2786 	}
2787 
2788 	nvme->n_intr_cnt = count;
2789 
2790 	ret = ddi_intr_get_pri(nvme->n_inth[0], &nvme->n_intr_pri);
2791 	if (ret != DDI_SUCCESS) {
2792 		dev_err(nvme->n_dip, CE_WARN, "!%s: ddi_intr_get_pri failed",
2793 		    __func__);
2794 		goto fail;
2795 	}
2796 
2797 	for (i = 0; i < count; i++) {
2798 		ret = ddi_intr_add_handler(nvme->n_inth[i], nvme_intr,
2799 		    (void *)nvme, (void *)(uintptr_t)i);
2800 		if (ret != DDI_SUCCESS) {
2801 			dev_err(nvme->n_dip, CE_WARN,
2802 			    "!%s: ddi_intr_add_handler failed", __func__);
2803 			goto fail;
2804 		}
2805 	}
2806 
2807 	(void) ddi_intr_get_cap(nvme->n_inth[0], &nvme->n_intr_cap);
2808 
2809 	for (i = 0; i < count; i++) {
2810 		if (nvme->n_intr_cap & DDI_INTR_FLAG_BLOCK)
2811 			ret = ddi_intr_block_enable(&nvme->n_inth[i], 1);
2812 		else
2813 			ret = ddi_intr_enable(nvme->n_inth[i]);
2814 
2815 		if (ret != DDI_SUCCESS) {
2816 			dev_err(nvme->n_dip, CE_WARN,
2817 			    "!%s: enabling interrupt %d failed", __func__, i);
2818 			goto fail;
2819 		}
2820 	}
2821 
2822 	nvme->n_intr_type = intr_type;
2823 
2824 	nvme->n_progress |= NVME_INTERRUPTS;
2825 
2826 	return (DDI_SUCCESS);
2827 
2828 fail:
2829 	nvme_release_interrupts(nvme);
2830 
2831 	return (ret);
2832 }
2833 
2834 static int
2835 nvme_fm_errcb(dev_info_t *dip, ddi_fm_error_t *fm_error, const void *arg)
2836 {
2837 	_NOTE(ARGUNUSED(arg));
2838 
2839 	pci_ereport_post(dip, fm_error, NULL);
2840 	return (fm_error->fme_status);
2841 }
2842 
2843 static int
2844 nvme_attach(dev_info_t *dip, ddi_attach_cmd_t cmd)
2845 {
2846 	nvme_t *nvme;
2847 	int instance;
2848 	int nregs;
2849 	off_t regsize;
2850 	int i;
2851 	char name[32];
2852 
2853 	if (cmd != DDI_ATTACH)
2854 		return (DDI_FAILURE);
2855 
2856 	instance = ddi_get_instance(dip);
2857 
2858 	if (ddi_soft_state_zalloc(nvme_state, instance) != DDI_SUCCESS)
2859 		return (DDI_FAILURE);
2860 
2861 	nvme = ddi_get_soft_state(nvme_state, instance);
2862 	ddi_set_driver_private(dip, nvme);
2863 	nvme->n_dip = dip;
2864 
2865 	mutex_init(&nvme->n_minor.nm_mutex, NULL, MUTEX_DRIVER, NULL);
2866 
2867 	nvme->n_strict_version = ddi_prop_get_int(DDI_DEV_T_ANY, dip,
2868 	    DDI_PROP_DONTPASS, "strict-version", 1) == 1 ? B_TRUE : B_FALSE;
2869 	nvme->n_ignore_unknown_vendor_status = ddi_prop_get_int(DDI_DEV_T_ANY,
2870 	    dip, DDI_PROP_DONTPASS, "ignore-unknown-vendor-status", 0) == 1 ?
2871 	    B_TRUE : B_FALSE;
2872 	nvme->n_admin_queue_len = ddi_prop_get_int(DDI_DEV_T_ANY, dip,
2873 	    DDI_PROP_DONTPASS, "admin-queue-len", NVME_DEFAULT_ADMIN_QUEUE_LEN);
2874 	nvme->n_io_queue_len = ddi_prop_get_int(DDI_DEV_T_ANY, dip,
2875 	    DDI_PROP_DONTPASS, "io-queue-len", NVME_DEFAULT_IO_QUEUE_LEN);
2876 	nvme->n_async_event_limit = ddi_prop_get_int(DDI_DEV_T_ANY, dip,
2877 	    DDI_PROP_DONTPASS, "async-event-limit",
2878 	    NVME_DEFAULT_ASYNC_EVENT_LIMIT);
2879 	nvme->n_write_cache_enabled = ddi_prop_get_int(DDI_DEV_T_ANY, dip,
2880 	    DDI_PROP_DONTPASS, "volatile-write-cache-enable", 1) != 0 ?
2881 	    B_TRUE : B_FALSE;
2882 	nvme->n_min_block_size = ddi_prop_get_int(DDI_DEV_T_ANY, dip,
2883 	    DDI_PROP_DONTPASS, "min-phys-block-size",
2884 	    NVME_DEFAULT_MIN_BLOCK_SIZE);
2885 
2886 	if (!ISP2(nvme->n_min_block_size) ||
2887 	    (nvme->n_min_block_size < NVME_DEFAULT_MIN_BLOCK_SIZE)) {
2888 		dev_err(dip, CE_WARN, "!min-phys-block-size %s, "
2889 		    "using default %d", ISP2(nvme->n_min_block_size) ?
2890 		    "too low" : "not a power of 2",
2891 		    NVME_DEFAULT_MIN_BLOCK_SIZE);
2892 		nvme->n_min_block_size = NVME_DEFAULT_MIN_BLOCK_SIZE;
2893 	}
2894 
2895 	if (nvme->n_admin_queue_len < NVME_MIN_ADMIN_QUEUE_LEN)
2896 		nvme->n_admin_queue_len = NVME_MIN_ADMIN_QUEUE_LEN;
2897 	else if (nvme->n_admin_queue_len > NVME_MAX_ADMIN_QUEUE_LEN)
2898 		nvme->n_admin_queue_len = NVME_MAX_ADMIN_QUEUE_LEN;
2899 
2900 	if (nvme->n_io_queue_len < NVME_MIN_IO_QUEUE_LEN)
2901 		nvme->n_io_queue_len = NVME_MIN_IO_QUEUE_LEN;
2902 
2903 	if (nvme->n_async_event_limit < 1)
2904 		nvme->n_async_event_limit = NVME_DEFAULT_ASYNC_EVENT_LIMIT;
2905 
2906 	nvme->n_reg_acc_attr = nvme_reg_acc_attr;
2907 	nvme->n_queue_dma_attr = nvme_queue_dma_attr;
2908 	nvme->n_prp_dma_attr = nvme_prp_dma_attr;
2909 	nvme->n_sgl_dma_attr = nvme_sgl_dma_attr;
2910 
2911 	/*
2912 	 * Setup FMA support.
2913 	 */
2914 	nvme->n_fm_cap = ddi_getprop(DDI_DEV_T_ANY, dip,
2915 	    DDI_PROP_CANSLEEP | DDI_PROP_DONTPASS, "fm-capable",
2916 	    DDI_FM_EREPORT_CAPABLE | DDI_FM_ACCCHK_CAPABLE |
2917 	    DDI_FM_DMACHK_CAPABLE | DDI_FM_ERRCB_CAPABLE);
2918 
2919 	ddi_fm_init(dip, &nvme->n_fm_cap, &nvme->n_fm_ibc);
2920 
2921 	if (nvme->n_fm_cap) {
2922 		if (nvme->n_fm_cap & DDI_FM_ACCCHK_CAPABLE)
2923 			nvme->n_reg_acc_attr.devacc_attr_access =
2924 			    DDI_FLAGERR_ACC;
2925 
2926 		if (nvme->n_fm_cap & DDI_FM_DMACHK_CAPABLE) {
2927 			nvme->n_prp_dma_attr.dma_attr_flags |= DDI_DMA_FLAGERR;
2928 			nvme->n_sgl_dma_attr.dma_attr_flags |= DDI_DMA_FLAGERR;
2929 		}
2930 
2931 		if (DDI_FM_EREPORT_CAP(nvme->n_fm_cap) ||
2932 		    DDI_FM_ERRCB_CAP(nvme->n_fm_cap))
2933 			pci_ereport_setup(dip);
2934 
2935 		if (DDI_FM_ERRCB_CAP(nvme->n_fm_cap))
2936 			ddi_fm_handler_register(dip, nvme_fm_errcb,
2937 			    (void *)nvme);
2938 	}
2939 
2940 	nvme->n_progress |= NVME_FMA_INIT;
2941 
2942 	/*
2943 	 * The spec defines several register sets. Only the controller
2944 	 * registers (set 1) are currently used.
2945 	 */
2946 	if (ddi_dev_nregs(dip, &nregs) == DDI_FAILURE ||
2947 	    nregs < 2 ||
2948 	    ddi_dev_regsize(dip, 1, &regsize) == DDI_FAILURE)
2949 		goto fail;
2950 
2951 	if (ddi_regs_map_setup(dip, 1, &nvme->n_regs, 0, regsize,
2952 	    &nvme->n_reg_acc_attr, &nvme->n_regh) != DDI_SUCCESS) {
2953 		dev_err(dip, CE_WARN, "!failed to map regset 1");
2954 		goto fail;
2955 	}
2956 
2957 	nvme->n_progress |= NVME_REGS_MAPPED;
2958 
2959 	/*
2960 	 * Create taskq for command completion.
2961 	 */
2962 	(void) snprintf(name, sizeof (name), "%s%d_cmd_taskq",
2963 	    ddi_driver_name(dip), ddi_get_instance(dip));
2964 	nvme->n_cmd_taskq = ddi_taskq_create(dip, name, MIN(UINT16_MAX, ncpus),
2965 	    TASKQ_DEFAULTPRI, 0);
2966 	if (nvme->n_cmd_taskq == NULL) {
2967 		dev_err(dip, CE_WARN, "!failed to create cmd taskq");
2968 		goto fail;
2969 	}
2970 
2971 	/*
2972 	 * Create PRP DMA cache
2973 	 */
2974 	(void) snprintf(name, sizeof (name), "%s%d_prp_cache",
2975 	    ddi_driver_name(dip), ddi_get_instance(dip));
2976 	nvme->n_prp_cache = kmem_cache_create(name, sizeof (nvme_dma_t),
2977 	    0, nvme_prp_dma_constructor, nvme_prp_dma_destructor,
2978 	    NULL, (void *)nvme, NULL, 0);
2979 
2980 	if (nvme_init(nvme) != DDI_SUCCESS)
2981 		goto fail;
2982 
2983 	/*
2984 	 * Attach the blkdev driver for each namespace.
2985 	 */
2986 	for (i = 0; i != nvme->n_namespace_count; i++) {
2987 		if (ddi_create_minor_node(nvme->n_dip, nvme->n_ns[i].ns_name,
2988 		    S_IFCHR, NVME_MINOR(ddi_get_instance(nvme->n_dip), i + 1),
2989 		    DDI_NT_NVME_ATTACHMENT_POINT, 0) != DDI_SUCCESS) {
2990 			dev_err(dip, CE_WARN,
2991 			    "!failed to create minor node for namespace %d", i);
2992 			goto fail;
2993 		}
2994 
2995 		if (nvme->n_ns[i].ns_ignore)
2996 			continue;
2997 
2998 		nvme->n_ns[i].ns_bd_hdl = bd_alloc_handle(&nvme->n_ns[i],
2999 		    &nvme_bd_ops, &nvme->n_prp_dma_attr, KM_SLEEP);
3000 
3001 		if (nvme->n_ns[i].ns_bd_hdl == NULL) {
3002 			dev_err(dip, CE_WARN,
3003 			    "!failed to get blkdev handle for namespace %d", i);
3004 			goto fail;
3005 		}
3006 
3007 		if (bd_attach_handle(dip, nvme->n_ns[i].ns_bd_hdl)
3008 		    != DDI_SUCCESS) {
3009 			dev_err(dip, CE_WARN,
3010 			    "!failed to attach blkdev handle for namespace %d",
3011 			    i);
3012 			goto fail;
3013 		}
3014 	}
3015 
3016 	if (ddi_create_minor_node(dip, "devctl", S_IFCHR,
3017 	    NVME_MINOR(ddi_get_instance(dip), 0), DDI_NT_NVME_NEXUS, 0)
3018 	    != DDI_SUCCESS) {
3019 		dev_err(dip, CE_WARN, "nvme_attach: "
3020 		    "cannot create devctl minor node");
3021 		goto fail;
3022 	}
3023 
3024 	return (DDI_SUCCESS);
3025 
3026 fail:
3027 	/* attach successful anyway so that FMA can retire the device */
3028 	if (nvme->n_dead)
3029 		return (DDI_SUCCESS);
3030 
3031 	(void) nvme_detach(dip, DDI_DETACH);
3032 
3033 	return (DDI_FAILURE);
3034 }
3035 
3036 static int
3037 nvme_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
3038 {
3039 	int instance, i;
3040 	nvme_t *nvme;
3041 
3042 	if (cmd != DDI_DETACH)
3043 		return (DDI_FAILURE);
3044 
3045 	instance = ddi_get_instance(dip);
3046 
3047 	nvme = ddi_get_soft_state(nvme_state, instance);
3048 
3049 	if (nvme == NULL)
3050 		return (DDI_FAILURE);
3051 
3052 	ddi_remove_minor_node(dip, "devctl");
3053 	mutex_destroy(&nvme->n_minor.nm_mutex);
3054 
3055 	if (nvme->n_ns) {
3056 		for (i = 0; i != nvme->n_namespace_count; i++) {
3057 			ddi_remove_minor_node(dip, nvme->n_ns[i].ns_name);
3058 			mutex_destroy(&nvme->n_ns[i].ns_minor.nm_mutex);
3059 
3060 			if (nvme->n_ns[i].ns_bd_hdl) {
3061 				(void) bd_detach_handle(
3062 				    nvme->n_ns[i].ns_bd_hdl);
3063 				bd_free_handle(nvme->n_ns[i].ns_bd_hdl);
3064 			}
3065 
3066 			if (nvme->n_ns[i].ns_idns)
3067 				kmem_free(nvme->n_ns[i].ns_idns,
3068 				    sizeof (nvme_identify_nsid_t));
3069 			if (nvme->n_ns[i].ns_devid)
3070 				strfree(nvme->n_ns[i].ns_devid);
3071 		}
3072 
3073 		kmem_free(nvme->n_ns, sizeof (nvme_namespace_t) *
3074 		    nvme->n_namespace_count);
3075 	}
3076 
3077 	if (nvme->n_progress & NVME_INTERRUPTS)
3078 		nvme_release_interrupts(nvme);
3079 
3080 	if (nvme->n_cmd_taskq)
3081 		ddi_taskq_wait(nvme->n_cmd_taskq);
3082 
3083 	if (nvme->n_ioq_count > 0) {
3084 		for (i = 1; i != nvme->n_ioq_count + 1; i++) {
3085 			if (nvme->n_ioq[i] != NULL) {
3086 				/* TODO: send destroy queue commands */
3087 				nvme_free_qpair(nvme->n_ioq[i]);
3088 			}
3089 		}
3090 
3091 		kmem_free(nvme->n_ioq, sizeof (nvme_qpair_t *) *
3092 		    (nvme->n_ioq_count + 1));
3093 	}
3094 
3095 	if (nvme->n_prp_cache != NULL) {
3096 		kmem_cache_destroy(nvme->n_prp_cache);
3097 	}
3098 
3099 	if (nvme->n_progress & NVME_REGS_MAPPED) {
3100 		nvme_shutdown(nvme, NVME_CC_SHN_NORMAL, B_FALSE);
3101 		(void) nvme_reset(nvme, B_FALSE);
3102 	}
3103 
3104 	if (nvme->n_cmd_taskq)
3105 		ddi_taskq_destroy(nvme->n_cmd_taskq);
3106 
3107 	if (nvme->n_progress & NVME_CTRL_LIMITS)
3108 		sema_destroy(&nvme->n_abort_sema);
3109 
3110 	if (nvme->n_progress & NVME_ADMIN_QUEUE)
3111 		nvme_free_qpair(nvme->n_adminq);
3112 
3113 	if (nvme->n_idctl)
3114 		kmem_free(nvme->n_idctl, NVME_IDENTIFY_BUFSIZE);
3115 
3116 	if (nvme->n_progress & NVME_REGS_MAPPED)
3117 		ddi_regs_map_free(&nvme->n_regh);
3118 
3119 	if (nvme->n_progress & NVME_FMA_INIT) {
3120 		if (DDI_FM_ERRCB_CAP(nvme->n_fm_cap))
3121 			ddi_fm_handler_unregister(nvme->n_dip);
3122 
3123 		if (DDI_FM_EREPORT_CAP(nvme->n_fm_cap) ||
3124 		    DDI_FM_ERRCB_CAP(nvme->n_fm_cap))
3125 			pci_ereport_teardown(nvme->n_dip);
3126 
3127 		ddi_fm_fini(nvme->n_dip);
3128 	}
3129 
3130 	if (nvme->n_vendor != NULL)
3131 		strfree(nvme->n_vendor);
3132 
3133 	if (nvme->n_product != NULL)
3134 		strfree(nvme->n_product);
3135 
3136 	ddi_soft_state_free(nvme_state, instance);
3137 
3138 	return (DDI_SUCCESS);
3139 }
3140 
3141 static int
3142 nvme_quiesce(dev_info_t *dip)
3143 {
3144 	int instance;
3145 	nvme_t *nvme;
3146 
3147 	instance = ddi_get_instance(dip);
3148 
3149 	nvme = ddi_get_soft_state(nvme_state, instance);
3150 
3151 	if (nvme == NULL)
3152 		return (DDI_FAILURE);
3153 
3154 	nvme_shutdown(nvme, NVME_CC_SHN_ABRUPT, B_TRUE);
3155 
3156 	(void) nvme_reset(nvme, B_TRUE);
3157 
3158 	return (DDI_FAILURE);
3159 }
3160 
3161 static int
3162 nvme_fill_prp(nvme_cmd_t *cmd, bd_xfer_t *xfer)
3163 {
3164 	nvme_t *nvme = cmd->nc_nvme;
3165 	int nprp_page, nprp;
3166 	uint64_t *prp;
3167 
3168 	if (xfer->x_ndmac == 0)
3169 		return (DDI_FAILURE);
3170 
3171 	cmd->nc_sqe.sqe_dptr.d_prp[0] = xfer->x_dmac.dmac_laddress;
3172 	ddi_dma_nextcookie(xfer->x_dmah, &xfer->x_dmac);
3173 
3174 	if (xfer->x_ndmac == 1) {
3175 		cmd->nc_sqe.sqe_dptr.d_prp[1] = 0;
3176 		return (DDI_SUCCESS);
3177 	} else if (xfer->x_ndmac == 2) {
3178 		cmd->nc_sqe.sqe_dptr.d_prp[1] = xfer->x_dmac.dmac_laddress;
3179 		return (DDI_SUCCESS);
3180 	}
3181 
3182 	xfer->x_ndmac--;
3183 
3184 	nprp_page = nvme->n_pagesize / sizeof (uint64_t) - 1;
3185 	ASSERT(nprp_page > 0);
3186 	nprp = (xfer->x_ndmac + nprp_page - 1) / nprp_page;
3187 
3188 	/*
3189 	 * We currently don't support chained PRPs and set up our DMA
3190 	 * attributes to reflect that. If we still get an I/O request
3191 	 * that needs a chained PRP something is very wrong.
3192 	 */
3193 	VERIFY(nprp == 1);
3194 
3195 	cmd->nc_dma = kmem_cache_alloc(nvme->n_prp_cache, KM_SLEEP);
3196 	bzero(cmd->nc_dma->nd_memp, cmd->nc_dma->nd_len);
3197 
3198 	cmd->nc_sqe.sqe_dptr.d_prp[1] = cmd->nc_dma->nd_cookie.dmac_laddress;
3199 
3200 	/*LINTED: E_PTR_BAD_CAST_ALIGN*/
3201 	for (prp = (uint64_t *)cmd->nc_dma->nd_memp;
3202 	    xfer->x_ndmac > 0;
3203 	    prp++, xfer->x_ndmac--) {
3204 		*prp = xfer->x_dmac.dmac_laddress;
3205 		ddi_dma_nextcookie(xfer->x_dmah, &xfer->x_dmac);
3206 	}
3207 
3208 	(void) ddi_dma_sync(cmd->nc_dma->nd_dmah, 0, cmd->nc_dma->nd_len,
3209 	    DDI_DMA_SYNC_FORDEV);
3210 	return (DDI_SUCCESS);
3211 }
3212 
3213 static nvme_cmd_t *
3214 nvme_create_nvm_cmd(nvme_namespace_t *ns, uint8_t opc, bd_xfer_t *xfer)
3215 {
3216 	nvme_t *nvme = ns->ns_nvme;
3217 	nvme_cmd_t *cmd;
3218 
3219 	/*
3220 	 * Blkdev only sets BD_XFER_POLL when dumping, so don't sleep.
3221 	 */
3222 	cmd = nvme_alloc_cmd(nvme, (xfer->x_flags & BD_XFER_POLL) ?
3223 	    KM_NOSLEEP : KM_SLEEP);
3224 
3225 	if (cmd == NULL)
3226 		return (NULL);
3227 
3228 	cmd->nc_sqe.sqe_opc = opc;
3229 	cmd->nc_callback = nvme_bd_xfer_done;
3230 	cmd->nc_xfer = xfer;
3231 
3232 	switch (opc) {
3233 	case NVME_OPC_NVM_WRITE:
3234 	case NVME_OPC_NVM_READ:
3235 		VERIFY(xfer->x_nblks <= 0x10000);
3236 
3237 		cmd->nc_sqe.sqe_nsid = ns->ns_id;
3238 
3239 		cmd->nc_sqe.sqe_cdw10 = xfer->x_blkno & 0xffffffffu;
3240 		cmd->nc_sqe.sqe_cdw11 = (xfer->x_blkno >> 32);
3241 		cmd->nc_sqe.sqe_cdw12 = (uint16_t)(xfer->x_nblks - 1);
3242 
3243 		if (nvme_fill_prp(cmd, xfer) != DDI_SUCCESS)
3244 			goto fail;
3245 		break;
3246 
3247 	case NVME_OPC_NVM_FLUSH:
3248 		cmd->nc_sqe.sqe_nsid = ns->ns_id;
3249 		break;
3250 
3251 	default:
3252 		goto fail;
3253 	}
3254 
3255 	return (cmd);
3256 
3257 fail:
3258 	nvme_free_cmd(cmd);
3259 	return (NULL);
3260 }
3261 
3262 static void
3263 nvme_bd_xfer_done(void *arg)
3264 {
3265 	nvme_cmd_t *cmd = arg;
3266 	bd_xfer_t *xfer = cmd->nc_xfer;
3267 	int error = 0;
3268 
3269 	error = nvme_check_cmd_status(cmd);
3270 	nvme_free_cmd(cmd);
3271 
3272 	bd_xfer_done(xfer, error);
3273 }
3274 
3275 static void
3276 nvme_bd_driveinfo(void *arg, bd_drive_t *drive)
3277 {
3278 	nvme_namespace_t *ns = arg;
3279 	nvme_t *nvme = ns->ns_nvme;
3280 
3281 	/*
3282 	 * blkdev maintains one queue size per instance (namespace),
3283 	 * but all namespace share the I/O queues.
3284 	 * TODO: need to figure out a sane default, or use per-NS I/O queues,
3285 	 * or change blkdev to handle EAGAIN
3286 	 */
3287 	drive->d_qsize = nvme->n_ioq_count * nvme->n_io_queue_len
3288 	    / nvme->n_namespace_count;
3289 
3290 	/*
3291 	 * d_maxxfer is not set, which means the value is taken from the DMA
3292 	 * attributes specified to bd_alloc_handle.
3293 	 */
3294 
3295 	drive->d_removable = B_FALSE;
3296 	drive->d_hotpluggable = B_FALSE;
3297 
3298 	bcopy(ns->ns_eui64, drive->d_eui64, sizeof (drive->d_eui64));
3299 	drive->d_target = ns->ns_id;
3300 	drive->d_lun = 0;
3301 
3302 	drive->d_model = nvme->n_idctl->id_model;
3303 	drive->d_model_len = sizeof (nvme->n_idctl->id_model);
3304 	drive->d_vendor = nvme->n_vendor;
3305 	drive->d_vendor_len = strlen(nvme->n_vendor);
3306 	drive->d_product = nvme->n_product;
3307 	drive->d_product_len = strlen(nvme->n_product);
3308 	drive->d_serial = nvme->n_idctl->id_serial;
3309 	drive->d_serial_len = sizeof (nvme->n_idctl->id_serial);
3310 	drive->d_revision = nvme->n_idctl->id_fwrev;
3311 	drive->d_revision_len = sizeof (nvme->n_idctl->id_fwrev);
3312 }
3313 
3314 static int
3315 nvme_bd_mediainfo(void *arg, bd_media_t *media)
3316 {
3317 	nvme_namespace_t *ns = arg;
3318 
3319 	media->m_nblks = ns->ns_block_count;
3320 	media->m_blksize = ns->ns_block_size;
3321 	media->m_readonly = B_FALSE;
3322 	media->m_solidstate = B_TRUE;
3323 
3324 	media->m_pblksize = ns->ns_best_block_size;
3325 
3326 	return (0);
3327 }
3328 
3329 static int
3330 nvme_bd_cmd(nvme_namespace_t *ns, bd_xfer_t *xfer, uint8_t opc)
3331 {
3332 	nvme_t *nvme = ns->ns_nvme;
3333 	nvme_cmd_t *cmd;
3334 	nvme_qpair_t *ioq;
3335 	boolean_t poll;
3336 	int ret;
3337 
3338 	if (nvme->n_dead)
3339 		return (EIO);
3340 
3341 	cmd = nvme_create_nvm_cmd(ns, opc, xfer);
3342 	if (cmd == NULL)
3343 		return (ENOMEM);
3344 
3345 	cmd->nc_sqid = (CPU->cpu_id % nvme->n_ioq_count) + 1;
3346 	ASSERT(cmd->nc_sqid <= nvme->n_ioq_count);
3347 	ioq = nvme->n_ioq[cmd->nc_sqid];
3348 
3349 	/*
3350 	 * Get the polling flag before submitting the command. The command may
3351 	 * complete immediately after it was submitted, which means we must
3352 	 * treat both cmd and xfer as if they have been freed already.
3353 	 */
3354 	poll = (xfer->x_flags & BD_XFER_POLL) != 0;
3355 
3356 	ret = nvme_submit_io_cmd(ioq, cmd);
3357 
3358 	if (ret != 0)
3359 		return (ret);
3360 
3361 	if (!poll)
3362 		return (0);
3363 
3364 	do {
3365 		cmd = nvme_retrieve_cmd(nvme, ioq);
3366 		if (cmd != NULL)
3367 			nvme_bd_xfer_done(cmd);
3368 		else
3369 			drv_usecwait(10);
3370 	} while (ioq->nq_active_cmds != 0);
3371 
3372 	return (0);
3373 }
3374 
3375 static int
3376 nvme_bd_read(void *arg, bd_xfer_t *xfer)
3377 {
3378 	nvme_namespace_t *ns = arg;
3379 
3380 	return (nvme_bd_cmd(ns, xfer, NVME_OPC_NVM_READ));
3381 }
3382 
3383 static int
3384 nvme_bd_write(void *arg, bd_xfer_t *xfer)
3385 {
3386 	nvme_namespace_t *ns = arg;
3387 
3388 	return (nvme_bd_cmd(ns, xfer, NVME_OPC_NVM_WRITE));
3389 }
3390 
3391 static int
3392 nvme_bd_sync(void *arg, bd_xfer_t *xfer)
3393 {
3394 	nvme_namespace_t *ns = arg;
3395 
3396 	if (ns->ns_nvme->n_dead)
3397 		return (EIO);
3398 
3399 	/*
3400 	 * If the volatile write cache is not present or not enabled the FLUSH
3401 	 * command is a no-op, so we can take a shortcut here.
3402 	 */
3403 	if (!ns->ns_nvme->n_write_cache_present) {
3404 		bd_xfer_done(xfer, ENOTSUP);
3405 		return (0);
3406 	}
3407 
3408 	if (!ns->ns_nvme->n_write_cache_enabled) {
3409 		bd_xfer_done(xfer, 0);
3410 		return (0);
3411 	}
3412 
3413 	return (nvme_bd_cmd(ns, xfer, NVME_OPC_NVM_FLUSH));
3414 }
3415 
3416 static int
3417 nvme_bd_devid(void *arg, dev_info_t *devinfo, ddi_devid_t *devid)
3418 {
3419 	nvme_namespace_t *ns = arg;
3420 
3421 	/*LINTED: E_BAD_PTR_CAST_ALIGN*/
3422 	if (*(uint64_t *)ns->ns_eui64 != 0) {
3423 		return (ddi_devid_init(devinfo, DEVID_SCSI3_WWN,
3424 		    sizeof (ns->ns_eui64), ns->ns_eui64, devid));
3425 	} else {
3426 		return (ddi_devid_init(devinfo, DEVID_ENCAP,
3427 		    strlen(ns->ns_devid), ns->ns_devid, devid));
3428 	}
3429 }
3430 
3431 static int
3432 nvme_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
3433 {
3434 #ifndef __lock_lint
3435 	_NOTE(ARGUNUSED(cred_p));
3436 #endif
3437 	minor_t minor = getminor(*devp);
3438 	nvme_t *nvme = ddi_get_soft_state(nvme_state, NVME_MINOR_INST(minor));
3439 	int nsid = NVME_MINOR_NSID(minor);
3440 	nvme_minor_state_t *nm;
3441 	int rv = 0;
3442 
3443 	if (otyp != OTYP_CHR)
3444 		return (EINVAL);
3445 
3446 	if (nvme == NULL)
3447 		return (ENXIO);
3448 
3449 	if (nsid > nvme->n_namespace_count)
3450 		return (ENXIO);
3451 
3452 	if (nvme->n_dead)
3453 		return (EIO);
3454 
3455 	nm = nsid == 0 ? &nvme->n_minor : &nvme->n_ns[nsid - 1].ns_minor;
3456 
3457 	mutex_enter(&nm->nm_mutex);
3458 	if (nm->nm_oexcl) {
3459 		rv = EBUSY;
3460 		goto out;
3461 	}
3462 
3463 	if (flag & FEXCL) {
3464 		if (nm->nm_ocnt != 0) {
3465 			rv = EBUSY;
3466 			goto out;
3467 		}
3468 		nm->nm_oexcl = B_TRUE;
3469 	}
3470 
3471 	nm->nm_ocnt++;
3472 
3473 out:
3474 	mutex_exit(&nm->nm_mutex);
3475 	return (rv);
3476 
3477 }
3478 
3479 static int
3480 nvme_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
3481 {
3482 #ifndef __lock_lint
3483 	_NOTE(ARGUNUSED(cred_p));
3484 	_NOTE(ARGUNUSED(flag));
3485 #endif
3486 	minor_t minor = getminor(dev);
3487 	nvme_t *nvme = ddi_get_soft_state(nvme_state, NVME_MINOR_INST(minor));
3488 	int nsid = NVME_MINOR_NSID(minor);
3489 	nvme_minor_state_t *nm;
3490 
3491 	if (otyp != OTYP_CHR)
3492 		return (ENXIO);
3493 
3494 	if (nvme == NULL)
3495 		return (ENXIO);
3496 
3497 	if (nsid > nvme->n_namespace_count)
3498 		return (ENXIO);
3499 
3500 	nm = nsid == 0 ? &nvme->n_minor : &nvme->n_ns[nsid - 1].ns_minor;
3501 
3502 	mutex_enter(&nm->nm_mutex);
3503 	if (nm->nm_oexcl)
3504 		nm->nm_oexcl = B_FALSE;
3505 
3506 	ASSERT(nm->nm_ocnt > 0);
3507 	nm->nm_ocnt--;
3508 	mutex_exit(&nm->nm_mutex);
3509 
3510 	return (0);
3511 }
3512 
3513 static int
3514 nvme_ioctl_identify(nvme_t *nvme, int nsid, nvme_ioctl_t *nioc, int mode,
3515     cred_t *cred_p)
3516 {
3517 	_NOTE(ARGUNUSED(cred_p));
3518 	int rv = 0;
3519 	void *idctl;
3520 
3521 	if ((mode & FREAD) == 0)
3522 		return (EPERM);
3523 
3524 	if (nioc->n_len < NVME_IDENTIFY_BUFSIZE)
3525 		return (EINVAL);
3526 
3527 	if ((rv = nvme_identify(nvme, nsid, (void **)&idctl)) != 0)
3528 		return (rv);
3529 
3530 	if (ddi_copyout(idctl, (void *)nioc->n_buf, NVME_IDENTIFY_BUFSIZE, mode)
3531 	    != 0)
3532 		rv = EFAULT;
3533 
3534 	kmem_free(idctl, NVME_IDENTIFY_BUFSIZE);
3535 
3536 	return (rv);
3537 }
3538 
3539 static int
3540 nvme_ioctl_capabilities(nvme_t *nvme, int nsid, nvme_ioctl_t *nioc,
3541     int mode, cred_t *cred_p)
3542 {
3543 	_NOTE(ARGUNUSED(nsid, cred_p));
3544 	int rv = 0;
3545 	nvme_reg_cap_t cap = { 0 };
3546 	nvme_capabilities_t nc;
3547 
3548 	if ((mode & FREAD) == 0)
3549 		return (EPERM);
3550 
3551 	if (nioc->n_len < sizeof (nc))
3552 		return (EINVAL);
3553 
3554 	cap.r = nvme_get64(nvme, NVME_REG_CAP);
3555 
3556 	/*
3557 	 * The MPSMIN and MPSMAX fields in the CAP register use 0 to
3558 	 * specify the base page size of 4k (1<<12), so add 12 here to
3559 	 * get the real page size value.
3560 	 */
3561 	nc.mpsmax = 1 << (12 + cap.b.cap_mpsmax);
3562 	nc.mpsmin = 1 << (12 + cap.b.cap_mpsmin);
3563 
3564 	if (ddi_copyout(&nc, (void *)nioc->n_buf, sizeof (nc), mode) != 0)
3565 		rv = EFAULT;
3566 
3567 	return (rv);
3568 }
3569 
3570 static int
3571 nvme_ioctl_get_logpage(nvme_t *nvme, int nsid, nvme_ioctl_t *nioc,
3572     int mode, cred_t *cred_p)
3573 {
3574 	_NOTE(ARGUNUSED(cred_p));
3575 	void *log = NULL;
3576 	size_t bufsize = 0;
3577 	int rv = 0;
3578 
3579 	if ((mode & FREAD) == 0)
3580 		return (EPERM);
3581 
3582 	switch (nioc->n_arg) {
3583 	case NVME_LOGPAGE_ERROR:
3584 		if (nsid != 0)
3585 			return (EINVAL);
3586 		break;
3587 	case NVME_LOGPAGE_HEALTH:
3588 		if (nsid != 0 && nvme->n_idctl->id_lpa.lp_smart == 0)
3589 			return (EINVAL);
3590 
3591 		if (nsid == 0)
3592 			nsid = (uint32_t)-1;
3593 
3594 		break;
3595 	case NVME_LOGPAGE_FWSLOT:
3596 		if (nsid != 0)
3597 			return (EINVAL);
3598 		break;
3599 	default:
3600 		return (EINVAL);
3601 	}
3602 
3603 	if (nvme_get_logpage(nvme, &log, &bufsize, nioc->n_arg, nsid)
3604 	    != DDI_SUCCESS)
3605 		return (EIO);
3606 
3607 	if (nioc->n_len < bufsize) {
3608 		kmem_free(log, bufsize);
3609 		return (EINVAL);
3610 	}
3611 
3612 	if (ddi_copyout(log, (void *)nioc->n_buf, bufsize, mode) != 0)
3613 		rv = EFAULT;
3614 
3615 	nioc->n_len = bufsize;
3616 	kmem_free(log, bufsize);
3617 
3618 	return (rv);
3619 }
3620 
3621 static int
3622 nvme_ioctl_get_features(nvme_t *nvme, int nsid, nvme_ioctl_t *nioc,
3623     int mode, cred_t *cred_p)
3624 {
3625 	_NOTE(ARGUNUSED(cred_p));
3626 	void *buf = NULL;
3627 	size_t bufsize = 0;
3628 	uint32_t res = 0;
3629 	uint8_t feature;
3630 	int rv = 0;
3631 
3632 	if ((mode & FREAD) == 0)
3633 		return (EPERM);
3634 
3635 	if ((nioc->n_arg >> 32) > 0xff)
3636 		return (EINVAL);
3637 
3638 	feature = (uint8_t)(nioc->n_arg >> 32);
3639 
3640 	switch (feature) {
3641 	case NVME_FEAT_ARBITRATION:
3642 	case NVME_FEAT_POWER_MGMT:
3643 	case NVME_FEAT_TEMPERATURE:
3644 	case NVME_FEAT_ERROR:
3645 	case NVME_FEAT_NQUEUES:
3646 	case NVME_FEAT_INTR_COAL:
3647 	case NVME_FEAT_WRITE_ATOM:
3648 	case NVME_FEAT_ASYNC_EVENT:
3649 	case NVME_FEAT_PROGRESS:
3650 		if (nsid != 0)
3651 			return (EINVAL);
3652 		break;
3653 
3654 	case NVME_FEAT_INTR_VECT:
3655 		if (nsid != 0)
3656 			return (EINVAL);
3657 
3658 		res = nioc->n_arg & 0xffffffffUL;
3659 		if (res >= nvme->n_intr_cnt)
3660 			return (EINVAL);
3661 		break;
3662 
3663 	case NVME_FEAT_LBA_RANGE:
3664 		if (nvme->n_lba_range_supported == B_FALSE)
3665 			return (EINVAL);
3666 
3667 		if (nsid == 0 ||
3668 		    nsid > nvme->n_namespace_count)
3669 			return (EINVAL);
3670 
3671 		break;
3672 
3673 	case NVME_FEAT_WRITE_CACHE:
3674 		if (nsid != 0)
3675 			return (EINVAL);
3676 
3677 		if (!nvme->n_write_cache_present)
3678 			return (EINVAL);
3679 
3680 		break;
3681 
3682 	case NVME_FEAT_AUTO_PST:
3683 		if (nsid != 0)
3684 			return (EINVAL);
3685 
3686 		if (!nvme->n_auto_pst_supported)
3687 			return (EINVAL);
3688 
3689 		break;
3690 
3691 	default:
3692 		return (EINVAL);
3693 	}
3694 
3695 	rv = nvme_get_features(nvme, nsid, feature, &res, &buf, &bufsize);
3696 	if (rv != 0)
3697 		return (rv);
3698 
3699 	if (nioc->n_len < bufsize) {
3700 		kmem_free(buf, bufsize);
3701 		return (EINVAL);
3702 	}
3703 
3704 	if (buf && ddi_copyout(buf, (void*)nioc->n_buf, bufsize, mode) != 0)
3705 		rv = EFAULT;
3706 
3707 	kmem_free(buf, bufsize);
3708 	nioc->n_arg = res;
3709 	nioc->n_len = bufsize;
3710 
3711 	return (rv);
3712 }
3713 
3714 static int
3715 nvme_ioctl_intr_cnt(nvme_t *nvme, int nsid, nvme_ioctl_t *nioc, int mode,
3716     cred_t *cred_p)
3717 {
3718 	_NOTE(ARGUNUSED(nsid, mode, cred_p));
3719 
3720 	if ((mode & FREAD) == 0)
3721 		return (EPERM);
3722 
3723 	nioc->n_arg = nvme->n_intr_cnt;
3724 	return (0);
3725 }
3726 
3727 static int
3728 nvme_ioctl_version(nvme_t *nvme, int nsid, nvme_ioctl_t *nioc, int mode,
3729     cred_t *cred_p)
3730 {
3731 	_NOTE(ARGUNUSED(nsid, cred_p));
3732 	int rv = 0;
3733 
3734 	if ((mode & FREAD) == 0)
3735 		return (EPERM);
3736 
3737 	if (nioc->n_len < sizeof (nvme->n_version))
3738 		return (ENOMEM);
3739 
3740 	if (ddi_copyout(&nvme->n_version, (void *)nioc->n_buf,
3741 	    sizeof (nvme->n_version), mode) != 0)
3742 		rv = EFAULT;
3743 
3744 	return (rv);
3745 }
3746 
3747 static int
3748 nvme_ioctl_format(nvme_t *nvme, int nsid, nvme_ioctl_t *nioc, int mode,
3749     cred_t *cred_p)
3750 {
3751 	_NOTE(ARGUNUSED(mode));
3752 	nvme_format_nvm_t frmt = { 0 };
3753 	int c_nsid = nsid != 0 ? nsid - 1 : 0;
3754 
3755 	if ((mode & FWRITE) == 0 || secpolicy_sys_config(cred_p, B_FALSE) != 0)
3756 		return (EPERM);
3757 
3758 	frmt.r = nioc->n_arg & 0xffffffff;
3759 
3760 	/*
3761 	 * Check whether the FORMAT NVM command is supported.
3762 	 */
3763 	if (nvme->n_idctl->id_oacs.oa_format == 0)
3764 		return (EINVAL);
3765 
3766 	/*
3767 	 * Don't allow format or secure erase of individual namespace if that
3768 	 * would cause a format or secure erase of all namespaces.
3769 	 */
3770 	if (nsid != 0 && nvme->n_idctl->id_fna.fn_format != 0)
3771 		return (EINVAL);
3772 
3773 	if (nsid != 0 && frmt.b.fm_ses != NVME_FRMT_SES_NONE &&
3774 	    nvme->n_idctl->id_fna.fn_sec_erase != 0)
3775 		return (EINVAL);
3776 
3777 	/*
3778 	 * Don't allow formatting with Protection Information.
3779 	 */
3780 	if (frmt.b.fm_pi != 0 || frmt.b.fm_pil != 0 || frmt.b.fm_ms != 0)
3781 		return (EINVAL);
3782 
3783 	/*
3784 	 * Don't allow formatting using an illegal LBA format, or any LBA format
3785 	 * that uses metadata.
3786 	 */
3787 	if (frmt.b.fm_lbaf > nvme->n_ns[c_nsid].ns_idns->id_nlbaf ||
3788 	    nvme->n_ns[c_nsid].ns_idns->id_lbaf[frmt.b.fm_lbaf].lbaf_ms != 0)
3789 		return (EINVAL);
3790 
3791 	/*
3792 	 * Don't allow formatting using an illegal Secure Erase setting.
3793 	 */
3794 	if (frmt.b.fm_ses > NVME_FRMT_MAX_SES ||
3795 	    (frmt.b.fm_ses == NVME_FRMT_SES_CRYPTO &&
3796 	    nvme->n_idctl->id_fna.fn_crypt_erase == 0))
3797 		return (EINVAL);
3798 
3799 	if (nsid == 0)
3800 		nsid = (uint32_t)-1;
3801 
3802 	return (nvme_format_nvm(nvme, nsid, frmt.b.fm_lbaf, B_FALSE, 0, B_FALSE,
3803 	    frmt.b.fm_ses));
3804 }
3805 
3806 static int
3807 nvme_ioctl_detach(nvme_t *nvme, int nsid, nvme_ioctl_t *nioc, int mode,
3808     cred_t *cred_p)
3809 {
3810 	_NOTE(ARGUNUSED(nioc, mode));
3811 	int rv = 0;
3812 
3813 	if ((mode & FWRITE) == 0 || secpolicy_sys_config(cred_p, B_FALSE) != 0)
3814 		return (EPERM);
3815 
3816 	if (nsid == 0)
3817 		return (EINVAL);
3818 
3819 	rv = bd_detach_handle(nvme->n_ns[nsid - 1].ns_bd_hdl);
3820 	if (rv != DDI_SUCCESS)
3821 		rv = EBUSY;
3822 
3823 	return (rv);
3824 }
3825 
3826 static int
3827 nvme_ioctl_attach(nvme_t *nvme, int nsid, nvme_ioctl_t *nioc, int mode,
3828     cred_t *cred_p)
3829 {
3830 	_NOTE(ARGUNUSED(nioc, mode));
3831 	nvme_identify_nsid_t *idns;
3832 	int rv = 0;
3833 
3834 	if ((mode & FWRITE) == 0 || secpolicy_sys_config(cred_p, B_FALSE) != 0)
3835 		return (EPERM);
3836 
3837 	if (nsid == 0)
3838 		return (EINVAL);
3839 
3840 	/*
3841 	 * Identify namespace again, free old identify data.
3842 	 */
3843 	idns = nvme->n_ns[nsid - 1].ns_idns;
3844 	if (nvme_init_ns(nvme, nsid) != DDI_SUCCESS)
3845 		return (EIO);
3846 
3847 	kmem_free(idns, sizeof (nvme_identify_nsid_t));
3848 
3849 	rv = bd_attach_handle(nvme->n_dip, nvme->n_ns[nsid - 1].ns_bd_hdl);
3850 	if (rv != DDI_SUCCESS)
3851 		rv = EBUSY;
3852 
3853 	return (rv);
3854 }
3855 
3856 static int
3857 nvme_ioctl(dev_t dev, int cmd, intptr_t arg, int mode, cred_t *cred_p,
3858     int *rval_p)
3859 {
3860 #ifndef __lock_lint
3861 	_NOTE(ARGUNUSED(rval_p));
3862 #endif
3863 	minor_t minor = getminor(dev);
3864 	nvme_t *nvme = ddi_get_soft_state(nvme_state, NVME_MINOR_INST(minor));
3865 	int nsid = NVME_MINOR_NSID(minor);
3866 	int rv = 0;
3867 	nvme_ioctl_t nioc;
3868 
3869 	int (*nvme_ioctl[])(nvme_t *, int, nvme_ioctl_t *, int, cred_t *) = {
3870 		NULL,
3871 		nvme_ioctl_identify,
3872 		nvme_ioctl_identify,
3873 		nvme_ioctl_capabilities,
3874 		nvme_ioctl_get_logpage,
3875 		nvme_ioctl_get_features,
3876 		nvme_ioctl_intr_cnt,
3877 		nvme_ioctl_version,
3878 		nvme_ioctl_format,
3879 		nvme_ioctl_detach,
3880 		nvme_ioctl_attach
3881 	};
3882 
3883 	if (nvme == NULL)
3884 		return (ENXIO);
3885 
3886 	if (nsid > nvme->n_namespace_count)
3887 		return (ENXIO);
3888 
3889 	if (IS_DEVCTL(cmd))
3890 		return (ndi_devctl_ioctl(nvme->n_dip, cmd, arg, mode, 0));
3891 
3892 #ifdef _MULTI_DATAMODEL
3893 	switch (ddi_model_convert_from(mode & FMODELS)) {
3894 	case DDI_MODEL_ILP32: {
3895 		nvme_ioctl32_t nioc32;
3896 		if (ddi_copyin((void*)arg, &nioc32, sizeof (nvme_ioctl32_t),
3897 		    mode) != 0)
3898 			return (EFAULT);
3899 		nioc.n_len = nioc32.n_len;
3900 		nioc.n_buf = nioc32.n_buf;
3901 		nioc.n_arg = nioc32.n_arg;
3902 		break;
3903 	}
3904 	case DDI_MODEL_NONE:
3905 #endif
3906 		if (ddi_copyin((void*)arg, &nioc, sizeof (nvme_ioctl_t), mode)
3907 		    != 0)
3908 			return (EFAULT);
3909 #ifdef _MULTI_DATAMODEL
3910 		break;
3911 	}
3912 #endif
3913 
3914 	if (nvme->n_dead && cmd != NVME_IOC_DETACH)
3915 		return (EIO);
3916 
3917 
3918 	if (cmd == NVME_IOC_IDENTIFY_CTRL) {
3919 		/*
3920 		 * This makes NVME_IOC_IDENTIFY_CTRL work the same on devctl and
3921 		 * attachment point nodes.
3922 		 */
3923 		nsid = 0;
3924 	} else if (cmd == NVME_IOC_IDENTIFY_NSID && nsid == 0) {
3925 		/*
3926 		 * This makes NVME_IOC_IDENTIFY_NSID work on a devctl node, it
3927 		 * will always return identify data for namespace 1.
3928 		 */
3929 		nsid = 1;
3930 	}
3931 
3932 	if (IS_NVME_IOC(cmd) && nvme_ioctl[NVME_IOC_CMD(cmd)] != NULL)
3933 		rv = nvme_ioctl[NVME_IOC_CMD(cmd)](nvme, nsid, &nioc, mode,
3934 		    cred_p);
3935 	else
3936 		rv = EINVAL;
3937 
3938 #ifdef _MULTI_DATAMODEL
3939 	switch (ddi_model_convert_from(mode & FMODELS)) {
3940 	case DDI_MODEL_ILP32: {
3941 		nvme_ioctl32_t nioc32;
3942 
3943 		nioc32.n_len = (size32_t)nioc.n_len;
3944 		nioc32.n_buf = (uintptr32_t)nioc.n_buf;
3945 		nioc32.n_arg = nioc.n_arg;
3946 
3947 		if (ddi_copyout(&nioc32, (void *)arg, sizeof (nvme_ioctl32_t),
3948 		    mode) != 0)
3949 			return (EFAULT);
3950 		break;
3951 	}
3952 	case DDI_MODEL_NONE:
3953 #endif
3954 		if (ddi_copyout(&nioc, (void *)arg, sizeof (nvme_ioctl_t), mode)
3955 		    != 0)
3956 			return (EFAULT);
3957 #ifdef _MULTI_DATAMODEL
3958 		break;
3959 	}
3960 #endif
3961 
3962 	return (rv);
3963 }
3964