xref: /illumos-gate/usr/src/uts/common/io/nvme/nvme.c (revision ef86b95d1dbad918184489140320c3b7dd69e4ec)
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 (c) 2016 The MathWorks, Inc.  All rights reserved.
14  * Copyright 2019 Unix Software Ltd.
15  * Copyright 2020 Joyent, Inc.
16  * Copyright 2020 Racktop Systems.
17  * Copyright 2024 Oxide Computer Company.
18  * Copyright 2022 OmniOS Community Edition (OmniOSce) Association.
19  * Copyright 2022 Tintri by DDN, Inc. All rights reserved.
20  */
21 
22 /*
23  * blkdev driver for NVMe compliant storage devices
24  *
25  * This driver targets and is designed to support all NVMe 1.x and NVMe 2.x
26  * devices. Features are added to the driver as we encounter devices that
27  * require them and our needs, so some commands or log pages may not take
28  * advantage of newer features that devices support at this time. When you
29  * encounter such a case, it is generally fine to add that support to the driver
30  * as long as you take care to ensure that the requisite device version is met
31  * before using it.
32  *
33  * The driver has only been tested on x86 systems and will not work on big-
34  * endian systems without changes to the code accessing registers and data
35  * structures used by the hardware.
36  *
37  *
38  * Interrupt Usage:
39  *
40  * The driver will use a single interrupt while configuring the device as the
41  * specification requires, but contrary to the specification it will try to use
42  * a single-message MSI(-X) or FIXED interrupt. Later in the attach process it
43  * will switch to multiple-message MSI(-X) if supported. The driver wants to
44  * have one interrupt vector per CPU, but it will work correctly if less are
45  * available. Interrupts can be shared by queues, the interrupt handler will
46  * iterate through the I/O queue array by steps of n_intr_cnt. Usually only
47  * the admin queue will share an interrupt with one I/O queue. The interrupt
48  * handler will retrieve completed commands from all queues sharing an interrupt
49  * vector and will post them to a taskq for completion processing.
50  *
51  *
52  * Command Processing:
53  *
54  * NVMe devices can have up to 65535 I/O queue pairs, with each queue holding up
55  * to 65536 I/O commands. The driver will configure one I/O queue pair per
56  * available interrupt vector, with the queue length usually much smaller than
57  * the maximum of 65536. If the hardware doesn't provide enough queues, fewer
58  * interrupt vectors will be used.
59  *
60  * Additionally the hardware provides a single special admin queue pair that can
61  * hold up to 4096 admin commands.
62  *
63  * From the hardware perspective both queues of a queue pair are independent,
64  * but they share some driver state: the command array (holding pointers to
65  * commands currently being processed by the hardware) and the active command
66  * counter. Access to a submission queue and the shared state is protected by
67  * nq_mutex; completion queue is protected by ncq_mutex.
68  *
69  * When a command is submitted to a queue pair the active command counter is
70  * incremented and a pointer to the command is stored in the command array. The
71  * array index is used as command identifier (CID) in the submission queue
72  * entry. Some commands may take a very long time to complete, and if the queue
73  * wraps around in that time a submission may find the next array slot to still
74  * be used by a long-running command. In this case the array is sequentially
75  * searched for the next free slot. The length of the command array is the same
76  * as the configured queue length. Queue overrun is prevented by the semaphore,
77  * so a command submission may block if the queue is full.
78  *
79  *
80  * Polled I/O Support:
81  *
82  * For kernel core dump support the driver can do polled I/O. As interrupts are
83  * turned off while dumping the driver will just submit a command in the regular
84  * way, and then repeatedly attempt a command retrieval until it gets the
85  * command back.
86  *
87  *
88  * Namespace Support:
89  *
90  * NVMe devices can have multiple namespaces, each being a independent data
91  * store. The driver supports multiple namespaces and creates a blkdev interface
92  * for each namespace found. Namespaces can have various attributes to support
93  * protection information. This driver does not support any of this and ignores
94  * namespaces that have these attributes.
95  *
96  * As of NVMe 1.1 namespaces can have an 64bit Extended Unique Identifier
97  * (EUI64), and NVMe 1.2 introduced an additional 128bit Namespace Globally
98  * Unique Identifier (NGUID). This driver uses either the NGUID or the EUI64
99  * if present to generate the devid, and passes the EUI64 to blkdev to use it
100  * in the device node names.
101  *
102  * We currently support only (2 << NVME_MINOR_INST_SHIFT) - 2 namespaces in a
103  * single controller. This is an artificial limit imposed by the driver to be
104  * able to address a reasonable number of controllers and namespaces while
105  * fitting within the constraints of MAXMIN32, aka a 32-bit device number which
106  * only has 18-bits for the minor number. See the minor node section for more
107  * information.
108  *
109  *
110  * Minor nodes:
111  *
112  * For each NVMe device the driver exposes one minor node for the controller and
113  * one minor node for each namespace. The only operations supported by those
114  * minor nodes are open(9E), close(9E), and ioctl(9E). This serves as the
115  * primary control interface for the devices. The character device is a private
116  * interface and we attempt stability through libnvme and more so nvmeadm.
117  *
118  * The controller minor node is much more flexible than the namespace minor node
119  * and should be preferred. The controller node allows one to target any
120  * namespace that the device has, while the namespace is limited in what it can
121  * acquire. While the namespace minor exists, it should not be relied upon and
122  * is not by libnvme.
123  *
124  * The minor number space is split in two. We use the lower part to support the
125  * controller and namespaces as described above in the 'Namespace Support'
126  * section. The second set is used for cloning opens. We set aside one million
127  * minors for this purpose. We utilize a cloning open so that way we can have
128  * per-file_t state. This is how we end up implementing and tracking locking
129  * state and related.
130  *
131  * When we have this cloned open, then we allocate a new nvme_minor_t which gets
132  * its minor number from the nvme_open_minors id_space_t and is stored in the
133  * nvme_open_minors_avl. While someone calls open on a controller or namespace
134  * minor, everything else occurs in the context of one of these ephemeral
135  * minors.
136  *
137  *
138  * ioctls, Errors, and Exclusive Access:
139  *
140  * All of the logical commands that one can issue are driven through the
141  * ioctl(9E) interface. All of our ioctls have a similar shape where they
142  * all include the 'nvme_ioctl_common_t' as their first member.
143  *
144  * This common ioctl structure is used to communicate the namespace that should
145  * be targeted. When the namespace is left as 0, then that indicates that it
146  * should target whatever the default is of the minor node. For a namespace
147  * minor, that will be transparently rewritten to the namespace's namespace id.
148  *
149  * In addition, the nvme_ioctl_common_t structure also has a standard error
150  * return. Our goal in our ioctl path is to ensure that we have useful semantic
151  * errors as much as possible. EINVAL, EIO, etc. are all overloaded. Instead as
152  * long as we can copy in our structure, then we will set a semantic error. If
153  * we have an error from the controller, then that will be included there.
154  *
155  * Each command has a specific policy that controls whether or not it is allowed
156  * on the namespace or controller minor, whether the broadcast namespace is
157  * allowed, various settings around what kind of exclusive access is allowed,
158  * and more. Each of these is wrapped up in a bit of policy described by the
159  * 'nvme_ioctl_check_t' structure.
160  *
161  * The device provides a form of exclusion in the form of both a
162  * controller-level and namespace-level read and write lock. Most operations do
163  * not require a lock (e.g. get log page, identify, etc.), but a few do (e.g.
164  * format nvm, firmware related activity, etc.). A read lock guarantees that you
165  * can complete your operation without interference, but read locks are not
166  * required. If you don't take a read lock and someone comes in with a write
167  * lock, then subsequent operations will fail with a semantic error indicating
168  * that you were blocked due to this.
169  *
170  * Here are some of the rules that govern our locks:
171  *
172  * 1. Writers starve readers. Any readers are allowed to finish when there is a
173  *    pending writer; however, all subsequent readers will be blocked upon that
174  *    writer.
175  * 2. A controller write lock takes priority over all other locks. Put
176  *    differently a controller writer not only starves subsequent controller
177  *    readers, but also all namespace read and write locks.
178  * 3. Each namespace lock is independent.
179  * 4. At most a single namespace lock may be owned.
180  * 5. If you own a namespace lock, you may not take a controller lock (to help
181  *    with lock ordering).
182  * 6. In a similar spirit, if you own a controller write lock, you may not take
183  *    any namespace lock. Someone with the controller write lock can perform any
184  *    operations that they need to. However, if you have a controller read lock
185  *    you may take any namespace lock.
186  * 7. There is no ability to upgrade a read lock to a write lock.
187  * 8. There is no recursive locking.
188  *
189  * While there's a lot there to keep track of, the goals of these are to
190  * constrain things so as to avoid deadlock. This is more complex than the
191  * original implementation in the driver which only allowed for an exclusive
192  * open that was tied to the thread. The first issue with tying this to the
193  * thread was that that didn't work well for software that utilized thread
194  * pools, like complex daemons. The second issue is that we want the ability for
195  * daemons, such as a FRU monitor, to be able to retain a file descriptor to the
196  * device without blocking others from taking action except during critical
197  * periods.
198  *
199  * In particular to enable something like libnvme, we didn't want someone to
200  * have to open and close the file descriptor to change what kind of exclusive
201  * access they desired.
202  *
203  * There are two different sets of data structures that we employ for tracking
204  * locking information:
205  *
206  * 1) The nvme_lock_t structure is contained in both the nvme_t and the
207  * nvme_namespace_t and tracks the current writer, readers, and pending writers
208  * and readers. Each of these lists or the writer pointer all refer to our
209  * second data structure.
210  *
211  * When a lock is owned by a single writer, then the nl_writer field is set to a
212  * specific minor's lock data structure. If instead readers are present, then
213  * the nl_readers list_t is not empty. An invariant of the system is that if
214  * nl_writer is non-NULL, nl_readers must be empty and conversely, if nl_readers
215  * is not empty, nl_writer must be NULL.
216  *
217  * 2) The nvme_minor_lock_info_t exists in the nvme_minor_t. There is one
218  * information structure which represents the minor's controller lock and a
219  * second one that represents the minor's namespace lock. The members of this
220  * are broken into tracking what the current lock is and what it targets. It
221  * also several members that are intended for debugging (nli_last_change,
222  * nli_acq_kthread, etc.).
223  *
224  * While the minor has two different lock information structures, our rules
225  * ensure that only one of the two can be pending and that they shouldn't result
226  * in a deadlock. When a lock is pending, the caller is sleeping on the minor's
227  * nm_cv member.
228  *
229  * These relationships are represented in the following image which shows a
230  * controller write lock being held with a pending readers on the controller
231  * lock and pending writers on one of the controller's namespaces.
232  *
233  *  +---------+
234  *  | nvme_t  |
235  *  |         |
236  *  | n_lock -|-------+
237  *  | n_ns -+ |       |                          +-----------------------------+
238  *  +-------|-+   +-----------------+            | nvme_minor_t                |
239  *          |     | nvme_lock_t     |            |                             |
240  *          |     |                 |            |  +------------------------+ |
241  *          |     | writer        --|-------------->| nvme_minor_lock_info_t | |
242  *          |     | reader list     |            |  | nm_ctrl_lock           | |
243  *          |     | pending writers |            |  +------------------------+ |
244  *          |     | pending readers |------+     |  +------------------------+ |
245  *          |     +-----------------+      |     |  | nvme_minor_lock_info_t | |
246  *          |                              |     |  | nm_ns_lock             | |
247  *          |                              |     |  +------------------------+ |
248  *          |                              |     +-----------------------------+
249  *  +------------------+                   |                 +-----------------+
250  *  | nvme_namespace_t |                   |                 | nvme_minor_t    |
251  *  |                  |                   |                 |                 |
252  *  | ns_lock ---+     |                   |                 | +-------------+ |
253  *  +------------|-----+                   +-----------------|>|nm_ctrl_lock | |
254  *               |                                           | +-------------+ |
255  *               v                                           +-----------------+
256  *     +------------------+                                         ...
257  *     | nvme_lock_t      |                                  +-----------------+
258  *     |                  |                                  | nvme_minor_t    |
259  *     | writer           |                                  |                 |
260  *     | reader list      |                                  | +-------------+ |
261  *     | pending writers -|-----------------+                | |nm_ctrl_lock | |
262  *     | pending readers  |                 |                | +-------------+ |
263  *     +------------------+                 |                +-----------------+
264  *         +-----------------------------+  |  +-----------------------------+
265  *         | nvme_minor_t                |  |  | nvme_minor_t                |
266  *         |                             |  |  |                             |
267  *         |  +------------------------+ |  |  |  +------------------------+ |
268  *         |  | nvme_minor_lock_info_t | |  |  |  | nvme_minor_lock_info_t | |
269  *         |  | nm_ctrl_lock           | |  |  |  | nm_ctrl_lock           | |
270  *         |  +------------------------+ |  |  |  +------------------------+ |
271  *         |  +------------------------+ |  v  |  +------------------------+ |
272  *         |  | nvme_minor_lock_info_t |-|-----|->| nvme_minor_lock_info_t | |
273  *         |  | nm_ns_lock             | |     |  | nm_ns_lock             | |
274  *         |  +------------------------+ |     |  +------------------------+ |
275  *         +-----------------------------+     +-----------------------------+
276  *
277  * Blkdev Interface:
278  *
279  * This driver uses blkdev to do all the heavy lifting involved with presenting
280  * a disk device to the system. As a result, the processing of I/O requests is
281  * relatively simple as blkdev takes care of partitioning, boundary checks, DMA
282  * setup, and splitting of transfers into manageable chunks.
283  *
284  * I/O requests coming in from blkdev are turned into NVM commands and posted to
285  * an I/O queue. The queue is selected by taking the CPU id modulo the number of
286  * queues. There is currently no timeout handling of I/O commands.
287  *
288  * Blkdev also supports querying device/media information and generating a
289  * devid. The driver reports the best block size as determined by the namespace
290  * format back to blkdev as physical block size to support partition and block
291  * alignment. The devid is either based on the namespace GUID or EUI64, if
292  * present, or composed using the device vendor ID, model number, serial number,
293  * and the namespace ID.
294  *
295  *
296  * Error Handling:
297  *
298  * Error handling is currently limited to detecting fatal hardware errors,
299  * either by asynchronous events, or synchronously through command status or
300  * admin command timeouts. In case of severe errors the device is fenced off,
301  * all further requests will return EIO. FMA is then called to fault the device.
302  *
303  * The hardware has a limit for outstanding asynchronous event requests. Before
304  * this limit is known the driver assumes it is at least 1 and posts a single
305  * asynchronous request. Later when the limit is known more asynchronous event
306  * requests are posted to allow quicker reception of error information. When an
307  * asynchronous event is posted by the hardware the driver will parse the error
308  * status fields and log information or fault the device, depending on the
309  * severity of the asynchronous event. The asynchronous event request is then
310  * reused and posted to the admin queue again.
311  *
312  * On command completion the command status is checked for errors. In case of
313  * errors indicating a driver bug the driver panics. Almost all other error
314  * status values just cause EIO to be returned.
315  *
316  * Command timeouts are currently detected for all admin commands except
317  * asynchronous event requests. If a command times out and the hardware appears
318  * to be healthy the driver attempts to abort the command. The abort command
319  * timeout is a separate tunable but the original command timeout will be used
320  * if it is greater. If the abort times out too the driver assumes the device
321  * to be dead, fences it off, and calls FMA to retire it. In all other cases
322  * the aborted command should return immediately with a status indicating it
323  * was aborted, and the driver will wait indefinitely for that to happen. No
324  * timeout handling of normal I/O commands is presently done.
325  *
326  * Any command that times out due to the controller dropping dead will be put on
327  * nvme_lost_cmds list if it references DMA memory. This will prevent the DMA
328  * memory being reused by the system and later being written to by a "dead"
329  * NVMe controller.
330  *
331  *
332  * Locking:
333  *
334  * Each queue pair has a nq_mutex and ncq_mutex. The nq_mutex must be held
335  * when accessing shared state and submission queue registers, ncq_mutex
336  * is held when accessing completion queue state and registers.
337  * Callers of nvme_unqueue_cmd() must make sure that nq_mutex is held, while
338  * nvme_submit_{admin,io}_cmd() and nvme_retrieve_cmd() take care of both
339  * mutexes themselves.
340  *
341  * Each command also has its own nc_mutex, which is associated with the
342  * condition variable nc_cv. It is only used on admin commands which are run
343  * synchronously. In that case it must be held across calls to
344  * nvme_submit_{admin,io}_cmd() and nvme_wait_cmd(), which is taken care of by
345  * nvme_admin_cmd(). It must also be held whenever the completion state of the
346  * command is changed or while an admin command timeout is handled.
347  *
348  * If both nc_mutex and nq_mutex must be held, nc_mutex must be acquired first.
349  * More than one nc_mutex may only be held when aborting commands. In this case,
350  * the nc_mutex of the command to be aborted must be held across the call to
351  * nvme_abort_cmd() to prevent the command from completing while the abort is in
352  * progress.
353  *
354  * If both nq_mutex and ncq_mutex need to be held, ncq_mutex must be
355  * acquired first. More than one nq_mutex is never held by a single thread.
356  * The ncq_mutex is only held by nvme_retrieve_cmd() and
357  * nvme_process_iocq(). nvme_process_iocq() is only called from the
358  * interrupt thread and nvme_retrieve_cmd() during polled I/O, so the
359  * mutex is non-contentious but is required for implementation completeness
360  * and safety.
361  *
362  * Each nvme_t has an n_admin_stat_mutex that protects the admin command
363  * statistics structure. If this is taken in conjunction with any other locks,
364  * then it must be taken last.
365  *
366  * There is one mutex n_minor_mutex which protects all open flags nm_open and
367  * exclusive-open thread pointers nm_oexcl of each minor node associated with a
368  * controller and its namespaces.
369  *
370  * In addition, there is a logical namespace management mutex which protects the
371  * data about namespaces. When interrogating the metadata of any namespace, this
372  * lock must be held. This gets tricky as we need to call into blkdev, which may
373  * issue callbacks into us which want this and it is illegal to hold locks
374  * across those blkdev calls as otherwise they might lead to deadlock (blkdev
375  * leverages ndi_devi_enter()).
376  *
377  * The lock exposes two levels, one that we call 'NVME' and one 'BDRO' or blkdev
378  * read-only. The idea is that most callers will use the NVME level which says
379  * this is a full traditional mutex operation. The BDRO level is used by blkdev
380  * callback functions and is a promise to only only read the data. When a blkdev
381  * operation starts, the lock holder will use nvme_mgmt_bd_start(). This
382  * strictly speaking drops the mutex, but records that the lock is logically
383  * held by the thread that did the start() operation.
384  *
385  * During this time, other threads (or even the same one) may end up calling
386  * into nvme_mgmt_lock(). Only one person may still hold the lock at any time;
387  * however, the BRDO level will be allowed to proceed during this time. This
388  * allows us to make consistent progress and honor the blkdev lock ordering
389  * requirements, albeit it is not as straightforward as a simple mutex.
390  *
391  * Quiesce / Fast Reboot:
392  *
393  * The driver currently does not support fast reboot. A quiesce(9E) entry point
394  * is still provided which is used to send a shutdown notification to the
395  * device.
396  *
397  *
398  * NVMe Hotplug:
399  *
400  * The driver supports hot removal. The driver uses the NDI event framework
401  * to register a callback, nvme_remove_callback, to clean up when a disk is
402  * removed. In particular, the driver will unqueue outstanding I/O commands and
403  * set n_dead on the softstate to true so that other operations, such as ioctls
404  * and command submissions, fail as well.
405  *
406  * While the callback registration relies on the NDI event framework, the
407  * removal event itself is kicked off in the PCIe hotplug framework, when the
408  * PCIe bridge driver ("pcieb") gets a hotplug interrupt indicating that a
409  * device was removed from the slot.
410  *
411  * The NVMe driver instance itself will remain until the final close of the
412  * device.
413  *
414  *
415  * DDI UFM Support
416  *
417  * The driver supports the DDI UFM framework for reporting information about
418  * the device's firmware image and slot configuration. This data can be
419  * queried by userland software via ioctls to the ufm driver. For more
420  * information, see ddi_ufm(9E).
421  *
422  *
423  * Driver Configuration:
424  *
425  * The following driver properties can be changed to control some aspects of the
426  * drivers operation:
427  * - strict-version: can be set to 0 to allow devices conforming to newer
428  *   major versions to be used
429  * - ignore-unknown-vendor-status: can be set to 1 to not handle any vendor
430  *   specific command status as a fatal error leading device faulting
431  * - admin-queue-len: the maximum length of the admin queue (16-4096)
432  * - io-squeue-len: the maximum length of the I/O submission queues (16-65536)
433  * - io-cqueue-len: the maximum length of the I/O completion queues (16-65536)
434  * - async-event-limit: the maximum number of asynchronous event requests to be
435  *   posted by the driver
436  * - volatile-write-cache-enable: can be set to 0 to disable the volatile write
437  *   cache
438  * - min-phys-block-size: the minimum physical block size to report to blkdev,
439  *   which is among other things the basis for ZFS vdev ashift
440  * - max-submission-queues: the maximum number of I/O submission queues.
441  * - max-completion-queues: the maximum number of I/O completion queues,
442  *   can be less than max-submission-queues, in which case the completion
443  *   queues are shared.
444  *
445  * In addition to the above properties, some device-specific tunables can be
446  * configured using the nvme-config-list global property. The value of this
447  * property is a list of triplets. The formal syntax is:
448  *
449  *   nvme-config-list ::= <triplet> [, <triplet>]* ;
450  *   <triplet>        ::= "<model>" , "<rev-list>" , "<tuple-list>"
451  *   <rev-list>       ::= [ <fwrev> [, <fwrev>]*]
452  *   <tuple-list>     ::= <tunable> [, <tunable>]*
453  *   <tunable>        ::= <name> : <value>
454  *
455  * The <model> and <fwrev> are the strings in nvme_identify_ctrl_t`id_model and
456  * nvme_identify_ctrl_t`id_fwrev, respectively. The remainder of <tuple-list>
457  * contains one or more tunables to apply to all controllers that match the
458  * specified model number and optionally firmware revision. Each <tunable> is a
459  * <name> : <value> pair.  Supported tunables are:
460  *
461  * - ignore-unknown-vendor-status:  can be set to "on" to not handle any vendor
462  *   specific command status as a fatal error leading device faulting
463  *
464  * - min-phys-block-size: the minimum physical block size to report to blkdev,
465  *   which is among other things the basis for ZFS vdev ashift
466  *
467  * - volatile-write-cache: can be set to "on" or "off" to enable or disable the
468  *   volatile write cache, if present
469  *
470  *
471  * TODO:
472  * - figure out sane default for I/O queue depth reported to blkdev
473  * - FMA handling of media errors
474  * - support for devices supporting very large I/O requests using chained PRPs
475  * - support for configuring hardware parameters like interrupt coalescing
476  * - support for media formatting and hard partitioning into namespaces
477  * - support for big-endian systems
478  * - support for fast reboot
479  * - support for NVMe Subsystem Reset (1.1)
480  * - support for Scatter/Gather lists (1.1)
481  * - support for Reservations (1.1)
482  * - support for power management
483  */
484 
485 #include <sys/byteorder.h>
486 #ifdef _BIG_ENDIAN
487 #error nvme driver needs porting for big-endian platforms
488 #endif
489 
490 #include <sys/modctl.h>
491 #include <sys/conf.h>
492 #include <sys/devops.h>
493 #include <sys/ddi.h>
494 #include <sys/ddi_ufm.h>
495 #include <sys/sunddi.h>
496 #include <sys/sunndi.h>
497 #include <sys/bitmap.h>
498 #include <sys/sysmacros.h>
499 #include <sys/param.h>
500 #include <sys/varargs.h>
501 #include <sys/cpuvar.h>
502 #include <sys/disp.h>
503 #include <sys/blkdev.h>
504 #include <sys/atomic.h>
505 #include <sys/archsystm.h>
506 #include <sys/sata/sata_hba.h>
507 #include <sys/stat.h>
508 #include <sys/policy.h>
509 #include <sys/list.h>
510 #include <sys/dkio.h>
511 #include <sys/pci.h>
512 #include <sys/mkdev.h>
513 
514 #include <sys/nvme.h>
515 
516 #ifdef __x86
517 #include <sys/x86_archext.h>
518 #endif
519 
520 #include "nvme_reg.h"
521 #include "nvme_var.h"
522 
523 /*
524  * Assertions to make sure that we've properly captured various aspects of the
525  * packed structures and haven't broken them during updates.
526  */
527 CTASSERT(sizeof (nvme_identify_ctrl_t) == NVME_IDENTIFY_BUFSIZE);
528 CTASSERT(offsetof(nvme_identify_ctrl_t, id_oacs) == 256);
529 CTASSERT(offsetof(nvme_identify_ctrl_t, id_sqes) == 512);
530 CTASSERT(offsetof(nvme_identify_ctrl_t, id_oncs) == 520);
531 CTASSERT(offsetof(nvme_identify_ctrl_t, id_subnqn) == 768);
532 CTASSERT(offsetof(nvme_identify_ctrl_t, id_nvmof) == 1792);
533 CTASSERT(offsetof(nvme_identify_ctrl_t, id_psd) == 2048);
534 CTASSERT(offsetof(nvme_identify_ctrl_t, id_vs) == 3072);
535 
536 CTASSERT(sizeof (nvme_identify_nsid_t) == NVME_IDENTIFY_BUFSIZE);
537 CTASSERT(offsetof(nvme_identify_nsid_t, id_fpi) == 32);
538 CTASSERT(offsetof(nvme_identify_nsid_t, id_anagrpid) == 92);
539 CTASSERT(offsetof(nvme_identify_nsid_t, id_nguid) == 104);
540 CTASSERT(offsetof(nvme_identify_nsid_t, id_lbaf) == 128);
541 CTASSERT(offsetof(nvme_identify_nsid_t, id_vs) == 384);
542 
543 CTASSERT(sizeof (nvme_identify_nsid_list_t) == NVME_IDENTIFY_BUFSIZE);
544 CTASSERT(sizeof (nvme_identify_ctrl_list_t) == NVME_IDENTIFY_BUFSIZE);
545 
546 CTASSERT(sizeof (nvme_identify_primary_caps_t) == NVME_IDENTIFY_BUFSIZE);
547 CTASSERT(offsetof(nvme_identify_primary_caps_t, nipc_vqfrt) == 32);
548 CTASSERT(offsetof(nvme_identify_primary_caps_t, nipc_vifrt) == 64);
549 
550 CTASSERT(sizeof (nvme_nschange_list_t) == 4096);
551 
552 /* NVMe spec version supported */
553 static const int nvme_version_major = 2;
554 
555 /* Tunable for FORMAT NVM command timeout in seconds, default is 600s */
556 uint32_t nvme_format_cmd_timeout = 600;
557 
558 /* Tunable for firmware commit with NVME_FWC_SAVE, default is 15s */
559 uint32_t nvme_commit_save_cmd_timeout = 15;
560 
561 /*
562  * Tunable for the admin command timeout used for commands other than those
563  * with their own timeouts defined above; in seconds. While most commands are
564  * expected to complete very quickly (sub-second), experience has shown that
565  * some controllers can occasionally be a bit slower, and not always consistent
566  * in the time taken - times of up to around 4.2s have been observed. Setting
567  * this to 15s by default provides headroom.
568  */
569 uint32_t nvme_admin_cmd_timeout = 15;
570 
571 /*
572  * Tunable for abort command timeout in seconds, default is 60s. This timeout
573  * is used when issuing an abort command, currently only in response to a
574  * different admin command timing out. Aborts always complete after the command
575  * that they are attempting to abort so we need to allow enough time for the
576  * controller to process the long running command that we are attempting to
577  * abort. The abort timeout here is only used if it is greater than the timeout
578  * for the command that is being aborted.
579  */
580 uint32_t nvme_abort_cmd_timeout = 60;
581 
582 /*
583  * Tunable for the size of arbitrary vendor specific admin commands,
584  * default is 16MiB.
585  */
586 uint32_t nvme_vendor_specific_admin_cmd_size = 1 << 24;
587 
588 /*
589  * Tunable for the max timeout of arbitary vendor specific admin commands,
590  * default is 60s.
591  */
592 uint_t nvme_vendor_specific_admin_cmd_max_timeout = 60;
593 
594 /*
595  * This ID space, AVL, and lock are used for keeping track of minor state across
596  * opens between different devices.
597  */
598 static id_space_t *nvme_open_minors;
599 static avl_tree_t nvme_open_minors_avl;
600 kmutex_t nvme_open_minors_mutex;
601 
602 /*
603  * Removal taskq used for n_dead callback processing.
604  */
605 taskq_t *nvme_dead_taskq;
606 
607 /*
608  * This enumeration is used in tandem with nvme_mgmt_lock() to describe which
609  * form of the lock is being taken. See the theory statement for more context.
610  */
611 typedef enum {
612 	/*
613 	 * This is the primary form of taking the management lock and indicates
614 	 * that the user intends to do a read/write of it. This should always be
615 	 * used for any ioctl paths or truly anything other than a blkdev
616 	 * information operation.
617 	 */
618 	NVME_MGMT_LOCK_NVME,
619 	/*
620 	 * This is a subordinate form of the lock whereby the user is in blkdev
621 	 * callback context and will only intend to read the namespace data.
622 	 */
623 	NVME_MGMT_LOCK_BDRO
624 } nvme_mgmt_lock_level_t;
625 
626 static int nvme_attach(dev_info_t *, ddi_attach_cmd_t);
627 static int nvme_detach(dev_info_t *, ddi_detach_cmd_t);
628 static int nvme_quiesce(dev_info_t *);
629 static int nvme_fm_errcb(dev_info_t *, ddi_fm_error_t *, const void *);
630 static int nvme_setup_interrupts(nvme_t *, int, int);
631 static void nvme_release_interrupts(nvme_t *);
632 static uint_t nvme_intr(caddr_t, caddr_t);
633 
634 static void nvme_shutdown(nvme_t *, boolean_t);
635 static boolean_t nvme_reset(nvme_t *, boolean_t);
636 static int nvme_init(nvme_t *);
637 static nvme_cmd_t *nvme_alloc_cmd(nvme_t *, int);
638 static void nvme_free_cmd(nvme_cmd_t *);
639 static nvme_cmd_t *nvme_create_nvm_cmd(nvme_namespace_t *, uint8_t,
640     bd_xfer_t *);
641 static void nvme_admin_cmd(nvme_cmd_t *, uint32_t);
642 static void nvme_submit_admin_cmd(nvme_qpair_t *, nvme_cmd_t *, uint32_t *);
643 static int nvme_submit_io_cmd(nvme_qpair_t *, nvme_cmd_t *);
644 static void nvme_submit_cmd_common(nvme_qpair_t *, nvme_cmd_t *, uint32_t *);
645 static nvme_cmd_t *nvme_unqueue_cmd(nvme_t *, nvme_qpair_t *, int);
646 static nvme_cmd_t *nvme_retrieve_cmd(nvme_t *, nvme_qpair_t *);
647 static void nvme_wait_cmd(nvme_cmd_t *, uint_t);
648 static void nvme_wakeup_cmd(void *);
649 static void nvme_async_event_task(void *);
650 
651 static int nvme_check_unknown_cmd_status(nvme_cmd_t *);
652 static int nvme_check_vendor_cmd_status(nvme_cmd_t *);
653 static int nvme_check_integrity_cmd_status(nvme_cmd_t *);
654 static int nvme_check_specific_cmd_status(nvme_cmd_t *);
655 static int nvme_check_generic_cmd_status(nvme_cmd_t *);
656 static inline int nvme_check_cmd_status(nvme_cmd_t *);
657 static boolean_t nvme_check_cmd_status_ioctl(nvme_cmd_t *,
658     nvme_ioctl_common_t *);
659 
660 static int nvme_abort_cmd(nvme_cmd_t *, const uint32_t);
661 static void nvme_async_event(nvme_t *);
662 static boolean_t nvme_format_nvm(nvme_t *, nvme_ioctl_format_t *);
663 static boolean_t nvme_get_logpage_int(nvme_t *, boolean_t, void **, size_t *,
664     uint8_t);
665 static boolean_t nvme_identify(nvme_t *, boolean_t, nvme_ioctl_identify_t *,
666     void **);
667 static boolean_t nvme_identify_int(nvme_t *, uint32_t, uint8_t, void **);
668 static int nvme_set_features(nvme_t *, boolean_t, uint32_t, uint8_t, uint32_t,
669     uint32_t *);
670 static int nvme_write_cache_set(nvme_t *, boolean_t);
671 static int nvme_set_nqueues(nvme_t *);
672 
673 static void nvme_free_dma(nvme_dma_t *);
674 static int nvme_zalloc_dma(nvme_t *, size_t, uint_t, ddi_dma_attr_t *,
675     nvme_dma_t **);
676 static int nvme_zalloc_queue_dma(nvme_t *, uint32_t, uint16_t, uint_t,
677     nvme_dma_t **);
678 static void nvme_free_qpair(nvme_qpair_t *);
679 static int nvme_alloc_qpair(nvme_t *, uint32_t, nvme_qpair_t **, uint_t);
680 static int nvme_create_io_qpair(nvme_t *, nvme_qpair_t *, uint16_t);
681 
682 static inline void nvme_put64(nvme_t *, uintptr_t, uint64_t);
683 static inline void nvme_put32(nvme_t *, uintptr_t, uint32_t);
684 static inline uint64_t nvme_get64(nvme_t *, uintptr_t);
685 static inline uint32_t nvme_get32(nvme_t *, uintptr_t);
686 
687 static boolean_t nvme_check_regs_hdl(nvme_t *);
688 static boolean_t nvme_check_dma_hdl(nvme_dma_t *);
689 
690 static int nvme_fill_prp(nvme_cmd_t *, ddi_dma_handle_t);
691 
692 static void nvme_bd_xfer_done(void *);
693 static void nvme_bd_driveinfo(void *, bd_drive_t *);
694 static int nvme_bd_mediainfo(void *, bd_media_t *);
695 static int nvme_bd_cmd(nvme_namespace_t *, bd_xfer_t *, uint8_t);
696 static int nvme_bd_read(void *, bd_xfer_t *);
697 static int nvme_bd_write(void *, bd_xfer_t *);
698 static int nvme_bd_sync(void *, bd_xfer_t *);
699 static int nvme_bd_devid(void *, dev_info_t *, ddi_devid_t *);
700 static int nvme_bd_free_space(void *, bd_xfer_t *);
701 
702 static int nvme_prp_dma_constructor(void *, void *, int);
703 static void nvme_prp_dma_destructor(void *, void *);
704 
705 static void nvme_prepare_devid(nvme_t *, uint32_t);
706 
707 /* DDI UFM callbacks */
708 static int nvme_ufm_fill_image(ddi_ufm_handle_t *, void *, uint_t,
709     ddi_ufm_image_t *);
710 static int nvme_ufm_fill_slot(ddi_ufm_handle_t *, void *, uint_t, uint_t,
711     ddi_ufm_slot_t *);
712 static int nvme_ufm_getcaps(ddi_ufm_handle_t *, void *, ddi_ufm_cap_t *);
713 
714 static int nvme_open(dev_t *, int, int, cred_t *);
715 static int nvme_close(dev_t, int, int, cred_t *);
716 static int nvme_ioctl(dev_t, int, intptr_t, int, cred_t *, int *);
717 
718 static int nvme_init_ns(nvme_t *, uint32_t);
719 static boolean_t nvme_attach_ns(nvme_t *, nvme_ioctl_common_t *);
720 static boolean_t nvme_detach_ns(nvme_t *, nvme_ioctl_common_t *);
721 
722 static int nvme_minor_comparator(const void *, const void *);
723 
724 static ddi_ufm_ops_t nvme_ufm_ops = {
725 	NULL,
726 	nvme_ufm_fill_image,
727 	nvme_ufm_fill_slot,
728 	nvme_ufm_getcaps
729 };
730 
731 /*
732  * Minor numbers are split amongst those used for controllers and for device
733  * opens. The number of controller minors are limited based open MAXMIN32 per
734  * the theory statement. We allocate 1 million minors as a total guess at a
735  * number that'll probably be enough. The starting point of the open minors can
736  * be shifted to accommodate future expansion of the NVMe device minors.
737  */
738 #define	NVME_MINOR_INST_SHIFT	9
739 #define	NVME_MINOR(inst, nsid)	(((inst) << NVME_MINOR_INST_SHIFT) | (nsid))
740 #define	NVME_MINOR_INST(minor)	((minor) >> NVME_MINOR_INST_SHIFT)
741 #define	NVME_MINOR_NSID(minor)	((minor) & ((1 << NVME_MINOR_INST_SHIFT) - 1))
742 #define	NVME_MINOR_MAX		(NVME_MINOR(1, 0) - 2)
743 
744 #define	NVME_OPEN_NMINORS		(1024 * 1024)
745 #define	NVME_OPEN_MINOR_MIN		(MAXMIN32 + 1)
746 #define	NVME_OPEN_MINOR_MAX_EXCL	(NVME_OPEN_MINOR_MIN + \
747     NVME_OPEN_NMINORS)
748 
749 #define	NVME_BUMP_STAT(nvme, stat)	\
750 	atomic_inc_64(&nvme->n_device_stat.nds_ ## stat.value.ui64)
751 
752 static void *nvme_state;
753 static kmem_cache_t *nvme_cmd_cache;
754 
755 /*
756  * DMA attributes for queue DMA memory
757  *
758  * Queue DMA memory must be page aligned. The maximum length of a queue is
759  * 65536 entries, and an entry can be 64 bytes long.
760  */
761 static const ddi_dma_attr_t nvme_queue_dma_attr = {
762 	.dma_attr_version	= DMA_ATTR_V0,
763 	.dma_attr_addr_lo	= 0,
764 	.dma_attr_addr_hi	= 0xffffffffffffffffULL,
765 	.dma_attr_count_max	= (UINT16_MAX + 1) * sizeof (nvme_sqe_t) - 1,
766 	.dma_attr_align		= 0x1000,
767 	.dma_attr_burstsizes	= 0x7ff,
768 	.dma_attr_minxfer	= 0x1000,
769 	.dma_attr_maxxfer	= (UINT16_MAX + 1) * sizeof (nvme_sqe_t),
770 	.dma_attr_seg		= 0xffffffffffffffffULL,
771 	.dma_attr_sgllen	= 1,
772 	.dma_attr_granular	= 1,
773 	.dma_attr_flags		= 0,
774 };
775 
776 /*
777  * DMA attributes for transfers using Physical Region Page (PRP) entries
778  *
779  * A PRP entry describes one page of DMA memory using the page size specified
780  * in the controller configuration's memory page size register (CC.MPS). It uses
781  * a 64bit base address aligned to this page size. There is no limitation on
782  * chaining PRPs together for arbitrarily large DMA transfers. These DMA
783  * attributes will be copied into the nvme_t during nvme_attach() and the
784  * dma_attr_maxxfer will be updated.
785  */
786 static const ddi_dma_attr_t nvme_prp_dma_attr = {
787 	.dma_attr_version	= DMA_ATTR_V0,
788 	.dma_attr_addr_lo	= 0,
789 	.dma_attr_addr_hi	= 0xffffffffffffffffULL,
790 	.dma_attr_count_max	= 0xfff,
791 	.dma_attr_align		= 0x1000,
792 	.dma_attr_burstsizes	= 0x7ff,
793 	.dma_attr_minxfer	= 0x1000,
794 	.dma_attr_maxxfer	= 0x1000,
795 	.dma_attr_seg		= 0xfff,
796 	.dma_attr_sgllen	= -1,
797 	.dma_attr_granular	= 1,
798 	.dma_attr_flags		= 0,
799 };
800 
801 /*
802  * DMA attributes for transfers using scatter/gather lists
803  *
804  * A SGL entry describes a chunk of DMA memory using a 64bit base address and a
805  * 32bit length field. SGL Segment and SGL Last Segment entries require the
806  * length to be a multiple of 16 bytes. While the SGL DMA attributes are copied
807  * into the nvme_t, they are not currently used for any I/O.
808  */
809 static const ddi_dma_attr_t nvme_sgl_dma_attr = {
810 	.dma_attr_version	= DMA_ATTR_V0,
811 	.dma_attr_addr_lo	= 0,
812 	.dma_attr_addr_hi	= 0xffffffffffffffffULL,
813 	.dma_attr_count_max	= 0xffffffffUL,
814 	.dma_attr_align		= 1,
815 	.dma_attr_burstsizes	= 0x7ff,
816 	.dma_attr_minxfer	= 0x10,
817 	.dma_attr_maxxfer	= 0xfffffffffULL,
818 	.dma_attr_seg		= 0xffffffffffffffffULL,
819 	.dma_attr_sgllen	= -1,
820 	.dma_attr_granular	= 0x10,
821 	.dma_attr_flags		= 0
822 };
823 
824 static ddi_device_acc_attr_t nvme_reg_acc_attr = {
825 	.devacc_attr_version	= DDI_DEVICE_ATTR_V0,
826 	.devacc_attr_endian_flags = DDI_STRUCTURE_LE_ACC,
827 	.devacc_attr_dataorder	= DDI_STRICTORDER_ACC
828 };
829 
830 /*
831  * ioctl validation policies. These are policies that determine which namespaces
832  * are allowed or disallowed for various operations. Note, all policy items
833  * should be explicitly listed here to help make it clear what our intent is.
834  * That is also why some of these are identical or repeated when they cover
835  * different ioctls.
836  */
837 
838 /*
839  * The controller information ioctl generally contains read-only information
840  * about the controller that is sourced from multiple different pieces of
841  * information. This does not operate on a namespace and none are accepted.
842  */
843 static const nvme_ioctl_check_t nvme_check_ctrl_info = {
844 	.nck_ns_ok = B_FALSE, .nck_ns_minor_ok = B_FALSE,
845 	.nck_skip_ctrl = B_FALSE, .nck_ctrl_rewrite = B_FALSE,
846 	.nck_bcast_ok = B_FALSE, .nck_excl = NVME_IOCTL_EXCL_NONE
847 };
848 
849 /*
850  * The kernel namespace information requires a namespace ID to be specified. It
851  * does not allow for the broadcast ID to be specified.
852  */
853 static const nvme_ioctl_check_t nvme_check_ns_info = {
854 	.nck_ns_ok = B_TRUE, .nck_ns_minor_ok = B_TRUE,
855 	.nck_skip_ctrl = B_FALSE, .nck_ctrl_rewrite = B_FALSE,
856 	.nck_bcast_ok = B_FALSE, .nck_excl = NVME_IOCTL_EXCL_NONE
857 };
858 
859 /*
860  * Identify commands are allowed to operate on a namespace minor. Unfortunately,
861  * the namespace field in identify commands is a bit, weird. In particular, some
862  * commands need a valid namespace, while others are namespace listing
863  * operations, which means illegal namespaces like zero are allowed.
864  */
865 static const nvme_ioctl_check_t nvme_check_identify = {
866 	.nck_ns_ok = B_TRUE, .nck_ns_minor_ok = B_TRUE,
867 	.nck_skip_ctrl = B_TRUE, .nck_ctrl_rewrite = B_FALSE,
868 	.nck_bcast_ok = B_TRUE, .nck_excl = NVME_IOCTL_EXCL_NONE
869 };
870 
871 /*
872  * The get log page command requires the ability to specify namespaces. When
873  * targeting the controller, one must use the broadcast NSID.
874  */
875 static const nvme_ioctl_check_t nvme_check_get_logpage = {
876 	.nck_ns_ok = B_TRUE, .nck_ns_minor_ok = B_TRUE,
877 	.nck_skip_ctrl = B_FALSE, .nck_ctrl_rewrite = B_TRUE,
878 	.nck_bcast_ok = B_TRUE, .nck_excl = NVME_IOCTL_EXCL_NONE
879 };
880 
881 /*
882  * When getting a feature, we do not want rewriting behavior as most features do
883  * not require a namespace to be specified. Specific instances are checked in
884  * nvme_validate_get_feature().
885  */
886 static const nvme_ioctl_check_t nvme_check_get_feature = {
887 	.nck_ns_ok = B_TRUE, .nck_ns_minor_ok = B_TRUE,
888 	.nck_skip_ctrl = B_FALSE, .nck_ctrl_rewrite = B_FALSE,
889 	.nck_bcast_ok = B_TRUE, .nck_excl = NVME_IOCTL_EXCL_NONE
890 };
891 
892 /*
893  * Format commands must target a namespace. The broadcast namespace must be used
894  * when referring to the controller.
895  */
896 static const nvme_ioctl_check_t nvme_check_format = {
897 	.nck_ns_ok = B_TRUE, .nck_ns_minor_ok = B_TRUE,
898 	.nck_skip_ctrl = B_FALSE, .nck_ctrl_rewrite = B_TRUE,
899 	.nck_bcast_ok = B_TRUE, .nck_excl = NVME_IOCTL_EXCL_WRITE
900 };
901 
902 /*
903  * Attach and detach must always target a minor. However, the broadcast
904  * namespace is not allowed. We still perform rewriting so that way specifying
905  * the controller node with 0 will be caught.
906  */
907 static const nvme_ioctl_check_t nvme_check_attach_detach = {
908 	.nck_ns_ok = B_TRUE, .nck_ns_minor_ok = B_TRUE,
909 	.nck_skip_ctrl = B_FALSE, .nck_ctrl_rewrite = B_TRUE,
910 	.nck_bcast_ok = B_FALSE, .nck_excl = NVME_IOCTL_EXCL_WRITE
911 };
912 
913 /*
914  * Firmware operations must not target a namespace and are only allowed from the
915  * controller.
916  */
917 static const nvme_ioctl_check_t nvme_check_firmware = {
918 	.nck_ns_ok = B_FALSE, .nck_ns_minor_ok = B_FALSE,
919 	.nck_skip_ctrl = B_FALSE, .nck_ctrl_rewrite = B_FALSE,
920 	.nck_bcast_ok = B_FALSE, .nck_excl = NVME_IOCTL_EXCL_WRITE
921 };
922 
923 /*
924  * Passthru commands are an odd set. We only allow them from the primary
925  * controller; however, we allow a namespace to be specified in them and allow
926  * the broadcast namespace. We do not perform rewriting because we don't know
927  * what the semantics are. We explicitly exempt passthru commands from needing
928  * an exclusive lock and leave it up to them to tell us the impact of the
929  * command and semantics. As this is a privileged interface and the semantics
930  * are arbitrary, there's not much we can do without some assistance from the
931  * consumer.
932  */
933 static const nvme_ioctl_check_t nvme_check_passthru = {
934 	.nck_ns_ok = B_TRUE, .nck_ns_minor_ok = B_FALSE,
935 	.nck_skip_ctrl = B_FALSE, .nck_ctrl_rewrite = B_FALSE,
936 	.nck_bcast_ok = B_TRUE, .nck_excl = NVME_IOCTL_EXCL_NONE
937 };
938 
939 /*
940  * Lock operations are allowed to target a namespace, but must not be rewritten.
941  * There is no support for the broadcast namespace. This is the only ioctl that
942  * should skip exclusive checking as it's used to grant it.
943  */
944 static const nvme_ioctl_check_t nvme_check_locking = {
945 	.nck_ns_ok = B_TRUE, .nck_ns_minor_ok = B_TRUE,
946 	.nck_skip_ctrl = B_FALSE, .nck_ctrl_rewrite = B_FALSE,
947 	.nck_bcast_ok = B_FALSE, .nck_excl = NVME_IOCTL_EXCL_SKIP
948 };
949 
950 static struct cb_ops nvme_cb_ops = {
951 	.cb_open	= nvme_open,
952 	.cb_close	= nvme_close,
953 	.cb_strategy	= nodev,
954 	.cb_print	= nodev,
955 	.cb_dump	= nodev,
956 	.cb_read	= nodev,
957 	.cb_write	= nodev,
958 	.cb_ioctl	= nvme_ioctl,
959 	.cb_devmap	= nodev,
960 	.cb_mmap	= nodev,
961 	.cb_segmap	= nodev,
962 	.cb_chpoll	= nochpoll,
963 	.cb_prop_op	= ddi_prop_op,
964 	.cb_str		= 0,
965 	.cb_flag	= D_NEW | D_MP,
966 	.cb_rev		= CB_REV,
967 	.cb_aread	= nodev,
968 	.cb_awrite	= nodev
969 };
970 
971 static struct dev_ops nvme_dev_ops = {
972 	.devo_rev	= DEVO_REV,
973 	.devo_refcnt	= 0,
974 	.devo_getinfo	= ddi_no_info,
975 	.devo_identify	= nulldev,
976 	.devo_probe	= nulldev,
977 	.devo_attach	= nvme_attach,
978 	.devo_detach	= nvme_detach,
979 	.devo_reset	= nodev,
980 	.devo_cb_ops	= &nvme_cb_ops,
981 	.devo_bus_ops	= NULL,
982 	.devo_power	= NULL,
983 	.devo_quiesce	= nvme_quiesce,
984 };
985 
986 static struct modldrv nvme_modldrv = {
987 	.drv_modops	= &mod_driverops,
988 	.drv_linkinfo	= "NVMe driver",
989 	.drv_dev_ops	= &nvme_dev_ops
990 };
991 
992 static struct modlinkage nvme_modlinkage = {
993 	.ml_rev		= MODREV_1,
994 	.ml_linkage	= { &nvme_modldrv, NULL }
995 };
996 
997 static bd_ops_t nvme_bd_ops = {
998 	.o_version	= BD_OPS_CURRENT_VERSION,
999 	.o_drive_info	= nvme_bd_driveinfo,
1000 	.o_media_info	= nvme_bd_mediainfo,
1001 	.o_devid_init	= nvme_bd_devid,
1002 	.o_sync_cache	= nvme_bd_sync,
1003 	.o_read		= nvme_bd_read,
1004 	.o_write	= nvme_bd_write,
1005 	.o_free_space	= nvme_bd_free_space,
1006 };
1007 
1008 /*
1009  * This list will hold commands that have timed out and couldn't be aborted.
1010  * As we don't know what the hardware may still do with the DMA memory we can't
1011  * free them, so we'll keep them forever on this list where we can easily look
1012  * at them with mdb.
1013  */
1014 static struct list nvme_lost_cmds;
1015 static kmutex_t nvme_lc_mutex;
1016 
1017 int
1018 _init(void)
1019 {
1020 	int error;
1021 
1022 	error = ddi_soft_state_init(&nvme_state, sizeof (nvme_t), 1);
1023 	if (error != DDI_SUCCESS)
1024 		return (error);
1025 
1026 	if ((nvme_open_minors = id_space_create("nvme_open_minors",
1027 	    NVME_OPEN_MINOR_MIN, NVME_OPEN_MINOR_MAX_EXCL)) == NULL) {
1028 		ddi_soft_state_fini(&nvme_state);
1029 		return (ENOMEM);
1030 	}
1031 
1032 	nvme_cmd_cache = kmem_cache_create("nvme_cmd_cache",
1033 	    sizeof (nvme_cmd_t), 64, NULL, NULL, NULL, NULL, NULL, 0);
1034 
1035 	mutex_init(&nvme_lc_mutex, NULL, MUTEX_DRIVER, NULL);
1036 	list_create(&nvme_lost_cmds, sizeof (nvme_cmd_t),
1037 	    offsetof(nvme_cmd_t, nc_list));
1038 
1039 	mutex_init(&nvme_open_minors_mutex, NULL, MUTEX_DRIVER, NULL);
1040 	avl_create(&nvme_open_minors_avl, nvme_minor_comparator,
1041 	    sizeof (nvme_minor_t), offsetof(nvme_minor_t, nm_avl));
1042 
1043 	nvme_dead_taskq = taskq_create("nvme_dead_taskq", 1, minclsyspri, 1, 1,
1044 	    TASKQ_PREPOPULATE);
1045 
1046 	bd_mod_init(&nvme_dev_ops);
1047 
1048 	error = mod_install(&nvme_modlinkage);
1049 	if (error != DDI_SUCCESS) {
1050 		ddi_soft_state_fini(&nvme_state);
1051 		id_space_destroy(nvme_open_minors);
1052 		mutex_destroy(&nvme_lc_mutex);
1053 		list_destroy(&nvme_lost_cmds);
1054 		bd_mod_fini(&nvme_dev_ops);
1055 		mutex_destroy(&nvme_open_minors_mutex);
1056 		avl_destroy(&nvme_open_minors_avl);
1057 		taskq_destroy(nvme_dead_taskq);
1058 	}
1059 
1060 	return (error);
1061 }
1062 
1063 int
1064 _fini(void)
1065 {
1066 	int error;
1067 
1068 	if (!list_is_empty(&nvme_lost_cmds))
1069 		return (DDI_FAILURE);
1070 
1071 	error = mod_remove(&nvme_modlinkage);
1072 	if (error == DDI_SUCCESS) {
1073 		ddi_soft_state_fini(&nvme_state);
1074 		id_space_destroy(nvme_open_minors);
1075 		kmem_cache_destroy(nvme_cmd_cache);
1076 		mutex_destroy(&nvme_lc_mutex);
1077 		list_destroy(&nvme_lost_cmds);
1078 		bd_mod_fini(&nvme_dev_ops);
1079 		mutex_destroy(&nvme_open_minors_mutex);
1080 		avl_destroy(&nvme_open_minors_avl);
1081 		taskq_destroy(nvme_dead_taskq);
1082 	}
1083 
1084 	return (error);
1085 }
1086 
1087 int
1088 _info(struct modinfo *modinfop)
1089 {
1090 	return (mod_info(&nvme_modlinkage, modinfop));
1091 }
1092 
1093 static inline void
1094 nvme_put64(nvme_t *nvme, uintptr_t reg, uint64_t val)
1095 {
1096 	ASSERT(((uintptr_t)(nvme->n_regs + reg) & 0x7) == 0);
1097 
1098 	/*LINTED: E_BAD_PTR_CAST_ALIGN*/
1099 	ddi_put64(nvme->n_regh, (uint64_t *)(nvme->n_regs + reg), val);
1100 }
1101 
1102 static inline void
1103 nvme_put32(nvme_t *nvme, uintptr_t reg, uint32_t val)
1104 {
1105 	ASSERT(((uintptr_t)(nvme->n_regs + reg) & 0x3) == 0);
1106 
1107 	/*LINTED: E_BAD_PTR_CAST_ALIGN*/
1108 	ddi_put32(nvme->n_regh, (uint32_t *)(nvme->n_regs + reg), val);
1109 }
1110 
1111 static inline uint64_t
1112 nvme_get64(nvme_t *nvme, uintptr_t reg)
1113 {
1114 	uint64_t val;
1115 
1116 	ASSERT(((uintptr_t)(nvme->n_regs + reg) & 0x7) == 0);
1117 
1118 	/*LINTED: E_BAD_PTR_CAST_ALIGN*/
1119 	val = ddi_get64(nvme->n_regh, (uint64_t *)(nvme->n_regs + reg));
1120 
1121 	return (val);
1122 }
1123 
1124 static inline uint32_t
1125 nvme_get32(nvme_t *nvme, uintptr_t reg)
1126 {
1127 	uint32_t val;
1128 
1129 	ASSERT(((uintptr_t)(nvme->n_regs + reg) & 0x3) == 0);
1130 
1131 	/*LINTED: E_BAD_PTR_CAST_ALIGN*/
1132 	val = ddi_get32(nvme->n_regh, (uint32_t *)(nvme->n_regs + reg));
1133 
1134 	return (val);
1135 }
1136 
1137 static void
1138 nvme_mgmt_lock_fini(nvme_mgmt_lock_t *lock)
1139 {
1140 	ASSERT3U(lock->nml_bd_own, ==, 0);
1141 	mutex_destroy(&lock->nml_lock);
1142 	cv_destroy(&lock->nml_cv);
1143 }
1144 
1145 static void
1146 nvme_mgmt_lock_init(nvme_mgmt_lock_t *lock)
1147 {
1148 	mutex_init(&lock->nml_lock, NULL, MUTEX_DRIVER, NULL);
1149 	cv_init(&lock->nml_cv, NULL, CV_DRIVER, NULL);
1150 	lock->nml_bd_own = 0;
1151 }
1152 
1153 static void
1154 nvme_mgmt_unlock(nvme_t *nvme)
1155 {
1156 	nvme_mgmt_lock_t *lock = &nvme->n_mgmt;
1157 
1158 	cv_broadcast(&lock->nml_cv);
1159 	mutex_exit(&lock->nml_lock);
1160 }
1161 
1162 #ifdef	DEBUG
1163 static boolean_t
1164 nvme_mgmt_lock_held(nvme_t *nvme)
1165 {
1166 	return (MUTEX_HELD(&nvme->n_mgmt.nml_lock) != 0);
1167 }
1168 #endif	/* DEBUG */
1169 
1170 static void
1171 nvme_mgmt_lock(nvme_t *nvme, nvme_mgmt_lock_level_t level)
1172 {
1173 	nvme_mgmt_lock_t *lock = &nvme->n_mgmt;
1174 	mutex_enter(&lock->nml_lock);
1175 	while (lock->nml_bd_own != 0) {
1176 		if (level == NVME_MGMT_LOCK_BDRO)
1177 			break;
1178 		cv_wait(&lock->nml_cv, &lock->nml_lock);
1179 	}
1180 }
1181 
1182 /*
1183  * This and nvme_mgmt_bd_end() are used to indicate that the driver is going to
1184  * be calling into a re-entrant blkdev related function. We cannot hold the lock
1185  * across such an operation and therefore must indicate that this is logically
1186  * held, while allowing other operations to proceed. This nvme_mgmt_bd_end() may
1187  * only be called by a thread that already holds the nmve_mgmt_lock().
1188  */
1189 static void
1190 nvme_mgmt_bd_start(nvme_t *nvme)
1191 {
1192 	nvme_mgmt_lock_t *lock = &nvme->n_mgmt;
1193 
1194 	VERIFY(MUTEX_HELD(&lock->nml_lock));
1195 	VERIFY3U(lock->nml_bd_own, ==, 0);
1196 	lock->nml_bd_own = (uintptr_t)curthread;
1197 	mutex_exit(&lock->nml_lock);
1198 }
1199 
1200 static void
1201 nvme_mgmt_bd_end(nvme_t *nvme)
1202 {
1203 	nvme_mgmt_lock_t *lock = &nvme->n_mgmt;
1204 
1205 	mutex_enter(&lock->nml_lock);
1206 	VERIFY3U(lock->nml_bd_own, ==, (uintptr_t)curthread);
1207 	lock->nml_bd_own = 0;
1208 }
1209 
1210 /*
1211  * This is a central clearing house for marking an NVMe controller dead and/or
1212  * removed. This takes care of setting the flag, taking care of outstanding
1213  * blocked locks, and sending a DDI FMA impact. This is called from a precarious
1214  * place where locking is suspect. The only guarantee we have is that the nvme_t
1215  * is valid and won't disappear until we return.
1216  *
1217  * This should only be used after attach has been called.
1218  */
1219 static void
1220 nvme_ctrl_mark_dead(nvme_t *nvme, boolean_t removed)
1221 {
1222 	boolean_t was_dead;
1223 
1224 	/*
1225 	 * See if we win the race to set things up here. If someone beat us to
1226 	 * it, we do not do anything.
1227 	 */
1228 	was_dead = atomic_cas_32((volatile uint32_t *)&nvme->n_dead, B_FALSE,
1229 	    B_TRUE);
1230 	if (was_dead) {
1231 		return;
1232 	}
1233 
1234 	/*
1235 	 * If this was removed, there is no reason to change the service impact.
1236 	 * However, then we need to change our default return code that we use
1237 	 * here to indicate that it was gone versus that it is dead.
1238 	 */
1239 	if (removed) {
1240 		nvme->n_dead_status = NVME_IOCTL_E_CTRL_GONE;
1241 	} else {
1242 		ASSERT3U(nvme->n_dead_status, ==, NVME_IOCTL_E_CTRL_DEAD);
1243 		ddi_fm_service_impact(nvme->n_dip, DDI_SERVICE_LOST);
1244 	}
1245 
1246 	taskq_dispatch_ent(nvme_dead_taskq, nvme_rwlock_ctrl_dead, nvme,
1247 	    TQ_NOSLEEP, &nvme->n_dead_tqent);
1248 }
1249 
1250 static boolean_t
1251 nvme_check_regs_hdl(nvme_t *nvme)
1252 {
1253 	ddi_fm_error_t error;
1254 
1255 	ddi_fm_acc_err_get(nvme->n_regh, &error, DDI_FME_VERSION);
1256 
1257 	if (error.fme_status != DDI_FM_OK)
1258 		return (B_TRUE);
1259 
1260 	return (B_FALSE);
1261 }
1262 
1263 static boolean_t
1264 nvme_check_dma_hdl(nvme_dma_t *dma)
1265 {
1266 	ddi_fm_error_t error;
1267 
1268 	if (dma == NULL)
1269 		return (B_FALSE);
1270 
1271 	ddi_fm_dma_err_get(dma->nd_dmah, &error, DDI_FME_VERSION);
1272 
1273 	if (error.fme_status != DDI_FM_OK)
1274 		return (B_TRUE);
1275 
1276 	return (B_FALSE);
1277 }
1278 
1279 static void
1280 nvme_free_dma_common(nvme_dma_t *dma)
1281 {
1282 	if (dma->nd_dmah != NULL)
1283 		(void) ddi_dma_unbind_handle(dma->nd_dmah);
1284 	if (dma->nd_acch != NULL)
1285 		ddi_dma_mem_free(&dma->nd_acch);
1286 	if (dma->nd_dmah != NULL)
1287 		ddi_dma_free_handle(&dma->nd_dmah);
1288 }
1289 
1290 static void
1291 nvme_free_dma(nvme_dma_t *dma)
1292 {
1293 	nvme_free_dma_common(dma);
1294 	kmem_free(dma, sizeof (*dma));
1295 }
1296 
1297 static void
1298 nvme_prp_dma_destructor(void *buf, void *private __unused)
1299 {
1300 	nvme_dma_t *dma = (nvme_dma_t *)buf;
1301 
1302 	nvme_free_dma_common(dma);
1303 }
1304 
1305 static int
1306 nvme_alloc_dma_common(nvme_t *nvme, nvme_dma_t *dma,
1307     size_t len, uint_t flags, ddi_dma_attr_t *dma_attr)
1308 {
1309 	if (ddi_dma_alloc_handle(nvme->n_dip, dma_attr, DDI_DMA_SLEEP, NULL,
1310 	    &dma->nd_dmah) != DDI_SUCCESS) {
1311 		/*
1312 		 * Due to DDI_DMA_SLEEP this can't be DDI_DMA_NORESOURCES, and
1313 		 * the only other possible error is DDI_DMA_BADATTR which
1314 		 * indicates a driver bug which should cause a panic.
1315 		 */
1316 		dev_err(nvme->n_dip, CE_PANIC,
1317 		    "!failed to get DMA handle, check DMA attributes");
1318 		return (DDI_FAILURE);
1319 	}
1320 
1321 	/*
1322 	 * ddi_dma_mem_alloc() can only fail when DDI_DMA_NOSLEEP is specified
1323 	 * or the flags are conflicting, which isn't the case here.
1324 	 */
1325 	(void) ddi_dma_mem_alloc(dma->nd_dmah, len, &nvme->n_reg_acc_attr,
1326 	    DDI_DMA_CONSISTENT, DDI_DMA_SLEEP, NULL, &dma->nd_memp,
1327 	    &dma->nd_len, &dma->nd_acch);
1328 
1329 	if (ddi_dma_addr_bind_handle(dma->nd_dmah, NULL, dma->nd_memp,
1330 	    dma->nd_len, flags | DDI_DMA_CONSISTENT, DDI_DMA_SLEEP, NULL,
1331 	    &dma->nd_cookie, &dma->nd_ncookie) != DDI_DMA_MAPPED) {
1332 		dev_err(nvme->n_dip, CE_WARN,
1333 		    "!failed to bind DMA memory");
1334 		NVME_BUMP_STAT(nvme, dma_bind_err);
1335 		nvme_free_dma_common(dma);
1336 		return (DDI_FAILURE);
1337 	}
1338 
1339 	return (DDI_SUCCESS);
1340 }
1341 
1342 static int
1343 nvme_zalloc_dma(nvme_t *nvme, size_t len, uint_t flags,
1344     ddi_dma_attr_t *dma_attr, nvme_dma_t **ret)
1345 {
1346 	nvme_dma_t *dma = kmem_zalloc(sizeof (nvme_dma_t), KM_SLEEP);
1347 
1348 	if (nvme_alloc_dma_common(nvme, dma, len, flags, dma_attr) !=
1349 	    DDI_SUCCESS) {
1350 		*ret = NULL;
1351 		kmem_free(dma, sizeof (nvme_dma_t));
1352 		return (DDI_FAILURE);
1353 	}
1354 
1355 	bzero(dma->nd_memp, dma->nd_len);
1356 
1357 	*ret = dma;
1358 	return (DDI_SUCCESS);
1359 }
1360 
1361 static int
1362 nvme_prp_dma_constructor(void *buf, void *private, int flags __unused)
1363 {
1364 	nvme_dma_t *dma = (nvme_dma_t *)buf;
1365 	nvme_t *nvme = (nvme_t *)private;
1366 
1367 	dma->nd_dmah = NULL;
1368 	dma->nd_acch = NULL;
1369 
1370 	if (nvme_alloc_dma_common(nvme, dma, nvme->n_pagesize,
1371 	    DDI_DMA_READ, &nvme->n_prp_dma_attr) != DDI_SUCCESS) {
1372 		return (-1);
1373 	}
1374 
1375 	ASSERT(dma->nd_ncookie == 1);
1376 
1377 	dma->nd_cached = B_TRUE;
1378 
1379 	return (0);
1380 }
1381 
1382 static int
1383 nvme_zalloc_queue_dma(nvme_t *nvme, uint32_t nentry, uint16_t qe_len,
1384     uint_t flags, nvme_dma_t **dma)
1385 {
1386 	uint32_t len = nentry * qe_len;
1387 	ddi_dma_attr_t q_dma_attr = nvme->n_queue_dma_attr;
1388 
1389 	len = roundup(len, nvme->n_pagesize);
1390 
1391 	if (nvme_zalloc_dma(nvme, len, flags, &q_dma_attr, dma)
1392 	    != DDI_SUCCESS) {
1393 		dev_err(nvme->n_dip, CE_WARN,
1394 		    "!failed to get DMA memory for queue");
1395 		goto fail;
1396 	}
1397 
1398 	if ((*dma)->nd_ncookie != 1) {
1399 		dev_err(nvme->n_dip, CE_WARN,
1400 		    "!got too many cookies for queue DMA");
1401 		goto fail;
1402 	}
1403 
1404 	return (DDI_SUCCESS);
1405 
1406 fail:
1407 	if (*dma) {
1408 		nvme_free_dma(*dma);
1409 		*dma = NULL;
1410 	}
1411 
1412 	return (DDI_FAILURE);
1413 }
1414 
1415 static void
1416 nvme_free_cq(nvme_cq_t *cq)
1417 {
1418 	mutex_destroy(&cq->ncq_mutex);
1419 
1420 	if (cq->ncq_cmd_taskq != NULL)
1421 		taskq_destroy(cq->ncq_cmd_taskq);
1422 
1423 	if (cq->ncq_dma != NULL)
1424 		nvme_free_dma(cq->ncq_dma);
1425 
1426 	kmem_free(cq, sizeof (*cq));
1427 }
1428 
1429 static void
1430 nvme_free_qpair(nvme_qpair_t *qp)
1431 {
1432 	int i;
1433 
1434 	mutex_destroy(&qp->nq_mutex);
1435 	sema_destroy(&qp->nq_sema);
1436 
1437 	if (qp->nq_sqdma != NULL)
1438 		nvme_free_dma(qp->nq_sqdma);
1439 
1440 	if (qp->nq_active_cmds > 0)
1441 		for (i = 0; i != qp->nq_nentry; i++)
1442 			if (qp->nq_cmd[i] != NULL)
1443 				nvme_free_cmd(qp->nq_cmd[i]);
1444 
1445 	if (qp->nq_cmd != NULL)
1446 		kmem_free(qp->nq_cmd, sizeof (nvme_cmd_t *) * qp->nq_nentry);
1447 
1448 	kmem_free(qp, sizeof (nvme_qpair_t));
1449 }
1450 
1451 /*
1452  * Destroy the pre-allocated cq array, but only free individual completion
1453  * queues from the given starting index.
1454  */
1455 static void
1456 nvme_destroy_cq_array(nvme_t *nvme, uint_t start)
1457 {
1458 	uint_t i;
1459 
1460 	for (i = start; i < nvme->n_cq_count; i++)
1461 		if (nvme->n_cq[i] != NULL)
1462 			nvme_free_cq(nvme->n_cq[i]);
1463 
1464 	kmem_free(nvme->n_cq, sizeof (*nvme->n_cq) * nvme->n_cq_count);
1465 }
1466 
1467 static int
1468 nvme_alloc_cq(nvme_t *nvme, uint32_t nentry, nvme_cq_t **cqp, uint16_t idx,
1469     uint_t nthr)
1470 {
1471 	nvme_cq_t *cq = kmem_zalloc(sizeof (*cq), KM_SLEEP);
1472 	char name[64];		/* large enough for the taskq name */
1473 
1474 	mutex_init(&cq->ncq_mutex, NULL, MUTEX_DRIVER,
1475 	    DDI_INTR_PRI(nvme->n_intr_pri));
1476 
1477 	if (nvme_zalloc_queue_dma(nvme, nentry, sizeof (nvme_cqe_t),
1478 	    DDI_DMA_READ, &cq->ncq_dma) != DDI_SUCCESS)
1479 		goto fail;
1480 
1481 	cq->ncq_cq = (nvme_cqe_t *)cq->ncq_dma->nd_memp;
1482 	cq->ncq_nentry = nentry;
1483 	cq->ncq_id = idx;
1484 	cq->ncq_hdbl = NVME_REG_CQHDBL(nvme, idx);
1485 
1486 	/*
1487 	 * Each completion queue has its own command taskq.
1488 	 */
1489 	(void) snprintf(name, sizeof (name), "%s%d_cmd_taskq%u",
1490 	    ddi_driver_name(nvme->n_dip), ddi_get_instance(nvme->n_dip), idx);
1491 
1492 	cq->ncq_cmd_taskq = taskq_create(name, nthr, minclsyspri, 64, INT_MAX,
1493 	    TASKQ_PREPOPULATE);
1494 
1495 	if (cq->ncq_cmd_taskq == NULL) {
1496 		dev_err(nvme->n_dip, CE_WARN, "!failed to create cmd "
1497 		    "taskq for cq %u", idx);
1498 		goto fail;
1499 	}
1500 
1501 	*cqp = cq;
1502 	return (DDI_SUCCESS);
1503 
1504 fail:
1505 	nvme_free_cq(cq);
1506 	*cqp = NULL;
1507 
1508 	return (DDI_FAILURE);
1509 }
1510 
1511 /*
1512  * Create the n_cq array big enough to hold "ncq" completion queues.
1513  * If the array already exists it will be re-sized (but only larger).
1514  * The admin queue is included in this array, which boosts the
1515  * max number of entries to UINT16_MAX + 1.
1516  */
1517 static int
1518 nvme_create_cq_array(nvme_t *nvme, uint_t ncq, uint32_t nentry, uint_t nthr)
1519 {
1520 	nvme_cq_t **cq;
1521 	uint_t i, cq_count;
1522 
1523 	ASSERT3U(ncq, >, nvme->n_cq_count);
1524 
1525 	cq = nvme->n_cq;
1526 	cq_count = nvme->n_cq_count;
1527 
1528 	nvme->n_cq = kmem_zalloc(sizeof (*nvme->n_cq) * ncq, KM_SLEEP);
1529 	nvme->n_cq_count = ncq;
1530 
1531 	for (i = 0; i < cq_count; i++)
1532 		nvme->n_cq[i] = cq[i];
1533 
1534 	for (; i < nvme->n_cq_count; i++)
1535 		if (nvme_alloc_cq(nvme, nentry, &nvme->n_cq[i], i, nthr) !=
1536 		    DDI_SUCCESS)
1537 			goto fail;
1538 
1539 	if (cq != NULL)
1540 		kmem_free(cq, sizeof (*cq) * cq_count);
1541 
1542 	return (DDI_SUCCESS);
1543 
1544 fail:
1545 	nvme_destroy_cq_array(nvme, cq_count);
1546 	/*
1547 	 * Restore the original array
1548 	 */
1549 	nvme->n_cq_count = cq_count;
1550 	nvme->n_cq = cq;
1551 
1552 	return (DDI_FAILURE);
1553 }
1554 
1555 static int
1556 nvme_alloc_qpair(nvme_t *nvme, uint32_t nentry, nvme_qpair_t **nqp,
1557     uint_t idx)
1558 {
1559 	nvme_qpair_t *qp = kmem_zalloc(sizeof (*qp), KM_SLEEP);
1560 	uint_t cq_idx;
1561 
1562 	mutex_init(&qp->nq_mutex, NULL, MUTEX_DRIVER,
1563 	    DDI_INTR_PRI(nvme->n_intr_pri));
1564 
1565 	/*
1566 	 * The NVMe spec defines that a full queue has one empty (unused) slot;
1567 	 * initialize the semaphore accordingly.
1568 	 */
1569 	sema_init(&qp->nq_sema, nentry - 1, NULL, SEMA_DRIVER, NULL);
1570 
1571 	if (nvme_zalloc_queue_dma(nvme, nentry, sizeof (nvme_sqe_t),
1572 	    DDI_DMA_WRITE, &qp->nq_sqdma) != DDI_SUCCESS)
1573 		goto fail;
1574 
1575 	/*
1576 	 * idx == 0 is adminq, those above 0 are shared io completion queues.
1577 	 */
1578 	cq_idx = idx == 0 ? 0 : 1 + (idx - 1) % (nvme->n_cq_count - 1);
1579 	qp->nq_cq = nvme->n_cq[cq_idx];
1580 	qp->nq_sq = (nvme_sqe_t *)qp->nq_sqdma->nd_memp;
1581 	qp->nq_nentry = nentry;
1582 
1583 	qp->nq_sqtdbl = NVME_REG_SQTDBL(nvme, idx);
1584 
1585 	qp->nq_cmd = kmem_zalloc(sizeof (nvme_cmd_t *) * nentry, KM_SLEEP);
1586 	qp->nq_next_cmd = 0;
1587 
1588 	*nqp = qp;
1589 	return (DDI_SUCCESS);
1590 
1591 fail:
1592 	nvme_free_qpair(qp);
1593 	*nqp = NULL;
1594 
1595 	return (DDI_FAILURE);
1596 }
1597 
1598 /*
1599  * One might reasonably consider that the nvme_cmd_cache should have a cache
1600  * constructor and destructor that takes care of the mutex/cv init/destroy, and
1601  * that nvme_free_cmd should reset more fields such that allocation becomes
1602  * simpler. This is not currently implemented as:
1603  * - nvme_cmd_cache is a global cache, shared across nvme instances and
1604  *   therefore there is no easy access to the corresponding nvme_t in the
1605  *   constructor to determine the required interrupt priority.
1606  * - Most fields in nvme_cmd_t would need to be zeroed in nvme_free_cmd while
1607  *   preserving the mutex/cv. It is easier to able to zero the entire
1608  *   structure and then init the mutex/cv only in the unlikely event that we
1609  *   want an admin command.
1610  */
1611 static nvme_cmd_t *
1612 nvme_alloc_cmd(nvme_t *nvme, int kmflag)
1613 {
1614 	nvme_cmd_t *cmd = kmem_cache_alloc(nvme_cmd_cache, kmflag);
1615 
1616 	if (cmd != NULL) {
1617 		bzero(cmd, sizeof (nvme_cmd_t));
1618 		cmd->nc_nvme = nvme;
1619 	}
1620 
1621 	return (cmd);
1622 }
1623 
1624 static nvme_cmd_t *
1625 nvme_alloc_admin_cmd(nvme_t *nvme, int kmflag)
1626 {
1627 	nvme_cmd_t *cmd = nvme_alloc_cmd(nvme, kmflag);
1628 
1629 	if (cmd != NULL) {
1630 		cmd->nc_flags |= NVME_CMD_F_USELOCK;
1631 		mutex_init(&cmd->nc_mutex, NULL, MUTEX_DRIVER,
1632 		    DDI_INTR_PRI(nvme->n_intr_pri));
1633 		cv_init(&cmd->nc_cv, NULL, CV_DRIVER, NULL);
1634 	}
1635 
1636 	return (cmd);
1637 }
1638 
1639 static void
1640 nvme_free_cmd(nvme_cmd_t *cmd)
1641 {
1642 	/* Don't free commands on the lost commands list. */
1643 	if (list_link_active(&cmd->nc_list))
1644 		return;
1645 
1646 	if (cmd->nc_dma) {
1647 		nvme_free_dma(cmd->nc_dma);
1648 		cmd->nc_dma = NULL;
1649 	}
1650 
1651 	if (cmd->nc_prp) {
1652 		kmem_cache_free(cmd->nc_nvme->n_prp_cache, cmd->nc_prp);
1653 		cmd->nc_prp = NULL;
1654 	}
1655 
1656 	if ((cmd->nc_flags & NVME_CMD_F_USELOCK) != 0) {
1657 		cv_destroy(&cmd->nc_cv);
1658 		mutex_destroy(&cmd->nc_mutex);
1659 	}
1660 
1661 	kmem_cache_free(nvme_cmd_cache, cmd);
1662 }
1663 
1664 static void
1665 nvme_submit_admin_cmd(nvme_qpair_t *qp, nvme_cmd_t *cmd, uint32_t *qtimeoutp)
1666 {
1667 	sema_p(&qp->nq_sema);
1668 	nvme_submit_cmd_common(qp, cmd, qtimeoutp);
1669 }
1670 
1671 static int
1672 nvme_submit_io_cmd(nvme_qpair_t *qp, nvme_cmd_t *cmd)
1673 {
1674 	if (cmd->nc_nvme->n_dead) {
1675 		return (EIO);
1676 	}
1677 
1678 	if (sema_tryp(&qp->nq_sema) == 0)
1679 		return (EAGAIN);
1680 
1681 	nvme_submit_cmd_common(qp, cmd, NULL);
1682 	return (0);
1683 }
1684 
1685 /*
1686  * Common command submission routine. If `qtimeoutp` is not NULL then it will
1687  * be set to the sum of the timeouts of any active commands ahead of the one
1688  * being submitted.
1689  */
1690 static void
1691 nvme_submit_cmd_common(nvme_qpair_t *qp, nvme_cmd_t *cmd, uint32_t *qtimeoutp)
1692 {
1693 	nvme_reg_sqtdbl_t tail = { 0 };
1694 
1695 	/*
1696 	 * We don't need to take a lock on cmd since it is not yet enqueued.
1697 	 */
1698 	cmd->nc_submit_ts = gethrtime();
1699 	cmd->nc_state = NVME_CMD_SUBMITTED;
1700 
1701 	mutex_enter(&qp->nq_mutex);
1702 
1703 	/*
1704 	 * Now that we hold the queue pair lock, we must check whether or not
1705 	 * the controller has been listed as dead (e.g. was removed due to
1706 	 * hotplug). This is necessary as otherwise we could race with
1707 	 * nvme_remove_callback(). Because this has not been enqueued, we don't
1708 	 * call nvme_unqueue_cmd(), which is why we must manually decrement the
1709 	 * semaphore.
1710 	 */
1711 	if (cmd->nc_nvme->n_dead) {
1712 		cmd->nc_queue_ts = gethrtime();
1713 		cmd->nc_state = NVME_CMD_QUEUED;
1714 		taskq_dispatch_ent(qp->nq_cq->ncq_cmd_taskq, cmd->nc_callback,
1715 		    cmd, TQ_NOSLEEP, &cmd->nc_tqent);
1716 		sema_v(&qp->nq_sema);
1717 		mutex_exit(&qp->nq_mutex);
1718 		return;
1719 	}
1720 
1721 	/*
1722 	 * Try to insert the cmd into the active cmd array at the nq_next_cmd
1723 	 * slot. If the slot is already occupied advance to the next slot and
1724 	 * try again. This can happen for long running commands like async event
1725 	 * requests.
1726 	 */
1727 	while (qp->nq_cmd[qp->nq_next_cmd] != NULL)
1728 		qp->nq_next_cmd = (qp->nq_next_cmd + 1) % qp->nq_nentry;
1729 	qp->nq_cmd[qp->nq_next_cmd] = cmd;
1730 
1731 	/*
1732 	 * We keep track of the number of active commands in this queue, and
1733 	 * the sum of the timeouts for those active commands.
1734 	 */
1735 	qp->nq_active_cmds++;
1736 	if (qtimeoutp != NULL)
1737 		*qtimeoutp = qp->nq_active_timeout;
1738 	qp->nq_active_timeout += cmd->nc_timeout;
1739 
1740 	cmd->nc_sqe.sqe_cid = qp->nq_next_cmd;
1741 	bcopy(&cmd->nc_sqe, &qp->nq_sq[qp->nq_sqtail], sizeof (nvme_sqe_t));
1742 	(void) ddi_dma_sync(qp->nq_sqdma->nd_dmah,
1743 	    sizeof (nvme_sqe_t) * qp->nq_sqtail,
1744 	    sizeof (nvme_sqe_t), DDI_DMA_SYNC_FORDEV);
1745 	qp->nq_next_cmd = (qp->nq_next_cmd + 1) % qp->nq_nentry;
1746 
1747 	tail.b.sqtdbl_sqt = qp->nq_sqtail = (qp->nq_sqtail + 1) % qp->nq_nentry;
1748 	nvme_put32(cmd->nc_nvme, qp->nq_sqtdbl, tail.r);
1749 
1750 	mutex_exit(&qp->nq_mutex);
1751 }
1752 
1753 static nvme_cmd_t *
1754 nvme_unqueue_cmd(nvme_t *nvme, nvme_qpair_t *qp, int cid)
1755 {
1756 	nvme_cmd_t *cmd;
1757 
1758 	ASSERT(mutex_owned(&qp->nq_mutex));
1759 	ASSERT3S(cid, <, qp->nq_nentry);
1760 
1761 	cmd = qp->nq_cmd[cid];
1762 	/*
1763 	 * Some controllers will erroneously add things to the completion queue
1764 	 * for which there is no matching outstanding command. If this happens,
1765 	 * it is almost certainly a controller firmware bug since nq_mutex
1766 	 * is held across command submission and ringing the queue doorbell,
1767 	 * and is also held in this function.
1768 	 *
1769 	 * If we see such an unexpected command, there is not much we can do.
1770 	 * These will be logged and counted in nvme_get_completed(), but
1771 	 * otherwise ignored.
1772 	 */
1773 	if (cmd == NULL)
1774 		return (NULL);
1775 	qp->nq_cmd[cid] = NULL;
1776 	ASSERT3U(qp->nq_active_cmds, >, 0);
1777 	qp->nq_active_cmds--;
1778 	ASSERT3U(qp->nq_active_timeout, >=, cmd->nc_timeout);
1779 	qp->nq_active_timeout -= cmd->nc_timeout;
1780 	sema_v(&qp->nq_sema);
1781 
1782 	ASSERT3P(cmd, !=, NULL);
1783 	ASSERT3P(cmd->nc_nvme, ==, nvme);
1784 	ASSERT3S(cmd->nc_sqe.sqe_cid, ==, cid);
1785 
1786 	return (cmd);
1787 }
1788 
1789 /*
1790  * This is called when an admin abort has failed to complete, once for the
1791  * original command and once for the abort itself. At this point the controller
1792  * has been marked dead. The commands are considered lost, de-queued if
1793  * possible, and placed on a global lost commands list so that they cannot be
1794  * freed and so that any DMA memory they have have is not re-used.
1795  */
1796 static void
1797 nvme_lost_cmd(nvme_t *nvme, nvme_cmd_t *cmd)
1798 {
1799 	ASSERT(mutex_owned(&cmd->nc_mutex));
1800 
1801 	switch (cmd->nc_state) {
1802 	case NVME_CMD_SUBMITTED: {
1803 		nvme_qpair_t *qp = nvme->n_ioq[cmd->nc_sqid];
1804 
1805 		/*
1806 		 * The command is still in the submitted state, meaning that we
1807 		 * have not processed a completion queue entry for it. De-queue
1808 		 * should be successful and if the hardware does later report
1809 		 * completion we'll skip it as a command for which we aren't
1810 		 * expecting a response (see nvme_unqueue_cmd()).
1811 		 */
1812 		mutex_enter(&qp->nq_mutex);
1813 		(void) nvme_unqueue_cmd(nvme, qp, cmd->nc_sqe.sqe_cid);
1814 		mutex_exit(&qp->nq_mutex);
1815 	}
1816 	case NVME_CMD_ALLOCATED:
1817 	case NVME_CMD_COMPLETED:
1818 		/*
1819 		 * If the command has not been submitted, or has completed,
1820 		 * there is nothing to do here. In the event of an abort
1821 		 * command timeout, we can end up here in the process of
1822 		 * "losing" the original command. It's possible that command
1823 		 * has actually completed (or been queued on the taskq) in the
1824 		 * interim.
1825 		 */
1826 		break;
1827 	case NVME_CMD_QUEUED:
1828 		/*
1829 		 * The command is on the taskq, awaiting callback. This should
1830 		 * be fairly rapid so wait for completion.
1831 		 */
1832 		while (cmd->nc_state != NVME_CMD_COMPLETED)
1833 			cv_wait(&cmd->nc_cv, &cmd->nc_mutex);
1834 		break;
1835 	case NVME_CMD_LOST:
1836 		dev_err(cmd->nc_nvme->n_dip, CE_PANIC,
1837 		    "%s: command %p already lost", __func__, (void *)cmd);
1838 		break;
1839 	}
1840 
1841 	cmd->nc_state = NVME_CMD_LOST;
1842 
1843 	mutex_enter(&nvme_lc_mutex);
1844 	list_insert_head(&nvme_lost_cmds, cmd);
1845 	mutex_exit(&nvme_lc_mutex);
1846 }
1847 
1848 /*
1849  * Get the command tied to the next completed cqe and bump along completion
1850  * queue head counter.
1851  */
1852 static nvme_cmd_t *
1853 nvme_get_completed(nvme_t *nvme, nvme_cq_t *cq)
1854 {
1855 	nvme_qpair_t *qp;
1856 	nvme_cqe_t *cqe;
1857 	nvme_cmd_t *cmd;
1858 
1859 	ASSERT(mutex_owned(&cq->ncq_mutex));
1860 
1861 retry:
1862 	cqe = &cq->ncq_cq[cq->ncq_head];
1863 
1864 	/* Check phase tag of CQE. Hardware inverts it for new entries. */
1865 	if (cqe->cqe_sf.sf_p == cq->ncq_phase)
1866 		return (NULL);
1867 
1868 	qp = nvme->n_ioq[cqe->cqe_sqid];
1869 
1870 	mutex_enter(&qp->nq_mutex);
1871 	cmd = nvme_unqueue_cmd(nvme, qp, cqe->cqe_cid);
1872 	mutex_exit(&qp->nq_mutex);
1873 
1874 	qp->nq_sqhead = cqe->cqe_sqhd;
1875 	cq->ncq_head = (cq->ncq_head + 1) % cq->ncq_nentry;
1876 
1877 	/* Toggle phase on wrap-around. */
1878 	if (cq->ncq_head == 0)
1879 		cq->ncq_phase = cq->ncq_phase != 0 ? 0 : 1;
1880 
1881 	if (cmd == NULL) {
1882 		dev_err(nvme->n_dip, CE_WARN,
1883 		    "!received completion for unknown cid 0x%x", cqe->cqe_cid);
1884 		NVME_BUMP_STAT(nvme, unknown_cid);
1885 		/*
1886 		 * We want to ignore this unexpected completion entry as it
1887 		 * is most likely a result of a bug in the controller firmware.
1888 		 * However, if we return NULL, then callers will assume there
1889 		 * are no more pending commands for this wakeup. Retry to keep
1890 		 * enumerating commands until the phase tag indicates there are
1891 		 * no more and we are really done.
1892 		 */
1893 		goto retry;
1894 	}
1895 
1896 	ASSERT3U(cmd->nc_sqid, ==, cqe->cqe_sqid);
1897 	bcopy(cqe, &cmd->nc_cqe, sizeof (nvme_cqe_t));
1898 
1899 	return (cmd);
1900 }
1901 
1902 /*
1903  * Process all completed commands on the io completion queue.
1904  */
1905 static uint_t
1906 nvme_process_iocq(nvme_t *nvme, nvme_cq_t *cq)
1907 {
1908 	nvme_reg_cqhdbl_t head = { 0 };
1909 	nvme_cmd_t *cmd;
1910 	uint_t completed = 0;
1911 
1912 	if (ddi_dma_sync(cq->ncq_dma->nd_dmah, 0, 0, DDI_DMA_SYNC_FORKERNEL) !=
1913 	    DDI_SUCCESS)
1914 		dev_err(nvme->n_dip, CE_WARN, "!ddi_dma_sync() failed in %s",
1915 		    __func__);
1916 
1917 	mutex_enter(&cq->ncq_mutex);
1918 
1919 	while ((cmd = nvme_get_completed(nvme, cq)) != NULL) {
1920 		/*
1921 		 * NVME_CMD_F_USELOCK is applied to all commands which are
1922 		 * going to be waited for by another thread in nvme_wait_cmd
1923 		 * and indicates that the lock should be taken before modifying
1924 		 * protected fields, and that the mutex has been initialised.
1925 		 * Commands which do not require the mutex to be held have not
1926 		 * initialised it (to reduce overhead).
1927 		 */
1928 		if ((cmd->nc_flags & NVME_CMD_F_USELOCK) != 0) {
1929 			mutex_enter(&cmd->nc_mutex);
1930 			/*
1931 			 * The command could have been de-queued as lost while
1932 			 * we waited on the lock, in which case we drop it.
1933 			 */
1934 			if (cmd->nc_state == NVME_CMD_LOST) {
1935 				mutex_exit(&cmd->nc_mutex);
1936 				completed++;
1937 				continue;
1938 			}
1939 		}
1940 		cmd->nc_queue_ts = gethrtime();
1941 		cmd->nc_state = NVME_CMD_QUEUED;
1942 		if ((cmd->nc_flags & NVME_CMD_F_USELOCK) != 0)
1943 			mutex_exit(&cmd->nc_mutex);
1944 		taskq_dispatch_ent(cq->ncq_cmd_taskq, cmd->nc_callback, cmd,
1945 		    TQ_NOSLEEP, &cmd->nc_tqent);
1946 
1947 		completed++;
1948 	}
1949 
1950 	if (completed > 0) {
1951 		/*
1952 		 * Update the completion queue head doorbell.
1953 		 */
1954 		head.b.cqhdbl_cqh = cq->ncq_head;
1955 		nvme_put32(nvme, cq->ncq_hdbl, head.r);
1956 	}
1957 
1958 	mutex_exit(&cq->ncq_mutex);
1959 
1960 	return (completed);
1961 }
1962 
1963 static nvme_cmd_t *
1964 nvme_retrieve_cmd(nvme_t *nvme, nvme_qpair_t *qp)
1965 {
1966 	nvme_cq_t *cq = qp->nq_cq;
1967 	nvme_reg_cqhdbl_t head = { 0 };
1968 	nvme_cmd_t *cmd;
1969 
1970 	if (ddi_dma_sync(cq->ncq_dma->nd_dmah, 0, 0, DDI_DMA_SYNC_FORKERNEL) !=
1971 	    DDI_SUCCESS)
1972 		dev_err(nvme->n_dip, CE_WARN, "!ddi_dma_sync() failed in %s",
1973 		    __func__);
1974 
1975 	mutex_enter(&cq->ncq_mutex);
1976 
1977 	if ((cmd = nvme_get_completed(nvme, cq)) != NULL) {
1978 		head.b.cqhdbl_cqh = cq->ncq_head;
1979 		nvme_put32(nvme, cq->ncq_hdbl, head.r);
1980 	}
1981 
1982 	mutex_exit(&cq->ncq_mutex);
1983 
1984 	return (cmd);
1985 }
1986 
1987 static int
1988 nvme_check_unknown_cmd_status(nvme_cmd_t *cmd)
1989 {
1990 	nvme_cqe_t *cqe = &cmd->nc_cqe;
1991 
1992 	dev_err(cmd->nc_nvme->n_dip, CE_WARN,
1993 	    "!unknown command status received: opc = %x, sqid = %d, cid = %d, "
1994 	    "sc = %x, sct = %x, dnr = %d, m = %d", cmd->nc_sqe.sqe_opc,
1995 	    cqe->cqe_sqid, cqe->cqe_cid, cqe->cqe_sf.sf_sc, cqe->cqe_sf.sf_sct,
1996 	    cqe->cqe_sf.sf_dnr, cqe->cqe_sf.sf_m);
1997 
1998 	if (cmd->nc_xfer != NULL)
1999 		bd_error(cmd->nc_xfer, BD_ERR_ILLRQ);
2000 
2001 	if (cmd->nc_nvme->n_strict_version) {
2002 		nvme_ctrl_mark_dead(cmd->nc_nvme, B_FALSE);
2003 	}
2004 
2005 	return (EIO);
2006 }
2007 
2008 static int
2009 nvme_check_vendor_cmd_status(nvme_cmd_t *cmd)
2010 {
2011 	nvme_cqe_t *cqe = &cmd->nc_cqe;
2012 
2013 	dev_err(cmd->nc_nvme->n_dip, CE_WARN,
2014 	    "!unknown command status received: opc = %x, sqid = %d, cid = %d, "
2015 	    "sc = %x, sct = %x, dnr = %d, m = %d", cmd->nc_sqe.sqe_opc,
2016 	    cqe->cqe_sqid, cqe->cqe_cid, cqe->cqe_sf.sf_sc, cqe->cqe_sf.sf_sct,
2017 	    cqe->cqe_sf.sf_dnr, cqe->cqe_sf.sf_m);
2018 	if (!cmd->nc_nvme->n_ignore_unknown_vendor_status) {
2019 		nvme_ctrl_mark_dead(cmd->nc_nvme, B_FALSE);
2020 	}
2021 
2022 	return (EIO);
2023 }
2024 
2025 static int
2026 nvme_check_integrity_cmd_status(nvme_cmd_t *cmd)
2027 {
2028 	nvme_cqe_t *cqe = &cmd->nc_cqe;
2029 
2030 	switch (cqe->cqe_sf.sf_sc) {
2031 	case NVME_CQE_SC_INT_NVM_WRITE:
2032 		/* write fail */
2033 		/* TODO: post ereport */
2034 		if (cmd->nc_xfer != NULL)
2035 			bd_error(cmd->nc_xfer, BD_ERR_MEDIA);
2036 		return (EIO);
2037 
2038 	case NVME_CQE_SC_INT_NVM_READ:
2039 		/* read fail */
2040 		/* TODO: post ereport */
2041 		if (cmd->nc_xfer != NULL)
2042 			bd_error(cmd->nc_xfer, BD_ERR_MEDIA);
2043 		return (EIO);
2044 
2045 	default:
2046 		return (nvme_check_unknown_cmd_status(cmd));
2047 	}
2048 }
2049 
2050 static int
2051 nvme_check_generic_cmd_status(nvme_cmd_t *cmd)
2052 {
2053 	nvme_cqe_t *cqe = &cmd->nc_cqe;
2054 
2055 	switch (cqe->cqe_sf.sf_sc) {
2056 	case NVME_CQE_SC_GEN_SUCCESS:
2057 		return (0);
2058 
2059 	/*
2060 	 * Errors indicating a bug in the driver should cause a panic.
2061 	 */
2062 	case NVME_CQE_SC_GEN_INV_OPC:
2063 		/* Invalid Command Opcode */
2064 		NVME_BUMP_STAT(cmd->nc_nvme, inv_cmd_err);
2065 		if ((cmd->nc_flags & NVME_CMD_F_DONTPANIC) == 0) {
2066 			dev_err(cmd->nc_nvme->n_dip, CE_PANIC,
2067 			    "programming error: invalid opcode in cmd %p",
2068 			    (void *)cmd);
2069 		}
2070 		return (EINVAL);
2071 
2072 	case NVME_CQE_SC_GEN_INV_FLD:
2073 		/* Invalid Field in Command */
2074 		NVME_BUMP_STAT(cmd->nc_nvme, inv_field_err);
2075 		if ((cmd->nc_flags & NVME_CMD_F_DONTPANIC) == 0) {
2076 			dev_err(cmd->nc_nvme->n_dip, CE_PANIC,
2077 			    "programming error: invalid field in cmd %p",
2078 			    (void *)cmd);
2079 		}
2080 		return (EIO);
2081 
2082 	case NVME_CQE_SC_GEN_ID_CNFL:
2083 		/* Command ID Conflict */
2084 		dev_err(cmd->nc_nvme->n_dip, CE_PANIC, "programming error: "
2085 		    "cmd ID conflict in cmd %p", (void *)cmd);
2086 		return (0);
2087 
2088 	case NVME_CQE_SC_GEN_INV_NS:
2089 		/* Invalid Namespace or Format */
2090 		NVME_BUMP_STAT(cmd->nc_nvme, inv_nsfmt_err);
2091 		if ((cmd->nc_flags & NVME_CMD_F_DONTPANIC) == 0) {
2092 			dev_err(cmd->nc_nvme->n_dip, CE_PANIC,
2093 			    "programming error: invalid NS/format in cmd %p",
2094 			    (void *)cmd);
2095 		}
2096 		return (EINVAL);
2097 
2098 	case NVME_CQE_SC_GEN_NVM_LBA_RANGE:
2099 		/* LBA Out Of Range */
2100 		dev_err(cmd->nc_nvme->n_dip, CE_PANIC, "programming error: "
2101 		    "LBA out of range in cmd %p", (void *)cmd);
2102 		return (0);
2103 
2104 	/*
2105 	 * Non-fatal errors, handle gracefully.
2106 	 */
2107 	case NVME_CQE_SC_GEN_DATA_XFR_ERR:
2108 		/* Data Transfer Error (DMA) */
2109 		/* TODO: post ereport */
2110 		NVME_BUMP_STAT(cmd->nc_nvme, data_xfr_err);
2111 		if (cmd->nc_xfer != NULL)
2112 			bd_error(cmd->nc_xfer, BD_ERR_NTRDY);
2113 		return (EIO);
2114 
2115 	case NVME_CQE_SC_GEN_INTERNAL_ERR:
2116 		/*
2117 		 * Internal Error. The spec (v1.0, section 4.5.1.2) says
2118 		 * detailed error information is returned as async event,
2119 		 * so we pretty much ignore the error here and handle it
2120 		 * in the async event handler.
2121 		 */
2122 		NVME_BUMP_STAT(cmd->nc_nvme, internal_err);
2123 		if (cmd->nc_xfer != NULL)
2124 			bd_error(cmd->nc_xfer, BD_ERR_NTRDY);
2125 		return (EIO);
2126 
2127 	case NVME_CQE_SC_GEN_ABORT_REQUEST:
2128 		/*
2129 		 * Command Abort Requested. This normally happens only when a
2130 		 * command times out.
2131 		 */
2132 		/* TODO: post ereport or change blkdev to handle this? */
2133 		NVME_BUMP_STAT(cmd->nc_nvme, abort_rq_err);
2134 		return (ECANCELED);
2135 
2136 	case NVME_CQE_SC_GEN_ABORT_PWRLOSS:
2137 		/* Command Aborted due to Power Loss Notification */
2138 		NVME_BUMP_STAT(cmd->nc_nvme, abort_pwrloss_err);
2139 		nvme_ctrl_mark_dead(cmd->nc_nvme, B_FALSE);
2140 		return (EIO);
2141 
2142 	case NVME_CQE_SC_GEN_ABORT_SQ_DEL:
2143 		/* Command Aborted due to SQ Deletion */
2144 		NVME_BUMP_STAT(cmd->nc_nvme, abort_sq_del);
2145 		return (EIO);
2146 
2147 	case NVME_CQE_SC_GEN_NVM_CAP_EXC:
2148 		/* Capacity Exceeded */
2149 		NVME_BUMP_STAT(cmd->nc_nvme, nvm_cap_exc);
2150 		if (cmd->nc_xfer != NULL)
2151 			bd_error(cmd->nc_xfer, BD_ERR_MEDIA);
2152 		return (EIO);
2153 
2154 	case NVME_CQE_SC_GEN_NVM_NS_NOTRDY:
2155 		/* Namespace Not Ready */
2156 		NVME_BUMP_STAT(cmd->nc_nvme, nvm_ns_notrdy);
2157 		if (cmd->nc_xfer != NULL)
2158 			bd_error(cmd->nc_xfer, BD_ERR_NTRDY);
2159 		return (EIO);
2160 
2161 	case NVME_CQE_SC_GEN_NVM_FORMATTING:
2162 		/* Format in progress (1.2) */
2163 		if (!NVME_VERSION_ATLEAST(&cmd->nc_nvme->n_version, 1, 2))
2164 			return (nvme_check_unknown_cmd_status(cmd));
2165 		NVME_BUMP_STAT(cmd->nc_nvme, nvm_ns_formatting);
2166 		if (cmd->nc_xfer != NULL)
2167 			bd_error(cmd->nc_xfer, BD_ERR_NTRDY);
2168 		return (EIO);
2169 
2170 	default:
2171 		return (nvme_check_unknown_cmd_status(cmd));
2172 	}
2173 }
2174 
2175 static int
2176 nvme_check_specific_cmd_status(nvme_cmd_t *cmd)
2177 {
2178 	nvme_cqe_t *cqe = &cmd->nc_cqe;
2179 
2180 	switch (cqe->cqe_sf.sf_sc) {
2181 	case NVME_CQE_SC_SPC_INV_CQ:
2182 		/* Completion Queue Invalid */
2183 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_CREATE_SQUEUE);
2184 		NVME_BUMP_STAT(cmd->nc_nvme, inv_cq_err);
2185 		return (EINVAL);
2186 
2187 	case NVME_CQE_SC_SPC_INV_QID:
2188 		/* Invalid Queue Identifier */
2189 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_CREATE_SQUEUE ||
2190 		    cmd->nc_sqe.sqe_opc == NVME_OPC_DELETE_SQUEUE ||
2191 		    cmd->nc_sqe.sqe_opc == NVME_OPC_CREATE_CQUEUE ||
2192 		    cmd->nc_sqe.sqe_opc == NVME_OPC_DELETE_CQUEUE);
2193 		NVME_BUMP_STAT(cmd->nc_nvme, inv_qid_err);
2194 		return (EINVAL);
2195 
2196 	case NVME_CQE_SC_SPC_MAX_QSZ_EXC:
2197 		/* Max Queue Size Exceeded */
2198 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_CREATE_SQUEUE ||
2199 		    cmd->nc_sqe.sqe_opc == NVME_OPC_CREATE_CQUEUE);
2200 		NVME_BUMP_STAT(cmd->nc_nvme, max_qsz_exc);
2201 		return (EINVAL);
2202 
2203 	case NVME_CQE_SC_SPC_ABRT_CMD_EXC:
2204 		/* Abort Command Limit Exceeded */
2205 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_ABORT);
2206 		dev_err(cmd->nc_nvme->n_dip, CE_PANIC, "programming error: "
2207 		    "abort command limit exceeded in cmd %p", (void *)cmd);
2208 		return (0);
2209 
2210 	case NVME_CQE_SC_SPC_ASYNC_EVREQ_EXC:
2211 		/* Async Event Request Limit Exceeded */
2212 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_ASYNC_EVENT);
2213 		dev_err(cmd->nc_nvme->n_dip, CE_PANIC, "programming error: "
2214 		    "async event request limit exceeded in cmd %p",
2215 		    (void *)cmd);
2216 		return (0);
2217 
2218 	case NVME_CQE_SC_SPC_INV_INT_VECT:
2219 		/* Invalid Interrupt Vector */
2220 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_CREATE_CQUEUE);
2221 		NVME_BUMP_STAT(cmd->nc_nvme, inv_int_vect);
2222 		return (EINVAL);
2223 
2224 	case NVME_CQE_SC_SPC_INV_LOG_PAGE:
2225 		/* Invalid Log Page */
2226 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_GET_LOG_PAGE);
2227 		NVME_BUMP_STAT(cmd->nc_nvme, inv_log_page);
2228 		return (EINVAL);
2229 
2230 	case NVME_CQE_SC_SPC_INV_FORMAT:
2231 		/* Invalid Format */
2232 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_NVM_FORMAT);
2233 		NVME_BUMP_STAT(cmd->nc_nvme, inv_format);
2234 		if (cmd->nc_xfer != NULL)
2235 			bd_error(cmd->nc_xfer, BD_ERR_ILLRQ);
2236 		return (EINVAL);
2237 
2238 	case NVME_CQE_SC_SPC_INV_Q_DEL:
2239 		/* Invalid Queue Deletion */
2240 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_DELETE_CQUEUE);
2241 		NVME_BUMP_STAT(cmd->nc_nvme, inv_q_del);
2242 		return (EINVAL);
2243 
2244 	case NVME_CQE_SC_SPC_NVM_CNFL_ATTR:
2245 		/* Conflicting Attributes */
2246 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_NVM_DSET_MGMT ||
2247 		    cmd->nc_sqe.sqe_opc == NVME_OPC_NVM_READ ||
2248 		    cmd->nc_sqe.sqe_opc == NVME_OPC_NVM_WRITE);
2249 		NVME_BUMP_STAT(cmd->nc_nvme, cnfl_attr);
2250 		if (cmd->nc_xfer != NULL)
2251 			bd_error(cmd->nc_xfer, BD_ERR_ILLRQ);
2252 		return (EINVAL);
2253 
2254 	case NVME_CQE_SC_SPC_NVM_INV_PROT:
2255 		/* Invalid Protection Information */
2256 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_NVM_COMPARE ||
2257 		    cmd->nc_sqe.sqe_opc == NVME_OPC_NVM_READ ||
2258 		    cmd->nc_sqe.sqe_opc == NVME_OPC_NVM_WRITE);
2259 		NVME_BUMP_STAT(cmd->nc_nvme, inv_prot);
2260 		if (cmd->nc_xfer != NULL)
2261 			bd_error(cmd->nc_xfer, BD_ERR_ILLRQ);
2262 		return (EINVAL);
2263 
2264 	case NVME_CQE_SC_SPC_NVM_READONLY:
2265 		/* Write to Read Only Range */
2266 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_NVM_WRITE);
2267 		NVME_BUMP_STAT(cmd->nc_nvme, readonly);
2268 		if (cmd->nc_xfer != NULL)
2269 			bd_error(cmd->nc_xfer, BD_ERR_ILLRQ);
2270 		return (EROFS);
2271 
2272 	case NVME_CQE_SC_SPC_INV_FW_SLOT:
2273 		/* Invalid Firmware Slot */
2274 		NVME_BUMP_STAT(cmd->nc_nvme, inv_fwslot);
2275 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_FW_ACTIVATE);
2276 		return (EINVAL);
2277 
2278 	case NVME_CQE_SC_SPC_INV_FW_IMG:
2279 		/* Invalid Firmware Image */
2280 		NVME_BUMP_STAT(cmd->nc_nvme, inv_fwimg);
2281 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_FW_ACTIVATE);
2282 		return (EINVAL);
2283 
2284 	case NVME_CQE_SC_SPC_FW_RESET:
2285 		/* Conventional Reset Required */
2286 		NVME_BUMP_STAT(cmd->nc_nvme, fwact_creset);
2287 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_FW_ACTIVATE);
2288 		return (0);
2289 
2290 	case NVME_CQE_SC_SPC_FW_NSSR:
2291 		/* NVMe Subsystem Reset Required */
2292 		NVME_BUMP_STAT(cmd->nc_nvme, fwact_nssr);
2293 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_FW_ACTIVATE);
2294 		return (0);
2295 
2296 	case NVME_CQE_SC_SPC_FW_NEXT_RESET:
2297 		/* Activation Requires Reset */
2298 		NVME_BUMP_STAT(cmd->nc_nvme, fwact_reset);
2299 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_FW_ACTIVATE);
2300 		return (0);
2301 
2302 	case NVME_CQE_SC_SPC_FW_MTFA:
2303 		/* Activation Requires Maximum Time Violation */
2304 		NVME_BUMP_STAT(cmd->nc_nvme, fwact_mtfa);
2305 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_FW_ACTIVATE);
2306 		return (EAGAIN);
2307 
2308 	case NVME_CQE_SC_SPC_FW_PROHIBITED:
2309 		/* Activation Prohibited */
2310 		NVME_BUMP_STAT(cmd->nc_nvme, fwact_prohibited);
2311 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_FW_ACTIVATE);
2312 		return (EINVAL);
2313 
2314 	case NVME_CQE_SC_SPC_FW_OVERLAP:
2315 		/* Overlapping Firmware Ranges */
2316 		NVME_BUMP_STAT(cmd->nc_nvme, fw_overlap);
2317 		ASSERT(cmd->nc_sqe.sqe_opc == NVME_OPC_FW_IMAGE_LOAD ||
2318 		    cmd->nc_sqe.sqe_opc == NVME_OPC_FW_ACTIVATE);
2319 		return (EINVAL);
2320 
2321 	default:
2322 		return (nvme_check_unknown_cmd_status(cmd));
2323 	}
2324 }
2325 
2326 static inline int
2327 nvme_check_cmd_status(nvme_cmd_t *cmd)
2328 {
2329 	nvme_cqe_t *cqe = &cmd->nc_cqe;
2330 
2331 	/*
2332 	 * Take a shortcut if the controller is dead, or if
2333 	 * command status indicates no error.
2334 	 */
2335 	if (cmd->nc_nvme->n_dead)
2336 		return (EIO);
2337 
2338 	if (cqe->cqe_sf.sf_sct == NVME_CQE_SCT_GENERIC &&
2339 	    cqe->cqe_sf.sf_sc == NVME_CQE_SC_GEN_SUCCESS)
2340 		return (0);
2341 
2342 	if (cqe->cqe_sf.sf_sct == NVME_CQE_SCT_GENERIC)
2343 		return (nvme_check_generic_cmd_status(cmd));
2344 	else if (cqe->cqe_sf.sf_sct == NVME_CQE_SCT_SPECIFIC)
2345 		return (nvme_check_specific_cmd_status(cmd));
2346 	else if (cqe->cqe_sf.sf_sct == NVME_CQE_SCT_INTEGRITY)
2347 		return (nvme_check_integrity_cmd_status(cmd));
2348 	else if (cqe->cqe_sf.sf_sct == NVME_CQE_SCT_VENDOR)
2349 		return (nvme_check_vendor_cmd_status(cmd));
2350 
2351 	return (nvme_check_unknown_cmd_status(cmd));
2352 }
2353 
2354 /*
2355  * Check the command status as used by an ioctl path and do not convert it to an
2356  * errno. We still allow all the command status checking to occur, but otherwise
2357  * will pass back the controller error as is.
2358  */
2359 static boolean_t
2360 nvme_check_cmd_status_ioctl(nvme_cmd_t *cmd, nvme_ioctl_common_t *ioc)
2361 {
2362 	nvme_cqe_t *cqe = &cmd->nc_cqe;
2363 	nvme_t *nvme = cmd->nc_nvme;
2364 
2365 	if (nvme->n_dead) {
2366 		return (nvme_ioctl_error(ioc, nvme->n_dead_status, 0, 0));
2367 	}
2368 
2369 	if (cqe->cqe_sf.sf_sct == NVME_CQE_SCT_GENERIC &&
2370 	    cqe->cqe_sf.sf_sc == NVME_CQE_SC_GEN_SUCCESS)
2371 		return (B_TRUE);
2372 
2373 	if (cqe->cqe_sf.sf_sct == NVME_CQE_SCT_GENERIC) {
2374 		(void) nvme_check_generic_cmd_status(cmd);
2375 	} else if (cqe->cqe_sf.sf_sct == NVME_CQE_SCT_SPECIFIC) {
2376 		(void) nvme_check_specific_cmd_status(cmd);
2377 	} else if (cqe->cqe_sf.sf_sct == NVME_CQE_SCT_INTEGRITY) {
2378 		(void) nvme_check_integrity_cmd_status(cmd);
2379 	} else if (cqe->cqe_sf.sf_sct == NVME_CQE_SCT_VENDOR) {
2380 		(void) nvme_check_vendor_cmd_status(cmd);
2381 	} else {
2382 		(void) nvme_check_unknown_cmd_status(cmd);
2383 	}
2384 
2385 	return (nvme_ioctl_error(ioc, NVME_IOCTL_E_CTRL_ERROR,
2386 	    cqe->cqe_sf.sf_sct, cqe->cqe_sf.sf_sc));
2387 }
2388 
2389 static int
2390 nvme_abort_cmd(nvme_cmd_t *cmd, const uint32_t sec)
2391 {
2392 	nvme_t *nvme = cmd->nc_nvme;
2393 	nvme_cmd_t *abort_cmd = nvme_alloc_admin_cmd(nvme, KM_SLEEP);
2394 	nvme_abort_cmd_t ac = { 0 };
2395 	int ret = 0;
2396 
2397 	sema_p(&nvme->n_abort_sema);
2398 
2399 	ac.b.ac_cid = cmd->nc_sqe.sqe_cid;
2400 	ac.b.ac_sqid = cmd->nc_sqid;
2401 
2402 	abort_cmd->nc_sqid = 0;
2403 	abort_cmd->nc_sqe.sqe_opc = NVME_OPC_ABORT;
2404 	abort_cmd->nc_callback = nvme_wakeup_cmd;
2405 	abort_cmd->nc_sqe.sqe_cdw10 = ac.r;
2406 
2407 	/*
2408 	 * Send the ABORT to the hardware. The ABORT command will return _after_
2409 	 * the aborted command has completed (aborted or otherwise) so we must
2410 	 * drop the aborted command's lock to allow it to complete.
2411 	 * We want to allow at least `nvme_abort_cmd_timeout` seconds for the
2412 	 * abort to be processed, but more if we are aborting a long-running
2413 	 * command to give that time to complete/abort too.
2414 	 */
2415 	mutex_exit(&cmd->nc_mutex);
2416 	nvme_admin_cmd(abort_cmd, MAX(nvme_abort_cmd_timeout, sec));
2417 	mutex_enter(&cmd->nc_mutex);
2418 
2419 	sema_v(&nvme->n_abort_sema);
2420 
2421 	/*
2422 	 * If the abort command itself has timed out, it will have been
2423 	 * de-queued so that its callback will not be called after this point,
2424 	 * and its state will be NVME_CMD_LOST.
2425 	 *
2426 	 * nvme_admin_cmd(abort_cmd)
2427 	 *   -> nvme_wait_cmd(abort_cmd)
2428 	 *     -> nvme_cmd(abort_cmd)
2429 	 *     | -> nvme_admin_cmd(cmd)
2430 	 *     |   -> nvme_wait_cmd(cmd)
2431 	 *     |     -> nvme_ctrl_mark_dead()
2432 	 *     |     -> nvme_lost_cmd(cmd)
2433 	 *     |       -> cmd->nc_stat = NVME_CMD_LOST
2434 	 *     and here we are.
2435 	 */
2436 	if (abort_cmd->nc_state == NVME_CMD_LOST) {
2437 		dev_err(nvme->n_dip, CE_WARN,
2438 		    "!ABORT of command %d/%d timed out",
2439 		    cmd->nc_sqe.sqe_cid, cmd->nc_sqid);
2440 		NVME_BUMP_STAT(nvme, abort_timeout);
2441 		ret = EIO;
2442 	} else if ((ret = nvme_check_cmd_status(abort_cmd)) != 0) {
2443 		dev_err(nvme->n_dip, CE_WARN,
2444 		    "!ABORT of command %d/%d "
2445 		    "failed with sct = %x, sc = %x",
2446 		    cmd->nc_sqe.sqe_cid, cmd->nc_sqid,
2447 		    abort_cmd->nc_cqe.cqe_sf.sf_sct,
2448 		    abort_cmd->nc_cqe.cqe_sf.sf_sc);
2449 		NVME_BUMP_STAT(nvme, abort_failed);
2450 	} else {
2451 		boolean_t success = ((abort_cmd->nc_cqe.cqe_dw0 & 1) == 0);
2452 
2453 		dev_err(nvme->n_dip, CE_WARN,
2454 		    "!ABORT of command %d/%d %ssuccessful",
2455 		    cmd->nc_sqe.sqe_cid, cmd->nc_sqid,
2456 		    success ? "" : "un");
2457 
2458 		if (success) {
2459 			NVME_BUMP_STAT(nvme, abort_successful);
2460 		} else {
2461 			NVME_BUMP_STAT(nvme, abort_unsuccessful);
2462 		}
2463 	}
2464 
2465 	/*
2466 	 * This abort abort_cmd has either completed or been de-queued as
2467 	 * lost in nvme_wait_cmd. Either way it's safe to free it here.
2468 	 */
2469 	nvme_free_cmd(abort_cmd);
2470 
2471 	return (ret);
2472 }
2473 
2474 /*
2475  * nvme_wait_cmd -- wait for command completion or timeout
2476  *
2477  * In case of a serious error or a timeout of the abort command the hardware
2478  * will be declared dead and FMA will be notified.
2479  */
2480 static void
2481 nvme_wait_cmd(nvme_cmd_t *cmd, uint32_t sec)
2482 {
2483 	nvme_t *nvme = cmd->nc_nvme;
2484 	nvme_reg_csts_t csts;
2485 
2486 	ASSERT(mutex_owned(&cmd->nc_mutex));
2487 
2488 	while (cmd->nc_state != NVME_CMD_COMPLETED) {
2489 		clock_t timeout = ddi_get_lbolt() +
2490 		    drv_usectohz((long)sec * MICROSEC);
2491 
2492 		if (cv_timedwait(&cmd->nc_cv, &cmd->nc_mutex, timeout) == -1) {
2493 			/*
2494 			 * If this command is on the task queue then we don't
2495 			 * consider it to have timed out. We are waiting for
2496 			 * the callback to be invoked, the timing of which can
2497 			 * be affected by system load and should not count
2498 			 * against the device; continue to wait.
2499 			 * While this doesn't help deal with the possibility of
2500 			 * a command timing out between being placed on the CQ
2501 			 * and arriving on the taskq, we expect interrupts to
2502 			 * run fairly promptly making this a small window.
2503 			 */
2504 			if (cmd->nc_state != NVME_CMD_QUEUED)
2505 				break;
2506 		}
2507 	}
2508 
2509 	if (cmd->nc_state == NVME_CMD_COMPLETED) {
2510 		DTRACE_PROBE1(nvme_admin_cmd_completed, nvme_cmd_t *, cmd);
2511 		nvme_admin_stat_cmd(nvme, cmd);
2512 		return;
2513 	}
2514 
2515 	/*
2516 	 * The command timed out.
2517 	 */
2518 
2519 	DTRACE_PROBE1(nvme_admin_cmd_timeout, nvme_cmd_t *, cmd);
2520 	csts.r = nvme_get32(nvme, NVME_REG_CSTS);
2521 	dev_err(nvme->n_dip, CE_WARN, "!command %d/%d timeout, "
2522 	    "OPC = %x, CFS = %d", cmd->nc_sqe.sqe_cid, cmd->nc_sqid,
2523 	    cmd->nc_sqe.sqe_opc, csts.b.csts_cfs);
2524 	NVME_BUMP_STAT(nvme, cmd_timeout);
2525 
2526 	/*
2527 	 * Check controller for fatal status, any errors associated with the
2528 	 * register or DMA handle, or for a double timeout (abort command timed
2529 	 * out). If necessary log a warning and call FMA.
2530 	 */
2531 	if (csts.b.csts_cfs ||
2532 	    nvme_check_regs_hdl(nvme) ||
2533 	    nvme_check_dma_hdl(cmd->nc_dma) ||
2534 	    cmd->nc_sqe.sqe_opc == NVME_OPC_ABORT) {
2535 		nvme_ctrl_mark_dead(cmd->nc_nvme, B_FALSE);
2536 		nvme_lost_cmd(nvme, cmd);
2537 		return;
2538 	}
2539 
2540 	/* Issue an abort for the command that has timed out */
2541 	if (nvme_abort_cmd(cmd, sec) == 0) {
2542 		/*
2543 		 * If the abort completed, whether or not it was
2544 		 * successful in aborting the command, that command
2545 		 * will also have completed with an appropriate
2546 		 * status.
2547 		 */
2548 		while (cmd->nc_state != NVME_CMD_COMPLETED)
2549 			cv_wait(&cmd->nc_cv, &cmd->nc_mutex);
2550 		return;
2551 	}
2552 
2553 	/*
2554 	 * Otherwise, the abort has also timed out or failed, which
2555 	 * will have marked the controller dead. De-queue the original command
2556 	 * and add it to the lost commands list.
2557 	 */
2558 	VERIFY(cmd->nc_nvme->n_dead);
2559 	nvme_lost_cmd(nvme, cmd);
2560 }
2561 
2562 static void
2563 nvme_wakeup_cmd(void *arg)
2564 {
2565 	nvme_cmd_t *cmd = arg;
2566 
2567 	ASSERT(cmd->nc_flags & NVME_CMD_F_USELOCK);
2568 
2569 	mutex_enter(&cmd->nc_mutex);
2570 	cmd->nc_state = NVME_CMD_COMPLETED;
2571 	cv_signal(&cmd->nc_cv);
2572 	mutex_exit(&cmd->nc_mutex);
2573 }
2574 
2575 static void
2576 nvme_async_event_task(void *arg)
2577 {
2578 	nvme_cmd_t *cmd = arg;
2579 	nvme_t *nvme = cmd->nc_nvme;
2580 	nvme_error_log_entry_t *error_log = NULL;
2581 	nvme_health_log_t *health_log = NULL;
2582 	nvme_nschange_list_t *nslist = NULL;
2583 	size_t logsize = 0;
2584 	nvme_async_event_t event;
2585 
2586 	/*
2587 	 * Check for errors associated with the async request itself. The only
2588 	 * command-specific error is "async event limit exceeded", which
2589 	 * indicates a programming error in the driver and causes a panic in
2590 	 * nvme_check_cmd_status().
2591 	 *
2592 	 * Other possible errors are various scenarios where the async request
2593 	 * was aborted, or internal errors in the device. Internal errors are
2594 	 * reported to FMA, the command aborts need no special handling here.
2595 	 *
2596 	 * And finally, at least qemu nvme does not support async events,
2597 	 * and will return NVME_CQE_SC_GEN_INV_OPC | DNR. If so, we
2598 	 * will avoid posting async events.
2599 	 */
2600 
2601 	if (nvme_check_cmd_status(cmd) != 0) {
2602 		dev_err(cmd->nc_nvme->n_dip, CE_WARN,
2603 		    "!async event request returned failure, sct = 0x%x, "
2604 		    "sc = 0x%x, dnr = %d, m = %d", cmd->nc_cqe.cqe_sf.sf_sct,
2605 		    cmd->nc_cqe.cqe_sf.sf_sc, cmd->nc_cqe.cqe_sf.sf_dnr,
2606 		    cmd->nc_cqe.cqe_sf.sf_m);
2607 
2608 		if (cmd->nc_cqe.cqe_sf.sf_sct == NVME_CQE_SCT_GENERIC &&
2609 		    cmd->nc_cqe.cqe_sf.sf_sc == NVME_CQE_SC_GEN_INTERNAL_ERR) {
2610 			nvme_ctrl_mark_dead(cmd->nc_nvme, B_FALSE);
2611 		}
2612 
2613 		if (cmd->nc_cqe.cqe_sf.sf_sct == NVME_CQE_SCT_GENERIC &&
2614 		    cmd->nc_cqe.cqe_sf.sf_sc == NVME_CQE_SC_GEN_INV_OPC &&
2615 		    cmd->nc_cqe.cqe_sf.sf_dnr == 1) {
2616 			nvme->n_async_event_supported = B_FALSE;
2617 		}
2618 
2619 		nvme_free_cmd(cmd);
2620 		return;
2621 	}
2622 
2623 	event.r = cmd->nc_cqe.cqe_dw0;
2624 
2625 	/* Clear CQE and re-submit the async request. */
2626 	bzero(&cmd->nc_cqe, sizeof (nvme_cqe_t));
2627 	nvme_submit_admin_cmd(nvme->n_adminq, cmd, NULL);
2628 	cmd = NULL;	/* cmd can no longer be used after resubmission */
2629 
2630 	switch (event.b.ae_type) {
2631 	case NVME_ASYNC_TYPE_ERROR:
2632 		if (event.b.ae_logpage == NVME_LOGPAGE_ERROR) {
2633 			if (!nvme_get_logpage_int(nvme, B_FALSE,
2634 			    (void **)&error_log, &logsize,
2635 			    NVME_LOGPAGE_ERROR)) {
2636 				return;
2637 			}
2638 		} else {
2639 			dev_err(nvme->n_dip, CE_WARN, "!wrong logpage in "
2640 			    "async event reply: type=0x%x logpage=0x%x",
2641 			    event.b.ae_type, event.b.ae_logpage);
2642 			NVME_BUMP_STAT(nvme, wrong_logpage);
2643 			return;
2644 		}
2645 
2646 		switch (event.b.ae_info) {
2647 		case NVME_ASYNC_ERROR_INV_SQ:
2648 			dev_err(nvme->n_dip, CE_PANIC, "programming error: "
2649 			    "invalid submission queue");
2650 			return;
2651 
2652 		case NVME_ASYNC_ERROR_INV_DBL:
2653 			dev_err(nvme->n_dip, CE_PANIC, "programming error: "
2654 			    "invalid doorbell write value");
2655 			return;
2656 
2657 		case NVME_ASYNC_ERROR_DIAGFAIL:
2658 			dev_err(nvme->n_dip, CE_WARN, "!diagnostic failure");
2659 			nvme_ctrl_mark_dead(cmd->nc_nvme, B_FALSE);
2660 			NVME_BUMP_STAT(nvme, diagfail_event);
2661 			break;
2662 
2663 		case NVME_ASYNC_ERROR_PERSISTENT:
2664 			dev_err(nvme->n_dip, CE_WARN, "!persistent internal "
2665 			    "device error");
2666 			nvme_ctrl_mark_dead(cmd->nc_nvme, B_FALSE);
2667 			NVME_BUMP_STAT(nvme, persistent_event);
2668 			break;
2669 
2670 		case NVME_ASYNC_ERROR_TRANSIENT:
2671 			dev_err(nvme->n_dip, CE_WARN, "!transient internal "
2672 			    "device error");
2673 			/* TODO: send ereport */
2674 			NVME_BUMP_STAT(nvme, transient_event);
2675 			break;
2676 
2677 		case NVME_ASYNC_ERROR_FW_LOAD:
2678 			dev_err(nvme->n_dip, CE_WARN,
2679 			    "!firmware image load error");
2680 			NVME_BUMP_STAT(nvme, fw_load_event);
2681 			break;
2682 		}
2683 		break;
2684 
2685 	case NVME_ASYNC_TYPE_HEALTH:
2686 		if (event.b.ae_logpage == NVME_LOGPAGE_HEALTH) {
2687 			if (!nvme_get_logpage_int(nvme, B_FALSE,
2688 			    (void **)&health_log, &logsize,
2689 			    NVME_LOGPAGE_HEALTH)) {
2690 				return;
2691 			}
2692 		} else {
2693 			dev_err(nvme->n_dip, CE_WARN, "!wrong logpage in "
2694 			    "type=0x%x logpage=0x%x", event.b.ae_type,
2695 			    event.b.ae_logpage);
2696 			NVME_BUMP_STAT(nvme, wrong_logpage);
2697 			return;
2698 		}
2699 
2700 		switch (event.b.ae_info) {
2701 		case NVME_ASYNC_HEALTH_RELIABILITY:
2702 			dev_err(nvme->n_dip, CE_WARN,
2703 			    "!device reliability compromised");
2704 			/* TODO: send ereport */
2705 			NVME_BUMP_STAT(nvme, reliability_event);
2706 			break;
2707 
2708 		case NVME_ASYNC_HEALTH_TEMPERATURE:
2709 			dev_err(nvme->n_dip, CE_WARN,
2710 			    "!temperature above threshold");
2711 			/* TODO: send ereport */
2712 			NVME_BUMP_STAT(nvme, temperature_event);
2713 			break;
2714 
2715 		case NVME_ASYNC_HEALTH_SPARE:
2716 			dev_err(nvme->n_dip, CE_WARN,
2717 			    "!spare space below threshold");
2718 			/* TODO: send ereport */
2719 			NVME_BUMP_STAT(nvme, spare_event);
2720 			break;
2721 		}
2722 		break;
2723 
2724 	case NVME_ASYNC_TYPE_NOTICE:
2725 		switch (event.b.ae_info) {
2726 		case NVME_ASYNC_NOTICE_NS_CHANGE:
2727 			if (event.b.ae_logpage != NVME_LOGPAGE_NSCHANGE) {
2728 				dev_err(nvme->n_dip, CE_WARN,
2729 				    "!wrong logpage in async event reply: "
2730 				    "type=0x%x logpage=0x%x",
2731 				    event.b.ae_type, event.b.ae_logpage);
2732 				NVME_BUMP_STAT(nvme, wrong_logpage);
2733 				break;
2734 			}
2735 
2736 			dev_err(nvme->n_dip, CE_NOTE,
2737 			    "namespace attribute change event, "
2738 			    "logpage = 0x%x", event.b.ae_logpage);
2739 			NVME_BUMP_STAT(nvme, notice_event);
2740 
2741 			if (!nvme_get_logpage_int(nvme, B_FALSE,
2742 			    (void **)&nslist, &logsize,
2743 			    NVME_LOGPAGE_NSCHANGE)) {
2744 				break;
2745 			}
2746 
2747 			if (nslist->nscl_ns[0] == UINT32_MAX) {
2748 				dev_err(nvme->n_dip, CE_CONT,
2749 				    "more than %u namespaces have changed.\n",
2750 				    NVME_NSCHANGE_LIST_SIZE);
2751 				break;
2752 			}
2753 
2754 			nvme_mgmt_lock(nvme, NVME_MGMT_LOCK_NVME);
2755 			for (uint_t i = 0; i < NVME_NSCHANGE_LIST_SIZE; i++) {
2756 				uint32_t nsid = nslist->nscl_ns[i];
2757 
2758 				if (nsid == 0)	/* end of list */
2759 					break;
2760 
2761 				dev_err(nvme->n_dip, CE_NOTE,
2762 				    "!namespace nvme%d/%u has changed.",
2763 				    ddi_get_instance(nvme->n_dip), nsid);
2764 
2765 				if (nvme_init_ns(nvme, nsid) != DDI_SUCCESS)
2766 					continue;
2767 
2768 				nvme_mgmt_bd_start(nvme);
2769 				bd_state_change(nvme_nsid2ns(nvme,
2770 				    nsid)->ns_bd_hdl);
2771 				nvme_mgmt_bd_end(nvme);
2772 			}
2773 			nvme_mgmt_unlock(nvme);
2774 
2775 			break;
2776 
2777 		case NVME_ASYNC_NOTICE_FW_ACTIVATE:
2778 			dev_err(nvme->n_dip, CE_NOTE,
2779 			    "firmware activation starting, "
2780 			    "logpage = 0x%x", event.b.ae_logpage);
2781 			NVME_BUMP_STAT(nvme, notice_event);
2782 			break;
2783 
2784 		case NVME_ASYNC_NOTICE_TELEMETRY:
2785 			dev_err(nvme->n_dip, CE_NOTE,
2786 			    "telemetry log changed, "
2787 			    "logpage = 0x%x", event.b.ae_logpage);
2788 			NVME_BUMP_STAT(nvme, notice_event);
2789 			break;
2790 
2791 		case NVME_ASYNC_NOTICE_NS_ASYMM:
2792 			dev_err(nvme->n_dip, CE_NOTE,
2793 			    "asymmetric namespace access change, "
2794 			    "logpage = 0x%x", event.b.ae_logpage);
2795 			NVME_BUMP_STAT(nvme, notice_event);
2796 			break;
2797 
2798 		case NVME_ASYNC_NOTICE_LATENCYLOG:
2799 			dev_err(nvme->n_dip, CE_NOTE,
2800 			    "predictable latency event aggregate log change, "
2801 			    "logpage = 0x%x", event.b.ae_logpage);
2802 			NVME_BUMP_STAT(nvme, notice_event);
2803 			break;
2804 
2805 		case NVME_ASYNC_NOTICE_LBASTATUS:
2806 			dev_err(nvme->n_dip, CE_NOTE,
2807 			    "LBA status information alert, "
2808 			    "logpage = 0x%x", event.b.ae_logpage);
2809 			NVME_BUMP_STAT(nvme, notice_event);
2810 			break;
2811 
2812 		case NVME_ASYNC_NOTICE_ENDURANCELOG:
2813 			dev_err(nvme->n_dip, CE_NOTE,
2814 			    "endurance group event aggregate log page change, "
2815 			    "logpage = 0x%x", event.b.ae_logpage);
2816 			NVME_BUMP_STAT(nvme, notice_event);
2817 			break;
2818 
2819 		default:
2820 			dev_err(nvme->n_dip, CE_WARN,
2821 			    "!unknown notice async event received, "
2822 			    "info = 0x%x, logpage = 0x%x", event.b.ae_info,
2823 			    event.b.ae_logpage);
2824 			NVME_BUMP_STAT(nvme, unknown_event);
2825 			break;
2826 		}
2827 		break;
2828 
2829 	case NVME_ASYNC_TYPE_VENDOR:
2830 		dev_err(nvme->n_dip, CE_WARN, "!vendor specific async event "
2831 		    "received, info = 0x%x, logpage = 0x%x", event.b.ae_info,
2832 		    event.b.ae_logpage);
2833 		NVME_BUMP_STAT(nvme, vendor_event);
2834 		break;
2835 
2836 	default:
2837 		dev_err(nvme->n_dip, CE_WARN, "!unknown async event received, "
2838 		    "type = 0x%x, info = 0x%x, logpage = 0x%x", event.b.ae_type,
2839 		    event.b.ae_info, event.b.ae_logpage);
2840 		NVME_BUMP_STAT(nvme, unknown_event);
2841 		break;
2842 	}
2843 
2844 	if (error_log != NULL)
2845 		kmem_free(error_log, logsize);
2846 
2847 	if (health_log != NULL)
2848 		kmem_free(health_log, logsize);
2849 
2850 	if (nslist != NULL)
2851 		kmem_free(nslist, logsize);
2852 }
2853 
2854 static void
2855 nvme_admin_cmd(nvme_cmd_t *cmd, uint32_t sec)
2856 {
2857 	uint32_t qtimeout;
2858 
2859 	ASSERT(cmd->nc_flags & NVME_CMD_F_USELOCK);
2860 
2861 	mutex_enter(&cmd->nc_mutex);
2862 	cmd->nc_timeout = sec;
2863 	nvme_submit_admin_cmd(cmd->nc_nvme->n_adminq, cmd, &qtimeout);
2864 	/*
2865 	 * We will wait for a total of this command's specified timeout plus
2866 	 * the sum of the timeouts of any commands queued ahead of this one. If
2867 	 * we aren't first in the queue, this will inflate the timeout somewhat
2868 	 * but these times are not critical and it means that if we get stuck
2869 	 * behind a long running command such as a namespace format then we
2870 	 * won't time out and trigger an abort.
2871 	 */
2872 	nvme_wait_cmd(cmd, sec + qtimeout);
2873 	mutex_exit(&cmd->nc_mutex);
2874 }
2875 
2876 static void
2877 nvme_async_event(nvme_t *nvme)
2878 {
2879 	nvme_cmd_t *cmd;
2880 
2881 	cmd = nvme_alloc_admin_cmd(nvme, KM_SLEEP);
2882 	cmd->nc_sqid = 0;
2883 	cmd->nc_sqe.sqe_opc = NVME_OPC_ASYNC_EVENT;
2884 	cmd->nc_callback = nvme_async_event_task;
2885 	cmd->nc_flags |= NVME_CMD_F_DONTPANIC;
2886 
2887 	nvme_submit_admin_cmd(nvme->n_adminq, cmd, NULL);
2888 }
2889 
2890 /*
2891  * There are commands such as format or vendor unique commands that are going to
2892  * manipulate the data in a namespace or destroy them, we make sure that none of
2893  * the ones that will be impacted are actually attached.
2894  */
2895 static boolean_t
2896 nvme_no_blkdev_attached(nvme_t *nvme, uint32_t nsid)
2897 {
2898 	ASSERT(nvme_mgmt_lock_held(nvme));
2899 	ASSERT3U(nsid, !=, 0);
2900 
2901 	if (nsid != NVME_NSID_BCAST) {
2902 		nvme_namespace_t *ns = nvme_nsid2ns(nvme, nsid);
2903 		return (!ns->ns_attached);
2904 	}
2905 
2906 	for (uint32_t i = 1; i <= nvme->n_namespace_count; i++) {
2907 		nvme_namespace_t *ns = nvme_nsid2ns(nvme, i);
2908 
2909 		if (ns->ns_attached) {
2910 			return (B_FALSE);
2911 		}
2912 	}
2913 
2914 	return (B_TRUE);
2915 }
2916 
2917 static boolean_t
2918 nvme_format_nvm(nvme_t *nvme, nvme_ioctl_format_t *ioc)
2919 {
2920 	nvme_cmd_t *cmd = nvme_alloc_admin_cmd(nvme, KM_SLEEP);
2921 	nvme_format_nvm_t format_nvm = { 0 };
2922 	boolean_t ret;
2923 
2924 	format_nvm.b.fm_lbaf = bitx32(ioc->nif_lbaf, 3, 0);
2925 	format_nvm.b.fm_ses = bitx32(ioc->nif_ses, 2, 0);
2926 
2927 	cmd->nc_sqid = 0;
2928 	cmd->nc_callback = nvme_wakeup_cmd;
2929 	cmd->nc_sqe.sqe_nsid = ioc->nif_common.nioc_nsid;
2930 	cmd->nc_sqe.sqe_opc = NVME_OPC_NVM_FORMAT;
2931 	cmd->nc_sqe.sqe_cdw10 = format_nvm.r;
2932 
2933 	/*
2934 	 * We don't want to panic on any format commands. There are two reasons
2935 	 * for this:
2936 	 *
2937 	 * 1) All format commands are initiated by users. We don't want to panic
2938 	 * on user commands.
2939 	 *
2940 	 * 2) Several devices like the Samsung SM951 don't allow formatting of
2941 	 * all namespaces in one command and we'd prefer to handle that
2942 	 * gracefully.
2943 	 */
2944 	cmd->nc_flags |= NVME_CMD_F_DONTPANIC;
2945 
2946 	nvme_admin_cmd(cmd, nvme_format_cmd_timeout);
2947 
2948 	if (!nvme_check_cmd_status_ioctl(cmd, &ioc->nif_common) != 0) {
2949 		dev_err(nvme->n_dip, CE_WARN,
2950 		    "!FORMAT failed with sct = %x, sc = %x",
2951 		    cmd->nc_cqe.cqe_sf.sf_sct, cmd->nc_cqe.cqe_sf.sf_sc);
2952 		ret = B_FALSE;
2953 		goto fail;
2954 	}
2955 
2956 	ret = B_TRUE;
2957 fail:
2958 	nvme_free_cmd(cmd);
2959 	return (ret);
2960 }
2961 
2962 /*
2963  * Retrieve a specific log page. The contents of the log page request should
2964  * have already been validated by the system.
2965  */
2966 static boolean_t
2967 nvme_get_logpage(nvme_t *nvme, boolean_t user, nvme_ioctl_get_logpage_t *log,
2968     void **buf)
2969 {
2970 	nvme_cmd_t *cmd = nvme_alloc_admin_cmd(nvme, KM_SLEEP);
2971 	nvme_getlogpage_dw10_t dw10;
2972 	uint32_t offlo, offhi;
2973 	nvme_getlogpage_dw11_t dw11;
2974 	nvme_getlogpage_dw14_t dw14;
2975 	uint32_t ndw;
2976 	boolean_t ret = B_FALSE;
2977 
2978 	bzero(&dw10, sizeof (dw10));
2979 	bzero(&dw11, sizeof (dw11));
2980 	bzero(&dw14, sizeof (dw14));
2981 
2982 	cmd->nc_sqid = 0;
2983 	cmd->nc_callback = nvme_wakeup_cmd;
2984 	cmd->nc_sqe.sqe_opc = NVME_OPC_GET_LOG_PAGE;
2985 	cmd->nc_sqe.sqe_nsid = log->nigl_common.nioc_nsid;
2986 
2987 	if (user)
2988 		cmd->nc_flags |= NVME_CMD_F_DONTPANIC;
2989 
2990 	/*
2991 	 * The size field is the number of double words, but is a zeros based
2992 	 * value. We need to store our actual value minus one.
2993 	 */
2994 	ndw = (uint32_t)(log->nigl_len / 4);
2995 	ASSERT3U(ndw, >, 0);
2996 	ndw--;
2997 
2998 	dw10.b.lp_lid = bitx32(log->nigl_lid, 7, 0);
2999 	dw10.b.lp_lsp = bitx32(log->nigl_lsp, 6, 0);
3000 	dw10.b.lp_rae = bitx32(log->nigl_lsp, 0, 0);
3001 	dw10.b.lp_lnumdl = bitx32(ndw, 15, 0);
3002 
3003 	dw11.b.lp_numdu = bitx32(ndw, 31, 16);
3004 	dw11.b.lp_lsi = bitx32(log->nigl_lsi, 15, 0);
3005 
3006 	offlo = bitx64(log->nigl_offset, 31, 0);
3007 	offhi = bitx64(log->nigl_offset, 63, 32);
3008 
3009 	dw14.b.lp_csi = bitx32(log->nigl_csi, 7, 0);
3010 
3011 	cmd->nc_sqe.sqe_cdw10 = dw10.r;
3012 	cmd->nc_sqe.sqe_cdw11 = dw11.r;
3013 	cmd->nc_sqe.sqe_cdw12 = offlo;
3014 	cmd->nc_sqe.sqe_cdw13 = offhi;
3015 	cmd->nc_sqe.sqe_cdw14 = dw14.r;
3016 
3017 	if (nvme_zalloc_dma(nvme, log->nigl_len, DDI_DMA_READ,
3018 	    &nvme->n_prp_dma_attr, &cmd->nc_dma) != DDI_SUCCESS) {
3019 		dev_err(nvme->n_dip, CE_WARN,
3020 		    "!nvme_zalloc_dma failed for GET LOG PAGE");
3021 		ret = nvme_ioctl_error(&log->nigl_common,
3022 		    NVME_IOCTL_E_NO_DMA_MEM, 0, 0);
3023 		goto fail;
3024 	}
3025 
3026 	if (nvme_fill_prp(cmd, cmd->nc_dma->nd_dmah) != 0) {
3027 		ret = nvme_ioctl_error(&log->nigl_common,
3028 		    NVME_IOCTL_E_NO_DMA_MEM, 0, 0);
3029 		goto fail;
3030 	}
3031 	nvme_admin_cmd(cmd, nvme_admin_cmd_timeout);
3032 
3033 	if (!nvme_check_cmd_status_ioctl(cmd, &log->nigl_common)) {
3034 		if (!user) {
3035 			dev_err(nvme->n_dip, CE_WARN,
3036 			    "!GET LOG PAGE failed with sct = %x, sc = %x",
3037 			    cmd->nc_cqe.cqe_sf.sf_sct,
3038 			    cmd->nc_cqe.cqe_sf.sf_sc);
3039 		}
3040 		ret = B_FALSE;
3041 		goto fail;
3042 	}
3043 
3044 	*buf = kmem_alloc(log->nigl_len, KM_SLEEP);
3045 	bcopy(cmd->nc_dma->nd_memp, *buf, log->nigl_len);
3046 
3047 	ret = B_TRUE;
3048 fail:
3049 	nvme_free_cmd(cmd);
3050 
3051 	return (ret);
3052 }
3053 
3054 /*
3055  * This is an internal wrapper for when the kernel wants to get a log page.
3056  * Currently this assumes that the only thing that is required is the log page
3057  * ID. If more information is required, we'll be better served to just use the
3058  * general ioctl interface.
3059  */
3060 static boolean_t
3061 nvme_get_logpage_int(nvme_t *nvme, boolean_t user, void **buf, size_t *bufsize,
3062     uint8_t lid)
3063 {
3064 	const nvme_log_page_info_t *info = NULL;
3065 	nvme_ioctl_get_logpage_t log;
3066 	nvme_valid_ctrl_data_t data;
3067 	boolean_t bret;
3068 	bool var;
3069 
3070 	for (size_t i = 0; i < nvme_std_log_npages; i++) {
3071 		if (nvme_std_log_pages[i].nlpi_lid == lid &&
3072 		    nvme_std_log_pages[i].nlpi_csi == NVME_CSI_NVM) {
3073 			info = &nvme_std_log_pages[i];
3074 			break;
3075 		}
3076 	}
3077 
3078 	if (info == NULL) {
3079 		return (B_FALSE);
3080 	}
3081 
3082 	data.vcd_vers = &nvme->n_version;
3083 	data.vcd_id = nvme->n_idctl;
3084 	bzero(&log, sizeof (log));
3085 	log.nigl_common.nioc_nsid = NVME_NSID_BCAST;
3086 	log.nigl_csi = info->nlpi_csi;
3087 	log.nigl_lid = info->nlpi_lid;
3088 	log.nigl_len = nvme_log_page_info_size(info, &data, &var);
3089 
3090 	/*
3091 	 * We only support getting standard fixed-length log pages through the
3092 	 * kernel interface at this time. If a log page either has an unknown
3093 	 * size or has a variable length, then we cannot get it.
3094 	 */
3095 	if (log.nigl_len == 0 || var) {
3096 		return (B_FALSE);
3097 	}
3098 
3099 	bret = nvme_get_logpage(nvme, user, &log, buf);
3100 	if (!bret) {
3101 		return (B_FALSE);
3102 	}
3103 
3104 	*bufsize = log.nigl_len;
3105 	return (B_TRUE);
3106 }
3107 
3108 static boolean_t
3109 nvme_identify(nvme_t *nvme, boolean_t user, nvme_ioctl_identify_t *ioc,
3110     void **buf)
3111 {
3112 	nvme_cmd_t *cmd = nvme_alloc_admin_cmd(nvme, KM_SLEEP);
3113 	boolean_t ret = B_FALSE;
3114 	nvme_identify_dw10_t dw10;
3115 
3116 	ASSERT3P(buf, !=, NULL);
3117 
3118 	bzero(&dw10, sizeof (dw10));
3119 
3120 	cmd->nc_sqid = 0;
3121 	cmd->nc_callback = nvme_wakeup_cmd;
3122 	cmd->nc_sqe.sqe_opc = NVME_OPC_IDENTIFY;
3123 	cmd->nc_sqe.sqe_nsid = ioc->nid_common.nioc_nsid;
3124 
3125 	dw10.b.id_cns = bitx32(ioc->nid_cns, 7, 0);
3126 	dw10.b.id_cntid = bitx32(ioc->nid_ctrlid, 15, 0);
3127 
3128 	cmd->nc_sqe.sqe_cdw10 = dw10.r;
3129 
3130 	if (nvme_zalloc_dma(nvme, NVME_IDENTIFY_BUFSIZE, DDI_DMA_READ,
3131 	    &nvme->n_prp_dma_attr, &cmd->nc_dma) != DDI_SUCCESS) {
3132 		dev_err(nvme->n_dip, CE_WARN,
3133 		    "!nvme_zalloc_dma failed for IDENTIFY");
3134 		ret = nvme_ioctl_error(&ioc->nid_common,
3135 		    NVME_IOCTL_E_NO_DMA_MEM, 0, 0);
3136 		goto fail;
3137 	}
3138 
3139 	if (cmd->nc_dma->nd_ncookie > 2) {
3140 		dev_err(nvme->n_dip, CE_WARN,
3141 		    "!too many DMA cookies for IDENTIFY");
3142 		NVME_BUMP_STAT(nvme, too_many_cookies);
3143 		ret = nvme_ioctl_error(&ioc->nid_common,
3144 		    NVME_IOCTL_E_BAD_PRP, 0, 0);
3145 		goto fail;
3146 	}
3147 
3148 	cmd->nc_sqe.sqe_dptr.d_prp[0] = cmd->nc_dma->nd_cookie.dmac_laddress;
3149 	if (cmd->nc_dma->nd_ncookie > 1) {
3150 		ddi_dma_nextcookie(cmd->nc_dma->nd_dmah,
3151 		    &cmd->nc_dma->nd_cookie);
3152 		cmd->nc_sqe.sqe_dptr.d_prp[1] =
3153 		    cmd->nc_dma->nd_cookie.dmac_laddress;
3154 	}
3155 
3156 	if (user)
3157 		cmd->nc_flags |= NVME_CMD_F_DONTPANIC;
3158 
3159 	nvme_admin_cmd(cmd, nvme_admin_cmd_timeout);
3160 
3161 	if (!nvme_check_cmd_status_ioctl(cmd, &ioc->nid_common)) {
3162 		dev_err(nvme->n_dip, CE_WARN,
3163 		    "!IDENTIFY failed with sct = %x, sc = %x",
3164 		    cmd->nc_cqe.cqe_sf.sf_sct, cmd->nc_cqe.cqe_sf.sf_sc);
3165 		ret = B_FALSE;
3166 		goto fail;
3167 	}
3168 
3169 	*buf = kmem_alloc(NVME_IDENTIFY_BUFSIZE, KM_SLEEP);
3170 	bcopy(cmd->nc_dma->nd_memp, *buf, NVME_IDENTIFY_BUFSIZE);
3171 	ret = B_TRUE;
3172 
3173 fail:
3174 	nvme_free_cmd(cmd);
3175 
3176 	return (ret);
3177 }
3178 
3179 static boolean_t
3180 nvme_identify_int(nvme_t *nvme, uint32_t nsid, uint8_t cns, void **buf)
3181 {
3182 	nvme_ioctl_identify_t id;
3183 
3184 	bzero(&id, sizeof (nvme_ioctl_identify_t));
3185 	id.nid_common.nioc_nsid = nsid;
3186 	id.nid_cns = cns;
3187 
3188 	return (nvme_identify(nvme, B_FALSE, &id, buf));
3189 }
3190 
3191 static int
3192 nvme_set_features(nvme_t *nvme, boolean_t user, uint32_t nsid, uint8_t feature,
3193     uint32_t val, uint32_t *res)
3194 {
3195 	_NOTE(ARGUNUSED(nsid));
3196 	nvme_cmd_t *cmd = nvme_alloc_admin_cmd(nvme, KM_SLEEP);
3197 	int ret = EINVAL;
3198 
3199 	ASSERT(res != NULL);
3200 
3201 	cmd->nc_sqid = 0;
3202 	cmd->nc_callback = nvme_wakeup_cmd;
3203 	cmd->nc_sqe.sqe_opc = NVME_OPC_SET_FEATURES;
3204 	cmd->nc_sqe.sqe_cdw10 = feature;
3205 	cmd->nc_sqe.sqe_cdw11 = val;
3206 
3207 	if (user)
3208 		cmd->nc_flags |= NVME_CMD_F_DONTPANIC;
3209 
3210 	switch (feature) {
3211 	case NVME_FEAT_WRITE_CACHE:
3212 		if (!nvme->n_write_cache_present)
3213 			goto fail;
3214 		break;
3215 
3216 	case NVME_FEAT_NQUEUES:
3217 		break;
3218 
3219 	default:
3220 		goto fail;
3221 	}
3222 
3223 	nvme_admin_cmd(cmd, nvme_admin_cmd_timeout);
3224 
3225 	if ((ret = nvme_check_cmd_status(cmd)) != 0) {
3226 		dev_err(nvme->n_dip, CE_WARN,
3227 		    "!SET FEATURES %d failed with sct = %x, sc = %x",
3228 		    feature, cmd->nc_cqe.cqe_sf.sf_sct,
3229 		    cmd->nc_cqe.cqe_sf.sf_sc);
3230 		goto fail;
3231 	}
3232 
3233 	*res = cmd->nc_cqe.cqe_dw0;
3234 
3235 fail:
3236 	nvme_free_cmd(cmd);
3237 	return (ret);
3238 }
3239 
3240 static int
3241 nvme_write_cache_set(nvme_t *nvme, boolean_t enable)
3242 {
3243 	nvme_write_cache_t nwc = { 0 };
3244 
3245 	if (enable)
3246 		nwc.b.wc_wce = 1;
3247 
3248 	/*
3249 	 * We've seen some cases where this fails due to us being told we've
3250 	 * specified an invalid namespace when operating against the Xen xcp-ng
3251 	 * qemu NVMe virtual device. As such, we generally ensure that trying to
3252 	 * enable this doesn't lead us to panic. It's not completely clear why
3253 	 * specifying namespace zero here fails, but not when we're setting the
3254 	 * number of queues below.
3255 	 */
3256 	return (nvme_set_features(nvme, B_TRUE, 0, NVME_FEAT_WRITE_CACHE,
3257 	    nwc.r, &nwc.r));
3258 }
3259 
3260 static int
3261 nvme_set_nqueues(nvme_t *nvme)
3262 {
3263 	nvme_nqueues_t nq = { 0 };
3264 	int ret;
3265 
3266 	/*
3267 	 * The default is to allocate one completion queue per vector.
3268 	 */
3269 	if (nvme->n_completion_queues == -1)
3270 		nvme->n_completion_queues = nvme->n_intr_cnt;
3271 
3272 	/*
3273 	 * There is no point in having more completion queues than
3274 	 * interrupt vectors.
3275 	 */
3276 	nvme->n_completion_queues = MIN(nvme->n_completion_queues,
3277 	    nvme->n_intr_cnt);
3278 
3279 	/*
3280 	 * The default is to use one submission queue per completion queue.
3281 	 */
3282 	if (nvme->n_submission_queues == -1)
3283 		nvme->n_submission_queues = nvme->n_completion_queues;
3284 
3285 	/*
3286 	 * There is no point in having more completion queues than
3287 	 * submission queues.
3288 	 */
3289 	nvme->n_completion_queues = MIN(nvme->n_completion_queues,
3290 	    nvme->n_submission_queues);
3291 
3292 	ASSERT(nvme->n_submission_queues > 0);
3293 	ASSERT(nvme->n_completion_queues > 0);
3294 
3295 	nq.b.nq_nsq = nvme->n_submission_queues - 1;
3296 	nq.b.nq_ncq = nvme->n_completion_queues - 1;
3297 
3298 	ret = nvme_set_features(nvme, B_FALSE, 0, NVME_FEAT_NQUEUES, nq.r,
3299 	    &nq.r);
3300 
3301 	if (ret == 0) {
3302 		/*
3303 		 * Never use more than the requested number of queues.
3304 		 */
3305 		nvme->n_submission_queues = MIN(nvme->n_submission_queues,
3306 		    nq.b.nq_nsq + 1);
3307 		nvme->n_completion_queues = MIN(nvme->n_completion_queues,
3308 		    nq.b.nq_ncq + 1);
3309 	}
3310 
3311 	return (ret);
3312 }
3313 
3314 static int
3315 nvme_create_completion_queue(nvme_t *nvme, nvme_cq_t *cq)
3316 {
3317 	nvme_cmd_t *cmd = nvme_alloc_admin_cmd(nvme, KM_SLEEP);
3318 	nvme_create_queue_dw10_t dw10 = { 0 };
3319 	nvme_create_cq_dw11_t c_dw11 = { 0 };
3320 	int ret;
3321 
3322 	dw10.b.q_qid = cq->ncq_id;
3323 	dw10.b.q_qsize = cq->ncq_nentry - 1;
3324 
3325 	c_dw11.b.cq_pc = 1;
3326 	c_dw11.b.cq_ien = 1;
3327 	c_dw11.b.cq_iv = cq->ncq_id % nvme->n_intr_cnt;
3328 
3329 	cmd->nc_sqid = 0;
3330 	cmd->nc_callback = nvme_wakeup_cmd;
3331 	cmd->nc_sqe.sqe_opc = NVME_OPC_CREATE_CQUEUE;
3332 	cmd->nc_sqe.sqe_cdw10 = dw10.r;
3333 	cmd->nc_sqe.sqe_cdw11 = c_dw11.r;
3334 	cmd->nc_sqe.sqe_dptr.d_prp[0] = cq->ncq_dma->nd_cookie.dmac_laddress;
3335 
3336 	nvme_admin_cmd(cmd, nvme_admin_cmd_timeout);
3337 
3338 	if ((ret = nvme_check_cmd_status(cmd)) != 0) {
3339 		dev_err(nvme->n_dip, CE_WARN,
3340 		    "!CREATE CQUEUE failed with sct = %x, sc = %x",
3341 		    cmd->nc_cqe.cqe_sf.sf_sct, cmd->nc_cqe.cqe_sf.sf_sc);
3342 	}
3343 
3344 	nvme_free_cmd(cmd);
3345 
3346 	return (ret);
3347 }
3348 
3349 static int
3350 nvme_create_io_qpair(nvme_t *nvme, nvme_qpair_t *qp, uint16_t idx)
3351 {
3352 	nvme_cq_t *cq = qp->nq_cq;
3353 	nvme_cmd_t *cmd;
3354 	nvme_create_queue_dw10_t dw10 = { 0 };
3355 	nvme_create_sq_dw11_t s_dw11 = { 0 };
3356 	int ret;
3357 
3358 	/*
3359 	 * It is possible to have more qpairs than completion queues,
3360 	 * and when the idx > ncq_id, that completion queue is shared
3361 	 * and has already been created.
3362 	 */
3363 	if (idx <= cq->ncq_id &&
3364 	    nvme_create_completion_queue(nvme, cq) != DDI_SUCCESS)
3365 		return (DDI_FAILURE);
3366 
3367 	dw10.b.q_qid = idx;
3368 	dw10.b.q_qsize = qp->nq_nentry - 1;
3369 
3370 	s_dw11.b.sq_pc = 1;
3371 	s_dw11.b.sq_cqid = cq->ncq_id;
3372 
3373 	cmd = nvme_alloc_admin_cmd(nvme, KM_SLEEP);
3374 	cmd->nc_sqid = 0;
3375 	cmd->nc_callback = nvme_wakeup_cmd;
3376 	cmd->nc_sqe.sqe_opc = NVME_OPC_CREATE_SQUEUE;
3377 	cmd->nc_sqe.sqe_cdw10 = dw10.r;
3378 	cmd->nc_sqe.sqe_cdw11 = s_dw11.r;
3379 	cmd->nc_sqe.sqe_dptr.d_prp[0] = qp->nq_sqdma->nd_cookie.dmac_laddress;
3380 
3381 	nvme_admin_cmd(cmd, nvme_admin_cmd_timeout);
3382 
3383 	if ((ret = nvme_check_cmd_status(cmd)) != 0) {
3384 		dev_err(nvme->n_dip, CE_WARN,
3385 		    "!CREATE SQUEUE failed with sct = %x, sc = %x",
3386 		    cmd->nc_cqe.cqe_sf.sf_sct, cmd->nc_cqe.cqe_sf.sf_sc);
3387 	}
3388 
3389 	nvme_free_cmd(cmd);
3390 
3391 	return (ret);
3392 }
3393 
3394 static boolean_t
3395 nvme_reset(nvme_t *nvme, boolean_t quiesce)
3396 {
3397 	nvme_reg_csts_t csts;
3398 	int i;
3399 
3400 	nvme_put32(nvme, NVME_REG_CC, 0);
3401 
3402 	csts.r = nvme_get32(nvme, NVME_REG_CSTS);
3403 	if (csts.b.csts_rdy == 1) {
3404 		nvme_put32(nvme, NVME_REG_CC, 0);
3405 
3406 		/*
3407 		 * The timeout value is from the Controller Capabilities
3408 		 * register (CAP.TO, section 3.1.1). This is the worst case
3409 		 * time to wait for CSTS.RDY to transition from 1 to 0 after
3410 		 * CC.EN transitions from 1 to 0.
3411 		 *
3412 		 * The timeout units are in 500 ms units, and we are delaying
3413 		 * in 50ms chunks, hence counting to n_timeout * 10.
3414 		 */
3415 		for (i = 0; i < nvme->n_timeout * 10; i++) {
3416 			csts.r = nvme_get32(nvme, NVME_REG_CSTS);
3417 			if (csts.b.csts_rdy == 0)
3418 				break;
3419 
3420 			/*
3421 			 * Quiescing drivers should not use locks or timeouts,
3422 			 * so if this is the quiesce path, use a quiesce-safe
3423 			 * delay.
3424 			 */
3425 			if (quiesce) {
3426 				drv_usecwait(50000);
3427 			} else {
3428 				delay(drv_usectohz(50000));
3429 			}
3430 		}
3431 	}
3432 
3433 	nvme_put32(nvme, NVME_REG_AQA, 0);
3434 	nvme_put32(nvme, NVME_REG_ASQ, 0);
3435 	nvme_put32(nvme, NVME_REG_ACQ, 0);
3436 
3437 	csts.r = nvme_get32(nvme, NVME_REG_CSTS);
3438 	return (csts.b.csts_rdy == 0 ? B_TRUE : B_FALSE);
3439 }
3440 
3441 static void
3442 nvme_shutdown(nvme_t *nvme, boolean_t quiesce)
3443 {
3444 	nvme_reg_cc_t cc;
3445 	nvme_reg_csts_t csts;
3446 	int i;
3447 
3448 	cc.r = nvme_get32(nvme, NVME_REG_CC);
3449 	cc.b.cc_shn = NVME_CC_SHN_NORMAL;
3450 	nvme_put32(nvme, NVME_REG_CC, cc.r);
3451 
3452 	for (i = 0; i < 10; i++) {
3453 		csts.r = nvme_get32(nvme, NVME_REG_CSTS);
3454 		if (csts.b.csts_shst == NVME_CSTS_SHN_COMPLETE)
3455 			break;
3456 
3457 		if (quiesce) {
3458 			drv_usecwait(100000);
3459 		} else {
3460 			delay(drv_usectohz(100000));
3461 		}
3462 	}
3463 }
3464 
3465 /*
3466  * Return length of string without trailing spaces.
3467  */
3468 static int
3469 nvme_strlen(const char *str, int len)
3470 {
3471 	if (len <= 0)
3472 		return (0);
3473 
3474 	while (str[--len] == ' ')
3475 		;
3476 
3477 	return (++len);
3478 }
3479 
3480 static void
3481 nvme_config_min_block_size(nvme_t *nvme, char *model, char *val)
3482 {
3483 	ulong_t bsize = 0;
3484 	char *msg = "";
3485 
3486 	if (ddi_strtoul(val, NULL, 0, &bsize) != 0)
3487 		goto err;
3488 
3489 	if (!ISP2(bsize)) {
3490 		msg = ": not a power of 2";
3491 		goto err;
3492 	}
3493 
3494 	if (bsize < NVME_DEFAULT_MIN_BLOCK_SIZE) {
3495 		msg = ": too low";
3496 		goto err;
3497 	}
3498 
3499 	nvme->n_min_block_size = bsize;
3500 	return;
3501 
3502 err:
3503 	dev_err(nvme->n_dip, CE_WARN,
3504 	    "!nvme-config-list: ignoring invalid min-phys-block-size '%s' "
3505 	    "for model '%s'%s", val, model, msg);
3506 
3507 	nvme->n_min_block_size = NVME_DEFAULT_MIN_BLOCK_SIZE;
3508 }
3509 
3510 static void
3511 nvme_config_boolean(nvme_t *nvme, char *model, char *name, char *val,
3512     boolean_t *b)
3513 {
3514 	if (strcmp(val, "on") == 0 ||
3515 	    strcmp(val, "true") == 0)
3516 		*b = B_TRUE;
3517 	else if (strcmp(val, "off") == 0 ||
3518 	    strcmp(val, "false") == 0)
3519 		*b = B_FALSE;
3520 	else
3521 		dev_err(nvme->n_dip, CE_WARN,
3522 		    "!nvme-config-list: invalid value for %s '%s'"
3523 		    " for model '%s', ignoring", name, val, model);
3524 }
3525 
3526 static void
3527 nvme_config_list(nvme_t *nvme)
3528 {
3529 	char	**config_list;
3530 	uint_t	nelem;
3531 	int	rv, i;
3532 
3533 	/*
3534 	 * We're following the pattern of 'sd-config-list' here, but extend it.
3535 	 * Instead of two we have three separate strings for "model", "fwrev",
3536 	 * and "name-value-list".
3537 	 */
3538 	rv = ddi_prop_lookup_string_array(DDI_DEV_T_ANY, nvme->n_dip,
3539 	    DDI_PROP_DONTPASS, "nvme-config-list", &config_list, &nelem);
3540 
3541 	if (rv != DDI_PROP_SUCCESS) {
3542 		if (rv == DDI_PROP_CANNOT_DECODE) {
3543 			dev_err(nvme->n_dip, CE_WARN,
3544 			    "!nvme-config-list: cannot be decoded");
3545 		}
3546 
3547 		return;
3548 	}
3549 
3550 	if ((nelem % 3) != 0) {
3551 		dev_err(nvme->n_dip, CE_WARN, "!nvme-config-list: must be "
3552 		    "triplets of <model>/<fwrev>/<name-value-list> strings ");
3553 		goto out;
3554 	}
3555 
3556 	for (i = 0; i < nelem; i += 3) {
3557 		char	*model = config_list[i];
3558 		char	*fwrev = config_list[i + 1];
3559 		char	*nvp, *save_nv;
3560 		int	id_model_len, id_fwrev_len;
3561 
3562 		id_model_len = nvme_strlen(nvme->n_idctl->id_model,
3563 		    sizeof (nvme->n_idctl->id_model));
3564 
3565 		if (strlen(model) != id_model_len)
3566 			continue;
3567 
3568 		if (strncmp(model, nvme->n_idctl->id_model, id_model_len) != 0)
3569 			continue;
3570 
3571 		id_fwrev_len = nvme_strlen(nvme->n_idctl->id_fwrev,
3572 		    sizeof (nvme->n_idctl->id_fwrev));
3573 
3574 		if (strlen(fwrev) != 0) {
3575 			boolean_t match = B_FALSE;
3576 			char *fwr, *last_fw;
3577 
3578 			for (fwr = strtok_r(fwrev, ",", &last_fw);
3579 			    fwr != NULL;
3580 			    fwr = strtok_r(NULL, ",", &last_fw)) {
3581 				if (strlen(fwr) != id_fwrev_len)
3582 					continue;
3583 
3584 				if (strncmp(fwr, nvme->n_idctl->id_fwrev,
3585 				    id_fwrev_len) == 0)
3586 					match = B_TRUE;
3587 			}
3588 
3589 			if (!match)
3590 				continue;
3591 		}
3592 
3593 		/*
3594 		 * We should now have a comma-separated list of name:value
3595 		 * pairs.
3596 		 */
3597 		for (nvp = strtok_r(config_list[i + 2], ",", &save_nv);
3598 		    nvp != NULL; nvp = strtok_r(NULL, ",", &save_nv)) {
3599 			char	*name = nvp;
3600 			char	*val = strchr(nvp, ':');
3601 
3602 			if (val == NULL || name == val) {
3603 				dev_err(nvme->n_dip, CE_WARN,
3604 				    "!nvme-config-list: <name-value-list> "
3605 				    "for model '%s' is malformed", model);
3606 				goto out;
3607 			}
3608 
3609 			/*
3610 			 * Null-terminate 'name', move 'val' past ':' sep.
3611 			 */
3612 			*val++ = '\0';
3613 
3614 			/*
3615 			 * Process the name:val pairs that we know about.
3616 			 */
3617 			if (strcmp(name, "ignore-unknown-vendor-status") == 0) {
3618 				nvme_config_boolean(nvme, model, name, val,
3619 				    &nvme->n_ignore_unknown_vendor_status);
3620 			} else if (strcmp(name, "min-phys-block-size") == 0) {
3621 				nvme_config_min_block_size(nvme, model, val);
3622 			} else if (strcmp(name, "volatile-write-cache") == 0) {
3623 				nvme_config_boolean(nvme, model, name, val,
3624 				    &nvme->n_write_cache_enabled);
3625 			} else {
3626 				/*
3627 				 * Unknown 'name'.
3628 				 */
3629 				dev_err(nvme->n_dip, CE_WARN,
3630 				    "!nvme-config-list: unknown config '%s' "
3631 				    "for model '%s', ignoring", name, model);
3632 			}
3633 		}
3634 	}
3635 
3636 out:
3637 	ddi_prop_free(config_list);
3638 }
3639 
3640 static void
3641 nvme_prepare_devid(nvme_t *nvme, uint32_t nsid)
3642 {
3643 	/*
3644 	 * Section 7.7 of the spec describes how to get a unique ID for
3645 	 * the controller: the vendor ID, the model name and the serial
3646 	 * number shall be unique when combined.
3647 	 *
3648 	 * If a namespace has no EUI64 we use the above and add the hex
3649 	 * namespace ID to get a unique ID for the namespace.
3650 	 */
3651 	char model[sizeof (nvme->n_idctl->id_model) + 1];
3652 	char serial[sizeof (nvme->n_idctl->id_serial) + 1];
3653 
3654 	bcopy(nvme->n_idctl->id_model, model, sizeof (nvme->n_idctl->id_model));
3655 	bcopy(nvme->n_idctl->id_serial, serial,
3656 	    sizeof (nvme->n_idctl->id_serial));
3657 
3658 	model[sizeof (nvme->n_idctl->id_model)] = '\0';
3659 	serial[sizeof (nvme->n_idctl->id_serial)] = '\0';
3660 
3661 	nvme_nsid2ns(nvme, nsid)->ns_devid = kmem_asprintf("%4X-%s-%s-%X",
3662 	    nvme->n_idctl->id_vid, model, serial, nsid);
3663 }
3664 
3665 static nvme_identify_nsid_list_t *
3666 nvme_update_nsid_list(nvme_t *nvme, int cns)
3667 {
3668 	nvme_identify_nsid_list_t *nslist;
3669 
3670 	/*
3671 	 * We currently don't handle cases where there are more than
3672 	 * 1024 active namespaces, requiring several IDENTIFY commands.
3673 	 */
3674 	if (nvme_identify_int(nvme, 0, cns, (void **)&nslist))
3675 		return (nslist);
3676 
3677 	return (NULL);
3678 }
3679 
3680 nvme_namespace_t *
3681 nvme_nsid2ns(nvme_t *nvme, uint32_t nsid)
3682 {
3683 	ASSERT3U(nsid, !=, 0);
3684 	ASSERT3U(nsid, <=, nvme->n_namespace_count);
3685 	return (&nvme->n_ns[nsid - 1]);
3686 }
3687 
3688 static boolean_t
3689 nvme_allocated_ns(nvme_namespace_t *ns)
3690 {
3691 	nvme_t *nvme = ns->ns_nvme;
3692 	uint32_t i;
3693 
3694 	ASSERT(nvme_mgmt_lock_held(nvme));
3695 
3696 	/*
3697 	 * If supported, update the list of allocated namespace IDs.
3698 	 */
3699 	if (NVME_VERSION_ATLEAST(&nvme->n_version, 1, 2) &&
3700 	    nvme->n_idctl->id_oacs.oa_nsmgmt != 0) {
3701 		nvme_identify_nsid_list_t *nslist = nvme_update_nsid_list(nvme,
3702 		    NVME_IDENTIFY_NSID_ALLOC_LIST);
3703 		boolean_t found = B_FALSE;
3704 
3705 		/*
3706 		 * When namespace management is supported, this really shouldn't
3707 		 * be NULL. Treat all namespaces as allocated if it is.
3708 		 */
3709 		if (nslist == NULL)
3710 			return (B_TRUE);
3711 
3712 		for (i = 0; i < ARRAY_SIZE(nslist->nl_nsid); i++) {
3713 			if (ns->ns_id == 0)
3714 				break;
3715 
3716 			if (ns->ns_id == nslist->nl_nsid[i])
3717 				found = B_TRUE;
3718 		}
3719 
3720 		kmem_free(nslist, NVME_IDENTIFY_BUFSIZE);
3721 		return (found);
3722 	} else {
3723 		/*
3724 		 * If namespace management isn't supported, report all
3725 		 * namespaces as allocated.
3726 		 */
3727 		return (B_TRUE);
3728 	}
3729 }
3730 
3731 static boolean_t
3732 nvme_active_ns(nvme_namespace_t *ns)
3733 {
3734 	nvme_t *nvme = ns->ns_nvme;
3735 	uint64_t *ptr;
3736 	uint32_t i;
3737 
3738 	ASSERT(nvme_mgmt_lock_held(nvme));
3739 
3740 	/*
3741 	 * If supported, update the list of active namespace IDs.
3742 	 */
3743 	if (NVME_VERSION_ATLEAST(&nvme->n_version, 1, 1)) {
3744 		nvme_identify_nsid_list_t *nslist = nvme_update_nsid_list(nvme,
3745 		    NVME_IDENTIFY_NSID_LIST);
3746 		boolean_t found = B_FALSE;
3747 
3748 		/*
3749 		 * When namespace management is supported, this really shouldn't
3750 		 * be NULL. Treat all namespaces as allocated if it is.
3751 		 */
3752 		if (nslist == NULL)
3753 			return (B_TRUE);
3754 
3755 		for (i = 0; i < ARRAY_SIZE(nslist->nl_nsid); i++) {
3756 			if (ns->ns_id == 0)
3757 				break;
3758 
3759 			if (ns->ns_id == nslist->nl_nsid[i])
3760 				found = B_TRUE;
3761 		}
3762 
3763 		kmem_free(nslist, NVME_IDENTIFY_BUFSIZE);
3764 		return (found);
3765 	}
3766 
3767 	/*
3768 	 * Workaround for revision 1.0:
3769 	 * Check whether the IDENTIFY NAMESPACE data is zero-filled.
3770 	 */
3771 	for (ptr = (uint64_t *)ns->ns_idns;
3772 	    ptr != (uint64_t *)(ns->ns_idns + 1);
3773 	    ptr++) {
3774 		if (*ptr != 0) {
3775 			return (B_TRUE);
3776 		}
3777 	}
3778 
3779 	return (B_FALSE);
3780 }
3781 
3782 static int
3783 nvme_init_ns(nvme_t *nvme, uint32_t nsid)
3784 {
3785 	nvme_namespace_t *ns = nvme_nsid2ns(nvme, nsid);
3786 	nvme_identify_nsid_t *idns;
3787 	boolean_t was_ignored;
3788 	int last_rp;
3789 
3790 	ns->ns_nvme = nvme;
3791 
3792 	ASSERT(nvme_mgmt_lock_held(nvme));
3793 
3794 	/*
3795 	 * Because we might rescan a namespace and this will fail after boot
3796 	 * that'd leave us in a bad spot. We need to do something about this
3797 	 * longer term, but it's not clear how exactly we would recover right
3798 	 * now.
3799 	 */
3800 	if (!nvme_identify_int(nvme, nsid, NVME_IDENTIFY_NSID,
3801 	    (void **)&idns)) {
3802 		dev_err(nvme->n_dip, CE_WARN,
3803 		    "!failed to identify namespace %d", nsid);
3804 		return (DDI_FAILURE);
3805 	}
3806 
3807 	if (ns->ns_idns != NULL)
3808 		kmem_free(ns->ns_idns, sizeof (nvme_identify_nsid_t));
3809 
3810 	ns->ns_idns = idns;
3811 	ns->ns_id = nsid;
3812 
3813 	was_ignored = ns->ns_ignore;
3814 
3815 	ns->ns_allocated = nvme_allocated_ns(ns);
3816 	ns->ns_active = nvme_active_ns(ns);
3817 
3818 	ns->ns_block_count = idns->id_nsize;
3819 	ns->ns_block_size =
3820 	    1 << idns->id_lbaf[idns->id_flbas.lba_format].lbaf_lbads;
3821 	ns->ns_best_block_size = ns->ns_block_size;
3822 
3823 	/*
3824 	 * Get the EUI64 if present.
3825 	 */
3826 	if (NVME_VERSION_ATLEAST(&nvme->n_version, 1, 1))
3827 		bcopy(idns->id_eui64, ns->ns_eui64, sizeof (ns->ns_eui64));
3828 
3829 	/*
3830 	 * Get the NGUID if present.
3831 	 */
3832 	if (NVME_VERSION_ATLEAST(&nvme->n_version, 1, 2))
3833 		bcopy(idns->id_nguid, ns->ns_nguid, sizeof (ns->ns_nguid));
3834 
3835 	/*LINTED: E_BAD_PTR_CAST_ALIGN*/
3836 	if (*(uint64_t *)ns->ns_eui64 == 0)
3837 		nvme_prepare_devid(nvme, ns->ns_id);
3838 
3839 	(void) snprintf(ns->ns_name, sizeof (ns->ns_name), "%u", ns->ns_id);
3840 
3841 	/*
3842 	 * Find the LBA format with no metadata and the best relative
3843 	 * performance. A value of 3 means "degraded", 0 is best.
3844 	 */
3845 	last_rp = 3;
3846 	for (int j = 0; j <= idns->id_nlbaf; j++) {
3847 		if (idns->id_lbaf[j].lbaf_lbads == 0)
3848 			break;
3849 		if (idns->id_lbaf[j].lbaf_ms != 0)
3850 			continue;
3851 		if (idns->id_lbaf[j].lbaf_rp >= last_rp)
3852 			continue;
3853 		last_rp = idns->id_lbaf[j].lbaf_rp;
3854 		ns->ns_best_block_size =
3855 		    1 << idns->id_lbaf[j].lbaf_lbads;
3856 	}
3857 
3858 	if (ns->ns_best_block_size < nvme->n_min_block_size)
3859 		ns->ns_best_block_size = nvme->n_min_block_size;
3860 
3861 	was_ignored = ns->ns_ignore;
3862 
3863 	/*
3864 	 * We currently don't support namespaces that are inactive, or use
3865 	 * either:
3866 	 * - protection information
3867 	 * - illegal block size (< 512)
3868 	 */
3869 	if (!ns->ns_active) {
3870 		ns->ns_ignore = B_TRUE;
3871 	} else if (idns->id_dps.dp_pinfo) {
3872 		dev_err(nvme->n_dip, CE_WARN,
3873 		    "!ignoring namespace %d, unsupported feature: "
3874 		    "pinfo = %d", nsid, idns->id_dps.dp_pinfo);
3875 		ns->ns_ignore = B_TRUE;
3876 	} else if (ns->ns_block_size < 512) {
3877 		dev_err(nvme->n_dip, CE_WARN,
3878 		    "!ignoring namespace %d, unsupported block size %"PRIu64,
3879 		    nsid, (uint64_t)ns->ns_block_size);
3880 		ns->ns_ignore = B_TRUE;
3881 	} else {
3882 		ns->ns_ignore = B_FALSE;
3883 	}
3884 
3885 	/*
3886 	 * Keep a count of namespaces which are attachable.
3887 	 * See comments in nvme_bd_driveinfo() to understand its effect.
3888 	 */
3889 	if (was_ignored) {
3890 		/*
3891 		 * Previously ignored, but now not. Count it.
3892 		 */
3893 		if (!ns->ns_ignore)
3894 			nvme->n_namespaces_attachable++;
3895 	} else {
3896 		/*
3897 		 * Wasn't ignored previously, but now needs to be.
3898 		 * Discount it.
3899 		 */
3900 		if (ns->ns_ignore)
3901 			nvme->n_namespaces_attachable--;
3902 	}
3903 
3904 	return (DDI_SUCCESS);
3905 }
3906 
3907 static boolean_t
3908 nvme_attach_ns(nvme_t *nvme, nvme_ioctl_common_t *com)
3909 {
3910 	nvme_namespace_t *ns = nvme_nsid2ns(nvme, com->nioc_nsid);
3911 	int ret;
3912 
3913 	ASSERT(nvme_mgmt_lock_held(nvme));
3914 
3915 	if (ns->ns_ignore) {
3916 		return (nvme_ioctl_error(com, NVME_IOCTL_E_UNSUP_ATTACH_NS,
3917 		    0, 0));
3918 	}
3919 
3920 	if (ns->ns_bd_hdl == NULL) {
3921 		bd_ops_t ops = nvme_bd_ops;
3922 
3923 		if (!nvme->n_idctl->id_oncs.on_dset_mgmt)
3924 			ops.o_free_space = NULL;
3925 
3926 		ns->ns_bd_hdl = bd_alloc_handle(ns, &ops, &nvme->n_prp_dma_attr,
3927 		    KM_SLEEP);
3928 
3929 		if (ns->ns_bd_hdl == NULL) {
3930 			dev_err(nvme->n_dip, CE_WARN, "!Failed to get blkdev "
3931 			    "handle for namespace id %u", com->nioc_nsid);
3932 			return (nvme_ioctl_error(com,
3933 			    NVME_IOCTL_E_BLKDEV_ATTACH, 0, 0));
3934 		}
3935 	}
3936 
3937 	nvme_mgmt_bd_start(nvme);
3938 	ret = bd_attach_handle(nvme->n_dip, ns->ns_bd_hdl);
3939 	nvme_mgmt_bd_end(nvme);
3940 	if (ret != DDI_SUCCESS) {
3941 		return (nvme_ioctl_error(com, NVME_IOCTL_E_BLKDEV_ATTACH,
3942 		    0, 0));
3943 	}
3944 
3945 	ns->ns_attached = B_TRUE;
3946 
3947 	return (B_TRUE);
3948 }
3949 
3950 static boolean_t
3951 nvme_detach_ns(nvme_t *nvme, nvme_ioctl_common_t *com)
3952 {
3953 	nvme_namespace_t *ns = nvme_nsid2ns(nvme, com->nioc_nsid);
3954 	int ret;
3955 
3956 	ASSERT(nvme_mgmt_lock_held(nvme));
3957 
3958 	if (ns->ns_ignore || !ns->ns_attached)
3959 		return (B_TRUE);
3960 
3961 	nvme_mgmt_bd_start(nvme);
3962 	ASSERT3P(ns->ns_bd_hdl, !=, NULL);
3963 	ret = bd_detach_handle(ns->ns_bd_hdl);
3964 	nvme_mgmt_bd_end(nvme);
3965 
3966 	if (ret != DDI_SUCCESS) {
3967 		return (nvme_ioctl_error(com, NVME_IOCTL_E_BLKDEV_DETACH, 0,
3968 		    0));
3969 	}
3970 
3971 	ns->ns_attached = B_FALSE;
3972 	return (B_TRUE);
3973 
3974 }
3975 
3976 /*
3977  * Rescan the namespace information associated with the namespaces indicated by
3978  * ioc. They should not be attached to blkdev right now.
3979  */
3980 static void
3981 nvme_rescan_ns(nvme_t *nvme, uint32_t nsid)
3982 {
3983 	ASSERT(nvme_mgmt_lock_held(nvme));
3984 	ASSERT3U(nsid, !=, 0);
3985 
3986 	if (nsid != NVME_NSID_BCAST) {
3987 		nvme_namespace_t *ns = nvme_nsid2ns(nvme, nsid);
3988 
3989 		ASSERT3U(ns->ns_attached, ==, B_FALSE);
3990 		(void) nvme_init_ns(nvme, nsid);
3991 		return;
3992 	}
3993 
3994 	for (uint32_t i = 1; i <= nvme->n_namespace_count; i++) {
3995 		nvme_namespace_t *ns = nvme_nsid2ns(nvme, i);
3996 
3997 		ASSERT3U(ns->ns_attached, ==, B_FALSE);
3998 		(void) nvme_init_ns(nvme, i);
3999 	}
4000 }
4001 
4002 typedef struct nvme_quirk_table {
4003 	uint16_t nq_vendor_id;
4004 	uint16_t nq_device_id;
4005 	nvme_quirk_t nq_quirks;
4006 } nvme_quirk_table_t;
4007 
4008 static const nvme_quirk_table_t nvme_quirks[] = {
4009 	{ 0x1987, 0x5018, NVME_QUIRK_START_CID },	/* Phison E18 */
4010 };
4011 
4012 static void
4013 nvme_detect_quirks(nvme_t *nvme)
4014 {
4015 	for (uint_t i = 0; i < ARRAY_SIZE(nvme_quirks); i++) {
4016 		const nvme_quirk_table_t *nqt = &nvme_quirks[i];
4017 
4018 		if (nqt->nq_vendor_id == nvme->n_vendor_id &&
4019 		    nqt->nq_device_id == nvme->n_device_id) {
4020 			nvme->n_quirks = nqt->nq_quirks;
4021 			return;
4022 		}
4023 	}
4024 }
4025 
4026 static int
4027 nvme_init(nvme_t *nvme)
4028 {
4029 	nvme_reg_cc_t cc = { 0 };
4030 	nvme_reg_aqa_t aqa = { 0 };
4031 	nvme_reg_asq_t asq = { 0 };
4032 	nvme_reg_acq_t acq = { 0 };
4033 	nvme_reg_cap_t cap;
4034 	nvme_reg_vs_t vs;
4035 	nvme_reg_csts_t csts;
4036 	int i = 0;
4037 	uint16_t nqueues;
4038 	uint_t tq_threads;
4039 	char model[sizeof (nvme->n_idctl->id_model) + 1];
4040 	char *vendor, *product;
4041 	uint32_t nsid;
4042 
4043 	/* Check controller version */
4044 	vs.r = nvme_get32(nvme, NVME_REG_VS);
4045 	nvme->n_version.v_major = vs.b.vs_mjr;
4046 	nvme->n_version.v_minor = vs.b.vs_mnr;
4047 	dev_err(nvme->n_dip, CE_CONT, "?NVMe spec version %d.%d\n",
4048 	    nvme->n_version.v_major, nvme->n_version.v_minor);
4049 
4050 	if (nvme->n_version.v_major > nvme_version_major) {
4051 		dev_err(nvme->n_dip, CE_WARN, "!no support for version > %d.x",
4052 		    nvme_version_major);
4053 		if (nvme->n_strict_version)
4054 			goto fail;
4055 	}
4056 
4057 	/* retrieve controller configuration */
4058 	cap.r = nvme_get64(nvme, NVME_REG_CAP);
4059 
4060 	if ((cap.b.cap_css & NVME_CAP_CSS_NVM) == 0) {
4061 		dev_err(nvme->n_dip, CE_WARN,
4062 		    "!NVM command set not supported by hardware");
4063 		goto fail;
4064 	}
4065 
4066 	nvme->n_nssr_supported = cap.b.cap_nssrs;
4067 	nvme->n_doorbell_stride = 4 << cap.b.cap_dstrd;
4068 	nvme->n_timeout = cap.b.cap_to;
4069 	nvme->n_arbitration_mechanisms = cap.b.cap_ams;
4070 	nvme->n_cont_queues_reqd = cap.b.cap_cqr;
4071 	nvme->n_max_queue_entries = cap.b.cap_mqes + 1;
4072 
4073 	/*
4074 	 * The MPSMIN and MPSMAX fields in the CAP register use 0 to specify
4075 	 * the base page size of 4k (1<<12), so add 12 here to get the real
4076 	 * page size value.
4077 	 */
4078 	nvme->n_pageshift = MIN(MAX(cap.b.cap_mpsmin + 12, PAGESHIFT),
4079 	    cap.b.cap_mpsmax + 12);
4080 	nvme->n_pagesize = 1UL << (nvme->n_pageshift);
4081 
4082 	/*
4083 	 * Set up Queue DMA to transfer at least 1 page-aligned page at a time.
4084 	 */
4085 	nvme->n_queue_dma_attr.dma_attr_align = nvme->n_pagesize;
4086 	nvme->n_queue_dma_attr.dma_attr_minxfer = nvme->n_pagesize;
4087 
4088 	/*
4089 	 * Set up PRP DMA to transfer 1 page-aligned page at a time.
4090 	 * Maxxfer may be increased after we identified the controller limits.
4091 	 */
4092 	nvme->n_prp_dma_attr.dma_attr_maxxfer = nvme->n_pagesize;
4093 	nvme->n_prp_dma_attr.dma_attr_minxfer = nvme->n_pagesize;
4094 	nvme->n_prp_dma_attr.dma_attr_align = nvme->n_pagesize;
4095 	nvme->n_prp_dma_attr.dma_attr_seg = nvme->n_pagesize - 1;
4096 
4097 	/*
4098 	 * Reset controller if it's still in ready state.
4099 	 */
4100 	if (nvme_reset(nvme, B_FALSE) == B_FALSE) {
4101 		dev_err(nvme->n_dip, CE_WARN, "!unable to reset controller");
4102 		ddi_fm_service_impact(nvme->n_dip, DDI_SERVICE_LOST);
4103 		nvme->n_dead = B_TRUE;
4104 		goto fail;
4105 	}
4106 
4107 	/*
4108 	 * Create the cq array with one completion queue to be assigned
4109 	 * to the admin queue pair and a limited number of taskqs (4).
4110 	 */
4111 	if (nvme_create_cq_array(nvme, 1, nvme->n_admin_queue_len, 4) !=
4112 	    DDI_SUCCESS) {
4113 		dev_err(nvme->n_dip, CE_WARN,
4114 		    "!failed to pre-allocate admin completion queue");
4115 		goto fail;
4116 	}
4117 	/*
4118 	 * Create the admin queue pair.
4119 	 */
4120 	if (nvme_alloc_qpair(nvme, nvme->n_admin_queue_len, &nvme->n_adminq, 0)
4121 	    != DDI_SUCCESS) {
4122 		dev_err(nvme->n_dip, CE_WARN,
4123 		    "!unable to allocate admin qpair");
4124 		goto fail;
4125 	}
4126 	nvme->n_ioq = kmem_alloc(sizeof (nvme_qpair_t *), KM_SLEEP);
4127 	nvme->n_ioq[0] = nvme->n_adminq;
4128 
4129 	if (nvme->n_quirks & NVME_QUIRK_START_CID)
4130 		nvme->n_adminq->nq_next_cmd++;
4131 
4132 	nvme->n_progress |= NVME_ADMIN_QUEUE;
4133 
4134 	(void) ddi_prop_update_int(DDI_DEV_T_NONE, nvme->n_dip,
4135 	    "admin-queue-len", nvme->n_admin_queue_len);
4136 
4137 	aqa.b.aqa_asqs = aqa.b.aqa_acqs = nvme->n_admin_queue_len - 1;
4138 	asq = nvme->n_adminq->nq_sqdma->nd_cookie.dmac_laddress;
4139 	acq = nvme->n_adminq->nq_cq->ncq_dma->nd_cookie.dmac_laddress;
4140 
4141 	ASSERT((asq & (nvme->n_pagesize - 1)) == 0);
4142 	ASSERT((acq & (nvme->n_pagesize - 1)) == 0);
4143 
4144 	nvme_put32(nvme, NVME_REG_AQA, aqa.r);
4145 	nvme_put64(nvme, NVME_REG_ASQ, asq);
4146 	nvme_put64(nvme, NVME_REG_ACQ, acq);
4147 
4148 	cc.b.cc_ams = 0;	/* use Round-Robin arbitration */
4149 	cc.b.cc_css = 0;	/* use NVM command set */
4150 	cc.b.cc_mps = nvme->n_pageshift - 12;
4151 	cc.b.cc_shn = 0;	/* no shutdown in progress */
4152 	cc.b.cc_en = 1;		/* enable controller */
4153 	cc.b.cc_iosqes = 6;	/* submission queue entry is 2^6 bytes long */
4154 	cc.b.cc_iocqes = 4;	/* completion queue entry is 2^4 bytes long */
4155 
4156 	nvme_put32(nvme, NVME_REG_CC, cc.r);
4157 
4158 	/*
4159 	 * Wait for the controller to become ready.
4160 	 */
4161 	csts.r = nvme_get32(nvme, NVME_REG_CSTS);
4162 	if (csts.b.csts_rdy == 0) {
4163 		for (i = 0; i != nvme->n_timeout * 10; i++) {
4164 			delay(drv_usectohz(50000));
4165 			csts.r = nvme_get32(nvme, NVME_REG_CSTS);
4166 
4167 			if (csts.b.csts_cfs == 1) {
4168 				dev_err(nvme->n_dip, CE_WARN,
4169 				    "!controller fatal status at init");
4170 				ddi_fm_service_impact(nvme->n_dip,
4171 				    DDI_SERVICE_LOST);
4172 				nvme->n_dead = B_TRUE;
4173 				goto fail;
4174 			}
4175 
4176 			if (csts.b.csts_rdy == 1)
4177 				break;
4178 		}
4179 	}
4180 
4181 	if (csts.b.csts_rdy == 0) {
4182 		dev_err(nvme->n_dip, CE_WARN, "!controller not ready");
4183 		ddi_fm_service_impact(nvme->n_dip, DDI_SERVICE_LOST);
4184 		nvme->n_dead = B_TRUE;
4185 		goto fail;
4186 	}
4187 
4188 	/*
4189 	 * Assume an abort command limit of 1. We'll destroy and re-init
4190 	 * that later when we know the true abort command limit.
4191 	 */
4192 	sema_init(&nvme->n_abort_sema, 1, NULL, SEMA_DRIVER, NULL);
4193 
4194 	/*
4195 	 * Set up initial interrupt for admin queue.
4196 	 */
4197 	if ((nvme_setup_interrupts(nvme, DDI_INTR_TYPE_MSIX, 1)
4198 	    != DDI_SUCCESS) &&
4199 	    (nvme_setup_interrupts(nvme, DDI_INTR_TYPE_MSI, 1)
4200 	    != DDI_SUCCESS) &&
4201 	    (nvme_setup_interrupts(nvme, DDI_INTR_TYPE_FIXED, 1)
4202 	    != DDI_SUCCESS)) {
4203 		dev_err(nvme->n_dip, CE_WARN,
4204 		    "!failed to set up initial interrupt");
4205 		goto fail;
4206 	}
4207 
4208 	/*
4209 	 * Post an asynchronous event command to catch errors.
4210 	 * We assume the asynchronous events are supported as required by
4211 	 * specification (Figure 40 in section 5 of NVMe 1.2).
4212 	 * However, since at least qemu does not follow the specification,
4213 	 * we need a mechanism to protect ourselves.
4214 	 */
4215 	nvme->n_async_event_supported = B_TRUE;
4216 	nvme_async_event(nvme);
4217 
4218 	/*
4219 	 * Identify Controller
4220 	 */
4221 	if (!nvme_identify_int(nvme, 0, NVME_IDENTIFY_CTRL,
4222 	    (void **)&nvme->n_idctl)) {
4223 		dev_err(nvme->n_dip, CE_WARN, "!failed to identify controller");
4224 		goto fail;
4225 	}
4226 
4227 	/*
4228 	 * Get the common namespace information if available. If not, we use the
4229 	 * information for nsid 1.
4230 	 */
4231 	if (nvme_ctrl_atleast(nvme, &nvme_vers_1v2) &&
4232 	    nvme->n_idctl->id_oacs.oa_nsmgmt != 0) {
4233 		nsid = NVME_NSID_BCAST;
4234 	} else {
4235 		nsid = 1;
4236 	}
4237 
4238 	if (!nvme_identify_int(nvme, nsid, NVME_IDENTIFY_NSID,
4239 	    (void **)&nvme->n_idcomns)) {
4240 		dev_err(nvme->n_dip, CE_WARN, "!failed to identify common "
4241 		    "namespace information");
4242 		goto fail;
4243 	}
4244 	/*
4245 	 * Process nvme-config-list (if present) in nvme.conf.
4246 	 */
4247 	nvme_config_list(nvme);
4248 
4249 	/*
4250 	 * Get Vendor & Product ID
4251 	 */
4252 	bcopy(nvme->n_idctl->id_model, model, sizeof (nvme->n_idctl->id_model));
4253 	model[sizeof (nvme->n_idctl->id_model)] = '\0';
4254 	sata_split_model(model, &vendor, &product);
4255 
4256 	if (vendor == NULL)
4257 		nvme->n_vendor = strdup("NVMe");
4258 	else
4259 		nvme->n_vendor = strdup(vendor);
4260 
4261 	nvme->n_product = strdup(product);
4262 
4263 	/*
4264 	 * Get controller limits.
4265 	 */
4266 	nvme->n_async_event_limit = MAX(NVME_MIN_ASYNC_EVENT_LIMIT,
4267 	    MIN(nvme->n_admin_queue_len / 10,
4268 	    MIN(nvme->n_idctl->id_aerl + 1, nvme->n_async_event_limit)));
4269 
4270 	(void) ddi_prop_update_int(DDI_DEV_T_NONE, nvme->n_dip,
4271 	    "async-event-limit", nvme->n_async_event_limit);
4272 
4273 	nvme->n_abort_command_limit = nvme->n_idctl->id_acl + 1;
4274 
4275 	/*
4276 	 * Reinitialize the semaphore with the true abort command limit
4277 	 * supported by the hardware. It's not necessary to disable interrupts
4278 	 * as only command aborts use the semaphore, and no commands are
4279 	 * executed or aborted while we're here.
4280 	 */
4281 	sema_destroy(&nvme->n_abort_sema);
4282 	sema_init(&nvme->n_abort_sema, nvme->n_abort_command_limit - 1, NULL,
4283 	    SEMA_DRIVER, NULL);
4284 
4285 	nvme->n_progress |= NVME_CTRL_LIMITS;
4286 
4287 	if (nvme->n_idctl->id_mdts == 0)
4288 		nvme->n_max_data_transfer_size = nvme->n_pagesize * 65536;
4289 	else
4290 		nvme->n_max_data_transfer_size =
4291 		    1ull << (nvme->n_pageshift + nvme->n_idctl->id_mdts);
4292 
4293 	nvme->n_error_log_len = nvme->n_idctl->id_elpe + 1;
4294 
4295 	/*
4296 	 * Limit n_max_data_transfer_size to what we can handle in one PRP.
4297 	 * Chained PRPs are currently unsupported.
4298 	 *
4299 	 * This is a no-op on hardware which doesn't support a transfer size
4300 	 * big enough to require chained PRPs.
4301 	 */
4302 	nvme->n_max_data_transfer_size = MIN(nvme->n_max_data_transfer_size,
4303 	    (nvme->n_pagesize / sizeof (uint64_t) * nvme->n_pagesize));
4304 
4305 	nvme->n_prp_dma_attr.dma_attr_maxxfer = nvme->n_max_data_transfer_size;
4306 
4307 	/*
4308 	 * Make sure the minimum/maximum queue entry sizes are not
4309 	 * larger/smaller than the default.
4310 	 */
4311 
4312 	if (((1 << nvme->n_idctl->id_sqes.qes_min) > sizeof (nvme_sqe_t)) ||
4313 	    ((1 << nvme->n_idctl->id_sqes.qes_max) < sizeof (nvme_sqe_t)) ||
4314 	    ((1 << nvme->n_idctl->id_cqes.qes_min) > sizeof (nvme_cqe_t)) ||
4315 	    ((1 << nvme->n_idctl->id_cqes.qes_max) < sizeof (nvme_cqe_t)))
4316 		goto fail;
4317 
4318 	/*
4319 	 * Check for the presence of a Volatile Write Cache. If present,
4320 	 * enable or disable based on the value of the property
4321 	 * volatile-write-cache-enable (default is enabled).
4322 	 */
4323 	nvme->n_write_cache_present =
4324 	    nvme->n_idctl->id_vwc.vwc_present == 0 ? B_FALSE : B_TRUE;
4325 
4326 	(void) ddi_prop_update_int(DDI_DEV_T_NONE, nvme->n_dip,
4327 	    "volatile-write-cache-present",
4328 	    nvme->n_write_cache_present ? 1 : 0);
4329 
4330 	if (!nvme->n_write_cache_present) {
4331 		nvme->n_write_cache_enabled = B_FALSE;
4332 	} else if (nvme_write_cache_set(nvme, nvme->n_write_cache_enabled)
4333 	    != 0) {
4334 		dev_err(nvme->n_dip, CE_WARN,
4335 		    "!failed to %sable volatile write cache",
4336 		    nvme->n_write_cache_enabled ? "en" : "dis");
4337 		/*
4338 		 * Assume the cache is (still) enabled.
4339 		 */
4340 		nvme->n_write_cache_enabled = B_TRUE;
4341 	}
4342 
4343 	(void) ddi_prop_update_int(DDI_DEV_T_NONE, nvme->n_dip,
4344 	    "volatile-write-cache-enable",
4345 	    nvme->n_write_cache_enabled ? 1 : 0);
4346 
4347 	/*
4348 	 * Get number of supported namespaces and allocate namespace array.
4349 	 */
4350 	nvme->n_namespace_count = nvme->n_idctl->id_nn;
4351 
4352 	if (nvme->n_namespace_count == 0) {
4353 		dev_err(nvme->n_dip, CE_WARN,
4354 		    "!controllers without namespaces are not supported");
4355 		goto fail;
4356 	}
4357 
4358 	if (nvme->n_namespace_count > NVME_MINOR_MAX) {
4359 		dev_err(nvme->n_dip, CE_WARN,
4360 		    "!too many namespaces: %d, limiting to %d\n",
4361 		    nvme->n_namespace_count, NVME_MINOR_MAX);
4362 		nvme->n_namespace_count = NVME_MINOR_MAX;
4363 	}
4364 
4365 	nvme->n_ns = kmem_zalloc(sizeof (nvme_namespace_t) *
4366 	    nvme->n_namespace_count, KM_SLEEP);
4367 
4368 	/*
4369 	 * Try to set up MSI/MSI-X interrupts.
4370 	 */
4371 	if ((nvme->n_intr_types & (DDI_INTR_TYPE_MSI | DDI_INTR_TYPE_MSIX))
4372 	    != 0) {
4373 		nvme_release_interrupts(nvme);
4374 
4375 		nqueues = MIN(UINT16_MAX, ncpus);
4376 
4377 		if ((nvme_setup_interrupts(nvme, DDI_INTR_TYPE_MSIX,
4378 		    nqueues) != DDI_SUCCESS) &&
4379 		    (nvme_setup_interrupts(nvme, DDI_INTR_TYPE_MSI,
4380 		    nqueues) != DDI_SUCCESS)) {
4381 			dev_err(nvme->n_dip, CE_WARN,
4382 			    "!failed to set up MSI/MSI-X interrupts");
4383 			goto fail;
4384 		}
4385 	}
4386 
4387 	/*
4388 	 * Create I/O queue pairs.
4389 	 */
4390 
4391 	if (nvme_set_nqueues(nvme) != 0) {
4392 		dev_err(nvme->n_dip, CE_WARN,
4393 		    "!failed to set number of I/O queues to %d",
4394 		    nvme->n_intr_cnt);
4395 		goto fail;
4396 	}
4397 
4398 	/*
4399 	 * Reallocate I/O queue array
4400 	 */
4401 	kmem_free(nvme->n_ioq, sizeof (nvme_qpair_t *));
4402 	nvme->n_ioq = kmem_zalloc(sizeof (nvme_qpair_t *) *
4403 	    (nvme->n_submission_queues + 1), KM_SLEEP);
4404 	nvme->n_ioq[0] = nvme->n_adminq;
4405 
4406 	/*
4407 	 * There should always be at least as many submission queues
4408 	 * as completion queues.
4409 	 */
4410 	ASSERT(nvme->n_submission_queues >= nvme->n_completion_queues);
4411 
4412 	nvme->n_ioq_count = nvme->n_submission_queues;
4413 
4414 	nvme->n_io_squeue_len =
4415 	    MIN(nvme->n_io_squeue_len, nvme->n_max_queue_entries);
4416 
4417 	(void) ddi_prop_update_int(DDI_DEV_T_NONE, nvme->n_dip, "io-squeue-len",
4418 	    nvme->n_io_squeue_len);
4419 
4420 	/*
4421 	 * Pre-allocate completion queues.
4422 	 * When there are the same number of submission and completion
4423 	 * queues there is no value in having a larger completion
4424 	 * queue length.
4425 	 */
4426 	if (nvme->n_submission_queues == nvme->n_completion_queues)
4427 		nvme->n_io_cqueue_len = MIN(nvme->n_io_cqueue_len,
4428 		    nvme->n_io_squeue_len);
4429 
4430 	nvme->n_io_cqueue_len = MIN(nvme->n_io_cqueue_len,
4431 	    nvme->n_max_queue_entries);
4432 
4433 	(void) ddi_prop_update_int(DDI_DEV_T_NONE, nvme->n_dip, "io-cqueue-len",
4434 	    nvme->n_io_cqueue_len);
4435 
4436 	/*
4437 	 * Assign the equal quantity of taskq threads to each completion
4438 	 * queue, capping the total number of threads to the number
4439 	 * of CPUs.
4440 	 */
4441 	tq_threads = MIN(UINT16_MAX, ncpus) / nvme->n_completion_queues;
4442 
4443 	/*
4444 	 * In case the calculation above is zero, we need at least one
4445 	 * thread per completion queue.
4446 	 */
4447 	tq_threads = MAX(1, tq_threads);
4448 
4449 	if (nvme_create_cq_array(nvme, nvme->n_completion_queues + 1,
4450 	    nvme->n_io_cqueue_len, tq_threads) != DDI_SUCCESS) {
4451 		dev_err(nvme->n_dip, CE_WARN,
4452 		    "!failed to pre-allocate completion queues");
4453 		goto fail;
4454 	}
4455 
4456 	/*
4457 	 * If we use less completion queues than interrupt vectors return
4458 	 * some of the interrupt vectors back to the system.
4459 	 */
4460 	if (nvme->n_completion_queues + 1 < nvme->n_intr_cnt) {
4461 		nvme_release_interrupts(nvme);
4462 
4463 		if (nvme_setup_interrupts(nvme, nvme->n_intr_type,
4464 		    nvme->n_completion_queues + 1) != DDI_SUCCESS) {
4465 			dev_err(nvme->n_dip, CE_WARN,
4466 			    "!failed to reduce number of interrupts");
4467 			goto fail;
4468 		}
4469 	}
4470 
4471 	/*
4472 	 * Alloc & register I/O queue pairs
4473 	 */
4474 
4475 	for (i = 1; i != nvme->n_ioq_count + 1; i++) {
4476 		if (nvme_alloc_qpair(nvme, nvme->n_io_squeue_len,
4477 		    &nvme->n_ioq[i], i) != DDI_SUCCESS) {
4478 			dev_err(nvme->n_dip, CE_WARN,
4479 			    "!unable to allocate I/O qpair %d", i);
4480 			goto fail;
4481 		}
4482 
4483 		if (nvme_create_io_qpair(nvme, nvme->n_ioq[i], i) != 0) {
4484 			dev_err(nvme->n_dip, CE_WARN,
4485 			    "!unable to create I/O qpair %d", i);
4486 			goto fail;
4487 		}
4488 	}
4489 
4490 	/*
4491 	 * Post more asynchronous events commands to reduce event reporting
4492 	 * latency as suggested by the spec.
4493 	 */
4494 	if (nvme->n_async_event_supported) {
4495 		for (i = 1; i != nvme->n_async_event_limit; i++)
4496 			nvme_async_event(nvme);
4497 	}
4498 
4499 	return (DDI_SUCCESS);
4500 
4501 fail:
4502 	(void) nvme_reset(nvme, B_FALSE);
4503 	return (DDI_FAILURE);
4504 }
4505 
4506 static uint_t
4507 nvme_intr(caddr_t arg1, caddr_t arg2)
4508 {
4509 	nvme_t *nvme = (nvme_t *)arg1;
4510 	int inum = (int)(uintptr_t)arg2;
4511 	int ccnt = 0;
4512 	int qnum;
4513 
4514 	if (inum >= nvme->n_intr_cnt)
4515 		return (DDI_INTR_UNCLAIMED);
4516 
4517 	if (nvme->n_dead) {
4518 		return (nvme->n_intr_type == DDI_INTR_TYPE_FIXED ?
4519 		    DDI_INTR_UNCLAIMED : DDI_INTR_CLAIMED);
4520 	}
4521 
4522 	/*
4523 	 * The interrupt vector a queue uses is calculated as queue_idx %
4524 	 * intr_cnt in nvme_create_io_qpair(). Iterate through the queue array
4525 	 * in steps of n_intr_cnt to process all queues using this vector.
4526 	 */
4527 	for (qnum = inum;
4528 	    qnum < nvme->n_cq_count && nvme->n_cq[qnum] != NULL;
4529 	    qnum += nvme->n_intr_cnt) {
4530 		ccnt += nvme_process_iocq(nvme, nvme->n_cq[qnum]);
4531 	}
4532 
4533 	return (ccnt > 0 ? DDI_INTR_CLAIMED : DDI_INTR_UNCLAIMED);
4534 }
4535 
4536 static void
4537 nvme_release_interrupts(nvme_t *nvme)
4538 {
4539 	int i;
4540 
4541 	for (i = 0; i < nvme->n_intr_cnt; i++) {
4542 		if (nvme->n_inth[i] == NULL)
4543 			break;
4544 
4545 		if (nvme->n_intr_cap & DDI_INTR_FLAG_BLOCK)
4546 			(void) ddi_intr_block_disable(&nvme->n_inth[i], 1);
4547 		else
4548 			(void) ddi_intr_disable(nvme->n_inth[i]);
4549 
4550 		(void) ddi_intr_remove_handler(nvme->n_inth[i]);
4551 		(void) ddi_intr_free(nvme->n_inth[i]);
4552 	}
4553 
4554 	kmem_free(nvme->n_inth, nvme->n_inth_sz);
4555 	nvme->n_inth = NULL;
4556 	nvme->n_inth_sz = 0;
4557 
4558 	nvme->n_progress &= ~NVME_INTERRUPTS;
4559 }
4560 
4561 static int
4562 nvme_setup_interrupts(nvme_t *nvme, int intr_type, int nqpairs)
4563 {
4564 	int nintrs, navail, count;
4565 	int ret;
4566 	int i;
4567 
4568 	if (nvme->n_intr_types == 0) {
4569 		ret = ddi_intr_get_supported_types(nvme->n_dip,
4570 		    &nvme->n_intr_types);
4571 		if (ret != DDI_SUCCESS) {
4572 			dev_err(nvme->n_dip, CE_WARN,
4573 			    "!%s: ddi_intr_get_supported types failed",
4574 			    __func__);
4575 			return (ret);
4576 		}
4577 #ifdef __x86
4578 		if (get_hwenv() == HW_VMWARE)
4579 			nvme->n_intr_types &= ~DDI_INTR_TYPE_MSIX;
4580 #endif
4581 	}
4582 
4583 	if ((nvme->n_intr_types & intr_type) == 0)
4584 		return (DDI_FAILURE);
4585 
4586 	ret = ddi_intr_get_nintrs(nvme->n_dip, intr_type, &nintrs);
4587 	if (ret != DDI_SUCCESS) {
4588 		dev_err(nvme->n_dip, CE_WARN, "!%s: ddi_intr_get_nintrs failed",
4589 		    __func__);
4590 		return (ret);
4591 	}
4592 
4593 	ret = ddi_intr_get_navail(nvme->n_dip, intr_type, &navail);
4594 	if (ret != DDI_SUCCESS) {
4595 		dev_err(nvme->n_dip, CE_WARN, "!%s: ddi_intr_get_navail failed",
4596 		    __func__);
4597 		return (ret);
4598 	}
4599 
4600 	/* We want at most one interrupt per queue pair. */
4601 	if (navail > nqpairs)
4602 		navail = nqpairs;
4603 
4604 	nvme->n_inth_sz = sizeof (ddi_intr_handle_t) * navail;
4605 	nvme->n_inth = kmem_zalloc(nvme->n_inth_sz, KM_SLEEP);
4606 
4607 	ret = ddi_intr_alloc(nvme->n_dip, nvme->n_inth, intr_type, 0, navail,
4608 	    &count, 0);
4609 	if (ret != DDI_SUCCESS) {
4610 		dev_err(nvme->n_dip, CE_WARN, "!%s: ddi_intr_alloc failed",
4611 		    __func__);
4612 		goto fail;
4613 	}
4614 
4615 	nvme->n_intr_cnt = count;
4616 
4617 	ret = ddi_intr_get_pri(nvme->n_inth[0], &nvme->n_intr_pri);
4618 	if (ret != DDI_SUCCESS) {
4619 		dev_err(nvme->n_dip, CE_WARN, "!%s: ddi_intr_get_pri failed",
4620 		    __func__);
4621 		goto fail;
4622 	}
4623 
4624 	for (i = 0; i < count; i++) {
4625 		ret = ddi_intr_add_handler(nvme->n_inth[i], nvme_intr,
4626 		    (void *)nvme, (void *)(uintptr_t)i);
4627 		if (ret != DDI_SUCCESS) {
4628 			dev_err(nvme->n_dip, CE_WARN,
4629 			    "!%s: ddi_intr_add_handler failed", __func__);
4630 			goto fail;
4631 		}
4632 	}
4633 
4634 	(void) ddi_intr_get_cap(nvme->n_inth[0], &nvme->n_intr_cap);
4635 
4636 	for (i = 0; i < count; i++) {
4637 		if (nvme->n_intr_cap & DDI_INTR_FLAG_BLOCK)
4638 			ret = ddi_intr_block_enable(&nvme->n_inth[i], 1);
4639 		else
4640 			ret = ddi_intr_enable(nvme->n_inth[i]);
4641 
4642 		if (ret != DDI_SUCCESS) {
4643 			dev_err(nvme->n_dip, CE_WARN,
4644 			    "!%s: enabling interrupt %d failed", __func__, i);
4645 			goto fail;
4646 		}
4647 	}
4648 
4649 	nvme->n_intr_type = intr_type;
4650 
4651 	nvme->n_progress |= NVME_INTERRUPTS;
4652 
4653 	return (DDI_SUCCESS);
4654 
4655 fail:
4656 	nvme_release_interrupts(nvme);
4657 
4658 	return (ret);
4659 }
4660 
4661 static int
4662 nvme_fm_errcb(dev_info_t *dip, ddi_fm_error_t *fm_error, const void *arg)
4663 {
4664 	_NOTE(ARGUNUSED(arg));
4665 
4666 	pci_ereport_post(dip, fm_error, NULL);
4667 	return (fm_error->fme_status);
4668 }
4669 
4670 static void
4671 nvme_remove_callback(dev_info_t *dip, ddi_eventcookie_t cookie, void *a,
4672     void *b)
4673 {
4674 	nvme_t *nvme = a;
4675 
4676 	nvme_ctrl_mark_dead(nvme, B_TRUE);
4677 
4678 	/*
4679 	 * Fail all outstanding commands, including those in the admin queue
4680 	 * (queue 0).
4681 	 */
4682 	for (uint_t i = 0; i < nvme->n_ioq_count + 1; i++) {
4683 		nvme_qpair_t *qp = nvme->n_ioq[i];
4684 
4685 		mutex_enter(&qp->nq_mutex);
4686 		for (size_t j = 0; j < qp->nq_nentry; j++) {
4687 			nvme_cmd_t *cmd = qp->nq_cmd[j];
4688 			nvme_cmd_t *u_cmd;
4689 
4690 			if (cmd == NULL) {
4691 				continue;
4692 			}
4693 
4694 			/*
4695 			 * Since we have the queue lock held the entire time we
4696 			 * iterate over it, it's not possible for the queue to
4697 			 * change underneath us. Thus, we don't need to check
4698 			 * that the return value of nvme_unqueue_cmd matches the
4699 			 * requested cmd to unqueue.
4700 			 */
4701 			u_cmd = nvme_unqueue_cmd(nvme, qp, cmd->nc_sqe.sqe_cid);
4702 			taskq_dispatch_ent(qp->nq_cq->ncq_cmd_taskq,
4703 			    cmd->nc_callback, cmd, TQ_NOSLEEP, &cmd->nc_tqent);
4704 
4705 			ASSERT3P(u_cmd, ==, cmd);
4706 		}
4707 		mutex_exit(&qp->nq_mutex);
4708 	}
4709 }
4710 
4711 /*
4712  * Open minor management
4713  */
4714 static int
4715 nvme_minor_comparator(const void *l, const void *r)
4716 {
4717 	const nvme_minor_t *lm = l;
4718 	const nvme_minor_t *rm = r;
4719 
4720 	if (lm->nm_minor > rm->nm_minor) {
4721 		return (1);
4722 	} else if (lm->nm_minor < rm->nm_minor) {
4723 		return (-1);
4724 	} else {
4725 		return (0);
4726 	}
4727 }
4728 
4729 static void
4730 nvme_minor_free(nvme_minor_t *minor)
4731 {
4732 	if (minor->nm_minor > 0) {
4733 		ASSERT3S(minor->nm_minor, >=, NVME_OPEN_MINOR_MIN);
4734 		id_free(nvme_open_minors, minor->nm_minor);
4735 		minor->nm_minor = 0;
4736 	}
4737 	VERIFY0(list_link_active(&minor->nm_ctrl_lock.nli_node));
4738 	VERIFY0(list_link_active(&minor->nm_ns_lock.nli_node));
4739 	cv_destroy(&minor->nm_cv);
4740 	kmem_free(minor, sizeof (nvme_minor_t));
4741 }
4742 
4743 static nvme_minor_t *
4744 nvme_minor_find_by_dev(dev_t dev)
4745 {
4746 	id_t id = (id_t)getminor(dev);
4747 	nvme_minor_t search = { .nm_minor = id };
4748 	nvme_minor_t *ret;
4749 
4750 	mutex_enter(&nvme_open_minors_mutex);
4751 	ret = avl_find(&nvme_open_minors_avl, &search, NULL);
4752 	mutex_exit(&nvme_open_minors_mutex);
4753 
4754 	return (ret);
4755 }
4756 
4757 static int
4758 nvme_attach(dev_info_t *dip, ddi_attach_cmd_t cmd)
4759 {
4760 	nvme_t *nvme;
4761 	int instance;
4762 	int nregs;
4763 	off_t regsize;
4764 	char name[32];
4765 	boolean_t attached_ns;
4766 
4767 	if (cmd != DDI_ATTACH)
4768 		return (DDI_FAILURE);
4769 
4770 	instance = ddi_get_instance(dip);
4771 
4772 	if (ddi_soft_state_zalloc(nvme_state, instance) != DDI_SUCCESS)
4773 		return (DDI_FAILURE);
4774 
4775 	nvme = ddi_get_soft_state(nvme_state, instance);
4776 	ddi_set_driver_private(dip, nvme);
4777 	nvme->n_dip = dip;
4778 
4779 	/*
4780 	 * Map PCI config space
4781 	 */
4782 	if (pci_config_setup(dip, &nvme->n_pcicfg_handle) != DDI_SUCCESS) {
4783 		dev_err(dip, CE_WARN, "!failed to map PCI config space");
4784 		goto fail;
4785 	}
4786 	nvme->n_progress |= NVME_PCI_CONFIG;
4787 
4788 	/*
4789 	 * Get the various PCI IDs from config space
4790 	 */
4791 	nvme->n_vendor_id =
4792 	    pci_config_get16(nvme->n_pcicfg_handle, PCI_CONF_VENID);
4793 	nvme->n_device_id =
4794 	    pci_config_get16(nvme->n_pcicfg_handle, PCI_CONF_DEVID);
4795 	nvme->n_revision_id =
4796 	    pci_config_get8(nvme->n_pcicfg_handle, PCI_CONF_REVID);
4797 	nvme->n_subsystem_device_id =
4798 	    pci_config_get16(nvme->n_pcicfg_handle, PCI_CONF_SUBSYSID);
4799 	nvme->n_subsystem_vendor_id =
4800 	    pci_config_get16(nvme->n_pcicfg_handle, PCI_CONF_SUBVENID);
4801 
4802 	nvme_detect_quirks(nvme);
4803 
4804 	/*
4805 	 * Set up event handlers for hot removal. While npe(4D) supports the hot
4806 	 * removal event being injected for devices, the same is not true of all
4807 	 * of our possible parents (i.e. pci(4D) as of this writing). The most
4808 	 * common case this shows up is in some virtualization environments. We
4809 	 * should treat this as non-fatal so that way devices work but leave
4810 	 * this set up in such a way that if a nexus does grow support for this
4811 	 * we're good to go.
4812 	 */
4813 	if (ddi_get_eventcookie(nvme->n_dip, DDI_DEVI_REMOVE_EVENT,
4814 	    &nvme->n_rm_cookie) == DDI_SUCCESS) {
4815 		if (ddi_add_event_handler(nvme->n_dip, nvme->n_rm_cookie,
4816 		    nvme_remove_callback, nvme, &nvme->n_ev_rm_cb_id) !=
4817 		    DDI_SUCCESS) {
4818 			goto fail;
4819 		}
4820 	} else {
4821 		nvme->n_ev_rm_cb_id = NULL;
4822 	}
4823 
4824 	mutex_init(&nvme->n_minor_mutex, NULL, MUTEX_DRIVER, NULL);
4825 	nvme->n_progress |= NVME_MUTEX_INIT;
4826 
4827 	nvme->n_strict_version = ddi_prop_get_int(DDI_DEV_T_ANY, dip,
4828 	    DDI_PROP_DONTPASS, "strict-version", 1) == 1 ? B_TRUE : B_FALSE;
4829 	nvme->n_ignore_unknown_vendor_status = ddi_prop_get_int(DDI_DEV_T_ANY,
4830 	    dip, DDI_PROP_DONTPASS, "ignore-unknown-vendor-status", 0) == 1 ?
4831 	    B_TRUE : B_FALSE;
4832 	nvme->n_admin_queue_len = ddi_prop_get_int(DDI_DEV_T_ANY, dip,
4833 	    DDI_PROP_DONTPASS, "admin-queue-len", NVME_DEFAULT_ADMIN_QUEUE_LEN);
4834 	nvme->n_io_squeue_len = ddi_prop_get_int(DDI_DEV_T_ANY, dip,
4835 	    DDI_PROP_DONTPASS, "io-squeue-len", NVME_DEFAULT_IO_QUEUE_LEN);
4836 	/*
4837 	 * Double up the default for completion queues in case of
4838 	 * queue sharing.
4839 	 */
4840 	nvme->n_io_cqueue_len = ddi_prop_get_int(DDI_DEV_T_ANY, dip,
4841 	    DDI_PROP_DONTPASS, "io-cqueue-len", 2 * NVME_DEFAULT_IO_QUEUE_LEN);
4842 	nvme->n_async_event_limit = ddi_prop_get_int(DDI_DEV_T_ANY, dip,
4843 	    DDI_PROP_DONTPASS, "async-event-limit",
4844 	    NVME_DEFAULT_ASYNC_EVENT_LIMIT);
4845 	nvme->n_write_cache_enabled = ddi_prop_get_int(DDI_DEV_T_ANY, dip,
4846 	    DDI_PROP_DONTPASS, "volatile-write-cache-enable", 1) != 0 ?
4847 	    B_TRUE : B_FALSE;
4848 	nvme->n_min_block_size = ddi_prop_get_int(DDI_DEV_T_ANY, dip,
4849 	    DDI_PROP_DONTPASS, "min-phys-block-size",
4850 	    NVME_DEFAULT_MIN_BLOCK_SIZE);
4851 	nvme->n_submission_queues = ddi_prop_get_int(DDI_DEV_T_ANY, dip,
4852 	    DDI_PROP_DONTPASS, "max-submission-queues", -1);
4853 	nvme->n_completion_queues = ddi_prop_get_int(DDI_DEV_T_ANY, dip,
4854 	    DDI_PROP_DONTPASS, "max-completion-queues", -1);
4855 
4856 	if (!ISP2(nvme->n_min_block_size) ||
4857 	    (nvme->n_min_block_size < NVME_DEFAULT_MIN_BLOCK_SIZE)) {
4858 		dev_err(dip, CE_WARN, "!min-phys-block-size %s, "
4859 		    "using default %d", ISP2(nvme->n_min_block_size) ?
4860 		    "too low" : "not a power of 2",
4861 		    NVME_DEFAULT_MIN_BLOCK_SIZE);
4862 		nvme->n_min_block_size = NVME_DEFAULT_MIN_BLOCK_SIZE;
4863 	}
4864 
4865 	if (nvme->n_submission_queues != -1 &&
4866 	    (nvme->n_submission_queues < 1 ||
4867 	    nvme->n_submission_queues > UINT16_MAX)) {
4868 		dev_err(dip, CE_WARN, "!\"submission-queues\"=%d is not "
4869 		    "valid. Must be [1..%d]", nvme->n_submission_queues,
4870 		    UINT16_MAX);
4871 		nvme->n_submission_queues = -1;
4872 	}
4873 
4874 	if (nvme->n_completion_queues != -1 &&
4875 	    (nvme->n_completion_queues < 1 ||
4876 	    nvme->n_completion_queues > UINT16_MAX)) {
4877 		dev_err(dip, CE_WARN, "!\"completion-queues\"=%d is not "
4878 		    "valid. Must be [1..%d]", nvme->n_completion_queues,
4879 		    UINT16_MAX);
4880 		nvme->n_completion_queues = -1;
4881 	}
4882 
4883 	if (nvme->n_admin_queue_len < NVME_MIN_ADMIN_QUEUE_LEN)
4884 		nvme->n_admin_queue_len = NVME_MIN_ADMIN_QUEUE_LEN;
4885 	else if (nvme->n_admin_queue_len > NVME_MAX_ADMIN_QUEUE_LEN)
4886 		nvme->n_admin_queue_len = NVME_MAX_ADMIN_QUEUE_LEN;
4887 
4888 	if (nvme->n_io_squeue_len < NVME_MIN_IO_QUEUE_LEN)
4889 		nvme->n_io_squeue_len = NVME_MIN_IO_QUEUE_LEN;
4890 	if (nvme->n_io_cqueue_len < NVME_MIN_IO_QUEUE_LEN)
4891 		nvme->n_io_cqueue_len = NVME_MIN_IO_QUEUE_LEN;
4892 
4893 	if (nvme->n_async_event_limit < 1)
4894 		nvme->n_async_event_limit = NVME_DEFAULT_ASYNC_EVENT_LIMIT;
4895 
4896 	nvme->n_reg_acc_attr = nvme_reg_acc_attr;
4897 	nvme->n_queue_dma_attr = nvme_queue_dma_attr;
4898 	nvme->n_prp_dma_attr = nvme_prp_dma_attr;
4899 	nvme->n_sgl_dma_attr = nvme_sgl_dma_attr;
4900 
4901 	/*
4902 	 * Set up FMA support.
4903 	 */
4904 	nvme->n_fm_cap = ddi_getprop(DDI_DEV_T_ANY, dip,
4905 	    DDI_PROP_CANSLEEP | DDI_PROP_DONTPASS, "fm-capable",
4906 	    DDI_FM_EREPORT_CAPABLE | DDI_FM_ACCCHK_CAPABLE |
4907 	    DDI_FM_DMACHK_CAPABLE | DDI_FM_ERRCB_CAPABLE);
4908 
4909 	ddi_fm_init(dip, &nvme->n_fm_cap, &nvme->n_fm_ibc);
4910 
4911 	if (nvme->n_fm_cap) {
4912 		if (nvme->n_fm_cap & DDI_FM_ACCCHK_CAPABLE)
4913 			nvme->n_reg_acc_attr.devacc_attr_access =
4914 			    DDI_FLAGERR_ACC;
4915 
4916 		if (nvme->n_fm_cap & DDI_FM_DMACHK_CAPABLE) {
4917 			nvme->n_prp_dma_attr.dma_attr_flags |= DDI_DMA_FLAGERR;
4918 			nvme->n_sgl_dma_attr.dma_attr_flags |= DDI_DMA_FLAGERR;
4919 		}
4920 
4921 		if (DDI_FM_EREPORT_CAP(nvme->n_fm_cap) ||
4922 		    DDI_FM_ERRCB_CAP(nvme->n_fm_cap))
4923 			pci_ereport_setup(dip);
4924 
4925 		if (DDI_FM_ERRCB_CAP(nvme->n_fm_cap))
4926 			ddi_fm_handler_register(dip, nvme_fm_errcb,
4927 			    (void *)nvme);
4928 	}
4929 
4930 	nvme->n_progress |= NVME_FMA_INIT;
4931 
4932 	/*
4933 	 * The spec defines several register sets. Only the controller
4934 	 * registers (set 1) are currently used.
4935 	 */
4936 	if (ddi_dev_nregs(dip, &nregs) == DDI_FAILURE ||
4937 	    nregs < 2 ||
4938 	    ddi_dev_regsize(dip, 1, &regsize) == DDI_FAILURE)
4939 		goto fail;
4940 
4941 	if (ddi_regs_map_setup(dip, 1, &nvme->n_regs, 0, regsize,
4942 	    &nvme->n_reg_acc_attr, &nvme->n_regh) != DDI_SUCCESS) {
4943 		dev_err(dip, CE_WARN, "!failed to map regset 1");
4944 		goto fail;
4945 	}
4946 
4947 	nvme->n_progress |= NVME_REGS_MAPPED;
4948 
4949 	/*
4950 	 * Set up kstats
4951 	 */
4952 	if (!nvme_stat_init(nvme)) {
4953 		dev_err(dip, CE_WARN, "!failed to create device kstats");
4954 		goto fail;
4955 	}
4956 	nvme->n_progress |= NVME_STAT_INIT;
4957 
4958 	/*
4959 	 * Create PRP DMA cache
4960 	 */
4961 	(void) snprintf(name, sizeof (name), "%s%d_prp_cache",
4962 	    ddi_driver_name(dip), ddi_get_instance(dip));
4963 	nvme->n_prp_cache = kmem_cache_create(name, sizeof (nvme_dma_t),
4964 	    0, nvme_prp_dma_constructor, nvme_prp_dma_destructor,
4965 	    NULL, (void *)nvme, NULL, 0);
4966 
4967 	if (nvme_init(nvme) != DDI_SUCCESS)
4968 		goto fail;
4969 
4970 	/*
4971 	 * Initialize the driver with the UFM subsystem
4972 	 */
4973 	if (ddi_ufm_init(dip, DDI_UFM_CURRENT_VERSION, &nvme_ufm_ops,
4974 	    &nvme->n_ufmh, nvme) != 0) {
4975 		dev_err(dip, CE_WARN, "!failed to initialize UFM subsystem");
4976 		goto fail;
4977 	}
4978 	mutex_init(&nvme->n_fwslot_mutex, NULL, MUTEX_DRIVER, NULL);
4979 	ddi_ufm_update(nvme->n_ufmh);
4980 	nvme->n_progress |= NVME_UFM_INIT;
4981 
4982 	nvme_mgmt_lock_init(&nvme->n_mgmt);
4983 	nvme_lock_init(&nvme->n_lock);
4984 	nvme->n_progress |= NVME_MGMT_INIT;
4985 	nvme->n_dead_status = NVME_IOCTL_E_CTRL_DEAD;
4986 
4987 	/*
4988 	 * Identify namespaces.
4989 	 */
4990 	nvme_mgmt_lock(nvme, NVME_MGMT_LOCK_NVME);
4991 
4992 	for (uint32_t i = 1; i <= nvme->n_namespace_count; i++) {
4993 		nvme_namespace_t *ns = nvme_nsid2ns(nvme, i);
4994 
4995 		nvme_lock_init(&ns->ns_lock);
4996 		ns->ns_progress |= NVME_NS_LOCK;
4997 
4998 		/*
4999 		 * Namespaces start out ignored. When nvme_init_ns() checks
5000 		 * their properties and finds they can be used, it will set
5001 		 * ns_ignore to B_FALSE. It will also use this state change
5002 		 * to keep an accurate count of attachable namespaces.
5003 		 */
5004 		ns->ns_ignore = B_TRUE;
5005 		if (nvme_init_ns(nvme, i) != 0) {
5006 			nvme_mgmt_unlock(nvme);
5007 			goto fail;
5008 		}
5009 
5010 		if (ddi_create_minor_node(nvme->n_dip, ns->ns_name, S_IFCHR,
5011 		    NVME_MINOR(ddi_get_instance(nvme->n_dip), i),
5012 		    DDI_NT_NVME_ATTACHMENT_POINT, 0) != DDI_SUCCESS) {
5013 			nvme_mgmt_unlock(nvme);
5014 			dev_err(dip, CE_WARN,
5015 			    "!failed to create minor node for namespace %d", i);
5016 			goto fail;
5017 		}
5018 	}
5019 
5020 	if (ddi_create_minor_node(dip, "devctl", S_IFCHR,
5021 	    NVME_MINOR(ddi_get_instance(dip), 0), DDI_NT_NVME_NEXUS, 0) !=
5022 	    DDI_SUCCESS) {
5023 		nvme_mgmt_unlock(nvme);
5024 		dev_err(dip, CE_WARN, "nvme_attach: "
5025 		    "cannot create devctl minor node");
5026 		goto fail;
5027 	}
5028 
5029 	attached_ns = B_FALSE;
5030 	for (uint32_t i = 1; i <= nvme->n_namespace_count; i++) {
5031 		nvme_ioctl_common_t com = { .nioc_nsid = i };
5032 
5033 		if (nvme_attach_ns(nvme, &com)) {
5034 			attached_ns = B_TRUE;
5035 		} else if (com.nioc_drv_err != NVME_IOCTL_E_UNSUP_ATTACH_NS) {
5036 			dev_err(nvme->n_dip, CE_WARN, "!failed to attach "
5037 			    "namespace %d due to blkdev error", i);
5038 			/*
5039 			 * Once we have successfully attached a namespace we
5040 			 * can no longer fail the driver attach as there is now
5041 			 * a blkdev child node linked to this device, and
5042 			 * our node is not yet in the attached state.
5043 			 */
5044 			if (!attached_ns) {
5045 				nvme_mgmt_unlock(nvme);
5046 				goto fail;
5047 			}
5048 		}
5049 	}
5050 
5051 	nvme_mgmt_unlock(nvme);
5052 
5053 	return (DDI_SUCCESS);
5054 
5055 fail:
5056 	/* attach successful anyway so that FMA can retire the device */
5057 	if (nvme->n_dead)
5058 		return (DDI_SUCCESS);
5059 
5060 	(void) nvme_detach(dip, DDI_DETACH);
5061 
5062 	return (DDI_FAILURE);
5063 }
5064 
5065 static int
5066 nvme_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
5067 {
5068 	int instance;
5069 	nvme_t *nvme;
5070 
5071 	if (cmd != DDI_DETACH)
5072 		return (DDI_FAILURE);
5073 
5074 	instance = ddi_get_instance(dip);
5075 
5076 	nvme = ddi_get_soft_state(nvme_state, instance);
5077 
5078 	if (nvme == NULL)
5079 		return (DDI_FAILURE);
5080 
5081 	/*
5082 	 * Remove all minor nodes from the device regardless of the source in
5083 	 * one swoop.
5084 	 */
5085 	ddi_remove_minor_node(dip, NULL);
5086 
5087 	/*
5088 	 * We need to remove the event handler as one of the first things that
5089 	 * we do. If we proceed with other teardown without removing the event
5090 	 * handler, we could end up in a very unfortunate race with ourselves.
5091 	 * The DDI does not serialize these with detach (just like timeout(9F)
5092 	 * and others).
5093 	 */
5094 	if (nvme->n_ev_rm_cb_id != NULL) {
5095 		(void) ddi_remove_event_handler(nvme->n_ev_rm_cb_id);
5096 	}
5097 	nvme->n_ev_rm_cb_id = NULL;
5098 
5099 	/*
5100 	 * If the controller was marked dead, there is a slight chance that we
5101 	 * are asynchronusly processing the removal taskq. Because we have
5102 	 * removed the callback handler above and all minor nodes and commands
5103 	 * are closed, there is no other way to get in here. As such, we wait on
5104 	 * the nvme_dead_taskq to complete so we can avoid tracking if it's
5105 	 * running or not.
5106 	 */
5107 	taskq_wait(nvme_dead_taskq);
5108 
5109 	if (nvme->n_ns) {
5110 		for (uint32_t i = 1; i <= nvme->n_namespace_count; i++) {
5111 			nvme_namespace_t *ns = nvme_nsid2ns(nvme, i);
5112 
5113 			if (ns->ns_bd_hdl) {
5114 				(void) bd_detach_handle(ns->ns_bd_hdl);
5115 				bd_free_handle(ns->ns_bd_hdl);
5116 			}
5117 
5118 			if (ns->ns_idns)
5119 				kmem_free(ns->ns_idns,
5120 				    sizeof (nvme_identify_nsid_t));
5121 			if (ns->ns_devid)
5122 				strfree(ns->ns_devid);
5123 
5124 			if ((ns->ns_progress & NVME_NS_LOCK) != 0)
5125 				nvme_lock_fini(&ns->ns_lock);
5126 		}
5127 
5128 		kmem_free(nvme->n_ns, sizeof (nvme_namespace_t) *
5129 		    nvme->n_namespace_count);
5130 	}
5131 
5132 	if (nvme->n_progress & NVME_MGMT_INIT) {
5133 		nvme_lock_fini(&nvme->n_lock);
5134 		nvme_mgmt_lock_fini(&nvme->n_mgmt);
5135 	}
5136 
5137 	if (nvme->n_progress & NVME_UFM_INIT) {
5138 		ddi_ufm_fini(nvme->n_ufmh);
5139 		mutex_destroy(&nvme->n_fwslot_mutex);
5140 	}
5141 
5142 	if (nvme->n_progress & NVME_INTERRUPTS)
5143 		nvme_release_interrupts(nvme);
5144 
5145 	for (uint_t i = 0; i < nvme->n_cq_count; i++) {
5146 		if (nvme->n_cq[i]->ncq_cmd_taskq != NULL)
5147 			taskq_wait(nvme->n_cq[i]->ncq_cmd_taskq);
5148 	}
5149 
5150 	if (nvme->n_progress & NVME_MUTEX_INIT) {
5151 		mutex_destroy(&nvme->n_minor_mutex);
5152 	}
5153 
5154 	if (nvme->n_ioq_count > 0) {
5155 		for (uint_t i = 1; i != nvme->n_ioq_count + 1; i++) {
5156 			if (nvme->n_ioq[i] != NULL) {
5157 				/* TODO: send destroy queue commands */
5158 				nvme_free_qpair(nvme->n_ioq[i]);
5159 			}
5160 		}
5161 
5162 		kmem_free(nvme->n_ioq, sizeof (nvme_qpair_t *) *
5163 		    (nvme->n_ioq_count + 1));
5164 	}
5165 
5166 	if (nvme->n_prp_cache != NULL) {
5167 		kmem_cache_destroy(nvme->n_prp_cache);
5168 	}
5169 
5170 	if (nvme->n_progress & NVME_REGS_MAPPED) {
5171 		nvme_shutdown(nvme, B_FALSE);
5172 		(void) nvme_reset(nvme, B_FALSE);
5173 	}
5174 
5175 	if (nvme->n_progress & NVME_CTRL_LIMITS)
5176 		sema_destroy(&nvme->n_abort_sema);
5177 
5178 	if (nvme->n_progress & NVME_ADMIN_QUEUE)
5179 		nvme_free_qpair(nvme->n_adminq);
5180 
5181 	if (nvme->n_cq_count > 0) {
5182 		nvme_destroy_cq_array(nvme, 0);
5183 		nvme->n_cq = NULL;
5184 		nvme->n_cq_count = 0;
5185 	}
5186 
5187 	if (nvme->n_idcomns)
5188 		kmem_free(nvme->n_idcomns, NVME_IDENTIFY_BUFSIZE);
5189 
5190 	if (nvme->n_idctl)
5191 		kmem_free(nvme->n_idctl, NVME_IDENTIFY_BUFSIZE);
5192 
5193 	if (nvme->n_progress & NVME_REGS_MAPPED)
5194 		ddi_regs_map_free(&nvme->n_regh);
5195 
5196 	if (nvme->n_progress & NVME_STAT_INIT)
5197 		nvme_stat_cleanup(nvme);
5198 
5199 	if (nvme->n_progress & NVME_FMA_INIT) {
5200 		if (DDI_FM_ERRCB_CAP(nvme->n_fm_cap))
5201 			ddi_fm_handler_unregister(nvme->n_dip);
5202 
5203 		if (DDI_FM_EREPORT_CAP(nvme->n_fm_cap) ||
5204 		    DDI_FM_ERRCB_CAP(nvme->n_fm_cap))
5205 			pci_ereport_teardown(nvme->n_dip);
5206 
5207 		ddi_fm_fini(nvme->n_dip);
5208 	}
5209 
5210 	if (nvme->n_progress & NVME_PCI_CONFIG)
5211 		pci_config_teardown(&nvme->n_pcicfg_handle);
5212 
5213 	if (nvme->n_vendor != NULL)
5214 		strfree(nvme->n_vendor);
5215 
5216 	if (nvme->n_product != NULL)
5217 		strfree(nvme->n_product);
5218 
5219 	ddi_soft_state_free(nvme_state, instance);
5220 
5221 	return (DDI_SUCCESS);
5222 }
5223 
5224 static int
5225 nvme_quiesce(dev_info_t *dip)
5226 {
5227 	int instance;
5228 	nvme_t *nvme;
5229 
5230 	instance = ddi_get_instance(dip);
5231 
5232 	nvme = ddi_get_soft_state(nvme_state, instance);
5233 
5234 	if (nvme == NULL)
5235 		return (DDI_FAILURE);
5236 
5237 	nvme_shutdown(nvme, B_TRUE);
5238 
5239 	(void) nvme_reset(nvme, B_TRUE);
5240 
5241 	return (DDI_SUCCESS);
5242 }
5243 
5244 static int
5245 nvme_fill_prp(nvme_cmd_t *cmd, ddi_dma_handle_t dma)
5246 {
5247 	nvme_t *nvme = cmd->nc_nvme;
5248 	uint_t nprp_per_page, nprp;
5249 	uint64_t *prp;
5250 	const ddi_dma_cookie_t *cookie;
5251 	uint_t idx;
5252 	uint_t ncookies = ddi_dma_ncookies(dma);
5253 
5254 	if (ncookies == 0)
5255 		return (DDI_FAILURE);
5256 
5257 	if ((cookie = ddi_dma_cookie_get(dma, 0)) == NULL)
5258 		return (DDI_FAILURE);
5259 	cmd->nc_sqe.sqe_dptr.d_prp[0] = cookie->dmac_laddress;
5260 
5261 	if (ncookies == 1) {
5262 		cmd->nc_sqe.sqe_dptr.d_prp[1] = 0;
5263 		return (DDI_SUCCESS);
5264 	} else if (ncookies == 2) {
5265 		if ((cookie = ddi_dma_cookie_get(dma, 1)) == NULL)
5266 			return (DDI_FAILURE);
5267 		cmd->nc_sqe.sqe_dptr.d_prp[1] = cookie->dmac_laddress;
5268 		return (DDI_SUCCESS);
5269 	}
5270 
5271 	/*
5272 	 * At this point, we're always operating on cookies at
5273 	 * index >= 1 and writing the addresses of those cookies
5274 	 * into a new page. The address of that page is stored
5275 	 * as the second PRP entry.
5276 	 */
5277 	nprp_per_page = nvme->n_pagesize / sizeof (uint64_t);
5278 	ASSERT(nprp_per_page > 0);
5279 
5280 	/*
5281 	 * We currently don't support chained PRPs and set up our DMA
5282 	 * attributes to reflect that. If we still get an I/O request
5283 	 * that needs a chained PRP something is very wrong. Account
5284 	 * for the first cookie here, which we've placed in d_prp[0].
5285 	 */
5286 	nprp = howmany(ncookies - 1, nprp_per_page);
5287 	VERIFY(nprp == 1);
5288 
5289 	/*
5290 	 * Allocate a page of pointers, in which we'll write the
5291 	 * addresses of cookies 1 to `ncookies`.
5292 	 */
5293 	cmd->nc_prp = kmem_cache_alloc(nvme->n_prp_cache, KM_SLEEP);
5294 	bzero(cmd->nc_prp->nd_memp, cmd->nc_prp->nd_len);
5295 	cmd->nc_sqe.sqe_dptr.d_prp[1] = cmd->nc_prp->nd_cookie.dmac_laddress;
5296 
5297 	prp = (uint64_t *)cmd->nc_prp->nd_memp;
5298 	for (idx = 1; idx < ncookies; idx++) {
5299 		if ((cookie = ddi_dma_cookie_get(dma, idx)) == NULL)
5300 			return (DDI_FAILURE);
5301 		*prp++ = cookie->dmac_laddress;
5302 	}
5303 
5304 	(void) ddi_dma_sync(cmd->nc_prp->nd_dmah, 0, cmd->nc_prp->nd_len,
5305 	    DDI_DMA_SYNC_FORDEV);
5306 	return (DDI_SUCCESS);
5307 }
5308 
5309 /*
5310  * The maximum number of requests supported for a deallocate request is
5311  * NVME_DSET_MGMT_MAX_RANGES (256) -- this is from the NVMe 1.1 spec (and
5312  * unchanged through at least 1.4a). The definition of nvme_range_t is also
5313  * from the NVMe 1.1 spec. Together, the result is that all of the ranges for
5314  * a deallocate request will fit into the smallest supported namespace page
5315  * (4k).
5316  */
5317 CTASSERT(sizeof (nvme_range_t) * NVME_DSET_MGMT_MAX_RANGES == 4096);
5318 
5319 static int
5320 nvme_fill_ranges(nvme_cmd_t *cmd, bd_xfer_t *xfer, uint64_t blocksize,
5321     int allocflag)
5322 {
5323 	const dkioc_free_list_t *dfl = xfer->x_dfl;
5324 	const dkioc_free_list_ext_t *exts = dfl->dfl_exts;
5325 	nvme_t *nvme = cmd->nc_nvme;
5326 	nvme_range_t *ranges = NULL;
5327 	uint_t i;
5328 
5329 	/*
5330 	 * The number of ranges in the request is 0s based (that is
5331 	 * word10 == 0 -> 1 range, word10 == 1 -> 2 ranges, ...,
5332 	 * word10 == 255 -> 256 ranges). Therefore the allowed values are
5333 	 * [1..NVME_DSET_MGMT_MAX_RANGES]. If blkdev gives us a bad request,
5334 	 * we either provided bad info in nvme_bd_driveinfo() or there is a bug
5335 	 * in blkdev.
5336 	 */
5337 	VERIFY3U(dfl->dfl_num_exts, >, 0);
5338 	VERIFY3U(dfl->dfl_num_exts, <=, NVME_DSET_MGMT_MAX_RANGES);
5339 	cmd->nc_sqe.sqe_cdw10 = (dfl->dfl_num_exts - 1) & 0xff;
5340 
5341 	cmd->nc_sqe.sqe_cdw11 = NVME_DSET_MGMT_ATTR_DEALLOCATE;
5342 
5343 	cmd->nc_prp = kmem_cache_alloc(nvme->n_prp_cache, allocflag);
5344 	if (cmd->nc_prp == NULL)
5345 		return (DDI_FAILURE);
5346 
5347 	bzero(cmd->nc_prp->nd_memp, cmd->nc_prp->nd_len);
5348 	ranges = (nvme_range_t *)cmd->nc_prp->nd_memp;
5349 
5350 	cmd->nc_sqe.sqe_dptr.d_prp[0] = cmd->nc_prp->nd_cookie.dmac_laddress;
5351 	cmd->nc_sqe.sqe_dptr.d_prp[1] = 0;
5352 
5353 	for (i = 0; i < dfl->dfl_num_exts; i++) {
5354 		uint64_t lba, len;
5355 
5356 		lba = (dfl->dfl_offset + exts[i].dfle_start) / blocksize;
5357 		len = exts[i].dfle_length / blocksize;
5358 
5359 		VERIFY3U(len, <=, UINT32_MAX);
5360 
5361 		/* No context attributes for a deallocate request */
5362 		ranges[i].nr_ctxattr = 0;
5363 		ranges[i].nr_len = len;
5364 		ranges[i].nr_lba = lba;
5365 	}
5366 
5367 	(void) ddi_dma_sync(cmd->nc_prp->nd_dmah, 0, cmd->nc_prp->nd_len,
5368 	    DDI_DMA_SYNC_FORDEV);
5369 
5370 	return (DDI_SUCCESS);
5371 }
5372 
5373 static nvme_cmd_t *
5374 nvme_create_nvm_cmd(nvme_namespace_t *ns, uint8_t opc, bd_xfer_t *xfer)
5375 {
5376 	nvme_t *nvme = ns->ns_nvme;
5377 	nvme_cmd_t *cmd;
5378 	int allocflag;
5379 
5380 	/*
5381 	 * Blkdev only sets BD_XFER_POLL when dumping, so don't sleep.
5382 	 */
5383 	allocflag = (xfer->x_flags & BD_XFER_POLL) ? KM_NOSLEEP : KM_SLEEP;
5384 	cmd = nvme_alloc_cmd(nvme, allocflag);
5385 
5386 	if (cmd == NULL)
5387 		return (NULL);
5388 
5389 	cmd->nc_sqe.sqe_opc = opc;
5390 	cmd->nc_callback = nvme_bd_xfer_done;
5391 	cmd->nc_xfer = xfer;
5392 
5393 	switch (opc) {
5394 	case NVME_OPC_NVM_WRITE:
5395 	case NVME_OPC_NVM_READ:
5396 		VERIFY(xfer->x_nblks <= 0x10000);
5397 
5398 		cmd->nc_sqe.sqe_nsid = ns->ns_id;
5399 
5400 		cmd->nc_sqe.sqe_cdw10 = xfer->x_blkno & 0xffffffffu;
5401 		cmd->nc_sqe.sqe_cdw11 = (xfer->x_blkno >> 32);
5402 		cmd->nc_sqe.sqe_cdw12 = (uint16_t)(xfer->x_nblks - 1);
5403 
5404 		if (nvme_fill_prp(cmd, xfer->x_dmah) != DDI_SUCCESS)
5405 			goto fail;
5406 		break;
5407 
5408 	case NVME_OPC_NVM_FLUSH:
5409 		cmd->nc_sqe.sqe_nsid = ns->ns_id;
5410 		break;
5411 
5412 	case NVME_OPC_NVM_DSET_MGMT:
5413 		cmd->nc_sqe.sqe_nsid = ns->ns_id;
5414 
5415 		if (nvme_fill_ranges(cmd, xfer,
5416 		    (uint64_t)ns->ns_block_size, allocflag) != DDI_SUCCESS)
5417 			goto fail;
5418 		break;
5419 
5420 	default:
5421 		goto fail;
5422 	}
5423 
5424 	return (cmd);
5425 
5426 fail:
5427 	nvme_free_cmd(cmd);
5428 	return (NULL);
5429 }
5430 
5431 static void
5432 nvme_bd_xfer_done(void *arg)
5433 {
5434 	nvme_cmd_t *cmd = arg;
5435 	bd_xfer_t *xfer = cmd->nc_xfer;
5436 	int error = 0;
5437 
5438 	error = nvme_check_cmd_status(cmd);
5439 	nvme_free_cmd(cmd);
5440 
5441 	bd_xfer_done(xfer, error);
5442 }
5443 
5444 static void
5445 nvme_bd_driveinfo(void *arg, bd_drive_t *drive)
5446 {
5447 	nvme_namespace_t *ns = arg;
5448 	nvme_t *nvme = ns->ns_nvme;
5449 	uint_t ns_count = MAX(1, nvme->n_namespaces_attachable);
5450 
5451 	nvme_mgmt_lock(nvme, NVME_MGMT_LOCK_BDRO);
5452 
5453 	/*
5454 	 * Set the blkdev qcount to the number of submission queues.
5455 	 * It will then create one waitq/runq pair for each submission
5456 	 * queue and spread I/O requests across the queues.
5457 	 */
5458 	drive->d_qcount = nvme->n_ioq_count;
5459 
5460 	/*
5461 	 * I/O activity to individual namespaces is distributed across
5462 	 * each of the d_qcount blkdev queues (which has been set to
5463 	 * the number of nvme submission queues). d_qsize is the number
5464 	 * of submitted and not completed I/Os within each queue that blkdev
5465 	 * will allow before it starts holding them in the waitq.
5466 	 *
5467 	 * Each namespace will create a child blkdev instance, for each one
5468 	 * we try and set the d_qsize so that each namespace gets an
5469 	 * equal portion of the submission queue.
5470 	 *
5471 	 * If post instantiation of the nvme drive, n_namespaces_attachable
5472 	 * changes and a namespace is attached it could calculate a
5473 	 * different d_qsize. It may even be that the sum of the d_qsizes is
5474 	 * now beyond the submission queue size. Should that be the case
5475 	 * and the I/O rate is such that blkdev attempts to submit more
5476 	 * I/Os than the size of the submission queue, the excess I/Os
5477 	 * will be held behind the semaphore nq_sema.
5478 	 */
5479 	drive->d_qsize = nvme->n_io_squeue_len / ns_count;
5480 
5481 	/*
5482 	 * Don't let the queue size drop below the minimum, though.
5483 	 */
5484 	drive->d_qsize = MAX(drive->d_qsize, NVME_MIN_IO_QUEUE_LEN);
5485 
5486 	/*
5487 	 * d_maxxfer is not set, which means the value is taken from the DMA
5488 	 * attributes specified to bd_alloc_handle.
5489 	 */
5490 
5491 	drive->d_removable = B_FALSE;
5492 	drive->d_hotpluggable = B_FALSE;
5493 
5494 	bcopy(ns->ns_eui64, drive->d_eui64, sizeof (drive->d_eui64));
5495 	drive->d_target = ns->ns_id;
5496 	drive->d_lun = 0;
5497 
5498 	drive->d_model = nvme->n_idctl->id_model;
5499 	drive->d_model_len = sizeof (nvme->n_idctl->id_model);
5500 	drive->d_vendor = nvme->n_vendor;
5501 	drive->d_vendor_len = strlen(nvme->n_vendor);
5502 	drive->d_product = nvme->n_product;
5503 	drive->d_product_len = strlen(nvme->n_product);
5504 	drive->d_serial = nvme->n_idctl->id_serial;
5505 	drive->d_serial_len = sizeof (nvme->n_idctl->id_serial);
5506 	drive->d_revision = nvme->n_idctl->id_fwrev;
5507 	drive->d_revision_len = sizeof (nvme->n_idctl->id_fwrev);
5508 
5509 	/*
5510 	 * If we support the dataset management command, the only restrictions
5511 	 * on a discard request are the maximum number of ranges (segments)
5512 	 * per single request.
5513 	 */
5514 	if (nvme->n_idctl->id_oncs.on_dset_mgmt)
5515 		drive->d_max_free_seg = NVME_DSET_MGMT_MAX_RANGES;
5516 
5517 	nvme_mgmt_unlock(nvme);
5518 }
5519 
5520 static int
5521 nvme_bd_mediainfo(void *arg, bd_media_t *media)
5522 {
5523 	nvme_namespace_t *ns = arg;
5524 	nvme_t *nvme = ns->ns_nvme;
5525 
5526 	if (nvme->n_dead) {
5527 		return (EIO);
5528 	}
5529 
5530 	nvme_mgmt_lock(nvme, NVME_MGMT_LOCK_BDRO);
5531 
5532 	media->m_nblks = ns->ns_block_count;
5533 	media->m_blksize = ns->ns_block_size;
5534 	media->m_readonly = B_FALSE;
5535 	media->m_solidstate = B_TRUE;
5536 
5537 	media->m_pblksize = ns->ns_best_block_size;
5538 
5539 	nvme_mgmt_unlock(nvme);
5540 
5541 	return (0);
5542 }
5543 
5544 static int
5545 nvme_bd_cmd(nvme_namespace_t *ns, bd_xfer_t *xfer, uint8_t opc)
5546 {
5547 	nvme_t *nvme = ns->ns_nvme;
5548 	nvme_cmd_t *cmd;
5549 	nvme_qpair_t *ioq;
5550 	boolean_t poll;
5551 	int ret;
5552 
5553 	if (nvme->n_dead) {
5554 		return (EIO);
5555 	}
5556 
5557 	cmd = nvme_create_nvm_cmd(ns, opc, xfer);
5558 	if (cmd == NULL)
5559 		return (ENOMEM);
5560 
5561 	cmd->nc_sqid = xfer->x_qnum + 1;
5562 	ASSERT(cmd->nc_sqid <= nvme->n_ioq_count);
5563 	ioq = nvme->n_ioq[cmd->nc_sqid];
5564 
5565 	/*
5566 	 * Get the polling flag before submitting the command. The command may
5567 	 * complete immediately after it was submitted, which means we must
5568 	 * treat both cmd and xfer as if they have been freed already.
5569 	 */
5570 	poll = (xfer->x_flags & BD_XFER_POLL) != 0;
5571 
5572 	ret = nvme_submit_io_cmd(ioq, cmd);
5573 
5574 	if (ret != 0)
5575 		return (ret);
5576 
5577 	if (!poll)
5578 		return (0);
5579 
5580 	do {
5581 		cmd = nvme_retrieve_cmd(nvme, ioq);
5582 		if (cmd != NULL) {
5583 			ASSERT0(cmd->nc_flags & NVME_CMD_F_USELOCK);
5584 			cmd->nc_callback(cmd);
5585 		} else {
5586 			drv_usecwait(10);
5587 		}
5588 	} while (ioq->nq_active_cmds != 0);
5589 
5590 	return (0);
5591 }
5592 
5593 static int
5594 nvme_bd_read(void *arg, bd_xfer_t *xfer)
5595 {
5596 	nvme_namespace_t *ns = arg;
5597 
5598 	return (nvme_bd_cmd(ns, xfer, NVME_OPC_NVM_READ));
5599 }
5600 
5601 static int
5602 nvme_bd_write(void *arg, bd_xfer_t *xfer)
5603 {
5604 	nvme_namespace_t *ns = arg;
5605 
5606 	return (nvme_bd_cmd(ns, xfer, NVME_OPC_NVM_WRITE));
5607 }
5608 
5609 static int
5610 nvme_bd_sync(void *arg, bd_xfer_t *xfer)
5611 {
5612 	nvme_namespace_t *ns = arg;
5613 
5614 	if (ns->ns_nvme->n_dead)
5615 		return (EIO);
5616 
5617 	/*
5618 	 * If the volatile write cache is not present or not enabled the FLUSH
5619 	 * command is a no-op, so we can take a shortcut here.
5620 	 */
5621 	if (!ns->ns_nvme->n_write_cache_present) {
5622 		bd_xfer_done(xfer, ENOTSUP);
5623 		return (0);
5624 	}
5625 
5626 	if (!ns->ns_nvme->n_write_cache_enabled) {
5627 		bd_xfer_done(xfer, 0);
5628 		return (0);
5629 	}
5630 
5631 	return (nvme_bd_cmd(ns, xfer, NVME_OPC_NVM_FLUSH));
5632 }
5633 
5634 static int
5635 nvme_bd_devid(void *arg, dev_info_t *devinfo, ddi_devid_t *devid)
5636 {
5637 	nvme_namespace_t *ns = arg;
5638 	nvme_t *nvme = ns->ns_nvme;
5639 
5640 	if (nvme->n_dead) {
5641 		return (EIO);
5642 	}
5643 
5644 	if (*(uint64_t *)ns->ns_nguid != 0 ||
5645 	    *(uint64_t *)(ns->ns_nguid + 8) != 0) {
5646 		return (ddi_devid_init(devinfo, DEVID_NVME_NGUID,
5647 		    sizeof (ns->ns_nguid), ns->ns_nguid, devid));
5648 	} else if (*(uint64_t *)ns->ns_eui64 != 0) {
5649 		return (ddi_devid_init(devinfo, DEVID_NVME_EUI64,
5650 		    sizeof (ns->ns_eui64), ns->ns_eui64, devid));
5651 	} else {
5652 		return (ddi_devid_init(devinfo, DEVID_NVME_NSID,
5653 		    strlen(ns->ns_devid), ns->ns_devid, devid));
5654 	}
5655 }
5656 
5657 static int
5658 nvme_bd_free_space(void *arg, bd_xfer_t *xfer)
5659 {
5660 	nvme_namespace_t *ns = arg;
5661 
5662 	if (xfer->x_dfl == NULL)
5663 		return (EINVAL);
5664 
5665 	if (!ns->ns_nvme->n_idctl->id_oncs.on_dset_mgmt)
5666 		return (ENOTSUP);
5667 
5668 	return (nvme_bd_cmd(ns, xfer, NVME_OPC_NVM_DSET_MGMT));
5669 }
5670 
5671 static int
5672 nvme_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
5673 {
5674 #ifndef __lock_lint
5675 	_NOTE(ARGUNUSED(cred_p));
5676 #endif
5677 	nvme_t *nvme;
5678 	nvme_minor_t *minor = NULL;
5679 	uint32_t nsid;
5680 	minor_t m = getminor(*devp);
5681 	int rv = 0;
5682 
5683 	if (otyp != OTYP_CHR)
5684 		return (EINVAL);
5685 
5686 	if (m >= NVME_OPEN_MINOR_MIN)
5687 		return (ENXIO);
5688 
5689 	nvme = ddi_get_soft_state(nvme_state, NVME_MINOR_INST(m));
5690 	nsid = NVME_MINOR_NSID(m);
5691 
5692 	if (nvme == NULL)
5693 		return (ENXIO);
5694 
5695 	if (nsid > nvme->n_namespace_count)
5696 		return (ENXIO);
5697 
5698 	if (nvme->n_dead)
5699 		return (EIO);
5700 
5701 	/*
5702 	 * At this point, we're going to allow an open to proceed on this
5703 	 * device. We need to allocate a new instance for this (presuming one is
5704 	 * available).
5705 	 */
5706 	minor = kmem_zalloc(sizeof (nvme_minor_t), KM_NOSLEEP_LAZY);
5707 	if (minor == NULL) {
5708 		return (ENOMEM);
5709 	}
5710 
5711 	cv_init(&minor->nm_cv, NULL, CV_DRIVER, NULL);
5712 	list_link_init(&minor->nm_ctrl_lock.nli_node);
5713 	minor->nm_ctrl_lock.nli_nvme = nvme;
5714 	minor->nm_ctrl_lock.nli_minor = minor;
5715 	list_link_init(&minor->nm_ns_lock.nli_node);
5716 	minor->nm_ns_lock.nli_nvme = nvme;
5717 	minor->nm_ns_lock.nli_minor = minor;
5718 	minor->nm_minor = id_alloc_nosleep(nvme_open_minors);
5719 	if (minor->nm_minor == -1) {
5720 		nvme_minor_free(minor);
5721 		return (ENOSPC);
5722 	}
5723 
5724 	minor->nm_ctrl = nvme;
5725 	if (nsid != 0) {
5726 		minor->nm_ns = nvme_nsid2ns(nvme, nsid);
5727 	}
5728 
5729 	/*
5730 	 * Before we check for exclusive access and attempt a lock if requested,
5731 	 * ensure that this minor is persisted.
5732 	 */
5733 	mutex_enter(&nvme_open_minors_mutex);
5734 	avl_add(&nvme_open_minors_avl, minor);
5735 	mutex_exit(&nvme_open_minors_mutex);
5736 
5737 	/*
5738 	 * A request for opening this FEXCL, is translated into a non-blocking
5739 	 * write lock of the appropriate entity. This honors the original
5740 	 * semantics here. In the future, we should see if we can remove this
5741 	 * and turn a request for FEXCL at open into ENOTSUP.
5742 	 */
5743 	mutex_enter(&nvme->n_minor_mutex);
5744 	if ((flag & FEXCL) != 0) {
5745 		nvme_ioctl_lock_t lock = {
5746 			.nil_level = NVME_LOCK_L_WRITE,
5747 			.nil_flags = NVME_LOCK_F_DONT_BLOCK
5748 		};
5749 
5750 		if (minor->nm_ns != NULL) {
5751 			lock.nil_ent = NVME_LOCK_E_NS;
5752 			lock.nil_common.nioc_nsid = nsid;
5753 		} else {
5754 			lock.nil_ent = NVME_LOCK_E_CTRL;
5755 		}
5756 		nvme_rwlock(minor, &lock);
5757 		if (lock.nil_common.nioc_drv_err != NVME_IOCTL_E_OK) {
5758 			mutex_exit(&nvme->n_minor_mutex);
5759 
5760 			mutex_enter(&nvme_open_minors_mutex);
5761 			avl_remove(&nvme_open_minors_avl, minor);
5762 			mutex_exit(&nvme_open_minors_mutex);
5763 
5764 			nvme_minor_free(minor);
5765 			return (EBUSY);
5766 		}
5767 	}
5768 	mutex_exit(&nvme->n_minor_mutex);
5769 
5770 	*devp = makedevice(getmajor(*devp), (minor_t)minor->nm_minor);
5771 	return (rv);
5772 
5773 }
5774 
5775 static int
5776 nvme_close(dev_t dev, int flag __unused, int otyp, cred_t *cred_p __unused)
5777 {
5778 	nvme_minor_t *minor;
5779 	nvme_t *nvme;
5780 
5781 	if (otyp != OTYP_CHR) {
5782 		return (ENXIO);
5783 	}
5784 
5785 	minor = nvme_minor_find_by_dev(dev);
5786 	if (minor == NULL) {
5787 		return (ENXIO);
5788 	}
5789 
5790 	mutex_enter(&nvme_open_minors_mutex);
5791 	avl_remove(&nvme_open_minors_avl, minor);
5792 	mutex_exit(&nvme_open_minors_mutex);
5793 
5794 	/*
5795 	 * When this device is being closed, we must ensure that any locks held
5796 	 * by this are dealt with.
5797 	 */
5798 	nvme = minor->nm_ctrl;
5799 	mutex_enter(&nvme->n_minor_mutex);
5800 	ASSERT3U(minor->nm_ctrl_lock.nli_state, !=, NVME_LOCK_STATE_BLOCKED);
5801 	ASSERT3U(minor->nm_ns_lock.nli_state, !=, NVME_LOCK_STATE_BLOCKED);
5802 
5803 	if (minor->nm_ctrl_lock.nli_state == NVME_LOCK_STATE_ACQUIRED) {
5804 		VERIFY3P(minor->nm_ctrl_lock.nli_lock, !=, NULL);
5805 		nvme_rwunlock(&minor->nm_ctrl_lock,
5806 		    minor->nm_ctrl_lock.nli_lock);
5807 	}
5808 
5809 	if (minor->nm_ns_lock.nli_state == NVME_LOCK_STATE_ACQUIRED) {
5810 		VERIFY3P(minor->nm_ns_lock.nli_lock, !=, NULL);
5811 		nvme_rwunlock(&minor->nm_ns_lock, minor->nm_ns_lock.nli_lock);
5812 	}
5813 	mutex_exit(&nvme->n_minor_mutex);
5814 
5815 	nvme_minor_free(minor);
5816 
5817 	return (0);
5818 }
5819 
5820 void
5821 nvme_ioctl_success(nvme_ioctl_common_t *ioc)
5822 {
5823 	ioc->nioc_drv_err = NVME_IOCTL_E_OK;
5824 	ioc->nioc_ctrl_sc = NVME_CQE_SC_GEN_SUCCESS;
5825 	ioc->nioc_ctrl_sct = NVME_CQE_SCT_GENERIC;
5826 }
5827 
5828 boolean_t
5829 nvme_ioctl_error(nvme_ioctl_common_t *ioc, nvme_ioctl_errno_t err, uint32_t sct,
5830     uint32_t sc)
5831 {
5832 	ioc->nioc_drv_err = err;
5833 	ioc->nioc_ctrl_sct = sct;
5834 	ioc->nioc_ctrl_sc = sc;
5835 
5836 	return (B_FALSE);
5837 }
5838 
5839 static int
5840 nvme_ioctl_copyout_error(nvme_ioctl_errno_t err, intptr_t uaddr, int mode)
5841 {
5842 	nvme_ioctl_common_t ioc;
5843 
5844 	ASSERT3U(err, !=, NVME_IOCTL_E_CTRL_ERROR);
5845 	bzero(&ioc, sizeof (ioc));
5846 	if (ddi_copyout(&ioc, (void *)uaddr, sizeof (nvme_ioctl_common_t),
5847 	    mode & FKIOCTL) != 0) {
5848 		return (EFAULT);
5849 	}
5850 	return (0);
5851 }
5852 
5853 /*
5854  * The companion to the namespace checking. This occurs after any rewriting
5855  * occurs. This is the primary point that we attempt to enforce any operation's
5856  * exclusivity. Note, it is theoretically possible for an operation to be
5857  * ongoing and to have someone with an exclusive lock ask to unlock it for some
5858  * reason. This does not maintain the number of such events that are going on.
5859  * While perhaps this is leaving too much up to the user, by the same token we
5860  * don't try to stop them from issuing two different format NVM commands
5861  * targeting the whole device at the same time either, even though the
5862  * controller would really rather that didn't happen.
5863  */
5864 static boolean_t
5865 nvme_ioctl_excl_check(nvme_minor_t *minor, nvme_ioctl_common_t *ioc,
5866     const nvme_ioctl_check_t *check)
5867 {
5868 	nvme_t *const nvme = minor->nm_ctrl;
5869 	nvme_namespace_t *ns;
5870 	boolean_t have_ctrl, have_ns, ctrl_is_excl, ns_is_excl;
5871 
5872 	/*
5873 	 * If the command doesn't require anything, then we're done.
5874 	 */
5875 	if (check->nck_excl == NVME_IOCTL_EXCL_SKIP) {
5876 		return (B_TRUE);
5877 	}
5878 
5879 	if (ioc->nioc_nsid == 0 || ioc->nioc_nsid == NVME_NSID_BCAST) {
5880 		ns = NULL;
5881 	} else {
5882 		ns = nvme_nsid2ns(nvme, ioc->nioc_nsid);
5883 	}
5884 
5885 	mutex_enter(&nvme->n_minor_mutex);
5886 	ctrl_is_excl = nvme->n_lock.nl_writer != NULL;
5887 	have_ctrl = nvme->n_lock.nl_writer == &minor->nm_ctrl_lock;
5888 	if (ns != NULL) {
5889 		/*
5890 		 * We explicitly test the namespace lock's writer versus asking
5891 		 * the minor because the minor's namespace lock may apply to a
5892 		 * different namespace.
5893 		 */
5894 		ns_is_excl = ns->ns_lock.nl_writer != NULL;
5895 		have_ns = ns->ns_lock.nl_writer == &minor->nm_ns_lock;
5896 		ASSERT0(have_ctrl && have_ns);
5897 #ifdef	DEBUG
5898 		if (have_ns) {
5899 			ASSERT3P(minor->nm_ns_lock.nli_ns, ==, ns);
5900 		}
5901 #endif
5902 	} else {
5903 		ns_is_excl = B_FALSE;
5904 		have_ns = B_FALSE;
5905 	}
5906 	ASSERT0(ctrl_is_excl && ns_is_excl);
5907 	mutex_exit(&nvme->n_minor_mutex);
5908 
5909 	if (check->nck_excl == NVME_IOCTL_EXCL_WRITE) {
5910 		if (ns == NULL) {
5911 			if (have_ctrl) {
5912 				return (B_TRUE);
5913 			}
5914 			return (nvme_ioctl_error(ioc,
5915 			    NVME_IOCTL_E_NEED_CTRL_WRLOCK, 0, 0));
5916 		} else {
5917 			if (have_ctrl || have_ns) {
5918 				return (B_TRUE);
5919 			}
5920 			return (nvme_ioctl_error(ioc,
5921 			    NVME_IOCTL_E_NEED_NS_WRLOCK, 0, 0));
5922 		}
5923 	}
5924 
5925 	/*
5926 	 * Now we have an operation that does not require exclusive access. We
5927 	 * can proceed as long as no one else has it or if someone does it is
5928 	 * us. Regardless of what we target, a controller lock will stop us.
5929 	 */
5930 	if (ctrl_is_excl && !have_ctrl) {
5931 		return (nvme_ioctl_error(ioc, NVME_IOCTL_E_CTRL_LOCKED, 0, 0));
5932 	}
5933 
5934 	/*
5935 	 * Only check namespace exclusivity if we are targeting one.
5936 	 */
5937 	if (ns != NULL && ns_is_excl && !have_ns) {
5938 		return (nvme_ioctl_error(ioc, NVME_IOCTL_E_NS_LOCKED, 0, 0));
5939 	}
5940 
5941 	return (B_TRUE);
5942 }
5943 
5944 /*
5945  * Perform common checking as to whether or not an ioctl operation may proceed.
5946  * We check in this function various aspects of the namespace attributes that
5947  * it's calling on. Once the namespace attributes and any possible rewriting
5948  * have been performed, then we proceed to check whether or not the requisite
5949  * exclusive access is present in nvme_ioctl_excl_check().
5950  */
5951 static boolean_t
5952 nvme_ioctl_check(nvme_minor_t *minor, nvme_ioctl_common_t *ioc,
5953     const nvme_ioctl_check_t *check)
5954 {
5955 	/*
5956 	 * If the minor has a namespace pointer, then it is constrained to that
5957 	 * namespace. If a namespace is allowed, then there are only two valid
5958 	 * values that we can find. The first is matching the minor. The second
5959 	 * is our value zero, which will be transformed to the current
5960 	 * namespace.
5961 	 */
5962 	if (minor->nm_ns != NULL) {
5963 		if (!check->nck_ns_ok || !check->nck_ns_minor_ok) {
5964 			return (nvme_ioctl_error(ioc, NVME_IOCTL_E_NOT_CTRL, 0,
5965 			    0));
5966 		}
5967 
5968 		if (ioc->nioc_nsid == 0) {
5969 			ioc->nioc_nsid = minor->nm_ns->ns_id;
5970 		} else if (ioc->nioc_nsid != minor->nm_ns->ns_id) {
5971 			return (nvme_ioctl_error(ioc,
5972 			    NVME_IOCTL_E_MINOR_WRONG_NS, 0, 0));
5973 		}
5974 
5975 		return (nvme_ioctl_excl_check(minor, ioc, check));
5976 	}
5977 
5978 	/*
5979 	 * If we've been told to skip checking the controller, here's where we
5980 	 * do that. This should really only be for commands which use the
5981 	 * namespace ID for listing purposes and therefore can have
5982 	 * traditionally illegal values here.
5983 	 */
5984 	if (check->nck_skip_ctrl) {
5985 		return (nvme_ioctl_excl_check(minor, ioc, check));
5986 	}
5987 
5988 	/*
5989 	 * At this point, we know that we're on the controller's node. We first
5990 	 * deal with the simple case, is a namespace allowed at all or not. If
5991 	 * it is not allowed, then the only acceptable value is zero.
5992 	 */
5993 	if (!check->nck_ns_ok) {
5994 		if (ioc->nioc_nsid != 0) {
5995 			return (nvme_ioctl_error(ioc, NVME_IOCTL_E_NS_UNUSE, 0,
5996 			    0));
5997 		}
5998 
5999 		return (nvme_ioctl_excl_check(minor, ioc, check));
6000 	}
6001 
6002 	/*
6003 	 * At this point, we know that a controller is allowed to use a
6004 	 * namespace. If we haven't been given zero or the broadcast namespace,
6005 	 * check to see if it's actually a valid namespace ID. If is outside of
6006 	 * range, then it is an error. Next, if we have been requested to
6007 	 * rewrite 0 (the this controller indicator) as the broadcast namespace,
6008 	 * do so.
6009 	 *
6010 	 * While we validate that this namespace is within the valid range, we
6011 	 * do not check if it is active or inactive. That is left to our callers
6012 	 * to determine.
6013 	 */
6014 	if (ioc->nioc_nsid > minor->nm_ctrl->n_namespace_count &&
6015 	    ioc->nioc_nsid != NVME_NSID_BCAST) {
6016 		return (nvme_ioctl_error(ioc, NVME_IOCTL_E_NS_RANGE, 0, 0));
6017 	}
6018 
6019 	if (ioc->nioc_nsid == 0 && check->nck_ctrl_rewrite) {
6020 		ioc->nioc_nsid = NVME_NSID_BCAST;
6021 	}
6022 
6023 	/*
6024 	 * Finally, see if we have ended up with a broadcast namespace ID
6025 	 * whether through specification or rewriting. If that is not allowed,
6026 	 * then that is an error.
6027 	 */
6028 	if (!check->nck_bcast_ok && ioc->nioc_nsid == NVME_NSID_BCAST) {
6029 		return (nvme_ioctl_error(ioc, NVME_IOCTL_E_NO_BCAST_NS, 0, 0));
6030 	}
6031 
6032 	return (nvme_ioctl_excl_check(minor, ioc, check));
6033 }
6034 
6035 static int
6036 nvme_ioctl_ctrl_info(nvme_minor_t *minor, intptr_t arg, int mode,
6037     cred_t *cred_p)
6038 {
6039 	nvme_t *const nvme = minor->nm_ctrl;
6040 	nvme_ioctl_ctrl_info_t *info;
6041 	nvme_reg_cap_t cap = { 0 };
6042 	nvme_ioctl_identify_t id = { .nid_cns = NVME_IDENTIFY_CTRL };
6043 	void *idbuf;
6044 
6045 	if ((mode & FREAD) == 0)
6046 		return (EBADF);
6047 
6048 	info = kmem_alloc(sizeof (nvme_ioctl_ctrl_info_t), KM_NOSLEEP_LAZY);
6049 	if (info == NULL) {
6050 		return (nvme_ioctl_copyout_error(NVME_IOCTL_E_NO_KERN_MEM, arg,
6051 		    mode));
6052 	}
6053 
6054 	if (ddi_copyin((void *)arg, info, sizeof (nvme_ioctl_ctrl_info_t),
6055 	    mode & FKIOCTL) != 0) {
6056 		kmem_free(info, sizeof (nvme_ioctl_ctrl_info_t));
6057 		return (EFAULT);
6058 	}
6059 
6060 	if (!nvme_ioctl_check(minor, &info->nci_common,
6061 	    &nvme_check_ctrl_info)) {
6062 		goto copyout;
6063 	}
6064 
6065 	/*
6066 	 * We explicitly do not use the identify controller copy in the kernel
6067 	 * right now so that way we can get a snapshot of the controller's
6068 	 * current capacity and values. While it's tempting to try to use this
6069 	 * to refresh the kernel's version we don't just to simplify the rest of
6070 	 * the driver right now.
6071 	 */
6072 	if (!nvme_identify(nvme, B_TRUE, &id, &idbuf)) {
6073 		info->nci_common = id.nid_common;
6074 		goto copyout;
6075 	}
6076 	bcopy(idbuf, &info->nci_ctrl_id, sizeof (nvme_identify_ctrl_t));
6077 	kmem_free(idbuf, NVME_IDENTIFY_BUFSIZE);
6078 
6079 	/*
6080 	 * Use the kernel's cached common namespace information for this.
6081 	 */
6082 	bcopy(nvme->n_idcomns, &info->nci_common_ns,
6083 	    sizeof (nvme_identify_nsid_t));
6084 
6085 	info->nci_vers = nvme->n_version;
6086 
6087 	/*
6088 	 * The MPSMIN and MPSMAX fields in the CAP register use 0 to
6089 	 * specify the base page size of 4k (1<<12), so add 12 here to
6090 	 * get the real page size value.
6091 	 */
6092 	cap.r = nvme_get64(nvme, NVME_REG_CAP);
6093 	info->nci_caps.cap_mpsmax = 1 << (12 + cap.b.cap_mpsmax);
6094 	info->nci_caps.cap_mpsmin = 1 << (12 + cap.b.cap_mpsmin);
6095 
6096 	info->nci_nintrs = (uint32_t)nvme->n_intr_cnt;
6097 
6098 copyout:
6099 	if (ddi_copyout(info, (void *)arg, sizeof (nvme_ioctl_ctrl_info_t),
6100 	    mode & FKIOCTL) != 0) {
6101 		kmem_free(info, sizeof (nvme_ioctl_ctrl_info_t));
6102 		return (EFAULT);
6103 	}
6104 
6105 	kmem_free(info, sizeof (nvme_ioctl_ctrl_info_t));
6106 	return (0);
6107 }
6108 
6109 static int
6110 nvme_ioctl_ns_info(nvme_minor_t *minor, intptr_t arg, int mode, cred_t *cred_p)
6111 {
6112 	nvme_t *const nvme = minor->nm_ctrl;
6113 	nvme_ioctl_ns_info_t *ns_info;
6114 	nvme_namespace_t *ns;
6115 	nvme_ioctl_identify_t id = { .nid_cns = NVME_IDENTIFY_NSID };
6116 	void *idbuf;
6117 
6118 	if ((mode & FREAD) == 0)
6119 		return (EBADF);
6120 
6121 	ns_info = kmem_zalloc(sizeof (nvme_ioctl_ns_info_t), KM_NOSLEEP_LAZY);
6122 	if (ns_info == NULL) {
6123 		return (nvme_ioctl_copyout_error(NVME_IOCTL_E_NO_KERN_MEM, arg,
6124 		    mode));
6125 	}
6126 
6127 	if (ddi_copyin((void *)arg, ns_info, sizeof (nvme_ioctl_ns_info_t),
6128 	    mode & FKIOCTL) != 0) {
6129 		kmem_free(ns_info, sizeof (nvme_ioctl_ns_info_t));
6130 		return (EFAULT);
6131 	}
6132 
6133 	if (!nvme_ioctl_check(minor, &ns_info->nni_common,
6134 	    &nvme_check_ns_info)) {
6135 		goto copyout;
6136 	}
6137 
6138 	ASSERT3U(ns_info->nni_common.nioc_nsid, >, 0);
6139 	ns = nvme_nsid2ns(nvme, ns_info->nni_common.nioc_nsid);
6140 
6141 	/*
6142 	 * First fetch a fresh copy of the namespace information. Most callers
6143 	 * are using this because they will want a mostly accurate snapshot of
6144 	 * capacity and utilization.
6145 	 */
6146 	id.nid_common.nioc_nsid = ns_info->nni_common.nioc_nsid;
6147 	if (!nvme_identify(nvme, B_TRUE, &id, &idbuf)) {
6148 		ns_info->nni_common = id.nid_common;
6149 		goto copyout;
6150 	}
6151 	bcopy(idbuf, &ns_info->nni_id, sizeof (nvme_identify_nsid_t));
6152 	kmem_free(idbuf, NVME_IDENTIFY_BUFSIZE);
6153 
6154 	nvme_mgmt_lock(nvme, NVME_MGMT_LOCK_NVME);
6155 	if (ns->ns_allocated)
6156 		ns_info->nni_state |= NVME_NS_STATE_ALLOCATED;
6157 
6158 	if (ns->ns_active)
6159 		ns_info->nni_state |= NVME_NS_STATE_ACTIVE;
6160 
6161 	if (ns->ns_ignore)
6162 		ns_info->nni_state |= NVME_NS_STATE_IGNORED;
6163 
6164 	if (ns->ns_attached) {
6165 		const char *addr;
6166 
6167 		ns_info->nni_state |= NVME_NS_STATE_ATTACHED;
6168 		addr = bd_address(ns->ns_bd_hdl);
6169 		if (strlcpy(ns_info->nni_addr, addr,
6170 		    sizeof (ns_info->nni_addr)) >= sizeof (ns_info->nni_addr)) {
6171 			nvme_mgmt_unlock(nvme);
6172 			(void) nvme_ioctl_error(&ns_info->nni_common,
6173 			    NVME_IOCTL_E_BD_ADDR_OVER, 0, 0);
6174 			goto copyout;
6175 		}
6176 	}
6177 	nvme_mgmt_unlock(nvme);
6178 
6179 copyout:
6180 	if (ddi_copyout(ns_info, (void *)arg, sizeof (nvme_ioctl_ns_info_t),
6181 	    mode & FKIOCTL) != 0) {
6182 		kmem_free(ns_info, sizeof (nvme_ioctl_ns_info_t));
6183 		return (EFAULT);
6184 	}
6185 
6186 	kmem_free(ns_info, sizeof (nvme_ioctl_ns_info_t));
6187 	return (0);
6188 }
6189 
6190 static int
6191 nvme_ioctl_identify(nvme_minor_t *minor, intptr_t arg, int mode, cred_t *cred_p)
6192 {
6193 	_NOTE(ARGUNUSED(cred_p));
6194 	nvme_t *const nvme = minor->nm_ctrl;
6195 	void *idctl;
6196 	uint_t model;
6197 	nvme_ioctl_identify_t id;
6198 #ifdef	_MULTI_DATAMODEL
6199 	nvme_ioctl_identify32_t id32;
6200 #endif
6201 	boolean_t ns_minor;
6202 
6203 	if ((mode & FREAD) == 0)
6204 		return (EBADF);
6205 
6206 	model = ddi_model_convert_from(mode);
6207 	switch (model) {
6208 #ifdef	_MULTI_DATAMODEL
6209 	case DDI_MODEL_ILP32:
6210 		bzero(&id, sizeof (id));
6211 		if (ddi_copyin((void *)arg, &id32, sizeof (id32),
6212 		    mode & FKIOCTL) != 0) {
6213 			return (EFAULT);
6214 		}
6215 		id.nid_common.nioc_nsid = id32.nid_common.nioc_nsid;
6216 		id.nid_cns = id32.nid_cns;
6217 		id.nid_ctrlid = id32.nid_ctrlid;
6218 		id.nid_data = id32.nid_data;
6219 		break;
6220 #endif	/* _MULTI_DATAMODEL */
6221 	case DDI_MODEL_NONE:
6222 		if (ddi_copyin((void *)arg, &id, sizeof (id),
6223 		    mode & FKIOCTL) != 0) {
6224 			return (EFAULT);
6225 		}
6226 		break;
6227 	default:
6228 		return (ENOTSUP);
6229 	}
6230 
6231 	if (!nvme_ioctl_check(minor, &id.nid_common, &nvme_check_identify)) {
6232 		goto copyout;
6233 	}
6234 
6235 	ns_minor = minor->nm_ns != NULL;
6236 	if (!nvme_validate_identify(nvme, &id, ns_minor)) {
6237 		goto copyout;
6238 	}
6239 
6240 	if (nvme_identify(nvme, B_TRUE, &id, &idctl)) {
6241 		int ret = ddi_copyout(idctl, (void *)id.nid_data,
6242 		    NVME_IDENTIFY_BUFSIZE, mode & FKIOCTL);
6243 		kmem_free(idctl, NVME_IDENTIFY_BUFSIZE);
6244 		if (ret != 0) {
6245 			(void) nvme_ioctl_error(&id.nid_common,
6246 			    NVME_IOCTL_E_BAD_USER_DATA, 0, 0);
6247 			goto copyout;
6248 		}
6249 
6250 		nvme_ioctl_success(&id.nid_common);
6251 	}
6252 
6253 copyout:
6254 	switch (model) {
6255 #ifdef	_MULTI_DATAMODEL
6256 	case DDI_MODEL_ILP32:
6257 		id32.nid_common = id.nid_common;
6258 
6259 		if (ddi_copyout(&id32, (void *)arg, sizeof (id32),
6260 		    mode & FKIOCTL) != 0) {
6261 			return (EFAULT);
6262 		}
6263 		break;
6264 #endif	/* _MULTI_DATAMODEL */
6265 	case DDI_MODEL_NONE:
6266 		if (ddi_copyout(&id, (void *)arg, sizeof (id),
6267 		    mode & FKIOCTL) != 0) {
6268 			return (EFAULT);
6269 		}
6270 		break;
6271 	default:
6272 		return (ENOTSUP);
6273 	}
6274 
6275 	return (0);
6276 }
6277 
6278 /*
6279  * Execute commands on behalf of the various ioctls.
6280  *
6281  * If this returns true then the command completed successfully. Otherwise error
6282  * information is returned in the nvme_ioctl_common_t arguments.
6283  */
6284 typedef struct {
6285 	nvme_sqe_t *ica_sqe;
6286 	void *ica_data;
6287 	uint32_t ica_data_len;
6288 	uint_t ica_dma_flags;
6289 	int ica_copy_flags;
6290 	uint32_t ica_timeout;
6291 	uint32_t ica_cdw0;
6292 } nvme_ioc_cmd_args_t;
6293 
6294 static boolean_t
6295 nvme_ioc_cmd(nvme_t *nvme, nvme_ioctl_common_t *ioc, nvme_ioc_cmd_args_t *args)
6296 {
6297 	nvme_cmd_t *cmd;
6298 	boolean_t ret = B_FALSE;
6299 
6300 	cmd = nvme_alloc_admin_cmd(nvme, KM_SLEEP);
6301 	cmd->nc_sqid = 0;
6302 
6303 	/*
6304 	 * This function is used to facilitate requests from
6305 	 * userspace, so don't panic if the command fails. This
6306 	 * is especially true for admin passthru commands, where
6307 	 * the actual command data structure is entirely defined
6308 	 * by userspace.
6309 	 */
6310 	cmd->nc_flags |= NVME_CMD_F_DONTPANIC;
6311 
6312 	cmd->nc_callback = nvme_wakeup_cmd;
6313 	cmd->nc_sqe = *args->ica_sqe;
6314 
6315 	if ((args->ica_dma_flags & DDI_DMA_RDWR) != 0) {
6316 		if (args->ica_data == NULL) {
6317 			ret = nvme_ioctl_error(ioc, NVME_IOCTL_E_NO_DMA_MEM,
6318 			    0, 0);
6319 			goto free_cmd;
6320 		}
6321 
6322 		if (nvme_zalloc_dma(nvme, args->ica_data_len,
6323 		    args->ica_dma_flags, &nvme->n_prp_dma_attr, &cmd->nc_dma) !=
6324 		    DDI_SUCCESS) {
6325 			dev_err(nvme->n_dip, CE_WARN,
6326 			    "!nvme_zalloc_dma failed for nvme_ioc_cmd()");
6327 			ret = nvme_ioctl_error(ioc,
6328 			    NVME_IOCTL_E_NO_DMA_MEM, 0, 0);
6329 			goto free_cmd;
6330 		}
6331 
6332 		if (nvme_fill_prp(cmd, cmd->nc_dma->nd_dmah) != 0) {
6333 			ret = nvme_ioctl_error(ioc,
6334 			    NVME_IOCTL_E_NO_DMA_MEM, 0, 0);
6335 			goto free_cmd;
6336 		}
6337 
6338 		if ((args->ica_dma_flags & DDI_DMA_WRITE) != 0 &&
6339 		    ddi_copyin(args->ica_data, cmd->nc_dma->nd_memp,
6340 		    args->ica_data_len, args->ica_copy_flags) != 0) {
6341 			ret = nvme_ioctl_error(ioc, NVME_IOCTL_E_BAD_USER_DATA,
6342 			    0, 0);
6343 			goto free_cmd;
6344 		}
6345 	}
6346 
6347 	nvme_admin_cmd(cmd, args->ica_timeout);
6348 
6349 	if (!nvme_check_cmd_status_ioctl(cmd, ioc)) {
6350 		ret = B_FALSE;
6351 		goto free_cmd;
6352 	}
6353 
6354 	args->ica_cdw0 = cmd->nc_cqe.cqe_dw0;
6355 
6356 	if ((args->ica_dma_flags & DDI_DMA_READ) != 0 &&
6357 	    ddi_copyout(cmd->nc_dma->nd_memp, args->ica_data,
6358 	    args->ica_data_len, args->ica_copy_flags) != 0) {
6359 		ret = nvme_ioctl_error(ioc, NVME_IOCTL_E_BAD_USER_DATA, 0, 0);
6360 		goto free_cmd;
6361 	}
6362 
6363 	ret = B_TRUE;
6364 	nvme_ioctl_success(ioc);
6365 
6366 free_cmd:
6367 	nvme_free_cmd(cmd);
6368 
6369 	return (ret);
6370 }
6371 
6372 static int
6373 nvme_ioctl_get_logpage(nvme_minor_t *minor, intptr_t arg, int mode,
6374     cred_t *cred_p)
6375 {
6376 	nvme_t *const nvme = minor->nm_ctrl;
6377 	void *buf;
6378 	nvme_ioctl_get_logpage_t log;
6379 	uint_t model;
6380 #ifdef	_MULTI_DATAMODEL
6381 	nvme_ioctl_get_logpage32_t log32;
6382 #endif
6383 
6384 	if ((mode & FREAD) == 0) {
6385 		return (EBADF);
6386 	}
6387 
6388 	model = ddi_model_convert_from(mode);
6389 	switch (model) {
6390 #ifdef	_MULTI_DATAMODEL
6391 	case DDI_MODEL_ILP32:
6392 		bzero(&log, sizeof (log));
6393 		if (ddi_copyin((void *)arg, &log32, sizeof (log32),
6394 		    mode & FKIOCTL) != 0) {
6395 			return (EFAULT);
6396 		}
6397 
6398 		log.nigl_common.nioc_nsid = log32.nigl_common.nioc_nsid;
6399 		log.nigl_csi = log32.nigl_csi;
6400 		log.nigl_lid = log32.nigl_lid;
6401 		log.nigl_lsp = log32.nigl_lsp;
6402 		log.nigl_len = log32.nigl_len;
6403 		log.nigl_offset = log32.nigl_offset;
6404 		log.nigl_data = log32.nigl_data;
6405 		break;
6406 #endif	/* _MULTI_DATAMODEL */
6407 	case DDI_MODEL_NONE:
6408 		if (ddi_copyin((void *)arg, &log, sizeof (log),
6409 		    mode & FKIOCTL) != 0) {
6410 			return (EFAULT);
6411 		}
6412 		break;
6413 	default:
6414 		return (ENOTSUP);
6415 	}
6416 
6417 	/*
6418 	 * Eventually we'd like to do a soft lock on the namespaces from
6419 	 * changing out from us during this operation in the future. But we
6420 	 * haven't implemented that yet.
6421 	 */
6422 	if (!nvme_ioctl_check(minor, &log.nigl_common,
6423 	    &nvme_check_get_logpage)) {
6424 		goto copyout;
6425 	}
6426 
6427 	if (!nvme_validate_logpage(nvme, &log)) {
6428 		goto copyout;
6429 	}
6430 
6431 	if (nvme_get_logpage(nvme, B_TRUE, &log, &buf)) {
6432 		int copy;
6433 
6434 		copy = ddi_copyout(buf, (void *)log.nigl_data, log.nigl_len,
6435 		    mode & FKIOCTL);
6436 		kmem_free(buf, log.nigl_len);
6437 		if (copy != 0) {
6438 			(void) nvme_ioctl_error(&log.nigl_common,
6439 			    NVME_IOCTL_E_BAD_USER_DATA, 0, 0);
6440 			goto copyout;
6441 		}
6442 
6443 		nvme_ioctl_success(&log.nigl_common);
6444 	}
6445 
6446 copyout:
6447 	switch (model) {
6448 #ifdef	_MULTI_DATAMODEL
6449 	case DDI_MODEL_ILP32:
6450 		bzero(&log32, sizeof (log32));
6451 
6452 		log32.nigl_common = log.nigl_common;
6453 		log32.nigl_csi = log.nigl_csi;
6454 		log32.nigl_lid = log.nigl_lid;
6455 		log32.nigl_lsp = log.nigl_lsp;
6456 		log32.nigl_len = log.nigl_len;
6457 		log32.nigl_offset = log.nigl_offset;
6458 		log32.nigl_data = log.nigl_data;
6459 		if (ddi_copyout(&log32, (void *)arg, sizeof (log32),
6460 		    mode & FKIOCTL) != 0) {
6461 			return (EFAULT);
6462 		}
6463 		break;
6464 #endif	/* _MULTI_DATAMODEL */
6465 	case DDI_MODEL_NONE:
6466 		if (ddi_copyout(&log, (void *)arg, sizeof (log),
6467 		    mode & FKIOCTL) != 0) {
6468 			return (EFAULT);
6469 		}
6470 		break;
6471 	default:
6472 		return (ENOTSUP);
6473 	}
6474 
6475 	return (0);
6476 }
6477 
6478 static int
6479 nvme_ioctl_get_feature(nvme_minor_t *minor, intptr_t arg, int mode,
6480     cred_t *cred_p)
6481 {
6482 	nvme_t *const nvme = minor->nm_ctrl;
6483 	nvme_ioctl_get_feature_t feat;
6484 	uint_t model;
6485 #ifdef	_MULTI_DATAMODEL
6486 	nvme_ioctl_get_feature32_t feat32;
6487 #endif
6488 	nvme_get_features_dw10_t gf_dw10 = { 0 };
6489 	nvme_ioc_cmd_args_t args = { NULL };
6490 	nvme_sqe_t sqe = {
6491 	    .sqe_opc	= NVME_OPC_GET_FEATURES
6492 	};
6493 
6494 	if ((mode & FREAD) == 0) {
6495 		return (EBADF);
6496 	}
6497 
6498 	model = ddi_model_convert_from(mode);
6499 	switch (model) {
6500 #ifdef	_MULTI_DATAMODEL
6501 	case DDI_MODEL_ILP32:
6502 		bzero(&feat, sizeof (feat));
6503 		if (ddi_copyin((void *)arg, &feat32, sizeof (feat32),
6504 		    mode & FKIOCTL) != 0) {
6505 			return (EFAULT);
6506 		}
6507 
6508 		feat.nigf_common.nioc_nsid = feat32.nigf_common.nioc_nsid;
6509 		feat.nigf_fid = feat32.nigf_fid;
6510 		feat.nigf_sel = feat32.nigf_sel;
6511 		feat.nigf_cdw11 = feat32.nigf_cdw11;
6512 		feat.nigf_data = feat32.nigf_data;
6513 		feat.nigf_len = feat32.nigf_len;
6514 		break;
6515 #endif	/* _MULTI_DATAMODEL */
6516 	case DDI_MODEL_NONE:
6517 		if (ddi_copyin((void *)arg, &feat, sizeof (feat),
6518 		    mode & FKIOCTL) != 0) {
6519 			return (EFAULT);
6520 		}
6521 		break;
6522 	default:
6523 		return (ENOTSUP);
6524 	}
6525 
6526 	if (!nvme_ioctl_check(minor, &feat.nigf_common,
6527 	    &nvme_check_get_feature)) {
6528 		goto copyout;
6529 	}
6530 
6531 	if (!nvme_validate_get_feature(nvme, &feat)) {
6532 		goto copyout;
6533 	}
6534 
6535 	gf_dw10.b.gt_fid = bitx32(feat.nigf_fid, 7, 0);
6536 	gf_dw10.b.gt_sel = bitx32(feat.nigf_sel, 2, 0);
6537 	sqe.sqe_cdw10 = gf_dw10.r;
6538 	sqe.sqe_cdw11 = feat.nigf_cdw11;
6539 	sqe.sqe_nsid = feat.nigf_common.nioc_nsid;
6540 
6541 	args.ica_sqe = &sqe;
6542 	if (feat.nigf_len != 0) {
6543 		args.ica_data = (void *)feat.nigf_data;
6544 		args.ica_data_len = feat.nigf_len;
6545 		args.ica_dma_flags = DDI_DMA_READ;
6546 	}
6547 	args.ica_copy_flags = mode;
6548 	args.ica_timeout = nvme_admin_cmd_timeout;
6549 
6550 	if (!nvme_ioc_cmd(nvme, &feat.nigf_common, &args)) {
6551 		goto copyout;
6552 	}
6553 
6554 	feat.nigf_cdw0 = args.ica_cdw0;
6555 
6556 copyout:
6557 	switch (model) {
6558 #ifdef	_MULTI_DATAMODEL
6559 	case DDI_MODEL_ILP32:
6560 		bzero(&feat32, sizeof (feat32));
6561 
6562 		feat32.nigf_common = feat.nigf_common;
6563 		feat32.nigf_fid = feat.nigf_fid;
6564 		feat32.nigf_sel = feat.nigf_sel;
6565 		feat32.nigf_cdw11 = feat.nigf_cdw11;
6566 		feat32.nigf_data = feat.nigf_data;
6567 		feat32.nigf_len = feat.nigf_len;
6568 		feat32.nigf_cdw0 = feat.nigf_cdw0;
6569 		if (ddi_copyout(&feat32, (void *)arg, sizeof (feat32),
6570 		    mode & FKIOCTL) != 0) {
6571 			return (EFAULT);
6572 		}
6573 		break;
6574 #endif	/* _MULTI_DATAMODEL */
6575 	case DDI_MODEL_NONE:
6576 		if (ddi_copyout(&feat, (void *)arg, sizeof (feat),
6577 		    mode & FKIOCTL) != 0) {
6578 			return (EFAULT);
6579 		}
6580 		break;
6581 	default:
6582 		return (ENOTSUP);
6583 	}
6584 
6585 	return (0);
6586 }
6587 
6588 static int
6589 nvme_ioctl_format(nvme_minor_t *minor, intptr_t arg, int mode, cred_t *cred_p)
6590 {
6591 	nvme_t *const nvme = minor->nm_ctrl;
6592 	nvme_ioctl_format_t ioc;
6593 
6594 	if ((mode & FWRITE) == 0)
6595 		return (EBADF);
6596 
6597 	if (secpolicy_sys_config(cred_p, B_FALSE) != 0)
6598 		return (EPERM);
6599 
6600 	if (ddi_copyin((void *)(uintptr_t)arg, &ioc,
6601 	    sizeof (nvme_ioctl_format_t), mode & FKIOCTL) != 0)
6602 		return (EFAULT);
6603 
6604 	if (!nvme_ioctl_check(minor, &ioc.nif_common, &nvme_check_format)) {
6605 		goto copyout;
6606 	}
6607 
6608 	if (!nvme_validate_format(nvme, &ioc)) {
6609 		goto copyout;
6610 	}
6611 
6612 	nvme_mgmt_lock(nvme, NVME_MGMT_LOCK_NVME);
6613 	if (!nvme_no_blkdev_attached(nvme, ioc.nif_common.nioc_nsid)) {
6614 		nvme_mgmt_unlock(nvme);
6615 		(void) nvme_ioctl_error(&ioc.nif_common,
6616 		    NVME_IOCTL_E_NS_BLKDEV_ATTACH, 0, 0);
6617 		goto copyout;
6618 	}
6619 
6620 	if (nvme_format_nvm(nvme, &ioc)) {
6621 		nvme_ioctl_success(&ioc.nif_common);
6622 		nvme_rescan_ns(nvme, ioc.nif_common.nioc_nsid);
6623 	}
6624 	nvme_mgmt_unlock(nvme);
6625 
6626 copyout:
6627 	if (ddi_copyout(&ioc, (void *)(uintptr_t)arg, sizeof (ioc),
6628 	    mode & FKIOCTL) != 0) {
6629 		return (EFAULT);
6630 	}
6631 
6632 	return (0);
6633 }
6634 
6635 static int
6636 nvme_ioctl_detach(nvme_minor_t *minor, intptr_t arg, int mode, cred_t *cred_p)
6637 {
6638 	nvme_t *const nvme = minor->nm_ctrl;
6639 	nvme_ioctl_common_t com;
6640 
6641 	if ((mode & FWRITE) == 0)
6642 		return (EBADF);
6643 
6644 	if (secpolicy_sys_config(cred_p, B_FALSE) != 0)
6645 		return (EPERM);
6646 
6647 	if (ddi_copyin((void *)(uintptr_t)arg, &com, sizeof (com),
6648 	    mode & FKIOCTL) != 0) {
6649 		return (EFAULT);
6650 	}
6651 
6652 	if (!nvme_ioctl_check(minor, &com, &nvme_check_attach_detach)) {
6653 		goto copyout;
6654 	}
6655 
6656 	nvme_mgmt_lock(nvme, NVME_MGMT_LOCK_NVME);
6657 	if (nvme_detach_ns(nvme, &com)) {
6658 		nvme_ioctl_success(&com);
6659 	}
6660 	nvme_mgmt_unlock(nvme);
6661 
6662 copyout:
6663 	if (ddi_copyout(&com, (void *)(uintptr_t)arg, sizeof (com),
6664 	    mode & FKIOCTL) != 0) {
6665 		return (EFAULT);
6666 	}
6667 
6668 	return (0);
6669 }
6670 
6671 static int
6672 nvme_ioctl_attach(nvme_minor_t *minor, intptr_t arg, int mode,
6673     cred_t *cred_p)
6674 {
6675 	nvme_t *const nvme = minor->nm_ctrl;
6676 	nvme_ioctl_common_t com;
6677 	nvme_namespace_t *ns;
6678 
6679 	if ((mode & FWRITE) == 0)
6680 		return (EBADF);
6681 
6682 	if (secpolicy_sys_config(cred_p, B_FALSE) != 0)
6683 		return (EPERM);
6684 
6685 	if (ddi_copyin((void *)(uintptr_t)arg, &com, sizeof (com),
6686 	    mode & FKIOCTL) != 0) {
6687 		return (EFAULT);
6688 	}
6689 
6690 	if (!nvme_ioctl_check(minor, &com, &nvme_check_attach_detach)) {
6691 		goto copyout;
6692 	}
6693 
6694 	nvme_mgmt_lock(nvme, NVME_MGMT_LOCK_NVME);
6695 	ns = nvme_nsid2ns(nvme, com.nioc_nsid);
6696 
6697 	/*
6698 	 * Strictly speaking we shouldn't need to call nvme_init_ns() here as
6699 	 * we should be properly refreshing the internal state when we are
6700 	 * issuing commands that change things. However, we opt to still do so
6701 	 * as a bit of a safety check lest we give the kernel something bad or a
6702 	 * vendor unique command somehow did something behind our backs.
6703 	 */
6704 	if (!ns->ns_attached) {
6705 		(void) nvme_rescan_ns(nvme, com.nioc_nsid);
6706 		if (nvme_attach_ns(nvme, &com)) {
6707 			nvme_ioctl_success(&com);
6708 		}
6709 	} else {
6710 		nvme_ioctl_success(&com);
6711 	}
6712 	nvme_mgmt_unlock(nvme);
6713 
6714 copyout:
6715 	if (ddi_copyout(&com, (void *)(uintptr_t)arg, sizeof (com),
6716 	    mode & FKIOCTL) != 0) {
6717 		return (EFAULT);
6718 	}
6719 
6720 	return (0);
6721 }
6722 
6723 static void
6724 nvme_ufm_update(nvme_t *nvme)
6725 {
6726 	mutex_enter(&nvme->n_fwslot_mutex);
6727 	ddi_ufm_update(nvme->n_ufmh);
6728 	if (nvme->n_fwslot != NULL) {
6729 		kmem_free(nvme->n_fwslot, sizeof (nvme_fwslot_log_t));
6730 		nvme->n_fwslot = NULL;
6731 	}
6732 	mutex_exit(&nvme->n_fwslot_mutex);
6733 }
6734 
6735 /*
6736  * Download new firmware to the device's internal staging area. We do not call
6737  * nvme_ufm_update() here because after a firmware download, there has been no
6738  * change to any of the actual persistent firmware data. That requires a
6739  * subsequent ioctl (NVME_IOC_FIRMWARE_COMMIT) to commit the firmware to a slot
6740  * or to activate a slot.
6741  */
6742 static int
6743 nvme_ioctl_firmware_download(nvme_minor_t *minor, intptr_t arg, int mode,
6744     cred_t *cred_p)
6745 {
6746 	nvme_t *const nvme = minor->nm_ctrl;
6747 	nvme_ioctl_fw_load_t fw;
6748 	uint64_t len, maxcopy;
6749 	offset_t offset;
6750 	uint32_t gran;
6751 	nvme_valid_ctrl_data_t data;
6752 	uintptr_t buf;
6753 	nvme_sqe_t sqe = {
6754 	    .sqe_opc	= NVME_OPC_FW_IMAGE_LOAD
6755 	};
6756 
6757 	if ((mode & FWRITE) == 0)
6758 		return (EBADF);
6759 
6760 	if (secpolicy_sys_config(cred_p, B_FALSE) != 0)
6761 		return (EPERM);
6762 
6763 	if (ddi_copyin((void *)(uintptr_t)arg, &fw, sizeof (fw),
6764 	    mode & FKIOCTL) != 0) {
6765 		return (EFAULT);
6766 	}
6767 
6768 	if (!nvme_ioctl_check(minor, &fw.fwl_common, &nvme_check_firmware)) {
6769 		goto copyout;
6770 	}
6771 
6772 	if (!nvme_validate_fw_load(nvme, &fw)) {
6773 		goto copyout;
6774 	}
6775 
6776 	len = fw.fwl_len;
6777 	offset = fw.fwl_off;
6778 	buf = fw.fwl_buf;
6779 
6780 	/*
6781 	 * We need to determine the minimum and maximum amount of data that we
6782 	 * will send to the device in a given go. Starting in NMVe 1.3 this must
6783 	 * be a multiple of the firmware update granularity (FWUG), but must not
6784 	 * exceed the maximum data transfer that we've set. Many devices don't
6785 	 * report something here, which means we'll end up getting our default
6786 	 * value. Our policy is a little simple, but it's basically if the
6787 	 * maximum data transfer is evenly divided by the granularity, then use
6788 	 * it. Otherwise we use the granularity itself. The granularity is
6789 	 * always in page sized units, so trying to find another optimum point
6790 	 * isn't worth it. If we encounter a contradiction, then we will have to
6791 	 * error out.
6792 	 */
6793 	data.vcd_vers = &nvme->n_version;
6794 	data.vcd_id = nvme->n_idctl;
6795 	gran = nvme_fw_load_granularity(&data);
6796 
6797 	if ((nvme->n_max_data_transfer_size % gran) == 0) {
6798 		maxcopy = nvme->n_max_data_transfer_size;
6799 	} else if (gran <= nvme->n_max_data_transfer_size) {
6800 		maxcopy = gran;
6801 	} else {
6802 		(void) nvme_ioctl_error(&fw.fwl_common,
6803 		    NVME_IOCTL_E_FW_LOAD_IMPOS_GRAN, 0, 0);
6804 		goto copyout;
6805 	}
6806 
6807 	while (len > 0) {
6808 		nvme_ioc_cmd_args_t args = { NULL };
6809 		uint64_t copylen = MIN(maxcopy, len);
6810 
6811 		sqe.sqe_cdw10 = (uint32_t)(copylen >> NVME_DWORD_SHIFT) - 1;
6812 		sqe.sqe_cdw11 = (uint32_t)(offset >> NVME_DWORD_SHIFT);
6813 
6814 		args.ica_sqe = &sqe;
6815 		args.ica_data = (void *)buf;
6816 		args.ica_data_len = copylen;
6817 		args.ica_dma_flags = DDI_DMA_WRITE;
6818 		args.ica_copy_flags = mode;
6819 		args.ica_timeout = nvme_admin_cmd_timeout;
6820 
6821 		if (!nvme_ioc_cmd(nvme, &fw.fwl_common, &args)) {
6822 			break;
6823 		}
6824 
6825 		buf += copylen;
6826 		offset += copylen;
6827 		len -= copylen;
6828 	}
6829 
6830 copyout:
6831 	if (ddi_copyout(&fw, (void *)(uintptr_t)arg, sizeof (fw),
6832 	    mode & FKIOCTL) != 0) {
6833 		return (EFAULT);
6834 	}
6835 
6836 	return (0);
6837 }
6838 
6839 static int
6840 nvme_ioctl_firmware_commit(nvme_minor_t *minor, intptr_t arg, int mode,
6841     cred_t *cred_p)
6842 {
6843 	nvme_t *const nvme = minor->nm_ctrl;
6844 	nvme_ioctl_fw_commit_t fw;
6845 	nvme_firmware_commit_dw10_t fc_dw10 = { 0 };
6846 	nvme_ioc_cmd_args_t args = { NULL };
6847 	nvme_sqe_t sqe = {
6848 	    .sqe_opc	= NVME_OPC_FW_ACTIVATE
6849 	};
6850 
6851 	if ((mode & FWRITE) == 0)
6852 		return (EBADF);
6853 
6854 	if (secpolicy_sys_config(cred_p, B_FALSE) != 0)
6855 		return (EPERM);
6856 
6857 	if (ddi_copyin((void *)(uintptr_t)arg, &fw, sizeof (fw),
6858 	    mode & FKIOCTL) != 0) {
6859 		return (EFAULT);
6860 	}
6861 
6862 	if (!nvme_ioctl_check(minor, &fw.fwc_common, &nvme_check_firmware)) {
6863 		goto copyout;
6864 	}
6865 
6866 	if (!nvme_validate_fw_commit(nvme, &fw)) {
6867 		goto copyout;
6868 	}
6869 
6870 	fc_dw10.b.fc_slot = fw.fwc_slot;
6871 	fc_dw10.b.fc_action = fw.fwc_action;
6872 	sqe.sqe_cdw10 = fc_dw10.r;
6873 
6874 	args.ica_sqe = &sqe;
6875 	args.ica_timeout = nvme_commit_save_cmd_timeout;
6876 
6877 	/*
6878 	 * There are no conditional actions to take based on this succeeding or
6879 	 * failing. A failure is recorded in the ioctl structure returned to the
6880 	 * user.
6881 	 */
6882 	(void) nvme_ioc_cmd(nvme, &fw.fwc_common, &args);
6883 
6884 	/*
6885 	 * Let the DDI UFM subsystem know that the firmware information for
6886 	 * this device has changed. We perform this unconditionally as an
6887 	 * invalidation doesn't particularly hurt us.
6888 	 */
6889 	nvme_ufm_update(nvme);
6890 
6891 copyout:
6892 	if (ddi_copyout(&fw, (void *)(uintptr_t)arg, sizeof (fw),
6893 	    mode & FKIOCTL) != 0) {
6894 		return (EFAULT);
6895 	}
6896 
6897 	return (0);
6898 }
6899 
6900 /*
6901  * Helper to copy in a passthru command from userspace, handling
6902  * different data models.
6903  */
6904 static int
6905 nvme_passthru_copyin_cmd(const void *buf, nvme_ioctl_passthru_t *cmd, int mode)
6906 {
6907 	switch (ddi_model_convert_from(mode & FMODELS)) {
6908 #ifdef _MULTI_DATAMODEL
6909 	case DDI_MODEL_ILP32: {
6910 		nvme_ioctl_passthru32_t cmd32;
6911 
6912 		if (ddi_copyin(buf, (void*)&cmd32, sizeof (cmd32), mode) != 0)
6913 			return (EFAULT);
6914 
6915 		bzero(cmd, sizeof (nvme_ioctl_passthru_t));
6916 
6917 		cmd->npc_common.nioc_nsid = cmd32.npc_common.nioc_nsid;
6918 		cmd->npc_opcode = cmd32.npc_opcode;
6919 		cmd->npc_timeout = cmd32.npc_timeout;
6920 		cmd->npc_flags = cmd32.npc_flags;
6921 		cmd->npc_impact = cmd32.npc_impact;
6922 		cmd->npc_cdw12 = cmd32.npc_cdw12;
6923 		cmd->npc_cdw13 = cmd32.npc_cdw13;
6924 		cmd->npc_cdw14 = cmd32.npc_cdw14;
6925 		cmd->npc_cdw15 = cmd32.npc_cdw15;
6926 		cmd->npc_buflen = cmd32.npc_buflen;
6927 		cmd->npc_buf = cmd32.npc_buf;
6928 		break;
6929 	}
6930 #endif	/* _MULTI_DATAMODEL */
6931 	case DDI_MODEL_NONE:
6932 		if (ddi_copyin(buf, (void *)cmd, sizeof (nvme_ioctl_passthru_t),
6933 		    mode) != 0) {
6934 			return (EFAULT);
6935 		}
6936 		break;
6937 	default:
6938 		return (ENOTSUP);
6939 	}
6940 
6941 	return (0);
6942 }
6943 
6944 /*
6945  * Helper to copy out a passthru command result to userspace, handling
6946  * different data models.
6947  */
6948 static int
6949 nvme_passthru_copyout_cmd(const nvme_ioctl_passthru_t *cmd, void *buf, int mode)
6950 {
6951 	switch (ddi_model_convert_from(mode & FMODELS)) {
6952 #ifdef _MULTI_DATAMODEL
6953 	case DDI_MODEL_ILP32: {
6954 		nvme_ioctl_passthru32_t cmd32;
6955 
6956 		bzero(&cmd32, sizeof (nvme_ioctl_passthru32_t));
6957 
6958 		cmd32.npc_common = cmd->npc_common;
6959 		cmd32.npc_opcode = cmd->npc_opcode;
6960 		cmd32.npc_timeout = cmd->npc_timeout;
6961 		cmd32.npc_flags = cmd->npc_flags;
6962 		cmd32.npc_impact = cmd->npc_impact;
6963 		cmd32.npc_cdw0 = cmd->npc_cdw0;
6964 		cmd32.npc_cdw12 = cmd->npc_cdw12;
6965 		cmd32.npc_cdw13 = cmd->npc_cdw13;
6966 		cmd32.npc_cdw14 = cmd->npc_cdw14;
6967 		cmd32.npc_cdw15 = cmd->npc_cdw15;
6968 		cmd32.npc_buflen = (size32_t)cmd->npc_buflen;
6969 		cmd32.npc_buf = (uintptr32_t)cmd->npc_buf;
6970 		if (ddi_copyout(&cmd32, buf, sizeof (cmd32), mode) != 0)
6971 			return (EFAULT);
6972 		break;
6973 	}
6974 #endif	/* _MULTI_DATAMODEL */
6975 	case DDI_MODEL_NONE:
6976 		if (ddi_copyout(cmd, buf, sizeof (nvme_ioctl_passthru_t),
6977 		    mode) != 0) {
6978 			return (EFAULT);
6979 		}
6980 		break;
6981 	default:
6982 		return (ENOTSUP);
6983 	}
6984 	return (0);
6985 }
6986 
6987 /*
6988  * Run an arbitrary vendor-specific admin command on the device.
6989  */
6990 static int
6991 nvme_ioctl_passthru(nvme_minor_t *minor, intptr_t arg, int mode, cred_t *cred_p)
6992 {
6993 	nvme_t *const nvme = minor->nm_ctrl;
6994 	int rv;
6995 	nvme_ioctl_passthru_t pass;
6996 	nvme_sqe_t sqe;
6997 	nvme_ioc_cmd_args_t args = { NULL };
6998 
6999 	/*
7000 	 * Basic checks: permissions, data model, argument size.
7001 	 */
7002 	if ((mode & FWRITE) == 0)
7003 		return (EBADF);
7004 
7005 	if (secpolicy_sys_config(cred_p, B_FALSE) != 0)
7006 		return (EPERM);
7007 
7008 	if ((rv = nvme_passthru_copyin_cmd((void *)(uintptr_t)arg, &pass,
7009 	    mode)) != 0) {
7010 		return (rv);
7011 	}
7012 
7013 	if (!nvme_ioctl_check(minor, &pass.npc_common, &nvme_check_passthru)) {
7014 		goto copyout;
7015 	}
7016 
7017 	if (!nvme_validate_vuc(nvme, &pass)) {
7018 		goto copyout;
7019 	}
7020 
7021 	nvme_mgmt_lock(nvme, NVME_MGMT_LOCK_NVME);
7022 	if ((pass.npc_impact & NVME_IMPACT_NS) != 0) {
7023 		/*
7024 		 * We've been told this has ns impact. Right now force that to
7025 		 * be every ns until we have more use cases and reason to trust
7026 		 * the nsid field.
7027 		 */
7028 		if (!nvme_no_blkdev_attached(nvme, NVME_NSID_BCAST)) {
7029 			nvme_mgmt_unlock(nvme);
7030 			(void) nvme_ioctl_error(&pass.npc_common,
7031 			    NVME_IOCTL_E_NS_BLKDEV_ATTACH, 0, 0);
7032 			goto copyout;
7033 		}
7034 	}
7035 
7036 	bzero(&sqe, sizeof (sqe));
7037 
7038 	sqe.sqe_opc = pass.npc_opcode;
7039 	sqe.sqe_nsid = pass.npc_common.nioc_nsid;
7040 	sqe.sqe_cdw10 = (uint32_t)(pass.npc_buflen >> NVME_DWORD_SHIFT);
7041 	sqe.sqe_cdw12 = pass.npc_cdw12;
7042 	sqe.sqe_cdw13 = pass.npc_cdw13;
7043 	sqe.sqe_cdw14 = pass.npc_cdw14;
7044 	sqe.sqe_cdw15 = pass.npc_cdw15;
7045 
7046 	args.ica_sqe = &sqe;
7047 	args.ica_data = (void *)pass.npc_buf;
7048 	args.ica_data_len = pass.npc_buflen;
7049 	args.ica_copy_flags = mode;
7050 	args.ica_timeout = pass.npc_timeout;
7051 
7052 	if ((pass.npc_flags & NVME_PASSTHRU_READ) != 0)
7053 		args.ica_dma_flags |= DDI_DMA_READ;
7054 	else if ((pass.npc_flags & NVME_PASSTHRU_WRITE) != 0)
7055 		args.ica_dma_flags |= DDI_DMA_WRITE;
7056 
7057 	if (nvme_ioc_cmd(nvme, &pass.npc_common, &args)) {
7058 		pass.npc_cdw0 = args.ica_cdw0;
7059 		if ((pass.npc_impact & NVME_IMPACT_NS) != 0) {
7060 			nvme_rescan_ns(nvme, NVME_NSID_BCAST);
7061 		}
7062 	}
7063 	nvme_mgmt_unlock(nvme);
7064 
7065 copyout:
7066 	rv = nvme_passthru_copyout_cmd(&pass, (void *)(uintptr_t)arg,
7067 	    mode);
7068 
7069 	return (rv);
7070 }
7071 
7072 static int
7073 nvme_ioctl_lock(nvme_minor_t *minor, intptr_t arg, int mode,
7074     cred_t *cred_p)
7075 {
7076 	nvme_ioctl_lock_t lock;
7077 	const nvme_lock_flags_t all_flags = NVME_LOCK_F_DONT_BLOCK;
7078 	nvme_t *nvme = minor->nm_ctrl;
7079 
7080 	if ((mode & FWRITE) == 0)
7081 		return (EBADF);
7082 
7083 	if (secpolicy_sys_config(cred_p, B_FALSE) != 0)
7084 		return (EPERM);
7085 
7086 	if (ddi_copyin((void *)(uintptr_t)arg, &lock, sizeof (lock),
7087 	    mode & FKIOCTL) != 0) {
7088 		return (EFAULT);
7089 	}
7090 
7091 	if (lock.nil_ent != NVME_LOCK_E_CTRL &&
7092 	    lock.nil_ent != NVME_LOCK_E_NS) {
7093 		(void) nvme_ioctl_error(&lock.nil_common,
7094 		    NVME_IOCTL_E_BAD_LOCK_ENTITY, 0, 0);
7095 		goto copyout;
7096 	}
7097 
7098 	if (lock.nil_level != NVME_LOCK_L_READ &&
7099 	    lock.nil_level != NVME_LOCK_L_WRITE) {
7100 		(void) nvme_ioctl_error(&lock.nil_common,
7101 		    NVME_IOCTL_E_BAD_LOCK_LEVEL, 0, 0);
7102 		goto copyout;
7103 	}
7104 
7105 	if ((lock.nil_flags & ~all_flags) != 0) {
7106 		(void) nvme_ioctl_error(&lock.nil_common,
7107 		    NVME_IOCTL_E_BAD_LOCK_FLAGS, 0, 0);
7108 		goto copyout;
7109 	}
7110 
7111 	if (!nvme_ioctl_check(minor, &lock.nil_common, &nvme_check_locking)) {
7112 		goto copyout;
7113 	}
7114 
7115 	/*
7116 	 * If we're on a namespace, confirm that we're not asking for the
7117 	 * controller.
7118 	 */
7119 	if (lock.nil_common.nioc_nsid != 0 &&
7120 	    lock.nil_ent == NVME_LOCK_E_CTRL) {
7121 		(void) nvme_ioctl_error(&lock.nil_common,
7122 		    NVME_IOCTL_E_NS_CANNOT_LOCK_CTRL, 0, 0);
7123 		goto copyout;
7124 	}
7125 
7126 	/*
7127 	 * We've reached the point where we can no longer actually check things
7128 	 * without serializing state. First, we need to check to make sure that
7129 	 * none of our invariants are being broken for locking:
7130 	 *
7131 	 * 1) The caller isn't already blocking for a lock operation to
7132 	 * complete.
7133 	 *
7134 	 * 2) The caller is attempting to grab a lock that they already have.
7135 	 * While there are other rule violations that this might create, we opt
7136 	 * to check this ahead of it so we can have slightly better error
7137 	 * messages for our callers.
7138 	 *
7139 	 * 3) The caller is trying to grab a controller lock, while holding a
7140 	 * namespace lock.
7141 	 *
7142 	 * 4) The caller has a controller write lock and is trying to get a
7143 	 * namespace lock. For now, we disallow this case. Holding a controller
7144 	 * read lock is allowed, but the write lock allows you to operate on all
7145 	 * namespaces anyways. In addition, this simplifies the locking logic;
7146 	 * however, this constraint may be loosened in the future.
7147 	 *
7148 	 * 5) The caller is trying to acquire a second namespace lock when they
7149 	 * already have one.
7150 	 */
7151 	mutex_enter(&nvme->n_minor_mutex);
7152 	if (minor->nm_ctrl_lock.nli_state == NVME_LOCK_STATE_BLOCKED ||
7153 	    minor->nm_ns_lock.nli_state == NVME_LOCK_STATE_BLOCKED) {
7154 		(void) nvme_ioctl_error(&lock.nil_common,
7155 		    NVME_IOCTL_E_LOCK_PENDING, 0, 0);
7156 		mutex_exit(&nvme->n_minor_mutex);
7157 		goto copyout;
7158 	}
7159 
7160 	if ((lock.nil_ent == NVME_LOCK_E_CTRL &&
7161 	    minor->nm_ctrl_lock.nli_state == NVME_LOCK_STATE_ACQUIRED) ||
7162 	    (lock.nil_ent == NVME_LOCK_E_NS &&
7163 	    minor->nm_ns_lock.nli_state == NVME_LOCK_STATE_ACQUIRED &&
7164 	    minor->nm_ns_lock.nli_ns->ns_id == lock.nil_common.nioc_nsid)) {
7165 		(void) nvme_ioctl_error(&lock.nil_common,
7166 		    NVME_IOCTL_E_LOCK_ALREADY_HELD, 0, 0);
7167 		mutex_exit(&nvme->n_minor_mutex);
7168 		goto copyout;
7169 	}
7170 
7171 	if (lock.nil_ent == NVME_LOCK_E_CTRL &&
7172 	    minor->nm_ns_lock.nli_state != NVME_LOCK_STATE_UNLOCKED) {
7173 		(void) nvme_ioctl_error(&lock.nil_common,
7174 		    NVME_IOCTL_E_LOCK_NO_CTRL_WITH_NS, 0, 0);
7175 		mutex_exit(&nvme->n_minor_mutex);
7176 		goto copyout;
7177 	}
7178 
7179 	if (lock.nil_ent == NVME_LOCK_E_NS &&
7180 	    (minor->nm_ctrl_lock.nli_state == NVME_LOCK_STATE_ACQUIRED &&
7181 	    minor->nm_ctrl_lock.nli_curlevel == NVME_LOCK_L_WRITE)) {
7182 		(void) nvme_ioctl_error(&lock.nil_common,
7183 		    NVME_IOCTL_LOCK_NO_NS_WITH_CTRL_WRLOCK, 0, 0);
7184 		mutex_exit(&nvme->n_minor_mutex);
7185 		goto copyout;
7186 	}
7187 
7188 	if (lock.nil_ent == NVME_LOCK_E_NS &&
7189 	    minor->nm_ns_lock.nli_state != NVME_LOCK_STATE_UNLOCKED) {
7190 		(void) nvme_ioctl_error(&lock.nil_common,
7191 		    NVME_IOCTL_E_LOCK_NO_2ND_NS, 0, 0);
7192 		mutex_exit(&nvme->n_minor_mutex);
7193 		goto copyout;
7194 	}
7195 
7196 #ifdef	DEBUG
7197 	/*
7198 	 * This is a big block of sanity checks to make sure that we haven't
7199 	 * allowed anything bad to happen.
7200 	 */
7201 	if (lock.nil_ent == NVME_LOCK_E_NS) {
7202 		ASSERT3P(minor->nm_ns_lock.nli_lock, ==, NULL);
7203 		ASSERT3U(minor->nm_ns_lock.nli_state, ==,
7204 		    NVME_LOCK_STATE_UNLOCKED);
7205 		ASSERT3U(minor->nm_ns_lock.nli_curlevel, ==, 0);
7206 		ASSERT3P(minor->nm_ns_lock.nli_ns, ==, NULL);
7207 
7208 		if (minor->nm_ns != NULL) {
7209 			ASSERT3U(minor->nm_ns->ns_id, ==,
7210 			    lock.nil_common.nioc_nsid);
7211 		}
7212 
7213 		ASSERT0(list_link_active(&minor->nm_ns_lock.nli_node));
7214 	} else {
7215 		ASSERT3P(minor->nm_ctrl_lock.nli_lock, ==, NULL);
7216 		ASSERT3U(minor->nm_ctrl_lock.nli_state, ==,
7217 		    NVME_LOCK_STATE_UNLOCKED);
7218 		ASSERT3U(minor->nm_ctrl_lock.nli_curlevel, ==, 0);
7219 		ASSERT3P(minor->nm_ns_lock.nli_ns, ==, NULL);
7220 		ASSERT0(list_link_active(&minor->nm_ctrl_lock.nli_node));
7221 
7222 		ASSERT3P(minor->nm_ns_lock.nli_lock, ==, NULL);
7223 		ASSERT3U(minor->nm_ns_lock.nli_state, ==,
7224 		    NVME_LOCK_STATE_UNLOCKED);
7225 		ASSERT3U(minor->nm_ns_lock.nli_curlevel, ==, 0);
7226 		ASSERT3P(minor->nm_ns_lock.nli_ns, ==, NULL);
7227 		ASSERT0(list_link_active(&minor->nm_ns_lock.nli_node));
7228 	}
7229 #endif	/* DEBUG */
7230 
7231 	/*
7232 	 * At this point we should actually attempt a locking operation.
7233 	 */
7234 	nvme_rwlock(minor, &lock);
7235 	mutex_exit(&nvme->n_minor_mutex);
7236 
7237 copyout:
7238 	if (ddi_copyout(&lock, (void *)(uintptr_t)arg, sizeof (lock),
7239 	    mode & FKIOCTL) != 0) {
7240 		return (EFAULT);
7241 	}
7242 
7243 	return (0);
7244 }
7245 
7246 static int
7247 nvme_ioctl_unlock(nvme_minor_t *minor, intptr_t arg, int mode,
7248     cred_t *cred_p)
7249 {
7250 	nvme_ioctl_unlock_t unlock;
7251 	nvme_t *const nvme = minor->nm_ctrl;
7252 	boolean_t is_ctrl;
7253 	nvme_lock_t *lock;
7254 	nvme_minor_lock_info_t *info;
7255 
7256 	/*
7257 	 * Note, we explicitly don't check for privileges for unlock. The idea
7258 	 * being that if you have the lock, that's what matters. If you don't
7259 	 * have the lock, it doesn't matter what privileges that you have at
7260 	 * all.
7261 	 */
7262 	if ((mode & FWRITE) == 0)
7263 		return (EBADF);
7264 
7265 	if (ddi_copyin((void *)(uintptr_t)arg, &unlock, sizeof (unlock),
7266 	    mode & FKIOCTL) != 0) {
7267 		return (EFAULT);
7268 	}
7269 
7270 	if (unlock.niu_ent != NVME_LOCK_E_CTRL &&
7271 	    unlock.niu_ent != NVME_LOCK_E_NS) {
7272 		(void) nvme_ioctl_error(&unlock.niu_common,
7273 		    NVME_IOCTL_E_BAD_LOCK_ENTITY, 0, 0);
7274 		goto copyout;
7275 	}
7276 
7277 	if (!nvme_ioctl_check(minor, &unlock.niu_common, &nvme_check_locking)) {
7278 		goto copyout;
7279 	}
7280 
7281 	/*
7282 	 * If we're on a namespace, confirm that we're not asking for the
7283 	 * controller.
7284 	 */
7285 	if (unlock.niu_common.nioc_nsid != 0 &&
7286 	    unlock.niu_ent == NVME_LOCK_E_CTRL) {
7287 		(void) nvme_ioctl_error(&unlock.niu_common,
7288 		    NVME_IOCTL_E_NS_CANNOT_UNLOCK_CTRL, 0, 0);
7289 		goto copyout;
7290 	}
7291 
7292 	mutex_enter(&nvme->n_minor_mutex);
7293 	if (unlock.niu_ent == NVME_LOCK_E_CTRL) {
7294 		if (minor->nm_ctrl_lock.nli_state != NVME_LOCK_STATE_ACQUIRED) {
7295 			mutex_exit(&nvme->n_minor_mutex);
7296 			(void) nvme_ioctl_error(&unlock.niu_common,
7297 			    NVME_IOCTL_E_LOCK_NOT_HELD, 0, 0);
7298 			goto copyout;
7299 		}
7300 	} else {
7301 		if (minor->nm_ns_lock.nli_ns == NULL) {
7302 			mutex_exit(&nvme->n_minor_mutex);
7303 			(void) nvme_ioctl_error(&unlock.niu_common,
7304 			    NVME_IOCTL_E_LOCK_NOT_HELD, 0, 0);
7305 			goto copyout;
7306 		}
7307 
7308 		/*
7309 		 * Check that our unlock request corresponds to the namespace ID
7310 		 * that is currently locked. This could happen if we're using
7311 		 * the controller node and it specified a valid, but not locked,
7312 		 * namespace ID.
7313 		 */
7314 		if (minor->nm_ns_lock.nli_ns->ns_id !=
7315 		    unlock.niu_common.nioc_nsid) {
7316 			mutex_exit(&nvme->n_minor_mutex);
7317 			ASSERT3P(minor->nm_ns, ==, NULL);
7318 			(void) nvme_ioctl_error(&unlock.niu_common,
7319 			    NVME_IOCTL_E_LOCK_WRONG_NS, 0, 0);
7320 			goto copyout;
7321 		}
7322 
7323 		if (minor->nm_ns_lock.nli_state != NVME_LOCK_STATE_ACQUIRED) {
7324 			mutex_exit(&nvme->n_minor_mutex);
7325 			(void) nvme_ioctl_error(&unlock.niu_common,
7326 			    NVME_IOCTL_E_LOCK_NOT_HELD, 0, 0);
7327 			goto copyout;
7328 		}
7329 	}
7330 
7331 	/*
7332 	 * Finally, perform the unlock.
7333 	 */
7334 	is_ctrl = unlock.niu_ent == NVME_LOCK_E_CTRL;
7335 	if (is_ctrl) {
7336 		lock = &nvme->n_lock;
7337 		info = &minor->nm_ctrl_lock;
7338 	} else {
7339 		nvme_namespace_t *ns;
7340 		const uint32_t nsid = unlock.niu_common.nioc_nsid;
7341 
7342 		ns = nvme_nsid2ns(nvme, nsid);
7343 		lock = &ns->ns_lock;
7344 		info = &minor->nm_ns_lock;
7345 		VERIFY3P(ns, ==, info->nli_ns);
7346 	}
7347 	nvme_rwunlock(info, lock);
7348 	mutex_exit(&nvme->n_minor_mutex);
7349 	nvme_ioctl_success(&unlock.niu_common);
7350 
7351 copyout:
7352 	if (ddi_copyout(&unlock, (void *)(uintptr_t)arg, sizeof (unlock),
7353 	    mode & FKIOCTL) != 0) {
7354 		return (EFAULT);
7355 	}
7356 
7357 	return (0);
7358 }
7359 
7360 static int
7361 nvme_ioctl(dev_t dev, int cmd, intptr_t arg, int mode, cred_t *cred_p,
7362     int *rval_p)
7363 {
7364 #ifndef __lock_lint
7365 	_NOTE(ARGUNUSED(rval_p));
7366 #endif
7367 	nvme_minor_t *minor;
7368 	nvme_t *nvme;
7369 
7370 	minor = nvme_minor_find_by_dev(dev);
7371 	if (minor == NULL) {
7372 		return (ENXIO);
7373 	}
7374 
7375 	nvme = minor->nm_ctrl;
7376 	if (nvme == NULL)
7377 		return (ENXIO);
7378 
7379 	if (IS_DEVCTL(cmd))
7380 		return (ndi_devctl_ioctl(nvme->n_dip, cmd, arg, mode, 0));
7381 
7382 	if (nvme->n_dead && (cmd != NVME_IOC_DETACH && cmd !=
7383 	    NVME_IOC_UNLOCK)) {
7384 		if (IS_NVME_IOC(cmd) == 0) {
7385 			return (EIO);
7386 		}
7387 
7388 		return (nvme_ioctl_copyout_error(nvme->n_dead_status, arg,
7389 		    mode));
7390 	}
7391 
7392 	/*
7393 	 * ioctls that are no longer using the original ioctl structure.
7394 	 */
7395 	switch (cmd) {
7396 	case NVME_IOC_CTRL_INFO:
7397 		return (nvme_ioctl_ctrl_info(minor, arg, mode, cred_p));
7398 	case NVME_IOC_IDENTIFY:
7399 		return (nvme_ioctl_identify(minor, arg, mode, cred_p));
7400 	case NVME_IOC_GET_LOGPAGE:
7401 		return (nvme_ioctl_get_logpage(minor, arg, mode, cred_p));
7402 	case NVME_IOC_GET_FEATURE:
7403 		return (nvme_ioctl_get_feature(minor, arg, mode, cred_p));
7404 	case NVME_IOC_DETACH:
7405 		return (nvme_ioctl_detach(minor, arg, mode, cred_p));
7406 	case NVME_IOC_ATTACH:
7407 		return (nvme_ioctl_attach(minor, arg, mode, cred_p));
7408 	case NVME_IOC_FORMAT:
7409 		return (nvme_ioctl_format(minor, arg, mode, cred_p));
7410 	case NVME_IOC_FIRMWARE_DOWNLOAD:
7411 		return (nvme_ioctl_firmware_download(minor, arg, mode,
7412 		    cred_p));
7413 	case NVME_IOC_FIRMWARE_COMMIT:
7414 		return (nvme_ioctl_firmware_commit(minor, arg, mode,
7415 		    cred_p));
7416 	case NVME_IOC_NS_INFO:
7417 		return (nvme_ioctl_ns_info(minor, arg, mode, cred_p));
7418 	case NVME_IOC_PASSTHRU:
7419 		return (nvme_ioctl_passthru(minor, arg, mode, cred_p));
7420 	case NVME_IOC_LOCK:
7421 		return (nvme_ioctl_lock(minor, arg, mode, cred_p));
7422 	case NVME_IOC_UNLOCK:
7423 		return (nvme_ioctl_unlock(minor, arg, mode, cred_p));
7424 	default:
7425 		return (ENOTTY);
7426 	}
7427 }
7428 
7429 /*
7430  * DDI UFM Callbacks
7431  */
7432 static int
7433 nvme_ufm_fill_image(ddi_ufm_handle_t *ufmh, void *arg, uint_t imgno,
7434     ddi_ufm_image_t *img)
7435 {
7436 	nvme_t *nvme = arg;
7437 
7438 	if (imgno != 0)
7439 		return (EINVAL);
7440 
7441 	ddi_ufm_image_set_desc(img, "Firmware");
7442 	ddi_ufm_image_set_nslots(img, nvme->n_idctl->id_frmw.fw_nslot);
7443 
7444 	return (0);
7445 }
7446 
7447 /*
7448  * Fill out firmware slot information for the requested slot.  The firmware
7449  * slot information is gathered by requesting the Firmware Slot Information log
7450  * page.  The format of the page is described in section 5.10.1.3.
7451  *
7452  * We lazily cache the log page on the first call and then invalidate the cache
7453  * data after a successful firmware download or firmware commit command.
7454  * The cached data is protected by a mutex as the state can change
7455  * asynchronous to this callback.
7456  */
7457 static int
7458 nvme_ufm_fill_slot(ddi_ufm_handle_t *ufmh, void *arg, uint_t imgno,
7459     uint_t slotno, ddi_ufm_slot_t *slot)
7460 {
7461 	nvme_t *nvme = arg;
7462 	void *log = NULL;
7463 	size_t bufsize;
7464 	ddi_ufm_attr_t attr = 0;
7465 	char fw_ver[NVME_FWVER_SZ + 1];
7466 
7467 	if (imgno > 0 || slotno > (nvme->n_idctl->id_frmw.fw_nslot - 1))
7468 		return (EINVAL);
7469 
7470 	mutex_enter(&nvme->n_fwslot_mutex);
7471 	if (nvme->n_fwslot == NULL) {
7472 		if (!nvme_get_logpage_int(nvme, B_TRUE, &log, &bufsize,
7473 		    NVME_LOGPAGE_FWSLOT) ||
7474 		    bufsize != sizeof (nvme_fwslot_log_t)) {
7475 			if (log != NULL)
7476 				kmem_free(log, bufsize);
7477 			mutex_exit(&nvme->n_fwslot_mutex);
7478 			return (EIO);
7479 		}
7480 		nvme->n_fwslot = (nvme_fwslot_log_t *)log;
7481 	}
7482 
7483 	/*
7484 	 * NVMe numbers firmware slots starting at 1
7485 	 */
7486 	if (slotno == (nvme->n_fwslot->fw_afi - 1))
7487 		attr |= DDI_UFM_ATTR_ACTIVE;
7488 
7489 	if (slotno != 0 || nvme->n_idctl->id_frmw.fw_readonly == 0)
7490 		attr |= DDI_UFM_ATTR_WRITEABLE;
7491 
7492 	if (nvme->n_fwslot->fw_frs[slotno][0] == '\0') {
7493 		attr |= DDI_UFM_ATTR_EMPTY;
7494 	} else {
7495 		(void) strncpy(fw_ver, nvme->n_fwslot->fw_frs[slotno],
7496 		    NVME_FWVER_SZ);
7497 		fw_ver[NVME_FWVER_SZ] = '\0';
7498 		ddi_ufm_slot_set_version(slot, fw_ver);
7499 	}
7500 	mutex_exit(&nvme->n_fwslot_mutex);
7501 
7502 	ddi_ufm_slot_set_attrs(slot, attr);
7503 
7504 	return (0);
7505 }
7506 
7507 static int
7508 nvme_ufm_getcaps(ddi_ufm_handle_t *ufmh, void *arg, ddi_ufm_cap_t *caps)
7509 {
7510 	*caps = DDI_UFM_CAP_REPORT;
7511 	return (0);
7512 }
7513 
7514 boolean_t
7515 nvme_ctrl_atleast(nvme_t *nvme, const nvme_version_t *min)
7516 {
7517 	return (nvme_vers_atleast(&nvme->n_version, min) ? B_TRUE : B_FALSE);
7518 }
7519