xref: /linux/drivers/nvme/host/ioctl.c (revision fe833e4397fbdc3ae13a60202dfc7c335b032499)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (c) 2011-2014, Intel Corporation.
4  * Copyright (c) 2017-2021 Christoph Hellwig.
5  */
6 #include <linux/ptrace.h>	/* for force_successful_syscall_return */
7 #include <linux/nvme_ioctl.h>
8 #include <linux/io_uring/cmd.h>
9 #include "nvme.h"
10 
11 enum {
12 	NVME_IOCTL_VEC		= (1 << 0),
13 	NVME_IOCTL_PARTITION	= (1 << 1),
14 };
15 
16 static bool nvme_cmd_allowed(struct nvme_ns *ns, struct nvme_command *c,
17 		unsigned int flags, bool open_for_write)
18 {
19 	u32 effects;
20 
21 	/*
22 	 * Do not allow unprivileged passthrough on partitions, as that allows an
23 	 * escape from the containment of the partition.
24 	 */
25 	if (flags & NVME_IOCTL_PARTITION)
26 		goto admin;
27 
28 	/*
29 	 * Do not allow unprivileged processes to send vendor specific or fabrics
30 	 * commands as we can't be sure about their effects.
31 	 */
32 	if (c->common.opcode >= nvme_cmd_vendor_start ||
33 	    c->common.opcode == nvme_fabrics_command)
34 		goto admin;
35 
36 	/*
37 	 * Do not allow unprivileged passthrough of admin commands except
38 	 * for a subset of identify commands that contain information required
39 	 * to form proper I/O commands in userspace and do not expose any
40 	 * potentially sensitive information.
41 	 */
42 	if (!ns) {
43 		if (c->common.opcode == nvme_admin_identify) {
44 			switch (c->identify.cns) {
45 			case NVME_ID_CNS_NS:
46 			case NVME_ID_CNS_CS_NS:
47 			case NVME_ID_CNS_NS_CS_INDEP:
48 			case NVME_ID_CNS_CS_CTRL:
49 			case NVME_ID_CNS_CTRL:
50 				return true;
51 			}
52 		}
53 		goto admin;
54 	}
55 
56 	/*
57 	 * Check if the controller provides a Commands Supported and Effects log
58 	 * and marks this command as supported.  If not reject unprivileged
59 	 * passthrough.
60 	 */
61 	effects = nvme_command_effects(ns->ctrl, ns, c->common.opcode);
62 	if (!(effects & NVME_CMD_EFFECTS_CSUPP))
63 		goto admin;
64 
65 	/*
66 	 * Don't allow passthrough for command that have intrusive (or unknown)
67 	 * effects.
68 	 */
69 	if (effects & ~(NVME_CMD_EFFECTS_CSUPP | NVME_CMD_EFFECTS_LBCC |
70 			NVME_CMD_EFFECTS_UUID_SEL |
71 			NVME_CMD_EFFECTS_SCOPE_MASK))
72 		goto admin;
73 
74 	/*
75 	 * Only allow I/O commands that transfer data to the controller or that
76 	 * change the logical block contents if the file descriptor is open for
77 	 * writing.
78 	 */
79 	if ((nvme_is_write(c) || (effects & NVME_CMD_EFFECTS_LBCC)) &&
80 	    !open_for_write)
81 		goto admin;
82 
83 	return true;
84 admin:
85 	return capable(CAP_SYS_ADMIN);
86 }
87 
88 /*
89  * Convert integer values from ioctl structures to user pointers, silently
90  * ignoring the upper bits in the compat case to match behaviour of 32-bit
91  * kernels.
92  */
93 static void __user *nvme_to_user_ptr(uintptr_t ptrval)
94 {
95 	if (in_compat_syscall())
96 		ptrval = (compat_uptr_t)ptrval;
97 	return (void __user *)ptrval;
98 }
99 
100 static struct request *nvme_alloc_user_request(struct request_queue *q,
101 		struct nvme_command *cmd, blk_opf_t rq_flags,
102 		blk_mq_req_flags_t blk_flags)
103 {
104 	struct request *req;
105 
106 	req = blk_mq_alloc_request(q, nvme_req_op(cmd) | rq_flags, blk_flags);
107 	if (IS_ERR(req))
108 		return req;
109 	nvme_init_request(req, cmd);
110 	nvme_req(req)->flags |= NVME_REQ_USERCMD;
111 	return req;
112 }
113 
114 static int nvme_map_user_request(struct request *req, u64 ubuffer,
115 		unsigned bufflen, void __user *meta_buffer, unsigned meta_len,
116 		u32 meta_seed, struct io_uring_cmd *ioucmd, unsigned int flags)
117 {
118 	struct request_queue *q = req->q;
119 	struct nvme_ns *ns = q->queuedata;
120 	struct block_device *bdev = ns ? ns->disk->part0 : NULL;
121 	struct bio *bio = NULL;
122 	int ret;
123 
124 	if (ioucmd && (ioucmd->flags & IORING_URING_CMD_FIXED)) {
125 		struct iov_iter iter;
126 
127 		/* fixedbufs is only for non-vectored io */
128 		if (WARN_ON_ONCE(flags & NVME_IOCTL_VEC))
129 			return -EINVAL;
130 		ret = io_uring_cmd_import_fixed(ubuffer, bufflen,
131 				rq_data_dir(req), &iter, ioucmd);
132 		if (ret < 0)
133 			goto out;
134 		ret = blk_rq_map_user_iov(q, req, NULL, &iter, GFP_KERNEL);
135 	} else {
136 		ret = blk_rq_map_user_io(req, NULL, nvme_to_user_ptr(ubuffer),
137 				bufflen, GFP_KERNEL, flags & NVME_IOCTL_VEC, 0,
138 				0, rq_data_dir(req));
139 	}
140 
141 	if (ret)
142 		goto out;
143 
144 	bio = req->bio;
145 	if (bdev) {
146 		bio_set_dev(bio, bdev);
147 		if (meta_buffer && meta_len) {
148 			ret = bio_integrity_map_user(bio, meta_buffer, meta_len,
149 						     meta_seed);
150 			if (ret)
151 				goto out_unmap;
152 			req->cmd_flags |= REQ_INTEGRITY;
153 		}
154 	}
155 
156 	return ret;
157 
158 out_unmap:
159 	if (bio)
160 		blk_rq_unmap_user(bio);
161 out:
162 	blk_mq_free_request(req);
163 	return ret;
164 }
165 
166 static int nvme_submit_user_cmd(struct request_queue *q,
167 		struct nvme_command *cmd, u64 ubuffer, unsigned bufflen,
168 		void __user *meta_buffer, unsigned meta_len, u32 meta_seed,
169 		u64 *result, unsigned timeout, unsigned int flags)
170 {
171 	struct nvme_ns *ns = q->queuedata;
172 	struct nvme_ctrl *ctrl;
173 	struct request *req;
174 	struct bio *bio;
175 	u32 effects;
176 	int ret;
177 
178 	req = nvme_alloc_user_request(q, cmd, 0, 0);
179 	if (IS_ERR(req))
180 		return PTR_ERR(req);
181 
182 	req->timeout = timeout;
183 	if (ubuffer && bufflen) {
184 		ret = nvme_map_user_request(req, ubuffer, bufflen, meta_buffer,
185 				meta_len, meta_seed, NULL, flags);
186 		if (ret)
187 			return ret;
188 	}
189 
190 	bio = req->bio;
191 	ctrl = nvme_req(req)->ctrl;
192 
193 	effects = nvme_passthru_start(ctrl, ns, cmd->common.opcode);
194 	ret = nvme_execute_rq(req, false);
195 	if (result)
196 		*result = le64_to_cpu(nvme_req(req)->result.u64);
197 	if (bio)
198 		blk_rq_unmap_user(bio);
199 	blk_mq_free_request(req);
200 
201 	if (effects)
202 		nvme_passthru_end(ctrl, ns, effects, cmd, ret);
203 
204 	return ret;
205 }
206 
207 static int nvme_submit_io(struct nvme_ns *ns, struct nvme_user_io __user *uio)
208 {
209 	struct nvme_user_io io;
210 	struct nvme_command c;
211 	unsigned length, meta_len;
212 	void __user *metadata;
213 
214 	if (copy_from_user(&io, uio, sizeof(io)))
215 		return -EFAULT;
216 	if (io.flags)
217 		return -EINVAL;
218 
219 	switch (io.opcode) {
220 	case nvme_cmd_write:
221 	case nvme_cmd_read:
222 	case nvme_cmd_compare:
223 		break;
224 	default:
225 		return -EINVAL;
226 	}
227 
228 	length = (io.nblocks + 1) << ns->head->lba_shift;
229 
230 	if ((io.control & NVME_RW_PRINFO_PRACT) &&
231 	    (ns->head->ms == ns->head->pi_size)) {
232 		/*
233 		 * Protection information is stripped/inserted by the
234 		 * controller.
235 		 */
236 		if (nvme_to_user_ptr(io.metadata))
237 			return -EINVAL;
238 		meta_len = 0;
239 		metadata = NULL;
240 	} else {
241 		meta_len = (io.nblocks + 1) * ns->head->ms;
242 		metadata = nvme_to_user_ptr(io.metadata);
243 	}
244 
245 	if (ns->head->features & NVME_NS_EXT_LBAS) {
246 		length += meta_len;
247 		meta_len = 0;
248 	} else if (meta_len) {
249 		if ((io.metadata & 3) || !io.metadata)
250 			return -EINVAL;
251 	}
252 
253 	memset(&c, 0, sizeof(c));
254 	c.rw.opcode = io.opcode;
255 	c.rw.flags = io.flags;
256 	c.rw.nsid = cpu_to_le32(ns->head->ns_id);
257 	c.rw.slba = cpu_to_le64(io.slba);
258 	c.rw.length = cpu_to_le16(io.nblocks);
259 	c.rw.control = cpu_to_le16(io.control);
260 	c.rw.dsmgmt = cpu_to_le32(io.dsmgmt);
261 	c.rw.reftag = cpu_to_le32(io.reftag);
262 	c.rw.apptag = cpu_to_le16(io.apptag);
263 	c.rw.appmask = cpu_to_le16(io.appmask);
264 
265 	return nvme_submit_user_cmd(ns->queue, &c, io.addr, length, metadata,
266 			meta_len, lower_32_bits(io.slba), NULL, 0, 0);
267 }
268 
269 static bool nvme_validate_passthru_nsid(struct nvme_ctrl *ctrl,
270 					struct nvme_ns *ns, __u32 nsid)
271 {
272 	if (ns && nsid != ns->head->ns_id) {
273 		dev_err(ctrl->device,
274 			"%s: nsid (%u) in cmd does not match nsid (%u)"
275 			"of namespace\n",
276 			current->comm, nsid, ns->head->ns_id);
277 		return false;
278 	}
279 
280 	return true;
281 }
282 
283 static int nvme_user_cmd(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
284 		struct nvme_passthru_cmd __user *ucmd, unsigned int flags,
285 		bool open_for_write)
286 {
287 	struct nvme_passthru_cmd cmd;
288 	struct nvme_command c;
289 	unsigned timeout = 0;
290 	u64 result;
291 	int status;
292 
293 	if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
294 		return -EFAULT;
295 	if (cmd.flags)
296 		return -EINVAL;
297 	if (!nvme_validate_passthru_nsid(ctrl, ns, cmd.nsid))
298 		return -EINVAL;
299 
300 	memset(&c, 0, sizeof(c));
301 	c.common.opcode = cmd.opcode;
302 	c.common.flags = cmd.flags;
303 	c.common.nsid = cpu_to_le32(cmd.nsid);
304 	c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
305 	c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
306 	c.common.cdw10 = cpu_to_le32(cmd.cdw10);
307 	c.common.cdw11 = cpu_to_le32(cmd.cdw11);
308 	c.common.cdw12 = cpu_to_le32(cmd.cdw12);
309 	c.common.cdw13 = cpu_to_le32(cmd.cdw13);
310 	c.common.cdw14 = cpu_to_le32(cmd.cdw14);
311 	c.common.cdw15 = cpu_to_le32(cmd.cdw15);
312 
313 	if (!nvme_cmd_allowed(ns, &c, 0, open_for_write))
314 		return -EACCES;
315 
316 	if (cmd.timeout_ms)
317 		timeout = msecs_to_jiffies(cmd.timeout_ms);
318 
319 	status = nvme_submit_user_cmd(ns ? ns->queue : ctrl->admin_q, &c,
320 			cmd.addr, cmd.data_len, nvme_to_user_ptr(cmd.metadata),
321 			cmd.metadata_len, 0, &result, timeout, 0);
322 
323 	if (status >= 0) {
324 		if (put_user(result, &ucmd->result))
325 			return -EFAULT;
326 	}
327 
328 	return status;
329 }
330 
331 static int nvme_user_cmd64(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
332 		struct nvme_passthru_cmd64 __user *ucmd, unsigned int flags,
333 		bool open_for_write)
334 {
335 	struct nvme_passthru_cmd64 cmd;
336 	struct nvme_command c;
337 	unsigned timeout = 0;
338 	int status;
339 
340 	if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
341 		return -EFAULT;
342 	if (cmd.flags)
343 		return -EINVAL;
344 	if (!nvme_validate_passthru_nsid(ctrl, ns, cmd.nsid))
345 		return -EINVAL;
346 
347 	memset(&c, 0, sizeof(c));
348 	c.common.opcode = cmd.opcode;
349 	c.common.flags = cmd.flags;
350 	c.common.nsid = cpu_to_le32(cmd.nsid);
351 	c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
352 	c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
353 	c.common.cdw10 = cpu_to_le32(cmd.cdw10);
354 	c.common.cdw11 = cpu_to_le32(cmd.cdw11);
355 	c.common.cdw12 = cpu_to_le32(cmd.cdw12);
356 	c.common.cdw13 = cpu_to_le32(cmd.cdw13);
357 	c.common.cdw14 = cpu_to_le32(cmd.cdw14);
358 	c.common.cdw15 = cpu_to_le32(cmd.cdw15);
359 
360 	if (!nvme_cmd_allowed(ns, &c, flags, open_for_write))
361 		return -EACCES;
362 
363 	if (cmd.timeout_ms)
364 		timeout = msecs_to_jiffies(cmd.timeout_ms);
365 
366 	status = nvme_submit_user_cmd(ns ? ns->queue : ctrl->admin_q, &c,
367 			cmd.addr, cmd.data_len, nvme_to_user_ptr(cmd.metadata),
368 			cmd.metadata_len, 0, &cmd.result, timeout, flags);
369 
370 	if (status >= 0) {
371 		if (put_user(cmd.result, &ucmd->result))
372 			return -EFAULT;
373 	}
374 
375 	return status;
376 }
377 
378 struct nvme_uring_data {
379 	__u64	metadata;
380 	__u64	addr;
381 	__u32	data_len;
382 	__u32	metadata_len;
383 	__u32	timeout_ms;
384 };
385 
386 /*
387  * This overlays struct io_uring_cmd pdu.
388  * Expect build errors if this grows larger than that.
389  */
390 struct nvme_uring_cmd_pdu {
391 	struct request *req;
392 	struct bio *bio;
393 	u64 result;
394 	int status;
395 };
396 
397 static inline struct nvme_uring_cmd_pdu *nvme_uring_cmd_pdu(
398 		struct io_uring_cmd *ioucmd)
399 {
400 	return (struct nvme_uring_cmd_pdu *)&ioucmd->pdu;
401 }
402 
403 static void nvme_uring_task_cb(struct io_uring_cmd *ioucmd,
404 			       unsigned issue_flags)
405 {
406 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
407 
408 	if (pdu->bio)
409 		blk_rq_unmap_user(pdu->bio);
410 	io_uring_cmd_done(ioucmd, pdu->status, pdu->result, issue_flags);
411 }
412 
413 static enum rq_end_io_ret nvme_uring_cmd_end_io(struct request *req,
414 						blk_status_t err)
415 {
416 	struct io_uring_cmd *ioucmd = req->end_io_data;
417 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
418 
419 	if (nvme_req(req)->flags & NVME_REQ_CANCELLED)
420 		pdu->status = -EINTR;
421 	else
422 		pdu->status = nvme_req(req)->status;
423 	pdu->result = le64_to_cpu(nvme_req(req)->result.u64);
424 
425 	/*
426 	 * For iopoll, complete it directly. Note that using the uring_cmd
427 	 * helper for this is safe only because we check blk_rq_is_poll().
428 	 * As that returns false if we're NOT on a polled queue, then it's
429 	 * safe to use the polled completion helper.
430 	 *
431 	 * Otherwise, move the completion to task work.
432 	 */
433 	if (blk_rq_is_poll(req)) {
434 		if (pdu->bio)
435 			blk_rq_unmap_user(pdu->bio);
436 		io_uring_cmd_iopoll_done(ioucmd, pdu->result, pdu->status);
437 	} else {
438 		io_uring_cmd_do_in_task_lazy(ioucmd, nvme_uring_task_cb);
439 	}
440 
441 	return RQ_END_IO_FREE;
442 }
443 
444 static int nvme_uring_cmd_io(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
445 		struct io_uring_cmd *ioucmd, unsigned int issue_flags, bool vec)
446 {
447 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
448 	const struct nvme_uring_cmd *cmd = io_uring_sqe_cmd(ioucmd->sqe);
449 	struct request_queue *q = ns ? ns->queue : ctrl->admin_q;
450 	struct nvme_uring_data d;
451 	struct nvme_command c;
452 	struct request *req;
453 	blk_opf_t rq_flags = REQ_ALLOC_CACHE;
454 	blk_mq_req_flags_t blk_flags = 0;
455 	int ret;
456 
457 	c.common.opcode = READ_ONCE(cmd->opcode);
458 	c.common.flags = READ_ONCE(cmd->flags);
459 	if (c.common.flags)
460 		return -EINVAL;
461 
462 	c.common.command_id = 0;
463 	c.common.nsid = cpu_to_le32(cmd->nsid);
464 	if (!nvme_validate_passthru_nsid(ctrl, ns, le32_to_cpu(c.common.nsid)))
465 		return -EINVAL;
466 
467 	c.common.cdw2[0] = cpu_to_le32(READ_ONCE(cmd->cdw2));
468 	c.common.cdw2[1] = cpu_to_le32(READ_ONCE(cmd->cdw3));
469 	c.common.metadata = 0;
470 	c.common.dptr.prp1 = c.common.dptr.prp2 = 0;
471 	c.common.cdw10 = cpu_to_le32(READ_ONCE(cmd->cdw10));
472 	c.common.cdw11 = cpu_to_le32(READ_ONCE(cmd->cdw11));
473 	c.common.cdw12 = cpu_to_le32(READ_ONCE(cmd->cdw12));
474 	c.common.cdw13 = cpu_to_le32(READ_ONCE(cmd->cdw13));
475 	c.common.cdw14 = cpu_to_le32(READ_ONCE(cmd->cdw14));
476 	c.common.cdw15 = cpu_to_le32(READ_ONCE(cmd->cdw15));
477 
478 	if (!nvme_cmd_allowed(ns, &c, 0, ioucmd->file->f_mode & FMODE_WRITE))
479 		return -EACCES;
480 
481 	d.metadata = READ_ONCE(cmd->metadata);
482 	d.addr = READ_ONCE(cmd->addr);
483 	d.data_len = READ_ONCE(cmd->data_len);
484 	d.metadata_len = READ_ONCE(cmd->metadata_len);
485 	d.timeout_ms = READ_ONCE(cmd->timeout_ms);
486 
487 	if (issue_flags & IO_URING_F_NONBLOCK) {
488 		rq_flags |= REQ_NOWAIT;
489 		blk_flags = BLK_MQ_REQ_NOWAIT;
490 	}
491 	if (issue_flags & IO_URING_F_IOPOLL)
492 		rq_flags |= REQ_POLLED;
493 
494 	req = nvme_alloc_user_request(q, &c, rq_flags, blk_flags);
495 	if (IS_ERR(req))
496 		return PTR_ERR(req);
497 	req->timeout = d.timeout_ms ? msecs_to_jiffies(d.timeout_ms) : 0;
498 
499 	if (d.addr && d.data_len) {
500 		ret = nvme_map_user_request(req, d.addr,
501 			d.data_len, nvme_to_user_ptr(d.metadata),
502 			d.metadata_len, 0, ioucmd, vec);
503 		if (ret)
504 			return ret;
505 	}
506 
507 	/* to free bio on completion, as req->bio will be null at that time */
508 	pdu->bio = req->bio;
509 	pdu->req = req;
510 	req->end_io_data = ioucmd;
511 	req->end_io = nvme_uring_cmd_end_io;
512 	blk_execute_rq_nowait(req, false);
513 	return -EIOCBQUEUED;
514 }
515 
516 static bool is_ctrl_ioctl(unsigned int cmd)
517 {
518 	if (cmd == NVME_IOCTL_ADMIN_CMD || cmd == NVME_IOCTL_ADMIN64_CMD)
519 		return true;
520 	if (is_sed_ioctl(cmd))
521 		return true;
522 	return false;
523 }
524 
525 static int nvme_ctrl_ioctl(struct nvme_ctrl *ctrl, unsigned int cmd,
526 		void __user *argp, bool open_for_write)
527 {
528 	switch (cmd) {
529 	case NVME_IOCTL_ADMIN_CMD:
530 		return nvme_user_cmd(ctrl, NULL, argp, 0, open_for_write);
531 	case NVME_IOCTL_ADMIN64_CMD:
532 		return nvme_user_cmd64(ctrl, NULL, argp, 0, open_for_write);
533 	default:
534 		return sed_ioctl(ctrl->opal_dev, cmd, argp);
535 	}
536 }
537 
538 #ifdef COMPAT_FOR_U64_ALIGNMENT
539 struct nvme_user_io32 {
540 	__u8	opcode;
541 	__u8	flags;
542 	__u16	control;
543 	__u16	nblocks;
544 	__u16	rsvd;
545 	__u64	metadata;
546 	__u64	addr;
547 	__u64	slba;
548 	__u32	dsmgmt;
549 	__u32	reftag;
550 	__u16	apptag;
551 	__u16	appmask;
552 } __attribute__((__packed__));
553 #define NVME_IOCTL_SUBMIT_IO32	_IOW('N', 0x42, struct nvme_user_io32)
554 #endif /* COMPAT_FOR_U64_ALIGNMENT */
555 
556 static int nvme_ns_ioctl(struct nvme_ns *ns, unsigned int cmd,
557 		void __user *argp, unsigned int flags, bool open_for_write)
558 {
559 	switch (cmd) {
560 	case NVME_IOCTL_ID:
561 		force_successful_syscall_return();
562 		return ns->head->ns_id;
563 	case NVME_IOCTL_IO_CMD:
564 		return nvme_user_cmd(ns->ctrl, ns, argp, flags, open_for_write);
565 	/*
566 	 * struct nvme_user_io can have different padding on some 32-bit ABIs.
567 	 * Just accept the compat version as all fields that are used are the
568 	 * same size and at the same offset.
569 	 */
570 #ifdef COMPAT_FOR_U64_ALIGNMENT
571 	case NVME_IOCTL_SUBMIT_IO32:
572 #endif
573 	case NVME_IOCTL_SUBMIT_IO:
574 		return nvme_submit_io(ns, argp);
575 	case NVME_IOCTL_IO64_CMD_VEC:
576 		flags |= NVME_IOCTL_VEC;
577 		fallthrough;
578 	case NVME_IOCTL_IO64_CMD:
579 		return nvme_user_cmd64(ns->ctrl, ns, argp, flags,
580 				       open_for_write);
581 	default:
582 		return -ENOTTY;
583 	}
584 }
585 
586 int nvme_ioctl(struct block_device *bdev, blk_mode_t mode,
587 		unsigned int cmd, unsigned long arg)
588 {
589 	struct nvme_ns *ns = bdev->bd_disk->private_data;
590 	bool open_for_write = mode & BLK_OPEN_WRITE;
591 	void __user *argp = (void __user *)arg;
592 	unsigned int flags = 0;
593 
594 	if (bdev_is_partition(bdev))
595 		flags |= NVME_IOCTL_PARTITION;
596 
597 	if (is_ctrl_ioctl(cmd))
598 		return nvme_ctrl_ioctl(ns->ctrl, cmd, argp, open_for_write);
599 	return nvme_ns_ioctl(ns, cmd, argp, flags, open_for_write);
600 }
601 
602 long nvme_ns_chr_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
603 {
604 	struct nvme_ns *ns =
605 		container_of(file_inode(file)->i_cdev, struct nvme_ns, cdev);
606 	bool open_for_write = file->f_mode & FMODE_WRITE;
607 	void __user *argp = (void __user *)arg;
608 
609 	if (is_ctrl_ioctl(cmd))
610 		return nvme_ctrl_ioctl(ns->ctrl, cmd, argp, open_for_write);
611 	return nvme_ns_ioctl(ns, cmd, argp, 0, open_for_write);
612 }
613 
614 static int nvme_uring_cmd_checks(unsigned int issue_flags)
615 {
616 
617 	/* NVMe passthrough requires big SQE/CQE support */
618 	if ((issue_flags & (IO_URING_F_SQE128|IO_URING_F_CQE32)) !=
619 	    (IO_URING_F_SQE128|IO_URING_F_CQE32))
620 		return -EOPNOTSUPP;
621 	return 0;
622 }
623 
624 static int nvme_ns_uring_cmd(struct nvme_ns *ns, struct io_uring_cmd *ioucmd,
625 			     unsigned int issue_flags)
626 {
627 	struct nvme_ctrl *ctrl = ns->ctrl;
628 	int ret;
629 
630 	BUILD_BUG_ON(sizeof(struct nvme_uring_cmd_pdu) > sizeof(ioucmd->pdu));
631 
632 	ret = nvme_uring_cmd_checks(issue_flags);
633 	if (ret)
634 		return ret;
635 
636 	switch (ioucmd->cmd_op) {
637 	case NVME_URING_CMD_IO:
638 		ret = nvme_uring_cmd_io(ctrl, ns, ioucmd, issue_flags, false);
639 		break;
640 	case NVME_URING_CMD_IO_VEC:
641 		ret = nvme_uring_cmd_io(ctrl, ns, ioucmd, issue_flags, true);
642 		break;
643 	default:
644 		ret = -ENOTTY;
645 	}
646 
647 	return ret;
648 }
649 
650 int nvme_ns_chr_uring_cmd(struct io_uring_cmd *ioucmd, unsigned int issue_flags)
651 {
652 	struct nvme_ns *ns = container_of(file_inode(ioucmd->file)->i_cdev,
653 			struct nvme_ns, cdev);
654 
655 	return nvme_ns_uring_cmd(ns, ioucmd, issue_flags);
656 }
657 
658 int nvme_ns_chr_uring_cmd_iopoll(struct io_uring_cmd *ioucmd,
659 				 struct io_comp_batch *iob,
660 				 unsigned int poll_flags)
661 {
662 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
663 	struct request *req = pdu->req;
664 
665 	if (req && blk_rq_is_poll(req))
666 		return blk_rq_poll(req, iob, poll_flags);
667 	return 0;
668 }
669 #ifdef CONFIG_NVME_MULTIPATH
670 static int nvme_ns_head_ctrl_ioctl(struct nvme_ns *ns, unsigned int cmd,
671 		void __user *argp, struct nvme_ns_head *head, int srcu_idx,
672 		bool open_for_write)
673 	__releases(&head->srcu)
674 {
675 	struct nvme_ctrl *ctrl = ns->ctrl;
676 	int ret;
677 
678 	nvme_get_ctrl(ns->ctrl);
679 	srcu_read_unlock(&head->srcu, srcu_idx);
680 	ret = nvme_ctrl_ioctl(ns->ctrl, cmd, argp, open_for_write);
681 
682 	nvme_put_ctrl(ctrl);
683 	return ret;
684 }
685 
686 int nvme_ns_head_ioctl(struct block_device *bdev, blk_mode_t mode,
687 		unsigned int cmd, unsigned long arg)
688 {
689 	struct nvme_ns_head *head = bdev->bd_disk->private_data;
690 	bool open_for_write = mode & BLK_OPEN_WRITE;
691 	void __user *argp = (void __user *)arg;
692 	struct nvme_ns *ns;
693 	int srcu_idx, ret = -EWOULDBLOCK;
694 	unsigned int flags = 0;
695 
696 	if (bdev_is_partition(bdev))
697 		flags |= NVME_IOCTL_PARTITION;
698 
699 	srcu_idx = srcu_read_lock(&head->srcu);
700 	ns = nvme_find_path(head);
701 	if (!ns)
702 		goto out_unlock;
703 
704 	/*
705 	 * Handle ioctls that apply to the controller instead of the namespace
706 	 * seperately and drop the ns SRCU reference early.  This avoids a
707 	 * deadlock when deleting namespaces using the passthrough interface.
708 	 */
709 	if (is_ctrl_ioctl(cmd))
710 		return nvme_ns_head_ctrl_ioctl(ns, cmd, argp, head, srcu_idx,
711 					       open_for_write);
712 
713 	ret = nvme_ns_ioctl(ns, cmd, argp, flags, open_for_write);
714 out_unlock:
715 	srcu_read_unlock(&head->srcu, srcu_idx);
716 	return ret;
717 }
718 
719 long nvme_ns_head_chr_ioctl(struct file *file, unsigned int cmd,
720 		unsigned long arg)
721 {
722 	bool open_for_write = file->f_mode & FMODE_WRITE;
723 	struct cdev *cdev = file_inode(file)->i_cdev;
724 	struct nvme_ns_head *head =
725 		container_of(cdev, struct nvme_ns_head, cdev);
726 	void __user *argp = (void __user *)arg;
727 	struct nvme_ns *ns;
728 	int srcu_idx, ret = -EWOULDBLOCK;
729 
730 	srcu_idx = srcu_read_lock(&head->srcu);
731 	ns = nvme_find_path(head);
732 	if (!ns)
733 		goto out_unlock;
734 
735 	if (is_ctrl_ioctl(cmd))
736 		return nvme_ns_head_ctrl_ioctl(ns, cmd, argp, head, srcu_idx,
737 				open_for_write);
738 
739 	ret = nvme_ns_ioctl(ns, cmd, argp, 0, open_for_write);
740 out_unlock:
741 	srcu_read_unlock(&head->srcu, srcu_idx);
742 	return ret;
743 }
744 
745 int nvme_ns_head_chr_uring_cmd(struct io_uring_cmd *ioucmd,
746 		unsigned int issue_flags)
747 {
748 	struct cdev *cdev = file_inode(ioucmd->file)->i_cdev;
749 	struct nvme_ns_head *head = container_of(cdev, struct nvme_ns_head, cdev);
750 	int srcu_idx = srcu_read_lock(&head->srcu);
751 	struct nvme_ns *ns = nvme_find_path(head);
752 	int ret = -EINVAL;
753 
754 	if (ns)
755 		ret = nvme_ns_uring_cmd(ns, ioucmd, issue_flags);
756 	srcu_read_unlock(&head->srcu, srcu_idx);
757 	return ret;
758 }
759 #endif /* CONFIG_NVME_MULTIPATH */
760 
761 int nvme_dev_uring_cmd(struct io_uring_cmd *ioucmd, unsigned int issue_flags)
762 {
763 	struct nvme_ctrl *ctrl = ioucmd->file->private_data;
764 	int ret;
765 
766 	/* IOPOLL not supported yet */
767 	if (issue_flags & IO_URING_F_IOPOLL)
768 		return -EOPNOTSUPP;
769 
770 	ret = nvme_uring_cmd_checks(issue_flags);
771 	if (ret)
772 		return ret;
773 
774 	switch (ioucmd->cmd_op) {
775 	case NVME_URING_CMD_ADMIN:
776 		ret = nvme_uring_cmd_io(ctrl, NULL, ioucmd, issue_flags, false);
777 		break;
778 	case NVME_URING_CMD_ADMIN_VEC:
779 		ret = nvme_uring_cmd_io(ctrl, NULL, ioucmd, issue_flags, true);
780 		break;
781 	default:
782 		ret = -ENOTTY;
783 	}
784 
785 	return ret;
786 }
787 
788 static int nvme_dev_user_cmd(struct nvme_ctrl *ctrl, void __user *argp,
789 		bool open_for_write)
790 {
791 	struct nvme_ns *ns;
792 	int ret, srcu_idx;
793 
794 	srcu_idx = srcu_read_lock(&ctrl->srcu);
795 	if (list_empty(&ctrl->namespaces)) {
796 		ret = -ENOTTY;
797 		goto out_unlock;
798 	}
799 
800 	ns = list_first_or_null_rcu(&ctrl->namespaces, struct nvme_ns, list);
801 	if (ns != list_last_entry(&ctrl->namespaces, struct nvme_ns, list)) {
802 		dev_warn(ctrl->device,
803 			"NVME_IOCTL_IO_CMD not supported when multiple namespaces present!\n");
804 		ret = -EINVAL;
805 		goto out_unlock;
806 	}
807 
808 	dev_warn(ctrl->device,
809 		"using deprecated NVME_IOCTL_IO_CMD ioctl on the char device!\n");
810 	if (!nvme_get_ns(ns)) {
811 		ret = -ENXIO;
812 		goto out_unlock;
813 	}
814 	srcu_read_unlock(&ctrl->srcu, srcu_idx);
815 
816 	ret = nvme_user_cmd(ctrl, ns, argp, 0, open_for_write);
817 	nvme_put_ns(ns);
818 	return ret;
819 
820 out_unlock:
821 	srcu_read_unlock(&ctrl->srcu, srcu_idx);
822 	return ret;
823 }
824 
825 long nvme_dev_ioctl(struct file *file, unsigned int cmd,
826 		unsigned long arg)
827 {
828 	bool open_for_write = file->f_mode & FMODE_WRITE;
829 	struct nvme_ctrl *ctrl = file->private_data;
830 	void __user *argp = (void __user *)arg;
831 
832 	switch (cmd) {
833 	case NVME_IOCTL_ADMIN_CMD:
834 		return nvme_user_cmd(ctrl, NULL, argp, 0, open_for_write);
835 	case NVME_IOCTL_ADMIN64_CMD:
836 		return nvme_user_cmd64(ctrl, NULL, argp, 0, open_for_write);
837 	case NVME_IOCTL_IO_CMD:
838 		return nvme_dev_user_cmd(ctrl, argp, open_for_write);
839 	case NVME_IOCTL_RESET:
840 		if (!capable(CAP_SYS_ADMIN))
841 			return -EACCES;
842 		dev_warn(ctrl->device, "resetting controller\n");
843 		return nvme_reset_ctrl_sync(ctrl);
844 	case NVME_IOCTL_SUBSYS_RESET:
845 		if (!capable(CAP_SYS_ADMIN))
846 			return -EACCES;
847 		return nvme_reset_subsystem(ctrl);
848 	case NVME_IOCTL_RESCAN:
849 		if (!capable(CAP_SYS_ADMIN))
850 			return -EACCES;
851 		nvme_queue_scan(ctrl);
852 		return 0;
853 	default:
854 		return -ENOTTY;
855 	}
856 }
857