xref: /linux/drivers/nvme/host/ioctl.c (revision a3f143c461444c0b56360bbf468615fa814a8372)
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/blk-integrity.h>
7 #include <linux/ptrace.h>	/* for force_successful_syscall_return */
8 #include <linux/nvme_ioctl.h>
9 #include <linux/io_uring/cmd.h>
10 #include "nvme.h"
11 
12 enum {
13 	NVME_IOCTL_VEC		= (1 << 0),
14 	NVME_IOCTL_PARTITION	= (1 << 1),
15 };
16 
17 static bool nvme_cmd_allowed(struct nvme_ns *ns, struct nvme_command *c,
18 		unsigned int flags, bool open_for_write)
19 {
20 	u32 effects;
21 
22 	/*
23 	 * Do not allow unprivileged passthrough on partitions, as that allows an
24 	 * escape from the containment of the partition.
25 	 */
26 	if (flags & NVME_IOCTL_PARTITION)
27 		goto admin;
28 
29 	/*
30 	 * Do not allow unprivileged processes to send vendor specific or fabrics
31 	 * commands as we can't be sure about their effects.
32 	 */
33 	if (c->common.opcode >= nvme_cmd_vendor_start ||
34 	    c->common.opcode == nvme_fabrics_command)
35 		goto admin;
36 
37 	/*
38 	 * Do not allow unprivileged passthrough of admin commands except
39 	 * for a subset of identify commands that contain information required
40 	 * to form proper I/O commands in userspace and do not expose any
41 	 * potentially sensitive information.
42 	 */
43 	if (!ns) {
44 		if (c->common.opcode == nvme_admin_identify) {
45 			switch (c->identify.cns) {
46 			case NVME_ID_CNS_NS:
47 			case NVME_ID_CNS_CS_NS:
48 			case NVME_ID_CNS_NS_CS_INDEP:
49 			case NVME_ID_CNS_CS_CTRL:
50 			case NVME_ID_CNS_CTRL:
51 				return true;
52 			}
53 		}
54 		goto admin;
55 	}
56 
57 	/*
58 	 * Check if the controller provides a Commands Supported and Effects log
59 	 * and marks this command as supported.  If not reject unprivileged
60 	 * passthrough.
61 	 */
62 	effects = nvme_command_effects(ns->ctrl, ns, c->common.opcode);
63 	if (!(effects & NVME_CMD_EFFECTS_CSUPP))
64 		goto admin;
65 
66 	/*
67 	 * Don't allow passthrough for command that have intrusive (or unknown)
68 	 * effects.
69 	 */
70 	if (effects & ~(NVME_CMD_EFFECTS_CSUPP | NVME_CMD_EFFECTS_LBCC |
71 			NVME_CMD_EFFECTS_UUID_SEL |
72 			NVME_CMD_EFFECTS_SCOPE_MASK))
73 		goto admin;
74 
75 	/*
76 	 * Only allow I/O commands that transfer data to the controller or that
77 	 * change the logical block contents if the file descriptor is open for
78 	 * writing.
79 	 */
80 	if ((nvme_is_write(c) || (effects & NVME_CMD_EFFECTS_LBCC)) &&
81 	    !open_for_write)
82 		goto admin;
83 
84 	return true;
85 admin:
86 	return capable(CAP_SYS_ADMIN);
87 }
88 
89 /*
90  * Convert integer values from ioctl structures to user pointers, silently
91  * ignoring the upper bits in the compat case to match behaviour of 32-bit
92  * kernels.
93  */
94 static void __user *nvme_to_user_ptr(uintptr_t ptrval)
95 {
96 	if (in_compat_syscall())
97 		ptrval = (compat_uptr_t)ptrval;
98 	return (void __user *)ptrval;
99 }
100 
101 static struct request *nvme_alloc_user_request(struct request_queue *q,
102 		struct nvme_command *cmd, blk_opf_t rq_flags,
103 		blk_mq_req_flags_t blk_flags)
104 {
105 	struct request *req;
106 
107 	req = blk_mq_alloc_request(q, nvme_req_op(cmd) | rq_flags, blk_flags);
108 	if (IS_ERR(req))
109 		return req;
110 	nvme_init_request(req, cmd);
111 	nvme_req(req)->flags |= NVME_REQ_USERCMD;
112 	return req;
113 }
114 
115 static int nvme_map_user_request(struct request *req, u64 ubuffer,
116 		unsigned bufflen, void __user *meta_buffer, unsigned meta_len,
117 		struct io_uring_cmd *ioucmd, unsigned int flags)
118 {
119 	struct request_queue *q = req->q;
120 	struct nvme_ns *ns = q->queuedata;
121 	struct block_device *bdev = ns ? ns->disk->part0 : NULL;
122 	bool supports_metadata = bdev && blk_get_integrity(bdev->bd_disk);
123 	bool has_metadata = meta_buffer && meta_len;
124 	struct bio *bio = NULL;
125 	int ret;
126 
127 	if (has_metadata && !supports_metadata)
128 		return -EINVAL;
129 
130 	if (ioucmd && (ioucmd->flags & IORING_URING_CMD_FIXED)) {
131 		struct iov_iter iter;
132 
133 		/* fixedbufs is only for non-vectored io */
134 		if (WARN_ON_ONCE(flags & NVME_IOCTL_VEC))
135 			return -EINVAL;
136 		ret = io_uring_cmd_import_fixed(ubuffer, bufflen,
137 				rq_data_dir(req), &iter, ioucmd);
138 		if (ret < 0)
139 			goto out;
140 		ret = blk_rq_map_user_iov(q, req, NULL, &iter, GFP_KERNEL);
141 	} else {
142 		ret = blk_rq_map_user_io(req, NULL, nvme_to_user_ptr(ubuffer),
143 				bufflen, GFP_KERNEL, flags & NVME_IOCTL_VEC, 0,
144 				0, rq_data_dir(req));
145 	}
146 
147 	if (ret)
148 		goto out;
149 
150 	bio = req->bio;
151 	if (bdev)
152 		bio_set_dev(bio, bdev);
153 
154 	if (has_metadata) {
155 		ret = blk_rq_integrity_map_user(req, meta_buffer, meta_len);
156 		if (ret)
157 			goto out_unmap;
158 	}
159 
160 	return ret;
161 
162 out_unmap:
163 	if (bio)
164 		blk_rq_unmap_user(bio);
165 out:
166 	blk_mq_free_request(req);
167 	return ret;
168 }
169 
170 static int nvme_submit_user_cmd(struct request_queue *q,
171 		struct nvme_command *cmd, u64 ubuffer, unsigned bufflen,
172 		void __user *meta_buffer, unsigned meta_len,
173 		u64 *result, unsigned timeout, unsigned int flags)
174 {
175 	struct nvme_ns *ns = q->queuedata;
176 	struct nvme_ctrl *ctrl;
177 	struct request *req;
178 	struct bio *bio;
179 	u32 effects;
180 	int ret;
181 
182 	req = nvme_alloc_user_request(q, cmd, 0, 0);
183 	if (IS_ERR(req))
184 		return PTR_ERR(req);
185 
186 	req->timeout = timeout;
187 	if (ubuffer && bufflen) {
188 		ret = nvme_map_user_request(req, ubuffer, bufflen, meta_buffer,
189 				meta_len, NULL, flags);
190 		if (ret)
191 			return ret;
192 	}
193 
194 	bio = req->bio;
195 	ctrl = nvme_req(req)->ctrl;
196 
197 	effects = nvme_passthru_start(ctrl, ns, cmd->common.opcode);
198 	ret = nvme_execute_rq(req, false);
199 	if (result)
200 		*result = le64_to_cpu(nvme_req(req)->result.u64);
201 	if (bio)
202 		blk_rq_unmap_user(bio);
203 	blk_mq_free_request(req);
204 
205 	if (effects)
206 		nvme_passthru_end(ctrl, ns, effects, cmd, ret);
207 
208 	return ret;
209 }
210 
211 static int nvme_submit_io(struct nvme_ns *ns, struct nvme_user_io __user *uio)
212 {
213 	struct nvme_user_io io;
214 	struct nvme_command c;
215 	unsigned length, meta_len;
216 	void __user *metadata;
217 
218 	if (copy_from_user(&io, uio, sizeof(io)))
219 		return -EFAULT;
220 	if (io.flags)
221 		return -EINVAL;
222 
223 	switch (io.opcode) {
224 	case nvme_cmd_write:
225 	case nvme_cmd_read:
226 	case nvme_cmd_compare:
227 		break;
228 	default:
229 		return -EINVAL;
230 	}
231 
232 	length = (io.nblocks + 1) << ns->head->lba_shift;
233 
234 	if ((io.control & NVME_RW_PRINFO_PRACT) &&
235 	    (ns->head->ms == ns->head->pi_size)) {
236 		/*
237 		 * Protection information is stripped/inserted by the
238 		 * controller.
239 		 */
240 		if (nvme_to_user_ptr(io.metadata))
241 			return -EINVAL;
242 		meta_len = 0;
243 		metadata = NULL;
244 	} else {
245 		meta_len = (io.nblocks + 1) * ns->head->ms;
246 		metadata = nvme_to_user_ptr(io.metadata);
247 	}
248 
249 	if (ns->head->features & NVME_NS_EXT_LBAS) {
250 		length += meta_len;
251 		meta_len = 0;
252 	} else if (meta_len) {
253 		if ((io.metadata & 3) || !io.metadata)
254 			return -EINVAL;
255 	}
256 
257 	memset(&c, 0, sizeof(c));
258 	c.rw.opcode = io.opcode;
259 	c.rw.flags = io.flags;
260 	c.rw.nsid = cpu_to_le32(ns->head->ns_id);
261 	c.rw.slba = cpu_to_le64(io.slba);
262 	c.rw.length = cpu_to_le16(io.nblocks);
263 	c.rw.control = cpu_to_le16(io.control);
264 	c.rw.dsmgmt = cpu_to_le32(io.dsmgmt);
265 	c.rw.reftag = cpu_to_le32(io.reftag);
266 	c.rw.lbat = cpu_to_le16(io.apptag);
267 	c.rw.lbatm = cpu_to_le16(io.appmask);
268 
269 	return nvme_submit_user_cmd(ns->queue, &c, io.addr, length, metadata,
270 			meta_len, NULL, 0, 0);
271 }
272 
273 static bool nvme_validate_passthru_nsid(struct nvme_ctrl *ctrl,
274 					struct nvme_ns *ns, __u32 nsid)
275 {
276 	if (ns && nsid != ns->head->ns_id) {
277 		dev_err(ctrl->device,
278 			"%s: nsid (%u) in cmd does not match nsid (%u)"
279 			"of namespace\n",
280 			current->comm, nsid, ns->head->ns_id);
281 		return false;
282 	}
283 
284 	return true;
285 }
286 
287 static int nvme_user_cmd(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
288 		struct nvme_passthru_cmd __user *ucmd, unsigned int flags,
289 		bool open_for_write)
290 {
291 	struct nvme_passthru_cmd cmd;
292 	struct nvme_command c;
293 	unsigned timeout = 0;
294 	u64 result;
295 	int status;
296 
297 	if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
298 		return -EFAULT;
299 	if (cmd.flags)
300 		return -EINVAL;
301 	if (!nvme_validate_passthru_nsid(ctrl, ns, cmd.nsid))
302 		return -EINVAL;
303 
304 	memset(&c, 0, sizeof(c));
305 	c.common.opcode = cmd.opcode;
306 	c.common.flags = cmd.flags;
307 	c.common.nsid = cpu_to_le32(cmd.nsid);
308 	c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
309 	c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
310 	c.common.cdw10 = cpu_to_le32(cmd.cdw10);
311 	c.common.cdw11 = cpu_to_le32(cmd.cdw11);
312 	c.common.cdw12 = cpu_to_le32(cmd.cdw12);
313 	c.common.cdw13 = cpu_to_le32(cmd.cdw13);
314 	c.common.cdw14 = cpu_to_le32(cmd.cdw14);
315 	c.common.cdw15 = cpu_to_le32(cmd.cdw15);
316 
317 	if (!nvme_cmd_allowed(ns, &c, 0, open_for_write))
318 		return -EACCES;
319 
320 	if (cmd.timeout_ms)
321 		timeout = msecs_to_jiffies(cmd.timeout_ms);
322 
323 	status = nvme_submit_user_cmd(ns ? ns->queue : ctrl->admin_q, &c,
324 			cmd.addr, cmd.data_len, nvme_to_user_ptr(cmd.metadata),
325 			cmd.metadata_len, &result, timeout, 0);
326 
327 	if (status >= 0) {
328 		if (put_user(result, &ucmd->result))
329 			return -EFAULT;
330 	}
331 
332 	return status;
333 }
334 
335 static int nvme_user_cmd64(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
336 		struct nvme_passthru_cmd64 __user *ucmd, unsigned int flags,
337 		bool open_for_write)
338 {
339 	struct nvme_passthru_cmd64 cmd;
340 	struct nvme_command c;
341 	unsigned timeout = 0;
342 	int status;
343 
344 	if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
345 		return -EFAULT;
346 	if (cmd.flags)
347 		return -EINVAL;
348 	if (!nvme_validate_passthru_nsid(ctrl, ns, cmd.nsid))
349 		return -EINVAL;
350 
351 	memset(&c, 0, sizeof(c));
352 	c.common.opcode = cmd.opcode;
353 	c.common.flags = cmd.flags;
354 	c.common.nsid = cpu_to_le32(cmd.nsid);
355 	c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
356 	c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
357 	c.common.cdw10 = cpu_to_le32(cmd.cdw10);
358 	c.common.cdw11 = cpu_to_le32(cmd.cdw11);
359 	c.common.cdw12 = cpu_to_le32(cmd.cdw12);
360 	c.common.cdw13 = cpu_to_le32(cmd.cdw13);
361 	c.common.cdw14 = cpu_to_le32(cmd.cdw14);
362 	c.common.cdw15 = cpu_to_le32(cmd.cdw15);
363 
364 	if (!nvme_cmd_allowed(ns, &c, flags, open_for_write))
365 		return -EACCES;
366 
367 	if (cmd.timeout_ms)
368 		timeout = msecs_to_jiffies(cmd.timeout_ms);
369 
370 	status = nvme_submit_user_cmd(ns ? ns->queue : ctrl->admin_q, &c,
371 			cmd.addr, cmd.data_len, nvme_to_user_ptr(cmd.metadata),
372 			cmd.metadata_len, &cmd.result, timeout, flags);
373 
374 	if (status >= 0) {
375 		if (put_user(cmd.result, &ucmd->result))
376 			return -EFAULT;
377 	}
378 
379 	return status;
380 }
381 
382 struct nvme_uring_data {
383 	__u64	metadata;
384 	__u64	addr;
385 	__u32	data_len;
386 	__u32	metadata_len;
387 	__u32	timeout_ms;
388 };
389 
390 /*
391  * This overlays struct io_uring_cmd pdu.
392  * Expect build errors if this grows larger than that.
393  */
394 struct nvme_uring_cmd_pdu {
395 	struct request *req;
396 	struct bio *bio;
397 	u64 result;
398 	int status;
399 };
400 
401 static inline struct nvme_uring_cmd_pdu *nvme_uring_cmd_pdu(
402 		struct io_uring_cmd *ioucmd)
403 {
404 	return io_uring_cmd_to_pdu(ioucmd, struct nvme_uring_cmd_pdu);
405 }
406 
407 static void nvme_uring_task_cb(struct io_uring_cmd *ioucmd,
408 			       unsigned issue_flags)
409 {
410 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
411 
412 	if (pdu->bio)
413 		blk_rq_unmap_user(pdu->bio);
414 	io_uring_cmd_done(ioucmd, pdu->status, pdu->result, issue_flags);
415 }
416 
417 static enum rq_end_io_ret nvme_uring_cmd_end_io(struct request *req,
418 						blk_status_t err)
419 {
420 	struct io_uring_cmd *ioucmd = req->end_io_data;
421 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
422 
423 	if (nvme_req(req)->flags & NVME_REQ_CANCELLED)
424 		pdu->status = -EINTR;
425 	else
426 		pdu->status = nvme_req(req)->status;
427 	pdu->result = le64_to_cpu(nvme_req(req)->result.u64);
428 
429 	/*
430 	 * For iopoll, complete it directly. Note that using the uring_cmd
431 	 * helper for this is safe only because we check blk_rq_is_poll().
432 	 * As that returns false if we're NOT on a polled queue, then it's
433 	 * safe to use the polled completion helper.
434 	 *
435 	 * Otherwise, move the completion to task work.
436 	 */
437 	if (blk_rq_is_poll(req)) {
438 		if (pdu->bio)
439 			blk_rq_unmap_user(pdu->bio);
440 		io_uring_cmd_iopoll_done(ioucmd, pdu->result, pdu->status);
441 	} else {
442 		io_uring_cmd_do_in_task_lazy(ioucmd, nvme_uring_task_cb);
443 	}
444 
445 	return RQ_END_IO_FREE;
446 }
447 
448 static int nvme_uring_cmd_io(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
449 		struct io_uring_cmd *ioucmd, unsigned int issue_flags, bool vec)
450 {
451 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
452 	const struct nvme_uring_cmd *cmd = io_uring_sqe_cmd(ioucmd->sqe);
453 	struct request_queue *q = ns ? ns->queue : ctrl->admin_q;
454 	struct nvme_uring_data d;
455 	struct nvme_command c;
456 	struct request *req;
457 	blk_opf_t rq_flags = REQ_ALLOC_CACHE;
458 	blk_mq_req_flags_t blk_flags = 0;
459 	int ret;
460 
461 	c.common.opcode = READ_ONCE(cmd->opcode);
462 	c.common.flags = READ_ONCE(cmd->flags);
463 	if (c.common.flags)
464 		return -EINVAL;
465 
466 	c.common.command_id = 0;
467 	c.common.nsid = cpu_to_le32(cmd->nsid);
468 	if (!nvme_validate_passthru_nsid(ctrl, ns, le32_to_cpu(c.common.nsid)))
469 		return -EINVAL;
470 
471 	c.common.cdw2[0] = cpu_to_le32(READ_ONCE(cmd->cdw2));
472 	c.common.cdw2[1] = cpu_to_le32(READ_ONCE(cmd->cdw3));
473 	c.common.metadata = 0;
474 	c.common.dptr.prp1 = c.common.dptr.prp2 = 0;
475 	c.common.cdw10 = cpu_to_le32(READ_ONCE(cmd->cdw10));
476 	c.common.cdw11 = cpu_to_le32(READ_ONCE(cmd->cdw11));
477 	c.common.cdw12 = cpu_to_le32(READ_ONCE(cmd->cdw12));
478 	c.common.cdw13 = cpu_to_le32(READ_ONCE(cmd->cdw13));
479 	c.common.cdw14 = cpu_to_le32(READ_ONCE(cmd->cdw14));
480 	c.common.cdw15 = cpu_to_le32(READ_ONCE(cmd->cdw15));
481 
482 	if (!nvme_cmd_allowed(ns, &c, 0, ioucmd->file->f_mode & FMODE_WRITE))
483 		return -EACCES;
484 
485 	d.metadata = READ_ONCE(cmd->metadata);
486 	d.addr = READ_ONCE(cmd->addr);
487 	d.data_len = READ_ONCE(cmd->data_len);
488 	d.metadata_len = READ_ONCE(cmd->metadata_len);
489 	d.timeout_ms = READ_ONCE(cmd->timeout_ms);
490 
491 	if (issue_flags & IO_URING_F_NONBLOCK) {
492 		rq_flags |= REQ_NOWAIT;
493 		blk_flags = BLK_MQ_REQ_NOWAIT;
494 	}
495 	if (issue_flags & IO_URING_F_IOPOLL)
496 		rq_flags |= REQ_POLLED;
497 
498 	req = nvme_alloc_user_request(q, &c, rq_flags, blk_flags);
499 	if (IS_ERR(req))
500 		return PTR_ERR(req);
501 	req->timeout = d.timeout_ms ? msecs_to_jiffies(d.timeout_ms) : 0;
502 
503 	if (d.addr && d.data_len) {
504 		ret = nvme_map_user_request(req, d.addr,
505 			d.data_len, nvme_to_user_ptr(d.metadata),
506 			d.metadata_len, ioucmd, vec);
507 		if (ret)
508 			return ret;
509 	}
510 
511 	/* to free bio on completion, as req->bio will be null at that time */
512 	pdu->bio = req->bio;
513 	pdu->req = req;
514 	req->end_io_data = ioucmd;
515 	req->end_io = nvme_uring_cmd_end_io;
516 	blk_execute_rq_nowait(req, false);
517 	return -EIOCBQUEUED;
518 }
519 
520 static bool is_ctrl_ioctl(unsigned int cmd)
521 {
522 	if (cmd == NVME_IOCTL_ADMIN_CMD || cmd == NVME_IOCTL_ADMIN64_CMD)
523 		return true;
524 	if (is_sed_ioctl(cmd))
525 		return true;
526 	return false;
527 }
528 
529 static int nvme_ctrl_ioctl(struct nvme_ctrl *ctrl, unsigned int cmd,
530 		void __user *argp, bool open_for_write)
531 {
532 	switch (cmd) {
533 	case NVME_IOCTL_ADMIN_CMD:
534 		return nvme_user_cmd(ctrl, NULL, argp, 0, open_for_write);
535 	case NVME_IOCTL_ADMIN64_CMD:
536 		return nvme_user_cmd64(ctrl, NULL, argp, 0, open_for_write);
537 	default:
538 		return sed_ioctl(ctrl->opal_dev, cmd, argp);
539 	}
540 }
541 
542 #ifdef COMPAT_FOR_U64_ALIGNMENT
543 struct nvme_user_io32 {
544 	__u8	opcode;
545 	__u8	flags;
546 	__u16	control;
547 	__u16	nblocks;
548 	__u16	rsvd;
549 	__u64	metadata;
550 	__u64	addr;
551 	__u64	slba;
552 	__u32	dsmgmt;
553 	__u32	reftag;
554 	__u16	apptag;
555 	__u16	appmask;
556 } __attribute__((__packed__));
557 #define NVME_IOCTL_SUBMIT_IO32	_IOW('N', 0x42, struct nvme_user_io32)
558 #endif /* COMPAT_FOR_U64_ALIGNMENT */
559 
560 static int nvme_ns_ioctl(struct nvme_ns *ns, unsigned int cmd,
561 		void __user *argp, unsigned int flags, bool open_for_write)
562 {
563 	switch (cmd) {
564 	case NVME_IOCTL_ID:
565 		force_successful_syscall_return();
566 		return ns->head->ns_id;
567 	case NVME_IOCTL_IO_CMD:
568 		return nvme_user_cmd(ns->ctrl, ns, argp, flags, open_for_write);
569 	/*
570 	 * struct nvme_user_io can have different padding on some 32-bit ABIs.
571 	 * Just accept the compat version as all fields that are used are the
572 	 * same size and at the same offset.
573 	 */
574 #ifdef COMPAT_FOR_U64_ALIGNMENT
575 	case NVME_IOCTL_SUBMIT_IO32:
576 #endif
577 	case NVME_IOCTL_SUBMIT_IO:
578 		return nvme_submit_io(ns, argp);
579 	case NVME_IOCTL_IO64_CMD_VEC:
580 		flags |= NVME_IOCTL_VEC;
581 		fallthrough;
582 	case NVME_IOCTL_IO64_CMD:
583 		return nvme_user_cmd64(ns->ctrl, ns, argp, flags,
584 				       open_for_write);
585 	default:
586 		return -ENOTTY;
587 	}
588 }
589 
590 int nvme_ioctl(struct block_device *bdev, blk_mode_t mode,
591 		unsigned int cmd, unsigned long arg)
592 {
593 	struct nvme_ns *ns = bdev->bd_disk->private_data;
594 	bool open_for_write = mode & BLK_OPEN_WRITE;
595 	void __user *argp = (void __user *)arg;
596 	unsigned int flags = 0;
597 
598 	if (bdev_is_partition(bdev))
599 		flags |= NVME_IOCTL_PARTITION;
600 
601 	if (is_ctrl_ioctl(cmd))
602 		return nvme_ctrl_ioctl(ns->ctrl, cmd, argp, open_for_write);
603 	return nvme_ns_ioctl(ns, cmd, argp, flags, open_for_write);
604 }
605 
606 long nvme_ns_chr_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
607 {
608 	struct nvme_ns *ns =
609 		container_of(file_inode(file)->i_cdev, struct nvme_ns, cdev);
610 	bool open_for_write = file->f_mode & FMODE_WRITE;
611 	void __user *argp = (void __user *)arg;
612 
613 	if (is_ctrl_ioctl(cmd))
614 		return nvme_ctrl_ioctl(ns->ctrl, cmd, argp, open_for_write);
615 	return nvme_ns_ioctl(ns, cmd, argp, 0, open_for_write);
616 }
617 
618 static int nvme_uring_cmd_checks(unsigned int issue_flags)
619 {
620 
621 	/* NVMe passthrough requires big SQE/CQE support */
622 	if ((issue_flags & (IO_URING_F_SQE128|IO_URING_F_CQE32)) !=
623 	    (IO_URING_F_SQE128|IO_URING_F_CQE32))
624 		return -EOPNOTSUPP;
625 	return 0;
626 }
627 
628 static int nvme_ns_uring_cmd(struct nvme_ns *ns, struct io_uring_cmd *ioucmd,
629 			     unsigned int issue_flags)
630 {
631 	struct nvme_ctrl *ctrl = ns->ctrl;
632 	int ret;
633 
634 	ret = nvme_uring_cmd_checks(issue_flags);
635 	if (ret)
636 		return ret;
637 
638 	switch (ioucmd->cmd_op) {
639 	case NVME_URING_CMD_IO:
640 		ret = nvme_uring_cmd_io(ctrl, ns, ioucmd, issue_flags, false);
641 		break;
642 	case NVME_URING_CMD_IO_VEC:
643 		ret = nvme_uring_cmd_io(ctrl, ns, ioucmd, issue_flags, true);
644 		break;
645 	default:
646 		ret = -ENOTTY;
647 	}
648 
649 	return ret;
650 }
651 
652 int nvme_ns_chr_uring_cmd(struct io_uring_cmd *ioucmd, unsigned int issue_flags)
653 {
654 	struct nvme_ns *ns = container_of(file_inode(ioucmd->file)->i_cdev,
655 			struct nvme_ns, cdev);
656 
657 	return nvme_ns_uring_cmd(ns, ioucmd, issue_flags);
658 }
659 
660 int nvme_ns_chr_uring_cmd_iopoll(struct io_uring_cmd *ioucmd,
661 				 struct io_comp_batch *iob,
662 				 unsigned int poll_flags)
663 {
664 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
665 	struct request *req = pdu->req;
666 
667 	if (req && blk_rq_is_poll(req))
668 		return blk_rq_poll(req, iob, poll_flags);
669 	return 0;
670 }
671 #ifdef CONFIG_NVME_MULTIPATH
672 static int nvme_ns_head_ctrl_ioctl(struct nvme_ns *ns, unsigned int cmd,
673 		void __user *argp, struct nvme_ns_head *head, int srcu_idx,
674 		bool open_for_write)
675 	__releases(&head->srcu)
676 {
677 	struct nvme_ctrl *ctrl = ns->ctrl;
678 	int ret;
679 
680 	nvme_get_ctrl(ns->ctrl);
681 	srcu_read_unlock(&head->srcu, srcu_idx);
682 	ret = nvme_ctrl_ioctl(ns->ctrl, cmd, argp, open_for_write);
683 
684 	nvme_put_ctrl(ctrl);
685 	return ret;
686 }
687 
688 int nvme_ns_head_ioctl(struct block_device *bdev, blk_mode_t mode,
689 		unsigned int cmd, unsigned long arg)
690 {
691 	struct nvme_ns_head *head = bdev->bd_disk->private_data;
692 	bool open_for_write = mode & BLK_OPEN_WRITE;
693 	void __user *argp = (void __user *)arg;
694 	struct nvme_ns *ns;
695 	int srcu_idx, ret = -EWOULDBLOCK;
696 	unsigned int flags = 0;
697 
698 	if (bdev_is_partition(bdev))
699 		flags |= NVME_IOCTL_PARTITION;
700 
701 	srcu_idx = srcu_read_lock(&head->srcu);
702 	ns = nvme_find_path(head);
703 	if (!ns)
704 		goto out_unlock;
705 
706 	/*
707 	 * Handle ioctls that apply to the controller instead of the namespace
708 	 * seperately and drop the ns SRCU reference early.  This avoids a
709 	 * deadlock when deleting namespaces using the passthrough interface.
710 	 */
711 	if (is_ctrl_ioctl(cmd))
712 		return nvme_ns_head_ctrl_ioctl(ns, cmd, argp, head, srcu_idx,
713 					       open_for_write);
714 
715 	ret = nvme_ns_ioctl(ns, cmd, argp, flags, open_for_write);
716 out_unlock:
717 	srcu_read_unlock(&head->srcu, srcu_idx);
718 	return ret;
719 }
720 
721 long nvme_ns_head_chr_ioctl(struct file *file, unsigned int cmd,
722 		unsigned long arg)
723 {
724 	bool open_for_write = file->f_mode & FMODE_WRITE;
725 	struct cdev *cdev = file_inode(file)->i_cdev;
726 	struct nvme_ns_head *head =
727 		container_of(cdev, struct nvme_ns_head, cdev);
728 	void __user *argp = (void __user *)arg;
729 	struct nvme_ns *ns;
730 	int srcu_idx, ret = -EWOULDBLOCK;
731 
732 	srcu_idx = srcu_read_lock(&head->srcu);
733 	ns = nvme_find_path(head);
734 	if (!ns)
735 		goto out_unlock;
736 
737 	if (is_ctrl_ioctl(cmd))
738 		return nvme_ns_head_ctrl_ioctl(ns, cmd, argp, head, srcu_idx,
739 				open_for_write);
740 
741 	ret = nvme_ns_ioctl(ns, cmd, argp, 0, open_for_write);
742 out_unlock:
743 	srcu_read_unlock(&head->srcu, srcu_idx);
744 	return ret;
745 }
746 
747 int nvme_ns_head_chr_uring_cmd(struct io_uring_cmd *ioucmd,
748 		unsigned int issue_flags)
749 {
750 	struct cdev *cdev = file_inode(ioucmd->file)->i_cdev;
751 	struct nvme_ns_head *head = container_of(cdev, struct nvme_ns_head, cdev);
752 	int srcu_idx = srcu_read_lock(&head->srcu);
753 	struct nvme_ns *ns = nvme_find_path(head);
754 	int ret = -EINVAL;
755 
756 	if (ns)
757 		ret = nvme_ns_uring_cmd(ns, ioucmd, issue_flags);
758 	srcu_read_unlock(&head->srcu, srcu_idx);
759 	return ret;
760 }
761 #endif /* CONFIG_NVME_MULTIPATH */
762 
763 int nvme_dev_uring_cmd(struct io_uring_cmd *ioucmd, unsigned int issue_flags)
764 {
765 	struct nvme_ctrl *ctrl = ioucmd->file->private_data;
766 	int ret;
767 
768 	/* IOPOLL not supported yet */
769 	if (issue_flags & IO_URING_F_IOPOLL)
770 		return -EOPNOTSUPP;
771 
772 	ret = nvme_uring_cmd_checks(issue_flags);
773 	if (ret)
774 		return ret;
775 
776 	switch (ioucmd->cmd_op) {
777 	case NVME_URING_CMD_ADMIN:
778 		ret = nvme_uring_cmd_io(ctrl, NULL, ioucmd, issue_flags, false);
779 		break;
780 	case NVME_URING_CMD_ADMIN_VEC:
781 		ret = nvme_uring_cmd_io(ctrl, NULL, ioucmd, issue_flags, true);
782 		break;
783 	default:
784 		ret = -ENOTTY;
785 	}
786 
787 	return ret;
788 }
789 
790 static int nvme_dev_user_cmd(struct nvme_ctrl *ctrl, void __user *argp,
791 		bool open_for_write)
792 {
793 	struct nvme_ns *ns;
794 	int ret, srcu_idx;
795 
796 	srcu_idx = srcu_read_lock(&ctrl->srcu);
797 	if (list_empty(&ctrl->namespaces)) {
798 		ret = -ENOTTY;
799 		goto out_unlock;
800 	}
801 
802 	ns = list_first_or_null_rcu(&ctrl->namespaces, struct nvme_ns, list);
803 	if (ns != list_last_entry(&ctrl->namespaces, struct nvme_ns, list)) {
804 		dev_warn(ctrl->device,
805 			"NVME_IOCTL_IO_CMD not supported when multiple namespaces present!\n");
806 		ret = -EINVAL;
807 		goto out_unlock;
808 	}
809 
810 	dev_warn(ctrl->device,
811 		"using deprecated NVME_IOCTL_IO_CMD ioctl on the char device!\n");
812 	if (!nvme_get_ns(ns)) {
813 		ret = -ENXIO;
814 		goto out_unlock;
815 	}
816 	srcu_read_unlock(&ctrl->srcu, srcu_idx);
817 
818 	ret = nvme_user_cmd(ctrl, ns, argp, 0, open_for_write);
819 	nvme_put_ns(ns);
820 	return ret;
821 
822 out_unlock:
823 	srcu_read_unlock(&ctrl->srcu, srcu_idx);
824 	return ret;
825 }
826 
827 long nvme_dev_ioctl(struct file *file, unsigned int cmd,
828 		unsigned long arg)
829 {
830 	bool open_for_write = file->f_mode & FMODE_WRITE;
831 	struct nvme_ctrl *ctrl = file->private_data;
832 	void __user *argp = (void __user *)arg;
833 
834 	switch (cmd) {
835 	case NVME_IOCTL_ADMIN_CMD:
836 		return nvme_user_cmd(ctrl, NULL, argp, 0, open_for_write);
837 	case NVME_IOCTL_ADMIN64_CMD:
838 		return nvme_user_cmd64(ctrl, NULL, argp, 0, open_for_write);
839 	case NVME_IOCTL_IO_CMD:
840 		return nvme_dev_user_cmd(ctrl, argp, open_for_write);
841 	case NVME_IOCTL_RESET:
842 		if (!capable(CAP_SYS_ADMIN))
843 			return -EACCES;
844 		dev_warn(ctrl->device, "resetting controller\n");
845 		return nvme_reset_ctrl_sync(ctrl);
846 	case NVME_IOCTL_SUBSYS_RESET:
847 		if (!capable(CAP_SYS_ADMIN))
848 			return -EACCES;
849 		return nvme_reset_subsystem(ctrl);
850 	case NVME_IOCTL_RESCAN:
851 		if (!capable(CAP_SYS_ADMIN))
852 			return -EACCES;
853 		nvme_queue_scan(ctrl);
854 		return 0;
855 	default:
856 		return -ENOTTY;
857 	}
858 }
859