xref: /linux/drivers/nvme/host/ioctl.c (revision 11a299a7933e03c83818b431e6a1c53ad387423d)
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 		u32 meta_seed, 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 						meta_seed);
157 		if (ret)
158 			goto out_unmap;
159 	}
160 
161 	return ret;
162 
163 out_unmap:
164 	if (bio)
165 		blk_rq_unmap_user(bio);
166 out:
167 	blk_mq_free_request(req);
168 	return ret;
169 }
170 
171 static int nvme_submit_user_cmd(struct request_queue *q,
172 		struct nvme_command *cmd, u64 ubuffer, unsigned bufflen,
173 		void __user *meta_buffer, unsigned meta_len, u32 meta_seed,
174 		u64 *result, unsigned timeout, unsigned int flags)
175 {
176 	struct nvme_ns *ns = q->queuedata;
177 	struct nvme_ctrl *ctrl;
178 	struct request *req;
179 	struct bio *bio;
180 	u32 effects;
181 	int ret;
182 
183 	req = nvme_alloc_user_request(q, cmd, 0, 0);
184 	if (IS_ERR(req))
185 		return PTR_ERR(req);
186 
187 	req->timeout = timeout;
188 	if (ubuffer && bufflen) {
189 		ret = nvme_map_user_request(req, ubuffer, bufflen, meta_buffer,
190 				meta_len, meta_seed, NULL, flags);
191 		if (ret)
192 			return ret;
193 	}
194 
195 	bio = req->bio;
196 	ctrl = nvme_req(req)->ctrl;
197 
198 	effects = nvme_passthru_start(ctrl, ns, cmd->common.opcode);
199 	ret = nvme_execute_rq(req, false);
200 	if (result)
201 		*result = le64_to_cpu(nvme_req(req)->result.u64);
202 	if (bio)
203 		blk_rq_unmap_user(bio);
204 	blk_mq_free_request(req);
205 
206 	if (effects)
207 		nvme_passthru_end(ctrl, ns, effects, cmd, ret);
208 
209 	return ret;
210 }
211 
212 static int nvme_submit_io(struct nvme_ns *ns, struct nvme_user_io __user *uio)
213 {
214 	struct nvme_user_io io;
215 	struct nvme_command c;
216 	unsigned length, meta_len;
217 	void __user *metadata;
218 
219 	if (copy_from_user(&io, uio, sizeof(io)))
220 		return -EFAULT;
221 	if (io.flags)
222 		return -EINVAL;
223 
224 	switch (io.opcode) {
225 	case nvme_cmd_write:
226 	case nvme_cmd_read:
227 	case nvme_cmd_compare:
228 		break;
229 	default:
230 		return -EINVAL;
231 	}
232 
233 	length = (io.nblocks + 1) << ns->head->lba_shift;
234 
235 	if ((io.control & NVME_RW_PRINFO_PRACT) &&
236 	    (ns->head->ms == ns->head->pi_size)) {
237 		/*
238 		 * Protection information is stripped/inserted by the
239 		 * controller.
240 		 */
241 		if (nvme_to_user_ptr(io.metadata))
242 			return -EINVAL;
243 		meta_len = 0;
244 		metadata = NULL;
245 	} else {
246 		meta_len = (io.nblocks + 1) * ns->head->ms;
247 		metadata = nvme_to_user_ptr(io.metadata);
248 	}
249 
250 	if (ns->head->features & NVME_NS_EXT_LBAS) {
251 		length += meta_len;
252 		meta_len = 0;
253 	} else if (meta_len) {
254 		if ((io.metadata & 3) || !io.metadata)
255 			return -EINVAL;
256 	}
257 
258 	memset(&c, 0, sizeof(c));
259 	c.rw.opcode = io.opcode;
260 	c.rw.flags = io.flags;
261 	c.rw.nsid = cpu_to_le32(ns->head->ns_id);
262 	c.rw.slba = cpu_to_le64(io.slba);
263 	c.rw.length = cpu_to_le16(io.nblocks);
264 	c.rw.control = cpu_to_le16(io.control);
265 	c.rw.dsmgmt = cpu_to_le32(io.dsmgmt);
266 	c.rw.reftag = cpu_to_le32(io.reftag);
267 	c.rw.lbat = cpu_to_le16(io.apptag);
268 	c.rw.lbatm = cpu_to_le16(io.appmask);
269 
270 	return nvme_submit_user_cmd(ns->queue, &c, io.addr, length, metadata,
271 			meta_len, lower_32_bits(io.slba), NULL, 0, 0);
272 }
273 
274 static bool nvme_validate_passthru_nsid(struct nvme_ctrl *ctrl,
275 					struct nvme_ns *ns, __u32 nsid)
276 {
277 	if (ns && nsid != ns->head->ns_id) {
278 		dev_err(ctrl->device,
279 			"%s: nsid (%u) in cmd does not match nsid (%u)"
280 			"of namespace\n",
281 			current->comm, nsid, ns->head->ns_id);
282 		return false;
283 	}
284 
285 	return true;
286 }
287 
288 static int nvme_user_cmd(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
289 		struct nvme_passthru_cmd __user *ucmd, unsigned int flags,
290 		bool open_for_write)
291 {
292 	struct nvme_passthru_cmd cmd;
293 	struct nvme_command c;
294 	unsigned timeout = 0;
295 	u64 result;
296 	int status;
297 
298 	if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
299 		return -EFAULT;
300 	if (cmd.flags)
301 		return -EINVAL;
302 	if (!nvme_validate_passthru_nsid(ctrl, ns, cmd.nsid))
303 		return -EINVAL;
304 
305 	memset(&c, 0, sizeof(c));
306 	c.common.opcode = cmd.opcode;
307 	c.common.flags = cmd.flags;
308 	c.common.nsid = cpu_to_le32(cmd.nsid);
309 	c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
310 	c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
311 	c.common.cdw10 = cpu_to_le32(cmd.cdw10);
312 	c.common.cdw11 = cpu_to_le32(cmd.cdw11);
313 	c.common.cdw12 = cpu_to_le32(cmd.cdw12);
314 	c.common.cdw13 = cpu_to_le32(cmd.cdw13);
315 	c.common.cdw14 = cpu_to_le32(cmd.cdw14);
316 	c.common.cdw15 = cpu_to_le32(cmd.cdw15);
317 
318 	if (!nvme_cmd_allowed(ns, &c, 0, open_for_write))
319 		return -EACCES;
320 
321 	if (cmd.timeout_ms)
322 		timeout = msecs_to_jiffies(cmd.timeout_ms);
323 
324 	status = nvme_submit_user_cmd(ns ? ns->queue : ctrl->admin_q, &c,
325 			cmd.addr, cmd.data_len, nvme_to_user_ptr(cmd.metadata),
326 			cmd.metadata_len, 0, &result, timeout, 0);
327 
328 	if (status >= 0) {
329 		if (put_user(result, &ucmd->result))
330 			return -EFAULT;
331 	}
332 
333 	return status;
334 }
335 
336 static int nvme_user_cmd64(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
337 		struct nvme_passthru_cmd64 __user *ucmd, unsigned int flags,
338 		bool open_for_write)
339 {
340 	struct nvme_passthru_cmd64 cmd;
341 	struct nvme_command c;
342 	unsigned timeout = 0;
343 	int status;
344 
345 	if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
346 		return -EFAULT;
347 	if (cmd.flags)
348 		return -EINVAL;
349 	if (!nvme_validate_passthru_nsid(ctrl, ns, cmd.nsid))
350 		return -EINVAL;
351 
352 	memset(&c, 0, sizeof(c));
353 	c.common.opcode = cmd.opcode;
354 	c.common.flags = cmd.flags;
355 	c.common.nsid = cpu_to_le32(cmd.nsid);
356 	c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
357 	c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
358 	c.common.cdw10 = cpu_to_le32(cmd.cdw10);
359 	c.common.cdw11 = cpu_to_le32(cmd.cdw11);
360 	c.common.cdw12 = cpu_to_le32(cmd.cdw12);
361 	c.common.cdw13 = cpu_to_le32(cmd.cdw13);
362 	c.common.cdw14 = cpu_to_le32(cmd.cdw14);
363 	c.common.cdw15 = cpu_to_le32(cmd.cdw15);
364 
365 	if (!nvme_cmd_allowed(ns, &c, flags, open_for_write))
366 		return -EACCES;
367 
368 	if (cmd.timeout_ms)
369 		timeout = msecs_to_jiffies(cmd.timeout_ms);
370 
371 	status = nvme_submit_user_cmd(ns ? ns->queue : ctrl->admin_q, &c,
372 			cmd.addr, cmd.data_len, nvme_to_user_ptr(cmd.metadata),
373 			cmd.metadata_len, 0, &cmd.result, timeout, flags);
374 
375 	if (status >= 0) {
376 		if (put_user(cmd.result, &ucmd->result))
377 			return -EFAULT;
378 	}
379 
380 	return status;
381 }
382 
383 struct nvme_uring_data {
384 	__u64	metadata;
385 	__u64	addr;
386 	__u32	data_len;
387 	__u32	metadata_len;
388 	__u32	timeout_ms;
389 };
390 
391 /*
392  * This overlays struct io_uring_cmd pdu.
393  * Expect build errors if this grows larger than that.
394  */
395 struct nvme_uring_cmd_pdu {
396 	struct request *req;
397 	struct bio *bio;
398 	u64 result;
399 	int status;
400 };
401 
402 static inline struct nvme_uring_cmd_pdu *nvme_uring_cmd_pdu(
403 		struct io_uring_cmd *ioucmd)
404 {
405 	return (struct nvme_uring_cmd_pdu *)&ioucmd->pdu;
406 }
407 
408 static void nvme_uring_task_cb(struct io_uring_cmd *ioucmd,
409 			       unsigned issue_flags)
410 {
411 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
412 
413 	if (pdu->bio)
414 		blk_rq_unmap_user(pdu->bio);
415 	io_uring_cmd_done(ioucmd, pdu->status, pdu->result, issue_flags);
416 }
417 
418 static enum rq_end_io_ret nvme_uring_cmd_end_io(struct request *req,
419 						blk_status_t err)
420 {
421 	struct io_uring_cmd *ioucmd = req->end_io_data;
422 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
423 
424 	if (nvme_req(req)->flags & NVME_REQ_CANCELLED)
425 		pdu->status = -EINTR;
426 	else
427 		pdu->status = nvme_req(req)->status;
428 	pdu->result = le64_to_cpu(nvme_req(req)->result.u64);
429 
430 	/*
431 	 * For iopoll, complete it directly. Note that using the uring_cmd
432 	 * helper for this is safe only because we check blk_rq_is_poll().
433 	 * As that returns false if we're NOT on a polled queue, then it's
434 	 * safe to use the polled completion helper.
435 	 *
436 	 * Otherwise, move the completion to task work.
437 	 */
438 	if (blk_rq_is_poll(req)) {
439 		if (pdu->bio)
440 			blk_rq_unmap_user(pdu->bio);
441 		io_uring_cmd_iopoll_done(ioucmd, pdu->result, pdu->status);
442 	} else {
443 		io_uring_cmd_do_in_task_lazy(ioucmd, nvme_uring_task_cb);
444 	}
445 
446 	return RQ_END_IO_FREE;
447 }
448 
449 static int nvme_uring_cmd_io(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
450 		struct io_uring_cmd *ioucmd, unsigned int issue_flags, bool vec)
451 {
452 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
453 	const struct nvme_uring_cmd *cmd = io_uring_sqe_cmd(ioucmd->sqe);
454 	struct request_queue *q = ns ? ns->queue : ctrl->admin_q;
455 	struct nvme_uring_data d;
456 	struct nvme_command c;
457 	struct request *req;
458 	blk_opf_t rq_flags = REQ_ALLOC_CACHE;
459 	blk_mq_req_flags_t blk_flags = 0;
460 	int ret;
461 
462 	c.common.opcode = READ_ONCE(cmd->opcode);
463 	c.common.flags = READ_ONCE(cmd->flags);
464 	if (c.common.flags)
465 		return -EINVAL;
466 
467 	c.common.command_id = 0;
468 	c.common.nsid = cpu_to_le32(cmd->nsid);
469 	if (!nvme_validate_passthru_nsid(ctrl, ns, le32_to_cpu(c.common.nsid)))
470 		return -EINVAL;
471 
472 	c.common.cdw2[0] = cpu_to_le32(READ_ONCE(cmd->cdw2));
473 	c.common.cdw2[1] = cpu_to_le32(READ_ONCE(cmd->cdw3));
474 	c.common.metadata = 0;
475 	c.common.dptr.prp1 = c.common.dptr.prp2 = 0;
476 	c.common.cdw10 = cpu_to_le32(READ_ONCE(cmd->cdw10));
477 	c.common.cdw11 = cpu_to_le32(READ_ONCE(cmd->cdw11));
478 	c.common.cdw12 = cpu_to_le32(READ_ONCE(cmd->cdw12));
479 	c.common.cdw13 = cpu_to_le32(READ_ONCE(cmd->cdw13));
480 	c.common.cdw14 = cpu_to_le32(READ_ONCE(cmd->cdw14));
481 	c.common.cdw15 = cpu_to_le32(READ_ONCE(cmd->cdw15));
482 
483 	if (!nvme_cmd_allowed(ns, &c, 0, ioucmd->file->f_mode & FMODE_WRITE))
484 		return -EACCES;
485 
486 	d.metadata = READ_ONCE(cmd->metadata);
487 	d.addr = READ_ONCE(cmd->addr);
488 	d.data_len = READ_ONCE(cmd->data_len);
489 	d.metadata_len = READ_ONCE(cmd->metadata_len);
490 	d.timeout_ms = READ_ONCE(cmd->timeout_ms);
491 
492 	if (issue_flags & IO_URING_F_NONBLOCK) {
493 		rq_flags |= REQ_NOWAIT;
494 		blk_flags = BLK_MQ_REQ_NOWAIT;
495 	}
496 	if (issue_flags & IO_URING_F_IOPOLL)
497 		rq_flags |= REQ_POLLED;
498 
499 	req = nvme_alloc_user_request(q, &c, rq_flags, blk_flags);
500 	if (IS_ERR(req))
501 		return PTR_ERR(req);
502 	req->timeout = d.timeout_ms ? msecs_to_jiffies(d.timeout_ms) : 0;
503 
504 	if (d.addr && d.data_len) {
505 		ret = nvme_map_user_request(req, d.addr,
506 			d.data_len, nvme_to_user_ptr(d.metadata),
507 			d.metadata_len, 0, ioucmd, vec);
508 		if (ret)
509 			return ret;
510 	}
511 
512 	/* to free bio on completion, as req->bio will be null at that time */
513 	pdu->bio = req->bio;
514 	pdu->req = req;
515 	req->end_io_data = ioucmd;
516 	req->end_io = nvme_uring_cmd_end_io;
517 	blk_execute_rq_nowait(req, false);
518 	return -EIOCBQUEUED;
519 }
520 
521 static bool is_ctrl_ioctl(unsigned int cmd)
522 {
523 	if (cmd == NVME_IOCTL_ADMIN_CMD || cmd == NVME_IOCTL_ADMIN64_CMD)
524 		return true;
525 	if (is_sed_ioctl(cmd))
526 		return true;
527 	return false;
528 }
529 
530 static int nvme_ctrl_ioctl(struct nvme_ctrl *ctrl, unsigned int cmd,
531 		void __user *argp, bool open_for_write)
532 {
533 	switch (cmd) {
534 	case NVME_IOCTL_ADMIN_CMD:
535 		return nvme_user_cmd(ctrl, NULL, argp, 0, open_for_write);
536 	case NVME_IOCTL_ADMIN64_CMD:
537 		return nvme_user_cmd64(ctrl, NULL, argp, 0, open_for_write);
538 	default:
539 		return sed_ioctl(ctrl->opal_dev, cmd, argp);
540 	}
541 }
542 
543 #ifdef COMPAT_FOR_U64_ALIGNMENT
544 struct nvme_user_io32 {
545 	__u8	opcode;
546 	__u8	flags;
547 	__u16	control;
548 	__u16	nblocks;
549 	__u16	rsvd;
550 	__u64	metadata;
551 	__u64	addr;
552 	__u64	slba;
553 	__u32	dsmgmt;
554 	__u32	reftag;
555 	__u16	apptag;
556 	__u16	appmask;
557 } __attribute__((__packed__));
558 #define NVME_IOCTL_SUBMIT_IO32	_IOW('N', 0x42, struct nvme_user_io32)
559 #endif /* COMPAT_FOR_U64_ALIGNMENT */
560 
561 static int nvme_ns_ioctl(struct nvme_ns *ns, unsigned int cmd,
562 		void __user *argp, unsigned int flags, bool open_for_write)
563 {
564 	switch (cmd) {
565 	case NVME_IOCTL_ID:
566 		force_successful_syscall_return();
567 		return ns->head->ns_id;
568 	case NVME_IOCTL_IO_CMD:
569 		return nvme_user_cmd(ns->ctrl, ns, argp, flags, open_for_write);
570 	/*
571 	 * struct nvme_user_io can have different padding on some 32-bit ABIs.
572 	 * Just accept the compat version as all fields that are used are the
573 	 * same size and at the same offset.
574 	 */
575 #ifdef COMPAT_FOR_U64_ALIGNMENT
576 	case NVME_IOCTL_SUBMIT_IO32:
577 #endif
578 	case NVME_IOCTL_SUBMIT_IO:
579 		return nvme_submit_io(ns, argp);
580 	case NVME_IOCTL_IO64_CMD_VEC:
581 		flags |= NVME_IOCTL_VEC;
582 		fallthrough;
583 	case NVME_IOCTL_IO64_CMD:
584 		return nvme_user_cmd64(ns->ctrl, ns, argp, flags,
585 				       open_for_write);
586 	default:
587 		return -ENOTTY;
588 	}
589 }
590 
591 int nvme_ioctl(struct block_device *bdev, blk_mode_t mode,
592 		unsigned int cmd, unsigned long arg)
593 {
594 	struct nvme_ns *ns = bdev->bd_disk->private_data;
595 	bool open_for_write = mode & BLK_OPEN_WRITE;
596 	void __user *argp = (void __user *)arg;
597 	unsigned int flags = 0;
598 
599 	if (bdev_is_partition(bdev))
600 		flags |= NVME_IOCTL_PARTITION;
601 
602 	if (is_ctrl_ioctl(cmd))
603 		return nvme_ctrl_ioctl(ns->ctrl, cmd, argp, open_for_write);
604 	return nvme_ns_ioctl(ns, cmd, argp, flags, open_for_write);
605 }
606 
607 long nvme_ns_chr_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
608 {
609 	struct nvme_ns *ns =
610 		container_of(file_inode(file)->i_cdev, struct nvme_ns, cdev);
611 	bool open_for_write = file->f_mode & FMODE_WRITE;
612 	void __user *argp = (void __user *)arg;
613 
614 	if (is_ctrl_ioctl(cmd))
615 		return nvme_ctrl_ioctl(ns->ctrl, cmd, argp, open_for_write);
616 	return nvme_ns_ioctl(ns, cmd, argp, 0, open_for_write);
617 }
618 
619 static int nvme_uring_cmd_checks(unsigned int issue_flags)
620 {
621 
622 	/* NVMe passthrough requires big SQE/CQE support */
623 	if ((issue_flags & (IO_URING_F_SQE128|IO_URING_F_CQE32)) !=
624 	    (IO_URING_F_SQE128|IO_URING_F_CQE32))
625 		return -EOPNOTSUPP;
626 	return 0;
627 }
628 
629 static int nvme_ns_uring_cmd(struct nvme_ns *ns, struct io_uring_cmd *ioucmd,
630 			     unsigned int issue_flags)
631 {
632 	struct nvme_ctrl *ctrl = ns->ctrl;
633 	int ret;
634 
635 	BUILD_BUG_ON(sizeof(struct nvme_uring_cmd_pdu) > sizeof(ioucmd->pdu));
636 
637 	ret = nvme_uring_cmd_checks(issue_flags);
638 	if (ret)
639 		return ret;
640 
641 	switch (ioucmd->cmd_op) {
642 	case NVME_URING_CMD_IO:
643 		ret = nvme_uring_cmd_io(ctrl, ns, ioucmd, issue_flags, false);
644 		break;
645 	case NVME_URING_CMD_IO_VEC:
646 		ret = nvme_uring_cmd_io(ctrl, ns, ioucmd, issue_flags, true);
647 		break;
648 	default:
649 		ret = -ENOTTY;
650 	}
651 
652 	return ret;
653 }
654 
655 int nvme_ns_chr_uring_cmd(struct io_uring_cmd *ioucmd, unsigned int issue_flags)
656 {
657 	struct nvme_ns *ns = container_of(file_inode(ioucmd->file)->i_cdev,
658 			struct nvme_ns, cdev);
659 
660 	return nvme_ns_uring_cmd(ns, ioucmd, issue_flags);
661 }
662 
663 int nvme_ns_chr_uring_cmd_iopoll(struct io_uring_cmd *ioucmd,
664 				 struct io_comp_batch *iob,
665 				 unsigned int poll_flags)
666 {
667 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
668 	struct request *req = pdu->req;
669 
670 	if (req && blk_rq_is_poll(req))
671 		return blk_rq_poll(req, iob, poll_flags);
672 	return 0;
673 }
674 #ifdef CONFIG_NVME_MULTIPATH
675 static int nvme_ns_head_ctrl_ioctl(struct nvme_ns *ns, unsigned int cmd,
676 		void __user *argp, struct nvme_ns_head *head, int srcu_idx,
677 		bool open_for_write)
678 	__releases(&head->srcu)
679 {
680 	struct nvme_ctrl *ctrl = ns->ctrl;
681 	int ret;
682 
683 	nvme_get_ctrl(ns->ctrl);
684 	srcu_read_unlock(&head->srcu, srcu_idx);
685 	ret = nvme_ctrl_ioctl(ns->ctrl, cmd, argp, open_for_write);
686 
687 	nvme_put_ctrl(ctrl);
688 	return ret;
689 }
690 
691 int nvme_ns_head_ioctl(struct block_device *bdev, blk_mode_t mode,
692 		unsigned int cmd, unsigned long arg)
693 {
694 	struct nvme_ns_head *head = bdev->bd_disk->private_data;
695 	bool open_for_write = mode & BLK_OPEN_WRITE;
696 	void __user *argp = (void __user *)arg;
697 	struct nvme_ns *ns;
698 	int srcu_idx, ret = -EWOULDBLOCK;
699 	unsigned int flags = 0;
700 
701 	if (bdev_is_partition(bdev))
702 		flags |= NVME_IOCTL_PARTITION;
703 
704 	srcu_idx = srcu_read_lock(&head->srcu);
705 	ns = nvme_find_path(head);
706 	if (!ns)
707 		goto out_unlock;
708 
709 	/*
710 	 * Handle ioctls that apply to the controller instead of the namespace
711 	 * seperately and drop the ns SRCU reference early.  This avoids a
712 	 * deadlock when deleting namespaces using the passthrough interface.
713 	 */
714 	if (is_ctrl_ioctl(cmd))
715 		return nvme_ns_head_ctrl_ioctl(ns, cmd, argp, head, srcu_idx,
716 					       open_for_write);
717 
718 	ret = nvme_ns_ioctl(ns, cmd, argp, flags, open_for_write);
719 out_unlock:
720 	srcu_read_unlock(&head->srcu, srcu_idx);
721 	return ret;
722 }
723 
724 long nvme_ns_head_chr_ioctl(struct file *file, unsigned int cmd,
725 		unsigned long arg)
726 {
727 	bool open_for_write = file->f_mode & FMODE_WRITE;
728 	struct cdev *cdev = file_inode(file)->i_cdev;
729 	struct nvme_ns_head *head =
730 		container_of(cdev, struct nvme_ns_head, cdev);
731 	void __user *argp = (void __user *)arg;
732 	struct nvme_ns *ns;
733 	int srcu_idx, ret = -EWOULDBLOCK;
734 
735 	srcu_idx = srcu_read_lock(&head->srcu);
736 	ns = nvme_find_path(head);
737 	if (!ns)
738 		goto out_unlock;
739 
740 	if (is_ctrl_ioctl(cmd))
741 		return nvme_ns_head_ctrl_ioctl(ns, cmd, argp, head, srcu_idx,
742 				open_for_write);
743 
744 	ret = nvme_ns_ioctl(ns, cmd, argp, 0, open_for_write);
745 out_unlock:
746 	srcu_read_unlock(&head->srcu, srcu_idx);
747 	return ret;
748 }
749 
750 int nvme_ns_head_chr_uring_cmd(struct io_uring_cmd *ioucmd,
751 		unsigned int issue_flags)
752 {
753 	struct cdev *cdev = file_inode(ioucmd->file)->i_cdev;
754 	struct nvme_ns_head *head = container_of(cdev, struct nvme_ns_head, cdev);
755 	int srcu_idx = srcu_read_lock(&head->srcu);
756 	struct nvme_ns *ns = nvme_find_path(head);
757 	int ret = -EINVAL;
758 
759 	if (ns)
760 		ret = nvme_ns_uring_cmd(ns, ioucmd, issue_flags);
761 	srcu_read_unlock(&head->srcu, srcu_idx);
762 	return ret;
763 }
764 #endif /* CONFIG_NVME_MULTIPATH */
765 
766 int nvme_dev_uring_cmd(struct io_uring_cmd *ioucmd, unsigned int issue_flags)
767 {
768 	struct nvme_ctrl *ctrl = ioucmd->file->private_data;
769 	int ret;
770 
771 	/* IOPOLL not supported yet */
772 	if (issue_flags & IO_URING_F_IOPOLL)
773 		return -EOPNOTSUPP;
774 
775 	ret = nvme_uring_cmd_checks(issue_flags);
776 	if (ret)
777 		return ret;
778 
779 	switch (ioucmd->cmd_op) {
780 	case NVME_URING_CMD_ADMIN:
781 		ret = nvme_uring_cmd_io(ctrl, NULL, ioucmd, issue_flags, false);
782 		break;
783 	case NVME_URING_CMD_ADMIN_VEC:
784 		ret = nvme_uring_cmd_io(ctrl, NULL, ioucmd, issue_flags, true);
785 		break;
786 	default:
787 		ret = -ENOTTY;
788 	}
789 
790 	return ret;
791 }
792 
793 static int nvme_dev_user_cmd(struct nvme_ctrl *ctrl, void __user *argp,
794 		bool open_for_write)
795 {
796 	struct nvme_ns *ns;
797 	int ret, srcu_idx;
798 
799 	srcu_idx = srcu_read_lock(&ctrl->srcu);
800 	if (list_empty(&ctrl->namespaces)) {
801 		ret = -ENOTTY;
802 		goto out_unlock;
803 	}
804 
805 	ns = list_first_or_null_rcu(&ctrl->namespaces, struct nvme_ns, list);
806 	if (ns != list_last_entry(&ctrl->namespaces, struct nvme_ns, list)) {
807 		dev_warn(ctrl->device,
808 			"NVME_IOCTL_IO_CMD not supported when multiple namespaces present!\n");
809 		ret = -EINVAL;
810 		goto out_unlock;
811 	}
812 
813 	dev_warn(ctrl->device,
814 		"using deprecated NVME_IOCTL_IO_CMD ioctl on the char device!\n");
815 	if (!nvme_get_ns(ns)) {
816 		ret = -ENXIO;
817 		goto out_unlock;
818 	}
819 	srcu_read_unlock(&ctrl->srcu, srcu_idx);
820 
821 	ret = nvme_user_cmd(ctrl, ns, argp, 0, open_for_write);
822 	nvme_put_ns(ns);
823 	return ret;
824 
825 out_unlock:
826 	srcu_read_unlock(&ctrl->srcu, srcu_idx);
827 	return ret;
828 }
829 
830 long nvme_dev_ioctl(struct file *file, unsigned int cmd,
831 		unsigned long arg)
832 {
833 	bool open_for_write = file->f_mode & FMODE_WRITE;
834 	struct nvme_ctrl *ctrl = file->private_data;
835 	void __user *argp = (void __user *)arg;
836 
837 	switch (cmd) {
838 	case NVME_IOCTL_ADMIN_CMD:
839 		return nvme_user_cmd(ctrl, NULL, argp, 0, open_for_write);
840 	case NVME_IOCTL_ADMIN64_CMD:
841 		return nvme_user_cmd64(ctrl, NULL, argp, 0, open_for_write);
842 	case NVME_IOCTL_IO_CMD:
843 		return nvme_dev_user_cmd(ctrl, argp, open_for_write);
844 	case NVME_IOCTL_RESET:
845 		if (!capable(CAP_SYS_ADMIN))
846 			return -EACCES;
847 		dev_warn(ctrl->device, "resetting controller\n");
848 		return nvme_reset_ctrl_sync(ctrl);
849 	case NVME_IOCTL_SUBSYS_RESET:
850 		if (!capable(CAP_SYS_ADMIN))
851 			return -EACCES;
852 		return nvme_reset_subsystem(ctrl);
853 	case NVME_IOCTL_RESCAN:
854 		if (!capable(CAP_SYS_ADMIN))
855 			return -EACCES;
856 		nvme_queue_scan(ctrl);
857 		return 0;
858 	default:
859 		return -ENOTTY;
860 	}
861 }
862