xref: /linux/drivers/nvme/host/ioctl.c (revision f7bc22ca0d55bdcb59e3a4a028fb811d23e53959)
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 iov_iter *iter, 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 	struct nvme_ctrl *ctrl = nvme_req(req)->ctrl;
124 	bool has_metadata = meta_buffer && meta_len;
125 	struct bio *bio = NULL;
126 	int ret;
127 
128 	if (!nvme_ctrl_sgl_supported(ctrl))
129 		dev_warn_once(ctrl->device, "using unchecked data buffer\n");
130 	if (has_metadata) {
131 		if (!supports_metadata)
132 			return -EINVAL;
133 
134 		if (!nvme_ctrl_meta_sgl_supported(ctrl))
135 			dev_warn_once(ctrl->device,
136 				      "using unchecked metadata buffer\n");
137 	}
138 
139 	if (iter)
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 	if (ret)
146 		return ret;
147 
148 	if (has_metadata) {
149 		ret = blk_rq_integrity_map_user(req, meta_buffer, meta_len);
150 		if (ret)
151 			goto out_unmap;
152 	}
153 
154 	return ret;
155 
156 out_unmap:
157 	if (bio)
158 		blk_rq_unmap_user(bio);
159 	return ret;
160 }
161 
162 static int nvme_submit_user_cmd(struct request_queue *q,
163 		struct nvme_command *cmd, u64 ubuffer, unsigned bufflen,
164 		void __user *meta_buffer, unsigned meta_len,
165 		u64 *result, unsigned timeout, unsigned int flags)
166 {
167 	struct nvme_ns *ns = q->queuedata;
168 	struct nvme_ctrl *ctrl;
169 	struct request *req;
170 	struct bio *bio;
171 	u32 effects;
172 	int ret;
173 
174 	req = nvme_alloc_user_request(q, cmd, 0, 0);
175 	if (IS_ERR(req))
176 		return PTR_ERR(req);
177 
178 	req->timeout = timeout;
179 	if (ubuffer && bufflen) {
180 		ret = nvme_map_user_request(req, ubuffer, bufflen, meta_buffer,
181 				meta_len, NULL, flags);
182 		if (ret)
183 			goto out_free_req;
184 	}
185 
186 	bio = req->bio;
187 	ctrl = nvme_req(req)->ctrl;
188 
189 	effects = nvme_passthru_start(ctrl, ns, cmd->common.opcode);
190 	ret = nvme_execute_rq(req, false);
191 	if (result)
192 		*result = le64_to_cpu(nvme_req(req)->result.u64);
193 	if (bio)
194 		blk_rq_unmap_user(bio);
195 	blk_mq_free_request(req);
196 
197 	if (effects)
198 		nvme_passthru_end(ctrl, ns, effects, cmd, ret);
199 	return ret;
200 
201 out_free_req:
202 	blk_mq_free_request(req);
203 	return ret;
204 }
205 
206 static int nvme_submit_io(struct nvme_ns *ns, struct nvme_user_io __user *uio)
207 {
208 	struct nvme_user_io io;
209 	struct nvme_command c;
210 	unsigned length, meta_len;
211 	void __user *metadata;
212 
213 	if (copy_from_user(&io, uio, sizeof(io)))
214 		return -EFAULT;
215 	if (io.flags)
216 		return -EINVAL;
217 
218 	switch (io.opcode) {
219 	case nvme_cmd_write:
220 	case nvme_cmd_read:
221 	case nvme_cmd_compare:
222 		break;
223 	default:
224 		return -EINVAL;
225 	}
226 
227 	length = (io.nblocks + 1) << ns->head->lba_shift;
228 
229 	if ((io.control & NVME_RW_PRINFO_PRACT) &&
230 	    (ns->head->ms == ns->head->pi_size)) {
231 		/*
232 		 * Protection information is stripped/inserted by the
233 		 * controller.
234 		 */
235 		if (nvme_to_user_ptr(io.metadata))
236 			return -EINVAL;
237 		meta_len = 0;
238 		metadata = NULL;
239 	} else {
240 		meta_len = (io.nblocks + 1) * ns->head->ms;
241 		metadata = nvme_to_user_ptr(io.metadata);
242 	}
243 
244 	if (ns->head->features & NVME_NS_EXT_LBAS) {
245 		length += meta_len;
246 		meta_len = 0;
247 	} else if (meta_len) {
248 		if ((io.metadata & 3) || !io.metadata)
249 			return -EINVAL;
250 	}
251 
252 	memset(&c, 0, sizeof(c));
253 	c.rw.opcode = io.opcode;
254 	c.rw.flags = io.flags;
255 	c.rw.nsid = cpu_to_le32(ns->head->ns_id);
256 	c.rw.slba = cpu_to_le64(io.slba);
257 	c.rw.length = cpu_to_le16(io.nblocks);
258 	c.rw.control = cpu_to_le16(io.control);
259 	c.rw.dsmgmt = cpu_to_le32(io.dsmgmt);
260 	c.rw.reftag = cpu_to_le32(io.reftag);
261 	c.rw.lbat = cpu_to_le16(io.apptag);
262 	c.rw.lbatm = cpu_to_le16(io.appmask);
263 
264 	return nvme_submit_user_cmd(ns->queue, &c, io.addr, length, metadata,
265 			meta_len, NULL, 0, 0);
266 }
267 
268 static bool nvme_validate_passthru_nsid(struct nvme_ctrl *ctrl,
269 					struct nvme_ns *ns, __u32 nsid)
270 {
271 	if (ns && nsid != ns->head->ns_id) {
272 		dev_err(ctrl->device,
273 			"%s: nsid (%u) in cmd does not match nsid (%u) of namespace\n",
274 			current->comm, nsid, ns->head->ns_id);
275 		return false;
276 	}
277 
278 	return true;
279 }
280 
281 static int nvme_user_cmd(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
282 		struct nvme_passthru_cmd __user *ucmd, unsigned int flags,
283 		bool open_for_write)
284 {
285 	struct nvme_passthru_cmd cmd;
286 	struct nvme_command c;
287 	unsigned timeout = 0;
288 	u64 result;
289 	int status;
290 
291 	if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
292 		return -EFAULT;
293 	if (cmd.flags)
294 		return -EINVAL;
295 	if (!nvme_validate_passthru_nsid(ctrl, ns, cmd.nsid))
296 		return -EINVAL;
297 
298 	memset(&c, 0, sizeof(c));
299 	c.common.opcode = cmd.opcode;
300 	c.common.flags = cmd.flags;
301 	c.common.nsid = cpu_to_le32(cmd.nsid);
302 	c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
303 	c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
304 	c.common.cdw10 = cpu_to_le32(cmd.cdw10);
305 	c.common.cdw11 = cpu_to_le32(cmd.cdw11);
306 	c.common.cdw12 = cpu_to_le32(cmd.cdw12);
307 	c.common.cdw13 = cpu_to_le32(cmd.cdw13);
308 	c.common.cdw14 = cpu_to_le32(cmd.cdw14);
309 	c.common.cdw15 = cpu_to_le32(cmd.cdw15);
310 
311 	if (!nvme_cmd_allowed(ns, &c, 0, open_for_write))
312 		return -EACCES;
313 
314 	if (cmd.timeout_ms)
315 		timeout = msecs_to_jiffies(cmd.timeout_ms);
316 
317 	status = nvme_submit_user_cmd(ns ? ns->queue : ctrl->admin_q, &c,
318 			cmd.addr, cmd.data_len, nvme_to_user_ptr(cmd.metadata),
319 			cmd.metadata_len, &result, timeout, 0);
320 
321 	if (status >= 0) {
322 		if (put_user(result, &ucmd->result))
323 			return -EFAULT;
324 	}
325 
326 	return status;
327 }
328 
329 static int nvme_user_cmd64(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
330 		struct nvme_passthru_cmd64 __user *ucmd, unsigned int flags,
331 		bool open_for_write)
332 {
333 	struct nvme_passthru_cmd64 cmd;
334 	struct nvme_command c;
335 	unsigned timeout = 0;
336 	int status;
337 
338 	if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
339 		return -EFAULT;
340 	if (cmd.flags)
341 		return -EINVAL;
342 	if (!nvme_validate_passthru_nsid(ctrl, ns, cmd.nsid))
343 		return -EINVAL;
344 
345 	memset(&c, 0, sizeof(c));
346 	c.common.opcode = cmd.opcode;
347 	c.common.flags = cmd.flags;
348 	c.common.nsid = cpu_to_le32(cmd.nsid);
349 	c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
350 	c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
351 	c.common.cdw10 = cpu_to_le32(cmd.cdw10);
352 	c.common.cdw11 = cpu_to_le32(cmd.cdw11);
353 	c.common.cdw12 = cpu_to_le32(cmd.cdw12);
354 	c.common.cdw13 = cpu_to_le32(cmd.cdw13);
355 	c.common.cdw14 = cpu_to_le32(cmd.cdw14);
356 	c.common.cdw15 = cpu_to_le32(cmd.cdw15);
357 
358 	if (!nvme_cmd_allowed(ns, &c, flags, open_for_write))
359 		return -EACCES;
360 
361 	if (cmd.timeout_ms)
362 		timeout = msecs_to_jiffies(cmd.timeout_ms);
363 
364 	status = nvme_submit_user_cmd(ns ? ns->queue : ctrl->admin_q, &c,
365 			cmd.addr, cmd.data_len, nvme_to_user_ptr(cmd.metadata),
366 			cmd.metadata_len, &cmd.result, timeout, flags);
367 
368 	if (status >= 0) {
369 		if (put_user(cmd.result, &ucmd->result))
370 			return -EFAULT;
371 	}
372 
373 	return status;
374 }
375 
376 struct nvme_uring_data {
377 	__u64	metadata;
378 	__u64	addr;
379 	__u32	data_len;
380 	__u32	metadata_len;
381 	__u32	timeout_ms;
382 };
383 
384 /*
385  * This overlays struct io_uring_cmd pdu.
386  * Expect build errors if this grows larger than that.
387  */
388 struct nvme_uring_cmd_pdu {
389 	struct request *req;
390 	struct bio *bio;
391 	u64 result;
392 	int status;
393 };
394 
395 static inline struct nvme_uring_cmd_pdu *nvme_uring_cmd_pdu(
396 		struct io_uring_cmd *ioucmd)
397 {
398 	return io_uring_cmd_to_pdu(ioucmd, struct nvme_uring_cmd_pdu);
399 }
400 
401 static void nvme_uring_task_cb(struct io_tw_req tw_req, io_tw_token_t tw)
402 {
403 	struct io_uring_cmd *ioucmd = io_uring_cmd_from_tw(tw_req);
404 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
405 
406 	if (pdu->bio)
407 		blk_rq_unmap_user(pdu->bio);
408 	io_uring_cmd_done32(ioucmd, pdu->status, pdu->result,
409 			    IO_URING_CMD_TASK_WORK_ISSUE_FLAGS);
410 }
411 
412 static enum rq_end_io_ret nvme_uring_cmd_end_io(struct request *req,
413 						blk_status_t err,
414 						const struct io_comp_batch *iob)
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 		if (!pdu->status)
424 			pdu->status = blk_status_to_errno(err);
425 	}
426 	pdu->result = le64_to_cpu(nvme_req(req)->result.u64);
427 
428 	/*
429 	 * For IOPOLL, check if this completion is happening in the context
430 	 * of the same io_ring that owns the request (local context). If so,
431 	 * we can complete inline without task_work overhead. Otherwise, we
432 	 * must punt to task_work to ensure completion happens in the correct
433 	 * ring's context.
434 	 */
435 	if (blk_rq_is_poll(req) && iob &&
436 	    iob->poll_ctx == io_uring_cmd_ctx_handle(ioucmd)) {
437 		if (pdu->bio)
438 			blk_rq_unmap_user(pdu->bio);
439 		io_uring_cmd_done32(ioucmd, pdu->status, pdu->result, 0);
440 	} else {
441 		io_uring_cmd_do_in_task_lazy(ioucmd, nvme_uring_task_cb);
442 	}
443 	return RQ_END_IO_FREE;
444 }
445 
446 static int nvme_uring_cmd_io(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
447 		struct io_uring_cmd *ioucmd, unsigned int issue_flags, bool vec)
448 {
449 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
450 	const struct nvme_uring_cmd *cmd = io_uring_sqe_cmd(ioucmd->sqe);
451 	struct request_queue *q = ns ? ns->queue : ctrl->admin_q;
452 	struct nvme_uring_data d;
453 	struct nvme_command c;
454 	struct iov_iter iter;
455 	struct iov_iter *map_iter = NULL;
456 	struct request *req;
457 	blk_opf_t rq_flags = 0;
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 (d.data_len && (ioucmd->flags & IORING_URING_CMD_FIXED)) {
492 		int ddir = nvme_is_write(&c) ? WRITE : READ;
493 
494 		if (vec)
495 			ret = io_uring_cmd_import_fixed_vec(ioucmd,
496 					u64_to_user_ptr(d.addr), d.data_len,
497 					ddir, &iter, issue_flags);
498 		else
499 			ret = io_uring_cmd_import_fixed(d.addr, d.data_len,
500 					ddir, &iter, ioucmd, issue_flags);
501 		if (ret < 0)
502 			return ret;
503 
504 		map_iter = &iter;
505 	}
506 
507 	if (issue_flags & IO_URING_F_NONBLOCK) {
508 		rq_flags |= REQ_NOWAIT;
509 		blk_flags = BLK_MQ_REQ_NOWAIT;
510 	}
511 	if (issue_flags & IO_URING_F_IOPOLL)
512 		rq_flags |= REQ_POLLED;
513 
514 	req = nvme_alloc_user_request(q, &c, rq_flags, blk_flags);
515 	if (IS_ERR(req))
516 		return PTR_ERR(req);
517 	req->timeout = d.timeout_ms ? msecs_to_jiffies(d.timeout_ms) : 0;
518 
519 	if (d.data_len) {
520 		ret = nvme_map_user_request(req, d.addr, d.data_len,
521 			nvme_to_user_ptr(d.metadata), d.metadata_len,
522 			map_iter, vec ? NVME_IOCTL_VEC : 0);
523 		if (ret)
524 			goto out_free_req;
525 	}
526 
527 	/* to free bio on completion, as req->bio will be null at that time */
528 	pdu->bio = req->bio;
529 	pdu->req = req;
530 	req->end_io_data = ioucmd;
531 	req->end_io = nvme_uring_cmd_end_io;
532 	blk_execute_rq_nowait(req, false);
533 	return -EIOCBQUEUED;
534 
535 out_free_req:
536 	blk_mq_free_request(req);
537 	return ret;
538 }
539 
540 static bool is_ctrl_ioctl(unsigned int cmd)
541 {
542 	if (cmd == NVME_IOCTL_ADMIN_CMD || cmd == NVME_IOCTL_ADMIN64_CMD)
543 		return true;
544 	if (is_sed_ioctl(cmd))
545 		return true;
546 	return false;
547 }
548 
549 static int nvme_ctrl_ioctl(struct nvme_ctrl *ctrl, unsigned int cmd,
550 		void __user *argp, bool open_for_write)
551 {
552 	switch (cmd) {
553 	case NVME_IOCTL_ADMIN_CMD:
554 		return nvme_user_cmd(ctrl, NULL, argp, 0, open_for_write);
555 	case NVME_IOCTL_ADMIN64_CMD:
556 		return nvme_user_cmd64(ctrl, NULL, argp, 0, open_for_write);
557 	default:
558 		return sed_ioctl(ctrl->opal_dev, cmd, argp);
559 	}
560 }
561 
562 #ifdef COMPAT_FOR_U64_ALIGNMENT
563 struct nvme_user_io32 {
564 	__u8	opcode;
565 	__u8	flags;
566 	__u16	control;
567 	__u16	nblocks;
568 	__u16	rsvd;
569 	__u64	metadata;
570 	__u64	addr;
571 	__u64	slba;
572 	__u32	dsmgmt;
573 	__u32	reftag;
574 	__u16	apptag;
575 	__u16	appmask;
576 } __attribute__((__packed__));
577 #define NVME_IOCTL_SUBMIT_IO32	_IOW('N', 0x42, struct nvme_user_io32)
578 #endif /* COMPAT_FOR_U64_ALIGNMENT */
579 
580 static int nvme_ns_ioctl(struct nvme_ns *ns, unsigned int cmd,
581 		void __user *argp, unsigned int flags, bool open_for_write)
582 {
583 	switch (cmd) {
584 	case NVME_IOCTL_ID:
585 		force_successful_syscall_return();
586 		return ns->head->ns_id;
587 	case NVME_IOCTL_IO_CMD:
588 		return nvme_user_cmd(ns->ctrl, ns, argp, flags, open_for_write);
589 	/*
590 	 * struct nvme_user_io can have different padding on some 32-bit ABIs.
591 	 * Just accept the compat version as all fields that are used are the
592 	 * same size and at the same offset.
593 	 */
594 #ifdef COMPAT_FOR_U64_ALIGNMENT
595 	case NVME_IOCTL_SUBMIT_IO32:
596 #endif
597 	case NVME_IOCTL_SUBMIT_IO:
598 		return nvme_submit_io(ns, argp);
599 	case NVME_IOCTL_IO64_CMD_VEC:
600 		flags |= NVME_IOCTL_VEC;
601 		fallthrough;
602 	case NVME_IOCTL_IO64_CMD:
603 		return nvme_user_cmd64(ns->ctrl, ns, argp, flags,
604 				       open_for_write);
605 	default:
606 		return -ENOTTY;
607 	}
608 }
609 
610 int nvme_ioctl(struct block_device *bdev, blk_mode_t mode,
611 		unsigned int cmd, unsigned long arg)
612 {
613 	struct nvme_ns *ns = bdev->bd_disk->private_data;
614 	bool open_for_write = mode & BLK_OPEN_WRITE;
615 	void __user *argp = (void __user *)arg;
616 	unsigned int flags = 0;
617 
618 	if (bdev_is_partition(bdev))
619 		flags |= NVME_IOCTL_PARTITION;
620 
621 	if (is_ctrl_ioctl(cmd))
622 		return nvme_ctrl_ioctl(ns->ctrl, cmd, argp, open_for_write);
623 	return nvme_ns_ioctl(ns, cmd, argp, flags, open_for_write);
624 }
625 
626 long nvme_ns_chr_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
627 {
628 	struct nvme_ns *ns =
629 		container_of(file_inode(file)->i_cdev, struct nvme_ns, cdev);
630 	bool open_for_write = file->f_mode & FMODE_WRITE;
631 	void __user *argp = (void __user *)arg;
632 
633 	if (is_ctrl_ioctl(cmd))
634 		return nvme_ctrl_ioctl(ns->ctrl, cmd, argp, open_for_write);
635 	return nvme_ns_ioctl(ns, cmd, argp, 0, open_for_write);
636 }
637 
638 static int nvme_uring_cmd_checks(unsigned int issue_flags)
639 {
640 
641 	/* NVMe passthrough requires big SQE/CQE support */
642 	if ((issue_flags & (IO_URING_F_SQE128|IO_URING_F_CQE32)) !=
643 	    (IO_URING_F_SQE128|IO_URING_F_CQE32))
644 		return -EOPNOTSUPP;
645 	return 0;
646 }
647 
648 static int nvme_ns_uring_cmd(struct nvme_ns *ns, struct io_uring_cmd *ioucmd,
649 			     unsigned int issue_flags)
650 {
651 	struct nvme_ctrl *ctrl = ns->ctrl;
652 	int ret;
653 
654 	ret = nvme_uring_cmd_checks(issue_flags);
655 	if (ret)
656 		return ret;
657 
658 	switch (ioucmd->cmd_op) {
659 	case NVME_URING_CMD_IO:
660 		ret = nvme_uring_cmd_io(ctrl, ns, ioucmd, issue_flags, false);
661 		break;
662 	case NVME_URING_CMD_IO_VEC:
663 		ret = nvme_uring_cmd_io(ctrl, ns, ioucmd, issue_flags, true);
664 		break;
665 	default:
666 		ret = -ENOTTY;
667 	}
668 
669 	return ret;
670 }
671 
672 int nvme_ns_chr_uring_cmd(struct io_uring_cmd *ioucmd, unsigned int issue_flags)
673 {
674 	struct nvme_ns *ns = container_of(file_inode(ioucmd->file)->i_cdev,
675 			struct nvme_ns, cdev);
676 
677 	return nvme_ns_uring_cmd(ns, ioucmd, issue_flags);
678 }
679 
680 int nvme_ns_chr_uring_cmd_iopoll(struct io_uring_cmd *ioucmd,
681 				 struct io_comp_batch *iob,
682 				 unsigned int poll_flags)
683 {
684 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
685 	struct request *req = pdu->req;
686 
687 	if (req && blk_rq_is_poll(req))
688 		return blk_rq_poll(req, iob, poll_flags);
689 	return 0;
690 }
691 #ifdef CONFIG_NVME_MULTIPATH
692 static int nvme_ns_head_ctrl_ioctl(struct nvme_ns *ns, unsigned int cmd,
693 		void __user *argp, struct nvme_ns_head *head, int srcu_idx,
694 		bool open_for_write)
695 	__releases(&head->srcu)
696 {
697 	struct nvme_ctrl *ctrl = ns->ctrl;
698 	int ret;
699 
700 	nvme_get_ctrl(ns->ctrl);
701 	srcu_read_unlock(&head->srcu, srcu_idx);
702 	ret = nvme_ctrl_ioctl(ns->ctrl, cmd, argp, open_for_write);
703 
704 	nvme_put_ctrl(ctrl);
705 	return ret;
706 }
707 
708 int nvme_ns_head_ioctl(struct block_device *bdev, blk_mode_t mode,
709 		unsigned int cmd, unsigned long arg)
710 {
711 	struct nvme_ns_head *head = bdev->bd_disk->private_data;
712 	bool open_for_write = mode & BLK_OPEN_WRITE;
713 	void __user *argp = (void __user *)arg;
714 	struct nvme_ns *ns;
715 	int srcu_idx, ret = -EWOULDBLOCK;
716 	unsigned int flags = 0;
717 
718 	if (bdev_is_partition(bdev))
719 		flags |= NVME_IOCTL_PARTITION;
720 
721 	srcu_idx = srcu_read_lock(&head->srcu);
722 	ns = nvme_find_path(head);
723 	if (!ns)
724 		goto out_unlock;
725 
726 	/*
727 	 * Handle ioctls that apply to the controller instead of the namespace
728 	 * separately and drop the ns SRCU reference early.  This avoids a
729 	 * deadlock when deleting namespaces using the passthrough interface.
730 	 */
731 	if (is_ctrl_ioctl(cmd))
732 		return nvme_ns_head_ctrl_ioctl(ns, cmd, argp, head, srcu_idx,
733 					       open_for_write);
734 
735 	ret = nvme_ns_ioctl(ns, cmd, argp, flags, open_for_write);
736 out_unlock:
737 	srcu_read_unlock(&head->srcu, srcu_idx);
738 	return ret;
739 }
740 
741 long nvme_ns_head_chr_ioctl(struct file *file, unsigned int cmd,
742 		unsigned long arg)
743 {
744 	bool open_for_write = file->f_mode & FMODE_WRITE;
745 	struct cdev *cdev = file_inode(file)->i_cdev;
746 	struct nvme_ns_head *head =
747 		container_of(cdev, struct nvme_ns_head, cdev);
748 	void __user *argp = (void __user *)arg;
749 	struct nvme_ns *ns;
750 	int srcu_idx, ret = -EWOULDBLOCK;
751 
752 	srcu_idx = srcu_read_lock(&head->srcu);
753 	ns = nvme_find_path(head);
754 	if (!ns)
755 		goto out_unlock;
756 
757 	if (is_ctrl_ioctl(cmd))
758 		return nvme_ns_head_ctrl_ioctl(ns, cmd, argp, head, srcu_idx,
759 				open_for_write);
760 
761 	ret = nvme_ns_ioctl(ns, cmd, argp, 0, open_for_write);
762 out_unlock:
763 	srcu_read_unlock(&head->srcu, srcu_idx);
764 	return ret;
765 }
766 
767 int nvme_ns_head_chr_uring_cmd(struct io_uring_cmd *ioucmd,
768 		unsigned int issue_flags)
769 {
770 	struct cdev *cdev = file_inode(ioucmd->file)->i_cdev;
771 	struct nvme_ns_head *head = container_of(cdev, struct nvme_ns_head, cdev);
772 	int srcu_idx = srcu_read_lock(&head->srcu);
773 	struct nvme_ns *ns = nvme_find_path(head);
774 	int ret = -EINVAL;
775 
776 	if (ns)
777 		ret = nvme_ns_uring_cmd(ns, ioucmd, issue_flags);
778 	srcu_read_unlock(&head->srcu, srcu_idx);
779 	return ret;
780 }
781 #endif /* CONFIG_NVME_MULTIPATH */
782 
783 int nvme_dev_uring_cmd(struct io_uring_cmd *ioucmd, unsigned int issue_flags)
784 {
785 	struct nvme_ctrl *ctrl = ioucmd->file->private_data;
786 	int ret;
787 
788 	/* IOPOLL not supported yet */
789 	if (issue_flags & IO_URING_F_IOPOLL)
790 		return -EOPNOTSUPP;
791 
792 	ret = nvme_uring_cmd_checks(issue_flags);
793 	if (ret)
794 		return ret;
795 
796 	switch (ioucmd->cmd_op) {
797 	case NVME_URING_CMD_ADMIN:
798 		ret = nvme_uring_cmd_io(ctrl, NULL, ioucmd, issue_flags, false);
799 		break;
800 	case NVME_URING_CMD_ADMIN_VEC:
801 		ret = nvme_uring_cmd_io(ctrl, NULL, ioucmd, issue_flags, true);
802 		break;
803 	default:
804 		ret = -ENOTTY;
805 	}
806 
807 	return ret;
808 }
809 
810 static int nvme_dev_user_cmd(struct nvme_ctrl *ctrl, void __user *argp,
811 		bool open_for_write)
812 {
813 	struct nvme_ns *ns;
814 	int ret, srcu_idx;
815 
816 	srcu_idx = srcu_read_lock(&ctrl->srcu);
817 	if (list_empty(&ctrl->namespaces)) {
818 		ret = -ENOTTY;
819 		goto out_unlock;
820 	}
821 
822 	ns = list_first_or_null_rcu(&ctrl->namespaces, struct nvme_ns, list);
823 	if (ns != list_last_entry(&ctrl->namespaces, struct nvme_ns, list)) {
824 		dev_warn(ctrl->device,
825 			"NVME_IOCTL_IO_CMD not supported when multiple namespaces present!\n");
826 		ret = -EINVAL;
827 		goto out_unlock;
828 	}
829 
830 	dev_warn(ctrl->device,
831 		"using deprecated NVME_IOCTL_IO_CMD ioctl on the char device!\n");
832 	if (!nvme_get_ns(ns)) {
833 		ret = -ENXIO;
834 		goto out_unlock;
835 	}
836 	srcu_read_unlock(&ctrl->srcu, srcu_idx);
837 
838 	ret = nvme_user_cmd(ctrl, ns, argp, 0, open_for_write);
839 	nvme_put_ns(ns);
840 	return ret;
841 
842 out_unlock:
843 	srcu_read_unlock(&ctrl->srcu, srcu_idx);
844 	return ret;
845 }
846 
847 long nvme_dev_ioctl(struct file *file, unsigned int cmd,
848 		unsigned long arg)
849 {
850 	bool open_for_write = file->f_mode & FMODE_WRITE;
851 	struct nvme_ctrl *ctrl = file->private_data;
852 	void __user *argp = (void __user *)arg;
853 
854 	switch (cmd) {
855 	case NVME_IOCTL_ADMIN_CMD:
856 		return nvme_user_cmd(ctrl, NULL, argp, 0, open_for_write);
857 	case NVME_IOCTL_ADMIN64_CMD:
858 		return nvme_user_cmd64(ctrl, NULL, argp, 0, open_for_write);
859 	case NVME_IOCTL_IO_CMD:
860 		return nvme_dev_user_cmd(ctrl, argp, open_for_write);
861 	case NVME_IOCTL_RESET:
862 		if (!capable(CAP_SYS_ADMIN))
863 			return -EACCES;
864 		dev_warn(ctrl->device, "resetting controller\n");
865 		return nvme_reset_ctrl_sync(ctrl);
866 	case NVME_IOCTL_SUBSYS_RESET:
867 		if (!capable(CAP_SYS_ADMIN))
868 			return -EACCES;
869 		return nvme_reset_subsystem(ctrl);
870 	case NVME_IOCTL_RESCAN:
871 		if (!capable(CAP_SYS_ADMIN))
872 			return -EACCES;
873 		nvme_queue_scan(ctrl);
874 		return 0;
875 	default:
876 		return -ENOTTY;
877 	}
878 }
879