xref: /linux/drivers/nvme/host/ioctl.c (revision 7f71507851fc7764b36a3221839607d3a45c2025)
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 	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 		if (!nvme_ctrl_meta_sgl_supported(ctrl))
134 			dev_warn_once(ctrl->device,
135 				      "using unchecked metadata buffer\n");
136 	}
137 
138 	if (ioucmd && (ioucmd->flags & IORING_URING_CMD_FIXED)) {
139 		struct iov_iter iter;
140 
141 		/* fixedbufs is only for non-vectored io */
142 		if (WARN_ON_ONCE(flags & NVME_IOCTL_VEC))
143 			return -EINVAL;
144 		ret = io_uring_cmd_import_fixed(ubuffer, bufflen,
145 				rq_data_dir(req), &iter, ioucmd);
146 		if (ret < 0)
147 			goto out;
148 		ret = blk_rq_map_user_iov(q, req, NULL, &iter, GFP_KERNEL);
149 	} else {
150 		ret = blk_rq_map_user_io(req, NULL, nvme_to_user_ptr(ubuffer),
151 				bufflen, GFP_KERNEL, flags & NVME_IOCTL_VEC, 0,
152 				0, rq_data_dir(req));
153 	}
154 
155 	if (ret)
156 		goto out;
157 
158 	bio = req->bio;
159 	if (bdev)
160 		bio_set_dev(bio, bdev);
161 
162 	if (has_metadata) {
163 		ret = blk_rq_integrity_map_user(req, meta_buffer, meta_len);
164 		if (ret)
165 			goto out_unmap;
166 	}
167 
168 	return ret;
169 
170 out_unmap:
171 	if (bio)
172 		blk_rq_unmap_user(bio);
173 out:
174 	blk_mq_free_request(req);
175 	return ret;
176 }
177 
178 static int nvme_submit_user_cmd(struct request_queue *q,
179 		struct nvme_command *cmd, u64 ubuffer, unsigned bufflen,
180 		void __user *meta_buffer, unsigned meta_len,
181 		u64 *result, unsigned timeout, unsigned int flags)
182 {
183 	struct nvme_ns *ns = q->queuedata;
184 	struct nvme_ctrl *ctrl;
185 	struct request *req;
186 	struct bio *bio;
187 	u32 effects;
188 	int ret;
189 
190 	req = nvme_alloc_user_request(q, cmd, 0, 0);
191 	if (IS_ERR(req))
192 		return PTR_ERR(req);
193 
194 	req->timeout = timeout;
195 	if (ubuffer && bufflen) {
196 		ret = nvme_map_user_request(req, ubuffer, bufflen, meta_buffer,
197 				meta_len, NULL, flags);
198 		if (ret)
199 			return ret;
200 	}
201 
202 	bio = req->bio;
203 	ctrl = nvme_req(req)->ctrl;
204 
205 	effects = nvme_passthru_start(ctrl, ns, cmd->common.opcode);
206 	ret = nvme_execute_rq(req, false);
207 	if (result)
208 		*result = le64_to_cpu(nvme_req(req)->result.u64);
209 	if (bio)
210 		blk_rq_unmap_user(bio);
211 	blk_mq_free_request(req);
212 
213 	if (effects)
214 		nvme_passthru_end(ctrl, ns, effects, cmd, ret);
215 
216 	return ret;
217 }
218 
219 static int nvme_submit_io(struct nvme_ns *ns, struct nvme_user_io __user *uio)
220 {
221 	struct nvme_user_io io;
222 	struct nvme_command c;
223 	unsigned length, meta_len;
224 	void __user *metadata;
225 
226 	if (copy_from_user(&io, uio, sizeof(io)))
227 		return -EFAULT;
228 	if (io.flags)
229 		return -EINVAL;
230 
231 	switch (io.opcode) {
232 	case nvme_cmd_write:
233 	case nvme_cmd_read:
234 	case nvme_cmd_compare:
235 		break;
236 	default:
237 		return -EINVAL;
238 	}
239 
240 	length = (io.nblocks + 1) << ns->head->lba_shift;
241 
242 	if ((io.control & NVME_RW_PRINFO_PRACT) &&
243 	    (ns->head->ms == ns->head->pi_size)) {
244 		/*
245 		 * Protection information is stripped/inserted by the
246 		 * controller.
247 		 */
248 		if (nvme_to_user_ptr(io.metadata))
249 			return -EINVAL;
250 		meta_len = 0;
251 		metadata = NULL;
252 	} else {
253 		meta_len = (io.nblocks + 1) * ns->head->ms;
254 		metadata = nvme_to_user_ptr(io.metadata);
255 	}
256 
257 	if (ns->head->features & NVME_NS_EXT_LBAS) {
258 		length += meta_len;
259 		meta_len = 0;
260 	} else if (meta_len) {
261 		if ((io.metadata & 3) || !io.metadata)
262 			return -EINVAL;
263 	}
264 
265 	memset(&c, 0, sizeof(c));
266 	c.rw.opcode = io.opcode;
267 	c.rw.flags = io.flags;
268 	c.rw.nsid = cpu_to_le32(ns->head->ns_id);
269 	c.rw.slba = cpu_to_le64(io.slba);
270 	c.rw.length = cpu_to_le16(io.nblocks);
271 	c.rw.control = cpu_to_le16(io.control);
272 	c.rw.dsmgmt = cpu_to_le32(io.dsmgmt);
273 	c.rw.reftag = cpu_to_le32(io.reftag);
274 	c.rw.lbat = cpu_to_le16(io.apptag);
275 	c.rw.lbatm = cpu_to_le16(io.appmask);
276 
277 	return nvme_submit_user_cmd(ns->queue, &c, io.addr, length, metadata,
278 			meta_len, NULL, 0, 0);
279 }
280 
281 static bool nvme_validate_passthru_nsid(struct nvme_ctrl *ctrl,
282 					struct nvme_ns *ns, __u32 nsid)
283 {
284 	if (ns && nsid != ns->head->ns_id) {
285 		dev_err(ctrl->device,
286 			"%s: nsid (%u) in cmd does not match nsid (%u)"
287 			"of namespace\n",
288 			current->comm, nsid, ns->head->ns_id);
289 		return false;
290 	}
291 
292 	return true;
293 }
294 
295 static int nvme_user_cmd(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
296 		struct nvme_passthru_cmd __user *ucmd, unsigned int flags,
297 		bool open_for_write)
298 {
299 	struct nvme_passthru_cmd cmd;
300 	struct nvme_command c;
301 	unsigned timeout = 0;
302 	u64 result;
303 	int status;
304 
305 	if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
306 		return -EFAULT;
307 	if (cmd.flags)
308 		return -EINVAL;
309 	if (!nvme_validate_passthru_nsid(ctrl, ns, cmd.nsid))
310 		return -EINVAL;
311 
312 	memset(&c, 0, sizeof(c));
313 	c.common.opcode = cmd.opcode;
314 	c.common.flags = cmd.flags;
315 	c.common.nsid = cpu_to_le32(cmd.nsid);
316 	c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
317 	c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
318 	c.common.cdw10 = cpu_to_le32(cmd.cdw10);
319 	c.common.cdw11 = cpu_to_le32(cmd.cdw11);
320 	c.common.cdw12 = cpu_to_le32(cmd.cdw12);
321 	c.common.cdw13 = cpu_to_le32(cmd.cdw13);
322 	c.common.cdw14 = cpu_to_le32(cmd.cdw14);
323 	c.common.cdw15 = cpu_to_le32(cmd.cdw15);
324 
325 	if (!nvme_cmd_allowed(ns, &c, 0, open_for_write))
326 		return -EACCES;
327 
328 	if (cmd.timeout_ms)
329 		timeout = msecs_to_jiffies(cmd.timeout_ms);
330 
331 	status = nvme_submit_user_cmd(ns ? ns->queue : ctrl->admin_q, &c,
332 			cmd.addr, cmd.data_len, nvme_to_user_ptr(cmd.metadata),
333 			cmd.metadata_len, &result, timeout, 0);
334 
335 	if (status >= 0) {
336 		if (put_user(result, &ucmd->result))
337 			return -EFAULT;
338 	}
339 
340 	return status;
341 }
342 
343 static int nvme_user_cmd64(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
344 		struct nvme_passthru_cmd64 __user *ucmd, unsigned int flags,
345 		bool open_for_write)
346 {
347 	struct nvme_passthru_cmd64 cmd;
348 	struct nvme_command c;
349 	unsigned timeout = 0;
350 	int status;
351 
352 	if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
353 		return -EFAULT;
354 	if (cmd.flags)
355 		return -EINVAL;
356 	if (!nvme_validate_passthru_nsid(ctrl, ns, cmd.nsid))
357 		return -EINVAL;
358 
359 	memset(&c, 0, sizeof(c));
360 	c.common.opcode = cmd.opcode;
361 	c.common.flags = cmd.flags;
362 	c.common.nsid = cpu_to_le32(cmd.nsid);
363 	c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
364 	c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
365 	c.common.cdw10 = cpu_to_le32(cmd.cdw10);
366 	c.common.cdw11 = cpu_to_le32(cmd.cdw11);
367 	c.common.cdw12 = cpu_to_le32(cmd.cdw12);
368 	c.common.cdw13 = cpu_to_le32(cmd.cdw13);
369 	c.common.cdw14 = cpu_to_le32(cmd.cdw14);
370 	c.common.cdw15 = cpu_to_le32(cmd.cdw15);
371 
372 	if (!nvme_cmd_allowed(ns, &c, flags, open_for_write))
373 		return -EACCES;
374 
375 	if (cmd.timeout_ms)
376 		timeout = msecs_to_jiffies(cmd.timeout_ms);
377 
378 	status = nvme_submit_user_cmd(ns ? ns->queue : ctrl->admin_q, &c,
379 			cmd.addr, cmd.data_len, nvme_to_user_ptr(cmd.metadata),
380 			cmd.metadata_len, &cmd.result, timeout, flags);
381 
382 	if (status >= 0) {
383 		if (put_user(cmd.result, &ucmd->result))
384 			return -EFAULT;
385 	}
386 
387 	return status;
388 }
389 
390 struct nvme_uring_data {
391 	__u64	metadata;
392 	__u64	addr;
393 	__u32	data_len;
394 	__u32	metadata_len;
395 	__u32	timeout_ms;
396 };
397 
398 /*
399  * This overlays struct io_uring_cmd pdu.
400  * Expect build errors if this grows larger than that.
401  */
402 struct nvme_uring_cmd_pdu {
403 	struct request *req;
404 	struct bio *bio;
405 	u64 result;
406 	int status;
407 };
408 
409 static inline struct nvme_uring_cmd_pdu *nvme_uring_cmd_pdu(
410 		struct io_uring_cmd *ioucmd)
411 {
412 	return io_uring_cmd_to_pdu(ioucmd, struct nvme_uring_cmd_pdu);
413 }
414 
415 static void nvme_uring_task_cb(struct io_uring_cmd *ioucmd,
416 			       unsigned issue_flags)
417 {
418 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
419 
420 	if (pdu->bio)
421 		blk_rq_unmap_user(pdu->bio);
422 	io_uring_cmd_done(ioucmd, pdu->status, pdu->result, issue_flags);
423 }
424 
425 static enum rq_end_io_ret nvme_uring_cmd_end_io(struct request *req,
426 						blk_status_t err)
427 {
428 	struct io_uring_cmd *ioucmd = req->end_io_data;
429 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
430 
431 	if (nvme_req(req)->flags & NVME_REQ_CANCELLED) {
432 		pdu->status = -EINTR;
433 	} else {
434 		pdu->status = nvme_req(req)->status;
435 		if (!pdu->status)
436 			pdu->status = blk_status_to_errno(err);
437 	}
438 	pdu->result = le64_to_cpu(nvme_req(req)->result.u64);
439 
440 	/*
441 	 * For iopoll, complete it directly. Note that using the uring_cmd
442 	 * helper for this is safe only because we check blk_rq_is_poll().
443 	 * As that returns false if we're NOT on a polled queue, then it's
444 	 * safe to use the polled completion helper.
445 	 *
446 	 * Otherwise, move the completion to task work.
447 	 */
448 	if (blk_rq_is_poll(req)) {
449 		if (pdu->bio)
450 			blk_rq_unmap_user(pdu->bio);
451 		io_uring_cmd_iopoll_done(ioucmd, pdu->result, pdu->status);
452 	} else {
453 		io_uring_cmd_do_in_task_lazy(ioucmd, nvme_uring_task_cb);
454 	}
455 
456 	return RQ_END_IO_FREE;
457 }
458 
459 static int nvme_uring_cmd_io(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
460 		struct io_uring_cmd *ioucmd, unsigned int issue_flags, bool vec)
461 {
462 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
463 	const struct nvme_uring_cmd *cmd = io_uring_sqe_cmd(ioucmd->sqe);
464 	struct request_queue *q = ns ? ns->queue : ctrl->admin_q;
465 	struct nvme_uring_data d;
466 	struct nvme_command c;
467 	struct request *req;
468 	blk_opf_t rq_flags = REQ_ALLOC_CACHE;
469 	blk_mq_req_flags_t blk_flags = 0;
470 	int ret;
471 
472 	c.common.opcode = READ_ONCE(cmd->opcode);
473 	c.common.flags = READ_ONCE(cmd->flags);
474 	if (c.common.flags)
475 		return -EINVAL;
476 
477 	c.common.command_id = 0;
478 	c.common.nsid = cpu_to_le32(cmd->nsid);
479 	if (!nvme_validate_passthru_nsid(ctrl, ns, le32_to_cpu(c.common.nsid)))
480 		return -EINVAL;
481 
482 	c.common.cdw2[0] = cpu_to_le32(READ_ONCE(cmd->cdw2));
483 	c.common.cdw2[1] = cpu_to_le32(READ_ONCE(cmd->cdw3));
484 	c.common.metadata = 0;
485 	c.common.dptr.prp1 = c.common.dptr.prp2 = 0;
486 	c.common.cdw10 = cpu_to_le32(READ_ONCE(cmd->cdw10));
487 	c.common.cdw11 = cpu_to_le32(READ_ONCE(cmd->cdw11));
488 	c.common.cdw12 = cpu_to_le32(READ_ONCE(cmd->cdw12));
489 	c.common.cdw13 = cpu_to_le32(READ_ONCE(cmd->cdw13));
490 	c.common.cdw14 = cpu_to_le32(READ_ONCE(cmd->cdw14));
491 	c.common.cdw15 = cpu_to_le32(READ_ONCE(cmd->cdw15));
492 
493 	if (!nvme_cmd_allowed(ns, &c, 0, ioucmd->file->f_mode & FMODE_WRITE))
494 		return -EACCES;
495 
496 	d.metadata = READ_ONCE(cmd->metadata);
497 	d.addr = READ_ONCE(cmd->addr);
498 	d.data_len = READ_ONCE(cmd->data_len);
499 	d.metadata_len = READ_ONCE(cmd->metadata_len);
500 	d.timeout_ms = READ_ONCE(cmd->timeout_ms);
501 
502 	if (issue_flags & IO_URING_F_NONBLOCK) {
503 		rq_flags |= REQ_NOWAIT;
504 		blk_flags = BLK_MQ_REQ_NOWAIT;
505 	}
506 	if (issue_flags & IO_URING_F_IOPOLL)
507 		rq_flags |= REQ_POLLED;
508 
509 	req = nvme_alloc_user_request(q, &c, rq_flags, blk_flags);
510 	if (IS_ERR(req))
511 		return PTR_ERR(req);
512 	req->timeout = d.timeout_ms ? msecs_to_jiffies(d.timeout_ms) : 0;
513 
514 	if (d.addr && d.data_len) {
515 		ret = nvme_map_user_request(req, d.addr,
516 			d.data_len, nvme_to_user_ptr(d.metadata),
517 			d.metadata_len, ioucmd, vec);
518 		if (ret)
519 			return ret;
520 	}
521 
522 	/* to free bio on completion, as req->bio will be null at that time */
523 	pdu->bio = req->bio;
524 	pdu->req = req;
525 	req->end_io_data = ioucmd;
526 	req->end_io = nvme_uring_cmd_end_io;
527 	blk_execute_rq_nowait(req, false);
528 	return -EIOCBQUEUED;
529 }
530 
531 static bool is_ctrl_ioctl(unsigned int cmd)
532 {
533 	if (cmd == NVME_IOCTL_ADMIN_CMD || cmd == NVME_IOCTL_ADMIN64_CMD)
534 		return true;
535 	if (is_sed_ioctl(cmd))
536 		return true;
537 	return false;
538 }
539 
540 static int nvme_ctrl_ioctl(struct nvme_ctrl *ctrl, unsigned int cmd,
541 		void __user *argp, bool open_for_write)
542 {
543 	switch (cmd) {
544 	case NVME_IOCTL_ADMIN_CMD:
545 		return nvme_user_cmd(ctrl, NULL, argp, 0, open_for_write);
546 	case NVME_IOCTL_ADMIN64_CMD:
547 		return nvme_user_cmd64(ctrl, NULL, argp, 0, open_for_write);
548 	default:
549 		return sed_ioctl(ctrl->opal_dev, cmd, argp);
550 	}
551 }
552 
553 #ifdef COMPAT_FOR_U64_ALIGNMENT
554 struct nvme_user_io32 {
555 	__u8	opcode;
556 	__u8	flags;
557 	__u16	control;
558 	__u16	nblocks;
559 	__u16	rsvd;
560 	__u64	metadata;
561 	__u64	addr;
562 	__u64	slba;
563 	__u32	dsmgmt;
564 	__u32	reftag;
565 	__u16	apptag;
566 	__u16	appmask;
567 } __attribute__((__packed__));
568 #define NVME_IOCTL_SUBMIT_IO32	_IOW('N', 0x42, struct nvme_user_io32)
569 #endif /* COMPAT_FOR_U64_ALIGNMENT */
570 
571 static int nvme_ns_ioctl(struct nvme_ns *ns, unsigned int cmd,
572 		void __user *argp, unsigned int flags, bool open_for_write)
573 {
574 	switch (cmd) {
575 	case NVME_IOCTL_ID:
576 		force_successful_syscall_return();
577 		return ns->head->ns_id;
578 	case NVME_IOCTL_IO_CMD:
579 		return nvme_user_cmd(ns->ctrl, ns, argp, flags, open_for_write);
580 	/*
581 	 * struct nvme_user_io can have different padding on some 32-bit ABIs.
582 	 * Just accept the compat version as all fields that are used are the
583 	 * same size and at the same offset.
584 	 */
585 #ifdef COMPAT_FOR_U64_ALIGNMENT
586 	case NVME_IOCTL_SUBMIT_IO32:
587 #endif
588 	case NVME_IOCTL_SUBMIT_IO:
589 		return nvme_submit_io(ns, argp);
590 	case NVME_IOCTL_IO64_CMD_VEC:
591 		flags |= NVME_IOCTL_VEC;
592 		fallthrough;
593 	case NVME_IOCTL_IO64_CMD:
594 		return nvme_user_cmd64(ns->ctrl, ns, argp, flags,
595 				       open_for_write);
596 	default:
597 		return -ENOTTY;
598 	}
599 }
600 
601 int nvme_ioctl(struct block_device *bdev, blk_mode_t mode,
602 		unsigned int cmd, unsigned long arg)
603 {
604 	struct nvme_ns *ns = bdev->bd_disk->private_data;
605 	bool open_for_write = mode & BLK_OPEN_WRITE;
606 	void __user *argp = (void __user *)arg;
607 	unsigned int flags = 0;
608 
609 	if (bdev_is_partition(bdev))
610 		flags |= NVME_IOCTL_PARTITION;
611 
612 	if (is_ctrl_ioctl(cmd))
613 		return nvme_ctrl_ioctl(ns->ctrl, cmd, argp, open_for_write);
614 	return nvme_ns_ioctl(ns, cmd, argp, flags, open_for_write);
615 }
616 
617 long nvme_ns_chr_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
618 {
619 	struct nvme_ns *ns =
620 		container_of(file_inode(file)->i_cdev, struct nvme_ns, cdev);
621 	bool open_for_write = file->f_mode & FMODE_WRITE;
622 	void __user *argp = (void __user *)arg;
623 
624 	if (is_ctrl_ioctl(cmd))
625 		return nvme_ctrl_ioctl(ns->ctrl, cmd, argp, open_for_write);
626 	return nvme_ns_ioctl(ns, cmd, argp, 0, open_for_write);
627 }
628 
629 static int nvme_uring_cmd_checks(unsigned int issue_flags)
630 {
631 
632 	/* NVMe passthrough requires big SQE/CQE support */
633 	if ((issue_flags & (IO_URING_F_SQE128|IO_URING_F_CQE32)) !=
634 	    (IO_URING_F_SQE128|IO_URING_F_CQE32))
635 		return -EOPNOTSUPP;
636 	return 0;
637 }
638 
639 static int nvme_ns_uring_cmd(struct nvme_ns *ns, struct io_uring_cmd *ioucmd,
640 			     unsigned int issue_flags)
641 {
642 	struct nvme_ctrl *ctrl = ns->ctrl;
643 	int ret;
644 
645 	ret = nvme_uring_cmd_checks(issue_flags);
646 	if (ret)
647 		return ret;
648 
649 	switch (ioucmd->cmd_op) {
650 	case NVME_URING_CMD_IO:
651 		ret = nvme_uring_cmd_io(ctrl, ns, ioucmd, issue_flags, false);
652 		break;
653 	case NVME_URING_CMD_IO_VEC:
654 		ret = nvme_uring_cmd_io(ctrl, ns, ioucmd, issue_flags, true);
655 		break;
656 	default:
657 		ret = -ENOTTY;
658 	}
659 
660 	return ret;
661 }
662 
663 int nvme_ns_chr_uring_cmd(struct io_uring_cmd *ioucmd, unsigned int issue_flags)
664 {
665 	struct nvme_ns *ns = container_of(file_inode(ioucmd->file)->i_cdev,
666 			struct nvme_ns, cdev);
667 
668 	return nvme_ns_uring_cmd(ns, ioucmd, issue_flags);
669 }
670 
671 int nvme_ns_chr_uring_cmd_iopoll(struct io_uring_cmd *ioucmd,
672 				 struct io_comp_batch *iob,
673 				 unsigned int poll_flags)
674 {
675 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
676 	struct request *req = pdu->req;
677 
678 	if (req && blk_rq_is_poll(req))
679 		return blk_rq_poll(req, iob, poll_flags);
680 	return 0;
681 }
682 #ifdef CONFIG_NVME_MULTIPATH
683 static int nvme_ns_head_ctrl_ioctl(struct nvme_ns *ns, unsigned int cmd,
684 		void __user *argp, struct nvme_ns_head *head, int srcu_idx,
685 		bool open_for_write)
686 	__releases(&head->srcu)
687 {
688 	struct nvme_ctrl *ctrl = ns->ctrl;
689 	int ret;
690 
691 	nvme_get_ctrl(ns->ctrl);
692 	srcu_read_unlock(&head->srcu, srcu_idx);
693 	ret = nvme_ctrl_ioctl(ns->ctrl, cmd, argp, open_for_write);
694 
695 	nvme_put_ctrl(ctrl);
696 	return ret;
697 }
698 
699 int nvme_ns_head_ioctl(struct block_device *bdev, blk_mode_t mode,
700 		unsigned int cmd, unsigned long arg)
701 {
702 	struct nvme_ns_head *head = bdev->bd_disk->private_data;
703 	bool open_for_write = mode & BLK_OPEN_WRITE;
704 	void __user *argp = (void __user *)arg;
705 	struct nvme_ns *ns;
706 	int srcu_idx, ret = -EWOULDBLOCK;
707 	unsigned int flags = 0;
708 
709 	if (bdev_is_partition(bdev))
710 		flags |= NVME_IOCTL_PARTITION;
711 
712 	srcu_idx = srcu_read_lock(&head->srcu);
713 	ns = nvme_find_path(head);
714 	if (!ns)
715 		goto out_unlock;
716 
717 	/*
718 	 * Handle ioctls that apply to the controller instead of the namespace
719 	 * seperately and drop the ns SRCU reference early.  This avoids a
720 	 * deadlock when deleting namespaces using the passthrough interface.
721 	 */
722 	if (is_ctrl_ioctl(cmd))
723 		return nvme_ns_head_ctrl_ioctl(ns, cmd, argp, head, srcu_idx,
724 					       open_for_write);
725 
726 	ret = nvme_ns_ioctl(ns, cmd, argp, flags, open_for_write);
727 out_unlock:
728 	srcu_read_unlock(&head->srcu, srcu_idx);
729 	return ret;
730 }
731 
732 long nvme_ns_head_chr_ioctl(struct file *file, unsigned int cmd,
733 		unsigned long arg)
734 {
735 	bool open_for_write = file->f_mode & FMODE_WRITE;
736 	struct cdev *cdev = file_inode(file)->i_cdev;
737 	struct nvme_ns_head *head =
738 		container_of(cdev, struct nvme_ns_head, cdev);
739 	void __user *argp = (void __user *)arg;
740 	struct nvme_ns *ns;
741 	int srcu_idx, ret = -EWOULDBLOCK;
742 
743 	srcu_idx = srcu_read_lock(&head->srcu);
744 	ns = nvme_find_path(head);
745 	if (!ns)
746 		goto out_unlock;
747 
748 	if (is_ctrl_ioctl(cmd))
749 		return nvme_ns_head_ctrl_ioctl(ns, cmd, argp, head, srcu_idx,
750 				open_for_write);
751 
752 	ret = nvme_ns_ioctl(ns, cmd, argp, 0, open_for_write);
753 out_unlock:
754 	srcu_read_unlock(&head->srcu, srcu_idx);
755 	return ret;
756 }
757 
758 int nvme_ns_head_chr_uring_cmd(struct io_uring_cmd *ioucmd,
759 		unsigned int issue_flags)
760 {
761 	struct cdev *cdev = file_inode(ioucmd->file)->i_cdev;
762 	struct nvme_ns_head *head = container_of(cdev, struct nvme_ns_head, cdev);
763 	int srcu_idx = srcu_read_lock(&head->srcu);
764 	struct nvme_ns *ns = nvme_find_path(head);
765 	int ret = -EINVAL;
766 
767 	if (ns)
768 		ret = nvme_ns_uring_cmd(ns, ioucmd, issue_flags);
769 	srcu_read_unlock(&head->srcu, srcu_idx);
770 	return ret;
771 }
772 #endif /* CONFIG_NVME_MULTIPATH */
773 
774 int nvme_dev_uring_cmd(struct io_uring_cmd *ioucmd, unsigned int issue_flags)
775 {
776 	struct nvme_ctrl *ctrl = ioucmd->file->private_data;
777 	int ret;
778 
779 	/* IOPOLL not supported yet */
780 	if (issue_flags & IO_URING_F_IOPOLL)
781 		return -EOPNOTSUPP;
782 
783 	ret = nvme_uring_cmd_checks(issue_flags);
784 	if (ret)
785 		return ret;
786 
787 	switch (ioucmd->cmd_op) {
788 	case NVME_URING_CMD_ADMIN:
789 		ret = nvme_uring_cmd_io(ctrl, NULL, ioucmd, issue_flags, false);
790 		break;
791 	case NVME_URING_CMD_ADMIN_VEC:
792 		ret = nvme_uring_cmd_io(ctrl, NULL, ioucmd, issue_flags, true);
793 		break;
794 	default:
795 		ret = -ENOTTY;
796 	}
797 
798 	return ret;
799 }
800 
801 static int nvme_dev_user_cmd(struct nvme_ctrl *ctrl, void __user *argp,
802 		bool open_for_write)
803 {
804 	struct nvme_ns *ns;
805 	int ret, srcu_idx;
806 
807 	srcu_idx = srcu_read_lock(&ctrl->srcu);
808 	if (list_empty(&ctrl->namespaces)) {
809 		ret = -ENOTTY;
810 		goto out_unlock;
811 	}
812 
813 	ns = list_first_or_null_rcu(&ctrl->namespaces, struct nvme_ns, list);
814 	if (ns != list_last_entry(&ctrl->namespaces, struct nvme_ns, list)) {
815 		dev_warn(ctrl->device,
816 			"NVME_IOCTL_IO_CMD not supported when multiple namespaces present!\n");
817 		ret = -EINVAL;
818 		goto out_unlock;
819 	}
820 
821 	dev_warn(ctrl->device,
822 		"using deprecated NVME_IOCTL_IO_CMD ioctl on the char device!\n");
823 	if (!nvme_get_ns(ns)) {
824 		ret = -ENXIO;
825 		goto out_unlock;
826 	}
827 	srcu_read_unlock(&ctrl->srcu, srcu_idx);
828 
829 	ret = nvme_user_cmd(ctrl, ns, argp, 0, open_for_write);
830 	nvme_put_ns(ns);
831 	return ret;
832 
833 out_unlock:
834 	srcu_read_unlock(&ctrl->srcu, srcu_idx);
835 	return ret;
836 }
837 
838 long nvme_dev_ioctl(struct file *file, unsigned int cmd,
839 		unsigned long arg)
840 {
841 	bool open_for_write = file->f_mode & FMODE_WRITE;
842 	struct nvme_ctrl *ctrl = file->private_data;
843 	void __user *argp = (void __user *)arg;
844 
845 	switch (cmd) {
846 	case NVME_IOCTL_ADMIN_CMD:
847 		return nvme_user_cmd(ctrl, NULL, argp, 0, open_for_write);
848 	case NVME_IOCTL_ADMIN64_CMD:
849 		return nvme_user_cmd64(ctrl, NULL, argp, 0, open_for_write);
850 	case NVME_IOCTL_IO_CMD:
851 		return nvme_dev_user_cmd(ctrl, argp, open_for_write);
852 	case NVME_IOCTL_RESET:
853 		if (!capable(CAP_SYS_ADMIN))
854 			return -EACCES;
855 		dev_warn(ctrl->device, "resetting controller\n");
856 		return nvme_reset_ctrl_sync(ctrl);
857 	case NVME_IOCTL_SUBSYS_RESET:
858 		if (!capable(CAP_SYS_ADMIN))
859 			return -EACCES;
860 		return nvme_reset_subsystem(ctrl);
861 	case NVME_IOCTL_RESCAN:
862 		if (!capable(CAP_SYS_ADMIN))
863 			return -EACCES;
864 		nvme_queue_scan(ctrl);
865 		return 0;
866 	default:
867 		return -ENOTTY;
868 	}
869 }
870