xref: /linux/drivers/nvme/host/ioctl.c (revision fbf8fb328d1bfe3bd17d5c5626cb485a1ca1a50d)
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) of namespace\n",
287 			current->comm, nsid, ns->head->ns_id);
288 		return false;
289 	}
290 
291 	return true;
292 }
293 
294 static int nvme_user_cmd(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
295 		struct nvme_passthru_cmd __user *ucmd, unsigned int flags,
296 		bool open_for_write)
297 {
298 	struct nvme_passthru_cmd cmd;
299 	struct nvme_command c;
300 	unsigned timeout = 0;
301 	u64 result;
302 	int status;
303 
304 	if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
305 		return -EFAULT;
306 	if (cmd.flags)
307 		return -EINVAL;
308 	if (!nvme_validate_passthru_nsid(ctrl, ns, cmd.nsid))
309 		return -EINVAL;
310 
311 	memset(&c, 0, sizeof(c));
312 	c.common.opcode = cmd.opcode;
313 	c.common.flags = cmd.flags;
314 	c.common.nsid = cpu_to_le32(cmd.nsid);
315 	c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
316 	c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
317 	c.common.cdw10 = cpu_to_le32(cmd.cdw10);
318 	c.common.cdw11 = cpu_to_le32(cmd.cdw11);
319 	c.common.cdw12 = cpu_to_le32(cmd.cdw12);
320 	c.common.cdw13 = cpu_to_le32(cmd.cdw13);
321 	c.common.cdw14 = cpu_to_le32(cmd.cdw14);
322 	c.common.cdw15 = cpu_to_le32(cmd.cdw15);
323 
324 	if (!nvme_cmd_allowed(ns, &c, 0, open_for_write))
325 		return -EACCES;
326 
327 	if (cmd.timeout_ms)
328 		timeout = msecs_to_jiffies(cmd.timeout_ms);
329 
330 	status = nvme_submit_user_cmd(ns ? ns->queue : ctrl->admin_q, &c,
331 			cmd.addr, cmd.data_len, nvme_to_user_ptr(cmd.metadata),
332 			cmd.metadata_len, &result, timeout, 0);
333 
334 	if (status >= 0) {
335 		if (put_user(result, &ucmd->result))
336 			return -EFAULT;
337 	}
338 
339 	return status;
340 }
341 
342 static int nvme_user_cmd64(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
343 		struct nvme_passthru_cmd64 __user *ucmd, unsigned int flags,
344 		bool open_for_write)
345 {
346 	struct nvme_passthru_cmd64 cmd;
347 	struct nvme_command c;
348 	unsigned timeout = 0;
349 	int status;
350 
351 	if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
352 		return -EFAULT;
353 	if (cmd.flags)
354 		return -EINVAL;
355 	if (!nvme_validate_passthru_nsid(ctrl, ns, cmd.nsid))
356 		return -EINVAL;
357 
358 	memset(&c, 0, sizeof(c));
359 	c.common.opcode = cmd.opcode;
360 	c.common.flags = cmd.flags;
361 	c.common.nsid = cpu_to_le32(cmd.nsid);
362 	c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
363 	c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
364 	c.common.cdw10 = cpu_to_le32(cmd.cdw10);
365 	c.common.cdw11 = cpu_to_le32(cmd.cdw11);
366 	c.common.cdw12 = cpu_to_le32(cmd.cdw12);
367 	c.common.cdw13 = cpu_to_le32(cmd.cdw13);
368 	c.common.cdw14 = cpu_to_le32(cmd.cdw14);
369 	c.common.cdw15 = cpu_to_le32(cmd.cdw15);
370 
371 	if (!nvme_cmd_allowed(ns, &c, flags, open_for_write))
372 		return -EACCES;
373 
374 	if (cmd.timeout_ms)
375 		timeout = msecs_to_jiffies(cmd.timeout_ms);
376 
377 	status = nvme_submit_user_cmd(ns ? ns->queue : ctrl->admin_q, &c,
378 			cmd.addr, cmd.data_len, nvme_to_user_ptr(cmd.metadata),
379 			cmd.metadata_len, &cmd.result, timeout, flags);
380 
381 	if (status >= 0) {
382 		if (put_user(cmd.result, &ucmd->result))
383 			return -EFAULT;
384 	}
385 
386 	return status;
387 }
388 
389 struct nvme_uring_data {
390 	__u64	metadata;
391 	__u64	addr;
392 	__u32	data_len;
393 	__u32	metadata_len;
394 	__u32	timeout_ms;
395 };
396 
397 /*
398  * This overlays struct io_uring_cmd pdu.
399  * Expect build errors if this grows larger than that.
400  */
401 struct nvme_uring_cmd_pdu {
402 	struct request *req;
403 	struct bio *bio;
404 	u64 result;
405 	int status;
406 };
407 
408 static inline struct nvme_uring_cmd_pdu *nvme_uring_cmd_pdu(
409 		struct io_uring_cmd *ioucmd)
410 {
411 	return io_uring_cmd_to_pdu(ioucmd, struct nvme_uring_cmd_pdu);
412 }
413 
414 static void nvme_uring_task_cb(struct io_uring_cmd *ioucmd,
415 			       unsigned issue_flags)
416 {
417 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
418 
419 	if (pdu->bio)
420 		blk_rq_unmap_user(pdu->bio);
421 	io_uring_cmd_done(ioucmd, pdu->status, pdu->result, issue_flags);
422 }
423 
424 static enum rq_end_io_ret nvme_uring_cmd_end_io(struct request *req,
425 						blk_status_t err)
426 {
427 	struct io_uring_cmd *ioucmd = req->end_io_data;
428 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
429 
430 	if (nvme_req(req)->flags & NVME_REQ_CANCELLED) {
431 		pdu->status = -EINTR;
432 	} else {
433 		pdu->status = nvme_req(req)->status;
434 		if (!pdu->status)
435 			pdu->status = blk_status_to_errno(err);
436 	}
437 	pdu->result = le64_to_cpu(nvme_req(req)->result.u64);
438 
439 	/*
440 	 * For iopoll, complete it directly. Note that using the uring_cmd
441 	 * helper for this is safe only because we check blk_rq_is_poll().
442 	 * As that returns false if we're NOT on a polled queue, then it's
443 	 * safe to use the polled completion helper.
444 	 *
445 	 * Otherwise, move the completion to task work.
446 	 */
447 	if (blk_rq_is_poll(req)) {
448 		if (pdu->bio)
449 			blk_rq_unmap_user(pdu->bio);
450 		io_uring_cmd_iopoll_done(ioucmd, pdu->result, pdu->status);
451 	} else {
452 		io_uring_cmd_do_in_task_lazy(ioucmd, nvme_uring_task_cb);
453 	}
454 
455 	return RQ_END_IO_FREE;
456 }
457 
458 static int nvme_uring_cmd_io(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
459 		struct io_uring_cmd *ioucmd, unsigned int issue_flags, bool vec)
460 {
461 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
462 	const struct nvme_uring_cmd *cmd = io_uring_sqe_cmd(ioucmd->sqe);
463 	struct request_queue *q = ns ? ns->queue : ctrl->admin_q;
464 	struct nvme_uring_data d;
465 	struct nvme_command c;
466 	struct request *req;
467 	blk_opf_t rq_flags = REQ_ALLOC_CACHE;
468 	blk_mq_req_flags_t blk_flags = 0;
469 	int ret;
470 
471 	c.common.opcode = READ_ONCE(cmd->opcode);
472 	c.common.flags = READ_ONCE(cmd->flags);
473 	if (c.common.flags)
474 		return -EINVAL;
475 
476 	c.common.command_id = 0;
477 	c.common.nsid = cpu_to_le32(cmd->nsid);
478 	if (!nvme_validate_passthru_nsid(ctrl, ns, le32_to_cpu(c.common.nsid)))
479 		return -EINVAL;
480 
481 	c.common.cdw2[0] = cpu_to_le32(READ_ONCE(cmd->cdw2));
482 	c.common.cdw2[1] = cpu_to_le32(READ_ONCE(cmd->cdw3));
483 	c.common.metadata = 0;
484 	c.common.dptr.prp1 = c.common.dptr.prp2 = 0;
485 	c.common.cdw10 = cpu_to_le32(READ_ONCE(cmd->cdw10));
486 	c.common.cdw11 = cpu_to_le32(READ_ONCE(cmd->cdw11));
487 	c.common.cdw12 = cpu_to_le32(READ_ONCE(cmd->cdw12));
488 	c.common.cdw13 = cpu_to_le32(READ_ONCE(cmd->cdw13));
489 	c.common.cdw14 = cpu_to_le32(READ_ONCE(cmd->cdw14));
490 	c.common.cdw15 = cpu_to_le32(READ_ONCE(cmd->cdw15));
491 
492 	if (!nvme_cmd_allowed(ns, &c, 0, ioucmd->file->f_mode & FMODE_WRITE))
493 		return -EACCES;
494 
495 	d.metadata = READ_ONCE(cmd->metadata);
496 	d.addr = READ_ONCE(cmd->addr);
497 	d.data_len = READ_ONCE(cmd->data_len);
498 	d.metadata_len = READ_ONCE(cmd->metadata_len);
499 	d.timeout_ms = READ_ONCE(cmd->timeout_ms);
500 
501 	if (issue_flags & IO_URING_F_NONBLOCK) {
502 		rq_flags |= REQ_NOWAIT;
503 		blk_flags = BLK_MQ_REQ_NOWAIT;
504 	}
505 	if (issue_flags & IO_URING_F_IOPOLL)
506 		rq_flags |= REQ_POLLED;
507 
508 	req = nvme_alloc_user_request(q, &c, rq_flags, blk_flags);
509 	if (IS_ERR(req))
510 		return PTR_ERR(req);
511 	req->timeout = d.timeout_ms ? msecs_to_jiffies(d.timeout_ms) : 0;
512 
513 	if (d.addr && d.data_len) {
514 		ret = nvme_map_user_request(req, d.addr,
515 			d.data_len, nvme_to_user_ptr(d.metadata),
516 			d.metadata_len, ioucmd, vec);
517 		if (ret)
518 			return ret;
519 	}
520 
521 	/* to free bio on completion, as req->bio will be null at that time */
522 	pdu->bio = req->bio;
523 	pdu->req = req;
524 	req->end_io_data = ioucmd;
525 	req->end_io = nvme_uring_cmd_end_io;
526 	blk_execute_rq_nowait(req, false);
527 	return -EIOCBQUEUED;
528 }
529 
530 static bool is_ctrl_ioctl(unsigned int cmd)
531 {
532 	if (cmd == NVME_IOCTL_ADMIN_CMD || cmd == NVME_IOCTL_ADMIN64_CMD)
533 		return true;
534 	if (is_sed_ioctl(cmd))
535 		return true;
536 	return false;
537 }
538 
539 static int nvme_ctrl_ioctl(struct nvme_ctrl *ctrl, unsigned int cmd,
540 		void __user *argp, bool open_for_write)
541 {
542 	switch (cmd) {
543 	case NVME_IOCTL_ADMIN_CMD:
544 		return nvme_user_cmd(ctrl, NULL, argp, 0, open_for_write);
545 	case NVME_IOCTL_ADMIN64_CMD:
546 		return nvme_user_cmd64(ctrl, NULL, argp, 0, open_for_write);
547 	default:
548 		return sed_ioctl(ctrl->opal_dev, cmd, argp);
549 	}
550 }
551 
552 #ifdef COMPAT_FOR_U64_ALIGNMENT
553 struct nvme_user_io32 {
554 	__u8	opcode;
555 	__u8	flags;
556 	__u16	control;
557 	__u16	nblocks;
558 	__u16	rsvd;
559 	__u64	metadata;
560 	__u64	addr;
561 	__u64	slba;
562 	__u32	dsmgmt;
563 	__u32	reftag;
564 	__u16	apptag;
565 	__u16	appmask;
566 } __attribute__((__packed__));
567 #define NVME_IOCTL_SUBMIT_IO32	_IOW('N', 0x42, struct nvme_user_io32)
568 #endif /* COMPAT_FOR_U64_ALIGNMENT */
569 
570 static int nvme_ns_ioctl(struct nvme_ns *ns, unsigned int cmd,
571 		void __user *argp, unsigned int flags, bool open_for_write)
572 {
573 	switch (cmd) {
574 	case NVME_IOCTL_ID:
575 		force_successful_syscall_return();
576 		return ns->head->ns_id;
577 	case NVME_IOCTL_IO_CMD:
578 		return nvme_user_cmd(ns->ctrl, ns, argp, flags, open_for_write);
579 	/*
580 	 * struct nvme_user_io can have different padding on some 32-bit ABIs.
581 	 * Just accept the compat version as all fields that are used are the
582 	 * same size and at the same offset.
583 	 */
584 #ifdef COMPAT_FOR_U64_ALIGNMENT
585 	case NVME_IOCTL_SUBMIT_IO32:
586 #endif
587 	case NVME_IOCTL_SUBMIT_IO:
588 		return nvme_submit_io(ns, argp);
589 	case NVME_IOCTL_IO64_CMD_VEC:
590 		flags |= NVME_IOCTL_VEC;
591 		fallthrough;
592 	case NVME_IOCTL_IO64_CMD:
593 		return nvme_user_cmd64(ns->ctrl, ns, argp, flags,
594 				       open_for_write);
595 	default:
596 		return -ENOTTY;
597 	}
598 }
599 
600 int nvme_ioctl(struct block_device *bdev, blk_mode_t mode,
601 		unsigned int cmd, unsigned long arg)
602 {
603 	struct nvme_ns *ns = bdev->bd_disk->private_data;
604 	bool open_for_write = mode & BLK_OPEN_WRITE;
605 	void __user *argp = (void __user *)arg;
606 	unsigned int flags = 0;
607 
608 	if (bdev_is_partition(bdev))
609 		flags |= NVME_IOCTL_PARTITION;
610 
611 	if (is_ctrl_ioctl(cmd))
612 		return nvme_ctrl_ioctl(ns->ctrl, cmd, argp, open_for_write);
613 	return nvme_ns_ioctl(ns, cmd, argp, flags, open_for_write);
614 }
615 
616 long nvme_ns_chr_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
617 {
618 	struct nvme_ns *ns =
619 		container_of(file_inode(file)->i_cdev, struct nvme_ns, cdev);
620 	bool open_for_write = file->f_mode & FMODE_WRITE;
621 	void __user *argp = (void __user *)arg;
622 
623 	if (is_ctrl_ioctl(cmd))
624 		return nvme_ctrl_ioctl(ns->ctrl, cmd, argp, open_for_write);
625 	return nvme_ns_ioctl(ns, cmd, argp, 0, open_for_write);
626 }
627 
628 static int nvme_uring_cmd_checks(unsigned int issue_flags)
629 {
630 
631 	/* NVMe passthrough requires big SQE/CQE support */
632 	if ((issue_flags & (IO_URING_F_SQE128|IO_URING_F_CQE32)) !=
633 	    (IO_URING_F_SQE128|IO_URING_F_CQE32))
634 		return -EOPNOTSUPP;
635 	return 0;
636 }
637 
638 static int nvme_ns_uring_cmd(struct nvme_ns *ns, struct io_uring_cmd *ioucmd,
639 			     unsigned int issue_flags)
640 {
641 	struct nvme_ctrl *ctrl = ns->ctrl;
642 	int ret;
643 
644 	ret = nvme_uring_cmd_checks(issue_flags);
645 	if (ret)
646 		return ret;
647 
648 	switch (ioucmd->cmd_op) {
649 	case NVME_URING_CMD_IO:
650 		ret = nvme_uring_cmd_io(ctrl, ns, ioucmd, issue_flags, false);
651 		break;
652 	case NVME_URING_CMD_IO_VEC:
653 		ret = nvme_uring_cmd_io(ctrl, ns, ioucmd, issue_flags, true);
654 		break;
655 	default:
656 		ret = -ENOTTY;
657 	}
658 
659 	return ret;
660 }
661 
662 int nvme_ns_chr_uring_cmd(struct io_uring_cmd *ioucmd, unsigned int issue_flags)
663 {
664 	struct nvme_ns *ns = container_of(file_inode(ioucmd->file)->i_cdev,
665 			struct nvme_ns, cdev);
666 
667 	return nvme_ns_uring_cmd(ns, ioucmd, issue_flags);
668 }
669 
670 int nvme_ns_chr_uring_cmd_iopoll(struct io_uring_cmd *ioucmd,
671 				 struct io_comp_batch *iob,
672 				 unsigned int poll_flags)
673 {
674 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
675 	struct request *req = pdu->req;
676 
677 	if (req && blk_rq_is_poll(req))
678 		return blk_rq_poll(req, iob, poll_flags);
679 	return 0;
680 }
681 #ifdef CONFIG_NVME_MULTIPATH
682 static int nvme_ns_head_ctrl_ioctl(struct nvme_ns *ns, unsigned int cmd,
683 		void __user *argp, struct nvme_ns_head *head, int srcu_idx,
684 		bool open_for_write)
685 	__releases(&head->srcu)
686 {
687 	struct nvme_ctrl *ctrl = ns->ctrl;
688 	int ret;
689 
690 	nvme_get_ctrl(ns->ctrl);
691 	srcu_read_unlock(&head->srcu, srcu_idx);
692 	ret = nvme_ctrl_ioctl(ns->ctrl, cmd, argp, open_for_write);
693 
694 	nvme_put_ctrl(ctrl);
695 	return ret;
696 }
697 
698 int nvme_ns_head_ioctl(struct block_device *bdev, blk_mode_t mode,
699 		unsigned int cmd, unsigned long arg)
700 {
701 	struct nvme_ns_head *head = bdev->bd_disk->private_data;
702 	bool open_for_write = mode & BLK_OPEN_WRITE;
703 	void __user *argp = (void __user *)arg;
704 	struct nvme_ns *ns;
705 	int srcu_idx, ret = -EWOULDBLOCK;
706 	unsigned int flags = 0;
707 
708 	if (bdev_is_partition(bdev))
709 		flags |= NVME_IOCTL_PARTITION;
710 
711 	srcu_idx = srcu_read_lock(&head->srcu);
712 	ns = nvme_find_path(head);
713 	if (!ns)
714 		goto out_unlock;
715 
716 	/*
717 	 * Handle ioctls that apply to the controller instead of the namespace
718 	 * seperately and drop the ns SRCU reference early.  This avoids a
719 	 * deadlock when deleting namespaces using the passthrough interface.
720 	 */
721 	if (is_ctrl_ioctl(cmd))
722 		return nvme_ns_head_ctrl_ioctl(ns, cmd, argp, head, srcu_idx,
723 					       open_for_write);
724 
725 	ret = nvme_ns_ioctl(ns, cmd, argp, flags, open_for_write);
726 out_unlock:
727 	srcu_read_unlock(&head->srcu, srcu_idx);
728 	return ret;
729 }
730 
731 long nvme_ns_head_chr_ioctl(struct file *file, unsigned int cmd,
732 		unsigned long arg)
733 {
734 	bool open_for_write = file->f_mode & FMODE_WRITE;
735 	struct cdev *cdev = file_inode(file)->i_cdev;
736 	struct nvme_ns_head *head =
737 		container_of(cdev, struct nvme_ns_head, cdev);
738 	void __user *argp = (void __user *)arg;
739 	struct nvme_ns *ns;
740 	int srcu_idx, ret = -EWOULDBLOCK;
741 
742 	srcu_idx = srcu_read_lock(&head->srcu);
743 	ns = nvme_find_path(head);
744 	if (!ns)
745 		goto out_unlock;
746 
747 	if (is_ctrl_ioctl(cmd))
748 		return nvme_ns_head_ctrl_ioctl(ns, cmd, argp, head, srcu_idx,
749 				open_for_write);
750 
751 	ret = nvme_ns_ioctl(ns, cmd, argp, 0, open_for_write);
752 out_unlock:
753 	srcu_read_unlock(&head->srcu, srcu_idx);
754 	return ret;
755 }
756 
757 int nvme_ns_head_chr_uring_cmd(struct io_uring_cmd *ioucmd,
758 		unsigned int issue_flags)
759 {
760 	struct cdev *cdev = file_inode(ioucmd->file)->i_cdev;
761 	struct nvme_ns_head *head = container_of(cdev, struct nvme_ns_head, cdev);
762 	int srcu_idx = srcu_read_lock(&head->srcu);
763 	struct nvme_ns *ns = nvme_find_path(head);
764 	int ret = -EINVAL;
765 
766 	if (ns)
767 		ret = nvme_ns_uring_cmd(ns, ioucmd, issue_flags);
768 	srcu_read_unlock(&head->srcu, srcu_idx);
769 	return ret;
770 }
771 #endif /* CONFIG_NVME_MULTIPATH */
772 
773 int nvme_dev_uring_cmd(struct io_uring_cmd *ioucmd, unsigned int issue_flags)
774 {
775 	struct nvme_ctrl *ctrl = ioucmd->file->private_data;
776 	int ret;
777 
778 	/* IOPOLL not supported yet */
779 	if (issue_flags & IO_URING_F_IOPOLL)
780 		return -EOPNOTSUPP;
781 
782 	ret = nvme_uring_cmd_checks(issue_flags);
783 	if (ret)
784 		return ret;
785 
786 	switch (ioucmd->cmd_op) {
787 	case NVME_URING_CMD_ADMIN:
788 		ret = nvme_uring_cmd_io(ctrl, NULL, ioucmd, issue_flags, false);
789 		break;
790 	case NVME_URING_CMD_ADMIN_VEC:
791 		ret = nvme_uring_cmd_io(ctrl, NULL, ioucmd, issue_flags, true);
792 		break;
793 	default:
794 		ret = -ENOTTY;
795 	}
796 
797 	return ret;
798 }
799 
800 static int nvme_dev_user_cmd(struct nvme_ctrl *ctrl, void __user *argp,
801 		bool open_for_write)
802 {
803 	struct nvme_ns *ns;
804 	int ret, srcu_idx;
805 
806 	srcu_idx = srcu_read_lock(&ctrl->srcu);
807 	if (list_empty(&ctrl->namespaces)) {
808 		ret = -ENOTTY;
809 		goto out_unlock;
810 	}
811 
812 	ns = list_first_or_null_rcu(&ctrl->namespaces, struct nvme_ns, list);
813 	if (ns != list_last_entry(&ctrl->namespaces, struct nvme_ns, list)) {
814 		dev_warn(ctrl->device,
815 			"NVME_IOCTL_IO_CMD not supported when multiple namespaces present!\n");
816 		ret = -EINVAL;
817 		goto out_unlock;
818 	}
819 
820 	dev_warn(ctrl->device,
821 		"using deprecated NVME_IOCTL_IO_CMD ioctl on the char device!\n");
822 	if (!nvme_get_ns(ns)) {
823 		ret = -ENXIO;
824 		goto out_unlock;
825 	}
826 	srcu_read_unlock(&ctrl->srcu, srcu_idx);
827 
828 	ret = nvme_user_cmd(ctrl, ns, argp, 0, open_for_write);
829 	nvme_put_ns(ns);
830 	return ret;
831 
832 out_unlock:
833 	srcu_read_unlock(&ctrl->srcu, srcu_idx);
834 	return ret;
835 }
836 
837 long nvme_dev_ioctl(struct file *file, unsigned int cmd,
838 		unsigned long arg)
839 {
840 	bool open_for_write = file->f_mode & FMODE_WRITE;
841 	struct nvme_ctrl *ctrl = file->private_data;
842 	void __user *argp = (void __user *)arg;
843 
844 	switch (cmd) {
845 	case NVME_IOCTL_ADMIN_CMD:
846 		return nvme_user_cmd(ctrl, NULL, argp, 0, open_for_write);
847 	case NVME_IOCTL_ADMIN64_CMD:
848 		return nvme_user_cmd64(ctrl, NULL, argp, 0, open_for_write);
849 	case NVME_IOCTL_IO_CMD:
850 		return nvme_dev_user_cmd(ctrl, argp, open_for_write);
851 	case NVME_IOCTL_RESET:
852 		if (!capable(CAP_SYS_ADMIN))
853 			return -EACCES;
854 		dev_warn(ctrl->device, "resetting controller\n");
855 		return nvme_reset_ctrl_sync(ctrl);
856 	case NVME_IOCTL_SUBSYS_RESET:
857 		if (!capable(CAP_SYS_ADMIN))
858 			return -EACCES;
859 		return nvme_reset_subsystem(ctrl);
860 	case NVME_IOCTL_RESCAN:
861 		if (!capable(CAP_SYS_ADMIN))
862 			return -EACCES;
863 		nvme_queue_scan(ctrl);
864 		return 0;
865 	default:
866 		return -ENOTTY;
867 	}
868 }
869