xref: /openbmc/linux/drivers/nvme/host/ioctl.c (revision 6db6b729)
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/ptrace.h>	/* for force_successful_syscall_return */
7 #include <linux/nvme_ioctl.h>
8 #include <linux/io_uring.h>
9 #include "nvme.h"
10 
11 enum {
12 	NVME_IOCTL_VEC		= (1 << 0),
13 	NVME_IOCTL_PARTITION	= (1 << 1),
14 };
15 
16 static bool nvme_cmd_allowed(struct nvme_ns *ns, struct nvme_command *c,
17 		unsigned int flags, bool open_for_write)
18 {
19 	u32 effects;
20 
21 	if (capable(CAP_SYS_ADMIN))
22 		return true;
23 
24 	/*
25 	 * Do not allow unprivileged passthrough on partitions, as that allows an
26 	 * escape from the containment of the partition.
27 	 */
28 	if (flags & NVME_IOCTL_PARTITION)
29 		return false;
30 
31 	/*
32 	 * Do not allow unprivileged processes to send vendor specific or fabrics
33 	 * commands as we can't be sure about their effects.
34 	 */
35 	if (c->common.opcode >= nvme_cmd_vendor_start ||
36 	    c->common.opcode == nvme_fabrics_command)
37 		return false;
38 
39 	/*
40 	 * Do not allow unprivileged passthrough of admin commands except
41 	 * for a subset of identify commands that contain information required
42 	 * to form proper I/O commands in userspace and do not expose any
43 	 * potentially sensitive information.
44 	 */
45 	if (!ns) {
46 		if (c->common.opcode == nvme_admin_identify) {
47 			switch (c->identify.cns) {
48 			case NVME_ID_CNS_NS:
49 			case NVME_ID_CNS_CS_NS:
50 			case NVME_ID_CNS_NS_CS_INDEP:
51 			case NVME_ID_CNS_CS_CTRL:
52 			case NVME_ID_CNS_CTRL:
53 				return true;
54 			}
55 		}
56 		return false;
57 	}
58 
59 	/*
60 	 * Check if the controller provides a Commands Supported and Effects log
61 	 * and marks this command as supported.  If not reject unprivileged
62 	 * passthrough.
63 	 */
64 	effects = nvme_command_effects(ns->ctrl, ns, c->common.opcode);
65 	if (!(effects & NVME_CMD_EFFECTS_CSUPP))
66 		return false;
67 
68 	/*
69 	 * Don't allow passthrough for command that have intrusive (or unknown)
70 	 * effects.
71 	 */
72 	if (effects & ~(NVME_CMD_EFFECTS_CSUPP | NVME_CMD_EFFECTS_LBCC |
73 			NVME_CMD_EFFECTS_UUID_SEL |
74 			NVME_CMD_EFFECTS_SCOPE_MASK))
75 		return false;
76 
77 	/*
78 	 * Only allow I/O commands that transfer data to the controller or that
79 	 * change the logical block contents if the file descriptor is open for
80 	 * writing.
81 	 */
82 	if (nvme_is_write(c) || (effects & NVME_CMD_EFFECTS_LBCC))
83 		return open_for_write;
84 	return true;
85 }
86 
87 /*
88  * Convert integer values from ioctl structures to user pointers, silently
89  * ignoring the upper bits in the compat case to match behaviour of 32-bit
90  * kernels.
91  */
92 static void __user *nvme_to_user_ptr(uintptr_t ptrval)
93 {
94 	if (in_compat_syscall())
95 		ptrval = (compat_uptr_t)ptrval;
96 	return (void __user *)ptrval;
97 }
98 
99 static void *nvme_add_user_metadata(struct request *req, void __user *ubuf,
100 		unsigned len, u32 seed)
101 {
102 	struct bio_integrity_payload *bip;
103 	int ret = -ENOMEM;
104 	void *buf;
105 	struct bio *bio = req->bio;
106 
107 	buf = kmalloc(len, GFP_KERNEL);
108 	if (!buf)
109 		goto out;
110 
111 	if (req_op(req) == REQ_OP_DRV_OUT) {
112 		ret = -EFAULT;
113 		if (copy_from_user(buf, ubuf, len))
114 			goto out_free_meta;
115 	} else {
116 		memset(buf, 0, len);
117 	}
118 
119 	bip = bio_integrity_alloc(bio, GFP_KERNEL, 1);
120 	if (IS_ERR(bip)) {
121 		ret = PTR_ERR(bip);
122 		goto out_free_meta;
123 	}
124 
125 	bip->bip_iter.bi_sector = seed;
126 	ret = bio_integrity_add_page(bio, virt_to_page(buf), len,
127 			offset_in_page(buf));
128 	if (ret != len) {
129 		ret = -ENOMEM;
130 		goto out_free_meta;
131 	}
132 
133 	req->cmd_flags |= REQ_INTEGRITY;
134 	return buf;
135 out_free_meta:
136 	kfree(buf);
137 out:
138 	return ERR_PTR(ret);
139 }
140 
141 static int nvme_finish_user_metadata(struct request *req, void __user *ubuf,
142 		void *meta, unsigned len, int ret)
143 {
144 	if (!ret && req_op(req) == REQ_OP_DRV_IN &&
145 	    copy_to_user(ubuf, meta, len))
146 		ret = -EFAULT;
147 	kfree(meta);
148 	return ret;
149 }
150 
151 static struct request *nvme_alloc_user_request(struct request_queue *q,
152 		struct nvme_command *cmd, blk_opf_t rq_flags,
153 		blk_mq_req_flags_t blk_flags)
154 {
155 	struct request *req;
156 
157 	req = blk_mq_alloc_request(q, nvme_req_op(cmd) | rq_flags, blk_flags);
158 	if (IS_ERR(req))
159 		return req;
160 	nvme_init_request(req, cmd);
161 	nvme_req(req)->flags |= NVME_REQ_USERCMD;
162 	return req;
163 }
164 
165 static int nvme_map_user_request(struct request *req, u64 ubuffer,
166 		unsigned bufflen, void __user *meta_buffer, unsigned meta_len,
167 		u32 meta_seed, void **metap, struct io_uring_cmd *ioucmd,
168 		unsigned int flags)
169 {
170 	struct request_queue *q = req->q;
171 	struct nvme_ns *ns = q->queuedata;
172 	struct block_device *bdev = ns ? ns->disk->part0 : NULL;
173 	struct bio *bio = NULL;
174 	void *meta = NULL;
175 	int ret;
176 
177 	if (ioucmd && (ioucmd->flags & IORING_URING_CMD_FIXED)) {
178 		struct iov_iter iter;
179 
180 		/* fixedbufs is only for non-vectored io */
181 		if (WARN_ON_ONCE(flags & NVME_IOCTL_VEC))
182 			return -EINVAL;
183 		ret = io_uring_cmd_import_fixed(ubuffer, bufflen,
184 				rq_data_dir(req), &iter, ioucmd);
185 		if (ret < 0)
186 			goto out;
187 		ret = blk_rq_map_user_iov(q, req, NULL, &iter, GFP_KERNEL);
188 	} else {
189 		ret = blk_rq_map_user_io(req, NULL, nvme_to_user_ptr(ubuffer),
190 				bufflen, GFP_KERNEL, flags & NVME_IOCTL_VEC, 0,
191 				0, rq_data_dir(req));
192 	}
193 
194 	if (ret)
195 		goto out;
196 	bio = req->bio;
197 	if (bdev)
198 		bio_set_dev(bio, bdev);
199 
200 	if (bdev && meta_buffer && meta_len) {
201 		meta = nvme_add_user_metadata(req, meta_buffer, meta_len,
202 				meta_seed);
203 		if (IS_ERR(meta)) {
204 			ret = PTR_ERR(meta);
205 			goto out_unmap;
206 		}
207 		*metap = meta;
208 	}
209 
210 	return ret;
211 
212 out_unmap:
213 	if (bio)
214 		blk_rq_unmap_user(bio);
215 out:
216 	blk_mq_free_request(req);
217 	return ret;
218 }
219 
220 static int nvme_submit_user_cmd(struct request_queue *q,
221 		struct nvme_command *cmd, u64 ubuffer, unsigned bufflen,
222 		void __user *meta_buffer, unsigned meta_len, u32 meta_seed,
223 		u64 *result, unsigned timeout, unsigned int flags)
224 {
225 	struct nvme_ns *ns = q->queuedata;
226 	struct nvme_ctrl *ctrl;
227 	struct request *req;
228 	void *meta = NULL;
229 	struct bio *bio;
230 	u32 effects;
231 	int ret;
232 
233 	req = nvme_alloc_user_request(q, cmd, 0, 0);
234 	if (IS_ERR(req))
235 		return PTR_ERR(req);
236 
237 	req->timeout = timeout;
238 	if (ubuffer && bufflen) {
239 		ret = nvme_map_user_request(req, ubuffer, bufflen, meta_buffer,
240 				meta_len, meta_seed, &meta, NULL, flags);
241 		if (ret)
242 			return ret;
243 	}
244 
245 	bio = req->bio;
246 	ctrl = nvme_req(req)->ctrl;
247 
248 	effects = nvme_passthru_start(ctrl, ns, cmd->common.opcode);
249 	ret = nvme_execute_rq(req, false);
250 	if (result)
251 		*result = le64_to_cpu(nvme_req(req)->result.u64);
252 	if (meta)
253 		ret = nvme_finish_user_metadata(req, meta_buffer, meta,
254 						meta_len, ret);
255 	if (bio)
256 		blk_rq_unmap_user(bio);
257 	blk_mq_free_request(req);
258 
259 	if (effects)
260 		nvme_passthru_end(ctrl, ns, effects, cmd, ret);
261 
262 	return ret;
263 }
264 
265 static int nvme_submit_io(struct nvme_ns *ns, struct nvme_user_io __user *uio)
266 {
267 	struct nvme_user_io io;
268 	struct nvme_command c;
269 	unsigned length, meta_len;
270 	void __user *metadata;
271 
272 	if (copy_from_user(&io, uio, sizeof(io)))
273 		return -EFAULT;
274 	if (io.flags)
275 		return -EINVAL;
276 
277 	switch (io.opcode) {
278 	case nvme_cmd_write:
279 	case nvme_cmd_read:
280 	case nvme_cmd_compare:
281 		break;
282 	default:
283 		return -EINVAL;
284 	}
285 
286 	length = (io.nblocks + 1) << ns->lba_shift;
287 
288 	if ((io.control & NVME_RW_PRINFO_PRACT) &&
289 	    ns->ms == sizeof(struct t10_pi_tuple)) {
290 		/*
291 		 * Protection information is stripped/inserted by the
292 		 * controller.
293 		 */
294 		if (nvme_to_user_ptr(io.metadata))
295 			return -EINVAL;
296 		meta_len = 0;
297 		metadata = NULL;
298 	} else {
299 		meta_len = (io.nblocks + 1) * ns->ms;
300 		metadata = nvme_to_user_ptr(io.metadata);
301 	}
302 
303 	if (ns->features & NVME_NS_EXT_LBAS) {
304 		length += meta_len;
305 		meta_len = 0;
306 	} else if (meta_len) {
307 		if ((io.metadata & 3) || !io.metadata)
308 			return -EINVAL;
309 	}
310 
311 	memset(&c, 0, sizeof(c));
312 	c.rw.opcode = io.opcode;
313 	c.rw.flags = io.flags;
314 	c.rw.nsid = cpu_to_le32(ns->head->ns_id);
315 	c.rw.slba = cpu_to_le64(io.slba);
316 	c.rw.length = cpu_to_le16(io.nblocks);
317 	c.rw.control = cpu_to_le16(io.control);
318 	c.rw.dsmgmt = cpu_to_le32(io.dsmgmt);
319 	c.rw.reftag = cpu_to_le32(io.reftag);
320 	c.rw.apptag = cpu_to_le16(io.apptag);
321 	c.rw.appmask = cpu_to_le16(io.appmask);
322 
323 	return nvme_submit_user_cmd(ns->queue, &c, io.addr, length, metadata,
324 			meta_len, lower_32_bits(io.slba), NULL, 0, 0);
325 }
326 
327 static bool nvme_validate_passthru_nsid(struct nvme_ctrl *ctrl,
328 					struct nvme_ns *ns, __u32 nsid)
329 {
330 	if (ns && nsid != ns->head->ns_id) {
331 		dev_err(ctrl->device,
332 			"%s: nsid (%u) in cmd does not match nsid (%u)"
333 			"of namespace\n",
334 			current->comm, nsid, ns->head->ns_id);
335 		return false;
336 	}
337 
338 	return true;
339 }
340 
341 static int nvme_user_cmd(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
342 		struct nvme_passthru_cmd __user *ucmd, unsigned int flags,
343 		bool open_for_write)
344 {
345 	struct nvme_passthru_cmd cmd;
346 	struct nvme_command c;
347 	unsigned timeout = 0;
348 	u64 result;
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, 0, 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, 0, &result, timeout, 0);
380 
381 	if (status >= 0) {
382 		if (put_user(result, &ucmd->result))
383 			return -EFAULT;
384 	}
385 
386 	return status;
387 }
388 
389 static int nvme_user_cmd64(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
390 		struct nvme_passthru_cmd64 __user *ucmd, unsigned int flags,
391 		bool open_for_write)
392 {
393 	struct nvme_passthru_cmd64 cmd;
394 	struct nvme_command c;
395 	unsigned timeout = 0;
396 	int status;
397 
398 	if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
399 		return -EFAULT;
400 	if (cmd.flags)
401 		return -EINVAL;
402 	if (!nvme_validate_passthru_nsid(ctrl, ns, cmd.nsid))
403 		return -EINVAL;
404 
405 	memset(&c, 0, sizeof(c));
406 	c.common.opcode = cmd.opcode;
407 	c.common.flags = cmd.flags;
408 	c.common.nsid = cpu_to_le32(cmd.nsid);
409 	c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
410 	c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
411 	c.common.cdw10 = cpu_to_le32(cmd.cdw10);
412 	c.common.cdw11 = cpu_to_le32(cmd.cdw11);
413 	c.common.cdw12 = cpu_to_le32(cmd.cdw12);
414 	c.common.cdw13 = cpu_to_le32(cmd.cdw13);
415 	c.common.cdw14 = cpu_to_le32(cmd.cdw14);
416 	c.common.cdw15 = cpu_to_le32(cmd.cdw15);
417 
418 	if (!nvme_cmd_allowed(ns, &c, flags, open_for_write))
419 		return -EACCES;
420 
421 	if (cmd.timeout_ms)
422 		timeout = msecs_to_jiffies(cmd.timeout_ms);
423 
424 	status = nvme_submit_user_cmd(ns ? ns->queue : ctrl->admin_q, &c,
425 			cmd.addr, cmd.data_len, nvme_to_user_ptr(cmd.metadata),
426 			cmd.metadata_len, 0, &cmd.result, timeout, flags);
427 
428 	if (status >= 0) {
429 		if (put_user(cmd.result, &ucmd->result))
430 			return -EFAULT;
431 	}
432 
433 	return status;
434 }
435 
436 struct nvme_uring_data {
437 	__u64	metadata;
438 	__u64	addr;
439 	__u32	data_len;
440 	__u32	metadata_len;
441 	__u32	timeout_ms;
442 };
443 
444 /*
445  * This overlays struct io_uring_cmd pdu.
446  * Expect build errors if this grows larger than that.
447  */
448 struct nvme_uring_cmd_pdu {
449 	union {
450 		struct bio *bio;
451 		struct request *req;
452 	};
453 	u32 meta_len;
454 	u32 nvme_status;
455 	union {
456 		struct {
457 			void *meta; /* kernel-resident buffer */
458 			void __user *meta_buffer;
459 		};
460 		u64 result;
461 	} u;
462 };
463 
464 static inline struct nvme_uring_cmd_pdu *nvme_uring_cmd_pdu(
465 		struct io_uring_cmd *ioucmd)
466 {
467 	return (struct nvme_uring_cmd_pdu *)&ioucmd->pdu;
468 }
469 
470 static void nvme_uring_task_meta_cb(struct io_uring_cmd *ioucmd,
471 				    unsigned issue_flags)
472 {
473 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
474 	struct request *req = pdu->req;
475 	int status;
476 	u64 result;
477 
478 	if (nvme_req(req)->flags & NVME_REQ_CANCELLED)
479 		status = -EINTR;
480 	else
481 		status = nvme_req(req)->status;
482 
483 	result = le64_to_cpu(nvme_req(req)->result.u64);
484 
485 	if (pdu->meta_len)
486 		status = nvme_finish_user_metadata(req, pdu->u.meta_buffer,
487 					pdu->u.meta, pdu->meta_len, status);
488 	if (req->bio)
489 		blk_rq_unmap_user(req->bio);
490 	blk_mq_free_request(req);
491 
492 	io_uring_cmd_done(ioucmd, status, result, issue_flags);
493 }
494 
495 static void nvme_uring_task_cb(struct io_uring_cmd *ioucmd,
496 			       unsigned issue_flags)
497 {
498 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
499 
500 	if (pdu->bio)
501 		blk_rq_unmap_user(pdu->bio);
502 
503 	io_uring_cmd_done(ioucmd, pdu->nvme_status, pdu->u.result, issue_flags);
504 }
505 
506 static enum rq_end_io_ret nvme_uring_cmd_end_io(struct request *req,
507 						blk_status_t err)
508 {
509 	struct io_uring_cmd *ioucmd = req->end_io_data;
510 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
511 
512 	req->bio = pdu->bio;
513 	if (nvme_req(req)->flags & NVME_REQ_CANCELLED)
514 		pdu->nvme_status = -EINTR;
515 	else
516 		pdu->nvme_status = nvme_req(req)->status;
517 	pdu->u.result = le64_to_cpu(nvme_req(req)->result.u64);
518 
519 	/*
520 	 * For iopoll, complete it directly.
521 	 * Otherwise, move the completion to task work.
522 	 */
523 	if (blk_rq_is_poll(req)) {
524 		WRITE_ONCE(ioucmd->cookie, NULL);
525 		nvme_uring_task_cb(ioucmd, IO_URING_F_UNLOCKED);
526 	} else {
527 		io_uring_cmd_do_in_task_lazy(ioucmd, nvme_uring_task_cb);
528 	}
529 
530 	return RQ_END_IO_FREE;
531 }
532 
533 static enum rq_end_io_ret nvme_uring_cmd_end_io_meta(struct request *req,
534 						     blk_status_t err)
535 {
536 	struct io_uring_cmd *ioucmd = req->end_io_data;
537 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
538 
539 	req->bio = pdu->bio;
540 	pdu->req = req;
541 
542 	/*
543 	 * For iopoll, complete it directly.
544 	 * Otherwise, move the completion to task work.
545 	 */
546 	if (blk_rq_is_poll(req)) {
547 		WRITE_ONCE(ioucmd->cookie, NULL);
548 		nvme_uring_task_meta_cb(ioucmd, IO_URING_F_UNLOCKED);
549 	} else {
550 		io_uring_cmd_do_in_task_lazy(ioucmd, nvme_uring_task_meta_cb);
551 	}
552 
553 	return RQ_END_IO_NONE;
554 }
555 
556 static int nvme_uring_cmd_io(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
557 		struct io_uring_cmd *ioucmd, unsigned int issue_flags, bool vec)
558 {
559 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
560 	const struct nvme_uring_cmd *cmd = io_uring_sqe_cmd(ioucmd->sqe);
561 	struct request_queue *q = ns ? ns->queue : ctrl->admin_q;
562 	struct nvme_uring_data d;
563 	struct nvme_command c;
564 	struct request *req;
565 	blk_opf_t rq_flags = REQ_ALLOC_CACHE;
566 	blk_mq_req_flags_t blk_flags = 0;
567 	void *meta = NULL;
568 	int ret;
569 
570 	c.common.opcode = READ_ONCE(cmd->opcode);
571 	c.common.flags = READ_ONCE(cmd->flags);
572 	if (c.common.flags)
573 		return -EINVAL;
574 
575 	c.common.command_id = 0;
576 	c.common.nsid = cpu_to_le32(cmd->nsid);
577 	if (!nvme_validate_passthru_nsid(ctrl, ns, le32_to_cpu(c.common.nsid)))
578 		return -EINVAL;
579 
580 	c.common.cdw2[0] = cpu_to_le32(READ_ONCE(cmd->cdw2));
581 	c.common.cdw2[1] = cpu_to_le32(READ_ONCE(cmd->cdw3));
582 	c.common.metadata = 0;
583 	c.common.dptr.prp1 = c.common.dptr.prp2 = 0;
584 	c.common.cdw10 = cpu_to_le32(READ_ONCE(cmd->cdw10));
585 	c.common.cdw11 = cpu_to_le32(READ_ONCE(cmd->cdw11));
586 	c.common.cdw12 = cpu_to_le32(READ_ONCE(cmd->cdw12));
587 	c.common.cdw13 = cpu_to_le32(READ_ONCE(cmd->cdw13));
588 	c.common.cdw14 = cpu_to_le32(READ_ONCE(cmd->cdw14));
589 	c.common.cdw15 = cpu_to_le32(READ_ONCE(cmd->cdw15));
590 
591 	if (!nvme_cmd_allowed(ns, &c, 0, ioucmd->file->f_mode & FMODE_WRITE))
592 		return -EACCES;
593 
594 	d.metadata = READ_ONCE(cmd->metadata);
595 	d.addr = READ_ONCE(cmd->addr);
596 	d.data_len = READ_ONCE(cmd->data_len);
597 	d.metadata_len = READ_ONCE(cmd->metadata_len);
598 	d.timeout_ms = READ_ONCE(cmd->timeout_ms);
599 
600 	if (issue_flags & IO_URING_F_NONBLOCK) {
601 		rq_flags |= REQ_NOWAIT;
602 		blk_flags = BLK_MQ_REQ_NOWAIT;
603 	}
604 	if (issue_flags & IO_URING_F_IOPOLL)
605 		rq_flags |= REQ_POLLED;
606 
607 	req = nvme_alloc_user_request(q, &c, rq_flags, blk_flags);
608 	if (IS_ERR(req))
609 		return PTR_ERR(req);
610 	req->timeout = d.timeout_ms ? msecs_to_jiffies(d.timeout_ms) : 0;
611 
612 	if (d.addr && d.data_len) {
613 		ret = nvme_map_user_request(req, d.addr,
614 			d.data_len, nvme_to_user_ptr(d.metadata),
615 			d.metadata_len, 0, &meta, ioucmd, vec);
616 		if (ret)
617 			return ret;
618 	}
619 
620 	if (blk_rq_is_poll(req)) {
621 		ioucmd->flags |= IORING_URING_CMD_POLLED;
622 		WRITE_ONCE(ioucmd->cookie, req);
623 	}
624 
625 	/* to free bio on completion, as req->bio will be null at that time */
626 	pdu->bio = req->bio;
627 	pdu->meta_len = d.metadata_len;
628 	req->end_io_data = ioucmd;
629 	if (pdu->meta_len) {
630 		pdu->u.meta = meta;
631 		pdu->u.meta_buffer = nvme_to_user_ptr(d.metadata);
632 		req->end_io = nvme_uring_cmd_end_io_meta;
633 	} else {
634 		req->end_io = nvme_uring_cmd_end_io;
635 	}
636 	blk_execute_rq_nowait(req, false);
637 	return -EIOCBQUEUED;
638 }
639 
640 static bool is_ctrl_ioctl(unsigned int cmd)
641 {
642 	if (cmd == NVME_IOCTL_ADMIN_CMD || cmd == NVME_IOCTL_ADMIN64_CMD)
643 		return true;
644 	if (is_sed_ioctl(cmd))
645 		return true;
646 	return false;
647 }
648 
649 static int nvme_ctrl_ioctl(struct nvme_ctrl *ctrl, unsigned int cmd,
650 		void __user *argp, bool open_for_write)
651 {
652 	switch (cmd) {
653 	case NVME_IOCTL_ADMIN_CMD:
654 		return nvme_user_cmd(ctrl, NULL, argp, 0, open_for_write);
655 	case NVME_IOCTL_ADMIN64_CMD:
656 		return nvme_user_cmd64(ctrl, NULL, argp, 0, open_for_write);
657 	default:
658 		return sed_ioctl(ctrl->opal_dev, cmd, argp);
659 	}
660 }
661 
662 #ifdef COMPAT_FOR_U64_ALIGNMENT
663 struct nvme_user_io32 {
664 	__u8	opcode;
665 	__u8	flags;
666 	__u16	control;
667 	__u16	nblocks;
668 	__u16	rsvd;
669 	__u64	metadata;
670 	__u64	addr;
671 	__u64	slba;
672 	__u32	dsmgmt;
673 	__u32	reftag;
674 	__u16	apptag;
675 	__u16	appmask;
676 } __attribute__((__packed__));
677 #define NVME_IOCTL_SUBMIT_IO32	_IOW('N', 0x42, struct nvme_user_io32)
678 #endif /* COMPAT_FOR_U64_ALIGNMENT */
679 
680 static int nvme_ns_ioctl(struct nvme_ns *ns, unsigned int cmd,
681 		void __user *argp, unsigned int flags, bool open_for_write)
682 {
683 	switch (cmd) {
684 	case NVME_IOCTL_ID:
685 		force_successful_syscall_return();
686 		return ns->head->ns_id;
687 	case NVME_IOCTL_IO_CMD:
688 		return nvme_user_cmd(ns->ctrl, ns, argp, flags, open_for_write);
689 	/*
690 	 * struct nvme_user_io can have different padding on some 32-bit ABIs.
691 	 * Just accept the compat version as all fields that are used are the
692 	 * same size and at the same offset.
693 	 */
694 #ifdef COMPAT_FOR_U64_ALIGNMENT
695 	case NVME_IOCTL_SUBMIT_IO32:
696 #endif
697 	case NVME_IOCTL_SUBMIT_IO:
698 		return nvme_submit_io(ns, argp);
699 	case NVME_IOCTL_IO64_CMD_VEC:
700 		flags |= NVME_IOCTL_VEC;
701 		fallthrough;
702 	case NVME_IOCTL_IO64_CMD:
703 		return nvme_user_cmd64(ns->ctrl, ns, argp, flags,
704 				       open_for_write);
705 	default:
706 		return -ENOTTY;
707 	}
708 }
709 
710 int nvme_ioctl(struct block_device *bdev, blk_mode_t mode,
711 		unsigned int cmd, unsigned long arg)
712 {
713 	struct nvme_ns *ns = bdev->bd_disk->private_data;
714 	bool open_for_write = mode & BLK_OPEN_WRITE;
715 	void __user *argp = (void __user *)arg;
716 	unsigned int flags = 0;
717 
718 	if (bdev_is_partition(bdev))
719 		flags |= NVME_IOCTL_PARTITION;
720 
721 	if (is_ctrl_ioctl(cmd))
722 		return nvme_ctrl_ioctl(ns->ctrl, cmd, argp, open_for_write);
723 	return nvme_ns_ioctl(ns, cmd, argp, flags, open_for_write);
724 }
725 
726 long nvme_ns_chr_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
727 {
728 	struct nvme_ns *ns =
729 		container_of(file_inode(file)->i_cdev, struct nvme_ns, cdev);
730 	bool open_for_write = file->f_mode & FMODE_WRITE;
731 	void __user *argp = (void __user *)arg;
732 
733 	if (is_ctrl_ioctl(cmd))
734 		return nvme_ctrl_ioctl(ns->ctrl, cmd, argp, open_for_write);
735 	return nvme_ns_ioctl(ns, cmd, argp, 0, open_for_write);
736 }
737 
738 static int nvme_uring_cmd_checks(unsigned int issue_flags)
739 {
740 
741 	/* NVMe passthrough requires big SQE/CQE support */
742 	if ((issue_flags & (IO_URING_F_SQE128|IO_URING_F_CQE32)) !=
743 	    (IO_URING_F_SQE128|IO_URING_F_CQE32))
744 		return -EOPNOTSUPP;
745 	return 0;
746 }
747 
748 static int nvme_ns_uring_cmd(struct nvme_ns *ns, struct io_uring_cmd *ioucmd,
749 			     unsigned int issue_flags)
750 {
751 	struct nvme_ctrl *ctrl = ns->ctrl;
752 	int ret;
753 
754 	BUILD_BUG_ON(sizeof(struct nvme_uring_cmd_pdu) > sizeof(ioucmd->pdu));
755 
756 	ret = nvme_uring_cmd_checks(issue_flags);
757 	if (ret)
758 		return ret;
759 
760 	switch (ioucmd->cmd_op) {
761 	case NVME_URING_CMD_IO:
762 		ret = nvme_uring_cmd_io(ctrl, ns, ioucmd, issue_flags, false);
763 		break;
764 	case NVME_URING_CMD_IO_VEC:
765 		ret = nvme_uring_cmd_io(ctrl, ns, ioucmd, issue_flags, true);
766 		break;
767 	default:
768 		ret = -ENOTTY;
769 	}
770 
771 	return ret;
772 }
773 
774 int nvme_ns_chr_uring_cmd(struct io_uring_cmd *ioucmd, unsigned int issue_flags)
775 {
776 	struct nvme_ns *ns = container_of(file_inode(ioucmd->file)->i_cdev,
777 			struct nvme_ns, cdev);
778 
779 	return nvme_ns_uring_cmd(ns, ioucmd, issue_flags);
780 }
781 
782 int nvme_ns_chr_uring_cmd_iopoll(struct io_uring_cmd *ioucmd,
783 				 struct io_comp_batch *iob,
784 				 unsigned int poll_flags)
785 {
786 	struct request *req;
787 	int ret = 0;
788 
789 	if (!(ioucmd->flags & IORING_URING_CMD_POLLED))
790 		return 0;
791 
792 	req = READ_ONCE(ioucmd->cookie);
793 	if (req && blk_rq_is_poll(req))
794 		ret = blk_rq_poll(req, iob, poll_flags);
795 	return ret;
796 }
797 #ifdef CONFIG_NVME_MULTIPATH
798 static int nvme_ns_head_ctrl_ioctl(struct nvme_ns *ns, unsigned int cmd,
799 		void __user *argp, struct nvme_ns_head *head, int srcu_idx,
800 		bool open_for_write)
801 	__releases(&head->srcu)
802 {
803 	struct nvme_ctrl *ctrl = ns->ctrl;
804 	int ret;
805 
806 	nvme_get_ctrl(ns->ctrl);
807 	srcu_read_unlock(&head->srcu, srcu_idx);
808 	ret = nvme_ctrl_ioctl(ns->ctrl, cmd, argp, open_for_write);
809 
810 	nvme_put_ctrl(ctrl);
811 	return ret;
812 }
813 
814 int nvme_ns_head_ioctl(struct block_device *bdev, blk_mode_t mode,
815 		unsigned int cmd, unsigned long arg)
816 {
817 	struct nvme_ns_head *head = bdev->bd_disk->private_data;
818 	bool open_for_write = mode & BLK_OPEN_WRITE;
819 	void __user *argp = (void __user *)arg;
820 	struct nvme_ns *ns;
821 	int srcu_idx, ret = -EWOULDBLOCK;
822 	unsigned int flags = 0;
823 
824 	if (bdev_is_partition(bdev))
825 		flags |= NVME_IOCTL_PARTITION;
826 
827 	srcu_idx = srcu_read_lock(&head->srcu);
828 	ns = nvme_find_path(head);
829 	if (!ns)
830 		goto out_unlock;
831 
832 	/*
833 	 * Handle ioctls that apply to the controller instead of the namespace
834 	 * seperately and drop the ns SRCU reference early.  This avoids a
835 	 * deadlock when deleting namespaces using the passthrough interface.
836 	 */
837 	if (is_ctrl_ioctl(cmd))
838 		return nvme_ns_head_ctrl_ioctl(ns, cmd, argp, head, srcu_idx,
839 					       open_for_write);
840 
841 	ret = nvme_ns_ioctl(ns, cmd, argp, flags, open_for_write);
842 out_unlock:
843 	srcu_read_unlock(&head->srcu, srcu_idx);
844 	return ret;
845 }
846 
847 long nvme_ns_head_chr_ioctl(struct file *file, unsigned int cmd,
848 		unsigned long arg)
849 {
850 	bool open_for_write = file->f_mode & FMODE_WRITE;
851 	struct cdev *cdev = file_inode(file)->i_cdev;
852 	struct nvme_ns_head *head =
853 		container_of(cdev, struct nvme_ns_head, cdev);
854 	void __user *argp = (void __user *)arg;
855 	struct nvme_ns *ns;
856 	int srcu_idx, ret = -EWOULDBLOCK;
857 
858 	srcu_idx = srcu_read_lock(&head->srcu);
859 	ns = nvme_find_path(head);
860 	if (!ns)
861 		goto out_unlock;
862 
863 	if (is_ctrl_ioctl(cmd))
864 		return nvme_ns_head_ctrl_ioctl(ns, cmd, argp, head, srcu_idx,
865 				open_for_write);
866 
867 	ret = nvme_ns_ioctl(ns, cmd, argp, 0, open_for_write);
868 out_unlock:
869 	srcu_read_unlock(&head->srcu, srcu_idx);
870 	return ret;
871 }
872 
873 int nvme_ns_head_chr_uring_cmd(struct io_uring_cmd *ioucmd,
874 		unsigned int issue_flags)
875 {
876 	struct cdev *cdev = file_inode(ioucmd->file)->i_cdev;
877 	struct nvme_ns_head *head = container_of(cdev, struct nvme_ns_head, cdev);
878 	int srcu_idx = srcu_read_lock(&head->srcu);
879 	struct nvme_ns *ns = nvme_find_path(head);
880 	int ret = -EINVAL;
881 
882 	if (ns)
883 		ret = nvme_ns_uring_cmd(ns, ioucmd, issue_flags);
884 	srcu_read_unlock(&head->srcu, srcu_idx);
885 	return ret;
886 }
887 #endif /* CONFIG_NVME_MULTIPATH */
888 
889 int nvme_dev_uring_cmd(struct io_uring_cmd *ioucmd, unsigned int issue_flags)
890 {
891 	struct nvme_ctrl *ctrl = ioucmd->file->private_data;
892 	int ret;
893 
894 	/* IOPOLL not supported yet */
895 	if (issue_flags & IO_URING_F_IOPOLL)
896 		return -EOPNOTSUPP;
897 
898 	ret = nvme_uring_cmd_checks(issue_flags);
899 	if (ret)
900 		return ret;
901 
902 	switch (ioucmd->cmd_op) {
903 	case NVME_URING_CMD_ADMIN:
904 		ret = nvme_uring_cmd_io(ctrl, NULL, ioucmd, issue_flags, false);
905 		break;
906 	case NVME_URING_CMD_ADMIN_VEC:
907 		ret = nvme_uring_cmd_io(ctrl, NULL, ioucmd, issue_flags, true);
908 		break;
909 	default:
910 		ret = -ENOTTY;
911 	}
912 
913 	return ret;
914 }
915 
916 static int nvme_dev_user_cmd(struct nvme_ctrl *ctrl, void __user *argp,
917 		bool open_for_write)
918 {
919 	struct nvme_ns *ns;
920 	int ret;
921 
922 	down_read(&ctrl->namespaces_rwsem);
923 	if (list_empty(&ctrl->namespaces)) {
924 		ret = -ENOTTY;
925 		goto out_unlock;
926 	}
927 
928 	ns = list_first_entry(&ctrl->namespaces, struct nvme_ns, list);
929 	if (ns != list_last_entry(&ctrl->namespaces, struct nvme_ns, list)) {
930 		dev_warn(ctrl->device,
931 			"NVME_IOCTL_IO_CMD not supported when multiple namespaces present!\n");
932 		ret = -EINVAL;
933 		goto out_unlock;
934 	}
935 
936 	dev_warn(ctrl->device,
937 		"using deprecated NVME_IOCTL_IO_CMD ioctl on the char device!\n");
938 	kref_get(&ns->kref);
939 	up_read(&ctrl->namespaces_rwsem);
940 
941 	ret = nvme_user_cmd(ctrl, ns, argp, 0, open_for_write);
942 	nvme_put_ns(ns);
943 	return ret;
944 
945 out_unlock:
946 	up_read(&ctrl->namespaces_rwsem);
947 	return ret;
948 }
949 
950 long nvme_dev_ioctl(struct file *file, unsigned int cmd,
951 		unsigned long arg)
952 {
953 	bool open_for_write = file->f_mode & FMODE_WRITE;
954 	struct nvme_ctrl *ctrl = file->private_data;
955 	void __user *argp = (void __user *)arg;
956 
957 	switch (cmd) {
958 	case NVME_IOCTL_ADMIN_CMD:
959 		return nvme_user_cmd(ctrl, NULL, argp, 0, open_for_write);
960 	case NVME_IOCTL_ADMIN64_CMD:
961 		return nvme_user_cmd64(ctrl, NULL, argp, 0, open_for_write);
962 	case NVME_IOCTL_IO_CMD:
963 		return nvme_dev_user_cmd(ctrl, argp, open_for_write);
964 	case NVME_IOCTL_RESET:
965 		if (!capable(CAP_SYS_ADMIN))
966 			return -EACCES;
967 		dev_warn(ctrl->device, "resetting controller\n");
968 		return nvme_reset_ctrl_sync(ctrl);
969 	case NVME_IOCTL_SUBSYS_RESET:
970 		if (!capable(CAP_SYS_ADMIN))
971 			return -EACCES;
972 		return nvme_reset_subsystem(ctrl);
973 	case NVME_IOCTL_RESCAN:
974 		if (!capable(CAP_SYS_ADMIN))
975 			return -EACCES;
976 		nvme_queue_scan(ctrl);
977 		return 0;
978 	default:
979 		return -ENOTTY;
980 	}
981 }
982