xref: /openbmc/linux/drivers/nvme/host/ioctl.c (revision 313c08c7)
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, fmode_t mode)
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 mode & FMODE_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 	ret = -EFAULT;
112 	if ((req_op(req) == REQ_OP_DRV_OUT) && copy_from_user(buf, ubuf, len))
113 		goto out_free_meta;
114 
115 	bip = bio_integrity_alloc(bio, GFP_KERNEL, 1);
116 	if (IS_ERR(bip)) {
117 		ret = PTR_ERR(bip);
118 		goto out_free_meta;
119 	}
120 
121 	bip->bip_iter.bi_size = len;
122 	bip->bip_iter.bi_sector = seed;
123 	ret = bio_integrity_add_page(bio, virt_to_page(buf), len,
124 			offset_in_page(buf));
125 	if (ret != len) {
126 		ret = -ENOMEM;
127 		goto out_free_meta;
128 	}
129 
130 	req->cmd_flags |= REQ_INTEGRITY;
131 	return buf;
132 out_free_meta:
133 	kfree(buf);
134 out:
135 	return ERR_PTR(ret);
136 }
137 
138 static int nvme_finish_user_metadata(struct request *req, void __user *ubuf,
139 		void *meta, unsigned len, int ret)
140 {
141 	if (!ret && req_op(req) == REQ_OP_DRV_IN &&
142 	    copy_to_user(ubuf, meta, len))
143 		ret = -EFAULT;
144 	kfree(meta);
145 	return ret;
146 }
147 
148 static struct request *nvme_alloc_user_request(struct request_queue *q,
149 		struct nvme_command *cmd, blk_opf_t rq_flags,
150 		blk_mq_req_flags_t blk_flags)
151 {
152 	struct request *req;
153 
154 	req = blk_mq_alloc_request(q, nvme_req_op(cmd) | rq_flags, blk_flags);
155 	if (IS_ERR(req))
156 		return req;
157 	nvme_init_request(req, cmd);
158 	nvme_req(req)->flags |= NVME_REQ_USERCMD;
159 	return req;
160 }
161 
162 static int nvme_map_user_request(struct request *req, u64 ubuffer,
163 		unsigned bufflen, void __user *meta_buffer, unsigned meta_len,
164 		u32 meta_seed, void **metap, struct io_uring_cmd *ioucmd,
165 		unsigned int flags)
166 {
167 	struct request_queue *q = req->q;
168 	struct nvme_ns *ns = q->queuedata;
169 	struct block_device *bdev = ns ? ns->disk->part0 : NULL;
170 	struct bio *bio = NULL;
171 	void *meta = NULL;
172 	int ret;
173 
174 	if (ioucmd && (ioucmd->flags & IORING_URING_CMD_FIXED)) {
175 		struct iov_iter iter;
176 
177 		/* fixedbufs is only for non-vectored io */
178 		if (WARN_ON_ONCE(flags & NVME_IOCTL_VEC))
179 			return -EINVAL;
180 		ret = io_uring_cmd_import_fixed(ubuffer, bufflen,
181 				rq_data_dir(req), &iter, ioucmd);
182 		if (ret < 0)
183 			goto out;
184 		ret = blk_rq_map_user_iov(q, req, NULL, &iter, GFP_KERNEL);
185 	} else {
186 		ret = blk_rq_map_user_io(req, NULL, nvme_to_user_ptr(ubuffer),
187 				bufflen, GFP_KERNEL, flags & NVME_IOCTL_VEC, 0,
188 				0, rq_data_dir(req));
189 	}
190 
191 	if (ret)
192 		goto out;
193 	bio = req->bio;
194 	if (bdev)
195 		bio_set_dev(bio, bdev);
196 
197 	if (bdev && meta_buffer && meta_len) {
198 		meta = nvme_add_user_metadata(req, meta_buffer, meta_len,
199 				meta_seed);
200 		if (IS_ERR(meta)) {
201 			ret = PTR_ERR(meta);
202 			goto out_unmap;
203 		}
204 		*metap = meta;
205 	}
206 
207 	return ret;
208 
209 out_unmap:
210 	if (bio)
211 		blk_rq_unmap_user(bio);
212 out:
213 	blk_mq_free_request(req);
214 	return ret;
215 }
216 
217 static int nvme_submit_user_cmd(struct request_queue *q,
218 		struct nvme_command *cmd, u64 ubuffer, unsigned bufflen,
219 		void __user *meta_buffer, unsigned meta_len, u32 meta_seed,
220 		u64 *result, unsigned timeout, unsigned int flags)
221 {
222 	struct nvme_ctrl *ctrl;
223 	struct request *req;
224 	void *meta = NULL;
225 	struct bio *bio;
226 	u32 effects;
227 	int ret;
228 
229 	req = nvme_alloc_user_request(q, cmd, 0, 0);
230 	if (IS_ERR(req))
231 		return PTR_ERR(req);
232 
233 	req->timeout = timeout;
234 	if (ubuffer && bufflen) {
235 		ret = nvme_map_user_request(req, ubuffer, bufflen, meta_buffer,
236 				meta_len, meta_seed, &meta, NULL, flags);
237 		if (ret)
238 			return ret;
239 	}
240 
241 	bio = req->bio;
242 	ctrl = nvme_req(req)->ctrl;
243 
244 	ret = nvme_execute_passthru_rq(req, &effects);
245 
246 	if (result)
247 		*result = le64_to_cpu(nvme_req(req)->result.u64);
248 	if (meta)
249 		ret = nvme_finish_user_metadata(req, meta_buffer, meta,
250 						meta_len, ret);
251 	if (bio)
252 		blk_rq_unmap_user(bio);
253 	blk_mq_free_request(req);
254 
255 	if (effects)
256 		nvme_passthru_end(ctrl, effects, cmd, ret);
257 
258 	return ret;
259 }
260 
261 static int nvme_submit_io(struct nvme_ns *ns, struct nvme_user_io __user *uio)
262 {
263 	struct nvme_user_io io;
264 	struct nvme_command c;
265 	unsigned length, meta_len;
266 	void __user *metadata;
267 
268 	if (copy_from_user(&io, uio, sizeof(io)))
269 		return -EFAULT;
270 	if (io.flags)
271 		return -EINVAL;
272 
273 	switch (io.opcode) {
274 	case nvme_cmd_write:
275 	case nvme_cmd_read:
276 	case nvme_cmd_compare:
277 		break;
278 	default:
279 		return -EINVAL;
280 	}
281 
282 	length = (io.nblocks + 1) << ns->lba_shift;
283 
284 	if ((io.control & NVME_RW_PRINFO_PRACT) &&
285 	    ns->ms == sizeof(struct t10_pi_tuple)) {
286 		/*
287 		 * Protection information is stripped/inserted by the
288 		 * controller.
289 		 */
290 		if (nvme_to_user_ptr(io.metadata))
291 			return -EINVAL;
292 		meta_len = 0;
293 		metadata = NULL;
294 	} else {
295 		meta_len = (io.nblocks + 1) * ns->ms;
296 		metadata = nvme_to_user_ptr(io.metadata);
297 	}
298 
299 	if (ns->features & NVME_NS_EXT_LBAS) {
300 		length += meta_len;
301 		meta_len = 0;
302 	} else if (meta_len) {
303 		if ((io.metadata & 3) || !io.metadata)
304 			return -EINVAL;
305 	}
306 
307 	memset(&c, 0, sizeof(c));
308 	c.rw.opcode = io.opcode;
309 	c.rw.flags = io.flags;
310 	c.rw.nsid = cpu_to_le32(ns->head->ns_id);
311 	c.rw.slba = cpu_to_le64(io.slba);
312 	c.rw.length = cpu_to_le16(io.nblocks);
313 	c.rw.control = cpu_to_le16(io.control);
314 	c.rw.dsmgmt = cpu_to_le32(io.dsmgmt);
315 	c.rw.reftag = cpu_to_le32(io.reftag);
316 	c.rw.apptag = cpu_to_le16(io.apptag);
317 	c.rw.appmask = cpu_to_le16(io.appmask);
318 
319 	return nvme_submit_user_cmd(ns->queue, &c, io.addr, length, metadata,
320 			meta_len, lower_32_bits(io.slba), NULL, 0, 0);
321 }
322 
323 static bool nvme_validate_passthru_nsid(struct nvme_ctrl *ctrl,
324 					struct nvme_ns *ns, __u32 nsid)
325 {
326 	if (ns && nsid != ns->head->ns_id) {
327 		dev_err(ctrl->device,
328 			"%s: nsid (%u) in cmd does not match nsid (%u)"
329 			"of namespace\n",
330 			current->comm, nsid, ns->head->ns_id);
331 		return false;
332 	}
333 
334 	return true;
335 }
336 
337 static int nvme_user_cmd(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
338 		struct nvme_passthru_cmd __user *ucmd, unsigned int flags,
339 		fmode_t mode)
340 {
341 	struct nvme_passthru_cmd cmd;
342 	struct nvme_command c;
343 	unsigned timeout = 0;
344 	u64 result;
345 	int status;
346 
347 	if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
348 		return -EFAULT;
349 	if (cmd.flags)
350 		return -EINVAL;
351 	if (!nvme_validate_passthru_nsid(ctrl, ns, cmd.nsid))
352 		return -EINVAL;
353 
354 	memset(&c, 0, sizeof(c));
355 	c.common.opcode = cmd.opcode;
356 	c.common.flags = cmd.flags;
357 	c.common.nsid = cpu_to_le32(cmd.nsid);
358 	c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
359 	c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
360 	c.common.cdw10 = cpu_to_le32(cmd.cdw10);
361 	c.common.cdw11 = cpu_to_le32(cmd.cdw11);
362 	c.common.cdw12 = cpu_to_le32(cmd.cdw12);
363 	c.common.cdw13 = cpu_to_le32(cmd.cdw13);
364 	c.common.cdw14 = cpu_to_le32(cmd.cdw14);
365 	c.common.cdw15 = cpu_to_le32(cmd.cdw15);
366 
367 	if (!nvme_cmd_allowed(ns, &c, 0, mode))
368 		return -EACCES;
369 
370 	if (cmd.timeout_ms)
371 		timeout = msecs_to_jiffies(cmd.timeout_ms);
372 
373 	status = nvme_submit_user_cmd(ns ? ns->queue : ctrl->admin_q, &c,
374 			cmd.addr, cmd.data_len, nvme_to_user_ptr(cmd.metadata),
375 			cmd.metadata_len, 0, &result, timeout, 0);
376 
377 	if (status >= 0) {
378 		if (put_user(result, &ucmd->result))
379 			return -EFAULT;
380 	}
381 
382 	return status;
383 }
384 
385 static int nvme_user_cmd64(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
386 		struct nvme_passthru_cmd64 __user *ucmd, unsigned int flags,
387 		fmode_t mode)
388 {
389 	struct nvme_passthru_cmd64 cmd;
390 	struct nvme_command c;
391 	unsigned timeout = 0;
392 	int status;
393 
394 	if (copy_from_user(&cmd, ucmd, sizeof(cmd)))
395 		return -EFAULT;
396 	if (cmd.flags)
397 		return -EINVAL;
398 	if (!nvme_validate_passthru_nsid(ctrl, ns, cmd.nsid))
399 		return -EINVAL;
400 
401 	memset(&c, 0, sizeof(c));
402 	c.common.opcode = cmd.opcode;
403 	c.common.flags = cmd.flags;
404 	c.common.nsid = cpu_to_le32(cmd.nsid);
405 	c.common.cdw2[0] = cpu_to_le32(cmd.cdw2);
406 	c.common.cdw2[1] = cpu_to_le32(cmd.cdw3);
407 	c.common.cdw10 = cpu_to_le32(cmd.cdw10);
408 	c.common.cdw11 = cpu_to_le32(cmd.cdw11);
409 	c.common.cdw12 = cpu_to_le32(cmd.cdw12);
410 	c.common.cdw13 = cpu_to_le32(cmd.cdw13);
411 	c.common.cdw14 = cpu_to_le32(cmd.cdw14);
412 	c.common.cdw15 = cpu_to_le32(cmd.cdw15);
413 
414 	if (!nvme_cmd_allowed(ns, &c, flags, mode))
415 		return -EACCES;
416 
417 	if (cmd.timeout_ms)
418 		timeout = msecs_to_jiffies(cmd.timeout_ms);
419 
420 	status = nvme_submit_user_cmd(ns ? ns->queue : ctrl->admin_q, &c,
421 			cmd.addr, cmd.data_len, nvme_to_user_ptr(cmd.metadata),
422 			cmd.metadata_len, 0, &cmd.result, timeout, flags);
423 
424 	if (status >= 0) {
425 		if (put_user(cmd.result, &ucmd->result))
426 			return -EFAULT;
427 	}
428 
429 	return status;
430 }
431 
432 struct nvme_uring_data {
433 	__u64	metadata;
434 	__u64	addr;
435 	__u32	data_len;
436 	__u32	metadata_len;
437 	__u32	timeout_ms;
438 };
439 
440 /*
441  * This overlays struct io_uring_cmd pdu.
442  * Expect build errors if this grows larger than that.
443  */
444 struct nvme_uring_cmd_pdu {
445 	union {
446 		struct bio *bio;
447 		struct request *req;
448 	};
449 	u32 meta_len;
450 	u32 nvme_status;
451 	union {
452 		struct {
453 			void *meta; /* kernel-resident buffer */
454 			void __user *meta_buffer;
455 		};
456 		u64 result;
457 	} u;
458 };
459 
460 static inline struct nvme_uring_cmd_pdu *nvme_uring_cmd_pdu(
461 		struct io_uring_cmd *ioucmd)
462 {
463 	return (struct nvme_uring_cmd_pdu *)&ioucmd->pdu;
464 }
465 
466 static void nvme_uring_task_meta_cb(struct io_uring_cmd *ioucmd)
467 {
468 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
469 	struct request *req = pdu->req;
470 	int status;
471 	u64 result;
472 
473 	if (nvme_req(req)->flags & NVME_REQ_CANCELLED)
474 		status = -EINTR;
475 	else
476 		status = nvme_req(req)->status;
477 
478 	result = le64_to_cpu(nvme_req(req)->result.u64);
479 
480 	if (pdu->meta_len)
481 		status = nvme_finish_user_metadata(req, pdu->u.meta_buffer,
482 					pdu->u.meta, pdu->meta_len, status);
483 	if (req->bio)
484 		blk_rq_unmap_user(req->bio);
485 	blk_mq_free_request(req);
486 
487 	io_uring_cmd_done(ioucmd, status, result);
488 }
489 
490 static void nvme_uring_task_cb(struct io_uring_cmd *ioucmd)
491 {
492 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
493 
494 	if (pdu->bio)
495 		blk_rq_unmap_user(pdu->bio);
496 
497 	io_uring_cmd_done(ioucmd, pdu->nvme_status, pdu->u.result);
498 }
499 
500 static enum rq_end_io_ret nvme_uring_cmd_end_io(struct request *req,
501 						blk_status_t err)
502 {
503 	struct io_uring_cmd *ioucmd = req->end_io_data;
504 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
505 	void *cookie = READ_ONCE(ioucmd->cookie);
506 
507 	req->bio = pdu->bio;
508 	if (nvme_req(req)->flags & NVME_REQ_CANCELLED)
509 		pdu->nvme_status = -EINTR;
510 	else
511 		pdu->nvme_status = nvme_req(req)->status;
512 	pdu->u.result = le64_to_cpu(nvme_req(req)->result.u64);
513 
514 	/*
515 	 * For iopoll, complete it directly.
516 	 * Otherwise, move the completion to task work.
517 	 */
518 	if (cookie != NULL && blk_rq_is_poll(req))
519 		nvme_uring_task_cb(ioucmd);
520 	else
521 		io_uring_cmd_complete_in_task(ioucmd, nvme_uring_task_cb);
522 
523 	return RQ_END_IO_FREE;
524 }
525 
526 static enum rq_end_io_ret nvme_uring_cmd_end_io_meta(struct request *req,
527 						     blk_status_t err)
528 {
529 	struct io_uring_cmd *ioucmd = req->end_io_data;
530 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
531 	void *cookie = READ_ONCE(ioucmd->cookie);
532 
533 	req->bio = pdu->bio;
534 	pdu->req = req;
535 
536 	/*
537 	 * For iopoll, complete it directly.
538 	 * Otherwise, move the completion to task work.
539 	 */
540 	if (cookie != NULL && blk_rq_is_poll(req))
541 		nvme_uring_task_meta_cb(ioucmd);
542 	else
543 		io_uring_cmd_complete_in_task(ioucmd, nvme_uring_task_meta_cb);
544 
545 	return RQ_END_IO_NONE;
546 }
547 
548 static int nvme_uring_cmd_io(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
549 		struct io_uring_cmd *ioucmd, unsigned int issue_flags, bool vec)
550 {
551 	struct nvme_uring_cmd_pdu *pdu = nvme_uring_cmd_pdu(ioucmd);
552 	const struct nvme_uring_cmd *cmd = ioucmd->cmd;
553 	struct request_queue *q = ns ? ns->queue : ctrl->admin_q;
554 	struct nvme_uring_data d;
555 	struct nvme_command c;
556 	struct request *req;
557 	blk_opf_t rq_flags = 0;
558 	blk_mq_req_flags_t blk_flags = 0;
559 	void *meta = NULL;
560 	int ret;
561 
562 	c.common.opcode = READ_ONCE(cmd->opcode);
563 	c.common.flags = READ_ONCE(cmd->flags);
564 	if (c.common.flags)
565 		return -EINVAL;
566 
567 	c.common.command_id = 0;
568 	c.common.nsid = cpu_to_le32(cmd->nsid);
569 	if (!nvme_validate_passthru_nsid(ctrl, ns, le32_to_cpu(c.common.nsid)))
570 		return -EINVAL;
571 
572 	c.common.cdw2[0] = cpu_to_le32(READ_ONCE(cmd->cdw2));
573 	c.common.cdw2[1] = cpu_to_le32(READ_ONCE(cmd->cdw3));
574 	c.common.metadata = 0;
575 	c.common.dptr.prp1 = c.common.dptr.prp2 = 0;
576 	c.common.cdw10 = cpu_to_le32(READ_ONCE(cmd->cdw10));
577 	c.common.cdw11 = cpu_to_le32(READ_ONCE(cmd->cdw11));
578 	c.common.cdw12 = cpu_to_le32(READ_ONCE(cmd->cdw12));
579 	c.common.cdw13 = cpu_to_le32(READ_ONCE(cmd->cdw13));
580 	c.common.cdw14 = cpu_to_le32(READ_ONCE(cmd->cdw14));
581 	c.common.cdw15 = cpu_to_le32(READ_ONCE(cmd->cdw15));
582 
583 	if (!nvme_cmd_allowed(ns, &c, 0, ioucmd->file->f_mode))
584 		return -EACCES;
585 
586 	d.metadata = READ_ONCE(cmd->metadata);
587 	d.addr = READ_ONCE(cmd->addr);
588 	d.data_len = READ_ONCE(cmd->data_len);
589 	d.metadata_len = READ_ONCE(cmd->metadata_len);
590 	d.timeout_ms = READ_ONCE(cmd->timeout_ms);
591 
592 	if (issue_flags & IO_URING_F_NONBLOCK) {
593 		rq_flags = REQ_NOWAIT;
594 		blk_flags = BLK_MQ_REQ_NOWAIT;
595 	}
596 	if (issue_flags & IO_URING_F_IOPOLL)
597 		rq_flags |= REQ_POLLED;
598 
599 retry:
600 	req = nvme_alloc_user_request(q, &c, rq_flags, blk_flags);
601 	if (IS_ERR(req))
602 		return PTR_ERR(req);
603 	req->timeout = d.timeout_ms ? msecs_to_jiffies(d.timeout_ms) : 0;
604 
605 	if (d.addr && d.data_len) {
606 		ret = nvme_map_user_request(req, d.addr,
607 			d.data_len, nvme_to_user_ptr(d.metadata),
608 			d.metadata_len, 0, &meta, ioucmd, vec);
609 		if (ret)
610 			return ret;
611 	}
612 
613 	if (issue_flags & IO_URING_F_IOPOLL && rq_flags & REQ_POLLED) {
614 		if (unlikely(!req->bio)) {
615 			/* we can't poll this, so alloc regular req instead */
616 			blk_mq_free_request(req);
617 			rq_flags &= ~REQ_POLLED;
618 			goto retry;
619 		} else {
620 			WRITE_ONCE(ioucmd->cookie, req->bio);
621 			req->bio->bi_opf |= REQ_POLLED;
622 		}
623 	}
624 	/* to free bio on completion, as req->bio will be null at that time */
625 	pdu->bio = req->bio;
626 	pdu->meta_len = d.metadata_len;
627 	req->end_io_data = ioucmd;
628 	if (pdu->meta_len) {
629 		pdu->u.meta = meta;
630 		pdu->u.meta_buffer = nvme_to_user_ptr(d.metadata);
631 		req->end_io = nvme_uring_cmd_end_io_meta;
632 	} else {
633 		req->end_io = nvme_uring_cmd_end_io;
634 	}
635 	blk_execute_rq_nowait(req, false);
636 	return -EIOCBQUEUED;
637 }
638 
639 static bool is_ctrl_ioctl(unsigned int cmd)
640 {
641 	if (cmd == NVME_IOCTL_ADMIN_CMD || cmd == NVME_IOCTL_ADMIN64_CMD)
642 		return true;
643 	if (is_sed_ioctl(cmd))
644 		return true;
645 	return false;
646 }
647 
648 static int nvme_ctrl_ioctl(struct nvme_ctrl *ctrl, unsigned int cmd,
649 		void __user *argp, fmode_t mode)
650 {
651 	switch (cmd) {
652 	case NVME_IOCTL_ADMIN_CMD:
653 		return nvme_user_cmd(ctrl, NULL, argp, 0, mode);
654 	case NVME_IOCTL_ADMIN64_CMD:
655 		return nvme_user_cmd64(ctrl, NULL, argp, 0, mode);
656 	default:
657 		return sed_ioctl(ctrl->opal_dev, cmd, argp);
658 	}
659 }
660 
661 #ifdef COMPAT_FOR_U64_ALIGNMENT
662 struct nvme_user_io32 {
663 	__u8	opcode;
664 	__u8	flags;
665 	__u16	control;
666 	__u16	nblocks;
667 	__u16	rsvd;
668 	__u64	metadata;
669 	__u64	addr;
670 	__u64	slba;
671 	__u32	dsmgmt;
672 	__u32	reftag;
673 	__u16	apptag;
674 	__u16	appmask;
675 } __attribute__((__packed__));
676 #define NVME_IOCTL_SUBMIT_IO32	_IOW('N', 0x42, struct nvme_user_io32)
677 #endif /* COMPAT_FOR_U64_ALIGNMENT */
678 
679 static int nvme_ns_ioctl(struct nvme_ns *ns, unsigned int cmd,
680 		void __user *argp, unsigned int flags, fmode_t mode)
681 {
682 	switch (cmd) {
683 	case NVME_IOCTL_ID:
684 		force_successful_syscall_return();
685 		return ns->head->ns_id;
686 	case NVME_IOCTL_IO_CMD:
687 		return nvme_user_cmd(ns->ctrl, ns, argp, flags, mode);
688 	/*
689 	 * struct nvme_user_io can have different padding on some 32-bit ABIs.
690 	 * Just accept the compat version as all fields that are used are the
691 	 * same size and at the same offset.
692 	 */
693 #ifdef COMPAT_FOR_U64_ALIGNMENT
694 	case NVME_IOCTL_SUBMIT_IO32:
695 #endif
696 	case NVME_IOCTL_SUBMIT_IO:
697 		return nvme_submit_io(ns, argp);
698 	case NVME_IOCTL_IO64_CMD_VEC:
699 		flags |= NVME_IOCTL_VEC;
700 		fallthrough;
701 	case NVME_IOCTL_IO64_CMD:
702 		return nvme_user_cmd64(ns->ctrl, ns, argp, flags, mode);
703 	default:
704 		return -ENOTTY;
705 	}
706 }
707 
708 int nvme_ioctl(struct block_device *bdev, fmode_t mode,
709 		unsigned int cmd, unsigned long arg)
710 {
711 	struct nvme_ns *ns = bdev->bd_disk->private_data;
712 	void __user *argp = (void __user *)arg;
713 	unsigned int flags = 0;
714 
715 	if (bdev_is_partition(bdev))
716 		flags |= NVME_IOCTL_PARTITION;
717 
718 	if (is_ctrl_ioctl(cmd))
719 		return nvme_ctrl_ioctl(ns->ctrl, cmd, argp, mode);
720 	return nvme_ns_ioctl(ns, cmd, argp, flags, mode);
721 }
722 
723 long nvme_ns_chr_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
724 {
725 	struct nvme_ns *ns =
726 		container_of(file_inode(file)->i_cdev, struct nvme_ns, cdev);
727 	void __user *argp = (void __user *)arg;
728 
729 	if (is_ctrl_ioctl(cmd))
730 		return nvme_ctrl_ioctl(ns->ctrl, cmd, argp, file->f_mode);
731 	return nvme_ns_ioctl(ns, cmd, argp, 0, file->f_mode);
732 }
733 
734 static int nvme_uring_cmd_checks(unsigned int issue_flags)
735 {
736 
737 	/* NVMe passthrough requires big SQE/CQE support */
738 	if ((issue_flags & (IO_URING_F_SQE128|IO_URING_F_CQE32)) !=
739 	    (IO_URING_F_SQE128|IO_URING_F_CQE32))
740 		return -EOPNOTSUPP;
741 	return 0;
742 }
743 
744 static int nvme_ns_uring_cmd(struct nvme_ns *ns, struct io_uring_cmd *ioucmd,
745 			     unsigned int issue_flags)
746 {
747 	struct nvme_ctrl *ctrl = ns->ctrl;
748 	int ret;
749 
750 	BUILD_BUG_ON(sizeof(struct nvme_uring_cmd_pdu) > sizeof(ioucmd->pdu));
751 
752 	ret = nvme_uring_cmd_checks(issue_flags);
753 	if (ret)
754 		return ret;
755 
756 	switch (ioucmd->cmd_op) {
757 	case NVME_URING_CMD_IO:
758 		ret = nvme_uring_cmd_io(ctrl, ns, ioucmd, issue_flags, false);
759 		break;
760 	case NVME_URING_CMD_IO_VEC:
761 		ret = nvme_uring_cmd_io(ctrl, ns, ioucmd, issue_flags, true);
762 		break;
763 	default:
764 		ret = -ENOTTY;
765 	}
766 
767 	return ret;
768 }
769 
770 int nvme_ns_chr_uring_cmd(struct io_uring_cmd *ioucmd, unsigned int issue_flags)
771 {
772 	struct nvme_ns *ns = container_of(file_inode(ioucmd->file)->i_cdev,
773 			struct nvme_ns, cdev);
774 
775 	return nvme_ns_uring_cmd(ns, ioucmd, issue_flags);
776 }
777 
778 int nvme_ns_chr_uring_cmd_iopoll(struct io_uring_cmd *ioucmd,
779 				 struct io_comp_batch *iob,
780 				 unsigned int poll_flags)
781 {
782 	struct bio *bio;
783 	int ret = 0;
784 	struct nvme_ns *ns;
785 	struct request_queue *q;
786 
787 	rcu_read_lock();
788 	bio = READ_ONCE(ioucmd->cookie);
789 	ns = container_of(file_inode(ioucmd->file)->i_cdev,
790 			struct nvme_ns, cdev);
791 	q = ns->queue;
792 	if (test_bit(QUEUE_FLAG_POLL, &q->queue_flags) && bio && bio->bi_bdev)
793 		ret = bio_poll(bio, iob, poll_flags);
794 	rcu_read_unlock();
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 		fmode_t mode)
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, mode);
809 
810 	nvme_put_ctrl(ctrl);
811 	return ret;
812 }
813 
814 int nvme_ns_head_ioctl(struct block_device *bdev, fmode_t mode,
815 		unsigned int cmd, unsigned long arg)
816 {
817 	struct nvme_ns_head *head = bdev->bd_disk->private_data;
818 	void __user *argp = (void __user *)arg;
819 	struct nvme_ns *ns;
820 	int srcu_idx, ret = -EWOULDBLOCK;
821 	unsigned int flags = 0;
822 
823 	if (bdev_is_partition(bdev))
824 		flags |= NVME_IOCTL_PARTITION;
825 
826 	srcu_idx = srcu_read_lock(&head->srcu);
827 	ns = nvme_find_path(head);
828 	if (!ns)
829 		goto out_unlock;
830 
831 	/*
832 	 * Handle ioctls that apply to the controller instead of the namespace
833 	 * seperately and drop the ns SRCU reference early.  This avoids a
834 	 * deadlock when deleting namespaces using the passthrough interface.
835 	 */
836 	if (is_ctrl_ioctl(cmd))
837 		return nvme_ns_head_ctrl_ioctl(ns, cmd, argp, head, srcu_idx,
838 					mode);
839 
840 	ret = nvme_ns_ioctl(ns, cmd, argp, flags, mode);
841 out_unlock:
842 	srcu_read_unlock(&head->srcu, srcu_idx);
843 	return ret;
844 }
845 
846 long nvme_ns_head_chr_ioctl(struct file *file, unsigned int cmd,
847 		unsigned long arg)
848 {
849 	struct cdev *cdev = file_inode(file)->i_cdev;
850 	struct nvme_ns_head *head =
851 		container_of(cdev, struct nvme_ns_head, cdev);
852 	void __user *argp = (void __user *)arg;
853 	struct nvme_ns *ns;
854 	int srcu_idx, ret = -EWOULDBLOCK;
855 
856 	srcu_idx = srcu_read_lock(&head->srcu);
857 	ns = nvme_find_path(head);
858 	if (!ns)
859 		goto out_unlock;
860 
861 	if (is_ctrl_ioctl(cmd))
862 		return nvme_ns_head_ctrl_ioctl(ns, cmd, argp, head, srcu_idx,
863 				file->f_mode);
864 
865 	ret = nvme_ns_ioctl(ns, cmd, argp, 0, file->f_mode);
866 out_unlock:
867 	srcu_read_unlock(&head->srcu, srcu_idx);
868 	return ret;
869 }
870 
871 int nvme_ns_head_chr_uring_cmd(struct io_uring_cmd *ioucmd,
872 		unsigned int issue_flags)
873 {
874 	struct cdev *cdev = file_inode(ioucmd->file)->i_cdev;
875 	struct nvme_ns_head *head = container_of(cdev, struct nvme_ns_head, cdev);
876 	int srcu_idx = srcu_read_lock(&head->srcu);
877 	struct nvme_ns *ns = nvme_find_path(head);
878 	int ret = -EINVAL;
879 
880 	if (ns)
881 		ret = nvme_ns_uring_cmd(ns, ioucmd, issue_flags);
882 	srcu_read_unlock(&head->srcu, srcu_idx);
883 	return ret;
884 }
885 
886 int nvme_ns_head_chr_uring_cmd_iopoll(struct io_uring_cmd *ioucmd,
887 				      struct io_comp_batch *iob,
888 				      unsigned int poll_flags)
889 {
890 	struct cdev *cdev = file_inode(ioucmd->file)->i_cdev;
891 	struct nvme_ns_head *head = container_of(cdev, struct nvme_ns_head, cdev);
892 	int srcu_idx = srcu_read_lock(&head->srcu);
893 	struct nvme_ns *ns = nvme_find_path(head);
894 	struct bio *bio;
895 	int ret = 0;
896 	struct request_queue *q;
897 
898 	if (ns) {
899 		rcu_read_lock();
900 		bio = READ_ONCE(ioucmd->cookie);
901 		q = ns->queue;
902 		if (test_bit(QUEUE_FLAG_POLL, &q->queue_flags) && bio
903 				&& bio->bi_bdev)
904 			ret = bio_poll(bio, iob, poll_flags);
905 		rcu_read_unlock();
906 	}
907 	srcu_read_unlock(&head->srcu, srcu_idx);
908 	return ret;
909 }
910 #endif /* CONFIG_NVME_MULTIPATH */
911 
912 int nvme_dev_uring_cmd(struct io_uring_cmd *ioucmd, unsigned int issue_flags)
913 {
914 	struct nvme_ctrl *ctrl = ioucmd->file->private_data;
915 	int ret;
916 
917 	/* IOPOLL not supported yet */
918 	if (issue_flags & IO_URING_F_IOPOLL)
919 		return -EOPNOTSUPP;
920 
921 	ret = nvme_uring_cmd_checks(issue_flags);
922 	if (ret)
923 		return ret;
924 
925 	switch (ioucmd->cmd_op) {
926 	case NVME_URING_CMD_ADMIN:
927 		ret = nvme_uring_cmd_io(ctrl, NULL, ioucmd, issue_flags, false);
928 		break;
929 	case NVME_URING_CMD_ADMIN_VEC:
930 		ret = nvme_uring_cmd_io(ctrl, NULL, ioucmd, issue_flags, true);
931 		break;
932 	default:
933 		ret = -ENOTTY;
934 	}
935 
936 	return ret;
937 }
938 
939 static int nvme_dev_user_cmd(struct nvme_ctrl *ctrl, void __user *argp,
940 		fmode_t mode)
941 {
942 	struct nvme_ns *ns;
943 	int ret;
944 
945 	down_read(&ctrl->namespaces_rwsem);
946 	if (list_empty(&ctrl->namespaces)) {
947 		ret = -ENOTTY;
948 		goto out_unlock;
949 	}
950 
951 	ns = list_first_entry(&ctrl->namespaces, struct nvme_ns, list);
952 	if (ns != list_last_entry(&ctrl->namespaces, struct nvme_ns, list)) {
953 		dev_warn(ctrl->device,
954 			"NVME_IOCTL_IO_CMD not supported when multiple namespaces present!\n");
955 		ret = -EINVAL;
956 		goto out_unlock;
957 	}
958 
959 	dev_warn(ctrl->device,
960 		"using deprecated NVME_IOCTL_IO_CMD ioctl on the char device!\n");
961 	kref_get(&ns->kref);
962 	up_read(&ctrl->namespaces_rwsem);
963 
964 	ret = nvme_user_cmd(ctrl, ns, argp, 0, mode);
965 	nvme_put_ns(ns);
966 	return ret;
967 
968 out_unlock:
969 	up_read(&ctrl->namespaces_rwsem);
970 	return ret;
971 }
972 
973 long nvme_dev_ioctl(struct file *file, unsigned int cmd,
974 		unsigned long arg)
975 {
976 	struct nvme_ctrl *ctrl = file->private_data;
977 	void __user *argp = (void __user *)arg;
978 
979 	switch (cmd) {
980 	case NVME_IOCTL_ADMIN_CMD:
981 		return nvme_user_cmd(ctrl, NULL, argp, 0, file->f_mode);
982 	case NVME_IOCTL_ADMIN64_CMD:
983 		return nvme_user_cmd64(ctrl, NULL, argp, 0, file->f_mode);
984 	case NVME_IOCTL_IO_CMD:
985 		return nvme_dev_user_cmd(ctrl, argp, file->f_mode);
986 	case NVME_IOCTL_RESET:
987 		if (!capable(CAP_SYS_ADMIN))
988 			return -EACCES;
989 		dev_warn(ctrl->device, "resetting controller\n");
990 		return nvme_reset_ctrl_sync(ctrl);
991 	case NVME_IOCTL_SUBSYS_RESET:
992 		if (!capable(CAP_SYS_ADMIN))
993 			return -EACCES;
994 		return nvme_reset_subsystem(ctrl);
995 	case NVME_IOCTL_RESCAN:
996 		if (!capable(CAP_SYS_ADMIN))
997 			return -EACCES;
998 		nvme_queue_scan(ctrl);
999 		return 0;
1000 	default:
1001 		return -ENOTTY;
1002 	}
1003 }
1004