xref: /linux/drivers/nvme/target/passthru.c (revision f8eacd8ad7a658b805c635f8ffad7913981f863c)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * NVMe Over Fabrics Target Passthrough command implementation.
4  *
5  * Copyright (c) 2017-2018 Western Digital Corporation or its
6  * affiliates.
7  * Copyright (c) 2019-2020, Eideticom Inc.
8  *
9  */
10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11 #include <linux/module.h>
12 
13 #include "../host/nvme.h"
14 #include "nvmet.h"
15 
16 MODULE_IMPORT_NS(NVME_TARGET_PASSTHRU);
17 
18 /*
19  * xarray to maintain one passthru subsystem per nvme controller.
20  */
21 static DEFINE_XARRAY(passthru_subsystems);
22 
nvmet_passthrough_override_cap(struct nvmet_ctrl * ctrl)23 void nvmet_passthrough_override_cap(struct nvmet_ctrl *ctrl)
24 {
25 	/*
26 	 * Multiple command set support can only be declared if the underlying
27 	 * controller actually supports it.
28 	 */
29 	if (!nvme_multi_css(ctrl->subsys->passthru_ctrl))
30 		ctrl->cap &= ~(1ULL << 43);
31 }
32 
nvmet_passthru_override_id_descs(struct nvmet_req * req)33 static u16 nvmet_passthru_override_id_descs(struct nvmet_req *req)
34 {
35 	struct nvmet_ctrl *ctrl = req->sq->ctrl;
36 	u16 status = NVME_SC_SUCCESS;
37 	int pos, len;
38 	bool csi_seen = false;
39 	void *data;
40 	u8 csi;
41 
42 	if (!ctrl->subsys->clear_ids)
43 		return status;
44 
45 	data = kzalloc(NVME_IDENTIFY_DATA_SIZE, GFP_KERNEL);
46 	if (!data)
47 		return NVME_SC_INTERNAL;
48 
49 	status = nvmet_copy_from_sgl(req, 0, data, NVME_IDENTIFY_DATA_SIZE);
50 	if (status)
51 		goto out_free;
52 
53 	for (pos = 0; pos < NVME_IDENTIFY_DATA_SIZE; pos += len) {
54 		struct nvme_ns_id_desc *cur = data + pos;
55 
56 		if (cur->nidl == 0)
57 			break;
58 		if (cur->nidt == NVME_NIDT_CSI) {
59 			memcpy(&csi, cur + 1, NVME_NIDT_CSI_LEN);
60 			csi_seen = true;
61 			break;
62 		}
63 		len = sizeof(struct nvme_ns_id_desc) + cur->nidl;
64 	}
65 
66 	memset(data, 0, NVME_IDENTIFY_DATA_SIZE);
67 	if (csi_seen) {
68 		struct nvme_ns_id_desc *cur = data;
69 
70 		cur->nidt = NVME_NIDT_CSI;
71 		cur->nidl = NVME_NIDT_CSI_LEN;
72 		memcpy(cur + 1, &csi, NVME_NIDT_CSI_LEN);
73 	}
74 	status = nvmet_copy_to_sgl(req, 0, data, NVME_IDENTIFY_DATA_SIZE);
75 out_free:
76 	kfree(data);
77 	return status;
78 }
79 
nvmet_passthru_override_id_ctrl(struct nvmet_req * req)80 static u16 nvmet_passthru_override_id_ctrl(struct nvmet_req *req)
81 {
82 	struct nvmet_ctrl *ctrl = req->sq->ctrl;
83 	struct nvme_ctrl *pctrl = ctrl->subsys->passthru_ctrl;
84 	u16 status = NVME_SC_SUCCESS;
85 	struct nvme_id_ctrl *id;
86 	unsigned int max_hw_sectors;
87 	int page_shift;
88 
89 	id = kzalloc(sizeof(*id), GFP_KERNEL);
90 	if (!id)
91 		return NVME_SC_INTERNAL;
92 
93 	status = nvmet_copy_from_sgl(req, 0, id, sizeof(*id));
94 	if (status)
95 		goto out_free;
96 
97 	id->cntlid = cpu_to_le16(ctrl->cntlid);
98 	id->ver = cpu_to_le32(ctrl->subsys->ver);
99 
100 	/*
101 	 * The passthru NVMe driver may have a limit on the number of segments
102 	 * which depends on the host's memory fragementation. To solve this,
103 	 * ensure mdts is limited to the pages equal to the number of segments.
104 	 */
105 	max_hw_sectors = min_not_zero(pctrl->max_segments << PAGE_SECTORS_SHIFT,
106 				      pctrl->max_hw_sectors);
107 
108 	/*
109 	 * nvmet_passthru_map_sg is limitted to using a single bio so limit
110 	 * the mdts based on BIO_MAX_VECS as well
111 	 */
112 	max_hw_sectors = min_not_zero(BIO_MAX_VECS << PAGE_SECTORS_SHIFT,
113 				      max_hw_sectors);
114 
115 	page_shift = NVME_CAP_MPSMIN(ctrl->cap) + 12;
116 
117 	id->mdts = ilog2(max_hw_sectors) + 9 - page_shift;
118 
119 	id->acl = 3;
120 	/*
121 	 * We export aerl limit for the fabrics controller, update this when
122 	 * passthru based aerl support is added.
123 	 */
124 	id->aerl = NVMET_ASYNC_EVENTS - 1;
125 
126 	/* emulate kas as most of the PCIe ctrl don't have a support for kas */
127 	id->kas = cpu_to_le16(NVMET_KAS);
128 
129 	/* don't support host memory buffer */
130 	id->hmpre = 0;
131 	id->hmmin = 0;
132 
133 	id->sqes = min_t(__u8, ((0x6 << 4) | 0x6), id->sqes);
134 	id->cqes = min_t(__u8, ((0x4 << 4) | 0x4), id->cqes);
135 	id->maxcmd = cpu_to_le16(NVMET_MAX_CMD(ctrl));
136 
137 	/* don't support fuse commands */
138 	id->fuses = 0;
139 
140 	id->sgls = cpu_to_le32(1 << 0); /* we always support SGLs */
141 	if (ctrl->ops->flags & NVMF_KEYED_SGLS)
142 		id->sgls |= cpu_to_le32(1 << 2);
143 	if (req->port->inline_data_size)
144 		id->sgls |= cpu_to_le32(1 << 20);
145 
146 	/*
147 	 * When passthru controller is setup using nvme-loop transport it will
148 	 * export the passthru ctrl subsysnqn (PCIe NVMe ctrl) and will fail in
149 	 * the nvme/host/core.c in the nvme_init_subsystem()->nvme_active_ctrl()
150 	 * code path with duplicate ctr subsynqn. In order to prevent that we
151 	 * mask the passthru-ctrl subsysnqn with the target ctrl subsysnqn.
152 	 */
153 	memcpy(id->subnqn, ctrl->subsysnqn, sizeof(id->subnqn));
154 
155 	/* use fabric id-ctrl values */
156 	id->ioccsz = cpu_to_le32((sizeof(struct nvme_command) +
157 				req->port->inline_data_size) / 16);
158 	id->iorcsz = cpu_to_le32(sizeof(struct nvme_completion) / 16);
159 
160 	id->msdbd = ctrl->ops->msdbd;
161 
162 	/* Support multipath connections with fabrics */
163 	id->cmic |= 1 << 1;
164 
165 	/* Disable reservations, see nvmet_parse_passthru_io_cmd() */
166 	id->oncs &= cpu_to_le16(~NVME_CTRL_ONCS_RESERVATIONS);
167 
168 	status = nvmet_copy_to_sgl(req, 0, id, sizeof(struct nvme_id_ctrl));
169 
170 out_free:
171 	kfree(id);
172 	return status;
173 }
174 
nvmet_passthru_override_id_ns(struct nvmet_req * req)175 static u16 nvmet_passthru_override_id_ns(struct nvmet_req *req)
176 {
177 	u16 status = NVME_SC_SUCCESS;
178 	struct nvme_id_ns *id;
179 	int i;
180 
181 	id = kzalloc(sizeof(*id), GFP_KERNEL);
182 	if (!id)
183 		return NVME_SC_INTERNAL;
184 
185 	status = nvmet_copy_from_sgl(req, 0, id, sizeof(struct nvme_id_ns));
186 	if (status)
187 		goto out_free;
188 
189 	for (i = 0; i < (id->nlbaf + 1); i++)
190 		if (id->lbaf[i].ms)
191 			memset(&id->lbaf[i], 0, sizeof(id->lbaf[i]));
192 
193 	id->flbas = id->flbas & ~(1 << 4);
194 
195 	/*
196 	 * Presently the NVMEof target code does not support sending
197 	 * metadata, so we must disable it here. This should be updated
198 	 * once target starts supporting metadata.
199 	 */
200 	id->mc = 0;
201 
202 	if (req->sq->ctrl->subsys->clear_ids) {
203 		memset(id->nguid, 0, NVME_NIDT_NGUID_LEN);
204 		memset(id->eui64, 0, NVME_NIDT_EUI64_LEN);
205 	}
206 
207 	status = nvmet_copy_to_sgl(req, 0, id, sizeof(*id));
208 
209 out_free:
210 	kfree(id);
211 	return status;
212 }
213 
nvmet_passthru_execute_cmd_work(struct work_struct * w)214 static void nvmet_passthru_execute_cmd_work(struct work_struct *w)
215 {
216 	struct nvmet_req *req = container_of(w, struct nvmet_req, p.work);
217 	struct request *rq = req->p.rq;
218 	struct nvme_ctrl *ctrl = nvme_req(rq)->ctrl;
219 	struct nvme_ns *ns = rq->q->queuedata;
220 	u32 effects;
221 	int status;
222 
223 	effects = nvme_passthru_start(ctrl, ns, req->cmd->common.opcode);
224 	status = nvme_execute_rq(rq, false);
225 	if (status == NVME_SC_SUCCESS &&
226 	    req->cmd->common.opcode == nvme_admin_identify) {
227 		switch (req->cmd->identify.cns) {
228 		case NVME_ID_CNS_CTRL:
229 			status = nvmet_passthru_override_id_ctrl(req);
230 			break;
231 		case NVME_ID_CNS_NS:
232 			status = nvmet_passthru_override_id_ns(req);
233 			break;
234 		case NVME_ID_CNS_NS_DESC_LIST:
235 			status = nvmet_passthru_override_id_descs(req);
236 			break;
237 		}
238 	} else if (status < 0)
239 		status = NVME_SC_INTERNAL;
240 
241 	req->cqe->result = nvme_req(rq)->result;
242 	nvmet_req_complete(req, status);
243 	blk_mq_free_request(rq);
244 
245 	if (effects)
246 		nvme_passthru_end(ctrl, ns, effects, req->cmd, status);
247 }
248 
nvmet_passthru_req_done(struct request * rq,blk_status_t blk_status)249 static enum rq_end_io_ret nvmet_passthru_req_done(struct request *rq,
250 						  blk_status_t blk_status)
251 {
252 	struct nvmet_req *req = rq->end_io_data;
253 
254 	req->cqe->result = nvme_req(rq)->result;
255 	nvmet_req_complete(req, nvme_req(rq)->status);
256 	blk_mq_free_request(rq);
257 	return RQ_END_IO_NONE;
258 }
259 
nvmet_passthru_map_sg(struct nvmet_req * req,struct request * rq)260 static int nvmet_passthru_map_sg(struct nvmet_req *req, struct request *rq)
261 {
262 	struct scatterlist *sg;
263 	struct bio *bio;
264 	int i;
265 
266 	if (req->sg_cnt > BIO_MAX_VECS)
267 		return -EINVAL;
268 
269 	if (nvmet_use_inline_bvec(req)) {
270 		bio = &req->p.inline_bio;
271 		bio_init(bio, NULL, req->inline_bvec,
272 			 ARRAY_SIZE(req->inline_bvec), req_op(rq));
273 	} else {
274 		bio = bio_alloc(NULL, bio_max_segs(req->sg_cnt), req_op(rq),
275 				GFP_KERNEL);
276 		bio->bi_end_io = bio_put;
277 	}
278 
279 	for_each_sg(req->sg, sg, req->sg_cnt, i) {
280 		if (bio_add_pc_page(rq->q, bio, sg_page(sg), sg->length,
281 				    sg->offset) < sg->length) {
282 			nvmet_req_bio_put(req, bio);
283 			return -EINVAL;
284 		}
285 	}
286 
287 	blk_rq_bio_prep(rq, bio, req->sg_cnt);
288 
289 	return 0;
290 }
291 
nvmet_passthru_execute_cmd(struct nvmet_req * req)292 static void nvmet_passthru_execute_cmd(struct nvmet_req *req)
293 {
294 	struct nvme_ctrl *ctrl = nvmet_req_subsys(req)->passthru_ctrl;
295 	struct request_queue *q = ctrl->admin_q;
296 	struct nvme_ns *ns = NULL;
297 	struct request *rq = NULL;
298 	unsigned int timeout;
299 	u32 effects;
300 	u16 status;
301 	int ret;
302 
303 	if (likely(req->sq->qid != 0)) {
304 		u32 nsid = le32_to_cpu(req->cmd->common.nsid);
305 
306 		ns = nvme_find_get_ns(ctrl, nsid);
307 		if (unlikely(!ns)) {
308 			pr_err("failed to get passthru ns nsid:%u\n", nsid);
309 			status = NVME_SC_INVALID_NS | NVME_STATUS_DNR;
310 			goto out;
311 		}
312 
313 		q = ns->queue;
314 		timeout = nvmet_req_subsys(req)->io_timeout;
315 	} else {
316 		timeout = nvmet_req_subsys(req)->admin_timeout;
317 	}
318 
319 	rq = blk_mq_alloc_request(q, nvme_req_op(req->cmd), 0);
320 	if (IS_ERR(rq)) {
321 		status = NVME_SC_INTERNAL;
322 		goto out_put_ns;
323 	}
324 	nvme_init_request(rq, req->cmd);
325 
326 	if (timeout)
327 		rq->timeout = timeout;
328 
329 	if (req->sg_cnt) {
330 		ret = nvmet_passthru_map_sg(req, rq);
331 		if (unlikely(ret)) {
332 			status = NVME_SC_INTERNAL;
333 			goto out_put_req;
334 		}
335 	}
336 
337 	/*
338 	 * If a command needs post-execution fixups, or there are any
339 	 * non-trivial effects, make sure to execute the command synchronously
340 	 * in a workqueue so that nvme_passthru_end gets called.
341 	 */
342 	effects = nvme_command_effects(ctrl, ns, req->cmd->common.opcode);
343 	if (req->p.use_workqueue ||
344 	    (effects & ~(NVME_CMD_EFFECTS_CSUPP | NVME_CMD_EFFECTS_LBCC))) {
345 		INIT_WORK(&req->p.work, nvmet_passthru_execute_cmd_work);
346 		req->p.rq = rq;
347 		queue_work(nvmet_wq, &req->p.work);
348 	} else {
349 		rq->end_io = nvmet_passthru_req_done;
350 		rq->end_io_data = req;
351 		blk_execute_rq_nowait(rq, false);
352 	}
353 
354 	if (ns)
355 		nvme_put_ns(ns);
356 
357 	return;
358 
359 out_put_req:
360 	blk_mq_free_request(rq);
361 out_put_ns:
362 	if (ns)
363 		nvme_put_ns(ns);
364 out:
365 	nvmet_req_complete(req, status);
366 }
367 
368 /*
369  * We need to emulate set host behaviour to ensure that any requested
370  * behaviour of the target's host matches the requested behaviour
371  * of the device's host and fail otherwise.
372  */
nvmet_passthru_set_host_behaviour(struct nvmet_req * req)373 static void nvmet_passthru_set_host_behaviour(struct nvmet_req *req)
374 {
375 	struct nvme_ctrl *ctrl = nvmet_req_subsys(req)->passthru_ctrl;
376 	struct nvme_feat_host_behavior *host;
377 	u16 status = NVME_SC_INTERNAL;
378 	int ret;
379 
380 	host = kzalloc(sizeof(*host) * 2, GFP_KERNEL);
381 	if (!host)
382 		goto out_complete_req;
383 
384 	ret = nvme_get_features(ctrl, NVME_FEAT_HOST_BEHAVIOR, 0,
385 				host, sizeof(*host), NULL);
386 	if (ret)
387 		goto out_free_host;
388 
389 	status = nvmet_copy_from_sgl(req, 0, &host[1], sizeof(*host));
390 	if (status)
391 		goto out_free_host;
392 
393 	if (memcmp(&host[0], &host[1], sizeof(host[0]))) {
394 		pr_warn("target host has requested different behaviour from the local host\n");
395 		status = NVME_SC_INTERNAL;
396 	}
397 
398 out_free_host:
399 	kfree(host);
400 out_complete_req:
401 	nvmet_req_complete(req, status);
402 }
403 
nvmet_setup_passthru_command(struct nvmet_req * req)404 static u16 nvmet_setup_passthru_command(struct nvmet_req *req)
405 {
406 	req->p.use_workqueue = false;
407 	req->execute = nvmet_passthru_execute_cmd;
408 	return NVME_SC_SUCCESS;
409 }
410 
nvmet_parse_passthru_io_cmd(struct nvmet_req * req)411 u16 nvmet_parse_passthru_io_cmd(struct nvmet_req *req)
412 {
413 	/* Reject any commands with non-sgl flags set (ie. fused commands) */
414 	if (req->cmd->common.flags & ~NVME_CMD_SGL_ALL)
415 		return NVME_SC_INVALID_FIELD;
416 
417 	switch (req->cmd->common.opcode) {
418 	case nvme_cmd_resv_register:
419 	case nvme_cmd_resv_report:
420 	case nvme_cmd_resv_acquire:
421 	case nvme_cmd_resv_release:
422 		/*
423 		 * Reservations cannot be supported properly because the
424 		 * underlying device has no way of differentiating different
425 		 * hosts that connect via fabrics. This could potentially be
426 		 * emulated in the future if regular targets grow support for
427 		 * this feature.
428 		 */
429 		return NVME_SC_INVALID_OPCODE | NVME_STATUS_DNR;
430 	}
431 
432 	return nvmet_setup_passthru_command(req);
433 }
434 
435 /*
436  * Only features that are emulated or specifically allowed in the list  are
437  * passed down to the controller. This function implements the allow list for
438  * both get and set features.
439  */
nvmet_passthru_get_set_features(struct nvmet_req * req)440 static u16 nvmet_passthru_get_set_features(struct nvmet_req *req)
441 {
442 	switch (le32_to_cpu(req->cmd->features.fid)) {
443 	case NVME_FEAT_ARBITRATION:
444 	case NVME_FEAT_POWER_MGMT:
445 	case NVME_FEAT_LBA_RANGE:
446 	case NVME_FEAT_TEMP_THRESH:
447 	case NVME_FEAT_ERR_RECOVERY:
448 	case NVME_FEAT_VOLATILE_WC:
449 	case NVME_FEAT_WRITE_ATOMIC:
450 	case NVME_FEAT_AUTO_PST:
451 	case NVME_FEAT_TIMESTAMP:
452 	case NVME_FEAT_HCTM:
453 	case NVME_FEAT_NOPSC:
454 	case NVME_FEAT_RRL:
455 	case NVME_FEAT_PLM_CONFIG:
456 	case NVME_FEAT_PLM_WINDOW:
457 	case NVME_FEAT_HOST_BEHAVIOR:
458 	case NVME_FEAT_SANITIZE:
459 	case NVME_FEAT_VENDOR_START ... NVME_FEAT_VENDOR_END:
460 		return nvmet_setup_passthru_command(req);
461 
462 	case NVME_FEAT_ASYNC_EVENT:
463 		/* There is no support for forwarding ASYNC events */
464 	case NVME_FEAT_IRQ_COALESCE:
465 	case NVME_FEAT_IRQ_CONFIG:
466 		/* The IRQ settings will not apply to the target controller */
467 	case NVME_FEAT_HOST_MEM_BUF:
468 		/*
469 		 * Any HMB that's set will not be passed through and will
470 		 * not work as expected
471 		 */
472 	case NVME_FEAT_SW_PROGRESS:
473 		/*
474 		 * The Pre-Boot Software Load Count doesn't make much
475 		 * sense for a target to export
476 		 */
477 	case NVME_FEAT_RESV_MASK:
478 	case NVME_FEAT_RESV_PERSIST:
479 		/* No reservations, see nvmet_parse_passthru_io_cmd() */
480 	default:
481 		return NVME_SC_INVALID_OPCODE | NVME_STATUS_DNR;
482 	}
483 }
484 
nvmet_parse_passthru_admin_cmd(struct nvmet_req * req)485 u16 nvmet_parse_passthru_admin_cmd(struct nvmet_req *req)
486 {
487 	/* Reject any commands with non-sgl flags set (ie. fused commands) */
488 	if (req->cmd->common.flags & ~NVME_CMD_SGL_ALL)
489 		return NVME_SC_INVALID_FIELD;
490 
491 	/*
492 	 * Passthru all vendor specific commands
493 	 */
494 	if (req->cmd->common.opcode >= nvme_admin_vendor_start)
495 		return nvmet_setup_passthru_command(req);
496 
497 	switch (req->cmd->common.opcode) {
498 	case nvme_admin_async_event:
499 		req->execute = nvmet_execute_async_event;
500 		return NVME_SC_SUCCESS;
501 	case nvme_admin_keep_alive:
502 		/*
503 		 * Most PCIe ctrls don't support keep alive cmd, we route keep
504 		 * alive to the non-passthru mode. In future please change this
505 		 * code when PCIe ctrls with keep alive support available.
506 		 */
507 		req->execute = nvmet_execute_keep_alive;
508 		return NVME_SC_SUCCESS;
509 	case nvme_admin_set_features:
510 		switch (le32_to_cpu(req->cmd->features.fid)) {
511 		case NVME_FEAT_ASYNC_EVENT:
512 		case NVME_FEAT_KATO:
513 		case NVME_FEAT_NUM_QUEUES:
514 		case NVME_FEAT_HOST_ID:
515 			req->execute = nvmet_execute_set_features;
516 			return NVME_SC_SUCCESS;
517 		case NVME_FEAT_HOST_BEHAVIOR:
518 			req->execute = nvmet_passthru_set_host_behaviour;
519 			return NVME_SC_SUCCESS;
520 		default:
521 			return nvmet_passthru_get_set_features(req);
522 		}
523 		break;
524 	case nvme_admin_get_features:
525 		switch (le32_to_cpu(req->cmd->features.fid)) {
526 		case NVME_FEAT_ASYNC_EVENT:
527 		case NVME_FEAT_KATO:
528 		case NVME_FEAT_NUM_QUEUES:
529 		case NVME_FEAT_HOST_ID:
530 			req->execute = nvmet_execute_get_features;
531 			return NVME_SC_SUCCESS;
532 		default:
533 			return nvmet_passthru_get_set_features(req);
534 		}
535 		break;
536 	case nvme_admin_identify:
537 		switch (req->cmd->identify.cns) {
538 		case NVME_ID_CNS_CS_CTRL:
539 			switch (req->cmd->identify.csi) {
540 			case NVME_CSI_ZNS:
541 				req->execute = nvmet_passthru_execute_cmd;
542 				req->p.use_workqueue = true;
543 				return NVME_SC_SUCCESS;
544 			}
545 			return NVME_SC_INVALID_OPCODE | NVME_STATUS_DNR;
546 		case NVME_ID_CNS_CTRL:
547 		case NVME_ID_CNS_NS:
548 		case NVME_ID_CNS_NS_DESC_LIST:
549 			req->execute = nvmet_passthru_execute_cmd;
550 			req->p.use_workqueue = true;
551 			return NVME_SC_SUCCESS;
552 		case NVME_ID_CNS_CS_NS:
553 			switch (req->cmd->identify.csi) {
554 			case NVME_CSI_ZNS:
555 				req->execute = nvmet_passthru_execute_cmd;
556 				req->p.use_workqueue = true;
557 				return NVME_SC_SUCCESS;
558 			}
559 			return NVME_SC_INVALID_OPCODE | NVME_STATUS_DNR;
560 		default:
561 			return nvmet_setup_passthru_command(req);
562 		}
563 	case nvme_admin_get_log_page:
564 		return nvmet_setup_passthru_command(req);
565 	default:
566 		/* Reject commands not in the allowlist above */
567 		return nvmet_report_invalid_opcode(req);
568 	}
569 }
570 
nvmet_passthru_ctrl_enable(struct nvmet_subsys * subsys)571 int nvmet_passthru_ctrl_enable(struct nvmet_subsys *subsys)
572 {
573 	struct nvme_ctrl *ctrl;
574 	struct file *file;
575 	int ret = -EINVAL;
576 	void *old;
577 
578 	mutex_lock(&subsys->lock);
579 	if (!subsys->passthru_ctrl_path)
580 		goto out_unlock;
581 	if (subsys->passthru_ctrl)
582 		goto out_unlock;
583 
584 	if (subsys->nr_namespaces) {
585 		pr_info("cannot enable both passthru and regular namespaces for a single subsystem");
586 		goto out_unlock;
587 	}
588 
589 	file = filp_open(subsys->passthru_ctrl_path, O_RDWR, 0);
590 	if (IS_ERR(file)) {
591 		ret = PTR_ERR(file);
592 		goto out_unlock;
593 	}
594 
595 	ctrl = nvme_ctrl_from_file(file);
596 	if (!ctrl) {
597 		pr_err("failed to open nvme controller %s\n",
598 		       subsys->passthru_ctrl_path);
599 
600 		goto out_put_file;
601 	}
602 
603 	old = xa_cmpxchg(&passthru_subsystems, ctrl->instance, NULL,
604 			 subsys, GFP_KERNEL);
605 	if (xa_is_err(old)) {
606 		ret = xa_err(old);
607 		goto out_put_file;
608 	}
609 
610 	if (old)
611 		goto out_put_file;
612 
613 	subsys->passthru_ctrl = ctrl;
614 	subsys->ver = ctrl->vs;
615 
616 	if (subsys->ver < NVME_VS(1, 2, 1)) {
617 		pr_warn("nvme controller version is too old: %llu.%llu.%llu, advertising 1.2.1\n",
618 			NVME_MAJOR(subsys->ver), NVME_MINOR(subsys->ver),
619 			NVME_TERTIARY(subsys->ver));
620 		subsys->ver = NVME_VS(1, 2, 1);
621 	}
622 	nvme_get_ctrl(ctrl);
623 	__module_get(subsys->passthru_ctrl->ops->module);
624 	ret = 0;
625 
626 out_put_file:
627 	filp_close(file, NULL);
628 out_unlock:
629 	mutex_unlock(&subsys->lock);
630 	return ret;
631 }
632 
__nvmet_passthru_ctrl_disable(struct nvmet_subsys * subsys)633 static void __nvmet_passthru_ctrl_disable(struct nvmet_subsys *subsys)
634 {
635 	if (subsys->passthru_ctrl) {
636 		xa_erase(&passthru_subsystems, subsys->passthru_ctrl->instance);
637 		module_put(subsys->passthru_ctrl->ops->module);
638 		nvme_put_ctrl(subsys->passthru_ctrl);
639 	}
640 	subsys->passthru_ctrl = NULL;
641 	subsys->ver = NVMET_DEFAULT_VS;
642 }
643 
nvmet_passthru_ctrl_disable(struct nvmet_subsys * subsys)644 void nvmet_passthru_ctrl_disable(struct nvmet_subsys *subsys)
645 {
646 	mutex_lock(&subsys->lock);
647 	__nvmet_passthru_ctrl_disable(subsys);
648 	mutex_unlock(&subsys->lock);
649 }
650 
nvmet_passthru_subsys_free(struct nvmet_subsys * subsys)651 void nvmet_passthru_subsys_free(struct nvmet_subsys *subsys)
652 {
653 	mutex_lock(&subsys->lock);
654 	__nvmet_passthru_ctrl_disable(subsys);
655 	mutex_unlock(&subsys->lock);
656 	kfree(subsys->passthru_ctrl_path);
657 }
658