xref: /linux/drivers/nvme/target/passthru.c (revision bf4afc53b77aeaa48b5409da5c8da6bb4eff7f43)
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_obj(*id);
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 fragmentation. 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 limited 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 ctrl subsysnqn. In order to prevent that we
151 	 * mask the passthru-ctrl subsysnqn with the target ctrl subsysnqn.
152 	 */
153 	strscpy(id->subnqn, ctrl->subsys->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_obj(*id);
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,const struct io_comp_batch * iob)249 static enum rq_end_io_ret nvmet_passthru_req_done(struct request *rq,
250 						  blk_status_t blk_status,
251 						  const struct io_comp_batch *iob)
252 {
253 	struct nvmet_req *req = rq->end_io_data;
254 
255 	req->cqe->result = nvme_req(rq)->result;
256 	nvmet_req_complete(req, nvme_req(rq)->status);
257 	blk_mq_free_request(rq);
258 	return RQ_END_IO_NONE;
259 }
260 
nvmet_passthru_map_sg(struct nvmet_req * req,struct request * rq)261 static int nvmet_passthru_map_sg(struct nvmet_req *req, struct request *rq)
262 {
263 	struct scatterlist *sg;
264 	struct bio *bio;
265 	int ret = -EINVAL;
266 	int i;
267 
268 	if (req->sg_cnt > BIO_MAX_VECS)
269 		return -EINVAL;
270 
271 	if (nvmet_use_inline_bvec(req)) {
272 		bio = &req->p.inline_bio;
273 		bio_init(bio, NULL, req->inline_bvec,
274 			 ARRAY_SIZE(req->inline_bvec), req_op(rq));
275 	} else {
276 		bio = bio_alloc(NULL, bio_max_segs(req->sg_cnt), req_op(rq),
277 				GFP_KERNEL);
278 		bio->bi_end_io = bio_put;
279 	}
280 
281 	for_each_sg(req->sg, sg, req->sg_cnt, i) {
282 		if (bio_add_page(bio, sg_page(sg), sg->length, sg->offset) <
283 				sg->length)
284 			goto out_bio_put;
285 	}
286 
287 	ret = blk_rq_append_bio(rq, bio);
288 	if (ret)
289 		goto out_bio_put;
290 	return 0;
291 
292 out_bio_put:
293 	nvmet_req_bio_put(req, bio);
294 	return ret;
295 }
296 
nvmet_passthru_execute_cmd(struct nvmet_req * req)297 static void nvmet_passthru_execute_cmd(struct nvmet_req *req)
298 {
299 	struct nvme_ctrl *ctrl = nvmet_req_subsys(req)->passthru_ctrl;
300 	struct request_queue *q = ctrl->admin_q;
301 	struct nvme_ns *ns = NULL;
302 	struct request *rq = NULL;
303 	unsigned int timeout;
304 	u32 effects;
305 	u16 status;
306 	int ret;
307 
308 	if (likely(req->sq->qid != 0)) {
309 		u32 nsid = le32_to_cpu(req->cmd->common.nsid);
310 
311 		ns = nvme_find_get_ns(ctrl, nsid);
312 		if (unlikely(!ns)) {
313 			pr_err("failed to get passthru ns nsid:%u\n", nsid);
314 			status = NVME_SC_INVALID_NS | NVME_STATUS_DNR;
315 			goto out;
316 		}
317 
318 		q = ns->queue;
319 		timeout = nvmet_req_subsys(req)->io_timeout;
320 	} else {
321 		timeout = nvmet_req_subsys(req)->admin_timeout;
322 	}
323 
324 	rq = blk_mq_alloc_request(q, nvme_req_op(req->cmd), 0);
325 	if (IS_ERR(rq)) {
326 		status = NVME_SC_INTERNAL;
327 		goto out_put_ns;
328 	}
329 	nvme_init_request(rq, req->cmd);
330 
331 	if (timeout)
332 		rq->timeout = timeout;
333 
334 	if (req->sg_cnt) {
335 		ret = nvmet_passthru_map_sg(req, rq);
336 		if (unlikely(ret)) {
337 			status = NVME_SC_INTERNAL;
338 			goto out_put_req;
339 		}
340 	}
341 
342 	/*
343 	 * If a command needs post-execution fixups, or there are any
344 	 * non-trivial effects, make sure to execute the command synchronously
345 	 * in a workqueue so that nvme_passthru_end gets called.
346 	 */
347 	effects = nvme_command_effects(ctrl, ns, req->cmd->common.opcode);
348 	if (req->p.use_workqueue ||
349 	    (effects & ~(NVME_CMD_EFFECTS_CSUPP | NVME_CMD_EFFECTS_LBCC))) {
350 		INIT_WORK(&req->p.work, nvmet_passthru_execute_cmd_work);
351 		req->p.rq = rq;
352 		queue_work(nvmet_wq, &req->p.work);
353 	} else {
354 		rq->end_io = nvmet_passthru_req_done;
355 		rq->end_io_data = req;
356 		blk_execute_rq_nowait(rq, false);
357 	}
358 
359 	if (ns)
360 		nvme_put_ns(ns);
361 
362 	return;
363 
364 out_put_req:
365 	blk_mq_free_request(rq);
366 out_put_ns:
367 	if (ns)
368 		nvme_put_ns(ns);
369 out:
370 	nvmet_req_complete(req, status);
371 }
372 
373 /*
374  * We need to emulate set host behaviour to ensure that any requested
375  * behaviour of the target's host matches the requested behaviour
376  * of the device's host and fail otherwise.
377  */
nvmet_passthru_set_host_behaviour(struct nvmet_req * req)378 static void nvmet_passthru_set_host_behaviour(struct nvmet_req *req)
379 {
380 	struct nvme_ctrl *ctrl = nvmet_req_subsys(req)->passthru_ctrl;
381 	struct nvme_feat_host_behavior *host;
382 	u16 status = NVME_SC_INTERNAL;
383 	int ret;
384 
385 	host = kzalloc(sizeof(*host) * 2, GFP_KERNEL);
386 	if (!host)
387 		goto out_complete_req;
388 
389 	ret = nvme_get_features(ctrl, NVME_FEAT_HOST_BEHAVIOR, 0,
390 				host, sizeof(*host), NULL);
391 	if (ret)
392 		goto out_free_host;
393 
394 	status = nvmet_copy_from_sgl(req, 0, &host[1], sizeof(*host));
395 	if (status)
396 		goto out_free_host;
397 
398 	if (memcmp(&host[0], &host[1], sizeof(host[0]))) {
399 		pr_warn("target host has requested different behaviour from the local host\n");
400 		status = NVME_SC_INTERNAL;
401 	}
402 
403 out_free_host:
404 	kfree(host);
405 out_complete_req:
406 	nvmet_req_complete(req, status);
407 }
408 
nvmet_setup_passthru_command(struct nvmet_req * req)409 static u16 nvmet_setup_passthru_command(struct nvmet_req *req)
410 {
411 	req->p.use_workqueue = false;
412 	req->execute = nvmet_passthru_execute_cmd;
413 	return NVME_SC_SUCCESS;
414 }
415 
nvmet_parse_passthru_io_cmd(struct nvmet_req * req)416 u16 nvmet_parse_passthru_io_cmd(struct nvmet_req *req)
417 {
418 	/* Reject any commands with non-sgl flags set (ie. fused commands) */
419 	if (req->cmd->common.flags & ~NVME_CMD_SGL_ALL)
420 		return NVME_SC_INVALID_FIELD;
421 
422 	switch (req->cmd->common.opcode) {
423 	case nvme_cmd_resv_register:
424 	case nvme_cmd_resv_report:
425 	case nvme_cmd_resv_acquire:
426 	case nvme_cmd_resv_release:
427 		/*
428 		 * Reservations cannot be supported properly because the
429 		 * underlying device has no way of differentiating different
430 		 * hosts that connect via fabrics. This could potentially be
431 		 * emulated in the future if regular targets grow support for
432 		 * this feature.
433 		 */
434 		return NVME_SC_INVALID_OPCODE | NVME_STATUS_DNR;
435 	}
436 
437 	return nvmet_setup_passthru_command(req);
438 }
439 
440 /*
441  * Only features that are emulated or specifically allowed in the list  are
442  * passed down to the controller. This function implements the allow list for
443  * both get and set features.
444  */
nvmet_passthru_get_set_features(struct nvmet_req * req)445 static u16 nvmet_passthru_get_set_features(struct nvmet_req *req)
446 {
447 	switch (le32_to_cpu(req->cmd->features.fid)) {
448 	case NVME_FEAT_ARBITRATION:
449 	case NVME_FEAT_POWER_MGMT:
450 	case NVME_FEAT_LBA_RANGE:
451 	case NVME_FEAT_TEMP_THRESH:
452 	case NVME_FEAT_ERR_RECOVERY:
453 	case NVME_FEAT_VOLATILE_WC:
454 	case NVME_FEAT_WRITE_ATOMIC:
455 	case NVME_FEAT_AUTO_PST:
456 	case NVME_FEAT_TIMESTAMP:
457 	case NVME_FEAT_HCTM:
458 	case NVME_FEAT_NOPSC:
459 	case NVME_FEAT_RRL:
460 	case NVME_FEAT_PLM_CONFIG:
461 	case NVME_FEAT_PLM_WINDOW:
462 	case NVME_FEAT_HOST_BEHAVIOR:
463 	case NVME_FEAT_SANITIZE:
464 	case NVME_FEAT_VENDOR_START ... NVME_FEAT_VENDOR_END:
465 		return nvmet_setup_passthru_command(req);
466 
467 	case NVME_FEAT_ASYNC_EVENT:
468 		/* There is no support for forwarding ASYNC events */
469 	case NVME_FEAT_IRQ_COALESCE:
470 	case NVME_FEAT_IRQ_CONFIG:
471 		/* The IRQ settings will not apply to the target controller */
472 	case NVME_FEAT_HOST_MEM_BUF:
473 		/*
474 		 * Any HMB that's set will not be passed through and will
475 		 * not work as expected
476 		 */
477 	case NVME_FEAT_SW_PROGRESS:
478 		/*
479 		 * The Pre-Boot Software Load Count doesn't make much
480 		 * sense for a target to export
481 		 */
482 	case NVME_FEAT_RESV_MASK:
483 	case NVME_FEAT_RESV_PERSIST:
484 		/* No reservations, see nvmet_parse_passthru_io_cmd() */
485 	default:
486 		return NVME_SC_INVALID_OPCODE | NVME_STATUS_DNR;
487 	}
488 }
489 
nvmet_parse_passthru_admin_cmd(struct nvmet_req * req)490 u16 nvmet_parse_passthru_admin_cmd(struct nvmet_req *req)
491 {
492 	/* Reject any commands with non-sgl flags set (ie. fused commands) */
493 	if (req->cmd->common.flags & ~NVME_CMD_SGL_ALL)
494 		return NVME_SC_INVALID_FIELD;
495 
496 	/*
497 	 * Passthru all vendor specific commands
498 	 */
499 	if (req->cmd->common.opcode >= nvme_admin_vendor_start)
500 		return nvmet_setup_passthru_command(req);
501 
502 	switch (req->cmd->common.opcode) {
503 	case nvme_admin_async_event:
504 		req->execute = nvmet_execute_async_event;
505 		return NVME_SC_SUCCESS;
506 	case nvme_admin_keep_alive:
507 		/*
508 		 * Most PCIe ctrls don't support keep alive cmd, we route keep
509 		 * alive to the non-passthru mode. In future please change this
510 		 * code when PCIe ctrls with keep alive support available.
511 		 */
512 		req->execute = nvmet_execute_keep_alive;
513 		return NVME_SC_SUCCESS;
514 	case nvme_admin_set_features:
515 		switch (le32_to_cpu(req->cmd->features.fid)) {
516 		case NVME_FEAT_ASYNC_EVENT:
517 		case NVME_FEAT_KATO:
518 		case NVME_FEAT_NUM_QUEUES:
519 		case NVME_FEAT_HOST_ID:
520 			req->execute = nvmet_execute_set_features;
521 			return NVME_SC_SUCCESS;
522 		case NVME_FEAT_HOST_BEHAVIOR:
523 			req->execute = nvmet_passthru_set_host_behaviour;
524 			return NVME_SC_SUCCESS;
525 		default:
526 			return nvmet_passthru_get_set_features(req);
527 		}
528 		break;
529 	case nvme_admin_get_features:
530 		switch (le32_to_cpu(req->cmd->features.fid)) {
531 		case NVME_FEAT_ASYNC_EVENT:
532 		case NVME_FEAT_KATO:
533 		case NVME_FEAT_NUM_QUEUES:
534 		case NVME_FEAT_HOST_ID:
535 			req->execute = nvmet_execute_get_features;
536 			return NVME_SC_SUCCESS;
537 		case NVME_FEAT_FDP:
538 			return nvmet_setup_passthru_command(req);
539 		default:
540 			return nvmet_passthru_get_set_features(req);
541 		}
542 		break;
543 	case nvme_admin_identify:
544 		switch (req->cmd->identify.cns) {
545 		case NVME_ID_CNS_CS_CTRL:
546 			switch (req->cmd->identify.csi) {
547 			case NVME_CSI_ZNS:
548 				req->execute = nvmet_passthru_execute_cmd;
549 				req->p.use_workqueue = true;
550 				return NVME_SC_SUCCESS;
551 			}
552 			return NVME_SC_INVALID_OPCODE | NVME_STATUS_DNR;
553 		case NVME_ID_CNS_CTRL:
554 		case NVME_ID_CNS_NS:
555 		case NVME_ID_CNS_NS_DESC_LIST:
556 			req->execute = nvmet_passthru_execute_cmd;
557 			req->p.use_workqueue = true;
558 			return NVME_SC_SUCCESS;
559 		case NVME_ID_CNS_CS_NS:
560 			switch (req->cmd->identify.csi) {
561 			case NVME_CSI_ZNS:
562 				req->execute = nvmet_passthru_execute_cmd;
563 				req->p.use_workqueue = true;
564 				return NVME_SC_SUCCESS;
565 			}
566 			return NVME_SC_INVALID_OPCODE | NVME_STATUS_DNR;
567 		default:
568 			return nvmet_setup_passthru_command(req);
569 		}
570 	case nvme_admin_get_log_page:
571 		return nvmet_setup_passthru_command(req);
572 	default:
573 		/* Reject commands not in the allowlist above */
574 		return nvmet_report_invalid_opcode(req);
575 	}
576 }
577 
nvmet_passthru_ctrl_enable(struct nvmet_subsys * subsys)578 int nvmet_passthru_ctrl_enable(struct nvmet_subsys *subsys)
579 {
580 	struct nvme_ctrl *ctrl;
581 	struct file *file;
582 	int ret = -EINVAL;
583 	void *old;
584 
585 	mutex_lock(&subsys->lock);
586 	if (!subsys->passthru_ctrl_path)
587 		goto out_unlock;
588 	if (subsys->passthru_ctrl)
589 		goto out_unlock;
590 
591 	if (subsys->nr_namespaces) {
592 		pr_info("cannot enable both passthru and regular namespaces for a single subsystem");
593 		goto out_unlock;
594 	}
595 
596 	file = filp_open(subsys->passthru_ctrl_path, O_RDWR, 0);
597 	if (IS_ERR(file)) {
598 		ret = PTR_ERR(file);
599 		goto out_unlock;
600 	}
601 
602 	ctrl = nvme_ctrl_from_file(file);
603 	if (!ctrl) {
604 		pr_err("failed to open nvme controller %s\n",
605 		       subsys->passthru_ctrl_path);
606 
607 		goto out_put_file;
608 	}
609 
610 	old = xa_cmpxchg(&passthru_subsystems, ctrl->instance, NULL,
611 			 subsys, GFP_KERNEL);
612 	if (xa_is_err(old)) {
613 		ret = xa_err(old);
614 		goto out_put_file;
615 	}
616 
617 	if (old)
618 		goto out_put_file;
619 
620 	subsys->passthru_ctrl = ctrl;
621 	subsys->ver = ctrl->vs;
622 
623 	if (subsys->ver < NVME_VS(1, 2, 1)) {
624 		pr_warn("nvme controller version is too old: %llu.%llu.%llu, advertising 1.2.1\n",
625 			NVME_MAJOR(subsys->ver), NVME_MINOR(subsys->ver),
626 			NVME_TERTIARY(subsys->ver));
627 		subsys->ver = NVME_VS(1, 2, 1);
628 	}
629 	nvme_get_ctrl(ctrl);
630 	__module_get(subsys->passthru_ctrl->ops->module);
631 	ret = 0;
632 
633 out_put_file:
634 	filp_close(file, NULL);
635 out_unlock:
636 	mutex_unlock(&subsys->lock);
637 	return ret;
638 }
639 
__nvmet_passthru_ctrl_disable(struct nvmet_subsys * subsys)640 static void __nvmet_passthru_ctrl_disable(struct nvmet_subsys *subsys)
641 {
642 	if (subsys->passthru_ctrl) {
643 		xa_erase(&passthru_subsystems, subsys->passthru_ctrl->instance);
644 		module_put(subsys->passthru_ctrl->ops->module);
645 		nvme_put_ctrl(subsys->passthru_ctrl);
646 	}
647 	subsys->passthru_ctrl = NULL;
648 	subsys->ver = NVMET_DEFAULT_VS;
649 }
650 
nvmet_passthru_ctrl_disable(struct nvmet_subsys * subsys)651 void nvmet_passthru_ctrl_disable(struct nvmet_subsys *subsys)
652 {
653 	mutex_lock(&subsys->lock);
654 	__nvmet_passthru_ctrl_disable(subsys);
655 	mutex_unlock(&subsys->lock);
656 }
657 
nvmet_passthru_subsys_free(struct nvmet_subsys * subsys)658 void nvmet_passthru_subsys_free(struct nvmet_subsys *subsys)
659 {
660 	mutex_lock(&subsys->lock);
661 	__nvmet_passthru_ctrl_disable(subsys);
662 	mutex_unlock(&subsys->lock);
663 	kfree(subsys->passthru_ctrl_path);
664 }
665