1 // SPDX-License-Identifier: GPL-2.0
3 * NVMe Over Fabrics Target Passthrough command implementation.
5 * Copyright (c) 2017-2018 Western Digital Corporation or its
7 * Copyright (c) 2019-2020, Eideticom Inc.
10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11 #include <linux/module.h>
13 #include "../host/nvme.h"
16 MODULE_IMPORT_NS(NVME_TARGET_PASSTHRU);
19 * xarray to maintain one passthru subsystem per nvme controller.
21 static DEFINE_XARRAY(passthru_subsystems);
23 static u16 nvmet_passthru_override_id_ctrl(struct nvmet_req *req)
25 struct nvmet_ctrl *ctrl = req->sq->ctrl;
26 struct nvme_ctrl *pctrl = ctrl->subsys->passthru_ctrl;
27 u16 status = NVME_SC_SUCCESS;
28 struct nvme_id_ctrl *id;
29 unsigned int max_hw_sectors;
32 id = kzalloc(sizeof(*id), GFP_KERNEL);
34 return NVME_SC_INTERNAL;
36 status = nvmet_copy_from_sgl(req, 0, id, sizeof(*id));
40 id->cntlid = cpu_to_le16(ctrl->cntlid);
41 id->ver = cpu_to_le32(ctrl->subsys->ver);
44 * The passthru NVMe driver may have a limit on the number of segments
45 * which depends on the host's memory fragementation. To solve this,
46 * ensure mdts is limited to the pages equal to the number of segments.
48 max_hw_sectors = min_not_zero(pctrl->max_segments << (PAGE_SHIFT - 9),
49 pctrl->max_hw_sectors);
52 * nvmet_passthru_map_sg is limitted to using a single bio so limit
53 * the mdts based on BIO_MAX_VECS as well
55 max_hw_sectors = min_not_zero(BIO_MAX_VECS << (PAGE_SHIFT - 9),
58 page_shift = NVME_CAP_MPSMIN(ctrl->cap) + 12;
60 id->mdts = ilog2(max_hw_sectors) + 9 - page_shift;
64 * We export aerl limit for the fabrics controller, update this when
65 * passthru based aerl support is added.
67 id->aerl = NVMET_ASYNC_EVENTS - 1;
69 /* emulate kas as most of the PCIe ctrl don't have a support for kas */
70 id->kas = cpu_to_le16(NVMET_KAS);
72 /* don't support host memory buffer */
76 id->sqes = min_t(__u8, ((0x6 << 4) | 0x6), id->sqes);
77 id->cqes = min_t(__u8, ((0x4 << 4) | 0x4), id->cqes);
78 id->maxcmd = cpu_to_le16(NVMET_MAX_CMD);
80 /* don't support fuse commands */
83 id->sgls = cpu_to_le32(1 << 0); /* we always support SGLs */
84 if (ctrl->ops->flags & NVMF_KEYED_SGLS)
85 id->sgls |= cpu_to_le32(1 << 2);
86 if (req->port->inline_data_size)
87 id->sgls |= cpu_to_le32(1 << 20);
90 * When passsthru controller is setup using nvme-loop transport it will
91 * export the passthru ctrl subsysnqn (PCIe NVMe ctrl) and will fail in
92 * the nvme/host/core.c in the nvme_init_subsystem()->nvme_active_ctrl()
93 * code path with duplicate ctr subsynqn. In order to prevent that we
94 * mask the passthru-ctrl subsysnqn with the target ctrl subsysnqn.
96 memcpy(id->subnqn, ctrl->subsysnqn, sizeof(id->subnqn));
98 /* use fabric id-ctrl values */
99 id->ioccsz = cpu_to_le32((sizeof(struct nvme_command) +
100 req->port->inline_data_size) / 16);
101 id->iorcsz = cpu_to_le32(sizeof(struct nvme_completion) / 16);
103 id->msdbd = ctrl->ops->msdbd;
105 /* Support multipath connections with fabrics */
108 /* Disable reservations, see nvmet_parse_passthru_io_cmd() */
109 id->oncs &= cpu_to_le16(~NVME_CTRL_ONCS_RESERVATIONS);
111 status = nvmet_copy_to_sgl(req, 0, id, sizeof(struct nvme_id_ctrl));
118 static u16 nvmet_passthru_override_id_ns(struct nvmet_req *req)
120 u16 status = NVME_SC_SUCCESS;
121 struct nvme_id_ns *id;
124 id = kzalloc(sizeof(*id), GFP_KERNEL);
126 return NVME_SC_INTERNAL;
128 status = nvmet_copy_from_sgl(req, 0, id, sizeof(struct nvme_id_ns));
132 for (i = 0; i < (id->nlbaf + 1); i++)
134 memset(&id->lbaf[i], 0, sizeof(id->lbaf[i]));
136 id->flbas = id->flbas & ~(1 << 4);
139 * Presently the NVMEof target code does not support sending
140 * metadata, so we must disable it here. This should be updated
141 * once target starts supporting metadata.
145 status = nvmet_copy_to_sgl(req, 0, id, sizeof(*id));
152 static void nvmet_passthru_execute_cmd_work(struct work_struct *w)
154 struct nvmet_req *req = container_of(w, struct nvmet_req, p.work);
155 struct request *rq = req->p.rq;
158 status = nvme_execute_passthru_rq(rq);
160 if (status == NVME_SC_SUCCESS &&
161 req->cmd->common.opcode == nvme_admin_identify) {
162 switch (req->cmd->identify.cns) {
163 case NVME_ID_CNS_CTRL:
164 nvmet_passthru_override_id_ctrl(req);
167 nvmet_passthru_override_id_ns(req);
170 } else if (status < 0)
171 status = NVME_SC_INTERNAL;
173 req->cqe->result = nvme_req(rq)->result;
174 nvmet_req_complete(req, status);
175 blk_mq_free_request(rq);
178 static void nvmet_passthru_req_done(struct request *rq,
179 blk_status_t blk_status)
181 struct nvmet_req *req = rq->end_io_data;
183 req->cqe->result = nvme_req(rq)->result;
184 nvmet_req_complete(req, nvme_req(rq)->status);
185 blk_mq_free_request(rq);
188 static int nvmet_passthru_map_sg(struct nvmet_req *req, struct request *rq)
190 struct scatterlist *sg;
194 if (req->sg_cnt > BIO_MAX_VECS)
197 if (nvmet_use_inline_bvec(req)) {
198 bio = &req->p.inline_bio;
199 bio_init(bio, req->inline_bvec, ARRAY_SIZE(req->inline_bvec));
201 bio = bio_alloc(GFP_KERNEL, bio_max_segs(req->sg_cnt));
202 bio->bi_end_io = bio_put;
204 bio->bi_opf = req_op(rq);
206 for_each_sg(req->sg, sg, req->sg_cnt, i) {
207 if (bio_add_pc_page(rq->q, bio, sg_page(sg), sg->length,
208 sg->offset) < sg->length) {
209 nvmet_req_bio_put(req, bio);
214 blk_rq_bio_prep(rq, bio, req->sg_cnt);
219 static void nvmet_passthru_execute_cmd(struct nvmet_req *req)
221 struct nvme_ctrl *ctrl = nvmet_req_passthru_ctrl(req);
222 struct request_queue *q = ctrl->admin_q;
223 struct nvme_ns *ns = NULL;
224 struct request *rq = NULL;
225 unsigned int timeout;
230 if (likely(req->sq->qid != 0)) {
231 u32 nsid = le32_to_cpu(req->cmd->common.nsid);
233 ns = nvme_find_get_ns(ctrl, nsid);
235 pr_err("failed to get passthru ns nsid:%u\n", nsid);
236 status = NVME_SC_INVALID_NS | NVME_SC_DNR;
241 timeout = nvmet_req_subsys(req)->io_timeout;
243 timeout = nvmet_req_subsys(req)->admin_timeout;
246 rq = nvme_alloc_request(q, req->cmd, 0);
248 status = NVME_SC_INTERNAL;
253 rq->timeout = timeout;
256 ret = nvmet_passthru_map_sg(req, rq);
258 status = NVME_SC_INTERNAL;
264 * If there are effects for the command we are about to execute, or
265 * an end_req function we need to use nvme_execute_passthru_rq()
266 * synchronously in a work item seeing the end_req function and
267 * nvme_passthru_end() can't be called in the request done callback
268 * which is typically in interrupt context.
270 effects = nvme_command_effects(ctrl, ns, req->cmd->common.opcode);
271 if (req->p.use_workqueue || effects) {
272 INIT_WORK(&req->p.work, nvmet_passthru_execute_cmd_work);
274 schedule_work(&req->p.work);
276 rq->end_io_data = req;
277 blk_execute_rq_nowait(ns ? ns->disk : NULL, rq, 0,
278 nvmet_passthru_req_done);
287 blk_mq_free_request(rq);
292 nvmet_req_complete(req, status);
296 * We need to emulate set host behaviour to ensure that any requested
297 * behaviour of the target's host matches the requested behaviour
298 * of the device's host and fail otherwise.
300 static void nvmet_passthru_set_host_behaviour(struct nvmet_req *req)
302 struct nvme_ctrl *ctrl = nvmet_req_passthru_ctrl(req);
303 struct nvme_feat_host_behavior *host;
304 u16 status = NVME_SC_INTERNAL;
307 host = kzalloc(sizeof(*host) * 2, GFP_KERNEL);
309 goto out_complete_req;
311 ret = nvme_get_features(ctrl, NVME_FEAT_HOST_BEHAVIOR, 0,
312 host, sizeof(*host), NULL);
316 status = nvmet_copy_from_sgl(req, 0, &host[1], sizeof(*host));
320 if (memcmp(&host[0], &host[1], sizeof(host[0]))) {
321 pr_warn("target host has requested different behaviour from the local host\n");
322 status = NVME_SC_INTERNAL;
328 nvmet_req_complete(req, status);
331 static u16 nvmet_setup_passthru_command(struct nvmet_req *req)
333 req->p.use_workqueue = false;
334 req->execute = nvmet_passthru_execute_cmd;
335 return NVME_SC_SUCCESS;
338 u16 nvmet_parse_passthru_io_cmd(struct nvmet_req *req)
340 /* Reject any commands with non-sgl flags set (ie. fused commands) */
341 if (req->cmd->common.flags & ~NVME_CMD_SGL_ALL)
342 return NVME_SC_INVALID_FIELD;
344 switch (req->cmd->common.opcode) {
345 case nvme_cmd_resv_register:
346 case nvme_cmd_resv_report:
347 case nvme_cmd_resv_acquire:
348 case nvme_cmd_resv_release:
350 * Reservations cannot be supported properly because the
351 * underlying device has no way of differentiating different
352 * hosts that connect via fabrics. This could potentially be
353 * emulated in the future if regular targets grow support for
356 return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
359 return nvmet_setup_passthru_command(req);
363 * Only features that are emulated or specifically allowed in the list are
364 * passed down to the controller. This function implements the allow list for
365 * both get and set features.
367 static u16 nvmet_passthru_get_set_features(struct nvmet_req *req)
369 switch (le32_to_cpu(req->cmd->features.fid)) {
370 case NVME_FEAT_ARBITRATION:
371 case NVME_FEAT_POWER_MGMT:
372 case NVME_FEAT_LBA_RANGE:
373 case NVME_FEAT_TEMP_THRESH:
374 case NVME_FEAT_ERR_RECOVERY:
375 case NVME_FEAT_VOLATILE_WC:
376 case NVME_FEAT_WRITE_ATOMIC:
377 case NVME_FEAT_AUTO_PST:
378 case NVME_FEAT_TIMESTAMP:
380 case NVME_FEAT_NOPSC:
382 case NVME_FEAT_PLM_CONFIG:
383 case NVME_FEAT_PLM_WINDOW:
384 case NVME_FEAT_HOST_BEHAVIOR:
385 case NVME_FEAT_SANITIZE:
386 case NVME_FEAT_VENDOR_START ... NVME_FEAT_VENDOR_END:
387 return nvmet_setup_passthru_command(req);
389 case NVME_FEAT_ASYNC_EVENT:
390 /* There is no support for forwarding ASYNC events */
391 case NVME_FEAT_IRQ_COALESCE:
392 case NVME_FEAT_IRQ_CONFIG:
393 /* The IRQ settings will not apply to the target controller */
394 case NVME_FEAT_HOST_MEM_BUF:
396 * Any HMB that's set will not be passed through and will
397 * not work as expected
399 case NVME_FEAT_SW_PROGRESS:
401 * The Pre-Boot Software Load Count doesn't make much
402 * sense for a target to export
404 case NVME_FEAT_RESV_MASK:
405 case NVME_FEAT_RESV_PERSIST:
406 /* No reservations, see nvmet_parse_passthru_io_cmd() */
408 return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
412 u16 nvmet_parse_passthru_admin_cmd(struct nvmet_req *req)
414 /* Reject any commands with non-sgl flags set (ie. fused commands) */
415 if (req->cmd->common.flags & ~NVME_CMD_SGL_ALL)
416 return NVME_SC_INVALID_FIELD;
419 * Passthru all vendor specific commands
421 if (req->cmd->common.opcode >= nvme_admin_vendor_start)
422 return nvmet_setup_passthru_command(req);
424 switch (req->cmd->common.opcode) {
425 case nvme_admin_async_event:
426 req->execute = nvmet_execute_async_event;
427 return NVME_SC_SUCCESS;
428 case nvme_admin_keep_alive:
430 * Most PCIe ctrls don't support keep alive cmd, we route keep
431 * alive to the non-passthru mode. In future please change this
432 * code when PCIe ctrls with keep alive support available.
434 req->execute = nvmet_execute_keep_alive;
435 return NVME_SC_SUCCESS;
436 case nvme_admin_set_features:
437 switch (le32_to_cpu(req->cmd->features.fid)) {
438 case NVME_FEAT_ASYNC_EVENT:
440 case NVME_FEAT_NUM_QUEUES:
441 case NVME_FEAT_HOST_ID:
442 req->execute = nvmet_execute_set_features;
443 return NVME_SC_SUCCESS;
444 case NVME_FEAT_HOST_BEHAVIOR:
445 req->execute = nvmet_passthru_set_host_behaviour;
446 return NVME_SC_SUCCESS;
448 return nvmet_passthru_get_set_features(req);
451 case nvme_admin_get_features:
452 switch (le32_to_cpu(req->cmd->features.fid)) {
453 case NVME_FEAT_ASYNC_EVENT:
455 case NVME_FEAT_NUM_QUEUES:
456 case NVME_FEAT_HOST_ID:
457 req->execute = nvmet_execute_get_features;
458 return NVME_SC_SUCCESS;
460 return nvmet_passthru_get_set_features(req);
463 case nvme_admin_identify:
464 switch (req->cmd->identify.cns) {
465 case NVME_ID_CNS_CTRL:
466 req->execute = nvmet_passthru_execute_cmd;
467 req->p.use_workqueue = true;
468 return NVME_SC_SUCCESS;
469 case NVME_ID_CNS_CS_CTRL:
470 switch (req->cmd->identify.csi) {
472 req->execute = nvmet_passthru_execute_cmd;
473 req->p.use_workqueue = true;
474 return NVME_SC_SUCCESS;
476 return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
478 req->execute = nvmet_passthru_execute_cmd;
479 req->p.use_workqueue = true;
480 return NVME_SC_SUCCESS;
481 case NVME_ID_CNS_CS_NS:
482 switch (req->cmd->identify.csi) {
484 req->execute = nvmet_passthru_execute_cmd;
485 req->p.use_workqueue = true;
486 return NVME_SC_SUCCESS;
488 return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
490 return nvmet_setup_passthru_command(req);
492 case nvme_admin_get_log_page:
493 return nvmet_setup_passthru_command(req);
495 /* Reject commands not in the allowlist above */
496 return nvmet_report_invalid_opcode(req);
500 int nvmet_passthru_ctrl_enable(struct nvmet_subsys *subsys)
502 struct nvme_ctrl *ctrl;
507 mutex_lock(&subsys->lock);
508 if (!subsys->passthru_ctrl_path)
510 if (subsys->passthru_ctrl)
513 if (subsys->nr_namespaces) {
514 pr_info("cannot enable both passthru and regular namespaces for a single subsystem");
518 file = filp_open(subsys->passthru_ctrl_path, O_RDWR, 0);
524 ctrl = nvme_ctrl_from_file(file);
526 pr_err("failed to open nvme controller %s\n",
527 subsys->passthru_ctrl_path);
532 old = xa_cmpxchg(&passthru_subsystems, ctrl->cntlid, NULL,
534 if (xa_is_err(old)) {
542 subsys->passthru_ctrl = ctrl;
543 subsys->ver = ctrl->vs;
545 if (subsys->ver < NVME_VS(1, 2, 1)) {
546 pr_warn("nvme controller version is too old: %llu.%llu.%llu, advertising 1.2.1\n",
547 NVME_MAJOR(subsys->ver), NVME_MINOR(subsys->ver),
548 NVME_TERTIARY(subsys->ver));
549 subsys->ver = NVME_VS(1, 2, 1);
552 __module_get(subsys->passthru_ctrl->ops->module);
556 filp_close(file, NULL);
558 mutex_unlock(&subsys->lock);
562 static void __nvmet_passthru_ctrl_disable(struct nvmet_subsys *subsys)
564 if (subsys->passthru_ctrl) {
565 xa_erase(&passthru_subsystems, subsys->passthru_ctrl->cntlid);
566 module_put(subsys->passthru_ctrl->ops->module);
567 nvme_put_ctrl(subsys->passthru_ctrl);
569 subsys->passthru_ctrl = NULL;
570 subsys->ver = NVMET_DEFAULT_VS;
573 void nvmet_passthru_ctrl_disable(struct nvmet_subsys *subsys)
575 mutex_lock(&subsys->lock);
576 __nvmet_passthru_ctrl_disable(subsys);
577 mutex_unlock(&subsys->lock);
580 void nvmet_passthru_subsys_free(struct nvmet_subsys *subsys)
582 mutex_lock(&subsys->lock);
583 __nvmet_passthru_ctrl_disable(subsys);
584 mutex_unlock(&subsys->lock);
585 kfree(subsys->passthru_ctrl_path);