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 nvme_execute_passthru_rq(rq);
160 status = nvme_req(rq)->status;
161 if (status == NVME_SC_SUCCESS &&
162 req->cmd->common.opcode == nvme_admin_identify) {
163 switch (req->cmd->identify.cns) {
164 case NVME_ID_CNS_CTRL:
165 nvmet_passthru_override_id_ctrl(req);
168 nvmet_passthru_override_id_ns(req);
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 (req->transfer_len <= NVMET_MAX_INLINE_DATA_LEN) {
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 if (bio != &req->p.inline_bio)
215 blk_rq_bio_prep(rq, bio, req->sg_cnt);
220 static void nvmet_passthru_execute_cmd(struct nvmet_req *req)
222 struct nvme_ctrl *ctrl = nvmet_req_passthru_ctrl(req);
223 struct request_queue *q = ctrl->admin_q;
224 struct nvme_ns *ns = NULL;
225 struct request *rq = NULL;
226 unsigned int timeout;
231 if (likely(req->sq->qid != 0)) {
232 u32 nsid = le32_to_cpu(req->cmd->common.nsid);
234 ns = nvme_find_get_ns(ctrl, nsid);
236 pr_err("failed to get passthru ns nsid:%u\n", nsid);
237 status = NVME_SC_INVALID_NS | NVME_SC_DNR;
242 timeout = nvmet_req_subsys(req)->io_timeout;
244 timeout = nvmet_req_subsys(req)->admin_timeout;
247 rq = nvme_alloc_request(q, req->cmd, 0);
249 status = NVME_SC_INTERNAL;
254 rq->timeout = timeout;
257 ret = nvmet_passthru_map_sg(req, rq);
259 status = NVME_SC_INTERNAL;
265 * If there are effects for the command we are about to execute, or
266 * an end_req function we need to use nvme_execute_passthru_rq()
267 * synchronously in a work item seeing the end_req function and
268 * nvme_passthru_end() can't be called in the request done callback
269 * which is typically in interrupt context.
271 effects = nvme_command_effects(ctrl, ns, req->cmd->common.opcode);
272 if (req->p.use_workqueue || effects) {
273 INIT_WORK(&req->p.work, nvmet_passthru_execute_cmd_work);
275 schedule_work(&req->p.work);
277 rq->end_io_data = req;
278 blk_execute_rq_nowait(ns ? ns->disk : NULL, rq, 0,
279 nvmet_passthru_req_done);
288 blk_mq_free_request(rq);
293 nvmet_req_complete(req, status);
297 * We need to emulate set host behaviour to ensure that any requested
298 * behaviour of the target's host matches the requested behaviour
299 * of the device's host and fail otherwise.
301 static void nvmet_passthru_set_host_behaviour(struct nvmet_req *req)
303 struct nvme_ctrl *ctrl = nvmet_req_passthru_ctrl(req);
304 struct nvme_feat_host_behavior *host;
305 u16 status = NVME_SC_INTERNAL;
308 host = kzalloc(sizeof(*host) * 2, GFP_KERNEL);
310 goto out_complete_req;
312 ret = nvme_get_features(ctrl, NVME_FEAT_HOST_BEHAVIOR, 0,
313 host, sizeof(*host), NULL);
317 status = nvmet_copy_from_sgl(req, 0, &host[1], sizeof(*host));
321 if (memcmp(&host[0], &host[1], sizeof(host[0]))) {
322 pr_warn("target host has requested different behaviour from the local host\n");
323 status = NVME_SC_INTERNAL;
329 nvmet_req_complete(req, status);
332 static u16 nvmet_setup_passthru_command(struct nvmet_req *req)
334 req->p.use_workqueue = false;
335 req->execute = nvmet_passthru_execute_cmd;
336 return NVME_SC_SUCCESS;
339 u16 nvmet_parse_passthru_io_cmd(struct nvmet_req *req)
341 /* Reject any commands with non-sgl flags set (ie. fused commands) */
342 if (req->cmd->common.flags & ~NVME_CMD_SGL_ALL)
343 return NVME_SC_INVALID_FIELD;
345 switch (req->cmd->common.opcode) {
346 case nvme_cmd_resv_register:
347 case nvme_cmd_resv_report:
348 case nvme_cmd_resv_acquire:
349 case nvme_cmd_resv_release:
351 * Reservations cannot be supported properly because the
352 * underlying device has no way of differentiating different
353 * hosts that connect via fabrics. This could potentially be
354 * emulated in the future if regular targets grow support for
357 return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
360 return nvmet_setup_passthru_command(req);
364 * Only features that are emulated or specifically allowed in the list are
365 * passed down to the controller. This function implements the allow list for
366 * both get and set features.
368 static u16 nvmet_passthru_get_set_features(struct nvmet_req *req)
370 switch (le32_to_cpu(req->cmd->features.fid)) {
371 case NVME_FEAT_ARBITRATION:
372 case NVME_FEAT_POWER_MGMT:
373 case NVME_FEAT_LBA_RANGE:
374 case NVME_FEAT_TEMP_THRESH:
375 case NVME_FEAT_ERR_RECOVERY:
376 case NVME_FEAT_VOLATILE_WC:
377 case NVME_FEAT_WRITE_ATOMIC:
378 case NVME_FEAT_AUTO_PST:
379 case NVME_FEAT_TIMESTAMP:
381 case NVME_FEAT_NOPSC:
383 case NVME_FEAT_PLM_CONFIG:
384 case NVME_FEAT_PLM_WINDOW:
385 case NVME_FEAT_HOST_BEHAVIOR:
386 case NVME_FEAT_SANITIZE:
387 case NVME_FEAT_VENDOR_START ... NVME_FEAT_VENDOR_END:
388 return nvmet_setup_passthru_command(req);
390 case NVME_FEAT_ASYNC_EVENT:
391 /* There is no support for forwarding ASYNC events */
392 case NVME_FEAT_IRQ_COALESCE:
393 case NVME_FEAT_IRQ_CONFIG:
394 /* The IRQ settings will not apply to the target controller */
395 case NVME_FEAT_HOST_MEM_BUF:
397 * Any HMB that's set will not be passed through and will
398 * not work as expected
400 case NVME_FEAT_SW_PROGRESS:
402 * The Pre-Boot Software Load Count doesn't make much
403 * sense for a target to export
405 case NVME_FEAT_RESV_MASK:
406 case NVME_FEAT_RESV_PERSIST:
407 /* No reservations, see nvmet_parse_passthru_io_cmd() */
409 return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
413 u16 nvmet_parse_passthru_admin_cmd(struct nvmet_req *req)
415 /* Reject any commands with non-sgl flags set (ie. fused commands) */
416 if (req->cmd->common.flags & ~NVME_CMD_SGL_ALL)
417 return NVME_SC_INVALID_FIELD;
420 * Passthru all vendor specific commands
422 if (req->cmd->common.opcode >= nvme_admin_vendor_start)
423 return nvmet_setup_passthru_command(req);
425 switch (req->cmd->common.opcode) {
426 case nvme_admin_async_event:
427 req->execute = nvmet_execute_async_event;
428 return NVME_SC_SUCCESS;
429 case nvme_admin_keep_alive:
431 * Most PCIe ctrls don't support keep alive cmd, we route keep
432 * alive to the non-passthru mode. In future please change this
433 * code when PCIe ctrls with keep alive support available.
435 req->execute = nvmet_execute_keep_alive;
436 return NVME_SC_SUCCESS;
437 case nvme_admin_set_features:
438 switch (le32_to_cpu(req->cmd->features.fid)) {
439 case NVME_FEAT_ASYNC_EVENT:
441 case NVME_FEAT_NUM_QUEUES:
442 case NVME_FEAT_HOST_ID:
443 req->execute = nvmet_execute_set_features;
444 return NVME_SC_SUCCESS;
445 case NVME_FEAT_HOST_BEHAVIOR:
446 req->execute = nvmet_passthru_set_host_behaviour;
447 return NVME_SC_SUCCESS;
449 return nvmet_passthru_get_set_features(req);
452 case nvme_admin_get_features:
453 switch (le32_to_cpu(req->cmd->features.fid)) {
454 case NVME_FEAT_ASYNC_EVENT:
456 case NVME_FEAT_NUM_QUEUES:
457 case NVME_FEAT_HOST_ID:
458 req->execute = nvmet_execute_get_features;
459 return NVME_SC_SUCCESS;
461 return nvmet_passthru_get_set_features(req);
464 case nvme_admin_identify:
465 switch (req->cmd->identify.cns) {
466 case NVME_ID_CNS_CTRL:
467 req->execute = nvmet_passthru_execute_cmd;
468 req->p.use_workqueue = true;
469 return NVME_SC_SUCCESS;
470 case NVME_ID_CNS_CS_CTRL:
471 switch (req->cmd->identify.csi) {
473 req->execute = nvmet_passthru_execute_cmd;
474 req->p.use_workqueue = true;
475 return NVME_SC_SUCCESS;
477 return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
479 req->execute = nvmet_passthru_execute_cmd;
480 req->p.use_workqueue = true;
481 return NVME_SC_SUCCESS;
482 case NVME_ID_CNS_CS_NS:
483 switch (req->cmd->identify.csi) {
485 req->execute = nvmet_passthru_execute_cmd;
486 req->p.use_workqueue = true;
487 return NVME_SC_SUCCESS;
489 return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
491 return nvmet_setup_passthru_command(req);
493 case nvme_admin_get_log_page:
494 return nvmet_setup_passthru_command(req);
496 /* Reject commands not in the allowlist above */
497 return nvmet_report_invalid_opcode(req);
501 int nvmet_passthru_ctrl_enable(struct nvmet_subsys *subsys)
503 struct nvme_ctrl *ctrl;
508 mutex_lock(&subsys->lock);
509 if (!subsys->passthru_ctrl_path)
511 if (subsys->passthru_ctrl)
514 if (subsys->nr_namespaces) {
515 pr_info("cannot enable both passthru and regular namespaces for a single subsystem");
519 file = filp_open(subsys->passthru_ctrl_path, O_RDWR, 0);
525 ctrl = nvme_ctrl_from_file(file);
527 pr_err("failed to open nvme controller %s\n",
528 subsys->passthru_ctrl_path);
533 old = xa_cmpxchg(&passthru_subsystems, ctrl->cntlid, NULL,
535 if (xa_is_err(old)) {
543 subsys->passthru_ctrl = ctrl;
544 subsys->ver = ctrl->vs;
546 if (subsys->ver < NVME_VS(1, 2, 1)) {
547 pr_warn("nvme controller version is too old: %llu.%llu.%llu, advertising 1.2.1\n",
548 NVME_MAJOR(subsys->ver), NVME_MINOR(subsys->ver),
549 NVME_TERTIARY(subsys->ver));
550 subsys->ver = NVME_VS(1, 2, 1);
553 __module_get(subsys->passthru_ctrl->ops->module);
557 filp_close(file, NULL);
559 mutex_unlock(&subsys->lock);
563 static void __nvmet_passthru_ctrl_disable(struct nvmet_subsys *subsys)
565 if (subsys->passthru_ctrl) {
566 xa_erase(&passthru_subsystems, subsys->passthru_ctrl->cntlid);
567 module_put(subsys->passthru_ctrl->ops->module);
568 nvme_put_ctrl(subsys->passthru_ctrl);
570 subsys->passthru_ctrl = NULL;
571 subsys->ver = NVMET_DEFAULT_VS;
574 void nvmet_passthru_ctrl_disable(struct nvmet_subsys *subsys)
576 mutex_lock(&subsys->lock);
577 __nvmet_passthru_ctrl_disable(subsys);
578 mutex_unlock(&subsys->lock);
581 void nvmet_passthru_subsys_free(struct nvmet_subsys *subsys)
583 mutex_lock(&subsys->lock);
584 __nvmet_passthru_ctrl_disable(subsys);
585 mutex_unlock(&subsys->lock);
586 kfree(subsys->passthru_ctrl_path);