1 // SPDX-License-Identifier: GPL-2.0
3 * Common code for the NVMe target.
4 * Copyright (c) 2015-2016 HGST, a Western Digital Company.
6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7 #include <linux/module.h>
8 #include <linux/random.h>
9 #include <linux/rculist.h>
10 #include <linux/pci-p2pdma.h>
11 #include <linux/scatterlist.h>
13 #define CREATE_TRACE_POINTS
18 struct workqueue_struct *buffered_io_wq;
19 static const struct nvmet_fabrics_ops *nvmet_transports[NVMF_TRTYPE_MAX];
20 static DEFINE_IDA(cntlid_ida);
23 * This read/write semaphore is used to synchronize access to configuration
24 * information on a target system that will result in discovery log page
25 * information change for at least one host.
26 * The full list of resources to protected by this semaphore is:
29 * - per-subsystem allowed hosts list
30 * - allow_any_host subsystem attribute
32 * - the nvmet_transports array
34 * When updating any of those lists/structures write lock should be obtained,
35 * while when reading (popolating discovery log page or checking host-subsystem
36 * link) read lock is obtained to allow concurrent reads.
38 DECLARE_RWSEM(nvmet_config_sem);
40 u32 nvmet_ana_group_enabled[NVMET_MAX_ANAGRPS + 1];
42 DECLARE_RWSEM(nvmet_ana_sem);
44 inline u16 errno_to_nvme_status(struct nvmet_req *req, int errno)
50 status = NVME_SC_SUCCESS;
53 req->error_loc = offsetof(struct nvme_rw_command, length);
54 status = NVME_SC_CAP_EXCEEDED | NVME_SC_DNR;
57 req->error_loc = offsetof(struct nvme_rw_command, slba);
58 status = NVME_SC_LBA_RANGE | NVME_SC_DNR;
61 req->error_loc = offsetof(struct nvme_common_command, opcode);
62 switch (req->cmd->common.opcode) {
64 case nvme_cmd_write_zeroes:
65 status = NVME_SC_ONCS_NOT_SUPPORTED | NVME_SC_DNR;
68 status = NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
72 req->error_loc = offsetof(struct nvme_rw_command, nsid);
73 status = NVME_SC_ACCESS_DENIED;
78 req->error_loc = offsetof(struct nvme_common_command, opcode);
79 status = NVME_SC_INTERNAL | NVME_SC_DNR;
85 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
86 const char *subsysnqn);
88 u16 nvmet_copy_to_sgl(struct nvmet_req *req, off_t off, const void *buf,
91 if (sg_pcopy_from_buffer(req->sg, req->sg_cnt, buf, len, off) != len) {
92 req->error_loc = offsetof(struct nvme_common_command, dptr);
93 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
98 u16 nvmet_copy_from_sgl(struct nvmet_req *req, off_t off, void *buf, size_t len)
100 if (sg_pcopy_to_buffer(req->sg, req->sg_cnt, buf, len, off) != len) {
101 req->error_loc = offsetof(struct nvme_common_command, dptr);
102 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
107 u16 nvmet_zero_sgl(struct nvmet_req *req, off_t off, size_t len)
109 if (sg_zero_buffer(req->sg, req->sg_cnt, len, off) != len) {
110 req->error_loc = offsetof(struct nvme_common_command, dptr);
111 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
116 static unsigned int nvmet_max_nsid(struct nvmet_subsys *subsys)
120 if (list_empty(&subsys->namespaces))
123 ns = list_last_entry(&subsys->namespaces, struct nvmet_ns, dev_link);
127 static u32 nvmet_async_event_result(struct nvmet_async_event *aen)
129 return aen->event_type | (aen->event_info << 8) | (aen->log_page << 16);
132 static void nvmet_async_events_free(struct nvmet_ctrl *ctrl)
134 struct nvmet_req *req;
137 mutex_lock(&ctrl->lock);
138 if (!ctrl->nr_async_event_cmds) {
139 mutex_unlock(&ctrl->lock);
143 req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
144 mutex_unlock(&ctrl->lock);
145 nvmet_req_complete(req, NVME_SC_INTERNAL | NVME_SC_DNR);
149 static void nvmet_async_event_work(struct work_struct *work)
151 struct nvmet_ctrl *ctrl =
152 container_of(work, struct nvmet_ctrl, async_event_work);
153 struct nvmet_async_event *aen;
154 struct nvmet_req *req;
157 mutex_lock(&ctrl->lock);
158 aen = list_first_entry_or_null(&ctrl->async_events,
159 struct nvmet_async_event, entry);
160 if (!aen || !ctrl->nr_async_event_cmds) {
161 mutex_unlock(&ctrl->lock);
165 req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
166 nvmet_set_result(req, nvmet_async_event_result(aen));
168 list_del(&aen->entry);
171 mutex_unlock(&ctrl->lock);
172 nvmet_req_complete(req, 0);
176 void nvmet_add_async_event(struct nvmet_ctrl *ctrl, u8 event_type,
177 u8 event_info, u8 log_page)
179 struct nvmet_async_event *aen;
181 aen = kmalloc(sizeof(*aen), GFP_KERNEL);
185 aen->event_type = event_type;
186 aen->event_info = event_info;
187 aen->log_page = log_page;
189 mutex_lock(&ctrl->lock);
190 list_add_tail(&aen->entry, &ctrl->async_events);
191 mutex_unlock(&ctrl->lock);
193 schedule_work(&ctrl->async_event_work);
196 static void nvmet_add_to_changed_ns_log(struct nvmet_ctrl *ctrl, __le32 nsid)
200 mutex_lock(&ctrl->lock);
201 if (ctrl->nr_changed_ns > NVME_MAX_CHANGED_NAMESPACES)
204 for (i = 0; i < ctrl->nr_changed_ns; i++) {
205 if (ctrl->changed_ns_list[i] == nsid)
209 if (ctrl->nr_changed_ns == NVME_MAX_CHANGED_NAMESPACES) {
210 ctrl->changed_ns_list[0] = cpu_to_le32(0xffffffff);
211 ctrl->nr_changed_ns = U32_MAX;
215 ctrl->changed_ns_list[ctrl->nr_changed_ns++] = nsid;
217 mutex_unlock(&ctrl->lock);
220 void nvmet_ns_changed(struct nvmet_subsys *subsys, u32 nsid)
222 struct nvmet_ctrl *ctrl;
224 lockdep_assert_held(&subsys->lock);
226 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
227 nvmet_add_to_changed_ns_log(ctrl, cpu_to_le32(nsid));
228 if (nvmet_aen_bit_disabled(ctrl, NVME_AEN_BIT_NS_ATTR))
230 nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE,
231 NVME_AER_NOTICE_NS_CHANGED,
232 NVME_LOG_CHANGED_NS);
236 void nvmet_send_ana_event(struct nvmet_subsys *subsys,
237 struct nvmet_port *port)
239 struct nvmet_ctrl *ctrl;
241 mutex_lock(&subsys->lock);
242 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
243 if (port && ctrl->port != port)
245 if (nvmet_aen_bit_disabled(ctrl, NVME_AEN_BIT_ANA_CHANGE))
247 nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE,
248 NVME_AER_NOTICE_ANA, NVME_LOG_ANA);
250 mutex_unlock(&subsys->lock);
253 void nvmet_port_send_ana_event(struct nvmet_port *port)
255 struct nvmet_subsys_link *p;
257 down_read(&nvmet_config_sem);
258 list_for_each_entry(p, &port->subsystems, entry)
259 nvmet_send_ana_event(p->subsys, port);
260 up_read(&nvmet_config_sem);
263 int nvmet_register_transport(const struct nvmet_fabrics_ops *ops)
267 down_write(&nvmet_config_sem);
268 if (nvmet_transports[ops->type])
271 nvmet_transports[ops->type] = ops;
272 up_write(&nvmet_config_sem);
276 EXPORT_SYMBOL_GPL(nvmet_register_transport);
278 void nvmet_unregister_transport(const struct nvmet_fabrics_ops *ops)
280 down_write(&nvmet_config_sem);
281 nvmet_transports[ops->type] = NULL;
282 up_write(&nvmet_config_sem);
284 EXPORT_SYMBOL_GPL(nvmet_unregister_transport);
286 void nvmet_port_del_ctrls(struct nvmet_port *port, struct nvmet_subsys *subsys)
288 struct nvmet_ctrl *ctrl;
290 mutex_lock(&subsys->lock);
291 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
292 if (ctrl->port == port)
293 ctrl->ops->delete_ctrl(ctrl);
295 mutex_unlock(&subsys->lock);
298 int nvmet_enable_port(struct nvmet_port *port)
300 const struct nvmet_fabrics_ops *ops;
303 lockdep_assert_held(&nvmet_config_sem);
305 ops = nvmet_transports[port->disc_addr.trtype];
307 up_write(&nvmet_config_sem);
308 request_module("nvmet-transport-%d", port->disc_addr.trtype);
309 down_write(&nvmet_config_sem);
310 ops = nvmet_transports[port->disc_addr.trtype];
312 pr_err("transport type %d not supported\n",
313 port->disc_addr.trtype);
318 if (!try_module_get(ops->owner))
321 ret = ops->add_port(port);
323 module_put(ops->owner);
327 /* If the transport didn't set inline_data_size, then disable it. */
328 if (port->inline_data_size < 0)
329 port->inline_data_size = 0;
331 port->enabled = true;
336 void nvmet_disable_port(struct nvmet_port *port)
338 const struct nvmet_fabrics_ops *ops;
340 lockdep_assert_held(&nvmet_config_sem);
342 port->enabled = false;
345 ops = nvmet_transports[port->disc_addr.trtype];
346 ops->remove_port(port);
347 module_put(ops->owner);
350 static void nvmet_keep_alive_timer(struct work_struct *work)
352 struct nvmet_ctrl *ctrl = container_of(to_delayed_work(work),
353 struct nvmet_ctrl, ka_work);
354 bool cmd_seen = ctrl->cmd_seen;
356 ctrl->cmd_seen = false;
358 pr_debug("ctrl %d reschedule traffic based keep-alive timer\n",
360 schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
364 pr_err("ctrl %d keep-alive timer (%d seconds) expired!\n",
365 ctrl->cntlid, ctrl->kato);
367 nvmet_ctrl_fatal_error(ctrl);
370 static void nvmet_start_keep_alive_timer(struct nvmet_ctrl *ctrl)
372 pr_debug("ctrl %d start keep-alive timer for %d secs\n",
373 ctrl->cntlid, ctrl->kato);
375 INIT_DELAYED_WORK(&ctrl->ka_work, nvmet_keep_alive_timer);
376 schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
379 static void nvmet_stop_keep_alive_timer(struct nvmet_ctrl *ctrl)
381 pr_debug("ctrl %d stop keep-alive\n", ctrl->cntlid);
383 cancel_delayed_work_sync(&ctrl->ka_work);
386 static struct nvmet_ns *__nvmet_find_namespace(struct nvmet_ctrl *ctrl,
391 list_for_each_entry_rcu(ns, &ctrl->subsys->namespaces, dev_link) {
392 if (ns->nsid == le32_to_cpu(nsid))
399 struct nvmet_ns *nvmet_find_namespace(struct nvmet_ctrl *ctrl, __le32 nsid)
404 ns = __nvmet_find_namespace(ctrl, nsid);
406 percpu_ref_get(&ns->ref);
412 static void nvmet_destroy_namespace(struct percpu_ref *ref)
414 struct nvmet_ns *ns = container_of(ref, struct nvmet_ns, ref);
416 complete(&ns->disable_done);
419 void nvmet_put_namespace(struct nvmet_ns *ns)
421 percpu_ref_put(&ns->ref);
424 static void nvmet_ns_dev_disable(struct nvmet_ns *ns)
426 nvmet_bdev_ns_disable(ns);
427 nvmet_file_ns_disable(ns);
430 static int nvmet_p2pmem_ns_enable(struct nvmet_ns *ns)
433 struct pci_dev *p2p_dev;
439 pr_err("peer-to-peer DMA is not supported by non-block device namespaces\n");
443 if (!blk_queue_pci_p2pdma(ns->bdev->bd_queue)) {
444 pr_err("peer-to-peer DMA is not supported by the driver of %s\n",
450 ret = pci_p2pdma_distance(ns->p2p_dev, nvmet_ns_dev(ns), true);
455 * Right now we just check that there is p2pmem available so
456 * we can report an error to the user right away if there
457 * is not. We'll find the actual device to use once we
458 * setup the controller when the port's device is available.
461 p2p_dev = pci_p2pmem_find(nvmet_ns_dev(ns));
463 pr_err("no peer-to-peer memory is available for %s\n",
468 pci_dev_put(p2p_dev);
475 * Note: ctrl->subsys->lock should be held when calling this function
477 static void nvmet_p2pmem_ns_add_p2p(struct nvmet_ctrl *ctrl,
480 struct device *clients[2];
481 struct pci_dev *p2p_dev;
484 if (!ctrl->p2p_client || !ns->use_p2pmem)
488 ret = pci_p2pdma_distance(ns->p2p_dev, ctrl->p2p_client, true);
492 p2p_dev = pci_dev_get(ns->p2p_dev);
494 clients[0] = ctrl->p2p_client;
495 clients[1] = nvmet_ns_dev(ns);
497 p2p_dev = pci_p2pmem_find_many(clients, ARRAY_SIZE(clients));
499 pr_err("no peer-to-peer memory is available that's supported by %s and %s\n",
500 dev_name(ctrl->p2p_client), ns->device_path);
505 ret = radix_tree_insert(&ctrl->p2p_ns_map, ns->nsid, p2p_dev);
507 pci_dev_put(p2p_dev);
509 pr_info("using p2pmem on %s for nsid %d\n", pci_name(p2p_dev),
513 int nvmet_ns_enable(struct nvmet_ns *ns)
515 struct nvmet_subsys *subsys = ns->subsys;
516 struct nvmet_ctrl *ctrl;
519 mutex_lock(&subsys->lock);
525 if (subsys->nr_namespaces == NVMET_MAX_NAMESPACES)
528 ret = nvmet_bdev_ns_enable(ns);
530 ret = nvmet_file_ns_enable(ns);
534 ret = nvmet_p2pmem_ns_enable(ns);
536 goto out_dev_disable;
538 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
539 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
541 ret = percpu_ref_init(&ns->ref, nvmet_destroy_namespace,
546 if (ns->nsid > subsys->max_nsid)
547 subsys->max_nsid = ns->nsid;
550 * The namespaces list needs to be sorted to simplify the implementation
551 * of the Identify Namepace List subcommand.
553 if (list_empty(&subsys->namespaces)) {
554 list_add_tail_rcu(&ns->dev_link, &subsys->namespaces);
556 struct nvmet_ns *old;
558 list_for_each_entry_rcu(old, &subsys->namespaces, dev_link,
559 lockdep_is_held(&subsys->lock)) {
560 BUG_ON(ns->nsid == old->nsid);
561 if (ns->nsid < old->nsid)
565 list_add_tail_rcu(&ns->dev_link, &old->dev_link);
567 subsys->nr_namespaces++;
569 nvmet_ns_changed(subsys, ns->nsid);
573 mutex_unlock(&subsys->lock);
576 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
577 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
579 nvmet_ns_dev_disable(ns);
583 void nvmet_ns_disable(struct nvmet_ns *ns)
585 struct nvmet_subsys *subsys = ns->subsys;
586 struct nvmet_ctrl *ctrl;
588 mutex_lock(&subsys->lock);
593 list_del_rcu(&ns->dev_link);
594 if (ns->nsid == subsys->max_nsid)
595 subsys->max_nsid = nvmet_max_nsid(subsys);
597 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
598 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
600 mutex_unlock(&subsys->lock);
603 * Now that we removed the namespaces from the lookup list, we
604 * can kill the per_cpu ref and wait for any remaining references
605 * to be dropped, as well as a RCU grace period for anyone only
606 * using the namepace under rcu_read_lock(). Note that we can't
607 * use call_rcu here as we need to ensure the namespaces have
608 * been fully destroyed before unloading the module.
610 percpu_ref_kill(&ns->ref);
612 wait_for_completion(&ns->disable_done);
613 percpu_ref_exit(&ns->ref);
615 mutex_lock(&subsys->lock);
617 subsys->nr_namespaces--;
618 nvmet_ns_changed(subsys, ns->nsid);
619 nvmet_ns_dev_disable(ns);
621 mutex_unlock(&subsys->lock);
624 void nvmet_ns_free(struct nvmet_ns *ns)
626 nvmet_ns_disable(ns);
628 down_write(&nvmet_ana_sem);
629 nvmet_ana_group_enabled[ns->anagrpid]--;
630 up_write(&nvmet_ana_sem);
632 kfree(ns->device_path);
636 struct nvmet_ns *nvmet_ns_alloc(struct nvmet_subsys *subsys, u32 nsid)
640 ns = kzalloc(sizeof(*ns), GFP_KERNEL);
644 INIT_LIST_HEAD(&ns->dev_link);
645 init_completion(&ns->disable_done);
650 down_write(&nvmet_ana_sem);
651 ns->anagrpid = NVMET_DEFAULT_ANA_GRPID;
652 nvmet_ana_group_enabled[ns->anagrpid]++;
653 up_write(&nvmet_ana_sem);
656 ns->buffered_io = false;
661 static void nvmet_update_sq_head(struct nvmet_req *req)
664 u32 old_sqhd, new_sqhd;
667 old_sqhd = req->sq->sqhd;
668 new_sqhd = (old_sqhd + 1) % req->sq->size;
669 } while (cmpxchg(&req->sq->sqhd, old_sqhd, new_sqhd) !=
672 req->cqe->sq_head = cpu_to_le16(req->sq->sqhd & 0x0000FFFF);
675 static void nvmet_set_error(struct nvmet_req *req, u16 status)
677 struct nvmet_ctrl *ctrl = req->sq->ctrl;
678 struct nvme_error_slot *new_error_slot;
681 req->cqe->status = cpu_to_le16(status << 1);
683 if (!ctrl || req->error_loc == NVMET_NO_ERROR_LOC)
686 spin_lock_irqsave(&ctrl->error_lock, flags);
689 &ctrl->slots[ctrl->err_counter % NVMET_ERROR_LOG_SLOTS];
691 new_error_slot->error_count = cpu_to_le64(ctrl->err_counter);
692 new_error_slot->sqid = cpu_to_le16(req->sq->qid);
693 new_error_slot->cmdid = cpu_to_le16(req->cmd->common.command_id);
694 new_error_slot->status_field = cpu_to_le16(status << 1);
695 new_error_slot->param_error_location = cpu_to_le16(req->error_loc);
696 new_error_slot->lba = cpu_to_le64(req->error_slba);
697 new_error_slot->nsid = req->cmd->common.nsid;
698 spin_unlock_irqrestore(&ctrl->error_lock, flags);
700 /* set the more bit for this request */
701 req->cqe->status |= cpu_to_le16(1 << 14);
704 static void __nvmet_req_complete(struct nvmet_req *req, u16 status)
706 if (!req->sq->sqhd_disabled)
707 nvmet_update_sq_head(req);
708 req->cqe->sq_id = cpu_to_le16(req->sq->qid);
709 req->cqe->command_id = req->cmd->common.command_id;
711 if (unlikely(status))
712 nvmet_set_error(req, status);
714 trace_nvmet_req_complete(req);
717 nvmet_put_namespace(req->ns);
718 req->ops->queue_response(req);
721 void nvmet_req_complete(struct nvmet_req *req, u16 status)
723 __nvmet_req_complete(req, status);
724 percpu_ref_put(&req->sq->ref);
726 EXPORT_SYMBOL_GPL(nvmet_req_complete);
728 void nvmet_cq_setup(struct nvmet_ctrl *ctrl, struct nvmet_cq *cq,
737 void nvmet_sq_setup(struct nvmet_ctrl *ctrl, struct nvmet_sq *sq,
747 static void nvmet_confirm_sq(struct percpu_ref *ref)
749 struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
751 complete(&sq->confirm_done);
754 void nvmet_sq_destroy(struct nvmet_sq *sq)
757 * If this is the admin queue, complete all AERs so that our
758 * queue doesn't have outstanding requests on it.
760 if (sq->ctrl && sq->ctrl->sqs && sq->ctrl->sqs[0] == sq)
761 nvmet_async_events_free(sq->ctrl);
762 percpu_ref_kill_and_confirm(&sq->ref, nvmet_confirm_sq);
763 wait_for_completion(&sq->confirm_done);
764 wait_for_completion(&sq->free_done);
765 percpu_ref_exit(&sq->ref);
768 nvmet_ctrl_put(sq->ctrl);
769 sq->ctrl = NULL; /* allows reusing the queue later */
772 EXPORT_SYMBOL_GPL(nvmet_sq_destroy);
774 static void nvmet_sq_free(struct percpu_ref *ref)
776 struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
778 complete(&sq->free_done);
781 int nvmet_sq_init(struct nvmet_sq *sq)
785 ret = percpu_ref_init(&sq->ref, nvmet_sq_free, 0, GFP_KERNEL);
787 pr_err("percpu_ref init failed!\n");
790 init_completion(&sq->free_done);
791 init_completion(&sq->confirm_done);
795 EXPORT_SYMBOL_GPL(nvmet_sq_init);
797 static inline u16 nvmet_check_ana_state(struct nvmet_port *port,
800 enum nvme_ana_state state = port->ana_state[ns->anagrpid];
802 if (unlikely(state == NVME_ANA_INACCESSIBLE))
803 return NVME_SC_ANA_INACCESSIBLE;
804 if (unlikely(state == NVME_ANA_PERSISTENT_LOSS))
805 return NVME_SC_ANA_PERSISTENT_LOSS;
806 if (unlikely(state == NVME_ANA_CHANGE))
807 return NVME_SC_ANA_TRANSITION;
811 static inline u16 nvmet_io_cmd_check_access(struct nvmet_req *req)
813 if (unlikely(req->ns->readonly)) {
814 switch (req->cmd->common.opcode) {
819 return NVME_SC_NS_WRITE_PROTECTED;
826 static u16 nvmet_parse_io_cmd(struct nvmet_req *req)
828 struct nvme_command *cmd = req->cmd;
831 ret = nvmet_check_ctrl_status(req, cmd);
835 req->ns = nvmet_find_namespace(req->sq->ctrl, cmd->rw.nsid);
836 if (unlikely(!req->ns)) {
837 req->error_loc = offsetof(struct nvme_common_command, nsid);
838 return NVME_SC_INVALID_NS | NVME_SC_DNR;
840 ret = nvmet_check_ana_state(req->port, req->ns);
842 req->error_loc = offsetof(struct nvme_common_command, nsid);
845 ret = nvmet_io_cmd_check_access(req);
847 req->error_loc = offsetof(struct nvme_common_command, nsid);
852 return nvmet_file_parse_io_cmd(req);
854 return nvmet_bdev_parse_io_cmd(req);
857 bool nvmet_req_init(struct nvmet_req *req, struct nvmet_cq *cq,
858 struct nvmet_sq *sq, const struct nvmet_fabrics_ops *ops)
860 u8 flags = req->cmd->common.flags;
868 req->transfer_len = 0;
869 req->cqe->status = 0;
870 req->cqe->sq_head = 0;
872 req->error_loc = NVMET_NO_ERROR_LOC;
875 trace_nvmet_req_init(req, req->cmd);
877 /* no support for fused commands yet */
878 if (unlikely(flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND))) {
879 req->error_loc = offsetof(struct nvme_common_command, flags);
880 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
885 * For fabrics, PSDT field shall describe metadata pointer (MPTR) that
886 * contains an address of a single contiguous physical buffer that is
889 if (unlikely((flags & NVME_CMD_SGL_ALL) != NVME_CMD_SGL_METABUF)) {
890 req->error_loc = offsetof(struct nvme_common_command, flags);
891 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
895 if (unlikely(!req->sq->ctrl))
896 /* will return an error for any non-connect command: */
897 status = nvmet_parse_connect_cmd(req);
898 else if (likely(req->sq->qid != 0))
899 status = nvmet_parse_io_cmd(req);
901 status = nvmet_parse_admin_cmd(req);
906 if (unlikely(!percpu_ref_tryget_live(&sq->ref))) {
907 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
912 sq->ctrl->cmd_seen = true;
917 __nvmet_req_complete(req, status);
920 EXPORT_SYMBOL_GPL(nvmet_req_init);
922 void nvmet_req_uninit(struct nvmet_req *req)
924 percpu_ref_put(&req->sq->ref);
926 nvmet_put_namespace(req->ns);
928 EXPORT_SYMBOL_GPL(nvmet_req_uninit);
930 bool nvmet_check_data_len(struct nvmet_req *req, size_t data_len)
932 if (unlikely(data_len != req->transfer_len)) {
933 req->error_loc = offsetof(struct nvme_common_command, dptr);
934 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
940 EXPORT_SYMBOL_GPL(nvmet_check_data_len);
942 bool nvmet_check_data_len_lte(struct nvmet_req *req, size_t data_len)
944 if (unlikely(data_len > req->transfer_len)) {
945 req->error_loc = offsetof(struct nvme_common_command, dptr);
946 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
953 int nvmet_req_alloc_sgl(struct nvmet_req *req)
955 struct pci_dev *p2p_dev = NULL;
957 if (IS_ENABLED(CONFIG_PCI_P2PDMA)) {
958 if (req->sq->ctrl && req->ns)
959 p2p_dev = radix_tree_lookup(&req->sq->ctrl->p2p_ns_map,
963 if (req->sq->qid && p2p_dev) {
964 req->sg = pci_p2pmem_alloc_sgl(p2p_dev, &req->sg_cnt,
967 req->p2p_dev = p2p_dev;
973 * If no P2P memory was available we fallback to using
978 req->sg = sgl_alloc(req->transfer_len, GFP_KERNEL, &req->sg_cnt);
979 if (unlikely(!req->sg))
984 EXPORT_SYMBOL_GPL(nvmet_req_alloc_sgl);
986 void nvmet_req_free_sgl(struct nvmet_req *req)
989 pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
996 EXPORT_SYMBOL_GPL(nvmet_req_free_sgl);
998 static inline bool nvmet_cc_en(u32 cc)
1000 return (cc >> NVME_CC_EN_SHIFT) & 0x1;
1003 static inline u8 nvmet_cc_css(u32 cc)
1005 return (cc >> NVME_CC_CSS_SHIFT) & 0x7;
1008 static inline u8 nvmet_cc_mps(u32 cc)
1010 return (cc >> NVME_CC_MPS_SHIFT) & 0xf;
1013 static inline u8 nvmet_cc_ams(u32 cc)
1015 return (cc >> NVME_CC_AMS_SHIFT) & 0x7;
1018 static inline u8 nvmet_cc_shn(u32 cc)
1020 return (cc >> NVME_CC_SHN_SHIFT) & 0x3;
1023 static inline u8 nvmet_cc_iosqes(u32 cc)
1025 return (cc >> NVME_CC_IOSQES_SHIFT) & 0xf;
1028 static inline u8 nvmet_cc_iocqes(u32 cc)
1030 return (cc >> NVME_CC_IOCQES_SHIFT) & 0xf;
1033 static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl)
1035 lockdep_assert_held(&ctrl->lock);
1037 if (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES ||
1038 nvmet_cc_iocqes(ctrl->cc) != NVME_NVM_IOCQES ||
1039 nvmet_cc_mps(ctrl->cc) != 0 ||
1040 nvmet_cc_ams(ctrl->cc) != 0 ||
1041 nvmet_cc_css(ctrl->cc) != 0) {
1042 ctrl->csts = NVME_CSTS_CFS;
1046 ctrl->csts = NVME_CSTS_RDY;
1049 * Controllers that are not yet enabled should not really enforce the
1050 * keep alive timeout, but we still want to track a timeout and cleanup
1051 * in case a host died before it enabled the controller. Hence, simply
1052 * reset the keep alive timer when the controller is enabled.
1054 mod_delayed_work(system_wq, &ctrl->ka_work, ctrl->kato * HZ);
1057 static void nvmet_clear_ctrl(struct nvmet_ctrl *ctrl)
1059 lockdep_assert_held(&ctrl->lock);
1061 /* XXX: tear down queues? */
1062 ctrl->csts &= ~NVME_CSTS_RDY;
1066 void nvmet_update_cc(struct nvmet_ctrl *ctrl, u32 new)
1070 mutex_lock(&ctrl->lock);
1074 if (nvmet_cc_en(new) && !nvmet_cc_en(old))
1075 nvmet_start_ctrl(ctrl);
1076 if (!nvmet_cc_en(new) && nvmet_cc_en(old))
1077 nvmet_clear_ctrl(ctrl);
1078 if (nvmet_cc_shn(new) && !nvmet_cc_shn(old)) {
1079 nvmet_clear_ctrl(ctrl);
1080 ctrl->csts |= NVME_CSTS_SHST_CMPLT;
1082 if (!nvmet_cc_shn(new) && nvmet_cc_shn(old))
1083 ctrl->csts &= ~NVME_CSTS_SHST_CMPLT;
1084 mutex_unlock(&ctrl->lock);
1087 static void nvmet_init_cap(struct nvmet_ctrl *ctrl)
1089 /* command sets supported: NVMe command set: */
1090 ctrl->cap = (1ULL << 37);
1091 /* CC.EN timeout in 500msec units: */
1092 ctrl->cap |= (15ULL << 24);
1093 /* maximum queue entries supported: */
1094 ctrl->cap |= NVMET_QUEUE_SIZE - 1;
1097 u16 nvmet_ctrl_find_get(const char *subsysnqn, const char *hostnqn, u16 cntlid,
1098 struct nvmet_req *req, struct nvmet_ctrl **ret)
1100 struct nvmet_subsys *subsys;
1101 struct nvmet_ctrl *ctrl;
1104 subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1106 pr_warn("connect request for invalid subsystem %s!\n",
1108 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1109 return NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1112 mutex_lock(&subsys->lock);
1113 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
1114 if (ctrl->cntlid == cntlid) {
1115 if (strncmp(hostnqn, ctrl->hostnqn, NVMF_NQN_SIZE)) {
1116 pr_warn("hostnqn mismatch.\n");
1119 if (!kref_get_unless_zero(&ctrl->ref))
1127 pr_warn("could not find controller %d for subsys %s / host %s\n",
1128 cntlid, subsysnqn, hostnqn);
1129 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(cntlid);
1130 status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1133 mutex_unlock(&subsys->lock);
1134 nvmet_subsys_put(subsys);
1138 u16 nvmet_check_ctrl_status(struct nvmet_req *req, struct nvme_command *cmd)
1140 if (unlikely(!(req->sq->ctrl->cc & NVME_CC_ENABLE))) {
1141 pr_err("got cmd %d while CC.EN == 0 on qid = %d\n",
1142 cmd->common.opcode, req->sq->qid);
1143 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1146 if (unlikely(!(req->sq->ctrl->csts & NVME_CSTS_RDY))) {
1147 pr_err("got cmd %d while CSTS.RDY == 0 on qid = %d\n",
1148 cmd->common.opcode, req->sq->qid);
1149 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1154 bool nvmet_host_allowed(struct nvmet_subsys *subsys, const char *hostnqn)
1156 struct nvmet_host_link *p;
1158 lockdep_assert_held(&nvmet_config_sem);
1160 if (subsys->allow_any_host)
1163 if (subsys->type == NVME_NQN_DISC) /* allow all access to disc subsys */
1166 list_for_each_entry(p, &subsys->hosts, entry) {
1167 if (!strcmp(nvmet_host_name(p->host), hostnqn))
1175 * Note: ctrl->subsys->lock should be held when calling this function
1177 static void nvmet_setup_p2p_ns_map(struct nvmet_ctrl *ctrl,
1178 struct nvmet_req *req)
1180 struct nvmet_ns *ns;
1182 if (!req->p2p_client)
1185 ctrl->p2p_client = get_device(req->p2p_client);
1187 list_for_each_entry_rcu(ns, &ctrl->subsys->namespaces, dev_link,
1188 lockdep_is_held(&ctrl->subsys->lock))
1189 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
1193 * Note: ctrl->subsys->lock should be held when calling this function
1195 static void nvmet_release_p2p_ns_map(struct nvmet_ctrl *ctrl)
1197 struct radix_tree_iter iter;
1200 radix_tree_for_each_slot(slot, &ctrl->p2p_ns_map, &iter, 0)
1201 pci_dev_put(radix_tree_deref_slot(slot));
1203 put_device(ctrl->p2p_client);
1206 static void nvmet_fatal_error_handler(struct work_struct *work)
1208 struct nvmet_ctrl *ctrl =
1209 container_of(work, struct nvmet_ctrl, fatal_err_work);
1211 pr_err("ctrl %d fatal error occurred!\n", ctrl->cntlid);
1212 ctrl->ops->delete_ctrl(ctrl);
1215 u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn,
1216 struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp)
1218 struct nvmet_subsys *subsys;
1219 struct nvmet_ctrl *ctrl;
1223 status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1224 subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1226 pr_warn("connect request for invalid subsystem %s!\n",
1228 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1232 status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1233 down_read(&nvmet_config_sem);
1234 if (!nvmet_host_allowed(subsys, hostnqn)) {
1235 pr_info("connect by host %s for subsystem %s not allowed\n",
1236 hostnqn, subsysnqn);
1237 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(hostnqn);
1238 up_read(&nvmet_config_sem);
1239 status = NVME_SC_CONNECT_INVALID_HOST | NVME_SC_DNR;
1240 goto out_put_subsystem;
1242 up_read(&nvmet_config_sem);
1244 status = NVME_SC_INTERNAL;
1245 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1247 goto out_put_subsystem;
1248 mutex_init(&ctrl->lock);
1250 nvmet_init_cap(ctrl);
1252 ctrl->port = req->port;
1254 INIT_WORK(&ctrl->async_event_work, nvmet_async_event_work);
1255 INIT_LIST_HEAD(&ctrl->async_events);
1256 INIT_RADIX_TREE(&ctrl->p2p_ns_map, GFP_KERNEL);
1257 INIT_WORK(&ctrl->fatal_err_work, nvmet_fatal_error_handler);
1259 memcpy(ctrl->subsysnqn, subsysnqn, NVMF_NQN_SIZE);
1260 memcpy(ctrl->hostnqn, hostnqn, NVMF_NQN_SIZE);
1262 kref_init(&ctrl->ref);
1263 ctrl->subsys = subsys;
1264 WRITE_ONCE(ctrl->aen_enabled, NVMET_AEN_CFG_OPTIONAL);
1266 ctrl->changed_ns_list = kmalloc_array(NVME_MAX_CHANGED_NAMESPACES,
1267 sizeof(__le32), GFP_KERNEL);
1268 if (!ctrl->changed_ns_list)
1271 ctrl->cqs = kcalloc(subsys->max_qid + 1,
1272 sizeof(struct nvmet_cq *),
1275 goto out_free_changed_ns_list;
1277 ctrl->sqs = kcalloc(subsys->max_qid + 1,
1278 sizeof(struct nvmet_sq *),
1283 ret = ida_simple_get(&cntlid_ida,
1284 NVME_CNTLID_MIN, NVME_CNTLID_MAX,
1287 status = NVME_SC_CONNECT_CTRL_BUSY | NVME_SC_DNR;
1292 ctrl->ops = req->ops;
1295 * Discovery controllers may use some arbitrary high value
1296 * in order to cleanup stale discovery sessions
1298 if ((ctrl->subsys->type == NVME_NQN_DISC) && !kato)
1299 kato = NVMET_DISC_KATO_MS;
1301 /* keep-alive timeout in seconds */
1302 ctrl->kato = DIV_ROUND_UP(kato, 1000);
1304 ctrl->err_counter = 0;
1305 spin_lock_init(&ctrl->error_lock);
1307 nvmet_start_keep_alive_timer(ctrl);
1309 mutex_lock(&subsys->lock);
1310 list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
1311 nvmet_setup_p2p_ns_map(ctrl, req);
1312 mutex_unlock(&subsys->lock);
1321 out_free_changed_ns_list:
1322 kfree(ctrl->changed_ns_list);
1326 nvmet_subsys_put(subsys);
1331 static void nvmet_ctrl_free(struct kref *ref)
1333 struct nvmet_ctrl *ctrl = container_of(ref, struct nvmet_ctrl, ref);
1334 struct nvmet_subsys *subsys = ctrl->subsys;
1336 mutex_lock(&subsys->lock);
1337 nvmet_release_p2p_ns_map(ctrl);
1338 list_del(&ctrl->subsys_entry);
1339 mutex_unlock(&subsys->lock);
1341 nvmet_stop_keep_alive_timer(ctrl);
1343 flush_work(&ctrl->async_event_work);
1344 cancel_work_sync(&ctrl->fatal_err_work);
1346 ida_simple_remove(&cntlid_ida, ctrl->cntlid);
1350 kfree(ctrl->changed_ns_list);
1353 nvmet_subsys_put(subsys);
1356 void nvmet_ctrl_put(struct nvmet_ctrl *ctrl)
1358 kref_put(&ctrl->ref, nvmet_ctrl_free);
1361 void nvmet_ctrl_fatal_error(struct nvmet_ctrl *ctrl)
1363 mutex_lock(&ctrl->lock);
1364 if (!(ctrl->csts & NVME_CSTS_CFS)) {
1365 ctrl->csts |= NVME_CSTS_CFS;
1366 schedule_work(&ctrl->fatal_err_work);
1368 mutex_unlock(&ctrl->lock);
1370 EXPORT_SYMBOL_GPL(nvmet_ctrl_fatal_error);
1372 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
1373 const char *subsysnqn)
1375 struct nvmet_subsys_link *p;
1380 if (!strcmp(NVME_DISC_SUBSYS_NAME, subsysnqn)) {
1381 if (!kref_get_unless_zero(&nvmet_disc_subsys->ref))
1383 return nvmet_disc_subsys;
1386 down_read(&nvmet_config_sem);
1387 list_for_each_entry(p, &port->subsystems, entry) {
1388 if (!strncmp(p->subsys->subsysnqn, subsysnqn,
1390 if (!kref_get_unless_zero(&p->subsys->ref))
1392 up_read(&nvmet_config_sem);
1396 up_read(&nvmet_config_sem);
1400 struct nvmet_subsys *nvmet_subsys_alloc(const char *subsysnqn,
1401 enum nvme_subsys_type type)
1403 struct nvmet_subsys *subsys;
1405 subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
1407 return ERR_PTR(-ENOMEM);
1409 subsys->ver = NVME_VS(1, 3, 0); /* NVMe 1.3.0 */
1410 /* generate a random serial number as our controllers are ephemeral: */
1411 get_random_bytes(&subsys->serial, sizeof(subsys->serial));
1415 subsys->max_qid = NVMET_NR_QUEUES;
1418 subsys->max_qid = 0;
1421 pr_err("%s: Unknown Subsystem type - %d\n", __func__, type);
1423 return ERR_PTR(-EINVAL);
1425 subsys->type = type;
1426 subsys->subsysnqn = kstrndup(subsysnqn, NVMF_NQN_SIZE,
1428 if (!subsys->subsysnqn) {
1430 return ERR_PTR(-ENOMEM);
1433 kref_init(&subsys->ref);
1435 mutex_init(&subsys->lock);
1436 INIT_LIST_HEAD(&subsys->namespaces);
1437 INIT_LIST_HEAD(&subsys->ctrls);
1438 INIT_LIST_HEAD(&subsys->hosts);
1443 static void nvmet_subsys_free(struct kref *ref)
1445 struct nvmet_subsys *subsys =
1446 container_of(ref, struct nvmet_subsys, ref);
1448 WARN_ON_ONCE(!list_empty(&subsys->namespaces));
1450 kfree(subsys->subsysnqn);
1454 void nvmet_subsys_del_ctrls(struct nvmet_subsys *subsys)
1456 struct nvmet_ctrl *ctrl;
1458 mutex_lock(&subsys->lock);
1459 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
1460 ctrl->ops->delete_ctrl(ctrl);
1461 mutex_unlock(&subsys->lock);
1464 void nvmet_subsys_put(struct nvmet_subsys *subsys)
1466 kref_put(&subsys->ref, nvmet_subsys_free);
1469 static int __init nvmet_init(void)
1473 nvmet_ana_group_enabled[NVMET_DEFAULT_ANA_GRPID] = 1;
1475 buffered_io_wq = alloc_workqueue("nvmet-buffered-io-wq",
1477 if (!buffered_io_wq) {
1482 error = nvmet_init_discovery();
1484 goto out_free_work_queue;
1486 error = nvmet_init_configfs();
1488 goto out_exit_discovery;
1492 nvmet_exit_discovery();
1493 out_free_work_queue:
1494 destroy_workqueue(buffered_io_wq);
1499 static void __exit nvmet_exit(void)
1501 nvmet_exit_configfs();
1502 nvmet_exit_discovery();
1503 ida_destroy(&cntlid_ida);
1504 destroy_workqueue(buffered_io_wq);
1506 BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_entry) != 1024);
1507 BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_hdr) != 1024);
1510 module_init(nvmet_init);
1511 module_exit(nvmet_exit);
1513 MODULE_LICENSE("GPL v2");