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 BUG_ON(ns->nsid == old->nsid);
560 if (ns->nsid < old->nsid)
564 list_add_tail_rcu(&ns->dev_link, &old->dev_link);
566 subsys->nr_namespaces++;
568 nvmet_ns_changed(subsys, ns->nsid);
572 mutex_unlock(&subsys->lock);
575 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
576 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
578 nvmet_ns_dev_disable(ns);
582 void nvmet_ns_disable(struct nvmet_ns *ns)
584 struct nvmet_subsys *subsys = ns->subsys;
585 struct nvmet_ctrl *ctrl;
587 mutex_lock(&subsys->lock);
592 list_del_rcu(&ns->dev_link);
593 if (ns->nsid == subsys->max_nsid)
594 subsys->max_nsid = nvmet_max_nsid(subsys);
596 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
597 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
599 mutex_unlock(&subsys->lock);
602 * Now that we removed the namespaces from the lookup list, we
603 * can kill the per_cpu ref and wait for any remaining references
604 * to be dropped, as well as a RCU grace period for anyone only
605 * using the namepace under rcu_read_lock(). Note that we can't
606 * use call_rcu here as we need to ensure the namespaces have
607 * been fully destroyed before unloading the module.
609 percpu_ref_kill(&ns->ref);
611 wait_for_completion(&ns->disable_done);
612 percpu_ref_exit(&ns->ref);
614 mutex_lock(&subsys->lock);
616 subsys->nr_namespaces--;
617 nvmet_ns_changed(subsys, ns->nsid);
618 nvmet_ns_dev_disable(ns);
620 mutex_unlock(&subsys->lock);
623 void nvmet_ns_free(struct nvmet_ns *ns)
625 nvmet_ns_disable(ns);
627 down_write(&nvmet_ana_sem);
628 nvmet_ana_group_enabled[ns->anagrpid]--;
629 up_write(&nvmet_ana_sem);
631 kfree(ns->device_path);
635 struct nvmet_ns *nvmet_ns_alloc(struct nvmet_subsys *subsys, u32 nsid)
639 ns = kzalloc(sizeof(*ns), GFP_KERNEL);
643 INIT_LIST_HEAD(&ns->dev_link);
644 init_completion(&ns->disable_done);
649 down_write(&nvmet_ana_sem);
650 ns->anagrpid = NVMET_DEFAULT_ANA_GRPID;
651 nvmet_ana_group_enabled[ns->anagrpid]++;
652 up_write(&nvmet_ana_sem);
655 ns->buffered_io = false;
660 static void nvmet_update_sq_head(struct nvmet_req *req)
663 u32 old_sqhd, new_sqhd;
666 old_sqhd = req->sq->sqhd;
667 new_sqhd = (old_sqhd + 1) % req->sq->size;
668 } while (cmpxchg(&req->sq->sqhd, old_sqhd, new_sqhd) !=
671 req->cqe->sq_head = cpu_to_le16(req->sq->sqhd & 0x0000FFFF);
674 static void nvmet_set_error(struct nvmet_req *req, u16 status)
676 struct nvmet_ctrl *ctrl = req->sq->ctrl;
677 struct nvme_error_slot *new_error_slot;
680 req->cqe->status = cpu_to_le16(status << 1);
682 if (!ctrl || req->error_loc == NVMET_NO_ERROR_LOC)
685 spin_lock_irqsave(&ctrl->error_lock, flags);
688 &ctrl->slots[ctrl->err_counter % NVMET_ERROR_LOG_SLOTS];
690 new_error_slot->error_count = cpu_to_le64(ctrl->err_counter);
691 new_error_slot->sqid = cpu_to_le16(req->sq->qid);
692 new_error_slot->cmdid = cpu_to_le16(req->cmd->common.command_id);
693 new_error_slot->status_field = cpu_to_le16(status << 1);
694 new_error_slot->param_error_location = cpu_to_le16(req->error_loc);
695 new_error_slot->lba = cpu_to_le64(req->error_slba);
696 new_error_slot->nsid = req->cmd->common.nsid;
697 spin_unlock_irqrestore(&ctrl->error_lock, flags);
699 /* set the more bit for this request */
700 req->cqe->status |= cpu_to_le16(1 << 14);
703 static void __nvmet_req_complete(struct nvmet_req *req, u16 status)
705 if (!req->sq->sqhd_disabled)
706 nvmet_update_sq_head(req);
707 req->cqe->sq_id = cpu_to_le16(req->sq->qid);
708 req->cqe->command_id = req->cmd->common.command_id;
710 if (unlikely(status))
711 nvmet_set_error(req, status);
713 trace_nvmet_req_complete(req);
716 nvmet_put_namespace(req->ns);
717 req->ops->queue_response(req);
720 void nvmet_req_complete(struct nvmet_req *req, u16 status)
722 __nvmet_req_complete(req, status);
723 percpu_ref_put(&req->sq->ref);
725 EXPORT_SYMBOL_GPL(nvmet_req_complete);
727 void nvmet_cq_setup(struct nvmet_ctrl *ctrl, struct nvmet_cq *cq,
736 void nvmet_sq_setup(struct nvmet_ctrl *ctrl, struct nvmet_sq *sq,
746 static void nvmet_confirm_sq(struct percpu_ref *ref)
748 struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
750 complete(&sq->confirm_done);
753 void nvmet_sq_destroy(struct nvmet_sq *sq)
756 * If this is the admin queue, complete all AERs so that our
757 * queue doesn't have outstanding requests on it.
759 if (sq->ctrl && sq->ctrl->sqs && sq->ctrl->sqs[0] == sq)
760 nvmet_async_events_free(sq->ctrl);
761 percpu_ref_kill_and_confirm(&sq->ref, nvmet_confirm_sq);
762 wait_for_completion(&sq->confirm_done);
763 wait_for_completion(&sq->free_done);
764 percpu_ref_exit(&sq->ref);
767 nvmet_ctrl_put(sq->ctrl);
768 sq->ctrl = NULL; /* allows reusing the queue later */
771 EXPORT_SYMBOL_GPL(nvmet_sq_destroy);
773 static void nvmet_sq_free(struct percpu_ref *ref)
775 struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
777 complete(&sq->free_done);
780 int nvmet_sq_init(struct nvmet_sq *sq)
784 ret = percpu_ref_init(&sq->ref, nvmet_sq_free, 0, GFP_KERNEL);
786 pr_err("percpu_ref init failed!\n");
789 init_completion(&sq->free_done);
790 init_completion(&sq->confirm_done);
794 EXPORT_SYMBOL_GPL(nvmet_sq_init);
796 static inline u16 nvmet_check_ana_state(struct nvmet_port *port,
799 enum nvme_ana_state state = port->ana_state[ns->anagrpid];
801 if (unlikely(state == NVME_ANA_INACCESSIBLE))
802 return NVME_SC_ANA_INACCESSIBLE;
803 if (unlikely(state == NVME_ANA_PERSISTENT_LOSS))
804 return NVME_SC_ANA_PERSISTENT_LOSS;
805 if (unlikely(state == NVME_ANA_CHANGE))
806 return NVME_SC_ANA_TRANSITION;
810 static inline u16 nvmet_io_cmd_check_access(struct nvmet_req *req)
812 if (unlikely(req->ns->readonly)) {
813 switch (req->cmd->common.opcode) {
818 return NVME_SC_NS_WRITE_PROTECTED;
825 static u16 nvmet_parse_io_cmd(struct nvmet_req *req)
827 struct nvme_command *cmd = req->cmd;
830 ret = nvmet_check_ctrl_status(req, cmd);
834 req->ns = nvmet_find_namespace(req->sq->ctrl, cmd->rw.nsid);
835 if (unlikely(!req->ns)) {
836 req->error_loc = offsetof(struct nvme_common_command, nsid);
837 return NVME_SC_INVALID_NS | NVME_SC_DNR;
839 ret = nvmet_check_ana_state(req->port, req->ns);
841 req->error_loc = offsetof(struct nvme_common_command, nsid);
844 ret = nvmet_io_cmd_check_access(req);
846 req->error_loc = offsetof(struct nvme_common_command, nsid);
851 return nvmet_file_parse_io_cmd(req);
853 return nvmet_bdev_parse_io_cmd(req);
856 bool nvmet_req_init(struct nvmet_req *req, struct nvmet_cq *cq,
857 struct nvmet_sq *sq, const struct nvmet_fabrics_ops *ops)
859 u8 flags = req->cmd->common.flags;
867 req->transfer_len = 0;
868 req->cqe->status = 0;
869 req->cqe->sq_head = 0;
871 req->error_loc = NVMET_NO_ERROR_LOC;
874 trace_nvmet_req_init(req, req->cmd);
876 /* no support for fused commands yet */
877 if (unlikely(flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND))) {
878 req->error_loc = offsetof(struct nvme_common_command, flags);
879 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
884 * For fabrics, PSDT field shall describe metadata pointer (MPTR) that
885 * contains an address of a single contiguous physical buffer that is
888 if (unlikely((flags & NVME_CMD_SGL_ALL) != NVME_CMD_SGL_METABUF)) {
889 req->error_loc = offsetof(struct nvme_common_command, flags);
890 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
894 if (unlikely(!req->sq->ctrl))
895 /* will return an error for any Non-connect command: */
896 status = nvmet_parse_connect_cmd(req);
897 else if (likely(req->sq->qid != 0))
898 status = nvmet_parse_io_cmd(req);
899 else if (nvme_is_fabrics(req->cmd))
900 status = nvmet_parse_fabrics_cmd(req);
901 else if (req->sq->ctrl->subsys->type == NVME_NQN_DISC)
902 status = nvmet_parse_discovery_cmd(req);
904 status = nvmet_parse_admin_cmd(req);
909 if (unlikely(!percpu_ref_tryget_live(&sq->ref))) {
910 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
915 sq->ctrl->cmd_seen = true;
920 __nvmet_req_complete(req, status);
923 EXPORT_SYMBOL_GPL(nvmet_req_init);
925 void nvmet_req_uninit(struct nvmet_req *req)
927 percpu_ref_put(&req->sq->ref);
929 nvmet_put_namespace(req->ns);
931 EXPORT_SYMBOL_GPL(nvmet_req_uninit);
933 void nvmet_req_execute(struct nvmet_req *req)
935 if (unlikely(req->data_len != req->transfer_len)) {
936 req->error_loc = offsetof(struct nvme_common_command, dptr);
937 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
941 EXPORT_SYMBOL_GPL(nvmet_req_execute);
943 int nvmet_req_alloc_sgl(struct nvmet_req *req)
945 struct pci_dev *p2p_dev = NULL;
947 if (IS_ENABLED(CONFIG_PCI_P2PDMA)) {
948 if (req->sq->ctrl && req->ns)
949 p2p_dev = radix_tree_lookup(&req->sq->ctrl->p2p_ns_map,
953 if (req->sq->qid && p2p_dev) {
954 req->sg = pci_p2pmem_alloc_sgl(p2p_dev, &req->sg_cnt,
957 req->p2p_dev = p2p_dev;
963 * If no P2P memory was available we fallback to using
968 req->sg = sgl_alloc(req->transfer_len, GFP_KERNEL, &req->sg_cnt);
974 EXPORT_SYMBOL_GPL(nvmet_req_alloc_sgl);
976 void nvmet_req_free_sgl(struct nvmet_req *req)
979 pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
986 EXPORT_SYMBOL_GPL(nvmet_req_free_sgl);
988 static inline bool nvmet_cc_en(u32 cc)
990 return (cc >> NVME_CC_EN_SHIFT) & 0x1;
993 static inline u8 nvmet_cc_css(u32 cc)
995 return (cc >> NVME_CC_CSS_SHIFT) & 0x7;
998 static inline u8 nvmet_cc_mps(u32 cc)
1000 return (cc >> NVME_CC_MPS_SHIFT) & 0xf;
1003 static inline u8 nvmet_cc_ams(u32 cc)
1005 return (cc >> NVME_CC_AMS_SHIFT) & 0x7;
1008 static inline u8 nvmet_cc_shn(u32 cc)
1010 return (cc >> NVME_CC_SHN_SHIFT) & 0x3;
1013 static inline u8 nvmet_cc_iosqes(u32 cc)
1015 return (cc >> NVME_CC_IOSQES_SHIFT) & 0xf;
1018 static inline u8 nvmet_cc_iocqes(u32 cc)
1020 return (cc >> NVME_CC_IOCQES_SHIFT) & 0xf;
1023 static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl)
1025 lockdep_assert_held(&ctrl->lock);
1027 if (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES ||
1028 nvmet_cc_iocqes(ctrl->cc) != NVME_NVM_IOCQES ||
1029 nvmet_cc_mps(ctrl->cc) != 0 ||
1030 nvmet_cc_ams(ctrl->cc) != 0 ||
1031 nvmet_cc_css(ctrl->cc) != 0) {
1032 ctrl->csts = NVME_CSTS_CFS;
1036 ctrl->csts = NVME_CSTS_RDY;
1039 * Controllers that are not yet enabled should not really enforce the
1040 * keep alive timeout, but we still want to track a timeout and cleanup
1041 * in case a host died before it enabled the controller. Hence, simply
1042 * reset the keep alive timer when the controller is enabled.
1044 mod_delayed_work(system_wq, &ctrl->ka_work, ctrl->kato * HZ);
1047 static void nvmet_clear_ctrl(struct nvmet_ctrl *ctrl)
1049 lockdep_assert_held(&ctrl->lock);
1051 /* XXX: tear down queues? */
1052 ctrl->csts &= ~NVME_CSTS_RDY;
1056 void nvmet_update_cc(struct nvmet_ctrl *ctrl, u32 new)
1060 mutex_lock(&ctrl->lock);
1064 if (nvmet_cc_en(new) && !nvmet_cc_en(old))
1065 nvmet_start_ctrl(ctrl);
1066 if (!nvmet_cc_en(new) && nvmet_cc_en(old))
1067 nvmet_clear_ctrl(ctrl);
1068 if (nvmet_cc_shn(new) && !nvmet_cc_shn(old)) {
1069 nvmet_clear_ctrl(ctrl);
1070 ctrl->csts |= NVME_CSTS_SHST_CMPLT;
1072 if (!nvmet_cc_shn(new) && nvmet_cc_shn(old))
1073 ctrl->csts &= ~NVME_CSTS_SHST_CMPLT;
1074 mutex_unlock(&ctrl->lock);
1077 static void nvmet_init_cap(struct nvmet_ctrl *ctrl)
1079 /* command sets supported: NVMe command set: */
1080 ctrl->cap = (1ULL << 37);
1081 /* CC.EN timeout in 500msec units: */
1082 ctrl->cap |= (15ULL << 24);
1083 /* maximum queue entries supported: */
1084 ctrl->cap |= NVMET_QUEUE_SIZE - 1;
1087 u16 nvmet_ctrl_find_get(const char *subsysnqn, const char *hostnqn, u16 cntlid,
1088 struct nvmet_req *req, struct nvmet_ctrl **ret)
1090 struct nvmet_subsys *subsys;
1091 struct nvmet_ctrl *ctrl;
1094 subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1096 pr_warn("connect request for invalid subsystem %s!\n",
1098 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1099 return NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1102 mutex_lock(&subsys->lock);
1103 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
1104 if (ctrl->cntlid == cntlid) {
1105 if (strncmp(hostnqn, ctrl->hostnqn, NVMF_NQN_SIZE)) {
1106 pr_warn("hostnqn mismatch.\n");
1109 if (!kref_get_unless_zero(&ctrl->ref))
1117 pr_warn("could not find controller %d for subsys %s / host %s\n",
1118 cntlid, subsysnqn, hostnqn);
1119 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(cntlid);
1120 status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1123 mutex_unlock(&subsys->lock);
1124 nvmet_subsys_put(subsys);
1128 u16 nvmet_check_ctrl_status(struct nvmet_req *req, struct nvme_command *cmd)
1130 if (unlikely(!(req->sq->ctrl->cc & NVME_CC_ENABLE))) {
1131 pr_err("got cmd %d while CC.EN == 0 on qid = %d\n",
1132 cmd->common.opcode, req->sq->qid);
1133 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1136 if (unlikely(!(req->sq->ctrl->csts & NVME_CSTS_RDY))) {
1137 pr_err("got cmd %d while CSTS.RDY == 0 on qid = %d\n",
1138 cmd->common.opcode, req->sq->qid);
1139 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1144 bool nvmet_host_allowed(struct nvmet_subsys *subsys, const char *hostnqn)
1146 struct nvmet_host_link *p;
1148 lockdep_assert_held(&nvmet_config_sem);
1150 if (subsys->allow_any_host)
1153 if (subsys->type == NVME_NQN_DISC) /* allow all access to disc subsys */
1156 list_for_each_entry(p, &subsys->hosts, entry) {
1157 if (!strcmp(nvmet_host_name(p->host), hostnqn))
1165 * Note: ctrl->subsys->lock should be held when calling this function
1167 static void nvmet_setup_p2p_ns_map(struct nvmet_ctrl *ctrl,
1168 struct nvmet_req *req)
1170 struct nvmet_ns *ns;
1172 if (!req->p2p_client)
1175 ctrl->p2p_client = get_device(req->p2p_client);
1177 list_for_each_entry_rcu(ns, &ctrl->subsys->namespaces, dev_link)
1178 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
1182 * Note: ctrl->subsys->lock should be held when calling this function
1184 static void nvmet_release_p2p_ns_map(struct nvmet_ctrl *ctrl)
1186 struct radix_tree_iter iter;
1189 radix_tree_for_each_slot(slot, &ctrl->p2p_ns_map, &iter, 0)
1190 pci_dev_put(radix_tree_deref_slot(slot));
1192 put_device(ctrl->p2p_client);
1195 static void nvmet_fatal_error_handler(struct work_struct *work)
1197 struct nvmet_ctrl *ctrl =
1198 container_of(work, struct nvmet_ctrl, fatal_err_work);
1200 pr_err("ctrl %d fatal error occurred!\n", ctrl->cntlid);
1201 ctrl->ops->delete_ctrl(ctrl);
1204 u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn,
1205 struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp)
1207 struct nvmet_subsys *subsys;
1208 struct nvmet_ctrl *ctrl;
1212 status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1213 subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1215 pr_warn("connect request for invalid subsystem %s!\n",
1217 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1221 status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1222 down_read(&nvmet_config_sem);
1223 if (!nvmet_host_allowed(subsys, hostnqn)) {
1224 pr_info("connect by host %s for subsystem %s not allowed\n",
1225 hostnqn, subsysnqn);
1226 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(hostnqn);
1227 up_read(&nvmet_config_sem);
1228 status = NVME_SC_CONNECT_INVALID_HOST | NVME_SC_DNR;
1229 goto out_put_subsystem;
1231 up_read(&nvmet_config_sem);
1233 status = NVME_SC_INTERNAL;
1234 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1236 goto out_put_subsystem;
1237 mutex_init(&ctrl->lock);
1239 nvmet_init_cap(ctrl);
1241 ctrl->port = req->port;
1243 INIT_WORK(&ctrl->async_event_work, nvmet_async_event_work);
1244 INIT_LIST_HEAD(&ctrl->async_events);
1245 INIT_RADIX_TREE(&ctrl->p2p_ns_map, GFP_KERNEL);
1246 INIT_WORK(&ctrl->fatal_err_work, nvmet_fatal_error_handler);
1248 memcpy(ctrl->subsysnqn, subsysnqn, NVMF_NQN_SIZE);
1249 memcpy(ctrl->hostnqn, hostnqn, NVMF_NQN_SIZE);
1251 kref_init(&ctrl->ref);
1252 ctrl->subsys = subsys;
1253 WRITE_ONCE(ctrl->aen_enabled, NVMET_AEN_CFG_OPTIONAL);
1255 ctrl->changed_ns_list = kmalloc_array(NVME_MAX_CHANGED_NAMESPACES,
1256 sizeof(__le32), GFP_KERNEL);
1257 if (!ctrl->changed_ns_list)
1260 ctrl->cqs = kcalloc(subsys->max_qid + 1,
1261 sizeof(struct nvmet_cq *),
1264 goto out_free_changed_ns_list;
1266 ctrl->sqs = kcalloc(subsys->max_qid + 1,
1267 sizeof(struct nvmet_sq *),
1272 ret = ida_simple_get(&cntlid_ida,
1273 NVME_CNTLID_MIN, NVME_CNTLID_MAX,
1276 status = NVME_SC_CONNECT_CTRL_BUSY | NVME_SC_DNR;
1281 ctrl->ops = req->ops;
1284 * Discovery controllers may use some arbitrary high value
1285 * in order to cleanup stale discovery sessions
1287 if ((ctrl->subsys->type == NVME_NQN_DISC) && !kato)
1288 kato = NVMET_DISC_KATO_MS;
1290 /* keep-alive timeout in seconds */
1291 ctrl->kato = DIV_ROUND_UP(kato, 1000);
1293 ctrl->err_counter = 0;
1294 spin_lock_init(&ctrl->error_lock);
1296 nvmet_start_keep_alive_timer(ctrl);
1298 mutex_lock(&subsys->lock);
1299 list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
1300 nvmet_setup_p2p_ns_map(ctrl, req);
1301 mutex_unlock(&subsys->lock);
1310 out_free_changed_ns_list:
1311 kfree(ctrl->changed_ns_list);
1315 nvmet_subsys_put(subsys);
1320 static void nvmet_ctrl_free(struct kref *ref)
1322 struct nvmet_ctrl *ctrl = container_of(ref, struct nvmet_ctrl, ref);
1323 struct nvmet_subsys *subsys = ctrl->subsys;
1325 mutex_lock(&subsys->lock);
1326 nvmet_release_p2p_ns_map(ctrl);
1327 list_del(&ctrl->subsys_entry);
1328 mutex_unlock(&subsys->lock);
1330 nvmet_stop_keep_alive_timer(ctrl);
1332 flush_work(&ctrl->async_event_work);
1333 cancel_work_sync(&ctrl->fatal_err_work);
1335 ida_simple_remove(&cntlid_ida, ctrl->cntlid);
1339 kfree(ctrl->changed_ns_list);
1342 nvmet_subsys_put(subsys);
1345 void nvmet_ctrl_put(struct nvmet_ctrl *ctrl)
1347 kref_put(&ctrl->ref, nvmet_ctrl_free);
1350 void nvmet_ctrl_fatal_error(struct nvmet_ctrl *ctrl)
1352 mutex_lock(&ctrl->lock);
1353 if (!(ctrl->csts & NVME_CSTS_CFS)) {
1354 ctrl->csts |= NVME_CSTS_CFS;
1355 schedule_work(&ctrl->fatal_err_work);
1357 mutex_unlock(&ctrl->lock);
1359 EXPORT_SYMBOL_GPL(nvmet_ctrl_fatal_error);
1361 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
1362 const char *subsysnqn)
1364 struct nvmet_subsys_link *p;
1369 if (!strcmp(NVME_DISC_SUBSYS_NAME, subsysnqn)) {
1370 if (!kref_get_unless_zero(&nvmet_disc_subsys->ref))
1372 return nvmet_disc_subsys;
1375 down_read(&nvmet_config_sem);
1376 list_for_each_entry(p, &port->subsystems, entry) {
1377 if (!strncmp(p->subsys->subsysnqn, subsysnqn,
1379 if (!kref_get_unless_zero(&p->subsys->ref))
1381 up_read(&nvmet_config_sem);
1385 up_read(&nvmet_config_sem);
1389 struct nvmet_subsys *nvmet_subsys_alloc(const char *subsysnqn,
1390 enum nvme_subsys_type type)
1392 struct nvmet_subsys *subsys;
1394 subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
1396 return ERR_PTR(-ENOMEM);
1398 subsys->ver = NVME_VS(1, 3, 0); /* NVMe 1.3.0 */
1399 /* generate a random serial number as our controllers are ephemeral: */
1400 get_random_bytes(&subsys->serial, sizeof(subsys->serial));
1404 subsys->max_qid = NVMET_NR_QUEUES;
1407 subsys->max_qid = 0;
1410 pr_err("%s: Unknown Subsystem type - %d\n", __func__, type);
1412 return ERR_PTR(-EINVAL);
1414 subsys->type = type;
1415 subsys->subsysnqn = kstrndup(subsysnqn, NVMF_NQN_SIZE,
1417 if (!subsys->subsysnqn) {
1419 return ERR_PTR(-ENOMEM);
1422 kref_init(&subsys->ref);
1424 mutex_init(&subsys->lock);
1425 INIT_LIST_HEAD(&subsys->namespaces);
1426 INIT_LIST_HEAD(&subsys->ctrls);
1427 INIT_LIST_HEAD(&subsys->hosts);
1432 static void nvmet_subsys_free(struct kref *ref)
1434 struct nvmet_subsys *subsys =
1435 container_of(ref, struct nvmet_subsys, ref);
1437 WARN_ON_ONCE(!list_empty(&subsys->namespaces));
1439 kfree(subsys->subsysnqn);
1443 void nvmet_subsys_del_ctrls(struct nvmet_subsys *subsys)
1445 struct nvmet_ctrl *ctrl;
1447 mutex_lock(&subsys->lock);
1448 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
1449 ctrl->ops->delete_ctrl(ctrl);
1450 mutex_unlock(&subsys->lock);
1453 void nvmet_subsys_put(struct nvmet_subsys *subsys)
1455 kref_put(&subsys->ref, nvmet_subsys_free);
1458 static int __init nvmet_init(void)
1462 nvmet_ana_group_enabled[NVMET_DEFAULT_ANA_GRPID] = 1;
1464 buffered_io_wq = alloc_workqueue("nvmet-buffered-io-wq",
1466 if (!buffered_io_wq) {
1471 error = nvmet_init_discovery();
1473 goto out_free_work_queue;
1475 error = nvmet_init_configfs();
1477 goto out_exit_discovery;
1481 nvmet_exit_discovery();
1482 out_free_work_queue:
1483 destroy_workqueue(buffered_io_wq);
1488 static void __exit nvmet_exit(void)
1490 nvmet_exit_configfs();
1491 nvmet_exit_discovery();
1492 ida_destroy(&cntlid_ida);
1493 destroy_workqueue(buffered_io_wq);
1495 BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_entry) != 1024);
1496 BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_hdr) != 1024);
1499 module_init(nvmet_init);
1500 module_exit(nvmet_exit);
1502 MODULE_LICENSE("GPL v2");