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 u16 nvmet_report_invalid_opcode(struct nvmet_req *req)
87 pr_debug("unhandled cmd %d on qid %d\n", req->cmd->common.opcode,
90 req->error_loc = offsetof(struct nvme_common_command, opcode);
91 return NVME_SC_INVALID_OPCODE | NVME_SC_DNR;
94 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
95 const char *subsysnqn);
97 u16 nvmet_copy_to_sgl(struct nvmet_req *req, off_t off, const void *buf,
100 if (sg_pcopy_from_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_copy_from_sgl(struct nvmet_req *req, off_t off, void *buf, size_t len)
109 if (sg_pcopy_to_buffer(req->sg, req->sg_cnt, buf, len, off) != len) {
110 req->error_loc = offsetof(struct nvme_common_command, dptr);
111 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
116 u16 nvmet_zero_sgl(struct nvmet_req *req, off_t off, size_t len)
118 if (sg_zero_buffer(req->sg, req->sg_cnt, len, off) != len) {
119 req->error_loc = offsetof(struct nvme_common_command, dptr);
120 return NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR;
125 static unsigned int nvmet_max_nsid(struct nvmet_subsys *subsys)
127 unsigned long nsid = 0;
128 struct nvmet_ns *cur;
131 xa_for_each(&subsys->namespaces, idx, cur)
137 static u32 nvmet_async_event_result(struct nvmet_async_event *aen)
139 return aen->event_type | (aen->event_info << 8) | (aen->log_page << 16);
142 static void nvmet_async_events_failall(struct nvmet_ctrl *ctrl)
144 u16 status = NVME_SC_INTERNAL | NVME_SC_DNR;
145 struct nvmet_req *req;
147 mutex_lock(&ctrl->lock);
148 while (ctrl->nr_async_event_cmds) {
149 req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
150 mutex_unlock(&ctrl->lock);
151 nvmet_req_complete(req, status);
152 mutex_lock(&ctrl->lock);
154 mutex_unlock(&ctrl->lock);
157 static void nvmet_async_events_process(struct nvmet_ctrl *ctrl)
159 struct nvmet_async_event *aen;
160 struct nvmet_req *req;
162 mutex_lock(&ctrl->lock);
163 while (ctrl->nr_async_event_cmds && !list_empty(&ctrl->async_events)) {
164 aen = list_first_entry(&ctrl->async_events,
165 struct nvmet_async_event, entry);
166 req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
167 nvmet_set_result(req, nvmet_async_event_result(aen));
169 list_del(&aen->entry);
172 mutex_unlock(&ctrl->lock);
173 trace_nvmet_async_event(ctrl, req->cqe->result.u32);
174 nvmet_req_complete(req, 0);
175 mutex_lock(&ctrl->lock);
177 mutex_unlock(&ctrl->lock);
180 static void nvmet_async_events_free(struct nvmet_ctrl *ctrl)
182 struct nvmet_async_event *aen, *tmp;
184 mutex_lock(&ctrl->lock);
185 list_for_each_entry_safe(aen, tmp, &ctrl->async_events, entry) {
186 list_del(&aen->entry);
189 mutex_unlock(&ctrl->lock);
192 static void nvmet_async_event_work(struct work_struct *work)
194 struct nvmet_ctrl *ctrl =
195 container_of(work, struct nvmet_ctrl, async_event_work);
197 nvmet_async_events_process(ctrl);
200 void nvmet_add_async_event(struct nvmet_ctrl *ctrl, u8 event_type,
201 u8 event_info, u8 log_page)
203 struct nvmet_async_event *aen;
205 aen = kmalloc(sizeof(*aen), GFP_KERNEL);
209 aen->event_type = event_type;
210 aen->event_info = event_info;
211 aen->log_page = log_page;
213 mutex_lock(&ctrl->lock);
214 list_add_tail(&aen->entry, &ctrl->async_events);
215 mutex_unlock(&ctrl->lock);
217 schedule_work(&ctrl->async_event_work);
220 static void nvmet_add_to_changed_ns_log(struct nvmet_ctrl *ctrl, __le32 nsid)
224 mutex_lock(&ctrl->lock);
225 if (ctrl->nr_changed_ns > NVME_MAX_CHANGED_NAMESPACES)
228 for (i = 0; i < ctrl->nr_changed_ns; i++) {
229 if (ctrl->changed_ns_list[i] == nsid)
233 if (ctrl->nr_changed_ns == NVME_MAX_CHANGED_NAMESPACES) {
234 ctrl->changed_ns_list[0] = cpu_to_le32(0xffffffff);
235 ctrl->nr_changed_ns = U32_MAX;
239 ctrl->changed_ns_list[ctrl->nr_changed_ns++] = nsid;
241 mutex_unlock(&ctrl->lock);
244 void nvmet_ns_changed(struct nvmet_subsys *subsys, u32 nsid)
246 struct nvmet_ctrl *ctrl;
248 lockdep_assert_held(&subsys->lock);
250 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
251 nvmet_add_to_changed_ns_log(ctrl, cpu_to_le32(nsid));
252 if (nvmet_aen_bit_disabled(ctrl, NVME_AEN_BIT_NS_ATTR))
254 nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE,
255 NVME_AER_NOTICE_NS_CHANGED,
256 NVME_LOG_CHANGED_NS);
260 void nvmet_send_ana_event(struct nvmet_subsys *subsys,
261 struct nvmet_port *port)
263 struct nvmet_ctrl *ctrl;
265 mutex_lock(&subsys->lock);
266 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
267 if (port && ctrl->port != port)
269 if (nvmet_aen_bit_disabled(ctrl, NVME_AEN_BIT_ANA_CHANGE))
271 nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE,
272 NVME_AER_NOTICE_ANA, NVME_LOG_ANA);
274 mutex_unlock(&subsys->lock);
277 void nvmet_port_send_ana_event(struct nvmet_port *port)
279 struct nvmet_subsys_link *p;
281 down_read(&nvmet_config_sem);
282 list_for_each_entry(p, &port->subsystems, entry)
283 nvmet_send_ana_event(p->subsys, port);
284 up_read(&nvmet_config_sem);
287 int nvmet_register_transport(const struct nvmet_fabrics_ops *ops)
291 down_write(&nvmet_config_sem);
292 if (nvmet_transports[ops->type])
295 nvmet_transports[ops->type] = ops;
296 up_write(&nvmet_config_sem);
300 EXPORT_SYMBOL_GPL(nvmet_register_transport);
302 void nvmet_unregister_transport(const struct nvmet_fabrics_ops *ops)
304 down_write(&nvmet_config_sem);
305 nvmet_transports[ops->type] = NULL;
306 up_write(&nvmet_config_sem);
308 EXPORT_SYMBOL_GPL(nvmet_unregister_transport);
310 void nvmet_port_del_ctrls(struct nvmet_port *port, struct nvmet_subsys *subsys)
312 struct nvmet_ctrl *ctrl;
314 mutex_lock(&subsys->lock);
315 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
316 if (ctrl->port == port)
317 ctrl->ops->delete_ctrl(ctrl);
319 mutex_unlock(&subsys->lock);
322 int nvmet_enable_port(struct nvmet_port *port)
324 const struct nvmet_fabrics_ops *ops;
327 lockdep_assert_held(&nvmet_config_sem);
329 ops = nvmet_transports[port->disc_addr.trtype];
331 up_write(&nvmet_config_sem);
332 request_module("nvmet-transport-%d", port->disc_addr.trtype);
333 down_write(&nvmet_config_sem);
334 ops = nvmet_transports[port->disc_addr.trtype];
336 pr_err("transport type %d not supported\n",
337 port->disc_addr.trtype);
342 if (!try_module_get(ops->owner))
346 * If the user requested PI support and the transport isn't pi capable,
347 * don't enable the port.
349 if (port->pi_enable && !(ops->flags & NVMF_METADATA_SUPPORTED)) {
350 pr_err("T10-PI is not supported by transport type %d\n",
351 port->disc_addr.trtype);
356 ret = ops->add_port(port);
360 /* If the transport didn't set inline_data_size, then disable it. */
361 if (port->inline_data_size < 0)
362 port->inline_data_size = 0;
364 port->enabled = true;
369 module_put(ops->owner);
373 void nvmet_disable_port(struct nvmet_port *port)
375 const struct nvmet_fabrics_ops *ops;
377 lockdep_assert_held(&nvmet_config_sem);
379 port->enabled = false;
382 ops = nvmet_transports[port->disc_addr.trtype];
383 ops->remove_port(port);
384 module_put(ops->owner);
387 static void nvmet_keep_alive_timer(struct work_struct *work)
389 struct nvmet_ctrl *ctrl = container_of(to_delayed_work(work),
390 struct nvmet_ctrl, ka_work);
391 bool reset_tbkas = ctrl->reset_tbkas;
393 ctrl->reset_tbkas = false;
395 pr_debug("ctrl %d reschedule traffic based keep-alive timer\n",
397 schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
401 pr_err("ctrl %d keep-alive timer (%d seconds) expired!\n",
402 ctrl->cntlid, ctrl->kato);
404 nvmet_ctrl_fatal_error(ctrl);
407 void nvmet_start_keep_alive_timer(struct nvmet_ctrl *ctrl)
409 if (unlikely(ctrl->kato == 0))
412 pr_debug("ctrl %d start keep-alive timer for %d secs\n",
413 ctrl->cntlid, ctrl->kato);
415 INIT_DELAYED_WORK(&ctrl->ka_work, nvmet_keep_alive_timer);
416 schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
419 void nvmet_stop_keep_alive_timer(struct nvmet_ctrl *ctrl)
421 if (unlikely(ctrl->kato == 0))
424 pr_debug("ctrl %d stop keep-alive\n", ctrl->cntlid);
426 cancel_delayed_work_sync(&ctrl->ka_work);
429 u16 nvmet_req_find_ns(struct nvmet_req *req)
431 u32 nsid = le32_to_cpu(req->cmd->common.nsid);
433 req->ns = xa_load(&nvmet_req_subsys(req)->namespaces, nsid);
434 if (unlikely(!req->ns)) {
435 req->error_loc = offsetof(struct nvme_common_command, nsid);
436 return NVME_SC_INVALID_NS | NVME_SC_DNR;
439 percpu_ref_get(&req->ns->ref);
440 return NVME_SC_SUCCESS;
443 static void nvmet_destroy_namespace(struct percpu_ref *ref)
445 struct nvmet_ns *ns = container_of(ref, struct nvmet_ns, ref);
447 complete(&ns->disable_done);
450 void nvmet_put_namespace(struct nvmet_ns *ns)
452 percpu_ref_put(&ns->ref);
455 static void nvmet_ns_dev_disable(struct nvmet_ns *ns)
457 nvmet_bdev_ns_disable(ns);
458 nvmet_file_ns_disable(ns);
461 static int nvmet_p2pmem_ns_enable(struct nvmet_ns *ns)
464 struct pci_dev *p2p_dev;
470 pr_err("peer-to-peer DMA is not supported by non-block device namespaces\n");
474 if (!blk_queue_pci_p2pdma(ns->bdev->bd_disk->queue)) {
475 pr_err("peer-to-peer DMA is not supported by the driver of %s\n",
481 ret = pci_p2pdma_distance(ns->p2p_dev, nvmet_ns_dev(ns), true);
486 * Right now we just check that there is p2pmem available so
487 * we can report an error to the user right away if there
488 * is not. We'll find the actual device to use once we
489 * setup the controller when the port's device is available.
492 p2p_dev = pci_p2pmem_find(nvmet_ns_dev(ns));
494 pr_err("no peer-to-peer memory is available for %s\n",
499 pci_dev_put(p2p_dev);
506 * Note: ctrl->subsys->lock should be held when calling this function
508 static void nvmet_p2pmem_ns_add_p2p(struct nvmet_ctrl *ctrl,
511 struct device *clients[2];
512 struct pci_dev *p2p_dev;
515 if (!ctrl->p2p_client || !ns->use_p2pmem)
519 ret = pci_p2pdma_distance(ns->p2p_dev, ctrl->p2p_client, true);
523 p2p_dev = pci_dev_get(ns->p2p_dev);
525 clients[0] = ctrl->p2p_client;
526 clients[1] = nvmet_ns_dev(ns);
528 p2p_dev = pci_p2pmem_find_many(clients, ARRAY_SIZE(clients));
530 pr_err("no peer-to-peer memory is available that's supported by %s and %s\n",
531 dev_name(ctrl->p2p_client), ns->device_path);
536 ret = radix_tree_insert(&ctrl->p2p_ns_map, ns->nsid, p2p_dev);
538 pci_dev_put(p2p_dev);
540 pr_info("using p2pmem on %s for nsid %d\n", pci_name(p2p_dev),
544 void nvmet_ns_revalidate(struct nvmet_ns *ns)
546 loff_t oldsize = ns->size;
549 nvmet_bdev_ns_revalidate(ns);
551 nvmet_file_ns_revalidate(ns);
553 if (oldsize != ns->size)
554 nvmet_ns_changed(ns->subsys, ns->nsid);
557 int nvmet_ns_enable(struct nvmet_ns *ns)
559 struct nvmet_subsys *subsys = ns->subsys;
560 struct nvmet_ctrl *ctrl;
563 mutex_lock(&subsys->lock);
566 if (nvmet_passthru_ctrl(subsys)) {
567 pr_info("cannot enable both passthru and regular namespaces for a single subsystem");
575 if (subsys->nr_namespaces == NVMET_MAX_NAMESPACES)
578 ret = nvmet_bdev_ns_enable(ns);
580 ret = nvmet_file_ns_enable(ns);
584 ret = nvmet_p2pmem_ns_enable(ns);
586 goto out_dev_disable;
588 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
589 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
591 ret = percpu_ref_init(&ns->ref, nvmet_destroy_namespace,
596 if (ns->nsid > subsys->max_nsid)
597 subsys->max_nsid = ns->nsid;
599 ret = xa_insert(&subsys->namespaces, ns->nsid, ns, GFP_KERNEL);
601 goto out_restore_subsys_maxnsid;
603 subsys->nr_namespaces++;
605 nvmet_ns_changed(subsys, ns->nsid);
609 mutex_unlock(&subsys->lock);
612 out_restore_subsys_maxnsid:
613 subsys->max_nsid = nvmet_max_nsid(subsys);
614 percpu_ref_exit(&ns->ref);
616 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
617 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
619 nvmet_ns_dev_disable(ns);
623 void nvmet_ns_disable(struct nvmet_ns *ns)
625 struct nvmet_subsys *subsys = ns->subsys;
626 struct nvmet_ctrl *ctrl;
628 mutex_lock(&subsys->lock);
633 xa_erase(&ns->subsys->namespaces, ns->nsid);
634 if (ns->nsid == subsys->max_nsid)
635 subsys->max_nsid = nvmet_max_nsid(subsys);
637 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
638 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
640 mutex_unlock(&subsys->lock);
643 * Now that we removed the namespaces from the lookup list, we
644 * can kill the per_cpu ref and wait for any remaining references
645 * to be dropped, as well as a RCU grace period for anyone only
646 * using the namepace under rcu_read_lock(). Note that we can't
647 * use call_rcu here as we need to ensure the namespaces have
648 * been fully destroyed before unloading the module.
650 percpu_ref_kill(&ns->ref);
652 wait_for_completion(&ns->disable_done);
653 percpu_ref_exit(&ns->ref);
655 mutex_lock(&subsys->lock);
657 subsys->nr_namespaces--;
658 nvmet_ns_changed(subsys, ns->nsid);
659 nvmet_ns_dev_disable(ns);
661 mutex_unlock(&subsys->lock);
664 void nvmet_ns_free(struct nvmet_ns *ns)
666 nvmet_ns_disable(ns);
668 down_write(&nvmet_ana_sem);
669 nvmet_ana_group_enabled[ns->anagrpid]--;
670 up_write(&nvmet_ana_sem);
672 kfree(ns->device_path);
676 struct nvmet_ns *nvmet_ns_alloc(struct nvmet_subsys *subsys, u32 nsid)
680 ns = kzalloc(sizeof(*ns), GFP_KERNEL);
684 init_completion(&ns->disable_done);
689 down_write(&nvmet_ana_sem);
690 ns->anagrpid = NVMET_DEFAULT_ANA_GRPID;
691 nvmet_ana_group_enabled[ns->anagrpid]++;
692 up_write(&nvmet_ana_sem);
695 ns->buffered_io = false;
700 static void nvmet_update_sq_head(struct nvmet_req *req)
703 u32 old_sqhd, new_sqhd;
706 old_sqhd = req->sq->sqhd;
707 new_sqhd = (old_sqhd + 1) % req->sq->size;
708 } while (cmpxchg(&req->sq->sqhd, old_sqhd, new_sqhd) !=
711 req->cqe->sq_head = cpu_to_le16(req->sq->sqhd & 0x0000FFFF);
714 static void nvmet_set_error(struct nvmet_req *req, u16 status)
716 struct nvmet_ctrl *ctrl = req->sq->ctrl;
717 struct nvme_error_slot *new_error_slot;
720 req->cqe->status = cpu_to_le16(status << 1);
722 if (!ctrl || req->error_loc == NVMET_NO_ERROR_LOC)
725 spin_lock_irqsave(&ctrl->error_lock, flags);
728 &ctrl->slots[ctrl->err_counter % NVMET_ERROR_LOG_SLOTS];
730 new_error_slot->error_count = cpu_to_le64(ctrl->err_counter);
731 new_error_slot->sqid = cpu_to_le16(req->sq->qid);
732 new_error_slot->cmdid = cpu_to_le16(req->cmd->common.command_id);
733 new_error_slot->status_field = cpu_to_le16(status << 1);
734 new_error_slot->param_error_location = cpu_to_le16(req->error_loc);
735 new_error_slot->lba = cpu_to_le64(req->error_slba);
736 new_error_slot->nsid = req->cmd->common.nsid;
737 spin_unlock_irqrestore(&ctrl->error_lock, flags);
739 /* set the more bit for this request */
740 req->cqe->status |= cpu_to_le16(1 << 14);
743 static void __nvmet_req_complete(struct nvmet_req *req, u16 status)
745 if (!req->sq->sqhd_disabled)
746 nvmet_update_sq_head(req);
747 req->cqe->sq_id = cpu_to_le16(req->sq->qid);
748 req->cqe->command_id = req->cmd->common.command_id;
750 if (unlikely(status))
751 nvmet_set_error(req, status);
753 trace_nvmet_req_complete(req);
756 nvmet_put_namespace(req->ns);
757 req->ops->queue_response(req);
760 void nvmet_req_complete(struct nvmet_req *req, u16 status)
762 __nvmet_req_complete(req, status);
763 percpu_ref_put(&req->sq->ref);
765 EXPORT_SYMBOL_GPL(nvmet_req_complete);
767 void nvmet_cq_setup(struct nvmet_ctrl *ctrl, struct nvmet_cq *cq,
774 void nvmet_sq_setup(struct nvmet_ctrl *ctrl, struct nvmet_sq *sq,
784 static void nvmet_confirm_sq(struct percpu_ref *ref)
786 struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
788 complete(&sq->confirm_done);
791 void nvmet_sq_destroy(struct nvmet_sq *sq)
793 struct nvmet_ctrl *ctrl = sq->ctrl;
796 * If this is the admin queue, complete all AERs so that our
797 * queue doesn't have outstanding requests on it.
799 if (ctrl && ctrl->sqs && ctrl->sqs[0] == sq)
800 nvmet_async_events_failall(ctrl);
801 percpu_ref_kill_and_confirm(&sq->ref, nvmet_confirm_sq);
802 wait_for_completion(&sq->confirm_done);
803 wait_for_completion(&sq->free_done);
804 percpu_ref_exit(&sq->ref);
808 * The teardown flow may take some time, and the host may not
809 * send us keep-alive during this period, hence reset the
810 * traffic based keep-alive timer so we don't trigger a
811 * controller teardown as a result of a keep-alive expiration.
813 ctrl->reset_tbkas = true;
814 nvmet_ctrl_put(ctrl);
815 sq->ctrl = NULL; /* allows reusing the queue later */
818 EXPORT_SYMBOL_GPL(nvmet_sq_destroy);
820 static void nvmet_sq_free(struct percpu_ref *ref)
822 struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
824 complete(&sq->free_done);
827 int nvmet_sq_init(struct nvmet_sq *sq)
831 ret = percpu_ref_init(&sq->ref, nvmet_sq_free, 0, GFP_KERNEL);
833 pr_err("percpu_ref init failed!\n");
836 init_completion(&sq->free_done);
837 init_completion(&sq->confirm_done);
841 EXPORT_SYMBOL_GPL(nvmet_sq_init);
843 static inline u16 nvmet_check_ana_state(struct nvmet_port *port,
846 enum nvme_ana_state state = port->ana_state[ns->anagrpid];
848 if (unlikely(state == NVME_ANA_INACCESSIBLE))
849 return NVME_SC_ANA_INACCESSIBLE;
850 if (unlikely(state == NVME_ANA_PERSISTENT_LOSS))
851 return NVME_SC_ANA_PERSISTENT_LOSS;
852 if (unlikely(state == NVME_ANA_CHANGE))
853 return NVME_SC_ANA_TRANSITION;
857 static inline u16 nvmet_io_cmd_check_access(struct nvmet_req *req)
859 if (unlikely(req->ns->readonly)) {
860 switch (req->cmd->common.opcode) {
865 return NVME_SC_NS_WRITE_PROTECTED;
872 static u16 nvmet_parse_io_cmd(struct nvmet_req *req)
876 ret = nvmet_check_ctrl_status(req);
880 if (nvmet_req_passthru_ctrl(req))
881 return nvmet_parse_passthru_io_cmd(req);
883 ret = nvmet_req_find_ns(req);
887 ret = nvmet_check_ana_state(req->port, req->ns);
889 req->error_loc = offsetof(struct nvme_common_command, nsid);
892 ret = nvmet_io_cmd_check_access(req);
894 req->error_loc = offsetof(struct nvme_common_command, nsid);
899 return nvmet_file_parse_io_cmd(req);
901 return nvmet_bdev_parse_io_cmd(req);
904 bool nvmet_req_init(struct nvmet_req *req, struct nvmet_cq *cq,
905 struct nvmet_sq *sq, const struct nvmet_fabrics_ops *ops)
907 u8 flags = req->cmd->common.flags;
914 req->metadata_sg = NULL;
916 req->metadata_sg_cnt = 0;
917 req->transfer_len = 0;
918 req->metadata_len = 0;
919 req->cqe->status = 0;
920 req->cqe->sq_head = 0;
922 req->error_loc = NVMET_NO_ERROR_LOC;
925 /* no support for fused commands yet */
926 if (unlikely(flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND))) {
927 req->error_loc = offsetof(struct nvme_common_command, flags);
928 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
933 * For fabrics, PSDT field shall describe metadata pointer (MPTR) that
934 * contains an address of a single contiguous physical buffer that is
937 if (unlikely((flags & NVME_CMD_SGL_ALL) != NVME_CMD_SGL_METABUF)) {
938 req->error_loc = offsetof(struct nvme_common_command, flags);
939 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
943 if (unlikely(!req->sq->ctrl))
944 /* will return an error for any non-connect command: */
945 status = nvmet_parse_connect_cmd(req);
946 else if (likely(req->sq->qid != 0))
947 status = nvmet_parse_io_cmd(req);
949 status = nvmet_parse_admin_cmd(req);
954 trace_nvmet_req_init(req, req->cmd);
956 if (unlikely(!percpu_ref_tryget_live(&sq->ref))) {
957 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
962 sq->ctrl->reset_tbkas = true;
967 __nvmet_req_complete(req, status);
970 EXPORT_SYMBOL_GPL(nvmet_req_init);
972 void nvmet_req_uninit(struct nvmet_req *req)
974 percpu_ref_put(&req->sq->ref);
976 nvmet_put_namespace(req->ns);
978 EXPORT_SYMBOL_GPL(nvmet_req_uninit);
980 bool nvmet_check_transfer_len(struct nvmet_req *req, size_t len)
982 if (unlikely(len != req->transfer_len)) {
983 req->error_loc = offsetof(struct nvme_common_command, dptr);
984 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
990 EXPORT_SYMBOL_GPL(nvmet_check_transfer_len);
992 bool nvmet_check_data_len_lte(struct nvmet_req *req, size_t data_len)
994 if (unlikely(data_len > req->transfer_len)) {
995 req->error_loc = offsetof(struct nvme_common_command, dptr);
996 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
1003 static unsigned int nvmet_data_transfer_len(struct nvmet_req *req)
1005 return req->transfer_len - req->metadata_len;
1008 static int nvmet_req_alloc_p2pmem_sgls(struct pci_dev *p2p_dev,
1009 struct nvmet_req *req)
1011 req->sg = pci_p2pmem_alloc_sgl(p2p_dev, &req->sg_cnt,
1012 nvmet_data_transfer_len(req));
1016 if (req->metadata_len) {
1017 req->metadata_sg = pci_p2pmem_alloc_sgl(p2p_dev,
1018 &req->metadata_sg_cnt, req->metadata_len);
1019 if (!req->metadata_sg)
1023 req->p2p_dev = p2p_dev;
1027 pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
1032 static struct pci_dev *nvmet_req_find_p2p_dev(struct nvmet_req *req)
1034 if (!IS_ENABLED(CONFIG_PCI_P2PDMA) ||
1035 !req->sq->ctrl || !req->sq->qid || !req->ns)
1037 return radix_tree_lookup(&req->sq->ctrl->p2p_ns_map, req->ns->nsid);
1040 int nvmet_req_alloc_sgls(struct nvmet_req *req)
1042 struct pci_dev *p2p_dev = nvmet_req_find_p2p_dev(req);
1044 if (p2p_dev && !nvmet_req_alloc_p2pmem_sgls(p2p_dev, req))
1047 req->sg = sgl_alloc(nvmet_data_transfer_len(req), GFP_KERNEL,
1049 if (unlikely(!req->sg))
1052 if (req->metadata_len) {
1053 req->metadata_sg = sgl_alloc(req->metadata_len, GFP_KERNEL,
1054 &req->metadata_sg_cnt);
1055 if (unlikely(!req->metadata_sg))
1065 EXPORT_SYMBOL_GPL(nvmet_req_alloc_sgls);
1067 void nvmet_req_free_sgls(struct nvmet_req *req)
1070 pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
1071 if (req->metadata_sg)
1072 pci_p2pmem_free_sgl(req->p2p_dev, req->metadata_sg);
1073 req->p2p_dev = NULL;
1076 if (req->metadata_sg)
1077 sgl_free(req->metadata_sg);
1081 req->metadata_sg = NULL;
1083 req->metadata_sg_cnt = 0;
1085 EXPORT_SYMBOL_GPL(nvmet_req_free_sgls);
1087 static inline bool nvmet_cc_en(u32 cc)
1089 return (cc >> NVME_CC_EN_SHIFT) & 0x1;
1092 static inline u8 nvmet_cc_css(u32 cc)
1094 return (cc >> NVME_CC_CSS_SHIFT) & 0x7;
1097 static inline u8 nvmet_cc_mps(u32 cc)
1099 return (cc >> NVME_CC_MPS_SHIFT) & 0xf;
1102 static inline u8 nvmet_cc_ams(u32 cc)
1104 return (cc >> NVME_CC_AMS_SHIFT) & 0x7;
1107 static inline u8 nvmet_cc_shn(u32 cc)
1109 return (cc >> NVME_CC_SHN_SHIFT) & 0x3;
1112 static inline u8 nvmet_cc_iosqes(u32 cc)
1114 return (cc >> NVME_CC_IOSQES_SHIFT) & 0xf;
1117 static inline u8 nvmet_cc_iocqes(u32 cc)
1119 return (cc >> NVME_CC_IOCQES_SHIFT) & 0xf;
1122 static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl)
1124 lockdep_assert_held(&ctrl->lock);
1127 * Only I/O controllers should verify iosqes,iocqes.
1128 * Strictly speaking, the spec says a discovery controller
1129 * should verify iosqes,iocqes are zeroed, however that
1130 * would break backwards compatibility, so don't enforce it.
1132 if (ctrl->subsys->type != NVME_NQN_DISC &&
1133 (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES ||
1134 nvmet_cc_iocqes(ctrl->cc) != NVME_NVM_IOCQES)) {
1135 ctrl->csts = NVME_CSTS_CFS;
1139 if (nvmet_cc_mps(ctrl->cc) != 0 ||
1140 nvmet_cc_ams(ctrl->cc) != 0 ||
1141 nvmet_cc_css(ctrl->cc) != 0) {
1142 ctrl->csts = NVME_CSTS_CFS;
1146 ctrl->csts = NVME_CSTS_RDY;
1149 * Controllers that are not yet enabled should not really enforce the
1150 * keep alive timeout, but we still want to track a timeout and cleanup
1151 * in case a host died before it enabled the controller. Hence, simply
1152 * reset the keep alive timer when the controller is enabled.
1155 mod_delayed_work(system_wq, &ctrl->ka_work, ctrl->kato * HZ);
1158 static void nvmet_clear_ctrl(struct nvmet_ctrl *ctrl)
1160 lockdep_assert_held(&ctrl->lock);
1162 /* XXX: tear down queues? */
1163 ctrl->csts &= ~NVME_CSTS_RDY;
1167 void nvmet_update_cc(struct nvmet_ctrl *ctrl, u32 new)
1171 mutex_lock(&ctrl->lock);
1175 if (nvmet_cc_en(new) && !nvmet_cc_en(old))
1176 nvmet_start_ctrl(ctrl);
1177 if (!nvmet_cc_en(new) && nvmet_cc_en(old))
1178 nvmet_clear_ctrl(ctrl);
1179 if (nvmet_cc_shn(new) && !nvmet_cc_shn(old)) {
1180 nvmet_clear_ctrl(ctrl);
1181 ctrl->csts |= NVME_CSTS_SHST_CMPLT;
1183 if (!nvmet_cc_shn(new) && nvmet_cc_shn(old))
1184 ctrl->csts &= ~NVME_CSTS_SHST_CMPLT;
1185 mutex_unlock(&ctrl->lock);
1188 static void nvmet_init_cap(struct nvmet_ctrl *ctrl)
1190 /* command sets supported: NVMe command set: */
1191 ctrl->cap = (1ULL << 37);
1192 /* CC.EN timeout in 500msec units: */
1193 ctrl->cap |= (15ULL << 24);
1194 /* maximum queue entries supported: */
1195 ctrl->cap |= NVMET_QUEUE_SIZE - 1;
1198 struct nvmet_ctrl *nvmet_ctrl_find_get(const char *subsysnqn,
1199 const char *hostnqn, u16 cntlid,
1200 struct nvmet_req *req)
1202 struct nvmet_ctrl *ctrl = NULL;
1203 struct nvmet_subsys *subsys;
1205 subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1207 pr_warn("connect request for invalid subsystem %s!\n",
1209 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1213 mutex_lock(&subsys->lock);
1214 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
1215 if (ctrl->cntlid == cntlid) {
1216 if (strncmp(hostnqn, ctrl->hostnqn, NVMF_NQN_SIZE)) {
1217 pr_warn("hostnqn mismatch.\n");
1220 if (!kref_get_unless_zero(&ctrl->ref))
1228 ctrl = NULL; /* ctrl not found */
1229 pr_warn("could not find controller %d for subsys %s / host %s\n",
1230 cntlid, subsysnqn, hostnqn);
1231 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(cntlid);
1234 mutex_unlock(&subsys->lock);
1235 nvmet_subsys_put(subsys);
1240 u16 nvmet_check_ctrl_status(struct nvmet_req *req)
1242 if (unlikely(!(req->sq->ctrl->cc & NVME_CC_ENABLE))) {
1243 pr_err("got cmd %d while CC.EN == 0 on qid = %d\n",
1244 req->cmd->common.opcode, req->sq->qid);
1245 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1248 if (unlikely(!(req->sq->ctrl->csts & NVME_CSTS_RDY))) {
1249 pr_err("got cmd %d while CSTS.RDY == 0 on qid = %d\n",
1250 req->cmd->common.opcode, req->sq->qid);
1251 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1256 bool nvmet_host_allowed(struct nvmet_subsys *subsys, const char *hostnqn)
1258 struct nvmet_host_link *p;
1260 lockdep_assert_held(&nvmet_config_sem);
1262 if (subsys->allow_any_host)
1265 if (subsys->type == NVME_NQN_DISC) /* allow all access to disc subsys */
1268 list_for_each_entry(p, &subsys->hosts, entry) {
1269 if (!strcmp(nvmet_host_name(p->host), hostnqn))
1277 * Note: ctrl->subsys->lock should be held when calling this function
1279 static void nvmet_setup_p2p_ns_map(struct nvmet_ctrl *ctrl,
1280 struct nvmet_req *req)
1282 struct nvmet_ns *ns;
1285 if (!req->p2p_client)
1288 ctrl->p2p_client = get_device(req->p2p_client);
1290 xa_for_each(&ctrl->subsys->namespaces, idx, ns)
1291 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
1295 * Note: ctrl->subsys->lock should be held when calling this function
1297 static void nvmet_release_p2p_ns_map(struct nvmet_ctrl *ctrl)
1299 struct radix_tree_iter iter;
1302 radix_tree_for_each_slot(slot, &ctrl->p2p_ns_map, &iter, 0)
1303 pci_dev_put(radix_tree_deref_slot(slot));
1305 put_device(ctrl->p2p_client);
1308 static void nvmet_fatal_error_handler(struct work_struct *work)
1310 struct nvmet_ctrl *ctrl =
1311 container_of(work, struct nvmet_ctrl, fatal_err_work);
1313 pr_err("ctrl %d fatal error occurred!\n", ctrl->cntlid);
1314 ctrl->ops->delete_ctrl(ctrl);
1317 u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn,
1318 struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp)
1320 struct nvmet_subsys *subsys;
1321 struct nvmet_ctrl *ctrl;
1325 status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1326 subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1328 pr_warn("connect request for invalid subsystem %s!\n",
1330 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1331 req->error_loc = offsetof(struct nvme_common_command, dptr);
1335 down_read(&nvmet_config_sem);
1336 if (!nvmet_host_allowed(subsys, hostnqn)) {
1337 pr_info("connect by host %s for subsystem %s not allowed\n",
1338 hostnqn, subsysnqn);
1339 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(hostnqn);
1340 up_read(&nvmet_config_sem);
1341 status = NVME_SC_CONNECT_INVALID_HOST | NVME_SC_DNR;
1342 req->error_loc = offsetof(struct nvme_common_command, dptr);
1343 goto out_put_subsystem;
1345 up_read(&nvmet_config_sem);
1347 status = NVME_SC_INTERNAL;
1348 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1350 goto out_put_subsystem;
1351 mutex_init(&ctrl->lock);
1353 nvmet_init_cap(ctrl);
1355 ctrl->port = req->port;
1357 INIT_WORK(&ctrl->async_event_work, nvmet_async_event_work);
1358 INIT_LIST_HEAD(&ctrl->async_events);
1359 INIT_RADIX_TREE(&ctrl->p2p_ns_map, GFP_KERNEL);
1360 INIT_WORK(&ctrl->fatal_err_work, nvmet_fatal_error_handler);
1362 memcpy(ctrl->subsysnqn, subsysnqn, NVMF_NQN_SIZE);
1363 memcpy(ctrl->hostnqn, hostnqn, NVMF_NQN_SIZE);
1365 kref_init(&ctrl->ref);
1366 ctrl->subsys = subsys;
1367 WRITE_ONCE(ctrl->aen_enabled, NVMET_AEN_CFG_OPTIONAL);
1369 ctrl->changed_ns_list = kmalloc_array(NVME_MAX_CHANGED_NAMESPACES,
1370 sizeof(__le32), GFP_KERNEL);
1371 if (!ctrl->changed_ns_list)
1374 ctrl->sqs = kcalloc(subsys->max_qid + 1,
1375 sizeof(struct nvmet_sq *),
1378 goto out_free_changed_ns_list;
1380 if (subsys->cntlid_min > subsys->cntlid_max)
1383 ret = ida_simple_get(&cntlid_ida,
1384 subsys->cntlid_min, subsys->cntlid_max,
1387 status = NVME_SC_CONNECT_CTRL_BUSY | NVME_SC_DNR;
1392 ctrl->ops = req->ops;
1395 * Discovery controllers may use some arbitrary high value
1396 * in order to cleanup stale discovery sessions
1398 if ((ctrl->subsys->type == NVME_NQN_DISC) && !kato)
1399 kato = NVMET_DISC_KATO_MS;
1401 /* keep-alive timeout in seconds */
1402 ctrl->kato = DIV_ROUND_UP(kato, 1000);
1404 ctrl->err_counter = 0;
1405 spin_lock_init(&ctrl->error_lock);
1407 nvmet_start_keep_alive_timer(ctrl);
1409 mutex_lock(&subsys->lock);
1410 list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
1411 nvmet_setup_p2p_ns_map(ctrl, req);
1412 mutex_unlock(&subsys->lock);
1419 out_free_changed_ns_list:
1420 kfree(ctrl->changed_ns_list);
1424 nvmet_subsys_put(subsys);
1429 static void nvmet_ctrl_free(struct kref *ref)
1431 struct nvmet_ctrl *ctrl = container_of(ref, struct nvmet_ctrl, ref);
1432 struct nvmet_subsys *subsys = ctrl->subsys;
1434 mutex_lock(&subsys->lock);
1435 nvmet_release_p2p_ns_map(ctrl);
1436 list_del(&ctrl->subsys_entry);
1437 mutex_unlock(&subsys->lock);
1439 nvmet_stop_keep_alive_timer(ctrl);
1441 flush_work(&ctrl->async_event_work);
1442 cancel_work_sync(&ctrl->fatal_err_work);
1444 ida_simple_remove(&cntlid_ida, ctrl->cntlid);
1446 nvmet_async_events_free(ctrl);
1448 kfree(ctrl->changed_ns_list);
1451 nvmet_subsys_put(subsys);
1454 void nvmet_ctrl_put(struct nvmet_ctrl *ctrl)
1456 kref_put(&ctrl->ref, nvmet_ctrl_free);
1459 void nvmet_ctrl_fatal_error(struct nvmet_ctrl *ctrl)
1461 mutex_lock(&ctrl->lock);
1462 if (!(ctrl->csts & NVME_CSTS_CFS)) {
1463 ctrl->csts |= NVME_CSTS_CFS;
1464 schedule_work(&ctrl->fatal_err_work);
1466 mutex_unlock(&ctrl->lock);
1468 EXPORT_SYMBOL_GPL(nvmet_ctrl_fatal_error);
1470 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
1471 const char *subsysnqn)
1473 struct nvmet_subsys_link *p;
1478 if (!strcmp(NVME_DISC_SUBSYS_NAME, subsysnqn)) {
1479 if (!kref_get_unless_zero(&nvmet_disc_subsys->ref))
1481 return nvmet_disc_subsys;
1484 down_read(&nvmet_config_sem);
1485 list_for_each_entry(p, &port->subsystems, entry) {
1486 if (!strncmp(p->subsys->subsysnqn, subsysnqn,
1488 if (!kref_get_unless_zero(&p->subsys->ref))
1490 up_read(&nvmet_config_sem);
1494 up_read(&nvmet_config_sem);
1498 struct nvmet_subsys *nvmet_subsys_alloc(const char *subsysnqn,
1499 enum nvme_subsys_type type)
1501 struct nvmet_subsys *subsys;
1503 subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
1505 return ERR_PTR(-ENOMEM);
1507 subsys->ver = NVMET_DEFAULT_VS;
1508 /* generate a random serial number as our controllers are ephemeral: */
1509 get_random_bytes(&subsys->serial, sizeof(subsys->serial));
1513 subsys->max_qid = NVMET_NR_QUEUES;
1516 subsys->max_qid = 0;
1519 pr_err("%s: Unknown Subsystem type - %d\n", __func__, type);
1521 return ERR_PTR(-EINVAL);
1523 subsys->type = type;
1524 subsys->subsysnqn = kstrndup(subsysnqn, NVMF_NQN_SIZE,
1526 if (!subsys->subsysnqn) {
1528 return ERR_PTR(-ENOMEM);
1530 subsys->cntlid_min = NVME_CNTLID_MIN;
1531 subsys->cntlid_max = NVME_CNTLID_MAX;
1532 kref_init(&subsys->ref);
1534 mutex_init(&subsys->lock);
1535 xa_init(&subsys->namespaces);
1536 INIT_LIST_HEAD(&subsys->ctrls);
1537 INIT_LIST_HEAD(&subsys->hosts);
1542 static void nvmet_subsys_free(struct kref *ref)
1544 struct nvmet_subsys *subsys =
1545 container_of(ref, struct nvmet_subsys, ref);
1547 WARN_ON_ONCE(!xa_empty(&subsys->namespaces));
1549 xa_destroy(&subsys->namespaces);
1550 nvmet_passthru_subsys_free(subsys);
1552 kfree(subsys->subsysnqn);
1553 kfree(subsys->model_number);
1557 void nvmet_subsys_del_ctrls(struct nvmet_subsys *subsys)
1559 struct nvmet_ctrl *ctrl;
1561 mutex_lock(&subsys->lock);
1562 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
1563 ctrl->ops->delete_ctrl(ctrl);
1564 mutex_unlock(&subsys->lock);
1567 void nvmet_subsys_put(struct nvmet_subsys *subsys)
1569 kref_put(&subsys->ref, nvmet_subsys_free);
1572 static int __init nvmet_init(void)
1576 nvmet_ana_group_enabled[NVMET_DEFAULT_ANA_GRPID] = 1;
1578 buffered_io_wq = alloc_workqueue("nvmet-buffered-io-wq",
1580 if (!buffered_io_wq) {
1585 error = nvmet_init_discovery();
1587 goto out_free_work_queue;
1589 error = nvmet_init_configfs();
1591 goto out_exit_discovery;
1595 nvmet_exit_discovery();
1596 out_free_work_queue:
1597 destroy_workqueue(buffered_io_wq);
1602 static void __exit nvmet_exit(void)
1604 nvmet_exit_configfs();
1605 nvmet_exit_discovery();
1606 ida_destroy(&cntlid_ida);
1607 destroy_workqueue(buffered_io_wq);
1609 BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_entry) != 1024);
1610 BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_hdr) != 1024);
1613 module_init(nvmet_init);
1614 module_exit(nvmet_exit);
1616 MODULE_LICENSE("GPL v2");