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)
118 unsigned long nsid = 0;
119 struct nvmet_ns *cur;
122 xa_for_each(&subsys->namespaces, idx, cur)
128 static u32 nvmet_async_event_result(struct nvmet_async_event *aen)
130 return aen->event_type | (aen->event_info << 8) | (aen->log_page << 16);
133 static void nvmet_async_events_failall(struct nvmet_ctrl *ctrl)
135 u16 status = NVME_SC_INTERNAL | NVME_SC_DNR;
136 struct nvmet_req *req;
138 mutex_lock(&ctrl->lock);
139 while (ctrl->nr_async_event_cmds) {
140 req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
141 mutex_unlock(&ctrl->lock);
142 nvmet_req_complete(req, status);
143 mutex_lock(&ctrl->lock);
145 mutex_unlock(&ctrl->lock);
148 static void nvmet_async_events_process(struct nvmet_ctrl *ctrl)
150 struct nvmet_async_event *aen;
151 struct nvmet_req *req;
153 mutex_lock(&ctrl->lock);
154 while (ctrl->nr_async_event_cmds && !list_empty(&ctrl->async_events)) {
155 aen = list_first_entry(&ctrl->async_events,
156 struct nvmet_async_event, entry);
157 req = ctrl->async_event_cmds[--ctrl->nr_async_event_cmds];
158 nvmet_set_result(req, nvmet_async_event_result(aen));
160 list_del(&aen->entry);
163 mutex_unlock(&ctrl->lock);
164 trace_nvmet_async_event(ctrl, req->cqe->result.u32);
165 nvmet_req_complete(req, 0);
166 mutex_lock(&ctrl->lock);
168 mutex_unlock(&ctrl->lock);
171 static void nvmet_async_events_free(struct nvmet_ctrl *ctrl)
173 struct nvmet_async_event *aen, *tmp;
175 mutex_lock(&ctrl->lock);
176 list_for_each_entry_safe(aen, tmp, &ctrl->async_events, entry) {
177 list_del(&aen->entry);
180 mutex_unlock(&ctrl->lock);
183 static void nvmet_async_event_work(struct work_struct *work)
185 struct nvmet_ctrl *ctrl =
186 container_of(work, struct nvmet_ctrl, async_event_work);
188 nvmet_async_events_process(ctrl);
191 void nvmet_add_async_event(struct nvmet_ctrl *ctrl, u8 event_type,
192 u8 event_info, u8 log_page)
194 struct nvmet_async_event *aen;
196 aen = kmalloc(sizeof(*aen), GFP_KERNEL);
200 aen->event_type = event_type;
201 aen->event_info = event_info;
202 aen->log_page = log_page;
204 mutex_lock(&ctrl->lock);
205 list_add_tail(&aen->entry, &ctrl->async_events);
206 mutex_unlock(&ctrl->lock);
208 schedule_work(&ctrl->async_event_work);
211 static void nvmet_add_to_changed_ns_log(struct nvmet_ctrl *ctrl, __le32 nsid)
215 mutex_lock(&ctrl->lock);
216 if (ctrl->nr_changed_ns > NVME_MAX_CHANGED_NAMESPACES)
219 for (i = 0; i < ctrl->nr_changed_ns; i++) {
220 if (ctrl->changed_ns_list[i] == nsid)
224 if (ctrl->nr_changed_ns == NVME_MAX_CHANGED_NAMESPACES) {
225 ctrl->changed_ns_list[0] = cpu_to_le32(0xffffffff);
226 ctrl->nr_changed_ns = U32_MAX;
230 ctrl->changed_ns_list[ctrl->nr_changed_ns++] = nsid;
232 mutex_unlock(&ctrl->lock);
235 void nvmet_ns_changed(struct nvmet_subsys *subsys, u32 nsid)
237 struct nvmet_ctrl *ctrl;
239 lockdep_assert_held(&subsys->lock);
241 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
242 nvmet_add_to_changed_ns_log(ctrl, cpu_to_le32(nsid));
243 if (nvmet_aen_bit_disabled(ctrl, NVME_AEN_BIT_NS_ATTR))
245 nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE,
246 NVME_AER_NOTICE_NS_CHANGED,
247 NVME_LOG_CHANGED_NS);
251 void nvmet_send_ana_event(struct nvmet_subsys *subsys,
252 struct nvmet_port *port)
254 struct nvmet_ctrl *ctrl;
256 mutex_lock(&subsys->lock);
257 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
258 if (port && ctrl->port != port)
260 if (nvmet_aen_bit_disabled(ctrl, NVME_AEN_BIT_ANA_CHANGE))
262 nvmet_add_async_event(ctrl, NVME_AER_TYPE_NOTICE,
263 NVME_AER_NOTICE_ANA, NVME_LOG_ANA);
265 mutex_unlock(&subsys->lock);
268 void nvmet_port_send_ana_event(struct nvmet_port *port)
270 struct nvmet_subsys_link *p;
272 down_read(&nvmet_config_sem);
273 list_for_each_entry(p, &port->subsystems, entry)
274 nvmet_send_ana_event(p->subsys, port);
275 up_read(&nvmet_config_sem);
278 int nvmet_register_transport(const struct nvmet_fabrics_ops *ops)
282 down_write(&nvmet_config_sem);
283 if (nvmet_transports[ops->type])
286 nvmet_transports[ops->type] = ops;
287 up_write(&nvmet_config_sem);
291 EXPORT_SYMBOL_GPL(nvmet_register_transport);
293 void nvmet_unregister_transport(const struct nvmet_fabrics_ops *ops)
295 down_write(&nvmet_config_sem);
296 nvmet_transports[ops->type] = NULL;
297 up_write(&nvmet_config_sem);
299 EXPORT_SYMBOL_GPL(nvmet_unregister_transport);
301 void nvmet_port_del_ctrls(struct nvmet_port *port, struct nvmet_subsys *subsys)
303 struct nvmet_ctrl *ctrl;
305 mutex_lock(&subsys->lock);
306 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
307 if (ctrl->port == port)
308 ctrl->ops->delete_ctrl(ctrl);
310 mutex_unlock(&subsys->lock);
313 int nvmet_enable_port(struct nvmet_port *port)
315 const struct nvmet_fabrics_ops *ops;
318 lockdep_assert_held(&nvmet_config_sem);
320 ops = nvmet_transports[port->disc_addr.trtype];
322 up_write(&nvmet_config_sem);
323 request_module("nvmet-transport-%d", port->disc_addr.trtype);
324 down_write(&nvmet_config_sem);
325 ops = nvmet_transports[port->disc_addr.trtype];
327 pr_err("transport type %d not supported\n",
328 port->disc_addr.trtype);
333 if (!try_module_get(ops->owner))
337 * If the user requested PI support and the transport isn't pi capable,
338 * don't enable the port.
340 if (port->pi_enable && !(ops->flags & NVMF_METADATA_SUPPORTED)) {
341 pr_err("T10-PI is not supported by transport type %d\n",
342 port->disc_addr.trtype);
347 ret = ops->add_port(port);
351 /* If the transport didn't set inline_data_size, then disable it. */
352 if (port->inline_data_size < 0)
353 port->inline_data_size = 0;
355 port->enabled = true;
360 module_put(ops->owner);
364 void nvmet_disable_port(struct nvmet_port *port)
366 const struct nvmet_fabrics_ops *ops;
368 lockdep_assert_held(&nvmet_config_sem);
370 port->enabled = false;
373 ops = nvmet_transports[port->disc_addr.trtype];
374 ops->remove_port(port);
375 module_put(ops->owner);
378 static void nvmet_keep_alive_timer(struct work_struct *work)
380 struct nvmet_ctrl *ctrl = container_of(to_delayed_work(work),
381 struct nvmet_ctrl, ka_work);
382 bool reset_tbkas = ctrl->reset_tbkas;
384 ctrl->reset_tbkas = false;
386 pr_debug("ctrl %d reschedule traffic based keep-alive timer\n",
388 schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
392 pr_err("ctrl %d keep-alive timer (%d seconds) expired!\n",
393 ctrl->cntlid, ctrl->kato);
395 nvmet_ctrl_fatal_error(ctrl);
398 void nvmet_start_keep_alive_timer(struct nvmet_ctrl *ctrl)
400 if (unlikely(ctrl->kato == 0))
403 pr_debug("ctrl %d start keep-alive timer for %d secs\n",
404 ctrl->cntlid, ctrl->kato);
406 INIT_DELAYED_WORK(&ctrl->ka_work, nvmet_keep_alive_timer);
407 schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
410 void nvmet_stop_keep_alive_timer(struct nvmet_ctrl *ctrl)
412 if (unlikely(ctrl->kato == 0))
415 pr_debug("ctrl %d stop keep-alive\n", ctrl->cntlid);
417 cancel_delayed_work_sync(&ctrl->ka_work);
420 struct nvmet_ns *nvmet_find_namespace(struct nvmet_ctrl *ctrl, __le32 nsid)
424 ns = xa_load(&ctrl->subsys->namespaces, le32_to_cpu(nsid));
426 percpu_ref_get(&ns->ref);
431 static void nvmet_destroy_namespace(struct percpu_ref *ref)
433 struct nvmet_ns *ns = container_of(ref, struct nvmet_ns, ref);
435 complete(&ns->disable_done);
438 void nvmet_put_namespace(struct nvmet_ns *ns)
440 percpu_ref_put(&ns->ref);
443 static void nvmet_ns_dev_disable(struct nvmet_ns *ns)
445 nvmet_bdev_ns_disable(ns);
446 nvmet_file_ns_disable(ns);
449 static int nvmet_p2pmem_ns_enable(struct nvmet_ns *ns)
452 struct pci_dev *p2p_dev;
458 pr_err("peer-to-peer DMA is not supported by non-block device namespaces\n");
462 if (!blk_queue_pci_p2pdma(ns->bdev->bd_disk->queue)) {
463 pr_err("peer-to-peer DMA is not supported by the driver of %s\n",
469 ret = pci_p2pdma_distance(ns->p2p_dev, nvmet_ns_dev(ns), true);
474 * Right now we just check that there is p2pmem available so
475 * we can report an error to the user right away if there
476 * is not. We'll find the actual device to use once we
477 * setup the controller when the port's device is available.
480 p2p_dev = pci_p2pmem_find(nvmet_ns_dev(ns));
482 pr_err("no peer-to-peer memory is available for %s\n",
487 pci_dev_put(p2p_dev);
494 * Note: ctrl->subsys->lock should be held when calling this function
496 static void nvmet_p2pmem_ns_add_p2p(struct nvmet_ctrl *ctrl,
499 struct device *clients[2];
500 struct pci_dev *p2p_dev;
503 if (!ctrl->p2p_client || !ns->use_p2pmem)
507 ret = pci_p2pdma_distance(ns->p2p_dev, ctrl->p2p_client, true);
511 p2p_dev = pci_dev_get(ns->p2p_dev);
513 clients[0] = ctrl->p2p_client;
514 clients[1] = nvmet_ns_dev(ns);
516 p2p_dev = pci_p2pmem_find_many(clients, ARRAY_SIZE(clients));
518 pr_err("no peer-to-peer memory is available that's supported by %s and %s\n",
519 dev_name(ctrl->p2p_client), ns->device_path);
524 ret = radix_tree_insert(&ctrl->p2p_ns_map, ns->nsid, p2p_dev);
526 pci_dev_put(p2p_dev);
528 pr_info("using p2pmem on %s for nsid %d\n", pci_name(p2p_dev),
532 void nvmet_ns_revalidate(struct nvmet_ns *ns)
534 loff_t oldsize = ns->size;
537 nvmet_bdev_ns_revalidate(ns);
539 nvmet_file_ns_revalidate(ns);
541 if (oldsize != ns->size)
542 nvmet_ns_changed(ns->subsys, ns->nsid);
545 int nvmet_ns_enable(struct nvmet_ns *ns)
547 struct nvmet_subsys *subsys = ns->subsys;
548 struct nvmet_ctrl *ctrl;
551 mutex_lock(&subsys->lock);
554 if (nvmet_passthru_ctrl(subsys)) {
555 pr_info("cannot enable both passthru and regular namespaces for a single subsystem");
563 if (subsys->nr_namespaces == NVMET_MAX_NAMESPACES)
566 ret = nvmet_bdev_ns_enable(ns);
568 ret = nvmet_file_ns_enable(ns);
572 ret = nvmet_p2pmem_ns_enable(ns);
574 goto out_dev_disable;
576 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
577 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
579 ret = percpu_ref_init(&ns->ref, nvmet_destroy_namespace,
584 if (ns->nsid > subsys->max_nsid)
585 subsys->max_nsid = ns->nsid;
587 ret = xa_insert(&subsys->namespaces, ns->nsid, ns, GFP_KERNEL);
589 goto out_restore_subsys_maxnsid;
591 subsys->nr_namespaces++;
593 nvmet_ns_changed(subsys, ns->nsid);
597 mutex_unlock(&subsys->lock);
600 out_restore_subsys_maxnsid:
601 subsys->max_nsid = nvmet_max_nsid(subsys);
602 percpu_ref_exit(&ns->ref);
604 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
605 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
607 nvmet_ns_dev_disable(ns);
611 void nvmet_ns_disable(struct nvmet_ns *ns)
613 struct nvmet_subsys *subsys = ns->subsys;
614 struct nvmet_ctrl *ctrl;
616 mutex_lock(&subsys->lock);
621 xa_erase(&ns->subsys->namespaces, ns->nsid);
622 if (ns->nsid == subsys->max_nsid)
623 subsys->max_nsid = nvmet_max_nsid(subsys);
625 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
626 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
628 mutex_unlock(&subsys->lock);
631 * Now that we removed the namespaces from the lookup list, we
632 * can kill the per_cpu ref and wait for any remaining references
633 * to be dropped, as well as a RCU grace period for anyone only
634 * using the namepace under rcu_read_lock(). Note that we can't
635 * use call_rcu here as we need to ensure the namespaces have
636 * been fully destroyed before unloading the module.
638 percpu_ref_kill(&ns->ref);
640 wait_for_completion(&ns->disable_done);
641 percpu_ref_exit(&ns->ref);
643 mutex_lock(&subsys->lock);
645 subsys->nr_namespaces--;
646 nvmet_ns_changed(subsys, ns->nsid);
647 nvmet_ns_dev_disable(ns);
649 mutex_unlock(&subsys->lock);
652 void nvmet_ns_free(struct nvmet_ns *ns)
654 nvmet_ns_disable(ns);
656 down_write(&nvmet_ana_sem);
657 nvmet_ana_group_enabled[ns->anagrpid]--;
658 up_write(&nvmet_ana_sem);
660 kfree(ns->device_path);
664 struct nvmet_ns *nvmet_ns_alloc(struct nvmet_subsys *subsys, u32 nsid)
668 ns = kzalloc(sizeof(*ns), GFP_KERNEL);
672 init_completion(&ns->disable_done);
677 down_write(&nvmet_ana_sem);
678 ns->anagrpid = NVMET_DEFAULT_ANA_GRPID;
679 nvmet_ana_group_enabled[ns->anagrpid]++;
680 up_write(&nvmet_ana_sem);
683 ns->buffered_io = false;
688 static void nvmet_update_sq_head(struct nvmet_req *req)
691 u32 old_sqhd, new_sqhd;
694 old_sqhd = req->sq->sqhd;
695 new_sqhd = (old_sqhd + 1) % req->sq->size;
696 } while (cmpxchg(&req->sq->sqhd, old_sqhd, new_sqhd) !=
699 req->cqe->sq_head = cpu_to_le16(req->sq->sqhd & 0x0000FFFF);
702 static void nvmet_set_error(struct nvmet_req *req, u16 status)
704 struct nvmet_ctrl *ctrl = req->sq->ctrl;
705 struct nvme_error_slot *new_error_slot;
708 req->cqe->status = cpu_to_le16(status << 1);
710 if (!ctrl || req->error_loc == NVMET_NO_ERROR_LOC)
713 spin_lock_irqsave(&ctrl->error_lock, flags);
716 &ctrl->slots[ctrl->err_counter % NVMET_ERROR_LOG_SLOTS];
718 new_error_slot->error_count = cpu_to_le64(ctrl->err_counter);
719 new_error_slot->sqid = cpu_to_le16(req->sq->qid);
720 new_error_slot->cmdid = cpu_to_le16(req->cmd->common.command_id);
721 new_error_slot->status_field = cpu_to_le16(status << 1);
722 new_error_slot->param_error_location = cpu_to_le16(req->error_loc);
723 new_error_slot->lba = cpu_to_le64(req->error_slba);
724 new_error_slot->nsid = req->cmd->common.nsid;
725 spin_unlock_irqrestore(&ctrl->error_lock, flags);
727 /* set the more bit for this request */
728 req->cqe->status |= cpu_to_le16(1 << 14);
731 static void __nvmet_req_complete(struct nvmet_req *req, u16 status)
733 struct nvmet_ns *ns = req->ns;
735 if (!req->sq->sqhd_disabled)
736 nvmet_update_sq_head(req);
737 req->cqe->sq_id = cpu_to_le16(req->sq->qid);
738 req->cqe->command_id = req->cmd->common.command_id;
740 if (unlikely(status))
741 nvmet_set_error(req, status);
743 trace_nvmet_req_complete(req);
745 req->ops->queue_response(req);
747 nvmet_put_namespace(ns);
750 void nvmet_req_complete(struct nvmet_req *req, u16 status)
752 struct nvmet_sq *sq = req->sq;
754 __nvmet_req_complete(req, status);
755 percpu_ref_put(&sq->ref);
757 EXPORT_SYMBOL_GPL(nvmet_req_complete);
759 void nvmet_cq_setup(struct nvmet_ctrl *ctrl, struct nvmet_cq *cq,
766 void nvmet_sq_setup(struct nvmet_ctrl *ctrl, struct nvmet_sq *sq,
776 static void nvmet_confirm_sq(struct percpu_ref *ref)
778 struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
780 complete(&sq->confirm_done);
783 void nvmet_sq_destroy(struct nvmet_sq *sq)
785 struct nvmet_ctrl *ctrl = sq->ctrl;
788 * If this is the admin queue, complete all AERs so that our
789 * queue doesn't have outstanding requests on it.
791 if (ctrl && ctrl->sqs && ctrl->sqs[0] == sq)
792 nvmet_async_events_failall(ctrl);
793 percpu_ref_kill_and_confirm(&sq->ref, nvmet_confirm_sq);
794 wait_for_completion(&sq->confirm_done);
795 wait_for_completion(&sq->free_done);
796 percpu_ref_exit(&sq->ref);
800 * The teardown flow may take some time, and the host may not
801 * send us keep-alive during this period, hence reset the
802 * traffic based keep-alive timer so we don't trigger a
803 * controller teardown as a result of a keep-alive expiration.
805 ctrl->reset_tbkas = true;
806 nvmet_ctrl_put(ctrl);
807 sq->ctrl = NULL; /* allows reusing the queue later */
810 EXPORT_SYMBOL_GPL(nvmet_sq_destroy);
812 static void nvmet_sq_free(struct percpu_ref *ref)
814 struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
816 complete(&sq->free_done);
819 int nvmet_sq_init(struct nvmet_sq *sq)
823 ret = percpu_ref_init(&sq->ref, nvmet_sq_free, 0, GFP_KERNEL);
825 pr_err("percpu_ref init failed!\n");
828 init_completion(&sq->free_done);
829 init_completion(&sq->confirm_done);
833 EXPORT_SYMBOL_GPL(nvmet_sq_init);
835 static inline u16 nvmet_check_ana_state(struct nvmet_port *port,
838 enum nvme_ana_state state = port->ana_state[ns->anagrpid];
840 if (unlikely(state == NVME_ANA_INACCESSIBLE))
841 return NVME_SC_ANA_INACCESSIBLE;
842 if (unlikely(state == NVME_ANA_PERSISTENT_LOSS))
843 return NVME_SC_ANA_PERSISTENT_LOSS;
844 if (unlikely(state == NVME_ANA_CHANGE))
845 return NVME_SC_ANA_TRANSITION;
849 static inline u16 nvmet_io_cmd_check_access(struct nvmet_req *req)
851 if (unlikely(req->ns->readonly)) {
852 switch (req->cmd->common.opcode) {
857 return NVME_SC_NS_WRITE_PROTECTED;
864 static u16 nvmet_parse_io_cmd(struct nvmet_req *req)
866 struct nvme_command *cmd = req->cmd;
869 ret = nvmet_check_ctrl_status(req, cmd);
873 if (nvmet_req_passthru_ctrl(req))
874 return nvmet_parse_passthru_io_cmd(req);
876 req->ns = nvmet_find_namespace(req->sq->ctrl, cmd->rw.nsid);
877 if (unlikely(!req->ns)) {
878 req->error_loc = offsetof(struct nvme_common_command, nsid);
879 return NVME_SC_INVALID_NS | NVME_SC_DNR;
881 ret = nvmet_check_ana_state(req->port, req->ns);
883 req->error_loc = offsetof(struct nvme_common_command, nsid);
886 ret = nvmet_io_cmd_check_access(req);
888 req->error_loc = offsetof(struct nvme_common_command, nsid);
893 return nvmet_file_parse_io_cmd(req);
895 return nvmet_bdev_parse_io_cmd(req);
898 bool nvmet_req_init(struct nvmet_req *req, struct nvmet_cq *cq,
899 struct nvmet_sq *sq, const struct nvmet_fabrics_ops *ops)
901 u8 flags = req->cmd->common.flags;
908 req->metadata_sg = NULL;
910 req->metadata_sg_cnt = 0;
911 req->transfer_len = 0;
912 req->metadata_len = 0;
913 req->cqe->status = 0;
914 req->cqe->sq_head = 0;
916 req->error_loc = NVMET_NO_ERROR_LOC;
919 /* no support for fused commands yet */
920 if (unlikely(flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND))) {
921 req->error_loc = offsetof(struct nvme_common_command, flags);
922 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
927 * For fabrics, PSDT field shall describe metadata pointer (MPTR) that
928 * contains an address of a single contiguous physical buffer that is
931 if (unlikely((flags & NVME_CMD_SGL_ALL) != NVME_CMD_SGL_METABUF)) {
932 req->error_loc = offsetof(struct nvme_common_command, flags);
933 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
937 if (unlikely(!req->sq->ctrl))
938 /* will return an error for any non-connect command: */
939 status = nvmet_parse_connect_cmd(req);
940 else if (likely(req->sq->qid != 0))
941 status = nvmet_parse_io_cmd(req);
943 status = nvmet_parse_admin_cmd(req);
948 trace_nvmet_req_init(req, req->cmd);
950 if (unlikely(!percpu_ref_tryget_live(&sq->ref))) {
951 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
956 sq->ctrl->reset_tbkas = true;
961 __nvmet_req_complete(req, status);
964 EXPORT_SYMBOL_GPL(nvmet_req_init);
966 void nvmet_req_uninit(struct nvmet_req *req)
968 percpu_ref_put(&req->sq->ref);
970 nvmet_put_namespace(req->ns);
972 EXPORT_SYMBOL_GPL(nvmet_req_uninit);
974 bool nvmet_check_transfer_len(struct nvmet_req *req, size_t len)
976 if (unlikely(len != req->transfer_len)) {
977 req->error_loc = offsetof(struct nvme_common_command, dptr);
978 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
984 EXPORT_SYMBOL_GPL(nvmet_check_transfer_len);
986 bool nvmet_check_data_len_lte(struct nvmet_req *req, size_t data_len)
988 if (unlikely(data_len > req->transfer_len)) {
989 req->error_loc = offsetof(struct nvme_common_command, dptr);
990 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
997 static unsigned int nvmet_data_transfer_len(struct nvmet_req *req)
999 return req->transfer_len - req->metadata_len;
1002 static int nvmet_req_alloc_p2pmem_sgls(struct pci_dev *p2p_dev,
1003 struct nvmet_req *req)
1005 req->sg = pci_p2pmem_alloc_sgl(p2p_dev, &req->sg_cnt,
1006 nvmet_data_transfer_len(req));
1010 if (req->metadata_len) {
1011 req->metadata_sg = pci_p2pmem_alloc_sgl(p2p_dev,
1012 &req->metadata_sg_cnt, req->metadata_len);
1013 if (!req->metadata_sg)
1017 req->p2p_dev = p2p_dev;
1021 pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
1026 static struct pci_dev *nvmet_req_find_p2p_dev(struct nvmet_req *req)
1028 if (!IS_ENABLED(CONFIG_PCI_P2PDMA) ||
1029 !req->sq->ctrl || !req->sq->qid || !req->ns)
1031 return radix_tree_lookup(&req->sq->ctrl->p2p_ns_map, req->ns->nsid);
1034 int nvmet_req_alloc_sgls(struct nvmet_req *req)
1036 struct pci_dev *p2p_dev = nvmet_req_find_p2p_dev(req);
1038 if (p2p_dev && !nvmet_req_alloc_p2pmem_sgls(p2p_dev, req))
1041 req->sg = sgl_alloc(nvmet_data_transfer_len(req), GFP_KERNEL,
1043 if (unlikely(!req->sg))
1046 if (req->metadata_len) {
1047 req->metadata_sg = sgl_alloc(req->metadata_len, GFP_KERNEL,
1048 &req->metadata_sg_cnt);
1049 if (unlikely(!req->metadata_sg))
1059 EXPORT_SYMBOL_GPL(nvmet_req_alloc_sgls);
1061 void nvmet_req_free_sgls(struct nvmet_req *req)
1064 pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
1065 if (req->metadata_sg)
1066 pci_p2pmem_free_sgl(req->p2p_dev, req->metadata_sg);
1067 req->p2p_dev = NULL;
1070 if (req->metadata_sg)
1071 sgl_free(req->metadata_sg);
1075 req->metadata_sg = NULL;
1077 req->metadata_sg_cnt = 0;
1079 EXPORT_SYMBOL_GPL(nvmet_req_free_sgls);
1081 static inline bool nvmet_cc_en(u32 cc)
1083 return (cc >> NVME_CC_EN_SHIFT) & 0x1;
1086 static inline u8 nvmet_cc_css(u32 cc)
1088 return (cc >> NVME_CC_CSS_SHIFT) & 0x7;
1091 static inline u8 nvmet_cc_mps(u32 cc)
1093 return (cc >> NVME_CC_MPS_SHIFT) & 0xf;
1096 static inline u8 nvmet_cc_ams(u32 cc)
1098 return (cc >> NVME_CC_AMS_SHIFT) & 0x7;
1101 static inline u8 nvmet_cc_shn(u32 cc)
1103 return (cc >> NVME_CC_SHN_SHIFT) & 0x3;
1106 static inline u8 nvmet_cc_iosqes(u32 cc)
1108 return (cc >> NVME_CC_IOSQES_SHIFT) & 0xf;
1111 static inline u8 nvmet_cc_iocqes(u32 cc)
1113 return (cc >> NVME_CC_IOCQES_SHIFT) & 0xf;
1116 static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl)
1118 lockdep_assert_held(&ctrl->lock);
1121 * Only I/O controllers should verify iosqes,iocqes.
1122 * Strictly speaking, the spec says a discovery controller
1123 * should verify iosqes,iocqes are zeroed, however that
1124 * would break backwards compatibility, so don't enforce it.
1126 if (ctrl->subsys->type != NVME_NQN_DISC &&
1127 (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES ||
1128 nvmet_cc_iocqes(ctrl->cc) != NVME_NVM_IOCQES)) {
1129 ctrl->csts = NVME_CSTS_CFS;
1133 if (nvmet_cc_mps(ctrl->cc) != 0 ||
1134 nvmet_cc_ams(ctrl->cc) != 0 ||
1135 nvmet_cc_css(ctrl->cc) != 0) {
1136 ctrl->csts = NVME_CSTS_CFS;
1140 ctrl->csts = NVME_CSTS_RDY;
1143 * Controllers that are not yet enabled should not really enforce the
1144 * keep alive timeout, but we still want to track a timeout and cleanup
1145 * in case a host died before it enabled the controller. Hence, simply
1146 * reset the keep alive timer when the controller is enabled.
1149 mod_delayed_work(system_wq, &ctrl->ka_work, ctrl->kato * HZ);
1152 static void nvmet_clear_ctrl(struct nvmet_ctrl *ctrl)
1154 lockdep_assert_held(&ctrl->lock);
1156 /* XXX: tear down queues? */
1157 ctrl->csts &= ~NVME_CSTS_RDY;
1161 void nvmet_update_cc(struct nvmet_ctrl *ctrl, u32 new)
1165 mutex_lock(&ctrl->lock);
1169 if (nvmet_cc_en(new) && !nvmet_cc_en(old))
1170 nvmet_start_ctrl(ctrl);
1171 if (!nvmet_cc_en(new) && nvmet_cc_en(old))
1172 nvmet_clear_ctrl(ctrl);
1173 if (nvmet_cc_shn(new) && !nvmet_cc_shn(old)) {
1174 nvmet_clear_ctrl(ctrl);
1175 ctrl->csts |= NVME_CSTS_SHST_CMPLT;
1177 if (!nvmet_cc_shn(new) && nvmet_cc_shn(old))
1178 ctrl->csts &= ~NVME_CSTS_SHST_CMPLT;
1179 mutex_unlock(&ctrl->lock);
1182 static void nvmet_init_cap(struct nvmet_ctrl *ctrl)
1184 /* command sets supported: NVMe command set: */
1185 ctrl->cap = (1ULL << 37);
1186 /* CC.EN timeout in 500msec units: */
1187 ctrl->cap |= (15ULL << 24);
1188 /* maximum queue entries supported: */
1189 ctrl->cap |= NVMET_QUEUE_SIZE - 1;
1192 struct nvmet_ctrl *nvmet_ctrl_find_get(const char *subsysnqn,
1193 const char *hostnqn, u16 cntlid,
1194 struct nvmet_req *req)
1196 struct nvmet_ctrl *ctrl = NULL;
1197 struct nvmet_subsys *subsys;
1199 subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1201 pr_warn("connect request for invalid subsystem %s!\n",
1203 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1207 mutex_lock(&subsys->lock);
1208 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
1209 if (ctrl->cntlid == cntlid) {
1210 if (strncmp(hostnqn, ctrl->hostnqn, NVMF_NQN_SIZE)) {
1211 pr_warn("hostnqn mismatch.\n");
1214 if (!kref_get_unless_zero(&ctrl->ref))
1222 ctrl = NULL; /* ctrl not found */
1223 pr_warn("could not find controller %d for subsys %s / host %s\n",
1224 cntlid, subsysnqn, hostnqn);
1225 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(cntlid);
1228 mutex_unlock(&subsys->lock);
1229 nvmet_subsys_put(subsys);
1234 u16 nvmet_check_ctrl_status(struct nvmet_req *req, struct nvme_command *cmd)
1236 if (unlikely(!(req->sq->ctrl->cc & NVME_CC_ENABLE))) {
1237 pr_err("got cmd %d while CC.EN == 0 on qid = %d\n",
1238 cmd->common.opcode, req->sq->qid);
1239 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1242 if (unlikely(!(req->sq->ctrl->csts & NVME_CSTS_RDY))) {
1243 pr_err("got cmd %d while CSTS.RDY == 0 on qid = %d\n",
1244 cmd->common.opcode, req->sq->qid);
1245 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1250 bool nvmet_host_allowed(struct nvmet_subsys *subsys, const char *hostnqn)
1252 struct nvmet_host_link *p;
1254 lockdep_assert_held(&nvmet_config_sem);
1256 if (subsys->allow_any_host)
1259 if (subsys->type == NVME_NQN_DISC) /* allow all access to disc subsys */
1262 list_for_each_entry(p, &subsys->hosts, entry) {
1263 if (!strcmp(nvmet_host_name(p->host), hostnqn))
1271 * Note: ctrl->subsys->lock should be held when calling this function
1273 static void nvmet_setup_p2p_ns_map(struct nvmet_ctrl *ctrl,
1274 struct nvmet_req *req)
1276 struct nvmet_ns *ns;
1279 if (!req->p2p_client)
1282 ctrl->p2p_client = get_device(req->p2p_client);
1284 xa_for_each(&ctrl->subsys->namespaces, idx, ns)
1285 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
1289 * Note: ctrl->subsys->lock should be held when calling this function
1291 static void nvmet_release_p2p_ns_map(struct nvmet_ctrl *ctrl)
1293 struct radix_tree_iter iter;
1296 radix_tree_for_each_slot(slot, &ctrl->p2p_ns_map, &iter, 0)
1297 pci_dev_put(radix_tree_deref_slot(slot));
1299 put_device(ctrl->p2p_client);
1302 static void nvmet_fatal_error_handler(struct work_struct *work)
1304 struct nvmet_ctrl *ctrl =
1305 container_of(work, struct nvmet_ctrl, fatal_err_work);
1307 pr_err("ctrl %d fatal error occurred!\n", ctrl->cntlid);
1308 ctrl->ops->delete_ctrl(ctrl);
1311 u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn,
1312 struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp)
1314 struct nvmet_subsys *subsys;
1315 struct nvmet_ctrl *ctrl;
1319 status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1320 subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1322 pr_warn("connect request for invalid subsystem %s!\n",
1324 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1328 status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1329 down_read(&nvmet_config_sem);
1330 if (!nvmet_host_allowed(subsys, hostnqn)) {
1331 pr_info("connect by host %s for subsystem %s not allowed\n",
1332 hostnqn, subsysnqn);
1333 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(hostnqn);
1334 up_read(&nvmet_config_sem);
1335 status = NVME_SC_CONNECT_INVALID_HOST | NVME_SC_DNR;
1336 goto out_put_subsystem;
1338 up_read(&nvmet_config_sem);
1340 status = NVME_SC_INTERNAL;
1341 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1343 goto out_put_subsystem;
1344 mutex_init(&ctrl->lock);
1346 nvmet_init_cap(ctrl);
1348 ctrl->port = req->port;
1350 INIT_WORK(&ctrl->async_event_work, nvmet_async_event_work);
1351 INIT_LIST_HEAD(&ctrl->async_events);
1352 INIT_RADIX_TREE(&ctrl->p2p_ns_map, GFP_KERNEL);
1353 INIT_WORK(&ctrl->fatal_err_work, nvmet_fatal_error_handler);
1355 memcpy(ctrl->subsysnqn, subsysnqn, NVMF_NQN_SIZE);
1356 memcpy(ctrl->hostnqn, hostnqn, NVMF_NQN_SIZE);
1358 kref_init(&ctrl->ref);
1359 ctrl->subsys = subsys;
1360 WRITE_ONCE(ctrl->aen_enabled, NVMET_AEN_CFG_OPTIONAL);
1362 ctrl->changed_ns_list = kmalloc_array(NVME_MAX_CHANGED_NAMESPACES,
1363 sizeof(__le32), GFP_KERNEL);
1364 if (!ctrl->changed_ns_list)
1367 ctrl->sqs = kcalloc(subsys->max_qid + 1,
1368 sizeof(struct nvmet_sq *),
1371 goto out_free_changed_ns_list;
1373 if (subsys->cntlid_min > subsys->cntlid_max)
1376 ret = ida_simple_get(&cntlid_ida,
1377 subsys->cntlid_min, subsys->cntlid_max,
1380 status = NVME_SC_CONNECT_CTRL_BUSY | NVME_SC_DNR;
1385 ctrl->ops = req->ops;
1388 * Discovery controllers may use some arbitrary high value
1389 * in order to cleanup stale discovery sessions
1391 if ((ctrl->subsys->type == NVME_NQN_DISC) && !kato)
1392 kato = NVMET_DISC_KATO_MS;
1394 /* keep-alive timeout in seconds */
1395 ctrl->kato = DIV_ROUND_UP(kato, 1000);
1397 ctrl->err_counter = 0;
1398 spin_lock_init(&ctrl->error_lock);
1400 nvmet_start_keep_alive_timer(ctrl);
1402 mutex_lock(&subsys->lock);
1403 list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
1404 nvmet_setup_p2p_ns_map(ctrl, req);
1405 mutex_unlock(&subsys->lock);
1412 out_free_changed_ns_list:
1413 kfree(ctrl->changed_ns_list);
1417 nvmet_subsys_put(subsys);
1422 static void nvmet_ctrl_free(struct kref *ref)
1424 struct nvmet_ctrl *ctrl = container_of(ref, struct nvmet_ctrl, ref);
1425 struct nvmet_subsys *subsys = ctrl->subsys;
1427 mutex_lock(&subsys->lock);
1428 nvmet_release_p2p_ns_map(ctrl);
1429 list_del(&ctrl->subsys_entry);
1430 mutex_unlock(&subsys->lock);
1432 nvmet_stop_keep_alive_timer(ctrl);
1434 flush_work(&ctrl->async_event_work);
1435 cancel_work_sync(&ctrl->fatal_err_work);
1437 ida_simple_remove(&cntlid_ida, ctrl->cntlid);
1439 nvmet_async_events_free(ctrl);
1441 kfree(ctrl->changed_ns_list);
1444 nvmet_subsys_put(subsys);
1447 void nvmet_ctrl_put(struct nvmet_ctrl *ctrl)
1449 kref_put(&ctrl->ref, nvmet_ctrl_free);
1452 void nvmet_ctrl_fatal_error(struct nvmet_ctrl *ctrl)
1454 mutex_lock(&ctrl->lock);
1455 if (!(ctrl->csts & NVME_CSTS_CFS)) {
1456 ctrl->csts |= NVME_CSTS_CFS;
1457 schedule_work(&ctrl->fatal_err_work);
1459 mutex_unlock(&ctrl->lock);
1461 EXPORT_SYMBOL_GPL(nvmet_ctrl_fatal_error);
1463 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
1464 const char *subsysnqn)
1466 struct nvmet_subsys_link *p;
1471 if (!strcmp(NVME_DISC_SUBSYS_NAME, subsysnqn)) {
1472 if (!kref_get_unless_zero(&nvmet_disc_subsys->ref))
1474 return nvmet_disc_subsys;
1477 down_read(&nvmet_config_sem);
1478 list_for_each_entry(p, &port->subsystems, entry) {
1479 if (!strncmp(p->subsys->subsysnqn, subsysnqn,
1481 if (!kref_get_unless_zero(&p->subsys->ref))
1483 up_read(&nvmet_config_sem);
1487 up_read(&nvmet_config_sem);
1491 struct nvmet_subsys *nvmet_subsys_alloc(const char *subsysnqn,
1492 enum nvme_subsys_type type)
1494 struct nvmet_subsys *subsys;
1496 subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
1498 return ERR_PTR(-ENOMEM);
1500 subsys->ver = NVMET_DEFAULT_VS;
1501 /* generate a random serial number as our controllers are ephemeral: */
1502 get_random_bytes(&subsys->serial, sizeof(subsys->serial));
1506 subsys->max_qid = NVMET_NR_QUEUES;
1509 subsys->max_qid = 0;
1512 pr_err("%s: Unknown Subsystem type - %d\n", __func__, type);
1514 return ERR_PTR(-EINVAL);
1516 subsys->type = type;
1517 subsys->subsysnqn = kstrndup(subsysnqn, NVMF_NQN_SIZE,
1519 if (!subsys->subsysnqn) {
1521 return ERR_PTR(-ENOMEM);
1523 subsys->cntlid_min = NVME_CNTLID_MIN;
1524 subsys->cntlid_max = NVME_CNTLID_MAX;
1525 kref_init(&subsys->ref);
1527 mutex_init(&subsys->lock);
1528 xa_init(&subsys->namespaces);
1529 INIT_LIST_HEAD(&subsys->ctrls);
1530 INIT_LIST_HEAD(&subsys->hosts);
1535 static void nvmet_subsys_free(struct kref *ref)
1537 struct nvmet_subsys *subsys =
1538 container_of(ref, struct nvmet_subsys, ref);
1540 WARN_ON_ONCE(!xa_empty(&subsys->namespaces));
1542 xa_destroy(&subsys->namespaces);
1543 nvmet_passthru_subsys_free(subsys);
1545 kfree(subsys->subsysnqn);
1546 kfree_rcu(subsys->model, rcuhead);
1550 void nvmet_subsys_del_ctrls(struct nvmet_subsys *subsys)
1552 struct nvmet_ctrl *ctrl;
1554 mutex_lock(&subsys->lock);
1555 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
1556 ctrl->ops->delete_ctrl(ctrl);
1557 mutex_unlock(&subsys->lock);
1560 void nvmet_subsys_put(struct nvmet_subsys *subsys)
1562 kref_put(&subsys->ref, nvmet_subsys_free);
1565 static int __init nvmet_init(void)
1569 nvmet_ana_group_enabled[NVMET_DEFAULT_ANA_GRPID] = 1;
1571 buffered_io_wq = alloc_workqueue("nvmet-buffered-io-wq",
1573 if (!buffered_io_wq) {
1578 error = nvmet_init_discovery();
1580 goto out_free_work_queue;
1582 error = nvmet_init_configfs();
1584 goto out_exit_discovery;
1588 nvmet_exit_discovery();
1589 out_free_work_queue:
1590 destroy_workqueue(buffered_io_wq);
1595 static void __exit nvmet_exit(void)
1597 nvmet_exit_configfs();
1598 nvmet_exit_discovery();
1599 ida_destroy(&cntlid_ida);
1600 destroy_workqueue(buffered_io_wq);
1602 BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_entry) != 1024);
1603 BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_hdr) != 1024);
1606 module_init(nvmet_init);
1607 module_exit(nvmet_exit);
1609 MODULE_LICENSE("GPL v2");