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 __nvmet_req_complete(req, status);
753 percpu_ref_put(&req->sq->ref);
755 EXPORT_SYMBOL_GPL(nvmet_req_complete);
757 void nvmet_cq_setup(struct nvmet_ctrl *ctrl, struct nvmet_cq *cq,
764 void nvmet_sq_setup(struct nvmet_ctrl *ctrl, struct nvmet_sq *sq,
774 static void nvmet_confirm_sq(struct percpu_ref *ref)
776 struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
778 complete(&sq->confirm_done);
781 void nvmet_sq_destroy(struct nvmet_sq *sq)
783 struct nvmet_ctrl *ctrl = sq->ctrl;
786 * If this is the admin queue, complete all AERs so that our
787 * queue doesn't have outstanding requests on it.
789 if (ctrl && ctrl->sqs && ctrl->sqs[0] == sq)
790 nvmet_async_events_failall(ctrl);
791 percpu_ref_kill_and_confirm(&sq->ref, nvmet_confirm_sq);
792 wait_for_completion(&sq->confirm_done);
793 wait_for_completion(&sq->free_done);
794 percpu_ref_exit(&sq->ref);
798 * The teardown flow may take some time, and the host may not
799 * send us keep-alive during this period, hence reset the
800 * traffic based keep-alive timer so we don't trigger a
801 * controller teardown as a result of a keep-alive expiration.
803 ctrl->reset_tbkas = true;
804 nvmet_ctrl_put(ctrl);
805 sq->ctrl = NULL; /* allows reusing the queue later */
808 EXPORT_SYMBOL_GPL(nvmet_sq_destroy);
810 static void nvmet_sq_free(struct percpu_ref *ref)
812 struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
814 complete(&sq->free_done);
817 int nvmet_sq_init(struct nvmet_sq *sq)
821 ret = percpu_ref_init(&sq->ref, nvmet_sq_free, 0, GFP_KERNEL);
823 pr_err("percpu_ref init failed!\n");
826 init_completion(&sq->free_done);
827 init_completion(&sq->confirm_done);
831 EXPORT_SYMBOL_GPL(nvmet_sq_init);
833 static inline u16 nvmet_check_ana_state(struct nvmet_port *port,
836 enum nvme_ana_state state = port->ana_state[ns->anagrpid];
838 if (unlikely(state == NVME_ANA_INACCESSIBLE))
839 return NVME_SC_ANA_INACCESSIBLE;
840 if (unlikely(state == NVME_ANA_PERSISTENT_LOSS))
841 return NVME_SC_ANA_PERSISTENT_LOSS;
842 if (unlikely(state == NVME_ANA_CHANGE))
843 return NVME_SC_ANA_TRANSITION;
847 static inline u16 nvmet_io_cmd_check_access(struct nvmet_req *req)
849 if (unlikely(req->ns->readonly)) {
850 switch (req->cmd->common.opcode) {
855 return NVME_SC_NS_WRITE_PROTECTED;
862 static u16 nvmet_parse_io_cmd(struct nvmet_req *req)
864 struct nvme_command *cmd = req->cmd;
867 ret = nvmet_check_ctrl_status(req, cmd);
871 if (nvmet_req_passthru_ctrl(req))
872 return nvmet_parse_passthru_io_cmd(req);
874 req->ns = nvmet_find_namespace(req->sq->ctrl, cmd->rw.nsid);
875 if (unlikely(!req->ns)) {
876 req->error_loc = offsetof(struct nvme_common_command, nsid);
877 return NVME_SC_INVALID_NS | NVME_SC_DNR;
879 ret = nvmet_check_ana_state(req->port, req->ns);
881 req->error_loc = offsetof(struct nvme_common_command, nsid);
884 ret = nvmet_io_cmd_check_access(req);
886 req->error_loc = offsetof(struct nvme_common_command, nsid);
891 return nvmet_file_parse_io_cmd(req);
893 return nvmet_bdev_parse_io_cmd(req);
896 bool nvmet_req_init(struct nvmet_req *req, struct nvmet_cq *cq,
897 struct nvmet_sq *sq, const struct nvmet_fabrics_ops *ops)
899 u8 flags = req->cmd->common.flags;
906 req->metadata_sg = NULL;
908 req->metadata_sg_cnt = 0;
909 req->transfer_len = 0;
910 req->metadata_len = 0;
911 req->cqe->status = 0;
912 req->cqe->sq_head = 0;
914 req->error_loc = NVMET_NO_ERROR_LOC;
917 /* no support for fused commands yet */
918 if (unlikely(flags & (NVME_CMD_FUSE_FIRST | NVME_CMD_FUSE_SECOND))) {
919 req->error_loc = offsetof(struct nvme_common_command, flags);
920 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
925 * For fabrics, PSDT field shall describe metadata pointer (MPTR) that
926 * contains an address of a single contiguous physical buffer that is
929 if (unlikely((flags & NVME_CMD_SGL_ALL) != NVME_CMD_SGL_METABUF)) {
930 req->error_loc = offsetof(struct nvme_common_command, flags);
931 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
935 if (unlikely(!req->sq->ctrl))
936 /* will return an error for any non-connect command: */
937 status = nvmet_parse_connect_cmd(req);
938 else if (likely(req->sq->qid != 0))
939 status = nvmet_parse_io_cmd(req);
941 status = nvmet_parse_admin_cmd(req);
946 trace_nvmet_req_init(req, req->cmd);
948 if (unlikely(!percpu_ref_tryget_live(&sq->ref))) {
949 status = NVME_SC_INVALID_FIELD | NVME_SC_DNR;
954 sq->ctrl->reset_tbkas = true;
959 __nvmet_req_complete(req, status);
962 EXPORT_SYMBOL_GPL(nvmet_req_init);
964 void nvmet_req_uninit(struct nvmet_req *req)
966 percpu_ref_put(&req->sq->ref);
968 nvmet_put_namespace(req->ns);
970 EXPORT_SYMBOL_GPL(nvmet_req_uninit);
972 bool nvmet_check_transfer_len(struct nvmet_req *req, size_t len)
974 if (unlikely(len != req->transfer_len)) {
975 req->error_loc = offsetof(struct nvme_common_command, dptr);
976 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
982 EXPORT_SYMBOL_GPL(nvmet_check_transfer_len);
984 bool nvmet_check_data_len_lte(struct nvmet_req *req, size_t data_len)
986 if (unlikely(data_len > req->transfer_len)) {
987 req->error_loc = offsetof(struct nvme_common_command, dptr);
988 nvmet_req_complete(req, NVME_SC_SGL_INVALID_DATA | NVME_SC_DNR);
995 static unsigned int nvmet_data_transfer_len(struct nvmet_req *req)
997 return req->transfer_len - req->metadata_len;
1000 static int nvmet_req_alloc_p2pmem_sgls(struct pci_dev *p2p_dev,
1001 struct nvmet_req *req)
1003 req->sg = pci_p2pmem_alloc_sgl(p2p_dev, &req->sg_cnt,
1004 nvmet_data_transfer_len(req));
1008 if (req->metadata_len) {
1009 req->metadata_sg = pci_p2pmem_alloc_sgl(p2p_dev,
1010 &req->metadata_sg_cnt, req->metadata_len);
1011 if (!req->metadata_sg)
1015 req->p2p_dev = p2p_dev;
1019 pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
1024 static struct pci_dev *nvmet_req_find_p2p_dev(struct nvmet_req *req)
1026 if (!IS_ENABLED(CONFIG_PCI_P2PDMA) ||
1027 !req->sq->ctrl || !req->sq->qid || !req->ns)
1029 return radix_tree_lookup(&req->sq->ctrl->p2p_ns_map, req->ns->nsid);
1032 int nvmet_req_alloc_sgls(struct nvmet_req *req)
1034 struct pci_dev *p2p_dev = nvmet_req_find_p2p_dev(req);
1036 if (p2p_dev && !nvmet_req_alloc_p2pmem_sgls(p2p_dev, req))
1039 req->sg = sgl_alloc(nvmet_data_transfer_len(req), GFP_KERNEL,
1041 if (unlikely(!req->sg))
1044 if (req->metadata_len) {
1045 req->metadata_sg = sgl_alloc(req->metadata_len, GFP_KERNEL,
1046 &req->metadata_sg_cnt);
1047 if (unlikely(!req->metadata_sg))
1057 EXPORT_SYMBOL_GPL(nvmet_req_alloc_sgls);
1059 void nvmet_req_free_sgls(struct nvmet_req *req)
1062 pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
1063 if (req->metadata_sg)
1064 pci_p2pmem_free_sgl(req->p2p_dev, req->metadata_sg);
1065 req->p2p_dev = NULL;
1068 if (req->metadata_sg)
1069 sgl_free(req->metadata_sg);
1073 req->metadata_sg = NULL;
1075 req->metadata_sg_cnt = 0;
1077 EXPORT_SYMBOL_GPL(nvmet_req_free_sgls);
1079 static inline bool nvmet_cc_en(u32 cc)
1081 return (cc >> NVME_CC_EN_SHIFT) & 0x1;
1084 static inline u8 nvmet_cc_css(u32 cc)
1086 return (cc >> NVME_CC_CSS_SHIFT) & 0x7;
1089 static inline u8 nvmet_cc_mps(u32 cc)
1091 return (cc >> NVME_CC_MPS_SHIFT) & 0xf;
1094 static inline u8 nvmet_cc_ams(u32 cc)
1096 return (cc >> NVME_CC_AMS_SHIFT) & 0x7;
1099 static inline u8 nvmet_cc_shn(u32 cc)
1101 return (cc >> NVME_CC_SHN_SHIFT) & 0x3;
1104 static inline u8 nvmet_cc_iosqes(u32 cc)
1106 return (cc >> NVME_CC_IOSQES_SHIFT) & 0xf;
1109 static inline u8 nvmet_cc_iocqes(u32 cc)
1111 return (cc >> NVME_CC_IOCQES_SHIFT) & 0xf;
1114 static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl)
1116 lockdep_assert_held(&ctrl->lock);
1119 * Only I/O controllers should verify iosqes,iocqes.
1120 * Strictly speaking, the spec says a discovery controller
1121 * should verify iosqes,iocqes are zeroed, however that
1122 * would break backwards compatibility, so don't enforce it.
1124 if (ctrl->subsys->type != NVME_NQN_DISC &&
1125 (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES ||
1126 nvmet_cc_iocqes(ctrl->cc) != NVME_NVM_IOCQES)) {
1127 ctrl->csts = NVME_CSTS_CFS;
1131 if (nvmet_cc_mps(ctrl->cc) != 0 ||
1132 nvmet_cc_ams(ctrl->cc) != 0 ||
1133 nvmet_cc_css(ctrl->cc) != 0) {
1134 ctrl->csts = NVME_CSTS_CFS;
1138 ctrl->csts = NVME_CSTS_RDY;
1141 * Controllers that are not yet enabled should not really enforce the
1142 * keep alive timeout, but we still want to track a timeout and cleanup
1143 * in case a host died before it enabled the controller. Hence, simply
1144 * reset the keep alive timer when the controller is enabled.
1147 mod_delayed_work(system_wq, &ctrl->ka_work, ctrl->kato * HZ);
1150 static void nvmet_clear_ctrl(struct nvmet_ctrl *ctrl)
1152 lockdep_assert_held(&ctrl->lock);
1154 /* XXX: tear down queues? */
1155 ctrl->csts &= ~NVME_CSTS_RDY;
1159 void nvmet_update_cc(struct nvmet_ctrl *ctrl, u32 new)
1163 mutex_lock(&ctrl->lock);
1167 if (nvmet_cc_en(new) && !nvmet_cc_en(old))
1168 nvmet_start_ctrl(ctrl);
1169 if (!nvmet_cc_en(new) && nvmet_cc_en(old))
1170 nvmet_clear_ctrl(ctrl);
1171 if (nvmet_cc_shn(new) && !nvmet_cc_shn(old)) {
1172 nvmet_clear_ctrl(ctrl);
1173 ctrl->csts |= NVME_CSTS_SHST_CMPLT;
1175 if (!nvmet_cc_shn(new) && nvmet_cc_shn(old))
1176 ctrl->csts &= ~NVME_CSTS_SHST_CMPLT;
1177 mutex_unlock(&ctrl->lock);
1180 static void nvmet_init_cap(struct nvmet_ctrl *ctrl)
1182 /* command sets supported: NVMe command set: */
1183 ctrl->cap = (1ULL << 37);
1184 /* CC.EN timeout in 500msec units: */
1185 ctrl->cap |= (15ULL << 24);
1186 /* maximum queue entries supported: */
1187 ctrl->cap |= NVMET_QUEUE_SIZE - 1;
1190 u16 nvmet_ctrl_find_get(const char *subsysnqn, const char *hostnqn, u16 cntlid,
1191 struct nvmet_req *req, struct nvmet_ctrl **ret)
1193 struct nvmet_subsys *subsys;
1194 struct nvmet_ctrl *ctrl;
1197 subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1199 pr_warn("connect request for invalid subsystem %s!\n",
1201 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1202 return NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1205 mutex_lock(&subsys->lock);
1206 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
1207 if (ctrl->cntlid == cntlid) {
1208 if (strncmp(hostnqn, ctrl->hostnqn, NVMF_NQN_SIZE)) {
1209 pr_warn("hostnqn mismatch.\n");
1212 if (!kref_get_unless_zero(&ctrl->ref))
1220 pr_warn("could not find controller %d for subsys %s / host %s\n",
1221 cntlid, subsysnqn, hostnqn);
1222 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(cntlid);
1223 status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1226 mutex_unlock(&subsys->lock);
1227 nvmet_subsys_put(subsys);
1231 u16 nvmet_check_ctrl_status(struct nvmet_req *req, struct nvme_command *cmd)
1233 if (unlikely(!(req->sq->ctrl->cc & NVME_CC_ENABLE))) {
1234 pr_err("got cmd %d while CC.EN == 0 on qid = %d\n",
1235 cmd->common.opcode, req->sq->qid);
1236 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1239 if (unlikely(!(req->sq->ctrl->csts & NVME_CSTS_RDY))) {
1240 pr_err("got cmd %d while CSTS.RDY == 0 on qid = %d\n",
1241 cmd->common.opcode, req->sq->qid);
1242 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1247 bool nvmet_host_allowed(struct nvmet_subsys *subsys, const char *hostnqn)
1249 struct nvmet_host_link *p;
1251 lockdep_assert_held(&nvmet_config_sem);
1253 if (subsys->allow_any_host)
1256 if (subsys->type == NVME_NQN_DISC) /* allow all access to disc subsys */
1259 list_for_each_entry(p, &subsys->hosts, entry) {
1260 if (!strcmp(nvmet_host_name(p->host), hostnqn))
1268 * Note: ctrl->subsys->lock should be held when calling this function
1270 static void nvmet_setup_p2p_ns_map(struct nvmet_ctrl *ctrl,
1271 struct nvmet_req *req)
1273 struct nvmet_ns *ns;
1276 if (!req->p2p_client)
1279 ctrl->p2p_client = get_device(req->p2p_client);
1281 xa_for_each(&ctrl->subsys->namespaces, idx, ns)
1282 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
1286 * Note: ctrl->subsys->lock should be held when calling this function
1288 static void nvmet_release_p2p_ns_map(struct nvmet_ctrl *ctrl)
1290 struct radix_tree_iter iter;
1293 radix_tree_for_each_slot(slot, &ctrl->p2p_ns_map, &iter, 0)
1294 pci_dev_put(radix_tree_deref_slot(slot));
1296 put_device(ctrl->p2p_client);
1299 static void nvmet_fatal_error_handler(struct work_struct *work)
1301 struct nvmet_ctrl *ctrl =
1302 container_of(work, struct nvmet_ctrl, fatal_err_work);
1304 pr_err("ctrl %d fatal error occurred!\n", ctrl->cntlid);
1305 ctrl->ops->delete_ctrl(ctrl);
1308 u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn,
1309 struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp)
1311 struct nvmet_subsys *subsys;
1312 struct nvmet_ctrl *ctrl;
1316 status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1317 subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1319 pr_warn("connect request for invalid subsystem %s!\n",
1321 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1325 status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1326 down_read(&nvmet_config_sem);
1327 if (!nvmet_host_allowed(subsys, hostnqn)) {
1328 pr_info("connect by host %s for subsystem %s not allowed\n",
1329 hostnqn, subsysnqn);
1330 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(hostnqn);
1331 up_read(&nvmet_config_sem);
1332 status = NVME_SC_CONNECT_INVALID_HOST | NVME_SC_DNR;
1333 goto out_put_subsystem;
1335 up_read(&nvmet_config_sem);
1337 status = NVME_SC_INTERNAL;
1338 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1340 goto out_put_subsystem;
1341 mutex_init(&ctrl->lock);
1343 nvmet_init_cap(ctrl);
1345 ctrl->port = req->port;
1347 INIT_WORK(&ctrl->async_event_work, nvmet_async_event_work);
1348 INIT_LIST_HEAD(&ctrl->async_events);
1349 INIT_RADIX_TREE(&ctrl->p2p_ns_map, GFP_KERNEL);
1350 INIT_WORK(&ctrl->fatal_err_work, nvmet_fatal_error_handler);
1352 memcpy(ctrl->subsysnqn, subsysnqn, NVMF_NQN_SIZE);
1353 memcpy(ctrl->hostnqn, hostnqn, NVMF_NQN_SIZE);
1355 kref_init(&ctrl->ref);
1356 ctrl->subsys = subsys;
1357 WRITE_ONCE(ctrl->aen_enabled, NVMET_AEN_CFG_OPTIONAL);
1359 ctrl->changed_ns_list = kmalloc_array(NVME_MAX_CHANGED_NAMESPACES,
1360 sizeof(__le32), GFP_KERNEL);
1361 if (!ctrl->changed_ns_list)
1364 ctrl->sqs = kcalloc(subsys->max_qid + 1,
1365 sizeof(struct nvmet_sq *),
1368 goto out_free_changed_ns_list;
1370 if (subsys->cntlid_min > subsys->cntlid_max)
1373 ret = ida_simple_get(&cntlid_ida,
1374 subsys->cntlid_min, subsys->cntlid_max,
1377 status = NVME_SC_CONNECT_CTRL_BUSY | NVME_SC_DNR;
1382 ctrl->ops = req->ops;
1385 * Discovery controllers may use some arbitrary high value
1386 * in order to cleanup stale discovery sessions
1388 if ((ctrl->subsys->type == NVME_NQN_DISC) && !kato)
1389 kato = NVMET_DISC_KATO_MS;
1391 /* keep-alive timeout in seconds */
1392 ctrl->kato = DIV_ROUND_UP(kato, 1000);
1394 ctrl->err_counter = 0;
1395 spin_lock_init(&ctrl->error_lock);
1397 nvmet_start_keep_alive_timer(ctrl);
1399 mutex_lock(&subsys->lock);
1400 list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
1401 nvmet_setup_p2p_ns_map(ctrl, req);
1402 mutex_unlock(&subsys->lock);
1409 out_free_changed_ns_list:
1410 kfree(ctrl->changed_ns_list);
1414 nvmet_subsys_put(subsys);
1419 static void nvmet_ctrl_free(struct kref *ref)
1421 struct nvmet_ctrl *ctrl = container_of(ref, struct nvmet_ctrl, ref);
1422 struct nvmet_subsys *subsys = ctrl->subsys;
1424 mutex_lock(&subsys->lock);
1425 nvmet_release_p2p_ns_map(ctrl);
1426 list_del(&ctrl->subsys_entry);
1427 mutex_unlock(&subsys->lock);
1429 nvmet_stop_keep_alive_timer(ctrl);
1431 flush_work(&ctrl->async_event_work);
1432 cancel_work_sync(&ctrl->fatal_err_work);
1434 ida_simple_remove(&cntlid_ida, ctrl->cntlid);
1436 nvmet_async_events_free(ctrl);
1438 kfree(ctrl->changed_ns_list);
1441 nvmet_subsys_put(subsys);
1444 void nvmet_ctrl_put(struct nvmet_ctrl *ctrl)
1446 kref_put(&ctrl->ref, nvmet_ctrl_free);
1449 void nvmet_ctrl_fatal_error(struct nvmet_ctrl *ctrl)
1451 mutex_lock(&ctrl->lock);
1452 if (!(ctrl->csts & NVME_CSTS_CFS)) {
1453 ctrl->csts |= NVME_CSTS_CFS;
1454 schedule_work(&ctrl->fatal_err_work);
1456 mutex_unlock(&ctrl->lock);
1458 EXPORT_SYMBOL_GPL(nvmet_ctrl_fatal_error);
1460 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
1461 const char *subsysnqn)
1463 struct nvmet_subsys_link *p;
1468 if (!strcmp(NVME_DISC_SUBSYS_NAME, subsysnqn)) {
1469 if (!kref_get_unless_zero(&nvmet_disc_subsys->ref))
1471 return nvmet_disc_subsys;
1474 down_read(&nvmet_config_sem);
1475 list_for_each_entry(p, &port->subsystems, entry) {
1476 if (!strncmp(p->subsys->subsysnqn, subsysnqn,
1478 if (!kref_get_unless_zero(&p->subsys->ref))
1480 up_read(&nvmet_config_sem);
1484 up_read(&nvmet_config_sem);
1488 struct nvmet_subsys *nvmet_subsys_alloc(const char *subsysnqn,
1489 enum nvme_subsys_type type)
1491 struct nvmet_subsys *subsys;
1493 subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
1495 return ERR_PTR(-ENOMEM);
1497 subsys->ver = NVMET_DEFAULT_VS;
1498 /* generate a random serial number as our controllers are ephemeral: */
1499 get_random_bytes(&subsys->serial, sizeof(subsys->serial));
1503 subsys->max_qid = NVMET_NR_QUEUES;
1506 subsys->max_qid = 0;
1509 pr_err("%s: Unknown Subsystem type - %d\n", __func__, type);
1511 return ERR_PTR(-EINVAL);
1513 subsys->type = type;
1514 subsys->subsysnqn = kstrndup(subsysnqn, NVMF_NQN_SIZE,
1516 if (!subsys->subsysnqn) {
1518 return ERR_PTR(-ENOMEM);
1520 subsys->cntlid_min = NVME_CNTLID_MIN;
1521 subsys->cntlid_max = NVME_CNTLID_MAX;
1522 kref_init(&subsys->ref);
1524 mutex_init(&subsys->lock);
1525 xa_init(&subsys->namespaces);
1526 INIT_LIST_HEAD(&subsys->ctrls);
1527 INIT_LIST_HEAD(&subsys->hosts);
1532 static void nvmet_subsys_free(struct kref *ref)
1534 struct nvmet_subsys *subsys =
1535 container_of(ref, struct nvmet_subsys, ref);
1537 WARN_ON_ONCE(!xa_empty(&subsys->namespaces));
1539 xa_destroy(&subsys->namespaces);
1540 nvmet_passthru_subsys_free(subsys);
1542 kfree(subsys->subsysnqn);
1543 kfree_rcu(subsys->model, rcuhead);
1547 void nvmet_subsys_del_ctrls(struct nvmet_subsys *subsys)
1549 struct nvmet_ctrl *ctrl;
1551 mutex_lock(&subsys->lock);
1552 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
1553 ctrl->ops->delete_ctrl(ctrl);
1554 mutex_unlock(&subsys->lock);
1557 void nvmet_subsys_put(struct nvmet_subsys *subsys)
1559 kref_put(&subsys->ref, nvmet_subsys_free);
1562 static int __init nvmet_init(void)
1566 nvmet_ana_group_enabled[NVMET_DEFAULT_ANA_GRPID] = 1;
1568 buffered_io_wq = alloc_workqueue("nvmet-buffered-io-wq",
1570 if (!buffered_io_wq) {
1575 error = nvmet_init_discovery();
1577 goto out_free_work_queue;
1579 error = nvmet_init_configfs();
1581 goto out_exit_discovery;
1585 nvmet_exit_discovery();
1586 out_free_work_queue:
1587 destroy_workqueue(buffered_io_wq);
1592 static void __exit nvmet_exit(void)
1594 nvmet_exit_configfs();
1595 nvmet_exit_discovery();
1596 ida_destroy(&cntlid_ida);
1597 destroy_workqueue(buffered_io_wq);
1599 BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_entry) != 1024);
1600 BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_hdr) != 1024);
1603 module_init(nvmet_init);
1604 module_exit(nvmet_exit);
1606 MODULE_LICENSE("GPL v2");