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 cmd_seen = ctrl->cmd_seen;
393 ctrl->cmd_seen = 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 schedule_delayed_work(&ctrl->ka_work, ctrl->kato * HZ);
418 void nvmet_stop_keep_alive_timer(struct nvmet_ctrl *ctrl)
420 if (unlikely(ctrl->kato == 0))
423 pr_debug("ctrl %d stop keep-alive\n", ctrl->cntlid);
425 cancel_delayed_work_sync(&ctrl->ka_work);
428 u16 nvmet_req_find_ns(struct nvmet_req *req)
430 u32 nsid = le32_to_cpu(req->cmd->common.nsid);
432 req->ns = xa_load(&nvmet_req_subsys(req)->namespaces, nsid);
433 if (unlikely(!req->ns)) {
434 req->error_loc = offsetof(struct nvme_common_command, nsid);
435 return NVME_SC_INVALID_NS | NVME_SC_DNR;
438 percpu_ref_get(&req->ns->ref);
439 return NVME_SC_SUCCESS;
442 static void nvmet_destroy_namespace(struct percpu_ref *ref)
444 struct nvmet_ns *ns = container_of(ref, struct nvmet_ns, ref);
446 complete(&ns->disable_done);
449 void nvmet_put_namespace(struct nvmet_ns *ns)
451 percpu_ref_put(&ns->ref);
454 static void nvmet_ns_dev_disable(struct nvmet_ns *ns)
456 nvmet_bdev_ns_disable(ns);
457 nvmet_file_ns_disable(ns);
460 static int nvmet_p2pmem_ns_enable(struct nvmet_ns *ns)
463 struct pci_dev *p2p_dev;
469 pr_err("peer-to-peer DMA is not supported by non-block device namespaces\n");
473 if (!blk_queue_pci_p2pdma(ns->bdev->bd_disk->queue)) {
474 pr_err("peer-to-peer DMA is not supported by the driver of %s\n",
480 ret = pci_p2pdma_distance(ns->p2p_dev, nvmet_ns_dev(ns), true);
485 * Right now we just check that there is p2pmem available so
486 * we can report an error to the user right away if there
487 * is not. We'll find the actual device to use once we
488 * setup the controller when the port's device is available.
491 p2p_dev = pci_p2pmem_find(nvmet_ns_dev(ns));
493 pr_err("no peer-to-peer memory is available for %s\n",
498 pci_dev_put(p2p_dev);
505 * Note: ctrl->subsys->lock should be held when calling this function
507 static void nvmet_p2pmem_ns_add_p2p(struct nvmet_ctrl *ctrl,
510 struct device *clients[2];
511 struct pci_dev *p2p_dev;
514 if (!ctrl->p2p_client || !ns->use_p2pmem)
518 ret = pci_p2pdma_distance(ns->p2p_dev, ctrl->p2p_client, true);
522 p2p_dev = pci_dev_get(ns->p2p_dev);
524 clients[0] = ctrl->p2p_client;
525 clients[1] = nvmet_ns_dev(ns);
527 p2p_dev = pci_p2pmem_find_many(clients, ARRAY_SIZE(clients));
529 pr_err("no peer-to-peer memory is available that's supported by %s and %s\n",
530 dev_name(ctrl->p2p_client), ns->device_path);
535 ret = radix_tree_insert(&ctrl->p2p_ns_map, ns->nsid, p2p_dev);
537 pci_dev_put(p2p_dev);
539 pr_info("using p2pmem on %s for nsid %d\n", pci_name(p2p_dev),
543 void nvmet_ns_revalidate(struct nvmet_ns *ns)
545 loff_t oldsize = ns->size;
548 nvmet_bdev_ns_revalidate(ns);
550 nvmet_file_ns_revalidate(ns);
552 if (oldsize != ns->size)
553 nvmet_ns_changed(ns->subsys, ns->nsid);
556 int nvmet_ns_enable(struct nvmet_ns *ns)
558 struct nvmet_subsys *subsys = ns->subsys;
559 struct nvmet_ctrl *ctrl;
562 mutex_lock(&subsys->lock);
565 if (nvmet_passthru_ctrl(subsys)) {
566 pr_info("cannot enable both passthru and regular namespaces for a single subsystem");
574 if (subsys->nr_namespaces == NVMET_MAX_NAMESPACES)
577 ret = nvmet_bdev_ns_enable(ns);
579 ret = nvmet_file_ns_enable(ns);
583 ret = nvmet_p2pmem_ns_enable(ns);
585 goto out_dev_disable;
587 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
588 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
590 ret = percpu_ref_init(&ns->ref, nvmet_destroy_namespace,
595 if (ns->nsid > subsys->max_nsid)
596 subsys->max_nsid = ns->nsid;
598 ret = xa_insert(&subsys->namespaces, ns->nsid, ns, GFP_KERNEL);
600 goto out_restore_subsys_maxnsid;
602 subsys->nr_namespaces++;
604 nvmet_ns_changed(subsys, ns->nsid);
608 mutex_unlock(&subsys->lock);
611 out_restore_subsys_maxnsid:
612 subsys->max_nsid = nvmet_max_nsid(subsys);
613 percpu_ref_exit(&ns->ref);
615 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
616 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
618 nvmet_ns_dev_disable(ns);
622 void nvmet_ns_disable(struct nvmet_ns *ns)
624 struct nvmet_subsys *subsys = ns->subsys;
625 struct nvmet_ctrl *ctrl;
627 mutex_lock(&subsys->lock);
632 xa_erase(&ns->subsys->namespaces, ns->nsid);
633 if (ns->nsid == subsys->max_nsid)
634 subsys->max_nsid = nvmet_max_nsid(subsys);
636 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
637 pci_dev_put(radix_tree_delete(&ctrl->p2p_ns_map, ns->nsid));
639 mutex_unlock(&subsys->lock);
642 * Now that we removed the namespaces from the lookup list, we
643 * can kill the per_cpu ref and wait for any remaining references
644 * to be dropped, as well as a RCU grace period for anyone only
645 * using the namepace under rcu_read_lock(). Note that we can't
646 * use call_rcu here as we need to ensure the namespaces have
647 * been fully destroyed before unloading the module.
649 percpu_ref_kill(&ns->ref);
651 wait_for_completion(&ns->disable_done);
652 percpu_ref_exit(&ns->ref);
654 mutex_lock(&subsys->lock);
656 subsys->nr_namespaces--;
657 nvmet_ns_changed(subsys, ns->nsid);
658 nvmet_ns_dev_disable(ns);
660 mutex_unlock(&subsys->lock);
663 void nvmet_ns_free(struct nvmet_ns *ns)
665 nvmet_ns_disable(ns);
667 down_write(&nvmet_ana_sem);
668 nvmet_ana_group_enabled[ns->anagrpid]--;
669 up_write(&nvmet_ana_sem);
671 kfree(ns->device_path);
675 struct nvmet_ns *nvmet_ns_alloc(struct nvmet_subsys *subsys, u32 nsid)
679 ns = kzalloc(sizeof(*ns), GFP_KERNEL);
683 init_completion(&ns->disable_done);
688 down_write(&nvmet_ana_sem);
689 ns->anagrpid = NVMET_DEFAULT_ANA_GRPID;
690 nvmet_ana_group_enabled[ns->anagrpid]++;
691 up_write(&nvmet_ana_sem);
694 ns->buffered_io = false;
699 static void nvmet_update_sq_head(struct nvmet_req *req)
702 u32 old_sqhd, new_sqhd;
705 old_sqhd = req->sq->sqhd;
706 new_sqhd = (old_sqhd + 1) % req->sq->size;
707 } while (cmpxchg(&req->sq->sqhd, old_sqhd, new_sqhd) !=
710 req->cqe->sq_head = cpu_to_le16(req->sq->sqhd & 0x0000FFFF);
713 static void nvmet_set_error(struct nvmet_req *req, u16 status)
715 struct nvmet_ctrl *ctrl = req->sq->ctrl;
716 struct nvme_error_slot *new_error_slot;
719 req->cqe->status = cpu_to_le16(status << 1);
721 if (!ctrl || req->error_loc == NVMET_NO_ERROR_LOC)
724 spin_lock_irqsave(&ctrl->error_lock, flags);
727 &ctrl->slots[ctrl->err_counter % NVMET_ERROR_LOG_SLOTS];
729 new_error_slot->error_count = cpu_to_le64(ctrl->err_counter);
730 new_error_slot->sqid = cpu_to_le16(req->sq->qid);
731 new_error_slot->cmdid = cpu_to_le16(req->cmd->common.command_id);
732 new_error_slot->status_field = cpu_to_le16(status << 1);
733 new_error_slot->param_error_location = cpu_to_le16(req->error_loc);
734 new_error_slot->lba = cpu_to_le64(req->error_slba);
735 new_error_slot->nsid = req->cmd->common.nsid;
736 spin_unlock_irqrestore(&ctrl->error_lock, flags);
738 /* set the more bit for this request */
739 req->cqe->status |= cpu_to_le16(1 << 14);
742 static void __nvmet_req_complete(struct nvmet_req *req, u16 status)
744 if (!req->sq->sqhd_disabled)
745 nvmet_update_sq_head(req);
746 req->cqe->sq_id = cpu_to_le16(req->sq->qid);
747 req->cqe->command_id = req->cmd->common.command_id;
749 if (unlikely(status))
750 nvmet_set_error(req, status);
752 trace_nvmet_req_complete(req);
755 nvmet_put_namespace(req->ns);
756 req->ops->queue_response(req);
759 void nvmet_req_complete(struct nvmet_req *req, u16 status)
761 __nvmet_req_complete(req, status);
762 percpu_ref_put(&req->sq->ref);
764 EXPORT_SYMBOL_GPL(nvmet_req_complete);
766 void nvmet_cq_setup(struct nvmet_ctrl *ctrl, struct nvmet_cq *cq,
773 void nvmet_sq_setup(struct nvmet_ctrl *ctrl, struct nvmet_sq *sq,
783 static void nvmet_confirm_sq(struct percpu_ref *ref)
785 struct nvmet_sq *sq = container_of(ref, struct nvmet_sq, ref);
787 complete(&sq->confirm_done);
790 void nvmet_sq_destroy(struct nvmet_sq *sq)
792 struct nvmet_ctrl *ctrl = sq->ctrl;
795 * If this is the admin queue, complete all AERs so that our
796 * queue doesn't have outstanding requests on it.
798 if (ctrl && ctrl->sqs && ctrl->sqs[0] == sq)
799 nvmet_async_events_failall(ctrl);
800 percpu_ref_kill_and_confirm(&sq->ref, nvmet_confirm_sq);
801 wait_for_completion(&sq->confirm_done);
802 wait_for_completion(&sq->free_done);
803 percpu_ref_exit(&sq->ref);
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)
868 ret = nvmet_check_ctrl_status(req);
872 if (nvmet_req_passthru_ctrl(req))
873 return nvmet_parse_passthru_io_cmd(req);
875 ret = nvmet_req_find_ns(req);
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->cmd_seen = 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 nvmet_req *req)
1002 req->sg = pci_p2pmem_alloc_sgl(req->p2p_dev, &req->sg_cnt,
1003 nvmet_data_transfer_len(req));
1007 if (req->metadata_len) {
1008 req->metadata_sg = pci_p2pmem_alloc_sgl(req->p2p_dev,
1009 &req->metadata_sg_cnt, req->metadata_len);
1010 if (!req->metadata_sg)
1015 pci_p2pmem_free_sgl(req->p2p_dev, req->sg);
1020 static bool nvmet_req_find_p2p_dev(struct nvmet_req *req)
1022 if (!IS_ENABLED(CONFIG_PCI_P2PDMA))
1025 if (req->sq->ctrl && req->sq->qid && req->ns) {
1026 req->p2p_dev = radix_tree_lookup(&req->sq->ctrl->p2p_ns_map,
1032 req->p2p_dev = NULL;
1036 int nvmet_req_alloc_sgls(struct nvmet_req *req)
1038 if (nvmet_req_find_p2p_dev(req) && !nvmet_req_alloc_p2pmem_sgls(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);
1069 if (req->metadata_sg)
1070 sgl_free(req->metadata_sg);
1074 req->metadata_sg = NULL;
1076 req->metadata_sg_cnt = 0;
1078 EXPORT_SYMBOL_GPL(nvmet_req_free_sgls);
1080 static inline bool nvmet_cc_en(u32 cc)
1082 return (cc >> NVME_CC_EN_SHIFT) & 0x1;
1085 static inline u8 nvmet_cc_css(u32 cc)
1087 return (cc >> NVME_CC_CSS_SHIFT) & 0x7;
1090 static inline u8 nvmet_cc_mps(u32 cc)
1092 return (cc >> NVME_CC_MPS_SHIFT) & 0xf;
1095 static inline u8 nvmet_cc_ams(u32 cc)
1097 return (cc >> NVME_CC_AMS_SHIFT) & 0x7;
1100 static inline u8 nvmet_cc_shn(u32 cc)
1102 return (cc >> NVME_CC_SHN_SHIFT) & 0x3;
1105 static inline u8 nvmet_cc_iosqes(u32 cc)
1107 return (cc >> NVME_CC_IOSQES_SHIFT) & 0xf;
1110 static inline u8 nvmet_cc_iocqes(u32 cc)
1112 return (cc >> NVME_CC_IOCQES_SHIFT) & 0xf;
1115 static void nvmet_start_ctrl(struct nvmet_ctrl *ctrl)
1117 lockdep_assert_held(&ctrl->lock);
1120 * Only I/O controllers should verify iosqes,iocqes.
1121 * Strictly speaking, the spec says a discovery controller
1122 * should verify iosqes,iocqes are zeroed, however that
1123 * would break backwards compatibility, so don't enforce it.
1125 if (ctrl->subsys->type != NVME_NQN_DISC &&
1126 (nvmet_cc_iosqes(ctrl->cc) != NVME_NVM_IOSQES ||
1127 nvmet_cc_iocqes(ctrl->cc) != NVME_NVM_IOCQES)) {
1128 ctrl->csts = NVME_CSTS_CFS;
1132 if (nvmet_cc_mps(ctrl->cc) != 0 ||
1133 nvmet_cc_ams(ctrl->cc) != 0 ||
1134 nvmet_cc_css(ctrl->cc) != 0) {
1135 ctrl->csts = NVME_CSTS_CFS;
1139 ctrl->csts = NVME_CSTS_RDY;
1142 * Controllers that are not yet enabled should not really enforce the
1143 * keep alive timeout, but we still want to track a timeout and cleanup
1144 * in case a host died before it enabled the controller. Hence, simply
1145 * reset the keep alive timer when the controller is enabled.
1148 mod_delayed_work(system_wq, &ctrl->ka_work, ctrl->kato * HZ);
1151 static void nvmet_clear_ctrl(struct nvmet_ctrl *ctrl)
1153 lockdep_assert_held(&ctrl->lock);
1155 /* XXX: tear down queues? */
1156 ctrl->csts &= ~NVME_CSTS_RDY;
1160 void nvmet_update_cc(struct nvmet_ctrl *ctrl, u32 new)
1164 mutex_lock(&ctrl->lock);
1168 if (nvmet_cc_en(new) && !nvmet_cc_en(old))
1169 nvmet_start_ctrl(ctrl);
1170 if (!nvmet_cc_en(new) && nvmet_cc_en(old))
1171 nvmet_clear_ctrl(ctrl);
1172 if (nvmet_cc_shn(new) && !nvmet_cc_shn(old)) {
1173 nvmet_clear_ctrl(ctrl);
1174 ctrl->csts |= NVME_CSTS_SHST_CMPLT;
1176 if (!nvmet_cc_shn(new) && nvmet_cc_shn(old))
1177 ctrl->csts &= ~NVME_CSTS_SHST_CMPLT;
1178 mutex_unlock(&ctrl->lock);
1181 static void nvmet_init_cap(struct nvmet_ctrl *ctrl)
1183 /* command sets supported: NVMe command set: */
1184 ctrl->cap = (1ULL << 37);
1185 /* CC.EN timeout in 500msec units: */
1186 ctrl->cap |= (15ULL << 24);
1187 /* maximum queue entries supported: */
1188 ctrl->cap |= NVMET_QUEUE_SIZE - 1;
1191 struct nvmet_ctrl *nvmet_ctrl_find_get(const char *subsysnqn,
1192 const char *hostnqn, u16 cntlid,
1193 struct nvmet_req *req)
1195 struct nvmet_ctrl *ctrl = NULL;
1196 struct nvmet_subsys *subsys;
1198 subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1200 pr_warn("connect request for invalid subsystem %s!\n",
1202 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1206 mutex_lock(&subsys->lock);
1207 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
1208 if (ctrl->cntlid == cntlid) {
1209 if (strncmp(hostnqn, ctrl->hostnqn, NVMF_NQN_SIZE)) {
1210 pr_warn("hostnqn mismatch.\n");
1213 if (!kref_get_unless_zero(&ctrl->ref))
1221 ctrl = NULL; /* ctrl not found */
1222 pr_warn("could not find controller %d for subsys %s / host %s\n",
1223 cntlid, subsysnqn, hostnqn);
1224 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(cntlid);
1227 mutex_unlock(&subsys->lock);
1228 nvmet_subsys_put(subsys);
1233 u16 nvmet_check_ctrl_status(struct nvmet_req *req)
1235 if (unlikely(!(req->sq->ctrl->cc & NVME_CC_ENABLE))) {
1236 pr_err("got cmd %d while CC.EN == 0 on qid = %d\n",
1237 req->cmd->common.opcode, req->sq->qid);
1238 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1241 if (unlikely(!(req->sq->ctrl->csts & NVME_CSTS_RDY))) {
1242 pr_err("got cmd %d while CSTS.RDY == 0 on qid = %d\n",
1243 req->cmd->common.opcode, req->sq->qid);
1244 return NVME_SC_CMD_SEQ_ERROR | NVME_SC_DNR;
1249 bool nvmet_host_allowed(struct nvmet_subsys *subsys, const char *hostnqn)
1251 struct nvmet_host_link *p;
1253 lockdep_assert_held(&nvmet_config_sem);
1255 if (subsys->allow_any_host)
1258 if (subsys->type == NVME_NQN_DISC) /* allow all access to disc subsys */
1261 list_for_each_entry(p, &subsys->hosts, entry) {
1262 if (!strcmp(nvmet_host_name(p->host), hostnqn))
1270 * Note: ctrl->subsys->lock should be held when calling this function
1272 static void nvmet_setup_p2p_ns_map(struct nvmet_ctrl *ctrl,
1273 struct nvmet_req *req)
1275 struct nvmet_ns *ns;
1278 if (!req->p2p_client)
1281 ctrl->p2p_client = get_device(req->p2p_client);
1283 xa_for_each(&ctrl->subsys->namespaces, idx, ns)
1284 nvmet_p2pmem_ns_add_p2p(ctrl, ns);
1288 * Note: ctrl->subsys->lock should be held when calling this function
1290 static void nvmet_release_p2p_ns_map(struct nvmet_ctrl *ctrl)
1292 struct radix_tree_iter iter;
1295 radix_tree_for_each_slot(slot, &ctrl->p2p_ns_map, &iter, 0)
1296 pci_dev_put(radix_tree_deref_slot(slot));
1298 put_device(ctrl->p2p_client);
1301 static void nvmet_fatal_error_handler(struct work_struct *work)
1303 struct nvmet_ctrl *ctrl =
1304 container_of(work, struct nvmet_ctrl, fatal_err_work);
1306 pr_err("ctrl %d fatal error occurred!\n", ctrl->cntlid);
1307 ctrl->ops->delete_ctrl(ctrl);
1310 u16 nvmet_alloc_ctrl(const char *subsysnqn, const char *hostnqn,
1311 struct nvmet_req *req, u32 kato, struct nvmet_ctrl **ctrlp)
1313 struct nvmet_subsys *subsys;
1314 struct nvmet_ctrl *ctrl;
1318 status = NVME_SC_CONNECT_INVALID_PARAM | NVME_SC_DNR;
1319 subsys = nvmet_find_get_subsys(req->port, subsysnqn);
1321 pr_warn("connect request for invalid subsystem %s!\n",
1323 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(subsysnqn);
1324 req->error_loc = offsetof(struct nvme_common_command, dptr);
1328 down_read(&nvmet_config_sem);
1329 if (!nvmet_host_allowed(subsys, hostnqn)) {
1330 pr_info("connect by host %s for subsystem %s not allowed\n",
1331 hostnqn, subsysnqn);
1332 req->cqe->result.u32 = IPO_IATTR_CONNECT_DATA(hostnqn);
1333 up_read(&nvmet_config_sem);
1334 status = NVME_SC_CONNECT_INVALID_HOST | NVME_SC_DNR;
1335 req->error_loc = offsetof(struct nvme_common_command, dptr);
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);
1354 INIT_DELAYED_WORK(&ctrl->ka_work, nvmet_keep_alive_timer);
1356 memcpy(ctrl->subsysnqn, subsysnqn, NVMF_NQN_SIZE);
1357 memcpy(ctrl->hostnqn, hostnqn, NVMF_NQN_SIZE);
1359 kref_init(&ctrl->ref);
1360 ctrl->subsys = subsys;
1361 WRITE_ONCE(ctrl->aen_enabled, NVMET_AEN_CFG_OPTIONAL);
1363 ctrl->changed_ns_list = kmalloc_array(NVME_MAX_CHANGED_NAMESPACES,
1364 sizeof(__le32), GFP_KERNEL);
1365 if (!ctrl->changed_ns_list)
1368 ctrl->sqs = kcalloc(subsys->max_qid + 1,
1369 sizeof(struct nvmet_sq *),
1372 goto out_free_changed_ns_list;
1374 if (subsys->cntlid_min > subsys->cntlid_max)
1377 ret = ida_simple_get(&cntlid_ida,
1378 subsys->cntlid_min, subsys->cntlid_max,
1381 status = NVME_SC_CONNECT_CTRL_BUSY | NVME_SC_DNR;
1386 ctrl->ops = req->ops;
1389 * Discovery controllers may use some arbitrary high value
1390 * in order to cleanup stale discovery sessions
1392 if ((ctrl->subsys->type == NVME_NQN_DISC) && !kato)
1393 kato = NVMET_DISC_KATO_MS;
1395 /* keep-alive timeout in seconds */
1396 ctrl->kato = DIV_ROUND_UP(kato, 1000);
1398 ctrl->err_counter = 0;
1399 spin_lock_init(&ctrl->error_lock);
1401 nvmet_start_keep_alive_timer(ctrl);
1403 mutex_lock(&subsys->lock);
1404 list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
1405 nvmet_setup_p2p_ns_map(ctrl, req);
1406 mutex_unlock(&subsys->lock);
1413 out_free_changed_ns_list:
1414 kfree(ctrl->changed_ns_list);
1418 nvmet_subsys_put(subsys);
1423 static void nvmet_ctrl_free(struct kref *ref)
1425 struct nvmet_ctrl *ctrl = container_of(ref, struct nvmet_ctrl, ref);
1426 struct nvmet_subsys *subsys = ctrl->subsys;
1428 mutex_lock(&subsys->lock);
1429 nvmet_release_p2p_ns_map(ctrl);
1430 list_del(&ctrl->subsys_entry);
1431 mutex_unlock(&subsys->lock);
1433 nvmet_stop_keep_alive_timer(ctrl);
1435 flush_work(&ctrl->async_event_work);
1436 cancel_work_sync(&ctrl->fatal_err_work);
1438 ida_simple_remove(&cntlid_ida, ctrl->cntlid);
1440 nvmet_async_events_free(ctrl);
1442 kfree(ctrl->changed_ns_list);
1445 nvmet_subsys_put(subsys);
1448 void nvmet_ctrl_put(struct nvmet_ctrl *ctrl)
1450 kref_put(&ctrl->ref, nvmet_ctrl_free);
1453 void nvmet_ctrl_fatal_error(struct nvmet_ctrl *ctrl)
1455 mutex_lock(&ctrl->lock);
1456 if (!(ctrl->csts & NVME_CSTS_CFS)) {
1457 ctrl->csts |= NVME_CSTS_CFS;
1458 schedule_work(&ctrl->fatal_err_work);
1460 mutex_unlock(&ctrl->lock);
1462 EXPORT_SYMBOL_GPL(nvmet_ctrl_fatal_error);
1464 static struct nvmet_subsys *nvmet_find_get_subsys(struct nvmet_port *port,
1465 const char *subsysnqn)
1467 struct nvmet_subsys_link *p;
1472 if (!strcmp(NVME_DISC_SUBSYS_NAME, subsysnqn)) {
1473 if (!kref_get_unless_zero(&nvmet_disc_subsys->ref))
1475 return nvmet_disc_subsys;
1478 down_read(&nvmet_config_sem);
1479 list_for_each_entry(p, &port->subsystems, entry) {
1480 if (!strncmp(p->subsys->subsysnqn, subsysnqn,
1482 if (!kref_get_unless_zero(&p->subsys->ref))
1484 up_read(&nvmet_config_sem);
1488 up_read(&nvmet_config_sem);
1492 struct nvmet_subsys *nvmet_subsys_alloc(const char *subsysnqn,
1493 enum nvme_subsys_type type)
1495 struct nvmet_subsys *subsys;
1496 char serial[NVMET_SN_MAX_SIZE / 2];
1499 subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
1501 return ERR_PTR(-ENOMEM);
1503 subsys->ver = NVMET_DEFAULT_VS;
1504 /* generate a random serial number as our controllers are ephemeral: */
1505 get_random_bytes(&serial, sizeof(serial));
1506 bin2hex(subsys->serial, &serial, sizeof(serial));
1508 subsys->model_number = kstrdup(NVMET_DEFAULT_CTRL_MODEL, GFP_KERNEL);
1509 if (!subsys->model_number) {
1516 subsys->max_qid = NVMET_NR_QUEUES;
1519 subsys->max_qid = 0;
1522 pr_err("%s: Unknown Subsystem type - %d\n", __func__, type);
1526 subsys->type = type;
1527 subsys->subsysnqn = kstrndup(subsysnqn, NVMF_NQN_SIZE,
1529 if (!subsys->subsysnqn) {
1533 subsys->cntlid_min = NVME_CNTLID_MIN;
1534 subsys->cntlid_max = NVME_CNTLID_MAX;
1535 kref_init(&subsys->ref);
1537 mutex_init(&subsys->lock);
1538 xa_init(&subsys->namespaces);
1539 INIT_LIST_HEAD(&subsys->ctrls);
1540 INIT_LIST_HEAD(&subsys->hosts);
1545 kfree(subsys->model_number);
1548 return ERR_PTR(ret);
1551 static void nvmet_subsys_free(struct kref *ref)
1553 struct nvmet_subsys *subsys =
1554 container_of(ref, struct nvmet_subsys, ref);
1556 WARN_ON_ONCE(!xa_empty(&subsys->namespaces));
1558 xa_destroy(&subsys->namespaces);
1559 nvmet_passthru_subsys_free(subsys);
1561 kfree(subsys->subsysnqn);
1562 kfree(subsys->model_number);
1566 void nvmet_subsys_del_ctrls(struct nvmet_subsys *subsys)
1568 struct nvmet_ctrl *ctrl;
1570 mutex_lock(&subsys->lock);
1571 list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
1572 ctrl->ops->delete_ctrl(ctrl);
1573 mutex_unlock(&subsys->lock);
1576 void nvmet_subsys_put(struct nvmet_subsys *subsys)
1578 kref_put(&subsys->ref, nvmet_subsys_free);
1581 static int __init nvmet_init(void)
1585 nvmet_ana_group_enabled[NVMET_DEFAULT_ANA_GRPID] = 1;
1587 buffered_io_wq = alloc_workqueue("nvmet-buffered-io-wq",
1589 if (!buffered_io_wq) {
1594 error = nvmet_init_discovery();
1596 goto out_free_work_queue;
1598 error = nvmet_init_configfs();
1600 goto out_exit_discovery;
1604 nvmet_exit_discovery();
1605 out_free_work_queue:
1606 destroy_workqueue(buffered_io_wq);
1611 static void __exit nvmet_exit(void)
1613 nvmet_exit_configfs();
1614 nvmet_exit_discovery();
1615 ida_destroy(&cntlid_ida);
1616 destroy_workqueue(buffered_io_wq);
1618 BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_entry) != 1024);
1619 BUILD_BUG_ON(sizeof(struct nvmf_disc_rsp_page_hdr) != 1024);
1622 module_init(nvmet_init);
1623 module_exit(nvmet_exit);
1625 MODULE_LICENSE("GPL v2");