1 // SPDX-License-Identifier: GPL-2.0
3 * NVM Express device driver
4 * Copyright (c) 2011-2014, Intel Corporation.
7 #include <linux/blkdev.h>
8 #include <linux/blk-mq.h>
9 #include <linux/blk-integrity.h>
10 #include <linux/compat.h>
11 #include <linux/delay.h>
12 #include <linux/errno.h>
13 #include <linux/hdreg.h>
14 #include <linux/kernel.h>
15 #include <linux/module.h>
16 #include <linux/backing-dev.h>
17 #include <linux/slab.h>
18 #include <linux/types.h>
20 #include <linux/ptrace.h>
21 #include <linux/nvme_ioctl.h>
22 #include <linux/pm_qos.h>
23 #include <linux/ratelimit.h>
24 #include <asm/unaligned.h>
28 #include <linux/nvme-auth.h>
30 #define CREATE_TRACE_POINTS
33 #define NVME_MINORS (1U << MINORBITS)
36 struct nvme_ns_ids ids;
45 unsigned int admin_timeout = 60;
46 module_param(admin_timeout, uint, 0644);
47 MODULE_PARM_DESC(admin_timeout, "timeout in seconds for admin commands");
48 EXPORT_SYMBOL_GPL(admin_timeout);
50 unsigned int nvme_io_timeout = 30;
51 module_param_named(io_timeout, nvme_io_timeout, uint, 0644);
52 MODULE_PARM_DESC(io_timeout, "timeout in seconds for I/O");
53 EXPORT_SYMBOL_GPL(nvme_io_timeout);
55 static unsigned char shutdown_timeout = 5;
56 module_param(shutdown_timeout, byte, 0644);
57 MODULE_PARM_DESC(shutdown_timeout, "timeout in seconds for controller shutdown");
59 static u8 nvme_max_retries = 5;
60 module_param_named(max_retries, nvme_max_retries, byte, 0644);
61 MODULE_PARM_DESC(max_retries, "max number of retries a command may have");
63 static unsigned long default_ps_max_latency_us = 100000;
64 module_param(default_ps_max_latency_us, ulong, 0644);
65 MODULE_PARM_DESC(default_ps_max_latency_us,
66 "max power saving latency for new devices; use PM QOS to change per device");
68 static bool force_apst;
69 module_param(force_apst, bool, 0644);
70 MODULE_PARM_DESC(force_apst, "allow APST for newly enumerated devices even if quirked off");
72 static unsigned long apst_primary_timeout_ms = 100;
73 module_param(apst_primary_timeout_ms, ulong, 0644);
74 MODULE_PARM_DESC(apst_primary_timeout_ms,
75 "primary APST timeout in ms");
77 static unsigned long apst_secondary_timeout_ms = 2000;
78 module_param(apst_secondary_timeout_ms, ulong, 0644);
79 MODULE_PARM_DESC(apst_secondary_timeout_ms,
80 "secondary APST timeout in ms");
82 static unsigned long apst_primary_latency_tol_us = 15000;
83 module_param(apst_primary_latency_tol_us, ulong, 0644);
84 MODULE_PARM_DESC(apst_primary_latency_tol_us,
85 "primary APST latency tolerance in us");
87 static unsigned long apst_secondary_latency_tol_us = 100000;
88 module_param(apst_secondary_latency_tol_us, ulong, 0644);
89 MODULE_PARM_DESC(apst_secondary_latency_tol_us,
90 "secondary APST latency tolerance in us");
93 * nvme_wq - hosts nvme related works that are not reset or delete
94 * nvme_reset_wq - hosts nvme reset works
95 * nvme_delete_wq - hosts nvme delete works
97 * nvme_wq will host works such as scan, aen handling, fw activation,
98 * keep-alive, periodic reconnects etc. nvme_reset_wq
99 * runs reset works which also flush works hosted on nvme_wq for
100 * serialization purposes. nvme_delete_wq host controller deletion
101 * works which flush reset works for serialization.
103 struct workqueue_struct *nvme_wq;
104 EXPORT_SYMBOL_GPL(nvme_wq);
106 struct workqueue_struct *nvme_reset_wq;
107 EXPORT_SYMBOL_GPL(nvme_reset_wq);
109 struct workqueue_struct *nvme_delete_wq;
110 EXPORT_SYMBOL_GPL(nvme_delete_wq);
112 static LIST_HEAD(nvme_subsystems);
113 static DEFINE_MUTEX(nvme_subsystems_lock);
115 static DEFINE_IDA(nvme_instance_ida);
116 static dev_t nvme_ctrl_base_chr_devt;
117 static int nvme_class_uevent(const struct device *dev, struct kobj_uevent_env *env);
118 static const struct class nvme_class = {
120 .dev_uevent = nvme_class_uevent,
123 static const struct class nvme_subsys_class = {
124 .name = "nvme-subsystem",
127 static DEFINE_IDA(nvme_ns_chr_minor_ida);
128 static dev_t nvme_ns_chr_devt;
129 static const struct class nvme_ns_chr_class = {
130 .name = "nvme-generic",
133 static void nvme_put_subsystem(struct nvme_subsystem *subsys);
134 static void nvme_remove_invalid_namespaces(struct nvme_ctrl *ctrl,
136 static void nvme_update_keep_alive(struct nvme_ctrl *ctrl,
137 struct nvme_command *cmd);
139 void nvme_queue_scan(struct nvme_ctrl *ctrl)
142 * Only new queue scan work when admin and IO queues are both alive
144 if (nvme_ctrl_state(ctrl) == NVME_CTRL_LIVE && ctrl->tagset)
145 queue_work(nvme_wq, &ctrl->scan_work);
149 * Use this function to proceed with scheduling reset_work for a controller
150 * that had previously been set to the resetting state. This is intended for
151 * code paths that can't be interrupted by other reset attempts. A hot removal
152 * may prevent this from succeeding.
154 int nvme_try_sched_reset(struct nvme_ctrl *ctrl)
156 if (nvme_ctrl_state(ctrl) != NVME_CTRL_RESETTING)
158 if (!queue_work(nvme_reset_wq, &ctrl->reset_work))
162 EXPORT_SYMBOL_GPL(nvme_try_sched_reset);
164 static void nvme_failfast_work(struct work_struct *work)
166 struct nvme_ctrl *ctrl = container_of(to_delayed_work(work),
167 struct nvme_ctrl, failfast_work);
169 if (nvme_ctrl_state(ctrl) != NVME_CTRL_CONNECTING)
172 set_bit(NVME_CTRL_FAILFAST_EXPIRED, &ctrl->flags);
173 dev_info(ctrl->device, "failfast expired\n");
174 nvme_kick_requeue_lists(ctrl);
177 static inline void nvme_start_failfast_work(struct nvme_ctrl *ctrl)
179 if (!ctrl->opts || ctrl->opts->fast_io_fail_tmo == -1)
182 schedule_delayed_work(&ctrl->failfast_work,
183 ctrl->opts->fast_io_fail_tmo * HZ);
186 static inline void nvme_stop_failfast_work(struct nvme_ctrl *ctrl)
191 cancel_delayed_work_sync(&ctrl->failfast_work);
192 clear_bit(NVME_CTRL_FAILFAST_EXPIRED, &ctrl->flags);
196 int nvme_reset_ctrl(struct nvme_ctrl *ctrl)
198 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING))
200 if (!queue_work(nvme_reset_wq, &ctrl->reset_work))
204 EXPORT_SYMBOL_GPL(nvme_reset_ctrl);
206 int nvme_reset_ctrl_sync(struct nvme_ctrl *ctrl)
210 ret = nvme_reset_ctrl(ctrl);
212 flush_work(&ctrl->reset_work);
213 if (nvme_ctrl_state(ctrl) != NVME_CTRL_LIVE)
220 static void nvme_do_delete_ctrl(struct nvme_ctrl *ctrl)
222 dev_info(ctrl->device,
223 "Removing ctrl: NQN \"%s\"\n", nvmf_ctrl_subsysnqn(ctrl));
225 flush_work(&ctrl->reset_work);
226 nvme_stop_ctrl(ctrl);
227 nvme_remove_namespaces(ctrl);
228 ctrl->ops->delete_ctrl(ctrl);
229 nvme_uninit_ctrl(ctrl);
232 static void nvme_delete_ctrl_work(struct work_struct *work)
234 struct nvme_ctrl *ctrl =
235 container_of(work, struct nvme_ctrl, delete_work);
237 nvme_do_delete_ctrl(ctrl);
240 int nvme_delete_ctrl(struct nvme_ctrl *ctrl)
242 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_DELETING))
244 if (!queue_work(nvme_delete_wq, &ctrl->delete_work))
248 EXPORT_SYMBOL_GPL(nvme_delete_ctrl);
250 void nvme_delete_ctrl_sync(struct nvme_ctrl *ctrl)
253 * Keep a reference until nvme_do_delete_ctrl() complete,
254 * since ->delete_ctrl can free the controller.
257 if (nvme_change_ctrl_state(ctrl, NVME_CTRL_DELETING))
258 nvme_do_delete_ctrl(ctrl);
262 static blk_status_t nvme_error_status(u16 status)
264 switch (status & 0x7ff) {
265 case NVME_SC_SUCCESS:
267 case NVME_SC_CAP_EXCEEDED:
268 return BLK_STS_NOSPC;
269 case NVME_SC_LBA_RANGE:
270 case NVME_SC_CMD_INTERRUPTED:
271 case NVME_SC_NS_NOT_READY:
272 return BLK_STS_TARGET;
273 case NVME_SC_BAD_ATTRIBUTES:
274 case NVME_SC_ONCS_NOT_SUPPORTED:
275 case NVME_SC_INVALID_OPCODE:
276 case NVME_SC_INVALID_FIELD:
277 case NVME_SC_INVALID_NS:
278 return BLK_STS_NOTSUPP;
279 case NVME_SC_WRITE_FAULT:
280 case NVME_SC_READ_ERROR:
281 case NVME_SC_UNWRITTEN_BLOCK:
282 case NVME_SC_ACCESS_DENIED:
283 case NVME_SC_READ_ONLY:
284 case NVME_SC_COMPARE_FAILED:
285 return BLK_STS_MEDIUM;
286 case NVME_SC_GUARD_CHECK:
287 case NVME_SC_APPTAG_CHECK:
288 case NVME_SC_REFTAG_CHECK:
289 case NVME_SC_INVALID_PI:
290 return BLK_STS_PROTECTION;
291 case NVME_SC_RESERVATION_CONFLICT:
292 return BLK_STS_RESV_CONFLICT;
293 case NVME_SC_HOST_PATH_ERROR:
294 return BLK_STS_TRANSPORT;
295 case NVME_SC_ZONE_TOO_MANY_ACTIVE:
296 return BLK_STS_ZONE_ACTIVE_RESOURCE;
297 case NVME_SC_ZONE_TOO_MANY_OPEN:
298 return BLK_STS_ZONE_OPEN_RESOURCE;
300 return BLK_STS_IOERR;
304 static void nvme_retry_req(struct request *req)
306 unsigned long delay = 0;
309 /* The mask and shift result must be <= 3 */
310 crd = (nvme_req(req)->status & NVME_SC_CRD) >> 11;
312 delay = nvme_req(req)->ctrl->crdt[crd - 1] * 100;
314 nvme_req(req)->retries++;
315 blk_mq_requeue_request(req, false);
316 blk_mq_delay_kick_requeue_list(req->q, delay);
319 static void nvme_log_error(struct request *req)
321 struct nvme_ns *ns = req->q->queuedata;
322 struct nvme_request *nr = nvme_req(req);
325 pr_err_ratelimited("%s: %s(0x%x) @ LBA %llu, %u blocks, %s (sct 0x%x / sc 0x%x) %s%s\n",
326 ns->disk ? ns->disk->disk_name : "?",
327 nvme_get_opcode_str(nr->cmd->common.opcode),
328 nr->cmd->common.opcode,
329 nvme_sect_to_lba(ns->head, blk_rq_pos(req)),
330 blk_rq_bytes(req) >> ns->head->lba_shift,
331 nvme_get_error_status_str(nr->status),
332 nr->status >> 8 & 7, /* Status Code Type */
333 nr->status & 0xff, /* Status Code */
334 nr->status & NVME_SC_MORE ? "MORE " : "",
335 nr->status & NVME_SC_DNR ? "DNR " : "");
339 pr_err_ratelimited("%s: %s(0x%x), %s (sct 0x%x / sc 0x%x) %s%s\n",
340 dev_name(nr->ctrl->device),
341 nvme_get_admin_opcode_str(nr->cmd->common.opcode),
342 nr->cmd->common.opcode,
343 nvme_get_error_status_str(nr->status),
344 nr->status >> 8 & 7, /* Status Code Type */
345 nr->status & 0xff, /* Status Code */
346 nr->status & NVME_SC_MORE ? "MORE " : "",
347 nr->status & NVME_SC_DNR ? "DNR " : "");
350 static void nvme_log_err_passthru(struct request *req)
352 struct nvme_ns *ns = req->q->queuedata;
353 struct nvme_request *nr = nvme_req(req);
355 pr_err_ratelimited("%s: %s(0x%x), %s (sct 0x%x / sc 0x%x) %s%s"
356 "cdw10=0x%x cdw11=0x%x cdw12=0x%x cdw13=0x%x cdw14=0x%x cdw15=0x%x\n",
357 ns ? ns->disk->disk_name : dev_name(nr->ctrl->device),
358 ns ? nvme_get_opcode_str(nr->cmd->common.opcode) :
359 nvme_get_admin_opcode_str(nr->cmd->common.opcode),
360 nr->cmd->common.opcode,
361 nvme_get_error_status_str(nr->status),
362 nr->status >> 8 & 7, /* Status Code Type */
363 nr->status & 0xff, /* Status Code */
364 nr->status & NVME_SC_MORE ? "MORE " : "",
365 nr->status & NVME_SC_DNR ? "DNR " : "",
366 nr->cmd->common.cdw10,
367 nr->cmd->common.cdw11,
368 nr->cmd->common.cdw12,
369 nr->cmd->common.cdw13,
370 nr->cmd->common.cdw14,
371 nr->cmd->common.cdw14);
374 enum nvme_disposition {
381 static inline enum nvme_disposition nvme_decide_disposition(struct request *req)
383 if (likely(nvme_req(req)->status == 0))
386 if ((nvme_req(req)->status & 0x7ff) == NVME_SC_AUTH_REQUIRED)
389 if (blk_noretry_request(req) ||
390 (nvme_req(req)->status & NVME_SC_DNR) ||
391 nvme_req(req)->retries >= nvme_max_retries)
394 if (req->cmd_flags & REQ_NVME_MPATH) {
395 if (nvme_is_path_error(nvme_req(req)->status) ||
396 blk_queue_dying(req->q))
399 if (blk_queue_dying(req->q))
406 static inline void nvme_end_req_zoned(struct request *req)
408 if (IS_ENABLED(CONFIG_BLK_DEV_ZONED) &&
409 req_op(req) == REQ_OP_ZONE_APPEND) {
410 struct nvme_ns *ns = req->q->queuedata;
412 req->__sector = nvme_lba_to_sect(ns->head,
413 le64_to_cpu(nvme_req(req)->result.u64));
417 static inline void nvme_end_req(struct request *req)
419 blk_status_t status = nvme_error_status(nvme_req(req)->status);
421 if (unlikely(nvme_req(req)->status && !(req->rq_flags & RQF_QUIET))) {
422 if (blk_rq_is_passthrough(req))
423 nvme_log_err_passthru(req);
427 nvme_end_req_zoned(req);
428 nvme_trace_bio_complete(req);
429 if (req->cmd_flags & REQ_NVME_MPATH)
430 nvme_mpath_end_request(req);
431 blk_mq_end_request(req, status);
434 void nvme_complete_rq(struct request *req)
436 struct nvme_ctrl *ctrl = nvme_req(req)->ctrl;
438 trace_nvme_complete_rq(req);
439 nvme_cleanup_cmd(req);
442 * Completions of long-running commands should not be able to
443 * defer sending of periodic keep alives, since the controller
444 * may have completed processing such commands a long time ago
445 * (arbitrarily close to command submission time).
446 * req->deadline - req->timeout is the command submission time
450 req->deadline - req->timeout >= ctrl->ka_last_check_time)
451 ctrl->comp_seen = true;
453 switch (nvme_decide_disposition(req)) {
461 nvme_failover_req(req);
464 #ifdef CONFIG_NVME_HOST_AUTH
465 queue_work(nvme_wq, &ctrl->dhchap_auth_work);
473 EXPORT_SYMBOL_GPL(nvme_complete_rq);
475 void nvme_complete_batch_req(struct request *req)
477 trace_nvme_complete_rq(req);
478 nvme_cleanup_cmd(req);
479 nvme_end_req_zoned(req);
481 EXPORT_SYMBOL_GPL(nvme_complete_batch_req);
484 * Called to unwind from ->queue_rq on a failed command submission so that the
485 * multipathing code gets called to potentially failover to another path.
486 * The caller needs to unwind all transport specific resource allocations and
487 * must return propagate the return value.
489 blk_status_t nvme_host_path_error(struct request *req)
491 nvme_req(req)->status = NVME_SC_HOST_PATH_ERROR;
492 blk_mq_set_request_complete(req);
493 nvme_complete_rq(req);
496 EXPORT_SYMBOL_GPL(nvme_host_path_error);
498 bool nvme_cancel_request(struct request *req, void *data)
500 dev_dbg_ratelimited(((struct nvme_ctrl *) data)->device,
501 "Cancelling I/O %d", req->tag);
503 /* don't abort one completed or idle request */
504 if (blk_mq_rq_state(req) != MQ_RQ_IN_FLIGHT)
507 nvme_req(req)->status = NVME_SC_HOST_ABORTED_CMD;
508 nvme_req(req)->flags |= NVME_REQ_CANCELLED;
509 blk_mq_complete_request(req);
512 EXPORT_SYMBOL_GPL(nvme_cancel_request);
514 void nvme_cancel_tagset(struct nvme_ctrl *ctrl)
517 blk_mq_tagset_busy_iter(ctrl->tagset,
518 nvme_cancel_request, ctrl);
519 blk_mq_tagset_wait_completed_request(ctrl->tagset);
522 EXPORT_SYMBOL_GPL(nvme_cancel_tagset);
524 void nvme_cancel_admin_tagset(struct nvme_ctrl *ctrl)
526 if (ctrl->admin_tagset) {
527 blk_mq_tagset_busy_iter(ctrl->admin_tagset,
528 nvme_cancel_request, ctrl);
529 blk_mq_tagset_wait_completed_request(ctrl->admin_tagset);
532 EXPORT_SYMBOL_GPL(nvme_cancel_admin_tagset);
534 bool nvme_change_ctrl_state(struct nvme_ctrl *ctrl,
535 enum nvme_ctrl_state new_state)
537 enum nvme_ctrl_state old_state;
539 bool changed = false;
541 spin_lock_irqsave(&ctrl->lock, flags);
543 old_state = nvme_ctrl_state(ctrl);
548 case NVME_CTRL_RESETTING:
549 case NVME_CTRL_CONNECTING:
556 case NVME_CTRL_RESETTING:
566 case NVME_CTRL_CONNECTING:
569 case NVME_CTRL_RESETTING:
576 case NVME_CTRL_DELETING:
579 case NVME_CTRL_RESETTING:
580 case NVME_CTRL_CONNECTING:
587 case NVME_CTRL_DELETING_NOIO:
589 case NVME_CTRL_DELETING:
599 case NVME_CTRL_DELETING:
611 WRITE_ONCE(ctrl->state, new_state);
612 wake_up_all(&ctrl->state_wq);
615 spin_unlock_irqrestore(&ctrl->lock, flags);
619 if (new_state == NVME_CTRL_LIVE) {
620 if (old_state == NVME_CTRL_CONNECTING)
621 nvme_stop_failfast_work(ctrl);
622 nvme_kick_requeue_lists(ctrl);
623 } else if (new_state == NVME_CTRL_CONNECTING &&
624 old_state == NVME_CTRL_RESETTING) {
625 nvme_start_failfast_work(ctrl);
629 EXPORT_SYMBOL_GPL(nvme_change_ctrl_state);
632 * Returns true for sink states that can't ever transition back to live.
634 static bool nvme_state_terminal(struct nvme_ctrl *ctrl)
636 switch (nvme_ctrl_state(ctrl)) {
639 case NVME_CTRL_RESETTING:
640 case NVME_CTRL_CONNECTING:
642 case NVME_CTRL_DELETING:
643 case NVME_CTRL_DELETING_NOIO:
647 WARN_ONCE(1, "Unhandled ctrl state:%d", ctrl->state);
653 * Waits for the controller state to be resetting, or returns false if it is
654 * not possible to ever transition to that state.
656 bool nvme_wait_reset(struct nvme_ctrl *ctrl)
658 wait_event(ctrl->state_wq,
659 nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING) ||
660 nvme_state_terminal(ctrl));
661 return nvme_ctrl_state(ctrl) == NVME_CTRL_RESETTING;
663 EXPORT_SYMBOL_GPL(nvme_wait_reset);
665 static void nvme_free_ns_head(struct kref *ref)
667 struct nvme_ns_head *head =
668 container_of(ref, struct nvme_ns_head, ref);
670 nvme_mpath_remove_disk(head);
671 ida_free(&head->subsys->ns_ida, head->instance);
672 cleanup_srcu_struct(&head->srcu);
673 nvme_put_subsystem(head->subsys);
677 bool nvme_tryget_ns_head(struct nvme_ns_head *head)
679 return kref_get_unless_zero(&head->ref);
682 void nvme_put_ns_head(struct nvme_ns_head *head)
684 kref_put(&head->ref, nvme_free_ns_head);
687 static void nvme_free_ns(struct kref *kref)
689 struct nvme_ns *ns = container_of(kref, struct nvme_ns, kref);
692 nvme_put_ns_head(ns->head);
693 nvme_put_ctrl(ns->ctrl);
697 static inline bool nvme_get_ns(struct nvme_ns *ns)
699 return kref_get_unless_zero(&ns->kref);
702 void nvme_put_ns(struct nvme_ns *ns)
704 kref_put(&ns->kref, nvme_free_ns);
706 EXPORT_SYMBOL_NS_GPL(nvme_put_ns, NVME_TARGET_PASSTHRU);
708 static inline void nvme_clear_nvme_request(struct request *req)
710 nvme_req(req)->status = 0;
711 nvme_req(req)->retries = 0;
712 nvme_req(req)->flags = 0;
713 req->rq_flags |= RQF_DONTPREP;
716 /* initialize a passthrough request */
717 void nvme_init_request(struct request *req, struct nvme_command *cmd)
719 struct nvme_request *nr = nvme_req(req);
720 bool logging_enabled;
722 if (req->q->queuedata) {
723 struct nvme_ns *ns = req->q->disk->private_data;
725 logging_enabled = ns->head->passthru_err_log_enabled;
726 req->timeout = NVME_IO_TIMEOUT;
727 } else { /* no queuedata implies admin queue */
728 logging_enabled = nr->ctrl->passthru_err_log_enabled;
729 req->timeout = NVME_ADMIN_TIMEOUT;
732 if (!logging_enabled)
733 req->rq_flags |= RQF_QUIET;
735 /* passthru commands should let the driver set the SGL flags */
736 cmd->common.flags &= ~NVME_CMD_SGL_ALL;
738 req->cmd_flags |= REQ_FAILFAST_DRIVER;
739 if (req->mq_hctx->type == HCTX_TYPE_POLL)
740 req->cmd_flags |= REQ_POLLED;
741 nvme_clear_nvme_request(req);
742 memcpy(nr->cmd, cmd, sizeof(*cmd));
744 EXPORT_SYMBOL_GPL(nvme_init_request);
747 * For something we're not in a state to send to the device the default action
748 * is to busy it and retry it after the controller state is recovered. However,
749 * if the controller is deleting or if anything is marked for failfast or
750 * nvme multipath it is immediately failed.
752 * Note: commands used to initialize the controller will be marked for failfast.
753 * Note: nvme cli/ioctl commands are marked for failfast.
755 blk_status_t nvme_fail_nonready_command(struct nvme_ctrl *ctrl,
758 enum nvme_ctrl_state state = nvme_ctrl_state(ctrl);
760 if (state != NVME_CTRL_DELETING_NOIO &&
761 state != NVME_CTRL_DELETING &&
762 state != NVME_CTRL_DEAD &&
763 !test_bit(NVME_CTRL_FAILFAST_EXPIRED, &ctrl->flags) &&
764 !blk_noretry_request(rq) && !(rq->cmd_flags & REQ_NVME_MPATH))
765 return BLK_STS_RESOURCE;
766 return nvme_host_path_error(rq);
768 EXPORT_SYMBOL_GPL(nvme_fail_nonready_command);
770 bool __nvme_check_ready(struct nvme_ctrl *ctrl, struct request *rq,
771 bool queue_live, enum nvme_ctrl_state state)
773 struct nvme_request *req = nvme_req(rq);
776 * currently we have a problem sending passthru commands
777 * on the admin_q if the controller is not LIVE because we can't
778 * make sure that they are going out after the admin connect,
779 * controller enable and/or other commands in the initialization
780 * sequence. until the controller will be LIVE, fail with
781 * BLK_STS_RESOURCE so that they will be rescheduled.
783 if (rq->q == ctrl->admin_q && (req->flags & NVME_REQ_USERCMD))
786 if (ctrl->ops->flags & NVME_F_FABRICS) {
788 * Only allow commands on a live queue, except for the connect
789 * command, which is require to set the queue live in the
790 * appropinquate states.
793 case NVME_CTRL_CONNECTING:
794 if (blk_rq_is_passthrough(rq) && nvme_is_fabrics(req->cmd) &&
795 (req->cmd->fabrics.fctype == nvme_fabrics_type_connect ||
796 req->cmd->fabrics.fctype == nvme_fabrics_type_auth_send ||
797 req->cmd->fabrics.fctype == nvme_fabrics_type_auth_receive))
809 EXPORT_SYMBOL_GPL(__nvme_check_ready);
811 static inline void nvme_setup_flush(struct nvme_ns *ns,
812 struct nvme_command *cmnd)
814 memset(cmnd, 0, sizeof(*cmnd));
815 cmnd->common.opcode = nvme_cmd_flush;
816 cmnd->common.nsid = cpu_to_le32(ns->head->ns_id);
819 static blk_status_t nvme_setup_discard(struct nvme_ns *ns, struct request *req,
820 struct nvme_command *cmnd)
822 unsigned short segments = blk_rq_nr_discard_segments(req), n = 0;
823 struct nvme_dsm_range *range;
827 * Some devices do not consider the DSM 'Number of Ranges' field when
828 * determining how much data to DMA. Always allocate memory for maximum
829 * number of segments to prevent device reading beyond end of buffer.
831 static const size_t alloc_size = sizeof(*range) * NVME_DSM_MAX_RANGES;
833 range = kzalloc(alloc_size, GFP_ATOMIC | __GFP_NOWARN);
836 * If we fail allocation our range, fallback to the controller
837 * discard page. If that's also busy, it's safe to return
838 * busy, as we know we can make progress once that's freed.
840 if (test_and_set_bit_lock(0, &ns->ctrl->discard_page_busy))
841 return BLK_STS_RESOURCE;
843 range = page_address(ns->ctrl->discard_page);
846 if (queue_max_discard_segments(req->q) == 1) {
847 u64 slba = nvme_sect_to_lba(ns->head, blk_rq_pos(req));
848 u32 nlb = blk_rq_sectors(req) >> (ns->head->lba_shift - 9);
850 range[0].cattr = cpu_to_le32(0);
851 range[0].nlb = cpu_to_le32(nlb);
852 range[0].slba = cpu_to_le64(slba);
855 __rq_for_each_bio(bio, req) {
856 u64 slba = nvme_sect_to_lba(ns->head,
857 bio->bi_iter.bi_sector);
858 u32 nlb = bio->bi_iter.bi_size >> ns->head->lba_shift;
861 range[n].cattr = cpu_to_le32(0);
862 range[n].nlb = cpu_to_le32(nlb);
863 range[n].slba = cpu_to_le64(slba);
869 if (WARN_ON_ONCE(n != segments)) {
870 if (virt_to_page(range) == ns->ctrl->discard_page)
871 clear_bit_unlock(0, &ns->ctrl->discard_page_busy);
874 return BLK_STS_IOERR;
877 memset(cmnd, 0, sizeof(*cmnd));
878 cmnd->dsm.opcode = nvme_cmd_dsm;
879 cmnd->dsm.nsid = cpu_to_le32(ns->head->ns_id);
880 cmnd->dsm.nr = cpu_to_le32(segments - 1);
881 cmnd->dsm.attributes = cpu_to_le32(NVME_DSMGMT_AD);
883 bvec_set_virt(&req->special_vec, range, alloc_size);
884 req->rq_flags |= RQF_SPECIAL_PAYLOAD;
889 static void nvme_set_ref_tag(struct nvme_ns *ns, struct nvme_command *cmnd,
895 /* both rw and write zeroes share the same reftag format */
896 switch (ns->head->guard_type) {
897 case NVME_NVM_NS_16B_GUARD:
898 cmnd->rw.reftag = cpu_to_le32(t10_pi_ref_tag(req));
900 case NVME_NVM_NS_64B_GUARD:
901 ref48 = ext_pi_ref_tag(req);
902 lower = lower_32_bits(ref48);
903 upper = upper_32_bits(ref48);
905 cmnd->rw.reftag = cpu_to_le32(lower);
906 cmnd->rw.cdw3 = cpu_to_le32(upper);
913 static inline blk_status_t nvme_setup_write_zeroes(struct nvme_ns *ns,
914 struct request *req, struct nvme_command *cmnd)
916 memset(cmnd, 0, sizeof(*cmnd));
918 if (ns->ctrl->quirks & NVME_QUIRK_DEALLOCATE_ZEROES)
919 return nvme_setup_discard(ns, req, cmnd);
921 cmnd->write_zeroes.opcode = nvme_cmd_write_zeroes;
922 cmnd->write_zeroes.nsid = cpu_to_le32(ns->head->ns_id);
923 cmnd->write_zeroes.slba =
924 cpu_to_le64(nvme_sect_to_lba(ns->head, blk_rq_pos(req)));
925 cmnd->write_zeroes.length =
926 cpu_to_le16((blk_rq_bytes(req) >> ns->head->lba_shift) - 1);
928 if (!(req->cmd_flags & REQ_NOUNMAP) &&
929 (ns->head->features & NVME_NS_DEAC))
930 cmnd->write_zeroes.control |= cpu_to_le16(NVME_WZ_DEAC);
932 if (nvme_ns_has_pi(ns->head)) {
933 cmnd->write_zeroes.control |= cpu_to_le16(NVME_RW_PRINFO_PRACT);
935 switch (ns->head->pi_type) {
936 case NVME_NS_DPS_PI_TYPE1:
937 case NVME_NS_DPS_PI_TYPE2:
938 nvme_set_ref_tag(ns, cmnd, req);
946 static inline blk_status_t nvme_setup_rw(struct nvme_ns *ns,
947 struct request *req, struct nvme_command *cmnd,
953 if (req->cmd_flags & REQ_FUA)
954 control |= NVME_RW_FUA;
955 if (req->cmd_flags & (REQ_FAILFAST_DEV | REQ_RAHEAD))
956 control |= NVME_RW_LR;
958 if (req->cmd_flags & REQ_RAHEAD)
959 dsmgmt |= NVME_RW_DSM_FREQ_PREFETCH;
961 cmnd->rw.opcode = op;
963 cmnd->rw.nsid = cpu_to_le32(ns->head->ns_id);
966 cmnd->rw.metadata = 0;
968 cpu_to_le64(nvme_sect_to_lba(ns->head, blk_rq_pos(req)));
970 cpu_to_le16((blk_rq_bytes(req) >> ns->head->lba_shift) - 1);
973 cmnd->rw.appmask = 0;
977 * If formated with metadata, the block layer always provides a
978 * metadata buffer if CONFIG_BLK_DEV_INTEGRITY is enabled. Else
979 * we enable the PRACT bit for protection information or set the
980 * namespace capacity to zero to prevent any I/O.
982 if (!blk_integrity_rq(req)) {
983 if (WARN_ON_ONCE(!nvme_ns_has_pi(ns->head)))
984 return BLK_STS_NOTSUPP;
985 control |= NVME_RW_PRINFO_PRACT;
988 switch (ns->head->pi_type) {
989 case NVME_NS_DPS_PI_TYPE3:
990 control |= NVME_RW_PRINFO_PRCHK_GUARD;
992 case NVME_NS_DPS_PI_TYPE1:
993 case NVME_NS_DPS_PI_TYPE2:
994 control |= NVME_RW_PRINFO_PRCHK_GUARD |
995 NVME_RW_PRINFO_PRCHK_REF;
996 if (op == nvme_cmd_zone_append)
997 control |= NVME_RW_APPEND_PIREMAP;
998 nvme_set_ref_tag(ns, cmnd, req);
1003 cmnd->rw.control = cpu_to_le16(control);
1004 cmnd->rw.dsmgmt = cpu_to_le32(dsmgmt);
1008 void nvme_cleanup_cmd(struct request *req)
1010 if (req->rq_flags & RQF_SPECIAL_PAYLOAD) {
1011 struct nvme_ctrl *ctrl = nvme_req(req)->ctrl;
1013 if (req->special_vec.bv_page == ctrl->discard_page)
1014 clear_bit_unlock(0, &ctrl->discard_page_busy);
1016 kfree(bvec_virt(&req->special_vec));
1019 EXPORT_SYMBOL_GPL(nvme_cleanup_cmd);
1021 blk_status_t nvme_setup_cmd(struct nvme_ns *ns, struct request *req)
1023 struct nvme_command *cmd = nvme_req(req)->cmd;
1024 blk_status_t ret = BLK_STS_OK;
1026 if (!(req->rq_flags & RQF_DONTPREP))
1027 nvme_clear_nvme_request(req);
1029 switch (req_op(req)) {
1031 case REQ_OP_DRV_OUT:
1032 /* these are setup prior to execution in nvme_init_request() */
1035 nvme_setup_flush(ns, cmd);
1037 case REQ_OP_ZONE_RESET_ALL:
1038 case REQ_OP_ZONE_RESET:
1039 ret = nvme_setup_zone_mgmt_send(ns, req, cmd, NVME_ZONE_RESET);
1041 case REQ_OP_ZONE_OPEN:
1042 ret = nvme_setup_zone_mgmt_send(ns, req, cmd, NVME_ZONE_OPEN);
1044 case REQ_OP_ZONE_CLOSE:
1045 ret = nvme_setup_zone_mgmt_send(ns, req, cmd, NVME_ZONE_CLOSE);
1047 case REQ_OP_ZONE_FINISH:
1048 ret = nvme_setup_zone_mgmt_send(ns, req, cmd, NVME_ZONE_FINISH);
1050 case REQ_OP_WRITE_ZEROES:
1051 ret = nvme_setup_write_zeroes(ns, req, cmd);
1053 case REQ_OP_DISCARD:
1054 ret = nvme_setup_discard(ns, req, cmd);
1057 ret = nvme_setup_rw(ns, req, cmd, nvme_cmd_read);
1060 ret = nvme_setup_rw(ns, req, cmd, nvme_cmd_write);
1062 case REQ_OP_ZONE_APPEND:
1063 ret = nvme_setup_rw(ns, req, cmd, nvme_cmd_zone_append);
1067 return BLK_STS_IOERR;
1070 cmd->common.command_id = nvme_cid(req);
1071 trace_nvme_setup_cmd(req, cmd);
1074 EXPORT_SYMBOL_GPL(nvme_setup_cmd);
1079 * >0: nvme controller's cqe status response
1080 * <0: kernel error in lieu of controller response
1082 int nvme_execute_rq(struct request *rq, bool at_head)
1084 blk_status_t status;
1086 status = blk_execute_rq(rq, at_head);
1087 if (nvme_req(rq)->flags & NVME_REQ_CANCELLED)
1089 if (nvme_req(rq)->status)
1090 return nvme_req(rq)->status;
1091 return blk_status_to_errno(status);
1093 EXPORT_SYMBOL_NS_GPL(nvme_execute_rq, NVME_TARGET_PASSTHRU);
1096 * Returns 0 on success. If the result is negative, it's a Linux error code;
1097 * if the result is positive, it's an NVM Express status code
1099 int __nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
1100 union nvme_result *result, void *buffer, unsigned bufflen,
1101 int qid, nvme_submit_flags_t flags)
1103 struct request *req;
1105 blk_mq_req_flags_t blk_flags = 0;
1107 if (flags & NVME_SUBMIT_NOWAIT)
1108 blk_flags |= BLK_MQ_REQ_NOWAIT;
1109 if (flags & NVME_SUBMIT_RESERVED)
1110 blk_flags |= BLK_MQ_REQ_RESERVED;
1111 if (qid == NVME_QID_ANY)
1112 req = blk_mq_alloc_request(q, nvme_req_op(cmd), blk_flags);
1114 req = blk_mq_alloc_request_hctx(q, nvme_req_op(cmd), blk_flags,
1118 return PTR_ERR(req);
1119 nvme_init_request(req, cmd);
1120 if (flags & NVME_SUBMIT_RETRY)
1121 req->cmd_flags &= ~REQ_FAILFAST_DRIVER;
1123 if (buffer && bufflen) {
1124 ret = blk_rq_map_kern(q, req, buffer, bufflen, GFP_KERNEL);
1129 ret = nvme_execute_rq(req, flags & NVME_SUBMIT_AT_HEAD);
1130 if (result && ret >= 0)
1131 *result = nvme_req(req)->result;
1133 blk_mq_free_request(req);
1136 EXPORT_SYMBOL_GPL(__nvme_submit_sync_cmd);
1138 int nvme_submit_sync_cmd(struct request_queue *q, struct nvme_command *cmd,
1139 void *buffer, unsigned bufflen)
1141 return __nvme_submit_sync_cmd(q, cmd, NULL, buffer, bufflen,
1144 EXPORT_SYMBOL_GPL(nvme_submit_sync_cmd);
1146 u32 nvme_command_effects(struct nvme_ctrl *ctrl, struct nvme_ns *ns, u8 opcode)
1151 effects = le32_to_cpu(ns->head->effects->iocs[opcode]);
1152 if (effects & ~(NVME_CMD_EFFECTS_CSUPP | NVME_CMD_EFFECTS_LBCC))
1153 dev_warn_once(ctrl->device,
1154 "IO command:%02x has unusual effects:%08x\n",
1158 * NVME_CMD_EFFECTS_CSE_MASK causes a freeze all I/O queues,
1159 * which would deadlock when done on an I/O command. Note that
1160 * We already warn about an unusual effect above.
1162 effects &= ~NVME_CMD_EFFECTS_CSE_MASK;
1164 effects = le32_to_cpu(ctrl->effects->acs[opcode]);
1166 /* Ignore execution restrictions if any relaxation bits are set */
1167 if (effects & NVME_CMD_EFFECTS_CSER_MASK)
1168 effects &= ~NVME_CMD_EFFECTS_CSE_MASK;
1173 EXPORT_SYMBOL_NS_GPL(nvme_command_effects, NVME_TARGET_PASSTHRU);
1175 u32 nvme_passthru_start(struct nvme_ctrl *ctrl, struct nvme_ns *ns, u8 opcode)
1177 u32 effects = nvme_command_effects(ctrl, ns, opcode);
1180 * For simplicity, IO to all namespaces is quiesced even if the command
1181 * effects say only one namespace is affected.
1183 if (effects & NVME_CMD_EFFECTS_CSE_MASK) {
1184 mutex_lock(&ctrl->scan_lock);
1185 mutex_lock(&ctrl->subsys->lock);
1186 nvme_mpath_start_freeze(ctrl->subsys);
1187 nvme_mpath_wait_freeze(ctrl->subsys);
1188 nvme_start_freeze(ctrl);
1189 nvme_wait_freeze(ctrl);
1193 EXPORT_SYMBOL_NS_GPL(nvme_passthru_start, NVME_TARGET_PASSTHRU);
1195 void nvme_passthru_end(struct nvme_ctrl *ctrl, struct nvme_ns *ns, u32 effects,
1196 struct nvme_command *cmd, int status)
1198 if (effects & NVME_CMD_EFFECTS_CSE_MASK) {
1199 nvme_unfreeze(ctrl);
1200 nvme_mpath_unfreeze(ctrl->subsys);
1201 mutex_unlock(&ctrl->subsys->lock);
1202 mutex_unlock(&ctrl->scan_lock);
1204 if (effects & NVME_CMD_EFFECTS_CCC) {
1205 if (!test_and_set_bit(NVME_CTRL_DIRTY_CAPABILITY,
1207 dev_info(ctrl->device,
1208 "controller capabilities changed, reset may be required to take effect.\n");
1211 if (effects & (NVME_CMD_EFFECTS_NIC | NVME_CMD_EFFECTS_NCC)) {
1212 nvme_queue_scan(ctrl);
1213 flush_work(&ctrl->scan_work);
1218 switch (cmd->common.opcode) {
1219 case nvme_admin_set_features:
1220 switch (le32_to_cpu(cmd->common.cdw10) & 0xFF) {
1221 case NVME_FEAT_KATO:
1223 * Keep alive commands interval on the host should be
1224 * updated when KATO is modified by Set Features
1228 nvme_update_keep_alive(ctrl, cmd);
1238 EXPORT_SYMBOL_NS_GPL(nvme_passthru_end, NVME_TARGET_PASSTHRU);
1241 * Recommended frequency for KATO commands per NVMe 1.4 section 7.12.1:
1243 * The host should send Keep Alive commands at half of the Keep Alive Timeout
1244 * accounting for transport roundtrip times [..].
1246 static unsigned long nvme_keep_alive_work_period(struct nvme_ctrl *ctrl)
1248 unsigned long delay = ctrl->kato * HZ / 2;
1251 * When using Traffic Based Keep Alive, we need to run
1252 * nvme_keep_alive_work at twice the normal frequency, as one
1253 * command completion can postpone sending a keep alive command
1254 * by up to twice the delay between runs.
1256 if (ctrl->ctratt & NVME_CTRL_ATTR_TBKAS)
1261 static void nvme_queue_keep_alive_work(struct nvme_ctrl *ctrl)
1263 unsigned long now = jiffies;
1264 unsigned long delay = nvme_keep_alive_work_period(ctrl);
1265 unsigned long ka_next_check_tm = ctrl->ka_last_check_time + delay;
1267 if (time_after(now, ka_next_check_tm))
1270 delay = ka_next_check_tm - now;
1272 queue_delayed_work(nvme_wq, &ctrl->ka_work, delay);
1275 static enum rq_end_io_ret nvme_keep_alive_end_io(struct request *rq,
1276 blk_status_t status)
1278 struct nvme_ctrl *ctrl = rq->end_io_data;
1279 unsigned long flags;
1280 bool startka = false;
1281 unsigned long rtt = jiffies - (rq->deadline - rq->timeout);
1282 unsigned long delay = nvme_keep_alive_work_period(ctrl);
1285 * Subtract off the keepalive RTT so nvme_keep_alive_work runs
1286 * at the desired frequency.
1291 dev_warn(ctrl->device, "long keepalive RTT (%u ms)\n",
1292 jiffies_to_msecs(rtt));
1296 blk_mq_free_request(rq);
1299 dev_err(ctrl->device,
1300 "failed nvme_keep_alive_end_io error=%d\n",
1302 return RQ_END_IO_NONE;
1305 ctrl->ka_last_check_time = jiffies;
1306 ctrl->comp_seen = false;
1307 spin_lock_irqsave(&ctrl->lock, flags);
1308 if (ctrl->state == NVME_CTRL_LIVE ||
1309 ctrl->state == NVME_CTRL_CONNECTING)
1311 spin_unlock_irqrestore(&ctrl->lock, flags);
1313 queue_delayed_work(nvme_wq, &ctrl->ka_work, delay);
1314 return RQ_END_IO_NONE;
1317 static void nvme_keep_alive_work(struct work_struct *work)
1319 struct nvme_ctrl *ctrl = container_of(to_delayed_work(work),
1320 struct nvme_ctrl, ka_work);
1321 bool comp_seen = ctrl->comp_seen;
1324 ctrl->ka_last_check_time = jiffies;
1326 if ((ctrl->ctratt & NVME_CTRL_ATTR_TBKAS) && comp_seen) {
1327 dev_dbg(ctrl->device,
1328 "reschedule traffic based keep-alive timer\n");
1329 ctrl->comp_seen = false;
1330 nvme_queue_keep_alive_work(ctrl);
1334 rq = blk_mq_alloc_request(ctrl->admin_q, nvme_req_op(&ctrl->ka_cmd),
1335 BLK_MQ_REQ_RESERVED | BLK_MQ_REQ_NOWAIT);
1337 /* allocation failure, reset the controller */
1338 dev_err(ctrl->device, "keep-alive failed: %ld\n", PTR_ERR(rq));
1339 nvme_reset_ctrl(ctrl);
1342 nvme_init_request(rq, &ctrl->ka_cmd);
1344 rq->timeout = ctrl->kato * HZ;
1345 rq->end_io = nvme_keep_alive_end_io;
1346 rq->end_io_data = ctrl;
1347 blk_execute_rq_nowait(rq, false);
1350 static void nvme_start_keep_alive(struct nvme_ctrl *ctrl)
1352 if (unlikely(ctrl->kato == 0))
1355 nvme_queue_keep_alive_work(ctrl);
1358 void nvme_stop_keep_alive(struct nvme_ctrl *ctrl)
1360 if (unlikely(ctrl->kato == 0))
1363 cancel_delayed_work_sync(&ctrl->ka_work);
1365 EXPORT_SYMBOL_GPL(nvme_stop_keep_alive);
1367 static void nvme_update_keep_alive(struct nvme_ctrl *ctrl,
1368 struct nvme_command *cmd)
1370 unsigned int new_kato =
1371 DIV_ROUND_UP(le32_to_cpu(cmd->common.cdw11), 1000);
1373 dev_info(ctrl->device,
1374 "keep alive interval updated from %u ms to %u ms\n",
1375 ctrl->kato * 1000 / 2, new_kato * 1000 / 2);
1377 nvme_stop_keep_alive(ctrl);
1378 ctrl->kato = new_kato;
1379 nvme_start_keep_alive(ctrl);
1383 * In NVMe 1.0 the CNS field was just a binary controller or namespace
1384 * flag, thus sending any new CNS opcodes has a big chance of not working.
1385 * Qemu unfortunately had that bug after reporting a 1.1 version compliance
1386 * (but not for any later version).
1388 static bool nvme_ctrl_limited_cns(struct nvme_ctrl *ctrl)
1390 if (ctrl->quirks & NVME_QUIRK_IDENTIFY_CNS)
1391 return ctrl->vs < NVME_VS(1, 2, 0);
1392 return ctrl->vs < NVME_VS(1, 1, 0);
1395 static int nvme_identify_ctrl(struct nvme_ctrl *dev, struct nvme_id_ctrl **id)
1397 struct nvme_command c = { };
1400 /* gcc-4.4.4 (at least) has issues with initializers and anon unions */
1401 c.identify.opcode = nvme_admin_identify;
1402 c.identify.cns = NVME_ID_CNS_CTRL;
1404 *id = kmalloc(sizeof(struct nvme_id_ctrl), GFP_KERNEL);
1408 error = nvme_submit_sync_cmd(dev->admin_q, &c, *id,
1409 sizeof(struct nvme_id_ctrl));
1417 static int nvme_process_ns_desc(struct nvme_ctrl *ctrl, struct nvme_ns_ids *ids,
1418 struct nvme_ns_id_desc *cur, bool *csi_seen)
1420 const char *warn_str = "ctrl returned bogus length:";
1423 switch (cur->nidt) {
1424 case NVME_NIDT_EUI64:
1425 if (cur->nidl != NVME_NIDT_EUI64_LEN) {
1426 dev_warn(ctrl->device, "%s %d for NVME_NIDT_EUI64\n",
1427 warn_str, cur->nidl);
1430 if (ctrl->quirks & NVME_QUIRK_BOGUS_NID)
1431 return NVME_NIDT_EUI64_LEN;
1432 memcpy(ids->eui64, data + sizeof(*cur), NVME_NIDT_EUI64_LEN);
1433 return NVME_NIDT_EUI64_LEN;
1434 case NVME_NIDT_NGUID:
1435 if (cur->nidl != NVME_NIDT_NGUID_LEN) {
1436 dev_warn(ctrl->device, "%s %d for NVME_NIDT_NGUID\n",
1437 warn_str, cur->nidl);
1440 if (ctrl->quirks & NVME_QUIRK_BOGUS_NID)
1441 return NVME_NIDT_NGUID_LEN;
1442 memcpy(ids->nguid, data + sizeof(*cur), NVME_NIDT_NGUID_LEN);
1443 return NVME_NIDT_NGUID_LEN;
1444 case NVME_NIDT_UUID:
1445 if (cur->nidl != NVME_NIDT_UUID_LEN) {
1446 dev_warn(ctrl->device, "%s %d for NVME_NIDT_UUID\n",
1447 warn_str, cur->nidl);
1450 if (ctrl->quirks & NVME_QUIRK_BOGUS_NID)
1451 return NVME_NIDT_UUID_LEN;
1452 uuid_copy(&ids->uuid, data + sizeof(*cur));
1453 return NVME_NIDT_UUID_LEN;
1455 if (cur->nidl != NVME_NIDT_CSI_LEN) {
1456 dev_warn(ctrl->device, "%s %d for NVME_NIDT_CSI\n",
1457 warn_str, cur->nidl);
1460 memcpy(&ids->csi, data + sizeof(*cur), NVME_NIDT_CSI_LEN);
1462 return NVME_NIDT_CSI_LEN;
1464 /* Skip unknown types */
1469 static int nvme_identify_ns_descs(struct nvme_ctrl *ctrl,
1470 struct nvme_ns_info *info)
1472 struct nvme_command c = { };
1473 bool csi_seen = false;
1474 int status, pos, len;
1477 if (ctrl->vs < NVME_VS(1, 3, 0) && !nvme_multi_css(ctrl))
1479 if (ctrl->quirks & NVME_QUIRK_NO_NS_DESC_LIST)
1482 c.identify.opcode = nvme_admin_identify;
1483 c.identify.nsid = cpu_to_le32(info->nsid);
1484 c.identify.cns = NVME_ID_CNS_NS_DESC_LIST;
1486 data = kzalloc(NVME_IDENTIFY_DATA_SIZE, GFP_KERNEL);
1490 status = nvme_submit_sync_cmd(ctrl->admin_q, &c, data,
1491 NVME_IDENTIFY_DATA_SIZE);
1493 dev_warn(ctrl->device,
1494 "Identify Descriptors failed (nsid=%u, status=0x%x)\n",
1495 info->nsid, status);
1499 for (pos = 0; pos < NVME_IDENTIFY_DATA_SIZE; pos += len) {
1500 struct nvme_ns_id_desc *cur = data + pos;
1505 len = nvme_process_ns_desc(ctrl, &info->ids, cur, &csi_seen);
1509 len += sizeof(*cur);
1512 if (nvme_multi_css(ctrl) && !csi_seen) {
1513 dev_warn(ctrl->device, "Command set not reported for nsid:%d\n",
1523 int nvme_identify_ns(struct nvme_ctrl *ctrl, unsigned nsid,
1524 struct nvme_id_ns **id)
1526 struct nvme_command c = { };
1529 /* gcc-4.4.4 (at least) has issues with initializers and anon unions */
1530 c.identify.opcode = nvme_admin_identify;
1531 c.identify.nsid = cpu_to_le32(nsid);
1532 c.identify.cns = NVME_ID_CNS_NS;
1534 *id = kmalloc(sizeof(**id), GFP_KERNEL);
1538 error = nvme_submit_sync_cmd(ctrl->admin_q, &c, *id, sizeof(**id));
1540 dev_warn(ctrl->device, "Identify namespace failed (%d)\n", error);
1547 static int nvme_ns_info_from_identify(struct nvme_ctrl *ctrl,
1548 struct nvme_ns_info *info)
1550 struct nvme_ns_ids *ids = &info->ids;
1551 struct nvme_id_ns *id;
1554 ret = nvme_identify_ns(ctrl, info->nsid, &id);
1558 if (id->ncap == 0) {
1559 /* namespace not allocated or attached */
1560 info->is_removed = true;
1565 info->anagrpid = id->anagrpid;
1566 info->is_shared = id->nmic & NVME_NS_NMIC_SHARED;
1567 info->is_readonly = id->nsattr & NVME_NS_ATTR_RO;
1568 info->is_ready = true;
1569 if (ctrl->quirks & NVME_QUIRK_BOGUS_NID) {
1570 dev_info(ctrl->device,
1571 "Ignoring bogus Namespace Identifiers\n");
1573 if (ctrl->vs >= NVME_VS(1, 1, 0) &&
1574 !memchr_inv(ids->eui64, 0, sizeof(ids->eui64)))
1575 memcpy(ids->eui64, id->eui64, sizeof(ids->eui64));
1576 if (ctrl->vs >= NVME_VS(1, 2, 0) &&
1577 !memchr_inv(ids->nguid, 0, sizeof(ids->nguid)))
1578 memcpy(ids->nguid, id->nguid, sizeof(ids->nguid));
1586 static int nvme_ns_info_from_id_cs_indep(struct nvme_ctrl *ctrl,
1587 struct nvme_ns_info *info)
1589 struct nvme_id_ns_cs_indep *id;
1590 struct nvme_command c = {
1591 .identify.opcode = nvme_admin_identify,
1592 .identify.nsid = cpu_to_le32(info->nsid),
1593 .identify.cns = NVME_ID_CNS_NS_CS_INDEP,
1597 id = kmalloc(sizeof(*id), GFP_KERNEL);
1601 ret = nvme_submit_sync_cmd(ctrl->admin_q, &c, id, sizeof(*id));
1603 info->anagrpid = id->anagrpid;
1604 info->is_shared = id->nmic & NVME_NS_NMIC_SHARED;
1605 info->is_readonly = id->nsattr & NVME_NS_ATTR_RO;
1606 info->is_ready = id->nstat & NVME_NSTAT_NRDY;
1612 static int nvme_features(struct nvme_ctrl *dev, u8 op, unsigned int fid,
1613 unsigned int dword11, void *buffer, size_t buflen, u32 *result)
1615 union nvme_result res = { 0 };
1616 struct nvme_command c = { };
1619 c.features.opcode = op;
1620 c.features.fid = cpu_to_le32(fid);
1621 c.features.dword11 = cpu_to_le32(dword11);
1623 ret = __nvme_submit_sync_cmd(dev->admin_q, &c, &res,
1624 buffer, buflen, NVME_QID_ANY, 0);
1625 if (ret >= 0 && result)
1626 *result = le32_to_cpu(res.u32);
1630 int nvme_set_features(struct nvme_ctrl *dev, unsigned int fid,
1631 unsigned int dword11, void *buffer, size_t buflen,
1634 return nvme_features(dev, nvme_admin_set_features, fid, dword11, buffer,
1637 EXPORT_SYMBOL_GPL(nvme_set_features);
1639 int nvme_get_features(struct nvme_ctrl *dev, unsigned int fid,
1640 unsigned int dword11, void *buffer, size_t buflen,
1643 return nvme_features(dev, nvme_admin_get_features, fid, dword11, buffer,
1646 EXPORT_SYMBOL_GPL(nvme_get_features);
1648 int nvme_set_queue_count(struct nvme_ctrl *ctrl, int *count)
1650 u32 q_count = (*count - 1) | ((*count - 1) << 16);
1652 int status, nr_io_queues;
1654 status = nvme_set_features(ctrl, NVME_FEAT_NUM_QUEUES, q_count, NULL, 0,
1660 * Degraded controllers might return an error when setting the queue
1661 * count. We still want to be able to bring them online and offer
1662 * access to the admin queue, as that might be only way to fix them up.
1665 dev_err(ctrl->device, "Could not set queue count (%d)\n", status);
1668 nr_io_queues = min(result & 0xffff, result >> 16) + 1;
1669 *count = min(*count, nr_io_queues);
1674 EXPORT_SYMBOL_GPL(nvme_set_queue_count);
1676 #define NVME_AEN_SUPPORTED \
1677 (NVME_AEN_CFG_NS_ATTR | NVME_AEN_CFG_FW_ACT | \
1678 NVME_AEN_CFG_ANA_CHANGE | NVME_AEN_CFG_DISC_CHANGE)
1680 static void nvme_enable_aen(struct nvme_ctrl *ctrl)
1682 u32 result, supported_aens = ctrl->oaes & NVME_AEN_SUPPORTED;
1685 if (!supported_aens)
1688 status = nvme_set_features(ctrl, NVME_FEAT_ASYNC_EVENT, supported_aens,
1691 dev_warn(ctrl->device, "Failed to configure AEN (cfg %x)\n",
1694 queue_work(nvme_wq, &ctrl->async_event_work);
1697 static int nvme_ns_open(struct nvme_ns *ns)
1700 /* should never be called due to GENHD_FL_HIDDEN */
1701 if (WARN_ON_ONCE(nvme_ns_head_multipath(ns->head)))
1703 if (!nvme_get_ns(ns))
1705 if (!try_module_get(ns->ctrl->ops->module))
1716 static void nvme_ns_release(struct nvme_ns *ns)
1719 module_put(ns->ctrl->ops->module);
1723 static int nvme_open(struct gendisk *disk, blk_mode_t mode)
1725 return nvme_ns_open(disk->private_data);
1728 static void nvme_release(struct gendisk *disk)
1730 nvme_ns_release(disk->private_data);
1733 int nvme_getgeo(struct block_device *bdev, struct hd_geometry *geo)
1735 /* some standard values */
1736 geo->heads = 1 << 6;
1737 geo->sectors = 1 << 5;
1738 geo->cylinders = get_capacity(bdev->bd_disk) >> 11;
1742 static bool nvme_init_integrity(struct gendisk *disk, struct nvme_ns_head *head)
1744 struct blk_integrity integrity = { };
1746 blk_integrity_unregister(disk);
1752 * PI can always be supported as we can ask the controller to simply
1753 * insert/strip it, which is not possible for other kinds of metadata.
1755 if (!IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY) ||
1756 !(head->features & NVME_NS_METADATA_SUPPORTED))
1757 return nvme_ns_has_pi(head);
1759 switch (head->pi_type) {
1760 case NVME_NS_DPS_PI_TYPE3:
1761 switch (head->guard_type) {
1762 case NVME_NVM_NS_16B_GUARD:
1763 integrity.profile = &t10_pi_type3_crc;
1764 integrity.tag_size = sizeof(u16) + sizeof(u32);
1765 integrity.flags |= BLK_INTEGRITY_DEVICE_CAPABLE;
1767 case NVME_NVM_NS_64B_GUARD:
1768 integrity.profile = &ext_pi_type3_crc64;
1769 integrity.tag_size = sizeof(u16) + 6;
1770 integrity.flags |= BLK_INTEGRITY_DEVICE_CAPABLE;
1773 integrity.profile = NULL;
1777 case NVME_NS_DPS_PI_TYPE1:
1778 case NVME_NS_DPS_PI_TYPE2:
1779 switch (head->guard_type) {
1780 case NVME_NVM_NS_16B_GUARD:
1781 integrity.profile = &t10_pi_type1_crc;
1782 integrity.tag_size = sizeof(u16);
1783 integrity.flags |= BLK_INTEGRITY_DEVICE_CAPABLE;
1785 case NVME_NVM_NS_64B_GUARD:
1786 integrity.profile = &ext_pi_type1_crc64;
1787 integrity.tag_size = sizeof(u16);
1788 integrity.flags |= BLK_INTEGRITY_DEVICE_CAPABLE;
1791 integrity.profile = NULL;
1796 integrity.profile = NULL;
1800 integrity.tuple_size = head->ms;
1801 integrity.pi_offset = head->pi_offset;
1802 blk_integrity_register(disk, &integrity);
1806 static void nvme_config_discard(struct nvme_ns *ns, struct queue_limits *lim)
1808 struct nvme_ctrl *ctrl = ns->ctrl;
1810 if (ctrl->dmrsl && ctrl->dmrsl <= nvme_sect_to_lba(ns->head, UINT_MAX))
1811 lim->max_hw_discard_sectors =
1812 nvme_lba_to_sect(ns->head, ctrl->dmrsl);
1813 else if (ctrl->oncs & NVME_CTRL_ONCS_DSM)
1814 lim->max_hw_discard_sectors = UINT_MAX;
1816 lim->max_hw_discard_sectors = 0;
1818 lim->discard_granularity = lim->logical_block_size;
1821 lim->max_discard_segments = ctrl->dmrl;
1823 lim->max_discard_segments = NVME_DSM_MAX_RANGES;
1826 static bool nvme_ns_ids_equal(struct nvme_ns_ids *a, struct nvme_ns_ids *b)
1828 return uuid_equal(&a->uuid, &b->uuid) &&
1829 memcmp(&a->nguid, &b->nguid, sizeof(a->nguid)) == 0 &&
1830 memcmp(&a->eui64, &b->eui64, sizeof(a->eui64)) == 0 &&
1834 static int nvme_identify_ns_nvm(struct nvme_ctrl *ctrl, unsigned int nsid,
1835 struct nvme_id_ns_nvm **nvmp)
1837 struct nvme_command c = {
1838 .identify.opcode = nvme_admin_identify,
1839 .identify.nsid = cpu_to_le32(nsid),
1840 .identify.cns = NVME_ID_CNS_CS_NS,
1841 .identify.csi = NVME_CSI_NVM,
1843 struct nvme_id_ns_nvm *nvm;
1846 nvm = kzalloc(sizeof(*nvm), GFP_KERNEL);
1850 ret = nvme_submit_sync_cmd(ctrl->admin_q, &c, nvm, sizeof(*nvm));
1858 static void nvme_configure_pi_elbas(struct nvme_ns_head *head,
1859 struct nvme_id_ns *id, struct nvme_id_ns_nvm *nvm)
1861 u32 elbaf = le32_to_cpu(nvm->elbaf[nvme_lbaf_index(id->flbas)]);
1863 /* no support for storage tag formats right now */
1864 if (nvme_elbaf_sts(elbaf))
1867 head->guard_type = nvme_elbaf_guard_type(elbaf);
1868 switch (head->guard_type) {
1869 case NVME_NVM_NS_64B_GUARD:
1870 head->pi_size = sizeof(struct crc64_pi_tuple);
1872 case NVME_NVM_NS_16B_GUARD:
1873 head->pi_size = sizeof(struct t10_pi_tuple);
1880 static void nvme_configure_metadata(struct nvme_ctrl *ctrl,
1881 struct nvme_ns_head *head, struct nvme_id_ns *id,
1882 struct nvme_id_ns_nvm *nvm)
1884 head->features &= ~(NVME_NS_METADATA_SUPPORTED | NVME_NS_EXT_LBAS);
1887 head->pi_offset = 0;
1888 head->ms = le16_to_cpu(id->lbaf[nvme_lbaf_index(id->flbas)].ms);
1889 if (!head->ms || !(ctrl->ops->flags & NVME_F_METADATA_SUPPORTED))
1892 if (nvm && (ctrl->ctratt & NVME_CTRL_ATTR_ELBAS)) {
1893 nvme_configure_pi_elbas(head, id, nvm);
1895 head->pi_size = sizeof(struct t10_pi_tuple);
1896 head->guard_type = NVME_NVM_NS_16B_GUARD;
1899 if (head->pi_size && head->ms >= head->pi_size)
1900 head->pi_type = id->dps & NVME_NS_DPS_PI_MASK;
1901 if (!(id->dps & NVME_NS_DPS_PI_FIRST))
1902 head->pi_offset = head->ms - head->pi_size;
1904 if (ctrl->ops->flags & NVME_F_FABRICS) {
1906 * The NVMe over Fabrics specification only supports metadata as
1907 * part of the extended data LBA. We rely on HCA/HBA support to
1908 * remap the separate metadata buffer from the block layer.
1910 if (WARN_ON_ONCE(!(id->flbas & NVME_NS_FLBAS_META_EXT)))
1913 head->features |= NVME_NS_EXT_LBAS;
1916 * The current fabrics transport drivers support namespace
1917 * metadata formats only if nvme_ns_has_pi() returns true.
1918 * Suppress support for all other formats so the namespace will
1919 * have a 0 capacity and not be usable through the block stack.
1921 * Note, this check will need to be modified if any drivers
1922 * gain the ability to use other metadata formats.
1924 if (ctrl->max_integrity_segments && nvme_ns_has_pi(head))
1925 head->features |= NVME_NS_METADATA_SUPPORTED;
1928 * For PCIe controllers, we can't easily remap the separate
1929 * metadata buffer from the block layer and thus require a
1930 * separate metadata buffer for block layer metadata/PI support.
1931 * We allow extended LBAs for the passthrough interface, though.
1933 if (id->flbas & NVME_NS_FLBAS_META_EXT)
1934 head->features |= NVME_NS_EXT_LBAS;
1936 head->features |= NVME_NS_METADATA_SUPPORTED;
1940 static u32 nvme_max_drv_segments(struct nvme_ctrl *ctrl)
1942 return ctrl->max_hw_sectors / (NVME_CTRL_PAGE_SIZE >> SECTOR_SHIFT) + 1;
1945 static void nvme_set_ctrl_limits(struct nvme_ctrl *ctrl,
1946 struct queue_limits *lim)
1948 lim->max_hw_sectors = ctrl->max_hw_sectors;
1949 lim->max_segments = min_t(u32, USHRT_MAX,
1950 min_not_zero(nvme_max_drv_segments(ctrl), ctrl->max_segments));
1951 lim->max_integrity_segments = ctrl->max_integrity_segments;
1952 lim->virt_boundary_mask = NVME_CTRL_PAGE_SIZE - 1;
1953 lim->max_segment_size = UINT_MAX;
1954 lim->dma_alignment = 3;
1957 static bool nvme_update_disk_info(struct nvme_ns *ns, struct nvme_id_ns *id,
1958 struct queue_limits *lim)
1960 struct nvme_ns_head *head = ns->head;
1961 u32 bs = 1U << head->lba_shift;
1962 u32 atomic_bs, phys_bs, io_opt = 0;
1966 * The block layer can't support LBA sizes larger than the page size
1967 * or smaller than a sector size yet, so catch this early and don't
1970 if (head->lba_shift > PAGE_SHIFT || head->lba_shift < SECTOR_SHIFT) {
1975 atomic_bs = phys_bs = bs;
1976 if (id->nabo == 0) {
1978 * Bit 1 indicates whether NAWUPF is defined for this namespace
1979 * and whether it should be used instead of AWUPF. If NAWUPF ==
1980 * 0 then AWUPF must be used instead.
1982 if (id->nsfeat & NVME_NS_FEAT_ATOMICS && id->nawupf)
1983 atomic_bs = (1 + le16_to_cpu(id->nawupf)) * bs;
1985 atomic_bs = (1 + ns->ctrl->subsys->awupf) * bs;
1988 if (id->nsfeat & NVME_NS_FEAT_IO_OPT) {
1989 /* NPWG = Namespace Preferred Write Granularity */
1990 phys_bs = bs * (1 + le16_to_cpu(id->npwg));
1991 /* NOWS = Namespace Optimal Write Size */
1992 io_opt = bs * (1 + le16_to_cpu(id->nows));
1996 * Linux filesystems assume writing a single physical block is
1997 * an atomic operation. Hence limit the physical block size to the
1998 * value of the Atomic Write Unit Power Fail parameter.
2000 lim->logical_block_size = bs;
2001 lim->physical_block_size = min(phys_bs, atomic_bs);
2002 lim->io_min = phys_bs;
2003 lim->io_opt = io_opt;
2004 if (ns->ctrl->quirks & NVME_QUIRK_DEALLOCATE_ZEROES)
2005 lim->max_write_zeroes_sectors = UINT_MAX;
2007 lim->max_write_zeroes_sectors = ns->ctrl->max_zeroes_sectors;
2011 static bool nvme_ns_is_readonly(struct nvme_ns *ns, struct nvme_ns_info *info)
2013 return info->is_readonly || test_bit(NVME_NS_FORCE_RO, &ns->flags);
2016 static inline bool nvme_first_scan(struct gendisk *disk)
2018 /* nvme_alloc_ns() scans the disk prior to adding it */
2019 return !disk_live(disk);
2022 static void nvme_set_chunk_sectors(struct nvme_ns *ns, struct nvme_id_ns *id,
2023 struct queue_limits *lim)
2025 struct nvme_ctrl *ctrl = ns->ctrl;
2028 if ((ctrl->quirks & NVME_QUIRK_STRIPE_SIZE) &&
2029 is_power_of_2(ctrl->max_hw_sectors))
2030 iob = ctrl->max_hw_sectors;
2032 iob = nvme_lba_to_sect(ns->head, le16_to_cpu(id->noiob));
2037 if (!is_power_of_2(iob)) {
2038 if (nvme_first_scan(ns->disk))
2039 pr_warn("%s: ignoring unaligned IO boundary:%u\n",
2040 ns->disk->disk_name, iob);
2044 if (blk_queue_is_zoned(ns->disk->queue)) {
2045 if (nvme_first_scan(ns->disk))
2046 pr_warn("%s: ignoring zoned namespace IO boundary\n",
2047 ns->disk->disk_name);
2051 lim->chunk_sectors = iob;
2054 static int nvme_update_ns_info_generic(struct nvme_ns *ns,
2055 struct nvme_ns_info *info)
2057 struct queue_limits lim;
2060 blk_mq_freeze_queue(ns->disk->queue);
2061 lim = queue_limits_start_update(ns->disk->queue);
2062 nvme_set_ctrl_limits(ns->ctrl, &lim);
2063 ret = queue_limits_commit_update(ns->disk->queue, &lim);
2064 set_disk_ro(ns->disk, nvme_ns_is_readonly(ns, info));
2065 blk_mq_unfreeze_queue(ns->disk->queue);
2067 /* Hide the block-interface for these devices */
2073 static int nvme_update_ns_info_block(struct nvme_ns *ns,
2074 struct nvme_ns_info *info)
2076 bool vwc = ns->ctrl->vwc & NVME_CTRL_VWC_PRESENT;
2077 struct queue_limits lim;
2078 struct nvme_id_ns_nvm *nvm = NULL;
2079 struct nvme_zone_info zi = {};
2080 struct nvme_id_ns *id;
2085 ret = nvme_identify_ns(ns->ctrl, info->nsid, &id);
2089 if (id->ncap == 0) {
2090 /* namespace not allocated or attached */
2091 info->is_removed = true;
2095 lbaf = nvme_lbaf_index(id->flbas);
2097 if (ns->ctrl->ctratt & NVME_CTRL_ATTR_ELBAS) {
2098 ret = nvme_identify_ns_nvm(ns->ctrl, info->nsid, &nvm);
2103 if (IS_ENABLED(CONFIG_BLK_DEV_ZONED) &&
2104 ns->head->ids.csi == NVME_CSI_ZNS) {
2105 ret = nvme_query_zone_info(ns, lbaf, &zi);
2110 blk_mq_freeze_queue(ns->disk->queue);
2111 ns->head->lba_shift = id->lbaf[lbaf].ds;
2112 ns->head->nuse = le64_to_cpu(id->nuse);
2113 capacity = nvme_lba_to_sect(ns->head, le64_to_cpu(id->nsze));
2115 lim = queue_limits_start_update(ns->disk->queue);
2116 nvme_set_ctrl_limits(ns->ctrl, &lim);
2117 nvme_configure_metadata(ns->ctrl, ns->head, id, nvm);
2118 nvme_set_chunk_sectors(ns, id, &lim);
2119 if (!nvme_update_disk_info(ns, id, &lim))
2121 nvme_config_discard(ns, &lim);
2122 if (IS_ENABLED(CONFIG_BLK_DEV_ZONED) &&
2123 ns->head->ids.csi == NVME_CSI_ZNS)
2124 nvme_update_zone_info(ns, &lim, &zi);
2125 ret = queue_limits_commit_update(ns->disk->queue, &lim);
2127 blk_mq_unfreeze_queue(ns->disk->queue);
2132 * Register a metadata profile for PI, or the plain non-integrity NVMe
2133 * metadata masquerading as Type 0 if supported, otherwise reject block
2134 * I/O to namespaces with metadata except when the namespace supports
2135 * PI, as it can strip/insert in that case.
2137 if (!nvme_init_integrity(ns->disk, ns->head))
2140 set_capacity_and_notify(ns->disk, capacity);
2143 * Only set the DEAC bit if the device guarantees that reads from
2144 * deallocated data return zeroes. While the DEAC bit does not
2145 * require that, it must be a no-op if reads from deallocated data
2146 * do not return zeroes.
2148 if ((id->dlfeat & 0x7) == 0x1 && (id->dlfeat & (1 << 3)))
2149 ns->head->features |= NVME_NS_DEAC;
2150 set_disk_ro(ns->disk, nvme_ns_is_readonly(ns, info));
2151 blk_queue_write_cache(ns->disk->queue, vwc, vwc);
2152 set_bit(NVME_NS_READY, &ns->flags);
2153 blk_mq_unfreeze_queue(ns->disk->queue);
2155 if (blk_queue_is_zoned(ns->queue)) {
2156 ret = blk_revalidate_disk_zones(ns->disk, NULL);
2157 if (ret && !nvme_first_scan(ns->disk))
2168 static int nvme_update_ns_info(struct nvme_ns *ns, struct nvme_ns_info *info)
2170 bool unsupported = false;
2173 switch (info->ids.csi) {
2175 if (!IS_ENABLED(CONFIG_BLK_DEV_ZONED)) {
2176 dev_info(ns->ctrl->device,
2177 "block device for nsid %u not supported without CONFIG_BLK_DEV_ZONED\n",
2179 ret = nvme_update_ns_info_generic(ns, info);
2182 ret = nvme_update_ns_info_block(ns, info);
2185 ret = nvme_update_ns_info_block(ns, info);
2188 dev_info(ns->ctrl->device,
2189 "block device for nsid %u not supported (csi %u)\n",
2190 info->nsid, info->ids.csi);
2191 ret = nvme_update_ns_info_generic(ns, info);
2196 * If probing fails due an unsupported feature, hide the block device,
2197 * but still allow other access.
2199 if (ret == -ENODEV) {
2200 ns->disk->flags |= GENHD_FL_HIDDEN;
2201 set_bit(NVME_NS_READY, &ns->flags);
2206 if (!ret && nvme_ns_head_multipath(ns->head)) {
2207 struct queue_limits *ns_lim = &ns->disk->queue->limits;
2208 struct queue_limits lim;
2210 blk_mq_freeze_queue(ns->head->disk->queue);
2212 ns->head->disk->flags |= GENHD_FL_HIDDEN;
2214 nvme_init_integrity(ns->head->disk, ns->head);
2215 set_capacity_and_notify(ns->head->disk, get_capacity(ns->disk));
2216 set_disk_ro(ns->head->disk, nvme_ns_is_readonly(ns, info));
2217 nvme_mpath_revalidate_paths(ns);
2220 * queue_limits mixes values that are the hardware limitations
2221 * for bio splitting with what is the device configuration.
2223 * For NVMe the device configuration can change after e.g. a
2224 * Format command, and we really want to pick up the new format
2225 * value here. But we must still stack the queue limits to the
2226 * least common denominator for multipathing to split the bios
2229 * To work around this, we explicitly set the device
2230 * configuration to those that we just queried, but only stack
2231 * the splitting limits in to make sure we still obey possibly
2232 * lower limitations of other controllers.
2234 lim = queue_limits_start_update(ns->head->disk->queue);
2235 lim.logical_block_size = ns_lim->logical_block_size;
2236 lim.physical_block_size = ns_lim->physical_block_size;
2237 lim.io_min = ns_lim->io_min;
2238 lim.io_opt = ns_lim->io_opt;
2239 queue_limits_stack_bdev(&lim, ns->disk->part0, 0,
2240 ns->head->disk->disk_name);
2241 ret = queue_limits_commit_update(ns->head->disk->queue, &lim);
2242 blk_mq_unfreeze_queue(ns->head->disk->queue);
2248 #ifdef CONFIG_BLK_SED_OPAL
2249 static int nvme_sec_submit(void *data, u16 spsp, u8 secp, void *buffer, size_t len,
2252 struct nvme_ctrl *ctrl = data;
2253 struct nvme_command cmd = { };
2256 cmd.common.opcode = nvme_admin_security_send;
2258 cmd.common.opcode = nvme_admin_security_recv;
2259 cmd.common.nsid = 0;
2260 cmd.common.cdw10 = cpu_to_le32(((u32)secp) << 24 | ((u32)spsp) << 8);
2261 cmd.common.cdw11 = cpu_to_le32(len);
2263 return __nvme_submit_sync_cmd(ctrl->admin_q, &cmd, NULL, buffer, len,
2264 NVME_QID_ANY, NVME_SUBMIT_AT_HEAD);
2267 static void nvme_configure_opal(struct nvme_ctrl *ctrl, bool was_suspended)
2269 if (ctrl->oacs & NVME_CTRL_OACS_SEC_SUPP) {
2270 if (!ctrl->opal_dev)
2271 ctrl->opal_dev = init_opal_dev(ctrl, &nvme_sec_submit);
2272 else if (was_suspended)
2273 opal_unlock_from_suspend(ctrl->opal_dev);
2275 free_opal_dev(ctrl->opal_dev);
2276 ctrl->opal_dev = NULL;
2280 static void nvme_configure_opal(struct nvme_ctrl *ctrl, bool was_suspended)
2283 #endif /* CONFIG_BLK_SED_OPAL */
2285 #ifdef CONFIG_BLK_DEV_ZONED
2286 static int nvme_report_zones(struct gendisk *disk, sector_t sector,
2287 unsigned int nr_zones, report_zones_cb cb, void *data)
2289 return nvme_ns_report_zones(disk->private_data, sector, nr_zones, cb,
2293 #define nvme_report_zones NULL
2294 #endif /* CONFIG_BLK_DEV_ZONED */
2296 const struct block_device_operations nvme_bdev_ops = {
2297 .owner = THIS_MODULE,
2298 .ioctl = nvme_ioctl,
2299 .compat_ioctl = blkdev_compat_ptr_ioctl,
2301 .release = nvme_release,
2302 .getgeo = nvme_getgeo,
2303 .report_zones = nvme_report_zones,
2304 .pr_ops = &nvme_pr_ops,
2307 static int nvme_wait_ready(struct nvme_ctrl *ctrl, u32 mask, u32 val,
2308 u32 timeout, const char *op)
2310 unsigned long timeout_jiffies = jiffies + timeout * HZ;
2314 while ((ret = ctrl->ops->reg_read32(ctrl, NVME_REG_CSTS, &csts)) == 0) {
2317 if ((csts & mask) == val)
2320 usleep_range(1000, 2000);
2321 if (fatal_signal_pending(current))
2323 if (time_after(jiffies, timeout_jiffies)) {
2324 dev_err(ctrl->device,
2325 "Device not ready; aborting %s, CSTS=0x%x\n",
2334 int nvme_disable_ctrl(struct nvme_ctrl *ctrl, bool shutdown)
2338 ctrl->ctrl_config &= ~NVME_CC_SHN_MASK;
2340 ctrl->ctrl_config |= NVME_CC_SHN_NORMAL;
2342 ctrl->ctrl_config &= ~NVME_CC_ENABLE;
2344 ret = ctrl->ops->reg_write32(ctrl, NVME_REG_CC, ctrl->ctrl_config);
2349 return nvme_wait_ready(ctrl, NVME_CSTS_SHST_MASK,
2350 NVME_CSTS_SHST_CMPLT,
2351 ctrl->shutdown_timeout, "shutdown");
2353 if (ctrl->quirks & NVME_QUIRK_DELAY_BEFORE_CHK_RDY)
2354 msleep(NVME_QUIRK_DELAY_AMOUNT);
2355 return nvme_wait_ready(ctrl, NVME_CSTS_RDY, 0,
2356 (NVME_CAP_TIMEOUT(ctrl->cap) + 1) / 2, "reset");
2358 EXPORT_SYMBOL_GPL(nvme_disable_ctrl);
2360 int nvme_enable_ctrl(struct nvme_ctrl *ctrl)
2362 unsigned dev_page_min;
2366 ret = ctrl->ops->reg_read64(ctrl, NVME_REG_CAP, &ctrl->cap);
2368 dev_err(ctrl->device, "Reading CAP failed (%d)\n", ret);
2371 dev_page_min = NVME_CAP_MPSMIN(ctrl->cap) + 12;
2373 if (NVME_CTRL_PAGE_SHIFT < dev_page_min) {
2374 dev_err(ctrl->device,
2375 "Minimum device page size %u too large for host (%u)\n",
2376 1 << dev_page_min, 1 << NVME_CTRL_PAGE_SHIFT);
2380 if (NVME_CAP_CSS(ctrl->cap) & NVME_CAP_CSS_CSI)
2381 ctrl->ctrl_config = NVME_CC_CSS_CSI;
2383 ctrl->ctrl_config = NVME_CC_CSS_NVM;
2385 if (ctrl->cap & NVME_CAP_CRMS_CRWMS && ctrl->cap & NVME_CAP_CRMS_CRIMS)
2386 ctrl->ctrl_config |= NVME_CC_CRIME;
2388 ctrl->ctrl_config |= (NVME_CTRL_PAGE_SHIFT - 12) << NVME_CC_MPS_SHIFT;
2389 ctrl->ctrl_config |= NVME_CC_AMS_RR | NVME_CC_SHN_NONE;
2390 ctrl->ctrl_config |= NVME_CC_IOSQES | NVME_CC_IOCQES;
2391 ret = ctrl->ops->reg_write32(ctrl, NVME_REG_CC, ctrl->ctrl_config);
2395 /* Flush write to device (required if transport is PCI) */
2396 ret = ctrl->ops->reg_read32(ctrl, NVME_REG_CC, &ctrl->ctrl_config);
2400 /* CAP value may change after initial CC write */
2401 ret = ctrl->ops->reg_read64(ctrl, NVME_REG_CAP, &ctrl->cap);
2405 timeout = NVME_CAP_TIMEOUT(ctrl->cap);
2406 if (ctrl->cap & NVME_CAP_CRMS_CRWMS) {
2407 u32 crto, ready_timeout;
2409 ret = ctrl->ops->reg_read32(ctrl, NVME_REG_CRTO, &crto);
2411 dev_err(ctrl->device, "Reading CRTO failed (%d)\n",
2417 * CRTO should always be greater or equal to CAP.TO, but some
2418 * devices are known to get this wrong. Use the larger of the
2421 if (ctrl->ctrl_config & NVME_CC_CRIME)
2422 ready_timeout = NVME_CRTO_CRIMT(crto);
2424 ready_timeout = NVME_CRTO_CRWMT(crto);
2426 if (ready_timeout < timeout)
2427 dev_warn_once(ctrl->device, "bad crto:%x cap:%llx\n",
2430 timeout = ready_timeout;
2433 ctrl->ctrl_config |= NVME_CC_ENABLE;
2434 ret = ctrl->ops->reg_write32(ctrl, NVME_REG_CC, ctrl->ctrl_config);
2437 return nvme_wait_ready(ctrl, NVME_CSTS_RDY, NVME_CSTS_RDY,
2438 (timeout + 1) / 2, "initialisation");
2440 EXPORT_SYMBOL_GPL(nvme_enable_ctrl);
2442 static int nvme_configure_timestamp(struct nvme_ctrl *ctrl)
2447 if (!(ctrl->oncs & NVME_CTRL_ONCS_TIMESTAMP))
2450 ts = cpu_to_le64(ktime_to_ms(ktime_get_real()));
2451 ret = nvme_set_features(ctrl, NVME_FEAT_TIMESTAMP, 0, &ts, sizeof(ts),
2454 dev_warn_once(ctrl->device,
2455 "could not set timestamp (%d)\n", ret);
2459 static int nvme_configure_host_options(struct nvme_ctrl *ctrl)
2461 struct nvme_feat_host_behavior *host;
2462 u8 acre = 0, lbafee = 0;
2465 /* Don't bother enabling the feature if retry delay is not reported */
2467 acre = NVME_ENABLE_ACRE;
2468 if (ctrl->ctratt & NVME_CTRL_ATTR_ELBAS)
2469 lbafee = NVME_ENABLE_LBAFEE;
2471 if (!acre && !lbafee)
2474 host = kzalloc(sizeof(*host), GFP_KERNEL);
2479 host->lbafee = lbafee;
2480 ret = nvme_set_features(ctrl, NVME_FEAT_HOST_BEHAVIOR, 0,
2481 host, sizeof(*host), NULL);
2487 * The function checks whether the given total (exlat + enlat) latency of
2488 * a power state allows the latter to be used as an APST transition target.
2489 * It does so by comparing the latency to the primary and secondary latency
2490 * tolerances defined by module params. If there's a match, the corresponding
2491 * timeout value is returned and the matching tolerance index (1 or 2) is
2494 static bool nvme_apst_get_transition_time(u64 total_latency,
2495 u64 *transition_time, unsigned *last_index)
2497 if (total_latency <= apst_primary_latency_tol_us) {
2498 if (*last_index == 1)
2501 *transition_time = apst_primary_timeout_ms;
2504 if (apst_secondary_timeout_ms &&
2505 total_latency <= apst_secondary_latency_tol_us) {
2506 if (*last_index <= 2)
2509 *transition_time = apst_secondary_timeout_ms;
2516 * APST (Autonomous Power State Transition) lets us program a table of power
2517 * state transitions that the controller will perform automatically.
2519 * Depending on module params, one of the two supported techniques will be used:
2521 * - If the parameters provide explicit timeouts and tolerances, they will be
2522 * used to build a table with up to 2 non-operational states to transition to.
2523 * The default parameter values were selected based on the values used by
2524 * Microsoft's and Intel's NVMe drivers. Yet, since we don't implement dynamic
2525 * regeneration of the APST table in the event of switching between external
2526 * and battery power, the timeouts and tolerances reflect a compromise
2527 * between values used by Microsoft for AC and battery scenarios.
2528 * - If not, we'll configure the table with a simple heuristic: we are willing
2529 * to spend at most 2% of the time transitioning between power states.
2530 * Therefore, when running in any given state, we will enter the next
2531 * lower-power non-operational state after waiting 50 * (enlat + exlat)
2532 * microseconds, as long as that state's exit latency is under the requested
2535 * We will not autonomously enter any non-operational state for which the total
2536 * latency exceeds ps_max_latency_us.
2538 * Users can set ps_max_latency_us to zero to turn off APST.
2540 static int nvme_configure_apst(struct nvme_ctrl *ctrl)
2542 struct nvme_feat_auto_pst *table;
2549 unsigned last_lt_index = UINT_MAX;
2552 * If APST isn't supported or if we haven't been initialized yet,
2553 * then don't do anything.
2558 if (ctrl->npss > 31) {
2559 dev_warn(ctrl->device, "NPSS is invalid; not using APST\n");
2563 table = kzalloc(sizeof(*table), GFP_KERNEL);
2567 if (!ctrl->apst_enabled || ctrl->ps_max_latency_us == 0) {
2568 /* Turn off APST. */
2569 dev_dbg(ctrl->device, "APST disabled\n");
2574 * Walk through all states from lowest- to highest-power.
2575 * According to the spec, lower-numbered states use more power. NPSS,
2576 * despite the name, is the index of the lowest-power state, not the
2579 for (state = (int)ctrl->npss; state >= 0; state--) {
2580 u64 total_latency_us, exit_latency_us, transition_ms;
2583 table->entries[state] = target;
2586 * Don't allow transitions to the deepest state if it's quirked
2589 if (state == ctrl->npss &&
2590 (ctrl->quirks & NVME_QUIRK_NO_DEEPEST_PS))
2594 * Is this state a useful non-operational state for higher-power
2595 * states to autonomously transition to?
2597 if (!(ctrl->psd[state].flags & NVME_PS_FLAGS_NON_OP_STATE))
2600 exit_latency_us = (u64)le32_to_cpu(ctrl->psd[state].exit_lat);
2601 if (exit_latency_us > ctrl->ps_max_latency_us)
2604 total_latency_us = exit_latency_us +
2605 le32_to_cpu(ctrl->psd[state].entry_lat);
2608 * This state is good. It can be used as the APST idle target
2609 * for higher power states.
2611 if (apst_primary_timeout_ms && apst_primary_latency_tol_us) {
2612 if (!nvme_apst_get_transition_time(total_latency_us,
2613 &transition_ms, &last_lt_index))
2616 transition_ms = total_latency_us + 19;
2617 do_div(transition_ms, 20);
2618 if (transition_ms > (1 << 24) - 1)
2619 transition_ms = (1 << 24) - 1;
2622 target = cpu_to_le64((state << 3) | (transition_ms << 8));
2625 if (total_latency_us > max_lat_us)
2626 max_lat_us = total_latency_us;
2630 dev_dbg(ctrl->device, "APST enabled but no non-operational states are available\n");
2632 dev_dbg(ctrl->device, "APST enabled: max PS = %d, max round-trip latency = %lluus, table = %*phN\n",
2633 max_ps, max_lat_us, (int)sizeof(*table), table);
2637 ret = nvme_set_features(ctrl, NVME_FEAT_AUTO_PST, apste,
2638 table, sizeof(*table), NULL);
2640 dev_err(ctrl->device, "failed to set APST feature (%d)\n", ret);
2645 static void nvme_set_latency_tolerance(struct device *dev, s32 val)
2647 struct nvme_ctrl *ctrl = dev_get_drvdata(dev);
2651 case PM_QOS_LATENCY_TOLERANCE_NO_CONSTRAINT:
2652 case PM_QOS_LATENCY_ANY:
2660 if (ctrl->ps_max_latency_us != latency) {
2661 ctrl->ps_max_latency_us = latency;
2662 if (nvme_ctrl_state(ctrl) == NVME_CTRL_LIVE)
2663 nvme_configure_apst(ctrl);
2667 struct nvme_core_quirk_entry {
2669 * NVMe model and firmware strings are padded with spaces. For
2670 * simplicity, strings in the quirk table are padded with NULLs
2676 unsigned long quirks;
2679 static const struct nvme_core_quirk_entry core_quirks[] = {
2682 * This Toshiba device seems to die using any APST states. See:
2683 * https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1678184/comments/11
2686 .mn = "THNSF5256GPUK TOSHIBA",
2687 .quirks = NVME_QUIRK_NO_APST,
2691 * This LiteON CL1-3D*-Q11 firmware version has a race
2692 * condition associated with actions related to suspend to idle
2693 * LiteON has resolved the problem in future firmware
2697 .quirks = NVME_QUIRK_SIMPLE_SUSPEND,
2701 * This Kioxia CD6-V Series / HPE PE8030 device times out and
2702 * aborts I/O during any load, but more easily reproducible
2703 * with discards (fstrim).
2705 * The device is left in a state where it is also not possible
2706 * to use "nvme set-feature" to disable APST, but booting with
2707 * nvme_core.default_ps_max_latency=0 works.
2710 .mn = "KCD6XVUL6T40",
2711 .quirks = NVME_QUIRK_NO_APST,
2715 * The external Samsung X5 SSD fails initialization without a
2716 * delay before checking if it is ready and has a whole set of
2717 * other problems. To make this even more interesting, it
2718 * shares the PCI ID with internal Samsung 970 Evo Plus that
2719 * does not need or want these quirks.
2722 .mn = "Samsung Portable SSD X5",
2723 .quirks = NVME_QUIRK_DELAY_BEFORE_CHK_RDY |
2724 NVME_QUIRK_NO_DEEPEST_PS |
2725 NVME_QUIRK_IGNORE_DEV_SUBNQN,
2729 /* match is null-terminated but idstr is space-padded. */
2730 static bool string_matches(const char *idstr, const char *match, size_t len)
2737 matchlen = strlen(match);
2738 WARN_ON_ONCE(matchlen > len);
2740 if (memcmp(idstr, match, matchlen))
2743 for (; matchlen < len; matchlen++)
2744 if (idstr[matchlen] != ' ')
2750 static bool quirk_matches(const struct nvme_id_ctrl *id,
2751 const struct nvme_core_quirk_entry *q)
2753 return q->vid == le16_to_cpu(id->vid) &&
2754 string_matches(id->mn, q->mn, sizeof(id->mn)) &&
2755 string_matches(id->fr, q->fr, sizeof(id->fr));
2758 static void nvme_init_subnqn(struct nvme_subsystem *subsys, struct nvme_ctrl *ctrl,
2759 struct nvme_id_ctrl *id)
2764 if(!(ctrl->quirks & NVME_QUIRK_IGNORE_DEV_SUBNQN)) {
2765 nqnlen = strnlen(id->subnqn, NVMF_NQN_SIZE);
2766 if (nqnlen > 0 && nqnlen < NVMF_NQN_SIZE) {
2767 strscpy(subsys->subnqn, id->subnqn, NVMF_NQN_SIZE);
2771 if (ctrl->vs >= NVME_VS(1, 2, 1))
2772 dev_warn(ctrl->device, "missing or invalid SUBNQN field.\n");
2776 * Generate a "fake" NQN similar to the one in Section 4.5 of the NVMe
2777 * Base Specification 2.0. It is slightly different from the format
2778 * specified there due to historic reasons, and we can't change it now.
2780 off = snprintf(subsys->subnqn, NVMF_NQN_SIZE,
2781 "nqn.2014.08.org.nvmexpress:%04x%04x",
2782 le16_to_cpu(id->vid), le16_to_cpu(id->ssvid));
2783 memcpy(subsys->subnqn + off, id->sn, sizeof(id->sn));
2784 off += sizeof(id->sn);
2785 memcpy(subsys->subnqn + off, id->mn, sizeof(id->mn));
2786 off += sizeof(id->mn);
2787 memset(subsys->subnqn + off, 0, sizeof(subsys->subnqn) - off);
2790 static void nvme_release_subsystem(struct device *dev)
2792 struct nvme_subsystem *subsys =
2793 container_of(dev, struct nvme_subsystem, dev);
2795 if (subsys->instance >= 0)
2796 ida_free(&nvme_instance_ida, subsys->instance);
2800 static void nvme_destroy_subsystem(struct kref *ref)
2802 struct nvme_subsystem *subsys =
2803 container_of(ref, struct nvme_subsystem, ref);
2805 mutex_lock(&nvme_subsystems_lock);
2806 list_del(&subsys->entry);
2807 mutex_unlock(&nvme_subsystems_lock);
2809 ida_destroy(&subsys->ns_ida);
2810 device_del(&subsys->dev);
2811 put_device(&subsys->dev);
2814 static void nvme_put_subsystem(struct nvme_subsystem *subsys)
2816 kref_put(&subsys->ref, nvme_destroy_subsystem);
2819 static struct nvme_subsystem *__nvme_find_get_subsystem(const char *subsysnqn)
2821 struct nvme_subsystem *subsys;
2823 lockdep_assert_held(&nvme_subsystems_lock);
2826 * Fail matches for discovery subsystems. This results
2827 * in each discovery controller bound to a unique subsystem.
2828 * This avoids issues with validating controller values
2829 * that can only be true when there is a single unique subsystem.
2830 * There may be multiple and completely independent entities
2831 * that provide discovery controllers.
2833 if (!strcmp(subsysnqn, NVME_DISC_SUBSYS_NAME))
2836 list_for_each_entry(subsys, &nvme_subsystems, entry) {
2837 if (strcmp(subsys->subnqn, subsysnqn))
2839 if (!kref_get_unless_zero(&subsys->ref))
2847 static inline bool nvme_discovery_ctrl(struct nvme_ctrl *ctrl)
2849 return ctrl->opts && ctrl->opts->discovery_nqn;
2852 static bool nvme_validate_cntlid(struct nvme_subsystem *subsys,
2853 struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
2855 struct nvme_ctrl *tmp;
2857 lockdep_assert_held(&nvme_subsystems_lock);
2859 list_for_each_entry(tmp, &subsys->ctrls, subsys_entry) {
2860 if (nvme_state_terminal(tmp))
2863 if (tmp->cntlid == ctrl->cntlid) {
2864 dev_err(ctrl->device,
2865 "Duplicate cntlid %u with %s, subsys %s, rejecting\n",
2866 ctrl->cntlid, dev_name(tmp->device),
2871 if ((id->cmic & NVME_CTRL_CMIC_MULTI_CTRL) ||
2872 nvme_discovery_ctrl(ctrl))
2875 dev_err(ctrl->device,
2876 "Subsystem does not support multiple controllers\n");
2883 static int nvme_init_subsystem(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
2885 struct nvme_subsystem *subsys, *found;
2888 subsys = kzalloc(sizeof(*subsys), GFP_KERNEL);
2892 subsys->instance = -1;
2893 mutex_init(&subsys->lock);
2894 kref_init(&subsys->ref);
2895 INIT_LIST_HEAD(&subsys->ctrls);
2896 INIT_LIST_HEAD(&subsys->nsheads);
2897 nvme_init_subnqn(subsys, ctrl, id);
2898 memcpy(subsys->serial, id->sn, sizeof(subsys->serial));
2899 memcpy(subsys->model, id->mn, sizeof(subsys->model));
2900 subsys->vendor_id = le16_to_cpu(id->vid);
2901 subsys->cmic = id->cmic;
2903 /* Versions prior to 1.4 don't necessarily report a valid type */
2904 if (id->cntrltype == NVME_CTRL_DISC ||
2905 !strcmp(subsys->subnqn, NVME_DISC_SUBSYS_NAME))
2906 subsys->subtype = NVME_NQN_DISC;
2908 subsys->subtype = NVME_NQN_NVME;
2910 if (nvme_discovery_ctrl(ctrl) && subsys->subtype != NVME_NQN_DISC) {
2911 dev_err(ctrl->device,
2912 "Subsystem %s is not a discovery controller",
2917 subsys->awupf = le16_to_cpu(id->awupf);
2918 nvme_mpath_default_iopolicy(subsys);
2920 subsys->dev.class = &nvme_subsys_class;
2921 subsys->dev.release = nvme_release_subsystem;
2922 subsys->dev.groups = nvme_subsys_attrs_groups;
2923 dev_set_name(&subsys->dev, "nvme-subsys%d", ctrl->instance);
2924 device_initialize(&subsys->dev);
2926 mutex_lock(&nvme_subsystems_lock);
2927 found = __nvme_find_get_subsystem(subsys->subnqn);
2929 put_device(&subsys->dev);
2932 if (!nvme_validate_cntlid(subsys, ctrl, id)) {
2934 goto out_put_subsystem;
2937 ret = device_add(&subsys->dev);
2939 dev_err(ctrl->device,
2940 "failed to register subsystem device.\n");
2941 put_device(&subsys->dev);
2944 ida_init(&subsys->ns_ida);
2945 list_add_tail(&subsys->entry, &nvme_subsystems);
2948 ret = sysfs_create_link(&subsys->dev.kobj, &ctrl->device->kobj,
2949 dev_name(ctrl->device));
2951 dev_err(ctrl->device,
2952 "failed to create sysfs link from subsystem.\n");
2953 goto out_put_subsystem;
2957 subsys->instance = ctrl->instance;
2958 ctrl->subsys = subsys;
2959 list_add_tail(&ctrl->subsys_entry, &subsys->ctrls);
2960 mutex_unlock(&nvme_subsystems_lock);
2964 nvme_put_subsystem(subsys);
2966 mutex_unlock(&nvme_subsystems_lock);
2970 int nvme_get_log(struct nvme_ctrl *ctrl, u32 nsid, u8 log_page, u8 lsp, u8 csi,
2971 void *log, size_t size, u64 offset)
2973 struct nvme_command c = { };
2974 u32 dwlen = nvme_bytes_to_numd(size);
2976 c.get_log_page.opcode = nvme_admin_get_log_page;
2977 c.get_log_page.nsid = cpu_to_le32(nsid);
2978 c.get_log_page.lid = log_page;
2979 c.get_log_page.lsp = lsp;
2980 c.get_log_page.numdl = cpu_to_le16(dwlen & ((1 << 16) - 1));
2981 c.get_log_page.numdu = cpu_to_le16(dwlen >> 16);
2982 c.get_log_page.lpol = cpu_to_le32(lower_32_bits(offset));
2983 c.get_log_page.lpou = cpu_to_le32(upper_32_bits(offset));
2984 c.get_log_page.csi = csi;
2986 return nvme_submit_sync_cmd(ctrl->admin_q, &c, log, size);
2989 static int nvme_get_effects_log(struct nvme_ctrl *ctrl, u8 csi,
2990 struct nvme_effects_log **log)
2992 struct nvme_effects_log *cel = xa_load(&ctrl->cels, csi);
2998 cel = kzalloc(sizeof(*cel), GFP_KERNEL);
3002 ret = nvme_get_log(ctrl, 0x00, NVME_LOG_CMD_EFFECTS, 0, csi,
3003 cel, sizeof(*cel), 0);
3009 xa_store(&ctrl->cels, csi, cel, GFP_KERNEL);
3015 static inline u32 nvme_mps_to_sectors(struct nvme_ctrl *ctrl, u32 units)
3017 u32 page_shift = NVME_CAP_MPSMIN(ctrl->cap) + 12, val;
3019 if (check_shl_overflow(1U, units + page_shift - 9, &val))
3024 static int nvme_init_non_mdts_limits(struct nvme_ctrl *ctrl)
3026 struct nvme_command c = { };
3027 struct nvme_id_ctrl_nvm *id;
3031 * Even though NVMe spec explicitly states that MDTS is not applicable
3032 * to the write-zeroes, we are cautious and limit the size to the
3033 * controllers max_hw_sectors value, which is based on the MDTS field
3034 * and possibly other limiting factors.
3036 if ((ctrl->oncs & NVME_CTRL_ONCS_WRITE_ZEROES) &&
3037 !(ctrl->quirks & NVME_QUIRK_DISABLE_WRITE_ZEROES))
3038 ctrl->max_zeroes_sectors = ctrl->max_hw_sectors;
3040 ctrl->max_zeroes_sectors = 0;
3042 if (ctrl->subsys->subtype != NVME_NQN_NVME ||
3043 nvme_ctrl_limited_cns(ctrl) ||
3044 test_bit(NVME_CTRL_SKIP_ID_CNS_CS, &ctrl->flags))
3047 id = kzalloc(sizeof(*id), GFP_KERNEL);
3051 c.identify.opcode = nvme_admin_identify;
3052 c.identify.cns = NVME_ID_CNS_CS_CTRL;
3053 c.identify.csi = NVME_CSI_NVM;
3055 ret = nvme_submit_sync_cmd(ctrl->admin_q, &c, id, sizeof(*id));
3059 ctrl->dmrl = id->dmrl;
3060 ctrl->dmrsl = le32_to_cpu(id->dmrsl);
3062 ctrl->max_zeroes_sectors = nvme_mps_to_sectors(ctrl, id->wzsl);
3066 set_bit(NVME_CTRL_SKIP_ID_CNS_CS, &ctrl->flags);
3071 static void nvme_init_known_nvm_effects(struct nvme_ctrl *ctrl)
3073 struct nvme_effects_log *log = ctrl->effects;
3075 log->acs[nvme_admin_format_nvm] |= cpu_to_le32(NVME_CMD_EFFECTS_LBCC |
3076 NVME_CMD_EFFECTS_NCC |
3077 NVME_CMD_EFFECTS_CSE_MASK);
3078 log->acs[nvme_admin_sanitize_nvm] |= cpu_to_le32(NVME_CMD_EFFECTS_LBCC |
3079 NVME_CMD_EFFECTS_CSE_MASK);
3082 * The spec says the result of a security receive command depends on
3083 * the previous security send command. As such, many vendors log this
3084 * command as one to submitted only when no other commands to the same
3085 * namespace are outstanding. The intention is to tell the host to
3086 * prevent mixing security send and receive.
3088 * This driver can only enforce such exclusive access against IO
3089 * queues, though. We are not readily able to enforce such a rule for
3090 * two commands to the admin queue, which is the only queue that
3091 * matters for this command.
3093 * Rather than blindly freezing the IO queues for this effect that
3094 * doesn't even apply to IO, mask it off.
3096 log->acs[nvme_admin_security_recv] &= cpu_to_le32(~NVME_CMD_EFFECTS_CSE_MASK);
3098 log->iocs[nvme_cmd_write] |= cpu_to_le32(NVME_CMD_EFFECTS_LBCC);
3099 log->iocs[nvme_cmd_write_zeroes] |= cpu_to_le32(NVME_CMD_EFFECTS_LBCC);
3100 log->iocs[nvme_cmd_write_uncor] |= cpu_to_le32(NVME_CMD_EFFECTS_LBCC);
3103 static int nvme_init_effects(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
3110 if (id->lpa & NVME_CTRL_LPA_CMD_EFFECTS_LOG) {
3111 ret = nvme_get_effects_log(ctrl, NVME_CSI_NVM, &ctrl->effects);
3116 if (!ctrl->effects) {
3117 ctrl->effects = kzalloc(sizeof(*ctrl->effects), GFP_KERNEL);
3120 xa_store(&ctrl->cels, NVME_CSI_NVM, ctrl->effects, GFP_KERNEL);
3123 nvme_init_known_nvm_effects(ctrl);
3127 static int nvme_check_ctrl_fabric_info(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
3130 * In fabrics we need to verify the cntlid matches the
3133 if (ctrl->cntlid != le16_to_cpu(id->cntlid)) {
3134 dev_err(ctrl->device,
3135 "Mismatching cntlid: Connect %u vs Identify %u, rejecting\n",
3136 ctrl->cntlid, le16_to_cpu(id->cntlid));
3140 if (!nvme_discovery_ctrl(ctrl) && !ctrl->kas) {
3141 dev_err(ctrl->device,
3142 "keep-alive support is mandatory for fabrics\n");
3146 if (!nvme_discovery_ctrl(ctrl) && ctrl->ioccsz < 4) {
3147 dev_err(ctrl->device,
3148 "I/O queue command capsule supported size %d < 4\n",
3153 if (!nvme_discovery_ctrl(ctrl) && ctrl->iorcsz < 1) {
3154 dev_err(ctrl->device,
3155 "I/O queue response capsule supported size %d < 1\n",
3160 if (!ctrl->maxcmd) {
3161 dev_err(ctrl->device, "Maximum outstanding commands is 0\n");
3168 static int nvme_init_identify(struct nvme_ctrl *ctrl)
3170 struct queue_limits lim;
3171 struct nvme_id_ctrl *id;
3173 bool prev_apst_enabled;
3176 ret = nvme_identify_ctrl(ctrl, &id);
3178 dev_err(ctrl->device, "Identify Controller failed (%d)\n", ret);
3182 if (!(ctrl->ops->flags & NVME_F_FABRICS))
3183 ctrl->cntlid = le16_to_cpu(id->cntlid);
3185 if (!ctrl->identified) {
3189 * Check for quirks. Quirk can depend on firmware version,
3190 * so, in principle, the set of quirks present can change
3191 * across a reset. As a possible future enhancement, we
3192 * could re-scan for quirks every time we reinitialize
3193 * the device, but we'd have to make sure that the driver
3194 * behaves intelligently if the quirks change.
3196 for (i = 0; i < ARRAY_SIZE(core_quirks); i++) {
3197 if (quirk_matches(id, &core_quirks[i]))
3198 ctrl->quirks |= core_quirks[i].quirks;
3201 ret = nvme_init_subsystem(ctrl, id);
3205 ret = nvme_init_effects(ctrl, id);
3209 memcpy(ctrl->subsys->firmware_rev, id->fr,
3210 sizeof(ctrl->subsys->firmware_rev));
3212 if (force_apst && (ctrl->quirks & NVME_QUIRK_NO_DEEPEST_PS)) {
3213 dev_warn(ctrl->device, "forcibly allowing all power states due to nvme_core.force_apst -- use at your own risk\n");
3214 ctrl->quirks &= ~NVME_QUIRK_NO_DEEPEST_PS;
3217 ctrl->crdt[0] = le16_to_cpu(id->crdt1);
3218 ctrl->crdt[1] = le16_to_cpu(id->crdt2);
3219 ctrl->crdt[2] = le16_to_cpu(id->crdt3);
3221 ctrl->oacs = le16_to_cpu(id->oacs);
3222 ctrl->oncs = le16_to_cpu(id->oncs);
3223 ctrl->mtfa = le16_to_cpu(id->mtfa);
3224 ctrl->oaes = le32_to_cpu(id->oaes);
3225 ctrl->wctemp = le16_to_cpu(id->wctemp);
3226 ctrl->cctemp = le16_to_cpu(id->cctemp);
3228 atomic_set(&ctrl->abort_limit, id->acl + 1);
3229 ctrl->vwc = id->vwc;
3231 max_hw_sectors = nvme_mps_to_sectors(ctrl, id->mdts);
3233 max_hw_sectors = UINT_MAX;
3234 ctrl->max_hw_sectors =
3235 min_not_zero(ctrl->max_hw_sectors, max_hw_sectors);
3237 lim = queue_limits_start_update(ctrl->admin_q);
3238 nvme_set_ctrl_limits(ctrl, &lim);
3239 ret = queue_limits_commit_update(ctrl->admin_q, &lim);
3243 ctrl->sgls = le32_to_cpu(id->sgls);
3244 ctrl->kas = le16_to_cpu(id->kas);
3245 ctrl->max_namespaces = le32_to_cpu(id->mnan);
3246 ctrl->ctratt = le32_to_cpu(id->ctratt);
3248 ctrl->cntrltype = id->cntrltype;
3249 ctrl->dctype = id->dctype;
3253 u32 transition_time = le32_to_cpu(id->rtd3e) / USEC_PER_SEC;
3255 ctrl->shutdown_timeout = clamp_t(unsigned int, transition_time,
3256 shutdown_timeout, 60);
3258 if (ctrl->shutdown_timeout != shutdown_timeout)
3259 dev_info(ctrl->device,
3260 "D3 entry latency set to %u seconds\n",
3261 ctrl->shutdown_timeout);
3263 ctrl->shutdown_timeout = shutdown_timeout;
3265 ctrl->npss = id->npss;
3266 ctrl->apsta = id->apsta;
3267 prev_apst_enabled = ctrl->apst_enabled;
3268 if (ctrl->quirks & NVME_QUIRK_NO_APST) {
3269 if (force_apst && id->apsta) {
3270 dev_warn(ctrl->device, "forcibly allowing APST due to nvme_core.force_apst -- use at your own risk\n");
3271 ctrl->apst_enabled = true;
3273 ctrl->apst_enabled = false;
3276 ctrl->apst_enabled = id->apsta;
3278 memcpy(ctrl->psd, id->psd, sizeof(ctrl->psd));
3280 if (ctrl->ops->flags & NVME_F_FABRICS) {
3281 ctrl->icdoff = le16_to_cpu(id->icdoff);
3282 ctrl->ioccsz = le32_to_cpu(id->ioccsz);
3283 ctrl->iorcsz = le32_to_cpu(id->iorcsz);
3284 ctrl->maxcmd = le16_to_cpu(id->maxcmd);
3286 ret = nvme_check_ctrl_fabric_info(ctrl, id);
3290 ctrl->hmpre = le32_to_cpu(id->hmpre);
3291 ctrl->hmmin = le32_to_cpu(id->hmmin);
3292 ctrl->hmminds = le32_to_cpu(id->hmminds);
3293 ctrl->hmmaxd = le16_to_cpu(id->hmmaxd);
3296 ret = nvme_mpath_init_identify(ctrl, id);
3300 if (ctrl->apst_enabled && !prev_apst_enabled)
3301 dev_pm_qos_expose_latency_tolerance(ctrl->device);
3302 else if (!ctrl->apst_enabled && prev_apst_enabled)
3303 dev_pm_qos_hide_latency_tolerance(ctrl->device);
3311 * Initialize the cached copies of the Identify data and various controller
3312 * register in our nvme_ctrl structure. This should be called as soon as
3313 * the admin queue is fully up and running.
3315 int nvme_init_ctrl_finish(struct nvme_ctrl *ctrl, bool was_suspended)
3319 ret = ctrl->ops->reg_read32(ctrl, NVME_REG_VS, &ctrl->vs);
3321 dev_err(ctrl->device, "Reading VS failed (%d)\n", ret);
3325 ctrl->sqsize = min_t(u16, NVME_CAP_MQES(ctrl->cap), ctrl->sqsize);
3327 if (ctrl->vs >= NVME_VS(1, 1, 0))
3328 ctrl->subsystem = NVME_CAP_NSSRC(ctrl->cap);
3330 ret = nvme_init_identify(ctrl);
3334 ret = nvme_configure_apst(ctrl);
3338 ret = nvme_configure_timestamp(ctrl);
3342 ret = nvme_configure_host_options(ctrl);
3346 nvme_configure_opal(ctrl, was_suspended);
3348 if (!ctrl->identified && !nvme_discovery_ctrl(ctrl)) {
3350 * Do not return errors unless we are in a controller reset,
3351 * the controller works perfectly fine without hwmon.
3353 ret = nvme_hwmon_init(ctrl);
3358 clear_bit(NVME_CTRL_DIRTY_CAPABILITY, &ctrl->flags);
3359 ctrl->identified = true;
3361 nvme_start_keep_alive(ctrl);
3365 EXPORT_SYMBOL_GPL(nvme_init_ctrl_finish);
3367 static int nvme_dev_open(struct inode *inode, struct file *file)
3369 struct nvme_ctrl *ctrl =
3370 container_of(inode->i_cdev, struct nvme_ctrl, cdev);
3372 switch (nvme_ctrl_state(ctrl)) {
3373 case NVME_CTRL_LIVE:
3376 return -EWOULDBLOCK;
3379 nvme_get_ctrl(ctrl);
3380 if (!try_module_get(ctrl->ops->module)) {
3381 nvme_put_ctrl(ctrl);
3385 file->private_data = ctrl;
3389 static int nvme_dev_release(struct inode *inode, struct file *file)
3391 struct nvme_ctrl *ctrl =
3392 container_of(inode->i_cdev, struct nvme_ctrl, cdev);
3394 module_put(ctrl->ops->module);
3395 nvme_put_ctrl(ctrl);
3399 static const struct file_operations nvme_dev_fops = {
3400 .owner = THIS_MODULE,
3401 .open = nvme_dev_open,
3402 .release = nvme_dev_release,
3403 .unlocked_ioctl = nvme_dev_ioctl,
3404 .compat_ioctl = compat_ptr_ioctl,
3405 .uring_cmd = nvme_dev_uring_cmd,
3408 static struct nvme_ns_head *nvme_find_ns_head(struct nvme_ctrl *ctrl,
3411 struct nvme_ns_head *h;
3413 lockdep_assert_held(&ctrl->subsys->lock);
3415 list_for_each_entry(h, &ctrl->subsys->nsheads, entry) {
3417 * Private namespaces can share NSIDs under some conditions.
3418 * In that case we can't use the same ns_head for namespaces
3419 * with the same NSID.
3421 if (h->ns_id != nsid || !nvme_is_unique_nsid(ctrl, h))
3423 if (!list_empty(&h->list) && nvme_tryget_ns_head(h))
3430 static int nvme_subsys_check_duplicate_ids(struct nvme_subsystem *subsys,
3431 struct nvme_ns_ids *ids)
3433 bool has_uuid = !uuid_is_null(&ids->uuid);
3434 bool has_nguid = memchr_inv(ids->nguid, 0, sizeof(ids->nguid));
3435 bool has_eui64 = memchr_inv(ids->eui64, 0, sizeof(ids->eui64));
3436 struct nvme_ns_head *h;
3438 lockdep_assert_held(&subsys->lock);
3440 list_for_each_entry(h, &subsys->nsheads, entry) {
3441 if (has_uuid && uuid_equal(&ids->uuid, &h->ids.uuid))
3444 memcmp(&ids->nguid, &h->ids.nguid, sizeof(ids->nguid)) == 0)
3447 memcmp(&ids->eui64, &h->ids.eui64, sizeof(ids->eui64)) == 0)
3454 static void nvme_cdev_rel(struct device *dev)
3456 ida_free(&nvme_ns_chr_minor_ida, MINOR(dev->devt));
3459 void nvme_cdev_del(struct cdev *cdev, struct device *cdev_device)
3461 cdev_device_del(cdev, cdev_device);
3462 put_device(cdev_device);
3465 int nvme_cdev_add(struct cdev *cdev, struct device *cdev_device,
3466 const struct file_operations *fops, struct module *owner)
3470 minor = ida_alloc(&nvme_ns_chr_minor_ida, GFP_KERNEL);
3473 cdev_device->devt = MKDEV(MAJOR(nvme_ns_chr_devt), minor);
3474 cdev_device->class = &nvme_ns_chr_class;
3475 cdev_device->release = nvme_cdev_rel;
3476 device_initialize(cdev_device);
3477 cdev_init(cdev, fops);
3478 cdev->owner = owner;
3479 ret = cdev_device_add(cdev, cdev_device);
3481 put_device(cdev_device);
3486 static int nvme_ns_chr_open(struct inode *inode, struct file *file)
3488 return nvme_ns_open(container_of(inode->i_cdev, struct nvme_ns, cdev));
3491 static int nvme_ns_chr_release(struct inode *inode, struct file *file)
3493 nvme_ns_release(container_of(inode->i_cdev, struct nvme_ns, cdev));
3497 static const struct file_operations nvme_ns_chr_fops = {
3498 .owner = THIS_MODULE,
3499 .open = nvme_ns_chr_open,
3500 .release = nvme_ns_chr_release,
3501 .unlocked_ioctl = nvme_ns_chr_ioctl,
3502 .compat_ioctl = compat_ptr_ioctl,
3503 .uring_cmd = nvme_ns_chr_uring_cmd,
3504 .uring_cmd_iopoll = nvme_ns_chr_uring_cmd_iopoll,
3507 static int nvme_add_ns_cdev(struct nvme_ns *ns)
3511 ns->cdev_device.parent = ns->ctrl->device;
3512 ret = dev_set_name(&ns->cdev_device, "ng%dn%d",
3513 ns->ctrl->instance, ns->head->instance);
3517 return nvme_cdev_add(&ns->cdev, &ns->cdev_device, &nvme_ns_chr_fops,
3518 ns->ctrl->ops->module);
3521 static struct nvme_ns_head *nvme_alloc_ns_head(struct nvme_ctrl *ctrl,
3522 struct nvme_ns_info *info)
3524 struct nvme_ns_head *head;
3525 size_t size = sizeof(*head);
3528 #ifdef CONFIG_NVME_MULTIPATH
3529 size += num_possible_nodes() * sizeof(struct nvme_ns *);
3532 head = kzalloc(size, GFP_KERNEL);
3535 ret = ida_alloc_min(&ctrl->subsys->ns_ida, 1, GFP_KERNEL);
3538 head->instance = ret;
3539 INIT_LIST_HEAD(&head->list);
3540 ret = init_srcu_struct(&head->srcu);
3542 goto out_ida_remove;
3543 head->subsys = ctrl->subsys;
3544 head->ns_id = info->nsid;
3545 head->ids = info->ids;
3546 head->shared = info->is_shared;
3547 ratelimit_state_init(&head->rs_nuse, 5 * HZ, 1);
3548 ratelimit_set_flags(&head->rs_nuse, RATELIMIT_MSG_ON_RELEASE);
3549 kref_init(&head->ref);
3551 if (head->ids.csi) {
3552 ret = nvme_get_effects_log(ctrl, head->ids.csi, &head->effects);
3554 goto out_cleanup_srcu;
3556 head->effects = ctrl->effects;
3558 ret = nvme_mpath_alloc_disk(ctrl, head);
3560 goto out_cleanup_srcu;
3562 list_add_tail(&head->entry, &ctrl->subsys->nsheads);
3564 kref_get(&ctrl->subsys->ref);
3568 cleanup_srcu_struct(&head->srcu);
3570 ida_free(&ctrl->subsys->ns_ida, head->instance);
3575 ret = blk_status_to_errno(nvme_error_status(ret));
3576 return ERR_PTR(ret);
3579 static int nvme_global_check_duplicate_ids(struct nvme_subsystem *this,
3580 struct nvme_ns_ids *ids)
3582 struct nvme_subsystem *s;
3586 * Note that this check is racy as we try to avoid holding the global
3587 * lock over the whole ns_head creation. But it is only intended as
3588 * a sanity check anyway.
3590 mutex_lock(&nvme_subsystems_lock);
3591 list_for_each_entry(s, &nvme_subsystems, entry) {
3594 mutex_lock(&s->lock);
3595 ret = nvme_subsys_check_duplicate_ids(s, ids);
3596 mutex_unlock(&s->lock);
3600 mutex_unlock(&nvme_subsystems_lock);
3605 static int nvme_init_ns_head(struct nvme_ns *ns, struct nvme_ns_info *info)
3607 struct nvme_ctrl *ctrl = ns->ctrl;
3608 struct nvme_ns_head *head = NULL;
3611 ret = nvme_global_check_duplicate_ids(ctrl->subsys, &info->ids);
3614 * We've found two different namespaces on two different
3615 * subsystems that report the same ID. This is pretty nasty
3616 * for anything that actually requires unique device
3617 * identification. In the kernel we need this for multipathing,
3618 * and in user space the /dev/disk/by-id/ links rely on it.
3620 * If the device also claims to be multi-path capable back off
3621 * here now and refuse the probe the second device as this is a
3622 * recipe for data corruption. If not this is probably a
3623 * cheap consumer device if on the PCIe bus, so let the user
3624 * proceed and use the shiny toy, but warn that with changing
3625 * probing order (which due to our async probing could just be
3626 * device taking longer to startup) the other device could show
3629 nvme_print_device_info(ctrl);
3630 if ((ns->ctrl->ops->flags & NVME_F_FABRICS) || /* !PCIe */
3631 ((ns->ctrl->subsys->cmic & NVME_CTRL_CMIC_MULTI_CTRL) &&
3633 dev_err(ctrl->device,
3634 "ignoring nsid %d because of duplicate IDs\n",
3639 dev_err(ctrl->device,
3640 "clearing duplicate IDs for nsid %d\n", info->nsid);
3641 dev_err(ctrl->device,
3642 "use of /dev/disk/by-id/ may cause data corruption\n");
3643 memset(&info->ids.nguid, 0, sizeof(info->ids.nguid));
3644 memset(&info->ids.uuid, 0, sizeof(info->ids.uuid));
3645 memset(&info->ids.eui64, 0, sizeof(info->ids.eui64));
3646 ctrl->quirks |= NVME_QUIRK_BOGUS_NID;
3649 mutex_lock(&ctrl->subsys->lock);
3650 head = nvme_find_ns_head(ctrl, info->nsid);
3652 ret = nvme_subsys_check_duplicate_ids(ctrl->subsys, &info->ids);
3654 dev_err(ctrl->device,
3655 "duplicate IDs in subsystem for nsid %d\n",
3659 head = nvme_alloc_ns_head(ctrl, info);
3661 ret = PTR_ERR(head);
3666 if (!info->is_shared || !head->shared) {
3667 dev_err(ctrl->device,
3668 "Duplicate unshared namespace %d\n",
3670 goto out_put_ns_head;
3672 if (!nvme_ns_ids_equal(&head->ids, &info->ids)) {
3673 dev_err(ctrl->device,
3674 "IDs don't match for shared namespace %d\n",
3676 goto out_put_ns_head;
3680 dev_warn(ctrl->device,
3681 "Found shared namespace %d, but multipathing not supported.\n",
3683 dev_warn_once(ctrl->device,
3684 "Support for shared namespaces without CONFIG_NVME_MULTIPATH is deprecated and will be removed in Linux 6.0\n.");
3688 list_add_tail_rcu(&ns->siblings, &head->list);
3690 mutex_unlock(&ctrl->subsys->lock);
3694 nvme_put_ns_head(head);
3696 mutex_unlock(&ctrl->subsys->lock);
3700 struct nvme_ns *nvme_find_get_ns(struct nvme_ctrl *ctrl, unsigned nsid)
3702 struct nvme_ns *ns, *ret = NULL;
3704 down_read(&ctrl->namespaces_rwsem);
3705 list_for_each_entry(ns, &ctrl->namespaces, list) {
3706 if (ns->head->ns_id == nsid) {
3707 if (!nvme_get_ns(ns))
3712 if (ns->head->ns_id > nsid)
3715 up_read(&ctrl->namespaces_rwsem);
3718 EXPORT_SYMBOL_NS_GPL(nvme_find_get_ns, NVME_TARGET_PASSTHRU);
3721 * Add the namespace to the controller list while keeping the list ordered.
3723 static void nvme_ns_add_to_ctrl_list(struct nvme_ns *ns)
3725 struct nvme_ns *tmp;
3727 list_for_each_entry_reverse(tmp, &ns->ctrl->namespaces, list) {
3728 if (tmp->head->ns_id < ns->head->ns_id) {
3729 list_add(&ns->list, &tmp->list);
3733 list_add(&ns->list, &ns->ctrl->namespaces);
3736 static void nvme_alloc_ns(struct nvme_ctrl *ctrl, struct nvme_ns_info *info)
3739 struct gendisk *disk;
3740 int node = ctrl->numa_node;
3742 ns = kzalloc_node(sizeof(*ns), GFP_KERNEL, node);
3746 disk = blk_mq_alloc_disk(ctrl->tagset, NULL, ns);
3749 disk->fops = &nvme_bdev_ops;
3750 disk->private_data = ns;
3753 ns->queue = disk->queue;
3755 if (ctrl->opts && ctrl->opts->data_digest)
3756 blk_queue_flag_set(QUEUE_FLAG_STABLE_WRITES, ns->queue);
3758 blk_queue_flag_set(QUEUE_FLAG_NONROT, ns->queue);
3759 if (ctrl->ops->supports_pci_p2pdma &&
3760 ctrl->ops->supports_pci_p2pdma(ctrl))
3761 blk_queue_flag_set(QUEUE_FLAG_PCI_P2PDMA, ns->queue);
3764 kref_init(&ns->kref);
3766 if (nvme_init_ns_head(ns, info))
3767 goto out_cleanup_disk;
3770 * If multipathing is enabled, the device name for all disks and not
3771 * just those that represent shared namespaces needs to be based on the
3772 * subsystem instance. Using the controller instance for private
3773 * namespaces could lead to naming collisions between shared and private
3774 * namespaces if they don't use a common numbering scheme.
3776 * If multipathing is not enabled, disk names must use the controller
3777 * instance as shared namespaces will show up as multiple block
3780 if (nvme_ns_head_multipath(ns->head)) {
3781 sprintf(disk->disk_name, "nvme%dc%dn%d", ctrl->subsys->instance,
3782 ctrl->instance, ns->head->instance);
3783 disk->flags |= GENHD_FL_HIDDEN;
3784 } else if (multipath) {
3785 sprintf(disk->disk_name, "nvme%dn%d", ctrl->subsys->instance,
3786 ns->head->instance);
3788 sprintf(disk->disk_name, "nvme%dn%d", ctrl->instance,
3789 ns->head->instance);
3792 if (nvme_update_ns_info(ns, info))
3795 down_write(&ctrl->namespaces_rwsem);
3797 * Ensure that no namespaces are added to the ctrl list after the queues
3798 * are frozen, thereby avoiding a deadlock between scan and reset.
3800 if (test_bit(NVME_CTRL_FROZEN, &ctrl->flags)) {
3801 up_write(&ctrl->namespaces_rwsem);
3804 nvme_ns_add_to_ctrl_list(ns);
3805 up_write(&ctrl->namespaces_rwsem);
3806 nvme_get_ctrl(ctrl);
3808 if (device_add_disk(ctrl->device, ns->disk, nvme_ns_attr_groups))
3809 goto out_cleanup_ns_from_list;
3811 if (!nvme_ns_head_multipath(ns->head))
3812 nvme_add_ns_cdev(ns);
3814 nvme_mpath_add_disk(ns, info->anagrpid);
3815 nvme_fault_inject_init(&ns->fault_inject, ns->disk->disk_name);
3818 * Set ns->disk->device->driver_data to ns so we can access
3819 * ns->head->passthru_err_log_enabled in
3820 * nvme_io_passthru_err_log_enabled_[store | show]().
3822 dev_set_drvdata(disk_to_dev(ns->disk), ns);
3826 out_cleanup_ns_from_list:
3827 nvme_put_ctrl(ctrl);
3828 down_write(&ctrl->namespaces_rwsem);
3829 list_del_init(&ns->list);
3830 up_write(&ctrl->namespaces_rwsem);
3832 mutex_lock(&ctrl->subsys->lock);
3833 list_del_rcu(&ns->siblings);
3834 if (list_empty(&ns->head->list))
3835 list_del_init(&ns->head->entry);
3836 mutex_unlock(&ctrl->subsys->lock);
3837 nvme_put_ns_head(ns->head);
3844 static void nvme_ns_remove(struct nvme_ns *ns)
3846 bool last_path = false;
3848 if (test_and_set_bit(NVME_NS_REMOVING, &ns->flags))
3851 clear_bit(NVME_NS_READY, &ns->flags);
3852 set_capacity(ns->disk, 0);
3853 nvme_fault_inject_fini(&ns->fault_inject);
3856 * Ensure that !NVME_NS_READY is seen by other threads to prevent
3857 * this ns going back into current_path.
3859 synchronize_srcu(&ns->head->srcu);
3861 /* wait for concurrent submissions */
3862 if (nvme_mpath_clear_current_path(ns))
3863 synchronize_srcu(&ns->head->srcu);
3865 mutex_lock(&ns->ctrl->subsys->lock);
3866 list_del_rcu(&ns->siblings);
3867 if (list_empty(&ns->head->list)) {
3868 list_del_init(&ns->head->entry);
3871 mutex_unlock(&ns->ctrl->subsys->lock);
3873 /* guarantee not available in head->list */
3874 synchronize_srcu(&ns->head->srcu);
3876 if (!nvme_ns_head_multipath(ns->head))
3877 nvme_cdev_del(&ns->cdev, &ns->cdev_device);
3878 del_gendisk(ns->disk);
3880 down_write(&ns->ctrl->namespaces_rwsem);
3881 list_del_init(&ns->list);
3882 up_write(&ns->ctrl->namespaces_rwsem);
3885 nvme_mpath_shutdown_disk(ns->head);
3889 static void nvme_ns_remove_by_nsid(struct nvme_ctrl *ctrl, u32 nsid)
3891 struct nvme_ns *ns = nvme_find_get_ns(ctrl, nsid);
3899 static void nvme_validate_ns(struct nvme_ns *ns, struct nvme_ns_info *info)
3901 int ret = NVME_SC_INVALID_NS | NVME_SC_DNR;
3903 if (!nvme_ns_ids_equal(&ns->head->ids, &info->ids)) {
3904 dev_err(ns->ctrl->device,
3905 "identifiers changed for nsid %d\n", ns->head->ns_id);
3909 ret = nvme_update_ns_info(ns, info);
3912 * Only remove the namespace if we got a fatal error back from the
3913 * device, otherwise ignore the error and just move on.
3915 * TODO: we should probably schedule a delayed retry here.
3917 if (ret > 0 && (ret & NVME_SC_DNR))
3921 static void nvme_scan_ns(struct nvme_ctrl *ctrl, unsigned nsid)
3923 struct nvme_ns_info info = { .nsid = nsid };
3927 if (nvme_identify_ns_descs(ctrl, &info))
3930 if (info.ids.csi != NVME_CSI_NVM && !nvme_multi_css(ctrl)) {
3931 dev_warn(ctrl->device,
3932 "command set not reported for nsid: %d\n", nsid);
3937 * If available try to use the Command Set Idependent Identify Namespace
3938 * data structure to find all the generic information that is needed to
3939 * set up a namespace. If not fall back to the legacy version.
3941 if ((ctrl->cap & NVME_CAP_CRMS_CRIMS) ||
3942 (info.ids.csi != NVME_CSI_NVM && info.ids.csi != NVME_CSI_ZNS))
3943 ret = nvme_ns_info_from_id_cs_indep(ctrl, &info);
3945 ret = nvme_ns_info_from_identify(ctrl, &info);
3947 if (info.is_removed)
3948 nvme_ns_remove_by_nsid(ctrl, nsid);
3951 * Ignore the namespace if it is not ready. We will get an AEN once it
3952 * becomes ready and restart the scan.
3954 if (ret || !info.is_ready)
3957 ns = nvme_find_get_ns(ctrl, nsid);
3959 nvme_validate_ns(ns, &info);
3962 nvme_alloc_ns(ctrl, &info);
3966 static void nvme_remove_invalid_namespaces(struct nvme_ctrl *ctrl,
3969 struct nvme_ns *ns, *next;
3972 down_write(&ctrl->namespaces_rwsem);
3973 list_for_each_entry_safe(ns, next, &ctrl->namespaces, list) {
3974 if (ns->head->ns_id > nsid)
3975 list_move_tail(&ns->list, &rm_list);
3977 up_write(&ctrl->namespaces_rwsem);
3979 list_for_each_entry_safe(ns, next, &rm_list, list)
3984 static int nvme_scan_ns_list(struct nvme_ctrl *ctrl)
3986 const int nr_entries = NVME_IDENTIFY_DATA_SIZE / sizeof(__le32);
3991 ns_list = kzalloc(NVME_IDENTIFY_DATA_SIZE, GFP_KERNEL);
3996 struct nvme_command cmd = {
3997 .identify.opcode = nvme_admin_identify,
3998 .identify.cns = NVME_ID_CNS_NS_ACTIVE_LIST,
3999 .identify.nsid = cpu_to_le32(prev),
4002 ret = nvme_submit_sync_cmd(ctrl->admin_q, &cmd, ns_list,
4003 NVME_IDENTIFY_DATA_SIZE);
4005 dev_warn(ctrl->device,
4006 "Identify NS List failed (status=0x%x)\n", ret);
4010 for (i = 0; i < nr_entries; i++) {
4011 u32 nsid = le32_to_cpu(ns_list[i]);
4013 if (!nsid) /* end of the list? */
4015 nvme_scan_ns(ctrl, nsid);
4016 while (++prev < nsid)
4017 nvme_ns_remove_by_nsid(ctrl, prev);
4021 nvme_remove_invalid_namespaces(ctrl, prev);
4027 static void nvme_scan_ns_sequential(struct nvme_ctrl *ctrl)
4029 struct nvme_id_ctrl *id;
4032 if (nvme_identify_ctrl(ctrl, &id))
4034 nn = le32_to_cpu(id->nn);
4037 for (i = 1; i <= nn; i++)
4038 nvme_scan_ns(ctrl, i);
4040 nvme_remove_invalid_namespaces(ctrl, nn);
4043 static void nvme_clear_changed_ns_log(struct nvme_ctrl *ctrl)
4045 size_t log_size = NVME_MAX_CHANGED_NAMESPACES * sizeof(__le32);
4049 log = kzalloc(log_size, GFP_KERNEL);
4054 * We need to read the log to clear the AEN, but we don't want to rely
4055 * on it for the changed namespace information as userspace could have
4056 * raced with us in reading the log page, which could cause us to miss
4059 error = nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_CHANGED_NS, 0,
4060 NVME_CSI_NVM, log, log_size, 0);
4062 dev_warn(ctrl->device,
4063 "reading changed ns log failed: %d\n", error);
4068 static void nvme_scan_work(struct work_struct *work)
4070 struct nvme_ctrl *ctrl =
4071 container_of(work, struct nvme_ctrl, scan_work);
4074 /* No tagset on a live ctrl means IO queues could not created */
4075 if (nvme_ctrl_state(ctrl) != NVME_CTRL_LIVE || !ctrl->tagset)
4079 * Identify controller limits can change at controller reset due to
4080 * new firmware download, even though it is not common we cannot ignore
4081 * such scenario. Controller's non-mdts limits are reported in the unit
4082 * of logical blocks that is dependent on the format of attached
4083 * namespace. Hence re-read the limits at the time of ns allocation.
4085 ret = nvme_init_non_mdts_limits(ctrl);
4087 dev_warn(ctrl->device,
4088 "reading non-mdts-limits failed: %d\n", ret);
4092 if (test_and_clear_bit(NVME_AER_NOTICE_NS_CHANGED, &ctrl->events)) {
4093 dev_info(ctrl->device, "rescanning namespaces.\n");
4094 nvme_clear_changed_ns_log(ctrl);
4097 mutex_lock(&ctrl->scan_lock);
4098 if (nvme_ctrl_limited_cns(ctrl)) {
4099 nvme_scan_ns_sequential(ctrl);
4102 * Fall back to sequential scan if DNR is set to handle broken
4103 * devices which should support Identify NS List (as per the VS
4104 * they report) but don't actually support it.
4106 ret = nvme_scan_ns_list(ctrl);
4107 if (ret > 0 && ret & NVME_SC_DNR)
4108 nvme_scan_ns_sequential(ctrl);
4110 mutex_unlock(&ctrl->scan_lock);
4114 * This function iterates the namespace list unlocked to allow recovery from
4115 * controller failure. It is up to the caller to ensure the namespace list is
4116 * not modified by scan work while this function is executing.
4118 void nvme_remove_namespaces(struct nvme_ctrl *ctrl)
4120 struct nvme_ns *ns, *next;
4124 * make sure to requeue I/O to all namespaces as these
4125 * might result from the scan itself and must complete
4126 * for the scan_work to make progress
4128 nvme_mpath_clear_ctrl_paths(ctrl);
4131 * Unquiesce io queues so any pending IO won't hang, especially
4132 * those submitted from scan work
4134 nvme_unquiesce_io_queues(ctrl);
4136 /* prevent racing with ns scanning */
4137 flush_work(&ctrl->scan_work);
4140 * The dead states indicates the controller was not gracefully
4141 * disconnected. In that case, we won't be able to flush any data while
4142 * removing the namespaces' disks; fail all the queues now to avoid
4143 * potentially having to clean up the failed sync later.
4145 if (nvme_ctrl_state(ctrl) == NVME_CTRL_DEAD)
4146 nvme_mark_namespaces_dead(ctrl);
4148 /* this is a no-op when called from the controller reset handler */
4149 nvme_change_ctrl_state(ctrl, NVME_CTRL_DELETING_NOIO);
4151 down_write(&ctrl->namespaces_rwsem);
4152 list_splice_init(&ctrl->namespaces, &ns_list);
4153 up_write(&ctrl->namespaces_rwsem);
4155 list_for_each_entry_safe(ns, next, &ns_list, list)
4158 EXPORT_SYMBOL_GPL(nvme_remove_namespaces);
4160 static int nvme_class_uevent(const struct device *dev, struct kobj_uevent_env *env)
4162 const struct nvme_ctrl *ctrl =
4163 container_of(dev, struct nvme_ctrl, ctrl_device);
4164 struct nvmf_ctrl_options *opts = ctrl->opts;
4167 ret = add_uevent_var(env, "NVME_TRTYPE=%s", ctrl->ops->name);
4172 ret = add_uevent_var(env, "NVME_TRADDR=%s", opts->traddr);
4176 ret = add_uevent_var(env, "NVME_TRSVCID=%s",
4177 opts->trsvcid ?: "none");
4181 ret = add_uevent_var(env, "NVME_HOST_TRADDR=%s",
4182 opts->host_traddr ?: "none");
4186 ret = add_uevent_var(env, "NVME_HOST_IFACE=%s",
4187 opts->host_iface ?: "none");
4192 static void nvme_change_uevent(struct nvme_ctrl *ctrl, char *envdata)
4194 char *envp[2] = { envdata, NULL };
4196 kobject_uevent_env(&ctrl->device->kobj, KOBJ_CHANGE, envp);
4199 static void nvme_aen_uevent(struct nvme_ctrl *ctrl)
4201 char *envp[2] = { NULL, NULL };
4202 u32 aen_result = ctrl->aen_result;
4204 ctrl->aen_result = 0;
4208 envp[0] = kasprintf(GFP_KERNEL, "NVME_AEN=%#08x", aen_result);
4211 kobject_uevent_env(&ctrl->device->kobj, KOBJ_CHANGE, envp);
4215 static void nvme_async_event_work(struct work_struct *work)
4217 struct nvme_ctrl *ctrl =
4218 container_of(work, struct nvme_ctrl, async_event_work);
4220 nvme_aen_uevent(ctrl);
4223 * The transport drivers must guarantee AER submission here is safe by
4224 * flushing ctrl async_event_work after changing the controller state
4225 * from LIVE and before freeing the admin queue.
4227 if (nvme_ctrl_state(ctrl) == NVME_CTRL_LIVE)
4228 ctrl->ops->submit_async_event(ctrl);
4231 static bool nvme_ctrl_pp_status(struct nvme_ctrl *ctrl)
4236 if (ctrl->ops->reg_read32(ctrl, NVME_REG_CSTS, &csts))
4242 return ((ctrl->ctrl_config & NVME_CC_ENABLE) && (csts & NVME_CSTS_PP));
4245 static void nvme_get_fw_slot_info(struct nvme_ctrl *ctrl)
4247 struct nvme_fw_slot_info_log *log;
4248 u8 next_fw_slot, cur_fw_slot;
4250 log = kmalloc(sizeof(*log), GFP_KERNEL);
4254 if (nvme_get_log(ctrl, NVME_NSID_ALL, NVME_LOG_FW_SLOT, 0, NVME_CSI_NVM,
4255 log, sizeof(*log), 0)) {
4256 dev_warn(ctrl->device, "Get FW SLOT INFO log error\n");
4260 cur_fw_slot = log->afi & 0x7;
4261 next_fw_slot = (log->afi & 0x70) >> 4;
4262 if (!cur_fw_slot || (next_fw_slot && (cur_fw_slot != next_fw_slot))) {
4263 dev_info(ctrl->device,
4264 "Firmware is activated after next Controller Level Reset\n");
4268 memcpy(ctrl->subsys->firmware_rev, &log->frs[cur_fw_slot - 1],
4269 sizeof(ctrl->subsys->firmware_rev));
4275 static void nvme_fw_act_work(struct work_struct *work)
4277 struct nvme_ctrl *ctrl = container_of(work,
4278 struct nvme_ctrl, fw_act_work);
4279 unsigned long fw_act_timeout;
4281 nvme_auth_stop(ctrl);
4284 fw_act_timeout = jiffies +
4285 msecs_to_jiffies(ctrl->mtfa * 100);
4287 fw_act_timeout = jiffies +
4288 msecs_to_jiffies(admin_timeout * 1000);
4290 nvme_quiesce_io_queues(ctrl);
4291 while (nvme_ctrl_pp_status(ctrl)) {
4292 if (time_after(jiffies, fw_act_timeout)) {
4293 dev_warn(ctrl->device,
4294 "Fw activation timeout, reset controller\n");
4295 nvme_try_sched_reset(ctrl);
4301 if (!nvme_change_ctrl_state(ctrl, NVME_CTRL_LIVE))
4304 nvme_unquiesce_io_queues(ctrl);
4305 /* read FW slot information to clear the AER */
4306 nvme_get_fw_slot_info(ctrl);
4308 queue_work(nvme_wq, &ctrl->async_event_work);
4311 static u32 nvme_aer_type(u32 result)
4313 return result & 0x7;
4316 static u32 nvme_aer_subtype(u32 result)
4318 return (result & 0xff00) >> 8;
4321 static bool nvme_handle_aen_notice(struct nvme_ctrl *ctrl, u32 result)
4323 u32 aer_notice_type = nvme_aer_subtype(result);
4324 bool requeue = true;
4326 switch (aer_notice_type) {
4327 case NVME_AER_NOTICE_NS_CHANGED:
4328 set_bit(NVME_AER_NOTICE_NS_CHANGED, &ctrl->events);
4329 nvme_queue_scan(ctrl);
4331 case NVME_AER_NOTICE_FW_ACT_STARTING:
4333 * We are (ab)using the RESETTING state to prevent subsequent
4334 * recovery actions from interfering with the controller's
4335 * firmware activation.
4337 if (nvme_change_ctrl_state(ctrl, NVME_CTRL_RESETTING)) {
4339 queue_work(nvme_wq, &ctrl->fw_act_work);
4342 #ifdef CONFIG_NVME_MULTIPATH
4343 case NVME_AER_NOTICE_ANA:
4344 if (!ctrl->ana_log_buf)
4346 queue_work(nvme_wq, &ctrl->ana_work);
4349 case NVME_AER_NOTICE_DISC_CHANGED:
4350 ctrl->aen_result = result;
4353 dev_warn(ctrl->device, "async event result %08x\n", result);
4358 static void nvme_handle_aer_persistent_error(struct nvme_ctrl *ctrl)
4360 dev_warn(ctrl->device, "resetting controller due to AER\n");
4361 nvme_reset_ctrl(ctrl);
4364 void nvme_complete_async_event(struct nvme_ctrl *ctrl, __le16 status,
4365 volatile union nvme_result *res)
4367 u32 result = le32_to_cpu(res->u32);
4368 u32 aer_type = nvme_aer_type(result);
4369 u32 aer_subtype = nvme_aer_subtype(result);
4370 bool requeue = true;
4372 if (le16_to_cpu(status) >> 1 != NVME_SC_SUCCESS)
4375 trace_nvme_async_event(ctrl, result);
4377 case NVME_AER_NOTICE:
4378 requeue = nvme_handle_aen_notice(ctrl, result);
4380 case NVME_AER_ERROR:
4382 * For a persistent internal error, don't run async_event_work
4383 * to submit a new AER. The controller reset will do it.
4385 if (aer_subtype == NVME_AER_ERROR_PERSIST_INT_ERR) {
4386 nvme_handle_aer_persistent_error(ctrl);
4390 case NVME_AER_SMART:
4393 ctrl->aen_result = result;
4400 queue_work(nvme_wq, &ctrl->async_event_work);
4402 EXPORT_SYMBOL_GPL(nvme_complete_async_event);
4404 int nvme_alloc_admin_tag_set(struct nvme_ctrl *ctrl, struct blk_mq_tag_set *set,
4405 const struct blk_mq_ops *ops, unsigned int cmd_size)
4407 struct queue_limits lim = {};
4410 memset(set, 0, sizeof(*set));
4412 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
4413 if (ctrl->ops->flags & NVME_F_FABRICS)
4414 /* Reserved for fabric connect and keep alive */
4415 set->reserved_tags = 2;
4416 set->numa_node = ctrl->numa_node;
4417 set->flags = BLK_MQ_F_NO_SCHED;
4418 if (ctrl->ops->flags & NVME_F_BLOCKING)
4419 set->flags |= BLK_MQ_F_BLOCKING;
4420 set->cmd_size = cmd_size;
4421 set->driver_data = ctrl;
4422 set->nr_hw_queues = 1;
4423 set->timeout = NVME_ADMIN_TIMEOUT;
4424 ret = blk_mq_alloc_tag_set(set);
4428 ctrl->admin_q = blk_mq_alloc_queue(set, &lim, NULL);
4429 if (IS_ERR(ctrl->admin_q)) {
4430 ret = PTR_ERR(ctrl->admin_q);
4431 goto out_free_tagset;
4434 if (ctrl->ops->flags & NVME_F_FABRICS) {
4435 ctrl->fabrics_q = blk_mq_alloc_queue(set, NULL, NULL);
4436 if (IS_ERR(ctrl->fabrics_q)) {
4437 ret = PTR_ERR(ctrl->fabrics_q);
4438 goto out_cleanup_admin_q;
4442 ctrl->admin_tagset = set;
4445 out_cleanup_admin_q:
4446 blk_mq_destroy_queue(ctrl->admin_q);
4447 blk_put_queue(ctrl->admin_q);
4449 blk_mq_free_tag_set(set);
4450 ctrl->admin_q = NULL;
4451 ctrl->fabrics_q = NULL;
4454 EXPORT_SYMBOL_GPL(nvme_alloc_admin_tag_set);
4456 void nvme_remove_admin_tag_set(struct nvme_ctrl *ctrl)
4458 blk_mq_destroy_queue(ctrl->admin_q);
4459 blk_put_queue(ctrl->admin_q);
4460 if (ctrl->ops->flags & NVME_F_FABRICS) {
4461 blk_mq_destroy_queue(ctrl->fabrics_q);
4462 blk_put_queue(ctrl->fabrics_q);
4464 blk_mq_free_tag_set(ctrl->admin_tagset);
4466 EXPORT_SYMBOL_GPL(nvme_remove_admin_tag_set);
4468 int nvme_alloc_io_tag_set(struct nvme_ctrl *ctrl, struct blk_mq_tag_set *set,
4469 const struct blk_mq_ops *ops, unsigned int nr_maps,
4470 unsigned int cmd_size)
4474 memset(set, 0, sizeof(*set));
4476 set->queue_depth = min_t(unsigned, ctrl->sqsize, BLK_MQ_MAX_DEPTH - 1);
4478 * Some Apple controllers requires tags to be unique across admin and
4479 * the (only) I/O queue, so reserve the first 32 tags of the I/O queue.
4481 if (ctrl->quirks & NVME_QUIRK_SHARED_TAGS)
4482 set->reserved_tags = NVME_AQ_DEPTH;
4483 else if (ctrl->ops->flags & NVME_F_FABRICS)
4484 /* Reserved for fabric connect */
4485 set->reserved_tags = 1;
4486 set->numa_node = ctrl->numa_node;
4487 set->flags = BLK_MQ_F_SHOULD_MERGE;
4488 if (ctrl->ops->flags & NVME_F_BLOCKING)
4489 set->flags |= BLK_MQ_F_BLOCKING;
4490 set->cmd_size = cmd_size,
4491 set->driver_data = ctrl;
4492 set->nr_hw_queues = ctrl->queue_count - 1;
4493 set->timeout = NVME_IO_TIMEOUT;
4494 set->nr_maps = nr_maps;
4495 ret = blk_mq_alloc_tag_set(set);
4499 if (ctrl->ops->flags & NVME_F_FABRICS) {
4500 ctrl->connect_q = blk_mq_alloc_queue(set, NULL, NULL);
4501 if (IS_ERR(ctrl->connect_q)) {
4502 ret = PTR_ERR(ctrl->connect_q);
4503 goto out_free_tag_set;
4505 blk_queue_flag_set(QUEUE_FLAG_SKIP_TAGSET_QUIESCE,
4513 blk_mq_free_tag_set(set);
4514 ctrl->connect_q = NULL;
4517 EXPORT_SYMBOL_GPL(nvme_alloc_io_tag_set);
4519 void nvme_remove_io_tag_set(struct nvme_ctrl *ctrl)
4521 if (ctrl->ops->flags & NVME_F_FABRICS) {
4522 blk_mq_destroy_queue(ctrl->connect_q);
4523 blk_put_queue(ctrl->connect_q);
4525 blk_mq_free_tag_set(ctrl->tagset);
4527 EXPORT_SYMBOL_GPL(nvme_remove_io_tag_set);
4529 void nvme_stop_ctrl(struct nvme_ctrl *ctrl)
4531 nvme_mpath_stop(ctrl);
4532 nvme_auth_stop(ctrl);
4533 nvme_stop_keep_alive(ctrl);
4534 nvme_stop_failfast_work(ctrl);
4535 flush_work(&ctrl->async_event_work);
4536 cancel_work_sync(&ctrl->fw_act_work);
4537 if (ctrl->ops->stop_ctrl)
4538 ctrl->ops->stop_ctrl(ctrl);
4540 EXPORT_SYMBOL_GPL(nvme_stop_ctrl);
4542 void nvme_start_ctrl(struct nvme_ctrl *ctrl)
4544 nvme_enable_aen(ctrl);
4547 * persistent discovery controllers need to send indication to userspace
4548 * to re-read the discovery log page to learn about possible changes
4549 * that were missed. We identify persistent discovery controllers by
4550 * checking that they started once before, hence are reconnecting back.
4552 if (test_bit(NVME_CTRL_STARTED_ONCE, &ctrl->flags) &&
4553 nvme_discovery_ctrl(ctrl))
4554 nvme_change_uevent(ctrl, "NVME_EVENT=rediscover");
4556 if (ctrl->queue_count > 1) {
4557 nvme_queue_scan(ctrl);
4558 nvme_unquiesce_io_queues(ctrl);
4559 nvme_mpath_update(ctrl);
4562 nvme_change_uevent(ctrl, "NVME_EVENT=connected");
4563 set_bit(NVME_CTRL_STARTED_ONCE, &ctrl->flags);
4565 EXPORT_SYMBOL_GPL(nvme_start_ctrl);
4567 void nvme_uninit_ctrl(struct nvme_ctrl *ctrl)
4569 nvme_hwmon_exit(ctrl);
4570 nvme_fault_inject_fini(&ctrl->fault_inject);
4571 dev_pm_qos_hide_latency_tolerance(ctrl->device);
4572 cdev_device_del(&ctrl->cdev, ctrl->device);
4573 nvme_put_ctrl(ctrl);
4575 EXPORT_SYMBOL_GPL(nvme_uninit_ctrl);
4577 static void nvme_free_cels(struct nvme_ctrl *ctrl)
4579 struct nvme_effects_log *cel;
4582 xa_for_each(&ctrl->cels, i, cel) {
4583 xa_erase(&ctrl->cels, i);
4587 xa_destroy(&ctrl->cels);
4590 static void nvme_free_ctrl(struct device *dev)
4592 struct nvme_ctrl *ctrl =
4593 container_of(dev, struct nvme_ctrl, ctrl_device);
4594 struct nvme_subsystem *subsys = ctrl->subsys;
4596 if (!subsys || ctrl->instance != subsys->instance)
4597 ida_free(&nvme_instance_ida, ctrl->instance);
4598 key_put(ctrl->tls_key);
4599 nvme_free_cels(ctrl);
4600 nvme_mpath_uninit(ctrl);
4601 nvme_auth_stop(ctrl);
4602 nvme_auth_free(ctrl);
4603 __free_page(ctrl->discard_page);
4604 free_opal_dev(ctrl->opal_dev);
4607 mutex_lock(&nvme_subsystems_lock);
4608 list_del(&ctrl->subsys_entry);
4609 sysfs_remove_link(&subsys->dev.kobj, dev_name(ctrl->device));
4610 mutex_unlock(&nvme_subsystems_lock);
4613 ctrl->ops->free_ctrl(ctrl);
4616 nvme_put_subsystem(subsys);
4620 * Initialize a NVMe controller structures. This needs to be called during
4621 * earliest initialization so that we have the initialized structured around
4624 int nvme_init_ctrl(struct nvme_ctrl *ctrl, struct device *dev,
4625 const struct nvme_ctrl_ops *ops, unsigned long quirks)
4629 WRITE_ONCE(ctrl->state, NVME_CTRL_NEW);
4630 ctrl->passthru_err_log_enabled = false;
4631 clear_bit(NVME_CTRL_FAILFAST_EXPIRED, &ctrl->flags);
4632 spin_lock_init(&ctrl->lock);
4633 mutex_init(&ctrl->scan_lock);
4634 INIT_LIST_HEAD(&ctrl->namespaces);
4635 xa_init(&ctrl->cels);
4636 init_rwsem(&ctrl->namespaces_rwsem);
4639 ctrl->quirks = quirks;
4640 ctrl->numa_node = NUMA_NO_NODE;
4641 INIT_WORK(&ctrl->scan_work, nvme_scan_work);
4642 INIT_WORK(&ctrl->async_event_work, nvme_async_event_work);
4643 INIT_WORK(&ctrl->fw_act_work, nvme_fw_act_work);
4644 INIT_WORK(&ctrl->delete_work, nvme_delete_ctrl_work);
4645 init_waitqueue_head(&ctrl->state_wq);
4647 INIT_DELAYED_WORK(&ctrl->ka_work, nvme_keep_alive_work);
4648 INIT_DELAYED_WORK(&ctrl->failfast_work, nvme_failfast_work);
4649 memset(&ctrl->ka_cmd, 0, sizeof(ctrl->ka_cmd));
4650 ctrl->ka_cmd.common.opcode = nvme_admin_keep_alive;
4651 ctrl->ka_last_check_time = jiffies;
4653 BUILD_BUG_ON(NVME_DSM_MAX_RANGES * sizeof(struct nvme_dsm_range) >
4655 ctrl->discard_page = alloc_page(GFP_KERNEL);
4656 if (!ctrl->discard_page) {
4661 ret = ida_alloc(&nvme_instance_ida, GFP_KERNEL);
4664 ctrl->instance = ret;
4666 device_initialize(&ctrl->ctrl_device);
4667 ctrl->device = &ctrl->ctrl_device;
4668 ctrl->device->devt = MKDEV(MAJOR(nvme_ctrl_base_chr_devt),
4670 ctrl->device->class = &nvme_class;
4671 ctrl->device->parent = ctrl->dev;
4672 if (ops->dev_attr_groups)
4673 ctrl->device->groups = ops->dev_attr_groups;
4675 ctrl->device->groups = nvme_dev_attr_groups;
4676 ctrl->device->release = nvme_free_ctrl;
4677 dev_set_drvdata(ctrl->device, ctrl);
4678 ret = dev_set_name(ctrl->device, "nvme%d", ctrl->instance);
4680 goto out_release_instance;
4682 nvme_get_ctrl(ctrl);
4683 cdev_init(&ctrl->cdev, &nvme_dev_fops);
4684 ctrl->cdev.owner = ops->module;
4685 ret = cdev_device_add(&ctrl->cdev, ctrl->device);
4690 * Initialize latency tolerance controls. The sysfs files won't
4691 * be visible to userspace unless the device actually supports APST.
4693 ctrl->device->power.set_latency_tolerance = nvme_set_latency_tolerance;
4694 dev_pm_qos_update_user_latency_tolerance(ctrl->device,
4695 min(default_ps_max_latency_us, (unsigned long)S32_MAX));
4697 nvme_fault_inject_init(&ctrl->fault_inject, dev_name(ctrl->device));
4698 nvme_mpath_init_ctrl(ctrl);
4699 ret = nvme_auth_init_ctrl(ctrl);
4705 nvme_fault_inject_fini(&ctrl->fault_inject);
4706 dev_pm_qos_hide_latency_tolerance(ctrl->device);
4707 cdev_device_del(&ctrl->cdev, ctrl->device);
4709 nvme_put_ctrl(ctrl);
4710 kfree_const(ctrl->device->kobj.name);
4711 out_release_instance:
4712 ida_free(&nvme_instance_ida, ctrl->instance);
4714 if (ctrl->discard_page)
4715 __free_page(ctrl->discard_page);
4718 EXPORT_SYMBOL_GPL(nvme_init_ctrl);
4720 /* let I/O to all namespaces fail in preparation for surprise removal */
4721 void nvme_mark_namespaces_dead(struct nvme_ctrl *ctrl)
4725 down_read(&ctrl->namespaces_rwsem);
4726 list_for_each_entry(ns, &ctrl->namespaces, list)
4727 blk_mark_disk_dead(ns->disk);
4728 up_read(&ctrl->namespaces_rwsem);
4730 EXPORT_SYMBOL_GPL(nvme_mark_namespaces_dead);
4732 void nvme_unfreeze(struct nvme_ctrl *ctrl)
4736 down_read(&ctrl->namespaces_rwsem);
4737 list_for_each_entry(ns, &ctrl->namespaces, list)
4738 blk_mq_unfreeze_queue(ns->queue);
4739 up_read(&ctrl->namespaces_rwsem);
4740 clear_bit(NVME_CTRL_FROZEN, &ctrl->flags);
4742 EXPORT_SYMBOL_GPL(nvme_unfreeze);
4744 int nvme_wait_freeze_timeout(struct nvme_ctrl *ctrl, long timeout)
4748 down_read(&ctrl->namespaces_rwsem);
4749 list_for_each_entry(ns, &ctrl->namespaces, list) {
4750 timeout = blk_mq_freeze_queue_wait_timeout(ns->queue, timeout);
4754 up_read(&ctrl->namespaces_rwsem);
4757 EXPORT_SYMBOL_GPL(nvme_wait_freeze_timeout);
4759 void nvme_wait_freeze(struct nvme_ctrl *ctrl)
4763 down_read(&ctrl->namespaces_rwsem);
4764 list_for_each_entry(ns, &ctrl->namespaces, list)
4765 blk_mq_freeze_queue_wait(ns->queue);
4766 up_read(&ctrl->namespaces_rwsem);
4768 EXPORT_SYMBOL_GPL(nvme_wait_freeze);
4770 void nvme_start_freeze(struct nvme_ctrl *ctrl)
4774 set_bit(NVME_CTRL_FROZEN, &ctrl->flags);
4775 down_read(&ctrl->namespaces_rwsem);
4776 list_for_each_entry(ns, &ctrl->namespaces, list)
4777 blk_freeze_queue_start(ns->queue);
4778 up_read(&ctrl->namespaces_rwsem);
4780 EXPORT_SYMBOL_GPL(nvme_start_freeze);
4782 void nvme_quiesce_io_queues(struct nvme_ctrl *ctrl)
4786 if (!test_and_set_bit(NVME_CTRL_STOPPED, &ctrl->flags))
4787 blk_mq_quiesce_tagset(ctrl->tagset);
4789 blk_mq_wait_quiesce_done(ctrl->tagset);
4791 EXPORT_SYMBOL_GPL(nvme_quiesce_io_queues);
4793 void nvme_unquiesce_io_queues(struct nvme_ctrl *ctrl)
4797 if (test_and_clear_bit(NVME_CTRL_STOPPED, &ctrl->flags))
4798 blk_mq_unquiesce_tagset(ctrl->tagset);
4800 EXPORT_SYMBOL_GPL(nvme_unquiesce_io_queues);
4802 void nvme_quiesce_admin_queue(struct nvme_ctrl *ctrl)
4804 if (!test_and_set_bit(NVME_CTRL_ADMIN_Q_STOPPED, &ctrl->flags))
4805 blk_mq_quiesce_queue(ctrl->admin_q);
4807 blk_mq_wait_quiesce_done(ctrl->admin_q->tag_set);
4809 EXPORT_SYMBOL_GPL(nvme_quiesce_admin_queue);
4811 void nvme_unquiesce_admin_queue(struct nvme_ctrl *ctrl)
4813 if (test_and_clear_bit(NVME_CTRL_ADMIN_Q_STOPPED, &ctrl->flags))
4814 blk_mq_unquiesce_queue(ctrl->admin_q);
4816 EXPORT_SYMBOL_GPL(nvme_unquiesce_admin_queue);
4818 void nvme_sync_io_queues(struct nvme_ctrl *ctrl)
4822 down_read(&ctrl->namespaces_rwsem);
4823 list_for_each_entry(ns, &ctrl->namespaces, list)
4824 blk_sync_queue(ns->queue);
4825 up_read(&ctrl->namespaces_rwsem);
4827 EXPORT_SYMBOL_GPL(nvme_sync_io_queues);
4829 void nvme_sync_queues(struct nvme_ctrl *ctrl)
4831 nvme_sync_io_queues(ctrl);
4833 blk_sync_queue(ctrl->admin_q);
4835 EXPORT_SYMBOL_GPL(nvme_sync_queues);
4837 struct nvme_ctrl *nvme_ctrl_from_file(struct file *file)
4839 if (file->f_op != &nvme_dev_fops)
4841 return file->private_data;
4843 EXPORT_SYMBOL_NS_GPL(nvme_ctrl_from_file, NVME_TARGET_PASSTHRU);
4846 * Check we didn't inadvertently grow the command structure sizes:
4848 static inline void _nvme_check_size(void)
4850 BUILD_BUG_ON(sizeof(struct nvme_common_command) != 64);
4851 BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
4852 BUILD_BUG_ON(sizeof(struct nvme_identify) != 64);
4853 BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
4854 BUILD_BUG_ON(sizeof(struct nvme_download_firmware) != 64);
4855 BUILD_BUG_ON(sizeof(struct nvme_format_cmd) != 64);
4856 BUILD_BUG_ON(sizeof(struct nvme_dsm_cmd) != 64);
4857 BUILD_BUG_ON(sizeof(struct nvme_write_zeroes_cmd) != 64);
4858 BUILD_BUG_ON(sizeof(struct nvme_abort_cmd) != 64);
4859 BUILD_BUG_ON(sizeof(struct nvme_get_log_page_command) != 64);
4860 BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
4861 BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != NVME_IDENTIFY_DATA_SIZE);
4862 BUILD_BUG_ON(sizeof(struct nvme_id_ns) != NVME_IDENTIFY_DATA_SIZE);
4863 BUILD_BUG_ON(sizeof(struct nvme_id_ns_cs_indep) !=
4864 NVME_IDENTIFY_DATA_SIZE);
4865 BUILD_BUG_ON(sizeof(struct nvme_id_ns_zns) != NVME_IDENTIFY_DATA_SIZE);
4866 BUILD_BUG_ON(sizeof(struct nvme_id_ns_nvm) != NVME_IDENTIFY_DATA_SIZE);
4867 BUILD_BUG_ON(sizeof(struct nvme_id_ctrl_zns) != NVME_IDENTIFY_DATA_SIZE);
4868 BUILD_BUG_ON(sizeof(struct nvme_id_ctrl_nvm) != NVME_IDENTIFY_DATA_SIZE);
4869 BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
4870 BUILD_BUG_ON(sizeof(struct nvme_smart_log) != 512);
4871 BUILD_BUG_ON(sizeof(struct nvme_dbbuf) != 64);
4872 BUILD_BUG_ON(sizeof(struct nvme_directive_cmd) != 64);
4873 BUILD_BUG_ON(sizeof(struct nvme_feat_host_behavior) != 512);
4877 static int __init nvme_core_init(void)
4879 int result = -ENOMEM;
4883 nvme_wq = alloc_workqueue("nvme-wq",
4884 WQ_UNBOUND | WQ_MEM_RECLAIM | WQ_SYSFS, 0);
4888 nvme_reset_wq = alloc_workqueue("nvme-reset-wq",
4889 WQ_UNBOUND | WQ_MEM_RECLAIM | WQ_SYSFS, 0);
4893 nvme_delete_wq = alloc_workqueue("nvme-delete-wq",
4894 WQ_UNBOUND | WQ_MEM_RECLAIM | WQ_SYSFS, 0);
4895 if (!nvme_delete_wq)
4896 goto destroy_reset_wq;
4898 result = alloc_chrdev_region(&nvme_ctrl_base_chr_devt, 0,
4899 NVME_MINORS, "nvme");
4901 goto destroy_delete_wq;
4903 result = class_register(&nvme_class);
4905 goto unregister_chrdev;
4907 result = class_register(&nvme_subsys_class);
4911 result = alloc_chrdev_region(&nvme_ns_chr_devt, 0, NVME_MINORS,
4914 goto destroy_subsys_class;
4916 result = class_register(&nvme_ns_chr_class);
4918 goto unregister_generic_ns;
4920 result = nvme_init_auth();
4922 goto destroy_ns_chr;
4926 class_unregister(&nvme_ns_chr_class);
4927 unregister_generic_ns:
4928 unregister_chrdev_region(nvme_ns_chr_devt, NVME_MINORS);
4929 destroy_subsys_class:
4930 class_unregister(&nvme_subsys_class);
4932 class_unregister(&nvme_class);
4934 unregister_chrdev_region(nvme_ctrl_base_chr_devt, NVME_MINORS);
4936 destroy_workqueue(nvme_delete_wq);
4938 destroy_workqueue(nvme_reset_wq);
4940 destroy_workqueue(nvme_wq);
4945 static void __exit nvme_core_exit(void)
4948 class_unregister(&nvme_ns_chr_class);
4949 class_unregister(&nvme_subsys_class);
4950 class_unregister(&nvme_class);
4951 unregister_chrdev_region(nvme_ns_chr_devt, NVME_MINORS);
4952 unregister_chrdev_region(nvme_ctrl_base_chr_devt, NVME_MINORS);
4953 destroy_workqueue(nvme_delete_wq);
4954 destroy_workqueue(nvme_reset_wq);
4955 destroy_workqueue(nvme_wq);
4956 ida_destroy(&nvme_ns_chr_minor_ida);
4957 ida_destroy(&nvme_instance_ida);
4960 MODULE_LICENSE("GPL");
4961 MODULE_VERSION("1.0");
4962 MODULE_DESCRIPTION("NVMe host core framework");
4963 module_init(nvme_core_init);
4964 module_exit(nvme_core_exit);