2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (C) 2012-2016 Intel Corporation
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
34 #include <sys/param.h>
35 #include <sys/systm.h>
39 #include <sys/ioccom.h>
43 #include <sys/endian.h>
46 #include "nvme_private.h"
48 #define B4_CHK_RDY_DELAY_MS 2300 /* work around controller bug */
50 static void nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
51 struct nvme_async_event_request *aer);
54 nvme_ctrlr_construct_admin_qpair(struct nvme_controller *ctrlr)
56 struct nvme_qpair *qpair;
60 qpair = &ctrlr->adminq;
62 qpair->cpu = CPU_FFS(&cpuset_domain[ctrlr->domain]) - 1;
63 qpair->domain = ctrlr->domain;
65 num_entries = NVME_ADMIN_ENTRIES;
66 TUNABLE_INT_FETCH("hw.nvme.admin_entries", &num_entries);
68 * If admin_entries was overridden to an invalid value, revert it
69 * back to our default value.
71 if (num_entries < NVME_MIN_ADMIN_ENTRIES ||
72 num_entries > NVME_MAX_ADMIN_ENTRIES) {
73 nvme_printf(ctrlr, "invalid hw.nvme.admin_entries=%d "
74 "specified\n", num_entries);
75 num_entries = NVME_ADMIN_ENTRIES;
79 * The admin queue's max xfer size is treated differently than the
80 * max I/O xfer size. 16KB is sufficient here - maybe even less?
82 error = nvme_qpair_construct(qpair, num_entries, NVME_ADMIN_TRACKERS,
87 #define QP(ctrlr, c) ((c) * (ctrlr)->num_io_queues / mp_ncpus)
90 nvme_ctrlr_construct_io_qpairs(struct nvme_controller *ctrlr)
92 struct nvme_qpair *qpair;
96 int num_entries, num_trackers, max_entries;
99 * NVMe spec sets a hard limit of 64K max entries, but devices may
100 * specify a smaller limit, so we need to check the MQES field in the
101 * capabilities register. We have to cap the number of entries to the
102 * current stride allows for in BAR 0/1, otherwise the remainder entries
103 * are inaccessable. MQES should reflect this, and this is just a
107 (rman_get_size(ctrlr->resource) - nvme_mmio_offsetof(doorbell[0])) /
108 (1 << (ctrlr->dstrd + 1));
109 num_entries = NVME_IO_ENTRIES;
110 TUNABLE_INT_FETCH("hw.nvme.io_entries", &num_entries);
111 cap_lo = nvme_mmio_read_4(ctrlr, cap_lo);
112 mqes = NVME_CAP_LO_MQES(cap_lo);
113 num_entries = min(num_entries, mqes + 1);
114 num_entries = min(num_entries, max_entries);
116 num_trackers = NVME_IO_TRACKERS;
117 TUNABLE_INT_FETCH("hw.nvme.io_trackers", &num_trackers);
119 num_trackers = max(num_trackers, NVME_MIN_IO_TRACKERS);
120 num_trackers = min(num_trackers, NVME_MAX_IO_TRACKERS);
122 * No need to have more trackers than entries in the submit queue. Note
123 * also that for a queue size of N, we can only have (N-1) commands
124 * outstanding, hence the "-1" here.
126 num_trackers = min(num_trackers, (num_entries-1));
129 * Our best estimate for the maximum number of I/Os that we should
130 * normally have in flight at one time. This should be viewed as a hint,
131 * not a hard limit and will need to be revisited when the upper layers
132 * of the storage system grows multi-queue support.
134 ctrlr->max_hw_pend_io = num_trackers * ctrlr->num_io_queues * 3 / 4;
136 ctrlr->ioq = malloc(ctrlr->num_io_queues * sizeof(struct nvme_qpair),
137 M_NVME, M_ZERO | M_WAITOK);
139 for (i = c = n = 0; i < ctrlr->num_io_queues; i++, c += n) {
140 qpair = &ctrlr->ioq[i];
143 * Admin queue has ID=0. IO queues start at ID=1 -
144 * hence the 'i+1' here.
147 if (ctrlr->num_io_queues > 1) {
148 /* Find number of CPUs served by this queue. */
149 for (n = 1; QP(ctrlr, c + n) == i; n++)
151 /* Shuffle multiple NVMe devices between CPUs. */
152 qpair->cpu = c + (device_get_unit(ctrlr->dev)+n/2) % n;
153 qpair->domain = pcpu_find(qpair->cpu)->pc_domain;
155 qpair->cpu = CPU_FFS(&cpuset_domain[ctrlr->domain]) - 1;
156 qpair->domain = ctrlr->domain;
160 * For I/O queues, use the controller-wide max_xfer_size
161 * calculated in nvme_attach().
163 error = nvme_qpair_construct(qpair, num_entries, num_trackers,
169 * Do not bother binding interrupts if we only have one I/O
170 * interrupt thread for this controller.
172 if (ctrlr->num_io_queues > 1)
173 bus_bind_intr(ctrlr->dev, qpair->res, qpair->cpu);
180 nvme_ctrlr_fail(struct nvme_controller *ctrlr)
184 ctrlr->is_failed = TRUE;
185 nvme_admin_qpair_disable(&ctrlr->adminq);
186 nvme_qpair_fail(&ctrlr->adminq);
187 if (ctrlr->ioq != NULL) {
188 for (i = 0; i < ctrlr->num_io_queues; i++) {
189 nvme_io_qpair_disable(&ctrlr->ioq[i]);
190 nvme_qpair_fail(&ctrlr->ioq[i]);
193 nvme_notify_fail_consumers(ctrlr);
197 nvme_ctrlr_post_failed_request(struct nvme_controller *ctrlr,
198 struct nvme_request *req)
201 mtx_lock(&ctrlr->lock);
202 STAILQ_INSERT_TAIL(&ctrlr->fail_req, req, stailq);
203 mtx_unlock(&ctrlr->lock);
204 taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->fail_req_task);
208 nvme_ctrlr_fail_req_task(void *arg, int pending)
210 struct nvme_controller *ctrlr = arg;
211 struct nvme_request *req;
213 mtx_lock(&ctrlr->lock);
214 while ((req = STAILQ_FIRST(&ctrlr->fail_req)) != NULL) {
215 STAILQ_REMOVE_HEAD(&ctrlr->fail_req, stailq);
216 mtx_unlock(&ctrlr->lock);
217 nvme_qpair_manual_complete_request(req->qpair, req,
218 NVME_SCT_GENERIC, NVME_SC_ABORTED_BY_REQUEST);
219 mtx_lock(&ctrlr->lock);
221 mtx_unlock(&ctrlr->lock);
225 nvme_ctrlr_wait_for_ready(struct nvme_controller *ctrlr, int desired_val)
232 csts = nvme_mmio_read_4(ctrlr, csts);
233 if (csts == 0xffffffff) /* Hot unplug. */
235 if (((csts >> NVME_CSTS_REG_RDY_SHIFT) & NVME_CSTS_REG_RDY_MASK)
238 if (ms_waited++ > ctrlr->ready_timeout_in_ms) {
239 nvme_printf(ctrlr, "controller ready did not become %d "
240 "within %d ms\n", desired_val, ctrlr->ready_timeout_in_ms);
250 nvme_ctrlr_disable(struct nvme_controller *ctrlr)
257 cc = nvme_mmio_read_4(ctrlr, cc);
258 csts = nvme_mmio_read_4(ctrlr, csts);
260 en = (cc >> NVME_CC_REG_EN_SHIFT) & NVME_CC_REG_EN_MASK;
261 rdy = (csts >> NVME_CSTS_REG_RDY_SHIFT) & NVME_CSTS_REG_RDY_MASK;
264 * Per 3.1.5 in NVME 1.3 spec, transitioning CC.EN from 0 to 1
265 * when CSTS.RDY is 1 or transitioning CC.EN from 1 to 0 when
266 * CSTS.RDY is 0 "has undefined results" So make sure that CSTS.RDY
267 * isn't the desired value. Short circuit if we're already disabled.
271 /* EN == 1, wait for RDY == 1 or fail */
272 err = nvme_ctrlr_wait_for_ready(ctrlr, 1);
277 /* EN == 0 already wait for RDY == 0 */
281 return (nvme_ctrlr_wait_for_ready(ctrlr, 0));
284 cc &= ~NVME_CC_REG_EN_MASK;
285 nvme_mmio_write_4(ctrlr, cc, cc);
287 * Some drives have issues with accessing the mmio after we
288 * disable, so delay for a bit after we write the bit to
289 * cope with these issues.
291 if (ctrlr->quirks & QUIRK_DELAY_B4_CHK_RDY)
292 pause("nvmeR", B4_CHK_RDY_DELAY_MS * hz / 1000);
293 return (nvme_ctrlr_wait_for_ready(ctrlr, 0));
297 nvme_ctrlr_enable(struct nvme_controller *ctrlr)
306 cc = nvme_mmio_read_4(ctrlr, cc);
307 csts = nvme_mmio_read_4(ctrlr, csts);
309 en = (cc >> NVME_CC_REG_EN_SHIFT) & NVME_CC_REG_EN_MASK;
310 rdy = (csts >> NVME_CSTS_REG_RDY_SHIFT) & NVME_CSTS_REG_RDY_MASK;
313 * See note in nvme_ctrlr_disable. Short circuit if we're already enabled.
319 return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
321 /* EN == 0 already wait for RDY == 0 or fail */
322 err = nvme_ctrlr_wait_for_ready(ctrlr, 0);
327 nvme_mmio_write_8(ctrlr, asq, ctrlr->adminq.cmd_bus_addr);
329 nvme_mmio_write_8(ctrlr, acq, ctrlr->adminq.cpl_bus_addr);
332 /* acqs and asqs are 0-based. */
333 qsize = ctrlr->adminq.num_entries - 1;
336 aqa = (qsize & NVME_AQA_REG_ACQS_MASK) << NVME_AQA_REG_ACQS_SHIFT;
337 aqa |= (qsize & NVME_AQA_REG_ASQS_MASK) << NVME_AQA_REG_ASQS_SHIFT;
338 nvme_mmio_write_4(ctrlr, aqa, aqa);
341 /* Initialization values for CC */
343 cc |= 1 << NVME_CC_REG_EN_SHIFT;
344 cc |= 0 << NVME_CC_REG_CSS_SHIFT;
345 cc |= 0 << NVME_CC_REG_AMS_SHIFT;
346 cc |= 0 << NVME_CC_REG_SHN_SHIFT;
347 cc |= 6 << NVME_CC_REG_IOSQES_SHIFT; /* SQ entry size == 64 == 2^6 */
348 cc |= 4 << NVME_CC_REG_IOCQES_SHIFT; /* CQ entry size == 16 == 2^4 */
350 /* This evaluates to 0, which is according to spec. */
351 cc |= (PAGE_SIZE >> 13) << NVME_CC_REG_MPS_SHIFT;
353 nvme_mmio_write_4(ctrlr, cc, cc);
355 return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
359 nvme_ctrlr_disable_qpairs(struct nvme_controller *ctrlr)
363 nvme_admin_qpair_disable(&ctrlr->adminq);
365 * I/O queues are not allocated before the initial HW
366 * reset, so do not try to disable them. Use is_initialized
367 * to determine if this is the initial HW reset.
369 if (ctrlr->is_initialized) {
370 for (i = 0; i < ctrlr->num_io_queues; i++)
371 nvme_io_qpair_disable(&ctrlr->ioq[i]);
376 nvme_ctrlr_hw_reset(struct nvme_controller *ctrlr)
380 nvme_ctrlr_disable_qpairs(ctrlr);
384 err = nvme_ctrlr_disable(ctrlr);
387 return (nvme_ctrlr_enable(ctrlr));
391 nvme_ctrlr_reset(struct nvme_controller *ctrlr)
395 cmpset = atomic_cmpset_32(&ctrlr->is_resetting, 0, 1);
397 if (cmpset == 0 || ctrlr->is_failed)
399 * Controller is already resetting or has failed. Return
400 * immediately since there is no need to kick off another
401 * reset in these cases.
405 taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->reset_task);
409 nvme_ctrlr_identify(struct nvme_controller *ctrlr)
411 struct nvme_completion_poll_status status;
414 nvme_ctrlr_cmd_identify_controller(ctrlr, &ctrlr->cdata,
415 nvme_completion_poll_cb, &status);
416 nvme_completion_poll(&status);
417 if (nvme_completion_is_error(&status.cpl)) {
418 nvme_printf(ctrlr, "nvme_identify_controller failed!\n");
422 /* Convert data to host endian */
423 nvme_controller_data_swapbytes(&ctrlr->cdata);
426 * Use MDTS to ensure our default max_xfer_size doesn't exceed what the
427 * controller supports.
429 if (ctrlr->cdata.mdts > 0)
430 ctrlr->max_xfer_size = min(ctrlr->max_xfer_size,
431 ctrlr->min_page_size * (1 << (ctrlr->cdata.mdts)));
437 nvme_ctrlr_set_num_qpairs(struct nvme_controller *ctrlr)
439 struct nvme_completion_poll_status status;
440 int cq_allocated, sq_allocated;
443 nvme_ctrlr_cmd_set_num_queues(ctrlr, ctrlr->num_io_queues,
444 nvme_completion_poll_cb, &status);
445 nvme_completion_poll(&status);
446 if (nvme_completion_is_error(&status.cpl)) {
447 nvme_printf(ctrlr, "nvme_ctrlr_set_num_qpairs failed!\n");
452 * Data in cdw0 is 0-based.
453 * Lower 16-bits indicate number of submission queues allocated.
454 * Upper 16-bits indicate number of completion queues allocated.
456 sq_allocated = (status.cpl.cdw0 & 0xFFFF) + 1;
457 cq_allocated = (status.cpl.cdw0 >> 16) + 1;
460 * Controller may allocate more queues than we requested,
461 * so use the minimum of the number requested and what was
462 * actually allocated.
464 ctrlr->num_io_queues = min(ctrlr->num_io_queues, sq_allocated);
465 ctrlr->num_io_queues = min(ctrlr->num_io_queues, cq_allocated);
466 if (ctrlr->num_io_queues > vm_ndomains)
467 ctrlr->num_io_queues -= ctrlr->num_io_queues % vm_ndomains;
473 nvme_ctrlr_create_qpairs(struct nvme_controller *ctrlr)
475 struct nvme_completion_poll_status status;
476 struct nvme_qpair *qpair;
479 for (i = 0; i < ctrlr->num_io_queues; i++) {
480 qpair = &ctrlr->ioq[i];
483 nvme_ctrlr_cmd_create_io_cq(ctrlr, qpair,
484 nvme_completion_poll_cb, &status);
485 nvme_completion_poll(&status);
486 if (nvme_completion_is_error(&status.cpl)) {
487 nvme_printf(ctrlr, "nvme_create_io_cq failed!\n");
492 nvme_ctrlr_cmd_create_io_sq(qpair->ctrlr, qpair,
493 nvme_completion_poll_cb, &status);
494 nvme_completion_poll(&status);
495 if (nvme_completion_is_error(&status.cpl)) {
496 nvme_printf(ctrlr, "nvme_create_io_sq failed!\n");
505 nvme_ctrlr_delete_qpairs(struct nvme_controller *ctrlr)
507 struct nvme_completion_poll_status status;
508 struct nvme_qpair *qpair;
510 for (int i = 0; i < ctrlr->num_io_queues; i++) {
511 qpair = &ctrlr->ioq[i];
514 nvme_ctrlr_cmd_delete_io_sq(ctrlr, qpair,
515 nvme_completion_poll_cb, &status);
516 nvme_completion_poll(&status);
517 if (nvme_completion_is_error(&status.cpl)) {
518 nvme_printf(ctrlr, "nvme_destroy_io_sq failed!\n");
523 nvme_ctrlr_cmd_delete_io_cq(ctrlr, qpair,
524 nvme_completion_poll_cb, &status);
525 nvme_completion_poll(&status);
526 if (nvme_completion_is_error(&status.cpl)) {
527 nvme_printf(ctrlr, "nvme_destroy_io_cq failed!\n");
536 nvme_ctrlr_construct_namespaces(struct nvme_controller *ctrlr)
538 struct nvme_namespace *ns;
541 for (i = 0; i < min(ctrlr->cdata.nn, NVME_MAX_NAMESPACES); i++) {
543 nvme_ns_construct(ns, i+1, ctrlr);
550 is_log_page_id_valid(uint8_t page_id)
555 case NVME_LOG_HEALTH_INFORMATION:
556 case NVME_LOG_FIRMWARE_SLOT:
557 case NVME_LOG_CHANGED_NAMESPACE:
558 case NVME_LOG_COMMAND_EFFECT:
559 case NVME_LOG_RES_NOTIFICATION:
560 case NVME_LOG_SANITIZE_STATUS:
568 nvme_ctrlr_get_log_page_size(struct nvme_controller *ctrlr, uint8_t page_id)
570 uint32_t log_page_size;
575 sizeof(struct nvme_error_information_entry) *
576 (ctrlr->cdata.elpe + 1), NVME_MAX_AER_LOG_SIZE);
578 case NVME_LOG_HEALTH_INFORMATION:
579 log_page_size = sizeof(struct nvme_health_information_page);
581 case NVME_LOG_FIRMWARE_SLOT:
582 log_page_size = sizeof(struct nvme_firmware_page);
584 case NVME_LOG_CHANGED_NAMESPACE:
585 log_page_size = sizeof(struct nvme_ns_list);
587 case NVME_LOG_COMMAND_EFFECT:
588 log_page_size = sizeof(struct nvme_command_effects_page);
590 case NVME_LOG_RES_NOTIFICATION:
591 log_page_size = sizeof(struct nvme_res_notification_page);
593 case NVME_LOG_SANITIZE_STATUS:
594 log_page_size = sizeof(struct nvme_sanitize_status_page);
601 return (log_page_size);
605 nvme_ctrlr_log_critical_warnings(struct nvme_controller *ctrlr,
609 if (state & NVME_CRIT_WARN_ST_AVAILABLE_SPARE)
610 nvme_printf(ctrlr, "available spare space below threshold\n");
612 if (state & NVME_CRIT_WARN_ST_TEMPERATURE)
613 nvme_printf(ctrlr, "temperature above threshold\n");
615 if (state & NVME_CRIT_WARN_ST_DEVICE_RELIABILITY)
616 nvme_printf(ctrlr, "device reliability degraded\n");
618 if (state & NVME_CRIT_WARN_ST_READ_ONLY)
619 nvme_printf(ctrlr, "media placed in read only mode\n");
621 if (state & NVME_CRIT_WARN_ST_VOLATILE_MEMORY_BACKUP)
622 nvme_printf(ctrlr, "volatile memory backup device failed\n");
624 if (state & NVME_CRIT_WARN_ST_RESERVED_MASK)
626 "unknown critical warning(s): state = 0x%02x\n", state);
630 nvme_ctrlr_async_event_log_page_cb(void *arg, const struct nvme_completion *cpl)
632 struct nvme_async_event_request *aer = arg;
633 struct nvme_health_information_page *health_info;
634 struct nvme_ns_list *nsl;
635 struct nvme_error_information_entry *err;
639 * If the log page fetch for some reason completed with an error,
640 * don't pass log page data to the consumers. In practice, this case
641 * should never happen.
643 if (nvme_completion_is_error(cpl))
644 nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
645 aer->log_page_id, NULL, 0);
647 /* Convert data to host endian */
648 switch (aer->log_page_id) {
650 err = (struct nvme_error_information_entry *)aer->log_page_buffer;
651 for (i = 0; i < (aer->ctrlr->cdata.elpe + 1); i++)
652 nvme_error_information_entry_swapbytes(err++);
654 case NVME_LOG_HEALTH_INFORMATION:
655 nvme_health_information_page_swapbytes(
656 (struct nvme_health_information_page *)aer->log_page_buffer);
658 case NVME_LOG_FIRMWARE_SLOT:
659 nvme_firmware_page_swapbytes(
660 (struct nvme_firmware_page *)aer->log_page_buffer);
662 case NVME_LOG_CHANGED_NAMESPACE:
663 nvme_ns_list_swapbytes(
664 (struct nvme_ns_list *)aer->log_page_buffer);
666 case NVME_LOG_COMMAND_EFFECT:
667 nvme_command_effects_page_swapbytes(
668 (struct nvme_command_effects_page *)aer->log_page_buffer);
670 case NVME_LOG_RES_NOTIFICATION:
671 nvme_res_notification_page_swapbytes(
672 (struct nvme_res_notification_page *)aer->log_page_buffer);
674 case NVME_LOG_SANITIZE_STATUS:
675 nvme_sanitize_status_page_swapbytes(
676 (struct nvme_sanitize_status_page *)aer->log_page_buffer);
678 case INTEL_LOG_TEMP_STATS:
679 intel_log_temp_stats_swapbytes(
680 (struct intel_log_temp_stats *)aer->log_page_buffer);
686 if (aer->log_page_id == NVME_LOG_HEALTH_INFORMATION) {
687 health_info = (struct nvme_health_information_page *)
688 aer->log_page_buffer;
689 nvme_ctrlr_log_critical_warnings(aer->ctrlr,
690 health_info->critical_warning);
692 * Critical warnings reported through the
693 * SMART/health log page are persistent, so
694 * clear the associated bits in the async event
695 * config so that we do not receive repeated
696 * notifications for the same event.
698 aer->ctrlr->async_event_config &=
699 ~health_info->critical_warning;
700 nvme_ctrlr_cmd_set_async_event_config(aer->ctrlr,
701 aer->ctrlr->async_event_config, NULL, NULL);
702 } else if (aer->log_page_id == NVME_LOG_CHANGED_NAMESPACE &&
704 nsl = (struct nvme_ns_list *)aer->log_page_buffer;
705 for (i = 0; i < nitems(nsl->ns) && nsl->ns[i] != 0; i++) {
706 if (nsl->ns[i] > NVME_MAX_NAMESPACES)
708 nvme_notify_ns(aer->ctrlr, nsl->ns[i]);
714 * Pass the cpl data from the original async event completion,
715 * not the log page fetch.
717 nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
718 aer->log_page_id, aer->log_page_buffer, aer->log_page_size);
722 * Repost another asynchronous event request to replace the one
723 * that just completed.
725 nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
729 nvme_ctrlr_async_event_cb(void *arg, const struct nvme_completion *cpl)
731 struct nvme_async_event_request *aer = arg;
733 if (nvme_completion_is_error(cpl)) {
735 * Do not retry failed async event requests. This avoids
736 * infinite loops where a new async event request is submitted
737 * to replace the one just failed, only to fail again and
738 * perpetuate the loop.
743 /* Associated log page is in bits 23:16 of completion entry dw0. */
744 aer->log_page_id = (cpl->cdw0 & 0xFF0000) >> 16;
746 nvme_printf(aer->ctrlr, "async event occurred (type 0x%x, info 0x%02x,"
747 " page 0x%02x)\n", (cpl->cdw0 & 0x07), (cpl->cdw0 & 0xFF00) >> 8,
750 if (is_log_page_id_valid(aer->log_page_id)) {
751 aer->log_page_size = nvme_ctrlr_get_log_page_size(aer->ctrlr,
753 memcpy(&aer->cpl, cpl, sizeof(*cpl));
754 nvme_ctrlr_cmd_get_log_page(aer->ctrlr, aer->log_page_id,
755 NVME_GLOBAL_NAMESPACE_TAG, aer->log_page_buffer,
756 aer->log_page_size, nvme_ctrlr_async_event_log_page_cb,
758 /* Wait to notify consumers until after log page is fetched. */
760 nvme_notify_async_consumers(aer->ctrlr, cpl, aer->log_page_id,
764 * Repost another asynchronous event request to replace the one
765 * that just completed.
767 nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
772 nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
773 struct nvme_async_event_request *aer)
775 struct nvme_request *req;
778 req = nvme_allocate_request_null(nvme_ctrlr_async_event_cb, aer);
782 * Disable timeout here, since asynchronous event requests should by
783 * nature never be timed out.
785 req->timeout = FALSE;
786 req->cmd.opc = NVME_OPC_ASYNC_EVENT_REQUEST;
787 nvme_ctrlr_submit_admin_request(ctrlr, req);
791 nvme_ctrlr_configure_aer(struct nvme_controller *ctrlr)
793 struct nvme_completion_poll_status status;
794 struct nvme_async_event_request *aer;
797 ctrlr->async_event_config = NVME_CRIT_WARN_ST_AVAILABLE_SPARE |
798 NVME_CRIT_WARN_ST_DEVICE_RELIABILITY |
799 NVME_CRIT_WARN_ST_READ_ONLY |
800 NVME_CRIT_WARN_ST_VOLATILE_MEMORY_BACKUP;
801 if (ctrlr->cdata.ver >= NVME_REV(1, 2))
802 ctrlr->async_event_config |= 0x300;
805 nvme_ctrlr_cmd_get_feature(ctrlr, NVME_FEAT_TEMPERATURE_THRESHOLD,
806 0, NULL, 0, nvme_completion_poll_cb, &status);
807 nvme_completion_poll(&status);
808 if (nvme_completion_is_error(&status.cpl) ||
809 (status.cpl.cdw0 & 0xFFFF) == 0xFFFF ||
810 (status.cpl.cdw0 & 0xFFFF) == 0x0000) {
811 nvme_printf(ctrlr, "temperature threshold not supported\n");
813 ctrlr->async_event_config |= NVME_CRIT_WARN_ST_TEMPERATURE;
815 nvme_ctrlr_cmd_set_async_event_config(ctrlr,
816 ctrlr->async_event_config, NULL, NULL);
818 /* aerl is a zero-based value, so we need to add 1 here. */
819 ctrlr->num_aers = min(NVME_MAX_ASYNC_EVENTS, (ctrlr->cdata.aerl+1));
821 for (i = 0; i < ctrlr->num_aers; i++) {
822 aer = &ctrlr->aer[i];
823 nvme_ctrlr_construct_and_submit_aer(ctrlr, aer);
828 nvme_ctrlr_configure_int_coalescing(struct nvme_controller *ctrlr)
831 ctrlr->int_coal_time = 0;
832 TUNABLE_INT_FETCH("hw.nvme.int_coal_time",
833 &ctrlr->int_coal_time);
835 ctrlr->int_coal_threshold = 0;
836 TUNABLE_INT_FETCH("hw.nvme.int_coal_threshold",
837 &ctrlr->int_coal_threshold);
839 nvme_ctrlr_cmd_set_interrupt_coalescing(ctrlr, ctrlr->int_coal_time,
840 ctrlr->int_coal_threshold, NULL, NULL);
844 nvme_ctrlr_start(void *ctrlr_arg, bool resetting)
846 struct nvme_controller *ctrlr = ctrlr_arg;
847 uint32_t old_num_io_queues;
851 * Only reset adminq here when we are restarting the
852 * controller after a reset. During initialization,
853 * we have already submitted admin commands to get
854 * the number of I/O queues supported, so cannot reset
855 * the adminq again here.
858 nvme_qpair_reset(&ctrlr->adminq);
860 for (i = 0; i < ctrlr->num_io_queues; i++)
861 nvme_qpair_reset(&ctrlr->ioq[i]);
863 nvme_admin_qpair_enable(&ctrlr->adminq);
865 if (nvme_ctrlr_identify(ctrlr) != 0) {
866 nvme_ctrlr_fail(ctrlr);
871 * The number of qpairs are determined during controller initialization,
872 * including using NVMe SET_FEATURES/NUMBER_OF_QUEUES to determine the
873 * HW limit. We call SET_FEATURES again here so that it gets called
874 * after any reset for controllers that depend on the driver to
875 * explicit specify how many queues it will use. This value should
876 * never change between resets, so panic if somehow that does happen.
879 old_num_io_queues = ctrlr->num_io_queues;
880 if (nvme_ctrlr_set_num_qpairs(ctrlr) != 0) {
881 nvme_ctrlr_fail(ctrlr);
885 if (old_num_io_queues != ctrlr->num_io_queues) {
886 panic("num_io_queues changed from %u to %u",
887 old_num_io_queues, ctrlr->num_io_queues);
891 if (nvme_ctrlr_create_qpairs(ctrlr) != 0) {
892 nvme_ctrlr_fail(ctrlr);
896 if (nvme_ctrlr_construct_namespaces(ctrlr) != 0) {
897 nvme_ctrlr_fail(ctrlr);
901 nvme_ctrlr_configure_aer(ctrlr);
902 nvme_ctrlr_configure_int_coalescing(ctrlr);
904 for (i = 0; i < ctrlr->num_io_queues; i++)
905 nvme_io_qpair_enable(&ctrlr->ioq[i]);
909 nvme_ctrlr_start_config_hook(void *arg)
911 struct nvme_controller *ctrlr = arg;
913 nvme_qpair_reset(&ctrlr->adminq);
914 nvme_admin_qpair_enable(&ctrlr->adminq);
916 if (nvme_ctrlr_set_num_qpairs(ctrlr) == 0 &&
917 nvme_ctrlr_construct_io_qpairs(ctrlr) == 0)
918 nvme_ctrlr_start(ctrlr, false);
920 nvme_ctrlr_fail(ctrlr);
922 nvme_sysctl_initialize_ctrlr(ctrlr);
923 config_intrhook_disestablish(&ctrlr->config_hook);
925 ctrlr->is_initialized = 1;
926 nvme_notify_new_controller(ctrlr);
930 nvme_ctrlr_reset_task(void *arg, int pending)
932 struct nvme_controller *ctrlr = arg;
935 nvme_printf(ctrlr, "resetting controller\n");
936 status = nvme_ctrlr_hw_reset(ctrlr);
938 * Use pause instead of DELAY, so that we yield to any nvme interrupt
939 * handlers on this CPU that were blocked on a qpair lock. We want
940 * all nvme interrupts completed before proceeding with restarting the
943 * XXX - any way to guarantee the interrupt handlers have quiesced?
945 pause("nvmereset", hz / 10);
947 nvme_ctrlr_start(ctrlr, true);
949 nvme_ctrlr_fail(ctrlr);
951 atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
955 * Poll all the queues enabled on the device for completion.
958 nvme_ctrlr_poll(struct nvme_controller *ctrlr)
962 nvme_qpair_process_completions(&ctrlr->adminq);
964 for (i = 0; i < ctrlr->num_io_queues; i++)
965 if (ctrlr->ioq && ctrlr->ioq[i].cpl)
966 nvme_qpair_process_completions(&ctrlr->ioq[i]);
970 * Poll the single-vector interrupt case: num_io_queues will be 1 and
971 * there's only a single vector. While we're polling, we mask further
972 * interrupts in the controller.
975 nvme_ctrlr_intx_handler(void *arg)
977 struct nvme_controller *ctrlr = arg;
979 nvme_mmio_write_4(ctrlr, intms, 1);
980 nvme_ctrlr_poll(ctrlr);
981 nvme_mmio_write_4(ctrlr, intmc, 1);
985 nvme_pt_done(void *arg, const struct nvme_completion *cpl)
987 struct nvme_pt_command *pt = arg;
988 struct mtx *mtx = pt->driver_lock;
991 bzero(&pt->cpl, sizeof(pt->cpl));
992 pt->cpl.cdw0 = cpl->cdw0;
994 status = cpl->status;
995 status &= ~NVME_STATUS_P_MASK;
996 pt->cpl.status = status;
999 pt->driver_lock = NULL;
1005 nvme_ctrlr_passthrough_cmd(struct nvme_controller *ctrlr,
1006 struct nvme_pt_command *pt, uint32_t nsid, int is_user_buffer,
1009 struct nvme_request *req;
1011 struct buf *buf = NULL;
1013 vm_offset_t addr, end;
1017 * vmapbuf calls vm_fault_quick_hold_pages which only maps full
1018 * pages. Ensure this request has fewer than MAXPHYS bytes when
1019 * extended to full pages.
1021 addr = (vm_offset_t)pt->buf;
1022 end = round_page(addr + pt->len);
1023 addr = trunc_page(addr);
1024 if (end - addr > MAXPHYS)
1027 if (pt->len > ctrlr->max_xfer_size) {
1028 nvme_printf(ctrlr, "pt->len (%d) "
1029 "exceeds max_xfer_size (%d)\n", pt->len,
1030 ctrlr->max_xfer_size);
1033 if (is_user_buffer) {
1035 * Ensure the user buffer is wired for the duration of
1036 * this pass-through command.
1039 buf = getpbuf(NULL);
1040 buf->b_data = pt->buf;
1041 buf->b_bufsize = pt->len;
1042 buf->b_iocmd = pt->is_read ? BIO_READ : BIO_WRITE;
1043 if (vmapbuf(buf, 1) < 0) {
1047 req = nvme_allocate_request_vaddr(buf->b_data, pt->len,
1050 req = nvme_allocate_request_vaddr(pt->buf, pt->len,
1053 req = nvme_allocate_request_null(nvme_pt_done, pt);
1055 /* Assume user space already converted to little-endian */
1056 req->cmd.opc = pt->cmd.opc;
1057 req->cmd.fuse = pt->cmd.fuse;
1058 req->cmd.rsvd2 = pt->cmd.rsvd2;
1059 req->cmd.rsvd3 = pt->cmd.rsvd3;
1060 req->cmd.cdw10 = pt->cmd.cdw10;
1061 req->cmd.cdw11 = pt->cmd.cdw11;
1062 req->cmd.cdw12 = pt->cmd.cdw12;
1063 req->cmd.cdw13 = pt->cmd.cdw13;
1064 req->cmd.cdw14 = pt->cmd.cdw14;
1065 req->cmd.cdw15 = pt->cmd.cdw15;
1067 req->cmd.nsid = htole32(nsid);
1069 mtx = mtx_pool_find(mtxpool_sleep, pt);
1070 pt->driver_lock = mtx;
1073 nvme_ctrlr_submit_admin_request(ctrlr, req);
1075 nvme_ctrlr_submit_io_request(ctrlr, req);
1078 while (pt->driver_lock != NULL)
1079 mtx_sleep(pt, mtx, PRIBIO, "nvme_pt", 0);
1092 nvme_ctrlr_ioctl(struct cdev *cdev, u_long cmd, caddr_t arg, int flag,
1095 struct nvme_controller *ctrlr;
1096 struct nvme_pt_command *pt;
1098 ctrlr = cdev->si_drv1;
1101 case NVME_RESET_CONTROLLER:
1102 nvme_ctrlr_reset(ctrlr);
1104 case NVME_PASSTHROUGH_CMD:
1105 pt = (struct nvme_pt_command *)arg;
1106 return (nvme_ctrlr_passthrough_cmd(ctrlr, pt, le32toh(pt->cmd.nsid),
1107 1 /* is_user_buffer */, 1 /* is_admin_cmd */));
1110 struct nvme_get_nsid *gnsid = (struct nvme_get_nsid *)arg;
1111 strncpy(gnsid->cdev, device_get_nameunit(ctrlr->dev),
1112 sizeof(gnsid->cdev));
1123 static struct cdevsw nvme_ctrlr_cdevsw = {
1124 .d_version = D_VERSION,
1126 .d_ioctl = nvme_ctrlr_ioctl
1130 nvme_ctrlr_construct(struct nvme_controller *ctrlr, device_t dev)
1132 struct make_dev_args md_args;
1137 int status, timeout_period;
1141 mtx_init(&ctrlr->lock, "nvme ctrlr lock", NULL, MTX_DEF);
1142 if (bus_get_domain(dev, &ctrlr->domain) != 0)
1145 cap_hi = nvme_mmio_read_4(ctrlr, cap_hi);
1146 ctrlr->dstrd = NVME_CAP_HI_DSTRD(cap_hi) + 2;
1148 mpsmin = NVME_CAP_HI_MPSMIN(cap_hi);
1149 ctrlr->min_page_size = 1 << (12 + mpsmin);
1151 /* Get ready timeout value from controller, in units of 500ms. */
1152 cap_lo = nvme_mmio_read_4(ctrlr, cap_lo);
1153 to = NVME_CAP_LO_TO(cap_lo) + 1;
1154 ctrlr->ready_timeout_in_ms = to * 500;
1156 timeout_period = NVME_DEFAULT_TIMEOUT_PERIOD;
1157 TUNABLE_INT_FETCH("hw.nvme.timeout_period", &timeout_period);
1158 timeout_period = min(timeout_period, NVME_MAX_TIMEOUT_PERIOD);
1159 timeout_period = max(timeout_period, NVME_MIN_TIMEOUT_PERIOD);
1160 ctrlr->timeout_period = timeout_period;
1162 nvme_retry_count = NVME_DEFAULT_RETRY_COUNT;
1163 TUNABLE_INT_FETCH("hw.nvme.retry_count", &nvme_retry_count);
1165 ctrlr->enable_aborts = 0;
1166 TUNABLE_INT_FETCH("hw.nvme.enable_aborts", &ctrlr->enable_aborts);
1168 ctrlr->max_xfer_size = NVME_MAX_XFER_SIZE;
1169 if (nvme_ctrlr_construct_admin_qpair(ctrlr) != 0)
1172 ctrlr->taskqueue = taskqueue_create("nvme_taskq", M_WAITOK,
1173 taskqueue_thread_enqueue, &ctrlr->taskqueue);
1174 taskqueue_start_threads(&ctrlr->taskqueue, 1, PI_DISK, "nvme taskq");
1176 ctrlr->is_resetting = 0;
1177 ctrlr->is_initialized = 0;
1178 ctrlr->notification_sent = 0;
1179 TASK_INIT(&ctrlr->reset_task, 0, nvme_ctrlr_reset_task, ctrlr);
1180 TASK_INIT(&ctrlr->fail_req_task, 0, nvme_ctrlr_fail_req_task, ctrlr);
1181 STAILQ_INIT(&ctrlr->fail_req);
1182 ctrlr->is_failed = FALSE;
1184 make_dev_args_init(&md_args);
1185 md_args.mda_devsw = &nvme_ctrlr_cdevsw;
1186 md_args.mda_uid = UID_ROOT;
1187 md_args.mda_gid = GID_WHEEL;
1188 md_args.mda_mode = 0600;
1189 md_args.mda_unit = device_get_unit(dev);
1190 md_args.mda_si_drv1 = (void *)ctrlr;
1191 status = make_dev_s(&md_args, &ctrlr->cdev, "nvme%d",
1192 device_get_unit(dev));
1200 nvme_ctrlr_destruct(struct nvme_controller *ctrlr, device_t dev)
1204 if (ctrlr->resource == NULL)
1208 * Check whether it is a hot unplug or a clean driver detach.
1209 * If device is not there any more, skip any shutdown commands.
1211 gone = (nvme_mmio_read_4(ctrlr, csts) == 0xffffffff);
1213 nvme_ctrlr_fail(ctrlr);
1215 nvme_notify_fail_consumers(ctrlr);
1217 for (i = 0; i < NVME_MAX_NAMESPACES; i++)
1218 nvme_ns_destruct(&ctrlr->ns[i]);
1221 destroy_dev(ctrlr->cdev);
1223 if (ctrlr->is_initialized) {
1225 nvme_ctrlr_delete_qpairs(ctrlr);
1226 for (i = 0; i < ctrlr->num_io_queues; i++)
1227 nvme_io_qpair_destroy(&ctrlr->ioq[i]);
1228 free(ctrlr->ioq, M_NVME);
1229 nvme_admin_qpair_destroy(&ctrlr->adminq);
1233 * Notify the controller of a shutdown, even though this is due to
1234 * a driver unload, not a system shutdown (this path is not invoked
1235 * during shutdown). This ensures the controller receives a
1236 * shutdown notification in case the system is shutdown before
1237 * reloading the driver.
1240 nvme_ctrlr_shutdown(ctrlr);
1243 nvme_ctrlr_disable(ctrlr);
1245 if (ctrlr->taskqueue)
1246 taskqueue_free(ctrlr->taskqueue);
1249 bus_teardown_intr(ctrlr->dev, ctrlr->res, ctrlr->tag);
1252 bus_release_resource(ctrlr->dev, SYS_RES_IRQ,
1253 rman_get_rid(ctrlr->res), ctrlr->res);
1255 if (ctrlr->bar4_resource != NULL) {
1256 bus_release_resource(dev, SYS_RES_MEMORY,
1257 ctrlr->bar4_resource_id, ctrlr->bar4_resource);
1260 bus_release_resource(dev, SYS_RES_MEMORY,
1261 ctrlr->resource_id, ctrlr->resource);
1264 mtx_destroy(&ctrlr->lock);
1268 nvme_ctrlr_shutdown(struct nvme_controller *ctrlr)
1274 cc = nvme_mmio_read_4(ctrlr, cc);
1275 cc &= ~(NVME_CC_REG_SHN_MASK << NVME_CC_REG_SHN_SHIFT);
1276 cc |= NVME_SHN_NORMAL << NVME_CC_REG_SHN_SHIFT;
1277 nvme_mmio_write_4(ctrlr, cc, cc);
1280 csts = nvme_mmio_read_4(ctrlr, csts);
1281 if (csts == 0xffffffff) /* Hot unplug. */
1283 if (NVME_CSTS_GET_SHST(csts) == NVME_SHST_COMPLETE)
1285 if (ticks++ > 5*hz) {
1286 nvme_printf(ctrlr, "did not complete shutdown within"
1287 " 5 seconds of notification\n");
1290 pause("nvme shn", 1);
1295 nvme_ctrlr_submit_admin_request(struct nvme_controller *ctrlr,
1296 struct nvme_request *req)
1299 nvme_qpair_submit_request(&ctrlr->adminq, req);
1303 nvme_ctrlr_submit_io_request(struct nvme_controller *ctrlr,
1304 struct nvme_request *req)
1306 struct nvme_qpair *qpair;
1308 qpair = &ctrlr->ioq[QP(ctrlr, curcpu)];
1309 nvme_qpair_submit_request(qpair, req);
1313 nvme_ctrlr_get_device(struct nvme_controller *ctrlr)
1316 return (ctrlr->dev);
1319 const struct nvme_controller_data *
1320 nvme_ctrlr_get_data(struct nvme_controller *ctrlr)
1323 return (&ctrlr->cdata);
1327 nvme_ctrlr_suspend(struct nvme_controller *ctrlr)
1332 * Can't touch failed controllers, so it's already suspended.
1334 if (ctrlr->is_failed)
1338 * We don't want the reset taskqueue running, since it does similar
1339 * things, so prevent it from running after we start. Wait for any reset
1340 * that may have been started to complete. The reset process we follow
1341 * will ensure that any new I/O will queue and be given to the hardware
1342 * after we resume (though there should be none).
1344 while (atomic_cmpset_32(&ctrlr->is_resetting, 0, 1) == 0 && to-- > 0)
1345 pause("nvmesusp", 1);
1348 "Competing reset task didn't finish. Try again later.\n");
1349 return (EWOULDBLOCK);
1353 * Per Section 7.6.2 of NVMe spec 1.4, to properly suspend, we need to
1354 * delete the hardware I/O queues, and then shutdown. This properly
1355 * flushes any metadata the drive may have stored so it can survive
1356 * having its power removed and prevents the unsafe shutdown count from
1357 * incriminating. Once we delete the qpairs, we have to disable them
1358 * before shutting down. The delay is out of paranoia in
1359 * nvme_ctrlr_hw_reset, and is repeated here (though we should have no
1360 * pending I/O that the delay copes with).
1362 nvme_ctrlr_delete_qpairs(ctrlr);
1363 nvme_ctrlr_disable_qpairs(ctrlr);
1365 nvme_ctrlr_shutdown(ctrlr);
1371 nvme_ctrlr_resume(struct nvme_controller *ctrlr)
1375 * Can't touch failed controllers, so nothing to do to resume.
1377 if (ctrlr->is_failed)
1381 * Have to reset the hardware twice, just like we do on attach. See
1382 * nmve_attach() for why.
1384 if (nvme_ctrlr_hw_reset(ctrlr) != 0)
1386 if (nvme_ctrlr_hw_reset(ctrlr) != 0)
1390 * Now that we're reset the hardware, we can restart the controller. Any
1391 * I/O that was pending is requeued. Any admin commands are aborted with
1392 * an error. Once we've restarted, take the controller out of reset.
1394 nvme_ctrlr_start(ctrlr, true);
1395 atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
1400 * Since we can't bring the controller out of reset, announce and fail
1401 * the controller. However, we have to return success for the resume
1402 * itself, due to questionable APIs.
1404 nvme_printf(ctrlr, "Failed to reset on resume, failing.\n");
1405 nvme_ctrlr_fail(ctrlr);
1406 atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);