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>
45 #include <dev/pci/pcireg.h>
46 #include <dev/pci/pcivar.h>
48 #include "nvme_private.h"
50 #define B4_CHK_RDY_DELAY_MS 2300 /* work around controller bug */
52 static void nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
53 struct nvme_async_event_request *aer);
54 static void nvme_ctrlr_setup_interrupts(struct nvme_controller *ctrlr);
57 nvme_ctrlr_allocate_bar(struct nvme_controller *ctrlr)
60 ctrlr->resource_id = PCIR_BAR(0);
62 ctrlr->resource = bus_alloc_resource_any(ctrlr->dev, SYS_RES_MEMORY,
63 &ctrlr->resource_id, RF_ACTIVE);
65 if(ctrlr->resource == NULL) {
66 nvme_printf(ctrlr, "unable to allocate pci resource\n");
70 ctrlr->bus_tag = rman_get_bustag(ctrlr->resource);
71 ctrlr->bus_handle = rman_get_bushandle(ctrlr->resource);
72 ctrlr->regs = (struct nvme_registers *)ctrlr->bus_handle;
75 * The NVMe spec allows for the MSI-X table to be placed behind
76 * BAR 4/5, separate from the control/doorbell registers. Always
77 * try to map this bar, because it must be mapped prior to calling
78 * pci_alloc_msix(). If the table isn't behind BAR 4/5,
79 * bus_alloc_resource() will just return NULL which is OK.
81 ctrlr->bar4_resource_id = PCIR_BAR(4);
82 ctrlr->bar4_resource = bus_alloc_resource_any(ctrlr->dev, SYS_RES_MEMORY,
83 &ctrlr->bar4_resource_id, RF_ACTIVE);
89 nvme_ctrlr_construct_admin_qpair(struct nvme_controller *ctrlr)
91 struct nvme_qpair *qpair;
95 qpair = &ctrlr->adminq;
97 num_entries = NVME_ADMIN_ENTRIES;
98 TUNABLE_INT_FETCH("hw.nvme.admin_entries", &num_entries);
100 * If admin_entries was overridden to an invalid value, revert it
101 * back to our default value.
103 if (num_entries < NVME_MIN_ADMIN_ENTRIES ||
104 num_entries > NVME_MAX_ADMIN_ENTRIES) {
105 nvme_printf(ctrlr, "invalid hw.nvme.admin_entries=%d "
106 "specified\n", num_entries);
107 num_entries = NVME_ADMIN_ENTRIES;
111 * The admin queue's max xfer size is treated differently than the
112 * max I/O xfer size. 16KB is sufficient here - maybe even less?
114 error = nvme_qpair_construct(qpair,
124 nvme_ctrlr_construct_io_qpairs(struct nvme_controller *ctrlr)
126 struct nvme_qpair *qpair;
129 int i, error, num_entries, num_trackers;
131 num_entries = NVME_IO_ENTRIES;
132 TUNABLE_INT_FETCH("hw.nvme.io_entries", &num_entries);
135 * NVMe spec sets a hard limit of 64K max entries, but
136 * devices may specify a smaller limit, so we need to check
137 * the MQES field in the capabilities register.
139 cap_lo = nvme_mmio_read_4(ctrlr, cap_lo);
140 mqes = NVME_CAP_LO_MQES(cap_lo);
141 num_entries = min(num_entries, mqes + 1);
143 num_trackers = NVME_IO_TRACKERS;
144 TUNABLE_INT_FETCH("hw.nvme.io_trackers", &num_trackers);
146 num_trackers = max(num_trackers, NVME_MIN_IO_TRACKERS);
147 num_trackers = min(num_trackers, NVME_MAX_IO_TRACKERS);
149 * No need to have more trackers than entries in the submit queue.
150 * Note also that for a queue size of N, we can only have (N-1)
151 * commands outstanding, hence the "-1" here.
153 num_trackers = min(num_trackers, (num_entries-1));
156 * Our best estimate for the maximum number of I/Os that we should
157 * noramlly have in flight at one time. This should be viewed as a hint,
158 * not a hard limit and will need to be revisitted when the upper layers
159 * of the storage system grows multi-queue support.
161 ctrlr->max_hw_pend_io = num_trackers * ctrlr->num_io_queues * 3 / 4;
164 * This was calculated previously when setting up interrupts, but
165 * a controller could theoretically support fewer I/O queues than
166 * MSI-X vectors. So calculate again here just to be safe.
168 ctrlr->num_cpus_per_ioq = howmany(mp_ncpus, ctrlr->num_io_queues);
170 ctrlr->ioq = malloc(ctrlr->num_io_queues * sizeof(struct nvme_qpair),
171 M_NVME, M_ZERO | M_WAITOK);
173 for (i = 0; i < ctrlr->num_io_queues; i++) {
174 qpair = &ctrlr->ioq[i];
177 * Admin queue has ID=0. IO queues start at ID=1 -
178 * hence the 'i+1' here.
180 * For I/O queues, use the controller-wide max_xfer_size
181 * calculated in nvme_attach().
183 error = nvme_qpair_construct(qpair,
185 ctrlr->msix_enabled ? i+1 : 0, /* vector */
193 * Do not bother binding interrupts if we only have one I/O
194 * interrupt thread for this controller.
196 if (ctrlr->num_io_queues > 1)
197 bus_bind_intr(ctrlr->dev, qpair->res,
198 i * ctrlr->num_cpus_per_ioq);
205 nvme_ctrlr_fail(struct nvme_controller *ctrlr)
209 ctrlr->is_failed = TRUE;
210 nvme_qpair_fail(&ctrlr->adminq);
211 if (ctrlr->ioq != NULL) {
212 for (i = 0; i < ctrlr->num_io_queues; i++)
213 nvme_qpair_fail(&ctrlr->ioq[i]);
215 nvme_notify_fail_consumers(ctrlr);
219 nvme_ctrlr_post_failed_request(struct nvme_controller *ctrlr,
220 struct nvme_request *req)
223 mtx_lock(&ctrlr->lock);
224 STAILQ_INSERT_TAIL(&ctrlr->fail_req, req, stailq);
225 mtx_unlock(&ctrlr->lock);
226 taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->fail_req_task);
230 nvme_ctrlr_fail_req_task(void *arg, int pending)
232 struct nvme_controller *ctrlr = arg;
233 struct nvme_request *req;
235 mtx_lock(&ctrlr->lock);
236 while ((req = STAILQ_FIRST(&ctrlr->fail_req)) != NULL) {
237 STAILQ_REMOVE_HEAD(&ctrlr->fail_req, stailq);
238 mtx_unlock(&ctrlr->lock);
239 nvme_qpair_manual_complete_request(req->qpair, req,
240 NVME_SCT_GENERIC, NVME_SC_ABORTED_BY_REQUEST);
241 mtx_lock(&ctrlr->lock);
243 mtx_unlock(&ctrlr->lock);
247 nvme_ctrlr_wait_for_ready(struct nvme_controller *ctrlr, int desired_val)
252 csts = nvme_mmio_read_4(ctrlr, csts);
255 while (((csts >> NVME_CSTS_REG_RDY_SHIFT) & NVME_CSTS_REG_RDY_MASK) != desired_val) {
256 if (ms_waited++ > ctrlr->ready_timeout_in_ms) {
257 nvme_printf(ctrlr, "controller ready did not become %d "
258 "within %d ms\n", desired_val, ctrlr->ready_timeout_in_ms);
262 csts = nvme_mmio_read_4(ctrlr, csts);
269 nvme_ctrlr_disable(struct nvme_controller *ctrlr)
276 cc = nvme_mmio_read_4(ctrlr, cc);
277 csts = nvme_mmio_read_4(ctrlr, csts);
279 en = (cc >> NVME_CC_REG_EN_SHIFT) & NVME_CC_REG_EN_MASK;
280 rdy = (csts >> NVME_CSTS_REG_RDY_SHIFT) & NVME_CSTS_REG_RDY_MASK;
283 * Per 3.1.5 in NVME 1.3 spec, transitioning CC.EN from 0 to 1
284 * when CSTS.RDY is 1 or transitioning CC.EN from 1 to 0 when
285 * CSTS.RDY is 0 "has undefined results" So make sure that CSTS.RDY
286 * isn't the desired value. Short circuit if we're already disabled.
290 /* EN == 1, wait for RDY == 1 or fail */
291 err = nvme_ctrlr_wait_for_ready(ctrlr, 1);
296 /* EN == 0 already wait for RDY == 0 */
300 return (nvme_ctrlr_wait_for_ready(ctrlr, 0));
303 cc &= ~NVME_CC_REG_EN_MASK;
304 nvme_mmio_write_4(ctrlr, cc, cc);
306 * Some drives have issues with accessing the mmio after we
307 * disable, so delay for a bit after we write the bit to
308 * cope with these issues.
310 if (ctrlr->quirks & QUIRK_DELAY_B4_CHK_RDY)
311 pause("nvmeR", B4_CHK_RDY_DELAY_MS * hz / 1000);
312 return (nvme_ctrlr_wait_for_ready(ctrlr, 0));
316 nvme_ctrlr_enable(struct nvme_controller *ctrlr)
325 cc = nvme_mmio_read_4(ctrlr, cc);
326 csts = nvme_mmio_read_4(ctrlr, csts);
328 en = (cc >> NVME_CC_REG_EN_SHIFT) & NVME_CC_REG_EN_MASK;
329 rdy = (csts >> NVME_CSTS_REG_RDY_SHIFT) & NVME_CSTS_REG_RDY_MASK;
332 * See note in nvme_ctrlr_disable. Short circuit if we're already enabled.
338 return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
340 /* EN == 0 already wait for RDY == 0 or fail */
341 err = nvme_ctrlr_wait_for_ready(ctrlr, 0);
346 nvme_mmio_write_8(ctrlr, asq, ctrlr->adminq.cmd_bus_addr);
348 nvme_mmio_write_8(ctrlr, acq, ctrlr->adminq.cpl_bus_addr);
351 /* acqs and asqs are 0-based. */
352 qsize = ctrlr->adminq.num_entries - 1;
355 aqa = (qsize & NVME_AQA_REG_ACQS_MASK) << NVME_AQA_REG_ACQS_SHIFT;
356 aqa |= (qsize & NVME_AQA_REG_ASQS_MASK) << NVME_AQA_REG_ASQS_SHIFT;
357 nvme_mmio_write_4(ctrlr, aqa, aqa);
360 /* Initialization values for CC */
362 cc |= 1 << NVME_CC_REG_EN_SHIFT;
363 cc |= 0 << NVME_CC_REG_CSS_SHIFT;
364 cc |= 0 << NVME_CC_REG_AMS_SHIFT;
365 cc |= 0 << NVME_CC_REG_SHN_SHIFT;
366 cc |= 6 << NVME_CC_REG_IOSQES_SHIFT; /* SQ entry size == 64 == 2^6 */
367 cc |= 4 << NVME_CC_REG_IOCQES_SHIFT; /* CQ entry size == 16 == 2^4 */
369 /* This evaluates to 0, which is according to spec. */
370 cc |= (PAGE_SIZE >> 13) << NVME_CC_REG_MPS_SHIFT;
372 nvme_mmio_write_4(ctrlr, cc, cc);
374 return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
378 nvme_ctrlr_hw_reset(struct nvme_controller *ctrlr)
382 nvme_admin_qpair_disable(&ctrlr->adminq);
384 * I/O queues are not allocated before the initial HW
385 * reset, so do not try to disable them. Use is_initialized
386 * to determine if this is the initial HW reset.
388 if (ctrlr->is_initialized) {
389 for (i = 0; i < ctrlr->num_io_queues; i++)
390 nvme_io_qpair_disable(&ctrlr->ioq[i]);
395 err = nvme_ctrlr_disable(ctrlr);
398 return (nvme_ctrlr_enable(ctrlr));
402 nvme_ctrlr_reset(struct nvme_controller *ctrlr)
406 cmpset = atomic_cmpset_32(&ctrlr->is_resetting, 0, 1);
408 if (cmpset == 0 || ctrlr->is_failed)
410 * Controller is already resetting or has failed. Return
411 * immediately since there is no need to kick off another
412 * reset in these cases.
416 taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->reset_task);
420 nvme_ctrlr_identify(struct nvme_controller *ctrlr)
422 struct nvme_completion_poll_status status;
425 nvme_ctrlr_cmd_identify_controller(ctrlr, &ctrlr->cdata,
426 nvme_completion_poll_cb, &status);
427 while (!atomic_load_acq_int(&status.done))
429 if (nvme_completion_is_error(&status.cpl)) {
430 nvme_printf(ctrlr, "nvme_identify_controller failed!\n");
434 /* Convert data to host endian */
435 nvme_controller_data_swapbytes(&ctrlr->cdata);
438 * Use MDTS to ensure our default max_xfer_size doesn't exceed what the
439 * controller supports.
441 if (ctrlr->cdata.mdts > 0)
442 ctrlr->max_xfer_size = min(ctrlr->max_xfer_size,
443 ctrlr->min_page_size * (1 << (ctrlr->cdata.mdts)));
449 nvme_ctrlr_set_num_qpairs(struct nvme_controller *ctrlr)
451 struct nvme_completion_poll_status status;
452 int cq_allocated, sq_allocated;
455 nvme_ctrlr_cmd_set_num_queues(ctrlr, ctrlr->num_io_queues,
456 nvme_completion_poll_cb, &status);
457 while (!atomic_load_acq_int(&status.done))
459 if (nvme_completion_is_error(&status.cpl)) {
460 nvme_printf(ctrlr, "nvme_ctrlr_set_num_qpairs failed!\n");
465 * Data in cdw0 is 0-based.
466 * Lower 16-bits indicate number of submission queues allocated.
467 * Upper 16-bits indicate number of completion queues allocated.
469 sq_allocated = (status.cpl.cdw0 & 0xFFFF) + 1;
470 cq_allocated = (status.cpl.cdw0 >> 16) + 1;
473 * Controller may allocate more queues than we requested,
474 * so use the minimum of the number requested and what was
475 * actually allocated.
477 ctrlr->num_io_queues = min(ctrlr->num_io_queues, sq_allocated);
478 ctrlr->num_io_queues = min(ctrlr->num_io_queues, cq_allocated);
484 nvme_ctrlr_create_qpairs(struct nvme_controller *ctrlr)
486 struct nvme_completion_poll_status status;
487 struct nvme_qpair *qpair;
490 for (i = 0; i < ctrlr->num_io_queues; i++) {
491 qpair = &ctrlr->ioq[i];
494 nvme_ctrlr_cmd_create_io_cq(ctrlr, qpair, qpair->vector,
495 nvme_completion_poll_cb, &status);
496 while (!atomic_load_acq_int(&status.done))
498 if (nvme_completion_is_error(&status.cpl)) {
499 nvme_printf(ctrlr, "nvme_create_io_cq failed!\n");
504 nvme_ctrlr_cmd_create_io_sq(qpair->ctrlr, qpair,
505 nvme_completion_poll_cb, &status);
506 while (!atomic_load_acq_int(&status.done))
508 if (nvme_completion_is_error(&status.cpl)) {
509 nvme_printf(ctrlr, "nvme_create_io_sq failed!\n");
518 nvme_ctrlr_destroy_qpairs(struct nvme_controller *ctrlr)
520 struct nvme_completion_poll_status status;
521 struct nvme_qpair *qpair;
523 for (int i = 0; i < ctrlr->num_io_queues; i++) {
524 qpair = &ctrlr->ioq[i];
527 nvme_ctrlr_cmd_delete_io_sq(ctrlr, qpair,
528 nvme_completion_poll_cb, &status);
529 while (!atomic_load_acq_int(&status.done))
531 if (nvme_completion_is_error(&status.cpl)) {
532 nvme_printf(ctrlr, "nvme_destroy_io_sq failed!\n");
537 nvme_ctrlr_cmd_delete_io_cq(ctrlr, qpair,
538 nvme_completion_poll_cb, &status);
539 while (!atomic_load_acq_int(&status.done))
541 if (nvme_completion_is_error(&status.cpl)) {
542 nvme_printf(ctrlr, "nvme_destroy_io_cq failed!\n");
551 nvme_ctrlr_construct_namespaces(struct nvme_controller *ctrlr)
553 struct nvme_namespace *ns;
556 for (i = 0; i < min(ctrlr->cdata.nn, NVME_MAX_NAMESPACES); i++) {
558 nvme_ns_construct(ns, i+1, ctrlr);
565 is_log_page_id_valid(uint8_t page_id)
570 case NVME_LOG_HEALTH_INFORMATION:
571 case NVME_LOG_FIRMWARE_SLOT:
572 case NVME_LOG_CHANGED_NAMESPACE:
573 case NVME_LOG_COMMAND_EFFECT:
574 case NVME_LOG_RES_NOTIFICATION:
575 case NVME_LOG_SANITIZE_STATUS:
583 nvme_ctrlr_get_log_page_size(struct nvme_controller *ctrlr, uint8_t page_id)
585 uint32_t log_page_size;
590 sizeof(struct nvme_error_information_entry) *
591 (ctrlr->cdata.elpe + 1), NVME_MAX_AER_LOG_SIZE);
593 case NVME_LOG_HEALTH_INFORMATION:
594 log_page_size = sizeof(struct nvme_health_information_page);
596 case NVME_LOG_FIRMWARE_SLOT:
597 log_page_size = sizeof(struct nvme_firmware_page);
599 case NVME_LOG_CHANGED_NAMESPACE:
600 log_page_size = sizeof(struct nvme_ns_list);
602 case NVME_LOG_COMMAND_EFFECT:
603 log_page_size = sizeof(struct nvme_command_effects_page);
605 case NVME_LOG_RES_NOTIFICATION:
606 log_page_size = sizeof(struct nvme_res_notification_page);
608 case NVME_LOG_SANITIZE_STATUS:
609 log_page_size = sizeof(struct nvme_sanitize_status_page);
616 return (log_page_size);
620 nvme_ctrlr_log_critical_warnings(struct nvme_controller *ctrlr,
624 if (state & NVME_CRIT_WARN_ST_AVAILABLE_SPARE)
625 nvme_printf(ctrlr, "available spare space below threshold\n");
627 if (state & NVME_CRIT_WARN_ST_TEMPERATURE)
628 nvme_printf(ctrlr, "temperature above threshold\n");
630 if (state & NVME_CRIT_WARN_ST_DEVICE_RELIABILITY)
631 nvme_printf(ctrlr, "device reliability degraded\n");
633 if (state & NVME_CRIT_WARN_ST_READ_ONLY)
634 nvme_printf(ctrlr, "media placed in read only mode\n");
636 if (state & NVME_CRIT_WARN_ST_VOLATILE_MEMORY_BACKUP)
637 nvme_printf(ctrlr, "volatile memory backup device failed\n");
639 if (state & NVME_CRIT_WARN_ST_RESERVED_MASK)
641 "unknown critical warning(s): state = 0x%02x\n", state);
645 nvme_ctrlr_async_event_log_page_cb(void *arg, const struct nvme_completion *cpl)
647 struct nvme_async_event_request *aer = arg;
648 struct nvme_health_information_page *health_info;
649 struct nvme_ns_list *nsl;
650 struct nvme_error_information_entry *err;
654 * If the log page fetch for some reason completed with an error,
655 * don't pass log page data to the consumers. In practice, this case
656 * should never happen.
658 if (nvme_completion_is_error(cpl))
659 nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
660 aer->log_page_id, NULL, 0);
662 /* Convert data to host endian */
663 switch (aer->log_page_id) {
665 err = (struct nvme_error_information_entry *)aer->log_page_buffer;
666 for (i = 0; i < (aer->ctrlr->cdata.elpe + 1); i++)
667 nvme_error_information_entry_swapbytes(err++);
669 case NVME_LOG_HEALTH_INFORMATION:
670 nvme_health_information_page_swapbytes(
671 (struct nvme_health_information_page *)aer->log_page_buffer);
673 case NVME_LOG_FIRMWARE_SLOT:
674 nvme_firmware_page_swapbytes(
675 (struct nvme_firmware_page *)aer->log_page_buffer);
677 case NVME_LOG_CHANGED_NAMESPACE:
678 nvme_ns_list_swapbytes(
679 (struct nvme_ns_list *)aer->log_page_buffer);
681 case NVME_LOG_COMMAND_EFFECT:
682 nvme_command_effects_page_swapbytes(
683 (struct nvme_command_effects_page *)aer->log_page_buffer);
685 case NVME_LOG_RES_NOTIFICATION:
686 nvme_res_notification_page_swapbytes(
687 (struct nvme_res_notification_page *)aer->log_page_buffer);
689 case NVME_LOG_SANITIZE_STATUS:
690 nvme_sanitize_status_page_swapbytes(
691 (struct nvme_sanitize_status_page *)aer->log_page_buffer);
693 case INTEL_LOG_TEMP_STATS:
694 intel_log_temp_stats_swapbytes(
695 (struct intel_log_temp_stats *)aer->log_page_buffer);
701 if (aer->log_page_id == NVME_LOG_HEALTH_INFORMATION) {
702 health_info = (struct nvme_health_information_page *)
703 aer->log_page_buffer;
704 nvme_ctrlr_log_critical_warnings(aer->ctrlr,
705 health_info->critical_warning);
707 * Critical warnings reported through the
708 * SMART/health log page are persistent, so
709 * clear the associated bits in the async event
710 * config so that we do not receive repeated
711 * notifications for the same event.
713 aer->ctrlr->async_event_config &=
714 ~health_info->critical_warning;
715 nvme_ctrlr_cmd_set_async_event_config(aer->ctrlr,
716 aer->ctrlr->async_event_config, NULL, NULL);
717 } else if (aer->log_page_id == NVME_LOG_CHANGED_NAMESPACE &&
719 nsl = (struct nvme_ns_list *)aer->log_page_buffer;
720 for (i = 0; i < nitems(nsl->ns) && nsl->ns[i] != 0; i++) {
721 if (nsl->ns[i] > NVME_MAX_NAMESPACES)
723 nvme_notify_ns(aer->ctrlr, nsl->ns[i]);
729 * Pass the cpl data from the original async event completion,
730 * not the log page fetch.
732 nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
733 aer->log_page_id, aer->log_page_buffer, aer->log_page_size);
737 * Repost another asynchronous event request to replace the one
738 * that just completed.
740 nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
744 nvme_ctrlr_async_event_cb(void *arg, const struct nvme_completion *cpl)
746 struct nvme_async_event_request *aer = arg;
748 if (nvme_completion_is_error(cpl)) {
750 * Do not retry failed async event requests. This avoids
751 * infinite loops where a new async event request is submitted
752 * to replace the one just failed, only to fail again and
753 * perpetuate the loop.
758 /* Associated log page is in bits 23:16 of completion entry dw0. */
759 aer->log_page_id = (cpl->cdw0 & 0xFF0000) >> 16;
761 nvme_printf(aer->ctrlr, "async event occurred (type 0x%x, info 0x%02x,"
762 " page 0x%02x)\n", (cpl->cdw0 & 0x07), (cpl->cdw0 & 0xFF00) >> 8,
765 if (is_log_page_id_valid(aer->log_page_id)) {
766 aer->log_page_size = nvme_ctrlr_get_log_page_size(aer->ctrlr,
768 memcpy(&aer->cpl, cpl, sizeof(*cpl));
769 nvme_ctrlr_cmd_get_log_page(aer->ctrlr, aer->log_page_id,
770 NVME_GLOBAL_NAMESPACE_TAG, aer->log_page_buffer,
771 aer->log_page_size, nvme_ctrlr_async_event_log_page_cb,
773 /* Wait to notify consumers until after log page is fetched. */
775 nvme_notify_async_consumers(aer->ctrlr, cpl, aer->log_page_id,
779 * Repost another asynchronous event request to replace the one
780 * that just completed.
782 nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
787 nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
788 struct nvme_async_event_request *aer)
790 struct nvme_request *req;
793 req = nvme_allocate_request_null(nvme_ctrlr_async_event_cb, aer);
797 * Disable timeout here, since asynchronous event requests should by
798 * nature never be timed out.
800 req->timeout = FALSE;
801 req->cmd.opc = NVME_OPC_ASYNC_EVENT_REQUEST;
802 nvme_ctrlr_submit_admin_request(ctrlr, req);
806 nvme_ctrlr_configure_aer(struct nvme_controller *ctrlr)
808 struct nvme_completion_poll_status status;
809 struct nvme_async_event_request *aer;
812 ctrlr->async_event_config = NVME_CRIT_WARN_ST_AVAILABLE_SPARE |
813 NVME_CRIT_WARN_ST_DEVICE_RELIABILITY |
814 NVME_CRIT_WARN_ST_READ_ONLY |
815 NVME_CRIT_WARN_ST_VOLATILE_MEMORY_BACKUP;
816 if (ctrlr->cdata.ver >= NVME_REV(1, 2))
817 ctrlr->async_event_config |= 0x300;
820 nvme_ctrlr_cmd_get_feature(ctrlr, NVME_FEAT_TEMPERATURE_THRESHOLD,
821 0, NULL, 0, nvme_completion_poll_cb, &status);
822 while (!atomic_load_acq_int(&status.done))
824 if (nvme_completion_is_error(&status.cpl) ||
825 (status.cpl.cdw0 & 0xFFFF) == 0xFFFF ||
826 (status.cpl.cdw0 & 0xFFFF) == 0x0000) {
827 nvme_printf(ctrlr, "temperature threshold not supported\n");
829 ctrlr->async_event_config |= NVME_CRIT_WARN_ST_TEMPERATURE;
831 nvme_ctrlr_cmd_set_async_event_config(ctrlr,
832 ctrlr->async_event_config, NULL, NULL);
834 /* aerl is a zero-based value, so we need to add 1 here. */
835 ctrlr->num_aers = min(NVME_MAX_ASYNC_EVENTS, (ctrlr->cdata.aerl+1));
837 for (i = 0; i < ctrlr->num_aers; i++) {
838 aer = &ctrlr->aer[i];
839 nvme_ctrlr_construct_and_submit_aer(ctrlr, aer);
844 nvme_ctrlr_configure_int_coalescing(struct nvme_controller *ctrlr)
847 ctrlr->int_coal_time = 0;
848 TUNABLE_INT_FETCH("hw.nvme.int_coal_time",
849 &ctrlr->int_coal_time);
851 ctrlr->int_coal_threshold = 0;
852 TUNABLE_INT_FETCH("hw.nvme.int_coal_threshold",
853 &ctrlr->int_coal_threshold);
855 nvme_ctrlr_cmd_set_interrupt_coalescing(ctrlr, ctrlr->int_coal_time,
856 ctrlr->int_coal_threshold, NULL, NULL);
860 nvme_ctrlr_start(void *ctrlr_arg)
862 struct nvme_controller *ctrlr = ctrlr_arg;
863 uint32_t old_num_io_queues;
867 * Only reset adminq here when we are restarting the
868 * controller after a reset. During initialization,
869 * we have already submitted admin commands to get
870 * the number of I/O queues supported, so cannot reset
871 * the adminq again here.
873 if (ctrlr->is_resetting) {
874 nvme_qpair_reset(&ctrlr->adminq);
877 for (i = 0; i < ctrlr->num_io_queues; i++)
878 nvme_qpair_reset(&ctrlr->ioq[i]);
880 nvme_admin_qpair_enable(&ctrlr->adminq);
882 if (nvme_ctrlr_identify(ctrlr) != 0) {
883 nvme_ctrlr_fail(ctrlr);
888 * The number of qpairs are determined during controller initialization,
889 * including using NVMe SET_FEATURES/NUMBER_OF_QUEUES to determine the
890 * HW limit. We call SET_FEATURES again here so that it gets called
891 * after any reset for controllers that depend on the driver to
892 * explicit specify how many queues it will use. This value should
893 * never change between resets, so panic if somehow that does happen.
895 if (ctrlr->is_resetting) {
896 old_num_io_queues = ctrlr->num_io_queues;
897 if (nvme_ctrlr_set_num_qpairs(ctrlr) != 0) {
898 nvme_ctrlr_fail(ctrlr);
902 if (old_num_io_queues != ctrlr->num_io_queues) {
903 panic("num_io_queues changed from %u to %u",
904 old_num_io_queues, ctrlr->num_io_queues);
908 if (nvme_ctrlr_create_qpairs(ctrlr) != 0) {
909 nvme_ctrlr_fail(ctrlr);
913 if (nvme_ctrlr_construct_namespaces(ctrlr) != 0) {
914 nvme_ctrlr_fail(ctrlr);
918 nvme_ctrlr_configure_aer(ctrlr);
919 nvme_ctrlr_configure_int_coalescing(ctrlr);
921 for (i = 0; i < ctrlr->num_io_queues; i++)
922 nvme_io_qpair_enable(&ctrlr->ioq[i]);
926 nvme_ctrlr_start_config_hook(void *arg)
928 struct nvme_controller *ctrlr = arg;
930 nvme_qpair_reset(&ctrlr->adminq);
931 nvme_admin_qpair_enable(&ctrlr->adminq);
933 if (nvme_ctrlr_set_num_qpairs(ctrlr) == 0 &&
934 nvme_ctrlr_construct_io_qpairs(ctrlr) == 0)
935 nvme_ctrlr_start(ctrlr);
937 nvme_ctrlr_fail(ctrlr);
939 nvme_sysctl_initialize_ctrlr(ctrlr);
940 config_intrhook_disestablish(&ctrlr->config_hook);
942 ctrlr->is_initialized = 1;
943 nvme_notify_new_controller(ctrlr);
947 nvme_ctrlr_reset_task(void *arg, int pending)
949 struct nvme_controller *ctrlr = arg;
952 nvme_printf(ctrlr, "resetting controller\n");
953 status = nvme_ctrlr_hw_reset(ctrlr);
955 * Use pause instead of DELAY, so that we yield to any nvme interrupt
956 * handlers on this CPU that were blocked on a qpair lock. We want
957 * all nvme interrupts completed before proceeding with restarting the
960 * XXX - any way to guarantee the interrupt handlers have quiesced?
962 pause("nvmereset", hz / 10);
964 nvme_ctrlr_start(ctrlr);
966 nvme_ctrlr_fail(ctrlr);
968 atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
972 * Poll all the queues enabled on the device for completion.
975 nvme_ctrlr_poll(struct nvme_controller *ctrlr)
979 nvme_qpair_process_completions(&ctrlr->adminq);
981 for (i = 0; i < ctrlr->num_io_queues; i++)
982 if (ctrlr->ioq && ctrlr->ioq[i].cpl)
983 nvme_qpair_process_completions(&ctrlr->ioq[i]);
987 * Poll the single-vector intertrupt case: num_io_queues will be 1 and
988 * there's only a single vector. While we're polling, we mask further
989 * interrupts in the controller.
992 nvme_ctrlr_intx_handler(void *arg)
994 struct nvme_controller *ctrlr = arg;
996 nvme_mmio_write_4(ctrlr, intms, 1);
997 nvme_ctrlr_poll(ctrlr);
998 nvme_mmio_write_4(ctrlr, intmc, 1);
1002 nvme_ctrlr_configure_intx(struct nvme_controller *ctrlr)
1005 ctrlr->msix_enabled = 0;
1006 ctrlr->num_io_queues = 1;
1007 ctrlr->num_cpus_per_ioq = mp_ncpus;
1009 ctrlr->res = bus_alloc_resource_any(ctrlr->dev, SYS_RES_IRQ,
1010 &ctrlr->rid, RF_SHAREABLE | RF_ACTIVE);
1012 if (ctrlr->res == NULL) {
1013 nvme_printf(ctrlr, "unable to allocate shared IRQ\n");
1017 bus_setup_intr(ctrlr->dev, ctrlr->res,
1018 INTR_TYPE_MISC | INTR_MPSAFE, NULL, nvme_ctrlr_intx_handler,
1019 ctrlr, &ctrlr->tag);
1021 if (ctrlr->tag == NULL) {
1022 nvme_printf(ctrlr, "unable to setup intx handler\n");
1030 nvme_pt_done(void *arg, const struct nvme_completion *cpl)
1032 struct nvme_pt_command *pt = arg;
1033 struct mtx *mtx = pt->driver_lock;
1036 bzero(&pt->cpl, sizeof(pt->cpl));
1037 pt->cpl.cdw0 = cpl->cdw0;
1039 status = cpl->status;
1040 status &= ~NVME_STATUS_P_MASK;
1041 pt->cpl.status = status;
1044 pt->driver_lock = NULL;
1050 nvme_ctrlr_passthrough_cmd(struct nvme_controller *ctrlr,
1051 struct nvme_pt_command *pt, uint32_t nsid, int is_user_buffer,
1054 struct nvme_request *req;
1056 struct buf *buf = NULL;
1058 vm_offset_t addr, end;
1062 * vmapbuf calls vm_fault_quick_hold_pages which only maps full
1063 * pages. Ensure this request has fewer than MAXPHYS bytes when
1064 * extended to full pages.
1066 addr = (vm_offset_t)pt->buf;
1067 end = round_page(addr + pt->len);
1068 addr = trunc_page(addr);
1069 if (end - addr > MAXPHYS)
1072 if (pt->len > ctrlr->max_xfer_size) {
1073 nvme_printf(ctrlr, "pt->len (%d) "
1074 "exceeds max_xfer_size (%d)\n", pt->len,
1075 ctrlr->max_xfer_size);
1078 if (is_user_buffer) {
1080 * Ensure the user buffer is wired for the duration of
1081 * this passthrough command.
1084 buf = uma_zalloc(pbuf_zone, M_WAITOK);
1085 buf->b_data = pt->buf;
1086 buf->b_bufsize = pt->len;
1087 buf->b_iocmd = pt->is_read ? BIO_READ : BIO_WRITE;
1088 if (vmapbuf(buf, 1) < 0) {
1092 req = nvme_allocate_request_vaddr(buf->b_data, pt->len,
1095 req = nvme_allocate_request_vaddr(pt->buf, pt->len,
1098 req = nvme_allocate_request_null(nvme_pt_done, pt);
1100 /* Assume userspace already converted to little-endian */
1101 req->cmd.opc = pt->cmd.opc;
1102 req->cmd.fuse = pt->cmd.fuse;
1103 req->cmd.rsvd2 = pt->cmd.rsvd2;
1104 req->cmd.rsvd3 = pt->cmd.rsvd3;
1105 req->cmd.cdw10 = pt->cmd.cdw10;
1106 req->cmd.cdw11 = pt->cmd.cdw11;
1107 req->cmd.cdw12 = pt->cmd.cdw12;
1108 req->cmd.cdw13 = pt->cmd.cdw13;
1109 req->cmd.cdw14 = pt->cmd.cdw14;
1110 req->cmd.cdw15 = pt->cmd.cdw15;
1112 req->cmd.nsid = htole32(nsid);
1114 mtx = mtx_pool_find(mtxpool_sleep, pt);
1115 pt->driver_lock = mtx;
1118 nvme_ctrlr_submit_admin_request(ctrlr, req);
1120 nvme_ctrlr_submit_io_request(ctrlr, req);
1123 while (pt->driver_lock != NULL)
1124 mtx_sleep(pt, mtx, PRIBIO, "nvme_pt", 0);
1129 uma_zfree(pbuf_zone, buf);
1137 nvme_ctrlr_ioctl(struct cdev *cdev, u_long cmd, caddr_t arg, int flag,
1140 struct nvme_controller *ctrlr;
1141 struct nvme_pt_command *pt;
1143 ctrlr = cdev->si_drv1;
1146 case NVME_RESET_CONTROLLER:
1147 nvme_ctrlr_reset(ctrlr);
1149 case NVME_PASSTHROUGH_CMD:
1150 pt = (struct nvme_pt_command *)arg;
1151 return (nvme_ctrlr_passthrough_cmd(ctrlr, pt, le32toh(pt->cmd.nsid),
1152 1 /* is_user_buffer */, 1 /* is_admin_cmd */));
1155 struct nvme_get_nsid *gnsid = (struct nvme_get_nsid *)arg;
1156 strncpy(gnsid->cdev, device_get_nameunit(ctrlr->dev),
1157 sizeof(gnsid->cdev));
1168 static struct cdevsw nvme_ctrlr_cdevsw = {
1169 .d_version = D_VERSION,
1171 .d_ioctl = nvme_ctrlr_ioctl
1175 nvme_ctrlr_setup_interrupts(struct nvme_controller *ctrlr)
1178 int per_cpu_io_queues;
1179 int min_cpus_per_ioq;
1180 int num_vectors_requested, num_vectors_allocated;
1181 int num_vectors_available;
1184 min_cpus_per_ioq = 1;
1185 TUNABLE_INT_FETCH("hw.nvme.min_cpus_per_ioq", &min_cpus_per_ioq);
1187 if (min_cpus_per_ioq < 1) {
1188 min_cpus_per_ioq = 1;
1189 } else if (min_cpus_per_ioq > mp_ncpus) {
1190 min_cpus_per_ioq = mp_ncpus;
1193 per_cpu_io_queues = 1;
1194 TUNABLE_INT_FETCH("hw.nvme.per_cpu_io_queues", &per_cpu_io_queues);
1196 if (per_cpu_io_queues == 0) {
1197 min_cpus_per_ioq = mp_ncpus;
1200 ctrlr->force_intx = 0;
1201 TUNABLE_INT_FETCH("hw.nvme.force_intx", &ctrlr->force_intx);
1204 * FreeBSD currently cannot allocate more than about 190 vectors at
1205 * boot, meaning that systems with high core count and many devices
1206 * requesting per-CPU interrupt vectors will not get their full
1207 * allotment. So first, try to allocate as many as we may need to
1208 * understand what is available, then immediately release them.
1209 * Then figure out how many of those we will actually use, based on
1210 * assigning an equal number of cores to each I/O queue.
1213 /* One vector for per core I/O queue, plus one vector for admin queue. */
1214 num_vectors_available = min(pci_msix_count(dev), mp_ncpus + 1);
1215 if (pci_alloc_msix(dev, &num_vectors_available) != 0) {
1216 num_vectors_available = 0;
1218 pci_release_msi(dev);
1220 if (ctrlr->force_intx || num_vectors_available < 2) {
1221 nvme_ctrlr_configure_intx(ctrlr);
1226 * Do not use all vectors for I/O queues - one must be saved for the
1229 ctrlr->num_cpus_per_ioq = max(min_cpus_per_ioq,
1230 howmany(mp_ncpus, num_vectors_available - 1));
1232 ctrlr->num_io_queues = howmany(mp_ncpus, ctrlr->num_cpus_per_ioq);
1233 num_vectors_requested = ctrlr->num_io_queues + 1;
1234 num_vectors_allocated = num_vectors_requested;
1237 * Now just allocate the number of vectors we need. This should
1238 * succeed, since we previously called pci_alloc_msix()
1239 * successfully returning at least this many vectors, but just to
1240 * be safe, if something goes wrong just revert to INTx.
1242 if (pci_alloc_msix(dev, &num_vectors_allocated) != 0) {
1243 nvme_ctrlr_configure_intx(ctrlr);
1247 if (num_vectors_allocated < num_vectors_requested) {
1248 pci_release_msi(dev);
1249 nvme_ctrlr_configure_intx(ctrlr);
1253 ctrlr->msix_enabled = 1;
1257 nvme_ctrlr_construct(struct nvme_controller *ctrlr, device_t dev)
1259 struct make_dev_args md_args;
1265 int status, timeout_period;
1269 mtx_init(&ctrlr->lock, "nvme ctrlr lock", NULL, MTX_DEF);
1271 status = nvme_ctrlr_allocate_bar(ctrlr);
1277 * Software emulators may set the doorbell stride to something
1278 * other than zero, but this driver is not set up to handle that.
1280 cap_hi = nvme_mmio_read_4(ctrlr, cap_hi);
1281 dstrd = NVME_CAP_HI_DSTRD(cap_hi);
1285 mpsmin = NVME_CAP_HI_MPSMIN(cap_hi);
1286 ctrlr->min_page_size = 1 << (12 + mpsmin);
1288 /* Get ready timeout value from controller, in units of 500ms. */
1289 cap_lo = nvme_mmio_read_4(ctrlr, cap_lo);
1290 to = NVME_CAP_LO_TO(cap_lo) + 1;
1291 ctrlr->ready_timeout_in_ms = to * 500;
1293 timeout_period = NVME_DEFAULT_TIMEOUT_PERIOD;
1294 TUNABLE_INT_FETCH("hw.nvme.timeout_period", &timeout_period);
1295 timeout_period = min(timeout_period, NVME_MAX_TIMEOUT_PERIOD);
1296 timeout_period = max(timeout_period, NVME_MIN_TIMEOUT_PERIOD);
1297 ctrlr->timeout_period = timeout_period;
1299 nvme_retry_count = NVME_DEFAULT_RETRY_COUNT;
1300 TUNABLE_INT_FETCH("hw.nvme.retry_count", &nvme_retry_count);
1302 ctrlr->enable_aborts = 0;
1303 TUNABLE_INT_FETCH("hw.nvme.enable_aborts", &ctrlr->enable_aborts);
1305 nvme_ctrlr_setup_interrupts(ctrlr);
1307 ctrlr->max_xfer_size = NVME_MAX_XFER_SIZE;
1308 if (nvme_ctrlr_construct_admin_qpair(ctrlr) != 0)
1311 ctrlr->taskqueue = taskqueue_create("nvme_taskq", M_WAITOK,
1312 taskqueue_thread_enqueue, &ctrlr->taskqueue);
1313 taskqueue_start_threads(&ctrlr->taskqueue, 1, PI_DISK, "nvme taskq");
1315 ctrlr->is_resetting = 0;
1316 ctrlr->is_initialized = 0;
1317 ctrlr->notification_sent = 0;
1318 TASK_INIT(&ctrlr->reset_task, 0, nvme_ctrlr_reset_task, ctrlr);
1319 TASK_INIT(&ctrlr->fail_req_task, 0, nvme_ctrlr_fail_req_task, ctrlr);
1320 STAILQ_INIT(&ctrlr->fail_req);
1321 ctrlr->is_failed = FALSE;
1323 make_dev_args_init(&md_args);
1324 md_args.mda_devsw = &nvme_ctrlr_cdevsw;
1325 md_args.mda_uid = UID_ROOT;
1326 md_args.mda_gid = GID_WHEEL;
1327 md_args.mda_mode = 0600;
1328 md_args.mda_unit = device_get_unit(dev);
1329 md_args.mda_si_drv1 = (void *)ctrlr;
1330 status = make_dev_s(&md_args, &ctrlr->cdev, "nvme%d",
1331 device_get_unit(dev));
1339 nvme_ctrlr_destruct(struct nvme_controller *ctrlr, device_t dev)
1343 if (ctrlr->resource == NULL)
1346 nvme_notify_fail_consumers(ctrlr);
1348 for (i = 0; i < NVME_MAX_NAMESPACES; i++)
1349 nvme_ns_destruct(&ctrlr->ns[i]);
1352 destroy_dev(ctrlr->cdev);
1354 nvme_ctrlr_destroy_qpairs(ctrlr);
1355 for (i = 0; i < ctrlr->num_io_queues; i++) {
1356 nvme_io_qpair_destroy(&ctrlr->ioq[i]);
1358 free(ctrlr->ioq, M_NVME);
1360 nvme_admin_qpair_destroy(&ctrlr->adminq);
1363 * Notify the controller of a shutdown, even though this is due to
1364 * a driver unload, not a system shutdown (this path is not invoked
1365 * during shutdown). This ensures the controller receives a
1366 * shutdown notification in case the system is shutdown before
1367 * reloading the driver.
1369 nvme_ctrlr_shutdown(ctrlr);
1371 nvme_ctrlr_disable(ctrlr);
1373 if (ctrlr->taskqueue)
1374 taskqueue_free(ctrlr->taskqueue);
1377 bus_teardown_intr(ctrlr->dev, ctrlr->res, ctrlr->tag);
1380 bus_release_resource(ctrlr->dev, SYS_RES_IRQ,
1381 rman_get_rid(ctrlr->res), ctrlr->res);
1383 if (ctrlr->msix_enabled)
1384 pci_release_msi(dev);
1386 if (ctrlr->bar4_resource != NULL) {
1387 bus_release_resource(dev, SYS_RES_MEMORY,
1388 ctrlr->bar4_resource_id, ctrlr->bar4_resource);
1391 bus_release_resource(dev, SYS_RES_MEMORY,
1392 ctrlr->resource_id, ctrlr->resource);
1395 mtx_destroy(&ctrlr->lock);
1399 nvme_ctrlr_shutdown(struct nvme_controller *ctrlr)
1405 cc = nvme_mmio_read_4(ctrlr, cc);
1406 cc &= ~(NVME_CC_REG_SHN_MASK << NVME_CC_REG_SHN_SHIFT);
1407 cc |= NVME_SHN_NORMAL << NVME_CC_REG_SHN_SHIFT;
1408 nvme_mmio_write_4(ctrlr, cc, cc);
1410 csts = nvme_mmio_read_4(ctrlr, csts);
1411 while ((NVME_CSTS_GET_SHST(csts) != NVME_SHST_COMPLETE) && (ticks++ < 5*hz)) {
1412 pause("nvme shn", 1);
1413 csts = nvme_mmio_read_4(ctrlr, csts);
1415 if (NVME_CSTS_GET_SHST(csts) != NVME_SHST_COMPLETE)
1416 nvme_printf(ctrlr, "did not complete shutdown within 5 seconds "
1417 "of notification\n");
1421 nvme_ctrlr_submit_admin_request(struct nvme_controller *ctrlr,
1422 struct nvme_request *req)
1425 nvme_qpair_submit_request(&ctrlr->adminq, req);
1429 nvme_ctrlr_submit_io_request(struct nvme_controller *ctrlr,
1430 struct nvme_request *req)
1432 struct nvme_qpair *qpair;
1434 qpair = &ctrlr->ioq[curcpu / ctrlr->num_cpus_per_ioq];
1435 nvme_qpair_submit_request(qpair, req);
1439 nvme_ctrlr_get_device(struct nvme_controller *ctrlr)
1442 return (ctrlr->dev);
1445 const struct nvme_controller_data *
1446 nvme_ctrlr_get_data(struct nvme_controller *ctrlr)
1449 return (&ctrlr->cdata);