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>
44 #include <dev/pci/pcireg.h>
45 #include <dev/pci/pcivar.h>
47 #include "nvme_private.h"
49 static void nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
50 struct nvme_async_event_request *aer);
51 static void nvme_ctrlr_setup_interrupts(struct nvme_controller *ctrlr);
54 nvme_ctrlr_allocate_bar(struct nvme_controller *ctrlr)
57 ctrlr->resource_id = PCIR_BAR(0);
59 ctrlr->resource = bus_alloc_resource_any(ctrlr->dev, SYS_RES_MEMORY,
60 &ctrlr->resource_id, RF_ACTIVE);
62 if(ctrlr->resource == NULL) {
63 nvme_printf(ctrlr, "unable to allocate pci resource\n");
67 ctrlr->bus_tag = rman_get_bustag(ctrlr->resource);
68 ctrlr->bus_handle = rman_get_bushandle(ctrlr->resource);
69 ctrlr->regs = (struct nvme_registers *)ctrlr->bus_handle;
72 * The NVMe spec allows for the MSI-X table to be placed behind
73 * BAR 4/5, separate from the control/doorbell registers. Always
74 * try to map this bar, because it must be mapped prior to calling
75 * pci_alloc_msix(). If the table isn't behind BAR 4/5,
76 * bus_alloc_resource() will just return NULL which is OK.
78 ctrlr->bar4_resource_id = PCIR_BAR(4);
79 ctrlr->bar4_resource = bus_alloc_resource_any(ctrlr->dev, SYS_RES_MEMORY,
80 &ctrlr->bar4_resource_id, RF_ACTIVE);
86 nvme_ctrlr_construct_admin_qpair(struct nvme_controller *ctrlr)
88 struct nvme_qpair *qpair;
92 qpair = &ctrlr->adminq;
94 num_entries = NVME_ADMIN_ENTRIES;
95 TUNABLE_INT_FETCH("hw.nvme.admin_entries", &num_entries);
97 * If admin_entries was overridden to an invalid value, revert it
98 * back to our default value.
100 if (num_entries < NVME_MIN_ADMIN_ENTRIES ||
101 num_entries > NVME_MAX_ADMIN_ENTRIES) {
102 nvme_printf(ctrlr, "invalid hw.nvme.admin_entries=%d "
103 "specified\n", num_entries);
104 num_entries = NVME_ADMIN_ENTRIES;
108 * The admin queue's max xfer size is treated differently than the
109 * max I/O xfer size. 16KB is sufficient here - maybe even less?
111 error = nvme_qpair_construct(qpair,
121 nvme_ctrlr_construct_io_qpairs(struct nvme_controller *ctrlr)
123 struct nvme_qpair *qpair;
124 union cap_lo_register cap_lo;
125 int i, error, num_entries, num_trackers;
127 num_entries = NVME_IO_ENTRIES;
128 TUNABLE_INT_FETCH("hw.nvme.io_entries", &num_entries);
131 * NVMe spec sets a hard limit of 64K max entries, but
132 * devices may specify a smaller limit, so we need to check
133 * the MQES field in the capabilities register.
135 cap_lo.raw = nvme_mmio_read_4(ctrlr, cap_lo);
136 num_entries = min(num_entries, cap_lo.bits.mqes+1);
138 num_trackers = NVME_IO_TRACKERS;
139 TUNABLE_INT_FETCH("hw.nvme.io_trackers", &num_trackers);
141 num_trackers = max(num_trackers, NVME_MIN_IO_TRACKERS);
142 num_trackers = min(num_trackers, NVME_MAX_IO_TRACKERS);
144 * No need to have more trackers than entries in the submit queue.
145 * Note also that for a queue size of N, we can only have (N-1)
146 * commands outstanding, hence the "-1" here.
148 num_trackers = min(num_trackers, (num_entries-1));
151 * Our best estimate for the maximum number of I/Os that we should
152 * noramlly have in flight at one time. This should be viewed as a hint,
153 * not a hard limit and will need to be revisitted when the upper layers
154 * of the storage system grows multi-queue support.
156 ctrlr->max_hw_pend_io = num_trackers * ctrlr->num_io_queues * 3 / 4;
159 * This was calculated previously when setting up interrupts, but
160 * a controller could theoretically support fewer I/O queues than
161 * MSI-X vectors. So calculate again here just to be safe.
163 ctrlr->num_cpus_per_ioq = howmany(mp_ncpus, ctrlr->num_io_queues);
165 ctrlr->ioq = malloc(ctrlr->num_io_queues * sizeof(struct nvme_qpair),
166 M_NVME, M_ZERO | M_WAITOK);
168 for (i = 0; i < ctrlr->num_io_queues; i++) {
169 qpair = &ctrlr->ioq[i];
172 * Admin queue has ID=0. IO queues start at ID=1 -
173 * hence the 'i+1' here.
175 * For I/O queues, use the controller-wide max_xfer_size
176 * calculated in nvme_attach().
178 error = nvme_qpair_construct(qpair,
180 ctrlr->msix_enabled ? i+1 : 0, /* vector */
188 * Do not bother binding interrupts if we only have one I/O
189 * interrupt thread for this controller.
191 if (ctrlr->num_io_queues > 1)
192 bus_bind_intr(ctrlr->dev, qpair->res,
193 i * ctrlr->num_cpus_per_ioq);
200 nvme_ctrlr_fail(struct nvme_controller *ctrlr)
204 ctrlr->is_failed = TRUE;
205 nvme_qpair_fail(&ctrlr->adminq);
206 if (ctrlr->ioq != NULL) {
207 for (i = 0; i < ctrlr->num_io_queues; i++)
208 nvme_qpair_fail(&ctrlr->ioq[i]);
210 nvme_notify_fail_consumers(ctrlr);
214 nvme_ctrlr_post_failed_request(struct nvme_controller *ctrlr,
215 struct nvme_request *req)
218 mtx_lock(&ctrlr->lock);
219 STAILQ_INSERT_TAIL(&ctrlr->fail_req, req, stailq);
220 mtx_unlock(&ctrlr->lock);
221 taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->fail_req_task);
225 nvme_ctrlr_fail_req_task(void *arg, int pending)
227 struct nvme_controller *ctrlr = arg;
228 struct nvme_request *req;
230 mtx_lock(&ctrlr->lock);
231 while (!STAILQ_EMPTY(&ctrlr->fail_req)) {
232 req = STAILQ_FIRST(&ctrlr->fail_req);
233 STAILQ_REMOVE_HEAD(&ctrlr->fail_req, stailq);
234 nvme_qpair_manual_complete_request(req->qpair, req,
235 NVME_SCT_GENERIC, NVME_SC_ABORTED_BY_REQUEST, TRUE);
237 mtx_unlock(&ctrlr->lock);
241 nvme_ctrlr_wait_for_ready(struct nvme_controller *ctrlr, int desired_val)
244 union cc_register cc;
245 union csts_register csts;
247 cc.raw = nvme_mmio_read_4(ctrlr, cc);
248 csts.raw = nvme_mmio_read_4(ctrlr, csts);
250 if (cc.bits.en != desired_val) {
251 nvme_printf(ctrlr, "%s called with desired_val = %d "
252 "but cc.en = %d\n", __func__, desired_val, cc.bits.en);
258 while (csts.bits.rdy != desired_val) {
260 if (ms_waited++ > ctrlr->ready_timeout_in_ms) {
261 nvme_printf(ctrlr, "controller ready did not become %d "
262 "within %d ms\n", desired_val, ctrlr->ready_timeout_in_ms);
265 csts.raw = nvme_mmio_read_4(ctrlr, csts);
272 nvme_ctrlr_disable(struct nvme_controller *ctrlr)
274 union cc_register cc;
275 union csts_register csts;
277 cc.raw = nvme_mmio_read_4(ctrlr, cc);
278 csts.raw = nvme_mmio_read_4(ctrlr, csts);
280 if (cc.bits.en == 1 && csts.bits.rdy == 0)
281 nvme_ctrlr_wait_for_ready(ctrlr, 1);
284 nvme_mmio_write_4(ctrlr, cc, cc.raw);
286 nvme_ctrlr_wait_for_ready(ctrlr, 0);
290 nvme_ctrlr_enable(struct nvme_controller *ctrlr)
292 union cc_register cc;
293 union csts_register csts;
294 union aqa_register aqa;
296 cc.raw = nvme_mmio_read_4(ctrlr, cc);
297 csts.raw = nvme_mmio_read_4(ctrlr, csts);
299 if (cc.bits.en == 1) {
300 if (csts.bits.rdy == 1)
303 return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
306 nvme_mmio_write_8(ctrlr, asq, ctrlr->adminq.cmd_bus_addr);
308 nvme_mmio_write_8(ctrlr, acq, ctrlr->adminq.cpl_bus_addr);
312 /* acqs and asqs are 0-based. */
313 aqa.bits.acqs = ctrlr->adminq.num_entries-1;
314 aqa.bits.asqs = ctrlr->adminq.num_entries-1;
315 nvme_mmio_write_4(ctrlr, aqa, aqa.raw);
322 cc.bits.iosqes = 6; /* SQ entry size == 64 == 2^6 */
323 cc.bits.iocqes = 4; /* CQ entry size == 16 == 2^4 */
325 /* This evaluates to 0, which is according to spec. */
326 cc.bits.mps = (PAGE_SIZE >> 13);
328 nvme_mmio_write_4(ctrlr, cc, cc.raw);
331 return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
335 nvme_ctrlr_hw_reset(struct nvme_controller *ctrlr)
339 nvme_admin_qpair_disable(&ctrlr->adminq);
341 * I/O queues are not allocated before the initial HW
342 * reset, so do not try to disable them. Use is_initialized
343 * to determine if this is the initial HW reset.
345 if (ctrlr->is_initialized) {
346 for (i = 0; i < ctrlr->num_io_queues; i++)
347 nvme_io_qpair_disable(&ctrlr->ioq[i]);
352 nvme_ctrlr_disable(ctrlr);
353 return (nvme_ctrlr_enable(ctrlr));
357 nvme_ctrlr_reset(struct nvme_controller *ctrlr)
361 cmpset = atomic_cmpset_32(&ctrlr->is_resetting, 0, 1);
363 if (cmpset == 0 || ctrlr->is_failed)
365 * Controller is already resetting or has failed. Return
366 * immediately since there is no need to kick off another
367 * reset in these cases.
371 taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->reset_task);
375 nvme_ctrlr_identify(struct nvme_controller *ctrlr)
377 struct nvme_completion_poll_status status;
380 nvme_ctrlr_cmd_identify_controller(ctrlr, &ctrlr->cdata,
381 nvme_completion_poll_cb, &status);
382 while (status.done == FALSE)
384 if (nvme_completion_is_error(&status.cpl)) {
385 nvme_printf(ctrlr, "nvme_identify_controller failed!\n");
390 * Use MDTS to ensure our default max_xfer_size doesn't exceed what the
391 * controller supports.
393 if (ctrlr->cdata.mdts > 0)
394 ctrlr->max_xfer_size = min(ctrlr->max_xfer_size,
395 ctrlr->min_page_size * (1 << (ctrlr->cdata.mdts)));
401 nvme_ctrlr_set_num_qpairs(struct nvme_controller *ctrlr)
403 struct nvme_completion_poll_status status;
404 int cq_allocated, sq_allocated;
407 nvme_ctrlr_cmd_set_num_queues(ctrlr, ctrlr->num_io_queues,
408 nvme_completion_poll_cb, &status);
409 while (status.done == FALSE)
411 if (nvme_completion_is_error(&status.cpl)) {
412 nvme_printf(ctrlr, "nvme_ctrlr_set_num_qpairs failed!\n");
417 * Data in cdw0 is 0-based.
418 * Lower 16-bits indicate number of submission queues allocated.
419 * Upper 16-bits indicate number of completion queues allocated.
421 sq_allocated = (status.cpl.cdw0 & 0xFFFF) + 1;
422 cq_allocated = (status.cpl.cdw0 >> 16) + 1;
425 * Controller may allocate more queues than we requested,
426 * so use the minimum of the number requested and what was
427 * actually allocated.
429 ctrlr->num_io_queues = min(ctrlr->num_io_queues, sq_allocated);
430 ctrlr->num_io_queues = min(ctrlr->num_io_queues, cq_allocated);
436 nvme_ctrlr_create_qpairs(struct nvme_controller *ctrlr)
438 struct nvme_completion_poll_status status;
439 struct nvme_qpair *qpair;
442 for (i = 0; i < ctrlr->num_io_queues; i++) {
443 qpair = &ctrlr->ioq[i];
446 nvme_ctrlr_cmd_create_io_cq(ctrlr, qpair, qpair->vector,
447 nvme_completion_poll_cb, &status);
448 while (status.done == FALSE)
450 if (nvme_completion_is_error(&status.cpl)) {
451 nvme_printf(ctrlr, "nvme_create_io_cq failed!\n");
456 nvme_ctrlr_cmd_create_io_sq(qpair->ctrlr, qpair,
457 nvme_completion_poll_cb, &status);
458 while (status.done == FALSE)
460 if (nvme_completion_is_error(&status.cpl)) {
461 nvme_printf(ctrlr, "nvme_create_io_sq failed!\n");
470 nvme_ctrlr_construct_namespaces(struct nvme_controller *ctrlr)
472 struct nvme_namespace *ns;
475 for (i = 0; i < min(ctrlr->cdata.nn, NVME_MAX_NAMESPACES); i++) {
477 nvme_ns_construct(ns, i+1, ctrlr);
484 is_log_page_id_valid(uint8_t page_id)
489 case NVME_LOG_HEALTH_INFORMATION:
490 case NVME_LOG_FIRMWARE_SLOT:
498 nvme_ctrlr_get_log_page_size(struct nvme_controller *ctrlr, uint8_t page_id)
500 uint32_t log_page_size;
505 sizeof(struct nvme_error_information_entry) *
507 NVME_MAX_AER_LOG_SIZE);
509 case NVME_LOG_HEALTH_INFORMATION:
510 log_page_size = sizeof(struct nvme_health_information_page);
512 case NVME_LOG_FIRMWARE_SLOT:
513 log_page_size = sizeof(struct nvme_firmware_page);
520 return (log_page_size);
524 nvme_ctrlr_log_critical_warnings(struct nvme_controller *ctrlr,
525 union nvme_critical_warning_state state)
528 if (state.bits.available_spare == 1)
529 nvme_printf(ctrlr, "available spare space below threshold\n");
531 if (state.bits.temperature == 1)
532 nvme_printf(ctrlr, "temperature above threshold\n");
534 if (state.bits.device_reliability == 1)
535 nvme_printf(ctrlr, "device reliability degraded\n");
537 if (state.bits.read_only == 1)
538 nvme_printf(ctrlr, "media placed in read only mode\n");
540 if (state.bits.volatile_memory_backup == 1)
541 nvme_printf(ctrlr, "volatile memory backup device failed\n");
543 if (state.bits.reserved != 0)
545 "unknown critical warning(s): state = 0x%02x\n", state.raw);
549 nvme_ctrlr_async_event_log_page_cb(void *arg, const struct nvme_completion *cpl)
551 struct nvme_async_event_request *aer = arg;
552 struct nvme_health_information_page *health_info;
555 * If the log page fetch for some reason completed with an error,
556 * don't pass log page data to the consumers. In practice, this case
557 * should never happen.
559 if (nvme_completion_is_error(cpl))
560 nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
561 aer->log_page_id, NULL, 0);
563 if (aer->log_page_id == NVME_LOG_HEALTH_INFORMATION) {
564 health_info = (struct nvme_health_information_page *)
565 aer->log_page_buffer;
566 nvme_ctrlr_log_critical_warnings(aer->ctrlr,
567 health_info->critical_warning);
569 * Critical warnings reported through the
570 * SMART/health log page are persistent, so
571 * clear the associated bits in the async event
572 * config so that we do not receive repeated
573 * notifications for the same event.
575 aer->ctrlr->async_event_config.raw &=
576 ~health_info->critical_warning.raw;
577 nvme_ctrlr_cmd_set_async_event_config(aer->ctrlr,
578 aer->ctrlr->async_event_config, NULL, NULL);
583 * Pass the cpl data from the original async event completion,
584 * not the log page fetch.
586 nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
587 aer->log_page_id, aer->log_page_buffer, aer->log_page_size);
591 * Repost another asynchronous event request to replace the one
592 * that just completed.
594 nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
598 nvme_ctrlr_async_event_cb(void *arg, const struct nvme_completion *cpl)
600 struct nvme_async_event_request *aer = arg;
602 if (nvme_completion_is_error(cpl)) {
604 * Do not retry failed async event requests. This avoids
605 * infinite loops where a new async event request is submitted
606 * to replace the one just failed, only to fail again and
607 * perpetuate the loop.
612 /* Associated log page is in bits 23:16 of completion entry dw0. */
613 aer->log_page_id = (cpl->cdw0 & 0xFF0000) >> 16;
615 nvme_printf(aer->ctrlr, "async event occurred (log page id=0x%x)\n",
618 if (is_log_page_id_valid(aer->log_page_id)) {
619 aer->log_page_size = nvme_ctrlr_get_log_page_size(aer->ctrlr,
621 memcpy(&aer->cpl, cpl, sizeof(*cpl));
622 nvme_ctrlr_cmd_get_log_page(aer->ctrlr, aer->log_page_id,
623 NVME_GLOBAL_NAMESPACE_TAG, aer->log_page_buffer,
624 aer->log_page_size, nvme_ctrlr_async_event_log_page_cb,
626 /* Wait to notify consumers until after log page is fetched. */
628 nvme_notify_async_consumers(aer->ctrlr, cpl, aer->log_page_id,
632 * Repost another asynchronous event request to replace the one
633 * that just completed.
635 nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
640 nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
641 struct nvme_async_event_request *aer)
643 struct nvme_request *req;
646 req = nvme_allocate_request_null(nvme_ctrlr_async_event_cb, aer);
650 * Disable timeout here, since asynchronous event requests should by
651 * nature never be timed out.
653 req->timeout = FALSE;
654 req->cmd.opc = NVME_OPC_ASYNC_EVENT_REQUEST;
655 nvme_ctrlr_submit_admin_request(ctrlr, req);
659 nvme_ctrlr_configure_aer(struct nvme_controller *ctrlr)
661 struct nvme_completion_poll_status status;
662 struct nvme_async_event_request *aer;
665 ctrlr->async_event_config.raw = 0xFF;
666 ctrlr->async_event_config.bits.reserved = 0;
669 nvme_ctrlr_cmd_get_feature(ctrlr, NVME_FEAT_TEMPERATURE_THRESHOLD,
670 0, NULL, 0, nvme_completion_poll_cb, &status);
671 while (status.done == FALSE)
673 if (nvme_completion_is_error(&status.cpl) ||
674 (status.cpl.cdw0 & 0xFFFF) == 0xFFFF ||
675 (status.cpl.cdw0 & 0xFFFF) == 0x0000) {
676 nvme_printf(ctrlr, "temperature threshold not supported\n");
677 ctrlr->async_event_config.bits.temperature = 0;
680 nvme_ctrlr_cmd_set_async_event_config(ctrlr,
681 ctrlr->async_event_config, NULL, NULL);
683 /* aerl is a zero-based value, so we need to add 1 here. */
684 ctrlr->num_aers = min(NVME_MAX_ASYNC_EVENTS, (ctrlr->cdata.aerl+1));
686 for (i = 0; i < ctrlr->num_aers; i++) {
687 aer = &ctrlr->aer[i];
688 nvme_ctrlr_construct_and_submit_aer(ctrlr, aer);
693 nvme_ctrlr_configure_int_coalescing(struct nvme_controller *ctrlr)
696 ctrlr->int_coal_time = 0;
697 TUNABLE_INT_FETCH("hw.nvme.int_coal_time",
698 &ctrlr->int_coal_time);
700 ctrlr->int_coal_threshold = 0;
701 TUNABLE_INT_FETCH("hw.nvme.int_coal_threshold",
702 &ctrlr->int_coal_threshold);
704 nvme_ctrlr_cmd_set_interrupt_coalescing(ctrlr, ctrlr->int_coal_time,
705 ctrlr->int_coal_threshold, NULL, NULL);
709 nvme_ctrlr_start(void *ctrlr_arg)
711 struct nvme_controller *ctrlr = ctrlr_arg;
712 uint32_t old_num_io_queues;
716 * Only reset adminq here when we are restarting the
717 * controller after a reset. During initialization,
718 * we have already submitted admin commands to get
719 * the number of I/O queues supported, so cannot reset
720 * the adminq again here.
722 if (ctrlr->is_resetting) {
723 nvme_qpair_reset(&ctrlr->adminq);
726 for (i = 0; i < ctrlr->num_io_queues; i++)
727 nvme_qpair_reset(&ctrlr->ioq[i]);
729 nvme_admin_qpair_enable(&ctrlr->adminq);
731 if (nvme_ctrlr_identify(ctrlr) != 0) {
732 nvme_ctrlr_fail(ctrlr);
737 * The number of qpairs are determined during controller initialization,
738 * including using NVMe SET_FEATURES/NUMBER_OF_QUEUES to determine the
739 * HW limit. We call SET_FEATURES again here so that it gets called
740 * after any reset for controllers that depend on the driver to
741 * explicit specify how many queues it will use. This value should
742 * never change between resets, so panic if somehow that does happen.
744 if (ctrlr->is_resetting) {
745 old_num_io_queues = ctrlr->num_io_queues;
746 if (nvme_ctrlr_set_num_qpairs(ctrlr) != 0) {
747 nvme_ctrlr_fail(ctrlr);
751 if (old_num_io_queues != ctrlr->num_io_queues) {
752 panic("num_io_queues changed from %u to %u",
753 old_num_io_queues, ctrlr->num_io_queues);
757 if (nvme_ctrlr_create_qpairs(ctrlr) != 0) {
758 nvme_ctrlr_fail(ctrlr);
762 if (nvme_ctrlr_construct_namespaces(ctrlr) != 0) {
763 nvme_ctrlr_fail(ctrlr);
767 nvme_ctrlr_configure_aer(ctrlr);
768 nvme_ctrlr_configure_int_coalescing(ctrlr);
770 for (i = 0; i < ctrlr->num_io_queues; i++)
771 nvme_io_qpair_enable(&ctrlr->ioq[i]);
775 nvme_ctrlr_start_config_hook(void *arg)
777 struct nvme_controller *ctrlr = arg;
779 nvme_qpair_reset(&ctrlr->adminq);
780 nvme_admin_qpair_enable(&ctrlr->adminq);
782 if (nvme_ctrlr_set_num_qpairs(ctrlr) == 0 &&
783 nvme_ctrlr_construct_io_qpairs(ctrlr) == 0)
784 nvme_ctrlr_start(ctrlr);
786 nvme_ctrlr_fail(ctrlr);
788 nvme_sysctl_initialize_ctrlr(ctrlr);
789 config_intrhook_disestablish(&ctrlr->config_hook);
791 ctrlr->is_initialized = 1;
792 nvme_notify_new_controller(ctrlr);
796 nvme_ctrlr_reset_task(void *arg, int pending)
798 struct nvme_controller *ctrlr = arg;
801 nvme_printf(ctrlr, "resetting controller\n");
802 status = nvme_ctrlr_hw_reset(ctrlr);
804 * Use pause instead of DELAY, so that we yield to any nvme interrupt
805 * handlers on this CPU that were blocked on a qpair lock. We want
806 * all nvme interrupts completed before proceeding with restarting the
809 * XXX - any way to guarantee the interrupt handlers have quiesced?
811 pause("nvmereset", hz / 10);
813 nvme_ctrlr_start(ctrlr);
815 nvme_ctrlr_fail(ctrlr);
817 atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
821 * Poll all the queues enabled on the device for completion.
824 nvme_ctrlr_poll(struct nvme_controller *ctrlr)
828 nvme_qpair_process_completions(&ctrlr->adminq);
830 for (i = 0; i < ctrlr->num_io_queues; i++)
831 if (ctrlr->ioq && ctrlr->ioq[i].cpl)
832 nvme_qpair_process_completions(&ctrlr->ioq[i]);
836 * Poll the single-vector intertrupt case: num_io_queues will be 1 and
837 * there's only a single vector. While we're polling, we mask further
838 * interrupts in the controller.
841 nvme_ctrlr_intx_handler(void *arg)
843 struct nvme_controller *ctrlr = arg;
845 nvme_mmio_write_4(ctrlr, intms, 1);
846 nvme_ctrlr_poll(ctrlr);
847 nvme_mmio_write_4(ctrlr, intmc, 1);
851 nvme_ctrlr_configure_intx(struct nvme_controller *ctrlr)
854 ctrlr->msix_enabled = 0;
855 ctrlr->num_io_queues = 1;
856 ctrlr->num_cpus_per_ioq = mp_ncpus;
858 ctrlr->res = bus_alloc_resource_any(ctrlr->dev, SYS_RES_IRQ,
859 &ctrlr->rid, RF_SHAREABLE | RF_ACTIVE);
861 if (ctrlr->res == NULL) {
862 nvme_printf(ctrlr, "unable to allocate shared IRQ\n");
866 bus_setup_intr(ctrlr->dev, ctrlr->res,
867 INTR_TYPE_MISC | INTR_MPSAFE, NULL, nvme_ctrlr_intx_handler,
870 if (ctrlr->tag == NULL) {
871 nvme_printf(ctrlr, "unable to setup intx handler\n");
879 nvme_pt_done(void *arg, const struct nvme_completion *cpl)
881 struct nvme_pt_command *pt = arg;
883 bzero(&pt->cpl, sizeof(pt->cpl));
884 pt->cpl.cdw0 = cpl->cdw0;
885 pt->cpl.status = cpl->status;
886 pt->cpl.status.p = 0;
888 mtx_lock(pt->driver_lock);
890 mtx_unlock(pt->driver_lock);
894 nvme_ctrlr_passthrough_cmd(struct nvme_controller *ctrlr,
895 struct nvme_pt_command *pt, uint32_t nsid, int is_user_buffer,
898 struct nvme_request *req;
900 struct buf *buf = NULL;
902 vm_offset_t addr, end;
906 * vmapbuf calls vm_fault_quick_hold_pages which only maps full
907 * pages. Ensure this request has fewer than MAXPHYS bytes when
908 * extended to full pages.
910 addr = (vm_offset_t)pt->buf;
911 end = round_page(addr + pt->len);
912 addr = trunc_page(addr);
913 if (end - addr > MAXPHYS)
916 if (pt->len > ctrlr->max_xfer_size) {
917 nvme_printf(ctrlr, "pt->len (%d) "
918 "exceeds max_xfer_size (%d)\n", pt->len,
919 ctrlr->max_xfer_size);
922 if (is_user_buffer) {
924 * Ensure the user buffer is wired for the duration of
925 * this passthrough command.
929 buf->b_data = pt->buf;
930 buf->b_bufsize = pt->len;
931 buf->b_iocmd = pt->is_read ? BIO_READ : BIO_WRITE;
932 #ifdef NVME_UNMAPPED_BIO_SUPPORT
933 if (vmapbuf(buf, 1) < 0) {
935 if (vmapbuf(buf) < 0) {
940 req = nvme_allocate_request_vaddr(buf->b_data, pt->len,
943 req = nvme_allocate_request_vaddr(pt->buf, pt->len,
946 req = nvme_allocate_request_null(nvme_pt_done, pt);
948 req->cmd.opc = pt->cmd.opc;
949 req->cmd.cdw10 = pt->cmd.cdw10;
950 req->cmd.cdw11 = pt->cmd.cdw11;
951 req->cmd.cdw12 = pt->cmd.cdw12;
952 req->cmd.cdw13 = pt->cmd.cdw13;
953 req->cmd.cdw14 = pt->cmd.cdw14;
954 req->cmd.cdw15 = pt->cmd.cdw15;
956 req->cmd.nsid = nsid;
961 mtx = &ctrlr->ns[nsid-1].lock;
964 pt->driver_lock = mtx;
967 nvme_ctrlr_submit_admin_request(ctrlr, req);
969 nvme_ctrlr_submit_io_request(ctrlr, req);
971 mtx_sleep(pt, mtx, PRIBIO, "nvme_pt", 0);
974 pt->driver_lock = NULL;
986 nvme_ctrlr_ioctl(struct cdev *cdev, u_long cmd, caddr_t arg, int flag,
989 struct nvme_controller *ctrlr;
990 struct nvme_pt_command *pt;
992 ctrlr = cdev->si_drv1;
995 case NVME_RESET_CONTROLLER:
996 nvme_ctrlr_reset(ctrlr);
998 case NVME_PASSTHROUGH_CMD:
999 pt = (struct nvme_pt_command *)arg;
1000 return (nvme_ctrlr_passthrough_cmd(ctrlr, pt, pt->cmd.nsid,
1001 1 /* is_user_buffer */, 1 /* is_admin_cmd */));
1009 static struct cdevsw nvme_ctrlr_cdevsw = {
1010 .d_version = D_VERSION,
1012 .d_ioctl = nvme_ctrlr_ioctl
1016 nvme_ctrlr_setup_interrupts(struct nvme_controller *ctrlr)
1019 int per_cpu_io_queues;
1020 int min_cpus_per_ioq;
1021 int num_vectors_requested, num_vectors_allocated;
1022 int num_vectors_available;
1025 min_cpus_per_ioq = 1;
1026 TUNABLE_INT_FETCH("hw.nvme.min_cpus_per_ioq", &min_cpus_per_ioq);
1028 if (min_cpus_per_ioq < 1) {
1029 min_cpus_per_ioq = 1;
1030 } else if (min_cpus_per_ioq > mp_ncpus) {
1031 min_cpus_per_ioq = mp_ncpus;
1034 per_cpu_io_queues = 1;
1035 TUNABLE_INT_FETCH("hw.nvme.per_cpu_io_queues", &per_cpu_io_queues);
1037 if (per_cpu_io_queues == 0) {
1038 min_cpus_per_ioq = mp_ncpus;
1041 ctrlr->force_intx = 0;
1042 TUNABLE_INT_FETCH("hw.nvme.force_intx", &ctrlr->force_intx);
1045 * FreeBSD currently cannot allocate more than about 190 vectors at
1046 * boot, meaning that systems with high core count and many devices
1047 * requesting per-CPU interrupt vectors will not get their full
1048 * allotment. So first, try to allocate as many as we may need to
1049 * understand what is available, then immediately release them.
1050 * Then figure out how many of those we will actually use, based on
1051 * assigning an equal number of cores to each I/O queue.
1054 /* One vector for per core I/O queue, plus one vector for admin queue. */
1055 num_vectors_available = min(pci_msix_count(dev), mp_ncpus + 1);
1056 if (pci_alloc_msix(dev, &num_vectors_available) != 0) {
1057 num_vectors_available = 0;
1059 pci_release_msi(dev);
1061 if (ctrlr->force_intx || num_vectors_available < 2) {
1062 nvme_ctrlr_configure_intx(ctrlr);
1067 * Do not use all vectors for I/O queues - one must be saved for the
1070 ctrlr->num_cpus_per_ioq = max(min_cpus_per_ioq,
1071 howmany(mp_ncpus, num_vectors_available - 1));
1073 ctrlr->num_io_queues = howmany(mp_ncpus, ctrlr->num_cpus_per_ioq);
1074 num_vectors_requested = ctrlr->num_io_queues + 1;
1075 num_vectors_allocated = num_vectors_requested;
1078 * Now just allocate the number of vectors we need. This should
1079 * succeed, since we previously called pci_alloc_msix()
1080 * successfully returning at least this many vectors, but just to
1081 * be safe, if something goes wrong just revert to INTx.
1083 if (pci_alloc_msix(dev, &num_vectors_allocated) != 0) {
1084 nvme_ctrlr_configure_intx(ctrlr);
1088 if (num_vectors_allocated < num_vectors_requested) {
1089 pci_release_msi(dev);
1090 nvme_ctrlr_configure_intx(ctrlr);
1094 ctrlr->msix_enabled = 1;
1098 nvme_ctrlr_construct(struct nvme_controller *ctrlr, device_t dev)
1100 union cap_lo_register cap_lo;
1101 union cap_hi_register cap_hi;
1102 int status, timeout_period;
1106 mtx_init(&ctrlr->lock, "nvme ctrlr lock", NULL, MTX_DEF);
1108 status = nvme_ctrlr_allocate_bar(ctrlr);
1114 * Software emulators may set the doorbell stride to something
1115 * other than zero, but this driver is not set up to handle that.
1117 cap_hi.raw = nvme_mmio_read_4(ctrlr, cap_hi);
1118 if (cap_hi.bits.dstrd != 0)
1121 ctrlr->min_page_size = 1 << (12 + cap_hi.bits.mpsmin);
1123 /* Get ready timeout value from controller, in units of 500ms. */
1124 cap_lo.raw = nvme_mmio_read_4(ctrlr, cap_lo);
1125 ctrlr->ready_timeout_in_ms = cap_lo.bits.to * 500;
1127 timeout_period = NVME_DEFAULT_TIMEOUT_PERIOD;
1128 TUNABLE_INT_FETCH("hw.nvme.timeout_period", &timeout_period);
1129 timeout_period = min(timeout_period, NVME_MAX_TIMEOUT_PERIOD);
1130 timeout_period = max(timeout_period, NVME_MIN_TIMEOUT_PERIOD);
1131 ctrlr->timeout_period = timeout_period;
1133 nvme_retry_count = NVME_DEFAULT_RETRY_COUNT;
1134 TUNABLE_INT_FETCH("hw.nvme.retry_count", &nvme_retry_count);
1136 ctrlr->enable_aborts = 0;
1137 TUNABLE_INT_FETCH("hw.nvme.enable_aborts", &ctrlr->enable_aborts);
1139 nvme_ctrlr_setup_interrupts(ctrlr);
1141 ctrlr->max_xfer_size = NVME_MAX_XFER_SIZE;
1142 if (nvme_ctrlr_construct_admin_qpair(ctrlr) != 0)
1145 ctrlr->cdev = make_dev(&nvme_ctrlr_cdevsw, device_get_unit(dev),
1146 UID_ROOT, GID_WHEEL, 0600, "nvme%d", device_get_unit(dev));
1148 if (ctrlr->cdev == NULL)
1151 ctrlr->cdev->si_drv1 = (void *)ctrlr;
1153 ctrlr->taskqueue = taskqueue_create("nvme_taskq", M_WAITOK,
1154 taskqueue_thread_enqueue, &ctrlr->taskqueue);
1155 taskqueue_start_threads(&ctrlr->taskqueue, 1, PI_DISK, "nvme taskq");
1157 ctrlr->is_resetting = 0;
1158 ctrlr->is_initialized = 0;
1159 ctrlr->notification_sent = 0;
1160 TASK_INIT(&ctrlr->reset_task, 0, nvme_ctrlr_reset_task, ctrlr);
1162 TASK_INIT(&ctrlr->fail_req_task, 0, nvme_ctrlr_fail_req_task, ctrlr);
1163 STAILQ_INIT(&ctrlr->fail_req);
1164 ctrlr->is_failed = FALSE;
1170 nvme_ctrlr_destruct(struct nvme_controller *ctrlr, device_t dev)
1175 * Notify the controller of a shutdown, even though this is due to
1176 * a driver unload, not a system shutdown (this path is not invoked
1177 * during shutdown). This ensures the controller receives a
1178 * shutdown notification in case the system is shutdown before
1179 * reloading the driver.
1181 nvme_ctrlr_shutdown(ctrlr);
1183 nvme_ctrlr_disable(ctrlr);
1184 taskqueue_free(ctrlr->taskqueue);
1186 for (i = 0; i < NVME_MAX_NAMESPACES; i++)
1187 nvme_ns_destruct(&ctrlr->ns[i]);
1190 destroy_dev(ctrlr->cdev);
1192 for (i = 0; i < ctrlr->num_io_queues; i++) {
1193 nvme_io_qpair_destroy(&ctrlr->ioq[i]);
1196 free(ctrlr->ioq, M_NVME);
1198 nvme_admin_qpair_destroy(&ctrlr->adminq);
1200 if (ctrlr->resource != NULL) {
1201 bus_release_resource(dev, SYS_RES_MEMORY,
1202 ctrlr->resource_id, ctrlr->resource);
1205 if (ctrlr->bar4_resource != NULL) {
1206 bus_release_resource(dev, SYS_RES_MEMORY,
1207 ctrlr->bar4_resource_id, ctrlr->bar4_resource);
1211 bus_teardown_intr(ctrlr->dev, ctrlr->res, ctrlr->tag);
1214 bus_release_resource(ctrlr->dev, SYS_RES_IRQ,
1215 rman_get_rid(ctrlr->res), ctrlr->res);
1217 if (ctrlr->msix_enabled)
1218 pci_release_msi(dev);
1222 nvme_ctrlr_shutdown(struct nvme_controller *ctrlr)
1224 union cc_register cc;
1225 union csts_register csts;
1228 cc.raw = nvme_mmio_read_4(ctrlr, cc);
1229 cc.bits.shn = NVME_SHN_NORMAL;
1230 nvme_mmio_write_4(ctrlr, cc, cc.raw);
1231 csts.raw = nvme_mmio_read_4(ctrlr, csts);
1232 while ((csts.bits.shst != NVME_SHST_COMPLETE) && (ticks++ < 5*hz)) {
1233 pause("nvme shn", 1);
1234 csts.raw = nvme_mmio_read_4(ctrlr, csts);
1236 if (csts.bits.shst != NVME_SHST_COMPLETE)
1237 nvme_printf(ctrlr, "did not complete shutdown within 5 seconds "
1238 "of notification\n");
1242 nvme_ctrlr_submit_admin_request(struct nvme_controller *ctrlr,
1243 struct nvme_request *req)
1246 nvme_qpair_submit_request(&ctrlr->adminq, req);
1250 nvme_ctrlr_submit_io_request(struct nvme_controller *ctrlr,
1251 struct nvme_request *req)
1253 struct nvme_qpair *qpair;
1255 qpair = &ctrlr->ioq[curcpu / ctrlr->num_cpus_per_ioq];
1256 nvme_qpair_submit_request(qpair, req);
1260 nvme_ctrlr_get_device(struct nvme_controller *ctrlr)
1263 return (ctrlr->dev);
1266 const struct nvme_controller_data *
1267 nvme_ctrlr_get_data(struct nvme_controller *ctrlr)
1270 return (&ctrlr->cdata);