2 * Copyright (C) 2012-2016 Intel Corporation
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
30 #include <sys/param.h>
31 #include <sys/systm.h>
35 #include <sys/ioccom.h>
40 #include <dev/pci/pcireg.h>
41 #include <dev/pci/pcivar.h>
43 #include "nvme_private.h"
45 static void nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
46 struct nvme_async_event_request *aer);
47 static void nvme_ctrlr_setup_interrupts(struct nvme_controller *ctrlr);
50 nvme_ctrlr_allocate_bar(struct nvme_controller *ctrlr)
53 ctrlr->resource_id = PCIR_BAR(0);
55 ctrlr->resource = bus_alloc_resource(ctrlr->dev, SYS_RES_MEMORY,
56 &ctrlr->resource_id, 0, ~0, 1, RF_ACTIVE);
58 if(ctrlr->resource == NULL) {
59 nvme_printf(ctrlr, "unable to allocate pci resource\n");
63 ctrlr->bus_tag = rman_get_bustag(ctrlr->resource);
64 ctrlr->bus_handle = rman_get_bushandle(ctrlr->resource);
65 ctrlr->regs = (struct nvme_registers *)ctrlr->bus_handle;
68 * The NVMe spec allows for the MSI-X table to be placed behind
69 * BAR 4/5, separate from the control/doorbell registers. Always
70 * try to map this bar, because it must be mapped prior to calling
71 * pci_alloc_msix(). If the table isn't behind BAR 4/5,
72 * bus_alloc_resource() will just return NULL which is OK.
74 ctrlr->bar4_resource_id = PCIR_BAR(4);
75 ctrlr->bar4_resource = bus_alloc_resource(ctrlr->dev, SYS_RES_MEMORY,
76 &ctrlr->bar4_resource_id, 0, ~0, 1, RF_ACTIVE);
82 nvme_ctrlr_construct_admin_qpair(struct nvme_controller *ctrlr)
84 struct nvme_qpair *qpair;
87 qpair = &ctrlr->adminq;
89 num_entries = NVME_ADMIN_ENTRIES;
90 TUNABLE_INT_FETCH("hw.nvme.admin_entries", &num_entries);
92 * If admin_entries was overridden to an invalid value, revert it
93 * back to our default value.
95 if (num_entries < NVME_MIN_ADMIN_ENTRIES ||
96 num_entries > NVME_MAX_ADMIN_ENTRIES) {
97 nvme_printf(ctrlr, "invalid hw.nvme.admin_entries=%d "
98 "specified\n", num_entries);
99 num_entries = NVME_ADMIN_ENTRIES;
103 * The admin queue's max xfer size is treated differently than the
104 * max I/O xfer size. 16KB is sufficient here - maybe even less?
106 nvme_qpair_construct(qpair,
115 nvme_ctrlr_construct_io_qpairs(struct nvme_controller *ctrlr)
117 struct nvme_qpair *qpair;
118 union cap_lo_register cap_lo;
119 int i, num_entries, num_trackers;
121 num_entries = NVME_IO_ENTRIES;
122 TUNABLE_INT_FETCH("hw.nvme.io_entries", &num_entries);
125 * NVMe spec sets a hard limit of 64K max entries, but
126 * devices may specify a smaller limit, so we need to check
127 * the MQES field in the capabilities register.
129 cap_lo.raw = nvme_mmio_read_4(ctrlr, cap_lo);
130 num_entries = min(num_entries, cap_lo.bits.mqes+1);
132 num_trackers = NVME_IO_TRACKERS;
133 TUNABLE_INT_FETCH("hw.nvme.io_trackers", &num_trackers);
135 num_trackers = max(num_trackers, NVME_MIN_IO_TRACKERS);
136 num_trackers = min(num_trackers, NVME_MAX_IO_TRACKERS);
138 * No need to have more trackers than entries in the submit queue.
139 * Note also that for a queue size of N, we can only have (N-1)
140 * commands outstanding, hence the "-1" here.
142 num_trackers = min(num_trackers, (num_entries-1));
145 * This was calculated previously when setting up interrupts, but
146 * a controller could theoretically support fewer I/O queues than
147 * MSI-X vectors. So calculate again here just to be safe.
149 ctrlr->num_cpus_per_ioq = howmany(mp_ncpus, ctrlr->num_io_queues);
151 ctrlr->ioq = malloc(ctrlr->num_io_queues * sizeof(struct nvme_qpair),
152 M_NVME, M_ZERO | M_WAITOK);
154 for (i = 0; i < ctrlr->num_io_queues; i++) {
155 qpair = &ctrlr->ioq[i];
158 * Admin queue has ID=0. IO queues start at ID=1 -
159 * hence the 'i+1' here.
161 * For I/O queues, use the controller-wide max_xfer_size
162 * calculated in nvme_attach().
164 nvme_qpair_construct(qpair,
166 ctrlr->msix_enabled ? i+1 : 0, /* vector */
172 * Do not bother binding interrupts if we only have one I/O
173 * interrupt thread for this controller.
175 if (ctrlr->num_io_queues > 1)
176 bus_bind_intr(ctrlr->dev, qpair->res,
177 i * ctrlr->num_cpus_per_ioq);
184 nvme_ctrlr_fail(struct nvme_controller *ctrlr)
188 ctrlr->is_failed = TRUE;
189 nvme_qpair_fail(&ctrlr->adminq);
190 for (i = 0; i < ctrlr->num_io_queues; i++)
191 nvme_qpair_fail(&ctrlr->ioq[i]);
192 nvme_notify_fail_consumers(ctrlr);
196 nvme_ctrlr_post_failed_request(struct nvme_controller *ctrlr,
197 struct nvme_request *req)
200 mtx_lock(&ctrlr->lock);
201 STAILQ_INSERT_TAIL(&ctrlr->fail_req, req, stailq);
202 mtx_unlock(&ctrlr->lock);
203 taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->fail_req_task);
207 nvme_ctrlr_fail_req_task(void *arg, int pending)
209 struct nvme_controller *ctrlr = arg;
210 struct nvme_request *req;
212 mtx_lock(&ctrlr->lock);
213 while (!STAILQ_EMPTY(&ctrlr->fail_req)) {
214 req = STAILQ_FIRST(&ctrlr->fail_req);
215 STAILQ_REMOVE_HEAD(&ctrlr->fail_req, stailq);
216 nvme_qpair_manual_complete_request(req->qpair, req,
217 NVME_SCT_GENERIC, NVME_SC_ABORTED_BY_REQUEST, TRUE);
219 mtx_unlock(&ctrlr->lock);
223 nvme_ctrlr_wait_for_ready(struct nvme_controller *ctrlr, int desired_val)
226 union cc_register cc;
227 union csts_register csts;
229 cc.raw = nvme_mmio_read_4(ctrlr, cc);
230 csts.raw = nvme_mmio_read_4(ctrlr, csts);
232 if (cc.bits.en != desired_val) {
233 nvme_printf(ctrlr, "%s called with desired_val = %d "
234 "but cc.en = %d\n", __func__, desired_val, cc.bits.en);
240 while (csts.bits.rdy != desired_val) {
242 if (ms_waited++ > ctrlr->ready_timeout_in_ms) {
243 nvme_printf(ctrlr, "controller ready did not become %d "
244 "within %d ms\n", desired_val, ctrlr->ready_timeout_in_ms);
247 csts.raw = nvme_mmio_read_4(ctrlr, csts);
254 nvme_ctrlr_disable(struct nvme_controller *ctrlr)
256 union cc_register cc;
257 union csts_register csts;
259 cc.raw = nvme_mmio_read_4(ctrlr, cc);
260 csts.raw = nvme_mmio_read_4(ctrlr, csts);
262 if (cc.bits.en == 1 && csts.bits.rdy == 0)
263 nvme_ctrlr_wait_for_ready(ctrlr, 1);
266 nvme_mmio_write_4(ctrlr, cc, cc.raw);
268 nvme_ctrlr_wait_for_ready(ctrlr, 0);
272 nvme_ctrlr_enable(struct nvme_controller *ctrlr)
274 union cc_register cc;
275 union csts_register csts;
276 union aqa_register aqa;
278 cc.raw = nvme_mmio_read_4(ctrlr, cc);
279 csts.raw = nvme_mmio_read_4(ctrlr, csts);
281 if (cc.bits.en == 1) {
282 if (csts.bits.rdy == 1)
285 return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
288 nvme_mmio_write_8(ctrlr, asq, ctrlr->adminq.cmd_bus_addr);
290 nvme_mmio_write_8(ctrlr, acq, ctrlr->adminq.cpl_bus_addr);
294 /* acqs and asqs are 0-based. */
295 aqa.bits.acqs = ctrlr->adminq.num_entries-1;
296 aqa.bits.asqs = ctrlr->adminq.num_entries-1;
297 nvme_mmio_write_4(ctrlr, aqa, aqa.raw);
304 cc.bits.iosqes = 6; /* SQ entry size == 64 == 2^6 */
305 cc.bits.iocqes = 4; /* CQ entry size == 16 == 2^4 */
307 /* This evaluates to 0, which is according to spec. */
308 cc.bits.mps = (PAGE_SIZE >> 13);
310 nvme_mmio_write_4(ctrlr, cc, cc.raw);
313 return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
317 nvme_ctrlr_hw_reset(struct nvme_controller *ctrlr)
321 nvme_admin_qpair_disable(&ctrlr->adminq);
323 * I/O queues are not allocated before the initial HW
324 * reset, so do not try to disable them. Use is_initialized
325 * to determine if this is the initial HW reset.
327 if (ctrlr->is_initialized) {
328 for (i = 0; i < ctrlr->num_io_queues; i++)
329 nvme_io_qpair_disable(&ctrlr->ioq[i]);
334 nvme_ctrlr_disable(ctrlr);
335 return (nvme_ctrlr_enable(ctrlr));
339 nvme_ctrlr_reset(struct nvme_controller *ctrlr)
343 cmpset = atomic_cmpset_32(&ctrlr->is_resetting, 0, 1);
345 if (cmpset == 0 || ctrlr->is_failed)
347 * Controller is already resetting or has failed. Return
348 * immediately since there is no need to kick off another
349 * reset in these cases.
353 taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->reset_task);
357 nvme_ctrlr_identify(struct nvme_controller *ctrlr)
359 struct nvme_completion_poll_status status;
362 nvme_ctrlr_cmd_identify_controller(ctrlr, &ctrlr->cdata,
363 nvme_completion_poll_cb, &status);
364 while (status.done == FALSE)
366 if (nvme_completion_is_error(&status.cpl)) {
367 nvme_printf(ctrlr, "nvme_identify_controller failed!\n");
372 * Use MDTS to ensure our default max_xfer_size doesn't exceed what the
373 * controller supports.
375 if (ctrlr->cdata.mdts > 0)
376 ctrlr->max_xfer_size = min(ctrlr->max_xfer_size,
377 ctrlr->min_page_size * (1 << (ctrlr->cdata.mdts)));
383 nvme_ctrlr_set_num_qpairs(struct nvme_controller *ctrlr)
385 struct nvme_completion_poll_status status;
386 int cq_allocated, sq_allocated;
389 nvme_ctrlr_cmd_set_num_queues(ctrlr, ctrlr->num_io_queues,
390 nvme_completion_poll_cb, &status);
391 while (status.done == FALSE)
393 if (nvme_completion_is_error(&status.cpl)) {
394 nvme_printf(ctrlr, "nvme_set_num_queues failed!\n");
399 * Data in cdw0 is 0-based.
400 * Lower 16-bits indicate number of submission queues allocated.
401 * Upper 16-bits indicate number of completion queues allocated.
403 sq_allocated = (status.cpl.cdw0 & 0xFFFF) + 1;
404 cq_allocated = (status.cpl.cdw0 >> 16) + 1;
407 * Controller may allocate more queues than we requested,
408 * so use the minimum of the number requested and what was
409 * actually allocated.
411 ctrlr->num_io_queues = min(ctrlr->num_io_queues, sq_allocated);
412 ctrlr->num_io_queues = min(ctrlr->num_io_queues, cq_allocated);
418 nvme_ctrlr_create_qpairs(struct nvme_controller *ctrlr)
420 struct nvme_completion_poll_status status;
421 struct nvme_qpair *qpair;
424 for (i = 0; i < ctrlr->num_io_queues; i++) {
425 qpair = &ctrlr->ioq[i];
428 nvme_ctrlr_cmd_create_io_cq(ctrlr, qpair, qpair->vector,
429 nvme_completion_poll_cb, &status);
430 while (status.done == FALSE)
432 if (nvme_completion_is_error(&status.cpl)) {
433 nvme_printf(ctrlr, "nvme_create_io_cq failed!\n");
438 nvme_ctrlr_cmd_create_io_sq(qpair->ctrlr, qpair,
439 nvme_completion_poll_cb, &status);
440 while (status.done == FALSE)
442 if (nvme_completion_is_error(&status.cpl)) {
443 nvme_printf(ctrlr, "nvme_create_io_sq failed!\n");
452 nvme_ctrlr_construct_namespaces(struct nvme_controller *ctrlr)
454 struct nvme_namespace *ns;
457 for (i = 0; i < ctrlr->cdata.nn; i++) {
459 status = nvme_ns_construct(ns, i+1, ctrlr);
468 is_log_page_id_valid(uint8_t page_id)
473 case NVME_LOG_HEALTH_INFORMATION:
474 case NVME_LOG_FIRMWARE_SLOT:
482 nvme_ctrlr_get_log_page_size(struct nvme_controller *ctrlr, uint8_t page_id)
484 uint32_t log_page_size;
489 sizeof(struct nvme_error_information_entry) *
491 NVME_MAX_AER_LOG_SIZE);
493 case NVME_LOG_HEALTH_INFORMATION:
494 log_page_size = sizeof(struct nvme_health_information_page);
496 case NVME_LOG_FIRMWARE_SLOT:
497 log_page_size = sizeof(struct nvme_firmware_page);
504 return (log_page_size);
508 nvme_ctrlr_log_critical_warnings(struct nvme_controller *ctrlr,
509 union nvme_critical_warning_state state)
512 if (state.bits.available_spare == 1)
513 nvme_printf(ctrlr, "available spare space below threshold\n");
515 if (state.bits.temperature == 1)
516 nvme_printf(ctrlr, "temperature above threshold\n");
518 if (state.bits.device_reliability == 1)
519 nvme_printf(ctrlr, "device reliability degraded\n");
521 if (state.bits.read_only == 1)
522 nvme_printf(ctrlr, "media placed in read only mode\n");
524 if (state.bits.volatile_memory_backup == 1)
525 nvme_printf(ctrlr, "volatile memory backup device failed\n");
527 if (state.bits.reserved != 0)
529 "unknown critical warning(s): state = 0x%02x\n", state.raw);
533 nvme_ctrlr_async_event_log_page_cb(void *arg, const struct nvme_completion *cpl)
535 struct nvme_async_event_request *aer = arg;
536 struct nvme_health_information_page *health_info;
539 * If the log page fetch for some reason completed with an error,
540 * don't pass log page data to the consumers. In practice, this case
541 * should never happen.
543 if (nvme_completion_is_error(cpl))
544 nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
545 aer->log_page_id, NULL, 0);
547 if (aer->log_page_id == NVME_LOG_HEALTH_INFORMATION) {
548 health_info = (struct nvme_health_information_page *)
549 aer->log_page_buffer;
550 nvme_ctrlr_log_critical_warnings(aer->ctrlr,
551 health_info->critical_warning);
553 * Critical warnings reported through the
554 * SMART/health log page are persistent, so
555 * clear the associated bits in the async event
556 * config so that we do not receive repeated
557 * notifications for the same event.
559 aer->ctrlr->async_event_config.raw &=
560 ~health_info->critical_warning.raw;
561 nvme_ctrlr_cmd_set_async_event_config(aer->ctrlr,
562 aer->ctrlr->async_event_config, NULL, NULL);
567 * Pass the cpl data from the original async event completion,
568 * not the log page fetch.
570 nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
571 aer->log_page_id, aer->log_page_buffer, aer->log_page_size);
575 * Repost another asynchronous event request to replace the one
576 * that just completed.
578 nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
582 nvme_ctrlr_async_event_cb(void *arg, const struct nvme_completion *cpl)
584 struct nvme_async_event_request *aer = arg;
586 if (nvme_completion_is_error(cpl)) {
588 * Do not retry failed async event requests. This avoids
589 * infinite loops where a new async event request is submitted
590 * to replace the one just failed, only to fail again and
591 * perpetuate the loop.
596 /* Associated log page is in bits 23:16 of completion entry dw0. */
597 aer->log_page_id = (cpl->cdw0 & 0xFF0000) >> 16;
599 nvme_printf(aer->ctrlr, "async event occurred (log page id=0x%x)\n",
602 if (is_log_page_id_valid(aer->log_page_id)) {
603 aer->log_page_size = nvme_ctrlr_get_log_page_size(aer->ctrlr,
605 memcpy(&aer->cpl, cpl, sizeof(*cpl));
606 nvme_ctrlr_cmd_get_log_page(aer->ctrlr, aer->log_page_id,
607 NVME_GLOBAL_NAMESPACE_TAG, aer->log_page_buffer,
608 aer->log_page_size, nvme_ctrlr_async_event_log_page_cb,
610 /* Wait to notify consumers until after log page is fetched. */
612 nvme_notify_async_consumers(aer->ctrlr, cpl, aer->log_page_id,
616 * Repost another asynchronous event request to replace the one
617 * that just completed.
619 nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
624 nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
625 struct nvme_async_event_request *aer)
627 struct nvme_request *req;
630 req = nvme_allocate_request_null(nvme_ctrlr_async_event_cb, aer);
634 * Disable timeout here, since asynchronous event requests should by
635 * nature never be timed out.
637 req->timeout = FALSE;
638 req->cmd.opc = NVME_OPC_ASYNC_EVENT_REQUEST;
639 nvme_ctrlr_submit_admin_request(ctrlr, req);
643 nvme_ctrlr_configure_aer(struct nvme_controller *ctrlr)
645 struct nvme_completion_poll_status status;
646 struct nvme_async_event_request *aer;
649 ctrlr->async_event_config.raw = 0xFF;
650 ctrlr->async_event_config.bits.reserved = 0;
653 nvme_ctrlr_cmd_get_feature(ctrlr, NVME_FEAT_TEMPERATURE_THRESHOLD,
654 0, NULL, 0, nvme_completion_poll_cb, &status);
655 while (status.done == FALSE)
657 if (nvme_completion_is_error(&status.cpl) ||
658 (status.cpl.cdw0 & 0xFFFF) == 0xFFFF ||
659 (status.cpl.cdw0 & 0xFFFF) == 0x0000) {
660 nvme_printf(ctrlr, "temperature threshold not supported\n");
661 ctrlr->async_event_config.bits.temperature = 0;
664 nvme_ctrlr_cmd_set_async_event_config(ctrlr,
665 ctrlr->async_event_config, NULL, NULL);
667 /* aerl is a zero-based value, so we need to add 1 here. */
668 ctrlr->num_aers = min(NVME_MAX_ASYNC_EVENTS, (ctrlr->cdata.aerl+1));
670 for (i = 0; i < ctrlr->num_aers; i++) {
671 aer = &ctrlr->aer[i];
672 nvme_ctrlr_construct_and_submit_aer(ctrlr, aer);
677 nvme_ctrlr_configure_int_coalescing(struct nvme_controller *ctrlr)
680 ctrlr->int_coal_time = 0;
681 TUNABLE_INT_FETCH("hw.nvme.int_coal_time",
682 &ctrlr->int_coal_time);
684 ctrlr->int_coal_threshold = 0;
685 TUNABLE_INT_FETCH("hw.nvme.int_coal_threshold",
686 &ctrlr->int_coal_threshold);
688 nvme_ctrlr_cmd_set_interrupt_coalescing(ctrlr, ctrlr->int_coal_time,
689 ctrlr->int_coal_threshold, NULL, NULL);
693 nvme_ctrlr_start(void *ctrlr_arg)
695 struct nvme_controller *ctrlr = ctrlr_arg;
696 uint32_t old_num_io_queues;
700 * Only reset adminq here when we are restarting the
701 * controller after a reset. During initialization,
702 * we have already submitted admin commands to get
703 * the number of I/O queues supported, so cannot reset
704 * the adminq again here.
706 if (ctrlr->is_resetting) {
707 nvme_qpair_reset(&ctrlr->adminq);
710 for (i = 0; i < ctrlr->num_io_queues; i++)
711 nvme_qpair_reset(&ctrlr->ioq[i]);
713 nvme_admin_qpair_enable(&ctrlr->adminq);
715 if (nvme_ctrlr_identify(ctrlr) != 0) {
716 nvme_ctrlr_fail(ctrlr);
721 * The number of qpairs are determined during controller initialization,
722 * including using NVMe SET_FEATURES/NUMBER_OF_QUEUES to determine the
723 * HW limit. We call SET_FEATURES again here so that it gets called
724 * after any reset for controllers that depend on the driver to
725 * explicit specify how many queues it will use. This value should
726 * never change between resets, so panic if somehow that does happen.
728 if (ctrlr->is_resetting) {
729 old_num_io_queues = ctrlr->num_io_queues;
730 if (nvme_ctrlr_set_num_qpairs(ctrlr) != 0) {
731 nvme_ctrlr_fail(ctrlr);
735 if (old_num_io_queues != ctrlr->num_io_queues) {
736 panic("num_io_queues changed from %u to %u",
737 old_num_io_queues, ctrlr->num_io_queues);
741 if (nvme_ctrlr_create_qpairs(ctrlr) != 0) {
742 nvme_ctrlr_fail(ctrlr);
746 if (nvme_ctrlr_construct_namespaces(ctrlr) != 0) {
747 nvme_ctrlr_fail(ctrlr);
751 nvme_ctrlr_configure_aer(ctrlr);
752 nvme_ctrlr_configure_int_coalescing(ctrlr);
754 for (i = 0; i < ctrlr->num_io_queues; i++)
755 nvme_io_qpair_enable(&ctrlr->ioq[i]);
759 nvme_ctrlr_start_config_hook(void *arg)
761 struct nvme_controller *ctrlr = arg;
763 nvme_qpair_reset(&ctrlr->adminq);
764 nvme_admin_qpair_enable(&ctrlr->adminq);
766 if (nvme_ctrlr_set_num_qpairs(ctrlr) == 0 &&
767 nvme_ctrlr_construct_io_qpairs(ctrlr) == 0)
768 nvme_ctrlr_start(ctrlr);
770 nvme_ctrlr_fail(ctrlr);
772 nvme_sysctl_initialize_ctrlr(ctrlr);
773 config_intrhook_disestablish(&ctrlr->config_hook);
775 ctrlr->is_initialized = 1;
776 nvme_notify_new_controller(ctrlr);
780 nvme_ctrlr_reset_task(void *arg, int pending)
782 struct nvme_controller *ctrlr = arg;
785 nvme_printf(ctrlr, "resetting controller\n");
786 status = nvme_ctrlr_hw_reset(ctrlr);
788 * Use pause instead of DELAY, so that we yield to any nvme interrupt
789 * handlers on this CPU that were blocked on a qpair lock. We want
790 * all nvme interrupts completed before proceeding with restarting the
793 * XXX - any way to guarantee the interrupt handlers have quiesced?
795 pause("nvmereset", hz / 10);
797 nvme_ctrlr_start(ctrlr);
799 nvme_ctrlr_fail(ctrlr);
801 atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
805 nvme_ctrlr_intx_handler(void *arg)
807 struct nvme_controller *ctrlr = arg;
809 nvme_mmio_write_4(ctrlr, intms, 1);
811 nvme_qpair_process_completions(&ctrlr->adminq);
813 if (ctrlr->ioq && ctrlr->ioq[0].cpl)
814 nvme_qpair_process_completions(&ctrlr->ioq[0]);
816 nvme_mmio_write_4(ctrlr, intmc, 1);
820 nvme_ctrlr_configure_intx(struct nvme_controller *ctrlr)
823 ctrlr->msix_enabled = 0;
824 ctrlr->num_io_queues = 1;
825 ctrlr->num_cpus_per_ioq = mp_ncpus;
827 ctrlr->res = bus_alloc_resource_any(ctrlr->dev, SYS_RES_IRQ,
828 &ctrlr->rid, RF_SHAREABLE | RF_ACTIVE);
830 if (ctrlr->res == NULL) {
831 nvme_printf(ctrlr, "unable to allocate shared IRQ\n");
835 bus_setup_intr(ctrlr->dev, ctrlr->res,
836 INTR_TYPE_MISC | INTR_MPSAFE, NULL, nvme_ctrlr_intx_handler,
839 if (ctrlr->tag == NULL) {
840 nvme_printf(ctrlr, "unable to setup intx handler\n");
848 nvme_pt_done(void *arg, const struct nvme_completion *cpl)
850 struct nvme_pt_command *pt = arg;
852 bzero(&pt->cpl, sizeof(pt->cpl));
853 pt->cpl.cdw0 = cpl->cdw0;
854 pt->cpl.status = cpl->status;
855 pt->cpl.status.p = 0;
857 mtx_lock(pt->driver_lock);
859 mtx_unlock(pt->driver_lock);
863 nvme_ctrlr_passthrough_cmd(struct nvme_controller *ctrlr,
864 struct nvme_pt_command *pt, uint32_t nsid, int is_user_buffer,
867 struct nvme_request *req;
869 struct buf *buf = NULL;
873 if (pt->len > ctrlr->max_xfer_size) {
874 nvme_printf(ctrlr, "pt->len (%d) "
875 "exceeds max_xfer_size (%d)\n", pt->len,
876 ctrlr->max_xfer_size);
879 if (is_user_buffer) {
881 * Ensure the user buffer is wired for the duration of
882 * this passthrough command.
886 buf->b_saveaddr = buf->b_data;
887 buf->b_data = pt->buf;
888 buf->b_bufsize = pt->len;
889 buf->b_iocmd = pt->is_read ? BIO_READ : BIO_WRITE;
890 #ifdef NVME_UNMAPPED_BIO_SUPPORT
891 if (vmapbuf(buf, 1) < 0) {
893 if (vmapbuf(buf) < 0) {
898 req = nvme_allocate_request_vaddr(buf->b_data, pt->len,
901 req = nvme_allocate_request_vaddr(pt->buf, pt->len,
904 req = nvme_allocate_request_null(nvme_pt_done, pt);
906 req->cmd.opc = pt->cmd.opc;
907 req->cmd.cdw10 = pt->cmd.cdw10;
908 req->cmd.cdw11 = pt->cmd.cdw11;
909 req->cmd.cdw12 = pt->cmd.cdw12;
910 req->cmd.cdw13 = pt->cmd.cdw13;
911 req->cmd.cdw14 = pt->cmd.cdw14;
912 req->cmd.cdw15 = pt->cmd.cdw15;
914 req->cmd.nsid = nsid;
919 mtx = &ctrlr->ns[nsid-1].lock;
922 pt->driver_lock = mtx;
925 nvme_ctrlr_submit_admin_request(ctrlr, req);
927 nvme_ctrlr_submit_io_request(ctrlr, req);
929 mtx_sleep(pt, mtx, PRIBIO, "nvme_pt", 0);
932 pt->driver_lock = NULL;
944 nvme_ctrlr_ioctl(struct cdev *cdev, u_long cmd, caddr_t arg, int flag,
947 struct nvme_controller *ctrlr;
948 struct nvme_pt_command *pt;
950 ctrlr = cdev->si_drv1;
953 case NVME_RESET_CONTROLLER:
954 nvme_ctrlr_reset(ctrlr);
956 case NVME_PASSTHROUGH_CMD:
957 pt = (struct nvme_pt_command *)arg;
958 return (nvme_ctrlr_passthrough_cmd(ctrlr, pt, pt->cmd.nsid,
959 1 /* is_user_buffer */, 1 /* is_admin_cmd */));
967 static struct cdevsw nvme_ctrlr_cdevsw = {
968 .d_version = D_VERSION,
970 .d_ioctl = nvme_ctrlr_ioctl
974 nvme_ctrlr_setup_interrupts(struct nvme_controller *ctrlr)
977 int per_cpu_io_queues;
978 int min_cpus_per_ioq;
979 int num_vectors_requested, num_vectors_allocated;
980 int num_vectors_available;
983 min_cpus_per_ioq = 1;
984 TUNABLE_INT_FETCH("hw.nvme.min_cpus_per_ioq", &min_cpus_per_ioq);
986 if (min_cpus_per_ioq < 1) {
987 min_cpus_per_ioq = 1;
988 } else if (min_cpus_per_ioq > mp_ncpus) {
989 min_cpus_per_ioq = mp_ncpus;
992 per_cpu_io_queues = 1;
993 TUNABLE_INT_FETCH("hw.nvme.per_cpu_io_queues", &per_cpu_io_queues);
995 if (per_cpu_io_queues == 0) {
996 min_cpus_per_ioq = mp_ncpus;
999 ctrlr->force_intx = 0;
1000 TUNABLE_INT_FETCH("hw.nvme.force_intx", &ctrlr->force_intx);
1003 * FreeBSD currently cannot allocate more than about 190 vectors at
1004 * boot, meaning that systems with high core count and many devices
1005 * requesting per-CPU interrupt vectors will not get their full
1006 * allotment. So first, try to allocate as many as we may need to
1007 * understand what is available, then immediately release them.
1008 * Then figure out how many of those we will actually use, based on
1009 * assigning an equal number of cores to each I/O queue.
1012 /* One vector for per core I/O queue, plus one vector for admin queue. */
1013 num_vectors_available = min(pci_msix_count(dev), mp_ncpus + 1);
1014 if (pci_alloc_msix(dev, &num_vectors_available) != 0) {
1015 num_vectors_available = 0;
1017 pci_release_msi(dev);
1019 if (ctrlr->force_intx || num_vectors_available < 2) {
1020 nvme_ctrlr_configure_intx(ctrlr);
1025 * Do not use all vectors for I/O queues - one must be saved for the
1028 ctrlr->num_cpus_per_ioq = max(min_cpus_per_ioq,
1029 howmany(mp_ncpus, num_vectors_available - 1));
1031 ctrlr->num_io_queues = howmany(mp_ncpus, ctrlr->num_cpus_per_ioq);
1032 num_vectors_requested = ctrlr->num_io_queues + 1;
1033 num_vectors_allocated = num_vectors_requested;
1036 * Now just allocate the number of vectors we need. This should
1037 * succeed, since we previously called pci_alloc_msix()
1038 * successfully returning at least this many vectors, but just to
1039 * be safe, if something goes wrong just revert to INTx.
1041 if (pci_alloc_msix(dev, &num_vectors_allocated) != 0) {
1042 nvme_ctrlr_configure_intx(ctrlr);
1046 if (num_vectors_allocated < num_vectors_requested) {
1047 pci_release_msi(dev);
1048 nvme_ctrlr_configure_intx(ctrlr);
1052 ctrlr->msix_enabled = 1;
1056 nvme_ctrlr_construct(struct nvme_controller *ctrlr, device_t dev)
1058 union cap_lo_register cap_lo;
1059 union cap_hi_register cap_hi;
1060 int status, timeout_period;
1064 mtx_init(&ctrlr->lock, "nvme ctrlr lock", NULL, MTX_DEF);
1066 status = nvme_ctrlr_allocate_bar(ctrlr);
1072 * Software emulators may set the doorbell stride to something
1073 * other than zero, but this driver is not set up to handle that.
1075 cap_hi.raw = nvme_mmio_read_4(ctrlr, cap_hi);
1076 if (cap_hi.bits.dstrd != 0)
1079 ctrlr->min_page_size = 1 << (12 + cap_hi.bits.mpsmin);
1081 /* Get ready timeout value from controller, in units of 500ms. */
1082 cap_lo.raw = nvme_mmio_read_4(ctrlr, cap_lo);
1083 ctrlr->ready_timeout_in_ms = cap_lo.bits.to * 500;
1085 timeout_period = NVME_DEFAULT_TIMEOUT_PERIOD;
1086 TUNABLE_INT_FETCH("hw.nvme.timeout_period", &timeout_period);
1087 timeout_period = min(timeout_period, NVME_MAX_TIMEOUT_PERIOD);
1088 timeout_period = max(timeout_period, NVME_MIN_TIMEOUT_PERIOD);
1089 ctrlr->timeout_period = timeout_period;
1091 nvme_retry_count = NVME_DEFAULT_RETRY_COUNT;
1092 TUNABLE_INT_FETCH("hw.nvme.retry_count", &nvme_retry_count);
1094 ctrlr->enable_aborts = 0;
1095 TUNABLE_INT_FETCH("hw.nvme.enable_aborts", &ctrlr->enable_aborts);
1097 nvme_ctrlr_setup_interrupts(ctrlr);
1099 ctrlr->max_xfer_size = NVME_MAX_XFER_SIZE;
1100 nvme_ctrlr_construct_admin_qpair(ctrlr);
1102 ctrlr->cdev = make_dev(&nvme_ctrlr_cdevsw, device_get_unit(dev),
1103 UID_ROOT, GID_WHEEL, 0600, "nvme%d", device_get_unit(dev));
1105 if (ctrlr->cdev == NULL)
1108 ctrlr->cdev->si_drv1 = (void *)ctrlr;
1110 ctrlr->taskqueue = taskqueue_create("nvme_taskq", M_WAITOK,
1111 taskqueue_thread_enqueue, &ctrlr->taskqueue);
1112 taskqueue_start_threads(&ctrlr->taskqueue, 1, PI_DISK, "nvme taskq");
1114 ctrlr->is_resetting = 0;
1115 ctrlr->is_initialized = 0;
1116 ctrlr->notification_sent = 0;
1117 TASK_INIT(&ctrlr->reset_task, 0, nvme_ctrlr_reset_task, ctrlr);
1119 TASK_INIT(&ctrlr->fail_req_task, 0, nvme_ctrlr_fail_req_task, ctrlr);
1120 STAILQ_INIT(&ctrlr->fail_req);
1121 ctrlr->is_failed = FALSE;
1127 nvme_ctrlr_destruct(struct nvme_controller *ctrlr, device_t dev)
1132 * Notify the controller of a shutdown, even though this is due to
1133 * a driver unload, not a system shutdown (this path is not invoked
1134 * during shutdown). This ensures the controller receives a
1135 * shutdown notification in case the system is shutdown before
1136 * reloading the driver.
1138 nvme_ctrlr_shutdown(ctrlr);
1140 nvme_ctrlr_disable(ctrlr);
1141 taskqueue_free(ctrlr->taskqueue);
1143 for (i = 0; i < NVME_MAX_NAMESPACES; i++)
1144 nvme_ns_destruct(&ctrlr->ns[i]);
1147 destroy_dev(ctrlr->cdev);
1149 for (i = 0; i < ctrlr->num_io_queues; i++) {
1150 nvme_io_qpair_destroy(&ctrlr->ioq[i]);
1153 free(ctrlr->ioq, M_NVME);
1155 nvme_admin_qpair_destroy(&ctrlr->adminq);
1157 if (ctrlr->resource != NULL) {
1158 bus_release_resource(dev, SYS_RES_MEMORY,
1159 ctrlr->resource_id, ctrlr->resource);
1162 if (ctrlr->bar4_resource != NULL) {
1163 bus_release_resource(dev, SYS_RES_MEMORY,
1164 ctrlr->bar4_resource_id, ctrlr->bar4_resource);
1168 bus_teardown_intr(ctrlr->dev, ctrlr->res, ctrlr->tag);
1171 bus_release_resource(ctrlr->dev, SYS_RES_IRQ,
1172 rman_get_rid(ctrlr->res), ctrlr->res);
1174 if (ctrlr->msix_enabled)
1175 pci_release_msi(dev);
1179 nvme_ctrlr_shutdown(struct nvme_controller *ctrlr)
1181 union cc_register cc;
1182 union csts_register csts;
1185 cc.raw = nvme_mmio_read_4(ctrlr, cc);
1186 cc.bits.shn = NVME_SHN_NORMAL;
1187 nvme_mmio_write_4(ctrlr, cc, cc.raw);
1188 csts.raw = nvme_mmio_read_4(ctrlr, csts);
1189 while ((csts.bits.shst != NVME_SHST_COMPLETE) && (ticks++ < 5*hz)) {
1190 pause("nvme shn", 1);
1191 csts.raw = nvme_mmio_read_4(ctrlr, csts);
1193 if (csts.bits.shst != NVME_SHST_COMPLETE)
1194 nvme_printf(ctrlr, "did not complete shutdown within 5 seconds "
1195 "of notification\n");
1199 nvme_ctrlr_submit_admin_request(struct nvme_controller *ctrlr,
1200 struct nvme_request *req)
1203 nvme_qpair_submit_request(&ctrlr->adminq, req);
1207 nvme_ctrlr_submit_io_request(struct nvme_controller *ctrlr,
1208 struct nvme_request *req)
1210 struct nvme_qpair *qpair;
1212 qpair = &ctrlr->ioq[curcpu / ctrlr->num_cpus_per_ioq];
1213 nvme_qpair_submit_request(qpair, req);
1217 nvme_ctrlr_get_device(struct nvme_controller *ctrlr)
1220 return (ctrlr->dev);
1223 const struct nvme_controller_data *
1224 nvme_ctrlr_get_data(struct nvme_controller *ctrlr)
1227 return (&ctrlr->cdata);