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("$FreeBSD$");
32 #include <sys/param.h>
33 #include <sys/systm.h>
37 #include <sys/ioccom.h>
42 #include <dev/pci/pcireg.h>
43 #include <dev/pci/pcivar.h>
45 #include "nvme_private.h"
47 static void nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
48 struct nvme_async_event_request *aer);
49 static void nvme_ctrlr_setup_interrupts(struct nvme_controller *ctrlr);
52 nvme_ctrlr_allocate_bar(struct nvme_controller *ctrlr)
55 ctrlr->resource_id = PCIR_BAR(0);
57 ctrlr->resource = bus_alloc_resource_any(ctrlr->dev, SYS_RES_MEMORY,
58 &ctrlr->resource_id, RF_ACTIVE);
60 if(ctrlr->resource == NULL) {
61 nvme_printf(ctrlr, "unable to allocate pci resource\n");
65 ctrlr->bus_tag = rman_get_bustag(ctrlr->resource);
66 ctrlr->bus_handle = rman_get_bushandle(ctrlr->resource);
67 ctrlr->regs = (struct nvme_registers *)ctrlr->bus_handle;
70 * The NVMe spec allows for the MSI-X table to be placed behind
71 * BAR 4/5, separate from the control/doorbell registers. Always
72 * try to map this bar, because it must be mapped prior to calling
73 * pci_alloc_msix(). If the table isn't behind BAR 4/5,
74 * bus_alloc_resource() will just return NULL which is OK.
76 ctrlr->bar4_resource_id = PCIR_BAR(4);
77 ctrlr->bar4_resource = bus_alloc_resource_any(ctrlr->dev, SYS_RES_MEMORY,
78 &ctrlr->bar4_resource_id, RF_ACTIVE);
84 nvme_ctrlr_construct_admin_qpair(struct nvme_controller *ctrlr)
86 struct nvme_qpair *qpair;
90 qpair = &ctrlr->adminq;
92 num_entries = NVME_ADMIN_ENTRIES;
93 TUNABLE_INT_FETCH("hw.nvme.admin_entries", &num_entries);
95 * If admin_entries was overridden to an invalid value, revert it
96 * back to our default value.
98 if (num_entries < NVME_MIN_ADMIN_ENTRIES ||
99 num_entries > NVME_MAX_ADMIN_ENTRIES) {
100 nvme_printf(ctrlr, "invalid hw.nvme.admin_entries=%d "
101 "specified\n", num_entries);
102 num_entries = NVME_ADMIN_ENTRIES;
106 * The admin queue's max xfer size is treated differently than the
107 * max I/O xfer size. 16KB is sufficient here - maybe even less?
109 error = nvme_qpair_construct(qpair,
119 nvme_ctrlr_construct_io_qpairs(struct nvme_controller *ctrlr)
121 struct nvme_qpair *qpair;
122 union cap_lo_register cap_lo;
123 int i, error, num_entries, num_trackers;
125 num_entries = NVME_IO_ENTRIES;
126 TUNABLE_INT_FETCH("hw.nvme.io_entries", &num_entries);
129 * NVMe spec sets a hard limit of 64K max entries, but
130 * devices may specify a smaller limit, so we need to check
131 * the MQES field in the capabilities register.
133 cap_lo.raw = nvme_mmio_read_4(ctrlr, cap_lo);
134 num_entries = min(num_entries, cap_lo.bits.mqes+1);
136 num_trackers = NVME_IO_TRACKERS;
137 TUNABLE_INT_FETCH("hw.nvme.io_trackers", &num_trackers);
139 num_trackers = max(num_trackers, NVME_MIN_IO_TRACKERS);
140 num_trackers = min(num_trackers, NVME_MAX_IO_TRACKERS);
142 * No need to have more trackers than entries in the submit queue.
143 * Note also that for a queue size of N, we can only have (N-1)
144 * commands outstanding, hence the "-1" here.
146 num_trackers = min(num_trackers, (num_entries-1));
149 * Our best estimate for the maximum number of I/Os that we should
150 * noramlly have in flight at one time. This should be viewed as a hint,
151 * not a hard limit and will need to be revisitted when the upper layers
152 * of the storage system grows multi-queue support.
154 ctrlr->max_hw_pend_io = num_trackers * ctrlr->num_io_queues * 3 / 4;
157 * This was calculated previously when setting up interrupts, but
158 * a controller could theoretically support fewer I/O queues than
159 * MSI-X vectors. So calculate again here just to be safe.
161 ctrlr->num_cpus_per_ioq = howmany(mp_ncpus, ctrlr->num_io_queues);
163 ctrlr->ioq = malloc(ctrlr->num_io_queues * sizeof(struct nvme_qpair),
164 M_NVME, M_ZERO | M_WAITOK);
166 for (i = 0; i < ctrlr->num_io_queues; i++) {
167 qpair = &ctrlr->ioq[i];
170 * Admin queue has ID=0. IO queues start at ID=1 -
171 * hence the 'i+1' here.
173 * For I/O queues, use the controller-wide max_xfer_size
174 * calculated in nvme_attach().
176 error = nvme_qpair_construct(qpair,
178 ctrlr->msix_enabled ? i+1 : 0, /* vector */
186 * Do not bother binding interrupts if we only have one I/O
187 * interrupt thread for this controller.
189 if (ctrlr->num_io_queues > 1)
190 bus_bind_intr(ctrlr->dev, qpair->res,
191 i * ctrlr->num_cpus_per_ioq);
198 nvme_ctrlr_fail(struct nvme_controller *ctrlr)
202 ctrlr->is_failed = TRUE;
203 nvme_qpair_fail(&ctrlr->adminq);
204 if (ctrlr->ioq != NULL) {
205 for (i = 0; i < ctrlr->num_io_queues; i++)
206 nvme_qpair_fail(&ctrlr->ioq[i]);
208 nvme_notify_fail_consumers(ctrlr);
212 nvme_ctrlr_post_failed_request(struct nvme_controller *ctrlr,
213 struct nvme_request *req)
216 mtx_lock(&ctrlr->lock);
217 STAILQ_INSERT_TAIL(&ctrlr->fail_req, req, stailq);
218 mtx_unlock(&ctrlr->lock);
219 taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->fail_req_task);
223 nvme_ctrlr_fail_req_task(void *arg, int pending)
225 struct nvme_controller *ctrlr = arg;
226 struct nvme_request *req;
228 mtx_lock(&ctrlr->lock);
229 while (!STAILQ_EMPTY(&ctrlr->fail_req)) {
230 req = STAILQ_FIRST(&ctrlr->fail_req);
231 STAILQ_REMOVE_HEAD(&ctrlr->fail_req, stailq);
232 nvme_qpair_manual_complete_request(req->qpair, req,
233 NVME_SCT_GENERIC, NVME_SC_ABORTED_BY_REQUEST, TRUE);
235 mtx_unlock(&ctrlr->lock);
239 nvme_ctrlr_wait_for_ready(struct nvme_controller *ctrlr, int desired_val)
242 union cc_register cc;
243 union csts_register csts;
245 cc.raw = nvme_mmio_read_4(ctrlr, cc);
246 csts.raw = nvme_mmio_read_4(ctrlr, csts);
248 if (cc.bits.en != desired_val) {
249 nvme_printf(ctrlr, "%s called with desired_val = %d "
250 "but cc.en = %d\n", __func__, desired_val, cc.bits.en);
256 while (csts.bits.rdy != desired_val) {
258 if (ms_waited++ > ctrlr->ready_timeout_in_ms) {
259 nvme_printf(ctrlr, "controller ready did not become %d "
260 "within %d ms\n", desired_val, ctrlr->ready_timeout_in_ms);
263 csts.raw = nvme_mmio_read_4(ctrlr, csts);
270 nvme_ctrlr_disable(struct nvme_controller *ctrlr)
272 union cc_register cc;
273 union csts_register csts;
275 cc.raw = nvme_mmio_read_4(ctrlr, cc);
276 csts.raw = nvme_mmio_read_4(ctrlr, csts);
278 if (cc.bits.en == 1 && csts.bits.rdy == 0)
279 nvme_ctrlr_wait_for_ready(ctrlr, 1);
282 nvme_mmio_write_4(ctrlr, cc, cc.raw);
284 nvme_ctrlr_wait_for_ready(ctrlr, 0);
288 nvme_ctrlr_enable(struct nvme_controller *ctrlr)
290 union cc_register cc;
291 union csts_register csts;
292 union aqa_register aqa;
294 cc.raw = nvme_mmio_read_4(ctrlr, cc);
295 csts.raw = nvme_mmio_read_4(ctrlr, csts);
297 if (cc.bits.en == 1) {
298 if (csts.bits.rdy == 1)
301 return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
304 nvme_mmio_write_8(ctrlr, asq, ctrlr->adminq.cmd_bus_addr);
306 nvme_mmio_write_8(ctrlr, acq, ctrlr->adminq.cpl_bus_addr);
310 /* acqs and asqs are 0-based. */
311 aqa.bits.acqs = ctrlr->adminq.num_entries-1;
312 aqa.bits.asqs = ctrlr->adminq.num_entries-1;
313 nvme_mmio_write_4(ctrlr, aqa, aqa.raw);
320 cc.bits.iosqes = 6; /* SQ entry size == 64 == 2^6 */
321 cc.bits.iocqes = 4; /* CQ entry size == 16 == 2^4 */
323 /* This evaluates to 0, which is according to spec. */
324 cc.bits.mps = (PAGE_SIZE >> 13);
326 nvme_mmio_write_4(ctrlr, cc, cc.raw);
329 return (nvme_ctrlr_wait_for_ready(ctrlr, 1));
333 nvme_ctrlr_hw_reset(struct nvme_controller *ctrlr)
337 nvme_admin_qpair_disable(&ctrlr->adminq);
339 * I/O queues are not allocated before the initial HW
340 * reset, so do not try to disable them. Use is_initialized
341 * to determine if this is the initial HW reset.
343 if (ctrlr->is_initialized) {
344 for (i = 0; i < ctrlr->num_io_queues; i++)
345 nvme_io_qpair_disable(&ctrlr->ioq[i]);
350 nvme_ctrlr_disable(ctrlr);
351 return (nvme_ctrlr_enable(ctrlr));
355 nvme_ctrlr_reset(struct nvme_controller *ctrlr)
359 cmpset = atomic_cmpset_32(&ctrlr->is_resetting, 0, 1);
361 if (cmpset == 0 || ctrlr->is_failed)
363 * Controller is already resetting or has failed. Return
364 * immediately since there is no need to kick off another
365 * reset in these cases.
369 taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->reset_task);
373 nvme_ctrlr_identify(struct nvme_controller *ctrlr)
375 struct nvme_completion_poll_status status;
378 nvme_ctrlr_cmd_identify_controller(ctrlr, &ctrlr->cdata,
379 nvme_completion_poll_cb, &status);
380 while (status.done == FALSE)
382 if (nvme_completion_is_error(&status.cpl)) {
383 nvme_printf(ctrlr, "nvme_identify_controller failed!\n");
388 * Use MDTS to ensure our default max_xfer_size doesn't exceed what the
389 * controller supports.
391 if (ctrlr->cdata.mdts > 0)
392 ctrlr->max_xfer_size = min(ctrlr->max_xfer_size,
393 ctrlr->min_page_size * (1 << (ctrlr->cdata.mdts)));
399 nvme_ctrlr_set_num_qpairs(struct nvme_controller *ctrlr)
401 struct nvme_completion_poll_status status;
402 int cq_allocated, sq_allocated;
405 nvme_ctrlr_cmd_set_num_queues(ctrlr, ctrlr->num_io_queues,
406 nvme_completion_poll_cb, &status);
407 while (status.done == FALSE)
409 if (nvme_completion_is_error(&status.cpl)) {
410 nvme_printf(ctrlr, "nvme_ctrlr_set_num_qpairs failed!\n");
415 * Data in cdw0 is 0-based.
416 * Lower 16-bits indicate number of submission queues allocated.
417 * Upper 16-bits indicate number of completion queues allocated.
419 sq_allocated = (status.cpl.cdw0 & 0xFFFF) + 1;
420 cq_allocated = (status.cpl.cdw0 >> 16) + 1;
423 * Controller may allocate more queues than we requested,
424 * so use the minimum of the number requested and what was
425 * actually allocated.
427 ctrlr->num_io_queues = min(ctrlr->num_io_queues, sq_allocated);
428 ctrlr->num_io_queues = min(ctrlr->num_io_queues, cq_allocated);
434 nvme_ctrlr_create_qpairs(struct nvme_controller *ctrlr)
436 struct nvme_completion_poll_status status;
437 struct nvme_qpair *qpair;
440 for (i = 0; i < ctrlr->num_io_queues; i++) {
441 qpair = &ctrlr->ioq[i];
444 nvme_ctrlr_cmd_create_io_cq(ctrlr, qpair, qpair->vector,
445 nvme_completion_poll_cb, &status);
446 while (status.done == FALSE)
448 if (nvme_completion_is_error(&status.cpl)) {
449 nvme_printf(ctrlr, "nvme_create_io_cq failed!\n");
454 nvme_ctrlr_cmd_create_io_sq(qpair->ctrlr, qpair,
455 nvme_completion_poll_cb, &status);
456 while (status.done == FALSE)
458 if (nvme_completion_is_error(&status.cpl)) {
459 nvme_printf(ctrlr, "nvme_create_io_sq failed!\n");
468 nvme_ctrlr_construct_namespaces(struct nvme_controller *ctrlr)
470 struct nvme_namespace *ns;
473 for (i = 0; i < min(ctrlr->cdata.nn, NVME_MAX_NAMESPACES); i++) {
475 nvme_ns_construct(ns, i+1, ctrlr);
482 is_log_page_id_valid(uint8_t page_id)
487 case NVME_LOG_HEALTH_INFORMATION:
488 case NVME_LOG_FIRMWARE_SLOT:
496 nvme_ctrlr_get_log_page_size(struct nvme_controller *ctrlr, uint8_t page_id)
498 uint32_t log_page_size;
503 sizeof(struct nvme_error_information_entry) *
505 NVME_MAX_AER_LOG_SIZE);
507 case NVME_LOG_HEALTH_INFORMATION:
508 log_page_size = sizeof(struct nvme_health_information_page);
510 case NVME_LOG_FIRMWARE_SLOT:
511 log_page_size = sizeof(struct nvme_firmware_page);
518 return (log_page_size);
522 nvme_ctrlr_log_critical_warnings(struct nvme_controller *ctrlr,
523 union nvme_critical_warning_state state)
526 if (state.bits.available_spare == 1)
527 nvme_printf(ctrlr, "available spare space below threshold\n");
529 if (state.bits.temperature == 1)
530 nvme_printf(ctrlr, "temperature above threshold\n");
532 if (state.bits.device_reliability == 1)
533 nvme_printf(ctrlr, "device reliability degraded\n");
535 if (state.bits.read_only == 1)
536 nvme_printf(ctrlr, "media placed in read only mode\n");
538 if (state.bits.volatile_memory_backup == 1)
539 nvme_printf(ctrlr, "volatile memory backup device failed\n");
541 if (state.bits.reserved != 0)
543 "unknown critical warning(s): state = 0x%02x\n", state.raw);
547 nvme_ctrlr_async_event_log_page_cb(void *arg, const struct nvme_completion *cpl)
549 struct nvme_async_event_request *aer = arg;
550 struct nvme_health_information_page *health_info;
553 * If the log page fetch for some reason completed with an error,
554 * don't pass log page data to the consumers. In practice, this case
555 * should never happen.
557 if (nvme_completion_is_error(cpl))
558 nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
559 aer->log_page_id, NULL, 0);
561 if (aer->log_page_id == NVME_LOG_HEALTH_INFORMATION) {
562 health_info = (struct nvme_health_information_page *)
563 aer->log_page_buffer;
564 nvme_ctrlr_log_critical_warnings(aer->ctrlr,
565 health_info->critical_warning);
567 * Critical warnings reported through the
568 * SMART/health log page are persistent, so
569 * clear the associated bits in the async event
570 * config so that we do not receive repeated
571 * notifications for the same event.
573 aer->ctrlr->async_event_config.raw &=
574 ~health_info->critical_warning.raw;
575 nvme_ctrlr_cmd_set_async_event_config(aer->ctrlr,
576 aer->ctrlr->async_event_config, NULL, NULL);
581 * Pass the cpl data from the original async event completion,
582 * not the log page fetch.
584 nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
585 aer->log_page_id, aer->log_page_buffer, aer->log_page_size);
589 * Repost another asynchronous event request to replace the one
590 * that just completed.
592 nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
596 nvme_ctrlr_async_event_cb(void *arg, const struct nvme_completion *cpl)
598 struct nvme_async_event_request *aer = arg;
600 if (nvme_completion_is_error(cpl)) {
602 * Do not retry failed async event requests. This avoids
603 * infinite loops where a new async event request is submitted
604 * to replace the one just failed, only to fail again and
605 * perpetuate the loop.
610 /* Associated log page is in bits 23:16 of completion entry dw0. */
611 aer->log_page_id = (cpl->cdw0 & 0xFF0000) >> 16;
613 nvme_printf(aer->ctrlr, "async event occurred (log page id=0x%x)\n",
616 if (is_log_page_id_valid(aer->log_page_id)) {
617 aer->log_page_size = nvme_ctrlr_get_log_page_size(aer->ctrlr,
619 memcpy(&aer->cpl, cpl, sizeof(*cpl));
620 nvme_ctrlr_cmd_get_log_page(aer->ctrlr, aer->log_page_id,
621 NVME_GLOBAL_NAMESPACE_TAG, aer->log_page_buffer,
622 aer->log_page_size, nvme_ctrlr_async_event_log_page_cb,
624 /* Wait to notify consumers until after log page is fetched. */
626 nvme_notify_async_consumers(aer->ctrlr, cpl, aer->log_page_id,
630 * Repost another asynchronous event request to replace the one
631 * that just completed.
633 nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
638 nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
639 struct nvme_async_event_request *aer)
641 struct nvme_request *req;
644 req = nvme_allocate_request_null(nvme_ctrlr_async_event_cb, aer);
648 * Disable timeout here, since asynchronous event requests should by
649 * nature never be timed out.
651 req->timeout = FALSE;
652 req->cmd.opc = NVME_OPC_ASYNC_EVENT_REQUEST;
653 nvme_ctrlr_submit_admin_request(ctrlr, req);
657 nvme_ctrlr_configure_aer(struct nvme_controller *ctrlr)
659 struct nvme_completion_poll_status status;
660 struct nvme_async_event_request *aer;
663 ctrlr->async_event_config.raw = 0xFF;
664 ctrlr->async_event_config.bits.reserved = 0;
667 nvme_ctrlr_cmd_get_feature(ctrlr, NVME_FEAT_TEMPERATURE_THRESHOLD,
668 0, NULL, 0, nvme_completion_poll_cb, &status);
669 while (status.done == FALSE)
671 if (nvme_completion_is_error(&status.cpl) ||
672 (status.cpl.cdw0 & 0xFFFF) == 0xFFFF ||
673 (status.cpl.cdw0 & 0xFFFF) == 0x0000) {
674 nvme_printf(ctrlr, "temperature threshold not supported\n");
675 ctrlr->async_event_config.bits.temperature = 0;
678 nvme_ctrlr_cmd_set_async_event_config(ctrlr,
679 ctrlr->async_event_config, NULL, NULL);
681 /* aerl is a zero-based value, so we need to add 1 here. */
682 ctrlr->num_aers = min(NVME_MAX_ASYNC_EVENTS, (ctrlr->cdata.aerl+1));
684 for (i = 0; i < ctrlr->num_aers; i++) {
685 aer = &ctrlr->aer[i];
686 nvme_ctrlr_construct_and_submit_aer(ctrlr, aer);
691 nvme_ctrlr_configure_int_coalescing(struct nvme_controller *ctrlr)
694 ctrlr->int_coal_time = 0;
695 TUNABLE_INT_FETCH("hw.nvme.int_coal_time",
696 &ctrlr->int_coal_time);
698 ctrlr->int_coal_threshold = 0;
699 TUNABLE_INT_FETCH("hw.nvme.int_coal_threshold",
700 &ctrlr->int_coal_threshold);
702 nvme_ctrlr_cmd_set_interrupt_coalescing(ctrlr, ctrlr->int_coal_time,
703 ctrlr->int_coal_threshold, NULL, NULL);
707 nvme_ctrlr_start(void *ctrlr_arg)
709 struct nvme_controller *ctrlr = ctrlr_arg;
710 uint32_t old_num_io_queues;
714 * Only reset adminq here when we are restarting the
715 * controller after a reset. During initialization,
716 * we have already submitted admin commands to get
717 * the number of I/O queues supported, so cannot reset
718 * the adminq again here.
720 if (ctrlr->is_resetting) {
721 nvme_qpair_reset(&ctrlr->adminq);
724 for (i = 0; i < ctrlr->num_io_queues; i++)
725 nvme_qpair_reset(&ctrlr->ioq[i]);
727 nvme_admin_qpair_enable(&ctrlr->adminq);
729 if (nvme_ctrlr_identify(ctrlr) != 0) {
730 nvme_ctrlr_fail(ctrlr);
735 * The number of qpairs are determined during controller initialization,
736 * including using NVMe SET_FEATURES/NUMBER_OF_QUEUES to determine the
737 * HW limit. We call SET_FEATURES again here so that it gets called
738 * after any reset for controllers that depend on the driver to
739 * explicit specify how many queues it will use. This value should
740 * never change between resets, so panic if somehow that does happen.
742 if (ctrlr->is_resetting) {
743 old_num_io_queues = ctrlr->num_io_queues;
744 if (nvme_ctrlr_set_num_qpairs(ctrlr) != 0) {
745 nvme_ctrlr_fail(ctrlr);
749 if (old_num_io_queues != ctrlr->num_io_queues) {
750 panic("num_io_queues changed from %u to %u",
751 old_num_io_queues, ctrlr->num_io_queues);
755 if (nvme_ctrlr_create_qpairs(ctrlr) != 0) {
756 nvme_ctrlr_fail(ctrlr);
760 if (nvme_ctrlr_construct_namespaces(ctrlr) != 0) {
761 nvme_ctrlr_fail(ctrlr);
765 nvme_ctrlr_configure_aer(ctrlr);
766 nvme_ctrlr_configure_int_coalescing(ctrlr);
768 for (i = 0; i < ctrlr->num_io_queues; i++)
769 nvme_io_qpair_enable(&ctrlr->ioq[i]);
773 nvme_ctrlr_start_config_hook(void *arg)
775 struct nvme_controller *ctrlr = arg;
777 nvme_qpair_reset(&ctrlr->adminq);
778 nvme_admin_qpair_enable(&ctrlr->adminq);
780 if (nvme_ctrlr_set_num_qpairs(ctrlr) == 0 &&
781 nvme_ctrlr_construct_io_qpairs(ctrlr) == 0)
782 nvme_ctrlr_start(ctrlr);
784 nvme_ctrlr_fail(ctrlr);
786 nvme_sysctl_initialize_ctrlr(ctrlr);
787 config_intrhook_disestablish(&ctrlr->config_hook);
789 ctrlr->is_initialized = 1;
790 nvme_notify_new_controller(ctrlr);
794 nvme_ctrlr_reset_task(void *arg, int pending)
796 struct nvme_controller *ctrlr = arg;
799 nvme_printf(ctrlr, "resetting controller\n");
800 status = nvme_ctrlr_hw_reset(ctrlr);
802 * Use pause instead of DELAY, so that we yield to any nvme interrupt
803 * handlers on this CPU that were blocked on a qpair lock. We want
804 * all nvme interrupts completed before proceeding with restarting the
807 * XXX - any way to guarantee the interrupt handlers have quiesced?
809 pause("nvmereset", hz / 10);
811 nvme_ctrlr_start(ctrlr);
813 nvme_ctrlr_fail(ctrlr);
815 atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
819 * Poll all the queues enabled on the device for completion.
822 nvme_ctrlr_poll(struct nvme_controller *ctrlr)
826 nvme_qpair_process_completions(&ctrlr->adminq);
828 for (i = 0; i < ctrlr->num_io_queues; i++)
829 if (ctrlr->ioq && ctrlr->ioq[i].cpl)
830 nvme_qpair_process_completions(&ctrlr->ioq[i]);
834 * Poll the single-vector intertrupt case: num_io_queues will be 1 and
835 * there's only a single vector. While we're polling, we mask further
836 * interrupts in the controller.
839 nvme_ctrlr_intx_handler(void *arg)
841 struct nvme_controller *ctrlr = arg;
843 nvme_mmio_write_4(ctrlr, intms, 1);
844 nvme_ctrlr_poll(ctrlr);
845 nvme_mmio_write_4(ctrlr, intmc, 1);
849 nvme_ctrlr_configure_intx(struct nvme_controller *ctrlr)
852 ctrlr->msix_enabled = 0;
853 ctrlr->num_io_queues = 1;
854 ctrlr->num_cpus_per_ioq = mp_ncpus;
856 ctrlr->res = bus_alloc_resource_any(ctrlr->dev, SYS_RES_IRQ,
857 &ctrlr->rid, RF_SHAREABLE | RF_ACTIVE);
859 if (ctrlr->res == NULL) {
860 nvme_printf(ctrlr, "unable to allocate shared IRQ\n");
864 bus_setup_intr(ctrlr->dev, ctrlr->res,
865 INTR_TYPE_MISC | INTR_MPSAFE, NULL, nvme_ctrlr_intx_handler,
868 if (ctrlr->tag == NULL) {
869 nvme_printf(ctrlr, "unable to setup intx handler\n");
877 nvme_pt_done(void *arg, const struct nvme_completion *cpl)
879 struct nvme_pt_command *pt = arg;
881 bzero(&pt->cpl, sizeof(pt->cpl));
882 pt->cpl.cdw0 = cpl->cdw0;
883 pt->cpl.status = cpl->status;
884 pt->cpl.status.p = 0;
886 mtx_lock(pt->driver_lock);
888 mtx_unlock(pt->driver_lock);
892 nvme_ctrlr_passthrough_cmd(struct nvme_controller *ctrlr,
893 struct nvme_pt_command *pt, uint32_t nsid, int is_user_buffer,
896 struct nvme_request *req;
898 struct buf *buf = NULL;
900 vm_offset_t addr, end;
904 * vmapbuf calls vm_fault_quick_hold_pages which only maps full
905 * pages. Ensure this request has fewer than MAXPHYS bytes when
906 * extended to full pages.
908 addr = (vm_offset_t)pt->buf;
909 end = round_page(addr + pt->len);
910 addr = trunc_page(addr);
911 if (end - addr > MAXPHYS)
914 if (pt->len > ctrlr->max_xfer_size) {
915 nvme_printf(ctrlr, "pt->len (%d) "
916 "exceeds max_xfer_size (%d)\n", pt->len,
917 ctrlr->max_xfer_size);
920 if (is_user_buffer) {
922 * Ensure the user buffer is wired for the duration of
923 * this passthrough command.
927 buf->b_data = pt->buf;
928 buf->b_bufsize = pt->len;
929 buf->b_iocmd = pt->is_read ? BIO_READ : BIO_WRITE;
930 #ifdef NVME_UNMAPPED_BIO_SUPPORT
931 if (vmapbuf(buf, 1) < 0) {
933 if (vmapbuf(buf) < 0) {
938 req = nvme_allocate_request_vaddr(buf->b_data, pt->len,
941 req = nvme_allocate_request_vaddr(pt->buf, pt->len,
944 req = nvme_allocate_request_null(nvme_pt_done, pt);
946 req->cmd.opc = pt->cmd.opc;
947 req->cmd.cdw10 = pt->cmd.cdw10;
948 req->cmd.cdw11 = pt->cmd.cdw11;
949 req->cmd.cdw12 = pt->cmd.cdw12;
950 req->cmd.cdw13 = pt->cmd.cdw13;
951 req->cmd.cdw14 = pt->cmd.cdw14;
952 req->cmd.cdw15 = pt->cmd.cdw15;
954 req->cmd.nsid = nsid;
959 mtx = &ctrlr->ns[nsid-1].lock;
962 pt->driver_lock = mtx;
965 nvme_ctrlr_submit_admin_request(ctrlr, req);
967 nvme_ctrlr_submit_io_request(ctrlr, req);
969 mtx_sleep(pt, mtx, PRIBIO, "nvme_pt", 0);
972 pt->driver_lock = NULL;
984 nvme_ctrlr_ioctl(struct cdev *cdev, u_long cmd, caddr_t arg, int flag,
987 struct nvme_controller *ctrlr;
988 struct nvme_pt_command *pt;
990 ctrlr = cdev->si_drv1;
993 case NVME_RESET_CONTROLLER:
994 nvme_ctrlr_reset(ctrlr);
996 case NVME_PASSTHROUGH_CMD:
997 pt = (struct nvme_pt_command *)arg;
998 return (nvme_ctrlr_passthrough_cmd(ctrlr, pt, pt->cmd.nsid,
999 1 /* is_user_buffer */, 1 /* is_admin_cmd */));
1007 static struct cdevsw nvme_ctrlr_cdevsw = {
1008 .d_version = D_VERSION,
1010 .d_ioctl = nvme_ctrlr_ioctl
1014 nvme_ctrlr_setup_interrupts(struct nvme_controller *ctrlr)
1017 int per_cpu_io_queues;
1018 int min_cpus_per_ioq;
1019 int num_vectors_requested, num_vectors_allocated;
1020 int num_vectors_available;
1023 min_cpus_per_ioq = 1;
1024 TUNABLE_INT_FETCH("hw.nvme.min_cpus_per_ioq", &min_cpus_per_ioq);
1026 if (min_cpus_per_ioq < 1) {
1027 min_cpus_per_ioq = 1;
1028 } else if (min_cpus_per_ioq > mp_ncpus) {
1029 min_cpus_per_ioq = mp_ncpus;
1032 per_cpu_io_queues = 1;
1033 TUNABLE_INT_FETCH("hw.nvme.per_cpu_io_queues", &per_cpu_io_queues);
1035 if (per_cpu_io_queues == 0) {
1036 min_cpus_per_ioq = mp_ncpus;
1039 ctrlr->force_intx = 0;
1040 TUNABLE_INT_FETCH("hw.nvme.force_intx", &ctrlr->force_intx);
1043 * FreeBSD currently cannot allocate more than about 190 vectors at
1044 * boot, meaning that systems with high core count and many devices
1045 * requesting per-CPU interrupt vectors will not get their full
1046 * allotment. So first, try to allocate as many as we may need to
1047 * understand what is available, then immediately release them.
1048 * Then figure out how many of those we will actually use, based on
1049 * assigning an equal number of cores to each I/O queue.
1052 /* One vector for per core I/O queue, plus one vector for admin queue. */
1053 num_vectors_available = min(pci_msix_count(dev), mp_ncpus + 1);
1054 if (pci_alloc_msix(dev, &num_vectors_available) != 0) {
1055 num_vectors_available = 0;
1057 pci_release_msi(dev);
1059 if (ctrlr->force_intx || num_vectors_available < 2) {
1060 nvme_ctrlr_configure_intx(ctrlr);
1065 * Do not use all vectors for I/O queues - one must be saved for the
1068 ctrlr->num_cpus_per_ioq = max(min_cpus_per_ioq,
1069 howmany(mp_ncpus, num_vectors_available - 1));
1071 ctrlr->num_io_queues = howmany(mp_ncpus, ctrlr->num_cpus_per_ioq);
1072 num_vectors_requested = ctrlr->num_io_queues + 1;
1073 num_vectors_allocated = num_vectors_requested;
1076 * Now just allocate the number of vectors we need. This should
1077 * succeed, since we previously called pci_alloc_msix()
1078 * successfully returning at least this many vectors, but just to
1079 * be safe, if something goes wrong just revert to INTx.
1081 if (pci_alloc_msix(dev, &num_vectors_allocated) != 0) {
1082 nvme_ctrlr_configure_intx(ctrlr);
1086 if (num_vectors_allocated < num_vectors_requested) {
1087 pci_release_msi(dev);
1088 nvme_ctrlr_configure_intx(ctrlr);
1092 ctrlr->msix_enabled = 1;
1096 nvme_ctrlr_construct(struct nvme_controller *ctrlr, device_t dev)
1098 union cap_lo_register cap_lo;
1099 union cap_hi_register cap_hi;
1100 int status, timeout_period;
1104 mtx_init(&ctrlr->lock, "nvme ctrlr lock", NULL, MTX_DEF);
1106 status = nvme_ctrlr_allocate_bar(ctrlr);
1112 * Software emulators may set the doorbell stride to something
1113 * other than zero, but this driver is not set up to handle that.
1115 cap_hi.raw = nvme_mmio_read_4(ctrlr, cap_hi);
1116 if (cap_hi.bits.dstrd != 0)
1119 ctrlr->min_page_size = 1 << (12 + cap_hi.bits.mpsmin);
1121 /* Get ready timeout value from controller, in units of 500ms. */
1122 cap_lo.raw = nvme_mmio_read_4(ctrlr, cap_lo);
1123 ctrlr->ready_timeout_in_ms = cap_lo.bits.to * 500;
1125 timeout_period = NVME_DEFAULT_TIMEOUT_PERIOD;
1126 TUNABLE_INT_FETCH("hw.nvme.timeout_period", &timeout_period);
1127 timeout_period = min(timeout_period, NVME_MAX_TIMEOUT_PERIOD);
1128 timeout_period = max(timeout_period, NVME_MIN_TIMEOUT_PERIOD);
1129 ctrlr->timeout_period = timeout_period;
1131 nvme_retry_count = NVME_DEFAULT_RETRY_COUNT;
1132 TUNABLE_INT_FETCH("hw.nvme.retry_count", &nvme_retry_count);
1134 ctrlr->enable_aborts = 0;
1135 TUNABLE_INT_FETCH("hw.nvme.enable_aborts", &ctrlr->enable_aborts);
1137 nvme_ctrlr_setup_interrupts(ctrlr);
1139 ctrlr->max_xfer_size = NVME_MAX_XFER_SIZE;
1140 if (nvme_ctrlr_construct_admin_qpair(ctrlr) != 0)
1143 ctrlr->cdev = make_dev(&nvme_ctrlr_cdevsw, device_get_unit(dev),
1144 UID_ROOT, GID_WHEEL, 0600, "nvme%d", device_get_unit(dev));
1146 if (ctrlr->cdev == NULL)
1149 ctrlr->cdev->si_drv1 = (void *)ctrlr;
1151 ctrlr->taskqueue = taskqueue_create("nvme_taskq", M_WAITOK,
1152 taskqueue_thread_enqueue, &ctrlr->taskqueue);
1153 taskqueue_start_threads(&ctrlr->taskqueue, 1, PI_DISK, "nvme taskq");
1155 ctrlr->is_resetting = 0;
1156 ctrlr->is_initialized = 0;
1157 ctrlr->notification_sent = 0;
1158 TASK_INIT(&ctrlr->reset_task, 0, nvme_ctrlr_reset_task, ctrlr);
1160 TASK_INIT(&ctrlr->fail_req_task, 0, nvme_ctrlr_fail_req_task, ctrlr);
1161 STAILQ_INIT(&ctrlr->fail_req);
1162 ctrlr->is_failed = FALSE;
1168 nvme_ctrlr_destruct(struct nvme_controller *ctrlr, device_t dev)
1173 * Notify the controller of a shutdown, even though this is due to
1174 * a driver unload, not a system shutdown (this path is not invoked
1175 * during shutdown). This ensures the controller receives a
1176 * shutdown notification in case the system is shutdown before
1177 * reloading the driver.
1179 nvme_ctrlr_shutdown(ctrlr);
1181 nvme_ctrlr_disable(ctrlr);
1182 taskqueue_free(ctrlr->taskqueue);
1184 for (i = 0; i < NVME_MAX_NAMESPACES; i++)
1185 nvme_ns_destruct(&ctrlr->ns[i]);
1188 destroy_dev(ctrlr->cdev);
1190 for (i = 0; i < ctrlr->num_io_queues; i++) {
1191 nvme_io_qpair_destroy(&ctrlr->ioq[i]);
1194 free(ctrlr->ioq, M_NVME);
1196 nvme_admin_qpair_destroy(&ctrlr->adminq);
1198 if (ctrlr->resource != NULL) {
1199 bus_release_resource(dev, SYS_RES_MEMORY,
1200 ctrlr->resource_id, ctrlr->resource);
1203 if (ctrlr->bar4_resource != NULL) {
1204 bus_release_resource(dev, SYS_RES_MEMORY,
1205 ctrlr->bar4_resource_id, ctrlr->bar4_resource);
1209 bus_teardown_intr(ctrlr->dev, ctrlr->res, ctrlr->tag);
1212 bus_release_resource(ctrlr->dev, SYS_RES_IRQ,
1213 rman_get_rid(ctrlr->res), ctrlr->res);
1215 if (ctrlr->msix_enabled)
1216 pci_release_msi(dev);
1220 nvme_ctrlr_shutdown(struct nvme_controller *ctrlr)
1222 union cc_register cc;
1223 union csts_register csts;
1226 cc.raw = nvme_mmio_read_4(ctrlr, cc);
1227 cc.bits.shn = NVME_SHN_NORMAL;
1228 nvme_mmio_write_4(ctrlr, cc, cc.raw);
1229 csts.raw = nvme_mmio_read_4(ctrlr, csts);
1230 while ((csts.bits.shst != NVME_SHST_COMPLETE) && (ticks++ < 5*hz)) {
1231 pause("nvme shn", 1);
1232 csts.raw = nvme_mmio_read_4(ctrlr, csts);
1234 if (csts.bits.shst != NVME_SHST_COMPLETE)
1235 nvme_printf(ctrlr, "did not complete shutdown within 5 seconds "
1236 "of notification\n");
1240 nvme_ctrlr_submit_admin_request(struct nvme_controller *ctrlr,
1241 struct nvme_request *req)
1244 nvme_qpair_submit_request(&ctrlr->adminq, req);
1248 nvme_ctrlr_submit_io_request(struct nvme_controller *ctrlr,
1249 struct nvme_request *req)
1251 struct nvme_qpair *qpair;
1253 qpair = &ctrlr->ioq[curcpu / ctrlr->num_cpus_per_ioq];
1254 nvme_qpair_submit_request(qpair, req);
1258 nvme_ctrlr_get_device(struct nvme_controller *ctrlr)
1261 return (ctrlr->dev);
1264 const struct nvme_controller_data *
1265 nvme_ctrlr_get_data(struct nvme_controller *ctrlr)
1268 return (&ctrlr->cdata);