2 * Copyright (C) 2012-2014 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);
49 nvme_ctrlr_allocate_bar(struct nvme_controller *ctrlr)
52 /* Chatham puts the NVMe MMRs behind BAR 2/3, not BAR 0/1. */
53 if (pci_get_devid(ctrlr->dev) == CHATHAM_PCI_ID)
54 ctrlr->resource_id = PCIR_BAR(2);
56 ctrlr->resource_id = PCIR_BAR(0);
58 ctrlr->resource = bus_alloc_resource(ctrlr->dev, SYS_RES_MEMORY,
59 &ctrlr->resource_id, 0, ~0, 1, RF_ACTIVE);
61 if(ctrlr->resource == NULL) {
62 nvme_printf(ctrlr, "unable to allocate pci resource\n");
66 ctrlr->bus_tag = rman_get_bustag(ctrlr->resource);
67 ctrlr->bus_handle = rman_get_bushandle(ctrlr->resource);
68 ctrlr->regs = (struct nvme_registers *)ctrlr->bus_handle;
71 * The NVMe spec allows for the MSI-X table to be placed behind
72 * BAR 4/5, separate from the control/doorbell registers. Always
73 * try to map this bar, because it must be mapped prior to calling
74 * pci_alloc_msix(). If the table isn't behind BAR 4/5,
75 * bus_alloc_resource() will just return NULL which is OK.
77 ctrlr->bar4_resource_id = PCIR_BAR(4);
78 ctrlr->bar4_resource = bus_alloc_resource(ctrlr->dev, SYS_RES_MEMORY,
79 &ctrlr->bar4_resource_id, 0, ~0, 1, RF_ACTIVE);
86 nvme_ctrlr_allocate_chatham_bar(struct nvme_controller *ctrlr)
89 ctrlr->chatham_resource_id = PCIR_BAR(CHATHAM_CONTROL_BAR);
90 ctrlr->chatham_resource = bus_alloc_resource(ctrlr->dev,
91 SYS_RES_MEMORY, &ctrlr->chatham_resource_id, 0, ~0, 1,
94 if(ctrlr->chatham_resource == NULL) {
95 nvme_printf(ctrlr, "unable to alloc pci resource\n");
99 ctrlr->chatham_bus_tag = rman_get_bustag(ctrlr->chatham_resource);
100 ctrlr->chatham_bus_handle =
101 rman_get_bushandle(ctrlr->chatham_resource);
107 nvme_ctrlr_setup_chatham(struct nvme_controller *ctrlr)
109 uint64_t reg1, reg2, reg3;
110 uint64_t temp1, temp2;
112 uint32_t use_flash_timings = 0;
116 temp3 = chatham_read_4(ctrlr, 0x8080);
118 device_printf(ctrlr->dev, "Chatham version: 0x%x\n", temp3);
120 ctrlr->chatham_lbas = chatham_read_4(ctrlr, 0x8068) - 0x110;
121 ctrlr->chatham_size = ctrlr->chatham_lbas * 512;
123 device_printf(ctrlr->dev, "Chatham size: %jd\n",
124 (intmax_t)ctrlr->chatham_size);
126 reg1 = reg2 = reg3 = ctrlr->chatham_size - 1;
128 TUNABLE_INT_FETCH("hw.nvme.use_flash_timings", &use_flash_timings);
129 if (use_flash_timings) {
130 device_printf(ctrlr->dev, "Chatham: using flash timings\n");
131 temp1 = 0x00001b58000007d0LL;
132 temp2 = 0x000000cb00000131LL;
134 device_printf(ctrlr->dev, "Chatham: using DDR timings\n");
135 temp1 = temp2 = 0x0LL;
138 chatham_write_8(ctrlr, 0x8000, reg1);
139 chatham_write_8(ctrlr, 0x8008, reg2);
140 chatham_write_8(ctrlr, 0x8010, reg3);
142 chatham_write_8(ctrlr, 0x8020, temp1);
143 temp3 = chatham_read_4(ctrlr, 0x8020);
145 chatham_write_8(ctrlr, 0x8028, temp2);
146 temp3 = chatham_read_4(ctrlr, 0x8028);
148 chatham_write_8(ctrlr, 0x8030, temp1);
149 chatham_write_8(ctrlr, 0x8038, temp2);
150 chatham_write_8(ctrlr, 0x8040, temp1);
151 chatham_write_8(ctrlr, 0x8048, temp2);
152 chatham_write_8(ctrlr, 0x8050, temp1);
153 chatham_write_8(ctrlr, 0x8058, temp2);
159 nvme_chatham_populate_cdata(struct nvme_controller *ctrlr)
161 struct nvme_controller_data *cdata;
163 cdata = &ctrlr->cdata;
166 cdata->ssvid = 0x2011;
169 * Chatham2 puts garbage data in these fields when we
170 * invoke IDENTIFY_CONTROLLER, so we need to re-zero
171 * the fields before calling bcopy().
173 memset(cdata->sn, 0, sizeof(cdata->sn));
174 memcpy(cdata->sn, "2012", strlen("2012"));
175 memset(cdata->mn, 0, sizeof(cdata->mn));
176 memcpy(cdata->mn, "CHATHAM2", strlen("CHATHAM2"));
177 memset(cdata->fr, 0, sizeof(cdata->fr));
178 memcpy(cdata->fr, "0", strlen("0"));
181 cdata->lpa.ns_smart = 1;
188 /* Chatham2 doesn't support DSM command */
191 cdata->vwc.present = 1;
193 #endif /* CHATHAM2 */
196 nvme_ctrlr_construct_admin_qpair(struct nvme_controller *ctrlr)
198 struct nvme_qpair *qpair;
199 uint32_t num_entries;
201 qpair = &ctrlr->adminq;
203 num_entries = NVME_ADMIN_ENTRIES;
204 TUNABLE_INT_FETCH("hw.nvme.admin_entries", &num_entries);
206 * If admin_entries was overridden to an invalid value, revert it
207 * back to our default value.
209 if (num_entries < NVME_MIN_ADMIN_ENTRIES ||
210 num_entries > NVME_MAX_ADMIN_ENTRIES) {
211 nvme_printf(ctrlr, "invalid hw.nvme.admin_entries=%d "
212 "specified\n", num_entries);
213 num_entries = NVME_ADMIN_ENTRIES;
217 * The admin queue's max xfer size is treated differently than the
218 * max I/O xfer size. 16KB is sufficient here - maybe even less?
220 nvme_qpair_construct(qpair,
229 nvme_ctrlr_construct_io_qpairs(struct nvme_controller *ctrlr)
231 struct nvme_qpair *qpair;
232 union cap_lo_register cap_lo;
233 int i, num_entries, num_trackers;
235 num_entries = NVME_IO_ENTRIES;
236 TUNABLE_INT_FETCH("hw.nvme.io_entries", &num_entries);
239 * NVMe spec sets a hard limit of 64K max entries, but
240 * devices may specify a smaller limit, so we need to check
241 * the MQES field in the capabilities register.
243 cap_lo.raw = nvme_mmio_read_4(ctrlr, cap_lo);
244 num_entries = min(num_entries, cap_lo.bits.mqes+1);
246 num_trackers = NVME_IO_TRACKERS;
247 TUNABLE_INT_FETCH("hw.nvme.io_trackers", &num_trackers);
249 num_trackers = max(num_trackers, NVME_MIN_IO_TRACKERS);
250 num_trackers = min(num_trackers, NVME_MAX_IO_TRACKERS);
252 * No need to have more trackers than entries in the submit queue.
253 * Note also that for a queue size of N, we can only have (N-1)
254 * commands outstanding, hence the "-1" here.
256 num_trackers = min(num_trackers, (num_entries-1));
258 ctrlr->ioq = malloc(ctrlr->num_io_queues * sizeof(struct nvme_qpair),
259 M_NVME, M_ZERO | M_WAITOK);
261 for (i = 0; i < ctrlr->num_io_queues; i++) {
262 qpair = &ctrlr->ioq[i];
265 * Admin queue has ID=0. IO queues start at ID=1 -
266 * hence the 'i+1' here.
268 * For I/O queues, use the controller-wide max_xfer_size
269 * calculated in nvme_attach().
271 nvme_qpair_construct(qpair,
273 ctrlr->msix_enabled ? i+1 : 0, /* vector */
278 if (ctrlr->per_cpu_io_queues)
279 bus_bind_intr(ctrlr->dev, qpair->res, i);
286 nvme_ctrlr_fail(struct nvme_controller *ctrlr)
290 ctrlr->is_failed = TRUE;
291 nvme_qpair_fail(&ctrlr->adminq);
292 for (i = 0; i < ctrlr->num_io_queues; i++)
293 nvme_qpair_fail(&ctrlr->ioq[i]);
294 nvme_notify_fail_consumers(ctrlr);
298 nvme_ctrlr_post_failed_request(struct nvme_controller *ctrlr,
299 struct nvme_request *req)
302 mtx_lock(&ctrlr->lock);
303 STAILQ_INSERT_TAIL(&ctrlr->fail_req, req, stailq);
304 mtx_unlock(&ctrlr->lock);
305 taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->fail_req_task);
309 nvme_ctrlr_fail_req_task(void *arg, int pending)
311 struct nvme_controller *ctrlr = arg;
312 struct nvme_request *req;
314 mtx_lock(&ctrlr->lock);
315 while (!STAILQ_EMPTY(&ctrlr->fail_req)) {
316 req = STAILQ_FIRST(&ctrlr->fail_req);
317 STAILQ_REMOVE_HEAD(&ctrlr->fail_req, stailq);
318 nvme_qpair_manual_complete_request(req->qpair, req,
319 NVME_SCT_GENERIC, NVME_SC_ABORTED_BY_REQUEST, TRUE);
321 mtx_unlock(&ctrlr->lock);
325 nvme_ctrlr_wait_for_ready(struct nvme_controller *ctrlr)
328 union cc_register cc;
329 union csts_register csts;
331 cc.raw = nvme_mmio_read_4(ctrlr, cc);
332 csts.raw = nvme_mmio_read_4(ctrlr, csts);
335 nvme_printf(ctrlr, "%s called with cc.en = 0\n", __func__);
341 while (!csts.bits.rdy) {
343 if (ms_waited++ > ctrlr->ready_timeout_in_ms) {
344 nvme_printf(ctrlr, "controller did not become ready "
345 "within %d ms\n", ctrlr->ready_timeout_in_ms);
348 csts.raw = nvme_mmio_read_4(ctrlr, csts);
355 nvme_ctrlr_disable(struct nvme_controller *ctrlr)
357 union cc_register cc;
358 union csts_register csts;
360 cc.raw = nvme_mmio_read_4(ctrlr, cc);
361 csts.raw = nvme_mmio_read_4(ctrlr, csts);
363 if (cc.bits.en == 1 && csts.bits.rdy == 0)
364 nvme_ctrlr_wait_for_ready(ctrlr);
367 nvme_mmio_write_4(ctrlr, cc, cc.raw);
372 nvme_ctrlr_enable(struct nvme_controller *ctrlr)
374 union cc_register cc;
375 union csts_register csts;
376 union aqa_register aqa;
378 cc.raw = nvme_mmio_read_4(ctrlr, cc);
379 csts.raw = nvme_mmio_read_4(ctrlr, csts);
381 if (cc.bits.en == 1) {
382 if (csts.bits.rdy == 1)
385 return (nvme_ctrlr_wait_for_ready(ctrlr));
388 nvme_mmio_write_8(ctrlr, asq, ctrlr->adminq.cmd_bus_addr);
390 nvme_mmio_write_8(ctrlr, acq, ctrlr->adminq.cpl_bus_addr);
394 /* acqs and asqs are 0-based. */
395 aqa.bits.acqs = ctrlr->adminq.num_entries-1;
396 aqa.bits.asqs = ctrlr->adminq.num_entries-1;
397 nvme_mmio_write_4(ctrlr, aqa, aqa.raw);
404 cc.bits.iosqes = 6; /* SQ entry size == 64 == 2^6 */
405 cc.bits.iocqes = 4; /* CQ entry size == 16 == 2^4 */
407 /* This evaluates to 0, which is according to spec. */
408 cc.bits.mps = (PAGE_SIZE >> 13);
410 nvme_mmio_write_4(ctrlr, cc, cc.raw);
413 return (nvme_ctrlr_wait_for_ready(ctrlr));
417 nvme_ctrlr_hw_reset(struct nvme_controller *ctrlr)
421 nvme_admin_qpair_disable(&ctrlr->adminq);
422 for (i = 0; i < ctrlr->num_io_queues; i++)
423 nvme_io_qpair_disable(&ctrlr->ioq[i]);
427 nvme_ctrlr_disable(ctrlr);
428 return (nvme_ctrlr_enable(ctrlr));
432 nvme_ctrlr_reset(struct nvme_controller *ctrlr)
436 cmpset = atomic_cmpset_32(&ctrlr->is_resetting, 0, 1);
438 if (cmpset == 0 || ctrlr->is_failed)
440 * Controller is already resetting or has failed. Return
441 * immediately since there is no need to kick off another
442 * reset in these cases.
446 taskqueue_enqueue(ctrlr->taskqueue, &ctrlr->reset_task);
450 nvme_ctrlr_identify(struct nvme_controller *ctrlr)
452 struct nvme_completion_poll_status status;
455 nvme_ctrlr_cmd_identify_controller(ctrlr, &ctrlr->cdata,
456 nvme_completion_poll_cb, &status);
457 while (status.done == FALSE)
459 if (nvme_completion_is_error(&status.cpl)) {
460 nvme_printf(ctrlr, "nvme_identify_controller failed!\n");
465 if (pci_get_devid(ctrlr->dev) == CHATHAM_PCI_ID)
466 nvme_chatham_populate_cdata(ctrlr);
470 * Use MDTS to ensure our default max_xfer_size doesn't exceed what the
471 * controller supports.
473 if (ctrlr->cdata.mdts > 0)
474 ctrlr->max_xfer_size = min(ctrlr->max_xfer_size,
475 ctrlr->min_page_size * (1 << (ctrlr->cdata.mdts)));
481 nvme_ctrlr_set_num_qpairs(struct nvme_controller *ctrlr)
483 struct nvme_completion_poll_status status;
484 int cq_allocated, i, sq_allocated;
487 nvme_ctrlr_cmd_set_num_queues(ctrlr, ctrlr->num_io_queues,
488 nvme_completion_poll_cb, &status);
489 while (status.done == FALSE)
491 if (nvme_completion_is_error(&status.cpl)) {
492 nvme_printf(ctrlr, "nvme_set_num_queues failed!\n");
497 * Data in cdw0 is 0-based.
498 * Lower 16-bits indicate number of submission queues allocated.
499 * Upper 16-bits indicate number of completion queues allocated.
501 sq_allocated = (status.cpl.cdw0 & 0xFFFF) + 1;
502 cq_allocated = (status.cpl.cdw0 >> 16) + 1;
505 * Check that the controller was able to allocate the number of
506 * queues we requested. If not, revert to one IO queue pair.
508 if (sq_allocated < ctrlr->num_io_queues ||
509 cq_allocated < ctrlr->num_io_queues) {
512 * Destroy extra IO queue pairs that were created at
513 * controller construction time but are no longer
514 * needed. This will only happen when a controller
515 * supports fewer queues than MSI-X vectors. This
516 * is not the normal case, but does occur with the
517 * Chatham prototype board.
519 for (i = 1; i < ctrlr->num_io_queues; i++)
520 nvme_io_qpair_destroy(&ctrlr->ioq[i]);
522 ctrlr->num_io_queues = 1;
523 ctrlr->per_cpu_io_queues = 0;
530 nvme_ctrlr_create_qpairs(struct nvme_controller *ctrlr)
532 struct nvme_completion_poll_status status;
533 struct nvme_qpair *qpair;
536 for (i = 0; i < ctrlr->num_io_queues; i++) {
537 qpair = &ctrlr->ioq[i];
540 nvme_ctrlr_cmd_create_io_cq(ctrlr, qpair, qpair->vector,
541 nvme_completion_poll_cb, &status);
542 while (status.done == FALSE)
544 if (nvme_completion_is_error(&status.cpl)) {
545 nvme_printf(ctrlr, "nvme_create_io_cq failed!\n");
550 nvme_ctrlr_cmd_create_io_sq(qpair->ctrlr, qpair,
551 nvme_completion_poll_cb, &status);
552 while (status.done == FALSE)
554 if (nvme_completion_is_error(&status.cpl)) {
555 nvme_printf(ctrlr, "nvme_create_io_sq failed!\n");
564 nvme_ctrlr_construct_namespaces(struct nvme_controller *ctrlr)
566 struct nvme_namespace *ns;
569 for (i = 0; i < ctrlr->cdata.nn; i++) {
571 status = nvme_ns_construct(ns, i+1, ctrlr);
580 is_log_page_id_valid(uint8_t page_id)
585 case NVME_LOG_HEALTH_INFORMATION:
586 case NVME_LOG_FIRMWARE_SLOT:
594 nvme_ctrlr_get_log_page_size(struct nvme_controller *ctrlr, uint8_t page_id)
596 uint32_t log_page_size;
601 sizeof(struct nvme_error_information_entry) *
603 NVME_MAX_AER_LOG_SIZE);
605 case NVME_LOG_HEALTH_INFORMATION:
606 log_page_size = sizeof(struct nvme_health_information_page);
608 case NVME_LOG_FIRMWARE_SLOT:
609 log_page_size = sizeof(struct nvme_firmware_page);
616 return (log_page_size);
620 nvme_ctrlr_log_critical_warnings(struct nvme_controller *ctrlr,
621 union nvme_critical_warning_state state)
624 if (state.bits.available_spare == 1)
625 nvme_printf(ctrlr, "available spare space below threshold\n");
627 if (state.bits.temperature == 1)
628 nvme_printf(ctrlr, "temperature above threshold\n");
630 if (state.bits.device_reliability == 1)
631 nvme_printf(ctrlr, "device reliability degraded\n");
633 if (state.bits.read_only == 1)
634 nvme_printf(ctrlr, "media placed in read only mode\n");
636 if (state.bits.volatile_memory_backup == 1)
637 nvme_printf(ctrlr, "volatile memory backup device failed\n");
639 if (state.bits.reserved != 0)
641 "unknown critical warning(s): state = 0x%02x\n", state.raw);
645 nvme_ctrlr_async_event_log_page_cb(void *arg, const struct nvme_completion *cpl)
647 struct nvme_async_event_request *aer = arg;
648 struct nvme_health_information_page *health_info;
651 * If the log page fetch for some reason completed with an error,
652 * don't pass log page data to the consumers. In practice, this case
653 * should never happen.
655 if (nvme_completion_is_error(cpl))
656 nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
657 aer->log_page_id, NULL, 0);
659 if (aer->log_page_id == NVME_LOG_HEALTH_INFORMATION) {
660 health_info = (struct nvme_health_information_page *)
661 aer->log_page_buffer;
662 nvme_ctrlr_log_critical_warnings(aer->ctrlr,
663 health_info->critical_warning);
665 * Critical warnings reported through the
666 * SMART/health log page are persistent, so
667 * clear the associated bits in the async event
668 * config so that we do not receive repeated
669 * notifications for the same event.
671 aer->ctrlr->async_event_config.raw &=
672 ~health_info->critical_warning.raw;
673 nvme_ctrlr_cmd_set_async_event_config(aer->ctrlr,
674 aer->ctrlr->async_event_config, NULL, NULL);
679 * Pass the cpl data from the original async event completion,
680 * not the log page fetch.
682 nvme_notify_async_consumers(aer->ctrlr, &aer->cpl,
683 aer->log_page_id, aer->log_page_buffer, aer->log_page_size);
687 * Repost another asynchronous event request to replace the one
688 * that just completed.
690 nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
694 nvme_ctrlr_async_event_cb(void *arg, const struct nvme_completion *cpl)
696 struct nvme_async_event_request *aer = arg;
698 if (nvme_completion_is_error(cpl)) {
700 * Do not retry failed async event requests. This avoids
701 * infinite loops where a new async event request is submitted
702 * to replace the one just failed, only to fail again and
703 * perpetuate the loop.
708 /* Associated log page is in bits 23:16 of completion entry dw0. */
709 aer->log_page_id = (cpl->cdw0 & 0xFF0000) >> 16;
711 nvme_printf(aer->ctrlr, "async event occurred (log page id=0x%x)\n",
714 if (is_log_page_id_valid(aer->log_page_id)) {
715 aer->log_page_size = nvme_ctrlr_get_log_page_size(aer->ctrlr,
717 memcpy(&aer->cpl, cpl, sizeof(*cpl));
718 nvme_ctrlr_cmd_get_log_page(aer->ctrlr, aer->log_page_id,
719 NVME_GLOBAL_NAMESPACE_TAG, aer->log_page_buffer,
720 aer->log_page_size, nvme_ctrlr_async_event_log_page_cb,
722 /* Wait to notify consumers until after log page is fetched. */
724 nvme_notify_async_consumers(aer->ctrlr, cpl, aer->log_page_id,
728 * Repost another asynchronous event request to replace the one
729 * that just completed.
731 nvme_ctrlr_construct_and_submit_aer(aer->ctrlr, aer);
736 nvme_ctrlr_construct_and_submit_aer(struct nvme_controller *ctrlr,
737 struct nvme_async_event_request *aer)
739 struct nvme_request *req;
742 req = nvme_allocate_request_null(nvme_ctrlr_async_event_cb, aer);
746 * Disable timeout here, since asynchronous event requests should by
747 * nature never be timed out.
749 req->timeout = FALSE;
750 req->cmd.opc = NVME_OPC_ASYNC_EVENT_REQUEST;
751 nvme_ctrlr_submit_admin_request(ctrlr, req);
755 nvme_ctrlr_configure_aer(struct nvme_controller *ctrlr)
757 struct nvme_completion_poll_status status;
758 struct nvme_async_event_request *aer;
761 ctrlr->async_event_config.raw = 0xFF;
762 ctrlr->async_event_config.bits.reserved = 0;
765 nvme_ctrlr_cmd_get_feature(ctrlr, NVME_FEAT_TEMPERATURE_THRESHOLD,
766 0, NULL, 0, nvme_completion_poll_cb, &status);
767 while (status.done == FALSE)
769 if (nvme_completion_is_error(&status.cpl) ||
770 (status.cpl.cdw0 & 0xFFFF) == 0xFFFF ||
771 (status.cpl.cdw0 & 0xFFFF) == 0x0000) {
772 nvme_printf(ctrlr, "temperature threshold not supported\n");
773 ctrlr->async_event_config.bits.temperature = 0;
776 nvme_ctrlr_cmd_set_async_event_config(ctrlr,
777 ctrlr->async_event_config, NULL, NULL);
779 /* aerl is a zero-based value, so we need to add 1 here. */
780 ctrlr->num_aers = min(NVME_MAX_ASYNC_EVENTS, (ctrlr->cdata.aerl+1));
782 /* Chatham doesn't support AERs. */
783 if (pci_get_devid(ctrlr->dev) == CHATHAM_PCI_ID)
786 for (i = 0; i < ctrlr->num_aers; i++) {
787 aer = &ctrlr->aer[i];
788 nvme_ctrlr_construct_and_submit_aer(ctrlr, aer);
793 nvme_ctrlr_configure_int_coalescing(struct nvme_controller *ctrlr)
796 ctrlr->int_coal_time = 0;
797 TUNABLE_INT_FETCH("hw.nvme.int_coal_time",
798 &ctrlr->int_coal_time);
800 ctrlr->int_coal_threshold = 0;
801 TUNABLE_INT_FETCH("hw.nvme.int_coal_threshold",
802 &ctrlr->int_coal_threshold);
804 nvme_ctrlr_cmd_set_interrupt_coalescing(ctrlr, ctrlr->int_coal_time,
805 ctrlr->int_coal_threshold, NULL, NULL);
809 nvme_ctrlr_start(void *ctrlr_arg)
811 struct nvme_controller *ctrlr = ctrlr_arg;
814 nvme_qpair_reset(&ctrlr->adminq);
815 for (i = 0; i < ctrlr->num_io_queues; i++)
816 nvme_qpair_reset(&ctrlr->ioq[i]);
818 nvme_admin_qpair_enable(&ctrlr->adminq);
820 if (nvme_ctrlr_identify(ctrlr) != 0) {
821 nvme_ctrlr_fail(ctrlr);
825 if (nvme_ctrlr_set_num_qpairs(ctrlr) != 0) {
826 nvme_ctrlr_fail(ctrlr);
830 if (nvme_ctrlr_create_qpairs(ctrlr) != 0) {
831 nvme_ctrlr_fail(ctrlr);
835 if (nvme_ctrlr_construct_namespaces(ctrlr) != 0) {
836 nvme_ctrlr_fail(ctrlr);
840 nvme_ctrlr_configure_aer(ctrlr);
841 nvme_ctrlr_configure_int_coalescing(ctrlr);
843 for (i = 0; i < ctrlr->num_io_queues; i++)
844 nvme_io_qpair_enable(&ctrlr->ioq[i]);
848 nvme_ctrlr_start_config_hook(void *arg)
850 struct nvme_controller *ctrlr = arg;
852 nvme_ctrlr_start(ctrlr);
853 config_intrhook_disestablish(&ctrlr->config_hook);
855 ctrlr->is_initialized = 1;
856 nvme_notify_new_controller(ctrlr);
860 nvme_ctrlr_reset_task(void *arg, int pending)
862 struct nvme_controller *ctrlr = arg;
865 nvme_printf(ctrlr, "resetting controller\n");
866 status = nvme_ctrlr_hw_reset(ctrlr);
868 * Use pause instead of DELAY, so that we yield to any nvme interrupt
869 * handlers on this CPU that were blocked on a qpair lock. We want
870 * all nvme interrupts completed before proceeding with restarting the
873 * XXX - any way to guarantee the interrupt handlers have quiesced?
875 pause("nvmereset", hz / 10);
877 nvme_ctrlr_start(ctrlr);
879 nvme_ctrlr_fail(ctrlr);
881 atomic_cmpset_32(&ctrlr->is_resetting, 1, 0);
885 nvme_ctrlr_intx_handler(void *arg)
887 struct nvme_controller *ctrlr = arg;
889 nvme_mmio_write_4(ctrlr, intms, 1);
891 nvme_qpair_process_completions(&ctrlr->adminq);
893 if (ctrlr->ioq[0].cpl)
894 nvme_qpair_process_completions(&ctrlr->ioq[0]);
896 nvme_mmio_write_4(ctrlr, intmc, 1);
900 nvme_ctrlr_configure_intx(struct nvme_controller *ctrlr)
903 ctrlr->num_io_queues = 1;
904 ctrlr->per_cpu_io_queues = 0;
906 ctrlr->res = bus_alloc_resource_any(ctrlr->dev, SYS_RES_IRQ,
907 &ctrlr->rid, RF_SHAREABLE | RF_ACTIVE);
909 if (ctrlr->res == NULL) {
910 nvme_printf(ctrlr, "unable to allocate shared IRQ\n");
914 bus_setup_intr(ctrlr->dev, ctrlr->res,
915 INTR_TYPE_MISC | INTR_MPSAFE, NULL, nvme_ctrlr_intx_handler,
918 if (ctrlr->tag == NULL) {
919 nvme_printf(ctrlr, "unable to setup intx handler\n");
927 nvme_pt_done(void *arg, const struct nvme_completion *cpl)
929 struct nvme_pt_command *pt = arg;
931 bzero(&pt->cpl, sizeof(pt->cpl));
932 pt->cpl.cdw0 = cpl->cdw0;
933 pt->cpl.status = cpl->status;
934 pt->cpl.status.p = 0;
936 mtx_lock(pt->driver_lock);
938 mtx_unlock(pt->driver_lock);
942 nvme_ctrlr_passthrough_cmd(struct nvme_controller *ctrlr,
943 struct nvme_pt_command *pt, uint32_t nsid, int is_user_buffer,
946 struct nvme_request *req;
948 struct buf *buf = NULL;
952 if (pt->len > ctrlr->max_xfer_size) {
953 nvme_printf(ctrlr, "pt->len (%d) "
954 "exceeds max_xfer_size (%d)\n", pt->len,
955 ctrlr->max_xfer_size);
958 if (is_user_buffer) {
960 * Ensure the user buffer is wired for the duration of
961 * this passthrough command.
965 buf->b_saveaddr = buf->b_data;
966 buf->b_data = pt->buf;
967 buf->b_bufsize = pt->len;
968 buf->b_iocmd = pt->is_read ? BIO_READ : BIO_WRITE;
969 #ifdef NVME_UNMAPPED_BIO_SUPPORT
970 if (vmapbuf(buf, 1) < 0) {
972 if (vmapbuf(buf) < 0) {
977 req = nvme_allocate_request_vaddr(buf->b_data, pt->len,
980 req = nvme_allocate_request_vaddr(pt->buf, pt->len,
983 req = nvme_allocate_request_null(nvme_pt_done, pt);
985 req->cmd.opc = pt->cmd.opc;
986 req->cmd.cdw10 = pt->cmd.cdw10;
987 req->cmd.cdw11 = pt->cmd.cdw11;
988 req->cmd.cdw12 = pt->cmd.cdw12;
989 req->cmd.cdw13 = pt->cmd.cdw13;
990 req->cmd.cdw14 = pt->cmd.cdw14;
991 req->cmd.cdw15 = pt->cmd.cdw15;
993 req->cmd.nsid = nsid;
998 mtx = &ctrlr->ns[nsid-1].lock;
1001 pt->driver_lock = mtx;
1004 nvme_ctrlr_submit_admin_request(ctrlr, req);
1006 nvme_ctrlr_submit_io_request(ctrlr, req);
1008 mtx_sleep(pt, mtx, PRIBIO, "nvme_pt", 0);
1011 pt->driver_lock = NULL;
1023 nvme_ctrlr_ioctl(struct cdev *cdev, u_long cmd, caddr_t arg, int flag,
1026 struct nvme_controller *ctrlr;
1027 struct nvme_pt_command *pt;
1029 ctrlr = cdev->si_drv1;
1032 case NVME_RESET_CONTROLLER:
1033 nvme_ctrlr_reset(ctrlr);
1035 case NVME_PASSTHROUGH_CMD:
1036 pt = (struct nvme_pt_command *)arg;
1039 * Chatham IDENTIFY data is spoofed, so copy the spoofed data
1040 * rather than issuing the command to the Chatham controller.
1042 if (pci_get_devid(ctrlr->dev) == CHATHAM_PCI_ID &&
1043 pt->cmd.opc == NVME_OPC_IDENTIFY) {
1044 if (pt->cmd.cdw10 == 1) {
1045 if (pt->len != sizeof(ctrlr->cdata))
1047 return (copyout(&ctrlr->cdata, pt->buf,
1050 if (pt->len != sizeof(ctrlr->ns[0].data) ||
1053 return (copyout(&ctrlr->ns[0].data, pt->buf,
1058 return (nvme_ctrlr_passthrough_cmd(ctrlr, pt, pt->cmd.nsid,
1059 1 /* is_user_buffer */, 1 /* is_admin_cmd */));
1067 static struct cdevsw nvme_ctrlr_cdevsw = {
1068 .d_version = D_VERSION,
1070 .d_ioctl = nvme_ctrlr_ioctl
1074 nvme_ctrlr_construct(struct nvme_controller *ctrlr, device_t dev)
1076 union cap_lo_register cap_lo;
1077 union cap_hi_register cap_hi;
1078 int i, num_vectors, per_cpu_io_queues, rid;
1079 int status, timeout_period;
1083 mtx_init(&ctrlr->lock, "nvme ctrlr lock", NULL, MTX_DEF);
1085 status = nvme_ctrlr_allocate_bar(ctrlr);
1091 if (pci_get_devid(dev) == CHATHAM_PCI_ID) {
1092 status = nvme_ctrlr_allocate_chatham_bar(ctrlr);
1095 nvme_ctrlr_setup_chatham(ctrlr);
1100 * Software emulators may set the doorbell stride to something
1101 * other than zero, but this driver is not set up to handle that.
1103 cap_hi.raw = nvme_mmio_read_4(ctrlr, cap_hi);
1104 if (cap_hi.bits.dstrd != 0)
1107 ctrlr->min_page_size = 1 << (12 + cap_hi.bits.mpsmin);
1109 /* Get ready timeout value from controller, in units of 500ms. */
1110 cap_lo.raw = nvme_mmio_read_4(ctrlr, cap_lo);
1111 ctrlr->ready_timeout_in_ms = cap_lo.bits.to * 500;
1113 timeout_period = NVME_DEFAULT_TIMEOUT_PERIOD;
1114 TUNABLE_INT_FETCH("hw.nvme.timeout_period", &timeout_period);
1115 timeout_period = min(timeout_period, NVME_MAX_TIMEOUT_PERIOD);
1116 timeout_period = max(timeout_period, NVME_MIN_TIMEOUT_PERIOD);
1117 ctrlr->timeout_period = timeout_period;
1119 nvme_retry_count = NVME_DEFAULT_RETRY_COUNT;
1120 TUNABLE_INT_FETCH("hw.nvme.retry_count", &nvme_retry_count);
1122 per_cpu_io_queues = 1;
1123 TUNABLE_INT_FETCH("hw.nvme.per_cpu_io_queues", &per_cpu_io_queues);
1124 ctrlr->per_cpu_io_queues = per_cpu_io_queues ? TRUE : FALSE;
1126 if (ctrlr->per_cpu_io_queues)
1127 ctrlr->num_io_queues = mp_ncpus;
1129 ctrlr->num_io_queues = 1;
1131 ctrlr->force_intx = 0;
1132 TUNABLE_INT_FETCH("hw.nvme.force_intx", &ctrlr->force_intx);
1134 ctrlr->enable_aborts = 0;
1135 TUNABLE_INT_FETCH("hw.nvme.enable_aborts", &ctrlr->enable_aborts);
1137 ctrlr->msix_enabled = 1;
1139 if (ctrlr->force_intx) {
1140 ctrlr->msix_enabled = 0;
1144 /* One vector per IO queue, plus one vector for admin queue. */
1145 num_vectors = ctrlr->num_io_queues + 1;
1147 if (pci_msix_count(dev) < num_vectors) {
1148 ctrlr->msix_enabled = 0;
1152 if (pci_alloc_msix(dev, &num_vectors) != 0) {
1153 ctrlr->msix_enabled = 0;
1158 * On earlier FreeBSD releases, there are reports that
1159 * pci_alloc_msix() can return successfully with all vectors
1160 * requested, but a subsequent bus_alloc_resource_any()
1161 * for one of those vectors fails. This issue occurs more
1162 * readily with multiple devices using per-CPU vectors.
1163 * To workaround this issue, try to allocate the resources now,
1164 * and fall back to INTx if we cannot allocate all of them.
1165 * This issue cannot be reproduced on more recent versions of
1166 * FreeBSD which have increased the maximum number of MSI-X
1167 * vectors, but adding the workaround makes it easier for
1168 * vendors wishing to import this driver into kernels based on
1169 * older versions of FreeBSD.
1171 for (i = 0; i < num_vectors; i++) {
1173 ctrlr->msi_res[i] = bus_alloc_resource_any(ctrlr->dev,
1174 SYS_RES_IRQ, &rid, RF_ACTIVE);
1176 if (ctrlr->msi_res[i] == NULL) {
1177 ctrlr->msix_enabled = 0;
1180 bus_release_resource(ctrlr->dev,
1182 rman_get_rid(ctrlr->msi_res[i]),
1185 pci_release_msi(dev);
1186 nvme_printf(ctrlr, "could not obtain all MSI-X "
1187 "resources, reverting to intx\n");
1194 if (!ctrlr->msix_enabled)
1195 nvme_ctrlr_configure_intx(ctrlr);
1197 ctrlr->max_xfer_size = NVME_MAX_XFER_SIZE;
1198 nvme_ctrlr_construct_admin_qpair(ctrlr);
1199 status = nvme_ctrlr_construct_io_qpairs(ctrlr);
1204 ctrlr->cdev = make_dev(&nvme_ctrlr_cdevsw, device_get_unit(dev),
1205 UID_ROOT, GID_WHEEL, 0600, "nvme%d", device_get_unit(dev));
1207 if (ctrlr->cdev == NULL)
1210 ctrlr->cdev->si_drv1 = (void *)ctrlr;
1212 ctrlr->taskqueue = taskqueue_create("nvme_taskq", M_WAITOK,
1213 taskqueue_thread_enqueue, &ctrlr->taskqueue);
1214 taskqueue_start_threads(&ctrlr->taskqueue, 1, PI_DISK, "nvme taskq");
1216 ctrlr->is_resetting = 0;
1217 ctrlr->is_initialized = 0;
1218 ctrlr->notification_sent = 0;
1219 TASK_INIT(&ctrlr->reset_task, 0, nvme_ctrlr_reset_task, ctrlr);
1221 TASK_INIT(&ctrlr->fail_req_task, 0, nvme_ctrlr_fail_req_task, ctrlr);
1222 STAILQ_INIT(&ctrlr->fail_req);
1223 ctrlr->is_failed = FALSE;
1229 nvme_ctrlr_destruct(struct nvme_controller *ctrlr, device_t dev)
1234 * Notify the controller of a shutdown, even though this is due to
1235 * a driver unload, not a system shutdown (this path is not invoked
1236 * during shutdown). This ensures the controller receives a
1237 * shutdown notification in case the system is shutdown before
1238 * reloading the driver.
1240 * Chatham does not let you re-enable the controller after shutdown
1241 * notification has been received, so do not send it in this case.
1242 * This is OK because Chatham does not depend on the shutdown
1243 * notification anyways.
1245 if (pci_get_devid(ctrlr->dev) != CHATHAM_PCI_ID)
1246 nvme_ctrlr_shutdown(ctrlr);
1248 nvme_ctrlr_disable(ctrlr);
1249 taskqueue_free(ctrlr->taskqueue);
1251 for (i = 0; i < NVME_MAX_NAMESPACES; i++)
1252 nvme_ns_destruct(&ctrlr->ns[i]);
1255 destroy_dev(ctrlr->cdev);
1257 for (i = 0; i < ctrlr->num_io_queues; i++) {
1258 nvme_io_qpair_destroy(&ctrlr->ioq[i]);
1261 free(ctrlr->ioq, M_NVME);
1263 nvme_admin_qpair_destroy(&ctrlr->adminq);
1265 if (ctrlr->resource != NULL) {
1266 bus_release_resource(dev, SYS_RES_MEMORY,
1267 ctrlr->resource_id, ctrlr->resource);
1270 if (ctrlr->bar4_resource != NULL) {
1271 bus_release_resource(dev, SYS_RES_MEMORY,
1272 ctrlr->bar4_resource_id, ctrlr->bar4_resource);
1276 if (ctrlr->chatham_resource != NULL) {
1277 bus_release_resource(dev, SYS_RES_MEMORY,
1278 ctrlr->chatham_resource_id, ctrlr->chatham_resource);
1283 bus_teardown_intr(ctrlr->dev, ctrlr->res, ctrlr->tag);
1286 bus_release_resource(ctrlr->dev, SYS_RES_IRQ,
1287 rman_get_rid(ctrlr->res), ctrlr->res);
1289 if (ctrlr->msix_enabled)
1290 pci_release_msi(dev);
1294 nvme_ctrlr_shutdown(struct nvme_controller *ctrlr)
1296 union cc_register cc;
1297 union csts_register csts;
1300 cc.raw = nvme_mmio_read_4(ctrlr, cc);
1301 cc.bits.shn = NVME_SHN_NORMAL;
1302 nvme_mmio_write_4(ctrlr, cc, cc.raw);
1303 csts.raw = nvme_mmio_read_4(ctrlr, csts);
1304 while ((csts.bits.shst != NVME_SHST_COMPLETE) && (ticks++ < 5*hz)) {
1305 pause("nvme shn", 1);
1306 csts.raw = nvme_mmio_read_4(ctrlr, csts);
1308 if (csts.bits.shst != NVME_SHST_COMPLETE)
1309 nvme_printf(ctrlr, "did not complete shutdown within 5 seconds "
1310 "of notification\n");
1314 nvme_ctrlr_submit_admin_request(struct nvme_controller *ctrlr,
1315 struct nvme_request *req)
1318 nvme_qpair_submit_request(&ctrlr->adminq, req);
1322 nvme_ctrlr_submit_io_request(struct nvme_controller *ctrlr,
1323 struct nvme_request *req)
1325 struct nvme_qpair *qpair;
1327 if (ctrlr->per_cpu_io_queues)
1328 qpair = &ctrlr->ioq[curcpu];
1330 qpair = &ctrlr->ioq[0];
1332 nvme_qpair_submit_request(qpair, req);
1336 nvme_ctrlr_get_device(struct nvme_controller *ctrlr)
1339 return (ctrlr->dev);
1342 const struct nvme_controller_data *
1343 nvme_ctrlr_get_data(struct nvme_controller *ctrlr)
1346 return (&ctrlr->cdata);