2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2017 Shunsuke Mie
5 * Copyright (c) 2018 Leon Dang
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 * bhyve PCIe-NVMe device emulation.
33 * -s <n>,nvme,devpath,maxq=#,qsz=#,ioslots=#,sectsz=#,ser=A-Z
40 * maxq = max number of queues
41 * qsz = max elements in each queue
42 * ioslots = max number of concurrent io requests
43 * sectsz = sector size (defaults to blockif sector size)
44 * ser = serial number (20-chars max)
49 - create async event for smart and log
53 #include <sys/cdefs.h>
54 __FBSDID("$FreeBSD$");
56 #include <sys/types.h>
60 #include <semaphore.h>
68 #include <machine/atomic.h>
69 #include <machine/vmm.h>
72 #include <dev/nvme/nvme.h>
79 static int nvme_debug = 0;
80 #define DPRINTF(params) if (nvme_debug) printf params
81 #define WPRINTF(params) printf params
83 /* defaults; can be overridden */
84 #define NVME_MSIX_BAR 4
86 #define NVME_IOSLOTS 8
88 /* The NVMe spec defines bits 13:4 in BAR0 as reserved */
89 #define NVME_MMIO_SPACE_MIN (1 << 14)
91 #define NVME_QUEUES 16
92 #define NVME_MAX_QENTRIES 2048
94 #define NVME_PRP2_ITEMS (PAGE_SIZE/sizeof(uint64_t))
95 #define NVME_MAX_BLOCKIOVS 512
99 /* Convert a zero-based value into a one-based value */
100 #define ONE_BASED(zero) ((zero) + 1)
101 /* Convert a one-based value into a zero-based value */
102 #define ZERO_BASED(one) ((one) - 1)
104 /* Encode number of SQ's and CQ's for Set/Get Features */
105 #define NVME_FEATURE_NUM_QUEUES(sc) \
106 (ZERO_BASED((sc)->num_squeues) & 0xffff) | \
107 (ZERO_BASED((sc)->num_cqueues) & 0xffff) << 16;
109 #define NVME_DOORBELL_OFFSET offsetof(struct nvme_registers, doorbell)
111 enum nvme_controller_register_offsets {
112 NVME_CR_CAP_LOW = 0x00,
113 NVME_CR_CAP_HI = 0x04,
115 NVME_CR_INTMS = 0x0c,
116 NVME_CR_INTMC = 0x10,
121 NVME_CR_ASQ_LOW = 0x28,
122 NVME_CR_ASQ_HI = 0x2c,
123 NVME_CR_ACQ_LOW = 0x30,
124 NVME_CR_ACQ_HI = 0x34,
127 enum nvme_cmd_cdw11 {
128 NVME_CMD_CDW11_PC = 0x0001,
129 NVME_CMD_CDW11_IEN = 0x0002,
130 NVME_CMD_CDW11_IV = 0xFFFF0000,
133 #define NVME_CQ_INTEN 0x01
134 #define NVME_CQ_INTCOAL 0x02
136 struct nvme_completion_queue {
137 struct nvme_completion *qbase;
139 uint16_t tail; /* nvme progress */
140 uint16_t head; /* guest progress */
146 struct nvme_submission_queue {
147 struct nvme_command *qbase;
149 uint16_t head; /* nvme progress */
150 uint16_t tail; /* guest progress */
151 uint16_t cqid; /* completion queue id */
152 int busy; /* queue is being processed */
156 enum nvme_storage_type {
157 NVME_STOR_BLOCKIF = 0,
161 struct pci_nvme_blockstore {
162 enum nvme_storage_type type;
166 uint32_t sectsz_bits;
169 struct pci_nvme_ioreq {
170 struct pci_nvme_softc *sc;
171 struct pci_nvme_ioreq *next;
172 struct nvme_submission_queue *nvme_sq;
175 /* command information */
180 uint64_t prev_gpaddr;
184 * lock if all iovs consumed (big IO);
185 * complete transaction before continuing
190 struct blockif_req io_req;
192 /* pad to fit up to 512 page descriptors from guest IO request */
193 struct iovec iovpadding[NVME_MAX_BLOCKIOVS-BLOCKIF_IOV_MAX];
196 struct pci_nvme_softc {
197 struct pci_devinst *nsc_pi;
201 struct nvme_registers regs;
203 struct nvme_namespace_data nsdata;
204 struct nvme_controller_data ctrldata;
206 struct pci_nvme_blockstore nvstore;
208 uint16_t max_qentries; /* max entries per queue */
209 uint32_t max_queues; /* max number of IO SQ's or CQ's */
210 uint32_t num_cqueues;
211 uint32_t num_squeues;
213 struct pci_nvme_ioreq *ioreqs;
214 struct pci_nvme_ioreq *ioreqs_free; /* free list of ioreqs */
215 uint32_t pending_ios;
220 * Memory mapped Submission and Completion queues
221 * Each array includes both Admin and IO queues
223 struct nvme_completion_queue *compl_queues;
224 struct nvme_submission_queue *submit_queues;
226 /* controller features */
227 uint32_t intr_coales_aggr_time; /* 0x08: uS to delay intr */
228 uint32_t intr_coales_aggr_thresh; /* 0x08: compl-Q entries */
229 uint32_t async_ev_config; /* 0x0B: async event config */
233 static void pci_nvme_io_partial(struct blockif_req *br, int err);
235 /* Controller Configuration utils */
236 #define NVME_CC_GET_EN(cc) \
237 ((cc) >> NVME_CC_REG_EN_SHIFT & NVME_CC_REG_EN_MASK)
238 #define NVME_CC_GET_CSS(cc) \
239 ((cc) >> NVME_CC_REG_CSS_SHIFT & NVME_CC_REG_CSS_MASK)
240 #define NVME_CC_GET_SHN(cc) \
241 ((cc) >> NVME_CC_REG_SHN_SHIFT & NVME_CC_REG_SHN_MASK)
242 #define NVME_CC_GET_IOSQES(cc) \
243 ((cc) >> NVME_CC_REG_IOSQES_SHIFT & NVME_CC_REG_IOSQES_MASK)
244 #define NVME_CC_GET_IOCQES(cc) \
245 ((cc) >> NVME_CC_REG_IOCQES_SHIFT & NVME_CC_REG_IOCQES_MASK)
247 #define NVME_CC_WRITE_MASK \
248 ((NVME_CC_REG_EN_MASK << NVME_CC_REG_EN_SHIFT) | \
249 (NVME_CC_REG_IOSQES_MASK << NVME_CC_REG_IOSQES_SHIFT) | \
250 (NVME_CC_REG_IOCQES_MASK << NVME_CC_REG_IOCQES_SHIFT))
252 #define NVME_CC_NEN_WRITE_MASK \
253 ((NVME_CC_REG_CSS_MASK << NVME_CC_REG_CSS_SHIFT) | \
254 (NVME_CC_REG_MPS_MASK << NVME_CC_REG_MPS_SHIFT) | \
255 (NVME_CC_REG_AMS_MASK << NVME_CC_REG_AMS_SHIFT))
257 /* Controller Status utils */
258 #define NVME_CSTS_GET_RDY(sts) \
259 ((sts) >> NVME_CSTS_REG_RDY_SHIFT & NVME_CSTS_REG_RDY_MASK)
261 #define NVME_CSTS_RDY (1 << NVME_CSTS_REG_RDY_SHIFT)
263 /* Completion Queue status word utils */
264 #define NVME_STATUS_P (1 << NVME_STATUS_P_SHIFT)
265 #define NVME_STATUS_MASK \
266 ((NVME_STATUS_SCT_MASK << NVME_STATUS_SCT_SHIFT) |\
267 (NVME_STATUS_SC_MASK << NVME_STATUS_SC_SHIFT))
270 cpywithpad(char *dst, size_t dst_size, const char *src, char pad)
274 len = strnlen(src, dst_size);
275 memset(dst, pad, dst_size);
276 memcpy(dst, src, len);
280 pci_nvme_status_tc(uint16_t *status, uint16_t type, uint16_t code)
283 *status &= ~NVME_STATUS_MASK;
284 *status |= (type & NVME_STATUS_SCT_MASK) << NVME_STATUS_SCT_SHIFT |
285 (code & NVME_STATUS_SC_MASK) << NVME_STATUS_SC_SHIFT;
289 pci_nvme_status_genc(uint16_t *status, uint16_t code)
292 pci_nvme_status_tc(status, NVME_SCT_GENERIC, code);
296 pci_nvme_toggle_phase(uint16_t *status, int prev)
300 *status &= ~NVME_STATUS_P;
302 *status |= NVME_STATUS_P;
306 pci_nvme_init_ctrldata(struct pci_nvme_softc *sc)
308 struct nvme_controller_data *cd = &sc->ctrldata;
313 cpywithpad((char *)cd->mn, sizeof(cd->mn), "bhyve-NVMe", ' ');
314 cpywithpad((char *)cd->fr, sizeof(cd->fr), "1.0", ' ');
316 /* Num of submission commands that we can handle at a time (2^rab) */
326 cd->mdts = 9; /* max data transfer size (2^mdts * CAP.MPSMIN) */
328 cd->ver = 0x00010300;
330 cd->oacs = 1 << NVME_CTRLR_DATA_OACS_FORMAT_SHIFT;
334 cd->lpa = 0; /* TODO: support some simple things like SMART */
335 cd->elpe = 0; /* max error log page entries */
336 cd->npss = 1; /* number of power states support */
338 /* Warning Composite Temperature Threshold */
341 cd->sqes = (6 << NVME_CTRLR_DATA_SQES_MAX_SHIFT) |
342 (6 << NVME_CTRLR_DATA_SQES_MIN_SHIFT);
343 cd->cqes = (4 << NVME_CTRLR_DATA_CQES_MAX_SHIFT) |
344 (4 << NVME_CTRLR_DATA_CQES_MIN_SHIFT);
345 cd->nn = 1; /* number of namespaces */
349 cd->power_state[0].mp = 10;
353 pci_nvme_init_nsdata(struct pci_nvme_softc *sc)
355 struct nvme_namespace_data *nd;
359 nd->nsze = sc->nvstore.size / sc->nvstore.sectsz;
363 /* Get LBA and backstore information from backing store */
364 nd->nlbaf = 0; /* NLBAF is a 0's based value (i.e. 1 LBA Format) */
365 /* LBA data-sz = 2^lbads */
366 nd->lbaf[0] = sc->nvstore.sectsz_bits << NVME_NS_DATA_LBAF_LBADS_SHIFT;
372 pci_nvme_reset_locked(struct pci_nvme_softc *sc)
374 DPRINTF(("%s\r\n", __func__));
376 sc->regs.cap_lo = (ZERO_BASED(sc->max_qentries) & NVME_CAP_LO_REG_MQES_MASK) |
377 (1 << NVME_CAP_LO_REG_CQR_SHIFT) |
378 (60 << NVME_CAP_LO_REG_TO_SHIFT);
380 sc->regs.cap_hi = 1 << NVME_CAP_HI_REG_CSS_NVM_SHIFT;
382 sc->regs.vs = 0x00010300; /* NVMe v1.3 */
387 sc->num_cqueues = sc->num_squeues = sc->max_queues;
388 if (sc->submit_queues != NULL) {
389 for (int i = 0; i < sc->num_squeues + 1; i++) {
391 * The Admin Submission Queue is at index 0.
392 * It must not be changed at reset otherwise the
393 * emulation will be out of sync with the guest.
396 sc->submit_queues[i].qbase = NULL;
397 sc->submit_queues[i].size = 0;
398 sc->submit_queues[i].cqid = 0;
400 sc->submit_queues[i].tail = 0;
401 sc->submit_queues[i].head = 0;
402 sc->submit_queues[i].busy = 0;
405 sc->submit_queues = calloc(sc->num_squeues + 1,
406 sizeof(struct nvme_submission_queue));
408 if (sc->compl_queues != NULL) {
409 for (int i = 0; i < sc->num_cqueues + 1; i++) {
410 /* See Admin Submission Queue note above */
412 sc->compl_queues[i].qbase = NULL;
413 sc->compl_queues[i].size = 0;
416 sc->compl_queues[i].tail = 0;
417 sc->compl_queues[i].head = 0;
420 sc->compl_queues = calloc(sc->num_cqueues + 1,
421 sizeof(struct nvme_completion_queue));
423 for (int i = 0; i < sc->num_cqueues + 1; i++)
424 pthread_mutex_init(&sc->compl_queues[i].mtx, NULL);
429 pci_nvme_reset(struct pci_nvme_softc *sc)
431 pthread_mutex_lock(&sc->mtx);
432 pci_nvme_reset_locked(sc);
433 pthread_mutex_unlock(&sc->mtx);
437 pci_nvme_init_controller(struct vmctx *ctx, struct pci_nvme_softc *sc)
441 DPRINTF(("%s\r\n", __func__));
443 asqs = (sc->regs.aqa & NVME_AQA_REG_ASQS_MASK) + 1;
444 sc->submit_queues[0].size = asqs;
445 sc->submit_queues[0].qbase = vm_map_gpa(ctx, sc->regs.asq,
446 sizeof(struct nvme_command) * asqs);
448 DPRINTF(("%s mapping Admin-SQ guest 0x%lx, host: %p\r\n",
449 __func__, sc->regs.asq, sc->submit_queues[0].qbase));
451 acqs = ((sc->regs.aqa >> NVME_AQA_REG_ACQS_SHIFT) &
452 NVME_AQA_REG_ACQS_MASK) + 1;
453 sc->compl_queues[0].size = acqs;
454 sc->compl_queues[0].qbase = vm_map_gpa(ctx, sc->regs.acq,
455 sizeof(struct nvme_completion) * acqs);
456 DPRINTF(("%s mapping Admin-CQ guest 0x%lx, host: %p\r\n",
457 __func__, sc->regs.acq, sc->compl_queues[0].qbase));
461 nvme_opc_delete_io_sq(struct pci_nvme_softc* sc, struct nvme_command* command,
462 struct nvme_completion* compl)
464 uint16_t qid = command->cdw10 & 0xffff;
466 DPRINTF(("%s DELETE_IO_SQ %u\r\n", __func__, qid));
467 if (qid == 0 || qid > sc->num_squeues) {
468 WPRINTF(("%s NOT PERMITTED queue id %u / num_squeues %u\r\n",
469 __func__, qid, sc->num_squeues));
470 pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
471 NVME_SC_INVALID_QUEUE_IDENTIFIER);
475 sc->submit_queues[qid].qbase = NULL;
476 pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
481 nvme_opc_create_io_sq(struct pci_nvme_softc* sc, struct nvme_command* command,
482 struct nvme_completion* compl)
484 if (command->cdw11 & NVME_CMD_CDW11_PC) {
485 uint16_t qid = command->cdw10 & 0xffff;
486 struct nvme_submission_queue *nsq;
488 if ((qid == 0) || (qid > sc->num_squeues)) {
489 WPRINTF(("%s queue index %u > num_squeues %u\r\n",
490 __func__, qid, sc->num_squeues));
491 pci_nvme_status_tc(&compl->status,
492 NVME_SCT_COMMAND_SPECIFIC,
493 NVME_SC_INVALID_QUEUE_IDENTIFIER);
497 nsq = &sc->submit_queues[qid];
498 nsq->size = ONE_BASED((command->cdw10 >> 16) & 0xffff);
500 nsq->qbase = vm_map_gpa(sc->nsc_pi->pi_vmctx, command->prp1,
501 sizeof(struct nvme_command) * (size_t)nsq->size);
502 nsq->cqid = (command->cdw11 >> 16) & 0xffff;
503 nsq->qpriority = (command->cdw11 >> 1) & 0x03;
505 DPRINTF(("%s sq %u size %u gaddr %p cqid %u\r\n", __func__,
506 qid, nsq->size, nsq->qbase, nsq->cqid));
508 pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
510 DPRINTF(("%s completed creating IOSQ qid %u\r\n",
514 * Guest sent non-cont submission queue request.
515 * This setting is unsupported by this emulation.
517 WPRINTF(("%s unsupported non-contig (list-based) "
518 "create i/o submission queue\r\n", __func__));
520 pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
526 nvme_opc_delete_io_cq(struct pci_nvme_softc* sc, struct nvme_command* command,
527 struct nvme_completion* compl)
529 uint16_t qid = command->cdw10 & 0xffff;
531 DPRINTF(("%s DELETE_IO_CQ %u\r\n", __func__, qid));
532 if (qid == 0 || qid > sc->num_cqueues) {
533 WPRINTF(("%s queue index %u / num_cqueues %u\r\n",
534 __func__, qid, sc->num_cqueues));
535 pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
536 NVME_SC_INVALID_QUEUE_IDENTIFIER);
540 sc->compl_queues[qid].qbase = NULL;
541 pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
546 nvme_opc_create_io_cq(struct pci_nvme_softc* sc, struct nvme_command* command,
547 struct nvme_completion* compl)
549 if (command->cdw11 & NVME_CMD_CDW11_PC) {
550 uint16_t qid = command->cdw10 & 0xffff;
551 struct nvme_completion_queue *ncq;
553 if ((qid == 0) || (qid > sc->num_cqueues)) {
554 WPRINTF(("%s queue index %u > num_cqueues %u\r\n",
555 __func__, qid, sc->num_cqueues));
556 pci_nvme_status_tc(&compl->status,
557 NVME_SCT_COMMAND_SPECIFIC,
558 NVME_SC_INVALID_QUEUE_IDENTIFIER);
562 ncq = &sc->compl_queues[qid];
563 ncq->intr_en = (command->cdw11 & NVME_CMD_CDW11_IEN) >> 1;
564 ncq->intr_vec = (command->cdw11 >> 16) & 0xffff;
565 ncq->size = ONE_BASED((command->cdw10 >> 16) & 0xffff);
567 ncq->qbase = vm_map_gpa(sc->nsc_pi->pi_vmctx,
569 sizeof(struct nvme_command) * (size_t)ncq->size);
571 pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
574 * Non-contig completion queue unsupported.
576 WPRINTF(("%s unsupported non-contig (list-based) "
577 "create i/o completion queue\r\n",
580 /* 0x12 = Invalid Use of Controller Memory Buffer */
581 pci_nvme_status_genc(&compl->status, 0x12);
588 nvme_opc_get_log_page(struct pci_nvme_softc* sc, struct nvme_command* command,
589 struct nvme_completion* compl)
591 uint32_t logsize = (1 + ((command->cdw10 >> 16) & 0xFFF)) * 2;
592 uint8_t logpage = command->cdw10 & 0xFF;
595 DPRINTF(("%s log page %u len %u\r\n", __func__, logpage, logsize));
597 if (logpage >= 1 && logpage <= 3)
598 data = vm_map_gpa(sc->nsc_pi->pi_vmctx, command->prp1,
601 pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
604 case 0x01: /* Error information */
605 memset(data, 0, logsize > PAGE_SIZE ? PAGE_SIZE : logsize);
607 case 0x02: /* SMART/Health information */
608 /* TODO: present some smart info */
609 memset(data, 0, logsize > PAGE_SIZE ? PAGE_SIZE : logsize);
611 case 0x03: /* Firmware slot information */
612 memset(data, 0, logsize > PAGE_SIZE ? PAGE_SIZE : logsize);
615 WPRINTF(("%s get log page %x command not supported\r\n",
618 pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
619 NVME_SC_INVALID_LOG_PAGE);
626 nvme_opc_identify(struct pci_nvme_softc* sc, struct nvme_command* command,
627 struct nvme_completion* compl)
631 DPRINTF(("%s identify 0x%x nsid 0x%x\r\n", __func__,
632 command->cdw10 & 0xFF, command->nsid));
634 switch (command->cdw10 & 0xFF) {
635 case 0x00: /* return Identify Namespace data structure */
636 dest = vm_map_gpa(sc->nsc_pi->pi_vmctx, command->prp1,
638 memcpy(dest, &sc->nsdata, sizeof(sc->nsdata));
640 case 0x01: /* return Identify Controller data structure */
641 dest = vm_map_gpa(sc->nsc_pi->pi_vmctx, command->prp1,
642 sizeof(sc->ctrldata));
643 memcpy(dest, &sc->ctrldata, sizeof(sc->ctrldata));
645 case 0x02: /* list of 1024 active NSIDs > CDW1.NSID */
646 dest = vm_map_gpa(sc->nsc_pi->pi_vmctx, command->prp1,
647 sizeof(uint32_t) * 1024);
648 ((uint32_t *)dest)[0] = 1;
649 ((uint32_t *)dest)[1] = 0;
652 pci_nvme_status_genc(&compl->status,
653 NVME_SC_INVALID_NAMESPACE_OR_FORMAT);
655 case 0x03: /* list of NSID structures in CDW1.NSID, 4096 bytes */
662 DPRINTF(("%s unsupported identify command requested 0x%x\r\n",
663 __func__, command->cdw10 & 0xFF));
664 pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
668 pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
673 nvme_set_feature_queues(struct pci_nvme_softc* sc, struct nvme_command* command,
674 struct nvme_completion* compl)
676 uint16_t nqr; /* Number of Queues Requested */
678 nqr = command->cdw11 & 0xFFFF;
680 WPRINTF(("%s: Illegal NSQR value %#x\n", __func__, nqr));
681 pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
685 sc->num_squeues = ONE_BASED(nqr);
686 if (sc->num_squeues > sc->max_queues) {
687 DPRINTF(("NSQR=%u is greater than max %u\n", sc->num_squeues,
689 sc->num_squeues = sc->max_queues;
692 nqr = (command->cdw11 >> 16) & 0xFFFF;
694 WPRINTF(("%s: Illegal NCQR value %#x\n", __func__, nqr));
695 pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
699 sc->num_cqueues = ONE_BASED(nqr);
700 if (sc->num_cqueues > sc->max_queues) {
701 DPRINTF(("NCQR=%u is greater than max %u\n", sc->num_cqueues,
703 sc->num_cqueues = sc->max_queues;
706 compl->cdw0 = NVME_FEATURE_NUM_QUEUES(sc);
712 nvme_opc_set_features(struct pci_nvme_softc* sc, struct nvme_command* command,
713 struct nvme_completion* compl)
715 int feature = command->cdw10 & 0xFF;
718 DPRINTF(("%s feature 0x%x\r\n", __func__, feature));
722 case NVME_FEAT_ARBITRATION:
723 DPRINTF((" arbitration 0x%x\r\n", command->cdw11));
725 case NVME_FEAT_POWER_MANAGEMENT:
726 DPRINTF((" power management 0x%x\r\n", command->cdw11));
728 case NVME_FEAT_LBA_RANGE_TYPE:
729 DPRINTF((" lba range 0x%x\r\n", command->cdw11));
731 case NVME_FEAT_TEMPERATURE_THRESHOLD:
732 DPRINTF((" temperature threshold 0x%x\r\n", command->cdw11));
734 case NVME_FEAT_ERROR_RECOVERY:
735 DPRINTF((" error recovery 0x%x\r\n", command->cdw11));
737 case NVME_FEAT_VOLATILE_WRITE_CACHE:
738 DPRINTF((" volatile write cache 0x%x\r\n", command->cdw11));
740 case NVME_FEAT_NUMBER_OF_QUEUES:
741 nvme_set_feature_queues(sc, command, compl);
743 case NVME_FEAT_INTERRUPT_COALESCING:
744 DPRINTF((" interrupt coalescing 0x%x\r\n", command->cdw11));
747 sc->intr_coales_aggr_time = ((command->cdw11 >> 8) & 0xFF)*100;
749 sc->intr_coales_aggr_thresh = command->cdw11 & 0xFF;
751 case NVME_FEAT_INTERRUPT_VECTOR_CONFIGURATION:
752 iv = command->cdw11 & 0xFFFF;
754 DPRINTF((" interrupt vector configuration 0x%x\r\n",
757 for (uint32_t i = 0; i < sc->num_cqueues + 1; i++) {
758 if (sc->compl_queues[i].intr_vec == iv) {
759 if (command->cdw11 & (1 << 16))
760 sc->compl_queues[i].intr_en |=
763 sc->compl_queues[i].intr_en &=
768 case NVME_FEAT_WRITE_ATOMICITY:
769 DPRINTF((" write atomicity 0x%x\r\n", command->cdw11));
771 case NVME_FEAT_ASYNC_EVENT_CONFIGURATION:
772 DPRINTF((" async event configuration 0x%x\r\n",
774 sc->async_ev_config = command->cdw11;
776 case NVME_FEAT_SOFTWARE_PROGRESS_MARKER:
777 DPRINTF((" software progress marker 0x%x\r\n",
781 DPRINTF((" autonomous power state transition 0x%x\r\n",
785 WPRINTF(("%s invalid feature\r\n", __func__));
786 pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
790 pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
795 nvme_opc_get_features(struct pci_nvme_softc* sc, struct nvme_command* command,
796 struct nvme_completion* compl)
798 int feature = command->cdw10 & 0xFF;
800 DPRINTF(("%s feature 0x%x\r\n", __func__, feature));
805 case NVME_FEAT_ARBITRATION:
806 DPRINTF((" arbitration\r\n"));
808 case NVME_FEAT_POWER_MANAGEMENT:
809 DPRINTF((" power management\r\n"));
811 case NVME_FEAT_LBA_RANGE_TYPE:
812 DPRINTF((" lba range\r\n"));
814 case NVME_FEAT_TEMPERATURE_THRESHOLD:
815 DPRINTF((" temperature threshold\r\n"));
816 switch ((command->cdw11 >> 20) & 0x3) {
818 /* Over temp threshold */
819 compl->cdw0 = 0xFFFF;
822 /* Under temp threshold */
826 WPRINTF((" invalid threshold type select\r\n"));
827 pci_nvme_status_genc(&compl->status,
828 NVME_SC_INVALID_FIELD);
832 case NVME_FEAT_ERROR_RECOVERY:
833 DPRINTF((" error recovery\r\n"));
835 case NVME_FEAT_VOLATILE_WRITE_CACHE:
836 DPRINTF((" volatile write cache\r\n"));
838 case NVME_FEAT_NUMBER_OF_QUEUES:
839 compl->cdw0 = NVME_FEATURE_NUM_QUEUES(sc);
841 DPRINTF((" number of queues (submit %u, completion %u)\r\n",
842 compl->cdw0 & 0xFFFF,
843 (compl->cdw0 >> 16) & 0xFFFF));
846 case NVME_FEAT_INTERRUPT_COALESCING:
847 DPRINTF((" interrupt coalescing\r\n"));
849 case NVME_FEAT_INTERRUPT_VECTOR_CONFIGURATION:
850 DPRINTF((" interrupt vector configuration\r\n"));
852 case NVME_FEAT_WRITE_ATOMICITY:
853 DPRINTF((" write atomicity\r\n"));
855 case NVME_FEAT_ASYNC_EVENT_CONFIGURATION:
856 DPRINTF((" async event configuration\r\n"));
857 sc->async_ev_config = command->cdw11;
859 case NVME_FEAT_SOFTWARE_PROGRESS_MARKER:
860 DPRINTF((" software progress marker\r\n"));
863 DPRINTF((" autonomous power state transition\r\n"));
866 WPRINTF(("%s invalid feature 0x%x\r\n", __func__, feature));
867 pci_nvme_status_genc(&compl->status, NVME_SC_INVALID_FIELD);
871 pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
876 nvme_opc_abort(struct pci_nvme_softc* sc, struct nvme_command* command,
877 struct nvme_completion* compl)
879 DPRINTF(("%s submission queue %u, command ID 0x%x\r\n", __func__,
880 command->cdw10 & 0xFFFF, (command->cdw10 >> 16) & 0xFFFF));
882 /* TODO: search for the command ID and abort it */
885 pci_nvme_status_genc(&compl->status, NVME_SC_SUCCESS);
890 nvme_opc_async_event_req(struct pci_nvme_softc* sc,
891 struct nvme_command* command, struct nvme_completion* compl)
893 DPRINTF(("%s async event request 0x%x\r\n", __func__, command->cdw11));
896 * TODO: raise events when they happen based on the Set Features cmd.
897 * These events happen async, so only set completion successful if
898 * there is an event reflective of the request to get event.
900 pci_nvme_status_tc(&compl->status, NVME_SCT_COMMAND_SPECIFIC,
901 NVME_SC_ASYNC_EVENT_REQUEST_LIMIT_EXCEEDED);
906 pci_nvme_handle_admin_cmd(struct pci_nvme_softc* sc, uint64_t value)
908 struct nvme_completion compl;
909 struct nvme_command *cmd;
910 struct nvme_submission_queue *sq;
911 struct nvme_completion_queue *cq;
915 DPRINTF(("%s index %u\r\n", __func__, (uint32_t)value));
917 sq = &sc->submit_queues[0];
919 sqhead = atomic_load_acq_short(&sq->head);
921 if (atomic_testandset_int(&sq->busy, 1)) {
922 DPRINTF(("%s SQ busy, head %u, tail %u\r\n",
923 __func__, sqhead, sq->tail));
927 DPRINTF(("sqhead %u, tail %u\r\n", sqhead, sq->tail));
929 while (sqhead != atomic_load_acq_short(&sq->tail)) {
930 cmd = &(sq->qbase)[sqhead];
934 case NVME_OPC_DELETE_IO_SQ:
935 DPRINTF(("%s command DELETE_IO_SQ\r\n", __func__));
936 do_intr |= nvme_opc_delete_io_sq(sc, cmd, &compl);
938 case NVME_OPC_CREATE_IO_SQ:
939 DPRINTF(("%s command CREATE_IO_SQ\r\n", __func__));
940 do_intr |= nvme_opc_create_io_sq(sc, cmd, &compl);
942 case NVME_OPC_DELETE_IO_CQ:
943 DPRINTF(("%s command DELETE_IO_CQ\r\n", __func__));
944 do_intr |= nvme_opc_delete_io_cq(sc, cmd, &compl);
946 case NVME_OPC_CREATE_IO_CQ:
947 DPRINTF(("%s command CREATE_IO_CQ\r\n", __func__));
948 do_intr |= nvme_opc_create_io_cq(sc, cmd, &compl);
950 case NVME_OPC_GET_LOG_PAGE:
951 DPRINTF(("%s command GET_LOG_PAGE\r\n", __func__));
952 do_intr |= nvme_opc_get_log_page(sc, cmd, &compl);
954 case NVME_OPC_IDENTIFY:
955 DPRINTF(("%s command IDENTIFY\r\n", __func__));
956 do_intr |= nvme_opc_identify(sc, cmd, &compl);
959 DPRINTF(("%s command ABORT\r\n", __func__));
960 do_intr |= nvme_opc_abort(sc, cmd, &compl);
962 case NVME_OPC_SET_FEATURES:
963 DPRINTF(("%s command SET_FEATURES\r\n", __func__));
964 do_intr |= nvme_opc_set_features(sc, cmd, &compl);
966 case NVME_OPC_GET_FEATURES:
967 DPRINTF(("%s command GET_FEATURES\r\n", __func__));
968 do_intr |= nvme_opc_get_features(sc, cmd, &compl);
970 case NVME_OPC_ASYNC_EVENT_REQUEST:
971 DPRINTF(("%s command ASYNC_EVENT_REQ\r\n", __func__));
972 /* XXX dont care, unhandled for now
973 do_intr |= nvme_opc_async_event_req(sc, cmd, &compl);
977 WPRINTF(("0x%x command is not implemented\r\n",
981 /* for now skip async event generation */
982 if (cmd->opc != NVME_OPC_ASYNC_EVENT_REQUEST) {
983 struct nvme_completion *cp;
986 cq = &sc->compl_queues[0];
988 cp = &(cq->qbase)[cq->tail];
989 cp->cdw0 = compl.cdw0;
994 phase = NVME_STATUS_GET_P(cp->status);
995 cp->status = compl.status;
996 pci_nvme_toggle_phase(&cp->status, phase);
998 cq->tail = (cq->tail + 1) % cq->size;
1000 sqhead = (sqhead + 1) % sq->size;
1003 DPRINTF(("setting sqhead %u\r\n", sqhead));
1004 atomic_store_short(&sq->head, sqhead);
1005 atomic_store_int(&sq->busy, 0);
1008 pci_generate_msix(sc->nsc_pi, 0);
1013 pci_nvme_append_iov_req(struct pci_nvme_softc *sc, struct pci_nvme_ioreq *req,
1014 uint64_t gpaddr, size_t size, int do_write, uint64_t lba)
1019 /* concatenate contig block-iovs to minimize number of iovs */
1020 if ((req->prev_gpaddr + req->prev_size) == gpaddr) {
1021 iovidx = req->io_req.br_iovcnt - 1;
1023 req->io_req.br_iov[iovidx].iov_base =
1024 paddr_guest2host(req->sc->nsc_pi->pi_vmctx,
1025 req->prev_gpaddr, size);
1027 req->prev_size += size;
1028 req->io_req.br_resid += size;
1030 req->io_req.br_iov[iovidx].iov_len = req->prev_size;
1032 pthread_mutex_lock(&req->mtx);
1034 iovidx = req->io_req.br_iovcnt;
1035 if (iovidx == NVME_MAX_BLOCKIOVS) {
1038 DPRINTF(("large I/O, doing partial req\r\n"));
1041 req->io_req.br_iovcnt = 0;
1043 req->io_req.br_callback = pci_nvme_io_partial;
1046 err = blockif_read(sc->nvstore.ctx,
1049 err = blockif_write(sc->nvstore.ctx,
1052 /* wait until req completes before cont */
1054 pthread_cond_wait(&req->cv, &req->mtx);
1057 req->io_req.br_offset = lba;
1058 req->io_req.br_resid = 0;
1059 req->io_req.br_param = req;
1062 req->io_req.br_iov[iovidx].iov_base =
1063 paddr_guest2host(req->sc->nsc_pi->pi_vmctx,
1066 req->io_req.br_iov[iovidx].iov_len = size;
1068 req->prev_gpaddr = gpaddr;
1069 req->prev_size = size;
1070 req->io_req.br_resid += size;
1072 req->io_req.br_iovcnt++;
1074 pthread_mutex_unlock(&req->mtx);
1077 /* RAM buffer: read/write directly */
1078 void *p = sc->nvstore.ctx;
1081 if ((lba + size) > sc->nvstore.size) {
1082 WPRINTF(("%s write would overflow RAM\r\n", __func__));
1086 p = (void *)((uintptr_t)p + (uintptr_t)lba);
1087 gptr = paddr_guest2host(sc->nsc_pi->pi_vmctx, gpaddr, size);
1089 memcpy(p, gptr, size);
1091 memcpy(gptr, p, size);
1097 pci_nvme_set_completion(struct pci_nvme_softc *sc,
1098 struct nvme_submission_queue *sq, int sqid, uint16_t cid,
1099 uint32_t cdw0, uint16_t status, int ignore_busy)
1101 struct nvme_completion_queue *cq = &sc->compl_queues[sq->cqid];
1102 struct nvme_completion *compl;
1106 DPRINTF(("%s sqid %d cqid %u cid %u status: 0x%x 0x%x\r\n",
1107 __func__, sqid, sq->cqid, cid, NVME_STATUS_GET_SCT(status),
1108 NVME_STATUS_GET_SC(status)));
1110 pthread_mutex_lock(&cq->mtx);
1112 assert(cq->qbase != NULL);
1114 compl = &cq->qbase[cq->tail];
1116 compl->sqhd = atomic_load_acq_short(&sq->head);
1121 phase = NVME_STATUS_GET_P(compl->status);
1122 compl->status = status;
1123 pci_nvme_toggle_phase(&compl->status, phase);
1125 cq->tail = (cq->tail + 1) % cq->size;
1127 if (cq->intr_en & NVME_CQ_INTEN)
1130 pthread_mutex_unlock(&cq->mtx);
1132 if (ignore_busy || !atomic_load_acq_int(&sq->busy))
1134 pci_generate_msix(sc->nsc_pi, cq->intr_vec);
1138 pci_nvme_release_ioreq(struct pci_nvme_softc *sc, struct pci_nvme_ioreq *req)
1141 req->nvme_sq = NULL;
1144 pthread_mutex_lock(&sc->mtx);
1146 req->next = sc->ioreqs_free;
1147 sc->ioreqs_free = req;
1150 /* when no more IO pending, can set to ready if device reset/enabled */
1151 if (sc->pending_ios == 0 &&
1152 NVME_CC_GET_EN(sc->regs.cc) && !(NVME_CSTS_GET_RDY(sc->regs.csts)))
1153 sc->regs.csts |= NVME_CSTS_RDY;
1155 pthread_mutex_unlock(&sc->mtx);
1157 sem_post(&sc->iosemlock);
1160 static struct pci_nvme_ioreq *
1161 pci_nvme_get_ioreq(struct pci_nvme_softc *sc)
1163 struct pci_nvme_ioreq *req = NULL;;
1165 sem_wait(&sc->iosemlock);
1166 pthread_mutex_lock(&sc->mtx);
1168 req = sc->ioreqs_free;
1169 assert(req != NULL);
1171 sc->ioreqs_free = req->next;
1178 pthread_mutex_unlock(&sc->mtx);
1180 req->io_req.br_iovcnt = 0;
1181 req->io_req.br_offset = 0;
1182 req->io_req.br_resid = 0;
1183 req->io_req.br_param = req;
1184 req->prev_gpaddr = 0;
1191 pci_nvme_io_done(struct blockif_req *br, int err)
1193 struct pci_nvme_ioreq *req = br->br_param;
1194 struct nvme_submission_queue *sq = req->nvme_sq;
1195 uint16_t code, status;
1197 DPRINTF(("%s error %d %s\r\n", __func__, err, strerror(err)));
1199 /* TODO return correct error */
1200 code = err ? NVME_SC_DATA_TRANSFER_ERROR : NVME_SC_SUCCESS;
1201 pci_nvme_status_genc(&status, code);
1203 pci_nvme_set_completion(req->sc, sq, req->sqid, req->cid, 0, status, 0);
1204 pci_nvme_release_ioreq(req->sc, req);
1208 pci_nvme_io_partial(struct blockif_req *br, int err)
1210 struct pci_nvme_ioreq *req = br->br_param;
1212 DPRINTF(("%s error %d %s\r\n", __func__, err, strerror(err)));
1214 pthread_cond_signal(&req->cv);
1219 pci_nvme_handle_io_cmd(struct pci_nvme_softc* sc, uint16_t idx)
1221 struct nvme_submission_queue *sq;
1226 /* handle all submissions up to sq->tail index */
1227 sq = &sc->submit_queues[idx];
1229 if (atomic_testandset_int(&sq->busy, 1)) {
1230 DPRINTF(("%s sqid %u busy\r\n", __func__, idx));
1234 sqhead = atomic_load_acq_short(&sq->head);
1236 DPRINTF(("nvme_handle_io qid %u head %u tail %u cmdlist %p\r\n",
1237 idx, sqhead, sq->tail, sq->qbase));
1239 while (sqhead != atomic_load_acq_short(&sq->tail)) {
1240 struct nvme_command *cmd;
1241 struct pci_nvme_ioreq *req = NULL;
1243 uint64_t nblocks, bytes, size, cpsz;
1245 /* TODO: support scatter gather list handling */
1247 cmd = &sq->qbase[sqhead];
1248 sqhead = (sqhead + 1) % sq->size;
1250 lba = ((uint64_t)cmd->cdw11 << 32) | cmd->cdw10;
1252 if (cmd->opc == NVME_OPC_FLUSH) {
1253 pci_nvme_status_genc(&status, NVME_SC_SUCCESS);
1254 pci_nvme_set_completion(sc, sq, idx, cmd->cid, 0,
1258 } else if (cmd->opc == 0x08) {
1259 /* TODO: write zeroes */
1260 WPRINTF(("%s write zeroes lba 0x%lx blocks %u\r\n",
1261 __func__, lba, cmd->cdw12 & 0xFFFF));
1262 pci_nvme_status_genc(&status, NVME_SC_SUCCESS);
1263 pci_nvme_set_completion(sc, sq, idx, cmd->cid, 0,
1269 nblocks = (cmd->cdw12 & 0xFFFF) + 1;
1271 bytes = nblocks * sc->nvstore.sectsz;
1273 if (sc->nvstore.type == NVME_STOR_BLOCKIF) {
1274 req = pci_nvme_get_ioreq(sc);
1280 * If data starts mid-page and flows into the next page, then
1281 * increase page count
1284 DPRINTF(("[h%u:t%u:n%u] %s starting LBA 0x%lx blocks %lu "
1286 sqhead==0 ? sq->size-1 : sqhead-1, sq->tail, sq->size,
1287 cmd->opc == NVME_OPC_WRITE ?
1289 lba, nblocks, bytes));
1291 cmd->prp1 &= ~(0x03UL);
1292 cmd->prp2 &= ~(0x03UL);
1294 DPRINTF((" prp1 0x%lx prp2 0x%lx\r\n", cmd->prp1, cmd->prp2));
1297 lba *= sc->nvstore.sectsz;
1299 cpsz = PAGE_SIZE - (cmd->prp1 % PAGE_SIZE);
1305 req->io_req.br_offset = ((uint64_t)cmd->cdw11 << 32) |
1307 req->opc = cmd->opc;
1308 req->cid = cmd->cid;
1309 req->nsid = cmd->nsid;
1312 err = pci_nvme_append_iov_req(sc, req, cmd->prp1, cpsz,
1313 cmd->opc == NVME_OPC_WRITE, lba);
1320 if (size <= PAGE_SIZE) {
1321 /* prp2 is second (and final) page in transfer */
1323 err = pci_nvme_append_iov_req(sc, req, cmd->prp2,
1325 cmd->opc == NVME_OPC_WRITE,
1331 /* prp2 is pointer to a physical region page list */
1332 prp_list = paddr_guest2host(sc->nsc_pi->pi_vmctx,
1333 cmd->prp2, PAGE_SIZE);
1337 cpsz = MIN(size, PAGE_SIZE);
1340 * Move to linked physical region page list
1343 if (i == (NVME_PRP2_ITEMS-1) &&
1345 assert((prp_list[i] & (PAGE_SIZE-1)) == 0);
1346 prp_list = paddr_guest2host(
1347 sc->nsc_pi->pi_vmctx,
1348 prp_list[i], PAGE_SIZE);
1351 if (prp_list[i] == 0) {
1352 WPRINTF(("PRP2[%d] = 0 !!!\r\n", i));
1357 err = pci_nvme_append_iov_req(sc, req,
1359 cmd->opc == NVME_OPC_WRITE, lba);
1370 if (sc->nvstore.type == NVME_STOR_RAM) {
1371 uint16_t code, status;
1373 code = err ? NVME_SC_LBA_OUT_OF_RANGE :
1375 pci_nvme_status_genc(&status, code);
1377 pci_nvme_set_completion(sc, sq, idx, cmd->cid, 0,
1387 req->io_req.br_callback = pci_nvme_io_done;
1392 err = blockif_read(sc->nvstore.ctx, &req->io_req);
1394 case NVME_OPC_WRITE:
1395 err = blockif_write(sc->nvstore.ctx, &req->io_req);
1398 WPRINTF(("%s unhandled io command 0x%x\r\n",
1399 __func__, cmd->opc));
1407 pci_nvme_status_genc(&status,
1408 NVME_SC_DATA_TRANSFER_ERROR);
1410 pci_nvme_set_completion(sc, sq, idx, cmd->cid, 0,
1412 pci_nvme_release_ioreq(sc, req);
1416 atomic_store_short(&sq->head, sqhead);
1417 atomic_store_int(&sq->busy, 0);
1421 pci_nvme_handle_doorbell(struct vmctx *ctx, struct pci_nvme_softc* sc,
1422 uint64_t idx, int is_sq, uint64_t value)
1424 DPRINTF(("nvme doorbell %lu, %s, val 0x%lx\r\n",
1425 idx, is_sq ? "SQ" : "CQ", value & 0xFFFF));
1428 atomic_store_short(&sc->submit_queues[idx].tail,
1432 pci_nvme_handle_admin_cmd(sc, value);
1434 /* submission queue; handle new entries in SQ */
1435 if (idx > sc->num_squeues) {
1436 WPRINTF(("%s SQ index %lu overflow from "
1437 "guest (max %u)\r\n",
1438 __func__, idx, sc->num_squeues));
1441 pci_nvme_handle_io_cmd(sc, (uint16_t)idx);
1444 if (idx > sc->num_cqueues) {
1445 WPRINTF(("%s queue index %lu overflow from "
1446 "guest (max %u)\r\n",
1447 __func__, idx, sc->num_cqueues));
1451 sc->compl_queues[idx].head = (uint16_t)value;
1456 pci_nvme_bar0_reg_dumps(const char *func, uint64_t offset, int iswrite)
1458 const char *s = iswrite ? "WRITE" : "READ";
1461 case NVME_CR_CAP_LOW:
1462 DPRINTF(("%s %s NVME_CR_CAP_LOW\r\n", func, s));
1464 case NVME_CR_CAP_HI:
1465 DPRINTF(("%s %s NVME_CR_CAP_HI\r\n", func, s));
1468 DPRINTF(("%s %s NVME_CR_VS\r\n", func, s));
1471 DPRINTF(("%s %s NVME_CR_INTMS\r\n", func, s));
1474 DPRINTF(("%s %s NVME_CR_INTMC\r\n", func, s));
1477 DPRINTF(("%s %s NVME_CR_CC\r\n", func, s));
1480 DPRINTF(("%s %s NVME_CR_CSTS\r\n", func, s));
1483 DPRINTF(("%s %s NVME_CR_NSSR\r\n", func, s));
1486 DPRINTF(("%s %s NVME_CR_AQA\r\n", func, s));
1488 case NVME_CR_ASQ_LOW:
1489 DPRINTF(("%s %s NVME_CR_ASQ_LOW\r\n", func, s));
1491 case NVME_CR_ASQ_HI:
1492 DPRINTF(("%s %s NVME_CR_ASQ_HI\r\n", func, s));
1494 case NVME_CR_ACQ_LOW:
1495 DPRINTF(("%s %s NVME_CR_ACQ_LOW\r\n", func, s));
1497 case NVME_CR_ACQ_HI:
1498 DPRINTF(("%s %s NVME_CR_ACQ_HI\r\n", func, s));
1501 DPRINTF(("unknown nvme bar-0 offset 0x%lx\r\n", offset));
1507 pci_nvme_write_bar_0(struct vmctx *ctx, struct pci_nvme_softc* sc,
1508 uint64_t offset, int size, uint64_t value)
1512 if (offset >= NVME_DOORBELL_OFFSET) {
1513 uint64_t belloffset = offset - NVME_DOORBELL_OFFSET;
1514 uint64_t idx = belloffset / 8; /* door bell size = 2*int */
1515 int is_sq = (belloffset % 8) < 4;
1517 if (belloffset > ((sc->max_queues+1) * 8 - 4)) {
1518 WPRINTF(("guest attempted an overflow write offset "
1519 "0x%lx, val 0x%lx in %s",
1520 offset, value, __func__));
1524 pci_nvme_handle_doorbell(ctx, sc, idx, is_sq, value);
1528 DPRINTF(("nvme-write offset 0x%lx, size %d, value 0x%lx\r\n",
1529 offset, size, value));
1532 WPRINTF(("guest wrote invalid size %d (offset 0x%lx, "
1533 "val 0x%lx) to bar0 in %s",
1534 size, offset, value, __func__));
1535 /* TODO: shutdown device */
1539 pci_nvme_bar0_reg_dumps(__func__, offset, 1);
1541 pthread_mutex_lock(&sc->mtx);
1544 case NVME_CR_CAP_LOW:
1545 case NVME_CR_CAP_HI:
1552 /* MSI-X, so ignore */
1555 /* MSI-X, so ignore */
1558 ccreg = (uint32_t)value;
1560 DPRINTF(("%s NVME_CR_CC en %x css %x shn %x iosqes %u "
1563 NVME_CC_GET_EN(ccreg), NVME_CC_GET_CSS(ccreg),
1564 NVME_CC_GET_SHN(ccreg), NVME_CC_GET_IOSQES(ccreg),
1565 NVME_CC_GET_IOCQES(ccreg)));
1567 if (NVME_CC_GET_SHN(ccreg)) {
1568 /* perform shutdown - flush out data to backend */
1569 sc->regs.csts &= ~(NVME_CSTS_REG_SHST_MASK <<
1570 NVME_CSTS_REG_SHST_SHIFT);
1571 sc->regs.csts |= NVME_SHST_COMPLETE <<
1572 NVME_CSTS_REG_SHST_SHIFT;
1574 if (NVME_CC_GET_EN(ccreg) != NVME_CC_GET_EN(sc->regs.cc)) {
1575 if (NVME_CC_GET_EN(ccreg) == 0)
1576 /* transition 1-> causes controller reset */
1577 pci_nvme_reset_locked(sc);
1579 pci_nvme_init_controller(ctx, sc);
1582 /* Insert the iocqes, iosqes and en bits from the write */
1583 sc->regs.cc &= ~NVME_CC_WRITE_MASK;
1584 sc->regs.cc |= ccreg & NVME_CC_WRITE_MASK;
1585 if (NVME_CC_GET_EN(ccreg) == 0) {
1586 /* Insert the ams, mps and css bit fields */
1587 sc->regs.cc &= ~NVME_CC_NEN_WRITE_MASK;
1588 sc->regs.cc |= ccreg & NVME_CC_NEN_WRITE_MASK;
1589 sc->regs.csts &= ~NVME_CSTS_RDY;
1590 } else if (sc->pending_ios == 0) {
1591 sc->regs.csts |= NVME_CSTS_RDY;
1597 /* ignore writes; don't support subsystem reset */
1600 sc->regs.aqa = (uint32_t)value;
1602 case NVME_CR_ASQ_LOW:
1603 sc->regs.asq = (sc->regs.asq & (0xFFFFFFFF00000000)) |
1604 (0xFFFFF000 & value);
1606 case NVME_CR_ASQ_HI:
1607 sc->regs.asq = (sc->regs.asq & (0x00000000FFFFFFFF)) |
1610 case NVME_CR_ACQ_LOW:
1611 sc->regs.acq = (sc->regs.acq & (0xFFFFFFFF00000000)) |
1612 (0xFFFFF000 & value);
1614 case NVME_CR_ACQ_HI:
1615 sc->regs.acq = (sc->regs.acq & (0x00000000FFFFFFFF)) |
1619 DPRINTF(("%s unknown offset 0x%lx, value 0x%lx size %d\r\n",
1620 __func__, offset, value, size));
1622 pthread_mutex_unlock(&sc->mtx);
1626 pci_nvme_write(struct vmctx *ctx, int vcpu, struct pci_devinst *pi,
1627 int baridx, uint64_t offset, int size, uint64_t value)
1629 struct pci_nvme_softc* sc = pi->pi_arg;
1631 if (baridx == pci_msix_table_bar(pi) ||
1632 baridx == pci_msix_pba_bar(pi)) {
1633 DPRINTF(("nvme-write baridx %d, msix: off 0x%lx, size %d, "
1634 " value 0x%lx\r\n", baridx, offset, size, value));
1636 pci_emul_msix_twrite(pi, offset, size, value);
1642 pci_nvme_write_bar_0(ctx, sc, offset, size, value);
1646 DPRINTF(("%s unknown baridx %d, val 0x%lx\r\n",
1647 __func__, baridx, value));
1651 static uint64_t pci_nvme_read_bar_0(struct pci_nvme_softc* sc,
1652 uint64_t offset, int size)
1656 pci_nvme_bar0_reg_dumps(__func__, offset, 0);
1658 if (offset < NVME_DOORBELL_OFFSET) {
1659 void *p = &(sc->regs);
1660 pthread_mutex_lock(&sc->mtx);
1661 memcpy(&value, (void *)((uintptr_t)p + offset), size);
1662 pthread_mutex_unlock(&sc->mtx);
1665 WPRINTF(("pci_nvme: read invalid offset %ld\r\n", offset));
1676 value &= 0xFFFFFFFF;
1680 DPRINTF((" nvme-read offset 0x%lx, size %d -> value 0x%x\r\n",
1681 offset, size, (uint32_t)value));
1689 pci_nvme_read(struct vmctx *ctx, int vcpu, struct pci_devinst *pi, int baridx,
1690 uint64_t offset, int size)
1692 struct pci_nvme_softc* sc = pi->pi_arg;
1694 if (baridx == pci_msix_table_bar(pi) ||
1695 baridx == pci_msix_pba_bar(pi)) {
1696 DPRINTF(("nvme-read bar: %d, msix: regoff 0x%lx, size %d\r\n",
1697 baridx, offset, size));
1699 return pci_emul_msix_tread(pi, offset, size);
1704 return pci_nvme_read_bar_0(sc, offset, size);
1707 DPRINTF(("unknown bar %d, 0x%lx\r\n", baridx, offset));
1715 pci_nvme_parse_opts(struct pci_nvme_softc *sc, char *opts)
1717 char bident[sizeof("XX:X:X")];
1718 char *uopt, *xopts, *config;
1722 sc->max_queues = NVME_QUEUES;
1723 sc->max_qentries = NVME_MAX_QENTRIES;
1724 sc->ioslots = NVME_IOSLOTS;
1725 sc->num_squeues = sc->max_queues;
1726 sc->num_cqueues = sc->max_queues;
1729 uopt = strdup(opts);
1731 snprintf(sc->ctrldata.sn, sizeof(sc->ctrldata.sn),
1732 "NVME-%d-%d", sc->nsc_pi->pi_slot, sc->nsc_pi->pi_func);
1733 for (xopts = strtok(uopt, ",");
1735 xopts = strtok(NULL, ",")) {
1737 if ((config = strchr(xopts, '=')) != NULL)
1740 if (!strcmp("maxq", xopts)) {
1741 sc->max_queues = atoi(config);
1742 } else if (!strcmp("qsz", xopts)) {
1743 sc->max_qentries = atoi(config);
1744 } else if (!strcmp("ioslots", xopts)) {
1745 sc->ioslots = atoi(config);
1746 } else if (!strcmp("sectsz", xopts)) {
1747 sectsz = atoi(config);
1748 } else if (!strcmp("ser", xopts)) {
1750 * This field indicates the Product Serial Number in
1751 * 7-bit ASCII, unused bytes should be space characters.
1754 cpywithpad((char *)sc->ctrldata.sn,
1755 sizeof(sc->ctrldata.sn), config, ' ');
1756 } else if (!strcmp("ram", xopts)) {
1757 uint64_t sz = strtoull(&xopts[4], NULL, 10);
1759 sc->nvstore.type = NVME_STOR_RAM;
1760 sc->nvstore.size = sz * 1024 * 1024;
1761 sc->nvstore.ctx = calloc(1, sc->nvstore.size);
1762 sc->nvstore.sectsz = 4096;
1763 sc->nvstore.sectsz_bits = 12;
1764 if (sc->nvstore.ctx == NULL) {
1765 perror("Unable to allocate RAM");
1769 } else if (optidx == 0) {
1770 snprintf(bident, sizeof(bident), "%d:%d",
1771 sc->nsc_pi->pi_slot, sc->nsc_pi->pi_func);
1772 sc->nvstore.ctx = blockif_open(xopts, bident);
1773 if (sc->nvstore.ctx == NULL) {
1774 perror("Could not open backing file");
1778 sc->nvstore.type = NVME_STOR_BLOCKIF;
1779 sc->nvstore.size = blockif_size(sc->nvstore.ctx);
1781 fprintf(stderr, "Invalid option %s\n", xopts);
1790 if (sc->nvstore.ctx == NULL || sc->nvstore.size == 0) {
1791 fprintf(stderr, "backing store not specified\n");
1794 if (sectsz == 512 || sectsz == 4096 || sectsz == 8192)
1795 sc->nvstore.sectsz = sectsz;
1796 else if (sc->nvstore.type != NVME_STOR_RAM)
1797 sc->nvstore.sectsz = blockif_sectsz(sc->nvstore.ctx);
1798 for (sc->nvstore.sectsz_bits = 9;
1799 (1 << sc->nvstore.sectsz_bits) < sc->nvstore.sectsz;
1800 sc->nvstore.sectsz_bits++);
1802 if (sc->max_queues <= 0 || sc->max_queues > NVME_QUEUES)
1803 sc->max_queues = NVME_QUEUES;
1805 if (sc->max_qentries <= 0) {
1806 fprintf(stderr, "Invalid qsz option\n");
1809 if (sc->ioslots <= 0) {
1810 fprintf(stderr, "Invalid ioslots option\n");
1818 pci_nvme_init(struct vmctx *ctx, struct pci_devinst *pi, char *opts)
1820 struct pci_nvme_softc *sc;
1821 uint32_t pci_membar_sz;
1826 sc = calloc(1, sizeof(struct pci_nvme_softc));
1830 error = pci_nvme_parse_opts(sc, opts);
1836 sc->ioreqs = calloc(sc->ioslots, sizeof(struct pci_nvme_ioreq));
1837 for (int i = 0; i < sc->ioslots; i++) {
1838 if (i < (sc->ioslots-1))
1839 sc->ioreqs[i].next = &sc->ioreqs[i+1];
1840 pthread_mutex_init(&sc->ioreqs[i].mtx, NULL);
1841 pthread_cond_init(&sc->ioreqs[i].cv, NULL);
1843 sc->ioreqs_free = sc->ioreqs;
1844 sc->intr_coales_aggr_thresh = 1;
1846 pci_set_cfgdata16(pi, PCIR_DEVICE, 0x0A0A);
1847 pci_set_cfgdata16(pi, PCIR_VENDOR, 0xFB5D);
1848 pci_set_cfgdata8(pi, PCIR_CLASS, PCIC_STORAGE);
1849 pci_set_cfgdata8(pi, PCIR_SUBCLASS, PCIS_STORAGE_NVM);
1850 pci_set_cfgdata8(pi, PCIR_PROGIF,
1851 PCIP_STORAGE_NVM_ENTERPRISE_NVMHCI_1_0);
1854 * Allocate size of NVMe registers + doorbell space for all queues.
1856 * The specification requires a minimum memory I/O window size of 16K.
1857 * The Windows driver will refuse to start a device with a smaller
1860 pci_membar_sz = sizeof(struct nvme_registers) +
1861 2 * sizeof(uint32_t) * (sc->max_queues + 1);
1862 pci_membar_sz = MAX(pci_membar_sz, NVME_MMIO_SPACE_MIN);
1864 DPRINTF(("nvme membar size: %u\r\n", pci_membar_sz));
1866 error = pci_emul_alloc_bar(pi, 0, PCIBAR_MEM64, pci_membar_sz);
1868 WPRINTF(("%s pci alloc mem bar failed\r\n", __func__));
1872 error = pci_emul_add_msixcap(pi, sc->max_queues + 1, NVME_MSIX_BAR);
1874 WPRINTF(("%s pci add msixcap failed\r\n", __func__));
1878 pthread_mutex_init(&sc->mtx, NULL);
1879 sem_init(&sc->iosemlock, 0, sc->ioslots);
1882 pci_nvme_init_ctrldata(sc);
1883 pci_nvme_init_nsdata(sc);
1885 pci_lintr_request(pi);
1892 struct pci_devemu pci_de_nvme = {
1894 .pe_init = pci_nvme_init,
1895 .pe_barwrite = pci_nvme_write,
1896 .pe_barread = pci_nvme_read
1898 PCI_EMUL_SET(pci_de_nvme);