1 // SPDX-License-Identifier: GPL-2.0
3 * NVMe over Fabrics RDMA target.
4 * Copyright (c) 2015-2016 HGST, a Western Digital Company.
6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7 #include <linux/atomic.h>
8 #include <linux/ctype.h>
9 #include <linux/delay.h>
10 #include <linux/err.h>
11 #include <linux/init.h>
12 #include <linux/module.h>
13 #include <linux/nvme.h>
14 #include <linux/slab.h>
15 #include <linux/string.h>
16 #include <linux/wait.h>
17 #include <linux/inet.h>
18 #include <asm/unaligned.h>
20 #include <rdma/ib_verbs.h>
21 #include <rdma/rdma_cm.h>
23 #include <rdma/ib_cm.h>
25 #include <linux/nvme-rdma.h>
29 * We allow at least 1 page, up to 4 SGEs, and up to 16KB of inline data
31 #define NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE PAGE_SIZE
32 #define NVMET_RDMA_MAX_INLINE_SGE 4
33 #define NVMET_RDMA_MAX_INLINE_DATA_SIZE max_t(int, SZ_16K, PAGE_SIZE)
35 /* Assume mpsmin == device_page_size == 4KB */
36 #define NVMET_RDMA_MAX_MDTS 8
37 #define NVMET_RDMA_MAX_METADATA_MDTS 5
39 struct nvmet_rdma_srq;
41 struct nvmet_rdma_cmd {
42 struct ib_sge sge[NVMET_RDMA_MAX_INLINE_SGE + 1];
45 struct scatterlist inline_sg[NVMET_RDMA_MAX_INLINE_SGE];
46 struct nvme_command *nvme_cmd;
47 struct nvmet_rdma_queue *queue;
48 struct nvmet_rdma_srq *nsrq;
52 NVMET_RDMA_REQ_INLINE_DATA = (1 << 0),
53 NVMET_RDMA_REQ_INVALIDATE_RKEY = (1 << 1),
56 struct nvmet_rdma_rsp {
57 struct ib_sge send_sge;
58 struct ib_cqe send_cqe;
59 struct ib_send_wr send_wr;
61 struct nvmet_rdma_cmd *cmd;
62 struct nvmet_rdma_queue *queue;
64 struct ib_cqe read_cqe;
65 struct ib_cqe write_cqe;
66 struct rdma_rw_ctx rw;
75 struct list_head wait_list;
76 struct list_head free_list;
79 enum nvmet_rdma_queue_state {
80 NVMET_RDMA_Q_CONNECTING,
82 NVMET_RDMA_Q_DISCONNECTING,
85 struct nvmet_rdma_queue {
86 struct rdma_cm_id *cm_id;
88 struct nvmet_port *port;
91 struct nvmet_rdma_device *dev;
92 struct nvmet_rdma_srq *nsrq;
93 spinlock_t state_lock;
94 enum nvmet_rdma_queue_state state;
95 struct nvmet_cq nvme_cq;
96 struct nvmet_sq nvme_sq;
98 struct nvmet_rdma_rsp *rsps;
99 struct list_head free_rsps;
100 spinlock_t rsps_lock;
101 struct nvmet_rdma_cmd *cmds;
103 struct work_struct release_work;
104 struct list_head rsp_wait_list;
105 struct list_head rsp_wr_wait_list;
106 spinlock_t rsp_wr_wait_lock;
114 struct list_head queue_list;
117 struct nvmet_rdma_port {
118 struct nvmet_port *nport;
119 struct sockaddr_storage addr;
120 struct rdma_cm_id *cm_id;
121 struct delayed_work repair_work;
124 struct nvmet_rdma_srq {
126 struct nvmet_rdma_cmd *cmds;
127 struct nvmet_rdma_device *ndev;
130 struct nvmet_rdma_device {
131 struct ib_device *device;
133 struct nvmet_rdma_srq **srqs;
137 struct list_head entry;
138 int inline_data_size;
139 int inline_page_count;
142 static bool nvmet_rdma_use_srq;
143 module_param_named(use_srq, nvmet_rdma_use_srq, bool, 0444);
144 MODULE_PARM_DESC(use_srq, "Use shared receive queue.");
146 static int srq_size_set(const char *val, const struct kernel_param *kp);
147 static const struct kernel_param_ops srq_size_ops = {
149 .get = param_get_int,
152 static int nvmet_rdma_srq_size = 1024;
153 module_param_cb(srq_size, &srq_size_ops, &nvmet_rdma_srq_size, 0644);
154 MODULE_PARM_DESC(srq_size, "set Shared Receive Queue (SRQ) size, should >= 256 (default: 1024)");
156 static DEFINE_IDA(nvmet_rdma_queue_ida);
157 static LIST_HEAD(nvmet_rdma_queue_list);
158 static DEFINE_MUTEX(nvmet_rdma_queue_mutex);
160 static LIST_HEAD(device_list);
161 static DEFINE_MUTEX(device_list_mutex);
163 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp);
164 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc);
165 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
166 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc);
167 static void nvmet_rdma_write_data_done(struct ib_cq *cq, struct ib_wc *wc);
168 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv);
169 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue);
170 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
171 struct nvmet_rdma_rsp *r);
172 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
173 struct nvmet_rdma_rsp *r);
175 static const struct nvmet_fabrics_ops nvmet_rdma_ops;
177 static int srq_size_set(const char *val, const struct kernel_param *kp)
181 ret = kstrtoint(val, 10, &n);
182 if (ret != 0 || n < 256)
185 return param_set_int(val, kp);
188 static int num_pages(int len)
190 return 1 + (((len - 1) & PAGE_MASK) >> PAGE_SHIFT);
193 static inline bool nvmet_rdma_need_data_in(struct nvmet_rdma_rsp *rsp)
195 return nvme_is_write(rsp->req.cmd) &&
196 rsp->req.transfer_len &&
197 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
200 static inline bool nvmet_rdma_need_data_out(struct nvmet_rdma_rsp *rsp)
202 return !nvme_is_write(rsp->req.cmd) &&
203 rsp->req.transfer_len &&
204 !rsp->req.cqe->status &&
205 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
208 static inline struct nvmet_rdma_rsp *
209 nvmet_rdma_get_rsp(struct nvmet_rdma_queue *queue)
211 struct nvmet_rdma_rsp *rsp;
214 spin_lock_irqsave(&queue->rsps_lock, flags);
215 rsp = list_first_entry_or_null(&queue->free_rsps,
216 struct nvmet_rdma_rsp, free_list);
218 list_del(&rsp->free_list);
219 spin_unlock_irqrestore(&queue->rsps_lock, flags);
221 if (unlikely(!rsp)) {
224 rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
227 ret = nvmet_rdma_alloc_rsp(queue->dev, rsp);
233 rsp->allocated = true;
240 nvmet_rdma_put_rsp(struct nvmet_rdma_rsp *rsp)
244 if (unlikely(rsp->allocated)) {
245 nvmet_rdma_free_rsp(rsp->queue->dev, rsp);
250 spin_lock_irqsave(&rsp->queue->rsps_lock, flags);
251 list_add_tail(&rsp->free_list, &rsp->queue->free_rsps);
252 spin_unlock_irqrestore(&rsp->queue->rsps_lock, flags);
255 static void nvmet_rdma_free_inline_pages(struct nvmet_rdma_device *ndev,
256 struct nvmet_rdma_cmd *c)
258 struct scatterlist *sg;
262 if (!ndev->inline_data_size)
268 for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
270 ib_dma_unmap_page(ndev->device, sge->addr,
271 sge->length, DMA_FROM_DEVICE);
273 __free_page(sg_page(sg));
277 static int nvmet_rdma_alloc_inline_pages(struct nvmet_rdma_device *ndev,
278 struct nvmet_rdma_cmd *c)
280 struct scatterlist *sg;
286 if (!ndev->inline_data_size)
290 sg_init_table(sg, ndev->inline_page_count);
292 len = ndev->inline_data_size;
294 for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
295 pg = alloc_page(GFP_KERNEL);
298 sg_assign_page(sg, pg);
299 sge->addr = ib_dma_map_page(ndev->device,
300 pg, 0, PAGE_SIZE, DMA_FROM_DEVICE);
301 if (ib_dma_mapping_error(ndev->device, sge->addr))
303 sge->length = min_t(int, len, PAGE_SIZE);
304 sge->lkey = ndev->pd->local_dma_lkey;
310 for (; i >= 0; i--, sg--, sge--) {
312 ib_dma_unmap_page(ndev->device, sge->addr,
313 sge->length, DMA_FROM_DEVICE);
315 __free_page(sg_page(sg));
320 static int nvmet_rdma_alloc_cmd(struct nvmet_rdma_device *ndev,
321 struct nvmet_rdma_cmd *c, bool admin)
323 /* NVMe command / RDMA RECV */
324 c->nvme_cmd = kmalloc(sizeof(*c->nvme_cmd), GFP_KERNEL);
328 c->sge[0].addr = ib_dma_map_single(ndev->device, c->nvme_cmd,
329 sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
330 if (ib_dma_mapping_error(ndev->device, c->sge[0].addr))
333 c->sge[0].length = sizeof(*c->nvme_cmd);
334 c->sge[0].lkey = ndev->pd->local_dma_lkey;
336 if (!admin && nvmet_rdma_alloc_inline_pages(ndev, c))
339 c->cqe.done = nvmet_rdma_recv_done;
341 c->wr.wr_cqe = &c->cqe;
342 c->wr.sg_list = c->sge;
343 c->wr.num_sge = admin ? 1 : ndev->inline_page_count + 1;
348 ib_dma_unmap_single(ndev->device, c->sge[0].addr,
349 sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
357 static void nvmet_rdma_free_cmd(struct nvmet_rdma_device *ndev,
358 struct nvmet_rdma_cmd *c, bool admin)
361 nvmet_rdma_free_inline_pages(ndev, c);
362 ib_dma_unmap_single(ndev->device, c->sge[0].addr,
363 sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
367 static struct nvmet_rdma_cmd *
368 nvmet_rdma_alloc_cmds(struct nvmet_rdma_device *ndev,
369 int nr_cmds, bool admin)
371 struct nvmet_rdma_cmd *cmds;
372 int ret = -EINVAL, i;
374 cmds = kcalloc(nr_cmds, sizeof(struct nvmet_rdma_cmd), GFP_KERNEL);
378 for (i = 0; i < nr_cmds; i++) {
379 ret = nvmet_rdma_alloc_cmd(ndev, cmds + i, admin);
388 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
394 static void nvmet_rdma_free_cmds(struct nvmet_rdma_device *ndev,
395 struct nvmet_rdma_cmd *cmds, int nr_cmds, bool admin)
399 for (i = 0; i < nr_cmds; i++)
400 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
404 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
405 struct nvmet_rdma_rsp *r)
407 /* NVMe CQE / RDMA SEND */
408 r->req.cqe = kmalloc(sizeof(*r->req.cqe), GFP_KERNEL);
412 r->send_sge.addr = ib_dma_map_single(ndev->device, r->req.cqe,
413 sizeof(*r->req.cqe), DMA_TO_DEVICE);
414 if (ib_dma_mapping_error(ndev->device, r->send_sge.addr))
417 if (!ib_uses_virt_dma(ndev->device))
418 r->req.p2p_client = &ndev->device->dev;
419 r->send_sge.length = sizeof(*r->req.cqe);
420 r->send_sge.lkey = ndev->pd->local_dma_lkey;
422 r->send_cqe.done = nvmet_rdma_send_done;
424 r->send_wr.wr_cqe = &r->send_cqe;
425 r->send_wr.sg_list = &r->send_sge;
426 r->send_wr.num_sge = 1;
427 r->send_wr.send_flags = IB_SEND_SIGNALED;
429 /* Data In / RDMA READ */
430 r->read_cqe.done = nvmet_rdma_read_data_done;
431 /* Data Out / RDMA WRITE */
432 r->write_cqe.done = nvmet_rdma_write_data_done;
442 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
443 struct nvmet_rdma_rsp *r)
445 ib_dma_unmap_single(ndev->device, r->send_sge.addr,
446 sizeof(*r->req.cqe), DMA_TO_DEVICE);
451 nvmet_rdma_alloc_rsps(struct nvmet_rdma_queue *queue)
453 struct nvmet_rdma_device *ndev = queue->dev;
454 int nr_rsps = queue->recv_queue_size * 2;
455 int ret = -EINVAL, i;
457 queue->rsps = kcalloc(nr_rsps, sizeof(struct nvmet_rdma_rsp),
462 for (i = 0; i < nr_rsps; i++) {
463 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
465 ret = nvmet_rdma_alloc_rsp(ndev, rsp);
469 list_add_tail(&rsp->free_list, &queue->free_rsps);
476 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
478 list_del(&rsp->free_list);
479 nvmet_rdma_free_rsp(ndev, rsp);
486 static void nvmet_rdma_free_rsps(struct nvmet_rdma_queue *queue)
488 struct nvmet_rdma_device *ndev = queue->dev;
489 int i, nr_rsps = queue->recv_queue_size * 2;
491 for (i = 0; i < nr_rsps; i++) {
492 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
494 list_del(&rsp->free_list);
495 nvmet_rdma_free_rsp(ndev, rsp);
500 static int nvmet_rdma_post_recv(struct nvmet_rdma_device *ndev,
501 struct nvmet_rdma_cmd *cmd)
505 ib_dma_sync_single_for_device(ndev->device,
506 cmd->sge[0].addr, cmd->sge[0].length,
510 ret = ib_post_srq_recv(cmd->nsrq->srq, &cmd->wr, NULL);
512 ret = ib_post_recv(cmd->queue->qp, &cmd->wr, NULL);
515 pr_err("post_recv cmd failed\n");
520 static void nvmet_rdma_process_wr_wait_list(struct nvmet_rdma_queue *queue)
522 spin_lock(&queue->rsp_wr_wait_lock);
523 while (!list_empty(&queue->rsp_wr_wait_list)) {
524 struct nvmet_rdma_rsp *rsp;
527 rsp = list_entry(queue->rsp_wr_wait_list.next,
528 struct nvmet_rdma_rsp, wait_list);
529 list_del(&rsp->wait_list);
531 spin_unlock(&queue->rsp_wr_wait_lock);
532 ret = nvmet_rdma_execute_command(rsp);
533 spin_lock(&queue->rsp_wr_wait_lock);
536 list_add(&rsp->wait_list, &queue->rsp_wr_wait_list);
540 spin_unlock(&queue->rsp_wr_wait_lock);
543 static u16 nvmet_rdma_check_pi_status(struct ib_mr *sig_mr)
545 struct ib_mr_status mr_status;
549 ret = ib_check_mr_status(sig_mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
551 pr_err("ib_check_mr_status failed, ret %d\n", ret);
552 return NVME_SC_INVALID_PI;
555 if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
556 switch (mr_status.sig_err.err_type) {
557 case IB_SIG_BAD_GUARD:
558 status = NVME_SC_GUARD_CHECK;
560 case IB_SIG_BAD_REFTAG:
561 status = NVME_SC_REFTAG_CHECK;
563 case IB_SIG_BAD_APPTAG:
564 status = NVME_SC_APPTAG_CHECK;
567 pr_err("PI error found type %d expected 0x%x vs actual 0x%x\n",
568 mr_status.sig_err.err_type,
569 mr_status.sig_err.expected,
570 mr_status.sig_err.actual);
576 static void nvmet_rdma_set_sig_domain(struct blk_integrity *bi,
577 struct nvme_command *cmd, struct ib_sig_domain *domain,
578 u16 control, u8 pi_type)
580 domain->sig_type = IB_SIG_TYPE_T10_DIF;
581 domain->sig.dif.bg_type = IB_T10DIF_CRC;
582 domain->sig.dif.pi_interval = 1 << bi->interval_exp;
583 domain->sig.dif.ref_tag = le32_to_cpu(cmd->rw.reftag);
584 if (control & NVME_RW_PRINFO_PRCHK_REF)
585 domain->sig.dif.ref_remap = true;
587 domain->sig.dif.app_tag = le16_to_cpu(cmd->rw.apptag);
588 domain->sig.dif.apptag_check_mask = le16_to_cpu(cmd->rw.appmask);
589 domain->sig.dif.app_escape = true;
590 if (pi_type == NVME_NS_DPS_PI_TYPE3)
591 domain->sig.dif.ref_escape = true;
594 static void nvmet_rdma_set_sig_attrs(struct nvmet_req *req,
595 struct ib_sig_attrs *sig_attrs)
597 struct nvme_command *cmd = req->cmd;
598 u16 control = le16_to_cpu(cmd->rw.control);
599 u8 pi_type = req->ns->pi_type;
600 struct blk_integrity *bi;
602 bi = bdev_get_integrity(req->ns->bdev);
604 memset(sig_attrs, 0, sizeof(*sig_attrs));
606 if (control & NVME_RW_PRINFO_PRACT) {
607 /* for WRITE_INSERT/READ_STRIP no wire domain */
608 sig_attrs->wire.sig_type = IB_SIG_TYPE_NONE;
609 nvmet_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control,
611 /* Clear the PRACT bit since HCA will generate/verify the PI */
612 control &= ~NVME_RW_PRINFO_PRACT;
613 cmd->rw.control = cpu_to_le16(control);
614 /* PI is added by the HW */
615 req->transfer_len += req->metadata_len;
617 /* for WRITE_PASS/READ_PASS both wire/memory domains exist */
618 nvmet_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
620 nvmet_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control,
624 if (control & NVME_RW_PRINFO_PRCHK_REF)
625 sig_attrs->check_mask |= IB_SIG_CHECK_REFTAG;
626 if (control & NVME_RW_PRINFO_PRCHK_GUARD)
627 sig_attrs->check_mask |= IB_SIG_CHECK_GUARD;
628 if (control & NVME_RW_PRINFO_PRCHK_APP)
629 sig_attrs->check_mask |= IB_SIG_CHECK_APPTAG;
632 static int nvmet_rdma_rw_ctx_init(struct nvmet_rdma_rsp *rsp, u64 addr, u32 key,
633 struct ib_sig_attrs *sig_attrs)
635 struct rdma_cm_id *cm_id = rsp->queue->cm_id;
636 struct nvmet_req *req = &rsp->req;
639 if (req->metadata_len)
640 ret = rdma_rw_ctx_signature_init(&rsp->rw, cm_id->qp,
641 cm_id->port_num, req->sg, req->sg_cnt,
642 req->metadata_sg, req->metadata_sg_cnt, sig_attrs,
643 addr, key, nvmet_data_dir(req));
645 ret = rdma_rw_ctx_init(&rsp->rw, cm_id->qp, cm_id->port_num,
646 req->sg, req->sg_cnt, 0, addr, key,
647 nvmet_data_dir(req));
652 static void nvmet_rdma_rw_ctx_destroy(struct nvmet_rdma_rsp *rsp)
654 struct rdma_cm_id *cm_id = rsp->queue->cm_id;
655 struct nvmet_req *req = &rsp->req;
657 if (req->metadata_len)
658 rdma_rw_ctx_destroy_signature(&rsp->rw, cm_id->qp,
659 cm_id->port_num, req->sg, req->sg_cnt,
660 req->metadata_sg, req->metadata_sg_cnt,
661 nvmet_data_dir(req));
663 rdma_rw_ctx_destroy(&rsp->rw, cm_id->qp, cm_id->port_num,
664 req->sg, req->sg_cnt, nvmet_data_dir(req));
667 static void nvmet_rdma_release_rsp(struct nvmet_rdma_rsp *rsp)
669 struct nvmet_rdma_queue *queue = rsp->queue;
671 atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
674 nvmet_rdma_rw_ctx_destroy(rsp);
676 if (rsp->req.sg != rsp->cmd->inline_sg)
677 nvmet_req_free_sgls(&rsp->req);
679 if (unlikely(!list_empty_careful(&queue->rsp_wr_wait_list)))
680 nvmet_rdma_process_wr_wait_list(queue);
682 nvmet_rdma_put_rsp(rsp);
685 static void nvmet_rdma_error_comp(struct nvmet_rdma_queue *queue)
687 if (queue->nvme_sq.ctrl) {
688 nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
691 * we didn't setup the controller yet in case
692 * of admin connect error, just disconnect and
695 nvmet_rdma_queue_disconnect(queue);
699 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
701 struct nvmet_rdma_rsp *rsp =
702 container_of(wc->wr_cqe, struct nvmet_rdma_rsp, send_cqe);
703 struct nvmet_rdma_queue *queue = wc->qp->qp_context;
705 nvmet_rdma_release_rsp(rsp);
707 if (unlikely(wc->status != IB_WC_SUCCESS &&
708 wc->status != IB_WC_WR_FLUSH_ERR)) {
709 pr_err("SEND for CQE 0x%p failed with status %s (%d).\n",
710 wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
711 nvmet_rdma_error_comp(queue);
715 static void nvmet_rdma_queue_response(struct nvmet_req *req)
717 struct nvmet_rdma_rsp *rsp =
718 container_of(req, struct nvmet_rdma_rsp, req);
719 struct rdma_cm_id *cm_id = rsp->queue->cm_id;
720 struct ib_send_wr *first_wr;
722 if (rsp->flags & NVMET_RDMA_REQ_INVALIDATE_RKEY) {
723 rsp->send_wr.opcode = IB_WR_SEND_WITH_INV;
724 rsp->send_wr.ex.invalidate_rkey = rsp->invalidate_rkey;
726 rsp->send_wr.opcode = IB_WR_SEND;
729 if (nvmet_rdma_need_data_out(rsp)) {
730 if (rsp->req.metadata_len)
731 first_wr = rdma_rw_ctx_wrs(&rsp->rw, cm_id->qp,
732 cm_id->port_num, &rsp->write_cqe, NULL);
734 first_wr = rdma_rw_ctx_wrs(&rsp->rw, cm_id->qp,
735 cm_id->port_num, NULL, &rsp->send_wr);
737 first_wr = &rsp->send_wr;
740 nvmet_rdma_post_recv(rsp->queue->dev, rsp->cmd);
742 ib_dma_sync_single_for_device(rsp->queue->dev->device,
743 rsp->send_sge.addr, rsp->send_sge.length,
746 if (unlikely(ib_post_send(cm_id->qp, first_wr, NULL))) {
747 pr_err("sending cmd response failed\n");
748 nvmet_rdma_release_rsp(rsp);
752 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc)
754 struct nvmet_rdma_rsp *rsp =
755 container_of(wc->wr_cqe, struct nvmet_rdma_rsp, read_cqe);
756 struct nvmet_rdma_queue *queue = wc->qp->qp_context;
759 WARN_ON(rsp->n_rdma <= 0);
760 atomic_add(rsp->n_rdma, &queue->sq_wr_avail);
763 if (unlikely(wc->status != IB_WC_SUCCESS)) {
764 nvmet_rdma_rw_ctx_destroy(rsp);
765 nvmet_req_uninit(&rsp->req);
766 nvmet_rdma_release_rsp(rsp);
767 if (wc->status != IB_WC_WR_FLUSH_ERR) {
768 pr_info("RDMA READ for CQE 0x%p failed with status %s (%d).\n",
769 wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
770 nvmet_rdma_error_comp(queue);
775 if (rsp->req.metadata_len)
776 status = nvmet_rdma_check_pi_status(rsp->rw.reg->mr);
777 nvmet_rdma_rw_ctx_destroy(rsp);
779 if (unlikely(status))
780 nvmet_req_complete(&rsp->req, status);
782 rsp->req.execute(&rsp->req);
785 static void nvmet_rdma_write_data_done(struct ib_cq *cq, struct ib_wc *wc)
787 struct nvmet_rdma_rsp *rsp =
788 container_of(wc->wr_cqe, struct nvmet_rdma_rsp, write_cqe);
789 struct nvmet_rdma_queue *queue = wc->qp->qp_context;
790 struct rdma_cm_id *cm_id = rsp->queue->cm_id;
793 if (!IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY))
796 WARN_ON(rsp->n_rdma <= 0);
797 atomic_add(rsp->n_rdma, &queue->sq_wr_avail);
800 if (unlikely(wc->status != IB_WC_SUCCESS)) {
801 nvmet_rdma_rw_ctx_destroy(rsp);
802 nvmet_req_uninit(&rsp->req);
803 nvmet_rdma_release_rsp(rsp);
804 if (wc->status != IB_WC_WR_FLUSH_ERR) {
805 pr_info("RDMA WRITE for CQE failed with status %s (%d).\n",
806 ib_wc_status_msg(wc->status), wc->status);
807 nvmet_rdma_error_comp(queue);
813 * Upon RDMA completion check the signature status
814 * - if succeeded send good NVMe response
815 * - if failed send bad NVMe response with appropriate error
817 status = nvmet_rdma_check_pi_status(rsp->rw.reg->mr);
818 if (unlikely(status))
819 rsp->req.cqe->status = cpu_to_le16(status << 1);
820 nvmet_rdma_rw_ctx_destroy(rsp);
822 if (unlikely(ib_post_send(cm_id->qp, &rsp->send_wr, NULL))) {
823 pr_err("sending cmd response failed\n");
824 nvmet_rdma_release_rsp(rsp);
828 static void nvmet_rdma_use_inline_sg(struct nvmet_rdma_rsp *rsp, u32 len,
831 int sg_count = num_pages(len);
832 struct scatterlist *sg;
835 sg = rsp->cmd->inline_sg;
836 for (i = 0; i < sg_count; i++, sg++) {
837 if (i < sg_count - 1)
842 sg->length = min_t(int, len, PAGE_SIZE - off);
848 rsp->req.sg = rsp->cmd->inline_sg;
849 rsp->req.sg_cnt = sg_count;
852 static u16 nvmet_rdma_map_sgl_inline(struct nvmet_rdma_rsp *rsp)
854 struct nvme_sgl_desc *sgl = &rsp->req.cmd->common.dptr.sgl;
855 u64 off = le64_to_cpu(sgl->addr);
856 u32 len = le32_to_cpu(sgl->length);
858 if (!nvme_is_write(rsp->req.cmd)) {
860 offsetof(struct nvme_common_command, opcode);
861 return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
864 if (off + len > rsp->queue->dev->inline_data_size) {
865 pr_err("invalid inline data offset!\n");
866 return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
869 /* no data command? */
873 nvmet_rdma_use_inline_sg(rsp, len, off);
874 rsp->flags |= NVMET_RDMA_REQ_INLINE_DATA;
875 rsp->req.transfer_len += len;
879 static u16 nvmet_rdma_map_sgl_keyed(struct nvmet_rdma_rsp *rsp,
880 struct nvme_keyed_sgl_desc *sgl, bool invalidate)
882 u64 addr = le64_to_cpu(sgl->addr);
883 u32 key = get_unaligned_le32(sgl->key);
884 struct ib_sig_attrs sig_attrs;
887 rsp->req.transfer_len = get_unaligned_le24(sgl->length);
889 /* no data command? */
890 if (!rsp->req.transfer_len)
893 if (rsp->req.metadata_len)
894 nvmet_rdma_set_sig_attrs(&rsp->req, &sig_attrs);
896 ret = nvmet_req_alloc_sgls(&rsp->req);
897 if (unlikely(ret < 0))
900 ret = nvmet_rdma_rw_ctx_init(rsp, addr, key, &sig_attrs);
901 if (unlikely(ret < 0))
906 rsp->invalidate_rkey = key;
907 rsp->flags |= NVMET_RDMA_REQ_INVALIDATE_RKEY;
913 rsp->req.transfer_len = 0;
914 return NVME_SC_INTERNAL;
917 static u16 nvmet_rdma_map_sgl(struct nvmet_rdma_rsp *rsp)
919 struct nvme_keyed_sgl_desc *sgl = &rsp->req.cmd->common.dptr.ksgl;
921 switch (sgl->type >> 4) {
922 case NVME_SGL_FMT_DATA_DESC:
923 switch (sgl->type & 0xf) {
924 case NVME_SGL_FMT_OFFSET:
925 return nvmet_rdma_map_sgl_inline(rsp);
927 pr_err("invalid SGL subtype: %#x\n", sgl->type);
929 offsetof(struct nvme_common_command, dptr);
930 return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
932 case NVME_KEY_SGL_FMT_DATA_DESC:
933 switch (sgl->type & 0xf) {
934 case NVME_SGL_FMT_ADDRESS | NVME_SGL_FMT_INVALIDATE:
935 return nvmet_rdma_map_sgl_keyed(rsp, sgl, true);
936 case NVME_SGL_FMT_ADDRESS:
937 return nvmet_rdma_map_sgl_keyed(rsp, sgl, false);
939 pr_err("invalid SGL subtype: %#x\n", sgl->type);
941 offsetof(struct nvme_common_command, dptr);
942 return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
945 pr_err("invalid SGL type: %#x\n", sgl->type);
946 rsp->req.error_loc = offsetof(struct nvme_common_command, dptr);
947 return NVME_SC_SGL_INVALID_TYPE | NVME_SC_DNR;
951 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp)
953 struct nvmet_rdma_queue *queue = rsp->queue;
955 if (unlikely(atomic_sub_return(1 + rsp->n_rdma,
956 &queue->sq_wr_avail) < 0)) {
957 pr_debug("IB send queue full (needed %d): queue %u cntlid %u\n",
958 1 + rsp->n_rdma, queue->idx,
959 queue->nvme_sq.ctrl->cntlid);
960 atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
964 if (nvmet_rdma_need_data_in(rsp)) {
965 if (rdma_rw_ctx_post(&rsp->rw, queue->qp,
966 queue->cm_id->port_num, &rsp->read_cqe, NULL))
967 nvmet_req_complete(&rsp->req, NVME_SC_DATA_XFER_ERROR);
969 rsp->req.execute(&rsp->req);
975 static void nvmet_rdma_handle_command(struct nvmet_rdma_queue *queue,
976 struct nvmet_rdma_rsp *cmd)
980 ib_dma_sync_single_for_cpu(queue->dev->device,
981 cmd->cmd->sge[0].addr, cmd->cmd->sge[0].length,
983 ib_dma_sync_single_for_cpu(queue->dev->device,
984 cmd->send_sge.addr, cmd->send_sge.length,
987 if (!nvmet_req_init(&cmd->req, &queue->nvme_cq,
988 &queue->nvme_sq, &nvmet_rdma_ops))
991 status = nvmet_rdma_map_sgl(cmd);
995 if (unlikely(!nvmet_rdma_execute_command(cmd))) {
996 spin_lock(&queue->rsp_wr_wait_lock);
997 list_add_tail(&cmd->wait_list, &queue->rsp_wr_wait_list);
998 spin_unlock(&queue->rsp_wr_wait_lock);
1004 nvmet_req_complete(&cmd->req, status);
1007 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1009 struct nvmet_rdma_cmd *cmd =
1010 container_of(wc->wr_cqe, struct nvmet_rdma_cmd, cqe);
1011 struct nvmet_rdma_queue *queue = wc->qp->qp_context;
1012 struct nvmet_rdma_rsp *rsp;
1014 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1015 if (wc->status != IB_WC_WR_FLUSH_ERR) {
1016 pr_err("RECV for CQE 0x%p failed with status %s (%d)\n",
1017 wc->wr_cqe, ib_wc_status_msg(wc->status),
1019 nvmet_rdma_error_comp(queue);
1024 if (unlikely(wc->byte_len < sizeof(struct nvme_command))) {
1025 pr_err("Ctrl Fatal Error: capsule size less than 64 bytes\n");
1026 nvmet_rdma_error_comp(queue);
1031 rsp = nvmet_rdma_get_rsp(queue);
1032 if (unlikely(!rsp)) {
1034 * we get here only under memory pressure,
1035 * silently drop and have the host retry
1036 * as we can't even fail it.
1038 nvmet_rdma_post_recv(queue->dev, cmd);
1044 rsp->req.cmd = cmd->nvme_cmd;
1045 rsp->req.port = queue->port;
1048 if (unlikely(queue->state != NVMET_RDMA_Q_LIVE)) {
1049 unsigned long flags;
1051 spin_lock_irqsave(&queue->state_lock, flags);
1052 if (queue->state == NVMET_RDMA_Q_CONNECTING)
1053 list_add_tail(&rsp->wait_list, &queue->rsp_wait_list);
1055 nvmet_rdma_put_rsp(rsp);
1056 spin_unlock_irqrestore(&queue->state_lock, flags);
1060 nvmet_rdma_handle_command(queue, rsp);
1063 static void nvmet_rdma_destroy_srq(struct nvmet_rdma_srq *nsrq)
1065 nvmet_rdma_free_cmds(nsrq->ndev, nsrq->cmds, nsrq->ndev->srq_size,
1067 ib_destroy_srq(nsrq->srq);
1072 static void nvmet_rdma_destroy_srqs(struct nvmet_rdma_device *ndev)
1079 for (i = 0; i < ndev->srq_count; i++)
1080 nvmet_rdma_destroy_srq(ndev->srqs[i]);
1085 static struct nvmet_rdma_srq *
1086 nvmet_rdma_init_srq(struct nvmet_rdma_device *ndev)
1088 struct ib_srq_init_attr srq_attr = { NULL, };
1089 size_t srq_size = ndev->srq_size;
1090 struct nvmet_rdma_srq *nsrq;
1094 nsrq = kzalloc(sizeof(*nsrq), GFP_KERNEL);
1096 return ERR_PTR(-ENOMEM);
1098 srq_attr.attr.max_wr = srq_size;
1099 srq_attr.attr.max_sge = 1 + ndev->inline_page_count;
1100 srq_attr.attr.srq_limit = 0;
1101 srq_attr.srq_type = IB_SRQT_BASIC;
1102 srq = ib_create_srq(ndev->pd, &srq_attr);
1108 nsrq->cmds = nvmet_rdma_alloc_cmds(ndev, srq_size, false);
1109 if (IS_ERR(nsrq->cmds)) {
1110 ret = PTR_ERR(nsrq->cmds);
1111 goto out_destroy_srq;
1117 for (i = 0; i < srq_size; i++) {
1118 nsrq->cmds[i].nsrq = nsrq;
1119 ret = nvmet_rdma_post_recv(ndev, &nsrq->cmds[i]);
1127 nvmet_rdma_free_cmds(ndev, nsrq->cmds, srq_size, false);
1129 ib_destroy_srq(srq);
1132 return ERR_PTR(ret);
1135 static int nvmet_rdma_init_srqs(struct nvmet_rdma_device *ndev)
1139 if (!ndev->device->attrs.max_srq_wr || !ndev->device->attrs.max_srq) {
1141 * If SRQs aren't supported we just go ahead and use normal
1142 * non-shared receive queues.
1144 pr_info("SRQ requested but not supported.\n");
1148 ndev->srq_size = min(ndev->device->attrs.max_srq_wr,
1149 nvmet_rdma_srq_size);
1150 ndev->srq_count = min(ndev->device->num_comp_vectors,
1151 ndev->device->attrs.max_srq);
1153 ndev->srqs = kcalloc(ndev->srq_count, sizeof(*ndev->srqs), GFP_KERNEL);
1157 for (i = 0; i < ndev->srq_count; i++) {
1158 ndev->srqs[i] = nvmet_rdma_init_srq(ndev);
1159 if (IS_ERR(ndev->srqs[i])) {
1160 ret = PTR_ERR(ndev->srqs[i]);
1169 nvmet_rdma_destroy_srq(ndev->srqs[i]);
1174 static void nvmet_rdma_free_dev(struct kref *ref)
1176 struct nvmet_rdma_device *ndev =
1177 container_of(ref, struct nvmet_rdma_device, ref);
1179 mutex_lock(&device_list_mutex);
1180 list_del(&ndev->entry);
1181 mutex_unlock(&device_list_mutex);
1183 nvmet_rdma_destroy_srqs(ndev);
1184 ib_dealloc_pd(ndev->pd);
1189 static struct nvmet_rdma_device *
1190 nvmet_rdma_find_get_device(struct rdma_cm_id *cm_id)
1192 struct nvmet_rdma_port *port = cm_id->context;
1193 struct nvmet_port *nport = port->nport;
1194 struct nvmet_rdma_device *ndev;
1195 int inline_page_count;
1196 int inline_sge_count;
1199 mutex_lock(&device_list_mutex);
1200 list_for_each_entry(ndev, &device_list, entry) {
1201 if (ndev->device->node_guid == cm_id->device->node_guid &&
1202 kref_get_unless_zero(&ndev->ref))
1206 ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
1210 inline_page_count = num_pages(nport->inline_data_size);
1211 inline_sge_count = max(cm_id->device->attrs.max_sge_rd,
1212 cm_id->device->attrs.max_recv_sge) - 1;
1213 if (inline_page_count > inline_sge_count) {
1214 pr_warn("inline_data_size %d cannot be supported by device %s. Reducing to %lu.\n",
1215 nport->inline_data_size, cm_id->device->name,
1216 inline_sge_count * PAGE_SIZE);
1217 nport->inline_data_size = inline_sge_count * PAGE_SIZE;
1218 inline_page_count = inline_sge_count;
1220 ndev->inline_data_size = nport->inline_data_size;
1221 ndev->inline_page_count = inline_page_count;
1223 if (nport->pi_enable && !(cm_id->device->attrs.device_cap_flags &
1224 IB_DEVICE_INTEGRITY_HANDOVER)) {
1225 pr_warn("T10-PI is not supported by device %s. Disabling it\n",
1226 cm_id->device->name);
1227 nport->pi_enable = false;
1230 ndev->device = cm_id->device;
1231 kref_init(&ndev->ref);
1233 ndev->pd = ib_alloc_pd(ndev->device, 0);
1234 if (IS_ERR(ndev->pd))
1237 if (nvmet_rdma_use_srq) {
1238 ret = nvmet_rdma_init_srqs(ndev);
1243 list_add(&ndev->entry, &device_list);
1245 mutex_unlock(&device_list_mutex);
1246 pr_debug("added %s.\n", ndev->device->name);
1250 ib_dealloc_pd(ndev->pd);
1254 mutex_unlock(&device_list_mutex);
1258 static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue)
1260 struct ib_qp_init_attr qp_attr;
1261 struct nvmet_rdma_device *ndev = queue->dev;
1262 int nr_cqe, ret, i, factor;
1265 * Reserve CQ slots for RECV + RDMA_READ/RDMA_WRITE + RDMA_SEND.
1267 nr_cqe = queue->recv_queue_size + 2 * queue->send_queue_size;
1269 queue->cq = ib_cq_pool_get(ndev->device, nr_cqe + 1,
1270 queue->comp_vector, IB_POLL_WORKQUEUE);
1271 if (IS_ERR(queue->cq)) {
1272 ret = PTR_ERR(queue->cq);
1273 pr_err("failed to create CQ cqe= %d ret= %d\n",
1278 memset(&qp_attr, 0, sizeof(qp_attr));
1279 qp_attr.qp_context = queue;
1280 qp_attr.event_handler = nvmet_rdma_qp_event;
1281 qp_attr.send_cq = queue->cq;
1282 qp_attr.recv_cq = queue->cq;
1283 qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
1284 qp_attr.qp_type = IB_QPT_RC;
1286 qp_attr.cap.max_send_wr = queue->send_queue_size + 1;
1287 factor = rdma_rw_mr_factor(ndev->device, queue->cm_id->port_num,
1288 1 << NVMET_RDMA_MAX_MDTS);
1289 qp_attr.cap.max_rdma_ctxs = queue->send_queue_size * factor;
1290 qp_attr.cap.max_send_sge = max(ndev->device->attrs.max_sge_rd,
1291 ndev->device->attrs.max_send_sge);
1294 qp_attr.srq = queue->nsrq->srq;
1297 qp_attr.cap.max_recv_wr = 1 + queue->recv_queue_size;
1298 qp_attr.cap.max_recv_sge = 1 + ndev->inline_page_count;
1301 if (queue->port->pi_enable && queue->host_qid)
1302 qp_attr.create_flags |= IB_QP_CREATE_INTEGRITY_EN;
1304 ret = rdma_create_qp(queue->cm_id, ndev->pd, &qp_attr);
1306 pr_err("failed to create_qp ret= %d\n", ret);
1307 goto err_destroy_cq;
1309 queue->qp = queue->cm_id->qp;
1311 atomic_set(&queue->sq_wr_avail, qp_attr.cap.max_send_wr);
1313 pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
1314 __func__, queue->cq->cqe, qp_attr.cap.max_send_sge,
1315 qp_attr.cap.max_send_wr, queue->cm_id);
1318 for (i = 0; i < queue->recv_queue_size; i++) {
1319 queue->cmds[i].queue = queue;
1320 ret = nvmet_rdma_post_recv(ndev, &queue->cmds[i]);
1322 goto err_destroy_qp;
1330 rdma_destroy_qp(queue->cm_id);
1332 ib_cq_pool_put(queue->cq, nr_cqe + 1);
1336 static void nvmet_rdma_destroy_queue_ib(struct nvmet_rdma_queue *queue)
1338 ib_drain_qp(queue->qp);
1340 rdma_destroy_id(queue->cm_id);
1341 ib_destroy_qp(queue->qp);
1342 ib_cq_pool_put(queue->cq, queue->recv_queue_size + 2 *
1343 queue->send_queue_size + 1);
1346 static void nvmet_rdma_free_queue(struct nvmet_rdma_queue *queue)
1348 pr_debug("freeing queue %d\n", queue->idx);
1350 nvmet_sq_destroy(&queue->nvme_sq);
1352 nvmet_rdma_destroy_queue_ib(queue);
1354 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1355 queue->recv_queue_size,
1358 nvmet_rdma_free_rsps(queue);
1359 ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1363 static void nvmet_rdma_release_queue_work(struct work_struct *w)
1365 struct nvmet_rdma_queue *queue =
1366 container_of(w, struct nvmet_rdma_queue, release_work);
1367 struct nvmet_rdma_device *dev = queue->dev;
1369 nvmet_rdma_free_queue(queue);
1371 kref_put(&dev->ref, nvmet_rdma_free_dev);
1375 nvmet_rdma_parse_cm_connect_req(struct rdma_conn_param *conn,
1376 struct nvmet_rdma_queue *queue)
1378 struct nvme_rdma_cm_req *req;
1380 req = (struct nvme_rdma_cm_req *)conn->private_data;
1381 if (!req || conn->private_data_len == 0)
1382 return NVME_RDMA_CM_INVALID_LEN;
1384 if (le16_to_cpu(req->recfmt) != NVME_RDMA_CM_FMT_1_0)
1385 return NVME_RDMA_CM_INVALID_RECFMT;
1387 queue->host_qid = le16_to_cpu(req->qid);
1390 * req->hsqsize corresponds to our recv queue size plus 1
1391 * req->hrqsize corresponds to our send queue size
1393 queue->recv_queue_size = le16_to_cpu(req->hsqsize) + 1;
1394 queue->send_queue_size = le16_to_cpu(req->hrqsize);
1396 if (!queue->host_qid && queue->recv_queue_size > NVME_AQ_DEPTH)
1397 return NVME_RDMA_CM_INVALID_HSQSIZE;
1399 /* XXX: Should we enforce some kind of max for IO queues? */
1404 static int nvmet_rdma_cm_reject(struct rdma_cm_id *cm_id,
1405 enum nvme_rdma_cm_status status)
1407 struct nvme_rdma_cm_rej rej;
1409 pr_debug("rejecting connect request: status %d (%s)\n",
1410 status, nvme_rdma_cm_msg(status));
1412 rej.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1413 rej.sts = cpu_to_le16(status);
1415 return rdma_reject(cm_id, (void *)&rej, sizeof(rej),
1416 IB_CM_REJ_CONSUMER_DEFINED);
1419 static struct nvmet_rdma_queue *
1420 nvmet_rdma_alloc_queue(struct nvmet_rdma_device *ndev,
1421 struct rdma_cm_id *cm_id,
1422 struct rdma_cm_event *event)
1424 struct nvmet_rdma_port *port = cm_id->context;
1425 struct nvmet_rdma_queue *queue;
1428 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1430 ret = NVME_RDMA_CM_NO_RSC;
1434 ret = nvmet_sq_init(&queue->nvme_sq);
1436 ret = NVME_RDMA_CM_NO_RSC;
1437 goto out_free_queue;
1440 ret = nvmet_rdma_parse_cm_connect_req(&event->param.conn, queue);
1442 goto out_destroy_sq;
1445 * Schedules the actual release because calling rdma_destroy_id from
1446 * inside a CM callback would trigger a deadlock. (great API design..)
1448 INIT_WORK(&queue->release_work, nvmet_rdma_release_queue_work);
1450 queue->cm_id = cm_id;
1451 queue->port = port->nport;
1453 spin_lock_init(&queue->state_lock);
1454 queue->state = NVMET_RDMA_Q_CONNECTING;
1455 INIT_LIST_HEAD(&queue->rsp_wait_list);
1456 INIT_LIST_HEAD(&queue->rsp_wr_wait_list);
1457 spin_lock_init(&queue->rsp_wr_wait_lock);
1458 INIT_LIST_HEAD(&queue->free_rsps);
1459 spin_lock_init(&queue->rsps_lock);
1460 INIT_LIST_HEAD(&queue->queue_list);
1462 queue->idx = ida_simple_get(&nvmet_rdma_queue_ida, 0, 0, GFP_KERNEL);
1463 if (queue->idx < 0) {
1464 ret = NVME_RDMA_CM_NO_RSC;
1465 goto out_destroy_sq;
1469 * Spread the io queues across completion vectors,
1470 * but still keep all admin queues on vector 0.
1472 queue->comp_vector = !queue->host_qid ? 0 :
1473 queue->idx % ndev->device->num_comp_vectors;
1476 ret = nvmet_rdma_alloc_rsps(queue);
1478 ret = NVME_RDMA_CM_NO_RSC;
1479 goto out_ida_remove;
1483 queue->nsrq = ndev->srqs[queue->comp_vector % ndev->srq_count];
1485 queue->cmds = nvmet_rdma_alloc_cmds(ndev,
1486 queue->recv_queue_size,
1488 if (IS_ERR(queue->cmds)) {
1489 ret = NVME_RDMA_CM_NO_RSC;
1490 goto out_free_responses;
1494 ret = nvmet_rdma_create_queue_ib(queue);
1496 pr_err("%s: creating RDMA queue failed (%d).\n",
1498 ret = NVME_RDMA_CM_NO_RSC;
1506 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1507 queue->recv_queue_size,
1511 nvmet_rdma_free_rsps(queue);
1513 ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1515 nvmet_sq_destroy(&queue->nvme_sq);
1519 nvmet_rdma_cm_reject(cm_id, ret);
1523 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv)
1525 struct nvmet_rdma_queue *queue = priv;
1527 switch (event->event) {
1528 case IB_EVENT_COMM_EST:
1529 rdma_notify(queue->cm_id, event->event);
1531 case IB_EVENT_QP_LAST_WQE_REACHED:
1532 pr_debug("received last WQE reached event for queue=0x%p\n",
1536 pr_err("received IB QP event: %s (%d)\n",
1537 ib_event_msg(event->event), event->event);
1542 static int nvmet_rdma_cm_accept(struct rdma_cm_id *cm_id,
1543 struct nvmet_rdma_queue *queue,
1544 struct rdma_conn_param *p)
1546 struct rdma_conn_param param = { };
1547 struct nvme_rdma_cm_rep priv = { };
1550 param.rnr_retry_count = 7;
1551 param.flow_control = 1;
1552 param.initiator_depth = min_t(u8, p->initiator_depth,
1553 queue->dev->device->attrs.max_qp_init_rd_atom);
1554 param.private_data = &priv;
1555 param.private_data_len = sizeof(priv);
1556 priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1557 priv.crqsize = cpu_to_le16(queue->recv_queue_size);
1559 ret = rdma_accept(cm_id, ¶m);
1561 pr_err("rdma_accept failed (error code = %d)\n", ret);
1566 static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
1567 struct rdma_cm_event *event)
1569 struct nvmet_rdma_device *ndev;
1570 struct nvmet_rdma_queue *queue;
1573 ndev = nvmet_rdma_find_get_device(cm_id);
1575 nvmet_rdma_cm_reject(cm_id, NVME_RDMA_CM_NO_RSC);
1576 return -ECONNREFUSED;
1579 queue = nvmet_rdma_alloc_queue(ndev, cm_id, event);
1585 if (queue->host_qid == 0) {
1586 /* Let inflight controller teardown complete */
1587 flush_scheduled_work();
1590 ret = nvmet_rdma_cm_accept(cm_id, queue, &event->param.conn);
1593 * Don't destroy the cm_id in free path, as we implicitly
1594 * destroy the cm_id here with non-zero ret code.
1596 queue->cm_id = NULL;
1600 mutex_lock(&nvmet_rdma_queue_mutex);
1601 list_add_tail(&queue->queue_list, &nvmet_rdma_queue_list);
1602 mutex_unlock(&nvmet_rdma_queue_mutex);
1607 nvmet_rdma_free_queue(queue);
1609 kref_put(&ndev->ref, nvmet_rdma_free_dev);
1614 static void nvmet_rdma_queue_established(struct nvmet_rdma_queue *queue)
1616 unsigned long flags;
1618 spin_lock_irqsave(&queue->state_lock, flags);
1619 if (queue->state != NVMET_RDMA_Q_CONNECTING) {
1620 pr_warn("trying to establish a connected queue\n");
1623 queue->state = NVMET_RDMA_Q_LIVE;
1625 while (!list_empty(&queue->rsp_wait_list)) {
1626 struct nvmet_rdma_rsp *cmd;
1628 cmd = list_first_entry(&queue->rsp_wait_list,
1629 struct nvmet_rdma_rsp, wait_list);
1630 list_del(&cmd->wait_list);
1632 spin_unlock_irqrestore(&queue->state_lock, flags);
1633 nvmet_rdma_handle_command(queue, cmd);
1634 spin_lock_irqsave(&queue->state_lock, flags);
1638 spin_unlock_irqrestore(&queue->state_lock, flags);
1641 static void __nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1643 bool disconnect = false;
1644 unsigned long flags;
1646 pr_debug("cm_id= %p queue->state= %d\n", queue->cm_id, queue->state);
1648 spin_lock_irqsave(&queue->state_lock, flags);
1649 switch (queue->state) {
1650 case NVMET_RDMA_Q_CONNECTING:
1651 while (!list_empty(&queue->rsp_wait_list)) {
1652 struct nvmet_rdma_rsp *rsp;
1654 rsp = list_first_entry(&queue->rsp_wait_list,
1655 struct nvmet_rdma_rsp,
1657 list_del(&rsp->wait_list);
1658 nvmet_rdma_put_rsp(rsp);
1661 case NVMET_RDMA_Q_LIVE:
1662 queue->state = NVMET_RDMA_Q_DISCONNECTING;
1665 case NVMET_RDMA_Q_DISCONNECTING:
1668 spin_unlock_irqrestore(&queue->state_lock, flags);
1671 rdma_disconnect(queue->cm_id);
1672 schedule_work(&queue->release_work);
1676 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1678 bool disconnect = false;
1680 mutex_lock(&nvmet_rdma_queue_mutex);
1681 if (!list_empty(&queue->queue_list)) {
1682 list_del_init(&queue->queue_list);
1685 mutex_unlock(&nvmet_rdma_queue_mutex);
1688 __nvmet_rdma_queue_disconnect(queue);
1691 static void nvmet_rdma_queue_connect_fail(struct rdma_cm_id *cm_id,
1692 struct nvmet_rdma_queue *queue)
1694 WARN_ON_ONCE(queue->state != NVMET_RDMA_Q_CONNECTING);
1696 mutex_lock(&nvmet_rdma_queue_mutex);
1697 if (!list_empty(&queue->queue_list))
1698 list_del_init(&queue->queue_list);
1699 mutex_unlock(&nvmet_rdma_queue_mutex);
1701 pr_err("failed to connect queue %d\n", queue->idx);
1702 schedule_work(&queue->release_work);
1706 * nvme_rdma_device_removal() - Handle RDMA device removal
1707 * @cm_id: rdma_cm id, used for nvmet port
1708 * @queue: nvmet rdma queue (cm id qp_context)
1710 * DEVICE_REMOVAL event notifies us that the RDMA device is about
1711 * to unplug. Note that this event can be generated on a normal
1712 * queue cm_id and/or a device bound listener cm_id (where in this
1713 * case queue will be null).
1715 * We registered an ib_client to handle device removal for queues,
1716 * so we only need to handle the listening port cm_ids. In this case
1717 * we nullify the priv to prevent double cm_id destruction and destroying
1718 * the cm_id implicitely by returning a non-zero rc to the callout.
1720 static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
1721 struct nvmet_rdma_queue *queue)
1723 struct nvmet_rdma_port *port;
1727 * This is a queue cm_id. we have registered
1728 * an ib_client to handle queues removal
1729 * so don't interfear and just return.
1734 port = cm_id->context;
1737 * This is a listener cm_id. Make sure that
1738 * future remove_port won't invoke a double
1739 * cm_id destroy. use atomic xchg to make sure
1740 * we don't compete with remove_port.
1742 if (xchg(&port->cm_id, NULL) != cm_id)
1746 * We need to return 1 so that the core will destroy
1747 * it's own ID. What a great API design..
1752 static int nvmet_rdma_cm_handler(struct rdma_cm_id *cm_id,
1753 struct rdma_cm_event *event)
1755 struct nvmet_rdma_queue *queue = NULL;
1759 queue = cm_id->qp->qp_context;
1761 pr_debug("%s (%d): status %d id %p\n",
1762 rdma_event_msg(event->event), event->event,
1763 event->status, cm_id);
1765 switch (event->event) {
1766 case RDMA_CM_EVENT_CONNECT_REQUEST:
1767 ret = nvmet_rdma_queue_connect(cm_id, event);
1769 case RDMA_CM_EVENT_ESTABLISHED:
1770 nvmet_rdma_queue_established(queue);
1772 case RDMA_CM_EVENT_ADDR_CHANGE:
1774 struct nvmet_rdma_port *port = cm_id->context;
1776 schedule_delayed_work(&port->repair_work, 0);
1780 case RDMA_CM_EVENT_DISCONNECTED:
1781 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1782 nvmet_rdma_queue_disconnect(queue);
1784 case RDMA_CM_EVENT_DEVICE_REMOVAL:
1785 ret = nvmet_rdma_device_removal(cm_id, queue);
1787 case RDMA_CM_EVENT_REJECTED:
1788 pr_debug("Connection rejected: %s\n",
1789 rdma_reject_msg(cm_id, event->status));
1791 case RDMA_CM_EVENT_UNREACHABLE:
1792 case RDMA_CM_EVENT_CONNECT_ERROR:
1793 nvmet_rdma_queue_connect_fail(cm_id, queue);
1796 pr_err("received unrecognized RDMA CM event %d\n",
1804 static void nvmet_rdma_delete_ctrl(struct nvmet_ctrl *ctrl)
1806 struct nvmet_rdma_queue *queue;
1809 mutex_lock(&nvmet_rdma_queue_mutex);
1810 list_for_each_entry(queue, &nvmet_rdma_queue_list, queue_list) {
1811 if (queue->nvme_sq.ctrl == ctrl) {
1812 list_del_init(&queue->queue_list);
1813 mutex_unlock(&nvmet_rdma_queue_mutex);
1815 __nvmet_rdma_queue_disconnect(queue);
1819 mutex_unlock(&nvmet_rdma_queue_mutex);
1822 static void nvmet_rdma_disable_port(struct nvmet_rdma_port *port)
1824 struct rdma_cm_id *cm_id = xchg(&port->cm_id, NULL);
1827 rdma_destroy_id(cm_id);
1830 static int nvmet_rdma_enable_port(struct nvmet_rdma_port *port)
1832 struct sockaddr *addr = (struct sockaddr *)&port->addr;
1833 struct rdma_cm_id *cm_id;
1836 cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
1837 RDMA_PS_TCP, IB_QPT_RC);
1838 if (IS_ERR(cm_id)) {
1839 pr_err("CM ID creation failed\n");
1840 return PTR_ERR(cm_id);
1844 * Allow both IPv4 and IPv6 sockets to bind a single port
1847 ret = rdma_set_afonly(cm_id, 1);
1849 pr_err("rdma_set_afonly failed (%d)\n", ret);
1850 goto out_destroy_id;
1853 ret = rdma_bind_addr(cm_id, addr);
1855 pr_err("binding CM ID to %pISpcs failed (%d)\n", addr, ret);
1856 goto out_destroy_id;
1859 ret = rdma_listen(cm_id, 128);
1861 pr_err("listening to %pISpcs failed (%d)\n", addr, ret);
1862 goto out_destroy_id;
1865 port->cm_id = cm_id;
1869 rdma_destroy_id(cm_id);
1873 static void nvmet_rdma_repair_port_work(struct work_struct *w)
1875 struct nvmet_rdma_port *port = container_of(to_delayed_work(w),
1876 struct nvmet_rdma_port, repair_work);
1879 nvmet_rdma_disable_port(port);
1880 ret = nvmet_rdma_enable_port(port);
1882 schedule_delayed_work(&port->repair_work, 5 * HZ);
1885 static int nvmet_rdma_add_port(struct nvmet_port *nport)
1887 struct nvmet_rdma_port *port;
1888 __kernel_sa_family_t af;
1891 port = kzalloc(sizeof(*port), GFP_KERNEL);
1896 port->nport = nport;
1897 INIT_DELAYED_WORK(&port->repair_work, nvmet_rdma_repair_port_work);
1899 switch (nport->disc_addr.adrfam) {
1900 case NVMF_ADDR_FAMILY_IP4:
1903 case NVMF_ADDR_FAMILY_IP6:
1907 pr_err("address family %d not supported\n",
1908 nport->disc_addr.adrfam);
1913 if (nport->inline_data_size < 0) {
1914 nport->inline_data_size = NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
1915 } else if (nport->inline_data_size > NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
1916 pr_warn("inline_data_size %u is too large, reducing to %u\n",
1917 nport->inline_data_size,
1918 NVMET_RDMA_MAX_INLINE_DATA_SIZE);
1919 nport->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
1922 ret = inet_pton_with_scope(&init_net, af, nport->disc_addr.traddr,
1923 nport->disc_addr.trsvcid, &port->addr);
1925 pr_err("malformed ip/port passed: %s:%s\n",
1926 nport->disc_addr.traddr, nport->disc_addr.trsvcid);
1930 ret = nvmet_rdma_enable_port(port);
1934 pr_info("enabling port %d (%pISpcs)\n",
1935 le16_to_cpu(nport->disc_addr.portid),
1936 (struct sockaddr *)&port->addr);
1945 static void nvmet_rdma_remove_port(struct nvmet_port *nport)
1947 struct nvmet_rdma_port *port = nport->priv;
1949 cancel_delayed_work_sync(&port->repair_work);
1950 nvmet_rdma_disable_port(port);
1954 static void nvmet_rdma_disc_port_addr(struct nvmet_req *req,
1955 struct nvmet_port *nport, char *traddr)
1957 struct nvmet_rdma_port *port = nport->priv;
1958 struct rdma_cm_id *cm_id = port->cm_id;
1960 if (inet_addr_is_any((struct sockaddr *)&cm_id->route.addr.src_addr)) {
1961 struct nvmet_rdma_rsp *rsp =
1962 container_of(req, struct nvmet_rdma_rsp, req);
1963 struct rdma_cm_id *req_cm_id = rsp->queue->cm_id;
1964 struct sockaddr *addr = (void *)&req_cm_id->route.addr.src_addr;
1966 sprintf(traddr, "%pISc", addr);
1968 memcpy(traddr, nport->disc_addr.traddr, NVMF_TRADDR_SIZE);
1972 static u8 nvmet_rdma_get_mdts(const struct nvmet_ctrl *ctrl)
1974 if (ctrl->pi_support)
1975 return NVMET_RDMA_MAX_METADATA_MDTS;
1976 return NVMET_RDMA_MAX_MDTS;
1979 static const struct nvmet_fabrics_ops nvmet_rdma_ops = {
1980 .owner = THIS_MODULE,
1981 .type = NVMF_TRTYPE_RDMA,
1983 .flags = NVMF_KEYED_SGLS | NVMF_METADATA_SUPPORTED,
1984 .add_port = nvmet_rdma_add_port,
1985 .remove_port = nvmet_rdma_remove_port,
1986 .queue_response = nvmet_rdma_queue_response,
1987 .delete_ctrl = nvmet_rdma_delete_ctrl,
1988 .disc_traddr = nvmet_rdma_disc_port_addr,
1989 .get_mdts = nvmet_rdma_get_mdts,
1992 static void nvmet_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1994 struct nvmet_rdma_queue *queue, *tmp;
1995 struct nvmet_rdma_device *ndev;
1998 mutex_lock(&device_list_mutex);
1999 list_for_each_entry(ndev, &device_list, entry) {
2000 if (ndev->device == ib_device) {
2005 mutex_unlock(&device_list_mutex);
2011 * IB Device that is used by nvmet controllers is being removed,
2012 * delete all queues using this device.
2014 mutex_lock(&nvmet_rdma_queue_mutex);
2015 list_for_each_entry_safe(queue, tmp, &nvmet_rdma_queue_list,
2017 if (queue->dev->device != ib_device)
2020 pr_info("Removing queue %d\n", queue->idx);
2021 list_del_init(&queue->queue_list);
2022 __nvmet_rdma_queue_disconnect(queue);
2024 mutex_unlock(&nvmet_rdma_queue_mutex);
2026 flush_scheduled_work();
2029 static struct ib_client nvmet_rdma_ib_client = {
2030 .name = "nvmet_rdma",
2031 .remove = nvmet_rdma_remove_one
2034 static int __init nvmet_rdma_init(void)
2038 ret = ib_register_client(&nvmet_rdma_ib_client);
2042 ret = nvmet_register_transport(&nvmet_rdma_ops);
2049 ib_unregister_client(&nvmet_rdma_ib_client);
2053 static void __exit nvmet_rdma_exit(void)
2055 nvmet_unregister_transport(&nvmet_rdma_ops);
2056 ib_unregister_client(&nvmet_rdma_ib_client);
2057 WARN_ON_ONCE(!list_empty(&nvmet_rdma_queue_list));
2058 ida_destroy(&nvmet_rdma_queue_ida);
2061 module_init(nvmet_rdma_init);
2062 module_exit(nvmet_rdma_exit);
2064 MODULE_LICENSE("GPL v2");
2065 MODULE_ALIAS("nvmet-transport-1"); /* 1 == NVMF_TRTYPE_RDMA */