1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
3 * Copyright 2018-2020 Amazon.com, Inc. or its affiliates. All rights reserved.
6 #include <linux/vmalloc.h>
7 #include <linux/log2.h>
9 #include <rdma/ib_addr.h>
10 #include <rdma/ib_umem.h>
11 #include <rdma/ib_user_verbs.h>
12 #include <rdma/ib_verbs.h>
13 #include <rdma/uverbs_ioctl.h>
18 EFA_MMAP_DMA_PAGE = 0,
23 #define EFA_AENQ_ENABLED_GROUPS \
24 (BIT(EFA_ADMIN_FATAL_ERROR) | BIT(EFA_ADMIN_WARNING) | \
25 BIT(EFA_ADMIN_NOTIFICATION) | BIT(EFA_ADMIN_KEEP_ALIVE))
27 struct efa_user_mmap_entry {
28 struct rdma_user_mmap_entry rdma_entry;
33 #define EFA_DEFINE_STATS(op) \
34 op(EFA_TX_BYTES, "tx_bytes") \
35 op(EFA_TX_PKTS, "tx_pkts") \
36 op(EFA_RX_BYTES, "rx_bytes") \
37 op(EFA_RX_PKTS, "rx_pkts") \
38 op(EFA_RX_DROPS, "rx_drops") \
39 op(EFA_SEND_BYTES, "send_bytes") \
40 op(EFA_SEND_WRS, "send_wrs") \
41 op(EFA_RECV_BYTES, "recv_bytes") \
42 op(EFA_RECV_WRS, "recv_wrs") \
43 op(EFA_RDMA_READ_WRS, "rdma_read_wrs") \
44 op(EFA_RDMA_READ_BYTES, "rdma_read_bytes") \
45 op(EFA_RDMA_READ_WR_ERR, "rdma_read_wr_err") \
46 op(EFA_RDMA_READ_RESP_BYTES, "rdma_read_resp_bytes") \
47 op(EFA_SUBMITTED_CMDS, "submitted_cmds") \
48 op(EFA_COMPLETED_CMDS, "completed_cmds") \
49 op(EFA_CMDS_ERR, "cmds_err") \
50 op(EFA_NO_COMPLETION_CMDS, "no_completion_cmds") \
51 op(EFA_KEEP_ALIVE_RCVD, "keep_alive_rcvd") \
52 op(EFA_ALLOC_PD_ERR, "alloc_pd_err") \
53 op(EFA_CREATE_QP_ERR, "create_qp_err") \
54 op(EFA_CREATE_CQ_ERR, "create_cq_err") \
55 op(EFA_REG_MR_ERR, "reg_mr_err") \
56 op(EFA_ALLOC_UCONTEXT_ERR, "alloc_ucontext_err") \
57 op(EFA_CREATE_AH_ERR, "create_ah_err") \
58 op(EFA_MMAP_ERR, "mmap_err")
60 #define EFA_STATS_ENUM(ename, name) ename,
61 #define EFA_STATS_STR(ename, name) [ename] = name,
64 EFA_DEFINE_STATS(EFA_STATS_ENUM)
67 static const char *const efa_stats_names[] = {
68 EFA_DEFINE_STATS(EFA_STATS_STR)
71 #define EFA_CHUNK_PAYLOAD_SHIFT 12
72 #define EFA_CHUNK_PAYLOAD_SIZE BIT(EFA_CHUNK_PAYLOAD_SHIFT)
73 #define EFA_CHUNK_PAYLOAD_PTR_SIZE 8
75 #define EFA_CHUNK_SHIFT 12
76 #define EFA_CHUNK_SIZE BIT(EFA_CHUNK_SHIFT)
77 #define EFA_CHUNK_PTR_SIZE sizeof(struct efa_com_ctrl_buff_info)
79 #define EFA_PTRS_PER_CHUNK \
80 ((EFA_CHUNK_SIZE - EFA_CHUNK_PTR_SIZE) / EFA_CHUNK_PAYLOAD_PTR_SIZE)
82 #define EFA_CHUNK_USED_SIZE \
83 ((EFA_PTRS_PER_CHUNK * EFA_CHUNK_PAYLOAD_PTR_SIZE) + EFA_CHUNK_PTR_SIZE)
91 struct pbl_chunk_list {
92 struct pbl_chunk *chunks;
102 u32 pbl_buf_size_in_pages;
103 struct scatterlist *sgl;
105 struct pbl_chunk_list chunk_list;
109 u32 pbl_buf_size_in_bytes;
110 u8 physically_continuous;
113 static inline struct efa_dev *to_edev(struct ib_device *ibdev)
115 return container_of(ibdev, struct efa_dev, ibdev);
118 static inline struct efa_ucontext *to_eucontext(struct ib_ucontext *ibucontext)
120 return container_of(ibucontext, struct efa_ucontext, ibucontext);
123 static inline struct efa_pd *to_epd(struct ib_pd *ibpd)
125 return container_of(ibpd, struct efa_pd, ibpd);
128 static inline struct efa_mr *to_emr(struct ib_mr *ibmr)
130 return container_of(ibmr, struct efa_mr, ibmr);
133 static inline struct efa_qp *to_eqp(struct ib_qp *ibqp)
135 return container_of(ibqp, struct efa_qp, ibqp);
138 static inline struct efa_cq *to_ecq(struct ib_cq *ibcq)
140 return container_of(ibcq, struct efa_cq, ibcq);
143 static inline struct efa_ah *to_eah(struct ib_ah *ibah)
145 return container_of(ibah, struct efa_ah, ibah);
148 static inline struct efa_user_mmap_entry *
149 to_emmap(struct rdma_user_mmap_entry *rdma_entry)
151 return container_of(rdma_entry, struct efa_user_mmap_entry, rdma_entry);
154 #define EFA_DEV_CAP(dev, cap) \
155 ((dev)->dev_attr.device_caps & \
156 EFA_ADMIN_FEATURE_DEVICE_ATTR_DESC_##cap##_MASK)
158 #define is_reserved_cleared(reserved) \
159 !memchr_inv(reserved, 0, sizeof(reserved))
161 static void *efa_zalloc_mapped(struct efa_dev *dev, dma_addr_t *dma_addr,
162 size_t size, enum dma_data_direction dir)
166 addr = alloc_pages_exact(size, GFP_KERNEL | __GFP_ZERO);
170 *dma_addr = dma_map_single(&dev->pdev->dev, addr, size, dir);
171 if (dma_mapping_error(&dev->pdev->dev, *dma_addr)) {
172 ibdev_err(&dev->ibdev, "Failed to map DMA address\n");
173 free_pages_exact(addr, size);
180 static void efa_free_mapped(struct efa_dev *dev, void *cpu_addr,
182 size_t size, enum dma_data_direction dir)
184 dma_unmap_single(&dev->pdev->dev, dma_addr, size, dir);
185 free_pages_exact(cpu_addr, size);
188 int efa_query_device(struct ib_device *ibdev,
189 struct ib_device_attr *props,
190 struct ib_udata *udata)
192 struct efa_com_get_device_attr_result *dev_attr;
193 struct efa_ibv_ex_query_device_resp resp = {};
194 struct efa_dev *dev = to_edev(ibdev);
197 if (udata && udata->inlen &&
198 !ib_is_udata_cleared(udata, 0, udata->inlen)) {
200 "Incompatible ABI params, udata not cleared\n");
204 dev_attr = &dev->dev_attr;
206 memset(props, 0, sizeof(*props));
207 props->max_mr_size = dev_attr->max_mr_pages * PAGE_SIZE;
208 props->page_size_cap = dev_attr->page_size_cap;
209 props->vendor_id = dev->pdev->vendor;
210 props->vendor_part_id = dev->pdev->device;
211 props->hw_ver = dev->pdev->subsystem_device;
212 props->max_qp = dev_attr->max_qp;
213 props->max_cq = dev_attr->max_cq;
214 props->max_pd = dev_attr->max_pd;
215 props->max_mr = dev_attr->max_mr;
216 props->max_ah = dev_attr->max_ah;
217 props->max_cqe = dev_attr->max_cq_depth;
218 props->max_qp_wr = min_t(u32, dev_attr->max_sq_depth,
219 dev_attr->max_rq_depth);
220 props->max_send_sge = dev_attr->max_sq_sge;
221 props->max_recv_sge = dev_attr->max_rq_sge;
222 props->max_sge_rd = dev_attr->max_wr_rdma_sge;
223 props->max_pkeys = 1;
225 if (udata && udata->outlen) {
226 resp.max_sq_sge = dev_attr->max_sq_sge;
227 resp.max_rq_sge = dev_attr->max_rq_sge;
228 resp.max_sq_wr = dev_attr->max_sq_depth;
229 resp.max_rq_wr = dev_attr->max_rq_depth;
230 resp.max_rdma_size = dev_attr->max_rdma_size;
232 if (EFA_DEV_CAP(dev, RDMA_READ))
233 resp.device_caps |= EFA_QUERY_DEVICE_CAPS_RDMA_READ;
235 if (EFA_DEV_CAP(dev, RNR_RETRY))
236 resp.device_caps |= EFA_QUERY_DEVICE_CAPS_RNR_RETRY;
238 err = ib_copy_to_udata(udata, &resp,
239 min(sizeof(resp), udata->outlen));
242 "Failed to copy udata for query_device\n");
250 int efa_query_port(struct ib_device *ibdev, u8 port,
251 struct ib_port_attr *props)
253 struct efa_dev *dev = to_edev(ibdev);
257 props->state = IB_PORT_ACTIVE;
258 props->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
259 props->gid_tbl_len = 1;
260 props->pkey_tbl_len = 1;
261 props->active_speed = IB_SPEED_EDR;
262 props->active_width = IB_WIDTH_4X;
263 props->max_mtu = ib_mtu_int_to_enum(dev->dev_attr.mtu);
264 props->active_mtu = ib_mtu_int_to_enum(dev->dev_attr.mtu);
265 props->max_msg_sz = dev->dev_attr.mtu;
266 props->max_vl_num = 1;
271 int efa_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
273 struct ib_qp_init_attr *qp_init_attr)
275 struct efa_dev *dev = to_edev(ibqp->device);
276 struct efa_com_query_qp_params params = {};
277 struct efa_com_query_qp_result result;
278 struct efa_qp *qp = to_eqp(ibqp);
281 #define EFA_QUERY_QP_SUPP_MASK \
282 (IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT | \
283 IB_QP_QKEY | IB_QP_SQ_PSN | IB_QP_CAP | IB_QP_RNR_RETRY)
285 if (qp_attr_mask & ~EFA_QUERY_QP_SUPP_MASK) {
286 ibdev_dbg(&dev->ibdev,
287 "Unsupported qp_attr_mask[%#x] supported[%#x]\n",
288 qp_attr_mask, EFA_QUERY_QP_SUPP_MASK);
292 memset(qp_attr, 0, sizeof(*qp_attr));
293 memset(qp_init_attr, 0, sizeof(*qp_init_attr));
295 params.qp_handle = qp->qp_handle;
296 err = efa_com_query_qp(&dev->edev, ¶ms, &result);
300 qp_attr->qp_state = result.qp_state;
301 qp_attr->qkey = result.qkey;
302 qp_attr->sq_psn = result.sq_psn;
303 qp_attr->sq_draining = result.sq_draining;
304 qp_attr->port_num = 1;
305 qp_attr->rnr_retry = result.rnr_retry;
307 qp_attr->cap.max_send_wr = qp->max_send_wr;
308 qp_attr->cap.max_recv_wr = qp->max_recv_wr;
309 qp_attr->cap.max_send_sge = qp->max_send_sge;
310 qp_attr->cap.max_recv_sge = qp->max_recv_sge;
311 qp_attr->cap.max_inline_data = qp->max_inline_data;
313 qp_init_attr->qp_type = ibqp->qp_type;
314 qp_init_attr->recv_cq = ibqp->recv_cq;
315 qp_init_attr->send_cq = ibqp->send_cq;
316 qp_init_attr->qp_context = ibqp->qp_context;
317 qp_init_attr->cap = qp_attr->cap;
322 int efa_query_gid(struct ib_device *ibdev, u8 port, int index,
325 struct efa_dev *dev = to_edev(ibdev);
327 memcpy(gid->raw, dev->dev_attr.addr, sizeof(dev->dev_attr.addr));
332 int efa_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
342 static int efa_pd_dealloc(struct efa_dev *dev, u16 pdn)
344 struct efa_com_dealloc_pd_params params = {
348 return efa_com_dealloc_pd(&dev->edev, ¶ms);
351 int efa_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
353 struct efa_dev *dev = to_edev(ibpd->device);
354 struct efa_ibv_alloc_pd_resp resp = {};
355 struct efa_com_alloc_pd_result result;
356 struct efa_pd *pd = to_epd(ibpd);
360 !ib_is_udata_cleared(udata, 0, udata->inlen)) {
361 ibdev_dbg(&dev->ibdev,
362 "Incompatible ABI params, udata not cleared\n");
367 err = efa_com_alloc_pd(&dev->edev, &result);
371 pd->pdn = result.pdn;
372 resp.pdn = result.pdn;
375 err = ib_copy_to_udata(udata, &resp,
376 min(sizeof(resp), udata->outlen));
378 ibdev_dbg(&dev->ibdev,
379 "Failed to copy udata for alloc_pd\n");
384 ibdev_dbg(&dev->ibdev, "Allocated pd[%d]\n", pd->pdn);
389 efa_pd_dealloc(dev, result.pdn);
391 atomic64_inc(&dev->stats.alloc_pd_err);
395 int efa_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
397 struct efa_dev *dev = to_edev(ibpd->device);
398 struct efa_pd *pd = to_epd(ibpd);
400 ibdev_dbg(&dev->ibdev, "Dealloc pd[%d]\n", pd->pdn);
401 efa_pd_dealloc(dev, pd->pdn);
405 static int efa_destroy_qp_handle(struct efa_dev *dev, u32 qp_handle)
407 struct efa_com_destroy_qp_params params = { .qp_handle = qp_handle };
409 return efa_com_destroy_qp(&dev->edev, ¶ms);
412 static void efa_qp_user_mmap_entries_remove(struct efa_qp *qp)
414 rdma_user_mmap_entry_remove(qp->rq_mmap_entry);
415 rdma_user_mmap_entry_remove(qp->rq_db_mmap_entry);
416 rdma_user_mmap_entry_remove(qp->llq_desc_mmap_entry);
417 rdma_user_mmap_entry_remove(qp->sq_db_mmap_entry);
420 int efa_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
422 struct efa_dev *dev = to_edev(ibqp->pd->device);
423 struct efa_qp *qp = to_eqp(ibqp);
426 ibdev_dbg(&dev->ibdev, "Destroy qp[%u]\n", ibqp->qp_num);
428 efa_qp_user_mmap_entries_remove(qp);
430 err = efa_destroy_qp_handle(dev, qp->qp_handle);
434 if (qp->rq_cpu_addr) {
435 ibdev_dbg(&dev->ibdev,
436 "qp->cpu_addr[0x%p] freed: size[%lu], dma[%pad]\n",
437 qp->rq_cpu_addr, qp->rq_size,
439 efa_free_mapped(dev, qp->rq_cpu_addr, qp->rq_dma_addr,
440 qp->rq_size, DMA_TO_DEVICE);
447 static struct rdma_user_mmap_entry*
448 efa_user_mmap_entry_insert(struct ib_ucontext *ucontext,
449 u64 address, size_t length,
450 u8 mmap_flag, u64 *offset)
452 struct efa_user_mmap_entry *entry = kzalloc(sizeof(*entry), GFP_KERNEL);
458 entry->address = address;
459 entry->mmap_flag = mmap_flag;
461 err = rdma_user_mmap_entry_insert(ucontext, &entry->rdma_entry,
467 *offset = rdma_user_mmap_get_offset(&entry->rdma_entry);
469 return &entry->rdma_entry;
472 static int qp_mmap_entries_setup(struct efa_qp *qp,
474 struct efa_ucontext *ucontext,
475 struct efa_com_create_qp_params *params,
476 struct efa_ibv_create_qp_resp *resp)
481 address = dev->db_bar_addr + resp->sq_db_offset;
482 qp->sq_db_mmap_entry =
483 efa_user_mmap_entry_insert(&ucontext->ibucontext,
485 PAGE_SIZE, EFA_MMAP_IO_NC,
486 &resp->sq_db_mmap_key);
487 if (!qp->sq_db_mmap_entry)
490 resp->sq_db_offset &= ~PAGE_MASK;
492 address = dev->mem_bar_addr + resp->llq_desc_offset;
493 length = PAGE_ALIGN(params->sq_ring_size_in_bytes +
494 (resp->llq_desc_offset & ~PAGE_MASK));
496 qp->llq_desc_mmap_entry =
497 efa_user_mmap_entry_insert(&ucontext->ibucontext,
500 &resp->llq_desc_mmap_key);
501 if (!qp->llq_desc_mmap_entry)
502 goto err_remove_mmap;
504 resp->llq_desc_offset &= ~PAGE_MASK;
507 address = dev->db_bar_addr + resp->rq_db_offset;
509 qp->rq_db_mmap_entry =
510 efa_user_mmap_entry_insert(&ucontext->ibucontext,
513 &resp->rq_db_mmap_key);
514 if (!qp->rq_db_mmap_entry)
515 goto err_remove_mmap;
517 resp->rq_db_offset &= ~PAGE_MASK;
519 address = virt_to_phys(qp->rq_cpu_addr);
521 efa_user_mmap_entry_insert(&ucontext->ibucontext,
522 address, qp->rq_size,
525 if (!qp->rq_mmap_entry)
526 goto err_remove_mmap;
528 resp->rq_mmap_size = qp->rq_size;
534 efa_qp_user_mmap_entries_remove(qp);
539 static int efa_qp_validate_cap(struct efa_dev *dev,
540 struct ib_qp_init_attr *init_attr)
542 if (init_attr->cap.max_send_wr > dev->dev_attr.max_sq_depth) {
543 ibdev_dbg(&dev->ibdev,
544 "qp: requested send wr[%u] exceeds the max[%u]\n",
545 init_attr->cap.max_send_wr,
546 dev->dev_attr.max_sq_depth);
549 if (init_attr->cap.max_recv_wr > dev->dev_attr.max_rq_depth) {
550 ibdev_dbg(&dev->ibdev,
551 "qp: requested receive wr[%u] exceeds the max[%u]\n",
552 init_attr->cap.max_recv_wr,
553 dev->dev_attr.max_rq_depth);
556 if (init_attr->cap.max_send_sge > dev->dev_attr.max_sq_sge) {
557 ibdev_dbg(&dev->ibdev,
558 "qp: requested sge send[%u] exceeds the max[%u]\n",
559 init_attr->cap.max_send_sge, dev->dev_attr.max_sq_sge);
562 if (init_attr->cap.max_recv_sge > dev->dev_attr.max_rq_sge) {
563 ibdev_dbg(&dev->ibdev,
564 "qp: requested sge recv[%u] exceeds the max[%u]\n",
565 init_attr->cap.max_recv_sge, dev->dev_attr.max_rq_sge);
568 if (init_attr->cap.max_inline_data > dev->dev_attr.inline_buf_size) {
569 ibdev_dbg(&dev->ibdev,
570 "qp: requested inline data[%u] exceeds the max[%u]\n",
571 init_attr->cap.max_inline_data,
572 dev->dev_attr.inline_buf_size);
579 static int efa_qp_validate_attr(struct efa_dev *dev,
580 struct ib_qp_init_attr *init_attr)
582 if (init_attr->qp_type != IB_QPT_DRIVER &&
583 init_attr->qp_type != IB_QPT_UD) {
584 ibdev_dbg(&dev->ibdev,
585 "Unsupported qp type %d\n", init_attr->qp_type);
589 if (init_attr->srq) {
590 ibdev_dbg(&dev->ibdev, "SRQ is not supported\n");
594 if (init_attr->create_flags) {
595 ibdev_dbg(&dev->ibdev, "Unsupported create flags\n");
602 struct ib_qp *efa_create_qp(struct ib_pd *ibpd,
603 struct ib_qp_init_attr *init_attr,
604 struct ib_udata *udata)
606 struct efa_com_create_qp_params create_qp_params = {};
607 struct efa_com_create_qp_result create_qp_resp;
608 struct efa_dev *dev = to_edev(ibpd->device);
609 struct efa_ibv_create_qp_resp resp = {};
610 struct efa_ibv_create_qp cmd = {};
611 struct efa_ucontext *ucontext;
615 ucontext = rdma_udata_to_drv_context(udata, struct efa_ucontext,
618 err = efa_qp_validate_cap(dev, init_attr);
622 err = efa_qp_validate_attr(dev, init_attr);
626 if (offsetofend(typeof(cmd), driver_qp_type) > udata->inlen) {
627 ibdev_dbg(&dev->ibdev,
628 "Incompatible ABI params, no input udata\n");
633 if (udata->inlen > sizeof(cmd) &&
634 !ib_is_udata_cleared(udata, sizeof(cmd),
635 udata->inlen - sizeof(cmd))) {
636 ibdev_dbg(&dev->ibdev,
637 "Incompatible ABI params, unknown fields in udata\n");
642 err = ib_copy_from_udata(&cmd, udata,
643 min(sizeof(cmd), udata->inlen));
645 ibdev_dbg(&dev->ibdev,
646 "Cannot copy udata for create_qp\n");
651 ibdev_dbg(&dev->ibdev,
652 "Incompatible ABI params, unknown fields in udata\n");
657 qp = kzalloc(sizeof(*qp), GFP_KERNEL);
663 create_qp_params.uarn = ucontext->uarn;
664 create_qp_params.pd = to_epd(ibpd)->pdn;
666 if (init_attr->qp_type == IB_QPT_UD) {
667 create_qp_params.qp_type = EFA_ADMIN_QP_TYPE_UD;
668 } else if (cmd.driver_qp_type == EFA_QP_DRIVER_TYPE_SRD) {
669 create_qp_params.qp_type = EFA_ADMIN_QP_TYPE_SRD;
671 ibdev_dbg(&dev->ibdev,
672 "Unsupported qp type %d driver qp type %d\n",
673 init_attr->qp_type, cmd.driver_qp_type);
678 ibdev_dbg(&dev->ibdev, "Create QP: qp type %d driver qp type %#x\n",
679 init_attr->qp_type, cmd.driver_qp_type);
680 create_qp_params.send_cq_idx = to_ecq(init_attr->send_cq)->cq_idx;
681 create_qp_params.recv_cq_idx = to_ecq(init_attr->recv_cq)->cq_idx;
682 create_qp_params.sq_depth = init_attr->cap.max_send_wr;
683 create_qp_params.sq_ring_size_in_bytes = cmd.sq_ring_size;
685 create_qp_params.rq_depth = init_attr->cap.max_recv_wr;
686 create_qp_params.rq_ring_size_in_bytes = cmd.rq_ring_size;
687 qp->rq_size = PAGE_ALIGN(create_qp_params.rq_ring_size_in_bytes);
689 qp->rq_cpu_addr = efa_zalloc_mapped(dev, &qp->rq_dma_addr,
690 qp->rq_size, DMA_TO_DEVICE);
691 if (!qp->rq_cpu_addr) {
696 ibdev_dbg(&dev->ibdev,
697 "qp->cpu_addr[0x%p] allocated: size[%lu], dma[%pad]\n",
698 qp->rq_cpu_addr, qp->rq_size, &qp->rq_dma_addr);
699 create_qp_params.rq_base_addr = qp->rq_dma_addr;
702 err = efa_com_create_qp(&dev->edev, &create_qp_params,
705 goto err_free_mapped;
707 resp.sq_db_offset = create_qp_resp.sq_db_offset;
708 resp.rq_db_offset = create_qp_resp.rq_db_offset;
709 resp.llq_desc_offset = create_qp_resp.llq_descriptors_offset;
710 resp.send_sub_cq_idx = create_qp_resp.send_sub_cq_idx;
711 resp.recv_sub_cq_idx = create_qp_resp.recv_sub_cq_idx;
713 err = qp_mmap_entries_setup(qp, dev, ucontext, &create_qp_params,
718 qp->qp_handle = create_qp_resp.qp_handle;
719 qp->ibqp.qp_num = create_qp_resp.qp_num;
720 qp->ibqp.qp_type = init_attr->qp_type;
721 qp->max_send_wr = init_attr->cap.max_send_wr;
722 qp->max_recv_wr = init_attr->cap.max_recv_wr;
723 qp->max_send_sge = init_attr->cap.max_send_sge;
724 qp->max_recv_sge = init_attr->cap.max_recv_sge;
725 qp->max_inline_data = init_attr->cap.max_inline_data;
728 err = ib_copy_to_udata(udata, &resp,
729 min(sizeof(resp), udata->outlen));
731 ibdev_dbg(&dev->ibdev,
732 "Failed to copy udata for qp[%u]\n",
733 create_qp_resp.qp_num);
734 goto err_remove_mmap_entries;
738 ibdev_dbg(&dev->ibdev, "Created qp[%d]\n", qp->ibqp.qp_num);
742 err_remove_mmap_entries:
743 efa_qp_user_mmap_entries_remove(qp);
745 efa_destroy_qp_handle(dev, create_qp_resp.qp_handle);
748 efa_free_mapped(dev, qp->rq_cpu_addr, qp->rq_dma_addr,
749 qp->rq_size, DMA_TO_DEVICE);
753 atomic64_inc(&dev->stats.create_qp_err);
757 static const struct {
759 enum ib_qp_attr_mask req_param;
760 enum ib_qp_attr_mask opt_param;
761 } srd_qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
763 [IB_QPS_RESET] = { .valid = 1 },
766 .req_param = IB_QP_PKEY_INDEX |
772 [IB_QPS_RESET] = { .valid = 1 },
773 [IB_QPS_ERR] = { .valid = 1 },
776 .opt_param = IB_QP_PKEY_INDEX |
782 .opt_param = IB_QP_PKEY_INDEX |
787 [IB_QPS_RESET] = { .valid = 1 },
788 [IB_QPS_ERR] = { .valid = 1 },
791 .req_param = IB_QP_SQ_PSN,
792 .opt_param = IB_QP_CUR_STATE |
799 [IB_QPS_RESET] = { .valid = 1 },
800 [IB_QPS_ERR] = { .valid = 1 },
803 .opt_param = IB_QP_CUR_STATE |
808 .opt_param = IB_QP_EN_SQD_ASYNC_NOTIFY,
812 [IB_QPS_RESET] = { .valid = 1 },
813 [IB_QPS_ERR] = { .valid = 1 },
816 .opt_param = IB_QP_CUR_STATE |
821 .opt_param = IB_QP_PKEY_INDEX |
826 [IB_QPS_RESET] = { .valid = 1 },
827 [IB_QPS_ERR] = { .valid = 1 },
830 .opt_param = IB_QP_CUR_STATE |
835 [IB_QPS_RESET] = { .valid = 1 },
836 [IB_QPS_ERR] = { .valid = 1 },
840 static bool efa_modify_srd_qp_is_ok(enum ib_qp_state cur_state,
841 enum ib_qp_state next_state,
842 enum ib_qp_attr_mask mask)
844 enum ib_qp_attr_mask req_param, opt_param;
846 if (mask & IB_QP_CUR_STATE &&
847 cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
848 cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
851 if (!srd_qp_state_table[cur_state][next_state].valid)
854 req_param = srd_qp_state_table[cur_state][next_state].req_param;
855 opt_param = srd_qp_state_table[cur_state][next_state].opt_param;
857 if ((mask & req_param) != req_param)
860 if (mask & ~(req_param | opt_param | IB_QP_STATE))
866 static int efa_modify_qp_validate(struct efa_dev *dev, struct efa_qp *qp,
867 struct ib_qp_attr *qp_attr, int qp_attr_mask,
868 enum ib_qp_state cur_state,
869 enum ib_qp_state new_state)
873 #define EFA_MODIFY_QP_SUPP_MASK \
874 (IB_QP_STATE | IB_QP_CUR_STATE | IB_QP_EN_SQD_ASYNC_NOTIFY | \
875 IB_QP_PKEY_INDEX | IB_QP_PORT | IB_QP_QKEY | IB_QP_SQ_PSN | \
878 if (qp_attr_mask & ~EFA_MODIFY_QP_SUPP_MASK) {
879 ibdev_dbg(&dev->ibdev,
880 "Unsupported qp_attr_mask[%#x] supported[%#x]\n",
881 qp_attr_mask, EFA_MODIFY_QP_SUPP_MASK);
885 if (qp->ibqp.qp_type == IB_QPT_DRIVER)
886 err = !efa_modify_srd_qp_is_ok(cur_state, new_state,
889 err = !ib_modify_qp_is_ok(cur_state, new_state, IB_QPT_UD,
893 ibdev_dbg(&dev->ibdev, "Invalid modify QP parameters\n");
897 if ((qp_attr_mask & IB_QP_PORT) && qp_attr->port_num != 1) {
898 ibdev_dbg(&dev->ibdev, "Can't change port num\n");
902 if ((qp_attr_mask & IB_QP_PKEY_INDEX) && qp_attr->pkey_index) {
903 ibdev_dbg(&dev->ibdev, "Can't change pkey index\n");
910 int efa_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
911 int qp_attr_mask, struct ib_udata *udata)
913 struct efa_dev *dev = to_edev(ibqp->device);
914 struct efa_com_modify_qp_params params = {};
915 struct efa_qp *qp = to_eqp(ibqp);
916 enum ib_qp_state cur_state;
917 enum ib_qp_state new_state;
921 !ib_is_udata_cleared(udata, 0, udata->inlen)) {
922 ibdev_dbg(&dev->ibdev,
923 "Incompatible ABI params, udata not cleared\n");
927 cur_state = qp_attr_mask & IB_QP_CUR_STATE ? qp_attr->cur_qp_state :
929 new_state = qp_attr_mask & IB_QP_STATE ? qp_attr->qp_state : cur_state;
931 err = efa_modify_qp_validate(dev, qp, qp_attr, qp_attr_mask, cur_state,
936 params.qp_handle = qp->qp_handle;
938 if (qp_attr_mask & IB_QP_STATE) {
939 EFA_SET(¶ms.modify_mask, EFA_ADMIN_MODIFY_QP_CMD_QP_STATE,
941 EFA_SET(¶ms.modify_mask,
942 EFA_ADMIN_MODIFY_QP_CMD_CUR_QP_STATE, 1);
943 params.cur_qp_state = cur_state;
944 params.qp_state = new_state;
947 if (qp_attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY) {
948 EFA_SET(¶ms.modify_mask,
949 EFA_ADMIN_MODIFY_QP_CMD_SQ_DRAINED_ASYNC_NOTIFY, 1);
950 params.sq_drained_async_notify = qp_attr->en_sqd_async_notify;
953 if (qp_attr_mask & IB_QP_QKEY) {
954 EFA_SET(¶ms.modify_mask, EFA_ADMIN_MODIFY_QP_CMD_QKEY, 1);
955 params.qkey = qp_attr->qkey;
958 if (qp_attr_mask & IB_QP_SQ_PSN) {
959 EFA_SET(¶ms.modify_mask, EFA_ADMIN_MODIFY_QP_CMD_SQ_PSN, 1);
960 params.sq_psn = qp_attr->sq_psn;
963 if (qp_attr_mask & IB_QP_RNR_RETRY) {
964 EFA_SET(¶ms.modify_mask, EFA_ADMIN_MODIFY_QP_CMD_RNR_RETRY,
966 params.rnr_retry = qp_attr->rnr_retry;
969 err = efa_com_modify_qp(&dev->edev, ¶ms);
973 qp->state = new_state;
978 static int efa_destroy_cq_idx(struct efa_dev *dev, int cq_idx)
980 struct efa_com_destroy_cq_params params = { .cq_idx = cq_idx };
982 return efa_com_destroy_cq(&dev->edev, ¶ms);
985 int efa_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
987 struct efa_dev *dev = to_edev(ibcq->device);
988 struct efa_cq *cq = to_ecq(ibcq);
990 ibdev_dbg(&dev->ibdev,
991 "Destroy cq[%d] virt[0x%p] freed: size[%lu], dma[%pad]\n",
992 cq->cq_idx, cq->cpu_addr, cq->size, &cq->dma_addr);
994 rdma_user_mmap_entry_remove(cq->mmap_entry);
995 efa_destroy_cq_idx(dev, cq->cq_idx);
996 efa_free_mapped(dev, cq->cpu_addr, cq->dma_addr, cq->size,
1001 static int cq_mmap_entries_setup(struct efa_dev *dev, struct efa_cq *cq,
1002 struct efa_ibv_create_cq_resp *resp)
1004 resp->q_mmap_size = cq->size;
1005 cq->mmap_entry = efa_user_mmap_entry_insert(&cq->ucontext->ibucontext,
1006 virt_to_phys(cq->cpu_addr),
1007 cq->size, EFA_MMAP_DMA_PAGE,
1009 if (!cq->mmap_entry)
1015 int efa_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
1016 struct ib_udata *udata)
1018 struct efa_ucontext *ucontext = rdma_udata_to_drv_context(
1019 udata, struct efa_ucontext, ibucontext);
1020 struct efa_ibv_create_cq_resp resp = {};
1021 struct efa_com_create_cq_params params;
1022 struct efa_com_create_cq_result result;
1023 struct ib_device *ibdev = ibcq->device;
1024 struct efa_dev *dev = to_edev(ibdev);
1025 struct efa_ibv_create_cq cmd = {};
1026 struct efa_cq *cq = to_ecq(ibcq);
1027 int entries = attr->cqe;
1030 ibdev_dbg(ibdev, "create_cq entries %d\n", entries);
1032 if (entries < 1 || entries > dev->dev_attr.max_cq_depth) {
1034 "cq: requested entries[%u] non-positive or greater than max[%u]\n",
1035 entries, dev->dev_attr.max_cq_depth);
1040 if (offsetofend(typeof(cmd), num_sub_cqs) > udata->inlen) {
1042 "Incompatible ABI params, no input udata\n");
1047 if (udata->inlen > sizeof(cmd) &&
1048 !ib_is_udata_cleared(udata, sizeof(cmd),
1049 udata->inlen - sizeof(cmd))) {
1051 "Incompatible ABI params, unknown fields in udata\n");
1056 err = ib_copy_from_udata(&cmd, udata,
1057 min(sizeof(cmd), udata->inlen));
1059 ibdev_dbg(ibdev, "Cannot copy udata for create_cq\n");
1063 if (cmd.comp_mask || !is_reserved_cleared(cmd.reserved_50)) {
1065 "Incompatible ABI params, unknown fields in udata\n");
1070 if (!cmd.cq_entry_size) {
1072 "Invalid entry size [%u]\n", cmd.cq_entry_size);
1077 if (cmd.num_sub_cqs != dev->dev_attr.sub_cqs_per_cq) {
1079 "Invalid number of sub cqs[%u] expected[%u]\n",
1080 cmd.num_sub_cqs, dev->dev_attr.sub_cqs_per_cq);
1085 cq->ucontext = ucontext;
1086 cq->size = PAGE_ALIGN(cmd.cq_entry_size * entries * cmd.num_sub_cqs);
1087 cq->cpu_addr = efa_zalloc_mapped(dev, &cq->dma_addr, cq->size,
1089 if (!cq->cpu_addr) {
1094 params.uarn = cq->ucontext->uarn;
1095 params.cq_depth = entries;
1096 params.dma_addr = cq->dma_addr;
1097 params.entry_size_in_bytes = cmd.cq_entry_size;
1098 params.num_sub_cqs = cmd.num_sub_cqs;
1099 err = efa_com_create_cq(&dev->edev, ¶ms, &result);
1101 goto err_free_mapped;
1103 resp.cq_idx = result.cq_idx;
1104 cq->cq_idx = result.cq_idx;
1105 cq->ibcq.cqe = result.actual_depth;
1106 WARN_ON_ONCE(entries != result.actual_depth);
1108 err = cq_mmap_entries_setup(dev, cq, &resp);
1110 ibdev_dbg(ibdev, "Could not setup cq[%u] mmap entries\n",
1112 goto err_destroy_cq;
1115 if (udata->outlen) {
1116 err = ib_copy_to_udata(udata, &resp,
1117 min(sizeof(resp), udata->outlen));
1120 "Failed to copy udata for create_cq\n");
1121 goto err_remove_mmap;
1125 ibdev_dbg(ibdev, "Created cq[%d], cq depth[%u]. dma[%pad] virt[0x%p]\n",
1126 cq->cq_idx, result.actual_depth, &cq->dma_addr, cq->cpu_addr);
1131 rdma_user_mmap_entry_remove(cq->mmap_entry);
1133 efa_destroy_cq_idx(dev, cq->cq_idx);
1135 efa_free_mapped(dev, cq->cpu_addr, cq->dma_addr, cq->size,
1139 atomic64_inc(&dev->stats.create_cq_err);
1143 static int umem_to_page_list(struct efa_dev *dev,
1144 struct ib_umem *umem,
1149 u32 pages_in_hp = BIT(hp_shift - PAGE_SHIFT);
1150 struct ib_block_iter biter;
1151 unsigned int hp_idx = 0;
1153 ibdev_dbg(&dev->ibdev, "hp_cnt[%u], pages_in_hp[%u]\n",
1154 hp_cnt, pages_in_hp);
1156 rdma_umem_for_each_dma_block(umem, &biter, BIT(hp_shift))
1157 page_list[hp_idx++] = rdma_block_iter_dma_address(&biter);
1162 static struct scatterlist *efa_vmalloc_buf_to_sg(u64 *buf, int page_cnt)
1164 struct scatterlist *sglist;
1168 sglist = kmalloc_array(page_cnt, sizeof(*sglist), GFP_KERNEL);
1171 sg_init_table(sglist, page_cnt);
1172 for (i = 0; i < page_cnt; i++) {
1173 pg = vmalloc_to_page(buf);
1176 sg_set_page(&sglist[i], pg, PAGE_SIZE, 0);
1177 buf += PAGE_SIZE / sizeof(*buf);
1187 * create a chunk list of physical pages dma addresses from the supplied
1188 * scatter gather list
1190 static int pbl_chunk_list_create(struct efa_dev *dev, struct pbl_context *pbl)
1192 struct pbl_chunk_list *chunk_list = &pbl->phys.indirect.chunk_list;
1193 int page_cnt = pbl->phys.indirect.pbl_buf_size_in_pages;
1194 struct scatterlist *pages_sgl = pbl->phys.indirect.sgl;
1195 unsigned int chunk_list_size, chunk_idx, payload_idx;
1196 int sg_dma_cnt = pbl->phys.indirect.sg_dma_cnt;
1197 struct efa_com_ctrl_buff_info *ctrl_buf;
1198 u64 *cur_chunk_buf, *prev_chunk_buf;
1199 struct ib_block_iter biter;
1200 dma_addr_t dma_addr;
1203 /* allocate a chunk list that consists of 4KB chunks */
1204 chunk_list_size = DIV_ROUND_UP(page_cnt, EFA_PTRS_PER_CHUNK);
1206 chunk_list->size = chunk_list_size;
1207 chunk_list->chunks = kcalloc(chunk_list_size,
1208 sizeof(*chunk_list->chunks),
1210 if (!chunk_list->chunks)
1213 ibdev_dbg(&dev->ibdev,
1214 "chunk_list_size[%u] - pages[%u]\n", chunk_list_size,
1217 /* allocate chunk buffers: */
1218 for (i = 0; i < chunk_list_size; i++) {
1219 chunk_list->chunks[i].buf = kzalloc(EFA_CHUNK_SIZE, GFP_KERNEL);
1220 if (!chunk_list->chunks[i].buf)
1221 goto chunk_list_dealloc;
1223 chunk_list->chunks[i].length = EFA_CHUNK_USED_SIZE;
1225 chunk_list->chunks[chunk_list_size - 1].length =
1226 ((page_cnt % EFA_PTRS_PER_CHUNK) * EFA_CHUNK_PAYLOAD_PTR_SIZE) +
1229 /* fill the dma addresses of sg list pages to chunks: */
1232 cur_chunk_buf = chunk_list->chunks[0].buf;
1233 rdma_for_each_block(pages_sgl, &biter, sg_dma_cnt,
1234 EFA_CHUNK_PAYLOAD_SIZE) {
1235 cur_chunk_buf[payload_idx++] =
1236 rdma_block_iter_dma_address(&biter);
1238 if (payload_idx == EFA_PTRS_PER_CHUNK) {
1240 cur_chunk_buf = chunk_list->chunks[chunk_idx].buf;
1245 /* map chunks to dma and fill chunks next ptrs */
1246 for (i = chunk_list_size - 1; i >= 0; i--) {
1247 dma_addr = dma_map_single(&dev->pdev->dev,
1248 chunk_list->chunks[i].buf,
1249 chunk_list->chunks[i].length,
1251 if (dma_mapping_error(&dev->pdev->dev, dma_addr)) {
1252 ibdev_err(&dev->ibdev,
1253 "chunk[%u] dma_map_failed\n", i);
1254 goto chunk_list_unmap;
1257 chunk_list->chunks[i].dma_addr = dma_addr;
1258 ibdev_dbg(&dev->ibdev,
1259 "chunk[%u] mapped at [%pad]\n", i, &dma_addr);
1264 prev_chunk_buf = chunk_list->chunks[i - 1].buf;
1266 ctrl_buf = (struct efa_com_ctrl_buff_info *)
1267 &prev_chunk_buf[EFA_PTRS_PER_CHUNK];
1268 ctrl_buf->length = chunk_list->chunks[i].length;
1270 efa_com_set_dma_addr(dma_addr,
1271 &ctrl_buf->address.mem_addr_high,
1272 &ctrl_buf->address.mem_addr_low);
1278 for (; i < chunk_list_size; i++) {
1279 dma_unmap_single(&dev->pdev->dev, chunk_list->chunks[i].dma_addr,
1280 chunk_list->chunks[i].length, DMA_TO_DEVICE);
1283 for (i = 0; i < chunk_list_size; i++)
1284 kfree(chunk_list->chunks[i].buf);
1286 kfree(chunk_list->chunks);
1290 static void pbl_chunk_list_destroy(struct efa_dev *dev, struct pbl_context *pbl)
1292 struct pbl_chunk_list *chunk_list = &pbl->phys.indirect.chunk_list;
1295 for (i = 0; i < chunk_list->size; i++) {
1296 dma_unmap_single(&dev->pdev->dev, chunk_list->chunks[i].dma_addr,
1297 chunk_list->chunks[i].length, DMA_TO_DEVICE);
1298 kfree(chunk_list->chunks[i].buf);
1301 kfree(chunk_list->chunks);
1304 /* initialize pbl continuous mode: map pbl buffer to a dma address. */
1305 static int pbl_continuous_initialize(struct efa_dev *dev,
1306 struct pbl_context *pbl)
1308 dma_addr_t dma_addr;
1310 dma_addr = dma_map_single(&dev->pdev->dev, pbl->pbl_buf,
1311 pbl->pbl_buf_size_in_bytes, DMA_TO_DEVICE);
1312 if (dma_mapping_error(&dev->pdev->dev, dma_addr)) {
1313 ibdev_err(&dev->ibdev, "Unable to map pbl to DMA address\n");
1317 pbl->phys.continuous.dma_addr = dma_addr;
1318 ibdev_dbg(&dev->ibdev,
1319 "pbl continuous - dma_addr = %pad, size[%u]\n",
1320 &dma_addr, pbl->pbl_buf_size_in_bytes);
1326 * initialize pbl indirect mode:
1327 * create a chunk list out of the dma addresses of the physical pages of
1330 static int pbl_indirect_initialize(struct efa_dev *dev, struct pbl_context *pbl)
1332 u32 size_in_pages = DIV_ROUND_UP(pbl->pbl_buf_size_in_bytes, PAGE_SIZE);
1333 struct scatterlist *sgl;
1334 int sg_dma_cnt, err;
1336 BUILD_BUG_ON(EFA_CHUNK_PAYLOAD_SIZE > PAGE_SIZE);
1337 sgl = efa_vmalloc_buf_to_sg(pbl->pbl_buf, size_in_pages);
1341 sg_dma_cnt = dma_map_sg(&dev->pdev->dev, sgl, size_in_pages, DMA_TO_DEVICE);
1347 pbl->phys.indirect.pbl_buf_size_in_pages = size_in_pages;
1348 pbl->phys.indirect.sgl = sgl;
1349 pbl->phys.indirect.sg_dma_cnt = sg_dma_cnt;
1350 err = pbl_chunk_list_create(dev, pbl);
1352 ibdev_dbg(&dev->ibdev,
1353 "chunk_list creation failed[%d]\n", err);
1357 ibdev_dbg(&dev->ibdev,
1358 "pbl indirect - size[%u], chunks[%u]\n",
1359 pbl->pbl_buf_size_in_bytes,
1360 pbl->phys.indirect.chunk_list.size);
1365 dma_unmap_sg(&dev->pdev->dev, sgl, size_in_pages, DMA_TO_DEVICE);
1371 static void pbl_indirect_terminate(struct efa_dev *dev, struct pbl_context *pbl)
1373 pbl_chunk_list_destroy(dev, pbl);
1374 dma_unmap_sg(&dev->pdev->dev, pbl->phys.indirect.sgl,
1375 pbl->phys.indirect.pbl_buf_size_in_pages, DMA_TO_DEVICE);
1376 kfree(pbl->phys.indirect.sgl);
1379 /* create a page buffer list from a mapped user memory region */
1380 static int pbl_create(struct efa_dev *dev,
1381 struct pbl_context *pbl,
1382 struct ib_umem *umem,
1388 pbl->pbl_buf_size_in_bytes = hp_cnt * EFA_CHUNK_PAYLOAD_PTR_SIZE;
1389 pbl->pbl_buf = kvzalloc(pbl->pbl_buf_size_in_bytes, GFP_KERNEL);
1393 if (is_vmalloc_addr(pbl->pbl_buf)) {
1394 pbl->physically_continuous = 0;
1395 err = umem_to_page_list(dev, umem, pbl->pbl_buf, hp_cnt,
1400 err = pbl_indirect_initialize(dev, pbl);
1404 pbl->physically_continuous = 1;
1405 err = umem_to_page_list(dev, umem, pbl->pbl_buf, hp_cnt,
1410 err = pbl_continuous_initialize(dev, pbl);
1415 ibdev_dbg(&dev->ibdev,
1416 "user_pbl_created: user_pages[%u], continuous[%u]\n",
1417 hp_cnt, pbl->physically_continuous);
1422 kvfree(pbl->pbl_buf);
1426 static void pbl_destroy(struct efa_dev *dev, struct pbl_context *pbl)
1428 if (pbl->physically_continuous)
1429 dma_unmap_single(&dev->pdev->dev, pbl->phys.continuous.dma_addr,
1430 pbl->pbl_buf_size_in_bytes, DMA_TO_DEVICE);
1432 pbl_indirect_terminate(dev, pbl);
1434 kvfree(pbl->pbl_buf);
1437 static int efa_create_inline_pbl(struct efa_dev *dev, struct efa_mr *mr,
1438 struct efa_com_reg_mr_params *params)
1442 params->inline_pbl = 1;
1443 err = umem_to_page_list(dev, mr->umem, params->pbl.inline_pbl_array,
1444 params->page_num, params->page_shift);
1448 ibdev_dbg(&dev->ibdev,
1449 "inline_pbl_array - pages[%u]\n", params->page_num);
1454 static int efa_create_pbl(struct efa_dev *dev,
1455 struct pbl_context *pbl,
1457 struct efa_com_reg_mr_params *params)
1461 err = pbl_create(dev, pbl, mr->umem, params->page_num,
1462 params->page_shift);
1464 ibdev_dbg(&dev->ibdev, "Failed to create pbl[%d]\n", err);
1468 params->inline_pbl = 0;
1469 params->indirect = !pbl->physically_continuous;
1470 if (pbl->physically_continuous) {
1471 params->pbl.pbl.length = pbl->pbl_buf_size_in_bytes;
1473 efa_com_set_dma_addr(pbl->phys.continuous.dma_addr,
1474 ¶ms->pbl.pbl.address.mem_addr_high,
1475 ¶ms->pbl.pbl.address.mem_addr_low);
1477 params->pbl.pbl.length =
1478 pbl->phys.indirect.chunk_list.chunks[0].length;
1480 efa_com_set_dma_addr(pbl->phys.indirect.chunk_list.chunks[0].dma_addr,
1481 ¶ms->pbl.pbl.address.mem_addr_high,
1482 ¶ms->pbl.pbl.address.mem_addr_low);
1488 struct ib_mr *efa_reg_mr(struct ib_pd *ibpd, u64 start, u64 length,
1489 u64 virt_addr, int access_flags,
1490 struct ib_udata *udata)
1492 struct efa_dev *dev = to_edev(ibpd->device);
1493 struct efa_com_reg_mr_params params = {};
1494 struct efa_com_reg_mr_result result = {};
1495 struct pbl_context pbl;
1496 int supp_access_flags;
1502 if (udata && udata->inlen &&
1503 !ib_is_udata_cleared(udata, 0, sizeof(udata->inlen))) {
1504 ibdev_dbg(&dev->ibdev,
1505 "Incompatible ABI params, udata not cleared\n");
1511 IB_ACCESS_LOCAL_WRITE |
1512 (EFA_DEV_CAP(dev, RDMA_READ) ? IB_ACCESS_REMOTE_READ : 0);
1514 access_flags &= ~IB_ACCESS_OPTIONAL;
1515 if (access_flags & ~supp_access_flags) {
1516 ibdev_dbg(&dev->ibdev,
1517 "Unsupported access flags[%#x], supported[%#x]\n",
1518 access_flags, supp_access_flags);
1523 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1529 mr->umem = ib_umem_get(ibpd->device, start, length, access_flags);
1530 if (IS_ERR(mr->umem)) {
1531 err = PTR_ERR(mr->umem);
1532 ibdev_dbg(&dev->ibdev,
1533 "Failed to pin and map user space memory[%d]\n", err);
1537 params.pd = to_epd(ibpd)->pdn;
1538 params.iova = virt_addr;
1539 params.mr_length_in_bytes = length;
1540 params.permissions = access_flags;
1542 pg_sz = ib_umem_find_best_pgsz(mr->umem,
1543 dev->dev_attr.page_size_cap,
1547 ibdev_dbg(&dev->ibdev, "Failed to find a suitable page size in page_size_cap %#llx\n",
1548 dev->dev_attr.page_size_cap);
1552 params.page_shift = order_base_2(pg_sz);
1553 params.page_num = ib_umem_num_dma_blocks(mr->umem, pg_sz);
1555 ibdev_dbg(&dev->ibdev,
1556 "start %#llx length %#llx params.page_shift %u params.page_num %u\n",
1557 start, length, params.page_shift, params.page_num);
1559 inline_size = ARRAY_SIZE(params.pbl.inline_pbl_array);
1560 if (params.page_num <= inline_size) {
1561 err = efa_create_inline_pbl(dev, mr, ¶ms);
1565 err = efa_com_register_mr(&dev->edev, ¶ms, &result);
1569 err = efa_create_pbl(dev, &pbl, mr, ¶ms);
1573 err = efa_com_register_mr(&dev->edev, ¶ms, &result);
1574 pbl_destroy(dev, &pbl);
1580 mr->ibmr.lkey = result.l_key;
1581 mr->ibmr.rkey = result.r_key;
1582 mr->ibmr.length = length;
1583 ibdev_dbg(&dev->ibdev, "Registered mr[%d]\n", mr->ibmr.lkey);
1588 ib_umem_release(mr->umem);
1592 atomic64_inc(&dev->stats.reg_mr_err);
1593 return ERR_PTR(err);
1596 int efa_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
1598 struct efa_dev *dev = to_edev(ibmr->device);
1599 struct efa_com_dereg_mr_params params;
1600 struct efa_mr *mr = to_emr(ibmr);
1603 ibdev_dbg(&dev->ibdev, "Deregister mr[%d]\n", ibmr->lkey);
1605 params.l_key = mr->ibmr.lkey;
1606 err = efa_com_dereg_mr(&dev->edev, ¶ms);
1610 ib_umem_release(mr->umem);
1616 int efa_get_port_immutable(struct ib_device *ibdev, u8 port_num,
1617 struct ib_port_immutable *immutable)
1619 struct ib_port_attr attr;
1622 err = ib_query_port(ibdev, port_num, &attr);
1624 ibdev_dbg(ibdev, "Couldn't query port err[%d]\n", err);
1628 immutable->pkey_tbl_len = attr.pkey_tbl_len;
1629 immutable->gid_tbl_len = attr.gid_tbl_len;
1634 static int efa_dealloc_uar(struct efa_dev *dev, u16 uarn)
1636 struct efa_com_dealloc_uar_params params = {
1640 return efa_com_dealloc_uar(&dev->edev, ¶ms);
1643 #define EFA_CHECK_USER_COMP(_dev, _comp_mask, _attr, _mask, _attr_str) \
1644 (_attr_str = (!(_dev)->dev_attr._attr || ((_comp_mask) & (_mask))) ? \
1647 static int efa_user_comp_handshake(const struct ib_ucontext *ibucontext,
1648 const struct efa_ibv_alloc_ucontext_cmd *cmd)
1650 struct efa_dev *dev = to_edev(ibucontext->device);
1653 if (EFA_CHECK_USER_COMP(dev, cmd->comp_mask, max_tx_batch,
1654 EFA_ALLOC_UCONTEXT_CMD_COMP_TX_BATCH, attr_str))
1657 if (EFA_CHECK_USER_COMP(dev, cmd->comp_mask, min_sq_depth,
1658 EFA_ALLOC_UCONTEXT_CMD_COMP_MIN_SQ_WR,
1665 ibdev_dbg(&dev->ibdev, "Userspace handshake failed for %s attribute\n",
1670 int efa_alloc_ucontext(struct ib_ucontext *ibucontext, struct ib_udata *udata)
1672 struct efa_ucontext *ucontext = to_eucontext(ibucontext);
1673 struct efa_dev *dev = to_edev(ibucontext->device);
1674 struct efa_ibv_alloc_ucontext_resp resp = {};
1675 struct efa_ibv_alloc_ucontext_cmd cmd = {};
1676 struct efa_com_alloc_uar_result result;
1680 * it's fine if the driver does not know all request fields,
1681 * we will ack input fields in our response.
1684 err = ib_copy_from_udata(&cmd, udata,
1685 min(sizeof(cmd), udata->inlen));
1687 ibdev_dbg(&dev->ibdev,
1688 "Cannot copy udata for alloc_ucontext\n");
1692 err = efa_user_comp_handshake(ibucontext, &cmd);
1696 err = efa_com_alloc_uar(&dev->edev, &result);
1700 ucontext->uarn = result.uarn;
1702 resp.cmds_supp_udata_mask |= EFA_USER_CMDS_SUPP_UDATA_QUERY_DEVICE;
1703 resp.cmds_supp_udata_mask |= EFA_USER_CMDS_SUPP_UDATA_CREATE_AH;
1704 resp.sub_cqs_per_cq = dev->dev_attr.sub_cqs_per_cq;
1705 resp.inline_buf_size = dev->dev_attr.inline_buf_size;
1706 resp.max_llq_size = dev->dev_attr.max_llq_size;
1707 resp.max_tx_batch = dev->dev_attr.max_tx_batch;
1708 resp.min_sq_wr = dev->dev_attr.min_sq_depth;
1710 err = ib_copy_to_udata(udata, &resp,
1711 min(sizeof(resp), udata->outlen));
1713 goto err_dealloc_uar;
1718 efa_dealloc_uar(dev, result.uarn);
1720 atomic64_inc(&dev->stats.alloc_ucontext_err);
1724 void efa_dealloc_ucontext(struct ib_ucontext *ibucontext)
1726 struct efa_ucontext *ucontext = to_eucontext(ibucontext);
1727 struct efa_dev *dev = to_edev(ibucontext->device);
1729 efa_dealloc_uar(dev, ucontext->uarn);
1732 void efa_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
1734 struct efa_user_mmap_entry *entry = to_emmap(rdma_entry);
1739 static int __efa_mmap(struct efa_dev *dev, struct efa_ucontext *ucontext,
1740 struct vm_area_struct *vma)
1742 struct rdma_user_mmap_entry *rdma_entry;
1743 struct efa_user_mmap_entry *entry;
1748 rdma_entry = rdma_user_mmap_entry_get(&ucontext->ibucontext, vma);
1750 ibdev_dbg(&dev->ibdev,
1751 "pgoff[%#lx] does not have valid entry\n",
1753 atomic64_inc(&dev->stats.mmap_err);
1756 entry = to_emmap(rdma_entry);
1758 ibdev_dbg(&dev->ibdev,
1759 "Mapping address[%#llx], length[%#zx], mmap_flag[%d]\n",
1760 entry->address, rdma_entry->npages * PAGE_SIZE,
1763 pfn = entry->address >> PAGE_SHIFT;
1764 switch (entry->mmap_flag) {
1765 case EFA_MMAP_IO_NC:
1766 err = rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn,
1767 entry->rdma_entry.npages * PAGE_SIZE,
1768 pgprot_noncached(vma->vm_page_prot),
1771 case EFA_MMAP_IO_WC:
1772 err = rdma_user_mmap_io(&ucontext->ibucontext, vma, pfn,
1773 entry->rdma_entry.npages * PAGE_SIZE,
1774 pgprot_writecombine(vma->vm_page_prot),
1777 case EFA_MMAP_DMA_PAGE:
1778 for (va = vma->vm_start; va < vma->vm_end;
1779 va += PAGE_SIZE, pfn++) {
1780 err = vm_insert_page(vma, va, pfn_to_page(pfn));
1792 "Couldn't mmap address[%#llx] length[%#zx] mmap_flag[%d] err[%d]\n",
1793 entry->address, rdma_entry->npages * PAGE_SIZE,
1794 entry->mmap_flag, err);
1795 atomic64_inc(&dev->stats.mmap_err);
1798 rdma_user_mmap_entry_put(rdma_entry);
1802 int efa_mmap(struct ib_ucontext *ibucontext,
1803 struct vm_area_struct *vma)
1805 struct efa_ucontext *ucontext = to_eucontext(ibucontext);
1806 struct efa_dev *dev = to_edev(ibucontext->device);
1807 size_t length = vma->vm_end - vma->vm_start;
1809 ibdev_dbg(&dev->ibdev,
1810 "start %#lx, end %#lx, length = %#zx, pgoff = %#lx\n",
1811 vma->vm_start, vma->vm_end, length, vma->vm_pgoff);
1813 return __efa_mmap(dev, ucontext, vma);
1816 static int efa_ah_destroy(struct efa_dev *dev, struct efa_ah *ah)
1818 struct efa_com_destroy_ah_params params = {
1820 .pdn = to_epd(ah->ibah.pd)->pdn,
1823 return efa_com_destroy_ah(&dev->edev, ¶ms);
1826 int efa_create_ah(struct ib_ah *ibah,
1827 struct rdma_ah_init_attr *init_attr,
1828 struct ib_udata *udata)
1830 struct rdma_ah_attr *ah_attr = init_attr->ah_attr;
1831 struct efa_dev *dev = to_edev(ibah->device);
1832 struct efa_com_create_ah_params params = {};
1833 struct efa_ibv_create_ah_resp resp = {};
1834 struct efa_com_create_ah_result result;
1835 struct efa_ah *ah = to_eah(ibah);
1838 if (!(init_attr->flags & RDMA_CREATE_AH_SLEEPABLE)) {
1839 ibdev_dbg(&dev->ibdev,
1840 "Create address handle is not supported in atomic context\n");
1846 !ib_is_udata_cleared(udata, 0, udata->inlen)) {
1847 ibdev_dbg(&dev->ibdev, "Incompatible ABI params\n");
1852 memcpy(params.dest_addr, ah_attr->grh.dgid.raw,
1853 sizeof(params.dest_addr));
1854 params.pdn = to_epd(ibah->pd)->pdn;
1855 err = efa_com_create_ah(&dev->edev, ¶ms, &result);
1859 memcpy(ah->id, ah_attr->grh.dgid.raw, sizeof(ah->id));
1862 resp.efa_address_handle = result.ah;
1864 if (udata->outlen) {
1865 err = ib_copy_to_udata(udata, &resp,
1866 min(sizeof(resp), udata->outlen));
1868 ibdev_dbg(&dev->ibdev,
1869 "Failed to copy udata for create_ah response\n");
1870 goto err_destroy_ah;
1873 ibdev_dbg(&dev->ibdev, "Created ah[%d]\n", ah->ah);
1878 efa_ah_destroy(dev, ah);
1880 atomic64_inc(&dev->stats.create_ah_err);
1884 int efa_destroy_ah(struct ib_ah *ibah, u32 flags)
1886 struct efa_dev *dev = to_edev(ibah->pd->device);
1887 struct efa_ah *ah = to_eah(ibah);
1889 ibdev_dbg(&dev->ibdev, "Destroy ah[%d]\n", ah->ah);
1891 if (!(flags & RDMA_DESTROY_AH_SLEEPABLE)) {
1892 ibdev_dbg(&dev->ibdev,
1893 "Destroy address handle is not supported in atomic context\n");
1897 efa_ah_destroy(dev, ah);
1901 struct rdma_hw_stats *efa_alloc_hw_stats(struct ib_device *ibdev, u8 port_num)
1903 return rdma_alloc_hw_stats_struct(efa_stats_names,
1904 ARRAY_SIZE(efa_stats_names),
1905 RDMA_HW_STATS_DEFAULT_LIFESPAN);
1908 int efa_get_hw_stats(struct ib_device *ibdev, struct rdma_hw_stats *stats,
1909 u8 port_num, int index)
1911 struct efa_com_get_stats_params params = {};
1912 union efa_com_get_stats_result result;
1913 struct efa_dev *dev = to_edev(ibdev);
1914 struct efa_com_rdma_read_stats *rrs;
1915 struct efa_com_messages_stats *ms;
1916 struct efa_com_basic_stats *bs;
1917 struct efa_com_stats_admin *as;
1918 struct efa_stats *s;
1921 params.scope = EFA_ADMIN_GET_STATS_SCOPE_ALL;
1922 params.type = EFA_ADMIN_GET_STATS_TYPE_BASIC;
1924 err = efa_com_get_stats(&dev->edev, ¶ms, &result);
1928 bs = &result.basic_stats;
1929 stats->value[EFA_TX_BYTES] = bs->tx_bytes;
1930 stats->value[EFA_TX_PKTS] = bs->tx_pkts;
1931 stats->value[EFA_RX_BYTES] = bs->rx_bytes;
1932 stats->value[EFA_RX_PKTS] = bs->rx_pkts;
1933 stats->value[EFA_RX_DROPS] = bs->rx_drops;
1935 params.type = EFA_ADMIN_GET_STATS_TYPE_MESSAGES;
1936 err = efa_com_get_stats(&dev->edev, ¶ms, &result);
1940 ms = &result.messages_stats;
1941 stats->value[EFA_SEND_BYTES] = ms->send_bytes;
1942 stats->value[EFA_SEND_WRS] = ms->send_wrs;
1943 stats->value[EFA_RECV_BYTES] = ms->recv_bytes;
1944 stats->value[EFA_RECV_WRS] = ms->recv_wrs;
1946 params.type = EFA_ADMIN_GET_STATS_TYPE_RDMA_READ;
1947 err = efa_com_get_stats(&dev->edev, ¶ms, &result);
1951 rrs = &result.rdma_read_stats;
1952 stats->value[EFA_RDMA_READ_WRS] = rrs->read_wrs;
1953 stats->value[EFA_RDMA_READ_BYTES] = rrs->read_bytes;
1954 stats->value[EFA_RDMA_READ_WR_ERR] = rrs->read_wr_err;
1955 stats->value[EFA_RDMA_READ_RESP_BYTES] = rrs->read_resp_bytes;
1957 as = &dev->edev.aq.stats;
1958 stats->value[EFA_SUBMITTED_CMDS] = atomic64_read(&as->submitted_cmd);
1959 stats->value[EFA_COMPLETED_CMDS] = atomic64_read(&as->completed_cmd);
1960 stats->value[EFA_CMDS_ERR] = atomic64_read(&as->cmd_err);
1961 stats->value[EFA_NO_COMPLETION_CMDS] = atomic64_read(&as->no_completion);
1964 stats->value[EFA_KEEP_ALIVE_RCVD] = atomic64_read(&s->keep_alive_rcvd);
1965 stats->value[EFA_ALLOC_PD_ERR] = atomic64_read(&s->alloc_pd_err);
1966 stats->value[EFA_CREATE_QP_ERR] = atomic64_read(&s->create_qp_err);
1967 stats->value[EFA_CREATE_CQ_ERR] = atomic64_read(&s->create_cq_err);
1968 stats->value[EFA_REG_MR_ERR] = atomic64_read(&s->reg_mr_err);
1969 stats->value[EFA_ALLOC_UCONTEXT_ERR] =
1970 atomic64_read(&s->alloc_ucontext_err);
1971 stats->value[EFA_CREATE_AH_ERR] = atomic64_read(&s->create_ah_err);
1972 stats->value[EFA_MMAP_ERR] = atomic64_read(&s->mmap_err);
1974 return ARRAY_SIZE(efa_stats_names);
1977 enum rdma_link_layer efa_port_link_layer(struct ib_device *ibdev,
1980 return IB_LINK_LAYER_UNSPECIFIED;