1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2
3 /* Authors: Cheng Xu <[email protected]> */
4 /* Kai Shen <[email protected]> */
5 /* Copyright (c) 2020-2022, Alibaba Group. */
6
7 #include "erdma_verbs.h"
8
get_next_valid_cqe(struct erdma_cq * cq)9 static void *get_next_valid_cqe(struct erdma_cq *cq)
10 {
11 __be32 *cqe = get_queue_entry(cq->kern_cq.qbuf, cq->kern_cq.ci,
12 cq->depth, CQE_SHIFT);
13 u32 owner = FIELD_GET(ERDMA_CQE_HDR_OWNER_MASK,
14 be32_to_cpu(READ_ONCE(*cqe)));
15
16 return owner ^ !!(cq->kern_cq.ci & cq->depth) ? cqe : NULL;
17 }
18
notify_cq(struct erdma_cq * cq,u8 solcitied)19 static void notify_cq(struct erdma_cq *cq, u8 solcitied)
20 {
21 u64 db_data =
22 FIELD_PREP(ERDMA_CQDB_IDX_MASK, (cq->kern_cq.notify_cnt)) |
23 FIELD_PREP(ERDMA_CQDB_CQN_MASK, cq->cqn) |
24 FIELD_PREP(ERDMA_CQDB_ARM_MASK, 1) |
25 FIELD_PREP(ERDMA_CQDB_SOL_MASK, solcitied) |
26 FIELD_PREP(ERDMA_CQDB_CMDSN_MASK, cq->kern_cq.cmdsn) |
27 FIELD_PREP(ERDMA_CQDB_CI_MASK, cq->kern_cq.ci);
28
29 *cq->kern_cq.dbrec = db_data;
30 writeq(db_data, cq->kern_cq.db);
31 }
32
erdma_req_notify_cq(struct ib_cq * ibcq,enum ib_cq_notify_flags flags)33 int erdma_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
34 {
35 struct erdma_cq *cq = to_ecq(ibcq);
36 unsigned long irq_flags;
37 int ret = 0;
38
39 spin_lock_irqsave(&cq->kern_cq.lock, irq_flags);
40
41 notify_cq(cq, (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED);
42
43 if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && get_next_valid_cqe(cq))
44 ret = 1;
45
46 cq->kern_cq.notify_cnt++;
47
48 spin_unlock_irqrestore(&cq->kern_cq.lock, irq_flags);
49
50 return ret;
51 }
52
53 static const enum ib_wc_opcode wc_mapping_table[ERDMA_NUM_OPCODES] = {
54 [ERDMA_OP_WRITE] = IB_WC_RDMA_WRITE,
55 [ERDMA_OP_READ] = IB_WC_RDMA_READ,
56 [ERDMA_OP_SEND] = IB_WC_SEND,
57 [ERDMA_OP_SEND_WITH_IMM] = IB_WC_SEND,
58 [ERDMA_OP_RECEIVE] = IB_WC_RECV,
59 [ERDMA_OP_RECV_IMM] = IB_WC_RECV_RDMA_WITH_IMM,
60 [ERDMA_OP_RECV_INV] = IB_WC_RECV,
61 [ERDMA_OP_WRITE_WITH_IMM] = IB_WC_RDMA_WRITE,
62 [ERDMA_OP_RSP_SEND_IMM] = IB_WC_RECV,
63 [ERDMA_OP_SEND_WITH_INV] = IB_WC_SEND,
64 [ERDMA_OP_REG_MR] = IB_WC_REG_MR,
65 [ERDMA_OP_LOCAL_INV] = IB_WC_LOCAL_INV,
66 [ERDMA_OP_READ_WITH_INV] = IB_WC_RDMA_READ,
67 [ERDMA_OP_ATOMIC_CAS] = IB_WC_COMP_SWAP,
68 [ERDMA_OP_ATOMIC_FAA] = IB_WC_FETCH_ADD,
69 };
70
71 static const struct {
72 enum erdma_wc_status erdma;
73 enum ib_wc_status base;
74 enum erdma_vendor_err vendor;
75 } map_cqe_status[ERDMA_NUM_WC_STATUS] = {
76 { ERDMA_WC_SUCCESS, IB_WC_SUCCESS, ERDMA_WC_VENDOR_NO_ERR },
77 { ERDMA_WC_GENERAL_ERR, IB_WC_GENERAL_ERR, ERDMA_WC_VENDOR_NO_ERR },
78 { ERDMA_WC_RECV_WQE_FORMAT_ERR, IB_WC_GENERAL_ERR,
79 ERDMA_WC_VENDOR_INVALID_RQE },
80 { ERDMA_WC_RECV_STAG_INVALID_ERR, IB_WC_REM_ACCESS_ERR,
81 ERDMA_WC_VENDOR_RQE_INVALID_STAG },
82 { ERDMA_WC_RECV_ADDR_VIOLATION_ERR, IB_WC_REM_ACCESS_ERR,
83 ERDMA_WC_VENDOR_RQE_ADDR_VIOLATION },
84 { ERDMA_WC_RECV_RIGHT_VIOLATION_ERR, IB_WC_REM_ACCESS_ERR,
85 ERDMA_WC_VENDOR_RQE_ACCESS_RIGHT_ERR },
86 { ERDMA_WC_RECV_PDID_ERR, IB_WC_REM_ACCESS_ERR,
87 ERDMA_WC_VENDOR_RQE_INVALID_PD },
88 { ERDMA_WC_RECV_WARRPING_ERR, IB_WC_REM_ACCESS_ERR,
89 ERDMA_WC_VENDOR_RQE_WRAP_ERR },
90 { ERDMA_WC_SEND_WQE_FORMAT_ERR, IB_WC_LOC_QP_OP_ERR,
91 ERDMA_WC_VENDOR_INVALID_SQE },
92 { ERDMA_WC_SEND_WQE_ORD_EXCEED, IB_WC_GENERAL_ERR,
93 ERDMA_WC_VENDOR_ZERO_ORD },
94 { ERDMA_WC_SEND_STAG_INVALID_ERR, IB_WC_LOC_ACCESS_ERR,
95 ERDMA_WC_VENDOR_SQE_INVALID_STAG },
96 { ERDMA_WC_SEND_ADDR_VIOLATION_ERR, IB_WC_LOC_ACCESS_ERR,
97 ERDMA_WC_VENDOR_SQE_ADDR_VIOLATION },
98 { ERDMA_WC_SEND_RIGHT_VIOLATION_ERR, IB_WC_LOC_ACCESS_ERR,
99 ERDMA_WC_VENDOR_SQE_ACCESS_ERR },
100 { ERDMA_WC_SEND_PDID_ERR, IB_WC_LOC_ACCESS_ERR,
101 ERDMA_WC_VENDOR_SQE_INVALID_PD },
102 { ERDMA_WC_SEND_WARRPING_ERR, IB_WC_LOC_ACCESS_ERR,
103 ERDMA_WC_VENDOR_SQE_WARP_ERR },
104 { ERDMA_WC_FLUSH_ERR, IB_WC_WR_FLUSH_ERR, ERDMA_WC_VENDOR_NO_ERR },
105 { ERDMA_WC_RETRY_EXC_ERR, IB_WC_RETRY_EXC_ERR, ERDMA_WC_VENDOR_NO_ERR },
106 };
107
erdma_process_ud_cqe(struct erdma_cqe * cqe,struct ib_wc * wc)108 static void erdma_process_ud_cqe(struct erdma_cqe *cqe, struct ib_wc *wc)
109 {
110 u32 ud_info;
111
112 wc->wc_flags |= (IB_WC_GRH | IB_WC_WITH_NETWORK_HDR_TYPE);
113 ud_info = be32_to_cpu(cqe->ud.info);
114 wc->network_hdr_type = FIELD_GET(ERDMA_CQE_NTYPE_MASK, ud_info);
115 if (wc->network_hdr_type == ERDMA_NETWORK_TYPE_IPV4)
116 wc->network_hdr_type = RDMA_NETWORK_IPV4;
117 else
118 wc->network_hdr_type = RDMA_NETWORK_IPV6;
119 wc->src_qp = FIELD_GET(ERDMA_CQE_SQPN_MASK, ud_info);
120 wc->sl = FIELD_GET(ERDMA_CQE_SL_MASK, ud_info);
121 wc->pkey_index = 0;
122 }
123
124 #define ERDMA_POLLCQ_NO_QP 1
125
erdma_poll_one_cqe(struct erdma_cq * cq,struct ib_wc * wc)126 static int erdma_poll_one_cqe(struct erdma_cq *cq, struct ib_wc *wc)
127 {
128 struct erdma_dev *dev = to_edev(cq->ibcq.device);
129 u8 opcode, syndrome, qtype;
130 struct erdma_kqp *kern_qp;
131 struct erdma_cqe *cqe;
132 struct erdma_qp *qp;
133 u16 wqe_idx, depth;
134 u32 qpn, cqe_hdr;
135 u64 *id_table;
136 u64 *wqe_hdr;
137
138 cqe = get_next_valid_cqe(cq);
139 if (!cqe)
140 return -EAGAIN;
141
142 cq->kern_cq.ci++;
143
144 /* cqbuf should be ready when we poll */
145 dma_rmb();
146
147 qpn = be32_to_cpu(cqe->qpn);
148 wqe_idx = be32_to_cpu(cqe->qe_idx);
149 cqe_hdr = be32_to_cpu(cqe->hdr);
150
151 qp = find_qp_by_qpn(dev, qpn);
152 if (!qp)
153 return ERDMA_POLLCQ_NO_QP;
154
155 kern_qp = &qp->kern_qp;
156
157 qtype = FIELD_GET(ERDMA_CQE_HDR_QTYPE_MASK, cqe_hdr);
158 syndrome = FIELD_GET(ERDMA_CQE_HDR_SYNDROME_MASK, cqe_hdr);
159 opcode = FIELD_GET(ERDMA_CQE_HDR_OPCODE_MASK, cqe_hdr);
160
161 if (qtype == ERDMA_CQE_QTYPE_SQ) {
162 id_table = kern_qp->swr_tbl;
163 depth = qp->attrs.sq_size;
164 wqe_hdr = get_queue_entry(qp->kern_qp.sq_buf, wqe_idx,
165 qp->attrs.sq_size, SQEBB_SHIFT);
166 kern_qp->sq_ci =
167 FIELD_GET(ERDMA_SQE_HDR_WQEBB_CNT_MASK, *wqe_hdr) +
168 wqe_idx + 1;
169 } else {
170 id_table = kern_qp->rwr_tbl;
171 depth = qp->attrs.rq_size;
172 }
173 wc->wr_id = id_table[wqe_idx & (depth - 1)];
174 wc->byte_len = be32_to_cpu(cqe->size);
175
176 wc->wc_flags = 0;
177
178 wc->opcode = wc_mapping_table[opcode];
179 if (opcode == ERDMA_OP_RECV_IMM || opcode == ERDMA_OP_RSP_SEND_IMM) {
180 wc->ex.imm_data = cpu_to_be32(le32_to_cpu(cqe->imm_data));
181 wc->wc_flags |= IB_WC_WITH_IMM;
182 } else if (opcode == ERDMA_OP_RECV_INV) {
183 wc->ex.invalidate_rkey = be32_to_cpu(cqe->inv_rkey);
184 wc->wc_flags |= IB_WC_WITH_INVALIDATE;
185 }
186
187 if (erdma_device_rocev2(dev) &&
188 (qp->ibqp.qp_type == IB_QPT_UD || qp->ibqp.qp_type == IB_QPT_GSI))
189 erdma_process_ud_cqe(cqe, wc);
190
191 if (syndrome >= ERDMA_NUM_WC_STATUS)
192 syndrome = ERDMA_WC_GENERAL_ERR;
193
194 wc->status = map_cqe_status[syndrome].base;
195 wc->vendor_err = map_cqe_status[syndrome].vendor;
196 wc->qp = &qp->ibqp;
197
198 return 0;
199 }
200
erdma_poll_cq(struct ib_cq * ibcq,int num_entries,struct ib_wc * wc)201 int erdma_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
202 {
203 struct erdma_cq *cq = to_ecq(ibcq);
204 unsigned long flags;
205 int npolled, ret;
206
207 spin_lock_irqsave(&cq->kern_cq.lock, flags);
208
209 for (npolled = 0; npolled < num_entries;) {
210 ret = erdma_poll_one_cqe(cq, wc + npolled);
211
212 if (ret == -EAGAIN) /* no received new CQEs. */
213 break;
214 else if (ret) /* ignore invalid CQEs. */
215 continue;
216
217 npolled++;
218 }
219
220 spin_unlock_irqrestore(&cq->kern_cq.lock, flags);
221
222 return npolled;
223 }
224
erdma_remove_cqes_of_qp(struct ib_cq * ibcq,u32 qpn)225 void erdma_remove_cqes_of_qp(struct ib_cq *ibcq, u32 qpn)
226 {
227 struct erdma_cq *cq = to_ecq(ibcq);
228 struct erdma_cqe *cqe, *dst_cqe;
229 u32 prev_cq_ci, cur_cq_ci;
230 u32 ncqe = 0, nqp_cqe = 0;
231 unsigned long flags;
232 u8 owner;
233
234 spin_lock_irqsave(&cq->kern_cq.lock, flags);
235
236 prev_cq_ci = cq->kern_cq.ci;
237
238 while (ncqe < cq->depth && (cqe = get_next_valid_cqe(cq)) != NULL) {
239 ++cq->kern_cq.ci;
240 ++ncqe;
241 }
242
243 while (ncqe > 0) {
244 cur_cq_ci = prev_cq_ci + ncqe - 1;
245 cqe = get_queue_entry(cq->kern_cq.qbuf, cur_cq_ci, cq->depth,
246 CQE_SHIFT);
247
248 if (be32_to_cpu(cqe->qpn) == qpn) {
249 ++nqp_cqe;
250 } else if (nqp_cqe) {
251 dst_cqe = get_queue_entry(cq->kern_cq.qbuf,
252 cur_cq_ci + nqp_cqe,
253 cq->depth, CQE_SHIFT);
254 owner = FIELD_GET(ERDMA_CQE_HDR_OWNER_MASK,
255 be32_to_cpu(dst_cqe->hdr));
256 cqe->hdr = cpu_to_be32(
257 (be32_to_cpu(cqe->hdr) &
258 ~ERDMA_CQE_HDR_OWNER_MASK) |
259 FIELD_PREP(ERDMA_CQE_HDR_OWNER_MASK, owner));
260 memcpy(dst_cqe, cqe, sizeof(*cqe));
261 }
262
263 --ncqe;
264 }
265
266 cq->kern_cq.ci = prev_cq_ci + nqp_cqe;
267 spin_unlock_irqrestore(&cq->kern_cq.lock, flags);
268 }
269