1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB 2 /* Copyright (c) 2015 - 2021 Intel Corporation */ 3 #include "osdep.h" 4 #include "defs.h" 5 #include "user.h" 6 #include "irdma.h" 7 8 /** 9 * irdma_set_fragment - set fragment in wqe 10 * @wqe: wqe for setting fragment 11 * @offset: offset value 12 * @sge: sge length and stag 13 * @valid: The wqe valid 14 */ 15 static void irdma_set_fragment(__le64 *wqe, u32 offset, struct ib_sge *sge, 16 u8 valid) 17 { 18 if (sge) { 19 set_64bit_val(wqe, offset, 20 FIELD_PREP(IRDMAQPSQ_FRAG_TO, sge->addr)); 21 set_64bit_val(wqe, offset + 8, 22 FIELD_PREP(IRDMAQPSQ_VALID, valid) | 23 FIELD_PREP(IRDMAQPSQ_FRAG_LEN, sge->length) | 24 FIELD_PREP(IRDMAQPSQ_FRAG_STAG, sge->lkey)); 25 } else { 26 set_64bit_val(wqe, offset, 0); 27 set_64bit_val(wqe, offset + 8, 28 FIELD_PREP(IRDMAQPSQ_VALID, valid)); 29 } 30 } 31 32 /** 33 * irdma_set_fragment_gen_1 - set fragment in wqe 34 * @wqe: wqe for setting fragment 35 * @offset: offset value 36 * @sge: sge length and stag 37 * @valid: wqe valid flag 38 */ 39 static void irdma_set_fragment_gen_1(__le64 *wqe, u32 offset, 40 struct ib_sge *sge, u8 valid) 41 { 42 if (sge) { 43 set_64bit_val(wqe, offset, 44 FIELD_PREP(IRDMAQPSQ_FRAG_TO, sge->addr)); 45 set_64bit_val(wqe, offset + 8, 46 FIELD_PREP(IRDMAQPSQ_GEN1_FRAG_LEN, sge->length) | 47 FIELD_PREP(IRDMAQPSQ_GEN1_FRAG_STAG, sge->lkey)); 48 } else { 49 set_64bit_val(wqe, offset, 0); 50 set_64bit_val(wqe, offset + 8, 0); 51 } 52 } 53 54 /** 55 * irdma_nop_1 - insert a NOP wqe 56 * @qp: hw qp ptr 57 */ 58 static int irdma_nop_1(struct irdma_qp_uk *qp) 59 { 60 u64 hdr; 61 __le64 *wqe; 62 u32 wqe_idx; 63 bool signaled = false; 64 65 if (!qp->sq_ring.head) 66 return -EINVAL; 67 68 wqe_idx = IRDMA_RING_CURRENT_HEAD(qp->sq_ring); 69 wqe = qp->sq_base[wqe_idx].elem; 70 71 qp->sq_wrtrk_array[wqe_idx].quanta = IRDMA_QP_WQE_MIN_QUANTA; 72 73 set_64bit_val(wqe, 0, 0); 74 set_64bit_val(wqe, 8, 0); 75 set_64bit_val(wqe, 16, 0); 76 77 hdr = FIELD_PREP(IRDMAQPSQ_OPCODE, IRDMAQP_OP_NOP) | 78 FIELD_PREP(IRDMAQPSQ_SIGCOMPL, signaled) | 79 FIELD_PREP(IRDMAQPSQ_VALID, qp->swqe_polarity); 80 81 /* make sure WQE is written before valid bit is set */ 82 dma_wmb(); 83 84 set_64bit_val(wqe, 24, hdr); 85 86 return 0; 87 } 88 89 /** 90 * irdma_clr_wqes - clear next 128 sq entries 91 * @qp: hw qp ptr 92 * @qp_wqe_idx: wqe_idx 93 */ 94 void irdma_clr_wqes(struct irdma_qp_uk *qp, u32 qp_wqe_idx) 95 { 96 struct irdma_qp_quanta *sq; 97 u32 wqe_idx; 98 99 if (!(qp_wqe_idx & 0x7F)) { 100 wqe_idx = (qp_wqe_idx + 128) % qp->sq_ring.size; 101 sq = qp->sq_base + wqe_idx; 102 if (wqe_idx) 103 memset(sq, qp->swqe_polarity ? 0 : 0xFF, 104 128 * sizeof(*sq)); 105 else 106 memset(sq, qp->swqe_polarity ? 0xFF : 0, 107 128 * sizeof(*sq)); 108 } 109 } 110 111 /** 112 * irdma_uk_qp_post_wr - ring doorbell 113 * @qp: hw qp ptr 114 */ 115 void irdma_uk_qp_post_wr(struct irdma_qp_uk *qp) 116 { 117 u64 temp; 118 u32 hw_sq_tail; 119 u32 sw_sq_head; 120 121 /* valid bit is written and loads completed before reading shadow */ 122 mb(); 123 124 /* read the doorbell shadow area */ 125 get_64bit_val(qp->shadow_area, 0, &temp); 126 127 hw_sq_tail = (u32)FIELD_GET(IRDMA_QP_DBSA_HW_SQ_TAIL, temp); 128 sw_sq_head = IRDMA_RING_CURRENT_HEAD(qp->sq_ring); 129 if (sw_sq_head != qp->initial_ring.head) { 130 if (sw_sq_head != hw_sq_tail) { 131 if (sw_sq_head > qp->initial_ring.head) { 132 if (hw_sq_tail >= qp->initial_ring.head && 133 hw_sq_tail < sw_sq_head) 134 writel(qp->qp_id, qp->wqe_alloc_db); 135 } else { 136 if (hw_sq_tail >= qp->initial_ring.head || 137 hw_sq_tail < sw_sq_head) 138 writel(qp->qp_id, qp->wqe_alloc_db); 139 } 140 } 141 } 142 143 qp->initial_ring.head = qp->sq_ring.head; 144 } 145 146 /** 147 * irdma_qp_get_next_send_wqe - pad with NOP if needed, return where next WR should go 148 * @qp: hw qp ptr 149 * @wqe_idx: return wqe index 150 * @quanta: size of WR in quanta 151 * @total_size: size of WR in bytes 152 * @info: info on WR 153 */ 154 __le64 *irdma_qp_get_next_send_wqe(struct irdma_qp_uk *qp, u32 *wqe_idx, 155 u16 quanta, u32 total_size, 156 struct irdma_post_sq_info *info) 157 { 158 __le64 *wqe; 159 __le64 *wqe_0 = NULL; 160 u16 avail_quanta; 161 u16 i; 162 163 avail_quanta = qp->uk_attrs->max_hw_sq_chunk - 164 (IRDMA_RING_CURRENT_HEAD(qp->sq_ring) % 165 qp->uk_attrs->max_hw_sq_chunk); 166 if (quanta <= avail_quanta) { 167 /* WR fits in current chunk */ 168 if (quanta > IRDMA_SQ_RING_FREE_QUANTA(qp->sq_ring)) 169 return NULL; 170 } else { 171 /* Need to pad with NOP */ 172 if (quanta + avail_quanta > 173 IRDMA_SQ_RING_FREE_QUANTA(qp->sq_ring)) 174 return NULL; 175 176 for (i = 0; i < avail_quanta; i++) { 177 irdma_nop_1(qp); 178 IRDMA_RING_MOVE_HEAD_NOCHECK(qp->sq_ring); 179 } 180 } 181 182 *wqe_idx = IRDMA_RING_CURRENT_HEAD(qp->sq_ring); 183 if (!*wqe_idx) 184 qp->swqe_polarity = !qp->swqe_polarity; 185 186 IRDMA_RING_MOVE_HEAD_BY_COUNT_NOCHECK(qp->sq_ring, quanta); 187 188 wqe = qp->sq_base[*wqe_idx].elem; 189 if (qp->uk_attrs->hw_rev == IRDMA_GEN_1 && quanta == 1 && 190 (IRDMA_RING_CURRENT_HEAD(qp->sq_ring) & 1)) { 191 wqe_0 = qp->sq_base[IRDMA_RING_CURRENT_HEAD(qp->sq_ring)].elem; 192 wqe_0[3] = cpu_to_le64(FIELD_PREP(IRDMAQPSQ_VALID, qp->swqe_polarity ? 0 : 1)); 193 } 194 qp->sq_wrtrk_array[*wqe_idx].wrid = info->wr_id; 195 qp->sq_wrtrk_array[*wqe_idx].wr_len = total_size; 196 qp->sq_wrtrk_array[*wqe_idx].quanta = quanta; 197 198 return wqe; 199 } 200 201 /** 202 * irdma_qp_get_next_recv_wqe - get next qp's rcv wqe 203 * @qp: hw qp ptr 204 * @wqe_idx: return wqe index 205 */ 206 __le64 *irdma_qp_get_next_recv_wqe(struct irdma_qp_uk *qp, u32 *wqe_idx) 207 { 208 __le64 *wqe; 209 int ret_code; 210 211 if (IRDMA_RING_FULL_ERR(qp->rq_ring)) 212 return NULL; 213 214 IRDMA_ATOMIC_RING_MOVE_HEAD(qp->rq_ring, *wqe_idx, ret_code); 215 if (ret_code) 216 return NULL; 217 218 if (!*wqe_idx) 219 qp->rwqe_polarity = !qp->rwqe_polarity; 220 /* rq_wqe_size_multiplier is no of 32 byte quanta in one rq wqe */ 221 wqe = qp->rq_base[*wqe_idx * qp->rq_wqe_size_multiplier].elem; 222 223 return wqe; 224 } 225 226 /** 227 * irdma_uk_rdma_write - rdma write operation 228 * @qp: hw qp ptr 229 * @info: post sq information 230 * @post_sq: flag to post sq 231 */ 232 int irdma_uk_rdma_write(struct irdma_qp_uk *qp, struct irdma_post_sq_info *info, 233 bool post_sq) 234 { 235 u64 hdr; 236 __le64 *wqe; 237 struct irdma_rdma_write *op_info; 238 u32 i, wqe_idx; 239 u32 total_size = 0, byte_off; 240 int ret_code; 241 u32 frag_cnt, addl_frag_cnt; 242 bool read_fence = false; 243 u16 quanta; 244 245 op_info = &info->op.rdma_write; 246 if (op_info->num_lo_sges > qp->max_sq_frag_cnt) 247 return -EINVAL; 248 249 for (i = 0; i < op_info->num_lo_sges; i++) 250 total_size += op_info->lo_sg_list[i].length; 251 252 read_fence |= info->read_fence; 253 254 if (info->imm_data_valid) 255 frag_cnt = op_info->num_lo_sges + 1; 256 else 257 frag_cnt = op_info->num_lo_sges; 258 addl_frag_cnt = frag_cnt > 1 ? (frag_cnt - 1) : 0; 259 ret_code = irdma_fragcnt_to_quanta_sq(frag_cnt, &quanta); 260 if (ret_code) 261 return ret_code; 262 263 wqe = irdma_qp_get_next_send_wqe(qp, &wqe_idx, quanta, total_size, 264 info); 265 if (!wqe) 266 return -ENOMEM; 267 268 irdma_clr_wqes(qp, wqe_idx); 269 270 set_64bit_val(wqe, 16, 271 FIELD_PREP(IRDMAQPSQ_FRAG_TO, op_info->rem_addr.addr)); 272 273 if (info->imm_data_valid) { 274 set_64bit_val(wqe, 0, 275 FIELD_PREP(IRDMAQPSQ_IMMDATA, info->imm_data)); 276 i = 0; 277 } else { 278 qp->wqe_ops.iw_set_fragment(wqe, 0, 279 op_info->lo_sg_list, 280 qp->swqe_polarity); 281 i = 1; 282 } 283 284 for (byte_off = 32; i < op_info->num_lo_sges; i++) { 285 qp->wqe_ops.iw_set_fragment(wqe, byte_off, 286 &op_info->lo_sg_list[i], 287 qp->swqe_polarity); 288 byte_off += 16; 289 } 290 291 /* if not an odd number set valid bit in next fragment */ 292 if (qp->uk_attrs->hw_rev >= IRDMA_GEN_2 && !(frag_cnt & 0x01) && 293 frag_cnt) { 294 qp->wqe_ops.iw_set_fragment(wqe, byte_off, NULL, 295 qp->swqe_polarity); 296 if (qp->uk_attrs->hw_rev == IRDMA_GEN_2) 297 ++addl_frag_cnt; 298 } 299 300 hdr = FIELD_PREP(IRDMAQPSQ_REMSTAG, op_info->rem_addr.lkey) | 301 FIELD_PREP(IRDMAQPSQ_OPCODE, info->op_type) | 302 FIELD_PREP(IRDMAQPSQ_IMMDATAFLAG, info->imm_data_valid) | 303 FIELD_PREP(IRDMAQPSQ_REPORTRTT, info->report_rtt) | 304 FIELD_PREP(IRDMAQPSQ_ADDFRAGCNT, addl_frag_cnt) | 305 FIELD_PREP(IRDMAQPSQ_READFENCE, read_fence) | 306 FIELD_PREP(IRDMAQPSQ_LOCALFENCE, info->local_fence) | 307 FIELD_PREP(IRDMAQPSQ_SIGCOMPL, info->signaled) | 308 FIELD_PREP(IRDMAQPSQ_VALID, qp->swqe_polarity); 309 310 dma_wmb(); /* make sure WQE is populated before valid bit is set */ 311 312 set_64bit_val(wqe, 24, hdr); 313 314 if (post_sq) 315 irdma_uk_qp_post_wr(qp); 316 317 return 0; 318 } 319 320 /** 321 * irdma_uk_rdma_read - rdma read command 322 * @qp: hw qp ptr 323 * @info: post sq information 324 * @inv_stag: flag for inv_stag 325 * @post_sq: flag to post sq 326 */ 327 int irdma_uk_rdma_read(struct irdma_qp_uk *qp, struct irdma_post_sq_info *info, 328 bool inv_stag, bool post_sq) 329 { 330 struct irdma_rdma_read *op_info; 331 int ret_code; 332 u32 i, byte_off, total_size = 0; 333 bool local_fence = false; 334 u32 addl_frag_cnt; 335 __le64 *wqe; 336 u32 wqe_idx; 337 u16 quanta; 338 u64 hdr; 339 340 op_info = &info->op.rdma_read; 341 if (qp->max_sq_frag_cnt < op_info->num_lo_sges) 342 return -EINVAL; 343 344 for (i = 0; i < op_info->num_lo_sges; i++) 345 total_size += op_info->lo_sg_list[i].length; 346 347 ret_code = irdma_fragcnt_to_quanta_sq(op_info->num_lo_sges, &quanta); 348 if (ret_code) 349 return ret_code; 350 351 wqe = irdma_qp_get_next_send_wqe(qp, &wqe_idx, quanta, total_size, 352 info); 353 if (!wqe) 354 return -ENOMEM; 355 356 irdma_clr_wqes(qp, wqe_idx); 357 358 addl_frag_cnt = op_info->num_lo_sges > 1 ? 359 (op_info->num_lo_sges - 1) : 0; 360 local_fence |= info->local_fence; 361 362 qp->wqe_ops.iw_set_fragment(wqe, 0, op_info->lo_sg_list, 363 qp->swqe_polarity); 364 for (i = 1, byte_off = 32; i < op_info->num_lo_sges; ++i) { 365 qp->wqe_ops.iw_set_fragment(wqe, byte_off, 366 &op_info->lo_sg_list[i], 367 qp->swqe_polarity); 368 byte_off += 16; 369 } 370 371 /* if not an odd number set valid bit in next fragment */ 372 if (qp->uk_attrs->hw_rev >= IRDMA_GEN_2 && 373 !(op_info->num_lo_sges & 0x01) && op_info->num_lo_sges) { 374 qp->wqe_ops.iw_set_fragment(wqe, byte_off, NULL, 375 qp->swqe_polarity); 376 if (qp->uk_attrs->hw_rev == IRDMA_GEN_2) 377 ++addl_frag_cnt; 378 } 379 set_64bit_val(wqe, 16, 380 FIELD_PREP(IRDMAQPSQ_FRAG_TO, op_info->rem_addr.addr)); 381 hdr = FIELD_PREP(IRDMAQPSQ_REMSTAG, op_info->rem_addr.lkey) | 382 FIELD_PREP(IRDMAQPSQ_REPORTRTT, (info->report_rtt ? 1 : 0)) | 383 FIELD_PREP(IRDMAQPSQ_ADDFRAGCNT, addl_frag_cnt) | 384 FIELD_PREP(IRDMAQPSQ_OPCODE, 385 (inv_stag ? IRDMAQP_OP_RDMA_READ_LOC_INV : IRDMAQP_OP_RDMA_READ)) | 386 FIELD_PREP(IRDMAQPSQ_READFENCE, info->read_fence) | 387 FIELD_PREP(IRDMAQPSQ_LOCALFENCE, local_fence) | 388 FIELD_PREP(IRDMAQPSQ_SIGCOMPL, info->signaled) | 389 FIELD_PREP(IRDMAQPSQ_VALID, qp->swqe_polarity); 390 391 dma_wmb(); /* make sure WQE is populated before valid bit is set */ 392 393 set_64bit_val(wqe, 24, hdr); 394 395 if (post_sq) 396 irdma_uk_qp_post_wr(qp); 397 398 return 0; 399 } 400 401 /** 402 * irdma_uk_send - rdma send command 403 * @qp: hw qp ptr 404 * @info: post sq information 405 * @post_sq: flag to post sq 406 */ 407 int irdma_uk_send(struct irdma_qp_uk *qp, struct irdma_post_sq_info *info, 408 bool post_sq) 409 { 410 __le64 *wqe; 411 struct irdma_post_send *op_info; 412 u64 hdr; 413 u32 i, wqe_idx, total_size = 0, byte_off; 414 int ret_code; 415 u32 frag_cnt, addl_frag_cnt; 416 bool read_fence = false; 417 u16 quanta; 418 419 op_info = &info->op.send; 420 if (qp->max_sq_frag_cnt < op_info->num_sges) 421 return -EINVAL; 422 423 for (i = 0; i < op_info->num_sges; i++) 424 total_size += op_info->sg_list[i].length; 425 426 if (info->imm_data_valid) 427 frag_cnt = op_info->num_sges + 1; 428 else 429 frag_cnt = op_info->num_sges; 430 ret_code = irdma_fragcnt_to_quanta_sq(frag_cnt, &quanta); 431 if (ret_code) 432 return ret_code; 433 434 wqe = irdma_qp_get_next_send_wqe(qp, &wqe_idx, quanta, total_size, 435 info); 436 if (!wqe) 437 return -ENOMEM; 438 439 irdma_clr_wqes(qp, wqe_idx); 440 441 read_fence |= info->read_fence; 442 addl_frag_cnt = frag_cnt > 1 ? (frag_cnt - 1) : 0; 443 if (info->imm_data_valid) { 444 set_64bit_val(wqe, 0, 445 FIELD_PREP(IRDMAQPSQ_IMMDATA, info->imm_data)); 446 i = 0; 447 } else { 448 qp->wqe_ops.iw_set_fragment(wqe, 0, 449 frag_cnt ? op_info->sg_list : NULL, 450 qp->swqe_polarity); 451 i = 1; 452 } 453 454 for (byte_off = 32; i < op_info->num_sges; i++) { 455 qp->wqe_ops.iw_set_fragment(wqe, byte_off, &op_info->sg_list[i], 456 qp->swqe_polarity); 457 byte_off += 16; 458 } 459 460 /* if not an odd number set valid bit in next fragment */ 461 if (qp->uk_attrs->hw_rev >= IRDMA_GEN_2 && !(frag_cnt & 0x01) && 462 frag_cnt) { 463 qp->wqe_ops.iw_set_fragment(wqe, byte_off, NULL, 464 qp->swqe_polarity); 465 if (qp->uk_attrs->hw_rev == IRDMA_GEN_2) 466 ++addl_frag_cnt; 467 } 468 469 set_64bit_val(wqe, 16, 470 FIELD_PREP(IRDMAQPSQ_DESTQKEY, op_info->qkey) | 471 FIELD_PREP(IRDMAQPSQ_DESTQPN, op_info->dest_qp)); 472 hdr = FIELD_PREP(IRDMAQPSQ_REMSTAG, info->stag_to_inv) | 473 FIELD_PREP(IRDMAQPSQ_AHID, op_info->ah_id) | 474 FIELD_PREP(IRDMAQPSQ_IMMDATAFLAG, 475 (info->imm_data_valid ? 1 : 0)) | 476 FIELD_PREP(IRDMAQPSQ_REPORTRTT, (info->report_rtt ? 1 : 0)) | 477 FIELD_PREP(IRDMAQPSQ_OPCODE, info->op_type) | 478 FIELD_PREP(IRDMAQPSQ_ADDFRAGCNT, addl_frag_cnt) | 479 FIELD_PREP(IRDMAQPSQ_READFENCE, read_fence) | 480 FIELD_PREP(IRDMAQPSQ_LOCALFENCE, info->local_fence) | 481 FIELD_PREP(IRDMAQPSQ_SIGCOMPL, info->signaled) | 482 FIELD_PREP(IRDMAQPSQ_UDPHEADER, info->udp_hdr) | 483 FIELD_PREP(IRDMAQPSQ_L4LEN, info->l4len) | 484 FIELD_PREP(IRDMAQPSQ_VALID, qp->swqe_polarity); 485 486 dma_wmb(); /* make sure WQE is populated before valid bit is set */ 487 488 set_64bit_val(wqe, 24, hdr); 489 490 if (post_sq) 491 irdma_uk_qp_post_wr(qp); 492 493 return 0; 494 } 495 496 /** 497 * irdma_set_mw_bind_wqe_gen_1 - set mw bind wqe 498 * @wqe: wqe for setting fragment 499 * @op_info: info for setting bind wqe values 500 */ 501 static void irdma_set_mw_bind_wqe_gen_1(__le64 *wqe, 502 struct irdma_bind_window *op_info) 503 { 504 set_64bit_val(wqe, 0, (uintptr_t)op_info->va); 505 set_64bit_val(wqe, 8, 506 FIELD_PREP(IRDMAQPSQ_PARENTMRSTAG, op_info->mw_stag) | 507 FIELD_PREP(IRDMAQPSQ_MWSTAG, op_info->mr_stag)); 508 set_64bit_val(wqe, 16, op_info->bind_len); 509 } 510 511 /** 512 * irdma_copy_inline_data_gen_1 - Copy inline data to wqe 513 * @wqe: pointer to wqe 514 * @sge_list: table of pointers to inline data 515 * @num_sges: Total inline data length 516 * @polarity: compatibility parameter 517 */ 518 static void irdma_copy_inline_data_gen_1(u8 *wqe, struct ib_sge *sge_list, 519 u32 num_sges, u8 polarity) 520 { 521 u32 quanta_bytes_remaining = 16; 522 int i; 523 524 for (i = 0; i < num_sges; i++) { 525 u8 *cur_sge = (u8 *)(uintptr_t)sge_list[i].addr; 526 u32 sge_len = sge_list[i].length; 527 528 while (sge_len) { 529 u32 bytes_copied; 530 531 bytes_copied = min(sge_len, quanta_bytes_remaining); 532 memcpy(wqe, cur_sge, bytes_copied); 533 wqe += bytes_copied; 534 cur_sge += bytes_copied; 535 quanta_bytes_remaining -= bytes_copied; 536 sge_len -= bytes_copied; 537 538 if (!quanta_bytes_remaining) { 539 /* Remaining inline bytes reside after hdr */ 540 wqe += 16; 541 quanta_bytes_remaining = 32; 542 } 543 } 544 } 545 } 546 547 /** 548 * irdma_inline_data_size_to_quanta_gen_1 - based on inline data, quanta 549 * @data_size: data size for inline 550 * 551 * Gets the quanta based on inline and immediate data. 552 */ 553 static inline u16 irdma_inline_data_size_to_quanta_gen_1(u32 data_size) 554 { 555 return data_size <= 16 ? IRDMA_QP_WQE_MIN_QUANTA : 2; 556 } 557 558 /** 559 * irdma_set_mw_bind_wqe - set mw bind in wqe 560 * @wqe: wqe for setting mw bind 561 * @op_info: info for setting wqe values 562 */ 563 static void irdma_set_mw_bind_wqe(__le64 *wqe, 564 struct irdma_bind_window *op_info) 565 { 566 set_64bit_val(wqe, 0, (uintptr_t)op_info->va); 567 set_64bit_val(wqe, 8, 568 FIELD_PREP(IRDMAQPSQ_PARENTMRSTAG, op_info->mr_stag) | 569 FIELD_PREP(IRDMAQPSQ_MWSTAG, op_info->mw_stag)); 570 set_64bit_val(wqe, 16, op_info->bind_len); 571 } 572 573 /** 574 * irdma_copy_inline_data - Copy inline data to wqe 575 * @wqe: pointer to wqe 576 * @sge_list: table of pointers to inline data 577 * @num_sges: number of SGE's 578 * @polarity: polarity of wqe valid bit 579 */ 580 static void irdma_copy_inline_data(u8 *wqe, struct ib_sge *sge_list, 581 u32 num_sges, u8 polarity) 582 { 583 u8 inline_valid = polarity << IRDMA_INLINE_VALID_S; 584 u32 quanta_bytes_remaining = 8; 585 bool first_quanta = true; 586 int i; 587 588 wqe += 8; 589 590 for (i = 0; i < num_sges; i++) { 591 u8 *cur_sge = (u8 *)(uintptr_t)sge_list[i].addr; 592 u32 sge_len = sge_list[i].length; 593 594 while (sge_len) { 595 u32 bytes_copied; 596 597 bytes_copied = min(sge_len, quanta_bytes_remaining); 598 memcpy(wqe, cur_sge, bytes_copied); 599 wqe += bytes_copied; 600 cur_sge += bytes_copied; 601 quanta_bytes_remaining -= bytes_copied; 602 sge_len -= bytes_copied; 603 604 if (!quanta_bytes_remaining) { 605 quanta_bytes_remaining = 31; 606 607 /* Remaining inline bytes reside after hdr */ 608 if (first_quanta) { 609 first_quanta = false; 610 wqe += 16; 611 } else { 612 *wqe = inline_valid; 613 wqe++; 614 } 615 } 616 } 617 } 618 if (!first_quanta && quanta_bytes_remaining < 31) 619 *(wqe + quanta_bytes_remaining) = inline_valid; 620 } 621 622 /** 623 * irdma_inline_data_size_to_quanta - based on inline data, quanta 624 * @data_size: data size for inline 625 * 626 * Gets the quanta based on inline and immediate data. 627 */ 628 static u16 irdma_inline_data_size_to_quanta(u32 data_size) 629 { 630 if (data_size <= 8) 631 return IRDMA_QP_WQE_MIN_QUANTA; 632 else if (data_size <= 39) 633 return 2; 634 else if (data_size <= 70) 635 return 3; 636 else if (data_size <= 101) 637 return 4; 638 else if (data_size <= 132) 639 return 5; 640 else if (data_size <= 163) 641 return 6; 642 else if (data_size <= 194) 643 return 7; 644 else 645 return 8; 646 } 647 648 /** 649 * irdma_uk_inline_rdma_write - inline rdma write operation 650 * @qp: hw qp ptr 651 * @info: post sq information 652 * @post_sq: flag to post sq 653 */ 654 int irdma_uk_inline_rdma_write(struct irdma_qp_uk *qp, 655 struct irdma_post_sq_info *info, bool post_sq) 656 { 657 __le64 *wqe; 658 struct irdma_rdma_write *op_info; 659 u64 hdr = 0; 660 u32 wqe_idx; 661 bool read_fence = false; 662 u32 i, total_size = 0; 663 u16 quanta; 664 665 op_info = &info->op.rdma_write; 666 667 if (unlikely(qp->max_sq_frag_cnt < op_info->num_lo_sges)) 668 return -EINVAL; 669 670 for (i = 0; i < op_info->num_lo_sges; i++) 671 total_size += op_info->lo_sg_list[i].length; 672 673 if (unlikely(total_size > qp->max_inline_data)) 674 return -EINVAL; 675 676 quanta = qp->wqe_ops.iw_inline_data_size_to_quanta(total_size); 677 wqe = irdma_qp_get_next_send_wqe(qp, &wqe_idx, quanta, total_size, 678 info); 679 if (!wqe) 680 return -ENOMEM; 681 682 irdma_clr_wqes(qp, wqe_idx); 683 684 read_fence |= info->read_fence; 685 set_64bit_val(wqe, 16, 686 FIELD_PREP(IRDMAQPSQ_FRAG_TO, op_info->rem_addr.addr)); 687 688 hdr = FIELD_PREP(IRDMAQPSQ_REMSTAG, op_info->rem_addr.lkey) | 689 FIELD_PREP(IRDMAQPSQ_OPCODE, info->op_type) | 690 FIELD_PREP(IRDMAQPSQ_INLINEDATALEN, total_size) | 691 FIELD_PREP(IRDMAQPSQ_REPORTRTT, info->report_rtt ? 1 : 0) | 692 FIELD_PREP(IRDMAQPSQ_INLINEDATAFLAG, 1) | 693 FIELD_PREP(IRDMAQPSQ_IMMDATAFLAG, info->imm_data_valid ? 1 : 0) | 694 FIELD_PREP(IRDMAQPSQ_READFENCE, read_fence) | 695 FIELD_PREP(IRDMAQPSQ_LOCALFENCE, info->local_fence) | 696 FIELD_PREP(IRDMAQPSQ_SIGCOMPL, info->signaled) | 697 FIELD_PREP(IRDMAQPSQ_VALID, qp->swqe_polarity); 698 699 if (info->imm_data_valid) 700 set_64bit_val(wqe, 0, 701 FIELD_PREP(IRDMAQPSQ_IMMDATA, info->imm_data)); 702 703 qp->wqe_ops.iw_copy_inline_data((u8 *)wqe, op_info->lo_sg_list, 704 op_info->num_lo_sges, 705 qp->swqe_polarity); 706 dma_wmb(); /* make sure WQE is populated before valid bit is set */ 707 708 set_64bit_val(wqe, 24, hdr); 709 710 if (post_sq) 711 irdma_uk_qp_post_wr(qp); 712 713 return 0; 714 } 715 716 /** 717 * irdma_uk_inline_send - inline send operation 718 * @qp: hw qp ptr 719 * @info: post sq information 720 * @post_sq: flag to post sq 721 */ 722 int irdma_uk_inline_send(struct irdma_qp_uk *qp, 723 struct irdma_post_sq_info *info, bool post_sq) 724 { 725 __le64 *wqe; 726 struct irdma_post_send *op_info; 727 u64 hdr; 728 u32 wqe_idx; 729 bool read_fence = false; 730 u32 i, total_size = 0; 731 u16 quanta; 732 733 op_info = &info->op.send; 734 735 if (unlikely(qp->max_sq_frag_cnt < op_info->num_sges)) 736 return -EINVAL; 737 738 for (i = 0; i < op_info->num_sges; i++) 739 total_size += op_info->sg_list[i].length; 740 741 if (unlikely(total_size > qp->max_inline_data)) 742 return -EINVAL; 743 744 quanta = qp->wqe_ops.iw_inline_data_size_to_quanta(total_size); 745 wqe = irdma_qp_get_next_send_wqe(qp, &wqe_idx, quanta, total_size, 746 info); 747 if (!wqe) 748 return -ENOMEM; 749 750 irdma_clr_wqes(qp, wqe_idx); 751 752 set_64bit_val(wqe, 16, 753 FIELD_PREP(IRDMAQPSQ_DESTQKEY, op_info->qkey) | 754 FIELD_PREP(IRDMAQPSQ_DESTQPN, op_info->dest_qp)); 755 756 read_fence |= info->read_fence; 757 hdr = FIELD_PREP(IRDMAQPSQ_REMSTAG, info->stag_to_inv) | 758 FIELD_PREP(IRDMAQPSQ_AHID, op_info->ah_id) | 759 FIELD_PREP(IRDMAQPSQ_OPCODE, info->op_type) | 760 FIELD_PREP(IRDMAQPSQ_INLINEDATALEN, total_size) | 761 FIELD_PREP(IRDMAQPSQ_IMMDATAFLAG, 762 (info->imm_data_valid ? 1 : 0)) | 763 FIELD_PREP(IRDMAQPSQ_REPORTRTT, (info->report_rtt ? 1 : 0)) | 764 FIELD_PREP(IRDMAQPSQ_INLINEDATAFLAG, 1) | 765 FIELD_PREP(IRDMAQPSQ_READFENCE, read_fence) | 766 FIELD_PREP(IRDMAQPSQ_LOCALFENCE, info->local_fence) | 767 FIELD_PREP(IRDMAQPSQ_SIGCOMPL, info->signaled) | 768 FIELD_PREP(IRDMAQPSQ_UDPHEADER, info->udp_hdr) | 769 FIELD_PREP(IRDMAQPSQ_L4LEN, info->l4len) | 770 FIELD_PREP(IRDMAQPSQ_VALID, qp->swqe_polarity); 771 772 if (info->imm_data_valid) 773 set_64bit_val(wqe, 0, 774 FIELD_PREP(IRDMAQPSQ_IMMDATA, info->imm_data)); 775 qp->wqe_ops.iw_copy_inline_data((u8 *)wqe, op_info->sg_list, 776 op_info->num_sges, qp->swqe_polarity); 777 778 dma_wmb(); /* make sure WQE is populated before valid bit is set */ 779 780 set_64bit_val(wqe, 24, hdr); 781 782 if (post_sq) 783 irdma_uk_qp_post_wr(qp); 784 785 return 0; 786 } 787 788 /** 789 * irdma_uk_stag_local_invalidate - stag invalidate operation 790 * @qp: hw qp ptr 791 * @info: post sq information 792 * @post_sq: flag to post sq 793 */ 794 int irdma_uk_stag_local_invalidate(struct irdma_qp_uk *qp, 795 struct irdma_post_sq_info *info, 796 bool post_sq) 797 { 798 __le64 *wqe; 799 struct irdma_inv_local_stag *op_info; 800 u64 hdr; 801 u32 wqe_idx; 802 bool local_fence = false; 803 struct ib_sge sge = {}; 804 805 op_info = &info->op.inv_local_stag; 806 local_fence = info->local_fence; 807 808 wqe = irdma_qp_get_next_send_wqe(qp, &wqe_idx, IRDMA_QP_WQE_MIN_QUANTA, 809 0, info); 810 if (!wqe) 811 return -ENOMEM; 812 813 irdma_clr_wqes(qp, wqe_idx); 814 815 sge.lkey = op_info->target_stag; 816 qp->wqe_ops.iw_set_fragment(wqe, 0, &sge, 0); 817 818 set_64bit_val(wqe, 16, 0); 819 820 hdr = FIELD_PREP(IRDMAQPSQ_OPCODE, IRDMA_OP_TYPE_INV_STAG) | 821 FIELD_PREP(IRDMAQPSQ_READFENCE, info->read_fence) | 822 FIELD_PREP(IRDMAQPSQ_LOCALFENCE, local_fence) | 823 FIELD_PREP(IRDMAQPSQ_SIGCOMPL, info->signaled) | 824 FIELD_PREP(IRDMAQPSQ_VALID, qp->swqe_polarity); 825 826 dma_wmb(); /* make sure WQE is populated before valid bit is set */ 827 828 set_64bit_val(wqe, 24, hdr); 829 830 if (post_sq) 831 irdma_uk_qp_post_wr(qp); 832 833 return 0; 834 } 835 836 /** 837 * irdma_uk_post_receive - post receive wqe 838 * @qp: hw qp ptr 839 * @info: post rq information 840 */ 841 int irdma_uk_post_receive(struct irdma_qp_uk *qp, 842 struct irdma_post_rq_info *info) 843 { 844 u32 wqe_idx, i, byte_off; 845 u32 addl_frag_cnt; 846 __le64 *wqe; 847 u64 hdr; 848 849 if (qp->max_rq_frag_cnt < info->num_sges) 850 return -EINVAL; 851 852 wqe = irdma_qp_get_next_recv_wqe(qp, &wqe_idx); 853 if (!wqe) 854 return -ENOMEM; 855 856 qp->rq_wrid_array[wqe_idx] = info->wr_id; 857 addl_frag_cnt = info->num_sges > 1 ? (info->num_sges - 1) : 0; 858 qp->wqe_ops.iw_set_fragment(wqe, 0, info->sg_list, 859 qp->rwqe_polarity); 860 861 for (i = 1, byte_off = 32; i < info->num_sges; i++) { 862 qp->wqe_ops.iw_set_fragment(wqe, byte_off, &info->sg_list[i], 863 qp->rwqe_polarity); 864 byte_off += 16; 865 } 866 867 /* if not an odd number set valid bit in next fragment */ 868 if (qp->uk_attrs->hw_rev >= IRDMA_GEN_2 && !(info->num_sges & 0x01) && 869 info->num_sges) { 870 qp->wqe_ops.iw_set_fragment(wqe, byte_off, NULL, 871 qp->rwqe_polarity); 872 if (qp->uk_attrs->hw_rev == IRDMA_GEN_2) 873 ++addl_frag_cnt; 874 } 875 876 set_64bit_val(wqe, 16, 0); 877 hdr = FIELD_PREP(IRDMAQPSQ_ADDFRAGCNT, addl_frag_cnt) | 878 FIELD_PREP(IRDMAQPSQ_VALID, qp->rwqe_polarity); 879 880 dma_wmb(); /* make sure WQE is populated before valid bit is set */ 881 882 set_64bit_val(wqe, 24, hdr); 883 884 return 0; 885 } 886 887 /** 888 * irdma_uk_cq_resize - reset the cq buffer info 889 * @cq: cq to resize 890 * @cq_base: new cq buffer addr 891 * @cq_size: number of cqes 892 */ 893 void irdma_uk_cq_resize(struct irdma_cq_uk *cq, void *cq_base, int cq_size) 894 { 895 cq->cq_base = cq_base; 896 cq->cq_size = cq_size; 897 IRDMA_RING_INIT(cq->cq_ring, cq->cq_size); 898 cq->polarity = 1; 899 } 900 901 /** 902 * irdma_uk_cq_set_resized_cnt - record the count of the resized buffers 903 * @cq: cq to resize 904 * @cq_cnt: the count of the resized cq buffers 905 */ 906 void irdma_uk_cq_set_resized_cnt(struct irdma_cq_uk *cq, u16 cq_cnt) 907 { 908 u64 temp_val; 909 u16 sw_cq_sel; 910 u8 arm_next_se; 911 u8 arm_next; 912 u8 arm_seq_num; 913 914 get_64bit_val(cq->shadow_area, 32, &temp_val); 915 916 sw_cq_sel = (u16)FIELD_GET(IRDMA_CQ_DBSA_SW_CQ_SELECT, temp_val); 917 sw_cq_sel += cq_cnt; 918 919 arm_seq_num = (u8)FIELD_GET(IRDMA_CQ_DBSA_ARM_SEQ_NUM, temp_val); 920 arm_next_se = (u8)FIELD_GET(IRDMA_CQ_DBSA_ARM_NEXT_SE, temp_val); 921 arm_next = (u8)FIELD_GET(IRDMA_CQ_DBSA_ARM_NEXT, temp_val); 922 923 temp_val = FIELD_PREP(IRDMA_CQ_DBSA_ARM_SEQ_NUM, arm_seq_num) | 924 FIELD_PREP(IRDMA_CQ_DBSA_SW_CQ_SELECT, sw_cq_sel) | 925 FIELD_PREP(IRDMA_CQ_DBSA_ARM_NEXT_SE, arm_next_se) | 926 FIELD_PREP(IRDMA_CQ_DBSA_ARM_NEXT, arm_next); 927 928 set_64bit_val(cq->shadow_area, 32, temp_val); 929 } 930 931 /** 932 * irdma_uk_cq_request_notification - cq notification request (door bell) 933 * @cq: hw cq 934 * @cq_notify: notification type 935 */ 936 void irdma_uk_cq_request_notification(struct irdma_cq_uk *cq, 937 enum irdma_cmpl_notify cq_notify) 938 { 939 u64 temp_val; 940 u16 sw_cq_sel; 941 u8 arm_next_se = 0; 942 u8 arm_next = 0; 943 u8 arm_seq_num; 944 945 get_64bit_val(cq->shadow_area, 32, &temp_val); 946 arm_seq_num = (u8)FIELD_GET(IRDMA_CQ_DBSA_ARM_SEQ_NUM, temp_val); 947 arm_seq_num++; 948 sw_cq_sel = (u16)FIELD_GET(IRDMA_CQ_DBSA_SW_CQ_SELECT, temp_val); 949 arm_next_se = (u8)FIELD_GET(IRDMA_CQ_DBSA_ARM_NEXT_SE, temp_val); 950 arm_next_se |= 1; 951 if (cq_notify == IRDMA_CQ_COMPL_EVENT) 952 arm_next = 1; 953 temp_val = FIELD_PREP(IRDMA_CQ_DBSA_ARM_SEQ_NUM, arm_seq_num) | 954 FIELD_PREP(IRDMA_CQ_DBSA_SW_CQ_SELECT, sw_cq_sel) | 955 FIELD_PREP(IRDMA_CQ_DBSA_ARM_NEXT_SE, arm_next_se) | 956 FIELD_PREP(IRDMA_CQ_DBSA_ARM_NEXT, arm_next); 957 958 set_64bit_val(cq->shadow_area, 32, temp_val); 959 960 dma_wmb(); /* make sure WQE is populated before valid bit is set */ 961 962 writel(cq->cq_id, cq->cqe_alloc_db); 963 } 964 965 /** 966 * irdma_uk_cq_poll_cmpl - get cq completion info 967 * @cq: hw cq 968 * @info: cq poll information returned 969 */ 970 int irdma_uk_cq_poll_cmpl(struct irdma_cq_uk *cq, 971 struct irdma_cq_poll_info *info) 972 { 973 u64 comp_ctx, qword0, qword2, qword3; 974 __le64 *cqe; 975 struct irdma_qp_uk *qp; 976 struct irdma_ring *pring = NULL; 977 u32 wqe_idx; 978 int ret_code; 979 bool move_cq_head = true; 980 u8 polarity; 981 bool ext_valid; 982 __le64 *ext_cqe; 983 984 if (cq->avoid_mem_cflct) 985 cqe = IRDMA_GET_CURRENT_EXTENDED_CQ_ELEM(cq); 986 else 987 cqe = IRDMA_GET_CURRENT_CQ_ELEM(cq); 988 989 get_64bit_val(cqe, 24, &qword3); 990 polarity = (u8)FIELD_GET(IRDMA_CQ_VALID, qword3); 991 if (polarity != cq->polarity) 992 return -ENOENT; 993 994 /* Ensure CQE contents are read after valid bit is checked */ 995 dma_rmb(); 996 997 ext_valid = (bool)FIELD_GET(IRDMA_CQ_EXTCQE, qword3); 998 if (ext_valid) { 999 u64 qword6, qword7; 1000 u32 peek_head; 1001 1002 if (cq->avoid_mem_cflct) { 1003 ext_cqe = (__le64 *)((u8 *)cqe + 32); 1004 get_64bit_val(ext_cqe, 24, &qword7); 1005 polarity = (u8)FIELD_GET(IRDMA_CQ_VALID, qword7); 1006 } else { 1007 peek_head = (cq->cq_ring.head + 1) % cq->cq_ring.size; 1008 ext_cqe = cq->cq_base[peek_head].buf; 1009 get_64bit_val(ext_cqe, 24, &qword7); 1010 polarity = (u8)FIELD_GET(IRDMA_CQ_VALID, qword7); 1011 if (!peek_head) 1012 polarity ^= 1; 1013 } 1014 if (polarity != cq->polarity) 1015 return -ENOENT; 1016 1017 /* Ensure ext CQE contents are read after ext valid bit is checked */ 1018 dma_rmb(); 1019 1020 info->imm_valid = (bool)FIELD_GET(IRDMA_CQ_IMMVALID, qword7); 1021 if (info->imm_valid) { 1022 u64 qword4; 1023 1024 get_64bit_val(ext_cqe, 0, &qword4); 1025 info->imm_data = (u32)FIELD_GET(IRDMA_CQ_IMMDATALOW32, qword4); 1026 } 1027 info->ud_smac_valid = (bool)FIELD_GET(IRDMA_CQ_UDSMACVALID, qword7); 1028 info->ud_vlan_valid = (bool)FIELD_GET(IRDMA_CQ_UDVLANVALID, qword7); 1029 if (info->ud_smac_valid || info->ud_vlan_valid) { 1030 get_64bit_val(ext_cqe, 16, &qword6); 1031 if (info->ud_vlan_valid) 1032 info->ud_vlan = (u16)FIELD_GET(IRDMA_CQ_UDVLAN, qword6); 1033 if (info->ud_smac_valid) { 1034 info->ud_smac[5] = qword6 & 0xFF; 1035 info->ud_smac[4] = (qword6 >> 8) & 0xFF; 1036 info->ud_smac[3] = (qword6 >> 16) & 0xFF; 1037 info->ud_smac[2] = (qword6 >> 24) & 0xFF; 1038 info->ud_smac[1] = (qword6 >> 32) & 0xFF; 1039 info->ud_smac[0] = (qword6 >> 40) & 0xFF; 1040 } 1041 } 1042 } else { 1043 info->imm_valid = false; 1044 info->ud_smac_valid = false; 1045 info->ud_vlan_valid = false; 1046 } 1047 1048 info->q_type = (u8)FIELD_GET(IRDMA_CQ_SQ, qword3); 1049 info->error = (bool)FIELD_GET(IRDMA_CQ_ERROR, qword3); 1050 info->ipv4 = (bool)FIELD_GET(IRDMACQ_IPV4, qword3); 1051 if (info->error) { 1052 info->major_err = FIELD_GET(IRDMA_CQ_MAJERR, qword3); 1053 info->minor_err = FIELD_GET(IRDMA_CQ_MINERR, qword3); 1054 if (info->major_err == IRDMA_FLUSH_MAJOR_ERR) { 1055 info->comp_status = IRDMA_COMPL_STATUS_FLUSHED; 1056 /* Set the min error to standard flush error code for remaining cqes */ 1057 if (info->minor_err != FLUSH_GENERAL_ERR) { 1058 qword3 &= ~IRDMA_CQ_MINERR; 1059 qword3 |= FIELD_PREP(IRDMA_CQ_MINERR, FLUSH_GENERAL_ERR); 1060 set_64bit_val(cqe, 24, qword3); 1061 } 1062 } else { 1063 info->comp_status = IRDMA_COMPL_STATUS_UNKNOWN; 1064 } 1065 } else { 1066 info->comp_status = IRDMA_COMPL_STATUS_SUCCESS; 1067 } 1068 1069 get_64bit_val(cqe, 0, &qword0); 1070 get_64bit_val(cqe, 16, &qword2); 1071 1072 info->tcp_seq_num_rtt = (u32)FIELD_GET(IRDMACQ_TCPSEQNUMRTT, qword0); 1073 info->qp_id = (u32)FIELD_GET(IRDMACQ_QPID, qword2); 1074 info->ud_src_qpn = (u32)FIELD_GET(IRDMACQ_UDSRCQPN, qword2); 1075 1076 get_64bit_val(cqe, 8, &comp_ctx); 1077 1078 info->solicited_event = (bool)FIELD_GET(IRDMACQ_SOEVENT, qword3); 1079 qp = (struct irdma_qp_uk *)(unsigned long)comp_ctx; 1080 if (!qp || qp->destroy_pending) { 1081 ret_code = -EFAULT; 1082 goto exit; 1083 } 1084 wqe_idx = (u32)FIELD_GET(IRDMA_CQ_WQEIDX, qword3); 1085 info->qp_handle = (irdma_qp_handle)(unsigned long)qp; 1086 info->op_type = (u8)FIELD_GET(IRDMACQ_OP, qword3); 1087 1088 if (info->q_type == IRDMA_CQE_QTYPE_RQ) { 1089 u32 array_idx; 1090 1091 array_idx = wqe_idx / qp->rq_wqe_size_multiplier; 1092 1093 if (info->comp_status == IRDMA_COMPL_STATUS_FLUSHED || 1094 info->comp_status == IRDMA_COMPL_STATUS_UNKNOWN) { 1095 if (!IRDMA_RING_MORE_WORK(qp->rq_ring)) { 1096 ret_code = -ENOENT; 1097 goto exit; 1098 } 1099 1100 info->wr_id = qp->rq_wrid_array[qp->rq_ring.tail]; 1101 array_idx = qp->rq_ring.tail; 1102 } else { 1103 info->wr_id = qp->rq_wrid_array[array_idx]; 1104 } 1105 1106 info->bytes_xfered = (u32)FIELD_GET(IRDMACQ_PAYLDLEN, qword0); 1107 1108 if (qword3 & IRDMACQ_STAG) { 1109 info->stag_invalid_set = true; 1110 info->inv_stag = (u32)FIELD_GET(IRDMACQ_INVSTAG, qword2); 1111 } else { 1112 info->stag_invalid_set = false; 1113 } 1114 IRDMA_RING_SET_TAIL(qp->rq_ring, array_idx + 1); 1115 if (info->comp_status == IRDMA_COMPL_STATUS_FLUSHED) { 1116 qp->rq_flush_seen = true; 1117 if (!IRDMA_RING_MORE_WORK(qp->rq_ring)) 1118 qp->rq_flush_complete = true; 1119 else 1120 move_cq_head = false; 1121 } 1122 pring = &qp->rq_ring; 1123 } else { /* q_type is IRDMA_CQE_QTYPE_SQ */ 1124 if (qp->first_sq_wq) { 1125 if (wqe_idx + 1 >= qp->conn_wqes) 1126 qp->first_sq_wq = false; 1127 1128 if (wqe_idx < qp->conn_wqes && qp->sq_ring.head == qp->sq_ring.tail) { 1129 IRDMA_RING_MOVE_HEAD_NOCHECK(cq->cq_ring); 1130 IRDMA_RING_MOVE_TAIL(cq->cq_ring); 1131 set_64bit_val(cq->shadow_area, 0, 1132 IRDMA_RING_CURRENT_HEAD(cq->cq_ring)); 1133 memset(info, 0, 1134 sizeof(struct irdma_cq_poll_info)); 1135 return irdma_uk_cq_poll_cmpl(cq, info); 1136 } 1137 } 1138 if (info->comp_status != IRDMA_COMPL_STATUS_FLUSHED) { 1139 info->wr_id = qp->sq_wrtrk_array[wqe_idx].wrid; 1140 if (!info->comp_status) 1141 info->bytes_xfered = qp->sq_wrtrk_array[wqe_idx].wr_len; 1142 info->op_type = (u8)FIELD_GET(IRDMACQ_OP, qword3); 1143 IRDMA_RING_SET_TAIL(qp->sq_ring, 1144 wqe_idx + qp->sq_wrtrk_array[wqe_idx].quanta); 1145 } else { 1146 if (!IRDMA_RING_MORE_WORK(qp->sq_ring)) { 1147 ret_code = -ENOENT; 1148 goto exit; 1149 } 1150 1151 do { 1152 __le64 *sw_wqe; 1153 u64 wqe_qword; 1154 u32 tail; 1155 1156 tail = qp->sq_ring.tail; 1157 sw_wqe = qp->sq_base[tail].elem; 1158 get_64bit_val(sw_wqe, 24, 1159 &wqe_qword); 1160 info->op_type = (u8)FIELD_GET(IRDMAQPSQ_OPCODE, 1161 wqe_qword); 1162 IRDMA_RING_SET_TAIL(qp->sq_ring, 1163 tail + qp->sq_wrtrk_array[tail].quanta); 1164 if (info->op_type != IRDMAQP_OP_NOP) { 1165 info->wr_id = qp->sq_wrtrk_array[tail].wrid; 1166 info->bytes_xfered = qp->sq_wrtrk_array[tail].wr_len; 1167 break; 1168 } 1169 } while (1); 1170 if (info->op_type == IRDMA_OP_TYPE_BIND_MW && 1171 info->minor_err == FLUSH_PROT_ERR) 1172 info->minor_err = FLUSH_MW_BIND_ERR; 1173 qp->sq_flush_seen = true; 1174 if (!IRDMA_RING_MORE_WORK(qp->sq_ring)) 1175 qp->sq_flush_complete = true; 1176 } 1177 pring = &qp->sq_ring; 1178 } 1179 1180 ret_code = 0; 1181 1182 exit: 1183 if (!ret_code && info->comp_status == IRDMA_COMPL_STATUS_FLUSHED) 1184 if (pring && IRDMA_RING_MORE_WORK(*pring)) 1185 move_cq_head = false; 1186 1187 if (move_cq_head) { 1188 IRDMA_RING_MOVE_HEAD_NOCHECK(cq->cq_ring); 1189 if (!IRDMA_RING_CURRENT_HEAD(cq->cq_ring)) 1190 cq->polarity ^= 1; 1191 1192 if (ext_valid && !cq->avoid_mem_cflct) { 1193 IRDMA_RING_MOVE_HEAD_NOCHECK(cq->cq_ring); 1194 if (!IRDMA_RING_CURRENT_HEAD(cq->cq_ring)) 1195 cq->polarity ^= 1; 1196 } 1197 1198 IRDMA_RING_MOVE_TAIL(cq->cq_ring); 1199 if (!cq->avoid_mem_cflct && ext_valid) 1200 IRDMA_RING_MOVE_TAIL(cq->cq_ring); 1201 set_64bit_val(cq->shadow_area, 0, 1202 IRDMA_RING_CURRENT_HEAD(cq->cq_ring)); 1203 } else { 1204 qword3 &= ~IRDMA_CQ_WQEIDX; 1205 qword3 |= FIELD_PREP(IRDMA_CQ_WQEIDX, pring->tail); 1206 set_64bit_val(cqe, 24, qword3); 1207 } 1208 1209 return ret_code; 1210 } 1211 1212 /** 1213 * irdma_qp_round_up - return round up qp wq depth 1214 * @wqdepth: wq depth in quanta to round up 1215 */ 1216 static int irdma_qp_round_up(u32 wqdepth) 1217 { 1218 int scount = 1; 1219 1220 for (wqdepth--; scount <= 16; scount *= 2) 1221 wqdepth |= wqdepth >> scount; 1222 1223 return ++wqdepth; 1224 } 1225 1226 /** 1227 * irdma_get_wqe_shift - get shift count for maximum wqe size 1228 * @uk_attrs: qp HW attributes 1229 * @sge: Maximum Scatter Gather Elements wqe 1230 * @inline_data: Maximum inline data size 1231 * @shift: Returns the shift needed based on sge 1232 * 1233 * Shift can be used to left shift the wqe size based on number of SGEs and inlind data size. 1234 * For 1 SGE or inline data <= 8, shift = 0 (wqe size of 32 1235 * bytes). For 2 or 3 SGEs or inline data <= 39, shift = 1 (wqe 1236 * size of 64 bytes). 1237 * For 4-7 SGE's and inline <= 101 Shift of 2 otherwise (wqe 1238 * size of 256 bytes). 1239 */ 1240 void irdma_get_wqe_shift(struct irdma_uk_attrs *uk_attrs, u32 sge, 1241 u32 inline_data, u8 *shift) 1242 { 1243 *shift = 0; 1244 if (uk_attrs->hw_rev >= IRDMA_GEN_2) { 1245 if (sge > 1 || inline_data > 8) { 1246 if (sge < 4 && inline_data <= 39) 1247 *shift = 1; 1248 else if (sge < 8 && inline_data <= 101) 1249 *shift = 2; 1250 else 1251 *shift = 3; 1252 } 1253 } else if (sge > 1 || inline_data > 16) { 1254 *shift = (sge < 4 && inline_data <= 48) ? 1 : 2; 1255 } 1256 } 1257 1258 /* 1259 * irdma_get_sqdepth - get SQ depth (quanta) 1260 * @uk_attrs: qp HW attributes 1261 * @sq_size: SQ size 1262 * @shift: shift which determines size of WQE 1263 * @sqdepth: depth of SQ 1264 * 1265 */ 1266 int irdma_get_sqdepth(struct irdma_uk_attrs *uk_attrs, u32 sq_size, u8 shift, 1267 u32 *sqdepth) 1268 { 1269 u32 min_size = (u32)uk_attrs->min_hw_wq_size << shift; 1270 1271 *sqdepth = irdma_qp_round_up((sq_size << shift) + IRDMA_SQ_RSVD); 1272 1273 if (*sqdepth < min_size) 1274 *sqdepth = min_size; 1275 else if (*sqdepth > uk_attrs->max_hw_wq_quanta) 1276 return -EINVAL; 1277 1278 return 0; 1279 } 1280 1281 /* 1282 * irdma_get_rqdepth - get RQ depth (quanta) 1283 * @uk_attrs: qp HW attributes 1284 * @rq_size: RQ size 1285 * @shift: shift which determines size of WQE 1286 * @rqdepth: depth of RQ 1287 */ 1288 int irdma_get_rqdepth(struct irdma_uk_attrs *uk_attrs, u32 rq_size, u8 shift, 1289 u32 *rqdepth) 1290 { 1291 u32 min_size = (u32)uk_attrs->min_hw_wq_size << shift; 1292 1293 *rqdepth = irdma_qp_round_up((rq_size << shift) + IRDMA_RQ_RSVD); 1294 1295 if (*rqdepth < min_size) 1296 *rqdepth = min_size; 1297 else if (*rqdepth > uk_attrs->max_hw_rq_quanta) 1298 return -EINVAL; 1299 1300 return 0; 1301 } 1302 1303 static const struct irdma_wqe_uk_ops iw_wqe_uk_ops = { 1304 .iw_copy_inline_data = irdma_copy_inline_data, 1305 .iw_inline_data_size_to_quanta = irdma_inline_data_size_to_quanta, 1306 .iw_set_fragment = irdma_set_fragment, 1307 .iw_set_mw_bind_wqe = irdma_set_mw_bind_wqe, 1308 }; 1309 1310 static const struct irdma_wqe_uk_ops iw_wqe_uk_ops_gen_1 = { 1311 .iw_copy_inline_data = irdma_copy_inline_data_gen_1, 1312 .iw_inline_data_size_to_quanta = irdma_inline_data_size_to_quanta_gen_1, 1313 .iw_set_fragment = irdma_set_fragment_gen_1, 1314 .iw_set_mw_bind_wqe = irdma_set_mw_bind_wqe_gen_1, 1315 }; 1316 1317 /** 1318 * irdma_setup_connection_wqes - setup WQEs necessary to complete 1319 * connection. 1320 * @qp: hw qp (user and kernel) 1321 * @info: qp initialization info 1322 */ 1323 static void irdma_setup_connection_wqes(struct irdma_qp_uk *qp, 1324 struct irdma_qp_uk_init_info *info) 1325 { 1326 u16 move_cnt = 1; 1327 1328 if (!info->legacy_mode && 1329 (qp->uk_attrs->feature_flags & IRDMA_FEATURE_RTS_AE)) 1330 move_cnt = 3; 1331 1332 qp->conn_wqes = move_cnt; 1333 IRDMA_RING_MOVE_HEAD_BY_COUNT_NOCHECK(qp->sq_ring, move_cnt); 1334 IRDMA_RING_MOVE_TAIL_BY_COUNT(qp->sq_ring, move_cnt); 1335 IRDMA_RING_MOVE_HEAD_BY_COUNT_NOCHECK(qp->initial_ring, move_cnt); 1336 } 1337 1338 /** 1339 * irdma_uk_calc_shift_wq - calculate WQE shift for both SQ and RQ 1340 * @ukinfo: qp initialization info 1341 * @sq_shift: Returns shift of SQ 1342 * @rq_shift: Returns shift of RQ 1343 */ 1344 void irdma_uk_calc_shift_wq(struct irdma_qp_uk_init_info *ukinfo, u8 *sq_shift, 1345 u8 *rq_shift) 1346 { 1347 bool imm_support = ukinfo->uk_attrs->hw_rev >= IRDMA_GEN_2; 1348 1349 irdma_get_wqe_shift(ukinfo->uk_attrs, 1350 imm_support ? ukinfo->max_sq_frag_cnt + 1 : 1351 ukinfo->max_sq_frag_cnt, 1352 ukinfo->max_inline_data, sq_shift); 1353 1354 irdma_get_wqe_shift(ukinfo->uk_attrs, ukinfo->max_rq_frag_cnt, 0, 1355 rq_shift); 1356 1357 if (ukinfo->uk_attrs->hw_rev == IRDMA_GEN_1) { 1358 if (ukinfo->abi_ver > 4) 1359 *rq_shift = IRDMA_MAX_RQ_WQE_SHIFT_GEN1; 1360 } 1361 } 1362 1363 /** 1364 * irdma_uk_calc_depth_shift_sq - calculate depth and shift for SQ size. 1365 * @ukinfo: qp initialization info 1366 * @sq_depth: Returns depth of SQ 1367 * @sq_shift: Returns shift of SQ 1368 */ 1369 int irdma_uk_calc_depth_shift_sq(struct irdma_qp_uk_init_info *ukinfo, 1370 u32 *sq_depth, u8 *sq_shift) 1371 { 1372 bool imm_support = ukinfo->uk_attrs->hw_rev >= IRDMA_GEN_2; 1373 int status; 1374 1375 irdma_get_wqe_shift(ukinfo->uk_attrs, 1376 imm_support ? ukinfo->max_sq_frag_cnt + 1 : 1377 ukinfo->max_sq_frag_cnt, 1378 ukinfo->max_inline_data, sq_shift); 1379 status = irdma_get_sqdepth(ukinfo->uk_attrs, ukinfo->sq_size, 1380 *sq_shift, sq_depth); 1381 1382 return status; 1383 } 1384 1385 /** 1386 * irdma_uk_calc_depth_shift_rq - calculate depth and shift for RQ size. 1387 * @ukinfo: qp initialization info 1388 * @rq_depth: Returns depth of RQ 1389 * @rq_shift: Returns shift of RQ 1390 */ 1391 int irdma_uk_calc_depth_shift_rq(struct irdma_qp_uk_init_info *ukinfo, 1392 u32 *rq_depth, u8 *rq_shift) 1393 { 1394 int status; 1395 1396 irdma_get_wqe_shift(ukinfo->uk_attrs, ukinfo->max_rq_frag_cnt, 0, 1397 rq_shift); 1398 1399 if (ukinfo->uk_attrs->hw_rev == IRDMA_GEN_1) { 1400 if (ukinfo->abi_ver > 4) 1401 *rq_shift = IRDMA_MAX_RQ_WQE_SHIFT_GEN1; 1402 } 1403 1404 status = irdma_get_rqdepth(ukinfo->uk_attrs, ukinfo->rq_size, 1405 *rq_shift, rq_depth); 1406 1407 return status; 1408 } 1409 1410 /** 1411 * irdma_uk_qp_init - initialize shared qp 1412 * @qp: hw qp (user and kernel) 1413 * @info: qp initialization info 1414 * 1415 * initializes the vars used in both user and kernel mode. 1416 * size of the wqe depends on numbers of max. fragements 1417 * allowed. Then size of wqe * the number of wqes should be the 1418 * amount of memory allocated for sq and rq. 1419 */ 1420 int irdma_uk_qp_init(struct irdma_qp_uk *qp, struct irdma_qp_uk_init_info *info) 1421 { 1422 int ret_code = 0; 1423 u32 sq_ring_size; 1424 1425 qp->uk_attrs = info->uk_attrs; 1426 if (info->max_sq_frag_cnt > qp->uk_attrs->max_hw_wq_frags || 1427 info->max_rq_frag_cnt > qp->uk_attrs->max_hw_wq_frags) 1428 return -EINVAL; 1429 1430 qp->qp_caps = info->qp_caps; 1431 qp->sq_base = info->sq; 1432 qp->rq_base = info->rq; 1433 qp->qp_type = info->type ? info->type : IRDMA_QP_TYPE_IWARP; 1434 qp->shadow_area = info->shadow_area; 1435 qp->sq_wrtrk_array = info->sq_wrtrk_array; 1436 1437 qp->rq_wrid_array = info->rq_wrid_array; 1438 qp->wqe_alloc_db = info->wqe_alloc_db; 1439 qp->qp_id = info->qp_id; 1440 qp->sq_size = info->sq_size; 1441 qp->max_sq_frag_cnt = info->max_sq_frag_cnt; 1442 sq_ring_size = qp->sq_size << info->sq_shift; 1443 IRDMA_RING_INIT(qp->sq_ring, sq_ring_size); 1444 IRDMA_RING_INIT(qp->initial_ring, sq_ring_size); 1445 if (info->first_sq_wq) { 1446 irdma_setup_connection_wqes(qp, info); 1447 qp->swqe_polarity = 1; 1448 qp->first_sq_wq = true; 1449 } else { 1450 qp->swqe_polarity = 0; 1451 } 1452 qp->swqe_polarity_deferred = 1; 1453 qp->rwqe_polarity = 0; 1454 qp->rq_size = info->rq_size; 1455 qp->max_rq_frag_cnt = info->max_rq_frag_cnt; 1456 qp->max_inline_data = info->max_inline_data; 1457 qp->rq_wqe_size = info->rq_shift; 1458 IRDMA_RING_INIT(qp->rq_ring, qp->rq_size); 1459 qp->rq_wqe_size_multiplier = 1 << info->rq_shift; 1460 if (qp->uk_attrs->hw_rev == IRDMA_GEN_1) 1461 qp->wqe_ops = iw_wqe_uk_ops_gen_1; 1462 else 1463 qp->wqe_ops = iw_wqe_uk_ops; 1464 return ret_code; 1465 } 1466 1467 /** 1468 * irdma_uk_cq_init - initialize shared cq (user and kernel) 1469 * @cq: hw cq 1470 * @info: hw cq initialization info 1471 */ 1472 void irdma_uk_cq_init(struct irdma_cq_uk *cq, 1473 struct irdma_cq_uk_init_info *info) 1474 { 1475 cq->cq_base = info->cq_base; 1476 cq->cq_id = info->cq_id; 1477 cq->cq_size = info->cq_size; 1478 cq->cqe_alloc_db = info->cqe_alloc_db; 1479 cq->cq_ack_db = info->cq_ack_db; 1480 cq->shadow_area = info->shadow_area; 1481 cq->avoid_mem_cflct = info->avoid_mem_cflct; 1482 IRDMA_RING_INIT(cq->cq_ring, cq->cq_size); 1483 cq->polarity = 1; 1484 } 1485 1486 /** 1487 * irdma_uk_clean_cq - clean cq entries 1488 * @q: completion context 1489 * @cq: cq to clean 1490 */ 1491 void irdma_uk_clean_cq(void *q, struct irdma_cq_uk *cq) 1492 { 1493 __le64 *cqe; 1494 u64 qword3, comp_ctx; 1495 u32 cq_head; 1496 u8 polarity, temp; 1497 1498 cq_head = cq->cq_ring.head; 1499 temp = cq->polarity; 1500 do { 1501 if (cq->avoid_mem_cflct) 1502 cqe = ((struct irdma_extended_cqe *)(cq->cq_base))[cq_head].buf; 1503 else 1504 cqe = cq->cq_base[cq_head].buf; 1505 get_64bit_val(cqe, 24, &qword3); 1506 polarity = (u8)FIELD_GET(IRDMA_CQ_VALID, qword3); 1507 1508 if (polarity != temp) 1509 break; 1510 1511 /* Ensure CQE contents are read after valid bit is checked */ 1512 dma_rmb(); 1513 1514 get_64bit_val(cqe, 8, &comp_ctx); 1515 if ((void *)(unsigned long)comp_ctx == q) 1516 set_64bit_val(cqe, 8, 0); 1517 1518 cq_head = (cq_head + 1) % cq->cq_ring.size; 1519 if (!cq_head) 1520 temp ^= 1; 1521 } while (true); 1522 } 1523 1524 /** 1525 * irdma_nop - post a nop 1526 * @qp: hw qp ptr 1527 * @wr_id: work request id 1528 * @signaled: signaled for completion 1529 * @post_sq: ring doorbell 1530 */ 1531 int irdma_nop(struct irdma_qp_uk *qp, u64 wr_id, bool signaled, bool post_sq) 1532 { 1533 __le64 *wqe; 1534 u64 hdr; 1535 u32 wqe_idx; 1536 struct irdma_post_sq_info info = {}; 1537 1538 info.wr_id = wr_id; 1539 wqe = irdma_qp_get_next_send_wqe(qp, &wqe_idx, IRDMA_QP_WQE_MIN_QUANTA, 1540 0, &info); 1541 if (!wqe) 1542 return -ENOMEM; 1543 1544 irdma_clr_wqes(qp, wqe_idx); 1545 1546 set_64bit_val(wqe, 0, 0); 1547 set_64bit_val(wqe, 8, 0); 1548 set_64bit_val(wqe, 16, 0); 1549 1550 hdr = FIELD_PREP(IRDMAQPSQ_OPCODE, IRDMAQP_OP_NOP) | 1551 FIELD_PREP(IRDMAQPSQ_SIGCOMPL, signaled) | 1552 FIELD_PREP(IRDMAQPSQ_VALID, qp->swqe_polarity); 1553 1554 dma_wmb(); /* make sure WQE is populated before valid bit is set */ 1555 1556 set_64bit_val(wqe, 24, hdr); 1557 if (post_sq) 1558 irdma_uk_qp_post_wr(qp); 1559 1560 return 0; 1561 } 1562 1563 /** 1564 * irdma_fragcnt_to_quanta_sq - calculate quanta based on fragment count for SQ 1565 * @frag_cnt: number of fragments 1566 * @quanta: quanta for frag_cnt 1567 */ 1568 int irdma_fragcnt_to_quanta_sq(u32 frag_cnt, u16 *quanta) 1569 { 1570 switch (frag_cnt) { 1571 case 0: 1572 case 1: 1573 *quanta = IRDMA_QP_WQE_MIN_QUANTA; 1574 break; 1575 case 2: 1576 case 3: 1577 *quanta = 2; 1578 break; 1579 case 4: 1580 case 5: 1581 *quanta = 3; 1582 break; 1583 case 6: 1584 case 7: 1585 *quanta = 4; 1586 break; 1587 case 8: 1588 case 9: 1589 *quanta = 5; 1590 break; 1591 case 10: 1592 case 11: 1593 *quanta = 6; 1594 break; 1595 case 12: 1596 case 13: 1597 *quanta = 7; 1598 break; 1599 case 14: 1600 case 15: /* when immediate data is present */ 1601 *quanta = 8; 1602 break; 1603 default: 1604 return -EINVAL; 1605 } 1606 1607 return 0; 1608 } 1609 1610 /** 1611 * irdma_fragcnt_to_wqesize_rq - calculate wqe size based on fragment count for RQ 1612 * @frag_cnt: number of fragments 1613 * @wqe_size: size in bytes given frag_cnt 1614 */ 1615 int irdma_fragcnt_to_wqesize_rq(u32 frag_cnt, u16 *wqe_size) 1616 { 1617 switch (frag_cnt) { 1618 case 0: 1619 case 1: 1620 *wqe_size = 32; 1621 break; 1622 case 2: 1623 case 3: 1624 *wqe_size = 64; 1625 break; 1626 case 4: 1627 case 5: 1628 case 6: 1629 case 7: 1630 *wqe_size = 128; 1631 break; 1632 case 8: 1633 case 9: 1634 case 10: 1635 case 11: 1636 case 12: 1637 case 13: 1638 case 14: 1639 *wqe_size = 256; 1640 break; 1641 default: 1642 return -EINVAL; 1643 } 1644 1645 return 0; 1646 } 1647