1 // SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause 2 3 /* Authors: Bernard Metzler <bmt@zurich.ibm.com> */ 4 /* Copyright (c) 2008-2019, IBM Corporation */ 5 6 #include <linux/errno.h> 7 #include <linux/types.h> 8 #include <linux/net.h> 9 #include <linux/scatterlist.h> 10 #include <linux/highmem.h> 11 12 #include <rdma/iw_cm.h> 13 #include <rdma/ib_verbs.h> 14 15 #include "siw.h" 16 #include "siw_verbs.h" 17 #include "siw_mem.h" 18 19 /* 20 * siw_rx_umem() 21 * 22 * Receive data of @len into target referenced by @dest_addr. 23 * 24 * @srx: Receive Context 25 * @umem: siw representation of target memory 26 * @dest_addr: user virtual address 27 * @len: number of bytes to place 28 */ 29 static int siw_rx_umem(struct siw_rx_stream *srx, struct siw_umem *umem, 30 u64 dest_addr, int len) 31 { 32 int copied = 0; 33 34 while (len) { 35 struct page *p; 36 int pg_off, bytes, rv; 37 void *dest; 38 39 p = siw_get_upage(umem, dest_addr); 40 if (unlikely(!p)) { 41 pr_warn("siw: %s: [QP %u]: bogus addr: %pK, %pK\n", 42 __func__, qp_id(rx_qp(srx)), 43 (void *)(uintptr_t)dest_addr, 44 (void *)(uintptr_t)umem->fp_addr); 45 /* siw internal error */ 46 srx->skb_copied += copied; 47 srx->skb_new -= copied; 48 49 return -EFAULT; 50 } 51 pg_off = dest_addr & ~PAGE_MASK; 52 bytes = min(len, (int)PAGE_SIZE - pg_off); 53 54 siw_dbg_qp(rx_qp(srx), "page %pK, bytes=%u\n", p, bytes); 55 56 dest = kmap_atomic(p); 57 rv = skb_copy_bits(srx->skb, srx->skb_offset, dest + pg_off, 58 bytes); 59 60 if (unlikely(rv)) { 61 kunmap_atomic(dest); 62 srx->skb_copied += copied; 63 srx->skb_new -= copied; 64 65 pr_warn("siw: [QP %u]: %s, len %d, page %p, rv %d\n", 66 qp_id(rx_qp(srx)), __func__, len, p, rv); 67 68 return -EFAULT; 69 } 70 if (srx->mpa_crc_hd) { 71 if (rdma_is_kernel_res(&rx_qp(srx)->base_qp.res)) { 72 crypto_shash_update(srx->mpa_crc_hd, 73 (u8 *)(dest + pg_off), bytes); 74 kunmap_atomic(dest); 75 } else { 76 kunmap_atomic(dest); 77 /* 78 * Do CRC on original, not target buffer. 79 * Some user land applications may 80 * concurrently write the target buffer, 81 * which would yield a broken CRC. 82 * Walking the skb twice is very ineffcient. 83 * Folding the CRC into skb_copy_bits() 84 * would be much better, but is currently 85 * not supported. 86 */ 87 siw_crc_skb(srx, bytes); 88 } 89 } else { 90 kunmap_atomic(dest); 91 } 92 srx->skb_offset += bytes; 93 copied += bytes; 94 len -= bytes; 95 dest_addr += bytes; 96 pg_off = 0; 97 } 98 srx->skb_copied += copied; 99 srx->skb_new -= copied; 100 101 return copied; 102 } 103 104 static int siw_rx_kva(struct siw_rx_stream *srx, void *kva, int len) 105 { 106 int rv; 107 108 siw_dbg_qp(rx_qp(srx), "kva: 0x%pK, len: %u\n", kva, len); 109 110 rv = skb_copy_bits(srx->skb, srx->skb_offset, kva, len); 111 if (unlikely(rv)) { 112 pr_warn("siw: [QP %u]: %s, len %d, kva 0x%pK, rv %d\n", 113 qp_id(rx_qp(srx)), __func__, len, kva, rv); 114 115 return rv; 116 } 117 if (srx->mpa_crc_hd) 118 crypto_shash_update(srx->mpa_crc_hd, (u8 *)kva, len); 119 120 srx->skb_offset += len; 121 srx->skb_copied += len; 122 srx->skb_new -= len; 123 124 return len; 125 } 126 127 static int siw_rx_pbl(struct siw_rx_stream *srx, int *pbl_idx, 128 struct siw_mem *mem, u64 addr, int len) 129 { 130 struct siw_pbl *pbl = mem->pbl; 131 u64 offset = addr - mem->va; 132 int copied = 0; 133 134 while (len) { 135 int bytes; 136 dma_addr_t buf_addr = 137 siw_pbl_get_buffer(pbl, offset, &bytes, pbl_idx); 138 if (!buf_addr) 139 break; 140 141 bytes = min(bytes, len); 142 if (siw_rx_kva(srx, (void *)(uintptr_t)buf_addr, bytes) == 143 bytes) { 144 copied += bytes; 145 offset += bytes; 146 len -= bytes; 147 } else { 148 break; 149 } 150 } 151 return copied; 152 } 153 154 /* 155 * siw_rresp_check_ntoh() 156 * 157 * Check incoming RRESP fragment header against expected 158 * header values and update expected values for potential next 159 * fragment. 160 * 161 * NOTE: This function must be called only if a RRESP DDP segment 162 * starts but not for fragmented consecutive pieces of an 163 * already started DDP segment. 164 */ 165 static int siw_rresp_check_ntoh(struct siw_rx_stream *srx, 166 struct siw_rx_fpdu *frx) 167 { 168 struct iwarp_rdma_rresp *rresp = &srx->hdr.rresp; 169 struct siw_wqe *wqe = &frx->wqe_active; 170 enum ddp_ecode ecode; 171 172 u32 sink_stag = be32_to_cpu(rresp->sink_stag); 173 u64 sink_to = be64_to_cpu(rresp->sink_to); 174 175 if (frx->first_ddp_seg) { 176 srx->ddp_stag = wqe->sqe.sge[0].lkey; 177 srx->ddp_to = wqe->sqe.sge[0].laddr; 178 frx->pbl_idx = 0; 179 } 180 /* Below checks extend beyond the semantics of DDP, and 181 * into RDMAP: 182 * We check if the read response matches exactly the 183 * read request which was send to the remote peer to 184 * trigger this read response. RFC5040/5041 do not 185 * always have a proper error code for the detected 186 * error cases. We choose 'base or bounds error' for 187 * cases where the inbound STag is valid, but offset 188 * or length do not match our response receive state. 189 */ 190 if (unlikely(srx->ddp_stag != sink_stag)) { 191 pr_warn("siw: [QP %u]: rresp stag: %08x != %08x\n", 192 qp_id(rx_qp(srx)), sink_stag, srx->ddp_stag); 193 ecode = DDP_ECODE_T_INVALID_STAG; 194 goto error; 195 } 196 if (unlikely(srx->ddp_to != sink_to)) { 197 pr_warn("siw: [QP %u]: rresp off: %016llx != %016llx\n", 198 qp_id(rx_qp(srx)), (unsigned long long)sink_to, 199 (unsigned long long)srx->ddp_to); 200 ecode = DDP_ECODE_T_BASE_BOUNDS; 201 goto error; 202 } 203 if (unlikely(!frx->more_ddp_segs && 204 (wqe->processed + srx->fpdu_part_rem != wqe->bytes))) { 205 pr_warn("siw: [QP %u]: rresp len: %d != %d\n", 206 qp_id(rx_qp(srx)), 207 wqe->processed + srx->fpdu_part_rem, wqe->bytes); 208 ecode = DDP_ECODE_T_BASE_BOUNDS; 209 goto error; 210 } 211 return 0; 212 error: 213 siw_init_terminate(rx_qp(srx), TERM_ERROR_LAYER_DDP, 214 DDP_ETYPE_TAGGED_BUF, ecode, 0); 215 return -EINVAL; 216 } 217 218 /* 219 * siw_write_check_ntoh() 220 * 221 * Check incoming WRITE fragment header against expected 222 * header values and update expected values for potential next 223 * fragment 224 * 225 * NOTE: This function must be called only if a WRITE DDP segment 226 * starts but not for fragmented consecutive pieces of an 227 * already started DDP segment. 228 */ 229 static int siw_write_check_ntoh(struct siw_rx_stream *srx, 230 struct siw_rx_fpdu *frx) 231 { 232 struct iwarp_rdma_write *write = &srx->hdr.rwrite; 233 enum ddp_ecode ecode; 234 235 u32 sink_stag = be32_to_cpu(write->sink_stag); 236 u64 sink_to = be64_to_cpu(write->sink_to); 237 238 if (frx->first_ddp_seg) { 239 srx->ddp_stag = sink_stag; 240 srx->ddp_to = sink_to; 241 frx->pbl_idx = 0; 242 } else { 243 if (unlikely(srx->ddp_stag != sink_stag)) { 244 pr_warn("siw: [QP %u]: write stag: %08x != %08x\n", 245 qp_id(rx_qp(srx)), sink_stag, 246 srx->ddp_stag); 247 ecode = DDP_ECODE_T_INVALID_STAG; 248 goto error; 249 } 250 if (unlikely(srx->ddp_to != sink_to)) { 251 pr_warn("siw: [QP %u]: write off: %016llx != %016llx\n", 252 qp_id(rx_qp(srx)), 253 (unsigned long long)sink_to, 254 (unsigned long long)srx->ddp_to); 255 ecode = DDP_ECODE_T_BASE_BOUNDS; 256 goto error; 257 } 258 } 259 return 0; 260 error: 261 siw_init_terminate(rx_qp(srx), TERM_ERROR_LAYER_DDP, 262 DDP_ETYPE_TAGGED_BUF, ecode, 0); 263 return -EINVAL; 264 } 265 266 /* 267 * siw_send_check_ntoh() 268 * 269 * Check incoming SEND fragment header against expected 270 * header values and update expected MSN if no next 271 * fragment expected 272 * 273 * NOTE: This function must be called only if a SEND DDP segment 274 * starts but not for fragmented consecutive pieces of an 275 * already started DDP segment. 276 */ 277 static int siw_send_check_ntoh(struct siw_rx_stream *srx, 278 struct siw_rx_fpdu *frx) 279 { 280 struct iwarp_send_inv *send = &srx->hdr.send_inv; 281 struct siw_wqe *wqe = &frx->wqe_active; 282 enum ddp_ecode ecode; 283 284 u32 ddp_msn = be32_to_cpu(send->ddp_msn); 285 u32 ddp_mo = be32_to_cpu(send->ddp_mo); 286 u32 ddp_qn = be32_to_cpu(send->ddp_qn); 287 288 if (unlikely(ddp_qn != RDMAP_UNTAGGED_QN_SEND)) { 289 pr_warn("siw: [QP %u]: invalid ddp qn %d for send\n", 290 qp_id(rx_qp(srx)), ddp_qn); 291 ecode = DDP_ECODE_UT_INVALID_QN; 292 goto error; 293 } 294 if (unlikely(ddp_msn != srx->ddp_msn[RDMAP_UNTAGGED_QN_SEND])) { 295 pr_warn("siw: [QP %u]: send msn: %u != %u\n", 296 qp_id(rx_qp(srx)), ddp_msn, 297 srx->ddp_msn[RDMAP_UNTAGGED_QN_SEND]); 298 ecode = DDP_ECODE_UT_INVALID_MSN_RANGE; 299 goto error; 300 } 301 if (unlikely(ddp_mo != wqe->processed)) { 302 pr_warn("siw: [QP %u], send mo: %u != %u\n", 303 qp_id(rx_qp(srx)), ddp_mo, wqe->processed); 304 ecode = DDP_ECODE_UT_INVALID_MO; 305 goto error; 306 } 307 if (frx->first_ddp_seg) { 308 /* initialize user memory write position */ 309 frx->sge_idx = 0; 310 frx->sge_off = 0; 311 frx->pbl_idx = 0; 312 313 /* only valid for SEND_INV and SEND_SE_INV operations */ 314 srx->inval_stag = be32_to_cpu(send->inval_stag); 315 } 316 if (unlikely(wqe->bytes < wqe->processed + srx->fpdu_part_rem)) { 317 siw_dbg_qp(rx_qp(srx), "receive space short: %d - %d < %d\n", 318 wqe->bytes, wqe->processed, srx->fpdu_part_rem); 319 wqe->wc_status = SIW_WC_LOC_LEN_ERR; 320 ecode = DDP_ECODE_UT_INVALID_MSN_NOBUF; 321 goto error; 322 } 323 return 0; 324 error: 325 siw_init_terminate(rx_qp(srx), TERM_ERROR_LAYER_DDP, 326 DDP_ETYPE_UNTAGGED_BUF, ecode, 0); 327 return -EINVAL; 328 } 329 330 static struct siw_wqe *siw_rqe_get(struct siw_qp *qp) 331 { 332 struct siw_rqe *rqe; 333 struct siw_srq *srq; 334 struct siw_wqe *wqe = NULL; 335 bool srq_event = false; 336 unsigned long flags; 337 338 srq = qp->srq; 339 if (srq) { 340 spin_lock_irqsave(&srq->lock, flags); 341 if (unlikely(!srq->num_rqe)) 342 goto out; 343 344 rqe = &srq->recvq[srq->rq_get % srq->num_rqe]; 345 } else { 346 if (unlikely(!qp->recvq)) 347 goto out; 348 349 rqe = &qp->recvq[qp->rq_get % qp->attrs.rq_size]; 350 } 351 if (likely(rqe->flags == SIW_WQE_VALID)) { 352 int num_sge = rqe->num_sge; 353 354 if (likely(num_sge <= SIW_MAX_SGE)) { 355 int i = 0; 356 357 wqe = rx_wqe(&qp->rx_untagged); 358 rx_type(wqe) = SIW_OP_RECEIVE; 359 wqe->wr_status = SIW_WR_INPROGRESS; 360 wqe->bytes = 0; 361 wqe->processed = 0; 362 363 wqe->rqe.id = rqe->id; 364 wqe->rqe.num_sge = num_sge; 365 366 while (i < num_sge) { 367 wqe->rqe.sge[i].laddr = rqe->sge[i].laddr; 368 wqe->rqe.sge[i].lkey = rqe->sge[i].lkey; 369 wqe->rqe.sge[i].length = rqe->sge[i].length; 370 wqe->bytes += wqe->rqe.sge[i].length; 371 wqe->mem[i] = NULL; 372 i++; 373 } 374 /* can be re-used by appl */ 375 smp_store_mb(rqe->flags, 0); 376 } else { 377 siw_dbg_qp(qp, "too many sge's: %d\n", rqe->num_sge); 378 if (srq) 379 spin_unlock_irqrestore(&srq->lock, flags); 380 return NULL; 381 } 382 if (!srq) { 383 qp->rq_get++; 384 } else { 385 if (srq->armed) { 386 /* Test SRQ limit */ 387 u32 off = (srq->rq_get + srq->limit) % 388 srq->num_rqe; 389 struct siw_rqe *rqe2 = &srq->recvq[off]; 390 391 if (!(rqe2->flags & SIW_WQE_VALID)) { 392 srq->armed = false; 393 srq_event = true; 394 } 395 } 396 srq->rq_get++; 397 } 398 } 399 out: 400 if (srq) { 401 spin_unlock_irqrestore(&srq->lock, flags); 402 if (srq_event) 403 siw_srq_event(srq, IB_EVENT_SRQ_LIMIT_REACHED); 404 } 405 return wqe; 406 } 407 408 /* 409 * siw_proc_send: 410 * 411 * Process one incoming SEND and place data into memory referenced by 412 * receive wqe. 413 * 414 * Function supports partially received sends (suspending/resuming 415 * current receive wqe processing) 416 * 417 * return value: 418 * 0: reached the end of a DDP segment 419 * -EAGAIN: to be called again to finish the DDP segment 420 */ 421 int siw_proc_send(struct siw_qp *qp) 422 { 423 struct siw_rx_stream *srx = &qp->rx_stream; 424 struct siw_rx_fpdu *frx = &qp->rx_untagged; 425 struct siw_wqe *wqe; 426 u32 data_bytes; /* all data bytes available */ 427 u32 rcvd_bytes; /* sum of data bytes rcvd */ 428 int rv = 0; 429 430 if (frx->first_ddp_seg) { 431 wqe = siw_rqe_get(qp); 432 if (unlikely(!wqe)) { 433 siw_init_terminate(qp, TERM_ERROR_LAYER_DDP, 434 DDP_ETYPE_UNTAGGED_BUF, 435 DDP_ECODE_UT_INVALID_MSN_NOBUF, 0); 436 return -ENOENT; 437 } 438 } else { 439 wqe = rx_wqe(frx); 440 } 441 if (srx->state == SIW_GET_DATA_START) { 442 rv = siw_send_check_ntoh(srx, frx); 443 if (unlikely(rv)) { 444 siw_qp_event(qp, IB_EVENT_QP_FATAL); 445 return rv; 446 } 447 if (!srx->fpdu_part_rem) /* zero length SEND */ 448 return 0; 449 } 450 data_bytes = min(srx->fpdu_part_rem, srx->skb_new); 451 rcvd_bytes = 0; 452 453 /* A zero length SEND will skip below loop */ 454 while (data_bytes) { 455 struct ib_pd *pd; 456 struct siw_mem **mem, *mem_p; 457 struct siw_sge *sge; 458 u32 sge_bytes; /* data bytes avail for SGE */ 459 460 sge = &wqe->rqe.sge[frx->sge_idx]; 461 462 if (!sge->length) { 463 /* just skip empty sge's */ 464 frx->sge_idx++; 465 frx->sge_off = 0; 466 frx->pbl_idx = 0; 467 continue; 468 } 469 sge_bytes = min(data_bytes, sge->length - frx->sge_off); 470 mem = &wqe->mem[frx->sge_idx]; 471 472 /* 473 * check with QP's PD if no SRQ present, SRQ's PD otherwise 474 */ 475 pd = qp->srq == NULL ? qp->pd : qp->srq->base_srq.pd; 476 477 rv = siw_check_sge(pd, sge, mem, IB_ACCESS_LOCAL_WRITE, 478 frx->sge_off, sge_bytes); 479 if (unlikely(rv)) { 480 siw_init_terminate(qp, TERM_ERROR_LAYER_DDP, 481 DDP_ETYPE_CATASTROPHIC, 482 DDP_ECODE_CATASTROPHIC, 0); 483 484 siw_qp_event(qp, IB_EVENT_QP_ACCESS_ERR); 485 break; 486 } 487 mem_p = *mem; 488 if (mem_p->mem_obj == NULL) 489 rv = siw_rx_kva(srx, 490 (void *)(uintptr_t)(sge->laddr + frx->sge_off), 491 sge_bytes); 492 else if (!mem_p->is_pbl) 493 rv = siw_rx_umem(srx, mem_p->umem, 494 sge->laddr + frx->sge_off, sge_bytes); 495 else 496 rv = siw_rx_pbl(srx, &frx->pbl_idx, mem_p, 497 sge->laddr + frx->sge_off, sge_bytes); 498 499 if (unlikely(rv != sge_bytes)) { 500 wqe->processed += rcvd_bytes; 501 502 siw_init_terminate(qp, TERM_ERROR_LAYER_DDP, 503 DDP_ETYPE_CATASTROPHIC, 504 DDP_ECODE_CATASTROPHIC, 0); 505 return -EINVAL; 506 } 507 frx->sge_off += rv; 508 509 if (frx->sge_off == sge->length) { 510 frx->sge_idx++; 511 frx->sge_off = 0; 512 frx->pbl_idx = 0; 513 } 514 data_bytes -= rv; 515 rcvd_bytes += rv; 516 517 srx->fpdu_part_rem -= rv; 518 srx->fpdu_part_rcvd += rv; 519 } 520 wqe->processed += rcvd_bytes; 521 522 if (!srx->fpdu_part_rem) 523 return 0; 524 525 return (rv < 0) ? rv : -EAGAIN; 526 } 527 528 /* 529 * siw_proc_write: 530 * 531 * Place incoming WRITE after referencing and checking target buffer 532 533 * Function supports partially received WRITEs (suspending/resuming 534 * current receive processing) 535 * 536 * return value: 537 * 0: reached the end of a DDP segment 538 * -EAGAIN: to be called again to finish the DDP segment 539 */ 540 int siw_proc_write(struct siw_qp *qp) 541 { 542 struct siw_rx_stream *srx = &qp->rx_stream; 543 struct siw_rx_fpdu *frx = &qp->rx_tagged; 544 struct siw_mem *mem; 545 int bytes, rv; 546 547 if (srx->state == SIW_GET_DATA_START) { 548 if (!srx->fpdu_part_rem) /* zero length WRITE */ 549 return 0; 550 551 rv = siw_write_check_ntoh(srx, frx); 552 if (unlikely(rv)) { 553 siw_qp_event(qp, IB_EVENT_QP_FATAL); 554 return rv; 555 } 556 } 557 bytes = min(srx->fpdu_part_rem, srx->skb_new); 558 559 if (frx->first_ddp_seg) { 560 struct siw_wqe *wqe = rx_wqe(frx); 561 562 rx_mem(frx) = siw_mem_id2obj(qp->sdev, srx->ddp_stag >> 8); 563 if (unlikely(!rx_mem(frx))) { 564 siw_dbg_qp(qp, 565 "sink stag not found/invalid, stag 0x%08x\n", 566 srx->ddp_stag); 567 568 siw_init_terminate(qp, TERM_ERROR_LAYER_DDP, 569 DDP_ETYPE_TAGGED_BUF, 570 DDP_ECODE_T_INVALID_STAG, 0); 571 return -EINVAL; 572 } 573 wqe->rqe.num_sge = 1; 574 rx_type(wqe) = SIW_OP_WRITE; 575 wqe->wr_status = SIW_WR_INPROGRESS; 576 } 577 mem = rx_mem(frx); 578 579 /* 580 * Check if application re-registered memory with different 581 * key field of STag. 582 */ 583 if (unlikely(mem->stag != srx->ddp_stag)) { 584 siw_init_terminate(qp, TERM_ERROR_LAYER_DDP, 585 DDP_ETYPE_TAGGED_BUF, 586 DDP_ECODE_T_INVALID_STAG, 0); 587 return -EINVAL; 588 } 589 rv = siw_check_mem(qp->pd, mem, srx->ddp_to + srx->fpdu_part_rcvd, 590 IB_ACCESS_REMOTE_WRITE, bytes); 591 if (unlikely(rv)) { 592 siw_init_terminate(qp, TERM_ERROR_LAYER_DDP, 593 DDP_ETYPE_TAGGED_BUF, siw_tagged_error(-rv), 594 0); 595 596 siw_qp_event(qp, IB_EVENT_QP_ACCESS_ERR); 597 598 return -EINVAL; 599 } 600 601 if (mem->mem_obj == NULL) 602 rv = siw_rx_kva(srx, 603 (void *)(uintptr_t)(srx->ddp_to + srx->fpdu_part_rcvd), 604 bytes); 605 else if (!mem->is_pbl) 606 rv = siw_rx_umem(srx, mem->umem, 607 srx->ddp_to + srx->fpdu_part_rcvd, bytes); 608 else 609 rv = siw_rx_pbl(srx, &frx->pbl_idx, mem, 610 srx->ddp_to + srx->fpdu_part_rcvd, bytes); 611 612 if (unlikely(rv != bytes)) { 613 siw_init_terminate(qp, TERM_ERROR_LAYER_DDP, 614 DDP_ETYPE_CATASTROPHIC, 615 DDP_ECODE_CATASTROPHIC, 0); 616 return -EINVAL; 617 } 618 srx->fpdu_part_rem -= rv; 619 srx->fpdu_part_rcvd += rv; 620 621 if (!srx->fpdu_part_rem) { 622 srx->ddp_to += srx->fpdu_part_rcvd; 623 return 0; 624 } 625 return -EAGAIN; 626 } 627 628 /* 629 * Inbound RREQ's cannot carry user data. 630 */ 631 int siw_proc_rreq(struct siw_qp *qp) 632 { 633 struct siw_rx_stream *srx = &qp->rx_stream; 634 635 if (!srx->fpdu_part_rem) 636 return 0; 637 638 pr_warn("siw: [QP %u]: rreq with mpa len %d\n", qp_id(qp), 639 be16_to_cpu(srx->hdr.ctrl.mpa_len)); 640 641 return -EPROTO; 642 } 643 644 /* 645 * siw_init_rresp: 646 * 647 * Process inbound RDMA READ REQ. Produce a pseudo READ RESPONSE WQE. 648 * Put it at the tail of the IRQ, if there is another WQE currently in 649 * transmit processing. If not, make it the current WQE to be processed 650 * and schedule transmit processing. 651 * 652 * Can be called from softirq context and from process 653 * context (RREAD socket loopback case!) 654 * 655 * return value: 656 * 0: success, 657 * failure code otherwise 658 */ 659 660 static int siw_init_rresp(struct siw_qp *qp, struct siw_rx_stream *srx) 661 { 662 struct siw_wqe *tx_work = tx_wqe(qp); 663 struct siw_sqe *resp; 664 665 uint64_t raddr = be64_to_cpu(srx->hdr.rreq.sink_to), 666 laddr = be64_to_cpu(srx->hdr.rreq.source_to); 667 uint32_t length = be32_to_cpu(srx->hdr.rreq.read_size), 668 lkey = be32_to_cpu(srx->hdr.rreq.source_stag), 669 rkey = be32_to_cpu(srx->hdr.rreq.sink_stag), 670 msn = be32_to_cpu(srx->hdr.rreq.ddp_msn); 671 672 int run_sq = 1, rv = 0; 673 unsigned long flags; 674 675 if (unlikely(msn != srx->ddp_msn[RDMAP_UNTAGGED_QN_RDMA_READ])) { 676 siw_init_terminate(qp, TERM_ERROR_LAYER_DDP, 677 DDP_ETYPE_UNTAGGED_BUF, 678 DDP_ECODE_UT_INVALID_MSN_RANGE, 0); 679 return -EPROTO; 680 } 681 spin_lock_irqsave(&qp->sq_lock, flags); 682 683 if (tx_work->wr_status == SIW_WR_IDLE) { 684 /* 685 * immediately schedule READ response w/o 686 * consuming IRQ entry: IRQ must be empty. 687 */ 688 tx_work->processed = 0; 689 tx_work->mem[0] = NULL; 690 tx_work->wr_status = SIW_WR_QUEUED; 691 resp = &tx_work->sqe; 692 } else { 693 resp = irq_alloc_free(qp); 694 run_sq = 0; 695 } 696 if (likely(resp)) { 697 resp->opcode = SIW_OP_READ_RESPONSE; 698 699 resp->sge[0].length = length; 700 resp->sge[0].laddr = laddr; 701 resp->sge[0].lkey = lkey; 702 703 /* Keep aside message sequence number for potential 704 * error reporting during Read Response generation. 705 */ 706 resp->sge[1].length = msn; 707 708 resp->raddr = raddr; 709 resp->rkey = rkey; 710 resp->num_sge = length ? 1 : 0; 711 712 /* RRESP now valid as current TX wqe or placed into IRQ */ 713 smp_store_mb(resp->flags, SIW_WQE_VALID); 714 } else { 715 pr_warn("siw: [QP %u]: irq %d exceeded %d\n", qp_id(qp), 716 qp->irq_put % qp->attrs.irq_size, qp->attrs.irq_size); 717 718 siw_init_terminate(qp, TERM_ERROR_LAYER_RDMAP, 719 RDMAP_ETYPE_REMOTE_OPERATION, 720 RDMAP_ECODE_CATASTROPHIC_STREAM, 0); 721 rv = -EPROTO; 722 } 723 724 spin_unlock_irqrestore(&qp->sq_lock, flags); 725 726 if (run_sq) 727 rv = siw_sq_start(qp); 728 729 return rv; 730 } 731 732 /* 733 * Only called at start of Read.Resonse processing. 734 * Transfer pending Read from tip of ORQ into currrent rx wqe, 735 * but keep ORQ entry valid until Read.Response processing done. 736 * No Queue locking needed. 737 */ 738 static int siw_orqe_start_rx(struct siw_qp *qp) 739 { 740 struct siw_sqe *orqe; 741 struct siw_wqe *wqe = NULL; 742 743 /* make sure ORQ indices are current */ 744 smp_mb(); 745 746 orqe = orq_get_current(qp); 747 if (READ_ONCE(orqe->flags) & SIW_WQE_VALID) { 748 /* RRESP is a TAGGED RDMAP operation */ 749 wqe = rx_wqe(&qp->rx_tagged); 750 wqe->sqe.id = orqe->id; 751 wqe->sqe.opcode = orqe->opcode; 752 wqe->sqe.sge[0].laddr = orqe->sge[0].laddr; 753 wqe->sqe.sge[0].lkey = orqe->sge[0].lkey; 754 wqe->sqe.sge[0].length = orqe->sge[0].length; 755 wqe->sqe.flags = orqe->flags; 756 wqe->sqe.num_sge = 1; 757 wqe->bytes = orqe->sge[0].length; 758 wqe->processed = 0; 759 wqe->mem[0] = NULL; 760 /* make sure WQE is completely written before valid */ 761 smp_wmb(); 762 wqe->wr_status = SIW_WR_INPROGRESS; 763 764 return 0; 765 } 766 return -EPROTO; 767 } 768 769 /* 770 * siw_proc_rresp: 771 * 772 * Place incoming RRESP data into memory referenced by RREQ WQE 773 * which is at the tip of the ORQ 774 * 775 * Function supports partially received RRESP's (suspending/resuming 776 * current receive processing) 777 */ 778 int siw_proc_rresp(struct siw_qp *qp) 779 { 780 struct siw_rx_stream *srx = &qp->rx_stream; 781 struct siw_rx_fpdu *frx = &qp->rx_tagged; 782 struct siw_wqe *wqe = rx_wqe(frx); 783 struct siw_mem **mem, *mem_p; 784 struct siw_sge *sge; 785 int bytes, rv; 786 787 if (frx->first_ddp_seg) { 788 if (unlikely(wqe->wr_status != SIW_WR_IDLE)) { 789 pr_warn("siw: [QP %u]: proc RRESP: status %d, op %d\n", 790 qp_id(qp), wqe->wr_status, wqe->sqe.opcode); 791 rv = -EPROTO; 792 goto error_term; 793 } 794 /* 795 * fetch pending RREQ from orq 796 */ 797 rv = siw_orqe_start_rx(qp); 798 if (rv) { 799 pr_warn("siw: [QP %u]: ORQ empty at idx %d\n", 800 qp_id(qp), qp->orq_get % qp->attrs.orq_size); 801 goto error_term; 802 } 803 rv = siw_rresp_check_ntoh(srx, frx); 804 if (unlikely(rv)) { 805 siw_qp_event(qp, IB_EVENT_QP_FATAL); 806 return rv; 807 } 808 } else { 809 if (unlikely(wqe->wr_status != SIW_WR_INPROGRESS)) { 810 pr_warn("siw: [QP %u]: resume RRESP: status %d\n", 811 qp_id(qp), wqe->wr_status); 812 rv = -EPROTO; 813 goto error_term; 814 } 815 } 816 if (!srx->fpdu_part_rem) /* zero length RRESPONSE */ 817 return 0; 818 819 sge = wqe->sqe.sge; /* there is only one */ 820 mem = &wqe->mem[0]; 821 822 if (!(*mem)) { 823 /* 824 * check target memory which resolves memory on first fragment 825 */ 826 rv = siw_check_sge(qp->pd, sge, mem, IB_ACCESS_LOCAL_WRITE, 0, 827 wqe->bytes); 828 if (unlikely(rv)) { 829 siw_dbg_qp(qp, "target mem check: %d\n", rv); 830 wqe->wc_status = SIW_WC_LOC_PROT_ERR; 831 832 siw_init_terminate(qp, TERM_ERROR_LAYER_DDP, 833 DDP_ETYPE_TAGGED_BUF, 834 siw_tagged_error(-rv), 0); 835 836 siw_qp_event(qp, IB_EVENT_QP_ACCESS_ERR); 837 838 return -EINVAL; 839 } 840 } 841 mem_p = *mem; 842 843 bytes = min(srx->fpdu_part_rem, srx->skb_new); 844 845 if (mem_p->mem_obj == NULL) 846 rv = siw_rx_kva(srx, 847 (void *)(uintptr_t)(sge->laddr + wqe->processed), 848 bytes); 849 else if (!mem_p->is_pbl) 850 rv = siw_rx_umem(srx, mem_p->umem, sge->laddr + wqe->processed, 851 bytes); 852 else 853 rv = siw_rx_pbl(srx, &frx->pbl_idx, mem_p, 854 sge->laddr + wqe->processed, bytes); 855 if (rv != bytes) { 856 wqe->wc_status = SIW_WC_GENERAL_ERR; 857 rv = -EINVAL; 858 goto error_term; 859 } 860 srx->fpdu_part_rem -= rv; 861 srx->fpdu_part_rcvd += rv; 862 wqe->processed += rv; 863 864 if (!srx->fpdu_part_rem) { 865 srx->ddp_to += srx->fpdu_part_rcvd; 866 return 0; 867 } 868 return -EAGAIN; 869 870 error_term: 871 siw_init_terminate(qp, TERM_ERROR_LAYER_DDP, DDP_ETYPE_CATASTROPHIC, 872 DDP_ECODE_CATASTROPHIC, 0); 873 return rv; 874 } 875 876 int siw_proc_terminate(struct siw_qp *qp) 877 { 878 struct siw_rx_stream *srx = &qp->rx_stream; 879 struct sk_buff *skb = srx->skb; 880 struct iwarp_terminate *term = &srx->hdr.terminate; 881 union iwarp_hdr term_info; 882 u8 *infop = (u8 *)&term_info; 883 enum rdma_opcode op; 884 u16 to_copy = sizeof(struct iwarp_ctrl); 885 886 pr_warn("siw: got TERMINATE. layer %d, type %d, code %d\n", 887 __rdmap_term_layer(term), __rdmap_term_etype(term), 888 __rdmap_term_ecode(term)); 889 890 if (be32_to_cpu(term->ddp_qn) != RDMAP_UNTAGGED_QN_TERMINATE || 891 be32_to_cpu(term->ddp_msn) != 892 qp->rx_stream.ddp_msn[RDMAP_UNTAGGED_QN_TERMINATE] || 893 be32_to_cpu(term->ddp_mo) != 0) { 894 pr_warn("siw: rx bogus TERM [QN x%08x, MSN x%08x, MO x%08x]\n", 895 be32_to_cpu(term->ddp_qn), be32_to_cpu(term->ddp_msn), 896 be32_to_cpu(term->ddp_mo)); 897 return -ECONNRESET; 898 } 899 /* 900 * Receive remaining pieces of TERM if indicated 901 */ 902 if (!term->flag_m) 903 return -ECONNRESET; 904 905 /* Do not take the effort to reassemble a network fragmented 906 * TERM message 907 */ 908 if (srx->skb_new < sizeof(struct iwarp_ctrl_tagged)) 909 return -ECONNRESET; 910 911 memset(infop, 0, sizeof(term_info)); 912 913 skb_copy_bits(skb, srx->skb_offset, infop, to_copy); 914 915 op = __rdmap_get_opcode(&term_info.ctrl); 916 if (op >= RDMAP_TERMINATE) 917 goto out; 918 919 infop += to_copy; 920 srx->skb_offset += to_copy; 921 srx->skb_new -= to_copy; 922 srx->skb_copied += to_copy; 923 srx->fpdu_part_rcvd += to_copy; 924 srx->fpdu_part_rem -= to_copy; 925 926 to_copy = iwarp_pktinfo[op].hdr_len - to_copy; 927 928 /* Again, no network fragmented TERM's */ 929 if (to_copy + MPA_CRC_SIZE > srx->skb_new) 930 return -ECONNRESET; 931 932 skb_copy_bits(skb, srx->skb_offset, infop, to_copy); 933 934 if (term->flag_r) { 935 siw_dbg_qp(qp, "TERM reports RDMAP hdr type %u, len %u (%s)\n", 936 op, be16_to_cpu(term_info.ctrl.mpa_len), 937 term->flag_m ? "valid" : "invalid"); 938 } else if (term->flag_d) { 939 siw_dbg_qp(qp, "TERM reports DDP hdr type %u, len %u (%s)\n", 940 op, be16_to_cpu(term_info.ctrl.mpa_len), 941 term->flag_m ? "valid" : "invalid"); 942 } 943 out: 944 srx->skb_new -= to_copy; 945 srx->skb_offset += to_copy; 946 srx->skb_copied += to_copy; 947 srx->fpdu_part_rcvd += to_copy; 948 srx->fpdu_part_rem -= to_copy; 949 950 return -ECONNRESET; 951 } 952 953 static int siw_get_trailer(struct siw_qp *qp, struct siw_rx_stream *srx) 954 { 955 struct sk_buff *skb = srx->skb; 956 u8 *tbuf = (u8 *)&srx->trailer.crc - srx->pad; 957 __wsum crc_in, crc_own = 0; 958 959 siw_dbg_qp(qp, "expected %d, available %d, pad %u\n", 960 srx->fpdu_part_rem, srx->skb_new, srx->pad); 961 962 if (srx->skb_new < srx->fpdu_part_rem) 963 return -EAGAIN; 964 965 skb_copy_bits(skb, srx->skb_offset, tbuf, srx->fpdu_part_rem); 966 967 if (srx->mpa_crc_hd && srx->pad) 968 crypto_shash_update(srx->mpa_crc_hd, tbuf, srx->pad); 969 970 srx->skb_new -= srx->fpdu_part_rem; 971 srx->skb_offset += srx->fpdu_part_rem; 972 srx->skb_copied += srx->fpdu_part_rem; 973 974 if (!srx->mpa_crc_hd) 975 return 0; 976 977 /* 978 * CRC32 is computed, transmitted and received directly in NBO, 979 * so there's never a reason to convert byte order. 980 */ 981 crypto_shash_final(srx->mpa_crc_hd, (u8 *)&crc_own); 982 crc_in = (__force __wsum)srx->trailer.crc; 983 984 if (unlikely(crc_in != crc_own)) { 985 pr_warn("siw: crc error. in: %08x, own %08x, op %u\n", 986 crc_in, crc_own, qp->rx_stream.rdmap_op); 987 988 siw_init_terminate(qp, TERM_ERROR_LAYER_LLP, 989 LLP_ETYPE_MPA, 990 LLP_ECODE_RECEIVED_CRC, 0); 991 return -EINVAL; 992 } 993 return 0; 994 } 995 996 #define MIN_DDP_HDR sizeof(struct iwarp_ctrl_tagged) 997 998 static int siw_get_hdr(struct siw_rx_stream *srx) 999 { 1000 struct sk_buff *skb = srx->skb; 1001 struct siw_qp *qp = rx_qp(srx); 1002 struct iwarp_ctrl *c_hdr = &srx->hdr.ctrl; 1003 struct siw_rx_fpdu *frx; 1004 u8 opcode; 1005 int bytes; 1006 1007 if (srx->fpdu_part_rcvd < MIN_DDP_HDR) { 1008 /* 1009 * copy a mimimum sized (tagged) DDP frame control part 1010 */ 1011 bytes = min_t(int, srx->skb_new, 1012 MIN_DDP_HDR - srx->fpdu_part_rcvd); 1013 1014 skb_copy_bits(skb, srx->skb_offset, 1015 (char *)c_hdr + srx->fpdu_part_rcvd, bytes); 1016 1017 srx->fpdu_part_rcvd += bytes; 1018 1019 srx->skb_new -= bytes; 1020 srx->skb_offset += bytes; 1021 srx->skb_copied += bytes; 1022 1023 if (srx->fpdu_part_rcvd < MIN_DDP_HDR) 1024 return -EAGAIN; 1025 1026 if (unlikely(__ddp_get_version(c_hdr) != DDP_VERSION)) { 1027 enum ddp_etype etype; 1028 enum ddp_ecode ecode; 1029 1030 pr_warn("siw: received ddp version unsupported %d\n", 1031 __ddp_get_version(c_hdr)); 1032 1033 if (c_hdr->ddp_rdmap_ctrl & DDP_FLAG_TAGGED) { 1034 etype = DDP_ETYPE_TAGGED_BUF; 1035 ecode = DDP_ECODE_T_VERSION; 1036 } else { 1037 etype = DDP_ETYPE_UNTAGGED_BUF; 1038 ecode = DDP_ECODE_UT_VERSION; 1039 } 1040 siw_init_terminate(rx_qp(srx), TERM_ERROR_LAYER_DDP, 1041 etype, ecode, 0); 1042 return -EINVAL; 1043 } 1044 if (unlikely(__rdmap_get_version(c_hdr) != RDMAP_VERSION)) { 1045 pr_warn("siw: received rdmap version unsupported %d\n", 1046 __rdmap_get_version(c_hdr)); 1047 1048 siw_init_terminate(rx_qp(srx), TERM_ERROR_LAYER_RDMAP, 1049 RDMAP_ETYPE_REMOTE_OPERATION, 1050 RDMAP_ECODE_VERSION, 0); 1051 return -EINVAL; 1052 } 1053 opcode = __rdmap_get_opcode(c_hdr); 1054 1055 if (opcode > RDMAP_TERMINATE) { 1056 pr_warn("siw: received unknown packet type %u\n", 1057 opcode); 1058 1059 siw_init_terminate(rx_qp(srx), TERM_ERROR_LAYER_RDMAP, 1060 RDMAP_ETYPE_REMOTE_OPERATION, 1061 RDMAP_ECODE_OPCODE, 0); 1062 return -EINVAL; 1063 } 1064 siw_dbg_qp(rx_qp(srx), "new header, opcode %u\n", opcode); 1065 } else { 1066 opcode = __rdmap_get_opcode(c_hdr); 1067 } 1068 set_rx_fpdu_context(qp, opcode); 1069 frx = qp->rx_fpdu; 1070 1071 /* 1072 * Figure out len of current hdr: variable length of 1073 * iwarp hdr may force us to copy hdr information in 1074 * two steps. Only tagged DDP messages are already 1075 * completely received. 1076 */ 1077 if (iwarp_pktinfo[opcode].hdr_len > sizeof(struct iwarp_ctrl_tagged)) { 1078 bytes = iwarp_pktinfo[opcode].hdr_len - MIN_DDP_HDR; 1079 1080 if (srx->skb_new < bytes) 1081 return -EAGAIN; 1082 1083 skb_copy_bits(skb, srx->skb_offset, 1084 (char *)c_hdr + srx->fpdu_part_rcvd, bytes); 1085 1086 srx->fpdu_part_rcvd += bytes; 1087 1088 srx->skb_new -= bytes; 1089 srx->skb_offset += bytes; 1090 srx->skb_copied += bytes; 1091 } 1092 1093 /* 1094 * DDP/RDMAP header receive completed. Check if the current 1095 * DDP segment starts a new RDMAP message or continues a previously 1096 * started RDMAP message. 1097 * 1098 * Alternating reception of DDP segments (or FPDUs) from incomplete 1099 * tagged and untagged RDMAP messages is supported, as long as 1100 * the current tagged or untagged message gets eventually completed 1101 * w/o intersection from another message of the same type 1102 * (tagged/untagged). E.g., a WRITE can get intersected by a SEND, 1103 * but not by a READ RESPONSE etc. 1104 */ 1105 if (srx->mpa_crc_hd) { 1106 /* 1107 * Restart CRC computation 1108 */ 1109 crypto_shash_init(srx->mpa_crc_hd); 1110 crypto_shash_update(srx->mpa_crc_hd, (u8 *)c_hdr, 1111 srx->fpdu_part_rcvd); 1112 } 1113 if (frx->more_ddp_segs) { 1114 frx->first_ddp_seg = 0; 1115 if (frx->prev_rdmap_op != opcode) { 1116 pr_warn("siw: packet intersection: %u : %u\n", 1117 frx->prev_rdmap_op, opcode); 1118 /* 1119 * The last inbound RDMA operation of same type 1120 * (tagged or untagged) is left unfinished. 1121 * To complete it in error, make it the current 1122 * operation again, even with the header already 1123 * overwritten. For error handling, only the opcode 1124 * and current rx context are relevant. 1125 */ 1126 set_rx_fpdu_context(qp, frx->prev_rdmap_op); 1127 __rdmap_set_opcode(c_hdr, frx->prev_rdmap_op); 1128 return -EPROTO; 1129 } 1130 } else { 1131 frx->prev_rdmap_op = opcode; 1132 frx->first_ddp_seg = 1; 1133 } 1134 frx->more_ddp_segs = c_hdr->ddp_rdmap_ctrl & DDP_FLAG_LAST ? 0 : 1; 1135 1136 return 0; 1137 } 1138 1139 static int siw_check_tx_fence(struct siw_qp *qp) 1140 { 1141 struct siw_wqe *tx_waiting = tx_wqe(qp); 1142 struct siw_sqe *rreq; 1143 int resume_tx = 0, rv = 0; 1144 unsigned long flags; 1145 1146 spin_lock_irqsave(&qp->orq_lock, flags); 1147 1148 rreq = orq_get_current(qp); 1149 1150 /* free current orq entry */ 1151 WRITE_ONCE(rreq->flags, 0); 1152 1153 if (qp->tx_ctx.orq_fence) { 1154 if (unlikely(tx_waiting->wr_status != SIW_WR_QUEUED)) { 1155 pr_warn("siw: [QP %u]: fence resume: bad status %d\n", 1156 qp_id(qp), tx_waiting->wr_status); 1157 rv = -EPROTO; 1158 goto out; 1159 } 1160 /* resume SQ processing */ 1161 if (tx_waiting->sqe.opcode == SIW_OP_READ || 1162 tx_waiting->sqe.opcode == SIW_OP_READ_LOCAL_INV) { 1163 rreq = orq_get_tail(qp); 1164 if (unlikely(!rreq)) { 1165 pr_warn("siw: [QP %u]: no ORQE\n", qp_id(qp)); 1166 rv = -EPROTO; 1167 goto out; 1168 } 1169 siw_read_to_orq(rreq, &tx_waiting->sqe); 1170 1171 qp->orq_put++; 1172 qp->tx_ctx.orq_fence = 0; 1173 resume_tx = 1; 1174 1175 } else if (siw_orq_empty(qp)) { 1176 qp->tx_ctx.orq_fence = 0; 1177 resume_tx = 1; 1178 } else { 1179 pr_warn("siw: [QP %u]: fence resume: orq idx: %d:%d\n", 1180 qp_id(qp), qp->orq_get, qp->orq_put); 1181 rv = -EPROTO; 1182 } 1183 } 1184 qp->orq_get++; 1185 out: 1186 spin_unlock_irqrestore(&qp->orq_lock, flags); 1187 1188 if (resume_tx) 1189 rv = siw_sq_start(qp); 1190 1191 return rv; 1192 } 1193 1194 /* 1195 * siw_rdmap_complete() 1196 * 1197 * Complete processing of an RDMA message after receiving all 1198 * DDP segmens or ABort processing after encountering error case. 1199 * 1200 * o SENDs + RRESPs will need for completion, 1201 * o RREQs need for READ RESPONSE initialization 1202 * o WRITEs need memory dereferencing 1203 * 1204 * TODO: Failed WRITEs need local error to be surfaced. 1205 */ 1206 static int siw_rdmap_complete(struct siw_qp *qp, int error) 1207 { 1208 struct siw_rx_stream *srx = &qp->rx_stream; 1209 struct siw_wqe *wqe = rx_wqe(qp->rx_fpdu); 1210 enum siw_wc_status wc_status = wqe->wc_status; 1211 u8 opcode = __rdmap_get_opcode(&srx->hdr.ctrl); 1212 int rv = 0; 1213 1214 switch (opcode) { 1215 case RDMAP_SEND_SE: 1216 case RDMAP_SEND_SE_INVAL: 1217 wqe->rqe.flags |= SIW_WQE_SOLICITED; 1218 fallthrough; 1219 1220 case RDMAP_SEND: 1221 case RDMAP_SEND_INVAL: 1222 if (wqe->wr_status == SIW_WR_IDLE) 1223 break; 1224 1225 srx->ddp_msn[RDMAP_UNTAGGED_QN_SEND]++; 1226 1227 if (error != 0 && wc_status == SIW_WC_SUCCESS) 1228 wc_status = SIW_WC_GENERAL_ERR; 1229 /* 1230 * Handle STag invalidation request 1231 */ 1232 if (wc_status == SIW_WC_SUCCESS && 1233 (opcode == RDMAP_SEND_INVAL || 1234 opcode == RDMAP_SEND_SE_INVAL)) { 1235 rv = siw_invalidate_stag(qp->pd, srx->inval_stag); 1236 if (rv) { 1237 siw_init_terminate( 1238 qp, TERM_ERROR_LAYER_RDMAP, 1239 rv == -EACCES ? 1240 RDMAP_ETYPE_REMOTE_PROTECTION : 1241 RDMAP_ETYPE_REMOTE_OPERATION, 1242 RDMAP_ECODE_CANNOT_INVALIDATE, 0); 1243 1244 wc_status = SIW_WC_REM_INV_REQ_ERR; 1245 } 1246 rv = siw_rqe_complete(qp, &wqe->rqe, wqe->processed, 1247 rv ? 0 : srx->inval_stag, 1248 wc_status); 1249 } else { 1250 rv = siw_rqe_complete(qp, &wqe->rqe, wqe->processed, 1251 0, wc_status); 1252 } 1253 siw_wqe_put_mem(wqe, SIW_OP_RECEIVE); 1254 break; 1255 1256 case RDMAP_RDMA_READ_RESP: 1257 if (wqe->wr_status == SIW_WR_IDLE) 1258 break; 1259 1260 if (error != 0) { 1261 if ((srx->state == SIW_GET_HDR && 1262 qp->rx_fpdu->first_ddp_seg) || error == -ENODATA) 1263 /* possible RREQ in ORQ left untouched */ 1264 break; 1265 1266 if (wc_status == SIW_WC_SUCCESS) 1267 wc_status = SIW_WC_GENERAL_ERR; 1268 } else if (rdma_is_kernel_res(&qp->base_qp.res) && 1269 rx_type(wqe) == SIW_OP_READ_LOCAL_INV) { 1270 /* 1271 * Handle any STag invalidation request 1272 */ 1273 rv = siw_invalidate_stag(qp->pd, wqe->sqe.sge[0].lkey); 1274 if (rv) { 1275 siw_init_terminate(qp, TERM_ERROR_LAYER_RDMAP, 1276 RDMAP_ETYPE_CATASTROPHIC, 1277 RDMAP_ECODE_UNSPECIFIED, 0); 1278 1279 if (wc_status == SIW_WC_SUCCESS) { 1280 wc_status = SIW_WC_GENERAL_ERR; 1281 error = rv; 1282 } 1283 } 1284 } 1285 /* 1286 * All errors turn the wqe into signalled. 1287 */ 1288 if ((wqe->sqe.flags & SIW_WQE_SIGNALLED) || error != 0) 1289 rv = siw_sqe_complete(qp, &wqe->sqe, wqe->processed, 1290 wc_status); 1291 siw_wqe_put_mem(wqe, SIW_OP_READ); 1292 1293 if (!error) 1294 rv = siw_check_tx_fence(qp); 1295 else 1296 /* Disable current ORQ eleement */ 1297 WRITE_ONCE(orq_get_current(qp)->flags, 0); 1298 break; 1299 1300 case RDMAP_RDMA_READ_REQ: 1301 if (!error) { 1302 rv = siw_init_rresp(qp, srx); 1303 srx->ddp_msn[RDMAP_UNTAGGED_QN_RDMA_READ]++; 1304 } 1305 break; 1306 1307 case RDMAP_RDMA_WRITE: 1308 if (wqe->wr_status == SIW_WR_IDLE) 1309 break; 1310 1311 /* 1312 * Free References from memory object if 1313 * attached to receive context (inbound WRITE). 1314 * While a zero-length WRITE is allowed, 1315 * no memory reference got created. 1316 */ 1317 if (rx_mem(&qp->rx_tagged)) { 1318 siw_mem_put(rx_mem(&qp->rx_tagged)); 1319 rx_mem(&qp->rx_tagged) = NULL; 1320 } 1321 break; 1322 1323 default: 1324 break; 1325 } 1326 wqe->wr_status = SIW_WR_IDLE; 1327 1328 return rv; 1329 } 1330 1331 /* 1332 * siw_tcp_rx_data() 1333 * 1334 * Main routine to consume inbound TCP payload 1335 * 1336 * @rd_desc: read descriptor 1337 * @skb: socket buffer 1338 * @off: offset in skb 1339 * @len: skb->len - offset : payload in skb 1340 */ 1341 int siw_tcp_rx_data(read_descriptor_t *rd_desc, struct sk_buff *skb, 1342 unsigned int off, size_t len) 1343 { 1344 struct siw_qp *qp = rd_desc->arg.data; 1345 struct siw_rx_stream *srx = &qp->rx_stream; 1346 int rv; 1347 1348 srx->skb = skb; 1349 srx->skb_new = skb->len - off; 1350 srx->skb_offset = off; 1351 srx->skb_copied = 0; 1352 1353 siw_dbg_qp(qp, "new data, len %d\n", srx->skb_new); 1354 1355 while (srx->skb_new) { 1356 int run_completion = 1; 1357 1358 if (unlikely(srx->rx_suspend)) { 1359 /* Do not process any more data */ 1360 srx->skb_copied += srx->skb_new; 1361 break; 1362 } 1363 switch (srx->state) { 1364 case SIW_GET_HDR: 1365 rv = siw_get_hdr(srx); 1366 if (!rv) { 1367 srx->fpdu_part_rem = 1368 be16_to_cpu(srx->hdr.ctrl.mpa_len) - 1369 srx->fpdu_part_rcvd + MPA_HDR_SIZE; 1370 1371 if (srx->fpdu_part_rem) 1372 srx->pad = -srx->fpdu_part_rem & 0x3; 1373 else 1374 srx->pad = 0; 1375 1376 srx->state = SIW_GET_DATA_START; 1377 srx->fpdu_part_rcvd = 0; 1378 } 1379 break; 1380 1381 case SIW_GET_DATA_MORE: 1382 /* 1383 * Another data fragment of the same DDP segment. 1384 * Setting first_ddp_seg = 0 avoids repeating 1385 * initializations that shall occur only once per 1386 * DDP segment. 1387 */ 1388 qp->rx_fpdu->first_ddp_seg = 0; 1389 fallthrough; 1390 1391 case SIW_GET_DATA_START: 1392 /* 1393 * Headers will be checked by the opcode-specific 1394 * data receive function below. 1395 */ 1396 rv = iwarp_pktinfo[qp->rx_stream.rdmap_op].rx_data(qp); 1397 if (!rv) { 1398 int mpa_len = 1399 be16_to_cpu(srx->hdr.ctrl.mpa_len) 1400 + MPA_HDR_SIZE; 1401 1402 srx->fpdu_part_rem = (-mpa_len & 0x3) 1403 + MPA_CRC_SIZE; 1404 srx->fpdu_part_rcvd = 0; 1405 srx->state = SIW_GET_TRAILER; 1406 } else { 1407 if (unlikely(rv == -ECONNRESET)) 1408 run_completion = 0; 1409 else 1410 srx->state = SIW_GET_DATA_MORE; 1411 } 1412 break; 1413 1414 case SIW_GET_TRAILER: 1415 /* 1416 * read CRC + any padding 1417 */ 1418 rv = siw_get_trailer(qp, srx); 1419 if (likely(!rv)) { 1420 /* 1421 * FPDU completed. 1422 * complete RDMAP message if last fragment 1423 */ 1424 srx->state = SIW_GET_HDR; 1425 srx->fpdu_part_rcvd = 0; 1426 1427 if (!(srx->hdr.ctrl.ddp_rdmap_ctrl & 1428 DDP_FLAG_LAST)) 1429 /* more frags */ 1430 break; 1431 1432 rv = siw_rdmap_complete(qp, 0); 1433 run_completion = 0; 1434 } 1435 break; 1436 1437 default: 1438 pr_warn("QP[%u]: RX out of state\n", qp_id(qp)); 1439 rv = -EPROTO; 1440 run_completion = 0; 1441 } 1442 if (unlikely(rv != 0 && rv != -EAGAIN)) { 1443 if ((srx->state > SIW_GET_HDR || 1444 qp->rx_fpdu->more_ddp_segs) && run_completion) 1445 siw_rdmap_complete(qp, rv); 1446 1447 siw_dbg_qp(qp, "rx error %d, rx state %d\n", rv, 1448 srx->state); 1449 1450 siw_qp_cm_drop(qp, 1); 1451 1452 break; 1453 } 1454 if (rv) { 1455 siw_dbg_qp(qp, "fpdu fragment, state %d, missing %d\n", 1456 srx->state, srx->fpdu_part_rem); 1457 break; 1458 } 1459 } 1460 return srx->skb_copied; 1461 } 1462