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 #include <net/tcp.h> 12 13 #include <rdma/iw_cm.h> 14 #include <rdma/ib_verbs.h> 15 #include <rdma/ib_user_verbs.h> 16 17 #include "siw.h" 18 #include "siw_verbs.h" 19 #include "siw_mem.h" 20 21 #define MAX_HDR_INLINE \ 22 (((uint32_t)(sizeof(struct siw_rreq_pkt) - \ 23 sizeof(struct iwarp_send))) & 0xF8) 24 25 static struct page *siw_get_pblpage(struct siw_mem *mem, u64 addr, int *idx) 26 { 27 struct siw_pbl *pbl = mem->pbl; 28 u64 offset = addr - mem->va; 29 dma_addr_t paddr = siw_pbl_get_buffer(pbl, offset, NULL, idx); 30 31 if (paddr) 32 return ib_virt_dma_to_page(paddr); 33 34 return NULL; 35 } 36 37 /* 38 * Copy short payload at provided destination payload address 39 */ 40 static int siw_try_1seg(struct siw_iwarp_tx *c_tx, void *paddr) 41 { 42 struct siw_wqe *wqe = &c_tx->wqe_active; 43 struct siw_sge *sge = &wqe->sqe.sge[0]; 44 u32 bytes = sge->length; 45 46 if (bytes > MAX_HDR_INLINE || wqe->sqe.num_sge != 1) 47 return MAX_HDR_INLINE + 1; 48 49 if (!bytes) 50 return 0; 51 52 if (tx_flags(wqe) & SIW_WQE_INLINE) { 53 memcpy(paddr, &wqe->sqe.sge[1], bytes); 54 } else { 55 struct siw_mem *mem = wqe->mem[0]; 56 57 if (!mem->mem_obj) { 58 /* Kernel client using kva */ 59 memcpy(paddr, ib_virt_dma_to_ptr(sge->laddr), bytes); 60 } else if (c_tx->in_syscall) { 61 if (copy_from_user(paddr, u64_to_user_ptr(sge->laddr), 62 bytes)) 63 return -EFAULT; 64 } else { 65 unsigned int off = sge->laddr & ~PAGE_MASK; 66 struct page *p; 67 char *buffer; 68 int pbl_idx = 0; 69 70 if (!mem->is_pbl) 71 p = siw_get_upage(mem->umem, sge->laddr); 72 else 73 p = siw_get_pblpage(mem, sge->laddr, &pbl_idx); 74 75 if (unlikely(!p)) 76 return -EFAULT; 77 78 buffer = kmap_local_page(p); 79 80 if (likely(PAGE_SIZE - off >= bytes)) { 81 memcpy(paddr, buffer + off, bytes); 82 } else { 83 unsigned long part = bytes - (PAGE_SIZE - off); 84 85 memcpy(paddr, buffer + off, part); 86 kunmap_local(buffer); 87 88 if (!mem->is_pbl) 89 p = siw_get_upage(mem->umem, 90 sge->laddr + part); 91 else 92 p = siw_get_pblpage(mem, 93 sge->laddr + part, 94 &pbl_idx); 95 if (unlikely(!p)) 96 return -EFAULT; 97 98 buffer = kmap_local_page(p); 99 memcpy(paddr + part, buffer, bytes - part); 100 } 101 kunmap_local(buffer); 102 } 103 } 104 return (int)bytes; 105 } 106 107 #define PKT_FRAGMENTED 1 108 #define PKT_COMPLETE 0 109 110 /* 111 * siw_qp_prepare_tx() 112 * 113 * Prepare tx state for sending out one fpdu. Builds complete pkt 114 * if no user data or only immediate data are present. 115 * 116 * returns PKT_COMPLETE if complete pkt built, PKT_FRAGMENTED otherwise. 117 */ 118 static int siw_qp_prepare_tx(struct siw_iwarp_tx *c_tx) 119 { 120 struct siw_wqe *wqe = &c_tx->wqe_active; 121 char *crc = NULL; 122 int data = 0; 123 124 switch (tx_type(wqe)) { 125 case SIW_OP_READ: 126 case SIW_OP_READ_LOCAL_INV: 127 memcpy(&c_tx->pkt.ctrl, 128 &iwarp_pktinfo[RDMAP_RDMA_READ_REQ].ctrl, 129 sizeof(struct iwarp_ctrl)); 130 131 c_tx->pkt.rreq.rsvd = 0; 132 c_tx->pkt.rreq.ddp_qn = htonl(RDMAP_UNTAGGED_QN_RDMA_READ); 133 c_tx->pkt.rreq.ddp_msn = 134 htonl(++c_tx->ddp_msn[RDMAP_UNTAGGED_QN_RDMA_READ]); 135 c_tx->pkt.rreq.ddp_mo = 0; 136 c_tx->pkt.rreq.sink_stag = htonl(wqe->sqe.sge[0].lkey); 137 c_tx->pkt.rreq.sink_to = 138 cpu_to_be64(wqe->sqe.sge[0].laddr); 139 c_tx->pkt.rreq.source_stag = htonl(wqe->sqe.rkey); 140 c_tx->pkt.rreq.source_to = cpu_to_be64(wqe->sqe.raddr); 141 c_tx->pkt.rreq.read_size = htonl(wqe->sqe.sge[0].length); 142 143 c_tx->ctrl_len = sizeof(struct iwarp_rdma_rreq); 144 crc = (char *)&c_tx->pkt.rreq_pkt.crc; 145 break; 146 147 case SIW_OP_SEND: 148 if (tx_flags(wqe) & SIW_WQE_SOLICITED) 149 memcpy(&c_tx->pkt.ctrl, 150 &iwarp_pktinfo[RDMAP_SEND_SE].ctrl, 151 sizeof(struct iwarp_ctrl)); 152 else 153 memcpy(&c_tx->pkt.ctrl, &iwarp_pktinfo[RDMAP_SEND].ctrl, 154 sizeof(struct iwarp_ctrl)); 155 156 c_tx->pkt.send.ddp_qn = RDMAP_UNTAGGED_QN_SEND; 157 c_tx->pkt.send.ddp_msn = 158 htonl(++c_tx->ddp_msn[RDMAP_UNTAGGED_QN_SEND]); 159 c_tx->pkt.send.ddp_mo = 0; 160 161 c_tx->pkt.send_inv.inval_stag = 0; 162 163 c_tx->ctrl_len = sizeof(struct iwarp_send); 164 165 crc = (char *)&c_tx->pkt.send_pkt.crc; 166 data = siw_try_1seg(c_tx, crc); 167 break; 168 169 case SIW_OP_SEND_REMOTE_INV: 170 if (tx_flags(wqe) & SIW_WQE_SOLICITED) 171 memcpy(&c_tx->pkt.ctrl, 172 &iwarp_pktinfo[RDMAP_SEND_SE_INVAL].ctrl, 173 sizeof(struct iwarp_ctrl)); 174 else 175 memcpy(&c_tx->pkt.ctrl, 176 &iwarp_pktinfo[RDMAP_SEND_INVAL].ctrl, 177 sizeof(struct iwarp_ctrl)); 178 179 c_tx->pkt.send.ddp_qn = RDMAP_UNTAGGED_QN_SEND; 180 c_tx->pkt.send.ddp_msn = 181 htonl(++c_tx->ddp_msn[RDMAP_UNTAGGED_QN_SEND]); 182 c_tx->pkt.send.ddp_mo = 0; 183 184 c_tx->pkt.send_inv.inval_stag = cpu_to_be32(wqe->sqe.rkey); 185 186 c_tx->ctrl_len = sizeof(struct iwarp_send_inv); 187 188 crc = (char *)&c_tx->pkt.send_pkt.crc; 189 data = siw_try_1seg(c_tx, crc); 190 break; 191 192 case SIW_OP_WRITE: 193 memcpy(&c_tx->pkt.ctrl, &iwarp_pktinfo[RDMAP_RDMA_WRITE].ctrl, 194 sizeof(struct iwarp_ctrl)); 195 196 c_tx->pkt.rwrite.sink_stag = htonl(wqe->sqe.rkey); 197 c_tx->pkt.rwrite.sink_to = cpu_to_be64(wqe->sqe.raddr); 198 c_tx->ctrl_len = sizeof(struct iwarp_rdma_write); 199 200 crc = (char *)&c_tx->pkt.write_pkt.crc; 201 data = siw_try_1seg(c_tx, crc); 202 break; 203 204 case SIW_OP_READ_RESPONSE: 205 memcpy(&c_tx->pkt.ctrl, 206 &iwarp_pktinfo[RDMAP_RDMA_READ_RESP].ctrl, 207 sizeof(struct iwarp_ctrl)); 208 209 /* NBO */ 210 c_tx->pkt.rresp.sink_stag = cpu_to_be32(wqe->sqe.rkey); 211 c_tx->pkt.rresp.sink_to = cpu_to_be64(wqe->sqe.raddr); 212 213 c_tx->ctrl_len = sizeof(struct iwarp_rdma_rresp); 214 215 crc = (char *)&c_tx->pkt.write_pkt.crc; 216 data = siw_try_1seg(c_tx, crc); 217 break; 218 219 default: 220 siw_dbg_qp(tx_qp(c_tx), "stale wqe type %d\n", tx_type(wqe)); 221 return -EOPNOTSUPP; 222 } 223 if (unlikely(data < 0)) 224 return data; 225 226 c_tx->ctrl_sent = 0; 227 228 if (data <= MAX_HDR_INLINE) { 229 if (data) { 230 wqe->processed = data; 231 232 c_tx->pkt.ctrl.mpa_len = 233 htons(c_tx->ctrl_len + data - MPA_HDR_SIZE); 234 235 /* Add pad, if needed */ 236 data += -(int)data & 0x3; 237 /* advance CRC location after payload */ 238 crc += data; 239 c_tx->ctrl_len += data; 240 241 if (!(c_tx->pkt.ctrl.ddp_rdmap_ctrl & DDP_FLAG_TAGGED)) 242 c_tx->pkt.c_untagged.ddp_mo = 0; 243 else 244 c_tx->pkt.c_tagged.ddp_to = 245 cpu_to_be64(wqe->sqe.raddr); 246 } 247 248 *(u32 *)crc = 0; 249 /* 250 * Do complete CRC if enabled and short packet 251 */ 252 if (c_tx->mpa_crc_hd) { 253 crypto_shash_init(c_tx->mpa_crc_hd); 254 if (crypto_shash_update(c_tx->mpa_crc_hd, 255 (u8 *)&c_tx->pkt, 256 c_tx->ctrl_len)) 257 return -EINVAL; 258 crypto_shash_final(c_tx->mpa_crc_hd, (u8 *)crc); 259 } 260 c_tx->ctrl_len += MPA_CRC_SIZE; 261 262 return PKT_COMPLETE; 263 } 264 c_tx->ctrl_len += MPA_CRC_SIZE; 265 c_tx->sge_idx = 0; 266 c_tx->sge_off = 0; 267 c_tx->pbl_idx = 0; 268 269 /* 270 * Allow direct sending out of user buffer if WR is non signalled 271 * and payload is over threshold. 272 * Per RDMA verbs, the application should not change the send buffer 273 * until the work completed. In iWarp, work completion is only 274 * local delivery to TCP. TCP may reuse the buffer for 275 * retransmission. Changing unsent data also breaks the CRC, 276 * if applied. 277 */ 278 if (c_tx->zcopy_tx && wqe->bytes >= SENDPAGE_THRESH && 279 !(tx_flags(wqe) & SIW_WQE_SIGNALLED)) 280 c_tx->use_sendpage = 1; 281 else 282 c_tx->use_sendpage = 0; 283 284 return PKT_FRAGMENTED; 285 } 286 287 /* 288 * Send out one complete control type FPDU, or header of FPDU carrying 289 * data. Used for fixed sized packets like Read.Requests or zero length 290 * SENDs, WRITEs, READ.Responses, or header only. 291 */ 292 static int siw_tx_ctrl(struct siw_iwarp_tx *c_tx, struct socket *s, 293 int flags) 294 { 295 struct msghdr msg = { .msg_flags = flags }; 296 struct kvec iov = { .iov_base = 297 (char *)&c_tx->pkt.ctrl + c_tx->ctrl_sent, 298 .iov_len = c_tx->ctrl_len - c_tx->ctrl_sent }; 299 300 int rv = kernel_sendmsg(s, &msg, &iov, 1, 301 c_tx->ctrl_len - c_tx->ctrl_sent); 302 303 if (rv >= 0) { 304 c_tx->ctrl_sent += rv; 305 306 if (c_tx->ctrl_sent == c_tx->ctrl_len) 307 rv = 0; 308 else 309 rv = -EAGAIN; 310 } 311 return rv; 312 } 313 314 /* 315 * 0copy TCP transmit interface: Use MSG_SPLICE_PAGES. 316 * 317 * Using sendpage to push page by page appears to be less efficient 318 * than using sendmsg, even if data are copied. 319 * 320 * A general performance limitation might be the extra four bytes 321 * trailer checksum segment to be pushed after user data. 322 */ 323 static int siw_tcp_sendpages(struct socket *s, struct page **page, int offset, 324 size_t size) 325 { 326 struct bio_vec bvec; 327 struct msghdr msg = { 328 .msg_flags = (MSG_MORE | MSG_DONTWAIT | MSG_SPLICE_PAGES), 329 }; 330 struct sock *sk = s->sk; 331 int i = 0, rv = 0, sent = 0; 332 333 while (size) { 334 size_t bytes = min_t(size_t, PAGE_SIZE - offset, size); 335 336 if (size + offset <= PAGE_SIZE) 337 msg.msg_flags &= ~MSG_MORE; 338 339 tcp_rate_check_app_limited(sk); 340 bvec_set_page(&bvec, page[i], bytes, offset); 341 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, size); 342 343 try_page_again: 344 lock_sock(sk); 345 rv = tcp_sendmsg_locked(sk, &msg, size); 346 release_sock(sk); 347 348 if (rv > 0) { 349 size -= rv; 350 sent += rv; 351 if (rv != bytes) { 352 offset += rv; 353 bytes -= rv; 354 goto try_page_again; 355 } 356 offset = 0; 357 } else { 358 if (rv == -EAGAIN || rv == 0) 359 break; 360 return rv; 361 } 362 i++; 363 } 364 return sent; 365 } 366 367 /* 368 * siw_0copy_tx() 369 * 370 * Pushes list of pages to TCP socket. If pages from multiple 371 * SGE's, all referenced pages of each SGE are pushed in one 372 * shot. 373 */ 374 static int siw_0copy_tx(struct socket *s, struct page **page, 375 struct siw_sge *sge, unsigned int offset, 376 unsigned int size) 377 { 378 int i = 0, sent = 0, rv; 379 int sge_bytes = min(sge->length - offset, size); 380 381 offset = (sge->laddr + offset) & ~PAGE_MASK; 382 383 while (sent != size) { 384 rv = siw_tcp_sendpages(s, &page[i], offset, sge_bytes); 385 if (rv >= 0) { 386 sent += rv; 387 if (size == sent || sge_bytes > rv) 388 break; 389 390 i += PAGE_ALIGN(sge_bytes + offset) >> PAGE_SHIFT; 391 sge++; 392 sge_bytes = min(sge->length, size - sent); 393 offset = sge->laddr & ~PAGE_MASK; 394 } else { 395 sent = rv; 396 break; 397 } 398 } 399 return sent; 400 } 401 402 #define MAX_TRAILER (MPA_CRC_SIZE + 4) 403 404 static void siw_unmap_pages(struct kvec *iov, unsigned long kmap_mask, int len) 405 { 406 int i; 407 408 /* 409 * Work backwards through the array to honor the kmap_local_page() 410 * ordering requirements. 411 */ 412 for (i = (len-1); i >= 0; i--) { 413 if (kmap_mask & BIT(i)) { 414 unsigned long addr = (unsigned long)iov[i].iov_base; 415 416 kunmap_local((void *)(addr & PAGE_MASK)); 417 } 418 } 419 } 420 421 /* 422 * siw_tx_hdt() tries to push a complete packet to TCP where all 423 * packet fragments are referenced by the elements of one iovec. 424 * For the data portion, each involved page must be referenced by 425 * one extra element. All sge's data can be non-aligned to page 426 * boundaries. Two more elements are referencing iWARP header 427 * and trailer: 428 * MAX_ARRAY = 64KB/PAGE_SIZE + 1 + (2 * (SIW_MAX_SGE - 1) + HDR + TRL 429 */ 430 #define MAX_ARRAY ((0xffff / PAGE_SIZE) + 1 + (2 * (SIW_MAX_SGE - 1) + 2)) 431 432 /* 433 * Write out iov referencing hdr, data and trailer of current FPDU. 434 * Update transmit state dependent on write return status 435 */ 436 static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s) 437 { 438 struct siw_wqe *wqe = &c_tx->wqe_active; 439 struct siw_sge *sge = &wqe->sqe.sge[c_tx->sge_idx]; 440 struct kvec iov[MAX_ARRAY]; 441 struct page *page_array[MAX_ARRAY]; 442 struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_EOR }; 443 444 int seg = 0, do_crc = c_tx->do_crc, is_kva = 0, rv; 445 unsigned int data_len = c_tx->bytes_unsent, hdr_len = 0, trl_len = 0, 446 sge_off = c_tx->sge_off, sge_idx = c_tx->sge_idx, 447 pbl_idx = c_tx->pbl_idx; 448 unsigned long kmap_mask = 0L; 449 450 if (c_tx->state == SIW_SEND_HDR) { 451 if (c_tx->use_sendpage) { 452 rv = siw_tx_ctrl(c_tx, s, MSG_DONTWAIT | MSG_MORE); 453 if (rv) 454 goto done; 455 456 c_tx->state = SIW_SEND_DATA; 457 } else { 458 iov[0].iov_base = 459 (char *)&c_tx->pkt.ctrl + c_tx->ctrl_sent; 460 iov[0].iov_len = hdr_len = 461 c_tx->ctrl_len - c_tx->ctrl_sent; 462 seg = 1; 463 } 464 } 465 466 wqe->processed += data_len; 467 468 while (data_len) { /* walk the list of SGE's */ 469 unsigned int sge_len = min(sge->length - sge_off, data_len); 470 unsigned int fp_off = (sge->laddr + sge_off) & ~PAGE_MASK; 471 struct siw_mem *mem; 472 473 if (!(tx_flags(wqe) & SIW_WQE_INLINE)) { 474 mem = wqe->mem[sge_idx]; 475 is_kva = mem->mem_obj == NULL ? 1 : 0; 476 } else { 477 is_kva = 1; 478 } 479 if (is_kva && !c_tx->use_sendpage) { 480 /* 481 * tx from kernel virtual address: either inline data 482 * or memory region with assigned kernel buffer 483 */ 484 iov[seg].iov_base = 485 ib_virt_dma_to_ptr(sge->laddr + sge_off); 486 iov[seg].iov_len = sge_len; 487 488 if (do_crc) 489 crypto_shash_update(c_tx->mpa_crc_hd, 490 iov[seg].iov_base, 491 sge_len); 492 sge_off += sge_len; 493 data_len -= sge_len; 494 seg++; 495 goto sge_done; 496 } 497 498 while (sge_len) { 499 size_t plen = min((int)PAGE_SIZE - fp_off, sge_len); 500 void *kaddr; 501 502 if (!is_kva) { 503 struct page *p; 504 505 if (mem->is_pbl) 506 p = siw_get_pblpage( 507 mem, sge->laddr + sge_off, 508 &pbl_idx); 509 else 510 p = siw_get_upage(mem->umem, 511 sge->laddr + sge_off); 512 if (unlikely(!p)) { 513 siw_unmap_pages(iov, kmap_mask, seg); 514 wqe->processed -= c_tx->bytes_unsent; 515 rv = -EFAULT; 516 goto done_crc; 517 } 518 page_array[seg] = p; 519 520 if (!c_tx->use_sendpage) { 521 void *kaddr = kmap_local_page(p); 522 523 /* Remember for later kunmap() */ 524 kmap_mask |= BIT(seg); 525 iov[seg].iov_base = kaddr + fp_off; 526 iov[seg].iov_len = plen; 527 528 if (do_crc) 529 crypto_shash_update( 530 c_tx->mpa_crc_hd, 531 iov[seg].iov_base, 532 plen); 533 } else if (do_crc) { 534 kaddr = kmap_local_page(p); 535 crypto_shash_update(c_tx->mpa_crc_hd, 536 kaddr + fp_off, 537 plen); 538 kunmap_local(kaddr); 539 } 540 } else { 541 /* 542 * Cast to an uintptr_t to preserve all 64 bits 543 * in sge->laddr. 544 */ 545 u64 va = sge->laddr + sge_off; 546 547 page_array[seg] = ib_virt_dma_to_page(va); 548 if (do_crc) 549 crypto_shash_update( 550 c_tx->mpa_crc_hd, 551 ib_virt_dma_to_ptr(va), 552 plen); 553 } 554 555 sge_len -= plen; 556 sge_off += plen; 557 data_len -= plen; 558 fp_off = 0; 559 560 if (++seg >= (int)MAX_ARRAY) { 561 siw_dbg_qp(tx_qp(c_tx), "to many fragments\n"); 562 siw_unmap_pages(iov, kmap_mask, seg-1); 563 wqe->processed -= c_tx->bytes_unsent; 564 rv = -EMSGSIZE; 565 goto done_crc; 566 } 567 } 568 sge_done: 569 /* Update SGE variables at end of SGE */ 570 if (sge_off == sge->length && 571 (data_len != 0 || wqe->processed < wqe->bytes)) { 572 sge_idx++; 573 sge++; 574 sge_off = 0; 575 } 576 } 577 /* trailer */ 578 if (likely(c_tx->state != SIW_SEND_TRAILER)) { 579 iov[seg].iov_base = &c_tx->trailer.pad[4 - c_tx->pad]; 580 iov[seg].iov_len = trl_len = MAX_TRAILER - (4 - c_tx->pad); 581 } else { 582 iov[seg].iov_base = &c_tx->trailer.pad[c_tx->ctrl_sent]; 583 iov[seg].iov_len = trl_len = MAX_TRAILER - c_tx->ctrl_sent; 584 } 585 586 if (c_tx->pad) { 587 *(u32 *)c_tx->trailer.pad = 0; 588 if (do_crc) 589 crypto_shash_update(c_tx->mpa_crc_hd, 590 (u8 *)&c_tx->trailer.crc - c_tx->pad, 591 c_tx->pad); 592 } 593 if (!c_tx->mpa_crc_hd) 594 c_tx->trailer.crc = 0; 595 else if (do_crc) 596 crypto_shash_final(c_tx->mpa_crc_hd, (u8 *)&c_tx->trailer.crc); 597 598 data_len = c_tx->bytes_unsent; 599 600 if (c_tx->use_sendpage) { 601 rv = siw_0copy_tx(s, page_array, &wqe->sqe.sge[c_tx->sge_idx], 602 c_tx->sge_off, data_len); 603 if (rv == data_len) { 604 rv = kernel_sendmsg(s, &msg, &iov[seg], 1, trl_len); 605 if (rv > 0) 606 rv += data_len; 607 else 608 rv = data_len; 609 } 610 } else { 611 rv = kernel_sendmsg(s, &msg, iov, seg + 1, 612 hdr_len + data_len + trl_len); 613 siw_unmap_pages(iov, kmap_mask, seg); 614 } 615 if (rv < (int)hdr_len) { 616 /* Not even complete hdr pushed or negative rv */ 617 wqe->processed -= data_len; 618 if (rv >= 0) { 619 c_tx->ctrl_sent += rv; 620 rv = -EAGAIN; 621 } 622 goto done_crc; 623 } 624 rv -= hdr_len; 625 626 if (rv >= (int)data_len) { 627 /* all user data pushed to TCP or no data to push */ 628 if (data_len > 0 && wqe->processed < wqe->bytes) { 629 /* Save the current state for next tx */ 630 c_tx->sge_idx = sge_idx; 631 c_tx->sge_off = sge_off; 632 c_tx->pbl_idx = pbl_idx; 633 } 634 rv -= data_len; 635 636 if (rv == trl_len) /* all pushed */ 637 rv = 0; 638 else { 639 c_tx->state = SIW_SEND_TRAILER; 640 c_tx->ctrl_len = MAX_TRAILER; 641 c_tx->ctrl_sent = rv + 4 - c_tx->pad; 642 c_tx->bytes_unsent = 0; 643 rv = -EAGAIN; 644 } 645 646 } else if (data_len > 0) { 647 /* Maybe some user data pushed to TCP */ 648 c_tx->state = SIW_SEND_DATA; 649 wqe->processed -= data_len - rv; 650 651 if (rv) { 652 /* 653 * Some bytes out. Recompute tx state based 654 * on old state and bytes pushed 655 */ 656 unsigned int sge_unsent; 657 658 c_tx->bytes_unsent -= rv; 659 sge = &wqe->sqe.sge[c_tx->sge_idx]; 660 sge_unsent = sge->length - c_tx->sge_off; 661 662 while (sge_unsent <= rv) { 663 rv -= sge_unsent; 664 c_tx->sge_idx++; 665 c_tx->sge_off = 0; 666 sge++; 667 sge_unsent = sge->length; 668 } 669 c_tx->sge_off += rv; 670 } 671 rv = -EAGAIN; 672 } 673 done_crc: 674 c_tx->do_crc = 0; 675 done: 676 return rv; 677 } 678 679 static void siw_update_tcpseg(struct siw_iwarp_tx *c_tx, 680 struct socket *s) 681 { 682 struct tcp_sock *tp = tcp_sk(s->sk); 683 684 if (tp->gso_segs) { 685 if (c_tx->gso_seg_limit == 0) 686 c_tx->tcp_seglen = tp->mss_cache * tp->gso_segs; 687 else 688 c_tx->tcp_seglen = 689 tp->mss_cache * 690 min_t(u16, c_tx->gso_seg_limit, tp->gso_segs); 691 } else { 692 c_tx->tcp_seglen = tp->mss_cache; 693 } 694 /* Loopback may give odd numbers */ 695 c_tx->tcp_seglen &= 0xfffffff8; 696 } 697 698 /* 699 * siw_prepare_fpdu() 700 * 701 * Prepares transmit context to send out one FPDU if FPDU will contain 702 * user data and user data are not immediate data. 703 * Computes maximum FPDU length to fill up TCP MSS if possible. 704 * 705 * @qp: QP from which to transmit 706 * @wqe: Current WQE causing transmission 707 * 708 * TODO: Take into account real available sendspace on socket 709 * to avoid header misalignment due to send pausing within 710 * fpdu transmission 711 */ 712 static void siw_prepare_fpdu(struct siw_qp *qp, struct siw_wqe *wqe) 713 { 714 struct siw_iwarp_tx *c_tx = &qp->tx_ctx; 715 int data_len; 716 717 c_tx->ctrl_len = 718 iwarp_pktinfo[__rdmap_get_opcode(&c_tx->pkt.ctrl)].hdr_len; 719 c_tx->ctrl_sent = 0; 720 721 /* 722 * Update target buffer offset if any 723 */ 724 if (!(c_tx->pkt.ctrl.ddp_rdmap_ctrl & DDP_FLAG_TAGGED)) 725 /* Untagged message */ 726 c_tx->pkt.c_untagged.ddp_mo = cpu_to_be32(wqe->processed); 727 else /* Tagged message */ 728 c_tx->pkt.c_tagged.ddp_to = 729 cpu_to_be64(wqe->sqe.raddr + wqe->processed); 730 731 data_len = wqe->bytes - wqe->processed; 732 if (data_len + c_tx->ctrl_len + MPA_CRC_SIZE > c_tx->tcp_seglen) { 733 /* Trim DDP payload to fit into current TCP segment */ 734 data_len = c_tx->tcp_seglen - (c_tx->ctrl_len + MPA_CRC_SIZE); 735 c_tx->pkt.ctrl.ddp_rdmap_ctrl &= ~DDP_FLAG_LAST; 736 c_tx->pad = 0; 737 } else { 738 c_tx->pkt.ctrl.ddp_rdmap_ctrl |= DDP_FLAG_LAST; 739 c_tx->pad = -data_len & 0x3; 740 } 741 c_tx->bytes_unsent = data_len; 742 743 c_tx->pkt.ctrl.mpa_len = 744 htons(c_tx->ctrl_len + data_len - MPA_HDR_SIZE); 745 746 /* 747 * Init MPA CRC computation 748 */ 749 if (c_tx->mpa_crc_hd) { 750 crypto_shash_init(c_tx->mpa_crc_hd); 751 crypto_shash_update(c_tx->mpa_crc_hd, (u8 *)&c_tx->pkt, 752 c_tx->ctrl_len); 753 c_tx->do_crc = 1; 754 } 755 } 756 757 /* 758 * siw_check_sgl_tx() 759 * 760 * Check permissions for a list of SGE's (SGL). 761 * A successful check will have all memory referenced 762 * for transmission resolved and assigned to the WQE. 763 * 764 * @pd: Protection Domain SGL should belong to 765 * @wqe: WQE to be checked 766 * @perms: requested access permissions 767 * 768 */ 769 770 static int siw_check_sgl_tx(struct ib_pd *pd, struct siw_wqe *wqe, 771 enum ib_access_flags perms) 772 { 773 struct siw_sge *sge = &wqe->sqe.sge[0]; 774 int i, len, num_sge = wqe->sqe.num_sge; 775 776 if (unlikely(num_sge > SIW_MAX_SGE)) 777 return -EINVAL; 778 779 for (i = 0, len = 0; num_sge; num_sge--, i++, sge++) { 780 /* 781 * rdma verbs: do not check stag for a zero length sge 782 */ 783 if (sge->length) { 784 int rv = siw_check_sge(pd, sge, &wqe->mem[i], perms, 0, 785 sge->length); 786 787 if (unlikely(rv != E_ACCESS_OK)) 788 return rv; 789 } 790 len += sge->length; 791 } 792 return len; 793 } 794 795 /* 796 * siw_qp_sq_proc_tx() 797 * 798 * Process one WQE which needs transmission on the wire. 799 */ 800 static int siw_qp_sq_proc_tx(struct siw_qp *qp, struct siw_wqe *wqe) 801 { 802 struct siw_iwarp_tx *c_tx = &qp->tx_ctx; 803 struct socket *s = qp->attrs.sk; 804 int rv = 0, burst_len = qp->tx_ctx.burst; 805 enum rdmap_ecode ecode = RDMAP_ECODE_CATASTROPHIC_STREAM; 806 807 if (unlikely(wqe->wr_status == SIW_WR_IDLE)) 808 return 0; 809 810 if (!burst_len) 811 burst_len = SQ_USER_MAXBURST; 812 813 if (wqe->wr_status == SIW_WR_QUEUED) { 814 if (!(wqe->sqe.flags & SIW_WQE_INLINE)) { 815 if (tx_type(wqe) == SIW_OP_READ_RESPONSE) 816 wqe->sqe.num_sge = 1; 817 818 if (tx_type(wqe) != SIW_OP_READ && 819 tx_type(wqe) != SIW_OP_READ_LOCAL_INV) { 820 /* 821 * Reference memory to be tx'd w/o checking 822 * access for LOCAL_READ permission, since 823 * not defined in RDMA core. 824 */ 825 rv = siw_check_sgl_tx(qp->pd, wqe, 0); 826 if (rv < 0) { 827 if (tx_type(wqe) == 828 SIW_OP_READ_RESPONSE) 829 ecode = siw_rdmap_error(-rv); 830 rv = -EINVAL; 831 goto tx_error; 832 } 833 wqe->bytes = rv; 834 } else { 835 wqe->bytes = 0; 836 } 837 } else { 838 wqe->bytes = wqe->sqe.sge[0].length; 839 if (!rdma_is_kernel_res(&qp->base_qp.res)) { 840 if (wqe->bytes > SIW_MAX_INLINE) { 841 rv = -EINVAL; 842 goto tx_error; 843 } 844 wqe->sqe.sge[0].laddr = 845 (u64)(uintptr_t)&wqe->sqe.sge[1]; 846 } 847 } 848 wqe->wr_status = SIW_WR_INPROGRESS; 849 wqe->processed = 0; 850 851 siw_update_tcpseg(c_tx, s); 852 853 rv = siw_qp_prepare_tx(c_tx); 854 if (rv == PKT_FRAGMENTED) { 855 c_tx->state = SIW_SEND_HDR; 856 siw_prepare_fpdu(qp, wqe); 857 } else if (rv == PKT_COMPLETE) { 858 c_tx->state = SIW_SEND_SHORT_FPDU; 859 } else { 860 goto tx_error; 861 } 862 } 863 864 next_segment: 865 siw_dbg_qp(qp, "wr type %d, state %d, data %u, sent %u, id %llx\n", 866 tx_type(wqe), wqe->wr_status, wqe->bytes, wqe->processed, 867 wqe->sqe.id); 868 869 if (--burst_len == 0) { 870 rv = -EINPROGRESS; 871 goto tx_done; 872 } 873 if (c_tx->state == SIW_SEND_SHORT_FPDU) { 874 enum siw_opcode tx_type = tx_type(wqe); 875 unsigned int msg_flags; 876 877 if (siw_sq_empty(qp) || !siw_tcp_nagle || burst_len == 1) 878 /* 879 * End current TCP segment, if SQ runs empty, 880 * or siw_tcp_nagle is not set, or we bail out 881 * soon due to no burst credit left. 882 */ 883 msg_flags = MSG_DONTWAIT; 884 else 885 msg_flags = MSG_DONTWAIT | MSG_MORE; 886 887 rv = siw_tx_ctrl(c_tx, s, msg_flags); 888 889 if (!rv && tx_type != SIW_OP_READ && 890 tx_type != SIW_OP_READ_LOCAL_INV) 891 wqe->processed = wqe->bytes; 892 893 goto tx_done; 894 895 } else { 896 rv = siw_tx_hdt(c_tx, s); 897 } 898 if (!rv) { 899 /* 900 * One segment sent. Processing completed if last 901 * segment, Do next segment otherwise. 902 */ 903 if (unlikely(c_tx->tx_suspend)) { 904 /* 905 * Verbs, 6.4.: Try stopping sending after a full 906 * DDP segment if the connection goes down 907 * (== peer halfclose) 908 */ 909 rv = -ECONNABORTED; 910 goto tx_done; 911 } 912 if (c_tx->pkt.ctrl.ddp_rdmap_ctrl & DDP_FLAG_LAST) { 913 siw_dbg_qp(qp, "WQE completed\n"); 914 goto tx_done; 915 } 916 c_tx->state = SIW_SEND_HDR; 917 918 siw_update_tcpseg(c_tx, s); 919 920 siw_prepare_fpdu(qp, wqe); 921 goto next_segment; 922 } 923 tx_done: 924 qp->tx_ctx.burst = burst_len; 925 return rv; 926 927 tx_error: 928 if (ecode != RDMAP_ECODE_CATASTROPHIC_STREAM) 929 siw_init_terminate(qp, TERM_ERROR_LAYER_RDMAP, 930 RDMAP_ETYPE_REMOTE_PROTECTION, ecode, 1); 931 else 932 siw_init_terminate(qp, TERM_ERROR_LAYER_RDMAP, 933 RDMAP_ETYPE_CATASTROPHIC, 934 RDMAP_ECODE_UNSPECIFIED, 1); 935 return rv; 936 } 937 938 static int siw_fastreg_mr(struct ib_pd *pd, struct siw_sqe *sqe) 939 { 940 struct ib_mr *base_mr = (struct ib_mr *)(uintptr_t)sqe->base_mr; 941 struct siw_device *sdev = to_siw_dev(pd->device); 942 struct siw_mem *mem; 943 int rv = 0; 944 945 siw_dbg_pd(pd, "STag 0x%08x\n", sqe->rkey); 946 947 if (unlikely(!base_mr)) { 948 pr_warn("siw: fastreg: STag 0x%08x unknown\n", sqe->rkey); 949 return -EINVAL; 950 } 951 952 if (unlikely(base_mr->rkey >> 8 != sqe->rkey >> 8)) { 953 pr_warn("siw: fastreg: STag 0x%08x: bad MR\n", sqe->rkey); 954 return -EINVAL; 955 } 956 957 mem = siw_mem_id2obj(sdev, sqe->rkey >> 8); 958 if (unlikely(!mem)) { 959 pr_warn("siw: fastreg: STag 0x%08x unknown\n", sqe->rkey); 960 return -EINVAL; 961 } 962 963 if (unlikely(mem->pd != pd)) { 964 pr_warn("siw: fastreg: PD mismatch\n"); 965 rv = -EINVAL; 966 goto out; 967 } 968 if (unlikely(mem->stag_valid)) { 969 pr_warn("siw: fastreg: STag 0x%08x already valid\n", sqe->rkey); 970 rv = -EINVAL; 971 goto out; 972 } 973 /* Refresh STag since user may have changed key part */ 974 mem->stag = sqe->rkey; 975 mem->perms = sqe->access; 976 977 siw_dbg_mem(mem, "STag 0x%08x now valid\n", sqe->rkey); 978 mem->va = base_mr->iova; 979 mem->stag_valid = 1; 980 out: 981 siw_mem_put(mem); 982 return rv; 983 } 984 985 static int siw_qp_sq_proc_local(struct siw_qp *qp, struct siw_wqe *wqe) 986 { 987 int rv; 988 989 switch (tx_type(wqe)) { 990 case SIW_OP_REG_MR: 991 rv = siw_fastreg_mr(qp->pd, &wqe->sqe); 992 break; 993 994 case SIW_OP_INVAL_STAG: 995 rv = siw_invalidate_stag(qp->pd, wqe->sqe.rkey); 996 break; 997 998 default: 999 rv = -EINVAL; 1000 } 1001 return rv; 1002 } 1003 1004 /* 1005 * siw_qp_sq_process() 1006 * 1007 * Core TX path routine for RDMAP/DDP/MPA using a TCP kernel socket. 1008 * Sends RDMAP payload for the current SQ WR @wqe of @qp in one or more 1009 * MPA FPDUs, each containing a DDP segment. 1010 * 1011 * SQ processing may occur in user context as a result of posting 1012 * new WQE's or from siw_sq_work_handler() context. Processing in 1013 * user context is limited to non-kernel verbs users. 1014 * 1015 * SQ processing may get paused anytime, possibly in the middle of a WR 1016 * or FPDU, if insufficient send space is available. SQ processing 1017 * gets resumed from siw_sq_work_handler(), if send space becomes 1018 * available again. 1019 * 1020 * Must be called with the QP state read-locked. 1021 * 1022 * Note: 1023 * An outbound RREQ can be satisfied by the corresponding RRESP 1024 * _before_ it gets assigned to the ORQ. This happens regularly 1025 * in RDMA READ via loopback case. Since both outbound RREQ and 1026 * inbound RRESP can be handled by the same CPU, locking the ORQ 1027 * is dead-lock prone and thus not an option. With that, the 1028 * RREQ gets assigned to the ORQ _before_ being sent - see 1029 * siw_activate_tx() - and pulled back in case of send failure. 1030 */ 1031 int siw_qp_sq_process(struct siw_qp *qp) 1032 { 1033 struct siw_wqe *wqe = tx_wqe(qp); 1034 enum siw_opcode tx_type; 1035 unsigned long flags; 1036 int rv = 0; 1037 1038 siw_dbg_qp(qp, "enter for type %d\n", tx_type(wqe)); 1039 1040 next_wqe: 1041 /* 1042 * Stop QP processing if SQ state changed 1043 */ 1044 if (unlikely(qp->tx_ctx.tx_suspend)) { 1045 siw_dbg_qp(qp, "tx suspended\n"); 1046 goto done; 1047 } 1048 tx_type = tx_type(wqe); 1049 1050 if (tx_type <= SIW_OP_READ_RESPONSE) 1051 rv = siw_qp_sq_proc_tx(qp, wqe); 1052 else 1053 rv = siw_qp_sq_proc_local(qp, wqe); 1054 1055 if (!rv) { 1056 /* 1057 * WQE processing done 1058 */ 1059 switch (tx_type) { 1060 case SIW_OP_SEND: 1061 case SIW_OP_SEND_REMOTE_INV: 1062 case SIW_OP_WRITE: 1063 siw_wqe_put_mem(wqe, tx_type); 1064 fallthrough; 1065 1066 case SIW_OP_INVAL_STAG: 1067 case SIW_OP_REG_MR: 1068 if (tx_flags(wqe) & SIW_WQE_SIGNALLED) 1069 siw_sqe_complete(qp, &wqe->sqe, wqe->bytes, 1070 SIW_WC_SUCCESS); 1071 break; 1072 1073 case SIW_OP_READ: 1074 case SIW_OP_READ_LOCAL_INV: 1075 /* 1076 * already enqueued to ORQ queue 1077 */ 1078 break; 1079 1080 case SIW_OP_READ_RESPONSE: 1081 siw_wqe_put_mem(wqe, tx_type); 1082 break; 1083 1084 default: 1085 WARN(1, "undefined WQE type %d\n", tx_type); 1086 rv = -EINVAL; 1087 goto done; 1088 } 1089 1090 spin_lock_irqsave(&qp->sq_lock, flags); 1091 wqe->wr_status = SIW_WR_IDLE; 1092 rv = siw_activate_tx(qp); 1093 spin_unlock_irqrestore(&qp->sq_lock, flags); 1094 1095 if (rv <= 0) 1096 goto done; 1097 1098 goto next_wqe; 1099 1100 } else if (rv == -EAGAIN) { 1101 siw_dbg_qp(qp, "sq paused: hd/tr %d of %d, data %d\n", 1102 qp->tx_ctx.ctrl_sent, qp->tx_ctx.ctrl_len, 1103 qp->tx_ctx.bytes_unsent); 1104 rv = 0; 1105 goto done; 1106 } else if (rv == -EINPROGRESS) { 1107 rv = siw_sq_start(qp); 1108 goto done; 1109 } else { 1110 /* 1111 * WQE processing failed. 1112 * Verbs 8.3.2: 1113 * o It turns any WQE into a signalled WQE. 1114 * o Local catastrophic error must be surfaced 1115 * o QP must be moved into Terminate state: done by code 1116 * doing socket state change processing 1117 * 1118 * o TODO: Termination message must be sent. 1119 * o TODO: Implement more precise work completion errors, 1120 * see enum ib_wc_status in ib_verbs.h 1121 */ 1122 siw_dbg_qp(qp, "wqe type %d processing failed: %d\n", 1123 tx_type(wqe), rv); 1124 1125 spin_lock_irqsave(&qp->sq_lock, flags); 1126 /* 1127 * RREQ may have already been completed by inbound RRESP! 1128 */ 1129 if ((tx_type == SIW_OP_READ || 1130 tx_type == SIW_OP_READ_LOCAL_INV) && qp->attrs.orq_size) { 1131 /* Cleanup pending entry in ORQ */ 1132 qp->orq_put--; 1133 qp->orq[qp->orq_put % qp->attrs.orq_size].flags = 0; 1134 } 1135 spin_unlock_irqrestore(&qp->sq_lock, flags); 1136 /* 1137 * immediately suspends further TX processing 1138 */ 1139 if (!qp->tx_ctx.tx_suspend) 1140 siw_qp_cm_drop(qp, 0); 1141 1142 switch (tx_type) { 1143 case SIW_OP_SEND: 1144 case SIW_OP_SEND_REMOTE_INV: 1145 case SIW_OP_SEND_WITH_IMM: 1146 case SIW_OP_WRITE: 1147 case SIW_OP_READ: 1148 case SIW_OP_READ_LOCAL_INV: 1149 siw_wqe_put_mem(wqe, tx_type); 1150 fallthrough; 1151 1152 case SIW_OP_INVAL_STAG: 1153 case SIW_OP_REG_MR: 1154 siw_sqe_complete(qp, &wqe->sqe, wqe->bytes, 1155 SIW_WC_LOC_QP_OP_ERR); 1156 1157 siw_qp_event(qp, IB_EVENT_QP_FATAL); 1158 1159 break; 1160 1161 case SIW_OP_READ_RESPONSE: 1162 siw_dbg_qp(qp, "proc. read.response failed: %d\n", rv); 1163 1164 siw_qp_event(qp, IB_EVENT_QP_REQ_ERR); 1165 1166 siw_wqe_put_mem(wqe, SIW_OP_READ_RESPONSE); 1167 1168 break; 1169 1170 default: 1171 WARN(1, "undefined WQE type %d\n", tx_type); 1172 rv = -EINVAL; 1173 } 1174 wqe->wr_status = SIW_WR_IDLE; 1175 } 1176 done: 1177 return rv; 1178 } 1179 1180 static void siw_sq_resume(struct siw_qp *qp) 1181 { 1182 if (down_read_trylock(&qp->state_lock)) { 1183 if (likely(qp->attrs.state == SIW_QP_STATE_RTS && 1184 !qp->tx_ctx.tx_suspend)) { 1185 int rv = siw_qp_sq_process(qp); 1186 1187 up_read(&qp->state_lock); 1188 1189 if (unlikely(rv < 0)) { 1190 siw_dbg_qp(qp, "SQ task failed: err %d\n", rv); 1191 1192 if (!qp->tx_ctx.tx_suspend) 1193 siw_qp_cm_drop(qp, 0); 1194 } 1195 } else { 1196 up_read(&qp->state_lock); 1197 } 1198 } else { 1199 siw_dbg_qp(qp, "Resume SQ while QP locked\n"); 1200 } 1201 siw_qp_put(qp); 1202 } 1203 1204 struct tx_task_t { 1205 struct llist_head active; 1206 wait_queue_head_t waiting; 1207 }; 1208 1209 static DEFINE_PER_CPU(struct tx_task_t, siw_tx_task_g); 1210 1211 void siw_stop_tx_thread(int nr_cpu) 1212 { 1213 kthread_stop(siw_tx_thread[nr_cpu]); 1214 wake_up(&per_cpu(siw_tx_task_g, nr_cpu).waiting); 1215 } 1216 1217 int siw_run_sq(void *data) 1218 { 1219 const int nr_cpu = (unsigned int)(long)data; 1220 struct llist_node *active; 1221 struct siw_qp *qp; 1222 struct tx_task_t *tx_task = &per_cpu(siw_tx_task_g, nr_cpu); 1223 1224 init_llist_head(&tx_task->active); 1225 init_waitqueue_head(&tx_task->waiting); 1226 1227 while (1) { 1228 struct llist_node *fifo_list = NULL; 1229 1230 wait_event_interruptible(tx_task->waiting, 1231 !llist_empty(&tx_task->active) || 1232 kthread_should_stop()); 1233 1234 if (kthread_should_stop()) 1235 break; 1236 1237 active = llist_del_all(&tx_task->active); 1238 /* 1239 * llist_del_all returns a list with newest entry first. 1240 * Re-order list for fairness among QP's. 1241 */ 1242 while (active) { 1243 struct llist_node *tmp = active; 1244 1245 active = llist_next(active); 1246 tmp->next = fifo_list; 1247 fifo_list = tmp; 1248 } 1249 while (fifo_list) { 1250 qp = container_of(fifo_list, struct siw_qp, tx_list); 1251 fifo_list = llist_next(fifo_list); 1252 qp->tx_list.next = NULL; 1253 1254 siw_sq_resume(qp); 1255 } 1256 } 1257 active = llist_del_all(&tx_task->active); 1258 if (active) { 1259 llist_for_each_entry(qp, active, tx_list) { 1260 qp->tx_list.next = NULL; 1261 siw_sq_resume(qp); 1262 } 1263 } 1264 return 0; 1265 } 1266 1267 int siw_sq_start(struct siw_qp *qp) 1268 { 1269 if (tx_wqe(qp)->wr_status == SIW_WR_IDLE) 1270 return 0; 1271 1272 if (unlikely(!cpu_online(qp->tx_cpu))) { 1273 siw_put_tx_cpu(qp->tx_cpu); 1274 qp->tx_cpu = siw_get_tx_cpu(qp->sdev); 1275 if (qp->tx_cpu < 0) { 1276 pr_warn("siw: no tx cpu available\n"); 1277 1278 return -EIO; 1279 } 1280 } 1281 siw_qp_get(qp); 1282 1283 llist_add(&qp->tx_list, &per_cpu(siw_tx_task_g, qp->tx_cpu).active); 1284 1285 wake_up(&per_cpu(siw_tx_task_g, qp->tx_cpu).waiting); 1286 1287 return 0; 1288 } 1289