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_SENDPAGE_NOTLAST | 329 MSG_SPLICE_PAGES), 330 }; 331 struct sock *sk = s->sk; 332 int i = 0, rv = 0, sent = 0; 333 334 while (size) { 335 size_t bytes = min_t(size_t, PAGE_SIZE - offset, size); 336 337 if (size + offset <= PAGE_SIZE) 338 msg.msg_flags &= ~MSG_SENDPAGE_NOTLAST; 339 340 tcp_rate_check_app_limited(sk); 341 bvec_set_page(&bvec, page[i], bytes, offset); 342 iov_iter_bvec(&msg.msg_iter, ITER_SOURCE, &bvec, 1, size); 343 344 try_page_again: 345 lock_sock(sk); 346 rv = tcp_sendmsg_locked(sk, &msg, size); 347 release_sock(sk); 348 349 if (rv > 0) { 350 size -= rv; 351 sent += rv; 352 if (rv != bytes) { 353 offset += rv; 354 bytes -= rv; 355 goto try_page_again; 356 } 357 offset = 0; 358 } else { 359 if (rv == -EAGAIN || rv == 0) 360 break; 361 return rv; 362 } 363 i++; 364 } 365 return sent; 366 } 367 368 /* 369 * siw_0copy_tx() 370 * 371 * Pushes list of pages to TCP socket. If pages from multiple 372 * SGE's, all referenced pages of each SGE are pushed in one 373 * shot. 374 */ 375 static int siw_0copy_tx(struct socket *s, struct page **page, 376 struct siw_sge *sge, unsigned int offset, 377 unsigned int size) 378 { 379 int i = 0, sent = 0, rv; 380 int sge_bytes = min(sge->length - offset, size); 381 382 offset = (sge->laddr + offset) & ~PAGE_MASK; 383 384 while (sent != size) { 385 rv = siw_tcp_sendpages(s, &page[i], offset, sge_bytes); 386 if (rv >= 0) { 387 sent += rv; 388 if (size == sent || sge_bytes > rv) 389 break; 390 391 i += PAGE_ALIGN(sge_bytes + offset) >> PAGE_SHIFT; 392 sge++; 393 sge_bytes = min(sge->length, size - sent); 394 offset = sge->laddr & ~PAGE_MASK; 395 } else { 396 sent = rv; 397 break; 398 } 399 } 400 return sent; 401 } 402 403 #define MAX_TRAILER (MPA_CRC_SIZE + 4) 404 405 static void siw_unmap_pages(struct kvec *iov, unsigned long kmap_mask, int len) 406 { 407 int i; 408 409 /* 410 * Work backwards through the array to honor the kmap_local_page() 411 * ordering requirements. 412 */ 413 for (i = (len-1); i >= 0; i--) { 414 if (kmap_mask & BIT(i)) { 415 unsigned long addr = (unsigned long)iov[i].iov_base; 416 417 kunmap_local((void *)(addr & PAGE_MASK)); 418 } 419 } 420 } 421 422 /* 423 * siw_tx_hdt() tries to push a complete packet to TCP where all 424 * packet fragments are referenced by the elements of one iovec. 425 * For the data portion, each involved page must be referenced by 426 * one extra element. All sge's data can be non-aligned to page 427 * boundaries. Two more elements are referencing iWARP header 428 * and trailer: 429 * MAX_ARRAY = 64KB/PAGE_SIZE + 1 + (2 * (SIW_MAX_SGE - 1) + HDR + TRL 430 */ 431 #define MAX_ARRAY ((0xffff / PAGE_SIZE) + 1 + (2 * (SIW_MAX_SGE - 1) + 2)) 432 433 /* 434 * Write out iov referencing hdr, data and trailer of current FPDU. 435 * Update transmit state dependent on write return status 436 */ 437 static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s) 438 { 439 struct siw_wqe *wqe = &c_tx->wqe_active; 440 struct siw_sge *sge = &wqe->sqe.sge[c_tx->sge_idx]; 441 struct kvec iov[MAX_ARRAY]; 442 struct page *page_array[MAX_ARRAY]; 443 struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_EOR }; 444 445 int seg = 0, do_crc = c_tx->do_crc, is_kva = 0, rv; 446 unsigned int data_len = c_tx->bytes_unsent, hdr_len = 0, trl_len = 0, 447 sge_off = c_tx->sge_off, sge_idx = c_tx->sge_idx, 448 pbl_idx = c_tx->pbl_idx; 449 unsigned long kmap_mask = 0L; 450 451 if (c_tx->state == SIW_SEND_HDR) { 452 if (c_tx->use_sendpage) { 453 rv = siw_tx_ctrl(c_tx, s, MSG_DONTWAIT | MSG_MORE); 454 if (rv) 455 goto done; 456 457 c_tx->state = SIW_SEND_DATA; 458 } else { 459 iov[0].iov_base = 460 (char *)&c_tx->pkt.ctrl + c_tx->ctrl_sent; 461 iov[0].iov_len = hdr_len = 462 c_tx->ctrl_len - c_tx->ctrl_sent; 463 seg = 1; 464 } 465 } 466 467 wqe->processed += data_len; 468 469 while (data_len) { /* walk the list of SGE's */ 470 unsigned int sge_len = min(sge->length - sge_off, data_len); 471 unsigned int fp_off = (sge->laddr + sge_off) & ~PAGE_MASK; 472 struct siw_mem *mem; 473 474 if (!(tx_flags(wqe) & SIW_WQE_INLINE)) { 475 mem = wqe->mem[sge_idx]; 476 is_kva = mem->mem_obj == NULL ? 1 : 0; 477 } else { 478 is_kva = 1; 479 } 480 if (is_kva && !c_tx->use_sendpage) { 481 /* 482 * tx from kernel virtual address: either inline data 483 * or memory region with assigned kernel buffer 484 */ 485 iov[seg].iov_base = 486 ib_virt_dma_to_ptr(sge->laddr + sge_off); 487 iov[seg].iov_len = sge_len; 488 489 if (do_crc) 490 crypto_shash_update(c_tx->mpa_crc_hd, 491 iov[seg].iov_base, 492 sge_len); 493 sge_off += sge_len; 494 data_len -= sge_len; 495 seg++; 496 goto sge_done; 497 } 498 499 while (sge_len) { 500 size_t plen = min((int)PAGE_SIZE - fp_off, sge_len); 501 void *kaddr; 502 503 if (!is_kva) { 504 struct page *p; 505 506 if (mem->is_pbl) 507 p = siw_get_pblpage( 508 mem, sge->laddr + sge_off, 509 &pbl_idx); 510 else 511 p = siw_get_upage(mem->umem, 512 sge->laddr + sge_off); 513 if (unlikely(!p)) { 514 siw_unmap_pages(iov, kmap_mask, seg); 515 wqe->processed -= c_tx->bytes_unsent; 516 rv = -EFAULT; 517 goto done_crc; 518 } 519 page_array[seg] = p; 520 521 if (!c_tx->use_sendpage) { 522 void *kaddr = kmap_local_page(p); 523 524 /* Remember for later kunmap() */ 525 kmap_mask |= BIT(seg); 526 iov[seg].iov_base = kaddr + fp_off; 527 iov[seg].iov_len = plen; 528 529 if (do_crc) 530 crypto_shash_update( 531 c_tx->mpa_crc_hd, 532 iov[seg].iov_base, 533 plen); 534 } else if (do_crc) { 535 kaddr = kmap_local_page(p); 536 crypto_shash_update(c_tx->mpa_crc_hd, 537 kaddr + fp_off, 538 plen); 539 kunmap_local(kaddr); 540 } 541 } else { 542 /* 543 * Cast to an uintptr_t to preserve all 64 bits 544 * in sge->laddr. 545 */ 546 u64 va = sge->laddr + sge_off; 547 548 page_array[seg] = ib_virt_dma_to_page(va); 549 if (do_crc) 550 crypto_shash_update( 551 c_tx->mpa_crc_hd, 552 ib_virt_dma_to_ptr(va), 553 plen); 554 } 555 556 sge_len -= plen; 557 sge_off += plen; 558 data_len -= plen; 559 fp_off = 0; 560 561 if (++seg >= (int)MAX_ARRAY) { 562 siw_dbg_qp(tx_qp(c_tx), "to many fragments\n"); 563 siw_unmap_pages(iov, kmap_mask, seg-1); 564 wqe->processed -= c_tx->bytes_unsent; 565 rv = -EMSGSIZE; 566 goto done_crc; 567 } 568 } 569 sge_done: 570 /* Update SGE variables at end of SGE */ 571 if (sge_off == sge->length && 572 (data_len != 0 || wqe->processed < wqe->bytes)) { 573 sge_idx++; 574 sge++; 575 sge_off = 0; 576 } 577 } 578 /* trailer */ 579 if (likely(c_tx->state != SIW_SEND_TRAILER)) { 580 iov[seg].iov_base = &c_tx->trailer.pad[4 - c_tx->pad]; 581 iov[seg].iov_len = trl_len = MAX_TRAILER - (4 - c_tx->pad); 582 } else { 583 iov[seg].iov_base = &c_tx->trailer.pad[c_tx->ctrl_sent]; 584 iov[seg].iov_len = trl_len = MAX_TRAILER - c_tx->ctrl_sent; 585 } 586 587 if (c_tx->pad) { 588 *(u32 *)c_tx->trailer.pad = 0; 589 if (do_crc) 590 crypto_shash_update(c_tx->mpa_crc_hd, 591 (u8 *)&c_tx->trailer.crc - c_tx->pad, 592 c_tx->pad); 593 } 594 if (!c_tx->mpa_crc_hd) 595 c_tx->trailer.crc = 0; 596 else if (do_crc) 597 crypto_shash_final(c_tx->mpa_crc_hd, (u8 *)&c_tx->trailer.crc); 598 599 data_len = c_tx->bytes_unsent; 600 601 if (c_tx->use_sendpage) { 602 rv = siw_0copy_tx(s, page_array, &wqe->sqe.sge[c_tx->sge_idx], 603 c_tx->sge_off, data_len); 604 if (rv == data_len) { 605 rv = kernel_sendmsg(s, &msg, &iov[seg], 1, trl_len); 606 if (rv > 0) 607 rv += data_len; 608 else 609 rv = data_len; 610 } 611 } else { 612 rv = kernel_sendmsg(s, &msg, iov, seg + 1, 613 hdr_len + data_len + trl_len); 614 siw_unmap_pages(iov, kmap_mask, seg); 615 } 616 if (rv < (int)hdr_len) { 617 /* Not even complete hdr pushed or negative rv */ 618 wqe->processed -= data_len; 619 if (rv >= 0) { 620 c_tx->ctrl_sent += rv; 621 rv = -EAGAIN; 622 } 623 goto done_crc; 624 } 625 rv -= hdr_len; 626 627 if (rv >= (int)data_len) { 628 /* all user data pushed to TCP or no data to push */ 629 if (data_len > 0 && wqe->processed < wqe->bytes) { 630 /* Save the current state for next tx */ 631 c_tx->sge_idx = sge_idx; 632 c_tx->sge_off = sge_off; 633 c_tx->pbl_idx = pbl_idx; 634 } 635 rv -= data_len; 636 637 if (rv == trl_len) /* all pushed */ 638 rv = 0; 639 else { 640 c_tx->state = SIW_SEND_TRAILER; 641 c_tx->ctrl_len = MAX_TRAILER; 642 c_tx->ctrl_sent = rv + 4 - c_tx->pad; 643 c_tx->bytes_unsent = 0; 644 rv = -EAGAIN; 645 } 646 647 } else if (data_len > 0) { 648 /* Maybe some user data pushed to TCP */ 649 c_tx->state = SIW_SEND_DATA; 650 wqe->processed -= data_len - rv; 651 652 if (rv) { 653 /* 654 * Some bytes out. Recompute tx state based 655 * on old state and bytes pushed 656 */ 657 unsigned int sge_unsent; 658 659 c_tx->bytes_unsent -= rv; 660 sge = &wqe->sqe.sge[c_tx->sge_idx]; 661 sge_unsent = sge->length - c_tx->sge_off; 662 663 while (sge_unsent <= rv) { 664 rv -= sge_unsent; 665 c_tx->sge_idx++; 666 c_tx->sge_off = 0; 667 sge++; 668 sge_unsent = sge->length; 669 } 670 c_tx->sge_off += rv; 671 } 672 rv = -EAGAIN; 673 } 674 done_crc: 675 c_tx->do_crc = 0; 676 done: 677 return rv; 678 } 679 680 static void siw_update_tcpseg(struct siw_iwarp_tx *c_tx, 681 struct socket *s) 682 { 683 struct tcp_sock *tp = tcp_sk(s->sk); 684 685 if (tp->gso_segs) { 686 if (c_tx->gso_seg_limit == 0) 687 c_tx->tcp_seglen = tp->mss_cache * tp->gso_segs; 688 else 689 c_tx->tcp_seglen = 690 tp->mss_cache * 691 min_t(u16, c_tx->gso_seg_limit, tp->gso_segs); 692 } else { 693 c_tx->tcp_seglen = tp->mss_cache; 694 } 695 /* Loopback may give odd numbers */ 696 c_tx->tcp_seglen &= 0xfffffff8; 697 } 698 699 /* 700 * siw_prepare_fpdu() 701 * 702 * Prepares transmit context to send out one FPDU if FPDU will contain 703 * user data and user data are not immediate data. 704 * Computes maximum FPDU length to fill up TCP MSS if possible. 705 * 706 * @qp: QP from which to transmit 707 * @wqe: Current WQE causing transmission 708 * 709 * TODO: Take into account real available sendspace on socket 710 * to avoid header misalignment due to send pausing within 711 * fpdu transmission 712 */ 713 static void siw_prepare_fpdu(struct siw_qp *qp, struct siw_wqe *wqe) 714 { 715 struct siw_iwarp_tx *c_tx = &qp->tx_ctx; 716 int data_len; 717 718 c_tx->ctrl_len = 719 iwarp_pktinfo[__rdmap_get_opcode(&c_tx->pkt.ctrl)].hdr_len; 720 c_tx->ctrl_sent = 0; 721 722 /* 723 * Update target buffer offset if any 724 */ 725 if (!(c_tx->pkt.ctrl.ddp_rdmap_ctrl & DDP_FLAG_TAGGED)) 726 /* Untagged message */ 727 c_tx->pkt.c_untagged.ddp_mo = cpu_to_be32(wqe->processed); 728 else /* Tagged message */ 729 c_tx->pkt.c_tagged.ddp_to = 730 cpu_to_be64(wqe->sqe.raddr + wqe->processed); 731 732 data_len = wqe->bytes - wqe->processed; 733 if (data_len + c_tx->ctrl_len + MPA_CRC_SIZE > c_tx->tcp_seglen) { 734 /* Trim DDP payload to fit into current TCP segment */ 735 data_len = c_tx->tcp_seglen - (c_tx->ctrl_len + MPA_CRC_SIZE); 736 c_tx->pkt.ctrl.ddp_rdmap_ctrl &= ~DDP_FLAG_LAST; 737 c_tx->pad = 0; 738 } else { 739 c_tx->pkt.ctrl.ddp_rdmap_ctrl |= DDP_FLAG_LAST; 740 c_tx->pad = -data_len & 0x3; 741 } 742 c_tx->bytes_unsent = data_len; 743 744 c_tx->pkt.ctrl.mpa_len = 745 htons(c_tx->ctrl_len + data_len - MPA_HDR_SIZE); 746 747 /* 748 * Init MPA CRC computation 749 */ 750 if (c_tx->mpa_crc_hd) { 751 crypto_shash_init(c_tx->mpa_crc_hd); 752 crypto_shash_update(c_tx->mpa_crc_hd, (u8 *)&c_tx->pkt, 753 c_tx->ctrl_len); 754 c_tx->do_crc = 1; 755 } 756 } 757 758 /* 759 * siw_check_sgl_tx() 760 * 761 * Check permissions for a list of SGE's (SGL). 762 * A successful check will have all memory referenced 763 * for transmission resolved and assigned to the WQE. 764 * 765 * @pd: Protection Domain SGL should belong to 766 * @wqe: WQE to be checked 767 * @perms: requested access permissions 768 * 769 */ 770 771 static int siw_check_sgl_tx(struct ib_pd *pd, struct siw_wqe *wqe, 772 enum ib_access_flags perms) 773 { 774 struct siw_sge *sge = &wqe->sqe.sge[0]; 775 int i, len, num_sge = wqe->sqe.num_sge; 776 777 if (unlikely(num_sge > SIW_MAX_SGE)) 778 return -EINVAL; 779 780 for (i = 0, len = 0; num_sge; num_sge--, i++, sge++) { 781 /* 782 * rdma verbs: do not check stag for a zero length sge 783 */ 784 if (sge->length) { 785 int rv = siw_check_sge(pd, sge, &wqe->mem[i], perms, 0, 786 sge->length); 787 788 if (unlikely(rv != E_ACCESS_OK)) 789 return rv; 790 } 791 len += sge->length; 792 } 793 return len; 794 } 795 796 /* 797 * siw_qp_sq_proc_tx() 798 * 799 * Process one WQE which needs transmission on the wire. 800 */ 801 static int siw_qp_sq_proc_tx(struct siw_qp *qp, struct siw_wqe *wqe) 802 { 803 struct siw_iwarp_tx *c_tx = &qp->tx_ctx; 804 struct socket *s = qp->attrs.sk; 805 int rv = 0, burst_len = qp->tx_ctx.burst; 806 enum rdmap_ecode ecode = RDMAP_ECODE_CATASTROPHIC_STREAM; 807 808 if (unlikely(wqe->wr_status == SIW_WR_IDLE)) 809 return 0; 810 811 if (!burst_len) 812 burst_len = SQ_USER_MAXBURST; 813 814 if (wqe->wr_status == SIW_WR_QUEUED) { 815 if (!(wqe->sqe.flags & SIW_WQE_INLINE)) { 816 if (tx_type(wqe) == SIW_OP_READ_RESPONSE) 817 wqe->sqe.num_sge = 1; 818 819 if (tx_type(wqe) != SIW_OP_READ && 820 tx_type(wqe) != SIW_OP_READ_LOCAL_INV) { 821 /* 822 * Reference memory to be tx'd w/o checking 823 * access for LOCAL_READ permission, since 824 * not defined in RDMA core. 825 */ 826 rv = siw_check_sgl_tx(qp->pd, wqe, 0); 827 if (rv < 0) { 828 if (tx_type(wqe) == 829 SIW_OP_READ_RESPONSE) 830 ecode = siw_rdmap_error(-rv); 831 rv = -EINVAL; 832 goto tx_error; 833 } 834 wqe->bytes = rv; 835 } else { 836 wqe->bytes = 0; 837 } 838 } else { 839 wqe->bytes = wqe->sqe.sge[0].length; 840 if (!rdma_is_kernel_res(&qp->base_qp.res)) { 841 if (wqe->bytes > SIW_MAX_INLINE) { 842 rv = -EINVAL; 843 goto tx_error; 844 } 845 wqe->sqe.sge[0].laddr = 846 (u64)(uintptr_t)&wqe->sqe.sge[1]; 847 } 848 } 849 wqe->wr_status = SIW_WR_INPROGRESS; 850 wqe->processed = 0; 851 852 siw_update_tcpseg(c_tx, s); 853 854 rv = siw_qp_prepare_tx(c_tx); 855 if (rv == PKT_FRAGMENTED) { 856 c_tx->state = SIW_SEND_HDR; 857 siw_prepare_fpdu(qp, wqe); 858 } else if (rv == PKT_COMPLETE) { 859 c_tx->state = SIW_SEND_SHORT_FPDU; 860 } else { 861 goto tx_error; 862 } 863 } 864 865 next_segment: 866 siw_dbg_qp(qp, "wr type %d, state %d, data %u, sent %u, id %llx\n", 867 tx_type(wqe), wqe->wr_status, wqe->bytes, wqe->processed, 868 wqe->sqe.id); 869 870 if (--burst_len == 0) { 871 rv = -EINPROGRESS; 872 goto tx_done; 873 } 874 if (c_tx->state == SIW_SEND_SHORT_FPDU) { 875 enum siw_opcode tx_type = tx_type(wqe); 876 unsigned int msg_flags; 877 878 if (siw_sq_empty(qp) || !siw_tcp_nagle || burst_len == 1) 879 /* 880 * End current TCP segment, if SQ runs empty, 881 * or siw_tcp_nagle is not set, or we bail out 882 * soon due to no burst credit left. 883 */ 884 msg_flags = MSG_DONTWAIT; 885 else 886 msg_flags = MSG_DONTWAIT | MSG_MORE; 887 888 rv = siw_tx_ctrl(c_tx, s, msg_flags); 889 890 if (!rv && tx_type != SIW_OP_READ && 891 tx_type != SIW_OP_READ_LOCAL_INV) 892 wqe->processed = wqe->bytes; 893 894 goto tx_done; 895 896 } else { 897 rv = siw_tx_hdt(c_tx, s); 898 } 899 if (!rv) { 900 /* 901 * One segment sent. Processing completed if last 902 * segment, Do next segment otherwise. 903 */ 904 if (unlikely(c_tx->tx_suspend)) { 905 /* 906 * Verbs, 6.4.: Try stopping sending after a full 907 * DDP segment if the connection goes down 908 * (== peer halfclose) 909 */ 910 rv = -ECONNABORTED; 911 goto tx_done; 912 } 913 if (c_tx->pkt.ctrl.ddp_rdmap_ctrl & DDP_FLAG_LAST) { 914 siw_dbg_qp(qp, "WQE completed\n"); 915 goto tx_done; 916 } 917 c_tx->state = SIW_SEND_HDR; 918 919 siw_update_tcpseg(c_tx, s); 920 921 siw_prepare_fpdu(qp, wqe); 922 goto next_segment; 923 } 924 tx_done: 925 qp->tx_ctx.burst = burst_len; 926 return rv; 927 928 tx_error: 929 if (ecode != RDMAP_ECODE_CATASTROPHIC_STREAM) 930 siw_init_terminate(qp, TERM_ERROR_LAYER_RDMAP, 931 RDMAP_ETYPE_REMOTE_PROTECTION, ecode, 1); 932 else 933 siw_init_terminate(qp, TERM_ERROR_LAYER_RDMAP, 934 RDMAP_ETYPE_CATASTROPHIC, 935 RDMAP_ECODE_UNSPECIFIED, 1); 936 return rv; 937 } 938 939 static int siw_fastreg_mr(struct ib_pd *pd, struct siw_sqe *sqe) 940 { 941 struct ib_mr *base_mr = (struct ib_mr *)(uintptr_t)sqe->base_mr; 942 struct siw_device *sdev = to_siw_dev(pd->device); 943 struct siw_mem *mem; 944 int rv = 0; 945 946 siw_dbg_pd(pd, "STag 0x%08x\n", sqe->rkey); 947 948 if (unlikely(!base_mr)) { 949 pr_warn("siw: fastreg: STag 0x%08x unknown\n", sqe->rkey); 950 return -EINVAL; 951 } 952 953 if (unlikely(base_mr->rkey >> 8 != sqe->rkey >> 8)) { 954 pr_warn("siw: fastreg: STag 0x%08x: bad MR\n", sqe->rkey); 955 return -EINVAL; 956 } 957 958 mem = siw_mem_id2obj(sdev, sqe->rkey >> 8); 959 if (unlikely(!mem)) { 960 pr_warn("siw: fastreg: STag 0x%08x unknown\n", sqe->rkey); 961 return -EINVAL; 962 } 963 964 if (unlikely(mem->pd != pd)) { 965 pr_warn("siw: fastreg: PD mismatch\n"); 966 rv = -EINVAL; 967 goto out; 968 } 969 if (unlikely(mem->stag_valid)) { 970 pr_warn("siw: fastreg: STag 0x%08x already valid\n", sqe->rkey); 971 rv = -EINVAL; 972 goto out; 973 } 974 /* Refresh STag since user may have changed key part */ 975 mem->stag = sqe->rkey; 976 mem->perms = sqe->access; 977 978 siw_dbg_mem(mem, "STag 0x%08x now valid\n", sqe->rkey); 979 mem->va = base_mr->iova; 980 mem->stag_valid = 1; 981 out: 982 siw_mem_put(mem); 983 return rv; 984 } 985 986 static int siw_qp_sq_proc_local(struct siw_qp *qp, struct siw_wqe *wqe) 987 { 988 int rv; 989 990 switch (tx_type(wqe)) { 991 case SIW_OP_REG_MR: 992 rv = siw_fastreg_mr(qp->pd, &wqe->sqe); 993 break; 994 995 case SIW_OP_INVAL_STAG: 996 rv = siw_invalidate_stag(qp->pd, wqe->sqe.rkey); 997 break; 998 999 default: 1000 rv = -EINVAL; 1001 } 1002 return rv; 1003 } 1004 1005 /* 1006 * siw_qp_sq_process() 1007 * 1008 * Core TX path routine for RDMAP/DDP/MPA using a TCP kernel socket. 1009 * Sends RDMAP payload for the current SQ WR @wqe of @qp in one or more 1010 * MPA FPDUs, each containing a DDP segment. 1011 * 1012 * SQ processing may occur in user context as a result of posting 1013 * new WQE's or from siw_sq_work_handler() context. Processing in 1014 * user context is limited to non-kernel verbs users. 1015 * 1016 * SQ processing may get paused anytime, possibly in the middle of a WR 1017 * or FPDU, if insufficient send space is available. SQ processing 1018 * gets resumed from siw_sq_work_handler(), if send space becomes 1019 * available again. 1020 * 1021 * Must be called with the QP state read-locked. 1022 * 1023 * Note: 1024 * An outbound RREQ can be satisfied by the corresponding RRESP 1025 * _before_ it gets assigned to the ORQ. This happens regularly 1026 * in RDMA READ via loopback case. Since both outbound RREQ and 1027 * inbound RRESP can be handled by the same CPU, locking the ORQ 1028 * is dead-lock prone and thus not an option. With that, the 1029 * RREQ gets assigned to the ORQ _before_ being sent - see 1030 * siw_activate_tx() - and pulled back in case of send failure. 1031 */ 1032 int siw_qp_sq_process(struct siw_qp *qp) 1033 { 1034 struct siw_wqe *wqe = tx_wqe(qp); 1035 enum siw_opcode tx_type; 1036 unsigned long flags; 1037 int rv = 0; 1038 1039 siw_dbg_qp(qp, "enter for type %d\n", tx_type(wqe)); 1040 1041 next_wqe: 1042 /* 1043 * Stop QP processing if SQ state changed 1044 */ 1045 if (unlikely(qp->tx_ctx.tx_suspend)) { 1046 siw_dbg_qp(qp, "tx suspended\n"); 1047 goto done; 1048 } 1049 tx_type = tx_type(wqe); 1050 1051 if (tx_type <= SIW_OP_READ_RESPONSE) 1052 rv = siw_qp_sq_proc_tx(qp, wqe); 1053 else 1054 rv = siw_qp_sq_proc_local(qp, wqe); 1055 1056 if (!rv) { 1057 /* 1058 * WQE processing done 1059 */ 1060 switch (tx_type) { 1061 case SIW_OP_SEND: 1062 case SIW_OP_SEND_REMOTE_INV: 1063 case SIW_OP_WRITE: 1064 siw_wqe_put_mem(wqe, tx_type); 1065 fallthrough; 1066 1067 case SIW_OP_INVAL_STAG: 1068 case SIW_OP_REG_MR: 1069 if (tx_flags(wqe) & SIW_WQE_SIGNALLED) 1070 siw_sqe_complete(qp, &wqe->sqe, wqe->bytes, 1071 SIW_WC_SUCCESS); 1072 break; 1073 1074 case SIW_OP_READ: 1075 case SIW_OP_READ_LOCAL_INV: 1076 /* 1077 * already enqueued to ORQ queue 1078 */ 1079 break; 1080 1081 case SIW_OP_READ_RESPONSE: 1082 siw_wqe_put_mem(wqe, tx_type); 1083 break; 1084 1085 default: 1086 WARN(1, "undefined WQE type %d\n", tx_type); 1087 rv = -EINVAL; 1088 goto done; 1089 } 1090 1091 spin_lock_irqsave(&qp->sq_lock, flags); 1092 wqe->wr_status = SIW_WR_IDLE; 1093 rv = siw_activate_tx(qp); 1094 spin_unlock_irqrestore(&qp->sq_lock, flags); 1095 1096 if (rv <= 0) 1097 goto done; 1098 1099 goto next_wqe; 1100 1101 } else if (rv == -EAGAIN) { 1102 siw_dbg_qp(qp, "sq paused: hd/tr %d of %d, data %d\n", 1103 qp->tx_ctx.ctrl_sent, qp->tx_ctx.ctrl_len, 1104 qp->tx_ctx.bytes_unsent); 1105 rv = 0; 1106 goto done; 1107 } else if (rv == -EINPROGRESS) { 1108 rv = siw_sq_start(qp); 1109 goto done; 1110 } else { 1111 /* 1112 * WQE processing failed. 1113 * Verbs 8.3.2: 1114 * o It turns any WQE into a signalled WQE. 1115 * o Local catastrophic error must be surfaced 1116 * o QP must be moved into Terminate state: done by code 1117 * doing socket state change processing 1118 * 1119 * o TODO: Termination message must be sent. 1120 * o TODO: Implement more precise work completion errors, 1121 * see enum ib_wc_status in ib_verbs.h 1122 */ 1123 siw_dbg_qp(qp, "wqe type %d processing failed: %d\n", 1124 tx_type(wqe), rv); 1125 1126 spin_lock_irqsave(&qp->sq_lock, flags); 1127 /* 1128 * RREQ may have already been completed by inbound RRESP! 1129 */ 1130 if ((tx_type == SIW_OP_READ || 1131 tx_type == SIW_OP_READ_LOCAL_INV) && qp->attrs.orq_size) { 1132 /* Cleanup pending entry in ORQ */ 1133 qp->orq_put--; 1134 qp->orq[qp->orq_put % qp->attrs.orq_size].flags = 0; 1135 } 1136 spin_unlock_irqrestore(&qp->sq_lock, flags); 1137 /* 1138 * immediately suspends further TX processing 1139 */ 1140 if (!qp->tx_ctx.tx_suspend) 1141 siw_qp_cm_drop(qp, 0); 1142 1143 switch (tx_type) { 1144 case SIW_OP_SEND: 1145 case SIW_OP_SEND_REMOTE_INV: 1146 case SIW_OP_SEND_WITH_IMM: 1147 case SIW_OP_WRITE: 1148 case SIW_OP_READ: 1149 case SIW_OP_READ_LOCAL_INV: 1150 siw_wqe_put_mem(wqe, tx_type); 1151 fallthrough; 1152 1153 case SIW_OP_INVAL_STAG: 1154 case SIW_OP_REG_MR: 1155 siw_sqe_complete(qp, &wqe->sqe, wqe->bytes, 1156 SIW_WC_LOC_QP_OP_ERR); 1157 1158 siw_qp_event(qp, IB_EVENT_QP_FATAL); 1159 1160 break; 1161 1162 case SIW_OP_READ_RESPONSE: 1163 siw_dbg_qp(qp, "proc. read.response failed: %d\n", rv); 1164 1165 siw_qp_event(qp, IB_EVENT_QP_REQ_ERR); 1166 1167 siw_wqe_put_mem(wqe, SIW_OP_READ_RESPONSE); 1168 1169 break; 1170 1171 default: 1172 WARN(1, "undefined WQE type %d\n", tx_type); 1173 rv = -EINVAL; 1174 } 1175 wqe->wr_status = SIW_WR_IDLE; 1176 } 1177 done: 1178 return rv; 1179 } 1180 1181 static void siw_sq_resume(struct siw_qp *qp) 1182 { 1183 if (down_read_trylock(&qp->state_lock)) { 1184 if (likely(qp->attrs.state == SIW_QP_STATE_RTS && 1185 !qp->tx_ctx.tx_suspend)) { 1186 int rv = siw_qp_sq_process(qp); 1187 1188 up_read(&qp->state_lock); 1189 1190 if (unlikely(rv < 0)) { 1191 siw_dbg_qp(qp, "SQ task failed: err %d\n", rv); 1192 1193 if (!qp->tx_ctx.tx_suspend) 1194 siw_qp_cm_drop(qp, 0); 1195 } 1196 } else { 1197 up_read(&qp->state_lock); 1198 } 1199 } else { 1200 siw_dbg_qp(qp, "Resume SQ while QP locked\n"); 1201 } 1202 siw_qp_put(qp); 1203 } 1204 1205 struct tx_task_t { 1206 struct llist_head active; 1207 wait_queue_head_t waiting; 1208 }; 1209 1210 static DEFINE_PER_CPU(struct tx_task_t, siw_tx_task_g); 1211 1212 void siw_stop_tx_thread(int nr_cpu) 1213 { 1214 kthread_stop(siw_tx_thread[nr_cpu]); 1215 wake_up(&per_cpu(siw_tx_task_g, nr_cpu).waiting); 1216 } 1217 1218 int siw_run_sq(void *data) 1219 { 1220 const int nr_cpu = (unsigned int)(long)data; 1221 struct llist_node *active; 1222 struct siw_qp *qp; 1223 struct tx_task_t *tx_task = &per_cpu(siw_tx_task_g, nr_cpu); 1224 1225 init_llist_head(&tx_task->active); 1226 init_waitqueue_head(&tx_task->waiting); 1227 1228 while (1) { 1229 struct llist_node *fifo_list = NULL; 1230 1231 wait_event_interruptible(tx_task->waiting, 1232 !llist_empty(&tx_task->active) || 1233 kthread_should_stop()); 1234 1235 if (kthread_should_stop()) 1236 break; 1237 1238 active = llist_del_all(&tx_task->active); 1239 /* 1240 * llist_del_all returns a list with newest entry first. 1241 * Re-order list for fairness among QP's. 1242 */ 1243 while (active) { 1244 struct llist_node *tmp = active; 1245 1246 active = llist_next(active); 1247 tmp->next = fifo_list; 1248 fifo_list = tmp; 1249 } 1250 while (fifo_list) { 1251 qp = container_of(fifo_list, struct siw_qp, tx_list); 1252 fifo_list = llist_next(fifo_list); 1253 qp->tx_list.next = NULL; 1254 1255 siw_sq_resume(qp); 1256 } 1257 } 1258 active = llist_del_all(&tx_task->active); 1259 if (active) { 1260 llist_for_each_entry(qp, active, tx_list) { 1261 qp->tx_list.next = NULL; 1262 siw_sq_resume(qp); 1263 } 1264 } 1265 return 0; 1266 } 1267 1268 int siw_sq_start(struct siw_qp *qp) 1269 { 1270 if (tx_wqe(qp)->wr_status == SIW_WR_IDLE) 1271 return 0; 1272 1273 if (unlikely(!cpu_online(qp->tx_cpu))) { 1274 siw_put_tx_cpu(qp->tx_cpu); 1275 qp->tx_cpu = siw_get_tx_cpu(qp->sdev); 1276 if (qp->tx_cpu < 0) { 1277 pr_warn("siw: no tx cpu available\n"); 1278 1279 return -EIO; 1280 } 1281 } 1282 siw_qp_get(qp); 1283 1284 llist_add(&qp->tx_list, &per_cpu(siw_tx_task_g, qp->tx_cpu).active); 1285 1286 wake_up(&per_cpu(siw_tx_task_g, qp->tx_cpu).waiting); 1287 1288 return 0; 1289 } 1290