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