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 do_tcp_sendpages. 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 sock *sk = s->sk; 327 int i = 0, rv = 0, sent = 0, 328 flags = MSG_MORE | MSG_DONTWAIT | MSG_SENDPAGE_NOTLAST; 329 330 while (size) { 331 size_t bytes = min_t(size_t, PAGE_SIZE - offset, size); 332 333 if (size + offset <= PAGE_SIZE) 334 flags = MSG_MORE | MSG_DONTWAIT; 335 336 tcp_rate_check_app_limited(sk); 337 try_page_again: 338 lock_sock(sk); 339 rv = do_tcp_sendpages(sk, page[i], offset, bytes, flags); 340 release_sock(sk); 341 342 if (rv > 0) { 343 size -= rv; 344 sent += rv; 345 if (rv != bytes) { 346 offset += rv; 347 bytes -= rv; 348 goto try_page_again; 349 } 350 offset = 0; 351 } else { 352 if (rv == -EAGAIN || rv == 0) 353 break; 354 return rv; 355 } 356 i++; 357 } 358 return sent; 359 } 360 361 /* 362 * siw_0copy_tx() 363 * 364 * Pushes list of pages to TCP socket. If pages from multiple 365 * SGE's, all referenced pages of each SGE are pushed in one 366 * shot. 367 */ 368 static int siw_0copy_tx(struct socket *s, struct page **page, 369 struct siw_sge *sge, unsigned int offset, 370 unsigned int size) 371 { 372 int i = 0, sent = 0, rv; 373 int sge_bytes = min(sge->length - offset, size); 374 375 offset = (sge->laddr + offset) & ~PAGE_MASK; 376 377 while (sent != size) { 378 rv = siw_tcp_sendpages(s, &page[i], offset, sge_bytes); 379 if (rv >= 0) { 380 sent += rv; 381 if (size == sent || sge_bytes > rv) 382 break; 383 384 i += PAGE_ALIGN(sge_bytes + offset) >> PAGE_SHIFT; 385 sge++; 386 sge_bytes = min(sge->length, size - sent); 387 offset = sge->laddr & ~PAGE_MASK; 388 } else { 389 sent = rv; 390 break; 391 } 392 } 393 return sent; 394 } 395 396 #define MAX_TRAILER (MPA_CRC_SIZE + 4) 397 398 static void siw_unmap_pages(struct kvec *iov, unsigned long kmap_mask, int len) 399 { 400 int i; 401 402 /* 403 * Work backwards through the array to honor the kmap_local_page() 404 * ordering requirements. 405 */ 406 for (i = (len-1); i >= 0; i--) { 407 if (kmap_mask & BIT(i)) { 408 unsigned long addr = (unsigned long)iov[i].iov_base; 409 410 kunmap_local((void *)(addr & PAGE_MASK)); 411 } 412 } 413 } 414 415 /* 416 * siw_tx_hdt() tries to push a complete packet to TCP where all 417 * packet fragments are referenced by the elements of one iovec. 418 * For the data portion, each involved page must be referenced by 419 * one extra element. All sge's data can be non-aligned to page 420 * boundaries. Two more elements are referencing iWARP header 421 * and trailer: 422 * MAX_ARRAY = 64KB/PAGE_SIZE + 1 + (2 * (SIW_MAX_SGE - 1) + HDR + TRL 423 */ 424 #define MAX_ARRAY ((0xffff / PAGE_SIZE) + 1 + (2 * (SIW_MAX_SGE - 1) + 2)) 425 426 /* 427 * Write out iov referencing hdr, data and trailer of current FPDU. 428 * Update transmit state dependent on write return status 429 */ 430 static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s) 431 { 432 struct siw_wqe *wqe = &c_tx->wqe_active; 433 struct siw_sge *sge = &wqe->sqe.sge[c_tx->sge_idx]; 434 struct kvec iov[MAX_ARRAY]; 435 struct page *page_array[MAX_ARRAY]; 436 struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_EOR }; 437 438 int seg = 0, do_crc = c_tx->do_crc, is_kva = 0, rv; 439 unsigned int data_len = c_tx->bytes_unsent, hdr_len = 0, trl_len = 0, 440 sge_off = c_tx->sge_off, sge_idx = c_tx->sge_idx, 441 pbl_idx = c_tx->pbl_idx; 442 unsigned long kmap_mask = 0L; 443 444 if (c_tx->state == SIW_SEND_HDR) { 445 if (c_tx->use_sendpage) { 446 rv = siw_tx_ctrl(c_tx, s, MSG_DONTWAIT | MSG_MORE); 447 if (rv) 448 goto done; 449 450 c_tx->state = SIW_SEND_DATA; 451 } else { 452 iov[0].iov_base = 453 (char *)&c_tx->pkt.ctrl + c_tx->ctrl_sent; 454 iov[0].iov_len = hdr_len = 455 c_tx->ctrl_len - c_tx->ctrl_sent; 456 seg = 1; 457 } 458 } 459 460 wqe->processed += data_len; 461 462 while (data_len) { /* walk the list of SGE's */ 463 unsigned int sge_len = min(sge->length - sge_off, data_len); 464 unsigned int fp_off = (sge->laddr + sge_off) & ~PAGE_MASK; 465 struct siw_mem *mem; 466 467 if (!(tx_flags(wqe) & SIW_WQE_INLINE)) { 468 mem = wqe->mem[sge_idx]; 469 is_kva = mem->mem_obj == NULL ? 1 : 0; 470 } else { 471 is_kva = 1; 472 } 473 if (is_kva && !c_tx->use_sendpage) { 474 /* 475 * tx from kernel virtual address: either inline data 476 * or memory region with assigned kernel buffer 477 */ 478 iov[seg].iov_base = 479 ib_virt_dma_to_ptr(sge->laddr + sge_off); 480 iov[seg].iov_len = sge_len; 481 482 if (do_crc) 483 crypto_shash_update(c_tx->mpa_crc_hd, 484 iov[seg].iov_base, 485 sge_len); 486 sge_off += sge_len; 487 data_len -= sge_len; 488 seg++; 489 goto sge_done; 490 } 491 492 while (sge_len) { 493 size_t plen = min((int)PAGE_SIZE - fp_off, sge_len); 494 void *kaddr; 495 496 if (!is_kva) { 497 struct page *p; 498 499 if (mem->is_pbl) 500 p = siw_get_pblpage( 501 mem, sge->laddr + sge_off, 502 &pbl_idx); 503 else 504 p = siw_get_upage(mem->umem, 505 sge->laddr + sge_off); 506 if (unlikely(!p)) { 507 siw_unmap_pages(iov, kmap_mask, seg); 508 wqe->processed -= c_tx->bytes_unsent; 509 rv = -EFAULT; 510 goto done_crc; 511 } 512 page_array[seg] = p; 513 514 if (!c_tx->use_sendpage) { 515 void *kaddr = kmap_local_page(p); 516 517 /* Remember for later kunmap() */ 518 kmap_mask |= BIT(seg); 519 iov[seg].iov_base = kaddr + fp_off; 520 iov[seg].iov_len = plen; 521 522 if (do_crc) 523 crypto_shash_update( 524 c_tx->mpa_crc_hd, 525 iov[seg].iov_base, 526 plen); 527 } else if (do_crc) { 528 kaddr = kmap_local_page(p); 529 crypto_shash_update(c_tx->mpa_crc_hd, 530 kaddr + fp_off, 531 plen); 532 kunmap_local(kaddr); 533 } 534 } else { 535 /* 536 * Cast to an uintptr_t to preserve all 64 bits 537 * in sge->laddr. 538 */ 539 u64 va = sge->laddr + sge_off; 540 541 page_array[seg] = ib_virt_dma_to_page(va); 542 if (do_crc) 543 crypto_shash_update( 544 c_tx->mpa_crc_hd, 545 ib_virt_dma_to_ptr(va), 546 plen); 547 } 548 549 sge_len -= plen; 550 sge_off += plen; 551 data_len -= plen; 552 fp_off = 0; 553 554 if (++seg >= (int)MAX_ARRAY) { 555 siw_dbg_qp(tx_qp(c_tx), "to many fragments\n"); 556 siw_unmap_pages(iov, kmap_mask, seg-1); 557 wqe->processed -= c_tx->bytes_unsent; 558 rv = -EMSGSIZE; 559 goto done_crc; 560 } 561 } 562 sge_done: 563 /* Update SGE variables at end of SGE */ 564 if (sge_off == sge->length && 565 (data_len != 0 || wqe->processed < wqe->bytes)) { 566 sge_idx++; 567 sge++; 568 sge_off = 0; 569 } 570 } 571 /* trailer */ 572 if (likely(c_tx->state != SIW_SEND_TRAILER)) { 573 iov[seg].iov_base = &c_tx->trailer.pad[4 - c_tx->pad]; 574 iov[seg].iov_len = trl_len = MAX_TRAILER - (4 - c_tx->pad); 575 } else { 576 iov[seg].iov_base = &c_tx->trailer.pad[c_tx->ctrl_sent]; 577 iov[seg].iov_len = trl_len = MAX_TRAILER - c_tx->ctrl_sent; 578 } 579 580 if (c_tx->pad) { 581 *(u32 *)c_tx->trailer.pad = 0; 582 if (do_crc) 583 crypto_shash_update(c_tx->mpa_crc_hd, 584 (u8 *)&c_tx->trailer.crc - c_tx->pad, 585 c_tx->pad); 586 } 587 if (!c_tx->mpa_crc_hd) 588 c_tx->trailer.crc = 0; 589 else if (do_crc) 590 crypto_shash_final(c_tx->mpa_crc_hd, (u8 *)&c_tx->trailer.crc); 591 592 data_len = c_tx->bytes_unsent; 593 594 if (c_tx->use_sendpage) { 595 rv = siw_0copy_tx(s, page_array, &wqe->sqe.sge[c_tx->sge_idx], 596 c_tx->sge_off, data_len); 597 if (rv == data_len) { 598 rv = kernel_sendmsg(s, &msg, &iov[seg], 1, trl_len); 599 if (rv > 0) 600 rv += data_len; 601 else 602 rv = data_len; 603 } 604 } else { 605 rv = kernel_sendmsg(s, &msg, iov, seg + 1, 606 hdr_len + data_len + trl_len); 607 siw_unmap_pages(iov, kmap_mask, seg); 608 } 609 if (rv < (int)hdr_len) { 610 /* Not even complete hdr pushed or negative rv */ 611 wqe->processed -= data_len; 612 if (rv >= 0) { 613 c_tx->ctrl_sent += rv; 614 rv = -EAGAIN; 615 } 616 goto done_crc; 617 } 618 rv -= hdr_len; 619 620 if (rv >= (int)data_len) { 621 /* all user data pushed to TCP or no data to push */ 622 if (data_len > 0 && wqe->processed < wqe->bytes) { 623 /* Save the current state for next tx */ 624 c_tx->sge_idx = sge_idx; 625 c_tx->sge_off = sge_off; 626 c_tx->pbl_idx = pbl_idx; 627 } 628 rv -= data_len; 629 630 if (rv == trl_len) /* all pushed */ 631 rv = 0; 632 else { 633 c_tx->state = SIW_SEND_TRAILER; 634 c_tx->ctrl_len = MAX_TRAILER; 635 c_tx->ctrl_sent = rv + 4 - c_tx->pad; 636 c_tx->bytes_unsent = 0; 637 rv = -EAGAIN; 638 } 639 640 } else if (data_len > 0) { 641 /* Maybe some user data pushed to TCP */ 642 c_tx->state = SIW_SEND_DATA; 643 wqe->processed -= data_len - rv; 644 645 if (rv) { 646 /* 647 * Some bytes out. Recompute tx state based 648 * on old state and bytes pushed 649 */ 650 unsigned int sge_unsent; 651 652 c_tx->bytes_unsent -= rv; 653 sge = &wqe->sqe.sge[c_tx->sge_idx]; 654 sge_unsent = sge->length - c_tx->sge_off; 655 656 while (sge_unsent <= rv) { 657 rv -= sge_unsent; 658 c_tx->sge_idx++; 659 c_tx->sge_off = 0; 660 sge++; 661 sge_unsent = sge->length; 662 } 663 c_tx->sge_off += rv; 664 } 665 rv = -EAGAIN; 666 } 667 done_crc: 668 c_tx->do_crc = 0; 669 done: 670 return rv; 671 } 672 673 static void siw_update_tcpseg(struct siw_iwarp_tx *c_tx, 674 struct socket *s) 675 { 676 struct tcp_sock *tp = tcp_sk(s->sk); 677 678 if (tp->gso_segs) { 679 if (c_tx->gso_seg_limit == 0) 680 c_tx->tcp_seglen = tp->mss_cache * tp->gso_segs; 681 else 682 c_tx->tcp_seglen = 683 tp->mss_cache * 684 min_t(u16, c_tx->gso_seg_limit, tp->gso_segs); 685 } else { 686 c_tx->tcp_seglen = tp->mss_cache; 687 } 688 /* Loopback may give odd numbers */ 689 c_tx->tcp_seglen &= 0xfffffff8; 690 } 691 692 /* 693 * siw_prepare_fpdu() 694 * 695 * Prepares transmit context to send out one FPDU if FPDU will contain 696 * user data and user data are not immediate data. 697 * Computes maximum FPDU length to fill up TCP MSS if possible. 698 * 699 * @qp: QP from which to transmit 700 * @wqe: Current WQE causing transmission 701 * 702 * TODO: Take into account real available sendspace on socket 703 * to avoid header misalignment due to send pausing within 704 * fpdu transmission 705 */ 706 static void siw_prepare_fpdu(struct siw_qp *qp, struct siw_wqe *wqe) 707 { 708 struct siw_iwarp_tx *c_tx = &qp->tx_ctx; 709 int data_len; 710 711 c_tx->ctrl_len = 712 iwarp_pktinfo[__rdmap_get_opcode(&c_tx->pkt.ctrl)].hdr_len; 713 c_tx->ctrl_sent = 0; 714 715 /* 716 * Update target buffer offset if any 717 */ 718 if (!(c_tx->pkt.ctrl.ddp_rdmap_ctrl & DDP_FLAG_TAGGED)) 719 /* Untagged message */ 720 c_tx->pkt.c_untagged.ddp_mo = cpu_to_be32(wqe->processed); 721 else /* Tagged message */ 722 c_tx->pkt.c_tagged.ddp_to = 723 cpu_to_be64(wqe->sqe.raddr + wqe->processed); 724 725 data_len = wqe->bytes - wqe->processed; 726 if (data_len + c_tx->ctrl_len + MPA_CRC_SIZE > c_tx->tcp_seglen) { 727 /* Trim DDP payload to fit into current TCP segment */ 728 data_len = c_tx->tcp_seglen - (c_tx->ctrl_len + MPA_CRC_SIZE); 729 c_tx->pkt.ctrl.ddp_rdmap_ctrl &= ~DDP_FLAG_LAST; 730 c_tx->pad = 0; 731 } else { 732 c_tx->pkt.ctrl.ddp_rdmap_ctrl |= DDP_FLAG_LAST; 733 c_tx->pad = -data_len & 0x3; 734 } 735 c_tx->bytes_unsent = data_len; 736 737 c_tx->pkt.ctrl.mpa_len = 738 htons(c_tx->ctrl_len + data_len - MPA_HDR_SIZE); 739 740 /* 741 * Init MPA CRC computation 742 */ 743 if (c_tx->mpa_crc_hd) { 744 crypto_shash_init(c_tx->mpa_crc_hd); 745 crypto_shash_update(c_tx->mpa_crc_hd, (u8 *)&c_tx->pkt, 746 c_tx->ctrl_len); 747 c_tx->do_crc = 1; 748 } 749 } 750 751 /* 752 * siw_check_sgl_tx() 753 * 754 * Check permissions for a list of SGE's (SGL). 755 * A successful check will have all memory referenced 756 * for transmission resolved and assigned to the WQE. 757 * 758 * @pd: Protection Domain SGL should belong to 759 * @wqe: WQE to be checked 760 * @perms: requested access permissions 761 * 762 */ 763 764 static int siw_check_sgl_tx(struct ib_pd *pd, struct siw_wqe *wqe, 765 enum ib_access_flags perms) 766 { 767 struct siw_sge *sge = &wqe->sqe.sge[0]; 768 int i, len, num_sge = wqe->sqe.num_sge; 769 770 if (unlikely(num_sge > SIW_MAX_SGE)) 771 return -EINVAL; 772 773 for (i = 0, len = 0; num_sge; num_sge--, i++, sge++) { 774 /* 775 * rdma verbs: do not check stag for a zero length sge 776 */ 777 if (sge->length) { 778 int rv = siw_check_sge(pd, sge, &wqe->mem[i], perms, 0, 779 sge->length); 780 781 if (unlikely(rv != E_ACCESS_OK)) 782 return rv; 783 } 784 len += sge->length; 785 } 786 return len; 787 } 788 789 /* 790 * siw_qp_sq_proc_tx() 791 * 792 * Process one WQE which needs transmission on the wire. 793 */ 794 static int siw_qp_sq_proc_tx(struct siw_qp *qp, struct siw_wqe *wqe) 795 { 796 struct siw_iwarp_tx *c_tx = &qp->tx_ctx; 797 struct socket *s = qp->attrs.sk; 798 int rv = 0, burst_len = qp->tx_ctx.burst; 799 enum rdmap_ecode ecode = RDMAP_ECODE_CATASTROPHIC_STREAM; 800 801 if (unlikely(wqe->wr_status == SIW_WR_IDLE)) 802 return 0; 803 804 if (!burst_len) 805 burst_len = SQ_USER_MAXBURST; 806 807 if (wqe->wr_status == SIW_WR_QUEUED) { 808 if (!(wqe->sqe.flags & SIW_WQE_INLINE)) { 809 if (tx_type(wqe) == SIW_OP_READ_RESPONSE) 810 wqe->sqe.num_sge = 1; 811 812 if (tx_type(wqe) != SIW_OP_READ && 813 tx_type(wqe) != SIW_OP_READ_LOCAL_INV) { 814 /* 815 * Reference memory to be tx'd w/o checking 816 * access for LOCAL_READ permission, since 817 * not defined in RDMA core. 818 */ 819 rv = siw_check_sgl_tx(qp->pd, wqe, 0); 820 if (rv < 0) { 821 if (tx_type(wqe) == 822 SIW_OP_READ_RESPONSE) 823 ecode = siw_rdmap_error(-rv); 824 rv = -EINVAL; 825 goto tx_error; 826 } 827 wqe->bytes = rv; 828 } else { 829 wqe->bytes = 0; 830 } 831 } else { 832 wqe->bytes = wqe->sqe.sge[0].length; 833 if (!rdma_is_kernel_res(&qp->base_qp.res)) { 834 if (wqe->bytes > SIW_MAX_INLINE) { 835 rv = -EINVAL; 836 goto tx_error; 837 } 838 wqe->sqe.sge[0].laddr = 839 (u64)(uintptr_t)&wqe->sqe.sge[1]; 840 } 841 } 842 wqe->wr_status = SIW_WR_INPROGRESS; 843 wqe->processed = 0; 844 845 siw_update_tcpseg(c_tx, s); 846 847 rv = siw_qp_prepare_tx(c_tx); 848 if (rv == PKT_FRAGMENTED) { 849 c_tx->state = SIW_SEND_HDR; 850 siw_prepare_fpdu(qp, wqe); 851 } else if (rv == PKT_COMPLETE) { 852 c_tx->state = SIW_SEND_SHORT_FPDU; 853 } else { 854 goto tx_error; 855 } 856 } 857 858 next_segment: 859 siw_dbg_qp(qp, "wr type %d, state %d, data %u, sent %u, id %llx\n", 860 tx_type(wqe), wqe->wr_status, wqe->bytes, wqe->processed, 861 wqe->sqe.id); 862 863 if (--burst_len == 0) { 864 rv = -EINPROGRESS; 865 goto tx_done; 866 } 867 if (c_tx->state == SIW_SEND_SHORT_FPDU) { 868 enum siw_opcode tx_type = tx_type(wqe); 869 unsigned int msg_flags; 870 871 if (siw_sq_empty(qp) || !siw_tcp_nagle || burst_len == 1) 872 /* 873 * End current TCP segment, if SQ runs empty, 874 * or siw_tcp_nagle is not set, or we bail out 875 * soon due to no burst credit left. 876 */ 877 msg_flags = MSG_DONTWAIT; 878 else 879 msg_flags = MSG_DONTWAIT | MSG_MORE; 880 881 rv = siw_tx_ctrl(c_tx, s, msg_flags); 882 883 if (!rv && tx_type != SIW_OP_READ && 884 tx_type != SIW_OP_READ_LOCAL_INV) 885 wqe->processed = wqe->bytes; 886 887 goto tx_done; 888 889 } else { 890 rv = siw_tx_hdt(c_tx, s); 891 } 892 if (!rv) { 893 /* 894 * One segment sent. Processing completed if last 895 * segment, Do next segment otherwise. 896 */ 897 if (unlikely(c_tx->tx_suspend)) { 898 /* 899 * Verbs, 6.4.: Try stopping sending after a full 900 * DDP segment if the connection goes down 901 * (== peer halfclose) 902 */ 903 rv = -ECONNABORTED; 904 goto tx_done; 905 } 906 if (c_tx->pkt.ctrl.ddp_rdmap_ctrl & DDP_FLAG_LAST) { 907 siw_dbg_qp(qp, "WQE completed\n"); 908 goto tx_done; 909 } 910 c_tx->state = SIW_SEND_HDR; 911 912 siw_update_tcpseg(c_tx, s); 913 914 siw_prepare_fpdu(qp, wqe); 915 goto next_segment; 916 } 917 tx_done: 918 qp->tx_ctx.burst = burst_len; 919 return rv; 920 921 tx_error: 922 if (ecode != RDMAP_ECODE_CATASTROPHIC_STREAM) 923 siw_init_terminate(qp, TERM_ERROR_LAYER_RDMAP, 924 RDMAP_ETYPE_REMOTE_PROTECTION, ecode, 1); 925 else 926 siw_init_terminate(qp, TERM_ERROR_LAYER_RDMAP, 927 RDMAP_ETYPE_CATASTROPHIC, 928 RDMAP_ECODE_UNSPECIFIED, 1); 929 return rv; 930 } 931 932 static int siw_fastreg_mr(struct ib_pd *pd, struct siw_sqe *sqe) 933 { 934 struct ib_mr *base_mr = (struct ib_mr *)(uintptr_t)sqe->base_mr; 935 struct siw_device *sdev = to_siw_dev(pd->device); 936 struct siw_mem *mem; 937 int rv = 0; 938 939 siw_dbg_pd(pd, "STag 0x%08x\n", sqe->rkey); 940 941 if (unlikely(!base_mr)) { 942 pr_warn("siw: fastreg: STag 0x%08x unknown\n", sqe->rkey); 943 return -EINVAL; 944 } 945 946 if (unlikely(base_mr->rkey >> 8 != sqe->rkey >> 8)) { 947 pr_warn("siw: fastreg: STag 0x%08x: bad MR\n", sqe->rkey); 948 return -EINVAL; 949 } 950 951 mem = siw_mem_id2obj(sdev, sqe->rkey >> 8); 952 if (unlikely(!mem)) { 953 pr_warn("siw: fastreg: STag 0x%08x unknown\n", sqe->rkey); 954 return -EINVAL; 955 } 956 957 if (unlikely(mem->pd != pd)) { 958 pr_warn("siw: fastreg: PD mismatch\n"); 959 rv = -EINVAL; 960 goto out; 961 } 962 if (unlikely(mem->stag_valid)) { 963 pr_warn("siw: fastreg: STag 0x%08x already valid\n", sqe->rkey); 964 rv = -EINVAL; 965 goto out; 966 } 967 /* Refresh STag since user may have changed key part */ 968 mem->stag = sqe->rkey; 969 mem->perms = sqe->access; 970 971 siw_dbg_mem(mem, "STag 0x%08x now valid\n", sqe->rkey); 972 mem->va = base_mr->iova; 973 mem->stag_valid = 1; 974 out: 975 siw_mem_put(mem); 976 return rv; 977 } 978 979 static int siw_qp_sq_proc_local(struct siw_qp *qp, struct siw_wqe *wqe) 980 { 981 int rv; 982 983 switch (tx_type(wqe)) { 984 case SIW_OP_REG_MR: 985 rv = siw_fastreg_mr(qp->pd, &wqe->sqe); 986 break; 987 988 case SIW_OP_INVAL_STAG: 989 rv = siw_invalidate_stag(qp->pd, wqe->sqe.rkey); 990 break; 991 992 default: 993 rv = -EINVAL; 994 } 995 return rv; 996 } 997 998 /* 999 * siw_qp_sq_process() 1000 * 1001 * Core TX path routine for RDMAP/DDP/MPA using a TCP kernel socket. 1002 * Sends RDMAP payload for the current SQ WR @wqe of @qp in one or more 1003 * MPA FPDUs, each containing a DDP segment. 1004 * 1005 * SQ processing may occur in user context as a result of posting 1006 * new WQE's or from siw_sq_work_handler() context. Processing in 1007 * user context is limited to non-kernel verbs users. 1008 * 1009 * SQ processing may get paused anytime, possibly in the middle of a WR 1010 * or FPDU, if insufficient send space is available. SQ processing 1011 * gets resumed from siw_sq_work_handler(), if send space becomes 1012 * available again. 1013 * 1014 * Must be called with the QP state read-locked. 1015 * 1016 * Note: 1017 * An outbound RREQ can be satisfied by the corresponding RRESP 1018 * _before_ it gets assigned to the ORQ. This happens regularly 1019 * in RDMA READ via loopback case. Since both outbound RREQ and 1020 * inbound RRESP can be handled by the same CPU, locking the ORQ 1021 * is dead-lock prone and thus not an option. With that, the 1022 * RREQ gets assigned to the ORQ _before_ being sent - see 1023 * siw_activate_tx() - and pulled back in case of send failure. 1024 */ 1025 int siw_qp_sq_process(struct siw_qp *qp) 1026 { 1027 struct siw_wqe *wqe = tx_wqe(qp); 1028 enum siw_opcode tx_type; 1029 unsigned long flags; 1030 int rv = 0; 1031 1032 siw_dbg_qp(qp, "enter for type %d\n", tx_type(wqe)); 1033 1034 next_wqe: 1035 /* 1036 * Stop QP processing if SQ state changed 1037 */ 1038 if (unlikely(qp->tx_ctx.tx_suspend)) { 1039 siw_dbg_qp(qp, "tx suspended\n"); 1040 goto done; 1041 } 1042 tx_type = tx_type(wqe); 1043 1044 if (tx_type <= SIW_OP_READ_RESPONSE) 1045 rv = siw_qp_sq_proc_tx(qp, wqe); 1046 else 1047 rv = siw_qp_sq_proc_local(qp, wqe); 1048 1049 if (!rv) { 1050 /* 1051 * WQE processing done 1052 */ 1053 switch (tx_type) { 1054 case SIW_OP_SEND: 1055 case SIW_OP_SEND_REMOTE_INV: 1056 case SIW_OP_WRITE: 1057 siw_wqe_put_mem(wqe, tx_type); 1058 fallthrough; 1059 1060 case SIW_OP_INVAL_STAG: 1061 case SIW_OP_REG_MR: 1062 if (tx_flags(wqe) & SIW_WQE_SIGNALLED) 1063 siw_sqe_complete(qp, &wqe->sqe, wqe->bytes, 1064 SIW_WC_SUCCESS); 1065 break; 1066 1067 case SIW_OP_READ: 1068 case SIW_OP_READ_LOCAL_INV: 1069 /* 1070 * already enqueued to ORQ queue 1071 */ 1072 break; 1073 1074 case SIW_OP_READ_RESPONSE: 1075 siw_wqe_put_mem(wqe, tx_type); 1076 break; 1077 1078 default: 1079 WARN(1, "undefined WQE type %d\n", tx_type); 1080 rv = -EINVAL; 1081 goto done; 1082 } 1083 1084 spin_lock_irqsave(&qp->sq_lock, flags); 1085 wqe->wr_status = SIW_WR_IDLE; 1086 rv = siw_activate_tx(qp); 1087 spin_unlock_irqrestore(&qp->sq_lock, flags); 1088 1089 if (rv <= 0) 1090 goto done; 1091 1092 goto next_wqe; 1093 1094 } else if (rv == -EAGAIN) { 1095 siw_dbg_qp(qp, "sq paused: hd/tr %d of %d, data %d\n", 1096 qp->tx_ctx.ctrl_sent, qp->tx_ctx.ctrl_len, 1097 qp->tx_ctx.bytes_unsent); 1098 rv = 0; 1099 goto done; 1100 } else if (rv == -EINPROGRESS) { 1101 rv = siw_sq_start(qp); 1102 goto done; 1103 } else { 1104 /* 1105 * WQE processing failed. 1106 * Verbs 8.3.2: 1107 * o It turns any WQE into a signalled WQE. 1108 * o Local catastrophic error must be surfaced 1109 * o QP must be moved into Terminate state: done by code 1110 * doing socket state change processing 1111 * 1112 * o TODO: Termination message must be sent. 1113 * o TODO: Implement more precise work completion errors, 1114 * see enum ib_wc_status in ib_verbs.h 1115 */ 1116 siw_dbg_qp(qp, "wqe type %d processing failed: %d\n", 1117 tx_type(wqe), rv); 1118 1119 spin_lock_irqsave(&qp->sq_lock, flags); 1120 /* 1121 * RREQ may have already been completed by inbound RRESP! 1122 */ 1123 if ((tx_type == SIW_OP_READ || 1124 tx_type == SIW_OP_READ_LOCAL_INV) && qp->attrs.orq_size) { 1125 /* Cleanup pending entry in ORQ */ 1126 qp->orq_put--; 1127 qp->orq[qp->orq_put % qp->attrs.orq_size].flags = 0; 1128 } 1129 spin_unlock_irqrestore(&qp->sq_lock, flags); 1130 /* 1131 * immediately suspends further TX processing 1132 */ 1133 if (!qp->tx_ctx.tx_suspend) 1134 siw_qp_cm_drop(qp, 0); 1135 1136 switch (tx_type) { 1137 case SIW_OP_SEND: 1138 case SIW_OP_SEND_REMOTE_INV: 1139 case SIW_OP_SEND_WITH_IMM: 1140 case SIW_OP_WRITE: 1141 case SIW_OP_READ: 1142 case SIW_OP_READ_LOCAL_INV: 1143 siw_wqe_put_mem(wqe, tx_type); 1144 fallthrough; 1145 1146 case SIW_OP_INVAL_STAG: 1147 case SIW_OP_REG_MR: 1148 siw_sqe_complete(qp, &wqe->sqe, wqe->bytes, 1149 SIW_WC_LOC_QP_OP_ERR); 1150 1151 siw_qp_event(qp, IB_EVENT_QP_FATAL); 1152 1153 break; 1154 1155 case SIW_OP_READ_RESPONSE: 1156 siw_dbg_qp(qp, "proc. read.response failed: %d\n", rv); 1157 1158 siw_qp_event(qp, IB_EVENT_QP_REQ_ERR); 1159 1160 siw_wqe_put_mem(wqe, SIW_OP_READ_RESPONSE); 1161 1162 break; 1163 1164 default: 1165 WARN(1, "undefined WQE type %d\n", tx_type); 1166 rv = -EINVAL; 1167 } 1168 wqe->wr_status = SIW_WR_IDLE; 1169 } 1170 done: 1171 return rv; 1172 } 1173 1174 static void siw_sq_resume(struct siw_qp *qp) 1175 { 1176 if (down_read_trylock(&qp->state_lock)) { 1177 if (likely(qp->attrs.state == SIW_QP_STATE_RTS && 1178 !qp->tx_ctx.tx_suspend)) { 1179 int rv = siw_qp_sq_process(qp); 1180 1181 up_read(&qp->state_lock); 1182 1183 if (unlikely(rv < 0)) { 1184 siw_dbg_qp(qp, "SQ task failed: err %d\n", rv); 1185 1186 if (!qp->tx_ctx.tx_suspend) 1187 siw_qp_cm_drop(qp, 0); 1188 } 1189 } else { 1190 up_read(&qp->state_lock); 1191 } 1192 } else { 1193 siw_dbg_qp(qp, "Resume SQ while QP locked\n"); 1194 } 1195 siw_qp_put(qp); 1196 } 1197 1198 struct tx_task_t { 1199 struct llist_head active; 1200 wait_queue_head_t waiting; 1201 }; 1202 1203 static DEFINE_PER_CPU(struct tx_task_t, siw_tx_task_g); 1204 1205 void siw_stop_tx_thread(int nr_cpu) 1206 { 1207 kthread_stop(siw_tx_thread[nr_cpu]); 1208 wake_up(&per_cpu(siw_tx_task_g, nr_cpu).waiting); 1209 } 1210 1211 int siw_run_sq(void *data) 1212 { 1213 const int nr_cpu = (unsigned int)(long)data; 1214 struct llist_node *active; 1215 struct siw_qp *qp; 1216 struct tx_task_t *tx_task = &per_cpu(siw_tx_task_g, nr_cpu); 1217 1218 init_llist_head(&tx_task->active); 1219 init_waitqueue_head(&tx_task->waiting); 1220 1221 while (1) { 1222 struct llist_node *fifo_list = NULL; 1223 1224 wait_event_interruptible(tx_task->waiting, 1225 !llist_empty(&tx_task->active) || 1226 kthread_should_stop()); 1227 1228 if (kthread_should_stop()) 1229 break; 1230 1231 active = llist_del_all(&tx_task->active); 1232 /* 1233 * llist_del_all returns a list with newest entry first. 1234 * Re-order list for fairness among QP's. 1235 */ 1236 while (active) { 1237 struct llist_node *tmp = active; 1238 1239 active = llist_next(active); 1240 tmp->next = fifo_list; 1241 fifo_list = tmp; 1242 } 1243 while (fifo_list) { 1244 qp = container_of(fifo_list, struct siw_qp, tx_list); 1245 fifo_list = llist_next(fifo_list); 1246 qp->tx_list.next = NULL; 1247 1248 siw_sq_resume(qp); 1249 } 1250 } 1251 active = llist_del_all(&tx_task->active); 1252 if (active) { 1253 llist_for_each_entry(qp, active, tx_list) { 1254 qp->tx_list.next = NULL; 1255 siw_sq_resume(qp); 1256 } 1257 } 1258 return 0; 1259 } 1260 1261 int siw_sq_start(struct siw_qp *qp) 1262 { 1263 if (tx_wqe(qp)->wr_status == SIW_WR_IDLE) 1264 return 0; 1265 1266 if (unlikely(!cpu_online(qp->tx_cpu))) { 1267 siw_put_tx_cpu(qp->tx_cpu); 1268 qp->tx_cpu = siw_get_tx_cpu(qp->sdev); 1269 if (qp->tx_cpu < 0) { 1270 pr_warn("siw: no tx cpu available\n"); 1271 1272 return -EIO; 1273 } 1274 } 1275 siw_qp_get(qp); 1276 1277 llist_add(&qp->tx_list, &per_cpu(siw_tx_task_g, qp->tx_cpu).active); 1278 1279 wake_up(&per_cpu(siw_tx_task_g, qp->tx_cpu).waiting); 1280 1281 return 0; 1282 } 1283