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_local_page(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_local(buffer); 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_local_page(p); 100 memcpy(paddr + part, buffer, bytes - part); 101 } 102 kunmap_local(buffer); 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 kvec *iov, unsigned long kmap_mask, int len) 400 { 401 int i; 402 403 /* 404 * Work backwards through the array to honor the kmap_local_page() 405 * ordering requirements. 406 */ 407 for (i = (len-1); i >= 0; i--) { 408 if (kmap_mask & BIT(i)) { 409 unsigned long addr = (unsigned long)iov[i].iov_base; 410 411 kunmap_local((void *)(addr & PAGE_MASK)); 412 } 413 } 414 } 415 416 /* 417 * siw_tx_hdt() tries to push a complete packet to TCP where all 418 * packet fragments are referenced by the elements of one iovec. 419 * For the data portion, each involved page must be referenced by 420 * one extra element. All sge's data can be non-aligned to page 421 * boundaries. Two more elements are referencing iWARP header 422 * and trailer: 423 * MAX_ARRAY = 64KB/PAGE_SIZE + 1 + (2 * (SIW_MAX_SGE - 1) + HDR + TRL 424 */ 425 #define MAX_ARRAY ((0xffff / PAGE_SIZE) + 1 + (2 * (SIW_MAX_SGE - 1) + 2)) 426 427 /* 428 * Write out iov referencing hdr, data and trailer of current FPDU. 429 * Update transmit state dependent on write return status 430 */ 431 static int siw_tx_hdt(struct siw_iwarp_tx *c_tx, struct socket *s) 432 { 433 struct siw_wqe *wqe = &c_tx->wqe_active; 434 struct siw_sge *sge = &wqe->sqe.sge[c_tx->sge_idx]; 435 struct kvec iov[MAX_ARRAY]; 436 struct page *page_array[MAX_ARRAY]; 437 struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_EOR }; 438 439 int seg = 0, do_crc = c_tx->do_crc, is_kva = 0, rv; 440 unsigned int data_len = c_tx->bytes_unsent, hdr_len = 0, trl_len = 0, 441 sge_off = c_tx->sge_off, sge_idx = c_tx->sge_idx, 442 pbl_idx = c_tx->pbl_idx; 443 unsigned long kmap_mask = 0L; 444 445 if (c_tx->state == SIW_SEND_HDR) { 446 if (c_tx->use_sendpage) { 447 rv = siw_tx_ctrl(c_tx, s, MSG_DONTWAIT | MSG_MORE); 448 if (rv) 449 goto done; 450 451 c_tx->state = SIW_SEND_DATA; 452 } else { 453 iov[0].iov_base = 454 (char *)&c_tx->pkt.ctrl + c_tx->ctrl_sent; 455 iov[0].iov_len = hdr_len = 456 c_tx->ctrl_len - c_tx->ctrl_sent; 457 seg = 1; 458 } 459 } 460 461 wqe->processed += data_len; 462 463 while (data_len) { /* walk the list of SGE's */ 464 unsigned int sge_len = min(sge->length - sge_off, data_len); 465 unsigned int fp_off = (sge->laddr + sge_off) & ~PAGE_MASK; 466 struct siw_mem *mem; 467 468 if (!(tx_flags(wqe) & SIW_WQE_INLINE)) { 469 mem = wqe->mem[sge_idx]; 470 is_kva = mem->mem_obj == NULL ? 1 : 0; 471 } else { 472 is_kva = 1; 473 } 474 if (is_kva && !c_tx->use_sendpage) { 475 /* 476 * tx from kernel virtual address: either inline data 477 * or memory region with assigned kernel buffer 478 */ 479 iov[seg].iov_base = 480 (void *)(uintptr_t)(sge->laddr + sge_off); 481 iov[seg].iov_len = sge_len; 482 483 if (do_crc) 484 crypto_shash_update(c_tx->mpa_crc_hd, 485 iov[seg].iov_base, 486 sge_len); 487 sge_off += sge_len; 488 data_len -= sge_len; 489 seg++; 490 goto sge_done; 491 } 492 493 while (sge_len) { 494 size_t plen = min((int)PAGE_SIZE - fp_off, sge_len); 495 void *kaddr; 496 497 if (!is_kva) { 498 struct page *p; 499 500 if (mem->is_pbl) 501 p = siw_get_pblpage( 502 mem, sge->laddr + sge_off, 503 &pbl_idx); 504 else 505 p = siw_get_upage(mem->umem, 506 sge->laddr + sge_off); 507 if (unlikely(!p)) { 508 siw_unmap_pages(iov, kmap_mask, seg); 509 wqe->processed -= c_tx->bytes_unsent; 510 rv = -EFAULT; 511 goto done_crc; 512 } 513 page_array[seg] = p; 514 515 if (!c_tx->use_sendpage) { 516 void *kaddr = kmap_local_page(p); 517 518 /* Remember for later kunmap() */ 519 kmap_mask |= BIT(seg); 520 iov[seg].iov_base = kaddr + fp_off; 521 iov[seg].iov_len = plen; 522 523 if (do_crc) 524 crypto_shash_update( 525 c_tx->mpa_crc_hd, 526 iov[seg].iov_base, 527 plen); 528 } else if (do_crc) { 529 kaddr = kmap_local_page(p); 530 crypto_shash_update(c_tx->mpa_crc_hd, 531 kaddr + fp_off, 532 plen); 533 kunmap_local(kaddr); 534 } 535 } else { 536 u64 va = sge->laddr + sge_off; 537 538 page_array[seg] = virt_to_page(va & PAGE_MASK); 539 if (do_crc) 540 crypto_shash_update( 541 c_tx->mpa_crc_hd, 542 (void *)(uintptr_t)va, 543 plen); 544 } 545 546 sge_len -= plen; 547 sge_off += plen; 548 data_len -= plen; 549 fp_off = 0; 550 551 if (++seg > (int)MAX_ARRAY) { 552 siw_dbg_qp(tx_qp(c_tx), "to many fragments\n"); 553 siw_unmap_pages(iov, kmap_mask, seg-1); 554 wqe->processed -= c_tx->bytes_unsent; 555 rv = -EMSGSIZE; 556 goto done_crc; 557 } 558 } 559 sge_done: 560 /* Update SGE variables at end of SGE */ 561 if (sge_off == sge->length && 562 (data_len != 0 || wqe->processed < wqe->bytes)) { 563 sge_idx++; 564 sge++; 565 sge_off = 0; 566 } 567 } 568 /* trailer */ 569 if (likely(c_tx->state != SIW_SEND_TRAILER)) { 570 iov[seg].iov_base = &c_tx->trailer.pad[4 - c_tx->pad]; 571 iov[seg].iov_len = trl_len = MAX_TRAILER - (4 - c_tx->pad); 572 } else { 573 iov[seg].iov_base = &c_tx->trailer.pad[c_tx->ctrl_sent]; 574 iov[seg].iov_len = trl_len = MAX_TRAILER - c_tx->ctrl_sent; 575 } 576 577 if (c_tx->pad) { 578 *(u32 *)c_tx->trailer.pad = 0; 579 if (do_crc) 580 crypto_shash_update(c_tx->mpa_crc_hd, 581 (u8 *)&c_tx->trailer.crc - c_tx->pad, 582 c_tx->pad); 583 } 584 if (!c_tx->mpa_crc_hd) 585 c_tx->trailer.crc = 0; 586 else if (do_crc) 587 crypto_shash_final(c_tx->mpa_crc_hd, (u8 *)&c_tx->trailer.crc); 588 589 data_len = c_tx->bytes_unsent; 590 591 if (c_tx->use_sendpage) { 592 rv = siw_0copy_tx(s, page_array, &wqe->sqe.sge[c_tx->sge_idx], 593 c_tx->sge_off, data_len); 594 if (rv == data_len) { 595 rv = kernel_sendmsg(s, &msg, &iov[seg], 1, trl_len); 596 if (rv > 0) 597 rv += data_len; 598 else 599 rv = data_len; 600 } 601 } else { 602 rv = kernel_sendmsg(s, &msg, iov, seg + 1, 603 hdr_len + data_len + trl_len); 604 siw_unmap_pages(iov, kmap_mask, seg); 605 } 606 if (rv < (int)hdr_len) { 607 /* Not even complete hdr pushed or negative rv */ 608 wqe->processed -= data_len; 609 if (rv >= 0) { 610 c_tx->ctrl_sent += rv; 611 rv = -EAGAIN; 612 } 613 goto done_crc; 614 } 615 rv -= hdr_len; 616 617 if (rv >= (int)data_len) { 618 /* all user data pushed to TCP or no data to push */ 619 if (data_len > 0 && wqe->processed < wqe->bytes) { 620 /* Save the current state for next tx */ 621 c_tx->sge_idx = sge_idx; 622 c_tx->sge_off = sge_off; 623 c_tx->pbl_idx = pbl_idx; 624 } 625 rv -= data_len; 626 627 if (rv == trl_len) /* all pushed */ 628 rv = 0; 629 else { 630 c_tx->state = SIW_SEND_TRAILER; 631 c_tx->ctrl_len = MAX_TRAILER; 632 c_tx->ctrl_sent = rv + 4 - c_tx->pad; 633 c_tx->bytes_unsent = 0; 634 rv = -EAGAIN; 635 } 636 637 } else if (data_len > 0) { 638 /* Maybe some user data pushed to TCP */ 639 c_tx->state = SIW_SEND_DATA; 640 wqe->processed -= data_len - rv; 641 642 if (rv) { 643 /* 644 * Some bytes out. Recompute tx state based 645 * on old state and bytes pushed 646 */ 647 unsigned int sge_unsent; 648 649 c_tx->bytes_unsent -= rv; 650 sge = &wqe->sqe.sge[c_tx->sge_idx]; 651 sge_unsent = sge->length - c_tx->sge_off; 652 653 while (sge_unsent <= rv) { 654 rv -= sge_unsent; 655 c_tx->sge_idx++; 656 c_tx->sge_off = 0; 657 sge++; 658 sge_unsent = sge->length; 659 } 660 c_tx->sge_off += rv; 661 } 662 rv = -EAGAIN; 663 } 664 done_crc: 665 c_tx->do_crc = 0; 666 done: 667 return rv; 668 } 669 670 static void siw_update_tcpseg(struct siw_iwarp_tx *c_tx, 671 struct socket *s) 672 { 673 struct tcp_sock *tp = tcp_sk(s->sk); 674 675 if (tp->gso_segs) { 676 if (c_tx->gso_seg_limit == 0) 677 c_tx->tcp_seglen = tp->mss_cache * tp->gso_segs; 678 else 679 c_tx->tcp_seglen = 680 tp->mss_cache * 681 min_t(u16, c_tx->gso_seg_limit, tp->gso_segs); 682 } else { 683 c_tx->tcp_seglen = tp->mss_cache; 684 } 685 /* Loopback may give odd numbers */ 686 c_tx->tcp_seglen &= 0xfffffff8; 687 } 688 689 /* 690 * siw_prepare_fpdu() 691 * 692 * Prepares transmit context to send out one FPDU if FPDU will contain 693 * user data and user data are not immediate data. 694 * Computes maximum FPDU length to fill up TCP MSS if possible. 695 * 696 * @qp: QP from which to transmit 697 * @wqe: Current WQE causing transmission 698 * 699 * TODO: Take into account real available sendspace on socket 700 * to avoid header misalignment due to send pausing within 701 * fpdu transmission 702 */ 703 static void siw_prepare_fpdu(struct siw_qp *qp, struct siw_wqe *wqe) 704 { 705 struct siw_iwarp_tx *c_tx = &qp->tx_ctx; 706 int data_len; 707 708 c_tx->ctrl_len = 709 iwarp_pktinfo[__rdmap_get_opcode(&c_tx->pkt.ctrl)].hdr_len; 710 c_tx->ctrl_sent = 0; 711 712 /* 713 * Update target buffer offset if any 714 */ 715 if (!(c_tx->pkt.ctrl.ddp_rdmap_ctrl & DDP_FLAG_TAGGED)) 716 /* Untagged message */ 717 c_tx->pkt.c_untagged.ddp_mo = cpu_to_be32(wqe->processed); 718 else /* Tagged message */ 719 c_tx->pkt.c_tagged.ddp_to = 720 cpu_to_be64(wqe->sqe.raddr + wqe->processed); 721 722 data_len = wqe->bytes - wqe->processed; 723 if (data_len + c_tx->ctrl_len + MPA_CRC_SIZE > c_tx->tcp_seglen) { 724 /* Trim DDP payload to fit into current TCP segment */ 725 data_len = c_tx->tcp_seglen - (c_tx->ctrl_len + MPA_CRC_SIZE); 726 c_tx->pkt.ctrl.ddp_rdmap_ctrl &= ~DDP_FLAG_LAST; 727 c_tx->pad = 0; 728 } else { 729 c_tx->pkt.ctrl.ddp_rdmap_ctrl |= DDP_FLAG_LAST; 730 c_tx->pad = -data_len & 0x3; 731 } 732 c_tx->bytes_unsent = data_len; 733 734 c_tx->pkt.ctrl.mpa_len = 735 htons(c_tx->ctrl_len + data_len - MPA_HDR_SIZE); 736 737 /* 738 * Init MPA CRC computation 739 */ 740 if (c_tx->mpa_crc_hd) { 741 crypto_shash_init(c_tx->mpa_crc_hd); 742 crypto_shash_update(c_tx->mpa_crc_hd, (u8 *)&c_tx->pkt, 743 c_tx->ctrl_len); 744 c_tx->do_crc = 1; 745 } 746 } 747 748 /* 749 * siw_check_sgl_tx() 750 * 751 * Check permissions for a list of SGE's (SGL). 752 * A successful check will have all memory referenced 753 * for transmission resolved and assigned to the WQE. 754 * 755 * @pd: Protection Domain SGL should belong to 756 * @wqe: WQE to be checked 757 * @perms: requested access permissions 758 * 759 */ 760 761 static int siw_check_sgl_tx(struct ib_pd *pd, struct siw_wqe *wqe, 762 enum ib_access_flags perms) 763 { 764 struct siw_sge *sge = &wqe->sqe.sge[0]; 765 int i, len, num_sge = wqe->sqe.num_sge; 766 767 if (unlikely(num_sge > SIW_MAX_SGE)) 768 return -EINVAL; 769 770 for (i = 0, len = 0; num_sge; num_sge--, i++, sge++) { 771 /* 772 * rdma verbs: do not check stag for a zero length sge 773 */ 774 if (sge->length) { 775 int rv = siw_check_sge(pd, sge, &wqe->mem[i], perms, 0, 776 sge->length); 777 778 if (unlikely(rv != E_ACCESS_OK)) 779 return rv; 780 } 781 len += sge->length; 782 } 783 return len; 784 } 785 786 /* 787 * siw_qp_sq_proc_tx() 788 * 789 * Process one WQE which needs transmission on the wire. 790 */ 791 static int siw_qp_sq_proc_tx(struct siw_qp *qp, struct siw_wqe *wqe) 792 { 793 struct siw_iwarp_tx *c_tx = &qp->tx_ctx; 794 struct socket *s = qp->attrs.sk; 795 int rv = 0, burst_len = qp->tx_ctx.burst; 796 enum rdmap_ecode ecode = RDMAP_ECODE_CATASTROPHIC_STREAM; 797 798 if (unlikely(wqe->wr_status == SIW_WR_IDLE)) 799 return 0; 800 801 if (!burst_len) 802 burst_len = SQ_USER_MAXBURST; 803 804 if (wqe->wr_status == SIW_WR_QUEUED) { 805 if (!(wqe->sqe.flags & SIW_WQE_INLINE)) { 806 if (tx_type(wqe) == SIW_OP_READ_RESPONSE) 807 wqe->sqe.num_sge = 1; 808 809 if (tx_type(wqe) != SIW_OP_READ && 810 tx_type(wqe) != SIW_OP_READ_LOCAL_INV) { 811 /* 812 * Reference memory to be tx'd w/o checking 813 * access for LOCAL_READ permission, since 814 * not defined in RDMA core. 815 */ 816 rv = siw_check_sgl_tx(qp->pd, wqe, 0); 817 if (rv < 0) { 818 if (tx_type(wqe) == 819 SIW_OP_READ_RESPONSE) 820 ecode = siw_rdmap_error(-rv); 821 rv = -EINVAL; 822 goto tx_error; 823 } 824 wqe->bytes = rv; 825 } else { 826 wqe->bytes = 0; 827 } 828 } else { 829 wqe->bytes = wqe->sqe.sge[0].length; 830 if (!rdma_is_kernel_res(&qp->base_qp.res)) { 831 if (wqe->bytes > SIW_MAX_INLINE) { 832 rv = -EINVAL; 833 goto tx_error; 834 } 835 wqe->sqe.sge[0].laddr = 836 (u64)(uintptr_t)&wqe->sqe.sge[1]; 837 } 838 } 839 wqe->wr_status = SIW_WR_INPROGRESS; 840 wqe->processed = 0; 841 842 siw_update_tcpseg(c_tx, s); 843 844 rv = siw_qp_prepare_tx(c_tx); 845 if (rv == PKT_FRAGMENTED) { 846 c_tx->state = SIW_SEND_HDR; 847 siw_prepare_fpdu(qp, wqe); 848 } else if (rv == PKT_COMPLETE) { 849 c_tx->state = SIW_SEND_SHORT_FPDU; 850 } else { 851 goto tx_error; 852 } 853 } 854 855 next_segment: 856 siw_dbg_qp(qp, "wr type %d, state %d, data %u, sent %u, id %llx\n", 857 tx_type(wqe), wqe->wr_status, wqe->bytes, wqe->processed, 858 wqe->sqe.id); 859 860 if (--burst_len == 0) { 861 rv = -EINPROGRESS; 862 goto tx_done; 863 } 864 if (c_tx->state == SIW_SEND_SHORT_FPDU) { 865 enum siw_opcode tx_type = tx_type(wqe); 866 unsigned int msg_flags; 867 868 if (siw_sq_empty(qp) || !siw_tcp_nagle || burst_len == 1) 869 /* 870 * End current TCP segment, if SQ runs empty, 871 * or siw_tcp_nagle is not set, or we bail out 872 * soon due to no burst credit left. 873 */ 874 msg_flags = MSG_DONTWAIT; 875 else 876 msg_flags = MSG_DONTWAIT | MSG_MORE; 877 878 rv = siw_tx_ctrl(c_tx, s, msg_flags); 879 880 if (!rv && tx_type != SIW_OP_READ && 881 tx_type != SIW_OP_READ_LOCAL_INV) 882 wqe->processed = wqe->bytes; 883 884 goto tx_done; 885 886 } else { 887 rv = siw_tx_hdt(c_tx, s); 888 } 889 if (!rv) { 890 /* 891 * One segment sent. Processing completed if last 892 * segment, Do next segment otherwise. 893 */ 894 if (unlikely(c_tx->tx_suspend)) { 895 /* 896 * Verbs, 6.4.: Try stopping sending after a full 897 * DDP segment if the connection goes down 898 * (== peer halfclose) 899 */ 900 rv = -ECONNABORTED; 901 goto tx_done; 902 } 903 if (c_tx->pkt.ctrl.ddp_rdmap_ctrl & DDP_FLAG_LAST) { 904 siw_dbg_qp(qp, "WQE completed\n"); 905 goto tx_done; 906 } 907 c_tx->state = SIW_SEND_HDR; 908 909 siw_update_tcpseg(c_tx, s); 910 911 siw_prepare_fpdu(qp, wqe); 912 goto next_segment; 913 } 914 tx_done: 915 qp->tx_ctx.burst = burst_len; 916 return rv; 917 918 tx_error: 919 if (ecode != RDMAP_ECODE_CATASTROPHIC_STREAM) 920 siw_init_terminate(qp, TERM_ERROR_LAYER_RDMAP, 921 RDMAP_ETYPE_REMOTE_PROTECTION, ecode, 1); 922 else 923 siw_init_terminate(qp, TERM_ERROR_LAYER_RDMAP, 924 RDMAP_ETYPE_CATASTROPHIC, 925 RDMAP_ECODE_UNSPECIFIED, 1); 926 return rv; 927 } 928 929 static int siw_fastreg_mr(struct ib_pd *pd, struct siw_sqe *sqe) 930 { 931 struct ib_mr *base_mr = (struct ib_mr *)(uintptr_t)sqe->base_mr; 932 struct siw_device *sdev = to_siw_dev(pd->device); 933 struct siw_mem *mem; 934 int rv = 0; 935 936 siw_dbg_pd(pd, "STag 0x%08x\n", sqe->rkey); 937 938 if (unlikely(!base_mr)) { 939 pr_warn("siw: fastreg: STag 0x%08x unknown\n", sqe->rkey); 940 return -EINVAL; 941 } 942 943 if (unlikely(base_mr->rkey >> 8 != sqe->rkey >> 8)) { 944 pr_warn("siw: fastreg: STag 0x%08x: bad MR\n", sqe->rkey); 945 return -EINVAL; 946 } 947 948 mem = siw_mem_id2obj(sdev, sqe->rkey >> 8); 949 if (unlikely(!mem)) { 950 pr_warn("siw: fastreg: STag 0x%08x unknown\n", sqe->rkey); 951 return -EINVAL; 952 } 953 954 if (unlikely(mem->pd != pd)) { 955 pr_warn("siw: fastreg: PD mismatch\n"); 956 rv = -EINVAL; 957 goto out; 958 } 959 if (unlikely(mem->stag_valid)) { 960 pr_warn("siw: fastreg: STag 0x%08x already valid\n", sqe->rkey); 961 rv = -EINVAL; 962 goto out; 963 } 964 /* Refresh STag since user may have changed key part */ 965 mem->stag = sqe->rkey; 966 mem->perms = sqe->access; 967 968 siw_dbg_mem(mem, "STag 0x%08x now valid\n", sqe->rkey); 969 mem->va = base_mr->iova; 970 mem->stag_valid = 1; 971 out: 972 siw_mem_put(mem); 973 return rv; 974 } 975 976 static int siw_qp_sq_proc_local(struct siw_qp *qp, struct siw_wqe *wqe) 977 { 978 int rv; 979 980 switch (tx_type(wqe)) { 981 case SIW_OP_REG_MR: 982 rv = siw_fastreg_mr(qp->pd, &wqe->sqe); 983 break; 984 985 case SIW_OP_INVAL_STAG: 986 rv = siw_invalidate_stag(qp->pd, wqe->sqe.rkey); 987 break; 988 989 default: 990 rv = -EINVAL; 991 } 992 return rv; 993 } 994 995 /* 996 * siw_qp_sq_process() 997 * 998 * Core TX path routine for RDMAP/DDP/MPA using a TCP kernel socket. 999 * Sends RDMAP payload for the current SQ WR @wqe of @qp in one or more 1000 * MPA FPDUs, each containing a DDP segment. 1001 * 1002 * SQ processing may occur in user context as a result of posting 1003 * new WQE's or from siw_sq_work_handler() context. Processing in 1004 * user context is limited to non-kernel verbs users. 1005 * 1006 * SQ processing may get paused anytime, possibly in the middle of a WR 1007 * or FPDU, if insufficient send space is available. SQ processing 1008 * gets resumed from siw_sq_work_handler(), if send space becomes 1009 * available again. 1010 * 1011 * Must be called with the QP state read-locked. 1012 * 1013 * Note: 1014 * An outbound RREQ can be satisfied by the corresponding RRESP 1015 * _before_ it gets assigned to the ORQ. This happens regularly 1016 * in RDMA READ via loopback case. Since both outbound RREQ and 1017 * inbound RRESP can be handled by the same CPU, locking the ORQ 1018 * is dead-lock prone and thus not an option. With that, the 1019 * RREQ gets assigned to the ORQ _before_ being sent - see 1020 * siw_activate_tx() - and pulled back in case of send failure. 1021 */ 1022 int siw_qp_sq_process(struct siw_qp *qp) 1023 { 1024 struct siw_wqe *wqe = tx_wqe(qp); 1025 enum siw_opcode tx_type; 1026 unsigned long flags; 1027 int rv = 0; 1028 1029 siw_dbg_qp(qp, "enter for type %d\n", tx_type(wqe)); 1030 1031 next_wqe: 1032 /* 1033 * Stop QP processing if SQ state changed 1034 */ 1035 if (unlikely(qp->tx_ctx.tx_suspend)) { 1036 siw_dbg_qp(qp, "tx suspended\n"); 1037 goto done; 1038 } 1039 tx_type = tx_type(wqe); 1040 1041 if (tx_type <= SIW_OP_READ_RESPONSE) 1042 rv = siw_qp_sq_proc_tx(qp, wqe); 1043 else 1044 rv = siw_qp_sq_proc_local(qp, wqe); 1045 1046 if (!rv) { 1047 /* 1048 * WQE processing done 1049 */ 1050 switch (tx_type) { 1051 case SIW_OP_SEND: 1052 case SIW_OP_SEND_REMOTE_INV: 1053 case SIW_OP_WRITE: 1054 siw_wqe_put_mem(wqe, tx_type); 1055 fallthrough; 1056 1057 case SIW_OP_INVAL_STAG: 1058 case SIW_OP_REG_MR: 1059 if (tx_flags(wqe) & SIW_WQE_SIGNALLED) 1060 siw_sqe_complete(qp, &wqe->sqe, wqe->bytes, 1061 SIW_WC_SUCCESS); 1062 break; 1063 1064 case SIW_OP_READ: 1065 case SIW_OP_READ_LOCAL_INV: 1066 /* 1067 * already enqueued to ORQ queue 1068 */ 1069 break; 1070 1071 case SIW_OP_READ_RESPONSE: 1072 siw_wqe_put_mem(wqe, tx_type); 1073 break; 1074 1075 default: 1076 WARN(1, "undefined WQE type %d\n", tx_type); 1077 rv = -EINVAL; 1078 goto done; 1079 } 1080 1081 spin_lock_irqsave(&qp->sq_lock, flags); 1082 wqe->wr_status = SIW_WR_IDLE; 1083 rv = siw_activate_tx(qp); 1084 spin_unlock_irqrestore(&qp->sq_lock, flags); 1085 1086 if (rv <= 0) 1087 goto done; 1088 1089 goto next_wqe; 1090 1091 } else if (rv == -EAGAIN) { 1092 siw_dbg_qp(qp, "sq paused: hd/tr %d of %d, data %d\n", 1093 qp->tx_ctx.ctrl_sent, qp->tx_ctx.ctrl_len, 1094 qp->tx_ctx.bytes_unsent); 1095 rv = 0; 1096 goto done; 1097 } else if (rv == -EINPROGRESS) { 1098 rv = siw_sq_start(qp); 1099 goto done; 1100 } else { 1101 /* 1102 * WQE processing failed. 1103 * Verbs 8.3.2: 1104 * o It turns any WQE into a signalled WQE. 1105 * o Local catastrophic error must be surfaced 1106 * o QP must be moved into Terminate state: done by code 1107 * doing socket state change processing 1108 * 1109 * o TODO: Termination message must be sent. 1110 * o TODO: Implement more precise work completion errors, 1111 * see enum ib_wc_status in ib_verbs.h 1112 */ 1113 siw_dbg_qp(qp, "wqe type %d processing failed: %d\n", 1114 tx_type(wqe), rv); 1115 1116 spin_lock_irqsave(&qp->sq_lock, flags); 1117 /* 1118 * RREQ may have already been completed by inbound RRESP! 1119 */ 1120 if ((tx_type == SIW_OP_READ || 1121 tx_type == SIW_OP_READ_LOCAL_INV) && qp->attrs.orq_size) { 1122 /* Cleanup pending entry in ORQ */ 1123 qp->orq_put--; 1124 qp->orq[qp->orq_put % qp->attrs.orq_size].flags = 0; 1125 } 1126 spin_unlock_irqrestore(&qp->sq_lock, flags); 1127 /* 1128 * immediately suspends further TX processing 1129 */ 1130 if (!qp->tx_ctx.tx_suspend) 1131 siw_qp_cm_drop(qp, 0); 1132 1133 switch (tx_type) { 1134 case SIW_OP_SEND: 1135 case SIW_OP_SEND_REMOTE_INV: 1136 case SIW_OP_SEND_WITH_IMM: 1137 case SIW_OP_WRITE: 1138 case SIW_OP_READ: 1139 case SIW_OP_READ_LOCAL_INV: 1140 siw_wqe_put_mem(wqe, tx_type); 1141 fallthrough; 1142 1143 case SIW_OP_INVAL_STAG: 1144 case SIW_OP_REG_MR: 1145 siw_sqe_complete(qp, &wqe->sqe, wqe->bytes, 1146 SIW_WC_LOC_QP_OP_ERR); 1147 1148 siw_qp_event(qp, IB_EVENT_QP_FATAL); 1149 1150 break; 1151 1152 case SIW_OP_READ_RESPONSE: 1153 siw_dbg_qp(qp, "proc. read.response failed: %d\n", rv); 1154 1155 siw_qp_event(qp, IB_EVENT_QP_REQ_ERR); 1156 1157 siw_wqe_put_mem(wqe, SIW_OP_READ_RESPONSE); 1158 1159 break; 1160 1161 default: 1162 WARN(1, "undefined WQE type %d\n", tx_type); 1163 rv = -EINVAL; 1164 } 1165 wqe->wr_status = SIW_WR_IDLE; 1166 } 1167 done: 1168 return rv; 1169 } 1170 1171 static void siw_sq_resume(struct siw_qp *qp) 1172 { 1173 if (down_read_trylock(&qp->state_lock)) { 1174 if (likely(qp->attrs.state == SIW_QP_STATE_RTS && 1175 !qp->tx_ctx.tx_suspend)) { 1176 int rv = siw_qp_sq_process(qp); 1177 1178 up_read(&qp->state_lock); 1179 1180 if (unlikely(rv < 0)) { 1181 siw_dbg_qp(qp, "SQ task failed: err %d\n", rv); 1182 1183 if (!qp->tx_ctx.tx_suspend) 1184 siw_qp_cm_drop(qp, 0); 1185 } 1186 } else { 1187 up_read(&qp->state_lock); 1188 } 1189 } else { 1190 siw_dbg_qp(qp, "Resume SQ while QP locked\n"); 1191 } 1192 siw_qp_put(qp); 1193 } 1194 1195 struct tx_task_t { 1196 struct llist_head active; 1197 wait_queue_head_t waiting; 1198 }; 1199 1200 static DEFINE_PER_CPU(struct tx_task_t, siw_tx_task_g); 1201 1202 void siw_stop_tx_thread(int nr_cpu) 1203 { 1204 kthread_stop(siw_tx_thread[nr_cpu]); 1205 wake_up(&per_cpu(siw_tx_task_g, nr_cpu).waiting); 1206 } 1207 1208 int siw_run_sq(void *data) 1209 { 1210 const int nr_cpu = (unsigned int)(long)data; 1211 struct llist_node *active; 1212 struct siw_qp *qp; 1213 struct tx_task_t *tx_task = &per_cpu(siw_tx_task_g, nr_cpu); 1214 1215 init_llist_head(&tx_task->active); 1216 init_waitqueue_head(&tx_task->waiting); 1217 1218 while (1) { 1219 struct llist_node *fifo_list = NULL; 1220 1221 wait_event_interruptible(tx_task->waiting, 1222 !llist_empty(&tx_task->active) || 1223 kthread_should_stop()); 1224 1225 if (kthread_should_stop()) 1226 break; 1227 1228 active = llist_del_all(&tx_task->active); 1229 /* 1230 * llist_del_all returns a list with newest entry first. 1231 * Re-order list for fairness among QP's. 1232 */ 1233 while (active) { 1234 struct llist_node *tmp = active; 1235 1236 active = llist_next(active); 1237 tmp->next = fifo_list; 1238 fifo_list = tmp; 1239 } 1240 while (fifo_list) { 1241 qp = container_of(fifo_list, struct siw_qp, tx_list); 1242 fifo_list = llist_next(fifo_list); 1243 qp->tx_list.next = NULL; 1244 1245 siw_sq_resume(qp); 1246 } 1247 } 1248 active = llist_del_all(&tx_task->active); 1249 if (active) { 1250 llist_for_each_entry(qp, active, tx_list) { 1251 qp->tx_list.next = NULL; 1252 siw_sq_resume(qp); 1253 } 1254 } 1255 return 0; 1256 } 1257 1258 int siw_sq_start(struct siw_qp *qp) 1259 { 1260 if (tx_wqe(qp)->wr_status == SIW_WR_IDLE) 1261 return 0; 1262 1263 if (unlikely(!cpu_online(qp->tx_cpu))) { 1264 siw_put_tx_cpu(qp->tx_cpu); 1265 qp->tx_cpu = siw_get_tx_cpu(qp->sdev); 1266 if (qp->tx_cpu < 0) { 1267 pr_warn("siw: no tx cpu available\n"); 1268 1269 return -EIO; 1270 } 1271 } 1272 siw_qp_get(qp); 1273 1274 llist_add(&qp->tx_list, &per_cpu(siw_tx_task_g, qp->tx_cpu).active); 1275 1276 wake_up(&per_cpu(siw_tx_task_g, qp->tx_cpu).waiting); 1277 1278 return 0; 1279 } 1280