1 /*- 2 * Copyright (c) 2012 Chelsio Communications, Inc. 3 * All rights reserved. 4 * Written by: Navdeep Parhar <np@FreeBSD.org> 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 1. Redistributions of source code must retain the above copyright 10 * notice, this list of conditions and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25 * SUCH DAMAGE. 26 */ 27 28 #include <sys/cdefs.h> 29 __FBSDID("$FreeBSD$"); 30 31 #include "opt_inet.h" 32 33 #include <sys/param.h> 34 #include <sys/types.h> 35 #include <sys/systm.h> 36 #include <sys/kernel.h> 37 #include <sys/ktr.h> 38 #include <sys/module.h> 39 #include <sys/protosw.h> 40 #include <sys/proc.h> 41 #include <sys/domain.h> 42 #include <sys/socket.h> 43 #include <sys/socketvar.h> 44 #include <sys/uio.h> 45 #include <netinet/in.h> 46 #include <netinet/in_pcb.h> 47 #include <netinet/ip.h> 48 #include <netinet/tcp_var.h> 49 #define TCPSTATES 50 #include <netinet/tcp_fsm.h> 51 #include <netinet/toecore.h> 52 53 #include <vm/vm.h> 54 #include <vm/vm_extern.h> 55 #include <vm/vm_param.h> 56 #include <vm/pmap.h> 57 #include <vm/vm_map.h> 58 #include <vm/vm_page.h> 59 #include <vm/vm_object.h> 60 61 #ifdef TCP_OFFLOAD 62 #include "common/common.h" 63 #include "common/t4_msg.h" 64 #include "common/t4_regs.h" 65 #include "common/t4_tcb.h" 66 #include "tom/t4_tom.h" 67 68 VNET_DECLARE(int, tcp_do_autorcvbuf); 69 #define V_tcp_do_autorcvbuf VNET(tcp_do_autorcvbuf) 70 VNET_DECLARE(int, tcp_autorcvbuf_inc); 71 #define V_tcp_autorcvbuf_inc VNET(tcp_autorcvbuf_inc) 72 VNET_DECLARE(int, tcp_autorcvbuf_max); 73 #define V_tcp_autorcvbuf_max VNET(tcp_autorcvbuf_max) 74 75 static struct mbuf *get_ddp_mbuf(int len); 76 77 #define PPOD_SZ(n) ((n) * sizeof(struct pagepod)) 78 #define PPOD_SIZE (PPOD_SZ(1)) 79 80 /* XXX: must match A_ULP_RX_TDDP_PSZ */ 81 static int t4_ddp_pgsz[] = {4096, 4096 << 2, 4096 << 4, 4096 << 6}; 82 83 #if 0 84 static void 85 t4_dump_tcb(struct adapter *sc, int tid) 86 { 87 uint32_t tcb_base, off, i, j; 88 89 /* Dump TCB for the tid */ 90 tcb_base = t4_read_reg(sc, A_TP_CMM_TCB_BASE); 91 t4_write_reg(sc, PCIE_MEM_ACCESS_REG(A_PCIE_MEM_ACCESS_OFFSET, 2), 92 tcb_base + tid * TCB_SIZE); 93 t4_read_reg(sc, PCIE_MEM_ACCESS_REG(A_PCIE_MEM_ACCESS_OFFSET, 2)); 94 off = 0; 95 printf("\n"); 96 for (i = 0; i < 4; i++) { 97 uint32_t buf[8]; 98 for (j = 0; j < 8; j++, off += 4) 99 buf[j] = htonl(t4_read_reg(sc, MEMWIN2_BASE + off)); 100 101 printf("%08x %08x %08x %08x %08x %08x %08x %08x\n", 102 buf[0], buf[1], buf[2], buf[3], buf[4], buf[5], buf[6], 103 buf[7]); 104 } 105 } 106 #endif 107 108 #define MAX_DDP_BUFFER_SIZE (M_TCB_RX_DDP_BUF0_LEN) 109 static int 110 alloc_ppods(struct tom_data *td, int n, u_int *ppod_addr) 111 { 112 vmem_addr_t v; 113 int rc; 114 115 MPASS(n > 0); 116 117 rc = vmem_alloc(td->ppod_arena, PPOD_SZ(n), M_NOWAIT | M_FIRSTFIT, &v); 118 *ppod_addr = (u_int)v; 119 120 return (rc); 121 } 122 123 static void 124 free_ppods(struct tom_data *td, u_int ppod_addr, int n) 125 { 126 127 MPASS(n > 0); 128 129 vmem_free(td->ppod_arena, (vmem_addr_t)ppod_addr, PPOD_SZ(n)); 130 } 131 132 static inline int 133 pages_to_nppods(int npages, int ddp_pgsz) 134 { 135 int nsegs = npages * PAGE_SIZE / ddp_pgsz; 136 137 return (howmany(nsegs, PPOD_PAGES)); 138 } 139 140 static void 141 free_ddp_buffer(struct tom_data *td, struct ddp_buffer *db) 142 { 143 144 if (db == NULL) 145 return; 146 147 if (db->pages) 148 free(db->pages, M_CXGBE); 149 150 if (db->nppods > 0) 151 free_ppods(td, db->ppod_addr, db->nppods); 152 153 free(db, M_CXGBE); 154 } 155 156 void 157 release_ddp_resources(struct toepcb *toep) 158 { 159 int i; 160 161 for (i = 0; i < nitems(toep->db); i++) { 162 if (toep->db[i] != NULL) { 163 free_ddp_buffer(toep->td, toep->db[i]); 164 toep->db[i] = NULL; 165 } 166 } 167 } 168 169 /* XXX: handle_ddp_data code duplication */ 170 void 171 insert_ddp_data(struct toepcb *toep, uint32_t n) 172 { 173 struct inpcb *inp = toep->inp; 174 struct tcpcb *tp = intotcpcb(inp); 175 struct sockbuf *sb = &inp->inp_socket->so_rcv; 176 struct mbuf *m; 177 178 INP_WLOCK_ASSERT(inp); 179 SOCKBUF_LOCK_ASSERT(sb); 180 181 m = get_ddp_mbuf(n); 182 tp->rcv_nxt += n; 183 #ifndef USE_DDP_RX_FLOW_CONTROL 184 KASSERT(tp->rcv_wnd >= n, ("%s: negative window size", __func__)); 185 tp->rcv_wnd -= n; 186 #endif 187 188 KASSERT(toep->sb_cc >= sbused(sb), 189 ("%s: sb %p has more data (%d) than last time (%d).", 190 __func__, sb, sbused(sb), toep->sb_cc)); 191 toep->rx_credits += toep->sb_cc - sbused(sb); 192 #ifdef USE_DDP_RX_FLOW_CONTROL 193 toep->rx_credits -= n; /* adjust for F_RX_FC_DDP */ 194 #endif 195 sbappendstream_locked(sb, m, 0); 196 toep->sb_cc = sbused(sb); 197 } 198 199 /* SET_TCB_FIELD sent as a ULP command looks like this */ 200 #define LEN__SET_TCB_FIELD_ULP (sizeof(struct ulp_txpkt) + \ 201 sizeof(struct ulptx_idata) + sizeof(struct cpl_set_tcb_field_core)) 202 203 /* RX_DATA_ACK sent as a ULP command looks like this */ 204 #define LEN__RX_DATA_ACK_ULP (sizeof(struct ulp_txpkt) + \ 205 sizeof(struct ulptx_idata) + sizeof(struct cpl_rx_data_ack_core)) 206 207 static inline void * 208 mk_set_tcb_field_ulp(struct ulp_txpkt *ulpmc, struct toepcb *toep, 209 uint64_t word, uint64_t mask, uint64_t val) 210 { 211 struct ulptx_idata *ulpsc; 212 struct cpl_set_tcb_field_core *req; 213 214 ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0)); 215 ulpmc->len = htobe32(howmany(LEN__SET_TCB_FIELD_ULP, 16)); 216 217 ulpsc = (struct ulptx_idata *)(ulpmc + 1); 218 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM)); 219 ulpsc->len = htobe32(sizeof(*req)); 220 221 req = (struct cpl_set_tcb_field_core *)(ulpsc + 1); 222 OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_SET_TCB_FIELD, toep->tid)); 223 req->reply_ctrl = htobe16(V_NO_REPLY(1) | 224 V_QUEUENO(toep->ofld_rxq->iq.abs_id)); 225 req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0)); 226 req->mask = htobe64(mask); 227 req->val = htobe64(val); 228 229 ulpsc = (struct ulptx_idata *)(req + 1); 230 if (LEN__SET_TCB_FIELD_ULP % 16) { 231 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP)); 232 ulpsc->len = htobe32(0); 233 return (ulpsc + 1); 234 } 235 return (ulpsc); 236 } 237 238 static inline void * 239 mk_rx_data_ack_ulp(struct ulp_txpkt *ulpmc, struct toepcb *toep) 240 { 241 struct ulptx_idata *ulpsc; 242 struct cpl_rx_data_ack_core *req; 243 244 ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0)); 245 ulpmc->len = htobe32(howmany(LEN__RX_DATA_ACK_ULP, 16)); 246 247 ulpsc = (struct ulptx_idata *)(ulpmc + 1); 248 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM)); 249 ulpsc->len = htobe32(sizeof(*req)); 250 251 req = (struct cpl_rx_data_ack_core *)(ulpsc + 1); 252 OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_RX_DATA_ACK, toep->tid)); 253 req->credit_dack = htobe32(F_RX_MODULATE_RX); 254 255 ulpsc = (struct ulptx_idata *)(req + 1); 256 if (LEN__RX_DATA_ACK_ULP % 16) { 257 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP)); 258 ulpsc->len = htobe32(0); 259 return (ulpsc + 1); 260 } 261 return (ulpsc); 262 } 263 264 static inline uint64_t 265 select_ddp_flags(struct socket *so, int flags, int db_idx) 266 { 267 uint64_t ddp_flags = V_TF_DDP_INDICATE_OUT(0); 268 int waitall = flags & MSG_WAITALL; 269 int nb = so->so_state & SS_NBIO || flags & (MSG_DONTWAIT | MSG_NBIO); 270 271 KASSERT(db_idx == 0 || db_idx == 1, 272 ("%s: bad DDP buffer index %d", __func__, db_idx)); 273 274 if (db_idx == 0) { 275 ddp_flags |= V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_ACTIVE_BUF(0); 276 if (waitall) 277 ddp_flags |= V_TF_DDP_PUSH_DISABLE_0(1); 278 else if (nb) 279 ddp_flags |= V_TF_DDP_BUF0_FLUSH(1); 280 else 281 ddp_flags |= V_TF_DDP_BUF0_FLUSH(0); 282 } else { 283 ddp_flags |= V_TF_DDP_BUF1_VALID(1) | V_TF_DDP_ACTIVE_BUF(1); 284 if (waitall) 285 ddp_flags |= V_TF_DDP_PUSH_DISABLE_1(1); 286 else if (nb) 287 ddp_flags |= V_TF_DDP_BUF1_FLUSH(1); 288 else 289 ddp_flags |= V_TF_DDP_BUF1_FLUSH(0); 290 } 291 292 return (ddp_flags); 293 } 294 295 static struct wrqe * 296 mk_update_tcb_for_ddp(struct adapter *sc, struct toepcb *toep, int db_idx, 297 int offset, uint64_t ddp_flags) 298 { 299 struct ddp_buffer *db = toep->db[db_idx]; 300 struct wrqe *wr; 301 struct work_request_hdr *wrh; 302 struct ulp_txpkt *ulpmc; 303 int len; 304 305 KASSERT(db_idx == 0 || db_idx == 1, 306 ("%s: bad DDP buffer index %d", __func__, db_idx)); 307 308 /* 309 * We'll send a compound work request that has 3 SET_TCB_FIELDs and an 310 * RX_DATA_ACK (with RX_MODULATE to speed up delivery). 311 * 312 * The work request header is 16B and always ends at a 16B boundary. 313 * The ULPTX master commands that follow must all end at 16B boundaries 314 * too so we round up the size to 16. 315 */ 316 len = sizeof(*wrh) + 3 * roundup2(LEN__SET_TCB_FIELD_ULP, 16) + 317 roundup2(LEN__RX_DATA_ACK_ULP, 16); 318 319 wr = alloc_wrqe(len, toep->ctrlq); 320 if (wr == NULL) 321 return (NULL); 322 wrh = wrtod(wr); 323 INIT_ULPTX_WRH(wrh, len, 1, 0); /* atomic */ 324 ulpmc = (struct ulp_txpkt *)(wrh + 1); 325 326 /* Write the buffer's tag */ 327 ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, 328 W_TCB_RX_DDP_BUF0_TAG + db_idx, 329 V_TCB_RX_DDP_BUF0_TAG(M_TCB_RX_DDP_BUF0_TAG), 330 V_TCB_RX_DDP_BUF0_TAG(db->tag)); 331 332 /* Update the current offset in the DDP buffer and its total length */ 333 if (db_idx == 0) 334 ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, 335 W_TCB_RX_DDP_BUF0_OFFSET, 336 V_TCB_RX_DDP_BUF0_OFFSET(M_TCB_RX_DDP_BUF0_OFFSET) | 337 V_TCB_RX_DDP_BUF0_LEN(M_TCB_RX_DDP_BUF0_LEN), 338 V_TCB_RX_DDP_BUF0_OFFSET(offset) | 339 V_TCB_RX_DDP_BUF0_LEN(db->len)); 340 else 341 ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, 342 W_TCB_RX_DDP_BUF1_OFFSET, 343 V_TCB_RX_DDP_BUF1_OFFSET(M_TCB_RX_DDP_BUF1_OFFSET) | 344 V_TCB_RX_DDP_BUF1_LEN((u64)M_TCB_RX_DDP_BUF1_LEN << 32), 345 V_TCB_RX_DDP_BUF1_OFFSET(offset) | 346 V_TCB_RX_DDP_BUF1_LEN((u64)db->len << 32)); 347 348 /* Update DDP flags */ 349 ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, W_TCB_RX_DDP_FLAGS, 350 V_TF_DDP_BUF0_FLUSH(1) | V_TF_DDP_BUF1_FLUSH(1) | 351 V_TF_DDP_PUSH_DISABLE_0(1) | V_TF_DDP_PUSH_DISABLE_1(1) | 352 V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_BUF1_VALID(1) | 353 V_TF_DDP_ACTIVE_BUF(1) | V_TF_DDP_INDICATE_OUT(1), ddp_flags); 354 355 /* Gratuitous RX_DATA_ACK with RX_MODULATE set to speed up delivery. */ 356 ulpmc = mk_rx_data_ack_ulp(ulpmc, toep); 357 358 return (wr); 359 } 360 361 static void 362 discourage_ddp(struct toepcb *toep) 363 { 364 365 if (toep->ddp_score && --toep->ddp_score == 0) { 366 toep->ddp_flags &= ~DDP_OK; 367 toep->ddp_disabled = time_uptime; 368 CTR3(KTR_CXGBE, "%s: tid %u !DDP_OK @ %u", 369 __func__, toep->tid, time_uptime); 370 } 371 } 372 373 static int 374 handle_ddp_data(struct toepcb *toep, __be32 ddp_report, __be32 rcv_nxt, int len) 375 { 376 uint32_t report = be32toh(ddp_report); 377 unsigned int db_flag; 378 struct inpcb *inp = toep->inp; 379 struct tcpcb *tp; 380 struct socket *so; 381 struct sockbuf *sb; 382 struct mbuf *m; 383 384 db_flag = report & F_DDP_BUF_IDX ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE; 385 386 if (__predict_false(!(report & F_DDP_INV))) 387 CXGBE_UNIMPLEMENTED("DDP buffer still valid"); 388 389 INP_WLOCK(inp); 390 so = inp_inpcbtosocket(inp); 391 sb = &so->so_rcv; 392 if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))) { 393 394 /* 395 * XXX: think a bit more. 396 * tcpcb probably gone, but socket should still be around 397 * because we always wait for DDP completion in soreceive no 398 * matter what. Just wake it up and let it clean up. 399 */ 400 401 CTR5(KTR_CXGBE, "%s: tid %u, seq 0x%x, len %d, inp_flags 0x%x", 402 __func__, toep->tid, be32toh(rcv_nxt), len, inp->inp_flags); 403 SOCKBUF_LOCK(sb); 404 goto wakeup; 405 } 406 407 tp = intotcpcb(inp); 408 len += be32toh(rcv_nxt) - tp->rcv_nxt; 409 tp->rcv_nxt += len; 410 tp->t_rcvtime = ticks; 411 #ifndef USE_DDP_RX_FLOW_CONTROL 412 KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__)); 413 tp->rcv_wnd -= len; 414 #endif 415 m = get_ddp_mbuf(len); 416 417 SOCKBUF_LOCK(sb); 418 if (report & F_DDP_BUF_COMPLETE) 419 toep->ddp_score = DDP_HIGH_SCORE; 420 else 421 discourage_ddp(toep); 422 423 /* receive buffer autosize */ 424 if (sb->sb_flags & SB_AUTOSIZE && 425 V_tcp_do_autorcvbuf && 426 sb->sb_hiwat < V_tcp_autorcvbuf_max && 427 len > (sbspace(sb) / 8 * 7)) { 428 unsigned int hiwat = sb->sb_hiwat; 429 unsigned int newsize = min(hiwat + V_tcp_autorcvbuf_inc, 430 V_tcp_autorcvbuf_max); 431 432 if (!sbreserve_locked(sb, newsize, so, NULL)) 433 sb->sb_flags &= ~SB_AUTOSIZE; 434 else 435 toep->rx_credits += newsize - hiwat; 436 } 437 438 KASSERT(toep->sb_cc >= sbused(sb), 439 ("%s: sb %p has more data (%d) than last time (%d).", 440 __func__, sb, sbused(sb), toep->sb_cc)); 441 toep->rx_credits += toep->sb_cc - sbused(sb); 442 #ifdef USE_DDP_RX_FLOW_CONTROL 443 toep->rx_credits -= len; /* adjust for F_RX_FC_DDP */ 444 #endif 445 sbappendstream_locked(sb, m, 0); 446 toep->sb_cc = sbused(sb); 447 wakeup: 448 KASSERT(toep->ddp_flags & db_flag, 449 ("%s: DDP buffer not active. toep %p, ddp_flags 0x%x, report 0x%x", 450 __func__, toep, toep->ddp_flags, report)); 451 toep->ddp_flags &= ~db_flag; 452 sorwakeup_locked(so); 453 SOCKBUF_UNLOCK_ASSERT(sb); 454 455 INP_WUNLOCK(inp); 456 return (0); 457 } 458 459 void 460 handle_ddp_close(struct toepcb *toep, struct tcpcb *tp, struct sockbuf *sb, 461 __be32 rcv_nxt) 462 { 463 struct mbuf *m; 464 int len; 465 466 SOCKBUF_LOCK_ASSERT(sb); 467 INP_WLOCK_ASSERT(toep->inp); 468 len = be32toh(rcv_nxt) - tp->rcv_nxt; 469 470 /* Signal handle_ddp() to break out of its sleep loop. */ 471 toep->ddp_flags &= ~(DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE); 472 if (len == 0) 473 return; 474 475 tp->rcv_nxt += len; 476 KASSERT(toep->sb_cc >= sbused(sb), 477 ("%s: sb %p has more data (%d) than last time (%d).", 478 __func__, sb, sbused(sb), toep->sb_cc)); 479 toep->rx_credits += toep->sb_cc - sbused(sb); 480 #ifdef USE_DDP_RX_FLOW_CONTROL 481 toep->rx_credits -= len; /* adjust for F_RX_FC_DDP */ 482 #endif 483 484 m = get_ddp_mbuf(len); 485 486 sbappendstream_locked(sb, m, 0); 487 toep->sb_cc = sbused(sb); 488 } 489 490 #define DDP_ERR (F_DDP_PPOD_MISMATCH | F_DDP_LLIMIT_ERR | F_DDP_ULIMIT_ERR |\ 491 F_DDP_PPOD_PARITY_ERR | F_DDP_PADDING_ERR | F_DDP_OFFSET_ERR |\ 492 F_DDP_INVALID_TAG | F_DDP_COLOR_ERR | F_DDP_TID_MISMATCH |\ 493 F_DDP_INVALID_PPOD | F_DDP_HDRCRC_ERR | F_DDP_DATACRC_ERR) 494 495 static int 496 do_rx_data_ddp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 497 { 498 struct adapter *sc = iq->adapter; 499 const struct cpl_rx_data_ddp *cpl = (const void *)(rss + 1); 500 unsigned int tid = GET_TID(cpl); 501 uint32_t vld; 502 struct toepcb *toep = lookup_tid(sc, tid); 503 struct tom_data *td = toep->td; 504 505 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 506 KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__)); 507 KASSERT(!(toep->flags & TPF_SYNQE), 508 ("%s: toep %p claims to be a synq entry", __func__, toep)); 509 510 vld = be32toh(cpl->ddpvld); 511 if (__predict_false(vld & DDP_ERR)) { 512 panic("%s: DDP error 0x%x (tid %d, toep %p)", 513 __func__, vld, tid, toep); 514 } 515 if (toep->ulp_mode == ULP_MODE_ISCSI) { 516 m = m_get(M_NOWAIT, MT_DATA); 517 if (m == NULL) 518 CXGBE_UNIMPLEMENTED("mbuf alloc failure"); 519 memcpy(mtod(m, unsigned char *), cpl, 520 sizeof(struct cpl_rx_data_ddp)); 521 if (!t4_cpl_iscsi_callback(td, toep, m, CPL_RX_DATA_DDP)) 522 return (0); 523 m_freem(m); 524 } 525 526 handle_ddp_data(toep, cpl->u.ddp_report, cpl->seq, be16toh(cpl->len)); 527 528 return (0); 529 } 530 531 static int 532 do_rx_ddp_complete(struct sge_iq *iq, const struct rss_header *rss, 533 struct mbuf *m) 534 { 535 struct adapter *sc = iq->adapter; 536 const struct cpl_rx_ddp_complete *cpl = (const void *)(rss + 1); 537 unsigned int tid = GET_TID(cpl); 538 struct toepcb *toep = lookup_tid(sc, tid); 539 540 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 541 KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__)); 542 KASSERT(!(toep->flags & TPF_SYNQE), 543 ("%s: toep %p claims to be a synq entry", __func__, toep)); 544 545 handle_ddp_data(toep, cpl->ddp_report, cpl->rcv_nxt, 0); 546 547 return (0); 548 } 549 550 void 551 enable_ddp(struct adapter *sc, struct toepcb *toep) 552 { 553 554 KASSERT((toep->ddp_flags & (DDP_ON | DDP_OK | DDP_SC_REQ)) == DDP_OK, 555 ("%s: toep %p has bad ddp_flags 0x%x", 556 __func__, toep, toep->ddp_flags)); 557 558 CTR3(KTR_CXGBE, "%s: tid %u (time %u)", 559 __func__, toep->tid, time_uptime); 560 561 toep->ddp_flags |= DDP_SC_REQ; 562 t4_set_tcb_field(sc, toep, 1, W_TCB_RX_DDP_FLAGS, 563 V_TF_DDP_OFF(1) | V_TF_DDP_INDICATE_OUT(1) | 564 V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1) | 565 V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_BUF1_VALID(1), 566 V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1)); 567 t4_set_tcb_field(sc, toep, 1, W_TCB_T_FLAGS, 568 V_TF_RCV_COALESCE_ENABLE(1), 0); 569 } 570 571 static inline void 572 disable_ddp(struct adapter *sc, struct toepcb *toep) 573 { 574 575 KASSERT((toep->ddp_flags & (DDP_ON | DDP_SC_REQ)) == DDP_ON, 576 ("%s: toep %p has bad ddp_flags 0x%x", 577 __func__, toep, toep->ddp_flags)); 578 579 CTR3(KTR_CXGBE, "%s: tid %u (time %u)", 580 __func__, toep->tid, time_uptime); 581 582 toep->ddp_flags |= DDP_SC_REQ; 583 t4_set_tcb_field(sc, toep, 1, W_TCB_T_FLAGS, 584 V_TF_RCV_COALESCE_ENABLE(1), V_TF_RCV_COALESCE_ENABLE(1)); 585 t4_set_tcb_field(sc, toep, 1, W_TCB_RX_DDP_FLAGS, V_TF_DDP_OFF(1), 586 V_TF_DDP_OFF(1)); 587 } 588 589 static int 590 hold_uio(struct uio *uio, vm_page_t **ppages, int *pnpages) 591 { 592 struct vm_map *map; 593 struct iovec *iov; 594 vm_offset_t start, end; 595 vm_page_t *pp; 596 int n; 597 598 KASSERT(uio->uio_iovcnt == 1, 599 ("%s: uio_iovcnt %d", __func__, uio->uio_iovcnt)); 600 KASSERT(uio->uio_td->td_proc == curproc, 601 ("%s: uio proc (%p) is not curproc (%p)", 602 __func__, uio->uio_td->td_proc, curproc)); 603 604 map = &curproc->p_vmspace->vm_map; 605 iov = &uio->uio_iov[0]; 606 start = trunc_page((uintptr_t)iov->iov_base); 607 end = round_page((vm_offset_t)iov->iov_base + iov->iov_len); 608 n = howmany(end - start, PAGE_SIZE); 609 610 if (end - start > MAX_DDP_BUFFER_SIZE) 611 return (E2BIG); 612 613 pp = malloc(n * sizeof(vm_page_t), M_CXGBE, M_NOWAIT); 614 if (pp == NULL) 615 return (ENOMEM); 616 617 if (vm_fault_quick_hold_pages(map, (vm_offset_t)iov->iov_base, 618 iov->iov_len, VM_PROT_WRITE, pp, n) < 0) { 619 free(pp, M_CXGBE); 620 return (EFAULT); 621 } 622 623 *ppages = pp; 624 *pnpages = n; 625 626 return (0); 627 } 628 629 static int 630 bufcmp(struct ddp_buffer *db, vm_page_t *pages, int npages, int offset, int len) 631 { 632 int i; 633 634 if (db == NULL || db->npages != npages || db->offset != offset || 635 db->len != len) 636 return (1); 637 638 for (i = 0; i < npages; i++) { 639 if (pages[i]->phys_addr != db->pages[i]->phys_addr) 640 return (1); 641 } 642 643 return (0); 644 } 645 646 static int 647 calculate_hcf(int n1, int n2) 648 { 649 int a, b, t; 650 651 if (n1 <= n2) { 652 a = n1; 653 b = n2; 654 } else { 655 a = n2; 656 b = n1; 657 } 658 659 while (a != 0) { 660 t = a; 661 a = b % a; 662 b = t; 663 } 664 665 return (b); 666 } 667 668 static struct ddp_buffer * 669 alloc_ddp_buffer(struct tom_data *td, vm_page_t *pages, int npages, int offset, 670 int len) 671 { 672 int i, hcf, seglen, idx, ppod, nppods; 673 struct ddp_buffer *db; 674 675 /* 676 * The DDP page size is unrelated to the VM page size. We combine 677 * contiguous physical pages into larger segments to get the best DDP 678 * page size possible. This is the largest of the four sizes in 679 * A_ULP_RX_TDDP_PSZ that evenly divides the HCF of the segment sizes in 680 * the page list. 681 */ 682 hcf = 0; 683 for (i = 0; i < npages; i++) { 684 seglen = PAGE_SIZE; 685 while (i < npages - 1 && 686 pages[i]->phys_addr + PAGE_SIZE == pages[i + 1]->phys_addr) { 687 seglen += PAGE_SIZE; 688 i++; 689 } 690 691 hcf = calculate_hcf(hcf, seglen); 692 if (hcf < t4_ddp_pgsz[1]) { 693 idx = 0; 694 goto have_pgsz; /* give up, short circuit */ 695 } 696 } 697 698 if (hcf % t4_ddp_pgsz[0] != 0) { 699 /* hmmm. This could only happen when PAGE_SIZE < 4K */ 700 KASSERT(PAGE_SIZE < 4096, 701 ("%s: PAGE_SIZE %d, hcf %d", __func__, PAGE_SIZE, hcf)); 702 CTR3(KTR_CXGBE, "%s: PAGE_SIZE %d, hcf %d", 703 __func__, PAGE_SIZE, hcf); 704 return (NULL); 705 } 706 707 for (idx = nitems(t4_ddp_pgsz) - 1; idx > 0; idx--) { 708 if (hcf % t4_ddp_pgsz[idx] == 0) 709 break; 710 } 711 have_pgsz: 712 MPASS(idx <= M_PPOD_PGSZ); 713 714 db = malloc(sizeof(*db), M_CXGBE, M_NOWAIT); 715 if (db == NULL) { 716 CTR1(KTR_CXGBE, "%s: malloc failed.", __func__); 717 return (NULL); 718 } 719 720 nppods = pages_to_nppods(npages, t4_ddp_pgsz[idx]); 721 if (alloc_ppods(td, nppods, &db->ppod_addr) != 0) { 722 free(db, M_CXGBE); 723 CTR4(KTR_CXGBE, "%s: no pods, nppods %d, resid %d, pgsz %d", 724 __func__, nppods, len, t4_ddp_pgsz[idx]); 725 return (NULL); 726 } 727 ppod = (db->ppod_addr - td->ppod_start) / PPOD_SIZE; 728 729 db->tag = V_PPOD_PGSZ(idx) | V_PPOD_TAG(ppod); 730 db->nppods = nppods; 731 db->npages = npages; 732 db->pages = pages; 733 db->offset = offset; 734 db->len = len; 735 736 CTR6(KTR_CXGBE, "New DDP buffer. " 737 "ddp_pgsz %d, ppod 0x%x, npages %d, nppods %d, offset %d, len %d", 738 t4_ddp_pgsz[idx], ppod, db->npages, db->nppods, db->offset, 739 db->len); 740 741 return (db); 742 } 743 744 #define NUM_ULP_TX_SC_IMM_PPODS (256 / PPOD_SIZE) 745 746 static int 747 write_page_pods(struct adapter *sc, struct toepcb *toep, struct ddp_buffer *db) 748 { 749 struct wrqe *wr; 750 struct ulp_mem_io *ulpmc; 751 struct ulptx_idata *ulpsc; 752 struct pagepod *ppod; 753 int i, j, k, n, chunk, len, ddp_pgsz, idx; 754 u_int ppod_addr; 755 uint32_t cmd; 756 757 cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE)); 758 if (is_t4(sc)) 759 cmd |= htobe32(F_ULP_MEMIO_ORDER); 760 else 761 cmd |= htobe32(F_T5_ULP_MEMIO_IMM); 762 ddp_pgsz = t4_ddp_pgsz[G_PPOD_PGSZ(db->tag)]; 763 ppod_addr = db->ppod_addr; 764 for (i = 0; i < db->nppods; ppod_addr += chunk) { 765 766 /* How many page pods are we writing in this cycle */ 767 n = min(db->nppods - i, NUM_ULP_TX_SC_IMM_PPODS); 768 chunk = PPOD_SZ(n); 769 len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16); 770 771 wr = alloc_wrqe(len, toep->ctrlq); 772 if (wr == NULL) 773 return (ENOMEM); /* ok to just bail out */ 774 ulpmc = wrtod(wr); 775 776 INIT_ULPTX_WR(ulpmc, len, 0, 0); 777 ulpmc->cmd = cmd; 778 ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32)); 779 ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16)); 780 ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5)); 781 782 ulpsc = (struct ulptx_idata *)(ulpmc + 1); 783 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM)); 784 ulpsc->len = htobe32(chunk); 785 786 ppod = (struct pagepod *)(ulpsc + 1); 787 for (j = 0; j < n; i++, j++, ppod++) { 788 ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID | 789 V_PPOD_TID(toep->tid) | db->tag); 790 ppod->len_offset = htobe64(V_PPOD_LEN(db->len) | 791 V_PPOD_OFST(db->offset)); 792 ppod->rsvd = 0; 793 idx = i * PPOD_PAGES * (ddp_pgsz / PAGE_SIZE); 794 for (k = 0; k < nitems(ppod->addr); k++) { 795 if (idx < db->npages) { 796 ppod->addr[k] = 797 htobe64(db->pages[idx]->phys_addr); 798 idx += ddp_pgsz / PAGE_SIZE; 799 } else 800 ppod->addr[k] = 0; 801 #if 0 802 CTR5(KTR_CXGBE, 803 "%s: tid %d ppod[%d]->addr[%d] = %p", 804 __func__, toep->tid, i, k, 805 htobe64(ppod->addr[k])); 806 #endif 807 } 808 809 } 810 811 t4_wrq_tx(sc, wr); 812 } 813 814 return (0); 815 } 816 817 /* 818 * Reuse, or allocate (and program the page pods for) a new DDP buffer. The 819 * "pages" array is handed over to this function and should not be used in any 820 * way by the caller after that. 821 */ 822 static int 823 select_ddp_buffer(struct adapter *sc, struct toepcb *toep, vm_page_t *pages, 824 int npages, int db_off, int db_len) 825 { 826 struct ddp_buffer *db; 827 struct tom_data *td = sc->tom_softc; 828 int i, empty_slot = -1; 829 830 /* Try to reuse */ 831 for (i = 0; i < nitems(toep->db); i++) { 832 if (bufcmp(toep->db[i], pages, npages, db_off, db_len) == 0) { 833 free(pages, M_CXGBE); 834 return (i); /* pages still held */ 835 } else if (toep->db[i] == NULL && empty_slot < 0) 836 empty_slot = i; 837 } 838 839 /* Allocate new buffer, write its page pods. */ 840 db = alloc_ddp_buffer(td, pages, npages, db_off, db_len); 841 if (db == NULL) { 842 vm_page_unhold_pages(pages, npages); 843 free(pages, M_CXGBE); 844 return (-1); 845 } 846 if (write_page_pods(sc, toep, db) != 0) { 847 vm_page_unhold_pages(pages, npages); 848 free_ddp_buffer(td, db); 849 return (-1); 850 } 851 852 i = empty_slot; 853 if (i < 0) { 854 i = arc4random() % nitems(toep->db); 855 free_ddp_buffer(td, toep->db[i]); 856 } 857 toep->db[i] = db; 858 859 CTR5(KTR_CXGBE, "%s: tid %d, DDP buffer[%d] = %p (tag 0x%x)", 860 __func__, toep->tid, i, db, db->tag); 861 862 return (i); 863 } 864 865 static void 866 wire_ddp_buffer(struct ddp_buffer *db) 867 { 868 int i; 869 vm_page_t p; 870 871 for (i = 0; i < db->npages; i++) { 872 p = db->pages[i]; 873 vm_page_lock(p); 874 vm_page_wire(p); 875 vm_page_unhold(p); 876 vm_page_unlock(p); 877 } 878 } 879 880 static void 881 unwire_ddp_buffer(struct ddp_buffer *db) 882 { 883 int i; 884 vm_page_t p; 885 886 for (i = 0; i < db->npages; i++) { 887 p = db->pages[i]; 888 vm_page_lock(p); 889 vm_page_unwire(p, PQ_INACTIVE); 890 vm_page_unlock(p); 891 } 892 } 893 894 static int 895 handle_ddp(struct socket *so, struct uio *uio, int flags, int error) 896 { 897 struct sockbuf *sb = &so->so_rcv; 898 struct tcpcb *tp = so_sototcpcb(so); 899 struct toepcb *toep = tp->t_toe; 900 struct adapter *sc = td_adapter(toep->td); 901 vm_page_t *pages; 902 int npages, db_idx, rc, buf_flag; 903 struct ddp_buffer *db; 904 struct wrqe *wr; 905 uint64_t ddp_flags; 906 907 SOCKBUF_LOCK_ASSERT(sb); 908 909 #if 0 910 if (sbused(sb) + sc->tt.ddp_thres > uio->uio_resid) { 911 CTR4(KTR_CXGBE, "%s: sb_cc %d, threshold %d, resid %d", 912 __func__, sbused(sb), sc->tt.ddp_thres, uio->uio_resid); 913 } 914 #endif 915 916 /* XXX: too eager to disable DDP, could handle NBIO better than this. */ 917 if (sbused(sb) >= uio->uio_resid || uio->uio_resid < sc->tt.ddp_thres || 918 uio->uio_resid > MAX_DDP_BUFFER_SIZE || uio->uio_iovcnt > 1 || 919 so->so_state & SS_NBIO || flags & (MSG_DONTWAIT | MSG_NBIO) || 920 error || so->so_error || sb->sb_state & SBS_CANTRCVMORE) 921 goto no_ddp; 922 923 /* 924 * Fault in and then hold the pages of the uio buffers. We'll wire them 925 * a bit later if everything else works out. 926 */ 927 SOCKBUF_UNLOCK(sb); 928 if (hold_uio(uio, &pages, &npages) != 0) { 929 SOCKBUF_LOCK(sb); 930 goto no_ddp; 931 } 932 SOCKBUF_LOCK(sb); 933 if (__predict_false(so->so_error || sb->sb_state & SBS_CANTRCVMORE)) { 934 vm_page_unhold_pages(pages, npages); 935 free(pages, M_CXGBE); 936 goto no_ddp; 937 } 938 939 /* 940 * Figure out which one of the two DDP buffers to use this time. 941 */ 942 db_idx = select_ddp_buffer(sc, toep, pages, npages, 943 (uintptr_t)uio->uio_iov->iov_base & PAGE_MASK, uio->uio_resid); 944 pages = NULL; /* handed off to select_ddp_buffer */ 945 if (db_idx < 0) 946 goto no_ddp; 947 db = toep->db[db_idx]; 948 buf_flag = db_idx == 0 ? DDP_BUF0_ACTIVE : DDP_BUF1_ACTIVE; 949 950 /* 951 * Build the compound work request that tells the chip where to DMA the 952 * payload. 953 */ 954 ddp_flags = select_ddp_flags(so, flags, db_idx); 955 wr = mk_update_tcb_for_ddp(sc, toep, db_idx, sbused(sb), ddp_flags); 956 if (wr == NULL) { 957 /* 958 * Just unhold the pages. The DDP buffer's software state is 959 * left as-is in the toep. The page pods were written 960 * successfully and we may have an opportunity to use it in the 961 * future. 962 */ 963 vm_page_unhold_pages(db->pages, db->npages); 964 goto no_ddp; 965 } 966 967 /* Wire (and then unhold) the pages, and give the chip the go-ahead. */ 968 wire_ddp_buffer(db); 969 t4_wrq_tx(sc, wr); 970 sb->sb_flags &= ~SB_DDP_INDICATE; 971 toep->ddp_flags |= buf_flag; 972 973 /* 974 * Wait for the DDP operation to complete and then unwire the pages. 975 * The return code from the sbwait will be the final return code of this 976 * function. But we do need to wait for DDP no matter what. 977 */ 978 rc = sbwait(sb); 979 while (toep->ddp_flags & buf_flag) { 980 /* XXXGL: shouldn't here be sbwait() call? */ 981 sb->sb_flags |= SB_WAIT; 982 msleep(&sb->sb_acc, &sb->sb_mtx, PSOCK , "sbwait", 0); 983 } 984 unwire_ddp_buffer(db); 985 return (rc); 986 no_ddp: 987 disable_ddp(sc, toep); 988 discourage_ddp(toep); 989 sb->sb_flags &= ~SB_DDP_INDICATE; 990 return (0); 991 } 992 993 void 994 t4_init_ddp(struct adapter *sc, struct tom_data *td) 995 { 996 997 td->ppod_start = sc->vres.ddp.start; 998 td->ppod_arena = vmem_create("DDP page pods", sc->vres.ddp.start, 999 sc->vres.ddp.size, 1, 32, M_FIRSTFIT | M_NOWAIT); 1000 1001 t4_register_cpl_handler(sc, CPL_RX_DATA_DDP, do_rx_data_ddp); 1002 t4_register_cpl_handler(sc, CPL_RX_DDP_COMPLETE, do_rx_ddp_complete); 1003 } 1004 1005 void 1006 t4_uninit_ddp(struct adapter *sc __unused, struct tom_data *td) 1007 { 1008 1009 if (td->ppod_arena != NULL) { 1010 vmem_destroy(td->ppod_arena); 1011 td->ppod_arena = NULL; 1012 } 1013 } 1014 1015 #define VNET_SO_ASSERT(so) \ 1016 VNET_ASSERT(curvnet != NULL, \ 1017 ("%s:%d curvnet is NULL, so=%p", __func__, __LINE__, (so))); 1018 #define SBLOCKWAIT(f) (((f) & MSG_DONTWAIT) ? 0 : SBL_WAIT) 1019 static int 1020 soreceive_rcvoob(struct socket *so, struct uio *uio, int flags) 1021 { 1022 1023 CXGBE_UNIMPLEMENTED(__func__); 1024 } 1025 1026 static char ddp_magic_str[] = "nothing to see here"; 1027 1028 static struct mbuf * 1029 get_ddp_mbuf(int len) 1030 { 1031 struct mbuf *m; 1032 1033 m = m_get(M_NOWAIT, MT_DATA); 1034 if (m == NULL) 1035 CXGBE_UNIMPLEMENTED("mbuf alloc failure"); 1036 m->m_len = len; 1037 m->m_data = &ddp_magic_str[0]; 1038 1039 return (m); 1040 } 1041 1042 static inline int 1043 is_ddp_mbuf(struct mbuf *m) 1044 { 1045 1046 return (m->m_data == &ddp_magic_str[0]); 1047 } 1048 1049 /* 1050 * Copy an mbuf chain into a uio limited by len if set. 1051 */ 1052 static int 1053 m_mbuftouio_ddp(struct uio *uio, struct mbuf *m, int len) 1054 { 1055 int error, length, total; 1056 int progress = 0; 1057 1058 if (len > 0) 1059 total = min(uio->uio_resid, len); 1060 else 1061 total = uio->uio_resid; 1062 1063 /* Fill the uio with data from the mbufs. */ 1064 for (; m != NULL; m = m->m_next) { 1065 length = min(m->m_len, total - progress); 1066 1067 if (is_ddp_mbuf(m)) { 1068 enum uio_seg segflag = uio->uio_segflg; 1069 1070 uio->uio_segflg = UIO_NOCOPY; 1071 error = uiomove(mtod(m, void *), length, uio); 1072 uio->uio_segflg = segflag; 1073 } else 1074 error = uiomove(mtod(m, void *), length, uio); 1075 if (error) 1076 return (error); 1077 1078 progress += length; 1079 } 1080 1081 return (0); 1082 } 1083 1084 /* 1085 * Based on soreceive_stream() in uipc_socket.c 1086 */ 1087 int 1088 t4_soreceive_ddp(struct socket *so, struct sockaddr **psa, struct uio *uio, 1089 struct mbuf **mp0, struct mbuf **controlp, int *flagsp) 1090 { 1091 int len = 0, error = 0, flags, oresid, ddp_handled = 0; 1092 struct sockbuf *sb; 1093 struct mbuf *m, *n = NULL; 1094 1095 /* We only do stream sockets. */ 1096 if (so->so_type != SOCK_STREAM) 1097 return (EINVAL); 1098 if (psa != NULL) 1099 *psa = NULL; 1100 if (controlp != NULL) 1101 return (EINVAL); 1102 if (flagsp != NULL) 1103 flags = *flagsp &~ MSG_EOR; 1104 else 1105 flags = 0; 1106 if (flags & MSG_OOB) 1107 return (soreceive_rcvoob(so, uio, flags)); 1108 if (mp0 != NULL) 1109 *mp0 = NULL; 1110 1111 sb = &so->so_rcv; 1112 1113 /* Prevent other readers from entering the socket. */ 1114 error = sblock(sb, SBLOCKWAIT(flags)); 1115 SOCKBUF_LOCK(sb); 1116 if (error) 1117 goto out; 1118 1119 /* Easy one, no space to copyout anything. */ 1120 if (uio->uio_resid == 0) { 1121 error = EINVAL; 1122 goto out; 1123 } 1124 oresid = uio->uio_resid; 1125 1126 /* We will never ever get anything unless we are or were connected. */ 1127 if (!(so->so_state & (SS_ISCONNECTED|SS_ISDISCONNECTED))) { 1128 error = ENOTCONN; 1129 goto out; 1130 } 1131 1132 restart: 1133 SOCKBUF_LOCK_ASSERT(&so->so_rcv); 1134 1135 if (sb->sb_flags & SB_DDP_INDICATE && !ddp_handled) { 1136 1137 /* uio should be just as it was at entry */ 1138 KASSERT(oresid == uio->uio_resid, 1139 ("%s: oresid = %d, uio_resid = %zd, sbavail = %d", 1140 __func__, oresid, uio->uio_resid, sbavail(sb))); 1141 1142 error = handle_ddp(so, uio, flags, 0); 1143 ddp_handled = 1; 1144 if (error) 1145 goto out; 1146 } 1147 1148 /* Abort if socket has reported problems. */ 1149 if (so->so_error) { 1150 if (sbavail(sb)) 1151 goto deliver; 1152 if (oresid > uio->uio_resid) 1153 goto out; 1154 error = so->so_error; 1155 if (!(flags & MSG_PEEK)) 1156 so->so_error = 0; 1157 goto out; 1158 } 1159 1160 /* Door is closed. Deliver what is left, if any. */ 1161 if (sb->sb_state & SBS_CANTRCVMORE) { 1162 if (sbavail(sb)) 1163 goto deliver; 1164 else 1165 goto out; 1166 } 1167 1168 /* Socket buffer is empty and we shall not block. */ 1169 if (sbavail(sb) == 0 && 1170 ((so->so_state & SS_NBIO) || (flags & (MSG_DONTWAIT|MSG_NBIO)))) { 1171 error = EAGAIN; 1172 goto out; 1173 } 1174 1175 /* Socket buffer got some data that we shall deliver now. */ 1176 if (sbavail(sb) > 0 && !(flags & MSG_WAITALL) && 1177 ((so->so_state & SS_NBIO) || 1178 (flags & (MSG_DONTWAIT|MSG_NBIO)) || 1179 sbavail(sb) >= sb->sb_lowat || 1180 sbavail(sb) >= uio->uio_resid || 1181 sbavail(sb) >= sb->sb_hiwat) ) { 1182 goto deliver; 1183 } 1184 1185 /* On MSG_WAITALL we must wait until all data or error arrives. */ 1186 if ((flags & MSG_WAITALL) && 1187 (sbavail(sb) >= uio->uio_resid || sbavail(sb) >= sb->sb_lowat)) 1188 goto deliver; 1189 1190 /* 1191 * Wait and block until (more) data comes in. 1192 * NB: Drops the sockbuf lock during wait. 1193 */ 1194 error = sbwait(sb); 1195 if (error) { 1196 if (sb->sb_flags & SB_DDP_INDICATE && !ddp_handled) { 1197 (void) handle_ddp(so, uio, flags, 1); 1198 ddp_handled = 1; 1199 } 1200 goto out; 1201 } 1202 goto restart; 1203 1204 deliver: 1205 SOCKBUF_LOCK_ASSERT(&so->so_rcv); 1206 KASSERT(sbavail(sb) > 0, ("%s: sockbuf empty", __func__)); 1207 KASSERT(sb->sb_mb != NULL, ("%s: sb_mb == NULL", __func__)); 1208 1209 if (sb->sb_flags & SB_DDP_INDICATE && !ddp_handled) 1210 goto restart; 1211 1212 /* Statistics. */ 1213 if (uio->uio_td) 1214 uio->uio_td->td_ru.ru_msgrcv++; 1215 1216 /* Fill uio until full or current end of socket buffer is reached. */ 1217 len = min(uio->uio_resid, sbavail(sb)); 1218 if (mp0 != NULL) { 1219 /* Dequeue as many mbufs as possible. */ 1220 if (!(flags & MSG_PEEK) && len >= sb->sb_mb->m_len) { 1221 for (*mp0 = m = sb->sb_mb; 1222 m != NULL && m->m_len <= len; 1223 m = m->m_next) { 1224 len -= m->m_len; 1225 uio->uio_resid -= m->m_len; 1226 sbfree(sb, m); 1227 n = m; 1228 } 1229 sb->sb_mb = m; 1230 if (sb->sb_mb == NULL) 1231 SB_EMPTY_FIXUP(sb); 1232 n->m_next = NULL; 1233 } 1234 /* Copy the remainder. */ 1235 if (len > 0) { 1236 KASSERT(sb->sb_mb != NULL, 1237 ("%s: len > 0 && sb->sb_mb empty", __func__)); 1238 1239 m = m_copym(sb->sb_mb, 0, len, M_NOWAIT); 1240 if (m == NULL) 1241 len = 0; /* Don't flush data from sockbuf. */ 1242 else 1243 uio->uio_resid -= m->m_len; 1244 if (*mp0 != NULL) 1245 n->m_next = m; 1246 else 1247 *mp0 = m; 1248 if (*mp0 == NULL) { 1249 error = ENOBUFS; 1250 goto out; 1251 } 1252 } 1253 } else { 1254 /* NB: Must unlock socket buffer as uiomove may sleep. */ 1255 SOCKBUF_UNLOCK(sb); 1256 error = m_mbuftouio_ddp(uio, sb->sb_mb, len); 1257 SOCKBUF_LOCK(sb); 1258 if (error) 1259 goto out; 1260 } 1261 SBLASTRECORDCHK(sb); 1262 SBLASTMBUFCHK(sb); 1263 1264 /* 1265 * Remove the delivered data from the socket buffer unless we 1266 * were only peeking. 1267 */ 1268 if (!(flags & MSG_PEEK)) { 1269 if (len > 0) 1270 sbdrop_locked(sb, len); 1271 1272 /* Notify protocol that we drained some data. */ 1273 if ((so->so_proto->pr_flags & PR_WANTRCVD) && 1274 (((flags & MSG_WAITALL) && uio->uio_resid > 0) || 1275 !(flags & MSG_SOCALLBCK))) { 1276 SOCKBUF_UNLOCK(sb); 1277 VNET_SO_ASSERT(so); 1278 (*so->so_proto->pr_usrreqs->pru_rcvd)(so, flags); 1279 SOCKBUF_LOCK(sb); 1280 } 1281 } 1282 1283 /* 1284 * For MSG_WAITALL we may have to loop again and wait for 1285 * more data to come in. 1286 */ 1287 if ((flags & MSG_WAITALL) && uio->uio_resid > 0) 1288 goto restart; 1289 out: 1290 SOCKBUF_LOCK_ASSERT(sb); 1291 SBLASTRECORDCHK(sb); 1292 SBLASTMBUFCHK(sb); 1293 SOCKBUF_UNLOCK(sb); 1294 sbunlock(sb); 1295 return (error); 1296 } 1297 1298 #endif 1299