1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2012 Chelsio Communications, Inc. 5 * All rights reserved. 6 * Written by: Navdeep Parhar <np@FreeBSD.org> 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27 * SUCH DAMAGE. 28 */ 29 30 #include <sys/cdefs.h> 31 __FBSDID("$FreeBSD$"); 32 33 #include "opt_inet.h" 34 35 #include <sys/param.h> 36 #include <sys/aio.h> 37 #include <sys/file.h> 38 #include <sys/systm.h> 39 #include <sys/kernel.h> 40 #include <sys/ktr.h> 41 #include <sys/module.h> 42 #include <sys/protosw.h> 43 #include <sys/proc.h> 44 #include <sys/domain.h> 45 #include <sys/socket.h> 46 #include <sys/socketvar.h> 47 #include <sys/taskqueue.h> 48 #include <sys/uio.h> 49 #include <netinet/in.h> 50 #include <netinet/in_pcb.h> 51 #include <netinet/ip.h> 52 #include <netinet/tcp_var.h> 53 #define TCPSTATES 54 #include <netinet/tcp_fsm.h> 55 #include <netinet/toecore.h> 56 57 #include <vm/vm.h> 58 #include <vm/vm_extern.h> 59 #include <vm/vm_param.h> 60 #include <vm/pmap.h> 61 #include <vm/vm_map.h> 62 #include <vm/vm_page.h> 63 #include <vm/vm_object.h> 64 65 #ifdef TCP_OFFLOAD 66 #include "common/common.h" 67 #include "common/t4_msg.h" 68 #include "common/t4_regs.h" 69 #include "common/t4_tcb.h" 70 #include "tom/t4_tom.h" 71 72 /* 73 * Use the 'backend3' field in AIO jobs to store the amount of data 74 * received by the AIO job so far. 75 */ 76 #define aio_received backend3 77 78 static void aio_ddp_requeue_task(void *context, int pending); 79 static void ddp_complete_all(struct toepcb *toep, int error); 80 static void t4_aio_cancel_active(struct kaiocb *job); 81 static void t4_aio_cancel_queued(struct kaiocb *job); 82 83 static TAILQ_HEAD(, pageset) ddp_orphan_pagesets; 84 static struct mtx ddp_orphan_pagesets_lock; 85 static struct task ddp_orphan_task; 86 87 #define MAX_DDP_BUFFER_SIZE (M_TCB_RX_DDP_BUF0_LEN) 88 89 /* 90 * A page set holds information about a buffer used for DDP. The page 91 * set holds resources such as the VM pages backing the buffer (either 92 * held or wired) and the page pods associated with the buffer. 93 * Recently used page sets are cached to allow for efficient reuse of 94 * buffers (avoiding the need to re-fault in pages, hold them, etc.). 95 * Note that cached page sets keep the backing pages wired. The 96 * number of wired pages is capped by only allowing for two wired 97 * pagesets per connection. This is not a perfect cap, but is a 98 * trade-off for performance. 99 * 100 * If an application ping-pongs two buffers for a connection via 101 * aio_read(2) then those buffers should remain wired and expensive VM 102 * fault lookups should be avoided after each buffer has been used 103 * once. If an application uses more than two buffers then this will 104 * fall back to doing expensive VM fault lookups for each operation. 105 */ 106 static void 107 free_pageset(struct tom_data *td, struct pageset *ps) 108 { 109 vm_page_t p; 110 int i; 111 112 if (ps->prsv.prsv_nppods > 0) 113 t4_free_page_pods(&ps->prsv); 114 115 if (ps->flags & PS_WIRED) { 116 for (i = 0; i < ps->npages; i++) { 117 p = ps->pages[i]; 118 vm_page_lock(p); 119 vm_page_unwire(p, PQ_INACTIVE); 120 vm_page_unlock(p); 121 } 122 } else 123 vm_page_unhold_pages(ps->pages, ps->npages); 124 mtx_lock(&ddp_orphan_pagesets_lock); 125 TAILQ_INSERT_TAIL(&ddp_orphan_pagesets, ps, link); 126 taskqueue_enqueue(taskqueue_thread, &ddp_orphan_task); 127 mtx_unlock(&ddp_orphan_pagesets_lock); 128 } 129 130 static void 131 ddp_free_orphan_pagesets(void *context, int pending) 132 { 133 struct pageset *ps; 134 135 mtx_lock(&ddp_orphan_pagesets_lock); 136 while (!TAILQ_EMPTY(&ddp_orphan_pagesets)) { 137 ps = TAILQ_FIRST(&ddp_orphan_pagesets); 138 TAILQ_REMOVE(&ddp_orphan_pagesets, ps, link); 139 mtx_unlock(&ddp_orphan_pagesets_lock); 140 if (ps->vm) 141 vmspace_free(ps->vm); 142 free(ps, M_CXGBE); 143 mtx_lock(&ddp_orphan_pagesets_lock); 144 } 145 mtx_unlock(&ddp_orphan_pagesets_lock); 146 } 147 148 static void 149 recycle_pageset(struct toepcb *toep, struct pageset *ps) 150 { 151 152 DDP_ASSERT_LOCKED(toep); 153 if (!(toep->ddp.flags & DDP_DEAD) && ps->flags & PS_WIRED) { 154 KASSERT(toep->ddp.cached_count + toep->ddp.active_count < 155 nitems(toep->ddp.db), ("too many wired pagesets")); 156 TAILQ_INSERT_HEAD(&toep->ddp.cached_pagesets, ps, link); 157 toep->ddp.cached_count++; 158 } else 159 free_pageset(toep->td, ps); 160 } 161 162 static void 163 ddp_complete_one(struct kaiocb *job, int error) 164 { 165 long copied; 166 167 /* 168 * If this job had copied data out of the socket buffer before 169 * it was cancelled, report it as a short read rather than an 170 * error. 171 */ 172 copied = job->aio_received; 173 if (copied != 0 || error == 0) 174 aio_complete(job, copied, 0); 175 else 176 aio_complete(job, -1, error); 177 } 178 179 static void 180 free_ddp_buffer(struct tom_data *td, struct ddp_buffer *db) 181 { 182 183 if (db->job) { 184 /* 185 * XXX: If we are un-offloading the socket then we 186 * should requeue these on the socket somehow. If we 187 * got a FIN from the remote end, then this completes 188 * any remaining requests with an EOF read. 189 */ 190 if (!aio_clear_cancel_function(db->job)) 191 ddp_complete_one(db->job, 0); 192 } 193 194 if (db->ps) 195 free_pageset(td, db->ps); 196 } 197 198 void 199 ddp_init_toep(struct toepcb *toep) 200 { 201 202 TAILQ_INIT(&toep->ddp.aiojobq); 203 TASK_INIT(&toep->ddp.requeue_task, 0, aio_ddp_requeue_task, toep); 204 toep->ddp.flags = DDP_OK; 205 toep->ddp.active_id = -1; 206 mtx_init(&toep->ddp.lock, "t4 ddp", NULL, MTX_DEF); 207 } 208 209 void 210 ddp_uninit_toep(struct toepcb *toep) 211 { 212 213 mtx_destroy(&toep->ddp.lock); 214 } 215 216 void 217 release_ddp_resources(struct toepcb *toep) 218 { 219 struct pageset *ps; 220 int i; 221 222 DDP_LOCK(toep); 223 toep->flags |= DDP_DEAD; 224 for (i = 0; i < nitems(toep->ddp.db); i++) { 225 free_ddp_buffer(toep->td, &toep->ddp.db[i]); 226 } 227 while ((ps = TAILQ_FIRST(&toep->ddp.cached_pagesets)) != NULL) { 228 TAILQ_REMOVE(&toep->ddp.cached_pagesets, ps, link); 229 free_pageset(toep->td, ps); 230 } 231 ddp_complete_all(toep, 0); 232 DDP_UNLOCK(toep); 233 } 234 235 #ifdef INVARIANTS 236 void 237 ddp_assert_empty(struct toepcb *toep) 238 { 239 int i; 240 241 MPASS(!(toep->ddp.flags & DDP_TASK_ACTIVE)); 242 for (i = 0; i < nitems(toep->ddp.db); i++) { 243 MPASS(toep->ddp.db[i].job == NULL); 244 MPASS(toep->ddp.db[i].ps == NULL); 245 } 246 MPASS(TAILQ_EMPTY(&toep->ddp.cached_pagesets)); 247 MPASS(TAILQ_EMPTY(&toep->ddp.aiojobq)); 248 } 249 #endif 250 251 static void 252 complete_ddp_buffer(struct toepcb *toep, struct ddp_buffer *db, 253 unsigned int db_idx) 254 { 255 unsigned int db_flag; 256 257 toep->ddp.active_count--; 258 if (toep->ddp.active_id == db_idx) { 259 if (toep->ddp.active_count == 0) { 260 KASSERT(toep->ddp.db[db_idx ^ 1].job == NULL, 261 ("%s: active_count mismatch", __func__)); 262 toep->ddp.active_id = -1; 263 } else 264 toep->ddp.active_id ^= 1; 265 #ifdef VERBOSE_TRACES 266 CTR2(KTR_CXGBE, "%s: ddp_active_id = %d", __func__, 267 toep->ddp.active_id); 268 #endif 269 } else { 270 KASSERT(toep->ddp.active_count != 0 && 271 toep->ddp.active_id != -1, 272 ("%s: active count mismatch", __func__)); 273 } 274 275 db->cancel_pending = 0; 276 db->job = NULL; 277 recycle_pageset(toep, db->ps); 278 db->ps = NULL; 279 280 db_flag = db_idx == 1 ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE; 281 KASSERT(toep->ddp.flags & db_flag, 282 ("%s: DDP buffer not active. toep %p, ddp_flags 0x%x", 283 __func__, toep, toep->ddp.flags)); 284 toep->ddp.flags &= ~db_flag; 285 } 286 287 /* XXX: handle_ddp_data code duplication */ 288 void 289 insert_ddp_data(struct toepcb *toep, uint32_t n) 290 { 291 struct inpcb *inp = toep->inp; 292 struct tcpcb *tp = intotcpcb(inp); 293 struct ddp_buffer *db; 294 struct kaiocb *job; 295 size_t placed; 296 long copied; 297 unsigned int db_flag, db_idx; 298 299 INP_WLOCK_ASSERT(inp); 300 DDP_ASSERT_LOCKED(toep); 301 302 tp->rcv_nxt += n; 303 #ifndef USE_DDP_RX_FLOW_CONTROL 304 KASSERT(tp->rcv_wnd >= n, ("%s: negative window size", __func__)); 305 tp->rcv_wnd -= n; 306 #endif 307 #ifndef USE_DDP_RX_FLOW_CONTROL 308 toep->rx_credits += n; 309 #endif 310 CTR2(KTR_CXGBE, "%s: placed %u bytes before falling out of DDP", 311 __func__, n); 312 while (toep->ddp.active_count > 0) { 313 MPASS(toep->ddp.active_id != -1); 314 db_idx = toep->ddp.active_id; 315 db_flag = db_idx == 1 ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE; 316 MPASS((toep->ddp.flags & db_flag) != 0); 317 db = &toep->ddp.db[db_idx]; 318 job = db->job; 319 copied = job->aio_received; 320 placed = n; 321 if (placed > job->uaiocb.aio_nbytes - copied) 322 placed = job->uaiocb.aio_nbytes - copied; 323 if (placed > 0) 324 job->msgrcv = 1; 325 if (!aio_clear_cancel_function(job)) { 326 /* 327 * Update the copied length for when 328 * t4_aio_cancel_active() completes this 329 * request. 330 */ 331 job->aio_received += placed; 332 } else if (copied + placed != 0) { 333 CTR4(KTR_CXGBE, 334 "%s: completing %p (copied %ld, placed %lu)", 335 __func__, job, copied, placed); 336 /* XXX: This always completes if there is some data. */ 337 aio_complete(job, copied + placed, 0); 338 } else if (aio_set_cancel_function(job, t4_aio_cancel_queued)) { 339 TAILQ_INSERT_HEAD(&toep->ddp.aiojobq, job, list); 340 toep->ddp.waiting_count++; 341 } else 342 aio_cancel(job); 343 n -= placed; 344 complete_ddp_buffer(toep, db, db_idx); 345 } 346 347 MPASS(n == 0); 348 } 349 350 /* SET_TCB_FIELD sent as a ULP command looks like this */ 351 #define LEN__SET_TCB_FIELD_ULP (sizeof(struct ulp_txpkt) + \ 352 sizeof(struct ulptx_idata) + sizeof(struct cpl_set_tcb_field_core)) 353 354 /* RX_DATA_ACK sent as a ULP command looks like this */ 355 #define LEN__RX_DATA_ACK_ULP (sizeof(struct ulp_txpkt) + \ 356 sizeof(struct ulptx_idata) + sizeof(struct cpl_rx_data_ack_core)) 357 358 static inline void * 359 mk_set_tcb_field_ulp(struct ulp_txpkt *ulpmc, struct toepcb *toep, 360 uint64_t word, uint64_t mask, uint64_t val) 361 { 362 struct ulptx_idata *ulpsc; 363 struct cpl_set_tcb_field_core *req; 364 365 ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0)); 366 ulpmc->len = htobe32(howmany(LEN__SET_TCB_FIELD_ULP, 16)); 367 368 ulpsc = (struct ulptx_idata *)(ulpmc + 1); 369 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM)); 370 ulpsc->len = htobe32(sizeof(*req)); 371 372 req = (struct cpl_set_tcb_field_core *)(ulpsc + 1); 373 OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_SET_TCB_FIELD, toep->tid)); 374 req->reply_ctrl = htobe16(V_NO_REPLY(1) | 375 V_QUEUENO(toep->ofld_rxq->iq.abs_id)); 376 req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0)); 377 req->mask = htobe64(mask); 378 req->val = htobe64(val); 379 380 ulpsc = (struct ulptx_idata *)(req + 1); 381 if (LEN__SET_TCB_FIELD_ULP % 16) { 382 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP)); 383 ulpsc->len = htobe32(0); 384 return (ulpsc + 1); 385 } 386 return (ulpsc); 387 } 388 389 static inline void * 390 mk_rx_data_ack_ulp(struct ulp_txpkt *ulpmc, struct toepcb *toep) 391 { 392 struct ulptx_idata *ulpsc; 393 struct cpl_rx_data_ack_core *req; 394 395 ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0)); 396 ulpmc->len = htobe32(howmany(LEN__RX_DATA_ACK_ULP, 16)); 397 398 ulpsc = (struct ulptx_idata *)(ulpmc + 1); 399 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM)); 400 ulpsc->len = htobe32(sizeof(*req)); 401 402 req = (struct cpl_rx_data_ack_core *)(ulpsc + 1); 403 OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_RX_DATA_ACK, toep->tid)); 404 req->credit_dack = htobe32(F_RX_MODULATE_RX); 405 406 ulpsc = (struct ulptx_idata *)(req + 1); 407 if (LEN__RX_DATA_ACK_ULP % 16) { 408 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP)); 409 ulpsc->len = htobe32(0); 410 return (ulpsc + 1); 411 } 412 return (ulpsc); 413 } 414 415 static struct wrqe * 416 mk_update_tcb_for_ddp(struct adapter *sc, struct toepcb *toep, int db_idx, 417 struct pageset *ps, int offset, uint64_t ddp_flags, uint64_t ddp_flags_mask) 418 { 419 struct wrqe *wr; 420 struct work_request_hdr *wrh; 421 struct ulp_txpkt *ulpmc; 422 int len; 423 424 KASSERT(db_idx == 0 || db_idx == 1, 425 ("%s: bad DDP buffer index %d", __func__, db_idx)); 426 427 /* 428 * We'll send a compound work request that has 3 SET_TCB_FIELDs and an 429 * RX_DATA_ACK (with RX_MODULATE to speed up delivery). 430 * 431 * The work request header is 16B and always ends at a 16B boundary. 432 * The ULPTX master commands that follow must all end at 16B boundaries 433 * too so we round up the size to 16. 434 */ 435 len = sizeof(*wrh) + 3 * roundup2(LEN__SET_TCB_FIELD_ULP, 16) + 436 roundup2(LEN__RX_DATA_ACK_ULP, 16); 437 438 wr = alloc_wrqe(len, toep->ctrlq); 439 if (wr == NULL) 440 return (NULL); 441 wrh = wrtod(wr); 442 INIT_ULPTX_WRH(wrh, len, 1, 0); /* atomic */ 443 ulpmc = (struct ulp_txpkt *)(wrh + 1); 444 445 /* Write the buffer's tag */ 446 ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, 447 W_TCB_RX_DDP_BUF0_TAG + db_idx, 448 V_TCB_RX_DDP_BUF0_TAG(M_TCB_RX_DDP_BUF0_TAG), 449 V_TCB_RX_DDP_BUF0_TAG(ps->prsv.prsv_tag)); 450 451 /* Update the current offset in the DDP buffer and its total length */ 452 if (db_idx == 0) 453 ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, 454 W_TCB_RX_DDP_BUF0_OFFSET, 455 V_TCB_RX_DDP_BUF0_OFFSET(M_TCB_RX_DDP_BUF0_OFFSET) | 456 V_TCB_RX_DDP_BUF0_LEN(M_TCB_RX_DDP_BUF0_LEN), 457 V_TCB_RX_DDP_BUF0_OFFSET(offset) | 458 V_TCB_RX_DDP_BUF0_LEN(ps->len)); 459 else 460 ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, 461 W_TCB_RX_DDP_BUF1_OFFSET, 462 V_TCB_RX_DDP_BUF1_OFFSET(M_TCB_RX_DDP_BUF1_OFFSET) | 463 V_TCB_RX_DDP_BUF1_LEN((u64)M_TCB_RX_DDP_BUF1_LEN << 32), 464 V_TCB_RX_DDP_BUF1_OFFSET(offset) | 465 V_TCB_RX_DDP_BUF1_LEN((u64)ps->len << 32)); 466 467 /* Update DDP flags */ 468 ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, W_TCB_RX_DDP_FLAGS, 469 ddp_flags_mask, ddp_flags); 470 471 /* Gratuitous RX_DATA_ACK with RX_MODULATE set to speed up delivery. */ 472 ulpmc = mk_rx_data_ack_ulp(ulpmc, toep); 473 474 return (wr); 475 } 476 477 static int 478 handle_ddp_data(struct toepcb *toep, __be32 ddp_report, __be32 rcv_nxt, int len) 479 { 480 uint32_t report = be32toh(ddp_report); 481 unsigned int db_idx; 482 struct inpcb *inp = toep->inp; 483 struct ddp_buffer *db; 484 struct tcpcb *tp; 485 struct socket *so; 486 struct sockbuf *sb; 487 struct kaiocb *job; 488 long copied; 489 490 db_idx = report & F_DDP_BUF_IDX ? 1 : 0; 491 492 if (__predict_false(!(report & F_DDP_INV))) 493 CXGBE_UNIMPLEMENTED("DDP buffer still valid"); 494 495 INP_WLOCK(inp); 496 so = inp_inpcbtosocket(inp); 497 sb = &so->so_rcv; 498 DDP_LOCK(toep); 499 500 KASSERT(toep->ddp.active_id == db_idx, 501 ("completed DDP buffer (%d) != active_id (%d) for tid %d", db_idx, 502 toep->ddp.active_id, toep->tid)); 503 db = &toep->ddp.db[db_idx]; 504 job = db->job; 505 506 if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))) { 507 /* 508 * This can happen due to an administrative tcpdrop(8). 509 * Just fail the request with ECONNRESET. 510 */ 511 CTR5(KTR_CXGBE, "%s: tid %u, seq 0x%x, len %d, inp_flags 0x%x", 512 __func__, toep->tid, be32toh(rcv_nxt), len, inp->inp_flags); 513 if (aio_clear_cancel_function(job)) 514 ddp_complete_one(job, ECONNRESET); 515 goto completed; 516 } 517 518 tp = intotcpcb(inp); 519 520 /* 521 * For RX_DDP_COMPLETE, len will be zero and rcv_nxt is the 522 * sequence number of the next byte to receive. The length of 523 * the data received for this message must be computed by 524 * comparing the new and old values of rcv_nxt. 525 * 526 * For RX_DATA_DDP, len might be non-zero, but it is only the 527 * length of the most recent DMA. It does not include the 528 * total length of the data received since the previous update 529 * for this DDP buffer. rcv_nxt is the sequence number of the 530 * first received byte from the most recent DMA. 531 */ 532 len += be32toh(rcv_nxt) - tp->rcv_nxt; 533 tp->rcv_nxt += len; 534 tp->t_rcvtime = ticks; 535 #ifndef USE_DDP_RX_FLOW_CONTROL 536 KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__)); 537 tp->rcv_wnd -= len; 538 #endif 539 #ifdef VERBOSE_TRACES 540 CTR4(KTR_CXGBE, "%s: DDP[%d] placed %d bytes (%#x)", __func__, db_idx, 541 len, report); 542 #endif 543 544 /* receive buffer autosize */ 545 MPASS(toep->vnet == so->so_vnet); 546 CURVNET_SET(toep->vnet); 547 SOCKBUF_LOCK(sb); 548 if (sb->sb_flags & SB_AUTOSIZE && 549 V_tcp_do_autorcvbuf && 550 sb->sb_hiwat < V_tcp_autorcvbuf_max && 551 len > (sbspace(sb) / 8 * 7)) { 552 unsigned int hiwat = sb->sb_hiwat; 553 unsigned int newsize = min(hiwat + V_tcp_autorcvbuf_inc, 554 V_tcp_autorcvbuf_max); 555 556 if (!sbreserve_locked(sb, newsize, so, NULL)) 557 sb->sb_flags &= ~SB_AUTOSIZE; 558 else 559 toep->rx_credits += newsize - hiwat; 560 } 561 SOCKBUF_UNLOCK(sb); 562 CURVNET_RESTORE(); 563 564 #ifndef USE_DDP_RX_FLOW_CONTROL 565 toep->rx_credits += len; 566 #endif 567 568 job->msgrcv = 1; 569 if (db->cancel_pending) { 570 /* 571 * Update the job's length but defer completion to the 572 * TCB_RPL callback. 573 */ 574 job->aio_received += len; 575 goto out; 576 } else if (!aio_clear_cancel_function(job)) { 577 /* 578 * Update the copied length for when 579 * t4_aio_cancel_active() completes this request. 580 */ 581 job->aio_received += len; 582 } else { 583 copied = job->aio_received; 584 #ifdef VERBOSE_TRACES 585 CTR4(KTR_CXGBE, "%s: completing %p (copied %ld, placed %d)", 586 __func__, job, copied, len); 587 #endif 588 aio_complete(job, copied + len, 0); 589 t4_rcvd(&toep->td->tod, tp); 590 } 591 592 completed: 593 complete_ddp_buffer(toep, db, db_idx); 594 if (toep->ddp.waiting_count > 0) 595 ddp_queue_toep(toep); 596 out: 597 DDP_UNLOCK(toep); 598 INP_WUNLOCK(inp); 599 600 return (0); 601 } 602 603 void 604 handle_ddp_indicate(struct toepcb *toep) 605 { 606 607 DDP_ASSERT_LOCKED(toep); 608 MPASS(toep->ddp.active_count == 0); 609 MPASS((toep->ddp.flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE)) == 0); 610 if (toep->ddp.waiting_count == 0) { 611 /* 612 * The pending requests that triggered the request for an 613 * an indicate were cancelled. Those cancels should have 614 * already disabled DDP. Just ignore this as the data is 615 * going into the socket buffer anyway. 616 */ 617 return; 618 } 619 CTR3(KTR_CXGBE, "%s: tid %d indicated (%d waiting)", __func__, 620 toep->tid, toep->ddp.waiting_count); 621 ddp_queue_toep(toep); 622 } 623 624 enum { 625 DDP_BUF0_INVALIDATED = 0x2, 626 DDP_BUF1_INVALIDATED 627 }; 628 629 CTASSERT(DDP_BUF0_INVALIDATED == CPL_COOKIE_DDP0); 630 631 static int 632 do_ddp_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 633 { 634 struct adapter *sc = iq->adapter; 635 const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1); 636 unsigned int tid = GET_TID(cpl); 637 unsigned int db_idx; 638 struct toepcb *toep; 639 struct inpcb *inp; 640 struct ddp_buffer *db; 641 struct kaiocb *job; 642 long copied; 643 644 if (cpl->status != CPL_ERR_NONE) 645 panic("XXX: tcp_rpl failed: %d", cpl->status); 646 647 toep = lookup_tid(sc, tid); 648 inp = toep->inp; 649 switch (cpl->cookie) { 650 case V_WORD(W_TCB_RX_DDP_FLAGS) | V_COOKIE(DDP_BUF0_INVALIDATED): 651 case V_WORD(W_TCB_RX_DDP_FLAGS) | V_COOKIE(DDP_BUF1_INVALIDATED): 652 /* 653 * XXX: This duplicates a lot of code with handle_ddp_data(). 654 */ 655 db_idx = G_COOKIE(cpl->cookie) - DDP_BUF0_INVALIDATED; 656 MPASS(db_idx < nitems(toep->ddp.db)); 657 INP_WLOCK(inp); 658 DDP_LOCK(toep); 659 db = &toep->ddp.db[db_idx]; 660 661 /* 662 * handle_ddp_data() should leave the job around until 663 * this callback runs once a cancel is pending. 664 */ 665 MPASS(db != NULL); 666 MPASS(db->job != NULL); 667 MPASS(db->cancel_pending); 668 669 /* 670 * XXX: It's not clear what happens if there is data 671 * placed when the buffer is invalidated. I suspect we 672 * need to read the TCB to see how much data was placed. 673 * 674 * For now this just pretends like nothing was placed. 675 * 676 * XXX: Note that if we did check the PCB we would need to 677 * also take care of updating the tp, etc. 678 */ 679 job = db->job; 680 copied = job->aio_received; 681 if (copied == 0) { 682 CTR2(KTR_CXGBE, "%s: cancelling %p", __func__, job); 683 aio_cancel(job); 684 } else { 685 CTR3(KTR_CXGBE, "%s: completing %p (copied %ld)", 686 __func__, job, copied); 687 aio_complete(job, copied, 0); 688 t4_rcvd(&toep->td->tod, intotcpcb(inp)); 689 } 690 691 complete_ddp_buffer(toep, db, db_idx); 692 if (toep->ddp.waiting_count > 0) 693 ddp_queue_toep(toep); 694 DDP_UNLOCK(toep); 695 INP_WUNLOCK(inp); 696 break; 697 default: 698 panic("XXX: unknown tcb_rpl offset %#x, cookie %#x", 699 G_WORD(cpl->cookie), G_COOKIE(cpl->cookie)); 700 } 701 702 return (0); 703 } 704 705 void 706 handle_ddp_close(struct toepcb *toep, struct tcpcb *tp, __be32 rcv_nxt) 707 { 708 struct ddp_buffer *db; 709 struct kaiocb *job; 710 long copied; 711 unsigned int db_flag, db_idx; 712 int len, placed; 713 714 INP_WLOCK_ASSERT(toep->inp); 715 DDP_ASSERT_LOCKED(toep); 716 len = be32toh(rcv_nxt) - tp->rcv_nxt; 717 718 tp->rcv_nxt += len; 719 #ifndef USE_DDP_RX_FLOW_CONTROL 720 toep->rx_credits += len; 721 #endif 722 723 while (toep->ddp.active_count > 0) { 724 MPASS(toep->ddp.active_id != -1); 725 db_idx = toep->ddp.active_id; 726 db_flag = db_idx == 1 ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE; 727 MPASS((toep->ddp.flags & db_flag) != 0); 728 db = &toep->ddp.db[db_idx]; 729 job = db->job; 730 copied = job->aio_received; 731 placed = len; 732 if (placed > job->uaiocb.aio_nbytes - copied) 733 placed = job->uaiocb.aio_nbytes - copied; 734 if (placed > 0) 735 job->msgrcv = 1; 736 if (!aio_clear_cancel_function(job)) { 737 /* 738 * Update the copied length for when 739 * t4_aio_cancel_active() completes this 740 * request. 741 */ 742 job->aio_received += placed; 743 } else { 744 CTR4(KTR_CXGBE, "%s: tid %d completed buf %d len %d", 745 __func__, toep->tid, db_idx, placed); 746 aio_complete(job, copied + placed, 0); 747 } 748 len -= placed; 749 complete_ddp_buffer(toep, db, db_idx); 750 } 751 752 MPASS(len == 0); 753 ddp_complete_all(toep, 0); 754 } 755 756 #define DDP_ERR (F_DDP_PPOD_MISMATCH | F_DDP_LLIMIT_ERR | F_DDP_ULIMIT_ERR |\ 757 F_DDP_PPOD_PARITY_ERR | F_DDP_PADDING_ERR | F_DDP_OFFSET_ERR |\ 758 F_DDP_INVALID_TAG | F_DDP_COLOR_ERR | F_DDP_TID_MISMATCH |\ 759 F_DDP_INVALID_PPOD | F_DDP_HDRCRC_ERR | F_DDP_DATACRC_ERR) 760 761 extern cpl_handler_t t4_cpl_handler[]; 762 763 static int 764 do_rx_data_ddp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 765 { 766 struct adapter *sc = iq->adapter; 767 const struct cpl_rx_data_ddp *cpl = (const void *)(rss + 1); 768 unsigned int tid = GET_TID(cpl); 769 uint32_t vld; 770 struct toepcb *toep = lookup_tid(sc, tid); 771 772 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 773 KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__)); 774 KASSERT(!(toep->flags & TPF_SYNQE), 775 ("%s: toep %p claims to be a synq entry", __func__, toep)); 776 777 vld = be32toh(cpl->ddpvld); 778 if (__predict_false(vld & DDP_ERR)) { 779 panic("%s: DDP error 0x%x (tid %d, toep %p)", 780 __func__, vld, tid, toep); 781 } 782 783 if (toep->ulp_mode == ULP_MODE_ISCSI) { 784 t4_cpl_handler[CPL_RX_ISCSI_DDP](iq, rss, m); 785 return (0); 786 } 787 788 handle_ddp_data(toep, cpl->u.ddp_report, cpl->seq, be16toh(cpl->len)); 789 790 return (0); 791 } 792 793 static int 794 do_rx_ddp_complete(struct sge_iq *iq, const struct rss_header *rss, 795 struct mbuf *m) 796 { 797 struct adapter *sc = iq->adapter; 798 const struct cpl_rx_ddp_complete *cpl = (const void *)(rss + 1); 799 unsigned int tid = GET_TID(cpl); 800 struct toepcb *toep = lookup_tid(sc, tid); 801 802 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 803 KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__)); 804 KASSERT(!(toep->flags & TPF_SYNQE), 805 ("%s: toep %p claims to be a synq entry", __func__, toep)); 806 807 handle_ddp_data(toep, cpl->ddp_report, cpl->rcv_nxt, 0); 808 809 return (0); 810 } 811 812 static void 813 enable_ddp(struct adapter *sc, struct toepcb *toep) 814 { 815 816 KASSERT((toep->ddp.flags & (DDP_ON | DDP_OK | DDP_SC_REQ)) == DDP_OK, 817 ("%s: toep %p has bad ddp_flags 0x%x", 818 __func__, toep, toep->ddp.flags)); 819 820 CTR3(KTR_CXGBE, "%s: tid %u (time %u)", 821 __func__, toep->tid, time_uptime); 822 823 DDP_ASSERT_LOCKED(toep); 824 toep->ddp.flags |= DDP_SC_REQ; 825 t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_RX_DDP_FLAGS, 826 V_TF_DDP_OFF(1) | V_TF_DDP_INDICATE_OUT(1) | 827 V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1) | 828 V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_BUF1_VALID(1), 829 V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1), 0, 0); 830 t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_T_FLAGS, 831 V_TF_RCV_COALESCE_ENABLE(1), 0, 0, 0); 832 } 833 834 static int 835 calculate_hcf(int n1, int n2) 836 { 837 int a, b, t; 838 839 if (n1 <= n2) { 840 a = n1; 841 b = n2; 842 } else { 843 a = n2; 844 b = n1; 845 } 846 847 while (a != 0) { 848 t = a; 849 a = b % a; 850 b = t; 851 } 852 853 return (b); 854 } 855 856 static inline int 857 pages_to_nppods(int npages, int ddp_page_shift) 858 { 859 860 MPASS(ddp_page_shift >= PAGE_SHIFT); 861 862 return (howmany(npages >> (ddp_page_shift - PAGE_SHIFT), PPOD_PAGES)); 863 } 864 865 static int 866 alloc_page_pods(struct ppod_region *pr, u_int nppods, u_int pgsz_idx, 867 struct ppod_reservation *prsv) 868 { 869 vmem_addr_t addr; /* relative to start of region */ 870 871 if (vmem_alloc(pr->pr_arena, PPOD_SZ(nppods), M_NOWAIT | M_FIRSTFIT, 872 &addr) != 0) 873 return (ENOMEM); 874 875 CTR5(KTR_CXGBE, "%-17s arena %p, addr 0x%08x, nppods %d, pgsz %d", 876 __func__, pr->pr_arena, (uint32_t)addr & pr->pr_tag_mask, 877 nppods, 1 << pr->pr_page_shift[pgsz_idx]); 878 879 /* 880 * The hardware tagmask includes an extra invalid bit but the arena was 881 * seeded with valid values only. An allocation out of this arena will 882 * fit inside the tagmask but won't have the invalid bit set. 883 */ 884 MPASS((addr & pr->pr_tag_mask) == addr); 885 MPASS((addr & pr->pr_invalid_bit) == 0); 886 887 prsv->prsv_pr = pr; 888 prsv->prsv_tag = V_PPOD_PGSZ(pgsz_idx) | addr; 889 prsv->prsv_nppods = nppods; 890 891 return (0); 892 } 893 894 int 895 t4_alloc_page_pods_for_ps(struct ppod_region *pr, struct pageset *ps) 896 { 897 int i, hcf, seglen, idx, nppods; 898 struct ppod_reservation *prsv = &ps->prsv; 899 900 KASSERT(prsv->prsv_nppods == 0, 901 ("%s: page pods already allocated", __func__)); 902 903 /* 904 * The DDP page size is unrelated to the VM page size. We combine 905 * contiguous physical pages into larger segments to get the best DDP 906 * page size possible. This is the largest of the four sizes in 907 * A_ULP_RX_TDDP_PSZ that evenly divides the HCF of the segment sizes in 908 * the page list. 909 */ 910 hcf = 0; 911 for (i = 0; i < ps->npages; i++) { 912 seglen = PAGE_SIZE; 913 while (i < ps->npages - 1 && 914 ps->pages[i]->phys_addr + PAGE_SIZE == 915 ps->pages[i + 1]->phys_addr) { 916 seglen += PAGE_SIZE; 917 i++; 918 } 919 920 hcf = calculate_hcf(hcf, seglen); 921 if (hcf < (1 << pr->pr_page_shift[1])) { 922 idx = 0; 923 goto have_pgsz; /* give up, short circuit */ 924 } 925 } 926 927 #define PR_PAGE_MASK(x) ((1 << pr->pr_page_shift[(x)]) - 1) 928 MPASS((hcf & PR_PAGE_MASK(0)) == 0); /* PAGE_SIZE is >= 4K everywhere */ 929 for (idx = nitems(pr->pr_page_shift) - 1; idx > 0; idx--) { 930 if ((hcf & PR_PAGE_MASK(idx)) == 0) 931 break; 932 } 933 #undef PR_PAGE_MASK 934 935 have_pgsz: 936 MPASS(idx <= M_PPOD_PGSZ); 937 938 nppods = pages_to_nppods(ps->npages, pr->pr_page_shift[idx]); 939 if (alloc_page_pods(pr, nppods, idx, prsv) != 0) 940 return (0); 941 MPASS(prsv->prsv_nppods > 0); 942 943 return (1); 944 } 945 946 int 947 t4_alloc_page_pods_for_buf(struct ppod_region *pr, vm_offset_t buf, int len, 948 struct ppod_reservation *prsv) 949 { 950 int hcf, seglen, idx, npages, nppods; 951 uintptr_t start_pva, end_pva, pva, p1; 952 953 MPASS(buf > 0); 954 MPASS(len > 0); 955 956 /* 957 * The DDP page size is unrelated to the VM page size. We combine 958 * contiguous physical pages into larger segments to get the best DDP 959 * page size possible. This is the largest of the four sizes in 960 * A_ULP_RX_ISCSI_PSZ that evenly divides the HCF of the segment sizes 961 * in the page list. 962 */ 963 hcf = 0; 964 start_pva = trunc_page(buf); 965 end_pva = trunc_page(buf + len - 1); 966 pva = start_pva; 967 while (pva <= end_pva) { 968 seglen = PAGE_SIZE; 969 p1 = pmap_kextract(pva); 970 pva += PAGE_SIZE; 971 while (pva <= end_pva && p1 + seglen == pmap_kextract(pva)) { 972 seglen += PAGE_SIZE; 973 pva += PAGE_SIZE; 974 } 975 976 hcf = calculate_hcf(hcf, seglen); 977 if (hcf < (1 << pr->pr_page_shift[1])) { 978 idx = 0; 979 goto have_pgsz; /* give up, short circuit */ 980 } 981 } 982 983 #define PR_PAGE_MASK(x) ((1 << pr->pr_page_shift[(x)]) - 1) 984 MPASS((hcf & PR_PAGE_MASK(0)) == 0); /* PAGE_SIZE is >= 4K everywhere */ 985 for (idx = nitems(pr->pr_page_shift) - 1; idx > 0; idx--) { 986 if ((hcf & PR_PAGE_MASK(idx)) == 0) 987 break; 988 } 989 #undef PR_PAGE_MASK 990 991 have_pgsz: 992 MPASS(idx <= M_PPOD_PGSZ); 993 994 npages = 1; 995 npages += (end_pva - start_pva) >> pr->pr_page_shift[idx]; 996 nppods = howmany(npages, PPOD_PAGES); 997 if (alloc_page_pods(pr, nppods, idx, prsv) != 0) 998 return (ENOMEM); 999 MPASS(prsv->prsv_nppods > 0); 1000 1001 return (0); 1002 } 1003 1004 void 1005 t4_free_page_pods(struct ppod_reservation *prsv) 1006 { 1007 struct ppod_region *pr = prsv->prsv_pr; 1008 vmem_addr_t addr; 1009 1010 MPASS(prsv != NULL); 1011 MPASS(prsv->prsv_nppods != 0); 1012 1013 addr = prsv->prsv_tag & pr->pr_tag_mask; 1014 MPASS((addr & pr->pr_invalid_bit) == 0); 1015 1016 CTR4(KTR_CXGBE, "%-17s arena %p, addr 0x%08x, nppods %d", __func__, 1017 pr->pr_arena, addr, prsv->prsv_nppods); 1018 1019 vmem_free(pr->pr_arena, addr, PPOD_SZ(prsv->prsv_nppods)); 1020 prsv->prsv_nppods = 0; 1021 } 1022 1023 #define NUM_ULP_TX_SC_IMM_PPODS (256 / PPOD_SIZE) 1024 1025 int 1026 t4_write_page_pods_for_ps(struct adapter *sc, struct sge_wrq *wrq, int tid, 1027 struct pageset *ps) 1028 { 1029 struct wrqe *wr; 1030 struct ulp_mem_io *ulpmc; 1031 struct ulptx_idata *ulpsc; 1032 struct pagepod *ppod; 1033 int i, j, k, n, chunk, len, ddp_pgsz, idx; 1034 u_int ppod_addr; 1035 uint32_t cmd; 1036 struct ppod_reservation *prsv = &ps->prsv; 1037 struct ppod_region *pr = prsv->prsv_pr; 1038 1039 KASSERT(!(ps->flags & PS_PPODS_WRITTEN), 1040 ("%s: page pods already written", __func__)); 1041 MPASS(prsv->prsv_nppods > 0); 1042 1043 cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE)); 1044 if (is_t4(sc)) 1045 cmd |= htobe32(F_ULP_MEMIO_ORDER); 1046 else 1047 cmd |= htobe32(F_T5_ULP_MEMIO_IMM); 1048 ddp_pgsz = 1 << pr->pr_page_shift[G_PPOD_PGSZ(prsv->prsv_tag)]; 1049 ppod_addr = pr->pr_start + (prsv->prsv_tag & pr->pr_tag_mask); 1050 for (i = 0; i < prsv->prsv_nppods; ppod_addr += chunk) { 1051 1052 /* How many page pods are we writing in this cycle */ 1053 n = min(prsv->prsv_nppods - i, NUM_ULP_TX_SC_IMM_PPODS); 1054 chunk = PPOD_SZ(n); 1055 len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16); 1056 1057 wr = alloc_wrqe(len, wrq); 1058 if (wr == NULL) 1059 return (ENOMEM); /* ok to just bail out */ 1060 ulpmc = wrtod(wr); 1061 1062 INIT_ULPTX_WR(ulpmc, len, 0, 0); 1063 ulpmc->cmd = cmd; 1064 ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32)); 1065 ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16)); 1066 ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5)); 1067 1068 ulpsc = (struct ulptx_idata *)(ulpmc + 1); 1069 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM)); 1070 ulpsc->len = htobe32(chunk); 1071 1072 ppod = (struct pagepod *)(ulpsc + 1); 1073 for (j = 0; j < n; i++, j++, ppod++) { 1074 ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID | 1075 V_PPOD_TID(tid) | prsv->prsv_tag); 1076 ppod->len_offset = htobe64(V_PPOD_LEN(ps->len) | 1077 V_PPOD_OFST(ps->offset)); 1078 ppod->rsvd = 0; 1079 idx = i * PPOD_PAGES * (ddp_pgsz / PAGE_SIZE); 1080 for (k = 0; k < nitems(ppod->addr); k++) { 1081 if (idx < ps->npages) { 1082 ppod->addr[k] = 1083 htobe64(ps->pages[idx]->phys_addr); 1084 idx += ddp_pgsz / PAGE_SIZE; 1085 } else 1086 ppod->addr[k] = 0; 1087 #if 0 1088 CTR5(KTR_CXGBE, 1089 "%s: tid %d ppod[%d]->addr[%d] = %p", 1090 __func__, toep->tid, i, k, 1091 htobe64(ppod->addr[k])); 1092 #endif 1093 } 1094 1095 } 1096 1097 t4_wrq_tx(sc, wr); 1098 } 1099 ps->flags |= PS_PPODS_WRITTEN; 1100 1101 return (0); 1102 } 1103 1104 int 1105 t4_write_page_pods_for_buf(struct adapter *sc, struct sge_wrq *wrq, int tid, 1106 struct ppod_reservation *prsv, vm_offset_t buf, int buflen) 1107 { 1108 struct wrqe *wr; 1109 struct ulp_mem_io *ulpmc; 1110 struct ulptx_idata *ulpsc; 1111 struct pagepod *ppod; 1112 int i, j, k, n, chunk, len, ddp_pgsz; 1113 u_int ppod_addr, offset; 1114 uint32_t cmd; 1115 struct ppod_region *pr = prsv->prsv_pr; 1116 uintptr_t end_pva, pva, pa; 1117 1118 cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE)); 1119 if (is_t4(sc)) 1120 cmd |= htobe32(F_ULP_MEMIO_ORDER); 1121 else 1122 cmd |= htobe32(F_T5_ULP_MEMIO_IMM); 1123 ddp_pgsz = 1 << pr->pr_page_shift[G_PPOD_PGSZ(prsv->prsv_tag)]; 1124 offset = buf & PAGE_MASK; 1125 ppod_addr = pr->pr_start + (prsv->prsv_tag & pr->pr_tag_mask); 1126 pva = trunc_page(buf); 1127 end_pva = trunc_page(buf + buflen - 1); 1128 for (i = 0; i < prsv->prsv_nppods; ppod_addr += chunk) { 1129 1130 /* How many page pods are we writing in this cycle */ 1131 n = min(prsv->prsv_nppods - i, NUM_ULP_TX_SC_IMM_PPODS); 1132 MPASS(n > 0); 1133 chunk = PPOD_SZ(n); 1134 len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16); 1135 1136 wr = alloc_wrqe(len, wrq); 1137 if (wr == NULL) 1138 return (ENOMEM); /* ok to just bail out */ 1139 ulpmc = wrtod(wr); 1140 1141 INIT_ULPTX_WR(ulpmc, len, 0, 0); 1142 ulpmc->cmd = cmd; 1143 ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32)); 1144 ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16)); 1145 ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5)); 1146 1147 ulpsc = (struct ulptx_idata *)(ulpmc + 1); 1148 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM)); 1149 ulpsc->len = htobe32(chunk); 1150 1151 ppod = (struct pagepod *)(ulpsc + 1); 1152 for (j = 0; j < n; i++, j++, ppod++) { 1153 ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID | 1154 V_PPOD_TID(tid) | 1155 (prsv->prsv_tag & ~V_PPOD_PGSZ(M_PPOD_PGSZ))); 1156 ppod->len_offset = htobe64(V_PPOD_LEN(buflen) | 1157 V_PPOD_OFST(offset)); 1158 ppod->rsvd = 0; 1159 1160 for (k = 0; k < nitems(ppod->addr); k++) { 1161 if (pva > end_pva) 1162 ppod->addr[k] = 0; 1163 else { 1164 pa = pmap_kextract(pva); 1165 ppod->addr[k] = htobe64(pa); 1166 pva += ddp_pgsz; 1167 } 1168 #if 0 1169 CTR5(KTR_CXGBE, 1170 "%s: tid %d ppod[%d]->addr[%d] = %p", 1171 __func__, tid, i, k, 1172 htobe64(ppod->addr[k])); 1173 #endif 1174 } 1175 1176 /* 1177 * Walk back 1 segment so that the first address in the 1178 * next pod is the same as the last one in the current 1179 * pod. 1180 */ 1181 pva -= ddp_pgsz; 1182 } 1183 1184 t4_wrq_tx(sc, wr); 1185 } 1186 1187 MPASS(pva <= end_pva); 1188 1189 return (0); 1190 } 1191 1192 static void 1193 wire_pageset(struct pageset *ps) 1194 { 1195 vm_page_t p; 1196 int i; 1197 1198 KASSERT(!(ps->flags & PS_WIRED), ("pageset already wired")); 1199 1200 for (i = 0; i < ps->npages; i++) { 1201 p = ps->pages[i]; 1202 vm_page_lock(p); 1203 vm_page_wire(p); 1204 vm_page_unhold(p); 1205 vm_page_unlock(p); 1206 } 1207 ps->flags |= PS_WIRED; 1208 } 1209 1210 /* 1211 * Prepare a pageset for DDP. This wires the pageset and sets up page 1212 * pods. 1213 */ 1214 static int 1215 prep_pageset(struct adapter *sc, struct toepcb *toep, struct pageset *ps) 1216 { 1217 struct tom_data *td = sc->tom_softc; 1218 1219 if (!(ps->flags & PS_WIRED)) 1220 wire_pageset(ps); 1221 if (ps->prsv.prsv_nppods == 0 && 1222 !t4_alloc_page_pods_for_ps(&td->pr, ps)) { 1223 return (0); 1224 } 1225 if (!(ps->flags & PS_PPODS_WRITTEN) && 1226 t4_write_page_pods_for_ps(sc, toep->ctrlq, toep->tid, ps) != 0) { 1227 return (0); 1228 } 1229 1230 return (1); 1231 } 1232 1233 int 1234 t4_init_ppod_region(struct ppod_region *pr, struct t4_range *r, u_int psz, 1235 const char *name) 1236 { 1237 int i; 1238 1239 MPASS(pr != NULL); 1240 MPASS(r->size > 0); 1241 1242 pr->pr_start = r->start; 1243 pr->pr_len = r->size; 1244 pr->pr_page_shift[0] = 12 + G_HPZ0(psz); 1245 pr->pr_page_shift[1] = 12 + G_HPZ1(psz); 1246 pr->pr_page_shift[2] = 12 + G_HPZ2(psz); 1247 pr->pr_page_shift[3] = 12 + G_HPZ3(psz); 1248 1249 /* The SGL -> page pod algorithm requires the sizes to be in order. */ 1250 for (i = 1; i < nitems(pr->pr_page_shift); i++) { 1251 if (pr->pr_page_shift[i] <= pr->pr_page_shift[i - 1]) 1252 return (ENXIO); 1253 } 1254 1255 pr->pr_tag_mask = ((1 << fls(r->size)) - 1) & V_PPOD_TAG(M_PPOD_TAG); 1256 pr->pr_alias_mask = V_PPOD_TAG(M_PPOD_TAG) & ~pr->pr_tag_mask; 1257 if (pr->pr_tag_mask == 0 || pr->pr_alias_mask == 0) 1258 return (ENXIO); 1259 pr->pr_alias_shift = fls(pr->pr_tag_mask); 1260 pr->pr_invalid_bit = 1 << (pr->pr_alias_shift - 1); 1261 1262 pr->pr_arena = vmem_create(name, 0, pr->pr_len, PPOD_SIZE, 0, 1263 M_FIRSTFIT | M_NOWAIT); 1264 if (pr->pr_arena == NULL) 1265 return (ENOMEM); 1266 1267 return (0); 1268 } 1269 1270 void 1271 t4_free_ppod_region(struct ppod_region *pr) 1272 { 1273 1274 MPASS(pr != NULL); 1275 1276 if (pr->pr_arena) 1277 vmem_destroy(pr->pr_arena); 1278 bzero(pr, sizeof(*pr)); 1279 } 1280 1281 static int 1282 pscmp(struct pageset *ps, struct vmspace *vm, vm_offset_t start, int npages, 1283 int pgoff, int len) 1284 { 1285 1286 if (ps->start != start || ps->npages != npages || 1287 ps->offset != pgoff || ps->len != len) 1288 return (1); 1289 1290 return (ps->vm != vm || ps->vm_timestamp != vm->vm_map.timestamp); 1291 } 1292 1293 static int 1294 hold_aio(struct toepcb *toep, struct kaiocb *job, struct pageset **pps) 1295 { 1296 struct vmspace *vm; 1297 vm_map_t map; 1298 vm_offset_t start, end, pgoff; 1299 struct pageset *ps; 1300 int n; 1301 1302 DDP_ASSERT_LOCKED(toep); 1303 1304 /* 1305 * The AIO subsystem will cancel and drain all requests before 1306 * permitting a process to exit or exec, so p_vmspace should 1307 * be stable here. 1308 */ 1309 vm = job->userproc->p_vmspace; 1310 map = &vm->vm_map; 1311 start = (uintptr_t)job->uaiocb.aio_buf; 1312 pgoff = start & PAGE_MASK; 1313 end = round_page(start + job->uaiocb.aio_nbytes); 1314 start = trunc_page(start); 1315 1316 if (end - start > MAX_DDP_BUFFER_SIZE) { 1317 /* 1318 * Truncate the request to a short read. 1319 * Alternatively, we could DDP in chunks to the larger 1320 * buffer, but that would be quite a bit more work. 1321 * 1322 * When truncating, round the request down to avoid 1323 * crossing a cache line on the final transaction. 1324 */ 1325 end = rounddown2(start + MAX_DDP_BUFFER_SIZE, CACHE_LINE_SIZE); 1326 #ifdef VERBOSE_TRACES 1327 CTR4(KTR_CXGBE, "%s: tid %d, truncating size from %lu to %lu", 1328 __func__, toep->tid, (unsigned long)job->uaiocb.aio_nbytes, 1329 (unsigned long)(end - (start + pgoff))); 1330 job->uaiocb.aio_nbytes = end - (start + pgoff); 1331 #endif 1332 end = round_page(end); 1333 } 1334 1335 n = atop(end - start); 1336 1337 /* 1338 * Try to reuse a cached pageset. 1339 */ 1340 TAILQ_FOREACH(ps, &toep->ddp.cached_pagesets, link) { 1341 if (pscmp(ps, vm, start, n, pgoff, 1342 job->uaiocb.aio_nbytes) == 0) { 1343 TAILQ_REMOVE(&toep->ddp.cached_pagesets, ps, link); 1344 toep->ddp.cached_count--; 1345 *pps = ps; 1346 return (0); 1347 } 1348 } 1349 1350 /* 1351 * If there are too many cached pagesets to create a new one, 1352 * free a pageset before creating a new one. 1353 */ 1354 KASSERT(toep->ddp.active_count + toep->ddp.cached_count <= 1355 nitems(toep->ddp.db), ("%s: too many wired pagesets", __func__)); 1356 if (toep->ddp.active_count + toep->ddp.cached_count == 1357 nitems(toep->ddp.db)) { 1358 KASSERT(toep->ddp.cached_count > 0, 1359 ("no cached pageset to free")); 1360 ps = TAILQ_LAST(&toep->ddp.cached_pagesets, pagesetq); 1361 TAILQ_REMOVE(&toep->ddp.cached_pagesets, ps, link); 1362 toep->ddp.cached_count--; 1363 free_pageset(toep->td, ps); 1364 } 1365 DDP_UNLOCK(toep); 1366 1367 /* Create a new pageset. */ 1368 ps = malloc(sizeof(*ps) + n * sizeof(vm_page_t), M_CXGBE, M_WAITOK | 1369 M_ZERO); 1370 ps->pages = (vm_page_t *)(ps + 1); 1371 ps->vm_timestamp = map->timestamp; 1372 ps->npages = vm_fault_quick_hold_pages(map, start, end - start, 1373 VM_PROT_WRITE, ps->pages, n); 1374 1375 DDP_LOCK(toep); 1376 if (ps->npages < 0) { 1377 free(ps, M_CXGBE); 1378 return (EFAULT); 1379 } 1380 1381 KASSERT(ps->npages == n, ("hold_aio: page count mismatch: %d vs %d", 1382 ps->npages, n)); 1383 1384 ps->offset = pgoff; 1385 ps->len = job->uaiocb.aio_nbytes; 1386 atomic_add_int(&vm->vm_refcnt, 1); 1387 ps->vm = vm; 1388 ps->start = start; 1389 1390 CTR5(KTR_CXGBE, "%s: tid %d, new pageset %p for job %p, npages %d", 1391 __func__, toep->tid, ps, job, ps->npages); 1392 *pps = ps; 1393 return (0); 1394 } 1395 1396 static void 1397 ddp_complete_all(struct toepcb *toep, int error) 1398 { 1399 struct kaiocb *job; 1400 1401 DDP_ASSERT_LOCKED(toep); 1402 while (!TAILQ_EMPTY(&toep->ddp.aiojobq)) { 1403 job = TAILQ_FIRST(&toep->ddp.aiojobq); 1404 TAILQ_REMOVE(&toep->ddp.aiojobq, job, list); 1405 toep->ddp.waiting_count--; 1406 if (aio_clear_cancel_function(job)) 1407 ddp_complete_one(job, error); 1408 } 1409 } 1410 1411 static void 1412 aio_ddp_cancel_one(struct kaiocb *job) 1413 { 1414 long copied; 1415 1416 /* 1417 * If this job had copied data out of the socket buffer before 1418 * it was cancelled, report it as a short read rather than an 1419 * error. 1420 */ 1421 copied = job->aio_received; 1422 if (copied != 0) 1423 aio_complete(job, copied, 0); 1424 else 1425 aio_cancel(job); 1426 } 1427 1428 /* 1429 * Called when the main loop wants to requeue a job to retry it later. 1430 * Deals with the race of the job being cancelled while it was being 1431 * examined. 1432 */ 1433 static void 1434 aio_ddp_requeue_one(struct toepcb *toep, struct kaiocb *job) 1435 { 1436 1437 DDP_ASSERT_LOCKED(toep); 1438 if (!(toep->ddp.flags & DDP_DEAD) && 1439 aio_set_cancel_function(job, t4_aio_cancel_queued)) { 1440 TAILQ_INSERT_HEAD(&toep->ddp.aiojobq, job, list); 1441 toep->ddp.waiting_count++; 1442 } else 1443 aio_ddp_cancel_one(job); 1444 } 1445 1446 static void 1447 aio_ddp_requeue(struct toepcb *toep) 1448 { 1449 struct adapter *sc = td_adapter(toep->td); 1450 struct socket *so; 1451 struct sockbuf *sb; 1452 struct inpcb *inp; 1453 struct kaiocb *job; 1454 struct ddp_buffer *db; 1455 size_t copied, offset, resid; 1456 struct pageset *ps; 1457 struct mbuf *m; 1458 uint64_t ddp_flags, ddp_flags_mask; 1459 struct wrqe *wr; 1460 int buf_flag, db_idx, error; 1461 1462 DDP_ASSERT_LOCKED(toep); 1463 1464 restart: 1465 if (toep->ddp.flags & DDP_DEAD) { 1466 MPASS(toep->ddp.waiting_count == 0); 1467 MPASS(toep->ddp.active_count == 0); 1468 return; 1469 } 1470 1471 if (toep->ddp.waiting_count == 0 || 1472 toep->ddp.active_count == nitems(toep->ddp.db)) { 1473 return; 1474 } 1475 1476 job = TAILQ_FIRST(&toep->ddp.aiojobq); 1477 so = job->fd_file->f_data; 1478 sb = &so->so_rcv; 1479 SOCKBUF_LOCK(sb); 1480 1481 /* We will never get anything unless we are or were connected. */ 1482 if (!(so->so_state & (SS_ISCONNECTED|SS_ISDISCONNECTED))) { 1483 SOCKBUF_UNLOCK(sb); 1484 ddp_complete_all(toep, ENOTCONN); 1485 return; 1486 } 1487 1488 KASSERT(toep->ddp.active_count == 0 || sbavail(sb) == 0, 1489 ("%s: pending sockbuf data and DDP is active", __func__)); 1490 1491 /* Abort if socket has reported problems. */ 1492 /* XXX: Wait for any queued DDP's to finish and/or flush them? */ 1493 if (so->so_error && sbavail(sb) == 0) { 1494 toep->ddp.waiting_count--; 1495 TAILQ_REMOVE(&toep->ddp.aiojobq, job, list); 1496 if (!aio_clear_cancel_function(job)) { 1497 SOCKBUF_UNLOCK(sb); 1498 goto restart; 1499 } 1500 1501 /* 1502 * If this job has previously copied some data, report 1503 * a short read and leave the error to be reported by 1504 * a future request. 1505 */ 1506 copied = job->aio_received; 1507 if (copied != 0) { 1508 SOCKBUF_UNLOCK(sb); 1509 aio_complete(job, copied, 0); 1510 goto restart; 1511 } 1512 error = so->so_error; 1513 so->so_error = 0; 1514 SOCKBUF_UNLOCK(sb); 1515 aio_complete(job, -1, error); 1516 goto restart; 1517 } 1518 1519 /* 1520 * Door is closed. If there is pending data in the socket buffer, 1521 * deliver it. If there are pending DDP requests, wait for those 1522 * to complete. Once they have completed, return EOF reads. 1523 */ 1524 if (sb->sb_state & SBS_CANTRCVMORE && sbavail(sb) == 0) { 1525 SOCKBUF_UNLOCK(sb); 1526 if (toep->ddp.active_count != 0) 1527 return; 1528 ddp_complete_all(toep, 0); 1529 return; 1530 } 1531 1532 /* 1533 * If DDP is not enabled and there is no pending socket buffer 1534 * data, try to enable DDP. 1535 */ 1536 if (sbavail(sb) == 0 && (toep->ddp.flags & DDP_ON) == 0) { 1537 SOCKBUF_UNLOCK(sb); 1538 1539 /* 1540 * Wait for the card to ACK that DDP is enabled before 1541 * queueing any buffers. Currently this waits for an 1542 * indicate to arrive. This could use a TCB_SET_FIELD_RPL 1543 * message to know that DDP was enabled instead of waiting 1544 * for the indicate which would avoid copying the indicate 1545 * if no data is pending. 1546 * 1547 * XXX: Might want to limit the indicate size to the size 1548 * of the first queued request. 1549 */ 1550 if ((toep->ddp.flags & DDP_SC_REQ) == 0) 1551 enable_ddp(sc, toep); 1552 return; 1553 } 1554 SOCKBUF_UNLOCK(sb); 1555 1556 /* 1557 * If another thread is queueing a buffer for DDP, let it 1558 * drain any work and return. 1559 */ 1560 if (toep->ddp.queueing != NULL) 1561 return; 1562 1563 /* Take the next job to prep it for DDP. */ 1564 toep->ddp.waiting_count--; 1565 TAILQ_REMOVE(&toep->ddp.aiojobq, job, list); 1566 if (!aio_clear_cancel_function(job)) 1567 goto restart; 1568 toep->ddp.queueing = job; 1569 1570 /* NB: This drops DDP_LOCK while it holds the backing VM pages. */ 1571 error = hold_aio(toep, job, &ps); 1572 if (error != 0) { 1573 ddp_complete_one(job, error); 1574 toep->ddp.queueing = NULL; 1575 goto restart; 1576 } 1577 1578 SOCKBUF_LOCK(sb); 1579 if (so->so_error && sbavail(sb) == 0) { 1580 copied = job->aio_received; 1581 if (copied != 0) { 1582 SOCKBUF_UNLOCK(sb); 1583 recycle_pageset(toep, ps); 1584 aio_complete(job, copied, 0); 1585 toep->ddp.queueing = NULL; 1586 goto restart; 1587 } 1588 1589 error = so->so_error; 1590 so->so_error = 0; 1591 SOCKBUF_UNLOCK(sb); 1592 recycle_pageset(toep, ps); 1593 aio_complete(job, -1, error); 1594 toep->ddp.queueing = NULL; 1595 goto restart; 1596 } 1597 1598 if (sb->sb_state & SBS_CANTRCVMORE && sbavail(sb) == 0) { 1599 SOCKBUF_UNLOCK(sb); 1600 recycle_pageset(toep, ps); 1601 if (toep->ddp.active_count != 0) { 1602 /* 1603 * The door is closed, but there are still pending 1604 * DDP buffers. Requeue. These jobs will all be 1605 * completed once those buffers drain. 1606 */ 1607 aio_ddp_requeue_one(toep, job); 1608 toep->ddp.queueing = NULL; 1609 return; 1610 } 1611 ddp_complete_one(job, 0); 1612 ddp_complete_all(toep, 0); 1613 toep->ddp.queueing = NULL; 1614 return; 1615 } 1616 1617 sbcopy: 1618 /* 1619 * If the toep is dead, there shouldn't be any data in the socket 1620 * buffer, so the above case should have handled this. 1621 */ 1622 MPASS(!(toep->ddp.flags & DDP_DEAD)); 1623 1624 /* 1625 * If there is pending data in the socket buffer (either 1626 * from before the requests were queued or a DDP indicate), 1627 * copy those mbufs out directly. 1628 */ 1629 copied = 0; 1630 offset = ps->offset + job->aio_received; 1631 MPASS(job->aio_received <= job->uaiocb.aio_nbytes); 1632 resid = job->uaiocb.aio_nbytes - job->aio_received; 1633 m = sb->sb_mb; 1634 KASSERT(m == NULL || toep->ddp.active_count == 0, 1635 ("%s: sockbuf data with active DDP", __func__)); 1636 while (m != NULL && resid > 0) { 1637 struct iovec iov[1]; 1638 struct uio uio; 1639 int error; 1640 1641 iov[0].iov_base = mtod(m, void *); 1642 iov[0].iov_len = m->m_len; 1643 if (iov[0].iov_len > resid) 1644 iov[0].iov_len = resid; 1645 uio.uio_iov = iov; 1646 uio.uio_iovcnt = 1; 1647 uio.uio_offset = 0; 1648 uio.uio_resid = iov[0].iov_len; 1649 uio.uio_segflg = UIO_SYSSPACE; 1650 uio.uio_rw = UIO_WRITE; 1651 error = uiomove_fromphys(ps->pages, offset + copied, 1652 uio.uio_resid, &uio); 1653 MPASS(error == 0 && uio.uio_resid == 0); 1654 copied += uio.uio_offset; 1655 resid -= uio.uio_offset; 1656 m = m->m_next; 1657 } 1658 if (copied != 0) { 1659 sbdrop_locked(sb, copied); 1660 job->aio_received += copied; 1661 job->msgrcv = 1; 1662 copied = job->aio_received; 1663 inp = sotoinpcb(so); 1664 if (!INP_TRY_WLOCK(inp)) { 1665 /* 1666 * The reference on the socket file descriptor in 1667 * the AIO job should keep 'sb' and 'inp' stable. 1668 * Our caller has a reference on the 'toep' that 1669 * keeps it stable. 1670 */ 1671 SOCKBUF_UNLOCK(sb); 1672 DDP_UNLOCK(toep); 1673 INP_WLOCK(inp); 1674 DDP_LOCK(toep); 1675 SOCKBUF_LOCK(sb); 1676 1677 /* 1678 * If the socket has been closed, we should detect 1679 * that and complete this request if needed on 1680 * the next trip around the loop. 1681 */ 1682 } 1683 t4_rcvd_locked(&toep->td->tod, intotcpcb(inp)); 1684 INP_WUNLOCK(inp); 1685 if (resid == 0 || toep->ddp.flags & DDP_DEAD) { 1686 /* 1687 * We filled the entire buffer with socket 1688 * data, DDP is not being used, or the socket 1689 * is being shut down, so complete the 1690 * request. 1691 */ 1692 SOCKBUF_UNLOCK(sb); 1693 recycle_pageset(toep, ps); 1694 aio_complete(job, copied, 0); 1695 toep->ddp.queueing = NULL; 1696 goto restart; 1697 } 1698 1699 /* 1700 * If DDP is not enabled, requeue this request and restart. 1701 * This will either enable DDP or wait for more data to 1702 * arrive on the socket buffer. 1703 */ 1704 if ((toep->ddp.flags & (DDP_ON | DDP_SC_REQ)) != DDP_ON) { 1705 SOCKBUF_UNLOCK(sb); 1706 recycle_pageset(toep, ps); 1707 aio_ddp_requeue_one(toep, job); 1708 toep->ddp.queueing = NULL; 1709 goto restart; 1710 } 1711 1712 /* 1713 * An indicate might have arrived and been added to 1714 * the socket buffer while it was unlocked after the 1715 * copy to lock the INP. If so, restart the copy. 1716 */ 1717 if (sbavail(sb) != 0) 1718 goto sbcopy; 1719 } 1720 SOCKBUF_UNLOCK(sb); 1721 1722 if (prep_pageset(sc, toep, ps) == 0) { 1723 recycle_pageset(toep, ps); 1724 aio_ddp_requeue_one(toep, job); 1725 toep->ddp.queueing = NULL; 1726 1727 /* 1728 * XXX: Need to retry this later. Mostly need a trigger 1729 * when page pods are freed up. 1730 */ 1731 printf("%s: prep_pageset failed\n", __func__); 1732 return; 1733 } 1734 1735 /* Determine which DDP buffer to use. */ 1736 if (toep->ddp.db[0].job == NULL) { 1737 db_idx = 0; 1738 } else { 1739 MPASS(toep->ddp.db[1].job == NULL); 1740 db_idx = 1; 1741 } 1742 1743 ddp_flags = 0; 1744 ddp_flags_mask = 0; 1745 if (db_idx == 0) { 1746 ddp_flags |= V_TF_DDP_BUF0_VALID(1); 1747 if (so->so_state & SS_NBIO) 1748 ddp_flags |= V_TF_DDP_BUF0_FLUSH(1); 1749 ddp_flags_mask |= V_TF_DDP_PSH_NO_INVALIDATE0(1) | 1750 V_TF_DDP_PUSH_DISABLE_0(1) | V_TF_DDP_PSHF_ENABLE_0(1) | 1751 V_TF_DDP_BUF0_FLUSH(1) | V_TF_DDP_BUF0_VALID(1); 1752 buf_flag = DDP_BUF0_ACTIVE; 1753 } else { 1754 ddp_flags |= V_TF_DDP_BUF1_VALID(1); 1755 if (so->so_state & SS_NBIO) 1756 ddp_flags |= V_TF_DDP_BUF1_FLUSH(1); 1757 ddp_flags_mask |= V_TF_DDP_PSH_NO_INVALIDATE1(1) | 1758 V_TF_DDP_PUSH_DISABLE_1(1) | V_TF_DDP_PSHF_ENABLE_1(1) | 1759 V_TF_DDP_BUF1_FLUSH(1) | V_TF_DDP_BUF1_VALID(1); 1760 buf_flag = DDP_BUF1_ACTIVE; 1761 } 1762 MPASS((toep->ddp.flags & buf_flag) == 0); 1763 if ((toep->ddp.flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE)) == 0) { 1764 MPASS(db_idx == 0); 1765 MPASS(toep->ddp.active_id == -1); 1766 MPASS(toep->ddp.active_count == 0); 1767 ddp_flags_mask |= V_TF_DDP_ACTIVE_BUF(1); 1768 } 1769 1770 /* 1771 * The TID for this connection should still be valid. If DDP_DEAD 1772 * is set, SBS_CANTRCVMORE should be set, so we shouldn't be 1773 * this far anyway. Even if the socket is closing on the other 1774 * end, the AIO job holds a reference on this end of the socket 1775 * which will keep it open and keep the TCP PCB attached until 1776 * after the job is completed. 1777 */ 1778 wr = mk_update_tcb_for_ddp(sc, toep, db_idx, ps, job->aio_received, 1779 ddp_flags, ddp_flags_mask); 1780 if (wr == NULL) { 1781 recycle_pageset(toep, ps); 1782 aio_ddp_requeue_one(toep, job); 1783 toep->ddp.queueing = NULL; 1784 1785 /* 1786 * XXX: Need a way to kick a retry here. 1787 * 1788 * XXX: We know the fixed size needed and could 1789 * preallocate this using a blocking request at the 1790 * start of the task to avoid having to handle this 1791 * edge case. 1792 */ 1793 printf("%s: mk_update_tcb_for_ddp failed\n", __func__); 1794 return; 1795 } 1796 1797 if (!aio_set_cancel_function(job, t4_aio_cancel_active)) { 1798 free_wrqe(wr); 1799 recycle_pageset(toep, ps); 1800 aio_ddp_cancel_one(job); 1801 toep->ddp.queueing = NULL; 1802 goto restart; 1803 } 1804 1805 #ifdef VERBOSE_TRACES 1806 CTR5(KTR_CXGBE, "%s: scheduling %p for DDP[%d] (flags %#lx/%#lx)", 1807 __func__, job, db_idx, ddp_flags, ddp_flags_mask); 1808 #endif 1809 /* Give the chip the go-ahead. */ 1810 t4_wrq_tx(sc, wr); 1811 db = &toep->ddp.db[db_idx]; 1812 db->cancel_pending = 0; 1813 db->job = job; 1814 db->ps = ps; 1815 toep->ddp.queueing = NULL; 1816 toep->ddp.flags |= buf_flag; 1817 toep->ddp.active_count++; 1818 if (toep->ddp.active_count == 1) { 1819 MPASS(toep->ddp.active_id == -1); 1820 toep->ddp.active_id = db_idx; 1821 CTR2(KTR_CXGBE, "%s: ddp_active_id = %d", __func__, 1822 toep->ddp.active_id); 1823 } 1824 goto restart; 1825 } 1826 1827 void 1828 ddp_queue_toep(struct toepcb *toep) 1829 { 1830 1831 DDP_ASSERT_LOCKED(toep); 1832 if (toep->ddp.flags & DDP_TASK_ACTIVE) 1833 return; 1834 toep->ddp.flags |= DDP_TASK_ACTIVE; 1835 hold_toepcb(toep); 1836 soaio_enqueue(&toep->ddp.requeue_task); 1837 } 1838 1839 static void 1840 aio_ddp_requeue_task(void *context, int pending) 1841 { 1842 struct toepcb *toep = context; 1843 1844 DDP_LOCK(toep); 1845 aio_ddp_requeue(toep); 1846 toep->ddp.flags &= ~DDP_TASK_ACTIVE; 1847 DDP_UNLOCK(toep); 1848 1849 free_toepcb(toep); 1850 } 1851 1852 static void 1853 t4_aio_cancel_active(struct kaiocb *job) 1854 { 1855 struct socket *so = job->fd_file->f_data; 1856 struct tcpcb *tp = so_sototcpcb(so); 1857 struct toepcb *toep = tp->t_toe; 1858 struct adapter *sc = td_adapter(toep->td); 1859 uint64_t valid_flag; 1860 int i; 1861 1862 DDP_LOCK(toep); 1863 if (aio_cancel_cleared(job)) { 1864 DDP_UNLOCK(toep); 1865 aio_ddp_cancel_one(job); 1866 return; 1867 } 1868 1869 for (i = 0; i < nitems(toep->ddp.db); i++) { 1870 if (toep->ddp.db[i].job == job) { 1871 /* Should only ever get one cancel request for a job. */ 1872 MPASS(toep->ddp.db[i].cancel_pending == 0); 1873 1874 /* 1875 * Invalidate this buffer. It will be 1876 * cancelled or partially completed once the 1877 * card ACKs the invalidate. 1878 */ 1879 valid_flag = i == 0 ? V_TF_DDP_BUF0_VALID(1) : 1880 V_TF_DDP_BUF1_VALID(1); 1881 t4_set_tcb_field(sc, toep->ctrlq, toep, 1882 W_TCB_RX_DDP_FLAGS, valid_flag, 0, 1, 1883 i + DDP_BUF0_INVALIDATED); 1884 toep->ddp.db[i].cancel_pending = 1; 1885 CTR2(KTR_CXGBE, "%s: request %p marked pending", 1886 __func__, job); 1887 break; 1888 } 1889 } 1890 DDP_UNLOCK(toep); 1891 } 1892 1893 static void 1894 t4_aio_cancel_queued(struct kaiocb *job) 1895 { 1896 struct socket *so = job->fd_file->f_data; 1897 struct tcpcb *tp = so_sototcpcb(so); 1898 struct toepcb *toep = tp->t_toe; 1899 1900 DDP_LOCK(toep); 1901 if (!aio_cancel_cleared(job)) { 1902 TAILQ_REMOVE(&toep->ddp.aiojobq, job, list); 1903 toep->ddp.waiting_count--; 1904 if (toep->ddp.waiting_count == 0) 1905 ddp_queue_toep(toep); 1906 } 1907 CTR2(KTR_CXGBE, "%s: request %p cancelled", __func__, job); 1908 DDP_UNLOCK(toep); 1909 1910 aio_ddp_cancel_one(job); 1911 } 1912 1913 int 1914 t4_aio_queue_ddp(struct socket *so, struct kaiocb *job) 1915 { 1916 struct tcpcb *tp = so_sototcpcb(so); 1917 struct toepcb *toep = tp->t_toe; 1918 1919 1920 /* Ignore writes. */ 1921 if (job->uaiocb.aio_lio_opcode != LIO_READ) 1922 return (EOPNOTSUPP); 1923 1924 DDP_LOCK(toep); 1925 1926 /* 1927 * XXX: Think about possibly returning errors for ENOTCONN, 1928 * etc. Perhaps the caller would only queue the request 1929 * if it failed with EOPNOTSUPP? 1930 */ 1931 1932 #ifdef VERBOSE_TRACES 1933 CTR2(KTR_CXGBE, "%s: queueing %p", __func__, job); 1934 #endif 1935 if (!aio_set_cancel_function(job, t4_aio_cancel_queued)) 1936 panic("new job was cancelled"); 1937 TAILQ_INSERT_TAIL(&toep->ddp.aiojobq, job, list); 1938 toep->ddp.waiting_count++; 1939 toep->ddp.flags |= DDP_OK; 1940 1941 /* 1942 * Try to handle this request synchronously. If this has 1943 * to block because the task is running, it will just bail 1944 * and let the task handle it instead. 1945 */ 1946 aio_ddp_requeue(toep); 1947 DDP_UNLOCK(toep); 1948 return (0); 1949 } 1950 1951 void 1952 t4_ddp_mod_load(void) 1953 { 1954 1955 t4_register_shared_cpl_handler(CPL_SET_TCB_RPL, do_ddp_tcb_rpl, 1956 CPL_COOKIE_DDP0); 1957 t4_register_shared_cpl_handler(CPL_SET_TCB_RPL, do_ddp_tcb_rpl, 1958 CPL_COOKIE_DDP1); 1959 t4_register_cpl_handler(CPL_RX_DATA_DDP, do_rx_data_ddp); 1960 t4_register_cpl_handler(CPL_RX_DDP_COMPLETE, do_rx_ddp_complete); 1961 TAILQ_INIT(&ddp_orphan_pagesets); 1962 mtx_init(&ddp_orphan_pagesets_lock, "ddp orphans", NULL, MTX_DEF); 1963 TASK_INIT(&ddp_orphan_task, 0, ddp_free_orphan_pagesets, NULL); 1964 } 1965 1966 void 1967 t4_ddp_mod_unload(void) 1968 { 1969 1970 taskqueue_drain(taskqueue_thread, &ddp_orphan_task); 1971 MPASS(TAILQ_EMPTY(&ddp_orphan_pagesets)); 1972 mtx_destroy(&ddp_orphan_pagesets_lock); 1973 t4_register_cpl_handler(CPL_RX_DATA_DDP, NULL); 1974 t4_register_cpl_handler(CPL_RX_DDP_COMPLETE, NULL); 1975 } 1976 #endif 1977