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/aio.h> 35 #include <sys/file.h> 36 #include <sys/systm.h> 37 #include <sys/kernel.h> 38 #include <sys/ktr.h> 39 #include <sys/module.h> 40 #include <sys/protosw.h> 41 #include <sys/proc.h> 42 #include <sys/domain.h> 43 #include <sys/socket.h> 44 #include <sys/socketvar.h> 45 #include <sys/taskqueue.h> 46 #include <sys/uio.h> 47 #include <netinet/in.h> 48 #include <netinet/in_pcb.h> 49 #include <netinet/ip.h> 50 #include <netinet/tcp_var.h> 51 #define TCPSTATES 52 #include <netinet/tcp_fsm.h> 53 #include <netinet/toecore.h> 54 55 #include <vm/vm.h> 56 #include <vm/vm_extern.h> 57 #include <vm/vm_param.h> 58 #include <vm/pmap.h> 59 #include <vm/vm_map.h> 60 #include <vm/vm_page.h> 61 #include <vm/vm_object.h> 62 63 #ifdef TCP_OFFLOAD 64 #include "common/common.h" 65 #include "common/t4_msg.h" 66 #include "common/t4_regs.h" 67 #include "common/t4_tcb.h" 68 #include "tom/t4_tom.h" 69 70 VNET_DECLARE(int, tcp_do_autorcvbuf); 71 #define V_tcp_do_autorcvbuf VNET(tcp_do_autorcvbuf) 72 VNET_DECLARE(int, tcp_autorcvbuf_inc); 73 #define V_tcp_autorcvbuf_inc VNET(tcp_autorcvbuf_inc) 74 VNET_DECLARE(int, tcp_autorcvbuf_max); 75 #define V_tcp_autorcvbuf_max VNET(tcp_autorcvbuf_max) 76 77 /* 78 * Use the 'backend3' field in AIO jobs to store the amount of data 79 * received by the AIO job so far. 80 */ 81 #define aio_received backend3 82 83 static void aio_ddp_requeue_task(void *context, int pending); 84 static void ddp_complete_all(struct toepcb *toep, int error); 85 static void t4_aio_cancel_active(struct kaiocb *job); 86 static void t4_aio_cancel_queued(struct kaiocb *job); 87 88 #define PPOD_SZ(n) ((n) * sizeof(struct pagepod)) 89 #define PPOD_SIZE (PPOD_SZ(1)) 90 91 /* XXX: must match A_ULP_RX_TDDP_PSZ */ 92 static int t4_ddp_pgsz[] = {4096, 4096 << 2, 4096 << 4, 4096 << 6}; 93 94 static TAILQ_HEAD(, pageset) ddp_orphan_pagesets; 95 static struct mtx ddp_orphan_pagesets_lock; 96 static struct task ddp_orphan_task; 97 98 #define MAX_DDP_BUFFER_SIZE (M_TCB_RX_DDP_BUF0_LEN) 99 static int 100 alloc_ppods(struct tom_data *td, int n, u_int *ppod_addr) 101 { 102 vmem_addr_t v; 103 int rc; 104 105 MPASS(n > 0); 106 107 rc = vmem_alloc(td->ppod_arena, PPOD_SZ(n), M_NOWAIT | M_FIRSTFIT, &v); 108 *ppod_addr = (u_int)v; 109 110 return (rc); 111 } 112 113 static void 114 free_ppods(struct tom_data *td, u_int ppod_addr, int n) 115 { 116 117 MPASS(n > 0); 118 119 vmem_free(td->ppod_arena, (vmem_addr_t)ppod_addr, PPOD_SZ(n)); 120 } 121 122 static inline int 123 pages_to_nppods(int npages, int ddp_pgsz) 124 { 125 int nsegs = npages * PAGE_SIZE / ddp_pgsz; 126 127 return (howmany(nsegs, PPOD_PAGES)); 128 } 129 130 /* 131 * A page set holds information about a buffer used for DDP. The page 132 * set holds resources such as the VM pages backing the buffer (either 133 * held or wired) and the page pods associated with the buffer. 134 * Recently used page sets are cached to allow for efficient reuse of 135 * buffers (avoiding the need to re-fault in pages, hold them, etc.). 136 * Note that cached page sets keep the backing pages wired. The 137 * number of wired pages is capped by only allowing for two wired 138 * pagesets per connection. This is not a perfect cap, but is a 139 * trade-off for performance. 140 * 141 * If an application ping-pongs two buffers for a connection via 142 * aio_read(2) then those buffers should remain wired and expensive VM 143 * fault lookups should be avoided after each buffer has been used 144 * once. If an application uses more than two buffers then this will 145 * fall back to doing expensive VM fault lookups for each operation. 146 */ 147 static void 148 free_pageset(struct tom_data *td, struct pageset *ps) 149 { 150 vm_page_t p; 151 int i; 152 153 if (ps->nppods > 0) 154 free_ppods(td, ps->ppod_addr, ps->nppods); 155 156 if (ps->flags & PS_WIRED) { 157 for (i = 0; i < ps->npages; i++) { 158 p = ps->pages[i]; 159 vm_page_lock(p); 160 vm_page_unwire(p, PQ_INACTIVE); 161 vm_page_unlock(p); 162 } 163 } else 164 vm_page_unhold_pages(ps->pages, ps->npages); 165 mtx_lock(&ddp_orphan_pagesets_lock); 166 TAILQ_INSERT_TAIL(&ddp_orphan_pagesets, ps, link); 167 taskqueue_enqueue(taskqueue_thread, &ddp_orphan_task); 168 mtx_unlock(&ddp_orphan_pagesets_lock); 169 } 170 171 static void 172 ddp_free_orphan_pagesets(void *context, int pending) 173 { 174 struct pageset *ps; 175 176 mtx_lock(&ddp_orphan_pagesets_lock); 177 while (!TAILQ_EMPTY(&ddp_orphan_pagesets)) { 178 ps = TAILQ_FIRST(&ddp_orphan_pagesets); 179 TAILQ_REMOVE(&ddp_orphan_pagesets, ps, link); 180 mtx_unlock(&ddp_orphan_pagesets_lock); 181 if (ps->vm) 182 vmspace_free(ps->vm); 183 free(ps, M_CXGBE); 184 mtx_lock(&ddp_orphan_pagesets_lock); 185 } 186 mtx_unlock(&ddp_orphan_pagesets_lock); 187 } 188 189 static void 190 recycle_pageset(struct toepcb *toep, struct pageset *ps) 191 { 192 193 DDP_ASSERT_LOCKED(toep); 194 if (!(toep->ddp_flags & DDP_DEAD) && ps->flags & PS_WIRED) { 195 KASSERT(toep->ddp_cached_count + toep->ddp_active_count < 196 nitems(toep->db), ("too many wired pagesets")); 197 TAILQ_INSERT_HEAD(&toep->ddp_cached_pagesets, ps, link); 198 toep->ddp_cached_count++; 199 } else 200 free_pageset(toep->td, ps); 201 } 202 203 static void 204 ddp_complete_one(struct kaiocb *job, int error) 205 { 206 long copied; 207 208 /* 209 * If this job had copied data out of the socket buffer before 210 * it was cancelled, report it as a short read rather than an 211 * error. 212 */ 213 copied = job->aio_received; 214 if (copied != 0 || error == 0) 215 aio_complete(job, copied, 0); 216 else 217 aio_complete(job, -1, error); 218 } 219 220 static void 221 free_ddp_buffer(struct tom_data *td, struct ddp_buffer *db) 222 { 223 224 if (db->job) { 225 /* 226 * XXX: If we are un-offloading the socket then we 227 * should requeue these on the socket somehow. If we 228 * got a FIN from the remote end, then this completes 229 * any remaining requests with an EOF read. 230 */ 231 if (!aio_clear_cancel_function(db->job)) 232 ddp_complete_one(db->job, 0); 233 } 234 235 if (db->ps) 236 free_pageset(td, db->ps); 237 } 238 239 void 240 ddp_init_toep(struct toepcb *toep) 241 { 242 243 TAILQ_INIT(&toep->ddp_aiojobq); 244 TASK_INIT(&toep->ddp_requeue_task, 0, aio_ddp_requeue_task, toep); 245 toep->ddp_active_id = -1; 246 mtx_init(&toep->ddp_lock, "t4 ddp", NULL, MTX_DEF); 247 } 248 249 void 250 ddp_uninit_toep(struct toepcb *toep) 251 { 252 253 mtx_destroy(&toep->ddp_lock); 254 } 255 256 void 257 release_ddp_resources(struct toepcb *toep) 258 { 259 struct pageset *ps; 260 int i; 261 262 DDP_LOCK(toep); 263 toep->flags |= DDP_DEAD; 264 for (i = 0; i < nitems(toep->db); i++) { 265 free_ddp_buffer(toep->td, &toep->db[i]); 266 } 267 while ((ps = TAILQ_FIRST(&toep->ddp_cached_pagesets)) != NULL) { 268 TAILQ_REMOVE(&toep->ddp_cached_pagesets, ps, link); 269 free_pageset(toep->td, ps); 270 } 271 ddp_complete_all(toep, 0); 272 DDP_UNLOCK(toep); 273 } 274 275 #ifdef INVARIANTS 276 void 277 ddp_assert_empty(struct toepcb *toep) 278 { 279 int i; 280 281 MPASS(!(toep->ddp_flags & DDP_TASK_ACTIVE)); 282 for (i = 0; i < nitems(toep->db); i++) { 283 MPASS(toep->db[i].job == NULL); 284 MPASS(toep->db[i].ps == NULL); 285 } 286 MPASS(TAILQ_EMPTY(&toep->ddp_cached_pagesets)); 287 MPASS(TAILQ_EMPTY(&toep->ddp_aiojobq)); 288 } 289 #endif 290 291 static void 292 complete_ddp_buffer(struct toepcb *toep, struct ddp_buffer *db, 293 unsigned int db_idx) 294 { 295 unsigned int db_flag; 296 297 toep->ddp_active_count--; 298 if (toep->ddp_active_id == db_idx) { 299 if (toep->ddp_active_count == 0) { 300 KASSERT(toep->db[db_idx ^ 1].job == NULL, 301 ("%s: active_count mismatch", __func__)); 302 toep->ddp_active_id = -1; 303 } else 304 toep->ddp_active_id ^= 1; 305 #ifdef VERBOSE_TRACES 306 CTR2(KTR_CXGBE, "%s: ddp_active_id = %d", __func__, 307 toep->ddp_active_id); 308 #endif 309 } else { 310 KASSERT(toep->ddp_active_count != 0 && 311 toep->ddp_active_id != -1, 312 ("%s: active count mismatch", __func__)); 313 } 314 315 db->cancel_pending = 0; 316 db->job = NULL; 317 recycle_pageset(toep, db->ps); 318 db->ps = NULL; 319 320 db_flag = db_idx == 1 ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE; 321 KASSERT(toep->ddp_flags & db_flag, 322 ("%s: DDP buffer not active. toep %p, ddp_flags 0x%x", 323 __func__, toep, toep->ddp_flags)); 324 toep->ddp_flags &= ~db_flag; 325 } 326 327 /* XXX: handle_ddp_data code duplication */ 328 void 329 insert_ddp_data(struct toepcb *toep, uint32_t n) 330 { 331 struct inpcb *inp = toep->inp; 332 struct tcpcb *tp = intotcpcb(inp); 333 struct ddp_buffer *db; 334 struct kaiocb *job; 335 size_t placed; 336 long copied; 337 unsigned int db_flag, db_idx; 338 339 INP_WLOCK_ASSERT(inp); 340 DDP_ASSERT_LOCKED(toep); 341 342 tp->rcv_nxt += n; 343 #ifndef USE_DDP_RX_FLOW_CONTROL 344 KASSERT(tp->rcv_wnd >= n, ("%s: negative window size", __func__)); 345 tp->rcv_wnd -= n; 346 #endif 347 #ifndef USE_DDP_RX_FLOW_CONTROL 348 toep->rx_credits += n; 349 #endif 350 CTR2(KTR_CXGBE, "%s: placed %u bytes before falling out of DDP", 351 __func__, n); 352 while (toep->ddp_active_count > 0) { 353 MPASS(toep->ddp_active_id != -1); 354 db_idx = toep->ddp_active_id; 355 db_flag = db_idx == 1 ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE; 356 MPASS((toep->ddp_flags & db_flag) != 0); 357 db = &toep->db[db_idx]; 358 job = db->job; 359 copied = job->aio_received; 360 placed = n; 361 if (placed > job->uaiocb.aio_nbytes - copied) 362 placed = job->uaiocb.aio_nbytes - copied; 363 if (placed > 0) 364 job->msgrcv = 1; 365 if (!aio_clear_cancel_function(job)) { 366 /* 367 * Update the copied length for when 368 * t4_aio_cancel_active() completes this 369 * request. 370 */ 371 job->aio_received += placed; 372 } else if (copied + placed != 0) { 373 CTR4(KTR_CXGBE, 374 "%s: completing %p (copied %ld, placed %lu)", 375 __func__, job, copied, placed); 376 /* XXX: This always completes if there is some data. */ 377 aio_complete(job, copied + placed, 0); 378 } else if (aio_set_cancel_function(job, t4_aio_cancel_queued)) { 379 TAILQ_INSERT_HEAD(&toep->ddp_aiojobq, job, list); 380 toep->ddp_waiting_count++; 381 } else 382 aio_cancel(job); 383 n -= placed; 384 complete_ddp_buffer(toep, db, db_idx); 385 } 386 387 MPASS(n == 0); 388 } 389 390 /* SET_TCB_FIELD sent as a ULP command looks like this */ 391 #define LEN__SET_TCB_FIELD_ULP (sizeof(struct ulp_txpkt) + \ 392 sizeof(struct ulptx_idata) + sizeof(struct cpl_set_tcb_field_core)) 393 394 /* RX_DATA_ACK sent as a ULP command looks like this */ 395 #define LEN__RX_DATA_ACK_ULP (sizeof(struct ulp_txpkt) + \ 396 sizeof(struct ulptx_idata) + sizeof(struct cpl_rx_data_ack_core)) 397 398 static inline void * 399 mk_set_tcb_field_ulp(struct ulp_txpkt *ulpmc, struct toepcb *toep, 400 uint64_t word, uint64_t mask, uint64_t val) 401 { 402 struct ulptx_idata *ulpsc; 403 struct cpl_set_tcb_field_core *req; 404 405 ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0)); 406 ulpmc->len = htobe32(howmany(LEN__SET_TCB_FIELD_ULP, 16)); 407 408 ulpsc = (struct ulptx_idata *)(ulpmc + 1); 409 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM)); 410 ulpsc->len = htobe32(sizeof(*req)); 411 412 req = (struct cpl_set_tcb_field_core *)(ulpsc + 1); 413 OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_SET_TCB_FIELD, toep->tid)); 414 req->reply_ctrl = htobe16(V_NO_REPLY(1) | 415 V_QUEUENO(toep->ofld_rxq->iq.abs_id)); 416 req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0)); 417 req->mask = htobe64(mask); 418 req->val = htobe64(val); 419 420 ulpsc = (struct ulptx_idata *)(req + 1); 421 if (LEN__SET_TCB_FIELD_ULP % 16) { 422 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP)); 423 ulpsc->len = htobe32(0); 424 return (ulpsc + 1); 425 } 426 return (ulpsc); 427 } 428 429 static inline void * 430 mk_rx_data_ack_ulp(struct ulp_txpkt *ulpmc, struct toepcb *toep) 431 { 432 struct ulptx_idata *ulpsc; 433 struct cpl_rx_data_ack_core *req; 434 435 ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0)); 436 ulpmc->len = htobe32(howmany(LEN__RX_DATA_ACK_ULP, 16)); 437 438 ulpsc = (struct ulptx_idata *)(ulpmc + 1); 439 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM)); 440 ulpsc->len = htobe32(sizeof(*req)); 441 442 req = (struct cpl_rx_data_ack_core *)(ulpsc + 1); 443 OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_RX_DATA_ACK, toep->tid)); 444 req->credit_dack = htobe32(F_RX_MODULATE_RX); 445 446 ulpsc = (struct ulptx_idata *)(req + 1); 447 if (LEN__RX_DATA_ACK_ULP % 16) { 448 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP)); 449 ulpsc->len = htobe32(0); 450 return (ulpsc + 1); 451 } 452 return (ulpsc); 453 } 454 455 static struct wrqe * 456 mk_update_tcb_for_ddp(struct adapter *sc, struct toepcb *toep, int db_idx, 457 struct pageset *ps, int offset, uint64_t ddp_flags, uint64_t ddp_flags_mask) 458 { 459 struct wrqe *wr; 460 struct work_request_hdr *wrh; 461 struct ulp_txpkt *ulpmc; 462 int len; 463 464 KASSERT(db_idx == 0 || db_idx == 1, 465 ("%s: bad DDP buffer index %d", __func__, db_idx)); 466 467 /* 468 * We'll send a compound work request that has 3 SET_TCB_FIELDs and an 469 * RX_DATA_ACK (with RX_MODULATE to speed up delivery). 470 * 471 * The work request header is 16B and always ends at a 16B boundary. 472 * The ULPTX master commands that follow must all end at 16B boundaries 473 * too so we round up the size to 16. 474 */ 475 len = sizeof(*wrh) + 3 * roundup2(LEN__SET_TCB_FIELD_ULP, 16) + 476 roundup2(LEN__RX_DATA_ACK_ULP, 16); 477 478 wr = alloc_wrqe(len, toep->ctrlq); 479 if (wr == NULL) 480 return (NULL); 481 wrh = wrtod(wr); 482 INIT_ULPTX_WRH(wrh, len, 1, 0); /* atomic */ 483 ulpmc = (struct ulp_txpkt *)(wrh + 1); 484 485 /* Write the buffer's tag */ 486 ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, 487 W_TCB_RX_DDP_BUF0_TAG + db_idx, 488 V_TCB_RX_DDP_BUF0_TAG(M_TCB_RX_DDP_BUF0_TAG), 489 V_TCB_RX_DDP_BUF0_TAG(ps->tag)); 490 491 /* Update the current offset in the DDP buffer and its total length */ 492 if (db_idx == 0) 493 ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, 494 W_TCB_RX_DDP_BUF0_OFFSET, 495 V_TCB_RX_DDP_BUF0_OFFSET(M_TCB_RX_DDP_BUF0_OFFSET) | 496 V_TCB_RX_DDP_BUF0_LEN(M_TCB_RX_DDP_BUF0_LEN), 497 V_TCB_RX_DDP_BUF0_OFFSET(offset) | 498 V_TCB_RX_DDP_BUF0_LEN(ps->len)); 499 else 500 ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, 501 W_TCB_RX_DDP_BUF1_OFFSET, 502 V_TCB_RX_DDP_BUF1_OFFSET(M_TCB_RX_DDP_BUF1_OFFSET) | 503 V_TCB_RX_DDP_BUF1_LEN((u64)M_TCB_RX_DDP_BUF1_LEN << 32), 504 V_TCB_RX_DDP_BUF1_OFFSET(offset) | 505 V_TCB_RX_DDP_BUF1_LEN((u64)ps->len << 32)); 506 507 /* Update DDP flags */ 508 ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, W_TCB_RX_DDP_FLAGS, 509 ddp_flags_mask, ddp_flags); 510 511 /* Gratuitous RX_DATA_ACK with RX_MODULATE set to speed up delivery. */ 512 ulpmc = mk_rx_data_ack_ulp(ulpmc, toep); 513 514 return (wr); 515 } 516 517 static int 518 handle_ddp_data(struct toepcb *toep, __be32 ddp_report, __be32 rcv_nxt, int len) 519 { 520 uint32_t report = be32toh(ddp_report); 521 unsigned int db_idx; 522 struct inpcb *inp = toep->inp; 523 struct ddp_buffer *db; 524 struct tcpcb *tp; 525 struct socket *so; 526 struct sockbuf *sb; 527 struct kaiocb *job; 528 long copied; 529 530 db_idx = report & F_DDP_BUF_IDX ? 1 : 0; 531 532 if (__predict_false(!(report & F_DDP_INV))) 533 CXGBE_UNIMPLEMENTED("DDP buffer still valid"); 534 535 INP_WLOCK(inp); 536 so = inp_inpcbtosocket(inp); 537 sb = &so->so_rcv; 538 DDP_LOCK(toep); 539 540 KASSERT(toep->ddp_active_id == db_idx, 541 ("completed DDP buffer (%d) != active_id (%d) for tid %d", db_idx, 542 toep->ddp_active_id, toep->tid)); 543 db = &toep->db[db_idx]; 544 job = db->job; 545 546 if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))) { 547 /* 548 * This can happen due to an administrative tcpdrop(8). 549 * Just fail the request with ECONNRESET. 550 */ 551 CTR5(KTR_CXGBE, "%s: tid %u, seq 0x%x, len %d, inp_flags 0x%x", 552 __func__, toep->tid, be32toh(rcv_nxt), len, inp->inp_flags); 553 if (aio_clear_cancel_function(job)) 554 ddp_complete_one(job, ECONNRESET); 555 goto completed; 556 } 557 558 tp = intotcpcb(inp); 559 560 /* 561 * For RX_DDP_COMPLETE, len will be zero and rcv_nxt is the 562 * sequence number of the next byte to receive. The length of 563 * the data received for this message must be computed by 564 * comparing the new and old values of rcv_nxt. 565 * 566 * For RX_DATA_DDP, len might be non-zero, but it is only the 567 * length of the most recent DMA. It does not include the 568 * total length of the data received since the previous update 569 * for this DDP buffer. rcv_nxt is the sequence number of the 570 * first received byte from the most recent DMA. 571 */ 572 len += be32toh(rcv_nxt) - tp->rcv_nxt; 573 tp->rcv_nxt += len; 574 tp->t_rcvtime = ticks; 575 #ifndef USE_DDP_RX_FLOW_CONTROL 576 KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__)); 577 tp->rcv_wnd -= len; 578 #endif 579 #ifdef VERBOSE_TRACES 580 CTR4(KTR_CXGBE, "%s: DDP[%d] placed %d bytes (%#x)", __func__, db_idx, 581 len, report); 582 #endif 583 584 /* receive buffer autosize */ 585 CURVNET_SET(so->so_vnet); 586 SOCKBUF_LOCK(sb); 587 if (sb->sb_flags & SB_AUTOSIZE && 588 V_tcp_do_autorcvbuf && 589 sb->sb_hiwat < V_tcp_autorcvbuf_max && 590 len > (sbspace(sb) / 8 * 7)) { 591 unsigned int hiwat = sb->sb_hiwat; 592 unsigned int newsize = min(hiwat + V_tcp_autorcvbuf_inc, 593 V_tcp_autorcvbuf_max); 594 595 if (!sbreserve_locked(sb, newsize, so, NULL)) 596 sb->sb_flags &= ~SB_AUTOSIZE; 597 else 598 toep->rx_credits += newsize - hiwat; 599 } 600 SOCKBUF_UNLOCK(sb); 601 CURVNET_RESTORE(); 602 603 #ifndef USE_DDP_RX_FLOW_CONTROL 604 toep->rx_credits += len; 605 #endif 606 607 job->msgrcv = 1; 608 if (db->cancel_pending) { 609 /* 610 * Update the job's length but defer completion to the 611 * TCB_RPL callback. 612 */ 613 job->aio_received += len; 614 goto out; 615 } else if (!aio_clear_cancel_function(job)) { 616 /* 617 * Update the copied length for when 618 * t4_aio_cancel_active() completes this request. 619 */ 620 job->aio_received += len; 621 } else { 622 copied = job->aio_received; 623 #ifdef VERBOSE_TRACES 624 CTR4(KTR_CXGBE, "%s: completing %p (copied %ld, placed %d)", 625 __func__, job, copied, len); 626 #endif 627 aio_complete(job, copied + len, 0); 628 t4_rcvd(&toep->td->tod, tp); 629 } 630 631 completed: 632 complete_ddp_buffer(toep, db, db_idx); 633 if (toep->ddp_waiting_count > 0) 634 ddp_queue_toep(toep); 635 out: 636 DDP_UNLOCK(toep); 637 INP_WUNLOCK(inp); 638 639 return (0); 640 } 641 642 void 643 handle_ddp_indicate(struct toepcb *toep) 644 { 645 646 DDP_ASSERT_LOCKED(toep); 647 MPASS(toep->ddp_active_count == 0); 648 MPASS((toep->ddp_flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE)) == 0); 649 if (toep->ddp_waiting_count == 0) { 650 /* 651 * The pending requests that triggered the request for an 652 * an indicate were cancelled. Those cancels should have 653 * already disabled DDP. Just ignore this as the data is 654 * going into the socket buffer anyway. 655 */ 656 return; 657 } 658 CTR3(KTR_CXGBE, "%s: tid %d indicated (%d waiting)", __func__, 659 toep->tid, toep->ddp_waiting_count); 660 ddp_queue_toep(toep); 661 } 662 663 enum { 664 DDP_BUF0_INVALIDATED = 0x2, 665 DDP_BUF1_INVALIDATED 666 }; 667 668 void 669 handle_ddp_tcb_rpl(struct toepcb *toep, const struct cpl_set_tcb_rpl *cpl) 670 { 671 unsigned int db_idx; 672 struct inpcb *inp = toep->inp; 673 struct ddp_buffer *db; 674 struct kaiocb *job; 675 long copied; 676 677 if (cpl->status != CPL_ERR_NONE) 678 panic("XXX: tcp_rpl failed: %d", cpl->status); 679 680 switch (cpl->cookie) { 681 case V_WORD(W_TCB_RX_DDP_FLAGS) | V_COOKIE(DDP_BUF0_INVALIDATED): 682 case V_WORD(W_TCB_RX_DDP_FLAGS) | V_COOKIE(DDP_BUF1_INVALIDATED): 683 /* 684 * XXX: This duplicates a lot of code with handle_ddp_data(). 685 */ 686 db_idx = G_COOKIE(cpl->cookie) - DDP_BUF0_INVALIDATED; 687 INP_WLOCK(inp); 688 DDP_LOCK(toep); 689 db = &toep->db[db_idx]; 690 691 /* 692 * handle_ddp_data() should leave the job around until 693 * this callback runs once a cancel is pending. 694 */ 695 MPASS(db != NULL); 696 MPASS(db->job != NULL); 697 MPASS(db->cancel_pending); 698 699 /* 700 * XXX: It's not clear what happens if there is data 701 * placed when the buffer is invalidated. I suspect we 702 * need to read the TCB to see how much data was placed. 703 * 704 * For now this just pretends like nothing was placed. 705 * 706 * XXX: Note that if we did check the PCB we would need to 707 * also take care of updating the tp, etc. 708 */ 709 job = db->job; 710 copied = job->aio_received; 711 if (copied == 0) { 712 CTR2(KTR_CXGBE, "%s: cancelling %p", __func__, job); 713 aio_cancel(job); 714 } else { 715 CTR3(KTR_CXGBE, "%s: completing %p (copied %ld)", 716 __func__, job, copied); 717 aio_complete(job, copied, 0); 718 t4_rcvd(&toep->td->tod, intotcpcb(inp)); 719 } 720 721 complete_ddp_buffer(toep, db, db_idx); 722 if (toep->ddp_waiting_count > 0) 723 ddp_queue_toep(toep); 724 DDP_UNLOCK(toep); 725 INP_WUNLOCK(inp); 726 break; 727 default: 728 panic("XXX: unknown tcb_rpl offset %#x, cookie %#x", 729 G_WORD(cpl->cookie), G_COOKIE(cpl->cookie)); 730 } 731 } 732 733 void 734 handle_ddp_close(struct toepcb *toep, struct tcpcb *tp, __be32 rcv_nxt) 735 { 736 struct ddp_buffer *db; 737 struct kaiocb *job; 738 long copied; 739 unsigned int db_flag, db_idx; 740 int len, placed; 741 742 INP_WLOCK_ASSERT(toep->inp); 743 DDP_ASSERT_LOCKED(toep); 744 len = be32toh(rcv_nxt) - tp->rcv_nxt; 745 746 tp->rcv_nxt += len; 747 #ifndef USE_DDP_RX_FLOW_CONTROL 748 toep->rx_credits += len; 749 #endif 750 751 while (toep->ddp_active_count > 0) { 752 MPASS(toep->ddp_active_id != -1); 753 db_idx = toep->ddp_active_id; 754 db_flag = db_idx == 1 ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE; 755 MPASS((toep->ddp_flags & db_flag) != 0); 756 db = &toep->db[db_idx]; 757 job = db->job; 758 copied = job->aio_received; 759 placed = len; 760 if (placed > job->uaiocb.aio_nbytes - copied) 761 placed = job->uaiocb.aio_nbytes - copied; 762 if (placed > 0) 763 job->msgrcv = 1; 764 if (!aio_clear_cancel_function(job)) { 765 /* 766 * Update the copied length for when 767 * t4_aio_cancel_active() completes this 768 * request. 769 */ 770 job->aio_received += placed; 771 } else { 772 CTR4(KTR_CXGBE, "%s: tid %d completed buf %d len %d", 773 __func__, toep->tid, db_idx, placed); 774 aio_complete(job, copied + placed, 0); 775 } 776 len -= placed; 777 complete_ddp_buffer(toep, db, db_idx); 778 } 779 780 MPASS(len == 0); 781 ddp_complete_all(toep, 0); 782 } 783 784 #define DDP_ERR (F_DDP_PPOD_MISMATCH | F_DDP_LLIMIT_ERR | F_DDP_ULIMIT_ERR |\ 785 F_DDP_PPOD_PARITY_ERR | F_DDP_PADDING_ERR | F_DDP_OFFSET_ERR |\ 786 F_DDP_INVALID_TAG | F_DDP_COLOR_ERR | F_DDP_TID_MISMATCH |\ 787 F_DDP_INVALID_PPOD | F_DDP_HDRCRC_ERR | F_DDP_DATACRC_ERR) 788 789 static int 790 do_rx_data_ddp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 791 { 792 struct adapter *sc = iq->adapter; 793 const struct cpl_rx_data_ddp *cpl = (const void *)(rss + 1); 794 unsigned int tid = GET_TID(cpl); 795 uint32_t vld; 796 struct toepcb *toep = lookup_tid(sc, tid); 797 798 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 799 KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__)); 800 KASSERT(!(toep->flags & TPF_SYNQE), 801 ("%s: toep %p claims to be a synq entry", __func__, toep)); 802 803 vld = be32toh(cpl->ddpvld); 804 if (__predict_false(vld & DDP_ERR)) { 805 panic("%s: DDP error 0x%x (tid %d, toep %p)", 806 __func__, vld, tid, toep); 807 } 808 809 if (toep->ulp_mode == ULP_MODE_ISCSI) { 810 sc->cpl_handler[CPL_RX_ISCSI_DDP](iq, rss, m); 811 return (0); 812 } 813 814 handle_ddp_data(toep, cpl->u.ddp_report, cpl->seq, be16toh(cpl->len)); 815 816 return (0); 817 } 818 819 static int 820 do_rx_ddp_complete(struct sge_iq *iq, const struct rss_header *rss, 821 struct mbuf *m) 822 { 823 struct adapter *sc = iq->adapter; 824 const struct cpl_rx_ddp_complete *cpl = (const void *)(rss + 1); 825 unsigned int tid = GET_TID(cpl); 826 struct toepcb *toep = lookup_tid(sc, tid); 827 828 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 829 KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__)); 830 KASSERT(!(toep->flags & TPF_SYNQE), 831 ("%s: toep %p claims to be a synq entry", __func__, toep)); 832 833 handle_ddp_data(toep, cpl->ddp_report, cpl->rcv_nxt, 0); 834 835 return (0); 836 } 837 838 static void 839 enable_ddp(struct adapter *sc, struct toepcb *toep) 840 { 841 842 KASSERT((toep->ddp_flags & (DDP_ON | DDP_OK | DDP_SC_REQ)) == DDP_OK, 843 ("%s: toep %p has bad ddp_flags 0x%x", 844 __func__, toep, toep->ddp_flags)); 845 846 CTR3(KTR_CXGBE, "%s: tid %u (time %u)", 847 __func__, toep->tid, time_uptime); 848 849 DDP_ASSERT_LOCKED(toep); 850 toep->ddp_flags |= DDP_SC_REQ; 851 t4_set_tcb_field(sc, toep, 1, W_TCB_RX_DDP_FLAGS, 852 V_TF_DDP_OFF(1) | V_TF_DDP_INDICATE_OUT(1) | 853 V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1) | 854 V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_BUF1_VALID(1), 855 V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1)); 856 t4_set_tcb_field(sc, toep, 1, W_TCB_T_FLAGS, 857 V_TF_RCV_COALESCE_ENABLE(1), 0); 858 } 859 860 static int 861 calculate_hcf(int n1, int n2) 862 { 863 int a, b, t; 864 865 if (n1 <= n2) { 866 a = n1; 867 b = n2; 868 } else { 869 a = n2; 870 b = n1; 871 } 872 873 while (a != 0) { 874 t = a; 875 a = b % a; 876 b = t; 877 } 878 879 return (b); 880 } 881 882 static int 883 alloc_page_pods(struct tom_data *td, struct pageset *ps) 884 { 885 int i, hcf, seglen, idx, ppod, nppods; 886 u_int ppod_addr; 887 888 KASSERT(ps->nppods == 0, ("%s: page pods already allocated", __func__)); 889 890 /* 891 * The DDP page size is unrelated to the VM page size. We combine 892 * contiguous physical pages into larger segments to get the best DDP 893 * page size possible. This is the largest of the four sizes in 894 * A_ULP_RX_TDDP_PSZ that evenly divides the HCF of the segment sizes in 895 * the page list. 896 */ 897 hcf = 0; 898 for (i = 0; i < ps->npages; i++) { 899 seglen = PAGE_SIZE; 900 while (i < ps->npages - 1 && 901 ps->pages[i]->phys_addr + PAGE_SIZE == 902 ps->pages[i + 1]->phys_addr) { 903 seglen += PAGE_SIZE; 904 i++; 905 } 906 907 hcf = calculate_hcf(hcf, seglen); 908 if (hcf < t4_ddp_pgsz[1]) { 909 idx = 0; 910 goto have_pgsz; /* give up, short circuit */ 911 } 912 } 913 914 if (hcf % t4_ddp_pgsz[0] != 0) { 915 /* hmmm. This could only happen when PAGE_SIZE < 4K */ 916 KASSERT(PAGE_SIZE < 4096, 917 ("%s: PAGE_SIZE %d, hcf %d", __func__, PAGE_SIZE, hcf)); 918 CTR3(KTR_CXGBE, "%s: PAGE_SIZE %d, hcf %d", 919 __func__, PAGE_SIZE, hcf); 920 return (0); 921 } 922 923 for (idx = nitems(t4_ddp_pgsz) - 1; idx > 0; idx--) { 924 if (hcf % t4_ddp_pgsz[idx] == 0) 925 break; 926 } 927 have_pgsz: 928 MPASS(idx <= M_PPOD_PGSZ); 929 930 nppods = pages_to_nppods(ps->npages, t4_ddp_pgsz[idx]); 931 if (alloc_ppods(td, nppods, &ppod_addr) != 0) { 932 CTR4(KTR_CXGBE, "%s: no pods, nppods %d, npages %d, pgsz %d", 933 __func__, nppods, ps->npages, t4_ddp_pgsz[idx]); 934 return (0); 935 } 936 937 ppod = (ppod_addr - td->ppod_start) / PPOD_SIZE; 938 ps->tag = V_PPOD_PGSZ(idx) | V_PPOD_TAG(ppod); 939 ps->ppod_addr = ppod_addr; 940 ps->nppods = nppods; 941 942 CTR5(KTR_CXGBE, "New page pods. " 943 "ps %p, ddp_pgsz %d, ppod 0x%x, npages %d, nppods %d", 944 ps, t4_ddp_pgsz[idx], ppod, ps->npages, ps->nppods); 945 946 return (1); 947 } 948 949 #define NUM_ULP_TX_SC_IMM_PPODS (256 / PPOD_SIZE) 950 951 static int 952 write_page_pods(struct adapter *sc, struct toepcb *toep, struct pageset *ps) 953 { 954 struct wrqe *wr; 955 struct ulp_mem_io *ulpmc; 956 struct ulptx_idata *ulpsc; 957 struct pagepod *ppod; 958 int i, j, k, n, chunk, len, ddp_pgsz, idx; 959 u_int ppod_addr; 960 uint32_t cmd; 961 962 KASSERT(!(ps->flags & PS_PPODS_WRITTEN), 963 ("%s: page pods already written", __func__)); 964 965 cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE)); 966 if (is_t4(sc)) 967 cmd |= htobe32(F_ULP_MEMIO_ORDER); 968 else 969 cmd |= htobe32(F_T5_ULP_MEMIO_IMM); 970 ddp_pgsz = t4_ddp_pgsz[G_PPOD_PGSZ(ps->tag)]; 971 ppod_addr = ps->ppod_addr; 972 for (i = 0; i < ps->nppods; ppod_addr += chunk) { 973 974 /* How many page pods are we writing in this cycle */ 975 n = min(ps->nppods - i, NUM_ULP_TX_SC_IMM_PPODS); 976 chunk = PPOD_SZ(n); 977 len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16); 978 979 wr = alloc_wrqe(len, toep->ctrlq); 980 if (wr == NULL) 981 return (ENOMEM); /* ok to just bail out */ 982 ulpmc = wrtod(wr); 983 984 INIT_ULPTX_WR(ulpmc, len, 0, 0); 985 ulpmc->cmd = cmd; 986 ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32)); 987 ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16)); 988 ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5)); 989 990 ulpsc = (struct ulptx_idata *)(ulpmc + 1); 991 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM)); 992 ulpsc->len = htobe32(chunk); 993 994 ppod = (struct pagepod *)(ulpsc + 1); 995 for (j = 0; j < n; i++, j++, ppod++) { 996 ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID | 997 V_PPOD_TID(toep->tid) | ps->tag); 998 ppod->len_offset = htobe64(V_PPOD_LEN(ps->len) | 999 V_PPOD_OFST(ps->offset)); 1000 ppod->rsvd = 0; 1001 idx = i * PPOD_PAGES * (ddp_pgsz / PAGE_SIZE); 1002 for (k = 0; k < nitems(ppod->addr); k++) { 1003 if (idx < ps->npages) { 1004 ppod->addr[k] = 1005 htobe64(ps->pages[idx]->phys_addr); 1006 idx += ddp_pgsz / PAGE_SIZE; 1007 } else 1008 ppod->addr[k] = 0; 1009 #if 0 1010 CTR5(KTR_CXGBE, 1011 "%s: tid %d ppod[%d]->addr[%d] = %p", 1012 __func__, toep->tid, i, k, 1013 htobe64(ppod->addr[k])); 1014 #endif 1015 } 1016 1017 } 1018 1019 t4_wrq_tx(sc, wr); 1020 } 1021 ps->flags |= PS_PPODS_WRITTEN; 1022 1023 return (0); 1024 } 1025 1026 static void 1027 wire_pageset(struct pageset *ps) 1028 { 1029 vm_page_t p; 1030 int i; 1031 1032 KASSERT(!(ps->flags & PS_WIRED), ("pageset already wired")); 1033 1034 for (i = 0; i < ps->npages; i++) { 1035 p = ps->pages[i]; 1036 vm_page_lock(p); 1037 vm_page_wire(p); 1038 vm_page_unhold(p); 1039 vm_page_unlock(p); 1040 } 1041 ps->flags |= PS_WIRED; 1042 } 1043 1044 /* 1045 * Prepare a pageset for DDP. This wires the pageset and sets up page 1046 * pods. 1047 */ 1048 static int 1049 prep_pageset(struct adapter *sc, struct toepcb *toep, struct pageset *ps) 1050 { 1051 struct tom_data *td = sc->tom_softc; 1052 1053 if (!(ps->flags & PS_WIRED)) 1054 wire_pageset(ps); 1055 if (ps->nppods == 0 && !alloc_page_pods(td, ps)) { 1056 return (0); 1057 } 1058 if (!(ps->flags & PS_PPODS_WRITTEN) && 1059 write_page_pods(sc, toep, ps) != 0) { 1060 return (0); 1061 } 1062 1063 return (1); 1064 } 1065 1066 void 1067 t4_init_ddp(struct adapter *sc, struct tom_data *td) 1068 { 1069 1070 td->ppod_start = sc->vres.ddp.start; 1071 td->ppod_arena = vmem_create("DDP page pods", sc->vres.ddp.start, 1072 sc->vres.ddp.size, 1, 32, M_FIRSTFIT | M_NOWAIT); 1073 1074 t4_register_cpl_handler(sc, CPL_RX_DATA_DDP, do_rx_data_ddp); 1075 t4_register_cpl_handler(sc, CPL_RX_DDP_COMPLETE, do_rx_ddp_complete); 1076 } 1077 1078 void 1079 t4_uninit_ddp(struct adapter *sc __unused, struct tom_data *td) 1080 { 1081 1082 if (td->ppod_arena != NULL) { 1083 vmem_destroy(td->ppod_arena); 1084 td->ppod_arena = NULL; 1085 } 1086 } 1087 1088 static int 1089 pscmp(struct pageset *ps, struct vmspace *vm, vm_offset_t start, int npages, 1090 int pgoff, int len) 1091 { 1092 1093 if (ps->npages != npages || ps->offset != pgoff || ps->len != len) 1094 return (1); 1095 1096 return (ps->vm != vm || ps->vm_timestamp != vm->vm_map.timestamp); 1097 } 1098 1099 static int 1100 hold_aio(struct toepcb *toep, struct kaiocb *job, struct pageset **pps) 1101 { 1102 struct vmspace *vm; 1103 vm_map_t map; 1104 vm_offset_t start, end, pgoff; 1105 struct pageset *ps; 1106 int n; 1107 1108 DDP_ASSERT_LOCKED(toep); 1109 1110 /* 1111 * The AIO subsystem will cancel and drain all requests before 1112 * permitting a process to exit or exec, so p_vmspace should 1113 * be stable here. 1114 */ 1115 vm = job->userproc->p_vmspace; 1116 map = &vm->vm_map; 1117 start = (uintptr_t)job->uaiocb.aio_buf; 1118 pgoff = start & PAGE_MASK; 1119 end = round_page(start + job->uaiocb.aio_nbytes); 1120 start = trunc_page(start); 1121 1122 if (end - start > MAX_DDP_BUFFER_SIZE) { 1123 /* 1124 * Truncate the request to a short read. 1125 * Alternatively, we could DDP in chunks to the larger 1126 * buffer, but that would be quite a bit more work. 1127 * 1128 * When truncating, round the request down to avoid 1129 * crossing a cache line on the final transaction. 1130 */ 1131 end = rounddown2(start + MAX_DDP_BUFFER_SIZE, CACHE_LINE_SIZE); 1132 #ifdef VERBOSE_TRACES 1133 CTR4(KTR_CXGBE, "%s: tid %d, truncating size from %lu to %lu", 1134 __func__, toep->tid, (unsigned long)job->uaiocb.aio_nbytes, 1135 (unsigned long)(end - (start + pgoff))); 1136 job->uaiocb.aio_nbytes = end - (start + pgoff); 1137 #endif 1138 end = round_page(end); 1139 } 1140 1141 n = atop(end - start); 1142 1143 /* 1144 * Try to reuse a cached pageset. 1145 */ 1146 TAILQ_FOREACH(ps, &toep->ddp_cached_pagesets, link) { 1147 if (pscmp(ps, vm, start, n, pgoff, 1148 job->uaiocb.aio_nbytes) == 0) { 1149 TAILQ_REMOVE(&toep->ddp_cached_pagesets, ps, link); 1150 toep->ddp_cached_count--; 1151 *pps = ps; 1152 return (0); 1153 } 1154 } 1155 1156 /* 1157 * If there are too many cached pagesets to create a new one, 1158 * free a pageset before creating a new one. 1159 */ 1160 KASSERT(toep->ddp_active_count + toep->ddp_cached_count <= 1161 nitems(toep->db), ("%s: too many wired pagesets", __func__)); 1162 if (toep->ddp_active_count + toep->ddp_cached_count == 1163 nitems(toep->db)) { 1164 KASSERT(toep->ddp_cached_count > 0, 1165 ("no cached pageset to free")); 1166 ps = TAILQ_LAST(&toep->ddp_cached_pagesets, pagesetq); 1167 TAILQ_REMOVE(&toep->ddp_cached_pagesets, ps, link); 1168 toep->ddp_cached_count--; 1169 free_pageset(toep->td, ps); 1170 } 1171 DDP_UNLOCK(toep); 1172 1173 /* Create a new pageset. */ 1174 ps = malloc(sizeof(*ps) + n * sizeof(vm_page_t), M_CXGBE, M_WAITOK | 1175 M_ZERO); 1176 ps->pages = (vm_page_t *)(ps + 1); 1177 ps->vm_timestamp = map->timestamp; 1178 ps->npages = vm_fault_quick_hold_pages(map, start, end - start, 1179 VM_PROT_WRITE, ps->pages, n); 1180 1181 DDP_LOCK(toep); 1182 if (ps->npages < 0) { 1183 free(ps, M_CXGBE); 1184 return (EFAULT); 1185 } 1186 1187 KASSERT(ps->npages == n, ("hold_aio: page count mismatch: %d vs %d", 1188 ps->npages, n)); 1189 1190 ps->offset = pgoff; 1191 ps->len = job->uaiocb.aio_nbytes; 1192 atomic_add_int(&vm->vm_refcnt, 1); 1193 ps->vm = vm; 1194 1195 CTR5(KTR_CXGBE, "%s: tid %d, new pageset %p for job %p, npages %d", 1196 __func__, toep->tid, ps, job, ps->npages); 1197 *pps = ps; 1198 return (0); 1199 } 1200 1201 static void 1202 ddp_complete_all(struct toepcb *toep, int error) 1203 { 1204 struct kaiocb *job; 1205 1206 DDP_ASSERT_LOCKED(toep); 1207 while (!TAILQ_EMPTY(&toep->ddp_aiojobq)) { 1208 job = TAILQ_FIRST(&toep->ddp_aiojobq); 1209 TAILQ_REMOVE(&toep->ddp_aiojobq, job, list); 1210 toep->ddp_waiting_count--; 1211 if (aio_clear_cancel_function(job)) 1212 ddp_complete_one(job, error); 1213 } 1214 } 1215 1216 static void 1217 aio_ddp_cancel_one(struct kaiocb *job) 1218 { 1219 long copied; 1220 1221 /* 1222 * If this job had copied data out of the socket buffer before 1223 * it was cancelled, report it as a short read rather than an 1224 * error. 1225 */ 1226 copied = job->aio_received; 1227 if (copied != 0) 1228 aio_complete(job, copied, 0); 1229 else 1230 aio_cancel(job); 1231 } 1232 1233 /* 1234 * Called when the main loop wants to requeue a job to retry it later. 1235 * Deals with the race of the job being cancelled while it was being 1236 * examined. 1237 */ 1238 static void 1239 aio_ddp_requeue_one(struct toepcb *toep, struct kaiocb *job) 1240 { 1241 1242 DDP_ASSERT_LOCKED(toep); 1243 if (!(toep->ddp_flags & DDP_DEAD) && 1244 aio_set_cancel_function(job, t4_aio_cancel_queued)) { 1245 TAILQ_INSERT_HEAD(&toep->ddp_aiojobq, job, list); 1246 toep->ddp_waiting_count++; 1247 } else 1248 aio_ddp_cancel_one(job); 1249 } 1250 1251 static void 1252 aio_ddp_requeue(struct toepcb *toep) 1253 { 1254 struct adapter *sc = td_adapter(toep->td); 1255 struct socket *so; 1256 struct sockbuf *sb; 1257 struct inpcb *inp; 1258 struct kaiocb *job; 1259 struct ddp_buffer *db; 1260 size_t copied, offset, resid; 1261 struct pageset *ps; 1262 struct mbuf *m; 1263 uint64_t ddp_flags, ddp_flags_mask; 1264 struct wrqe *wr; 1265 int buf_flag, db_idx, error; 1266 1267 DDP_ASSERT_LOCKED(toep); 1268 1269 restart: 1270 if (toep->ddp_flags & DDP_DEAD) { 1271 MPASS(toep->ddp_waiting_count == 0); 1272 MPASS(toep->ddp_active_count == 0); 1273 return; 1274 } 1275 1276 if (toep->ddp_waiting_count == 0 || 1277 toep->ddp_active_count == nitems(toep->db)) { 1278 return; 1279 } 1280 1281 job = TAILQ_FIRST(&toep->ddp_aiojobq); 1282 so = job->fd_file->f_data; 1283 sb = &so->so_rcv; 1284 SOCKBUF_LOCK(sb); 1285 1286 /* We will never get anything unless we are or were connected. */ 1287 if (!(so->so_state & (SS_ISCONNECTED|SS_ISDISCONNECTED))) { 1288 SOCKBUF_UNLOCK(sb); 1289 ddp_complete_all(toep, ENOTCONN); 1290 return; 1291 } 1292 1293 KASSERT(toep->ddp_active_count == 0 || sbavail(sb) == 0, 1294 ("%s: pending sockbuf data and DDP is active", __func__)); 1295 1296 /* Abort if socket has reported problems. */ 1297 /* XXX: Wait for any queued DDP's to finish and/or flush them? */ 1298 if (so->so_error && sbavail(sb) == 0) { 1299 toep->ddp_waiting_count--; 1300 TAILQ_REMOVE(&toep->ddp_aiojobq, job, list); 1301 if (!aio_clear_cancel_function(job)) { 1302 SOCKBUF_UNLOCK(sb); 1303 goto restart; 1304 } 1305 1306 /* 1307 * If this job has previously copied some data, report 1308 * a short read and leave the error to be reported by 1309 * a future request. 1310 */ 1311 copied = job->aio_received; 1312 if (copied != 0) { 1313 SOCKBUF_UNLOCK(sb); 1314 aio_complete(job, copied, 0); 1315 goto restart; 1316 } 1317 error = so->so_error; 1318 so->so_error = 0; 1319 SOCKBUF_UNLOCK(sb); 1320 aio_complete(job, -1, error); 1321 goto restart; 1322 } 1323 1324 /* 1325 * Door is closed. If there is pending data in the socket buffer, 1326 * deliver it. If there are pending DDP requests, wait for those 1327 * to complete. Once they have completed, return EOF reads. 1328 */ 1329 if (sb->sb_state & SBS_CANTRCVMORE && sbavail(sb) == 0) { 1330 SOCKBUF_UNLOCK(sb); 1331 if (toep->ddp_active_count != 0) 1332 return; 1333 ddp_complete_all(toep, 0); 1334 return; 1335 } 1336 1337 /* 1338 * If DDP is not enabled and there is no pending socket buffer 1339 * data, try to enable DDP. 1340 */ 1341 if (sbavail(sb) == 0 && (toep->ddp_flags & DDP_ON) == 0) { 1342 SOCKBUF_UNLOCK(sb); 1343 1344 /* 1345 * Wait for the card to ACK that DDP is enabled before 1346 * queueing any buffers. Currently this waits for an 1347 * indicate to arrive. This could use a TCB_SET_FIELD_RPL 1348 * message to know that DDP was enabled instead of waiting 1349 * for the indicate which would avoid copying the indicate 1350 * if no data is pending. 1351 * 1352 * XXX: Might want to limit the indicate size to the size 1353 * of the first queued request. 1354 */ 1355 if ((toep->ddp_flags & DDP_SC_REQ) == 0) 1356 enable_ddp(sc, toep); 1357 return; 1358 } 1359 SOCKBUF_UNLOCK(sb); 1360 1361 /* 1362 * If another thread is queueing a buffer for DDP, let it 1363 * drain any work and return. 1364 */ 1365 if (toep->ddp_queueing != NULL) 1366 return; 1367 1368 /* Take the next job to prep it for DDP. */ 1369 toep->ddp_waiting_count--; 1370 TAILQ_REMOVE(&toep->ddp_aiojobq, job, list); 1371 if (!aio_clear_cancel_function(job)) 1372 goto restart; 1373 toep->ddp_queueing = job; 1374 1375 /* NB: This drops DDP_LOCK while it holds the backing VM pages. */ 1376 error = hold_aio(toep, job, &ps); 1377 if (error != 0) { 1378 ddp_complete_one(job, error); 1379 toep->ddp_queueing = NULL; 1380 goto restart; 1381 } 1382 1383 SOCKBUF_LOCK(sb); 1384 if (so->so_error && sbavail(sb) == 0) { 1385 copied = job->aio_received; 1386 if (copied != 0) { 1387 SOCKBUF_UNLOCK(sb); 1388 recycle_pageset(toep, ps); 1389 aio_complete(job, copied, 0); 1390 toep->ddp_queueing = NULL; 1391 goto restart; 1392 } 1393 1394 error = so->so_error; 1395 so->so_error = 0; 1396 SOCKBUF_UNLOCK(sb); 1397 recycle_pageset(toep, ps); 1398 aio_complete(job, -1, error); 1399 toep->ddp_queueing = NULL; 1400 goto restart; 1401 } 1402 1403 if (sb->sb_state & SBS_CANTRCVMORE && sbavail(sb) == 0) { 1404 SOCKBUF_UNLOCK(sb); 1405 recycle_pageset(toep, ps); 1406 if (toep->ddp_active_count != 0) { 1407 /* 1408 * The door is closed, but there are still pending 1409 * DDP buffers. Requeue. These jobs will all be 1410 * completed once those buffers drain. 1411 */ 1412 aio_ddp_requeue_one(toep, job); 1413 toep->ddp_queueing = NULL; 1414 return; 1415 } 1416 ddp_complete_one(job, 0); 1417 ddp_complete_all(toep, 0); 1418 toep->ddp_queueing = NULL; 1419 return; 1420 } 1421 1422 sbcopy: 1423 /* 1424 * If the toep is dead, there shouldn't be any data in the socket 1425 * buffer, so the above case should have handled this. 1426 */ 1427 MPASS(!(toep->ddp_flags & DDP_DEAD)); 1428 1429 /* 1430 * If there is pending data in the socket buffer (either 1431 * from before the requests were queued or a DDP indicate), 1432 * copy those mbufs out directly. 1433 */ 1434 copied = 0; 1435 offset = ps->offset + job->aio_received; 1436 MPASS(job->aio_received <= job->uaiocb.aio_nbytes); 1437 resid = job->uaiocb.aio_nbytes - job->aio_received; 1438 m = sb->sb_mb; 1439 KASSERT(m == NULL || toep->ddp_active_count == 0, 1440 ("%s: sockbuf data with active DDP", __func__)); 1441 while (m != NULL && resid > 0) { 1442 struct iovec iov[1]; 1443 struct uio uio; 1444 int error; 1445 1446 iov[0].iov_base = mtod(m, void *); 1447 iov[0].iov_len = m->m_len; 1448 if (iov[0].iov_len > resid) 1449 iov[0].iov_len = resid; 1450 uio.uio_iov = iov; 1451 uio.uio_iovcnt = 1; 1452 uio.uio_offset = 0; 1453 uio.uio_resid = iov[0].iov_len; 1454 uio.uio_segflg = UIO_SYSSPACE; 1455 uio.uio_rw = UIO_WRITE; 1456 error = uiomove_fromphys(ps->pages, offset + copied, 1457 uio.uio_resid, &uio); 1458 MPASS(error == 0 && uio.uio_resid == 0); 1459 copied += uio.uio_offset; 1460 resid -= uio.uio_offset; 1461 m = m->m_next; 1462 } 1463 if (copied != 0) { 1464 sbdrop_locked(sb, copied); 1465 job->aio_received += copied; 1466 job->msgrcv = 1; 1467 copied = job->aio_received; 1468 inp = sotoinpcb(so); 1469 if (!INP_TRY_WLOCK(inp)) { 1470 /* 1471 * The reference on the socket file descriptor in 1472 * the AIO job should keep 'sb' and 'inp' stable. 1473 * Our caller has a reference on the 'toep' that 1474 * keeps it stable. 1475 */ 1476 SOCKBUF_UNLOCK(sb); 1477 DDP_UNLOCK(toep); 1478 INP_WLOCK(inp); 1479 DDP_LOCK(toep); 1480 SOCKBUF_LOCK(sb); 1481 1482 /* 1483 * If the socket has been closed, we should detect 1484 * that and complete this request if needed on 1485 * the next trip around the loop. 1486 */ 1487 } 1488 t4_rcvd_locked(&toep->td->tod, intotcpcb(inp)); 1489 INP_WUNLOCK(inp); 1490 if (resid == 0 || toep->ddp_flags & DDP_DEAD) { 1491 /* 1492 * We filled the entire buffer with socket 1493 * data, DDP is not being used, or the socket 1494 * is being shut down, so complete the 1495 * request. 1496 */ 1497 SOCKBUF_UNLOCK(sb); 1498 recycle_pageset(toep, ps); 1499 aio_complete(job, copied, 0); 1500 toep->ddp_queueing = NULL; 1501 goto restart; 1502 } 1503 1504 /* 1505 * If DDP is not enabled, requeue this request and restart. 1506 * This will either enable DDP or wait for more data to 1507 * arrive on the socket buffer. 1508 */ 1509 if ((toep->ddp_flags & (DDP_ON | DDP_SC_REQ)) != DDP_ON) { 1510 SOCKBUF_UNLOCK(sb); 1511 recycle_pageset(toep, ps); 1512 aio_ddp_requeue_one(toep, job); 1513 toep->ddp_queueing = NULL; 1514 goto restart; 1515 } 1516 1517 /* 1518 * An indicate might have arrived and been added to 1519 * the socket buffer while it was unlocked after the 1520 * copy to lock the INP. If so, restart the copy. 1521 */ 1522 if (sbavail(sb) != 0) 1523 goto sbcopy; 1524 } 1525 SOCKBUF_UNLOCK(sb); 1526 1527 if (prep_pageset(sc, toep, ps) == 0) { 1528 recycle_pageset(toep, ps); 1529 aio_ddp_requeue_one(toep, job); 1530 toep->ddp_queueing = NULL; 1531 1532 /* 1533 * XXX: Need to retry this later. Mostly need a trigger 1534 * when page pods are freed up. 1535 */ 1536 printf("%s: prep_pageset failed\n", __func__); 1537 return; 1538 } 1539 1540 /* Determine which DDP buffer to use. */ 1541 if (toep->db[0].job == NULL) { 1542 db_idx = 0; 1543 } else { 1544 MPASS(toep->db[1].job == NULL); 1545 db_idx = 1; 1546 } 1547 1548 ddp_flags = 0; 1549 ddp_flags_mask = 0; 1550 if (db_idx == 0) { 1551 ddp_flags |= V_TF_DDP_BUF0_VALID(1); 1552 if (so->so_state & SS_NBIO) 1553 ddp_flags |= V_TF_DDP_BUF0_FLUSH(1); 1554 ddp_flags_mask |= V_TF_DDP_PSH_NO_INVALIDATE0(1) | 1555 V_TF_DDP_PUSH_DISABLE_0(1) | V_TF_DDP_PSHF_ENABLE_0(1) | 1556 V_TF_DDP_BUF0_FLUSH(1) | V_TF_DDP_BUF0_VALID(1); 1557 buf_flag = DDP_BUF0_ACTIVE; 1558 } else { 1559 ddp_flags |= V_TF_DDP_BUF1_VALID(1); 1560 if (so->so_state & SS_NBIO) 1561 ddp_flags |= V_TF_DDP_BUF1_FLUSH(1); 1562 ddp_flags_mask |= V_TF_DDP_PSH_NO_INVALIDATE1(1) | 1563 V_TF_DDP_PUSH_DISABLE_1(1) | V_TF_DDP_PSHF_ENABLE_1(1) | 1564 V_TF_DDP_BUF1_FLUSH(1) | V_TF_DDP_BUF1_VALID(1); 1565 buf_flag = DDP_BUF1_ACTIVE; 1566 } 1567 MPASS((toep->ddp_flags & buf_flag) == 0); 1568 if ((toep->ddp_flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE)) == 0) { 1569 MPASS(db_idx == 0); 1570 MPASS(toep->ddp_active_id == -1); 1571 MPASS(toep->ddp_active_count == 0); 1572 ddp_flags_mask |= V_TF_DDP_ACTIVE_BUF(1); 1573 } 1574 1575 /* 1576 * The TID for this connection should still be valid. If DDP_DEAD 1577 * is set, SBS_CANTRCVMORE should be set, so we shouldn't be 1578 * this far anyway. Even if the socket is closing on the other 1579 * end, the AIO job holds a reference on this end of the socket 1580 * which will keep it open and keep the TCP PCB attached until 1581 * after the job is completed. 1582 */ 1583 wr = mk_update_tcb_for_ddp(sc, toep, db_idx, ps, job->aio_received, 1584 ddp_flags, ddp_flags_mask); 1585 if (wr == NULL) { 1586 recycle_pageset(toep, ps); 1587 aio_ddp_requeue_one(toep, job); 1588 toep->ddp_queueing = NULL; 1589 1590 /* 1591 * XXX: Need a way to kick a retry here. 1592 * 1593 * XXX: We know the fixed size needed and could 1594 * preallocate this using a blocking request at the 1595 * start of the task to avoid having to handle this 1596 * edge case. 1597 */ 1598 printf("%s: mk_update_tcb_for_ddp failed\n", __func__); 1599 return; 1600 } 1601 1602 if (!aio_set_cancel_function(job, t4_aio_cancel_active)) { 1603 free_wrqe(wr); 1604 recycle_pageset(toep, ps); 1605 aio_ddp_cancel_one(job); 1606 toep->ddp_queueing = NULL; 1607 goto restart; 1608 } 1609 1610 #ifdef VERBOSE_TRACES 1611 CTR5(KTR_CXGBE, "%s: scheduling %p for DDP[%d] (flags %#lx/%#lx)", 1612 __func__, job, db_idx, ddp_flags, ddp_flags_mask); 1613 #endif 1614 /* Give the chip the go-ahead. */ 1615 t4_wrq_tx(sc, wr); 1616 db = &toep->db[db_idx]; 1617 db->cancel_pending = 0; 1618 db->job = job; 1619 db->ps = ps; 1620 toep->ddp_queueing = NULL; 1621 toep->ddp_flags |= buf_flag; 1622 toep->ddp_active_count++; 1623 if (toep->ddp_active_count == 1) { 1624 MPASS(toep->ddp_active_id == -1); 1625 toep->ddp_active_id = db_idx; 1626 CTR2(KTR_CXGBE, "%s: ddp_active_id = %d", __func__, 1627 toep->ddp_active_id); 1628 } 1629 goto restart; 1630 } 1631 1632 void 1633 ddp_queue_toep(struct toepcb *toep) 1634 { 1635 1636 DDP_ASSERT_LOCKED(toep); 1637 if (toep->ddp_flags & DDP_TASK_ACTIVE) 1638 return; 1639 toep->ddp_flags |= DDP_TASK_ACTIVE; 1640 hold_toepcb(toep); 1641 soaio_enqueue(&toep->ddp_requeue_task); 1642 } 1643 1644 static void 1645 aio_ddp_requeue_task(void *context, int pending) 1646 { 1647 struct toepcb *toep = context; 1648 1649 DDP_LOCK(toep); 1650 aio_ddp_requeue(toep); 1651 toep->ddp_flags &= ~DDP_TASK_ACTIVE; 1652 DDP_UNLOCK(toep); 1653 1654 free_toepcb(toep); 1655 } 1656 1657 static void 1658 t4_aio_cancel_active(struct kaiocb *job) 1659 { 1660 struct socket *so = job->fd_file->f_data; 1661 struct tcpcb *tp = so_sototcpcb(so); 1662 struct toepcb *toep = tp->t_toe; 1663 struct adapter *sc = td_adapter(toep->td); 1664 uint64_t valid_flag; 1665 int i; 1666 1667 DDP_LOCK(toep); 1668 if (aio_cancel_cleared(job)) { 1669 DDP_UNLOCK(toep); 1670 aio_ddp_cancel_one(job); 1671 return; 1672 } 1673 1674 for (i = 0; i < nitems(toep->db); i++) { 1675 if (toep->db[i].job == job) { 1676 /* Should only ever get one cancel request for a job. */ 1677 MPASS(toep->db[i].cancel_pending == 0); 1678 1679 /* 1680 * Invalidate this buffer. It will be 1681 * cancelled or partially completed once the 1682 * card ACKs the invalidate. 1683 */ 1684 valid_flag = i == 0 ? V_TF_DDP_BUF0_VALID(1) : 1685 V_TF_DDP_BUF1_VALID(1); 1686 t4_set_tcb_field_rpl(sc, toep, 1, W_TCB_RX_DDP_FLAGS, 1687 valid_flag, 0, i + DDP_BUF0_INVALIDATED); 1688 toep->db[i].cancel_pending = 1; 1689 CTR2(KTR_CXGBE, "%s: request %p marked pending", 1690 __func__, job); 1691 break; 1692 } 1693 } 1694 DDP_UNLOCK(toep); 1695 } 1696 1697 static void 1698 t4_aio_cancel_queued(struct kaiocb *job) 1699 { 1700 struct socket *so = job->fd_file->f_data; 1701 struct tcpcb *tp = so_sototcpcb(so); 1702 struct toepcb *toep = tp->t_toe; 1703 1704 DDP_LOCK(toep); 1705 if (!aio_cancel_cleared(job)) { 1706 TAILQ_REMOVE(&toep->ddp_aiojobq, job, list); 1707 toep->ddp_waiting_count--; 1708 if (toep->ddp_waiting_count == 0) 1709 ddp_queue_toep(toep); 1710 } 1711 CTR2(KTR_CXGBE, "%s: request %p cancelled", __func__, job); 1712 DDP_UNLOCK(toep); 1713 1714 aio_ddp_cancel_one(job); 1715 } 1716 1717 int 1718 t4_aio_queue_ddp(struct socket *so, struct kaiocb *job) 1719 { 1720 struct tcpcb *tp = so_sototcpcb(so); 1721 struct toepcb *toep = tp->t_toe; 1722 1723 1724 /* Ignore writes. */ 1725 if (job->uaiocb.aio_lio_opcode != LIO_READ) 1726 return (EOPNOTSUPP); 1727 1728 DDP_LOCK(toep); 1729 1730 /* 1731 * XXX: Think about possibly returning errors for ENOTCONN, 1732 * etc. Perhaps the caller would only queue the request 1733 * if it failed with EOPNOTSUPP? 1734 */ 1735 1736 #ifdef VERBOSE_TRACES 1737 CTR2(KTR_CXGBE, "%s: queueing %p", __func__, job); 1738 #endif 1739 if (!aio_set_cancel_function(job, t4_aio_cancel_queued)) 1740 panic("new job was cancelled"); 1741 TAILQ_INSERT_TAIL(&toep->ddp_aiojobq, job, list); 1742 toep->ddp_waiting_count++; 1743 toep->ddp_flags |= DDP_OK; 1744 1745 /* 1746 * Try to handle this request synchronously. If this has 1747 * to block because the task is running, it will just bail 1748 * and let the task handle it instead. 1749 */ 1750 aio_ddp_requeue(toep); 1751 DDP_UNLOCK(toep); 1752 return (0); 1753 } 1754 1755 int 1756 t4_ddp_mod_load(void) 1757 { 1758 1759 TAILQ_INIT(&ddp_orphan_pagesets); 1760 mtx_init(&ddp_orphan_pagesets_lock, "ddp orphans", NULL, MTX_DEF); 1761 TASK_INIT(&ddp_orphan_task, 0, ddp_free_orphan_pagesets, NULL); 1762 return (0); 1763 } 1764 1765 void 1766 t4_ddp_mod_unload(void) 1767 { 1768 1769 taskqueue_drain(taskqueue_thread, &ddp_orphan_task); 1770 MPASS(TAILQ_EMPTY(&ddp_orphan_pagesets)); 1771 mtx_destroy(&ddp_orphan_pagesets_lock); 1772 } 1773 #endif 1774