sys/dev/cxgbe/tom/t4_ddp.c

   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 #include <cam/scsi/scsi_all.h>
  66 #include <cam/ctl/ctl_io.h>
  67
  68 #ifdef TCP_OFFLOAD
  69 #include "common/common.h"
  70 #include "common/t4_msg.h"
  71 #include "common/t4_regs.h"
  72 #include "common/t4_tcb.h"
  73 #include "tom/t4_tom.h"
  74
  75 /*
  76  * Use the 'backend3' field in AIO jobs to store the amount of data
  77  * received by the AIO job so far.
  78  */
  79 #define aio_received    backend3
  80
  81 static void aio_ddp_requeue_task(void *context, int pending);
  82 static void ddp_complete_all(struct toepcb *toep, int error);
  83 static void t4_aio_cancel_active(struct kaiocb *job);
  84 static void t4_aio_cancel_queued(struct kaiocb *job);
  85
  86 static TAILQ_HEAD(, pageset) ddp_orphan_pagesets;
  87 static struct mtx ddp_orphan_pagesets_lock;
  88 static struct task ddp_orphan_task;
  89
  90 #define MAX_DDP_BUFFER_SIZE             (M_TCB_RX_DDP_BUF0_LEN)
  91
  92 /*
  93  * A page set holds information about a buffer used for DDP.  The page
  94  * set holds resources such as the VM pages backing the buffer (either
  95  * held or wired) and the page pods associated with the buffer.
  96  * Recently used page sets are cached to allow for efficient reuse of
  97  * buffers (avoiding the need to re-fault in pages, hold them, etc.).
  98  * Note that cached page sets keep the backing pages wired.  The
  99  * number of wired pages is capped by only allowing for two wired
 100  * pagesets per connection.  This is not a perfect cap, but is a
 101  * trade-off for performance.
 102  *
 103  * If an application ping-pongs two buffers for a connection via
 104  * aio_read(2) then those buffers should remain wired and expensive VM
 105  * fault lookups should be avoided after each buffer has been used
 106  * once.  If an application uses more than two buffers then this will
 107  * fall back to doing expensive VM fault lookups for each operation.
 108  */
 109 static void
 110 free_pageset(struct tom_data *td, struct pageset *ps)
 111 {
 112         vm_page_t p;
 113         int i;
 114
 115         if (ps->prsv.prsv_nppods > 0)
 116                 t4_free_page_pods(&ps->prsv);
 117
 118         for (i = 0; i < ps->npages; i++) {
 119                 p = ps->pages[i];
 120                 vm_page_unwire(p, PQ_INACTIVE);
 121         }
 122         mtx_lock(&ddp_orphan_pagesets_lock);
 123         TAILQ_INSERT_TAIL(&ddp_orphan_pagesets, ps, link);
 124         taskqueue_enqueue(taskqueue_thread, &ddp_orphan_task);
 125         mtx_unlock(&ddp_orphan_pagesets_lock);
 126 }
 127
 128 static void
 129 ddp_free_orphan_pagesets(void *context, int pending)
 130 {
 131         struct pageset *ps;
 132
 133         mtx_lock(&ddp_orphan_pagesets_lock);
 134         while (!TAILQ_EMPTY(&ddp_orphan_pagesets)) {
 135                 ps = TAILQ_FIRST(&ddp_orphan_pagesets);
 136                 TAILQ_REMOVE(&ddp_orphan_pagesets, ps, link);
 137                 mtx_unlock(&ddp_orphan_pagesets_lock);
 138                 if (ps->vm)
 139                         vmspace_free(ps->vm);
 140                 free(ps, M_CXGBE);
 141                 mtx_lock(&ddp_orphan_pagesets_lock);
 142         }
 143         mtx_unlock(&ddp_orphan_pagesets_lock);
 144 }
 145
 146 static void
 147 recycle_pageset(struct toepcb *toep, struct pageset *ps)
 148 {
 149
 150         DDP_ASSERT_LOCKED(toep);
 151         if (!(toep->ddp.flags & DDP_DEAD)) {
 152                 KASSERT(toep->ddp.cached_count + toep->ddp.active_count <
 153                     nitems(toep->ddp.db), ("too many wired pagesets"));
 154                 TAILQ_INSERT_HEAD(&toep->ddp.cached_pagesets, ps, link);
 155                 toep->ddp.cached_count++;
 156         } else
 157                 free_pageset(toep->td, ps);
 158 }
 159
 160 static void
 161 ddp_complete_one(struct kaiocb *job, int error)
 162 {
 163         long copied;
 164
 165         /*
 166          * If this job had copied data out of the socket buffer before
 167          * it was cancelled, report it as a short read rather than an
 168          * error.
 169          */
 170         copied = job->aio_received;
 171         if (copied != 0 || error == 0)
 172                 aio_complete(job, copied, 0);
 173         else
 174                 aio_complete(job, -1, error);
 175 }
 176
 177 static void
 178 free_ddp_buffer(struct tom_data *td, struct ddp_buffer *db)
 179 {
 180
 181         if (db->job) {
 182                 /*
 183                  * XXX: If we are un-offloading the socket then we
 184                  * should requeue these on the socket somehow.  If we
 185                  * got a FIN from the remote end, then this completes
 186                  * any remaining requests with an EOF read.
 187                  */
 188                 if (!aio_clear_cancel_function(db->job))
 189                         ddp_complete_one(db->job, 0);
 190         }
 191
 192         if (db->ps)
 193                 free_pageset(td, db->ps);
 194 }
 195
 196 void
 197 ddp_init_toep(struct toepcb *toep)
 198 {
 199
 200         TAILQ_INIT(&toep->ddp.aiojobq);
 201         TASK_INIT(&toep->ddp.requeue_task, 0, aio_ddp_requeue_task, toep);
 202         toep->ddp.flags = DDP_OK;
 203         toep->ddp.active_id = -1;
 204         mtx_init(&toep->ddp.lock, "t4 ddp", NULL, MTX_DEF);
 205 }
 206
 207 void
 208 ddp_uninit_toep(struct toepcb *toep)
 209 {
 210
 211         mtx_destroy(&toep->ddp.lock);
 212 }
 213
 214 void
 215 release_ddp_resources(struct toepcb *toep)
 216 {
 217         struct pageset *ps;
 218         int i;
 219
 220         DDP_LOCK(toep);
 221         toep->ddp.flags |= DDP_DEAD;
 222         for (i = 0; i < nitems(toep->ddp.db); i++) {
 223                 free_ddp_buffer(toep->td, &toep->ddp.db[i]);
 224         }
 225         while ((ps = TAILQ_FIRST(&toep->ddp.cached_pagesets)) != NULL) {
 226                 TAILQ_REMOVE(&toep->ddp.cached_pagesets, ps, link);
 227                 free_pageset(toep->td, ps);
 228         }
 229         ddp_complete_all(toep, 0);
 230         DDP_UNLOCK(toep);
 231 }
 232
 233 #ifdef INVARIANTS
 234 void
 235 ddp_assert_empty(struct toepcb *toep)
 236 {
 237         int i;
 238
 239         MPASS(!(toep->ddp.flags & DDP_TASK_ACTIVE));
 240         for (i = 0; i < nitems(toep->ddp.db); i++) {
 241                 MPASS(toep->ddp.db[i].job == NULL);
 242                 MPASS(toep->ddp.db[i].ps == NULL);
 243         }
 244         MPASS(TAILQ_EMPTY(&toep->ddp.cached_pagesets));
 245         MPASS(TAILQ_EMPTY(&toep->ddp.aiojobq));
 246 }
 247 #endif
 248
 249 static void
 250 complete_ddp_buffer(struct toepcb *toep, struct ddp_buffer *db,
 251     unsigned int db_idx)
 252 {
 253         unsigned int db_flag;
 254
 255         toep->ddp.active_count--;
 256         if (toep->ddp.active_id == db_idx) {
 257                 if (toep->ddp.active_count == 0) {
 258                         KASSERT(toep->ddp.db[db_idx ^ 1].job == NULL,
 259                             ("%s: active_count mismatch", __func__));
 260                         toep->ddp.active_id = -1;
 261                 } else
 262                         toep->ddp.active_id ^= 1;
 263 #ifdef VERBOSE_TRACES
 264                 CTR3(KTR_CXGBE, "%s: tid %u, ddp_active_id = %d", __func__,
 265                     toep->tid, toep->ddp.active_id);
 266 #endif
 267         } else {
 268                 KASSERT(toep->ddp.active_count != 0 &&
 269                     toep->ddp.active_id != -1,
 270                     ("%s: active count mismatch", __func__));
 271         }
 272
 273         db->cancel_pending = 0;
 274         db->job = NULL;
 275         recycle_pageset(toep, db->ps);
 276         db->ps = NULL;
 277
 278         db_flag = db_idx == 1 ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE;
 279         KASSERT(toep->ddp.flags & db_flag,
 280             ("%s: DDP buffer not active. toep %p, ddp_flags 0x%x",
 281             __func__, toep, toep->ddp.flags));
 282         toep->ddp.flags &= ~db_flag;
 283 }
 284
 285 /* XXX: handle_ddp_data code duplication */
 286 void
 287 insert_ddp_data(struct toepcb *toep, uint32_t n)
 288 {
 289         struct inpcb *inp = toep->inp;
 290         struct tcpcb *tp = intotcpcb(inp);
 291         struct ddp_buffer *db;
 292         struct kaiocb *job;
 293         size_t placed;
 294         long copied;
 295         unsigned int db_idx;
 296 #ifdef INVARIANTS
 297         unsigned int db_flag;
 298 #endif
 299
 300         INP_WLOCK_ASSERT(inp);
 301         DDP_ASSERT_LOCKED(toep);
 302
 303         tp->rcv_nxt += n;
 304 #ifndef USE_DDP_RX_FLOW_CONTROL
 305         KASSERT(tp->rcv_wnd >= n, ("%s: negative window size", __func__));
 306         tp->rcv_wnd -= n;
 307 #endif
 308         CTR2(KTR_CXGBE, "%s: placed %u bytes before falling out of DDP",
 309             __func__, n);
 310         while (toep->ddp.active_count > 0) {
 311                 MPASS(toep->ddp.active_id != -1);
 312                 db_idx = toep->ddp.active_id;
 313 #ifdef INVARIANTS
 314                 db_flag = db_idx == 1 ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE;
 315 #endif
 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         CTR5(KTR_CXGBE, "%s: tid %u, DDP[%d] placed %d bytes (%#x)", __func__,
 541             toep->tid, db_idx, 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                 struct adapter *sc = td_adapter(toep->td);
 553                 unsigned int hiwat = sb->sb_hiwat;
 554                 unsigned int newsize = min(hiwat + sc->tt.autorcvbuf_inc,
 555                     V_tcp_autorcvbuf_max);
 556
 557                 if (!sbreserve_locked(sb, newsize, so, NULL))
 558                         sb->sb_flags &= ~SB_AUTOSIZE;
 559         }
 560         SOCKBUF_UNLOCK(sb);
 561         CURVNET_RESTORE();
 562
 563         job->msgrcv = 1;
 564         if (db->cancel_pending) {
 565                 /*
 566                  * Update the job's length but defer completion to the
 567                  * TCB_RPL callback.
 568                  */
 569                 job->aio_received += len;
 570                 goto out;
 571         } else if (!aio_clear_cancel_function(job)) {
 572                 /*
 573                  * Update the copied length for when
 574                  * t4_aio_cancel_active() completes this request.
 575                  */
 576                 job->aio_received += len;
 577         } else {
 578                 copied = job->aio_received;
 579 #ifdef VERBOSE_TRACES
 580                 CTR5(KTR_CXGBE,
 581                     "%s: tid %u, completing %p (copied %ld, placed %d)",
 582                     __func__, toep->tid, job, copied, len);
 583 #endif
 584                 aio_complete(job, copied + len, 0);
 585                 t4_rcvd(&toep->td->tod, tp);
 586         }
 587
 588 completed:
 589         complete_ddp_buffer(toep, db, db_idx);
 590         if (toep->ddp.waiting_count > 0)
 591                 ddp_queue_toep(toep);
 592 out:
 593         DDP_UNLOCK(toep);
 594         INP_WUNLOCK(inp);
 595
 596         return (0);
 597 }
 598
 599 void
 600 handle_ddp_indicate(struct toepcb *toep)
 601 {
 602
 603         DDP_ASSERT_LOCKED(toep);
 604         MPASS(toep->ddp.active_count == 0);
 605         MPASS((toep->ddp.flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE)) == 0);
 606         if (toep->ddp.waiting_count == 0) {
 607                 /*
 608                  * The pending requests that triggered the request for an
 609                  * an indicate were cancelled.  Those cancels should have
 610                  * already disabled DDP.  Just ignore this as the data is
 611                  * going into the socket buffer anyway.
 612                  */
 613                 return;
 614         }
 615         CTR3(KTR_CXGBE, "%s: tid %d indicated (%d waiting)", __func__,
 616             toep->tid, toep->ddp.waiting_count);
 617         ddp_queue_toep(toep);
 618 }
 619
 620 CTASSERT(CPL_COOKIE_DDP0 + 1 == CPL_COOKIE_DDP1);
 621
 622 static int
 623 do_ddp_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
 624 {
 625         struct adapter *sc = iq->adapter;
 626         const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
 627         unsigned int tid = GET_TID(cpl);
 628         unsigned int db_idx;
 629         struct toepcb *toep;
 630         struct inpcb *inp;
 631         struct ddp_buffer *db;
 632         struct kaiocb *job;
 633         long copied;
 634
 635         if (cpl->status != CPL_ERR_NONE)
 636                 panic("XXX: tcp_rpl failed: %d", cpl->status);
 637
 638         toep = lookup_tid(sc, tid);
 639         inp = toep->inp;
 640         switch (cpl->cookie) {
 641         case V_WORD(W_TCB_RX_DDP_FLAGS) | V_COOKIE(CPL_COOKIE_DDP0):
 642         case V_WORD(W_TCB_RX_DDP_FLAGS) | V_COOKIE(CPL_COOKIE_DDP1):
 643                 /*
 644                  * XXX: This duplicates a lot of code with handle_ddp_data().
 645                  */
 646                 db_idx = G_COOKIE(cpl->cookie) - CPL_COOKIE_DDP0;
 647                 MPASS(db_idx < nitems(toep->ddp.db));
 648                 INP_WLOCK(inp);
 649                 DDP_LOCK(toep);
 650                 db = &toep->ddp.db[db_idx];
 651
 652                 /*
 653                  * handle_ddp_data() should leave the job around until
 654                  * this callback runs once a cancel is pending.
 655                  */
 656                 MPASS(db != NULL);
 657                 MPASS(db->job != NULL);
 658                 MPASS(db->cancel_pending);
 659
 660                 /*
 661                  * XXX: It's not clear what happens if there is data
 662                  * placed when the buffer is invalidated.  I suspect we
 663                  * need to read the TCB to see how much data was placed.
 664                  *
 665                  * For now this just pretends like nothing was placed.
 666                  *
 667                  * XXX: Note that if we did check the PCB we would need to
 668                  * also take care of updating the tp, etc.
 669                  */
 670                 job = db->job;
 671                 copied = job->aio_received;
 672                 if (copied == 0) {
 673                         CTR2(KTR_CXGBE, "%s: cancelling %p", __func__, job);
 674                         aio_cancel(job);
 675                 } else {
 676                         CTR3(KTR_CXGBE, "%s: completing %p (copied %ld)",
 677                             __func__, job, copied);
 678                         aio_complete(job, copied, 0);
 679                         t4_rcvd(&toep->td->tod, intotcpcb(inp));
 680                 }
 681
 682                 complete_ddp_buffer(toep, db, db_idx);
 683                 if (toep->ddp.waiting_count > 0)
 684                         ddp_queue_toep(toep);
 685                 DDP_UNLOCK(toep);
 686                 INP_WUNLOCK(inp);
 687                 break;
 688         default:
 689                 panic("XXX: unknown tcb_rpl offset %#x, cookie %#x",
 690                     G_WORD(cpl->cookie), G_COOKIE(cpl->cookie));
 691         }
 692
 693         return (0);
 694 }
 695
 696 void
 697 handle_ddp_close(struct toepcb *toep, struct tcpcb *tp, __be32 rcv_nxt)
 698 {
 699         struct ddp_buffer *db;
 700         struct kaiocb *job;
 701         long copied;
 702         unsigned int db_idx;
 703 #ifdef INVARIANTS
 704         unsigned int db_flag;
 705 #endif
 706         int len, placed;
 707
 708         INP_WLOCK_ASSERT(toep->inp);
 709         DDP_ASSERT_LOCKED(toep);
 710
 711         /* - 1 is to ignore the byte for FIN */
 712         len = be32toh(rcv_nxt) - tp->rcv_nxt - 1;
 713         tp->rcv_nxt += len;
 714
 715         while (toep->ddp.active_count > 0) {
 716                 MPASS(toep->ddp.active_id != -1);
 717                 db_idx = toep->ddp.active_id;
 718 #ifdef INVARIANTS
 719                 db_flag = db_idx == 1 ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE;
 720 #endif
 721                 MPASS((toep->ddp.flags & db_flag) != 0);
 722                 db = &toep->ddp.db[db_idx];
 723                 job = db->job;
 724                 copied = job->aio_received;
 725                 placed = len;
 726                 if (placed > job->uaiocb.aio_nbytes - copied)
 727                         placed = job->uaiocb.aio_nbytes - copied;
 728                 if (placed > 0)
 729                         job->msgrcv = 1;
 730                 if (!aio_clear_cancel_function(job)) {
 731                         /*
 732                          * Update the copied length for when
 733                          * t4_aio_cancel_active() completes this
 734                          * request.
 735                          */
 736                         job->aio_received += placed;
 737                 } else {
 738                         CTR4(KTR_CXGBE, "%s: tid %d completed buf %d len %d",
 739                             __func__, toep->tid, db_idx, placed);
 740                         aio_complete(job, copied + placed, 0);
 741                 }
 742                 len -= placed;
 743                 complete_ddp_buffer(toep, db, db_idx);
 744         }
 745
 746         MPASS(len == 0);
 747         ddp_complete_all(toep, 0);
 748 }
 749
 750 #define DDP_ERR (F_DDP_PPOD_MISMATCH | F_DDP_LLIMIT_ERR | F_DDP_ULIMIT_ERR |\
 751          F_DDP_PPOD_PARITY_ERR | F_DDP_PADDING_ERR | F_DDP_OFFSET_ERR |\
 752          F_DDP_INVALID_TAG | F_DDP_COLOR_ERR | F_DDP_TID_MISMATCH |\
 753          F_DDP_INVALID_PPOD | F_DDP_HDRCRC_ERR | F_DDP_DATACRC_ERR)
 754
 755 extern cpl_handler_t t4_cpl_handler[];
 756
 757 static int
 758 do_rx_data_ddp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
 759 {
 760         struct adapter *sc = iq->adapter;
 761         const struct cpl_rx_data_ddp *cpl = (const void *)(rss + 1);
 762         unsigned int tid = GET_TID(cpl);
 763         uint32_t vld;
 764         struct toepcb *toep = lookup_tid(sc, tid);
 765
 766         KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
 767         KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__));
 768         KASSERT(!(toep->flags & TPF_SYNQE),
 769             ("%s: toep %p claims to be a synq entry", __func__, toep));
 770
 771         vld = be32toh(cpl->ddpvld);
 772         if (__predict_false(vld & DDP_ERR)) {
 773                 panic("%s: DDP error 0x%x (tid %d, toep %p)",
 774                     __func__, vld, tid, toep);
 775         }
 776
 777         if (ulp_mode(toep) == ULP_MODE_ISCSI) {
 778                 t4_cpl_handler[CPL_RX_ISCSI_DDP](iq, rss, m);
 779                 return (0);
 780         }
 781
 782         handle_ddp_data(toep, cpl->u.ddp_report, cpl->seq, be16toh(cpl->len));
 783
 784         return (0);
 785 }
 786
 787 static int
 788 do_rx_ddp_complete(struct sge_iq *iq, const struct rss_header *rss,
 789     struct mbuf *m)
 790 {
 791         struct adapter *sc = iq->adapter;
 792         const struct cpl_rx_ddp_complete *cpl = (const void *)(rss + 1);
 793         unsigned int tid = GET_TID(cpl);
 794         struct toepcb *toep = lookup_tid(sc, tid);
 795
 796         KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
 797         KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__));
 798         KASSERT(!(toep->flags & TPF_SYNQE),
 799             ("%s: toep %p claims to be a synq entry", __func__, toep));
 800
 801         handle_ddp_data(toep, cpl->ddp_report, cpl->rcv_nxt, 0);
 802
 803         return (0);
 804 }
 805
 806 static void
 807 enable_ddp(struct adapter *sc, struct toepcb *toep)
 808 {
 809
 810         KASSERT((toep->ddp.flags & (DDP_ON | DDP_OK | DDP_SC_REQ)) == DDP_OK,
 811             ("%s: toep %p has bad ddp_flags 0x%x",
 812             __func__, toep, toep->ddp.flags));
 813
 814         CTR3(KTR_CXGBE, "%s: tid %u (time %u)",
 815             __func__, toep->tid, time_uptime);
 816
 817         DDP_ASSERT_LOCKED(toep);
 818         toep->ddp.flags |= DDP_SC_REQ;
 819         t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_RX_DDP_FLAGS,
 820             V_TF_DDP_OFF(1) | V_TF_DDP_INDICATE_OUT(1) |
 821             V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1) |
 822             V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_BUF1_VALID(1),
 823             V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1), 0, 0);
 824         t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_T_FLAGS,
 825             V_TF_RCV_COALESCE_ENABLE(1), 0, 0, 0);
 826 }
 827
 828 static int
 829 calculate_hcf(int n1, int n2)
 830 {
 831         int a, b, t;
 832
 833         if (n1 <= n2) {
 834                 a = n1;
 835                 b = n2;
 836         } else {
 837                 a = n2;
 838                 b = n1;
 839         }
 840
 841         while (a != 0) {
 842                 t = a;
 843                 a = b % a;
 844                 b = t;
 845         }
 846
 847         return (b);
 848 }
 849
 850 static inline int
 851 pages_to_nppods(int npages, int ddp_page_shift)
 852 {
 853
 854         MPASS(ddp_page_shift >= PAGE_SHIFT);
 855
 856         return (howmany(npages >> (ddp_page_shift - PAGE_SHIFT), PPOD_PAGES));
 857 }
 858
 859 static int
 860 alloc_page_pods(struct ppod_region *pr, u_int nppods, u_int pgsz_idx,
 861     struct ppod_reservation *prsv)
 862 {
 863         vmem_addr_t addr;       /* relative to start of region */
 864
 865         if (vmem_alloc(pr->pr_arena, PPOD_SZ(nppods), M_NOWAIT | M_FIRSTFIT,
 866             &addr) != 0)
 867                 return (ENOMEM);
 868
 869 #ifdef VERBOSE_TRACES
 870         CTR5(KTR_CXGBE, "%-17s arena %p, addr 0x%08x, nppods %d, pgsz %d",
 871             __func__, pr->pr_arena, (uint32_t)addr & pr->pr_tag_mask,
 872             nppods, 1 << pr->pr_page_shift[pgsz_idx]);
 873 #endif
 874
 875         /*
 876          * The hardware tagmask includes an extra invalid bit but the arena was
 877          * seeded with valid values only.  An allocation out of this arena will
 878          * fit inside the tagmask but won't have the invalid bit set.
 879          */
 880         MPASS((addr & pr->pr_tag_mask) == addr);
 881         MPASS((addr & pr->pr_invalid_bit) == 0);
 882
 883         prsv->prsv_pr = pr;
 884         prsv->prsv_tag = V_PPOD_PGSZ(pgsz_idx) | addr;
 885         prsv->prsv_nppods = nppods;
 886
 887         return (0);
 888 }
 889
 890 int
 891 t4_alloc_page_pods_for_ps(struct ppod_region *pr, struct pageset *ps)
 892 {
 893         int i, hcf, seglen, idx, nppods;
 894         struct ppod_reservation *prsv = &ps->prsv;
 895
 896         KASSERT(prsv->prsv_nppods == 0,
 897             ("%s: page pods already allocated", __func__));
 898
 899         /*
 900          * The DDP page size is unrelated to the VM page size.  We combine
 901          * contiguous physical pages into larger segments to get the best DDP
 902          * page size possible.  This is the largest of the four sizes in
 903          * A_ULP_RX_TDDP_PSZ that evenly divides the HCF of the segment sizes in
 904          * the page list.
 905          */
 906         hcf = 0;
 907         for (i = 0; i < ps->npages; i++) {
 908                 seglen = PAGE_SIZE;
 909                 while (i < ps->npages - 1 &&
 910                     VM_PAGE_TO_PHYS(ps->pages[i]) + PAGE_SIZE ==
 911                     VM_PAGE_TO_PHYS(ps->pages[i + 1])) {
 912                         seglen += PAGE_SIZE;
 913                         i++;
 914                 }
 915
 916                 hcf = calculate_hcf(hcf, seglen);
 917                 if (hcf < (1 << pr->pr_page_shift[1])) {
 918                         idx = 0;
 919                         goto have_pgsz; /* give up, short circuit */
 920                 }
 921         }
 922
 923 #define PR_PAGE_MASK(x) ((1 << pr->pr_page_shift[(x)]) - 1)
 924         MPASS((hcf & PR_PAGE_MASK(0)) == 0); /* PAGE_SIZE is >= 4K everywhere */
 925         for (idx = nitems(pr->pr_page_shift) - 1; idx > 0; idx--) {
 926                 if ((hcf & PR_PAGE_MASK(idx)) == 0)
 927                         break;
 928         }
 929 #undef PR_PAGE_MASK
 930
 931 have_pgsz:
 932         MPASS(idx <= M_PPOD_PGSZ);
 933
 934         nppods = pages_to_nppods(ps->npages, pr->pr_page_shift[idx]);
 935         if (alloc_page_pods(pr, nppods, idx, prsv) != 0)
 936                 return (0);
 937         MPASS(prsv->prsv_nppods > 0);
 938
 939         return (1);
 940 }
 941
 942 int
 943 t4_alloc_page_pods_for_buf(struct ppod_region *pr, vm_offset_t buf, int len,
 944     struct ppod_reservation *prsv)
 945 {
 946         int hcf, seglen, idx, npages, nppods;
 947         uintptr_t start_pva, end_pva, pva, p1;
 948
 949         MPASS(buf > 0);
 950         MPASS(len > 0);
 951
 952         /*
 953          * The DDP page size is unrelated to the VM page size.  We combine
 954          * contiguous physical pages into larger segments to get the best DDP
 955          * page size possible.  This is the largest of the four sizes in
 956          * A_ULP_RX_ISCSI_PSZ that evenly divides the HCF of the segment sizes
 957          * in the page list.
 958          */
 959         hcf = 0;
 960         start_pva = trunc_page(buf);
 961         end_pva = trunc_page(buf + len - 1);
 962         pva = start_pva;
 963         while (pva <= end_pva) {
 964                 seglen = PAGE_SIZE;
 965                 p1 = pmap_kextract(pva);
 966                 pva += PAGE_SIZE;
 967                 while (pva <= end_pva && p1 + seglen == pmap_kextract(pva)) {
 968                         seglen += PAGE_SIZE;
 969                         pva += PAGE_SIZE;
 970                 }
 971
 972                 hcf = calculate_hcf(hcf, seglen);
 973                 if (hcf < (1 << pr->pr_page_shift[1])) {
 974                         idx = 0;
 975                         goto have_pgsz; /* give up, short circuit */
 976                 }
 977         }
 978
 979 #define PR_PAGE_MASK(x) ((1 << pr->pr_page_shift[(x)]) - 1)
 980         MPASS((hcf & PR_PAGE_MASK(0)) == 0); /* PAGE_SIZE is >= 4K everywhere */
 981         for (idx = nitems(pr->pr_page_shift) - 1; idx > 0; idx--) {
 982                 if ((hcf & PR_PAGE_MASK(idx)) == 0)
 983                         break;
 984         }
 985 #undef PR_PAGE_MASK
 986
 987 have_pgsz:
 988         MPASS(idx <= M_PPOD_PGSZ);
 989
 990         npages = 1;
 991         npages += (end_pva - start_pva) >> pr->pr_page_shift[idx];
 992         nppods = howmany(npages, PPOD_PAGES);
 993         if (alloc_page_pods(pr, nppods, idx, prsv) != 0)
 994                 return (ENOMEM);
 995         MPASS(prsv->prsv_nppods > 0);
 996
 997         return (0);
 998 }
 999
1000 int
1001 t4_alloc_page_pods_for_sgl(struct ppod_region *pr, struct ctl_sg_entry *sgl,
1002     int entries, struct ppod_reservation *prsv)
1003 {
1004         int hcf, seglen, idx = 0, npages, nppods, i, len;
1005         uintptr_t start_pva, end_pva, pva, p1 ;
1006         vm_offset_t buf;
1007         struct ctl_sg_entry *sge;
1008
1009         MPASS(entries > 0);
1010         MPASS(sgl);
1011
1012         /*
1013          * The DDP page size is unrelated to the VM page size.  We combine
1014          * contiguous physical pages into larger segments to get the best DDP
1015          * page size possible.  This is the largest of the four sizes in
1016          * A_ULP_RX_ISCSI_PSZ that evenly divides the HCF of the segment sizes
1017          * in the page list.
1018          */
1019         hcf = 0;
1020         for (i = entries - 1; i >= 0; i--) {
1021                 sge = sgl + i;
1022                 buf = (vm_offset_t)sge->addr;
1023                 len = sge->len;
1024                 start_pva = trunc_page(buf);
1025                 end_pva = trunc_page(buf + len - 1);
1026                 pva = start_pva;
1027                 while (pva <= end_pva) {
1028                         seglen = PAGE_SIZE;
1029                         p1 = pmap_kextract(pva);
1030                         pva += PAGE_SIZE;
1031                         while (pva <= end_pva && p1 + seglen ==
1032                             pmap_kextract(pva)) {
1033                                 seglen += PAGE_SIZE;
1034                                 pva += PAGE_SIZE;
1035                         }
1036
1037                         hcf = calculate_hcf(hcf, seglen);
1038                         if (hcf < (1 << pr->pr_page_shift[1])) {
1039                                 idx = 0;
1040                                 goto have_pgsz; /* give up, short circuit */
1041                         }
1042                 }
1043         }
1044 #define PR_PAGE_MASK(x) ((1 << pr->pr_page_shift[(x)]) - 1)
1045         MPASS((hcf & PR_PAGE_MASK(0)) == 0); /* PAGE_SIZE is >= 4K everywhere */
1046         for (idx = nitems(pr->pr_page_shift) - 1; idx > 0; idx--) {
1047                 if ((hcf & PR_PAGE_MASK(idx)) == 0)
1048                         break;
1049         }
1050 #undef PR_PAGE_MASK
1051
1052 have_pgsz:
1053         MPASS(idx <= M_PPOD_PGSZ);
1054
1055         npages = 0;
1056         while (entries--) {
1057                 npages++;
1058                 start_pva = trunc_page((vm_offset_t)sgl->addr);
1059                 end_pva = trunc_page((vm_offset_t)sgl->addr + sgl->len - 1);
1060                 npages += (end_pva - start_pva) >> pr->pr_page_shift[idx];
1061                 sgl = sgl + 1;
1062         }
1063         nppods = howmany(npages, PPOD_PAGES);
1064         if (alloc_page_pods(pr, nppods, idx, prsv) != 0)
1065                 return (ENOMEM);
1066         MPASS(prsv->prsv_nppods > 0);
1067         return (0);
1068 }
1069
1070 void
1071 t4_free_page_pods(struct ppod_reservation *prsv)
1072 {
1073         struct ppod_region *pr = prsv->prsv_pr;
1074         vmem_addr_t addr;
1075
1076         MPASS(prsv != NULL);
1077         MPASS(prsv->prsv_nppods != 0);
1078
1079         addr = prsv->prsv_tag & pr->pr_tag_mask;
1080         MPASS((addr & pr->pr_invalid_bit) == 0);
1081
1082 #ifdef VERBOSE_TRACES
1083         CTR4(KTR_CXGBE, "%-17s arena %p, addr 0x%08x, nppods %d", __func__,
1084             pr->pr_arena, addr, prsv->prsv_nppods);
1085 #endif
1086
1087         vmem_free(pr->pr_arena, addr, PPOD_SZ(prsv->prsv_nppods));
1088         prsv->prsv_nppods = 0;
1089 }
1090
1091 #define NUM_ULP_TX_SC_IMM_PPODS (256 / PPOD_SIZE)
1092
1093 int
1094 t4_write_page_pods_for_ps(struct adapter *sc, struct sge_wrq *wrq, int tid,
1095     struct pageset *ps)
1096 {
1097         struct wrqe *wr;
1098         struct ulp_mem_io *ulpmc;
1099         struct ulptx_idata *ulpsc;
1100         struct pagepod *ppod;
1101         int i, j, k, n, chunk, len, ddp_pgsz, idx;
1102         u_int ppod_addr;
1103         uint32_t cmd;
1104         struct ppod_reservation *prsv = &ps->prsv;
1105         struct ppod_region *pr = prsv->prsv_pr;
1106         vm_paddr_t pa;
1107
1108         KASSERT(!(ps->flags & PS_PPODS_WRITTEN),
1109             ("%s: page pods already written", __func__));
1110         MPASS(prsv->prsv_nppods > 0);
1111
1112         cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE));
1113         if (is_t4(sc))
1114                 cmd |= htobe32(F_ULP_MEMIO_ORDER);
1115         else
1116                 cmd |= htobe32(F_T5_ULP_MEMIO_IMM);
1117         ddp_pgsz = 1 << pr->pr_page_shift[G_PPOD_PGSZ(prsv->prsv_tag)];
1118         ppod_addr = pr->pr_start + (prsv->prsv_tag & pr->pr_tag_mask);
1119         for (i = 0; i < prsv->prsv_nppods; ppod_addr += chunk) {
1120
1121                 /* How many page pods are we writing in this cycle */
1122                 n = min(prsv->prsv_nppods - i, NUM_ULP_TX_SC_IMM_PPODS);
1123                 chunk = PPOD_SZ(n);
1124                 len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16);
1125
1126                 wr = alloc_wrqe(len, wrq);
1127                 if (wr == NULL)
1128                         return (ENOMEM);        /* ok to just bail out */
1129                 ulpmc = wrtod(wr);
1130
1131                 INIT_ULPTX_WR(ulpmc, len, 0, 0);
1132                 ulpmc->cmd = cmd;
1133                 ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32));
1134                 ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16));
1135                 ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5));
1136
1137                 ulpsc = (struct ulptx_idata *)(ulpmc + 1);
1138                 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
1139                 ulpsc->len = htobe32(chunk);
1140
1141                 ppod = (struct pagepod *)(ulpsc + 1);
1142                 for (j = 0; j < n; i++, j++, ppod++) {
1143                         ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID |
1144                             V_PPOD_TID(tid) | prsv->prsv_tag);
1145                         ppod->len_offset = htobe64(V_PPOD_LEN(ps->len) |
1146                             V_PPOD_OFST(ps->offset));
1147                         ppod->rsvd = 0;
1148                         idx = i * PPOD_PAGES * (ddp_pgsz / PAGE_SIZE);
1149                         for (k = 0; k < nitems(ppod->addr); k++) {
1150                                 if (idx < ps->npages) {
1151                                         pa = VM_PAGE_TO_PHYS(ps->pages[idx]);
1152                                         ppod->addr[k] = htobe64(pa);
1153                                         idx += ddp_pgsz / PAGE_SIZE;
1154                                 } else
1155                                         ppod->addr[k] = 0;
1156 #if 0
1157                                 CTR5(KTR_CXGBE,
1158                                     "%s: tid %d ppod[%d]->addr[%d] = %p",
1159                                     __func__, toep->tid, i, k,
1160                                     be64toh(ppod->addr[k]));
1161 #endif
1162                         }
1163
1164                 }
1165
1166                 t4_wrq_tx(sc, wr);
1167         }
1168         ps->flags |= PS_PPODS_WRITTEN;
1169
1170         return (0);
1171 }
1172
1173 static struct mbuf *
1174 alloc_raw_wr_mbuf(int len)
1175 {
1176         struct mbuf *m;
1177
1178         if (len <= MHLEN)
1179                 m = m_gethdr(M_NOWAIT, MT_DATA);
1180         else if (len <= MCLBYTES)
1181                 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1182         else
1183                 m = NULL;
1184         if (m == NULL)
1185                 return (NULL);
1186         m->m_pkthdr.len = len;
1187         m->m_len = len;
1188         set_mbuf_raw_wr(m, true);
1189         return (m);
1190 }
1191
1192 int
1193 t4_write_page_pods_for_buf(struct adapter *sc, struct toepcb *toep,
1194     struct ppod_reservation *prsv, vm_offset_t buf, int buflen,
1195     struct mbufq *wrq)
1196 {
1197         struct ulp_mem_io *ulpmc;
1198         struct ulptx_idata *ulpsc;
1199         struct pagepod *ppod;
1200         int i, j, k, n, chunk, len, ddp_pgsz;
1201         u_int ppod_addr, offset;
1202         uint32_t cmd;
1203         struct ppod_region *pr = prsv->prsv_pr;
1204         uintptr_t end_pva, pva;
1205         vm_paddr_t pa;
1206         struct mbuf *m;
1207
1208         cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE));
1209         if (is_t4(sc))
1210                 cmd |= htobe32(F_ULP_MEMIO_ORDER);
1211         else
1212                 cmd |= htobe32(F_T5_ULP_MEMIO_IMM);
1213         ddp_pgsz = 1 << pr->pr_page_shift[G_PPOD_PGSZ(prsv->prsv_tag)];
1214         offset = buf & PAGE_MASK;
1215         ppod_addr = pr->pr_start + (prsv->prsv_tag & pr->pr_tag_mask);
1216         pva = trunc_page(buf);
1217         end_pva = trunc_page(buf + buflen - 1);
1218         for (i = 0; i < prsv->prsv_nppods; ppod_addr += chunk) {
1219
1220                 /* How many page pods are we writing in this cycle */
1221                 n = min(prsv->prsv_nppods - i, NUM_ULP_TX_SC_IMM_PPODS);
1222                 MPASS(n > 0);
1223                 chunk = PPOD_SZ(n);
1224                 len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16);
1225
1226                 m = alloc_raw_wr_mbuf(len);
1227                 if (m == NULL)
1228                         return (ENOMEM);
1229                 ulpmc = mtod(m, struct ulp_mem_io *);
1230
1231                 INIT_ULPTX_WR(ulpmc, len, 0, toep->tid);
1232                 ulpmc->cmd = cmd;
1233                 ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32));
1234                 ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16));
1235                 ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5));
1236
1237                 ulpsc = (struct ulptx_idata *)(ulpmc + 1);
1238                 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
1239                 ulpsc->len = htobe32(chunk);
1240
1241                 ppod = (struct pagepod *)(ulpsc + 1);
1242                 for (j = 0; j < n; i++, j++, ppod++) {
1243                         ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID |
1244                             V_PPOD_TID(toep->tid) |
1245                             (prsv->prsv_tag & ~V_PPOD_PGSZ(M_PPOD_PGSZ)));
1246                         ppod->len_offset = htobe64(V_PPOD_LEN(buflen) |
1247                             V_PPOD_OFST(offset));
1248                         ppod->rsvd = 0;
1249
1250                         for (k = 0; k < nitems(ppod->addr); k++) {
1251                                 if (pva > end_pva)
1252                                         ppod->addr[k] = 0;
1253                                 else {
1254                                         pa = pmap_kextract(pva);
1255                                         ppod->addr[k] = htobe64(pa);
1256                                         pva += ddp_pgsz;
1257                                 }
1258 #if 0
1259                                 CTR5(KTR_CXGBE,
1260                                     "%s: tid %d ppod[%d]->addr[%d] = %p",
1261                                     __func__, toep->tid, i, k,
1262                                     be64toh(ppod->addr[k]));
1263 #endif
1264                         }
1265
1266                         /*
1267                          * Walk back 1 segment so that the first address in the
1268                          * next pod is the same as the last one in the current
1269                          * pod.
1270                          */
1271                         pva -= ddp_pgsz;
1272                 }
1273
1274                 mbufq_enqueue(wrq, m);
1275         }
1276
1277         MPASS(pva <= end_pva);
1278
1279         return (0);
1280 }
1281
1282 int
1283 t4_write_page_pods_for_sgl(struct adapter *sc, struct toepcb *toep,
1284     struct ppod_reservation *prsv, struct ctl_sg_entry *sgl, int entries,
1285     int xferlen, struct mbufq *wrq)
1286 {
1287         struct ulp_mem_io *ulpmc;
1288         struct ulptx_idata *ulpsc;
1289         struct pagepod *ppod;
1290         int i, j, k, n, chunk, len, ddp_pgsz;
1291         u_int ppod_addr, offset, sg_offset = 0;
1292         uint32_t cmd;
1293         struct ppod_region *pr = prsv->prsv_pr;
1294         uintptr_t pva;
1295         vm_paddr_t pa;
1296         struct mbuf *m;
1297
1298         MPASS(sgl != NULL);
1299         MPASS(entries > 0);
1300         cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE));
1301         if (is_t4(sc))
1302                 cmd |= htobe32(F_ULP_MEMIO_ORDER);
1303         else
1304                 cmd |= htobe32(F_T5_ULP_MEMIO_IMM);
1305         ddp_pgsz = 1 << pr->pr_page_shift[G_PPOD_PGSZ(prsv->prsv_tag)];
1306         offset = (vm_offset_t)sgl->addr & PAGE_MASK;
1307         ppod_addr = pr->pr_start + (prsv->prsv_tag & pr->pr_tag_mask);
1308         pva = trunc_page((vm_offset_t)sgl->addr);
1309         for (i = 0; i < prsv->prsv_nppods; ppod_addr += chunk) {
1310
1311                 /* How many page pods are we writing in this cycle */
1312                 n = min(prsv->prsv_nppods - i, NUM_ULP_TX_SC_IMM_PPODS);
1313                 MPASS(n > 0);
1314                 chunk = PPOD_SZ(n);
1315                 len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16);
1316
1317                 m = alloc_raw_wr_mbuf(len);
1318                 if (m == NULL)
1319                         return (ENOMEM);
1320                 ulpmc = mtod(m, struct ulp_mem_io *);
1321
1322                 INIT_ULPTX_WR(ulpmc, len, 0, toep->tid);
1323                 ulpmc->cmd = cmd;
1324                 ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32));
1325                 ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16));
1326                 ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5));
1327
1328                 ulpsc = (struct ulptx_idata *)(ulpmc + 1);
1329                 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
1330                 ulpsc->len = htobe32(chunk);
1331
1332                 ppod = (struct pagepod *)(ulpsc + 1);
1333                 for (j = 0; j < n; i++, j++, ppod++) {
1334                         ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID |
1335                             V_PPOD_TID(toep->tid) |
1336                             (prsv->prsv_tag & ~V_PPOD_PGSZ(M_PPOD_PGSZ)));
1337                         ppod->len_offset = htobe64(V_PPOD_LEN(xferlen) |
1338                             V_PPOD_OFST(offset));
1339                         ppod->rsvd = 0;
1340
1341                         for (k = 0; k < nitems(ppod->addr); k++) {
1342                                 if (entries != 0) {
1343                                         pa = pmap_kextract(pva + sg_offset);
1344                                         ppod->addr[k] = htobe64(pa);
1345                                 } else
1346                                         ppod->addr[k] = 0;
1347
1348 #if 0
1349                                 CTR5(KTR_CXGBE,
1350                                     "%s: tid %d ppod[%d]->addr[%d] = %p",
1351                                     __func__, toep->tid, i, k,
1352                                     be64toh(ppod->addr[k]));
1353 #endif
1354
1355                                 /*
1356                                  * If this is the last entry in a pod,
1357                                  * reuse the same entry for first address
1358                                  * in the next pod.
1359                                  */
1360                                 if (k + 1 == nitems(ppod->addr))
1361                                         break;
1362
1363                                 /*
1364                                  * Don't move to the next DDP page if the
1365                                  * sgl is already finished.
1366                                  */
1367                                 if (entries == 0)
1368                                         continue;
1369
1370                                 sg_offset += ddp_pgsz;
1371                                 if (sg_offset == sgl->len) {
1372                                         /*
1373                                          * This sgl entry is done.  Go
1374                                          * to the next.
1375                                          */
1376                                         entries--;
1377                                         sgl++;
1378                                         sg_offset = 0;
1379                                         if (entries != 0)
1380                                                 pva = trunc_page(
1381                                                     (vm_offset_t)sgl->addr);
1382                                 }
1383                         }
1384                 }
1385
1386                 mbufq_enqueue(wrq, m);
1387         }
1388
1389         return (0);
1390 }
1391
1392 /*
1393  * Prepare a pageset for DDP.  This sets up page pods.
1394  */
1395 static int
1396 prep_pageset(struct adapter *sc, struct toepcb *toep, struct pageset *ps)
1397 {
1398         struct tom_data *td = sc->tom_softc;
1399
1400         if (ps->prsv.prsv_nppods == 0 &&
1401             !t4_alloc_page_pods_for_ps(&td->pr, ps)) {
1402                 return (0);
1403         }
1404         if (!(ps->flags & PS_PPODS_WRITTEN) &&
1405             t4_write_page_pods_for_ps(sc, toep->ctrlq, toep->tid, ps) != 0) {
1406                 return (0);
1407         }
1408
1409         return (1);
1410 }
1411
1412 int
1413 t4_init_ppod_region(struct ppod_region *pr, struct t4_range *r, u_int psz,
1414     const char *name)
1415 {
1416         int i;
1417
1418         MPASS(pr != NULL);
1419         MPASS(r->size > 0);
1420
1421         pr->pr_start = r->start;
1422         pr->pr_len = r->size;
1423         pr->pr_page_shift[0] = 12 + G_HPZ0(psz);
1424         pr->pr_page_shift[1] = 12 + G_HPZ1(psz);
1425         pr->pr_page_shift[2] = 12 + G_HPZ2(psz);
1426         pr->pr_page_shift[3] = 12 + G_HPZ3(psz);
1427
1428         /* The SGL -> page pod algorithm requires the sizes to be in order. */
1429         for (i = 1; i < nitems(pr->pr_page_shift); i++) {
1430                 if (pr->pr_page_shift[i] <= pr->pr_page_shift[i - 1])
1431                         return (ENXIO);
1432         }
1433
1434         pr->pr_tag_mask = ((1 << fls(r->size)) - 1) & V_PPOD_TAG(M_PPOD_TAG);
1435         pr->pr_alias_mask = V_PPOD_TAG(M_PPOD_TAG) & ~pr->pr_tag_mask;
1436         if (pr->pr_tag_mask == 0 || pr->pr_alias_mask == 0)
1437                 return (ENXIO);
1438         pr->pr_alias_shift = fls(pr->pr_tag_mask);
1439         pr->pr_invalid_bit = 1 << (pr->pr_alias_shift - 1);
1440
1441         pr->pr_arena = vmem_create(name, 0, pr->pr_len, PPOD_SIZE, 0,
1442             M_FIRSTFIT | M_NOWAIT);
1443         if (pr->pr_arena == NULL)
1444                 return (ENOMEM);
1445
1446         return (0);
1447 }
1448
1449 void
1450 t4_free_ppod_region(struct ppod_region *pr)
1451 {
1452
1453         MPASS(pr != NULL);
1454
1455         if (pr->pr_arena)
1456                 vmem_destroy(pr->pr_arena);
1457         bzero(pr, sizeof(*pr));
1458 }
1459
1460 static int
1461 pscmp(struct pageset *ps, struct vmspace *vm, vm_offset_t start, int npages,
1462     int pgoff, int len)
1463 {
1464
1465         if (ps->start != start || ps->npages != npages ||
1466             ps->offset != pgoff || ps->len != len)
1467                 return (1);
1468
1469         return (ps->vm != vm || ps->vm_timestamp != vm->vm_map.timestamp);
1470 }
1471
1472 static int
1473 hold_aio(struct toepcb *toep, struct kaiocb *job, struct pageset **pps)
1474 {
1475         struct vmspace *vm;
1476         vm_map_t map;
1477         vm_offset_t start, end, pgoff;
1478         struct pageset *ps;
1479         int n;
1480
1481         DDP_ASSERT_LOCKED(toep);
1482
1483         /*
1484          * The AIO subsystem will cancel and drain all requests before
1485          * permitting a process to exit or exec, so p_vmspace should
1486          * be stable here.
1487          */
1488         vm = job->userproc->p_vmspace;
1489         map = &vm->vm_map;
1490         start = (uintptr_t)job->uaiocb.aio_buf;
1491         pgoff = start & PAGE_MASK;
1492         end = round_page(start + job->uaiocb.aio_nbytes);
1493         start = trunc_page(start);
1494
1495         if (end - start > MAX_DDP_BUFFER_SIZE) {
1496                 /*
1497                  * Truncate the request to a short read.
1498                  * Alternatively, we could DDP in chunks to the larger
1499                  * buffer, but that would be quite a bit more work.
1500                  *
1501                  * When truncating, round the request down to avoid
1502                  * crossing a cache line on the final transaction.
1503                  */
1504                 end = rounddown2(start + MAX_DDP_BUFFER_SIZE, CACHE_LINE_SIZE);
1505 #ifdef VERBOSE_TRACES
1506                 CTR4(KTR_CXGBE, "%s: tid %d, truncating size from %lu to %lu",
1507                     __func__, toep->tid, (unsigned long)job->uaiocb.aio_nbytes,
1508                     (unsigned long)(end - (start + pgoff)));
1509                 job->uaiocb.aio_nbytes = end - (start + pgoff);
1510 #endif
1511                 end = round_page(end);
1512         }
1513
1514         n = atop(end - start);
1515
1516         /*
1517          * Try to reuse a cached pageset.
1518          */
1519         TAILQ_FOREACH(ps, &toep->ddp.cached_pagesets, link) {
1520                 if (pscmp(ps, vm, start, n, pgoff,
1521                     job->uaiocb.aio_nbytes) == 0) {
1522                         TAILQ_REMOVE(&toep->ddp.cached_pagesets, ps, link);
1523                         toep->ddp.cached_count--;
1524                         *pps = ps;
1525                         return (0);
1526                 }
1527         }
1528
1529         /*
1530          * If there are too many cached pagesets to create a new one,
1531          * free a pageset before creating a new one.
1532          */
1533         KASSERT(toep->ddp.active_count + toep->ddp.cached_count <=
1534             nitems(toep->ddp.db), ("%s: too many wired pagesets", __func__));
1535         if (toep->ddp.active_count + toep->ddp.cached_count ==
1536             nitems(toep->ddp.db)) {
1537                 KASSERT(toep->ddp.cached_count > 0,
1538                     ("no cached pageset to free"));
1539                 ps = TAILQ_LAST(&toep->ddp.cached_pagesets, pagesetq);
1540                 TAILQ_REMOVE(&toep->ddp.cached_pagesets, ps, link);
1541                 toep->ddp.cached_count--;
1542                 free_pageset(toep->td, ps);
1543         }
1544         DDP_UNLOCK(toep);
1545
1546         /* Create a new pageset. */
1547         ps = malloc(sizeof(*ps) + n * sizeof(vm_page_t), M_CXGBE, M_WAITOK |
1548             M_ZERO);
1549         ps->pages = (vm_page_t *)(ps + 1);
1550         ps->vm_timestamp = map->timestamp;
1551         ps->npages = vm_fault_quick_hold_pages(map, start, end - start,
1552             VM_PROT_WRITE, ps->pages, n);
1553
1554         DDP_LOCK(toep);
1555         if (ps->npages < 0) {
1556                 free(ps, M_CXGBE);
1557                 return (EFAULT);
1558         }
1559
1560         KASSERT(ps->npages == n, ("hold_aio: page count mismatch: %d vs %d",
1561             ps->npages, n));
1562
1563         ps->offset = pgoff;
1564         ps->len = job->uaiocb.aio_nbytes;
1565         refcount_acquire(&vm->vm_refcnt);
1566         ps->vm = vm;
1567         ps->start = start;
1568
1569         CTR5(KTR_CXGBE, "%s: tid %d, new pageset %p for job %p, npages %d",
1570             __func__, toep->tid, ps, job, ps->npages);
1571         *pps = ps;
1572         return (0);
1573 }
1574
1575 static void
1576 ddp_complete_all(struct toepcb *toep, int error)
1577 {
1578         struct kaiocb *job;
1579
1580         DDP_ASSERT_LOCKED(toep);
1581         while (!TAILQ_EMPTY(&toep->ddp.aiojobq)) {
1582                 job = TAILQ_FIRST(&toep->ddp.aiojobq);
1583                 TAILQ_REMOVE(&toep->ddp.aiojobq, job, list);
1584                 toep->ddp.waiting_count--;
1585                 if (aio_clear_cancel_function(job))
1586                         ddp_complete_one(job, error);
1587         }
1588 }
1589
1590 static void
1591 aio_ddp_cancel_one(struct kaiocb *job)
1592 {
1593         long copied;
1594
1595         /*
1596          * If this job had copied data out of the socket buffer before
1597          * it was cancelled, report it as a short read rather than an
1598          * error.
1599          */
1600         copied = job->aio_received;
1601         if (copied != 0)
1602                 aio_complete(job, copied, 0);
1603         else
1604                 aio_cancel(job);
1605 }
1606
1607 /*
1608  * Called when the main loop wants to requeue a job to retry it later.
1609  * Deals with the race of the job being cancelled while it was being
1610  * examined.
1611  */
1612 static void
1613 aio_ddp_requeue_one(struct toepcb *toep, struct kaiocb *job)
1614 {
1615
1616         DDP_ASSERT_LOCKED(toep);
1617         if (!(toep->ddp.flags & DDP_DEAD) &&
1618             aio_set_cancel_function(job, t4_aio_cancel_queued)) {
1619                 TAILQ_INSERT_HEAD(&toep->ddp.aiojobq, job, list);
1620                 toep->ddp.waiting_count++;
1621         } else
1622                 aio_ddp_cancel_one(job);
1623 }
1624
1625 static void
1626 aio_ddp_requeue(struct toepcb *toep)
1627 {
1628         struct adapter *sc = td_adapter(toep->td);
1629         struct socket *so;
1630         struct sockbuf *sb;
1631         struct inpcb *inp;
1632         struct kaiocb *job;
1633         struct ddp_buffer *db;
1634         size_t copied, offset, resid;
1635         struct pageset *ps;
1636         struct mbuf *m;
1637         uint64_t ddp_flags, ddp_flags_mask;
1638         struct wrqe *wr;
1639         int buf_flag, db_idx, error;
1640
1641         DDP_ASSERT_LOCKED(toep);
1642
1643 restart:
1644         if (toep->ddp.flags & DDP_DEAD) {
1645                 MPASS(toep->ddp.waiting_count == 0);
1646                 MPASS(toep->ddp.active_count == 0);
1647                 return;
1648         }
1649
1650         if (toep->ddp.waiting_count == 0 ||
1651             toep->ddp.active_count == nitems(toep->ddp.db)) {
1652                 return;
1653         }
1654
1655         job = TAILQ_FIRST(&toep->ddp.aiojobq);
1656         so = job->fd_file->f_data;
1657         sb = &so->so_rcv;
1658         SOCKBUF_LOCK(sb);
1659
1660         /* We will never get anything unless we are or were connected. */
1661         if (!(so->so_state & (SS_ISCONNECTED|SS_ISDISCONNECTED))) {
1662                 SOCKBUF_UNLOCK(sb);
1663                 ddp_complete_all(toep, ENOTCONN);
1664                 return;
1665         }
1666
1667         KASSERT(toep->ddp.active_count == 0 || sbavail(sb) == 0,
1668             ("%s: pending sockbuf data and DDP is active", __func__));
1669
1670         /* Abort if socket has reported problems. */
1671         /* XXX: Wait for any queued DDP's to finish and/or flush them? */
1672         if (so->so_error && sbavail(sb) == 0) {
1673                 toep->ddp.waiting_count--;
1674                 TAILQ_REMOVE(&toep->ddp.aiojobq, job, list);
1675                 if (!aio_clear_cancel_function(job)) {
1676                         SOCKBUF_UNLOCK(sb);
1677                         goto restart;
1678                 }
1679
1680                 /*
1681                  * If this job has previously copied some data, report
1682                  * a short read and leave the error to be reported by
1683                  * a future request.
1684                  */
1685                 copied = job->aio_received;
1686                 if (copied != 0) {
1687                         SOCKBUF_UNLOCK(sb);
1688                         aio_complete(job, copied, 0);
1689                         goto restart;
1690                 }
1691                 error = so->so_error;
1692                 so->so_error = 0;
1693                 SOCKBUF_UNLOCK(sb);
1694                 aio_complete(job, -1, error);
1695                 goto restart;
1696         }
1697
1698         /*
1699          * Door is closed.  If there is pending data in the socket buffer,
1700          * deliver it.  If there are pending DDP requests, wait for those
1701          * to complete.  Once they have completed, return EOF reads.
1702          */
1703         if (sb->sb_state & SBS_CANTRCVMORE && sbavail(sb) == 0) {
1704                 SOCKBUF_UNLOCK(sb);
1705                 if (toep->ddp.active_count != 0)
1706                         return;
1707                 ddp_complete_all(toep, 0);
1708                 return;
1709         }
1710
1711         /*
1712          * If DDP is not enabled and there is no pending socket buffer
1713          * data, try to enable DDP.
1714          */
1715         if (sbavail(sb) == 0 && (toep->ddp.flags & DDP_ON) == 0) {
1716                 SOCKBUF_UNLOCK(sb);
1717
1718                 /*
1719                  * Wait for the card to ACK that DDP is enabled before
1720                  * queueing any buffers.  Currently this waits for an
1721                  * indicate to arrive.  This could use a TCB_SET_FIELD_RPL
1722                  * message to know that DDP was enabled instead of waiting
1723                  * for the indicate which would avoid copying the indicate
1724                  * if no data is pending.
1725                  *
1726                  * XXX: Might want to limit the indicate size to the size
1727                  * of the first queued request.
1728                  */
1729                 if ((toep->ddp.flags & DDP_SC_REQ) == 0)
1730                         enable_ddp(sc, toep);
1731                 return;
1732         }
1733         SOCKBUF_UNLOCK(sb);
1734
1735         /*
1736          * If another thread is queueing a buffer for DDP, let it
1737          * drain any work and return.
1738          */
1739         if (toep->ddp.queueing != NULL)
1740                 return;
1741
1742         /* Take the next job to prep it for DDP. */
1743         toep->ddp.waiting_count--;
1744         TAILQ_REMOVE(&toep->ddp.aiojobq, job, list);
1745         if (!aio_clear_cancel_function(job))
1746                 goto restart;
1747         toep->ddp.queueing = job;
1748
1749         /* NB: This drops DDP_LOCK while it holds the backing VM pages. */
1750         error = hold_aio(toep, job, &ps);
1751         if (error != 0) {
1752                 ddp_complete_one(job, error);
1753                 toep->ddp.queueing = NULL;
1754                 goto restart;
1755         }
1756
1757         SOCKBUF_LOCK(sb);
1758         if (so->so_error && sbavail(sb) == 0) {
1759                 copied = job->aio_received;
1760                 if (copied != 0) {
1761                         SOCKBUF_UNLOCK(sb);
1762                         recycle_pageset(toep, ps);
1763                         aio_complete(job, copied, 0);
1764                         toep->ddp.queueing = NULL;
1765                         goto restart;
1766                 }
1767
1768                 error = so->so_error;
1769                 so->so_error = 0;
1770                 SOCKBUF_UNLOCK(sb);
1771                 recycle_pageset(toep, ps);
1772                 aio_complete(job, -1, error);
1773                 toep->ddp.queueing = NULL;
1774                 goto restart;
1775         }
1776
1777         if (sb->sb_state & SBS_CANTRCVMORE && sbavail(sb) == 0) {
1778                 SOCKBUF_UNLOCK(sb);
1779                 recycle_pageset(toep, ps);
1780                 if (toep->ddp.active_count != 0) {
1781                         /*
1782                          * The door is closed, but there are still pending
1783                          * DDP buffers.  Requeue.  These jobs will all be
1784                          * completed once those buffers drain.
1785                          */
1786                         aio_ddp_requeue_one(toep, job);
1787                         toep->ddp.queueing = NULL;
1788                         return;
1789                 }
1790                 ddp_complete_one(job, 0);
1791                 ddp_complete_all(toep, 0);
1792                 toep->ddp.queueing = NULL;
1793                 return;
1794         }
1795
1796 sbcopy:
1797         /*
1798          * If the toep is dead, there shouldn't be any data in the socket
1799          * buffer, so the above case should have handled this.
1800          */
1801         MPASS(!(toep->ddp.flags & DDP_DEAD));
1802
1803         /*
1804          * If there is pending data in the socket buffer (either
1805          * from before the requests were queued or a DDP indicate),
1806          * copy those mbufs out directly.
1807          */
1808         copied = 0;
1809         offset = ps->offset + job->aio_received;
1810         MPASS(job->aio_received <= job->uaiocb.aio_nbytes);
1811         resid = job->uaiocb.aio_nbytes - job->aio_received;
1812         m = sb->sb_mb;
1813         KASSERT(m == NULL || toep->ddp.active_count == 0,
1814             ("%s: sockbuf data with active DDP", __func__));
1815         while (m != NULL && resid > 0) {
1816                 struct iovec iov[1];
1817                 struct uio uio;
1818 #ifdef INVARIANTS
1819                 int error;
1820 #endif
1821
1822                 iov[0].iov_base = mtod(m, void *);
1823                 iov[0].iov_len = m->m_len;
1824                 if (iov[0].iov_len > resid)
1825                         iov[0].iov_len = resid;
1826                 uio.uio_iov = iov;
1827                 uio.uio_iovcnt = 1;
1828                 uio.uio_offset = 0;
1829                 uio.uio_resid = iov[0].iov_len;
1830                 uio.uio_segflg = UIO_SYSSPACE;
1831                 uio.uio_rw = UIO_WRITE;
1832 #ifdef INVARIANTS
1833                 error = uiomove_fromphys(ps->pages, offset + copied,
1834                     uio.uio_resid, &uio);
1835 #else
1836                 uiomove_fromphys(ps->pages, offset + copied, uio.uio_resid, &uio);
1837 #endif
1838                 MPASS(error == 0 && uio.uio_resid == 0);
1839                 copied += uio.uio_offset;
1840                 resid -= uio.uio_offset;
1841                 m = m->m_next;
1842         }
1843         if (copied != 0) {
1844                 sbdrop_locked(sb, copied);
1845                 job->aio_received += copied;
1846                 job->msgrcv = 1;
1847                 copied = job->aio_received;
1848                 inp = sotoinpcb(so);
1849                 if (!INP_TRY_WLOCK(inp)) {
1850                         /*
1851                          * The reference on the socket file descriptor in
1852                          * the AIO job should keep 'sb' and 'inp' stable.
1853                          * Our caller has a reference on the 'toep' that
1854                          * keeps it stable.
1855                          */
1856                         SOCKBUF_UNLOCK(sb);
1857                         DDP_UNLOCK(toep);
1858                         INP_WLOCK(inp);
1859                         DDP_LOCK(toep);
1860                         SOCKBUF_LOCK(sb);
1861
1862                         /*
1863                          * If the socket has been closed, we should detect
1864                          * that and complete this request if needed on
1865                          * the next trip around the loop.
1866                          */
1867                 }
1868                 t4_rcvd_locked(&toep->td->tod, intotcpcb(inp));
1869                 INP_WUNLOCK(inp);
1870                 if (resid == 0 || toep->ddp.flags & DDP_DEAD) {
1871                         /*
1872                          * We filled the entire buffer with socket
1873                          * data, DDP is not being used, or the socket
1874                          * is being shut down, so complete the
1875                          * request.
1876                          */
1877                         SOCKBUF_UNLOCK(sb);
1878                         recycle_pageset(toep, ps);
1879                         aio_complete(job, copied, 0);
1880                         toep->ddp.queueing = NULL;
1881                         goto restart;
1882                 }
1883
1884                 /*
1885                  * If DDP is not enabled, requeue this request and restart.
1886                  * This will either enable DDP or wait for more data to
1887                  * arrive on the socket buffer.
1888                  */
1889                 if ((toep->ddp.flags & (DDP_ON | DDP_SC_REQ)) != DDP_ON) {
1890                         SOCKBUF_UNLOCK(sb);
1891                         recycle_pageset(toep, ps);
1892                         aio_ddp_requeue_one(toep, job);
1893                         toep->ddp.queueing = NULL;
1894                         goto restart;
1895                 }
1896
1897                 /*
1898                  * An indicate might have arrived and been added to
1899                  * the socket buffer while it was unlocked after the
1900                  * copy to lock the INP.  If so, restart the copy.
1901                  */
1902                 if (sbavail(sb) != 0)
1903                         goto sbcopy;
1904         }
1905         SOCKBUF_UNLOCK(sb);
1906
1907         if (prep_pageset(sc, toep, ps) == 0) {
1908                 recycle_pageset(toep, ps);
1909                 aio_ddp_requeue_one(toep, job);
1910                 toep->ddp.queueing = NULL;
1911
1912                 /*
1913                  * XXX: Need to retry this later.  Mostly need a trigger
1914                  * when page pods are freed up.
1915                  */
1916                 printf("%s: prep_pageset failed\n", __func__);
1917                 return;
1918         }
1919
1920         /* Determine which DDP buffer to use. */
1921         if (toep->ddp.db[0].job == NULL) {
1922                 db_idx = 0;
1923         } else {
1924                 MPASS(toep->ddp.db[1].job == NULL);
1925                 db_idx = 1;
1926         }
1927
1928         ddp_flags = 0;
1929         ddp_flags_mask = 0;
1930         if (db_idx == 0) {
1931                 ddp_flags |= V_TF_DDP_BUF0_VALID(1);
1932                 if (so->so_state & SS_NBIO)
1933                         ddp_flags |= V_TF_DDP_BUF0_FLUSH(1);
1934                 ddp_flags_mask |= V_TF_DDP_PSH_NO_INVALIDATE0(1) |
1935                     V_TF_DDP_PUSH_DISABLE_0(1) | V_TF_DDP_PSHF_ENABLE_0(1) |
1936                     V_TF_DDP_BUF0_FLUSH(1) | V_TF_DDP_BUF0_VALID(1);
1937                 buf_flag = DDP_BUF0_ACTIVE;
1938         } else {
1939                 ddp_flags |= V_TF_DDP_BUF1_VALID(1);
1940                 if (so->so_state & SS_NBIO)
1941                         ddp_flags |= V_TF_DDP_BUF1_FLUSH(1);
1942                 ddp_flags_mask |= V_TF_DDP_PSH_NO_INVALIDATE1(1) |
1943                     V_TF_DDP_PUSH_DISABLE_1(1) | V_TF_DDP_PSHF_ENABLE_1(1) |
1944                     V_TF_DDP_BUF1_FLUSH(1) | V_TF_DDP_BUF1_VALID(1);
1945                 buf_flag = DDP_BUF1_ACTIVE;
1946         }
1947         MPASS((toep->ddp.flags & buf_flag) == 0);
1948         if ((toep->ddp.flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE)) == 0) {
1949                 MPASS(db_idx == 0);
1950                 MPASS(toep->ddp.active_id == -1);
1951                 MPASS(toep->ddp.active_count == 0);
1952                 ddp_flags_mask |= V_TF_DDP_ACTIVE_BUF(1);
1953         }
1954
1955         /*
1956          * The TID for this connection should still be valid.  If DDP_DEAD
1957          * is set, SBS_CANTRCVMORE should be set, so we shouldn't be
1958          * this far anyway.  Even if the socket is closing on the other
1959          * end, the AIO job holds a reference on this end of the socket
1960          * which will keep it open and keep the TCP PCB attached until
1961          * after the job is completed.
1962          */
1963         wr = mk_update_tcb_for_ddp(sc, toep, db_idx, ps, job->aio_received,
1964             ddp_flags, ddp_flags_mask);
1965         if (wr == NULL) {
1966                 recycle_pageset(toep, ps);
1967                 aio_ddp_requeue_one(toep, job);
1968                 toep->ddp.queueing = NULL;
1969
1970                 /*
1971                  * XXX: Need a way to kick a retry here.
1972                  *
1973                  * XXX: We know the fixed size needed and could
1974                  * preallocate this using a blocking request at the
1975                  * start of the task to avoid having to handle this
1976                  * edge case.
1977                  */
1978                 printf("%s: mk_update_tcb_for_ddp failed\n", __func__);
1979                 return;
1980         }
1981
1982         if (!aio_set_cancel_function(job, t4_aio_cancel_active)) {
1983                 free_wrqe(wr);
1984                 recycle_pageset(toep, ps);
1985                 aio_ddp_cancel_one(job);
1986                 toep->ddp.queueing = NULL;
1987                 goto restart;
1988         }
1989
1990 #ifdef VERBOSE_TRACES
1991         CTR6(KTR_CXGBE,
1992             "%s: tid %u, scheduling %p for DDP[%d] (flags %#lx/%#lx)", __func__,
1993             toep->tid, job, db_idx, ddp_flags, ddp_flags_mask);
1994 #endif
1995         /* Give the chip the go-ahead. */
1996         t4_wrq_tx(sc, wr);
1997         db = &toep->ddp.db[db_idx];
1998         db->cancel_pending = 0;
1999         db->job = job;
2000         db->ps = ps;
2001         toep->ddp.queueing = NULL;
2002         toep->ddp.flags |= buf_flag;
2003         toep->ddp.active_count++;
2004         if (toep->ddp.active_count == 1) {
2005                 MPASS(toep->ddp.active_id == -1);
2006                 toep->ddp.active_id = db_idx;
2007                 CTR2(KTR_CXGBE, "%s: ddp_active_id = %d", __func__,
2008                     toep->ddp.active_id);
2009         }
2010         goto restart;
2011 }
2012
2013 void
2014 ddp_queue_toep(struct toepcb *toep)
2015 {
2016
2017         DDP_ASSERT_LOCKED(toep);
2018         if (toep->ddp.flags & DDP_TASK_ACTIVE)
2019                 return;
2020         toep->ddp.flags |= DDP_TASK_ACTIVE;
2021         hold_toepcb(toep);
2022         soaio_enqueue(&toep->ddp.requeue_task);
2023 }
2024
2025 static void
2026 aio_ddp_requeue_task(void *context, int pending)
2027 {
2028         struct toepcb *toep = context;
2029
2030         DDP_LOCK(toep);
2031         aio_ddp_requeue(toep);
2032         toep->ddp.flags &= ~DDP_TASK_ACTIVE;
2033         DDP_UNLOCK(toep);
2034
2035         free_toepcb(toep);
2036 }
2037
2038 static void
2039 t4_aio_cancel_active(struct kaiocb *job)
2040 {
2041         struct socket *so = job->fd_file->f_data;
2042         struct tcpcb *tp = so_sototcpcb(so);
2043         struct toepcb *toep = tp->t_toe;
2044         struct adapter *sc = td_adapter(toep->td);
2045         uint64_t valid_flag;
2046         int i;
2047
2048         DDP_LOCK(toep);
2049         if (aio_cancel_cleared(job)) {
2050                 DDP_UNLOCK(toep);
2051                 aio_ddp_cancel_one(job);
2052                 return;
2053         }
2054
2055         for (i = 0; i < nitems(toep->ddp.db); i++) {
2056                 if (toep->ddp.db[i].job == job) {
2057                         /* Should only ever get one cancel request for a job. */
2058                         MPASS(toep->ddp.db[i].cancel_pending == 0);
2059
2060                         /*
2061                          * Invalidate this buffer.  It will be
2062                          * cancelled or partially completed once the
2063                          * card ACKs the invalidate.
2064                          */
2065                         valid_flag = i == 0 ? V_TF_DDP_BUF0_VALID(1) :
2066                             V_TF_DDP_BUF1_VALID(1);
2067                         t4_set_tcb_field(sc, toep->ctrlq, toep,
2068                             W_TCB_RX_DDP_FLAGS, valid_flag, 0, 1,
2069                             CPL_COOKIE_DDP0 + i);
2070                         toep->ddp.db[i].cancel_pending = 1;
2071                         CTR2(KTR_CXGBE, "%s: request %p marked pending",
2072                             __func__, job);
2073                         break;
2074                 }
2075         }
2076         DDP_UNLOCK(toep);
2077 }
2078
2079 static void
2080 t4_aio_cancel_queued(struct kaiocb *job)
2081 {
2082         struct socket *so = job->fd_file->f_data;
2083         struct tcpcb *tp = so_sototcpcb(so);
2084         struct toepcb *toep = tp->t_toe;
2085
2086         DDP_LOCK(toep);
2087         if (!aio_cancel_cleared(job)) {
2088                 TAILQ_REMOVE(&toep->ddp.aiojobq, job, list);
2089                 toep->ddp.waiting_count--;
2090                 if (toep->ddp.waiting_count == 0)
2091                         ddp_queue_toep(toep);
2092         }
2093         CTR2(KTR_CXGBE, "%s: request %p cancelled", __func__, job);
2094         DDP_UNLOCK(toep);
2095
2096         aio_ddp_cancel_one(job);
2097 }
2098
2099 int
2100 t4_aio_queue_ddp(struct socket *so, struct kaiocb *job)
2101 {
2102         struct tcpcb *tp = so_sototcpcb(so);
2103         struct toepcb *toep = tp->t_toe;
2104
2105
2106         /* Ignore writes. */
2107         if (job->uaiocb.aio_lio_opcode != LIO_READ)
2108                 return (EOPNOTSUPP);
2109
2110         DDP_LOCK(toep);
2111
2112         /*
2113          * XXX: Think about possibly returning errors for ENOTCONN,
2114          * etc.  Perhaps the caller would only queue the request
2115          * if it failed with EOPNOTSUPP?
2116          */
2117
2118 #ifdef VERBOSE_TRACES
2119         CTR3(KTR_CXGBE, "%s: queueing %p for tid %u", __func__, job, toep->tid);
2120 #endif
2121         if (!aio_set_cancel_function(job, t4_aio_cancel_queued))
2122                 panic("new job was cancelled");
2123         TAILQ_INSERT_TAIL(&toep->ddp.aiojobq, job, list);
2124         toep->ddp.waiting_count++;
2125         toep->ddp.flags |= DDP_OK;
2126
2127         /*
2128          * Try to handle this request synchronously.  If this has
2129          * to block because the task is running, it will just bail
2130          * and let the task handle it instead.
2131          */
2132         aio_ddp_requeue(toep);
2133         DDP_UNLOCK(toep);
2134         return (0);
2135 }
2136
2137 void
2138 t4_ddp_mod_load(void)
2139 {
2140
2141         t4_register_shared_cpl_handler(CPL_SET_TCB_RPL, do_ddp_tcb_rpl,
2142             CPL_COOKIE_DDP0);
2143         t4_register_shared_cpl_handler(CPL_SET_TCB_RPL, do_ddp_tcb_rpl,
2144             CPL_COOKIE_DDP1);
2145         t4_register_cpl_handler(CPL_RX_DATA_DDP, do_rx_data_ddp);
2146         t4_register_cpl_handler(CPL_RX_DDP_COMPLETE, do_rx_ddp_complete);
2147         TAILQ_INIT(&ddp_orphan_pagesets);
2148         mtx_init(&ddp_orphan_pagesets_lock, "ddp orphans", NULL, MTX_DEF);
2149         TASK_INIT(&ddp_orphan_task, 0, ddp_free_orphan_pagesets, NULL);
2150 }
2151
2152 void
2153 t4_ddp_mod_unload(void)
2154 {
2155
2156         taskqueue_drain(taskqueue_thread, &ddp_orphan_task);
2157         MPASS(TAILQ_EMPTY(&ddp_orphan_pagesets));
2158         mtx_destroy(&ddp_orphan_pagesets_lock);
2159         t4_register_shared_cpl_handler(CPL_SET_TCB_RPL, NULL, CPL_COOKIE_DDP0);
2160         t4_register_shared_cpl_handler(CPL_SET_TCB_RPL, NULL, CPL_COOKIE_DDP1);
2161         t4_register_cpl_handler(CPL_RX_DATA_DDP, NULL);
2162         t4_register_cpl_handler(CPL_RX_DDP_COMPLETE, NULL);
2163 }
2164 #endif