sys/dev/cxgbe/tom/t4_ddp.c

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