]> CyberLeo.Net >> Repos - FreeBSD/FreeBSD.git/blob - sys/dev/cxgbe/tom/t4_ddp.c
MFV r350896: 6585 sha512, skein, and edonr have an unenforced dependency on extensibl...
[FreeBSD/FreeBSD.git] / 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 #ifdef TCP_OFFLOAD
66 #include "common/common.h"
67 #include "common/t4_msg.h"
68 #include "common/t4_regs.h"
69 #include "common/t4_tcb.h"
70 #include "tom/t4_tom.h"
71
72 /*
73  * Use the 'backend3' field in AIO jobs to store the amount of data
74  * received by the AIO job so far.
75  */
76 #define aio_received    backend3
77
78 static void aio_ddp_requeue_task(void *context, int pending);
79 static void ddp_complete_all(struct toepcb *toep, int error);
80 static void t4_aio_cancel_active(struct kaiocb *job);
81 static void t4_aio_cancel_queued(struct kaiocb *job);
82
83 static TAILQ_HEAD(, pageset) ddp_orphan_pagesets;
84 static struct mtx ddp_orphan_pagesets_lock;
85 static struct task ddp_orphan_task;
86
87 #define MAX_DDP_BUFFER_SIZE             (M_TCB_RX_DDP_BUF0_LEN)
88
89 /*
90  * A page set holds information about a buffer used for DDP.  The page
91  * set holds resources such as the VM pages backing the buffer (either
92  * held or wired) and the page pods associated with the buffer.
93  * Recently used page sets are cached to allow for efficient reuse of
94  * buffers (avoiding the need to re-fault in pages, hold them, etc.).
95  * Note that cached page sets keep the backing pages wired.  The
96  * number of wired pages is capped by only allowing for two wired
97  * pagesets per connection.  This is not a perfect cap, but is a
98  * trade-off for performance.
99  *
100  * If an application ping-pongs two buffers for a connection via
101  * aio_read(2) then those buffers should remain wired and expensive VM
102  * fault lookups should be avoided after each buffer has been used
103  * once.  If an application uses more than two buffers then this will
104  * fall back to doing expensive VM fault lookups for each operation.
105  */
106 static void
107 free_pageset(struct tom_data *td, struct pageset *ps)
108 {
109         vm_page_t p;
110         int i;
111
112         if (ps->prsv.prsv_nppods > 0)
113                 t4_free_page_pods(&ps->prsv);
114
115         for (i = 0; i < ps->npages; i++) {
116                 p = ps->pages[i];
117                 vm_page_lock(p);
118                 vm_page_unwire(p, PQ_INACTIVE);
119                 vm_page_unlock(p);
120         }
121         mtx_lock(&ddp_orphan_pagesets_lock);
122         TAILQ_INSERT_TAIL(&ddp_orphan_pagesets, ps, link);
123         taskqueue_enqueue(taskqueue_thread, &ddp_orphan_task);
124         mtx_unlock(&ddp_orphan_pagesets_lock);
125 }
126
127 static void
128 ddp_free_orphan_pagesets(void *context, int pending)
129 {
130         struct pageset *ps;
131
132         mtx_lock(&ddp_orphan_pagesets_lock);
133         while (!TAILQ_EMPTY(&ddp_orphan_pagesets)) {
134                 ps = TAILQ_FIRST(&ddp_orphan_pagesets);
135                 TAILQ_REMOVE(&ddp_orphan_pagesets, ps, link);
136                 mtx_unlock(&ddp_orphan_pagesets_lock);
137                 if (ps->vm)
138                         vmspace_free(ps->vm);
139                 free(ps, M_CXGBE);
140                 mtx_lock(&ddp_orphan_pagesets_lock);
141         }
142         mtx_unlock(&ddp_orphan_pagesets_lock);
143 }
144
145 static void
146 recycle_pageset(struct toepcb *toep, struct pageset *ps)
147 {
148
149         DDP_ASSERT_LOCKED(toep);
150         if (!(toep->ddp.flags & DDP_DEAD)) {
151                 KASSERT(toep->ddp.cached_count + toep->ddp.active_count <
152                     nitems(toep->ddp.db), ("too many wired pagesets"));
153                 TAILQ_INSERT_HEAD(&toep->ddp.cached_pagesets, ps, link);
154                 toep->ddp.cached_count++;
155         } else
156                 free_pageset(toep->td, ps);
157 }
158
159 static void
160 ddp_complete_one(struct kaiocb *job, int error)
161 {
162         long copied;
163
164         /*
165          * If this job had copied data out of the socket buffer before
166          * it was cancelled, report it as a short read rather than an
167          * error.
168          */
169         copied = job->aio_received;
170         if (copied != 0 || error == 0)
171                 aio_complete(job, copied, 0);
172         else
173                 aio_complete(job, -1, error);
174 }
175
176 static void
177 free_ddp_buffer(struct tom_data *td, struct ddp_buffer *db)
178 {
179
180         if (db->job) {
181                 /*
182                  * XXX: If we are un-offloading the socket then we
183                  * should requeue these on the socket somehow.  If we
184                  * got a FIN from the remote end, then this completes
185                  * any remaining requests with an EOF read.
186                  */
187                 if (!aio_clear_cancel_function(db->job))
188                         ddp_complete_one(db->job, 0);
189         }
190
191         if (db->ps)
192                 free_pageset(td, db->ps);
193 }
194
195 void
196 ddp_init_toep(struct toepcb *toep)
197 {
198
199         TAILQ_INIT(&toep->ddp.aiojobq);
200         TASK_INIT(&toep->ddp.requeue_task, 0, aio_ddp_requeue_task, toep);
201         toep->ddp.flags = DDP_OK;
202         toep->ddp.active_id = -1;
203         mtx_init(&toep->ddp.lock, "t4 ddp", NULL, MTX_DEF);
204 }
205
206 void
207 ddp_uninit_toep(struct toepcb *toep)
208 {
209
210         mtx_destroy(&toep->ddp.lock);
211 }
212
213 void
214 release_ddp_resources(struct toepcb *toep)
215 {
216         struct pageset *ps;
217         int i;
218
219         DDP_LOCK(toep);
220         toep->ddp.flags |= DDP_DEAD;
221         for (i = 0; i < nitems(toep->ddp.db); i++) {
222                 free_ddp_buffer(toep->td, &toep->ddp.db[i]);
223         }
224         while ((ps = TAILQ_FIRST(&toep->ddp.cached_pagesets)) != NULL) {
225                 TAILQ_REMOVE(&toep->ddp.cached_pagesets, ps, link);
226                 free_pageset(toep->td, ps);
227         }
228         ddp_complete_all(toep, 0);
229         DDP_UNLOCK(toep);
230 }
231
232 #ifdef INVARIANTS
233 void
234 ddp_assert_empty(struct toepcb *toep)
235 {
236         int i;
237
238         MPASS(!(toep->ddp.flags & DDP_TASK_ACTIVE));
239         for (i = 0; i < nitems(toep->ddp.db); i++) {
240                 MPASS(toep->ddp.db[i].job == NULL);
241                 MPASS(toep->ddp.db[i].ps == NULL);
242         }
243         MPASS(TAILQ_EMPTY(&toep->ddp.cached_pagesets));
244         MPASS(TAILQ_EMPTY(&toep->ddp.aiojobq));
245 }
246 #endif
247
248 static void
249 complete_ddp_buffer(struct toepcb *toep, struct ddp_buffer *db,
250     unsigned int db_idx)
251 {
252         unsigned int db_flag;
253
254         toep->ddp.active_count--;
255         if (toep->ddp.active_id == db_idx) {
256                 if (toep->ddp.active_count == 0) {
257                         KASSERT(toep->ddp.db[db_idx ^ 1].job == NULL,
258                             ("%s: active_count mismatch", __func__));
259                         toep->ddp.active_id = -1;
260                 } else
261                         toep->ddp.active_id ^= 1;
262 #ifdef VERBOSE_TRACES
263                 CTR3(KTR_CXGBE, "%s: tid %u, ddp_active_id = %d", __func__,
264                     toep->tid, toep->ddp.active_id);
265 #endif
266         } else {
267                 KASSERT(toep->ddp.active_count != 0 &&
268                     toep->ddp.active_id != -1,
269                     ("%s: active count mismatch", __func__));
270         }
271
272         db->cancel_pending = 0;
273         db->job = NULL;
274         recycle_pageset(toep, db->ps);
275         db->ps = NULL;
276
277         db_flag = db_idx == 1 ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE;
278         KASSERT(toep->ddp.flags & db_flag,
279             ("%s: DDP buffer not active. toep %p, ddp_flags 0x%x",
280             __func__, toep, toep->ddp.flags));
281         toep->ddp.flags &= ~db_flag;
282 }
283
284 /* XXX: handle_ddp_data code duplication */
285 void
286 insert_ddp_data(struct toepcb *toep, uint32_t n)
287 {
288         struct inpcb *inp = toep->inp;
289         struct tcpcb *tp = intotcpcb(inp);
290         struct ddp_buffer *db;
291         struct kaiocb *job;
292         size_t placed;
293         long copied;
294         unsigned int db_flag, db_idx;
295
296         INP_WLOCK_ASSERT(inp);
297         DDP_ASSERT_LOCKED(toep);
298
299         tp->rcv_nxt += n;
300 #ifndef USE_DDP_RX_FLOW_CONTROL
301         KASSERT(tp->rcv_wnd >= n, ("%s: negative window size", __func__));
302         tp->rcv_wnd -= n;
303 #endif
304         CTR2(KTR_CXGBE, "%s: placed %u bytes before falling out of DDP",
305             __func__, n);
306         while (toep->ddp.active_count > 0) {
307                 MPASS(toep->ddp.active_id != -1);
308                 db_idx = toep->ddp.active_id;
309                 db_flag = db_idx == 1 ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE;
310                 MPASS((toep->ddp.flags & db_flag) != 0);
311                 db = &toep->ddp.db[db_idx];
312                 job = db->job;
313                 copied = job->aio_received;
314                 placed = n;
315                 if (placed > job->uaiocb.aio_nbytes - copied)
316                         placed = job->uaiocb.aio_nbytes - copied;
317                 if (placed > 0)
318                         job->msgrcv = 1;
319                 if (!aio_clear_cancel_function(job)) {
320                         /*
321                          * Update the copied length for when
322                          * t4_aio_cancel_active() completes this
323                          * request.
324                          */
325                         job->aio_received += placed;
326                 } else if (copied + placed != 0) {
327                         CTR4(KTR_CXGBE,
328                             "%s: completing %p (copied %ld, placed %lu)",
329                             __func__, job, copied, placed);
330                         /* XXX: This always completes if there is some data. */
331                         aio_complete(job, copied + placed, 0);
332                 } else if (aio_set_cancel_function(job, t4_aio_cancel_queued)) {
333                         TAILQ_INSERT_HEAD(&toep->ddp.aiojobq, job, list);
334                         toep->ddp.waiting_count++;
335                 } else
336                         aio_cancel(job);
337                 n -= placed;
338                 complete_ddp_buffer(toep, db, db_idx);
339         }
340
341         MPASS(n == 0);
342 }
343
344 /* SET_TCB_FIELD sent as a ULP command looks like this */
345 #define LEN__SET_TCB_FIELD_ULP (sizeof(struct ulp_txpkt) + \
346     sizeof(struct ulptx_idata) + sizeof(struct cpl_set_tcb_field_core))
347
348 /* RX_DATA_ACK sent as a ULP command looks like this */
349 #define LEN__RX_DATA_ACK_ULP (sizeof(struct ulp_txpkt) + \
350     sizeof(struct ulptx_idata) + sizeof(struct cpl_rx_data_ack_core))
351
352 static inline void *
353 mk_set_tcb_field_ulp(struct ulp_txpkt *ulpmc, struct toepcb *toep,
354     uint64_t word, uint64_t mask, uint64_t val)
355 {
356         struct ulptx_idata *ulpsc;
357         struct cpl_set_tcb_field_core *req;
358
359         ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0));
360         ulpmc->len = htobe32(howmany(LEN__SET_TCB_FIELD_ULP, 16));
361
362         ulpsc = (struct ulptx_idata *)(ulpmc + 1);
363         ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
364         ulpsc->len = htobe32(sizeof(*req));
365
366         req = (struct cpl_set_tcb_field_core *)(ulpsc + 1);
367         OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_SET_TCB_FIELD, toep->tid));
368         req->reply_ctrl = htobe16(V_NO_REPLY(1) |
369             V_QUEUENO(toep->ofld_rxq->iq.abs_id));
370         req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0));
371         req->mask = htobe64(mask);
372         req->val = htobe64(val);
373
374         ulpsc = (struct ulptx_idata *)(req + 1);
375         if (LEN__SET_TCB_FIELD_ULP % 16) {
376                 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP));
377                 ulpsc->len = htobe32(0);
378                 return (ulpsc + 1);
379         }
380         return (ulpsc);
381 }
382
383 static inline void *
384 mk_rx_data_ack_ulp(struct ulp_txpkt *ulpmc, struct toepcb *toep)
385 {
386         struct ulptx_idata *ulpsc;
387         struct cpl_rx_data_ack_core *req;
388
389         ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0));
390         ulpmc->len = htobe32(howmany(LEN__RX_DATA_ACK_ULP, 16));
391
392         ulpsc = (struct ulptx_idata *)(ulpmc + 1);
393         ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
394         ulpsc->len = htobe32(sizeof(*req));
395
396         req = (struct cpl_rx_data_ack_core *)(ulpsc + 1);
397         OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_RX_DATA_ACK, toep->tid));
398         req->credit_dack = htobe32(F_RX_MODULATE_RX);
399
400         ulpsc = (struct ulptx_idata *)(req + 1);
401         if (LEN__RX_DATA_ACK_ULP % 16) {
402                 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP));
403                 ulpsc->len = htobe32(0);
404                 return (ulpsc + 1);
405         }
406         return (ulpsc);
407 }
408
409 static struct wrqe *
410 mk_update_tcb_for_ddp(struct adapter *sc, struct toepcb *toep, int db_idx,
411     struct pageset *ps, int offset, uint64_t ddp_flags, uint64_t ddp_flags_mask)
412 {
413         struct wrqe *wr;
414         struct work_request_hdr *wrh;
415         struct ulp_txpkt *ulpmc;
416         int len;
417
418         KASSERT(db_idx == 0 || db_idx == 1,
419             ("%s: bad DDP buffer index %d", __func__, db_idx));
420
421         /*
422          * We'll send a compound work request that has 3 SET_TCB_FIELDs and an
423          * RX_DATA_ACK (with RX_MODULATE to speed up delivery).
424          *
425          * The work request header is 16B and always ends at a 16B boundary.
426          * The ULPTX master commands that follow must all end at 16B boundaries
427          * too so we round up the size to 16.
428          */
429         len = sizeof(*wrh) + 3 * roundup2(LEN__SET_TCB_FIELD_ULP, 16) +
430             roundup2(LEN__RX_DATA_ACK_ULP, 16);
431
432         wr = alloc_wrqe(len, toep->ctrlq);
433         if (wr == NULL)
434                 return (NULL);
435         wrh = wrtod(wr);
436         INIT_ULPTX_WRH(wrh, len, 1, 0); /* atomic */
437         ulpmc = (struct ulp_txpkt *)(wrh + 1);
438
439         /* Write the buffer's tag */
440         ulpmc = mk_set_tcb_field_ulp(ulpmc, toep,
441             W_TCB_RX_DDP_BUF0_TAG + db_idx,
442             V_TCB_RX_DDP_BUF0_TAG(M_TCB_RX_DDP_BUF0_TAG),
443             V_TCB_RX_DDP_BUF0_TAG(ps->prsv.prsv_tag));
444
445         /* Update the current offset in the DDP buffer and its total length */
446         if (db_idx == 0)
447                 ulpmc = mk_set_tcb_field_ulp(ulpmc, toep,
448                     W_TCB_RX_DDP_BUF0_OFFSET,
449                     V_TCB_RX_DDP_BUF0_OFFSET(M_TCB_RX_DDP_BUF0_OFFSET) |
450                     V_TCB_RX_DDP_BUF0_LEN(M_TCB_RX_DDP_BUF0_LEN),
451                     V_TCB_RX_DDP_BUF0_OFFSET(offset) |
452                     V_TCB_RX_DDP_BUF0_LEN(ps->len));
453         else
454                 ulpmc = mk_set_tcb_field_ulp(ulpmc, toep,
455                     W_TCB_RX_DDP_BUF1_OFFSET,
456                     V_TCB_RX_DDP_BUF1_OFFSET(M_TCB_RX_DDP_BUF1_OFFSET) |
457                     V_TCB_RX_DDP_BUF1_LEN((u64)M_TCB_RX_DDP_BUF1_LEN << 32),
458                     V_TCB_RX_DDP_BUF1_OFFSET(offset) |
459                     V_TCB_RX_DDP_BUF1_LEN((u64)ps->len << 32));
460
461         /* Update DDP flags */
462         ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, W_TCB_RX_DDP_FLAGS,
463             ddp_flags_mask, ddp_flags);
464
465         /* Gratuitous RX_DATA_ACK with RX_MODULATE set to speed up delivery. */
466         ulpmc = mk_rx_data_ack_ulp(ulpmc, toep);
467
468         return (wr);
469 }
470
471 static int
472 handle_ddp_data(struct toepcb *toep, __be32 ddp_report, __be32 rcv_nxt, int len)
473 {
474         uint32_t report = be32toh(ddp_report);
475         unsigned int db_idx;
476         struct inpcb *inp = toep->inp;
477         struct ddp_buffer *db;
478         struct tcpcb *tp;
479         struct socket *so;
480         struct sockbuf *sb;
481         struct kaiocb *job;
482         long copied;
483
484         db_idx = report & F_DDP_BUF_IDX ? 1 : 0;
485
486         if (__predict_false(!(report & F_DDP_INV)))
487                 CXGBE_UNIMPLEMENTED("DDP buffer still valid");
488
489         INP_WLOCK(inp);
490         so = inp_inpcbtosocket(inp);
491         sb = &so->so_rcv;
492         DDP_LOCK(toep);
493
494         KASSERT(toep->ddp.active_id == db_idx,
495             ("completed DDP buffer (%d) != active_id (%d) for tid %d", db_idx,
496             toep->ddp.active_id, toep->tid));
497         db = &toep->ddp.db[db_idx];
498         job = db->job;
499
500         if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))) {
501                 /*
502                  * This can happen due to an administrative tcpdrop(8).
503                  * Just fail the request with ECONNRESET.
504                  */
505                 CTR5(KTR_CXGBE, "%s: tid %u, seq 0x%x, len %d, inp_flags 0x%x",
506                     __func__, toep->tid, be32toh(rcv_nxt), len, inp->inp_flags);
507                 if (aio_clear_cancel_function(job))
508                         ddp_complete_one(job, ECONNRESET);
509                 goto completed;
510         }
511
512         tp = intotcpcb(inp);
513
514         /*
515          * For RX_DDP_COMPLETE, len will be zero and rcv_nxt is the
516          * sequence number of the next byte to receive.  The length of
517          * the data received for this message must be computed by
518          * comparing the new and old values of rcv_nxt.
519          *
520          * For RX_DATA_DDP, len might be non-zero, but it is only the
521          * length of the most recent DMA.  It does not include the
522          * total length of the data received since the previous update
523          * for this DDP buffer.  rcv_nxt is the sequence number of the
524          * first received byte from the most recent DMA.
525          */
526         len += be32toh(rcv_nxt) - tp->rcv_nxt;
527         tp->rcv_nxt += len;
528         tp->t_rcvtime = ticks;
529 #ifndef USE_DDP_RX_FLOW_CONTROL
530         KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__));
531         tp->rcv_wnd -= len;
532 #endif
533 #ifdef VERBOSE_TRACES
534         CTR5(KTR_CXGBE, "%s: tid %u, DDP[%d] placed %d bytes (%#x)", __func__,
535             toep->tid, db_idx, len, report);
536 #endif
537
538         /* receive buffer autosize */
539         MPASS(toep->vnet == so->so_vnet);
540         CURVNET_SET(toep->vnet);
541         SOCKBUF_LOCK(sb);
542         if (sb->sb_flags & SB_AUTOSIZE &&
543             V_tcp_do_autorcvbuf &&
544             sb->sb_hiwat < V_tcp_autorcvbuf_max &&
545             len > (sbspace(sb) / 8 * 7)) {
546                 struct adapter *sc = td_adapter(toep->td);
547                 unsigned int hiwat = sb->sb_hiwat;
548                 unsigned int newsize = min(hiwat + sc->tt.autorcvbuf_inc,
549                     V_tcp_autorcvbuf_max);
550
551                 if (!sbreserve_locked(sb, newsize, so, NULL))
552                         sb->sb_flags &= ~SB_AUTOSIZE;
553         }
554         SOCKBUF_UNLOCK(sb);
555         CURVNET_RESTORE();
556
557         job->msgrcv = 1;
558         if (db->cancel_pending) {
559                 /*
560                  * Update the job's length but defer completion to the
561                  * TCB_RPL callback.
562                  */
563                 job->aio_received += len;
564                 goto out;
565         } else if (!aio_clear_cancel_function(job)) {
566                 /*
567                  * Update the copied length for when
568                  * t4_aio_cancel_active() completes this request.
569                  */
570                 job->aio_received += len;
571         } else {
572                 copied = job->aio_received;
573 #ifdef VERBOSE_TRACES
574                 CTR5(KTR_CXGBE,
575                     "%s: tid %u, completing %p (copied %ld, placed %d)",
576                     __func__, toep->tid, job, copied, len);
577 #endif
578                 aio_complete(job, copied + len, 0);
579                 t4_rcvd(&toep->td->tod, tp);
580         }
581
582 completed:
583         complete_ddp_buffer(toep, db, db_idx);
584         if (toep->ddp.waiting_count > 0)
585                 ddp_queue_toep(toep);
586 out:
587         DDP_UNLOCK(toep);
588         INP_WUNLOCK(inp);
589
590         return (0);
591 }
592
593 void
594 handle_ddp_indicate(struct toepcb *toep)
595 {
596
597         DDP_ASSERT_LOCKED(toep);
598         MPASS(toep->ddp.active_count == 0);
599         MPASS((toep->ddp.flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE)) == 0);
600         if (toep->ddp.waiting_count == 0) {
601                 /*
602                  * The pending requests that triggered the request for an
603                  * an indicate were cancelled.  Those cancels should have
604                  * already disabled DDP.  Just ignore this as the data is
605                  * going into the socket buffer anyway.
606                  */
607                 return;
608         }
609         CTR3(KTR_CXGBE, "%s: tid %d indicated (%d waiting)", __func__,
610             toep->tid, toep->ddp.waiting_count);
611         ddp_queue_toep(toep);
612 }
613
614 enum {
615         DDP_BUF0_INVALIDATED = 0x2,
616         DDP_BUF1_INVALIDATED
617 };
618
619 CTASSERT(DDP_BUF0_INVALIDATED == CPL_COOKIE_DDP0);
620
621 static int
622 do_ddp_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
623 {
624         struct adapter *sc = iq->adapter;
625         const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
626         unsigned int tid = GET_TID(cpl);
627         unsigned int db_idx;
628         struct toepcb *toep;
629         struct inpcb *inp;
630         struct ddp_buffer *db;
631         struct kaiocb *job;
632         long copied;
633
634         if (cpl->status != CPL_ERR_NONE)
635                 panic("XXX: tcp_rpl failed: %d", cpl->status);
636
637         toep = lookup_tid(sc, tid);
638         inp = toep->inp;
639         switch (cpl->cookie) {
640         case V_WORD(W_TCB_RX_DDP_FLAGS) | V_COOKIE(DDP_BUF0_INVALIDATED):
641         case V_WORD(W_TCB_RX_DDP_FLAGS) | V_COOKIE(DDP_BUF1_INVALIDATED):
642                 /*
643                  * XXX: This duplicates a lot of code with handle_ddp_data().
644                  */
645                 db_idx = G_COOKIE(cpl->cookie) - DDP_BUF0_INVALIDATED;
646                 MPASS(db_idx < nitems(toep->ddp.db));
647                 INP_WLOCK(inp);
648                 DDP_LOCK(toep);
649                 db = &toep->ddp.db[db_idx];
650
651                 /*
652                  * handle_ddp_data() should leave the job around until
653                  * this callback runs once a cancel is pending.
654                  */
655                 MPASS(db != NULL);
656                 MPASS(db->job != NULL);
657                 MPASS(db->cancel_pending);
658
659                 /*
660                  * XXX: It's not clear what happens if there is data
661                  * placed when the buffer is invalidated.  I suspect we
662                  * need to read the TCB to see how much data was placed.
663                  *
664                  * For now this just pretends like nothing was placed.
665                  *
666                  * XXX: Note that if we did check the PCB we would need to
667                  * also take care of updating the tp, etc.
668                  */
669                 job = db->job;
670                 copied = job->aio_received;
671                 if (copied == 0) {
672                         CTR2(KTR_CXGBE, "%s: cancelling %p", __func__, job);
673                         aio_cancel(job);
674                 } else {
675                         CTR3(KTR_CXGBE, "%s: completing %p (copied %ld)",
676                             __func__, job, copied);
677                         aio_complete(job, copied, 0);
678                         t4_rcvd(&toep->td->tod, intotcpcb(inp));
679                 }
680
681                 complete_ddp_buffer(toep, db, db_idx);
682                 if (toep->ddp.waiting_count > 0)
683                         ddp_queue_toep(toep);
684                 DDP_UNLOCK(toep);
685                 INP_WUNLOCK(inp);
686                 break;
687         default:
688                 panic("XXX: unknown tcb_rpl offset %#x, cookie %#x",
689                     G_WORD(cpl->cookie), G_COOKIE(cpl->cookie));
690         }
691
692         return (0);
693 }
694
695 void
696 handle_ddp_close(struct toepcb *toep, struct tcpcb *tp, __be32 rcv_nxt)
697 {
698         struct ddp_buffer *db;
699         struct kaiocb *job;
700         long copied;
701         unsigned int db_flag, db_idx;
702         int len, placed;
703
704         INP_WLOCK_ASSERT(toep->inp);
705         DDP_ASSERT_LOCKED(toep);
706
707         len = be32toh(rcv_nxt) - tp->rcv_nxt;
708         tp->rcv_nxt += len;
709
710         while (toep->ddp.active_count > 0) {
711                 MPASS(toep->ddp.active_id != -1);
712                 db_idx = toep->ddp.active_id;
713                 db_flag = db_idx == 1 ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE;
714                 MPASS((toep->ddp.flags & db_flag) != 0);
715                 db = &toep->ddp.db[db_idx];
716                 job = db->job;
717                 copied = job->aio_received;
718                 placed = len;
719                 if (placed > job->uaiocb.aio_nbytes - copied)
720                         placed = job->uaiocb.aio_nbytes - copied;
721                 if (placed > 0)
722                         job->msgrcv = 1;
723                 if (!aio_clear_cancel_function(job)) {
724                         /*
725                          * Update the copied length for when
726                          * t4_aio_cancel_active() completes this
727                          * request.
728                          */
729                         job->aio_received += placed;
730                 } else {
731                         CTR4(KTR_CXGBE, "%s: tid %d completed buf %d len %d",
732                             __func__, toep->tid, db_idx, placed);
733                         aio_complete(job, copied + placed, 0);
734                 }
735                 len -= placed;
736                 complete_ddp_buffer(toep, db, db_idx);
737         }
738
739         MPASS(len == 0);
740         ddp_complete_all(toep, 0);
741 }
742
743 #define DDP_ERR (F_DDP_PPOD_MISMATCH | F_DDP_LLIMIT_ERR | F_DDP_ULIMIT_ERR |\
744          F_DDP_PPOD_PARITY_ERR | F_DDP_PADDING_ERR | F_DDP_OFFSET_ERR |\
745          F_DDP_INVALID_TAG | F_DDP_COLOR_ERR | F_DDP_TID_MISMATCH |\
746          F_DDP_INVALID_PPOD | F_DDP_HDRCRC_ERR | F_DDP_DATACRC_ERR)
747
748 extern cpl_handler_t t4_cpl_handler[];
749
750 static int
751 do_rx_data_ddp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
752 {
753         struct adapter *sc = iq->adapter;
754         const struct cpl_rx_data_ddp *cpl = (const void *)(rss + 1);
755         unsigned int tid = GET_TID(cpl);
756         uint32_t vld;
757         struct toepcb *toep = lookup_tid(sc, tid);
758
759         KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
760         KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__));
761         KASSERT(!(toep->flags & TPF_SYNQE),
762             ("%s: toep %p claims to be a synq entry", __func__, toep));
763
764         vld = be32toh(cpl->ddpvld);
765         if (__predict_false(vld & DDP_ERR)) {
766                 panic("%s: DDP error 0x%x (tid %d, toep %p)",
767                     __func__, vld, tid, toep);
768         }
769
770         if (toep->ulp_mode == ULP_MODE_ISCSI) {
771                 t4_cpl_handler[CPL_RX_ISCSI_DDP](iq, rss, m);
772                 return (0);
773         }
774
775         handle_ddp_data(toep, cpl->u.ddp_report, cpl->seq, be16toh(cpl->len));
776
777         return (0);
778 }
779
780 static int
781 do_rx_ddp_complete(struct sge_iq *iq, const struct rss_header *rss,
782     struct mbuf *m)
783 {
784         struct adapter *sc = iq->adapter;
785         const struct cpl_rx_ddp_complete *cpl = (const void *)(rss + 1);
786         unsigned int tid = GET_TID(cpl);
787         struct toepcb *toep = lookup_tid(sc, tid);
788
789         KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
790         KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__));
791         KASSERT(!(toep->flags & TPF_SYNQE),
792             ("%s: toep %p claims to be a synq entry", __func__, toep));
793
794         handle_ddp_data(toep, cpl->ddp_report, cpl->rcv_nxt, 0);
795
796         return (0);
797 }
798
799 static void
800 enable_ddp(struct adapter *sc, struct toepcb *toep)
801 {
802
803         KASSERT((toep->ddp.flags & (DDP_ON | DDP_OK | DDP_SC_REQ)) == DDP_OK,
804             ("%s: toep %p has bad ddp_flags 0x%x",
805             __func__, toep, toep->ddp.flags));
806
807         CTR3(KTR_CXGBE, "%s: tid %u (time %u)",
808             __func__, toep->tid, time_uptime);
809
810         DDP_ASSERT_LOCKED(toep);
811         toep->ddp.flags |= DDP_SC_REQ;
812         t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_RX_DDP_FLAGS,
813             V_TF_DDP_OFF(1) | V_TF_DDP_INDICATE_OUT(1) |
814             V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1) |
815             V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_BUF1_VALID(1),
816             V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1), 0, 0);
817         t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_T_FLAGS,
818             V_TF_RCV_COALESCE_ENABLE(1), 0, 0, 0);
819 }
820
821 static int
822 calculate_hcf(int n1, int n2)
823 {
824         int a, b, t;
825
826         if (n1 <= n2) {
827                 a = n1;
828                 b = n2;
829         } else {
830                 a = n2;
831                 b = n1;
832         }
833
834         while (a != 0) {
835                 t = a;
836                 a = b % a;
837                 b = t;
838         }
839
840         return (b);
841 }
842
843 static inline int
844 pages_to_nppods(int npages, int ddp_page_shift)
845 {
846
847         MPASS(ddp_page_shift >= PAGE_SHIFT);
848
849         return (howmany(npages >> (ddp_page_shift - PAGE_SHIFT), PPOD_PAGES));
850 }
851
852 static int
853 alloc_page_pods(struct ppod_region *pr, u_int nppods, u_int pgsz_idx,
854     struct ppod_reservation *prsv)
855 {
856         vmem_addr_t addr;       /* relative to start of region */
857
858         if (vmem_alloc(pr->pr_arena, PPOD_SZ(nppods), M_NOWAIT | M_FIRSTFIT,
859             &addr) != 0)
860                 return (ENOMEM);
861
862         CTR5(KTR_CXGBE, "%-17s arena %p, addr 0x%08x, nppods %d, pgsz %d",
863             __func__, pr->pr_arena, (uint32_t)addr & pr->pr_tag_mask,
864             nppods, 1 << pr->pr_page_shift[pgsz_idx]);
865
866         /*
867          * The hardware tagmask includes an extra invalid bit but the arena was
868          * seeded with valid values only.  An allocation out of this arena will
869          * fit inside the tagmask but won't have the invalid bit set.
870          */
871         MPASS((addr & pr->pr_tag_mask) == addr);
872         MPASS((addr & pr->pr_invalid_bit) == 0);
873
874         prsv->prsv_pr = pr;
875         prsv->prsv_tag = V_PPOD_PGSZ(pgsz_idx) | addr;
876         prsv->prsv_nppods = nppods;
877
878         return (0);
879 }
880
881 int
882 t4_alloc_page_pods_for_ps(struct ppod_region *pr, struct pageset *ps)
883 {
884         int i, hcf, seglen, idx, nppods;
885         struct ppod_reservation *prsv = &ps->prsv;
886
887         KASSERT(prsv->prsv_nppods == 0,
888             ("%s: page pods already allocated", __func__));
889
890         /*
891          * The DDP page size is unrelated to the VM page size.  We combine
892          * contiguous physical pages into larger segments to get the best DDP
893          * page size possible.  This is the largest of the four sizes in
894          * A_ULP_RX_TDDP_PSZ that evenly divides the HCF of the segment sizes in
895          * the page list.
896          */
897         hcf = 0;
898         for (i = 0; i < ps->npages; i++) {
899                 seglen = PAGE_SIZE;
900                 while (i < ps->npages - 1 &&
901                     ps->pages[i]->phys_addr + PAGE_SIZE ==
902                     ps->pages[i + 1]->phys_addr) {
903                         seglen += PAGE_SIZE;
904                         i++;
905                 }
906
907                 hcf = calculate_hcf(hcf, seglen);
908                 if (hcf < (1 << pr->pr_page_shift[1])) {
909                         idx = 0;
910                         goto have_pgsz; /* give up, short circuit */
911                 }
912         }
913
914 #define PR_PAGE_MASK(x) ((1 << pr->pr_page_shift[(x)]) - 1)
915         MPASS((hcf & PR_PAGE_MASK(0)) == 0); /* PAGE_SIZE is >= 4K everywhere */
916         for (idx = nitems(pr->pr_page_shift) - 1; idx > 0; idx--) {
917                 if ((hcf & PR_PAGE_MASK(idx)) == 0)
918                         break;
919         }
920 #undef PR_PAGE_MASK
921
922 have_pgsz:
923         MPASS(idx <= M_PPOD_PGSZ);
924
925         nppods = pages_to_nppods(ps->npages, pr->pr_page_shift[idx]);
926         if (alloc_page_pods(pr, nppods, idx, prsv) != 0)
927                 return (0);
928         MPASS(prsv->prsv_nppods > 0);
929
930         return (1);
931 }
932
933 int
934 t4_alloc_page_pods_for_buf(struct ppod_region *pr, vm_offset_t buf, int len,
935     struct ppod_reservation *prsv)
936 {
937         int hcf, seglen, idx, npages, nppods;
938         uintptr_t start_pva, end_pva, pva, p1;
939
940         MPASS(buf > 0);
941         MPASS(len > 0);
942
943         /*
944          * The DDP page size is unrelated to the VM page size.  We combine
945          * contiguous physical pages into larger segments to get the best DDP
946          * page size possible.  This is the largest of the four sizes in
947          * A_ULP_RX_ISCSI_PSZ that evenly divides the HCF of the segment sizes
948          * in the page list.
949          */
950         hcf = 0;
951         start_pva = trunc_page(buf);
952         end_pva = trunc_page(buf + len - 1);
953         pva = start_pva;
954         while (pva <= end_pva) {
955                 seglen = PAGE_SIZE;
956                 p1 = pmap_kextract(pva);
957                 pva += PAGE_SIZE;
958                 while (pva <= end_pva && p1 + seglen == pmap_kextract(pva)) {
959                         seglen += PAGE_SIZE;
960                         pva += PAGE_SIZE;
961                 }
962
963                 hcf = calculate_hcf(hcf, seglen);
964                 if (hcf < (1 << pr->pr_page_shift[1])) {
965                         idx = 0;
966                         goto have_pgsz; /* give up, short circuit */
967                 }
968         }
969
970 #define PR_PAGE_MASK(x) ((1 << pr->pr_page_shift[(x)]) - 1)
971         MPASS((hcf & PR_PAGE_MASK(0)) == 0); /* PAGE_SIZE is >= 4K everywhere */
972         for (idx = nitems(pr->pr_page_shift) - 1; idx > 0; idx--) {
973                 if ((hcf & PR_PAGE_MASK(idx)) == 0)
974                         break;
975         }
976 #undef PR_PAGE_MASK
977
978 have_pgsz:
979         MPASS(idx <= M_PPOD_PGSZ);
980
981         npages = 1;
982         npages += (end_pva - start_pva) >> pr->pr_page_shift[idx];
983         nppods = howmany(npages, PPOD_PAGES);
984         if (alloc_page_pods(pr, nppods, idx, prsv) != 0)
985                 return (ENOMEM);
986         MPASS(prsv->prsv_nppods > 0);
987
988         return (0);
989 }
990
991 void
992 t4_free_page_pods(struct ppod_reservation *prsv)
993 {
994         struct ppod_region *pr = prsv->prsv_pr;
995         vmem_addr_t addr;
996
997         MPASS(prsv != NULL);
998         MPASS(prsv->prsv_nppods != 0);
999
1000         addr = prsv->prsv_tag & pr->pr_tag_mask;
1001         MPASS((addr & pr->pr_invalid_bit) == 0);
1002
1003         CTR4(KTR_CXGBE, "%-17s arena %p, addr 0x%08x, nppods %d", __func__,
1004             pr->pr_arena, addr, prsv->prsv_nppods);
1005
1006         vmem_free(pr->pr_arena, addr, PPOD_SZ(prsv->prsv_nppods));
1007         prsv->prsv_nppods = 0;
1008 }
1009
1010 #define NUM_ULP_TX_SC_IMM_PPODS (256 / PPOD_SIZE)
1011
1012 int
1013 t4_write_page_pods_for_ps(struct adapter *sc, struct sge_wrq *wrq, int tid,
1014     struct pageset *ps)
1015 {
1016         struct wrqe *wr;
1017         struct ulp_mem_io *ulpmc;
1018         struct ulptx_idata *ulpsc;
1019         struct pagepod *ppod;
1020         int i, j, k, n, chunk, len, ddp_pgsz, idx;
1021         u_int ppod_addr;
1022         uint32_t cmd;
1023         struct ppod_reservation *prsv = &ps->prsv;
1024         struct ppod_region *pr = prsv->prsv_pr;
1025
1026         KASSERT(!(ps->flags & PS_PPODS_WRITTEN),
1027             ("%s: page pods already written", __func__));
1028         MPASS(prsv->prsv_nppods > 0);
1029
1030         cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE));
1031         if (is_t4(sc))
1032                 cmd |= htobe32(F_ULP_MEMIO_ORDER);
1033         else
1034                 cmd |= htobe32(F_T5_ULP_MEMIO_IMM);
1035         ddp_pgsz = 1 << pr->pr_page_shift[G_PPOD_PGSZ(prsv->prsv_tag)];
1036         ppod_addr = pr->pr_start + (prsv->prsv_tag & pr->pr_tag_mask);
1037         for (i = 0; i < prsv->prsv_nppods; ppod_addr += chunk) {
1038
1039                 /* How many page pods are we writing in this cycle */
1040                 n = min(prsv->prsv_nppods - i, NUM_ULP_TX_SC_IMM_PPODS);
1041                 chunk = PPOD_SZ(n);
1042                 len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16);
1043
1044                 wr = alloc_wrqe(len, wrq);
1045                 if (wr == NULL)
1046                         return (ENOMEM);        /* ok to just bail out */
1047                 ulpmc = wrtod(wr);
1048
1049                 INIT_ULPTX_WR(ulpmc, len, 0, 0);
1050                 ulpmc->cmd = cmd;
1051                 ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32));
1052                 ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16));
1053                 ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5));
1054
1055                 ulpsc = (struct ulptx_idata *)(ulpmc + 1);
1056                 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
1057                 ulpsc->len = htobe32(chunk);
1058
1059                 ppod = (struct pagepod *)(ulpsc + 1);
1060                 for (j = 0; j < n; i++, j++, ppod++) {
1061                         ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID |
1062                             V_PPOD_TID(tid) | prsv->prsv_tag);
1063                         ppod->len_offset = htobe64(V_PPOD_LEN(ps->len) |
1064                             V_PPOD_OFST(ps->offset));
1065                         ppod->rsvd = 0;
1066                         idx = i * PPOD_PAGES * (ddp_pgsz / PAGE_SIZE);
1067                         for (k = 0; k < nitems(ppod->addr); k++) {
1068                                 if (idx < ps->npages) {
1069                                         ppod->addr[k] =
1070                                             htobe64(ps->pages[idx]->phys_addr);
1071                                         idx += ddp_pgsz / PAGE_SIZE;
1072                                 } else
1073                                         ppod->addr[k] = 0;
1074 #if 0
1075                                 CTR5(KTR_CXGBE,
1076                                     "%s: tid %d ppod[%d]->addr[%d] = %p",
1077                                     __func__, toep->tid, i, k,
1078                                     htobe64(ppod->addr[k]));
1079 #endif
1080                         }
1081
1082                 }
1083
1084                 t4_wrq_tx(sc, wr);
1085         }
1086         ps->flags |= PS_PPODS_WRITTEN;
1087
1088         return (0);
1089 }
1090
1091 int
1092 t4_write_page_pods_for_buf(struct adapter *sc, struct sge_wrq *wrq, int tid,
1093     struct ppod_reservation *prsv, vm_offset_t buf, int buflen)
1094 {
1095         struct wrqe *wr;
1096         struct ulp_mem_io *ulpmc;
1097         struct ulptx_idata *ulpsc;
1098         struct pagepod *ppod;
1099         int i, j, k, n, chunk, len, ddp_pgsz;
1100         u_int ppod_addr, offset;
1101         uint32_t cmd;
1102         struct ppod_region *pr = prsv->prsv_pr;
1103         uintptr_t end_pva, pva, pa;
1104
1105         cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE));
1106         if (is_t4(sc))
1107                 cmd |= htobe32(F_ULP_MEMIO_ORDER);
1108         else
1109                 cmd |= htobe32(F_T5_ULP_MEMIO_IMM);
1110         ddp_pgsz = 1 << pr->pr_page_shift[G_PPOD_PGSZ(prsv->prsv_tag)];
1111         offset = buf & PAGE_MASK;
1112         ppod_addr = pr->pr_start + (prsv->prsv_tag & pr->pr_tag_mask);
1113         pva = trunc_page(buf);
1114         end_pva = trunc_page(buf + buflen - 1);
1115         for (i = 0; i < prsv->prsv_nppods; ppod_addr += chunk) {
1116
1117                 /* How many page pods are we writing in this cycle */
1118                 n = min(prsv->prsv_nppods - i, NUM_ULP_TX_SC_IMM_PPODS);
1119                 MPASS(n > 0);
1120                 chunk = PPOD_SZ(n);
1121                 len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16);
1122
1123                 wr = alloc_wrqe(len, wrq);
1124                 if (wr == NULL)
1125                         return (ENOMEM);        /* ok to just bail out */
1126                 ulpmc = wrtod(wr);
1127
1128                 INIT_ULPTX_WR(ulpmc, len, 0, 0);
1129                 ulpmc->cmd = cmd;
1130                 ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32));
1131                 ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16));
1132                 ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5));
1133
1134                 ulpsc = (struct ulptx_idata *)(ulpmc + 1);
1135                 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
1136                 ulpsc->len = htobe32(chunk);
1137
1138                 ppod = (struct pagepod *)(ulpsc + 1);
1139                 for (j = 0; j < n; i++, j++, ppod++) {
1140                         ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID |
1141                             V_PPOD_TID(tid) |
1142                             (prsv->prsv_tag & ~V_PPOD_PGSZ(M_PPOD_PGSZ)));
1143                         ppod->len_offset = htobe64(V_PPOD_LEN(buflen) |
1144                             V_PPOD_OFST(offset));
1145                         ppod->rsvd = 0;
1146
1147                         for (k = 0; k < nitems(ppod->addr); k++) {
1148                                 if (pva > end_pva)
1149                                         ppod->addr[k] = 0;
1150                                 else {
1151                                         pa = pmap_kextract(pva);
1152                                         ppod->addr[k] = htobe64(pa);
1153                                         pva += ddp_pgsz;
1154                                 }
1155 #if 0
1156                                 CTR5(KTR_CXGBE,
1157                                     "%s: tid %d ppod[%d]->addr[%d] = %p",
1158                                     __func__, tid, i, k,
1159                                     htobe64(ppod->addr[k]));
1160 #endif
1161                         }
1162
1163                         /*
1164                          * Walk back 1 segment so that the first address in the
1165                          * next pod is the same as the last one in the current
1166                          * pod.
1167                          */
1168                         pva -= ddp_pgsz;
1169                 }
1170
1171                 t4_wrq_tx(sc, wr);
1172         }
1173
1174         MPASS(pva <= end_pva);
1175
1176         return (0);
1177 }
1178
1179 /*
1180  * Prepare a pageset for DDP.  This sets up page pods.
1181  */
1182 static int
1183 prep_pageset(struct adapter *sc, struct toepcb *toep, struct pageset *ps)
1184 {
1185         struct tom_data *td = sc->tom_softc;
1186
1187         if (ps->prsv.prsv_nppods == 0 &&
1188             !t4_alloc_page_pods_for_ps(&td->pr, ps)) {
1189                 return (0);
1190         }
1191         if (!(ps->flags & PS_PPODS_WRITTEN) &&
1192             t4_write_page_pods_for_ps(sc, toep->ctrlq, toep->tid, ps) != 0) {
1193                 return (0);
1194         }
1195
1196         return (1);
1197 }
1198
1199 int
1200 t4_init_ppod_region(struct ppod_region *pr, struct t4_range *r, u_int psz,
1201     const char *name)
1202 {
1203         int i;
1204
1205         MPASS(pr != NULL);
1206         MPASS(r->size > 0);
1207
1208         pr->pr_start = r->start;
1209         pr->pr_len = r->size;
1210         pr->pr_page_shift[0] = 12 + G_HPZ0(psz);
1211         pr->pr_page_shift[1] = 12 + G_HPZ1(psz);
1212         pr->pr_page_shift[2] = 12 + G_HPZ2(psz);
1213         pr->pr_page_shift[3] = 12 + G_HPZ3(psz);
1214
1215         /* The SGL -> page pod algorithm requires the sizes to be in order. */
1216         for (i = 1; i < nitems(pr->pr_page_shift); i++) {
1217                 if (pr->pr_page_shift[i] <= pr->pr_page_shift[i - 1])
1218                         return (ENXIO);
1219         }
1220
1221         pr->pr_tag_mask = ((1 << fls(r->size)) - 1) & V_PPOD_TAG(M_PPOD_TAG);
1222         pr->pr_alias_mask = V_PPOD_TAG(M_PPOD_TAG) & ~pr->pr_tag_mask;
1223         if (pr->pr_tag_mask == 0 || pr->pr_alias_mask == 0)
1224                 return (ENXIO);
1225         pr->pr_alias_shift = fls(pr->pr_tag_mask);
1226         pr->pr_invalid_bit = 1 << (pr->pr_alias_shift - 1);
1227
1228         pr->pr_arena = vmem_create(name, 0, pr->pr_len, PPOD_SIZE, 0,
1229             M_FIRSTFIT | M_NOWAIT);
1230         if (pr->pr_arena == NULL)
1231                 return (ENOMEM);
1232
1233         return (0);
1234 }
1235
1236 void
1237 t4_free_ppod_region(struct ppod_region *pr)
1238 {
1239
1240         MPASS(pr != NULL);
1241
1242         if (pr->pr_arena)
1243                 vmem_destroy(pr->pr_arena);
1244         bzero(pr, sizeof(*pr));
1245 }
1246
1247 static int
1248 pscmp(struct pageset *ps, struct vmspace *vm, vm_offset_t start, int npages,
1249     int pgoff, int len)
1250 {
1251
1252         if (ps->start != start || ps->npages != npages ||
1253             ps->offset != pgoff || ps->len != len)
1254                 return (1);
1255
1256         return (ps->vm != vm || ps->vm_timestamp != vm->vm_map.timestamp);
1257 }
1258
1259 static int
1260 hold_aio(struct toepcb *toep, struct kaiocb *job, struct pageset **pps)
1261 {
1262         struct vmspace *vm;
1263         vm_map_t map;
1264         vm_offset_t start, end, pgoff;
1265         struct pageset *ps;
1266         int n;
1267
1268         DDP_ASSERT_LOCKED(toep);
1269
1270         /*
1271          * The AIO subsystem will cancel and drain all requests before
1272          * permitting a process to exit or exec, so p_vmspace should
1273          * be stable here.
1274          */
1275         vm = job->userproc->p_vmspace;
1276         map = &vm->vm_map;
1277         start = (uintptr_t)job->uaiocb.aio_buf;
1278         pgoff = start & PAGE_MASK;
1279         end = round_page(start + job->uaiocb.aio_nbytes);
1280         start = trunc_page(start);
1281
1282         if (end - start > MAX_DDP_BUFFER_SIZE) {
1283                 /*
1284                  * Truncate the request to a short read.
1285                  * Alternatively, we could DDP in chunks to the larger
1286                  * buffer, but that would be quite a bit more work.
1287                  *
1288                  * When truncating, round the request down to avoid
1289                  * crossing a cache line on the final transaction.
1290                  */
1291                 end = rounddown2(start + MAX_DDP_BUFFER_SIZE, CACHE_LINE_SIZE);
1292 #ifdef VERBOSE_TRACES
1293                 CTR4(KTR_CXGBE, "%s: tid %d, truncating size from %lu to %lu",
1294                     __func__, toep->tid, (unsigned long)job->uaiocb.aio_nbytes,
1295                     (unsigned long)(end - (start + pgoff)));
1296                 job->uaiocb.aio_nbytes = end - (start + pgoff);
1297 #endif
1298                 end = round_page(end);
1299         }
1300
1301         n = atop(end - start);
1302
1303         /*
1304          * Try to reuse a cached pageset.
1305          */
1306         TAILQ_FOREACH(ps, &toep->ddp.cached_pagesets, link) {
1307                 if (pscmp(ps, vm, start, n, pgoff,
1308                     job->uaiocb.aio_nbytes) == 0) {
1309                         TAILQ_REMOVE(&toep->ddp.cached_pagesets, ps, link);
1310                         toep->ddp.cached_count--;
1311                         *pps = ps;
1312                         return (0);
1313                 }
1314         }
1315
1316         /*
1317          * If there are too many cached pagesets to create a new one,
1318          * free a pageset before creating a new one.
1319          */
1320         KASSERT(toep->ddp.active_count + toep->ddp.cached_count <=
1321             nitems(toep->ddp.db), ("%s: too many wired pagesets", __func__));
1322         if (toep->ddp.active_count + toep->ddp.cached_count ==
1323             nitems(toep->ddp.db)) {
1324                 KASSERT(toep->ddp.cached_count > 0,
1325                     ("no cached pageset to free"));
1326                 ps = TAILQ_LAST(&toep->ddp.cached_pagesets, pagesetq);
1327                 TAILQ_REMOVE(&toep->ddp.cached_pagesets, ps, link);
1328                 toep->ddp.cached_count--;
1329                 free_pageset(toep->td, ps);
1330         }
1331         DDP_UNLOCK(toep);
1332
1333         /* Create a new pageset. */
1334         ps = malloc(sizeof(*ps) + n * sizeof(vm_page_t), M_CXGBE, M_WAITOK |
1335             M_ZERO);
1336         ps->pages = (vm_page_t *)(ps + 1);
1337         ps->vm_timestamp = map->timestamp;
1338         ps->npages = vm_fault_quick_hold_pages(map, start, end - start,
1339             VM_PROT_WRITE, ps->pages, n);
1340
1341         DDP_LOCK(toep);
1342         if (ps->npages < 0) {
1343                 free(ps, M_CXGBE);
1344                 return (EFAULT);
1345         }
1346
1347         KASSERT(ps->npages == n, ("hold_aio: page count mismatch: %d vs %d",
1348             ps->npages, n));
1349
1350         ps->offset = pgoff;
1351         ps->len = job->uaiocb.aio_nbytes;
1352         atomic_add_int(&vm->vm_refcnt, 1);
1353         ps->vm = vm;
1354         ps->start = start;
1355
1356         CTR5(KTR_CXGBE, "%s: tid %d, new pageset %p for job %p, npages %d",
1357             __func__, toep->tid, ps, job, ps->npages);
1358         *pps = ps;
1359         return (0);
1360 }
1361
1362 static void
1363 ddp_complete_all(struct toepcb *toep, int error)
1364 {
1365         struct kaiocb *job;
1366
1367         DDP_ASSERT_LOCKED(toep);
1368         while (!TAILQ_EMPTY(&toep->ddp.aiojobq)) {
1369                 job = TAILQ_FIRST(&toep->ddp.aiojobq);
1370                 TAILQ_REMOVE(&toep->ddp.aiojobq, job, list);
1371                 toep->ddp.waiting_count--;
1372                 if (aio_clear_cancel_function(job))
1373                         ddp_complete_one(job, error);
1374         }
1375 }
1376
1377 static void
1378 aio_ddp_cancel_one(struct kaiocb *job)
1379 {
1380         long copied;
1381
1382         /*
1383          * If this job had copied data out of the socket buffer before
1384          * it was cancelled, report it as a short read rather than an
1385          * error.
1386          */
1387         copied = job->aio_received;
1388         if (copied != 0)
1389                 aio_complete(job, copied, 0);
1390         else
1391                 aio_cancel(job);
1392 }
1393
1394 /*
1395  * Called when the main loop wants to requeue a job to retry it later.
1396  * Deals with the race of the job being cancelled while it was being
1397  * examined.
1398  */
1399 static void
1400 aio_ddp_requeue_one(struct toepcb *toep, struct kaiocb *job)
1401 {
1402
1403         DDP_ASSERT_LOCKED(toep);
1404         if (!(toep->ddp.flags & DDP_DEAD) &&
1405             aio_set_cancel_function(job, t4_aio_cancel_queued)) {
1406                 TAILQ_INSERT_HEAD(&toep->ddp.aiojobq, job, list);
1407                 toep->ddp.waiting_count++;
1408         } else
1409                 aio_ddp_cancel_one(job);
1410 }
1411
1412 static void
1413 aio_ddp_requeue(struct toepcb *toep)
1414 {
1415         struct adapter *sc = td_adapter(toep->td);
1416         struct socket *so;
1417         struct sockbuf *sb;
1418         struct inpcb *inp;
1419         struct kaiocb *job;
1420         struct ddp_buffer *db;
1421         size_t copied, offset, resid;
1422         struct pageset *ps;
1423         struct mbuf *m;
1424         uint64_t ddp_flags, ddp_flags_mask;
1425         struct wrqe *wr;
1426         int buf_flag, db_idx, error;
1427
1428         DDP_ASSERT_LOCKED(toep);
1429
1430 restart:
1431         if (toep->ddp.flags & DDP_DEAD) {
1432                 MPASS(toep->ddp.waiting_count == 0);
1433                 MPASS(toep->ddp.active_count == 0);
1434                 return;
1435         }
1436
1437         if (toep->ddp.waiting_count == 0 ||
1438             toep->ddp.active_count == nitems(toep->ddp.db)) {
1439                 return;
1440         }
1441
1442         job = TAILQ_FIRST(&toep->ddp.aiojobq);
1443         so = job->fd_file->f_data;
1444         sb = &so->so_rcv;
1445         SOCKBUF_LOCK(sb);
1446
1447         /* We will never get anything unless we are or were connected. */
1448         if (!(so->so_state & (SS_ISCONNECTED|SS_ISDISCONNECTED))) {
1449                 SOCKBUF_UNLOCK(sb);
1450                 ddp_complete_all(toep, ENOTCONN);
1451                 return;
1452         }
1453
1454         KASSERT(toep->ddp.active_count == 0 || sbavail(sb) == 0,
1455             ("%s: pending sockbuf data and DDP is active", __func__));
1456
1457         /* Abort if socket has reported problems. */
1458         /* XXX: Wait for any queued DDP's to finish and/or flush them? */
1459         if (so->so_error && sbavail(sb) == 0) {
1460                 toep->ddp.waiting_count--;
1461                 TAILQ_REMOVE(&toep->ddp.aiojobq, job, list);
1462                 if (!aio_clear_cancel_function(job)) {
1463                         SOCKBUF_UNLOCK(sb);
1464                         goto restart;
1465                 }
1466
1467                 /*
1468                  * If this job has previously copied some data, report
1469                  * a short read and leave the error to be reported by
1470                  * a future request.
1471                  */
1472                 copied = job->aio_received;
1473                 if (copied != 0) {
1474                         SOCKBUF_UNLOCK(sb);
1475                         aio_complete(job, copied, 0);
1476                         goto restart;
1477                 }
1478                 error = so->so_error;
1479                 so->so_error = 0;
1480                 SOCKBUF_UNLOCK(sb);
1481                 aio_complete(job, -1, error);
1482                 goto restart;
1483         }
1484
1485         /*
1486          * Door is closed.  If there is pending data in the socket buffer,
1487          * deliver it.  If there are pending DDP requests, wait for those
1488          * to complete.  Once they have completed, return EOF reads.
1489          */
1490         if (sb->sb_state & SBS_CANTRCVMORE && sbavail(sb) == 0) {
1491                 SOCKBUF_UNLOCK(sb);
1492                 if (toep->ddp.active_count != 0)
1493                         return;
1494                 ddp_complete_all(toep, 0);
1495                 return;
1496         }
1497
1498         /*
1499          * If DDP is not enabled and there is no pending socket buffer
1500          * data, try to enable DDP.
1501          */
1502         if (sbavail(sb) == 0 && (toep->ddp.flags & DDP_ON) == 0) {
1503                 SOCKBUF_UNLOCK(sb);
1504
1505                 /*
1506                  * Wait for the card to ACK that DDP is enabled before
1507                  * queueing any buffers.  Currently this waits for an
1508                  * indicate to arrive.  This could use a TCB_SET_FIELD_RPL
1509                  * message to know that DDP was enabled instead of waiting
1510                  * for the indicate which would avoid copying the indicate
1511                  * if no data is pending.
1512                  *
1513                  * XXX: Might want to limit the indicate size to the size
1514                  * of the first queued request.
1515                  */
1516                 if ((toep->ddp.flags & DDP_SC_REQ) == 0)
1517                         enable_ddp(sc, toep);
1518                 return;
1519         }
1520         SOCKBUF_UNLOCK(sb);
1521
1522         /*
1523          * If another thread is queueing a buffer for DDP, let it
1524          * drain any work and return.
1525          */
1526         if (toep->ddp.queueing != NULL)
1527                 return;
1528
1529         /* Take the next job to prep it for DDP. */
1530         toep->ddp.waiting_count--;
1531         TAILQ_REMOVE(&toep->ddp.aiojobq, job, list);
1532         if (!aio_clear_cancel_function(job))
1533                 goto restart;
1534         toep->ddp.queueing = job;
1535
1536         /* NB: This drops DDP_LOCK while it holds the backing VM pages. */
1537         error = hold_aio(toep, job, &ps);
1538         if (error != 0) {
1539                 ddp_complete_one(job, error);
1540                 toep->ddp.queueing = NULL;
1541                 goto restart;
1542         }
1543
1544         SOCKBUF_LOCK(sb);
1545         if (so->so_error && sbavail(sb) == 0) {
1546                 copied = job->aio_received;
1547                 if (copied != 0) {
1548                         SOCKBUF_UNLOCK(sb);
1549                         recycle_pageset(toep, ps);
1550                         aio_complete(job, copied, 0);
1551                         toep->ddp.queueing = NULL;
1552                         goto restart;
1553                 }
1554
1555                 error = so->so_error;
1556                 so->so_error = 0;
1557                 SOCKBUF_UNLOCK(sb);
1558                 recycle_pageset(toep, ps);
1559                 aio_complete(job, -1, error);
1560                 toep->ddp.queueing = NULL;
1561                 goto restart;
1562         }
1563
1564         if (sb->sb_state & SBS_CANTRCVMORE && sbavail(sb) == 0) {
1565                 SOCKBUF_UNLOCK(sb);
1566                 recycle_pageset(toep, ps);
1567                 if (toep->ddp.active_count != 0) {
1568                         /*
1569                          * The door is closed, but there are still pending
1570                          * DDP buffers.  Requeue.  These jobs will all be
1571                          * completed once those buffers drain.
1572                          */
1573                         aio_ddp_requeue_one(toep, job);
1574                         toep->ddp.queueing = NULL;
1575                         return;
1576                 }
1577                 ddp_complete_one(job, 0);
1578                 ddp_complete_all(toep, 0);
1579                 toep->ddp.queueing = NULL;
1580                 return;
1581         }
1582
1583 sbcopy:
1584         /*
1585          * If the toep is dead, there shouldn't be any data in the socket
1586          * buffer, so the above case should have handled this.
1587          */
1588         MPASS(!(toep->ddp.flags & DDP_DEAD));
1589
1590         /*
1591          * If there is pending data in the socket buffer (either
1592          * from before the requests were queued or a DDP indicate),
1593          * copy those mbufs out directly.
1594          */
1595         copied = 0;
1596         offset = ps->offset + job->aio_received;
1597         MPASS(job->aio_received <= job->uaiocb.aio_nbytes);
1598         resid = job->uaiocb.aio_nbytes - job->aio_received;
1599         m = sb->sb_mb;
1600         KASSERT(m == NULL || toep->ddp.active_count == 0,
1601             ("%s: sockbuf data with active DDP", __func__));
1602         while (m != NULL && resid > 0) {
1603                 struct iovec iov[1];
1604                 struct uio uio;
1605                 int error;
1606
1607                 iov[0].iov_base = mtod(m, void *);
1608                 iov[0].iov_len = m->m_len;
1609                 if (iov[0].iov_len > resid)
1610                         iov[0].iov_len = resid;
1611                 uio.uio_iov = iov;
1612                 uio.uio_iovcnt = 1;
1613                 uio.uio_offset = 0;
1614                 uio.uio_resid = iov[0].iov_len;
1615                 uio.uio_segflg = UIO_SYSSPACE;
1616                 uio.uio_rw = UIO_WRITE;
1617                 error = uiomove_fromphys(ps->pages, offset + copied,
1618                     uio.uio_resid, &uio);
1619                 MPASS(error == 0 && uio.uio_resid == 0);
1620                 copied += uio.uio_offset;
1621                 resid -= uio.uio_offset;
1622                 m = m->m_next;
1623         }
1624         if (copied != 0) {
1625                 sbdrop_locked(sb, copied);
1626                 job->aio_received += copied;
1627                 job->msgrcv = 1;
1628                 copied = job->aio_received;
1629                 inp = sotoinpcb(so);
1630                 if (!INP_TRY_WLOCK(inp)) {
1631                         /*
1632                          * The reference on the socket file descriptor in
1633                          * the AIO job should keep 'sb' and 'inp' stable.
1634                          * Our caller has a reference on the 'toep' that
1635                          * keeps it stable.
1636                          */
1637                         SOCKBUF_UNLOCK(sb);
1638                         DDP_UNLOCK(toep);
1639                         INP_WLOCK(inp);
1640                         DDP_LOCK(toep);
1641                         SOCKBUF_LOCK(sb);
1642
1643                         /*
1644                          * If the socket has been closed, we should detect
1645                          * that and complete this request if needed on
1646                          * the next trip around the loop.
1647                          */
1648                 }
1649                 t4_rcvd_locked(&toep->td->tod, intotcpcb(inp));
1650                 INP_WUNLOCK(inp);
1651                 if (resid == 0 || toep->ddp.flags & DDP_DEAD) {
1652                         /*
1653                          * We filled the entire buffer with socket
1654                          * data, DDP is not being used, or the socket
1655                          * is being shut down, so complete the
1656                          * request.
1657                          */
1658                         SOCKBUF_UNLOCK(sb);
1659                         recycle_pageset(toep, ps);
1660                         aio_complete(job, copied, 0);
1661                         toep->ddp.queueing = NULL;
1662                         goto restart;
1663                 }
1664
1665                 /*
1666                  * If DDP is not enabled, requeue this request and restart.
1667                  * This will either enable DDP or wait for more data to
1668                  * arrive on the socket buffer.
1669                  */
1670                 if ((toep->ddp.flags & (DDP_ON | DDP_SC_REQ)) != DDP_ON) {
1671                         SOCKBUF_UNLOCK(sb);
1672                         recycle_pageset(toep, ps);
1673                         aio_ddp_requeue_one(toep, job);
1674                         toep->ddp.queueing = NULL;
1675                         goto restart;
1676                 }
1677
1678                 /*
1679                  * An indicate might have arrived and been added to
1680                  * the socket buffer while it was unlocked after the
1681                  * copy to lock the INP.  If so, restart the copy.
1682                  */
1683                 if (sbavail(sb) != 0)
1684                         goto sbcopy;
1685         }
1686         SOCKBUF_UNLOCK(sb);
1687
1688         if (prep_pageset(sc, toep, ps) == 0) {
1689                 recycle_pageset(toep, ps);
1690                 aio_ddp_requeue_one(toep, job);
1691                 toep->ddp.queueing = NULL;
1692
1693                 /*
1694                  * XXX: Need to retry this later.  Mostly need a trigger
1695                  * when page pods are freed up.
1696                  */
1697                 printf("%s: prep_pageset failed\n", __func__);
1698                 return;
1699         }
1700
1701         /* Determine which DDP buffer to use. */
1702         if (toep->ddp.db[0].job == NULL) {
1703                 db_idx = 0;
1704         } else {
1705                 MPASS(toep->ddp.db[1].job == NULL);
1706                 db_idx = 1;
1707         }
1708
1709         ddp_flags = 0;
1710         ddp_flags_mask = 0;
1711         if (db_idx == 0) {
1712                 ddp_flags |= V_TF_DDP_BUF0_VALID(1);
1713                 if (so->so_state & SS_NBIO)
1714                         ddp_flags |= V_TF_DDP_BUF0_FLUSH(1);
1715                 ddp_flags_mask |= V_TF_DDP_PSH_NO_INVALIDATE0(1) |
1716                     V_TF_DDP_PUSH_DISABLE_0(1) | V_TF_DDP_PSHF_ENABLE_0(1) |
1717                     V_TF_DDP_BUF0_FLUSH(1) | V_TF_DDP_BUF0_VALID(1);
1718                 buf_flag = DDP_BUF0_ACTIVE;
1719         } else {
1720                 ddp_flags |= V_TF_DDP_BUF1_VALID(1);
1721                 if (so->so_state & SS_NBIO)
1722                         ddp_flags |= V_TF_DDP_BUF1_FLUSH(1);
1723                 ddp_flags_mask |= V_TF_DDP_PSH_NO_INVALIDATE1(1) |
1724                     V_TF_DDP_PUSH_DISABLE_1(1) | V_TF_DDP_PSHF_ENABLE_1(1) |
1725                     V_TF_DDP_BUF1_FLUSH(1) | V_TF_DDP_BUF1_VALID(1);
1726                 buf_flag = DDP_BUF1_ACTIVE;
1727         }
1728         MPASS((toep->ddp.flags & buf_flag) == 0);
1729         if ((toep->ddp.flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE)) == 0) {
1730                 MPASS(db_idx == 0);
1731                 MPASS(toep->ddp.active_id == -1);
1732                 MPASS(toep->ddp.active_count == 0);
1733                 ddp_flags_mask |= V_TF_DDP_ACTIVE_BUF(1);
1734         }
1735
1736         /*
1737          * The TID for this connection should still be valid.  If DDP_DEAD
1738          * is set, SBS_CANTRCVMORE should be set, so we shouldn't be
1739          * this far anyway.  Even if the socket is closing on the other
1740          * end, the AIO job holds a reference on this end of the socket
1741          * which will keep it open and keep the TCP PCB attached until
1742          * after the job is completed.
1743          */
1744         wr = mk_update_tcb_for_ddp(sc, toep, db_idx, ps, job->aio_received,
1745             ddp_flags, ddp_flags_mask);
1746         if (wr == NULL) {
1747                 recycle_pageset(toep, ps);
1748                 aio_ddp_requeue_one(toep, job);
1749                 toep->ddp.queueing = NULL;
1750
1751                 /*
1752                  * XXX: Need a way to kick a retry here.
1753                  *
1754                  * XXX: We know the fixed size needed and could
1755                  * preallocate this using a blocking request at the
1756                  * start of the task to avoid having to handle this
1757                  * edge case.
1758                  */
1759                 printf("%s: mk_update_tcb_for_ddp failed\n", __func__);
1760                 return;
1761         }
1762
1763         if (!aio_set_cancel_function(job, t4_aio_cancel_active)) {
1764                 free_wrqe(wr);
1765                 recycle_pageset(toep, ps);
1766                 aio_ddp_cancel_one(job);
1767                 toep->ddp.queueing = NULL;
1768                 goto restart;
1769         }
1770
1771 #ifdef VERBOSE_TRACES
1772         CTR6(KTR_CXGBE,
1773             "%s: tid %u, scheduling %p for DDP[%d] (flags %#lx/%#lx)", __func__,
1774             toep->tid, job, db_idx, ddp_flags, ddp_flags_mask);
1775 #endif
1776         /* Give the chip the go-ahead. */
1777         t4_wrq_tx(sc, wr);
1778         db = &toep->ddp.db[db_idx];
1779         db->cancel_pending = 0;
1780         db->job = job;
1781         db->ps = ps;
1782         toep->ddp.queueing = NULL;
1783         toep->ddp.flags |= buf_flag;
1784         toep->ddp.active_count++;
1785         if (toep->ddp.active_count == 1) {
1786                 MPASS(toep->ddp.active_id == -1);
1787                 toep->ddp.active_id = db_idx;
1788                 CTR2(KTR_CXGBE, "%s: ddp_active_id = %d", __func__,
1789                     toep->ddp.active_id);
1790         }
1791         goto restart;
1792 }
1793
1794 void
1795 ddp_queue_toep(struct toepcb *toep)
1796 {
1797
1798         DDP_ASSERT_LOCKED(toep);
1799         if (toep->ddp.flags & DDP_TASK_ACTIVE)
1800                 return;
1801         toep->ddp.flags |= DDP_TASK_ACTIVE;
1802         hold_toepcb(toep);
1803         soaio_enqueue(&toep->ddp.requeue_task);
1804 }
1805
1806 static void
1807 aio_ddp_requeue_task(void *context, int pending)
1808 {
1809         struct toepcb *toep = context;
1810
1811         DDP_LOCK(toep);
1812         aio_ddp_requeue(toep);
1813         toep->ddp.flags &= ~DDP_TASK_ACTIVE;
1814         DDP_UNLOCK(toep);
1815
1816         free_toepcb(toep);
1817 }
1818
1819 static void
1820 t4_aio_cancel_active(struct kaiocb *job)
1821 {
1822         struct socket *so = job->fd_file->f_data;
1823         struct tcpcb *tp = so_sototcpcb(so);
1824         struct toepcb *toep = tp->t_toe;
1825         struct adapter *sc = td_adapter(toep->td);
1826         uint64_t valid_flag;
1827         int i;
1828
1829         DDP_LOCK(toep);
1830         if (aio_cancel_cleared(job)) {
1831                 DDP_UNLOCK(toep);
1832                 aio_ddp_cancel_one(job);
1833                 return;
1834         }
1835
1836         for (i = 0; i < nitems(toep->ddp.db); i++) {
1837                 if (toep->ddp.db[i].job == job) {
1838                         /* Should only ever get one cancel request for a job. */
1839                         MPASS(toep->ddp.db[i].cancel_pending == 0);
1840
1841                         /*
1842                          * Invalidate this buffer.  It will be
1843                          * cancelled or partially completed once the
1844                          * card ACKs the invalidate.
1845                          */
1846                         valid_flag = i == 0 ? V_TF_DDP_BUF0_VALID(1) :
1847                             V_TF_DDP_BUF1_VALID(1);
1848                         t4_set_tcb_field(sc, toep->ctrlq, toep,
1849                             W_TCB_RX_DDP_FLAGS, valid_flag, 0, 1,
1850                             i + DDP_BUF0_INVALIDATED);
1851                         toep->ddp.db[i].cancel_pending = 1;
1852                         CTR2(KTR_CXGBE, "%s: request %p marked pending",
1853                             __func__, job);
1854                         break;
1855                 }
1856         }
1857         DDP_UNLOCK(toep);
1858 }
1859
1860 static void
1861 t4_aio_cancel_queued(struct kaiocb *job)
1862 {
1863         struct socket *so = job->fd_file->f_data;
1864         struct tcpcb *tp = so_sototcpcb(so);
1865         struct toepcb *toep = tp->t_toe;
1866
1867         DDP_LOCK(toep);
1868         if (!aio_cancel_cleared(job)) {
1869                 TAILQ_REMOVE(&toep->ddp.aiojobq, job, list);
1870                 toep->ddp.waiting_count--;
1871                 if (toep->ddp.waiting_count == 0)
1872                         ddp_queue_toep(toep);
1873         }
1874         CTR2(KTR_CXGBE, "%s: request %p cancelled", __func__, job);
1875         DDP_UNLOCK(toep);
1876
1877         aio_ddp_cancel_one(job);
1878 }
1879
1880 int
1881 t4_aio_queue_ddp(struct socket *so, struct kaiocb *job)
1882 {
1883         struct tcpcb *tp = so_sototcpcb(so);
1884         struct toepcb *toep = tp->t_toe;
1885
1886
1887         /* Ignore writes. */
1888         if (job->uaiocb.aio_lio_opcode != LIO_READ)
1889                 return (EOPNOTSUPP);
1890
1891         DDP_LOCK(toep);
1892
1893         /*
1894          * XXX: Think about possibly returning errors for ENOTCONN,
1895          * etc.  Perhaps the caller would only queue the request
1896          * if it failed with EOPNOTSUPP?
1897          */
1898
1899 #ifdef VERBOSE_TRACES
1900         CTR3(KTR_CXGBE, "%s: queueing %p for tid %u", __func__, job, toep->tid);
1901 #endif
1902         if (!aio_set_cancel_function(job, t4_aio_cancel_queued))
1903                 panic("new job was cancelled");
1904         TAILQ_INSERT_TAIL(&toep->ddp.aiojobq, job, list);
1905         toep->ddp.waiting_count++;
1906         toep->ddp.flags |= DDP_OK;
1907
1908         /*
1909          * Try to handle this request synchronously.  If this has
1910          * to block because the task is running, it will just bail
1911          * and let the task handle it instead.
1912          */
1913         aio_ddp_requeue(toep);
1914         DDP_UNLOCK(toep);
1915         return (0);
1916 }
1917
1918 void
1919 t4_ddp_mod_load(void)
1920 {
1921
1922         t4_register_shared_cpl_handler(CPL_SET_TCB_RPL, do_ddp_tcb_rpl,
1923             CPL_COOKIE_DDP0);
1924         t4_register_shared_cpl_handler(CPL_SET_TCB_RPL, do_ddp_tcb_rpl,
1925             CPL_COOKIE_DDP1);
1926         t4_register_cpl_handler(CPL_RX_DATA_DDP, do_rx_data_ddp);
1927         t4_register_cpl_handler(CPL_RX_DDP_COMPLETE, do_rx_ddp_complete);
1928         TAILQ_INIT(&ddp_orphan_pagesets);
1929         mtx_init(&ddp_orphan_pagesets_lock, "ddp orphans", NULL, MTX_DEF);
1930         TASK_INIT(&ddp_orphan_task, 0, ddp_free_orphan_pagesets, NULL);
1931 }
1932
1933 void
1934 t4_ddp_mod_unload(void)
1935 {
1936
1937         taskqueue_drain(taskqueue_thread, &ddp_orphan_task);
1938         MPASS(TAILQ_EMPTY(&ddp_orphan_pagesets));
1939         mtx_destroy(&ddp_orphan_pagesets_lock);
1940         t4_register_shared_cpl_handler(CPL_SET_TCB_RPL, NULL, CPL_COOKIE_DDP0);
1941         t4_register_shared_cpl_handler(CPL_SET_TCB_RPL, NULL, CPL_COOKIE_DDP1);
1942         t4_register_cpl_handler(CPL_RX_DATA_DDP, NULL);
1943         t4_register_cpl_handler(CPL_RX_DDP_COMPLETE, NULL);
1944 }
1945 #endif