xref: /freebsd/sys/dev/cxgbe/tom/t4_ddp.c (revision ec965063070e5753c166cf592c9336444b74720a)
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/bio.h>
38 #include <sys/file.h>
39 #include <sys/systm.h>
40 #include <sys/kernel.h>
41 #include <sys/ktr.h>
42 #include <sys/module.h>
43 #include <sys/protosw.h>
44 #include <sys/proc.h>
45 #include <sys/domain.h>
46 #include <sys/socket.h>
47 #include <sys/socketvar.h>
48 #include <sys/taskqueue.h>
49 #include <sys/uio.h>
50 #include <netinet/in.h>
51 #include <netinet/in_pcb.h>
52 #include <netinet/ip.h>
53 #include <netinet/tcp_var.h>
54 #define TCPSTATES
55 #include <netinet/tcp_fsm.h>
56 #include <netinet/toecore.h>
57 
58 #include <vm/vm.h>
59 #include <vm/vm_extern.h>
60 #include <vm/vm_param.h>
61 #include <vm/pmap.h>
62 #include <vm/vm_map.h>
63 #include <vm/vm_page.h>
64 #include <vm/vm_object.h>
65 
66 #include <cam/scsi/scsi_all.h>
67 #include <cam/ctl/ctl_io.h>
68 
69 #ifdef TCP_OFFLOAD
70 #include "common/common.h"
71 #include "common/t4_msg.h"
72 #include "common/t4_regs.h"
73 #include "common/t4_tcb.h"
74 #include "tom/t4_tom.h"
75 
76 /*
77  * Use the 'backend3' field in AIO jobs to store the amount of data
78  * received by the AIO job so far.
79  */
80 #define	aio_received	backend3
81 
82 static void aio_ddp_requeue_task(void *context, int pending);
83 static void ddp_complete_all(struct toepcb *toep, int error);
84 static void t4_aio_cancel_active(struct kaiocb *job);
85 static void t4_aio_cancel_queued(struct kaiocb *job);
86 
87 static TAILQ_HEAD(, pageset) ddp_orphan_pagesets;
88 static struct mtx ddp_orphan_pagesets_lock;
89 static struct task ddp_orphan_task;
90 
91 #define MAX_DDP_BUFFER_SIZE		(M_TCB_RX_DDP_BUF0_LEN)
92 
93 /*
94  * A page set holds information about a buffer used for DDP.  The page
95  * set holds resources such as the VM pages backing the buffer (either
96  * held or wired) and the page pods associated with the buffer.
97  * Recently used page sets are cached to allow for efficient reuse of
98  * buffers (avoiding the need to re-fault in pages, hold them, etc.).
99  * Note that cached page sets keep the backing pages wired.  The
100  * number of wired pages is capped by only allowing for two wired
101  * pagesets per connection.  This is not a perfect cap, but is a
102  * trade-off for performance.
103  *
104  * If an application ping-pongs two buffers for a connection via
105  * aio_read(2) then those buffers should remain wired and expensive VM
106  * fault lookups should be avoided after each buffer has been used
107  * once.  If an application uses more than two buffers then this will
108  * fall back to doing expensive VM fault lookups for each operation.
109  */
110 static void
111 free_pageset(struct tom_data *td, struct pageset *ps)
112 {
113 	vm_page_t p;
114 	int i;
115 
116 	if (ps->prsv.prsv_nppods > 0)
117 		t4_free_page_pods(&ps->prsv);
118 
119 	for (i = 0; i < ps->npages; i++) {
120 		p = ps->pages[i];
121 		vm_page_unwire(p, PQ_INACTIVE);
122 	}
123 	mtx_lock(&ddp_orphan_pagesets_lock);
124 	TAILQ_INSERT_TAIL(&ddp_orphan_pagesets, ps, link);
125 	taskqueue_enqueue(taskqueue_thread, &ddp_orphan_task);
126 	mtx_unlock(&ddp_orphan_pagesets_lock);
127 }
128 
129 static void
130 ddp_free_orphan_pagesets(void *context, int pending)
131 {
132 	struct pageset *ps;
133 
134 	mtx_lock(&ddp_orphan_pagesets_lock);
135 	while (!TAILQ_EMPTY(&ddp_orphan_pagesets)) {
136 		ps = TAILQ_FIRST(&ddp_orphan_pagesets);
137 		TAILQ_REMOVE(&ddp_orphan_pagesets, ps, link);
138 		mtx_unlock(&ddp_orphan_pagesets_lock);
139 		if (ps->vm)
140 			vmspace_free(ps->vm);
141 		free(ps, M_CXGBE);
142 		mtx_lock(&ddp_orphan_pagesets_lock);
143 	}
144 	mtx_unlock(&ddp_orphan_pagesets_lock);
145 }
146 
147 static void
148 recycle_pageset(struct toepcb *toep, struct pageset *ps)
149 {
150 
151 	DDP_ASSERT_LOCKED(toep);
152 	if (!(toep->ddp.flags & DDP_DEAD)) {
153 		KASSERT(toep->ddp.cached_count + toep->ddp.active_count <
154 		    nitems(toep->ddp.db), ("too many wired pagesets"));
155 		TAILQ_INSERT_HEAD(&toep->ddp.cached_pagesets, ps, link);
156 		toep->ddp.cached_count++;
157 	} else
158 		free_pageset(toep->td, ps);
159 }
160 
161 static void
162 ddp_complete_one(struct kaiocb *job, int error)
163 {
164 	long copied;
165 
166 	/*
167 	 * If this job had copied data out of the socket buffer before
168 	 * it was cancelled, report it as a short read rather than an
169 	 * error.
170 	 */
171 	copied = job->aio_received;
172 	if (copied != 0 || error == 0)
173 		aio_complete(job, copied, 0);
174 	else
175 		aio_complete(job, -1, error);
176 }
177 
178 static void
179 free_ddp_buffer(struct tom_data *td, struct ddp_buffer *db)
180 {
181 
182 	if (db->job) {
183 		/*
184 		 * XXX: If we are un-offloading the socket then we
185 		 * should requeue these on the socket somehow.  If we
186 		 * got a FIN from the remote end, then this completes
187 		 * any remaining requests with an EOF read.
188 		 */
189 		if (!aio_clear_cancel_function(db->job))
190 			ddp_complete_one(db->job, 0);
191 	}
192 
193 	if (db->ps)
194 		free_pageset(td, db->ps);
195 }
196 
197 void
198 ddp_init_toep(struct toepcb *toep)
199 {
200 
201 	TAILQ_INIT(&toep->ddp.aiojobq);
202 	TASK_INIT(&toep->ddp.requeue_task, 0, aio_ddp_requeue_task, toep);
203 	toep->ddp.flags = DDP_OK;
204 	toep->ddp.active_id = -1;
205 	mtx_init(&toep->ddp.lock, "t4 ddp", NULL, MTX_DEF);
206 }
207 
208 void
209 ddp_uninit_toep(struct toepcb *toep)
210 {
211 
212 	mtx_destroy(&toep->ddp.lock);
213 }
214 
215 void
216 release_ddp_resources(struct toepcb *toep)
217 {
218 	struct pageset *ps;
219 	int i;
220 
221 	DDP_LOCK(toep);
222 	toep->ddp.flags |= DDP_DEAD;
223 	for (i = 0; i < nitems(toep->ddp.db); i++) {
224 		free_ddp_buffer(toep->td, &toep->ddp.db[i]);
225 	}
226 	while ((ps = TAILQ_FIRST(&toep->ddp.cached_pagesets)) != NULL) {
227 		TAILQ_REMOVE(&toep->ddp.cached_pagesets, ps, link);
228 		free_pageset(toep->td, ps);
229 	}
230 	ddp_complete_all(toep, 0);
231 	DDP_UNLOCK(toep);
232 }
233 
234 #ifdef INVARIANTS
235 void
236 ddp_assert_empty(struct toepcb *toep)
237 {
238 	int i;
239 
240 	MPASS(!(toep->ddp.flags & DDP_TASK_ACTIVE));
241 	for (i = 0; i < nitems(toep->ddp.db); i++) {
242 		MPASS(toep->ddp.db[i].job == NULL);
243 		MPASS(toep->ddp.db[i].ps == NULL);
244 	}
245 	MPASS(TAILQ_EMPTY(&toep->ddp.cached_pagesets));
246 	MPASS(TAILQ_EMPTY(&toep->ddp.aiojobq));
247 }
248 #endif
249 
250 static void
251 complete_ddp_buffer(struct toepcb *toep, struct ddp_buffer *db,
252     unsigned int db_idx)
253 {
254 	unsigned int db_flag;
255 
256 	toep->ddp.active_count--;
257 	if (toep->ddp.active_id == db_idx) {
258 		if (toep->ddp.active_count == 0) {
259 			KASSERT(toep->ddp.db[db_idx ^ 1].job == NULL,
260 			    ("%s: active_count mismatch", __func__));
261 			toep->ddp.active_id = -1;
262 		} else
263 			toep->ddp.active_id ^= 1;
264 #ifdef VERBOSE_TRACES
265 		CTR3(KTR_CXGBE, "%s: tid %u, ddp_active_id = %d", __func__,
266 		    toep->tid, toep->ddp.active_id);
267 #endif
268 	} else {
269 		KASSERT(toep->ddp.active_count != 0 &&
270 		    toep->ddp.active_id != -1,
271 		    ("%s: active count mismatch", __func__));
272 	}
273 
274 	db->cancel_pending = 0;
275 	db->job = NULL;
276 	recycle_pageset(toep, db->ps);
277 	db->ps = NULL;
278 
279 	db_flag = db_idx == 1 ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE;
280 	KASSERT(toep->ddp.flags & db_flag,
281 	    ("%s: DDP buffer not active. toep %p, ddp_flags 0x%x",
282 	    __func__, toep, toep->ddp.flags));
283 	toep->ddp.flags &= ~db_flag;
284 }
285 
286 /* XXX: handle_ddp_data code duplication */
287 void
288 insert_ddp_data(struct toepcb *toep, uint32_t n)
289 {
290 	struct inpcb *inp = toep->inp;
291 	struct tcpcb *tp = intotcpcb(inp);
292 	struct ddp_buffer *db;
293 	struct kaiocb *job;
294 	size_t placed;
295 	long copied;
296 	unsigned int db_idx;
297 #ifdef INVARIANTS
298 	unsigned int db_flag;
299 #endif
300 
301 	INP_WLOCK_ASSERT(inp);
302 	DDP_ASSERT_LOCKED(toep);
303 
304 	tp->rcv_nxt += n;
305 #ifndef USE_DDP_RX_FLOW_CONTROL
306 	KASSERT(tp->rcv_wnd >= n, ("%s: negative window size", __func__));
307 	tp->rcv_wnd -= n;
308 #endif
309 	CTR2(KTR_CXGBE, "%s: placed %u bytes before falling out of DDP",
310 	    __func__, n);
311 	while (toep->ddp.active_count > 0) {
312 		MPASS(toep->ddp.active_id != -1);
313 		db_idx = toep->ddp.active_id;
314 #ifdef INVARIANTS
315 		db_flag = db_idx == 1 ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE;
316 #endif
317 		MPASS((toep->ddp.flags & db_flag) != 0);
318 		db = &toep->ddp.db[db_idx];
319 		job = db->job;
320 		copied = job->aio_received;
321 		placed = n;
322 		if (placed > job->uaiocb.aio_nbytes - copied)
323 			placed = job->uaiocb.aio_nbytes - copied;
324 		if (placed > 0)
325 			job->msgrcv = 1;
326 		if (!aio_clear_cancel_function(job)) {
327 			/*
328 			 * Update the copied length for when
329 			 * t4_aio_cancel_active() completes this
330 			 * request.
331 			 */
332 			job->aio_received += placed;
333 		} else if (copied + placed != 0) {
334 			CTR4(KTR_CXGBE,
335 			    "%s: completing %p (copied %ld, placed %lu)",
336 			    __func__, job, copied, placed);
337 			/* XXX: This always completes if there is some data. */
338 			aio_complete(job, copied + placed, 0);
339 		} else if (aio_set_cancel_function(job, t4_aio_cancel_queued)) {
340 			TAILQ_INSERT_HEAD(&toep->ddp.aiojobq, job, list);
341 			toep->ddp.waiting_count++;
342 		} else
343 			aio_cancel(job);
344 		n -= placed;
345 		complete_ddp_buffer(toep, db, db_idx);
346 	}
347 
348 	MPASS(n == 0);
349 }
350 
351 /* SET_TCB_FIELD sent as a ULP command looks like this */
352 #define LEN__SET_TCB_FIELD_ULP (sizeof(struct ulp_txpkt) + \
353     sizeof(struct ulptx_idata) + sizeof(struct cpl_set_tcb_field_core))
354 
355 /* RX_DATA_ACK sent as a ULP command looks like this */
356 #define LEN__RX_DATA_ACK_ULP (sizeof(struct ulp_txpkt) + \
357     sizeof(struct ulptx_idata) + sizeof(struct cpl_rx_data_ack_core))
358 
359 static inline void *
360 mk_set_tcb_field_ulp(struct ulp_txpkt *ulpmc, struct toepcb *toep,
361     uint64_t word, uint64_t mask, uint64_t val)
362 {
363 	struct ulptx_idata *ulpsc;
364 	struct cpl_set_tcb_field_core *req;
365 
366 	ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0));
367 	ulpmc->len = htobe32(howmany(LEN__SET_TCB_FIELD_ULP, 16));
368 
369 	ulpsc = (struct ulptx_idata *)(ulpmc + 1);
370 	ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
371 	ulpsc->len = htobe32(sizeof(*req));
372 
373 	req = (struct cpl_set_tcb_field_core *)(ulpsc + 1);
374 	OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_SET_TCB_FIELD, toep->tid));
375 	req->reply_ctrl = htobe16(V_NO_REPLY(1) |
376 	    V_QUEUENO(toep->ofld_rxq->iq.abs_id));
377 	req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0));
378         req->mask = htobe64(mask);
379         req->val = htobe64(val);
380 
381 	ulpsc = (struct ulptx_idata *)(req + 1);
382 	if (LEN__SET_TCB_FIELD_ULP % 16) {
383 		ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP));
384 		ulpsc->len = htobe32(0);
385 		return (ulpsc + 1);
386 	}
387 	return (ulpsc);
388 }
389 
390 static inline void *
391 mk_rx_data_ack_ulp(struct ulp_txpkt *ulpmc, struct toepcb *toep)
392 {
393 	struct ulptx_idata *ulpsc;
394 	struct cpl_rx_data_ack_core *req;
395 
396 	ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0));
397 	ulpmc->len = htobe32(howmany(LEN__RX_DATA_ACK_ULP, 16));
398 
399 	ulpsc = (struct ulptx_idata *)(ulpmc + 1);
400 	ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
401 	ulpsc->len = htobe32(sizeof(*req));
402 
403 	req = (struct cpl_rx_data_ack_core *)(ulpsc + 1);
404 	OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_RX_DATA_ACK, toep->tid));
405 	req->credit_dack = htobe32(F_RX_MODULATE_RX);
406 
407 	ulpsc = (struct ulptx_idata *)(req + 1);
408 	if (LEN__RX_DATA_ACK_ULP % 16) {
409 		ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP));
410 		ulpsc->len = htobe32(0);
411 		return (ulpsc + 1);
412 	}
413 	return (ulpsc);
414 }
415 
416 static struct wrqe *
417 mk_update_tcb_for_ddp(struct adapter *sc, struct toepcb *toep, int db_idx,
418     struct pageset *ps, int offset, uint64_t ddp_flags, uint64_t ddp_flags_mask)
419 {
420 	struct wrqe *wr;
421 	struct work_request_hdr *wrh;
422 	struct ulp_txpkt *ulpmc;
423 	int len;
424 
425 	KASSERT(db_idx == 0 || db_idx == 1,
426 	    ("%s: bad DDP buffer index %d", __func__, db_idx));
427 
428 	/*
429 	 * We'll send a compound work request that has 3 SET_TCB_FIELDs and an
430 	 * RX_DATA_ACK (with RX_MODULATE to speed up delivery).
431 	 *
432 	 * The work request header is 16B and always ends at a 16B boundary.
433 	 * The ULPTX master commands that follow must all end at 16B boundaries
434 	 * too so we round up the size to 16.
435 	 */
436 	len = sizeof(*wrh) + 3 * roundup2(LEN__SET_TCB_FIELD_ULP, 16) +
437 	    roundup2(LEN__RX_DATA_ACK_ULP, 16);
438 
439 	wr = alloc_wrqe(len, toep->ctrlq);
440 	if (wr == NULL)
441 		return (NULL);
442 	wrh = wrtod(wr);
443 	INIT_ULPTX_WRH(wrh, len, 1, 0);	/* atomic */
444 	ulpmc = (struct ulp_txpkt *)(wrh + 1);
445 
446 	/* Write the buffer's tag */
447 	ulpmc = mk_set_tcb_field_ulp(ulpmc, toep,
448 	    W_TCB_RX_DDP_BUF0_TAG + db_idx,
449 	    V_TCB_RX_DDP_BUF0_TAG(M_TCB_RX_DDP_BUF0_TAG),
450 	    V_TCB_RX_DDP_BUF0_TAG(ps->prsv.prsv_tag));
451 
452 	/* Update the current offset in the DDP buffer and its total length */
453 	if (db_idx == 0)
454 		ulpmc = mk_set_tcb_field_ulp(ulpmc, toep,
455 		    W_TCB_RX_DDP_BUF0_OFFSET,
456 		    V_TCB_RX_DDP_BUF0_OFFSET(M_TCB_RX_DDP_BUF0_OFFSET) |
457 		    V_TCB_RX_DDP_BUF0_LEN(M_TCB_RX_DDP_BUF0_LEN),
458 		    V_TCB_RX_DDP_BUF0_OFFSET(offset) |
459 		    V_TCB_RX_DDP_BUF0_LEN(ps->len));
460 	else
461 		ulpmc = mk_set_tcb_field_ulp(ulpmc, toep,
462 		    W_TCB_RX_DDP_BUF1_OFFSET,
463 		    V_TCB_RX_DDP_BUF1_OFFSET(M_TCB_RX_DDP_BUF1_OFFSET) |
464 		    V_TCB_RX_DDP_BUF1_LEN((u64)M_TCB_RX_DDP_BUF1_LEN << 32),
465 		    V_TCB_RX_DDP_BUF1_OFFSET(offset) |
466 		    V_TCB_RX_DDP_BUF1_LEN((u64)ps->len << 32));
467 
468 	/* Update DDP flags */
469 	ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, W_TCB_RX_DDP_FLAGS,
470 	    ddp_flags_mask, ddp_flags);
471 
472 	/* Gratuitous RX_DATA_ACK with RX_MODULATE set to speed up delivery. */
473 	ulpmc = mk_rx_data_ack_ulp(ulpmc, toep);
474 
475 	return (wr);
476 }
477 
478 static int
479 handle_ddp_data(struct toepcb *toep, __be32 ddp_report, __be32 rcv_nxt, int len)
480 {
481 	uint32_t report = be32toh(ddp_report);
482 	unsigned int db_idx;
483 	struct inpcb *inp = toep->inp;
484 	struct ddp_buffer *db;
485 	struct tcpcb *tp;
486 	struct socket *so;
487 	struct sockbuf *sb;
488 	struct kaiocb *job;
489 	long copied;
490 
491 	db_idx = report & F_DDP_BUF_IDX ? 1 : 0;
492 
493 	if (__predict_false(!(report & F_DDP_INV)))
494 		CXGBE_UNIMPLEMENTED("DDP buffer still valid");
495 
496 	INP_WLOCK(inp);
497 	so = inp_inpcbtosocket(inp);
498 	sb = &so->so_rcv;
499 	DDP_LOCK(toep);
500 
501 	KASSERT(toep->ddp.active_id == db_idx,
502 	    ("completed DDP buffer (%d) != active_id (%d) for tid %d", db_idx,
503 	    toep->ddp.active_id, toep->tid));
504 	db = &toep->ddp.db[db_idx];
505 	job = db->job;
506 
507 	if (__predict_false(inp->inp_flags & INP_DROPPED)) {
508 		/*
509 		 * This can happen due to an administrative tcpdrop(8).
510 		 * Just fail the request with ECONNRESET.
511 		 */
512 		CTR5(KTR_CXGBE, "%s: tid %u, seq 0x%x, len %d, inp_flags 0x%x",
513 		    __func__, toep->tid, be32toh(rcv_nxt), len, inp->inp_flags);
514 		if (aio_clear_cancel_function(job))
515 			ddp_complete_one(job, ECONNRESET);
516 		goto completed;
517 	}
518 
519 	tp = intotcpcb(inp);
520 
521 	/*
522 	 * For RX_DDP_COMPLETE, len will be zero and rcv_nxt is the
523 	 * sequence number of the next byte to receive.  The length of
524 	 * the data received for this message must be computed by
525 	 * comparing the new and old values of rcv_nxt.
526 	 *
527 	 * For RX_DATA_DDP, len might be non-zero, but it is only the
528 	 * length of the most recent DMA.  It does not include the
529 	 * total length of the data received since the previous update
530 	 * for this DDP buffer.  rcv_nxt is the sequence number of the
531 	 * first received byte from the most recent DMA.
532 	 */
533 	len += be32toh(rcv_nxt) - tp->rcv_nxt;
534 	tp->rcv_nxt += len;
535 	tp->t_rcvtime = ticks;
536 #ifndef USE_DDP_RX_FLOW_CONTROL
537 	KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__));
538 	tp->rcv_wnd -= len;
539 #endif
540 #ifdef VERBOSE_TRACES
541 	CTR5(KTR_CXGBE, "%s: tid %u, DDP[%d] placed %d bytes (%#x)", __func__,
542 	    toep->tid, db_idx, len, report);
543 #endif
544 
545 	/* receive buffer autosize */
546 	MPASS(toep->vnet == so->so_vnet);
547 	CURVNET_SET(toep->vnet);
548 	SOCKBUF_LOCK(sb);
549 	if (sb->sb_flags & SB_AUTOSIZE &&
550 	    V_tcp_do_autorcvbuf &&
551 	    sb->sb_hiwat < V_tcp_autorcvbuf_max &&
552 	    len > (sbspace(sb) / 8 * 7)) {
553 		struct adapter *sc = td_adapter(toep->td);
554 		unsigned int hiwat = sb->sb_hiwat;
555 		unsigned int newsize = min(hiwat + sc->tt.autorcvbuf_inc,
556 		    V_tcp_autorcvbuf_max);
557 
558 		if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
559 			sb->sb_flags &= ~SB_AUTOSIZE;
560 	}
561 	SOCKBUF_UNLOCK(sb);
562 	CURVNET_RESTORE();
563 
564 	job->msgrcv = 1;
565 	if (db->cancel_pending) {
566 		/*
567 		 * Update the job's length but defer completion to the
568 		 * TCB_RPL callback.
569 		 */
570 		job->aio_received += len;
571 		goto out;
572 	} else if (!aio_clear_cancel_function(job)) {
573 		/*
574 		 * Update the copied length for when
575 		 * t4_aio_cancel_active() completes this request.
576 		 */
577 		job->aio_received += len;
578 	} else {
579 		copied = job->aio_received;
580 #ifdef VERBOSE_TRACES
581 		CTR5(KTR_CXGBE,
582 		    "%s: tid %u, completing %p (copied %ld, placed %d)",
583 		    __func__, toep->tid, job, copied, len);
584 #endif
585 		aio_complete(job, copied + len, 0);
586 		t4_rcvd(&toep->td->tod, tp);
587 	}
588 
589 completed:
590 	complete_ddp_buffer(toep, db, db_idx);
591 	if (toep->ddp.waiting_count > 0)
592 		ddp_queue_toep(toep);
593 out:
594 	DDP_UNLOCK(toep);
595 	INP_WUNLOCK(inp);
596 
597 	return (0);
598 }
599 
600 void
601 handle_ddp_indicate(struct toepcb *toep)
602 {
603 
604 	DDP_ASSERT_LOCKED(toep);
605 	MPASS(toep->ddp.active_count == 0);
606 	MPASS((toep->ddp.flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE)) == 0);
607 	if (toep->ddp.waiting_count == 0) {
608 		/*
609 		 * The pending requests that triggered the request for an
610 		 * an indicate were cancelled.  Those cancels should have
611 		 * already disabled DDP.  Just ignore this as the data is
612 		 * going into the socket buffer anyway.
613 		 */
614 		return;
615 	}
616 	CTR3(KTR_CXGBE, "%s: tid %d indicated (%d waiting)", __func__,
617 	    toep->tid, toep->ddp.waiting_count);
618 	ddp_queue_toep(toep);
619 }
620 
621 CTASSERT(CPL_COOKIE_DDP0 + 1 == CPL_COOKIE_DDP1);
622 
623 static int
624 do_ddp_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
625 {
626 	struct adapter *sc = iq->adapter;
627 	const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
628 	unsigned int tid = GET_TID(cpl);
629 	unsigned int db_idx;
630 	struct toepcb *toep;
631 	struct inpcb *inp;
632 	struct ddp_buffer *db;
633 	struct kaiocb *job;
634 	long copied;
635 
636 	if (cpl->status != CPL_ERR_NONE)
637 		panic("XXX: tcp_rpl failed: %d", cpl->status);
638 
639 	toep = lookup_tid(sc, tid);
640 	inp = toep->inp;
641 	switch (cpl->cookie) {
642 	case V_WORD(W_TCB_RX_DDP_FLAGS) | V_COOKIE(CPL_COOKIE_DDP0):
643 	case V_WORD(W_TCB_RX_DDP_FLAGS) | V_COOKIE(CPL_COOKIE_DDP1):
644 		/*
645 		 * XXX: This duplicates a lot of code with handle_ddp_data().
646 		 */
647 		db_idx = G_COOKIE(cpl->cookie) - CPL_COOKIE_DDP0;
648 		MPASS(db_idx < nitems(toep->ddp.db));
649 		INP_WLOCK(inp);
650 		DDP_LOCK(toep);
651 		db = &toep->ddp.db[db_idx];
652 
653 		/*
654 		 * handle_ddp_data() should leave the job around until
655 		 * this callback runs once a cancel is pending.
656 		 */
657 		MPASS(db != NULL);
658 		MPASS(db->job != NULL);
659 		MPASS(db->cancel_pending);
660 
661 		/*
662 		 * XXX: It's not clear what happens if there is data
663 		 * placed when the buffer is invalidated.  I suspect we
664 		 * need to read the TCB to see how much data was placed.
665 		 *
666 		 * For now this just pretends like nothing was placed.
667 		 *
668 		 * XXX: Note that if we did check the PCB we would need to
669 		 * also take care of updating the tp, etc.
670 		 */
671 		job = db->job;
672 		copied = job->aio_received;
673 		if (copied == 0) {
674 			CTR2(KTR_CXGBE, "%s: cancelling %p", __func__, job);
675 			aio_cancel(job);
676 		} else {
677 			CTR3(KTR_CXGBE, "%s: completing %p (copied %ld)",
678 			    __func__, job, copied);
679 			aio_complete(job, copied, 0);
680 			t4_rcvd(&toep->td->tod, intotcpcb(inp));
681 		}
682 
683 		complete_ddp_buffer(toep, db, db_idx);
684 		if (toep->ddp.waiting_count > 0)
685 			ddp_queue_toep(toep);
686 		DDP_UNLOCK(toep);
687 		INP_WUNLOCK(inp);
688 		break;
689 	default:
690 		panic("XXX: unknown tcb_rpl offset %#x, cookie %#x",
691 		    G_WORD(cpl->cookie), G_COOKIE(cpl->cookie));
692 	}
693 
694 	return (0);
695 }
696 
697 void
698 handle_ddp_close(struct toepcb *toep, struct tcpcb *tp, __be32 rcv_nxt)
699 {
700 	struct ddp_buffer *db;
701 	struct kaiocb *job;
702 	long copied;
703 	unsigned int db_idx;
704 #ifdef INVARIANTS
705 	unsigned int db_flag;
706 #endif
707 	int len, placed;
708 
709 	INP_WLOCK_ASSERT(toep->inp);
710 	DDP_ASSERT_LOCKED(toep);
711 
712 	/* - 1 is to ignore the byte for FIN */
713 	len = be32toh(rcv_nxt) - tp->rcv_nxt - 1;
714 	tp->rcv_nxt += len;
715 
716 	while (toep->ddp.active_count > 0) {
717 		MPASS(toep->ddp.active_id != -1);
718 		db_idx = toep->ddp.active_id;
719 #ifdef INVARIANTS
720 		db_flag = db_idx == 1 ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE;
721 #endif
722 		MPASS((toep->ddp.flags & db_flag) != 0);
723 		db = &toep->ddp.db[db_idx];
724 		job = db->job;
725 		copied = job->aio_received;
726 		placed = len;
727 		if (placed > job->uaiocb.aio_nbytes - copied)
728 			placed = job->uaiocb.aio_nbytes - copied;
729 		if (placed > 0)
730 			job->msgrcv = 1;
731 		if (!aio_clear_cancel_function(job)) {
732 			/*
733 			 * Update the copied length for when
734 			 * t4_aio_cancel_active() completes this
735 			 * request.
736 			 */
737 			job->aio_received += placed;
738 		} else {
739 			CTR4(KTR_CXGBE, "%s: tid %d completed buf %d len %d",
740 			    __func__, toep->tid, db_idx, placed);
741 			aio_complete(job, copied + placed, 0);
742 		}
743 		len -= placed;
744 		complete_ddp_buffer(toep, db, db_idx);
745 	}
746 
747 	MPASS(len == 0);
748 	ddp_complete_all(toep, 0);
749 }
750 
751 #define DDP_ERR (F_DDP_PPOD_MISMATCH | F_DDP_LLIMIT_ERR | F_DDP_ULIMIT_ERR |\
752 	 F_DDP_PPOD_PARITY_ERR | F_DDP_PADDING_ERR | F_DDP_OFFSET_ERR |\
753 	 F_DDP_INVALID_TAG | F_DDP_COLOR_ERR | F_DDP_TID_MISMATCH |\
754 	 F_DDP_INVALID_PPOD | F_DDP_HDRCRC_ERR | F_DDP_DATACRC_ERR)
755 
756 extern cpl_handler_t t4_cpl_handler[];
757 
758 static int
759 do_rx_data_ddp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
760 {
761 	struct adapter *sc = iq->adapter;
762 	const struct cpl_rx_data_ddp *cpl = (const void *)(rss + 1);
763 	unsigned int tid = GET_TID(cpl);
764 	uint32_t vld;
765 	struct toepcb *toep = lookup_tid(sc, tid);
766 
767 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
768 	KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__));
769 	KASSERT(!(toep->flags & TPF_SYNQE),
770 	    ("%s: toep %p claims to be a synq entry", __func__, toep));
771 
772 	vld = be32toh(cpl->ddpvld);
773 	if (__predict_false(vld & DDP_ERR)) {
774 		panic("%s: DDP error 0x%x (tid %d, toep %p)",
775 		    __func__, vld, tid, toep);
776 	}
777 
778 	if (ulp_mode(toep) == ULP_MODE_ISCSI) {
779 		t4_cpl_handler[CPL_RX_ISCSI_DDP](iq, rss, m);
780 		return (0);
781 	}
782 
783 	handle_ddp_data(toep, cpl->u.ddp_report, cpl->seq, be16toh(cpl->len));
784 
785 	return (0);
786 }
787 
788 static int
789 do_rx_ddp_complete(struct sge_iq *iq, const struct rss_header *rss,
790     struct mbuf *m)
791 {
792 	struct adapter *sc = iq->adapter;
793 	const struct cpl_rx_ddp_complete *cpl = (const void *)(rss + 1);
794 	unsigned int tid = GET_TID(cpl);
795 	struct toepcb *toep = lookup_tid(sc, tid);
796 
797 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
798 	KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__));
799 	KASSERT(!(toep->flags & TPF_SYNQE),
800 	    ("%s: toep %p claims to be a synq entry", __func__, toep));
801 
802 	handle_ddp_data(toep, cpl->ddp_report, cpl->rcv_nxt, 0);
803 
804 	return (0);
805 }
806 
807 static void
808 enable_ddp(struct adapter *sc, struct toepcb *toep)
809 {
810 
811 	KASSERT((toep->ddp.flags & (DDP_ON | DDP_OK | DDP_SC_REQ)) == DDP_OK,
812 	    ("%s: toep %p has bad ddp_flags 0x%x",
813 	    __func__, toep, toep->ddp.flags));
814 
815 	CTR3(KTR_CXGBE, "%s: tid %u (time %u)",
816 	    __func__, toep->tid, time_uptime);
817 
818 	DDP_ASSERT_LOCKED(toep);
819 	toep->ddp.flags |= DDP_SC_REQ;
820 	t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_RX_DDP_FLAGS,
821 	    V_TF_DDP_OFF(1) | V_TF_DDP_INDICATE_OUT(1) |
822 	    V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1) |
823 	    V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_BUF1_VALID(1),
824 	    V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1), 0, 0);
825 	t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_T_FLAGS,
826 	    V_TF_RCV_COALESCE_ENABLE(1), 0, 0, 0);
827 }
828 
829 static int
830 calculate_hcf(int n1, int n2)
831 {
832 	int a, b, t;
833 
834 	if (n1 <= n2) {
835 		a = n1;
836 		b = n2;
837 	} else {
838 		a = n2;
839 		b = n1;
840 	}
841 
842 	while (a != 0) {
843 		t = a;
844 		a = b % a;
845 		b = t;
846 	}
847 
848 	return (b);
849 }
850 
851 static inline int
852 pages_to_nppods(int npages, int ddp_page_shift)
853 {
854 
855 	MPASS(ddp_page_shift >= PAGE_SHIFT);
856 
857 	return (howmany(npages >> (ddp_page_shift - PAGE_SHIFT), PPOD_PAGES));
858 }
859 
860 static int
861 alloc_page_pods(struct ppod_region *pr, u_int nppods, u_int pgsz_idx,
862     struct ppod_reservation *prsv)
863 {
864 	vmem_addr_t addr;       /* relative to start of region */
865 
866 	if (vmem_alloc(pr->pr_arena, PPOD_SZ(nppods), M_NOWAIT | M_FIRSTFIT,
867 	    &addr) != 0)
868 		return (ENOMEM);
869 
870 #ifdef VERBOSE_TRACES
871 	CTR5(KTR_CXGBE, "%-17s arena %p, addr 0x%08x, nppods %d, pgsz %d",
872 	    __func__, pr->pr_arena, (uint32_t)addr & pr->pr_tag_mask,
873 	    nppods, 1 << pr->pr_page_shift[pgsz_idx]);
874 #endif
875 
876 	/*
877 	 * The hardware tagmask includes an extra invalid bit but the arena was
878 	 * seeded with valid values only.  An allocation out of this arena will
879 	 * fit inside the tagmask but won't have the invalid bit set.
880 	 */
881 	MPASS((addr & pr->pr_tag_mask) == addr);
882 	MPASS((addr & pr->pr_invalid_bit) == 0);
883 
884 	prsv->prsv_pr = pr;
885 	prsv->prsv_tag = V_PPOD_PGSZ(pgsz_idx) | addr;
886 	prsv->prsv_nppods = nppods;
887 
888 	return (0);
889 }
890 
891 static int
892 t4_alloc_page_pods_for_vmpages(struct ppod_region *pr, vm_page_t *pages,
893     int npages, struct ppod_reservation *prsv)
894 {
895 	int i, hcf, seglen, idx, nppods;
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 < npages; i++) {
906 		seglen = PAGE_SIZE;
907 		while (i < npages - 1 &&
908 		    VM_PAGE_TO_PHYS(pages[i]) + PAGE_SIZE ==
909 		    VM_PAGE_TO_PHYS(pages[i + 1])) {
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(npages, pr->pr_page_shift[idx]);
933 	if (alloc_page_pods(pr, nppods, idx, prsv) != 0)
934 		return (ENOMEM);
935 	MPASS(prsv->prsv_nppods > 0);
936 
937 	return (0);
938 }
939 
940 int
941 t4_alloc_page_pods_for_ps(struct ppod_region *pr, struct pageset *ps)
942 {
943 	struct ppod_reservation *prsv = &ps->prsv;
944 
945 	KASSERT(prsv->prsv_nppods == 0,
946 	    ("%s: page pods already allocated", __func__));
947 
948 	return (t4_alloc_page_pods_for_vmpages(pr, ps->pages, ps->npages,
949 	    prsv));
950 }
951 
952 int
953 t4_alloc_page_pods_for_bio(struct ppod_region *pr, struct bio *bp,
954     struct ppod_reservation *prsv)
955 {
956 
957 	MPASS(bp->bio_flags & BIO_UNMAPPED);
958 
959 	return (t4_alloc_page_pods_for_vmpages(pr, bp->bio_ma, bp->bio_ma_n,
960 	    prsv));
961 }
962 
963 int
964 t4_alloc_page_pods_for_buf(struct ppod_region *pr, vm_offset_t buf, int len,
965     struct ppod_reservation *prsv)
966 {
967 	int hcf, seglen, idx, npages, nppods;
968 	uintptr_t start_pva, end_pva, pva, p1;
969 
970 	MPASS(buf > 0);
971 	MPASS(len > 0);
972 
973 	/*
974 	 * The DDP page size is unrelated to the VM page size.  We combine
975 	 * contiguous physical pages into larger segments to get the best DDP
976 	 * page size possible.  This is the largest of the four sizes in
977 	 * A_ULP_RX_ISCSI_PSZ that evenly divides the HCF of the segment sizes
978 	 * in the page list.
979 	 */
980 	hcf = 0;
981 	start_pva = trunc_page(buf);
982 	end_pva = trunc_page(buf + len - 1);
983 	pva = start_pva;
984 	while (pva <= end_pva) {
985 		seglen = PAGE_SIZE;
986 		p1 = pmap_kextract(pva);
987 		pva += PAGE_SIZE;
988 		while (pva <= end_pva && p1 + seglen == pmap_kextract(pva)) {
989 			seglen += PAGE_SIZE;
990 			pva += PAGE_SIZE;
991 		}
992 
993 		hcf = calculate_hcf(hcf, seglen);
994 		if (hcf < (1 << pr->pr_page_shift[1])) {
995 			idx = 0;
996 			goto have_pgsz;	/* give up, short circuit */
997 		}
998 	}
999 
1000 #define PR_PAGE_MASK(x) ((1 << pr->pr_page_shift[(x)]) - 1)
1001 	MPASS((hcf & PR_PAGE_MASK(0)) == 0); /* PAGE_SIZE is >= 4K everywhere */
1002 	for (idx = nitems(pr->pr_page_shift) - 1; idx > 0; idx--) {
1003 		if ((hcf & PR_PAGE_MASK(idx)) == 0)
1004 			break;
1005 	}
1006 #undef PR_PAGE_MASK
1007 
1008 have_pgsz:
1009 	MPASS(idx <= M_PPOD_PGSZ);
1010 
1011 	npages = 1;
1012 	npages += (end_pva - start_pva) >> pr->pr_page_shift[idx];
1013 	nppods = howmany(npages, PPOD_PAGES);
1014 	if (alloc_page_pods(pr, nppods, idx, prsv) != 0)
1015 		return (ENOMEM);
1016 	MPASS(prsv->prsv_nppods > 0);
1017 
1018 	return (0);
1019 }
1020 
1021 int
1022 t4_alloc_page_pods_for_sgl(struct ppod_region *pr, struct ctl_sg_entry *sgl,
1023     int entries, struct ppod_reservation *prsv)
1024 {
1025 	int hcf, seglen, idx = 0, npages, nppods, i, len;
1026 	uintptr_t start_pva, end_pva, pva, p1 ;
1027 	vm_offset_t buf;
1028 	struct ctl_sg_entry *sge;
1029 
1030 	MPASS(entries > 0);
1031 	MPASS(sgl);
1032 
1033 	/*
1034 	 * The DDP page size is unrelated to the VM page size.	We combine
1035 	 * contiguous physical pages into larger segments to get the best DDP
1036 	 * page size possible.	This is the largest of the four sizes in
1037 	 * A_ULP_RX_ISCSI_PSZ that evenly divides the HCF of the segment sizes
1038 	 * in the page list.
1039 	 */
1040 	hcf = 0;
1041 	for (i = entries - 1; i >= 0; i--) {
1042 		sge = sgl + i;
1043 		buf = (vm_offset_t)sge->addr;
1044 		len = sge->len;
1045 		start_pva = trunc_page(buf);
1046 		end_pva = trunc_page(buf + len - 1);
1047 		pva = start_pva;
1048 		while (pva <= end_pva) {
1049 			seglen = PAGE_SIZE;
1050 			p1 = pmap_kextract(pva);
1051 			pva += PAGE_SIZE;
1052 			while (pva <= end_pva && p1 + seglen ==
1053 			    pmap_kextract(pva)) {
1054 				seglen += PAGE_SIZE;
1055 				pva += PAGE_SIZE;
1056 			}
1057 
1058 			hcf = calculate_hcf(hcf, seglen);
1059 			if (hcf < (1 << pr->pr_page_shift[1])) {
1060 				idx = 0;
1061 				goto have_pgsz; /* give up, short circuit */
1062 			}
1063 		}
1064 	}
1065 #define PR_PAGE_MASK(x) ((1 << pr->pr_page_shift[(x)]) - 1)
1066 	MPASS((hcf & PR_PAGE_MASK(0)) == 0); /* PAGE_SIZE is >= 4K everywhere */
1067 	for (idx = nitems(pr->pr_page_shift) - 1; idx > 0; idx--) {
1068 		if ((hcf & PR_PAGE_MASK(idx)) == 0)
1069 			break;
1070 	}
1071 #undef PR_PAGE_MASK
1072 
1073 have_pgsz:
1074 	MPASS(idx <= M_PPOD_PGSZ);
1075 
1076 	npages = 0;
1077 	while (entries--) {
1078 		npages++;
1079 		start_pva = trunc_page((vm_offset_t)sgl->addr);
1080 		end_pva = trunc_page((vm_offset_t)sgl->addr + sgl->len - 1);
1081 		npages += (end_pva - start_pva) >> pr->pr_page_shift[idx];
1082 		sgl = sgl + 1;
1083 	}
1084 	nppods = howmany(npages, PPOD_PAGES);
1085 	if (alloc_page_pods(pr, nppods, idx, prsv) != 0)
1086 		return (ENOMEM);
1087 	MPASS(prsv->prsv_nppods > 0);
1088 	return (0);
1089 }
1090 
1091 void
1092 t4_free_page_pods(struct ppod_reservation *prsv)
1093 {
1094 	struct ppod_region *pr = prsv->prsv_pr;
1095 	vmem_addr_t addr;
1096 
1097 	MPASS(prsv != NULL);
1098 	MPASS(prsv->prsv_nppods != 0);
1099 
1100 	addr = prsv->prsv_tag & pr->pr_tag_mask;
1101 	MPASS((addr & pr->pr_invalid_bit) == 0);
1102 
1103 #ifdef VERBOSE_TRACES
1104 	CTR4(KTR_CXGBE, "%-17s arena %p, addr 0x%08x, nppods %d", __func__,
1105 	    pr->pr_arena, addr, prsv->prsv_nppods);
1106 #endif
1107 
1108 	vmem_free(pr->pr_arena, addr, PPOD_SZ(prsv->prsv_nppods));
1109 	prsv->prsv_nppods = 0;
1110 }
1111 
1112 #define NUM_ULP_TX_SC_IMM_PPODS (256 / PPOD_SIZE)
1113 
1114 int
1115 t4_write_page_pods_for_ps(struct adapter *sc, struct sge_wrq *wrq, int tid,
1116     struct pageset *ps)
1117 {
1118 	struct wrqe *wr;
1119 	struct ulp_mem_io *ulpmc;
1120 	struct ulptx_idata *ulpsc;
1121 	struct pagepod *ppod;
1122 	int i, j, k, n, chunk, len, ddp_pgsz, idx;
1123 	u_int ppod_addr;
1124 	uint32_t cmd;
1125 	struct ppod_reservation *prsv = &ps->prsv;
1126 	struct ppod_region *pr = prsv->prsv_pr;
1127 	vm_paddr_t pa;
1128 
1129 	KASSERT(!(ps->flags & PS_PPODS_WRITTEN),
1130 	    ("%s: page pods already written", __func__));
1131 	MPASS(prsv->prsv_nppods > 0);
1132 
1133 	cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE));
1134 	if (is_t4(sc))
1135 		cmd |= htobe32(F_ULP_MEMIO_ORDER);
1136 	else
1137 		cmd |= htobe32(F_T5_ULP_MEMIO_IMM);
1138 	ddp_pgsz = 1 << pr->pr_page_shift[G_PPOD_PGSZ(prsv->prsv_tag)];
1139 	ppod_addr = pr->pr_start + (prsv->prsv_tag & pr->pr_tag_mask);
1140 	for (i = 0; i < prsv->prsv_nppods; ppod_addr += chunk) {
1141 
1142 		/* How many page pods are we writing in this cycle */
1143 		n = min(prsv->prsv_nppods - i, NUM_ULP_TX_SC_IMM_PPODS);
1144 		chunk = PPOD_SZ(n);
1145 		len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16);
1146 
1147 		wr = alloc_wrqe(len, wrq);
1148 		if (wr == NULL)
1149 			return (ENOMEM);	/* ok to just bail out */
1150 		ulpmc = wrtod(wr);
1151 
1152 		INIT_ULPTX_WR(ulpmc, len, 0, 0);
1153 		ulpmc->cmd = cmd;
1154 		ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32));
1155 		ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16));
1156 		ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5));
1157 
1158 		ulpsc = (struct ulptx_idata *)(ulpmc + 1);
1159 		ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
1160 		ulpsc->len = htobe32(chunk);
1161 
1162 		ppod = (struct pagepod *)(ulpsc + 1);
1163 		for (j = 0; j < n; i++, j++, ppod++) {
1164 			ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID |
1165 			    V_PPOD_TID(tid) | prsv->prsv_tag);
1166 			ppod->len_offset = htobe64(V_PPOD_LEN(ps->len) |
1167 			    V_PPOD_OFST(ps->offset));
1168 			ppod->rsvd = 0;
1169 			idx = i * PPOD_PAGES * (ddp_pgsz / PAGE_SIZE);
1170 			for (k = 0; k < nitems(ppod->addr); k++) {
1171 				if (idx < ps->npages) {
1172 					pa = VM_PAGE_TO_PHYS(ps->pages[idx]);
1173 					ppod->addr[k] = htobe64(pa);
1174 					idx += ddp_pgsz / PAGE_SIZE;
1175 				} else
1176 					ppod->addr[k] = 0;
1177 #if 0
1178 				CTR5(KTR_CXGBE,
1179 				    "%s: tid %d ppod[%d]->addr[%d] = %p",
1180 				    __func__, tid, i, k,
1181 				    be64toh(ppod->addr[k]));
1182 #endif
1183 			}
1184 
1185 		}
1186 
1187 		t4_wrq_tx(sc, wr);
1188 	}
1189 	ps->flags |= PS_PPODS_WRITTEN;
1190 
1191 	return (0);
1192 }
1193 
1194 static struct mbuf *
1195 alloc_raw_wr_mbuf(int len)
1196 {
1197 	struct mbuf *m;
1198 
1199 	if (len <= MHLEN)
1200 		m = m_gethdr(M_NOWAIT, MT_DATA);
1201 	else if (len <= MCLBYTES)
1202 		m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1203 	else
1204 		m = NULL;
1205 	if (m == NULL)
1206 		return (NULL);
1207 	m->m_pkthdr.len = len;
1208 	m->m_len = len;
1209 	set_mbuf_raw_wr(m, true);
1210 	return (m);
1211 }
1212 
1213 int
1214 t4_write_page_pods_for_bio(struct adapter *sc, struct toepcb *toep,
1215     struct ppod_reservation *prsv, struct bio *bp, struct mbufq *wrq)
1216 {
1217 	struct ulp_mem_io *ulpmc;
1218 	struct ulptx_idata *ulpsc;
1219 	struct pagepod *ppod;
1220 	int i, j, k, n, chunk, len, ddp_pgsz, idx;
1221 	u_int ppod_addr;
1222 	uint32_t cmd;
1223 	struct ppod_region *pr = prsv->prsv_pr;
1224 	vm_paddr_t pa;
1225 	struct mbuf *m;
1226 
1227 	MPASS(bp->bio_flags & BIO_UNMAPPED);
1228 
1229 	cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE));
1230 	if (is_t4(sc))
1231 		cmd |= htobe32(F_ULP_MEMIO_ORDER);
1232 	else
1233 		cmd |= htobe32(F_T5_ULP_MEMIO_IMM);
1234 	ddp_pgsz = 1 << pr->pr_page_shift[G_PPOD_PGSZ(prsv->prsv_tag)];
1235 	ppod_addr = pr->pr_start + (prsv->prsv_tag & pr->pr_tag_mask);
1236 	for (i = 0; i < prsv->prsv_nppods; ppod_addr += chunk) {
1237 
1238 		/* How many page pods are we writing in this cycle */
1239 		n = min(prsv->prsv_nppods - i, NUM_ULP_TX_SC_IMM_PPODS);
1240 		MPASS(n > 0);
1241 		chunk = PPOD_SZ(n);
1242 		len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16);
1243 
1244 		m = alloc_raw_wr_mbuf(len);
1245 		if (m == NULL)
1246 			return (ENOMEM);
1247 
1248 		ulpmc = mtod(m, struct ulp_mem_io *);
1249 		INIT_ULPTX_WR(ulpmc, len, 0, toep->tid);
1250 		ulpmc->cmd = cmd;
1251 		ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32));
1252 		ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16));
1253 		ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5));
1254 
1255 		ulpsc = (struct ulptx_idata *)(ulpmc + 1);
1256 		ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
1257 		ulpsc->len = htobe32(chunk);
1258 
1259 		ppod = (struct pagepod *)(ulpsc + 1);
1260 		for (j = 0; j < n; i++, j++, ppod++) {
1261 			ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID |
1262 			    V_PPOD_TID(toep->tid) |
1263 			    (prsv->prsv_tag & ~V_PPOD_PGSZ(M_PPOD_PGSZ)));
1264 			ppod->len_offset = htobe64(V_PPOD_LEN(bp->bio_bcount) |
1265 			    V_PPOD_OFST(bp->bio_ma_offset));
1266 			ppod->rsvd = 0;
1267 			idx = i * PPOD_PAGES * (ddp_pgsz / PAGE_SIZE);
1268 			for (k = 0; k < nitems(ppod->addr); k++) {
1269 				if (idx < bp->bio_ma_n) {
1270 					pa = VM_PAGE_TO_PHYS(bp->bio_ma[idx]);
1271 					ppod->addr[k] = htobe64(pa);
1272 					idx += ddp_pgsz / PAGE_SIZE;
1273 				} else
1274 					ppod->addr[k] = 0;
1275 #if 0
1276 				CTR5(KTR_CXGBE,
1277 				    "%s: tid %d ppod[%d]->addr[%d] = %p",
1278 				    __func__, toep->tid, i, k,
1279 				    be64toh(ppod->addr[k]));
1280 #endif
1281 			}
1282 		}
1283 
1284 		mbufq_enqueue(wrq, m);
1285 	}
1286 
1287 	return (0);
1288 }
1289 
1290 int
1291 t4_write_page_pods_for_buf(struct adapter *sc, struct toepcb *toep,
1292     struct ppod_reservation *prsv, vm_offset_t buf, int buflen,
1293     struct mbufq *wrq)
1294 {
1295 	struct ulp_mem_io *ulpmc;
1296 	struct ulptx_idata *ulpsc;
1297 	struct pagepod *ppod;
1298 	int i, j, k, n, chunk, len, ddp_pgsz;
1299 	u_int ppod_addr, offset;
1300 	uint32_t cmd;
1301 	struct ppod_region *pr = prsv->prsv_pr;
1302 	uintptr_t end_pva, pva;
1303 	vm_paddr_t pa;
1304 	struct mbuf *m;
1305 
1306 	cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE));
1307 	if (is_t4(sc))
1308 		cmd |= htobe32(F_ULP_MEMIO_ORDER);
1309 	else
1310 		cmd |= htobe32(F_T5_ULP_MEMIO_IMM);
1311 	ddp_pgsz = 1 << pr->pr_page_shift[G_PPOD_PGSZ(prsv->prsv_tag)];
1312 	offset = buf & PAGE_MASK;
1313 	ppod_addr = pr->pr_start + (prsv->prsv_tag & pr->pr_tag_mask);
1314 	pva = trunc_page(buf);
1315 	end_pva = trunc_page(buf + buflen - 1);
1316 	for (i = 0; i < prsv->prsv_nppods; ppod_addr += chunk) {
1317 
1318 		/* How many page pods are we writing in this cycle */
1319 		n = min(prsv->prsv_nppods - i, NUM_ULP_TX_SC_IMM_PPODS);
1320 		MPASS(n > 0);
1321 		chunk = PPOD_SZ(n);
1322 		len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16);
1323 
1324 		m = alloc_raw_wr_mbuf(len);
1325 		if (m == NULL)
1326 			return (ENOMEM);
1327 		ulpmc = mtod(m, struct ulp_mem_io *);
1328 
1329 		INIT_ULPTX_WR(ulpmc, len, 0, toep->tid);
1330 		ulpmc->cmd = cmd;
1331 		ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32));
1332 		ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16));
1333 		ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5));
1334 
1335 		ulpsc = (struct ulptx_idata *)(ulpmc + 1);
1336 		ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
1337 		ulpsc->len = htobe32(chunk);
1338 
1339 		ppod = (struct pagepod *)(ulpsc + 1);
1340 		for (j = 0; j < n; i++, j++, ppod++) {
1341 			ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID |
1342 			    V_PPOD_TID(toep->tid) |
1343 			    (prsv->prsv_tag & ~V_PPOD_PGSZ(M_PPOD_PGSZ)));
1344 			ppod->len_offset = htobe64(V_PPOD_LEN(buflen) |
1345 			    V_PPOD_OFST(offset));
1346 			ppod->rsvd = 0;
1347 
1348 			for (k = 0; k < nitems(ppod->addr); k++) {
1349 				if (pva > end_pva)
1350 					ppod->addr[k] = 0;
1351 				else {
1352 					pa = pmap_kextract(pva);
1353 					ppod->addr[k] = htobe64(pa);
1354 					pva += ddp_pgsz;
1355 				}
1356 #if 0
1357 				CTR5(KTR_CXGBE,
1358 				    "%s: tid %d ppod[%d]->addr[%d] = %p",
1359 				    __func__, toep->tid, i, k,
1360 				    be64toh(ppod->addr[k]));
1361 #endif
1362 			}
1363 
1364 			/*
1365 			 * Walk back 1 segment so that the first address in the
1366 			 * next pod is the same as the last one in the current
1367 			 * pod.
1368 			 */
1369 			pva -= ddp_pgsz;
1370 		}
1371 
1372 		mbufq_enqueue(wrq, m);
1373 	}
1374 
1375 	MPASS(pva <= end_pva);
1376 
1377 	return (0);
1378 }
1379 
1380 int
1381 t4_write_page_pods_for_sgl(struct adapter *sc, struct toepcb *toep,
1382     struct ppod_reservation *prsv, struct ctl_sg_entry *sgl, int entries,
1383     int xferlen, struct mbufq *wrq)
1384 {
1385 	struct ulp_mem_io *ulpmc;
1386 	struct ulptx_idata *ulpsc;
1387 	struct pagepod *ppod;
1388 	int i, j, k, n, chunk, len, ddp_pgsz;
1389 	u_int ppod_addr, offset, sg_offset = 0;
1390 	uint32_t cmd;
1391 	struct ppod_region *pr = prsv->prsv_pr;
1392 	uintptr_t pva;
1393 	vm_paddr_t pa;
1394 	struct mbuf *m;
1395 
1396 	MPASS(sgl != NULL);
1397 	MPASS(entries > 0);
1398 	cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE));
1399 	if (is_t4(sc))
1400 		cmd |= htobe32(F_ULP_MEMIO_ORDER);
1401 	else
1402 		cmd |= htobe32(F_T5_ULP_MEMIO_IMM);
1403 	ddp_pgsz = 1 << pr->pr_page_shift[G_PPOD_PGSZ(prsv->prsv_tag)];
1404 	offset = (vm_offset_t)sgl->addr & PAGE_MASK;
1405 	ppod_addr = pr->pr_start + (prsv->prsv_tag & pr->pr_tag_mask);
1406 	pva = trunc_page((vm_offset_t)sgl->addr);
1407 	for (i = 0; i < prsv->prsv_nppods; ppod_addr += chunk) {
1408 
1409 		/* How many page pods are we writing in this cycle */
1410 		n = min(prsv->prsv_nppods - i, NUM_ULP_TX_SC_IMM_PPODS);
1411 		MPASS(n > 0);
1412 		chunk = PPOD_SZ(n);
1413 		len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16);
1414 
1415 		m = alloc_raw_wr_mbuf(len);
1416 		if (m == NULL)
1417 			return (ENOMEM);
1418 		ulpmc = mtod(m, struct ulp_mem_io *);
1419 
1420 		INIT_ULPTX_WR(ulpmc, len, 0, toep->tid);
1421 		ulpmc->cmd = cmd;
1422 		ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32));
1423 		ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16));
1424 		ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5));
1425 
1426 		ulpsc = (struct ulptx_idata *)(ulpmc + 1);
1427 		ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
1428 		ulpsc->len = htobe32(chunk);
1429 
1430 		ppod = (struct pagepod *)(ulpsc + 1);
1431 		for (j = 0; j < n; i++, j++, ppod++) {
1432 			ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID |
1433 			    V_PPOD_TID(toep->tid) |
1434 			    (prsv->prsv_tag & ~V_PPOD_PGSZ(M_PPOD_PGSZ)));
1435 			ppod->len_offset = htobe64(V_PPOD_LEN(xferlen) |
1436 			    V_PPOD_OFST(offset));
1437 			ppod->rsvd = 0;
1438 
1439 			for (k = 0; k < nitems(ppod->addr); k++) {
1440 				if (entries != 0) {
1441 					pa = pmap_kextract(pva + sg_offset);
1442 					ppod->addr[k] = htobe64(pa);
1443 				} else
1444 					ppod->addr[k] = 0;
1445 
1446 #if 0
1447 				CTR5(KTR_CXGBE,
1448 				    "%s: tid %d ppod[%d]->addr[%d] = %p",
1449 				    __func__, toep->tid, i, k,
1450 				    be64toh(ppod->addr[k]));
1451 #endif
1452 
1453 				/*
1454 				 * If this is the last entry in a pod,
1455 				 * reuse the same entry for first address
1456 				 * in the next pod.
1457 				 */
1458 				if (k + 1 == nitems(ppod->addr))
1459 					break;
1460 
1461 				/*
1462 				 * Don't move to the next DDP page if the
1463 				 * sgl is already finished.
1464 				 */
1465 				if (entries == 0)
1466 					continue;
1467 
1468 				sg_offset += ddp_pgsz;
1469 				if (sg_offset == sgl->len) {
1470 					/*
1471 					 * This sgl entry is done.  Go
1472 					 * to the next.
1473 					 */
1474 					entries--;
1475 					sgl++;
1476 					sg_offset = 0;
1477 					if (entries != 0)
1478 						pva = trunc_page(
1479 						    (vm_offset_t)sgl->addr);
1480 				}
1481 			}
1482 		}
1483 
1484 		mbufq_enqueue(wrq, m);
1485 	}
1486 
1487 	return (0);
1488 }
1489 
1490 /*
1491  * Prepare a pageset for DDP.  This sets up page pods.
1492  */
1493 static int
1494 prep_pageset(struct adapter *sc, struct toepcb *toep, struct pageset *ps)
1495 {
1496 	struct tom_data *td = sc->tom_softc;
1497 
1498 	if (ps->prsv.prsv_nppods == 0 &&
1499 	    t4_alloc_page_pods_for_ps(&td->pr, ps) != 0) {
1500 		return (0);
1501 	}
1502 	if (!(ps->flags & PS_PPODS_WRITTEN) &&
1503 	    t4_write_page_pods_for_ps(sc, toep->ctrlq, toep->tid, ps) != 0) {
1504 		return (0);
1505 	}
1506 
1507 	return (1);
1508 }
1509 
1510 int
1511 t4_init_ppod_region(struct ppod_region *pr, struct t4_range *r, u_int psz,
1512     const char *name)
1513 {
1514 	int i;
1515 
1516 	MPASS(pr != NULL);
1517 	MPASS(r->size > 0);
1518 
1519 	pr->pr_start = r->start;
1520 	pr->pr_len = r->size;
1521 	pr->pr_page_shift[0] = 12 + G_HPZ0(psz);
1522 	pr->pr_page_shift[1] = 12 + G_HPZ1(psz);
1523 	pr->pr_page_shift[2] = 12 + G_HPZ2(psz);
1524 	pr->pr_page_shift[3] = 12 + G_HPZ3(psz);
1525 
1526 	/* The SGL -> page pod algorithm requires the sizes to be in order. */
1527 	for (i = 1; i < nitems(pr->pr_page_shift); i++) {
1528 		if (pr->pr_page_shift[i] <= pr->pr_page_shift[i - 1])
1529 			return (ENXIO);
1530 	}
1531 
1532 	pr->pr_tag_mask = ((1 << fls(r->size)) - 1) & V_PPOD_TAG(M_PPOD_TAG);
1533 	pr->pr_alias_mask = V_PPOD_TAG(M_PPOD_TAG) & ~pr->pr_tag_mask;
1534 	if (pr->pr_tag_mask == 0 || pr->pr_alias_mask == 0)
1535 		return (ENXIO);
1536 	pr->pr_alias_shift = fls(pr->pr_tag_mask);
1537 	pr->pr_invalid_bit = 1 << (pr->pr_alias_shift - 1);
1538 
1539 	pr->pr_arena = vmem_create(name, 0, pr->pr_len, PPOD_SIZE, 0,
1540 	    M_FIRSTFIT | M_NOWAIT);
1541 	if (pr->pr_arena == NULL)
1542 		return (ENOMEM);
1543 
1544 	return (0);
1545 }
1546 
1547 void
1548 t4_free_ppod_region(struct ppod_region *pr)
1549 {
1550 
1551 	MPASS(pr != NULL);
1552 
1553 	if (pr->pr_arena)
1554 		vmem_destroy(pr->pr_arena);
1555 	bzero(pr, sizeof(*pr));
1556 }
1557 
1558 static int
1559 pscmp(struct pageset *ps, struct vmspace *vm, vm_offset_t start, int npages,
1560     int pgoff, int len)
1561 {
1562 
1563 	if (ps->start != start || ps->npages != npages ||
1564 	    ps->offset != pgoff || ps->len != len)
1565 		return (1);
1566 
1567 	return (ps->vm != vm || ps->vm_timestamp != vm->vm_map.timestamp);
1568 }
1569 
1570 static int
1571 hold_aio(struct toepcb *toep, struct kaiocb *job, struct pageset **pps)
1572 {
1573 	struct vmspace *vm;
1574 	vm_map_t map;
1575 	vm_offset_t start, end, pgoff;
1576 	struct pageset *ps;
1577 	int n;
1578 
1579 	DDP_ASSERT_LOCKED(toep);
1580 
1581 	/*
1582 	 * The AIO subsystem will cancel and drain all requests before
1583 	 * permitting a process to exit or exec, so p_vmspace should
1584 	 * be stable here.
1585 	 */
1586 	vm = job->userproc->p_vmspace;
1587 	map = &vm->vm_map;
1588 	start = (uintptr_t)job->uaiocb.aio_buf;
1589 	pgoff = start & PAGE_MASK;
1590 	end = round_page(start + job->uaiocb.aio_nbytes);
1591 	start = trunc_page(start);
1592 
1593 	if (end - start > MAX_DDP_BUFFER_SIZE) {
1594 		/*
1595 		 * Truncate the request to a short read.
1596 		 * Alternatively, we could DDP in chunks to the larger
1597 		 * buffer, but that would be quite a bit more work.
1598 		 *
1599 		 * When truncating, round the request down to avoid
1600 		 * crossing a cache line on the final transaction.
1601 		 */
1602 		end = rounddown2(start + MAX_DDP_BUFFER_SIZE, CACHE_LINE_SIZE);
1603 #ifdef VERBOSE_TRACES
1604 		CTR4(KTR_CXGBE, "%s: tid %d, truncating size from %lu to %lu",
1605 		    __func__, toep->tid, (unsigned long)job->uaiocb.aio_nbytes,
1606 		    (unsigned long)(end - (start + pgoff)));
1607 		job->uaiocb.aio_nbytes = end - (start + pgoff);
1608 #endif
1609 		end = round_page(end);
1610 	}
1611 
1612 	n = atop(end - start);
1613 
1614 	/*
1615 	 * Try to reuse a cached pageset.
1616 	 */
1617 	TAILQ_FOREACH(ps, &toep->ddp.cached_pagesets, link) {
1618 		if (pscmp(ps, vm, start, n, pgoff,
1619 		    job->uaiocb.aio_nbytes) == 0) {
1620 			TAILQ_REMOVE(&toep->ddp.cached_pagesets, ps, link);
1621 			toep->ddp.cached_count--;
1622 			*pps = ps;
1623 			return (0);
1624 		}
1625 	}
1626 
1627 	/*
1628 	 * If there are too many cached pagesets to create a new one,
1629 	 * free a pageset before creating a new one.
1630 	 */
1631 	KASSERT(toep->ddp.active_count + toep->ddp.cached_count <=
1632 	    nitems(toep->ddp.db), ("%s: too many wired pagesets", __func__));
1633 	if (toep->ddp.active_count + toep->ddp.cached_count ==
1634 	    nitems(toep->ddp.db)) {
1635 		KASSERT(toep->ddp.cached_count > 0,
1636 		    ("no cached pageset to free"));
1637 		ps = TAILQ_LAST(&toep->ddp.cached_pagesets, pagesetq);
1638 		TAILQ_REMOVE(&toep->ddp.cached_pagesets, ps, link);
1639 		toep->ddp.cached_count--;
1640 		free_pageset(toep->td, ps);
1641 	}
1642 	DDP_UNLOCK(toep);
1643 
1644 	/* Create a new pageset. */
1645 	ps = malloc(sizeof(*ps) + n * sizeof(vm_page_t), M_CXGBE, M_WAITOK |
1646 	    M_ZERO);
1647 	ps->pages = (vm_page_t *)(ps + 1);
1648 	ps->vm_timestamp = map->timestamp;
1649 	ps->npages = vm_fault_quick_hold_pages(map, start, end - start,
1650 	    VM_PROT_WRITE, ps->pages, n);
1651 
1652 	DDP_LOCK(toep);
1653 	if (ps->npages < 0) {
1654 		free(ps, M_CXGBE);
1655 		return (EFAULT);
1656 	}
1657 
1658 	KASSERT(ps->npages == n, ("hold_aio: page count mismatch: %d vs %d",
1659 	    ps->npages, n));
1660 
1661 	ps->offset = pgoff;
1662 	ps->len = job->uaiocb.aio_nbytes;
1663 	refcount_acquire(&vm->vm_refcnt);
1664 	ps->vm = vm;
1665 	ps->start = start;
1666 
1667 	CTR5(KTR_CXGBE, "%s: tid %d, new pageset %p for job %p, npages %d",
1668 	    __func__, toep->tid, ps, job, ps->npages);
1669 	*pps = ps;
1670 	return (0);
1671 }
1672 
1673 static void
1674 ddp_complete_all(struct toepcb *toep, int error)
1675 {
1676 	struct kaiocb *job;
1677 
1678 	DDP_ASSERT_LOCKED(toep);
1679 	while (!TAILQ_EMPTY(&toep->ddp.aiojobq)) {
1680 		job = TAILQ_FIRST(&toep->ddp.aiojobq);
1681 		TAILQ_REMOVE(&toep->ddp.aiojobq, job, list);
1682 		toep->ddp.waiting_count--;
1683 		if (aio_clear_cancel_function(job))
1684 			ddp_complete_one(job, error);
1685 	}
1686 }
1687 
1688 static void
1689 aio_ddp_cancel_one(struct kaiocb *job)
1690 {
1691 	long copied;
1692 
1693 	/*
1694 	 * If this job had copied data out of the socket buffer before
1695 	 * it was cancelled, report it as a short read rather than an
1696 	 * error.
1697 	 */
1698 	copied = job->aio_received;
1699 	if (copied != 0)
1700 		aio_complete(job, copied, 0);
1701 	else
1702 		aio_cancel(job);
1703 }
1704 
1705 /*
1706  * Called when the main loop wants to requeue a job to retry it later.
1707  * Deals with the race of the job being cancelled while it was being
1708  * examined.
1709  */
1710 static void
1711 aio_ddp_requeue_one(struct toepcb *toep, struct kaiocb *job)
1712 {
1713 
1714 	DDP_ASSERT_LOCKED(toep);
1715 	if (!(toep->ddp.flags & DDP_DEAD) &&
1716 	    aio_set_cancel_function(job, t4_aio_cancel_queued)) {
1717 		TAILQ_INSERT_HEAD(&toep->ddp.aiojobq, job, list);
1718 		toep->ddp.waiting_count++;
1719 	} else
1720 		aio_ddp_cancel_one(job);
1721 }
1722 
1723 static void
1724 aio_ddp_requeue(struct toepcb *toep)
1725 {
1726 	struct adapter *sc = td_adapter(toep->td);
1727 	struct socket *so;
1728 	struct sockbuf *sb;
1729 	struct inpcb *inp;
1730 	struct kaiocb *job;
1731 	struct ddp_buffer *db;
1732 	size_t copied, offset, resid;
1733 	struct pageset *ps;
1734 	struct mbuf *m;
1735 	uint64_t ddp_flags, ddp_flags_mask;
1736 	struct wrqe *wr;
1737 	int buf_flag, db_idx, error;
1738 
1739 	DDP_ASSERT_LOCKED(toep);
1740 
1741 restart:
1742 	if (toep->ddp.flags & DDP_DEAD) {
1743 		MPASS(toep->ddp.waiting_count == 0);
1744 		MPASS(toep->ddp.active_count == 0);
1745 		return;
1746 	}
1747 
1748 	if (toep->ddp.waiting_count == 0 ||
1749 	    toep->ddp.active_count == nitems(toep->ddp.db)) {
1750 		return;
1751 	}
1752 
1753 	job = TAILQ_FIRST(&toep->ddp.aiojobq);
1754 	so = job->fd_file->f_data;
1755 	sb = &so->so_rcv;
1756 	SOCKBUF_LOCK(sb);
1757 
1758 	/* We will never get anything unless we are or were connected. */
1759 	if (!(so->so_state & (SS_ISCONNECTED|SS_ISDISCONNECTED))) {
1760 		SOCKBUF_UNLOCK(sb);
1761 		ddp_complete_all(toep, ENOTCONN);
1762 		return;
1763 	}
1764 
1765 	KASSERT(toep->ddp.active_count == 0 || sbavail(sb) == 0,
1766 	    ("%s: pending sockbuf data and DDP is active", __func__));
1767 
1768 	/* Abort if socket has reported problems. */
1769 	/* XXX: Wait for any queued DDP's to finish and/or flush them? */
1770 	if (so->so_error && sbavail(sb) == 0) {
1771 		toep->ddp.waiting_count--;
1772 		TAILQ_REMOVE(&toep->ddp.aiojobq, job, list);
1773 		if (!aio_clear_cancel_function(job)) {
1774 			SOCKBUF_UNLOCK(sb);
1775 			goto restart;
1776 		}
1777 
1778 		/*
1779 		 * If this job has previously copied some data, report
1780 		 * a short read and leave the error to be reported by
1781 		 * a future request.
1782 		 */
1783 		copied = job->aio_received;
1784 		if (copied != 0) {
1785 			SOCKBUF_UNLOCK(sb);
1786 			aio_complete(job, copied, 0);
1787 			goto restart;
1788 		}
1789 		error = so->so_error;
1790 		so->so_error = 0;
1791 		SOCKBUF_UNLOCK(sb);
1792 		aio_complete(job, -1, error);
1793 		goto restart;
1794 	}
1795 
1796 	/*
1797 	 * Door is closed.  If there is pending data in the socket buffer,
1798 	 * deliver it.  If there are pending DDP requests, wait for those
1799 	 * to complete.  Once they have completed, return EOF reads.
1800 	 */
1801 	if (sb->sb_state & SBS_CANTRCVMORE && sbavail(sb) == 0) {
1802 		SOCKBUF_UNLOCK(sb);
1803 		if (toep->ddp.active_count != 0)
1804 			return;
1805 		ddp_complete_all(toep, 0);
1806 		return;
1807 	}
1808 
1809 	/*
1810 	 * If DDP is not enabled and there is no pending socket buffer
1811 	 * data, try to enable DDP.
1812 	 */
1813 	if (sbavail(sb) == 0 && (toep->ddp.flags & DDP_ON) == 0) {
1814 		SOCKBUF_UNLOCK(sb);
1815 
1816 		/*
1817 		 * Wait for the card to ACK that DDP is enabled before
1818 		 * queueing any buffers.  Currently this waits for an
1819 		 * indicate to arrive.  This could use a TCB_SET_FIELD_RPL
1820 		 * message to know that DDP was enabled instead of waiting
1821 		 * for the indicate which would avoid copying the indicate
1822 		 * if no data is pending.
1823 		 *
1824 		 * XXX: Might want to limit the indicate size to the size
1825 		 * of the first queued request.
1826 		 */
1827 		if ((toep->ddp.flags & DDP_SC_REQ) == 0)
1828 			enable_ddp(sc, toep);
1829 		return;
1830 	}
1831 	SOCKBUF_UNLOCK(sb);
1832 
1833 	/*
1834 	 * If another thread is queueing a buffer for DDP, let it
1835 	 * drain any work and return.
1836 	 */
1837 	if (toep->ddp.queueing != NULL)
1838 		return;
1839 
1840 	/* Take the next job to prep it for DDP. */
1841 	toep->ddp.waiting_count--;
1842 	TAILQ_REMOVE(&toep->ddp.aiojobq, job, list);
1843 	if (!aio_clear_cancel_function(job))
1844 		goto restart;
1845 	toep->ddp.queueing = job;
1846 
1847 	/* NB: This drops DDP_LOCK while it holds the backing VM pages. */
1848 	error = hold_aio(toep, job, &ps);
1849 	if (error != 0) {
1850 		ddp_complete_one(job, error);
1851 		toep->ddp.queueing = NULL;
1852 		goto restart;
1853 	}
1854 
1855 	SOCKBUF_LOCK(sb);
1856 	if (so->so_error && sbavail(sb) == 0) {
1857 		copied = job->aio_received;
1858 		if (copied != 0) {
1859 			SOCKBUF_UNLOCK(sb);
1860 			recycle_pageset(toep, ps);
1861 			aio_complete(job, copied, 0);
1862 			toep->ddp.queueing = NULL;
1863 			goto restart;
1864 		}
1865 
1866 		error = so->so_error;
1867 		so->so_error = 0;
1868 		SOCKBUF_UNLOCK(sb);
1869 		recycle_pageset(toep, ps);
1870 		aio_complete(job, -1, error);
1871 		toep->ddp.queueing = NULL;
1872 		goto restart;
1873 	}
1874 
1875 	if (sb->sb_state & SBS_CANTRCVMORE && sbavail(sb) == 0) {
1876 		SOCKBUF_UNLOCK(sb);
1877 		recycle_pageset(toep, ps);
1878 		if (toep->ddp.active_count != 0) {
1879 			/*
1880 			 * The door is closed, but there are still pending
1881 			 * DDP buffers.  Requeue.  These jobs will all be
1882 			 * completed once those buffers drain.
1883 			 */
1884 			aio_ddp_requeue_one(toep, job);
1885 			toep->ddp.queueing = NULL;
1886 			return;
1887 		}
1888 		ddp_complete_one(job, 0);
1889 		ddp_complete_all(toep, 0);
1890 		toep->ddp.queueing = NULL;
1891 		return;
1892 	}
1893 
1894 sbcopy:
1895 	/*
1896 	 * If the toep is dead, there shouldn't be any data in the socket
1897 	 * buffer, so the above case should have handled this.
1898 	 */
1899 	MPASS(!(toep->ddp.flags & DDP_DEAD));
1900 
1901 	/*
1902 	 * If there is pending data in the socket buffer (either
1903 	 * from before the requests were queued or a DDP indicate),
1904 	 * copy those mbufs out directly.
1905 	 */
1906 	copied = 0;
1907 	offset = ps->offset + job->aio_received;
1908 	MPASS(job->aio_received <= job->uaiocb.aio_nbytes);
1909 	resid = job->uaiocb.aio_nbytes - job->aio_received;
1910 	m = sb->sb_mb;
1911 	KASSERT(m == NULL || toep->ddp.active_count == 0,
1912 	    ("%s: sockbuf data with active DDP", __func__));
1913 	while (m != NULL && resid > 0) {
1914 		struct iovec iov[1];
1915 		struct uio uio;
1916 #ifdef INVARIANTS
1917 		int error;
1918 #endif
1919 
1920 		iov[0].iov_base = mtod(m, void *);
1921 		iov[0].iov_len = m->m_len;
1922 		if (iov[0].iov_len > resid)
1923 			iov[0].iov_len = resid;
1924 		uio.uio_iov = iov;
1925 		uio.uio_iovcnt = 1;
1926 		uio.uio_offset = 0;
1927 		uio.uio_resid = iov[0].iov_len;
1928 		uio.uio_segflg = UIO_SYSSPACE;
1929 		uio.uio_rw = UIO_WRITE;
1930 #ifdef INVARIANTS
1931 		error = uiomove_fromphys(ps->pages, offset + copied,
1932 		    uio.uio_resid, &uio);
1933 #else
1934 		uiomove_fromphys(ps->pages, offset + copied, uio.uio_resid, &uio);
1935 #endif
1936 		MPASS(error == 0 && uio.uio_resid == 0);
1937 		copied += uio.uio_offset;
1938 		resid -= uio.uio_offset;
1939 		m = m->m_next;
1940 	}
1941 	if (copied != 0) {
1942 		sbdrop_locked(sb, copied);
1943 		job->aio_received += copied;
1944 		job->msgrcv = 1;
1945 		copied = job->aio_received;
1946 		inp = sotoinpcb(so);
1947 		if (!INP_TRY_WLOCK(inp)) {
1948 			/*
1949 			 * The reference on the socket file descriptor in
1950 			 * the AIO job should keep 'sb' and 'inp' stable.
1951 			 * Our caller has a reference on the 'toep' that
1952 			 * keeps it stable.
1953 			 */
1954 			SOCKBUF_UNLOCK(sb);
1955 			DDP_UNLOCK(toep);
1956 			INP_WLOCK(inp);
1957 			DDP_LOCK(toep);
1958 			SOCKBUF_LOCK(sb);
1959 
1960 			/*
1961 			 * If the socket has been closed, we should detect
1962 			 * that and complete this request if needed on
1963 			 * the next trip around the loop.
1964 			 */
1965 		}
1966 		t4_rcvd_locked(&toep->td->tod, intotcpcb(inp));
1967 		INP_WUNLOCK(inp);
1968 		if (resid == 0 || toep->ddp.flags & DDP_DEAD) {
1969 			/*
1970 			 * We filled the entire buffer with socket
1971 			 * data, DDP is not being used, or the socket
1972 			 * is being shut down, so complete the
1973 			 * request.
1974 			 */
1975 			SOCKBUF_UNLOCK(sb);
1976 			recycle_pageset(toep, ps);
1977 			aio_complete(job, copied, 0);
1978 			toep->ddp.queueing = NULL;
1979 			goto restart;
1980 		}
1981 
1982 		/*
1983 		 * If DDP is not enabled, requeue this request and restart.
1984 		 * This will either enable DDP or wait for more data to
1985 		 * arrive on the socket buffer.
1986 		 */
1987 		if ((toep->ddp.flags & (DDP_ON | DDP_SC_REQ)) != DDP_ON) {
1988 			SOCKBUF_UNLOCK(sb);
1989 			recycle_pageset(toep, ps);
1990 			aio_ddp_requeue_one(toep, job);
1991 			toep->ddp.queueing = NULL;
1992 			goto restart;
1993 		}
1994 
1995 		/*
1996 		 * An indicate might have arrived and been added to
1997 		 * the socket buffer while it was unlocked after the
1998 		 * copy to lock the INP.  If so, restart the copy.
1999 		 */
2000 		if (sbavail(sb) != 0)
2001 			goto sbcopy;
2002 	}
2003 	SOCKBUF_UNLOCK(sb);
2004 
2005 	if (prep_pageset(sc, toep, ps) == 0) {
2006 		recycle_pageset(toep, ps);
2007 		aio_ddp_requeue_one(toep, job);
2008 		toep->ddp.queueing = NULL;
2009 
2010 		/*
2011 		 * XXX: Need to retry this later.  Mostly need a trigger
2012 		 * when page pods are freed up.
2013 		 */
2014 		printf("%s: prep_pageset failed\n", __func__);
2015 		return;
2016 	}
2017 
2018 	/* Determine which DDP buffer to use. */
2019 	if (toep->ddp.db[0].job == NULL) {
2020 		db_idx = 0;
2021 	} else {
2022 		MPASS(toep->ddp.db[1].job == NULL);
2023 		db_idx = 1;
2024 	}
2025 
2026 	ddp_flags = 0;
2027 	ddp_flags_mask = 0;
2028 	if (db_idx == 0) {
2029 		ddp_flags |= V_TF_DDP_BUF0_VALID(1);
2030 		if (so->so_state & SS_NBIO)
2031 			ddp_flags |= V_TF_DDP_BUF0_FLUSH(1);
2032 		ddp_flags_mask |= V_TF_DDP_PSH_NO_INVALIDATE0(1) |
2033 		    V_TF_DDP_PUSH_DISABLE_0(1) | V_TF_DDP_PSHF_ENABLE_0(1) |
2034 		    V_TF_DDP_BUF0_FLUSH(1) | V_TF_DDP_BUF0_VALID(1);
2035 		buf_flag = DDP_BUF0_ACTIVE;
2036 	} else {
2037 		ddp_flags |= V_TF_DDP_BUF1_VALID(1);
2038 		if (so->so_state & SS_NBIO)
2039 			ddp_flags |= V_TF_DDP_BUF1_FLUSH(1);
2040 		ddp_flags_mask |= V_TF_DDP_PSH_NO_INVALIDATE1(1) |
2041 		    V_TF_DDP_PUSH_DISABLE_1(1) | V_TF_DDP_PSHF_ENABLE_1(1) |
2042 		    V_TF_DDP_BUF1_FLUSH(1) | V_TF_DDP_BUF1_VALID(1);
2043 		buf_flag = DDP_BUF1_ACTIVE;
2044 	}
2045 	MPASS((toep->ddp.flags & buf_flag) == 0);
2046 	if ((toep->ddp.flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE)) == 0) {
2047 		MPASS(db_idx == 0);
2048 		MPASS(toep->ddp.active_id == -1);
2049 		MPASS(toep->ddp.active_count == 0);
2050 		ddp_flags_mask |= V_TF_DDP_ACTIVE_BUF(1);
2051 	}
2052 
2053 	/*
2054 	 * The TID for this connection should still be valid.  If DDP_DEAD
2055 	 * is set, SBS_CANTRCVMORE should be set, so we shouldn't be
2056 	 * this far anyway.  Even if the socket is closing on the other
2057 	 * end, the AIO job holds a reference on this end of the socket
2058 	 * which will keep it open and keep the TCP PCB attached until
2059 	 * after the job is completed.
2060 	 */
2061 	wr = mk_update_tcb_for_ddp(sc, toep, db_idx, ps, job->aio_received,
2062 	    ddp_flags, ddp_flags_mask);
2063 	if (wr == NULL) {
2064 		recycle_pageset(toep, ps);
2065 		aio_ddp_requeue_one(toep, job);
2066 		toep->ddp.queueing = NULL;
2067 
2068 		/*
2069 		 * XXX: Need a way to kick a retry here.
2070 		 *
2071 		 * XXX: We know the fixed size needed and could
2072 		 * preallocate this using a blocking request at the
2073 		 * start of the task to avoid having to handle this
2074 		 * edge case.
2075 		 */
2076 		printf("%s: mk_update_tcb_for_ddp failed\n", __func__);
2077 		return;
2078 	}
2079 
2080 	if (!aio_set_cancel_function(job, t4_aio_cancel_active)) {
2081 		free_wrqe(wr);
2082 		recycle_pageset(toep, ps);
2083 		aio_ddp_cancel_one(job);
2084 		toep->ddp.queueing = NULL;
2085 		goto restart;
2086 	}
2087 
2088 #ifdef VERBOSE_TRACES
2089 	CTR6(KTR_CXGBE,
2090 	    "%s: tid %u, scheduling %p for DDP[%d] (flags %#lx/%#lx)", __func__,
2091 	    toep->tid, job, db_idx, ddp_flags, ddp_flags_mask);
2092 #endif
2093 	/* Give the chip the go-ahead. */
2094 	t4_wrq_tx(sc, wr);
2095 	db = &toep->ddp.db[db_idx];
2096 	db->cancel_pending = 0;
2097 	db->job = job;
2098 	db->ps = ps;
2099 	toep->ddp.queueing = NULL;
2100 	toep->ddp.flags |= buf_flag;
2101 	toep->ddp.active_count++;
2102 	if (toep->ddp.active_count == 1) {
2103 		MPASS(toep->ddp.active_id == -1);
2104 		toep->ddp.active_id = db_idx;
2105 		CTR2(KTR_CXGBE, "%s: ddp_active_id = %d", __func__,
2106 		    toep->ddp.active_id);
2107 	}
2108 	goto restart;
2109 }
2110 
2111 void
2112 ddp_queue_toep(struct toepcb *toep)
2113 {
2114 
2115 	DDP_ASSERT_LOCKED(toep);
2116 	if (toep->ddp.flags & DDP_TASK_ACTIVE)
2117 		return;
2118 	toep->ddp.flags |= DDP_TASK_ACTIVE;
2119 	hold_toepcb(toep);
2120 	soaio_enqueue(&toep->ddp.requeue_task);
2121 }
2122 
2123 static void
2124 aio_ddp_requeue_task(void *context, int pending)
2125 {
2126 	struct toepcb *toep = context;
2127 
2128 	DDP_LOCK(toep);
2129 	aio_ddp_requeue(toep);
2130 	toep->ddp.flags &= ~DDP_TASK_ACTIVE;
2131 	DDP_UNLOCK(toep);
2132 
2133 	free_toepcb(toep);
2134 }
2135 
2136 static void
2137 t4_aio_cancel_active(struct kaiocb *job)
2138 {
2139 	struct socket *so = job->fd_file->f_data;
2140 	struct tcpcb *tp = sototcpcb(so);
2141 	struct toepcb *toep = tp->t_toe;
2142 	struct adapter *sc = td_adapter(toep->td);
2143 	uint64_t valid_flag;
2144 	int i;
2145 
2146 	DDP_LOCK(toep);
2147 	if (aio_cancel_cleared(job)) {
2148 		DDP_UNLOCK(toep);
2149 		aio_ddp_cancel_one(job);
2150 		return;
2151 	}
2152 
2153 	for (i = 0; i < nitems(toep->ddp.db); i++) {
2154 		if (toep->ddp.db[i].job == job) {
2155 			/* Should only ever get one cancel request for a job. */
2156 			MPASS(toep->ddp.db[i].cancel_pending == 0);
2157 
2158 			/*
2159 			 * Invalidate this buffer.  It will be
2160 			 * cancelled or partially completed once the
2161 			 * card ACKs the invalidate.
2162 			 */
2163 			valid_flag = i == 0 ? V_TF_DDP_BUF0_VALID(1) :
2164 			    V_TF_DDP_BUF1_VALID(1);
2165 			t4_set_tcb_field(sc, toep->ctrlq, toep,
2166 			    W_TCB_RX_DDP_FLAGS, valid_flag, 0, 1,
2167 			    CPL_COOKIE_DDP0 + i);
2168 			toep->ddp.db[i].cancel_pending = 1;
2169 			CTR2(KTR_CXGBE, "%s: request %p marked pending",
2170 			    __func__, job);
2171 			break;
2172 		}
2173 	}
2174 	DDP_UNLOCK(toep);
2175 }
2176 
2177 static void
2178 t4_aio_cancel_queued(struct kaiocb *job)
2179 {
2180 	struct socket *so = job->fd_file->f_data;
2181 	struct tcpcb *tp = sototcpcb(so);
2182 	struct toepcb *toep = tp->t_toe;
2183 
2184 	DDP_LOCK(toep);
2185 	if (!aio_cancel_cleared(job)) {
2186 		TAILQ_REMOVE(&toep->ddp.aiojobq, job, list);
2187 		toep->ddp.waiting_count--;
2188 		if (toep->ddp.waiting_count == 0)
2189 			ddp_queue_toep(toep);
2190 	}
2191 	CTR2(KTR_CXGBE, "%s: request %p cancelled", __func__, job);
2192 	DDP_UNLOCK(toep);
2193 
2194 	aio_ddp_cancel_one(job);
2195 }
2196 
2197 int
2198 t4_aio_queue_ddp(struct socket *so, struct kaiocb *job)
2199 {
2200 	struct tcpcb *tp = sototcpcb(so);
2201 	struct toepcb *toep = tp->t_toe;
2202 
2203 
2204 	/* Ignore writes. */
2205 	if (job->uaiocb.aio_lio_opcode != LIO_READ)
2206 		return (EOPNOTSUPP);
2207 
2208 	DDP_LOCK(toep);
2209 
2210 	/*
2211 	 * XXX: Think about possibly returning errors for ENOTCONN,
2212 	 * etc.  Perhaps the caller would only queue the request
2213 	 * if it failed with EOPNOTSUPP?
2214 	 */
2215 
2216 #ifdef VERBOSE_TRACES
2217 	CTR3(KTR_CXGBE, "%s: queueing %p for tid %u", __func__, job, toep->tid);
2218 #endif
2219 	if (!aio_set_cancel_function(job, t4_aio_cancel_queued))
2220 		panic("new job was cancelled");
2221 	TAILQ_INSERT_TAIL(&toep->ddp.aiojobq, job, list);
2222 	toep->ddp.waiting_count++;
2223 	toep->ddp.flags |= DDP_OK;
2224 
2225 	/*
2226 	 * Try to handle this request synchronously.  If this has
2227 	 * to block because the task is running, it will just bail
2228 	 * and let the task handle it instead.
2229 	 */
2230 	aio_ddp_requeue(toep);
2231 	DDP_UNLOCK(toep);
2232 	return (0);
2233 }
2234 
2235 void
2236 t4_ddp_mod_load(void)
2237 {
2238 
2239 	t4_register_shared_cpl_handler(CPL_SET_TCB_RPL, do_ddp_tcb_rpl,
2240 	    CPL_COOKIE_DDP0);
2241 	t4_register_shared_cpl_handler(CPL_SET_TCB_RPL, do_ddp_tcb_rpl,
2242 	    CPL_COOKIE_DDP1);
2243 	t4_register_cpl_handler(CPL_RX_DATA_DDP, do_rx_data_ddp);
2244 	t4_register_cpl_handler(CPL_RX_DDP_COMPLETE, do_rx_ddp_complete);
2245 	TAILQ_INIT(&ddp_orphan_pagesets);
2246 	mtx_init(&ddp_orphan_pagesets_lock, "ddp orphans", NULL, MTX_DEF);
2247 	TASK_INIT(&ddp_orphan_task, 0, ddp_free_orphan_pagesets, NULL);
2248 }
2249 
2250 void
2251 t4_ddp_mod_unload(void)
2252 {
2253 
2254 	taskqueue_drain(taskqueue_thread, &ddp_orphan_task);
2255 	MPASS(TAILQ_EMPTY(&ddp_orphan_pagesets));
2256 	mtx_destroy(&ddp_orphan_pagesets_lock);
2257 	t4_register_shared_cpl_handler(CPL_SET_TCB_RPL, NULL, CPL_COOKIE_DDP0);
2258 	t4_register_shared_cpl_handler(CPL_SET_TCB_RPL, NULL, CPL_COOKIE_DDP1);
2259 	t4_register_cpl_handler(CPL_RX_DATA_DDP, NULL);
2260 	t4_register_cpl_handler(CPL_RX_DDP_COMPLETE, NULL);
2261 }
2262 #endif
2263