xref: /freebsd/sys/dev/cxgbe/tom/t4_ddp.c (revision 69718b786d3943ea9a99eeeb5f5f6162f11c78b7)
1 /*-
2  * Copyright (c) 2012 Chelsio Communications, Inc.
3  * All rights reserved.
4  * Written by: Navdeep Parhar <np@FreeBSD.org>
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 
33 #include <sys/param.h>
34 #include <sys/aio.h>
35 #include <sys/file.h>
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
38 #include <sys/ktr.h>
39 #include <sys/module.h>
40 #include <sys/protosw.h>
41 #include <sys/proc.h>
42 #include <sys/domain.h>
43 #include <sys/socket.h>
44 #include <sys/socketvar.h>
45 #include <sys/taskqueue.h>
46 #include <sys/uio.h>
47 #include <netinet/in.h>
48 #include <netinet/in_pcb.h>
49 #include <netinet/ip.h>
50 #include <netinet/tcp_var.h>
51 #define TCPSTATES
52 #include <netinet/tcp_fsm.h>
53 #include <netinet/toecore.h>
54 
55 #include <vm/vm.h>
56 #include <vm/vm_extern.h>
57 #include <vm/vm_param.h>
58 #include <vm/pmap.h>
59 #include <vm/vm_map.h>
60 #include <vm/vm_page.h>
61 #include <vm/vm_object.h>
62 
63 #ifdef TCP_OFFLOAD
64 #include "common/common.h"
65 #include "common/t4_msg.h"
66 #include "common/t4_regs.h"
67 #include "common/t4_tcb.h"
68 #include "tom/t4_tom.h"
69 
70 VNET_DECLARE(int, tcp_do_autorcvbuf);
71 #define V_tcp_do_autorcvbuf VNET(tcp_do_autorcvbuf)
72 VNET_DECLARE(int, tcp_autorcvbuf_inc);
73 #define V_tcp_autorcvbuf_inc VNET(tcp_autorcvbuf_inc)
74 VNET_DECLARE(int, tcp_autorcvbuf_max);
75 #define V_tcp_autorcvbuf_max VNET(tcp_autorcvbuf_max)
76 
77 /*
78  * Use the 'backend3' field in AIO jobs to store the amount of data
79  * received by the AIO job so far.
80  */
81 #define	aio_received	backend3
82 
83 static void aio_ddp_requeue_task(void *context, int pending);
84 static void ddp_complete_all(struct toepcb *toep, int error);
85 static void t4_aio_cancel_active(struct kaiocb *job);
86 static void t4_aio_cancel_queued(struct kaiocb *job);
87 
88 static TAILQ_HEAD(, pageset) ddp_orphan_pagesets;
89 static struct mtx ddp_orphan_pagesets_lock;
90 static struct task ddp_orphan_task;
91 
92 #define MAX_DDP_BUFFER_SIZE		(M_TCB_RX_DDP_BUF0_LEN)
93 
94 /*
95  * A page set holds information about a buffer used for DDP.  The page
96  * set holds resources such as the VM pages backing the buffer (either
97  * held or wired) and the page pods associated with the buffer.
98  * Recently used page sets are cached to allow for efficient reuse of
99  * buffers (avoiding the need to re-fault in pages, hold them, etc.).
100  * Note that cached page sets keep the backing pages wired.  The
101  * number of wired pages is capped by only allowing for two wired
102  * pagesets per connection.  This is not a perfect cap, but is a
103  * trade-off for performance.
104  *
105  * If an application ping-pongs two buffers for a connection via
106  * aio_read(2) then those buffers should remain wired and expensive VM
107  * fault lookups should be avoided after each buffer has been used
108  * once.  If an application uses more than two buffers then this will
109  * fall back to doing expensive VM fault lookups for each operation.
110  */
111 static void
112 free_pageset(struct tom_data *td, struct pageset *ps)
113 {
114 	vm_page_t p;
115 	int i;
116 
117 	if (ps->prsv.prsv_nppods > 0)
118 		t4_free_page_pods(&ps->prsv);
119 
120 	if (ps->flags & PS_WIRED) {
121 		for (i = 0; i < ps->npages; i++) {
122 			p = ps->pages[i];
123 			vm_page_lock(p);
124 			vm_page_unwire(p, PQ_INACTIVE);
125 			vm_page_unlock(p);
126 		}
127 	} else
128 		vm_page_unhold_pages(ps->pages, ps->npages);
129 	mtx_lock(&ddp_orphan_pagesets_lock);
130 	TAILQ_INSERT_TAIL(&ddp_orphan_pagesets, ps, link);
131 	taskqueue_enqueue(taskqueue_thread, &ddp_orphan_task);
132 	mtx_unlock(&ddp_orphan_pagesets_lock);
133 }
134 
135 static void
136 ddp_free_orphan_pagesets(void *context, int pending)
137 {
138 	struct pageset *ps;
139 
140 	mtx_lock(&ddp_orphan_pagesets_lock);
141 	while (!TAILQ_EMPTY(&ddp_orphan_pagesets)) {
142 		ps = TAILQ_FIRST(&ddp_orphan_pagesets);
143 		TAILQ_REMOVE(&ddp_orphan_pagesets, ps, link);
144 		mtx_unlock(&ddp_orphan_pagesets_lock);
145 		if (ps->vm)
146 			vmspace_free(ps->vm);
147 		free(ps, M_CXGBE);
148 		mtx_lock(&ddp_orphan_pagesets_lock);
149 	}
150 	mtx_unlock(&ddp_orphan_pagesets_lock);
151 }
152 
153 static void
154 recycle_pageset(struct toepcb *toep, struct pageset *ps)
155 {
156 
157 	DDP_ASSERT_LOCKED(toep);
158 	if (!(toep->ddp_flags & DDP_DEAD) && ps->flags & PS_WIRED) {
159 		KASSERT(toep->ddp_cached_count + toep->ddp_active_count <
160 		    nitems(toep->db), ("too many wired pagesets"));
161 		TAILQ_INSERT_HEAD(&toep->ddp_cached_pagesets, ps, link);
162 		toep->ddp_cached_count++;
163 	} else
164 		free_pageset(toep->td, ps);
165 }
166 
167 static void
168 ddp_complete_one(struct kaiocb *job, int error)
169 {
170 	long copied;
171 
172 	/*
173 	 * If this job had copied data out of the socket buffer before
174 	 * it was cancelled, report it as a short read rather than an
175 	 * error.
176 	 */
177 	copied = job->aio_received;
178 	if (copied != 0 || error == 0)
179 		aio_complete(job, copied, 0);
180 	else
181 		aio_complete(job, -1, error);
182 }
183 
184 static void
185 free_ddp_buffer(struct tom_data *td, struct ddp_buffer *db)
186 {
187 
188 	if (db->job) {
189 		/*
190 		 * XXX: If we are un-offloading the socket then we
191 		 * should requeue these on the socket somehow.  If we
192 		 * got a FIN from the remote end, then this completes
193 		 * any remaining requests with an EOF read.
194 		 */
195 		if (!aio_clear_cancel_function(db->job))
196 			ddp_complete_one(db->job, 0);
197 	}
198 
199 	if (db->ps)
200 		free_pageset(td, db->ps);
201 }
202 
203 void
204 ddp_init_toep(struct toepcb *toep)
205 {
206 
207 	TAILQ_INIT(&toep->ddp_aiojobq);
208 	TASK_INIT(&toep->ddp_requeue_task, 0, aio_ddp_requeue_task, toep);
209 	toep->ddp_active_id = -1;
210 	mtx_init(&toep->ddp_lock, "t4 ddp", NULL, MTX_DEF);
211 }
212 
213 void
214 ddp_uninit_toep(struct toepcb *toep)
215 {
216 
217 	mtx_destroy(&toep->ddp_lock);
218 }
219 
220 void
221 release_ddp_resources(struct toepcb *toep)
222 {
223 	struct pageset *ps;
224 	int i;
225 
226 	DDP_LOCK(toep);
227 	toep->flags |= DDP_DEAD;
228 	for (i = 0; i < nitems(toep->db); i++) {
229 		free_ddp_buffer(toep->td, &toep->db[i]);
230 	}
231 	while ((ps = TAILQ_FIRST(&toep->ddp_cached_pagesets)) != NULL) {
232 		TAILQ_REMOVE(&toep->ddp_cached_pagesets, ps, link);
233 		free_pageset(toep->td, ps);
234 	}
235 	ddp_complete_all(toep, 0);
236 	DDP_UNLOCK(toep);
237 }
238 
239 #ifdef INVARIANTS
240 void
241 ddp_assert_empty(struct toepcb *toep)
242 {
243 	int i;
244 
245 	MPASS(!(toep->ddp_flags & DDP_TASK_ACTIVE));
246 	for (i = 0; i < nitems(toep->db); i++) {
247 		MPASS(toep->db[i].job == NULL);
248 		MPASS(toep->db[i].ps == NULL);
249 	}
250 	MPASS(TAILQ_EMPTY(&toep->ddp_cached_pagesets));
251 	MPASS(TAILQ_EMPTY(&toep->ddp_aiojobq));
252 }
253 #endif
254 
255 static void
256 complete_ddp_buffer(struct toepcb *toep, struct ddp_buffer *db,
257     unsigned int db_idx)
258 {
259 	unsigned int db_flag;
260 
261 	toep->ddp_active_count--;
262 	if (toep->ddp_active_id == db_idx) {
263 		if (toep->ddp_active_count == 0) {
264 			KASSERT(toep->db[db_idx ^ 1].job == NULL,
265 			    ("%s: active_count mismatch", __func__));
266 			toep->ddp_active_id = -1;
267 		} else
268 			toep->ddp_active_id ^= 1;
269 #ifdef VERBOSE_TRACES
270 		CTR2(KTR_CXGBE, "%s: ddp_active_id = %d", __func__,
271 		    toep->ddp_active_id);
272 #endif
273 	} else {
274 		KASSERT(toep->ddp_active_count != 0 &&
275 		    toep->ddp_active_id != -1,
276 		    ("%s: active count mismatch", __func__));
277 	}
278 
279 	db->cancel_pending = 0;
280 	db->job = NULL;
281 	recycle_pageset(toep, db->ps);
282 	db->ps = NULL;
283 
284 	db_flag = db_idx == 1 ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE;
285 	KASSERT(toep->ddp_flags & db_flag,
286 	    ("%s: DDP buffer not active. toep %p, ddp_flags 0x%x",
287 	    __func__, toep, toep->ddp_flags));
288 	toep->ddp_flags &= ~db_flag;
289 }
290 
291 /* XXX: handle_ddp_data code duplication */
292 void
293 insert_ddp_data(struct toepcb *toep, uint32_t n)
294 {
295 	struct inpcb *inp = toep->inp;
296 	struct tcpcb *tp = intotcpcb(inp);
297 	struct ddp_buffer *db;
298 	struct kaiocb *job;
299 	size_t placed;
300 	long copied;
301 	unsigned int db_flag, db_idx;
302 
303 	INP_WLOCK_ASSERT(inp);
304 	DDP_ASSERT_LOCKED(toep);
305 
306 	tp->rcv_nxt += n;
307 #ifndef USE_DDP_RX_FLOW_CONTROL
308 	KASSERT(tp->rcv_wnd >= n, ("%s: negative window size", __func__));
309 	tp->rcv_wnd -= n;
310 #endif
311 #ifndef USE_DDP_RX_FLOW_CONTROL
312 	toep->rx_credits += n;
313 #endif
314 	CTR2(KTR_CXGBE, "%s: placed %u bytes before falling out of DDP",
315 	    __func__, n);
316 	while (toep->ddp_active_count > 0) {
317 		MPASS(toep->ddp_active_id != -1);
318 		db_idx = toep->ddp_active_id;
319 		db_flag = db_idx == 1 ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE;
320 		MPASS((toep->ddp_flags & db_flag) != 0);
321 		db = &toep->db[db_idx];
322 		job = db->job;
323 		copied = job->aio_received;
324 		placed = n;
325 		if (placed > job->uaiocb.aio_nbytes - copied)
326 			placed = job->uaiocb.aio_nbytes - copied;
327 		if (placed > 0)
328 			job->msgrcv = 1;
329 		if (!aio_clear_cancel_function(job)) {
330 			/*
331 			 * Update the copied length for when
332 			 * t4_aio_cancel_active() completes this
333 			 * request.
334 			 */
335 			job->aio_received += placed;
336 		} else if (copied + placed != 0) {
337 			CTR4(KTR_CXGBE,
338 			    "%s: completing %p (copied %ld, placed %lu)",
339 			    __func__, job, copied, placed);
340 			/* XXX: This always completes if there is some data. */
341 			aio_complete(job, copied + placed, 0);
342 		} else if (aio_set_cancel_function(job, t4_aio_cancel_queued)) {
343 			TAILQ_INSERT_HEAD(&toep->ddp_aiojobq, job, list);
344 			toep->ddp_waiting_count++;
345 		} else
346 			aio_cancel(job);
347 		n -= placed;
348 		complete_ddp_buffer(toep, db, db_idx);
349 	}
350 
351 	MPASS(n == 0);
352 }
353 
354 /* SET_TCB_FIELD sent as a ULP command looks like this */
355 #define LEN__SET_TCB_FIELD_ULP (sizeof(struct ulp_txpkt) + \
356     sizeof(struct ulptx_idata) + sizeof(struct cpl_set_tcb_field_core))
357 
358 /* RX_DATA_ACK sent as a ULP command looks like this */
359 #define LEN__RX_DATA_ACK_ULP (sizeof(struct ulp_txpkt) + \
360     sizeof(struct ulptx_idata) + sizeof(struct cpl_rx_data_ack_core))
361 
362 static inline void *
363 mk_set_tcb_field_ulp(struct ulp_txpkt *ulpmc, struct toepcb *toep,
364     uint64_t word, uint64_t mask, uint64_t val)
365 {
366 	struct ulptx_idata *ulpsc;
367 	struct cpl_set_tcb_field_core *req;
368 
369 	ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0));
370 	ulpmc->len = htobe32(howmany(LEN__SET_TCB_FIELD_ULP, 16));
371 
372 	ulpsc = (struct ulptx_idata *)(ulpmc + 1);
373 	ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
374 	ulpsc->len = htobe32(sizeof(*req));
375 
376 	req = (struct cpl_set_tcb_field_core *)(ulpsc + 1);
377 	OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_SET_TCB_FIELD, toep->tid));
378 	req->reply_ctrl = htobe16(V_NO_REPLY(1) |
379 	    V_QUEUENO(toep->ofld_rxq->iq.abs_id));
380 	req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0));
381         req->mask = htobe64(mask);
382         req->val = htobe64(val);
383 
384 	ulpsc = (struct ulptx_idata *)(req + 1);
385 	if (LEN__SET_TCB_FIELD_ULP % 16) {
386 		ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP));
387 		ulpsc->len = htobe32(0);
388 		return (ulpsc + 1);
389 	}
390 	return (ulpsc);
391 }
392 
393 static inline void *
394 mk_rx_data_ack_ulp(struct ulp_txpkt *ulpmc, struct toepcb *toep)
395 {
396 	struct ulptx_idata *ulpsc;
397 	struct cpl_rx_data_ack_core *req;
398 
399 	ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0));
400 	ulpmc->len = htobe32(howmany(LEN__RX_DATA_ACK_ULP, 16));
401 
402 	ulpsc = (struct ulptx_idata *)(ulpmc + 1);
403 	ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
404 	ulpsc->len = htobe32(sizeof(*req));
405 
406 	req = (struct cpl_rx_data_ack_core *)(ulpsc + 1);
407 	OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_RX_DATA_ACK, toep->tid));
408 	req->credit_dack = htobe32(F_RX_MODULATE_RX);
409 
410 	ulpsc = (struct ulptx_idata *)(req + 1);
411 	if (LEN__RX_DATA_ACK_ULP % 16) {
412 		ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP));
413 		ulpsc->len = htobe32(0);
414 		return (ulpsc + 1);
415 	}
416 	return (ulpsc);
417 }
418 
419 static struct wrqe *
420 mk_update_tcb_for_ddp(struct adapter *sc, struct toepcb *toep, int db_idx,
421     struct pageset *ps, int offset, uint64_t ddp_flags, uint64_t ddp_flags_mask)
422 {
423 	struct wrqe *wr;
424 	struct work_request_hdr *wrh;
425 	struct ulp_txpkt *ulpmc;
426 	int len;
427 
428 	KASSERT(db_idx == 0 || db_idx == 1,
429 	    ("%s: bad DDP buffer index %d", __func__, db_idx));
430 
431 	/*
432 	 * We'll send a compound work request that has 3 SET_TCB_FIELDs and an
433 	 * RX_DATA_ACK (with RX_MODULATE to speed up delivery).
434 	 *
435 	 * The work request header is 16B and always ends at a 16B boundary.
436 	 * The ULPTX master commands that follow must all end at 16B boundaries
437 	 * too so we round up the size to 16.
438 	 */
439 	len = sizeof(*wrh) + 3 * roundup2(LEN__SET_TCB_FIELD_ULP, 16) +
440 	    roundup2(LEN__RX_DATA_ACK_ULP, 16);
441 
442 	wr = alloc_wrqe(len, toep->ctrlq);
443 	if (wr == NULL)
444 		return (NULL);
445 	wrh = wrtod(wr);
446 	INIT_ULPTX_WRH(wrh, len, 1, 0);	/* atomic */
447 	ulpmc = (struct ulp_txpkt *)(wrh + 1);
448 
449 	/* Write the buffer's tag */
450 	ulpmc = mk_set_tcb_field_ulp(ulpmc, toep,
451 	    W_TCB_RX_DDP_BUF0_TAG + db_idx,
452 	    V_TCB_RX_DDP_BUF0_TAG(M_TCB_RX_DDP_BUF0_TAG),
453 	    V_TCB_RX_DDP_BUF0_TAG(ps->prsv.prsv_tag));
454 
455 	/* Update the current offset in the DDP buffer and its total length */
456 	if (db_idx == 0)
457 		ulpmc = mk_set_tcb_field_ulp(ulpmc, toep,
458 		    W_TCB_RX_DDP_BUF0_OFFSET,
459 		    V_TCB_RX_DDP_BUF0_OFFSET(M_TCB_RX_DDP_BUF0_OFFSET) |
460 		    V_TCB_RX_DDP_BUF0_LEN(M_TCB_RX_DDP_BUF0_LEN),
461 		    V_TCB_RX_DDP_BUF0_OFFSET(offset) |
462 		    V_TCB_RX_DDP_BUF0_LEN(ps->len));
463 	else
464 		ulpmc = mk_set_tcb_field_ulp(ulpmc, toep,
465 		    W_TCB_RX_DDP_BUF1_OFFSET,
466 		    V_TCB_RX_DDP_BUF1_OFFSET(M_TCB_RX_DDP_BUF1_OFFSET) |
467 		    V_TCB_RX_DDP_BUF1_LEN((u64)M_TCB_RX_DDP_BUF1_LEN << 32),
468 		    V_TCB_RX_DDP_BUF1_OFFSET(offset) |
469 		    V_TCB_RX_DDP_BUF1_LEN((u64)ps->len << 32));
470 
471 	/* Update DDP flags */
472 	ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, W_TCB_RX_DDP_FLAGS,
473 	    ddp_flags_mask, ddp_flags);
474 
475 	/* Gratuitous RX_DATA_ACK with RX_MODULATE set to speed up delivery. */
476 	ulpmc = mk_rx_data_ack_ulp(ulpmc, toep);
477 
478 	return (wr);
479 }
480 
481 static int
482 handle_ddp_data(struct toepcb *toep, __be32 ddp_report, __be32 rcv_nxt, int len)
483 {
484 	uint32_t report = be32toh(ddp_report);
485 	unsigned int db_idx;
486 	struct inpcb *inp = toep->inp;
487 	struct ddp_buffer *db;
488 	struct tcpcb *tp;
489 	struct socket *so;
490 	struct sockbuf *sb;
491 	struct kaiocb *job;
492 	long copied;
493 
494 	db_idx = report & F_DDP_BUF_IDX ? 1 : 0;
495 
496 	if (__predict_false(!(report & F_DDP_INV)))
497 		CXGBE_UNIMPLEMENTED("DDP buffer still valid");
498 
499 	INP_WLOCK(inp);
500 	so = inp_inpcbtosocket(inp);
501 	sb = &so->so_rcv;
502 	DDP_LOCK(toep);
503 
504 	KASSERT(toep->ddp_active_id == db_idx,
505 	    ("completed DDP buffer (%d) != active_id (%d) for tid %d", db_idx,
506 	    toep->ddp_active_id, toep->tid));
507 	db = &toep->db[db_idx];
508 	job = db->job;
509 
510 	if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))) {
511 		/*
512 		 * This can happen due to an administrative tcpdrop(8).
513 		 * Just fail the request with ECONNRESET.
514 		 */
515 		CTR5(KTR_CXGBE, "%s: tid %u, seq 0x%x, len %d, inp_flags 0x%x",
516 		    __func__, toep->tid, be32toh(rcv_nxt), len, inp->inp_flags);
517 		if (aio_clear_cancel_function(job))
518 			ddp_complete_one(job, ECONNRESET);
519 		goto completed;
520 	}
521 
522 	tp = intotcpcb(inp);
523 
524 	/*
525 	 * For RX_DDP_COMPLETE, len will be zero and rcv_nxt is the
526 	 * sequence number of the next byte to receive.  The length of
527 	 * the data received for this message must be computed by
528 	 * comparing the new and old values of rcv_nxt.
529 	 *
530 	 * For RX_DATA_DDP, len might be non-zero, but it is only the
531 	 * length of the most recent DMA.  It does not include the
532 	 * total length of the data received since the previous update
533 	 * for this DDP buffer.  rcv_nxt is the sequence number of the
534 	 * first received byte from the most recent DMA.
535 	 */
536 	len += be32toh(rcv_nxt) - tp->rcv_nxt;
537 	tp->rcv_nxt += len;
538 	tp->t_rcvtime = ticks;
539 #ifndef USE_DDP_RX_FLOW_CONTROL
540 	KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__));
541 	tp->rcv_wnd -= len;
542 #endif
543 #ifdef VERBOSE_TRACES
544 	CTR4(KTR_CXGBE, "%s: DDP[%d] placed %d bytes (%#x)", __func__, db_idx,
545 	    len, report);
546 #endif
547 
548 	/* receive buffer autosize */
549 	CURVNET_SET(so->so_vnet);
550 	SOCKBUF_LOCK(sb);
551 	if (sb->sb_flags & SB_AUTOSIZE &&
552 	    V_tcp_do_autorcvbuf &&
553 	    sb->sb_hiwat < V_tcp_autorcvbuf_max &&
554 	    len > (sbspace(sb) / 8 * 7)) {
555 		unsigned int hiwat = sb->sb_hiwat;
556 		unsigned int newsize = min(hiwat + V_tcp_autorcvbuf_inc,
557 		    V_tcp_autorcvbuf_max);
558 
559 		if (!sbreserve_locked(sb, newsize, so, NULL))
560 			sb->sb_flags &= ~SB_AUTOSIZE;
561 		else
562 			toep->rx_credits += newsize - hiwat;
563 	}
564 	SOCKBUF_UNLOCK(sb);
565 	CURVNET_RESTORE();
566 
567 #ifndef USE_DDP_RX_FLOW_CONTROL
568 	toep->rx_credits += len;
569 #endif
570 
571 	job->msgrcv = 1;
572 	if (db->cancel_pending) {
573 		/*
574 		 * Update the job's length but defer completion to the
575 		 * TCB_RPL callback.
576 		 */
577 		job->aio_received += len;
578 		goto out;
579 	} else if (!aio_clear_cancel_function(job)) {
580 		/*
581 		 * Update the copied length for when
582 		 * t4_aio_cancel_active() completes this request.
583 		 */
584 		job->aio_received += len;
585 	} else {
586 		copied = job->aio_received;
587 #ifdef VERBOSE_TRACES
588 		CTR4(KTR_CXGBE, "%s: completing %p (copied %ld, placed %d)",
589 		    __func__, job, copied, len);
590 #endif
591 		aio_complete(job, copied + len, 0);
592 		t4_rcvd(&toep->td->tod, tp);
593 	}
594 
595 completed:
596 	complete_ddp_buffer(toep, db, db_idx);
597 	if (toep->ddp_waiting_count > 0)
598 		ddp_queue_toep(toep);
599 out:
600 	DDP_UNLOCK(toep);
601 	INP_WUNLOCK(inp);
602 
603 	return (0);
604 }
605 
606 void
607 handle_ddp_indicate(struct toepcb *toep)
608 {
609 
610 	DDP_ASSERT_LOCKED(toep);
611 	MPASS(toep->ddp_active_count == 0);
612 	MPASS((toep->ddp_flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE)) == 0);
613 	if (toep->ddp_waiting_count == 0) {
614 		/*
615 		 * The pending requests that triggered the request for an
616 		 * an indicate were cancelled.  Those cancels should have
617 		 * already disabled DDP.  Just ignore this as the data is
618 		 * going into the socket buffer anyway.
619 		 */
620 		return;
621 	}
622 	CTR3(KTR_CXGBE, "%s: tid %d indicated (%d waiting)", __func__,
623 	    toep->tid, toep->ddp_waiting_count);
624 	ddp_queue_toep(toep);
625 }
626 
627 enum {
628 	DDP_BUF0_INVALIDATED = 0x2,
629 	DDP_BUF1_INVALIDATED
630 };
631 
632 void
633 handle_ddp_tcb_rpl(struct toepcb *toep, const struct cpl_set_tcb_rpl *cpl)
634 {
635 	unsigned int db_idx;
636 	struct inpcb *inp = toep->inp;
637 	struct ddp_buffer *db;
638 	struct kaiocb *job;
639 	long copied;
640 
641 	if (cpl->status != CPL_ERR_NONE)
642 		panic("XXX: tcp_rpl failed: %d", cpl->status);
643 
644 	switch (cpl->cookie) {
645 	case V_WORD(W_TCB_RX_DDP_FLAGS) | V_COOKIE(DDP_BUF0_INVALIDATED):
646 	case V_WORD(W_TCB_RX_DDP_FLAGS) | V_COOKIE(DDP_BUF1_INVALIDATED):
647 		/*
648 		 * XXX: This duplicates a lot of code with handle_ddp_data().
649 		 */
650 		db_idx = G_COOKIE(cpl->cookie) - DDP_BUF0_INVALIDATED;
651 		INP_WLOCK(inp);
652 		DDP_LOCK(toep);
653 		db = &toep->db[db_idx];
654 
655 		/*
656 		 * handle_ddp_data() should leave the job around until
657 		 * this callback runs once a cancel is pending.
658 		 */
659 		MPASS(db != NULL);
660 		MPASS(db->job != NULL);
661 		MPASS(db->cancel_pending);
662 
663 		/*
664 		 * XXX: It's not clear what happens if there is data
665 		 * placed when the buffer is invalidated.  I suspect we
666 		 * need to read the TCB to see how much data was placed.
667 		 *
668 		 * For now this just pretends like nothing was placed.
669 		 *
670 		 * XXX: Note that if we did check the PCB we would need to
671 		 * also take care of updating the tp, etc.
672 		 */
673 		job = db->job;
674 		copied = job->aio_received;
675 		if (copied == 0) {
676 			CTR2(KTR_CXGBE, "%s: cancelling %p", __func__, job);
677 			aio_cancel(job);
678 		} else {
679 			CTR3(KTR_CXGBE, "%s: completing %p (copied %ld)",
680 			    __func__, job, copied);
681 			aio_complete(job, copied, 0);
682 			t4_rcvd(&toep->td->tod, intotcpcb(inp));
683 		}
684 
685 		complete_ddp_buffer(toep, db, db_idx);
686 		if (toep->ddp_waiting_count > 0)
687 			ddp_queue_toep(toep);
688 		DDP_UNLOCK(toep);
689 		INP_WUNLOCK(inp);
690 		break;
691 	default:
692 		panic("XXX: unknown tcb_rpl offset %#x, cookie %#x",
693 		    G_WORD(cpl->cookie), G_COOKIE(cpl->cookie));
694 	}
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_flag, db_idx;
704 	int len, placed;
705 
706 	INP_WLOCK_ASSERT(toep->inp);
707 	DDP_ASSERT_LOCKED(toep);
708 	len = be32toh(rcv_nxt) - tp->rcv_nxt;
709 
710 	tp->rcv_nxt += len;
711 #ifndef USE_DDP_RX_FLOW_CONTROL
712 	toep->rx_credits += len;
713 #endif
714 
715 	while (toep->ddp_active_count > 0) {
716 		MPASS(toep->ddp_active_id != -1);
717 		db_idx = toep->ddp_active_id;
718 		db_flag = db_idx == 1 ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE;
719 		MPASS((toep->ddp_flags & db_flag) != 0);
720 		db = &toep->db[db_idx];
721 		job = db->job;
722 		copied = job->aio_received;
723 		placed = len;
724 		if (placed > job->uaiocb.aio_nbytes - copied)
725 			placed = job->uaiocb.aio_nbytes - copied;
726 		if (placed > 0)
727 			job->msgrcv = 1;
728 		if (!aio_clear_cancel_function(job)) {
729 			/*
730 			 * Update the copied length for when
731 			 * t4_aio_cancel_active() completes this
732 			 * request.
733 			 */
734 			job->aio_received += placed;
735 		} else {
736 			CTR4(KTR_CXGBE, "%s: tid %d completed buf %d len %d",
737 			    __func__, toep->tid, db_idx, placed);
738 			aio_complete(job, copied + placed, 0);
739 		}
740 		len -= placed;
741 		complete_ddp_buffer(toep, db, db_idx);
742 	}
743 
744 	MPASS(len == 0);
745 	ddp_complete_all(toep, 0);
746 }
747 
748 #define DDP_ERR (F_DDP_PPOD_MISMATCH | F_DDP_LLIMIT_ERR | F_DDP_ULIMIT_ERR |\
749 	 F_DDP_PPOD_PARITY_ERR | F_DDP_PADDING_ERR | F_DDP_OFFSET_ERR |\
750 	 F_DDP_INVALID_TAG | F_DDP_COLOR_ERR | F_DDP_TID_MISMATCH |\
751 	 F_DDP_INVALID_PPOD | F_DDP_HDRCRC_ERR | F_DDP_DATACRC_ERR)
752 
753 extern cpl_handler_t t4_cpl_handler[];
754 
755 static int
756 do_rx_data_ddp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
757 {
758 	struct adapter *sc = iq->adapter;
759 	const struct cpl_rx_data_ddp *cpl = (const void *)(rss + 1);
760 	unsigned int tid = GET_TID(cpl);
761 	uint32_t vld;
762 	struct toepcb *toep = lookup_tid(sc, tid);
763 
764 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
765 	KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__));
766 	KASSERT(!(toep->flags & TPF_SYNQE),
767 	    ("%s: toep %p claims to be a synq entry", __func__, toep));
768 
769 	vld = be32toh(cpl->ddpvld);
770 	if (__predict_false(vld & DDP_ERR)) {
771 		panic("%s: DDP error 0x%x (tid %d, toep %p)",
772 		    __func__, vld, tid, toep);
773 	}
774 
775 	if (toep->ulp_mode == ULP_MODE_ISCSI) {
776 		t4_cpl_handler[CPL_RX_ISCSI_DDP](iq, rss, m);
777 		return (0);
778 	}
779 
780 	handle_ddp_data(toep, cpl->u.ddp_report, cpl->seq, be16toh(cpl->len));
781 
782 	return (0);
783 }
784 
785 static int
786 do_rx_ddp_complete(struct sge_iq *iq, const struct rss_header *rss,
787     struct mbuf *m)
788 {
789 	struct adapter *sc = iq->adapter;
790 	const struct cpl_rx_ddp_complete *cpl = (const void *)(rss + 1);
791 	unsigned int tid = GET_TID(cpl);
792 	struct toepcb *toep = lookup_tid(sc, tid);
793 
794 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
795 	KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__));
796 	KASSERT(!(toep->flags & TPF_SYNQE),
797 	    ("%s: toep %p claims to be a synq entry", __func__, toep));
798 
799 	handle_ddp_data(toep, cpl->ddp_report, cpl->rcv_nxt, 0);
800 
801 	return (0);
802 }
803 
804 static void
805 enable_ddp(struct adapter *sc, struct toepcb *toep)
806 {
807 
808 	KASSERT((toep->ddp_flags & (DDP_ON | DDP_OK | DDP_SC_REQ)) == DDP_OK,
809 	    ("%s: toep %p has bad ddp_flags 0x%x",
810 	    __func__, toep, toep->ddp_flags));
811 
812 	CTR3(KTR_CXGBE, "%s: tid %u (time %u)",
813 	    __func__, toep->tid, time_uptime);
814 
815 	DDP_ASSERT_LOCKED(toep);
816 	toep->ddp_flags |= DDP_SC_REQ;
817 	t4_set_tcb_field(sc, toep->ctrlq, toep->tid, W_TCB_RX_DDP_FLAGS,
818 	    V_TF_DDP_OFF(1) | V_TF_DDP_INDICATE_OUT(1) |
819 	    V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1) |
820 	    V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_BUF1_VALID(1),
821 	    V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1), 0, 0,
822 	    toep->ofld_rxq->iq.abs_id);
823 	t4_set_tcb_field(sc, toep->ctrlq, toep->tid, W_TCB_T_FLAGS,
824 	    V_TF_RCV_COALESCE_ENABLE(1), 0, 0, 0, toep->ofld_rxq->iq.abs_id);
825 }
826 
827 static int
828 calculate_hcf(int n1, int n2)
829 {
830 	int a, b, t;
831 
832 	if (n1 <= n2) {
833 		a = n1;
834 		b = n2;
835 	} else {
836 		a = n2;
837 		b = n1;
838 	}
839 
840 	while (a != 0) {
841 		t = a;
842 		a = b % a;
843 		b = t;
844 	}
845 
846 	return (b);
847 }
848 
849 static inline int
850 pages_to_nppods(int npages, int ddp_page_shift)
851 {
852 
853 	MPASS(ddp_page_shift >= PAGE_SHIFT);
854 
855 	return (howmany(npages >> (ddp_page_shift - PAGE_SHIFT), PPOD_PAGES));
856 }
857 
858 static int
859 alloc_page_pods(struct ppod_region *pr, u_int nppods, u_int pgsz_idx,
860     struct ppod_reservation *prsv)
861 {
862 	vmem_addr_t addr;       /* relative to start of region */
863 
864 	if (vmem_alloc(pr->pr_arena, PPOD_SZ(nppods), M_NOWAIT | M_FIRSTFIT,
865 	    &addr) != 0)
866 		return (ENOMEM);
867 
868 	CTR5(KTR_CXGBE, "%-17s arena %p, addr 0x%08x, nppods %d, pgsz %d",
869 	    __func__, pr->pr_arena, (uint32_t)addr & pr->pr_tag_mask,
870 	    nppods, 1 << pr->pr_page_shift[pgsz_idx]);
871 
872 	/*
873 	 * The hardware tagmask includes an extra invalid bit but the arena was
874 	 * seeded with valid values only.  An allocation out of this arena will
875 	 * fit inside the tagmask but won't have the invalid bit set.
876 	 */
877 	MPASS((addr & pr->pr_tag_mask) == addr);
878 	MPASS((addr & pr->pr_invalid_bit) == 0);
879 
880 	prsv->prsv_pr = pr;
881 	prsv->prsv_tag = V_PPOD_PGSZ(pgsz_idx) | addr;
882 	prsv->prsv_nppods = nppods;
883 
884 	return (0);
885 }
886 
887 int
888 t4_alloc_page_pods_for_ps(struct ppod_region *pr, struct pageset *ps)
889 {
890 	int i, hcf, seglen, idx, nppods;
891 	struct ppod_reservation *prsv = &ps->prsv;
892 
893 	KASSERT(prsv->prsv_nppods == 0,
894 	    ("%s: page pods already allocated", __func__));
895 
896 	/*
897 	 * The DDP page size is unrelated to the VM page size.  We combine
898 	 * contiguous physical pages into larger segments to get the best DDP
899 	 * page size possible.  This is the largest of the four sizes in
900 	 * A_ULP_RX_TDDP_PSZ that evenly divides the HCF of the segment sizes in
901 	 * the page list.
902 	 */
903 	hcf = 0;
904 	for (i = 0; i < ps->npages; i++) {
905 		seglen = PAGE_SIZE;
906 		while (i < ps->npages - 1 &&
907 		    ps->pages[i]->phys_addr + PAGE_SIZE ==
908 		    ps->pages[i + 1]->phys_addr) {
909 			seglen += PAGE_SIZE;
910 			i++;
911 		}
912 
913 		hcf = calculate_hcf(hcf, seglen);
914 		if (hcf < (1 << pr->pr_page_shift[1])) {
915 			idx = 0;
916 			goto have_pgsz;	/* give up, short circuit */
917 		}
918 	}
919 
920 #define PR_PAGE_MASK(x) ((1 << pr->pr_page_shift[(x)]) - 1)
921 	MPASS((hcf & PR_PAGE_MASK(0)) == 0); /* PAGE_SIZE is >= 4K everywhere */
922 	for (idx = nitems(pr->pr_page_shift) - 1; idx > 0; idx--) {
923 		if ((hcf & PR_PAGE_MASK(idx)) == 0)
924 			break;
925 	}
926 #undef PR_PAGE_MASK
927 
928 have_pgsz:
929 	MPASS(idx <= M_PPOD_PGSZ);
930 
931 	nppods = pages_to_nppods(ps->npages, pr->pr_page_shift[idx]);
932 	if (alloc_page_pods(pr, nppods, idx, prsv) != 0)
933 		return (0);
934 	MPASS(prsv->prsv_nppods > 0);
935 
936 	return (1);
937 }
938 
939 int
940 t4_alloc_page_pods_for_buf(struct ppod_region *pr, vm_offset_t buf, int len,
941     struct ppod_reservation *prsv)
942 {
943 	int hcf, seglen, idx, npages, nppods;
944 	uintptr_t start_pva, end_pva, pva, p1;
945 
946 	MPASS(buf > 0);
947 	MPASS(len > 0);
948 
949 	/*
950 	 * The DDP page size is unrelated to the VM page size.  We combine
951 	 * contiguous physical pages into larger segments to get the best DDP
952 	 * page size possible.  This is the largest of the four sizes in
953 	 * A_ULP_RX_ISCSI_PSZ that evenly divides the HCF of the segment sizes
954 	 * in the page list.
955 	 */
956 	hcf = 0;
957 	start_pva = trunc_page(buf);
958 	end_pva = trunc_page(buf + len - 1);
959 	pva = start_pva;
960 	while (pva <= end_pva) {
961 		seglen = PAGE_SIZE;
962 		p1 = pmap_kextract(pva);
963 		pva += PAGE_SIZE;
964 		while (pva <= end_pva && p1 + seglen == pmap_kextract(pva)) {
965 			seglen += PAGE_SIZE;
966 			pva += PAGE_SIZE;
967 		}
968 
969 		hcf = calculate_hcf(hcf, seglen);
970 		if (hcf < (1 << pr->pr_page_shift[1])) {
971 			idx = 0;
972 			goto have_pgsz;	/* give up, short circuit */
973 		}
974 	}
975 
976 #define PR_PAGE_MASK(x) ((1 << pr->pr_page_shift[(x)]) - 1)
977 	MPASS((hcf & PR_PAGE_MASK(0)) == 0); /* PAGE_SIZE is >= 4K everywhere */
978 	for (idx = nitems(pr->pr_page_shift) - 1; idx > 0; idx--) {
979 		if ((hcf & PR_PAGE_MASK(idx)) == 0)
980 			break;
981 	}
982 #undef PR_PAGE_MASK
983 
984 have_pgsz:
985 	MPASS(idx <= M_PPOD_PGSZ);
986 
987 	npages = 1;
988 	npages += (end_pva - start_pva) >> pr->pr_page_shift[idx];
989 	nppods = howmany(npages, PPOD_PAGES);
990 	if (alloc_page_pods(pr, nppods, idx, prsv) != 0)
991 		return (ENOMEM);
992 	MPASS(prsv->prsv_nppods > 0);
993 
994 	return (0);
995 }
996 
997 void
998 t4_free_page_pods(struct ppod_reservation *prsv)
999 {
1000 	struct ppod_region *pr = prsv->prsv_pr;
1001 	vmem_addr_t addr;
1002 
1003 	MPASS(prsv != NULL);
1004 	MPASS(prsv->prsv_nppods != 0);
1005 
1006 	addr = prsv->prsv_tag & pr->pr_tag_mask;
1007 	MPASS((addr & pr->pr_invalid_bit) == 0);
1008 
1009 	CTR4(KTR_CXGBE, "%-17s arena %p, addr 0x%08x, nppods %d", __func__,
1010 	    pr->pr_arena, addr, prsv->prsv_nppods);
1011 
1012 	vmem_free(pr->pr_arena, addr, PPOD_SZ(prsv->prsv_nppods));
1013 	prsv->prsv_nppods = 0;
1014 }
1015 
1016 #define NUM_ULP_TX_SC_IMM_PPODS (256 / PPOD_SIZE)
1017 
1018 int
1019 t4_write_page_pods_for_ps(struct adapter *sc, struct sge_wrq *wrq, int tid,
1020     struct pageset *ps)
1021 {
1022 	struct wrqe *wr;
1023 	struct ulp_mem_io *ulpmc;
1024 	struct ulptx_idata *ulpsc;
1025 	struct pagepod *ppod;
1026 	int i, j, k, n, chunk, len, ddp_pgsz, idx;
1027 	u_int ppod_addr;
1028 	uint32_t cmd;
1029 	struct ppod_reservation *prsv = &ps->prsv;
1030 	struct ppod_region *pr = prsv->prsv_pr;
1031 
1032 	KASSERT(!(ps->flags & PS_PPODS_WRITTEN),
1033 	    ("%s: page pods already written", __func__));
1034 	MPASS(prsv->prsv_nppods > 0);
1035 
1036 	cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE));
1037 	if (is_t4(sc))
1038 		cmd |= htobe32(F_ULP_MEMIO_ORDER);
1039 	else
1040 		cmd |= htobe32(F_T5_ULP_MEMIO_IMM);
1041 	ddp_pgsz = 1 << pr->pr_page_shift[G_PPOD_PGSZ(prsv->prsv_tag)];
1042 	ppod_addr = pr->pr_start + (prsv->prsv_tag & pr->pr_tag_mask);
1043 	for (i = 0; i < prsv->prsv_nppods; ppod_addr += chunk) {
1044 
1045 		/* How many page pods are we writing in this cycle */
1046 		n = min(prsv->prsv_nppods - i, NUM_ULP_TX_SC_IMM_PPODS);
1047 		chunk = PPOD_SZ(n);
1048 		len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16);
1049 
1050 		wr = alloc_wrqe(len, wrq);
1051 		if (wr == NULL)
1052 			return (ENOMEM);	/* ok to just bail out */
1053 		ulpmc = wrtod(wr);
1054 
1055 		INIT_ULPTX_WR(ulpmc, len, 0, 0);
1056 		ulpmc->cmd = cmd;
1057 		ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32));
1058 		ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16));
1059 		ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5));
1060 
1061 		ulpsc = (struct ulptx_idata *)(ulpmc + 1);
1062 		ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
1063 		ulpsc->len = htobe32(chunk);
1064 
1065 		ppod = (struct pagepod *)(ulpsc + 1);
1066 		for (j = 0; j < n; i++, j++, ppod++) {
1067 			ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID |
1068 			    V_PPOD_TID(tid) | prsv->prsv_tag);
1069 			ppod->len_offset = htobe64(V_PPOD_LEN(ps->len) |
1070 			    V_PPOD_OFST(ps->offset));
1071 			ppod->rsvd = 0;
1072 			idx = i * PPOD_PAGES * (ddp_pgsz / PAGE_SIZE);
1073 			for (k = 0; k < nitems(ppod->addr); k++) {
1074 				if (idx < ps->npages) {
1075 					ppod->addr[k] =
1076 					    htobe64(ps->pages[idx]->phys_addr);
1077 					idx += ddp_pgsz / PAGE_SIZE;
1078 				} else
1079 					ppod->addr[k] = 0;
1080 #if 0
1081 				CTR5(KTR_CXGBE,
1082 				    "%s: tid %d ppod[%d]->addr[%d] = %p",
1083 				    __func__, toep->tid, i, k,
1084 				    htobe64(ppod->addr[k]));
1085 #endif
1086 			}
1087 
1088 		}
1089 
1090 		t4_wrq_tx(sc, wr);
1091 	}
1092 	ps->flags |= PS_PPODS_WRITTEN;
1093 
1094 	return (0);
1095 }
1096 
1097 int
1098 t4_write_page_pods_for_buf(struct adapter *sc, struct sge_wrq *wrq, int tid,
1099     struct ppod_reservation *prsv, vm_offset_t buf, int buflen)
1100 {
1101 	struct wrqe *wr;
1102 	struct ulp_mem_io *ulpmc;
1103 	struct ulptx_idata *ulpsc;
1104 	struct pagepod *ppod;
1105 	int i, j, k, n, chunk, len, ddp_pgsz;
1106 	u_int ppod_addr, offset;
1107 	uint32_t cmd;
1108 	struct ppod_region *pr = prsv->prsv_pr;
1109 	uintptr_t end_pva, pva, pa;
1110 
1111 	cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE));
1112 	if (is_t4(sc))
1113 		cmd |= htobe32(F_ULP_MEMIO_ORDER);
1114 	else
1115 		cmd |= htobe32(F_T5_ULP_MEMIO_IMM);
1116 	ddp_pgsz = 1 << pr->pr_page_shift[G_PPOD_PGSZ(prsv->prsv_tag)];
1117 	offset = buf & PAGE_MASK;
1118 	ppod_addr = pr->pr_start + (prsv->prsv_tag & pr->pr_tag_mask);
1119 	pva = trunc_page(buf);
1120 	end_pva = trunc_page(buf + buflen - 1);
1121 	for (i = 0; i < prsv->prsv_nppods; ppod_addr += chunk) {
1122 
1123 		/* How many page pods are we writing in this cycle */
1124 		n = min(prsv->prsv_nppods - i, NUM_ULP_TX_SC_IMM_PPODS);
1125 		MPASS(n > 0);
1126 		chunk = PPOD_SZ(n);
1127 		len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16);
1128 
1129 		wr = alloc_wrqe(len, wrq);
1130 		if (wr == NULL)
1131 			return (ENOMEM);	/* ok to just bail out */
1132 		ulpmc = wrtod(wr);
1133 
1134 		INIT_ULPTX_WR(ulpmc, len, 0, 0);
1135 		ulpmc->cmd = cmd;
1136 		ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32));
1137 		ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16));
1138 		ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5));
1139 
1140 		ulpsc = (struct ulptx_idata *)(ulpmc + 1);
1141 		ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
1142 		ulpsc->len = htobe32(chunk);
1143 
1144 		ppod = (struct pagepod *)(ulpsc + 1);
1145 		for (j = 0; j < n; i++, j++, ppod++) {
1146 			ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID |
1147 			    V_PPOD_TID(tid) |
1148 			    (prsv->prsv_tag & ~V_PPOD_PGSZ(M_PPOD_PGSZ)));
1149 			ppod->len_offset = htobe64(V_PPOD_LEN(buflen) |
1150 			    V_PPOD_OFST(offset));
1151 			ppod->rsvd = 0;
1152 
1153 			for (k = 0; k < nitems(ppod->addr); k++) {
1154 				if (pva > end_pva)
1155 					ppod->addr[k] = 0;
1156 				else {
1157 					pa = pmap_kextract(pva);
1158 					ppod->addr[k] = htobe64(pa);
1159 					pva += ddp_pgsz;
1160 				}
1161 #if 0
1162 				CTR5(KTR_CXGBE,
1163 				    "%s: tid %d ppod[%d]->addr[%d] = %p",
1164 				    __func__, tid, i, k,
1165 				    htobe64(ppod->addr[k]));
1166 #endif
1167 			}
1168 
1169 			/*
1170 			 * Walk back 1 segment so that the first address in the
1171 			 * next pod is the same as the last one in the current
1172 			 * pod.
1173 			 */
1174 			pva -= ddp_pgsz;
1175 		}
1176 
1177 		t4_wrq_tx(sc, wr);
1178 	}
1179 
1180 	MPASS(pva <= end_pva);
1181 
1182 	return (0);
1183 }
1184 
1185 static void
1186 wire_pageset(struct pageset *ps)
1187 {
1188 	vm_page_t p;
1189 	int i;
1190 
1191 	KASSERT(!(ps->flags & PS_WIRED), ("pageset already wired"));
1192 
1193 	for (i = 0; i < ps->npages; i++) {
1194 		p = ps->pages[i];
1195 		vm_page_lock(p);
1196 		vm_page_wire(p);
1197 		vm_page_unhold(p);
1198 		vm_page_unlock(p);
1199 	}
1200 	ps->flags |= PS_WIRED;
1201 }
1202 
1203 /*
1204  * Prepare a pageset for DDP.  This wires the pageset and sets up page
1205  * pods.
1206  */
1207 static int
1208 prep_pageset(struct adapter *sc, struct toepcb *toep, struct pageset *ps)
1209 {
1210 	struct tom_data *td = sc->tom_softc;
1211 
1212 	if (!(ps->flags & PS_WIRED))
1213 		wire_pageset(ps);
1214 	if (ps->prsv.prsv_nppods == 0 &&
1215 	    !t4_alloc_page_pods_for_ps(&td->pr, ps)) {
1216 		return (0);
1217 	}
1218 	if (!(ps->flags & PS_PPODS_WRITTEN) &&
1219 	    t4_write_page_pods_for_ps(sc, toep->ctrlq, toep->tid, ps) != 0) {
1220 		return (0);
1221 	}
1222 
1223 	return (1);
1224 }
1225 
1226 int
1227 t4_init_ppod_region(struct ppod_region *pr, struct t4_range *r, u_int psz,
1228     const char *name)
1229 {
1230 	int i;
1231 
1232 	MPASS(pr != NULL);
1233 	MPASS(r->size > 0);
1234 
1235 	pr->pr_start = r->start;
1236 	pr->pr_len = r->size;
1237 	pr->pr_page_shift[0] = 12 + G_HPZ0(psz);
1238 	pr->pr_page_shift[1] = 12 + G_HPZ1(psz);
1239 	pr->pr_page_shift[2] = 12 + G_HPZ2(psz);
1240 	pr->pr_page_shift[3] = 12 + G_HPZ3(psz);
1241 
1242 	/* The SGL -> page pod algorithm requires the sizes to be in order. */
1243 	for (i = 1; i < nitems(pr->pr_page_shift); i++) {
1244 		if (pr->pr_page_shift[i] <= pr->pr_page_shift[i - 1])
1245 			return (ENXIO);
1246 	}
1247 
1248 	pr->pr_tag_mask = ((1 << fls(r->size)) - 1) & V_PPOD_TAG(M_PPOD_TAG);
1249 	pr->pr_alias_mask = V_PPOD_TAG(M_PPOD_TAG) & ~pr->pr_tag_mask;
1250 	if (pr->pr_tag_mask == 0 || pr->pr_alias_mask == 0)
1251 		return (ENXIO);
1252 	pr->pr_alias_shift = fls(pr->pr_tag_mask);
1253 	pr->pr_invalid_bit = 1 << (pr->pr_alias_shift - 1);
1254 
1255 	pr->pr_arena = vmem_create(name, 0, pr->pr_len, PPOD_SIZE, 0,
1256 	    M_FIRSTFIT | M_NOWAIT);
1257 	if (pr->pr_arena == NULL)
1258 		return (ENOMEM);
1259 
1260 	return (0);
1261 }
1262 
1263 void
1264 t4_free_ppod_region(struct ppod_region *pr)
1265 {
1266 
1267 	MPASS(pr != NULL);
1268 
1269 	if (pr->pr_arena)
1270 		vmem_destroy(pr->pr_arena);
1271 	bzero(pr, sizeof(*pr));
1272 }
1273 
1274 static int
1275 pscmp(struct pageset *ps, struct vmspace *vm, vm_offset_t start, int npages,
1276     int pgoff, int len)
1277 {
1278 
1279 	if (ps->npages != npages || ps->offset != pgoff || ps->len != len)
1280 		return (1);
1281 
1282 	return (ps->vm != vm || ps->vm_timestamp != vm->vm_map.timestamp);
1283 }
1284 
1285 static int
1286 hold_aio(struct toepcb *toep, struct kaiocb *job, struct pageset **pps)
1287 {
1288 	struct vmspace *vm;
1289 	vm_map_t map;
1290 	vm_offset_t start, end, pgoff;
1291 	struct pageset *ps;
1292 	int n;
1293 
1294 	DDP_ASSERT_LOCKED(toep);
1295 
1296 	/*
1297 	 * The AIO subsystem will cancel and drain all requests before
1298 	 * permitting a process to exit or exec, so p_vmspace should
1299 	 * be stable here.
1300 	 */
1301 	vm = job->userproc->p_vmspace;
1302 	map = &vm->vm_map;
1303 	start = (uintptr_t)job->uaiocb.aio_buf;
1304 	pgoff = start & PAGE_MASK;
1305 	end = round_page(start + job->uaiocb.aio_nbytes);
1306 	start = trunc_page(start);
1307 
1308 	if (end - start > MAX_DDP_BUFFER_SIZE) {
1309 		/*
1310 		 * Truncate the request to a short read.
1311 		 * Alternatively, we could DDP in chunks to the larger
1312 		 * buffer, but that would be quite a bit more work.
1313 		 *
1314 		 * When truncating, round the request down to avoid
1315 		 * crossing a cache line on the final transaction.
1316 		 */
1317 		end = rounddown2(start + MAX_DDP_BUFFER_SIZE, CACHE_LINE_SIZE);
1318 #ifdef VERBOSE_TRACES
1319 		CTR4(KTR_CXGBE, "%s: tid %d, truncating size from %lu to %lu",
1320 		    __func__, toep->tid, (unsigned long)job->uaiocb.aio_nbytes,
1321 		    (unsigned long)(end - (start + pgoff)));
1322 		job->uaiocb.aio_nbytes = end - (start + pgoff);
1323 #endif
1324 		end = round_page(end);
1325 	}
1326 
1327 	n = atop(end - start);
1328 
1329 	/*
1330 	 * Try to reuse a cached pageset.
1331 	 */
1332 	TAILQ_FOREACH(ps, &toep->ddp_cached_pagesets, link) {
1333 		if (pscmp(ps, vm, start, n, pgoff,
1334 		    job->uaiocb.aio_nbytes) == 0) {
1335 			TAILQ_REMOVE(&toep->ddp_cached_pagesets, ps, link);
1336 			toep->ddp_cached_count--;
1337 			*pps = ps;
1338 			return (0);
1339 		}
1340 	}
1341 
1342 	/*
1343 	 * If there are too many cached pagesets to create a new one,
1344 	 * free a pageset before creating a new one.
1345 	 */
1346 	KASSERT(toep->ddp_active_count + toep->ddp_cached_count <=
1347 	    nitems(toep->db), ("%s: too many wired pagesets", __func__));
1348 	if (toep->ddp_active_count + toep->ddp_cached_count ==
1349 	    nitems(toep->db)) {
1350 		KASSERT(toep->ddp_cached_count > 0,
1351 		    ("no cached pageset to free"));
1352 		ps = TAILQ_LAST(&toep->ddp_cached_pagesets, pagesetq);
1353 		TAILQ_REMOVE(&toep->ddp_cached_pagesets, ps, link);
1354 		toep->ddp_cached_count--;
1355 		free_pageset(toep->td, ps);
1356 	}
1357 	DDP_UNLOCK(toep);
1358 
1359 	/* Create a new pageset. */
1360 	ps = malloc(sizeof(*ps) + n * sizeof(vm_page_t), M_CXGBE, M_WAITOK |
1361 	    M_ZERO);
1362 	ps->pages = (vm_page_t *)(ps + 1);
1363 	ps->vm_timestamp = map->timestamp;
1364 	ps->npages = vm_fault_quick_hold_pages(map, start, end - start,
1365 	    VM_PROT_WRITE, ps->pages, n);
1366 
1367 	DDP_LOCK(toep);
1368 	if (ps->npages < 0) {
1369 		free(ps, M_CXGBE);
1370 		return (EFAULT);
1371 	}
1372 
1373 	KASSERT(ps->npages == n, ("hold_aio: page count mismatch: %d vs %d",
1374 	    ps->npages, n));
1375 
1376 	ps->offset = pgoff;
1377 	ps->len = job->uaiocb.aio_nbytes;
1378 	atomic_add_int(&vm->vm_refcnt, 1);
1379 	ps->vm = vm;
1380 
1381 	CTR5(KTR_CXGBE, "%s: tid %d, new pageset %p for job %p, npages %d",
1382 	    __func__, toep->tid, ps, job, ps->npages);
1383 	*pps = ps;
1384 	return (0);
1385 }
1386 
1387 static void
1388 ddp_complete_all(struct toepcb *toep, int error)
1389 {
1390 	struct kaiocb *job;
1391 
1392 	DDP_ASSERT_LOCKED(toep);
1393 	while (!TAILQ_EMPTY(&toep->ddp_aiojobq)) {
1394 		job = TAILQ_FIRST(&toep->ddp_aiojobq);
1395 		TAILQ_REMOVE(&toep->ddp_aiojobq, job, list);
1396 		toep->ddp_waiting_count--;
1397 		if (aio_clear_cancel_function(job))
1398 			ddp_complete_one(job, error);
1399 	}
1400 }
1401 
1402 static void
1403 aio_ddp_cancel_one(struct kaiocb *job)
1404 {
1405 	long copied;
1406 
1407 	/*
1408 	 * If this job had copied data out of the socket buffer before
1409 	 * it was cancelled, report it as a short read rather than an
1410 	 * error.
1411 	 */
1412 	copied = job->aio_received;
1413 	if (copied != 0)
1414 		aio_complete(job, copied, 0);
1415 	else
1416 		aio_cancel(job);
1417 }
1418 
1419 /*
1420  * Called when the main loop wants to requeue a job to retry it later.
1421  * Deals with the race of the job being cancelled while it was being
1422  * examined.
1423  */
1424 static void
1425 aio_ddp_requeue_one(struct toepcb *toep, struct kaiocb *job)
1426 {
1427 
1428 	DDP_ASSERT_LOCKED(toep);
1429 	if (!(toep->ddp_flags & DDP_DEAD) &&
1430 	    aio_set_cancel_function(job, t4_aio_cancel_queued)) {
1431 		TAILQ_INSERT_HEAD(&toep->ddp_aiojobq, job, list);
1432 		toep->ddp_waiting_count++;
1433 	} else
1434 		aio_ddp_cancel_one(job);
1435 }
1436 
1437 static void
1438 aio_ddp_requeue(struct toepcb *toep)
1439 {
1440 	struct adapter *sc = td_adapter(toep->td);
1441 	struct socket *so;
1442 	struct sockbuf *sb;
1443 	struct inpcb *inp;
1444 	struct kaiocb *job;
1445 	struct ddp_buffer *db;
1446 	size_t copied, offset, resid;
1447 	struct pageset *ps;
1448 	struct mbuf *m;
1449 	uint64_t ddp_flags, ddp_flags_mask;
1450 	struct wrqe *wr;
1451 	int buf_flag, db_idx, error;
1452 
1453 	DDP_ASSERT_LOCKED(toep);
1454 
1455 restart:
1456 	if (toep->ddp_flags & DDP_DEAD) {
1457 		MPASS(toep->ddp_waiting_count == 0);
1458 		MPASS(toep->ddp_active_count == 0);
1459 		return;
1460 	}
1461 
1462 	if (toep->ddp_waiting_count == 0 ||
1463 	    toep->ddp_active_count == nitems(toep->db)) {
1464 		return;
1465 	}
1466 
1467 	job = TAILQ_FIRST(&toep->ddp_aiojobq);
1468 	so = job->fd_file->f_data;
1469 	sb = &so->so_rcv;
1470 	SOCKBUF_LOCK(sb);
1471 
1472 	/* We will never get anything unless we are or were connected. */
1473 	if (!(so->so_state & (SS_ISCONNECTED|SS_ISDISCONNECTED))) {
1474 		SOCKBUF_UNLOCK(sb);
1475 		ddp_complete_all(toep, ENOTCONN);
1476 		return;
1477 	}
1478 
1479 	KASSERT(toep->ddp_active_count == 0 || sbavail(sb) == 0,
1480 	    ("%s: pending sockbuf data and DDP is active", __func__));
1481 
1482 	/* Abort if socket has reported problems. */
1483 	/* XXX: Wait for any queued DDP's to finish and/or flush them? */
1484 	if (so->so_error && sbavail(sb) == 0) {
1485 		toep->ddp_waiting_count--;
1486 		TAILQ_REMOVE(&toep->ddp_aiojobq, job, list);
1487 		if (!aio_clear_cancel_function(job)) {
1488 			SOCKBUF_UNLOCK(sb);
1489 			goto restart;
1490 		}
1491 
1492 		/*
1493 		 * If this job has previously copied some data, report
1494 		 * a short read and leave the error to be reported by
1495 		 * a future request.
1496 		 */
1497 		copied = job->aio_received;
1498 		if (copied != 0) {
1499 			SOCKBUF_UNLOCK(sb);
1500 			aio_complete(job, copied, 0);
1501 			goto restart;
1502 		}
1503 		error = so->so_error;
1504 		so->so_error = 0;
1505 		SOCKBUF_UNLOCK(sb);
1506 		aio_complete(job, -1, error);
1507 		goto restart;
1508 	}
1509 
1510 	/*
1511 	 * Door is closed.  If there is pending data in the socket buffer,
1512 	 * deliver it.  If there are pending DDP requests, wait for those
1513 	 * to complete.  Once they have completed, return EOF reads.
1514 	 */
1515 	if (sb->sb_state & SBS_CANTRCVMORE && sbavail(sb) == 0) {
1516 		SOCKBUF_UNLOCK(sb);
1517 		if (toep->ddp_active_count != 0)
1518 			return;
1519 		ddp_complete_all(toep, 0);
1520 		return;
1521 	}
1522 
1523 	/*
1524 	 * If DDP is not enabled and there is no pending socket buffer
1525 	 * data, try to enable DDP.
1526 	 */
1527 	if (sbavail(sb) == 0 && (toep->ddp_flags & DDP_ON) == 0) {
1528 		SOCKBUF_UNLOCK(sb);
1529 
1530 		/*
1531 		 * Wait for the card to ACK that DDP is enabled before
1532 		 * queueing any buffers.  Currently this waits for an
1533 		 * indicate to arrive.  This could use a TCB_SET_FIELD_RPL
1534 		 * message to know that DDP was enabled instead of waiting
1535 		 * for the indicate which would avoid copying the indicate
1536 		 * if no data is pending.
1537 		 *
1538 		 * XXX: Might want to limit the indicate size to the size
1539 		 * of the first queued request.
1540 		 */
1541 		if ((toep->ddp_flags & DDP_SC_REQ) == 0)
1542 			enable_ddp(sc, toep);
1543 		return;
1544 	}
1545 	SOCKBUF_UNLOCK(sb);
1546 
1547 	/*
1548 	 * If another thread is queueing a buffer for DDP, let it
1549 	 * drain any work and return.
1550 	 */
1551 	if (toep->ddp_queueing != NULL)
1552 		return;
1553 
1554 	/* Take the next job to prep it for DDP. */
1555 	toep->ddp_waiting_count--;
1556 	TAILQ_REMOVE(&toep->ddp_aiojobq, job, list);
1557 	if (!aio_clear_cancel_function(job))
1558 		goto restart;
1559 	toep->ddp_queueing = job;
1560 
1561 	/* NB: This drops DDP_LOCK while it holds the backing VM pages. */
1562 	error = hold_aio(toep, job, &ps);
1563 	if (error != 0) {
1564 		ddp_complete_one(job, error);
1565 		toep->ddp_queueing = NULL;
1566 		goto restart;
1567 	}
1568 
1569 	SOCKBUF_LOCK(sb);
1570 	if (so->so_error && sbavail(sb) == 0) {
1571 		copied = job->aio_received;
1572 		if (copied != 0) {
1573 			SOCKBUF_UNLOCK(sb);
1574 			recycle_pageset(toep, ps);
1575 			aio_complete(job, copied, 0);
1576 			toep->ddp_queueing = NULL;
1577 			goto restart;
1578 		}
1579 
1580 		error = so->so_error;
1581 		so->so_error = 0;
1582 		SOCKBUF_UNLOCK(sb);
1583 		recycle_pageset(toep, ps);
1584 		aio_complete(job, -1, error);
1585 		toep->ddp_queueing = NULL;
1586 		goto restart;
1587 	}
1588 
1589 	if (sb->sb_state & SBS_CANTRCVMORE && sbavail(sb) == 0) {
1590 		SOCKBUF_UNLOCK(sb);
1591 		recycle_pageset(toep, ps);
1592 		if (toep->ddp_active_count != 0) {
1593 			/*
1594 			 * The door is closed, but there are still pending
1595 			 * DDP buffers.  Requeue.  These jobs will all be
1596 			 * completed once those buffers drain.
1597 			 */
1598 			aio_ddp_requeue_one(toep, job);
1599 			toep->ddp_queueing = NULL;
1600 			return;
1601 		}
1602 		ddp_complete_one(job, 0);
1603 		ddp_complete_all(toep, 0);
1604 		toep->ddp_queueing = NULL;
1605 		return;
1606 	}
1607 
1608 sbcopy:
1609 	/*
1610 	 * If the toep is dead, there shouldn't be any data in the socket
1611 	 * buffer, so the above case should have handled this.
1612 	 */
1613 	MPASS(!(toep->ddp_flags & DDP_DEAD));
1614 
1615 	/*
1616 	 * If there is pending data in the socket buffer (either
1617 	 * from before the requests were queued or a DDP indicate),
1618 	 * copy those mbufs out directly.
1619 	 */
1620 	copied = 0;
1621 	offset = ps->offset + job->aio_received;
1622 	MPASS(job->aio_received <= job->uaiocb.aio_nbytes);
1623 	resid = job->uaiocb.aio_nbytes - job->aio_received;
1624 	m = sb->sb_mb;
1625 	KASSERT(m == NULL || toep->ddp_active_count == 0,
1626 	    ("%s: sockbuf data with active DDP", __func__));
1627 	while (m != NULL && resid > 0) {
1628 		struct iovec iov[1];
1629 		struct uio uio;
1630 		int error;
1631 
1632 		iov[0].iov_base = mtod(m, void *);
1633 		iov[0].iov_len = m->m_len;
1634 		if (iov[0].iov_len > resid)
1635 			iov[0].iov_len = resid;
1636 		uio.uio_iov = iov;
1637 		uio.uio_iovcnt = 1;
1638 		uio.uio_offset = 0;
1639 		uio.uio_resid = iov[0].iov_len;
1640 		uio.uio_segflg = UIO_SYSSPACE;
1641 		uio.uio_rw = UIO_WRITE;
1642 		error = uiomove_fromphys(ps->pages, offset + copied,
1643 		    uio.uio_resid, &uio);
1644 		MPASS(error == 0 && uio.uio_resid == 0);
1645 		copied += uio.uio_offset;
1646 		resid -= uio.uio_offset;
1647 		m = m->m_next;
1648 	}
1649 	if (copied != 0) {
1650 		sbdrop_locked(sb, copied);
1651 		job->aio_received += copied;
1652 		job->msgrcv = 1;
1653 		copied = job->aio_received;
1654 		inp = sotoinpcb(so);
1655 		if (!INP_TRY_WLOCK(inp)) {
1656 			/*
1657 			 * The reference on the socket file descriptor in
1658 			 * the AIO job should keep 'sb' and 'inp' stable.
1659 			 * Our caller has a reference on the 'toep' that
1660 			 * keeps it stable.
1661 			 */
1662 			SOCKBUF_UNLOCK(sb);
1663 			DDP_UNLOCK(toep);
1664 			INP_WLOCK(inp);
1665 			DDP_LOCK(toep);
1666 			SOCKBUF_LOCK(sb);
1667 
1668 			/*
1669 			 * If the socket has been closed, we should detect
1670 			 * that and complete this request if needed on
1671 			 * the next trip around the loop.
1672 			 */
1673 		}
1674 		t4_rcvd_locked(&toep->td->tod, intotcpcb(inp));
1675 		INP_WUNLOCK(inp);
1676 		if (resid == 0 || toep->ddp_flags & DDP_DEAD) {
1677 			/*
1678 			 * We filled the entire buffer with socket
1679 			 * data, DDP is not being used, or the socket
1680 			 * is being shut down, so complete the
1681 			 * request.
1682 			 */
1683 			SOCKBUF_UNLOCK(sb);
1684 			recycle_pageset(toep, ps);
1685 			aio_complete(job, copied, 0);
1686 			toep->ddp_queueing = NULL;
1687 			goto restart;
1688 		}
1689 
1690 		/*
1691 		 * If DDP is not enabled, requeue this request and restart.
1692 		 * This will either enable DDP or wait for more data to
1693 		 * arrive on the socket buffer.
1694 		 */
1695 		if ((toep->ddp_flags & (DDP_ON | DDP_SC_REQ)) != DDP_ON) {
1696 			SOCKBUF_UNLOCK(sb);
1697 			recycle_pageset(toep, ps);
1698 			aio_ddp_requeue_one(toep, job);
1699 			toep->ddp_queueing = NULL;
1700 			goto restart;
1701 		}
1702 
1703 		/*
1704 		 * An indicate might have arrived and been added to
1705 		 * the socket buffer while it was unlocked after the
1706 		 * copy to lock the INP.  If so, restart the copy.
1707 		 */
1708 		if (sbavail(sb) != 0)
1709 			goto sbcopy;
1710 	}
1711 	SOCKBUF_UNLOCK(sb);
1712 
1713 	if (prep_pageset(sc, toep, ps) == 0) {
1714 		recycle_pageset(toep, ps);
1715 		aio_ddp_requeue_one(toep, job);
1716 		toep->ddp_queueing = NULL;
1717 
1718 		/*
1719 		 * XXX: Need to retry this later.  Mostly need a trigger
1720 		 * when page pods are freed up.
1721 		 */
1722 		printf("%s: prep_pageset failed\n", __func__);
1723 		return;
1724 	}
1725 
1726 	/* Determine which DDP buffer to use. */
1727 	if (toep->db[0].job == NULL) {
1728 		db_idx = 0;
1729 	} else {
1730 		MPASS(toep->db[1].job == NULL);
1731 		db_idx = 1;
1732 	}
1733 
1734 	ddp_flags = 0;
1735 	ddp_flags_mask = 0;
1736 	if (db_idx == 0) {
1737 		ddp_flags |= V_TF_DDP_BUF0_VALID(1);
1738 		if (so->so_state & SS_NBIO)
1739 			ddp_flags |= V_TF_DDP_BUF0_FLUSH(1);
1740 		ddp_flags_mask |= V_TF_DDP_PSH_NO_INVALIDATE0(1) |
1741 		    V_TF_DDP_PUSH_DISABLE_0(1) | V_TF_DDP_PSHF_ENABLE_0(1) |
1742 		    V_TF_DDP_BUF0_FLUSH(1) | V_TF_DDP_BUF0_VALID(1);
1743 		buf_flag = DDP_BUF0_ACTIVE;
1744 	} else {
1745 		ddp_flags |= V_TF_DDP_BUF1_VALID(1);
1746 		if (so->so_state & SS_NBIO)
1747 			ddp_flags |= V_TF_DDP_BUF1_FLUSH(1);
1748 		ddp_flags_mask |= V_TF_DDP_PSH_NO_INVALIDATE1(1) |
1749 		    V_TF_DDP_PUSH_DISABLE_1(1) | V_TF_DDP_PSHF_ENABLE_1(1) |
1750 		    V_TF_DDP_BUF1_FLUSH(1) | V_TF_DDP_BUF1_VALID(1);
1751 		buf_flag = DDP_BUF1_ACTIVE;
1752 	}
1753 	MPASS((toep->ddp_flags & buf_flag) == 0);
1754 	if ((toep->ddp_flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE)) == 0) {
1755 		MPASS(db_idx == 0);
1756 		MPASS(toep->ddp_active_id == -1);
1757 		MPASS(toep->ddp_active_count == 0);
1758 		ddp_flags_mask |= V_TF_DDP_ACTIVE_BUF(1);
1759 	}
1760 
1761 	/*
1762 	 * The TID for this connection should still be valid.  If DDP_DEAD
1763 	 * is set, SBS_CANTRCVMORE should be set, so we shouldn't be
1764 	 * this far anyway.  Even if the socket is closing on the other
1765 	 * end, the AIO job holds a reference on this end of the socket
1766 	 * which will keep it open and keep the TCP PCB attached until
1767 	 * after the job is completed.
1768 	 */
1769 	wr = mk_update_tcb_for_ddp(sc, toep, db_idx, ps, job->aio_received,
1770 	    ddp_flags, ddp_flags_mask);
1771 	if (wr == NULL) {
1772 		recycle_pageset(toep, ps);
1773 		aio_ddp_requeue_one(toep, job);
1774 		toep->ddp_queueing = NULL;
1775 
1776 		/*
1777 		 * XXX: Need a way to kick a retry here.
1778 		 *
1779 		 * XXX: We know the fixed size needed and could
1780 		 * preallocate this using a blocking request at the
1781 		 * start of the task to avoid having to handle this
1782 		 * edge case.
1783 		 */
1784 		printf("%s: mk_update_tcb_for_ddp failed\n", __func__);
1785 		return;
1786 	}
1787 
1788 	if (!aio_set_cancel_function(job, t4_aio_cancel_active)) {
1789 		free_wrqe(wr);
1790 		recycle_pageset(toep, ps);
1791 		aio_ddp_cancel_one(job);
1792 		toep->ddp_queueing = NULL;
1793 		goto restart;
1794 	}
1795 
1796 #ifdef VERBOSE_TRACES
1797 	CTR5(KTR_CXGBE, "%s: scheduling %p for DDP[%d] (flags %#lx/%#lx)",
1798 	    __func__, job, db_idx, ddp_flags, ddp_flags_mask);
1799 #endif
1800 	/* Give the chip the go-ahead. */
1801 	t4_wrq_tx(sc, wr);
1802 	db = &toep->db[db_idx];
1803 	db->cancel_pending = 0;
1804 	db->job = job;
1805 	db->ps = ps;
1806 	toep->ddp_queueing = NULL;
1807 	toep->ddp_flags |= buf_flag;
1808 	toep->ddp_active_count++;
1809 	if (toep->ddp_active_count == 1) {
1810 		MPASS(toep->ddp_active_id == -1);
1811 		toep->ddp_active_id = db_idx;
1812 		CTR2(KTR_CXGBE, "%s: ddp_active_id = %d", __func__,
1813 		    toep->ddp_active_id);
1814 	}
1815 	goto restart;
1816 }
1817 
1818 void
1819 ddp_queue_toep(struct toepcb *toep)
1820 {
1821 
1822 	DDP_ASSERT_LOCKED(toep);
1823 	if (toep->ddp_flags & DDP_TASK_ACTIVE)
1824 		return;
1825 	toep->ddp_flags |= DDP_TASK_ACTIVE;
1826 	hold_toepcb(toep);
1827 	soaio_enqueue(&toep->ddp_requeue_task);
1828 }
1829 
1830 static void
1831 aio_ddp_requeue_task(void *context, int pending)
1832 {
1833 	struct toepcb *toep = context;
1834 
1835 	DDP_LOCK(toep);
1836 	aio_ddp_requeue(toep);
1837 	toep->ddp_flags &= ~DDP_TASK_ACTIVE;
1838 	DDP_UNLOCK(toep);
1839 
1840 	free_toepcb(toep);
1841 }
1842 
1843 static void
1844 t4_aio_cancel_active(struct kaiocb *job)
1845 {
1846 	struct socket *so = job->fd_file->f_data;
1847 	struct tcpcb *tp = so_sototcpcb(so);
1848 	struct toepcb *toep = tp->t_toe;
1849 	struct adapter *sc = td_adapter(toep->td);
1850 	uint64_t valid_flag;
1851 	int i;
1852 
1853 	DDP_LOCK(toep);
1854 	if (aio_cancel_cleared(job)) {
1855 		DDP_UNLOCK(toep);
1856 		aio_ddp_cancel_one(job);
1857 		return;
1858 	}
1859 
1860 	for (i = 0; i < nitems(toep->db); i++) {
1861 		if (toep->db[i].job == job) {
1862 			/* Should only ever get one cancel request for a job. */
1863 			MPASS(toep->db[i].cancel_pending == 0);
1864 
1865 			/*
1866 			 * Invalidate this buffer.  It will be
1867 			 * cancelled or partially completed once the
1868 			 * card ACKs the invalidate.
1869 			 */
1870 			valid_flag = i == 0 ? V_TF_DDP_BUF0_VALID(1) :
1871 			    V_TF_DDP_BUF1_VALID(1);
1872 			t4_set_tcb_field(sc, toep->ctrlq, toep->tid,
1873 			    W_TCB_RX_DDP_FLAGS, valid_flag, 0, 1,
1874 			    i + DDP_BUF0_INVALIDATED,
1875 			    toep->ofld_rxq->iq.abs_id);
1876 			toep->db[i].cancel_pending = 1;
1877 			CTR2(KTR_CXGBE, "%s: request %p marked pending",
1878 			    __func__, job);
1879 			break;
1880 		}
1881 	}
1882 	DDP_UNLOCK(toep);
1883 }
1884 
1885 static void
1886 t4_aio_cancel_queued(struct kaiocb *job)
1887 {
1888 	struct socket *so = job->fd_file->f_data;
1889 	struct tcpcb *tp = so_sototcpcb(so);
1890 	struct toepcb *toep = tp->t_toe;
1891 
1892 	DDP_LOCK(toep);
1893 	if (!aio_cancel_cleared(job)) {
1894 		TAILQ_REMOVE(&toep->ddp_aiojobq, job, list);
1895 		toep->ddp_waiting_count--;
1896 		if (toep->ddp_waiting_count == 0)
1897 			ddp_queue_toep(toep);
1898 	}
1899 	CTR2(KTR_CXGBE, "%s: request %p cancelled", __func__, job);
1900 	DDP_UNLOCK(toep);
1901 
1902 	aio_ddp_cancel_one(job);
1903 }
1904 
1905 int
1906 t4_aio_queue_ddp(struct socket *so, struct kaiocb *job)
1907 {
1908 	struct tcpcb *tp = so_sototcpcb(so);
1909 	struct toepcb *toep = tp->t_toe;
1910 
1911 
1912 	/* Ignore writes. */
1913 	if (job->uaiocb.aio_lio_opcode != LIO_READ)
1914 		return (EOPNOTSUPP);
1915 
1916 	DDP_LOCK(toep);
1917 
1918 	/*
1919 	 * XXX: Think about possibly returning errors for ENOTCONN,
1920 	 * etc.  Perhaps the caller would only queue the request
1921 	 * if it failed with EOPNOTSUPP?
1922 	 */
1923 
1924 #ifdef VERBOSE_TRACES
1925 	CTR2(KTR_CXGBE, "%s: queueing %p", __func__, job);
1926 #endif
1927 	if (!aio_set_cancel_function(job, t4_aio_cancel_queued))
1928 		panic("new job was cancelled");
1929 	TAILQ_INSERT_TAIL(&toep->ddp_aiojobq, job, list);
1930 	toep->ddp_waiting_count++;
1931 	toep->ddp_flags |= DDP_OK;
1932 
1933 	/*
1934 	 * Try to handle this request synchronously.  If this has
1935 	 * to block because the task is running, it will just bail
1936 	 * and let the task handle it instead.
1937 	 */
1938 	aio_ddp_requeue(toep);
1939 	DDP_UNLOCK(toep);
1940 	return (0);
1941 }
1942 
1943 int
1944 t4_ddp_mod_load(void)
1945 {
1946 
1947 	t4_register_cpl_handler(CPL_RX_DATA_DDP, do_rx_data_ddp);
1948 	t4_register_cpl_handler(CPL_RX_DDP_COMPLETE, do_rx_ddp_complete);
1949 	TAILQ_INIT(&ddp_orphan_pagesets);
1950 	mtx_init(&ddp_orphan_pagesets_lock, "ddp orphans", NULL, MTX_DEF);
1951 	TASK_INIT(&ddp_orphan_task, 0, ddp_free_orphan_pagesets, NULL);
1952 	return (0);
1953 }
1954 
1955 void
1956 t4_ddp_mod_unload(void)
1957 {
1958 
1959 	taskqueue_drain(taskqueue_thread, &ddp_orphan_task);
1960 	MPASS(TAILQ_EMPTY(&ddp_orphan_pagesets));
1961 	mtx_destroy(&ddp_orphan_pagesets_lock);
1962 	t4_register_cpl_handler(CPL_RX_DATA_DDP, NULL);
1963 	t4_register_cpl_handler(CPL_RX_DDP_COMPLETE, NULL);
1964 }
1965 #endif
1966