xref: /freebsd/sys/dev/cxgbe/tom/t4_ddp.c (revision c4e127e24dc9f1322ebe7ade0991de7022010bf1)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2012 Chelsio Communications, Inc.
5  * All rights reserved.
6  * Written by: Navdeep Parhar <np@FreeBSD.org>
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include "opt_inet.h"
34 
35 #include <sys/param.h>
36 #include <sys/aio.h>
37 #include <sys/file.h>
38 #include <sys/systm.h>
39 #include <sys/kernel.h>
40 #include <sys/ktr.h>
41 #include <sys/module.h>
42 #include <sys/protosw.h>
43 #include <sys/proc.h>
44 #include <sys/domain.h>
45 #include <sys/socket.h>
46 #include <sys/socketvar.h>
47 #include <sys/taskqueue.h>
48 #include <sys/uio.h>
49 #include <netinet/in.h>
50 #include <netinet/in_pcb.h>
51 #include <netinet/ip.h>
52 #include <netinet/tcp_var.h>
53 #define TCPSTATES
54 #include <netinet/tcp_fsm.h>
55 #include <netinet/toecore.h>
56 
57 #include <vm/vm.h>
58 #include <vm/vm_extern.h>
59 #include <vm/vm_param.h>
60 #include <vm/pmap.h>
61 #include <vm/vm_map.h>
62 #include <vm/vm_page.h>
63 #include <vm/vm_object.h>
64 
65 #ifdef TCP_OFFLOAD
66 #include "common/common.h"
67 #include "common/t4_msg.h"
68 #include "common/t4_regs.h"
69 #include "common/t4_tcb.h"
70 #include "tom/t4_tom.h"
71 
72 /*
73  * Use the 'backend3' field in AIO jobs to store the amount of data
74  * received by the AIO job so far.
75  */
76 #define	aio_received	backend3
77 
78 static void aio_ddp_requeue_task(void *context, int pending);
79 static void ddp_complete_all(struct toepcb *toep, int error);
80 static void t4_aio_cancel_active(struct kaiocb *job);
81 static void t4_aio_cancel_queued(struct kaiocb *job);
82 
83 static TAILQ_HEAD(, pageset) ddp_orphan_pagesets;
84 static struct mtx ddp_orphan_pagesets_lock;
85 static struct task ddp_orphan_task;
86 
87 #define MAX_DDP_BUFFER_SIZE		(M_TCB_RX_DDP_BUF0_LEN)
88 
89 /*
90  * A page set holds information about a buffer used for DDP.  The page
91  * set holds resources such as the VM pages backing the buffer (either
92  * held or wired) and the page pods associated with the buffer.
93  * Recently used page sets are cached to allow for efficient reuse of
94  * buffers (avoiding the need to re-fault in pages, hold them, etc.).
95  * Note that cached page sets keep the backing pages wired.  The
96  * number of wired pages is capped by only allowing for two wired
97  * pagesets per connection.  This is not a perfect cap, but is a
98  * trade-off for performance.
99  *
100  * If an application ping-pongs two buffers for a connection via
101  * aio_read(2) then those buffers should remain wired and expensive VM
102  * fault lookups should be avoided after each buffer has been used
103  * once.  If an application uses more than two buffers then this will
104  * fall back to doing expensive VM fault lookups for each operation.
105  */
106 static void
107 free_pageset(struct tom_data *td, struct pageset *ps)
108 {
109 	vm_page_t p;
110 	int i;
111 
112 	if (ps->prsv.prsv_nppods > 0)
113 		t4_free_page_pods(&ps->prsv);
114 
115 	if (ps->flags & PS_WIRED) {
116 		for (i = 0; i < ps->npages; i++) {
117 			p = ps->pages[i];
118 			vm_page_lock(p);
119 			vm_page_unwire(p, PQ_INACTIVE);
120 			vm_page_unlock(p);
121 		}
122 	} else
123 		vm_page_unhold_pages(ps->pages, ps->npages);
124 	mtx_lock(&ddp_orphan_pagesets_lock);
125 	TAILQ_INSERT_TAIL(&ddp_orphan_pagesets, ps, link);
126 	taskqueue_enqueue(taskqueue_thread, &ddp_orphan_task);
127 	mtx_unlock(&ddp_orphan_pagesets_lock);
128 }
129 
130 static void
131 ddp_free_orphan_pagesets(void *context, int pending)
132 {
133 	struct pageset *ps;
134 
135 	mtx_lock(&ddp_orphan_pagesets_lock);
136 	while (!TAILQ_EMPTY(&ddp_orphan_pagesets)) {
137 		ps = TAILQ_FIRST(&ddp_orphan_pagesets);
138 		TAILQ_REMOVE(&ddp_orphan_pagesets, ps, link);
139 		mtx_unlock(&ddp_orphan_pagesets_lock);
140 		if (ps->vm)
141 			vmspace_free(ps->vm);
142 		free(ps, M_CXGBE);
143 		mtx_lock(&ddp_orphan_pagesets_lock);
144 	}
145 	mtx_unlock(&ddp_orphan_pagesets_lock);
146 }
147 
148 static void
149 recycle_pageset(struct toepcb *toep, struct pageset *ps)
150 {
151 
152 	DDP_ASSERT_LOCKED(toep);
153 	if (!(toep->ddp.flags & DDP_DEAD) && ps->flags & PS_WIRED) {
154 		KASSERT(toep->ddp.cached_count + toep->ddp.active_count <
155 		    nitems(toep->ddp.db), ("too many wired pagesets"));
156 		TAILQ_INSERT_HEAD(&toep->ddp.cached_pagesets, ps, link);
157 		toep->ddp.cached_count++;
158 	} else
159 		free_pageset(toep->td, ps);
160 }
161 
162 static void
163 ddp_complete_one(struct kaiocb *job, int error)
164 {
165 	long copied;
166 
167 	/*
168 	 * If this job had copied data out of the socket buffer before
169 	 * it was cancelled, report it as a short read rather than an
170 	 * error.
171 	 */
172 	copied = job->aio_received;
173 	if (copied != 0 || error == 0)
174 		aio_complete(job, copied, 0);
175 	else
176 		aio_complete(job, -1, error);
177 }
178 
179 static void
180 free_ddp_buffer(struct tom_data *td, struct ddp_buffer *db)
181 {
182 
183 	if (db->job) {
184 		/*
185 		 * XXX: If we are un-offloading the socket then we
186 		 * should requeue these on the socket somehow.  If we
187 		 * got a FIN from the remote end, then this completes
188 		 * any remaining requests with an EOF read.
189 		 */
190 		if (!aio_clear_cancel_function(db->job))
191 			ddp_complete_one(db->job, 0);
192 	}
193 
194 	if (db->ps)
195 		free_pageset(td, db->ps);
196 }
197 
198 void
199 ddp_init_toep(struct toepcb *toep)
200 {
201 
202 	TAILQ_INIT(&toep->ddp.aiojobq);
203 	TASK_INIT(&toep->ddp.requeue_task, 0, aio_ddp_requeue_task, toep);
204 	toep->ddp.flags = DDP_OK;
205 	toep->ddp.active_id = -1;
206 	mtx_init(&toep->ddp.lock, "t4 ddp", NULL, MTX_DEF);
207 }
208 
209 void
210 ddp_uninit_toep(struct toepcb *toep)
211 {
212 
213 	mtx_destroy(&toep->ddp.lock);
214 }
215 
216 void
217 release_ddp_resources(struct toepcb *toep)
218 {
219 	struct pageset *ps;
220 	int i;
221 
222 	DDP_LOCK(toep);
223 	toep->ddp.flags |= DDP_DEAD;
224 	for (i = 0; i < nitems(toep->ddp.db); i++) {
225 		free_ddp_buffer(toep->td, &toep->ddp.db[i]);
226 	}
227 	while ((ps = TAILQ_FIRST(&toep->ddp.cached_pagesets)) != NULL) {
228 		TAILQ_REMOVE(&toep->ddp.cached_pagesets, ps, link);
229 		free_pageset(toep->td, ps);
230 	}
231 	ddp_complete_all(toep, 0);
232 	DDP_UNLOCK(toep);
233 }
234 
235 #ifdef INVARIANTS
236 void
237 ddp_assert_empty(struct toepcb *toep)
238 {
239 	int i;
240 
241 	MPASS(!(toep->ddp.flags & DDP_TASK_ACTIVE));
242 	for (i = 0; i < nitems(toep->ddp.db); i++) {
243 		MPASS(toep->ddp.db[i].job == NULL);
244 		MPASS(toep->ddp.db[i].ps == NULL);
245 	}
246 	MPASS(TAILQ_EMPTY(&toep->ddp.cached_pagesets));
247 	MPASS(TAILQ_EMPTY(&toep->ddp.aiojobq));
248 }
249 #endif
250 
251 static void
252 complete_ddp_buffer(struct toepcb *toep, struct ddp_buffer *db,
253     unsigned int db_idx)
254 {
255 	unsigned int db_flag;
256 
257 	toep->ddp.active_count--;
258 	if (toep->ddp.active_id == db_idx) {
259 		if (toep->ddp.active_count == 0) {
260 			KASSERT(toep->ddp.db[db_idx ^ 1].job == NULL,
261 			    ("%s: active_count mismatch", __func__));
262 			toep->ddp.active_id = -1;
263 		} else
264 			toep->ddp.active_id ^= 1;
265 #ifdef VERBOSE_TRACES
266 		CTR2(KTR_CXGBE, "%s: ddp_active_id = %d", __func__,
267 		    toep->ddp.active_id);
268 #endif
269 	} else {
270 		KASSERT(toep->ddp.active_count != 0 &&
271 		    toep->ddp.active_id != -1,
272 		    ("%s: active count mismatch", __func__));
273 	}
274 
275 	db->cancel_pending = 0;
276 	db->job = NULL;
277 	recycle_pageset(toep, db->ps);
278 	db->ps = NULL;
279 
280 	db_flag = db_idx == 1 ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE;
281 	KASSERT(toep->ddp.flags & db_flag,
282 	    ("%s: DDP buffer not active. toep %p, ddp_flags 0x%x",
283 	    __func__, toep, toep->ddp.flags));
284 	toep->ddp.flags &= ~db_flag;
285 }
286 
287 /* XXX: handle_ddp_data code duplication */
288 void
289 insert_ddp_data(struct toepcb *toep, uint32_t n)
290 {
291 	struct inpcb *inp = toep->inp;
292 	struct tcpcb *tp = intotcpcb(inp);
293 	struct ddp_buffer *db;
294 	struct kaiocb *job;
295 	size_t placed;
296 	long copied;
297 	unsigned int db_flag, db_idx;
298 
299 	INP_WLOCK_ASSERT(inp);
300 	DDP_ASSERT_LOCKED(toep);
301 
302 	tp->rcv_nxt += n;
303 #ifndef USE_DDP_RX_FLOW_CONTROL
304 	KASSERT(tp->rcv_wnd >= n, ("%s: negative window size", __func__));
305 	tp->rcv_wnd -= n;
306 #endif
307 	CTR2(KTR_CXGBE, "%s: placed %u bytes before falling out of DDP",
308 	    __func__, n);
309 	while (toep->ddp.active_count > 0) {
310 		MPASS(toep->ddp.active_id != -1);
311 		db_idx = toep->ddp.active_id;
312 		db_flag = db_idx == 1 ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE;
313 		MPASS((toep->ddp.flags & db_flag) != 0);
314 		db = &toep->ddp.db[db_idx];
315 		job = db->job;
316 		copied = job->aio_received;
317 		placed = n;
318 		if (placed > job->uaiocb.aio_nbytes - copied)
319 			placed = job->uaiocb.aio_nbytes - copied;
320 		if (placed > 0)
321 			job->msgrcv = 1;
322 		if (!aio_clear_cancel_function(job)) {
323 			/*
324 			 * Update the copied length for when
325 			 * t4_aio_cancel_active() completes this
326 			 * request.
327 			 */
328 			job->aio_received += placed;
329 		} else if (copied + placed != 0) {
330 			CTR4(KTR_CXGBE,
331 			    "%s: completing %p (copied %ld, placed %lu)",
332 			    __func__, job, copied, placed);
333 			/* XXX: This always completes if there is some data. */
334 			aio_complete(job, copied + placed, 0);
335 		} else if (aio_set_cancel_function(job, t4_aio_cancel_queued)) {
336 			TAILQ_INSERT_HEAD(&toep->ddp.aiojobq, job, list);
337 			toep->ddp.waiting_count++;
338 		} else
339 			aio_cancel(job);
340 		n -= placed;
341 		complete_ddp_buffer(toep, db, db_idx);
342 	}
343 
344 	MPASS(n == 0);
345 }
346 
347 /* SET_TCB_FIELD sent as a ULP command looks like this */
348 #define LEN__SET_TCB_FIELD_ULP (sizeof(struct ulp_txpkt) + \
349     sizeof(struct ulptx_idata) + sizeof(struct cpl_set_tcb_field_core))
350 
351 /* RX_DATA_ACK sent as a ULP command looks like this */
352 #define LEN__RX_DATA_ACK_ULP (sizeof(struct ulp_txpkt) + \
353     sizeof(struct ulptx_idata) + sizeof(struct cpl_rx_data_ack_core))
354 
355 static inline void *
356 mk_set_tcb_field_ulp(struct ulp_txpkt *ulpmc, struct toepcb *toep,
357     uint64_t word, uint64_t mask, uint64_t val)
358 {
359 	struct ulptx_idata *ulpsc;
360 	struct cpl_set_tcb_field_core *req;
361 
362 	ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0));
363 	ulpmc->len = htobe32(howmany(LEN__SET_TCB_FIELD_ULP, 16));
364 
365 	ulpsc = (struct ulptx_idata *)(ulpmc + 1);
366 	ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
367 	ulpsc->len = htobe32(sizeof(*req));
368 
369 	req = (struct cpl_set_tcb_field_core *)(ulpsc + 1);
370 	OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_SET_TCB_FIELD, toep->tid));
371 	req->reply_ctrl = htobe16(V_NO_REPLY(1) |
372 	    V_QUEUENO(toep->ofld_rxq->iq.abs_id));
373 	req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0));
374         req->mask = htobe64(mask);
375         req->val = htobe64(val);
376 
377 	ulpsc = (struct ulptx_idata *)(req + 1);
378 	if (LEN__SET_TCB_FIELD_ULP % 16) {
379 		ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP));
380 		ulpsc->len = htobe32(0);
381 		return (ulpsc + 1);
382 	}
383 	return (ulpsc);
384 }
385 
386 static inline void *
387 mk_rx_data_ack_ulp(struct ulp_txpkt *ulpmc, struct toepcb *toep)
388 {
389 	struct ulptx_idata *ulpsc;
390 	struct cpl_rx_data_ack_core *req;
391 
392 	ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0));
393 	ulpmc->len = htobe32(howmany(LEN__RX_DATA_ACK_ULP, 16));
394 
395 	ulpsc = (struct ulptx_idata *)(ulpmc + 1);
396 	ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
397 	ulpsc->len = htobe32(sizeof(*req));
398 
399 	req = (struct cpl_rx_data_ack_core *)(ulpsc + 1);
400 	OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_RX_DATA_ACK, toep->tid));
401 	req->credit_dack = htobe32(F_RX_MODULATE_RX);
402 
403 	ulpsc = (struct ulptx_idata *)(req + 1);
404 	if (LEN__RX_DATA_ACK_ULP % 16) {
405 		ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP));
406 		ulpsc->len = htobe32(0);
407 		return (ulpsc + 1);
408 	}
409 	return (ulpsc);
410 }
411 
412 static struct wrqe *
413 mk_update_tcb_for_ddp(struct adapter *sc, struct toepcb *toep, int db_idx,
414     struct pageset *ps, int offset, uint64_t ddp_flags, uint64_t ddp_flags_mask)
415 {
416 	struct wrqe *wr;
417 	struct work_request_hdr *wrh;
418 	struct ulp_txpkt *ulpmc;
419 	int len;
420 
421 	KASSERT(db_idx == 0 || db_idx == 1,
422 	    ("%s: bad DDP buffer index %d", __func__, db_idx));
423 
424 	/*
425 	 * We'll send a compound work request that has 3 SET_TCB_FIELDs and an
426 	 * RX_DATA_ACK (with RX_MODULATE to speed up delivery).
427 	 *
428 	 * The work request header is 16B and always ends at a 16B boundary.
429 	 * The ULPTX master commands that follow must all end at 16B boundaries
430 	 * too so we round up the size to 16.
431 	 */
432 	len = sizeof(*wrh) + 3 * roundup2(LEN__SET_TCB_FIELD_ULP, 16) +
433 	    roundup2(LEN__RX_DATA_ACK_ULP, 16);
434 
435 	wr = alloc_wrqe(len, toep->ctrlq);
436 	if (wr == NULL)
437 		return (NULL);
438 	wrh = wrtod(wr);
439 	INIT_ULPTX_WRH(wrh, len, 1, 0);	/* atomic */
440 	ulpmc = (struct ulp_txpkt *)(wrh + 1);
441 
442 	/* Write the buffer's tag */
443 	ulpmc = mk_set_tcb_field_ulp(ulpmc, toep,
444 	    W_TCB_RX_DDP_BUF0_TAG + db_idx,
445 	    V_TCB_RX_DDP_BUF0_TAG(M_TCB_RX_DDP_BUF0_TAG),
446 	    V_TCB_RX_DDP_BUF0_TAG(ps->prsv.prsv_tag));
447 
448 	/* Update the current offset in the DDP buffer and its total length */
449 	if (db_idx == 0)
450 		ulpmc = mk_set_tcb_field_ulp(ulpmc, toep,
451 		    W_TCB_RX_DDP_BUF0_OFFSET,
452 		    V_TCB_RX_DDP_BUF0_OFFSET(M_TCB_RX_DDP_BUF0_OFFSET) |
453 		    V_TCB_RX_DDP_BUF0_LEN(M_TCB_RX_DDP_BUF0_LEN),
454 		    V_TCB_RX_DDP_BUF0_OFFSET(offset) |
455 		    V_TCB_RX_DDP_BUF0_LEN(ps->len));
456 	else
457 		ulpmc = mk_set_tcb_field_ulp(ulpmc, toep,
458 		    W_TCB_RX_DDP_BUF1_OFFSET,
459 		    V_TCB_RX_DDP_BUF1_OFFSET(M_TCB_RX_DDP_BUF1_OFFSET) |
460 		    V_TCB_RX_DDP_BUF1_LEN((u64)M_TCB_RX_DDP_BUF1_LEN << 32),
461 		    V_TCB_RX_DDP_BUF1_OFFSET(offset) |
462 		    V_TCB_RX_DDP_BUF1_LEN((u64)ps->len << 32));
463 
464 	/* Update DDP flags */
465 	ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, W_TCB_RX_DDP_FLAGS,
466 	    ddp_flags_mask, ddp_flags);
467 
468 	/* Gratuitous RX_DATA_ACK with RX_MODULATE set to speed up delivery. */
469 	ulpmc = mk_rx_data_ack_ulp(ulpmc, toep);
470 
471 	return (wr);
472 }
473 
474 static int
475 handle_ddp_data(struct toepcb *toep, __be32 ddp_report, __be32 rcv_nxt, int len)
476 {
477 	uint32_t report = be32toh(ddp_report);
478 	unsigned int db_idx;
479 	struct inpcb *inp = toep->inp;
480 	struct ddp_buffer *db;
481 	struct tcpcb *tp;
482 	struct socket *so;
483 	struct sockbuf *sb;
484 	struct kaiocb *job;
485 	long copied;
486 
487 	db_idx = report & F_DDP_BUF_IDX ? 1 : 0;
488 
489 	if (__predict_false(!(report & F_DDP_INV)))
490 		CXGBE_UNIMPLEMENTED("DDP buffer still valid");
491 
492 	INP_WLOCK(inp);
493 	so = inp_inpcbtosocket(inp);
494 	sb = &so->so_rcv;
495 	DDP_LOCK(toep);
496 
497 	KASSERT(toep->ddp.active_id == db_idx,
498 	    ("completed DDP buffer (%d) != active_id (%d) for tid %d", db_idx,
499 	    toep->ddp.active_id, toep->tid));
500 	db = &toep->ddp.db[db_idx];
501 	job = db->job;
502 
503 	if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))) {
504 		/*
505 		 * This can happen due to an administrative tcpdrop(8).
506 		 * Just fail the request with ECONNRESET.
507 		 */
508 		CTR5(KTR_CXGBE, "%s: tid %u, seq 0x%x, len %d, inp_flags 0x%x",
509 		    __func__, toep->tid, be32toh(rcv_nxt), len, inp->inp_flags);
510 		if (aio_clear_cancel_function(job))
511 			ddp_complete_one(job, ECONNRESET);
512 		goto completed;
513 	}
514 
515 	tp = intotcpcb(inp);
516 
517 	/*
518 	 * For RX_DDP_COMPLETE, len will be zero and rcv_nxt is the
519 	 * sequence number of the next byte to receive.  The length of
520 	 * the data received for this message must be computed by
521 	 * comparing the new and old values of rcv_nxt.
522 	 *
523 	 * For RX_DATA_DDP, len might be non-zero, but it is only the
524 	 * length of the most recent DMA.  It does not include the
525 	 * total length of the data received since the previous update
526 	 * for this DDP buffer.  rcv_nxt is the sequence number of the
527 	 * first received byte from the most recent DMA.
528 	 */
529 	len += be32toh(rcv_nxt) - tp->rcv_nxt;
530 	tp->rcv_nxt += len;
531 	tp->t_rcvtime = ticks;
532 #ifndef USE_DDP_RX_FLOW_CONTROL
533 	KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__));
534 	tp->rcv_wnd -= len;
535 #endif
536 #ifdef VERBOSE_TRACES
537 	CTR4(KTR_CXGBE, "%s: DDP[%d] placed %d bytes (%#x)", __func__, db_idx,
538 	    len, report);
539 #endif
540 
541 	/* receive buffer autosize */
542 	MPASS(toep->vnet == so->so_vnet);
543 	CURVNET_SET(toep->vnet);
544 	SOCKBUF_LOCK(sb);
545 	if (sb->sb_flags & SB_AUTOSIZE &&
546 	    V_tcp_do_autorcvbuf &&
547 	    sb->sb_hiwat < V_tcp_autorcvbuf_max &&
548 	    len > (sbspace(sb) / 8 * 7)) {
549 		struct adapter *sc = td_adapter(toep->td);
550 		unsigned int hiwat = sb->sb_hiwat;
551 		unsigned int newsize = min(hiwat + sc->tt.autorcvbuf_inc,
552 		    V_tcp_autorcvbuf_max);
553 
554 		if (!sbreserve_locked(sb, newsize, so, NULL))
555 			sb->sb_flags &= ~SB_AUTOSIZE;
556 	}
557 	SOCKBUF_UNLOCK(sb);
558 	CURVNET_RESTORE();
559 
560 	job->msgrcv = 1;
561 	if (db->cancel_pending) {
562 		/*
563 		 * Update the job's length but defer completion to the
564 		 * TCB_RPL callback.
565 		 */
566 		job->aio_received += len;
567 		goto out;
568 	} else if (!aio_clear_cancel_function(job)) {
569 		/*
570 		 * Update the copied length for when
571 		 * t4_aio_cancel_active() completes this request.
572 		 */
573 		job->aio_received += len;
574 	} else {
575 		copied = job->aio_received;
576 #ifdef VERBOSE_TRACES
577 		CTR4(KTR_CXGBE, "%s: completing %p (copied %ld, placed %d)",
578 		    __func__, job, copied, len);
579 #endif
580 		aio_complete(job, copied + len, 0);
581 		t4_rcvd(&toep->td->tod, tp);
582 	}
583 
584 completed:
585 	complete_ddp_buffer(toep, db, db_idx);
586 	if (toep->ddp.waiting_count > 0)
587 		ddp_queue_toep(toep);
588 out:
589 	DDP_UNLOCK(toep);
590 	INP_WUNLOCK(inp);
591 
592 	return (0);
593 }
594 
595 void
596 handle_ddp_indicate(struct toepcb *toep)
597 {
598 
599 	DDP_ASSERT_LOCKED(toep);
600 	MPASS(toep->ddp.active_count == 0);
601 	MPASS((toep->ddp.flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE)) == 0);
602 	if (toep->ddp.waiting_count == 0) {
603 		/*
604 		 * The pending requests that triggered the request for an
605 		 * an indicate were cancelled.  Those cancels should have
606 		 * already disabled DDP.  Just ignore this as the data is
607 		 * going into the socket buffer anyway.
608 		 */
609 		return;
610 	}
611 	CTR3(KTR_CXGBE, "%s: tid %d indicated (%d waiting)", __func__,
612 	    toep->tid, toep->ddp.waiting_count);
613 	ddp_queue_toep(toep);
614 }
615 
616 enum {
617 	DDP_BUF0_INVALIDATED = 0x2,
618 	DDP_BUF1_INVALIDATED
619 };
620 
621 CTASSERT(DDP_BUF0_INVALIDATED == CPL_COOKIE_DDP0);
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(DDP_BUF0_INVALIDATED):
643 	case V_WORD(W_TCB_RX_DDP_FLAGS) | V_COOKIE(DDP_BUF1_INVALIDATED):
644 		/*
645 		 * XXX: This duplicates a lot of code with handle_ddp_data().
646 		 */
647 		db_idx = G_COOKIE(cpl->cookie) - DDP_BUF0_INVALIDATED;
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_flag, db_idx;
704 	int len, placed;
705 
706 	INP_WLOCK_ASSERT(toep->inp);
707 	DDP_ASSERT_LOCKED(toep);
708 
709 	len = be32toh(rcv_nxt) - tp->rcv_nxt;
710 	tp->rcv_nxt += len;
711 
712 	while (toep->ddp.active_count > 0) {
713 		MPASS(toep->ddp.active_id != -1);
714 		db_idx = toep->ddp.active_id;
715 		db_flag = db_idx == 1 ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE;
716 		MPASS((toep->ddp.flags & db_flag) != 0);
717 		db = &toep->ddp.db[db_idx];
718 		job = db->job;
719 		copied = job->aio_received;
720 		placed = len;
721 		if (placed > job->uaiocb.aio_nbytes - copied)
722 			placed = job->uaiocb.aio_nbytes - copied;
723 		if (placed > 0)
724 			job->msgrcv = 1;
725 		if (!aio_clear_cancel_function(job)) {
726 			/*
727 			 * Update the copied length for when
728 			 * t4_aio_cancel_active() completes this
729 			 * request.
730 			 */
731 			job->aio_received += placed;
732 		} else {
733 			CTR4(KTR_CXGBE, "%s: tid %d completed buf %d len %d",
734 			    __func__, toep->tid, db_idx, placed);
735 			aio_complete(job, copied + placed, 0);
736 		}
737 		len -= placed;
738 		complete_ddp_buffer(toep, db, db_idx);
739 	}
740 
741 	MPASS(len == 0);
742 	ddp_complete_all(toep, 0);
743 }
744 
745 #define DDP_ERR (F_DDP_PPOD_MISMATCH | F_DDP_LLIMIT_ERR | F_DDP_ULIMIT_ERR |\
746 	 F_DDP_PPOD_PARITY_ERR | F_DDP_PADDING_ERR | F_DDP_OFFSET_ERR |\
747 	 F_DDP_INVALID_TAG | F_DDP_COLOR_ERR | F_DDP_TID_MISMATCH |\
748 	 F_DDP_INVALID_PPOD | F_DDP_HDRCRC_ERR | F_DDP_DATACRC_ERR)
749 
750 extern cpl_handler_t t4_cpl_handler[];
751 
752 static int
753 do_rx_data_ddp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
754 {
755 	struct adapter *sc = iq->adapter;
756 	const struct cpl_rx_data_ddp *cpl = (const void *)(rss + 1);
757 	unsigned int tid = GET_TID(cpl);
758 	uint32_t vld;
759 	struct toepcb *toep = lookup_tid(sc, tid);
760 
761 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
762 	KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__));
763 	KASSERT(!(toep->flags & TPF_SYNQE),
764 	    ("%s: toep %p claims to be a synq entry", __func__, toep));
765 
766 	vld = be32toh(cpl->ddpvld);
767 	if (__predict_false(vld & DDP_ERR)) {
768 		panic("%s: DDP error 0x%x (tid %d, toep %p)",
769 		    __func__, vld, tid, toep);
770 	}
771 
772 	if (toep->ulp_mode == ULP_MODE_ISCSI) {
773 		t4_cpl_handler[CPL_RX_ISCSI_DDP](iq, rss, m);
774 		return (0);
775 	}
776 
777 	handle_ddp_data(toep, cpl->u.ddp_report, cpl->seq, be16toh(cpl->len));
778 
779 	return (0);
780 }
781 
782 static int
783 do_rx_ddp_complete(struct sge_iq *iq, const struct rss_header *rss,
784     struct mbuf *m)
785 {
786 	struct adapter *sc = iq->adapter;
787 	const struct cpl_rx_ddp_complete *cpl = (const void *)(rss + 1);
788 	unsigned int tid = GET_TID(cpl);
789 	struct toepcb *toep = lookup_tid(sc, tid);
790 
791 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
792 	KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__));
793 	KASSERT(!(toep->flags & TPF_SYNQE),
794 	    ("%s: toep %p claims to be a synq entry", __func__, toep));
795 
796 	handle_ddp_data(toep, cpl->ddp_report, cpl->rcv_nxt, 0);
797 
798 	return (0);
799 }
800 
801 static void
802 enable_ddp(struct adapter *sc, struct toepcb *toep)
803 {
804 
805 	KASSERT((toep->ddp.flags & (DDP_ON | DDP_OK | DDP_SC_REQ)) == DDP_OK,
806 	    ("%s: toep %p has bad ddp_flags 0x%x",
807 	    __func__, toep, toep->ddp.flags));
808 
809 	CTR3(KTR_CXGBE, "%s: tid %u (time %u)",
810 	    __func__, toep->tid, time_uptime);
811 
812 	DDP_ASSERT_LOCKED(toep);
813 	toep->ddp.flags |= DDP_SC_REQ;
814 	t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_RX_DDP_FLAGS,
815 	    V_TF_DDP_OFF(1) | V_TF_DDP_INDICATE_OUT(1) |
816 	    V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1) |
817 	    V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_BUF1_VALID(1),
818 	    V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1), 0, 0);
819 	t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_T_FLAGS,
820 	    V_TF_RCV_COALESCE_ENABLE(1), 0, 0, 0);
821 }
822 
823 static int
824 calculate_hcf(int n1, int n2)
825 {
826 	int a, b, t;
827 
828 	if (n1 <= n2) {
829 		a = n1;
830 		b = n2;
831 	} else {
832 		a = n2;
833 		b = n1;
834 	}
835 
836 	while (a != 0) {
837 		t = a;
838 		a = b % a;
839 		b = t;
840 	}
841 
842 	return (b);
843 }
844 
845 static inline int
846 pages_to_nppods(int npages, int ddp_page_shift)
847 {
848 
849 	MPASS(ddp_page_shift >= PAGE_SHIFT);
850 
851 	return (howmany(npages >> (ddp_page_shift - PAGE_SHIFT), PPOD_PAGES));
852 }
853 
854 static int
855 alloc_page_pods(struct ppod_region *pr, u_int nppods, u_int pgsz_idx,
856     struct ppod_reservation *prsv)
857 {
858 	vmem_addr_t addr;       /* relative to start of region */
859 
860 	if (vmem_alloc(pr->pr_arena, PPOD_SZ(nppods), M_NOWAIT | M_FIRSTFIT,
861 	    &addr) != 0)
862 		return (ENOMEM);
863 
864 	CTR5(KTR_CXGBE, "%-17s arena %p, addr 0x%08x, nppods %d, pgsz %d",
865 	    __func__, pr->pr_arena, (uint32_t)addr & pr->pr_tag_mask,
866 	    nppods, 1 << pr->pr_page_shift[pgsz_idx]);
867 
868 	/*
869 	 * The hardware tagmask includes an extra invalid bit but the arena was
870 	 * seeded with valid values only.  An allocation out of this arena will
871 	 * fit inside the tagmask but won't have the invalid bit set.
872 	 */
873 	MPASS((addr & pr->pr_tag_mask) == addr);
874 	MPASS((addr & pr->pr_invalid_bit) == 0);
875 
876 	prsv->prsv_pr = pr;
877 	prsv->prsv_tag = V_PPOD_PGSZ(pgsz_idx) | addr;
878 	prsv->prsv_nppods = nppods;
879 
880 	return (0);
881 }
882 
883 int
884 t4_alloc_page_pods_for_ps(struct ppod_region *pr, struct pageset *ps)
885 {
886 	int i, hcf, seglen, idx, nppods;
887 	struct ppod_reservation *prsv = &ps->prsv;
888 
889 	KASSERT(prsv->prsv_nppods == 0,
890 	    ("%s: page pods already allocated", __func__));
891 
892 	/*
893 	 * The DDP page size is unrelated to the VM page size.  We combine
894 	 * contiguous physical pages into larger segments to get the best DDP
895 	 * page size possible.  This is the largest of the four sizes in
896 	 * A_ULP_RX_TDDP_PSZ that evenly divides the HCF of the segment sizes in
897 	 * the page list.
898 	 */
899 	hcf = 0;
900 	for (i = 0; i < ps->npages; i++) {
901 		seglen = PAGE_SIZE;
902 		while (i < ps->npages - 1 &&
903 		    ps->pages[i]->phys_addr + PAGE_SIZE ==
904 		    ps->pages[i + 1]->phys_addr) {
905 			seglen += PAGE_SIZE;
906 			i++;
907 		}
908 
909 		hcf = calculate_hcf(hcf, seglen);
910 		if (hcf < (1 << pr->pr_page_shift[1])) {
911 			idx = 0;
912 			goto have_pgsz;	/* give up, short circuit */
913 		}
914 	}
915 
916 #define PR_PAGE_MASK(x) ((1 << pr->pr_page_shift[(x)]) - 1)
917 	MPASS((hcf & PR_PAGE_MASK(0)) == 0); /* PAGE_SIZE is >= 4K everywhere */
918 	for (idx = nitems(pr->pr_page_shift) - 1; idx > 0; idx--) {
919 		if ((hcf & PR_PAGE_MASK(idx)) == 0)
920 			break;
921 	}
922 #undef PR_PAGE_MASK
923 
924 have_pgsz:
925 	MPASS(idx <= M_PPOD_PGSZ);
926 
927 	nppods = pages_to_nppods(ps->npages, pr->pr_page_shift[idx]);
928 	if (alloc_page_pods(pr, nppods, idx, prsv) != 0)
929 		return (0);
930 	MPASS(prsv->prsv_nppods > 0);
931 
932 	return (1);
933 }
934 
935 int
936 t4_alloc_page_pods_for_buf(struct ppod_region *pr, vm_offset_t buf, int len,
937     struct ppod_reservation *prsv)
938 {
939 	int hcf, seglen, idx, npages, nppods;
940 	uintptr_t start_pva, end_pva, pva, p1;
941 
942 	MPASS(buf > 0);
943 	MPASS(len > 0);
944 
945 	/*
946 	 * The DDP page size is unrelated to the VM page size.  We combine
947 	 * contiguous physical pages into larger segments to get the best DDP
948 	 * page size possible.  This is the largest of the four sizes in
949 	 * A_ULP_RX_ISCSI_PSZ that evenly divides the HCF of the segment sizes
950 	 * in the page list.
951 	 */
952 	hcf = 0;
953 	start_pva = trunc_page(buf);
954 	end_pva = trunc_page(buf + len - 1);
955 	pva = start_pva;
956 	while (pva <= end_pva) {
957 		seglen = PAGE_SIZE;
958 		p1 = pmap_kextract(pva);
959 		pva += PAGE_SIZE;
960 		while (pva <= end_pva && p1 + seglen == pmap_kextract(pva)) {
961 			seglen += PAGE_SIZE;
962 			pva += PAGE_SIZE;
963 		}
964 
965 		hcf = calculate_hcf(hcf, seglen);
966 		if (hcf < (1 << pr->pr_page_shift[1])) {
967 			idx = 0;
968 			goto have_pgsz;	/* give up, short circuit */
969 		}
970 	}
971 
972 #define PR_PAGE_MASK(x) ((1 << pr->pr_page_shift[(x)]) - 1)
973 	MPASS((hcf & PR_PAGE_MASK(0)) == 0); /* PAGE_SIZE is >= 4K everywhere */
974 	for (idx = nitems(pr->pr_page_shift) - 1; idx > 0; idx--) {
975 		if ((hcf & PR_PAGE_MASK(idx)) == 0)
976 			break;
977 	}
978 #undef PR_PAGE_MASK
979 
980 have_pgsz:
981 	MPASS(idx <= M_PPOD_PGSZ);
982 
983 	npages = 1;
984 	npages += (end_pva - start_pva) >> pr->pr_page_shift[idx];
985 	nppods = howmany(npages, PPOD_PAGES);
986 	if (alloc_page_pods(pr, nppods, idx, prsv) != 0)
987 		return (ENOMEM);
988 	MPASS(prsv->prsv_nppods > 0);
989 
990 	return (0);
991 }
992 
993 void
994 t4_free_page_pods(struct ppod_reservation *prsv)
995 {
996 	struct ppod_region *pr = prsv->prsv_pr;
997 	vmem_addr_t addr;
998 
999 	MPASS(prsv != NULL);
1000 	MPASS(prsv->prsv_nppods != 0);
1001 
1002 	addr = prsv->prsv_tag & pr->pr_tag_mask;
1003 	MPASS((addr & pr->pr_invalid_bit) == 0);
1004 
1005 	CTR4(KTR_CXGBE, "%-17s arena %p, addr 0x%08x, nppods %d", __func__,
1006 	    pr->pr_arena, addr, prsv->prsv_nppods);
1007 
1008 	vmem_free(pr->pr_arena, addr, PPOD_SZ(prsv->prsv_nppods));
1009 	prsv->prsv_nppods = 0;
1010 }
1011 
1012 #define NUM_ULP_TX_SC_IMM_PPODS (256 / PPOD_SIZE)
1013 
1014 int
1015 t4_write_page_pods_for_ps(struct adapter *sc, struct sge_wrq *wrq, int tid,
1016     struct pageset *ps)
1017 {
1018 	struct wrqe *wr;
1019 	struct ulp_mem_io *ulpmc;
1020 	struct ulptx_idata *ulpsc;
1021 	struct pagepod *ppod;
1022 	int i, j, k, n, chunk, len, ddp_pgsz, idx;
1023 	u_int ppod_addr;
1024 	uint32_t cmd;
1025 	struct ppod_reservation *prsv = &ps->prsv;
1026 	struct ppod_region *pr = prsv->prsv_pr;
1027 
1028 	KASSERT(!(ps->flags & PS_PPODS_WRITTEN),
1029 	    ("%s: page pods already written", __func__));
1030 	MPASS(prsv->prsv_nppods > 0);
1031 
1032 	cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE));
1033 	if (is_t4(sc))
1034 		cmd |= htobe32(F_ULP_MEMIO_ORDER);
1035 	else
1036 		cmd |= htobe32(F_T5_ULP_MEMIO_IMM);
1037 	ddp_pgsz = 1 << pr->pr_page_shift[G_PPOD_PGSZ(prsv->prsv_tag)];
1038 	ppod_addr = pr->pr_start + (prsv->prsv_tag & pr->pr_tag_mask);
1039 	for (i = 0; i < prsv->prsv_nppods; ppod_addr += chunk) {
1040 
1041 		/* How many page pods are we writing in this cycle */
1042 		n = min(prsv->prsv_nppods - i, NUM_ULP_TX_SC_IMM_PPODS);
1043 		chunk = PPOD_SZ(n);
1044 		len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16);
1045 
1046 		wr = alloc_wrqe(len, wrq);
1047 		if (wr == NULL)
1048 			return (ENOMEM);	/* ok to just bail out */
1049 		ulpmc = wrtod(wr);
1050 
1051 		INIT_ULPTX_WR(ulpmc, len, 0, 0);
1052 		ulpmc->cmd = cmd;
1053 		ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32));
1054 		ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16));
1055 		ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5));
1056 
1057 		ulpsc = (struct ulptx_idata *)(ulpmc + 1);
1058 		ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
1059 		ulpsc->len = htobe32(chunk);
1060 
1061 		ppod = (struct pagepod *)(ulpsc + 1);
1062 		for (j = 0; j < n; i++, j++, ppod++) {
1063 			ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID |
1064 			    V_PPOD_TID(tid) | prsv->prsv_tag);
1065 			ppod->len_offset = htobe64(V_PPOD_LEN(ps->len) |
1066 			    V_PPOD_OFST(ps->offset));
1067 			ppod->rsvd = 0;
1068 			idx = i * PPOD_PAGES * (ddp_pgsz / PAGE_SIZE);
1069 			for (k = 0; k < nitems(ppod->addr); k++) {
1070 				if (idx < ps->npages) {
1071 					ppod->addr[k] =
1072 					    htobe64(ps->pages[idx]->phys_addr);
1073 					idx += ddp_pgsz / PAGE_SIZE;
1074 				} else
1075 					ppod->addr[k] = 0;
1076 #if 0
1077 				CTR5(KTR_CXGBE,
1078 				    "%s: tid %d ppod[%d]->addr[%d] = %p",
1079 				    __func__, toep->tid, i, k,
1080 				    htobe64(ppod->addr[k]));
1081 #endif
1082 			}
1083 
1084 		}
1085 
1086 		t4_wrq_tx(sc, wr);
1087 	}
1088 	ps->flags |= PS_PPODS_WRITTEN;
1089 
1090 	return (0);
1091 }
1092 
1093 int
1094 t4_write_page_pods_for_buf(struct adapter *sc, struct sge_wrq *wrq, int tid,
1095     struct ppod_reservation *prsv, vm_offset_t buf, int buflen)
1096 {
1097 	struct wrqe *wr;
1098 	struct ulp_mem_io *ulpmc;
1099 	struct ulptx_idata *ulpsc;
1100 	struct pagepod *ppod;
1101 	int i, j, k, n, chunk, len, ddp_pgsz;
1102 	u_int ppod_addr, offset;
1103 	uint32_t cmd;
1104 	struct ppod_region *pr = prsv->prsv_pr;
1105 	uintptr_t end_pva, pva, pa;
1106 
1107 	cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE));
1108 	if (is_t4(sc))
1109 		cmd |= htobe32(F_ULP_MEMIO_ORDER);
1110 	else
1111 		cmd |= htobe32(F_T5_ULP_MEMIO_IMM);
1112 	ddp_pgsz = 1 << pr->pr_page_shift[G_PPOD_PGSZ(prsv->prsv_tag)];
1113 	offset = buf & PAGE_MASK;
1114 	ppod_addr = pr->pr_start + (prsv->prsv_tag & pr->pr_tag_mask);
1115 	pva = trunc_page(buf);
1116 	end_pva = trunc_page(buf + buflen - 1);
1117 	for (i = 0; i < prsv->prsv_nppods; ppod_addr += chunk) {
1118 
1119 		/* How many page pods are we writing in this cycle */
1120 		n = min(prsv->prsv_nppods - i, NUM_ULP_TX_SC_IMM_PPODS);
1121 		MPASS(n > 0);
1122 		chunk = PPOD_SZ(n);
1123 		len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16);
1124 
1125 		wr = alloc_wrqe(len, wrq);
1126 		if (wr == NULL)
1127 			return (ENOMEM);	/* ok to just bail out */
1128 		ulpmc = wrtod(wr);
1129 
1130 		INIT_ULPTX_WR(ulpmc, len, 0, 0);
1131 		ulpmc->cmd = cmd;
1132 		ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32));
1133 		ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16));
1134 		ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5));
1135 
1136 		ulpsc = (struct ulptx_idata *)(ulpmc + 1);
1137 		ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
1138 		ulpsc->len = htobe32(chunk);
1139 
1140 		ppod = (struct pagepod *)(ulpsc + 1);
1141 		for (j = 0; j < n; i++, j++, ppod++) {
1142 			ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID |
1143 			    V_PPOD_TID(tid) |
1144 			    (prsv->prsv_tag & ~V_PPOD_PGSZ(M_PPOD_PGSZ)));
1145 			ppod->len_offset = htobe64(V_PPOD_LEN(buflen) |
1146 			    V_PPOD_OFST(offset));
1147 			ppod->rsvd = 0;
1148 
1149 			for (k = 0; k < nitems(ppod->addr); k++) {
1150 				if (pva > end_pva)
1151 					ppod->addr[k] = 0;
1152 				else {
1153 					pa = pmap_kextract(pva);
1154 					ppod->addr[k] = htobe64(pa);
1155 					pva += ddp_pgsz;
1156 				}
1157 #if 0
1158 				CTR5(KTR_CXGBE,
1159 				    "%s: tid %d ppod[%d]->addr[%d] = %p",
1160 				    __func__, tid, i, k,
1161 				    htobe64(ppod->addr[k]));
1162 #endif
1163 			}
1164 
1165 			/*
1166 			 * Walk back 1 segment so that the first address in the
1167 			 * next pod is the same as the last one in the current
1168 			 * pod.
1169 			 */
1170 			pva -= ddp_pgsz;
1171 		}
1172 
1173 		t4_wrq_tx(sc, wr);
1174 	}
1175 
1176 	MPASS(pva <= end_pva);
1177 
1178 	return (0);
1179 }
1180 
1181 static void
1182 wire_pageset(struct pageset *ps)
1183 {
1184 	vm_page_t p;
1185 	int i;
1186 
1187 	KASSERT(!(ps->flags & PS_WIRED), ("pageset already wired"));
1188 
1189 	for (i = 0; i < ps->npages; i++) {
1190 		p = ps->pages[i];
1191 		vm_page_lock(p);
1192 		vm_page_wire(p);
1193 		vm_page_unhold(p);
1194 		vm_page_unlock(p);
1195 	}
1196 	ps->flags |= PS_WIRED;
1197 }
1198 
1199 /*
1200  * Prepare a pageset for DDP.  This wires the pageset and sets up page
1201  * pods.
1202  */
1203 static int
1204 prep_pageset(struct adapter *sc, struct toepcb *toep, struct pageset *ps)
1205 {
1206 	struct tom_data *td = sc->tom_softc;
1207 
1208 	if (!(ps->flags & PS_WIRED))
1209 		wire_pageset(ps);
1210 	if (ps->prsv.prsv_nppods == 0 &&
1211 	    !t4_alloc_page_pods_for_ps(&td->pr, ps)) {
1212 		return (0);
1213 	}
1214 	if (!(ps->flags & PS_PPODS_WRITTEN) &&
1215 	    t4_write_page_pods_for_ps(sc, toep->ctrlq, toep->tid, ps) != 0) {
1216 		return (0);
1217 	}
1218 
1219 	return (1);
1220 }
1221 
1222 int
1223 t4_init_ppod_region(struct ppod_region *pr, struct t4_range *r, u_int psz,
1224     const char *name)
1225 {
1226 	int i;
1227 
1228 	MPASS(pr != NULL);
1229 	MPASS(r->size > 0);
1230 
1231 	pr->pr_start = r->start;
1232 	pr->pr_len = r->size;
1233 	pr->pr_page_shift[0] = 12 + G_HPZ0(psz);
1234 	pr->pr_page_shift[1] = 12 + G_HPZ1(psz);
1235 	pr->pr_page_shift[2] = 12 + G_HPZ2(psz);
1236 	pr->pr_page_shift[3] = 12 + G_HPZ3(psz);
1237 
1238 	/* The SGL -> page pod algorithm requires the sizes to be in order. */
1239 	for (i = 1; i < nitems(pr->pr_page_shift); i++) {
1240 		if (pr->pr_page_shift[i] <= pr->pr_page_shift[i - 1])
1241 			return (ENXIO);
1242 	}
1243 
1244 	pr->pr_tag_mask = ((1 << fls(r->size)) - 1) & V_PPOD_TAG(M_PPOD_TAG);
1245 	pr->pr_alias_mask = V_PPOD_TAG(M_PPOD_TAG) & ~pr->pr_tag_mask;
1246 	if (pr->pr_tag_mask == 0 || pr->pr_alias_mask == 0)
1247 		return (ENXIO);
1248 	pr->pr_alias_shift = fls(pr->pr_tag_mask);
1249 	pr->pr_invalid_bit = 1 << (pr->pr_alias_shift - 1);
1250 
1251 	pr->pr_arena = vmem_create(name, 0, pr->pr_len, PPOD_SIZE, 0,
1252 	    M_FIRSTFIT | M_NOWAIT);
1253 	if (pr->pr_arena == NULL)
1254 		return (ENOMEM);
1255 
1256 	return (0);
1257 }
1258 
1259 void
1260 t4_free_ppod_region(struct ppod_region *pr)
1261 {
1262 
1263 	MPASS(pr != NULL);
1264 
1265 	if (pr->pr_arena)
1266 		vmem_destroy(pr->pr_arena);
1267 	bzero(pr, sizeof(*pr));
1268 }
1269 
1270 static int
1271 pscmp(struct pageset *ps, struct vmspace *vm, vm_offset_t start, int npages,
1272     int pgoff, int len)
1273 {
1274 
1275 	if (ps->start != start || ps->npages != npages ||
1276 	    ps->offset != pgoff || ps->len != len)
1277 		return (1);
1278 
1279 	return (ps->vm != vm || ps->vm_timestamp != vm->vm_map.timestamp);
1280 }
1281 
1282 static int
1283 hold_aio(struct toepcb *toep, struct kaiocb *job, struct pageset **pps)
1284 {
1285 	struct vmspace *vm;
1286 	vm_map_t map;
1287 	vm_offset_t start, end, pgoff;
1288 	struct pageset *ps;
1289 	int n;
1290 
1291 	DDP_ASSERT_LOCKED(toep);
1292 
1293 	/*
1294 	 * The AIO subsystem will cancel and drain all requests before
1295 	 * permitting a process to exit or exec, so p_vmspace should
1296 	 * be stable here.
1297 	 */
1298 	vm = job->userproc->p_vmspace;
1299 	map = &vm->vm_map;
1300 	start = (uintptr_t)job->uaiocb.aio_buf;
1301 	pgoff = start & PAGE_MASK;
1302 	end = round_page(start + job->uaiocb.aio_nbytes);
1303 	start = trunc_page(start);
1304 
1305 	if (end - start > MAX_DDP_BUFFER_SIZE) {
1306 		/*
1307 		 * Truncate the request to a short read.
1308 		 * Alternatively, we could DDP in chunks to the larger
1309 		 * buffer, but that would be quite a bit more work.
1310 		 *
1311 		 * When truncating, round the request down to avoid
1312 		 * crossing a cache line on the final transaction.
1313 		 */
1314 		end = rounddown2(start + MAX_DDP_BUFFER_SIZE, CACHE_LINE_SIZE);
1315 #ifdef VERBOSE_TRACES
1316 		CTR4(KTR_CXGBE, "%s: tid %d, truncating size from %lu to %lu",
1317 		    __func__, toep->tid, (unsigned long)job->uaiocb.aio_nbytes,
1318 		    (unsigned long)(end - (start + pgoff)));
1319 		job->uaiocb.aio_nbytes = end - (start + pgoff);
1320 #endif
1321 		end = round_page(end);
1322 	}
1323 
1324 	n = atop(end - start);
1325 
1326 	/*
1327 	 * Try to reuse a cached pageset.
1328 	 */
1329 	TAILQ_FOREACH(ps, &toep->ddp.cached_pagesets, link) {
1330 		if (pscmp(ps, vm, start, n, pgoff,
1331 		    job->uaiocb.aio_nbytes) == 0) {
1332 			TAILQ_REMOVE(&toep->ddp.cached_pagesets, ps, link);
1333 			toep->ddp.cached_count--;
1334 			*pps = ps;
1335 			return (0);
1336 		}
1337 	}
1338 
1339 	/*
1340 	 * If there are too many cached pagesets to create a new one,
1341 	 * free a pageset before creating a new one.
1342 	 */
1343 	KASSERT(toep->ddp.active_count + toep->ddp.cached_count <=
1344 	    nitems(toep->ddp.db), ("%s: too many wired pagesets", __func__));
1345 	if (toep->ddp.active_count + toep->ddp.cached_count ==
1346 	    nitems(toep->ddp.db)) {
1347 		KASSERT(toep->ddp.cached_count > 0,
1348 		    ("no cached pageset to free"));
1349 		ps = TAILQ_LAST(&toep->ddp.cached_pagesets, pagesetq);
1350 		TAILQ_REMOVE(&toep->ddp.cached_pagesets, ps, link);
1351 		toep->ddp.cached_count--;
1352 		free_pageset(toep->td, ps);
1353 	}
1354 	DDP_UNLOCK(toep);
1355 
1356 	/* Create a new pageset. */
1357 	ps = malloc(sizeof(*ps) + n * sizeof(vm_page_t), M_CXGBE, M_WAITOK |
1358 	    M_ZERO);
1359 	ps->pages = (vm_page_t *)(ps + 1);
1360 	ps->vm_timestamp = map->timestamp;
1361 	ps->npages = vm_fault_quick_hold_pages(map, start, end - start,
1362 	    VM_PROT_WRITE, ps->pages, n);
1363 
1364 	DDP_LOCK(toep);
1365 	if (ps->npages < 0) {
1366 		free(ps, M_CXGBE);
1367 		return (EFAULT);
1368 	}
1369 
1370 	KASSERT(ps->npages == n, ("hold_aio: page count mismatch: %d vs %d",
1371 	    ps->npages, n));
1372 
1373 	ps->offset = pgoff;
1374 	ps->len = job->uaiocb.aio_nbytes;
1375 	atomic_add_int(&vm->vm_refcnt, 1);
1376 	ps->vm = vm;
1377 	ps->start = start;
1378 
1379 	CTR5(KTR_CXGBE, "%s: tid %d, new pageset %p for job %p, npages %d",
1380 	    __func__, toep->tid, ps, job, ps->npages);
1381 	*pps = ps;
1382 	return (0);
1383 }
1384 
1385 static void
1386 ddp_complete_all(struct toepcb *toep, int error)
1387 {
1388 	struct kaiocb *job;
1389 
1390 	DDP_ASSERT_LOCKED(toep);
1391 	while (!TAILQ_EMPTY(&toep->ddp.aiojobq)) {
1392 		job = TAILQ_FIRST(&toep->ddp.aiojobq);
1393 		TAILQ_REMOVE(&toep->ddp.aiojobq, job, list);
1394 		toep->ddp.waiting_count--;
1395 		if (aio_clear_cancel_function(job))
1396 			ddp_complete_one(job, error);
1397 	}
1398 }
1399 
1400 static void
1401 aio_ddp_cancel_one(struct kaiocb *job)
1402 {
1403 	long copied;
1404 
1405 	/*
1406 	 * If this job had copied data out of the socket buffer before
1407 	 * it was cancelled, report it as a short read rather than an
1408 	 * error.
1409 	 */
1410 	copied = job->aio_received;
1411 	if (copied != 0)
1412 		aio_complete(job, copied, 0);
1413 	else
1414 		aio_cancel(job);
1415 }
1416 
1417 /*
1418  * Called when the main loop wants to requeue a job to retry it later.
1419  * Deals with the race of the job being cancelled while it was being
1420  * examined.
1421  */
1422 static void
1423 aio_ddp_requeue_one(struct toepcb *toep, struct kaiocb *job)
1424 {
1425 
1426 	DDP_ASSERT_LOCKED(toep);
1427 	if (!(toep->ddp.flags & DDP_DEAD) &&
1428 	    aio_set_cancel_function(job, t4_aio_cancel_queued)) {
1429 		TAILQ_INSERT_HEAD(&toep->ddp.aiojobq, job, list);
1430 		toep->ddp.waiting_count++;
1431 	} else
1432 		aio_ddp_cancel_one(job);
1433 }
1434 
1435 static void
1436 aio_ddp_requeue(struct toepcb *toep)
1437 {
1438 	struct adapter *sc = td_adapter(toep->td);
1439 	struct socket *so;
1440 	struct sockbuf *sb;
1441 	struct inpcb *inp;
1442 	struct kaiocb *job;
1443 	struct ddp_buffer *db;
1444 	size_t copied, offset, resid;
1445 	struct pageset *ps;
1446 	struct mbuf *m;
1447 	uint64_t ddp_flags, ddp_flags_mask;
1448 	struct wrqe *wr;
1449 	int buf_flag, db_idx, error;
1450 
1451 	DDP_ASSERT_LOCKED(toep);
1452 
1453 restart:
1454 	if (toep->ddp.flags & DDP_DEAD) {
1455 		MPASS(toep->ddp.waiting_count == 0);
1456 		MPASS(toep->ddp.active_count == 0);
1457 		return;
1458 	}
1459 
1460 	if (toep->ddp.waiting_count == 0 ||
1461 	    toep->ddp.active_count == nitems(toep->ddp.db)) {
1462 		return;
1463 	}
1464 
1465 	job = TAILQ_FIRST(&toep->ddp.aiojobq);
1466 	so = job->fd_file->f_data;
1467 	sb = &so->so_rcv;
1468 	SOCKBUF_LOCK(sb);
1469 
1470 	/* We will never get anything unless we are or were connected. */
1471 	if (!(so->so_state & (SS_ISCONNECTED|SS_ISDISCONNECTED))) {
1472 		SOCKBUF_UNLOCK(sb);
1473 		ddp_complete_all(toep, ENOTCONN);
1474 		return;
1475 	}
1476 
1477 	KASSERT(toep->ddp.active_count == 0 || sbavail(sb) == 0,
1478 	    ("%s: pending sockbuf data and DDP is active", __func__));
1479 
1480 	/* Abort if socket has reported problems. */
1481 	/* XXX: Wait for any queued DDP's to finish and/or flush them? */
1482 	if (so->so_error && sbavail(sb) == 0) {
1483 		toep->ddp.waiting_count--;
1484 		TAILQ_REMOVE(&toep->ddp.aiojobq, job, list);
1485 		if (!aio_clear_cancel_function(job)) {
1486 			SOCKBUF_UNLOCK(sb);
1487 			goto restart;
1488 		}
1489 
1490 		/*
1491 		 * If this job has previously copied some data, report
1492 		 * a short read and leave the error to be reported by
1493 		 * a future request.
1494 		 */
1495 		copied = job->aio_received;
1496 		if (copied != 0) {
1497 			SOCKBUF_UNLOCK(sb);
1498 			aio_complete(job, copied, 0);
1499 			goto restart;
1500 		}
1501 		error = so->so_error;
1502 		so->so_error = 0;
1503 		SOCKBUF_UNLOCK(sb);
1504 		aio_complete(job, -1, error);
1505 		goto restart;
1506 	}
1507 
1508 	/*
1509 	 * Door is closed.  If there is pending data in the socket buffer,
1510 	 * deliver it.  If there are pending DDP requests, wait for those
1511 	 * to complete.  Once they have completed, return EOF reads.
1512 	 */
1513 	if (sb->sb_state & SBS_CANTRCVMORE && sbavail(sb) == 0) {
1514 		SOCKBUF_UNLOCK(sb);
1515 		if (toep->ddp.active_count != 0)
1516 			return;
1517 		ddp_complete_all(toep, 0);
1518 		return;
1519 	}
1520 
1521 	/*
1522 	 * If DDP is not enabled and there is no pending socket buffer
1523 	 * data, try to enable DDP.
1524 	 */
1525 	if (sbavail(sb) == 0 && (toep->ddp.flags & DDP_ON) == 0) {
1526 		SOCKBUF_UNLOCK(sb);
1527 
1528 		/*
1529 		 * Wait for the card to ACK that DDP is enabled before
1530 		 * queueing any buffers.  Currently this waits for an
1531 		 * indicate to arrive.  This could use a TCB_SET_FIELD_RPL
1532 		 * message to know that DDP was enabled instead of waiting
1533 		 * for the indicate which would avoid copying the indicate
1534 		 * if no data is pending.
1535 		 *
1536 		 * XXX: Might want to limit the indicate size to the size
1537 		 * of the first queued request.
1538 		 */
1539 		if ((toep->ddp.flags & DDP_SC_REQ) == 0)
1540 			enable_ddp(sc, toep);
1541 		return;
1542 	}
1543 	SOCKBUF_UNLOCK(sb);
1544 
1545 	/*
1546 	 * If another thread is queueing a buffer for DDP, let it
1547 	 * drain any work and return.
1548 	 */
1549 	if (toep->ddp.queueing != NULL)
1550 		return;
1551 
1552 	/* Take the next job to prep it for DDP. */
1553 	toep->ddp.waiting_count--;
1554 	TAILQ_REMOVE(&toep->ddp.aiojobq, job, list);
1555 	if (!aio_clear_cancel_function(job))
1556 		goto restart;
1557 	toep->ddp.queueing = job;
1558 
1559 	/* NB: This drops DDP_LOCK while it holds the backing VM pages. */
1560 	error = hold_aio(toep, job, &ps);
1561 	if (error != 0) {
1562 		ddp_complete_one(job, error);
1563 		toep->ddp.queueing = NULL;
1564 		goto restart;
1565 	}
1566 
1567 	SOCKBUF_LOCK(sb);
1568 	if (so->so_error && sbavail(sb) == 0) {
1569 		copied = job->aio_received;
1570 		if (copied != 0) {
1571 			SOCKBUF_UNLOCK(sb);
1572 			recycle_pageset(toep, ps);
1573 			aio_complete(job, copied, 0);
1574 			toep->ddp.queueing = NULL;
1575 			goto restart;
1576 		}
1577 
1578 		error = so->so_error;
1579 		so->so_error = 0;
1580 		SOCKBUF_UNLOCK(sb);
1581 		recycle_pageset(toep, ps);
1582 		aio_complete(job, -1, error);
1583 		toep->ddp.queueing = NULL;
1584 		goto restart;
1585 	}
1586 
1587 	if (sb->sb_state & SBS_CANTRCVMORE && sbavail(sb) == 0) {
1588 		SOCKBUF_UNLOCK(sb);
1589 		recycle_pageset(toep, ps);
1590 		if (toep->ddp.active_count != 0) {
1591 			/*
1592 			 * The door is closed, but there are still pending
1593 			 * DDP buffers.  Requeue.  These jobs will all be
1594 			 * completed once those buffers drain.
1595 			 */
1596 			aio_ddp_requeue_one(toep, job);
1597 			toep->ddp.queueing = NULL;
1598 			return;
1599 		}
1600 		ddp_complete_one(job, 0);
1601 		ddp_complete_all(toep, 0);
1602 		toep->ddp.queueing = NULL;
1603 		return;
1604 	}
1605 
1606 sbcopy:
1607 	/*
1608 	 * If the toep is dead, there shouldn't be any data in the socket
1609 	 * buffer, so the above case should have handled this.
1610 	 */
1611 	MPASS(!(toep->ddp.flags & DDP_DEAD));
1612 
1613 	/*
1614 	 * If there is pending data in the socket buffer (either
1615 	 * from before the requests were queued or a DDP indicate),
1616 	 * copy those mbufs out directly.
1617 	 */
1618 	copied = 0;
1619 	offset = ps->offset + job->aio_received;
1620 	MPASS(job->aio_received <= job->uaiocb.aio_nbytes);
1621 	resid = job->uaiocb.aio_nbytes - job->aio_received;
1622 	m = sb->sb_mb;
1623 	KASSERT(m == NULL || toep->ddp.active_count == 0,
1624 	    ("%s: sockbuf data with active DDP", __func__));
1625 	while (m != NULL && resid > 0) {
1626 		struct iovec iov[1];
1627 		struct uio uio;
1628 		int error;
1629 
1630 		iov[0].iov_base = mtod(m, void *);
1631 		iov[0].iov_len = m->m_len;
1632 		if (iov[0].iov_len > resid)
1633 			iov[0].iov_len = resid;
1634 		uio.uio_iov = iov;
1635 		uio.uio_iovcnt = 1;
1636 		uio.uio_offset = 0;
1637 		uio.uio_resid = iov[0].iov_len;
1638 		uio.uio_segflg = UIO_SYSSPACE;
1639 		uio.uio_rw = UIO_WRITE;
1640 		error = uiomove_fromphys(ps->pages, offset + copied,
1641 		    uio.uio_resid, &uio);
1642 		MPASS(error == 0 && uio.uio_resid == 0);
1643 		copied += uio.uio_offset;
1644 		resid -= uio.uio_offset;
1645 		m = m->m_next;
1646 	}
1647 	if (copied != 0) {
1648 		sbdrop_locked(sb, copied);
1649 		job->aio_received += copied;
1650 		job->msgrcv = 1;
1651 		copied = job->aio_received;
1652 		inp = sotoinpcb(so);
1653 		if (!INP_TRY_WLOCK(inp)) {
1654 			/*
1655 			 * The reference on the socket file descriptor in
1656 			 * the AIO job should keep 'sb' and 'inp' stable.
1657 			 * Our caller has a reference on the 'toep' that
1658 			 * keeps it stable.
1659 			 */
1660 			SOCKBUF_UNLOCK(sb);
1661 			DDP_UNLOCK(toep);
1662 			INP_WLOCK(inp);
1663 			DDP_LOCK(toep);
1664 			SOCKBUF_LOCK(sb);
1665 
1666 			/*
1667 			 * If the socket has been closed, we should detect
1668 			 * that and complete this request if needed on
1669 			 * the next trip around the loop.
1670 			 */
1671 		}
1672 		t4_rcvd_locked(&toep->td->tod, intotcpcb(inp));
1673 		INP_WUNLOCK(inp);
1674 		if (resid == 0 || toep->ddp.flags & DDP_DEAD) {
1675 			/*
1676 			 * We filled the entire buffer with socket
1677 			 * data, DDP is not being used, or the socket
1678 			 * is being shut down, so complete the
1679 			 * request.
1680 			 */
1681 			SOCKBUF_UNLOCK(sb);
1682 			recycle_pageset(toep, ps);
1683 			aio_complete(job, copied, 0);
1684 			toep->ddp.queueing = NULL;
1685 			goto restart;
1686 		}
1687 
1688 		/*
1689 		 * If DDP is not enabled, requeue this request and restart.
1690 		 * This will either enable DDP or wait for more data to
1691 		 * arrive on the socket buffer.
1692 		 */
1693 		if ((toep->ddp.flags & (DDP_ON | DDP_SC_REQ)) != DDP_ON) {
1694 			SOCKBUF_UNLOCK(sb);
1695 			recycle_pageset(toep, ps);
1696 			aio_ddp_requeue_one(toep, job);
1697 			toep->ddp.queueing = NULL;
1698 			goto restart;
1699 		}
1700 
1701 		/*
1702 		 * An indicate might have arrived and been added to
1703 		 * the socket buffer while it was unlocked after the
1704 		 * copy to lock the INP.  If so, restart the copy.
1705 		 */
1706 		if (sbavail(sb) != 0)
1707 			goto sbcopy;
1708 	}
1709 	SOCKBUF_UNLOCK(sb);
1710 
1711 	if (prep_pageset(sc, toep, ps) == 0) {
1712 		recycle_pageset(toep, ps);
1713 		aio_ddp_requeue_one(toep, job);
1714 		toep->ddp.queueing = NULL;
1715 
1716 		/*
1717 		 * XXX: Need to retry this later.  Mostly need a trigger
1718 		 * when page pods are freed up.
1719 		 */
1720 		printf("%s: prep_pageset failed\n", __func__);
1721 		return;
1722 	}
1723 
1724 	/* Determine which DDP buffer to use. */
1725 	if (toep->ddp.db[0].job == NULL) {
1726 		db_idx = 0;
1727 	} else {
1728 		MPASS(toep->ddp.db[1].job == NULL);
1729 		db_idx = 1;
1730 	}
1731 
1732 	ddp_flags = 0;
1733 	ddp_flags_mask = 0;
1734 	if (db_idx == 0) {
1735 		ddp_flags |= V_TF_DDP_BUF0_VALID(1);
1736 		if (so->so_state & SS_NBIO)
1737 			ddp_flags |= V_TF_DDP_BUF0_FLUSH(1);
1738 		ddp_flags_mask |= V_TF_DDP_PSH_NO_INVALIDATE0(1) |
1739 		    V_TF_DDP_PUSH_DISABLE_0(1) | V_TF_DDP_PSHF_ENABLE_0(1) |
1740 		    V_TF_DDP_BUF0_FLUSH(1) | V_TF_DDP_BUF0_VALID(1);
1741 		buf_flag = DDP_BUF0_ACTIVE;
1742 	} else {
1743 		ddp_flags |= V_TF_DDP_BUF1_VALID(1);
1744 		if (so->so_state & SS_NBIO)
1745 			ddp_flags |= V_TF_DDP_BUF1_FLUSH(1);
1746 		ddp_flags_mask |= V_TF_DDP_PSH_NO_INVALIDATE1(1) |
1747 		    V_TF_DDP_PUSH_DISABLE_1(1) | V_TF_DDP_PSHF_ENABLE_1(1) |
1748 		    V_TF_DDP_BUF1_FLUSH(1) | V_TF_DDP_BUF1_VALID(1);
1749 		buf_flag = DDP_BUF1_ACTIVE;
1750 	}
1751 	MPASS((toep->ddp.flags & buf_flag) == 0);
1752 	if ((toep->ddp.flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE)) == 0) {
1753 		MPASS(db_idx == 0);
1754 		MPASS(toep->ddp.active_id == -1);
1755 		MPASS(toep->ddp.active_count == 0);
1756 		ddp_flags_mask |= V_TF_DDP_ACTIVE_BUF(1);
1757 	}
1758 
1759 	/*
1760 	 * The TID for this connection should still be valid.  If DDP_DEAD
1761 	 * is set, SBS_CANTRCVMORE should be set, so we shouldn't be
1762 	 * this far anyway.  Even if the socket is closing on the other
1763 	 * end, the AIO job holds a reference on this end of the socket
1764 	 * which will keep it open and keep the TCP PCB attached until
1765 	 * after the job is completed.
1766 	 */
1767 	wr = mk_update_tcb_for_ddp(sc, toep, db_idx, ps, job->aio_received,
1768 	    ddp_flags, ddp_flags_mask);
1769 	if (wr == NULL) {
1770 		recycle_pageset(toep, ps);
1771 		aio_ddp_requeue_one(toep, job);
1772 		toep->ddp.queueing = NULL;
1773 
1774 		/*
1775 		 * XXX: Need a way to kick a retry here.
1776 		 *
1777 		 * XXX: We know the fixed size needed and could
1778 		 * preallocate this using a blocking request at the
1779 		 * start of the task to avoid having to handle this
1780 		 * edge case.
1781 		 */
1782 		printf("%s: mk_update_tcb_for_ddp failed\n", __func__);
1783 		return;
1784 	}
1785 
1786 	if (!aio_set_cancel_function(job, t4_aio_cancel_active)) {
1787 		free_wrqe(wr);
1788 		recycle_pageset(toep, ps);
1789 		aio_ddp_cancel_one(job);
1790 		toep->ddp.queueing = NULL;
1791 		goto restart;
1792 	}
1793 
1794 #ifdef VERBOSE_TRACES
1795 	CTR5(KTR_CXGBE, "%s: scheduling %p for DDP[%d] (flags %#lx/%#lx)",
1796 	    __func__, job, db_idx, ddp_flags, ddp_flags_mask);
1797 #endif
1798 	/* Give the chip the go-ahead. */
1799 	t4_wrq_tx(sc, wr);
1800 	db = &toep->ddp.db[db_idx];
1801 	db->cancel_pending = 0;
1802 	db->job = job;
1803 	db->ps = ps;
1804 	toep->ddp.queueing = NULL;
1805 	toep->ddp.flags |= buf_flag;
1806 	toep->ddp.active_count++;
1807 	if (toep->ddp.active_count == 1) {
1808 		MPASS(toep->ddp.active_id == -1);
1809 		toep->ddp.active_id = db_idx;
1810 		CTR2(KTR_CXGBE, "%s: ddp_active_id = %d", __func__,
1811 		    toep->ddp.active_id);
1812 	}
1813 	goto restart;
1814 }
1815 
1816 void
1817 ddp_queue_toep(struct toepcb *toep)
1818 {
1819 
1820 	DDP_ASSERT_LOCKED(toep);
1821 	if (toep->ddp.flags & DDP_TASK_ACTIVE)
1822 		return;
1823 	toep->ddp.flags |= DDP_TASK_ACTIVE;
1824 	hold_toepcb(toep);
1825 	soaio_enqueue(&toep->ddp.requeue_task);
1826 }
1827 
1828 static void
1829 aio_ddp_requeue_task(void *context, int pending)
1830 {
1831 	struct toepcb *toep = context;
1832 
1833 	DDP_LOCK(toep);
1834 	aio_ddp_requeue(toep);
1835 	toep->ddp.flags &= ~DDP_TASK_ACTIVE;
1836 	DDP_UNLOCK(toep);
1837 
1838 	free_toepcb(toep);
1839 }
1840 
1841 static void
1842 t4_aio_cancel_active(struct kaiocb *job)
1843 {
1844 	struct socket *so = job->fd_file->f_data;
1845 	struct tcpcb *tp = so_sototcpcb(so);
1846 	struct toepcb *toep = tp->t_toe;
1847 	struct adapter *sc = td_adapter(toep->td);
1848 	uint64_t valid_flag;
1849 	int i;
1850 
1851 	DDP_LOCK(toep);
1852 	if (aio_cancel_cleared(job)) {
1853 		DDP_UNLOCK(toep);
1854 		aio_ddp_cancel_one(job);
1855 		return;
1856 	}
1857 
1858 	for (i = 0; i < nitems(toep->ddp.db); i++) {
1859 		if (toep->ddp.db[i].job == job) {
1860 			/* Should only ever get one cancel request for a job. */
1861 			MPASS(toep->ddp.db[i].cancel_pending == 0);
1862 
1863 			/*
1864 			 * Invalidate this buffer.  It will be
1865 			 * cancelled or partially completed once the
1866 			 * card ACKs the invalidate.
1867 			 */
1868 			valid_flag = i == 0 ? V_TF_DDP_BUF0_VALID(1) :
1869 			    V_TF_DDP_BUF1_VALID(1);
1870 			t4_set_tcb_field(sc, toep->ctrlq, toep,
1871 			    W_TCB_RX_DDP_FLAGS, valid_flag, 0, 1,
1872 			    i + DDP_BUF0_INVALIDATED);
1873 			toep->ddp.db[i].cancel_pending = 1;
1874 			CTR2(KTR_CXGBE, "%s: request %p marked pending",
1875 			    __func__, job);
1876 			break;
1877 		}
1878 	}
1879 	DDP_UNLOCK(toep);
1880 }
1881 
1882 static void
1883 t4_aio_cancel_queued(struct kaiocb *job)
1884 {
1885 	struct socket *so = job->fd_file->f_data;
1886 	struct tcpcb *tp = so_sototcpcb(so);
1887 	struct toepcb *toep = tp->t_toe;
1888 
1889 	DDP_LOCK(toep);
1890 	if (!aio_cancel_cleared(job)) {
1891 		TAILQ_REMOVE(&toep->ddp.aiojobq, job, list);
1892 		toep->ddp.waiting_count--;
1893 		if (toep->ddp.waiting_count == 0)
1894 			ddp_queue_toep(toep);
1895 	}
1896 	CTR2(KTR_CXGBE, "%s: request %p cancelled", __func__, job);
1897 	DDP_UNLOCK(toep);
1898 
1899 	aio_ddp_cancel_one(job);
1900 }
1901 
1902 int
1903 t4_aio_queue_ddp(struct socket *so, struct kaiocb *job)
1904 {
1905 	struct tcpcb *tp = so_sototcpcb(so);
1906 	struct toepcb *toep = tp->t_toe;
1907 
1908 
1909 	/* Ignore writes. */
1910 	if (job->uaiocb.aio_lio_opcode != LIO_READ)
1911 		return (EOPNOTSUPP);
1912 
1913 	DDP_LOCK(toep);
1914 
1915 	/*
1916 	 * XXX: Think about possibly returning errors for ENOTCONN,
1917 	 * etc.  Perhaps the caller would only queue the request
1918 	 * if it failed with EOPNOTSUPP?
1919 	 */
1920 
1921 #ifdef VERBOSE_TRACES
1922 	CTR2(KTR_CXGBE, "%s: queueing %p", __func__, job);
1923 #endif
1924 	if (!aio_set_cancel_function(job, t4_aio_cancel_queued))
1925 		panic("new job was cancelled");
1926 	TAILQ_INSERT_TAIL(&toep->ddp.aiojobq, job, list);
1927 	toep->ddp.waiting_count++;
1928 	toep->ddp.flags |= DDP_OK;
1929 
1930 	/*
1931 	 * Try to handle this request synchronously.  If this has
1932 	 * to block because the task is running, it will just bail
1933 	 * and let the task handle it instead.
1934 	 */
1935 	aio_ddp_requeue(toep);
1936 	DDP_UNLOCK(toep);
1937 	return (0);
1938 }
1939 
1940 void
1941 t4_ddp_mod_load(void)
1942 {
1943 
1944 	t4_register_shared_cpl_handler(CPL_SET_TCB_RPL, do_ddp_tcb_rpl,
1945 	    CPL_COOKIE_DDP0);
1946 	t4_register_shared_cpl_handler(CPL_SET_TCB_RPL, do_ddp_tcb_rpl,
1947 	    CPL_COOKIE_DDP1);
1948 	t4_register_cpl_handler(CPL_RX_DATA_DDP, do_rx_data_ddp);
1949 	t4_register_cpl_handler(CPL_RX_DDP_COMPLETE, do_rx_ddp_complete);
1950 	TAILQ_INIT(&ddp_orphan_pagesets);
1951 	mtx_init(&ddp_orphan_pagesets_lock, "ddp orphans", NULL, MTX_DEF);
1952 	TASK_INIT(&ddp_orphan_task, 0, ddp_free_orphan_pagesets, NULL);
1953 }
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_shared_cpl_handler(CPL_SET_TCB_RPL, NULL, CPL_COOKIE_DDP0);
1963 	t4_register_shared_cpl_handler(CPL_SET_TCB_RPL, NULL, CPL_COOKIE_DDP1);
1964 	t4_register_cpl_handler(CPL_RX_DATA_DDP, NULL);
1965 	t4_register_cpl_handler(CPL_RX_DDP_COMPLETE, NULL);
1966 }
1967 #endif
1968