xref: /freebsd/sys/dev/cxgbe/tom/t4_ddp.c (revision b1f9167f94059fd55c630891d359bcff987bd7eb)
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/types.h>
35 #include <sys/systm.h>
36 #include <sys/kernel.h>
37 #include <sys/ktr.h>
38 #include <sys/module.h>
39 #include <sys/protosw.h>
40 #include <sys/proc.h>
41 #include <sys/domain.h>
42 #include <sys/socket.h>
43 #include <sys/socketvar.h>
44 #include <sys/uio.h>
45 #include <netinet/in.h>
46 #include <netinet/in_pcb.h>
47 #include <netinet/ip.h>
48 #include <netinet/tcp_var.h>
49 #define TCPSTATES
50 #include <netinet/tcp_fsm.h>
51 #include <netinet/toecore.h>
52 
53 #include <vm/vm.h>
54 #include <vm/vm_extern.h>
55 #include <vm/vm_param.h>
56 #include <vm/pmap.h>
57 #include <vm/vm_map.h>
58 #include <vm/vm_page.h>
59 #include <vm/vm_object.h>
60 
61 #ifdef TCP_OFFLOAD
62 #include "common/common.h"
63 #include "common/t4_msg.h"
64 #include "common/t4_regs.h"
65 #include "common/t4_tcb.h"
66 #include "tom/t4_tom.h"
67 
68 #define PPOD_SZ(n)	((n) * sizeof(struct pagepod))
69 #define PPOD_SIZE	(PPOD_SZ(1))
70 
71 /* XXX: must match A_ULP_RX_TDDP_PSZ */
72 static int t4_ddp_pgsz[] = {4096, 4096 << 2, 4096 << 4, 4096 << 6};
73 
74 #if 0
75 static void
76 t4_dump_tcb(struct adapter *sc, int tid)
77 {
78 	uint32_t tcb_base, off, i, j;
79 
80 	/* Dump TCB for the tid */
81 	tcb_base = t4_read_reg(sc, A_TP_CMM_TCB_BASE);
82 	t4_write_reg(sc, PCIE_MEM_ACCESS_REG(A_PCIE_MEM_ACCESS_OFFSET, 2),
83 	    tcb_base + tid * TCB_SIZE);
84 	t4_read_reg(sc, PCIE_MEM_ACCESS_REG(A_PCIE_MEM_ACCESS_OFFSET, 2));
85 	off = 0;
86 	printf("\n");
87 	for (i = 0; i < 4; i++) {
88 		uint32_t buf[8];
89 		for (j = 0; j < 8; j++, off += 4)
90 			buf[j] = htonl(t4_read_reg(sc, MEMWIN2_BASE + off));
91 
92 		printf("%08x %08x %08x %08x %08x %08x %08x %08x\n",
93 		    buf[0], buf[1], buf[2], buf[3], buf[4], buf[5], buf[6],
94 		    buf[7]);
95 	}
96 }
97 #endif
98 
99 #define MAX_DDP_BUFFER_SIZE		(M_TCB_RX_DDP_BUF0_LEN)
100 static int
101 alloc_ppods(struct tom_data *td, int n, struct ppod_region *pr)
102 {
103 	int ppod;
104 
105 	KASSERT(n > 0, ("%s: nonsense allocation (%d)", __func__, n));
106 
107 	mtx_lock(&td->ppod_lock);
108 	if (n > td->nppods_free) {
109 		mtx_unlock(&td->ppod_lock);
110 		return (-1);
111 	}
112 
113 	if (td->nppods_free_head >= n) {
114 		td->nppods_free_head -= n;
115 		ppod = td->nppods_free_head;
116 		TAILQ_INSERT_HEAD(&td->ppods, pr, link);
117 	} else {
118 		struct ppod_region *p;
119 
120 		ppod = td->nppods_free_head;
121 		TAILQ_FOREACH(p, &td->ppods, link) {
122 			ppod += p->used + p->free;
123 			if (n <= p->free) {
124 				ppod -= n;
125 				p->free -= n;
126 				TAILQ_INSERT_AFTER(&td->ppods, p, pr, link);
127 				goto allocated;
128 			}
129 		}
130 
131 		if (__predict_false(ppod != td->nppods)) {
132 			panic("%s: ppods TAILQ (%p) corrupt."
133 			    "  At %d instead of %d at the end of the queue.",
134 			    __func__, &td->ppods, ppod, td->nppods);
135 		}
136 
137 		mtx_unlock(&td->ppod_lock);
138 		return (-1);
139 	}
140 
141 allocated:
142 	pr->used = n;
143 	pr->free = 0;
144 	td->nppods_free -= n;
145 	mtx_unlock(&td->ppod_lock);
146 
147 	return (ppod);
148 }
149 
150 static void
151 free_ppods(struct tom_data *td, struct ppod_region *pr)
152 {
153 	struct ppod_region *p;
154 
155 	KASSERT(pr->used > 0, ("%s: nonsense free (%d)", __func__, pr->used));
156 
157 	mtx_lock(&td->ppod_lock);
158 	p = TAILQ_PREV(pr, ppod_head, link);
159 	if (p != NULL)
160 		p->free += pr->used + pr->free;
161 	else
162 		td->nppods_free_head += pr->used + pr->free;
163 	td->nppods_free += pr->used;
164 	KASSERT(td->nppods_free <= td->nppods,
165 	    ("%s: nppods_free (%d) > nppods (%d).  %d freed this time.",
166 	    __func__, td->nppods_free, td->nppods, pr->used));
167 	TAILQ_REMOVE(&td->ppods, pr, link);
168 	mtx_unlock(&td->ppod_lock);
169 }
170 
171 static inline int
172 pages_to_nppods(int npages, int ddp_pgsz)
173 {
174 	int nsegs = npages * PAGE_SIZE / ddp_pgsz;
175 
176 	return (howmany(nsegs, PPOD_PAGES));
177 }
178 
179 static void
180 free_ddp_buffer(struct tom_data *td, struct ddp_buffer *db)
181 {
182 
183 	if (db == NULL)
184 		return;
185 
186 	if (db->pages)
187 		free(db->pages, M_CXGBE);
188 
189 	if (db->nppods > 0)
190 		free_ppods(td, &db->ppod_region);
191 
192 	free(db, M_CXGBE);
193 }
194 
195 void
196 release_ddp_resources(struct toepcb *toep)
197 {
198 	int i;
199 
200 	for (i = 0; i < nitems(toep->db); i++) {
201 		if (toep->db[i] != NULL) {
202 			free_ddp_buffer(toep->td, toep->db[i]);
203 			toep->db[i] = NULL;
204 		}
205 	}
206 }
207 
208 /* XXX: handle_ddp_data code duplication */
209 void
210 insert_ddp_data(struct toepcb *toep, uint32_t n)
211 {
212 	struct inpcb *inp = toep->inp;
213 	struct tcpcb *tp = intotcpcb(inp);
214 	struct sockbuf *sb = &inp->inp_socket->so_rcv;
215 	struct mbuf *m;
216 
217 	INP_WLOCK_ASSERT(inp);
218 	SOCKBUF_LOCK_ASSERT(sb);
219 
220 	m = get_ddp_mbuf(n);
221 	tp->rcv_nxt += n;
222 #ifndef USE_DDP_RX_FLOW_CONTROL
223 	KASSERT(tp->rcv_wnd >= n, ("%s: negative window size", __func__));
224 	tp->rcv_wnd -= n;
225 #endif
226 
227 	KASSERT(toep->sb_cc >= sb->sb_cc,
228 	    ("%s: sb %p has more data (%d) than last time (%d).",
229 	    __func__, sb, sb->sb_cc, toep->sb_cc));
230 	toep->rx_credits += toep->sb_cc - sb->sb_cc;
231 #ifdef USE_DDP_RX_FLOW_CONTROL
232 	toep->rx_credits -= n;	/* adjust for F_RX_FC_DDP */
233 #endif
234 	sbappendstream_locked(sb, m);
235 	toep->sb_cc = sb->sb_cc;
236 }
237 
238 /* SET_TCB_FIELD sent as a ULP command looks like this */
239 #define LEN__SET_TCB_FIELD_ULP (sizeof(struct ulp_txpkt) + \
240     sizeof(struct ulptx_idata) + sizeof(struct cpl_set_tcb_field_core))
241 
242 /* RX_DATA_ACK sent as a ULP command looks like this */
243 #define LEN__RX_DATA_ACK_ULP (sizeof(struct ulp_txpkt) + \
244     sizeof(struct ulptx_idata) + sizeof(struct cpl_rx_data_ack_core))
245 
246 static inline void *
247 mk_set_tcb_field_ulp(struct ulp_txpkt *ulpmc, struct toepcb *toep,
248     uint64_t word, uint64_t mask, uint64_t val)
249 {
250 	struct ulptx_idata *ulpsc;
251 	struct cpl_set_tcb_field_core *req;
252 
253 	ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0));
254 	ulpmc->len = htobe32(howmany(LEN__SET_TCB_FIELD_ULP, 16));
255 
256 	ulpsc = (struct ulptx_idata *)(ulpmc + 1);
257 	ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
258 	ulpsc->len = htobe32(sizeof(*req));
259 
260 	req = (struct cpl_set_tcb_field_core *)(ulpsc + 1);
261 	OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_SET_TCB_FIELD, toep->tid));
262 	req->reply_ctrl = htobe16(V_NO_REPLY(1) |
263 	    V_QUEUENO(toep->ofld_rxq->iq.abs_id));
264 	req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0));
265         req->mask = htobe64(mask);
266         req->val = htobe64(val);
267 
268 	ulpsc = (struct ulptx_idata *)(req + 1);
269 	if (LEN__SET_TCB_FIELD_ULP % 16) {
270 		ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP));
271 		ulpsc->len = htobe32(0);
272 		return (ulpsc + 1);
273 	}
274 	return (ulpsc);
275 }
276 
277 static inline void *
278 mk_rx_data_ack_ulp(struct ulp_txpkt *ulpmc, struct toepcb *toep)
279 {
280 	struct ulptx_idata *ulpsc;
281 	struct cpl_rx_data_ack_core *req;
282 
283 	ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0));
284 	ulpmc->len = htobe32(howmany(LEN__RX_DATA_ACK_ULP, 16));
285 
286 	ulpsc = (struct ulptx_idata *)(ulpmc + 1);
287 	ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
288 	ulpsc->len = htobe32(sizeof(*req));
289 
290 	req = (struct cpl_rx_data_ack_core *)(ulpsc + 1);
291 	OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_RX_DATA_ACK, toep->tid));
292 	req->credit_dack = htobe32(F_RX_MODULATE_RX);
293 
294 	ulpsc = (struct ulptx_idata *)(req + 1);
295 	if (LEN__RX_DATA_ACK_ULP % 16) {
296 		ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP));
297 		ulpsc->len = htobe32(0);
298 		return (ulpsc + 1);
299 	}
300 	return (ulpsc);
301 }
302 
303 static inline uint64_t
304 select_ddp_flags(struct socket *so, int flags, int db_idx)
305 {
306 	uint64_t ddp_flags = V_TF_DDP_INDICATE_OUT(0);
307 	int waitall = flags & MSG_WAITALL;
308 	int nb = so->so_state & SS_NBIO || flags & (MSG_DONTWAIT | MSG_NBIO);
309 
310 	KASSERT(db_idx == 0 || db_idx == 1,
311 	    ("%s: bad DDP buffer index %d", __func__, db_idx));
312 
313 	if (db_idx == 0) {
314 		ddp_flags |= V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_ACTIVE_BUF(0);
315 		if (waitall)
316 			ddp_flags |= V_TF_DDP_PUSH_DISABLE_0(1);
317 		else if (nb)
318 			ddp_flags |= V_TF_DDP_BUF0_FLUSH(1);
319 		else
320 			ddp_flags |= V_TF_DDP_BUF0_FLUSH(0);
321 	} else {
322 		ddp_flags |= V_TF_DDP_BUF1_VALID(1) | V_TF_DDP_ACTIVE_BUF(1);
323 		if (waitall)
324 			ddp_flags |= V_TF_DDP_PUSH_DISABLE_1(1);
325 		else if (nb)
326 			ddp_flags |= V_TF_DDP_BUF1_FLUSH(1);
327 		else
328 			ddp_flags |= V_TF_DDP_BUF1_FLUSH(0);
329 	}
330 
331 	return (ddp_flags);
332 }
333 
334 static struct wrqe *
335 mk_update_tcb_for_ddp(struct adapter *sc, struct toepcb *toep, int db_idx,
336     int offset, uint64_t ddp_flags)
337 {
338 	struct ddp_buffer *db = toep->db[db_idx];
339 	struct wrqe *wr;
340 	struct work_request_hdr *wrh;
341 	struct ulp_txpkt *ulpmc;
342 	int len;
343 
344 	KASSERT(db_idx == 0 || db_idx == 1,
345 	    ("%s: bad DDP buffer index %d", __func__, db_idx));
346 
347 	/*
348 	 * We'll send a compound work request that has 3 SET_TCB_FIELDs and an
349 	 * RX_DATA_ACK (with RX_MODULATE to speed up delivery).
350 	 *
351 	 * The work request header is 16B and always ends at a 16B boundary.
352 	 * The ULPTX master commands that follow must all end at 16B boundaries
353 	 * too so we round up the size to 16.
354 	 */
355 	len = sizeof(*wrh) + 3 * roundup2(LEN__SET_TCB_FIELD_ULP, 16) +
356 	    roundup2(LEN__RX_DATA_ACK_ULP, 16);
357 
358 	wr = alloc_wrqe(len, toep->ctrlq);
359 	if (wr == NULL)
360 		return (NULL);
361 	wrh = wrtod(wr);
362 	INIT_ULPTX_WRH(wrh, len, 1, 0);	/* atomic */
363 	ulpmc = (struct ulp_txpkt *)(wrh + 1);
364 
365 	/* Write the buffer's tag */
366 	ulpmc = mk_set_tcb_field_ulp(ulpmc, toep,
367 	    W_TCB_RX_DDP_BUF0_TAG + db_idx,
368 	    V_TCB_RX_DDP_BUF0_TAG(M_TCB_RX_DDP_BUF0_TAG),
369 	    V_TCB_RX_DDP_BUF0_TAG(db->tag));
370 
371 	/* Update the current offset in the DDP buffer and its total length */
372 	if (db_idx == 0)
373 		ulpmc = mk_set_tcb_field_ulp(ulpmc, toep,
374 		    W_TCB_RX_DDP_BUF0_OFFSET,
375 		    V_TCB_RX_DDP_BUF0_OFFSET(M_TCB_RX_DDP_BUF0_OFFSET) |
376 		    V_TCB_RX_DDP_BUF0_LEN(M_TCB_RX_DDP_BUF0_LEN),
377 		    V_TCB_RX_DDP_BUF0_OFFSET(offset) |
378 		    V_TCB_RX_DDP_BUF0_LEN(db->len));
379 	else
380 		ulpmc = mk_set_tcb_field_ulp(ulpmc, toep,
381 		    W_TCB_RX_DDP_BUF1_OFFSET,
382 		    V_TCB_RX_DDP_BUF1_OFFSET(M_TCB_RX_DDP_BUF1_OFFSET) |
383 		    V_TCB_RX_DDP_BUF1_LEN((u64)M_TCB_RX_DDP_BUF1_LEN << 32),
384 		    V_TCB_RX_DDP_BUF1_OFFSET(offset) |
385 		    V_TCB_RX_DDP_BUF1_LEN((u64)db->len << 32));
386 
387 	/* Update DDP flags */
388 	ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, W_TCB_RX_DDP_FLAGS,
389 	    V_TF_DDP_BUF0_FLUSH(1) | V_TF_DDP_BUF1_FLUSH(1) |
390 	    V_TF_DDP_PUSH_DISABLE_0(1) | V_TF_DDP_PUSH_DISABLE_1(1) |
391 	    V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_BUF1_VALID(1) |
392 	    V_TF_DDP_ACTIVE_BUF(1) | V_TF_DDP_INDICATE_OUT(1), ddp_flags);
393 
394 	/* Gratuitous RX_DATA_ACK with RX_MODULATE set to speed up delivery. */
395 	ulpmc = mk_rx_data_ack_ulp(ulpmc, toep);
396 
397 	return (wr);
398 }
399 
400 static void
401 discourage_ddp(struct toepcb *toep)
402 {
403 
404 	if (toep->ddp_score && --toep->ddp_score == 0) {
405 		toep->ddp_flags &= ~DDP_OK;
406 		toep->ddp_disabled = time_uptime;
407 		CTR3(KTR_CXGBE, "%s: tid %u !DDP_OK @ %u",
408 		    __func__, toep->tid, time_uptime);
409 	}
410 }
411 
412 static int
413 handle_ddp_data(struct toepcb *toep, __be32 ddp_report, __be32 rcv_nxt, int len)
414 {
415 	uint32_t report = be32toh(ddp_report);
416 	unsigned int db_flag;
417 	struct inpcb *inp = toep->inp;
418 	struct tcpcb *tp;
419 	struct socket *so;
420 	struct sockbuf *sb;
421 	struct mbuf *m;
422 
423 	db_flag = report & F_DDP_BUF_IDX ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE;
424 
425 	if (__predict_false(!(report & F_DDP_INV)))
426 		CXGBE_UNIMPLEMENTED("DDP buffer still valid");
427 
428 	INP_WLOCK(inp);
429 	so = inp_inpcbtosocket(inp);
430 	sb = &so->so_rcv;
431 	if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))) {
432 
433 		/*
434 		 * XXX: think a bit more.
435 		 * tcpcb probably gone, but socket should still be around
436 		 * because we always wait for DDP completion in soreceive no
437 		 * matter what.  Just wake it up and let it clean up.
438 		 */
439 
440 		CTR5(KTR_CXGBE, "%s: tid %u, seq 0x%x, len %d, inp_flags 0x%x",
441 		    __func__, toep->tid, be32toh(rcv_nxt), len, inp->inp_flags);
442 		SOCKBUF_LOCK(sb);
443 		goto wakeup;
444 	}
445 
446 	tp = intotcpcb(inp);
447 	len += be32toh(rcv_nxt) - tp->rcv_nxt;
448 	tp->rcv_nxt += len;
449 	tp->t_rcvtime = ticks;
450 #ifndef USE_DDP_RX_FLOW_CONTROL
451 	KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__));
452 	tp->rcv_wnd -= len;
453 #endif
454 	m = get_ddp_mbuf(len);
455 
456 	SOCKBUF_LOCK(sb);
457 	if (report & F_DDP_BUF_COMPLETE)
458 		toep->ddp_score = DDP_HIGH_SCORE;
459 	else
460 		discourage_ddp(toep);
461 
462 	KASSERT(toep->sb_cc >= sb->sb_cc,
463 	    ("%s: sb %p has more data (%d) than last time (%d).",
464 	    __func__, sb, sb->sb_cc, toep->sb_cc));
465 	toep->rx_credits += toep->sb_cc - sb->sb_cc;
466 #ifdef USE_DDP_RX_FLOW_CONTROL
467 	toep->rx_credits -= len;	/* adjust for F_RX_FC_DDP */
468 #endif
469 	sbappendstream_locked(sb, m);
470 	toep->sb_cc = sb->sb_cc;
471 wakeup:
472 	KASSERT(toep->ddp_flags & db_flag,
473 	    ("%s: DDP buffer not active. toep %p, ddp_flags 0x%x, report 0x%x",
474 	    __func__, toep, toep->ddp_flags, report));
475 	toep->ddp_flags &= ~db_flag;
476 	sorwakeup_locked(so);
477 	SOCKBUF_UNLOCK_ASSERT(sb);
478 
479 	INP_WUNLOCK(inp);
480 	return (0);
481 }
482 
483 #define DDP_ERR (F_DDP_PPOD_MISMATCH | F_DDP_LLIMIT_ERR | F_DDP_ULIMIT_ERR |\
484 	 F_DDP_PPOD_PARITY_ERR | F_DDP_PADDING_ERR | F_DDP_OFFSET_ERR |\
485 	 F_DDP_INVALID_TAG | F_DDP_COLOR_ERR | F_DDP_TID_MISMATCH |\
486 	 F_DDP_INVALID_PPOD | F_DDP_HDRCRC_ERR | F_DDP_DATACRC_ERR)
487 
488 static int
489 do_rx_data_ddp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
490 {
491 	struct adapter *sc = iq->adapter;
492 	const struct cpl_rx_data_ddp *cpl = (const void *)(rss + 1);
493 	unsigned int tid = GET_TID(cpl);
494 	uint32_t vld;
495 	struct toepcb *toep = lookup_tid(sc, tid);
496 
497 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
498 	KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__));
499 	KASSERT(!(toep->flags & TPF_SYNQE),
500 	    ("%s: toep %p claims to be a synq entry", __func__, toep));
501 
502 	vld = be32toh(cpl->ddpvld);
503 	if (__predict_false(vld & DDP_ERR)) {
504 		panic("%s: DDP error 0x%x (tid %d, toep %p)",
505 		    __func__, vld, tid, toep);
506 	}
507 
508 	handle_ddp_data(toep, cpl->u.ddp_report, cpl->seq, be16toh(cpl->len));
509 
510 	return (0);
511 }
512 
513 static int
514 do_rx_ddp_complete(struct sge_iq *iq, const struct rss_header *rss,
515     struct mbuf *m)
516 {
517 	struct adapter *sc = iq->adapter;
518 	const struct cpl_rx_ddp_complete *cpl = (const void *)(rss + 1);
519 	unsigned int tid = GET_TID(cpl);
520 	struct toepcb *toep = lookup_tid(sc, tid);
521 
522 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
523 	KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__));
524 	KASSERT(!(toep->flags & TPF_SYNQE),
525 	    ("%s: toep %p claims to be a synq entry", __func__, toep));
526 
527 	handle_ddp_data(toep, cpl->ddp_report, cpl->rcv_nxt, 0);
528 
529 	return (0);
530 }
531 
532 void
533 enable_ddp(struct adapter *sc, struct toepcb *toep)
534 {
535 
536 	KASSERT((toep->ddp_flags & (DDP_ON | DDP_OK | DDP_SC_REQ)) == DDP_OK,
537 	    ("%s: toep %p has bad ddp_flags 0x%x",
538 	    __func__, toep, toep->ddp_flags));
539 
540 	CTR3(KTR_CXGBE, "%s: tid %u (time %u)",
541 	    __func__, toep->tid, time_uptime);
542 
543 	toep->ddp_flags |= DDP_SC_REQ;
544 	t4_set_tcb_field(sc, toep, 1, W_TCB_RX_DDP_FLAGS,
545 	    V_TF_DDP_OFF(1) | V_TF_DDP_INDICATE_OUT(1) |
546 	    V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1) |
547 	    V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_BUF1_VALID(1),
548 	    V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1));
549 	t4_set_tcb_field(sc, toep, 1, W_TCB_T_FLAGS,
550 	    V_TF_RCV_COALESCE_ENABLE(1), 0);
551 }
552 
553 static inline void
554 disable_ddp(struct adapter *sc, struct toepcb *toep)
555 {
556 
557 	KASSERT((toep->ddp_flags & (DDP_ON | DDP_SC_REQ)) == DDP_ON,
558 	    ("%s: toep %p has bad ddp_flags 0x%x",
559 	    __func__, toep, toep->ddp_flags));
560 
561 	CTR3(KTR_CXGBE, "%s: tid %u (time %u)",
562 	    __func__, toep->tid, time_uptime);
563 
564 	toep->ddp_flags |= DDP_SC_REQ;
565 	t4_set_tcb_field(sc, toep, 1, W_TCB_T_FLAGS,
566 	    V_TF_RCV_COALESCE_ENABLE(1), V_TF_RCV_COALESCE_ENABLE(1));
567 	t4_set_tcb_field(sc, toep, 1, W_TCB_RX_DDP_FLAGS, V_TF_DDP_OFF(1),
568 	    V_TF_DDP_OFF(1));
569 }
570 
571 static int
572 hold_uio(struct uio *uio, vm_page_t **ppages, int *pnpages)
573 {
574 	struct vm_map *map;
575 	struct iovec *iov;
576 	vm_offset_t start, end;
577 	vm_page_t *pp;
578 	int n;
579 
580 	KASSERT(uio->uio_iovcnt == 1,
581 	    ("%s: uio_iovcnt %d", __func__, uio->uio_iovcnt));
582 	KASSERT(uio->uio_td->td_proc == curproc,
583 	    ("%s: uio proc (%p) is not curproc (%p)",
584 	    __func__, uio->uio_td->td_proc, curproc));
585 
586 	map = &curproc->p_vmspace->vm_map;
587 	iov = &uio->uio_iov[0];
588 	start = trunc_page((uintptr_t)iov->iov_base);
589 	end = round_page((vm_offset_t)iov->iov_base + iov->iov_len);
590 	n = howmany(end - start, PAGE_SIZE);
591 
592 	if (end - start > MAX_DDP_BUFFER_SIZE)
593 		return (E2BIG);
594 
595 	pp = malloc(n * sizeof(vm_page_t), M_CXGBE, M_NOWAIT);
596 	if (pp == NULL)
597 		return (ENOMEM);
598 
599 	if (vm_fault_quick_hold_pages(map, (vm_offset_t)iov->iov_base,
600 	    iov->iov_len, VM_PROT_WRITE, pp, n) < 0) {
601 		free(pp, M_CXGBE);
602 		return (EFAULT);
603 	}
604 
605 	*ppages = pp;
606 	*pnpages = n;
607 
608 	return (0);
609 }
610 
611 static int
612 bufcmp(struct ddp_buffer *db, vm_page_t *pages, int npages, int offset, int len)
613 {
614 	int i;
615 
616 	if (db == NULL || db->npages != npages || db->offset != offset ||
617 	    db->len != len)
618 		return (1);
619 
620 	for (i = 0; i < npages; i++) {
621 		if (pages[i]->phys_addr != db->pages[i]->phys_addr)
622 			return (1);
623 	}
624 
625 	return (0);
626 }
627 
628 static int
629 calculate_hcf(int n1, int n2)
630 {
631 	int a, b, t;
632 
633 	if (n1 <= n2) {
634 		a = n1;
635 		b = n2;
636 	} else {
637 		a = n2;
638 		b = n1;
639 	}
640 
641 	while (a != 0) {
642 		t = a;
643 		a = b % a;
644 		b = t;
645 	}
646 
647 	return (b);
648 }
649 
650 static struct ddp_buffer *
651 alloc_ddp_buffer(struct tom_data *td, vm_page_t *pages, int npages, int offset,
652     int len)
653 {
654 	int i, hcf, seglen, idx, ppod, nppods;
655 	struct ddp_buffer *db;
656 
657 	/*
658 	 * The DDP page size is unrelated to the VM page size.  We combine
659 	 * contiguous physical pages into larger segments to get the best DDP
660 	 * page size possible.  This is the largest of the four sizes in
661 	 * A_ULP_RX_TDDP_PSZ that evenly divides the HCF of the segment sizes in
662 	 * the page list.
663 	 */
664 	hcf = 0;
665 	for (i = 0; i < npages; i++) {
666 		seglen = PAGE_SIZE;
667 		while (i < npages - 1 &&
668 		    pages[i]->phys_addr + PAGE_SIZE == pages[i + 1]->phys_addr) {
669 			seglen += PAGE_SIZE;
670 			i++;
671 		}
672 
673 		hcf = calculate_hcf(hcf, seglen);
674 		if (hcf < t4_ddp_pgsz[1]) {
675 			idx = 0;
676 			goto have_pgsz;	/* give up, short circuit */
677 		}
678 	}
679 
680 	if (hcf % t4_ddp_pgsz[0] != 0) {
681 		/* hmmm.  This could only happen when PAGE_SIZE < 4K */
682 		KASSERT(PAGE_SIZE < 4096,
683 		    ("%s: PAGE_SIZE %d, hcf %d", __func__, PAGE_SIZE, hcf));
684 		CTR3(KTR_CXGBE, "%s: PAGE_SIZE %d, hcf %d",
685 		    __func__, PAGE_SIZE, hcf);
686 		return (NULL);
687 	}
688 
689 	for (idx = nitems(t4_ddp_pgsz) - 1; idx > 0; idx--) {
690 		if (hcf % t4_ddp_pgsz[idx] == 0)
691 			break;
692 	}
693 have_pgsz:
694 
695 	db = malloc(sizeof(*db), M_CXGBE, M_NOWAIT);
696 	if (db == NULL) {
697 		CTR1(KTR_CXGBE, "%s: malloc failed.", __func__);
698 		return (NULL);
699 	}
700 
701 	nppods = pages_to_nppods(npages, t4_ddp_pgsz[idx]);
702 	ppod = alloc_ppods(td, nppods, &db->ppod_region);
703 	if (ppod < 0) {
704 		free(db, M_CXGBE);
705 		CTR4(KTR_CXGBE, "%s: no pods, nppods %d, resid %d, pgsz %d",
706 		    __func__, nppods, len, t4_ddp_pgsz[idx]);
707 		return (NULL);
708 	}
709 
710 	KASSERT(idx <= M_PPOD_PGSZ && ppod <= M_PPOD_TAG,
711 	    ("%s: DDP pgsz_idx = %d, ppod = %d", __func__, idx, ppod));
712 
713 	db->tag = V_PPOD_PGSZ(idx) | V_PPOD_TAG(ppod);
714 	db->nppods = nppods;
715 	db->npages = npages;
716 	db->pages = pages;
717 	db->offset = offset;
718 	db->len = len;
719 
720 	CTR6(KTR_CXGBE, "New DDP buffer.  "
721 	    "ddp_pgsz %d, ppod 0x%x, npages %d, nppods %d, offset %d, len %d",
722 	    t4_ddp_pgsz[idx], ppod, db->npages, db->nppods, db->offset,
723 	    db->len);
724 
725 	return (db);
726 }
727 
728 #define NUM_ULP_TX_SC_IMM_PPODS (256 / PPOD_SIZE)
729 
730 static int
731 write_page_pods(struct adapter *sc, struct toepcb *toep, struct ddp_buffer *db)
732 {
733 	struct wrqe *wr;
734 	struct ulp_mem_io *ulpmc;
735 	struct ulptx_idata *ulpsc;
736 	struct pagepod *ppod;
737 	int i, j, k, n, chunk, len, ddp_pgsz, idx, ppod_addr;
738 	uint32_t cmd;
739 
740 	cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE));
741 	if (is_t4(sc))
742 		cmd |= htobe32(F_ULP_MEMIO_ORDER);
743 	else
744 		cmd |= htobe32(F_T5_ULP_MEMIO_IMM);
745 	ddp_pgsz = t4_ddp_pgsz[G_PPOD_PGSZ(db->tag)];
746 	ppod_addr = sc->vres.ddp.start + G_PPOD_TAG(db->tag) * PPOD_SIZE;
747 	for (i = 0; i < db->nppods; ppod_addr += chunk) {
748 
749 		/* How many page pods are we writing in this cycle */
750 		n = min(db->nppods - i, NUM_ULP_TX_SC_IMM_PPODS);
751 		chunk = PPOD_SZ(n);
752 		len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16);
753 
754 		wr = alloc_wrqe(len, toep->ctrlq);
755 		if (wr == NULL)
756 			return (ENOMEM);	/* ok to just bail out */
757 		ulpmc = wrtod(wr);
758 
759 		INIT_ULPTX_WR(ulpmc, len, 0, 0);
760 		ulpmc->cmd = cmd;
761 		ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32));
762 		ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16));
763 		ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5));
764 
765 		ulpsc = (struct ulptx_idata *)(ulpmc + 1);
766 		ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
767 		ulpsc->len = htobe32(chunk);
768 
769 		ppod = (struct pagepod *)(ulpsc + 1);
770 		for (j = 0; j < n; i++, j++, ppod++) {
771 			ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID |
772 			    V_PPOD_TID(toep->tid) | db->tag);
773 			ppod->len_offset = htobe64(V_PPOD_LEN(db->len) |
774 			    V_PPOD_OFST(db->offset));
775 			ppod->rsvd = 0;
776 			idx = i * PPOD_PAGES * (ddp_pgsz / PAGE_SIZE);
777 			for (k = 0; k < nitems(ppod->addr); k++) {
778 				if (idx < db->npages) {
779 					ppod->addr[k] =
780 					    htobe64(db->pages[idx]->phys_addr);
781 					idx += ddp_pgsz / PAGE_SIZE;
782 				} else
783 					ppod->addr[k] = 0;
784 #if 0
785 				CTR5(KTR_CXGBE,
786 				    "%s: tid %d ppod[%d]->addr[%d] = %p",
787 				    __func__, toep->tid, i, k,
788 				    htobe64(ppod->addr[k]));
789 #endif
790 			}
791 
792 		}
793 
794 		t4_wrq_tx(sc, wr);
795 	}
796 
797 	return (0);
798 }
799 
800 /*
801  * Reuse, or allocate (and program the page pods for) a new DDP buffer.  The
802  * "pages" array is handed over to this function and should not be used in any
803  * way by the caller after that.
804  */
805 static int
806 select_ddp_buffer(struct adapter *sc, struct toepcb *toep, vm_page_t *pages,
807     int npages, int db_off, int db_len)
808 {
809 	struct ddp_buffer *db;
810 	struct tom_data *td = sc->tom_softc;
811 	int i, empty_slot = -1;
812 
813 	/* Try to reuse */
814 	for (i = 0; i < nitems(toep->db); i++) {
815 		if (bufcmp(toep->db[i], pages, npages, db_off, db_len) == 0) {
816 			free(pages, M_CXGBE);
817 			return (i);	/* pages still held */
818 		} else if (toep->db[i] == NULL && empty_slot < 0)
819 			empty_slot = i;
820 	}
821 
822 	/* Allocate new buffer, write its page pods. */
823 	db = alloc_ddp_buffer(td, pages, npages, db_off, db_len);
824 	if (db == NULL) {
825 		vm_page_unhold_pages(pages, npages);
826 		free(pages, M_CXGBE);
827 		return (-1);
828 	}
829 	if (write_page_pods(sc, toep, db) != 0) {
830 		vm_page_unhold_pages(pages, npages);
831 		free_ddp_buffer(td, db);
832 		return (-1);
833 	}
834 
835 	i = empty_slot;
836 	if (i < 0) {
837 		i = arc4random() % nitems(toep->db);
838 		free_ddp_buffer(td, toep->db[i]);
839 	}
840 	toep->db[i] = db;
841 
842 	CTR5(KTR_CXGBE, "%s: tid %d, DDP buffer[%d] = %p (tag 0x%x)",
843 	    __func__, toep->tid, i, db, db->tag);
844 
845 	return (i);
846 }
847 
848 static void
849 wire_ddp_buffer(struct ddp_buffer *db)
850 {
851 	int i;
852 	vm_page_t p;
853 
854 	for (i = 0; i < db->npages; i++) {
855 		p = db->pages[i];
856 		vm_page_lock(p);
857 		vm_page_wire(p);
858 		vm_page_unhold(p);
859 		vm_page_unlock(p);
860 	}
861 }
862 
863 static void
864 unwire_ddp_buffer(struct ddp_buffer *db)
865 {
866 	int i;
867 	vm_page_t p;
868 
869 	for (i = 0; i < db->npages; i++) {
870 		p = db->pages[i];
871 		vm_page_lock(p);
872 		vm_page_unwire(p, PQ_INACTIVE);
873 		vm_page_unlock(p);
874 	}
875 }
876 
877 static int
878 handle_ddp(struct socket *so, struct uio *uio, int flags, int error)
879 {
880 	struct sockbuf *sb = &so->so_rcv;
881 	struct tcpcb *tp = so_sototcpcb(so);
882 	struct toepcb *toep = tp->t_toe;
883 	struct adapter *sc = td_adapter(toep->td);
884 	vm_page_t *pages;
885 	int npages, db_idx, rc, buf_flag;
886 	struct ddp_buffer *db;
887 	struct wrqe *wr;
888 	uint64_t ddp_flags;
889 
890 	SOCKBUF_LOCK_ASSERT(sb);
891 
892 #if 0
893 	if (sb->sb_cc + sc->tt.ddp_thres > uio->uio_resid) {
894 		CTR4(KTR_CXGBE, "%s: sb_cc %d, threshold %d, resid %d",
895 		    __func__, sb->sb_cc, sc->tt.ddp_thres, uio->uio_resid);
896 	}
897 #endif
898 
899 	/* XXX: too eager to disable DDP, could handle NBIO better than this. */
900 	if (sb->sb_cc >= uio->uio_resid || uio->uio_resid < sc->tt.ddp_thres ||
901 	    uio->uio_resid > MAX_DDP_BUFFER_SIZE || uio->uio_iovcnt > 1 ||
902 	    so->so_state & SS_NBIO || flags & (MSG_DONTWAIT | MSG_NBIO) ||
903 	    error || so->so_error || sb->sb_state & SBS_CANTRCVMORE)
904 		goto no_ddp;
905 
906 	/*
907 	 * Fault in and then hold the pages of the uio buffers.  We'll wire them
908 	 * a bit later if everything else works out.
909 	 */
910 	SOCKBUF_UNLOCK(sb);
911 	if (hold_uio(uio, &pages, &npages) != 0) {
912 		SOCKBUF_LOCK(sb);
913 		goto no_ddp;
914 	}
915 	SOCKBUF_LOCK(sb);
916 	if (__predict_false(so->so_error || sb->sb_state & SBS_CANTRCVMORE)) {
917 		vm_page_unhold_pages(pages, npages);
918 		free(pages, M_CXGBE);
919 		goto no_ddp;
920 	}
921 
922 	/*
923 	 * Figure out which one of the two DDP buffers to use this time.
924 	 */
925 	db_idx = select_ddp_buffer(sc, toep, pages, npages,
926 	    (uintptr_t)uio->uio_iov->iov_base & PAGE_MASK, uio->uio_resid);
927 	pages = NULL;	/* handed off to select_ddp_buffer */
928 	if (db_idx < 0)
929 		goto no_ddp;
930 	db = toep->db[db_idx];
931 	buf_flag = db_idx == 0 ? DDP_BUF0_ACTIVE : DDP_BUF1_ACTIVE;
932 
933 	/*
934 	 * Build the compound work request that tells the chip where to DMA the
935 	 * payload.
936 	 */
937 	ddp_flags = select_ddp_flags(so, flags, db_idx);
938 	wr = mk_update_tcb_for_ddp(sc, toep, db_idx, sb->sb_cc, ddp_flags);
939 	if (wr == NULL) {
940 		/*
941 		 * Just unhold the pages.  The DDP buffer's software state is
942 		 * left as-is in the toep.  The page pods were written
943 		 * successfully and we may have an opportunity to use it in the
944 		 * future.
945 		 */
946 		vm_page_unhold_pages(db->pages, db->npages);
947 		goto no_ddp;
948 	}
949 
950 	/* Wire (and then unhold) the pages, and give the chip the go-ahead. */
951 	wire_ddp_buffer(db);
952 	t4_wrq_tx(sc, wr);
953 	sb->sb_flags &= ~SB_DDP_INDICATE;
954 	toep->ddp_flags |= buf_flag;
955 
956 	/*
957 	 * Wait for the DDP operation to complete and then unwire the pages.
958 	 * The return code from the sbwait will be the final return code of this
959 	 * function.  But we do need to wait for DDP no matter what.
960 	 */
961 	rc = sbwait(sb);
962 	while (toep->ddp_flags & buf_flag) {
963 		sb->sb_flags |= SB_WAIT;
964 		msleep(&sb->sb_cc, &sb->sb_mtx, PSOCK , "sbwait", 0);
965 	}
966 	unwire_ddp_buffer(db);
967 	return (rc);
968 no_ddp:
969 	disable_ddp(sc, toep);
970 	discourage_ddp(toep);
971 	sb->sb_flags &= ~SB_DDP_INDICATE;
972 	return (0);
973 }
974 
975 void
976 t4_init_ddp(struct adapter *sc, struct tom_data *td)
977 {
978 	int nppods = sc->vres.ddp.size / PPOD_SIZE;
979 
980 	td->nppods = nppods;
981 	td->nppods_free = nppods;
982 	td->nppods_free_head = nppods;
983 	TAILQ_INIT(&td->ppods);
984 	mtx_init(&td->ppod_lock, "page pods", NULL, MTX_DEF);
985 
986 	t4_register_cpl_handler(sc, CPL_RX_DATA_DDP, do_rx_data_ddp);
987 	t4_register_cpl_handler(sc, CPL_RX_DDP_COMPLETE, do_rx_ddp_complete);
988 }
989 
990 void
991 t4_uninit_ddp(struct adapter *sc __unused, struct tom_data *td)
992 {
993 
994 	KASSERT(td->nppods == td->nppods_free,
995 	    ("%s: page pods still in use, nppods = %d, free = %d",
996 	    __func__, td->nppods, td->nppods_free));
997 
998 	if (mtx_initialized(&td->ppod_lock))
999 		mtx_destroy(&td->ppod_lock);
1000 }
1001 
1002 #define	VNET_SO_ASSERT(so)						\
1003 	VNET_ASSERT(curvnet != NULL,					\
1004 	    ("%s:%d curvnet is NULL, so=%p", __func__, __LINE__, (so)));
1005 #define	SBLOCKWAIT(f)	(((f) & MSG_DONTWAIT) ? 0 : SBL_WAIT)
1006 static int
1007 soreceive_rcvoob(struct socket *so, struct uio *uio, int flags)
1008 {
1009 
1010 	CXGBE_UNIMPLEMENTED(__func__);
1011 }
1012 
1013 static char ddp_magic_str[] = "nothing to see here";
1014 
1015 struct mbuf *
1016 get_ddp_mbuf(int len)
1017 {
1018 	struct mbuf *m;
1019 
1020 	m = m_get(M_NOWAIT, MT_DATA);
1021 	if (m == NULL)
1022 		CXGBE_UNIMPLEMENTED("mbuf alloc failure");
1023 	m->m_len = len;
1024 	m->m_data = &ddp_magic_str[0];
1025 
1026 	return (m);
1027 }
1028 
1029 static inline int
1030 is_ddp_mbuf(struct mbuf *m)
1031 {
1032 
1033 	return (m->m_data == &ddp_magic_str[0]);
1034 }
1035 
1036 /*
1037  * Copy an mbuf chain into a uio limited by len if set.
1038  */
1039 static int
1040 m_mbuftouio_ddp(struct uio *uio, struct mbuf *m, int len)
1041 {
1042 	int error, length, total;
1043 	int progress = 0;
1044 
1045 	if (len > 0)
1046 		total = min(uio->uio_resid, len);
1047 	else
1048 		total = uio->uio_resid;
1049 
1050 	/* Fill the uio with data from the mbufs. */
1051 	for (; m != NULL; m = m->m_next) {
1052 		length = min(m->m_len, total - progress);
1053 
1054 		if (is_ddp_mbuf(m)) {
1055 			enum uio_seg segflag = uio->uio_segflg;
1056 
1057 			uio->uio_segflg	= UIO_NOCOPY;
1058 			error = uiomove(mtod(m, void *), length, uio);
1059 			uio->uio_segflg	= segflag;
1060 		} else
1061 			error = uiomove(mtod(m, void *), length, uio);
1062 		if (error)
1063 			return (error);
1064 
1065 		progress += length;
1066 	}
1067 
1068 	return (0);
1069 }
1070 
1071 /*
1072  * Based on soreceive_stream() in uipc_socket.c
1073  */
1074 int
1075 t4_soreceive_ddp(struct socket *so, struct sockaddr **psa, struct uio *uio,
1076     struct mbuf **mp0, struct mbuf **controlp, int *flagsp)
1077 {
1078 	int len = 0, error = 0, flags, oresid, ddp_handled = 0;
1079 	struct sockbuf *sb;
1080 	struct mbuf *m, *n = NULL;
1081 
1082 	/* We only do stream sockets. */
1083 	if (so->so_type != SOCK_STREAM)
1084 		return (EINVAL);
1085 	if (psa != NULL)
1086 		*psa = NULL;
1087 	if (controlp != NULL)
1088 		return (EINVAL);
1089 	if (flagsp != NULL)
1090 		flags = *flagsp &~ MSG_EOR;
1091 	else
1092 		flags = 0;
1093 	if (flags & MSG_OOB)
1094 		return (soreceive_rcvoob(so, uio, flags));
1095 	if (mp0 != NULL)
1096 		*mp0 = NULL;
1097 
1098 	sb = &so->so_rcv;
1099 
1100 	/* Prevent other readers from entering the socket. */
1101 	error = sblock(sb, SBLOCKWAIT(flags));
1102 	if (error)
1103 		goto out;
1104 	SOCKBUF_LOCK(sb);
1105 
1106 	/* Easy one, no space to copyout anything. */
1107 	if (uio->uio_resid == 0) {
1108 		error = EINVAL;
1109 		goto out;
1110 	}
1111 	oresid = uio->uio_resid;
1112 
1113 	/* We will never ever get anything unless we are or were connected. */
1114 	if (!(so->so_state & (SS_ISCONNECTED|SS_ISDISCONNECTED))) {
1115 		error = ENOTCONN;
1116 		goto out;
1117 	}
1118 
1119 restart:
1120 	SOCKBUF_LOCK_ASSERT(&so->so_rcv);
1121 
1122 	if (sb->sb_flags & SB_DDP_INDICATE && !ddp_handled) {
1123 
1124 		/* uio should be just as it was at entry */
1125 		KASSERT(oresid == uio->uio_resid,
1126 		    ("%s: oresid = %d, uio_resid = %zd, sb_cc = %d",
1127 		    __func__, oresid, uio->uio_resid, sb->sb_cc));
1128 
1129 		error = handle_ddp(so, uio, flags, 0);
1130 		ddp_handled = 1;
1131 		if (error)
1132 			goto out;
1133 	}
1134 
1135 	/* Abort if socket has reported problems. */
1136 	if (so->so_error) {
1137 		if (sb->sb_cc > 0)
1138 			goto deliver;
1139 		if (oresid > uio->uio_resid)
1140 			goto out;
1141 		error = so->so_error;
1142 		if (!(flags & MSG_PEEK))
1143 			so->so_error = 0;
1144 		goto out;
1145 	}
1146 
1147 	/* Door is closed.  Deliver what is left, if any. */
1148 	if (sb->sb_state & SBS_CANTRCVMORE) {
1149 		if (sb->sb_cc > 0)
1150 			goto deliver;
1151 		else
1152 			goto out;
1153 	}
1154 
1155 	/* Socket buffer is empty and we shall not block. */
1156 	if (sb->sb_cc == 0 &&
1157 	    ((so->so_state & SS_NBIO) || (flags & (MSG_DONTWAIT|MSG_NBIO)))) {
1158 		error = EAGAIN;
1159 		goto out;
1160 	}
1161 
1162 	/* Socket buffer got some data that we shall deliver now. */
1163 	if (sb->sb_cc > 0 && !(flags & MSG_WAITALL) &&
1164 	    ((sb->sb_flags & SS_NBIO) ||
1165 	     (flags & (MSG_DONTWAIT|MSG_NBIO)) ||
1166 	     sb->sb_cc >= sb->sb_lowat ||
1167 	     sb->sb_cc >= uio->uio_resid ||
1168 	     sb->sb_cc >= sb->sb_hiwat) ) {
1169 		goto deliver;
1170 	}
1171 
1172 	/* On MSG_WAITALL we must wait until all data or error arrives. */
1173 	if ((flags & MSG_WAITALL) &&
1174 	    (sb->sb_cc >= uio->uio_resid || sb->sb_cc >= sb->sb_lowat))
1175 		goto deliver;
1176 
1177 	/*
1178 	 * Wait and block until (more) data comes in.
1179 	 * NB: Drops the sockbuf lock during wait.
1180 	 */
1181 	error = sbwait(sb);
1182 	if (error) {
1183 		if (sb->sb_flags & SB_DDP_INDICATE && !ddp_handled) {
1184 			(void) handle_ddp(so, uio, flags, 1);
1185 			ddp_handled = 1;
1186 		}
1187 		goto out;
1188 	}
1189 	goto restart;
1190 
1191 deliver:
1192 	SOCKBUF_LOCK_ASSERT(&so->so_rcv);
1193 	KASSERT(sb->sb_cc > 0, ("%s: sockbuf empty", __func__));
1194 	KASSERT(sb->sb_mb != NULL, ("%s: sb_mb == NULL", __func__));
1195 
1196 	if (sb->sb_flags & SB_DDP_INDICATE && !ddp_handled)
1197 		goto restart;
1198 
1199 	/* Statistics. */
1200 	if (uio->uio_td)
1201 		uio->uio_td->td_ru.ru_msgrcv++;
1202 
1203 	/* Fill uio until full or current end of socket buffer is reached. */
1204 	len = min(uio->uio_resid, sb->sb_cc);
1205 	if (mp0 != NULL) {
1206 		/* Dequeue as many mbufs as possible. */
1207 		if (!(flags & MSG_PEEK) && len >= sb->sb_mb->m_len) {
1208 			for (*mp0 = m = sb->sb_mb;
1209 			     m != NULL && m->m_len <= len;
1210 			     m = m->m_next) {
1211 				len -= m->m_len;
1212 				uio->uio_resid -= m->m_len;
1213 				sbfree(sb, m);
1214 				n = m;
1215 			}
1216 			sb->sb_mb = m;
1217 			if (sb->sb_mb == NULL)
1218 				SB_EMPTY_FIXUP(sb);
1219 			n->m_next = NULL;
1220 		}
1221 		/* Copy the remainder. */
1222 		if (len > 0) {
1223 			KASSERT(sb->sb_mb != NULL,
1224 			    ("%s: len > 0 && sb->sb_mb empty", __func__));
1225 
1226 			m = m_copym(sb->sb_mb, 0, len, M_NOWAIT);
1227 			if (m == NULL)
1228 				len = 0;	/* Don't flush data from sockbuf. */
1229 			else
1230 				uio->uio_resid -= m->m_len;
1231 			if (*mp0 != NULL)
1232 				n->m_next = m;
1233 			else
1234 				*mp0 = m;
1235 			if (*mp0 == NULL) {
1236 				error = ENOBUFS;
1237 				goto out;
1238 			}
1239 		}
1240 	} else {
1241 		/* NB: Must unlock socket buffer as uiomove may sleep. */
1242 		SOCKBUF_UNLOCK(sb);
1243 		error = m_mbuftouio_ddp(uio, sb->sb_mb, len);
1244 		SOCKBUF_LOCK(sb);
1245 		if (error)
1246 			goto out;
1247 	}
1248 	SBLASTRECORDCHK(sb);
1249 	SBLASTMBUFCHK(sb);
1250 
1251 	/*
1252 	 * Remove the delivered data from the socket buffer unless we
1253 	 * were only peeking.
1254 	 */
1255 	if (!(flags & MSG_PEEK)) {
1256 		if (len > 0)
1257 			sbdrop_locked(sb, len);
1258 
1259 		/* Notify protocol that we drained some data. */
1260 		if ((so->so_proto->pr_flags & PR_WANTRCVD) &&
1261 		    (((flags & MSG_WAITALL) && uio->uio_resid > 0) ||
1262 		     !(flags & MSG_SOCALLBCK))) {
1263 			SOCKBUF_UNLOCK(sb);
1264 			VNET_SO_ASSERT(so);
1265 			(*so->so_proto->pr_usrreqs->pru_rcvd)(so, flags);
1266 			SOCKBUF_LOCK(sb);
1267 		}
1268 	}
1269 
1270 	/*
1271 	 * For MSG_WAITALL we may have to loop again and wait for
1272 	 * more data to come in.
1273 	 */
1274 	if ((flags & MSG_WAITALL) && uio->uio_resid > 0)
1275 		goto restart;
1276 out:
1277 	SOCKBUF_LOCK_ASSERT(sb);
1278 	SBLASTRECORDCHK(sb);
1279 	SBLASTMBUFCHK(sb);
1280 	SOCKBUF_UNLOCK(sb);
1281 	sbunlock(sb);
1282 	return (error);
1283 }
1284 
1285 #endif
1286