xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision c2bce4a2fcf3083607e00a1734b47c249751c8a8)
1 /*-
2  * Copyright (c) 2011 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/types.h>
34 #include <sys/mbuf.h>
35 #include <sys/socket.h>
36 #include <sys/kernel.h>
37 #include <sys/malloc.h>
38 #include <sys/queue.h>
39 #include <sys/taskqueue.h>
40 #include <sys/sysctl.h>
41 #include <net/bpf.h>
42 #include <net/ethernet.h>
43 #include <net/if.h>
44 #include <net/if_vlan_var.h>
45 #include <netinet/in.h>
46 #include <netinet/ip.h>
47 #include <netinet/tcp.h>
48 
49 #include "common/common.h"
50 #include "common/t4_regs.h"
51 #include "common/t4_regs_values.h"
52 #include "common/t4_msg.h"
53 #include "common/t4fw_interface.h"
54 
55 struct fl_buf_info {
56 	int size;
57 	int type;
58 	uma_zone_t zone;
59 };
60 
61 /* Filled up by t4_sge_modload */
62 static struct fl_buf_info fl_buf_info[FL_BUF_SIZES];
63 
64 #define FL_BUF_SIZE(x)	(fl_buf_info[x].size)
65 #define FL_BUF_TYPE(x)	(fl_buf_info[x].type)
66 #define FL_BUF_ZONE(x)	(fl_buf_info[x].zone)
67 
68 enum {
69 	FL_PKTSHIFT = 2
70 };
71 
72 #define FL_ALIGN	min(CACHE_LINE_SIZE, 32)
73 #if CACHE_LINE_SIZE > 64
74 #define SPG_LEN		128
75 #else
76 #define SPG_LEN		64
77 #endif
78 
79 /* Used to track coalesced tx work request */
80 struct txpkts {
81 	uint64_t *flitp;	/* ptr to flit where next pkt should start */
82 	uint8_t npkt;		/* # of packets in this work request */
83 	uint8_t nflits;		/* # of flits used by this work request */
84 	uint16_t plen;		/* total payload (sum of all packets) */
85 };
86 
87 /* A packet's SGL.  This + m_pkthdr has all info needed for tx */
88 struct sgl {
89 	int nsegs;		/* # of segments in the SGL, 0 means imm. tx */
90 	int nflits;		/* # of flits needed for the SGL */
91 	bus_dma_segment_t seg[TX_SGL_SEGS];
92 };
93 
94 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int,
95     int, iq_intr_handler_t *, char *);
96 static inline void init_fl(struct sge_fl *, int, char *);
97 static inline void init_eq(struct sge_eq *, int, char *);
98 static int alloc_ring(struct adapter *, size_t, bus_dma_tag_t *, bus_dmamap_t *,
99     bus_addr_t *, void **);
100 static int free_ring(struct adapter *, bus_dma_tag_t, bus_dmamap_t, bus_addr_t,
101     void *);
102 static int alloc_iq_fl(struct port_info *, struct sge_iq *, struct sge_fl *,
103     int);
104 static int free_iq_fl(struct port_info *, struct sge_iq *, struct sge_fl *);
105 static int alloc_iq(struct sge_iq *, int);
106 static int free_iq(struct sge_iq *);
107 static int alloc_rxq(struct port_info *, struct sge_rxq *, int, int);
108 static int free_rxq(struct port_info *, struct sge_rxq *);
109 static int alloc_ctrlq(struct adapter *, struct sge_ctrlq *, int);
110 static int free_ctrlq(struct adapter *, struct sge_ctrlq *);
111 static int alloc_txq(struct port_info *, struct sge_txq *, int);
112 static int free_txq(struct port_info *, struct sge_txq *);
113 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int);
114 static inline bool is_new_response(const struct sge_iq *, struct rsp_ctrl **);
115 static inline void iq_next(struct sge_iq *);
116 static inline void ring_fl_db(struct adapter *, struct sge_fl *);
117 static void refill_fl(struct sge_fl *, int);
118 static int alloc_fl_sdesc(struct sge_fl *);
119 static void free_fl_sdesc(struct sge_fl *);
120 static int alloc_tx_maps(struct sge_txq *);
121 static void free_tx_maps(struct sge_txq *);
122 static void set_fl_tag_idx(struct sge_fl *, int);
123 
124 static int get_pkt_sgl(struct sge_txq *, struct mbuf **, struct sgl *, int);
125 static int free_pkt_sgl(struct sge_txq *, struct sgl *);
126 static int write_txpkt_wr(struct port_info *, struct sge_txq *, struct mbuf *,
127     struct sgl *);
128 static int add_to_txpkts(struct port_info *, struct sge_txq *, struct txpkts *,
129     struct mbuf *, struct sgl *);
130 static void write_txpkts_wr(struct sge_txq *, struct txpkts *);
131 static inline void write_ulp_cpl_sgl(struct port_info *, struct sge_txq *,
132     struct txpkts *, struct mbuf *, struct sgl *);
133 static int write_sgl_to_txd(struct sge_eq *, struct sgl *, caddr_t *);
134 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int);
135 static inline void ring_eq_db(struct adapter *, struct sge_eq *);
136 static inline int reclaimable(struct sge_eq *);
137 static int reclaim_tx_descs(struct sge_txq *, int, int);
138 static void write_eqflush_wr(struct sge_eq *);
139 static __be64 get_flit(bus_dma_segment_t *, int, int);
140 static int handle_sge_egr_update(struct adapter *,
141     const struct cpl_sge_egr_update *);
142 
143 static int ctrl_tx(struct adapter *, struct sge_ctrlq *, struct mbuf *);
144 
145 /*
146  * Called on MOD_LOAD and fills up fl_buf_info[].
147  */
148 void
149 t4_sge_modload(void)
150 {
151 	int i;
152 	int bufsize[FL_BUF_SIZES] = {
153 		MCLBYTES,
154 #if MJUMPAGESIZE != MCLBYTES
155 		MJUMPAGESIZE,
156 #endif
157 		MJUM9BYTES,
158 		MJUM16BYTES
159 	};
160 
161 	for (i = 0; i < FL_BUF_SIZES; i++) {
162 		FL_BUF_SIZE(i) = bufsize[i];
163 		FL_BUF_TYPE(i) = m_gettype(bufsize[i]);
164 		FL_BUF_ZONE(i) = m_getzone(bufsize[i]);
165 	}
166 }
167 
168 /**
169  *	t4_sge_init - initialize SGE
170  *	@sc: the adapter
171  *
172  *	Performs SGE initialization needed every time after a chip reset.
173  *	We do not initialize any of the queues here, instead the driver
174  *	top-level must request them individually.
175  */
176 void
177 t4_sge_init(struct adapter *sc)
178 {
179 	struct sge *s = &sc->sge;
180 	int i;
181 
182 	t4_set_reg_field(sc, A_SGE_CONTROL, V_PKTSHIFT(M_PKTSHIFT) |
183 			 V_INGPADBOUNDARY(M_INGPADBOUNDARY) |
184 			 F_EGRSTATUSPAGESIZE,
185 			 V_INGPADBOUNDARY(ilog2(FL_ALIGN) - 5) |
186 			 V_PKTSHIFT(FL_PKTSHIFT) |
187 			 F_RXPKTCPLMODE |
188 			 V_EGRSTATUSPAGESIZE(SPG_LEN == 128));
189 	t4_set_reg_field(sc, A_SGE_HOST_PAGE_SIZE,
190 			 V_HOSTPAGESIZEPF0(M_HOSTPAGESIZEPF0),
191 			 V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10));
192 
193 	for (i = 0; i < FL_BUF_SIZES; i++) {
194 		t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i),
195 		    FL_BUF_SIZE(i));
196 	}
197 
198 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD,
199 		     V_THRESHOLD_0(s->counter_val[0]) |
200 		     V_THRESHOLD_1(s->counter_val[1]) |
201 		     V_THRESHOLD_2(s->counter_val[2]) |
202 		     V_THRESHOLD_3(s->counter_val[3]));
203 
204 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1,
205 		     V_TIMERVALUE0(us_to_core_ticks(sc, s->timer_val[0])) |
206 		     V_TIMERVALUE1(us_to_core_ticks(sc, s->timer_val[1])));
207 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3,
208 		     V_TIMERVALUE2(us_to_core_ticks(sc, s->timer_val[2])) |
209 		     V_TIMERVALUE3(us_to_core_ticks(sc, s->timer_val[3])));
210 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5,
211 		     V_TIMERVALUE4(us_to_core_ticks(sc, s->timer_val[4])) |
212 		     V_TIMERVALUE5(us_to_core_ticks(sc, s->timer_val[5])));
213 }
214 
215 int
216 t4_create_dma_tag(struct adapter *sc)
217 {
218 	int rc;
219 
220 	rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
221 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
222 	    BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
223 	    NULL, &sc->dmat);
224 	if (rc != 0) {
225 		device_printf(sc->dev,
226 		    "failed to create main DMA tag: %d\n", rc);
227 	}
228 
229 	return (rc);
230 }
231 
232 int
233 t4_destroy_dma_tag(struct adapter *sc)
234 {
235 	if (sc->dmat)
236 		bus_dma_tag_destroy(sc->dmat);
237 
238 	return (0);
239 }
240 
241 /*
242  * Allocate and initialize the firmware event queue, control queues, and the
243  * forwarded interrupt queues (if any).  The adapter owns all these queues as
244  * they are not associated with any particular port.
245  *
246  * Returns errno on failure.  Resources allocated up to that point may still be
247  * allocated.  Caller is responsible for cleanup in case this function fails.
248  */
249 int
250 t4_setup_adapter_queues(struct adapter *sc)
251 {
252 	int i, rc;
253 	struct sge_iq *iq, *fwq;
254 	struct sge_ctrlq *ctrlq;
255 	iq_intr_handler_t *handler;
256 	char name[16];
257 
258 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
259 
260 	if (sysctl_ctx_init(&sc->ctx) == 0) {
261 		struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
262 		struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
263 
264 		sc->oid_ctrlq = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO,
265 		    "ctrlq", CTLFLAG_RD, NULL, "ctrl queues");
266 	}
267 
268 	fwq = &sc->sge.fwq;
269 	if (sc->flags & INTR_FWD) {
270 		iq = &sc->sge.fiq[0];
271 
272 		/*
273 		 * Forwarded interrupt queues - allocate 1 if there's only 1
274 		 * vector available, one less than the number of vectors
275 		 * otherwise (the first vector is reserved for the error
276 		 * interrupt in that case).
277 		 */
278 		i = sc->intr_count > 1 ? 1 : 0;
279 		for (; i < sc->intr_count; i++, iq++) {
280 
281 			snprintf(name, sizeof(name), "%s fiq%d",
282 			    device_get_nameunit(sc->dev), i);
283 			init_iq(iq, sc, 0, 0, (sc->sge.nrxq + 1) * 2, 16, NULL,
284 			    name);
285 
286 			rc = alloc_iq(iq, i);
287 			if (rc != 0) {
288 				device_printf(sc->dev,
289 				    "failed to create fwd intr queue %d: %d\n",
290 				    i, rc);
291 				return (rc);
292 			}
293 		}
294 
295 		handler = t4_evt_rx;
296 		i = 0;	/* forward fwq's interrupt to the first fiq */
297 	} else {
298 		handler = NULL;
299 		i = 1;	/* fwq should use vector 1 (0 is used by error) */
300 	}
301 
302 	snprintf(name, sizeof(name), "%s fwq", device_get_nameunit(sc->dev));
303 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE, FW_IQ_ESIZE, handler, name);
304 	rc = alloc_iq(fwq, i);
305 	if (rc != 0) {
306 		device_printf(sc->dev,
307 		    "failed to create firmware event queue: %d\n", rc);
308 
309 		return (rc);
310 	}
311 
312 	/*
313 	 * Control queues - one per hardware channel.
314 	 */
315 	ctrlq = &sc->sge.ctrlq[0];
316 	for (i = 0; i < NCHAN; i++, ctrlq++) {
317 		snprintf(name, sizeof(name), "%s ctrlq%d",
318 		    device_get_nameunit(sc->dev), i);
319 		init_eq(&ctrlq->eq, CTRL_EQ_QSIZE, name);
320 
321 		rc = alloc_ctrlq(sc, ctrlq, i);
322 		if (rc != 0) {
323 			device_printf(sc->dev,
324 			    "failed to create control queue %d: %d\n", i, rc);
325 			return (rc);
326 		}
327 	}
328 
329 	return (rc);
330 }
331 
332 /*
333  * Idempotent
334  */
335 int
336 t4_teardown_adapter_queues(struct adapter *sc)
337 {
338 	int i;
339 	struct sge_iq *iq;
340 
341 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
342 
343 	/* Do this before freeing the queues */
344 	if (sc->oid_ctrlq) {
345 		sysctl_ctx_free(&sc->ctx);
346 		sc->oid_ctrlq = NULL;
347 	}
348 
349 	for (i = 0; i < NCHAN; i++)
350 		free_ctrlq(sc, &sc->sge.ctrlq[i]);
351 
352 	iq = &sc->sge.fwq;
353 	free_iq(iq);
354 	if (sc->flags & INTR_FWD) {
355 		for (i = 0; i < NFIQ(sc); i++) {
356 			iq = &sc->sge.fiq[i];
357 			free_iq(iq);
358 		}
359 	}
360 
361 	return (0);
362 }
363 
364 int
365 t4_setup_eth_queues(struct port_info *pi)
366 {
367 	int rc = 0, i, intr_idx;
368 	struct sge_rxq *rxq;
369 	struct sge_txq *txq;
370 	char name[16];
371 	struct adapter *sc = pi->adapter;
372 
373 	if (sysctl_ctx_init(&pi->ctx) == 0) {
374 		struct sysctl_oid *oid = device_get_sysctl_tree(pi->dev);
375 		struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
376 
377 		pi->oid_rxq = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO,
378 		    "rxq", CTLFLAG_RD, NULL, "rx queues");
379 		pi->oid_txq = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO,
380 		    "txq", CTLFLAG_RD, NULL, "tx queues");
381 	}
382 
383 	for_each_rxq(pi, i, rxq) {
384 
385 		snprintf(name, sizeof(name), "%s rxq%d-iq",
386 		    device_get_nameunit(pi->dev), i);
387 		init_iq(&rxq->iq, sc, pi->tmr_idx, pi->pktc_idx,
388 		    pi->qsize_rxq, RX_IQ_ESIZE,
389 		    sc->flags & INTR_FWD ? t4_eth_rx : NULL, name);
390 
391 		snprintf(name, sizeof(name), "%s rxq%d-fl",
392 		    device_get_nameunit(pi->dev), i);
393 		init_fl(&rxq->fl, pi->qsize_rxq / 8, name);
394 
395 		if (sc->flags & INTR_FWD)
396 			intr_idx = (pi->first_rxq + i) % NFIQ(sc);
397 		else
398 			intr_idx = pi->first_rxq + i + 2;
399 
400 		rc = alloc_rxq(pi, rxq, intr_idx, i);
401 		if (rc != 0)
402 			goto done;
403 
404 		intr_idx++;
405 	}
406 
407 	for_each_txq(pi, i, txq) {
408 
409 		snprintf(name, sizeof(name), "%s txq%d",
410 		    device_get_nameunit(pi->dev), i);
411 		init_eq(&txq->eq, pi->qsize_txq, name);
412 
413 		rc = alloc_txq(pi, txq, i);
414 		if (rc != 0)
415 			goto done;
416 	}
417 
418 done:
419 	if (rc)
420 		t4_teardown_eth_queues(pi);
421 
422 	return (rc);
423 }
424 
425 /*
426  * Idempotent
427  */
428 int
429 t4_teardown_eth_queues(struct port_info *pi)
430 {
431 	int i;
432 	struct sge_rxq *rxq;
433 	struct sge_txq *txq;
434 
435 	/* Do this before freeing the queues */
436 	if (pi->oid_txq || pi->oid_rxq) {
437 		sysctl_ctx_free(&pi->ctx);
438 		pi->oid_txq = pi->oid_rxq = NULL;
439 	}
440 
441 	for_each_txq(pi, i, txq) {
442 		free_txq(pi, txq);
443 	}
444 
445 	for_each_rxq(pi, i, rxq) {
446 		free_rxq(pi, rxq);
447 	}
448 
449 	return (0);
450 }
451 
452 /* Deals with errors and forwarded interrupts */
453 void
454 t4_intr_all(void *arg)
455 {
456 	struct adapter *sc = arg;
457 
458 	t4_intr_err(arg);
459 	t4_intr_fwd(&sc->sge.fiq[0]);
460 }
461 
462 /* Deals with forwarded interrupts on the given ingress queue */
463 void
464 t4_intr_fwd(void *arg)
465 {
466 	struct sge_iq *iq = arg, *q;
467 	struct adapter *sc = iq->adapter;
468 	struct rsp_ctrl *ctrl;
469 	int ndesc_pending = 0, ndesc_total = 0;
470 	int qid;
471 
472 	if (!atomic_cmpset_32(&iq->state, IQS_IDLE, IQS_BUSY))
473 		return;
474 
475 	while (is_new_response(iq, &ctrl)) {
476 
477 		rmb();
478 
479 		/* Only interrupt muxing expected on this queue */
480 		KASSERT(G_RSPD_TYPE(ctrl->u.type_gen) == X_RSPD_TYPE_INTR,
481 		    ("unexpected event on forwarded interrupt queue: %x",
482 		    G_RSPD_TYPE(ctrl->u.type_gen)));
483 
484 		qid = ntohl(ctrl->pldbuflen_qid) - sc->sge.iq_start;
485 		q = sc->sge.iqmap[qid];
486 
487 		q->handler(q);
488 
489 		ndesc_total++;
490 		if (++ndesc_pending >= iq->qsize / 4) {
491 			t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
492 			    V_CIDXINC(ndesc_pending) |
493 			    V_INGRESSQID(iq->cntxt_id) |
494 			    V_SEINTARM(
495 				V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
496 			ndesc_pending = 0;
497 		}
498 
499 		iq_next(iq);
500 	}
501 
502 	if (ndesc_total > 0) {
503 		t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
504 		    V_CIDXINC(ndesc_pending) | V_INGRESSQID((u32)iq->cntxt_id) |
505 		    V_SEINTARM(iq->intr_params));
506 	}
507 
508 	atomic_cmpset_32(&iq->state, IQS_BUSY, IQS_IDLE);
509 }
510 
511 /* Deals with error interrupts */
512 void
513 t4_intr_err(void *arg)
514 {
515 	struct adapter *sc = arg;
516 
517 	if (sc->intr_type == INTR_INTX)
518 		t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
519 
520 	t4_slow_intr_handler(sc);
521 }
522 
523 /* Deals with the firmware event queue */
524 void
525 t4_intr_evt(void *arg)
526 {
527 	struct sge_iq *iq = arg;
528 
529 	if (!atomic_cmpset_32(&iq->state, IQS_IDLE, IQS_BUSY))
530 		return;
531 
532 	t4_evt_rx(arg);
533 
534 	atomic_cmpset_32(&iq->state, IQS_BUSY, IQS_IDLE);
535 }
536 
537 void
538 t4_intr_data(void *arg)
539 {
540 	struct sge_iq *iq = arg;
541 
542 	if (!atomic_cmpset_32(&iq->state, IQS_IDLE, IQS_BUSY))
543 		return;
544 
545 	t4_eth_rx(arg);
546 
547 	atomic_cmpset_32(&iq->state, IQS_BUSY, IQS_IDLE);
548 }
549 
550 void
551 t4_evt_rx(void *arg)
552 {
553 	struct sge_iq *iq = arg;
554 	struct adapter *sc = iq->adapter;
555 	struct rsp_ctrl *ctrl;
556 	const struct rss_header *rss;
557 	int ndesc_pending = 0, ndesc_total = 0;
558 
559 	KASSERT(iq == &sc->sge.fwq, ("%s: unexpected ingress queue", __func__));
560 
561 	while (is_new_response(iq, &ctrl)) {
562 
563 		rmb();
564 
565 		rss = (const void *)iq->cdesc;
566 
567 		/* Should only get CPL on this queue */
568 		KASSERT(G_RSPD_TYPE(ctrl->u.type_gen) == X_RSPD_TYPE_CPL,
569 		    ("%s: unexpected type %d", __func__,
570 		    G_RSPD_TYPE(ctrl->u.type_gen)));
571 
572 		switch (rss->opcode) {
573 		case CPL_FW4_MSG:
574 		case CPL_FW6_MSG: {
575 			const struct cpl_fw6_msg *cpl;
576 
577 			cpl = (const void *)(rss + 1);
578 			if (cpl->type == FW6_TYPE_CMD_RPL)
579 				t4_handle_fw_rpl(sc, cpl->data);
580 
581 			break;
582 			}
583 		case CPL_SGE_EGR_UPDATE:
584 			handle_sge_egr_update(sc, (const void *)(rss + 1));
585 			break;
586 
587 		default:
588 			device_printf(sc->dev,
589 			    "can't handle CPL opcode %d.", rss->opcode);
590 		}
591 
592 		ndesc_total++;
593 		if (++ndesc_pending >= iq->qsize / 4) {
594 			t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
595 			    V_CIDXINC(ndesc_pending) |
596 			    V_INGRESSQID(iq->cntxt_id) |
597 			    V_SEINTARM(
598 				V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
599 			ndesc_pending = 0;
600 		}
601 		iq_next(iq);
602 	}
603 
604 	if (ndesc_total > 0) {
605 		t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
606 		    V_CIDXINC(ndesc_pending) | V_INGRESSQID(iq->cntxt_id) |
607 		    V_SEINTARM(iq->intr_params));
608 	}
609 }
610 
611 void
612 t4_eth_rx(void *arg)
613 {
614 	struct sge_rxq *rxq = arg;
615 	struct sge_iq *iq = arg;
616 	struct adapter *sc = iq->adapter;
617 	struct rsp_ctrl *ctrl;
618 	struct ifnet *ifp = rxq->ifp;
619 	struct sge_fl *fl = &rxq->fl;
620 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx], *sd_next;
621 	const struct rss_header *rss;
622 	const struct cpl_rx_pkt *cpl;
623 	uint32_t len;
624 	int ndescs = 0, i;
625 	struct mbuf *m0, *m;
626 #ifdef INET
627 	struct lro_ctrl *lro = &rxq->lro;
628 	struct lro_entry *l;
629 #endif
630 
631 	prefetch(sd->m);
632 	prefetch(sd->cl);
633 
634 	iq->intr_next = iq->intr_params;
635 	while (is_new_response(iq, &ctrl)) {
636 
637 		rmb();
638 
639 		rss = (const void *)iq->cdesc;
640 		i = G_RSPD_TYPE(ctrl->u.type_gen);
641 
642 		if (__predict_false(i == X_RSPD_TYPE_CPL)) {
643 
644 			/* Can't be anything except an egress update */
645 			KASSERT(rss->opcode == CPL_SGE_EGR_UPDATE,
646 			    ("%s: unexpected CPL %x", __func__, rss->opcode));
647 
648 			handle_sge_egr_update(sc, (const void *)(rss + 1));
649 			goto nextdesc;
650 		}
651 		KASSERT(i == X_RSPD_TYPE_FLBUF && rss->opcode == CPL_RX_PKT,
652 		    ("%s: unexpected CPL %x rsp %d", __func__, rss->opcode, i));
653 
654 		sd_next = sd + 1;
655 		if (__predict_false(fl->cidx + 1 == fl->cap))
656 			sd_next = fl->sdesc;
657 		prefetch(sd_next->m);
658 		prefetch(sd_next->cl);
659 
660 		cpl = (const void *)(rss + 1);
661 
662 		m0 = sd->m;
663 		sd->m = NULL;	/* consumed */
664 
665 		len = be32toh(ctrl->pldbuflen_qid);
666 		if (__predict_false((len & F_RSPD_NEWBUF) == 0))
667 			panic("%s: cannot handle packed frames", __func__);
668 		len = G_RSPD_LEN(len);
669 
670 		bus_dmamap_sync(fl->tag[sd->tag_idx], sd->map,
671 		    BUS_DMASYNC_POSTREAD);
672 
673 		m_init(m0, NULL, 0, M_NOWAIT, MT_DATA, M_PKTHDR);
674 		if (len < MINCLSIZE) {
675 			/* copy data to mbuf, buffer will be recycled */
676 			bcopy(sd->cl, mtod(m0, caddr_t), len);
677 			m0->m_len = len;
678 		} else {
679 			bus_dmamap_unload(fl->tag[sd->tag_idx], sd->map);
680 			m_cljset(m0, sd->cl, FL_BUF_TYPE(sd->tag_idx));
681 			sd->cl = NULL;	/* consumed */
682 			m0->m_len = min(len, FL_BUF_SIZE(sd->tag_idx));
683 		}
684 
685 		len -= FL_PKTSHIFT;
686 		m0->m_len -= FL_PKTSHIFT;
687 		m0->m_data += FL_PKTSHIFT;
688 
689 		m0->m_pkthdr.len = len;
690 		m0->m_pkthdr.rcvif = ifp;
691 		m0->m_flags |= M_FLOWID;
692 		m0->m_pkthdr.flowid = rss->hash_val;
693 
694 		if (cpl->csum_calc && !cpl->err_vec &&
695 		    ifp->if_capenable & IFCAP_RXCSUM) {
696 			m0->m_pkthdr.csum_flags |= (CSUM_IP_CHECKED |
697 			    CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
698 			if (cpl->ip_frag)
699 				m0->m_pkthdr.csum_data = be16toh(cpl->csum);
700 			else
701 				m0->m_pkthdr.csum_data = 0xffff;
702 			rxq->rxcsum++;
703 		}
704 
705 		if (cpl->vlan_ex) {
706 			m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
707 			m0->m_flags |= M_VLANTAG;
708 			rxq->vlan_extraction++;
709 		}
710 
711 		i = 1;	/* # of fl sdesc used */
712 		sd = sd_next;
713 		if (__predict_false(++fl->cidx == fl->cap))
714 			fl->cidx = 0;
715 
716 		len -= m0->m_len;
717 		m = m0;
718 		while (len) {
719 			i++;
720 
721 			sd_next = sd + 1;
722 			if (__predict_false(fl->cidx + 1 == fl->cap))
723 				sd_next = fl->sdesc;
724 			prefetch(sd_next->m);
725 			prefetch(sd_next->cl);
726 
727 			m->m_next = sd->m;
728 			sd->m = NULL;	/* consumed */
729 			m = m->m_next;
730 
731 			bus_dmamap_sync(fl->tag[sd->tag_idx], sd->map,
732 			    BUS_DMASYNC_POSTREAD);
733 
734 			m_init(m, NULL, 0, M_NOWAIT, MT_DATA, 0);
735 			if (len <= MLEN) {
736 				bcopy(sd->cl, mtod(m, caddr_t), len);
737 				m->m_len = len;
738 			} else {
739 				bus_dmamap_unload(fl->tag[sd->tag_idx],
740 				    sd->map);
741 				m_cljset(m, sd->cl, FL_BUF_TYPE(sd->tag_idx));
742 				sd->cl = NULL;	/* consumed */
743 				m->m_len = min(len, FL_BUF_SIZE(sd->tag_idx));
744 			}
745 
746 			i++;
747 			sd = sd_next;
748 			if (__predict_false(++fl->cidx == fl->cap))
749 				fl->cidx = 0;
750 
751 			len -= m->m_len;
752 		}
753 
754 #ifdef INET
755 		if (cpl->l2info & htobe32(F_RXF_LRO) &&
756 		    rxq->flags & RXQ_LRO_ENABLED &&
757 		    tcp_lro_rx(lro, m0, 0) == 0) {
758 			/* queued for LRO */
759 		} else
760 #endif
761 		ifp->if_input(ifp, m0);
762 
763 		FL_LOCK(fl);
764 		fl->needed += i;
765 		if (fl->needed >= 32)
766 			refill_fl(fl, 64);
767 		if (fl->pending >= 32)
768 			ring_fl_db(sc, fl);
769 		FL_UNLOCK(fl);
770 
771 nextdesc:	ndescs++;
772 		iq_next(iq);
773 
774 		if (ndescs > 32) {
775 			t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
776 			    V_CIDXINC(ndescs) |
777 			    V_INGRESSQID((u32)iq->cntxt_id) |
778 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
779 			ndescs = 0;
780 		}
781 	}
782 
783 #ifdef INET
784 	while (!SLIST_EMPTY(&lro->lro_active)) {
785 		l = SLIST_FIRST(&lro->lro_active);
786 		SLIST_REMOVE_HEAD(&lro->lro_active, next);
787 		tcp_lro_flush(lro, l);
788 	}
789 #endif
790 
791 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_CIDXINC(ndescs) |
792 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_next));
793 
794 	FL_LOCK(fl);
795 	if (fl->needed >= 32)
796 		refill_fl(fl, 128);
797 	if (fl->pending >= 8)
798 		ring_fl_db(sc, fl);
799 	FL_UNLOCK(fl);
800 }
801 
802 int
803 t4_mgmt_tx(struct adapter *sc, struct mbuf *m)
804 {
805 	return ctrl_tx(sc, &sc->sge.ctrlq[0], m);
806 }
807 
808 /* Per-packet header in a coalesced tx WR, before the SGL starts (in flits) */
809 #define TXPKTS_PKT_HDR ((\
810     sizeof(struct ulp_txpkt) + \
811     sizeof(struct ulptx_idata) + \
812     sizeof(struct cpl_tx_pkt_core) \
813     ) / 8)
814 
815 /* Header of a coalesced tx WR, before SGL of first packet (in flits) */
816 #define TXPKTS_WR_HDR (\
817     sizeof(struct fw_eth_tx_pkts_wr) / 8 + \
818     TXPKTS_PKT_HDR)
819 
820 /* Header of a tx WR, before SGL of first packet (in flits) */
821 #define TXPKT_WR_HDR ((\
822     sizeof(struct fw_eth_tx_pkt_wr) + \
823     sizeof(struct cpl_tx_pkt_core) \
824     ) / 8 )
825 
826 /* Header of a tx LSO WR, before SGL of first packet (in flits) */
827 #define TXPKT_LSO_WR_HDR ((\
828     sizeof(struct fw_eth_tx_pkt_wr) + \
829     sizeof(struct cpl_tx_pkt_lso) + \
830     sizeof(struct cpl_tx_pkt_core) \
831     ) / 8 )
832 
833 int
834 t4_eth_tx(struct ifnet *ifp, struct sge_txq *txq, struct mbuf *m)
835 {
836 	struct port_info *pi = (void *)ifp->if_softc;
837 	struct adapter *sc = pi->adapter;
838 	struct sge_eq *eq = &txq->eq;
839 	struct buf_ring *br = txq->br;
840 	struct mbuf *next;
841 	int rc, coalescing, can_reclaim;
842 	struct txpkts txpkts;
843 	struct sgl sgl;
844 
845 	TXQ_LOCK_ASSERT_OWNED(txq);
846 	KASSERT(m, ("%s: called with nothing to do.", __func__));
847 
848 	prefetch(&eq->desc[eq->pidx]);
849 	prefetch(&txq->sdesc[eq->pidx]);
850 
851 	txpkts.npkt = 0;/* indicates there's nothing in txpkts */
852 	coalescing = 0;
853 
854 	if (eq->avail < 8)
855 		reclaim_tx_descs(txq, 0, 8);
856 
857 	for (; m; m = next ? next : drbr_dequeue(ifp, br)) {
858 
859 		if (eq->avail < 8)
860 			break;
861 
862 		next = m->m_nextpkt;
863 		m->m_nextpkt = NULL;
864 
865 		if (next || buf_ring_peek(br))
866 			coalescing = 1;
867 
868 		rc = get_pkt_sgl(txq, &m, &sgl, coalescing);
869 		if (rc != 0) {
870 			if (rc == ENOMEM) {
871 
872 				/* Short of resources, suspend tx */
873 
874 				m->m_nextpkt = next;
875 				break;
876 			}
877 
878 			/*
879 			 * Unrecoverable error for this packet, throw it away
880 			 * and move on to the next.  get_pkt_sgl may already
881 			 * have freed m (it will be NULL in that case and the
882 			 * m_freem here is still safe).
883 			 */
884 
885 			m_freem(m);
886 			continue;
887 		}
888 
889 		if (coalescing &&
890 		    add_to_txpkts(pi, txq, &txpkts, m, &sgl) == 0) {
891 
892 			/* Successfully absorbed into txpkts */
893 
894 			write_ulp_cpl_sgl(pi, txq, &txpkts, m, &sgl);
895 			goto doorbell;
896 		}
897 
898 		/*
899 		 * We weren't coalescing to begin with, or current frame could
900 		 * not be coalesced (add_to_txpkts flushes txpkts if a frame
901 		 * given to it can't be coalesced).  Either way there should be
902 		 * nothing in txpkts.
903 		 */
904 		KASSERT(txpkts.npkt == 0,
905 		    ("%s: txpkts not empty: %d", __func__, txpkts.npkt));
906 
907 		/* We're sending out individual packets now */
908 		coalescing = 0;
909 
910 		if (eq->avail < 8)
911 			reclaim_tx_descs(txq, 0, 8);
912 		rc = write_txpkt_wr(pi, txq, m, &sgl);
913 		if (rc != 0) {
914 
915 			/* Short of hardware descriptors, suspend tx */
916 
917 			/*
918 			 * This is an unlikely but expensive failure.  We've
919 			 * done all the hard work (DMA mappings etc.) and now we
920 			 * can't send out the packet.  What's worse, we have to
921 			 * spend even more time freeing up everything in sgl.
922 			 */
923 			txq->no_desc++;
924 			free_pkt_sgl(txq, &sgl);
925 
926 			m->m_nextpkt = next;
927 			break;
928 		}
929 
930 		ETHER_BPF_MTAP(ifp, m);
931 		if (sgl.nsegs == 0)
932 			m_freem(m);
933 
934 doorbell:
935 		/* Fewer and fewer doorbells as the queue fills up */
936 		if (eq->pending >= (1 << (fls(eq->qsize - eq->avail) / 2)))
937 		    ring_eq_db(sc, eq);
938 
939 		can_reclaim = reclaimable(eq);
940 		if (can_reclaim >= 32)
941 			reclaim_tx_descs(txq, can_reclaim, 32);
942 	}
943 
944 	if (txpkts.npkt > 0)
945 		write_txpkts_wr(txq, &txpkts);
946 
947 	/*
948 	 * m not NULL means there was an error but we haven't thrown it away.
949 	 * This can happen when we're short of tx descriptors (no_desc) or maybe
950 	 * even DMA maps (no_dmamap).  Either way, a credit flush and reclaim
951 	 * will get things going again.
952 	 *
953 	 * If eq->avail is already 0 we know a credit flush was requested in the
954 	 * WR that reduced it to 0 so we don't need another flush (we don't have
955 	 * any descriptor for a flush WR anyway, duh).
956 	 */
957 	if (m && eq->avail > 0 && !(eq->flags & EQ_CRFLUSHED)) {
958 		struct tx_sdesc *txsd = &txq->sdesc[eq->pidx];
959 
960 		txsd->desc_used = 1;
961 		txsd->credits = 0;
962 		write_eqflush_wr(eq);
963 	}
964 	txq->m = m;
965 
966 	if (eq->pending)
967 		ring_eq_db(sc, eq);
968 
969 	can_reclaim = reclaimable(eq);
970 	if (can_reclaim >= 32)
971 		reclaim_tx_descs(txq, can_reclaim, 128);
972 
973 	return (0);
974 }
975 
976 void
977 t4_update_fl_bufsize(struct ifnet *ifp)
978 {
979 	struct port_info *pi = ifp->if_softc;
980 	struct sge_rxq *rxq;
981 	struct sge_fl *fl;
982 	int i;
983 
984 	for_each_rxq(pi, i, rxq) {
985 		fl = &rxq->fl;
986 
987 		FL_LOCK(fl);
988 		set_fl_tag_idx(fl, ifp->if_mtu);
989 		FL_UNLOCK(fl);
990 	}
991 }
992 
993 /*
994  * A non-NULL handler indicates this iq will not receive direct interrupts, the
995  * handler will be invoked by a forwarded interrupt queue.
996  */
997 static inline void
998 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
999     int qsize, int esize, iq_intr_handler_t *handler, char *name)
1000 {
1001 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
1002 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
1003 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
1004 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
1005 
1006 	iq->flags = 0;
1007 	iq->adapter = sc;
1008 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx) |
1009 	    V_QINTR_CNT_EN(pktc_idx >= 0);
1010 	iq->intr_pktc_idx = pktc_idx;
1011 	iq->qsize = roundup(qsize, 16);		/* See FW_IQ_CMD/iqsize */
1012 	iq->esize = max(esize, 16);		/* See FW_IQ_CMD/iqesize */
1013 	iq->handler = handler;
1014 	strlcpy(iq->lockname, name, sizeof(iq->lockname));
1015 }
1016 
1017 static inline void
1018 init_fl(struct sge_fl *fl, int qsize, char *name)
1019 {
1020 	fl->qsize = qsize;
1021 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
1022 }
1023 
1024 static inline void
1025 init_eq(struct sge_eq *eq, int qsize, char *name)
1026 {
1027 	eq->qsize = qsize;
1028 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
1029 }
1030 
1031 static int
1032 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
1033     bus_dmamap_t *map, bus_addr_t *pa, void **va)
1034 {
1035 	int rc;
1036 
1037 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
1038 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
1039 	if (rc != 0) {
1040 		device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
1041 		goto done;
1042 	}
1043 
1044 	rc = bus_dmamem_alloc(*tag, va,
1045 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
1046 	if (rc != 0) {
1047 		device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
1048 		goto done;
1049 	}
1050 
1051 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
1052 	if (rc != 0) {
1053 		device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
1054 		goto done;
1055 	}
1056 done:
1057 	if (rc)
1058 		free_ring(sc, *tag, *map, *pa, *va);
1059 
1060 	return (rc);
1061 }
1062 
1063 static int
1064 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
1065     bus_addr_t pa, void *va)
1066 {
1067 	if (pa)
1068 		bus_dmamap_unload(tag, map);
1069 	if (va)
1070 		bus_dmamem_free(tag, va, map);
1071 	if (tag)
1072 		bus_dma_tag_destroy(tag);
1073 
1074 	return (0);
1075 }
1076 
1077 /*
1078  * Allocates the ring for an ingress queue and an optional freelist.  If the
1079  * freelist is specified it will be allocated and then associated with the
1080  * ingress queue.
1081  *
1082  * Returns errno on failure.  Resources allocated up to that point may still be
1083  * allocated.  Caller is responsible for cleanup in case this function fails.
1084  *
1085  * If the ingress queue will take interrupts directly (iq->handler == NULL) then
1086  * the intr_idx specifies the vector, starting from 0.  Otherwise it specifies
1087  * the index of the queue to which its interrupts will be forwarded.
1088  */
1089 static int
1090 alloc_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl,
1091     int intr_idx)
1092 {
1093 	int rc, i, cntxt_id;
1094 	size_t len;
1095 	struct fw_iq_cmd c;
1096 	struct adapter *sc = iq->adapter;
1097 	__be32 v = 0;
1098 
1099 	/* The adapter queues are nominally allocated in port[0]'s name */
1100 	if (pi == NULL)
1101 		pi = sc->port[0];
1102 
1103 	len = iq->qsize * iq->esize;
1104 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
1105 	    (void **)&iq->desc);
1106 	if (rc != 0)
1107 		return (rc);
1108 
1109 	bzero(&c, sizeof(c));
1110 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
1111 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
1112 	    V_FW_IQ_CMD_VFN(0));
1113 
1114 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
1115 	    FW_LEN16(c));
1116 
1117 	/* Special handling for firmware event queue */
1118 	if (iq == &sc->sge.fwq)
1119 		v |= F_FW_IQ_CMD_IQASYNCH;
1120 
1121 	if (iq->handler) {
1122 		KASSERT(intr_idx < NFIQ(sc),
1123 		    ("%s: invalid indirect intr_idx %d", __func__, intr_idx));
1124 		v |= F_FW_IQ_CMD_IQANDST;
1125 		v |= V_FW_IQ_CMD_IQANDSTINDEX(sc->sge.fiq[intr_idx].abs_id);
1126 	} else {
1127 		KASSERT(intr_idx < sc->intr_count,
1128 		    ("%s: invalid direct intr_idx %d", __func__, intr_idx));
1129 		v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
1130 	}
1131 
1132 	c.type_to_iqandstindex = htobe32(v |
1133 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
1134 	    V_FW_IQ_CMD_VIID(pi->viid) |
1135 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
1136 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
1137 	    F_FW_IQ_CMD_IQGTSMODE |
1138 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
1139 	    V_FW_IQ_CMD_IQESIZE(ilog2(iq->esize) - 4));
1140 	c.iqsize = htobe16(iq->qsize);
1141 	c.iqaddr = htobe64(iq->ba);
1142 
1143 	if (fl) {
1144 		mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
1145 
1146 		for (i = 0; i < FL_BUF_SIZES; i++) {
1147 
1148 			/*
1149 			 * A freelist buffer must be 16 byte aligned as the SGE
1150 			 * uses the low 4 bits of the bus addr to figure out the
1151 			 * buffer size.
1152 			 */
1153 			rc = bus_dma_tag_create(sc->dmat, 16, 0,
1154 			    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
1155 			    FL_BUF_SIZE(i), 1, FL_BUF_SIZE(i), BUS_DMA_ALLOCNOW,
1156 			    NULL, NULL, &fl->tag[i]);
1157 			if (rc != 0) {
1158 				device_printf(sc->dev,
1159 				    "failed to create fl DMA tag[%d]: %d\n",
1160 				    i, rc);
1161 				return (rc);
1162 			}
1163 		}
1164 		len = fl->qsize * RX_FL_ESIZE;
1165 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
1166 		    &fl->ba, (void **)&fl->desc);
1167 		if (rc)
1168 			return (rc);
1169 
1170 		/* Allocate space for one software descriptor per buffer. */
1171 		fl->cap = (fl->qsize - SPG_LEN / RX_FL_ESIZE) * 8;
1172 		FL_LOCK(fl);
1173 		set_fl_tag_idx(fl, pi->ifp->if_mtu);
1174 		rc = alloc_fl_sdesc(fl);
1175 		FL_UNLOCK(fl);
1176 		if (rc != 0) {
1177 			device_printf(sc->dev,
1178 			    "failed to setup fl software descriptors: %d\n",
1179 			    rc);
1180 			return (rc);
1181 		}
1182 		fl->needed = fl->cap - 1; /* one less to avoid cidx = pidx */
1183 
1184 		c.iqns_to_fl0congen =
1185 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE));
1186 		c.fl0dcaen_to_fl0cidxfthresh =
1187 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(X_FETCHBURSTMIN_64B) |
1188 			V_FW_IQ_CMD_FL0FBMAX(X_FETCHBURSTMAX_512B));
1189 		c.fl0size = htobe16(fl->qsize);
1190 		c.fl0addr = htobe64(fl->ba);
1191 	}
1192 
1193 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
1194 	if (rc != 0) {
1195 		device_printf(sc->dev,
1196 		    "failed to create ingress queue: %d\n", rc);
1197 		return (rc);
1198 	}
1199 
1200 	iq->cdesc = iq->desc;
1201 	iq->cidx = 0;
1202 	iq->gen = 1;
1203 	iq->intr_next = iq->intr_params;
1204 	iq->cntxt_id = be16toh(c.iqid);
1205 	iq->abs_id = be16toh(c.physiqid);
1206 	iq->flags |= (IQ_ALLOCATED | IQ_STARTED);
1207 
1208 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
1209 	KASSERT(cntxt_id < sc->sge.niq,
1210 	    ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
1211 	    cntxt_id, sc->sge.niq - 1));
1212 	sc->sge.iqmap[cntxt_id] = iq;
1213 
1214 	if (fl) {
1215 		fl->cntxt_id = be16toh(c.fl0id);
1216 		fl->pidx = fl->cidx = 0;
1217 
1218 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
1219 		KASSERT(cntxt_id < sc->sge.neq,
1220 		    ("%s: fl->cntxt_id (%d) more than the max (%d)", __func__,
1221 		    cntxt_id, sc->sge.neq - 1));
1222 		sc->sge.eqmap[cntxt_id] = (void *)fl;
1223 
1224 		FL_LOCK(fl);
1225 		refill_fl(fl, -1);
1226 		if (fl->pending >= 8)
1227 			ring_fl_db(sc, fl);
1228 		FL_UNLOCK(fl);
1229 	}
1230 
1231 	/* Enable IQ interrupts */
1232 	atomic_store_rel_32(&iq->state, IQS_IDLE);
1233 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_SEINTARM(iq->intr_params) |
1234 	    V_INGRESSQID(iq->cntxt_id));
1235 
1236 	return (0);
1237 }
1238 
1239 /*
1240  * This can be called with the iq/fl in any state - fully allocated and
1241  * functional, partially allocated, even all-zeroed out.
1242  */
1243 static int
1244 free_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl)
1245 {
1246 	int i, rc;
1247 	struct adapter *sc = iq->adapter;
1248 	device_t dev;
1249 
1250 	if (sc == NULL)
1251 		return (0);	/* nothing to do */
1252 
1253 	dev = pi ? pi->dev : sc->dev;
1254 
1255 	if (iq->flags & IQ_STARTED) {
1256 		rc = -t4_iq_start_stop(sc, sc->mbox, 0, sc->pf, 0,
1257 		    iq->cntxt_id, fl ? fl->cntxt_id : 0xffff, 0xffff);
1258 		if (rc != 0) {
1259 			device_printf(dev,
1260 			    "failed to stop queue %p: %d\n", iq, rc);
1261 			return (rc);
1262 		}
1263 		iq->flags &= ~IQ_STARTED;
1264 
1265 		/* Synchronize with the interrupt handler */
1266 		while (!atomic_cmpset_32(&iq->state, IQS_IDLE, IQS_DISABLED))
1267 			pause("iqfree", hz / 1000);
1268 	}
1269 
1270 	if (iq->flags & IQ_ALLOCATED) {
1271 
1272 		rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
1273 		    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
1274 		    fl ? fl->cntxt_id : 0xffff, 0xffff);
1275 		if (rc != 0) {
1276 			device_printf(dev,
1277 			    "failed to free queue %p: %d\n", iq, rc);
1278 			return (rc);
1279 		}
1280 		iq->flags &= ~IQ_ALLOCATED;
1281 	}
1282 
1283 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
1284 
1285 	bzero(iq, sizeof(*iq));
1286 
1287 	if (fl) {
1288 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
1289 		    fl->desc);
1290 
1291 		if (fl->sdesc) {
1292 			FL_LOCK(fl);
1293 			free_fl_sdesc(fl);
1294 			FL_UNLOCK(fl);
1295 		}
1296 
1297 		if (mtx_initialized(&fl->fl_lock))
1298 			mtx_destroy(&fl->fl_lock);
1299 
1300 		for (i = 0; i < FL_BUF_SIZES; i++) {
1301 			if (fl->tag[i])
1302 				bus_dma_tag_destroy(fl->tag[i]);
1303 		}
1304 
1305 		bzero(fl, sizeof(*fl));
1306 	}
1307 
1308 	return (0);
1309 }
1310 
1311 static int
1312 alloc_iq(struct sge_iq *iq, int intr_idx)
1313 {
1314 	return alloc_iq_fl(NULL, iq, NULL, intr_idx);
1315 }
1316 
1317 static int
1318 free_iq(struct sge_iq *iq)
1319 {
1320 	return free_iq_fl(NULL, iq, NULL);
1321 }
1322 
1323 static int
1324 alloc_rxq(struct port_info *pi, struct sge_rxq *rxq, int intr_idx, int idx)
1325 {
1326 	int rc;
1327 	struct sysctl_oid *oid;
1328 	struct sysctl_oid_list *children;
1329 	char name[16];
1330 
1331 	rc = alloc_iq_fl(pi, &rxq->iq, &rxq->fl, intr_idx);
1332 	if (rc != 0)
1333 		return (rc);
1334 
1335 #ifdef INET
1336 	rc = tcp_lro_init(&rxq->lro);
1337 	if (rc != 0)
1338 		return (rc);
1339 	rxq->lro.ifp = pi->ifp; /* also indicates LRO init'ed */
1340 
1341 	if (pi->ifp->if_capenable & IFCAP_LRO)
1342 		rxq->flags |= RXQ_LRO_ENABLED;
1343 #endif
1344 	rxq->ifp = pi->ifp;
1345 
1346 	children = SYSCTL_CHILDREN(pi->oid_rxq);
1347 
1348 	snprintf(name, sizeof(name), "%d", idx);
1349 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
1350 	    NULL, "rx queue");
1351 	children = SYSCTL_CHILDREN(oid);
1352 
1353 #ifdef INET
1354 	SYSCTL_ADD_INT(&pi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
1355 	    &rxq->lro.lro_queued, 0, NULL);
1356 	SYSCTL_ADD_INT(&pi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
1357 	    &rxq->lro.lro_flushed, 0, NULL);
1358 #endif
1359 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
1360 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
1361 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "vlan_extraction",
1362 	    CTLFLAG_RD, &rxq->vlan_extraction,
1363 	    "# of times hardware extracted 802.1Q tag");
1364 
1365 	return (rc);
1366 }
1367 
1368 static int
1369 free_rxq(struct port_info *pi, struct sge_rxq *rxq)
1370 {
1371 	int rc;
1372 
1373 #ifdef INET
1374 	if (rxq->lro.ifp) {
1375 		tcp_lro_free(&rxq->lro);
1376 		rxq->lro.ifp = NULL;
1377 	}
1378 #endif
1379 
1380 	rc = free_iq_fl(pi, &rxq->iq, &rxq->fl);
1381 	if (rc == 0)
1382 		bzero(rxq, sizeof(*rxq));
1383 
1384 	return (rc);
1385 }
1386 
1387 static int
1388 alloc_ctrlq(struct adapter *sc, struct sge_ctrlq *ctrlq, int idx)
1389 {
1390 	int rc, cntxt_id;
1391 	size_t len;
1392 	struct fw_eq_ctrl_cmd c;
1393 	struct sge_eq *eq = &ctrlq->eq;
1394 	char name[16];
1395 	struct sysctl_oid *oid;
1396 	struct sysctl_oid_list *children;
1397 
1398 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
1399 
1400 	len = eq->qsize * CTRL_EQ_ESIZE;
1401 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
1402 	    &eq->ba, (void **)&eq->desc);
1403 	if (rc)
1404 		return (rc);
1405 
1406 	eq->cap = eq->qsize - SPG_LEN / CTRL_EQ_ESIZE;
1407 	eq->spg = (void *)&eq->desc[eq->cap];
1408 	eq->avail = eq->cap - 1;	/* one less to avoid cidx = pidx */
1409 	eq->iqid = sc->sge.fwq.cntxt_id;
1410 
1411 	bzero(&c, sizeof(c));
1412 
1413 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
1414 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
1415 	    V_FW_EQ_CTRL_CMD_VFN(0));
1416 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
1417 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
1418 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid)); /* XXX */
1419 	c.physeqid_pkd = htobe32(0);
1420 	c.fetchszm_to_iqid =
1421 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
1422 		V_FW_EQ_CTRL_CMD_PCIECHN(idx) |
1423 		V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
1424 	c.dcaen_to_eqsize =
1425 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
1426 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
1427 		V_FW_EQ_CTRL_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
1428 		V_FW_EQ_CTRL_CMD_EQSIZE(eq->qsize));
1429 	c.eqaddr = htobe64(eq->ba);
1430 
1431 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
1432 	if (rc != 0) {
1433 		device_printf(sc->dev,
1434 		    "failed to create control queue %d: %d\n", idx, rc);
1435 		return (rc);
1436 	}
1437 
1438 	eq->pidx = eq->cidx = 0;
1439 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
1440 	eq->flags |= (EQ_ALLOCATED | EQ_STARTED);
1441 
1442 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
1443 	KASSERT(cntxt_id < sc->sge.neq,
1444 	    ("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
1445 	    cntxt_id, sc->sge.neq - 1));
1446 	sc->sge.eqmap[cntxt_id] = eq;
1447 
1448 	children = SYSCTL_CHILDREN(sc->oid_ctrlq);
1449 
1450 	snprintf(name, sizeof(name), "%d", idx);
1451 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, name, CTLFLAG_RD,
1452 	    NULL, "ctrl queue");
1453 	children = SYSCTL_CHILDREN(oid);
1454 
1455 	SYSCTL_ADD_UQUAD(&sc->ctx, children, OID_AUTO, "total_wrs", CTLFLAG_RD,
1456 	    &ctrlq->total_wrs, "total # of work requests");
1457 	SYSCTL_ADD_UINT(&sc->ctx, children, OID_AUTO, "no_desc", CTLFLAG_RD,
1458 	    &ctrlq->no_desc, 0,
1459 	    "# of times ctrlq ran out of hardware descriptors");
1460 	SYSCTL_ADD_UINT(&sc->ctx, children, OID_AUTO, "too_long", CTLFLAG_RD,
1461 	    &ctrlq->too_long, 0, "# of oversized work requests");
1462 
1463 	return (rc);
1464 }
1465 
1466 static int
1467 free_ctrlq(struct adapter *sc, struct sge_ctrlq *ctrlq)
1468 {
1469 	int rc;
1470 	struct sge_eq *eq = &ctrlq->eq;
1471 
1472 	if (eq->flags & (EQ_ALLOCATED | EQ_STARTED)) {
1473 		rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0, eq->cntxt_id);
1474 		if (rc != 0) {
1475 			device_printf(sc->dev,
1476 			    "failed to free ctrl queue %p: %d\n", eq, rc);
1477 			return (rc);
1478 		}
1479 		eq->flags &= ~(EQ_ALLOCATED | EQ_STARTED);
1480 	}
1481 
1482 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
1483 
1484 	if (mtx_initialized(&eq->eq_lock))
1485 		mtx_destroy(&eq->eq_lock);
1486 
1487 	bzero(ctrlq, sizeof(*ctrlq));
1488 	return (0);
1489 }
1490 
1491 static int
1492 alloc_txq(struct port_info *pi, struct sge_txq *txq, int idx)
1493 {
1494 	int rc, cntxt_id;
1495 	size_t len;
1496 	struct adapter *sc = pi->adapter;
1497 	struct fw_eq_eth_cmd c;
1498 	struct sge_eq *eq = &txq->eq;
1499 	char name[16];
1500 	struct sysctl_oid *oid;
1501 	struct sysctl_oid_list *children;
1502 
1503 	txq->ifp = pi->ifp;
1504 	TASK_INIT(&txq->resume_tx, 0, cxgbe_txq_start, txq);
1505 
1506 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
1507 
1508 	len = eq->qsize * TX_EQ_ESIZE;
1509 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
1510 	    &eq->ba, (void **)&eq->desc);
1511 	if (rc)
1512 		return (rc);
1513 
1514 	eq->cap = eq->qsize - SPG_LEN / TX_EQ_ESIZE;
1515 	eq->spg = (void *)&eq->desc[eq->cap];
1516 	eq->avail = eq->cap - 1;	/* one less to avoid cidx = pidx */
1517 	txq->sdesc = malloc(eq->cap * sizeof(struct tx_sdesc), M_CXGBE,
1518 	    M_ZERO | M_WAITOK);
1519 	txq->br = buf_ring_alloc(eq->qsize, M_CXGBE, M_WAITOK, &eq->eq_lock);
1520 	eq->iqid = sc->sge.rxq[pi->first_rxq].iq.cntxt_id;
1521 
1522 	rc = bus_dma_tag_create(sc->dmat, 1, 0, BUS_SPACE_MAXADDR,
1523 	    BUS_SPACE_MAXADDR, NULL, NULL, 64 * 1024, TX_SGL_SEGS,
1524 	    BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL, NULL, &txq->tx_tag);
1525 	if (rc != 0) {
1526 		device_printf(sc->dev,
1527 		    "failed to create tx DMA tag: %d\n", rc);
1528 		return (rc);
1529 	}
1530 
1531 	rc = alloc_tx_maps(txq);
1532 	if (rc != 0) {
1533 		device_printf(sc->dev, "failed to setup tx DMA maps: %d\n", rc);
1534 		return (rc);
1535 	}
1536 
1537 	bzero(&c, sizeof(c));
1538 
1539 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
1540 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
1541 	    V_FW_EQ_ETH_CMD_VFN(0));
1542 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
1543 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
1544 	c.viid_pkd = htobe32(V_FW_EQ_ETH_CMD_VIID(pi->viid));
1545 	c.fetchszm_to_iqid =
1546 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
1547 		V_FW_EQ_ETH_CMD_PCIECHN(pi->tx_chan) |
1548 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
1549 	c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
1550 		      V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
1551 		      V_FW_EQ_ETH_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
1552 		      V_FW_EQ_ETH_CMD_EQSIZE(eq->qsize));
1553 	c.eqaddr = htobe64(eq->ba);
1554 
1555 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
1556 	if (rc != 0) {
1557 		device_printf(pi->dev,
1558 		    "failed to create egress queue: %d\n", rc);
1559 		return (rc);
1560 	}
1561 
1562 	eq->pidx = eq->cidx = 0;
1563 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
1564 	eq->flags |= (EQ_ALLOCATED | EQ_STARTED);
1565 
1566 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
1567 	KASSERT(cntxt_id < sc->sge.neq,
1568 	    ("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
1569 	    cntxt_id, sc->sge.neq - 1));
1570 	sc->sge.eqmap[cntxt_id] = eq;
1571 
1572 	children = SYSCTL_CHILDREN(pi->oid_txq);
1573 
1574 	snprintf(name, sizeof(name), "%d", idx);
1575 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
1576 	    NULL, "tx queue");
1577 	children = SYSCTL_CHILDREN(oid);
1578 
1579 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
1580 	    &txq->txcsum, "# of times hardware assisted with checksum");
1581 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "vlan_insertion",
1582 	    CTLFLAG_RD, &txq->vlan_insertion,
1583 	    "# of times hardware inserted 802.1Q tag");
1584 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
1585 	    &txq->tso_wrs, "# of IPv4 TSO work requests");
1586 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
1587 	    &txq->imm_wrs, "# of work requests with immediate data");
1588 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
1589 	    &txq->sgl_wrs, "# of work requests with direct SGL");
1590 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
1591 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
1592 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txpkts_wrs", CTLFLAG_RD,
1593 	    &txq->txpkts_wrs, "# of txpkts work requests (multiple pkts/WR)");
1594 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txpkts_pkts", CTLFLAG_RD,
1595 	    &txq->txpkts_pkts, "# of frames tx'd using txpkts work requests");
1596 
1597 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "no_dmamap", CTLFLAG_RD,
1598 	    &txq->no_dmamap, 0, "# of times txq ran out of DMA maps");
1599 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "no_desc", CTLFLAG_RD,
1600 	    &txq->no_desc, 0, "# of times txq ran out of hardware descriptors");
1601 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "egr_update", CTLFLAG_RD,
1602 	    &txq->egr_update, 0, "egress update notifications from the SGE");
1603 
1604 	return (rc);
1605 }
1606 
1607 static int
1608 free_txq(struct port_info *pi, struct sge_txq *txq)
1609 {
1610 	int rc;
1611 	struct adapter *sc = pi->adapter;
1612 	struct sge_eq *eq = &txq->eq;
1613 
1614 	if (eq->flags & (EQ_ALLOCATED | EQ_STARTED)) {
1615 
1616 		/*
1617 		 * Wait for the response to a credit flush if there's one
1618 		 * pending.  Clearing the flag tells handle_sge_egr_update or
1619 		 * cxgbe_txq_start (depending on how far the response has made
1620 		 * it) that they should ignore the response and wake up free_txq
1621 		 * instead.
1622 		 *
1623 		 * The interface has been marked down by the time we get here
1624 		 * (both IFF_UP and IFF_DRV_RUNNING cleared).  qflush has
1625 		 * emptied the tx buf_rings and we know nothing new is being
1626 		 * queued for tx so we don't have to worry about a new credit
1627 		 * flush request.
1628 		 */
1629 		TXQ_LOCK(txq);
1630 		if (eq->flags & EQ_CRFLUSHED) {
1631 			eq->flags &= ~EQ_CRFLUSHED;
1632 			msleep(txq, &eq->eq_lock, 0, "crflush", 0);
1633 		}
1634 		TXQ_UNLOCK(txq);
1635 
1636 		rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0, eq->cntxt_id);
1637 		if (rc != 0) {
1638 			device_printf(pi->dev,
1639 			    "failed to free egress queue %p: %d\n", eq, rc);
1640 			return (rc);
1641 		}
1642 		eq->flags &= ~(EQ_ALLOCATED | EQ_STARTED);
1643 	}
1644 
1645 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
1646 
1647 	free(txq->sdesc, M_CXGBE);
1648 
1649 	if (txq->maps)
1650 		free_tx_maps(txq);
1651 
1652 	buf_ring_free(txq->br, M_CXGBE);
1653 
1654 	if (txq->tx_tag)
1655 		bus_dma_tag_destroy(txq->tx_tag);
1656 
1657 	if (mtx_initialized(&eq->eq_lock))
1658 		mtx_destroy(&eq->eq_lock);
1659 
1660 	bzero(txq, sizeof(*txq));
1661 	return (0);
1662 }
1663 
1664 static void
1665 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1666 {
1667 	bus_addr_t *ba = arg;
1668 
1669 	KASSERT(nseg == 1,
1670 	    ("%s meant for single segment mappings only.", __func__));
1671 
1672 	*ba = error ? 0 : segs->ds_addr;
1673 }
1674 
1675 static inline bool
1676 is_new_response(const struct sge_iq *iq, struct rsp_ctrl **ctrl)
1677 {
1678 	*ctrl = (void *)((uintptr_t)iq->cdesc +
1679 	    (iq->esize - sizeof(struct rsp_ctrl)));
1680 
1681 	return (((*ctrl)->u.type_gen >> S_RSPD_GEN) == iq->gen);
1682 }
1683 
1684 static inline void
1685 iq_next(struct sge_iq *iq)
1686 {
1687 	iq->cdesc = (void *) ((uintptr_t)iq->cdesc + iq->esize);
1688 	if (__predict_false(++iq->cidx == iq->qsize - 1)) {
1689 		iq->cidx = 0;
1690 		iq->gen ^= 1;
1691 		iq->cdesc = iq->desc;
1692 	}
1693 }
1694 
1695 static inline void
1696 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
1697 {
1698 	int ndesc = fl->pending / 8;
1699 
1700 	/* Caller responsible for ensuring there's something useful to do */
1701 	KASSERT(ndesc > 0, ("%s called with no useful work to do.", __func__));
1702 
1703 	wmb();
1704 
1705 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL), F_DBPRIO |
1706 	    V_QID(fl->cntxt_id) | V_PIDX(ndesc));
1707 
1708 	fl->pending &= 7;
1709 }
1710 
1711 static void
1712 refill_fl(struct sge_fl *fl, int nbufs)
1713 {
1714 	__be64 *d = &fl->desc[fl->pidx];
1715 	struct fl_sdesc *sd = &fl->sdesc[fl->pidx];
1716 	bus_dma_tag_t tag;
1717 	bus_addr_t pa;
1718 	caddr_t cl;
1719 	int rc;
1720 
1721 	FL_LOCK_ASSERT_OWNED(fl);
1722 
1723 	if (nbufs < 0 || nbufs > fl->needed)
1724 		nbufs = fl->needed;
1725 
1726 	while (nbufs--) {
1727 
1728 		if (sd->cl != NULL) {
1729 
1730 			/*
1731 			 * This happens when a frame small enough to fit
1732 			 * entirely in an mbuf was received in cl last time.
1733 			 * We'd held on to cl and can reuse it now.  Note that
1734 			 * we reuse a cluster of the old size if fl->tag_idx is
1735 			 * no longer the same as sd->tag_idx.
1736 			 */
1737 
1738 			KASSERT(*d == sd->ba_tag,
1739 			    ("%s: recyling problem at pidx %d",
1740 			    __func__, fl->pidx));
1741 
1742 			d++;
1743 			goto recycled;
1744 		}
1745 
1746 
1747 		if (fl->tag_idx != sd->tag_idx) {
1748 			bus_dmamap_t map;
1749 			bus_dma_tag_t newtag = fl->tag[fl->tag_idx];
1750 			bus_dma_tag_t oldtag = fl->tag[sd->tag_idx];
1751 
1752 			/*
1753 			 * An MTU change can get us here.  Discard the old map
1754 			 * which was created with the old tag, but only if
1755 			 * we're able to get a new one.
1756 			 */
1757 			rc = bus_dmamap_create(newtag, 0, &map);
1758 			if (rc == 0) {
1759 				bus_dmamap_destroy(oldtag, sd->map);
1760 				sd->map = map;
1761 				sd->tag_idx = fl->tag_idx;
1762 			}
1763 		}
1764 
1765 		tag = fl->tag[sd->tag_idx];
1766 
1767 		cl = m_cljget(NULL, M_NOWAIT, FL_BUF_SIZE(sd->tag_idx));
1768 		if (cl == NULL)
1769 			break;
1770 
1771 		rc = bus_dmamap_load(tag, sd->map, cl, FL_BUF_SIZE(sd->tag_idx),
1772 		    oneseg_dma_callback, &pa, 0);
1773 		if (rc != 0 || pa == 0) {
1774 			fl->dmamap_failed++;
1775 			uma_zfree(FL_BUF_ZONE(sd->tag_idx), cl);
1776 			break;
1777 		}
1778 
1779 		sd->cl = cl;
1780 		*d++ = htobe64(pa | sd->tag_idx);
1781 
1782 #ifdef INVARIANTS
1783 		sd->ba_tag = htobe64(pa | sd->tag_idx);
1784 #endif
1785 
1786 recycled:
1787 		/* sd->m is never recycled, should always be NULL */
1788 		KASSERT(sd->m == NULL, ("%s: stray mbuf", __func__));
1789 
1790 		sd->m = m_gethdr(M_NOWAIT, MT_NOINIT);
1791 		if (sd->m == NULL)
1792 			break;
1793 
1794 		fl->pending++;
1795 		fl->needed--;
1796 		sd++;
1797 		if (++fl->pidx == fl->cap) {
1798 			fl->pidx = 0;
1799 			sd = fl->sdesc;
1800 			d = fl->desc;
1801 		}
1802 	}
1803 }
1804 
1805 static int
1806 alloc_fl_sdesc(struct sge_fl *fl)
1807 {
1808 	struct fl_sdesc *sd;
1809 	bus_dma_tag_t tag;
1810 	int i, rc;
1811 
1812 	FL_LOCK_ASSERT_OWNED(fl);
1813 
1814 	fl->sdesc = malloc(fl->cap * sizeof(struct fl_sdesc), M_CXGBE,
1815 	    M_ZERO | M_WAITOK);
1816 
1817 	tag = fl->tag[fl->tag_idx];
1818 	sd = fl->sdesc;
1819 	for (i = 0; i < fl->cap; i++, sd++) {
1820 
1821 		sd->tag_idx = fl->tag_idx;
1822 		rc = bus_dmamap_create(tag, 0, &sd->map);
1823 		if (rc != 0)
1824 			goto failed;
1825 	}
1826 
1827 	return (0);
1828 failed:
1829 	while (--i >= 0) {
1830 		sd--;
1831 		bus_dmamap_destroy(tag, sd->map);
1832 		if (sd->m) {
1833 			m_init(sd->m, NULL, 0, M_NOWAIT, MT_DATA, 0);
1834 			m_free(sd->m);
1835 			sd->m = NULL;
1836 		}
1837 	}
1838 	KASSERT(sd == fl->sdesc, ("%s: EDOOFUS", __func__));
1839 
1840 	free(fl->sdesc, M_CXGBE);
1841 	fl->sdesc = NULL;
1842 
1843 	return (rc);
1844 }
1845 
1846 static void
1847 free_fl_sdesc(struct sge_fl *fl)
1848 {
1849 	struct fl_sdesc *sd;
1850 	int i;
1851 
1852 	FL_LOCK_ASSERT_OWNED(fl);
1853 
1854 	sd = fl->sdesc;
1855 	for (i = 0; i < fl->cap; i++, sd++) {
1856 
1857 		if (sd->m) {
1858 			m_init(sd->m, NULL, 0, M_NOWAIT, MT_DATA, 0);
1859 			m_free(sd->m);
1860 			sd->m = NULL;
1861 		}
1862 
1863 		if (sd->cl) {
1864 			bus_dmamap_unload(fl->tag[sd->tag_idx], sd->map);
1865 			uma_zfree(FL_BUF_ZONE(sd->tag_idx), sd->cl);
1866 			sd->cl = NULL;
1867 		}
1868 
1869 		bus_dmamap_destroy(fl->tag[sd->tag_idx], sd->map);
1870 	}
1871 
1872 	free(fl->sdesc, M_CXGBE);
1873 	fl->sdesc = NULL;
1874 }
1875 
1876 static int
1877 alloc_tx_maps(struct sge_txq *txq)
1878 {
1879 	struct tx_map *txm;
1880 	int i, rc, count;
1881 
1882 	/*
1883 	 * We can stuff ~10 frames in an 8-descriptor txpkts WR (8 is the SGE
1884 	 * limit for any WR).  txq->no_dmamap events shouldn't occur if maps is
1885 	 * sized for the worst case.
1886 	 */
1887 	count = txq->eq.qsize * 10 / 8;
1888 	txq->map_total = txq->map_avail = count;
1889 	txq->map_cidx = txq->map_pidx = 0;
1890 
1891 	txq->maps = malloc(count * sizeof(struct tx_map), M_CXGBE,
1892 	    M_ZERO | M_WAITOK);
1893 
1894 	txm = txq->maps;
1895 	for (i = 0; i < count; i++, txm++) {
1896 		rc = bus_dmamap_create(txq->tx_tag, 0, &txm->map);
1897 		if (rc != 0)
1898 			goto failed;
1899 	}
1900 
1901 	return (0);
1902 failed:
1903 	while (--i >= 0) {
1904 		txm--;
1905 		bus_dmamap_destroy(txq->tx_tag, txm->map);
1906 	}
1907 	KASSERT(txm == txq->maps, ("%s: EDOOFUS", __func__));
1908 
1909 	free(txq->maps, M_CXGBE);
1910 	txq->maps = NULL;
1911 
1912 	return (rc);
1913 }
1914 
1915 static void
1916 free_tx_maps(struct sge_txq *txq)
1917 {
1918 	struct tx_map *txm;
1919 	int i;
1920 
1921 	txm = txq->maps;
1922 	for (i = 0; i < txq->map_total; i++, txm++) {
1923 
1924 		if (txm->m) {
1925 			bus_dmamap_unload(txq->tx_tag, txm->map);
1926 			m_freem(txm->m);
1927 			txm->m = NULL;
1928 		}
1929 
1930 		bus_dmamap_destroy(txq->tx_tag, txm->map);
1931 	}
1932 
1933 	free(txq->maps, M_CXGBE);
1934 	txq->maps = NULL;
1935 }
1936 
1937 /*
1938  * We'll do immediate data tx for non-TSO, but only when not coalescing.  We're
1939  * willing to use upto 2 hardware descriptors which means a maximum of 96 bytes
1940  * of immediate data.
1941  */
1942 #define IMM_LEN ( \
1943       2 * TX_EQ_ESIZE \
1944     - sizeof(struct fw_eth_tx_pkt_wr) \
1945     - sizeof(struct cpl_tx_pkt_core))
1946 
1947 /*
1948  * Returns non-zero on failure, no need to cleanup anything in that case.
1949  *
1950  * Note 1: We always try to defrag the mbuf if required and return EFBIG only
1951  * if the resulting chain still won't fit in a tx descriptor.
1952  *
1953  * Note 2: We'll pullup the mbuf chain if TSO is requested and the first mbuf
1954  * does not have the TCP header in it.
1955  */
1956 static int
1957 get_pkt_sgl(struct sge_txq *txq, struct mbuf **fp, struct sgl *sgl,
1958     int sgl_only)
1959 {
1960 	struct mbuf *m = *fp;
1961 	struct tx_map *txm;
1962 	int rc, defragged = 0, n;
1963 
1964 	TXQ_LOCK_ASSERT_OWNED(txq);
1965 
1966 	if (m->m_pkthdr.tso_segsz)
1967 		sgl_only = 1;	/* Do not allow immediate data with LSO */
1968 
1969 start:	sgl->nsegs = 0;
1970 
1971 	if (m->m_pkthdr.len <= IMM_LEN && !sgl_only)
1972 		return (0);	/* nsegs = 0 tells caller to use imm. tx */
1973 
1974 	if (txq->map_avail == 0) {
1975 		txq->no_dmamap++;
1976 		return (ENOMEM);
1977 	}
1978 	txm = &txq->maps[txq->map_pidx];
1979 
1980 	if (m->m_pkthdr.tso_segsz && m->m_len < 50) {
1981 		*fp = m_pullup(m, 50);
1982 		m = *fp;
1983 		if (m == NULL)
1984 			return (ENOBUFS);
1985 	}
1986 
1987 	rc = bus_dmamap_load_mbuf_sg(txq->tx_tag, txm->map, m, sgl->seg,
1988 	    &sgl->nsegs, BUS_DMA_NOWAIT);
1989 	if (rc == EFBIG && defragged == 0) {
1990 		m = m_defrag(m, M_DONTWAIT);
1991 		if (m == NULL)
1992 			return (EFBIG);
1993 
1994 		defragged = 1;
1995 		*fp = m;
1996 		goto start;
1997 	}
1998 	if (rc != 0)
1999 		return (rc);
2000 
2001 	txm->m = m;
2002 	txq->map_avail--;
2003 	if (++txq->map_pidx == txq->map_total)
2004 		txq->map_pidx = 0;
2005 
2006 	KASSERT(sgl->nsegs > 0 && sgl->nsegs <= TX_SGL_SEGS,
2007 	    ("%s: bad DMA mapping (%d segments)", __func__, sgl->nsegs));
2008 
2009 	/*
2010 	 * Store the # of flits required to hold this frame's SGL in nflits.  An
2011 	 * SGL has a (ULPTX header + len0, addr0) tuple optionally followed by
2012 	 * multiple (len0 + len1, addr0, addr1) tuples.  If addr1 is not used
2013 	 * then len1 must be set to 0.
2014 	 */
2015 	n = sgl->nsegs - 1;
2016 	sgl->nflits = (3 * n) / 2 + (n & 1) + 2;
2017 
2018 	return (0);
2019 }
2020 
2021 
2022 /*
2023  * Releases all the txq resources used up in the specified sgl.
2024  */
2025 static int
2026 free_pkt_sgl(struct sge_txq *txq, struct sgl *sgl)
2027 {
2028 	struct tx_map *txm;
2029 
2030 	TXQ_LOCK_ASSERT_OWNED(txq);
2031 
2032 	if (sgl->nsegs == 0)
2033 		return (0);	/* didn't use any map */
2034 
2035 	/* 1 pkt uses exactly 1 map, back it out */
2036 
2037 	txq->map_avail++;
2038 	if (txq->map_pidx > 0)
2039 		txq->map_pidx--;
2040 	else
2041 		txq->map_pidx = txq->map_total - 1;
2042 
2043 	txm = &txq->maps[txq->map_pidx];
2044 	bus_dmamap_unload(txq->tx_tag, txm->map);
2045 	txm->m = NULL;
2046 
2047 	return (0);
2048 }
2049 
2050 static int
2051 write_txpkt_wr(struct port_info *pi, struct sge_txq *txq, struct mbuf *m,
2052     struct sgl *sgl)
2053 {
2054 	struct sge_eq *eq = &txq->eq;
2055 	struct fw_eth_tx_pkt_wr *wr;
2056 	struct cpl_tx_pkt_core *cpl;
2057 	uint32_t ctrl;	/* used in many unrelated places */
2058 	uint64_t ctrl1;
2059 	int nflits, ndesc, pktlen;
2060 	struct tx_sdesc *txsd;
2061 	caddr_t dst;
2062 
2063 	TXQ_LOCK_ASSERT_OWNED(txq);
2064 
2065 	pktlen = m->m_pkthdr.len;
2066 
2067 	/*
2068 	 * Do we have enough flits to send this frame out?
2069 	 */
2070 	ctrl = sizeof(struct cpl_tx_pkt_core);
2071 	if (m->m_pkthdr.tso_segsz) {
2072 		nflits = TXPKT_LSO_WR_HDR;
2073 		ctrl += sizeof(struct cpl_tx_pkt_lso);
2074 	} else
2075 		nflits = TXPKT_WR_HDR;
2076 	if (sgl->nsegs > 0)
2077 		nflits += sgl->nflits;
2078 	else {
2079 		nflits += howmany(pktlen, 8);
2080 		ctrl += pktlen;
2081 	}
2082 	ndesc = howmany(nflits, 8);
2083 	if (ndesc > eq->avail)
2084 		return (ENOMEM);
2085 
2086 	/* Firmware work request header */
2087 	wr = (void *)&eq->desc[eq->pidx];
2088 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
2089 	    V_FW_WR_IMMDLEN(ctrl));
2090 	ctrl = V_FW_WR_LEN16(howmany(nflits, 2));
2091 	if (eq->avail == ndesc && !(eq->flags & EQ_CRFLUSHED)) {
2092 		ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ;
2093 		eq->flags |= EQ_CRFLUSHED;
2094 	}
2095 
2096 	wr->equiq_to_len16 = htobe32(ctrl);
2097 	wr->r3 = 0;
2098 
2099 	if (m->m_pkthdr.tso_segsz) {
2100 		struct cpl_tx_pkt_lso *lso = (void *)(wr + 1);
2101 		struct ether_header *eh;
2102 		struct ip *ip;
2103 		struct tcphdr *tcp;
2104 
2105 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
2106 		    F_LSO_LAST_SLICE;
2107 
2108 		eh = mtod(m, struct ether_header *);
2109 		if (eh->ether_type == htons(ETHERTYPE_VLAN)) {
2110 			ctrl |= V_LSO_ETHHDR_LEN(1);
2111 			ip = (void *)((struct ether_vlan_header *)eh + 1);
2112 		} else
2113 			ip = (void *)(eh + 1);
2114 
2115 		tcp = (void *)((uintptr_t)ip + ip->ip_hl * 4);
2116 		ctrl |= V_LSO_IPHDR_LEN(ip->ip_hl) |
2117 		    V_LSO_TCPHDR_LEN(tcp->th_off);
2118 
2119 		lso->lso_ctrl = htobe32(ctrl);
2120 		lso->ipid_ofst = htobe16(0);
2121 		lso->mss = htobe16(m->m_pkthdr.tso_segsz);
2122 		lso->seqno_offset = htobe32(0);
2123 		lso->len = htobe32(pktlen);
2124 
2125 		cpl = (void *)(lso + 1);
2126 
2127 		txq->tso_wrs++;
2128 	} else
2129 		cpl = (void *)(wr + 1);
2130 
2131 	/* Checksum offload */
2132 	ctrl1 = 0;
2133 	if (!(m->m_pkthdr.csum_flags & CSUM_IP))
2134 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
2135 	if (!(m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP)))
2136 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
2137 	if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP))
2138 		txq->txcsum++;	/* some hardware assistance provided */
2139 
2140 	/* VLAN tag insertion */
2141 	if (m->m_flags & M_VLANTAG) {
2142 		ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
2143 		txq->vlan_insertion++;
2144 	}
2145 
2146 	/* CPL header */
2147 	cpl->ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
2148 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf));
2149 	cpl->pack = 0;
2150 	cpl->len = htobe16(pktlen);
2151 	cpl->ctrl1 = htobe64(ctrl1);
2152 
2153 	/* Software descriptor */
2154 	txsd = &txq->sdesc[eq->pidx];
2155 	txsd->desc_used = ndesc;
2156 
2157 	eq->pending += ndesc;
2158 	eq->avail -= ndesc;
2159 	eq->pidx += ndesc;
2160 	if (eq->pidx >= eq->cap)
2161 		eq->pidx -= eq->cap;
2162 
2163 	/* SGL */
2164 	dst = (void *)(cpl + 1);
2165 	if (sgl->nsegs > 0) {
2166 		txsd->credits = 1;
2167 		txq->sgl_wrs++;
2168 		write_sgl_to_txd(eq, sgl, &dst);
2169 	} else {
2170 		txsd->credits = 0;
2171 		txq->imm_wrs++;
2172 		for (; m; m = m->m_next) {
2173 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
2174 #ifdef INVARIANTS
2175 			pktlen -= m->m_len;
2176 #endif
2177 		}
2178 #ifdef INVARIANTS
2179 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
2180 #endif
2181 
2182 	}
2183 
2184 	txq->txpkt_wrs++;
2185 	return (0);
2186 }
2187 
2188 /*
2189  * Returns 0 to indicate that m has been accepted into a coalesced tx work
2190  * request.  It has either been folded into txpkts or txpkts was flushed and m
2191  * has started a new coalesced work request (as the first frame in a fresh
2192  * txpkts).
2193  *
2194  * Returns non-zero to indicate a failure - caller is responsible for
2195  * transmitting m, if there was anything in txpkts it has been flushed.
2196  */
2197 static int
2198 add_to_txpkts(struct port_info *pi, struct sge_txq *txq, struct txpkts *txpkts,
2199     struct mbuf *m, struct sgl *sgl)
2200 {
2201 	struct sge_eq *eq = &txq->eq;
2202 	int can_coalesce;
2203 	struct tx_sdesc *txsd;
2204 	int flits;
2205 
2206 	TXQ_LOCK_ASSERT_OWNED(txq);
2207 
2208 	if (txpkts->npkt > 0) {
2209 		flits = TXPKTS_PKT_HDR + sgl->nflits;
2210 		can_coalesce = m->m_pkthdr.tso_segsz == 0 &&
2211 		    txpkts->nflits + flits <= TX_WR_FLITS &&
2212 		    txpkts->nflits + flits <= eq->avail * 8 &&
2213 		    txpkts->plen + m->m_pkthdr.len < 65536;
2214 
2215 		if (can_coalesce) {
2216 			txpkts->npkt++;
2217 			txpkts->nflits += flits;
2218 			txpkts->plen += m->m_pkthdr.len;
2219 
2220 			txsd = &txq->sdesc[eq->pidx];
2221 			txsd->credits++;
2222 
2223 			return (0);
2224 		}
2225 
2226 		/*
2227 		 * Couldn't coalesce m into txpkts.  The first order of business
2228 		 * is to send txpkts on its way.  Then we'll revisit m.
2229 		 */
2230 		write_txpkts_wr(txq, txpkts);
2231 	}
2232 
2233 	/*
2234 	 * Check if we can start a new coalesced tx work request with m as
2235 	 * the first packet in it.
2236 	 */
2237 
2238 	KASSERT(txpkts->npkt == 0, ("%s: txpkts not empty", __func__));
2239 
2240 	flits = TXPKTS_WR_HDR + sgl->nflits;
2241 	can_coalesce = m->m_pkthdr.tso_segsz == 0 &&
2242 	    flits <= eq->avail * 8 && flits <= TX_WR_FLITS;
2243 
2244 	if (can_coalesce == 0)
2245 		return (EINVAL);
2246 
2247 	/*
2248 	 * Start a fresh coalesced tx WR with m as the first frame in it.
2249 	 */
2250 	txpkts->npkt = 1;
2251 	txpkts->nflits = flits;
2252 	txpkts->flitp = &eq->desc[eq->pidx].flit[2];
2253 	txpkts->plen = m->m_pkthdr.len;
2254 
2255 	txsd = &txq->sdesc[eq->pidx];
2256 	txsd->credits = 1;
2257 
2258 	return (0);
2259 }
2260 
2261 /*
2262  * Note that write_txpkts_wr can never run out of hardware descriptors (but
2263  * write_txpkt_wr can).  add_to_txpkts ensures that a frame is accepted for
2264  * coalescing only if sufficient hardware descriptors are available.
2265  */
2266 static void
2267 write_txpkts_wr(struct sge_txq *txq, struct txpkts *txpkts)
2268 {
2269 	struct sge_eq *eq = &txq->eq;
2270 	struct fw_eth_tx_pkts_wr *wr;
2271 	struct tx_sdesc *txsd;
2272 	uint32_t ctrl;
2273 	int ndesc;
2274 
2275 	TXQ_LOCK_ASSERT_OWNED(txq);
2276 
2277 	ndesc = howmany(txpkts->nflits, 8);
2278 
2279 	wr = (void *)&eq->desc[eq->pidx];
2280 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR) |
2281 	    V_FW_WR_IMMDLEN(0)); /* immdlen does not matter in this WR */
2282 	ctrl = V_FW_WR_LEN16(howmany(txpkts->nflits, 2));
2283 	if (eq->avail == ndesc && !(eq->flags & EQ_CRFLUSHED)) {
2284 		ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ;
2285 		eq->flags |= EQ_CRFLUSHED;
2286 	}
2287 	wr->equiq_to_len16 = htobe32(ctrl);
2288 	wr->plen = htobe16(txpkts->plen);
2289 	wr->npkt = txpkts->npkt;
2290 	wr->r3 = wr->r4 = 0;
2291 
2292 	/* Everything else already written */
2293 
2294 	txsd = &txq->sdesc[eq->pidx];
2295 	txsd->desc_used = ndesc;
2296 
2297 	KASSERT(eq->avail >= ndesc, ("%s: out of descriptors", __func__));
2298 
2299 	eq->pending += ndesc;
2300 	eq->avail -= ndesc;
2301 	eq->pidx += ndesc;
2302 	if (eq->pidx >= eq->cap)
2303 		eq->pidx -= eq->cap;
2304 
2305 	txq->txpkts_pkts += txpkts->npkt;
2306 	txq->txpkts_wrs++;
2307 	txpkts->npkt = 0;	/* emptied */
2308 }
2309 
2310 static inline void
2311 write_ulp_cpl_sgl(struct port_info *pi, struct sge_txq *txq,
2312     struct txpkts *txpkts, struct mbuf *m, struct sgl *sgl)
2313 {
2314 	struct ulp_txpkt *ulpmc;
2315 	struct ulptx_idata *ulpsc;
2316 	struct cpl_tx_pkt_core *cpl;
2317 	struct sge_eq *eq = &txq->eq;
2318 	uintptr_t flitp, start, end;
2319 	uint64_t ctrl;
2320 	caddr_t dst;
2321 
2322 	KASSERT(txpkts->npkt > 0, ("%s: txpkts is empty", __func__));
2323 
2324 	start = (uintptr_t)eq->desc;
2325 	end = (uintptr_t)eq->spg;
2326 
2327 	/* Checksum offload */
2328 	ctrl = 0;
2329 	if (!(m->m_pkthdr.csum_flags & CSUM_IP))
2330 		ctrl |= F_TXPKT_IPCSUM_DIS;
2331 	if (!(m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP)))
2332 		ctrl |= F_TXPKT_L4CSUM_DIS;
2333 	if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP))
2334 		txq->txcsum++;	/* some hardware assistance provided */
2335 
2336 	/* VLAN tag insertion */
2337 	if (m->m_flags & M_VLANTAG) {
2338 		ctrl |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
2339 		txq->vlan_insertion++;
2340 	}
2341 
2342 	/*
2343 	 * The previous packet's SGL must have ended at a 16 byte boundary (this
2344 	 * is required by the firmware/hardware).  It follows that flitp cannot
2345 	 * wrap around between the ULPTX master command and ULPTX subcommand (8
2346 	 * bytes each), and that it can not wrap around in the middle of the
2347 	 * cpl_tx_pkt_core either.
2348 	 */
2349 	flitp = (uintptr_t)txpkts->flitp;
2350 	KASSERT((flitp & 0xf) == 0,
2351 	    ("%s: last SGL did not end at 16 byte boundary: %p",
2352 	    __func__, txpkts->flitp));
2353 
2354 	/* ULP master command */
2355 	ulpmc = (void *)flitp;
2356 	ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0) |
2357 	    V_ULP_TXPKT_FID(eq->iqid));
2358 	ulpmc->len = htonl(howmany(sizeof(*ulpmc) + sizeof(*ulpsc) +
2359 	    sizeof(*cpl) + 8 * sgl->nflits, 16));
2360 
2361 	/* ULP subcommand */
2362 	ulpsc = (void *)(ulpmc + 1);
2363 	ulpsc->cmd_more = htobe32(V_ULPTX_CMD((u32)ULP_TX_SC_IMM) |
2364 	    F_ULP_TX_SC_MORE);
2365 	ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
2366 
2367 	flitp += sizeof(*ulpmc) + sizeof(*ulpsc);
2368 	if (flitp == end)
2369 		flitp = start;
2370 
2371 	/* CPL_TX_PKT */
2372 	cpl = (void *)flitp;
2373 	cpl->ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
2374 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf));
2375 	cpl->pack = 0;
2376 	cpl->len = htobe16(m->m_pkthdr.len);
2377 	cpl->ctrl1 = htobe64(ctrl);
2378 
2379 	flitp += sizeof(*cpl);
2380 	if (flitp == end)
2381 		flitp = start;
2382 
2383 	/* SGL for this frame */
2384 	dst = (caddr_t)flitp;
2385 	txpkts->nflits += write_sgl_to_txd(eq, sgl, &dst);
2386 	txpkts->flitp = (void *)dst;
2387 
2388 	KASSERT(((uintptr_t)dst & 0xf) == 0,
2389 	    ("%s: SGL ends at %p (not a 16 byte boundary)", __func__, dst));
2390 }
2391 
2392 /*
2393  * If the SGL ends on an address that is not 16 byte aligned, this function will
2394  * add a 0 filled flit at the end.  It returns 1 in that case.
2395  */
2396 static int
2397 write_sgl_to_txd(struct sge_eq *eq, struct sgl *sgl, caddr_t *to)
2398 {
2399 	__be64 *flitp, *end;
2400 	struct ulptx_sgl *usgl;
2401 	bus_dma_segment_t *seg;
2402 	int i, padded;
2403 
2404 	KASSERT(sgl->nsegs > 0 && sgl->nflits > 0,
2405 	    ("%s: bad SGL - nsegs=%d, nflits=%d",
2406 	    __func__, sgl->nsegs, sgl->nflits));
2407 
2408 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
2409 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
2410 
2411 	flitp = (__be64 *)(*to);
2412 	end = flitp + sgl->nflits;
2413 	seg = &sgl->seg[0];
2414 	usgl = (void *)flitp;
2415 
2416 	/*
2417 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
2418 	 * ring, so we're at least 16 bytes away from the status page.  There is
2419 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
2420 	 */
2421 
2422 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
2423 	    V_ULPTX_NSGE(sgl->nsegs));
2424 	usgl->len0 = htobe32(seg->ds_len);
2425 	usgl->addr0 = htobe64(seg->ds_addr);
2426 	seg++;
2427 
2428 	if ((uintptr_t)end <= (uintptr_t)eq->spg) {
2429 
2430 		/* Won't wrap around at all */
2431 
2432 		for (i = 0; i < sgl->nsegs - 1; i++, seg++) {
2433 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ds_len);
2434 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ds_addr);
2435 		}
2436 		if (i & 1)
2437 			usgl->sge[i / 2].len[1] = htobe32(0);
2438 	} else {
2439 
2440 		/* Will wrap somewhere in the rest of the SGL */
2441 
2442 		/* 2 flits already written, write the rest flit by flit */
2443 		flitp = (void *)(usgl + 1);
2444 		for (i = 0; i < sgl->nflits - 2; i++) {
2445 			if ((uintptr_t)flitp == (uintptr_t)eq->spg)
2446 				flitp = (void *)eq->desc;
2447 			*flitp++ = get_flit(seg, sgl->nsegs - 1, i);
2448 		}
2449 		end = flitp;
2450 	}
2451 
2452 	if ((uintptr_t)end & 0xf) {
2453 		*(uint64_t *)end = 0;
2454 		end++;
2455 		padded = 1;
2456 	} else
2457 		padded = 0;
2458 
2459 	if ((uintptr_t)end == (uintptr_t)eq->spg)
2460 		*to = (void *)eq->desc;
2461 	else
2462 		*to = (void *)end;
2463 
2464 	return (padded);
2465 }
2466 
2467 static inline void
2468 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
2469 {
2470 	if ((uintptr_t)(*to) + len <= (uintptr_t)eq->spg) {
2471 		bcopy(from, *to, len);
2472 		(*to) += len;
2473 	} else {
2474 		int portion = (uintptr_t)eq->spg - (uintptr_t)(*to);
2475 
2476 		bcopy(from, *to, portion);
2477 		from += portion;
2478 		portion = len - portion;	/* remaining */
2479 		bcopy(from, (void *)eq->desc, portion);
2480 		(*to) = (caddr_t)eq->desc + portion;
2481 	}
2482 }
2483 
2484 static inline void
2485 ring_eq_db(struct adapter *sc, struct sge_eq *eq)
2486 {
2487 	wmb();
2488 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
2489 	    V_QID(eq->cntxt_id) | V_PIDX(eq->pending));
2490 	eq->pending = 0;
2491 }
2492 
2493 static inline int
2494 reclaimable(struct sge_eq *eq)
2495 {
2496 	unsigned int cidx;
2497 
2498 	cidx = eq->spg->cidx;	/* stable snapshot */
2499 	cidx = be16_to_cpu(cidx);
2500 
2501 	if (cidx >= eq->cidx)
2502 		return (cidx - eq->cidx);
2503 	else
2504 		return (cidx + eq->cap - eq->cidx);
2505 }
2506 
2507 /*
2508  * There are "can_reclaim" tx descriptors ready to be reclaimed.  Reclaim as
2509  * many as possible but stop when there are around "n" mbufs to free.
2510  *
2511  * The actual number reclaimed is provided as the return value.
2512  */
2513 static int
2514 reclaim_tx_descs(struct sge_txq *txq, int can_reclaim, int n)
2515 {
2516 	struct tx_sdesc *txsd;
2517 	struct tx_map *txm;
2518 	unsigned int reclaimed, maps;
2519 	struct sge_eq *eq = &txq->eq;
2520 
2521 	EQ_LOCK_ASSERT_OWNED(eq);
2522 
2523 	if (can_reclaim == 0)
2524 		can_reclaim = reclaimable(eq);
2525 
2526 	maps = reclaimed = 0;
2527 	while (can_reclaim && maps < n) {
2528 		int ndesc;
2529 
2530 		txsd = &txq->sdesc[eq->cidx];
2531 		ndesc = txsd->desc_used;
2532 
2533 		/* Firmware doesn't return "partial" credits. */
2534 		KASSERT(can_reclaim >= ndesc,
2535 		    ("%s: unexpected number of credits: %d, %d",
2536 		    __func__, can_reclaim, ndesc));
2537 
2538 		maps += txsd->credits;
2539 
2540 		reclaimed += ndesc;
2541 		can_reclaim -= ndesc;
2542 
2543 		eq->cidx += ndesc;
2544 		if (__predict_false(eq->cidx >= eq->cap))
2545 			eq->cidx -= eq->cap;
2546 	}
2547 
2548 	txm = &txq->maps[txq->map_cidx];
2549 	if (maps)
2550 		prefetch(txm->m);
2551 
2552 	eq->avail += reclaimed;
2553 	KASSERT(eq->avail < eq->cap,	/* avail tops out at (cap - 1) */
2554 	    ("%s: too many descriptors available", __func__));
2555 
2556 	txq->map_avail += maps;
2557 	KASSERT(txq->map_avail <= txq->map_total,
2558 	    ("%s: too many maps available", __func__));
2559 
2560 	while (maps--) {
2561 		struct tx_map *next;
2562 
2563 		next = txm + 1;
2564 		if (__predict_false(txq->map_cidx + 1 == txq->map_total))
2565 			next = txq->maps;
2566 		prefetch(next->m);
2567 
2568 		bus_dmamap_unload(txq->tx_tag, txm->map);
2569 		m_freem(txm->m);
2570 		txm->m = NULL;
2571 
2572 		txm = next;
2573 		if (__predict_false(++txq->map_cidx == txq->map_total))
2574 			txq->map_cidx = 0;
2575 	}
2576 
2577 	return (reclaimed);
2578 }
2579 
2580 static void
2581 write_eqflush_wr(struct sge_eq *eq)
2582 {
2583 	struct fw_eq_flush_wr *wr;
2584 
2585 	EQ_LOCK_ASSERT_OWNED(eq);
2586 	KASSERT(eq->avail > 0, ("%s: no descriptors left.", __func__));
2587 
2588 	wr = (void *)&eq->desc[eq->pidx];
2589 	bzero(wr, sizeof(*wr));
2590 	wr->opcode = FW_EQ_FLUSH_WR;
2591 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(sizeof(*wr) / 16) |
2592 	    F_FW_WR_EQUEQ | F_FW_WR_EQUIQ);
2593 
2594 	eq->flags |= EQ_CRFLUSHED;
2595 	eq->pending++;
2596 	eq->avail--;
2597 	if (++eq->pidx == eq->cap)
2598 		eq->pidx = 0;
2599 }
2600 
2601 static __be64
2602 get_flit(bus_dma_segment_t *sgl, int nsegs, int idx)
2603 {
2604 	int i = (idx / 3) * 2;
2605 
2606 	switch (idx % 3) {
2607 	case 0: {
2608 		__be64 rc;
2609 
2610 		rc = htobe32(sgl[i].ds_len);
2611 		if (i + 1 < nsegs)
2612 			rc |= (uint64_t)htobe32(sgl[i + 1].ds_len) << 32;
2613 
2614 		return (rc);
2615 	}
2616 	case 1:
2617 		return htobe64(sgl[i].ds_addr);
2618 	case 2:
2619 		return htobe64(sgl[i + 1].ds_addr);
2620 	}
2621 
2622 	return (0);
2623 }
2624 
2625 static void
2626 set_fl_tag_idx(struct sge_fl *fl, int mtu)
2627 {
2628 	int i;
2629 
2630 	FL_LOCK_ASSERT_OWNED(fl);
2631 
2632 	for (i = 0; i < FL_BUF_SIZES - 1; i++) {
2633 		if (FL_BUF_SIZE(i) >= (mtu + FL_PKTSHIFT))
2634 			break;
2635 	}
2636 
2637 	fl->tag_idx = i;
2638 }
2639 
2640 static int
2641 handle_sge_egr_update(struct adapter *sc, const struct cpl_sge_egr_update *cpl)
2642 {
2643 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
2644 	struct sge *s = &sc->sge;
2645 	struct sge_txq *txq;
2646 	struct port_info *pi;
2647 
2648 	txq = (void *)s->eqmap[qid - s->eq_start];
2649 	TXQ_LOCK(txq);
2650 	if (txq->eq.flags & EQ_CRFLUSHED) {
2651 		pi = txq->ifp->if_softc;
2652 		taskqueue_enqueue(pi->tq, &txq->resume_tx);
2653 		txq->egr_update++;
2654 	} else
2655 		wakeup_one(txq);	/* txq is going away, wakeup free_txq */
2656 	TXQ_UNLOCK(txq);
2657 
2658 	return (0);
2659 }
2660 
2661 /*
2662  * m0 is freed on successful transmission.
2663  */
2664 static int
2665 ctrl_tx(struct adapter *sc, struct sge_ctrlq *ctrlq, struct mbuf *m0)
2666 {
2667 	struct sge_eq *eq = &ctrlq->eq;
2668 	int rc = 0, ndesc;
2669 	int can_reclaim;
2670 	caddr_t dst;
2671 	struct mbuf *m;
2672 
2673 	M_ASSERTPKTHDR(m0);
2674 
2675 	if (m0->m_pkthdr.len > SGE_MAX_WR_LEN) {
2676 		ctrlq->too_long++;
2677 		return (EMSGSIZE);
2678 	}
2679 	ndesc = howmany(m0->m_pkthdr.len, CTRL_EQ_ESIZE);
2680 
2681 	EQ_LOCK(eq);
2682 
2683 	can_reclaim = reclaimable(eq);
2684 	eq->cidx += can_reclaim;
2685 	eq->avail += can_reclaim;
2686 	if (__predict_false(eq->cidx >= eq->cap))
2687 		eq->cidx -= eq->cap;
2688 
2689 	if (eq->avail < ndesc) {
2690 		rc = EAGAIN;
2691 		ctrlq->no_desc++;
2692 		goto failed;
2693 	}
2694 
2695 	dst = (void *)&eq->desc[eq->pidx];
2696 	for (m = m0; m; m = m->m_next)
2697 		copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
2698 
2699 	eq->pidx += ndesc;
2700 	if (__predict_false(eq->pidx >= eq->cap))
2701 		eq->pidx -= eq->cap;
2702 
2703 	eq->pending += ndesc;
2704 	ctrlq->total_wrs++;
2705 	ring_eq_db(sc, eq);
2706 failed:
2707 	EQ_UNLOCK(eq);
2708 	if (rc == 0)
2709 		m_freem(m0);
2710 
2711 	return (rc);
2712 }
2713