xref: /illumos-gate/usr/src/uts/common/io/cxgbe/t4nex/t4_sge.c (revision ae389aa988ce154c9f8d5d4dbe4a63c3744339f4)
1 /*
2  * This file and its contents are supplied under the terms of the
3  * Common Development and Distribution License ("CDDL"), version 1.0.
4  * You may only use this file in accordance with the terms of version
5  * 1.0 of the CDDL.
6  *
7  * A full copy of the text of the CDDL should have accompanied this
8  * source. A copy of the CDDL is also available via the Internet at
9  * http://www.illumos.org/license/CDDL.
10  */
11 
12 /*
13  * This file is part of the Chelsio T4 support code.
14  *
15  * Copyright (C) 2010-2013 Chelsio Communications.  All rights reserved.
16  *
17  * This program is distributed in the hope that it will be useful, but WITHOUT
18  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
19  * FITNESS FOR A PARTICULAR PURPOSE.  See the LICENSE file included in this
20  * release for licensing terms and conditions.
21  */
22 
23 #include <sys/ddi.h>
24 #include <sys/sunddi.h>
25 #include <sys/sunndi.h>
26 #include <sys/atomic.h>
27 #include <sys/dlpi.h>
28 #include <sys/pattr.h>
29 #include <sys/strsubr.h>
30 #include <sys/stream.h>
31 #include <sys/strsun.h>
32 #include <inet/ip.h>
33 #include <inet/tcp.h>
34 
35 #include "version.h"
36 #include "common/common.h"
37 #include "common/t4_msg.h"
38 #include "common/t4_regs.h"
39 #include "common/t4_regs_values.h"
40 
41 /* TODO: Tune. */
42 int rx_buf_size = 8192;
43 int tx_copy_threshold = 256;
44 uint16_t rx_copy_threshold = 256;
45 
46 /* Used to track coalesced tx work request */
47 struct txpkts {
48 	mblk_t *tail;		/* head is in the software descriptor */
49 	uint64_t *flitp;	/* ptr to flit where next pkt should start */
50 	uint8_t npkt;		/* # of packets in this work request */
51 	uint8_t nflits;		/* # of flits used by this work request */
52 	uint16_t plen;		/* total payload (sum of all packets) */
53 };
54 
55 /* All information needed to tx a frame */
56 struct txinfo {
57 	uint32_t len;		/* Total length of frame */
58 	uint32_t flags;		/* Checksum and LSO flags */
59 	uint32_t mss;		/* MSS for LSO */
60 	uint8_t nsegs;		/* # of segments in the SGL, 0 means imm. tx */
61 	uint8_t nflits;		/* # of flits needed for the SGL */
62 	uint8_t hdls_used;	/* # of DMA handles used */
63 	uint32_t txb_used;	/* txb_space used */
64 	struct ulptx_sgl sgl __attribute__((aligned(8)));
65 	struct ulptx_sge_pair reserved[TX_SGL_SEGS / 2];
66 };
67 
68 static int service_iq(struct sge_iq *iq, int budget);
69 static inline void init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx,
70     int8_t pktc_idx, int qsize, uint8_t esize);
71 static inline void init_fl(struct sge_fl *fl, uint16_t qsize);
72 static inline void init_eq(struct adapter *sc, struct sge_eq *eq,
73     uint16_t eqtype, uint16_t qsize,uint8_t tx_chan, uint16_t iqid);
74 static int alloc_iq_fl(struct port_info *pi, struct sge_iq *iq,
75     struct sge_fl *fl, int intr_idx, int cong);
76 static int free_iq_fl(struct port_info *pi, struct sge_iq *iq,
77     struct sge_fl *fl);
78 static int alloc_fwq(struct adapter *sc);
79 static int free_fwq(struct adapter *sc);
80 #ifdef TCP_OFFLOAD_ENABLE
81 static int alloc_mgmtq(struct adapter *sc);
82 #endif
83 static int alloc_rxq(struct port_info *pi, struct sge_rxq *rxq, int intr_idx,
84     int i);
85 static int free_rxq(struct port_info *pi, struct sge_rxq *rxq);
86 #ifdef TCP_OFFLOAD_ENABLE
87 static int alloc_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq,
88 	int intr_idx);
89 static int free_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq);
90 #endif
91 static int ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq);
92 static int eth_eq_alloc(struct adapter *sc, struct port_info *pi,
93     struct sge_eq *eq);
94 #ifdef TCP_OFFLOAD_ENABLE
95 static int ofld_eq_alloc(struct adapter *sc, struct port_info *pi,
96     struct sge_eq *eq);
97 #endif
98 static int alloc_eq(struct adapter *sc, struct port_info *pi,
99     struct sge_eq *eq);
100 static int free_eq(struct adapter *sc, struct sge_eq *eq);
101 #ifdef TCP_OFFLOAD_ENABLE
102 static int alloc_wrq(struct adapter *sc, struct port_info *pi,
103     struct sge_wrq *wrq, int idx);
104 static int free_wrq(struct adapter *sc, struct sge_wrq *wrq);
105 #endif
106 static int alloc_txq(struct port_info *pi, struct sge_txq *txq, int idx);
107 static int free_txq(struct port_info *pi, struct sge_txq *txq);
108 static int alloc_dma_memory(struct adapter *sc, size_t len, int flags,
109     ddi_device_acc_attr_t *acc_attr, ddi_dma_attr_t *dma_attr,
110     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl, uint64_t *pba,
111     caddr_t *pva);
112 static int free_dma_memory(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl);
113 static int alloc_desc_ring(struct adapter *sc, size_t len, int rw,
114     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl, uint64_t *pba,
115     caddr_t *pva);
116 static int free_desc_ring(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl);
117 static int alloc_tx_copybuffer(struct adapter *sc, size_t len,
118     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl, uint64_t *pba,
119     caddr_t *pva);
120 static inline bool is_new_response(const struct sge_iq *iq,
121     struct rsp_ctrl **ctrl);
122 static inline void iq_next(struct sge_iq *iq);
123 static int refill_fl(struct adapter *sc, struct sge_fl *fl, int nbufs);
124 static void refill_sfl(void *arg);
125 static void add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl);
126 static void free_fl_bufs(struct sge_fl *fl);
127 static mblk_t *get_fl_payload(struct adapter *sc, struct sge_fl *fl,
128     uint32_t len_newbuf, int *fl_bufs_used);
129 static int get_frame_txinfo(struct sge_txq *txq, mblk_t **fp,
130     struct txinfo *txinfo, int sgl_only);
131 static inline int fits_in_txb(struct sge_txq *txq, int len, int *waste);
132 static inline int copy_into_txb(struct sge_txq *txq, mblk_t *m, int len,
133     struct txinfo *txinfo);
134 static inline void add_seg(struct txinfo *txinfo, uint64_t ba, uint32_t len);
135 static inline int add_mblk(struct sge_txq *txq, struct txinfo *txinfo,
136     mblk_t *m, int len);
137 static void free_txinfo_resources(struct sge_txq *txq, struct txinfo *txinfo);
138 static int add_to_txpkts(struct sge_txq *txq, struct txpkts *txpkts, mblk_t *m,
139     struct txinfo *txinfo);
140 static void write_txpkts_wr(struct sge_txq *txq, struct txpkts *txpkts);
141 static int write_txpkt_wr(struct port_info *pi, struct sge_txq *txq, mblk_t *m,
142     struct txinfo *txinfo);
143 static inline void write_ulp_cpl_sgl(struct port_info *pi, struct sge_txq *txq,
144     struct txpkts *txpkts, struct txinfo *txinfo);
145 static inline void copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to,
146     int len);
147 static inline void ring_tx_db(struct adapter *sc, struct sge_eq *eq);
148 static int reclaim_tx_descs(struct sge_txq *txq, int howmany);
149 static void write_txqflush_wr(struct sge_txq *txq);
150 static int t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss,
151     mblk_t *m);
152 static inline void ring_fl_db(struct adapter *sc, struct sge_fl *fl);
153 static kstat_t *setup_port_config_kstats(struct port_info *pi);
154 static kstat_t *setup_port_info_kstats(struct port_info *pi);
155 static kstat_t *setup_rxq_kstats(struct port_info *pi, struct sge_rxq *rxq,
156     int idx);
157 static int update_rxq_kstats(kstat_t *ksp, int rw);
158 static int update_port_info_kstats(kstat_t *ksp, int rw);
159 static kstat_t *setup_txq_kstats(struct port_info *pi, struct sge_txq *txq,
160     int idx);
161 static int update_txq_kstats(kstat_t *ksp, int rw);
162 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
163     mblk_t *);
164 static int handle_fw_rpl(struct sge_iq *iq, const struct rss_header *rss,
165     mblk_t *m);
166 
167 static inline int
168 reclaimable(struct sge_eq *eq)
169 {
170 	unsigned int cidx;
171 
172 	cidx = eq->spg->cidx;   /* stable snapshot */
173 	cidx = be16_to_cpu(cidx);
174 
175 	if (cidx >= eq->cidx)
176 		return (cidx - eq->cidx);
177 	else
178 		return (cidx + eq->cap - eq->cidx);
179 }
180 
181 void
182 t4_sge_init(struct adapter *sc)
183 {
184 	struct driver_properties *p = &sc->props;
185 	ddi_dma_attr_t *dma_attr;
186 	ddi_device_acc_attr_t *acc_attr;
187 	uint32_t sge_control, sge_conm_ctrl;
188 	int egress_threshold;
189 
190 	/*
191 	 * Device access and DMA attributes for descriptor rings
192 	 */
193 	acc_attr = &sc->sge.acc_attr_desc;
194 	acc_attr->devacc_attr_version = DDI_DEVICE_ATTR_V0;
195 	acc_attr->devacc_attr_endian_flags = DDI_NEVERSWAP_ACC;
196 	acc_attr->devacc_attr_dataorder = DDI_STRICTORDER_ACC;
197 
198 	dma_attr = &sc->sge.dma_attr_desc;
199 	dma_attr->dma_attr_version = DMA_ATTR_V0;
200 	dma_attr->dma_attr_addr_lo = 0;
201 	dma_attr->dma_attr_addr_hi = UINT64_MAX;
202 	dma_attr->dma_attr_count_max = UINT64_MAX;
203 	dma_attr->dma_attr_align = 512;
204 	dma_attr->dma_attr_burstsizes = 0xfff;
205 	dma_attr->dma_attr_minxfer = 1;
206 	dma_attr->dma_attr_maxxfer = UINT64_MAX;
207 	dma_attr->dma_attr_seg = UINT64_MAX;
208 	dma_attr->dma_attr_sgllen = 1;
209 	dma_attr->dma_attr_granular = 1;
210 	dma_attr->dma_attr_flags = 0;
211 
212 	/*
213 	 * Device access and DMA attributes for tx buffers
214 	 */
215 	acc_attr = &sc->sge.acc_attr_tx;
216 	acc_attr->devacc_attr_version = DDI_DEVICE_ATTR_V0;
217 	acc_attr->devacc_attr_endian_flags = DDI_NEVERSWAP_ACC;
218 
219 	dma_attr = &sc->sge.dma_attr_tx;
220 	dma_attr->dma_attr_version = DMA_ATTR_V0;
221 	dma_attr->dma_attr_addr_lo = 0;
222 	dma_attr->dma_attr_addr_hi = UINT64_MAX;
223 	dma_attr->dma_attr_count_max = UINT64_MAX;
224 	dma_attr->dma_attr_align = 1;
225 	dma_attr->dma_attr_burstsizes = 0xfff;
226 	dma_attr->dma_attr_minxfer = 1;
227 	dma_attr->dma_attr_maxxfer = UINT64_MAX;
228 	dma_attr->dma_attr_seg = UINT64_MAX;
229 	dma_attr->dma_attr_sgllen = TX_SGL_SEGS;
230 	dma_attr->dma_attr_granular = 1;
231 	dma_attr->dma_attr_flags = 0;
232 
233 	/*
234 	 * Ingress Padding Boundary and Egress Status Page Size are set up by
235 	 * t4_fixup_host_params().
236 	 */
237 	sge_control = t4_read_reg(sc, A_SGE_CONTROL);
238 	sc->sge.pktshift = G_PKTSHIFT(sge_control);
239 	sc->sge.stat_len = (sge_control & F_EGRSTATUSPAGESIZE) ? 128 : 64;
240 
241 	/* t4_nex uses FLM packed mode */
242 	sc->sge.fl_align = t4_fl_pkt_align(sc, true);
243 
244 	/*
245 	 * Device access and DMA attributes for rx buffers
246 	 */
247 	sc->sge.rxb_params.dip = sc->dip;
248 	sc->sge.rxb_params.buf_size = rx_buf_size;
249 
250 	acc_attr = &sc->sge.rxb_params.acc_attr_rx;
251 	acc_attr->devacc_attr_version = DDI_DEVICE_ATTR_V0;
252 	acc_attr->devacc_attr_endian_flags = DDI_NEVERSWAP_ACC;
253 
254 	dma_attr = &sc->sge.rxb_params.dma_attr_rx;
255 	dma_attr->dma_attr_version = DMA_ATTR_V0;
256 	dma_attr->dma_attr_addr_lo = 0;
257 	dma_attr->dma_attr_addr_hi = UINT64_MAX;
258 	dma_attr->dma_attr_count_max = UINT64_MAX;
259 	/*
260 	 * Low 4 bits of an rx buffer address have a special meaning to the SGE
261 	 * and an rx buf cannot have an address with any of these bits set.
262 	 * FL_ALIGN is >= 32 so we're sure things are ok.
263 	 */
264 	dma_attr->dma_attr_align = sc->sge.fl_align;
265 	dma_attr->dma_attr_burstsizes = 0xfff;
266 	dma_attr->dma_attr_minxfer = 1;
267 	dma_attr->dma_attr_maxxfer = UINT64_MAX;
268 	dma_attr->dma_attr_seg = UINT64_MAX;
269 	dma_attr->dma_attr_sgllen = 1;
270 	dma_attr->dma_attr_granular = 1;
271 	dma_attr->dma_attr_flags = 0;
272 
273 	sc->sge.rxbuf_cache = rxbuf_cache_create(&sc->sge.rxb_params);
274 
275 	/*
276 	 * A FL with <= fl_starve_thres buffers is starving and a periodic
277 	 * timer will attempt to refill it.  This needs to be larger than the
278 	 * SGE's Egress Congestion Threshold.  If it isn't, then we can get
279 	 * stuck waiting for new packets while the SGE is waiting for us to
280 	 * give it more Free List entries.  (Note that the SGE's Egress
281 	 * Congestion Threshold is in units of 2 Free List pointers.) For T4,
282 	 * there was only a single field to control this.  For T5 there's the
283 	 * original field which now only applies to Unpacked Mode Free List
284 	 * buffers and a new field which only applies to Packed Mode Free List
285 	 * buffers.
286 	 */
287 
288 	sge_conm_ctrl = t4_read_reg(sc, A_SGE_CONM_CTRL);
289 	switch (CHELSIO_CHIP_VERSION(sc->params.chip)) {
290 	case CHELSIO_T4:
291 		egress_threshold = G_EGRTHRESHOLD(sge_conm_ctrl);
292 		break;
293 	case CHELSIO_T5:
294 		egress_threshold = G_EGRTHRESHOLDPACKING(sge_conm_ctrl);
295 		break;
296 	case CHELSIO_T6:
297 	default:
298 		egress_threshold = G_T6_EGRTHRESHOLDPACKING(sge_conm_ctrl);
299 	}
300 	sc->sge.fl_starve_threshold = 2*egress_threshold + 1;
301 
302 	t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0, rx_buf_size);
303 
304 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD,
305 	    V_THRESHOLD_0(p->counter_val[0]) |
306 	    V_THRESHOLD_1(p->counter_val[1]) |
307 	    V_THRESHOLD_2(p->counter_val[2]) |
308 	    V_THRESHOLD_3(p->counter_val[3]));
309 
310 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1,
311 	    V_TIMERVALUE0(us_to_core_ticks(sc, p->timer_val[0])) |
312 	    V_TIMERVALUE1(us_to_core_ticks(sc, p->timer_val[1])));
313 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3,
314 	    V_TIMERVALUE2(us_to_core_ticks(sc, p->timer_val[2])) |
315 	    V_TIMERVALUE3(us_to_core_ticks(sc, p->timer_val[3])));
316 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5,
317 	    V_TIMERVALUE4(us_to_core_ticks(sc, p->timer_val[4])) |
318 	    V_TIMERVALUE5(us_to_core_ticks(sc, p->timer_val[5])));
319 
320 	(void) t4_register_cpl_handler(sc, CPL_FW4_MSG, handle_fw_rpl);
321 	(void) t4_register_cpl_handler(sc, CPL_FW6_MSG, handle_fw_rpl);
322 	(void) t4_register_cpl_handler(sc, CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
323 	(void) t4_register_cpl_handler(sc, CPL_RX_PKT, t4_eth_rx);
324 	(void) t4_register_fw_msg_handler(sc, FW6_TYPE_CMD_RPL,
325 		    t4_handle_fw_rpl);
326 }
327 
328 /*
329  * Allocate and initialize the firmware event queue and the forwarded interrupt
330  * queues, if any.  The adapter owns all these queues as they are not associated
331  * with any particular port.
332  *
333  * Returns errno on failure.  Resources allocated up to that point may still be
334  * allocated.  Caller is responsible for cleanup in case this function fails.
335  */
336 int
337 t4_setup_adapter_queues(struct adapter *sc)
338 {
339 	int rc;
340 
341 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
342 
343 	/*
344 	 * Firmware event queue
345 	 */
346 	rc = alloc_fwq(sc);
347 	if (rc != 0)
348 		return (rc);
349 
350 #ifdef TCP_OFFLOAD_ENABLE
351 	/*
352 	 * Management queue.  This is just a control queue that uses the fwq as
353 	 * its associated iq.
354 	 */
355 	rc = alloc_mgmtq(sc);
356 #endif
357 
358 	return (rc);
359 }
360 
361 /*
362  * Idempotent
363  */
364 int
365 t4_teardown_adapter_queues(struct adapter *sc)
366 {
367 
368 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
369 
370 	(void) free_fwq(sc);
371 
372 	return (0);
373 }
374 
375 static inline int
376 first_vector(struct port_info *pi)
377 {
378 	struct adapter *sc = pi->adapter;
379 	int rc = T4_EXTRA_INTR, i;
380 
381 	if (sc->intr_count == 1)
382 		return (0);
383 
384 	for_each_port(sc, i) {
385 		struct port_info *p = sc->port[i];
386 
387 		if (i == pi->port_id)
388 			break;
389 
390 #ifdef TCP_OFFLOAD_ENABLE
391 		if (!(sc->flags & INTR_FWD))
392 			rc += p->nrxq + p->nofldrxq;
393 		else
394 			rc += max(p->nrxq, p->nofldrxq);
395 #else
396 		/*
397 		 * Not compiled with offload support and intr_count > 1.  Only
398 		 * NIC queues exist and they'd better be taking direct
399 		 * interrupts.
400 		 */
401 		ASSERT(!(sc->flags & INTR_FWD));
402 		rc += p->nrxq;
403 #endif
404 	}
405 	return (rc);
406 }
407 
408 /*
409  * Given an arbitrary "index," come up with an iq that can be used by other
410  * queues (of this port) for interrupt forwarding, SGE egress updates, etc.
411  * The iq returned is guaranteed to be something that takes direct interrupts.
412  */
413 static struct sge_iq *
414 port_intr_iq(struct port_info *pi, int idx)
415 {
416 	struct adapter *sc = pi->adapter;
417 	struct sge *s = &sc->sge;
418 	struct sge_iq *iq = NULL;
419 
420 	if (sc->intr_count == 1)
421 		return (&sc->sge.fwq);
422 
423 #ifdef TCP_OFFLOAD_ENABLE
424 	if (!(sc->flags & INTR_FWD)) {
425 		idx %= pi->nrxq + pi->nofldrxq;
426 
427 		if (idx >= pi->nrxq) {
428 			idx -= pi->nrxq;
429 			iq = &s->ofld_rxq[pi->first_ofld_rxq + idx].iq;
430 		} else
431 			iq = &s->rxq[pi->first_rxq + idx].iq;
432 
433 	} else {
434 		idx %= max(pi->nrxq, pi->nofldrxq);
435 
436 		if (pi->nrxq >= pi->nofldrxq)
437 			iq = &s->rxq[pi->first_rxq + idx].iq;
438 		else
439 			iq = &s->ofld_rxq[pi->first_ofld_rxq + idx].iq;
440 	}
441 #else
442 	/*
443 	 * Not compiled with offload support and intr_count > 1.  Only NIC
444 	 * queues exist and they'd better be taking direct interrupts.
445 	 */
446 	ASSERT(!(sc->flags & INTR_FWD));
447 
448 	idx %= pi->nrxq;
449 	iq = &s->rxq[pi->first_rxq + idx].iq;
450 #endif
451 
452 	return (iq);
453 }
454 
455 int
456 t4_setup_port_queues(struct port_info *pi)
457 {
458 	int rc = 0, i, intr_idx, j;
459 	struct sge_rxq *rxq;
460 	struct sge_txq *txq;
461 #ifdef TCP_OFFLOAD_ENABLE
462 	int iqid;
463 	struct sge_wrq *ctrlq;
464 	struct sge_ofld_rxq *ofld_rxq;
465 	struct sge_wrq *ofld_txq;
466 #endif
467 	struct adapter *sc = pi->adapter;
468 	struct driver_properties *p = &sc->props;
469 
470 	pi->ksp_config = setup_port_config_kstats(pi);
471 	pi->ksp_info   = setup_port_info_kstats(pi);
472 
473 	/* Interrupt vector to start from (when using multiple vectors) */
474 	intr_idx = first_vector(pi);
475 
476 	/*
477 	 * First pass over all rx queues (NIC and TOE):
478 	 * a) initialize iq and fl
479 	 * b) allocate queue iff it will take direct interrupts.
480 	 */
481 
482 	for_each_rxq(pi, i, rxq) {
483 
484 		init_iq(&rxq->iq, sc, pi->tmr_idx, pi->pktc_idx, p->qsize_rxq,
485 		    RX_IQ_ESIZE);
486 
487 		init_fl(&rxq->fl, p->qsize_rxq / 8); /* 8 bufs in each entry */
488 
489 		if ((!(sc->flags & INTR_FWD))
490 #ifdef TCP_OFFLOAD_ENABLE
491 		    || (sc->intr_count > 1 && pi->nrxq >= pi->nofldrxq)
492 #else
493 		    || (sc->intr_count > 1 && pi->nrxq)
494 #endif
495 		   ) {
496 			rxq->iq.flags |= IQ_INTR;
497 			rc = alloc_rxq(pi, rxq, intr_idx, i);
498 			if (rc != 0)
499 				goto done;
500 			intr_idx++;
501 		}
502 
503 	}
504 
505 #ifdef TCP_OFFLOAD_ENABLE
506 	for_each_ofld_rxq(pi, i, ofld_rxq) {
507 
508 		init_iq(&ofld_rxq->iq, sc, pi->tmr_idx, pi->pktc_idx,
509 		    p->qsize_rxq, RX_IQ_ESIZE);
510 
511 		init_fl(&ofld_rxq->fl, p->qsize_rxq / 8);
512 
513 		if (!(sc->flags & INTR_FWD) ||
514 		    (sc->intr_count > 1 && pi->nofldrxq > pi->nrxq)) {
515 			ofld_rxq->iq.flags = IQ_INTR;
516 			rc = alloc_ofld_rxq(pi, ofld_rxq, intr_idx);
517 			if (rc != 0)
518 				goto done;
519 
520 			intr_idx++;
521 		}
522 	}
523 #endif
524 
525 	/*
526 	 * Second pass over all rx queues (NIC and TOE).  The queues forwarding
527 	 * their interrupts are allocated now.
528 	 */
529 	j = 0;
530 	for_each_rxq(pi, i, rxq) {
531 		if (rxq->iq.flags & IQ_INTR)
532 			continue;
533 
534 		intr_idx = port_intr_iq(pi, j)->abs_id;
535 
536 		rc = alloc_rxq(pi, rxq, intr_idx, i);
537 		if (rc != 0)
538 			goto done;
539 		j++;
540 	}
541 
542 #ifdef TCP_OFFLOAD_ENABLE
543 	for_each_ofld_rxq(pi, i, ofld_rxq) {
544 		if (ofld_rxq->iq.flags & IQ_INTR)
545 			continue;
546 
547 		intr_idx = port_intr_iq(pi, j)->abs_id;
548 		rc = alloc_ofld_rxq(pi, ofld_rxq, intr_idx);
549 		if (rc != 0)
550 			goto done;
551 		j++;
552 	}
553 #endif
554 	/*
555 	 * Now the tx queues.  Only one pass needed.
556 	 */
557 	j = 0;
558 	for_each_txq(pi, i, txq) {
559 		uint16_t iqid;
560 
561 		iqid = port_intr_iq(pi, j)->cntxt_id;
562 		init_eq(sc, &txq->eq, EQ_ETH, p->qsize_txq, pi->tx_chan, iqid);
563 		rc = alloc_txq(pi, txq, i);
564 		if (rc != 0)
565 			goto done;
566 	}
567 
568 #ifdef TCP_OFFLOAD_ENABLE
569 	for_each_ofld_txq(pi, i, ofld_txq) {
570 		uint16_t iqid;
571 
572 		iqid = port_intr_iq(pi, j)->cntxt_id;
573 		init_eq(sc, &ofld_txq->eq, EQ_OFLD, p->qsize_txq, pi->tx_chan,
574 		    iqid);
575 		rc = alloc_wrq(sc, pi, ofld_txq, i);
576 		if (rc != 0)
577 			goto done;
578 	}
579 
580 	/*
581 	 * Finally, the control queue.
582 	 */
583 	ctrlq = &sc->sge.ctrlq[pi->port_id];
584 	iqid = port_intr_iq(pi, 0)->cntxt_id;
585 	init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, pi->tx_chan, iqid);
586 	rc = alloc_wrq(sc, pi, ctrlq, 0);
587 #endif
588 
589 done:
590 	if (rc != 0)
591 		(void) t4_teardown_port_queues(pi);
592 
593 	return (rc);
594 }
595 
596 /*
597  * Idempotent
598  */
599 int
600 t4_teardown_port_queues(struct port_info *pi)
601 {
602 	int i;
603 	struct sge_rxq *rxq;
604 	struct sge_txq *txq;
605 #ifdef TCP_OFFLOAD_ENABLE
606 	struct adapter *sc = pi->adapter;
607 	struct sge_ofld_rxq *ofld_rxq;
608 	struct sge_wrq *ofld_txq;
609 #endif
610 
611 	if (pi->ksp_config != NULL) {
612 		kstat_delete(pi->ksp_config);
613 		pi->ksp_config = NULL;
614 	}
615 	if (pi->ksp_info != NULL) {
616 		kstat_delete(pi->ksp_info);
617 		pi->ksp_info = NULL;
618 	}
619 
620 #ifdef TCP_OFFLOAD_ENABLE
621 	(void) free_wrq(sc, &sc->sge.ctrlq[pi->port_id]);
622 #endif
623 
624 	for_each_txq(pi, i, txq) {
625 		(void) free_txq(pi, txq);
626 	}
627 
628 #ifdef TCP_OFFLOAD_ENABLE
629 	for_each_ofld_txq(pi, i, ofld_txq) {
630 		(void) free_wrq(sc, ofld_txq);
631 	}
632 
633 	for_each_ofld_rxq(pi, i, ofld_rxq) {
634 		if ((ofld_rxq->iq.flags & IQ_INTR) == 0)
635 			(void) free_ofld_rxq(pi, ofld_rxq);
636 	}
637 #endif
638 
639 	for_each_rxq(pi, i, rxq) {
640 		if ((rxq->iq.flags & IQ_INTR) == 0)
641 			(void) free_rxq(pi, rxq);
642 	}
643 
644 	/*
645 	 * Then take down the rx queues that take direct interrupts.
646 	 */
647 
648 	for_each_rxq(pi, i, rxq) {
649 		if (rxq->iq.flags & IQ_INTR)
650 			(void) free_rxq(pi, rxq);
651 	}
652 
653 #ifdef TCP_OFFLOAD_ENABLE
654 	for_each_ofld_rxq(pi, i, ofld_rxq) {
655 		if (ofld_rxq->iq.flags & IQ_INTR)
656 			(void) free_ofld_rxq(pi, ofld_rxq);
657 	}
658 #endif
659 
660 	return (0);
661 }
662 
663 /* Deals with errors and forwarded interrupts */
664 uint_t
665 t4_intr_all(caddr_t arg1, caddr_t arg2)
666 {
667 
668 	(void) t4_intr_err(arg1, arg2);
669 	(void) t4_intr(arg1, arg2);
670 
671 	return (DDI_INTR_CLAIMED);
672 }
673 
674 static void
675 t4_intr_rx_work(struct sge_iq *iq)
676 {
677 	mblk_t *mp = NULL;
678 	struct sge_rxq *rxq = iq_to_rxq(iq);	/* Use iff iq is part of rxq */
679 	RXQ_LOCK(rxq);
680 	if (!iq->polling) {
681 		mp = t4_ring_rx(rxq, iq->qsize/8);
682 		t4_write_reg(iq->adapter, MYPF_REG(A_SGE_PF_GTS),
683 		     V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_next));
684 	}
685 	RXQ_UNLOCK(rxq);
686 	if (mp != NULL)
687 		mac_rx_ring(rxq->port->mh, rxq->ring_handle, mp,
688 			    rxq->ring_gen_num);
689 }
690 
691 /* Deals with interrupts on the given ingress queue */
692 /* ARGSUSED */
693 uint_t
694 t4_intr(caddr_t arg1, caddr_t arg2)
695 {
696 	struct sge_iq *iq = (struct sge_iq *)arg2;
697 	int state;
698 
699 	/* Right now receive polling is only enabled for MSI-X and
700 	 * when we have enough msi-x vectors i.e no interrupt forwarding.
701 	 */
702 	if (iq->adapter->props.multi_rings) {
703 		t4_intr_rx_work(iq);
704 	} else {
705 		state = atomic_cas_uint(&iq->state, IQS_IDLE, IQS_BUSY);
706 		if (state == IQS_IDLE) {
707 			(void) service_iq(iq, 0);
708 			(void) atomic_cas_uint(&iq->state, IQS_BUSY, IQS_IDLE);
709 		}
710 	}
711 	return (DDI_INTR_CLAIMED);
712 }
713 
714 /* Deals with error interrupts */
715 /* ARGSUSED */
716 uint_t
717 t4_intr_err(caddr_t arg1, caddr_t arg2)
718 {
719 	/* LINTED: E_BAD_PTR_CAST_ALIGN */
720 	struct adapter *sc = (struct adapter *)arg1;
721 
722 	t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
723 	(void) t4_slow_intr_handler(sc);
724 
725 	return (DDI_INTR_CLAIMED);
726 }
727 
728 /*
729  * t4_ring_rx - Process responses from an SGE response queue.
730  *
731  * This function processes responses from an SGE response queue up to the supplied budget.
732  * Responses include received packets as well as control messages from FW
733  * or HW.
734  * It returns a chain of mblks containing the received data, to be
735  * passed up to mac_ring_rx().
736  */
737 mblk_t *
738 t4_ring_rx(struct sge_rxq *rxq, int budget)
739 {
740 	struct sge_iq *iq = &rxq->iq;
741 	struct sge_fl *fl = &rxq->fl;           /* Use iff IQ_HAS_FL */
742 	struct adapter *sc = iq->adapter;
743 	struct rsp_ctrl *ctrl;
744 	const struct rss_header *rss;
745 	int ndescs = 0, fl_bufs_used = 0;
746 	int rsp_type;
747 	uint32_t lq;
748 	mblk_t *mblk_head = NULL, **mblk_tail, *m;
749 	struct cpl_rx_pkt *cpl;
750 	uint32_t received_bytes = 0, pkt_len = 0;
751 	bool csum_ok;
752 	uint16_t err_vec;
753 
754 	mblk_tail = &mblk_head;
755 
756 	while (is_new_response(iq, &ctrl)) {
757 
758 		membar_consumer();
759 
760 		m = NULL;
761 		rsp_type = G_RSPD_TYPE(ctrl->u.type_gen);
762 		lq = be32_to_cpu(ctrl->pldbuflen_qid);
763 		rss = (const void *)iq->cdesc;
764 
765 		switch (rsp_type) {
766 		case X_RSPD_TYPE_FLBUF:
767 
768 			ASSERT(iq->flags & IQ_HAS_FL);
769 
770 			if (CPL_RX_PKT == rss->opcode) {
771 				cpl = (void *)(rss + 1);
772 				pkt_len = be16_to_cpu(cpl->len);
773 
774 				if (iq->polling && ((received_bytes + pkt_len) > budget))
775 					goto done;
776 
777 				m = get_fl_payload(sc, fl, lq, &fl_bufs_used);
778 				if (m == NULL)
779 					goto done;
780 
781 				iq->intr_next = iq->intr_params;
782 				m->b_rptr += sc->sge.pktshift;
783 				if (sc->params.tp.rx_pkt_encap)
784 				/* It is enabled only in T6 config file */
785 					err_vec = G_T6_COMPR_RXERR_VEC(ntohs(cpl->err_vec));
786 				else
787 					err_vec = ntohs(cpl->err_vec);
788 
789 				csum_ok = cpl->csum_calc && !err_vec;
790 
791 				/* TODO: what about cpl->ip_frag? */
792 				if (csum_ok && !cpl->ip_frag) {
793 					mac_hcksum_set(m, 0, 0, 0, 0xffff,
794 					    HCK_FULLCKSUM_OK | HCK_FULLCKSUM |
795 					    HCK_IPV4_HDRCKSUM_OK);
796 					rxq->rxcsum++;
797 				}
798 				rxq->rxpkts++;
799 				rxq->rxbytes += pkt_len;
800 				received_bytes += pkt_len;
801 
802 				*mblk_tail = m;
803 				mblk_tail = &m->b_next;
804 
805 				break;
806 			}
807 
808 			m = get_fl_payload(sc, fl, lq, &fl_bufs_used);
809 			if (m == NULL)
810 				goto done;
811 			/* FALLTHROUGH */
812 
813 		case X_RSPD_TYPE_CPL:
814 			ASSERT(rss->opcode < NUM_CPL_CMDS);
815 			sc->cpl_handler[rss->opcode](iq, rss, m);
816 			break;
817 
818 		default:
819 			break;
820 		}
821 		iq_next(iq);
822 		++ndescs;
823 		if (!iq->polling && (ndescs == budget))
824 			break;
825 	}
826 
827 done:
828 
829 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
830 		     V_CIDXINC(ndescs) | V_INGRESSQID(iq->cntxt_id) |
831 		     V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
832 
833 	if ((fl_bufs_used > 0) || (iq->flags & IQ_HAS_FL)) {
834 		int starved;
835 		FL_LOCK(fl);
836 		fl->needed += fl_bufs_used;
837 		starved = refill_fl(sc, fl, fl->cap / 8);
838 		FL_UNLOCK(fl);
839 		if (starved)
840 			add_fl_to_sfl(sc, fl);
841 	}
842 	return (mblk_head);
843 }
844 
845 /*
846  * Deals with anything and everything on the given ingress queue.
847  */
848 static int
849 service_iq(struct sge_iq *iq, int budget)
850 {
851 	struct sge_iq *q;
852 	struct sge_rxq *rxq = iq_to_rxq(iq);	/* Use iff iq is part of rxq */
853 	struct sge_fl *fl = &rxq->fl;		/* Use iff IQ_HAS_FL */
854 	struct adapter *sc = iq->adapter;
855 	struct rsp_ctrl *ctrl;
856 	const struct rss_header *rss;
857 	int ndescs = 0, limit, fl_bufs_used = 0;
858 	int rsp_type;
859 	uint32_t lq;
860 	int starved;
861 	mblk_t *m;
862 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
863 
864 	limit = budget ? budget : iq->qsize / 8;
865 
866 	/*
867 	 * We always come back and check the descriptor ring for new indirect
868 	 * interrupts and other responses after running a single handler.
869 	 */
870 	for (;;) {
871 		while (is_new_response(iq, &ctrl)) {
872 
873 			membar_consumer();
874 
875 			m = NULL;
876 			rsp_type = G_RSPD_TYPE(ctrl->u.type_gen);
877 			lq = be32_to_cpu(ctrl->pldbuflen_qid);
878 			rss = (const void *)iq->cdesc;
879 
880 			switch (rsp_type) {
881 			case X_RSPD_TYPE_FLBUF:
882 
883 				ASSERT(iq->flags & IQ_HAS_FL);
884 
885 				m = get_fl_payload(sc, fl, lq, &fl_bufs_used);
886 				if (m == NULL) {
887 					/*
888 					 * Rearm the iq with a
889 					 * longer-than-default timer
890 					 */
891 					t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_CIDXINC(ndescs) |
892 							V_INGRESSQID((u32)iq->cntxt_id) |
893 							V_SEINTARM(V_QINTR_TIMER_IDX(SGE_NTIMERS-1)));
894 					if (fl_bufs_used > 0) {
895 						ASSERT(iq->flags & IQ_HAS_FL);
896 						FL_LOCK(fl);
897 						fl->needed += fl_bufs_used;
898 						starved = refill_fl(sc, fl, fl->cap / 8);
899 						FL_UNLOCK(fl);
900 						if (starved)
901 							add_fl_to_sfl(sc, fl);
902 					}
903 					return (0);
904 				}
905 
906 			/* FALLTHRU */
907 			case X_RSPD_TYPE_CPL:
908 
909 				ASSERT(rss->opcode < NUM_CPL_CMDS);
910 				sc->cpl_handler[rss->opcode](iq, rss, m);
911 				break;
912 
913 			case X_RSPD_TYPE_INTR:
914 
915 				/*
916 				 * Interrupts should be forwarded only to queues
917 				 * that are not forwarding their interrupts.
918 				 * This means service_iq can recurse but only 1
919 				 * level deep.
920 				 */
921 				ASSERT(budget == 0);
922 
923 				q = sc->sge.iqmap[lq - sc->sge.iq_start];
924 				if (atomic_cas_uint(&q->state, IQS_IDLE,
925 				    IQS_BUSY) == IQS_IDLE) {
926 					if (service_iq(q, q->qsize / 8) == 0) {
927 						(void) atomic_cas_uint(
928 						    &q->state, IQS_BUSY,
929 						    IQS_IDLE);
930 					} else {
931 						STAILQ_INSERT_TAIL(&iql, q,
932 						    link);
933 					}
934 				}
935 				break;
936 
937 			default:
938 				break;
939 			}
940 
941 			iq_next(iq);
942 			if (++ndescs == limit) {
943 				t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
944 				    V_CIDXINC(ndescs) |
945 				    V_INGRESSQID(iq->cntxt_id) |
946 				    V_SEINTARM(V_QINTR_TIMER_IDX(
947 				    X_TIMERREG_UPDATE_CIDX)));
948 				ndescs = 0;
949 
950 				if (fl_bufs_used > 0) {
951 					ASSERT(iq->flags & IQ_HAS_FL);
952 					FL_LOCK(fl);
953 					fl->needed += fl_bufs_used;
954 					(void) refill_fl(sc, fl, fl->cap / 8);
955 					FL_UNLOCK(fl);
956 					fl_bufs_used = 0;
957 				}
958 
959 				if (budget != 0)
960 					return (EINPROGRESS);
961 			}
962 		}
963 
964 		if (STAILQ_EMPTY(&iql) != 0)
965 			break;
966 
967 		/*
968 		 * Process the head only, and send it to the back of the list if
969 		 * it's still not done.
970 		 */
971 		q = STAILQ_FIRST(&iql);
972 		STAILQ_REMOVE_HEAD(&iql, link);
973 		if (service_iq(q, q->qsize / 8) == 0)
974 			(void) atomic_cas_uint(&q->state, IQS_BUSY, IQS_IDLE);
975 		else
976 			STAILQ_INSERT_TAIL(&iql, q, link);
977 	}
978 
979 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_CIDXINC(ndescs) |
980 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_next));
981 
982 	if (iq->flags & IQ_HAS_FL) {
983 
984 		FL_LOCK(fl);
985 		fl->needed += fl_bufs_used;
986 		starved = refill_fl(sc, fl, fl->cap / 4);
987 		FL_UNLOCK(fl);
988 		if (starved != 0)
989 			add_fl_to_sfl(sc, fl);
990 	}
991 
992 	return (0);
993 }
994 
995 #ifdef TCP_OFFLOAD_ENABLE
996 int
997 t4_mgmt_tx(struct adapter *sc, mblk_t *m)
998 {
999 	return (t4_wrq_tx(sc, &sc->sge.mgmtq, m));
1000 }
1001 
1002 /*
1003  * Doesn't fail.  Holds on to work requests it can't send right away.
1004  */
1005 int
1006 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, mblk_t *m0)
1007 {
1008 	struct sge_eq *eq = &wrq->eq;
1009 	struct mblk_pair *wr_list = &wrq->wr_list;
1010 	int can_reclaim;
1011 	caddr_t dst;
1012 	mblk_t *wr, *next;
1013 
1014 	TXQ_LOCK_ASSERT_OWNED(wrq);
1015 #ifdef TCP_OFFLOAD_ENABLE
1016 	ASSERT((eq->flags & EQ_TYPEMASK) == EQ_OFLD ||
1017 	    (eq->flags & EQ_TYPEMASK) == EQ_CTRL);
1018 #else
1019 	ASSERT((eq->flags & EQ_TYPEMASK) == EQ_CTRL);
1020 #endif
1021 
1022 	if (m0 != NULL) {
1023 		if (wr_list->head != NULL)
1024 			wr_list->tail->b_next = m0;
1025 		else
1026 			wr_list->head = m0;
1027 		while (m0->b_next)
1028 			m0 = m0->b_next;
1029 		wr_list->tail = m0;
1030 	}
1031 
1032 	can_reclaim = reclaimable(eq);
1033 	eq->cidx += can_reclaim;
1034 	eq->avail += can_reclaim;
1035 	if (eq->cidx >= eq->cap)
1036 		eq->cidx -= eq->cap;
1037 
1038 	for (wr = wr_list->head; wr; wr = next) {
1039 		int ndesc, len = 0;
1040 		mblk_t *m;
1041 
1042 		next = wr->b_next;
1043 		wr->b_next = NULL;
1044 
1045 		for (m = wr; m; m = m->b_cont)
1046 			len += MBLKL(m);
1047 
1048 		ASSERT(len > 0 && (len & 0x7) == 0);
1049 		ASSERT(len <= SGE_MAX_WR_LEN);
1050 
1051 		ndesc = howmany(len, EQ_ESIZE);
1052 		if (eq->avail < ndesc) {
1053 			wr->b_next = next;
1054 			wrq->no_desc++;
1055 			break;
1056 		}
1057 
1058 		dst = (void *)&eq->desc[eq->pidx];
1059 		for (m = wr; m; m = m->b_cont)
1060 			copy_to_txd(eq, (void *)m->b_rptr, &dst, MBLKL(m));
1061 
1062 		eq->pidx += ndesc;
1063 		eq->avail -= ndesc;
1064 		if (eq->pidx >= eq->cap)
1065 			eq->pidx -= eq->cap;
1066 
1067 		eq->pending += ndesc;
1068 		if (eq->pending > 16)
1069 			ring_tx_db(sc, eq);
1070 
1071 		wrq->tx_wrs++;
1072 		freemsg(wr);
1073 
1074 		if (eq->avail < 8) {
1075 			can_reclaim = reclaimable(eq);
1076 			eq->cidx += can_reclaim;
1077 			eq->avail += can_reclaim;
1078 			if (eq->cidx >= eq->cap)
1079 				eq->cidx -= eq->cap;
1080 		}
1081 	}
1082 
1083 	if (eq->pending != 0)
1084 		ring_tx_db(sc, eq);
1085 
1086 	if (wr == NULL)
1087 		wr_list->head = wr_list->tail = NULL;
1088 	else {
1089 		wr_list->head = wr;
1090 
1091 		ASSERT(wr_list->tail->b_next == NULL);
1092 	}
1093 
1094 	return (0);
1095 }
1096 #endif
1097 
1098 /* Per-packet header in a coalesced tx WR, before the SGL starts (in flits) */
1099 #define	TXPKTS_PKT_HDR ((\
1100 	sizeof (struct ulp_txpkt) + \
1101 	sizeof (struct ulptx_idata) + \
1102 	sizeof (struct cpl_tx_pkt_core)) / 8)
1103 
1104 /* Header of a coalesced tx WR, before SGL of first packet (in flits) */
1105 #define	TXPKTS_WR_HDR (\
1106 	sizeof (struct fw_eth_tx_pkts_wr) / 8 + \
1107 	TXPKTS_PKT_HDR)
1108 
1109 /* Header of a tx WR, before SGL of first packet (in flits) */
1110 #define	TXPKT_WR_HDR ((\
1111 	sizeof (struct fw_eth_tx_pkt_wr) + \
1112 	sizeof (struct cpl_tx_pkt_core)) / 8)
1113 
1114 /* Header of a tx LSO WR, before SGL of first packet (in flits) */
1115 #define	TXPKT_LSO_WR_HDR ((\
1116 	sizeof (struct fw_eth_tx_pkt_wr) + \
1117 	sizeof(struct cpl_tx_pkt_lso_core) + \
1118 	sizeof (struct cpl_tx_pkt_core)) / 8)
1119 
1120 mblk_t *
1121 t4_eth_tx(void *arg, mblk_t *frame)
1122 {
1123 	struct sge_txq *txq = (struct sge_txq *) arg;
1124 	struct port_info *pi = txq->port;
1125 	struct adapter *sc = pi->adapter;
1126 	struct sge_eq *eq = &txq->eq;
1127 	mblk_t *next_frame;
1128 	int rc, coalescing;
1129 	struct txpkts txpkts;
1130 	struct txinfo txinfo;
1131 
1132 	txpkts.npkt = 0; /* indicates there's nothing in txpkts */
1133 	coalescing = 0;
1134 
1135 	TXQ_LOCK(txq);
1136 	if (eq->avail < 8)
1137 		(void) reclaim_tx_descs(txq, 8);
1138 	for (; frame; frame = next_frame) {
1139 
1140 		if (eq->avail < 8)
1141 			break;
1142 
1143 		next_frame = frame->b_next;
1144 		frame->b_next = NULL;
1145 
1146 		if (next_frame != NULL)
1147 			coalescing = 1;
1148 
1149 		rc = get_frame_txinfo(txq, &frame, &txinfo, coalescing);
1150 		if (rc != 0) {
1151 			if (rc == ENOMEM) {
1152 
1153 				/* Short of resources, suspend tx */
1154 
1155 				frame->b_next = next_frame;
1156 				break;
1157 			}
1158 
1159 			/*
1160 			 * Unrecoverable error for this frame, throw it
1161 			 * away and move on to the next.
1162 			 */
1163 
1164 			freemsg(frame);
1165 			continue;
1166 		}
1167 
1168 		if (coalescing != 0 &&
1169 		    add_to_txpkts(txq, &txpkts, frame, &txinfo) == 0) {
1170 
1171 			/* Successfully absorbed into txpkts */
1172 
1173 			write_ulp_cpl_sgl(pi, txq, &txpkts, &txinfo);
1174 			goto doorbell;
1175 		}
1176 
1177 		/*
1178 		 * We weren't coalescing to begin with, or current frame could
1179 		 * not be coalesced (add_to_txpkts flushes txpkts if a frame
1180 		 * given to it can't be coalesced).  Either way there should be
1181 		 * nothing in txpkts.
1182 		 */
1183 		ASSERT(txpkts.npkt == 0);
1184 
1185 		/* We're sending out individual frames now */
1186 		coalescing = 0;
1187 
1188 		if (eq->avail < 8)
1189 			(void) reclaim_tx_descs(txq, 8);
1190 		rc = write_txpkt_wr(pi, txq, frame, &txinfo);
1191 		if (rc != 0) {
1192 
1193 			/* Short of hardware descriptors, suspend tx */
1194 
1195 			/*
1196 			 * This is an unlikely but expensive failure.  We've
1197 			 * done all the hard work (DMA bindings etc.) and now we
1198 			 * can't send out the frame.  What's worse, we have to
1199 			 * spend even more time freeing up everything in txinfo.
1200 			 */
1201 			txq->qfull++;
1202 			free_txinfo_resources(txq, &txinfo);
1203 
1204 			frame->b_next = next_frame;
1205 			break;
1206 		}
1207 
1208 doorbell:
1209 		/* Fewer and fewer doorbells as the queue fills up */
1210 		if (eq->pending >= (1 << (fls(eq->qsize - eq->avail) / 2))) {
1211 			txq->txbytes += txinfo.len;
1212 			txq->txpkts++;
1213 			ring_tx_db(sc, eq);
1214 		}
1215 		(void) reclaim_tx_descs(txq, 32);
1216 	}
1217 
1218 	if (txpkts.npkt > 0)
1219 		write_txpkts_wr(txq, &txpkts);
1220 
1221 	/*
1222 	 * frame not NULL means there was an error but we haven't thrown it
1223 	 * away.  This can happen when we're short of tx descriptors (qfull) or
1224 	 * maybe even DMA handles (dma_hdl_failed).  Either way, a credit flush
1225 	 * and reclaim will get things going again.
1226 	 *
1227 	 * If eq->avail is already 0 we know a credit flush was requested in the
1228 	 * WR that reduced it to 0 so we don't need another flush (we don't have
1229 	 * any descriptor for a flush WR anyway, duh).
1230 	 */
1231 	if (frame && eq->avail > 0)
1232 		write_txqflush_wr(txq);
1233 
1234 	if (eq->pending != 0)
1235 		ring_tx_db(sc, eq);
1236 
1237 	(void) reclaim_tx_descs(txq, eq->qsize);
1238 	TXQ_UNLOCK(txq);
1239 
1240 	return (frame);
1241 }
1242 
1243 static inline void
1244 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int8_t pktc_idx,
1245 	int qsize, uint8_t esize)
1246 {
1247 	ASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS);
1248 	ASSERT(pktc_idx < SGE_NCOUNTERS);	/* -ve is ok, means don't use */
1249 
1250 	iq->flags = 0;
1251 	iq->adapter = sc;
1252 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
1253 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
1254 	if (pktc_idx >= 0) {
1255 		iq->intr_params |= F_QINTR_CNT_EN;
1256 		iq->intr_pktc_idx = pktc_idx;
1257 	}
1258 	iq->qsize = roundup(qsize, 16);		/* See FW_IQ_CMD/iqsize */
1259 	iq->esize = max(esize, 16);		/* See FW_IQ_CMD/iqesize */
1260 }
1261 
1262 static inline void
1263 init_fl(struct sge_fl *fl, uint16_t qsize)
1264 {
1265 
1266 	fl->qsize = qsize;
1267 	fl->allocb_fail = 0;
1268 }
1269 
1270 static inline void
1271 init_eq(struct adapter *sc, struct sge_eq *eq, uint16_t eqtype, uint16_t qsize,
1272     uint8_t tx_chan, uint16_t iqid)
1273 {
1274 	struct sge *s = &sc->sge;
1275 	uint32_t r;
1276 
1277 	ASSERT(tx_chan < NCHAN);
1278 	ASSERT(eqtype <= EQ_TYPEMASK);
1279 
1280 	if (is_t5(sc->params.chip)) {
1281 		r = t4_read_reg(sc, A_SGE_EGRESS_QUEUES_PER_PAGE_PF);
1282 		r >>= S_QUEUESPERPAGEPF0 +
1283 		    (S_QUEUESPERPAGEPF1 - S_QUEUESPERPAGEPF0) * sc->pf;
1284 		s->s_qpp = r & M_QUEUESPERPAGEPF0;
1285 	}
1286 
1287 	eq->flags = eqtype & EQ_TYPEMASK;
1288 	eq->tx_chan = tx_chan;
1289 	eq->iqid = iqid;
1290 	eq->qsize = qsize;
1291 }
1292 
1293 /*
1294  * Allocates the ring for an ingress queue and an optional freelist.  If the
1295  * freelist is specified it will be allocated and then associated with the
1296  * ingress queue.
1297  *
1298  * Returns errno on failure.  Resources allocated up to that point may still be
1299  * allocated.  Caller is responsible for cleanup in case this function fails.
1300  *
1301  * If the ingress queue will take interrupts directly (iq->flags & IQ_INTR) then
1302  * the intr_idx specifies the vector, starting from 0.  Otherwise it specifies
1303  * the index of the queue to which its interrupts will be forwarded.
1304  */
1305 static int
1306 alloc_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl,
1307     int intr_idx, int cong)
1308 {
1309 	int rc, i, cntxt_id;
1310 	size_t len;
1311 	struct fw_iq_cmd c;
1312 	struct adapter *sc = iq->adapter;
1313 	uint32_t v = 0;
1314 
1315 	len = iq->qsize * iq->esize;
1316 	rc = alloc_desc_ring(sc, len, DDI_DMA_READ, &iq->dhdl, &iq->ahdl,
1317 	    &iq->ba, (caddr_t *)&iq->desc);
1318 	if (rc != 0)
1319 		return (rc);
1320 
1321 	bzero(&c, sizeof (c));
1322 	c.op_to_vfn = cpu_to_be32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
1323 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
1324 	    V_FW_IQ_CMD_VFN(0));
1325 
1326 	c.alloc_to_len16 = cpu_to_be32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
1327 	    FW_LEN16(c));
1328 
1329 	/* Special handling for firmware event queue */
1330 	if (iq == &sc->sge.fwq)
1331 		v |= F_FW_IQ_CMD_IQASYNCH;
1332 
1333 	if (iq->flags & IQ_INTR)
1334 		ASSERT(intr_idx < sc->intr_count);
1335 	else
1336 		v |= F_FW_IQ_CMD_IQANDST;
1337 	v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
1338 
1339 	c.type_to_iqandstindex = cpu_to_be32(v |
1340 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
1341 	    V_FW_IQ_CMD_VIID(pi->viid) |
1342 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
1343 	c.iqdroprss_to_iqesize = cpu_to_be16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
1344 	    F_FW_IQ_CMD_IQGTSMODE |
1345 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
1346 	    V_FW_IQ_CMD_IQESIZE(ilog2(iq->esize) - 4));
1347 	c.iqsize = cpu_to_be16(iq->qsize);
1348 	c.iqaddr = cpu_to_be64(iq->ba);
1349 	if (cong >= 0)
1350 		c.iqns_to_fl0congen = BE_32(F_FW_IQ_CMD_IQFLINTCONGEN);
1351 
1352 	if (fl != NULL) {
1353 		unsigned int chip_ver = CHELSIO_CHIP_VERSION(sc->params.chip);
1354 
1355 		mutex_init(&fl->lock, NULL, MUTEX_DRIVER,
1356 		    DDI_INTR_PRI(sc->intr_pri));
1357 		fl->flags |= FL_MTX;
1358 
1359 		len = fl->qsize * RX_FL_ESIZE;
1360 		rc = alloc_desc_ring(sc, len, DDI_DMA_WRITE, &fl->dhdl,
1361 		    &fl->ahdl, &fl->ba, (caddr_t *)&fl->desc);
1362 		if (rc != 0)
1363 			return (rc);
1364 
1365 		/* Allocate space for one software descriptor per buffer. */
1366 		fl->cap = (fl->qsize - sc->sge.stat_len / RX_FL_ESIZE) * 8;
1367 		fl->sdesc = kmem_zalloc(sizeof (struct fl_sdesc) * fl->cap,
1368 		    KM_SLEEP);
1369 		fl->needed = fl->cap;
1370 		fl->lowat = roundup(sc->sge.fl_starve_threshold, 8);
1371 
1372 		c.iqns_to_fl0congen |=
1373 		    cpu_to_be32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
1374 		    F_FW_IQ_CMD_FL0PACKEN | F_FW_IQ_CMD_FL0PADEN);
1375 		if (cong >= 0) {
1376 			c.iqns_to_fl0congen |=
1377 			    BE_32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
1378 			    F_FW_IQ_CMD_FL0CONGCIF |
1379 			    F_FW_IQ_CMD_FL0CONGEN);
1380 		}
1381 
1382 		/* In T6, for egress queue type FL there is internal overhead
1383 		 * of 16B for header going into FLM module.  Hence the maximum
1384 		 * allowed burst size is 448 bytes.  For T4/T5, the hardware
1385 		 * doesn't coalesce fetch requests if more than 64 bytes of
1386 		 * Free List pointers are provided, so we use a 128-byte Fetch
1387 		 * Burst Minimum there (T6 implements coalescing so we can use
1388 		 * the smaller 64-byte value there).
1389 		 */
1390 
1391 		c.fl0dcaen_to_fl0cidxfthresh =
1392 		    cpu_to_be16(V_FW_IQ_CMD_FL0FBMIN(chip_ver <= CHELSIO_T5
1393 						     ? X_FETCHBURSTMIN_128B
1394 						     : X_FETCHBURSTMIN_64B) |
1395 		    V_FW_IQ_CMD_FL0FBMAX(chip_ver <= CHELSIO_T5
1396 					 ? X_FETCHBURSTMAX_512B
1397 					 : X_FETCHBURSTMAX_256B));
1398 		c.fl0size = cpu_to_be16(fl->qsize);
1399 		c.fl0addr = cpu_to_be64(fl->ba);
1400 	}
1401 
1402 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof (c), &c);
1403 	if (rc != 0) {
1404 		cxgb_printf(sc->dip, CE_WARN,
1405 		    "failed to create ingress queue: %d", rc);
1406 		return (rc);
1407 	}
1408 
1409 	iq->cdesc = iq->desc;
1410 	iq->cidx = 0;
1411 	iq->gen = 1;
1412 	iq->intr_next = iq->intr_params;
1413 	iq->adapter = sc;
1414 	iq->cntxt_id = be16_to_cpu(c.iqid);
1415 	iq->abs_id = be16_to_cpu(c.physiqid);
1416 	iq->flags |= IQ_ALLOCATED;
1417 	mutex_init(&iq->lock, NULL,
1418 		    MUTEX_DRIVER, DDI_INTR_PRI(DDI_INTR_PRI(sc->intr_pri)));
1419 	iq->polling = 0;
1420 
1421 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
1422 	if (cntxt_id >= sc->sge.niq) {
1423 		panic("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
1424 		    cntxt_id, sc->sge.niq - 1);
1425 	}
1426 	sc->sge.iqmap[cntxt_id] = iq;
1427 
1428 	if (fl != NULL) {
1429 		fl->cntxt_id = be16_to_cpu(c.fl0id);
1430 		fl->pidx = fl->cidx = 0;
1431 		fl->copy_threshold = rx_copy_threshold;
1432 
1433 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
1434 		if (cntxt_id >= sc->sge.neq) {
1435 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
1436 			    __func__, cntxt_id, sc->sge.neq - 1);
1437 		}
1438 		sc->sge.eqmap[cntxt_id] = (void *)fl;
1439 
1440 		FL_LOCK(fl);
1441 		(void) refill_fl(sc, fl, fl->lowat);
1442 		FL_UNLOCK(fl);
1443 
1444 		iq->flags |= IQ_HAS_FL;
1445 	}
1446 
1447 	if (is_t5(sc->params.chip) && cong >= 0) {
1448 		uint32_t param, val;
1449 
1450 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
1451 			V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
1452 			V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
1453 		if (cong == 0)
1454 			val = 1 << 19;
1455 		else {
1456 			val = 2 << 19;
1457 			for (i = 0; i < 4; i++) {
1458 				if (cong & (1 << i))
1459 					val |= 1 << (i << 2);
1460 			}
1461 		}
1462 
1463 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
1464 		if (rc != 0) {
1465 			/* report error but carry on */
1466 			cxgb_printf(sc->dip, CE_WARN,
1467 			    "failed to set congestion manager context for "
1468 			    "ingress queue %d: %d", iq->cntxt_id, rc);
1469 		}
1470 	}
1471 
1472 	/* Enable IQ interrupts */
1473 	iq->state = IQS_IDLE;
1474 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_SEINTARM(iq->intr_params) |
1475 	    V_INGRESSQID(iq->cntxt_id));
1476 
1477 	return (0);
1478 }
1479 
1480 static int
1481 free_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl)
1482 {
1483 	int rc;
1484 	struct adapter *sc = iq->adapter;
1485 	dev_info_t *dip;
1486 
1487 	dip = pi ? pi->dip : sc->dip;
1488 
1489 	if (iq != NULL) {
1490 		if (iq->flags & IQ_ALLOCATED) {
1491 			rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
1492 			    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
1493 			    fl ? fl->cntxt_id : 0xffff, 0xffff);
1494 			if (rc != 0) {
1495 				cxgb_printf(dip, CE_WARN,
1496 				    "failed to free queue %p: %d", iq, rc);
1497 				return (rc);
1498 			}
1499 			mutex_destroy(&iq->lock);
1500 			iq->flags &= ~IQ_ALLOCATED;
1501 		}
1502 
1503 		if (iq->desc != NULL) {
1504 			(void) free_desc_ring(&iq->dhdl, &iq->ahdl);
1505 			iq->desc = NULL;
1506 		}
1507 
1508 		bzero(iq, sizeof (*iq));
1509 	}
1510 
1511 	if (fl != NULL) {
1512 		if (fl->sdesc != NULL) {
1513 			FL_LOCK(fl);
1514 			free_fl_bufs(fl);
1515 			FL_UNLOCK(fl);
1516 
1517 			kmem_free(fl->sdesc, sizeof (struct fl_sdesc) *
1518 			    fl->cap);
1519 			fl->sdesc = NULL;
1520 		}
1521 
1522 		if (fl->desc != NULL) {
1523 			(void) free_desc_ring(&fl->dhdl, &fl->ahdl);
1524 			fl->desc = NULL;
1525 		}
1526 
1527 		if (fl->flags & FL_MTX) {
1528 			mutex_destroy(&fl->lock);
1529 			fl->flags &= ~FL_MTX;
1530 		}
1531 
1532 		bzero(fl, sizeof (struct sge_fl));
1533 	}
1534 
1535 	return (0);
1536 }
1537 
1538 static int
1539 alloc_fwq(struct adapter *sc)
1540 {
1541 	int rc, intr_idx;
1542 	struct sge_iq *fwq = &sc->sge.fwq;
1543 
1544 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE, FW_IQ_ESIZE);
1545 	fwq->flags |= IQ_INTR;	/* always */
1546 	intr_idx = sc->intr_count > 1 ? 1 : 0;
1547 	rc = alloc_iq_fl(sc->port[0], fwq, NULL, intr_idx, -1);
1548 	if (rc != 0) {
1549 		cxgb_printf(sc->dip, CE_WARN,
1550 		    "failed to create firmware event queue: %d.", rc);
1551 		return (rc);
1552 	}
1553 
1554 	return (0);
1555 }
1556 
1557 static int
1558 free_fwq(struct adapter *sc)
1559 {
1560 
1561 	return (free_iq_fl(NULL, &sc->sge.fwq, NULL));
1562 }
1563 
1564 #ifdef TCP_OFFLOAD_ENABLE
1565 static int
1566 alloc_mgmtq(struct adapter *sc)
1567 {
1568 	int rc;
1569 	struct sge_wrq *mgmtq = &sc->sge.mgmtq;
1570 
1571 	init_eq(sc, &mgmtq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[0]->tx_chan,
1572 	    sc->sge.fwq.cntxt_id);
1573 	rc = alloc_wrq(sc, NULL, mgmtq, 0);
1574 	if (rc != 0) {
1575 		cxgb_printf(sc->dip, CE_WARN,
1576 		    "failed to create management queue: %d\n", rc);
1577 		return (rc);
1578 	}
1579 
1580 	return (0);
1581 }
1582 #endif
1583 
1584 static int
1585 alloc_rxq(struct port_info *pi, struct sge_rxq *rxq, int intr_idx, int i)
1586 {
1587 	int rc;
1588 
1589 	rxq->port = pi;
1590 	rc = alloc_iq_fl(pi, &rxq->iq, &rxq->fl, intr_idx, 1 << pi->tx_chan);
1591 	if (rc != 0)
1592 		return (rc);
1593 
1594 	rxq->ksp = setup_rxq_kstats(pi, rxq, i);
1595 
1596 	return (rc);
1597 }
1598 
1599 static int
1600 free_rxq(struct port_info *pi, struct sge_rxq *rxq)
1601 {
1602 	int rc;
1603 
1604 	if (rxq->ksp != NULL) {
1605 		kstat_delete(rxq->ksp);
1606 		rxq->ksp = NULL;
1607 	}
1608 
1609 	rc = free_iq_fl(pi, &rxq->iq, &rxq->fl);
1610 	if (rc == 0)
1611 		bzero(&rxq->fl, sizeof (*rxq) - offsetof(struct sge_rxq, fl));
1612 
1613 	return (rc);
1614 }
1615 
1616 #ifdef TCP_OFFLOAD_ENABLE
1617 static int
1618 alloc_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq,
1619 	int intr_idx)
1620 {
1621 	int rc;
1622 
1623 	rc = alloc_iq_fl(pi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx,
1624 	    1 << pi->tx_chan);
1625 	if (rc != 0)
1626 		return (rc);
1627 
1628 	return (rc);
1629 }
1630 
1631 static int
1632 free_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq)
1633 {
1634 	int rc;
1635 
1636 	rc = free_iq_fl(pi, &ofld_rxq->iq, &ofld_rxq->fl);
1637 	if (rc == 0)
1638 		bzero(&ofld_rxq->fl, sizeof (*ofld_rxq) -
1639 		    offsetof(struct sge_ofld_rxq, fl));
1640 
1641 	return (rc);
1642 }
1643 #endif
1644 
1645 static int
1646 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
1647 {
1648 	int rc, cntxt_id;
1649 	struct fw_eq_ctrl_cmd c;
1650 
1651 	bzero(&c, sizeof (c));
1652 
1653 	c.op_to_vfn = BE_32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
1654 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
1655 	    V_FW_EQ_CTRL_CMD_VFN(0));
1656 	c.alloc_to_len16 = BE_32(F_FW_EQ_CTRL_CMD_ALLOC |
1657 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
1658 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid)); /* TODO */
1659 	c.physeqid_pkd = BE_32(0);
1660 	c.fetchszm_to_iqid =
1661 	    BE_32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
1662 	    V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
1663 	    F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
1664 	c.dcaen_to_eqsize =
1665 	    BE_32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
1666 	    V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
1667 	    V_FW_EQ_CTRL_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
1668 	    V_FW_EQ_CTRL_CMD_EQSIZE(eq->qsize));
1669 	c.eqaddr = BE_64(eq->ba);
1670 
1671 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof (c), &c);
1672 	if (rc != 0) {
1673 		cxgb_printf(sc->dip, CE_WARN,
1674 		    "failed to create control queue %d: %d", eq->tx_chan, rc);
1675 		return (rc);
1676 	}
1677 	eq->flags |= EQ_ALLOCATED;
1678 
1679 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(BE_32(c.cmpliqid_eqid));
1680 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
1681 	if (cntxt_id >= sc->sge.neq)
1682 		panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
1683 		    cntxt_id, sc->sge.neq - 1);
1684 	sc->sge.eqmap[cntxt_id] = eq;
1685 
1686 	return (rc);
1687 }
1688 
1689 static int
1690 eth_eq_alloc(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
1691 {
1692 	int rc, cntxt_id;
1693 	struct fw_eq_eth_cmd c;
1694 
1695 	bzero(&c, sizeof (c));
1696 
1697 	c.op_to_vfn = BE_32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
1698 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
1699 	    V_FW_EQ_ETH_CMD_VFN(0));
1700 	c.alloc_to_len16 = BE_32(F_FW_EQ_ETH_CMD_ALLOC |
1701 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
1702 	c.autoequiqe_to_viid = BE_32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
1703 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(pi->viid));
1704 	c.fetchszm_to_iqid =
1705 	    BE_32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
1706 	    V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
1707 	    V_FW_EQ_ETH_CMD_IQID(eq->iqid));
1708 	c.dcaen_to_eqsize = BE_32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
1709 	    V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
1710 	    V_FW_EQ_ETH_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
1711 	    V_FW_EQ_ETH_CMD_EQSIZE(eq->qsize));
1712 	c.eqaddr = BE_64(eq->ba);
1713 
1714 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof (c), &c);
1715 	if (rc != 0) {
1716 		cxgb_printf(pi->dip, CE_WARN,
1717 		    "failed to create Ethernet egress queue: %d", rc);
1718 		return (rc);
1719 	}
1720 	eq->flags |= EQ_ALLOCATED;
1721 
1722 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(BE_32(c.eqid_pkd));
1723 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
1724 	if (cntxt_id >= sc->sge.neq)
1725 		panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
1726 		    cntxt_id, sc->sge.neq - 1);
1727 	sc->sge.eqmap[cntxt_id] = eq;
1728 
1729 	return (rc);
1730 }
1731 
1732 #ifdef TCP_OFFLOAD_ENABLE
1733 static int
1734 ofld_eq_alloc(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
1735 {
1736 	int rc, cntxt_id;
1737 	struct fw_eq_ofld_cmd c;
1738 
1739 	bzero(&c, sizeof (c));
1740 
1741 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
1742 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
1743 	    V_FW_EQ_OFLD_CMD_VFN(0));
1744 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
1745 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
1746 	c.fetchszm_to_iqid =
1747 	    htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
1748 	    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
1749 	    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
1750 	c.dcaen_to_eqsize =
1751 	    BE_32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
1752 	    V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
1753 	    V_FW_EQ_OFLD_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
1754 	    V_FW_EQ_OFLD_CMD_EQSIZE(eq->qsize));
1755 	c.eqaddr = BE_64(eq->ba);
1756 
1757 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof (c), &c);
1758 	if (rc != 0) {
1759 		cxgb_printf(pi->dip, CE_WARN,
1760 		    "failed to create egress queue for TCP offload: %d", rc);
1761 		return (rc);
1762 	}
1763 	eq->flags |= EQ_ALLOCATED;
1764 
1765 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(BE_32(c.eqid_pkd));
1766 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
1767 	if (cntxt_id >= sc->sge.neq)
1768 		panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
1769 		    cntxt_id, sc->sge.neq - 1);
1770 	sc->sge.eqmap[cntxt_id] = eq;
1771 
1772 	return (rc);
1773 }
1774 #endif
1775 
1776 static int
1777 alloc_eq(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
1778 {
1779 	int rc;
1780 	size_t len;
1781 
1782 	mutex_init(&eq->lock, NULL, MUTEX_DRIVER, DDI_INTR_PRI(sc->intr_pri));
1783 	eq->flags |= EQ_MTX;
1784 
1785 	len = eq->qsize * EQ_ESIZE;
1786 	rc = alloc_desc_ring(sc, len, DDI_DMA_WRITE, &eq->desc_dhdl,
1787 	    &eq->desc_ahdl, &eq->ba, (caddr_t *)&eq->desc);
1788 	if (rc != 0)
1789 		return (rc);
1790 
1791 	eq->cap = eq->qsize - sc->sge.stat_len / EQ_ESIZE;
1792 	eq->spg = (void *)&eq->desc[eq->cap];
1793 	eq->avail = eq->cap - 1;	/* one less to avoid cidx = pidx */
1794 	eq->pidx = eq->cidx = 0;
1795 	eq->doorbells = sc->doorbells;
1796 
1797 	switch (eq->flags & EQ_TYPEMASK) {
1798 	case EQ_CTRL:
1799 		rc = ctrl_eq_alloc(sc, eq);
1800 		break;
1801 
1802 	case EQ_ETH:
1803 		rc = eth_eq_alloc(sc, pi, eq);
1804 		break;
1805 
1806 #ifdef TCP_OFFLOAD_ENABLE
1807 	case EQ_OFLD:
1808 		rc = ofld_eq_alloc(sc, pi, eq);
1809 		break;
1810 #endif
1811 
1812 	default:
1813 		panic("%s: invalid eq type %d.", __func__,
1814 		    eq->flags & EQ_TYPEMASK);
1815 	}
1816 
1817 	if (eq->doorbells &
1818 		(DOORBELL_UDB | DOORBELL_UDBWC | DOORBELL_WCWR)) {
1819 		uint32_t s_qpp = sc->sge.s_qpp;
1820 		uint32_t mask = (1 << s_qpp) - 1;
1821 		volatile uint8_t *udb;
1822 
1823 		udb = (volatile uint8_t *)sc->reg1p + UDBS_DB_OFFSET;
1824 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;   /* pg offset */
1825 		eq->udb_qid = eq->cntxt_id & mask;              /* id in page */
1826 		if (eq->udb_qid > PAGE_SIZE / UDBS_SEG_SIZE)
1827 			eq->doorbells &= ~DOORBELL_WCWR;
1828 		else {
1829 			udb += eq->udb_qid << UDBS_SEG_SHIFT;   /* seg offset */
1830 			eq->udb_qid = 0;
1831 		}
1832 		eq->udb = (volatile void *)udb;
1833 	}
1834 
1835 	if (rc != 0) {
1836 		cxgb_printf(sc->dip, CE_WARN,
1837 		    "failed to allocate egress queue(%d): %d",
1838 		    eq->flags & EQ_TYPEMASK, rc);
1839 	}
1840 
1841 	return (rc);
1842 }
1843 
1844 static int
1845 free_eq(struct adapter *sc, struct sge_eq *eq)
1846 {
1847 	int rc;
1848 
1849 	if (eq->flags & EQ_ALLOCATED) {
1850 		switch (eq->flags & EQ_TYPEMASK) {
1851 		case EQ_CTRL:
1852 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
1853 			    eq->cntxt_id);
1854 			break;
1855 
1856 		case EQ_ETH:
1857 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
1858 			    eq->cntxt_id);
1859 			break;
1860 
1861 #ifdef TCP_OFFLOAD_ENABLE
1862 		case EQ_OFLD:
1863 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
1864 			    eq->cntxt_id);
1865 			break;
1866 #endif
1867 
1868 		default:
1869 			panic("%s: invalid eq type %d.", __func__,
1870 			    eq->flags & EQ_TYPEMASK);
1871 		}
1872 		if (rc != 0) {
1873 			cxgb_printf(sc->dip, CE_WARN,
1874 			    "failed to free egress queue (%d): %d",
1875 			    eq->flags & EQ_TYPEMASK, rc);
1876 			return (rc);
1877 		}
1878 		eq->flags &= ~EQ_ALLOCATED;
1879 	}
1880 
1881 	if (eq->desc != NULL) {
1882 		(void) free_desc_ring(&eq->desc_dhdl, &eq->desc_ahdl);
1883 		eq->desc = NULL;
1884 	}
1885 
1886 	if (eq->flags & EQ_MTX)
1887 		mutex_destroy(&eq->lock);
1888 
1889 	bzero(eq, sizeof (*eq));
1890 	return (0);
1891 }
1892 
1893 #ifdef TCP_OFFLOAD_ENABLE
1894 /* ARGSUSED */
1895 static int
1896 alloc_wrq(struct adapter *sc, struct port_info *pi, struct sge_wrq *wrq,
1897     int idx)
1898 {
1899 	int rc;
1900 
1901 	rc = alloc_eq(sc, pi, &wrq->eq);
1902 	if (rc != 0)
1903 		return (rc);
1904 
1905 	wrq->adapter = sc;
1906 	wrq->wr_list.head = NULL;
1907 	wrq->wr_list.tail = NULL;
1908 
1909 	/*
1910 	 * TODO: use idx to figure out what kind of wrq this is and install
1911 	 * useful kstats for it.
1912 	 */
1913 
1914 	return (rc);
1915 }
1916 
1917 static int
1918 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
1919 {
1920 	int rc;
1921 
1922 	rc = free_eq(sc, &wrq->eq);
1923 	if (rc != 0)
1924 		return (rc);
1925 
1926 	bzero(wrq, sizeof (*wrq));
1927 	return (0);
1928 }
1929 #endif
1930 
1931 static int
1932 alloc_txq(struct port_info *pi, struct sge_txq *txq, int idx)
1933 {
1934 	int rc, i;
1935 	struct adapter *sc = pi->adapter;
1936 	struct sge_eq *eq = &txq->eq;
1937 
1938 	rc = alloc_eq(sc, pi, eq);
1939 	if (rc != 0)
1940 		return (rc);
1941 
1942 	txq->port = pi;
1943 	txq->sdesc = kmem_zalloc(sizeof (struct tx_sdesc) * eq->cap, KM_SLEEP);
1944 	txq->txb_size = eq->qsize * tx_copy_threshold;
1945 	rc = alloc_tx_copybuffer(sc, txq->txb_size, &txq->txb_dhdl,
1946 	    &txq->txb_ahdl, &txq->txb_ba, &txq->txb_va);
1947 	if (rc == 0)
1948 		txq->txb_avail = txq->txb_size;
1949 	else
1950 		txq->txb_avail = txq->txb_size = 0;
1951 
1952 	/*
1953 	 * TODO: is this too low?  Worst case would need around 4 times qsize
1954 	 * (all tx descriptors filled to the brim with SGLs, with each entry in
1955 	 * the SGL coming from a distinct DMA handle).  Increase tx_dhdl_total
1956 	 * if you see too many dma_hdl_failed.
1957 	 */
1958 	txq->tx_dhdl_total = eq->qsize * 2;
1959 	txq->tx_dhdl = kmem_zalloc(sizeof (ddi_dma_handle_t) *
1960 	    txq->tx_dhdl_total, KM_SLEEP);
1961 	for (i = 0; i < txq->tx_dhdl_total; i++) {
1962 		rc = ddi_dma_alloc_handle(sc->dip, &sc->sge.dma_attr_tx,
1963 		    DDI_DMA_SLEEP, 0, &txq->tx_dhdl[i]);
1964 		if (rc != DDI_SUCCESS) {
1965 			cxgb_printf(sc->dip, CE_WARN,
1966 			    "%s: failed to allocate DMA handle (%d)",
1967 			    __func__, rc);
1968 			return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EINVAL);
1969 		}
1970 		txq->tx_dhdl_avail++;
1971 	}
1972 
1973 	txq->ksp = setup_txq_kstats(pi, txq, idx);
1974 
1975 	return (rc);
1976 }
1977 
1978 static int
1979 free_txq(struct port_info *pi, struct sge_txq *txq)
1980 {
1981 	int i;
1982 	struct adapter *sc = pi->adapter;
1983 	struct sge_eq *eq = &txq->eq;
1984 
1985 	if (txq->ksp != NULL) {
1986 		kstat_delete(txq->ksp);
1987 		txq->ksp = NULL;
1988 	}
1989 
1990 	if (txq->txb_va != NULL) {
1991 		(void) free_desc_ring(&txq->txb_dhdl, &txq->txb_ahdl);
1992 		txq->txb_va = NULL;
1993 	}
1994 
1995 	if (txq->sdesc != NULL) {
1996 		struct tx_sdesc *sd;
1997 		ddi_dma_handle_t hdl;
1998 
1999 		TXQ_LOCK(txq);
2000 		while (eq->cidx != eq->pidx) {
2001 			sd = &txq->sdesc[eq->cidx];
2002 
2003 			for (i = sd->hdls_used; i; i--) {
2004 				hdl = txq->tx_dhdl[txq->tx_dhdl_cidx];
2005 				(void) ddi_dma_unbind_handle(hdl);
2006 				if (++txq->tx_dhdl_cidx == txq->tx_dhdl_total)
2007 					txq->tx_dhdl_cidx = 0;
2008 			}
2009 
2010 			ASSERT(sd->m);
2011 			freemsgchain(sd->m);
2012 
2013 			eq->cidx += sd->desc_used;
2014 			if (eq->cidx >= eq->cap)
2015 				eq->cidx -= eq->cap;
2016 
2017 			txq->txb_avail += txq->txb_used;
2018 		}
2019 		ASSERT(txq->tx_dhdl_cidx == txq->tx_dhdl_pidx);
2020 		ASSERT(txq->txb_avail == txq->txb_size);
2021 		TXQ_UNLOCK(txq);
2022 
2023 		kmem_free(txq->sdesc, sizeof (struct tx_sdesc) * eq->cap);
2024 		txq->sdesc = NULL;
2025 	}
2026 
2027 	if (txq->tx_dhdl != NULL) {
2028 		for (i = 0; i < txq->tx_dhdl_total; i++) {
2029 			if (txq->tx_dhdl[i] != NULL)
2030 				ddi_dma_free_handle(&txq->tx_dhdl[i]);
2031 		}
2032 	}
2033 
2034 	(void) free_eq(sc, &txq->eq);
2035 
2036 	bzero(txq, sizeof (*txq));
2037 	return (0);
2038 }
2039 
2040 /*
2041  * Allocates a block of contiguous memory for DMA.  Can be used to allocate
2042  * memory for descriptor rings or for tx/rx copy buffers.
2043  *
2044  * Caller does not have to clean up anything if this function fails, it cleans
2045  * up after itself.
2046  *
2047  * Caller provides the following:
2048  * len		length of the block of memory to allocate.
2049  * flags	DDI_DMA_* flags to use (CONSISTENT/STREAMING, READ/WRITE/RDWR)
2050  * acc_attr	device access attributes for the allocation.
2051  * dma_attr	DMA attributes for the allocation
2052  *
2053  * If the function is successful it fills up this information:
2054  * dma_hdl	DMA handle for the allocated memory
2055  * acc_hdl	access handle for the allocated memory
2056  * ba		bus address of the allocated memory
2057  * va		KVA of the allocated memory.
2058  */
2059 static int
2060 alloc_dma_memory(struct adapter *sc, size_t len, int flags,
2061     ddi_device_acc_attr_t *acc_attr, ddi_dma_attr_t *dma_attr,
2062     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl,
2063     uint64_t *pba, caddr_t *pva)
2064 {
2065 	int rc;
2066 	ddi_dma_handle_t dhdl;
2067 	ddi_acc_handle_t ahdl;
2068 	ddi_dma_cookie_t cookie;
2069 	uint_t ccount;
2070 	caddr_t va;
2071 	size_t real_len;
2072 
2073 	*pva = NULL;
2074 
2075 	/*
2076 	 * DMA handle.
2077 	 */
2078 	rc = ddi_dma_alloc_handle(sc->dip, dma_attr, DDI_DMA_SLEEP, 0, &dhdl);
2079 	if (rc != DDI_SUCCESS) {
2080 		cxgb_printf(sc->dip, CE_WARN,
2081 		    "failed to allocate DMA handle: %d", rc);
2082 
2083 		return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EINVAL);
2084 	}
2085 
2086 	/*
2087 	 * Memory suitable for DMA.
2088 	 */
2089 	rc = ddi_dma_mem_alloc(dhdl, len, acc_attr,
2090 	    flags & DDI_DMA_CONSISTENT ? DDI_DMA_CONSISTENT : DDI_DMA_STREAMING,
2091 	    DDI_DMA_SLEEP, 0, &va, &real_len, &ahdl);
2092 	if (rc != DDI_SUCCESS) {
2093 		cxgb_printf(sc->dip, CE_WARN,
2094 		    "failed to allocate DMA memory: %d", rc);
2095 
2096 		ddi_dma_free_handle(&dhdl);
2097 		return (ENOMEM);
2098 	}
2099 
2100 	if (len != real_len) {
2101 		cxgb_printf(sc->dip, CE_WARN,
2102 		    "%s: len (%u) != real_len (%u)\n", len, real_len);
2103 	}
2104 
2105 	/*
2106 	 * DMA bindings.
2107 	 */
2108 	rc = ddi_dma_addr_bind_handle(dhdl, NULL, va, real_len, flags, NULL,
2109 	    NULL, &cookie, &ccount);
2110 	if (rc != DDI_DMA_MAPPED) {
2111 		cxgb_printf(sc->dip, CE_WARN,
2112 		    "failed to map DMA memory: %d", rc);
2113 
2114 		ddi_dma_mem_free(&ahdl);
2115 		ddi_dma_free_handle(&dhdl);
2116 		return (ENOMEM);
2117 	}
2118 	if (ccount != 1) {
2119 		cxgb_printf(sc->dip, CE_WARN,
2120 		    "unusable DMA mapping (%d segments)", ccount);
2121 		(void) free_desc_ring(&dhdl, &ahdl);
2122 	}
2123 
2124 	bzero(va, real_len);
2125 	*dma_hdl = dhdl;
2126 	*acc_hdl = ahdl;
2127 	*pba = cookie.dmac_laddress;
2128 	*pva = va;
2129 
2130 	return (0);
2131 }
2132 
2133 static int
2134 free_dma_memory(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl)
2135 {
2136 	(void) ddi_dma_unbind_handle(*dhdl);
2137 	ddi_dma_mem_free(ahdl);
2138 	ddi_dma_free_handle(dhdl);
2139 
2140 	return (0);
2141 }
2142 
2143 static int
2144 alloc_desc_ring(struct adapter *sc, size_t len, int rw,
2145     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl,
2146     uint64_t *pba, caddr_t *pva)
2147 {
2148 	ddi_device_acc_attr_t *acc_attr = &sc->sge.acc_attr_desc;
2149 	ddi_dma_attr_t *dma_attr = &sc->sge.dma_attr_desc;
2150 
2151 	return (alloc_dma_memory(sc, len, DDI_DMA_CONSISTENT | rw, acc_attr,
2152 	    dma_attr, dma_hdl, acc_hdl, pba, pva));
2153 }
2154 
2155 static int
2156 free_desc_ring(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl)
2157 {
2158 	return (free_dma_memory(dhdl, ahdl));
2159 }
2160 
2161 static int
2162 alloc_tx_copybuffer(struct adapter *sc, size_t len,
2163     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl,
2164     uint64_t *pba, caddr_t *pva)
2165 {
2166 	ddi_device_acc_attr_t *acc_attr = &sc->sge.acc_attr_tx;
2167 	ddi_dma_attr_t *dma_attr = &sc->sge.dma_attr_desc; /* NOT dma_attr_tx */
2168 
2169 	return (alloc_dma_memory(sc, len, DDI_DMA_STREAMING | DDI_DMA_WRITE,
2170 	    acc_attr, dma_attr, dma_hdl, acc_hdl, pba, pva));
2171 }
2172 
2173 static inline bool
2174 is_new_response(const struct sge_iq *iq, struct rsp_ctrl **ctrl)
2175 {
2176 	(void) ddi_dma_sync(iq->dhdl, (uintptr_t)iq->cdesc -
2177 	    (uintptr_t)iq->desc, iq->esize, DDI_DMA_SYNC_FORKERNEL);
2178 
2179 	*ctrl = (void *)((uintptr_t)iq->cdesc +
2180 	    (iq->esize - sizeof (struct rsp_ctrl)));
2181 
2182 	return ((((*ctrl)->u.type_gen >> S_RSPD_GEN) == iq->gen));
2183 }
2184 
2185 static inline void
2186 iq_next(struct sge_iq *iq)
2187 {
2188 	iq->cdesc = (void *) ((uintptr_t)iq->cdesc + iq->esize);
2189 	if (++iq->cidx == iq->qsize - 1) {
2190 		iq->cidx = 0;
2191 		iq->gen ^= 1;
2192 		iq->cdesc = iq->desc;
2193 	}
2194 }
2195 
2196 /*
2197  * Fill up the freelist by upto nbufs and maybe ring its doorbell.
2198  *
2199  * Returns non-zero to indicate that it should be added to the list of starving
2200  * freelists.
2201  */
2202 static int
2203 refill_fl(struct adapter *sc, struct sge_fl *fl, int nbufs)
2204 {
2205 	uint64_t *d = &fl->desc[fl->pidx];
2206 	struct fl_sdesc *sd = &fl->sdesc[fl->pidx];
2207 
2208 	FL_LOCK_ASSERT_OWNED(fl);
2209 	ASSERT(nbufs >= 0);
2210 
2211 	if (nbufs > fl->needed)
2212 		nbufs = fl->needed;
2213 
2214 	while (nbufs--) {
2215 		if (sd->rxb != NULL) {
2216 			if (sd->rxb->ref_cnt == 1) {
2217 				/*
2218 				 * Buffer is available for recycling.  Two ways
2219 				 * this can happen:
2220 				 *
2221 				 * a) All the packets DMA'd into it last time
2222 				 *    around were within the rx_copy_threshold
2223 				 *    and no part of the buffer was ever passed
2224 				 *    up (ref_cnt never went over 1).
2225 				 *
2226 				 * b) Packets DMA'd into the buffer were passed
2227 				 *    up but have all been freed by the upper
2228 				 *    layers by now (ref_cnt went over 1 but is
2229 				 *    now back to 1).
2230 				 *
2231 				 * Either way the bus address in the descriptor
2232 				 * ring is already valid.
2233 				 */
2234 				ASSERT(*d == cpu_to_be64(sd->rxb->ba));
2235 				d++;
2236 				goto recycled;
2237 			} else {
2238 				/*
2239 				 * Buffer still in use and we need a
2240 				 * replacement. But first release our reference
2241 				 * on the existing buffer.
2242 				 */
2243 				rxbuf_free(sd->rxb);
2244 			}
2245 		}
2246 
2247 		sd->rxb = rxbuf_alloc(sc->sge.rxbuf_cache, KM_NOSLEEP, 1);
2248 		if (sd->rxb == NULL)
2249 			break;
2250 		*d++ = cpu_to_be64(sd->rxb->ba);
2251 
2252 recycled:	fl->pending++;
2253 		sd++;
2254 		fl->needed--;
2255 		if (++fl->pidx == fl->cap) {
2256 			fl->pidx = 0;
2257 			sd = fl->sdesc;
2258 			d = fl->desc;
2259 		}
2260 	}
2261 
2262 	if (fl->pending >= 8)
2263 		ring_fl_db(sc, fl);
2264 
2265 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
2266 }
2267 
2268 #ifndef TAILQ_FOREACH_SAFE
2269 #define	TAILQ_FOREACH_SAFE(var, head, field, tvar)			\
2270 	for ((var) = TAILQ_FIRST((head));				\
2271 	    (var) && ((tvar) = TAILQ_NEXT((var), field), 1);		\
2272 	    (var) = (tvar))
2273 #endif
2274 
2275 /*
2276  * Attempt to refill all starving freelists.
2277  */
2278 static void
2279 refill_sfl(void *arg)
2280 {
2281 	struct adapter *sc = arg;
2282 	struct sge_fl *fl, *fl_temp;
2283 
2284 	mutex_enter(&sc->sfl_lock);
2285 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
2286 		FL_LOCK(fl);
2287 		(void) refill_fl(sc, fl, 64);
2288 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
2289 			TAILQ_REMOVE(&sc->sfl, fl, link);
2290 			fl->flags &= ~FL_STARVING;
2291 		}
2292 		FL_UNLOCK(fl);
2293 	}
2294 
2295 	if (!TAILQ_EMPTY(&sc->sfl) != 0)
2296 		sc->sfl_timer =  timeout(refill_sfl, sc, drv_usectohz(100000));
2297 	mutex_exit(&sc->sfl_lock);
2298 }
2299 
2300 static void
2301 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
2302 {
2303 	mutex_enter(&sc->sfl_lock);
2304 	FL_LOCK(fl);
2305 	if ((fl->flags & FL_DOOMED) == 0) {
2306 		if (TAILQ_EMPTY(&sc->sfl) != 0) {
2307 			sc->sfl_timer = timeout(refill_sfl, sc,
2308 			    drv_usectohz(100000));
2309 		}
2310 		fl->flags |= FL_STARVING;
2311 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
2312 	}
2313 	FL_UNLOCK(fl);
2314 	mutex_exit(&sc->sfl_lock);
2315 }
2316 
2317 static void
2318 free_fl_bufs(struct sge_fl *fl)
2319 {
2320 	struct fl_sdesc *sd;
2321 	unsigned int i;
2322 
2323 	FL_LOCK_ASSERT_OWNED(fl);
2324 
2325 	for (i = 0; i < fl->cap; i++) {
2326 		sd = &fl->sdesc[i];
2327 
2328 		if (sd->rxb != NULL) {
2329 			rxbuf_free(sd->rxb);
2330 			sd->rxb = NULL;
2331 		}
2332 	}
2333 }
2334 
2335 /*
2336  * Note that fl->cidx and fl->offset are left unchanged in case of failure.
2337  */
2338 static mblk_t *
2339 get_fl_payload(struct adapter *sc, struct sge_fl *fl,
2340 	       uint32_t len_newbuf, int *fl_bufs_used)
2341 {
2342 	struct mblk_pair frame = {0};
2343 	struct rxbuf *rxb;
2344 	mblk_t *m = NULL;
2345 	uint_t nbuf = 0, len, copy, n;
2346 	uint32_t cidx, offset, rcidx, roffset;
2347 
2348 	/*
2349 	 * The SGE won't pack a new frame into the current buffer if the entire
2350 	 * payload doesn't fit in the remaining space.  Move on to the next buf
2351 	 * in that case.
2352 	 */
2353 	rcidx = fl->cidx;
2354 	roffset = fl->offset;
2355 	if (fl->offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
2356 		fl->offset = 0;
2357 		if (++fl->cidx == fl->cap)
2358 			fl->cidx = 0;
2359 		nbuf++;
2360 	}
2361 	cidx = fl->cidx;
2362 	offset = fl->offset;
2363 
2364 	len = G_RSPD_LEN(len_newbuf);	/* pktshift + payload length */
2365 	copy = (len <= fl->copy_threshold);
2366 	if (copy != 0) {
2367 		frame.head = m = allocb(len, BPRI_HI);
2368 		if (m == NULL) {
2369 			fl->allocb_fail++;
2370 			cmn_err(CE_WARN,"%s: mbuf allocation failure "
2371 					"count = %llu", __func__,
2372 					(unsigned long long)fl->allocb_fail);
2373 			fl->cidx = rcidx;
2374 			fl->offset = roffset;
2375 			return (NULL);
2376 		}
2377 	}
2378 
2379 	while (len) {
2380 		rxb = fl->sdesc[cidx].rxb;
2381 		n = min(len, rxb->buf_size - offset);
2382 
2383 		(void) ddi_dma_sync(rxb->dhdl, offset, n,
2384 		    DDI_DMA_SYNC_FORKERNEL);
2385 
2386 		if (copy != 0)
2387 			bcopy(rxb->va + offset, m->b_wptr, n);
2388 		else {
2389 			m = desballoc((unsigned char *)rxb->va + offset, n,
2390 			    BPRI_HI, &rxb->freefunc);
2391 			if (m == NULL) {
2392 				fl->allocb_fail++;
2393 				cmn_err(CE_WARN,
2394 					"%s: mbuf allocation failure "
2395 					"count = %llu", __func__,
2396 					(unsigned long long)fl->allocb_fail);
2397 				if (frame.head)
2398 					freemsgchain(frame.head);
2399 				fl->cidx = rcidx;
2400 				fl->offset = roffset;
2401 				return (NULL);
2402 			}
2403 			atomic_inc_uint(&rxb->ref_cnt);
2404 			if (frame.head != NULL)
2405 				frame.tail->b_cont = m;
2406 			else
2407 				frame.head = m;
2408 			frame.tail = m;
2409 		}
2410 		m->b_wptr += n;
2411 		len -= n;
2412 		offset += roundup(n, sc->sge.fl_align);
2413 		ASSERT(offset <= rxb->buf_size);
2414 		if (offset == rxb->buf_size) {
2415 			offset = 0;
2416 			if (++cidx == fl->cap)
2417 				cidx = 0;
2418 			nbuf++;
2419 		}
2420 	}
2421 
2422 	fl->cidx = cidx;
2423 	fl->offset = offset;
2424 	(*fl_bufs_used) += nbuf;
2425 
2426 	ASSERT(frame.head != NULL);
2427 	return (frame.head);
2428 }
2429 
2430 /*
2431  * We'll do immediate data tx for non-LSO, but only when not coalescing.  We're
2432  * willing to use upto 2 hardware descriptors which means a maximum of 96 bytes
2433  * of immediate data.
2434  */
2435 #define	IMM_LEN ( \
2436 	2 * EQ_ESIZE \
2437 	- sizeof (struct fw_eth_tx_pkt_wr) \
2438 	- sizeof (struct cpl_tx_pkt_core))
2439 
2440 /*
2441  * Returns non-zero on failure, no need to cleanup anything in that case.
2442  *
2443  * Note 1: We always try to pull up the mblk if required and return E2BIG only
2444  * if this fails.
2445  *
2446  * Note 2: We'll also pullup incoming mblk if HW_LSO is set and the first mblk
2447  * does not have the TCP header in it.
2448  */
2449 static int
2450 get_frame_txinfo(struct sge_txq *txq, mblk_t **fp, struct txinfo *txinfo,
2451     int sgl_only)
2452 {
2453 	uint32_t flags = 0, len, n;
2454 	mblk_t *m = *fp;
2455 	int rc;
2456 
2457 	TXQ_LOCK_ASSERT_OWNED(txq);	/* will manipulate txb and dma_hdls */
2458 
2459 	mac_hcksum_get(m, NULL, NULL, NULL, NULL, &flags);
2460 	txinfo->flags = flags;
2461 
2462 	mac_lso_get(m, &txinfo->mss, &flags);
2463 	txinfo->flags |= flags;
2464 
2465 	if (flags & HW_LSO)
2466 		sgl_only = 1;	/* Do not allow immediate data with LSO */
2467 
2468 start:	txinfo->nsegs = 0;
2469 	txinfo->hdls_used = 0;
2470 	txinfo->txb_used = 0;
2471 	txinfo->len = 0;
2472 
2473 	/* total length and a rough estimate of # of segments */
2474 	n = 0;
2475 	for (; m; m = m->b_cont) {
2476 		len = MBLKL(m);
2477 		n += (len / PAGE_SIZE) + 1;
2478 		txinfo->len += len;
2479 	}
2480 	m = *fp;
2481 
2482 	if (n >= TX_SGL_SEGS || (flags & HW_LSO && MBLKL(m) < 50)) {
2483 		txq->pullup_early++;
2484 		m = msgpullup(*fp, -1);
2485 		if (m == NULL) {
2486 			txq->pullup_failed++;
2487 			return (E2BIG);	/* (*fp) left as it was */
2488 		}
2489 		freemsg(*fp);
2490 		*fp = m;
2491 		mac_hcksum_set(m, 0, 0, 0, 0, txinfo->flags);
2492 	}
2493 
2494 	if (txinfo->len <= IMM_LEN && !sgl_only)
2495 		return (0);	/* nsegs = 0 tells caller to use imm. tx */
2496 
2497 	if (txinfo->len <= txq->copy_threshold &&
2498 	    copy_into_txb(txq, m, txinfo->len, txinfo) == 0)
2499 		goto done;
2500 
2501 	for (; m; m = m->b_cont) {
2502 
2503 		len = MBLKL(m);
2504 
2505 		/* Use tx copy buffer if this mblk is small enough */
2506 		if (len <= txq->copy_threshold &&
2507 		    copy_into_txb(txq, m, len, txinfo) == 0)
2508 			continue;
2509 
2510 		/* Add DMA bindings for this mblk to the SGL */
2511 		rc = add_mblk(txq, txinfo, m, len);
2512 
2513 		if (rc == E2BIG ||
2514 		    (txinfo->nsegs == TX_SGL_SEGS && m->b_cont)) {
2515 
2516 			txq->pullup_late++;
2517 			m = msgpullup(*fp, -1);
2518 			if (m != NULL) {
2519 				free_txinfo_resources(txq, txinfo);
2520 				freemsg(*fp);
2521 				*fp = m;
2522 				mac_hcksum_set(m, 0, 0, 0, 0, txinfo->flags);
2523 				goto start;
2524 			}
2525 
2526 			txq->pullup_failed++;
2527 			rc = E2BIG;
2528 		}
2529 
2530 		if (rc != 0) {
2531 			free_txinfo_resources(txq, txinfo);
2532 			return (rc);
2533 		}
2534 	}
2535 
2536 	ASSERT(txinfo->nsegs > 0 && txinfo->nsegs <= TX_SGL_SEGS);
2537 
2538 done:
2539 
2540 	/*
2541 	 * Store the # of flits required to hold this frame's SGL in nflits.  An
2542 	 * SGL has a (ULPTX header + len0, addr0) tuple optionally followed by
2543 	 * multiple (len0 + len1, addr0, addr1) tuples.  If addr1 is not used
2544 	 * then len1 must be set to 0.
2545 	 */
2546 	n = txinfo->nsegs - 1;
2547 	txinfo->nflits = (3 * n) / 2 + (n & 1) + 2;
2548 	if (n & 1)
2549 		txinfo->sgl.sge[n / 2].len[1] = cpu_to_be32(0);
2550 
2551 	txinfo->sgl.cmd_nsge = cpu_to_be32(V_ULPTX_CMD((u32)ULP_TX_SC_DSGL) |
2552 	    V_ULPTX_NSGE(txinfo->nsegs));
2553 
2554 	return (0);
2555 }
2556 
2557 static inline int
2558 fits_in_txb(struct sge_txq *txq, int len, int *waste)
2559 {
2560 	if (txq->txb_avail < len)
2561 		return (0);
2562 
2563 	if (txq->txb_next + len <= txq->txb_size) {
2564 		*waste = 0;
2565 		return (1);
2566 	}
2567 
2568 	*waste = txq->txb_size - txq->txb_next;
2569 
2570 	return (txq->txb_avail - *waste < len ? 0 : 1);
2571 }
2572 
2573 #define	TXB_CHUNK	64
2574 
2575 /*
2576  * Copies the specified # of bytes into txq's tx copy buffer and updates txinfo
2577  * and txq to indicate resources used.  Caller has to make sure that those many
2578  * bytes are available in the mblk chain (b_cont linked).
2579  */
2580 static inline int
2581 copy_into_txb(struct sge_txq *txq, mblk_t *m, int len, struct txinfo *txinfo)
2582 {
2583 	int waste, n;
2584 
2585 	TXQ_LOCK_ASSERT_OWNED(txq);	/* will manipulate txb */
2586 
2587 	if (!fits_in_txb(txq, len, &waste)) {
2588 		txq->txb_full++;
2589 		return (ENOMEM);
2590 	}
2591 
2592 	if (waste != 0) {
2593 		ASSERT((waste & (TXB_CHUNK - 1)) == 0);
2594 		txinfo->txb_used += waste;
2595 		txq->txb_avail -= waste;
2596 		txq->txb_next = 0;
2597 	}
2598 
2599 	for (n = 0; n < len; m = m->b_cont) {
2600 		bcopy(m->b_rptr, txq->txb_va + txq->txb_next + n, MBLKL(m));
2601 		n += MBLKL(m);
2602 	}
2603 
2604 	add_seg(txinfo, txq->txb_ba + txq->txb_next, len);
2605 
2606 	n = roundup(len, TXB_CHUNK);
2607 	txinfo->txb_used += n;
2608 	txq->txb_avail -= n;
2609 	txq->txb_next += n;
2610 	ASSERT(txq->txb_next <= txq->txb_size);
2611 	if (txq->txb_next == txq->txb_size)
2612 		txq->txb_next = 0;
2613 
2614 	return (0);
2615 }
2616 
2617 static inline void
2618 add_seg(struct txinfo *txinfo, uint64_t ba, uint32_t len)
2619 {
2620 	ASSERT(txinfo->nsegs < TX_SGL_SEGS);	/* must have room */
2621 
2622 	if (txinfo->nsegs != 0) {
2623 		int idx = txinfo->nsegs - 1;
2624 		txinfo->sgl.sge[idx / 2].len[idx & 1] = cpu_to_be32(len);
2625 		txinfo->sgl.sge[idx / 2].addr[idx & 1] = cpu_to_be64(ba);
2626 	} else {
2627 		txinfo->sgl.len0 = cpu_to_be32(len);
2628 		txinfo->sgl.addr0 = cpu_to_be64(ba);
2629 	}
2630 	txinfo->nsegs++;
2631 }
2632 
2633 /*
2634  * This function cleans up any partially allocated resources when it fails so
2635  * there's nothing for the caller to clean up in that case.
2636  *
2637  * EIO indicates permanent failure.  Caller should drop the frame containing
2638  * this mblk and continue.
2639  *
2640  * E2BIG indicates that the SGL length for this mblk exceeds the hardware
2641  * limit.  Caller should pull up the frame before trying to send it out.
2642  * (This error means our pullup_early heuristic did not work for this frame)
2643  *
2644  * ENOMEM indicates a temporary shortage of resources (DMA handles, other DMA
2645  * resources, etc.).  Caller should suspend the tx queue and wait for reclaim to
2646  * free up resources.
2647  */
2648 static inline int
2649 add_mblk(struct sge_txq *txq, struct txinfo *txinfo, mblk_t *m, int len)
2650 {
2651 	ddi_dma_handle_t dhdl;
2652 	ddi_dma_cookie_t cookie;
2653 	uint_t ccount = 0;
2654 	int rc;
2655 
2656 	TXQ_LOCK_ASSERT_OWNED(txq);	/* will manipulate dhdls */
2657 
2658 	if (txq->tx_dhdl_avail == 0) {
2659 		txq->dma_hdl_failed++;
2660 		return (ENOMEM);
2661 	}
2662 
2663 	dhdl = txq->tx_dhdl[txq->tx_dhdl_pidx];
2664 	rc = ddi_dma_addr_bind_handle(dhdl, NULL, (caddr_t)m->b_rptr, len,
2665 	    DDI_DMA_WRITE | DDI_DMA_STREAMING, DDI_DMA_DONTWAIT, NULL, &cookie,
2666 	    &ccount);
2667 	if (rc != DDI_DMA_MAPPED) {
2668 		txq->dma_map_failed++;
2669 
2670 		ASSERT(rc != DDI_DMA_INUSE && rc != DDI_DMA_PARTIAL_MAP);
2671 
2672 		return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EIO);
2673 	}
2674 
2675 	if (ccount + txinfo->nsegs > TX_SGL_SEGS) {
2676 		(void) ddi_dma_unbind_handle(dhdl);
2677 		return (E2BIG);
2678 	}
2679 
2680 	add_seg(txinfo, cookie.dmac_laddress, cookie.dmac_size);
2681 	while (--ccount) {
2682 		ddi_dma_nextcookie(dhdl, &cookie);
2683 		add_seg(txinfo, cookie.dmac_laddress, cookie.dmac_size);
2684 	}
2685 
2686 	if (++txq->tx_dhdl_pidx == txq->tx_dhdl_total)
2687 		txq->tx_dhdl_pidx = 0;
2688 	txq->tx_dhdl_avail--;
2689 	txinfo->hdls_used++;
2690 
2691 	return (0);
2692 }
2693 
2694 /*
2695  * Releases all the txq resources used up in the specified txinfo.
2696  */
2697 static void
2698 free_txinfo_resources(struct sge_txq *txq, struct txinfo *txinfo)
2699 {
2700 	int n;
2701 
2702 	TXQ_LOCK_ASSERT_OWNED(txq);	/* dhdls, txb */
2703 
2704 	n = txinfo->txb_used;
2705 	if (n > 0) {
2706 		txq->txb_avail += n;
2707 		if (n <= txq->txb_next)
2708 			txq->txb_next -= n;
2709 		else {
2710 			n -= txq->txb_next;
2711 			txq->txb_next = txq->txb_size - n;
2712 		}
2713 	}
2714 
2715 	for (n = txinfo->hdls_used; n > 0; n--) {
2716 		if (txq->tx_dhdl_pidx > 0)
2717 			txq->tx_dhdl_pidx--;
2718 		else
2719 			txq->tx_dhdl_pidx = txq->tx_dhdl_total - 1;
2720 		txq->tx_dhdl_avail++;
2721 		(void) ddi_dma_unbind_handle(txq->tx_dhdl[txq->tx_dhdl_pidx]);
2722 	}
2723 }
2724 
2725 /*
2726  * Returns 0 to indicate that m has been accepted into a coalesced tx work
2727  * request.  It has either been folded into txpkts or txpkts was flushed and m
2728  * has started a new coalesced work request (as the first frame in a fresh
2729  * txpkts).
2730  *
2731  * Returns non-zero to indicate a failure - caller is responsible for
2732  * transmitting m, if there was anything in txpkts it has been flushed.
2733  */
2734 static int
2735 add_to_txpkts(struct sge_txq *txq, struct txpkts *txpkts, mblk_t *m,
2736     struct txinfo *txinfo)
2737 {
2738 	struct sge_eq *eq = &txq->eq;
2739 	int can_coalesce;
2740 	struct tx_sdesc *txsd;
2741 	uint8_t flits;
2742 
2743 	TXQ_LOCK_ASSERT_OWNED(txq);
2744 
2745 	if (txpkts->npkt > 0) {
2746 		flits = TXPKTS_PKT_HDR + txinfo->nflits;
2747 		can_coalesce = (txinfo->flags & HW_LSO) == 0 &&
2748 		    txpkts->nflits + flits <= TX_WR_FLITS &&
2749 		    txpkts->nflits + flits <= eq->avail * 8 &&
2750 		    txpkts->plen + txinfo->len < 65536;
2751 
2752 		if (can_coalesce != 0) {
2753 			txpkts->tail->b_next = m;
2754 			txpkts->tail = m;
2755 			txpkts->npkt++;
2756 			txpkts->nflits += flits;
2757 			txpkts->plen += txinfo->len;
2758 
2759 			txsd = &txq->sdesc[eq->pidx];
2760 			txsd->txb_used += txinfo->txb_used;
2761 			txsd->hdls_used += txinfo->hdls_used;
2762 
2763 			return (0);
2764 		}
2765 
2766 		/*
2767 		 * Couldn't coalesce m into txpkts.  The first order of business
2768 		 * is to send txpkts on its way.  Then we'll revisit m.
2769 		 */
2770 		write_txpkts_wr(txq, txpkts);
2771 	}
2772 
2773 	/*
2774 	 * Check if we can start a new coalesced tx work request with m as
2775 	 * the first packet in it.
2776 	 */
2777 
2778 	ASSERT(txpkts->npkt == 0);
2779 	ASSERT(txinfo->len < 65536);
2780 
2781 	flits = TXPKTS_WR_HDR + txinfo->nflits;
2782 	can_coalesce = (txinfo->flags & HW_LSO) == 0 &&
2783 	    flits <= eq->avail * 8 && flits <= TX_WR_FLITS;
2784 
2785 	if (can_coalesce == 0)
2786 		return (EINVAL);
2787 
2788 	/*
2789 	 * Start a fresh coalesced tx WR with m as the first frame in it.
2790 	 */
2791 	txpkts->tail = m;
2792 	txpkts->npkt = 1;
2793 	txpkts->nflits = flits;
2794 	txpkts->flitp = &eq->desc[eq->pidx].flit[2];
2795 	txpkts->plen = txinfo->len;
2796 
2797 	txsd = &txq->sdesc[eq->pidx];
2798 	txsd->m = m;
2799 	txsd->txb_used = txinfo->txb_used;
2800 	txsd->hdls_used = txinfo->hdls_used;
2801 
2802 	return (0);
2803 }
2804 
2805 /*
2806  * Note that write_txpkts_wr can never run out of hardware descriptors (but
2807  * write_txpkt_wr can).  add_to_txpkts ensures that a frame is accepted for
2808  * coalescing only if sufficient hardware descriptors are available.
2809  */
2810 static void
2811 write_txpkts_wr(struct sge_txq *txq, struct txpkts *txpkts)
2812 {
2813 	struct sge_eq *eq = &txq->eq;
2814 	struct fw_eth_tx_pkts_wr *wr;
2815 	struct tx_sdesc *txsd;
2816 	uint32_t ctrl;
2817 	uint16_t ndesc;
2818 
2819 	TXQ_LOCK_ASSERT_OWNED(txq);	/* pidx, avail */
2820 
2821 	ndesc = howmany(txpkts->nflits, 8);
2822 
2823 	wr = (void *)&eq->desc[eq->pidx];
2824 	wr->op_pkd = cpu_to_be32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR) |
2825 	    V_FW_WR_IMMDLEN(0)); /* immdlen does not matter in this WR */
2826 	ctrl = V_FW_WR_LEN16(howmany(txpkts->nflits, 2));
2827 	if (eq->avail == ndesc)
2828 		ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ;
2829 	wr->equiq_to_len16 = cpu_to_be32(ctrl);
2830 	wr->plen = cpu_to_be16(txpkts->plen);
2831 	wr->npkt = txpkts->npkt;
2832 	wr->r3 = wr->type = 0;
2833 
2834 	/* Everything else already written */
2835 
2836 	txsd = &txq->sdesc[eq->pidx];
2837 	txsd->desc_used = ndesc;
2838 
2839 	txq->txb_used += txsd->txb_used / TXB_CHUNK;
2840 	txq->hdl_used += txsd->hdls_used;
2841 
2842 	ASSERT(eq->avail >= ndesc);
2843 
2844 	eq->pending += ndesc;
2845 	eq->avail -= ndesc;
2846 	eq->pidx += ndesc;
2847 	if (eq->pidx >= eq->cap)
2848 		eq->pidx -= eq->cap;
2849 
2850 	txq->txpkts_pkts += txpkts->npkt;
2851 	txq->txpkts_wrs++;
2852 	txpkts->npkt = 0;	/* emptied */
2853 }
2854 
2855 static int
2856 write_txpkt_wr(struct port_info *pi, struct sge_txq *txq, mblk_t *m,
2857     struct txinfo *txinfo)
2858 {
2859 	struct sge_eq *eq = &txq->eq;
2860 	struct fw_eth_tx_pkt_wr *wr;
2861 	struct cpl_tx_pkt_core *cpl;
2862 	uint32_t ctrl;	/* used in many unrelated places */
2863 	uint64_t ctrl1;
2864 	int nflits, ndesc;
2865 	struct tx_sdesc *txsd;
2866 	caddr_t dst;
2867 
2868 	TXQ_LOCK_ASSERT_OWNED(txq);	/* pidx, avail */
2869 
2870 	/*
2871 	 * Do we have enough flits to send this frame out?
2872 	 */
2873 	ctrl = sizeof (struct cpl_tx_pkt_core);
2874 	if (txinfo->flags & HW_LSO) {
2875 		nflits = TXPKT_LSO_WR_HDR;
2876 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
2877 	} else
2878 		nflits = TXPKT_WR_HDR;
2879 	if (txinfo->nsegs > 0)
2880 		nflits += txinfo->nflits;
2881 	else {
2882 		nflits += howmany(txinfo->len, 8);
2883 		ctrl += txinfo->len;
2884 	}
2885 	ndesc = howmany(nflits, 8);
2886 	if (ndesc > eq->avail)
2887 		return (ENOMEM);
2888 
2889 	/* Firmware work request header */
2890 	wr = (void *)&eq->desc[eq->pidx];
2891 	wr->op_immdlen = cpu_to_be32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
2892 	    V_FW_WR_IMMDLEN(ctrl));
2893 	ctrl = V_FW_WR_LEN16(howmany(nflits, 2));
2894 	if (eq->avail == ndesc)
2895 		ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ;
2896 	wr->equiq_to_len16 = cpu_to_be32(ctrl);
2897 	wr->r3 = 0;
2898 
2899 	if (txinfo->flags & HW_LSO) {
2900 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
2901 		char *p = (void *)m->b_rptr;
2902 		ctrl = V_LSO_OPCODE((u32)CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
2903 		    F_LSO_LAST_SLICE;
2904 
2905 		/* LINTED: E_BAD_PTR_CAST_ALIGN */
2906 		if (((struct ether_header *)p)->ether_type ==
2907 		    htons(ETHERTYPE_VLAN)) {
2908 			ctrl |= V_LSO_ETHHDR_LEN(1);
2909 			p += sizeof (struct ether_vlan_header);
2910 		} else
2911 			p += sizeof (struct ether_header);
2912 
2913 		/* LINTED: E_BAD_PTR_CAST_ALIGN for IPH_HDR_LENGTH() */
2914 		ctrl |= V_LSO_IPHDR_LEN(IPH_HDR_LENGTH(p) / 4);
2915 		/* LINTED: E_BAD_PTR_CAST_ALIGN for IPH_HDR_LENGTH() */
2916 		p += IPH_HDR_LENGTH(p);
2917 		ctrl |= V_LSO_TCPHDR_LEN(TCP_HDR_LENGTH((tcph_t *)p) / 4);
2918 
2919 		lso->lso_ctrl = cpu_to_be32(ctrl);
2920 		lso->ipid_ofst = cpu_to_be16(0);
2921 		lso->mss = cpu_to_be16(txinfo->mss);
2922 		lso->seqno_offset = cpu_to_be32(0);
2923 		if (is_t4(pi->adapter->params.chip))
2924 			lso->len = cpu_to_be32(txinfo->len);
2925 		else
2926 			lso->len = cpu_to_be32(V_LSO_T5_XFER_SIZE(txinfo->len));
2927 
2928 		cpl = (void *)(lso + 1);
2929 
2930 		txq->tso_wrs++;
2931 	} else
2932 		cpl = (void *)(wr + 1);
2933 
2934 	/* Checksum offload */
2935 	ctrl1 = 0;
2936 	if (!(txinfo->flags & HCK_IPV4_HDRCKSUM))
2937 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
2938 	if (!(txinfo->flags & HCK_FULLCKSUM))
2939 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
2940 	if (ctrl1 == 0)
2941 		txq->txcsum++;	/* some hardware assistance provided */
2942 
2943 	/* CPL header */
2944 	cpl->ctrl0 = cpu_to_be32(V_TXPKT_OPCODE(CPL_TX_PKT) |
2945 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf));
2946 	cpl->pack = 0;
2947 	cpl->len = cpu_to_be16(txinfo->len);
2948 	cpl->ctrl1 = cpu_to_be64(ctrl1);
2949 
2950 	/* Software descriptor */
2951 	txsd = &txq->sdesc[eq->pidx];
2952 	txsd->m = m;
2953 	txsd->txb_used = txinfo->txb_used;
2954 	txsd->hdls_used = txinfo->hdls_used;
2955 	/* LINTED: E_ASSIGN_NARROW_CONV */
2956 	txsd->desc_used = ndesc;
2957 
2958 	txq->txb_used += txinfo->txb_used / TXB_CHUNK;
2959 	txq->hdl_used += txinfo->hdls_used;
2960 
2961 	eq->pending += ndesc;
2962 	eq->avail -= ndesc;
2963 	eq->pidx += ndesc;
2964 	if (eq->pidx >= eq->cap)
2965 		eq->pidx -= eq->cap;
2966 
2967 	/* SGL */
2968 	dst = (void *)(cpl + 1);
2969 	if (txinfo->nsegs > 0) {
2970 		txq->sgl_wrs++;
2971 		copy_to_txd(eq, (void *)&txinfo->sgl, &dst, txinfo->nflits * 8);
2972 
2973 		/* Need to zero-pad to a 16 byte boundary if not on one */
2974 		if ((uintptr_t)dst & 0xf)
2975 			/* LINTED: E_BAD_PTR_CAST_ALIGN */
2976 			*(uint64_t *)dst = 0;
2977 
2978 	} else {
2979 		txq->imm_wrs++;
2980 #ifdef DEBUG
2981 		ctrl = txinfo->len;
2982 #endif
2983 		for (; m; m = m->b_cont) {
2984 			copy_to_txd(eq, (void *)m->b_rptr, &dst, MBLKL(m));
2985 #ifdef DEBUG
2986 			ctrl -= MBLKL(m);
2987 #endif
2988 		}
2989 		ASSERT(ctrl == 0);
2990 	}
2991 
2992 	txq->txpkt_wrs++;
2993 	return (0);
2994 }
2995 
2996 static inline void
2997 write_ulp_cpl_sgl(struct port_info *pi, struct sge_txq *txq,
2998     struct txpkts *txpkts, struct txinfo *txinfo)
2999 {
3000 	struct ulp_txpkt *ulpmc;
3001 	struct ulptx_idata *ulpsc;
3002 	struct cpl_tx_pkt_core *cpl;
3003 	uintptr_t flitp, start, end;
3004 	uint64_t ctrl;
3005 	caddr_t dst;
3006 
3007 	ASSERT(txpkts->npkt > 0);
3008 
3009 	start = (uintptr_t)txq->eq.desc;
3010 	end = (uintptr_t)txq->eq.spg;
3011 
3012 	/* Checksum offload */
3013 	ctrl = 0;
3014 	if (!(txinfo->flags & HCK_IPV4_HDRCKSUM))
3015 		ctrl |= F_TXPKT_IPCSUM_DIS;
3016 	if (!(txinfo->flags & HCK_FULLCKSUM))
3017 		ctrl |= F_TXPKT_L4CSUM_DIS;
3018 	if (ctrl == 0)
3019 		txq->txcsum++;	/* some hardware assistance provided */
3020 
3021 	/*
3022 	 * The previous packet's SGL must have ended at a 16 byte boundary (this
3023 	 * is required by the firmware/hardware).  It follows that flitp cannot
3024 	 * wrap around between the ULPTX master command and ULPTX subcommand (8
3025 	 * bytes each), and that it can not wrap around in the middle of the
3026 	 * cpl_tx_pkt_core either.
3027 	 */
3028 	flitp = (uintptr_t)txpkts->flitp;
3029 	ASSERT((flitp & 0xf) == 0);
3030 
3031 	/* ULP master command */
3032 	ulpmc = (void *)flitp;
3033 	ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0));
3034 	ulpmc->len = htonl(howmany(sizeof (*ulpmc) + sizeof (*ulpsc) +
3035 	    sizeof (*cpl) + 8 * txinfo->nflits, 16));
3036 
3037 	/* ULP subcommand */
3038 	ulpsc = (void *)(ulpmc + 1);
3039 	ulpsc->cmd_more = cpu_to_be32(V_ULPTX_CMD((u32)ULP_TX_SC_IMM) |
3040 	    F_ULP_TX_SC_MORE);
3041 	ulpsc->len = cpu_to_be32(sizeof (struct cpl_tx_pkt_core));
3042 
3043 	flitp += sizeof (*ulpmc) + sizeof (*ulpsc);
3044 	if (flitp == end)
3045 		flitp = start;
3046 
3047 	/* CPL_TX_PKT */
3048 	cpl = (void *)flitp;
3049 	cpl->ctrl0 = cpu_to_be32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3050 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf));
3051 	cpl->pack = 0;
3052 	cpl->len = cpu_to_be16(txinfo->len);
3053 	cpl->ctrl1 = cpu_to_be64(ctrl);
3054 
3055 	flitp += sizeof (*cpl);
3056 	if (flitp == end)
3057 		flitp = start;
3058 
3059 	/* SGL for this frame */
3060 	dst = (caddr_t)flitp;
3061 	copy_to_txd(&txq->eq, (void *)&txinfo->sgl, &dst, txinfo->nflits * 8);
3062 	flitp = (uintptr_t)dst;
3063 
3064 	/* Zero pad and advance to a 16 byte boundary if not already at one. */
3065 	if (flitp & 0xf) {
3066 
3067 		/* no matter what, flitp should be on an 8 byte boundary */
3068 		ASSERT((flitp & 0x7) == 0);
3069 
3070 		*(uint64_t *)flitp = 0;
3071 		flitp += sizeof (uint64_t);
3072 		txpkts->nflits++;
3073 	}
3074 
3075 	if (flitp == end)
3076 		flitp = start;
3077 
3078 	txpkts->flitp = (void *)flitp;
3079 }
3080 
3081 static inline void
3082 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
3083 {
3084 	if ((uintptr_t)(*to) + len <= (uintptr_t)eq->spg) {
3085 		bcopy(from, *to, len);
3086 		(*to) += len;
3087 	} else {
3088 		int portion = (uintptr_t)eq->spg - (uintptr_t)(*to);
3089 
3090 		bcopy(from, *to, portion);
3091 		from += portion;
3092 		portion = len - portion;	/* remaining */
3093 		bcopy(from, (void *)eq->desc, portion);
3094 		(*to) = (caddr_t)eq->desc + portion;
3095 	}
3096 }
3097 
3098 static inline void
3099 ring_tx_db(struct adapter *sc, struct sge_eq *eq)
3100 {
3101 	int val, db_mode;
3102 	u_int db = eq->doorbells;
3103 
3104 	if (eq->pending > 1)
3105 		db &= ~DOORBELL_WCWR;
3106 
3107 	if (eq->pending > eq->pidx) {
3108 		int offset = eq->cap - (eq->pending - eq->pidx);
3109 
3110 		/* pidx has wrapped around since last doorbell */
3111 
3112 		(void) ddi_dma_sync(eq->desc_dhdl,
3113 		    offset * sizeof (struct tx_desc), 0,
3114 		    DDI_DMA_SYNC_FORDEV);
3115 		(void) ddi_dma_sync(eq->desc_dhdl,
3116 		    0, eq->pidx * sizeof (struct tx_desc),
3117 		    DDI_DMA_SYNC_FORDEV);
3118 	} else if (eq->pending > 0) {
3119 		(void) ddi_dma_sync(eq->desc_dhdl,
3120 		    (eq->pidx - eq->pending) * sizeof (struct tx_desc),
3121 		    eq->pending * sizeof (struct tx_desc),
3122 		    DDI_DMA_SYNC_FORDEV);
3123 	}
3124 
3125 	membar_producer();
3126 
3127 	if (is_t4(sc->params.chip))
3128 		val = V_PIDX(eq->pending);
3129 	else
3130 		val = V_PIDX_T5(eq->pending);
3131 
3132 	db_mode = (1 << (ffs(db) - 1));
3133 	switch (db_mode) {
3134 		case DOORBELL_UDB:
3135 			*eq->udb = LE_32(V_QID(eq->udb_qid) | val);
3136 			break;
3137 
3138 		case DOORBELL_WCWR:
3139 			{
3140 				volatile uint64_t *dst, *src;
3141 				int i;
3142 				/*
3143 				 * Queues whose 128B doorbell segment fits in
3144 				 * the page do not use relative qid
3145 				 * (udb_qid is always 0).  Only queues with
3146 				 * doorbell segments can do WCWR.
3147 				 */
3148 				ASSERT(eq->udb_qid == 0 && eq->pending == 1);
3149 
3150 				dst = (volatile void *)((uintptr_t)eq->udb +
3151 				    UDBS_WR_OFFSET - UDBS_DB_OFFSET);
3152 				i = eq->pidx ? eq->pidx - 1 : eq->cap - 1;
3153 				src = (void *)&eq->desc[i];
3154 				while (src != (void *)&eq->desc[i + 1])
3155 				        *dst++ = *src++;
3156 				membar_producer();
3157 				break;
3158 			}
3159 
3160 		case DOORBELL_UDBWC:
3161 			*eq->udb = LE_32(V_QID(eq->udb_qid) | val);
3162 			membar_producer();
3163 			break;
3164 
3165 		case DOORBELL_KDB:
3166 			t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
3167 			    V_QID(eq->cntxt_id) | val);
3168 			break;
3169 	}
3170 
3171 	eq->pending = 0;
3172 }
3173 
3174 static int
3175 reclaim_tx_descs(struct sge_txq *txq, int howmany)
3176 {
3177 	struct tx_sdesc *txsd;
3178 	uint_t cidx, can_reclaim, reclaimed, txb_freed, hdls_freed;
3179 	struct sge_eq *eq = &txq->eq;
3180 
3181 	EQ_LOCK_ASSERT_OWNED(eq);
3182 
3183 	cidx = eq->spg->cidx;	/* stable snapshot */
3184 	cidx = be16_to_cpu(cidx);
3185 
3186 	if (cidx >= eq->cidx)
3187 		can_reclaim = cidx - eq->cidx;
3188 	else
3189 		can_reclaim = cidx + eq->cap - eq->cidx;
3190 
3191 	if (can_reclaim == 0)
3192 		return (0);
3193 
3194 	txb_freed = hdls_freed = reclaimed = 0;
3195 	do {
3196 		int ndesc;
3197 
3198 		txsd = &txq->sdesc[eq->cidx];
3199 		ndesc = txsd->desc_used;
3200 
3201 		/* Firmware doesn't return "partial" credits. */
3202 		ASSERT(can_reclaim >= ndesc);
3203 
3204 		/*
3205 		 * We always keep mblk around, even for immediate data.  If mblk
3206 		 * is NULL, this has to be the software descriptor for a credit
3207 		 * flush work request.
3208 		 */
3209 		if (txsd->m != NULL)
3210 			freemsgchain(txsd->m);
3211 #ifdef DEBUG
3212 		else {
3213 			ASSERT(txsd->txb_used == 0);
3214 			ASSERT(txsd->hdls_used == 0);
3215 			ASSERT(ndesc == 1);
3216 		}
3217 #endif
3218 
3219 		txb_freed += txsd->txb_used;
3220 		hdls_freed += txsd->hdls_used;
3221 		reclaimed += ndesc;
3222 
3223 		eq->cidx += ndesc;
3224 		if (eq->cidx >= eq->cap)
3225 			eq->cidx -= eq->cap;
3226 
3227 		can_reclaim -= ndesc;
3228 
3229 	} while (can_reclaim && reclaimed < howmany);
3230 
3231 	eq->avail += reclaimed;
3232 	ASSERT(eq->avail < eq->cap);	/* avail tops out at (cap - 1) */
3233 
3234 	txq->txb_avail += txb_freed;
3235 
3236 	txq->tx_dhdl_avail += hdls_freed;
3237 	ASSERT(txq->tx_dhdl_avail <= txq->tx_dhdl_total);
3238 	for (; hdls_freed; hdls_freed--) {
3239 		(void) ddi_dma_unbind_handle(txq->tx_dhdl[txq->tx_dhdl_cidx]);
3240 		if (++txq->tx_dhdl_cidx == txq->tx_dhdl_total)
3241 			txq->tx_dhdl_cidx = 0;
3242 	}
3243 
3244 	return (reclaimed);
3245 }
3246 
3247 static void
3248 write_txqflush_wr(struct sge_txq *txq)
3249 {
3250 	struct sge_eq *eq = &txq->eq;
3251 	struct fw_eq_flush_wr *wr;
3252 	struct tx_sdesc *txsd;
3253 
3254 	EQ_LOCK_ASSERT_OWNED(eq);
3255 	ASSERT(eq->avail > 0);
3256 
3257 	wr = (void *)&eq->desc[eq->pidx];
3258 	bzero(wr, sizeof (*wr));
3259 	wr->opcode = FW_EQ_FLUSH_WR;
3260 	wr->equiq_to_len16 = cpu_to_be32(V_FW_WR_LEN16(sizeof (*wr) / 16) |
3261 	    F_FW_WR_EQUEQ | F_FW_WR_EQUIQ);
3262 
3263 	txsd = &txq->sdesc[eq->pidx];
3264 	txsd->m = NULL;
3265 	txsd->txb_used = 0;
3266 	txsd->hdls_used = 0;
3267 	txsd->desc_used = 1;
3268 
3269 	eq->pending++;
3270 	eq->avail--;
3271 	if (++eq->pidx == eq->cap)
3272 		eq->pidx = 0;
3273 }
3274 
3275 static int
3276 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, mblk_t *m)
3277 {
3278 	bool csum_ok;
3279 	uint16_t err_vec;
3280 	struct sge_rxq *rxq = (void *)iq;
3281 	struct mblk_pair chain = {0};
3282 	struct adapter *sc = iq->adapter;
3283 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
3284 
3285 	iq->intr_next = iq->intr_params;
3286 
3287 	m->b_rptr += sc->sge.pktshift;
3288 
3289 	/* Compressed error vector is enabled for T6 only */
3290 	if (sc->params.tp.rx_pkt_encap)
3291 		/* It is enabled only in T6 config file */
3292 		err_vec = G_T6_COMPR_RXERR_VEC(ntohs(cpl->err_vec));
3293 	else
3294 		err_vec = ntohs(cpl->err_vec);
3295 
3296 	csum_ok = cpl->csum_calc && !err_vec;
3297 	/* TODO: what about cpl->ip_frag? */
3298 	if (csum_ok && !cpl->ip_frag) {
3299 		mac_hcksum_set(m, 0, 0, 0, 0xffff,
3300 		    HCK_FULLCKSUM_OK | HCK_FULLCKSUM |
3301 		    HCK_IPV4_HDRCKSUM_OK);
3302 		rxq->rxcsum++;
3303 	}
3304 
3305 	/* Add to the chain that we'll send up */
3306 	if (chain.head != NULL)
3307 		chain.tail->b_next = m;
3308 	else
3309 		chain.head = m;
3310 	chain.tail = m;
3311 
3312 	t4_mac_rx(rxq->port, rxq, chain.head);
3313 
3314 	rxq->rxpkts++;
3315 	rxq->rxbytes  += be16_to_cpu(cpl->len);
3316 	return (0);
3317 }
3318 
3319 #define	FL_HW_IDX(idx)	((idx) >> 3)
3320 
3321 static inline void
3322 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
3323 {
3324 	int desc_start, desc_last, ndesc;
3325 	uint32_t v = sc->params.arch.sge_fl_db ;
3326 
3327 	ndesc = FL_HW_IDX(fl->pending);
3328 
3329 	/* Hold back one credit if pidx = cidx */
3330 	if (FL_HW_IDX(fl->pidx) == FL_HW_IDX(fl->cidx))
3331 		ndesc--;
3332 
3333 	/*
3334 	 * There are chances of ndesc modified above (to avoid pidx = cidx).
3335 	 * If there is nothing to post, return.
3336 	 */
3337 	if (ndesc <= 0)
3338 		return;
3339 
3340 	desc_last = FL_HW_IDX(fl->pidx);
3341 
3342 	if (fl->pidx < fl->pending) {
3343 		/* There was a wrap */
3344 		desc_start = FL_HW_IDX(fl->pidx + fl->cap - fl->pending);
3345 
3346 		/* From desc_start to the end of list */
3347 		(void) ddi_dma_sync(fl->dhdl, desc_start * RX_FL_ESIZE, 0,
3348 		    DDI_DMA_SYNC_FORDEV);
3349 
3350 		/* From start of list to the desc_last */
3351 		if (desc_last != 0)
3352 			(void) ddi_dma_sync(fl->dhdl, 0, desc_last *
3353 			    RX_FL_ESIZE, DDI_DMA_SYNC_FORDEV);
3354 	} else {
3355 		/* There was no wrap, sync from start_desc to last_desc */
3356 		desc_start = FL_HW_IDX(fl->pidx - fl->pending);
3357 		(void) ddi_dma_sync(fl->dhdl, desc_start * RX_FL_ESIZE,
3358 		    ndesc * RX_FL_ESIZE, DDI_DMA_SYNC_FORDEV);
3359 	}
3360 
3361 	if (is_t4(sc->params.chip))
3362 		v |= V_PIDX(ndesc);
3363 	else
3364 		v |= V_PIDX_T5(ndesc);
3365 	v |= V_QID(fl->cntxt_id) | V_PIDX(ndesc);
3366 
3367 	membar_producer();
3368 
3369 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL), v);
3370 
3371 	/*
3372 	 * Update pending count:
3373 	 * Deduct the number of descriptors posted
3374 	 */
3375 	fl->pending -= ndesc * 8;
3376 }
3377 
3378 static void
3379 tx_reclaim_task(void *arg)
3380 {
3381 	struct sge_txq *txq = arg;
3382 
3383 	TXQ_LOCK(txq);
3384 	reclaim_tx_descs(txq, txq->eq.qsize);
3385 	TXQ_UNLOCK(txq);
3386 }
3387 
3388 /* ARGSUSED */
3389 static int
3390 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
3391                 mblk_t *m)
3392 {
3393 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
3394 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
3395 	struct adapter *sc = iq->adapter;
3396 	struct sge *s = &sc->sge;
3397 	struct sge_eq *eq;
3398 	struct sge_txq *txq;
3399 
3400 	txq = (void *)s->eqmap[qid - s->eq_start];
3401 	eq = &txq->eq;
3402 	txq->qflush++;
3403 	t4_mac_tx_update(txq->port, txq);
3404 
3405 	ddi_taskq_dispatch(sc->tq[eq->tx_chan], tx_reclaim_task,
3406 		(void *)txq, DDI_NOSLEEP);
3407 
3408 	return (0);
3409 }
3410 
3411 static int
3412 handle_fw_rpl(struct sge_iq *iq, const struct rss_header *rss, mblk_t *m)
3413 {
3414 	struct adapter *sc = iq->adapter;
3415 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
3416 
3417 	ASSERT(m == NULL);
3418 
3419 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
3420 		const struct rss_header *rss2;
3421 
3422 		rss2 = (const struct rss_header *)&cpl->data[0];
3423 		return (sc->cpl_handler[rss2->opcode](iq, rss2, m));
3424 	}
3425 	return (sc->fw_msg_handler[cpl->type](sc, &cpl->data[0]));
3426 }
3427 
3428 int
3429 t4_alloc_tx_maps(struct adapter *sc, struct tx_maps *txmaps, int count,
3430     int flags)
3431 {
3432 	int i, rc;
3433 
3434 	txmaps->map_total =  count;
3435 	txmaps->map_avail = txmaps->map_cidx = txmaps->map_pidx = 0;
3436 
3437 	txmaps->map =  kmem_zalloc(sizeof (ddi_dma_handle_t) *
3438 	    txmaps->map_total, flags);
3439 
3440 	for (i = 0; i < count; i++) {
3441 		rc = ddi_dma_alloc_handle(sc->dip, &sc->sge.dma_attr_tx,
3442 		    DDI_DMA_SLEEP, 0, &txmaps->map[i]);
3443 		if (rc != DDI_SUCCESS) {
3444 			cxgb_printf(sc->dip, CE_WARN,
3445 			    "%s: failed to allocate DMA handle (%d)",
3446 			    __func__, rc);
3447 			return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EINVAL);
3448 		}
3449 		txmaps->map_avail++;
3450 	}
3451 
3452 	return (0);
3453 }
3454 
3455 #define	KS_UINIT(x)	kstat_named_init(&kstatp->x, #x, KSTAT_DATA_ULONG)
3456 #define	KS_CINIT(x)	kstat_named_init(&kstatp->x, #x, KSTAT_DATA_CHAR)
3457 #define	KS_U_SET(x, y)	kstatp->x.value.ul = (y)
3458 #define	KS_U_FROM(x, y)	kstatp->x.value.ul = (y)->x
3459 #define	KS_C_SET(x, ...)	\
3460 			(void) snprintf(kstatp->x.value.c, 16,  __VA_ARGS__)
3461 
3462 /*
3463  * cxgbe:X:config
3464  */
3465 struct cxgbe_port_config_kstats {
3466 	kstat_named_t idx;
3467 	kstat_named_t nrxq;
3468 	kstat_named_t ntxq;
3469 	kstat_named_t first_rxq;
3470 	kstat_named_t first_txq;
3471 	kstat_named_t controller;
3472 	kstat_named_t factory_mac_address;
3473 };
3474 
3475 /*
3476  * cxgbe:X:info
3477  */
3478 struct cxgbe_port_info_kstats {
3479 	kstat_named_t transceiver;
3480 	kstat_named_t rx_ovflow0;
3481 	kstat_named_t rx_ovflow1;
3482 	kstat_named_t rx_ovflow2;
3483 	kstat_named_t rx_ovflow3;
3484 	kstat_named_t rx_trunc0;
3485 	kstat_named_t rx_trunc1;
3486 	kstat_named_t rx_trunc2;
3487 	kstat_named_t rx_trunc3;
3488 	kstat_named_t tx_pause;
3489 	kstat_named_t rx_pause;
3490 };
3491 
3492 static kstat_t *
3493 setup_port_config_kstats(struct port_info *pi)
3494 {
3495 	kstat_t *ksp;
3496 	struct cxgbe_port_config_kstats *kstatp;
3497 	int ndata;
3498 	dev_info_t *pdip = ddi_get_parent(pi->dip);
3499 	uint8_t *ma = &pi->hw_addr[0];
3500 
3501 	ndata = sizeof (struct cxgbe_port_config_kstats) /
3502 	    sizeof (kstat_named_t);
3503 
3504 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), "config",
3505 	    "net", KSTAT_TYPE_NAMED, ndata, 0);
3506 	if (ksp == NULL) {
3507 		cxgb_printf(pi->dip, CE_WARN, "failed to initialize kstats.");
3508 		return (NULL);
3509 	}
3510 
3511 	kstatp = (struct cxgbe_port_config_kstats *)ksp->ks_data;
3512 
3513 	KS_UINIT(idx);
3514 	KS_UINIT(nrxq);
3515 	KS_UINIT(ntxq);
3516 	KS_UINIT(first_rxq);
3517 	KS_UINIT(first_txq);
3518 	KS_CINIT(controller);
3519 	KS_CINIT(factory_mac_address);
3520 
3521 	KS_U_SET(idx, pi->port_id);
3522 	KS_U_SET(nrxq, pi->nrxq);
3523 	KS_U_SET(ntxq, pi->ntxq);
3524 	KS_U_SET(first_rxq, pi->first_rxq);
3525 	KS_U_SET(first_txq, pi->first_txq);
3526 	KS_C_SET(controller, "%s%d", ddi_driver_name(pdip),
3527 	    ddi_get_instance(pdip));
3528 	KS_C_SET(factory_mac_address, "%02X%02X%02X%02X%02X%02X",
3529 	    ma[0], ma[1], ma[2], ma[3], ma[4], ma[5]);
3530 
3531 	/* Do NOT set ksp->ks_update.  These kstats do not change. */
3532 
3533 	/* Install the kstat */
3534 	ksp->ks_private = (void *)pi;
3535 	kstat_install(ksp);
3536 
3537 	return (ksp);
3538 }
3539 
3540 static kstat_t *
3541 setup_port_info_kstats(struct port_info *pi)
3542 {
3543 	kstat_t *ksp;
3544 	struct cxgbe_port_info_kstats *kstatp;
3545 	int ndata;
3546 
3547 	ndata = sizeof (struct cxgbe_port_info_kstats) / sizeof (kstat_named_t);
3548 
3549 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), "info",
3550 	    "net", KSTAT_TYPE_NAMED, ndata, 0);
3551 	if (ksp == NULL) {
3552 		cxgb_printf(pi->dip, CE_WARN, "failed to initialize kstats.");
3553 		return (NULL);
3554 	}
3555 
3556 	kstatp = (struct cxgbe_port_info_kstats *)ksp->ks_data;
3557 
3558 	KS_CINIT(transceiver);
3559 	KS_UINIT(rx_ovflow0);
3560 	KS_UINIT(rx_ovflow1);
3561 	KS_UINIT(rx_ovflow2);
3562 	KS_UINIT(rx_ovflow3);
3563 	KS_UINIT(rx_trunc0);
3564 	KS_UINIT(rx_trunc1);
3565 	KS_UINIT(rx_trunc2);
3566 	KS_UINIT(rx_trunc3);
3567 	KS_UINIT(tx_pause);
3568 	KS_UINIT(rx_pause);
3569 
3570 	/* Install the kstat */
3571 	ksp->ks_update = update_port_info_kstats;
3572 	ksp->ks_private = (void *)pi;
3573 	kstat_install(ksp);
3574 
3575 	return (ksp);
3576 }
3577 
3578 static int
3579 update_port_info_kstats(kstat_t *ksp, int rw)
3580 {
3581 	struct cxgbe_port_info_kstats *kstatp =
3582 	    (struct cxgbe_port_info_kstats *)ksp->ks_data;
3583 	struct port_info *pi = ksp->ks_private;
3584 	static const char *mod_str[] = { NULL, "LR", "SR", "ER", "TWINAX",
3585 	    "active TWINAX", "LRM" };
3586 	uint32_t bgmap;
3587 
3588 	if (rw == KSTAT_WRITE)
3589 		return (0);
3590 
3591 	if (pi->mod_type == FW_PORT_MOD_TYPE_NONE)
3592 		KS_C_SET(transceiver, "unplugged");
3593 	else if (pi->mod_type == FW_PORT_MOD_TYPE_UNKNOWN)
3594 		KS_C_SET(transceiver, "unknown");
3595 	else if (pi->mod_type == FW_PORT_MOD_TYPE_NOTSUPPORTED)
3596 		KS_C_SET(transceiver, "unsupported");
3597 	else if (pi->mod_type > 0 && pi->mod_type < ARRAY_SIZE(mod_str))
3598 		KS_C_SET(transceiver, "%s", mod_str[pi->mod_type]);
3599 	else
3600 		KS_C_SET(transceiver, "type %d", pi->mod_type);
3601 
3602 #define	GET_STAT(name) t4_read_reg64(pi->adapter, \
3603 	    PORT_REG(pi->port_id, A_MPS_PORT_STAT_##name##_L))
3604 #define	GET_STAT_COM(name) t4_read_reg64(pi->adapter, \
3605 	    A_MPS_STAT_##name##_L)
3606 
3607 	bgmap = G_NUMPORTS(t4_read_reg(pi->adapter, A_MPS_CMN_CTL));
3608 	if (bgmap == 0)
3609 		bgmap = (pi->port_id == 0) ? 0xf : 0;
3610 	else if (bgmap == 1)
3611 		bgmap = (pi->port_id < 2) ? (3 << (2 * pi->port_id)) : 0;
3612 	else
3613 		bgmap = 1;
3614 
3615 	KS_U_SET(rx_ovflow0, (bgmap & 1) ?
3616 	    GET_STAT_COM(RX_BG_0_MAC_DROP_FRAME) : 0);
3617 	KS_U_SET(rx_ovflow1, (bgmap & 2) ?
3618 	    GET_STAT_COM(RX_BG_1_MAC_DROP_FRAME) : 0);
3619 	KS_U_SET(rx_ovflow2, (bgmap & 4) ?
3620 	    GET_STAT_COM(RX_BG_2_MAC_DROP_FRAME) : 0);
3621 	KS_U_SET(rx_ovflow3, (bgmap & 8) ?
3622 	    GET_STAT_COM(RX_BG_3_MAC_DROP_FRAME) : 0);
3623 	KS_U_SET(rx_trunc0,  (bgmap & 1) ?
3624 	    GET_STAT_COM(RX_BG_0_MAC_TRUNC_FRAME) : 0);
3625 	KS_U_SET(rx_trunc1,  (bgmap & 2) ?
3626 	    GET_STAT_COM(RX_BG_1_MAC_TRUNC_FRAME) : 0);
3627 	KS_U_SET(rx_trunc2,  (bgmap & 4) ?
3628 	    GET_STAT_COM(RX_BG_2_MAC_TRUNC_FRAME) : 0);
3629 	KS_U_SET(rx_trunc3,  (bgmap & 8) ?
3630 	    GET_STAT_COM(RX_BG_3_MAC_TRUNC_FRAME) : 0);
3631 
3632 	KS_U_SET(tx_pause, GET_STAT(TX_PORT_PAUSE));
3633 	KS_U_SET(rx_pause, GET_STAT(RX_PORT_PAUSE));
3634 
3635 	return (0);
3636 
3637 }
3638 
3639 /*
3640  * cxgbe:X:rxqY
3641  */
3642 struct rxq_kstats {
3643 	kstat_named_t rxcsum;
3644 	kstat_named_t rxpkts;
3645 	kstat_named_t rxbytes;
3646 	kstat_named_t nomem;
3647 };
3648 
3649 static kstat_t *
3650 setup_rxq_kstats(struct port_info *pi, struct sge_rxq *rxq, int idx)
3651 {
3652 	struct kstat *ksp;
3653 	struct rxq_kstats *kstatp;
3654 	int ndata;
3655 	char str[16];
3656 
3657 	ndata = sizeof (struct rxq_kstats) / sizeof (kstat_named_t);
3658 	(void) snprintf(str, sizeof (str), "rxq%u", idx);
3659 
3660 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), str, "rxq",
3661 	    KSTAT_TYPE_NAMED, ndata, 0);
3662 	if (ksp == NULL) {
3663 		cxgb_printf(pi->dip, CE_WARN,
3664 		    "%s: failed to initialize rxq kstats for queue %d.",
3665 		    __func__, idx);
3666 		return (NULL);
3667 	}
3668 
3669 	kstatp = (struct rxq_kstats *)ksp->ks_data;
3670 
3671 	KS_UINIT(rxcsum);
3672 	KS_UINIT(rxpkts);
3673 	KS_UINIT(rxbytes);
3674 	KS_UINIT(nomem);
3675 
3676 	ksp->ks_update = update_rxq_kstats;
3677 	ksp->ks_private = (void *)rxq;
3678 	kstat_install(ksp);
3679 
3680 	return (ksp);
3681 }
3682 
3683 static int
3684 update_rxq_kstats(kstat_t *ksp, int rw)
3685 {
3686 	struct rxq_kstats *kstatp = (struct rxq_kstats *)ksp->ks_data;
3687 	struct sge_rxq *rxq = ksp->ks_private;
3688 
3689 	if (rw == KSTAT_WRITE)
3690 		return (0);
3691 
3692 	KS_U_FROM(rxcsum, rxq);
3693 	KS_U_FROM(rxpkts, rxq);
3694 	KS_U_FROM(rxbytes, rxq);
3695 	KS_U_FROM(nomem, rxq);
3696 
3697 	return (0);
3698 }
3699 
3700 /*
3701  * cxgbe:X:txqY
3702  */
3703 struct txq_kstats {
3704 	kstat_named_t txcsum;
3705 	kstat_named_t tso_wrs;
3706 	kstat_named_t imm_wrs;
3707 	kstat_named_t sgl_wrs;
3708 	kstat_named_t txpkt_wrs;
3709 	kstat_named_t txpkts_wrs;
3710 	kstat_named_t txpkts_pkts;
3711 	kstat_named_t txb_used;
3712 	kstat_named_t hdl_used;
3713 	kstat_named_t txb_full;
3714 	kstat_named_t dma_hdl_failed;
3715 	kstat_named_t dma_map_failed;
3716 	kstat_named_t qfull;
3717 	kstat_named_t qflush;
3718 	kstat_named_t pullup_early;
3719 	kstat_named_t pullup_late;
3720 	kstat_named_t pullup_failed;
3721 };
3722 
3723 static kstat_t *
3724 setup_txq_kstats(struct port_info *pi, struct sge_txq *txq, int idx)
3725 {
3726 	struct kstat *ksp;
3727 	struct txq_kstats *kstatp;
3728 	int ndata;
3729 	char str[16];
3730 
3731 	ndata = sizeof (struct txq_kstats) / sizeof (kstat_named_t);
3732 	(void) snprintf(str, sizeof (str), "txq%u", idx);
3733 
3734 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), str, "txq",
3735 	    KSTAT_TYPE_NAMED, ndata, 0);
3736 	if (ksp == NULL) {
3737 		cxgb_printf(pi->dip, CE_WARN,
3738 		    "%s: failed to initialize txq kstats for queue %d.",
3739 		    __func__, idx);
3740 		return (NULL);
3741 	}
3742 
3743 	kstatp = (struct txq_kstats *)ksp->ks_data;
3744 
3745 	KS_UINIT(txcsum);
3746 	KS_UINIT(tso_wrs);
3747 	KS_UINIT(imm_wrs);
3748 	KS_UINIT(sgl_wrs);
3749 	KS_UINIT(txpkt_wrs);
3750 	KS_UINIT(txpkts_wrs);
3751 	KS_UINIT(txpkts_pkts);
3752 	KS_UINIT(txb_used);
3753 	KS_UINIT(hdl_used);
3754 	KS_UINIT(txb_full);
3755 	KS_UINIT(dma_hdl_failed);
3756 	KS_UINIT(dma_map_failed);
3757 	KS_UINIT(qfull);
3758 	KS_UINIT(qflush);
3759 	KS_UINIT(pullup_early);
3760 	KS_UINIT(pullup_late);
3761 	KS_UINIT(pullup_failed);
3762 
3763 	ksp->ks_update = update_txq_kstats;
3764 	ksp->ks_private = (void *)txq;
3765 	kstat_install(ksp);
3766 
3767 	return (ksp);
3768 }
3769 
3770 static int
3771 update_txq_kstats(kstat_t *ksp, int rw)
3772 {
3773 	struct txq_kstats *kstatp = (struct txq_kstats *)ksp->ks_data;
3774 	struct sge_txq *txq = ksp->ks_private;
3775 
3776 	if (rw == KSTAT_WRITE)
3777 		return (0);
3778 
3779 	KS_U_FROM(txcsum, txq);
3780 	KS_U_FROM(tso_wrs, txq);
3781 	KS_U_FROM(imm_wrs, txq);
3782 	KS_U_FROM(sgl_wrs, txq);
3783 	KS_U_FROM(txpkt_wrs, txq);
3784 	KS_U_FROM(txpkts_wrs, txq);
3785 	KS_U_FROM(txpkts_pkts, txq);
3786 	KS_U_FROM(txb_used, txq);
3787 	KS_U_FROM(hdl_used, txq);
3788 	KS_U_FROM(txb_full, txq);
3789 	KS_U_FROM(dma_hdl_failed, txq);
3790 	KS_U_FROM(dma_map_failed, txq);
3791 	KS_U_FROM(qfull, txq);
3792 	KS_U_FROM(qflush, txq);
3793 	KS_U_FROM(pullup_early, txq);
3794 	KS_U_FROM(pullup_late, txq);
3795 	KS_U_FROM(pullup_failed, txq);
3796 
3797 	return (0);
3798 }
3799