xref: /illumos-gate/usr/src/uts/common/io/cxgbe/t4nex/t4_sge.c (revision cd277642e0bdc71a7f57c550df1279b0c091d6e2)
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 enum {
47 	SPG_SIZE = 64,		/* Size of status page */
48 	FL_ALIGN = CACHE_LINE,	/* packet buffer alignment in FL buffers */
49 	FL_PKTSHIFT = 2		/* payload is at this offset in packet buffer */
50 };
51 
52 /* Used to track coalesced tx work request */
53 struct txpkts {
54 	mblk_t *tail;		/* head is in the software descriptor */
55 	uint64_t *flitp;	/* ptr to flit where next pkt should start */
56 	uint8_t npkt;		/* # of packets in this work request */
57 	uint8_t nflits;		/* # of flits used by this work request */
58 	uint16_t plen;		/* total payload (sum of all packets) */
59 };
60 
61 /* All information needed to tx a frame */
62 struct txinfo {
63 	uint32_t len;		/* Total length of frame */
64 	uint32_t flags;		/* Checksum and LSO flags */
65 	uint32_t mss;		/* MSS for LSO */
66 	uint8_t nsegs;		/* # of segments in the SGL, 0 means imm. tx */
67 	uint8_t nflits;		/* # of flits needed for the SGL */
68 	uint8_t hdls_used;	/* # of DMA handles used */
69 	uint32_t txb_used;	/* txb_space used */
70 	struct ulptx_sgl sgl __attribute__((aligned(8)));
71 	struct ulptx_sge_pair reserved[TX_SGL_SEGS / 2];
72 };
73 
74 static int service_iq(struct sge_iq *iq, int budget);
75 static inline void init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx,
76     int8_t pktc_idx, int qsize, uint8_t esize);
77 static inline void init_fl(struct sge_fl *fl, uint16_t qsize);
78 static inline void init_eq(struct adapter *sc, struct sge_eq *eq,
79     uint16_t eqtype, uint16_t qsize,uint8_t tx_chan, uint16_t iqid);
80 static int alloc_iq_fl(struct port_info *pi, struct sge_iq *iq,
81     struct sge_fl *fl, int intr_idx, int cong);
82 static int free_iq_fl(struct port_info *pi, struct sge_iq *iq,
83     struct sge_fl *fl);
84 static int alloc_fwq(struct adapter *sc);
85 static int free_fwq(struct adapter *sc);
86 static int alloc_mgmtq(struct adapter *sc);
87 static int alloc_rxq(struct port_info *pi, struct sge_rxq *rxq, int intr_idx,
88     int i);
89 static int free_rxq(struct port_info *pi, struct sge_rxq *rxq);
90 #ifndef TCP_OFFLOAD_DISABLE
91 static int alloc_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq,
92 	int intr_idx);
93 static int free_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq);
94 #endif
95 static int ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq);
96 static int eth_eq_alloc(struct adapter *sc, struct port_info *pi,
97     struct sge_eq *eq);
98 #ifndef TCP_OFFLOAD_DISABLE
99 static int ofld_eq_alloc(struct adapter *sc, struct port_info *pi,
100     struct sge_eq *eq);
101 #endif
102 static int alloc_eq(struct adapter *sc, struct port_info *pi,
103     struct sge_eq *eq);
104 static int free_eq(struct adapter *sc, struct sge_eq *eq);
105 static int alloc_wrq(struct adapter *sc, struct port_info *pi,
106     struct sge_wrq *wrq, int idx);
107 static int free_wrq(struct adapter *sc, struct sge_wrq *wrq);
108 static int alloc_txq(struct port_info *pi, struct sge_txq *txq, int idx);
109 static int free_txq(struct port_info *pi, struct sge_txq *txq);
110 static int alloc_dma_memory(struct adapter *sc, size_t len, int flags,
111     ddi_device_acc_attr_t *acc_attr, ddi_dma_attr_t *dma_attr,
112     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl, uint64_t *pba,
113     caddr_t *pva);
114 static int free_dma_memory(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl);
115 static int alloc_desc_ring(struct adapter *sc, size_t len, int rw,
116     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl, uint64_t *pba,
117     caddr_t *pva);
118 static int free_desc_ring(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl);
119 static int alloc_tx_copybuffer(struct adapter *sc, size_t len,
120     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl, uint64_t *pba,
121     caddr_t *pva);
122 static inline bool is_new_response(const struct sge_iq *iq,
123     struct rsp_ctrl **ctrl);
124 static inline void iq_next(struct sge_iq *iq);
125 static int refill_fl(struct adapter *sc, struct sge_fl *fl, int nbufs);
126 static void refill_sfl(void *arg);
127 static void add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl);
128 static void free_fl_bufs(struct sge_fl *fl);
129 static mblk_t *get_fl_payload(struct sge_fl *fl, uint32_t len_newbuf,
130     int *fl_bufs_used);
131 static int get_frame_txinfo(struct sge_txq *txq, mblk_t **fp,
132     struct txinfo *txinfo, int sgl_only);
133 static inline int fits_in_txb(struct sge_txq *txq, int len, int *waste);
134 static inline int copy_into_txb(struct sge_txq *txq, mblk_t *m, int len,
135     struct txinfo *txinfo);
136 static inline void add_seg(struct txinfo *txinfo, uint64_t ba, uint32_t len);
137 static inline int add_mblk(struct sge_txq *txq, struct txinfo *txinfo,
138     mblk_t *m, int len);
139 static void free_txinfo_resources(struct sge_txq *txq, struct txinfo *txinfo);
140 static int add_to_txpkts(struct sge_txq *txq, struct txpkts *txpkts, mblk_t *m,
141     struct txinfo *txinfo);
142 static void write_txpkts_wr(struct sge_txq *txq, struct txpkts *txpkts);
143 static int write_txpkt_wr(struct port_info *pi, struct sge_txq *txq, mblk_t *m,
144     struct txinfo *txinfo);
145 static inline void write_ulp_cpl_sgl(struct port_info *pi, struct sge_txq *txq,
146     struct txpkts *txpkts, struct txinfo *txinfo);
147 static inline void copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to,
148     int len);
149 static inline void ring_tx_db(struct adapter *sc, struct sge_eq *eq);
150 static int reclaim_tx_descs(struct sge_txq *txq, int howmany);
151 static void write_txqflush_wr(struct sge_txq *txq);
152 static int t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss,
153     mblk_t *m);
154 static inline void ring_fl_db(struct adapter *sc, struct sge_fl *fl);
155 static kstat_t *setup_port_config_kstats(struct port_info *pi);
156 static kstat_t *setup_port_info_kstats(struct port_info *pi);
157 static kstat_t *setup_rxq_kstats(struct port_info *pi, struct sge_rxq *rxq,
158     int idx);
159 static int update_rxq_kstats(kstat_t *ksp, int rw);
160 static int update_port_info_kstats(kstat_t *ksp, int rw);
161 static kstat_t *setup_txq_kstats(struct port_info *pi, struct sge_txq *txq,
162     int idx);
163 static int update_txq_kstats(kstat_t *ksp, int rw);
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 v;
188 
189 	/*
190 	 * Device access and DMA attributes for descriptor rings
191 	 */
192 	acc_attr = &sc->sge.acc_attr_desc;
193 	acc_attr->devacc_attr_version = DDI_DEVICE_ATTR_V0;
194 	acc_attr->devacc_attr_endian_flags = DDI_NEVERSWAP_ACC;
195 	acc_attr->devacc_attr_dataorder = DDI_STRICTORDER_ACC;
196 
197 	dma_attr = &sc->sge.dma_attr_desc;
198 	dma_attr->dma_attr_version = DMA_ATTR_V0;
199 	dma_attr->dma_attr_addr_lo = 0;
200 	dma_attr->dma_attr_addr_hi = UINT64_MAX;
201 	dma_attr->dma_attr_count_max = UINT64_MAX;
202 	dma_attr->dma_attr_align = 512;
203 	dma_attr->dma_attr_burstsizes = 0xfff;
204 	dma_attr->dma_attr_minxfer = 1;
205 	dma_attr->dma_attr_maxxfer = UINT64_MAX;
206 	dma_attr->dma_attr_seg = UINT64_MAX;
207 	dma_attr->dma_attr_sgllen = 1;
208 	dma_attr->dma_attr_granular = 1;
209 	dma_attr->dma_attr_flags = 0;
210 
211 	/*
212 	 * Device access and DMA attributes for tx buffers
213 	 */
214 	acc_attr = &sc->sge.acc_attr_tx;
215 	acc_attr->devacc_attr_version = DDI_DEVICE_ATTR_V0;
216 	acc_attr->devacc_attr_endian_flags = DDI_NEVERSWAP_ACC;
217 
218 	dma_attr = &sc->sge.dma_attr_tx;
219 	dma_attr->dma_attr_version = DMA_ATTR_V0;
220 	dma_attr->dma_attr_addr_lo = 0;
221 	dma_attr->dma_attr_addr_hi = UINT64_MAX;
222 	dma_attr->dma_attr_count_max = UINT64_MAX;
223 	dma_attr->dma_attr_align = 1;
224 	dma_attr->dma_attr_burstsizes = 0xfff;
225 	dma_attr->dma_attr_minxfer = 1;
226 	dma_attr->dma_attr_maxxfer = UINT64_MAX;
227 	dma_attr->dma_attr_seg = UINT64_MAX;
228 	dma_attr->dma_attr_sgllen = TX_SGL_SEGS;
229 	dma_attr->dma_attr_granular = 1;
230 	dma_attr->dma_attr_flags = 0;
231 
232 	/*
233 	 * Device access and DMA attributes for rx buffers
234 	 */
235 	sc->sge.rxb_params.dip = sc->dip;
236 	sc->sge.rxb_params.buf_size = rx_buf_size;
237 
238 	acc_attr = &sc->sge.rxb_params.acc_attr_rx;
239 	acc_attr->devacc_attr_version = DDI_DEVICE_ATTR_V0;
240 	acc_attr->devacc_attr_endian_flags = DDI_NEVERSWAP_ACC;
241 
242 	dma_attr = &sc->sge.rxb_params.dma_attr_rx;
243 	dma_attr->dma_attr_version = DMA_ATTR_V0;
244 	dma_attr->dma_attr_addr_lo = 0;
245 	dma_attr->dma_attr_addr_hi = UINT64_MAX;
246 	dma_attr->dma_attr_count_max = UINT64_MAX;
247 	/*
248 	 * Low 4 bits of an rx buffer address have a special meaning to the SGE
249 	 * and an rx buf cannot have an address with any of these bits set.
250 	 * FL_ALIGN is >= 32 so we're sure things are ok.
251 	 */
252 	dma_attr->dma_attr_align = FL_ALIGN;
253 	dma_attr->dma_attr_burstsizes = 0xfff;
254 	dma_attr->dma_attr_minxfer = 1;
255 	dma_attr->dma_attr_maxxfer = UINT64_MAX;
256 	dma_attr->dma_attr_seg = UINT64_MAX;
257 	dma_attr->dma_attr_sgllen = 1;
258 	dma_attr->dma_attr_granular = 1;
259 	dma_attr->dma_attr_flags = 0;
260 
261 	sc->sge.rxbuf_cache = rxbuf_cache_create(&sc->sge.rxb_params);
262 
263 	v = t4_read_reg(sc, A_SGE_CONM_CTRL);
264 	sc->sge.fl_starve_threshold = G_EGRTHRESHOLD(v) * 2 + 1;
265 
266 	t4_set_reg_field(sc, A_SGE_CONTROL, V_PKTSHIFT(M_PKTSHIFT) |
267 	    V_INGPADBOUNDARY(M_INGPADBOUNDARY) |
268 	    F_EGRSTATUSPAGESIZE, V_INGPADBOUNDARY(ilog2(FL_ALIGN) - 5) |
269 	    V_PKTSHIFT(FL_PKTSHIFT) |
270 	    F_RXPKTCPLMODE |
271 	    V_EGRSTATUSPAGESIZE(SPG_SIZE == 128));
272 
273 	t4_set_reg_field(sc, A_SGE_HOST_PAGE_SIZE,
274 	    V_HOSTPAGESIZEPF0(M_HOSTPAGESIZEPF0),
275 	    V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10));
276 
277 	t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0, rx_buf_size);
278 
279 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD,
280 	    V_THRESHOLD_0(p->counter_val[0]) |
281 	    V_THRESHOLD_1(p->counter_val[1]) |
282 	    V_THRESHOLD_2(p->counter_val[2]) |
283 	    V_THRESHOLD_3(p->counter_val[3]));
284 
285 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1,
286 	    V_TIMERVALUE0(us_to_core_ticks(sc, p->timer_val[0])) |
287 	    V_TIMERVALUE1(us_to_core_ticks(sc, p->timer_val[1])));
288 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3,
289 	    V_TIMERVALUE2(us_to_core_ticks(sc, p->timer_val[2])) |
290 	    V_TIMERVALUE3(us_to_core_ticks(sc, p->timer_val[3])));
291 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5,
292 	    V_TIMERVALUE4(us_to_core_ticks(sc, p->timer_val[4])) |
293 	    V_TIMERVALUE5(us_to_core_ticks(sc, p->timer_val[5])));
294 
295 	(void) t4_register_cpl_handler(sc, CPL_FW4_MSG, handle_fw_rpl);
296 	(void) t4_register_cpl_handler(sc, CPL_FW6_MSG, handle_fw_rpl);
297 	(void) t4_register_cpl_handler(sc, CPL_RX_PKT, t4_eth_rx);
298 	(void) t4_register_fw_msg_handler(sc, FW6_TYPE_CMD_RPL,
299 		    t4_handle_fw_rpl);
300 }
301 
302 /*
303  * Allocate and initialize the firmware event queue and the forwarded interrupt
304  * queues, if any.  The adapter owns all these queues as they are not associated
305  * with any particular port.
306  *
307  * Returns errno on failure.  Resources allocated up to that point may still be
308  * allocated.  Caller is responsible for cleanup in case this function fails.
309  */
310 int
311 t4_setup_adapter_queues(struct adapter *sc)
312 {
313 	int rc;
314 
315 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
316 
317 	/*
318 	 * Firmware event queue
319 	 */
320 	rc = alloc_fwq(sc);
321 	if (rc != 0)
322 		return (rc);
323 
324 	/*
325 	 * Management queue.  This is just a control queue that uses the fwq as
326 	 * its associated iq.
327 	 */
328 	rc = alloc_mgmtq(sc);
329 
330 	return (rc);
331 }
332 
333 /*
334  * Idempotent
335  */
336 int
337 t4_teardown_adapter_queues(struct adapter *sc)
338 {
339 
340 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
341 
342 	(void) free_fwq(sc);
343 
344 	return (0);
345 }
346 
347 static inline int
348 first_vector(struct port_info *pi)
349 {
350 	struct adapter *sc = pi->adapter;
351 	int rc = T4_EXTRA_INTR, i;
352 
353 	if (sc->intr_count == 1)
354 		return (0);
355 
356 	for_each_port(sc, i) {
357 		struct port_info *p = sc->port[i];
358 
359 		if (i == pi->port_id)
360 			break;
361 
362 #ifndef TCP_OFFLOAD_DISABLE
363 		if (!(sc->flags & INTR_FWD))
364 			rc += p->nrxq + p->nofldrxq;
365 		else
366 			rc += max(p->nrxq, p->nofldrxq);
367 #else
368 		/*
369 		 * Not compiled with offload support and intr_count > 1.  Only
370 		 * NIC queues exist and they'd better be taking direct
371 		 * interrupts.
372 		 */
373 		ASSERT(!(sc->flags & INTR_FWD));
374 		rc += p->nrxq;
375 #endif
376 	}
377 	return (rc);
378 }
379 
380 /*
381  * Given an arbitrary "index," come up with an iq that can be used by other
382  * queues (of this port) for interrupt forwarding, SGE egress updates, etc.
383  * The iq returned is guaranteed to be something that takes direct interrupts.
384  */
385 static struct sge_iq *
386 port_intr_iq(struct port_info *pi, int idx)
387 {
388 	struct adapter *sc = pi->adapter;
389 	struct sge *s = &sc->sge;
390 	struct sge_iq *iq = NULL;
391 
392 	if (sc->intr_count == 1)
393 		return (&sc->sge.fwq);
394 
395 #ifndef TCP_OFFLOAD_DISABLE
396 	if (!(sc->flags & INTR_FWD)) {
397 		idx %= pi->nrxq + pi->nofldrxq;
398 
399 		if (idx >= pi->nrxq) {
400 			idx -= pi->nrxq;
401 			iq = &s->ofld_rxq[pi->first_ofld_rxq + idx].iq;
402 		} else
403 			iq = &s->rxq[pi->first_rxq + idx].iq;
404 
405 	} else {
406 		idx %= max(pi->nrxq, pi->nofldrxq);
407 
408 		if (pi->nrxq >= pi->nofldrxq)
409 			iq = &s->rxq[pi->first_rxq + idx].iq;
410 		else
411 			iq = &s->ofld_rxq[pi->first_ofld_rxq + idx].iq;
412 	}
413 #else
414 	/*
415 	 * Not compiled with offload support and intr_count > 1.  Only NIC
416 	 * queues exist and they'd better be taking direct interrupts.
417 	 */
418 	ASSERT(!(sc->flags & INTR_FWD));
419 
420 	idx %= pi->nrxq;
421 	iq = &s->rxq[pi->first_rxq + idx].iq;
422 #endif
423 
424 	return (iq);
425 }
426 
427 int
428 t4_setup_port_queues(struct port_info *pi)
429 {
430 	int rc = 0, i, intr_idx, j, iqid;
431 	struct sge_rxq *rxq;
432 	struct sge_txq *txq;
433 #ifndef TCP_OFFLOAD_DISABLE
434 	struct sge_wrq *ctrlq;
435 	struct sge_ofld_rxq *ofld_rxq;
436 	struct sge_wrq *ofld_txq;
437 #endif
438 	struct adapter *sc = pi->adapter;
439 	struct driver_properties *p = &sc->props;
440 
441 	pi->ksp_config = setup_port_config_kstats(pi);
442 	pi->ksp_info   = setup_port_info_kstats(pi);
443 
444 	/* Interrupt vector to start from (when using multiple vectors) */
445 	intr_idx = first_vector(pi);
446 
447 	/*
448 	 * First pass over all rx queues (NIC and TOE):
449 	 * a) initialize iq and fl
450 	 * b) allocate queue iff it will take direct interrupts.
451 	 */
452 
453 	for_each_rxq(pi, i, rxq) {
454 
455 		init_iq(&rxq->iq, sc, pi->tmr_idx, pi->pktc_idx, p->qsize_rxq,
456 		    RX_IQ_ESIZE);
457 
458 		init_fl(&rxq->fl, p->qsize_rxq / 8); /* 8 bufs in each entry */
459 
460 #ifndef TCP_OFFLOAD_DISABLE
461 		if ((!(sc->flags & INTR_FWD)) ||
462 		    (sc->intr_count > 1 && pi->nrxq >= pi->nofldrxq))
463 #else
464 		if (!(sc->flags & INTR_FWD))
465 #endif
466 		{
467 			rxq->iq.flags |= IQ_INTR;
468 			rc = alloc_rxq(pi, rxq, intr_idx, i);
469 			if (rc != 0)
470 				goto done;
471 			intr_idx++;
472 		}
473 
474 	}
475 
476 #ifndef TCP_OFFLOAD_DISABLE
477 	for_each_ofld_rxq(pi, i, ofld_rxq) {
478 
479 		init_iq(&ofld_rxq->iq, sc, pi->tmr_idx, pi->pktc_idx,
480 		    p->qsize_rxq, RX_IQ_ESIZE);
481 
482 		init_fl(&ofld_rxq->fl, p->qsize_rxq / 8);
483 
484 		if (!(sc->flags & INTR_FWD) ||
485 		    (sc->intr_count > 1 && pi->nofldrxq > pi->nrxq)) {
486 			ofld_rxq->iq.flags = IQ_INTR;
487 			rc = alloc_ofld_rxq(pi, ofld_rxq, intr_idx);
488 			if (rc != 0)
489 				goto done;
490 
491 			intr_idx++;
492 		}
493 	}
494 #endif
495 
496 	/*
497 	 * Second pass over all rx queues (NIC and TOE).  The queues forwarding
498 	 * their interrupts are allocated now.
499 	 */
500 	j = 0;
501 	for_each_rxq(pi, i, rxq) {
502 		if (rxq->iq.flags & IQ_INTR)
503 			continue;
504 
505 		intr_idx = port_intr_iq(pi, j)->abs_id;
506 
507 		rc = alloc_rxq(pi, rxq, intr_idx, i);
508 		if (rc != 0)
509 			goto done;
510 		j++;
511 	}
512 
513 #ifndef TCP_OFFLOAD_DISABLE
514 	for_each_ofld_rxq(pi, i, ofld_rxq) {
515 		if (ofld_rxq->iq.flags & IQ_INTR)
516 			continue;
517 
518 		intr_idx = port_intr_iq(pi, j)->abs_id;
519 		rc = alloc_ofld_rxq(pi, ofld_rxq, intr_idx);
520 		if (rc != 0)
521 			goto done;
522 		j++;
523 	}
524 #endif
525 	/*
526 	 * Now the tx queues.  Only one pass needed.
527 	 */
528 	j = 0;
529 	for_each_txq(pi, i, txq) {
530 		uint16_t iqid;
531 
532 		iqid = port_intr_iq(pi, j)->cntxt_id;
533 		init_eq(sc, &txq->eq, EQ_ETH, p->qsize_txq, pi->tx_chan, iqid);
534 		rc = alloc_txq(pi, txq, i);
535 		if (rc != 0)
536 			goto done;
537 	}
538 
539 #ifndef TCP_OFFLOAD_DISABLE
540 	for_each_ofld_txq(pi, i, ofld_txq) {
541 		uint16_t iqid;
542 
543 		iqid = port_intr_iq(pi, j)->cntxt_id;
544 		init_eq(sc, &ofld_txq->eq, EQ_OFLD, p->qsize_txq, pi->tx_chan,
545 		    iqid);
546 		rc = alloc_wrq(sc, pi, ofld_txq, i);
547 		if (rc != 0)
548 			goto done;
549 	}
550 #endif
551 
552 	/*
553 	 * Finally, the control queue.
554 	 */
555 	ctrlq = &sc->sge.ctrlq[pi->port_id];
556 	iqid = port_intr_iq(pi, 0)->cntxt_id;
557 	init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, pi->tx_chan, iqid);
558 	rc = alloc_wrq(sc, pi, ctrlq, 0);
559 
560 done:
561 	if (rc != 0)
562 		(void) t4_teardown_port_queues(pi);
563 
564 	return (rc);
565 }
566 
567 /*
568  * Idempotent
569  */
570 int
571 t4_teardown_port_queues(struct port_info *pi)
572 {
573 	int i;
574 	struct sge_rxq *rxq;
575 	struct sge_txq *txq;
576 	struct adapter *sc = pi->adapter;
577 #ifndef TCP_OFFLOAD_DISABLE
578 	struct sge_ofld_rxq *ofld_rxq;
579 	struct sge_wrq *ofld_txq;
580 #endif
581 
582 	if (pi->ksp_config != NULL) {
583 		kstat_delete(pi->ksp_config);
584 		pi->ksp_config = NULL;
585 	}
586 	if (pi->ksp_info != NULL) {
587 		kstat_delete(pi->ksp_info);
588 		pi->ksp_info = NULL;
589 	}
590 
591 	(void) free_wrq(sc, &sc->sge.ctrlq[pi->port_id]);
592 
593 	for_each_txq(pi, i, txq) {
594 		(void) free_txq(pi, txq);
595 	}
596 
597 #ifndef TCP_OFFLOAD_DISABLE
598 	for_each_ofld_txq(pi, i, ofld_txq) {
599 		(void) free_wrq(sc, ofld_txq);
600 	}
601 
602 	for_each_ofld_rxq(pi, i, ofld_rxq) {
603 		if ((ofld_rxq->iq.flags & IQ_INTR) == 0)
604 			(void) free_ofld_rxq(pi, ofld_rxq);
605 	}
606 #endif
607 
608 	for_each_rxq(pi, i, rxq) {
609 		if ((rxq->iq.flags & IQ_INTR) == 0)
610 			(void) free_rxq(pi, rxq);
611 	}
612 
613 	/*
614 	 * Then take down the rx queues that take direct interrupts.
615 	 */
616 
617 	for_each_rxq(pi, i, rxq) {
618 		if (rxq->iq.flags & IQ_INTR)
619 			(void) free_rxq(pi, rxq);
620 	}
621 
622 #ifndef TCP_OFFLOAD_DISABLE
623 	for_each_ofld_rxq(pi, i, ofld_rxq) {
624 		if (ofld_rxq->iq.flags & IQ_INTR)
625 			(void) free_ofld_rxq(pi, ofld_rxq);
626 	}
627 #endif
628 
629 	return (0);
630 }
631 
632 /* Deals with errors and forwarded interrupts */
633 uint_t
634 t4_intr_all(caddr_t arg1, caddr_t arg2)
635 {
636 
637 	(void) t4_intr_err(arg1, arg2);
638 	(void) t4_intr(arg1, arg2);
639 
640 	return (DDI_INTR_CLAIMED);
641 }
642 
643 /* Deals with interrupts on the given ingress queue */
644 /* ARGSUSED */
645 uint_t
646 t4_intr(caddr_t arg1, caddr_t arg2)
647 {
648 	/* LINTED: E_BAD_PTR_CAST_ALIGN */
649 	struct sge_iq *iq = (struct sge_iq *)arg2;
650 
651 	if (atomic_cas_uint(&iq->state, IQS_IDLE, IQS_BUSY) == IQS_IDLE) {
652 		(void) service_iq(iq, 0);
653 		(void) atomic_cas_uint(&iq->state, IQS_BUSY, IQS_IDLE);
654 	}
655 	return (DDI_INTR_CLAIMED);
656 }
657 
658 /* Deals with error interrupts */
659 /* ARGSUSED */
660 uint_t
661 t4_intr_err(caddr_t arg1, caddr_t arg2)
662 {
663 	/* LINTED: E_BAD_PTR_CAST_ALIGN */
664 	struct adapter *sc = (struct adapter *)arg1;
665 
666 	t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
667 	(void) t4_slow_intr_handler(sc);
668 
669 	return (DDI_INTR_CLAIMED);
670 }
671 
672 /*
673  * Deals with anything and everything on the given ingress queue.
674  */
675 static int
676 service_iq(struct sge_iq *iq, int budget)
677 {
678 	struct sge_iq *q;
679 	struct sge_rxq *rxq = iq_to_rxq(iq);	/* Use iff iq is part of rxq */
680 	struct sge_fl *fl = &rxq->fl;		/* Use iff IQ_HAS_FL */
681 	struct adapter *sc = iq->adapter;
682 	struct rsp_ctrl *ctrl;
683 	const struct rss_header *rss;
684 	int ndescs = 0, limit, fl_bufs_used = 0;
685 	int rsp_type;
686 	uint32_t lq;
687 	mblk_t *m;
688 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
689 
690 	limit = budget ? budget : iq->qsize / 8;
691 
692 	/*
693 	 * We always come back and check the descriptor ring for new indirect
694 	 * interrupts and other responses after running a single handler.
695 	 */
696 	for (;;) {
697 		while (is_new_response(iq, &ctrl)) {
698 
699 			membar_consumer();
700 
701 			m = NULL;
702 			rsp_type = G_RSPD_TYPE(ctrl->u.type_gen);
703 			lq = be32_to_cpu(ctrl->pldbuflen_qid);
704 			rss = (const void *)iq->cdesc;
705 
706 			switch (rsp_type) {
707 			case X_RSPD_TYPE_FLBUF:
708 
709 				ASSERT(iq->flags & IQ_HAS_FL);
710 
711 				m = get_fl_payload(fl, lq, &fl_bufs_used);
712 				if (m == NULL) {
713 					panic("%s: line %d.", __func__,
714 					    __LINE__);
715 				}
716 
717 			/* FALLTHRU */
718 			case X_RSPD_TYPE_CPL:
719 
720 				ASSERT(rss->opcode < NUM_CPL_CMDS);
721 				sc->cpl_handler[rss->opcode](iq, rss, m);
722 				break;
723 
724 			case X_RSPD_TYPE_INTR:
725 
726 				/*
727 				 * Interrupts should be forwarded only to queues
728 				 * that are not forwarding their interrupts.
729 				 * This means service_iq can recurse but only 1
730 				 * level deep.
731 				 */
732 				ASSERT(budget == 0);
733 
734 				q = sc->sge.iqmap[lq - sc->sge.iq_start];
735 				if (atomic_cas_uint(&q->state, IQS_IDLE,
736 				    IQS_BUSY) == IQS_IDLE) {
737 					if (service_iq(q, q->qsize / 8) == 0) {
738 						(void) atomic_cas_uint(
739 						    &q->state, IQS_BUSY,
740 						    IQS_IDLE);
741 					} else {
742 						STAILQ_INSERT_TAIL(&iql, q,
743 						    link);
744 					}
745 				}
746 				break;
747 
748 			default:
749 				break;
750 			}
751 
752 			iq_next(iq);
753 			if (++ndescs == limit) {
754 				t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
755 				    V_CIDXINC(ndescs) |
756 				    V_INGRESSQID(iq->cntxt_id) |
757 				    V_SEINTARM(V_QINTR_TIMER_IDX(
758 				    X_TIMERREG_UPDATE_CIDX)));
759 				ndescs = 0;
760 
761 				if (fl_bufs_used > 0) {
762 					ASSERT(iq->flags & IQ_HAS_FL);
763 					FL_LOCK(fl);
764 					fl->needed += fl_bufs_used;
765 					(void) refill_fl(sc, fl, fl->cap / 8);
766 					FL_UNLOCK(fl);
767 					fl_bufs_used = 0;
768 				}
769 
770 				if (budget != 0)
771 					return (EINPROGRESS);
772 			}
773 		}
774 
775 		if (STAILQ_EMPTY(&iql) != 0)
776 			break;
777 
778 		/*
779 		 * Process the head only, and send it to the back of the list if
780 		 * it's still not done.
781 		 */
782 		q = STAILQ_FIRST(&iql);
783 		STAILQ_REMOVE_HEAD(&iql, link);
784 		if (service_iq(q, q->qsize / 8) == 0)
785 			(void) atomic_cas_uint(&q->state, IQS_BUSY, IQS_IDLE);
786 		else
787 			STAILQ_INSERT_TAIL(&iql, q, link);
788 	}
789 
790 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_CIDXINC(ndescs) |
791 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_next));
792 
793 	if (iq->flags & IQ_HAS_FL) {
794 		int starved;
795 
796 		FL_LOCK(fl);
797 		fl->needed += fl_bufs_used;
798 		starved = refill_fl(sc, fl, fl->cap / 4);
799 		FL_UNLOCK(fl);
800 		if (starved != 0)
801 			add_fl_to_sfl(sc, fl);
802 	}
803 
804 	return (0);
805 }
806 
807 int
808 t4_mgmt_tx(struct adapter *sc, mblk_t *m)
809 {
810 	return (t4_wrq_tx(sc, &sc->sge.mgmtq, m));
811 }
812 
813 /*
814  * Doesn't fail.  Holds on to work requests it can't send right away.
815  */
816 int
817 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, mblk_t *m0)
818 {
819 	struct sge_eq *eq = &wrq->eq;
820 	struct mblk_pair *wr_list = &wrq->wr_list;
821 	int can_reclaim;
822 	caddr_t dst;
823 	mblk_t *wr, *next;
824 
825 	TXQ_LOCK_ASSERT_OWNED(wrq);
826 #ifndef TCP_OFFLOAD_DISABLE
827 	ASSERT((eq->flags & EQ_TYPEMASK) == EQ_OFLD ||
828 	    (eq->flags & EQ_TYPEMASK) == EQ_CTRL);
829 #else
830 	ASSERT((eq->flags & EQ_TYPEMASK) == EQ_CTRL);
831 #endif
832 
833 	if (m0 != NULL) {
834 		if (wr_list->head != NULL)
835 			wr_list->tail->b_next = m0;
836 		else
837 			wr_list->head = m0;
838 		while (m0->b_next)
839 			m0 = m0->b_next;
840 		wr_list->tail = m0;
841 	}
842 
843 	can_reclaim = reclaimable(eq);
844 	eq->cidx += can_reclaim;
845 	eq->avail += can_reclaim;
846 	if (eq->cidx >= eq->cap)
847 		eq->cidx -= eq->cap;
848 
849 	for (wr = wr_list->head; wr; wr = next) {
850 		int ndesc, len = 0;
851 		mblk_t *m;
852 
853 		next = wr->b_next;
854 		wr->b_next = NULL;
855 
856 		for (m = wr; m; m = m->b_cont)
857 			len += MBLKL(m);
858 
859 		ASSERT(len > 0 && (len & 0x7) == 0);
860 		ASSERT(len <= SGE_MAX_WR_LEN);
861 
862 		ndesc = howmany(len, EQ_ESIZE);
863 		if (eq->avail < ndesc) {
864 			wr->b_next = next;
865 			wrq->no_desc++;
866 			break;
867 		}
868 
869 		dst = (void *)&eq->desc[eq->pidx];
870 		for (m = wr; m; m = m->b_cont)
871 			copy_to_txd(eq, (void *)m->b_rptr, &dst, MBLKL(m));
872 
873 		eq->pidx += ndesc;
874 		eq->avail -= ndesc;
875 		if (eq->pidx >= eq->cap)
876 			eq->pidx -= eq->cap;
877 
878 		eq->pending += ndesc;
879 		if (eq->pending > 16)
880 			ring_tx_db(sc, eq);
881 
882 		wrq->tx_wrs++;
883 		freemsg(wr);
884 
885 		if (eq->avail < 8) {
886 			can_reclaim = reclaimable(eq);
887 			eq->cidx += can_reclaim;
888 			eq->avail += can_reclaim;
889 			if (eq->cidx >= eq->cap)
890 				eq->cidx -= eq->cap;
891 		}
892 	}
893 
894 	if (eq->pending != 0)
895 		ring_tx_db(sc, eq);
896 
897 	if (wr == NULL)
898 		wr_list->head = wr_list->tail = NULL;
899 	else {
900 		wr_list->head = wr;
901 
902 		ASSERT(wr_list->tail->b_next == NULL);
903 	}
904 
905 	return (0);
906 }
907 
908 /* Per-packet header in a coalesced tx WR, before the SGL starts (in flits) */
909 #define	TXPKTS_PKT_HDR ((\
910 	sizeof (struct ulp_txpkt) + \
911 	sizeof (struct ulptx_idata) + \
912 	sizeof (struct cpl_tx_pkt_core)) / 8)
913 
914 /* Header of a coalesced tx WR, before SGL of first packet (in flits) */
915 #define	TXPKTS_WR_HDR (\
916 	sizeof (struct fw_eth_tx_pkts_wr) / 8 + \
917 	TXPKTS_PKT_HDR)
918 
919 /* Header of a tx WR, before SGL of first packet (in flits) */
920 #define	TXPKT_WR_HDR ((\
921 	sizeof (struct fw_eth_tx_pkt_wr) + \
922 	sizeof (struct cpl_tx_pkt_core)) / 8)
923 
924 /* Header of a tx LSO WR, before SGL of first packet (in flits) */
925 #define	TXPKT_LSO_WR_HDR ((\
926 	sizeof (struct fw_eth_tx_pkt_wr) + \
927 	sizeof(struct cpl_tx_pkt_lso_core) + \
928 	sizeof (struct cpl_tx_pkt_core)) / 8)
929 
930 mblk_t *
931 t4_eth_tx(struct port_info *pi, struct sge_txq *txq, mblk_t *frame)
932 {
933 	struct adapter *sc = pi->adapter;
934 	struct sge_eq *eq = &txq->eq;
935 	mblk_t *next_frame;
936 	int rc, coalescing;
937 	struct txpkts txpkts;
938 	struct txinfo txinfo;
939 
940 	txpkts.npkt = 0; /* indicates there's nothing in txpkts */
941 	coalescing = 0;
942 
943 	TXQ_LOCK(txq);
944 	if (eq->avail < 8)
945 		(void) reclaim_tx_descs(txq, 8);
946 	for (; frame; frame = next_frame) {
947 
948 		if (eq->avail < 8)
949 			break;
950 
951 		next_frame = frame->b_next;
952 		frame->b_next = NULL;
953 
954 		if (next_frame != NULL)
955 			coalescing = 1;
956 
957 		rc = get_frame_txinfo(txq, &frame, &txinfo, coalescing);
958 		if (rc != 0) {
959 			if (rc == ENOMEM) {
960 
961 				/* Short of resources, suspend tx */
962 
963 				frame->b_next = next_frame;
964 				break;
965 			}
966 
967 			/*
968 			 * Unrecoverable error for this frame, throw it
969 			 * away and move on to the next.
970 			 */
971 
972 			freemsg(frame);
973 			continue;
974 		}
975 
976 		if (coalescing != 0 &&
977 		    add_to_txpkts(txq, &txpkts, frame, &txinfo) == 0) {
978 
979 			/* Successfully absorbed into txpkts */
980 
981 			write_ulp_cpl_sgl(pi, txq, &txpkts, &txinfo);
982 			goto doorbell;
983 		}
984 
985 		/*
986 		 * We weren't coalescing to begin with, or current frame could
987 		 * not be coalesced (add_to_txpkts flushes txpkts if a frame
988 		 * given to it can't be coalesced).  Either way there should be
989 		 * nothing in txpkts.
990 		 */
991 		ASSERT(txpkts.npkt == 0);
992 
993 		/* We're sending out individual frames now */
994 		coalescing = 0;
995 
996 		if (eq->avail < 8)
997 			(void) reclaim_tx_descs(txq, 8);
998 		rc = write_txpkt_wr(pi, txq, frame, &txinfo);
999 		if (rc != 0) {
1000 
1001 			/* Short of hardware descriptors, suspend tx */
1002 
1003 			/*
1004 			 * This is an unlikely but expensive failure.  We've
1005 			 * done all the hard work (DMA bindings etc.) and now we
1006 			 * can't send out the frame.  What's worse, we have to
1007 			 * spend even more time freeing up everything in txinfo.
1008 			 */
1009 			txq->qfull++;
1010 			free_txinfo_resources(txq, &txinfo);
1011 
1012 			frame->b_next = next_frame;
1013 			break;
1014 		}
1015 
1016 doorbell:
1017 		/* Fewer and fewer doorbells as the queue fills up */
1018 		if (eq->pending >= (1 << (fls(eq->qsize - eq->avail) / 2)))
1019 			ring_tx_db(sc, eq);
1020 		(void) reclaim_tx_descs(txq, 32);
1021 	}
1022 
1023 	if (txpkts.npkt > 0)
1024 		write_txpkts_wr(txq, &txpkts);
1025 
1026 	/*
1027 	 * frame not NULL means there was an error but we haven't thrown it
1028 	 * away.  This can happen when we're short of tx descriptors (qfull) or
1029 	 * maybe even DMA handles (dma_hdl_failed).  Either way, a credit flush
1030 	 * and reclaim will get things going again.
1031 	 *
1032 	 * If eq->avail is already 0 we know a credit flush was requested in the
1033 	 * WR that reduced it to 0 so we don't need another flush (we don't have
1034 	 * any descriptor for a flush WR anyway, duh).
1035 	 */
1036 	if (frame && eq->avail > 0)
1037 		write_txqflush_wr(txq);
1038 
1039 	if (eq->pending != 0)
1040 		ring_tx_db(sc, eq);
1041 
1042 	(void) reclaim_tx_descs(txq, eq->qsize);
1043 	TXQ_UNLOCK(txq);
1044 
1045 	return (frame);
1046 }
1047 
1048 static inline void
1049 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int8_t pktc_idx, int
1050     qsize, uint8_t esize)
1051 {
1052 	ASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS);
1053 	ASSERT(pktc_idx < SGE_NCOUNTERS);	/* -ve is ok, means don't use */
1054 
1055 	iq->flags = 0;
1056 	iq->adapter = sc;
1057 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
1058 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
1059 	if (pktc_idx >= 0) {
1060 		iq->intr_params |= F_QINTR_CNT_EN;
1061 		iq->intr_pktc_idx = pktc_idx;
1062 	}
1063 	iq->qsize = roundup(qsize, 16);		/* See FW_IQ_CMD/iqsize */
1064 	iq->esize = max(esize, 16);		/* See FW_IQ_CMD/iqesize */
1065 }
1066 
1067 static inline void
1068 init_fl(struct sge_fl *fl, uint16_t qsize)
1069 {
1070 
1071 	fl->qsize = qsize;
1072 }
1073 
1074 static inline void
1075 init_eq(struct adapter *sc, struct sge_eq *eq, uint16_t eqtype, uint16_t qsize,
1076     uint8_t tx_chan, uint16_t iqid)
1077 {
1078 	struct sge *s = &sc->sge;
1079 	uint32_t r;
1080 
1081 	ASSERT(tx_chan < NCHAN);
1082 	ASSERT(eqtype <= EQ_TYPEMASK);
1083 
1084 	if (is_t5(sc->params.chip)) {
1085 		r = t4_read_reg(sc, A_SGE_EGRESS_QUEUES_PER_PAGE_PF);
1086 		r >>= S_QUEUESPERPAGEPF0 +
1087 		    (S_QUEUESPERPAGEPF1 - S_QUEUESPERPAGEPF0) * sc->pf;
1088 		s->s_qpp = r & M_QUEUESPERPAGEPF0;
1089 	}
1090 
1091 	eq->flags = eqtype & EQ_TYPEMASK;
1092 	eq->tx_chan = tx_chan;
1093 	eq->iqid = iqid;
1094 	eq->qsize = qsize;
1095 }
1096 
1097 /*
1098  * Allocates the ring for an ingress queue and an optional freelist.  If the
1099  * freelist is specified it will be allocated and then associated with the
1100  * ingress queue.
1101  *
1102  * Returns errno on failure.  Resources allocated up to that point may still be
1103  * allocated.  Caller is responsible for cleanup in case this function fails.
1104  *
1105  * If the ingress queue will take interrupts directly (iq->flags & IQ_INTR) then
1106  * the intr_idx specifies the vector, starting from 0.  Otherwise it specifies
1107  * the index of the queue to which its interrupts will be forwarded.
1108  */
1109 static int
1110 alloc_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl,
1111     int intr_idx, int cong)
1112 {
1113 	int rc, i, cntxt_id;
1114 	size_t len;
1115 	struct fw_iq_cmd c;
1116 	struct adapter *sc = iq->adapter;
1117 	uint32_t v = 0;
1118 
1119 	len = iq->qsize * iq->esize;
1120 	rc = alloc_desc_ring(sc, len, DDI_DMA_READ, &iq->dhdl, &iq->ahdl,
1121 	    &iq->ba, (caddr_t *)&iq->desc);
1122 	if (rc != 0)
1123 		return (rc);
1124 
1125 	bzero(&c, sizeof (c));
1126 	c.op_to_vfn = cpu_to_be32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
1127 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
1128 	    V_FW_IQ_CMD_VFN(0));
1129 
1130 	c.alloc_to_len16 = cpu_to_be32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
1131 	    FW_LEN16(c));
1132 
1133 	/* Special handling for firmware event queue */
1134 	if (iq == &sc->sge.fwq)
1135 		v |= F_FW_IQ_CMD_IQASYNCH;
1136 
1137 	if (iq->flags & IQ_INTR)
1138 		ASSERT(intr_idx < sc->intr_count);
1139 	else
1140 		v |= F_FW_IQ_CMD_IQANDST;
1141 	v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
1142 
1143 	c.type_to_iqandstindex = cpu_to_be32(v |
1144 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
1145 	    V_FW_IQ_CMD_VIID(pi->viid) |
1146 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
1147 	c.iqdroprss_to_iqesize = cpu_to_be16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
1148 	    F_FW_IQ_CMD_IQGTSMODE |
1149 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
1150 	    V_FW_IQ_CMD_IQESIZE(ilog2(iq->esize) - 4));
1151 	c.iqsize = cpu_to_be16(iq->qsize);
1152 	c.iqaddr = cpu_to_be64(iq->ba);
1153 	if (cong >= 0)
1154 		c.iqns_to_fl0congen = BE_32(F_FW_IQ_CMD_IQFLINTCONGEN);
1155 
1156 	if (fl != NULL) {
1157 		mutex_init(&fl->lock, NULL, MUTEX_DRIVER,
1158 		    DDI_INTR_PRI(sc->intr_pri));
1159 		fl->flags |= FL_MTX;
1160 
1161 		len = fl->qsize * RX_FL_ESIZE;
1162 		rc = alloc_desc_ring(sc, len, DDI_DMA_WRITE, &fl->dhdl,
1163 		    &fl->ahdl, &fl->ba, (caddr_t *)&fl->desc);
1164 		if (rc != 0)
1165 			return (rc);
1166 
1167 		/* Allocate space for one software descriptor per buffer. */
1168 		fl->cap = (fl->qsize - SPG_SIZE / RX_FL_ESIZE) * 8;
1169 		fl->sdesc = kmem_zalloc(sizeof (struct fl_sdesc) * fl->cap,
1170 		    KM_SLEEP);
1171 		fl->needed = fl->cap;
1172 		fl->lowat = roundup(sc->sge.fl_starve_threshold, 8);
1173 
1174 		c.iqns_to_fl0congen |=
1175 		    cpu_to_be32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
1176 		    F_FW_IQ_CMD_FL0PACKEN | F_FW_IQ_CMD_FL0PADEN);
1177 		if (cong >= 0) {
1178 			c.iqns_to_fl0congen |=
1179 			    BE_32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
1180 			    F_FW_IQ_CMD_FL0CONGCIF |
1181 			    F_FW_IQ_CMD_FL0CONGEN);
1182 		}
1183 		c.fl0dcaen_to_fl0cidxfthresh =
1184 		    cpu_to_be16(V_FW_IQ_CMD_FL0FBMIN(X_FETCHBURSTMIN_64B) |
1185 		    V_FW_IQ_CMD_FL0FBMAX(X_FETCHBURSTMAX_512B));
1186 		c.fl0size = cpu_to_be16(fl->qsize);
1187 		c.fl0addr = cpu_to_be64(fl->ba);
1188 	}
1189 
1190 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof (c), &c);
1191 	if (rc != 0) {
1192 		cxgb_printf(sc->dip, CE_WARN,
1193 		    "failed to create ingress queue: %d", rc);
1194 		return (rc);
1195 	}
1196 
1197 	iq->cdesc = iq->desc;
1198 	iq->cidx = 0;
1199 	iq->gen = 1;
1200 	iq->intr_next = iq->intr_params;
1201 	iq->adapter = sc;
1202 	iq->cntxt_id = be16_to_cpu(c.iqid);
1203 	iq->abs_id = be16_to_cpu(c.physiqid);
1204 	iq->flags |= IQ_ALLOCATED;
1205 
1206 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
1207 	if (cntxt_id >= sc->sge.niq) {
1208 		panic("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
1209 		    cntxt_id, sc->sge.niq - 1);
1210 	}
1211 	sc->sge.iqmap[cntxt_id] = iq;
1212 
1213 	if (fl != NULL) {
1214 		fl->cntxt_id = be16_to_cpu(c.fl0id);
1215 		fl->pidx = fl->cidx = 0;
1216 		fl->copy_threshold = rx_copy_threshold;
1217 
1218 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
1219 		if (cntxt_id >= sc->sge.neq) {
1220 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
1221 			    __func__, cntxt_id, sc->sge.neq - 1);
1222 		}
1223 		sc->sge.eqmap[cntxt_id] = (void *)fl;
1224 
1225 		FL_LOCK(fl);
1226 		(void) refill_fl(sc, fl, fl->lowat);
1227 		FL_UNLOCK(fl);
1228 
1229 		iq->flags |= IQ_HAS_FL;
1230 	}
1231 
1232 	if (is_t5(sc->params.chip) && cong >= 0) {
1233 		uint32_t param, val;
1234 
1235 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
1236 			V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
1237 			V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
1238 		if (cong == 0)
1239 			val = 1 << 19;
1240 		else {
1241 			val = 2 << 19;
1242 			for (i = 0; i < 4; i++) {
1243 				if (cong & (1 << i))
1244 					val |= 1 << (i << 2);
1245 			}
1246 		}
1247 
1248 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
1249 		if (rc != 0) {
1250 			/* report error but carry on */
1251 			cxgb_printf(sc->dip, CE_WARN,
1252 			    "failed to set congestion manager context for "
1253 			    "ingress queue %d: %d", iq->cntxt_id, rc);
1254 		}
1255 	}
1256 
1257 	/* Enable IQ interrupts */
1258 	iq->state = IQS_IDLE;
1259 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_SEINTARM(iq->intr_params) |
1260 	    V_INGRESSQID(iq->cntxt_id));
1261 
1262 	return (0);
1263 }
1264 
1265 static int
1266 free_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl)
1267 {
1268 	int rc;
1269 	struct adapter *sc = iq->adapter;
1270 	dev_info_t *dip;
1271 
1272 	dip = pi ? pi->dip : sc->dip;
1273 
1274 	if (iq != NULL) {
1275 		if (iq->flags & IQ_ALLOCATED) {
1276 			rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
1277 			    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
1278 			    fl ? fl->cntxt_id : 0xffff, 0xffff);
1279 			if (rc != 0) {
1280 				cxgb_printf(dip, CE_WARN,
1281 				    "failed to free queue %p: %d", iq, rc);
1282 				return (rc);
1283 			}
1284 			iq->flags &= ~IQ_ALLOCATED;
1285 		}
1286 
1287 		if (iq->desc != NULL) {
1288 			(void) free_desc_ring(&iq->dhdl, &iq->ahdl);
1289 			iq->desc = NULL;
1290 		}
1291 
1292 		bzero(iq, sizeof (*iq));
1293 	}
1294 
1295 	if (fl != NULL) {
1296 		if (fl->sdesc != NULL) {
1297 			FL_LOCK(fl);
1298 			free_fl_bufs(fl);
1299 			FL_UNLOCK(fl);
1300 
1301 			kmem_free(fl->sdesc, sizeof (struct fl_sdesc) *
1302 			    fl->cap);
1303 			fl->sdesc = NULL;
1304 		}
1305 
1306 		if (fl->desc != NULL) {
1307 			(void) free_desc_ring(&fl->dhdl, &fl->ahdl);
1308 			fl->desc = NULL;
1309 		}
1310 
1311 		if (fl->flags & FL_MTX) {
1312 			mutex_destroy(&fl->lock);
1313 			fl->flags &= ~FL_MTX;
1314 		}
1315 
1316 		bzero(fl, sizeof (struct sge_fl));
1317 	}
1318 
1319 	return (0);
1320 }
1321 
1322 static int
1323 alloc_fwq(struct adapter *sc)
1324 {
1325 	int rc, intr_idx;
1326 	struct sge_iq *fwq = &sc->sge.fwq;
1327 
1328 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE, FW_IQ_ESIZE);
1329 	fwq->flags |= IQ_INTR;	/* always */
1330 	intr_idx = sc->intr_count > 1 ? 1 : 0;
1331 	rc = alloc_iq_fl(sc->port[0], fwq, NULL, intr_idx, -1);
1332 	if (rc != 0) {
1333 		cxgb_printf(sc->dip, CE_WARN,
1334 		    "failed to create firmware event queue: %d.", rc);
1335 		return (rc);
1336 	}
1337 
1338 	return (0);
1339 }
1340 
1341 static int
1342 free_fwq(struct adapter *sc)
1343 {
1344 
1345 	return (free_iq_fl(NULL, &sc->sge.fwq, NULL));
1346 }
1347 
1348 static int
1349 alloc_mgmtq(struct adapter *sc)
1350 {
1351 	int rc;
1352 	struct sge_wrq *mgmtq = &sc->sge.mgmtq;
1353 
1354 	init_eq(sc, &mgmtq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[0]->tx_chan,
1355 	    sc->sge.fwq.cntxt_id);
1356 	rc = alloc_wrq(sc, NULL, mgmtq, 0);
1357 	if (rc != 0) {
1358 		cxgb_printf(sc->dip, CE_WARN,
1359 		    "failed to create management queue: %d\n", rc);
1360 		return (rc);
1361 	}
1362 
1363 	return (0);
1364 }
1365 
1366 static int
1367 alloc_rxq(struct port_info *pi, struct sge_rxq *rxq, int intr_idx, int i)
1368 {
1369 	int rc;
1370 
1371 	rxq->port = pi;
1372 	rc = alloc_iq_fl(pi, &rxq->iq, &rxq->fl, intr_idx, 1 << pi->tx_chan);
1373 	if (rc != 0)
1374 		return (rc);
1375 
1376 	rxq->ksp = setup_rxq_kstats(pi, rxq, i);
1377 
1378 	return (rc);
1379 }
1380 
1381 static int
1382 free_rxq(struct port_info *pi, struct sge_rxq *rxq)
1383 {
1384 	int rc;
1385 
1386 	if (rxq->ksp != NULL) {
1387 		kstat_delete(rxq->ksp);
1388 		rxq->ksp = NULL;
1389 	}
1390 
1391 	rc = free_iq_fl(pi, &rxq->iq, &rxq->fl);
1392 	if (rc == 0)
1393 		bzero(&rxq->fl, sizeof (*rxq) - offsetof(struct sge_rxq, fl));
1394 
1395 	return (rc);
1396 }
1397 
1398 #ifndef TCP_OFFLOAD_DISABLE
1399 static int
1400 alloc_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq,
1401 	int intr_idx)
1402 {
1403 	int rc;
1404 
1405 	rc = alloc_iq_fl(pi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx,
1406 	    1 << pi->tx_chan);
1407 	if (rc != 0)
1408 		return (rc);
1409 
1410 	return (rc);
1411 }
1412 
1413 static int
1414 free_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq)
1415 {
1416 	int rc;
1417 
1418 	rc = free_iq_fl(pi, &ofld_rxq->iq, &ofld_rxq->fl);
1419 	if (rc == 0)
1420 		bzero(&ofld_rxq->fl, sizeof (*ofld_rxq) -
1421 		    offsetof(struct sge_ofld_rxq, fl));
1422 
1423 	return (rc);
1424 }
1425 #endif
1426 
1427 static int
1428 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
1429 {
1430 	int rc, cntxt_id;
1431 	struct fw_eq_ctrl_cmd c;
1432 
1433 	bzero(&c, sizeof (c));
1434 
1435 	c.op_to_vfn = BE_32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
1436 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
1437 	    V_FW_EQ_CTRL_CMD_VFN(0));
1438 	c.alloc_to_len16 = BE_32(F_FW_EQ_CTRL_CMD_ALLOC |
1439 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
1440 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid)); /* TODO */
1441 	c.physeqid_pkd = BE_32(0);
1442 	c.fetchszm_to_iqid =
1443 	    BE_32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
1444 	    V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
1445 	    F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
1446 	c.dcaen_to_eqsize =
1447 	    BE_32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
1448 	    V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
1449 	    V_FW_EQ_CTRL_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
1450 	    V_FW_EQ_CTRL_CMD_EQSIZE(eq->qsize));
1451 	c.eqaddr = BE_64(eq->ba);
1452 
1453 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof (c), &c);
1454 	if (rc != 0) {
1455 		cxgb_printf(sc->dip, CE_WARN,
1456 		    "failed to create control queue %d: %d", eq->tx_chan, rc);
1457 		return (rc);
1458 	}
1459 	eq->flags |= EQ_ALLOCATED;
1460 
1461 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(BE_32(c.cmpliqid_eqid));
1462 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
1463 	if (cntxt_id >= sc->sge.neq)
1464 		panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
1465 		    cntxt_id, sc->sge.neq - 1);
1466 	sc->sge.eqmap[cntxt_id] = eq;
1467 
1468 	return (rc);
1469 }
1470 
1471 static int
1472 eth_eq_alloc(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
1473 {
1474 	int rc, cntxt_id;
1475 	struct fw_eq_eth_cmd c;
1476 
1477 	bzero(&c, sizeof (c));
1478 
1479 	c.op_to_vfn = BE_32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
1480 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
1481 	    V_FW_EQ_ETH_CMD_VFN(0));
1482 	c.alloc_to_len16 = BE_32(F_FW_EQ_ETH_CMD_ALLOC |
1483 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
1484 	c.viid_pkd = BE_32(V_FW_EQ_ETH_CMD_VIID(pi->viid));
1485 	c.fetchszm_to_iqid =
1486 	    BE_32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
1487 	    V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
1488 	    V_FW_EQ_ETH_CMD_IQID(eq->iqid));
1489 	c.dcaen_to_eqsize = BE_32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
1490 	    V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
1491 	    V_FW_EQ_ETH_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
1492 	    V_FW_EQ_ETH_CMD_EQSIZE(eq->qsize));
1493 	c.eqaddr = BE_64(eq->ba);
1494 
1495 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof (c), &c);
1496 	if (rc != 0) {
1497 		cxgb_printf(pi->dip, CE_WARN,
1498 		    "failed to create Ethernet egress queue: %d", rc);
1499 		return (rc);
1500 	}
1501 	eq->flags |= EQ_ALLOCATED;
1502 
1503 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(BE_32(c.eqid_pkd));
1504 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
1505 	if (cntxt_id >= sc->sge.neq)
1506 		panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
1507 		    cntxt_id, sc->sge.neq - 1);
1508 	sc->sge.eqmap[cntxt_id] = eq;
1509 
1510 	return (rc);
1511 }
1512 
1513 #ifndef TCP_OFFLOAD_DISABLE
1514 static int
1515 ofld_eq_alloc(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
1516 {
1517 	int rc, cntxt_id;
1518 	struct fw_eq_ofld_cmd c;
1519 
1520 	bzero(&c, sizeof (c));
1521 
1522 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
1523 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
1524 	    V_FW_EQ_OFLD_CMD_VFN(0));
1525 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
1526 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
1527 	c.fetchszm_to_iqid =
1528 	    htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
1529 	    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
1530 	    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
1531 	c.dcaen_to_eqsize =
1532 	    BE_32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
1533 	    V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
1534 	    V_FW_EQ_OFLD_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
1535 	    V_FW_EQ_OFLD_CMD_EQSIZE(eq->qsize));
1536 	c.eqaddr = BE_64(eq->ba);
1537 
1538 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof (c), &c);
1539 	if (rc != 0) {
1540 		cxgb_printf(pi->dip, CE_WARN,
1541 		    "failed to create egress queue for TCP offload: %d", rc);
1542 		return (rc);
1543 	}
1544 	eq->flags |= EQ_ALLOCATED;
1545 
1546 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(BE_32(c.eqid_pkd));
1547 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
1548 	if (cntxt_id >= sc->sge.neq)
1549 		panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
1550 		    cntxt_id, sc->sge.neq - 1);
1551 	sc->sge.eqmap[cntxt_id] = eq;
1552 
1553 	return (rc);
1554 }
1555 #endif
1556 
1557 static int
1558 alloc_eq(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
1559 {
1560 	int rc;
1561 	size_t len;
1562 
1563 	mutex_init(&eq->lock, NULL, MUTEX_DRIVER, DDI_INTR_PRI(sc->intr_pri));
1564 	eq->flags |= EQ_MTX;
1565 
1566 	len = eq->qsize * EQ_ESIZE;
1567 	rc = alloc_desc_ring(sc, len, DDI_DMA_WRITE, &eq->desc_dhdl,
1568 	    &eq->desc_ahdl, &eq->ba, (caddr_t *)&eq->desc);
1569 	if (rc != 0)
1570 		return (rc);
1571 
1572 	eq->cap = eq->qsize - SPG_SIZE / EQ_ESIZE;
1573 	eq->spg = (void *)&eq->desc[eq->cap];
1574 	eq->avail = eq->cap - 1;	/* one less to avoid cidx = pidx */
1575 	eq->pidx = eq->cidx = 0;
1576 	eq->doorbells = sc->doorbells;
1577 
1578 	switch (eq->flags & EQ_TYPEMASK) {
1579 	case EQ_CTRL:
1580 		rc = ctrl_eq_alloc(sc, eq);
1581 		break;
1582 
1583 	case EQ_ETH:
1584 		rc = eth_eq_alloc(sc, pi, eq);
1585 		break;
1586 
1587 #ifndef TCP_OFFLOAD_DISABLE
1588 	case EQ_OFLD:
1589 		rc = ofld_eq_alloc(sc, pi, eq);
1590 		break;
1591 #endif
1592 
1593 	default:
1594 		panic("%s: invalid eq type %d.", __func__,
1595 		    eq->flags & EQ_TYPEMASK);
1596 	}
1597 
1598 	if (eq->doorbells &
1599 		(DOORBELL_UDB | DOORBELL_UDBWC | DOORBELL_WCWR)) {
1600 		uint32_t s_qpp = sc->sge.s_qpp;
1601 		uint32_t mask = (1 << s_qpp) - 1;
1602 		volatile uint8_t *udb;
1603 
1604 		udb = (volatile uint8_t *)sc->reg1p + UDBS_DB_OFFSET;
1605 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;   /* pg offset */
1606 		eq->udb_qid = eq->cntxt_id & mask;              /* id in page */
1607 		if (eq->udb_qid > PAGE_SIZE / UDBS_SEG_SIZE)
1608 			eq->doorbells &= ~DOORBELL_WCWR;
1609 		else {
1610 			udb += eq->udb_qid << UDBS_SEG_SHIFT;   /* seg offset */
1611 			eq->udb_qid = 0;
1612 		}
1613 		eq->udb = (volatile void *)udb;
1614 	}
1615 
1616 	if (rc != 0) {
1617 		cxgb_printf(sc->dip, CE_WARN,
1618 		    "failed to allocate egress queue(%d): %d",
1619 		    eq->flags & EQ_TYPEMASK, rc);
1620 	}
1621 
1622 	return (rc);
1623 }
1624 
1625 static int
1626 free_eq(struct adapter *sc, struct sge_eq *eq)
1627 {
1628 	int rc;
1629 
1630 	if (eq->flags & EQ_ALLOCATED) {
1631 		switch (eq->flags & EQ_TYPEMASK) {
1632 		case EQ_CTRL:
1633 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
1634 			    eq->cntxt_id);
1635 			break;
1636 
1637 		case EQ_ETH:
1638 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
1639 			    eq->cntxt_id);
1640 			break;
1641 
1642 #ifndef TCP_OFFLOAD_DISABLE
1643 		case EQ_OFLD:
1644 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
1645 			    eq->cntxt_id);
1646 			break;
1647 #endif
1648 
1649 		default:
1650 			panic("%s: invalid eq type %d.", __func__,
1651 			    eq->flags & EQ_TYPEMASK);
1652 		}
1653 		if (rc != 0) {
1654 			cxgb_printf(sc->dip, CE_WARN,
1655 			    "failed to free egress queue (%d): %d",
1656 			    eq->flags & EQ_TYPEMASK, rc);
1657 			return (rc);
1658 		}
1659 		eq->flags &= ~EQ_ALLOCATED;
1660 	}
1661 
1662 	if (eq->desc != NULL) {
1663 		(void) free_desc_ring(&eq->desc_dhdl, &eq->desc_ahdl);
1664 		eq->desc = NULL;
1665 	}
1666 
1667 	if (eq->flags & EQ_MTX)
1668 		mutex_destroy(&eq->lock);
1669 
1670 	bzero(eq, sizeof (*eq));
1671 	return (0);
1672 }
1673 
1674 /* ARGSUSED */
1675 static int
1676 alloc_wrq(struct adapter *sc, struct port_info *pi, struct sge_wrq *wrq,
1677     int idx)
1678 {
1679 	int rc;
1680 
1681 	rc = alloc_eq(sc, pi, &wrq->eq);
1682 	if (rc != 0)
1683 		return (rc);
1684 
1685 	wrq->adapter = sc;
1686 	wrq->wr_list.head = NULL;
1687 	wrq->wr_list.tail = NULL;
1688 
1689 	/*
1690 	 * TODO: use idx to figure out what kind of wrq this is and install
1691 	 * useful kstats for it.
1692 	 */
1693 
1694 	return (rc);
1695 }
1696 
1697 static int
1698 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
1699 {
1700 	int rc;
1701 
1702 	rc = free_eq(sc, &wrq->eq);
1703 	if (rc != 0)
1704 		return (rc);
1705 
1706 	bzero(wrq, sizeof (*wrq));
1707 	return (0);
1708 }
1709 
1710 static int
1711 alloc_txq(struct port_info *pi, struct sge_txq *txq, int idx)
1712 {
1713 	int rc, i;
1714 	struct adapter *sc = pi->adapter;
1715 	struct sge_eq *eq = &txq->eq;
1716 
1717 	rc = alloc_eq(sc, pi, eq);
1718 	if (rc != 0)
1719 		return (rc);
1720 
1721 	txq->port = pi;
1722 	txq->sdesc = kmem_zalloc(sizeof (struct tx_sdesc) * eq->cap, KM_SLEEP);
1723 	txq->txb_size = eq->qsize * tx_copy_threshold;
1724 	rc = alloc_tx_copybuffer(sc, txq->txb_size, &txq->txb_dhdl,
1725 	    &txq->txb_ahdl, &txq->txb_ba, &txq->txb_va);
1726 	if (rc == 0)
1727 		txq->txb_avail = txq->txb_size;
1728 	else
1729 		txq->txb_avail = txq->txb_size = 0;
1730 
1731 	/*
1732 	 * TODO: is this too low?  Worst case would need around 4 times qsize
1733 	 * (all tx descriptors filled to the brim with SGLs, with each entry in
1734 	 * the SGL coming from a distinct DMA handle).  Increase tx_dhdl_total
1735 	 * if you see too many dma_hdl_failed.
1736 	 */
1737 	txq->tx_dhdl_total = eq->qsize * 2;
1738 	txq->tx_dhdl = kmem_zalloc(sizeof (ddi_dma_handle_t) *
1739 	    txq->tx_dhdl_total, KM_SLEEP);
1740 	for (i = 0; i < txq->tx_dhdl_total; i++) {
1741 		rc = ddi_dma_alloc_handle(sc->dip, &sc->sge.dma_attr_tx,
1742 		    DDI_DMA_SLEEP, 0, &txq->tx_dhdl[i]);
1743 		if (rc != DDI_SUCCESS) {
1744 			cxgb_printf(sc->dip, CE_WARN,
1745 			    "%s: failed to allocate DMA handle (%d)",
1746 			    __func__, rc);
1747 			return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EINVAL);
1748 		}
1749 		txq->tx_dhdl_avail++;
1750 	}
1751 
1752 	txq->ksp = setup_txq_kstats(pi, txq, idx);
1753 
1754 	return (rc);
1755 }
1756 
1757 static int
1758 free_txq(struct port_info *pi, struct sge_txq *txq)
1759 {
1760 	int i;
1761 	struct adapter *sc = pi->adapter;
1762 	struct sge_eq *eq = &txq->eq;
1763 
1764 	if (txq->ksp != NULL) {
1765 		kstat_delete(txq->ksp);
1766 		txq->ksp = NULL;
1767 	}
1768 
1769 	if (txq->txb_va != NULL) {
1770 		(void) free_desc_ring(&txq->txb_dhdl, &txq->txb_ahdl);
1771 		txq->txb_va = NULL;
1772 	}
1773 
1774 	if (txq->sdesc != NULL) {
1775 		struct tx_sdesc *sd;
1776 		ddi_dma_handle_t hdl;
1777 
1778 		TXQ_LOCK(txq);
1779 		while (eq->cidx != eq->pidx) {
1780 			sd = &txq->sdesc[eq->cidx];
1781 
1782 			for (i = sd->hdls_used; i; i--) {
1783 				hdl = txq->tx_dhdl[txq->tx_dhdl_cidx];
1784 				(void) ddi_dma_unbind_handle(hdl);
1785 				if (++txq->tx_dhdl_cidx == txq->tx_dhdl_total)
1786 					txq->tx_dhdl_cidx = 0;
1787 			}
1788 
1789 			ASSERT(sd->m);
1790 			freemsgchain(sd->m);
1791 
1792 			eq->cidx += sd->desc_used;
1793 			if (eq->cidx >= eq->cap)
1794 				eq->cidx -= eq->cap;
1795 
1796 			txq->txb_avail += txq->txb_used;
1797 		}
1798 		ASSERT(txq->tx_dhdl_cidx == txq->tx_dhdl_pidx);
1799 		ASSERT(txq->txb_avail == txq->txb_size);
1800 		TXQ_UNLOCK(txq);
1801 
1802 		kmem_free(txq->sdesc, sizeof (struct tx_sdesc) * eq->cap);
1803 		txq->sdesc = NULL;
1804 	}
1805 
1806 	if (txq->tx_dhdl != NULL) {
1807 		for (i = 0; i < txq->tx_dhdl_total; i++) {
1808 			if (txq->tx_dhdl[i] != NULL)
1809 				ddi_dma_free_handle(&txq->tx_dhdl[i]);
1810 		}
1811 	}
1812 
1813 	(void) free_eq(sc, &txq->eq);
1814 
1815 	bzero(txq, sizeof (*txq));
1816 	return (0);
1817 }
1818 
1819 /*
1820  * Allocates a block of contiguous memory for DMA.  Can be used to allocate
1821  * memory for descriptor rings or for tx/rx copy buffers.
1822  *
1823  * Caller does not have to clean up anything if this function fails, it cleans
1824  * up after itself.
1825  *
1826  * Caller provides the following:
1827  * len		length of the block of memory to allocate.
1828  * flags	DDI_DMA_* flags to use (CONSISTENT/STREAMING, READ/WRITE/RDWR)
1829  * acc_attr	device access attributes for the allocation.
1830  * dma_attr	DMA attributes for the allocation
1831  *
1832  * If the function is successful it fills up this information:
1833  * dma_hdl	DMA handle for the allocated memory
1834  * acc_hdl	access handle for the allocated memory
1835  * ba		bus address of the allocated memory
1836  * va		KVA of the allocated memory.
1837  */
1838 static int
1839 alloc_dma_memory(struct adapter *sc, size_t len, int flags,
1840     ddi_device_acc_attr_t *acc_attr, ddi_dma_attr_t *dma_attr,
1841     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl,
1842     uint64_t *pba, caddr_t *pva)
1843 {
1844 	int rc;
1845 	ddi_dma_handle_t dhdl;
1846 	ddi_acc_handle_t ahdl;
1847 	ddi_dma_cookie_t cookie;
1848 	uint_t ccount;
1849 	caddr_t va;
1850 	size_t real_len;
1851 
1852 	*pva = NULL;
1853 
1854 	/*
1855 	 * DMA handle.
1856 	 */
1857 	rc = ddi_dma_alloc_handle(sc->dip, dma_attr, DDI_DMA_SLEEP, 0, &dhdl);
1858 	if (rc != DDI_SUCCESS) {
1859 		cxgb_printf(sc->dip, CE_WARN,
1860 		    "failed to allocate DMA handle: %d", rc);
1861 
1862 		return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EINVAL);
1863 	}
1864 
1865 	/*
1866 	 * Memory suitable for DMA.
1867 	 */
1868 	rc = ddi_dma_mem_alloc(dhdl, len, acc_attr,
1869 	    flags & DDI_DMA_CONSISTENT ? DDI_DMA_CONSISTENT : DDI_DMA_STREAMING,
1870 	    DDI_DMA_SLEEP, 0, &va, &real_len, &ahdl);
1871 	if (rc != DDI_SUCCESS) {
1872 		cxgb_printf(sc->dip, CE_WARN,
1873 		    "failed to allocate DMA memory: %d", rc);
1874 
1875 		ddi_dma_free_handle(&dhdl);
1876 		return (ENOMEM);
1877 	}
1878 
1879 	if (len != real_len) {
1880 		cxgb_printf(sc->dip, CE_WARN,
1881 		    "%s: len (%u) != real_len (%u)\n", len, real_len);
1882 	}
1883 
1884 	/*
1885 	 * DMA bindings.
1886 	 */
1887 	rc = ddi_dma_addr_bind_handle(dhdl, NULL, va, real_len, flags, NULL,
1888 	    NULL, &cookie, &ccount);
1889 	if (rc != DDI_DMA_MAPPED) {
1890 		cxgb_printf(sc->dip, CE_WARN,
1891 		    "failed to map DMA memory: %d", rc);
1892 
1893 		ddi_dma_mem_free(&ahdl);
1894 		ddi_dma_free_handle(&dhdl);
1895 		return (ENOMEM);
1896 	}
1897 	if (ccount != 1) {
1898 		cxgb_printf(sc->dip, CE_WARN,
1899 		    "unusable DMA mapping (%d segments)", ccount);
1900 		(void) free_desc_ring(&dhdl, &ahdl);
1901 	}
1902 
1903 	bzero(va, real_len);
1904 	*dma_hdl = dhdl;
1905 	*acc_hdl = ahdl;
1906 	*pba = cookie.dmac_laddress;
1907 	*pva = va;
1908 
1909 	return (0);
1910 }
1911 
1912 static int
1913 free_dma_memory(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl)
1914 {
1915 	(void) ddi_dma_unbind_handle(*dhdl);
1916 	ddi_dma_mem_free(ahdl);
1917 	ddi_dma_free_handle(dhdl);
1918 
1919 	return (0);
1920 }
1921 
1922 static int
1923 alloc_desc_ring(struct adapter *sc, size_t len, int rw,
1924     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl,
1925     uint64_t *pba, caddr_t *pva)
1926 {
1927 	ddi_device_acc_attr_t *acc_attr = &sc->sge.acc_attr_desc;
1928 	ddi_dma_attr_t *dma_attr = &sc->sge.dma_attr_desc;
1929 
1930 	return (alloc_dma_memory(sc, len, DDI_DMA_CONSISTENT | rw, acc_attr,
1931 	    dma_attr, dma_hdl, acc_hdl, pba, pva));
1932 }
1933 
1934 static int
1935 free_desc_ring(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl)
1936 {
1937 	return (free_dma_memory(dhdl, ahdl));
1938 }
1939 
1940 static int
1941 alloc_tx_copybuffer(struct adapter *sc, size_t len,
1942     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl,
1943     uint64_t *pba, caddr_t *pva)
1944 {
1945 	ddi_device_acc_attr_t *acc_attr = &sc->sge.acc_attr_tx;
1946 	ddi_dma_attr_t *dma_attr = &sc->sge.dma_attr_desc; /* NOT dma_attr_tx */
1947 
1948 	return (alloc_dma_memory(sc, len, DDI_DMA_STREAMING | DDI_DMA_WRITE,
1949 	    acc_attr, dma_attr, dma_hdl, acc_hdl, pba, pva));
1950 }
1951 
1952 static inline bool
1953 is_new_response(const struct sge_iq *iq, struct rsp_ctrl **ctrl)
1954 {
1955 	(void) ddi_dma_sync(iq->dhdl, (uintptr_t)iq->cdesc -
1956 	    (uintptr_t)iq->desc, iq->esize, DDI_DMA_SYNC_FORKERNEL);
1957 
1958 	*ctrl = (void *)((uintptr_t)iq->cdesc +
1959 	    (iq->esize - sizeof (struct rsp_ctrl)));
1960 
1961 	return ((((*ctrl)->u.type_gen >> S_RSPD_GEN) == iq->gen));
1962 }
1963 
1964 static inline void
1965 iq_next(struct sge_iq *iq)
1966 {
1967 	iq->cdesc = (void *) ((uintptr_t)iq->cdesc + iq->esize);
1968 	if (++iq->cidx == iq->qsize - 1) {
1969 		iq->cidx = 0;
1970 		iq->gen ^= 1;
1971 		iq->cdesc = iq->desc;
1972 	}
1973 }
1974 
1975 /*
1976  * Fill up the freelist by upto nbufs and maybe ring its doorbell.
1977  *
1978  * Returns non-zero to indicate that it should be added to the list of starving
1979  * freelists.
1980  */
1981 static int
1982 refill_fl(struct adapter *sc, struct sge_fl *fl, int nbufs)
1983 {
1984 	uint64_t *d = &fl->desc[fl->pidx];
1985 	struct fl_sdesc *sd = &fl->sdesc[fl->pidx];
1986 
1987 	FL_LOCK_ASSERT_OWNED(fl);
1988 	ASSERT(nbufs >= 0);
1989 
1990 	if (nbufs > fl->needed)
1991 		nbufs = fl->needed;
1992 
1993 	while (nbufs--) {
1994 		if (sd->rxb != NULL) {
1995 			if (sd->rxb->ref_cnt == 1) {
1996 				/*
1997 				 * Buffer is available for recycling.  Two ways
1998 				 * this can happen:
1999 				 *
2000 				 * a) All the packets DMA'd into it last time
2001 				 *    around were within the rx_copy_threshold
2002 				 *    and no part of the buffer was ever passed
2003 				 *    up (ref_cnt never went over 1).
2004 				 *
2005 				 * b) Packets DMA'd into the buffer were passed
2006 				 *    up but have all been freed by the upper
2007 				 *    layers by now (ref_cnt went over 1 but is
2008 				 *    now back to 1).
2009 				 *
2010 				 * Either way the bus address in the descriptor
2011 				 * ring is already valid.
2012 				 */
2013 				ASSERT(*d == cpu_to_be64(sd->rxb->ba));
2014 				d++;
2015 				goto recycled;
2016 			} else {
2017 				/*
2018 				 * Buffer still in use and we need a
2019 				 * replacement. But first release our reference
2020 				 * on the existing buffer.
2021 				 */
2022 				rxbuf_free(sd->rxb);
2023 			}
2024 		}
2025 
2026 		sd->rxb = rxbuf_alloc(sc->sge.rxbuf_cache, KM_NOSLEEP, 1);
2027 		if (sd->rxb == NULL)
2028 			break;
2029 		*d++ = cpu_to_be64(sd->rxb->ba);
2030 
2031 recycled:	fl->pending++;
2032 		sd++;
2033 		fl->needed--;
2034 		if (++fl->pidx == fl->cap) {
2035 			fl->pidx = 0;
2036 			sd = fl->sdesc;
2037 			d = fl->desc;
2038 		}
2039 	}
2040 
2041 	if (fl->pending >= 8)
2042 		ring_fl_db(sc, fl);
2043 
2044 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
2045 }
2046 
2047 #ifndef TAILQ_FOREACH_SAFE
2048 #define	TAILQ_FOREACH_SAFE(var, head, field, tvar)			\
2049 	for ((var) = TAILQ_FIRST((head));				\
2050 	    (var) && ((tvar) = TAILQ_NEXT((var), field), 1);		\
2051 	    (var) = (tvar))
2052 #endif
2053 
2054 /*
2055  * Attempt to refill all starving freelists.
2056  */
2057 static void
2058 refill_sfl(void *arg)
2059 {
2060 	struct adapter *sc = arg;
2061 	struct sge_fl *fl, *fl_temp;
2062 
2063 	mutex_enter(&sc->sfl_lock);
2064 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
2065 		FL_LOCK(fl);
2066 		(void) refill_fl(sc, fl, 64);
2067 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
2068 			TAILQ_REMOVE(&sc->sfl, fl, link);
2069 			fl->flags &= ~FL_STARVING;
2070 		}
2071 		FL_UNLOCK(fl);
2072 	}
2073 
2074 	if (!TAILQ_EMPTY(&sc->sfl) != 0)
2075 		sc->sfl_timer =  timeout(refill_sfl, sc, drv_usectohz(100000));
2076 	mutex_exit(&sc->sfl_lock);
2077 }
2078 
2079 static void
2080 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
2081 {
2082 	mutex_enter(&sc->sfl_lock);
2083 	FL_LOCK(fl);
2084 	if ((fl->flags & FL_DOOMED) == 0) {
2085 		if (TAILQ_EMPTY(&sc->sfl) != 0) {
2086 			sc->sfl_timer = timeout(refill_sfl, sc,
2087 			    drv_usectohz(100000));
2088 		}
2089 		fl->flags |= FL_STARVING;
2090 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
2091 	}
2092 	FL_UNLOCK(fl);
2093 	mutex_exit(&sc->sfl_lock);
2094 }
2095 
2096 static void
2097 free_fl_bufs(struct sge_fl *fl)
2098 {
2099 	struct fl_sdesc *sd;
2100 	unsigned int i;
2101 
2102 	FL_LOCK_ASSERT_OWNED(fl);
2103 
2104 	for (i = 0; i < fl->cap; i++) {
2105 		sd = &fl->sdesc[i];
2106 
2107 		if (sd->rxb != NULL) {
2108 			rxbuf_free(sd->rxb);
2109 			sd->rxb = NULL;
2110 		}
2111 	}
2112 }
2113 
2114 /*
2115  * Note that fl->cidx and fl->offset are left unchanged in case of failure.
2116  */
2117 static mblk_t *
2118 get_fl_payload(struct sge_fl *fl, uint32_t len_newbuf, int *fl_bufs_used)
2119 {
2120 	struct mblk_pair frame = {0};
2121 	struct rxbuf *rxb;
2122 	mblk_t *m = NULL;
2123 	uint_t nbuf = 0, len, copy, n;
2124 	uint32_t cidx, offset;
2125 
2126 	/*
2127 	 * The SGE won't pack a new frame into the current buffer if the entire
2128 	 * payload doesn't fit in the remaining space.  Move on to the next buf
2129 	 * in that case.
2130 	 */
2131 	if (fl->offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
2132 		fl->offset = 0;
2133 		if (++fl->cidx == fl->cap)
2134 			fl->cidx = 0;
2135 		nbuf++;
2136 	}
2137 	cidx = fl->cidx;
2138 	offset = fl->offset;
2139 
2140 	len = G_RSPD_LEN(len_newbuf);	/* pktshift + payload length */
2141 	copy = (len <= fl->copy_threshold);
2142 	if (copy != 0) {
2143 		frame.head = m = allocb(len, BPRI_HI);
2144 		if (m == NULL)
2145 			return (NULL);
2146 	}
2147 
2148 	while (len) {
2149 		rxb = fl->sdesc[cidx].rxb;
2150 		n = min(len, rxb->buf_size - offset);
2151 
2152 		(void) ddi_dma_sync(rxb->dhdl, offset, n,
2153 		    DDI_DMA_SYNC_FORKERNEL);
2154 
2155 		if (copy != 0)
2156 			bcopy(rxb->va + offset, m->b_wptr, n);
2157 		else {
2158 			m = desballoc((unsigned char *)rxb->va + offset, n,
2159 			    BPRI_HI, &rxb->freefunc);
2160 			if (m == NULL) {
2161 				freemsg(frame.head);
2162 				return (NULL);
2163 			}
2164 			atomic_inc_uint(&rxb->ref_cnt);
2165 			if (frame.head != NULL)
2166 				frame.tail->b_cont = m;
2167 			else
2168 				frame.head = m;
2169 			frame.tail = m;
2170 		}
2171 		m->b_wptr += n;
2172 		len -= n;
2173 		offset += roundup(n, FL_ALIGN);
2174 		ASSERT(offset <= rxb->buf_size);
2175 		if (offset == rxb->buf_size) {
2176 			offset = 0;
2177 			if (++cidx == fl->cap)
2178 				cidx = 0;
2179 			nbuf++;
2180 		}
2181 	}
2182 
2183 	fl->cidx = cidx;
2184 	fl->offset = offset;
2185 	(*fl_bufs_used) += nbuf;
2186 
2187 	ASSERT(frame.head != NULL);
2188 	return (frame.head);
2189 }
2190 
2191 /*
2192  * We'll do immediate data tx for non-LSO, but only when not coalescing.  We're
2193  * willing to use upto 2 hardware descriptors which means a maximum of 96 bytes
2194  * of immediate data.
2195  */
2196 #define	IMM_LEN ( \
2197 	2 * EQ_ESIZE \
2198 	- sizeof (struct fw_eth_tx_pkt_wr) \
2199 	- sizeof (struct cpl_tx_pkt_core))
2200 
2201 /*
2202  * Returns non-zero on failure, no need to cleanup anything in that case.
2203  *
2204  * Note 1: We always try to pull up the mblk if required and return E2BIG only
2205  * if this fails.
2206  *
2207  * Note 2: We'll also pullup incoming mblk if HW_LSO is set and the first mblk
2208  * does not have the TCP header in it.
2209  */
2210 static int
2211 get_frame_txinfo(struct sge_txq *txq, mblk_t **fp, struct txinfo *txinfo,
2212     int sgl_only)
2213 {
2214 	uint32_t flags = 0, len, n;
2215 	mblk_t *m = *fp;
2216 	int rc;
2217 
2218 	TXQ_LOCK_ASSERT_OWNED(txq);	/* will manipulate txb and dma_hdls */
2219 
2220 	mac_hcksum_get(m, NULL, NULL, NULL, NULL, &flags);
2221 	txinfo->flags = flags;
2222 
2223 	mac_lso_get(m, &txinfo->mss, &flags);
2224 	txinfo->flags |= flags;
2225 
2226 	if (flags & HW_LSO)
2227 		sgl_only = 1;	/* Do not allow immediate data with LSO */
2228 
2229 start:	txinfo->nsegs = 0;
2230 	txinfo->hdls_used = 0;
2231 	txinfo->txb_used = 0;
2232 	txinfo->len = 0;
2233 
2234 	/* total length and a rough estimate of # of segments */
2235 	n = 0;
2236 	for (; m; m = m->b_cont) {
2237 		len = MBLKL(m);
2238 		n += (len / PAGE_SIZE) + 1;
2239 		txinfo->len += len;
2240 	}
2241 	m = *fp;
2242 
2243 	if (n >= TX_SGL_SEGS || (flags & HW_LSO && MBLKL(m) < 50)) {
2244 		txq->pullup_early++;
2245 		m = msgpullup(*fp, -1);
2246 		if (m == NULL) {
2247 			txq->pullup_failed++;
2248 			return (E2BIG);	/* (*fp) left as it was */
2249 		}
2250 		freemsg(*fp);
2251 		*fp = m;
2252 		mac_hcksum_set(m, NULL, NULL, NULL, NULL, txinfo->flags);
2253 	}
2254 
2255 	if (txinfo->len <= IMM_LEN && !sgl_only)
2256 		return (0);	/* nsegs = 0 tells caller to use imm. tx */
2257 
2258 	if (txinfo->len <= txq->copy_threshold &&
2259 	    copy_into_txb(txq, m, txinfo->len, txinfo) == 0)
2260 		goto done;
2261 
2262 	for (; m; m = m->b_cont) {
2263 
2264 		len = MBLKL(m);
2265 
2266 		/* Use tx copy buffer if this mblk is small enough */
2267 		if (len <= txq->copy_threshold &&
2268 		    copy_into_txb(txq, m, len, txinfo) == 0)
2269 			continue;
2270 
2271 		/* Add DMA bindings for this mblk to the SGL */
2272 		rc = add_mblk(txq, txinfo, m, len);
2273 
2274 		if (rc == E2BIG ||
2275 		    (txinfo->nsegs == TX_SGL_SEGS && m->b_cont)) {
2276 
2277 			txq->pullup_late++;
2278 			m = msgpullup(*fp, -1);
2279 			if (m != NULL) {
2280 				free_txinfo_resources(txq, txinfo);
2281 				freemsg(*fp);
2282 				*fp = m;
2283 				mac_hcksum_set(m, NULL, NULL, NULL, NULL,
2284 				    txinfo->flags);
2285 				goto start;
2286 			}
2287 
2288 			txq->pullup_failed++;
2289 			rc = E2BIG;
2290 		}
2291 
2292 		if (rc != 0) {
2293 			free_txinfo_resources(txq, txinfo);
2294 			return (rc);
2295 		}
2296 	}
2297 
2298 	ASSERT(txinfo->nsegs > 0 && txinfo->nsegs <= TX_SGL_SEGS);
2299 
2300 done:
2301 
2302 	/*
2303 	 * Store the # of flits required to hold this frame's SGL in nflits.  An
2304 	 * SGL has a (ULPTX header + len0, addr0) tuple optionally followed by
2305 	 * multiple (len0 + len1, addr0, addr1) tuples.  If addr1 is not used
2306 	 * then len1 must be set to 0.
2307 	 */
2308 	n = txinfo->nsegs - 1;
2309 	txinfo->nflits = (3 * n) / 2 + (n & 1) + 2;
2310 	if (n & 1)
2311 		txinfo->sgl.sge[n / 2].len[1] = cpu_to_be32(0);
2312 
2313 	txinfo->sgl.cmd_nsge = cpu_to_be32(V_ULPTX_CMD((u32)ULP_TX_SC_DSGL) |
2314 	    V_ULPTX_NSGE(txinfo->nsegs));
2315 
2316 	return (0);
2317 }
2318 
2319 static inline int
2320 fits_in_txb(struct sge_txq *txq, int len, int *waste)
2321 {
2322 	if (txq->txb_avail < len)
2323 		return (0);
2324 
2325 	if (txq->txb_next + len <= txq->txb_size) {
2326 		*waste = 0;
2327 		return (1);
2328 	}
2329 
2330 	*waste = txq->txb_size - txq->txb_next;
2331 
2332 	return (txq->txb_avail - *waste < len ? 0 : 1);
2333 }
2334 
2335 #define	TXB_CHUNK	64
2336 
2337 /*
2338  * Copies the specified # of bytes into txq's tx copy buffer and updates txinfo
2339  * and txq to indicate resources used.  Caller has to make sure that those many
2340  * bytes are available in the mblk chain (b_cont linked).
2341  */
2342 static inline int
2343 copy_into_txb(struct sge_txq *txq, mblk_t *m, int len, struct txinfo *txinfo)
2344 {
2345 	int waste, n;
2346 
2347 	TXQ_LOCK_ASSERT_OWNED(txq);	/* will manipulate txb */
2348 
2349 	if (!fits_in_txb(txq, len, &waste)) {
2350 		txq->txb_full++;
2351 		return (ENOMEM);
2352 	}
2353 
2354 	if (waste != 0) {
2355 		ASSERT((waste & (TXB_CHUNK - 1)) == 0);
2356 		txinfo->txb_used += waste;
2357 		txq->txb_avail -= waste;
2358 		txq->txb_next = 0;
2359 	}
2360 
2361 	for (n = 0; n < len; m = m->b_cont) {
2362 		bcopy(m->b_rptr, txq->txb_va + txq->txb_next + n, MBLKL(m));
2363 		n += MBLKL(m);
2364 	}
2365 
2366 	add_seg(txinfo, txq->txb_ba + txq->txb_next, len);
2367 
2368 	n = roundup(len, TXB_CHUNK);
2369 	txinfo->txb_used += n;
2370 	txq->txb_avail -= n;
2371 	txq->txb_next += n;
2372 	ASSERT(txq->txb_next <= txq->txb_size);
2373 	if (txq->txb_next == txq->txb_size)
2374 		txq->txb_next = 0;
2375 
2376 	return (0);
2377 }
2378 
2379 static inline void
2380 add_seg(struct txinfo *txinfo, uint64_t ba, uint32_t len)
2381 {
2382 	ASSERT(txinfo->nsegs < TX_SGL_SEGS);	/* must have room */
2383 
2384 	if (txinfo->nsegs != 0) {
2385 		int idx = txinfo->nsegs - 1;
2386 		txinfo->sgl.sge[idx / 2].len[idx & 1] = cpu_to_be32(len);
2387 		txinfo->sgl.sge[idx / 2].addr[idx & 1] = cpu_to_be64(ba);
2388 	} else {
2389 		txinfo->sgl.len0 = cpu_to_be32(len);
2390 		txinfo->sgl.addr0 = cpu_to_be64(ba);
2391 	}
2392 	txinfo->nsegs++;
2393 }
2394 
2395 /*
2396  * This function cleans up any partially allocated resources when it fails so
2397  * there's nothing for the caller to clean up in that case.
2398  *
2399  * EIO indicates permanent failure.  Caller should drop the frame containing
2400  * this mblk and continue.
2401  *
2402  * E2BIG indicates that the SGL length for this mblk exceeds the hardware
2403  * limit.  Caller should pull up the frame before trying to send it out.
2404  * (This error means our pullup_early heuristic did not work for this frame)
2405  *
2406  * ENOMEM indicates a temporary shortage of resources (DMA handles, other DMA
2407  * resources, etc.).  Caller should suspend the tx queue and wait for reclaim to
2408  * free up resources.
2409  */
2410 static inline int
2411 add_mblk(struct sge_txq *txq, struct txinfo *txinfo, mblk_t *m, int len)
2412 {
2413 	ddi_dma_handle_t dhdl;
2414 	ddi_dma_cookie_t cookie;
2415 	uint_t ccount = 0;
2416 	int rc;
2417 
2418 	TXQ_LOCK_ASSERT_OWNED(txq);	/* will manipulate dhdls */
2419 
2420 	if (txq->tx_dhdl_avail == 0) {
2421 		txq->dma_hdl_failed++;
2422 		return (ENOMEM);
2423 	}
2424 
2425 	dhdl = txq->tx_dhdl[txq->tx_dhdl_pidx];
2426 	rc = ddi_dma_addr_bind_handle(dhdl, NULL, (caddr_t)m->b_rptr, len,
2427 	    DDI_DMA_WRITE | DDI_DMA_STREAMING, DDI_DMA_DONTWAIT, NULL, &cookie,
2428 	    &ccount);
2429 	if (rc != DDI_DMA_MAPPED) {
2430 		txq->dma_map_failed++;
2431 
2432 		ASSERT(rc != DDI_DMA_INUSE && rc != DDI_DMA_PARTIAL_MAP);
2433 
2434 		return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EIO);
2435 	}
2436 
2437 	if (ccount + txinfo->nsegs > TX_SGL_SEGS) {
2438 		(void) ddi_dma_unbind_handle(dhdl);
2439 		return (E2BIG);
2440 	}
2441 
2442 	add_seg(txinfo, cookie.dmac_laddress, cookie.dmac_size);
2443 	while (--ccount) {
2444 		ddi_dma_nextcookie(dhdl, &cookie);
2445 		add_seg(txinfo, cookie.dmac_laddress, cookie.dmac_size);
2446 	}
2447 
2448 	if (++txq->tx_dhdl_pidx == txq->tx_dhdl_total)
2449 		txq->tx_dhdl_pidx = 0;
2450 	txq->tx_dhdl_avail--;
2451 	txinfo->hdls_used++;
2452 
2453 	return (0);
2454 }
2455 
2456 /*
2457  * Releases all the txq resources used up in the specified txinfo.
2458  */
2459 static void
2460 free_txinfo_resources(struct sge_txq *txq, struct txinfo *txinfo)
2461 {
2462 	int n;
2463 
2464 	TXQ_LOCK_ASSERT_OWNED(txq);	/* dhdls, txb */
2465 
2466 	n = txinfo->txb_used;
2467 	if (n > 0) {
2468 		txq->txb_avail += n;
2469 		if (n <= txq->txb_next)
2470 			txq->txb_next -= n;
2471 		else {
2472 			n -= txq->txb_next;
2473 			txq->txb_next = txq->txb_size - n;
2474 		}
2475 	}
2476 
2477 	for (n = txinfo->hdls_used; n > 0; n--) {
2478 		if (txq->tx_dhdl_pidx > 0)
2479 			txq->tx_dhdl_pidx--;
2480 		else
2481 			txq->tx_dhdl_pidx = txq->tx_dhdl_total - 1;
2482 		txq->tx_dhdl_avail++;
2483 		(void) ddi_dma_unbind_handle(txq->tx_dhdl[txq->tx_dhdl_pidx]);
2484 	}
2485 }
2486 
2487 /*
2488  * Returns 0 to indicate that m has been accepted into a coalesced tx work
2489  * request.  It has either been folded into txpkts or txpkts was flushed and m
2490  * has started a new coalesced work request (as the first frame in a fresh
2491  * txpkts).
2492  *
2493  * Returns non-zero to indicate a failure - caller is responsible for
2494  * transmitting m, if there was anything in txpkts it has been flushed.
2495  */
2496 static int
2497 add_to_txpkts(struct sge_txq *txq, struct txpkts *txpkts, mblk_t *m,
2498     struct txinfo *txinfo)
2499 {
2500 	struct sge_eq *eq = &txq->eq;
2501 	int can_coalesce;
2502 	struct tx_sdesc *txsd;
2503 	uint8_t flits;
2504 
2505 	TXQ_LOCK_ASSERT_OWNED(txq);
2506 
2507 	if (txpkts->npkt > 0) {
2508 		flits = TXPKTS_PKT_HDR + txinfo->nflits;
2509 		can_coalesce = (txinfo->flags & HW_LSO) == 0 &&
2510 		    txpkts->nflits + flits <= TX_WR_FLITS &&
2511 		    txpkts->nflits + flits <= eq->avail * 8 &&
2512 		    txpkts->plen + txinfo->len < 65536;
2513 
2514 		if (can_coalesce != 0) {
2515 			txpkts->tail->b_next = m;
2516 			txpkts->tail = m;
2517 			txpkts->npkt++;
2518 			txpkts->nflits += flits;
2519 			txpkts->plen += txinfo->len;
2520 
2521 			txsd = &txq->sdesc[eq->pidx];
2522 			txsd->txb_used += txinfo->txb_used;
2523 			txsd->hdls_used += txinfo->hdls_used;
2524 
2525 			return (0);
2526 		}
2527 
2528 		/*
2529 		 * Couldn't coalesce m into txpkts.  The first order of business
2530 		 * is to send txpkts on its way.  Then we'll revisit m.
2531 		 */
2532 		write_txpkts_wr(txq, txpkts);
2533 	}
2534 
2535 	/*
2536 	 * Check if we can start a new coalesced tx work request with m as
2537 	 * the first packet in it.
2538 	 */
2539 
2540 	ASSERT(txpkts->npkt == 0);
2541 	ASSERT(txinfo->len < 65536);
2542 
2543 	flits = TXPKTS_WR_HDR + txinfo->nflits;
2544 	can_coalesce = (txinfo->flags & HW_LSO) == 0 &&
2545 	    flits <= eq->avail * 8 && flits <= TX_WR_FLITS;
2546 
2547 	if (can_coalesce == 0)
2548 		return (EINVAL);
2549 
2550 	/*
2551 	 * Start a fresh coalesced tx WR with m as the first frame in it.
2552 	 */
2553 	txpkts->tail = m;
2554 	txpkts->npkt = 1;
2555 	txpkts->nflits = flits;
2556 	txpkts->flitp = &eq->desc[eq->pidx].flit[2];
2557 	txpkts->plen = txinfo->len;
2558 
2559 	txsd = &txq->sdesc[eq->pidx];
2560 	txsd->m = m;
2561 	txsd->txb_used = txinfo->txb_used;
2562 	txsd->hdls_used = txinfo->hdls_used;
2563 
2564 	return (0);
2565 }
2566 
2567 /*
2568  * Note that write_txpkts_wr can never run out of hardware descriptors (but
2569  * write_txpkt_wr can).  add_to_txpkts ensures that a frame is accepted for
2570  * coalescing only if sufficient hardware descriptors are available.
2571  */
2572 static void
2573 write_txpkts_wr(struct sge_txq *txq, struct txpkts *txpkts)
2574 {
2575 	struct sge_eq *eq = &txq->eq;
2576 	struct fw_eth_tx_pkts_wr *wr;
2577 	struct tx_sdesc *txsd;
2578 	uint32_t ctrl;
2579 	uint16_t ndesc;
2580 
2581 	TXQ_LOCK_ASSERT_OWNED(txq);	/* pidx, avail */
2582 
2583 	ndesc = howmany(txpkts->nflits, 8);
2584 
2585 	wr = (void *)&eq->desc[eq->pidx];
2586 	wr->op_pkd = cpu_to_be32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR) |
2587 	    V_FW_WR_IMMDLEN(0)); /* immdlen does not matter in this WR */
2588 	ctrl = V_FW_WR_LEN16(howmany(txpkts->nflits, 2));
2589 	if (eq->avail == ndesc)
2590 		ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ;
2591 	wr->equiq_to_len16 = cpu_to_be32(ctrl);
2592 	wr->plen = cpu_to_be16(txpkts->plen);
2593 	wr->npkt = txpkts->npkt;
2594 	wr->r3 = wr->type = 0;
2595 
2596 	/* Everything else already written */
2597 
2598 	txsd = &txq->sdesc[eq->pidx];
2599 	txsd->desc_used = ndesc;
2600 
2601 	txq->txb_used += txsd->txb_used / TXB_CHUNK;
2602 	txq->hdl_used += txsd->hdls_used;
2603 
2604 	ASSERT(eq->avail >= ndesc);
2605 
2606 	eq->pending += ndesc;
2607 	eq->avail -= ndesc;
2608 	eq->pidx += ndesc;
2609 	if (eq->pidx >= eq->cap)
2610 		eq->pidx -= eq->cap;
2611 
2612 	txq->txpkts_pkts += txpkts->npkt;
2613 	txq->txpkts_wrs++;
2614 	txpkts->npkt = 0;	/* emptied */
2615 }
2616 
2617 static int
2618 write_txpkt_wr(struct port_info *pi, struct sge_txq *txq, mblk_t *m,
2619     struct txinfo *txinfo)
2620 {
2621 	struct sge_eq *eq = &txq->eq;
2622 	struct fw_eth_tx_pkt_wr *wr;
2623 	struct cpl_tx_pkt_core *cpl;
2624 	uint32_t ctrl;	/* used in many unrelated places */
2625 	uint64_t ctrl1;
2626 	int nflits, ndesc;
2627 	struct tx_sdesc *txsd;
2628 	caddr_t dst;
2629 
2630 	TXQ_LOCK_ASSERT_OWNED(txq);	/* pidx, avail */
2631 
2632 	/*
2633 	 * Do we have enough flits to send this frame out?
2634 	 */
2635 	ctrl = sizeof (struct cpl_tx_pkt_core);
2636 	if (txinfo->flags & HW_LSO) {
2637 		nflits = TXPKT_LSO_WR_HDR;
2638 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
2639 	} else
2640 		nflits = TXPKT_WR_HDR;
2641 	if (txinfo->nsegs > 0)
2642 		nflits += txinfo->nflits;
2643 	else {
2644 		nflits += howmany(txinfo->len, 8);
2645 		ctrl += txinfo->len;
2646 	}
2647 	ndesc = howmany(nflits, 8);
2648 	if (ndesc > eq->avail)
2649 		return (ENOMEM);
2650 
2651 	/* Firmware work request header */
2652 	wr = (void *)&eq->desc[eq->pidx];
2653 	wr->op_immdlen = cpu_to_be32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
2654 	    V_FW_WR_IMMDLEN(ctrl));
2655 	ctrl = V_FW_WR_LEN16(howmany(nflits, 2));
2656 	if (eq->avail == ndesc)
2657 		ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ;
2658 	wr->equiq_to_len16 = cpu_to_be32(ctrl);
2659 	wr->r3 = 0;
2660 
2661 	if (txinfo->flags & HW_LSO) {
2662 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
2663 		char *p = (void *)m->b_rptr;
2664 		ctrl = V_LSO_OPCODE((u32)CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
2665 		    F_LSO_LAST_SLICE;
2666 
2667 		/* LINTED: E_BAD_PTR_CAST_ALIGN */
2668 		if (((struct ether_header *)p)->ether_type ==
2669 		    htons(ETHERTYPE_VLAN)) {
2670 			ctrl |= V_LSO_ETHHDR_LEN(1);
2671 			p += sizeof (struct ether_vlan_header);
2672 		} else
2673 			p += sizeof (struct ether_header);
2674 
2675 		/* LINTED: E_BAD_PTR_CAST_ALIGN for IPH_HDR_LENGTH() */
2676 		ctrl |= V_LSO_IPHDR_LEN(IPH_HDR_LENGTH(p) / 4);
2677 		/* LINTED: E_BAD_PTR_CAST_ALIGN for IPH_HDR_LENGTH() */
2678 		p += IPH_HDR_LENGTH(p);
2679 		ctrl |= V_LSO_TCPHDR_LEN(TCP_HDR_LENGTH((tcph_t *)p) / 4);
2680 
2681 		lso->lso_ctrl = cpu_to_be32(ctrl);
2682 		lso->ipid_ofst = cpu_to_be16(0);
2683 		lso->mss = cpu_to_be16(txinfo->mss);
2684 		lso->seqno_offset = cpu_to_be32(0);
2685 		if (is_t4(pi->adapter->params.chip))
2686 			lso->len = cpu_to_be32(txinfo->len);
2687 		else
2688 			lso->len = cpu_to_be32(V_LSO_T5_XFER_SIZE(txinfo->len));
2689 
2690 		cpl = (void *)(lso + 1);
2691 
2692 		txq->tso_wrs++;
2693 	} else
2694 		cpl = (void *)(wr + 1);
2695 
2696 	/* Checksum offload */
2697 	ctrl1 = 0;
2698 	if (!(txinfo->flags & HCK_IPV4_HDRCKSUM))
2699 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
2700 	if (!(txinfo->flags & HCK_FULLCKSUM))
2701 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
2702 	if (ctrl1 == 0)
2703 		txq->txcsum++;	/* some hardware assistance provided */
2704 
2705 	/* CPL header */
2706 	cpl->ctrl0 = cpu_to_be32(V_TXPKT_OPCODE(CPL_TX_PKT) |
2707 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf));
2708 	cpl->pack = 0;
2709 	cpl->len = cpu_to_be16(txinfo->len);
2710 	cpl->ctrl1 = cpu_to_be64(ctrl1);
2711 
2712 	/* Software descriptor */
2713 	txsd = &txq->sdesc[eq->pidx];
2714 	txsd->m = m;
2715 	txsd->txb_used = txinfo->txb_used;
2716 	txsd->hdls_used = txinfo->hdls_used;
2717 	/* LINTED: E_ASSIGN_NARROW_CONV */
2718 	txsd->desc_used = ndesc;
2719 
2720 	txq->txb_used += txinfo->txb_used / TXB_CHUNK;
2721 	txq->hdl_used += txinfo->hdls_used;
2722 
2723 	eq->pending += ndesc;
2724 	eq->avail -= ndesc;
2725 	eq->pidx += ndesc;
2726 	if (eq->pidx >= eq->cap)
2727 		eq->pidx -= eq->cap;
2728 
2729 	/* SGL */
2730 	dst = (void *)(cpl + 1);
2731 	if (txinfo->nsegs > 0) {
2732 		txq->sgl_wrs++;
2733 		copy_to_txd(eq, (void *)&txinfo->sgl, &dst, txinfo->nflits * 8);
2734 
2735 		/* Need to zero-pad to a 16 byte boundary if not on one */
2736 		if ((uintptr_t)dst & 0xf)
2737 			/* LINTED: E_BAD_PTR_CAST_ALIGN */
2738 			*(uint64_t *)dst = 0;
2739 
2740 	} else {
2741 		txq->imm_wrs++;
2742 #ifdef DEBUG
2743 		ctrl = txinfo->len;
2744 #endif
2745 		for (; m; m = m->b_cont) {
2746 			copy_to_txd(eq, (void *)m->b_rptr, &dst, MBLKL(m));
2747 #ifdef DEBUG
2748 			ctrl -= MBLKL(m);
2749 #endif
2750 		}
2751 		ASSERT(ctrl == 0);
2752 	}
2753 
2754 	txq->txpkt_wrs++;
2755 	return (0);
2756 }
2757 
2758 static inline void
2759 write_ulp_cpl_sgl(struct port_info *pi, struct sge_txq *txq,
2760     struct txpkts *txpkts, struct txinfo *txinfo)
2761 {
2762 	struct ulp_txpkt *ulpmc;
2763 	struct ulptx_idata *ulpsc;
2764 	struct cpl_tx_pkt_core *cpl;
2765 	uintptr_t flitp, start, end;
2766 	uint64_t ctrl;
2767 	caddr_t dst;
2768 
2769 	ASSERT(txpkts->npkt > 0);
2770 
2771 	start = (uintptr_t)txq->eq.desc;
2772 	end = (uintptr_t)txq->eq.spg;
2773 
2774 	/* Checksum offload */
2775 	ctrl = 0;
2776 	if (!(txinfo->flags & HCK_IPV4_HDRCKSUM))
2777 		ctrl |= F_TXPKT_IPCSUM_DIS;
2778 	if (!(txinfo->flags & HCK_FULLCKSUM))
2779 		ctrl |= F_TXPKT_L4CSUM_DIS;
2780 	if (ctrl == 0)
2781 		txq->txcsum++;	/* some hardware assistance provided */
2782 
2783 	/*
2784 	 * The previous packet's SGL must have ended at a 16 byte boundary (this
2785 	 * is required by the firmware/hardware).  It follows that flitp cannot
2786 	 * wrap around between the ULPTX master command and ULPTX subcommand (8
2787 	 * bytes each), and that it can not wrap around in the middle of the
2788 	 * cpl_tx_pkt_core either.
2789 	 */
2790 	flitp = (uintptr_t)txpkts->flitp;
2791 	ASSERT((flitp & 0xf) == 0);
2792 
2793 	/* ULP master command */
2794 	ulpmc = (void *)flitp;
2795 	ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0));
2796 	ulpmc->len = htonl(howmany(sizeof (*ulpmc) + sizeof (*ulpsc) +
2797 	    sizeof (*cpl) + 8 * txinfo->nflits, 16));
2798 
2799 	/* ULP subcommand */
2800 	ulpsc = (void *)(ulpmc + 1);
2801 	ulpsc->cmd_more = cpu_to_be32(V_ULPTX_CMD((u32)ULP_TX_SC_IMM) |
2802 	    F_ULP_TX_SC_MORE);
2803 	ulpsc->len = cpu_to_be32(sizeof (struct cpl_tx_pkt_core));
2804 
2805 	flitp += sizeof (*ulpmc) + sizeof (*ulpsc);
2806 	if (flitp == end)
2807 		flitp = start;
2808 
2809 	/* CPL_TX_PKT */
2810 	cpl = (void *)flitp;
2811 	cpl->ctrl0 = cpu_to_be32(V_TXPKT_OPCODE(CPL_TX_PKT) |
2812 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf));
2813 	cpl->pack = 0;
2814 	cpl->len = cpu_to_be16(txinfo->len);
2815 	cpl->ctrl1 = cpu_to_be64(ctrl);
2816 
2817 	flitp += sizeof (*cpl);
2818 	if (flitp == end)
2819 		flitp = start;
2820 
2821 	/* SGL for this frame */
2822 	dst = (caddr_t)flitp;
2823 	copy_to_txd(&txq->eq, (void *)&txinfo->sgl, &dst, txinfo->nflits * 8);
2824 	flitp = (uintptr_t)dst;
2825 
2826 	/* Zero pad and advance to a 16 byte boundary if not already at one. */
2827 	if (flitp & 0xf) {
2828 
2829 		/* no matter what, flitp should be on an 8 byte boundary */
2830 		ASSERT((flitp & 0x7) == 0);
2831 
2832 		*(uint64_t *)flitp = 0;
2833 		flitp += sizeof (uint64_t);
2834 		txpkts->nflits++;
2835 	}
2836 
2837 	if (flitp == end)
2838 		flitp = start;
2839 
2840 	txpkts->flitp = (void *)flitp;
2841 }
2842 
2843 static inline void
2844 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
2845 {
2846 	if ((uintptr_t)(*to) + len <= (uintptr_t)eq->spg) {
2847 		bcopy(from, *to, len);
2848 		(*to) += len;
2849 	} else {
2850 		int portion = (uintptr_t)eq->spg - (uintptr_t)(*to);
2851 
2852 		bcopy(from, *to, portion);
2853 		from += portion;
2854 		portion = len - portion;	/* remaining */
2855 		bcopy(from, (void *)eq->desc, portion);
2856 		(*to) = (caddr_t)eq->desc + portion;
2857 	}
2858 }
2859 
2860 static inline void
2861 ring_tx_db(struct adapter *sc, struct sge_eq *eq)
2862 {
2863 	int val, db_mode;
2864 	u_int db = eq->doorbells;
2865 
2866 	if (eq->pending > 1)
2867 		db &= ~DOORBELL_WCWR;
2868 
2869 	if (eq->pending > eq->pidx) {
2870 		int offset = eq->cap - (eq->pending - eq->pidx);
2871 
2872 		/* pidx has wrapped around since last doorbell */
2873 
2874 		(void) ddi_dma_sync(eq->desc_dhdl,
2875 		    offset * sizeof (struct tx_desc), 0,
2876 		    DDI_DMA_SYNC_FORDEV);
2877 		(void) ddi_dma_sync(eq->desc_dhdl,
2878 		    0, eq->pidx * sizeof (struct tx_desc),
2879 		    DDI_DMA_SYNC_FORDEV);
2880 	} else if (eq->pending > 0) {
2881 		(void) ddi_dma_sync(eq->desc_dhdl,
2882 		    (eq->pidx - eq->pending) * sizeof (struct tx_desc),
2883 		    eq->pending * sizeof (struct tx_desc),
2884 		    DDI_DMA_SYNC_FORDEV);
2885 	}
2886 
2887 	membar_producer();
2888 
2889 	if (is_t4(sc->params.chip))
2890 		val = V_PIDX(eq->pending);
2891 	else
2892 		val = V_PIDX_T5(eq->pending);
2893 
2894 	db_mode = (1 << (ffs(db) - 1));
2895 	switch (db_mode) {
2896 		case DOORBELL_UDB:
2897 			*eq->udb = LE_32(V_QID(eq->udb_qid) | val);
2898 			break;
2899 
2900 		case DOORBELL_WCWR:
2901 			{
2902 				volatile uint64_t *dst, *src;
2903 				int i;
2904 				/*
2905 				 * Queues whose 128B doorbell segment fits in
2906 				 * the page do not use relative qid
2907 				 * (udb_qid is always 0).  Only queues with
2908 				 * doorbell segments can do WCWR.
2909 				 */
2910 				ASSERT(eq->udb_qid == 0 && eq->pending == 1);
2911 
2912 				dst = (volatile void *)((uintptr_t)eq->udb +
2913 				    UDBS_WR_OFFSET - UDBS_DB_OFFSET);
2914 				i = eq->pidx ? eq->pidx - 1 : eq->cap - 1;
2915 				src = (void *)&eq->desc[i];
2916 				while (src != (void *)&eq->desc[i + 1])
2917 				        *dst++ = *src++;
2918 				membar_producer();
2919 				break;
2920 			}
2921 
2922 		case DOORBELL_UDBWC:
2923 			*eq->udb = LE_32(V_QID(eq->udb_qid) | val);
2924 			membar_producer();
2925 			break;
2926 
2927 		case DOORBELL_KDB:
2928 			t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
2929 			    V_QID(eq->cntxt_id) | val);
2930 			break;
2931 	}
2932 
2933 	eq->pending = 0;
2934 }
2935 
2936 static int
2937 reclaim_tx_descs(struct sge_txq *txq, int howmany)
2938 {
2939 	struct tx_sdesc *txsd;
2940 	uint_t cidx, can_reclaim, reclaimed, txb_freed, hdls_freed;
2941 	struct sge_eq *eq = &txq->eq;
2942 
2943 	EQ_LOCK_ASSERT_OWNED(eq);
2944 
2945 	cidx = eq->spg->cidx;	/* stable snapshot */
2946 	cidx = be16_to_cpu(cidx);
2947 
2948 	if (cidx >= eq->cidx)
2949 		can_reclaim = cidx - eq->cidx;
2950 	else
2951 		can_reclaim = cidx + eq->cap - eq->cidx;
2952 
2953 	if (can_reclaim == 0)
2954 		return (0);
2955 
2956 	txb_freed = hdls_freed = reclaimed = 0;
2957 	do {
2958 		int ndesc;
2959 
2960 		txsd = &txq->sdesc[eq->cidx];
2961 		ndesc = txsd->desc_used;
2962 
2963 		/* Firmware doesn't return "partial" credits. */
2964 		ASSERT(can_reclaim >= ndesc);
2965 
2966 		/*
2967 		 * We always keep mblk around, even for immediate data.  If mblk
2968 		 * is NULL, this has to be the software descriptor for a credit
2969 		 * flush work request.
2970 		 */
2971 		if (txsd->m != NULL)
2972 			freemsgchain(txsd->m);
2973 #ifdef DEBUG
2974 		else {
2975 			ASSERT(txsd->txb_used == 0);
2976 			ASSERT(txsd->hdls_used == 0);
2977 			ASSERT(ndesc == 1);
2978 		}
2979 #endif
2980 
2981 		txb_freed += txsd->txb_used;
2982 		hdls_freed += txsd->hdls_used;
2983 		reclaimed += ndesc;
2984 
2985 		eq->cidx += ndesc;
2986 		if (eq->cidx >= eq->cap)
2987 			eq->cidx -= eq->cap;
2988 
2989 		can_reclaim -= ndesc;
2990 
2991 	} while (can_reclaim && reclaimed < howmany);
2992 
2993 	eq->avail += reclaimed;
2994 	ASSERT(eq->avail < eq->cap);	/* avail tops out at (cap - 1) */
2995 
2996 	txq->txb_avail += txb_freed;
2997 
2998 	txq->tx_dhdl_avail += hdls_freed;
2999 	ASSERT(txq->tx_dhdl_avail <= txq->tx_dhdl_total);
3000 	for (; hdls_freed; hdls_freed--) {
3001 		(void) ddi_dma_unbind_handle(txq->tx_dhdl[txq->tx_dhdl_cidx]);
3002 		if (++txq->tx_dhdl_cidx == txq->tx_dhdl_total)
3003 			txq->tx_dhdl_cidx = 0;
3004 	}
3005 
3006 	return (reclaimed);
3007 }
3008 
3009 static void
3010 write_txqflush_wr(struct sge_txq *txq)
3011 {
3012 	struct sge_eq *eq = &txq->eq;
3013 	struct fw_eq_flush_wr *wr;
3014 	struct tx_sdesc *txsd;
3015 
3016 	EQ_LOCK_ASSERT_OWNED(eq);
3017 	ASSERT(eq->avail > 0);
3018 
3019 	wr = (void *)&eq->desc[eq->pidx];
3020 	bzero(wr, sizeof (*wr));
3021 	wr->opcode = FW_EQ_FLUSH_WR;
3022 	wr->equiq_to_len16 = cpu_to_be32(V_FW_WR_LEN16(sizeof (*wr) / 16) |
3023 	    F_FW_WR_EQUEQ | F_FW_WR_EQUIQ);
3024 
3025 	txsd = &txq->sdesc[eq->pidx];
3026 	txsd->m = NULL;
3027 	txsd->txb_used = 0;
3028 	txsd->hdls_used = 0;
3029 	txsd->desc_used = 1;
3030 
3031 	eq->pending++;
3032 	eq->avail--;
3033 	if (++eq->pidx == eq->cap)
3034 		eq->pidx = 0;
3035 }
3036 
3037 static int
3038 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, mblk_t *m)
3039 {
3040 	struct sge_rxq *rxq = (void *)iq;
3041 	struct mblk_pair chain = {0};
3042 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
3043 
3044 	iq->intr_next = iq->intr_params;
3045 
3046 	m->b_rptr += FL_PKTSHIFT;
3047 
3048 	/* TODO: what about cpl->ip_frag? */
3049 	if (cpl->csum_calc && !cpl->err_vec && !cpl->ip_frag) {
3050 		mac_hcksum_set(m, 0, 0, 0, 0xffff,
3051 		    HCK_FULLCKSUM_OK | HCK_FULLCKSUM |
3052 		    HCK_IPV4_HDRCKSUM_OK);
3053 		rxq->rxcsum++;
3054 	}
3055 
3056 	/* Add to the chain that we'll send up */
3057 	if (chain.head != NULL)
3058 		chain.tail->b_next = m;
3059 	else
3060 		chain.head = m;
3061 	chain.tail = m;
3062 
3063 	t4_mac_rx(rxq->port, rxq, chain.head);
3064 
3065 	return (0);
3066 }
3067 
3068 #define	FL_HW_IDX(idx)	((idx) >> 3)
3069 
3070 static inline void
3071 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
3072 {
3073 	int desc_start, desc_last, ndesc;
3074 	uint32_t v;
3075 
3076 	ndesc = FL_HW_IDX(fl->pending);
3077 
3078 	/* Hold back one credit if pidx = cidx */
3079 	if (FL_HW_IDX(fl->pidx) == FL_HW_IDX(fl->cidx))
3080 		ndesc--;
3081 
3082 	/*
3083 	 * There are chances of ndesc modified above (to avoid pidx = cidx).
3084 	 * If there is nothing to post, return.
3085 	 */
3086 	if (ndesc <= 0)
3087 		return;
3088 
3089 	desc_last = FL_HW_IDX(fl->pidx);
3090 
3091 	if (fl->pidx < fl->pending) {
3092 		/* There was a wrap */
3093 		desc_start = FL_HW_IDX(fl->pidx + fl->cap - fl->pending);
3094 
3095 		/* From desc_start to the end of list */
3096 		(void) ddi_dma_sync(fl->dhdl, desc_start * RX_FL_ESIZE, 0,
3097 		    DDI_DMA_SYNC_FORDEV);
3098 
3099 		/* From start of list to the desc_last */
3100 		if (desc_last != 0)
3101 			(void) ddi_dma_sync(fl->dhdl, 0, desc_last *
3102 			    RX_FL_ESIZE, DDI_DMA_SYNC_FORDEV);
3103 	} else {
3104 		/* There was no wrap, sync from start_desc to last_desc */
3105 		desc_start = FL_HW_IDX(fl->pidx - fl->pending);
3106 		(void) ddi_dma_sync(fl->dhdl, desc_start * RX_FL_ESIZE,
3107 		    ndesc * RX_FL_ESIZE, DDI_DMA_SYNC_FORDEV);
3108 	}
3109 
3110 	if (is_t4(sc->params.chip))
3111 		v = V_PIDX(ndesc);
3112 	else
3113 		v = V_PIDX_T5(ndesc) | F_DBTYPE;
3114 	v |= F_DBPRIO | V_QID(fl->cntxt_id) | V_PIDX(ndesc);
3115 
3116 	membar_producer();
3117 
3118 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL), v);
3119 
3120 	/*
3121 	 * Update pending count:
3122 	 * Deduct the number of descriptors posted
3123 	 */
3124 	fl->pending -= ndesc * 8;
3125 }
3126 
3127 /* ARGSUSED */
3128 static int
3129 handle_fw_rpl(struct sge_iq *iq, const struct rss_header *rss, mblk_t *m)
3130 {
3131 	struct adapter *sc = iq->adapter;
3132 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
3133 
3134 	ASSERT(m == NULL);
3135 
3136 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
3137 		const struct rss_header *rss2;
3138 
3139 		rss2 = (const struct rss_header *)&cpl->data[0];
3140 		return (sc->cpl_handler[rss2->opcode](iq, rss2, m));
3141 	}
3142 	return (sc->fw_msg_handler[cpl->type](sc, &cpl->data[0]));
3143 }
3144 
3145 int
3146 t4_alloc_tx_maps(struct adapter *sc, struct tx_maps *txmaps, int count,
3147     int flags)
3148 {
3149 	int i, rc;
3150 
3151 	txmaps->map_total =  count;
3152 	txmaps->map_avail = txmaps->map_cidx = txmaps->map_pidx = 0;
3153 
3154 	txmaps->map =  kmem_zalloc(sizeof (ddi_dma_handle_t) *
3155 	    txmaps->map_total, flags);
3156 
3157 	for (i = 0; i < count; i++) {
3158 		rc = ddi_dma_alloc_handle(sc->dip, &sc->sge.dma_attr_tx,
3159 		    DDI_DMA_SLEEP, 0, &txmaps->map[i]);
3160 		if (rc != DDI_SUCCESS) {
3161 			cxgb_printf(sc->dip, CE_WARN,
3162 			    "%s: failed to allocate DMA handle (%d)",
3163 			    __func__, rc);
3164 			return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EINVAL);
3165 		}
3166 		txmaps->map_avail++;
3167 	}
3168 
3169 	return (0);
3170 }
3171 
3172 #define	KS_UINIT(x)	kstat_named_init(&kstatp->x, #x, KSTAT_DATA_ULONG)
3173 #define	KS_CINIT(x)	kstat_named_init(&kstatp->x, #x, KSTAT_DATA_CHAR)
3174 #define	KS_U_SET(x, y)	kstatp->x.value.ul = (y)
3175 #define	KS_U_FROM(x, y)	kstatp->x.value.ul = (y)->x
3176 #define	KS_C_SET(x, ...)	\
3177 			(void) snprintf(kstatp->x.value.c, 16,  __VA_ARGS__)
3178 
3179 /*
3180  * cxgbe:X:config
3181  */
3182 struct cxgbe_port_config_kstats {
3183 	kstat_named_t idx;
3184 	kstat_named_t nrxq;
3185 	kstat_named_t ntxq;
3186 	kstat_named_t first_rxq;
3187 	kstat_named_t first_txq;
3188 	kstat_named_t controller;
3189 	kstat_named_t factory_mac_address;
3190 };
3191 
3192 /*
3193  * cxgbe:X:info
3194  */
3195 struct cxgbe_port_info_kstats {
3196 	kstat_named_t transceiver;
3197 	kstat_named_t rx_ovflow0;
3198 	kstat_named_t rx_ovflow1;
3199 	kstat_named_t rx_ovflow2;
3200 	kstat_named_t rx_ovflow3;
3201 	kstat_named_t rx_trunc0;
3202 	kstat_named_t rx_trunc1;
3203 	kstat_named_t rx_trunc2;
3204 	kstat_named_t rx_trunc3;
3205 	kstat_named_t tx_pause;
3206 	kstat_named_t rx_pause;
3207 };
3208 
3209 static kstat_t *
3210 setup_port_config_kstats(struct port_info *pi)
3211 {
3212 	kstat_t *ksp;
3213 	struct cxgbe_port_config_kstats *kstatp;
3214 	int ndata;
3215 	dev_info_t *pdip = ddi_get_parent(pi->dip);
3216 	uint8_t *ma = &pi->hw_addr[0];
3217 
3218 	ndata = sizeof (struct cxgbe_port_config_kstats) /
3219 	    sizeof (kstat_named_t);
3220 
3221 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), "config",
3222 	    "net", KSTAT_TYPE_NAMED, ndata, 0);
3223 	if (ksp == NULL) {
3224 		cxgb_printf(pi->dip, CE_WARN, "failed to initialize kstats.");
3225 		return (NULL);
3226 	}
3227 
3228 	kstatp = (struct cxgbe_port_config_kstats *)ksp->ks_data;
3229 
3230 	KS_UINIT(idx);
3231 	KS_UINIT(nrxq);
3232 	KS_UINIT(ntxq);
3233 	KS_UINIT(first_rxq);
3234 	KS_UINIT(first_txq);
3235 	KS_CINIT(controller);
3236 	KS_CINIT(factory_mac_address);
3237 
3238 	KS_U_SET(idx, pi->port_id);
3239 	KS_U_SET(nrxq, pi->nrxq);
3240 	KS_U_SET(ntxq, pi->ntxq);
3241 	KS_U_SET(first_rxq, pi->first_rxq);
3242 	KS_U_SET(first_txq, pi->first_txq);
3243 	KS_C_SET(controller, "%s%d", ddi_driver_name(pdip),
3244 	    ddi_get_instance(pdip));
3245 	KS_C_SET(factory_mac_address, "%02X%02X%02X%02X%02X%02X",
3246 	    ma[0], ma[1], ma[2], ma[3], ma[4], ma[5]);
3247 
3248 	/* Do NOT set ksp->ks_update.  These kstats do not change. */
3249 
3250 	/* Install the kstat */
3251 	ksp->ks_private = (void *)pi;
3252 	kstat_install(ksp);
3253 
3254 	return (ksp);
3255 }
3256 
3257 static kstat_t *
3258 setup_port_info_kstats(struct port_info *pi)
3259 {
3260 	kstat_t *ksp;
3261 	struct cxgbe_port_info_kstats *kstatp;
3262 	int ndata;
3263 
3264 	ndata = sizeof (struct cxgbe_port_info_kstats) / sizeof (kstat_named_t);
3265 
3266 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), "info",
3267 	    "net", KSTAT_TYPE_NAMED, ndata, 0);
3268 	if (ksp == NULL) {
3269 		cxgb_printf(pi->dip, CE_WARN, "failed to initialize kstats.");
3270 		return (NULL);
3271 	}
3272 
3273 	kstatp = (struct cxgbe_port_info_kstats *)ksp->ks_data;
3274 
3275 	KS_CINIT(transceiver);
3276 	KS_UINIT(rx_ovflow0);
3277 	KS_UINIT(rx_ovflow1);
3278 	KS_UINIT(rx_ovflow2);
3279 	KS_UINIT(rx_ovflow3);
3280 	KS_UINIT(rx_trunc0);
3281 	KS_UINIT(rx_trunc1);
3282 	KS_UINIT(rx_trunc2);
3283 	KS_UINIT(rx_trunc3);
3284 	KS_UINIT(tx_pause);
3285 	KS_UINIT(rx_pause);
3286 
3287 	/* Install the kstat */
3288 	ksp->ks_update = update_port_info_kstats;
3289 	ksp->ks_private = (void *)pi;
3290 	kstat_install(ksp);
3291 
3292 	return (ksp);
3293 }
3294 
3295 static int
3296 update_port_info_kstats(kstat_t *ksp, int rw)
3297 {
3298 	struct cxgbe_port_info_kstats *kstatp =
3299 	    (struct cxgbe_port_info_kstats *)ksp->ks_data;
3300 	struct port_info *pi = ksp->ks_private;
3301 	static const char *mod_str[] = { NULL, "LR", "SR", "ER", "TWINAX",
3302 	    "active TWINAX", "LRM" };
3303 	uint32_t bgmap;
3304 
3305 	if (rw == KSTAT_WRITE)
3306 		return (0);
3307 
3308 	if (pi->mod_type == FW_PORT_MOD_TYPE_NONE)
3309 		KS_C_SET(transceiver, "unplugged");
3310 	else if (pi->mod_type == FW_PORT_MOD_TYPE_UNKNOWN)
3311 		KS_C_SET(transceiver, "unknown");
3312 	else if (pi->mod_type == FW_PORT_MOD_TYPE_NOTSUPPORTED)
3313 		KS_C_SET(transceiver, "unsupported");
3314 	else if (pi->mod_type > 0 && pi->mod_type < ARRAY_SIZE(mod_str))
3315 		KS_C_SET(transceiver, "%s", mod_str[pi->mod_type]);
3316 	else
3317 		KS_C_SET(transceiver, "type %d", pi->mod_type);
3318 
3319 #define	GET_STAT(name) t4_read_reg64(pi->adapter, \
3320 	    PORT_REG(pi->port_id, A_MPS_PORT_STAT_##name##_L))
3321 #define	GET_STAT_COM(name) t4_read_reg64(pi->adapter, \
3322 	    A_MPS_STAT_##name##_L)
3323 
3324 	bgmap = G_NUMPORTS(t4_read_reg(pi->adapter, A_MPS_CMN_CTL));
3325 	if (bgmap == 0)
3326 		bgmap = (pi->port_id == 0) ? 0xf : 0;
3327 	else if (bgmap == 1)
3328 		bgmap = (pi->port_id < 2) ? (3 << (2 * pi->port_id)) : 0;
3329 	else
3330 		bgmap = 1;
3331 
3332 	KS_U_SET(rx_ovflow0, (bgmap & 1) ?
3333 	    GET_STAT_COM(RX_BG_0_MAC_DROP_FRAME) : 0);
3334 	KS_U_SET(rx_ovflow1, (bgmap & 2) ?
3335 	    GET_STAT_COM(RX_BG_1_MAC_DROP_FRAME) : 0);
3336 	KS_U_SET(rx_ovflow2, (bgmap & 4) ?
3337 	    GET_STAT_COM(RX_BG_2_MAC_DROP_FRAME) : 0);
3338 	KS_U_SET(rx_ovflow3, (bgmap & 8) ?
3339 	    GET_STAT_COM(RX_BG_3_MAC_DROP_FRAME) : 0);
3340 	KS_U_SET(rx_trunc0,  (bgmap & 1) ?
3341 	    GET_STAT_COM(RX_BG_0_MAC_TRUNC_FRAME) : 0);
3342 	KS_U_SET(rx_trunc1,  (bgmap & 2) ?
3343 	    GET_STAT_COM(RX_BG_1_MAC_TRUNC_FRAME) : 0);
3344 	KS_U_SET(rx_trunc2,  (bgmap & 4) ?
3345 	    GET_STAT_COM(RX_BG_2_MAC_TRUNC_FRAME) : 0);
3346 	KS_U_SET(rx_trunc3,  (bgmap & 8) ?
3347 	    GET_STAT_COM(RX_BG_3_MAC_TRUNC_FRAME) : 0);
3348 
3349 	KS_U_SET(tx_pause, GET_STAT(TX_PORT_PAUSE));
3350 	KS_U_SET(rx_pause, GET_STAT(RX_PORT_PAUSE));
3351 
3352 	return (0);
3353 
3354 }
3355 
3356 /*
3357  * cxgbe:X:rxqY
3358  */
3359 struct rxq_kstats {
3360 	kstat_named_t rxcsum;
3361 	kstat_named_t nomem;
3362 };
3363 
3364 static kstat_t *
3365 setup_rxq_kstats(struct port_info *pi, struct sge_rxq *rxq, int idx)
3366 {
3367 	struct kstat *ksp;
3368 	struct rxq_kstats *kstatp;
3369 	int ndata;
3370 	char str[16];
3371 
3372 	ndata = sizeof (struct rxq_kstats) / sizeof (kstat_named_t);
3373 	(void) snprintf(str, sizeof (str), "rxq%u", idx);
3374 
3375 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), str, "rxq",
3376 	    KSTAT_TYPE_NAMED, ndata, 0);
3377 	if (ksp == NULL) {
3378 		cxgb_printf(pi->dip, CE_WARN,
3379 		    "%s: failed to initialize rxq kstats for queue %d.",
3380 		    __func__, idx);
3381 		return (NULL);
3382 	}
3383 
3384 	kstatp = (struct rxq_kstats *)ksp->ks_data;
3385 
3386 	KS_UINIT(rxcsum);
3387 	KS_UINIT(nomem);
3388 
3389 	ksp->ks_update = update_rxq_kstats;
3390 	ksp->ks_private = (void *)rxq;
3391 	kstat_install(ksp);
3392 
3393 	return (ksp);
3394 }
3395 
3396 static int
3397 update_rxq_kstats(kstat_t *ksp, int rw)
3398 {
3399 	struct rxq_kstats *kstatp = (struct rxq_kstats *)ksp->ks_data;
3400 	struct sge_rxq *rxq = ksp->ks_private;
3401 
3402 	if (rw == KSTAT_WRITE)
3403 		return (0);
3404 
3405 	KS_U_FROM(rxcsum, rxq);
3406 	KS_U_FROM(nomem, rxq);
3407 
3408 	return (0);
3409 }
3410 
3411 /*
3412  * cxgbe:X:txqY
3413  */
3414 struct txq_kstats {
3415 	kstat_named_t txcsum;
3416 	kstat_named_t tso_wrs;
3417 	kstat_named_t imm_wrs;
3418 	kstat_named_t sgl_wrs;
3419 	kstat_named_t txpkt_wrs;
3420 	kstat_named_t txpkts_wrs;
3421 	kstat_named_t txpkts_pkts;
3422 	kstat_named_t txb_used;
3423 	kstat_named_t hdl_used;
3424 	kstat_named_t txb_full;
3425 	kstat_named_t dma_hdl_failed;
3426 	kstat_named_t dma_map_failed;
3427 	kstat_named_t qfull;
3428 	kstat_named_t qflush;
3429 	kstat_named_t pullup_early;
3430 	kstat_named_t pullup_late;
3431 	kstat_named_t pullup_failed;
3432 };
3433 
3434 static kstat_t *
3435 setup_txq_kstats(struct port_info *pi, struct sge_txq *txq, int idx)
3436 {
3437 	struct kstat *ksp;
3438 	struct txq_kstats *kstatp;
3439 	int ndata;
3440 	char str[16];
3441 
3442 	ndata = sizeof (struct txq_kstats) / sizeof (kstat_named_t);
3443 	(void) snprintf(str, sizeof (str), "txq%u", idx);
3444 
3445 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), str, "txq",
3446 	    KSTAT_TYPE_NAMED, ndata, 0);
3447 	if (ksp == NULL) {
3448 		cxgb_printf(pi->dip, CE_WARN,
3449 		    "%s: failed to initialize txq kstats for queue %d.",
3450 		    __func__, idx);
3451 		return (NULL);
3452 	}
3453 
3454 	kstatp = (struct txq_kstats *)ksp->ks_data;
3455 
3456 	KS_UINIT(txcsum);
3457 	KS_UINIT(tso_wrs);
3458 	KS_UINIT(imm_wrs);
3459 	KS_UINIT(sgl_wrs);
3460 	KS_UINIT(txpkt_wrs);
3461 	KS_UINIT(txpkts_wrs);
3462 	KS_UINIT(txpkts_pkts);
3463 	KS_UINIT(txb_used);
3464 	KS_UINIT(hdl_used);
3465 	KS_UINIT(txb_full);
3466 	KS_UINIT(dma_hdl_failed);
3467 	KS_UINIT(dma_map_failed);
3468 	KS_UINIT(qfull);
3469 	KS_UINIT(qflush);
3470 	KS_UINIT(pullup_early);
3471 	KS_UINIT(pullup_late);
3472 	KS_UINIT(pullup_failed);
3473 
3474 	ksp->ks_update = update_txq_kstats;
3475 	ksp->ks_private = (void *)txq;
3476 	kstat_install(ksp);
3477 
3478 	return (ksp);
3479 }
3480 
3481 static int
3482 update_txq_kstats(kstat_t *ksp, int rw)
3483 {
3484 	struct txq_kstats *kstatp = (struct txq_kstats *)ksp->ks_data;
3485 	struct sge_txq *txq = ksp->ks_private;
3486 
3487 	if (rw == KSTAT_WRITE)
3488 		return (0);
3489 
3490 	KS_U_FROM(txcsum, txq);
3491 	KS_U_FROM(tso_wrs, txq);
3492 	KS_U_FROM(imm_wrs, txq);
3493 	KS_U_FROM(sgl_wrs, txq);
3494 	KS_U_FROM(txpkt_wrs, txq);
3495 	KS_U_FROM(txpkts_wrs, txq);
3496 	KS_U_FROM(txpkts_pkts, txq);
3497 	KS_U_FROM(txb_used, txq);
3498 	KS_U_FROM(hdl_used, txq);
3499 	KS_U_FROM(txb_full, txq);
3500 	KS_U_FROM(dma_hdl_failed, txq);
3501 	KS_U_FROM(dma_map_failed, txq);
3502 	KS_U_FROM(qfull, txq);
3503 	KS_U_FROM(qflush, txq);
3504 	KS_U_FROM(pullup_early, txq);
3505 	KS_U_FROM(pullup_late, txq);
3506 	KS_U_FROM(pullup_failed, txq);
3507 
3508 	return (0);
3509 }
3510