xref: /illumos-gate/usr/src/uts/common/io/cxgbe/t4nex/t4_sge.c (revision ddb365bfc9e868ad24ccdcb0dc91af18b10df082)
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 /*
24  * Copyright 2021 Oxide Computer Company
25  */
26 
27 #include <sys/ddi.h>
28 #include <sys/sunddi.h>
29 #include <sys/sunndi.h>
30 #include <sys/atomic.h>
31 #include <sys/dlpi.h>
32 #include <sys/pattr.h>
33 #include <sys/strsubr.h>
34 #include <sys/stream.h>
35 #include <sys/strsun.h>
36 #include <inet/ip.h>
37 #include <inet/tcp.h>
38 
39 #include "version.h"
40 #include "common/common.h"
41 #include "common/t4_msg.h"
42 #include "common/t4_regs.h"
43 #include "common/t4_regs_values.h"
44 
45 /* TODO: Tune. */
46 int rx_buf_size = 8192;
47 int tx_copy_threshold = 256;
48 uint16_t rx_copy_threshold = 256;
49 
50 /* Used to track coalesced tx work request */
51 struct txpkts {
52 	mblk_t *tail;		/* head is in the software descriptor */
53 	uint64_t *flitp;	/* ptr to flit where next pkt should start */
54 	uint8_t npkt;		/* # of packets in this work request */
55 	uint8_t nflits;		/* # of flits used by this work request */
56 	uint16_t plen;		/* total payload (sum of all packets) */
57 };
58 
59 /* All information needed to tx a frame */
60 struct txinfo {
61 	uint32_t len;		/* Total length of frame */
62 	uint32_t flags;		/* Checksum and LSO flags */
63 	uint32_t mss;		/* MSS for LSO */
64 	uint8_t nsegs;		/* # of segments in the SGL, 0 means imm. tx */
65 	uint8_t nflits;		/* # of flits needed for the SGL */
66 	uint8_t hdls_used;	/* # of DMA handles used */
67 	uint32_t txb_used;	/* txb_space used */
68 	struct ulptx_sgl sgl __attribute__((aligned(8)));
69 	struct ulptx_sge_pair reserved[TX_SGL_SEGS / 2];
70 };
71 
72 static int service_iq(struct sge_iq *iq, int budget);
73 static inline void init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx,
74     int8_t pktc_idx, int qsize, uint8_t esize);
75 static inline void init_fl(struct sge_fl *fl, uint16_t qsize);
76 static inline void init_eq(struct adapter *sc, struct sge_eq *eq,
77     uint16_t eqtype, uint16_t qsize,uint8_t tx_chan, uint16_t iqid);
78 static int alloc_iq_fl(struct port_info *pi, struct sge_iq *iq,
79     struct sge_fl *fl, int intr_idx, int cong);
80 static int free_iq_fl(struct port_info *pi, struct sge_iq *iq,
81     struct sge_fl *fl);
82 static int alloc_fwq(struct adapter *sc);
83 static int free_fwq(struct adapter *sc);
84 #ifdef TCP_OFFLOAD_ENABLE
85 static int alloc_mgmtq(struct adapter *sc);
86 #endif
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 #ifdef TCP_OFFLOAD_ENABLE
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 #ifdef TCP_OFFLOAD_ENABLE
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 #ifdef TCP_OFFLOAD_ENABLE
106 static int alloc_wrq(struct adapter *sc, struct port_info *pi,
107     struct sge_wrq *wrq, int idx);
108 static int free_wrq(struct adapter *sc, struct sge_wrq *wrq);
109 #endif
110 static int alloc_txq(struct port_info *pi, struct sge_txq *txq, int idx);
111 static int free_txq(struct port_info *pi, struct sge_txq *txq);
112 static int alloc_dma_memory(struct adapter *sc, size_t len, int flags,
113     ddi_device_acc_attr_t *acc_attr, ddi_dma_attr_t *dma_attr,
114     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl, uint64_t *pba,
115     caddr_t *pva);
116 static int free_dma_memory(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl);
117 static int alloc_desc_ring(struct adapter *sc, size_t len, int rw,
118     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl, uint64_t *pba,
119     caddr_t *pva);
120 static int free_desc_ring(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl);
121 static int alloc_tx_copybuffer(struct adapter *sc, size_t len,
122     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl, uint64_t *pba,
123     caddr_t *pva);
124 static inline bool is_new_response(const struct sge_iq *iq,
125     struct rsp_ctrl **ctrl);
126 static inline void iq_next(struct sge_iq *iq);
127 static int refill_fl(struct adapter *sc, struct sge_fl *fl, int nbufs);
128 static void refill_sfl(void *arg);
129 static void add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl);
130 static void free_fl_bufs(struct sge_fl *fl);
131 static mblk_t *get_fl_payload(struct adapter *sc, struct sge_fl *fl,
132     uint32_t len_newbuf, int *fl_bufs_used);
133 static int get_frame_txinfo(struct sge_txq *txq, mblk_t **fp,
134     struct txinfo *txinfo, int sgl_only);
135 static inline int fits_in_txb(struct sge_txq *txq, int len, int *waste);
136 static inline int copy_into_txb(struct sge_txq *txq, mblk_t *m, int len,
137     struct txinfo *txinfo);
138 static inline void add_seg(struct txinfo *txinfo, uint64_t ba, uint32_t len);
139 static inline int add_mblk(struct sge_txq *txq, struct txinfo *txinfo,
140     mblk_t *m, int len);
141 static void free_txinfo_resources(struct sge_txq *txq, struct txinfo *txinfo);
142 static int add_to_txpkts(struct sge_txq *txq, struct txpkts *txpkts, mblk_t *m,
143     struct txinfo *txinfo);
144 static void write_txpkts_wr(struct sge_txq *txq, struct txpkts *txpkts);
145 static int write_txpkt_wr(struct port_info *pi, struct sge_txq *txq, mblk_t *m,
146     struct txinfo *txinfo);
147 static inline void write_ulp_cpl_sgl(struct port_info *pi, struct sge_txq *txq,
148     struct txpkts *txpkts, struct txinfo *txinfo);
149 static inline void copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to,
150     int len);
151 static inline void ring_tx_db(struct adapter *sc, struct sge_eq *eq);
152 static int reclaim_tx_descs(struct sge_txq *txq, int howmany);
153 static void write_txqflush_wr(struct sge_txq *txq);
154 static int t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss,
155     mblk_t *m);
156 static inline void ring_fl_db(struct adapter *sc, struct sge_fl *fl);
157 static kstat_t *setup_port_config_kstats(struct port_info *pi);
158 static kstat_t *setup_port_info_kstats(struct port_info *pi);
159 static kstat_t *setup_rxq_kstats(struct port_info *pi, struct sge_rxq *rxq,
160     int idx);
161 static int update_rxq_kstats(kstat_t *ksp, int rw);
162 static int update_port_info_kstats(kstat_t *ksp, int rw);
163 static kstat_t *setup_txq_kstats(struct port_info *pi, struct sge_txq *txq,
164     int idx);
165 static int update_txq_kstats(kstat_t *ksp, int rw);
166 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
167     mblk_t *);
168 static int handle_fw_rpl(struct sge_iq *iq, const struct rss_header *rss,
169     mblk_t *m);
170 
171 static inline int
172 reclaimable(struct sge_eq *eq)
173 {
174 	unsigned int cidx;
175 
176 	cidx = eq->spg->cidx;   /* stable snapshot */
177 	cidx = be16_to_cpu(cidx);
178 
179 	if (cidx >= eq->cidx)
180 		return (cidx - eq->cidx);
181 	else
182 		return (cidx + eq->cap - eq->cidx);
183 }
184 
185 void
186 t4_sge_init(struct adapter *sc)
187 {
188 	struct driver_properties *p = &sc->props;
189 	ddi_dma_attr_t *dma_attr;
190 	ddi_device_acc_attr_t *acc_attr;
191 	uint32_t sge_control, sge_conm_ctrl;
192 	int egress_threshold;
193 
194 	/*
195 	 * Device access and DMA attributes for descriptor rings
196 	 */
197 	acc_attr = &sc->sge.acc_attr_desc;
198 	acc_attr->devacc_attr_version = DDI_DEVICE_ATTR_V0;
199 	acc_attr->devacc_attr_endian_flags = DDI_NEVERSWAP_ACC;
200 	acc_attr->devacc_attr_dataorder = DDI_STRICTORDER_ACC;
201 
202 	dma_attr = &sc->sge.dma_attr_desc;
203 	dma_attr->dma_attr_version = DMA_ATTR_V0;
204 	dma_attr->dma_attr_addr_lo = 0;
205 	dma_attr->dma_attr_addr_hi = UINT64_MAX;
206 	dma_attr->dma_attr_count_max = UINT64_MAX;
207 	dma_attr->dma_attr_align = 512;
208 	dma_attr->dma_attr_burstsizes = 0xfff;
209 	dma_attr->dma_attr_minxfer = 1;
210 	dma_attr->dma_attr_maxxfer = UINT64_MAX;
211 	dma_attr->dma_attr_seg = UINT64_MAX;
212 	dma_attr->dma_attr_sgllen = 1;
213 	dma_attr->dma_attr_granular = 1;
214 	dma_attr->dma_attr_flags = 0;
215 
216 	/*
217 	 * Device access and DMA attributes for tx buffers
218 	 */
219 	acc_attr = &sc->sge.acc_attr_tx;
220 	acc_attr->devacc_attr_version = DDI_DEVICE_ATTR_V0;
221 	acc_attr->devacc_attr_endian_flags = DDI_NEVERSWAP_ACC;
222 
223 	dma_attr = &sc->sge.dma_attr_tx;
224 	dma_attr->dma_attr_version = DMA_ATTR_V0;
225 	dma_attr->dma_attr_addr_lo = 0;
226 	dma_attr->dma_attr_addr_hi = UINT64_MAX;
227 	dma_attr->dma_attr_count_max = UINT64_MAX;
228 	dma_attr->dma_attr_align = 1;
229 	dma_attr->dma_attr_burstsizes = 0xfff;
230 	dma_attr->dma_attr_minxfer = 1;
231 	dma_attr->dma_attr_maxxfer = UINT64_MAX;
232 	dma_attr->dma_attr_seg = UINT64_MAX;
233 	dma_attr->dma_attr_sgllen = TX_SGL_SEGS;
234 	dma_attr->dma_attr_granular = 1;
235 	dma_attr->dma_attr_flags = 0;
236 
237 	/*
238 	 * Ingress Padding Boundary and Egress Status Page Size are set up by
239 	 * t4_fixup_host_params().
240 	 */
241 	sge_control = t4_read_reg(sc, A_SGE_CONTROL);
242 	sc->sge.pktshift = G_PKTSHIFT(sge_control);
243 	sc->sge.stat_len = (sge_control & F_EGRSTATUSPAGESIZE) ? 128 : 64;
244 
245 	/* t4_nex uses FLM packed mode */
246 	sc->sge.fl_align = t4_fl_pkt_align(sc, true);
247 
248 	/*
249 	 * Device access and DMA attributes for rx buffers
250 	 */
251 	sc->sge.rxb_params.dip = sc->dip;
252 	sc->sge.rxb_params.buf_size = rx_buf_size;
253 
254 	acc_attr = &sc->sge.rxb_params.acc_attr_rx;
255 	acc_attr->devacc_attr_version = DDI_DEVICE_ATTR_V0;
256 	acc_attr->devacc_attr_endian_flags = DDI_NEVERSWAP_ACC;
257 
258 	dma_attr = &sc->sge.rxb_params.dma_attr_rx;
259 	dma_attr->dma_attr_version = DMA_ATTR_V0;
260 	dma_attr->dma_attr_addr_lo = 0;
261 	dma_attr->dma_attr_addr_hi = UINT64_MAX;
262 	dma_attr->dma_attr_count_max = UINT64_MAX;
263 	/*
264 	 * Low 4 bits of an rx buffer address have a special meaning to the SGE
265 	 * and an rx buf cannot have an address with any of these bits set.
266 	 * FL_ALIGN is >= 32 so we're sure things are ok.
267 	 */
268 	dma_attr->dma_attr_align = sc->sge.fl_align;
269 	dma_attr->dma_attr_burstsizes = 0xfff;
270 	dma_attr->dma_attr_minxfer = 1;
271 	dma_attr->dma_attr_maxxfer = UINT64_MAX;
272 	dma_attr->dma_attr_seg = UINT64_MAX;
273 	dma_attr->dma_attr_sgllen = 1;
274 	dma_attr->dma_attr_granular = 1;
275 	dma_attr->dma_attr_flags = 0;
276 
277 	sc->sge.rxbuf_cache = rxbuf_cache_create(&sc->sge.rxb_params);
278 
279 	/*
280 	 * A FL with <= fl_starve_thres buffers is starving and a periodic
281 	 * timer will attempt to refill it.  This needs to be larger than the
282 	 * SGE's Egress Congestion Threshold.  If it isn't, then we can get
283 	 * stuck waiting for new packets while the SGE is waiting for us to
284 	 * give it more Free List entries.  (Note that the SGE's Egress
285 	 * Congestion Threshold is in units of 2 Free List pointers.) For T4,
286 	 * there was only a single field to control this.  For T5 there's the
287 	 * original field which now only applies to Unpacked Mode Free List
288 	 * buffers and a new field which only applies to Packed Mode Free List
289 	 * buffers.
290 	 */
291 
292 	sge_conm_ctrl = t4_read_reg(sc, A_SGE_CONM_CTRL);
293 	switch (CHELSIO_CHIP_VERSION(sc->params.chip)) {
294 	case CHELSIO_T4:
295 		egress_threshold = G_EGRTHRESHOLD(sge_conm_ctrl);
296 		break;
297 	case CHELSIO_T5:
298 		egress_threshold = G_EGRTHRESHOLDPACKING(sge_conm_ctrl);
299 		break;
300 	case CHELSIO_T6:
301 	default:
302 		egress_threshold = G_T6_EGRTHRESHOLDPACKING(sge_conm_ctrl);
303 	}
304 	sc->sge.fl_starve_threshold = 2*egress_threshold + 1;
305 
306 	t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0, rx_buf_size);
307 
308 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD,
309 	    V_THRESHOLD_0(p->counter_val[0]) |
310 	    V_THRESHOLD_1(p->counter_val[1]) |
311 	    V_THRESHOLD_2(p->counter_val[2]) |
312 	    V_THRESHOLD_3(p->counter_val[3]));
313 
314 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1,
315 	    V_TIMERVALUE0(us_to_core_ticks(sc, p->timer_val[0])) |
316 	    V_TIMERVALUE1(us_to_core_ticks(sc, p->timer_val[1])));
317 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3,
318 	    V_TIMERVALUE2(us_to_core_ticks(sc, p->timer_val[2])) |
319 	    V_TIMERVALUE3(us_to_core_ticks(sc, p->timer_val[3])));
320 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5,
321 	    V_TIMERVALUE4(us_to_core_ticks(sc, p->timer_val[4])) |
322 	    V_TIMERVALUE5(us_to_core_ticks(sc, p->timer_val[5])));
323 
324 	(void) t4_register_cpl_handler(sc, CPL_FW4_MSG, handle_fw_rpl);
325 	(void) t4_register_cpl_handler(sc, CPL_FW6_MSG, handle_fw_rpl);
326 	(void) t4_register_cpl_handler(sc, CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
327 	(void) t4_register_cpl_handler(sc, CPL_RX_PKT, t4_eth_rx);
328 	(void) t4_register_fw_msg_handler(sc, FW6_TYPE_CMD_RPL,
329 		    t4_handle_fw_rpl);
330 }
331 
332 /*
333  * Allocate and initialize the firmware event queue and the forwarded interrupt
334  * queues, if any.  The adapter owns all these queues as they are not associated
335  * with any particular port.
336  *
337  * Returns errno on failure.  Resources allocated up to that point may still be
338  * allocated.  Caller is responsible for cleanup in case this function fails.
339  */
340 int
341 t4_setup_adapter_queues(struct adapter *sc)
342 {
343 	int rc;
344 
345 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
346 
347 	/*
348 	 * Firmware event queue
349 	 */
350 	rc = alloc_fwq(sc);
351 	if (rc != 0)
352 		return (rc);
353 
354 #ifdef TCP_OFFLOAD_ENABLE
355 	/*
356 	 * Management queue.  This is just a control queue that uses the fwq as
357 	 * its associated iq.
358 	 */
359 	rc = alloc_mgmtq(sc);
360 #endif
361 
362 	return (rc);
363 }
364 
365 /*
366  * Idempotent
367  */
368 int
369 t4_teardown_adapter_queues(struct adapter *sc)
370 {
371 
372 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
373 
374 	(void) free_fwq(sc);
375 
376 	return (0);
377 }
378 
379 static inline int
380 first_vector(struct port_info *pi)
381 {
382 	struct adapter *sc = pi->adapter;
383 	int rc = T4_EXTRA_INTR, i;
384 
385 	if (sc->intr_count == 1)
386 		return (0);
387 
388 	for_each_port(sc, i) {
389 		struct port_info *p = sc->port[i];
390 
391 		if (i == pi->port_id)
392 			break;
393 
394 #ifdef TCP_OFFLOAD_ENABLE
395 		if (!(sc->flags & INTR_FWD))
396 			rc += p->nrxq + p->nofldrxq;
397 		else
398 			rc += max(p->nrxq, p->nofldrxq);
399 #else
400 		/*
401 		 * Not compiled with offload support and intr_count > 1.  Only
402 		 * NIC queues exist and they'd better be taking direct
403 		 * interrupts.
404 		 */
405 		ASSERT(!(sc->flags & INTR_FWD));
406 		rc += p->nrxq;
407 #endif
408 	}
409 	return (rc);
410 }
411 
412 /*
413  * Given an arbitrary "index," come up with an iq that can be used by other
414  * queues (of this port) for interrupt forwarding, SGE egress updates, etc.
415  * The iq returned is guaranteed to be something that takes direct interrupts.
416  */
417 static struct sge_iq *
418 port_intr_iq(struct port_info *pi, int idx)
419 {
420 	struct adapter *sc = pi->adapter;
421 	struct sge *s = &sc->sge;
422 	struct sge_iq *iq = NULL;
423 
424 	if (sc->intr_count == 1)
425 		return (&sc->sge.fwq);
426 
427 #ifdef TCP_OFFLOAD_ENABLE
428 	if (!(sc->flags & INTR_FWD)) {
429 		idx %= pi->nrxq + pi->nofldrxq;
430 
431 		if (idx >= pi->nrxq) {
432 			idx -= pi->nrxq;
433 			iq = &s->ofld_rxq[pi->first_ofld_rxq + idx].iq;
434 		} else
435 			iq = &s->rxq[pi->first_rxq + idx].iq;
436 
437 	} else {
438 		idx %= max(pi->nrxq, pi->nofldrxq);
439 
440 		if (pi->nrxq >= pi->nofldrxq)
441 			iq = &s->rxq[pi->first_rxq + idx].iq;
442 		else
443 			iq = &s->ofld_rxq[pi->first_ofld_rxq + idx].iq;
444 	}
445 #else
446 	/*
447 	 * Not compiled with offload support and intr_count > 1.  Only NIC
448 	 * queues exist and they'd better be taking direct interrupts.
449 	 */
450 	ASSERT(!(sc->flags & INTR_FWD));
451 
452 	idx %= pi->nrxq;
453 	iq = &s->rxq[pi->first_rxq + idx].iq;
454 #endif
455 
456 	return (iq);
457 }
458 
459 int
460 t4_setup_port_queues(struct port_info *pi)
461 {
462 	int rc = 0, i, intr_idx, j;
463 	struct sge_rxq *rxq;
464 	struct sge_txq *txq;
465 #ifdef TCP_OFFLOAD_ENABLE
466 	int iqid;
467 	struct sge_wrq *ctrlq;
468 	struct sge_ofld_rxq *ofld_rxq;
469 	struct sge_wrq *ofld_txq;
470 #endif
471 	struct adapter *sc = pi->adapter;
472 	struct driver_properties *p = &sc->props;
473 
474 	pi->ksp_config = setup_port_config_kstats(pi);
475 	pi->ksp_info   = setup_port_info_kstats(pi);
476 
477 	/* Interrupt vector to start from (when using multiple vectors) */
478 	intr_idx = first_vector(pi);
479 
480 	/*
481 	 * First pass over all rx queues (NIC and TOE):
482 	 * a) initialize iq and fl
483 	 * b) allocate queue iff it will take direct interrupts.
484 	 */
485 
486 	for_each_rxq(pi, i, rxq) {
487 
488 		init_iq(&rxq->iq, sc, pi->tmr_idx, pi->pktc_idx, p->qsize_rxq,
489 		    RX_IQ_ESIZE);
490 
491 		init_fl(&rxq->fl, p->qsize_rxq / 8); /* 8 bufs in each entry */
492 
493 		if ((!(sc->flags & INTR_FWD))
494 #ifdef TCP_OFFLOAD_ENABLE
495 		    || (sc->intr_count > 1 && pi->nrxq >= pi->nofldrxq)
496 #else
497 		    || (sc->intr_count > 1 && pi->nrxq)
498 #endif
499 		   ) {
500 			rxq->iq.flags |= IQ_INTR;
501 			rc = alloc_rxq(pi, rxq, intr_idx, i);
502 			if (rc != 0)
503 				goto done;
504 			intr_idx++;
505 		}
506 
507 	}
508 
509 #ifdef TCP_OFFLOAD_ENABLE
510 	for_each_ofld_rxq(pi, i, ofld_rxq) {
511 
512 		init_iq(&ofld_rxq->iq, sc, pi->tmr_idx, pi->pktc_idx,
513 		    p->qsize_rxq, RX_IQ_ESIZE);
514 
515 		init_fl(&ofld_rxq->fl, p->qsize_rxq / 8);
516 
517 		if (!(sc->flags & INTR_FWD) ||
518 		    (sc->intr_count > 1 && pi->nofldrxq > pi->nrxq)) {
519 			ofld_rxq->iq.flags = IQ_INTR;
520 			rc = alloc_ofld_rxq(pi, ofld_rxq, intr_idx);
521 			if (rc != 0)
522 				goto done;
523 
524 			intr_idx++;
525 		}
526 	}
527 #endif
528 
529 	/*
530 	 * Second pass over all rx queues (NIC and TOE).  The queues forwarding
531 	 * their interrupts are allocated now.
532 	 */
533 	j = 0;
534 	for_each_rxq(pi, i, rxq) {
535 		if (rxq->iq.flags & IQ_INTR)
536 			continue;
537 
538 		intr_idx = port_intr_iq(pi, j)->abs_id;
539 
540 		rc = alloc_rxq(pi, rxq, intr_idx, i);
541 		if (rc != 0)
542 			goto done;
543 		j++;
544 	}
545 
546 #ifdef TCP_OFFLOAD_ENABLE
547 	for_each_ofld_rxq(pi, i, ofld_rxq) {
548 		if (ofld_rxq->iq.flags & IQ_INTR)
549 			continue;
550 
551 		intr_idx = port_intr_iq(pi, j)->abs_id;
552 		rc = alloc_ofld_rxq(pi, ofld_rxq, intr_idx);
553 		if (rc != 0)
554 			goto done;
555 		j++;
556 	}
557 #endif
558 	/*
559 	 * Now the tx queues.  Only one pass needed.
560 	 */
561 	j = 0;
562 	for_each_txq(pi, i, txq) {
563 		uint16_t iqid;
564 
565 		iqid = port_intr_iq(pi, j)->cntxt_id;
566 		init_eq(sc, &txq->eq, EQ_ETH, p->qsize_txq, pi->tx_chan, iqid);
567 		rc = alloc_txq(pi, txq, i);
568 		if (rc != 0)
569 			goto done;
570 	}
571 
572 #ifdef TCP_OFFLOAD_ENABLE
573 	for_each_ofld_txq(pi, i, ofld_txq) {
574 		uint16_t iqid;
575 
576 		iqid = port_intr_iq(pi, j)->cntxt_id;
577 		init_eq(sc, &ofld_txq->eq, EQ_OFLD, p->qsize_txq, pi->tx_chan,
578 		    iqid);
579 		rc = alloc_wrq(sc, pi, ofld_txq, i);
580 		if (rc != 0)
581 			goto done;
582 	}
583 
584 	/*
585 	 * Finally, the control queue.
586 	 */
587 	ctrlq = &sc->sge.ctrlq[pi->port_id];
588 	iqid = port_intr_iq(pi, 0)->cntxt_id;
589 	init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, pi->tx_chan, iqid);
590 	rc = alloc_wrq(sc, pi, ctrlq, 0);
591 #endif
592 
593 done:
594 	if (rc != 0)
595 		(void) t4_teardown_port_queues(pi);
596 
597 	return (rc);
598 }
599 
600 /*
601  * Idempotent
602  */
603 int
604 t4_teardown_port_queues(struct port_info *pi)
605 {
606 	int i;
607 	struct sge_rxq *rxq;
608 	struct sge_txq *txq;
609 #ifdef TCP_OFFLOAD_ENABLE
610 	struct adapter *sc = pi->adapter;
611 	struct sge_ofld_rxq *ofld_rxq;
612 	struct sge_wrq *ofld_txq;
613 #endif
614 
615 	if (pi->ksp_config != NULL) {
616 		kstat_delete(pi->ksp_config);
617 		pi->ksp_config = NULL;
618 	}
619 	if (pi->ksp_info != NULL) {
620 		kstat_delete(pi->ksp_info);
621 		pi->ksp_info = NULL;
622 	}
623 
624 #ifdef TCP_OFFLOAD_ENABLE
625 	(void) free_wrq(sc, &sc->sge.ctrlq[pi->port_id]);
626 #endif
627 
628 	for_each_txq(pi, i, txq) {
629 		(void) free_txq(pi, txq);
630 	}
631 
632 #ifdef TCP_OFFLOAD_ENABLE
633 	for_each_ofld_txq(pi, i, ofld_txq) {
634 		(void) free_wrq(sc, ofld_txq);
635 	}
636 
637 	for_each_ofld_rxq(pi, i, ofld_rxq) {
638 		if ((ofld_rxq->iq.flags & IQ_INTR) == 0)
639 			(void) free_ofld_rxq(pi, ofld_rxq);
640 	}
641 #endif
642 
643 	for_each_rxq(pi, i, rxq) {
644 		if ((rxq->iq.flags & IQ_INTR) == 0)
645 			(void) free_rxq(pi, rxq);
646 	}
647 
648 	/*
649 	 * Then take down the rx queues that take direct interrupts.
650 	 */
651 
652 	for_each_rxq(pi, i, rxq) {
653 		if (rxq->iq.flags & IQ_INTR)
654 			(void) free_rxq(pi, rxq);
655 	}
656 
657 #ifdef TCP_OFFLOAD_ENABLE
658 	for_each_ofld_rxq(pi, i, ofld_rxq) {
659 		if (ofld_rxq->iq.flags & IQ_INTR)
660 			(void) free_ofld_rxq(pi, ofld_rxq);
661 	}
662 #endif
663 
664 	return (0);
665 }
666 
667 /* Deals with errors and forwarded interrupts */
668 uint_t
669 t4_intr_all(caddr_t arg1, caddr_t arg2)
670 {
671 
672 	(void) t4_intr_err(arg1, arg2);
673 	(void) t4_intr(arg1, arg2);
674 
675 	return (DDI_INTR_CLAIMED);
676 }
677 
678 static void
679 t4_intr_rx_work(struct sge_iq *iq)
680 {
681 	mblk_t *mp = NULL;
682 	struct sge_rxq *rxq = iq_to_rxq(iq);	/* Use iff iq is part of rxq */
683 	RXQ_LOCK(rxq);
684 	if (!iq->polling) {
685 		mp = t4_ring_rx(rxq, iq->qsize/8);
686 		t4_write_reg(iq->adapter, MYPF_REG(A_SGE_PF_GTS),
687 		     V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_next));
688 	}
689 	RXQ_UNLOCK(rxq);
690 	if (mp != NULL)
691 		mac_rx_ring(rxq->port->mh, rxq->ring_handle, mp,
692 			    rxq->ring_gen_num);
693 }
694 
695 /* Deals with interrupts on the given ingress queue */
696 /* ARGSUSED */
697 uint_t
698 t4_intr(caddr_t arg1, caddr_t arg2)
699 {
700 	struct sge_iq *iq = (struct sge_iq *)arg2;
701 	int state;
702 
703 	/* Right now receive polling is only enabled for MSI-X and
704 	 * when we have enough msi-x vectors i.e no interrupt forwarding.
705 	 */
706 	if (iq->adapter->props.multi_rings) {
707 		t4_intr_rx_work(iq);
708 	} else {
709 		state = atomic_cas_uint(&iq->state, IQS_IDLE, IQS_BUSY);
710 		if (state == IQS_IDLE) {
711 			(void) service_iq(iq, 0);
712 			(void) atomic_cas_uint(&iq->state, IQS_BUSY, IQS_IDLE);
713 		}
714 	}
715 	return (DDI_INTR_CLAIMED);
716 }
717 
718 /* Deals with error interrupts */
719 /* ARGSUSED */
720 uint_t
721 t4_intr_err(caddr_t arg1, caddr_t arg2)
722 {
723 	/* LINTED: E_BAD_PTR_CAST_ALIGN */
724 	struct adapter *sc = (struct adapter *)arg1;
725 
726 	t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
727 	(void) t4_slow_intr_handler(sc);
728 
729 	return (DDI_INTR_CLAIMED);
730 }
731 
732 /*
733  * t4_ring_rx - Process responses from an SGE response queue.
734  *
735  * This function processes responses from an SGE response queue up to the supplied budget.
736  * Responses include received packets as well as control messages from FW
737  * or HW.
738  * It returns a chain of mblks containing the received data, to be
739  * passed up to mac_ring_rx().
740  */
741 mblk_t *
742 t4_ring_rx(struct sge_rxq *rxq, int budget)
743 {
744 	struct sge_iq *iq = &rxq->iq;
745 	struct sge_fl *fl = &rxq->fl;           /* Use iff IQ_HAS_FL */
746 	struct adapter *sc = iq->adapter;
747 	struct rsp_ctrl *ctrl;
748 	const struct rss_header *rss;
749 	int ndescs = 0, fl_bufs_used = 0;
750 	int rsp_type;
751 	uint32_t lq;
752 	mblk_t *mblk_head = NULL, **mblk_tail, *m;
753 	struct cpl_rx_pkt *cpl;
754 	uint32_t received_bytes = 0, pkt_len = 0;
755 	bool csum_ok;
756 	uint16_t err_vec;
757 
758 	mblk_tail = &mblk_head;
759 
760 	while (is_new_response(iq, &ctrl)) {
761 
762 		membar_consumer();
763 
764 		m = NULL;
765 		rsp_type = G_RSPD_TYPE(ctrl->u.type_gen);
766 		lq = be32_to_cpu(ctrl->pldbuflen_qid);
767 		rss = (const void *)iq->cdesc;
768 
769 		switch (rsp_type) {
770 		case X_RSPD_TYPE_FLBUF:
771 
772 			ASSERT(iq->flags & IQ_HAS_FL);
773 
774 			if (CPL_RX_PKT == rss->opcode) {
775 				cpl = (void *)(rss + 1);
776 				pkt_len = be16_to_cpu(cpl->len);
777 
778 				if (iq->polling && ((received_bytes + pkt_len) > budget))
779 					goto done;
780 
781 				m = get_fl_payload(sc, fl, lq, &fl_bufs_used);
782 				if (m == NULL)
783 					goto done;
784 
785 				iq->intr_next = iq->intr_params;
786 				m->b_rptr += sc->sge.pktshift;
787 				if (sc->params.tp.rx_pkt_encap)
788 				/* It is enabled only in T6 config file */
789 					err_vec = G_T6_COMPR_RXERR_VEC(ntohs(cpl->err_vec));
790 				else
791 					err_vec = ntohs(cpl->err_vec);
792 
793 				csum_ok = cpl->csum_calc && !err_vec;
794 
795 				/* TODO: what about cpl->ip_frag? */
796 				if (csum_ok && !cpl->ip_frag) {
797 					mac_hcksum_set(m, 0, 0, 0, 0xffff,
798 					    HCK_FULLCKSUM_OK | HCK_FULLCKSUM |
799 					    HCK_IPV4_HDRCKSUM_OK);
800 					rxq->rxcsum++;
801 				}
802 				rxq->rxpkts++;
803 				rxq->rxbytes += pkt_len;
804 				received_bytes += pkt_len;
805 
806 				*mblk_tail = m;
807 				mblk_tail = &m->b_next;
808 
809 				break;
810 			}
811 
812 			m = get_fl_payload(sc, fl, lq, &fl_bufs_used);
813 			if (m == NULL)
814 				goto done;
815 			/* FALLTHROUGH */
816 
817 		case X_RSPD_TYPE_CPL:
818 			ASSERT(rss->opcode < NUM_CPL_CMDS);
819 			sc->cpl_handler[rss->opcode](iq, rss, m);
820 			break;
821 
822 		default:
823 			break;
824 		}
825 		iq_next(iq);
826 		++ndescs;
827 		if (!iq->polling && (ndescs == budget))
828 			break;
829 	}
830 
831 done:
832 
833 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
834 		     V_CIDXINC(ndescs) | V_INGRESSQID(iq->cntxt_id) |
835 		     V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
836 
837 	if ((fl_bufs_used > 0) || (iq->flags & IQ_HAS_FL)) {
838 		int starved;
839 		FL_LOCK(fl);
840 		fl->needed += fl_bufs_used;
841 		starved = refill_fl(sc, fl, fl->cap / 8);
842 		FL_UNLOCK(fl);
843 		if (starved)
844 			add_fl_to_sfl(sc, fl);
845 	}
846 	return (mblk_head);
847 }
848 
849 /*
850  * Deals with anything and everything on the given ingress queue.
851  */
852 static int
853 service_iq(struct sge_iq *iq, int budget)
854 {
855 	struct sge_iq *q;
856 	struct sge_rxq *rxq = iq_to_rxq(iq);	/* Use iff iq is part of rxq */
857 	struct sge_fl *fl = &rxq->fl;		/* Use iff IQ_HAS_FL */
858 	struct adapter *sc = iq->adapter;
859 	struct rsp_ctrl *ctrl;
860 	const struct rss_header *rss;
861 	int ndescs = 0, limit, fl_bufs_used = 0;
862 	int rsp_type;
863 	uint32_t lq;
864 	int starved;
865 	mblk_t *m;
866 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
867 
868 	limit = budget ? budget : iq->qsize / 8;
869 
870 	/*
871 	 * We always come back and check the descriptor ring for new indirect
872 	 * interrupts and other responses after running a single handler.
873 	 */
874 	for (;;) {
875 		while (is_new_response(iq, &ctrl)) {
876 
877 			membar_consumer();
878 
879 			m = NULL;
880 			rsp_type = G_RSPD_TYPE(ctrl->u.type_gen);
881 			lq = be32_to_cpu(ctrl->pldbuflen_qid);
882 			rss = (const void *)iq->cdesc;
883 
884 			switch (rsp_type) {
885 			case X_RSPD_TYPE_FLBUF:
886 
887 				ASSERT(iq->flags & IQ_HAS_FL);
888 
889 				m = get_fl_payload(sc, fl, lq, &fl_bufs_used);
890 				if (m == NULL) {
891 					/*
892 					 * Rearm the iq with a
893 					 * longer-than-default timer
894 					 */
895 					t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_CIDXINC(ndescs) |
896 							V_INGRESSQID((u32)iq->cntxt_id) |
897 							V_SEINTARM(V_QINTR_TIMER_IDX(SGE_NTIMERS-1)));
898 					if (fl_bufs_used > 0) {
899 						ASSERT(iq->flags & IQ_HAS_FL);
900 						FL_LOCK(fl);
901 						fl->needed += fl_bufs_used;
902 						starved = refill_fl(sc, fl, fl->cap / 8);
903 						FL_UNLOCK(fl);
904 						if (starved)
905 							add_fl_to_sfl(sc, fl);
906 					}
907 					return (0);
908 				}
909 
910 			/* FALLTHRU */
911 			case X_RSPD_TYPE_CPL:
912 
913 				ASSERT(rss->opcode < NUM_CPL_CMDS);
914 				sc->cpl_handler[rss->opcode](iq, rss, m);
915 				break;
916 
917 			case X_RSPD_TYPE_INTR:
918 
919 				/*
920 				 * Interrupts should be forwarded only to queues
921 				 * that are not forwarding their interrupts.
922 				 * This means service_iq can recurse but only 1
923 				 * level deep.
924 				 */
925 				ASSERT(budget == 0);
926 
927 				q = sc->sge.iqmap[lq - sc->sge.iq_start];
928 				if (atomic_cas_uint(&q->state, IQS_IDLE,
929 				    IQS_BUSY) == IQS_IDLE) {
930 					if (service_iq(q, q->qsize / 8) == 0) {
931 						(void) atomic_cas_uint(
932 						    &q->state, IQS_BUSY,
933 						    IQS_IDLE);
934 					} else {
935 						STAILQ_INSERT_TAIL(&iql, q,
936 						    link);
937 					}
938 				}
939 				break;
940 
941 			default:
942 				break;
943 			}
944 
945 			iq_next(iq);
946 			if (++ndescs == limit) {
947 				t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
948 				    V_CIDXINC(ndescs) |
949 				    V_INGRESSQID(iq->cntxt_id) |
950 				    V_SEINTARM(V_QINTR_TIMER_IDX(
951 				    X_TIMERREG_UPDATE_CIDX)));
952 				ndescs = 0;
953 
954 				if (fl_bufs_used > 0) {
955 					ASSERT(iq->flags & IQ_HAS_FL);
956 					FL_LOCK(fl);
957 					fl->needed += fl_bufs_used;
958 					(void) refill_fl(sc, fl, fl->cap / 8);
959 					FL_UNLOCK(fl);
960 					fl_bufs_used = 0;
961 				}
962 
963 				if (budget != 0)
964 					return (EINPROGRESS);
965 			}
966 		}
967 
968 		if (STAILQ_EMPTY(&iql) != 0)
969 			break;
970 
971 		/*
972 		 * Process the head only, and send it to the back of the list if
973 		 * it's still not done.
974 		 */
975 		q = STAILQ_FIRST(&iql);
976 		STAILQ_REMOVE_HEAD(&iql, link);
977 		if (service_iq(q, q->qsize / 8) == 0)
978 			(void) atomic_cas_uint(&q->state, IQS_BUSY, IQS_IDLE);
979 		else
980 			STAILQ_INSERT_TAIL(&iql, q, link);
981 	}
982 
983 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_CIDXINC(ndescs) |
984 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_next));
985 
986 	if (iq->flags & IQ_HAS_FL) {
987 
988 		FL_LOCK(fl);
989 		fl->needed += fl_bufs_used;
990 		starved = refill_fl(sc, fl, fl->cap / 4);
991 		FL_UNLOCK(fl);
992 		if (starved != 0)
993 			add_fl_to_sfl(sc, fl);
994 	}
995 
996 	return (0);
997 }
998 
999 #ifdef TCP_OFFLOAD_ENABLE
1000 int
1001 t4_mgmt_tx(struct adapter *sc, mblk_t *m)
1002 {
1003 	return (t4_wrq_tx(sc, &sc->sge.mgmtq, m));
1004 }
1005 
1006 /*
1007  * Doesn't fail.  Holds on to work requests it can't send right away.
1008  */
1009 int
1010 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, mblk_t *m0)
1011 {
1012 	struct sge_eq *eq = &wrq->eq;
1013 	struct mblk_pair *wr_list = &wrq->wr_list;
1014 	int can_reclaim;
1015 	caddr_t dst;
1016 	mblk_t *wr, *next;
1017 
1018 	TXQ_LOCK_ASSERT_OWNED(wrq);
1019 #ifdef TCP_OFFLOAD_ENABLE
1020 	ASSERT((eq->flags & EQ_TYPEMASK) == EQ_OFLD ||
1021 	    (eq->flags & EQ_TYPEMASK) == EQ_CTRL);
1022 #else
1023 	ASSERT((eq->flags & EQ_TYPEMASK) == EQ_CTRL);
1024 #endif
1025 
1026 	if (m0 != NULL) {
1027 		if (wr_list->head != NULL)
1028 			wr_list->tail->b_next = m0;
1029 		else
1030 			wr_list->head = m0;
1031 		while (m0->b_next)
1032 			m0 = m0->b_next;
1033 		wr_list->tail = m0;
1034 	}
1035 
1036 	can_reclaim = reclaimable(eq);
1037 	eq->cidx += can_reclaim;
1038 	eq->avail += can_reclaim;
1039 	if (eq->cidx >= eq->cap)
1040 		eq->cidx -= eq->cap;
1041 
1042 	for (wr = wr_list->head; wr; wr = next) {
1043 		int ndesc, len = 0;
1044 		mblk_t *m;
1045 
1046 		next = wr->b_next;
1047 		wr->b_next = NULL;
1048 
1049 		for (m = wr; m; m = m->b_cont)
1050 			len += MBLKL(m);
1051 
1052 		ASSERT(len > 0 && (len & 0x7) == 0);
1053 		ASSERT(len <= SGE_MAX_WR_LEN);
1054 
1055 		ndesc = howmany(len, EQ_ESIZE);
1056 		if (eq->avail < ndesc) {
1057 			wr->b_next = next;
1058 			wrq->no_desc++;
1059 			break;
1060 		}
1061 
1062 		dst = (void *)&eq->desc[eq->pidx];
1063 		for (m = wr; m; m = m->b_cont)
1064 			copy_to_txd(eq, (void *)m->b_rptr, &dst, MBLKL(m));
1065 
1066 		eq->pidx += ndesc;
1067 		eq->avail -= ndesc;
1068 		if (eq->pidx >= eq->cap)
1069 			eq->pidx -= eq->cap;
1070 
1071 		eq->pending += ndesc;
1072 		if (eq->pending > 16)
1073 			ring_tx_db(sc, eq);
1074 
1075 		wrq->tx_wrs++;
1076 		freemsg(wr);
1077 
1078 		if (eq->avail < 8) {
1079 			can_reclaim = reclaimable(eq);
1080 			eq->cidx += can_reclaim;
1081 			eq->avail += can_reclaim;
1082 			if (eq->cidx >= eq->cap)
1083 				eq->cidx -= eq->cap;
1084 		}
1085 	}
1086 
1087 	if (eq->pending != 0)
1088 		ring_tx_db(sc, eq);
1089 
1090 	if (wr == NULL)
1091 		wr_list->head = wr_list->tail = NULL;
1092 	else {
1093 		wr_list->head = wr;
1094 
1095 		ASSERT(wr_list->tail->b_next == NULL);
1096 	}
1097 
1098 	return (0);
1099 }
1100 #endif
1101 
1102 /* Per-packet header in a coalesced tx WR, before the SGL starts (in flits) */
1103 #define	TXPKTS_PKT_HDR ((\
1104 	sizeof (struct ulp_txpkt) + \
1105 	sizeof (struct ulptx_idata) + \
1106 	sizeof (struct cpl_tx_pkt_core)) / 8)
1107 
1108 /* Header of a coalesced tx WR, before SGL of first packet (in flits) */
1109 #define	TXPKTS_WR_HDR (\
1110 	sizeof (struct fw_eth_tx_pkts_wr) / 8 + \
1111 	TXPKTS_PKT_HDR)
1112 
1113 /* Header of a tx WR, before SGL of first packet (in flits) */
1114 #define	TXPKT_WR_HDR ((\
1115 	sizeof (struct fw_eth_tx_pkt_wr) + \
1116 	sizeof (struct cpl_tx_pkt_core)) / 8)
1117 
1118 /* Header of a tx LSO WR, before SGL of first packet (in flits) */
1119 #define	TXPKT_LSO_WR_HDR ((\
1120 	sizeof (struct fw_eth_tx_pkt_wr) + \
1121 	sizeof(struct cpl_tx_pkt_lso_core) + \
1122 	sizeof (struct cpl_tx_pkt_core)) / 8)
1123 
1124 mblk_t *
1125 t4_eth_tx(void *arg, mblk_t *frame)
1126 {
1127 	struct sge_txq *txq = (struct sge_txq *) arg;
1128 	struct port_info *pi = txq->port;
1129 	struct adapter *sc = pi->adapter;
1130 	struct sge_eq *eq = &txq->eq;
1131 	mblk_t *next_frame;
1132 	int rc, coalescing;
1133 	struct txpkts txpkts;
1134 	struct txinfo txinfo;
1135 
1136 	txpkts.npkt = 0; /* indicates there's nothing in txpkts */
1137 	coalescing = 0;
1138 
1139 	TXQ_LOCK(txq);
1140 	if (eq->avail < 8)
1141 		(void) reclaim_tx_descs(txq, 8);
1142 	for (; frame; frame = next_frame) {
1143 
1144 		if (eq->avail < 8)
1145 			break;
1146 
1147 		next_frame = frame->b_next;
1148 		frame->b_next = NULL;
1149 
1150 		if (next_frame != NULL)
1151 			coalescing = 1;
1152 
1153 		rc = get_frame_txinfo(txq, &frame, &txinfo, coalescing);
1154 		if (rc != 0) {
1155 			if (rc == ENOMEM) {
1156 
1157 				/* Short of resources, suspend tx */
1158 
1159 				frame->b_next = next_frame;
1160 				break;
1161 			}
1162 
1163 			/*
1164 			 * Unrecoverable error for this frame, throw it
1165 			 * away and move on to the next.
1166 			 */
1167 
1168 			freemsg(frame);
1169 			continue;
1170 		}
1171 
1172 		if (coalescing != 0 &&
1173 		    add_to_txpkts(txq, &txpkts, frame, &txinfo) == 0) {
1174 
1175 			/* Successfully absorbed into txpkts */
1176 
1177 			write_ulp_cpl_sgl(pi, txq, &txpkts, &txinfo);
1178 			goto doorbell;
1179 		}
1180 
1181 		/*
1182 		 * We weren't coalescing to begin with, or current frame could
1183 		 * not be coalesced (add_to_txpkts flushes txpkts if a frame
1184 		 * given to it can't be coalesced).  Either way there should be
1185 		 * nothing in txpkts.
1186 		 */
1187 		ASSERT(txpkts.npkt == 0);
1188 
1189 		/* We're sending out individual frames now */
1190 		coalescing = 0;
1191 
1192 		if (eq->avail < 8)
1193 			(void) reclaim_tx_descs(txq, 8);
1194 		rc = write_txpkt_wr(pi, txq, frame, &txinfo);
1195 		if (rc != 0) {
1196 
1197 			/* Short of hardware descriptors, suspend tx */
1198 
1199 			/*
1200 			 * This is an unlikely but expensive failure.  We've
1201 			 * done all the hard work (DMA bindings etc.) and now we
1202 			 * can't send out the frame.  What's worse, we have to
1203 			 * spend even more time freeing up everything in txinfo.
1204 			 */
1205 			txq->qfull++;
1206 			free_txinfo_resources(txq, &txinfo);
1207 
1208 			frame->b_next = next_frame;
1209 			break;
1210 		}
1211 
1212 doorbell:
1213 		/* Fewer and fewer doorbells as the queue fills up */
1214 		if (eq->pending >= (1 << (fls(eq->qsize - eq->avail) / 2))) {
1215 			txq->txbytes += txinfo.len;
1216 			txq->txpkts++;
1217 			ring_tx_db(sc, eq);
1218 		}
1219 		(void) reclaim_tx_descs(txq, 32);
1220 	}
1221 
1222 	if (txpkts.npkt > 0)
1223 		write_txpkts_wr(txq, &txpkts);
1224 
1225 	/*
1226 	 * frame not NULL means there was an error but we haven't thrown it
1227 	 * away.  This can happen when we're short of tx descriptors (qfull) or
1228 	 * maybe even DMA handles (dma_hdl_failed).  Either way, a credit flush
1229 	 * and reclaim will get things going again.
1230 	 *
1231 	 * If eq->avail is already 0 we know a credit flush was requested in the
1232 	 * WR that reduced it to 0 so we don't need another flush (we don't have
1233 	 * any descriptor for a flush WR anyway, duh).
1234 	 */
1235 	if (frame && eq->avail > 0)
1236 		write_txqflush_wr(txq);
1237 
1238 	if (eq->pending != 0)
1239 		ring_tx_db(sc, eq);
1240 
1241 	(void) reclaim_tx_descs(txq, eq->qsize);
1242 	TXQ_UNLOCK(txq);
1243 
1244 	return (frame);
1245 }
1246 
1247 static inline void
1248 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int8_t pktc_idx,
1249 	int qsize, uint8_t esize)
1250 {
1251 	ASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS);
1252 	ASSERT(pktc_idx < SGE_NCOUNTERS);	/* -ve is ok, means don't use */
1253 
1254 	iq->flags = 0;
1255 	iq->adapter = sc;
1256 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
1257 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
1258 	if (pktc_idx >= 0) {
1259 		iq->intr_params |= F_QINTR_CNT_EN;
1260 		iq->intr_pktc_idx = pktc_idx;
1261 	}
1262 	iq->qsize = roundup(qsize, 16);		/* See FW_IQ_CMD/iqsize */
1263 	iq->esize = max(esize, 16);		/* See FW_IQ_CMD/iqesize */
1264 }
1265 
1266 static inline void
1267 init_fl(struct sge_fl *fl, uint16_t qsize)
1268 {
1269 
1270 	fl->qsize = qsize;
1271 	fl->allocb_fail = 0;
1272 }
1273 
1274 static inline void
1275 init_eq(struct adapter *sc, struct sge_eq *eq, uint16_t eqtype, uint16_t qsize,
1276     uint8_t tx_chan, uint16_t iqid)
1277 {
1278 	struct sge *s = &sc->sge;
1279 	uint32_t r;
1280 
1281 	ASSERT(tx_chan < NCHAN);
1282 	ASSERT(eqtype <= EQ_TYPEMASK);
1283 
1284 	if (is_t5(sc->params.chip)) {
1285 		r = t4_read_reg(sc, A_SGE_EGRESS_QUEUES_PER_PAGE_PF);
1286 		r >>= S_QUEUESPERPAGEPF0 +
1287 		    (S_QUEUESPERPAGEPF1 - S_QUEUESPERPAGEPF0) * sc->pf;
1288 		s->s_qpp = r & M_QUEUESPERPAGEPF0;
1289 	}
1290 
1291 	eq->flags = eqtype & EQ_TYPEMASK;
1292 	eq->tx_chan = tx_chan;
1293 	eq->iqid = iqid;
1294 	eq->qsize = qsize;
1295 }
1296 
1297 /*
1298  * Allocates the ring for an ingress queue and an optional freelist.  If the
1299  * freelist is specified it will be allocated and then associated with the
1300  * ingress queue.
1301  *
1302  * Returns errno on failure.  Resources allocated up to that point may still be
1303  * allocated.  Caller is responsible for cleanup in case this function fails.
1304  *
1305  * If the ingress queue will take interrupts directly (iq->flags & IQ_INTR) then
1306  * the intr_idx specifies the vector, starting from 0.  Otherwise it specifies
1307  * the index of the queue to which its interrupts will be forwarded.
1308  */
1309 static int
1310 alloc_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl,
1311     int intr_idx, int cong)
1312 {
1313 	int rc, i, cntxt_id;
1314 	size_t len;
1315 	struct fw_iq_cmd c;
1316 	struct adapter *sc = iq->adapter;
1317 	uint32_t v = 0;
1318 
1319 	len = iq->qsize * iq->esize;
1320 	rc = alloc_desc_ring(sc, len, DDI_DMA_READ, &iq->dhdl, &iq->ahdl,
1321 	    &iq->ba, (caddr_t *)&iq->desc);
1322 	if (rc != 0)
1323 		return (rc);
1324 
1325 	bzero(&c, sizeof (c));
1326 	c.op_to_vfn = cpu_to_be32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
1327 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
1328 	    V_FW_IQ_CMD_VFN(0));
1329 
1330 	c.alloc_to_len16 = cpu_to_be32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
1331 	    FW_LEN16(c));
1332 
1333 	/* Special handling for firmware event queue */
1334 	if (iq == &sc->sge.fwq)
1335 		v |= F_FW_IQ_CMD_IQASYNCH;
1336 
1337 	if (iq->flags & IQ_INTR)
1338 		ASSERT(intr_idx < sc->intr_count);
1339 	else
1340 		v |= F_FW_IQ_CMD_IQANDST;
1341 	v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
1342 
1343 	c.type_to_iqandstindex = cpu_to_be32(v |
1344 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
1345 	    V_FW_IQ_CMD_VIID(pi->viid) |
1346 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
1347 	c.iqdroprss_to_iqesize = cpu_to_be16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
1348 	    F_FW_IQ_CMD_IQGTSMODE |
1349 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
1350 	    V_FW_IQ_CMD_IQESIZE(ilog2(iq->esize) - 4));
1351 	c.iqsize = cpu_to_be16(iq->qsize);
1352 	c.iqaddr = cpu_to_be64(iq->ba);
1353 	if (cong >= 0)
1354 		c.iqns_to_fl0congen = BE_32(F_FW_IQ_CMD_IQFLINTCONGEN |
1355 					V_FW_IQ_CMD_IQTYPE(cong ?
1356 					FW_IQ_IQTYPE_NIC : FW_IQ_IQTYPE_OFLD));
1357 
1358 	if (fl != NULL) {
1359 		unsigned int chip_ver = CHELSIO_CHIP_VERSION(sc->params.chip);
1360 
1361 		mutex_init(&fl->lock, NULL, MUTEX_DRIVER,
1362 		    DDI_INTR_PRI(sc->intr_pri));
1363 		fl->flags |= FL_MTX;
1364 
1365 		len = fl->qsize * RX_FL_ESIZE;
1366 		rc = alloc_desc_ring(sc, len, DDI_DMA_WRITE, &fl->dhdl,
1367 		    &fl->ahdl, &fl->ba, (caddr_t *)&fl->desc);
1368 		if (rc != 0)
1369 			return (rc);
1370 
1371 		/* Allocate space for one software descriptor per buffer. */
1372 		fl->cap = (fl->qsize - sc->sge.stat_len / RX_FL_ESIZE) * 8;
1373 		fl->sdesc = kmem_zalloc(sizeof (struct fl_sdesc) * fl->cap,
1374 		    KM_SLEEP);
1375 		fl->needed = fl->cap;
1376 		fl->lowat = roundup(sc->sge.fl_starve_threshold, 8);
1377 
1378 		c.iqns_to_fl0congen |=
1379 		    cpu_to_be32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
1380 		    F_FW_IQ_CMD_FL0PACKEN | F_FW_IQ_CMD_FL0PADEN);
1381 		if (cong >= 0) {
1382 			c.iqns_to_fl0congen |=
1383 			    BE_32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
1384 			    F_FW_IQ_CMD_FL0CONGCIF |
1385 			    F_FW_IQ_CMD_FL0CONGEN);
1386 		}
1387 
1388 		/* In T6, for egress queue type FL there is internal overhead
1389 		 * of 16B for header going into FLM module.  Hence the maximum
1390 		 * allowed burst size is 448 bytes.  For T4/T5, the hardware
1391 		 * doesn't coalesce fetch requests if more than 64 bytes of
1392 		 * Free List pointers are provided, so we use a 128-byte Fetch
1393 		 * Burst Minimum there (T6 implements coalescing so we can use
1394 		 * the smaller 64-byte value there).
1395 		 */
1396 
1397 		c.fl0dcaen_to_fl0cidxfthresh =
1398 		    cpu_to_be16(V_FW_IQ_CMD_FL0FBMIN(chip_ver <= CHELSIO_T5
1399 						     ? X_FETCHBURSTMIN_128B
1400 						     : X_FETCHBURSTMIN_64B) |
1401 		    V_FW_IQ_CMD_FL0FBMAX(chip_ver <= CHELSIO_T5
1402 					 ? X_FETCHBURSTMAX_512B
1403 					 : X_FETCHBURSTMAX_256B));
1404 		c.fl0size = cpu_to_be16(fl->qsize);
1405 		c.fl0addr = cpu_to_be64(fl->ba);
1406 	}
1407 
1408 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof (c), &c);
1409 	if (rc != 0) {
1410 		cxgb_printf(sc->dip, CE_WARN,
1411 		    "failed to create ingress queue: %d", rc);
1412 		return (rc);
1413 	}
1414 
1415 	iq->cdesc = iq->desc;
1416 	iq->cidx = 0;
1417 	iq->gen = 1;
1418 	iq->intr_next = iq->intr_params;
1419 	iq->adapter = sc;
1420 	iq->cntxt_id = be16_to_cpu(c.iqid);
1421 	iq->abs_id = be16_to_cpu(c.physiqid);
1422 	iq->flags |= IQ_ALLOCATED;
1423 	mutex_init(&iq->lock, NULL,
1424 		    MUTEX_DRIVER, DDI_INTR_PRI(DDI_INTR_PRI(sc->intr_pri)));
1425 	iq->polling = 0;
1426 
1427 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
1428 	if (cntxt_id >= sc->sge.iqmap_sz) {
1429 		panic("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
1430 		      cntxt_id, sc->sge.iqmap_sz - 1);
1431 	}
1432 	sc->sge.iqmap[cntxt_id] = iq;
1433 
1434 	if (fl != NULL) {
1435 		fl->cntxt_id = be16_to_cpu(c.fl0id);
1436 		fl->pidx = fl->cidx = 0;
1437 		fl->copy_threshold = rx_copy_threshold;
1438 
1439 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
1440 		if (cntxt_id >= sc->sge.eqmap_sz) {
1441 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
1442 			      __func__, cntxt_id, sc->sge.eqmap_sz - 1);
1443 		}
1444 		sc->sge.eqmap[cntxt_id] = (void *)fl;
1445 
1446 		FL_LOCK(fl);
1447 		(void) refill_fl(sc, fl, fl->lowat);
1448 		FL_UNLOCK(fl);
1449 
1450 		iq->flags |= IQ_HAS_FL;
1451 	}
1452 
1453 	if (is_t5(sc->params.chip) && cong >= 0) {
1454 		uint32_t param, val;
1455 
1456 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
1457 			V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
1458 			V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
1459 		if (cong == 0)
1460 			val = 1 << 19;
1461 		else {
1462 			val = 2 << 19;
1463 			for (i = 0; i < 4; i++) {
1464 				if (cong & (1 << i))
1465 					val |= 1 << (i << 2);
1466 			}
1467 		}
1468 
1469 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
1470 		if (rc != 0) {
1471 			/* report error but carry on */
1472 			cxgb_printf(sc->dip, CE_WARN,
1473 			    "failed to set congestion manager context for "
1474 			    "ingress queue %d: %d", iq->cntxt_id, rc);
1475 		}
1476 	}
1477 
1478 	/* Enable IQ interrupts */
1479 	iq->state = IQS_IDLE;
1480 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_SEINTARM(iq->intr_params) |
1481 	    V_INGRESSQID(iq->cntxt_id));
1482 
1483 	return (0);
1484 }
1485 
1486 static int
1487 free_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl)
1488 {
1489 	int rc;
1490 
1491 	if (iq != NULL) {
1492 		struct adapter *sc = iq->adapter;
1493 		dev_info_t *dip;
1494 
1495 		dip = pi ? pi->dip : sc->dip;
1496 		if (iq->flags & IQ_ALLOCATED) {
1497 			rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
1498 			    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
1499 			    fl ? fl->cntxt_id : 0xffff, 0xffff);
1500 			if (rc != 0) {
1501 				cxgb_printf(dip, CE_WARN,
1502 				    "failed to free queue %p: %d", iq, rc);
1503 				return (rc);
1504 			}
1505 			mutex_destroy(&iq->lock);
1506 			iq->flags &= ~IQ_ALLOCATED;
1507 		}
1508 
1509 		if (iq->desc != NULL) {
1510 			(void) free_desc_ring(&iq->dhdl, &iq->ahdl);
1511 			iq->desc = NULL;
1512 		}
1513 
1514 		bzero(iq, sizeof (*iq));
1515 	}
1516 
1517 	if (fl != NULL) {
1518 		if (fl->sdesc != NULL) {
1519 			FL_LOCK(fl);
1520 			free_fl_bufs(fl);
1521 			FL_UNLOCK(fl);
1522 
1523 			kmem_free(fl->sdesc, sizeof (struct fl_sdesc) *
1524 			    fl->cap);
1525 			fl->sdesc = NULL;
1526 		}
1527 
1528 		if (fl->desc != NULL) {
1529 			(void) free_desc_ring(&fl->dhdl, &fl->ahdl);
1530 			fl->desc = NULL;
1531 		}
1532 
1533 		if (fl->flags & FL_MTX) {
1534 			mutex_destroy(&fl->lock);
1535 			fl->flags &= ~FL_MTX;
1536 		}
1537 
1538 		bzero(fl, sizeof (struct sge_fl));
1539 	}
1540 
1541 	return (0);
1542 }
1543 
1544 static int
1545 alloc_fwq(struct adapter *sc)
1546 {
1547 	int rc, intr_idx;
1548 	struct sge_iq *fwq = &sc->sge.fwq;
1549 
1550 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE, FW_IQ_ESIZE);
1551 	fwq->flags |= IQ_INTR;	/* always */
1552 	intr_idx = sc->intr_count > 1 ? 1 : 0;
1553 	rc = alloc_iq_fl(sc->port[0], fwq, NULL, intr_idx, -1);
1554 	if (rc != 0) {
1555 		cxgb_printf(sc->dip, CE_WARN,
1556 		    "failed to create firmware event queue: %d.", rc);
1557 		return (rc);
1558 	}
1559 
1560 	return (0);
1561 }
1562 
1563 static int
1564 free_fwq(struct adapter *sc)
1565 {
1566 
1567 	return (free_iq_fl(NULL, &sc->sge.fwq, NULL));
1568 }
1569 
1570 #ifdef TCP_OFFLOAD_ENABLE
1571 static int
1572 alloc_mgmtq(struct adapter *sc)
1573 {
1574 	int rc;
1575 	struct sge_wrq *mgmtq = &sc->sge.mgmtq;
1576 
1577 	init_eq(sc, &mgmtq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[0]->tx_chan,
1578 	    sc->sge.fwq.cntxt_id);
1579 	rc = alloc_wrq(sc, NULL, mgmtq, 0);
1580 	if (rc != 0) {
1581 		cxgb_printf(sc->dip, CE_WARN,
1582 		    "failed to create management queue: %d\n", rc);
1583 		return (rc);
1584 	}
1585 
1586 	return (0);
1587 }
1588 #endif
1589 
1590 static int
1591 alloc_rxq(struct port_info *pi, struct sge_rxq *rxq, int intr_idx, int i)
1592 {
1593 	int rc;
1594 
1595 	rxq->port = pi;
1596 	rc = alloc_iq_fl(pi, &rxq->iq, &rxq->fl, intr_idx,
1597 			 t4_get_tp_ch_map(pi->adapter, pi->tx_chan));
1598 	if (rc != 0)
1599 		return (rc);
1600 
1601 	rxq->ksp = setup_rxq_kstats(pi, rxq, i);
1602 
1603 	return (rc);
1604 }
1605 
1606 static int
1607 free_rxq(struct port_info *pi, struct sge_rxq *rxq)
1608 {
1609 	int rc;
1610 
1611 	if (rxq->ksp != NULL) {
1612 		kstat_delete(rxq->ksp);
1613 		rxq->ksp = NULL;
1614 	}
1615 
1616 	rc = free_iq_fl(pi, &rxq->iq, &rxq->fl);
1617 	if (rc == 0)
1618 		bzero(&rxq->fl, sizeof (*rxq) - offsetof(struct sge_rxq, fl));
1619 
1620 	return (rc);
1621 }
1622 
1623 #ifdef TCP_OFFLOAD_ENABLE
1624 static int
1625 alloc_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq,
1626 	int intr_idx)
1627 {
1628 	int rc;
1629 
1630 	rc = alloc_iq_fl(pi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx,
1631 	    t4_get_tp_ch_map(pi->adapter, pi->tx_chan));
1632 	if (rc != 0)
1633 		return (rc);
1634 
1635 	return (rc);
1636 }
1637 
1638 static int
1639 free_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq)
1640 {
1641 	int rc;
1642 
1643 	rc = free_iq_fl(pi, &ofld_rxq->iq, &ofld_rxq->fl);
1644 	if (rc == 0)
1645 		bzero(&ofld_rxq->fl, sizeof (*ofld_rxq) -
1646 		    offsetof(struct sge_ofld_rxq, fl));
1647 
1648 	return (rc);
1649 }
1650 #endif
1651 
1652 static int
1653 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
1654 {
1655 	int rc, cntxt_id;
1656 	struct fw_eq_ctrl_cmd c;
1657 
1658 	bzero(&c, sizeof (c));
1659 
1660 	c.op_to_vfn = BE_32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
1661 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
1662 	    V_FW_EQ_CTRL_CMD_VFN(0));
1663 	c.alloc_to_len16 = BE_32(F_FW_EQ_CTRL_CMD_ALLOC |
1664 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
1665 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid)); /* TODO */
1666 	c.physeqid_pkd = BE_32(0);
1667 	c.fetchszm_to_iqid =
1668 	    BE_32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
1669 	    V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
1670 	    F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
1671 	c.dcaen_to_eqsize =
1672 	    BE_32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
1673 	    V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
1674 	    V_FW_EQ_CTRL_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
1675 	    V_FW_EQ_CTRL_CMD_EQSIZE(eq->qsize));
1676 	c.eqaddr = BE_64(eq->ba);
1677 
1678 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof (c), &c);
1679 	if (rc != 0) {
1680 		cxgb_printf(sc->dip, CE_WARN,
1681 		    "failed to create control queue %d: %d", eq->tx_chan, rc);
1682 		return (rc);
1683 	}
1684 	eq->flags |= EQ_ALLOCATED;
1685 
1686 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(BE_32(c.cmpliqid_eqid));
1687 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
1688 	if (cntxt_id >= sc->sge.eqmap_sz)
1689 		panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
1690 		      cntxt_id, sc->sge.eqmap_sz - 1);
1691 	sc->sge.eqmap[cntxt_id] = eq;
1692 
1693 	return (rc);
1694 }
1695 
1696 static int
1697 eth_eq_alloc(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
1698 {
1699 	int rc, cntxt_id;
1700 	struct fw_eq_eth_cmd c;
1701 
1702 	bzero(&c, sizeof (c));
1703 
1704 	c.op_to_vfn = BE_32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
1705 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
1706 	    V_FW_EQ_ETH_CMD_VFN(0));
1707 	c.alloc_to_len16 = BE_32(F_FW_EQ_ETH_CMD_ALLOC |
1708 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
1709 	c.autoequiqe_to_viid = BE_32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
1710 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(pi->viid));
1711 	c.fetchszm_to_iqid =
1712 	    BE_32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
1713 	    V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
1714 	    V_FW_EQ_ETH_CMD_IQID(eq->iqid));
1715 	c.dcaen_to_eqsize = BE_32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
1716 	    V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
1717 	    V_FW_EQ_ETH_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
1718 	    V_FW_EQ_ETH_CMD_EQSIZE(eq->qsize));
1719 	c.eqaddr = BE_64(eq->ba);
1720 
1721 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof (c), &c);
1722 	if (rc != 0) {
1723 		cxgb_printf(pi->dip, CE_WARN,
1724 		    "failed to create Ethernet egress queue: %d", rc);
1725 		return (rc);
1726 	}
1727 	eq->flags |= EQ_ALLOCATED;
1728 
1729 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(BE_32(c.eqid_pkd));
1730 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
1731 	if (cntxt_id >= sc->sge.eqmap_sz)
1732 		panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
1733 		      cntxt_id, sc->sge.eqmap_sz - 1);
1734 	sc->sge.eqmap[cntxt_id] = eq;
1735 
1736 	return (rc);
1737 }
1738 
1739 #ifdef TCP_OFFLOAD_ENABLE
1740 static int
1741 ofld_eq_alloc(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
1742 {
1743 	int rc, cntxt_id;
1744 	struct fw_eq_ofld_cmd c;
1745 
1746 	bzero(&c, sizeof (c));
1747 
1748 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
1749 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
1750 	    V_FW_EQ_OFLD_CMD_VFN(0));
1751 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
1752 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
1753 	c.fetchszm_to_iqid =
1754 	    htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
1755 	    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
1756 	    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
1757 	c.dcaen_to_eqsize =
1758 	    BE_32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
1759 	    V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
1760 	    V_FW_EQ_OFLD_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
1761 	    V_FW_EQ_OFLD_CMD_EQSIZE(eq->qsize));
1762 	c.eqaddr = BE_64(eq->ba);
1763 
1764 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof (c), &c);
1765 	if (rc != 0) {
1766 		cxgb_printf(pi->dip, CE_WARN,
1767 		    "failed to create egress queue for TCP offload: %d", rc);
1768 		return (rc);
1769 	}
1770 	eq->flags |= EQ_ALLOCATED;
1771 
1772 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(BE_32(c.eqid_pkd));
1773 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
1774 	if (cntxt_id >= sc->sge.eqmap_sz)
1775 		panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
1776 		      cntxt_id, sc->sge.eqmap_sz - 1);
1777 	sc->sge.eqmap[cntxt_id] = eq;
1778 
1779 	return (rc);
1780 }
1781 #endif
1782 
1783 static int
1784 alloc_eq(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
1785 {
1786 	int rc;
1787 	size_t len;
1788 
1789 	mutex_init(&eq->lock, NULL, MUTEX_DRIVER, DDI_INTR_PRI(sc->intr_pri));
1790 	eq->flags |= EQ_MTX;
1791 
1792 	len = eq->qsize * EQ_ESIZE;
1793 	rc = alloc_desc_ring(sc, len, DDI_DMA_WRITE, &eq->desc_dhdl,
1794 	    &eq->desc_ahdl, &eq->ba, (caddr_t *)&eq->desc);
1795 	if (rc != 0)
1796 		return (rc);
1797 
1798 	eq->cap = eq->qsize - sc->sge.stat_len / EQ_ESIZE;
1799 	eq->spg = (void *)&eq->desc[eq->cap];
1800 	eq->avail = eq->cap - 1;	/* one less to avoid cidx = pidx */
1801 	eq->pidx = eq->cidx = 0;
1802 	eq->doorbells = sc->doorbells;
1803 
1804 	switch (eq->flags & EQ_TYPEMASK) {
1805 	case EQ_CTRL:
1806 		rc = ctrl_eq_alloc(sc, eq);
1807 		break;
1808 
1809 	case EQ_ETH:
1810 		rc = eth_eq_alloc(sc, pi, eq);
1811 		break;
1812 
1813 #ifdef TCP_OFFLOAD_ENABLE
1814 	case EQ_OFLD:
1815 		rc = ofld_eq_alloc(sc, pi, eq);
1816 		break;
1817 #endif
1818 
1819 	default:
1820 		panic("%s: invalid eq type %d.", __func__,
1821 		    eq->flags & EQ_TYPEMASK);
1822 	}
1823 
1824 	if (eq->doorbells &
1825 		(DOORBELL_UDB | DOORBELL_UDBWC | DOORBELL_WCWR)) {
1826 		uint32_t s_qpp = sc->sge.s_qpp;
1827 		uint32_t mask = (1 << s_qpp) - 1;
1828 		volatile uint8_t *udb;
1829 
1830 		udb = (volatile uint8_t *)sc->reg1p + UDBS_DB_OFFSET;
1831 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;   /* pg offset */
1832 		eq->udb_qid = eq->cntxt_id & mask;              /* id in page */
1833 		if (eq->udb_qid > PAGE_SIZE / UDBS_SEG_SIZE)
1834 			eq->doorbells &= ~DOORBELL_WCWR;
1835 		else {
1836 			udb += eq->udb_qid << UDBS_SEG_SHIFT;   /* seg offset */
1837 			eq->udb_qid = 0;
1838 		}
1839 		eq->udb = (volatile void *)udb;
1840 	}
1841 
1842 	if (rc != 0) {
1843 		cxgb_printf(sc->dip, CE_WARN,
1844 		    "failed to allocate egress queue(%d): %d",
1845 		    eq->flags & EQ_TYPEMASK, rc);
1846 	}
1847 
1848 	return (rc);
1849 }
1850 
1851 static int
1852 free_eq(struct adapter *sc, struct sge_eq *eq)
1853 {
1854 	int rc;
1855 
1856 	if (eq->flags & EQ_ALLOCATED) {
1857 		switch (eq->flags & EQ_TYPEMASK) {
1858 		case EQ_CTRL:
1859 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
1860 			    eq->cntxt_id);
1861 			break;
1862 
1863 		case EQ_ETH:
1864 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
1865 			    eq->cntxt_id);
1866 			break;
1867 
1868 #ifdef TCP_OFFLOAD_ENABLE
1869 		case EQ_OFLD:
1870 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
1871 			    eq->cntxt_id);
1872 			break;
1873 #endif
1874 
1875 		default:
1876 			panic("%s: invalid eq type %d.", __func__,
1877 			    eq->flags & EQ_TYPEMASK);
1878 		}
1879 		if (rc != 0) {
1880 			cxgb_printf(sc->dip, CE_WARN,
1881 			    "failed to free egress queue (%d): %d",
1882 			    eq->flags & EQ_TYPEMASK, rc);
1883 			return (rc);
1884 		}
1885 		eq->flags &= ~EQ_ALLOCATED;
1886 	}
1887 
1888 	if (eq->desc != NULL) {
1889 		(void) free_desc_ring(&eq->desc_dhdl, &eq->desc_ahdl);
1890 		eq->desc = NULL;
1891 	}
1892 
1893 	if (eq->flags & EQ_MTX)
1894 		mutex_destroy(&eq->lock);
1895 
1896 	bzero(eq, sizeof (*eq));
1897 	return (0);
1898 }
1899 
1900 #ifdef TCP_OFFLOAD_ENABLE
1901 /* ARGSUSED */
1902 static int
1903 alloc_wrq(struct adapter *sc, struct port_info *pi, struct sge_wrq *wrq,
1904     int idx)
1905 {
1906 	int rc;
1907 
1908 	rc = alloc_eq(sc, pi, &wrq->eq);
1909 	if (rc != 0)
1910 		return (rc);
1911 
1912 	wrq->adapter = sc;
1913 	wrq->wr_list.head = NULL;
1914 	wrq->wr_list.tail = NULL;
1915 
1916 	/*
1917 	 * TODO: use idx to figure out what kind of wrq this is and install
1918 	 * useful kstats for it.
1919 	 */
1920 
1921 	return (rc);
1922 }
1923 
1924 static int
1925 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
1926 {
1927 	int rc;
1928 
1929 	rc = free_eq(sc, &wrq->eq);
1930 	if (rc != 0)
1931 		return (rc);
1932 
1933 	bzero(wrq, sizeof (*wrq));
1934 	return (0);
1935 }
1936 #endif
1937 
1938 static int
1939 alloc_txq(struct port_info *pi, struct sge_txq *txq, int idx)
1940 {
1941 	int rc, i;
1942 	struct adapter *sc = pi->adapter;
1943 	struct sge_eq *eq = &txq->eq;
1944 
1945 	rc = alloc_eq(sc, pi, eq);
1946 	if (rc != 0)
1947 		return (rc);
1948 
1949 	txq->port = pi;
1950 	txq->sdesc = kmem_zalloc(sizeof (struct tx_sdesc) * eq->cap, KM_SLEEP);
1951 	txq->txb_size = eq->qsize * tx_copy_threshold;
1952 	rc = alloc_tx_copybuffer(sc, txq->txb_size, &txq->txb_dhdl,
1953 	    &txq->txb_ahdl, &txq->txb_ba, &txq->txb_va);
1954 	if (rc == 0)
1955 		txq->txb_avail = txq->txb_size;
1956 	else
1957 		txq->txb_avail = txq->txb_size = 0;
1958 
1959 	/*
1960 	 * TODO: is this too low?  Worst case would need around 4 times qsize
1961 	 * (all tx descriptors filled to the brim with SGLs, with each entry in
1962 	 * the SGL coming from a distinct DMA handle).  Increase tx_dhdl_total
1963 	 * if you see too many dma_hdl_failed.
1964 	 */
1965 	txq->tx_dhdl_total = eq->qsize * 2;
1966 	txq->tx_dhdl = kmem_zalloc(sizeof (ddi_dma_handle_t) *
1967 	    txq->tx_dhdl_total, KM_SLEEP);
1968 	for (i = 0; i < txq->tx_dhdl_total; i++) {
1969 		rc = ddi_dma_alloc_handle(sc->dip, &sc->sge.dma_attr_tx,
1970 		    DDI_DMA_SLEEP, 0, &txq->tx_dhdl[i]);
1971 		if (rc != DDI_SUCCESS) {
1972 			cxgb_printf(sc->dip, CE_WARN,
1973 			    "%s: failed to allocate DMA handle (%d)",
1974 			    __func__, rc);
1975 			return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EINVAL);
1976 		}
1977 		txq->tx_dhdl_avail++;
1978 	}
1979 
1980 	txq->ksp = setup_txq_kstats(pi, txq, idx);
1981 
1982 	return (rc);
1983 }
1984 
1985 static int
1986 free_txq(struct port_info *pi, struct sge_txq *txq)
1987 {
1988 	int i;
1989 	struct adapter *sc = pi->adapter;
1990 	struct sge_eq *eq = &txq->eq;
1991 
1992 	if (txq->ksp != NULL) {
1993 		kstat_delete(txq->ksp);
1994 		txq->ksp = NULL;
1995 	}
1996 
1997 	if (txq->txb_va != NULL) {
1998 		(void) free_desc_ring(&txq->txb_dhdl, &txq->txb_ahdl);
1999 		txq->txb_va = NULL;
2000 	}
2001 
2002 	if (txq->sdesc != NULL) {
2003 		struct tx_sdesc *sd;
2004 		ddi_dma_handle_t hdl;
2005 
2006 		TXQ_LOCK(txq);
2007 		while (eq->cidx != eq->pidx) {
2008 			sd = &txq->sdesc[eq->cidx];
2009 
2010 			for (i = sd->hdls_used; i; i--) {
2011 				hdl = txq->tx_dhdl[txq->tx_dhdl_cidx];
2012 				(void) ddi_dma_unbind_handle(hdl);
2013 				if (++txq->tx_dhdl_cidx == txq->tx_dhdl_total)
2014 					txq->tx_dhdl_cidx = 0;
2015 			}
2016 
2017 			ASSERT(sd->m);
2018 			freemsgchain(sd->m);
2019 
2020 			eq->cidx += sd->desc_used;
2021 			if (eq->cidx >= eq->cap)
2022 				eq->cidx -= eq->cap;
2023 
2024 			txq->txb_avail += txq->txb_used;
2025 		}
2026 		ASSERT(txq->tx_dhdl_cidx == txq->tx_dhdl_pidx);
2027 		ASSERT(txq->txb_avail == txq->txb_size);
2028 		TXQ_UNLOCK(txq);
2029 
2030 		kmem_free(txq->sdesc, sizeof (struct tx_sdesc) * eq->cap);
2031 		txq->sdesc = NULL;
2032 	}
2033 
2034 	if (txq->tx_dhdl != NULL) {
2035 		for (i = 0; i < txq->tx_dhdl_total; i++) {
2036 			if (txq->tx_dhdl[i] != NULL)
2037 				ddi_dma_free_handle(&txq->tx_dhdl[i]);
2038 		}
2039 	}
2040 
2041 	(void) free_eq(sc, &txq->eq);
2042 
2043 	bzero(txq, sizeof (*txq));
2044 	return (0);
2045 }
2046 
2047 /*
2048  * Allocates a block of contiguous memory for DMA.  Can be used to allocate
2049  * memory for descriptor rings or for tx/rx copy buffers.
2050  *
2051  * Caller does not have to clean up anything if this function fails, it cleans
2052  * up after itself.
2053  *
2054  * Caller provides the following:
2055  * len		length of the block of memory to allocate.
2056  * flags	DDI_DMA_* flags to use (CONSISTENT/STREAMING, READ/WRITE/RDWR)
2057  * acc_attr	device access attributes for the allocation.
2058  * dma_attr	DMA attributes for the allocation
2059  *
2060  * If the function is successful it fills up this information:
2061  * dma_hdl	DMA handle for the allocated memory
2062  * acc_hdl	access handle for the allocated memory
2063  * ba		bus address of the allocated memory
2064  * va		KVA of the allocated memory.
2065  */
2066 static int
2067 alloc_dma_memory(struct adapter *sc, size_t len, int flags,
2068     ddi_device_acc_attr_t *acc_attr, ddi_dma_attr_t *dma_attr,
2069     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl,
2070     uint64_t *pba, caddr_t *pva)
2071 {
2072 	int rc;
2073 	ddi_dma_handle_t dhdl;
2074 	ddi_acc_handle_t ahdl;
2075 	ddi_dma_cookie_t cookie;
2076 	uint_t ccount;
2077 	caddr_t va;
2078 	size_t real_len;
2079 
2080 	*pva = NULL;
2081 
2082 	/*
2083 	 * DMA handle.
2084 	 */
2085 	rc = ddi_dma_alloc_handle(sc->dip, dma_attr, DDI_DMA_SLEEP, 0, &dhdl);
2086 	if (rc != DDI_SUCCESS) {
2087 		cxgb_printf(sc->dip, CE_WARN,
2088 		    "failed to allocate DMA handle: %d", rc);
2089 
2090 		return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EINVAL);
2091 	}
2092 
2093 	/*
2094 	 * Memory suitable for DMA.
2095 	 */
2096 	rc = ddi_dma_mem_alloc(dhdl, len, acc_attr,
2097 	    flags & DDI_DMA_CONSISTENT ? DDI_DMA_CONSISTENT : DDI_DMA_STREAMING,
2098 	    DDI_DMA_SLEEP, 0, &va, &real_len, &ahdl);
2099 	if (rc != DDI_SUCCESS) {
2100 		cxgb_printf(sc->dip, CE_WARN,
2101 		    "failed to allocate DMA memory: %d", rc);
2102 
2103 		ddi_dma_free_handle(&dhdl);
2104 		return (ENOMEM);
2105 	}
2106 
2107 	if (len != real_len) {
2108 		cxgb_printf(sc->dip, CE_WARN,
2109 		    "%s: len (%u) != real_len (%u)\n", len, real_len);
2110 	}
2111 
2112 	/*
2113 	 * DMA bindings.
2114 	 */
2115 	rc = ddi_dma_addr_bind_handle(dhdl, NULL, va, real_len, flags, NULL,
2116 	    NULL, &cookie, &ccount);
2117 	if (rc != DDI_DMA_MAPPED) {
2118 		cxgb_printf(sc->dip, CE_WARN,
2119 		    "failed to map DMA memory: %d", rc);
2120 
2121 		ddi_dma_mem_free(&ahdl);
2122 		ddi_dma_free_handle(&dhdl);
2123 		return (ENOMEM);
2124 	}
2125 	if (ccount != 1) {
2126 		cxgb_printf(sc->dip, CE_WARN,
2127 		    "unusable DMA mapping (%d segments)", ccount);
2128 		(void) free_desc_ring(&dhdl, &ahdl);
2129 	}
2130 
2131 	bzero(va, real_len);
2132 	*dma_hdl = dhdl;
2133 	*acc_hdl = ahdl;
2134 	*pba = cookie.dmac_laddress;
2135 	*pva = va;
2136 
2137 	return (0);
2138 }
2139 
2140 static int
2141 free_dma_memory(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl)
2142 {
2143 	(void) ddi_dma_unbind_handle(*dhdl);
2144 	ddi_dma_mem_free(ahdl);
2145 	ddi_dma_free_handle(dhdl);
2146 
2147 	return (0);
2148 }
2149 
2150 static int
2151 alloc_desc_ring(struct adapter *sc, size_t len, int rw,
2152     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl,
2153     uint64_t *pba, caddr_t *pva)
2154 {
2155 	ddi_device_acc_attr_t *acc_attr = &sc->sge.acc_attr_desc;
2156 	ddi_dma_attr_t *dma_attr = &sc->sge.dma_attr_desc;
2157 
2158 	return (alloc_dma_memory(sc, len, DDI_DMA_CONSISTENT | rw, acc_attr,
2159 	    dma_attr, dma_hdl, acc_hdl, pba, pva));
2160 }
2161 
2162 static int
2163 free_desc_ring(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl)
2164 {
2165 	return (free_dma_memory(dhdl, ahdl));
2166 }
2167 
2168 static int
2169 alloc_tx_copybuffer(struct adapter *sc, size_t len,
2170     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl,
2171     uint64_t *pba, caddr_t *pva)
2172 {
2173 	ddi_device_acc_attr_t *acc_attr = &sc->sge.acc_attr_tx;
2174 	ddi_dma_attr_t *dma_attr = &sc->sge.dma_attr_desc; /* NOT dma_attr_tx */
2175 
2176 	return (alloc_dma_memory(sc, len, DDI_DMA_STREAMING | DDI_DMA_WRITE,
2177 	    acc_attr, dma_attr, dma_hdl, acc_hdl, pba, pva));
2178 }
2179 
2180 static inline bool
2181 is_new_response(const struct sge_iq *iq, struct rsp_ctrl **ctrl)
2182 {
2183 	(void) ddi_dma_sync(iq->dhdl, (uintptr_t)iq->cdesc -
2184 	    (uintptr_t)iq->desc, iq->esize, DDI_DMA_SYNC_FORKERNEL);
2185 
2186 	*ctrl = (void *)((uintptr_t)iq->cdesc +
2187 	    (iq->esize - sizeof (struct rsp_ctrl)));
2188 
2189 	return ((((*ctrl)->u.type_gen >> S_RSPD_GEN) == iq->gen));
2190 }
2191 
2192 static inline void
2193 iq_next(struct sge_iq *iq)
2194 {
2195 	iq->cdesc = (void *) ((uintptr_t)iq->cdesc + iq->esize);
2196 	if (++iq->cidx == iq->qsize - 1) {
2197 		iq->cidx = 0;
2198 		iq->gen ^= 1;
2199 		iq->cdesc = iq->desc;
2200 	}
2201 }
2202 
2203 /*
2204  * Fill up the freelist by upto nbufs and maybe ring its doorbell.
2205  *
2206  * Returns non-zero to indicate that it should be added to the list of starving
2207  * freelists.
2208  */
2209 static int
2210 refill_fl(struct adapter *sc, struct sge_fl *fl, int nbufs)
2211 {
2212 	uint64_t *d = &fl->desc[fl->pidx];
2213 	struct fl_sdesc *sd = &fl->sdesc[fl->pidx];
2214 
2215 	FL_LOCK_ASSERT_OWNED(fl);
2216 	ASSERT(nbufs >= 0);
2217 
2218 	if (nbufs > fl->needed)
2219 		nbufs = fl->needed;
2220 
2221 	while (nbufs--) {
2222 		if (sd->rxb != NULL) {
2223 			if (sd->rxb->ref_cnt == 1) {
2224 				/*
2225 				 * Buffer is available for recycling.  Two ways
2226 				 * this can happen:
2227 				 *
2228 				 * a) All the packets DMA'd into it last time
2229 				 *    around were within the rx_copy_threshold
2230 				 *    and no part of the buffer was ever passed
2231 				 *    up (ref_cnt never went over 1).
2232 				 *
2233 				 * b) Packets DMA'd into the buffer were passed
2234 				 *    up but have all been freed by the upper
2235 				 *    layers by now (ref_cnt went over 1 but is
2236 				 *    now back to 1).
2237 				 *
2238 				 * Either way the bus address in the descriptor
2239 				 * ring is already valid.
2240 				 */
2241 				ASSERT(*d == cpu_to_be64(sd->rxb->ba));
2242 				d++;
2243 				goto recycled;
2244 			} else {
2245 				/*
2246 				 * Buffer still in use and we need a
2247 				 * replacement. But first release our reference
2248 				 * on the existing buffer.
2249 				 */
2250 				rxbuf_free(sd->rxb);
2251 			}
2252 		}
2253 
2254 		sd->rxb = rxbuf_alloc(sc->sge.rxbuf_cache, KM_NOSLEEP, 1);
2255 		if (sd->rxb == NULL)
2256 			break;
2257 		*d++ = cpu_to_be64(sd->rxb->ba);
2258 
2259 recycled:	fl->pending++;
2260 		sd++;
2261 		fl->needed--;
2262 		if (++fl->pidx == fl->cap) {
2263 			fl->pidx = 0;
2264 			sd = fl->sdesc;
2265 			d = fl->desc;
2266 		}
2267 	}
2268 
2269 	if (fl->pending >= 8)
2270 		ring_fl_db(sc, fl);
2271 
2272 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
2273 }
2274 
2275 #ifndef TAILQ_FOREACH_SAFE
2276 #define	TAILQ_FOREACH_SAFE(var, head, field, tvar)			\
2277 	for ((var) = TAILQ_FIRST((head));				\
2278 	    (var) && ((tvar) = TAILQ_NEXT((var), field), 1);		\
2279 	    (var) = (tvar))
2280 #endif
2281 
2282 /*
2283  * Attempt to refill all starving freelists.
2284  */
2285 static void
2286 refill_sfl(void *arg)
2287 {
2288 	struct adapter *sc = arg;
2289 	struct sge_fl *fl, *fl_temp;
2290 
2291 	mutex_enter(&sc->sfl_lock);
2292 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
2293 		FL_LOCK(fl);
2294 		(void) refill_fl(sc, fl, 64);
2295 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
2296 			TAILQ_REMOVE(&sc->sfl, fl, link);
2297 			fl->flags &= ~FL_STARVING;
2298 		}
2299 		FL_UNLOCK(fl);
2300 	}
2301 
2302 	if (!TAILQ_EMPTY(&sc->sfl) != 0)
2303 		sc->sfl_timer =  timeout(refill_sfl, sc, drv_usectohz(100000));
2304 	mutex_exit(&sc->sfl_lock);
2305 }
2306 
2307 static void
2308 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
2309 {
2310 	mutex_enter(&sc->sfl_lock);
2311 	FL_LOCK(fl);
2312 	if ((fl->flags & FL_DOOMED) == 0) {
2313 		if (TAILQ_EMPTY(&sc->sfl) != 0) {
2314 			sc->sfl_timer = timeout(refill_sfl, sc,
2315 			    drv_usectohz(100000));
2316 		}
2317 		fl->flags |= FL_STARVING;
2318 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
2319 	}
2320 	FL_UNLOCK(fl);
2321 	mutex_exit(&sc->sfl_lock);
2322 }
2323 
2324 static void
2325 free_fl_bufs(struct sge_fl *fl)
2326 {
2327 	struct fl_sdesc *sd;
2328 	unsigned int i;
2329 
2330 	FL_LOCK_ASSERT_OWNED(fl);
2331 
2332 	for (i = 0; i < fl->cap; i++) {
2333 		sd = &fl->sdesc[i];
2334 
2335 		if (sd->rxb != NULL) {
2336 			rxbuf_free(sd->rxb);
2337 			sd->rxb = NULL;
2338 		}
2339 	}
2340 }
2341 
2342 /*
2343  * Note that fl->cidx and fl->offset are left unchanged in case of failure.
2344  */
2345 static mblk_t *
2346 get_fl_payload(struct adapter *sc, struct sge_fl *fl,
2347 	       uint32_t len_newbuf, int *fl_bufs_used)
2348 {
2349 	struct mblk_pair frame = {0};
2350 	struct rxbuf *rxb;
2351 	mblk_t *m = NULL;
2352 	uint_t nbuf = 0, len, copy, n;
2353 	uint32_t cidx, offset, rcidx, roffset;
2354 
2355 	/*
2356 	 * The SGE won't pack a new frame into the current buffer if the entire
2357 	 * payload doesn't fit in the remaining space.  Move on to the next buf
2358 	 * in that case.
2359 	 */
2360 	rcidx = fl->cidx;
2361 	roffset = fl->offset;
2362 	if (fl->offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
2363 		fl->offset = 0;
2364 		if (++fl->cidx == fl->cap)
2365 			fl->cidx = 0;
2366 		nbuf++;
2367 	}
2368 	cidx = fl->cidx;
2369 	offset = fl->offset;
2370 
2371 	len = G_RSPD_LEN(len_newbuf);	/* pktshift + payload length */
2372 	copy = (len <= fl->copy_threshold);
2373 	if (copy != 0) {
2374 		frame.head = m = allocb(len, BPRI_HI);
2375 		if (m == NULL) {
2376 			fl->allocb_fail++;
2377 			cmn_err(CE_WARN,"%s: mbuf allocation failure "
2378 					"count = %llu", __func__,
2379 					(unsigned long long)fl->allocb_fail);
2380 			fl->cidx = rcidx;
2381 			fl->offset = roffset;
2382 			return (NULL);
2383 		}
2384 	}
2385 
2386 	while (len) {
2387 		rxb = fl->sdesc[cidx].rxb;
2388 		n = min(len, rxb->buf_size - offset);
2389 
2390 		(void) ddi_dma_sync(rxb->dhdl, offset, n,
2391 		    DDI_DMA_SYNC_FORKERNEL);
2392 
2393 		if (copy != 0)
2394 			bcopy(rxb->va + offset, m->b_wptr, n);
2395 		else {
2396 			m = desballoc((unsigned char *)rxb->va + offset, n,
2397 			    BPRI_HI, &rxb->freefunc);
2398 			if (m == NULL) {
2399 				fl->allocb_fail++;
2400 				cmn_err(CE_WARN,
2401 					"%s: mbuf allocation failure "
2402 					"count = %llu", __func__,
2403 					(unsigned long long)fl->allocb_fail);
2404 				if (frame.head)
2405 					freemsgchain(frame.head);
2406 				fl->cidx = rcidx;
2407 				fl->offset = roffset;
2408 				return (NULL);
2409 			}
2410 			atomic_inc_uint(&rxb->ref_cnt);
2411 			if (frame.head != NULL)
2412 				frame.tail->b_cont = m;
2413 			else
2414 				frame.head = m;
2415 			frame.tail = m;
2416 		}
2417 		m->b_wptr += n;
2418 		len -= n;
2419 		offset += roundup(n, sc->sge.fl_align);
2420 		ASSERT(offset <= rxb->buf_size);
2421 		if (offset == rxb->buf_size) {
2422 			offset = 0;
2423 			if (++cidx == fl->cap)
2424 				cidx = 0;
2425 			nbuf++;
2426 		}
2427 	}
2428 
2429 	fl->cidx = cidx;
2430 	fl->offset = offset;
2431 	(*fl_bufs_used) += nbuf;
2432 
2433 	ASSERT(frame.head != NULL);
2434 	return (frame.head);
2435 }
2436 
2437 /*
2438  * We'll do immediate data tx for non-LSO, but only when not coalescing.  We're
2439  * willing to use upto 2 hardware descriptors which means a maximum of 96 bytes
2440  * of immediate data.
2441  */
2442 #define	IMM_LEN ( \
2443 	2 * EQ_ESIZE \
2444 	- sizeof (struct fw_eth_tx_pkt_wr) \
2445 	- sizeof (struct cpl_tx_pkt_core))
2446 
2447 /*
2448  * Returns non-zero on failure, no need to cleanup anything in that case.
2449  *
2450  * Note 1: We always try to pull up the mblk if required and return E2BIG only
2451  * if this fails.
2452  *
2453  * Note 2: We'll also pullup incoming mblk if HW_LSO is set and the first mblk
2454  * does not have the TCP header in it.
2455  */
2456 static int
2457 get_frame_txinfo(struct sge_txq *txq, mblk_t **fp, struct txinfo *txinfo,
2458     int sgl_only)
2459 {
2460 	uint32_t flags = 0, len, n;
2461 	mblk_t *m = *fp;
2462 	int rc;
2463 
2464 	TXQ_LOCK_ASSERT_OWNED(txq);	/* will manipulate txb and dma_hdls */
2465 
2466 	mac_hcksum_get(m, NULL, NULL, NULL, NULL, &flags);
2467 	txinfo->flags = flags;
2468 
2469 	mac_lso_get(m, &txinfo->mss, &flags);
2470 	txinfo->flags |= flags;
2471 
2472 	if (flags & HW_LSO)
2473 		sgl_only = 1;	/* Do not allow immediate data with LSO */
2474 
2475 start:	txinfo->nsegs = 0;
2476 	txinfo->hdls_used = 0;
2477 	txinfo->txb_used = 0;
2478 	txinfo->len = 0;
2479 
2480 	/* total length and a rough estimate of # of segments */
2481 	n = 0;
2482 	for (; m; m = m->b_cont) {
2483 		len = MBLKL(m);
2484 		n += (len / PAGE_SIZE) + 1;
2485 		txinfo->len += len;
2486 	}
2487 	m = *fp;
2488 
2489 	if (n >= TX_SGL_SEGS || (flags & HW_LSO && MBLKL(m) < 50)) {
2490 		txq->pullup_early++;
2491 		m = msgpullup(*fp, -1);
2492 		if (m == NULL) {
2493 			txq->pullup_failed++;
2494 			return (E2BIG);	/* (*fp) left as it was */
2495 		}
2496 		freemsg(*fp);
2497 		*fp = m;
2498 		mac_hcksum_set(m, 0, 0, 0, 0, txinfo->flags);
2499 	}
2500 
2501 	if (txinfo->len <= IMM_LEN && !sgl_only)
2502 		return (0);	/* nsegs = 0 tells caller to use imm. tx */
2503 
2504 	if (txinfo->len <= txq->copy_threshold &&
2505 	    copy_into_txb(txq, m, txinfo->len, txinfo) == 0)
2506 		goto done;
2507 
2508 	for (; m; m = m->b_cont) {
2509 
2510 		len = MBLKL(m);
2511 
2512 		/* Use tx copy buffer if this mblk is small enough */
2513 		if (len <= txq->copy_threshold &&
2514 		    copy_into_txb(txq, m, len, txinfo) == 0)
2515 			continue;
2516 
2517 		/* Add DMA bindings for this mblk to the SGL */
2518 		rc = add_mblk(txq, txinfo, m, len);
2519 
2520 		if (rc == E2BIG ||
2521 		    (txinfo->nsegs == TX_SGL_SEGS && m->b_cont)) {
2522 
2523 			txq->pullup_late++;
2524 			m = msgpullup(*fp, -1);
2525 			if (m != NULL) {
2526 				free_txinfo_resources(txq, txinfo);
2527 				freemsg(*fp);
2528 				*fp = m;
2529 				mac_hcksum_set(m, 0, 0, 0, 0, txinfo->flags);
2530 				goto start;
2531 			}
2532 
2533 			txq->pullup_failed++;
2534 			rc = E2BIG;
2535 		}
2536 
2537 		if (rc != 0) {
2538 			free_txinfo_resources(txq, txinfo);
2539 			return (rc);
2540 		}
2541 	}
2542 
2543 	ASSERT(txinfo->nsegs > 0 && txinfo->nsegs <= TX_SGL_SEGS);
2544 
2545 done:
2546 
2547 	/*
2548 	 * Store the # of flits required to hold this frame's SGL in nflits.  An
2549 	 * SGL has a (ULPTX header + len0, addr0) tuple optionally followed by
2550 	 * multiple (len0 + len1, addr0, addr1) tuples.  If addr1 is not used
2551 	 * then len1 must be set to 0.
2552 	 */
2553 	n = txinfo->nsegs - 1;
2554 	txinfo->nflits = (3 * n) / 2 + (n & 1) + 2;
2555 	if (n & 1)
2556 		txinfo->sgl.sge[n / 2].len[1] = cpu_to_be32(0);
2557 
2558 	txinfo->sgl.cmd_nsge = cpu_to_be32(V_ULPTX_CMD((u32)ULP_TX_SC_DSGL) |
2559 	    V_ULPTX_NSGE(txinfo->nsegs));
2560 
2561 	return (0);
2562 }
2563 
2564 static inline int
2565 fits_in_txb(struct sge_txq *txq, int len, int *waste)
2566 {
2567 	if (txq->txb_avail < len)
2568 		return (0);
2569 
2570 	if (txq->txb_next + len <= txq->txb_size) {
2571 		*waste = 0;
2572 		return (1);
2573 	}
2574 
2575 	*waste = txq->txb_size - txq->txb_next;
2576 
2577 	return (txq->txb_avail - *waste < len ? 0 : 1);
2578 }
2579 
2580 #define	TXB_CHUNK	64
2581 
2582 /*
2583  * Copies the specified # of bytes into txq's tx copy buffer and updates txinfo
2584  * and txq to indicate resources used.  Caller has to make sure that those many
2585  * bytes are available in the mblk chain (b_cont linked).
2586  */
2587 static inline int
2588 copy_into_txb(struct sge_txq *txq, mblk_t *m, int len, struct txinfo *txinfo)
2589 {
2590 	int waste, n;
2591 
2592 	TXQ_LOCK_ASSERT_OWNED(txq);	/* will manipulate txb */
2593 
2594 	if (!fits_in_txb(txq, len, &waste)) {
2595 		txq->txb_full++;
2596 		return (ENOMEM);
2597 	}
2598 
2599 	if (waste != 0) {
2600 		ASSERT((waste & (TXB_CHUNK - 1)) == 0);
2601 		txinfo->txb_used += waste;
2602 		txq->txb_avail -= waste;
2603 		txq->txb_next = 0;
2604 	}
2605 
2606 	for (n = 0; n < len; m = m->b_cont) {
2607 		bcopy(m->b_rptr, txq->txb_va + txq->txb_next + n, MBLKL(m));
2608 		n += MBLKL(m);
2609 	}
2610 
2611 	add_seg(txinfo, txq->txb_ba + txq->txb_next, len);
2612 
2613 	n = roundup(len, TXB_CHUNK);
2614 	txinfo->txb_used += n;
2615 	txq->txb_avail -= n;
2616 	txq->txb_next += n;
2617 	ASSERT(txq->txb_next <= txq->txb_size);
2618 	if (txq->txb_next == txq->txb_size)
2619 		txq->txb_next = 0;
2620 
2621 	return (0);
2622 }
2623 
2624 static inline void
2625 add_seg(struct txinfo *txinfo, uint64_t ba, uint32_t len)
2626 {
2627 	ASSERT(txinfo->nsegs < TX_SGL_SEGS);	/* must have room */
2628 
2629 	if (txinfo->nsegs != 0) {
2630 		int idx = txinfo->nsegs - 1;
2631 		txinfo->sgl.sge[idx / 2].len[idx & 1] = cpu_to_be32(len);
2632 		txinfo->sgl.sge[idx / 2].addr[idx & 1] = cpu_to_be64(ba);
2633 	} else {
2634 		txinfo->sgl.len0 = cpu_to_be32(len);
2635 		txinfo->sgl.addr0 = cpu_to_be64(ba);
2636 	}
2637 	txinfo->nsegs++;
2638 }
2639 
2640 /*
2641  * This function cleans up any partially allocated resources when it fails so
2642  * there's nothing for the caller to clean up in that case.
2643  *
2644  * EIO indicates permanent failure.  Caller should drop the frame containing
2645  * this mblk and continue.
2646  *
2647  * E2BIG indicates that the SGL length for this mblk exceeds the hardware
2648  * limit.  Caller should pull up the frame before trying to send it out.
2649  * (This error means our pullup_early heuristic did not work for this frame)
2650  *
2651  * ENOMEM indicates a temporary shortage of resources (DMA handles, other DMA
2652  * resources, etc.).  Caller should suspend the tx queue and wait for reclaim to
2653  * free up resources.
2654  */
2655 static inline int
2656 add_mblk(struct sge_txq *txq, struct txinfo *txinfo, mblk_t *m, int len)
2657 {
2658 	ddi_dma_handle_t dhdl;
2659 	ddi_dma_cookie_t cookie;
2660 	uint_t ccount = 0;
2661 	int rc;
2662 
2663 	TXQ_LOCK_ASSERT_OWNED(txq);	/* will manipulate dhdls */
2664 
2665 	if (txq->tx_dhdl_avail == 0) {
2666 		txq->dma_hdl_failed++;
2667 		return (ENOMEM);
2668 	}
2669 
2670 	dhdl = txq->tx_dhdl[txq->tx_dhdl_pidx];
2671 	rc = ddi_dma_addr_bind_handle(dhdl, NULL, (caddr_t)m->b_rptr, len,
2672 	    DDI_DMA_WRITE | DDI_DMA_STREAMING, DDI_DMA_DONTWAIT, NULL, &cookie,
2673 	    &ccount);
2674 	if (rc != DDI_DMA_MAPPED) {
2675 		txq->dma_map_failed++;
2676 
2677 		ASSERT(rc != DDI_DMA_INUSE && rc != DDI_DMA_PARTIAL_MAP);
2678 
2679 		return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EIO);
2680 	}
2681 
2682 	if (ccount + txinfo->nsegs > TX_SGL_SEGS) {
2683 		(void) ddi_dma_unbind_handle(dhdl);
2684 		return (E2BIG);
2685 	}
2686 
2687 	add_seg(txinfo, cookie.dmac_laddress, cookie.dmac_size);
2688 	while (--ccount) {
2689 		ddi_dma_nextcookie(dhdl, &cookie);
2690 		add_seg(txinfo, cookie.dmac_laddress, cookie.dmac_size);
2691 	}
2692 
2693 	if (++txq->tx_dhdl_pidx == txq->tx_dhdl_total)
2694 		txq->tx_dhdl_pidx = 0;
2695 	txq->tx_dhdl_avail--;
2696 	txinfo->hdls_used++;
2697 
2698 	return (0);
2699 }
2700 
2701 /*
2702  * Releases all the txq resources used up in the specified txinfo.
2703  */
2704 static void
2705 free_txinfo_resources(struct sge_txq *txq, struct txinfo *txinfo)
2706 {
2707 	int n;
2708 
2709 	TXQ_LOCK_ASSERT_OWNED(txq);	/* dhdls, txb */
2710 
2711 	n = txinfo->txb_used;
2712 	if (n > 0) {
2713 		txq->txb_avail += n;
2714 		if (n <= txq->txb_next)
2715 			txq->txb_next -= n;
2716 		else {
2717 			n -= txq->txb_next;
2718 			txq->txb_next = txq->txb_size - n;
2719 		}
2720 	}
2721 
2722 	for (n = txinfo->hdls_used; n > 0; n--) {
2723 		if (txq->tx_dhdl_pidx > 0)
2724 			txq->tx_dhdl_pidx--;
2725 		else
2726 			txq->tx_dhdl_pidx = txq->tx_dhdl_total - 1;
2727 		txq->tx_dhdl_avail++;
2728 		(void) ddi_dma_unbind_handle(txq->tx_dhdl[txq->tx_dhdl_pidx]);
2729 	}
2730 }
2731 
2732 /*
2733  * Returns 0 to indicate that m has been accepted into a coalesced tx work
2734  * request.  It has either been folded into txpkts or txpkts was flushed and m
2735  * has started a new coalesced work request (as the first frame in a fresh
2736  * txpkts).
2737  *
2738  * Returns non-zero to indicate a failure - caller is responsible for
2739  * transmitting m, if there was anything in txpkts it has been flushed.
2740  */
2741 static int
2742 add_to_txpkts(struct sge_txq *txq, struct txpkts *txpkts, mblk_t *m,
2743     struct txinfo *txinfo)
2744 {
2745 	struct sge_eq *eq = &txq->eq;
2746 	int can_coalesce;
2747 	struct tx_sdesc *txsd;
2748 	uint8_t flits;
2749 
2750 	TXQ_LOCK_ASSERT_OWNED(txq);
2751 
2752 	if (txpkts->npkt > 0) {
2753 		flits = TXPKTS_PKT_HDR + txinfo->nflits;
2754 		can_coalesce = (txinfo->flags & HW_LSO) == 0 &&
2755 		    txpkts->nflits + flits <= TX_WR_FLITS &&
2756 		    txpkts->nflits + flits <= eq->avail * 8 &&
2757 		    txpkts->plen + txinfo->len < 65536;
2758 
2759 		if (can_coalesce != 0) {
2760 			txpkts->tail->b_next = m;
2761 			txpkts->tail = m;
2762 			txpkts->npkt++;
2763 			txpkts->nflits += flits;
2764 			txpkts->plen += txinfo->len;
2765 
2766 			txsd = &txq->sdesc[eq->pidx];
2767 			txsd->txb_used += txinfo->txb_used;
2768 			txsd->hdls_used += txinfo->hdls_used;
2769 
2770 			return (0);
2771 		}
2772 
2773 		/*
2774 		 * Couldn't coalesce m into txpkts.  The first order of business
2775 		 * is to send txpkts on its way.  Then we'll revisit m.
2776 		 */
2777 		write_txpkts_wr(txq, txpkts);
2778 	}
2779 
2780 	/*
2781 	 * Check if we can start a new coalesced tx work request with m as
2782 	 * the first packet in it.
2783 	 */
2784 
2785 	ASSERT(txpkts->npkt == 0);
2786 	ASSERT(txinfo->len < 65536);
2787 
2788 	flits = TXPKTS_WR_HDR + txinfo->nflits;
2789 	can_coalesce = (txinfo->flags & HW_LSO) == 0 &&
2790 	    flits <= eq->avail * 8 && flits <= TX_WR_FLITS;
2791 
2792 	if (can_coalesce == 0)
2793 		return (EINVAL);
2794 
2795 	/*
2796 	 * Start a fresh coalesced tx WR with m as the first frame in it.
2797 	 */
2798 	txpkts->tail = m;
2799 	txpkts->npkt = 1;
2800 	txpkts->nflits = flits;
2801 	txpkts->flitp = &eq->desc[eq->pidx].flit[2];
2802 	txpkts->plen = txinfo->len;
2803 
2804 	txsd = &txq->sdesc[eq->pidx];
2805 	txsd->m = m;
2806 	txsd->txb_used = txinfo->txb_used;
2807 	txsd->hdls_used = txinfo->hdls_used;
2808 
2809 	return (0);
2810 }
2811 
2812 /*
2813  * Note that write_txpkts_wr can never run out of hardware descriptors (but
2814  * write_txpkt_wr can).  add_to_txpkts ensures that a frame is accepted for
2815  * coalescing only if sufficient hardware descriptors are available.
2816  */
2817 static void
2818 write_txpkts_wr(struct sge_txq *txq, struct txpkts *txpkts)
2819 {
2820 	struct sge_eq *eq = &txq->eq;
2821 	struct fw_eth_tx_pkts_wr *wr;
2822 	struct tx_sdesc *txsd;
2823 	uint32_t ctrl;
2824 	uint16_t ndesc;
2825 
2826 	TXQ_LOCK_ASSERT_OWNED(txq);	/* pidx, avail */
2827 
2828 	ndesc = howmany(txpkts->nflits, 8);
2829 
2830 	wr = (void *)&eq->desc[eq->pidx];
2831 	wr->op_pkd = cpu_to_be32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR) |
2832 	    V_FW_WR_IMMDLEN(0)); /* immdlen does not matter in this WR */
2833 	ctrl = V_FW_WR_LEN16(howmany(txpkts->nflits, 2));
2834 	if (eq->avail == ndesc)
2835 		ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ;
2836 	wr->equiq_to_len16 = cpu_to_be32(ctrl);
2837 	wr->plen = cpu_to_be16(txpkts->plen);
2838 	wr->npkt = txpkts->npkt;
2839 	wr->r3 = wr->type = 0;
2840 
2841 	/* Everything else already written */
2842 
2843 	txsd = &txq->sdesc[eq->pidx];
2844 	txsd->desc_used = ndesc;
2845 
2846 	txq->txb_used += txsd->txb_used / TXB_CHUNK;
2847 	txq->hdl_used += txsd->hdls_used;
2848 
2849 	ASSERT(eq->avail >= ndesc);
2850 
2851 	eq->pending += ndesc;
2852 	eq->avail -= ndesc;
2853 	eq->pidx += ndesc;
2854 	if (eq->pidx >= eq->cap)
2855 		eq->pidx -= eq->cap;
2856 
2857 	txq->txpkts_pkts += txpkts->npkt;
2858 	txq->txpkts_wrs++;
2859 	txpkts->npkt = 0;	/* emptied */
2860 }
2861 
2862 static int
2863 write_txpkt_wr(struct port_info *pi, struct sge_txq *txq, mblk_t *m,
2864     struct txinfo *txinfo)
2865 {
2866 	struct sge_eq *eq = &txq->eq;
2867 	struct fw_eth_tx_pkt_wr *wr;
2868 	struct cpl_tx_pkt_core *cpl;
2869 	uint32_t ctrl;	/* used in many unrelated places */
2870 	uint64_t ctrl1;
2871 	int nflits, ndesc;
2872 	struct tx_sdesc *txsd;
2873 	caddr_t dst;
2874 
2875 	TXQ_LOCK_ASSERT_OWNED(txq);	/* pidx, avail */
2876 
2877 	/*
2878 	 * Do we have enough flits to send this frame out?
2879 	 */
2880 	ctrl = sizeof (struct cpl_tx_pkt_core);
2881 	if (txinfo->flags & HW_LSO) {
2882 		nflits = TXPKT_LSO_WR_HDR;
2883 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
2884 	} else
2885 		nflits = TXPKT_WR_HDR;
2886 	if (txinfo->nsegs > 0)
2887 		nflits += txinfo->nflits;
2888 	else {
2889 		nflits += howmany(txinfo->len, 8);
2890 		ctrl += txinfo->len;
2891 	}
2892 	ndesc = howmany(nflits, 8);
2893 	if (ndesc > eq->avail)
2894 		return (ENOMEM);
2895 
2896 	/* Firmware work request header */
2897 	wr = (void *)&eq->desc[eq->pidx];
2898 	wr->op_immdlen = cpu_to_be32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
2899 	    V_FW_WR_IMMDLEN(ctrl));
2900 	ctrl = V_FW_WR_LEN16(howmany(nflits, 2));
2901 	if (eq->avail == ndesc)
2902 		ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ;
2903 	wr->equiq_to_len16 = cpu_to_be32(ctrl);
2904 	wr->r3 = 0;
2905 
2906 	if (txinfo->flags & HW_LSO) {
2907 		uint16_t etype;
2908 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
2909 		char *p = (void *)m->b_rptr;
2910 		ctrl = V_LSO_OPCODE((u32)CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
2911 		    F_LSO_LAST_SLICE;
2912 
2913 		etype = ntohs(((struct ether_header *)p)->ether_type);
2914 		if (etype == ETHERTYPE_VLAN) {
2915 			ctrl |= V_LSO_ETHHDR_LEN(1);
2916 			etype = ntohs(((struct ether_vlan_header *)p)->ether_type);
2917 			p += sizeof (struct ether_vlan_header);
2918 		} else {
2919 			p += sizeof (struct ether_header);
2920 		}
2921 
2922 		switch (etype) {
2923 		case ETHERTYPE_IP:
2924 			ctrl |= V_LSO_IPHDR_LEN(IPH_HDR_LENGTH(p) / 4);
2925 			p += IPH_HDR_LENGTH(p);
2926 			break;
2927 		case ETHERTYPE_IPV6:
2928 			ctrl |= F_LSO_IPV6;
2929 			ctrl |= V_LSO_IPHDR_LEN(sizeof (ip6_t) / 4);
2930 			p += sizeof (ip6_t);
2931 		default:
2932 			break;
2933 		}
2934 
2935 		ctrl |= V_LSO_TCPHDR_LEN(TCP_HDR_LENGTH((tcph_t *)p) / 4);
2936 
2937 		lso->lso_ctrl = cpu_to_be32(ctrl);
2938 		lso->ipid_ofst = cpu_to_be16(0);
2939 		lso->mss = cpu_to_be16(txinfo->mss);
2940 		lso->seqno_offset = cpu_to_be32(0);
2941 		if (is_t4(pi->adapter->params.chip))
2942 			lso->len = cpu_to_be32(txinfo->len);
2943 		else
2944 			lso->len = cpu_to_be32(V_LSO_T5_XFER_SIZE(txinfo->len));
2945 
2946 		cpl = (void *)(lso + 1);
2947 
2948 		txq->tso_wrs++;
2949 	} else
2950 		cpl = (void *)(wr + 1);
2951 
2952 	/* Checksum offload */
2953 	ctrl1 = 0;
2954 	if (!(txinfo->flags & HCK_IPV4_HDRCKSUM))
2955 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
2956 	if (!(txinfo->flags & HCK_FULLCKSUM))
2957 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
2958 	if (ctrl1 == 0)
2959 		txq->txcsum++;	/* some hardware assistance provided */
2960 
2961 	/* CPL header */
2962 	cpl->ctrl0 = cpu_to_be32(V_TXPKT_OPCODE(CPL_TX_PKT) |
2963 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf));
2964 	cpl->pack = 0;
2965 	cpl->len = cpu_to_be16(txinfo->len);
2966 	cpl->ctrl1 = cpu_to_be64(ctrl1);
2967 
2968 	/* Software descriptor */
2969 	txsd = &txq->sdesc[eq->pidx];
2970 	txsd->m = m;
2971 	txsd->txb_used = txinfo->txb_used;
2972 	txsd->hdls_used = txinfo->hdls_used;
2973 	/* LINTED: E_ASSIGN_NARROW_CONV */
2974 	txsd->desc_used = ndesc;
2975 
2976 	txq->txb_used += txinfo->txb_used / TXB_CHUNK;
2977 	txq->hdl_used += txinfo->hdls_used;
2978 
2979 	eq->pending += ndesc;
2980 	eq->avail -= ndesc;
2981 	eq->pidx += ndesc;
2982 	if (eq->pidx >= eq->cap)
2983 		eq->pidx -= eq->cap;
2984 
2985 	/* SGL */
2986 	dst = (void *)(cpl + 1);
2987 	if (txinfo->nsegs > 0) {
2988 		txq->sgl_wrs++;
2989 		copy_to_txd(eq, (void *)&txinfo->sgl, &dst, txinfo->nflits * 8);
2990 
2991 		/* Need to zero-pad to a 16 byte boundary if not on one */
2992 		if ((uintptr_t)dst & 0xf)
2993 			/* LINTED: E_BAD_PTR_CAST_ALIGN */
2994 			*(uint64_t *)dst = 0;
2995 
2996 	} else {
2997 		txq->imm_wrs++;
2998 #ifdef DEBUG
2999 		ctrl = txinfo->len;
3000 #endif
3001 		for (; m; m = m->b_cont) {
3002 			copy_to_txd(eq, (void *)m->b_rptr, &dst, MBLKL(m));
3003 #ifdef DEBUG
3004 			ctrl -= MBLKL(m);
3005 #endif
3006 		}
3007 		ASSERT(ctrl == 0);
3008 	}
3009 
3010 	txq->txpkt_wrs++;
3011 	return (0);
3012 }
3013 
3014 static inline void
3015 write_ulp_cpl_sgl(struct port_info *pi, struct sge_txq *txq,
3016     struct txpkts *txpkts, struct txinfo *txinfo)
3017 {
3018 	struct ulp_txpkt *ulpmc;
3019 	struct ulptx_idata *ulpsc;
3020 	struct cpl_tx_pkt_core *cpl;
3021 	uintptr_t flitp, start, end;
3022 	uint64_t ctrl;
3023 	caddr_t dst;
3024 
3025 	ASSERT(txpkts->npkt > 0);
3026 
3027 	start = (uintptr_t)txq->eq.desc;
3028 	end = (uintptr_t)txq->eq.spg;
3029 
3030 	/* Checksum offload */
3031 	ctrl = 0;
3032 	if (!(txinfo->flags & HCK_IPV4_HDRCKSUM))
3033 		ctrl |= F_TXPKT_IPCSUM_DIS;
3034 	if (!(txinfo->flags & HCK_FULLCKSUM))
3035 		ctrl |= F_TXPKT_L4CSUM_DIS;
3036 	if (ctrl == 0)
3037 		txq->txcsum++;	/* some hardware assistance provided */
3038 
3039 	/*
3040 	 * The previous packet's SGL must have ended at a 16 byte boundary (this
3041 	 * is required by the firmware/hardware).  It follows that flitp cannot
3042 	 * wrap around between the ULPTX master command and ULPTX subcommand (8
3043 	 * bytes each), and that it can not wrap around in the middle of the
3044 	 * cpl_tx_pkt_core either.
3045 	 */
3046 	flitp = (uintptr_t)txpkts->flitp;
3047 	ASSERT((flitp & 0xf) == 0);
3048 
3049 	/* ULP master command */
3050 	ulpmc = (void *)flitp;
3051 	ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0));
3052 	ulpmc->len = htonl(howmany(sizeof (*ulpmc) + sizeof (*ulpsc) +
3053 	    sizeof (*cpl) + 8 * txinfo->nflits, 16));
3054 
3055 	/* ULP subcommand */
3056 	ulpsc = (void *)(ulpmc + 1);
3057 	ulpsc->cmd_more = cpu_to_be32(V_ULPTX_CMD((u32)ULP_TX_SC_IMM) |
3058 	    F_ULP_TX_SC_MORE);
3059 	ulpsc->len = cpu_to_be32(sizeof (struct cpl_tx_pkt_core));
3060 
3061 	flitp += sizeof (*ulpmc) + sizeof (*ulpsc);
3062 	if (flitp == end)
3063 		flitp = start;
3064 
3065 	/* CPL_TX_PKT */
3066 	cpl = (void *)flitp;
3067 	cpl->ctrl0 = cpu_to_be32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3068 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf));
3069 	cpl->pack = 0;
3070 	cpl->len = cpu_to_be16(txinfo->len);
3071 	cpl->ctrl1 = cpu_to_be64(ctrl);
3072 
3073 	flitp += sizeof (*cpl);
3074 	if (flitp == end)
3075 		flitp = start;
3076 
3077 	/* SGL for this frame */
3078 	dst = (caddr_t)flitp;
3079 	copy_to_txd(&txq->eq, (void *)&txinfo->sgl, &dst, txinfo->nflits * 8);
3080 	flitp = (uintptr_t)dst;
3081 
3082 	/* Zero pad and advance to a 16 byte boundary if not already at one. */
3083 	if (flitp & 0xf) {
3084 
3085 		/* no matter what, flitp should be on an 8 byte boundary */
3086 		ASSERT((flitp & 0x7) == 0);
3087 
3088 		*(uint64_t *)flitp = 0;
3089 		flitp += sizeof (uint64_t);
3090 		txpkts->nflits++;
3091 	}
3092 
3093 	if (flitp == end)
3094 		flitp = start;
3095 
3096 	txpkts->flitp = (void *)flitp;
3097 }
3098 
3099 static inline void
3100 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
3101 {
3102 	if ((uintptr_t)(*to) + len <= (uintptr_t)eq->spg) {
3103 		bcopy(from, *to, len);
3104 		(*to) += len;
3105 	} else {
3106 		int portion = (uintptr_t)eq->spg - (uintptr_t)(*to);
3107 
3108 		bcopy(from, *to, portion);
3109 		from += portion;
3110 		portion = len - portion;	/* remaining */
3111 		bcopy(from, (void *)eq->desc, portion);
3112 		(*to) = (caddr_t)eq->desc + portion;
3113 	}
3114 }
3115 
3116 static inline void
3117 ring_tx_db(struct adapter *sc, struct sge_eq *eq)
3118 {
3119 	int val, db_mode;
3120 	u_int db = eq->doorbells;
3121 
3122 	if (eq->pending > 1)
3123 		db &= ~DOORBELL_WCWR;
3124 
3125 	if (eq->pending > eq->pidx) {
3126 		int offset = eq->cap - (eq->pending - eq->pidx);
3127 
3128 		/* pidx has wrapped around since last doorbell */
3129 
3130 		(void) ddi_dma_sync(eq->desc_dhdl,
3131 		    offset * sizeof (struct tx_desc), 0,
3132 		    DDI_DMA_SYNC_FORDEV);
3133 		(void) ddi_dma_sync(eq->desc_dhdl,
3134 		    0, eq->pidx * sizeof (struct tx_desc),
3135 		    DDI_DMA_SYNC_FORDEV);
3136 	} else if (eq->pending > 0) {
3137 		(void) ddi_dma_sync(eq->desc_dhdl,
3138 		    (eq->pidx - eq->pending) * sizeof (struct tx_desc),
3139 		    eq->pending * sizeof (struct tx_desc),
3140 		    DDI_DMA_SYNC_FORDEV);
3141 	}
3142 
3143 	membar_producer();
3144 
3145 	if (is_t4(sc->params.chip))
3146 		val = V_PIDX(eq->pending);
3147 	else
3148 		val = V_PIDX_T5(eq->pending);
3149 
3150 	db_mode = (1 << (ffs(db) - 1));
3151 	switch (db_mode) {
3152 		case DOORBELL_UDB:
3153 			*eq->udb = LE_32(V_QID(eq->udb_qid) | val);
3154 			break;
3155 
3156 		case DOORBELL_WCWR:
3157 			{
3158 				volatile uint64_t *dst, *src;
3159 				int i;
3160 				/*
3161 				 * Queues whose 128B doorbell segment fits in
3162 				 * the page do not use relative qid
3163 				 * (udb_qid is always 0).  Only queues with
3164 				 * doorbell segments can do WCWR.
3165 				 */
3166 				ASSERT(eq->udb_qid == 0 && eq->pending == 1);
3167 
3168 				dst = (volatile void *)((uintptr_t)eq->udb +
3169 				    UDBS_WR_OFFSET - UDBS_DB_OFFSET);
3170 				i = eq->pidx ? eq->pidx - 1 : eq->cap - 1;
3171 				src = (void *)&eq->desc[i];
3172 				while (src != (void *)&eq->desc[i + 1])
3173 				        *dst++ = *src++;
3174 				membar_producer();
3175 				break;
3176 			}
3177 
3178 		case DOORBELL_UDBWC:
3179 			*eq->udb = LE_32(V_QID(eq->udb_qid) | val);
3180 			membar_producer();
3181 			break;
3182 
3183 		case DOORBELL_KDB:
3184 			t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
3185 			    V_QID(eq->cntxt_id) | val);
3186 			break;
3187 	}
3188 
3189 	eq->pending = 0;
3190 }
3191 
3192 static int
3193 reclaim_tx_descs(struct sge_txq *txq, int howmany)
3194 {
3195 	struct tx_sdesc *txsd;
3196 	uint_t cidx, can_reclaim, reclaimed, txb_freed, hdls_freed;
3197 	struct sge_eq *eq = &txq->eq;
3198 
3199 	EQ_LOCK_ASSERT_OWNED(eq);
3200 
3201 	cidx = eq->spg->cidx;	/* stable snapshot */
3202 	cidx = be16_to_cpu(cidx);
3203 
3204 	if (cidx >= eq->cidx)
3205 		can_reclaim = cidx - eq->cidx;
3206 	else
3207 		can_reclaim = cidx + eq->cap - eq->cidx;
3208 
3209 	if (can_reclaim == 0)
3210 		return (0);
3211 
3212 	txb_freed = hdls_freed = reclaimed = 0;
3213 	do {
3214 		int ndesc;
3215 
3216 		txsd = &txq->sdesc[eq->cidx];
3217 		ndesc = txsd->desc_used;
3218 
3219 		/* Firmware doesn't return "partial" credits. */
3220 		ASSERT(can_reclaim >= ndesc);
3221 
3222 		/*
3223 		 * We always keep mblk around, even for immediate data.  If mblk
3224 		 * is NULL, this has to be the software descriptor for a credit
3225 		 * flush work request.
3226 		 */
3227 		if (txsd->m != NULL)
3228 			freemsgchain(txsd->m);
3229 #ifdef DEBUG
3230 		else {
3231 			ASSERT(txsd->txb_used == 0);
3232 			ASSERT(txsd->hdls_used == 0);
3233 			ASSERT(ndesc == 1);
3234 		}
3235 #endif
3236 
3237 		txb_freed += txsd->txb_used;
3238 		hdls_freed += txsd->hdls_used;
3239 		reclaimed += ndesc;
3240 
3241 		eq->cidx += ndesc;
3242 		if (eq->cidx >= eq->cap)
3243 			eq->cidx -= eq->cap;
3244 
3245 		can_reclaim -= ndesc;
3246 
3247 	} while (can_reclaim && reclaimed < howmany);
3248 
3249 	eq->avail += reclaimed;
3250 	ASSERT(eq->avail < eq->cap);	/* avail tops out at (cap - 1) */
3251 
3252 	txq->txb_avail += txb_freed;
3253 
3254 	txq->tx_dhdl_avail += hdls_freed;
3255 	ASSERT(txq->tx_dhdl_avail <= txq->tx_dhdl_total);
3256 	for (; hdls_freed; hdls_freed--) {
3257 		(void) ddi_dma_unbind_handle(txq->tx_dhdl[txq->tx_dhdl_cidx]);
3258 		if (++txq->tx_dhdl_cidx == txq->tx_dhdl_total)
3259 			txq->tx_dhdl_cidx = 0;
3260 	}
3261 
3262 	return (reclaimed);
3263 }
3264 
3265 static void
3266 write_txqflush_wr(struct sge_txq *txq)
3267 {
3268 	struct sge_eq *eq = &txq->eq;
3269 	struct fw_eq_flush_wr *wr;
3270 	struct tx_sdesc *txsd;
3271 
3272 	EQ_LOCK_ASSERT_OWNED(eq);
3273 	ASSERT(eq->avail > 0);
3274 
3275 	wr = (void *)&eq->desc[eq->pidx];
3276 	bzero(wr, sizeof (*wr));
3277 	wr->opcode = FW_EQ_FLUSH_WR;
3278 	wr->equiq_to_len16 = cpu_to_be32(V_FW_WR_LEN16(sizeof (*wr) / 16) |
3279 	    F_FW_WR_EQUEQ | F_FW_WR_EQUIQ);
3280 
3281 	txsd = &txq->sdesc[eq->pidx];
3282 	txsd->m = NULL;
3283 	txsd->txb_used = 0;
3284 	txsd->hdls_used = 0;
3285 	txsd->desc_used = 1;
3286 
3287 	eq->pending++;
3288 	eq->avail--;
3289 	if (++eq->pidx == eq->cap)
3290 		eq->pidx = 0;
3291 }
3292 
3293 static int
3294 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, mblk_t *m)
3295 {
3296 	bool csum_ok;
3297 	uint16_t err_vec;
3298 	struct sge_rxq *rxq = (void *)iq;
3299 	struct mblk_pair chain = {0};
3300 	struct adapter *sc = iq->adapter;
3301 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
3302 
3303 	iq->intr_next = iq->intr_params;
3304 
3305 	m->b_rptr += sc->sge.pktshift;
3306 
3307 	/* Compressed error vector is enabled for T6 only */
3308 	if (sc->params.tp.rx_pkt_encap)
3309 		/* It is enabled only in T6 config file */
3310 		err_vec = G_T6_COMPR_RXERR_VEC(ntohs(cpl->err_vec));
3311 	else
3312 		err_vec = ntohs(cpl->err_vec);
3313 
3314 	csum_ok = cpl->csum_calc && !err_vec;
3315 	/* TODO: what about cpl->ip_frag? */
3316 	if (csum_ok && !cpl->ip_frag) {
3317 		mac_hcksum_set(m, 0, 0, 0, 0xffff,
3318 		    HCK_FULLCKSUM_OK | HCK_FULLCKSUM |
3319 		    HCK_IPV4_HDRCKSUM_OK);
3320 		rxq->rxcsum++;
3321 	}
3322 
3323 	/* Add to the chain that we'll send up */
3324 	if (chain.head != NULL)
3325 		chain.tail->b_next = m;
3326 	else
3327 		chain.head = m;
3328 	chain.tail = m;
3329 
3330 	t4_mac_rx(rxq->port, rxq, chain.head);
3331 
3332 	rxq->rxpkts++;
3333 	rxq->rxbytes  += be16_to_cpu(cpl->len);
3334 	return (0);
3335 }
3336 
3337 #define	FL_HW_IDX(idx)	((idx) >> 3)
3338 
3339 static inline void
3340 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
3341 {
3342 	int desc_start, desc_last, ndesc;
3343 	uint32_t v = sc->params.arch.sge_fl_db ;
3344 
3345 	ndesc = FL_HW_IDX(fl->pending);
3346 
3347 	/* Hold back one credit if pidx = cidx */
3348 	if (FL_HW_IDX(fl->pidx) == FL_HW_IDX(fl->cidx))
3349 		ndesc--;
3350 
3351 	/*
3352 	 * There are chances of ndesc modified above (to avoid pidx = cidx).
3353 	 * If there is nothing to post, return.
3354 	 */
3355 	if (ndesc <= 0)
3356 		return;
3357 
3358 	desc_last = FL_HW_IDX(fl->pidx);
3359 
3360 	if (fl->pidx < fl->pending) {
3361 		/* There was a wrap */
3362 		desc_start = FL_HW_IDX(fl->pidx + fl->cap - fl->pending);
3363 
3364 		/* From desc_start to the end of list */
3365 		(void) ddi_dma_sync(fl->dhdl, desc_start * RX_FL_ESIZE, 0,
3366 		    DDI_DMA_SYNC_FORDEV);
3367 
3368 		/* From start of list to the desc_last */
3369 		if (desc_last != 0)
3370 			(void) ddi_dma_sync(fl->dhdl, 0, desc_last *
3371 			    RX_FL_ESIZE, DDI_DMA_SYNC_FORDEV);
3372 	} else {
3373 		/* There was no wrap, sync from start_desc to last_desc */
3374 		desc_start = FL_HW_IDX(fl->pidx - fl->pending);
3375 		(void) ddi_dma_sync(fl->dhdl, desc_start * RX_FL_ESIZE,
3376 		    ndesc * RX_FL_ESIZE, DDI_DMA_SYNC_FORDEV);
3377 	}
3378 
3379 	if (is_t4(sc->params.chip))
3380 		v |= V_PIDX(ndesc);
3381 	else
3382 		v |= V_PIDX_T5(ndesc);
3383 	v |= V_QID(fl->cntxt_id) | V_PIDX(ndesc);
3384 
3385 	membar_producer();
3386 
3387 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL), v);
3388 
3389 	/*
3390 	 * Update pending count:
3391 	 * Deduct the number of descriptors posted
3392 	 */
3393 	fl->pending -= ndesc * 8;
3394 }
3395 
3396 static void
3397 tx_reclaim_task(void *arg)
3398 {
3399 	struct sge_txq *txq = arg;
3400 
3401 	TXQ_LOCK(txq);
3402 	reclaim_tx_descs(txq, txq->eq.qsize);
3403 	TXQ_UNLOCK(txq);
3404 }
3405 
3406 /* ARGSUSED */
3407 static int
3408 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
3409                 mblk_t *m)
3410 {
3411 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
3412 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
3413 	struct adapter *sc = iq->adapter;
3414 	struct sge *s = &sc->sge;
3415 	struct sge_eq *eq;
3416 	struct sge_txq *txq;
3417 
3418 	txq = (void *)s->eqmap[qid - s->eq_start];
3419 	eq = &txq->eq;
3420 	txq->qflush++;
3421 	t4_mac_tx_update(txq->port, txq);
3422 
3423 	ddi_taskq_dispatch(sc->tq[eq->tx_chan], tx_reclaim_task,
3424 		(void *)txq, DDI_NOSLEEP);
3425 
3426 	return (0);
3427 }
3428 
3429 static int
3430 handle_fw_rpl(struct sge_iq *iq, const struct rss_header *rss, mblk_t *m)
3431 {
3432 	struct adapter *sc = iq->adapter;
3433 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
3434 
3435 	ASSERT(m == NULL);
3436 
3437 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
3438 		const struct rss_header *rss2;
3439 
3440 		rss2 = (const struct rss_header *)&cpl->data[0];
3441 		return (sc->cpl_handler[rss2->opcode](iq, rss2, m));
3442 	}
3443 	return (sc->fw_msg_handler[cpl->type](sc, &cpl->data[0]));
3444 }
3445 
3446 int
3447 t4_alloc_tx_maps(struct adapter *sc, struct tx_maps *txmaps, int count,
3448     int flags)
3449 {
3450 	int i, rc;
3451 
3452 	txmaps->map_total =  count;
3453 	txmaps->map_avail = txmaps->map_cidx = txmaps->map_pidx = 0;
3454 
3455 	txmaps->map =  kmem_zalloc(sizeof (ddi_dma_handle_t) *
3456 	    txmaps->map_total, flags);
3457 
3458 	for (i = 0; i < count; i++) {
3459 		rc = ddi_dma_alloc_handle(sc->dip, &sc->sge.dma_attr_tx,
3460 		    DDI_DMA_SLEEP, 0, &txmaps->map[i]);
3461 		if (rc != DDI_SUCCESS) {
3462 			cxgb_printf(sc->dip, CE_WARN,
3463 			    "%s: failed to allocate DMA handle (%d)",
3464 			    __func__, rc);
3465 			return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EINVAL);
3466 		}
3467 		txmaps->map_avail++;
3468 	}
3469 
3470 	return (0);
3471 }
3472 
3473 #define	KS_UINIT(x)	kstat_named_init(&kstatp->x, #x, KSTAT_DATA_ULONG)
3474 #define	KS_CINIT(x)	kstat_named_init(&kstatp->x, #x, KSTAT_DATA_CHAR)
3475 #define	KS_U_SET(x, y)	kstatp->x.value.ul = (y)
3476 #define	KS_U_FROM(x, y)	kstatp->x.value.ul = (y)->x
3477 #define	KS_C_SET(x, ...)	\
3478 			(void) snprintf(kstatp->x.value.c, 16,  __VA_ARGS__)
3479 
3480 /*
3481  * cxgbe:X:config
3482  */
3483 struct cxgbe_port_config_kstats {
3484 	kstat_named_t idx;
3485 	kstat_named_t nrxq;
3486 	kstat_named_t ntxq;
3487 	kstat_named_t first_rxq;
3488 	kstat_named_t first_txq;
3489 	kstat_named_t controller;
3490 	kstat_named_t factory_mac_address;
3491 };
3492 
3493 /*
3494  * cxgbe:X:info
3495  */
3496 struct cxgbe_port_info_kstats {
3497 	kstat_named_t transceiver;
3498 	kstat_named_t rx_ovflow0;
3499 	kstat_named_t rx_ovflow1;
3500 	kstat_named_t rx_ovflow2;
3501 	kstat_named_t rx_ovflow3;
3502 	kstat_named_t rx_trunc0;
3503 	kstat_named_t rx_trunc1;
3504 	kstat_named_t rx_trunc2;
3505 	kstat_named_t rx_trunc3;
3506 	kstat_named_t tx_pause;
3507 	kstat_named_t rx_pause;
3508 };
3509 
3510 static kstat_t *
3511 setup_port_config_kstats(struct port_info *pi)
3512 {
3513 	kstat_t *ksp;
3514 	struct cxgbe_port_config_kstats *kstatp;
3515 	int ndata;
3516 	dev_info_t *pdip = ddi_get_parent(pi->dip);
3517 	uint8_t *ma = &pi->hw_addr[0];
3518 
3519 	ndata = sizeof (struct cxgbe_port_config_kstats) /
3520 	    sizeof (kstat_named_t);
3521 
3522 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), "config",
3523 	    "net", KSTAT_TYPE_NAMED, ndata, 0);
3524 	if (ksp == NULL) {
3525 		cxgb_printf(pi->dip, CE_WARN, "failed to initialize kstats.");
3526 		return (NULL);
3527 	}
3528 
3529 	kstatp = (struct cxgbe_port_config_kstats *)ksp->ks_data;
3530 
3531 	KS_UINIT(idx);
3532 	KS_UINIT(nrxq);
3533 	KS_UINIT(ntxq);
3534 	KS_UINIT(first_rxq);
3535 	KS_UINIT(first_txq);
3536 	KS_CINIT(controller);
3537 	KS_CINIT(factory_mac_address);
3538 
3539 	KS_U_SET(idx, pi->port_id);
3540 	KS_U_SET(nrxq, pi->nrxq);
3541 	KS_U_SET(ntxq, pi->ntxq);
3542 	KS_U_SET(first_rxq, pi->first_rxq);
3543 	KS_U_SET(first_txq, pi->first_txq);
3544 	KS_C_SET(controller, "%s%d", ddi_driver_name(pdip),
3545 	    ddi_get_instance(pdip));
3546 	KS_C_SET(factory_mac_address, "%02X%02X%02X%02X%02X%02X",
3547 	    ma[0], ma[1], ma[2], ma[3], ma[4], ma[5]);
3548 
3549 	/* Do NOT set ksp->ks_update.  These kstats do not change. */
3550 
3551 	/* Install the kstat */
3552 	ksp->ks_private = (void *)pi;
3553 	kstat_install(ksp);
3554 
3555 	return (ksp);
3556 }
3557 
3558 static kstat_t *
3559 setup_port_info_kstats(struct port_info *pi)
3560 {
3561 	kstat_t *ksp;
3562 	struct cxgbe_port_info_kstats *kstatp;
3563 	int ndata;
3564 
3565 	ndata = sizeof (struct cxgbe_port_info_kstats) / sizeof (kstat_named_t);
3566 
3567 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), "info",
3568 	    "net", KSTAT_TYPE_NAMED, ndata, 0);
3569 	if (ksp == NULL) {
3570 		cxgb_printf(pi->dip, CE_WARN, "failed to initialize kstats.");
3571 		return (NULL);
3572 	}
3573 
3574 	kstatp = (struct cxgbe_port_info_kstats *)ksp->ks_data;
3575 
3576 	KS_CINIT(transceiver);
3577 	KS_UINIT(rx_ovflow0);
3578 	KS_UINIT(rx_ovflow1);
3579 	KS_UINIT(rx_ovflow2);
3580 	KS_UINIT(rx_ovflow3);
3581 	KS_UINIT(rx_trunc0);
3582 	KS_UINIT(rx_trunc1);
3583 	KS_UINIT(rx_trunc2);
3584 	KS_UINIT(rx_trunc3);
3585 	KS_UINIT(tx_pause);
3586 	KS_UINIT(rx_pause);
3587 
3588 	/* Install the kstat */
3589 	ksp->ks_update = update_port_info_kstats;
3590 	ksp->ks_private = (void *)pi;
3591 	kstat_install(ksp);
3592 
3593 	return (ksp);
3594 }
3595 
3596 static int
3597 update_port_info_kstats(kstat_t *ksp, int rw)
3598 {
3599 	struct cxgbe_port_info_kstats *kstatp =
3600 	    (struct cxgbe_port_info_kstats *)ksp->ks_data;
3601 	struct port_info *pi = ksp->ks_private;
3602 	static const char *mod_str[] = { NULL, "LR", "SR", "ER", "TWINAX",
3603 	    "active TWINAX", "LRM" };
3604 	uint32_t bgmap;
3605 
3606 	if (rw == KSTAT_WRITE)
3607 		return (0);
3608 
3609 	if (pi->mod_type == FW_PORT_MOD_TYPE_NONE)
3610 		KS_C_SET(transceiver, "unplugged");
3611 	else if (pi->mod_type == FW_PORT_MOD_TYPE_UNKNOWN)
3612 		KS_C_SET(transceiver, "unknown");
3613 	else if (pi->mod_type == FW_PORT_MOD_TYPE_NOTSUPPORTED)
3614 		KS_C_SET(transceiver, "unsupported");
3615 	else if (pi->mod_type > 0 && pi->mod_type < ARRAY_SIZE(mod_str))
3616 		KS_C_SET(transceiver, "%s", mod_str[pi->mod_type]);
3617 	else
3618 		KS_C_SET(transceiver, "type %d", pi->mod_type);
3619 
3620 #define	GET_STAT(name) t4_read_reg64(pi->adapter, \
3621 	    PORT_REG(pi->port_id, A_MPS_PORT_STAT_##name##_L))
3622 #define	GET_STAT_COM(name) t4_read_reg64(pi->adapter, \
3623 	    A_MPS_STAT_##name##_L)
3624 
3625 	bgmap = G_NUMPORTS(t4_read_reg(pi->adapter, A_MPS_CMN_CTL));
3626 	if (bgmap == 0)
3627 		bgmap = (pi->port_id == 0) ? 0xf : 0;
3628 	else if (bgmap == 1)
3629 		bgmap = (pi->port_id < 2) ? (3 << (2 * pi->port_id)) : 0;
3630 	else
3631 		bgmap = 1;
3632 
3633 	KS_U_SET(rx_ovflow0, (bgmap & 1) ?
3634 	    GET_STAT_COM(RX_BG_0_MAC_DROP_FRAME) : 0);
3635 	KS_U_SET(rx_ovflow1, (bgmap & 2) ?
3636 	    GET_STAT_COM(RX_BG_1_MAC_DROP_FRAME) : 0);
3637 	KS_U_SET(rx_ovflow2, (bgmap & 4) ?
3638 	    GET_STAT_COM(RX_BG_2_MAC_DROP_FRAME) : 0);
3639 	KS_U_SET(rx_ovflow3, (bgmap & 8) ?
3640 	    GET_STAT_COM(RX_BG_3_MAC_DROP_FRAME) : 0);
3641 	KS_U_SET(rx_trunc0,  (bgmap & 1) ?
3642 	    GET_STAT_COM(RX_BG_0_MAC_TRUNC_FRAME) : 0);
3643 	KS_U_SET(rx_trunc1,  (bgmap & 2) ?
3644 	    GET_STAT_COM(RX_BG_1_MAC_TRUNC_FRAME) : 0);
3645 	KS_U_SET(rx_trunc2,  (bgmap & 4) ?
3646 	    GET_STAT_COM(RX_BG_2_MAC_TRUNC_FRAME) : 0);
3647 	KS_U_SET(rx_trunc3,  (bgmap & 8) ?
3648 	    GET_STAT_COM(RX_BG_3_MAC_TRUNC_FRAME) : 0);
3649 
3650 	KS_U_SET(tx_pause, GET_STAT(TX_PORT_PAUSE));
3651 	KS_U_SET(rx_pause, GET_STAT(RX_PORT_PAUSE));
3652 
3653 	return (0);
3654 
3655 }
3656 
3657 /*
3658  * cxgbe:X:rxqY
3659  */
3660 struct rxq_kstats {
3661 	kstat_named_t rxcsum;
3662 	kstat_named_t rxpkts;
3663 	kstat_named_t rxbytes;
3664 	kstat_named_t nomem;
3665 };
3666 
3667 static kstat_t *
3668 setup_rxq_kstats(struct port_info *pi, struct sge_rxq *rxq, int idx)
3669 {
3670 	struct kstat *ksp;
3671 	struct rxq_kstats *kstatp;
3672 	int ndata;
3673 	char str[16];
3674 
3675 	ndata = sizeof (struct rxq_kstats) / sizeof (kstat_named_t);
3676 	(void) snprintf(str, sizeof (str), "rxq%u", idx);
3677 
3678 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), str, "rxq",
3679 	    KSTAT_TYPE_NAMED, ndata, 0);
3680 	if (ksp == NULL) {
3681 		cxgb_printf(pi->dip, CE_WARN,
3682 		    "%s: failed to initialize rxq kstats for queue %d.",
3683 		    __func__, idx);
3684 		return (NULL);
3685 	}
3686 
3687 	kstatp = (struct rxq_kstats *)ksp->ks_data;
3688 
3689 	KS_UINIT(rxcsum);
3690 	KS_UINIT(rxpkts);
3691 	KS_UINIT(rxbytes);
3692 	KS_UINIT(nomem);
3693 
3694 	ksp->ks_update = update_rxq_kstats;
3695 	ksp->ks_private = (void *)rxq;
3696 	kstat_install(ksp);
3697 
3698 	return (ksp);
3699 }
3700 
3701 static int
3702 update_rxq_kstats(kstat_t *ksp, int rw)
3703 {
3704 	struct rxq_kstats *kstatp = (struct rxq_kstats *)ksp->ks_data;
3705 	struct sge_rxq *rxq = ksp->ks_private;
3706 
3707 	if (rw == KSTAT_WRITE)
3708 		return (0);
3709 
3710 	KS_U_FROM(rxcsum, rxq);
3711 	KS_U_FROM(rxpkts, rxq);
3712 	KS_U_FROM(rxbytes, rxq);
3713 	KS_U_FROM(nomem, rxq);
3714 
3715 	return (0);
3716 }
3717 
3718 /*
3719  * cxgbe:X:txqY
3720  */
3721 struct txq_kstats {
3722 	kstat_named_t txcsum;
3723 	kstat_named_t tso_wrs;
3724 	kstat_named_t imm_wrs;
3725 	kstat_named_t sgl_wrs;
3726 	kstat_named_t txpkt_wrs;
3727 	kstat_named_t txpkts_wrs;
3728 	kstat_named_t txpkts_pkts;
3729 	kstat_named_t txb_used;
3730 	kstat_named_t hdl_used;
3731 	kstat_named_t txb_full;
3732 	kstat_named_t dma_hdl_failed;
3733 	kstat_named_t dma_map_failed;
3734 	kstat_named_t qfull;
3735 	kstat_named_t qflush;
3736 	kstat_named_t pullup_early;
3737 	kstat_named_t pullup_late;
3738 	kstat_named_t pullup_failed;
3739 };
3740 
3741 static kstat_t *
3742 setup_txq_kstats(struct port_info *pi, struct sge_txq *txq, int idx)
3743 {
3744 	struct kstat *ksp;
3745 	struct txq_kstats *kstatp;
3746 	int ndata;
3747 	char str[16];
3748 
3749 	ndata = sizeof (struct txq_kstats) / sizeof (kstat_named_t);
3750 	(void) snprintf(str, sizeof (str), "txq%u", idx);
3751 
3752 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), str, "txq",
3753 	    KSTAT_TYPE_NAMED, ndata, 0);
3754 	if (ksp == NULL) {
3755 		cxgb_printf(pi->dip, CE_WARN,
3756 		    "%s: failed to initialize txq kstats for queue %d.",
3757 		    __func__, idx);
3758 		return (NULL);
3759 	}
3760 
3761 	kstatp = (struct txq_kstats *)ksp->ks_data;
3762 
3763 	KS_UINIT(txcsum);
3764 	KS_UINIT(tso_wrs);
3765 	KS_UINIT(imm_wrs);
3766 	KS_UINIT(sgl_wrs);
3767 	KS_UINIT(txpkt_wrs);
3768 	KS_UINIT(txpkts_wrs);
3769 	KS_UINIT(txpkts_pkts);
3770 	KS_UINIT(txb_used);
3771 	KS_UINIT(hdl_used);
3772 	KS_UINIT(txb_full);
3773 	KS_UINIT(dma_hdl_failed);
3774 	KS_UINIT(dma_map_failed);
3775 	KS_UINIT(qfull);
3776 	KS_UINIT(qflush);
3777 	KS_UINIT(pullup_early);
3778 	KS_UINIT(pullup_late);
3779 	KS_UINIT(pullup_failed);
3780 
3781 	ksp->ks_update = update_txq_kstats;
3782 	ksp->ks_private = (void *)txq;
3783 	kstat_install(ksp);
3784 
3785 	return (ksp);
3786 }
3787 
3788 static int
3789 update_txq_kstats(kstat_t *ksp, int rw)
3790 {
3791 	struct txq_kstats *kstatp = (struct txq_kstats *)ksp->ks_data;
3792 	struct sge_txq *txq = ksp->ks_private;
3793 
3794 	if (rw == KSTAT_WRITE)
3795 		return (0);
3796 
3797 	KS_U_FROM(txcsum, txq);
3798 	KS_U_FROM(tso_wrs, txq);
3799 	KS_U_FROM(imm_wrs, txq);
3800 	KS_U_FROM(sgl_wrs, txq);
3801 	KS_U_FROM(txpkt_wrs, txq);
3802 	KS_U_FROM(txpkts_wrs, txq);
3803 	KS_U_FROM(txpkts_pkts, txq);
3804 	KS_U_FROM(txb_used, txq);
3805 	KS_U_FROM(hdl_used, txq);
3806 	KS_U_FROM(txb_full, txq);
3807 	KS_U_FROM(dma_hdl_failed, txq);
3808 	KS_U_FROM(dma_map_failed, txq);
3809 	KS_U_FROM(qfull, txq);
3810 	KS_U_FROM(qflush, txq);
3811 	KS_U_FROM(pullup_early, txq);
3812 	KS_U_FROM(pullup_late, txq);
3813 	KS_U_FROM(pullup_failed, txq);
3814 
3815 	return (0);
3816 }
3817