xref: /illumos-gate/usr/src/uts/common/io/cxgbe/t4nex/t4_sge.c (revision b1529121add3ff25fc7b58196363a5439c0f6b67)
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 2024 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 	mac_ether_offload_info_t meoi;	/* pkt hdr info for offloads */
69 	struct ulptx_sgl sgl __attribute__((aligned(8)));
70 	struct ulptx_sge_pair reserved[TX_SGL_SEGS / 2];
71 };
72 
73 static int service_iq(struct sge_iq *iq, int budget);
74 static inline void init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx,
75     int8_t pktc_idx, int qsize, uint8_t esize);
76 static inline void init_fl(struct sge_fl *fl, uint16_t qsize);
77 static inline void init_eq(struct adapter *sc, struct sge_eq *eq,
78     uint16_t eqtype, uint16_t qsize,uint8_t tx_chan, uint16_t iqid);
79 static int alloc_iq_fl(struct port_info *pi, struct sge_iq *iq,
80     struct sge_fl *fl, int intr_idx, int cong);
81 static int free_iq_fl(struct port_info *pi, struct sge_iq *iq,
82     struct sge_fl *fl);
83 static int alloc_fwq(struct adapter *sc);
84 static int free_fwq(struct adapter *sc);
85 #ifdef TCP_OFFLOAD_ENABLE
86 static int alloc_mgmtq(struct adapter *sc);
87 #endif
88 static int alloc_rxq(struct port_info *pi, struct sge_rxq *rxq, int intr_idx,
89     int i);
90 static int free_rxq(struct port_info *pi, struct sge_rxq *rxq);
91 #ifdef TCP_OFFLOAD_ENABLE
92 static int alloc_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq,
93 	int intr_idx);
94 static int free_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq);
95 #endif
96 static int ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq);
97 static int eth_eq_alloc(struct adapter *sc, struct port_info *pi,
98     struct sge_eq *eq);
99 #ifdef TCP_OFFLOAD_ENABLE
100 static int ofld_eq_alloc(struct adapter *sc, struct port_info *pi,
101     struct sge_eq *eq);
102 #endif
103 static int alloc_eq(struct adapter *sc, struct port_info *pi,
104     struct sge_eq *eq);
105 static int free_eq(struct adapter *sc, struct sge_eq *eq);
106 #ifdef TCP_OFFLOAD_ENABLE
107 static int alloc_wrq(struct adapter *sc, struct port_info *pi,
108     struct sge_wrq *wrq, int idx);
109 static int free_wrq(struct adapter *sc, struct sge_wrq *wrq);
110 #endif
111 static int alloc_txq(struct port_info *pi, struct sge_txq *txq, int idx);
112 static int free_txq(struct port_info *pi, struct sge_txq *txq);
113 static int alloc_dma_memory(struct adapter *sc, size_t len, int flags,
114     ddi_device_acc_attr_t *acc_attr, ddi_dma_attr_t *dma_attr,
115     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl, uint64_t *pba,
116     caddr_t *pva);
117 static int free_dma_memory(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl);
118 static int alloc_desc_ring(struct adapter *sc, size_t len, int rw,
119     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl, uint64_t *pba,
120     caddr_t *pva);
121 static int free_desc_ring(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl);
122 static int alloc_tx_copybuffer(struct adapter *sc, size_t len,
123     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl, uint64_t *pba,
124     caddr_t *pva);
125 static inline bool is_new_response(const struct sge_iq *iq,
126     struct rsp_ctrl **ctrl);
127 static inline void iq_next(struct sge_iq *iq);
128 static int refill_fl(struct adapter *sc, struct sge_fl *fl, int nbufs);
129 static void refill_sfl(void *arg);
130 static void add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl);
131 static void free_fl_bufs(struct sge_fl *fl);
132 static mblk_t *get_fl_payload(struct adapter *sc, struct sge_fl *fl,
133     uint32_t len_newbuf, int *fl_bufs_used);
134 static int get_frame_txinfo(struct sge_txq *txq, mblk_t **fp,
135     struct txinfo *txinfo, int sgl_only);
136 static inline int fits_in_txb(struct sge_txq *txq, int len, int *waste);
137 static inline int copy_into_txb(struct sge_txq *txq, mblk_t *m, int len,
138     struct txinfo *txinfo);
139 static inline void add_seg(struct txinfo *txinfo, uint64_t ba, uint32_t len);
140 static inline int add_mblk(struct sge_txq *txq, struct txinfo *txinfo,
141     mblk_t *m, int len);
142 static void free_txinfo_resources(struct sge_txq *txq, struct txinfo *txinfo);
143 static int add_to_txpkts(struct sge_txq *txq, struct txpkts *txpkts, mblk_t *m,
144     struct txinfo *txinfo);
145 static void write_txpkts_wr(struct sge_txq *txq, struct txpkts *txpkts);
146 static int write_txpkt_wr(struct port_info *pi, struct sge_txq *txq, mblk_t *m,
147     struct txinfo *txinfo);
148 static inline void write_ulp_cpl_sgl(struct port_info *pi, struct sge_txq *txq,
149     struct txpkts *txpkts, struct txinfo *txinfo);
150 static inline void copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to,
151     int len);
152 static inline void ring_tx_db(struct adapter *sc, struct sge_eq *eq);
153 static int reclaim_tx_descs(struct sge_txq *txq, int howmany);
154 static void write_txqflush_wr(struct sge_txq *txq);
155 static int t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss,
156     mblk_t *m);
157 static inline void ring_fl_db(struct adapter *sc, struct sge_fl *fl);
158 static kstat_t *setup_port_config_kstats(struct port_info *pi);
159 static kstat_t *setup_port_info_kstats(struct port_info *pi);
160 static kstat_t *setup_rxq_kstats(struct port_info *pi, struct sge_rxq *rxq,
161     int idx);
162 static int update_rxq_kstats(kstat_t *ksp, int rw);
163 static int update_port_info_kstats(kstat_t *ksp, int rw);
164 static kstat_t *setup_txq_kstats(struct port_info *pi, struct sge_txq *txq,
165     int idx);
166 static int update_txq_kstats(kstat_t *ksp, int rw);
167 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
168     mblk_t *);
169 static int handle_fw_rpl(struct sge_iq *iq, const struct rss_header *rss,
170     mblk_t *m);
171 
172 static inline int
173 reclaimable(struct sge_eq *eq)
174 {
175 	unsigned int cidx;
176 
177 	cidx = eq->spg->cidx;   /* stable snapshot */
178 	cidx = be16_to_cpu(cidx);
179 
180 	if (cidx >= eq->cidx)
181 		return (cidx - eq->cidx);
182 	else
183 		return (cidx + eq->cap - eq->cidx);
184 }
185 
186 void
187 t4_sge_init(struct adapter *sc)
188 {
189 	struct driver_properties *p = &sc->props;
190 	ddi_dma_attr_t *dma_attr;
191 	ddi_device_acc_attr_t *acc_attr;
192 	uint32_t sge_control, sge_conm_ctrl;
193 	int egress_threshold;
194 
195 	/*
196 	 * Device access and DMA attributes for descriptor rings
197 	 */
198 	acc_attr = &sc->sge.acc_attr_desc;
199 	acc_attr->devacc_attr_version = DDI_DEVICE_ATTR_V0;
200 	acc_attr->devacc_attr_endian_flags = DDI_NEVERSWAP_ACC;
201 	acc_attr->devacc_attr_dataorder = DDI_STRICTORDER_ACC;
202 
203 	dma_attr = &sc->sge.dma_attr_desc;
204 	dma_attr->dma_attr_version = DMA_ATTR_V0;
205 	dma_attr->dma_attr_addr_lo = 0;
206 	dma_attr->dma_attr_addr_hi = UINT64_MAX;
207 	dma_attr->dma_attr_count_max = UINT64_MAX;
208 	dma_attr->dma_attr_align = 512;
209 	dma_attr->dma_attr_burstsizes = 0xfff;
210 	dma_attr->dma_attr_minxfer = 1;
211 	dma_attr->dma_attr_maxxfer = UINT64_MAX;
212 	dma_attr->dma_attr_seg = UINT64_MAX;
213 	dma_attr->dma_attr_sgllen = 1;
214 	dma_attr->dma_attr_granular = 1;
215 	dma_attr->dma_attr_flags = 0;
216 
217 	/*
218 	 * Device access and DMA attributes for tx buffers
219 	 */
220 	acc_attr = &sc->sge.acc_attr_tx;
221 	acc_attr->devacc_attr_version = DDI_DEVICE_ATTR_V0;
222 	acc_attr->devacc_attr_endian_flags = DDI_NEVERSWAP_ACC;
223 
224 	dma_attr = &sc->sge.dma_attr_tx;
225 	dma_attr->dma_attr_version = DMA_ATTR_V0;
226 	dma_attr->dma_attr_addr_lo = 0;
227 	dma_attr->dma_attr_addr_hi = UINT64_MAX;
228 	dma_attr->dma_attr_count_max = UINT64_MAX;
229 	dma_attr->dma_attr_align = 1;
230 	dma_attr->dma_attr_burstsizes = 0xfff;
231 	dma_attr->dma_attr_minxfer = 1;
232 	dma_attr->dma_attr_maxxfer = UINT64_MAX;
233 	dma_attr->dma_attr_seg = UINT64_MAX;
234 	dma_attr->dma_attr_sgllen = TX_SGL_SEGS;
235 	dma_attr->dma_attr_granular = 1;
236 	dma_attr->dma_attr_flags = 0;
237 
238 	/*
239 	 * Ingress Padding Boundary and Egress Status Page Size are set up by
240 	 * t4_fixup_host_params().
241 	 */
242 	sge_control = t4_read_reg(sc, A_SGE_CONTROL);
243 	sc->sge.pktshift = G_PKTSHIFT(sge_control);
244 	sc->sge.stat_len = (sge_control & F_EGRSTATUSPAGESIZE) ? 128 : 64;
245 
246 	/* t4_nex uses FLM packed mode */
247 	sc->sge.fl_align = t4_fl_pkt_align(sc, true);
248 
249 	/*
250 	 * Device access and DMA attributes for rx buffers
251 	 */
252 	sc->sge.rxb_params.dip = sc->dip;
253 	sc->sge.rxb_params.buf_size = rx_buf_size;
254 
255 	acc_attr = &sc->sge.rxb_params.acc_attr_rx;
256 	acc_attr->devacc_attr_version = DDI_DEVICE_ATTR_V0;
257 	acc_attr->devacc_attr_endian_flags = DDI_NEVERSWAP_ACC;
258 
259 	dma_attr = &sc->sge.rxb_params.dma_attr_rx;
260 	dma_attr->dma_attr_version = DMA_ATTR_V0;
261 	dma_attr->dma_attr_addr_lo = 0;
262 	dma_attr->dma_attr_addr_hi = UINT64_MAX;
263 	dma_attr->dma_attr_count_max = UINT64_MAX;
264 	/*
265 	 * Low 4 bits of an rx buffer address have a special meaning to the SGE
266 	 * and an rx buf cannot have an address with any of these bits set.
267 	 * FL_ALIGN is >= 32 so we're sure things are ok.
268 	 */
269 	dma_attr->dma_attr_align = sc->sge.fl_align;
270 	dma_attr->dma_attr_burstsizes = 0xfff;
271 	dma_attr->dma_attr_minxfer = 1;
272 	dma_attr->dma_attr_maxxfer = UINT64_MAX;
273 	dma_attr->dma_attr_seg = UINT64_MAX;
274 	dma_attr->dma_attr_sgllen = 1;
275 	dma_attr->dma_attr_granular = 1;
276 	dma_attr->dma_attr_flags = 0;
277 
278 	sc->sge.rxbuf_cache = rxbuf_cache_create(&sc->sge.rxb_params);
279 
280 	/*
281 	 * A FL with <= fl_starve_thres buffers is starving and a periodic
282 	 * timer will attempt to refill it.  This needs to be larger than the
283 	 * SGE's Egress Congestion Threshold.  If it isn't, then we can get
284 	 * stuck waiting for new packets while the SGE is waiting for us to
285 	 * give it more Free List entries.  (Note that the SGE's Egress
286 	 * Congestion Threshold is in units of 2 Free List pointers.) For T4,
287 	 * there was only a single field to control this.  For T5 there's the
288 	 * original field which now only applies to Unpacked Mode Free List
289 	 * buffers and a new field which only applies to Packed Mode Free List
290 	 * buffers.
291 	 */
292 
293 	sge_conm_ctrl = t4_read_reg(sc, A_SGE_CONM_CTRL);
294 	switch (CHELSIO_CHIP_VERSION(sc->params.chip)) {
295 	case CHELSIO_T4:
296 		egress_threshold = G_EGRTHRESHOLD(sge_conm_ctrl);
297 		break;
298 	case CHELSIO_T5:
299 		egress_threshold = G_EGRTHRESHOLDPACKING(sge_conm_ctrl);
300 		break;
301 	case CHELSIO_T6:
302 	default:
303 		egress_threshold = G_T6_EGRTHRESHOLDPACKING(sge_conm_ctrl);
304 	}
305 	sc->sge.fl_starve_threshold = 2*egress_threshold + 1;
306 
307 	t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0, rx_buf_size);
308 
309 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD,
310 	    V_THRESHOLD_0(p->counter_val[0]) |
311 	    V_THRESHOLD_1(p->counter_val[1]) |
312 	    V_THRESHOLD_2(p->counter_val[2]) |
313 	    V_THRESHOLD_3(p->counter_val[3]));
314 
315 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1,
316 	    V_TIMERVALUE0(us_to_core_ticks(sc, p->timer_val[0])) |
317 	    V_TIMERVALUE1(us_to_core_ticks(sc, p->timer_val[1])));
318 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3,
319 	    V_TIMERVALUE2(us_to_core_ticks(sc, p->timer_val[2])) |
320 	    V_TIMERVALUE3(us_to_core_ticks(sc, p->timer_val[3])));
321 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5,
322 	    V_TIMERVALUE4(us_to_core_ticks(sc, p->timer_val[4])) |
323 	    V_TIMERVALUE5(us_to_core_ticks(sc, p->timer_val[5])));
324 
325 	(void) t4_register_cpl_handler(sc, CPL_FW4_MSG, handle_fw_rpl);
326 	(void) t4_register_cpl_handler(sc, CPL_FW6_MSG, handle_fw_rpl);
327 	(void) t4_register_cpl_handler(sc, CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
328 	(void) t4_register_cpl_handler(sc, CPL_RX_PKT, t4_eth_rx);
329 	(void) t4_register_fw_msg_handler(sc, FW6_TYPE_CMD_RPL,
330 		    t4_handle_fw_rpl);
331 }
332 
333 /*
334  * Allocate and initialize the firmware event queue and the forwarded interrupt
335  * queues, if any.  The adapter owns all these queues as they are not associated
336  * with any particular port.
337  *
338  * Returns errno on failure.  Resources allocated up to that point may still be
339  * allocated.  Caller is responsible for cleanup in case this function fails.
340  */
341 int
342 t4_setup_adapter_queues(struct adapter *sc)
343 {
344 	int rc;
345 
346 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
347 
348 	/*
349 	 * Firmware event queue
350 	 */
351 	rc = alloc_fwq(sc);
352 	if (rc != 0)
353 		return (rc);
354 
355 #ifdef TCP_OFFLOAD_ENABLE
356 	/*
357 	 * Management queue.  This is just a control queue that uses the fwq as
358 	 * its associated iq.
359 	 */
360 	rc = alloc_mgmtq(sc);
361 #endif
362 
363 	return (rc);
364 }
365 
366 /*
367  * Idempotent
368  */
369 int
370 t4_teardown_adapter_queues(struct adapter *sc)
371 {
372 
373 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
374 
375 	(void) free_fwq(sc);
376 
377 	return (0);
378 }
379 
380 static inline int
381 first_vector(struct port_info *pi)
382 {
383 	struct adapter *sc = pi->adapter;
384 	int rc = T4_EXTRA_INTR, i;
385 
386 	if (sc->intr_count == 1)
387 		return (0);
388 
389 	for_each_port(sc, i) {
390 		struct port_info *p = sc->port[i];
391 
392 		if (i == pi->port_id)
393 			break;
394 
395 #ifdef TCP_OFFLOAD_ENABLE
396 		if (!(sc->flags & INTR_FWD))
397 			rc += p->nrxq + p->nofldrxq;
398 		else
399 			rc += max(p->nrxq, p->nofldrxq);
400 #else
401 		/*
402 		 * Not compiled with offload support and intr_count > 1.  Only
403 		 * NIC queues exist and they'd better be taking direct
404 		 * interrupts.
405 		 */
406 		ASSERT(!(sc->flags & INTR_FWD));
407 		rc += p->nrxq;
408 #endif
409 	}
410 	return (rc);
411 }
412 
413 /*
414  * Given an arbitrary "index," come up with an iq that can be used by other
415  * queues (of this port) for interrupt forwarding, SGE egress updates, etc.
416  * The iq returned is guaranteed to be something that takes direct interrupts.
417  */
418 static struct sge_iq *
419 port_intr_iq(struct port_info *pi, int idx)
420 {
421 	struct adapter *sc = pi->adapter;
422 	struct sge *s = &sc->sge;
423 	struct sge_iq *iq = NULL;
424 
425 	if (sc->intr_count == 1)
426 		return (&sc->sge.fwq);
427 
428 #ifdef TCP_OFFLOAD_ENABLE
429 	if (!(sc->flags & INTR_FWD)) {
430 		idx %= pi->nrxq + pi->nofldrxq;
431 
432 		if (idx >= pi->nrxq) {
433 			idx -= pi->nrxq;
434 			iq = &s->ofld_rxq[pi->first_ofld_rxq + idx].iq;
435 		} else
436 			iq = &s->rxq[pi->first_rxq + idx].iq;
437 
438 	} else {
439 		idx %= max(pi->nrxq, pi->nofldrxq);
440 
441 		if (pi->nrxq >= pi->nofldrxq)
442 			iq = &s->rxq[pi->first_rxq + idx].iq;
443 		else
444 			iq = &s->ofld_rxq[pi->first_ofld_rxq + idx].iq;
445 	}
446 #else
447 	/*
448 	 * Not compiled with offload support and intr_count > 1.  Only NIC
449 	 * queues exist and they'd better be taking direct interrupts.
450 	 */
451 	ASSERT(!(sc->flags & INTR_FWD));
452 
453 	idx %= pi->nrxq;
454 	iq = &s->rxq[pi->first_rxq + idx].iq;
455 #endif
456 
457 	return (iq);
458 }
459 
460 int
461 t4_setup_port_queues(struct port_info *pi)
462 {
463 	int rc = 0, i, intr_idx, j;
464 	struct sge_rxq *rxq;
465 	struct sge_txq *txq;
466 #ifdef TCP_OFFLOAD_ENABLE
467 	int iqid;
468 	struct sge_wrq *ctrlq;
469 	struct sge_ofld_rxq *ofld_rxq;
470 	struct sge_wrq *ofld_txq;
471 #endif
472 	struct adapter *sc = pi->adapter;
473 	struct driver_properties *p = &sc->props;
474 
475 	pi->ksp_config = setup_port_config_kstats(pi);
476 	pi->ksp_info   = setup_port_info_kstats(pi);
477 
478 	/* Interrupt vector to start from (when using multiple vectors) */
479 	intr_idx = first_vector(pi);
480 
481 	/*
482 	 * First pass over all rx queues (NIC and TOE):
483 	 * a) initialize iq and fl
484 	 * b) allocate queue iff it will take direct interrupts.
485 	 */
486 
487 	for_each_rxq(pi, i, rxq) {
488 
489 		init_iq(&rxq->iq, sc, pi->tmr_idx, pi->pktc_idx, p->qsize_rxq,
490 		    RX_IQ_ESIZE);
491 
492 		init_fl(&rxq->fl, p->qsize_rxq / 8); /* 8 bufs in each entry */
493 
494 		if ((!(sc->flags & INTR_FWD))
495 #ifdef TCP_OFFLOAD_ENABLE
496 		    || (sc->intr_count > 1 && pi->nrxq >= pi->nofldrxq)
497 #else
498 		    || (sc->intr_count > 1 && pi->nrxq)
499 #endif
500 		   ) {
501 			rxq->iq.flags |= IQ_INTR;
502 			rc = alloc_rxq(pi, rxq, intr_idx, i);
503 			if (rc != 0)
504 				goto done;
505 			intr_idx++;
506 		}
507 
508 	}
509 
510 #ifdef TCP_OFFLOAD_ENABLE
511 	for_each_ofld_rxq(pi, i, ofld_rxq) {
512 
513 		init_iq(&ofld_rxq->iq, sc, pi->tmr_idx, pi->pktc_idx,
514 		    p->qsize_rxq, RX_IQ_ESIZE);
515 
516 		init_fl(&ofld_rxq->fl, p->qsize_rxq / 8);
517 
518 		if (!(sc->flags & INTR_FWD) ||
519 		    (sc->intr_count > 1 && pi->nofldrxq > pi->nrxq)) {
520 			ofld_rxq->iq.flags = IQ_INTR;
521 			rc = alloc_ofld_rxq(pi, ofld_rxq, intr_idx);
522 			if (rc != 0)
523 				goto done;
524 
525 			intr_idx++;
526 		}
527 	}
528 #endif
529 
530 	/*
531 	 * Second pass over all rx queues (NIC and TOE).  The queues forwarding
532 	 * their interrupts are allocated now.
533 	 */
534 	j = 0;
535 	for_each_rxq(pi, i, rxq) {
536 		if (rxq->iq.flags & IQ_INTR)
537 			continue;
538 
539 		intr_idx = port_intr_iq(pi, j)->abs_id;
540 
541 		rc = alloc_rxq(pi, rxq, intr_idx, i);
542 		if (rc != 0)
543 			goto done;
544 		j++;
545 	}
546 
547 #ifdef TCP_OFFLOAD_ENABLE
548 	for_each_ofld_rxq(pi, i, ofld_rxq) {
549 		if (ofld_rxq->iq.flags & IQ_INTR)
550 			continue;
551 
552 		intr_idx = port_intr_iq(pi, j)->abs_id;
553 		rc = alloc_ofld_rxq(pi, ofld_rxq, intr_idx);
554 		if (rc != 0)
555 			goto done;
556 		j++;
557 	}
558 #endif
559 	/*
560 	 * Now the tx queues.  Only one pass needed.
561 	 */
562 	j = 0;
563 	for_each_txq(pi, i, txq) {
564 		uint16_t iqid;
565 
566 		iqid = port_intr_iq(pi, j)->cntxt_id;
567 		init_eq(sc, &txq->eq, EQ_ETH, p->qsize_txq, pi->tx_chan, iqid);
568 		rc = alloc_txq(pi, txq, i);
569 		if (rc != 0)
570 			goto done;
571 	}
572 
573 #ifdef TCP_OFFLOAD_ENABLE
574 	for_each_ofld_txq(pi, i, ofld_txq) {
575 		uint16_t iqid;
576 
577 		iqid = port_intr_iq(pi, j)->cntxt_id;
578 		init_eq(sc, &ofld_txq->eq, EQ_OFLD, p->qsize_txq, pi->tx_chan,
579 		    iqid);
580 		rc = alloc_wrq(sc, pi, ofld_txq, i);
581 		if (rc != 0)
582 			goto done;
583 	}
584 
585 	/*
586 	 * Finally, the control queue.
587 	 */
588 	ctrlq = &sc->sge.ctrlq[pi->port_id];
589 	iqid = port_intr_iq(pi, 0)->cntxt_id;
590 	init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, pi->tx_chan, iqid);
591 	rc = alloc_wrq(sc, pi, ctrlq, 0);
592 #endif
593 
594 done:
595 	if (rc != 0)
596 		(void) t4_teardown_port_queues(pi);
597 
598 	return (rc);
599 }
600 
601 /*
602  * Idempotent
603  */
604 int
605 t4_teardown_port_queues(struct port_info *pi)
606 {
607 	int i;
608 	struct sge_rxq *rxq;
609 	struct sge_txq *txq;
610 #ifdef TCP_OFFLOAD_ENABLE
611 	struct adapter *sc = pi->adapter;
612 	struct sge_ofld_rxq *ofld_rxq;
613 	struct sge_wrq *ofld_txq;
614 #endif
615 
616 	if (pi->ksp_config != NULL) {
617 		kstat_delete(pi->ksp_config);
618 		pi->ksp_config = NULL;
619 	}
620 	if (pi->ksp_info != NULL) {
621 		kstat_delete(pi->ksp_info);
622 		pi->ksp_info = NULL;
623 	}
624 
625 #ifdef TCP_OFFLOAD_ENABLE
626 	(void) free_wrq(sc, &sc->sge.ctrlq[pi->port_id]);
627 #endif
628 
629 	for_each_txq(pi, i, txq) {
630 		(void) free_txq(pi, txq);
631 	}
632 
633 #ifdef TCP_OFFLOAD_ENABLE
634 	for_each_ofld_txq(pi, i, ofld_txq) {
635 		(void) free_wrq(sc, ofld_txq);
636 	}
637 
638 	for_each_ofld_rxq(pi, i, ofld_rxq) {
639 		if ((ofld_rxq->iq.flags & IQ_INTR) == 0)
640 			(void) free_ofld_rxq(pi, ofld_rxq);
641 	}
642 #endif
643 
644 	for_each_rxq(pi, i, rxq) {
645 		if ((rxq->iq.flags & IQ_INTR) == 0)
646 			(void) free_rxq(pi, rxq);
647 	}
648 
649 	/*
650 	 * Then take down the rx queues that take direct interrupts.
651 	 */
652 
653 	for_each_rxq(pi, i, rxq) {
654 		if (rxq->iq.flags & IQ_INTR)
655 			(void) free_rxq(pi, rxq);
656 	}
657 
658 #ifdef TCP_OFFLOAD_ENABLE
659 	for_each_ofld_rxq(pi, i, ofld_rxq) {
660 		if (ofld_rxq->iq.flags & IQ_INTR)
661 			(void) free_ofld_rxq(pi, ofld_rxq);
662 	}
663 #endif
664 
665 	return (0);
666 }
667 
668 /* Deals with errors and forwarded interrupts */
669 uint_t
670 t4_intr_all(caddr_t arg1, caddr_t arg2)
671 {
672 
673 	(void) t4_intr_err(arg1, arg2);
674 	(void) t4_intr(arg1, arg2);
675 
676 	return (DDI_INTR_CLAIMED);
677 }
678 
679 static void
680 t4_intr_rx_work(struct sge_iq *iq)
681 {
682 	mblk_t *mp = NULL;
683 	struct sge_rxq *rxq = iq_to_rxq(iq);	/* Use iff iq is part of rxq */
684 	RXQ_LOCK(rxq);
685 	if (!iq->polling) {
686 		mp = t4_ring_rx(rxq, iq->qsize/8);
687 		t4_write_reg(iq->adapter, MYPF_REG(A_SGE_PF_GTS),
688 		     V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_next));
689 	}
690 	RXQ_UNLOCK(rxq);
691 	if (mp != NULL)
692 		mac_rx_ring(rxq->port->mh, rxq->ring_handle, mp,
693 			    rxq->ring_gen_num);
694 }
695 
696 /* Deals with interrupts on the given ingress queue */
697 /* ARGSUSED */
698 uint_t
699 t4_intr(caddr_t arg1, caddr_t arg2)
700 {
701 	struct sge_iq *iq = (struct sge_iq *)arg2;
702 	int state;
703 
704 	/* Right now receive polling is only enabled for MSI-X and
705 	 * when we have enough msi-x vectors i.e no interrupt forwarding.
706 	 */
707 	if (iq->adapter->props.multi_rings) {
708 		t4_intr_rx_work(iq);
709 	} else {
710 		state = atomic_cas_uint(&iq->state, IQS_IDLE, IQS_BUSY);
711 		if (state == IQS_IDLE) {
712 			(void) service_iq(iq, 0);
713 			(void) atomic_cas_uint(&iq->state, IQS_BUSY, IQS_IDLE);
714 		}
715 	}
716 	return (DDI_INTR_CLAIMED);
717 }
718 
719 /* Deals with error interrupts */
720 /* ARGSUSED */
721 uint_t
722 t4_intr_err(caddr_t arg1, caddr_t arg2)
723 {
724 	/* LINTED: E_BAD_PTR_CAST_ALIGN */
725 	struct adapter *sc = (struct adapter *)arg1;
726 
727 	t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
728 	(void) t4_slow_intr_handler(sc);
729 
730 	return (DDI_INTR_CLAIMED);
731 }
732 
733 /*
734  * t4_ring_rx - Process responses from an SGE response queue.
735  *
736  * This function processes responses from an SGE response queue up to the supplied budget.
737  * Responses include received packets as well as control messages from FW
738  * or HW.
739  * It returns a chain of mblks containing the received data, to be
740  * passed up to mac_ring_rx().
741  */
742 mblk_t *
743 t4_ring_rx(struct sge_rxq *rxq, int budget)
744 {
745 	struct sge_iq *iq = &rxq->iq;
746 	struct sge_fl *fl = &rxq->fl;           /* Use iff IQ_HAS_FL */
747 	struct adapter *sc = iq->adapter;
748 	struct rsp_ctrl *ctrl;
749 	const struct rss_header *rss;
750 	int ndescs = 0, fl_bufs_used = 0;
751 	int rsp_type;
752 	uint32_t lq;
753 	mblk_t *mblk_head = NULL, **mblk_tail, *m;
754 	struct cpl_rx_pkt *cpl;
755 	uint32_t received_bytes = 0, pkt_len = 0;
756 	bool csum_ok;
757 	uint16_t err_vec;
758 
759 	mblk_tail = &mblk_head;
760 
761 	while (is_new_response(iq, &ctrl)) {
762 
763 		membar_consumer();
764 
765 		m = NULL;
766 		rsp_type = G_RSPD_TYPE(ctrl->u.type_gen);
767 		lq = be32_to_cpu(ctrl->pldbuflen_qid);
768 		rss = (const void *)iq->cdesc;
769 
770 		switch (rsp_type) {
771 		case X_RSPD_TYPE_FLBUF:
772 
773 			ASSERT(iq->flags & IQ_HAS_FL);
774 
775 			if (CPL_RX_PKT == rss->opcode) {
776 				cpl = (void *)(rss + 1);
777 				pkt_len = be16_to_cpu(cpl->len);
778 
779 				if (iq->polling && ((received_bytes + pkt_len) > budget))
780 					goto done;
781 
782 				m = get_fl_payload(sc, fl, lq, &fl_bufs_used);
783 				if (m == NULL)
784 					goto done;
785 
786 				iq->intr_next = iq->intr_params;
787 				m->b_rptr += sc->sge.pktshift;
788 				if (sc->params.tp.rx_pkt_encap)
789 				/* It is enabled only in T6 config file */
790 					err_vec = G_T6_COMPR_RXERR_VEC(ntohs(cpl->err_vec));
791 				else
792 					err_vec = ntohs(cpl->err_vec);
793 
794 				csum_ok = cpl->csum_calc && !err_vec;
795 
796 				/* TODO: what about cpl->ip_frag? */
797 				if (csum_ok && !cpl->ip_frag) {
798 					mac_hcksum_set(m, 0, 0, 0, 0xffff,
799 					    HCK_FULLCKSUM_OK | HCK_FULLCKSUM |
800 					    HCK_IPV4_HDRCKSUM_OK);
801 					rxq->rxcsum++;
802 				}
803 				rxq->rxpkts++;
804 				rxq->rxbytes += pkt_len;
805 				received_bytes += pkt_len;
806 
807 				*mblk_tail = m;
808 				mblk_tail = &m->b_next;
809 
810 				break;
811 			}
812 
813 			m = get_fl_payload(sc, fl, lq, &fl_bufs_used);
814 			if (m == NULL)
815 				goto done;
816 			/* FALLTHROUGH */
817 
818 		case X_RSPD_TYPE_CPL:
819 			ASSERT(rss->opcode < NUM_CPL_CMDS);
820 			sc->cpl_handler[rss->opcode](iq, rss, m);
821 			break;
822 
823 		default:
824 			break;
825 		}
826 		iq_next(iq);
827 		++ndescs;
828 		if (!iq->polling && (ndescs == budget))
829 			break;
830 	}
831 
832 done:
833 
834 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
835 		     V_CIDXINC(ndescs) | V_INGRESSQID(iq->cntxt_id) |
836 		     V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
837 
838 	if ((fl_bufs_used > 0) || (iq->flags & IQ_HAS_FL)) {
839 		int starved;
840 		FL_LOCK(fl);
841 		fl->needed += fl_bufs_used;
842 		starved = refill_fl(sc, fl, fl->cap / 8);
843 		FL_UNLOCK(fl);
844 		if (starved)
845 			add_fl_to_sfl(sc, fl);
846 	}
847 	return (mblk_head);
848 }
849 
850 /*
851  * Deals with anything and everything on the given ingress queue.
852  */
853 static int
854 service_iq(struct sge_iq *iq, int budget)
855 {
856 	struct sge_iq *q;
857 	struct sge_rxq *rxq = iq_to_rxq(iq);	/* Use iff iq is part of rxq */
858 	struct sge_fl *fl = &rxq->fl;		/* Use iff IQ_HAS_FL */
859 	struct adapter *sc = iq->adapter;
860 	struct rsp_ctrl *ctrl;
861 	const struct rss_header *rss;
862 	int ndescs = 0, limit, fl_bufs_used = 0;
863 	int rsp_type;
864 	uint32_t lq;
865 	int starved;
866 	mblk_t *m;
867 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
868 
869 	limit = budget ? budget : iq->qsize / 8;
870 
871 	/*
872 	 * We always come back and check the descriptor ring for new indirect
873 	 * interrupts and other responses after running a single handler.
874 	 */
875 	for (;;) {
876 		while (is_new_response(iq, &ctrl)) {
877 
878 			membar_consumer();
879 
880 			m = NULL;
881 			rsp_type = G_RSPD_TYPE(ctrl->u.type_gen);
882 			lq = be32_to_cpu(ctrl->pldbuflen_qid);
883 			rss = (const void *)iq->cdesc;
884 
885 			switch (rsp_type) {
886 			case X_RSPD_TYPE_FLBUF:
887 
888 				ASSERT(iq->flags & IQ_HAS_FL);
889 
890 				m = get_fl_payload(sc, fl, lq, &fl_bufs_used);
891 				if (m == NULL) {
892 					/*
893 					 * Rearm the iq with a
894 					 * longer-than-default timer
895 					 */
896 					t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_CIDXINC(ndescs) |
897 							V_INGRESSQID((u32)iq->cntxt_id) |
898 							V_SEINTARM(V_QINTR_TIMER_IDX(SGE_NTIMERS-1)));
899 					if (fl_bufs_used > 0) {
900 						ASSERT(iq->flags & IQ_HAS_FL);
901 						FL_LOCK(fl);
902 						fl->needed += fl_bufs_used;
903 						starved = refill_fl(sc, fl, fl->cap / 8);
904 						FL_UNLOCK(fl);
905 						if (starved)
906 							add_fl_to_sfl(sc, fl);
907 					}
908 					return (0);
909 				}
910 
911 			/* FALLTHRU */
912 			case X_RSPD_TYPE_CPL:
913 
914 				ASSERT(rss->opcode < NUM_CPL_CMDS);
915 				sc->cpl_handler[rss->opcode](iq, rss, m);
916 				break;
917 
918 			case X_RSPD_TYPE_INTR:
919 
920 				/*
921 				 * Interrupts should be forwarded only to queues
922 				 * that are not forwarding their interrupts.
923 				 * This means service_iq can recurse but only 1
924 				 * level deep.
925 				 */
926 				ASSERT(budget == 0);
927 
928 				q = sc->sge.iqmap[lq - sc->sge.iq_start];
929 				if (atomic_cas_uint(&q->state, IQS_IDLE,
930 				    IQS_BUSY) == IQS_IDLE) {
931 					if (service_iq(q, q->qsize / 8) == 0) {
932 						(void) atomic_cas_uint(
933 						    &q->state, IQS_BUSY,
934 						    IQS_IDLE);
935 					} else {
936 						STAILQ_INSERT_TAIL(&iql, q,
937 						    link);
938 					}
939 				}
940 				break;
941 
942 			default:
943 				break;
944 			}
945 
946 			iq_next(iq);
947 			if (++ndescs == limit) {
948 				t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
949 				    V_CIDXINC(ndescs) |
950 				    V_INGRESSQID(iq->cntxt_id) |
951 				    V_SEINTARM(V_QINTR_TIMER_IDX(
952 				    X_TIMERREG_UPDATE_CIDX)));
953 				ndescs = 0;
954 
955 				if (fl_bufs_used > 0) {
956 					ASSERT(iq->flags & IQ_HAS_FL);
957 					FL_LOCK(fl);
958 					fl->needed += fl_bufs_used;
959 					(void) refill_fl(sc, fl, fl->cap / 8);
960 					FL_UNLOCK(fl);
961 					fl_bufs_used = 0;
962 				}
963 
964 				if (budget != 0)
965 					return (EINPROGRESS);
966 			}
967 		}
968 
969 		if (STAILQ_EMPTY(&iql) != 0)
970 			break;
971 
972 		/*
973 		 * Process the head only, and send it to the back of the list if
974 		 * it's still not done.
975 		 */
976 		q = STAILQ_FIRST(&iql);
977 		STAILQ_REMOVE_HEAD(&iql, link);
978 		if (service_iq(q, q->qsize / 8) == 0)
979 			(void) atomic_cas_uint(&q->state, IQS_BUSY, IQS_IDLE);
980 		else
981 			STAILQ_INSERT_TAIL(&iql, q, link);
982 	}
983 
984 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_CIDXINC(ndescs) |
985 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_next));
986 
987 	if (iq->flags & IQ_HAS_FL) {
988 
989 		FL_LOCK(fl);
990 		fl->needed += fl_bufs_used;
991 		starved = refill_fl(sc, fl, fl->cap / 4);
992 		FL_UNLOCK(fl);
993 		if (starved != 0)
994 			add_fl_to_sfl(sc, fl);
995 	}
996 
997 	return (0);
998 }
999 
1000 #ifdef TCP_OFFLOAD_ENABLE
1001 int
1002 t4_mgmt_tx(struct adapter *sc, mblk_t *m)
1003 {
1004 	return (t4_wrq_tx(sc, &sc->sge.mgmtq, m));
1005 }
1006 
1007 /*
1008  * Doesn't fail.  Holds on to work requests it can't send right away.
1009  */
1010 int
1011 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, mblk_t *m0)
1012 {
1013 	struct sge_eq *eq = &wrq->eq;
1014 	struct mblk_pair *wr_list = &wrq->wr_list;
1015 	int can_reclaim;
1016 	caddr_t dst;
1017 	mblk_t *wr, *next;
1018 
1019 	TXQ_LOCK_ASSERT_OWNED(wrq);
1020 #ifdef TCP_OFFLOAD_ENABLE
1021 	ASSERT((eq->flags & EQ_TYPEMASK) == EQ_OFLD ||
1022 	    (eq->flags & EQ_TYPEMASK) == EQ_CTRL);
1023 #else
1024 	ASSERT((eq->flags & EQ_TYPEMASK) == EQ_CTRL);
1025 #endif
1026 
1027 	if (m0 != NULL) {
1028 		if (wr_list->head != NULL)
1029 			wr_list->tail->b_next = m0;
1030 		else
1031 			wr_list->head = m0;
1032 		while (m0->b_next)
1033 			m0 = m0->b_next;
1034 		wr_list->tail = m0;
1035 	}
1036 
1037 	can_reclaim = reclaimable(eq);
1038 	eq->cidx += can_reclaim;
1039 	eq->avail += can_reclaim;
1040 	if (eq->cidx >= eq->cap)
1041 		eq->cidx -= eq->cap;
1042 
1043 	for (wr = wr_list->head; wr; wr = next) {
1044 		int ndesc, len = 0;
1045 		mblk_t *m;
1046 
1047 		next = wr->b_next;
1048 		wr->b_next = NULL;
1049 
1050 		for (m = wr; m; m = m->b_cont)
1051 			len += MBLKL(m);
1052 
1053 		ASSERT(len > 0 && (len & 0x7) == 0);
1054 		ASSERT(len <= SGE_MAX_WR_LEN);
1055 
1056 		ndesc = howmany(len, EQ_ESIZE);
1057 		if (eq->avail < ndesc) {
1058 			wr->b_next = next;
1059 			wrq->no_desc++;
1060 			break;
1061 		}
1062 
1063 		dst = (void *)&eq->desc[eq->pidx];
1064 		for (m = wr; m; m = m->b_cont)
1065 			copy_to_txd(eq, (void *)m->b_rptr, &dst, MBLKL(m));
1066 
1067 		eq->pidx += ndesc;
1068 		eq->avail -= ndesc;
1069 		if (eq->pidx >= eq->cap)
1070 			eq->pidx -= eq->cap;
1071 
1072 		eq->pending += ndesc;
1073 		if (eq->pending > 16)
1074 			ring_tx_db(sc, eq);
1075 
1076 		wrq->tx_wrs++;
1077 		freemsg(wr);
1078 
1079 		if (eq->avail < 8) {
1080 			can_reclaim = reclaimable(eq);
1081 			eq->cidx += can_reclaim;
1082 			eq->avail += can_reclaim;
1083 			if (eq->cidx >= eq->cap)
1084 				eq->cidx -= eq->cap;
1085 		}
1086 	}
1087 
1088 	if (eq->pending != 0)
1089 		ring_tx_db(sc, eq);
1090 
1091 	if (wr == NULL)
1092 		wr_list->head = wr_list->tail = NULL;
1093 	else {
1094 		wr_list->head = wr;
1095 
1096 		ASSERT(wr_list->tail->b_next == NULL);
1097 	}
1098 
1099 	return (0);
1100 }
1101 #endif
1102 
1103 /* Per-packet header in a coalesced tx WR, before the SGL starts (in flits) */
1104 #define	TXPKTS_PKT_HDR ((\
1105 	sizeof (struct ulp_txpkt) + \
1106 	sizeof (struct ulptx_idata) + \
1107 	sizeof (struct cpl_tx_pkt_core)) / 8)
1108 
1109 /* Header of a coalesced tx WR, before SGL of first packet (in flits) */
1110 #define	TXPKTS_WR_HDR (\
1111 	sizeof (struct fw_eth_tx_pkts_wr) / 8 + \
1112 	TXPKTS_PKT_HDR)
1113 
1114 /* Header of a tx WR, before SGL of first packet (in flits) */
1115 #define	TXPKT_WR_HDR ((\
1116 	sizeof (struct fw_eth_tx_pkt_wr) + \
1117 	sizeof (struct cpl_tx_pkt_core)) / 8)
1118 
1119 /* Header of a tx LSO WR, before SGL of first packet (in flits) */
1120 #define	TXPKT_LSO_WR_HDR ((\
1121 	sizeof (struct fw_eth_tx_pkt_wr) + \
1122 	sizeof(struct cpl_tx_pkt_lso_core) + \
1123 	sizeof (struct cpl_tx_pkt_core)) / 8)
1124 
1125 mblk_t *
1126 t4_eth_tx(void *arg, mblk_t *frame)
1127 {
1128 	struct sge_txq *txq = (struct sge_txq *) arg;
1129 	struct port_info *pi = txq->port;
1130 	struct adapter *sc = pi->adapter;
1131 	struct sge_eq *eq = &txq->eq;
1132 	mblk_t *next_frame;
1133 	int rc, coalescing;
1134 	struct txpkts txpkts;
1135 	struct txinfo txinfo;
1136 
1137 	txpkts.npkt = 0; /* indicates there's nothing in txpkts */
1138 	coalescing = 0;
1139 
1140 	TXQ_LOCK(txq);
1141 	if (eq->avail < 8)
1142 		(void) reclaim_tx_descs(txq, 8);
1143 	for (; frame; frame = next_frame) {
1144 
1145 		if (eq->avail < 8)
1146 			break;
1147 
1148 		next_frame = frame->b_next;
1149 		frame->b_next = NULL;
1150 
1151 		if (next_frame != NULL)
1152 			coalescing = 1;
1153 
1154 		rc = get_frame_txinfo(txq, &frame, &txinfo, coalescing);
1155 		if (rc != 0) {
1156 			if (rc == ENOMEM) {
1157 
1158 				/* Short of resources, suspend tx */
1159 
1160 				frame->b_next = next_frame;
1161 				break;
1162 			}
1163 
1164 			/*
1165 			 * Unrecoverable error for this frame, throw it
1166 			 * away and move on to the next.
1167 			 */
1168 
1169 			freemsg(frame);
1170 			continue;
1171 		}
1172 
1173 		if (coalescing != 0 &&
1174 		    add_to_txpkts(txq, &txpkts, frame, &txinfo) == 0) {
1175 
1176 			/* Successfully absorbed into txpkts */
1177 
1178 			write_ulp_cpl_sgl(pi, txq, &txpkts, &txinfo);
1179 			goto doorbell;
1180 		}
1181 
1182 		/*
1183 		 * We weren't coalescing to begin with, or current frame could
1184 		 * not be coalesced (add_to_txpkts flushes txpkts if a frame
1185 		 * given to it can't be coalesced).  Either way there should be
1186 		 * nothing in txpkts.
1187 		 */
1188 		ASSERT(txpkts.npkt == 0);
1189 
1190 		/* We're sending out individual frames now */
1191 		coalescing = 0;
1192 
1193 		if (eq->avail < 8)
1194 			(void) reclaim_tx_descs(txq, 8);
1195 		rc = write_txpkt_wr(pi, txq, frame, &txinfo);
1196 		if (rc != 0) {
1197 
1198 			/* Short of hardware descriptors, suspend tx */
1199 
1200 			/*
1201 			 * This is an unlikely but expensive failure.  We've
1202 			 * done all the hard work (DMA bindings etc.) and now we
1203 			 * can't send out the frame.  What's worse, we have to
1204 			 * spend even more time freeing up everything in txinfo.
1205 			 */
1206 			txq->qfull++;
1207 			free_txinfo_resources(txq, &txinfo);
1208 
1209 			frame->b_next = next_frame;
1210 			break;
1211 		}
1212 
1213 doorbell:
1214 		/* Fewer and fewer doorbells as the queue fills up */
1215 		if (eq->pending >= (1 << (fls(eq->qsize - eq->avail) / 2))) {
1216 			txq->txbytes += txinfo.len;
1217 			txq->txpkts++;
1218 			ring_tx_db(sc, eq);
1219 		}
1220 		(void) reclaim_tx_descs(txq, 32);
1221 	}
1222 
1223 	if (txpkts.npkt > 0)
1224 		write_txpkts_wr(txq, &txpkts);
1225 
1226 	/*
1227 	 * frame not NULL means there was an error but we haven't thrown it
1228 	 * away.  This can happen when we're short of tx descriptors (qfull) or
1229 	 * maybe even DMA handles (dma_hdl_failed).  Either way, a credit flush
1230 	 * and reclaim will get things going again.
1231 	 *
1232 	 * If eq->avail is already 0 we know a credit flush was requested in the
1233 	 * WR that reduced it to 0 so we don't need another flush (we don't have
1234 	 * any descriptor for a flush WR anyway, duh).
1235 	 */
1236 	if (frame && eq->avail > 0)
1237 		write_txqflush_wr(txq);
1238 
1239 	if (eq->pending != 0)
1240 		ring_tx_db(sc, eq);
1241 
1242 	(void) reclaim_tx_descs(txq, eq->qsize);
1243 	TXQ_UNLOCK(txq);
1244 
1245 	return (frame);
1246 }
1247 
1248 static inline void
1249 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int8_t pktc_idx,
1250 	int qsize, uint8_t esize)
1251 {
1252 	ASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS);
1253 	ASSERT(pktc_idx < SGE_NCOUNTERS);	/* -ve is ok, means don't use */
1254 
1255 	iq->flags = 0;
1256 	iq->adapter = sc;
1257 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
1258 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
1259 	if (pktc_idx >= 0) {
1260 		iq->intr_params |= F_QINTR_CNT_EN;
1261 		iq->intr_pktc_idx = pktc_idx;
1262 	}
1263 	iq->qsize = roundup(qsize, 16);		/* See FW_IQ_CMD/iqsize */
1264 	iq->esize = max(esize, 16);		/* See FW_IQ_CMD/iqesize */
1265 }
1266 
1267 static inline void
1268 init_fl(struct sge_fl *fl, uint16_t qsize)
1269 {
1270 
1271 	fl->qsize = qsize;
1272 	fl->allocb_fail = 0;
1273 }
1274 
1275 static inline void
1276 init_eq(struct adapter *sc, struct sge_eq *eq, uint16_t eqtype, uint16_t qsize,
1277     uint8_t tx_chan, uint16_t iqid)
1278 {
1279 	struct sge *s = &sc->sge;
1280 	uint32_t r;
1281 
1282 	ASSERT(tx_chan < NCHAN);
1283 	ASSERT(eqtype <= EQ_TYPEMASK);
1284 
1285 	if (is_t5(sc->params.chip)) {
1286 		r = t4_read_reg(sc, A_SGE_EGRESS_QUEUES_PER_PAGE_PF);
1287 		r >>= S_QUEUESPERPAGEPF0 +
1288 		    (S_QUEUESPERPAGEPF1 - S_QUEUESPERPAGEPF0) * sc->pf;
1289 		s->s_qpp = r & M_QUEUESPERPAGEPF0;
1290 	}
1291 
1292 	eq->flags = eqtype & EQ_TYPEMASK;
1293 	eq->tx_chan = tx_chan;
1294 	eq->iqid = iqid;
1295 	eq->qsize = qsize;
1296 }
1297 
1298 /*
1299  * Allocates the ring for an ingress queue and an optional freelist.  If the
1300  * freelist is specified it will be allocated and then associated with the
1301  * ingress queue.
1302  *
1303  * Returns errno on failure.  Resources allocated up to that point may still be
1304  * allocated.  Caller is responsible for cleanup in case this function fails.
1305  *
1306  * If the ingress queue will take interrupts directly (iq->flags & IQ_INTR) then
1307  * the intr_idx specifies the vector, starting from 0.  Otherwise it specifies
1308  * the index of the queue to which its interrupts will be forwarded.
1309  */
1310 static int
1311 alloc_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl,
1312     int intr_idx, int cong)
1313 {
1314 	int rc, i, cntxt_id;
1315 	size_t len;
1316 	struct fw_iq_cmd c;
1317 	struct adapter *sc = iq->adapter;
1318 	uint32_t v = 0;
1319 
1320 	len = iq->qsize * iq->esize;
1321 	rc = alloc_desc_ring(sc, len, DDI_DMA_READ, &iq->dhdl, &iq->ahdl,
1322 	    &iq->ba, (caddr_t *)&iq->desc);
1323 	if (rc != 0)
1324 		return (rc);
1325 
1326 	bzero(&c, sizeof (c));
1327 	c.op_to_vfn = cpu_to_be32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
1328 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
1329 	    V_FW_IQ_CMD_VFN(0));
1330 
1331 	c.alloc_to_len16 = cpu_to_be32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
1332 	    FW_LEN16(c));
1333 
1334 	/* Special handling for firmware event queue */
1335 	if (iq == &sc->sge.fwq)
1336 		v |= F_FW_IQ_CMD_IQASYNCH;
1337 
1338 	if (iq->flags & IQ_INTR)
1339 		ASSERT(intr_idx < sc->intr_count);
1340 	else
1341 		v |= F_FW_IQ_CMD_IQANDST;
1342 	v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
1343 
1344 	c.type_to_iqandstindex = cpu_to_be32(v |
1345 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
1346 	    V_FW_IQ_CMD_VIID(pi->viid) |
1347 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
1348 	c.iqdroprss_to_iqesize = cpu_to_be16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
1349 	    F_FW_IQ_CMD_IQGTSMODE |
1350 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
1351 	    V_FW_IQ_CMD_IQESIZE(ilog2(iq->esize) - 4));
1352 	c.iqsize = cpu_to_be16(iq->qsize);
1353 	c.iqaddr = cpu_to_be64(iq->ba);
1354 	if (cong >= 0)
1355 		c.iqns_to_fl0congen = BE_32(F_FW_IQ_CMD_IQFLINTCONGEN |
1356 					V_FW_IQ_CMD_IQTYPE(cong ?
1357 					FW_IQ_IQTYPE_NIC : FW_IQ_IQTYPE_OFLD));
1358 
1359 	if (fl != NULL) {
1360 		unsigned int chip_ver = CHELSIO_CHIP_VERSION(sc->params.chip);
1361 
1362 		mutex_init(&fl->lock, NULL, MUTEX_DRIVER,
1363 		    DDI_INTR_PRI(sc->intr_pri));
1364 		fl->flags |= FL_MTX;
1365 
1366 		len = fl->qsize * RX_FL_ESIZE;
1367 		rc = alloc_desc_ring(sc, len, DDI_DMA_WRITE, &fl->dhdl,
1368 		    &fl->ahdl, &fl->ba, (caddr_t *)&fl->desc);
1369 		if (rc != 0)
1370 			return (rc);
1371 
1372 		/* Allocate space for one software descriptor per buffer. */
1373 		fl->cap = (fl->qsize - sc->sge.stat_len / RX_FL_ESIZE) * 8;
1374 		fl->sdesc = kmem_zalloc(sizeof (struct fl_sdesc) * fl->cap,
1375 		    KM_SLEEP);
1376 		fl->needed = fl->cap;
1377 		fl->lowat = roundup(sc->sge.fl_starve_threshold, 8);
1378 
1379 		c.iqns_to_fl0congen |=
1380 		    cpu_to_be32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
1381 		    F_FW_IQ_CMD_FL0PACKEN | F_FW_IQ_CMD_FL0PADEN);
1382 		if (cong >= 0) {
1383 			c.iqns_to_fl0congen |=
1384 			    BE_32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
1385 			    F_FW_IQ_CMD_FL0CONGCIF |
1386 			    F_FW_IQ_CMD_FL0CONGEN);
1387 		}
1388 
1389 		/* In T6, for egress queue type FL there is internal overhead
1390 		 * of 16B for header going into FLM module.  Hence the maximum
1391 		 * allowed burst size is 448 bytes.  For T4/T5, the hardware
1392 		 * doesn't coalesce fetch requests if more than 64 bytes of
1393 		 * Free List pointers are provided, so we use a 128-byte Fetch
1394 		 * Burst Minimum there (T6 implements coalescing so we can use
1395 		 * the smaller 64-byte value there).
1396 		 */
1397 
1398 		c.fl0dcaen_to_fl0cidxfthresh =
1399 		    cpu_to_be16(V_FW_IQ_CMD_FL0FBMIN(chip_ver <= CHELSIO_T5
1400 						     ? X_FETCHBURSTMIN_128B
1401 						     : X_FETCHBURSTMIN_64B) |
1402 		    V_FW_IQ_CMD_FL0FBMAX(chip_ver <= CHELSIO_T5
1403 					 ? X_FETCHBURSTMAX_512B
1404 					 : X_FETCHBURSTMAX_256B));
1405 		c.fl0size = cpu_to_be16(fl->qsize);
1406 		c.fl0addr = cpu_to_be64(fl->ba);
1407 	}
1408 
1409 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof (c), &c);
1410 	if (rc != 0) {
1411 		cxgb_printf(sc->dip, CE_WARN,
1412 		    "failed to create ingress queue: %d", rc);
1413 		return (rc);
1414 	}
1415 
1416 	iq->cdesc = iq->desc;
1417 	iq->cidx = 0;
1418 	iq->gen = 1;
1419 	iq->intr_next = iq->intr_params;
1420 	iq->adapter = sc;
1421 	iq->cntxt_id = be16_to_cpu(c.iqid);
1422 	iq->abs_id = be16_to_cpu(c.physiqid);
1423 	iq->flags |= IQ_ALLOCATED;
1424 	mutex_init(&iq->lock, NULL,
1425 		    MUTEX_DRIVER, DDI_INTR_PRI(DDI_INTR_PRI(sc->intr_pri)));
1426 	iq->polling = 0;
1427 
1428 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
1429 	if (cntxt_id >= sc->sge.iqmap_sz) {
1430 		panic("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
1431 		      cntxt_id, sc->sge.iqmap_sz - 1);
1432 	}
1433 	sc->sge.iqmap[cntxt_id] = iq;
1434 
1435 	if (fl != NULL) {
1436 		fl->cntxt_id = be16_to_cpu(c.fl0id);
1437 		fl->pidx = fl->cidx = 0;
1438 		fl->copy_threshold = rx_copy_threshold;
1439 
1440 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
1441 		if (cntxt_id >= sc->sge.eqmap_sz) {
1442 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
1443 			      __func__, cntxt_id, sc->sge.eqmap_sz - 1);
1444 		}
1445 		sc->sge.eqmap[cntxt_id] = (void *)fl;
1446 
1447 		FL_LOCK(fl);
1448 		(void) refill_fl(sc, fl, fl->lowat);
1449 		FL_UNLOCK(fl);
1450 
1451 		iq->flags |= IQ_HAS_FL;
1452 	}
1453 
1454 	if (is_t5(sc->params.chip) && cong >= 0) {
1455 		uint32_t param, val;
1456 
1457 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
1458 			V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
1459 			V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
1460 		if (cong == 0)
1461 			val = 1 << 19;
1462 		else {
1463 			val = 2 << 19;
1464 			for (i = 0; i < 4; i++) {
1465 				if (cong & (1 << i))
1466 					val |= 1 << (i << 2);
1467 			}
1468 		}
1469 
1470 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
1471 		if (rc != 0) {
1472 			/* report error but carry on */
1473 			cxgb_printf(sc->dip, CE_WARN,
1474 			    "failed to set congestion manager context for "
1475 			    "ingress queue %d: %d", iq->cntxt_id, rc);
1476 		}
1477 	}
1478 
1479 	/* Enable IQ interrupts */
1480 	iq->state = IQS_IDLE;
1481 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_SEINTARM(iq->intr_params) |
1482 	    V_INGRESSQID(iq->cntxt_id));
1483 
1484 	return (0);
1485 }
1486 
1487 static int
1488 free_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl)
1489 {
1490 	int rc;
1491 
1492 	if (iq != NULL) {
1493 		struct adapter *sc = iq->adapter;
1494 		dev_info_t *dip;
1495 
1496 		dip = pi ? pi->dip : sc->dip;
1497 		if (iq->flags & IQ_ALLOCATED) {
1498 			rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
1499 			    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
1500 			    fl ? fl->cntxt_id : 0xffff, 0xffff);
1501 			if (rc != 0) {
1502 				cxgb_printf(dip, CE_WARN,
1503 				    "failed to free queue %p: %d", iq, rc);
1504 				return (rc);
1505 			}
1506 			mutex_destroy(&iq->lock);
1507 			iq->flags &= ~IQ_ALLOCATED;
1508 		}
1509 
1510 		if (iq->desc != NULL) {
1511 			(void) free_desc_ring(&iq->dhdl, &iq->ahdl);
1512 			iq->desc = NULL;
1513 		}
1514 
1515 		bzero(iq, sizeof (*iq));
1516 	}
1517 
1518 	if (fl != NULL) {
1519 		if (fl->sdesc != NULL) {
1520 			FL_LOCK(fl);
1521 			free_fl_bufs(fl);
1522 			FL_UNLOCK(fl);
1523 
1524 			kmem_free(fl->sdesc, sizeof (struct fl_sdesc) *
1525 			    fl->cap);
1526 			fl->sdesc = NULL;
1527 		}
1528 
1529 		if (fl->desc != NULL) {
1530 			(void) free_desc_ring(&fl->dhdl, &fl->ahdl);
1531 			fl->desc = NULL;
1532 		}
1533 
1534 		if (fl->flags & FL_MTX) {
1535 			mutex_destroy(&fl->lock);
1536 			fl->flags &= ~FL_MTX;
1537 		}
1538 
1539 		bzero(fl, sizeof (struct sge_fl));
1540 	}
1541 
1542 	return (0);
1543 }
1544 
1545 static int
1546 alloc_fwq(struct adapter *sc)
1547 {
1548 	int rc, intr_idx;
1549 	struct sge_iq *fwq = &sc->sge.fwq;
1550 
1551 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE, FW_IQ_ESIZE);
1552 	fwq->flags |= IQ_INTR;	/* always */
1553 	intr_idx = sc->intr_count > 1 ? 1 : 0;
1554 	rc = alloc_iq_fl(sc->port[0], fwq, NULL, intr_idx, -1);
1555 	if (rc != 0) {
1556 		cxgb_printf(sc->dip, CE_WARN,
1557 		    "failed to create firmware event queue: %d.", rc);
1558 		return (rc);
1559 	}
1560 
1561 	return (0);
1562 }
1563 
1564 static int
1565 free_fwq(struct adapter *sc)
1566 {
1567 
1568 	return (free_iq_fl(NULL, &sc->sge.fwq, NULL));
1569 }
1570 
1571 #ifdef TCP_OFFLOAD_ENABLE
1572 static int
1573 alloc_mgmtq(struct adapter *sc)
1574 {
1575 	int rc;
1576 	struct sge_wrq *mgmtq = &sc->sge.mgmtq;
1577 
1578 	init_eq(sc, &mgmtq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[0]->tx_chan,
1579 	    sc->sge.fwq.cntxt_id);
1580 	rc = alloc_wrq(sc, NULL, mgmtq, 0);
1581 	if (rc != 0) {
1582 		cxgb_printf(sc->dip, CE_WARN,
1583 		    "failed to create management queue: %d\n", rc);
1584 		return (rc);
1585 	}
1586 
1587 	return (0);
1588 }
1589 #endif
1590 
1591 static int
1592 alloc_rxq(struct port_info *pi, struct sge_rxq *rxq, int intr_idx, int i)
1593 {
1594 	int rc;
1595 
1596 	rxq->port = pi;
1597 	rc = alloc_iq_fl(pi, &rxq->iq, &rxq->fl, intr_idx,
1598 			 t4_get_tp_ch_map(pi->adapter, pi->tx_chan));
1599 	if (rc != 0)
1600 		return (rc);
1601 
1602 	rxq->ksp = setup_rxq_kstats(pi, rxq, i);
1603 
1604 	return (rc);
1605 }
1606 
1607 static int
1608 free_rxq(struct port_info *pi, struct sge_rxq *rxq)
1609 {
1610 	int rc;
1611 
1612 	if (rxq->ksp != NULL) {
1613 		kstat_delete(rxq->ksp);
1614 		rxq->ksp = NULL;
1615 	}
1616 
1617 	rc = free_iq_fl(pi, &rxq->iq, &rxq->fl);
1618 	if (rc == 0)
1619 		bzero(&rxq->fl, sizeof (*rxq) - offsetof(struct sge_rxq, fl));
1620 
1621 	return (rc);
1622 }
1623 
1624 #ifdef TCP_OFFLOAD_ENABLE
1625 static int
1626 alloc_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq,
1627 	int intr_idx)
1628 {
1629 	int rc;
1630 
1631 	rc = alloc_iq_fl(pi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx,
1632 	    t4_get_tp_ch_map(pi->adapter, pi->tx_chan));
1633 	if (rc != 0)
1634 		return (rc);
1635 
1636 	return (rc);
1637 }
1638 
1639 static int
1640 free_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq)
1641 {
1642 	int rc;
1643 
1644 	rc = free_iq_fl(pi, &ofld_rxq->iq, &ofld_rxq->fl);
1645 	if (rc == 0)
1646 		bzero(&ofld_rxq->fl, sizeof (*ofld_rxq) -
1647 		    offsetof(struct sge_ofld_rxq, fl));
1648 
1649 	return (rc);
1650 }
1651 #endif
1652 
1653 static int
1654 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
1655 {
1656 	int rc, cntxt_id;
1657 	struct fw_eq_ctrl_cmd c;
1658 
1659 	bzero(&c, sizeof (c));
1660 
1661 	c.op_to_vfn = BE_32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
1662 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
1663 	    V_FW_EQ_CTRL_CMD_VFN(0));
1664 	c.alloc_to_len16 = BE_32(F_FW_EQ_CTRL_CMD_ALLOC |
1665 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
1666 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid)); /* TODO */
1667 	c.physeqid_pkd = BE_32(0);
1668 	c.fetchszm_to_iqid =
1669 	    BE_32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
1670 	    V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
1671 	    F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
1672 	c.dcaen_to_eqsize =
1673 	    BE_32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
1674 	    V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
1675 	    V_FW_EQ_CTRL_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
1676 	    V_FW_EQ_CTRL_CMD_EQSIZE(eq->qsize));
1677 	c.eqaddr = BE_64(eq->ba);
1678 
1679 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof (c), &c);
1680 	if (rc != 0) {
1681 		cxgb_printf(sc->dip, CE_WARN,
1682 		    "failed to create control queue %d: %d", eq->tx_chan, rc);
1683 		return (rc);
1684 	}
1685 	eq->flags |= EQ_ALLOCATED;
1686 
1687 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(BE_32(c.cmpliqid_eqid));
1688 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
1689 	if (cntxt_id >= sc->sge.eqmap_sz)
1690 		panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
1691 		      cntxt_id, sc->sge.eqmap_sz - 1);
1692 	sc->sge.eqmap[cntxt_id] = eq;
1693 
1694 	return (rc);
1695 }
1696 
1697 static int
1698 eth_eq_alloc(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
1699 {
1700 	int rc, cntxt_id;
1701 	struct fw_eq_eth_cmd c;
1702 
1703 	bzero(&c, sizeof (c));
1704 
1705 	c.op_to_vfn = BE_32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
1706 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
1707 	    V_FW_EQ_ETH_CMD_VFN(0));
1708 	c.alloc_to_len16 = BE_32(F_FW_EQ_ETH_CMD_ALLOC |
1709 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
1710 	c.autoequiqe_to_viid = BE_32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
1711 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(pi->viid));
1712 	c.fetchszm_to_iqid =
1713 	    BE_32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
1714 	    V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
1715 	    V_FW_EQ_ETH_CMD_IQID(eq->iqid));
1716 	c.dcaen_to_eqsize = BE_32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
1717 	    V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
1718 	    V_FW_EQ_ETH_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
1719 	    V_FW_EQ_ETH_CMD_EQSIZE(eq->qsize));
1720 	c.eqaddr = BE_64(eq->ba);
1721 
1722 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof (c), &c);
1723 	if (rc != 0) {
1724 		cxgb_printf(pi->dip, CE_WARN,
1725 		    "failed to create Ethernet egress queue: %d", rc);
1726 		return (rc);
1727 	}
1728 	eq->flags |= EQ_ALLOCATED;
1729 
1730 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(BE_32(c.eqid_pkd));
1731 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
1732 	if (cntxt_id >= sc->sge.eqmap_sz)
1733 		panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
1734 		      cntxt_id, sc->sge.eqmap_sz - 1);
1735 	sc->sge.eqmap[cntxt_id] = eq;
1736 
1737 	return (rc);
1738 }
1739 
1740 #ifdef TCP_OFFLOAD_ENABLE
1741 static int
1742 ofld_eq_alloc(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
1743 {
1744 	int rc, cntxt_id;
1745 	struct fw_eq_ofld_cmd c;
1746 
1747 	bzero(&c, sizeof (c));
1748 
1749 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
1750 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
1751 	    V_FW_EQ_OFLD_CMD_VFN(0));
1752 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
1753 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
1754 	c.fetchszm_to_iqid =
1755 	    htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
1756 	    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
1757 	    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
1758 	c.dcaen_to_eqsize =
1759 	    BE_32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
1760 	    V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
1761 	    V_FW_EQ_OFLD_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
1762 	    V_FW_EQ_OFLD_CMD_EQSIZE(eq->qsize));
1763 	c.eqaddr = BE_64(eq->ba);
1764 
1765 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof (c), &c);
1766 	if (rc != 0) {
1767 		cxgb_printf(pi->dip, CE_WARN,
1768 		    "failed to create egress queue for TCP offload: %d", rc);
1769 		return (rc);
1770 	}
1771 	eq->flags |= EQ_ALLOCATED;
1772 
1773 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(BE_32(c.eqid_pkd));
1774 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
1775 	if (cntxt_id >= sc->sge.eqmap_sz)
1776 		panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
1777 		      cntxt_id, sc->sge.eqmap_sz - 1);
1778 	sc->sge.eqmap[cntxt_id] = eq;
1779 
1780 	return (rc);
1781 }
1782 #endif
1783 
1784 static int
1785 alloc_eq(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
1786 {
1787 	int rc;
1788 	size_t len;
1789 
1790 	mutex_init(&eq->lock, NULL, MUTEX_DRIVER, DDI_INTR_PRI(sc->intr_pri));
1791 	eq->flags |= EQ_MTX;
1792 
1793 	len = eq->qsize * EQ_ESIZE;
1794 	rc = alloc_desc_ring(sc, len, DDI_DMA_WRITE, &eq->desc_dhdl,
1795 	    &eq->desc_ahdl, &eq->ba, (caddr_t *)&eq->desc);
1796 	if (rc != 0)
1797 		return (rc);
1798 
1799 	eq->cap = eq->qsize - sc->sge.stat_len / EQ_ESIZE;
1800 	eq->spg = (void *)&eq->desc[eq->cap];
1801 	eq->avail = eq->cap - 1;	/* one less to avoid cidx = pidx */
1802 	eq->pidx = eq->cidx = 0;
1803 	eq->doorbells = sc->doorbells;
1804 
1805 	switch (eq->flags & EQ_TYPEMASK) {
1806 	case EQ_CTRL:
1807 		rc = ctrl_eq_alloc(sc, eq);
1808 		break;
1809 
1810 	case EQ_ETH:
1811 		rc = eth_eq_alloc(sc, pi, eq);
1812 		break;
1813 
1814 #ifdef TCP_OFFLOAD_ENABLE
1815 	case EQ_OFLD:
1816 		rc = ofld_eq_alloc(sc, pi, eq);
1817 		break;
1818 #endif
1819 
1820 	default:
1821 		panic("%s: invalid eq type %d.", __func__,
1822 		    eq->flags & EQ_TYPEMASK);
1823 	}
1824 
1825 	if (eq->doorbells &
1826 		(DOORBELL_UDB | DOORBELL_UDBWC | DOORBELL_WCWR)) {
1827 		uint32_t s_qpp = sc->sge.s_qpp;
1828 		uint32_t mask = (1 << s_qpp) - 1;
1829 		volatile uint8_t *udb;
1830 
1831 		udb = (volatile uint8_t *)sc->reg1p + UDBS_DB_OFFSET;
1832 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;   /* pg offset */
1833 		eq->udb_qid = eq->cntxt_id & mask;              /* id in page */
1834 		if (eq->udb_qid > PAGE_SIZE / UDBS_SEG_SIZE)
1835 			eq->doorbells &= ~DOORBELL_WCWR;
1836 		else {
1837 			udb += eq->udb_qid << UDBS_SEG_SHIFT;   /* seg offset */
1838 			eq->udb_qid = 0;
1839 		}
1840 		eq->udb = (volatile void *)udb;
1841 	}
1842 
1843 	if (rc != 0) {
1844 		cxgb_printf(sc->dip, CE_WARN,
1845 		    "failed to allocate egress queue(%d): %d",
1846 		    eq->flags & EQ_TYPEMASK, rc);
1847 	}
1848 
1849 	return (rc);
1850 }
1851 
1852 static int
1853 free_eq(struct adapter *sc, struct sge_eq *eq)
1854 {
1855 	int rc;
1856 
1857 	if (eq->flags & EQ_ALLOCATED) {
1858 		switch (eq->flags & EQ_TYPEMASK) {
1859 		case EQ_CTRL:
1860 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
1861 			    eq->cntxt_id);
1862 			break;
1863 
1864 		case EQ_ETH:
1865 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
1866 			    eq->cntxt_id);
1867 			break;
1868 
1869 #ifdef TCP_OFFLOAD_ENABLE
1870 		case EQ_OFLD:
1871 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
1872 			    eq->cntxt_id);
1873 			break;
1874 #endif
1875 
1876 		default:
1877 			panic("%s: invalid eq type %d.", __func__,
1878 			    eq->flags & EQ_TYPEMASK);
1879 		}
1880 		if (rc != 0) {
1881 			cxgb_printf(sc->dip, CE_WARN,
1882 			    "failed to free egress queue (%d): %d",
1883 			    eq->flags & EQ_TYPEMASK, rc);
1884 			return (rc);
1885 		}
1886 		eq->flags &= ~EQ_ALLOCATED;
1887 	}
1888 
1889 	if (eq->desc != NULL) {
1890 		(void) free_desc_ring(&eq->desc_dhdl, &eq->desc_ahdl);
1891 		eq->desc = NULL;
1892 	}
1893 
1894 	if (eq->flags & EQ_MTX)
1895 		mutex_destroy(&eq->lock);
1896 
1897 	bzero(eq, sizeof (*eq));
1898 	return (0);
1899 }
1900 
1901 #ifdef TCP_OFFLOAD_ENABLE
1902 /* ARGSUSED */
1903 static int
1904 alloc_wrq(struct adapter *sc, struct port_info *pi, struct sge_wrq *wrq,
1905     int idx)
1906 {
1907 	int rc;
1908 
1909 	rc = alloc_eq(sc, pi, &wrq->eq);
1910 	if (rc != 0)
1911 		return (rc);
1912 
1913 	wrq->adapter = sc;
1914 	wrq->wr_list.head = NULL;
1915 	wrq->wr_list.tail = NULL;
1916 
1917 	/*
1918 	 * TODO: use idx to figure out what kind of wrq this is and install
1919 	 * useful kstats for it.
1920 	 */
1921 
1922 	return (rc);
1923 }
1924 
1925 static int
1926 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
1927 {
1928 	int rc;
1929 
1930 	rc = free_eq(sc, &wrq->eq);
1931 	if (rc != 0)
1932 		return (rc);
1933 
1934 	bzero(wrq, sizeof (*wrq));
1935 	return (0);
1936 }
1937 #endif
1938 
1939 static int
1940 alloc_txq(struct port_info *pi, struct sge_txq *txq, int idx)
1941 {
1942 	int rc, i;
1943 	struct adapter *sc = pi->adapter;
1944 	struct sge_eq *eq = &txq->eq;
1945 
1946 	rc = alloc_eq(sc, pi, eq);
1947 	if (rc != 0)
1948 		return (rc);
1949 
1950 	txq->port = pi;
1951 	txq->sdesc = kmem_zalloc(sizeof (struct tx_sdesc) * eq->cap, KM_SLEEP);
1952 	txq->txb_size = eq->qsize * tx_copy_threshold;
1953 	rc = alloc_tx_copybuffer(sc, txq->txb_size, &txq->txb_dhdl,
1954 	    &txq->txb_ahdl, &txq->txb_ba, &txq->txb_va);
1955 	if (rc == 0)
1956 		txq->txb_avail = txq->txb_size;
1957 	else
1958 		txq->txb_avail = txq->txb_size = 0;
1959 
1960 	/*
1961 	 * TODO: is this too low?  Worst case would need around 4 times qsize
1962 	 * (all tx descriptors filled to the brim with SGLs, with each entry in
1963 	 * the SGL coming from a distinct DMA handle).  Increase tx_dhdl_total
1964 	 * if you see too many dma_hdl_failed.
1965 	 */
1966 	txq->tx_dhdl_total = eq->qsize * 2;
1967 	txq->tx_dhdl = kmem_zalloc(sizeof (ddi_dma_handle_t) *
1968 	    txq->tx_dhdl_total, KM_SLEEP);
1969 	for (i = 0; i < txq->tx_dhdl_total; i++) {
1970 		rc = ddi_dma_alloc_handle(sc->dip, &sc->sge.dma_attr_tx,
1971 		    DDI_DMA_SLEEP, 0, &txq->tx_dhdl[i]);
1972 		if (rc != DDI_SUCCESS) {
1973 			cxgb_printf(sc->dip, CE_WARN,
1974 			    "%s: failed to allocate DMA handle (%d)",
1975 			    __func__, rc);
1976 			return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EINVAL);
1977 		}
1978 		txq->tx_dhdl_avail++;
1979 	}
1980 
1981 	txq->ksp = setup_txq_kstats(pi, txq, idx);
1982 
1983 	return (rc);
1984 }
1985 
1986 static int
1987 free_txq(struct port_info *pi, struct sge_txq *txq)
1988 {
1989 	int i;
1990 	struct adapter *sc = pi->adapter;
1991 	struct sge_eq *eq = &txq->eq;
1992 
1993 	if (txq->ksp != NULL) {
1994 		kstat_delete(txq->ksp);
1995 		txq->ksp = NULL;
1996 	}
1997 
1998 	if (txq->txb_va != NULL) {
1999 		(void) free_desc_ring(&txq->txb_dhdl, &txq->txb_ahdl);
2000 		txq->txb_va = NULL;
2001 	}
2002 
2003 	if (txq->sdesc != NULL) {
2004 		struct tx_sdesc *sd;
2005 		ddi_dma_handle_t hdl;
2006 
2007 		TXQ_LOCK(txq);
2008 		while (eq->cidx != eq->pidx) {
2009 			sd = &txq->sdesc[eq->cidx];
2010 
2011 			for (i = sd->hdls_used; i; i--) {
2012 				hdl = txq->tx_dhdl[txq->tx_dhdl_cidx];
2013 				(void) ddi_dma_unbind_handle(hdl);
2014 				if (++txq->tx_dhdl_cidx == txq->tx_dhdl_total)
2015 					txq->tx_dhdl_cidx = 0;
2016 			}
2017 
2018 			ASSERT(sd->m);
2019 			freemsgchain(sd->m);
2020 
2021 			eq->cidx += sd->desc_used;
2022 			if (eq->cidx >= eq->cap)
2023 				eq->cidx -= eq->cap;
2024 
2025 			txq->txb_avail += txq->txb_used;
2026 		}
2027 		ASSERT(txq->tx_dhdl_cidx == txq->tx_dhdl_pidx);
2028 		ASSERT(txq->txb_avail == txq->txb_size);
2029 		TXQ_UNLOCK(txq);
2030 
2031 		kmem_free(txq->sdesc, sizeof (struct tx_sdesc) * eq->cap);
2032 		txq->sdesc = NULL;
2033 	}
2034 
2035 	if (txq->tx_dhdl != NULL) {
2036 		for (i = 0; i < txq->tx_dhdl_total; i++) {
2037 			if (txq->tx_dhdl[i] != NULL)
2038 				ddi_dma_free_handle(&txq->tx_dhdl[i]);
2039 		}
2040 	}
2041 
2042 	(void) free_eq(sc, &txq->eq);
2043 
2044 	bzero(txq, sizeof (*txq));
2045 	return (0);
2046 }
2047 
2048 /*
2049  * Allocates a block of contiguous memory for DMA.  Can be used to allocate
2050  * memory for descriptor rings or for tx/rx copy buffers.
2051  *
2052  * Caller does not have to clean up anything if this function fails, it cleans
2053  * up after itself.
2054  *
2055  * Caller provides the following:
2056  * len		length of the block of memory to allocate.
2057  * flags	DDI_DMA_* flags to use (CONSISTENT/STREAMING, READ/WRITE/RDWR)
2058  * acc_attr	device access attributes for the allocation.
2059  * dma_attr	DMA attributes for the allocation
2060  *
2061  * If the function is successful it fills up this information:
2062  * dma_hdl	DMA handle for the allocated memory
2063  * acc_hdl	access handle for the allocated memory
2064  * ba		bus address of the allocated memory
2065  * va		KVA of the allocated memory.
2066  */
2067 static int
2068 alloc_dma_memory(struct adapter *sc, size_t len, int flags,
2069     ddi_device_acc_attr_t *acc_attr, ddi_dma_attr_t *dma_attr,
2070     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl,
2071     uint64_t *pba, caddr_t *pva)
2072 {
2073 	int rc;
2074 	ddi_dma_handle_t dhdl;
2075 	ddi_acc_handle_t ahdl;
2076 	ddi_dma_cookie_t cookie;
2077 	uint_t ccount;
2078 	caddr_t va;
2079 	size_t real_len;
2080 
2081 	*pva = NULL;
2082 
2083 	/*
2084 	 * DMA handle.
2085 	 */
2086 	rc = ddi_dma_alloc_handle(sc->dip, dma_attr, DDI_DMA_SLEEP, 0, &dhdl);
2087 	if (rc != DDI_SUCCESS) {
2088 		cxgb_printf(sc->dip, CE_WARN,
2089 		    "failed to allocate DMA handle: %d", rc);
2090 
2091 		return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EINVAL);
2092 	}
2093 
2094 	/*
2095 	 * Memory suitable for DMA.
2096 	 */
2097 	rc = ddi_dma_mem_alloc(dhdl, len, acc_attr,
2098 	    flags & DDI_DMA_CONSISTENT ? DDI_DMA_CONSISTENT : DDI_DMA_STREAMING,
2099 	    DDI_DMA_SLEEP, 0, &va, &real_len, &ahdl);
2100 	if (rc != DDI_SUCCESS) {
2101 		cxgb_printf(sc->dip, CE_WARN,
2102 		    "failed to allocate DMA memory: %d", rc);
2103 
2104 		ddi_dma_free_handle(&dhdl);
2105 		return (ENOMEM);
2106 	}
2107 
2108 	if (len != real_len) {
2109 		cxgb_printf(sc->dip, CE_WARN,
2110 		    "%s: len (%u) != real_len (%u)\n", len, real_len);
2111 	}
2112 
2113 	/*
2114 	 * DMA bindings.
2115 	 */
2116 	rc = ddi_dma_addr_bind_handle(dhdl, NULL, va, real_len, flags, NULL,
2117 	    NULL, &cookie, &ccount);
2118 	if (rc != DDI_DMA_MAPPED) {
2119 		cxgb_printf(sc->dip, CE_WARN,
2120 		    "failed to map DMA memory: %d", rc);
2121 
2122 		ddi_dma_mem_free(&ahdl);
2123 		ddi_dma_free_handle(&dhdl);
2124 		return (ENOMEM);
2125 	}
2126 	if (ccount != 1) {
2127 		cxgb_printf(sc->dip, CE_WARN,
2128 		    "unusable DMA mapping (%d segments)", ccount);
2129 		(void) free_desc_ring(&dhdl, &ahdl);
2130 	}
2131 
2132 	bzero(va, real_len);
2133 	*dma_hdl = dhdl;
2134 	*acc_hdl = ahdl;
2135 	*pba = cookie.dmac_laddress;
2136 	*pva = va;
2137 
2138 	return (0);
2139 }
2140 
2141 static int
2142 free_dma_memory(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl)
2143 {
2144 	(void) ddi_dma_unbind_handle(*dhdl);
2145 	ddi_dma_mem_free(ahdl);
2146 	ddi_dma_free_handle(dhdl);
2147 
2148 	return (0);
2149 }
2150 
2151 static int
2152 alloc_desc_ring(struct adapter *sc, size_t len, int rw,
2153     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl,
2154     uint64_t *pba, caddr_t *pva)
2155 {
2156 	ddi_device_acc_attr_t *acc_attr = &sc->sge.acc_attr_desc;
2157 	ddi_dma_attr_t *dma_attr = &sc->sge.dma_attr_desc;
2158 
2159 	return (alloc_dma_memory(sc, len, DDI_DMA_CONSISTENT | rw, acc_attr,
2160 	    dma_attr, dma_hdl, acc_hdl, pba, pva));
2161 }
2162 
2163 static int
2164 free_desc_ring(ddi_dma_handle_t *dhdl, ddi_acc_handle_t *ahdl)
2165 {
2166 	return (free_dma_memory(dhdl, ahdl));
2167 }
2168 
2169 static int
2170 alloc_tx_copybuffer(struct adapter *sc, size_t len,
2171     ddi_dma_handle_t *dma_hdl, ddi_acc_handle_t *acc_hdl,
2172     uint64_t *pba, caddr_t *pva)
2173 {
2174 	ddi_device_acc_attr_t *acc_attr = &sc->sge.acc_attr_tx;
2175 	ddi_dma_attr_t *dma_attr = &sc->sge.dma_attr_desc; /* NOT dma_attr_tx */
2176 
2177 	return (alloc_dma_memory(sc, len, DDI_DMA_STREAMING | DDI_DMA_WRITE,
2178 	    acc_attr, dma_attr, dma_hdl, acc_hdl, pba, pva));
2179 }
2180 
2181 static inline bool
2182 is_new_response(const struct sge_iq *iq, struct rsp_ctrl **ctrl)
2183 {
2184 	(void) ddi_dma_sync(iq->dhdl, (uintptr_t)iq->cdesc -
2185 	    (uintptr_t)iq->desc, iq->esize, DDI_DMA_SYNC_FORKERNEL);
2186 
2187 	*ctrl = (void *)((uintptr_t)iq->cdesc +
2188 	    (iq->esize - sizeof (struct rsp_ctrl)));
2189 
2190 	return ((((*ctrl)->u.type_gen >> S_RSPD_GEN) == iq->gen));
2191 }
2192 
2193 static inline void
2194 iq_next(struct sge_iq *iq)
2195 {
2196 	iq->cdesc = (void *) ((uintptr_t)iq->cdesc + iq->esize);
2197 	if (++iq->cidx == iq->qsize - 1) {
2198 		iq->cidx = 0;
2199 		iq->gen ^= 1;
2200 		iq->cdesc = iq->desc;
2201 	}
2202 }
2203 
2204 /*
2205  * Fill up the freelist by upto nbufs and maybe ring its doorbell.
2206  *
2207  * Returns non-zero to indicate that it should be added to the list of starving
2208  * freelists.
2209  */
2210 static int
2211 refill_fl(struct adapter *sc, struct sge_fl *fl, int nbufs)
2212 {
2213 	uint64_t *d = &fl->desc[fl->pidx];
2214 	struct fl_sdesc *sd = &fl->sdesc[fl->pidx];
2215 
2216 	FL_LOCK_ASSERT_OWNED(fl);
2217 	ASSERT(nbufs >= 0);
2218 
2219 	if (nbufs > fl->needed)
2220 		nbufs = fl->needed;
2221 
2222 	while (nbufs--) {
2223 		if (sd->rxb != NULL) {
2224 			if (sd->rxb->ref_cnt == 1) {
2225 				/*
2226 				 * Buffer is available for recycling.  Two ways
2227 				 * this can happen:
2228 				 *
2229 				 * a) All the packets DMA'd into it last time
2230 				 *    around were within the rx_copy_threshold
2231 				 *    and no part of the buffer was ever passed
2232 				 *    up (ref_cnt never went over 1).
2233 				 *
2234 				 * b) Packets DMA'd into the buffer were passed
2235 				 *    up but have all been freed by the upper
2236 				 *    layers by now (ref_cnt went over 1 but is
2237 				 *    now back to 1).
2238 				 *
2239 				 * Either way the bus address in the descriptor
2240 				 * ring is already valid.
2241 				 */
2242 				ASSERT(*d == cpu_to_be64(sd->rxb->ba));
2243 				d++;
2244 				goto recycled;
2245 			} else {
2246 				/*
2247 				 * Buffer still in use and we need a
2248 				 * replacement. But first release our reference
2249 				 * on the existing buffer.
2250 				 */
2251 				rxbuf_free(sd->rxb);
2252 			}
2253 		}
2254 
2255 		sd->rxb = rxbuf_alloc(sc->sge.rxbuf_cache, KM_NOSLEEP, 1);
2256 		if (sd->rxb == NULL)
2257 			break;
2258 		*d++ = cpu_to_be64(sd->rxb->ba);
2259 
2260 recycled:	fl->pending++;
2261 		sd++;
2262 		fl->needed--;
2263 		if (++fl->pidx == fl->cap) {
2264 			fl->pidx = 0;
2265 			sd = fl->sdesc;
2266 			d = fl->desc;
2267 		}
2268 	}
2269 
2270 	if (fl->pending >= 8)
2271 		ring_fl_db(sc, fl);
2272 
2273 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
2274 }
2275 
2276 #ifndef TAILQ_FOREACH_SAFE
2277 #define	TAILQ_FOREACH_SAFE(var, head, field, tvar)			\
2278 	for ((var) = TAILQ_FIRST((head));				\
2279 	    (var) && ((tvar) = TAILQ_NEXT((var), field), 1);		\
2280 	    (var) = (tvar))
2281 #endif
2282 
2283 /*
2284  * Attempt to refill all starving freelists.
2285  */
2286 static void
2287 refill_sfl(void *arg)
2288 {
2289 	struct adapter *sc = arg;
2290 	struct sge_fl *fl, *fl_temp;
2291 
2292 	mutex_enter(&sc->sfl_lock);
2293 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
2294 		FL_LOCK(fl);
2295 		(void) refill_fl(sc, fl, 64);
2296 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
2297 			TAILQ_REMOVE(&sc->sfl, fl, link);
2298 			fl->flags &= ~FL_STARVING;
2299 		}
2300 		FL_UNLOCK(fl);
2301 	}
2302 
2303 	if (!TAILQ_EMPTY(&sc->sfl) != 0)
2304 		sc->sfl_timer =  timeout(refill_sfl, sc, drv_usectohz(100000));
2305 	mutex_exit(&sc->sfl_lock);
2306 }
2307 
2308 static void
2309 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
2310 {
2311 	mutex_enter(&sc->sfl_lock);
2312 	FL_LOCK(fl);
2313 	if ((fl->flags & FL_DOOMED) == 0) {
2314 		if (TAILQ_EMPTY(&sc->sfl) != 0) {
2315 			sc->sfl_timer = timeout(refill_sfl, sc,
2316 			    drv_usectohz(100000));
2317 		}
2318 		fl->flags |= FL_STARVING;
2319 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
2320 	}
2321 	FL_UNLOCK(fl);
2322 	mutex_exit(&sc->sfl_lock);
2323 }
2324 
2325 static void
2326 free_fl_bufs(struct sge_fl *fl)
2327 {
2328 	struct fl_sdesc *sd;
2329 	unsigned int i;
2330 
2331 	FL_LOCK_ASSERT_OWNED(fl);
2332 
2333 	for (i = 0; i < fl->cap; i++) {
2334 		sd = &fl->sdesc[i];
2335 
2336 		if (sd->rxb != NULL) {
2337 			rxbuf_free(sd->rxb);
2338 			sd->rxb = NULL;
2339 		}
2340 	}
2341 }
2342 
2343 /*
2344  * Note that fl->cidx and fl->offset are left unchanged in case of failure.
2345  */
2346 static mblk_t *
2347 get_fl_payload(struct adapter *sc, struct sge_fl *fl,
2348 	       uint32_t len_newbuf, int *fl_bufs_used)
2349 {
2350 	struct mblk_pair frame = {0};
2351 	struct rxbuf *rxb;
2352 	mblk_t *m = NULL;
2353 	uint_t nbuf = 0, len, copy, n;
2354 	uint32_t cidx, offset, rcidx, roffset;
2355 
2356 	/*
2357 	 * The SGE won't pack a new frame into the current buffer if the entire
2358 	 * payload doesn't fit in the remaining space.  Move on to the next buf
2359 	 * in that case.
2360 	 */
2361 	rcidx = fl->cidx;
2362 	roffset = fl->offset;
2363 	if (fl->offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
2364 		fl->offset = 0;
2365 		if (++fl->cidx == fl->cap)
2366 			fl->cidx = 0;
2367 		nbuf++;
2368 	}
2369 	cidx = fl->cidx;
2370 	offset = fl->offset;
2371 
2372 	len = G_RSPD_LEN(len_newbuf);	/* pktshift + payload length */
2373 	copy = (len <= fl->copy_threshold);
2374 	if (copy != 0) {
2375 		frame.head = m = allocb(len, BPRI_HI);
2376 		if (m == NULL) {
2377 			fl->allocb_fail++;
2378 			cmn_err(CE_WARN,"%s: mbuf allocation failure "
2379 					"count = %llu", __func__,
2380 					(unsigned long long)fl->allocb_fail);
2381 			fl->cidx = rcidx;
2382 			fl->offset = roffset;
2383 			return (NULL);
2384 		}
2385 	}
2386 
2387 	while (len) {
2388 		rxb = fl->sdesc[cidx].rxb;
2389 		n = min(len, rxb->buf_size - offset);
2390 
2391 		(void) ddi_dma_sync(rxb->dhdl, offset, n,
2392 		    DDI_DMA_SYNC_FORKERNEL);
2393 
2394 		if (copy != 0)
2395 			bcopy(rxb->va + offset, m->b_wptr, n);
2396 		else {
2397 			m = desballoc((unsigned char *)rxb->va + offset, n,
2398 			    BPRI_HI, &rxb->freefunc);
2399 			if (m == NULL) {
2400 				fl->allocb_fail++;
2401 				cmn_err(CE_WARN,
2402 					"%s: mbuf allocation failure "
2403 					"count = %llu", __func__,
2404 					(unsigned long long)fl->allocb_fail);
2405 				if (frame.head)
2406 					freemsgchain(frame.head);
2407 				fl->cidx = rcidx;
2408 				fl->offset = roffset;
2409 				return (NULL);
2410 			}
2411 			atomic_inc_uint(&rxb->ref_cnt);
2412 			if (frame.head != NULL)
2413 				frame.tail->b_cont = m;
2414 			else
2415 				frame.head = m;
2416 			frame.tail = m;
2417 		}
2418 		m->b_wptr += n;
2419 		len -= n;
2420 		offset += roundup(n, sc->sge.fl_align);
2421 		ASSERT(offset <= rxb->buf_size);
2422 		if (offset == rxb->buf_size) {
2423 			offset = 0;
2424 			if (++cidx == fl->cap)
2425 				cidx = 0;
2426 			nbuf++;
2427 		}
2428 	}
2429 
2430 	fl->cidx = cidx;
2431 	fl->offset = offset;
2432 	(*fl_bufs_used) += nbuf;
2433 
2434 	ASSERT(frame.head != NULL);
2435 	return (frame.head);
2436 }
2437 
2438 /*
2439  * We'll do immediate data tx for non-LSO, but only when not coalescing.  We're
2440  * willing to use upto 2 hardware descriptors which means a maximum of 96 bytes
2441  * of immediate data.
2442  */
2443 #define	IMM_LEN ( \
2444 	2 * EQ_ESIZE \
2445 	- sizeof (struct fw_eth_tx_pkt_wr) \
2446 	- sizeof (struct cpl_tx_pkt_core))
2447 
2448 /*
2449  * Returns non-zero on failure, no need to cleanup anything in that case.
2450  *
2451  * Note 1: We always try to pull up the mblk if required and return E2BIG only
2452  * if this fails.
2453  *
2454  * Note 2: We'll also pullup incoming mblk if HW_LSO is set and the first mblk
2455  * does not have the TCP header in it.
2456  */
2457 static int
2458 get_frame_txinfo(struct sge_txq *txq, mblk_t **fp, struct txinfo *txinfo,
2459     int sgl_only)
2460 {
2461 	uint32_t flags = 0, len, n;
2462 	mblk_t *m = *fp;
2463 	int rc;
2464 
2465 	TXQ_LOCK_ASSERT_OWNED(txq);	/* will manipulate txb and dma_hdls */
2466 
2467 	mac_hcksum_get(m, NULL, NULL, NULL, NULL, &flags);
2468 	txinfo->flags = (flags & HCK_TX_FLAGS);
2469 
2470 	mac_lso_get(m, &txinfo->mss, &flags);
2471 	txinfo->flags |= (flags & HW_LSO_FLAGS);
2472 
2473 	if (flags & HW_LSO)
2474 		sgl_only = 1;	/* Do not allow immediate data with LSO */
2475 
2476 	/*
2477 	 * If checksum or segmentation offloads are requested, gather
2478 	 * information about the sizes and types of headers in the packet.
2479 	 */
2480 	if (txinfo->flags != 0) {
2481 		/*
2482 		 * Even if this fails, the meoi_flags field will be capable of
2483 		 * communicating the lack of useful packet information.
2484 		 */
2485 		(void) mac_ether_offload_info(m, &txinfo->meoi);
2486 	} else {
2487 		bzero(&txinfo->meoi, sizeof (txinfo->meoi));
2488 	}
2489 
2490 start:	txinfo->nsegs = 0;
2491 	txinfo->hdls_used = 0;
2492 	txinfo->txb_used = 0;
2493 	txinfo->len = 0;
2494 
2495 	/* total length and a rough estimate of # of segments */
2496 	n = 0;
2497 	for (; m; m = m->b_cont) {
2498 		len = MBLKL(m);
2499 		n += (len / PAGE_SIZE) + 1;
2500 		txinfo->len += len;
2501 	}
2502 	m = *fp;
2503 
2504 	if (n >= TX_SGL_SEGS || (flags & HW_LSO && MBLKL(m) < 50)) {
2505 		txq->pullup_early++;
2506 		m = msgpullup(*fp, -1);
2507 		if (m == NULL) {
2508 			txq->pullup_failed++;
2509 			return (E2BIG);	/* (*fp) left as it was */
2510 		}
2511 		freemsg(*fp);
2512 		*fp = m;
2513 		mac_hcksum_set(m, 0, 0, 0, 0, txinfo->flags);
2514 	}
2515 
2516 	if (txinfo->len <= IMM_LEN && !sgl_only)
2517 		return (0);	/* nsegs = 0 tells caller to use imm. tx */
2518 
2519 	if (txinfo->len <= txq->copy_threshold &&
2520 	    copy_into_txb(txq, m, txinfo->len, txinfo) == 0)
2521 		goto done;
2522 
2523 	for (; m; m = m->b_cont) {
2524 
2525 		len = MBLKL(m);
2526 
2527 		/* Use tx copy buffer if this mblk is small enough */
2528 		if (len <= txq->copy_threshold &&
2529 		    copy_into_txb(txq, m, len, txinfo) == 0)
2530 			continue;
2531 
2532 		/* Add DMA bindings for this mblk to the SGL */
2533 		rc = add_mblk(txq, txinfo, m, len);
2534 
2535 		if (rc == E2BIG ||
2536 		    (txinfo->nsegs == TX_SGL_SEGS && m->b_cont)) {
2537 
2538 			txq->pullup_late++;
2539 			m = msgpullup(*fp, -1);
2540 			if (m != NULL) {
2541 				free_txinfo_resources(txq, txinfo);
2542 				freemsg(*fp);
2543 				*fp = m;
2544 				mac_hcksum_set(m, 0, 0, 0, 0, txinfo->flags);
2545 				goto start;
2546 			}
2547 
2548 			txq->pullup_failed++;
2549 			rc = E2BIG;
2550 		}
2551 
2552 		if (rc != 0) {
2553 			free_txinfo_resources(txq, txinfo);
2554 			return (rc);
2555 		}
2556 	}
2557 
2558 	ASSERT(txinfo->nsegs > 0 && txinfo->nsegs <= TX_SGL_SEGS);
2559 
2560 done:
2561 
2562 	/*
2563 	 * Store the # of flits required to hold this frame's SGL in nflits.  An
2564 	 * SGL has a (ULPTX header + len0, addr0) tuple optionally followed by
2565 	 * multiple (len0 + len1, addr0, addr1) tuples.  If addr1 is not used
2566 	 * then len1 must be set to 0.
2567 	 */
2568 	n = txinfo->nsegs - 1;
2569 	txinfo->nflits = (3 * n) / 2 + (n & 1) + 2;
2570 	if (n & 1)
2571 		txinfo->sgl.sge[n / 2].len[1] = cpu_to_be32(0);
2572 
2573 	txinfo->sgl.cmd_nsge = cpu_to_be32(V_ULPTX_CMD((u32)ULP_TX_SC_DSGL) |
2574 	    V_ULPTX_NSGE(txinfo->nsegs));
2575 
2576 	return (0);
2577 }
2578 
2579 static inline int
2580 fits_in_txb(struct sge_txq *txq, int len, int *waste)
2581 {
2582 	if (txq->txb_avail < len)
2583 		return (0);
2584 
2585 	if (txq->txb_next + len <= txq->txb_size) {
2586 		*waste = 0;
2587 		return (1);
2588 	}
2589 
2590 	*waste = txq->txb_size - txq->txb_next;
2591 
2592 	return (txq->txb_avail - *waste < len ? 0 : 1);
2593 }
2594 
2595 #define	TXB_CHUNK	64
2596 
2597 /*
2598  * Copies the specified # of bytes into txq's tx copy buffer and updates txinfo
2599  * and txq to indicate resources used.  Caller has to make sure that those many
2600  * bytes are available in the mblk chain (b_cont linked).
2601  */
2602 static inline int
2603 copy_into_txb(struct sge_txq *txq, mblk_t *m, int len, struct txinfo *txinfo)
2604 {
2605 	int waste, n;
2606 
2607 	TXQ_LOCK_ASSERT_OWNED(txq);	/* will manipulate txb */
2608 
2609 	if (!fits_in_txb(txq, len, &waste)) {
2610 		txq->txb_full++;
2611 		return (ENOMEM);
2612 	}
2613 
2614 	if (waste != 0) {
2615 		ASSERT((waste & (TXB_CHUNK - 1)) == 0);
2616 		txinfo->txb_used += waste;
2617 		txq->txb_avail -= waste;
2618 		txq->txb_next = 0;
2619 	}
2620 
2621 	for (n = 0; n < len; m = m->b_cont) {
2622 		bcopy(m->b_rptr, txq->txb_va + txq->txb_next + n, MBLKL(m));
2623 		n += MBLKL(m);
2624 	}
2625 
2626 	add_seg(txinfo, txq->txb_ba + txq->txb_next, len);
2627 
2628 	n = roundup(len, TXB_CHUNK);
2629 	txinfo->txb_used += n;
2630 	txq->txb_avail -= n;
2631 	txq->txb_next += n;
2632 	ASSERT(txq->txb_next <= txq->txb_size);
2633 	if (txq->txb_next == txq->txb_size)
2634 		txq->txb_next = 0;
2635 
2636 	return (0);
2637 }
2638 
2639 static inline void
2640 add_seg(struct txinfo *txinfo, uint64_t ba, uint32_t len)
2641 {
2642 	ASSERT(txinfo->nsegs < TX_SGL_SEGS);	/* must have room */
2643 
2644 	if (txinfo->nsegs != 0) {
2645 		int idx = txinfo->nsegs - 1;
2646 		txinfo->sgl.sge[idx / 2].len[idx & 1] = cpu_to_be32(len);
2647 		txinfo->sgl.sge[idx / 2].addr[idx & 1] = cpu_to_be64(ba);
2648 	} else {
2649 		txinfo->sgl.len0 = cpu_to_be32(len);
2650 		txinfo->sgl.addr0 = cpu_to_be64(ba);
2651 	}
2652 	txinfo->nsegs++;
2653 }
2654 
2655 /*
2656  * This function cleans up any partially allocated resources when it fails so
2657  * there's nothing for the caller to clean up in that case.
2658  *
2659  * EIO indicates permanent failure.  Caller should drop the frame containing
2660  * this mblk and continue.
2661  *
2662  * E2BIG indicates that the SGL length for this mblk exceeds the hardware
2663  * limit.  Caller should pull up the frame before trying to send it out.
2664  * (This error means our pullup_early heuristic did not work for this frame)
2665  *
2666  * ENOMEM indicates a temporary shortage of resources (DMA handles, other DMA
2667  * resources, etc.).  Caller should suspend the tx queue and wait for reclaim to
2668  * free up resources.
2669  */
2670 static inline int
2671 add_mblk(struct sge_txq *txq, struct txinfo *txinfo, mblk_t *m, int len)
2672 {
2673 	ddi_dma_handle_t dhdl;
2674 	ddi_dma_cookie_t cookie;
2675 	uint_t ccount = 0;
2676 	int rc;
2677 
2678 	TXQ_LOCK_ASSERT_OWNED(txq);	/* will manipulate dhdls */
2679 
2680 	if (txq->tx_dhdl_avail == 0) {
2681 		txq->dma_hdl_failed++;
2682 		return (ENOMEM);
2683 	}
2684 
2685 	dhdl = txq->tx_dhdl[txq->tx_dhdl_pidx];
2686 	rc = ddi_dma_addr_bind_handle(dhdl, NULL, (caddr_t)m->b_rptr, len,
2687 	    DDI_DMA_WRITE | DDI_DMA_STREAMING, DDI_DMA_DONTWAIT, NULL, &cookie,
2688 	    &ccount);
2689 	if (rc != DDI_DMA_MAPPED) {
2690 		txq->dma_map_failed++;
2691 
2692 		ASSERT(rc != DDI_DMA_INUSE && rc != DDI_DMA_PARTIAL_MAP);
2693 
2694 		return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EIO);
2695 	}
2696 
2697 	if (ccount + txinfo->nsegs > TX_SGL_SEGS) {
2698 		(void) ddi_dma_unbind_handle(dhdl);
2699 		return (E2BIG);
2700 	}
2701 
2702 	add_seg(txinfo, cookie.dmac_laddress, cookie.dmac_size);
2703 	while (--ccount) {
2704 		ddi_dma_nextcookie(dhdl, &cookie);
2705 		add_seg(txinfo, cookie.dmac_laddress, cookie.dmac_size);
2706 	}
2707 
2708 	if (++txq->tx_dhdl_pidx == txq->tx_dhdl_total)
2709 		txq->tx_dhdl_pidx = 0;
2710 	txq->tx_dhdl_avail--;
2711 	txinfo->hdls_used++;
2712 
2713 	return (0);
2714 }
2715 
2716 /*
2717  * Releases all the txq resources used up in the specified txinfo.
2718  */
2719 static void
2720 free_txinfo_resources(struct sge_txq *txq, struct txinfo *txinfo)
2721 {
2722 	int n;
2723 
2724 	TXQ_LOCK_ASSERT_OWNED(txq);	/* dhdls, txb */
2725 
2726 	n = txinfo->txb_used;
2727 	if (n > 0) {
2728 		txq->txb_avail += n;
2729 		if (n <= txq->txb_next)
2730 			txq->txb_next -= n;
2731 		else {
2732 			n -= txq->txb_next;
2733 			txq->txb_next = txq->txb_size - n;
2734 		}
2735 	}
2736 
2737 	for (n = txinfo->hdls_used; n > 0; n--) {
2738 		if (txq->tx_dhdl_pidx > 0)
2739 			txq->tx_dhdl_pidx--;
2740 		else
2741 			txq->tx_dhdl_pidx = txq->tx_dhdl_total - 1;
2742 		txq->tx_dhdl_avail++;
2743 		(void) ddi_dma_unbind_handle(txq->tx_dhdl[txq->tx_dhdl_pidx]);
2744 	}
2745 }
2746 
2747 /*
2748  * Returns 0 to indicate that m has been accepted into a coalesced tx work
2749  * request.  It has either been folded into txpkts or txpkts was flushed and m
2750  * has started a new coalesced work request (as the first frame in a fresh
2751  * txpkts).
2752  *
2753  * Returns non-zero to indicate a failure - caller is responsible for
2754  * transmitting m, if there was anything in txpkts it has been flushed.
2755  */
2756 static int
2757 add_to_txpkts(struct sge_txq *txq, struct txpkts *txpkts, mblk_t *m,
2758     struct txinfo *txinfo)
2759 {
2760 	struct sge_eq *eq = &txq->eq;
2761 	int can_coalesce;
2762 	struct tx_sdesc *txsd;
2763 	uint8_t flits;
2764 
2765 	TXQ_LOCK_ASSERT_OWNED(txq);
2766 
2767 	if (txpkts->npkt > 0) {
2768 		flits = TXPKTS_PKT_HDR + txinfo->nflits;
2769 		can_coalesce = (txinfo->flags & HW_LSO) == 0 &&
2770 		    txpkts->nflits + flits <= TX_WR_FLITS &&
2771 		    txpkts->nflits + flits <= eq->avail * 8 &&
2772 		    txpkts->plen + txinfo->len < 65536;
2773 
2774 		if (can_coalesce != 0) {
2775 			txpkts->tail->b_next = m;
2776 			txpkts->tail = m;
2777 			txpkts->npkt++;
2778 			txpkts->nflits += flits;
2779 			txpkts->plen += txinfo->len;
2780 
2781 			txsd = &txq->sdesc[eq->pidx];
2782 			txsd->txb_used += txinfo->txb_used;
2783 			txsd->hdls_used += txinfo->hdls_used;
2784 
2785 			return (0);
2786 		}
2787 
2788 		/*
2789 		 * Couldn't coalesce m into txpkts.  The first order of business
2790 		 * is to send txpkts on its way.  Then we'll revisit m.
2791 		 */
2792 		write_txpkts_wr(txq, txpkts);
2793 	}
2794 
2795 	/*
2796 	 * Check if we can start a new coalesced tx work request with m as
2797 	 * the first packet in it.
2798 	 */
2799 
2800 	ASSERT(txpkts->npkt == 0);
2801 	ASSERT(txinfo->len < 65536);
2802 
2803 	flits = TXPKTS_WR_HDR + txinfo->nflits;
2804 	can_coalesce = (txinfo->flags & HW_LSO) == 0 &&
2805 	    flits <= eq->avail * 8 && flits <= TX_WR_FLITS;
2806 
2807 	if (can_coalesce == 0)
2808 		return (EINVAL);
2809 
2810 	/*
2811 	 * Start a fresh coalesced tx WR with m as the first frame in it.
2812 	 */
2813 	txpkts->tail = m;
2814 	txpkts->npkt = 1;
2815 	txpkts->nflits = flits;
2816 	txpkts->flitp = &eq->desc[eq->pidx].flit[2];
2817 	txpkts->plen = txinfo->len;
2818 
2819 	txsd = &txq->sdesc[eq->pidx];
2820 	txsd->m = m;
2821 	txsd->txb_used = txinfo->txb_used;
2822 	txsd->hdls_used = txinfo->hdls_used;
2823 
2824 	return (0);
2825 }
2826 
2827 /*
2828  * Note that write_txpkts_wr can never run out of hardware descriptors (but
2829  * write_txpkt_wr can).  add_to_txpkts ensures that a frame is accepted for
2830  * coalescing only if sufficient hardware descriptors are available.
2831  */
2832 static void
2833 write_txpkts_wr(struct sge_txq *txq, struct txpkts *txpkts)
2834 {
2835 	struct sge_eq *eq = &txq->eq;
2836 	struct fw_eth_tx_pkts_wr *wr;
2837 	struct tx_sdesc *txsd;
2838 	uint32_t ctrl;
2839 	uint16_t ndesc;
2840 
2841 	TXQ_LOCK_ASSERT_OWNED(txq);	/* pidx, avail */
2842 
2843 	ndesc = howmany(txpkts->nflits, 8);
2844 
2845 	wr = (void *)&eq->desc[eq->pidx];
2846 	wr->op_pkd = cpu_to_be32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR) |
2847 	    V_FW_WR_IMMDLEN(0)); /* immdlen does not matter in this WR */
2848 	ctrl = V_FW_WR_LEN16(howmany(txpkts->nflits, 2));
2849 	if (eq->avail == ndesc)
2850 		ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ;
2851 	wr->equiq_to_len16 = cpu_to_be32(ctrl);
2852 	wr->plen = cpu_to_be16(txpkts->plen);
2853 	wr->npkt = txpkts->npkt;
2854 	wr->r3 = wr->type = 0;
2855 
2856 	/* Everything else already written */
2857 
2858 	txsd = &txq->sdesc[eq->pidx];
2859 	txsd->desc_used = ndesc;
2860 
2861 	txq->txb_used += txsd->txb_used / TXB_CHUNK;
2862 	txq->hdl_used += txsd->hdls_used;
2863 
2864 	ASSERT(eq->avail >= ndesc);
2865 
2866 	eq->pending += ndesc;
2867 	eq->avail -= ndesc;
2868 	eq->pidx += ndesc;
2869 	if (eq->pidx >= eq->cap)
2870 		eq->pidx -= eq->cap;
2871 
2872 	txq->txpkts_pkts += txpkts->npkt;
2873 	txq->txpkts_wrs++;
2874 	txpkts->npkt = 0;	/* emptied */
2875 }
2876 
2877 typedef enum {
2878 	COS_SUCCESS,	/* ctrl flit contains proper bits for csum offload */
2879 	COS_IGNORE,	/* no csum offload requested */
2880 	COS_FAIL,	/* csum offload requested, but pkt data missing */
2881 } csum_offload_status_t;
2882 /*
2883  * Build a ctrl1 flit for checksum offload in CPL_TX_PKT_XT command
2884  */
2885 static csum_offload_status_t
2886 csum_to_ctrl(const struct txinfo *txinfo, uint32_t chip_version,
2887     uint64_t *ctrlp)
2888 {
2889 	const mac_ether_offload_info_t *meoi = &txinfo->meoi;
2890 	const uint32_t tx_flags = txinfo->flags;
2891 	const boolean_t needs_l3_csum = (tx_flags & HW_LSO) != 0 ||
2892 	    (tx_flags & HCK_IPV4_HDRCKSUM) != 0;
2893 	const boolean_t needs_l4_csum = (tx_flags & HW_LSO) != 0 ||
2894 	    (tx_flags & (HCK_FULLCKSUM | HCK_PARTIALCKSUM)) != 0;
2895 
2896 	/*
2897 	 * Default to disabling any checksumming both for cases where it is not
2898 	 * requested, but also if we cannot appropriately interrogate the
2899 	 * required information from the packet.
2900 	 */
2901 	uint64_t ctrl = F_TXPKT_L4CSUM_DIS | F_TXPKT_IPCSUM_DIS;
2902 	if (!needs_l3_csum && !needs_l4_csum) {
2903 		*ctrlp = ctrl;
2904 		return (COS_IGNORE);
2905 	}
2906 
2907 	if (needs_l3_csum) {
2908 		/* Only IPv4 checksums are supported (for L3) */
2909 		if ((meoi->meoi_flags & MEOI_L3INFO_SET) == 0 ||
2910 		    meoi->meoi_l3proto != ETHERTYPE_IP) {
2911 			*ctrlp = ctrl;
2912 			return (COS_FAIL);
2913 		}
2914 		ctrl &= ~F_TXPKT_IPCSUM_DIS;
2915 	}
2916 
2917 	if (needs_l4_csum) {
2918 		/*
2919 		 * We need at least all of the L3 header to make decisions about
2920 		 * the contained L4 protocol.  If not all of the L4 information
2921 		 * is present, we will leave it to the NIC to checksum all it is
2922 		 * able to.
2923 		 */
2924 		if ((meoi->meoi_flags & MEOI_L3INFO_SET) == 0) {
2925 			*ctrlp = ctrl;
2926 			return (COS_FAIL);
2927 		}
2928 
2929 		/*
2930 		 * Since we are parsing the packet anyways, make the checksum
2931 		 * decision based on the L4 protocol, rather than using the
2932 		 * Generic TCP/UDP checksum using start & end offsets in the
2933 		 * packet (like requested with PARTIALCKSUM).
2934 		 */
2935 		int csum_type = -1;
2936 		if (meoi->meoi_l3proto == ETHERTYPE_IP &&
2937 		    meoi->meoi_l4proto == IPPROTO_TCP) {
2938 			csum_type = TX_CSUM_TCPIP;
2939 		} else if (meoi->meoi_l3proto == ETHERTYPE_IPV6 &&
2940 		    meoi->meoi_l4proto == IPPROTO_TCP) {
2941 			csum_type = TX_CSUM_TCPIP6;
2942 		} else if (meoi->meoi_l3proto == ETHERTYPE_IP &&
2943 		    meoi->meoi_l4proto == IPPROTO_UDP) {
2944 			csum_type = TX_CSUM_UDPIP;
2945 		} else if (meoi->meoi_l3proto == ETHERTYPE_IPV6 &&
2946 		    meoi->meoi_l4proto == IPPROTO_UDP) {
2947 			csum_type = TX_CSUM_UDPIP6;
2948 		} else {
2949 			*ctrlp = ctrl;
2950 			return (COS_FAIL);
2951 		}
2952 
2953 		ASSERT(csum_type != -1);
2954 		ctrl &= ~F_TXPKT_L4CSUM_DIS;
2955 		ctrl |= V_TXPKT_CSUM_TYPE(csum_type);
2956 	}
2957 
2958 	if ((ctrl & F_TXPKT_IPCSUM_DIS) == 0 &&
2959 	    (ctrl & F_TXPKT_L4CSUM_DIS) != 0) {
2960 		/*
2961 		 * If only the IPv4 checksum is requested, we need to set an
2962 		 * appropriate type in the command for it.
2963 		 */
2964 		ctrl |= V_TXPKT_CSUM_TYPE(TX_CSUM_IP);
2965 	}
2966 
2967 	ASSERT(ctrl != (F_TXPKT_L4CSUM_DIS | F_TXPKT_IPCSUM_DIS));
2968 
2969 	/*
2970 	 * Fill in the requisite L2/L3 header length data.
2971 	 *
2972 	 * The Ethernet header length is recorded as 'size - 14 bytes'
2973 	 */
2974 	const uint8_t eth_len = meoi->meoi_l2hlen - 14;
2975 	if (chip_version >= CHELSIO_T6) {
2976 		ctrl |= V_T6_TXPKT_ETHHDR_LEN(eth_len);
2977 	} else {
2978 		ctrl |= V_TXPKT_ETHHDR_LEN(eth_len);
2979 	}
2980 	ctrl |= V_TXPKT_IPHDR_LEN(meoi->meoi_l3hlen);
2981 
2982 	*ctrlp = ctrl;
2983 	return (COS_SUCCESS);
2984 }
2985 
2986 static int
2987 write_txpkt_wr(struct port_info *pi, struct sge_txq *txq, mblk_t *m,
2988     struct txinfo *txinfo)
2989 {
2990 	struct sge_eq *eq = &txq->eq;
2991 	struct fw_eth_tx_pkt_wr *wr;
2992 	struct cpl_tx_pkt_core *cpl;
2993 	uint32_t ctrl;	/* used in many unrelated places */
2994 	uint64_t ctrl1;
2995 	int nflits, ndesc;
2996 	struct tx_sdesc *txsd;
2997 	caddr_t dst;
2998 	const mac_ether_offload_info_t *meoi = &txinfo->meoi;
2999 
3000 	TXQ_LOCK_ASSERT_OWNED(txq);	/* pidx, avail */
3001 
3002 	/*
3003 	 * Do we have enough flits to send this frame out?
3004 	 */
3005 	ctrl = sizeof (struct cpl_tx_pkt_core);
3006 	if (txinfo->flags & HW_LSO) {
3007 		nflits = TXPKT_LSO_WR_HDR;
3008 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
3009 	} else
3010 		nflits = TXPKT_WR_HDR;
3011 	if (txinfo->nsegs > 0)
3012 		nflits += txinfo->nflits;
3013 	else {
3014 		nflits += howmany(txinfo->len, 8);
3015 		ctrl += txinfo->len;
3016 	}
3017 	ndesc = howmany(nflits, 8);
3018 	if (ndesc > eq->avail)
3019 		return (ENOMEM);
3020 
3021 	/* Firmware work request header */
3022 	wr = (void *)&eq->desc[eq->pidx];
3023 	wr->op_immdlen = cpu_to_be32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
3024 	    V_FW_WR_IMMDLEN(ctrl));
3025 	ctrl = V_FW_WR_LEN16(howmany(nflits, 2));
3026 	if (eq->avail == ndesc)
3027 		ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ;
3028 	wr->equiq_to_len16 = cpu_to_be32(ctrl);
3029 	wr->r3 = 0;
3030 
3031 	if (txinfo->flags & HW_LSO &&
3032 	    (meoi->meoi_flags & MEOI_L4INFO_SET) != 0 &&
3033 	    meoi->meoi_l4proto == IPPROTO_TCP) {
3034 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
3035 
3036 		ctrl = V_LSO_OPCODE((u32)CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
3037 		    F_LSO_LAST_SLICE;
3038 
3039 		if (meoi->meoi_l2hlen > sizeof (struct ether_header)) {
3040 			/*
3041 			 * This presently assumes a standard VLAN header,
3042 			 * without support for Q-in-Q.
3043 			 */
3044 			ctrl |= V_LSO_ETHHDR_LEN(1);
3045 		}
3046 
3047 		switch (meoi->meoi_l3proto) {
3048 		case ETHERTYPE_IPV6:
3049 			ctrl |= F_LSO_IPV6;
3050 			/* FALLTHROUGH */
3051 		case ETHERTYPE_IP:
3052 			ctrl |= V_LSO_IPHDR_LEN(meoi->meoi_l3hlen / 4);
3053 			break;
3054 		default:
3055 			break;
3056 		}
3057 
3058 		ctrl |= V_LSO_TCPHDR_LEN(meoi->meoi_l4hlen / 4);
3059 
3060 		lso->lso_ctrl = cpu_to_be32(ctrl);
3061 		lso->ipid_ofst = cpu_to_be16(0);
3062 		lso->mss = cpu_to_be16(txinfo->mss);
3063 		lso->seqno_offset = cpu_to_be32(0);
3064 		if (is_t4(pi->adapter->params.chip))
3065 			lso->len = cpu_to_be32(txinfo->len);
3066 		else
3067 			lso->len = cpu_to_be32(V_LSO_T5_XFER_SIZE(txinfo->len));
3068 
3069 		cpl = (void *)(lso + 1);
3070 
3071 		txq->tso_wrs++;
3072 	} else {
3073 		cpl = (void *)(wr + 1);
3074 	}
3075 
3076 	/* Checksum offload */
3077 	switch (csum_to_ctrl(txinfo,
3078 	    CHELSIO_CHIP_VERSION(pi->adapter->params.chip), &ctrl1)) {
3079 	case COS_SUCCESS:
3080 		txq->txcsum++;
3081 		break;
3082 	case COS_FAIL:
3083 		/*
3084 		 * Packet will be going out with checksums which are probably
3085 		 * wrong but there is little we can do now.
3086 		 */
3087 		txq->csum_failed++;
3088 		break;
3089 	default:
3090 		break;
3091 	}
3092 
3093 	/* CPL header */
3094 	cpl->ctrl0 = cpu_to_be32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
3095 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf));
3096 	cpl->pack = 0;
3097 	cpl->len = cpu_to_be16(txinfo->len);
3098 	cpl->ctrl1 = cpu_to_be64(ctrl1);
3099 
3100 	/* Software descriptor */
3101 	txsd = &txq->sdesc[eq->pidx];
3102 	txsd->m = m;
3103 	txsd->txb_used = txinfo->txb_used;
3104 	txsd->hdls_used = txinfo->hdls_used;
3105 	/* LINTED: E_ASSIGN_NARROW_CONV */
3106 	txsd->desc_used = ndesc;
3107 
3108 	txq->txb_used += txinfo->txb_used / TXB_CHUNK;
3109 	txq->hdl_used += txinfo->hdls_used;
3110 
3111 	eq->pending += ndesc;
3112 	eq->avail -= ndesc;
3113 	eq->pidx += ndesc;
3114 	if (eq->pidx >= eq->cap)
3115 		eq->pidx -= eq->cap;
3116 
3117 	/* SGL */
3118 	dst = (void *)(cpl + 1);
3119 	if (txinfo->nsegs > 0) {
3120 		txq->sgl_wrs++;
3121 		copy_to_txd(eq, (void *)&txinfo->sgl, &dst, txinfo->nflits * 8);
3122 
3123 		/* Need to zero-pad to a 16 byte boundary if not on one */
3124 		if ((uintptr_t)dst & 0xf)
3125 			/* LINTED: E_BAD_PTR_CAST_ALIGN */
3126 			*(uint64_t *)dst = 0;
3127 
3128 	} else {
3129 		txq->imm_wrs++;
3130 #ifdef DEBUG
3131 		ctrl = txinfo->len;
3132 #endif
3133 		for (; m; m = m->b_cont) {
3134 			copy_to_txd(eq, (void *)m->b_rptr, &dst, MBLKL(m));
3135 #ifdef DEBUG
3136 			ctrl -= MBLKL(m);
3137 #endif
3138 		}
3139 		ASSERT(ctrl == 0);
3140 	}
3141 
3142 	txq->txpkt_wrs++;
3143 	return (0);
3144 }
3145 
3146 static inline void
3147 write_ulp_cpl_sgl(struct port_info *pi, struct sge_txq *txq,
3148     struct txpkts *txpkts, struct txinfo *txinfo)
3149 {
3150 	struct ulp_txpkt *ulpmc;
3151 	struct ulptx_idata *ulpsc;
3152 	struct cpl_tx_pkt_core *cpl;
3153 	uintptr_t flitp, start, end;
3154 	uint64_t ctrl;
3155 	caddr_t dst;
3156 
3157 	ASSERT(txpkts->npkt > 0);
3158 
3159 	start = (uintptr_t)txq->eq.desc;
3160 	end = (uintptr_t)txq->eq.spg;
3161 
3162 	/* Checksum offload */
3163 	switch (csum_to_ctrl(txinfo,
3164 	    CHELSIO_CHIP_VERSION(pi->adapter->params.chip), &ctrl)) {
3165 	case COS_SUCCESS:
3166 		txq->txcsum++;
3167 		break;
3168 	case COS_FAIL:
3169 		/*
3170 		 * Packet will be going out with checksums which are probably
3171 		 * wrong but there is little we can do now.
3172 		 */
3173 		txq->csum_failed++;
3174 		break;
3175 	default:
3176 		break;
3177 	}
3178 
3179 	/*
3180 	 * The previous packet's SGL must have ended at a 16 byte boundary (this
3181 	 * is required by the firmware/hardware).  It follows that flitp cannot
3182 	 * wrap around between the ULPTX master command and ULPTX subcommand (8
3183 	 * bytes each), and that it can not wrap around in the middle of the
3184 	 * cpl_tx_pkt_core either.
3185 	 */
3186 	flitp = (uintptr_t)txpkts->flitp;
3187 	ASSERT((flitp & 0xf) == 0);
3188 
3189 	/* ULP master command */
3190 	ulpmc = (void *)flitp;
3191 	ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0));
3192 	ulpmc->len = htonl(howmany(sizeof (*ulpmc) + sizeof (*ulpsc) +
3193 	    sizeof (*cpl) + 8 * txinfo->nflits, 16));
3194 
3195 	/* ULP subcommand */
3196 	ulpsc = (void *)(ulpmc + 1);
3197 	ulpsc->cmd_more = cpu_to_be32(V_ULPTX_CMD((u32)ULP_TX_SC_IMM) |
3198 	    F_ULP_TX_SC_MORE);
3199 	ulpsc->len = cpu_to_be32(sizeof (struct cpl_tx_pkt_core));
3200 
3201 	flitp += sizeof (*ulpmc) + sizeof (*ulpsc);
3202 	if (flitp == end)
3203 		flitp = start;
3204 
3205 	/* CPL_TX_PKT_XT */
3206 	cpl = (void *)flitp;
3207 	cpl->ctrl0 = cpu_to_be32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
3208 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf));
3209 	cpl->pack = 0;
3210 	cpl->len = cpu_to_be16(txinfo->len);
3211 	cpl->ctrl1 = cpu_to_be64(ctrl);
3212 
3213 	flitp += sizeof (*cpl);
3214 	if (flitp == end)
3215 		flitp = start;
3216 
3217 	/* SGL for this frame */
3218 	dst = (caddr_t)flitp;
3219 	copy_to_txd(&txq->eq, (void *)&txinfo->sgl, &dst, txinfo->nflits * 8);
3220 	flitp = (uintptr_t)dst;
3221 
3222 	/* Zero pad and advance to a 16 byte boundary if not already at one. */
3223 	if (flitp & 0xf) {
3224 
3225 		/* no matter what, flitp should be on an 8 byte boundary */
3226 		ASSERT((flitp & 0x7) == 0);
3227 
3228 		*(uint64_t *)flitp = 0;
3229 		flitp += sizeof (uint64_t);
3230 		txpkts->nflits++;
3231 	}
3232 
3233 	if (flitp == end)
3234 		flitp = start;
3235 
3236 	txpkts->flitp = (void *)flitp;
3237 }
3238 
3239 static inline void
3240 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
3241 {
3242 	if ((uintptr_t)(*to) + len <= (uintptr_t)eq->spg) {
3243 		bcopy(from, *to, len);
3244 		(*to) += len;
3245 	} else {
3246 		int portion = (uintptr_t)eq->spg - (uintptr_t)(*to);
3247 
3248 		bcopy(from, *to, portion);
3249 		from += portion;
3250 		portion = len - portion;	/* remaining */
3251 		bcopy(from, (void *)eq->desc, portion);
3252 		(*to) = (caddr_t)eq->desc + portion;
3253 	}
3254 }
3255 
3256 static inline void
3257 ring_tx_db(struct adapter *sc, struct sge_eq *eq)
3258 {
3259 	int val, db_mode;
3260 	u_int db = eq->doorbells;
3261 
3262 	if (eq->pending > 1)
3263 		db &= ~DOORBELL_WCWR;
3264 
3265 	if (eq->pending > eq->pidx) {
3266 		int offset = eq->cap - (eq->pending - eq->pidx);
3267 
3268 		/* pidx has wrapped around since last doorbell */
3269 
3270 		(void) ddi_dma_sync(eq->desc_dhdl,
3271 		    offset * sizeof (struct tx_desc), 0,
3272 		    DDI_DMA_SYNC_FORDEV);
3273 		(void) ddi_dma_sync(eq->desc_dhdl,
3274 		    0, eq->pidx * sizeof (struct tx_desc),
3275 		    DDI_DMA_SYNC_FORDEV);
3276 	} else if (eq->pending > 0) {
3277 		(void) ddi_dma_sync(eq->desc_dhdl,
3278 		    (eq->pidx - eq->pending) * sizeof (struct tx_desc),
3279 		    eq->pending * sizeof (struct tx_desc),
3280 		    DDI_DMA_SYNC_FORDEV);
3281 	}
3282 
3283 	membar_producer();
3284 
3285 	if (is_t4(sc->params.chip))
3286 		val = V_PIDX(eq->pending);
3287 	else
3288 		val = V_PIDX_T5(eq->pending);
3289 
3290 	db_mode = (1 << (ffs(db) - 1));
3291 	switch (db_mode) {
3292 		case DOORBELL_UDB:
3293 			*eq->udb = LE_32(V_QID(eq->udb_qid) | val);
3294 			break;
3295 
3296 		case DOORBELL_WCWR:
3297 			{
3298 				volatile uint64_t *dst, *src;
3299 				int i;
3300 				/*
3301 				 * Queues whose 128B doorbell segment fits in
3302 				 * the page do not use relative qid
3303 				 * (udb_qid is always 0).  Only queues with
3304 				 * doorbell segments can do WCWR.
3305 				 */
3306 				ASSERT(eq->udb_qid == 0 && eq->pending == 1);
3307 
3308 				dst = (volatile void *)((uintptr_t)eq->udb +
3309 				    UDBS_WR_OFFSET - UDBS_DB_OFFSET);
3310 				i = eq->pidx ? eq->pidx - 1 : eq->cap - 1;
3311 				src = (void *)&eq->desc[i];
3312 				while (src != (void *)&eq->desc[i + 1])
3313 				        *dst++ = *src++;
3314 				membar_producer();
3315 				break;
3316 			}
3317 
3318 		case DOORBELL_UDBWC:
3319 			*eq->udb = LE_32(V_QID(eq->udb_qid) | val);
3320 			membar_producer();
3321 			break;
3322 
3323 		case DOORBELL_KDB:
3324 			t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
3325 			    V_QID(eq->cntxt_id) | val);
3326 			break;
3327 	}
3328 
3329 	eq->pending = 0;
3330 }
3331 
3332 static int
3333 reclaim_tx_descs(struct sge_txq *txq, int howmany)
3334 {
3335 	struct tx_sdesc *txsd;
3336 	uint_t cidx, can_reclaim, reclaimed, txb_freed, hdls_freed;
3337 	struct sge_eq *eq = &txq->eq;
3338 
3339 	EQ_LOCK_ASSERT_OWNED(eq);
3340 
3341 	cidx = eq->spg->cidx;	/* stable snapshot */
3342 	cidx = be16_to_cpu(cidx);
3343 
3344 	if (cidx >= eq->cidx)
3345 		can_reclaim = cidx - eq->cidx;
3346 	else
3347 		can_reclaim = cidx + eq->cap - eq->cidx;
3348 
3349 	if (can_reclaim == 0)
3350 		return (0);
3351 
3352 	txb_freed = hdls_freed = reclaimed = 0;
3353 	do {
3354 		int ndesc;
3355 
3356 		txsd = &txq->sdesc[eq->cidx];
3357 		ndesc = txsd->desc_used;
3358 
3359 		/* Firmware doesn't return "partial" credits. */
3360 		ASSERT(can_reclaim >= ndesc);
3361 
3362 		/*
3363 		 * We always keep mblk around, even for immediate data.  If mblk
3364 		 * is NULL, this has to be the software descriptor for a credit
3365 		 * flush work request.
3366 		 */
3367 		if (txsd->m != NULL)
3368 			freemsgchain(txsd->m);
3369 #ifdef DEBUG
3370 		else {
3371 			ASSERT(txsd->txb_used == 0);
3372 			ASSERT(txsd->hdls_used == 0);
3373 			ASSERT(ndesc == 1);
3374 		}
3375 #endif
3376 
3377 		txb_freed += txsd->txb_used;
3378 		hdls_freed += txsd->hdls_used;
3379 		reclaimed += ndesc;
3380 
3381 		eq->cidx += ndesc;
3382 		if (eq->cidx >= eq->cap)
3383 			eq->cidx -= eq->cap;
3384 
3385 		can_reclaim -= ndesc;
3386 
3387 	} while (can_reclaim && reclaimed < howmany);
3388 
3389 	eq->avail += reclaimed;
3390 	ASSERT(eq->avail < eq->cap);	/* avail tops out at (cap - 1) */
3391 
3392 	txq->txb_avail += txb_freed;
3393 
3394 	txq->tx_dhdl_avail += hdls_freed;
3395 	ASSERT(txq->tx_dhdl_avail <= txq->tx_dhdl_total);
3396 	for (; hdls_freed; hdls_freed--) {
3397 		(void) ddi_dma_unbind_handle(txq->tx_dhdl[txq->tx_dhdl_cidx]);
3398 		if (++txq->tx_dhdl_cidx == txq->tx_dhdl_total)
3399 			txq->tx_dhdl_cidx = 0;
3400 	}
3401 
3402 	return (reclaimed);
3403 }
3404 
3405 static void
3406 write_txqflush_wr(struct sge_txq *txq)
3407 {
3408 	struct sge_eq *eq = &txq->eq;
3409 	struct fw_eq_flush_wr *wr;
3410 	struct tx_sdesc *txsd;
3411 
3412 	EQ_LOCK_ASSERT_OWNED(eq);
3413 	ASSERT(eq->avail > 0);
3414 
3415 	wr = (void *)&eq->desc[eq->pidx];
3416 	bzero(wr, sizeof (*wr));
3417 	wr->opcode = FW_EQ_FLUSH_WR;
3418 	wr->equiq_to_len16 = cpu_to_be32(V_FW_WR_LEN16(sizeof (*wr) / 16) |
3419 	    F_FW_WR_EQUEQ | F_FW_WR_EQUIQ);
3420 
3421 	txsd = &txq->sdesc[eq->pidx];
3422 	txsd->m = NULL;
3423 	txsd->txb_used = 0;
3424 	txsd->hdls_used = 0;
3425 	txsd->desc_used = 1;
3426 
3427 	eq->pending++;
3428 	eq->avail--;
3429 	if (++eq->pidx == eq->cap)
3430 		eq->pidx = 0;
3431 }
3432 
3433 static int
3434 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, mblk_t *m)
3435 {
3436 	bool csum_ok;
3437 	uint16_t err_vec;
3438 	struct sge_rxq *rxq = (void *)iq;
3439 	struct mblk_pair chain = {0};
3440 	struct adapter *sc = iq->adapter;
3441 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
3442 
3443 	iq->intr_next = iq->intr_params;
3444 
3445 	m->b_rptr += sc->sge.pktshift;
3446 
3447 	/* Compressed error vector is enabled for T6 only */
3448 	if (sc->params.tp.rx_pkt_encap)
3449 		/* It is enabled only in T6 config file */
3450 		err_vec = G_T6_COMPR_RXERR_VEC(ntohs(cpl->err_vec));
3451 	else
3452 		err_vec = ntohs(cpl->err_vec);
3453 
3454 	csum_ok = cpl->csum_calc && !err_vec;
3455 	/* TODO: what about cpl->ip_frag? */
3456 	if (csum_ok && !cpl->ip_frag) {
3457 		mac_hcksum_set(m, 0, 0, 0, 0xffff,
3458 		    HCK_FULLCKSUM_OK | HCK_FULLCKSUM |
3459 		    HCK_IPV4_HDRCKSUM_OK);
3460 		rxq->rxcsum++;
3461 	}
3462 
3463 	/* Add to the chain that we'll send up */
3464 	if (chain.head != NULL)
3465 		chain.tail->b_next = m;
3466 	else
3467 		chain.head = m;
3468 	chain.tail = m;
3469 
3470 	t4_mac_rx(rxq->port, rxq, chain.head);
3471 
3472 	rxq->rxpkts++;
3473 	rxq->rxbytes  += be16_to_cpu(cpl->len);
3474 	return (0);
3475 }
3476 
3477 #define	FL_HW_IDX(idx)	((idx) >> 3)
3478 
3479 static inline void
3480 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
3481 {
3482 	int desc_start, desc_last, ndesc;
3483 	uint32_t v = sc->params.arch.sge_fl_db ;
3484 
3485 	ndesc = FL_HW_IDX(fl->pending);
3486 
3487 	/* Hold back one credit if pidx = cidx */
3488 	if (FL_HW_IDX(fl->pidx) == FL_HW_IDX(fl->cidx))
3489 		ndesc--;
3490 
3491 	/*
3492 	 * There are chances of ndesc modified above (to avoid pidx = cidx).
3493 	 * If there is nothing to post, return.
3494 	 */
3495 	if (ndesc <= 0)
3496 		return;
3497 
3498 	desc_last = FL_HW_IDX(fl->pidx);
3499 
3500 	if (fl->pidx < fl->pending) {
3501 		/* There was a wrap */
3502 		desc_start = FL_HW_IDX(fl->pidx + fl->cap - fl->pending);
3503 
3504 		/* From desc_start to the end of list */
3505 		(void) ddi_dma_sync(fl->dhdl, desc_start * RX_FL_ESIZE, 0,
3506 		    DDI_DMA_SYNC_FORDEV);
3507 
3508 		/* From start of list to the desc_last */
3509 		if (desc_last != 0)
3510 			(void) ddi_dma_sync(fl->dhdl, 0, desc_last *
3511 			    RX_FL_ESIZE, DDI_DMA_SYNC_FORDEV);
3512 	} else {
3513 		/* There was no wrap, sync from start_desc to last_desc */
3514 		desc_start = FL_HW_IDX(fl->pidx - fl->pending);
3515 		(void) ddi_dma_sync(fl->dhdl, desc_start * RX_FL_ESIZE,
3516 		    ndesc * RX_FL_ESIZE, DDI_DMA_SYNC_FORDEV);
3517 	}
3518 
3519 	if (is_t4(sc->params.chip))
3520 		v |= V_PIDX(ndesc);
3521 	else
3522 		v |= V_PIDX_T5(ndesc);
3523 	v |= V_QID(fl->cntxt_id) | V_PIDX(ndesc);
3524 
3525 	membar_producer();
3526 
3527 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL), v);
3528 
3529 	/*
3530 	 * Update pending count:
3531 	 * Deduct the number of descriptors posted
3532 	 */
3533 	fl->pending -= ndesc * 8;
3534 }
3535 
3536 static void
3537 tx_reclaim_task(void *arg)
3538 {
3539 	struct sge_txq *txq = arg;
3540 
3541 	TXQ_LOCK(txq);
3542 	reclaim_tx_descs(txq, txq->eq.qsize);
3543 	TXQ_UNLOCK(txq);
3544 }
3545 
3546 /* ARGSUSED */
3547 static int
3548 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
3549                 mblk_t *m)
3550 {
3551 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
3552 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
3553 	struct adapter *sc = iq->adapter;
3554 	struct sge *s = &sc->sge;
3555 	struct sge_eq *eq;
3556 	struct sge_txq *txq;
3557 
3558 	txq = (void *)s->eqmap[qid - s->eq_start];
3559 	eq = &txq->eq;
3560 	txq->qflush++;
3561 	t4_mac_tx_update(txq->port, txq);
3562 
3563 	ddi_taskq_dispatch(sc->tq[eq->tx_chan], tx_reclaim_task,
3564 		(void *)txq, DDI_NOSLEEP);
3565 
3566 	return (0);
3567 }
3568 
3569 static int
3570 handle_fw_rpl(struct sge_iq *iq, const struct rss_header *rss, mblk_t *m)
3571 {
3572 	struct adapter *sc = iq->adapter;
3573 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
3574 
3575 	ASSERT(m == NULL);
3576 
3577 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
3578 		const struct rss_header *rss2;
3579 
3580 		rss2 = (const struct rss_header *)&cpl->data[0];
3581 		return (sc->cpl_handler[rss2->opcode](iq, rss2, m));
3582 	}
3583 	return (sc->fw_msg_handler[cpl->type](sc, &cpl->data[0]));
3584 }
3585 
3586 int
3587 t4_alloc_tx_maps(struct adapter *sc, struct tx_maps *txmaps, int count,
3588     int flags)
3589 {
3590 	int i, rc;
3591 
3592 	txmaps->map_total =  count;
3593 	txmaps->map_avail = txmaps->map_cidx = txmaps->map_pidx = 0;
3594 
3595 	txmaps->map =  kmem_zalloc(sizeof (ddi_dma_handle_t) *
3596 	    txmaps->map_total, flags);
3597 
3598 	for (i = 0; i < count; i++) {
3599 		rc = ddi_dma_alloc_handle(sc->dip, &sc->sge.dma_attr_tx,
3600 		    DDI_DMA_SLEEP, 0, &txmaps->map[i]);
3601 		if (rc != DDI_SUCCESS) {
3602 			cxgb_printf(sc->dip, CE_WARN,
3603 			    "%s: failed to allocate DMA handle (%d)",
3604 			    __func__, rc);
3605 			return (rc == DDI_DMA_NORESOURCES ? ENOMEM : EINVAL);
3606 		}
3607 		txmaps->map_avail++;
3608 	}
3609 
3610 	return (0);
3611 }
3612 
3613 #define	KS_UINIT(x)	kstat_named_init(&kstatp->x, #x, KSTAT_DATA_ULONG)
3614 #define	KS_CINIT(x)	kstat_named_init(&kstatp->x, #x, KSTAT_DATA_CHAR)
3615 #define	KS_U_SET(x, y)	kstatp->x.value.ul = (y)
3616 #define	KS_U_FROM(x, y)	kstatp->x.value.ul = (y)->x
3617 #define	KS_C_SET(x, ...)	\
3618 			(void) snprintf(kstatp->x.value.c, 16,  __VA_ARGS__)
3619 
3620 /*
3621  * cxgbe:X:config
3622  */
3623 struct cxgbe_port_config_kstats {
3624 	kstat_named_t idx;
3625 	kstat_named_t nrxq;
3626 	kstat_named_t ntxq;
3627 	kstat_named_t first_rxq;
3628 	kstat_named_t first_txq;
3629 	kstat_named_t controller;
3630 	kstat_named_t factory_mac_address;
3631 };
3632 
3633 /*
3634  * cxgbe:X:info
3635  */
3636 struct cxgbe_port_info_kstats {
3637 	kstat_named_t transceiver;
3638 	kstat_named_t rx_ovflow0;
3639 	kstat_named_t rx_ovflow1;
3640 	kstat_named_t rx_ovflow2;
3641 	kstat_named_t rx_ovflow3;
3642 	kstat_named_t rx_trunc0;
3643 	kstat_named_t rx_trunc1;
3644 	kstat_named_t rx_trunc2;
3645 	kstat_named_t rx_trunc3;
3646 	kstat_named_t tx_pause;
3647 	kstat_named_t rx_pause;
3648 };
3649 
3650 static kstat_t *
3651 setup_port_config_kstats(struct port_info *pi)
3652 {
3653 	kstat_t *ksp;
3654 	struct cxgbe_port_config_kstats *kstatp;
3655 	int ndata;
3656 	dev_info_t *pdip = ddi_get_parent(pi->dip);
3657 	uint8_t *ma = &pi->hw_addr[0];
3658 
3659 	ndata = sizeof (struct cxgbe_port_config_kstats) /
3660 	    sizeof (kstat_named_t);
3661 
3662 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), "config",
3663 	    "net", KSTAT_TYPE_NAMED, ndata, 0);
3664 	if (ksp == NULL) {
3665 		cxgb_printf(pi->dip, CE_WARN, "failed to initialize kstats.");
3666 		return (NULL);
3667 	}
3668 
3669 	kstatp = (struct cxgbe_port_config_kstats *)ksp->ks_data;
3670 
3671 	KS_UINIT(idx);
3672 	KS_UINIT(nrxq);
3673 	KS_UINIT(ntxq);
3674 	KS_UINIT(first_rxq);
3675 	KS_UINIT(first_txq);
3676 	KS_CINIT(controller);
3677 	KS_CINIT(factory_mac_address);
3678 
3679 	KS_U_SET(idx, pi->port_id);
3680 	KS_U_SET(nrxq, pi->nrxq);
3681 	KS_U_SET(ntxq, pi->ntxq);
3682 	KS_U_SET(first_rxq, pi->first_rxq);
3683 	KS_U_SET(first_txq, pi->first_txq);
3684 	KS_C_SET(controller, "%s%d", ddi_driver_name(pdip),
3685 	    ddi_get_instance(pdip));
3686 	KS_C_SET(factory_mac_address, "%02X%02X%02X%02X%02X%02X",
3687 	    ma[0], ma[1], ma[2], ma[3], ma[4], ma[5]);
3688 
3689 	/* Do NOT set ksp->ks_update.  These kstats do not change. */
3690 
3691 	/* Install the kstat */
3692 	ksp->ks_private = (void *)pi;
3693 	kstat_install(ksp);
3694 
3695 	return (ksp);
3696 }
3697 
3698 static kstat_t *
3699 setup_port_info_kstats(struct port_info *pi)
3700 {
3701 	kstat_t *ksp;
3702 	struct cxgbe_port_info_kstats *kstatp;
3703 	int ndata;
3704 
3705 	ndata = sizeof (struct cxgbe_port_info_kstats) / sizeof (kstat_named_t);
3706 
3707 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), "info",
3708 	    "net", KSTAT_TYPE_NAMED, ndata, 0);
3709 	if (ksp == NULL) {
3710 		cxgb_printf(pi->dip, CE_WARN, "failed to initialize kstats.");
3711 		return (NULL);
3712 	}
3713 
3714 	kstatp = (struct cxgbe_port_info_kstats *)ksp->ks_data;
3715 
3716 	KS_CINIT(transceiver);
3717 	KS_UINIT(rx_ovflow0);
3718 	KS_UINIT(rx_ovflow1);
3719 	KS_UINIT(rx_ovflow2);
3720 	KS_UINIT(rx_ovflow3);
3721 	KS_UINIT(rx_trunc0);
3722 	KS_UINIT(rx_trunc1);
3723 	KS_UINIT(rx_trunc2);
3724 	KS_UINIT(rx_trunc3);
3725 	KS_UINIT(tx_pause);
3726 	KS_UINIT(rx_pause);
3727 
3728 	/* Install the kstat */
3729 	ksp->ks_update = update_port_info_kstats;
3730 	ksp->ks_private = (void *)pi;
3731 	kstat_install(ksp);
3732 
3733 	return (ksp);
3734 }
3735 
3736 static int
3737 update_port_info_kstats(kstat_t *ksp, int rw)
3738 {
3739 	struct cxgbe_port_info_kstats *kstatp =
3740 	    (struct cxgbe_port_info_kstats *)ksp->ks_data;
3741 	struct port_info *pi = ksp->ks_private;
3742 	static const char *mod_str[] = { NULL, "LR", "SR", "ER", "TWINAX",
3743 	    "active TWINAX", "LRM" };
3744 	uint32_t bgmap;
3745 
3746 	if (rw == KSTAT_WRITE)
3747 		return (0);
3748 
3749 	if (pi->mod_type == FW_PORT_MOD_TYPE_NONE)
3750 		KS_C_SET(transceiver, "unplugged");
3751 	else if (pi->mod_type == FW_PORT_MOD_TYPE_UNKNOWN)
3752 		KS_C_SET(transceiver, "unknown");
3753 	else if (pi->mod_type == FW_PORT_MOD_TYPE_NOTSUPPORTED)
3754 		KS_C_SET(transceiver, "unsupported");
3755 	else if (pi->mod_type > 0 && pi->mod_type < ARRAY_SIZE(mod_str))
3756 		KS_C_SET(transceiver, "%s", mod_str[pi->mod_type]);
3757 	else
3758 		KS_C_SET(transceiver, "type %d", pi->mod_type);
3759 
3760 #define	GET_STAT(name) t4_read_reg64(pi->adapter, \
3761 	    PORT_REG(pi->port_id, A_MPS_PORT_STAT_##name##_L))
3762 #define	GET_STAT_COM(name) t4_read_reg64(pi->adapter, \
3763 	    A_MPS_STAT_##name##_L)
3764 
3765 	bgmap = G_NUMPORTS(t4_read_reg(pi->adapter, A_MPS_CMN_CTL));
3766 	if (bgmap == 0)
3767 		bgmap = (pi->port_id == 0) ? 0xf : 0;
3768 	else if (bgmap == 1)
3769 		bgmap = (pi->port_id < 2) ? (3 << (2 * pi->port_id)) : 0;
3770 	else
3771 		bgmap = 1;
3772 
3773 	KS_U_SET(rx_ovflow0, (bgmap & 1) ?
3774 	    GET_STAT_COM(RX_BG_0_MAC_DROP_FRAME) : 0);
3775 	KS_U_SET(rx_ovflow1, (bgmap & 2) ?
3776 	    GET_STAT_COM(RX_BG_1_MAC_DROP_FRAME) : 0);
3777 	KS_U_SET(rx_ovflow2, (bgmap & 4) ?
3778 	    GET_STAT_COM(RX_BG_2_MAC_DROP_FRAME) : 0);
3779 	KS_U_SET(rx_ovflow3, (bgmap & 8) ?
3780 	    GET_STAT_COM(RX_BG_3_MAC_DROP_FRAME) : 0);
3781 	KS_U_SET(rx_trunc0,  (bgmap & 1) ?
3782 	    GET_STAT_COM(RX_BG_0_MAC_TRUNC_FRAME) : 0);
3783 	KS_U_SET(rx_trunc1,  (bgmap & 2) ?
3784 	    GET_STAT_COM(RX_BG_1_MAC_TRUNC_FRAME) : 0);
3785 	KS_U_SET(rx_trunc2,  (bgmap & 4) ?
3786 	    GET_STAT_COM(RX_BG_2_MAC_TRUNC_FRAME) : 0);
3787 	KS_U_SET(rx_trunc3,  (bgmap & 8) ?
3788 	    GET_STAT_COM(RX_BG_3_MAC_TRUNC_FRAME) : 0);
3789 
3790 	KS_U_SET(tx_pause, GET_STAT(TX_PORT_PAUSE));
3791 	KS_U_SET(rx_pause, GET_STAT(RX_PORT_PAUSE));
3792 
3793 	return (0);
3794 
3795 }
3796 
3797 /*
3798  * cxgbe:X:rxqY
3799  */
3800 struct rxq_kstats {
3801 	kstat_named_t rxcsum;
3802 	kstat_named_t rxpkts;
3803 	kstat_named_t rxbytes;
3804 	kstat_named_t nomem;
3805 };
3806 
3807 static kstat_t *
3808 setup_rxq_kstats(struct port_info *pi, struct sge_rxq *rxq, int idx)
3809 {
3810 	struct kstat *ksp;
3811 	struct rxq_kstats *kstatp;
3812 	int ndata;
3813 	char str[16];
3814 
3815 	ndata = sizeof (struct rxq_kstats) / sizeof (kstat_named_t);
3816 	(void) snprintf(str, sizeof (str), "rxq%u", idx);
3817 
3818 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), str, "rxq",
3819 	    KSTAT_TYPE_NAMED, ndata, 0);
3820 	if (ksp == NULL) {
3821 		cxgb_printf(pi->dip, CE_WARN,
3822 		    "%s: failed to initialize rxq kstats for queue %d.",
3823 		    __func__, idx);
3824 		return (NULL);
3825 	}
3826 
3827 	kstatp = (struct rxq_kstats *)ksp->ks_data;
3828 
3829 	KS_UINIT(rxcsum);
3830 	KS_UINIT(rxpkts);
3831 	KS_UINIT(rxbytes);
3832 	KS_UINIT(nomem);
3833 
3834 	ksp->ks_update = update_rxq_kstats;
3835 	ksp->ks_private = (void *)rxq;
3836 	kstat_install(ksp);
3837 
3838 	return (ksp);
3839 }
3840 
3841 static int
3842 update_rxq_kstats(kstat_t *ksp, int rw)
3843 {
3844 	struct rxq_kstats *kstatp = (struct rxq_kstats *)ksp->ks_data;
3845 	struct sge_rxq *rxq = ksp->ks_private;
3846 
3847 	if (rw == KSTAT_WRITE)
3848 		return (0);
3849 
3850 	KS_U_FROM(rxcsum, rxq);
3851 	KS_U_FROM(rxpkts, rxq);
3852 	KS_U_FROM(rxbytes, rxq);
3853 	KS_U_FROM(nomem, rxq);
3854 
3855 	return (0);
3856 }
3857 
3858 /*
3859  * cxgbe:X:txqY
3860  */
3861 struct txq_kstats {
3862 	kstat_named_t txcsum;
3863 	kstat_named_t tso_wrs;
3864 	kstat_named_t imm_wrs;
3865 	kstat_named_t sgl_wrs;
3866 	kstat_named_t txpkt_wrs;
3867 	kstat_named_t txpkts_wrs;
3868 	kstat_named_t txpkts_pkts;
3869 	kstat_named_t txb_used;
3870 	kstat_named_t hdl_used;
3871 	kstat_named_t txb_full;
3872 	kstat_named_t dma_hdl_failed;
3873 	kstat_named_t dma_map_failed;
3874 	kstat_named_t qfull;
3875 	kstat_named_t qflush;
3876 	kstat_named_t pullup_early;
3877 	kstat_named_t pullup_late;
3878 	kstat_named_t pullup_failed;
3879 	kstat_named_t csum_failed;
3880 };
3881 
3882 static kstat_t *
3883 setup_txq_kstats(struct port_info *pi, struct sge_txq *txq, int idx)
3884 {
3885 	struct kstat *ksp;
3886 	struct txq_kstats *kstatp;
3887 	int ndata;
3888 	char str[16];
3889 
3890 	ndata = sizeof (struct txq_kstats) / sizeof (kstat_named_t);
3891 	(void) snprintf(str, sizeof (str), "txq%u", idx);
3892 
3893 	ksp = kstat_create(T4_PORT_NAME, ddi_get_instance(pi->dip), str, "txq",
3894 	    KSTAT_TYPE_NAMED, ndata, 0);
3895 	if (ksp == NULL) {
3896 		cxgb_printf(pi->dip, CE_WARN,
3897 		    "%s: failed to initialize txq kstats for queue %d.",
3898 		    __func__, idx);
3899 		return (NULL);
3900 	}
3901 
3902 	kstatp = (struct txq_kstats *)ksp->ks_data;
3903 
3904 	KS_UINIT(txcsum);
3905 	KS_UINIT(tso_wrs);
3906 	KS_UINIT(imm_wrs);
3907 	KS_UINIT(sgl_wrs);
3908 	KS_UINIT(txpkt_wrs);
3909 	KS_UINIT(txpkts_wrs);
3910 	KS_UINIT(txpkts_pkts);
3911 	KS_UINIT(txb_used);
3912 	KS_UINIT(hdl_used);
3913 	KS_UINIT(txb_full);
3914 	KS_UINIT(dma_hdl_failed);
3915 	KS_UINIT(dma_map_failed);
3916 	KS_UINIT(qfull);
3917 	KS_UINIT(qflush);
3918 	KS_UINIT(pullup_early);
3919 	KS_UINIT(pullup_late);
3920 	KS_UINIT(pullup_failed);
3921 	KS_UINIT(csum_failed);
3922 
3923 	ksp->ks_update = update_txq_kstats;
3924 	ksp->ks_private = (void *)txq;
3925 	kstat_install(ksp);
3926 
3927 	return (ksp);
3928 }
3929 
3930 static int
3931 update_txq_kstats(kstat_t *ksp, int rw)
3932 {
3933 	struct txq_kstats *kstatp = (struct txq_kstats *)ksp->ks_data;
3934 	struct sge_txq *txq = ksp->ks_private;
3935 
3936 	if (rw == KSTAT_WRITE)
3937 		return (0);
3938 
3939 	KS_U_FROM(txcsum, txq);
3940 	KS_U_FROM(tso_wrs, txq);
3941 	KS_U_FROM(imm_wrs, txq);
3942 	KS_U_FROM(sgl_wrs, txq);
3943 	KS_U_FROM(txpkt_wrs, txq);
3944 	KS_U_FROM(txpkts_wrs, txq);
3945 	KS_U_FROM(txpkts_pkts, txq);
3946 	KS_U_FROM(txb_used, txq);
3947 	KS_U_FROM(hdl_used, txq);
3948 	KS_U_FROM(txb_full, txq);
3949 	KS_U_FROM(dma_hdl_failed, txq);
3950 	KS_U_FROM(dma_map_failed, txq);
3951 	KS_U_FROM(qfull, txq);
3952 	KS_U_FROM(qflush, txq);
3953 	KS_U_FROM(pullup_early, txq);
3954 	KS_U_FROM(pullup_late, txq);
3955 	KS_U_FROM(pullup_failed, txq);
3956 	KS_U_FROM(csum_failed, txq);
3957 
3958 	return (0);
3959 }
3960