xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision 41840d7587afd6ce27e3725b80481dd4d8f26b1a)
1 /*-
2  * Copyright (c) 2011 Chelsio Communications, Inc.
3  * All rights reserved.
4  * Written by: Navdeep Parhar <np@FreeBSD.org>
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 
34 #include <sys/types.h>
35 #include <sys/mbuf.h>
36 #include <sys/socket.h>
37 #include <sys/kernel.h>
38 #include <sys/kdb.h>
39 #include <sys/malloc.h>
40 #include <sys/queue.h>
41 #include <sys/taskqueue.h>
42 #include <sys/sysctl.h>
43 #include <sys/smp.h>
44 #include <net/bpf.h>
45 #include <net/ethernet.h>
46 #include <net/if.h>
47 #include <net/if_vlan_var.h>
48 #include <netinet/in.h>
49 #include <netinet/ip.h>
50 #include <netinet/ip6.h>
51 #include <netinet/tcp.h>
52 
53 #include "common/common.h"
54 #include "common/t4_regs.h"
55 #include "common/t4_regs_values.h"
56 #include "common/t4_msg.h"
57 
58 struct fl_buf_info {
59 	int size;
60 	int type;
61 	uma_zone_t zone;
62 };
63 
64 /* Filled up by t4_sge_modload */
65 static struct fl_buf_info fl_buf_info[FL_BUF_SIZES];
66 
67 #define FL_BUF_SIZE(x)	(fl_buf_info[x].size)
68 #define FL_BUF_TYPE(x)	(fl_buf_info[x].type)
69 #define FL_BUF_ZONE(x)	(fl_buf_info[x].zone)
70 
71 #ifdef T4_PKT_TIMESTAMP
72 #define RX_COPY_THRESHOLD (MINCLSIZE - 8)
73 #else
74 #define RX_COPY_THRESHOLD MINCLSIZE
75 #endif
76 
77 /*
78  * Ethernet frames are DMA'd at this byte offset into the freelist buffer.
79  * 0-7 are valid values.
80  */
81 static int fl_pktshift = 2;
82 TUNABLE_INT("hw.cxgbe.fl_pktshift", &fl_pktshift);
83 
84 /*
85  * Pad ethernet payload up to this boundary.
86  * -1: driver should figure out a good value.
87  *  Any power of 2, from 32 to 4096 (both inclusive) is a valid value.
88  */
89 static int fl_pad = -1;
90 TUNABLE_INT("hw.cxgbe.fl_pad", &fl_pad);
91 
92 /*
93  * Status page length.
94  * -1: driver should figure out a good value.
95  *  64 or 128 are the only other valid values.
96  */
97 static int spg_len = -1;
98 TUNABLE_INT("hw.cxgbe.spg_len", &spg_len);
99 
100 /*
101  * Congestion drops.
102  * -1: no congestion feedback (not recommended).
103  *  0: backpressure the channel instead of dropping packets right away.
104  *  1: no backpressure, drop packets for the congested queue immediately.
105  */
106 static int cong_drop = 0;
107 TUNABLE_INT("hw.cxgbe.cong_drop", &cong_drop);
108 
109 /* Used to track coalesced tx work request */
110 struct txpkts {
111 	uint64_t *flitp;	/* ptr to flit where next pkt should start */
112 	uint8_t npkt;		/* # of packets in this work request */
113 	uint8_t nflits;		/* # of flits used by this work request */
114 	uint16_t plen;		/* total payload (sum of all packets) */
115 };
116 
117 /* A packet's SGL.  This + m_pkthdr has all info needed for tx */
118 struct sgl {
119 	int nsegs;		/* # of segments in the SGL, 0 means imm. tx */
120 	int nflits;		/* # of flits needed for the SGL */
121 	bus_dma_segment_t seg[TX_SGL_SEGS];
122 };
123 
124 static int service_iq(struct sge_iq *, int);
125 static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t,
126     int *);
127 static int t4_eth_rx(struct sge_iq *, const struct rss_header *, struct mbuf *);
128 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int,
129     int);
130 static inline void init_fl(struct sge_fl *, int, int, char *);
131 static inline void init_eq(struct sge_eq *, int, int, uint8_t, uint16_t,
132     char *);
133 static int alloc_ring(struct adapter *, size_t, bus_dma_tag_t *, bus_dmamap_t *,
134     bus_addr_t *, void **);
135 static int free_ring(struct adapter *, bus_dma_tag_t, bus_dmamap_t, bus_addr_t,
136     void *);
137 static int alloc_iq_fl(struct port_info *, struct sge_iq *, struct sge_fl *,
138     int, int);
139 static int free_iq_fl(struct port_info *, struct sge_iq *, struct sge_fl *);
140 static int alloc_fwq(struct adapter *);
141 static int free_fwq(struct adapter *);
142 static int alloc_mgmtq(struct adapter *);
143 static int free_mgmtq(struct adapter *);
144 static int alloc_rxq(struct port_info *, struct sge_rxq *, int, int,
145     struct sysctl_oid *);
146 static int free_rxq(struct port_info *, struct sge_rxq *);
147 #ifdef TCP_OFFLOAD
148 static int alloc_ofld_rxq(struct port_info *, struct sge_ofld_rxq *, int, int,
149     struct sysctl_oid *);
150 static int free_ofld_rxq(struct port_info *, struct sge_ofld_rxq *);
151 #endif
152 static int ctrl_eq_alloc(struct adapter *, struct sge_eq *);
153 static int eth_eq_alloc(struct adapter *, struct port_info *, struct sge_eq *);
154 #ifdef TCP_OFFLOAD
155 static int ofld_eq_alloc(struct adapter *, struct port_info *, struct sge_eq *);
156 #endif
157 static int alloc_eq(struct adapter *, struct port_info *, struct sge_eq *);
158 static int free_eq(struct adapter *, struct sge_eq *);
159 static int alloc_wrq(struct adapter *, struct port_info *, struct sge_wrq *,
160     struct sysctl_oid *);
161 static int free_wrq(struct adapter *, struct sge_wrq *);
162 static int alloc_txq(struct port_info *, struct sge_txq *, int,
163     struct sysctl_oid *);
164 static int free_txq(struct port_info *, struct sge_txq *);
165 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int);
166 static inline bool is_new_response(const struct sge_iq *, struct rsp_ctrl **);
167 static inline void iq_next(struct sge_iq *);
168 static inline void ring_fl_db(struct adapter *, struct sge_fl *);
169 static int refill_fl(struct adapter *, struct sge_fl *, int);
170 static void refill_sfl(void *);
171 static int alloc_fl_sdesc(struct sge_fl *);
172 static void free_fl_sdesc(struct sge_fl *);
173 static void set_fl_tag_idx(struct sge_fl *, int);
174 static void add_fl_to_sfl(struct adapter *, struct sge_fl *);
175 
176 static int get_pkt_sgl(struct sge_txq *, struct mbuf **, struct sgl *, int);
177 static int free_pkt_sgl(struct sge_txq *, struct sgl *);
178 static int write_txpkt_wr(struct port_info *, struct sge_txq *, struct mbuf *,
179     struct sgl *);
180 static int add_to_txpkts(struct port_info *, struct sge_txq *, struct txpkts *,
181     struct mbuf *, struct sgl *);
182 static void write_txpkts_wr(struct sge_txq *, struct txpkts *);
183 static inline void write_ulp_cpl_sgl(struct port_info *, struct sge_txq *,
184     struct txpkts *, struct mbuf *, struct sgl *);
185 static int write_sgl_to_txd(struct sge_eq *, struct sgl *, caddr_t *);
186 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int);
187 static inline void ring_eq_db(struct adapter *, struct sge_eq *);
188 static inline int reclaimable(struct sge_eq *);
189 static int reclaim_tx_descs(struct sge_txq *, int, int);
190 static void write_eqflush_wr(struct sge_eq *);
191 static __be64 get_flit(bus_dma_segment_t *, int, int);
192 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
193     struct mbuf *);
194 static int handle_fw_msg(struct sge_iq *, const struct rss_header *,
195     struct mbuf *);
196 
197 static int sysctl_uint16(SYSCTL_HANDLER_ARGS);
198 
199 #if defined(__i386__) || defined(__amd64__)
200 extern u_int cpu_clflush_line_size;
201 #endif
202 
203 /*
204  * Called on MOD_LOAD.  Fills up fl_buf_info[] and validates/calculates the SGE
205  * tunables.
206  */
207 void
208 t4_sge_modload(void)
209 {
210 	int i;
211 	int bufsize[FL_BUF_SIZES] = {
212 		MCLBYTES,
213 #if MJUMPAGESIZE != MCLBYTES
214 		MJUMPAGESIZE,
215 #endif
216 		MJUM9BYTES,
217 		MJUM16BYTES
218 	};
219 
220 	for (i = 0; i < FL_BUF_SIZES; i++) {
221 		FL_BUF_SIZE(i) = bufsize[i];
222 		FL_BUF_TYPE(i) = m_gettype(bufsize[i]);
223 		FL_BUF_ZONE(i) = m_getzone(bufsize[i]);
224 	}
225 
226 	if (fl_pktshift < 0 || fl_pktshift > 7) {
227 		printf("Invalid hw.cxgbe.fl_pktshift value (%d),"
228 		    " using 2 instead.\n", fl_pktshift);
229 		fl_pktshift = 2;
230 	}
231 
232 	if (fl_pad < 32 || fl_pad > 4096 || !powerof2(fl_pad)) {
233 		int pad;
234 
235 #if defined(__i386__) || defined(__amd64__)
236 		pad = max(cpu_clflush_line_size, 32);
237 #else
238 		pad = max(CACHE_LINE_SIZE, 32);
239 #endif
240 		pad = min(pad, 4096);
241 
242 		if (fl_pad != -1) {
243 			printf("Invalid hw.cxgbe.fl_pad value (%d),"
244 			    " using %d instead.\n", fl_pad, pad);
245 		}
246 		fl_pad = pad;
247 	}
248 
249 	if (spg_len != 64 && spg_len != 128) {
250 		int len;
251 
252 #if defined(__i386__) || defined(__amd64__)
253 		len = cpu_clflush_line_size > 64 ? 128 : 64;
254 #else
255 		len = 64;
256 #endif
257 		if (spg_len != -1) {
258 			printf("Invalid hw.cxgbe.spg_len value (%d),"
259 			    " using %d instead.\n", spg_len, len);
260 		}
261 		spg_len = len;
262 	}
263 
264 	if (cong_drop < -1 || cong_drop > 1) {
265 		printf("Invalid hw.cxgbe.cong_drop value (%d),"
266 		    " using 0 instead.\n", cong_drop);
267 		cong_drop = 0;
268 	}
269 }
270 
271 void
272 t4_init_sge_cpl_handlers(struct adapter *sc)
273 {
274 
275 	t4_register_cpl_handler(sc, CPL_FW4_MSG, handle_fw_msg);
276 	t4_register_cpl_handler(sc, CPL_FW6_MSG, handle_fw_msg);
277 	t4_register_cpl_handler(sc, CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
278 	t4_register_cpl_handler(sc, CPL_RX_PKT, t4_eth_rx);
279 	t4_register_fw_msg_handler(sc, FW6_TYPE_CMD_RPL, t4_handle_fw_rpl);
280 }
281 
282 /*
283  * adap->params.vpd.cclk must be set up before this is called.
284  */
285 void
286 t4_tweak_chip_settings(struct adapter *sc)
287 {
288 	int i;
289 	uint32_t v, m;
290 	int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200};
291 	int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk;
292 	int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */
293 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
294 
295 	KASSERT(sc->flags & MASTER_PF,
296 	    ("%s: trying to change chip settings when not master.", __func__));
297 
298 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE |
299 	    V_INGPADBOUNDARY(M_INGPADBOUNDARY) | F_EGRSTATUSPAGESIZE;
300 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
301 	    V_INGPADBOUNDARY(ilog2(fl_pad) - 5) |
302 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
303 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
304 
305 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
306 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
307 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
308 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
309 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
310 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
311 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
312 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
313 	t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v);
314 
315 	for (i = 0; i < FL_BUF_SIZES; i++) {
316 		t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i),
317 		    FL_BUF_SIZE(i));
318 	}
319 
320 	v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) |
321 	    V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]);
322 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v);
323 
324 	KASSERT(intr_timer[0] <= timer_max,
325 	    ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0],
326 	    timer_max));
327 	for (i = 1; i < nitems(intr_timer); i++) {
328 		KASSERT(intr_timer[i] >= intr_timer[i - 1],
329 		    ("%s: timers not listed in increasing order (%d)",
330 		    __func__, i));
331 
332 		while (intr_timer[i] > timer_max) {
333 			if (i == nitems(intr_timer) - 1) {
334 				intr_timer[i] = timer_max;
335 				break;
336 			}
337 			intr_timer[i] += intr_timer[i - 1];
338 			intr_timer[i] /= 2;
339 		}
340 	}
341 
342 	v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) |
343 	    V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1]));
344 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v);
345 	v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) |
346 	    V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3]));
347 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v);
348 	v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) |
349 	    V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5]));
350 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v);
351 
352 	if (cong_drop == 0) {
353 		m = F_TUNNELCNGDROP0 | F_TUNNELCNGDROP1 | F_TUNNELCNGDROP2 |
354 		    F_TUNNELCNGDROP3;
355 		t4_set_reg_field(sc, A_TP_PARA_REG3, m, 0);
356 	}
357 
358 	/* 4K, 16K, 64K, 256K DDP "page sizes" */
359 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
360 	t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v);
361 
362 	m = v = F_TDDPTAGTCB;
363 	t4_set_reg_field(sc, A_ULP_RX_CTL, m, v);
364 
365 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
366 	    F_RESETDDPOFFSET;
367 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
368 	t4_set_reg_field(sc, A_TP_PARA_REG5, m, v);
369 }
370 
371 /*
372  * XXX: driver really should be able to deal with unexpected settings.
373  */
374 int
375 t4_read_chip_settings(struct adapter *sc)
376 {
377 	struct sge *s = &sc->sge;
378 	int i, rc = 0;
379 	uint32_t m, v, r;
380 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
381 
382 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE |
383 	    V_INGPADBOUNDARY(M_INGPADBOUNDARY) | F_EGRSTATUSPAGESIZE;
384 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
385 	    V_INGPADBOUNDARY(ilog2(fl_pad) - 5) |
386 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
387 	r = t4_read_reg(sc, A_SGE_CONTROL);
388 	if ((r & m) != v) {
389 		device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r);
390 		rc = EINVAL;
391 	}
392 
393 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
394 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
395 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
396 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
397 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
398 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
399 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
400 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
401 	r = t4_read_reg(sc, A_SGE_HOST_PAGE_SIZE);
402 	if (r != v) {
403 		device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r);
404 		rc = EINVAL;
405 	}
406 
407 	for (i = 0; i < FL_BUF_SIZES; i++) {
408 		v = t4_read_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i));
409 		if (v != FL_BUF_SIZE(i)) {
410 			device_printf(sc->dev,
411 			    "invalid SGE_FL_BUFFER_SIZE[%d](0x%x)\n", i, v);
412 			rc = EINVAL;
413 		}
414 	}
415 
416 	r = t4_read_reg(sc, A_SGE_INGRESS_RX_THRESHOLD);
417 	s->counter_val[0] = G_THRESHOLD_0(r);
418 	s->counter_val[1] = G_THRESHOLD_1(r);
419 	s->counter_val[2] = G_THRESHOLD_2(r);
420 	s->counter_val[3] = G_THRESHOLD_3(r);
421 
422 	r = t4_read_reg(sc, A_SGE_TIMER_VALUE_0_AND_1);
423 	s->timer_val[0] = G_TIMERVALUE0(r) / core_ticks_per_usec(sc);
424 	s->timer_val[1] = G_TIMERVALUE1(r) / core_ticks_per_usec(sc);
425 	r = t4_read_reg(sc, A_SGE_TIMER_VALUE_2_AND_3);
426 	s->timer_val[2] = G_TIMERVALUE2(r) / core_ticks_per_usec(sc);
427 	s->timer_val[3] = G_TIMERVALUE3(r) / core_ticks_per_usec(sc);
428 	r = t4_read_reg(sc, A_SGE_TIMER_VALUE_4_AND_5);
429 	s->timer_val[4] = G_TIMERVALUE4(r) / core_ticks_per_usec(sc);
430 	s->timer_val[5] = G_TIMERVALUE5(r) / core_ticks_per_usec(sc);
431 
432 	if (cong_drop == 0) {
433 		m = F_TUNNELCNGDROP0 | F_TUNNELCNGDROP1 | F_TUNNELCNGDROP2 |
434 		    F_TUNNELCNGDROP3;
435 		r = t4_read_reg(sc, A_TP_PARA_REG3);
436 		if (r & m) {
437 			device_printf(sc->dev,
438 			    "invalid TP_PARA_REG3(0x%x)\n", r);
439 			rc = EINVAL;
440 		}
441 	}
442 
443 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
444 	r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ);
445 	if (r != v) {
446 		device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r);
447 		rc = EINVAL;
448 	}
449 
450 	m = v = F_TDDPTAGTCB;
451 	r = t4_read_reg(sc, A_ULP_RX_CTL);
452 	if ((r & m) != v) {
453 		device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r);
454 		rc = EINVAL;
455 	}
456 
457 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
458 	    F_RESETDDPOFFSET;
459 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
460 	r = t4_read_reg(sc, A_TP_PARA_REG5);
461 	if ((r & m) != v) {
462 		device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r);
463 		rc = EINVAL;
464 	}
465 
466 	r = t4_read_reg(sc, A_SGE_CONM_CTRL);
467 	s->fl_starve_threshold = G_EGRTHRESHOLD(r) * 2 + 1;
468 
469 	if (is_t5(sc)) {
470 		r = t4_read_reg(sc, A_SGE_EGRESS_QUEUES_PER_PAGE_PF);
471 		r >>= S_QUEUESPERPAGEPF0 +
472 		    (S_QUEUESPERPAGEPF1 - S_QUEUESPERPAGEPF0) * sc->pf;
473 		s->s_qpp = r & M_QUEUESPERPAGEPF0;
474 	}
475 
476 	t4_init_tp_params(sc);
477 
478 	t4_read_mtu_tbl(sc, sc->params.mtus, NULL);
479 	t4_load_mtus(sc, sc->params.mtus, sc->params.a_wnd, sc->params.b_wnd);
480 
481 	return (rc);
482 }
483 
484 int
485 t4_create_dma_tag(struct adapter *sc)
486 {
487 	int rc;
488 
489 	rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
490 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
491 	    BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
492 	    NULL, &sc->dmat);
493 	if (rc != 0) {
494 		device_printf(sc->dev,
495 		    "failed to create main DMA tag: %d\n", rc);
496 	}
497 
498 	return (rc);
499 }
500 
501 int
502 t4_destroy_dma_tag(struct adapter *sc)
503 {
504 	if (sc->dmat)
505 		bus_dma_tag_destroy(sc->dmat);
506 
507 	return (0);
508 }
509 
510 /*
511  * Allocate and initialize the firmware event queue and the management queue.
512  *
513  * Returns errno on failure.  Resources allocated up to that point may still be
514  * allocated.  Caller is responsible for cleanup in case this function fails.
515  */
516 int
517 t4_setup_adapter_queues(struct adapter *sc)
518 {
519 	int rc;
520 
521 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
522 
523 	sysctl_ctx_init(&sc->ctx);
524 	sc->flags |= ADAP_SYSCTL_CTX;
525 
526 	/*
527 	 * Firmware event queue
528 	 */
529 	rc = alloc_fwq(sc);
530 	if (rc != 0)
531 		return (rc);
532 
533 	/*
534 	 * Management queue.  This is just a control queue that uses the fwq as
535 	 * its associated iq.
536 	 */
537 	rc = alloc_mgmtq(sc);
538 
539 	return (rc);
540 }
541 
542 /*
543  * Idempotent
544  */
545 int
546 t4_teardown_adapter_queues(struct adapter *sc)
547 {
548 
549 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
550 
551 	/* Do this before freeing the queue */
552 	if (sc->flags & ADAP_SYSCTL_CTX) {
553 		sysctl_ctx_free(&sc->ctx);
554 		sc->flags &= ~ADAP_SYSCTL_CTX;
555 	}
556 
557 	free_mgmtq(sc);
558 	free_fwq(sc);
559 
560 	return (0);
561 }
562 
563 static inline int
564 first_vector(struct port_info *pi)
565 {
566 	struct adapter *sc = pi->adapter;
567 	int rc = T4_EXTRA_INTR, i;
568 
569 	if (sc->intr_count == 1)
570 		return (0);
571 
572 	for_each_port(sc, i) {
573 		struct port_info *p = sc->port[i];
574 
575 		if (i == pi->port_id)
576 			break;
577 
578 #ifdef TCP_OFFLOAD
579 		if (sc->flags & INTR_DIRECT)
580 			rc += p->nrxq + p->nofldrxq;
581 		else
582 			rc += max(p->nrxq, p->nofldrxq);
583 #else
584 		/*
585 		 * Not compiled with offload support and intr_count > 1.  Only
586 		 * NIC queues exist and they'd better be taking direct
587 		 * interrupts.
588 		 */
589 		KASSERT(sc->flags & INTR_DIRECT,
590 		    ("%s: intr_count %d, !INTR_DIRECT", __func__,
591 		    sc->intr_count));
592 
593 		rc += p->nrxq;
594 #endif
595 	}
596 
597 	return (rc);
598 }
599 
600 /*
601  * Given an arbitrary "index," come up with an iq that can be used by other
602  * queues (of this port) for interrupt forwarding, SGE egress updates, etc.
603  * The iq returned is guaranteed to be something that takes direct interrupts.
604  */
605 static struct sge_iq *
606 port_intr_iq(struct port_info *pi, int idx)
607 {
608 	struct adapter *sc = pi->adapter;
609 	struct sge *s = &sc->sge;
610 	struct sge_iq *iq = NULL;
611 
612 	if (sc->intr_count == 1)
613 		return (&sc->sge.fwq);
614 
615 #ifdef TCP_OFFLOAD
616 	if (sc->flags & INTR_DIRECT) {
617 		idx %= pi->nrxq + pi->nofldrxq;
618 
619 		if (idx >= pi->nrxq) {
620 			idx -= pi->nrxq;
621 			iq = &s->ofld_rxq[pi->first_ofld_rxq + idx].iq;
622 		} else
623 			iq = &s->rxq[pi->first_rxq + idx].iq;
624 
625 	} else {
626 		idx %= max(pi->nrxq, pi->nofldrxq);
627 
628 		if (pi->nrxq >= pi->nofldrxq)
629 			iq = &s->rxq[pi->first_rxq + idx].iq;
630 		else
631 			iq = &s->ofld_rxq[pi->first_ofld_rxq + idx].iq;
632 	}
633 #else
634 	/*
635 	 * Not compiled with offload support and intr_count > 1.  Only NIC
636 	 * queues exist and they'd better be taking direct interrupts.
637 	 */
638 	KASSERT(sc->flags & INTR_DIRECT,
639 	    ("%s: intr_count %d, !INTR_DIRECT", __func__, sc->intr_count));
640 
641 	idx %= pi->nrxq;
642 	iq = &s->rxq[pi->first_rxq + idx].iq;
643 #endif
644 
645 	KASSERT(iq->flags & IQ_INTR, ("%s: EDOOFUS", __func__));
646 	return (iq);
647 }
648 
649 static inline int
650 mtu_to_bufsize(int mtu)
651 {
652 	int bufsize;
653 
654 	/* large enough for a frame even when VLAN extraction is disabled */
655 	bufsize = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN + mtu;
656 	bufsize = roundup2(bufsize + fl_pktshift, fl_pad);
657 
658 	return (bufsize);
659 }
660 
661 #ifdef TCP_OFFLOAD
662 static inline int
663 mtu_to_bufsize_toe(struct adapter *sc, int mtu)
664 {
665 
666 	if (sc->tt.rx_coalesce)
667 		return (G_RXCOALESCESIZE(t4_read_reg(sc, A_TP_PARA_REG2)));
668 
669 	return (mtu);
670 }
671 #endif
672 
673 int
674 t4_setup_port_queues(struct port_info *pi)
675 {
676 	int rc = 0, i, j, intr_idx, iqid;
677 	struct sge_rxq *rxq;
678 	struct sge_txq *txq;
679 	struct sge_wrq *ctrlq;
680 #ifdef TCP_OFFLOAD
681 	struct sge_ofld_rxq *ofld_rxq;
682 	struct sge_wrq *ofld_txq;
683 	struct sysctl_oid *oid2 = NULL;
684 #endif
685 	char name[16];
686 	struct adapter *sc = pi->adapter;
687 	struct ifnet *ifp = pi->ifp;
688 	struct sysctl_oid *oid = device_get_sysctl_tree(pi->dev);
689 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
690 	int bufsize;
691 
692 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "rxq", CTLFLAG_RD,
693 	    NULL, "rx queues");
694 
695 #ifdef TCP_OFFLOAD
696 	if (is_offload(sc)) {
697 		oid2 = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "ofld_rxq",
698 		    CTLFLAG_RD, NULL,
699 		    "rx queues for offloaded TCP connections");
700 	}
701 #endif
702 
703 	/* Interrupt vector to start from (when using multiple vectors) */
704 	intr_idx = first_vector(pi);
705 
706 	/*
707 	 * First pass over all rx queues (NIC and TOE):
708 	 * a) initialize iq and fl
709 	 * b) allocate queue iff it will take direct interrupts.
710 	 */
711 	bufsize = mtu_to_bufsize(ifp->if_mtu);
712 	for_each_rxq(pi, i, rxq) {
713 
714 		init_iq(&rxq->iq, sc, pi->tmr_idx, pi->pktc_idx, pi->qsize_rxq,
715 		    RX_IQ_ESIZE);
716 
717 		snprintf(name, sizeof(name), "%s rxq%d-fl",
718 		    device_get_nameunit(pi->dev), i);
719 		init_fl(&rxq->fl, pi->qsize_rxq / 8, bufsize, name);
720 
721 		if (sc->flags & INTR_DIRECT
722 #ifdef TCP_OFFLOAD
723 		    || (sc->intr_count > 1 && pi->nrxq >= pi->nofldrxq)
724 #endif
725 		   ) {
726 			rxq->iq.flags |= IQ_INTR;
727 			rc = alloc_rxq(pi, rxq, intr_idx, i, oid);
728 			if (rc != 0)
729 				goto done;
730 			intr_idx++;
731 		}
732 	}
733 
734 #ifdef TCP_OFFLOAD
735 	bufsize = mtu_to_bufsize_toe(sc, ifp->if_mtu);
736 	for_each_ofld_rxq(pi, i, ofld_rxq) {
737 
738 		init_iq(&ofld_rxq->iq, sc, pi->tmr_idx, pi->pktc_idx,
739 		    pi->qsize_rxq, RX_IQ_ESIZE);
740 
741 		snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
742 		    device_get_nameunit(pi->dev), i);
743 		init_fl(&ofld_rxq->fl, pi->qsize_rxq / 8, bufsize, name);
744 
745 		if (sc->flags & INTR_DIRECT ||
746 		    (sc->intr_count > 1 && pi->nofldrxq > pi->nrxq)) {
747 			ofld_rxq->iq.flags |= IQ_INTR;
748 			rc = alloc_ofld_rxq(pi, ofld_rxq, intr_idx, i, oid2);
749 			if (rc != 0)
750 				goto done;
751 			intr_idx++;
752 		}
753 	}
754 #endif
755 
756 	/*
757 	 * Second pass over all rx queues (NIC and TOE).  The queues forwarding
758 	 * their interrupts are allocated now.
759 	 */
760 	j = 0;
761 	for_each_rxq(pi, i, rxq) {
762 		if (rxq->iq.flags & IQ_INTR)
763 			continue;
764 
765 		intr_idx = port_intr_iq(pi, j)->abs_id;
766 
767 		rc = alloc_rxq(pi, rxq, intr_idx, i, oid);
768 		if (rc != 0)
769 			goto done;
770 		j++;
771 	}
772 
773 #ifdef TCP_OFFLOAD
774 	for_each_ofld_rxq(pi, i, ofld_rxq) {
775 		if (ofld_rxq->iq.flags & IQ_INTR)
776 			continue;
777 
778 		intr_idx = port_intr_iq(pi, j)->abs_id;
779 
780 		rc = alloc_ofld_rxq(pi, ofld_rxq, intr_idx, i, oid2);
781 		if (rc != 0)
782 			goto done;
783 		j++;
784 	}
785 #endif
786 
787 	/*
788 	 * Now the tx queues.  Only one pass needed.
789 	 */
790 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "txq", CTLFLAG_RD,
791 	    NULL, "tx queues");
792 	j = 0;
793 	for_each_txq(pi, i, txq) {
794 		uint16_t iqid;
795 
796 		iqid = port_intr_iq(pi, j)->cntxt_id;
797 
798 		snprintf(name, sizeof(name), "%s txq%d",
799 		    device_get_nameunit(pi->dev), i);
800 		init_eq(&txq->eq, EQ_ETH, pi->qsize_txq, pi->tx_chan, iqid,
801 		    name);
802 
803 		rc = alloc_txq(pi, txq, i, oid);
804 		if (rc != 0)
805 			goto done;
806 		j++;
807 	}
808 
809 #ifdef TCP_OFFLOAD
810 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "ofld_txq",
811 	    CTLFLAG_RD, NULL, "tx queues for offloaded TCP connections");
812 	for_each_ofld_txq(pi, i, ofld_txq) {
813 		uint16_t iqid;
814 
815 		iqid = port_intr_iq(pi, j)->cntxt_id;
816 
817 		snprintf(name, sizeof(name), "%s ofld_txq%d",
818 		    device_get_nameunit(pi->dev), i);
819 		init_eq(&ofld_txq->eq, EQ_OFLD, pi->qsize_txq, pi->tx_chan,
820 		    iqid, name);
821 
822 		snprintf(name, sizeof(name), "%d", i);
823 		oid2 = SYSCTL_ADD_NODE(&pi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
824 		    name, CTLFLAG_RD, NULL, "offload tx queue");
825 
826 		rc = alloc_wrq(sc, pi, ofld_txq, oid2);
827 		if (rc != 0)
828 			goto done;
829 		j++;
830 	}
831 #endif
832 
833 	/*
834 	 * Finally, the control queue.
835 	 */
836 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "ctrlq", CTLFLAG_RD,
837 	    NULL, "ctrl queue");
838 	ctrlq = &sc->sge.ctrlq[pi->port_id];
839 	iqid = port_intr_iq(pi, 0)->cntxt_id;
840 	snprintf(name, sizeof(name), "%s ctrlq", device_get_nameunit(pi->dev));
841 	init_eq(&ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, pi->tx_chan, iqid, name);
842 	rc = alloc_wrq(sc, pi, ctrlq, oid);
843 
844 done:
845 	if (rc)
846 		t4_teardown_port_queues(pi);
847 
848 	return (rc);
849 }
850 
851 /*
852  * Idempotent
853  */
854 int
855 t4_teardown_port_queues(struct port_info *pi)
856 {
857 	int i;
858 	struct adapter *sc = pi->adapter;
859 	struct sge_rxq *rxq;
860 	struct sge_txq *txq;
861 #ifdef TCP_OFFLOAD
862 	struct sge_ofld_rxq *ofld_rxq;
863 	struct sge_wrq *ofld_txq;
864 #endif
865 
866 	/* Do this before freeing the queues */
867 	if (pi->flags & PORT_SYSCTL_CTX) {
868 		sysctl_ctx_free(&pi->ctx);
869 		pi->flags &= ~PORT_SYSCTL_CTX;
870 	}
871 
872 	/*
873 	 * Take down all the tx queues first, as they reference the rx queues
874 	 * (for egress updates, etc.).
875 	 */
876 
877 	free_wrq(sc, &sc->sge.ctrlq[pi->port_id]);
878 
879 	for_each_txq(pi, i, txq) {
880 		free_txq(pi, txq);
881 	}
882 
883 #ifdef TCP_OFFLOAD
884 	for_each_ofld_txq(pi, i, ofld_txq) {
885 		free_wrq(sc, ofld_txq);
886 	}
887 #endif
888 
889 	/*
890 	 * Then take down the rx queues that forward their interrupts, as they
891 	 * reference other rx queues.
892 	 */
893 
894 	for_each_rxq(pi, i, rxq) {
895 		if ((rxq->iq.flags & IQ_INTR) == 0)
896 			free_rxq(pi, rxq);
897 	}
898 
899 #ifdef TCP_OFFLOAD
900 	for_each_ofld_rxq(pi, i, ofld_rxq) {
901 		if ((ofld_rxq->iq.flags & IQ_INTR) == 0)
902 			free_ofld_rxq(pi, ofld_rxq);
903 	}
904 #endif
905 
906 	/*
907 	 * Then take down the rx queues that take direct interrupts.
908 	 */
909 
910 	for_each_rxq(pi, i, rxq) {
911 		if (rxq->iq.flags & IQ_INTR)
912 			free_rxq(pi, rxq);
913 	}
914 
915 #ifdef TCP_OFFLOAD
916 	for_each_ofld_rxq(pi, i, ofld_rxq) {
917 		if (ofld_rxq->iq.flags & IQ_INTR)
918 			free_ofld_rxq(pi, ofld_rxq);
919 	}
920 #endif
921 
922 	return (0);
923 }
924 
925 /*
926  * Deals with errors and the firmware event queue.  All data rx queues forward
927  * their interrupt to the firmware event queue.
928  */
929 void
930 t4_intr_all(void *arg)
931 {
932 	struct adapter *sc = arg;
933 	struct sge_iq *fwq = &sc->sge.fwq;
934 
935 	t4_intr_err(arg);
936 	if (atomic_cmpset_int(&fwq->state, IQS_IDLE, IQS_BUSY)) {
937 		service_iq(fwq, 0);
938 		atomic_cmpset_int(&fwq->state, IQS_BUSY, IQS_IDLE);
939 	}
940 }
941 
942 /* Deals with error interrupts */
943 void
944 t4_intr_err(void *arg)
945 {
946 	struct adapter *sc = arg;
947 
948 	t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
949 	t4_slow_intr_handler(sc);
950 }
951 
952 void
953 t4_intr_evt(void *arg)
954 {
955 	struct sge_iq *iq = arg;
956 
957 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
958 		service_iq(iq, 0);
959 		atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
960 	}
961 }
962 
963 void
964 t4_intr(void *arg)
965 {
966 	struct sge_iq *iq = arg;
967 
968 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
969 		service_iq(iq, 0);
970 		atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
971 	}
972 }
973 
974 /*
975  * Deals with anything and everything on the given ingress queue.
976  */
977 static int
978 service_iq(struct sge_iq *iq, int budget)
979 {
980 	struct sge_iq *q;
981 	struct sge_rxq *rxq = iq_to_rxq(iq);	/* Use iff iq is part of rxq */
982 	struct sge_fl *fl = &rxq->fl;		/* Use iff IQ_HAS_FL */
983 	struct adapter *sc = iq->adapter;
984 	struct rsp_ctrl *ctrl;
985 	const struct rss_header *rss;
986 	int ndescs = 0, limit, fl_bufs_used = 0;
987 	int rsp_type;
988 	uint32_t lq;
989 	struct mbuf *m0;
990 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
991 
992 	limit = budget ? budget : iq->qsize / 8;
993 
994 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
995 
996 	/*
997 	 * We always come back and check the descriptor ring for new indirect
998 	 * interrupts and other responses after running a single handler.
999 	 */
1000 	for (;;) {
1001 		while (is_new_response(iq, &ctrl)) {
1002 
1003 			rmb();
1004 
1005 			m0 = NULL;
1006 			rsp_type = G_RSPD_TYPE(ctrl->u.type_gen);
1007 			lq = be32toh(ctrl->pldbuflen_qid);
1008 			rss = (const void *)iq->cdesc;
1009 
1010 			switch (rsp_type) {
1011 			case X_RSPD_TYPE_FLBUF:
1012 
1013 				KASSERT(iq->flags & IQ_HAS_FL,
1014 				    ("%s: data for an iq (%p) with no freelist",
1015 				    __func__, iq));
1016 
1017 				m0 = get_fl_payload(sc, fl, lq, &fl_bufs_used);
1018 #ifdef T4_PKT_TIMESTAMP
1019 				/*
1020 				 * 60 bit timestamp for the payload is
1021 				 * *(uint64_t *)m0->m_pktdat.  Note that it is
1022 				 * in the leading free-space in the mbuf.  The
1023 				 * kernel can clobber it during a pullup,
1024 				 * m_copymdata, etc.  You need to make sure that
1025 				 * the mbuf reaches you unmolested if you care
1026 				 * about the timestamp.
1027 				 */
1028 				*(uint64_t *)m0->m_pktdat =
1029 				    be64toh(ctrl->u.last_flit) &
1030 				    0xfffffffffffffff;
1031 #endif
1032 
1033 				/* fall through */
1034 
1035 			case X_RSPD_TYPE_CPL:
1036 				KASSERT(rss->opcode < NUM_CPL_CMDS,
1037 				    ("%s: bad opcode %02x.", __func__,
1038 				    rss->opcode));
1039 				sc->cpl_handler[rss->opcode](iq, rss, m0);
1040 				break;
1041 
1042 			case X_RSPD_TYPE_INTR:
1043 
1044 				/*
1045 				 * Interrupts should be forwarded only to queues
1046 				 * that are not forwarding their interrupts.
1047 				 * This means service_iq can recurse but only 1
1048 				 * level deep.
1049 				 */
1050 				KASSERT(budget == 0,
1051 				    ("%s: budget %u, rsp_type %u", __func__,
1052 				    budget, rsp_type));
1053 
1054 				q = sc->sge.iqmap[lq - sc->sge.iq_start];
1055 				if (atomic_cmpset_int(&q->state, IQS_IDLE,
1056 				    IQS_BUSY)) {
1057 					if (service_iq(q, q->qsize / 8) == 0) {
1058 						atomic_cmpset_int(&q->state,
1059 						    IQS_BUSY, IQS_IDLE);
1060 					} else {
1061 						STAILQ_INSERT_TAIL(&iql, q,
1062 						    link);
1063 					}
1064 				}
1065 				break;
1066 
1067 			default:
1068 				sc->an_handler(iq, ctrl);
1069 				break;
1070 			}
1071 
1072 			iq_next(iq);
1073 			if (++ndescs == limit) {
1074 				t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
1075 				    V_CIDXINC(ndescs) |
1076 				    V_INGRESSQID(iq->cntxt_id) |
1077 				    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1078 				ndescs = 0;
1079 
1080 				if (fl_bufs_used > 0) {
1081 					FL_LOCK(fl);
1082 					fl->needed += fl_bufs_used;
1083 					refill_fl(sc, fl, fl->cap / 8);
1084 					FL_UNLOCK(fl);
1085 					fl_bufs_used = 0;
1086 				}
1087 
1088 				if (budget)
1089 					return (EINPROGRESS);
1090 			}
1091 		}
1092 
1093 		if (STAILQ_EMPTY(&iql))
1094 			break;
1095 
1096 		/*
1097 		 * Process the head only, and send it to the back of the list if
1098 		 * it's still not done.
1099 		 */
1100 		q = STAILQ_FIRST(&iql);
1101 		STAILQ_REMOVE_HEAD(&iql, link);
1102 		if (service_iq(q, q->qsize / 8) == 0)
1103 			atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1104 		else
1105 			STAILQ_INSERT_TAIL(&iql, q, link);
1106 	}
1107 
1108 #if defined(INET) || defined(INET6)
1109 	if (iq->flags & IQ_LRO_ENABLED) {
1110 		struct lro_ctrl *lro = &rxq->lro;
1111 		struct lro_entry *l;
1112 
1113 		while (!SLIST_EMPTY(&lro->lro_active)) {
1114 			l = SLIST_FIRST(&lro->lro_active);
1115 			SLIST_REMOVE_HEAD(&lro->lro_active, next);
1116 			tcp_lro_flush(lro, l);
1117 		}
1118 	}
1119 #endif
1120 
1121 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_CIDXINC(ndescs) |
1122 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1123 
1124 	if (iq->flags & IQ_HAS_FL) {
1125 		int starved;
1126 
1127 		FL_LOCK(fl);
1128 		fl->needed += fl_bufs_used;
1129 		starved = refill_fl(sc, fl, fl->cap / 4);
1130 		FL_UNLOCK(fl);
1131 		if (__predict_false(starved != 0))
1132 			add_fl_to_sfl(sc, fl);
1133 	}
1134 
1135 	return (0);
1136 }
1137 
1138 static struct mbuf *
1139 get_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf,
1140     int *fl_bufs_used)
1141 {
1142 	struct mbuf *m0, *m;
1143 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1144 	unsigned int nbuf, len;
1145 
1146 	/*
1147 	 * No assertion for the fl lock because we don't need it.  This routine
1148 	 * is called only from the rx interrupt handler and it only updates
1149 	 * fl->cidx.  (Contrast that with fl->pidx/fl->needed which could be
1150 	 * updated in the rx interrupt handler or the starvation helper routine.
1151 	 * That's why code that manipulates fl->pidx/fl->needed needs the fl
1152 	 * lock but this routine does not).
1153 	 */
1154 
1155 	if (__predict_false((len_newbuf & F_RSPD_NEWBUF) == 0))
1156 		panic("%s: cannot handle packed frames", __func__);
1157 	len = G_RSPD_LEN(len_newbuf);
1158 
1159 	m0 = sd->m;
1160 	sd->m = NULL;	/* consumed */
1161 
1162 	bus_dmamap_sync(fl->tag[sd->tag_idx], sd->map, BUS_DMASYNC_POSTREAD);
1163 	m_init(m0, NULL, 0, M_NOWAIT, MT_DATA, M_PKTHDR);
1164 #ifdef T4_PKT_TIMESTAMP
1165 	/* Leave room for a timestamp */
1166 	m0->m_data += 8;
1167 #endif
1168 
1169 	if (len < RX_COPY_THRESHOLD) {
1170 		/* copy data to mbuf, buffer will be recycled */
1171 		bcopy(sd->cl, mtod(m0, caddr_t), len);
1172 		m0->m_len = len;
1173 	} else {
1174 		bus_dmamap_unload(fl->tag[sd->tag_idx], sd->map);
1175 		m_cljset(m0, sd->cl, FL_BUF_TYPE(sd->tag_idx));
1176 		sd->cl = NULL;	/* consumed */
1177 		m0->m_len = min(len, FL_BUF_SIZE(sd->tag_idx));
1178 	}
1179 	m0->m_pkthdr.len = len;
1180 
1181 	sd++;
1182 	if (__predict_false(++fl->cidx == fl->cap)) {
1183 		sd = fl->sdesc;
1184 		fl->cidx = 0;
1185 	}
1186 
1187 	m = m0;
1188 	len -= m->m_len;
1189 	nbuf = 1;	/* # of fl buffers used */
1190 
1191 	while (len > 0) {
1192 		m->m_next = sd->m;
1193 		sd->m = NULL;	/* consumed */
1194 		m = m->m_next;
1195 
1196 		bus_dmamap_sync(fl->tag[sd->tag_idx], sd->map,
1197 		    BUS_DMASYNC_POSTREAD);
1198 
1199 		m_init(m, NULL, 0, M_NOWAIT, MT_DATA, 0);
1200 		if (len <= MLEN) {
1201 			bcopy(sd->cl, mtod(m, caddr_t), len);
1202 			m->m_len = len;
1203 		} else {
1204 			bus_dmamap_unload(fl->tag[sd->tag_idx],
1205 			    sd->map);
1206 			m_cljset(m, sd->cl, FL_BUF_TYPE(sd->tag_idx));
1207 			sd->cl = NULL;	/* consumed */
1208 			m->m_len = min(len, FL_BUF_SIZE(sd->tag_idx));
1209 		}
1210 
1211 		sd++;
1212 		if (__predict_false(++fl->cidx == fl->cap)) {
1213 			sd = fl->sdesc;
1214 			fl->cidx = 0;
1215 		}
1216 
1217 		len -= m->m_len;
1218 		nbuf++;
1219 	}
1220 
1221 	(*fl_bufs_used) += nbuf;
1222 
1223 	return (m0);
1224 }
1225 
1226 static int
1227 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
1228 {
1229 	struct sge_rxq *rxq = iq_to_rxq(iq);
1230 	struct ifnet *ifp = rxq->ifp;
1231 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
1232 #if defined(INET) || defined(INET6)
1233 	struct lro_ctrl *lro = &rxq->lro;
1234 #endif
1235 
1236 	KASSERT(m0 != NULL, ("%s: no payload with opcode %02x", __func__,
1237 	    rss->opcode));
1238 
1239 	m0->m_pkthdr.len -= fl_pktshift;
1240 	m0->m_len -= fl_pktshift;
1241 	m0->m_data += fl_pktshift;
1242 
1243 	m0->m_pkthdr.rcvif = ifp;
1244 	m0->m_flags |= M_FLOWID;
1245 	m0->m_pkthdr.flowid = rss->hash_val;
1246 
1247 	if (cpl->csum_calc && !cpl->err_vec) {
1248 		if (ifp->if_capenable & IFCAP_RXCSUM &&
1249 		    cpl->l2info & htobe32(F_RXF_IP)) {
1250 			m0->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
1251 			    CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
1252 			rxq->rxcsum++;
1253 		} else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
1254 		    cpl->l2info & htobe32(F_RXF_IP6)) {
1255 			m0->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
1256 			    CSUM_PSEUDO_HDR);
1257 			rxq->rxcsum++;
1258 		}
1259 
1260 		if (__predict_false(cpl->ip_frag))
1261 			m0->m_pkthdr.csum_data = be16toh(cpl->csum);
1262 		else
1263 			m0->m_pkthdr.csum_data = 0xffff;
1264 	}
1265 
1266 	if (cpl->vlan_ex) {
1267 		m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
1268 		m0->m_flags |= M_VLANTAG;
1269 		rxq->vlan_extraction++;
1270 	}
1271 
1272 #if defined(INET) || defined(INET6)
1273 	if (cpl->l2info & htobe32(F_RXF_LRO) &&
1274 	    iq->flags & IQ_LRO_ENABLED &&
1275 	    tcp_lro_rx(lro, m0, 0) == 0) {
1276 		/* queued for LRO */
1277 	} else
1278 #endif
1279 	ifp->if_input(ifp, m0);
1280 
1281 	return (0);
1282 }
1283 
1284 /*
1285  * Doesn't fail.  Holds on to work requests it can't send right away.
1286  */
1287 void
1288 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
1289 {
1290 	struct sge_eq *eq = &wrq->eq;
1291 	int can_reclaim;
1292 	caddr_t dst;
1293 
1294 	TXQ_LOCK_ASSERT_OWNED(wrq);
1295 #ifdef TCP_OFFLOAD
1296 	KASSERT((eq->flags & EQ_TYPEMASK) == EQ_OFLD ||
1297 	    (eq->flags & EQ_TYPEMASK) == EQ_CTRL,
1298 	    ("%s: eq type %d", __func__, eq->flags & EQ_TYPEMASK));
1299 #else
1300 	KASSERT((eq->flags & EQ_TYPEMASK) == EQ_CTRL,
1301 	    ("%s: eq type %d", __func__, eq->flags & EQ_TYPEMASK));
1302 #endif
1303 
1304 	if (__predict_true(wr != NULL))
1305 		STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
1306 
1307 	can_reclaim = reclaimable(eq);
1308 	if (__predict_false(eq->flags & EQ_STALLED)) {
1309 		if (can_reclaim < tx_resume_threshold(eq))
1310 			return;
1311 		eq->flags &= ~EQ_STALLED;
1312 		eq->unstalled++;
1313 	}
1314 	eq->cidx += can_reclaim;
1315 	eq->avail += can_reclaim;
1316 	if (__predict_false(eq->cidx >= eq->cap))
1317 		eq->cidx -= eq->cap;
1318 
1319 	while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL) {
1320 		int ndesc;
1321 
1322 		if (__predict_false(wr->wr_len < 0 ||
1323 		    wr->wr_len > SGE_MAX_WR_LEN || (wr->wr_len & 0x7))) {
1324 
1325 #ifdef INVARIANTS
1326 			panic("%s: work request with length %d", __func__,
1327 			    wr->wr_len);
1328 #endif
1329 #ifdef KDB
1330 			kdb_backtrace();
1331 #endif
1332 			log(LOG_ERR, "%s: %s work request with length %d",
1333 			    device_get_nameunit(sc->dev), __func__, wr->wr_len);
1334 			STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
1335 			free_wrqe(wr);
1336 			continue;
1337 		}
1338 
1339 		ndesc = howmany(wr->wr_len, EQ_ESIZE);
1340 		if (eq->avail < ndesc) {
1341 			wrq->no_desc++;
1342 			break;
1343 		}
1344 
1345 		dst = (void *)&eq->desc[eq->pidx];
1346 		copy_to_txd(eq, wrtod(wr), &dst, wr->wr_len);
1347 
1348 		eq->pidx += ndesc;
1349 		eq->avail -= ndesc;
1350 		if (__predict_false(eq->pidx >= eq->cap))
1351 			eq->pidx -= eq->cap;
1352 
1353 		eq->pending += ndesc;
1354 		if (eq->pending >= 8)
1355 			ring_eq_db(sc, eq);
1356 
1357 		wrq->tx_wrs++;
1358 		STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
1359 		free_wrqe(wr);
1360 
1361 		if (eq->avail < 8) {
1362 			can_reclaim = reclaimable(eq);
1363 			eq->cidx += can_reclaim;
1364 			eq->avail += can_reclaim;
1365 			if (__predict_false(eq->cidx >= eq->cap))
1366 				eq->cidx -= eq->cap;
1367 		}
1368 	}
1369 
1370 	if (eq->pending)
1371 		ring_eq_db(sc, eq);
1372 
1373 	if (wr != NULL) {
1374 		eq->flags |= EQ_STALLED;
1375 		if (callout_pending(&eq->tx_callout) == 0)
1376 			callout_reset(&eq->tx_callout, 1, t4_tx_callout, eq);
1377 	}
1378 }
1379 
1380 /* Per-packet header in a coalesced tx WR, before the SGL starts (in flits) */
1381 #define TXPKTS_PKT_HDR ((\
1382     sizeof(struct ulp_txpkt) + \
1383     sizeof(struct ulptx_idata) + \
1384     sizeof(struct cpl_tx_pkt_core) \
1385     ) / 8)
1386 
1387 /* Header of a coalesced tx WR, before SGL of first packet (in flits) */
1388 #define TXPKTS_WR_HDR (\
1389     sizeof(struct fw_eth_tx_pkts_wr) / 8 + \
1390     TXPKTS_PKT_HDR)
1391 
1392 /* Header of a tx WR, before SGL of first packet (in flits) */
1393 #define TXPKT_WR_HDR ((\
1394     sizeof(struct fw_eth_tx_pkt_wr) + \
1395     sizeof(struct cpl_tx_pkt_core) \
1396     ) / 8 )
1397 
1398 /* Header of a tx LSO WR, before SGL of first packet (in flits) */
1399 #define TXPKT_LSO_WR_HDR ((\
1400     sizeof(struct fw_eth_tx_pkt_wr) + \
1401     sizeof(struct cpl_tx_pkt_lso_core) + \
1402     sizeof(struct cpl_tx_pkt_core) \
1403     ) / 8 )
1404 
1405 int
1406 t4_eth_tx(struct ifnet *ifp, struct sge_txq *txq, struct mbuf *m)
1407 {
1408 	struct port_info *pi = (void *)ifp->if_softc;
1409 	struct adapter *sc = pi->adapter;
1410 	struct sge_eq *eq = &txq->eq;
1411 	struct buf_ring *br = txq->br;
1412 	struct mbuf *next;
1413 	int rc, coalescing, can_reclaim;
1414 	struct txpkts txpkts;
1415 	struct sgl sgl;
1416 
1417 	TXQ_LOCK_ASSERT_OWNED(txq);
1418 	KASSERT(m, ("%s: called with nothing to do.", __func__));
1419 	KASSERT((eq->flags & EQ_TYPEMASK) == EQ_ETH,
1420 	    ("%s: eq type %d", __func__, eq->flags & EQ_TYPEMASK));
1421 
1422 	prefetch(&eq->desc[eq->pidx]);
1423 	prefetch(&txq->sdesc[eq->pidx]);
1424 
1425 	txpkts.npkt = 0;/* indicates there's nothing in txpkts */
1426 	coalescing = 0;
1427 
1428 	can_reclaim = reclaimable(eq);
1429 	if (__predict_false(eq->flags & EQ_STALLED)) {
1430 		if (can_reclaim < tx_resume_threshold(eq)) {
1431 			txq->m = m;
1432 			return (0);
1433 		}
1434 		eq->flags &= ~EQ_STALLED;
1435 		eq->unstalled++;
1436 	}
1437 
1438 	if (__predict_false(eq->flags & EQ_DOOMED)) {
1439 		m_freem(m);
1440 		while ((m = buf_ring_dequeue_sc(txq->br)) != NULL)
1441 			m_freem(m);
1442 		return (ENETDOWN);
1443 	}
1444 
1445 	if (eq->avail < 8 && can_reclaim)
1446 		reclaim_tx_descs(txq, can_reclaim, 32);
1447 
1448 	for (; m; m = next ? next : drbr_dequeue(ifp, br)) {
1449 
1450 		if (eq->avail < 8)
1451 			break;
1452 
1453 		next = m->m_nextpkt;
1454 		m->m_nextpkt = NULL;
1455 
1456 		if (next || buf_ring_peek(br))
1457 			coalescing = 1;
1458 
1459 		rc = get_pkt_sgl(txq, &m, &sgl, coalescing);
1460 		if (rc != 0) {
1461 			if (rc == ENOMEM) {
1462 
1463 				/* Short of resources, suspend tx */
1464 
1465 				m->m_nextpkt = next;
1466 				break;
1467 			}
1468 
1469 			/*
1470 			 * Unrecoverable error for this packet, throw it away
1471 			 * and move on to the next.  get_pkt_sgl may already
1472 			 * have freed m (it will be NULL in that case and the
1473 			 * m_freem here is still safe).
1474 			 */
1475 
1476 			m_freem(m);
1477 			continue;
1478 		}
1479 
1480 		if (coalescing &&
1481 		    add_to_txpkts(pi, txq, &txpkts, m, &sgl) == 0) {
1482 
1483 			/* Successfully absorbed into txpkts */
1484 
1485 			write_ulp_cpl_sgl(pi, txq, &txpkts, m, &sgl);
1486 			goto doorbell;
1487 		}
1488 
1489 		/*
1490 		 * We weren't coalescing to begin with, or current frame could
1491 		 * not be coalesced (add_to_txpkts flushes txpkts if a frame
1492 		 * given to it can't be coalesced).  Either way there should be
1493 		 * nothing in txpkts.
1494 		 */
1495 		KASSERT(txpkts.npkt == 0,
1496 		    ("%s: txpkts not empty: %d", __func__, txpkts.npkt));
1497 
1498 		/* We're sending out individual packets now */
1499 		coalescing = 0;
1500 
1501 		if (eq->avail < 8)
1502 			reclaim_tx_descs(txq, 0, 8);
1503 		rc = write_txpkt_wr(pi, txq, m, &sgl);
1504 		if (rc != 0) {
1505 
1506 			/* Short of hardware descriptors, suspend tx */
1507 
1508 			/*
1509 			 * This is an unlikely but expensive failure.  We've
1510 			 * done all the hard work (DMA mappings etc.) and now we
1511 			 * can't send out the packet.  What's worse, we have to
1512 			 * spend even more time freeing up everything in sgl.
1513 			 */
1514 			txq->no_desc++;
1515 			free_pkt_sgl(txq, &sgl);
1516 
1517 			m->m_nextpkt = next;
1518 			break;
1519 		}
1520 
1521 		ETHER_BPF_MTAP(ifp, m);
1522 		if (sgl.nsegs == 0)
1523 			m_freem(m);
1524 doorbell:
1525 		if (eq->pending >= 8)
1526 			ring_eq_db(sc, eq);
1527 
1528 		can_reclaim = reclaimable(eq);
1529 		if (can_reclaim >= 32)
1530 			reclaim_tx_descs(txq, can_reclaim, 64);
1531 	}
1532 
1533 	if (txpkts.npkt > 0)
1534 		write_txpkts_wr(txq, &txpkts);
1535 
1536 	/*
1537 	 * m not NULL means there was an error but we haven't thrown it away.
1538 	 * This can happen when we're short of tx descriptors (no_desc) or maybe
1539 	 * even DMA maps (no_dmamap).  Either way, a credit flush and reclaim
1540 	 * will get things going again.
1541 	 */
1542 	if (m && !(eq->flags & EQ_CRFLUSHED)) {
1543 		struct tx_sdesc *txsd = &txq->sdesc[eq->pidx];
1544 
1545 		/*
1546 		 * If EQ_CRFLUSHED is not set then we know we have at least one
1547 		 * available descriptor because any WR that reduces eq->avail to
1548 		 * 0 also sets EQ_CRFLUSHED.
1549 		 */
1550 		KASSERT(eq->avail > 0, ("%s: no space for eqflush.", __func__));
1551 
1552 		txsd->desc_used = 1;
1553 		txsd->credits = 0;
1554 		write_eqflush_wr(eq);
1555 	}
1556 	txq->m = m;
1557 
1558 	if (eq->pending)
1559 		ring_eq_db(sc, eq);
1560 
1561 	reclaim_tx_descs(txq, 0, 128);
1562 
1563 	if (eq->flags & EQ_STALLED && callout_pending(&eq->tx_callout) == 0)
1564 		callout_reset(&eq->tx_callout, 1, t4_tx_callout, eq);
1565 
1566 	return (0);
1567 }
1568 
1569 void
1570 t4_update_fl_bufsize(struct ifnet *ifp)
1571 {
1572 	struct port_info *pi = ifp->if_softc;
1573 	struct sge_rxq *rxq;
1574 #ifdef TCP_OFFLOAD
1575 	struct sge_ofld_rxq *ofld_rxq;
1576 #endif
1577 	struct sge_fl *fl;
1578 	int i, bufsize;
1579 
1580 	bufsize = mtu_to_bufsize(ifp->if_mtu);
1581 	for_each_rxq(pi, i, rxq) {
1582 		fl = &rxq->fl;
1583 
1584 		FL_LOCK(fl);
1585 		set_fl_tag_idx(fl, bufsize);
1586 		FL_UNLOCK(fl);
1587 	}
1588 #ifdef TCP_OFFLOAD
1589 	bufsize = mtu_to_bufsize_toe(pi->adapter, ifp->if_mtu);
1590 	for_each_ofld_rxq(pi, i, ofld_rxq) {
1591 		fl = &ofld_rxq->fl;
1592 
1593 		FL_LOCK(fl);
1594 		set_fl_tag_idx(fl, bufsize);
1595 		FL_UNLOCK(fl);
1596 	}
1597 #endif
1598 }
1599 
1600 int
1601 can_resume_tx(struct sge_eq *eq)
1602 {
1603 	return (reclaimable(eq) >= tx_resume_threshold(eq));
1604 }
1605 
1606 static inline void
1607 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
1608     int qsize, int esize)
1609 {
1610 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
1611 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
1612 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
1613 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
1614 
1615 	iq->flags = 0;
1616 	iq->adapter = sc;
1617 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
1618 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
1619 	if (pktc_idx >= 0) {
1620 		iq->intr_params |= F_QINTR_CNT_EN;
1621 		iq->intr_pktc_idx = pktc_idx;
1622 	}
1623 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
1624 	iq->esize = max(esize, 16);		/* See FW_IQ_CMD/iqesize */
1625 }
1626 
1627 static inline void
1628 init_fl(struct sge_fl *fl, int qsize, int bufsize, char *name)
1629 {
1630 	fl->qsize = qsize;
1631 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
1632 	set_fl_tag_idx(fl, bufsize);
1633 }
1634 
1635 static inline void
1636 init_eq(struct sge_eq *eq, int eqtype, int qsize, uint8_t tx_chan,
1637     uint16_t iqid, char *name)
1638 {
1639 	KASSERT(tx_chan < NCHAN, ("%s: bad tx channel %d", __func__, tx_chan));
1640 	KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
1641 
1642 	eq->flags = eqtype & EQ_TYPEMASK;
1643 	eq->tx_chan = tx_chan;
1644 	eq->iqid = iqid;
1645 	eq->qsize = qsize;
1646 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
1647 
1648 	TASK_INIT(&eq->tx_task, 0, t4_tx_task, eq);
1649 	callout_init(&eq->tx_callout, CALLOUT_MPSAFE);
1650 }
1651 
1652 static int
1653 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
1654     bus_dmamap_t *map, bus_addr_t *pa, void **va)
1655 {
1656 	int rc;
1657 
1658 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
1659 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
1660 	if (rc != 0) {
1661 		device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
1662 		goto done;
1663 	}
1664 
1665 	rc = bus_dmamem_alloc(*tag, va,
1666 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
1667 	if (rc != 0) {
1668 		device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
1669 		goto done;
1670 	}
1671 
1672 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
1673 	if (rc != 0) {
1674 		device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
1675 		goto done;
1676 	}
1677 done:
1678 	if (rc)
1679 		free_ring(sc, *tag, *map, *pa, *va);
1680 
1681 	return (rc);
1682 }
1683 
1684 static int
1685 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
1686     bus_addr_t pa, void *va)
1687 {
1688 	if (pa)
1689 		bus_dmamap_unload(tag, map);
1690 	if (va)
1691 		bus_dmamem_free(tag, va, map);
1692 	if (tag)
1693 		bus_dma_tag_destroy(tag);
1694 
1695 	return (0);
1696 }
1697 
1698 /*
1699  * Allocates the ring for an ingress queue and an optional freelist.  If the
1700  * freelist is specified it will be allocated and then associated with the
1701  * ingress queue.
1702  *
1703  * Returns errno on failure.  Resources allocated up to that point may still be
1704  * allocated.  Caller is responsible for cleanup in case this function fails.
1705  *
1706  * If the ingress queue will take interrupts directly (iq->flags & IQ_INTR) then
1707  * the intr_idx specifies the vector, starting from 0.  Otherwise it specifies
1708  * the abs_id of the ingress queue to which its interrupts should be forwarded.
1709  */
1710 static int
1711 alloc_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl,
1712     int intr_idx, int cong)
1713 {
1714 	int rc, i, cntxt_id;
1715 	size_t len;
1716 	struct fw_iq_cmd c;
1717 	struct adapter *sc = iq->adapter;
1718 	__be32 v = 0;
1719 
1720 	len = iq->qsize * iq->esize;
1721 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
1722 	    (void **)&iq->desc);
1723 	if (rc != 0)
1724 		return (rc);
1725 
1726 	bzero(&c, sizeof(c));
1727 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
1728 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
1729 	    V_FW_IQ_CMD_VFN(0));
1730 
1731 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
1732 	    FW_LEN16(c));
1733 
1734 	/* Special handling for firmware event queue */
1735 	if (iq == &sc->sge.fwq)
1736 		v |= F_FW_IQ_CMD_IQASYNCH;
1737 
1738 	if (iq->flags & IQ_INTR) {
1739 		KASSERT(intr_idx < sc->intr_count,
1740 		    ("%s: invalid direct intr_idx %d", __func__, intr_idx));
1741 	} else
1742 		v |= F_FW_IQ_CMD_IQANDST;
1743 	v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
1744 
1745 	c.type_to_iqandstindex = htobe32(v |
1746 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
1747 	    V_FW_IQ_CMD_VIID(pi->viid) |
1748 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
1749 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
1750 	    F_FW_IQ_CMD_IQGTSMODE |
1751 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
1752 	    V_FW_IQ_CMD_IQESIZE(ilog2(iq->esize) - 4));
1753 	c.iqsize = htobe16(iq->qsize);
1754 	c.iqaddr = htobe64(iq->ba);
1755 	if (cong >= 0)
1756 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
1757 
1758 	if (fl) {
1759 		mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
1760 
1761 		for (i = 0; i < FL_BUF_SIZES; i++) {
1762 
1763 			/*
1764 			 * A freelist buffer must be 16 byte aligned as the SGE
1765 			 * uses the low 4 bits of the bus addr to figure out the
1766 			 * buffer size.
1767 			 */
1768 			rc = bus_dma_tag_create(sc->dmat, 16, 0,
1769 			    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
1770 			    FL_BUF_SIZE(i), 1, FL_BUF_SIZE(i), BUS_DMA_ALLOCNOW,
1771 			    NULL, NULL, &fl->tag[i]);
1772 			if (rc != 0) {
1773 				device_printf(sc->dev,
1774 				    "failed to create fl DMA tag[%d]: %d\n",
1775 				    i, rc);
1776 				return (rc);
1777 			}
1778 		}
1779 		len = fl->qsize * RX_FL_ESIZE;
1780 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
1781 		    &fl->ba, (void **)&fl->desc);
1782 		if (rc)
1783 			return (rc);
1784 
1785 		/* Allocate space for one software descriptor per buffer. */
1786 		fl->cap = (fl->qsize - spg_len / RX_FL_ESIZE) * 8;
1787 		FL_LOCK(fl);
1788 		rc = alloc_fl_sdesc(fl);
1789 		FL_UNLOCK(fl);
1790 		if (rc != 0) {
1791 			device_printf(sc->dev,
1792 			    "failed to setup fl software descriptors: %d\n",
1793 			    rc);
1794 			return (rc);
1795 		}
1796 		fl->needed = fl->cap;
1797 		fl->lowat = roundup2(sc->sge.fl_starve_threshold, 8);
1798 
1799 		c.iqns_to_fl0congen |=
1800 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
1801 			F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
1802 			F_FW_IQ_CMD_FL0PADEN);
1803 		if (cong >= 0) {
1804 			c.iqns_to_fl0congen |=
1805 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
1806 				    F_FW_IQ_CMD_FL0CONGCIF |
1807 				    F_FW_IQ_CMD_FL0CONGEN);
1808 		}
1809 		c.fl0dcaen_to_fl0cidxfthresh =
1810 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(X_FETCHBURSTMIN_64B) |
1811 			V_FW_IQ_CMD_FL0FBMAX(X_FETCHBURSTMAX_512B));
1812 		c.fl0size = htobe16(fl->qsize);
1813 		c.fl0addr = htobe64(fl->ba);
1814 	}
1815 
1816 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
1817 	if (rc != 0) {
1818 		device_printf(sc->dev,
1819 		    "failed to create ingress queue: %d\n", rc);
1820 		return (rc);
1821 	}
1822 
1823 	iq->cdesc = iq->desc;
1824 	iq->cidx = 0;
1825 	iq->gen = 1;
1826 	iq->intr_next = iq->intr_params;
1827 	iq->cntxt_id = be16toh(c.iqid);
1828 	iq->abs_id = be16toh(c.physiqid);
1829 	iq->flags |= IQ_ALLOCATED;
1830 
1831 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
1832 	if (cntxt_id >= sc->sge.niq) {
1833 		panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
1834 		    cntxt_id, sc->sge.niq - 1);
1835 	}
1836 	sc->sge.iqmap[cntxt_id] = iq;
1837 
1838 	if (fl) {
1839 		fl->cntxt_id = be16toh(c.fl0id);
1840 		fl->pidx = fl->cidx = 0;
1841 
1842 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
1843 		if (cntxt_id >= sc->sge.neq) {
1844 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
1845 			    __func__, cntxt_id, sc->sge.neq - 1);
1846 		}
1847 		sc->sge.eqmap[cntxt_id] = (void *)fl;
1848 
1849 		FL_LOCK(fl);
1850 		/* Enough to make sure the SGE doesn't think it's starved */
1851 		refill_fl(sc, fl, fl->lowat);
1852 		FL_UNLOCK(fl);
1853 
1854 		iq->flags |= IQ_HAS_FL;
1855 	}
1856 
1857 	/* Enable IQ interrupts */
1858 	atomic_store_rel_int(&iq->state, IQS_IDLE);
1859 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_SEINTARM(iq->intr_params) |
1860 	    V_INGRESSQID(iq->cntxt_id));
1861 
1862 	return (0);
1863 }
1864 
1865 static int
1866 free_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl)
1867 {
1868 	int i, rc;
1869 	struct adapter *sc = iq->adapter;
1870 	device_t dev;
1871 
1872 	if (sc == NULL)
1873 		return (0);	/* nothing to do */
1874 
1875 	dev = pi ? pi->dev : sc->dev;
1876 
1877 	if (iq->flags & IQ_ALLOCATED) {
1878 		rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
1879 		    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
1880 		    fl ? fl->cntxt_id : 0xffff, 0xffff);
1881 		if (rc != 0) {
1882 			device_printf(dev,
1883 			    "failed to free queue %p: %d\n", iq, rc);
1884 			return (rc);
1885 		}
1886 		iq->flags &= ~IQ_ALLOCATED;
1887 	}
1888 
1889 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
1890 
1891 	bzero(iq, sizeof(*iq));
1892 
1893 	if (fl) {
1894 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
1895 		    fl->desc);
1896 
1897 		if (fl->sdesc) {
1898 			FL_LOCK(fl);
1899 			free_fl_sdesc(fl);
1900 			FL_UNLOCK(fl);
1901 		}
1902 
1903 		if (mtx_initialized(&fl->fl_lock))
1904 			mtx_destroy(&fl->fl_lock);
1905 
1906 		for (i = 0; i < FL_BUF_SIZES; i++) {
1907 			if (fl->tag[i])
1908 				bus_dma_tag_destroy(fl->tag[i]);
1909 		}
1910 
1911 		bzero(fl, sizeof(*fl));
1912 	}
1913 
1914 	return (0);
1915 }
1916 
1917 static int
1918 alloc_fwq(struct adapter *sc)
1919 {
1920 	int rc, intr_idx;
1921 	struct sge_iq *fwq = &sc->sge.fwq;
1922 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
1923 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
1924 
1925 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE, FW_IQ_ESIZE);
1926 	fwq->flags |= IQ_INTR;	/* always */
1927 	intr_idx = sc->intr_count > 1 ? 1 : 0;
1928 	rc = alloc_iq_fl(sc->port[0], fwq, NULL, intr_idx, -1);
1929 	if (rc != 0) {
1930 		device_printf(sc->dev,
1931 		    "failed to create firmware event queue: %d\n", rc);
1932 		return (rc);
1933 	}
1934 
1935 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD,
1936 	    NULL, "firmware event queue");
1937 	children = SYSCTL_CHILDREN(oid);
1938 
1939 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "abs_id",
1940 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->abs_id, 0, sysctl_uint16, "I",
1941 	    "absolute id of the queue");
1942 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cntxt_id",
1943 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->cntxt_id, 0, sysctl_uint16, "I",
1944 	    "SGE context id of the queue");
1945 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cidx",
1946 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->cidx, 0, sysctl_uint16, "I",
1947 	    "consumer index");
1948 
1949 	return (0);
1950 }
1951 
1952 static int
1953 free_fwq(struct adapter *sc)
1954 {
1955 	return free_iq_fl(NULL, &sc->sge.fwq, NULL);
1956 }
1957 
1958 static int
1959 alloc_mgmtq(struct adapter *sc)
1960 {
1961 	int rc;
1962 	struct sge_wrq *mgmtq = &sc->sge.mgmtq;
1963 	char name[16];
1964 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
1965 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
1966 
1967 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "mgmtq", CTLFLAG_RD,
1968 	    NULL, "management queue");
1969 
1970 	snprintf(name, sizeof(name), "%s mgmtq", device_get_nameunit(sc->dev));
1971 	init_eq(&mgmtq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[0]->tx_chan,
1972 	    sc->sge.fwq.cntxt_id, name);
1973 	rc = alloc_wrq(sc, NULL, mgmtq, oid);
1974 	if (rc != 0) {
1975 		device_printf(sc->dev,
1976 		    "failed to create management queue: %d\n", rc);
1977 		return (rc);
1978 	}
1979 
1980 	return (0);
1981 }
1982 
1983 static int
1984 free_mgmtq(struct adapter *sc)
1985 {
1986 
1987 	return free_wrq(sc, &sc->sge.mgmtq);
1988 }
1989 
1990 static inline int
1991 tnl_cong(struct port_info *pi)
1992 {
1993 
1994 	if (cong_drop == -1)
1995 		return (-1);
1996 	else if (cong_drop == 1)
1997 		return (0);
1998 	else
1999 		return (1 << pi->tx_chan);
2000 }
2001 
2002 static int
2003 alloc_rxq(struct port_info *pi, struct sge_rxq *rxq, int intr_idx, int idx,
2004     struct sysctl_oid *oid)
2005 {
2006 	int rc;
2007 	struct sysctl_oid_list *children;
2008 	char name[16];
2009 
2010 	rc = alloc_iq_fl(pi, &rxq->iq, &rxq->fl, intr_idx, tnl_cong(pi));
2011 	if (rc != 0)
2012 		return (rc);
2013 
2014 	FL_LOCK(&rxq->fl);
2015 	refill_fl(pi->adapter, &rxq->fl, rxq->fl.needed / 8);
2016 	FL_UNLOCK(&rxq->fl);
2017 
2018 #if defined(INET) || defined(INET6)
2019 	rc = tcp_lro_init(&rxq->lro);
2020 	if (rc != 0)
2021 		return (rc);
2022 	rxq->lro.ifp = pi->ifp; /* also indicates LRO init'ed */
2023 
2024 	if (pi->ifp->if_capenable & IFCAP_LRO)
2025 		rxq->iq.flags |= IQ_LRO_ENABLED;
2026 #endif
2027 	rxq->ifp = pi->ifp;
2028 
2029 	children = SYSCTL_CHILDREN(oid);
2030 
2031 	snprintf(name, sizeof(name), "%d", idx);
2032 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
2033 	    NULL, "rx queue");
2034 	children = SYSCTL_CHILDREN(oid);
2035 
2036 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "abs_id",
2037 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.abs_id, 0, sysctl_uint16, "I",
2038 	    "absolute id of the queue");
2039 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cntxt_id",
2040 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cntxt_id, 0, sysctl_uint16, "I",
2041 	    "SGE context id of the queue");
2042 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cidx",
2043 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cidx, 0, sysctl_uint16, "I",
2044 	    "consumer index");
2045 #if defined(INET) || defined(INET6)
2046 	SYSCTL_ADD_INT(&pi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
2047 	    &rxq->lro.lro_queued, 0, NULL);
2048 	SYSCTL_ADD_INT(&pi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
2049 	    &rxq->lro.lro_flushed, 0, NULL);
2050 #endif
2051 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
2052 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
2053 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "vlan_extraction",
2054 	    CTLFLAG_RD, &rxq->vlan_extraction,
2055 	    "# of times hardware extracted 802.1Q tag");
2056 
2057 	children = SYSCTL_CHILDREN(oid);
2058 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "fl", CTLFLAG_RD,
2059 	    NULL, "freelist");
2060 	children = SYSCTL_CHILDREN(oid);
2061 
2062 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cntxt_id",
2063 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->fl.cntxt_id, 0, sysctl_uint16, "I",
2064 	    "SGE context id of the queue");
2065 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
2066 	    &rxq->fl.cidx, 0, "consumer index");
2067 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
2068 	    &rxq->fl.pidx, 0, "producer index");
2069 
2070 	return (rc);
2071 }
2072 
2073 static int
2074 free_rxq(struct port_info *pi, struct sge_rxq *rxq)
2075 {
2076 	int rc;
2077 
2078 #if defined(INET) || defined(INET6)
2079 	if (rxq->lro.ifp) {
2080 		tcp_lro_free(&rxq->lro);
2081 		rxq->lro.ifp = NULL;
2082 	}
2083 #endif
2084 
2085 	rc = free_iq_fl(pi, &rxq->iq, &rxq->fl);
2086 	if (rc == 0)
2087 		bzero(rxq, sizeof(*rxq));
2088 
2089 	return (rc);
2090 }
2091 
2092 #ifdef TCP_OFFLOAD
2093 static int
2094 alloc_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq,
2095     int intr_idx, int idx, struct sysctl_oid *oid)
2096 {
2097 	int rc;
2098 	struct sysctl_oid_list *children;
2099 	char name[16];
2100 
2101 	rc = alloc_iq_fl(pi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx,
2102 	    1 << pi->tx_chan);
2103 	if (rc != 0)
2104 		return (rc);
2105 
2106 	children = SYSCTL_CHILDREN(oid);
2107 
2108 	snprintf(name, sizeof(name), "%d", idx);
2109 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
2110 	    NULL, "rx queue");
2111 	children = SYSCTL_CHILDREN(oid);
2112 
2113 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "abs_id",
2114 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.abs_id, 0, sysctl_uint16,
2115 	    "I", "absolute id of the queue");
2116 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cntxt_id",
2117 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cntxt_id, 0, sysctl_uint16,
2118 	    "I", "SGE context id of the queue");
2119 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cidx",
2120 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cidx, 0, sysctl_uint16, "I",
2121 	    "consumer index");
2122 
2123 	children = SYSCTL_CHILDREN(oid);
2124 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "fl", CTLFLAG_RD,
2125 	    NULL, "freelist");
2126 	children = SYSCTL_CHILDREN(oid);
2127 
2128 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cntxt_id",
2129 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->fl.cntxt_id, 0, sysctl_uint16,
2130 	    "I", "SGE context id of the queue");
2131 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
2132 	    &ofld_rxq->fl.cidx, 0, "consumer index");
2133 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
2134 	    &ofld_rxq->fl.pidx, 0, "producer index");
2135 
2136 	return (rc);
2137 }
2138 
2139 static int
2140 free_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq)
2141 {
2142 	int rc;
2143 
2144 	rc = free_iq_fl(pi, &ofld_rxq->iq, &ofld_rxq->fl);
2145 	if (rc == 0)
2146 		bzero(ofld_rxq, sizeof(*ofld_rxq));
2147 
2148 	return (rc);
2149 }
2150 #endif
2151 
2152 static int
2153 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
2154 {
2155 	int rc, cntxt_id;
2156 	struct fw_eq_ctrl_cmd c;
2157 
2158 	bzero(&c, sizeof(c));
2159 
2160 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
2161 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
2162 	    V_FW_EQ_CTRL_CMD_VFN(0));
2163 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
2164 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
2165 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid)); /* XXX */
2166 	c.physeqid_pkd = htobe32(0);
2167 	c.fetchszm_to_iqid =
2168 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
2169 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
2170 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
2171 	c.dcaen_to_eqsize =
2172 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
2173 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
2174 		V_FW_EQ_CTRL_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
2175 		V_FW_EQ_CTRL_CMD_EQSIZE(eq->qsize));
2176 	c.eqaddr = htobe64(eq->ba);
2177 
2178 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
2179 	if (rc != 0) {
2180 		device_printf(sc->dev,
2181 		    "failed to create control queue %d: %d\n", eq->tx_chan, rc);
2182 		return (rc);
2183 	}
2184 	eq->flags |= EQ_ALLOCATED;
2185 
2186 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
2187 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
2188 	if (cntxt_id >= sc->sge.neq)
2189 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
2190 		cntxt_id, sc->sge.neq - 1);
2191 	sc->sge.eqmap[cntxt_id] = eq;
2192 
2193 	return (rc);
2194 }
2195 
2196 static int
2197 eth_eq_alloc(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
2198 {
2199 	int rc, cntxt_id;
2200 	struct fw_eq_eth_cmd c;
2201 
2202 	bzero(&c, sizeof(c));
2203 
2204 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
2205 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
2206 	    V_FW_EQ_ETH_CMD_VFN(0));
2207 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
2208 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
2209 	c.viid_pkd = htobe32(V_FW_EQ_ETH_CMD_VIID(pi->viid));
2210 	c.fetchszm_to_iqid =
2211 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
2212 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
2213 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
2214 	c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
2215 		      V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
2216 		      V_FW_EQ_ETH_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
2217 		      V_FW_EQ_ETH_CMD_EQSIZE(eq->qsize));
2218 	c.eqaddr = htobe64(eq->ba);
2219 
2220 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
2221 	if (rc != 0) {
2222 		device_printf(pi->dev,
2223 		    "failed to create Ethernet egress queue: %d\n", rc);
2224 		return (rc);
2225 	}
2226 	eq->flags |= EQ_ALLOCATED;
2227 
2228 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
2229 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
2230 	if (cntxt_id >= sc->sge.neq)
2231 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
2232 		cntxt_id, sc->sge.neq - 1);
2233 	sc->sge.eqmap[cntxt_id] = eq;
2234 
2235 	return (rc);
2236 }
2237 
2238 #ifdef TCP_OFFLOAD
2239 static int
2240 ofld_eq_alloc(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
2241 {
2242 	int rc, cntxt_id;
2243 	struct fw_eq_ofld_cmd c;
2244 
2245 	bzero(&c, sizeof(c));
2246 
2247 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
2248 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
2249 	    V_FW_EQ_OFLD_CMD_VFN(0));
2250 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
2251 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
2252 	c.fetchszm_to_iqid =
2253 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
2254 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
2255 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
2256 	c.dcaen_to_eqsize =
2257 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
2258 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
2259 		V_FW_EQ_OFLD_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
2260 		V_FW_EQ_OFLD_CMD_EQSIZE(eq->qsize));
2261 	c.eqaddr = htobe64(eq->ba);
2262 
2263 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
2264 	if (rc != 0) {
2265 		device_printf(pi->dev,
2266 		    "failed to create egress queue for TCP offload: %d\n", rc);
2267 		return (rc);
2268 	}
2269 	eq->flags |= EQ_ALLOCATED;
2270 
2271 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
2272 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
2273 	if (cntxt_id >= sc->sge.neq)
2274 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
2275 		cntxt_id, sc->sge.neq - 1);
2276 	sc->sge.eqmap[cntxt_id] = eq;
2277 
2278 	return (rc);
2279 }
2280 #endif
2281 
2282 static int
2283 alloc_eq(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
2284 {
2285 	int rc;
2286 	size_t len;
2287 
2288 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
2289 
2290 	len = eq->qsize * EQ_ESIZE;
2291 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
2292 	    &eq->ba, (void **)&eq->desc);
2293 	if (rc)
2294 		return (rc);
2295 
2296 	eq->cap = eq->qsize - spg_len / EQ_ESIZE;
2297 	eq->spg = (void *)&eq->desc[eq->cap];
2298 	eq->avail = eq->cap - 1;	/* one less to avoid cidx = pidx */
2299 	eq->pidx = eq->cidx = 0;
2300 	eq->doorbells = sc->doorbells;
2301 
2302 	switch (eq->flags & EQ_TYPEMASK) {
2303 	case EQ_CTRL:
2304 		rc = ctrl_eq_alloc(sc, eq);
2305 		break;
2306 
2307 	case EQ_ETH:
2308 		rc = eth_eq_alloc(sc, pi, eq);
2309 		break;
2310 
2311 #ifdef TCP_OFFLOAD
2312 	case EQ_OFLD:
2313 		rc = ofld_eq_alloc(sc, pi, eq);
2314 		break;
2315 #endif
2316 
2317 	default:
2318 		panic("%s: invalid eq type %d.", __func__,
2319 		    eq->flags & EQ_TYPEMASK);
2320 	}
2321 	if (rc != 0) {
2322 		device_printf(sc->dev,
2323 		    "failed to allocate egress queue(%d): %d",
2324 		    eq->flags & EQ_TYPEMASK, rc);
2325 	}
2326 
2327 	eq->tx_callout.c_cpu = eq->cntxt_id % mp_ncpus;
2328 
2329 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
2330 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
2331 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
2332 		uint32_t s_qpp = sc->sge.s_qpp;
2333 		uint32_t mask = (1 << s_qpp) - 1;
2334 		volatile uint8_t *udb;
2335 
2336 		udb = sc->udbs_base + UDBS_DB_OFFSET;
2337 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
2338 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
2339 		if (eq->udb_qid > PAGE_SIZE / UDBS_SEG_SIZE)
2340 	    		clrbit(&eq->doorbells, DOORBELL_WCWR);
2341 		else {
2342 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
2343 			eq->udb_qid = 0;
2344 		}
2345 		eq->udb = (volatile void *)udb;
2346 	}
2347 
2348 	return (rc);
2349 }
2350 
2351 static int
2352 free_eq(struct adapter *sc, struct sge_eq *eq)
2353 {
2354 	int rc;
2355 
2356 	if (eq->flags & EQ_ALLOCATED) {
2357 		switch (eq->flags & EQ_TYPEMASK) {
2358 		case EQ_CTRL:
2359 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
2360 			    eq->cntxt_id);
2361 			break;
2362 
2363 		case EQ_ETH:
2364 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
2365 			    eq->cntxt_id);
2366 			break;
2367 
2368 #ifdef TCP_OFFLOAD
2369 		case EQ_OFLD:
2370 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
2371 			    eq->cntxt_id);
2372 			break;
2373 #endif
2374 
2375 		default:
2376 			panic("%s: invalid eq type %d.", __func__,
2377 			    eq->flags & EQ_TYPEMASK);
2378 		}
2379 		if (rc != 0) {
2380 			device_printf(sc->dev,
2381 			    "failed to free egress queue (%d): %d\n",
2382 			    eq->flags & EQ_TYPEMASK, rc);
2383 			return (rc);
2384 		}
2385 		eq->flags &= ~EQ_ALLOCATED;
2386 	}
2387 
2388 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
2389 
2390 	if (mtx_initialized(&eq->eq_lock))
2391 		mtx_destroy(&eq->eq_lock);
2392 
2393 	bzero(eq, sizeof(*eq));
2394 	return (0);
2395 }
2396 
2397 static int
2398 alloc_wrq(struct adapter *sc, struct port_info *pi, struct sge_wrq *wrq,
2399     struct sysctl_oid *oid)
2400 {
2401 	int rc;
2402 	struct sysctl_ctx_list *ctx = pi ? &pi->ctx : &sc->ctx;
2403 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2404 
2405 	rc = alloc_eq(sc, pi, &wrq->eq);
2406 	if (rc)
2407 		return (rc);
2408 
2409 	wrq->adapter = sc;
2410 	STAILQ_INIT(&wrq->wr_list);
2411 
2412 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
2413 	    &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
2414 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
2415 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I",
2416 	    "consumer index");
2417 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx",
2418 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I",
2419 	    "producer index");
2420 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs", CTLFLAG_RD,
2421 	    &wrq->tx_wrs, "# of work requests");
2422 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "no_desc", CTLFLAG_RD,
2423 	    &wrq->no_desc, 0,
2424 	    "# of times queue ran out of hardware descriptors");
2425 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "unstalled", CTLFLAG_RD,
2426 	    &wrq->eq.unstalled, 0, "# of times queue recovered after stall");
2427 
2428 
2429 	return (rc);
2430 }
2431 
2432 static int
2433 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
2434 {
2435 	int rc;
2436 
2437 	rc = free_eq(sc, &wrq->eq);
2438 	if (rc)
2439 		return (rc);
2440 
2441 	bzero(wrq, sizeof(*wrq));
2442 	return (0);
2443 }
2444 
2445 static int
2446 alloc_txq(struct port_info *pi, struct sge_txq *txq, int idx,
2447     struct sysctl_oid *oid)
2448 {
2449 	int rc;
2450 	struct adapter *sc = pi->adapter;
2451 	struct sge_eq *eq = &txq->eq;
2452 	char name[16];
2453 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2454 
2455 	rc = alloc_eq(sc, pi, eq);
2456 	if (rc)
2457 		return (rc);
2458 
2459 	txq->ifp = pi->ifp;
2460 
2461 	txq->sdesc = malloc(eq->cap * sizeof(struct tx_sdesc), M_CXGBE,
2462 	    M_ZERO | M_WAITOK);
2463 	txq->br = buf_ring_alloc(eq->qsize, M_CXGBE, M_WAITOK, &eq->eq_lock);
2464 
2465 	rc = bus_dma_tag_create(sc->dmat, 1, 0, BUS_SPACE_MAXADDR,
2466 	    BUS_SPACE_MAXADDR, NULL, NULL, 64 * 1024, TX_SGL_SEGS,
2467 	    BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL, NULL, &txq->tx_tag);
2468 	if (rc != 0) {
2469 		device_printf(sc->dev,
2470 		    "failed to create tx DMA tag: %d\n", rc);
2471 		return (rc);
2472 	}
2473 
2474 	/*
2475 	 * We can stuff ~10 frames in an 8-descriptor txpkts WR (8 is the SGE
2476 	 * limit for any WR).  txq->no_dmamap events shouldn't occur if maps is
2477 	 * sized for the worst case.
2478 	 */
2479 	rc = t4_alloc_tx_maps(&txq->txmaps, txq->tx_tag, eq->qsize * 10 / 8,
2480 	    M_WAITOK);
2481 	if (rc != 0) {
2482 		device_printf(sc->dev, "failed to setup tx DMA maps: %d\n", rc);
2483 		return (rc);
2484 	}
2485 
2486 	snprintf(name, sizeof(name), "%d", idx);
2487 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
2488 	    NULL, "tx queue");
2489 	children = SYSCTL_CHILDREN(oid);
2490 
2491 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
2492 	    &eq->cntxt_id, 0, "SGE context id of the queue");
2493 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cidx",
2494 	    CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I",
2495 	    "consumer index");
2496 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "pidx",
2497 	    CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I",
2498 	    "producer index");
2499 
2500 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
2501 	    &txq->txcsum, "# of times hardware assisted with checksum");
2502 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "vlan_insertion",
2503 	    CTLFLAG_RD, &txq->vlan_insertion,
2504 	    "# of times hardware inserted 802.1Q tag");
2505 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
2506 	    &txq->tso_wrs, "# of TSO work requests");
2507 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
2508 	    &txq->imm_wrs, "# of work requests with immediate data");
2509 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
2510 	    &txq->sgl_wrs, "# of work requests with direct SGL");
2511 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
2512 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
2513 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txpkts_wrs", CTLFLAG_RD,
2514 	    &txq->txpkts_wrs, "# of txpkts work requests (multiple pkts/WR)");
2515 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txpkts_pkts", CTLFLAG_RD,
2516 	    &txq->txpkts_pkts, "# of frames tx'd using txpkts work requests");
2517 
2518 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "br_drops", CTLFLAG_RD,
2519 	    &txq->br->br_drops, "# of drops in the buf_ring for this queue");
2520 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "no_dmamap", CTLFLAG_RD,
2521 	    &txq->no_dmamap, 0, "# of times txq ran out of DMA maps");
2522 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "no_desc", CTLFLAG_RD,
2523 	    &txq->no_desc, 0, "# of times txq ran out of hardware descriptors");
2524 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "egr_update", CTLFLAG_RD,
2525 	    &eq->egr_update, 0, "egress update notifications from the SGE");
2526 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "unstalled", CTLFLAG_RD,
2527 	    &eq->unstalled, 0, "# of times txq recovered after stall");
2528 
2529 	return (rc);
2530 }
2531 
2532 static int
2533 free_txq(struct port_info *pi, struct sge_txq *txq)
2534 {
2535 	int rc;
2536 	struct adapter *sc = pi->adapter;
2537 	struct sge_eq *eq = &txq->eq;
2538 
2539 	rc = free_eq(sc, eq);
2540 	if (rc)
2541 		return (rc);
2542 
2543 	free(txq->sdesc, M_CXGBE);
2544 
2545 	if (txq->txmaps.maps)
2546 		t4_free_tx_maps(&txq->txmaps, txq->tx_tag);
2547 
2548 	buf_ring_free(txq->br, M_CXGBE);
2549 
2550 	if (txq->tx_tag)
2551 		bus_dma_tag_destroy(txq->tx_tag);
2552 
2553 	bzero(txq, sizeof(*txq));
2554 	return (0);
2555 }
2556 
2557 static void
2558 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
2559 {
2560 	bus_addr_t *ba = arg;
2561 
2562 	KASSERT(nseg == 1,
2563 	    ("%s meant for single segment mappings only.", __func__));
2564 
2565 	*ba = error ? 0 : segs->ds_addr;
2566 }
2567 
2568 static inline bool
2569 is_new_response(const struct sge_iq *iq, struct rsp_ctrl **ctrl)
2570 {
2571 	*ctrl = (void *)((uintptr_t)iq->cdesc +
2572 	    (iq->esize - sizeof(struct rsp_ctrl)));
2573 
2574 	return (((*ctrl)->u.type_gen >> S_RSPD_GEN) == iq->gen);
2575 }
2576 
2577 static inline void
2578 iq_next(struct sge_iq *iq)
2579 {
2580 	iq->cdesc = (void *) ((uintptr_t)iq->cdesc + iq->esize);
2581 	if (__predict_false(++iq->cidx == iq->qsize - 1)) {
2582 		iq->cidx = 0;
2583 		iq->gen ^= 1;
2584 		iq->cdesc = iq->desc;
2585 	}
2586 }
2587 
2588 #define FL_HW_IDX(x) ((x) >> 3)
2589 static inline void
2590 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
2591 {
2592 	int ndesc = fl->pending / 8;
2593 	uint32_t v;
2594 
2595 	if (FL_HW_IDX(fl->pidx) == FL_HW_IDX(fl->cidx))
2596 		ndesc--;	/* hold back one credit */
2597 
2598 	if (ndesc <= 0)
2599 		return;		/* nothing to do */
2600 
2601 	v = F_DBPRIO | V_QID(fl->cntxt_id) | V_PIDX(ndesc);
2602 	if (is_t5(sc))
2603 		v |= F_DBTYPE;
2604 
2605 	wmb();
2606 
2607 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL), v);
2608 	fl->pending -= ndesc * 8;
2609 }
2610 
2611 /*
2612  * Fill up the freelist by upto nbufs and maybe ring its doorbell.
2613  *
2614  * Returns non-zero to indicate that it should be added to the list of starving
2615  * freelists.
2616  */
2617 static int
2618 refill_fl(struct adapter *sc, struct sge_fl *fl, int nbufs)
2619 {
2620 	__be64 *d = &fl->desc[fl->pidx];
2621 	struct fl_sdesc *sd = &fl->sdesc[fl->pidx];
2622 	bus_dma_tag_t tag;
2623 	bus_addr_t pa;
2624 	caddr_t cl;
2625 	int rc;
2626 
2627 	FL_LOCK_ASSERT_OWNED(fl);
2628 
2629 	if (nbufs > fl->needed)
2630 		nbufs = fl->needed;
2631 
2632 	while (nbufs--) {
2633 
2634 		if (sd->cl != NULL) {
2635 
2636 			/*
2637 			 * This happens when a frame small enough to fit
2638 			 * entirely in an mbuf was received in cl last time.
2639 			 * We'd held on to cl and can reuse it now.  Note that
2640 			 * we reuse a cluster of the old size if fl->tag_idx is
2641 			 * no longer the same as sd->tag_idx.
2642 			 */
2643 
2644 			KASSERT(*d == sd->ba_tag,
2645 			    ("%s: recyling problem at pidx %d",
2646 			    __func__, fl->pidx));
2647 
2648 			d++;
2649 			goto recycled;
2650 		}
2651 
2652 
2653 		if (fl->tag_idx != sd->tag_idx) {
2654 			bus_dmamap_t map;
2655 			bus_dma_tag_t newtag = fl->tag[fl->tag_idx];
2656 			bus_dma_tag_t oldtag = fl->tag[sd->tag_idx];
2657 
2658 			/*
2659 			 * An MTU change can get us here.  Discard the old map
2660 			 * which was created with the old tag, but only if
2661 			 * we're able to get a new one.
2662 			 */
2663 			rc = bus_dmamap_create(newtag, 0, &map);
2664 			if (rc == 0) {
2665 				bus_dmamap_destroy(oldtag, sd->map);
2666 				sd->map = map;
2667 				sd->tag_idx = fl->tag_idx;
2668 			}
2669 		}
2670 
2671 		tag = fl->tag[sd->tag_idx];
2672 
2673 		cl = m_cljget(NULL, M_NOWAIT, FL_BUF_SIZE(sd->tag_idx));
2674 		if (cl == NULL)
2675 			break;
2676 
2677 		rc = bus_dmamap_load(tag, sd->map, cl, FL_BUF_SIZE(sd->tag_idx),
2678 		    oneseg_dma_callback, &pa, 0);
2679 		if (rc != 0 || pa == 0) {
2680 			fl->dmamap_failed++;
2681 			uma_zfree(FL_BUF_ZONE(sd->tag_idx), cl);
2682 			break;
2683 		}
2684 
2685 		sd->cl = cl;
2686 		*d++ = htobe64(pa | sd->tag_idx);
2687 
2688 #ifdef INVARIANTS
2689 		sd->ba_tag = htobe64(pa | sd->tag_idx);
2690 #endif
2691 
2692 recycled:
2693 		/* sd->m is never recycled, should always be NULL */
2694 		KASSERT(sd->m == NULL, ("%s: stray mbuf", __func__));
2695 
2696 		sd->m = m_gethdr(M_NOWAIT, MT_NOINIT);
2697 		if (sd->m == NULL)
2698 			break;
2699 
2700 		fl->pending++;
2701 		fl->needed--;
2702 		sd++;
2703 		if (++fl->pidx == fl->cap) {
2704 			fl->pidx = 0;
2705 			sd = fl->sdesc;
2706 			d = fl->desc;
2707 		}
2708 	}
2709 
2710 	if (fl->pending >= 8)
2711 		ring_fl_db(sc, fl);
2712 
2713 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
2714 }
2715 
2716 /*
2717  * Attempt to refill all starving freelists.
2718  */
2719 static void
2720 refill_sfl(void *arg)
2721 {
2722 	struct adapter *sc = arg;
2723 	struct sge_fl *fl, *fl_temp;
2724 
2725 	mtx_lock(&sc->sfl_lock);
2726 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
2727 		FL_LOCK(fl);
2728 		refill_fl(sc, fl, 64);
2729 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
2730 			TAILQ_REMOVE(&sc->sfl, fl, link);
2731 			fl->flags &= ~FL_STARVING;
2732 		}
2733 		FL_UNLOCK(fl);
2734 	}
2735 
2736 	if (!TAILQ_EMPTY(&sc->sfl))
2737 		callout_schedule(&sc->sfl_callout, hz / 5);
2738 	mtx_unlock(&sc->sfl_lock);
2739 }
2740 
2741 static int
2742 alloc_fl_sdesc(struct sge_fl *fl)
2743 {
2744 	struct fl_sdesc *sd;
2745 	bus_dma_tag_t tag;
2746 	int i, rc;
2747 
2748 	FL_LOCK_ASSERT_OWNED(fl);
2749 
2750 	fl->sdesc = malloc(fl->cap * sizeof(struct fl_sdesc), M_CXGBE,
2751 	    M_ZERO | M_WAITOK);
2752 
2753 	tag = fl->tag[fl->tag_idx];
2754 	sd = fl->sdesc;
2755 	for (i = 0; i < fl->cap; i++, sd++) {
2756 
2757 		sd->tag_idx = fl->tag_idx;
2758 		rc = bus_dmamap_create(tag, 0, &sd->map);
2759 		if (rc != 0)
2760 			goto failed;
2761 	}
2762 
2763 	return (0);
2764 failed:
2765 	while (--i >= 0) {
2766 		sd--;
2767 		bus_dmamap_destroy(tag, sd->map);
2768 		if (sd->m) {
2769 			m_init(sd->m, NULL, 0, M_NOWAIT, MT_DATA, 0);
2770 			m_free(sd->m);
2771 			sd->m = NULL;
2772 		}
2773 	}
2774 	KASSERT(sd == fl->sdesc, ("%s: EDOOFUS", __func__));
2775 
2776 	free(fl->sdesc, M_CXGBE);
2777 	fl->sdesc = NULL;
2778 
2779 	return (rc);
2780 }
2781 
2782 static void
2783 free_fl_sdesc(struct sge_fl *fl)
2784 {
2785 	struct fl_sdesc *sd;
2786 	int i;
2787 
2788 	FL_LOCK_ASSERT_OWNED(fl);
2789 
2790 	sd = fl->sdesc;
2791 	for (i = 0; i < fl->cap; i++, sd++) {
2792 
2793 		if (sd->m) {
2794 			m_init(sd->m, NULL, 0, M_NOWAIT, MT_DATA, 0);
2795 			m_free(sd->m);
2796 			sd->m = NULL;
2797 		}
2798 
2799 		if (sd->cl) {
2800 			bus_dmamap_unload(fl->tag[sd->tag_idx], sd->map);
2801 			uma_zfree(FL_BUF_ZONE(sd->tag_idx), sd->cl);
2802 			sd->cl = NULL;
2803 		}
2804 
2805 		bus_dmamap_destroy(fl->tag[sd->tag_idx], sd->map);
2806 	}
2807 
2808 	free(fl->sdesc, M_CXGBE);
2809 	fl->sdesc = NULL;
2810 }
2811 
2812 int
2813 t4_alloc_tx_maps(struct tx_maps *txmaps, bus_dma_tag_t tx_tag, int count,
2814     int flags)
2815 {
2816 	struct tx_map *txm;
2817 	int i, rc;
2818 
2819 	txmaps->map_total = txmaps->map_avail = count;
2820 	txmaps->map_cidx = txmaps->map_pidx = 0;
2821 
2822 	txmaps->maps = malloc(count * sizeof(struct tx_map), M_CXGBE,
2823 	    M_ZERO | flags);
2824 
2825 	txm = txmaps->maps;
2826 	for (i = 0; i < count; i++, txm++) {
2827 		rc = bus_dmamap_create(tx_tag, 0, &txm->map);
2828 		if (rc != 0)
2829 			goto failed;
2830 	}
2831 
2832 	return (0);
2833 failed:
2834 	while (--i >= 0) {
2835 		txm--;
2836 		bus_dmamap_destroy(tx_tag, txm->map);
2837 	}
2838 	KASSERT(txm == txmaps->maps, ("%s: EDOOFUS", __func__));
2839 
2840 	free(txmaps->maps, M_CXGBE);
2841 	txmaps->maps = NULL;
2842 
2843 	return (rc);
2844 }
2845 
2846 void
2847 t4_free_tx_maps(struct tx_maps *txmaps, bus_dma_tag_t tx_tag)
2848 {
2849 	struct tx_map *txm;
2850 	int i;
2851 
2852 	txm = txmaps->maps;
2853 	for (i = 0; i < txmaps->map_total; i++, txm++) {
2854 
2855 		if (txm->m) {
2856 			bus_dmamap_unload(tx_tag, txm->map);
2857 			m_freem(txm->m);
2858 			txm->m = NULL;
2859 		}
2860 
2861 		bus_dmamap_destroy(tx_tag, txm->map);
2862 	}
2863 
2864 	free(txmaps->maps, M_CXGBE);
2865 	txmaps->maps = NULL;
2866 }
2867 
2868 /*
2869  * We'll do immediate data tx for non-TSO, but only when not coalescing.  We're
2870  * willing to use upto 2 hardware descriptors which means a maximum of 96 bytes
2871  * of immediate data.
2872  */
2873 #define IMM_LEN ( \
2874       2 * EQ_ESIZE \
2875     - sizeof(struct fw_eth_tx_pkt_wr) \
2876     - sizeof(struct cpl_tx_pkt_core))
2877 
2878 /*
2879  * Returns non-zero on failure, no need to cleanup anything in that case.
2880  *
2881  * Note 1: We always try to defrag the mbuf if required and return EFBIG only
2882  * if the resulting chain still won't fit in a tx descriptor.
2883  *
2884  * Note 2: We'll pullup the mbuf chain if TSO is requested and the first mbuf
2885  * does not have the TCP header in it.
2886  */
2887 static int
2888 get_pkt_sgl(struct sge_txq *txq, struct mbuf **fp, struct sgl *sgl,
2889     int sgl_only)
2890 {
2891 	struct mbuf *m = *fp;
2892 	struct tx_maps *txmaps;
2893 	struct tx_map *txm;
2894 	int rc, defragged = 0, n;
2895 
2896 	TXQ_LOCK_ASSERT_OWNED(txq);
2897 
2898 	if (m->m_pkthdr.tso_segsz)
2899 		sgl_only = 1;	/* Do not allow immediate data with LSO */
2900 
2901 start:	sgl->nsegs = 0;
2902 
2903 	if (m->m_pkthdr.len <= IMM_LEN && !sgl_only)
2904 		return (0);	/* nsegs = 0 tells caller to use imm. tx */
2905 
2906 	txmaps = &txq->txmaps;
2907 	if (txmaps->map_avail == 0) {
2908 		txq->no_dmamap++;
2909 		return (ENOMEM);
2910 	}
2911 	txm = &txmaps->maps[txmaps->map_pidx];
2912 
2913 	if (m->m_pkthdr.tso_segsz && m->m_len < 50) {
2914 		*fp = m_pullup(m, 50);
2915 		m = *fp;
2916 		if (m == NULL)
2917 			return (ENOBUFS);
2918 	}
2919 
2920 	rc = bus_dmamap_load_mbuf_sg(txq->tx_tag, txm->map, m, sgl->seg,
2921 	    &sgl->nsegs, BUS_DMA_NOWAIT);
2922 	if (rc == EFBIG && defragged == 0) {
2923 		m = m_defrag(m, M_NOWAIT);
2924 		if (m == NULL)
2925 			return (EFBIG);
2926 
2927 		defragged = 1;
2928 		*fp = m;
2929 		goto start;
2930 	}
2931 	if (rc != 0)
2932 		return (rc);
2933 
2934 	txm->m = m;
2935 	txmaps->map_avail--;
2936 	if (++txmaps->map_pidx == txmaps->map_total)
2937 		txmaps->map_pidx = 0;
2938 
2939 	KASSERT(sgl->nsegs > 0 && sgl->nsegs <= TX_SGL_SEGS,
2940 	    ("%s: bad DMA mapping (%d segments)", __func__, sgl->nsegs));
2941 
2942 	/*
2943 	 * Store the # of flits required to hold this frame's SGL in nflits.  An
2944 	 * SGL has a (ULPTX header + len0, addr0) tuple optionally followed by
2945 	 * multiple (len0 + len1, addr0, addr1) tuples.  If addr1 is not used
2946 	 * then len1 must be set to 0.
2947 	 */
2948 	n = sgl->nsegs - 1;
2949 	sgl->nflits = (3 * n) / 2 + (n & 1) + 2;
2950 
2951 	return (0);
2952 }
2953 
2954 
2955 /*
2956  * Releases all the txq resources used up in the specified sgl.
2957  */
2958 static int
2959 free_pkt_sgl(struct sge_txq *txq, struct sgl *sgl)
2960 {
2961 	struct tx_maps *txmaps;
2962 	struct tx_map *txm;
2963 
2964 	TXQ_LOCK_ASSERT_OWNED(txq);
2965 
2966 	if (sgl->nsegs == 0)
2967 		return (0);	/* didn't use any map */
2968 
2969 	txmaps = &txq->txmaps;
2970 
2971 	/* 1 pkt uses exactly 1 map, back it out */
2972 
2973 	txmaps->map_avail++;
2974 	if (txmaps->map_pidx > 0)
2975 		txmaps->map_pidx--;
2976 	else
2977 		txmaps->map_pidx = txmaps->map_total - 1;
2978 
2979 	txm = &txmaps->maps[txmaps->map_pidx];
2980 	bus_dmamap_unload(txq->tx_tag, txm->map);
2981 	txm->m = NULL;
2982 
2983 	return (0);
2984 }
2985 
2986 static int
2987 write_txpkt_wr(struct port_info *pi, struct sge_txq *txq, struct mbuf *m,
2988     struct sgl *sgl)
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, pktlen;
2996 	struct tx_sdesc *txsd;
2997 	caddr_t dst;
2998 
2999 	TXQ_LOCK_ASSERT_OWNED(txq);
3000 
3001 	pktlen = m->m_pkthdr.len;
3002 
3003 	/*
3004 	 * Do we have enough flits to send this frame out?
3005 	 */
3006 	ctrl = sizeof(struct cpl_tx_pkt_core);
3007 	if (m->m_pkthdr.tso_segsz) {
3008 		nflits = TXPKT_LSO_WR_HDR;
3009 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
3010 	} else
3011 		nflits = TXPKT_WR_HDR;
3012 	if (sgl->nsegs > 0)
3013 		nflits += sgl->nflits;
3014 	else {
3015 		nflits += howmany(pktlen, 8);
3016 		ctrl += pktlen;
3017 	}
3018 	ndesc = howmany(nflits, 8);
3019 	if (ndesc > eq->avail)
3020 		return (ENOMEM);
3021 
3022 	/* Firmware work request header */
3023 	wr = (void *)&eq->desc[eq->pidx];
3024 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
3025 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
3026 	ctrl = V_FW_WR_LEN16(howmany(nflits, 2));
3027 	if (eq->avail == ndesc) {
3028 		if (!(eq->flags & EQ_CRFLUSHED)) {
3029 			ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ;
3030 			eq->flags |= EQ_CRFLUSHED;
3031 		}
3032 		eq->flags |= EQ_STALLED;
3033 	}
3034 
3035 	wr->equiq_to_len16 = htobe32(ctrl);
3036 	wr->r3 = 0;
3037 
3038 	if (m->m_pkthdr.tso_segsz) {
3039 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
3040 		struct ether_header *eh;
3041 		void *l3hdr;
3042 #if defined(INET) || defined(INET6)
3043 		struct tcphdr *tcp;
3044 #endif
3045 		uint16_t eh_type;
3046 
3047 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
3048 		    F_LSO_LAST_SLICE;
3049 
3050 		eh = mtod(m, struct ether_header *);
3051 		eh_type = ntohs(eh->ether_type);
3052 		if (eh_type == ETHERTYPE_VLAN) {
3053 			struct ether_vlan_header *evh = (void *)eh;
3054 
3055 			ctrl |= V_LSO_ETHHDR_LEN(1);
3056 			l3hdr = evh + 1;
3057 			eh_type = ntohs(evh->evl_proto);
3058 		} else
3059 			l3hdr = eh + 1;
3060 
3061 		switch (eh_type) {
3062 #ifdef INET6
3063 		case ETHERTYPE_IPV6:
3064 		{
3065 			struct ip6_hdr *ip6 = l3hdr;
3066 
3067 			/*
3068 			 * XXX-BZ For now we do not pretend to support
3069 			 * IPv6 extension headers.
3070 			 */
3071 			KASSERT(ip6->ip6_nxt == IPPROTO_TCP, ("%s: CSUM_TSO "
3072 			    "with ip6_nxt != TCP: %u", __func__, ip6->ip6_nxt));
3073 			tcp = (struct tcphdr *)(ip6 + 1);
3074 			ctrl |= F_LSO_IPV6;
3075 			ctrl |= V_LSO_IPHDR_LEN(sizeof(*ip6) >> 2) |
3076 			    V_LSO_TCPHDR_LEN(tcp->th_off);
3077 			break;
3078 		}
3079 #endif
3080 #ifdef INET
3081 		case ETHERTYPE_IP:
3082 		{
3083 			struct ip *ip = l3hdr;
3084 
3085 			tcp = (void *)((uintptr_t)ip + ip->ip_hl * 4);
3086 			ctrl |= V_LSO_IPHDR_LEN(ip->ip_hl) |
3087 			    V_LSO_TCPHDR_LEN(tcp->th_off);
3088 			break;
3089 		}
3090 #endif
3091 		default:
3092 			panic("%s: CSUM_TSO but no supported IP version "
3093 			    "(0x%04x)", __func__, eh_type);
3094 		}
3095 
3096 		lso->lso_ctrl = htobe32(ctrl);
3097 		lso->ipid_ofst = htobe16(0);
3098 		lso->mss = htobe16(m->m_pkthdr.tso_segsz);
3099 		lso->seqno_offset = htobe32(0);
3100 		lso->len = htobe32(pktlen);
3101 
3102 		cpl = (void *)(lso + 1);
3103 
3104 		txq->tso_wrs++;
3105 	} else
3106 		cpl = (void *)(wr + 1);
3107 
3108 	/* Checksum offload */
3109 	ctrl1 = 0;
3110 	if (!(m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO)))
3111 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
3112 	if (!(m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
3113 	    CSUM_TCP_IPV6 | CSUM_TSO)))
3114 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
3115 	if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
3116 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
3117 		txq->txcsum++;	/* some hardware assistance provided */
3118 
3119 	/* VLAN tag insertion */
3120 	if (m->m_flags & M_VLANTAG) {
3121 		ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
3122 		txq->vlan_insertion++;
3123 	}
3124 
3125 	/* CPL header */
3126 	cpl->ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3127 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf));
3128 	cpl->pack = 0;
3129 	cpl->len = htobe16(pktlen);
3130 	cpl->ctrl1 = htobe64(ctrl1);
3131 
3132 	/* Software descriptor */
3133 	txsd = &txq->sdesc[eq->pidx];
3134 	txsd->desc_used = ndesc;
3135 
3136 	eq->pending += ndesc;
3137 	eq->avail -= ndesc;
3138 	eq->pidx += ndesc;
3139 	if (eq->pidx >= eq->cap)
3140 		eq->pidx -= eq->cap;
3141 
3142 	/* SGL */
3143 	dst = (void *)(cpl + 1);
3144 	if (sgl->nsegs > 0) {
3145 		txsd->credits = 1;
3146 		txq->sgl_wrs++;
3147 		write_sgl_to_txd(eq, sgl, &dst);
3148 	} else {
3149 		txsd->credits = 0;
3150 		txq->imm_wrs++;
3151 		for (; m; m = m->m_next) {
3152 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
3153 #ifdef INVARIANTS
3154 			pktlen -= m->m_len;
3155 #endif
3156 		}
3157 #ifdef INVARIANTS
3158 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
3159 #endif
3160 
3161 	}
3162 
3163 	txq->txpkt_wrs++;
3164 	return (0);
3165 }
3166 
3167 /*
3168  * Returns 0 to indicate that m has been accepted into a coalesced tx work
3169  * request.  It has either been folded into txpkts or txpkts was flushed and m
3170  * has started a new coalesced work request (as the first frame in a fresh
3171  * txpkts).
3172  *
3173  * Returns non-zero to indicate a failure - caller is responsible for
3174  * transmitting m, if there was anything in txpkts it has been flushed.
3175  */
3176 static int
3177 add_to_txpkts(struct port_info *pi, struct sge_txq *txq, struct txpkts *txpkts,
3178     struct mbuf *m, struct sgl *sgl)
3179 {
3180 	struct sge_eq *eq = &txq->eq;
3181 	int can_coalesce;
3182 	struct tx_sdesc *txsd;
3183 	int flits;
3184 
3185 	TXQ_LOCK_ASSERT_OWNED(txq);
3186 
3187 	KASSERT(sgl->nsegs, ("%s: can't coalesce imm data", __func__));
3188 
3189 	if (txpkts->npkt > 0) {
3190 		flits = TXPKTS_PKT_HDR + sgl->nflits;
3191 		can_coalesce = m->m_pkthdr.tso_segsz == 0 &&
3192 		    txpkts->nflits + flits <= TX_WR_FLITS &&
3193 		    txpkts->nflits + flits <= eq->avail * 8 &&
3194 		    txpkts->plen + m->m_pkthdr.len < 65536;
3195 
3196 		if (can_coalesce) {
3197 			txpkts->npkt++;
3198 			txpkts->nflits += flits;
3199 			txpkts->plen += m->m_pkthdr.len;
3200 
3201 			txsd = &txq->sdesc[eq->pidx];
3202 			txsd->credits++;
3203 
3204 			return (0);
3205 		}
3206 
3207 		/*
3208 		 * Couldn't coalesce m into txpkts.  The first order of business
3209 		 * is to send txpkts on its way.  Then we'll revisit m.
3210 		 */
3211 		write_txpkts_wr(txq, txpkts);
3212 	}
3213 
3214 	/*
3215 	 * Check if we can start a new coalesced tx work request with m as
3216 	 * the first packet in it.
3217 	 */
3218 
3219 	KASSERT(txpkts->npkt == 0, ("%s: txpkts not empty", __func__));
3220 
3221 	flits = TXPKTS_WR_HDR + sgl->nflits;
3222 	can_coalesce = m->m_pkthdr.tso_segsz == 0 &&
3223 	    flits <= eq->avail * 8 && flits <= TX_WR_FLITS;
3224 
3225 	if (can_coalesce == 0)
3226 		return (EINVAL);
3227 
3228 	/*
3229 	 * Start a fresh coalesced tx WR with m as the first frame in it.
3230 	 */
3231 	txpkts->npkt = 1;
3232 	txpkts->nflits = flits;
3233 	txpkts->flitp = &eq->desc[eq->pidx].flit[2];
3234 	txpkts->plen = m->m_pkthdr.len;
3235 
3236 	txsd = &txq->sdesc[eq->pidx];
3237 	txsd->credits = 1;
3238 
3239 	return (0);
3240 }
3241 
3242 /*
3243  * Note that write_txpkts_wr can never run out of hardware descriptors (but
3244  * write_txpkt_wr can).  add_to_txpkts ensures that a frame is accepted for
3245  * coalescing only if sufficient hardware descriptors are available.
3246  */
3247 static void
3248 write_txpkts_wr(struct sge_txq *txq, struct txpkts *txpkts)
3249 {
3250 	struct sge_eq *eq = &txq->eq;
3251 	struct fw_eth_tx_pkts_wr *wr;
3252 	struct tx_sdesc *txsd;
3253 	uint32_t ctrl;
3254 	int ndesc;
3255 
3256 	TXQ_LOCK_ASSERT_OWNED(txq);
3257 
3258 	ndesc = howmany(txpkts->nflits, 8);
3259 
3260 	wr = (void *)&eq->desc[eq->pidx];
3261 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
3262 	ctrl = V_FW_WR_LEN16(howmany(txpkts->nflits, 2));
3263 	if (eq->avail == ndesc) {
3264 		if (!(eq->flags & EQ_CRFLUSHED)) {
3265 			ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ;
3266 			eq->flags |= EQ_CRFLUSHED;
3267 		}
3268 		eq->flags |= EQ_STALLED;
3269 	}
3270 	wr->equiq_to_len16 = htobe32(ctrl);
3271 	wr->plen = htobe16(txpkts->plen);
3272 	wr->npkt = txpkts->npkt;
3273 	wr->r3 = wr->type = 0;
3274 
3275 	/* Everything else already written */
3276 
3277 	txsd = &txq->sdesc[eq->pidx];
3278 	txsd->desc_used = ndesc;
3279 
3280 	KASSERT(eq->avail >= ndesc, ("%s: out of descriptors", __func__));
3281 
3282 	eq->pending += ndesc;
3283 	eq->avail -= ndesc;
3284 	eq->pidx += ndesc;
3285 	if (eq->pidx >= eq->cap)
3286 		eq->pidx -= eq->cap;
3287 
3288 	txq->txpkts_pkts += txpkts->npkt;
3289 	txq->txpkts_wrs++;
3290 	txpkts->npkt = 0;	/* emptied */
3291 }
3292 
3293 static inline void
3294 write_ulp_cpl_sgl(struct port_info *pi, struct sge_txq *txq,
3295     struct txpkts *txpkts, struct mbuf *m, struct sgl *sgl)
3296 {
3297 	struct ulp_txpkt *ulpmc;
3298 	struct ulptx_idata *ulpsc;
3299 	struct cpl_tx_pkt_core *cpl;
3300 	struct sge_eq *eq = &txq->eq;
3301 	uintptr_t flitp, start, end;
3302 	uint64_t ctrl;
3303 	caddr_t dst;
3304 
3305 	KASSERT(txpkts->npkt > 0, ("%s: txpkts is empty", __func__));
3306 
3307 	start = (uintptr_t)eq->desc;
3308 	end = (uintptr_t)eq->spg;
3309 
3310 	/* Checksum offload */
3311 	ctrl = 0;
3312 	if (!(m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO)))
3313 		ctrl |= F_TXPKT_IPCSUM_DIS;
3314 	if (!(m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
3315 	    CSUM_TCP_IPV6 | CSUM_TSO)))
3316 		ctrl |= F_TXPKT_L4CSUM_DIS;
3317 	if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
3318 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
3319 		txq->txcsum++;	/* some hardware assistance provided */
3320 
3321 	/* VLAN tag insertion */
3322 	if (m->m_flags & M_VLANTAG) {
3323 		ctrl |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
3324 		txq->vlan_insertion++;
3325 	}
3326 
3327 	/*
3328 	 * The previous packet's SGL must have ended at a 16 byte boundary (this
3329 	 * is required by the firmware/hardware).  It follows that flitp cannot
3330 	 * wrap around between the ULPTX master command and ULPTX subcommand (8
3331 	 * bytes each), and that it can not wrap around in the middle of the
3332 	 * cpl_tx_pkt_core either.
3333 	 */
3334 	flitp = (uintptr_t)txpkts->flitp;
3335 	KASSERT((flitp & 0xf) == 0,
3336 	    ("%s: last SGL did not end at 16 byte boundary: %p",
3337 	    __func__, txpkts->flitp));
3338 
3339 	/* ULP master command */
3340 	ulpmc = (void *)flitp;
3341 	ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0) |
3342 	    V_ULP_TXPKT_FID(eq->iqid));
3343 	ulpmc->len = htonl(howmany(sizeof(*ulpmc) + sizeof(*ulpsc) +
3344 	    sizeof(*cpl) + 8 * sgl->nflits, 16));
3345 
3346 	/* ULP subcommand */
3347 	ulpsc = (void *)(ulpmc + 1);
3348 	ulpsc->cmd_more = htobe32(V_ULPTX_CMD((u32)ULP_TX_SC_IMM) |
3349 	    F_ULP_TX_SC_MORE);
3350 	ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
3351 
3352 	flitp += sizeof(*ulpmc) + sizeof(*ulpsc);
3353 	if (flitp == end)
3354 		flitp = start;
3355 
3356 	/* CPL_TX_PKT */
3357 	cpl = (void *)flitp;
3358 	cpl->ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3359 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf));
3360 	cpl->pack = 0;
3361 	cpl->len = htobe16(m->m_pkthdr.len);
3362 	cpl->ctrl1 = htobe64(ctrl);
3363 
3364 	flitp += sizeof(*cpl);
3365 	if (flitp == end)
3366 		flitp = start;
3367 
3368 	/* SGL for this frame */
3369 	dst = (caddr_t)flitp;
3370 	txpkts->nflits += write_sgl_to_txd(eq, sgl, &dst);
3371 	txpkts->flitp = (void *)dst;
3372 
3373 	KASSERT(((uintptr_t)dst & 0xf) == 0,
3374 	    ("%s: SGL ends at %p (not a 16 byte boundary)", __func__, dst));
3375 }
3376 
3377 /*
3378  * If the SGL ends on an address that is not 16 byte aligned, this function will
3379  * add a 0 filled flit at the end.  It returns 1 in that case.
3380  */
3381 static int
3382 write_sgl_to_txd(struct sge_eq *eq, struct sgl *sgl, caddr_t *to)
3383 {
3384 	__be64 *flitp, *end;
3385 	struct ulptx_sgl *usgl;
3386 	bus_dma_segment_t *seg;
3387 	int i, padded;
3388 
3389 	KASSERT(sgl->nsegs > 0 && sgl->nflits > 0,
3390 	    ("%s: bad SGL - nsegs=%d, nflits=%d",
3391 	    __func__, sgl->nsegs, sgl->nflits));
3392 
3393 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
3394 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
3395 
3396 	flitp = (__be64 *)(*to);
3397 	end = flitp + sgl->nflits;
3398 	seg = &sgl->seg[0];
3399 	usgl = (void *)flitp;
3400 
3401 	/*
3402 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
3403 	 * ring, so we're at least 16 bytes away from the status page.  There is
3404 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
3405 	 */
3406 
3407 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
3408 	    V_ULPTX_NSGE(sgl->nsegs));
3409 	usgl->len0 = htobe32(seg->ds_len);
3410 	usgl->addr0 = htobe64(seg->ds_addr);
3411 	seg++;
3412 
3413 	if ((uintptr_t)end <= (uintptr_t)eq->spg) {
3414 
3415 		/* Won't wrap around at all */
3416 
3417 		for (i = 0; i < sgl->nsegs - 1; i++, seg++) {
3418 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ds_len);
3419 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ds_addr);
3420 		}
3421 		if (i & 1)
3422 			usgl->sge[i / 2].len[1] = htobe32(0);
3423 	} else {
3424 
3425 		/* Will wrap somewhere in the rest of the SGL */
3426 
3427 		/* 2 flits already written, write the rest flit by flit */
3428 		flitp = (void *)(usgl + 1);
3429 		for (i = 0; i < sgl->nflits - 2; i++) {
3430 			if ((uintptr_t)flitp == (uintptr_t)eq->spg)
3431 				flitp = (void *)eq->desc;
3432 			*flitp++ = get_flit(seg, sgl->nsegs - 1, i);
3433 		}
3434 		end = flitp;
3435 	}
3436 
3437 	if ((uintptr_t)end & 0xf) {
3438 		*(uint64_t *)end = 0;
3439 		end++;
3440 		padded = 1;
3441 	} else
3442 		padded = 0;
3443 
3444 	if ((uintptr_t)end == (uintptr_t)eq->spg)
3445 		*to = (void *)eq->desc;
3446 	else
3447 		*to = (void *)end;
3448 
3449 	return (padded);
3450 }
3451 
3452 static inline void
3453 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
3454 {
3455 	if (__predict_true((uintptr_t)(*to) + len <= (uintptr_t)eq->spg)) {
3456 		bcopy(from, *to, len);
3457 		(*to) += len;
3458 	} else {
3459 		int portion = (uintptr_t)eq->spg - (uintptr_t)(*to);
3460 
3461 		bcopy(from, *to, portion);
3462 		from += portion;
3463 		portion = len - portion;	/* remaining */
3464 		bcopy(from, (void *)eq->desc, portion);
3465 		(*to) = (caddr_t)eq->desc + portion;
3466 	}
3467 }
3468 
3469 static inline void
3470 ring_eq_db(struct adapter *sc, struct sge_eq *eq)
3471 {
3472 	u_int db, pending;
3473 
3474 	db = eq->doorbells;
3475 	pending = eq->pending;
3476 	if (pending > 1)
3477 		clrbit(&db, DOORBELL_WCWR);
3478 	eq->pending = 0;
3479 	wmb();
3480 
3481 	switch (ffs(db) - 1) {
3482 	case DOORBELL_UDB:
3483 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(pending));
3484 		return;
3485 
3486 	case DOORBELL_WCWR: {
3487 		volatile uint64_t *dst, *src;
3488 		int i;
3489 
3490 		/*
3491 		 * Queues whose 128B doorbell segment fits in the page do not
3492 		 * use relative qid (udb_qid is always 0).  Only queues with
3493 		 * doorbell segments can do WCWR.
3494 		 */
3495 		KASSERT(eq->udb_qid == 0 && pending == 1,
3496 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
3497 		    __func__, eq->doorbells, pending, eq->pidx, eq));
3498 
3499 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
3500 		    UDBS_DB_OFFSET);
3501 		i = eq->pidx ? eq->pidx - 1 : eq->cap - 1;
3502 		src = (void *)&eq->desc[i];
3503 		while (src != (void *)&eq->desc[i + 1])
3504 			*dst++ = *src++;
3505 		wmb();
3506 		return;
3507 	}
3508 
3509 	case DOORBELL_UDBWC:
3510 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(pending));
3511 		wmb();
3512 		return;
3513 
3514 	case DOORBELL_KDB:
3515 		t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
3516 		    V_QID(eq->cntxt_id) | V_PIDX(pending));
3517 		return;
3518 	}
3519 }
3520 
3521 static inline int
3522 reclaimable(struct sge_eq *eq)
3523 {
3524 	unsigned int cidx;
3525 
3526 	cidx = eq->spg->cidx;	/* stable snapshot */
3527 	cidx = be16toh(cidx);
3528 
3529 	if (cidx >= eq->cidx)
3530 		return (cidx - eq->cidx);
3531 	else
3532 		return (cidx + eq->cap - eq->cidx);
3533 }
3534 
3535 /*
3536  * There are "can_reclaim" tx descriptors ready to be reclaimed.  Reclaim as
3537  * many as possible but stop when there are around "n" mbufs to free.
3538  *
3539  * The actual number reclaimed is provided as the return value.
3540  */
3541 static int
3542 reclaim_tx_descs(struct sge_txq *txq, int can_reclaim, int n)
3543 {
3544 	struct tx_sdesc *txsd;
3545 	struct tx_maps *txmaps;
3546 	struct tx_map *txm;
3547 	unsigned int reclaimed, maps;
3548 	struct sge_eq *eq = &txq->eq;
3549 
3550 	TXQ_LOCK_ASSERT_OWNED(txq);
3551 
3552 	if (can_reclaim == 0)
3553 		can_reclaim = reclaimable(eq);
3554 
3555 	maps = reclaimed = 0;
3556 	while (can_reclaim && maps < n) {
3557 		int ndesc;
3558 
3559 		txsd = &txq->sdesc[eq->cidx];
3560 		ndesc = txsd->desc_used;
3561 
3562 		/* Firmware doesn't return "partial" credits. */
3563 		KASSERT(can_reclaim >= ndesc,
3564 		    ("%s: unexpected number of credits: %d, %d",
3565 		    __func__, can_reclaim, ndesc));
3566 
3567 		maps += txsd->credits;
3568 
3569 		reclaimed += ndesc;
3570 		can_reclaim -= ndesc;
3571 
3572 		eq->cidx += ndesc;
3573 		if (__predict_false(eq->cidx >= eq->cap))
3574 			eq->cidx -= eq->cap;
3575 	}
3576 
3577 	txmaps = &txq->txmaps;
3578 	txm = &txmaps->maps[txmaps->map_cidx];
3579 	if (maps)
3580 		prefetch(txm->m);
3581 
3582 	eq->avail += reclaimed;
3583 	KASSERT(eq->avail < eq->cap,	/* avail tops out at (cap - 1) */
3584 	    ("%s: too many descriptors available", __func__));
3585 
3586 	txmaps->map_avail += maps;
3587 	KASSERT(txmaps->map_avail <= txmaps->map_total,
3588 	    ("%s: too many maps available", __func__));
3589 
3590 	while (maps--) {
3591 		struct tx_map *next;
3592 
3593 		next = txm + 1;
3594 		if (__predict_false(txmaps->map_cidx + 1 == txmaps->map_total))
3595 			next = txmaps->maps;
3596 		prefetch(next->m);
3597 
3598 		bus_dmamap_unload(txq->tx_tag, txm->map);
3599 		m_freem(txm->m);
3600 		txm->m = NULL;
3601 
3602 		txm = next;
3603 		if (__predict_false(++txmaps->map_cidx == txmaps->map_total))
3604 			txmaps->map_cidx = 0;
3605 	}
3606 
3607 	return (reclaimed);
3608 }
3609 
3610 static void
3611 write_eqflush_wr(struct sge_eq *eq)
3612 {
3613 	struct fw_eq_flush_wr *wr;
3614 
3615 	EQ_LOCK_ASSERT_OWNED(eq);
3616 	KASSERT(eq->avail > 0, ("%s: no descriptors left.", __func__));
3617 	KASSERT(!(eq->flags & EQ_CRFLUSHED), ("%s: flushed already", __func__));
3618 
3619 	wr = (void *)&eq->desc[eq->pidx];
3620 	bzero(wr, sizeof(*wr));
3621 	wr->opcode = FW_EQ_FLUSH_WR;
3622 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(sizeof(*wr) / 16) |
3623 	    F_FW_WR_EQUEQ | F_FW_WR_EQUIQ);
3624 
3625 	eq->flags |= (EQ_CRFLUSHED | EQ_STALLED);
3626 	eq->pending++;
3627 	eq->avail--;
3628 	if (++eq->pidx == eq->cap)
3629 		eq->pidx = 0;
3630 }
3631 
3632 static __be64
3633 get_flit(bus_dma_segment_t *sgl, int nsegs, int idx)
3634 {
3635 	int i = (idx / 3) * 2;
3636 
3637 	switch (idx % 3) {
3638 	case 0: {
3639 		__be64 rc;
3640 
3641 		rc = htobe32(sgl[i].ds_len);
3642 		if (i + 1 < nsegs)
3643 			rc |= (uint64_t)htobe32(sgl[i + 1].ds_len) << 32;
3644 
3645 		return (rc);
3646 	}
3647 	case 1:
3648 		return htobe64(sgl[i].ds_addr);
3649 	case 2:
3650 		return htobe64(sgl[i + 1].ds_addr);
3651 	}
3652 
3653 	return (0);
3654 }
3655 
3656 static void
3657 set_fl_tag_idx(struct sge_fl *fl, int bufsize)
3658 {
3659 	int i;
3660 
3661 	for (i = 0; i < FL_BUF_SIZES - 1; i++) {
3662 		if (FL_BUF_SIZE(i) >= bufsize)
3663 			break;
3664 	}
3665 
3666 	fl->tag_idx = i;
3667 }
3668 
3669 static void
3670 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
3671 {
3672 	mtx_lock(&sc->sfl_lock);
3673 	FL_LOCK(fl);
3674 	if ((fl->flags & FL_DOOMED) == 0) {
3675 		fl->flags |= FL_STARVING;
3676 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
3677 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
3678 	}
3679 	FL_UNLOCK(fl);
3680 	mtx_unlock(&sc->sfl_lock);
3681 }
3682 
3683 static int
3684 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
3685     struct mbuf *m)
3686 {
3687 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
3688 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
3689 	struct adapter *sc = iq->adapter;
3690 	struct sge *s = &sc->sge;
3691 	struct sge_eq *eq;
3692 
3693 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
3694 	    rss->opcode));
3695 
3696 	eq = s->eqmap[qid - s->eq_start];
3697 	EQ_LOCK(eq);
3698 	KASSERT(eq->flags & EQ_CRFLUSHED,
3699 	    ("%s: unsolicited egress update", __func__));
3700 	eq->flags &= ~EQ_CRFLUSHED;
3701 	eq->egr_update++;
3702 
3703 	if (__predict_false(eq->flags & EQ_DOOMED))
3704 		wakeup_one(eq);
3705 	else if (eq->flags & EQ_STALLED && can_resume_tx(eq))
3706 		taskqueue_enqueue(sc->tq[eq->tx_chan], &eq->tx_task);
3707 	EQ_UNLOCK(eq);
3708 
3709 	return (0);
3710 }
3711 
3712 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
3713 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
3714     offsetof(struct cpl_fw6_msg, data));
3715 
3716 static int
3717 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
3718 {
3719 	struct adapter *sc = iq->adapter;
3720 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
3721 
3722 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
3723 	    rss->opcode));
3724 
3725 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
3726 		const struct rss_header *rss2;
3727 
3728 		rss2 = (const struct rss_header *)&cpl->data[0];
3729 		return (sc->cpl_handler[rss2->opcode](iq, rss2, m));
3730 	}
3731 
3732 	return (sc->fw_msg_handler[cpl->type](sc, &cpl->data[0]));
3733 }
3734 
3735 static int
3736 sysctl_uint16(SYSCTL_HANDLER_ARGS)
3737 {
3738 	uint16_t *id = arg1;
3739 	int i = *id;
3740 
3741 	return sysctl_handle_int(oidp, &i, 0, req);
3742 }
3743