xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision a18eacbefdfa1085ca3db829e86ece78cd416493)
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/eventhandler.h>
36 #include <sys/mbuf.h>
37 #include <sys/socket.h>
38 #include <sys/kernel.h>
39 #include <sys/kdb.h>
40 #include <sys/malloc.h>
41 #include <sys/queue.h>
42 #include <sys/taskqueue.h>
43 #include <sys/time.h>
44 #include <sys/sysctl.h>
45 #include <sys/smp.h>
46 #include <net/bpf.h>
47 #include <net/ethernet.h>
48 #include <net/if.h>
49 #include <net/if_vlan_var.h>
50 #include <netinet/in.h>
51 #include <netinet/ip.h>
52 #include <netinet/ip6.h>
53 #include <netinet/tcp.h>
54 #include <machine/md_var.h>
55 
56 #include "common/common.h"
57 #include "common/t4_regs.h"
58 #include "common/t4_regs_values.h"
59 #include "common/t4_msg.h"
60 
61 #ifdef T4_PKT_TIMESTAMP
62 #define RX_COPY_THRESHOLD (MINCLSIZE - 8)
63 #else
64 #define RX_COPY_THRESHOLD MINCLSIZE
65 #endif
66 
67 /*
68  * Ethernet frames are DMA'd at this byte offset into the freelist buffer.
69  * 0-7 are valid values.
70  */
71 static int fl_pktshift = 2;
72 TUNABLE_INT("hw.cxgbe.fl_pktshift", &fl_pktshift);
73 
74 /*
75  * Pad ethernet payload up to this boundary.
76  * -1: driver should figure out a good value.
77  *  0: disable padding.
78  *  Any power of 2 from 32 to 4096 (both inclusive) is also a valid value.
79  */
80 static int fl_pad = -1;
81 TUNABLE_INT("hw.cxgbe.fl_pad", &fl_pad);
82 
83 /*
84  * Status page length.
85  * -1: driver should figure out a good value.
86  *  64 or 128 are the only other valid values.
87  */
88 static int spg_len = -1;
89 TUNABLE_INT("hw.cxgbe.spg_len", &spg_len);
90 
91 /*
92  * Congestion drops.
93  * -1: no congestion feedback (not recommended).
94  *  0: backpressure the channel instead of dropping packets right away.
95  *  1: no backpressure, drop packets for the congested queue immediately.
96  */
97 static int cong_drop = 0;
98 TUNABLE_INT("hw.cxgbe.cong_drop", &cong_drop);
99 
100 /*
101  * Deliver multiple frames in the same free list buffer if they fit.
102  * -1: let the driver decide whether to enable buffer packing or not.
103  *  0: disable buffer packing.
104  *  1: enable buffer packing.
105  */
106 static int buffer_packing = -1;
107 TUNABLE_INT("hw.cxgbe.buffer_packing", &buffer_packing);
108 
109 /*
110  * Start next frame in a packed buffer at this boundary.
111  * -1: driver should figure out a good value.
112  * T4:
113  * ---
114  * if fl_pad != 0
115  * 	value specified here will be overridden by fl_pad.
116  * else
117  * 	power of 2 from 32 to 4096 (both inclusive) is a valid value here.
118  * T5:
119  * ---
120  * 16, or a power of 2 from 64 to 4096 (both inclusive) is a valid value.
121  */
122 static int fl_pack = -1;
123 static int t4_fl_pack;
124 static int t5_fl_pack;
125 TUNABLE_INT("hw.cxgbe.fl_pack", &fl_pack);
126 
127 /* Used to track coalesced tx work request */
128 struct txpkts {
129 	uint64_t *flitp;	/* ptr to flit where next pkt should start */
130 	uint8_t npkt;		/* # of packets in this work request */
131 	uint8_t nflits;		/* # of flits used by this work request */
132 	uint16_t plen;		/* total payload (sum of all packets) */
133 };
134 
135 /* A packet's SGL.  This + m_pkthdr has all info needed for tx */
136 struct sgl {
137 	int nsegs;		/* # of segments in the SGL, 0 means imm. tx */
138 	int nflits;		/* # of flits needed for the SGL */
139 	bus_dma_segment_t seg[TX_SGL_SEGS];
140 };
141 
142 static int service_iq(struct sge_iq *, int);
143 static struct mbuf *get_fl_payload1(struct adapter *, struct sge_fl *, uint32_t,
144     int *);
145 static struct mbuf *get_fl_payload2(struct adapter *, struct sge_fl *, uint32_t,
146     int *);
147 static int t4_eth_rx(struct sge_iq *, const struct rss_header *, struct mbuf *);
148 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int,
149     int);
150 static inline void init_fl(struct adapter *, struct sge_fl *, int, int, int,
151     char *);
152 static inline void init_eq(struct sge_eq *, int, int, uint8_t, uint16_t,
153     char *);
154 static int alloc_ring(struct adapter *, size_t, bus_dma_tag_t *, bus_dmamap_t *,
155     bus_addr_t *, void **);
156 static int free_ring(struct adapter *, bus_dma_tag_t, bus_dmamap_t, bus_addr_t,
157     void *);
158 static int alloc_iq_fl(struct port_info *, struct sge_iq *, struct sge_fl *,
159     int, int);
160 static int free_iq_fl(struct port_info *, struct sge_iq *, struct sge_fl *);
161 static int alloc_fwq(struct adapter *);
162 static int free_fwq(struct adapter *);
163 static int alloc_mgmtq(struct adapter *);
164 static int free_mgmtq(struct adapter *);
165 static int alloc_rxq(struct port_info *, struct sge_rxq *, int, int,
166     struct sysctl_oid *);
167 static int free_rxq(struct port_info *, struct sge_rxq *);
168 #ifdef TCP_OFFLOAD
169 static int alloc_ofld_rxq(struct port_info *, struct sge_ofld_rxq *, int, int,
170     struct sysctl_oid *);
171 static int free_ofld_rxq(struct port_info *, struct sge_ofld_rxq *);
172 #endif
173 static int ctrl_eq_alloc(struct adapter *, struct sge_eq *);
174 static int eth_eq_alloc(struct adapter *, struct port_info *, struct sge_eq *);
175 #ifdef TCP_OFFLOAD
176 static int ofld_eq_alloc(struct adapter *, struct port_info *, struct sge_eq *);
177 #endif
178 static int alloc_eq(struct adapter *, struct port_info *, struct sge_eq *);
179 static int free_eq(struct adapter *, struct sge_eq *);
180 static int alloc_wrq(struct adapter *, struct port_info *, struct sge_wrq *,
181     struct sysctl_oid *);
182 static int free_wrq(struct adapter *, struct sge_wrq *);
183 static int alloc_txq(struct port_info *, struct sge_txq *, int,
184     struct sysctl_oid *);
185 static int free_txq(struct port_info *, struct sge_txq *);
186 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int);
187 static inline bool is_new_response(const struct sge_iq *, struct rsp_ctrl **);
188 static inline void iq_next(struct sge_iq *);
189 static inline void ring_fl_db(struct adapter *, struct sge_fl *);
190 static int refill_fl(struct adapter *, struct sge_fl *, int);
191 static void refill_sfl(void *);
192 static int alloc_fl_sdesc(struct sge_fl *);
193 static void free_fl_sdesc(struct adapter *, struct sge_fl *);
194 static void set_fl_tag_idx(struct adapter *, struct sge_fl *, int);
195 static void add_fl_to_sfl(struct adapter *, struct sge_fl *);
196 
197 static int get_pkt_sgl(struct sge_txq *, struct mbuf **, struct sgl *, int);
198 static int free_pkt_sgl(struct sge_txq *, struct sgl *);
199 static int write_txpkt_wr(struct port_info *, struct sge_txq *, struct mbuf *,
200     struct sgl *);
201 static int add_to_txpkts(struct port_info *, struct sge_txq *, struct txpkts *,
202     struct mbuf *, struct sgl *);
203 static void write_txpkts_wr(struct sge_txq *, struct txpkts *);
204 static inline void write_ulp_cpl_sgl(struct port_info *, struct sge_txq *,
205     struct txpkts *, struct mbuf *, struct sgl *);
206 static int write_sgl_to_txd(struct sge_eq *, struct sgl *, caddr_t *);
207 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int);
208 static inline void ring_eq_db(struct adapter *, struct sge_eq *);
209 static inline int reclaimable(struct sge_eq *);
210 static int reclaim_tx_descs(struct sge_txq *, int, int);
211 static void write_eqflush_wr(struct sge_eq *);
212 static __be64 get_flit(bus_dma_segment_t *, int, int);
213 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
214     struct mbuf *);
215 static int handle_fw_msg(struct sge_iq *, const struct rss_header *,
216     struct mbuf *);
217 
218 static int sysctl_uint16(SYSCTL_HANDLER_ARGS);
219 
220 /*
221  * Called on MOD_LOAD.  Validates and calculates the SGE tunables.
222  */
223 void
224 t4_sge_modload(void)
225 {
226 	int pad;
227 
228 	/* set pad to a reasonable powerof2 between 16 and 4096 (inclusive) */
229 #if defined(__i386__) || defined(__amd64__)
230 	pad = max(cpu_clflush_line_size, 16);
231 #else
232 	pad = max(CACHE_LINE_SIZE, 16);
233 #endif
234 	pad = min(pad, 4096);
235 
236 	if (fl_pktshift < 0 || fl_pktshift > 7) {
237 		printf("Invalid hw.cxgbe.fl_pktshift value (%d),"
238 		    " using 2 instead.\n", fl_pktshift);
239 		fl_pktshift = 2;
240 	}
241 
242 	if (fl_pad != 0 &&
243 	    (fl_pad < 32 || fl_pad > 4096 || !powerof2(fl_pad))) {
244 
245 		if (fl_pad != -1) {
246 			printf("Invalid hw.cxgbe.fl_pad value (%d),"
247 			    " using %d instead.\n", fl_pad, max(pad, 32));
248 		}
249 		fl_pad = max(pad, 32);
250 	}
251 
252 	/*
253 	 * T4 has the same pad and pack boundary.  If a pad boundary is set,
254 	 * pack boundary must be set to the same value.  Otherwise take the
255 	 * specified value or auto-calculate something reasonable.
256 	 */
257 	if (fl_pad)
258 		t4_fl_pack = fl_pad;
259 	else if (fl_pack < 32 || fl_pack > 4096 || !powerof2(fl_pack))
260 		t4_fl_pack = max(pad, 32);
261 	else
262 		t4_fl_pack = fl_pack;
263 
264 	/* T5's pack boundary is independent of the pad boundary. */
265 	if (fl_pack < 16 || fl_pack == 32 || fl_pack > 4096 ||
266 	    !powerof2(fl_pack))
267 	       t5_fl_pack = max(pad, 64);
268 	else
269 	       t5_fl_pack = fl_pack;
270 
271 	if (spg_len != 64 && spg_len != 128) {
272 		int len;
273 
274 #if defined(__i386__) || defined(__amd64__)
275 		len = cpu_clflush_line_size > 64 ? 128 : 64;
276 #else
277 		len = 64;
278 #endif
279 		if (spg_len != -1) {
280 			printf("Invalid hw.cxgbe.spg_len value (%d),"
281 			    " using %d instead.\n", spg_len, len);
282 		}
283 		spg_len = len;
284 	}
285 
286 	if (cong_drop < -1 || cong_drop > 1) {
287 		printf("Invalid hw.cxgbe.cong_drop value (%d),"
288 		    " using 0 instead.\n", cong_drop);
289 		cong_drop = 0;
290 	}
291 }
292 
293 void
294 t4_init_sge_cpl_handlers(struct adapter *sc)
295 {
296 
297 	t4_register_cpl_handler(sc, CPL_FW4_MSG, handle_fw_msg);
298 	t4_register_cpl_handler(sc, CPL_FW6_MSG, handle_fw_msg);
299 	t4_register_cpl_handler(sc, CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
300 	t4_register_cpl_handler(sc, CPL_RX_PKT, t4_eth_rx);
301 	t4_register_fw_msg_handler(sc, FW6_TYPE_CMD_RPL, t4_handle_fw_rpl);
302 }
303 
304 /*
305  * adap->params.vpd.cclk must be set up before this is called.
306  */
307 void
308 t4_tweak_chip_settings(struct adapter *sc)
309 {
310 	int i;
311 	uint32_t v, m;
312 	int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200};
313 	int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk;
314 	int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */
315 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
316 	int sw_flbuf_sizes[] = {
317 		MCLBYTES,
318 #if MJUMPAGESIZE != MCLBYTES
319 		MJUMPAGESIZE,
320 #endif
321 		MJUM9BYTES,
322 		MJUM16BYTES,
323 		MJUMPAGESIZE - MSIZE
324 	};
325 
326 	KASSERT(sc->flags & MASTER_PF,
327 	    ("%s: trying to change chip settings when not master.", __func__));
328 
329 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
330 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
331 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
332 	if (is_t4(sc) && (fl_pad || buffer_packing)) {
333 		/* t4_fl_pack has the correct value even when fl_pad = 0 */
334 		m |= V_INGPADBOUNDARY(M_INGPADBOUNDARY);
335 		v |= V_INGPADBOUNDARY(ilog2(t4_fl_pack) - 5);
336 	} else if (is_t5(sc) && fl_pad) {
337 		m |= V_INGPADBOUNDARY(M_INGPADBOUNDARY);
338 		v |= V_INGPADBOUNDARY(ilog2(fl_pad) - 5);
339 	}
340 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
341 
342 	if (is_t5(sc) && buffer_packing) {
343 		m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY);
344 		if (t5_fl_pack == 16)
345 			v = V_INGPACKBOUNDARY(0);
346 		else
347 			v = V_INGPACKBOUNDARY(ilog2(t5_fl_pack) - 5);
348 		t4_set_reg_field(sc, A_SGE_CONTROL2, m, v);
349 	}
350 
351 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
352 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
353 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
354 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
355 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
356 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
357 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
358 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
359 	t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v);
360 
361 	for (i = 0; i < min(nitems(sw_flbuf_sizes), 16); i++) {
362 		t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i),
363 		    sw_flbuf_sizes[i]);
364 	}
365 
366 	v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) |
367 	    V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]);
368 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v);
369 
370 	KASSERT(intr_timer[0] <= timer_max,
371 	    ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0],
372 	    timer_max));
373 	for (i = 1; i < nitems(intr_timer); i++) {
374 		KASSERT(intr_timer[i] >= intr_timer[i - 1],
375 		    ("%s: timers not listed in increasing order (%d)",
376 		    __func__, i));
377 
378 		while (intr_timer[i] > timer_max) {
379 			if (i == nitems(intr_timer) - 1) {
380 				intr_timer[i] = timer_max;
381 				break;
382 			}
383 			intr_timer[i] += intr_timer[i - 1];
384 			intr_timer[i] /= 2;
385 		}
386 	}
387 
388 	v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) |
389 	    V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1]));
390 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v);
391 	v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) |
392 	    V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3]));
393 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v);
394 	v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) |
395 	    V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5]));
396 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v);
397 
398 	if (cong_drop == 0) {
399 		m = F_TUNNELCNGDROP0 | F_TUNNELCNGDROP1 | F_TUNNELCNGDROP2 |
400 		    F_TUNNELCNGDROP3;
401 		t4_set_reg_field(sc, A_TP_PARA_REG3, m, 0);
402 	}
403 
404 	/* 4K, 16K, 64K, 256K DDP "page sizes" */
405 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
406 	t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v);
407 
408 	m = v = F_TDDPTAGTCB;
409 	t4_set_reg_field(sc, A_ULP_RX_CTL, m, v);
410 
411 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
412 	    F_RESETDDPOFFSET;
413 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
414 	t4_set_reg_field(sc, A_TP_PARA_REG5, m, v);
415 }
416 
417 /*
418  * XXX: driver really should be able to deal with unexpected settings.
419  */
420 int
421 t4_read_chip_settings(struct adapter *sc)
422 {
423 	struct sge *s = &sc->sge;
424 	int i, j, n, rc = 0;
425 	uint32_t m, v, r;
426 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
427 	uint32_t sge_flbuf_sizes[16], sw_flbuf_sizes[] = {
428 		MCLBYTES,
429 #if MJUMPAGESIZE != MCLBYTES
430 		MJUMPAGESIZE,
431 #endif
432 		MJUM9BYTES,
433 		MJUM16BYTES
434 	};
435 
436 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
437 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
438 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
439 	if (is_t4(sc) && (fl_pad || buffer_packing)) {
440 		m |= V_INGPADBOUNDARY(M_INGPADBOUNDARY);
441 		v |= V_INGPADBOUNDARY(ilog2(t4_fl_pack) - 5);
442 	} else if (is_t5(sc) && fl_pad) {
443 		m |= V_INGPADBOUNDARY(M_INGPADBOUNDARY);
444 		v |= V_INGPADBOUNDARY(ilog2(fl_pad) - 5);
445 	}
446 	r = t4_read_reg(sc, A_SGE_CONTROL);
447 	if ((r & m) != v) {
448 		device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r);
449 		rc = EINVAL;
450 	}
451 
452 	if (is_t5(sc) && buffer_packing) {
453 		m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY);
454 		if (t5_fl_pack == 16)
455 			v = V_INGPACKBOUNDARY(0);
456 		else
457 			v = V_INGPACKBOUNDARY(ilog2(t5_fl_pack) - 5);
458 		r = t4_read_reg(sc, A_SGE_CONTROL2);
459 		if ((r & m) != v) {
460 			device_printf(sc->dev,
461 			    "invalid SGE_CONTROL2(0x%x)\n", r);
462 			rc = EINVAL;
463 		}
464 	}
465 
466 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
467 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
468 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
469 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
470 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
471 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
472 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
473 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
474 	r = t4_read_reg(sc, A_SGE_HOST_PAGE_SIZE);
475 	if (r != v) {
476 		device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r);
477 		rc = EINVAL;
478 	}
479 
480 	/*
481 	 * Make a list of SGE FL buffer sizes programmed in the chip and tally
482 	 * it with the FL buffer sizes that we'd like to use.
483 	 */
484 	n = 0;
485 	for (i = 0; i < nitems(sge_flbuf_sizes); i++) {
486 		r = t4_read_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i));
487 		sge_flbuf_sizes[i] = r;
488 		if (r == MJUMPAGESIZE - MSIZE &&
489 		    (sc->flags & BUF_PACKING_OK) == 0) {
490 			sc->flags |= BUF_PACKING_OK;
491 			FL_BUF_HWTAG(sc, n) = i;
492 			FL_BUF_SIZE(sc, n) = MJUMPAGESIZE - MSIZE;
493 			FL_BUF_TYPE(sc, n) = m_gettype(MJUMPAGESIZE);
494 			FL_BUF_ZONE(sc, n) = m_getzone(MJUMPAGESIZE);
495 			n++;
496 		}
497 	}
498 	for (i = 0; i < nitems(sw_flbuf_sizes); i++) {
499 		for (j = 0; j < nitems(sge_flbuf_sizes); j++) {
500 			if (sw_flbuf_sizes[i] != sge_flbuf_sizes[j])
501 				continue;
502 			FL_BUF_HWTAG(sc, n) = j;
503 			FL_BUF_SIZE(sc, n) = sw_flbuf_sizes[i];
504 			FL_BUF_TYPE(sc, n) = m_gettype(sw_flbuf_sizes[i]);
505 			FL_BUF_ZONE(sc, n) = m_getzone(sw_flbuf_sizes[i]);
506 			n++;
507 			break;
508 		}
509 	}
510 	if (n == 0) {
511 		device_printf(sc->dev, "no usable SGE FL buffer size.\n");
512 		rc = EINVAL;
513 	} else if (n == 1 && (sc->flags & BUF_PACKING_OK)) {
514 		device_printf(sc->dev,
515 		    "no usable SGE FL buffer size when not packing buffers.\n");
516 		rc = EINVAL;
517 	}
518 	FL_BUF_SIZES(sc) = n;
519 
520 	r = t4_read_reg(sc, A_SGE_INGRESS_RX_THRESHOLD);
521 	s->counter_val[0] = G_THRESHOLD_0(r);
522 	s->counter_val[1] = G_THRESHOLD_1(r);
523 	s->counter_val[2] = G_THRESHOLD_2(r);
524 	s->counter_val[3] = G_THRESHOLD_3(r);
525 
526 	r = t4_read_reg(sc, A_SGE_TIMER_VALUE_0_AND_1);
527 	s->timer_val[0] = G_TIMERVALUE0(r) / core_ticks_per_usec(sc);
528 	s->timer_val[1] = G_TIMERVALUE1(r) / core_ticks_per_usec(sc);
529 	r = t4_read_reg(sc, A_SGE_TIMER_VALUE_2_AND_3);
530 	s->timer_val[2] = G_TIMERVALUE2(r) / core_ticks_per_usec(sc);
531 	s->timer_val[3] = G_TIMERVALUE3(r) / core_ticks_per_usec(sc);
532 	r = t4_read_reg(sc, A_SGE_TIMER_VALUE_4_AND_5);
533 	s->timer_val[4] = G_TIMERVALUE4(r) / core_ticks_per_usec(sc);
534 	s->timer_val[5] = G_TIMERVALUE5(r) / core_ticks_per_usec(sc);
535 
536 	if (cong_drop == 0) {
537 		m = F_TUNNELCNGDROP0 | F_TUNNELCNGDROP1 | F_TUNNELCNGDROP2 |
538 		    F_TUNNELCNGDROP3;
539 		r = t4_read_reg(sc, A_TP_PARA_REG3);
540 		if (r & m) {
541 			device_printf(sc->dev,
542 			    "invalid TP_PARA_REG3(0x%x)\n", r);
543 			rc = EINVAL;
544 		}
545 	}
546 
547 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
548 	r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ);
549 	if (r != v) {
550 		device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r);
551 		rc = EINVAL;
552 	}
553 
554 	m = v = F_TDDPTAGTCB;
555 	r = t4_read_reg(sc, A_ULP_RX_CTL);
556 	if ((r & m) != v) {
557 		device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r);
558 		rc = EINVAL;
559 	}
560 
561 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
562 	    F_RESETDDPOFFSET;
563 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
564 	r = t4_read_reg(sc, A_TP_PARA_REG5);
565 	if ((r & m) != v) {
566 		device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r);
567 		rc = EINVAL;
568 	}
569 
570 	r = t4_read_reg(sc, A_SGE_CONM_CTRL);
571 	s->fl_starve_threshold = G_EGRTHRESHOLD(r) * 2 + 1;
572 
573 	/* egress queues: log2 of # of doorbells per BAR2 page */
574 	r = t4_read_reg(sc, A_SGE_EGRESS_QUEUES_PER_PAGE_PF);
575 	r >>= S_QUEUESPERPAGEPF0 +
576 	    (S_QUEUESPERPAGEPF1 - S_QUEUESPERPAGEPF0) * sc->pf;
577 	s->eq_s_qpp = r & M_QUEUESPERPAGEPF0;
578 
579 	/* ingress queues: log2 of # of doorbells per BAR2 page */
580 	r = t4_read_reg(sc, A_SGE_INGRESS_QUEUES_PER_PAGE_PF);
581 	r >>= S_QUEUESPERPAGEPF0 +
582 	    (S_QUEUESPERPAGEPF1 - S_QUEUESPERPAGEPF0) * sc->pf;
583 	s->iq_s_qpp = r & M_QUEUESPERPAGEPF0;
584 
585 	t4_init_tp_params(sc);
586 
587 	t4_read_mtu_tbl(sc, sc->params.mtus, NULL);
588 	t4_load_mtus(sc, sc->params.mtus, sc->params.a_wnd, sc->params.b_wnd);
589 
590 	return (rc);
591 }
592 
593 int
594 t4_create_dma_tag(struct adapter *sc)
595 {
596 	int rc;
597 
598 	rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
599 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
600 	    BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
601 	    NULL, &sc->dmat);
602 	if (rc != 0) {
603 		device_printf(sc->dev,
604 		    "failed to create main DMA tag: %d\n", rc);
605 	}
606 
607 	return (rc);
608 }
609 
610 static inline int
611 enable_buffer_packing(struct adapter *sc)
612 {
613 
614 	if (sc->flags & BUF_PACKING_OK &&
615 	    ((is_t5(sc) && buffer_packing) ||	/* 1 or -1 both ok for T5 */
616 	    (is_t4(sc) && buffer_packing == 1)))
617 		return (1);
618 	return (0);
619 }
620 
621 void
622 t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
623     struct sysctl_oid_list *children)
624 {
625 
626 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD,
627 	    NULL, fl_pktshift, "payload DMA offset in rx buffer (bytes)");
628 
629 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD,
630 	    NULL, fl_pad, "payload pad boundary (bytes)");
631 
632 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD,
633 	    NULL, spg_len, "status page size (bytes)");
634 
635 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD,
636 	    NULL, cong_drop, "congestion drop setting");
637 
638 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "buffer_packing", CTLFLAG_RD,
639 	    NULL, enable_buffer_packing(sc),
640 	    "pack multiple frames in one fl buffer");
641 
642 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD,
643 	    NULL, is_t5(sc) ? t5_fl_pack : t4_fl_pack,
644 	    "payload pack boundary (bytes)");
645 }
646 
647 int
648 t4_destroy_dma_tag(struct adapter *sc)
649 {
650 	if (sc->dmat)
651 		bus_dma_tag_destroy(sc->dmat);
652 
653 	return (0);
654 }
655 
656 /*
657  * Allocate and initialize the firmware event queue and the management queue.
658  *
659  * Returns errno on failure.  Resources allocated up to that point may still be
660  * allocated.  Caller is responsible for cleanup in case this function fails.
661  */
662 int
663 t4_setup_adapter_queues(struct adapter *sc)
664 {
665 	int rc;
666 
667 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
668 
669 	sysctl_ctx_init(&sc->ctx);
670 	sc->flags |= ADAP_SYSCTL_CTX;
671 
672 	/*
673 	 * Firmware event queue
674 	 */
675 	rc = alloc_fwq(sc);
676 	if (rc != 0)
677 		return (rc);
678 
679 	/*
680 	 * Management queue.  This is just a control queue that uses the fwq as
681 	 * its associated iq.
682 	 */
683 	rc = alloc_mgmtq(sc);
684 
685 	return (rc);
686 }
687 
688 /*
689  * Idempotent
690  */
691 int
692 t4_teardown_adapter_queues(struct adapter *sc)
693 {
694 
695 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
696 
697 	/* Do this before freeing the queue */
698 	if (sc->flags & ADAP_SYSCTL_CTX) {
699 		sysctl_ctx_free(&sc->ctx);
700 		sc->flags &= ~ADAP_SYSCTL_CTX;
701 	}
702 
703 	free_mgmtq(sc);
704 	free_fwq(sc);
705 
706 	return (0);
707 }
708 
709 static inline int
710 first_vector(struct port_info *pi)
711 {
712 	struct adapter *sc = pi->adapter;
713 	int rc = T4_EXTRA_INTR, i;
714 
715 	if (sc->intr_count == 1)
716 		return (0);
717 
718 	for_each_port(sc, i) {
719 		struct port_info *p = sc->port[i];
720 
721 		if (i == pi->port_id)
722 			break;
723 
724 #ifdef TCP_OFFLOAD
725 		if (sc->flags & INTR_DIRECT)
726 			rc += p->nrxq + p->nofldrxq;
727 		else
728 			rc += max(p->nrxq, p->nofldrxq);
729 #else
730 		/*
731 		 * Not compiled with offload support and intr_count > 1.  Only
732 		 * NIC queues exist and they'd better be taking direct
733 		 * interrupts.
734 		 */
735 		KASSERT(sc->flags & INTR_DIRECT,
736 		    ("%s: intr_count %d, !INTR_DIRECT", __func__,
737 		    sc->intr_count));
738 
739 		rc += p->nrxq;
740 #endif
741 	}
742 
743 	return (rc);
744 }
745 
746 /*
747  * Given an arbitrary "index," come up with an iq that can be used by other
748  * queues (of this port) for interrupt forwarding, SGE egress updates, etc.
749  * The iq returned is guaranteed to be something that takes direct interrupts.
750  */
751 static struct sge_iq *
752 port_intr_iq(struct port_info *pi, int idx)
753 {
754 	struct adapter *sc = pi->adapter;
755 	struct sge *s = &sc->sge;
756 	struct sge_iq *iq = NULL;
757 
758 	if (sc->intr_count == 1)
759 		return (&sc->sge.fwq);
760 
761 #ifdef TCP_OFFLOAD
762 	if (sc->flags & INTR_DIRECT) {
763 		idx %= pi->nrxq + pi->nofldrxq;
764 
765 		if (idx >= pi->nrxq) {
766 			idx -= pi->nrxq;
767 			iq = &s->ofld_rxq[pi->first_ofld_rxq + idx].iq;
768 		} else
769 			iq = &s->rxq[pi->first_rxq + idx].iq;
770 
771 	} else {
772 		idx %= max(pi->nrxq, pi->nofldrxq);
773 
774 		if (pi->nrxq >= pi->nofldrxq)
775 			iq = &s->rxq[pi->first_rxq + idx].iq;
776 		else
777 			iq = &s->ofld_rxq[pi->first_ofld_rxq + idx].iq;
778 	}
779 #else
780 	/*
781 	 * Not compiled with offload support and intr_count > 1.  Only NIC
782 	 * queues exist and they'd better be taking direct interrupts.
783 	 */
784 	KASSERT(sc->flags & INTR_DIRECT,
785 	    ("%s: intr_count %d, !INTR_DIRECT", __func__, sc->intr_count));
786 
787 	idx %= pi->nrxq;
788 	iq = &s->rxq[pi->first_rxq + idx].iq;
789 #endif
790 
791 	KASSERT(iq->flags & IQ_INTR, ("%s: EDOOFUS", __func__));
792 	return (iq);
793 }
794 
795 static inline int
796 mtu_to_bufsize(int mtu)
797 {
798 	int bufsize;
799 
800 	/* large enough for a frame even when VLAN extraction is disabled */
801 	bufsize = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN + mtu;
802 	bufsize = roundup2(bufsize + fl_pktshift, fl_pad);
803 
804 	return (bufsize);
805 }
806 
807 #ifdef TCP_OFFLOAD
808 static inline int
809 mtu_to_bufsize_toe(struct adapter *sc, int mtu)
810 {
811 
812 	if (sc->tt.rx_coalesce)
813 		return (G_RXCOALESCESIZE(t4_read_reg(sc, A_TP_PARA_REG2)));
814 
815 	return (mtu);
816 }
817 #endif
818 
819 int
820 t4_setup_port_queues(struct port_info *pi)
821 {
822 	int rc = 0, i, j, intr_idx, iqid;
823 	struct sge_rxq *rxq;
824 	struct sge_txq *txq;
825 	struct sge_wrq *ctrlq;
826 #ifdef TCP_OFFLOAD
827 	struct sge_ofld_rxq *ofld_rxq;
828 	struct sge_wrq *ofld_txq;
829 	struct sysctl_oid *oid2 = NULL;
830 #endif
831 	char name[16];
832 	struct adapter *sc = pi->adapter;
833 	struct ifnet *ifp = pi->ifp;
834 	struct sysctl_oid *oid = device_get_sysctl_tree(pi->dev);
835 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
836 	int bufsize, pack;
837 
838 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "rxq", CTLFLAG_RD,
839 	    NULL, "rx queues");
840 
841 #ifdef TCP_OFFLOAD
842 	if (is_offload(sc)) {
843 		oid2 = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "ofld_rxq",
844 		    CTLFLAG_RD, NULL,
845 		    "rx queues for offloaded TCP connections");
846 	}
847 #endif
848 
849 	/* Interrupt vector to start from (when using multiple vectors) */
850 	intr_idx = first_vector(pi);
851 
852 	/*
853 	 * First pass over all rx queues (NIC and TOE):
854 	 * a) initialize iq and fl
855 	 * b) allocate queue iff it will take direct interrupts.
856 	 */
857 	bufsize = mtu_to_bufsize(ifp->if_mtu);
858 	pack = enable_buffer_packing(sc);
859 	for_each_rxq(pi, i, rxq) {
860 
861 		init_iq(&rxq->iq, sc, pi->tmr_idx, pi->pktc_idx, pi->qsize_rxq,
862 		    RX_IQ_ESIZE);
863 
864 		snprintf(name, sizeof(name), "%s rxq%d-fl",
865 		    device_get_nameunit(pi->dev), i);
866 		init_fl(sc, &rxq->fl, pi->qsize_rxq / 8, bufsize, pack, name);
867 
868 		if (sc->flags & INTR_DIRECT
869 #ifdef TCP_OFFLOAD
870 		    || (sc->intr_count > 1 && pi->nrxq >= pi->nofldrxq)
871 #endif
872 		   ) {
873 			rxq->iq.flags |= IQ_INTR;
874 			rc = alloc_rxq(pi, rxq, intr_idx, i, oid);
875 			if (rc != 0)
876 				goto done;
877 			intr_idx++;
878 		}
879 	}
880 
881 #ifdef TCP_OFFLOAD
882 	bufsize = mtu_to_bufsize_toe(sc, ifp->if_mtu);
883 	pack = 0;	/* XXX: think about this some more */
884 	for_each_ofld_rxq(pi, i, ofld_rxq) {
885 
886 		init_iq(&ofld_rxq->iq, sc, pi->tmr_idx, pi->pktc_idx,
887 		    pi->qsize_rxq, RX_IQ_ESIZE);
888 
889 		snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
890 		    device_get_nameunit(pi->dev), i);
891 		init_fl(sc, &ofld_rxq->fl, pi->qsize_rxq / 8, bufsize, pack,
892 		    name);
893 
894 		if (sc->flags & INTR_DIRECT ||
895 		    (sc->intr_count > 1 && pi->nofldrxq > pi->nrxq)) {
896 			ofld_rxq->iq.flags |= IQ_INTR;
897 			rc = alloc_ofld_rxq(pi, ofld_rxq, intr_idx, i, oid2);
898 			if (rc != 0)
899 				goto done;
900 			intr_idx++;
901 		}
902 	}
903 #endif
904 
905 	/*
906 	 * Second pass over all rx queues (NIC and TOE).  The queues forwarding
907 	 * their interrupts are allocated now.
908 	 */
909 	j = 0;
910 	for_each_rxq(pi, i, rxq) {
911 		if (rxq->iq.flags & IQ_INTR)
912 			continue;
913 
914 		intr_idx = port_intr_iq(pi, j)->abs_id;
915 
916 		rc = alloc_rxq(pi, rxq, intr_idx, i, oid);
917 		if (rc != 0)
918 			goto done;
919 		j++;
920 	}
921 
922 #ifdef TCP_OFFLOAD
923 	for_each_ofld_rxq(pi, i, ofld_rxq) {
924 		if (ofld_rxq->iq.flags & IQ_INTR)
925 			continue;
926 
927 		intr_idx = port_intr_iq(pi, j)->abs_id;
928 
929 		rc = alloc_ofld_rxq(pi, ofld_rxq, intr_idx, i, oid2);
930 		if (rc != 0)
931 			goto done;
932 		j++;
933 	}
934 #endif
935 
936 	/*
937 	 * Now the tx queues.  Only one pass needed.
938 	 */
939 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "txq", CTLFLAG_RD,
940 	    NULL, "tx queues");
941 	j = 0;
942 	for_each_txq(pi, i, txq) {
943 		uint16_t iqid;
944 
945 		iqid = port_intr_iq(pi, j)->cntxt_id;
946 
947 		snprintf(name, sizeof(name), "%s txq%d",
948 		    device_get_nameunit(pi->dev), i);
949 		init_eq(&txq->eq, EQ_ETH, pi->qsize_txq, pi->tx_chan, iqid,
950 		    name);
951 
952 		rc = alloc_txq(pi, txq, i, oid);
953 		if (rc != 0)
954 			goto done;
955 		j++;
956 	}
957 
958 #ifdef TCP_OFFLOAD
959 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "ofld_txq",
960 	    CTLFLAG_RD, NULL, "tx queues for offloaded TCP connections");
961 	for_each_ofld_txq(pi, i, ofld_txq) {
962 		uint16_t iqid;
963 
964 		iqid = port_intr_iq(pi, j)->cntxt_id;
965 
966 		snprintf(name, sizeof(name), "%s ofld_txq%d",
967 		    device_get_nameunit(pi->dev), i);
968 		init_eq(&ofld_txq->eq, EQ_OFLD, pi->qsize_txq, pi->tx_chan,
969 		    iqid, name);
970 
971 		snprintf(name, sizeof(name), "%d", i);
972 		oid2 = SYSCTL_ADD_NODE(&pi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
973 		    name, CTLFLAG_RD, NULL, "offload tx queue");
974 
975 		rc = alloc_wrq(sc, pi, ofld_txq, oid2);
976 		if (rc != 0)
977 			goto done;
978 		j++;
979 	}
980 #endif
981 
982 	/*
983 	 * Finally, the control queue.
984 	 */
985 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "ctrlq", CTLFLAG_RD,
986 	    NULL, "ctrl queue");
987 	ctrlq = &sc->sge.ctrlq[pi->port_id];
988 	iqid = port_intr_iq(pi, 0)->cntxt_id;
989 	snprintf(name, sizeof(name), "%s ctrlq", device_get_nameunit(pi->dev));
990 	init_eq(&ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, pi->tx_chan, iqid, name);
991 	rc = alloc_wrq(sc, pi, ctrlq, oid);
992 
993 done:
994 	if (rc)
995 		t4_teardown_port_queues(pi);
996 
997 	return (rc);
998 }
999 
1000 /*
1001  * Idempotent
1002  */
1003 int
1004 t4_teardown_port_queues(struct port_info *pi)
1005 {
1006 	int i;
1007 	struct adapter *sc = pi->adapter;
1008 	struct sge_rxq *rxq;
1009 	struct sge_txq *txq;
1010 #ifdef TCP_OFFLOAD
1011 	struct sge_ofld_rxq *ofld_rxq;
1012 	struct sge_wrq *ofld_txq;
1013 #endif
1014 
1015 	/* Do this before freeing the queues */
1016 	if (pi->flags & PORT_SYSCTL_CTX) {
1017 		sysctl_ctx_free(&pi->ctx);
1018 		pi->flags &= ~PORT_SYSCTL_CTX;
1019 	}
1020 
1021 	/*
1022 	 * Take down all the tx queues first, as they reference the rx queues
1023 	 * (for egress updates, etc.).
1024 	 */
1025 
1026 	free_wrq(sc, &sc->sge.ctrlq[pi->port_id]);
1027 
1028 	for_each_txq(pi, i, txq) {
1029 		free_txq(pi, txq);
1030 	}
1031 
1032 #ifdef TCP_OFFLOAD
1033 	for_each_ofld_txq(pi, i, ofld_txq) {
1034 		free_wrq(sc, ofld_txq);
1035 	}
1036 #endif
1037 
1038 	/*
1039 	 * Then take down the rx queues that forward their interrupts, as they
1040 	 * reference other rx queues.
1041 	 */
1042 
1043 	for_each_rxq(pi, i, rxq) {
1044 		if ((rxq->iq.flags & IQ_INTR) == 0)
1045 			free_rxq(pi, rxq);
1046 	}
1047 
1048 #ifdef TCP_OFFLOAD
1049 	for_each_ofld_rxq(pi, i, ofld_rxq) {
1050 		if ((ofld_rxq->iq.flags & IQ_INTR) == 0)
1051 			free_ofld_rxq(pi, ofld_rxq);
1052 	}
1053 #endif
1054 
1055 	/*
1056 	 * Then take down the rx queues that take direct interrupts.
1057 	 */
1058 
1059 	for_each_rxq(pi, i, rxq) {
1060 		if (rxq->iq.flags & IQ_INTR)
1061 			free_rxq(pi, rxq);
1062 	}
1063 
1064 #ifdef TCP_OFFLOAD
1065 	for_each_ofld_rxq(pi, i, ofld_rxq) {
1066 		if (ofld_rxq->iq.flags & IQ_INTR)
1067 			free_ofld_rxq(pi, ofld_rxq);
1068 	}
1069 #endif
1070 
1071 	return (0);
1072 }
1073 
1074 /*
1075  * Deals with errors and the firmware event queue.  All data rx queues forward
1076  * their interrupt to the firmware event queue.
1077  */
1078 void
1079 t4_intr_all(void *arg)
1080 {
1081 	struct adapter *sc = arg;
1082 	struct sge_iq *fwq = &sc->sge.fwq;
1083 
1084 	t4_intr_err(arg);
1085 	if (atomic_cmpset_int(&fwq->state, IQS_IDLE, IQS_BUSY)) {
1086 		service_iq(fwq, 0);
1087 		atomic_cmpset_int(&fwq->state, IQS_BUSY, IQS_IDLE);
1088 	}
1089 }
1090 
1091 /* Deals with error interrupts */
1092 void
1093 t4_intr_err(void *arg)
1094 {
1095 	struct adapter *sc = arg;
1096 
1097 	t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
1098 	t4_slow_intr_handler(sc);
1099 }
1100 
1101 void
1102 t4_intr_evt(void *arg)
1103 {
1104 	struct sge_iq *iq = arg;
1105 
1106 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1107 		service_iq(iq, 0);
1108 		atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1109 	}
1110 }
1111 
1112 void
1113 t4_intr(void *arg)
1114 {
1115 	struct sge_iq *iq = arg;
1116 
1117 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1118 		service_iq(iq, 0);
1119 		atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1120 	}
1121 }
1122 
1123 /*
1124  * Deals with anything and everything on the given ingress queue.
1125  */
1126 static int
1127 service_iq(struct sge_iq *iq, int budget)
1128 {
1129 	struct sge_iq *q;
1130 	struct sge_rxq *rxq = iq_to_rxq(iq);	/* Use iff iq is part of rxq */
1131 	struct sge_fl *fl = &rxq->fl;		/* Use iff IQ_HAS_FL */
1132 	struct adapter *sc = iq->adapter;
1133 	struct rsp_ctrl *ctrl;
1134 	const struct rss_header *rss;
1135 	int ndescs = 0, limit, fl_bufs_used = 0;
1136 	int rsp_type;
1137 	uint32_t lq;
1138 	struct mbuf *m0;
1139 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1140 #if defined(INET) || defined(INET6)
1141 	const struct timeval lro_timeout = {0, sc->lro_timeout};
1142 #endif
1143 
1144 	limit = budget ? budget : iq->qsize / 8;
1145 
1146 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1147 
1148 	/*
1149 	 * We always come back and check the descriptor ring for new indirect
1150 	 * interrupts and other responses after running a single handler.
1151 	 */
1152 	for (;;) {
1153 		while (is_new_response(iq, &ctrl)) {
1154 
1155 			rmb();
1156 
1157 			m0 = NULL;
1158 			rsp_type = G_RSPD_TYPE(ctrl->u.type_gen);
1159 			lq = be32toh(ctrl->pldbuflen_qid);
1160 			rss = (const void *)iq->cdesc;
1161 
1162 			switch (rsp_type) {
1163 			case X_RSPD_TYPE_FLBUF:
1164 
1165 				KASSERT(iq->flags & IQ_HAS_FL,
1166 				    ("%s: data for an iq (%p) with no freelist",
1167 				    __func__, iq));
1168 
1169 				m0 = fl->flags & FL_BUF_PACKING ?
1170 				    get_fl_payload1(sc, fl, lq, &fl_bufs_used) :
1171 				    get_fl_payload2(sc, fl, lq, &fl_bufs_used);
1172 
1173 				if (__predict_false(m0 == NULL))
1174 					goto process_iql;
1175 #ifdef T4_PKT_TIMESTAMP
1176 				/*
1177 				 * 60 bit timestamp for the payload is
1178 				 * *(uint64_t *)m0->m_pktdat.  Note that it is
1179 				 * in the leading free-space in the mbuf.  The
1180 				 * kernel can clobber it during a pullup,
1181 				 * m_copymdata, etc.  You need to make sure that
1182 				 * the mbuf reaches you unmolested if you care
1183 				 * about the timestamp.
1184 				 */
1185 				*(uint64_t *)m0->m_pktdat =
1186 				    be64toh(ctrl->u.last_flit) &
1187 				    0xfffffffffffffff;
1188 #endif
1189 
1190 				/* fall through */
1191 
1192 			case X_RSPD_TYPE_CPL:
1193 				KASSERT(rss->opcode < NUM_CPL_CMDS,
1194 				    ("%s: bad opcode %02x.", __func__,
1195 				    rss->opcode));
1196 				sc->cpl_handler[rss->opcode](iq, rss, m0);
1197 				break;
1198 
1199 			case X_RSPD_TYPE_INTR:
1200 
1201 				/*
1202 				 * Interrupts should be forwarded only to queues
1203 				 * that are not forwarding their interrupts.
1204 				 * This means service_iq can recurse but only 1
1205 				 * level deep.
1206 				 */
1207 				KASSERT(budget == 0,
1208 				    ("%s: budget %u, rsp_type %u", __func__,
1209 				    budget, rsp_type));
1210 
1211 				/*
1212 				 * There are 1K interrupt-capable queues (qids 0
1213 				 * through 1023).  A response type indicating a
1214 				 * forwarded interrupt with a qid >= 1K is an
1215 				 * iWARP async notification.
1216 				 */
1217 				if (lq >= 1024) {
1218                                         sc->an_handler(iq, ctrl);
1219                                         break;
1220                                 }
1221 
1222 				q = sc->sge.iqmap[lq - sc->sge.iq_start];
1223 				if (atomic_cmpset_int(&q->state, IQS_IDLE,
1224 				    IQS_BUSY)) {
1225 					if (service_iq(q, q->qsize / 8) == 0) {
1226 						atomic_cmpset_int(&q->state,
1227 						    IQS_BUSY, IQS_IDLE);
1228 					} else {
1229 						STAILQ_INSERT_TAIL(&iql, q,
1230 						    link);
1231 					}
1232 				}
1233 				break;
1234 
1235 			default:
1236 				KASSERT(0,
1237 				    ("%s: illegal response type %d on iq %p",
1238 				    __func__, rsp_type, iq));
1239 				log(LOG_ERR,
1240 				    "%s: illegal response type %d on iq %p",
1241 				    device_get_nameunit(sc->dev), rsp_type, iq);
1242 				break;
1243 			}
1244 
1245 			iq_next(iq);
1246 			if (++ndescs == limit) {
1247 				t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
1248 				    V_CIDXINC(ndescs) |
1249 				    V_INGRESSQID(iq->cntxt_id) |
1250 				    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1251 				ndescs = 0;
1252 
1253 #if defined(INET) || defined(INET6)
1254 				if (iq->flags & IQ_LRO_ENABLED &&
1255 				    sc->lro_timeout != 0) {
1256 					tcp_lro_flush_inactive(&rxq->lro,
1257 					    &lro_timeout);
1258 				}
1259 #endif
1260 
1261 				if (fl_bufs_used > 0) {
1262 					FL_LOCK(fl);
1263 					fl->needed += fl_bufs_used;
1264 					refill_fl(sc, fl, fl->cap / 8);
1265 					FL_UNLOCK(fl);
1266 					fl_bufs_used = 0;
1267 				}
1268 
1269 				if (budget)
1270 					return (EINPROGRESS);
1271 			}
1272 		}
1273 
1274 process_iql:
1275 		if (STAILQ_EMPTY(&iql))
1276 			break;
1277 
1278 		/*
1279 		 * Process the head only, and send it to the back of the list if
1280 		 * it's still not done.
1281 		 */
1282 		q = STAILQ_FIRST(&iql);
1283 		STAILQ_REMOVE_HEAD(&iql, link);
1284 		if (service_iq(q, q->qsize / 8) == 0)
1285 			atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1286 		else
1287 			STAILQ_INSERT_TAIL(&iql, q, link);
1288 	}
1289 
1290 #if defined(INET) || defined(INET6)
1291 	if (iq->flags & IQ_LRO_ENABLED) {
1292 		struct lro_ctrl *lro = &rxq->lro;
1293 		struct lro_entry *l;
1294 
1295 		while (!SLIST_EMPTY(&lro->lro_active)) {
1296 			l = SLIST_FIRST(&lro->lro_active);
1297 			SLIST_REMOVE_HEAD(&lro->lro_active, next);
1298 			tcp_lro_flush(lro, l);
1299 		}
1300 	}
1301 #endif
1302 
1303 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_CIDXINC(ndescs) |
1304 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1305 
1306 	if (iq->flags & IQ_HAS_FL) {
1307 		int starved;
1308 
1309 		FL_LOCK(fl);
1310 		fl->needed += fl_bufs_used;
1311 		starved = refill_fl(sc, fl, fl->cap / 4);
1312 		FL_UNLOCK(fl);
1313 		if (__predict_false(starved != 0))
1314 			add_fl_to_sfl(sc, fl);
1315 	}
1316 
1317 	return (0);
1318 }
1319 
1320 static int
1321 fill_mbuf_stash(struct sge_fl *fl)
1322 {
1323 	int i;
1324 
1325 	for (i = 0; i < nitems(fl->mstash); i++) {
1326 		if (fl->mstash[i] == NULL) {
1327 			struct mbuf *m;
1328 			if ((m = m_get(M_NOWAIT, MT_NOINIT)) == NULL)
1329 				return (ENOBUFS);
1330 			fl->mstash[i] = m;
1331 		}
1332 	}
1333 	return (0);
1334 }
1335 
1336 static struct mbuf *
1337 get_mbuf_from_stash(struct sge_fl *fl)
1338 {
1339 	int i;
1340 
1341 	for (i = 0; i < nitems(fl->mstash); i++) {
1342 		if (fl->mstash[i] != NULL) {
1343 			struct mbuf *m;
1344 
1345 			m = fl->mstash[i];
1346 			fl->mstash[i] = NULL;
1347 			return (m);
1348 		} else
1349 			fl->mstash[i] = m_get(M_NOWAIT, MT_NOINIT);
1350 	}
1351 
1352 	return (m_get(M_NOWAIT, MT_NOINIT));
1353 }
1354 
1355 static void
1356 return_mbuf_to_stash(struct sge_fl *fl, struct mbuf *m)
1357 {
1358 	int i;
1359 
1360 	if (m == NULL)
1361 		return;
1362 
1363 	for (i = 0; i < nitems(fl->mstash); i++) {
1364 		if (fl->mstash[i] == NULL) {
1365 			fl->mstash[i] = m;
1366 			return;
1367 		}
1368 	}
1369 	m_init(m, NULL, 0, M_NOWAIT, MT_DATA, 0);
1370 	m_free(m);
1371 }
1372 
1373 /* buf can be any address within the buffer */
1374 static inline u_int *
1375 find_buf_refcnt(caddr_t buf)
1376 {
1377 	uintptr_t ptr = (uintptr_t)buf;
1378 
1379 	return ((u_int *)((ptr & ~(MJUMPAGESIZE - 1)) + MSIZE - sizeof(u_int)));
1380 }
1381 
1382 static inline struct mbuf *
1383 find_buf_mbuf(caddr_t buf)
1384 {
1385 	uintptr_t ptr = (uintptr_t)buf;
1386 
1387 	return ((struct mbuf *)(ptr & ~(MJUMPAGESIZE - 1)));
1388 }
1389 
1390 static int
1391 rxb_free(struct mbuf *m, void *arg1, void *arg2)
1392 {
1393 	uma_zone_t zone = arg1;
1394 	caddr_t cl = arg2;
1395 #ifdef INVARIANTS
1396 	u_int refcount;
1397 
1398 	refcount = *find_buf_refcnt(cl);
1399 	KASSERT(refcount == 0, ("%s: cl %p refcount is %u", __func__,
1400 	    cl - MSIZE, refcount));
1401 #endif
1402 	cl -= MSIZE;
1403 	uma_zfree(zone, cl);
1404 
1405 	return (EXT_FREE_OK);
1406 }
1407 
1408 static struct mbuf *
1409 get_fl_payload1(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf,
1410     int *fl_bufs_used)
1411 {
1412 	struct mbuf *m0, *m;
1413 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1414 	unsigned int nbuf, len;
1415 	int pack_boundary = is_t4(sc) ? t4_fl_pack : t5_fl_pack;
1416 
1417 	/*
1418 	 * No assertion for the fl lock because we don't need it.  This routine
1419 	 * is called only from the rx interrupt handler and it only updates
1420 	 * fl->cidx.  (Contrast that with fl->pidx/fl->needed which could be
1421 	 * updated in the rx interrupt handler or the starvation helper routine.
1422 	 * That's why code that manipulates fl->pidx/fl->needed needs the fl
1423 	 * lock but this routine does not).
1424 	 */
1425 
1426 	KASSERT(fl->flags & FL_BUF_PACKING,
1427 	    ("%s: buffer packing disabled for fl %p", __func__, fl));
1428 
1429 	len = G_RSPD_LEN(len_newbuf);
1430 
1431 	if ((len_newbuf & F_RSPD_NEWBUF) == 0) {
1432 		KASSERT(fl->rx_offset > 0,
1433 		    ("%s: packed frame but driver at offset=0", __func__));
1434 
1435 		/* A packed frame is guaranteed to fit entirely in this buf. */
1436 		KASSERT(FL_BUF_SIZE(sc, sd->tag_idx) - fl->rx_offset >= len,
1437 		    ("%s: packing error.  bufsz=%u, offset=%u, len=%u",
1438 		    __func__, FL_BUF_SIZE(sc, sd->tag_idx), fl->rx_offset,
1439 		    len));
1440 
1441 		m0 = get_mbuf_from_stash(fl);
1442 		if (m0 == NULL ||
1443 		    m_init(m0, NULL, 0, M_NOWAIT, MT_DATA, M_PKTHDR) != 0) {
1444 			return_mbuf_to_stash(fl, m0);
1445 			return (NULL);
1446 		}
1447 
1448 		bus_dmamap_sync(fl->tag[sd->tag_idx], sd->map,
1449 		    BUS_DMASYNC_POSTREAD);
1450 		if (len < RX_COPY_THRESHOLD) {
1451 #ifdef T4_PKT_TIMESTAMP
1452 			/* Leave room for a timestamp */
1453 			m0->m_data += 8;
1454 #endif
1455 			bcopy(sd->cl + fl->rx_offset, mtod(m0, caddr_t), len);
1456 			m0->m_pkthdr.len = len;
1457 			m0->m_len = len;
1458 		} else {
1459 			m0->m_pkthdr.len = len;
1460 			m0->m_len = len;
1461 			m_extaddref(m0, sd->cl + fl->rx_offset,
1462 			    roundup2(m0->m_len, fl_pad),
1463 			    find_buf_refcnt(sd->cl), rxb_free,
1464 			    FL_BUF_ZONE(sc, sd->tag_idx), sd->cl);
1465 		}
1466 		fl->rx_offset += len;
1467 		fl->rx_offset = roundup2(fl->rx_offset, fl_pad);
1468 		fl->rx_offset = roundup2(fl->rx_offset, pack_boundary);
1469 		if (fl->rx_offset >= FL_BUF_SIZE(sc, sd->tag_idx)) {
1470 			fl->rx_offset = 0;
1471 			(*fl_bufs_used) += 1;
1472 			if (__predict_false(++fl->cidx == fl->cap))
1473 				fl->cidx = 0;
1474 		}
1475 
1476 		return (m0);
1477 	}
1478 
1479 	KASSERT(len_newbuf & F_RSPD_NEWBUF,
1480 	    ("%s: only new buffer handled here", __func__));
1481 
1482 	nbuf = 0;
1483 
1484 	/*
1485 	 * Move to the start of the next buffer if we are still in the middle of
1486 	 * some buffer.  This is the case where there was some room left in the
1487 	 * previous buffer but not enough to fit this frame in its entirety.
1488 	 */
1489 	if (fl->rx_offset > 0) {
1490 		KASSERT(roundup2(len, fl_pad) > FL_BUF_SIZE(sc, sd->tag_idx) -
1491 		    fl->rx_offset, ("%s: frame (%u bytes) should have fit at "
1492 		    "cidx %u offset %u bufsize %u", __func__, len, fl->cidx,
1493 		    fl->rx_offset, FL_BUF_SIZE(sc, sd->tag_idx)));
1494 		nbuf++;
1495 		fl->rx_offset = 0;
1496 		sd++;
1497 		if (__predict_false(++fl->cidx == fl->cap)) {
1498 			sd = fl->sdesc;
1499 			fl->cidx = 0;
1500 		}
1501 	}
1502 
1503 	m0 = find_buf_mbuf(sd->cl);
1504 	if (m_init(m0, NULL, 0, M_NOWAIT, MT_DATA, M_PKTHDR | M_NOFREE))
1505 		goto done;
1506 	bus_dmamap_sync(fl->tag[sd->tag_idx], sd->map, BUS_DMASYNC_POSTREAD);
1507 	m0->m_len = min(len, FL_BUF_SIZE(sc, sd->tag_idx));
1508 	m_extaddref(m0, sd->cl, roundup2(m0->m_len, fl_pad),
1509 	    find_buf_refcnt(sd->cl), rxb_free, FL_BUF_ZONE(sc, sd->tag_idx),
1510 	    sd->cl);
1511 	m0->m_pkthdr.len = len;
1512 
1513 	fl->rx_offset = roundup2(m0->m_len, fl_pad);
1514 	fl->rx_offset = roundup2(fl->rx_offset, pack_boundary);
1515 	if (fl->rx_offset >= FL_BUF_SIZE(sc, sd->tag_idx)) {
1516 		fl->rx_offset = 0;
1517 		nbuf++;
1518 		sd++;
1519 		if (__predict_false(++fl->cidx == fl->cap)) {
1520 			sd = fl->sdesc;
1521 			fl->cidx = 0;
1522 		}
1523 	}
1524 
1525 	m = m0;
1526 	len -= m->m_len;
1527 
1528 	while (len > 0) {
1529 		m->m_next = find_buf_mbuf(sd->cl);
1530 		m = m->m_next;
1531 
1532 		bus_dmamap_sync(fl->tag[sd->tag_idx], sd->map,
1533 		    BUS_DMASYNC_POSTREAD);
1534 
1535 		/* m_init for !M_PKTHDR can't fail so don't bother */
1536 		m_init(m, NULL, 0, M_NOWAIT, MT_DATA, M_NOFREE);
1537 		m->m_len = min(len, FL_BUF_SIZE(sc, sd->tag_idx));
1538 		m_extaddref(m, sd->cl, roundup2(m->m_len, fl_pad),
1539 		    find_buf_refcnt(sd->cl), rxb_free,
1540 		    FL_BUF_ZONE(sc, sd->tag_idx), sd->cl);
1541 
1542 		fl->rx_offset = roundup2(m->m_len, fl_pad);
1543 		fl->rx_offset = roundup2(fl->rx_offset, pack_boundary);
1544 		if (fl->rx_offset >= FL_BUF_SIZE(sc, sd->tag_idx)) {
1545 			fl->rx_offset = 0;
1546 			nbuf++;
1547 			sd++;
1548 			if (__predict_false(++fl->cidx == fl->cap)) {
1549 				sd = fl->sdesc;
1550 				fl->cidx = 0;
1551 			}
1552 		}
1553 
1554 		len -= m->m_len;
1555 	}
1556 done:
1557 	(*fl_bufs_used) += nbuf;
1558 	return (m0);
1559 }
1560 
1561 static struct mbuf *
1562 get_fl_payload2(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf,
1563     int *fl_bufs_used)
1564 {
1565 	struct mbuf *m0, *m;
1566 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1567 	unsigned int nbuf, len;
1568 
1569 	/*
1570 	 * No assertion for the fl lock because we don't need it.  This routine
1571 	 * is called only from the rx interrupt handler and it only updates
1572 	 * fl->cidx.  (Contrast that with fl->pidx/fl->needed which could be
1573 	 * updated in the rx interrupt handler or the starvation helper routine.
1574 	 * That's why code that manipulates fl->pidx/fl->needed needs the fl
1575 	 * lock but this routine does not).
1576 	 */
1577 
1578 	KASSERT((fl->flags & FL_BUF_PACKING) == 0,
1579 	    ("%s: buffer packing enabled for fl %p", __func__, fl));
1580 	if (__predict_false((len_newbuf & F_RSPD_NEWBUF) == 0))
1581 		panic("%s: cannot handle packed frames", __func__);
1582 	len = G_RSPD_LEN(len_newbuf);
1583 
1584 	/*
1585 	 * We never want to run out of mbufs in between a frame when a frame
1586 	 * spans multiple fl buffers.  If the fl's mbuf stash isn't full and
1587 	 * can't be filled up to the brim then fail early.
1588 	 */
1589 	if (len > FL_BUF_SIZE(sc, sd->tag_idx) && fill_mbuf_stash(fl) != 0)
1590 		return (NULL);
1591 
1592 	m0 = get_mbuf_from_stash(fl);
1593 	if (m0 == NULL ||
1594 	    m_init(m0, NULL, 0, M_NOWAIT, MT_DATA, M_PKTHDR) != 0) {
1595 		return_mbuf_to_stash(fl, m0);
1596 		return (NULL);
1597 	}
1598 
1599 	bus_dmamap_sync(fl->tag[sd->tag_idx], sd->map, BUS_DMASYNC_POSTREAD);
1600 
1601 	if (len < RX_COPY_THRESHOLD) {
1602 #ifdef T4_PKT_TIMESTAMP
1603 		/* Leave room for a timestamp */
1604 		m0->m_data += 8;
1605 #endif
1606 		/* copy data to mbuf, buffer will be recycled */
1607 		bcopy(sd->cl, mtod(m0, caddr_t), len);
1608 		m0->m_len = len;
1609 	} else {
1610 		bus_dmamap_unload(fl->tag[sd->tag_idx], sd->map);
1611 		m_cljset(m0, sd->cl, FL_BUF_TYPE(sc, sd->tag_idx));
1612 		sd->cl = NULL;	/* consumed */
1613 		m0->m_len = min(len, FL_BUF_SIZE(sc, sd->tag_idx));
1614 	}
1615 	m0->m_pkthdr.len = len;
1616 
1617 	sd++;
1618 	if (__predict_false(++fl->cidx == fl->cap)) {
1619 		sd = fl->sdesc;
1620 		fl->cidx = 0;
1621 	}
1622 
1623 	m = m0;
1624 	len -= m->m_len;
1625 	nbuf = 1;	/* # of fl buffers used */
1626 
1627 	while (len > 0) {
1628 		/* Can't fail, we checked earlier that the stash was full. */
1629 		m->m_next = get_mbuf_from_stash(fl);
1630 		m = m->m_next;
1631 
1632 		bus_dmamap_sync(fl->tag[sd->tag_idx], sd->map,
1633 		    BUS_DMASYNC_POSTREAD);
1634 
1635 		/* m_init for !M_PKTHDR can't fail so don't bother */
1636 		m_init(m, NULL, 0, M_NOWAIT, MT_DATA, 0);
1637 		if (len <= MLEN) {
1638 			bcopy(sd->cl, mtod(m, caddr_t), len);
1639 			m->m_len = len;
1640 		} else {
1641 			bus_dmamap_unload(fl->tag[sd->tag_idx], sd->map);
1642 			m_cljset(m, sd->cl, FL_BUF_TYPE(sc, sd->tag_idx));
1643 			sd->cl = NULL;	/* consumed */
1644 			m->m_len = min(len, FL_BUF_SIZE(sc, sd->tag_idx));
1645 		}
1646 
1647 		sd++;
1648 		if (__predict_false(++fl->cidx == fl->cap)) {
1649 			sd = fl->sdesc;
1650 			fl->cidx = 0;
1651 		}
1652 
1653 		len -= m->m_len;
1654 		nbuf++;
1655 	}
1656 
1657 	(*fl_bufs_used) += nbuf;
1658 
1659 	return (m0);
1660 }
1661 
1662 static int
1663 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
1664 {
1665 	struct sge_rxq *rxq = iq_to_rxq(iq);
1666 	struct ifnet *ifp = rxq->ifp;
1667 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
1668 #if defined(INET) || defined(INET6)
1669 	struct lro_ctrl *lro = &rxq->lro;
1670 #endif
1671 
1672 	KASSERT(m0 != NULL, ("%s: no payload with opcode %02x", __func__,
1673 	    rss->opcode));
1674 
1675 	m0->m_pkthdr.len -= fl_pktshift;
1676 	m0->m_len -= fl_pktshift;
1677 	m0->m_data += fl_pktshift;
1678 
1679 	m0->m_pkthdr.rcvif = ifp;
1680 	m0->m_flags |= M_FLOWID;
1681 	m0->m_pkthdr.flowid = rss->hash_val;
1682 
1683 	if (cpl->csum_calc && !cpl->err_vec) {
1684 		if (ifp->if_capenable & IFCAP_RXCSUM &&
1685 		    cpl->l2info & htobe32(F_RXF_IP)) {
1686 			m0->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
1687 			    CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
1688 			rxq->rxcsum++;
1689 		} else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
1690 		    cpl->l2info & htobe32(F_RXF_IP6)) {
1691 			m0->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
1692 			    CSUM_PSEUDO_HDR);
1693 			rxq->rxcsum++;
1694 		}
1695 
1696 		if (__predict_false(cpl->ip_frag))
1697 			m0->m_pkthdr.csum_data = be16toh(cpl->csum);
1698 		else
1699 			m0->m_pkthdr.csum_data = 0xffff;
1700 	}
1701 
1702 	if (cpl->vlan_ex) {
1703 		m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
1704 		m0->m_flags |= M_VLANTAG;
1705 		rxq->vlan_extraction++;
1706 	}
1707 
1708 #if defined(INET) || defined(INET6)
1709 	if (cpl->l2info & htobe32(F_RXF_LRO) &&
1710 	    iq->flags & IQ_LRO_ENABLED &&
1711 	    tcp_lro_rx(lro, m0, 0) == 0) {
1712 		/* queued for LRO */
1713 	} else
1714 #endif
1715 	ifp->if_input(ifp, m0);
1716 
1717 	return (0);
1718 }
1719 
1720 /*
1721  * Doesn't fail.  Holds on to work requests it can't send right away.
1722  */
1723 void
1724 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
1725 {
1726 	struct sge_eq *eq = &wrq->eq;
1727 	int can_reclaim;
1728 	caddr_t dst;
1729 
1730 	TXQ_LOCK_ASSERT_OWNED(wrq);
1731 #ifdef TCP_OFFLOAD
1732 	KASSERT((eq->flags & EQ_TYPEMASK) == EQ_OFLD ||
1733 	    (eq->flags & EQ_TYPEMASK) == EQ_CTRL,
1734 	    ("%s: eq type %d", __func__, eq->flags & EQ_TYPEMASK));
1735 #else
1736 	KASSERT((eq->flags & EQ_TYPEMASK) == EQ_CTRL,
1737 	    ("%s: eq type %d", __func__, eq->flags & EQ_TYPEMASK));
1738 #endif
1739 
1740 	if (__predict_true(wr != NULL))
1741 		STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
1742 
1743 	can_reclaim = reclaimable(eq);
1744 	if (__predict_false(eq->flags & EQ_STALLED)) {
1745 		if (can_reclaim < tx_resume_threshold(eq))
1746 			return;
1747 		eq->flags &= ~EQ_STALLED;
1748 		eq->unstalled++;
1749 	}
1750 	eq->cidx += can_reclaim;
1751 	eq->avail += can_reclaim;
1752 	if (__predict_false(eq->cidx >= eq->cap))
1753 		eq->cidx -= eq->cap;
1754 
1755 	while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL) {
1756 		int ndesc;
1757 
1758 		if (__predict_false(wr->wr_len < 0 ||
1759 		    wr->wr_len > SGE_MAX_WR_LEN || (wr->wr_len & 0x7))) {
1760 
1761 #ifdef INVARIANTS
1762 			panic("%s: work request with length %d", __func__,
1763 			    wr->wr_len);
1764 #endif
1765 #ifdef KDB
1766 			kdb_backtrace();
1767 #endif
1768 			log(LOG_ERR, "%s: %s work request with length %d",
1769 			    device_get_nameunit(sc->dev), __func__, wr->wr_len);
1770 			STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
1771 			free_wrqe(wr);
1772 			continue;
1773 		}
1774 
1775 		ndesc = howmany(wr->wr_len, EQ_ESIZE);
1776 		if (eq->avail < ndesc) {
1777 			wrq->no_desc++;
1778 			break;
1779 		}
1780 
1781 		dst = (void *)&eq->desc[eq->pidx];
1782 		copy_to_txd(eq, wrtod(wr), &dst, wr->wr_len);
1783 
1784 		eq->pidx += ndesc;
1785 		eq->avail -= ndesc;
1786 		if (__predict_false(eq->pidx >= eq->cap))
1787 			eq->pidx -= eq->cap;
1788 
1789 		eq->pending += ndesc;
1790 		if (eq->pending >= 8)
1791 			ring_eq_db(sc, eq);
1792 
1793 		wrq->tx_wrs++;
1794 		STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
1795 		free_wrqe(wr);
1796 
1797 		if (eq->avail < 8) {
1798 			can_reclaim = reclaimable(eq);
1799 			eq->cidx += can_reclaim;
1800 			eq->avail += can_reclaim;
1801 			if (__predict_false(eq->cidx >= eq->cap))
1802 				eq->cidx -= eq->cap;
1803 		}
1804 	}
1805 
1806 	if (eq->pending)
1807 		ring_eq_db(sc, eq);
1808 
1809 	if (wr != NULL) {
1810 		eq->flags |= EQ_STALLED;
1811 		if (callout_pending(&eq->tx_callout) == 0)
1812 			callout_reset(&eq->tx_callout, 1, t4_tx_callout, eq);
1813 	}
1814 }
1815 
1816 /* Per-packet header in a coalesced tx WR, before the SGL starts (in flits) */
1817 #define TXPKTS_PKT_HDR ((\
1818     sizeof(struct ulp_txpkt) + \
1819     sizeof(struct ulptx_idata) + \
1820     sizeof(struct cpl_tx_pkt_core) \
1821     ) / 8)
1822 
1823 /* Header of a coalesced tx WR, before SGL of first packet (in flits) */
1824 #define TXPKTS_WR_HDR (\
1825     sizeof(struct fw_eth_tx_pkts_wr) / 8 + \
1826     TXPKTS_PKT_HDR)
1827 
1828 /* Header of a tx WR, before SGL of first packet (in flits) */
1829 #define TXPKT_WR_HDR ((\
1830     sizeof(struct fw_eth_tx_pkt_wr) + \
1831     sizeof(struct cpl_tx_pkt_core) \
1832     ) / 8 )
1833 
1834 /* Header of a tx LSO WR, before SGL of first packet (in flits) */
1835 #define TXPKT_LSO_WR_HDR ((\
1836     sizeof(struct fw_eth_tx_pkt_wr) + \
1837     sizeof(struct cpl_tx_pkt_lso_core) + \
1838     sizeof(struct cpl_tx_pkt_core) \
1839     ) / 8 )
1840 
1841 int
1842 t4_eth_tx(struct ifnet *ifp, struct sge_txq *txq, struct mbuf *m)
1843 {
1844 	struct port_info *pi = (void *)ifp->if_softc;
1845 	struct adapter *sc = pi->adapter;
1846 	struct sge_eq *eq = &txq->eq;
1847 	struct buf_ring *br = txq->br;
1848 	struct mbuf *next;
1849 	int rc, coalescing, can_reclaim;
1850 	struct txpkts txpkts;
1851 	struct sgl sgl;
1852 
1853 	TXQ_LOCK_ASSERT_OWNED(txq);
1854 	KASSERT(m, ("%s: called with nothing to do.", __func__));
1855 	KASSERT((eq->flags & EQ_TYPEMASK) == EQ_ETH,
1856 	    ("%s: eq type %d", __func__, eq->flags & EQ_TYPEMASK));
1857 
1858 	prefetch(&eq->desc[eq->pidx]);
1859 	prefetch(&txq->sdesc[eq->pidx]);
1860 
1861 	txpkts.npkt = 0;/* indicates there's nothing in txpkts */
1862 	coalescing = 0;
1863 
1864 	can_reclaim = reclaimable(eq);
1865 	if (__predict_false(eq->flags & EQ_STALLED)) {
1866 		if (can_reclaim < tx_resume_threshold(eq)) {
1867 			txq->m = m;
1868 			return (0);
1869 		}
1870 		eq->flags &= ~EQ_STALLED;
1871 		eq->unstalled++;
1872 	}
1873 
1874 	if (__predict_false(eq->flags & EQ_DOOMED)) {
1875 		m_freem(m);
1876 		while ((m = buf_ring_dequeue_sc(txq->br)) != NULL)
1877 			m_freem(m);
1878 		return (ENETDOWN);
1879 	}
1880 
1881 	if (eq->avail < 8 && can_reclaim)
1882 		reclaim_tx_descs(txq, can_reclaim, 32);
1883 
1884 	for (; m; m = next ? next : drbr_dequeue(ifp, br)) {
1885 
1886 		if (eq->avail < 8)
1887 			break;
1888 
1889 		next = m->m_nextpkt;
1890 		m->m_nextpkt = NULL;
1891 
1892 		if (next || buf_ring_peek(br))
1893 			coalescing = 1;
1894 
1895 		rc = get_pkt_sgl(txq, &m, &sgl, coalescing);
1896 		if (rc != 0) {
1897 			if (rc == ENOMEM) {
1898 
1899 				/* Short of resources, suspend tx */
1900 
1901 				m->m_nextpkt = next;
1902 				break;
1903 			}
1904 
1905 			/*
1906 			 * Unrecoverable error for this packet, throw it away
1907 			 * and move on to the next.  get_pkt_sgl may already
1908 			 * have freed m (it will be NULL in that case and the
1909 			 * m_freem here is still safe).
1910 			 */
1911 
1912 			m_freem(m);
1913 			continue;
1914 		}
1915 
1916 		if (coalescing &&
1917 		    add_to_txpkts(pi, txq, &txpkts, m, &sgl) == 0) {
1918 
1919 			/* Successfully absorbed into txpkts */
1920 
1921 			write_ulp_cpl_sgl(pi, txq, &txpkts, m, &sgl);
1922 			goto doorbell;
1923 		}
1924 
1925 		/*
1926 		 * We weren't coalescing to begin with, or current frame could
1927 		 * not be coalesced (add_to_txpkts flushes txpkts if a frame
1928 		 * given to it can't be coalesced).  Either way there should be
1929 		 * nothing in txpkts.
1930 		 */
1931 		KASSERT(txpkts.npkt == 0,
1932 		    ("%s: txpkts not empty: %d", __func__, txpkts.npkt));
1933 
1934 		/* We're sending out individual packets now */
1935 		coalescing = 0;
1936 
1937 		if (eq->avail < 8)
1938 			reclaim_tx_descs(txq, 0, 8);
1939 		rc = write_txpkt_wr(pi, txq, m, &sgl);
1940 		if (rc != 0) {
1941 
1942 			/* Short of hardware descriptors, suspend tx */
1943 
1944 			/*
1945 			 * This is an unlikely but expensive failure.  We've
1946 			 * done all the hard work (DMA mappings etc.) and now we
1947 			 * can't send out the packet.  What's worse, we have to
1948 			 * spend even more time freeing up everything in sgl.
1949 			 */
1950 			txq->no_desc++;
1951 			free_pkt_sgl(txq, &sgl);
1952 
1953 			m->m_nextpkt = next;
1954 			break;
1955 		}
1956 
1957 		ETHER_BPF_MTAP(ifp, m);
1958 		if (sgl.nsegs == 0)
1959 			m_freem(m);
1960 doorbell:
1961 		if (eq->pending >= 8)
1962 			ring_eq_db(sc, eq);
1963 
1964 		can_reclaim = reclaimable(eq);
1965 		if (can_reclaim >= 32)
1966 			reclaim_tx_descs(txq, can_reclaim, 64);
1967 	}
1968 
1969 	if (txpkts.npkt > 0)
1970 		write_txpkts_wr(txq, &txpkts);
1971 
1972 	/*
1973 	 * m not NULL means there was an error but we haven't thrown it away.
1974 	 * This can happen when we're short of tx descriptors (no_desc) or maybe
1975 	 * even DMA maps (no_dmamap).  Either way, a credit flush and reclaim
1976 	 * will get things going again.
1977 	 */
1978 	if (m && !(eq->flags & EQ_CRFLUSHED)) {
1979 		struct tx_sdesc *txsd = &txq->sdesc[eq->pidx];
1980 
1981 		/*
1982 		 * If EQ_CRFLUSHED is not set then we know we have at least one
1983 		 * available descriptor because any WR that reduces eq->avail to
1984 		 * 0 also sets EQ_CRFLUSHED.
1985 		 */
1986 		KASSERT(eq->avail > 0, ("%s: no space for eqflush.", __func__));
1987 
1988 		txsd->desc_used = 1;
1989 		txsd->credits = 0;
1990 		write_eqflush_wr(eq);
1991 	}
1992 	txq->m = m;
1993 
1994 	if (eq->pending)
1995 		ring_eq_db(sc, eq);
1996 
1997 	reclaim_tx_descs(txq, 0, 128);
1998 
1999 	if (eq->flags & EQ_STALLED && callout_pending(&eq->tx_callout) == 0)
2000 		callout_reset(&eq->tx_callout, 1, t4_tx_callout, eq);
2001 
2002 	return (0);
2003 }
2004 
2005 void
2006 t4_update_fl_bufsize(struct ifnet *ifp)
2007 {
2008 	struct port_info *pi = ifp->if_softc;
2009 	struct adapter *sc = pi->adapter;
2010 	struct sge_rxq *rxq;
2011 #ifdef TCP_OFFLOAD
2012 	struct sge_ofld_rxq *ofld_rxq;
2013 #endif
2014 	struct sge_fl *fl;
2015 	int i, bufsize;
2016 
2017 	bufsize = mtu_to_bufsize(ifp->if_mtu);
2018 	for_each_rxq(pi, i, rxq) {
2019 		fl = &rxq->fl;
2020 
2021 		FL_LOCK(fl);
2022 		set_fl_tag_idx(sc, fl, bufsize);
2023 		FL_UNLOCK(fl);
2024 	}
2025 #ifdef TCP_OFFLOAD
2026 	bufsize = mtu_to_bufsize_toe(pi->adapter, ifp->if_mtu);
2027 	for_each_ofld_rxq(pi, i, ofld_rxq) {
2028 		fl = &ofld_rxq->fl;
2029 
2030 		FL_LOCK(fl);
2031 		set_fl_tag_idx(sc, fl, bufsize);
2032 		FL_UNLOCK(fl);
2033 	}
2034 #endif
2035 }
2036 
2037 int
2038 can_resume_tx(struct sge_eq *eq)
2039 {
2040 	return (reclaimable(eq) >= tx_resume_threshold(eq));
2041 }
2042 
2043 static inline void
2044 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
2045     int qsize, int esize)
2046 {
2047 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
2048 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
2049 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
2050 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
2051 
2052 	iq->flags = 0;
2053 	iq->adapter = sc;
2054 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
2055 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
2056 	if (pktc_idx >= 0) {
2057 		iq->intr_params |= F_QINTR_CNT_EN;
2058 		iq->intr_pktc_idx = pktc_idx;
2059 	}
2060 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
2061 	iq->esize = max(esize, 16);		/* See FW_IQ_CMD/iqesize */
2062 }
2063 
2064 static inline void
2065 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int bufsize, int pack,
2066     char *name)
2067 {
2068 
2069 	fl->qsize = qsize;
2070 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
2071 	if (pack)
2072 		fl->flags |= FL_BUF_PACKING;
2073 	set_fl_tag_idx(sc, fl, bufsize);
2074 }
2075 
2076 static inline void
2077 init_eq(struct sge_eq *eq, int eqtype, int qsize, uint8_t tx_chan,
2078     uint16_t iqid, char *name)
2079 {
2080 	KASSERT(tx_chan < NCHAN, ("%s: bad tx channel %d", __func__, tx_chan));
2081 	KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
2082 
2083 	eq->flags = eqtype & EQ_TYPEMASK;
2084 	eq->tx_chan = tx_chan;
2085 	eq->iqid = iqid;
2086 	eq->qsize = qsize;
2087 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
2088 
2089 	TASK_INIT(&eq->tx_task, 0, t4_tx_task, eq);
2090 	callout_init(&eq->tx_callout, CALLOUT_MPSAFE);
2091 }
2092 
2093 static int
2094 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
2095     bus_dmamap_t *map, bus_addr_t *pa, void **va)
2096 {
2097 	int rc;
2098 
2099 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
2100 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
2101 	if (rc != 0) {
2102 		device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
2103 		goto done;
2104 	}
2105 
2106 	rc = bus_dmamem_alloc(*tag, va,
2107 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
2108 	if (rc != 0) {
2109 		device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
2110 		goto done;
2111 	}
2112 
2113 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
2114 	if (rc != 0) {
2115 		device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
2116 		goto done;
2117 	}
2118 done:
2119 	if (rc)
2120 		free_ring(sc, *tag, *map, *pa, *va);
2121 
2122 	return (rc);
2123 }
2124 
2125 static int
2126 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
2127     bus_addr_t pa, void *va)
2128 {
2129 	if (pa)
2130 		bus_dmamap_unload(tag, map);
2131 	if (va)
2132 		bus_dmamem_free(tag, va, map);
2133 	if (tag)
2134 		bus_dma_tag_destroy(tag);
2135 
2136 	return (0);
2137 }
2138 
2139 /*
2140  * Allocates the ring for an ingress queue and an optional freelist.  If the
2141  * freelist is specified it will be allocated and then associated with the
2142  * ingress queue.
2143  *
2144  * Returns errno on failure.  Resources allocated up to that point may still be
2145  * allocated.  Caller is responsible for cleanup in case this function fails.
2146  *
2147  * If the ingress queue will take interrupts directly (iq->flags & IQ_INTR) then
2148  * the intr_idx specifies the vector, starting from 0.  Otherwise it specifies
2149  * the abs_id of the ingress queue to which its interrupts should be forwarded.
2150  */
2151 static int
2152 alloc_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl,
2153     int intr_idx, int cong)
2154 {
2155 	int rc, i, cntxt_id;
2156 	size_t len;
2157 	struct fw_iq_cmd c;
2158 	struct adapter *sc = iq->adapter;
2159 	__be32 v = 0;
2160 
2161 	len = iq->qsize * iq->esize;
2162 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
2163 	    (void **)&iq->desc);
2164 	if (rc != 0)
2165 		return (rc);
2166 
2167 	bzero(&c, sizeof(c));
2168 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
2169 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
2170 	    V_FW_IQ_CMD_VFN(0));
2171 
2172 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
2173 	    FW_LEN16(c));
2174 
2175 	/* Special handling for firmware event queue */
2176 	if (iq == &sc->sge.fwq)
2177 		v |= F_FW_IQ_CMD_IQASYNCH;
2178 
2179 	if (iq->flags & IQ_INTR) {
2180 		KASSERT(intr_idx < sc->intr_count,
2181 		    ("%s: invalid direct intr_idx %d", __func__, intr_idx));
2182 	} else
2183 		v |= F_FW_IQ_CMD_IQANDST;
2184 	v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
2185 
2186 	c.type_to_iqandstindex = htobe32(v |
2187 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
2188 	    V_FW_IQ_CMD_VIID(pi->viid) |
2189 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
2190 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
2191 	    F_FW_IQ_CMD_IQGTSMODE |
2192 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
2193 	    V_FW_IQ_CMD_IQESIZE(ilog2(iq->esize) - 4));
2194 	c.iqsize = htobe16(iq->qsize);
2195 	c.iqaddr = htobe64(iq->ba);
2196 	if (cong >= 0)
2197 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
2198 
2199 	if (fl) {
2200 		mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
2201 
2202 		for (i = 0; i < FL_BUF_SIZES(sc); i++) {
2203 
2204 			/*
2205 			 * A freelist buffer must be 16 byte aligned as the SGE
2206 			 * uses the low 4 bits of the bus addr to figure out the
2207 			 * buffer size.
2208 			 */
2209 			rc = bus_dma_tag_create(sc->dmat, 16, 0,
2210 			    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL,
2211 			    FL_BUF_SIZE(sc, i), 1, FL_BUF_SIZE(sc, i),
2212 			    BUS_DMA_ALLOCNOW, NULL, NULL, &fl->tag[i]);
2213 			if (rc != 0) {
2214 				device_printf(sc->dev,
2215 				    "failed to create fl DMA tag[%d]: %d\n",
2216 				    i, rc);
2217 				return (rc);
2218 			}
2219 		}
2220 		len = fl->qsize * RX_FL_ESIZE;
2221 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
2222 		    &fl->ba, (void **)&fl->desc);
2223 		if (rc)
2224 			return (rc);
2225 
2226 		/* Allocate space for one software descriptor per buffer. */
2227 		fl->cap = (fl->qsize - spg_len / RX_FL_ESIZE) * 8;
2228 		rc = alloc_fl_sdesc(fl);
2229 		if (rc != 0) {
2230 			device_printf(sc->dev,
2231 			    "failed to setup fl software descriptors: %d\n",
2232 			    rc);
2233 			return (rc);
2234 		}
2235 		fl->needed = fl->cap;
2236 		fl->lowat = roundup2(sc->sge.fl_starve_threshold, 8);
2237 
2238 		c.iqns_to_fl0congen |=
2239 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
2240 			F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
2241 			(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
2242 			(fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
2243 			    0));
2244 		if (cong >= 0) {
2245 			c.iqns_to_fl0congen |=
2246 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
2247 				    F_FW_IQ_CMD_FL0CONGCIF |
2248 				    F_FW_IQ_CMD_FL0CONGEN);
2249 		}
2250 		c.fl0dcaen_to_fl0cidxfthresh =
2251 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(X_FETCHBURSTMIN_64B) |
2252 			V_FW_IQ_CMD_FL0FBMAX(X_FETCHBURSTMAX_512B));
2253 		c.fl0size = htobe16(fl->qsize);
2254 		c.fl0addr = htobe64(fl->ba);
2255 	}
2256 
2257 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
2258 	if (rc != 0) {
2259 		device_printf(sc->dev,
2260 		    "failed to create ingress queue: %d\n", rc);
2261 		return (rc);
2262 	}
2263 
2264 	iq->cdesc = iq->desc;
2265 	iq->cidx = 0;
2266 	iq->gen = 1;
2267 	iq->intr_next = iq->intr_params;
2268 	iq->cntxt_id = be16toh(c.iqid);
2269 	iq->abs_id = be16toh(c.physiqid);
2270 	iq->flags |= IQ_ALLOCATED;
2271 
2272 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
2273 	if (cntxt_id >= sc->sge.niq) {
2274 		panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
2275 		    cntxt_id, sc->sge.niq - 1);
2276 	}
2277 	sc->sge.iqmap[cntxt_id] = iq;
2278 
2279 	if (fl) {
2280 		fl->cntxt_id = be16toh(c.fl0id);
2281 		fl->pidx = fl->cidx = 0;
2282 
2283 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
2284 		if (cntxt_id >= sc->sge.neq) {
2285 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
2286 			    __func__, cntxt_id, sc->sge.neq - 1);
2287 		}
2288 		sc->sge.eqmap[cntxt_id] = (void *)fl;
2289 
2290 		FL_LOCK(fl);
2291 		/* Enough to make sure the SGE doesn't think it's starved */
2292 		refill_fl(sc, fl, fl->lowat);
2293 		FL_UNLOCK(fl);
2294 
2295 		iq->flags |= IQ_HAS_FL;
2296 	}
2297 
2298 	if (is_t5(sc) && cong >= 0) {
2299 		uint32_t param, val;
2300 
2301 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
2302 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
2303 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
2304 		if (cong == 0)
2305 			val = 1 << 19;
2306 		else {
2307 			val = 2 << 19;
2308 			for (i = 0; i < 4; i++) {
2309 				if (cong & (1 << i))
2310 					val |= 1 << (i << 2);
2311 			}
2312 		}
2313 
2314 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
2315 		if (rc != 0) {
2316 			/* report error but carry on */
2317 			device_printf(sc->dev,
2318 			    "failed to set congestion manager context for "
2319 			    "ingress queue %d: %d\n", iq->cntxt_id, rc);
2320 		}
2321 	}
2322 
2323 	/* Enable IQ interrupts */
2324 	atomic_store_rel_int(&iq->state, IQS_IDLE);
2325 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_SEINTARM(iq->intr_params) |
2326 	    V_INGRESSQID(iq->cntxt_id));
2327 
2328 	return (0);
2329 }
2330 
2331 static int
2332 free_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl)
2333 {
2334 	int i, rc;
2335 	struct adapter *sc = iq->adapter;
2336 	device_t dev;
2337 
2338 	if (sc == NULL)
2339 		return (0);	/* nothing to do */
2340 
2341 	dev = pi ? pi->dev : sc->dev;
2342 
2343 	if (iq->flags & IQ_ALLOCATED) {
2344 		rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
2345 		    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
2346 		    fl ? fl->cntxt_id : 0xffff, 0xffff);
2347 		if (rc != 0) {
2348 			device_printf(dev,
2349 			    "failed to free queue %p: %d\n", iq, rc);
2350 			return (rc);
2351 		}
2352 		iq->flags &= ~IQ_ALLOCATED;
2353 	}
2354 
2355 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
2356 
2357 	bzero(iq, sizeof(*iq));
2358 
2359 	if (fl) {
2360 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
2361 		    fl->desc);
2362 
2363 		if (fl->sdesc)
2364 			free_fl_sdesc(sc, fl);
2365 
2366 		for (i = 0; i < nitems(fl->mstash); i++) {
2367 			struct mbuf *m = fl->mstash[i];
2368 
2369 			if (m != NULL) {
2370 				m_init(m, NULL, 0, M_NOWAIT, MT_DATA, 0);
2371 				m_free(m);
2372 			}
2373 		}
2374 
2375 		if (mtx_initialized(&fl->fl_lock))
2376 			mtx_destroy(&fl->fl_lock);
2377 
2378 		for (i = 0; i < FL_BUF_SIZES(sc); i++) {
2379 			if (fl->tag[i])
2380 				bus_dma_tag_destroy(fl->tag[i]);
2381 		}
2382 
2383 		bzero(fl, sizeof(*fl));
2384 	}
2385 
2386 	return (0);
2387 }
2388 
2389 static int
2390 alloc_fwq(struct adapter *sc)
2391 {
2392 	int rc, intr_idx;
2393 	struct sge_iq *fwq = &sc->sge.fwq;
2394 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
2395 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2396 
2397 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE, FW_IQ_ESIZE);
2398 	fwq->flags |= IQ_INTR;	/* always */
2399 	intr_idx = sc->intr_count > 1 ? 1 : 0;
2400 	rc = alloc_iq_fl(sc->port[0], fwq, NULL, intr_idx, -1);
2401 	if (rc != 0) {
2402 		device_printf(sc->dev,
2403 		    "failed to create firmware event queue: %d\n", rc);
2404 		return (rc);
2405 	}
2406 
2407 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD,
2408 	    NULL, "firmware event queue");
2409 	children = SYSCTL_CHILDREN(oid);
2410 
2411 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "abs_id",
2412 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->abs_id, 0, sysctl_uint16, "I",
2413 	    "absolute id of the queue");
2414 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cntxt_id",
2415 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->cntxt_id, 0, sysctl_uint16, "I",
2416 	    "SGE context id of the queue");
2417 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cidx",
2418 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->cidx, 0, sysctl_uint16, "I",
2419 	    "consumer index");
2420 
2421 	return (0);
2422 }
2423 
2424 static int
2425 free_fwq(struct adapter *sc)
2426 {
2427 	return free_iq_fl(NULL, &sc->sge.fwq, NULL);
2428 }
2429 
2430 static int
2431 alloc_mgmtq(struct adapter *sc)
2432 {
2433 	int rc;
2434 	struct sge_wrq *mgmtq = &sc->sge.mgmtq;
2435 	char name[16];
2436 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
2437 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2438 
2439 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "mgmtq", CTLFLAG_RD,
2440 	    NULL, "management queue");
2441 
2442 	snprintf(name, sizeof(name), "%s mgmtq", device_get_nameunit(sc->dev));
2443 	init_eq(&mgmtq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[0]->tx_chan,
2444 	    sc->sge.fwq.cntxt_id, name);
2445 	rc = alloc_wrq(sc, NULL, mgmtq, oid);
2446 	if (rc != 0) {
2447 		device_printf(sc->dev,
2448 		    "failed to create management queue: %d\n", rc);
2449 		return (rc);
2450 	}
2451 
2452 	return (0);
2453 }
2454 
2455 static int
2456 free_mgmtq(struct adapter *sc)
2457 {
2458 
2459 	return free_wrq(sc, &sc->sge.mgmtq);
2460 }
2461 
2462 static inline int
2463 tnl_cong(struct port_info *pi)
2464 {
2465 
2466 	if (cong_drop == -1)
2467 		return (-1);
2468 	else if (cong_drop == 1)
2469 		return (0);
2470 	else
2471 		return (1 << pi->tx_chan);
2472 }
2473 
2474 static int
2475 alloc_rxq(struct port_info *pi, struct sge_rxq *rxq, int intr_idx, int idx,
2476     struct sysctl_oid *oid)
2477 {
2478 	int rc;
2479 	struct sysctl_oid_list *children;
2480 	char name[16];
2481 
2482 	rc = alloc_iq_fl(pi, &rxq->iq, &rxq->fl, intr_idx, tnl_cong(pi));
2483 	if (rc != 0)
2484 		return (rc);
2485 
2486 	FL_LOCK(&rxq->fl);
2487 	refill_fl(pi->adapter, &rxq->fl, rxq->fl.needed / 8);
2488 	FL_UNLOCK(&rxq->fl);
2489 
2490 #if defined(INET) || defined(INET6)
2491 	rc = tcp_lro_init(&rxq->lro);
2492 	if (rc != 0)
2493 		return (rc);
2494 	rxq->lro.ifp = pi->ifp; /* also indicates LRO init'ed */
2495 
2496 	if (pi->ifp->if_capenable & IFCAP_LRO)
2497 		rxq->iq.flags |= IQ_LRO_ENABLED;
2498 #endif
2499 	rxq->ifp = pi->ifp;
2500 
2501 	children = SYSCTL_CHILDREN(oid);
2502 
2503 	snprintf(name, sizeof(name), "%d", idx);
2504 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
2505 	    NULL, "rx queue");
2506 	children = SYSCTL_CHILDREN(oid);
2507 
2508 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "abs_id",
2509 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.abs_id, 0, sysctl_uint16, "I",
2510 	    "absolute id of the queue");
2511 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cntxt_id",
2512 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cntxt_id, 0, sysctl_uint16, "I",
2513 	    "SGE context id of the queue");
2514 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cidx",
2515 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cidx, 0, sysctl_uint16, "I",
2516 	    "consumer index");
2517 #if defined(INET) || defined(INET6)
2518 	SYSCTL_ADD_INT(&pi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
2519 	    &rxq->lro.lro_queued, 0, NULL);
2520 	SYSCTL_ADD_INT(&pi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
2521 	    &rxq->lro.lro_flushed, 0, NULL);
2522 #endif
2523 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
2524 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
2525 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "vlan_extraction",
2526 	    CTLFLAG_RD, &rxq->vlan_extraction,
2527 	    "# of times hardware extracted 802.1Q tag");
2528 
2529 	children = SYSCTL_CHILDREN(oid);
2530 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "fl", CTLFLAG_RD,
2531 	    NULL, "freelist");
2532 	children = SYSCTL_CHILDREN(oid);
2533 
2534 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cntxt_id",
2535 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->fl.cntxt_id, 0, sysctl_uint16, "I",
2536 	    "SGE context id of the queue");
2537 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
2538 	    &rxq->fl.cidx, 0, "consumer index");
2539 	if (rxq->fl.flags & FL_BUF_PACKING) {
2540 		SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "rx_offset",
2541 		    CTLFLAG_RD, &rxq->fl.rx_offset, 0, "packing rx offset");
2542 	}
2543 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
2544 	    &rxq->fl.pidx, 0, "producer index");
2545 
2546 	return (rc);
2547 }
2548 
2549 static int
2550 free_rxq(struct port_info *pi, struct sge_rxq *rxq)
2551 {
2552 	int rc;
2553 
2554 #if defined(INET) || defined(INET6)
2555 	if (rxq->lro.ifp) {
2556 		tcp_lro_free(&rxq->lro);
2557 		rxq->lro.ifp = NULL;
2558 	}
2559 #endif
2560 
2561 	rc = free_iq_fl(pi, &rxq->iq, &rxq->fl);
2562 	if (rc == 0)
2563 		bzero(rxq, sizeof(*rxq));
2564 
2565 	return (rc);
2566 }
2567 
2568 #ifdef TCP_OFFLOAD
2569 static int
2570 alloc_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq,
2571     int intr_idx, int idx, struct sysctl_oid *oid)
2572 {
2573 	int rc;
2574 	struct sysctl_oid_list *children;
2575 	char name[16];
2576 
2577 	rc = alloc_iq_fl(pi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx,
2578 	    1 << pi->tx_chan);
2579 	if (rc != 0)
2580 		return (rc);
2581 
2582 	children = SYSCTL_CHILDREN(oid);
2583 
2584 	snprintf(name, sizeof(name), "%d", idx);
2585 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
2586 	    NULL, "rx queue");
2587 	children = SYSCTL_CHILDREN(oid);
2588 
2589 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "abs_id",
2590 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.abs_id, 0, sysctl_uint16,
2591 	    "I", "absolute id of the queue");
2592 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cntxt_id",
2593 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cntxt_id, 0, sysctl_uint16,
2594 	    "I", "SGE context id of the queue");
2595 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cidx",
2596 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cidx, 0, sysctl_uint16, "I",
2597 	    "consumer index");
2598 
2599 	children = SYSCTL_CHILDREN(oid);
2600 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "fl", CTLFLAG_RD,
2601 	    NULL, "freelist");
2602 	children = SYSCTL_CHILDREN(oid);
2603 
2604 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cntxt_id",
2605 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->fl.cntxt_id, 0, sysctl_uint16,
2606 	    "I", "SGE context id of the queue");
2607 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
2608 	    &ofld_rxq->fl.cidx, 0, "consumer index");
2609 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
2610 	    &ofld_rxq->fl.pidx, 0, "producer index");
2611 
2612 	return (rc);
2613 }
2614 
2615 static int
2616 free_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq)
2617 {
2618 	int rc;
2619 
2620 	rc = free_iq_fl(pi, &ofld_rxq->iq, &ofld_rxq->fl);
2621 	if (rc == 0)
2622 		bzero(ofld_rxq, sizeof(*ofld_rxq));
2623 
2624 	return (rc);
2625 }
2626 #endif
2627 
2628 static int
2629 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
2630 {
2631 	int rc, cntxt_id;
2632 	struct fw_eq_ctrl_cmd c;
2633 
2634 	bzero(&c, sizeof(c));
2635 
2636 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
2637 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
2638 	    V_FW_EQ_CTRL_CMD_VFN(0));
2639 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
2640 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
2641 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid)); /* XXX */
2642 	c.physeqid_pkd = htobe32(0);
2643 	c.fetchszm_to_iqid =
2644 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
2645 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
2646 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
2647 	c.dcaen_to_eqsize =
2648 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
2649 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
2650 		V_FW_EQ_CTRL_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
2651 		V_FW_EQ_CTRL_CMD_EQSIZE(eq->qsize));
2652 	c.eqaddr = htobe64(eq->ba);
2653 
2654 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
2655 	if (rc != 0) {
2656 		device_printf(sc->dev,
2657 		    "failed to create control queue %d: %d\n", eq->tx_chan, rc);
2658 		return (rc);
2659 	}
2660 	eq->flags |= EQ_ALLOCATED;
2661 
2662 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
2663 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
2664 	if (cntxt_id >= sc->sge.neq)
2665 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
2666 		cntxt_id, sc->sge.neq - 1);
2667 	sc->sge.eqmap[cntxt_id] = eq;
2668 
2669 	return (rc);
2670 }
2671 
2672 static int
2673 eth_eq_alloc(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
2674 {
2675 	int rc, cntxt_id;
2676 	struct fw_eq_eth_cmd c;
2677 
2678 	bzero(&c, sizeof(c));
2679 
2680 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
2681 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
2682 	    V_FW_EQ_ETH_CMD_VFN(0));
2683 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
2684 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
2685 	c.viid_pkd = htobe32(V_FW_EQ_ETH_CMD_VIID(pi->viid));
2686 	c.fetchszm_to_iqid =
2687 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
2688 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
2689 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
2690 	c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
2691 		      V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
2692 		      V_FW_EQ_ETH_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
2693 		      V_FW_EQ_ETH_CMD_EQSIZE(eq->qsize));
2694 	c.eqaddr = htobe64(eq->ba);
2695 
2696 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
2697 	if (rc != 0) {
2698 		device_printf(pi->dev,
2699 		    "failed to create Ethernet egress queue: %d\n", rc);
2700 		return (rc);
2701 	}
2702 	eq->flags |= EQ_ALLOCATED;
2703 
2704 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
2705 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
2706 	if (cntxt_id >= sc->sge.neq)
2707 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
2708 		cntxt_id, sc->sge.neq - 1);
2709 	sc->sge.eqmap[cntxt_id] = eq;
2710 
2711 	return (rc);
2712 }
2713 
2714 #ifdef TCP_OFFLOAD
2715 static int
2716 ofld_eq_alloc(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
2717 {
2718 	int rc, cntxt_id;
2719 	struct fw_eq_ofld_cmd c;
2720 
2721 	bzero(&c, sizeof(c));
2722 
2723 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
2724 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
2725 	    V_FW_EQ_OFLD_CMD_VFN(0));
2726 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
2727 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
2728 	c.fetchszm_to_iqid =
2729 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
2730 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
2731 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
2732 	c.dcaen_to_eqsize =
2733 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
2734 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
2735 		V_FW_EQ_OFLD_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
2736 		V_FW_EQ_OFLD_CMD_EQSIZE(eq->qsize));
2737 	c.eqaddr = htobe64(eq->ba);
2738 
2739 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
2740 	if (rc != 0) {
2741 		device_printf(pi->dev,
2742 		    "failed to create egress queue for TCP offload: %d\n", rc);
2743 		return (rc);
2744 	}
2745 	eq->flags |= EQ_ALLOCATED;
2746 
2747 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
2748 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
2749 	if (cntxt_id >= sc->sge.neq)
2750 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
2751 		cntxt_id, sc->sge.neq - 1);
2752 	sc->sge.eqmap[cntxt_id] = eq;
2753 
2754 	return (rc);
2755 }
2756 #endif
2757 
2758 static int
2759 alloc_eq(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
2760 {
2761 	int rc;
2762 	size_t len;
2763 
2764 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
2765 
2766 	len = eq->qsize * EQ_ESIZE;
2767 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
2768 	    &eq->ba, (void **)&eq->desc);
2769 	if (rc)
2770 		return (rc);
2771 
2772 	eq->cap = eq->qsize - spg_len / EQ_ESIZE;
2773 	eq->spg = (void *)&eq->desc[eq->cap];
2774 	eq->avail = eq->cap - 1;	/* one less to avoid cidx = pidx */
2775 	eq->pidx = eq->cidx = 0;
2776 	eq->doorbells = sc->doorbells;
2777 
2778 	switch (eq->flags & EQ_TYPEMASK) {
2779 	case EQ_CTRL:
2780 		rc = ctrl_eq_alloc(sc, eq);
2781 		break;
2782 
2783 	case EQ_ETH:
2784 		rc = eth_eq_alloc(sc, pi, eq);
2785 		break;
2786 
2787 #ifdef TCP_OFFLOAD
2788 	case EQ_OFLD:
2789 		rc = ofld_eq_alloc(sc, pi, eq);
2790 		break;
2791 #endif
2792 
2793 	default:
2794 		panic("%s: invalid eq type %d.", __func__,
2795 		    eq->flags & EQ_TYPEMASK);
2796 	}
2797 	if (rc != 0) {
2798 		device_printf(sc->dev,
2799 		    "failed to allocate egress queue(%d): %d",
2800 		    eq->flags & EQ_TYPEMASK, rc);
2801 	}
2802 
2803 	eq->tx_callout.c_cpu = eq->cntxt_id % mp_ncpus;
2804 
2805 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
2806 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
2807 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
2808 		uint32_t s_qpp = sc->sge.eq_s_qpp;
2809 		uint32_t mask = (1 << s_qpp) - 1;
2810 		volatile uint8_t *udb;
2811 
2812 		udb = sc->udbs_base + UDBS_DB_OFFSET;
2813 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
2814 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
2815 		if (eq->udb_qid > PAGE_SIZE / UDBS_SEG_SIZE)
2816 	    		clrbit(&eq->doorbells, DOORBELL_WCWR);
2817 		else {
2818 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
2819 			eq->udb_qid = 0;
2820 		}
2821 		eq->udb = (volatile void *)udb;
2822 	}
2823 
2824 	return (rc);
2825 }
2826 
2827 static int
2828 free_eq(struct adapter *sc, struct sge_eq *eq)
2829 {
2830 	int rc;
2831 
2832 	if (eq->flags & EQ_ALLOCATED) {
2833 		switch (eq->flags & EQ_TYPEMASK) {
2834 		case EQ_CTRL:
2835 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
2836 			    eq->cntxt_id);
2837 			break;
2838 
2839 		case EQ_ETH:
2840 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
2841 			    eq->cntxt_id);
2842 			break;
2843 
2844 #ifdef TCP_OFFLOAD
2845 		case EQ_OFLD:
2846 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
2847 			    eq->cntxt_id);
2848 			break;
2849 #endif
2850 
2851 		default:
2852 			panic("%s: invalid eq type %d.", __func__,
2853 			    eq->flags & EQ_TYPEMASK);
2854 		}
2855 		if (rc != 0) {
2856 			device_printf(sc->dev,
2857 			    "failed to free egress queue (%d): %d\n",
2858 			    eq->flags & EQ_TYPEMASK, rc);
2859 			return (rc);
2860 		}
2861 		eq->flags &= ~EQ_ALLOCATED;
2862 	}
2863 
2864 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
2865 
2866 	if (mtx_initialized(&eq->eq_lock))
2867 		mtx_destroy(&eq->eq_lock);
2868 
2869 	bzero(eq, sizeof(*eq));
2870 	return (0);
2871 }
2872 
2873 static int
2874 alloc_wrq(struct adapter *sc, struct port_info *pi, struct sge_wrq *wrq,
2875     struct sysctl_oid *oid)
2876 {
2877 	int rc;
2878 	struct sysctl_ctx_list *ctx = pi ? &pi->ctx : &sc->ctx;
2879 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2880 
2881 	rc = alloc_eq(sc, pi, &wrq->eq);
2882 	if (rc)
2883 		return (rc);
2884 
2885 	wrq->adapter = sc;
2886 	STAILQ_INIT(&wrq->wr_list);
2887 
2888 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
2889 	    &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
2890 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
2891 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I",
2892 	    "consumer index");
2893 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx",
2894 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I",
2895 	    "producer index");
2896 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs", CTLFLAG_RD,
2897 	    &wrq->tx_wrs, "# of work requests");
2898 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "no_desc", CTLFLAG_RD,
2899 	    &wrq->no_desc, 0,
2900 	    "# of times queue ran out of hardware descriptors");
2901 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "unstalled", CTLFLAG_RD,
2902 	    &wrq->eq.unstalled, 0, "# of times queue recovered after stall");
2903 
2904 
2905 	return (rc);
2906 }
2907 
2908 static int
2909 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
2910 {
2911 	int rc;
2912 
2913 	rc = free_eq(sc, &wrq->eq);
2914 	if (rc)
2915 		return (rc);
2916 
2917 	bzero(wrq, sizeof(*wrq));
2918 	return (0);
2919 }
2920 
2921 static int
2922 alloc_txq(struct port_info *pi, struct sge_txq *txq, int idx,
2923     struct sysctl_oid *oid)
2924 {
2925 	int rc;
2926 	struct adapter *sc = pi->adapter;
2927 	struct sge_eq *eq = &txq->eq;
2928 	char name[16];
2929 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2930 
2931 	rc = alloc_eq(sc, pi, eq);
2932 	if (rc)
2933 		return (rc);
2934 
2935 	txq->ifp = pi->ifp;
2936 
2937 	txq->sdesc = malloc(eq->cap * sizeof(struct tx_sdesc), M_CXGBE,
2938 	    M_ZERO | M_WAITOK);
2939 	txq->br = buf_ring_alloc(eq->qsize, M_CXGBE, M_WAITOK, &eq->eq_lock);
2940 
2941 	rc = bus_dma_tag_create(sc->dmat, 1, 0, BUS_SPACE_MAXADDR,
2942 	    BUS_SPACE_MAXADDR, NULL, NULL, 64 * 1024, TX_SGL_SEGS,
2943 	    BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL, NULL, &txq->tx_tag);
2944 	if (rc != 0) {
2945 		device_printf(sc->dev,
2946 		    "failed to create tx DMA tag: %d\n", rc);
2947 		return (rc);
2948 	}
2949 
2950 	/*
2951 	 * We can stuff ~10 frames in an 8-descriptor txpkts WR (8 is the SGE
2952 	 * limit for any WR).  txq->no_dmamap events shouldn't occur if maps is
2953 	 * sized for the worst case.
2954 	 */
2955 	rc = t4_alloc_tx_maps(&txq->txmaps, txq->tx_tag, eq->qsize * 10 / 8,
2956 	    M_WAITOK);
2957 	if (rc != 0) {
2958 		device_printf(sc->dev, "failed to setup tx DMA maps: %d\n", rc);
2959 		return (rc);
2960 	}
2961 
2962 	snprintf(name, sizeof(name), "%d", idx);
2963 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
2964 	    NULL, "tx queue");
2965 	children = SYSCTL_CHILDREN(oid);
2966 
2967 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
2968 	    &eq->cntxt_id, 0, "SGE context id of the queue");
2969 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cidx",
2970 	    CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I",
2971 	    "consumer index");
2972 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "pidx",
2973 	    CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I",
2974 	    "producer index");
2975 
2976 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
2977 	    &txq->txcsum, "# of times hardware assisted with checksum");
2978 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "vlan_insertion",
2979 	    CTLFLAG_RD, &txq->vlan_insertion,
2980 	    "# of times hardware inserted 802.1Q tag");
2981 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
2982 	    &txq->tso_wrs, "# of TSO work requests");
2983 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
2984 	    &txq->imm_wrs, "# of work requests with immediate data");
2985 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
2986 	    &txq->sgl_wrs, "# of work requests with direct SGL");
2987 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
2988 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
2989 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txpkts_wrs", CTLFLAG_RD,
2990 	    &txq->txpkts_wrs, "# of txpkts work requests (multiple pkts/WR)");
2991 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txpkts_pkts", CTLFLAG_RD,
2992 	    &txq->txpkts_pkts, "# of frames tx'd using txpkts work requests");
2993 
2994 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "br_drops", CTLFLAG_RD,
2995 	    &txq->br->br_drops, "# of drops in the buf_ring for this queue");
2996 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "no_dmamap", CTLFLAG_RD,
2997 	    &txq->no_dmamap, 0, "# of times txq ran out of DMA maps");
2998 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "no_desc", CTLFLAG_RD,
2999 	    &txq->no_desc, 0, "# of times txq ran out of hardware descriptors");
3000 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "egr_update", CTLFLAG_RD,
3001 	    &eq->egr_update, 0, "egress update notifications from the SGE");
3002 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "unstalled", CTLFLAG_RD,
3003 	    &eq->unstalled, 0, "# of times txq recovered after stall");
3004 
3005 	return (rc);
3006 }
3007 
3008 static int
3009 free_txq(struct port_info *pi, struct sge_txq *txq)
3010 {
3011 	int rc;
3012 	struct adapter *sc = pi->adapter;
3013 	struct sge_eq *eq = &txq->eq;
3014 
3015 	rc = free_eq(sc, eq);
3016 	if (rc)
3017 		return (rc);
3018 
3019 	free(txq->sdesc, M_CXGBE);
3020 
3021 	if (txq->txmaps.maps)
3022 		t4_free_tx_maps(&txq->txmaps, txq->tx_tag);
3023 
3024 	buf_ring_free(txq->br, M_CXGBE);
3025 
3026 	if (txq->tx_tag)
3027 		bus_dma_tag_destroy(txq->tx_tag);
3028 
3029 	bzero(txq, sizeof(*txq));
3030 	return (0);
3031 }
3032 
3033 static void
3034 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
3035 {
3036 	bus_addr_t *ba = arg;
3037 
3038 	KASSERT(nseg == 1,
3039 	    ("%s meant for single segment mappings only.", __func__));
3040 
3041 	*ba = error ? 0 : segs->ds_addr;
3042 }
3043 
3044 static inline bool
3045 is_new_response(const struct sge_iq *iq, struct rsp_ctrl **ctrl)
3046 {
3047 	*ctrl = (void *)((uintptr_t)iq->cdesc +
3048 	    (iq->esize - sizeof(struct rsp_ctrl)));
3049 
3050 	return (((*ctrl)->u.type_gen >> S_RSPD_GEN) == iq->gen);
3051 }
3052 
3053 static inline void
3054 iq_next(struct sge_iq *iq)
3055 {
3056 	iq->cdesc = (void *) ((uintptr_t)iq->cdesc + iq->esize);
3057 	if (__predict_false(++iq->cidx == iq->qsize - 1)) {
3058 		iq->cidx = 0;
3059 		iq->gen ^= 1;
3060 		iq->cdesc = iq->desc;
3061 	}
3062 }
3063 
3064 #define FL_HW_IDX(x) ((x) >> 3)
3065 static inline void
3066 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
3067 {
3068 	int ndesc = fl->pending / 8;
3069 	uint32_t v;
3070 
3071 	if (FL_HW_IDX(fl->pidx) == FL_HW_IDX(fl->cidx))
3072 		ndesc--;	/* hold back one credit */
3073 
3074 	if (ndesc <= 0)
3075 		return;		/* nothing to do */
3076 
3077 	v = F_DBPRIO | V_QID(fl->cntxt_id) | V_PIDX(ndesc);
3078 	if (is_t5(sc))
3079 		v |= F_DBTYPE;
3080 
3081 	wmb();
3082 
3083 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL), v);
3084 	fl->pending -= ndesc * 8;
3085 }
3086 
3087 /*
3088  * Fill up the freelist by upto nbufs and maybe ring its doorbell.
3089  *
3090  * Returns non-zero to indicate that it should be added to the list of starving
3091  * freelists.
3092  */
3093 static int
3094 refill_fl(struct adapter *sc, struct sge_fl *fl, int nbufs)
3095 {
3096 	__be64 *d = &fl->desc[fl->pidx];
3097 	struct fl_sdesc *sd = &fl->sdesc[fl->pidx];
3098 	bus_dma_tag_t tag;
3099 	bus_addr_t pa;
3100 	caddr_t cl;
3101 	int rc;
3102 
3103 	FL_LOCK_ASSERT_OWNED(fl);
3104 #ifdef INVARIANTS
3105 	if (fl->flags & FL_BUF_PACKING)
3106 		KASSERT(sd->tag_idx == 0,
3107 		    ("%s: expected tag 0 but found tag %d at pidx %u instead",
3108 		    __func__, sd->tag_idx, fl->pidx));
3109 #endif
3110 
3111 	if (nbufs > fl->needed)
3112 		nbufs = fl->needed;
3113 
3114 	while (nbufs--) {
3115 
3116 		if (sd->cl != NULL) {
3117 
3118 			KASSERT(*d == sd->ba_hwtag,
3119 			    ("%s: recyling problem at pidx %d",
3120 			    __func__, fl->pidx));
3121 
3122 			if (fl->flags & FL_BUF_PACKING) {
3123 				u_int *refcount = find_buf_refcnt(sd->cl);
3124 
3125 				if (atomic_fetchadd_int(refcount, -1) == 1) {
3126 					*refcount = 1;	/* reinstate */
3127 					d++;
3128 					goto recycled;
3129 				}
3130 				sd->cl = NULL;	/* gave up my reference */
3131 			} else {
3132 				/*
3133 				 * This happens when a frame small enough to fit
3134 				 * entirely in an mbuf was received in cl last
3135 				 * time.  We'd held on to cl and can reuse it
3136 				 * now.  Note that we reuse a cluster of the old
3137 				 * size if fl->tag_idx is no longer the same as
3138 				 * sd->tag_idx.
3139 				 */
3140 				d++;
3141 				goto recycled;
3142 			}
3143 		}
3144 
3145 		if (__predict_false(fl->tag_idx != sd->tag_idx)) {
3146 			bus_dmamap_t map;
3147 			bus_dma_tag_t newtag = fl->tag[fl->tag_idx];
3148 			bus_dma_tag_t oldtag = fl->tag[sd->tag_idx];
3149 
3150 			/*
3151 			 * An MTU change can get us here.  Discard the old map
3152 			 * which was created with the old tag, but only if
3153 			 * we're able to get a new one.
3154 			 */
3155 			rc = bus_dmamap_create(newtag, 0, &map);
3156 			if (rc == 0) {
3157 				bus_dmamap_destroy(oldtag, sd->map);
3158 				sd->map = map;
3159 				sd->tag_idx = fl->tag_idx;
3160 			}
3161 		}
3162 
3163 		tag = fl->tag[sd->tag_idx];
3164 
3165 		cl = uma_zalloc(FL_BUF_ZONE(sc, sd->tag_idx), M_NOWAIT);
3166 		if (cl == NULL)
3167 			break;
3168 		if (fl->flags & FL_BUF_PACKING) {
3169 			*find_buf_refcnt(cl) = 1;
3170 			cl += MSIZE;
3171 		}
3172 
3173 		rc = bus_dmamap_load(tag, sd->map, cl,
3174 		    FL_BUF_SIZE(sc, sd->tag_idx), oneseg_dma_callback, &pa, 0);
3175 		if (rc != 0 || pa == 0) {
3176 			fl->dmamap_failed++;
3177 			if (fl->flags & FL_BUF_PACKING)
3178 				cl -= MSIZE;
3179 			uma_zfree(FL_BUF_ZONE(sc, sd->tag_idx), cl);
3180 			break;
3181 		}
3182 
3183 		sd->cl = cl;
3184 		*d++ = htobe64(pa | FL_BUF_HWTAG(sc, sd->tag_idx));
3185 
3186 #ifdef INVARIANTS
3187 		sd->ba_hwtag = htobe64(pa | FL_BUF_HWTAG(sc, sd->tag_idx));
3188 #endif
3189 
3190 recycled:
3191 		fl->pending++;
3192 		fl->needed--;
3193 		sd++;
3194 		if (++fl->pidx == fl->cap) {
3195 			fl->pidx = 0;
3196 			sd = fl->sdesc;
3197 			d = fl->desc;
3198 		}
3199 	}
3200 
3201 	if (fl->pending >= 8)
3202 		ring_fl_db(sc, fl);
3203 
3204 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
3205 }
3206 
3207 /*
3208  * Attempt to refill all starving freelists.
3209  */
3210 static void
3211 refill_sfl(void *arg)
3212 {
3213 	struct adapter *sc = arg;
3214 	struct sge_fl *fl, *fl_temp;
3215 
3216 	mtx_lock(&sc->sfl_lock);
3217 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
3218 		FL_LOCK(fl);
3219 		refill_fl(sc, fl, 64);
3220 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
3221 			TAILQ_REMOVE(&sc->sfl, fl, link);
3222 			fl->flags &= ~FL_STARVING;
3223 		}
3224 		FL_UNLOCK(fl);
3225 	}
3226 
3227 	if (!TAILQ_EMPTY(&sc->sfl))
3228 		callout_schedule(&sc->sfl_callout, hz / 5);
3229 	mtx_unlock(&sc->sfl_lock);
3230 }
3231 
3232 static int
3233 alloc_fl_sdesc(struct sge_fl *fl)
3234 {
3235 	struct fl_sdesc *sd;
3236 	bus_dma_tag_t tag;
3237 	int i, rc;
3238 
3239 	fl->sdesc = malloc(fl->cap * sizeof(struct fl_sdesc), M_CXGBE,
3240 	    M_ZERO | M_WAITOK);
3241 
3242 	tag = fl->tag[fl->tag_idx];
3243 	sd = fl->sdesc;
3244 	for (i = 0; i < fl->cap; i++, sd++) {
3245 
3246 		sd->tag_idx = fl->tag_idx;
3247 		rc = bus_dmamap_create(tag, 0, &sd->map);
3248 		if (rc != 0)
3249 			goto failed;
3250 	}
3251 
3252 	return (0);
3253 failed:
3254 	while (--i >= 0) {
3255 		sd--;
3256 		bus_dmamap_destroy(tag, sd->map);
3257 	}
3258 	KASSERT(sd == fl->sdesc, ("%s: EDOOFUS", __func__));
3259 
3260 	free(fl->sdesc, M_CXGBE);
3261 	fl->sdesc = NULL;
3262 
3263 	return (rc);
3264 }
3265 
3266 static void
3267 free_fl_sdesc(struct adapter *sc, struct sge_fl *fl)
3268 {
3269 	struct fl_sdesc *sd;
3270 	int i;
3271 
3272 	sd = fl->sdesc;
3273 	for (i = 0; i < fl->cap; i++, sd++) {
3274 
3275 		if (sd->cl) {
3276 			bus_dmamap_unload(fl->tag[sd->tag_idx], sd->map);
3277 			uma_zfree(FL_BUF_ZONE(sc, sd->tag_idx), sd->cl);
3278 			sd->cl = NULL;
3279 		}
3280 
3281 		bus_dmamap_destroy(fl->tag[sd->tag_idx], sd->map);
3282 	}
3283 
3284 	free(fl->sdesc, M_CXGBE);
3285 	fl->sdesc = NULL;
3286 }
3287 
3288 int
3289 t4_alloc_tx_maps(struct tx_maps *txmaps, bus_dma_tag_t tx_tag, int count,
3290     int flags)
3291 {
3292 	struct tx_map *txm;
3293 	int i, rc;
3294 
3295 	txmaps->map_total = txmaps->map_avail = count;
3296 	txmaps->map_cidx = txmaps->map_pidx = 0;
3297 
3298 	txmaps->maps = malloc(count * sizeof(struct tx_map), M_CXGBE,
3299 	    M_ZERO | flags);
3300 
3301 	txm = txmaps->maps;
3302 	for (i = 0; i < count; i++, txm++) {
3303 		rc = bus_dmamap_create(tx_tag, 0, &txm->map);
3304 		if (rc != 0)
3305 			goto failed;
3306 	}
3307 
3308 	return (0);
3309 failed:
3310 	while (--i >= 0) {
3311 		txm--;
3312 		bus_dmamap_destroy(tx_tag, txm->map);
3313 	}
3314 	KASSERT(txm == txmaps->maps, ("%s: EDOOFUS", __func__));
3315 
3316 	free(txmaps->maps, M_CXGBE);
3317 	txmaps->maps = NULL;
3318 
3319 	return (rc);
3320 }
3321 
3322 void
3323 t4_free_tx_maps(struct tx_maps *txmaps, bus_dma_tag_t tx_tag)
3324 {
3325 	struct tx_map *txm;
3326 	int i;
3327 
3328 	txm = txmaps->maps;
3329 	for (i = 0; i < txmaps->map_total; i++, txm++) {
3330 
3331 		if (txm->m) {
3332 			bus_dmamap_unload(tx_tag, txm->map);
3333 			m_freem(txm->m);
3334 			txm->m = NULL;
3335 		}
3336 
3337 		bus_dmamap_destroy(tx_tag, txm->map);
3338 	}
3339 
3340 	free(txmaps->maps, M_CXGBE);
3341 	txmaps->maps = NULL;
3342 }
3343 
3344 /*
3345  * We'll do immediate data tx for non-TSO, but only when not coalescing.  We're
3346  * willing to use upto 2 hardware descriptors which means a maximum of 96 bytes
3347  * of immediate data.
3348  */
3349 #define IMM_LEN ( \
3350       2 * EQ_ESIZE \
3351     - sizeof(struct fw_eth_tx_pkt_wr) \
3352     - sizeof(struct cpl_tx_pkt_core))
3353 
3354 /*
3355  * Returns non-zero on failure, no need to cleanup anything in that case.
3356  *
3357  * Note 1: We always try to defrag the mbuf if required and return EFBIG only
3358  * if the resulting chain still won't fit in a tx descriptor.
3359  *
3360  * Note 2: We'll pullup the mbuf chain if TSO is requested and the first mbuf
3361  * does not have the TCP header in it.
3362  */
3363 static int
3364 get_pkt_sgl(struct sge_txq *txq, struct mbuf **fp, struct sgl *sgl,
3365     int sgl_only)
3366 {
3367 	struct mbuf *m = *fp;
3368 	struct tx_maps *txmaps;
3369 	struct tx_map *txm;
3370 	int rc, defragged = 0, n;
3371 
3372 	TXQ_LOCK_ASSERT_OWNED(txq);
3373 
3374 	if (m->m_pkthdr.tso_segsz)
3375 		sgl_only = 1;	/* Do not allow immediate data with LSO */
3376 
3377 start:	sgl->nsegs = 0;
3378 
3379 	if (m->m_pkthdr.len <= IMM_LEN && !sgl_only)
3380 		return (0);	/* nsegs = 0 tells caller to use imm. tx */
3381 
3382 	txmaps = &txq->txmaps;
3383 	if (txmaps->map_avail == 0) {
3384 		txq->no_dmamap++;
3385 		return (ENOMEM);
3386 	}
3387 	txm = &txmaps->maps[txmaps->map_pidx];
3388 
3389 	if (m->m_pkthdr.tso_segsz && m->m_len < 50) {
3390 		*fp = m_pullup(m, 50);
3391 		m = *fp;
3392 		if (m == NULL)
3393 			return (ENOBUFS);
3394 	}
3395 
3396 	rc = bus_dmamap_load_mbuf_sg(txq->tx_tag, txm->map, m, sgl->seg,
3397 	    &sgl->nsegs, BUS_DMA_NOWAIT);
3398 	if (rc == EFBIG && defragged == 0) {
3399 		m = m_defrag(m, M_NOWAIT);
3400 		if (m == NULL)
3401 			return (EFBIG);
3402 
3403 		defragged = 1;
3404 		*fp = m;
3405 		goto start;
3406 	}
3407 	if (rc != 0)
3408 		return (rc);
3409 
3410 	txm->m = m;
3411 	txmaps->map_avail--;
3412 	if (++txmaps->map_pidx == txmaps->map_total)
3413 		txmaps->map_pidx = 0;
3414 
3415 	KASSERT(sgl->nsegs > 0 && sgl->nsegs <= TX_SGL_SEGS,
3416 	    ("%s: bad DMA mapping (%d segments)", __func__, sgl->nsegs));
3417 
3418 	/*
3419 	 * Store the # of flits required to hold this frame's SGL in nflits.  An
3420 	 * SGL has a (ULPTX header + len0, addr0) tuple optionally followed by
3421 	 * multiple (len0 + len1, addr0, addr1) tuples.  If addr1 is not used
3422 	 * then len1 must be set to 0.
3423 	 */
3424 	n = sgl->nsegs - 1;
3425 	sgl->nflits = (3 * n) / 2 + (n & 1) + 2;
3426 
3427 	return (0);
3428 }
3429 
3430 
3431 /*
3432  * Releases all the txq resources used up in the specified sgl.
3433  */
3434 static int
3435 free_pkt_sgl(struct sge_txq *txq, struct sgl *sgl)
3436 {
3437 	struct tx_maps *txmaps;
3438 	struct tx_map *txm;
3439 
3440 	TXQ_LOCK_ASSERT_OWNED(txq);
3441 
3442 	if (sgl->nsegs == 0)
3443 		return (0);	/* didn't use any map */
3444 
3445 	txmaps = &txq->txmaps;
3446 
3447 	/* 1 pkt uses exactly 1 map, back it out */
3448 
3449 	txmaps->map_avail++;
3450 	if (txmaps->map_pidx > 0)
3451 		txmaps->map_pidx--;
3452 	else
3453 		txmaps->map_pidx = txmaps->map_total - 1;
3454 
3455 	txm = &txmaps->maps[txmaps->map_pidx];
3456 	bus_dmamap_unload(txq->tx_tag, txm->map);
3457 	txm->m = NULL;
3458 
3459 	return (0);
3460 }
3461 
3462 static int
3463 write_txpkt_wr(struct port_info *pi, struct sge_txq *txq, struct mbuf *m,
3464     struct sgl *sgl)
3465 {
3466 	struct sge_eq *eq = &txq->eq;
3467 	struct fw_eth_tx_pkt_wr *wr;
3468 	struct cpl_tx_pkt_core *cpl;
3469 	uint32_t ctrl;	/* used in many unrelated places */
3470 	uint64_t ctrl1;
3471 	int nflits, ndesc, pktlen;
3472 	struct tx_sdesc *txsd;
3473 	caddr_t dst;
3474 
3475 	TXQ_LOCK_ASSERT_OWNED(txq);
3476 
3477 	pktlen = m->m_pkthdr.len;
3478 
3479 	/*
3480 	 * Do we have enough flits to send this frame out?
3481 	 */
3482 	ctrl = sizeof(struct cpl_tx_pkt_core);
3483 	if (m->m_pkthdr.tso_segsz) {
3484 		nflits = TXPKT_LSO_WR_HDR;
3485 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
3486 	} else
3487 		nflits = TXPKT_WR_HDR;
3488 	if (sgl->nsegs > 0)
3489 		nflits += sgl->nflits;
3490 	else {
3491 		nflits += howmany(pktlen, 8);
3492 		ctrl += pktlen;
3493 	}
3494 	ndesc = howmany(nflits, 8);
3495 	if (ndesc > eq->avail)
3496 		return (ENOMEM);
3497 
3498 	/* Firmware work request header */
3499 	wr = (void *)&eq->desc[eq->pidx];
3500 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
3501 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
3502 	ctrl = V_FW_WR_LEN16(howmany(nflits, 2));
3503 	if (eq->avail == ndesc) {
3504 		if (!(eq->flags & EQ_CRFLUSHED)) {
3505 			ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ;
3506 			eq->flags |= EQ_CRFLUSHED;
3507 		}
3508 		eq->flags |= EQ_STALLED;
3509 	}
3510 
3511 	wr->equiq_to_len16 = htobe32(ctrl);
3512 	wr->r3 = 0;
3513 
3514 	if (m->m_pkthdr.tso_segsz) {
3515 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
3516 		struct ether_header *eh;
3517 		void *l3hdr;
3518 #if defined(INET) || defined(INET6)
3519 		struct tcphdr *tcp;
3520 #endif
3521 		uint16_t eh_type;
3522 
3523 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
3524 		    F_LSO_LAST_SLICE;
3525 
3526 		eh = mtod(m, struct ether_header *);
3527 		eh_type = ntohs(eh->ether_type);
3528 		if (eh_type == ETHERTYPE_VLAN) {
3529 			struct ether_vlan_header *evh = (void *)eh;
3530 
3531 			ctrl |= V_LSO_ETHHDR_LEN(1);
3532 			l3hdr = evh + 1;
3533 			eh_type = ntohs(evh->evl_proto);
3534 		} else
3535 			l3hdr = eh + 1;
3536 
3537 		switch (eh_type) {
3538 #ifdef INET6
3539 		case ETHERTYPE_IPV6:
3540 		{
3541 			struct ip6_hdr *ip6 = l3hdr;
3542 
3543 			/*
3544 			 * XXX-BZ For now we do not pretend to support
3545 			 * IPv6 extension headers.
3546 			 */
3547 			KASSERT(ip6->ip6_nxt == IPPROTO_TCP, ("%s: CSUM_TSO "
3548 			    "with ip6_nxt != TCP: %u", __func__, ip6->ip6_nxt));
3549 			tcp = (struct tcphdr *)(ip6 + 1);
3550 			ctrl |= F_LSO_IPV6;
3551 			ctrl |= V_LSO_IPHDR_LEN(sizeof(*ip6) >> 2) |
3552 			    V_LSO_TCPHDR_LEN(tcp->th_off);
3553 			break;
3554 		}
3555 #endif
3556 #ifdef INET
3557 		case ETHERTYPE_IP:
3558 		{
3559 			struct ip *ip = l3hdr;
3560 
3561 			tcp = (void *)((uintptr_t)ip + ip->ip_hl * 4);
3562 			ctrl |= V_LSO_IPHDR_LEN(ip->ip_hl) |
3563 			    V_LSO_TCPHDR_LEN(tcp->th_off);
3564 			break;
3565 		}
3566 #endif
3567 		default:
3568 			panic("%s: CSUM_TSO but no supported IP version "
3569 			    "(0x%04x)", __func__, eh_type);
3570 		}
3571 
3572 		lso->lso_ctrl = htobe32(ctrl);
3573 		lso->ipid_ofst = htobe16(0);
3574 		lso->mss = htobe16(m->m_pkthdr.tso_segsz);
3575 		lso->seqno_offset = htobe32(0);
3576 		lso->len = htobe32(pktlen);
3577 
3578 		cpl = (void *)(lso + 1);
3579 
3580 		txq->tso_wrs++;
3581 	} else
3582 		cpl = (void *)(wr + 1);
3583 
3584 	/* Checksum offload */
3585 	ctrl1 = 0;
3586 	if (!(m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO)))
3587 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
3588 	if (!(m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
3589 	    CSUM_TCP_IPV6 | CSUM_TSO)))
3590 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
3591 	if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
3592 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
3593 		txq->txcsum++;	/* some hardware assistance provided */
3594 
3595 	/* VLAN tag insertion */
3596 	if (m->m_flags & M_VLANTAG) {
3597 		ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
3598 		txq->vlan_insertion++;
3599 	}
3600 
3601 	/* CPL header */
3602 	cpl->ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3603 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf));
3604 	cpl->pack = 0;
3605 	cpl->len = htobe16(pktlen);
3606 	cpl->ctrl1 = htobe64(ctrl1);
3607 
3608 	/* Software descriptor */
3609 	txsd = &txq->sdesc[eq->pidx];
3610 	txsd->desc_used = ndesc;
3611 
3612 	eq->pending += ndesc;
3613 	eq->avail -= ndesc;
3614 	eq->pidx += ndesc;
3615 	if (eq->pidx >= eq->cap)
3616 		eq->pidx -= eq->cap;
3617 
3618 	/* SGL */
3619 	dst = (void *)(cpl + 1);
3620 	if (sgl->nsegs > 0) {
3621 		txsd->credits = 1;
3622 		txq->sgl_wrs++;
3623 		write_sgl_to_txd(eq, sgl, &dst);
3624 	} else {
3625 		txsd->credits = 0;
3626 		txq->imm_wrs++;
3627 		for (; m; m = m->m_next) {
3628 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
3629 #ifdef INVARIANTS
3630 			pktlen -= m->m_len;
3631 #endif
3632 		}
3633 #ifdef INVARIANTS
3634 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
3635 #endif
3636 
3637 	}
3638 
3639 	txq->txpkt_wrs++;
3640 	return (0);
3641 }
3642 
3643 /*
3644  * Returns 0 to indicate that m has been accepted into a coalesced tx work
3645  * request.  It has either been folded into txpkts or txpkts was flushed and m
3646  * has started a new coalesced work request (as the first frame in a fresh
3647  * txpkts).
3648  *
3649  * Returns non-zero to indicate a failure - caller is responsible for
3650  * transmitting m, if there was anything in txpkts it has been flushed.
3651  */
3652 static int
3653 add_to_txpkts(struct port_info *pi, struct sge_txq *txq, struct txpkts *txpkts,
3654     struct mbuf *m, struct sgl *sgl)
3655 {
3656 	struct sge_eq *eq = &txq->eq;
3657 	int can_coalesce;
3658 	struct tx_sdesc *txsd;
3659 	int flits;
3660 
3661 	TXQ_LOCK_ASSERT_OWNED(txq);
3662 
3663 	KASSERT(sgl->nsegs, ("%s: can't coalesce imm data", __func__));
3664 
3665 	if (txpkts->npkt > 0) {
3666 		flits = TXPKTS_PKT_HDR + sgl->nflits;
3667 		can_coalesce = m->m_pkthdr.tso_segsz == 0 &&
3668 		    txpkts->nflits + flits <= TX_WR_FLITS &&
3669 		    txpkts->nflits + flits <= eq->avail * 8 &&
3670 		    txpkts->plen + m->m_pkthdr.len < 65536;
3671 
3672 		if (can_coalesce) {
3673 			txpkts->npkt++;
3674 			txpkts->nflits += flits;
3675 			txpkts->plen += m->m_pkthdr.len;
3676 
3677 			txsd = &txq->sdesc[eq->pidx];
3678 			txsd->credits++;
3679 
3680 			return (0);
3681 		}
3682 
3683 		/*
3684 		 * Couldn't coalesce m into txpkts.  The first order of business
3685 		 * is to send txpkts on its way.  Then we'll revisit m.
3686 		 */
3687 		write_txpkts_wr(txq, txpkts);
3688 	}
3689 
3690 	/*
3691 	 * Check if we can start a new coalesced tx work request with m as
3692 	 * the first packet in it.
3693 	 */
3694 
3695 	KASSERT(txpkts->npkt == 0, ("%s: txpkts not empty", __func__));
3696 
3697 	flits = TXPKTS_WR_HDR + sgl->nflits;
3698 	can_coalesce = m->m_pkthdr.tso_segsz == 0 &&
3699 	    flits <= eq->avail * 8 && flits <= TX_WR_FLITS;
3700 
3701 	if (can_coalesce == 0)
3702 		return (EINVAL);
3703 
3704 	/*
3705 	 * Start a fresh coalesced tx WR with m as the first frame in it.
3706 	 */
3707 	txpkts->npkt = 1;
3708 	txpkts->nflits = flits;
3709 	txpkts->flitp = &eq->desc[eq->pidx].flit[2];
3710 	txpkts->plen = m->m_pkthdr.len;
3711 
3712 	txsd = &txq->sdesc[eq->pidx];
3713 	txsd->credits = 1;
3714 
3715 	return (0);
3716 }
3717 
3718 /*
3719  * Note that write_txpkts_wr can never run out of hardware descriptors (but
3720  * write_txpkt_wr can).  add_to_txpkts ensures that a frame is accepted for
3721  * coalescing only if sufficient hardware descriptors are available.
3722  */
3723 static void
3724 write_txpkts_wr(struct sge_txq *txq, struct txpkts *txpkts)
3725 {
3726 	struct sge_eq *eq = &txq->eq;
3727 	struct fw_eth_tx_pkts_wr *wr;
3728 	struct tx_sdesc *txsd;
3729 	uint32_t ctrl;
3730 	int ndesc;
3731 
3732 	TXQ_LOCK_ASSERT_OWNED(txq);
3733 
3734 	ndesc = howmany(txpkts->nflits, 8);
3735 
3736 	wr = (void *)&eq->desc[eq->pidx];
3737 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
3738 	ctrl = V_FW_WR_LEN16(howmany(txpkts->nflits, 2));
3739 	if (eq->avail == ndesc) {
3740 		if (!(eq->flags & EQ_CRFLUSHED)) {
3741 			ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ;
3742 			eq->flags |= EQ_CRFLUSHED;
3743 		}
3744 		eq->flags |= EQ_STALLED;
3745 	}
3746 	wr->equiq_to_len16 = htobe32(ctrl);
3747 	wr->plen = htobe16(txpkts->plen);
3748 	wr->npkt = txpkts->npkt;
3749 	wr->r3 = wr->type = 0;
3750 
3751 	/* Everything else already written */
3752 
3753 	txsd = &txq->sdesc[eq->pidx];
3754 	txsd->desc_used = ndesc;
3755 
3756 	KASSERT(eq->avail >= ndesc, ("%s: out of descriptors", __func__));
3757 
3758 	eq->pending += ndesc;
3759 	eq->avail -= ndesc;
3760 	eq->pidx += ndesc;
3761 	if (eq->pidx >= eq->cap)
3762 		eq->pidx -= eq->cap;
3763 
3764 	txq->txpkts_pkts += txpkts->npkt;
3765 	txq->txpkts_wrs++;
3766 	txpkts->npkt = 0;	/* emptied */
3767 }
3768 
3769 static inline void
3770 write_ulp_cpl_sgl(struct port_info *pi, struct sge_txq *txq,
3771     struct txpkts *txpkts, struct mbuf *m, struct sgl *sgl)
3772 {
3773 	struct ulp_txpkt *ulpmc;
3774 	struct ulptx_idata *ulpsc;
3775 	struct cpl_tx_pkt_core *cpl;
3776 	struct sge_eq *eq = &txq->eq;
3777 	uintptr_t flitp, start, end;
3778 	uint64_t ctrl;
3779 	caddr_t dst;
3780 
3781 	KASSERT(txpkts->npkt > 0, ("%s: txpkts is empty", __func__));
3782 
3783 	start = (uintptr_t)eq->desc;
3784 	end = (uintptr_t)eq->spg;
3785 
3786 	/* Checksum offload */
3787 	ctrl = 0;
3788 	if (!(m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO)))
3789 		ctrl |= F_TXPKT_IPCSUM_DIS;
3790 	if (!(m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
3791 	    CSUM_TCP_IPV6 | CSUM_TSO)))
3792 		ctrl |= F_TXPKT_L4CSUM_DIS;
3793 	if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
3794 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
3795 		txq->txcsum++;	/* some hardware assistance provided */
3796 
3797 	/* VLAN tag insertion */
3798 	if (m->m_flags & M_VLANTAG) {
3799 		ctrl |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
3800 		txq->vlan_insertion++;
3801 	}
3802 
3803 	/*
3804 	 * The previous packet's SGL must have ended at a 16 byte boundary (this
3805 	 * is required by the firmware/hardware).  It follows that flitp cannot
3806 	 * wrap around between the ULPTX master command and ULPTX subcommand (8
3807 	 * bytes each), and that it can not wrap around in the middle of the
3808 	 * cpl_tx_pkt_core either.
3809 	 */
3810 	flitp = (uintptr_t)txpkts->flitp;
3811 	KASSERT((flitp & 0xf) == 0,
3812 	    ("%s: last SGL did not end at 16 byte boundary: %p",
3813 	    __func__, txpkts->flitp));
3814 
3815 	/* ULP master command */
3816 	ulpmc = (void *)flitp;
3817 	ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0) |
3818 	    V_ULP_TXPKT_FID(eq->iqid));
3819 	ulpmc->len = htonl(howmany(sizeof(*ulpmc) + sizeof(*ulpsc) +
3820 	    sizeof(*cpl) + 8 * sgl->nflits, 16));
3821 
3822 	/* ULP subcommand */
3823 	ulpsc = (void *)(ulpmc + 1);
3824 	ulpsc->cmd_more = htobe32(V_ULPTX_CMD((u32)ULP_TX_SC_IMM) |
3825 	    F_ULP_TX_SC_MORE);
3826 	ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
3827 
3828 	flitp += sizeof(*ulpmc) + sizeof(*ulpsc);
3829 	if (flitp == end)
3830 		flitp = start;
3831 
3832 	/* CPL_TX_PKT */
3833 	cpl = (void *)flitp;
3834 	cpl->ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3835 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf));
3836 	cpl->pack = 0;
3837 	cpl->len = htobe16(m->m_pkthdr.len);
3838 	cpl->ctrl1 = htobe64(ctrl);
3839 
3840 	flitp += sizeof(*cpl);
3841 	if (flitp == end)
3842 		flitp = start;
3843 
3844 	/* SGL for this frame */
3845 	dst = (caddr_t)flitp;
3846 	txpkts->nflits += write_sgl_to_txd(eq, sgl, &dst);
3847 	txpkts->flitp = (void *)dst;
3848 
3849 	KASSERT(((uintptr_t)dst & 0xf) == 0,
3850 	    ("%s: SGL ends at %p (not a 16 byte boundary)", __func__, dst));
3851 }
3852 
3853 /*
3854  * If the SGL ends on an address that is not 16 byte aligned, this function will
3855  * add a 0 filled flit at the end.  It returns 1 in that case.
3856  */
3857 static int
3858 write_sgl_to_txd(struct sge_eq *eq, struct sgl *sgl, caddr_t *to)
3859 {
3860 	__be64 *flitp, *end;
3861 	struct ulptx_sgl *usgl;
3862 	bus_dma_segment_t *seg;
3863 	int i, padded;
3864 
3865 	KASSERT(sgl->nsegs > 0 && sgl->nflits > 0,
3866 	    ("%s: bad SGL - nsegs=%d, nflits=%d",
3867 	    __func__, sgl->nsegs, sgl->nflits));
3868 
3869 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
3870 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
3871 
3872 	flitp = (__be64 *)(*to);
3873 	end = flitp + sgl->nflits;
3874 	seg = &sgl->seg[0];
3875 	usgl = (void *)flitp;
3876 
3877 	/*
3878 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
3879 	 * ring, so we're at least 16 bytes away from the status page.  There is
3880 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
3881 	 */
3882 
3883 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
3884 	    V_ULPTX_NSGE(sgl->nsegs));
3885 	usgl->len0 = htobe32(seg->ds_len);
3886 	usgl->addr0 = htobe64(seg->ds_addr);
3887 	seg++;
3888 
3889 	if ((uintptr_t)end <= (uintptr_t)eq->spg) {
3890 
3891 		/* Won't wrap around at all */
3892 
3893 		for (i = 0; i < sgl->nsegs - 1; i++, seg++) {
3894 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ds_len);
3895 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ds_addr);
3896 		}
3897 		if (i & 1)
3898 			usgl->sge[i / 2].len[1] = htobe32(0);
3899 	} else {
3900 
3901 		/* Will wrap somewhere in the rest of the SGL */
3902 
3903 		/* 2 flits already written, write the rest flit by flit */
3904 		flitp = (void *)(usgl + 1);
3905 		for (i = 0; i < sgl->nflits - 2; i++) {
3906 			if ((uintptr_t)flitp == (uintptr_t)eq->spg)
3907 				flitp = (void *)eq->desc;
3908 			*flitp++ = get_flit(seg, sgl->nsegs - 1, i);
3909 		}
3910 		end = flitp;
3911 	}
3912 
3913 	if ((uintptr_t)end & 0xf) {
3914 		*(uint64_t *)end = 0;
3915 		end++;
3916 		padded = 1;
3917 	} else
3918 		padded = 0;
3919 
3920 	if ((uintptr_t)end == (uintptr_t)eq->spg)
3921 		*to = (void *)eq->desc;
3922 	else
3923 		*to = (void *)end;
3924 
3925 	return (padded);
3926 }
3927 
3928 static inline void
3929 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
3930 {
3931 	if (__predict_true((uintptr_t)(*to) + len <= (uintptr_t)eq->spg)) {
3932 		bcopy(from, *to, len);
3933 		(*to) += len;
3934 	} else {
3935 		int portion = (uintptr_t)eq->spg - (uintptr_t)(*to);
3936 
3937 		bcopy(from, *to, portion);
3938 		from += portion;
3939 		portion = len - portion;	/* remaining */
3940 		bcopy(from, (void *)eq->desc, portion);
3941 		(*to) = (caddr_t)eq->desc + portion;
3942 	}
3943 }
3944 
3945 static inline void
3946 ring_eq_db(struct adapter *sc, struct sge_eq *eq)
3947 {
3948 	u_int db, pending;
3949 
3950 	db = eq->doorbells;
3951 	pending = eq->pending;
3952 	if (pending > 1)
3953 		clrbit(&db, DOORBELL_WCWR);
3954 	eq->pending = 0;
3955 	wmb();
3956 
3957 	switch (ffs(db) - 1) {
3958 	case DOORBELL_UDB:
3959 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(pending));
3960 		return;
3961 
3962 	case DOORBELL_WCWR: {
3963 		volatile uint64_t *dst, *src;
3964 		int i;
3965 
3966 		/*
3967 		 * Queues whose 128B doorbell segment fits in the page do not
3968 		 * use relative qid (udb_qid is always 0).  Only queues with
3969 		 * doorbell segments can do WCWR.
3970 		 */
3971 		KASSERT(eq->udb_qid == 0 && pending == 1,
3972 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
3973 		    __func__, eq->doorbells, pending, eq->pidx, eq));
3974 
3975 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
3976 		    UDBS_DB_OFFSET);
3977 		i = eq->pidx ? eq->pidx - 1 : eq->cap - 1;
3978 		src = (void *)&eq->desc[i];
3979 		while (src != (void *)&eq->desc[i + 1])
3980 			*dst++ = *src++;
3981 		wmb();
3982 		return;
3983 	}
3984 
3985 	case DOORBELL_UDBWC:
3986 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(pending));
3987 		wmb();
3988 		return;
3989 
3990 	case DOORBELL_KDB:
3991 		t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
3992 		    V_QID(eq->cntxt_id) | V_PIDX(pending));
3993 		return;
3994 	}
3995 }
3996 
3997 static inline int
3998 reclaimable(struct sge_eq *eq)
3999 {
4000 	unsigned int cidx;
4001 
4002 	cidx = eq->spg->cidx;	/* stable snapshot */
4003 	cidx = be16toh(cidx);
4004 
4005 	if (cidx >= eq->cidx)
4006 		return (cidx - eq->cidx);
4007 	else
4008 		return (cidx + eq->cap - eq->cidx);
4009 }
4010 
4011 /*
4012  * There are "can_reclaim" tx descriptors ready to be reclaimed.  Reclaim as
4013  * many as possible but stop when there are around "n" mbufs to free.
4014  *
4015  * The actual number reclaimed is provided as the return value.
4016  */
4017 static int
4018 reclaim_tx_descs(struct sge_txq *txq, int can_reclaim, int n)
4019 {
4020 	struct tx_sdesc *txsd;
4021 	struct tx_maps *txmaps;
4022 	struct tx_map *txm;
4023 	unsigned int reclaimed, maps;
4024 	struct sge_eq *eq = &txq->eq;
4025 
4026 	TXQ_LOCK_ASSERT_OWNED(txq);
4027 
4028 	if (can_reclaim == 0)
4029 		can_reclaim = reclaimable(eq);
4030 
4031 	maps = reclaimed = 0;
4032 	while (can_reclaim && maps < n) {
4033 		int ndesc;
4034 
4035 		txsd = &txq->sdesc[eq->cidx];
4036 		ndesc = txsd->desc_used;
4037 
4038 		/* Firmware doesn't return "partial" credits. */
4039 		KASSERT(can_reclaim >= ndesc,
4040 		    ("%s: unexpected number of credits: %d, %d",
4041 		    __func__, can_reclaim, ndesc));
4042 
4043 		maps += txsd->credits;
4044 
4045 		reclaimed += ndesc;
4046 		can_reclaim -= ndesc;
4047 
4048 		eq->cidx += ndesc;
4049 		if (__predict_false(eq->cidx >= eq->cap))
4050 			eq->cidx -= eq->cap;
4051 	}
4052 
4053 	txmaps = &txq->txmaps;
4054 	txm = &txmaps->maps[txmaps->map_cidx];
4055 	if (maps)
4056 		prefetch(txm->m);
4057 
4058 	eq->avail += reclaimed;
4059 	KASSERT(eq->avail < eq->cap,	/* avail tops out at (cap - 1) */
4060 	    ("%s: too many descriptors available", __func__));
4061 
4062 	txmaps->map_avail += maps;
4063 	KASSERT(txmaps->map_avail <= txmaps->map_total,
4064 	    ("%s: too many maps available", __func__));
4065 
4066 	while (maps--) {
4067 		struct tx_map *next;
4068 
4069 		next = txm + 1;
4070 		if (__predict_false(txmaps->map_cidx + 1 == txmaps->map_total))
4071 			next = txmaps->maps;
4072 		prefetch(next->m);
4073 
4074 		bus_dmamap_unload(txq->tx_tag, txm->map);
4075 		m_freem(txm->m);
4076 		txm->m = NULL;
4077 
4078 		txm = next;
4079 		if (__predict_false(++txmaps->map_cidx == txmaps->map_total))
4080 			txmaps->map_cidx = 0;
4081 	}
4082 
4083 	return (reclaimed);
4084 }
4085 
4086 static void
4087 write_eqflush_wr(struct sge_eq *eq)
4088 {
4089 	struct fw_eq_flush_wr *wr;
4090 
4091 	EQ_LOCK_ASSERT_OWNED(eq);
4092 	KASSERT(eq->avail > 0, ("%s: no descriptors left.", __func__));
4093 	KASSERT(!(eq->flags & EQ_CRFLUSHED), ("%s: flushed already", __func__));
4094 
4095 	wr = (void *)&eq->desc[eq->pidx];
4096 	bzero(wr, sizeof(*wr));
4097 	wr->opcode = FW_EQ_FLUSH_WR;
4098 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(sizeof(*wr) / 16) |
4099 	    F_FW_WR_EQUEQ | F_FW_WR_EQUIQ);
4100 
4101 	eq->flags |= (EQ_CRFLUSHED | EQ_STALLED);
4102 	eq->pending++;
4103 	eq->avail--;
4104 	if (++eq->pidx == eq->cap)
4105 		eq->pidx = 0;
4106 }
4107 
4108 static __be64
4109 get_flit(bus_dma_segment_t *sgl, int nsegs, int idx)
4110 {
4111 	int i = (idx / 3) * 2;
4112 
4113 	switch (idx % 3) {
4114 	case 0: {
4115 		__be64 rc;
4116 
4117 		rc = htobe32(sgl[i].ds_len);
4118 		if (i + 1 < nsegs)
4119 			rc |= (uint64_t)htobe32(sgl[i + 1].ds_len) << 32;
4120 
4121 		return (rc);
4122 	}
4123 	case 1:
4124 		return htobe64(sgl[i].ds_addr);
4125 	case 2:
4126 		return htobe64(sgl[i + 1].ds_addr);
4127 	}
4128 
4129 	return (0);
4130 }
4131 
4132 /*
4133  * Find an SGE FL buffer size to use for the given bufsize.  Look for the the
4134  * smallest size that is large enough to hold bufsize or pick the largest size
4135  * if all sizes are less than bufsize.
4136  */
4137 static void
4138 set_fl_tag_idx(struct adapter *sc, struct sge_fl *fl, int bufsize)
4139 {
4140 	int i, largest, best, delta, start;
4141 
4142 	if (fl->flags & FL_BUF_PACKING) {
4143 		fl->tag_idx = 0;	/* first tag is the one for packing */
4144 		return;
4145 	}
4146 
4147 	start = sc->flags & BUF_PACKING_OK ? 1 : 0;
4148 	delta = FL_BUF_SIZE(sc, start) - bufsize;
4149 	if (delta == 0) {
4150 		fl->tag_idx = start;	/* ideal fit, look no further */
4151 		return;
4152 	}
4153 	best = start;
4154 	largest = start;
4155 
4156 	for (i = start + 1; i < FL_BUF_SIZES(sc); i++) {
4157 		int d, fl_buf_size;
4158 
4159 		fl_buf_size = FL_BUF_SIZE(sc, i);
4160 		d = fl_buf_size - bufsize;
4161 
4162 		if (d == 0) {
4163 			fl->tag_idx = i;	/* ideal fit, look no further */
4164 			return;
4165 		}
4166 		if (fl_buf_size > FL_BUF_SIZE(sc, largest))
4167 			largest = i;
4168 		if (d > 0 && (delta < 0 || delta > d)) {
4169 			delta = d;
4170 			best = i;
4171 		}
4172 	}
4173 
4174 	if (delta > 0)
4175 		fl->tag_idx = best;	/* Found a buf bigger than bufsize */
4176 	else
4177 		fl->tag_idx = largest;	/* No buf large enough for bufsize */
4178 }
4179 
4180 static void
4181 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
4182 {
4183 	mtx_lock(&sc->sfl_lock);
4184 	FL_LOCK(fl);
4185 	if ((fl->flags & FL_DOOMED) == 0) {
4186 		fl->flags |= FL_STARVING;
4187 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
4188 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
4189 	}
4190 	FL_UNLOCK(fl);
4191 	mtx_unlock(&sc->sfl_lock);
4192 }
4193 
4194 static int
4195 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
4196     struct mbuf *m)
4197 {
4198 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
4199 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
4200 	struct adapter *sc = iq->adapter;
4201 	struct sge *s = &sc->sge;
4202 	struct sge_eq *eq;
4203 
4204 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
4205 	    rss->opcode));
4206 
4207 	eq = s->eqmap[qid - s->eq_start];
4208 	EQ_LOCK(eq);
4209 	KASSERT(eq->flags & EQ_CRFLUSHED,
4210 	    ("%s: unsolicited egress update", __func__));
4211 	eq->flags &= ~EQ_CRFLUSHED;
4212 	eq->egr_update++;
4213 
4214 	if (__predict_false(eq->flags & EQ_DOOMED))
4215 		wakeup_one(eq);
4216 	else if (eq->flags & EQ_STALLED && can_resume_tx(eq))
4217 		taskqueue_enqueue(sc->tq[eq->tx_chan], &eq->tx_task);
4218 	EQ_UNLOCK(eq);
4219 
4220 	return (0);
4221 }
4222 
4223 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
4224 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
4225     offsetof(struct cpl_fw6_msg, data));
4226 
4227 static int
4228 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
4229 {
4230 	struct adapter *sc = iq->adapter;
4231 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
4232 
4233 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
4234 	    rss->opcode));
4235 
4236 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
4237 		const struct rss_header *rss2;
4238 
4239 		rss2 = (const struct rss_header *)&cpl->data[0];
4240 		return (sc->cpl_handler[rss2->opcode](iq, rss2, m));
4241 	}
4242 
4243 	return (sc->fw_msg_handler[cpl->type](sc, &cpl->data[0]));
4244 }
4245 
4246 static int
4247 sysctl_uint16(SYSCTL_HANDLER_ARGS)
4248 {
4249 	uint16_t *id = arg1;
4250 	int i = *id;
4251 
4252 	return sysctl_handle_int(oidp, &i, 0, req);
4253 }
4254