xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision 49dae58b287906be26f56ba3e3dc693c3ba8cf37)
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/malloc.h>
40 #include <sys/queue.h>
41 #include <sys/sbuf.h>
42 #include <sys/taskqueue.h>
43 #include <sys/time.h>
44 #include <sys/sglist.h>
45 #include <sys/sysctl.h>
46 #include <sys/smp.h>
47 #include <sys/counter.h>
48 #include <net/bpf.h>
49 #include <net/ethernet.h>
50 #include <net/if.h>
51 #include <net/if_vlan_var.h>
52 #include <netinet/in.h>
53 #include <netinet/ip.h>
54 #include <netinet/ip6.h>
55 #include <netinet/tcp.h>
56 #include <machine/md_var.h>
57 #include <vm/vm.h>
58 #include <vm/pmap.h>
59 #ifdef DEV_NETMAP
60 #include <machine/bus.h>
61 #include <sys/selinfo.h>
62 #include <net/if_var.h>
63 #include <net/netmap.h>
64 #include <dev/netmap/netmap_kern.h>
65 #endif
66 
67 #include "common/common.h"
68 #include "common/t4_regs.h"
69 #include "common/t4_regs_values.h"
70 #include "common/t4_msg.h"
71 #include "t4_mp_ring.h"
72 
73 #ifdef T4_PKT_TIMESTAMP
74 #define RX_COPY_THRESHOLD (MINCLSIZE - 8)
75 #else
76 #define RX_COPY_THRESHOLD MINCLSIZE
77 #endif
78 
79 /*
80  * Ethernet frames are DMA'd at this byte offset into the freelist buffer.
81  * 0-7 are valid values.
82  */
83 int fl_pktshift = 2;
84 TUNABLE_INT("hw.cxgbe.fl_pktshift", &fl_pktshift);
85 
86 /*
87  * Pad ethernet payload up to this boundary.
88  * -1: driver should figure out a good value.
89  *  0: disable padding.
90  *  Any power of 2 from 32 to 4096 (both inclusive) is also a valid value.
91  */
92 int fl_pad = -1;
93 TUNABLE_INT("hw.cxgbe.fl_pad", &fl_pad);
94 
95 /*
96  * Status page length.
97  * -1: driver should figure out a good value.
98  *  64 or 128 are the only other valid values.
99  */
100 int spg_len = -1;
101 TUNABLE_INT("hw.cxgbe.spg_len", &spg_len);
102 
103 /*
104  * Congestion drops.
105  * -1: no congestion feedback (not recommended).
106  *  0: backpressure the channel instead of dropping packets right away.
107  *  1: no backpressure, drop packets for the congested queue immediately.
108  */
109 static int cong_drop = 0;
110 TUNABLE_INT("hw.cxgbe.cong_drop", &cong_drop);
111 
112 /*
113  * Deliver multiple frames in the same free list buffer if they fit.
114  * -1: let the driver decide whether to enable buffer packing or not.
115  *  0: disable buffer packing.
116  *  1: enable buffer packing.
117  */
118 static int buffer_packing = -1;
119 TUNABLE_INT("hw.cxgbe.buffer_packing", &buffer_packing);
120 
121 /*
122  * Start next frame in a packed buffer at this boundary.
123  * -1: driver should figure out a good value.
124  * T4: driver will ignore this and use the same value as fl_pad above.
125  * T5: 16, or a power of 2 from 64 to 4096 (both inclusive) is a valid value.
126  */
127 static int fl_pack = -1;
128 TUNABLE_INT("hw.cxgbe.fl_pack", &fl_pack);
129 
130 /*
131  * Allow the driver to create mbuf(s) in a cluster allocated for rx.
132  * 0: never; always allocate mbufs from the zone_mbuf UMA zone.
133  * 1: ok to create mbuf(s) within a cluster if there is room.
134  */
135 static int allow_mbufs_in_cluster = 1;
136 TUNABLE_INT("hw.cxgbe.allow_mbufs_in_cluster", &allow_mbufs_in_cluster);
137 
138 /*
139  * Largest rx cluster size that the driver is allowed to allocate.
140  */
141 static int largest_rx_cluster = MJUM16BYTES;
142 TUNABLE_INT("hw.cxgbe.largest_rx_cluster", &largest_rx_cluster);
143 
144 /*
145  * Size of cluster allocation that's most likely to succeed.  The driver will
146  * fall back to this size if it fails to allocate clusters larger than this.
147  */
148 static int safest_rx_cluster = PAGE_SIZE;
149 TUNABLE_INT("hw.cxgbe.safest_rx_cluster", &safest_rx_cluster);
150 
151 struct txpkts {
152 	u_int wr_type;		/* type 0 or type 1 */
153 	u_int npkt;		/* # of packets in this work request */
154 	u_int plen;		/* total payload (sum of all packets) */
155 	u_int len16;		/* # of 16B pieces used by this work request */
156 };
157 
158 /* A packet's SGL.  This + m_pkthdr has all info needed for tx */
159 struct sgl {
160 	struct sglist sg;
161 	struct sglist_seg seg[TX_SGL_SEGS];
162 };
163 
164 static int service_iq(struct sge_iq *, int);
165 static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t);
166 static int t4_eth_rx(struct sge_iq *, const struct rss_header *, struct mbuf *);
167 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int);
168 static inline void init_fl(struct adapter *, struct sge_fl *, int, int, char *);
169 static inline void init_eq(struct adapter *, struct sge_eq *, int, int, uint8_t,
170     uint16_t, char *);
171 static int alloc_ring(struct adapter *, size_t, bus_dma_tag_t *, bus_dmamap_t *,
172     bus_addr_t *, void **);
173 static int free_ring(struct adapter *, bus_dma_tag_t, bus_dmamap_t, bus_addr_t,
174     void *);
175 static int alloc_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *,
176     int, int);
177 static int free_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *);
178 static void add_fl_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
179     struct sge_fl *);
180 static int alloc_fwq(struct adapter *);
181 static int free_fwq(struct adapter *);
182 static int alloc_mgmtq(struct adapter *);
183 static int free_mgmtq(struct adapter *);
184 static int alloc_rxq(struct vi_info *, struct sge_rxq *, int, int,
185     struct sysctl_oid *);
186 static int free_rxq(struct vi_info *, struct sge_rxq *);
187 #ifdef TCP_OFFLOAD
188 static int alloc_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *, int, int,
189     struct sysctl_oid *);
190 static int free_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *);
191 #endif
192 #ifdef DEV_NETMAP
193 static int alloc_nm_rxq(struct vi_info *, struct sge_nm_rxq *, int, int,
194     struct sysctl_oid *);
195 static int free_nm_rxq(struct vi_info *, struct sge_nm_rxq *);
196 static int alloc_nm_txq(struct vi_info *, struct sge_nm_txq *, int, int,
197     struct sysctl_oid *);
198 static int free_nm_txq(struct vi_info *, struct sge_nm_txq *);
199 #endif
200 static int ctrl_eq_alloc(struct adapter *, struct sge_eq *);
201 static int eth_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
202 #ifdef TCP_OFFLOAD
203 static int ofld_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
204 #endif
205 static int alloc_eq(struct adapter *, struct vi_info *, struct sge_eq *);
206 static int free_eq(struct adapter *, struct sge_eq *);
207 static int alloc_wrq(struct adapter *, struct vi_info *, struct sge_wrq *,
208     struct sysctl_oid *);
209 static int free_wrq(struct adapter *, struct sge_wrq *);
210 static int alloc_txq(struct vi_info *, struct sge_txq *, int,
211     struct sysctl_oid *);
212 static int free_txq(struct vi_info *, struct sge_txq *);
213 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int);
214 static inline void ring_fl_db(struct adapter *, struct sge_fl *);
215 static int refill_fl(struct adapter *, struct sge_fl *, int);
216 static void refill_sfl(void *);
217 static int alloc_fl_sdesc(struct sge_fl *);
218 static void free_fl_sdesc(struct adapter *, struct sge_fl *);
219 static void find_best_refill_source(struct adapter *, struct sge_fl *, int);
220 static void find_safe_refill_source(struct adapter *, struct sge_fl *);
221 static void add_fl_to_sfl(struct adapter *, struct sge_fl *);
222 
223 static inline void get_pkt_gl(struct mbuf *, struct sglist *);
224 static inline u_int txpkt_len16(u_int, u_int);
225 static inline u_int txpkts0_len16(u_int);
226 static inline u_int txpkts1_len16(void);
227 static u_int write_txpkt_wr(struct sge_txq *, struct fw_eth_tx_pkt_wr *,
228     struct mbuf *, u_int);
229 static int try_txpkts(struct mbuf *, struct mbuf *, struct txpkts *, u_int);
230 static int add_to_txpkts(struct mbuf *, struct txpkts *, u_int);
231 static u_int write_txpkts_wr(struct sge_txq *, struct fw_eth_tx_pkts_wr *,
232     struct mbuf *, const struct txpkts *, u_int);
233 static void write_gl_to_txd(struct sge_txq *, struct mbuf *, caddr_t *, int);
234 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int);
235 static inline void ring_eq_db(struct adapter *, struct sge_eq *, u_int);
236 static inline uint16_t read_hw_cidx(struct sge_eq *);
237 static inline u_int reclaimable_tx_desc(struct sge_eq *);
238 static inline u_int total_available_tx_desc(struct sge_eq *);
239 static u_int reclaim_tx_descs(struct sge_txq *, u_int);
240 static void tx_reclaim(void *, int);
241 static __be64 get_flit(struct sglist_seg *, int, int);
242 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
243     struct mbuf *);
244 static int handle_fw_msg(struct sge_iq *, const struct rss_header *,
245     struct mbuf *);
246 static void wrq_tx_drain(void *, int);
247 static void drain_wrq_wr_list(struct adapter *, struct sge_wrq *);
248 
249 static int sysctl_uint16(SYSCTL_HANDLER_ARGS);
250 static int sysctl_bufsizes(SYSCTL_HANDLER_ARGS);
251 
252 static counter_u64_t extfree_refs;
253 static counter_u64_t extfree_rels;
254 
255 /*
256  * Called on MOD_LOAD.  Validates and calculates the SGE tunables.
257  */
258 void
259 t4_sge_modload(void)
260 {
261 
262 	if (fl_pktshift < 0 || fl_pktshift > 7) {
263 		printf("Invalid hw.cxgbe.fl_pktshift value (%d),"
264 		    " using 2 instead.\n", fl_pktshift);
265 		fl_pktshift = 2;
266 	}
267 
268 	if (spg_len != 64 && spg_len != 128) {
269 		int len;
270 
271 #if defined(__i386__) || defined(__amd64__)
272 		len = cpu_clflush_line_size > 64 ? 128 : 64;
273 #else
274 		len = 64;
275 #endif
276 		if (spg_len != -1) {
277 			printf("Invalid hw.cxgbe.spg_len value (%d),"
278 			    " using %d instead.\n", spg_len, len);
279 		}
280 		spg_len = len;
281 	}
282 
283 	if (cong_drop < -1 || cong_drop > 1) {
284 		printf("Invalid hw.cxgbe.cong_drop value (%d),"
285 		    " using 0 instead.\n", cong_drop);
286 		cong_drop = 0;
287 	}
288 
289 	extfree_refs = counter_u64_alloc(M_WAITOK);
290 	extfree_rels = counter_u64_alloc(M_WAITOK);
291 	counter_u64_zero(extfree_refs);
292 	counter_u64_zero(extfree_rels);
293 }
294 
295 void
296 t4_sge_modunload(void)
297 {
298 
299 	counter_u64_free(extfree_refs);
300 	counter_u64_free(extfree_rels);
301 }
302 
303 uint64_t
304 t4_sge_extfree_refs(void)
305 {
306 	uint64_t refs, rels;
307 
308 	rels = counter_u64_fetch(extfree_rels);
309 	refs = counter_u64_fetch(extfree_refs);
310 
311 	return (refs - rels);
312 }
313 
314 void
315 t4_init_sge_cpl_handlers(struct adapter *sc)
316 {
317 
318 	t4_register_cpl_handler(sc, CPL_FW4_MSG, handle_fw_msg);
319 	t4_register_cpl_handler(sc, CPL_FW6_MSG, handle_fw_msg);
320 	t4_register_cpl_handler(sc, CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
321 	t4_register_cpl_handler(sc, CPL_RX_PKT, t4_eth_rx);
322 	t4_register_fw_msg_handler(sc, FW6_TYPE_CMD_RPL, t4_handle_fw_rpl);
323 }
324 
325 static inline void
326 setup_pad_and_pack_boundaries(struct adapter *sc)
327 {
328 	uint32_t v, m;
329 	int pad, pack;
330 
331 	pad = fl_pad;
332 	if (fl_pad < 32 || fl_pad > 4096 || !powerof2(fl_pad)) {
333 		/*
334 		 * If there is any chance that we might use buffer packing and
335 		 * the chip is a T4, then pick 64 as the pad/pack boundary.  Set
336 		 * it to 32 in all other cases.
337 		 */
338 		pad = is_t4(sc) && buffer_packing ? 64 : 32;
339 
340 		/*
341 		 * For fl_pad = 0 we'll still write a reasonable value to the
342 		 * register but all the freelists will opt out of padding.
343 		 * We'll complain here only if the user tried to set it to a
344 		 * value greater than 0 that was invalid.
345 		 */
346 		if (fl_pad > 0) {
347 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pad value"
348 			    " (%d), using %d instead.\n", fl_pad, pad);
349 		}
350 	}
351 	m = V_INGPADBOUNDARY(M_INGPADBOUNDARY);
352 	v = V_INGPADBOUNDARY(ilog2(pad) - 5);
353 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
354 
355 	if (is_t4(sc)) {
356 		if (fl_pack != -1 && fl_pack != pad) {
357 			/* Complain but carry on. */
358 			device_printf(sc->dev, "hw.cxgbe.fl_pack (%d) ignored,"
359 			    " using %d instead.\n", fl_pack, pad);
360 		}
361 		return;
362 	}
363 
364 	pack = fl_pack;
365 	if (fl_pack < 16 || fl_pack == 32 || fl_pack > 4096 ||
366 	    !powerof2(fl_pack)) {
367 		pack = max(sc->params.pci.mps, CACHE_LINE_SIZE);
368 		MPASS(powerof2(pack));
369 		if (pack < 16)
370 			pack = 16;
371 		if (pack == 32)
372 			pack = 64;
373 		if (pack > 4096)
374 			pack = 4096;
375 		if (fl_pack != -1) {
376 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pack value"
377 			    " (%d), using %d instead.\n", fl_pack, pack);
378 		}
379 	}
380 	m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY);
381 	if (pack == 16)
382 		v = V_INGPACKBOUNDARY(0);
383 	else
384 		v = V_INGPACKBOUNDARY(ilog2(pack) - 5);
385 
386 	MPASS(!is_t4(sc));	/* T4 doesn't have SGE_CONTROL2 */
387 	t4_set_reg_field(sc, A_SGE_CONTROL2, m, v);
388 }
389 
390 /*
391  * adap->params.vpd.cclk must be set up before this is called.
392  */
393 void
394 t4_tweak_chip_settings(struct adapter *sc)
395 {
396 	int i;
397 	uint32_t v, m;
398 	int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200};
399 	int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk;
400 	int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */
401 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
402 	static int sge_flbuf_sizes[] = {
403 		MCLBYTES,
404 #if MJUMPAGESIZE != MCLBYTES
405 		MJUMPAGESIZE,
406 		MJUMPAGESIZE - CL_METADATA_SIZE,
407 		MJUMPAGESIZE - 2 * MSIZE - CL_METADATA_SIZE,
408 #endif
409 		MJUM9BYTES,
410 		MJUM16BYTES,
411 		MCLBYTES - MSIZE - CL_METADATA_SIZE,
412 		MJUM9BYTES - CL_METADATA_SIZE,
413 		MJUM16BYTES - CL_METADATA_SIZE,
414 	};
415 
416 	KASSERT(sc->flags & MASTER_PF,
417 	    ("%s: trying to change chip settings when not master.", __func__));
418 
419 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
420 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
421 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
422 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
423 
424 	setup_pad_and_pack_boundaries(sc);
425 
426 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
427 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
428 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
429 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
430 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
431 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
432 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
433 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
434 	t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v);
435 
436 	KASSERT(nitems(sge_flbuf_sizes) <= SGE_FLBUF_SIZES,
437 	    ("%s: hw buffer size table too big", __func__));
438 	for (i = 0; i < min(nitems(sge_flbuf_sizes), SGE_FLBUF_SIZES); i++) {
439 		t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i),
440 		    sge_flbuf_sizes[i]);
441 	}
442 
443 	v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) |
444 	    V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]);
445 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v);
446 
447 	KASSERT(intr_timer[0] <= timer_max,
448 	    ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0],
449 	    timer_max));
450 	for (i = 1; i < nitems(intr_timer); i++) {
451 		KASSERT(intr_timer[i] >= intr_timer[i - 1],
452 		    ("%s: timers not listed in increasing order (%d)",
453 		    __func__, i));
454 
455 		while (intr_timer[i] > timer_max) {
456 			if (i == nitems(intr_timer) - 1) {
457 				intr_timer[i] = timer_max;
458 				break;
459 			}
460 			intr_timer[i] += intr_timer[i - 1];
461 			intr_timer[i] /= 2;
462 		}
463 	}
464 
465 	v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) |
466 	    V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1]));
467 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v);
468 	v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) |
469 	    V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3]));
470 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v);
471 	v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) |
472 	    V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5]));
473 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v);
474 
475 	/* 4K, 16K, 64K, 256K DDP "page sizes" */
476 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
477 	t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v);
478 
479 	m = v = F_TDDPTAGTCB;
480 	t4_set_reg_field(sc, A_ULP_RX_CTL, m, v);
481 
482 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
483 	    F_RESETDDPOFFSET;
484 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
485 	t4_set_reg_field(sc, A_TP_PARA_REG5, m, v);
486 }
487 
488 /*
489  * SGE wants the buffer to be at least 64B and then a multiple of 16.  If
490  * padding is is use the buffer's start and end need to be aligned to the pad
491  * boundary as well.  We'll just make sure that the size is a multiple of the
492  * boundary here, it is up to the buffer allocation code to make sure the start
493  * of the buffer is aligned as well.
494  */
495 static inline int
496 hwsz_ok(struct adapter *sc, int hwsz)
497 {
498 	int mask = fl_pad ? sc->params.sge.pad_boundary - 1 : 16 - 1;
499 
500 	return (hwsz >= 64 && (hwsz & mask) == 0);
501 }
502 
503 /*
504  * XXX: driver really should be able to deal with unexpected settings.
505  */
506 int
507 t4_read_chip_settings(struct adapter *sc)
508 {
509 	struct sge *s = &sc->sge;
510 	struct sge_params *sp = &sc->params.sge;
511 	int i, j, n, rc = 0;
512 	uint32_t m, v, r;
513 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
514 	static int sw_buf_sizes[] = {	/* Sorted by size */
515 		MCLBYTES,
516 #if MJUMPAGESIZE != MCLBYTES
517 		MJUMPAGESIZE,
518 #endif
519 		MJUM9BYTES,
520 		MJUM16BYTES
521 	};
522 	struct sw_zone_info *swz, *safe_swz;
523 	struct hw_buf_info *hwb;
524 
525 	t4_init_sge_params(sc);
526 
527 	m = F_RXPKTCPLMODE;
528 	v = F_RXPKTCPLMODE;
529 	r = t4_read_reg(sc, A_SGE_CONTROL);
530 	if ((r & m) != v) {
531 		device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r);
532 		rc = EINVAL;
533 	}
534 
535 	/*
536 	 * If this changes then every single use of PAGE_SHIFT in the driver
537 	 * needs to be carefully reviewed for PAGE_SHIFT vs sp->page_shift.
538 	 */
539 	if (sp->page_shift != PAGE_SHIFT) {
540 		device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r);
541 		rc = EINVAL;
542 	}
543 
544 	/* Filter out unusable hw buffer sizes entirely (mark with -2). */
545 	hwb = &s->hw_buf_info[0];
546 	for (i = 0; i < nitems(s->hw_buf_info); i++, hwb++) {
547 		r = t4_read_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i));
548 		hwb->size = r;
549 		hwb->zidx = hwsz_ok(sc, r) ? -1 : -2;
550 		hwb->next = -1;
551 	}
552 
553 	/*
554 	 * Create a sorted list in decreasing order of hw buffer sizes (and so
555 	 * increasing order of spare area) for each software zone.
556 	 *
557 	 * If padding is enabled then the start and end of the buffer must align
558 	 * to the pad boundary; if packing is enabled then they must align with
559 	 * the pack boundary as well.  Allocations from the cluster zones are
560 	 * aligned to min(size, 4K), so the buffer starts at that alignment and
561 	 * ends at hwb->size alignment.  If mbuf inlining is allowed the
562 	 * starting alignment will be reduced to MSIZE and the driver will
563 	 * exercise appropriate caution when deciding on the best buffer layout
564 	 * to use.
565 	 */
566 	n = 0;	/* no usable buffer size to begin with */
567 	swz = &s->sw_zone_info[0];
568 	safe_swz = NULL;
569 	for (i = 0; i < SW_ZONE_SIZES; i++, swz++) {
570 		int8_t head = -1, tail = -1;
571 
572 		swz->size = sw_buf_sizes[i];
573 		swz->zone = m_getzone(swz->size);
574 		swz->type = m_gettype(swz->size);
575 
576 		if (swz->size < PAGE_SIZE) {
577 			MPASS(powerof2(swz->size));
578 			if (fl_pad && (swz->size % sp->pad_boundary != 0))
579 				continue;
580 		}
581 
582 		if (swz->size == safest_rx_cluster)
583 			safe_swz = swz;
584 
585 		hwb = &s->hw_buf_info[0];
586 		for (j = 0; j < SGE_FLBUF_SIZES; j++, hwb++) {
587 			if (hwb->zidx != -1 || hwb->size > swz->size)
588 				continue;
589 #ifdef INVARIANTS
590 			if (fl_pad)
591 				MPASS(hwb->size % sp->pad_boundary == 0);
592 #endif
593 			hwb->zidx = i;
594 			if (head == -1)
595 				head = tail = j;
596 			else if (hwb->size < s->hw_buf_info[tail].size) {
597 				s->hw_buf_info[tail].next = j;
598 				tail = j;
599 			} else {
600 				int8_t *cur;
601 				struct hw_buf_info *t;
602 
603 				for (cur = &head; *cur != -1; cur = &t->next) {
604 					t = &s->hw_buf_info[*cur];
605 					if (hwb->size == t->size) {
606 						hwb->zidx = -2;
607 						break;
608 					}
609 					if (hwb->size > t->size) {
610 						hwb->next = *cur;
611 						*cur = j;
612 						break;
613 					}
614 				}
615 			}
616 		}
617 		swz->head_hwidx = head;
618 		swz->tail_hwidx = tail;
619 
620 		if (tail != -1) {
621 			n++;
622 			if (swz->size - s->hw_buf_info[tail].size >=
623 			    CL_METADATA_SIZE)
624 				sc->flags |= BUF_PACKING_OK;
625 		}
626 	}
627 	if (n == 0) {
628 		device_printf(sc->dev, "no usable SGE FL buffer size.\n");
629 		rc = EINVAL;
630 	}
631 
632 	s->safe_hwidx1 = -1;
633 	s->safe_hwidx2 = -1;
634 	if (safe_swz != NULL) {
635 		s->safe_hwidx1 = safe_swz->head_hwidx;
636 		for (i = safe_swz->head_hwidx; i != -1; i = hwb->next) {
637 			int spare;
638 
639 			hwb = &s->hw_buf_info[i];
640 #ifdef INVARIANTS
641 			if (fl_pad)
642 				MPASS(hwb->size % sp->pad_boundary == 0);
643 #endif
644 			spare = safe_swz->size - hwb->size;
645 			if (spare >= CL_METADATA_SIZE) {
646 				s->safe_hwidx2 = i;
647 				break;
648 			}
649 		}
650 	}
651 
652 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
653 	r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ);
654 	if (r != v) {
655 		device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r);
656 		rc = EINVAL;
657 	}
658 
659 	m = v = F_TDDPTAGTCB;
660 	r = t4_read_reg(sc, A_ULP_RX_CTL);
661 	if ((r & m) != v) {
662 		device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r);
663 		rc = EINVAL;
664 	}
665 
666 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
667 	    F_RESETDDPOFFSET;
668 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
669 	r = t4_read_reg(sc, A_TP_PARA_REG5);
670 	if ((r & m) != v) {
671 		device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r);
672 		rc = EINVAL;
673 	}
674 
675 	t4_init_tp_params(sc);
676 
677 	t4_read_mtu_tbl(sc, sc->params.mtus, NULL);
678 	t4_load_mtus(sc, sc->params.mtus, sc->params.a_wnd, sc->params.b_wnd);
679 
680 	return (rc);
681 }
682 
683 int
684 t4_create_dma_tag(struct adapter *sc)
685 {
686 	int rc;
687 
688 	rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
689 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
690 	    BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
691 	    NULL, &sc->dmat);
692 	if (rc != 0) {
693 		device_printf(sc->dev,
694 		    "failed to create main DMA tag: %d\n", rc);
695 	}
696 
697 	return (rc);
698 }
699 
700 void
701 t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
702     struct sysctl_oid_list *children)
703 {
704 	struct sge_params *sp = &sc->params.sge;
705 
706 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "buffer_sizes",
707 	    CTLTYPE_STRING | CTLFLAG_RD, &sc->sge, 0, sysctl_bufsizes, "A",
708 	    "freelist buffer sizes");
709 
710 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD,
711 	    NULL, sp->fl_pktshift, "payload DMA offset in rx buffer (bytes)");
712 
713 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD,
714 	    NULL, sp->pad_boundary, "payload pad boundary (bytes)");
715 
716 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD,
717 	    NULL, sp->spg_len, "status page size (bytes)");
718 
719 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD,
720 	    NULL, cong_drop, "congestion drop setting");
721 
722 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD,
723 	    NULL, sp->pack_boundary, "payload pack boundary (bytes)");
724 }
725 
726 int
727 t4_destroy_dma_tag(struct adapter *sc)
728 {
729 	if (sc->dmat)
730 		bus_dma_tag_destroy(sc->dmat);
731 
732 	return (0);
733 }
734 
735 /*
736  * Allocate and initialize the firmware event queue and the management queue.
737  *
738  * Returns errno on failure.  Resources allocated up to that point may still be
739  * allocated.  Caller is responsible for cleanup in case this function fails.
740  */
741 int
742 t4_setup_adapter_queues(struct adapter *sc)
743 {
744 	int rc;
745 
746 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
747 
748 	sysctl_ctx_init(&sc->ctx);
749 	sc->flags |= ADAP_SYSCTL_CTX;
750 
751 	/*
752 	 * Firmware event queue
753 	 */
754 	rc = alloc_fwq(sc);
755 	if (rc != 0)
756 		return (rc);
757 
758 	/*
759 	 * Management queue.  This is just a control queue that uses the fwq as
760 	 * its associated iq.
761 	 */
762 	rc = alloc_mgmtq(sc);
763 
764 	return (rc);
765 }
766 
767 /*
768  * Idempotent
769  */
770 int
771 t4_teardown_adapter_queues(struct adapter *sc)
772 {
773 
774 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
775 
776 	/* Do this before freeing the queue */
777 	if (sc->flags & ADAP_SYSCTL_CTX) {
778 		sysctl_ctx_free(&sc->ctx);
779 		sc->flags &= ~ADAP_SYSCTL_CTX;
780 	}
781 
782 	free_mgmtq(sc);
783 	free_fwq(sc);
784 
785 	return (0);
786 }
787 
788 static inline int
789 first_vector(struct vi_info *vi)
790 {
791 	struct adapter *sc = vi->pi->adapter;
792 
793 	if (sc->intr_count == 1)
794 		return (0);
795 
796 	return (vi->first_intr);
797 }
798 
799 /*
800  * Given an arbitrary "index," come up with an iq that can be used by other
801  * queues (of this VI) for interrupt forwarding, SGE egress updates, etc.
802  * The iq returned is guaranteed to be something that takes direct interrupts.
803  */
804 static struct sge_iq *
805 vi_intr_iq(struct vi_info *vi, int idx)
806 {
807 	struct adapter *sc = vi->pi->adapter;
808 	struct sge *s = &sc->sge;
809 	struct sge_iq *iq = NULL;
810 	int nintr, i;
811 
812 	if (sc->intr_count == 1)
813 		return (&sc->sge.fwq);
814 
815 	KASSERT(!(vi->flags & VI_NETMAP),
816 	    ("%s: called on netmap VI", __func__));
817 	nintr = vi->nintr;
818 	KASSERT(nintr != 0,
819 	    ("%s: vi %p has no exclusive interrupts, total interrupts = %d",
820 	    __func__, vi, sc->intr_count));
821 	i = idx % nintr;
822 
823 	if (vi->flags & INTR_RXQ) {
824 	       	if (i < vi->nrxq) {
825 			iq = &s->rxq[vi->first_rxq + i].iq;
826 			goto done;
827 		}
828 		i -= vi->nrxq;
829 	}
830 #ifdef TCP_OFFLOAD
831 	if (vi->flags & INTR_OFLD_RXQ) {
832 	       	if (i < vi->nofldrxq) {
833 			iq = &s->ofld_rxq[vi->first_ofld_rxq + i].iq;
834 			goto done;
835 		}
836 		i -= vi->nofldrxq;
837 	}
838 #endif
839 	panic("%s: vi %p, intr_flags 0x%lx, idx %d, total intr %d\n", __func__,
840 	    vi, vi->flags & INTR_ALL, idx, nintr);
841 done:
842 	MPASS(iq != NULL);
843 	KASSERT(iq->flags & IQ_INTR,
844 	    ("%s: iq %p (vi %p, intr_flags 0x%lx, idx %d)", __func__, iq, vi,
845 	    vi->flags & INTR_ALL, idx));
846 	return (iq);
847 }
848 
849 /* Maximum payload that can be delivered with a single iq descriptor */
850 static inline int
851 mtu_to_max_payload(struct adapter *sc, int mtu, const int toe)
852 {
853 	int payload;
854 
855 #ifdef TCP_OFFLOAD
856 	if (toe) {
857 		payload = sc->tt.rx_coalesce ?
858 		    G_RXCOALESCESIZE(t4_read_reg(sc, A_TP_PARA_REG2)) : mtu;
859 	} else {
860 #endif
861 		/* large enough even when hw VLAN extraction is disabled */
862 		payload = sc->params.sge.fl_pktshift + ETHER_HDR_LEN +
863 		    ETHER_VLAN_ENCAP_LEN + mtu;
864 #ifdef TCP_OFFLOAD
865 	}
866 #endif
867 
868 	return (payload);
869 }
870 
871 int
872 t4_setup_vi_queues(struct vi_info *vi)
873 {
874 	int rc = 0, i, j, intr_idx, iqid;
875 	struct sge_rxq *rxq;
876 	struct sge_txq *txq;
877 	struct sge_wrq *ctrlq;
878 #ifdef TCP_OFFLOAD
879 	struct sge_ofld_rxq *ofld_rxq;
880 	struct sge_wrq *ofld_txq;
881 #endif
882 #ifdef DEV_NETMAP
883 	struct sge_nm_rxq *nm_rxq;
884 	struct sge_nm_txq *nm_txq;
885 #endif
886 	char name[16];
887 	struct port_info *pi = vi->pi;
888 	struct adapter *sc = pi->adapter;
889 	struct ifnet *ifp = vi->ifp;
890 	struct sysctl_oid *oid = device_get_sysctl_tree(vi->dev);
891 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
892 	int maxp, mtu = ifp->if_mtu;
893 
894 	/* Interrupt vector to start from (when using multiple vectors) */
895 	intr_idx = first_vector(vi);
896 
897 #ifdef DEV_NETMAP
898 	if (vi->flags & VI_NETMAP) {
899 		/*
900 		 * We don't have buffers to back the netmap rx queues
901 		 * right now so we create the queues in a way that
902 		 * doesn't set off any congestion signal in the chip.
903 		 */
904 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
905 		    CTLFLAG_RD, NULL, "rx queues");
906 		for_each_nm_rxq(vi, i, nm_rxq) {
907 			rc = alloc_nm_rxq(vi, nm_rxq, intr_idx, i, oid);
908 			if (rc != 0)
909 				goto done;
910 			intr_idx++;
911 		}
912 
913 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "txq",
914 		    CTLFLAG_RD, NULL, "tx queues");
915 		for_each_nm_txq(vi, i, nm_txq) {
916 			iqid = vi->first_rxq + (i % vi->nrxq);
917 			rc = alloc_nm_txq(vi, nm_txq, iqid, i, oid);
918 			if (rc != 0)
919 				goto done;
920 		}
921 		goto done;
922 	}
923 #endif
924 
925 	/*
926 	 * First pass over all NIC and TOE rx queues:
927 	 * a) initialize iq and fl
928 	 * b) allocate queue iff it will take direct interrupts.
929 	 */
930 	maxp = mtu_to_max_payload(sc, mtu, 0);
931 	if (vi->flags & INTR_RXQ) {
932 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
933 		    CTLFLAG_RD, NULL, "rx queues");
934 	}
935 	for_each_rxq(vi, i, rxq) {
936 
937 		init_iq(&rxq->iq, sc, vi->tmr_idx, vi->pktc_idx, vi->qsize_rxq);
938 
939 		snprintf(name, sizeof(name), "%s rxq%d-fl",
940 		    device_get_nameunit(vi->dev), i);
941 		init_fl(sc, &rxq->fl, vi->qsize_rxq / 8, maxp, name);
942 
943 		if (vi->flags & INTR_RXQ) {
944 			rxq->iq.flags |= IQ_INTR;
945 			rc = alloc_rxq(vi, rxq, intr_idx, i, oid);
946 			if (rc != 0)
947 				goto done;
948 			intr_idx++;
949 		}
950 	}
951 #ifdef TCP_OFFLOAD
952 	maxp = mtu_to_max_payload(sc, mtu, 1);
953 	if (vi->flags & INTR_OFLD_RXQ) {
954 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq",
955 		    CTLFLAG_RD, NULL,
956 		    "rx queues for offloaded TCP connections");
957 	}
958 	for_each_ofld_rxq(vi, i, ofld_rxq) {
959 
960 		init_iq(&ofld_rxq->iq, sc, vi->tmr_idx, vi->pktc_idx,
961 		    vi->qsize_rxq);
962 
963 		snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
964 		    device_get_nameunit(vi->dev), i);
965 		init_fl(sc, &ofld_rxq->fl, vi->qsize_rxq / 8, maxp, name);
966 
967 		if (vi->flags & INTR_OFLD_RXQ) {
968 			ofld_rxq->iq.flags |= IQ_INTR;
969 			rc = alloc_ofld_rxq(vi, ofld_rxq, intr_idx, i, oid);
970 			if (rc != 0)
971 				goto done;
972 			intr_idx++;
973 		}
974 	}
975 #endif
976 
977 	/*
978 	 * Second pass over all NIC and TOE rx queues.  The queues forwarding
979 	 * their interrupts are allocated now.
980 	 */
981 	j = 0;
982 	if (!(vi->flags & INTR_RXQ)) {
983 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
984 		    CTLFLAG_RD, NULL, "rx queues");
985 		for_each_rxq(vi, i, rxq) {
986 			MPASS(!(rxq->iq.flags & IQ_INTR));
987 
988 			intr_idx = vi_intr_iq(vi, j)->abs_id;
989 
990 			rc = alloc_rxq(vi, rxq, intr_idx, i, oid);
991 			if (rc != 0)
992 				goto done;
993 			j++;
994 		}
995 	}
996 #ifdef TCP_OFFLOAD
997 	if (vi->nofldrxq != 0 && !(vi->flags & INTR_OFLD_RXQ)) {
998 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq",
999 		    CTLFLAG_RD, NULL,
1000 		    "rx queues for offloaded TCP connections");
1001 		for_each_ofld_rxq(vi, i, ofld_rxq) {
1002 			MPASS(!(ofld_rxq->iq.flags & IQ_INTR));
1003 
1004 			intr_idx = vi_intr_iq(vi, j)->abs_id;
1005 
1006 			rc = alloc_ofld_rxq(vi, ofld_rxq, intr_idx, i, oid);
1007 			if (rc != 0)
1008 				goto done;
1009 			j++;
1010 		}
1011 	}
1012 #endif
1013 
1014 	/*
1015 	 * Now the tx queues.  Only one pass needed.
1016 	 */
1017 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "txq", CTLFLAG_RD,
1018 	    NULL, "tx queues");
1019 	j = 0;
1020 	for_each_txq(vi, i, txq) {
1021 		iqid = vi_intr_iq(vi, j)->cntxt_id;
1022 		snprintf(name, sizeof(name), "%s txq%d",
1023 		    device_get_nameunit(vi->dev), i);
1024 		init_eq(sc, &txq->eq, EQ_ETH, vi->qsize_txq, pi->tx_chan, iqid,
1025 		    name);
1026 
1027 		rc = alloc_txq(vi, txq, i, oid);
1028 		if (rc != 0)
1029 			goto done;
1030 		j++;
1031 	}
1032 #ifdef TCP_OFFLOAD
1033 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_txq",
1034 	    CTLFLAG_RD, NULL, "tx queues for offloaded TCP connections");
1035 	for_each_ofld_txq(vi, i, ofld_txq) {
1036 		struct sysctl_oid *oid2;
1037 
1038 		iqid = vi_intr_iq(vi, j)->cntxt_id;
1039 		snprintf(name, sizeof(name), "%s ofld_txq%d",
1040 		    device_get_nameunit(vi->dev), i);
1041 		init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq, pi->tx_chan,
1042 		    iqid, name);
1043 
1044 		snprintf(name, sizeof(name), "%d", i);
1045 		oid2 = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
1046 		    name, CTLFLAG_RD, NULL, "offload tx queue");
1047 
1048 		rc = alloc_wrq(sc, vi, ofld_txq, oid2);
1049 		if (rc != 0)
1050 			goto done;
1051 		j++;
1052 	}
1053 #endif
1054 
1055 	/*
1056 	 * Finally, the control queue.
1057 	 */
1058 	if (!IS_MAIN_VI(vi))
1059 		goto done;
1060 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ctrlq", CTLFLAG_RD,
1061 	    NULL, "ctrl queue");
1062 	ctrlq = &sc->sge.ctrlq[pi->port_id];
1063 	iqid = vi_intr_iq(vi, 0)->cntxt_id;
1064 	snprintf(name, sizeof(name), "%s ctrlq", device_get_nameunit(vi->dev));
1065 	init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, pi->tx_chan, iqid,
1066 	    name);
1067 	rc = alloc_wrq(sc, vi, ctrlq, oid);
1068 
1069 done:
1070 	if (rc)
1071 		t4_teardown_vi_queues(vi);
1072 
1073 	return (rc);
1074 }
1075 
1076 /*
1077  * Idempotent
1078  */
1079 int
1080 t4_teardown_vi_queues(struct vi_info *vi)
1081 {
1082 	int i;
1083 	struct port_info *pi = vi->pi;
1084 	struct adapter *sc = pi->adapter;
1085 	struct sge_rxq *rxq;
1086 	struct sge_txq *txq;
1087 #ifdef TCP_OFFLOAD
1088 	struct sge_ofld_rxq *ofld_rxq;
1089 	struct sge_wrq *ofld_txq;
1090 #endif
1091 #ifdef DEV_NETMAP
1092 	struct sge_nm_rxq *nm_rxq;
1093 	struct sge_nm_txq *nm_txq;
1094 #endif
1095 
1096 	/* Do this before freeing the queues */
1097 	if (vi->flags & VI_SYSCTL_CTX) {
1098 		sysctl_ctx_free(&vi->ctx);
1099 		vi->flags &= ~VI_SYSCTL_CTX;
1100 	}
1101 
1102 #ifdef DEV_NETMAP
1103 	if (vi->flags & VI_NETMAP) {
1104 		for_each_nm_txq(vi, i, nm_txq) {
1105 			free_nm_txq(vi, nm_txq);
1106 		}
1107 
1108 		for_each_nm_rxq(vi, i, nm_rxq) {
1109 			free_nm_rxq(vi, nm_rxq);
1110 		}
1111 		return (0);
1112 	}
1113 #endif
1114 
1115 	/*
1116 	 * Take down all the tx queues first, as they reference the rx queues
1117 	 * (for egress updates, etc.).
1118 	 */
1119 
1120 	if (IS_MAIN_VI(vi))
1121 		free_wrq(sc, &sc->sge.ctrlq[pi->port_id]);
1122 
1123 	for_each_txq(vi, i, txq) {
1124 		free_txq(vi, txq);
1125 	}
1126 #ifdef TCP_OFFLOAD
1127 	for_each_ofld_txq(vi, i, ofld_txq) {
1128 		free_wrq(sc, ofld_txq);
1129 	}
1130 #endif
1131 
1132 	/*
1133 	 * Then take down the rx queues that forward their interrupts, as they
1134 	 * reference other rx queues.
1135 	 */
1136 
1137 	for_each_rxq(vi, i, rxq) {
1138 		if ((rxq->iq.flags & IQ_INTR) == 0)
1139 			free_rxq(vi, rxq);
1140 	}
1141 #ifdef TCP_OFFLOAD
1142 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1143 		if ((ofld_rxq->iq.flags & IQ_INTR) == 0)
1144 			free_ofld_rxq(vi, ofld_rxq);
1145 	}
1146 #endif
1147 
1148 	/*
1149 	 * Then take down the rx queues that take direct interrupts.
1150 	 */
1151 
1152 	for_each_rxq(vi, i, rxq) {
1153 		if (rxq->iq.flags & IQ_INTR)
1154 			free_rxq(vi, rxq);
1155 	}
1156 #ifdef TCP_OFFLOAD
1157 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1158 		if (ofld_rxq->iq.flags & IQ_INTR)
1159 			free_ofld_rxq(vi, ofld_rxq);
1160 	}
1161 #endif
1162 
1163 	return (0);
1164 }
1165 
1166 /*
1167  * Deals with errors and the firmware event queue.  All data rx queues forward
1168  * their interrupt to the firmware event queue.
1169  */
1170 void
1171 t4_intr_all(void *arg)
1172 {
1173 	struct adapter *sc = arg;
1174 	struct sge_iq *fwq = &sc->sge.fwq;
1175 
1176 	t4_intr_err(arg);
1177 	if (atomic_cmpset_int(&fwq->state, IQS_IDLE, IQS_BUSY)) {
1178 		service_iq(fwq, 0);
1179 		atomic_cmpset_int(&fwq->state, IQS_BUSY, IQS_IDLE);
1180 	}
1181 }
1182 
1183 /* Deals with error interrupts */
1184 void
1185 t4_intr_err(void *arg)
1186 {
1187 	struct adapter *sc = arg;
1188 
1189 	t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
1190 	t4_slow_intr_handler(sc);
1191 }
1192 
1193 void
1194 t4_intr_evt(void *arg)
1195 {
1196 	struct sge_iq *iq = arg;
1197 
1198 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1199 		service_iq(iq, 0);
1200 		atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1201 	}
1202 }
1203 
1204 void
1205 t4_intr(void *arg)
1206 {
1207 	struct sge_iq *iq = arg;
1208 
1209 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1210 		service_iq(iq, 0);
1211 		atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1212 	}
1213 }
1214 
1215 /*
1216  * Deals with anything and everything on the given ingress queue.
1217  */
1218 static int
1219 service_iq(struct sge_iq *iq, int budget)
1220 {
1221 	struct sge_iq *q;
1222 	struct sge_rxq *rxq = iq_to_rxq(iq);	/* Use iff iq is part of rxq */
1223 	struct sge_fl *fl;			/* Use iff IQ_HAS_FL */
1224 	struct adapter *sc = iq->adapter;
1225 	struct iq_desc *d = &iq->desc[iq->cidx];
1226 	int ndescs = 0, limit;
1227 	int rsp_type, refill;
1228 	uint32_t lq;
1229 	uint16_t fl_hw_cidx;
1230 	struct mbuf *m0;
1231 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1232 #if defined(INET) || defined(INET6)
1233 	const struct timeval lro_timeout = {0, sc->lro_timeout};
1234 #endif
1235 
1236 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1237 
1238 	limit = budget ? budget : iq->qsize / 16;
1239 
1240 	if (iq->flags & IQ_HAS_FL) {
1241 		fl = &rxq->fl;
1242 		fl_hw_cidx = fl->hw_cidx;	/* stable snapshot */
1243 	} else {
1244 		fl = NULL;
1245 		fl_hw_cidx = 0;			/* to silence gcc warning */
1246 	}
1247 
1248 	/*
1249 	 * We always come back and check the descriptor ring for new indirect
1250 	 * interrupts and other responses after running a single handler.
1251 	 */
1252 	for (;;) {
1253 		while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1254 
1255 			rmb();
1256 
1257 			refill = 0;
1258 			m0 = NULL;
1259 			rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1260 			lq = be32toh(d->rsp.pldbuflen_qid);
1261 
1262 			switch (rsp_type) {
1263 			case X_RSPD_TYPE_FLBUF:
1264 
1265 				KASSERT(iq->flags & IQ_HAS_FL,
1266 				    ("%s: data for an iq (%p) with no freelist",
1267 				    __func__, iq));
1268 
1269 				m0 = get_fl_payload(sc, fl, lq);
1270 				if (__predict_false(m0 == NULL))
1271 					goto process_iql;
1272 				refill = IDXDIFF(fl->hw_cidx, fl_hw_cidx, fl->sidx) > 2;
1273 #ifdef T4_PKT_TIMESTAMP
1274 				/*
1275 				 * 60 bit timestamp for the payload is
1276 				 * *(uint64_t *)m0->m_pktdat.  Note that it is
1277 				 * in the leading free-space in the mbuf.  The
1278 				 * kernel can clobber it during a pullup,
1279 				 * m_copymdata, etc.  You need to make sure that
1280 				 * the mbuf reaches you unmolested if you care
1281 				 * about the timestamp.
1282 				 */
1283 				*(uint64_t *)m0->m_pktdat =
1284 				    be64toh(ctrl->u.last_flit) &
1285 				    0xfffffffffffffff;
1286 #endif
1287 
1288 				/* fall through */
1289 
1290 			case X_RSPD_TYPE_CPL:
1291 				KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1292 				    ("%s: bad opcode %02x.", __func__,
1293 				    d->rss.opcode));
1294 				sc->cpl_handler[d->rss.opcode](iq, &d->rss, m0);
1295 				break;
1296 
1297 			case X_RSPD_TYPE_INTR:
1298 
1299 				/*
1300 				 * Interrupts should be forwarded only to queues
1301 				 * that are not forwarding their interrupts.
1302 				 * This means service_iq can recurse but only 1
1303 				 * level deep.
1304 				 */
1305 				KASSERT(budget == 0,
1306 				    ("%s: budget %u, rsp_type %u", __func__,
1307 				    budget, rsp_type));
1308 
1309 				/*
1310 				 * There are 1K interrupt-capable queues (qids 0
1311 				 * through 1023).  A response type indicating a
1312 				 * forwarded interrupt with a qid >= 1K is an
1313 				 * iWARP async notification.
1314 				 */
1315 				if (lq >= 1024) {
1316                                         sc->an_handler(iq, &d->rsp);
1317                                         break;
1318                                 }
1319 
1320 				q = sc->sge.iqmap[lq - sc->sge.iq_start];
1321 				if (atomic_cmpset_int(&q->state, IQS_IDLE,
1322 				    IQS_BUSY)) {
1323 					if (service_iq(q, q->qsize / 16) == 0) {
1324 						atomic_cmpset_int(&q->state,
1325 						    IQS_BUSY, IQS_IDLE);
1326 					} else {
1327 						STAILQ_INSERT_TAIL(&iql, q,
1328 						    link);
1329 					}
1330 				}
1331 				break;
1332 
1333 			default:
1334 				KASSERT(0,
1335 				    ("%s: illegal response type %d on iq %p",
1336 				    __func__, rsp_type, iq));
1337 				log(LOG_ERR,
1338 				    "%s: illegal response type %d on iq %p",
1339 				    device_get_nameunit(sc->dev), rsp_type, iq);
1340 				break;
1341 			}
1342 
1343 			d++;
1344 			if (__predict_false(++iq->cidx == iq->sidx)) {
1345 				iq->cidx = 0;
1346 				iq->gen ^= F_RSPD_GEN;
1347 				d = &iq->desc[0];
1348 			}
1349 			if (__predict_false(++ndescs == limit)) {
1350 				t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
1351 				    V_CIDXINC(ndescs) |
1352 				    V_INGRESSQID(iq->cntxt_id) |
1353 				    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1354 				ndescs = 0;
1355 
1356 #if defined(INET) || defined(INET6)
1357 				if (iq->flags & IQ_LRO_ENABLED &&
1358 				    sc->lro_timeout != 0) {
1359 					tcp_lro_flush_inactive(&rxq->lro,
1360 					    &lro_timeout);
1361 				}
1362 #endif
1363 
1364 				if (budget) {
1365 					if (iq->flags & IQ_HAS_FL) {
1366 						FL_LOCK(fl);
1367 						refill_fl(sc, fl, 32);
1368 						FL_UNLOCK(fl);
1369 					}
1370 					return (EINPROGRESS);
1371 				}
1372 			}
1373 			if (refill) {
1374 				FL_LOCK(fl);
1375 				refill_fl(sc, fl, 32);
1376 				FL_UNLOCK(fl);
1377 				fl_hw_cidx = fl->hw_cidx;
1378 			}
1379 		}
1380 
1381 process_iql:
1382 		if (STAILQ_EMPTY(&iql))
1383 			break;
1384 
1385 		/*
1386 		 * Process the head only, and send it to the back of the list if
1387 		 * it's still not done.
1388 		 */
1389 		q = STAILQ_FIRST(&iql);
1390 		STAILQ_REMOVE_HEAD(&iql, link);
1391 		if (service_iq(q, q->qsize / 8) == 0)
1392 			atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1393 		else
1394 			STAILQ_INSERT_TAIL(&iql, q, link);
1395 	}
1396 
1397 #if defined(INET) || defined(INET6)
1398 	if (iq->flags & IQ_LRO_ENABLED) {
1399 		struct lro_ctrl *lro = &rxq->lro;
1400 
1401 		tcp_lro_flush_all(lro);
1402 	}
1403 #endif
1404 
1405 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_CIDXINC(ndescs) |
1406 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1407 
1408 	if (iq->flags & IQ_HAS_FL) {
1409 		int starved;
1410 
1411 		FL_LOCK(fl);
1412 		starved = refill_fl(sc, fl, 64);
1413 		FL_UNLOCK(fl);
1414 		if (__predict_false(starved != 0))
1415 			add_fl_to_sfl(sc, fl);
1416 	}
1417 
1418 	return (0);
1419 }
1420 
1421 static inline int
1422 cl_has_metadata(struct sge_fl *fl, struct cluster_layout *cll)
1423 {
1424 	int rc = fl->flags & FL_BUF_PACKING || cll->region1 > 0;
1425 
1426 	if (rc)
1427 		MPASS(cll->region3 >= CL_METADATA_SIZE);
1428 
1429 	return (rc);
1430 }
1431 
1432 static inline struct cluster_metadata *
1433 cl_metadata(struct adapter *sc, struct sge_fl *fl, struct cluster_layout *cll,
1434     caddr_t cl)
1435 {
1436 
1437 	if (cl_has_metadata(fl, cll)) {
1438 		struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1439 
1440 		return ((struct cluster_metadata *)(cl + swz->size) - 1);
1441 	}
1442 	return (NULL);
1443 }
1444 
1445 static void
1446 rxb_free(struct mbuf *m, void *arg1, void *arg2)
1447 {
1448 	uma_zone_t zone = arg1;
1449 	caddr_t cl = arg2;
1450 
1451 	uma_zfree(zone, cl);
1452 	counter_u64_add(extfree_rels, 1);
1453 }
1454 
1455 /*
1456  * The mbuf returned by this function could be allocated from zone_mbuf or
1457  * constructed in spare room in the cluster.
1458  *
1459  * The mbuf carries the payload in one of these ways
1460  * a) frame inside the mbuf (mbuf from zone_mbuf)
1461  * b) m_cljset (for clusters without metadata) zone_mbuf
1462  * c) m_extaddref (cluster with metadata) inline mbuf
1463  * d) m_extaddref (cluster with metadata) zone_mbuf
1464  */
1465 static struct mbuf *
1466 get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1467     int remaining)
1468 {
1469 	struct mbuf *m;
1470 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1471 	struct cluster_layout *cll = &sd->cll;
1472 	struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1473 	struct hw_buf_info *hwb = &sc->sge.hw_buf_info[cll->hwidx];
1474 	struct cluster_metadata *clm = cl_metadata(sc, fl, cll, sd->cl);
1475 	int len, blen;
1476 	caddr_t payload;
1477 
1478 	blen = hwb->size - fl->rx_offset;	/* max possible in this buf */
1479 	len = min(remaining, blen);
1480 	payload = sd->cl + cll->region1 + fl->rx_offset;
1481 	if (fl->flags & FL_BUF_PACKING) {
1482 		const u_int l = fr_offset + len;
1483 		const u_int pad = roundup2(l, fl->buf_boundary) - l;
1484 
1485 		if (fl->rx_offset + len + pad < hwb->size)
1486 			blen = len + pad;
1487 		MPASS(fl->rx_offset + blen <= hwb->size);
1488 	} else {
1489 		MPASS(fl->rx_offset == 0);	/* not packing */
1490 	}
1491 
1492 
1493 	if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) {
1494 
1495 		/*
1496 		 * Copy payload into a freshly allocated mbuf.
1497 		 */
1498 
1499 		m = fr_offset == 0 ?
1500 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1501 		if (m == NULL)
1502 			return (NULL);
1503 		fl->mbuf_allocated++;
1504 #ifdef T4_PKT_TIMESTAMP
1505 		/* Leave room for a timestamp */
1506 		m->m_data += 8;
1507 #endif
1508 		/* copy data to mbuf */
1509 		bcopy(payload, mtod(m, caddr_t), len);
1510 
1511 	} else if (sd->nmbuf * MSIZE < cll->region1) {
1512 
1513 		/*
1514 		 * There's spare room in the cluster for an mbuf.  Create one
1515 		 * and associate it with the payload that's in the cluster.
1516 		 */
1517 
1518 		MPASS(clm != NULL);
1519 		m = (struct mbuf *)(sd->cl + sd->nmbuf * MSIZE);
1520 		/* No bzero required */
1521 		if (m_init(m, M_NOWAIT, MT_DATA,
1522 		    fr_offset == 0 ? M_PKTHDR | M_NOFREE : M_NOFREE))
1523 			return (NULL);
1524 		fl->mbuf_inlined++;
1525 		m_extaddref(m, payload, blen, &clm->refcount, rxb_free,
1526 		    swz->zone, sd->cl);
1527 		if (sd->nmbuf++ == 0)
1528 			counter_u64_add(extfree_refs, 1);
1529 
1530 	} else {
1531 
1532 		/*
1533 		 * Grab an mbuf from zone_mbuf and associate it with the
1534 		 * payload in the cluster.
1535 		 */
1536 
1537 		m = fr_offset == 0 ?
1538 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1539 		if (m == NULL)
1540 			return (NULL);
1541 		fl->mbuf_allocated++;
1542 		if (clm != NULL) {
1543 			m_extaddref(m, payload, blen, &clm->refcount,
1544 			    rxb_free, swz->zone, sd->cl);
1545 			if (sd->nmbuf++ == 0)
1546 				counter_u64_add(extfree_refs, 1);
1547 		} else {
1548 			m_cljset(m, sd->cl, swz->type);
1549 			sd->cl = NULL;	/* consumed, not a recycle candidate */
1550 		}
1551 	}
1552 	if (fr_offset == 0)
1553 		m->m_pkthdr.len = remaining;
1554 	m->m_len = len;
1555 
1556 	if (fl->flags & FL_BUF_PACKING) {
1557 		fl->rx_offset += blen;
1558 		MPASS(fl->rx_offset <= hwb->size);
1559 		if (fl->rx_offset < hwb->size)
1560 			return (m);	/* without advancing the cidx */
1561 	}
1562 
1563 	if (__predict_false(++fl->cidx % 8 == 0)) {
1564 		uint16_t cidx = fl->cidx / 8;
1565 
1566 		if (__predict_false(cidx == fl->sidx))
1567 			fl->cidx = cidx = 0;
1568 		fl->hw_cidx = cidx;
1569 	}
1570 	fl->rx_offset = 0;
1571 
1572 	return (m);
1573 }
1574 
1575 static struct mbuf *
1576 get_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf)
1577 {
1578 	struct mbuf *m0, *m, **pnext;
1579 	u_int remaining;
1580 	const u_int total = G_RSPD_LEN(len_newbuf);
1581 
1582 	if (__predict_false(fl->flags & FL_BUF_RESUME)) {
1583 		M_ASSERTPKTHDR(fl->m0);
1584 		MPASS(fl->m0->m_pkthdr.len == total);
1585 		MPASS(fl->remaining < total);
1586 
1587 		m0 = fl->m0;
1588 		pnext = fl->pnext;
1589 		remaining = fl->remaining;
1590 		fl->flags &= ~FL_BUF_RESUME;
1591 		goto get_segment;
1592 	}
1593 
1594 	if (fl->rx_offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
1595 		fl->rx_offset = 0;
1596 		if (__predict_false(++fl->cidx % 8 == 0)) {
1597 			uint16_t cidx = fl->cidx / 8;
1598 
1599 			if (__predict_false(cidx == fl->sidx))
1600 				fl->cidx = cidx = 0;
1601 			fl->hw_cidx = cidx;
1602 		}
1603 	}
1604 
1605 	/*
1606 	 * Payload starts at rx_offset in the current hw buffer.  Its length is
1607 	 * 'len' and it may span multiple hw buffers.
1608 	 */
1609 
1610 	m0 = get_scatter_segment(sc, fl, 0, total);
1611 	if (m0 == NULL)
1612 		return (NULL);
1613 	remaining = total - m0->m_len;
1614 	pnext = &m0->m_next;
1615 	while (remaining > 0) {
1616 get_segment:
1617 		MPASS(fl->rx_offset == 0);
1618 		m = get_scatter_segment(sc, fl, total - remaining, remaining);
1619 		if (__predict_false(m == NULL)) {
1620 			fl->m0 = m0;
1621 			fl->pnext = pnext;
1622 			fl->remaining = remaining;
1623 			fl->flags |= FL_BUF_RESUME;
1624 			return (NULL);
1625 		}
1626 		*pnext = m;
1627 		pnext = &m->m_next;
1628 		remaining -= m->m_len;
1629 	}
1630 	*pnext = NULL;
1631 
1632 	M_ASSERTPKTHDR(m0);
1633 	return (m0);
1634 }
1635 
1636 static int
1637 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
1638 {
1639 	struct sge_rxq *rxq = iq_to_rxq(iq);
1640 	struct ifnet *ifp = rxq->ifp;
1641 	struct adapter *sc = iq->adapter;
1642 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
1643 #if defined(INET) || defined(INET6)
1644 	struct lro_ctrl *lro = &rxq->lro;
1645 #endif
1646 	static const int sw_hashtype[4][2] = {
1647 		{M_HASHTYPE_NONE, M_HASHTYPE_NONE},
1648 		{M_HASHTYPE_RSS_IPV4, M_HASHTYPE_RSS_IPV6},
1649 		{M_HASHTYPE_RSS_TCP_IPV4, M_HASHTYPE_RSS_TCP_IPV6},
1650 		{M_HASHTYPE_RSS_UDP_IPV4, M_HASHTYPE_RSS_UDP_IPV6},
1651 	};
1652 
1653 	KASSERT(m0 != NULL, ("%s: no payload with opcode %02x", __func__,
1654 	    rss->opcode));
1655 
1656 	m0->m_pkthdr.len -= sc->params.sge.fl_pktshift;
1657 	m0->m_len -= sc->params.sge.fl_pktshift;
1658 	m0->m_data += sc->params.sge.fl_pktshift;
1659 
1660 	m0->m_pkthdr.rcvif = ifp;
1661 	M_HASHTYPE_SET(m0, sw_hashtype[rss->hash_type][rss->ipv6]);
1662 	m0->m_pkthdr.flowid = be32toh(rss->hash_val);
1663 
1664 	if (cpl->csum_calc && !cpl->err_vec) {
1665 		if (ifp->if_capenable & IFCAP_RXCSUM &&
1666 		    cpl->l2info & htobe32(F_RXF_IP)) {
1667 			m0->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
1668 			    CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
1669 			rxq->rxcsum++;
1670 		} else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
1671 		    cpl->l2info & htobe32(F_RXF_IP6)) {
1672 			m0->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
1673 			    CSUM_PSEUDO_HDR);
1674 			rxq->rxcsum++;
1675 		}
1676 
1677 		if (__predict_false(cpl->ip_frag))
1678 			m0->m_pkthdr.csum_data = be16toh(cpl->csum);
1679 		else
1680 			m0->m_pkthdr.csum_data = 0xffff;
1681 	}
1682 
1683 	if (cpl->vlan_ex) {
1684 		m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
1685 		m0->m_flags |= M_VLANTAG;
1686 		rxq->vlan_extraction++;
1687 	}
1688 
1689 #if defined(INET) || defined(INET6)
1690 	if (cpl->l2info & htobe32(F_RXF_LRO) &&
1691 	    iq->flags & IQ_LRO_ENABLED &&
1692 	    tcp_lro_rx(lro, m0, 0) == 0) {
1693 		/* queued for LRO */
1694 	} else
1695 #endif
1696 	ifp->if_input(ifp, m0);
1697 
1698 	return (0);
1699 }
1700 
1701 /*
1702  * Must drain the wrq or make sure that someone else will.
1703  */
1704 static void
1705 wrq_tx_drain(void *arg, int n)
1706 {
1707 	struct sge_wrq *wrq = arg;
1708 	struct sge_eq *eq = &wrq->eq;
1709 
1710 	EQ_LOCK(eq);
1711 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
1712 		drain_wrq_wr_list(wrq->adapter, wrq);
1713 	EQ_UNLOCK(eq);
1714 }
1715 
1716 static void
1717 drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq)
1718 {
1719 	struct sge_eq *eq = &wrq->eq;
1720 	u_int available, dbdiff;	/* # of hardware descriptors */
1721 	u_int n;
1722 	struct wrqe *wr;
1723 	struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
1724 
1725 	EQ_LOCK_ASSERT_OWNED(eq);
1726 	MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
1727 	wr = STAILQ_FIRST(&wrq->wr_list);
1728 	MPASS(wr != NULL);	/* Must be called with something useful to do */
1729 	MPASS(eq->pidx == eq->dbidx);
1730 	dbdiff = 0;
1731 
1732 	do {
1733 		eq->cidx = read_hw_cidx(eq);
1734 		if (eq->pidx == eq->cidx)
1735 			available = eq->sidx - 1;
1736 		else
1737 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
1738 
1739 		MPASS(wr->wrq == wrq);
1740 		n = howmany(wr->wr_len, EQ_ESIZE);
1741 		if (available < n)
1742 			break;
1743 
1744 		dst = (void *)&eq->desc[eq->pidx];
1745 		if (__predict_true(eq->sidx - eq->pidx > n)) {
1746 			/* Won't wrap, won't end exactly at the status page. */
1747 			bcopy(&wr->wr[0], dst, wr->wr_len);
1748 			eq->pidx += n;
1749 		} else {
1750 			int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE;
1751 
1752 			bcopy(&wr->wr[0], dst, first_portion);
1753 			if (wr->wr_len > first_portion) {
1754 				bcopy(&wr->wr[first_portion], &eq->desc[0],
1755 				    wr->wr_len - first_portion);
1756 			}
1757 			eq->pidx = n - (eq->sidx - eq->pidx);
1758 		}
1759 
1760 		if (available < eq->sidx / 4 &&
1761 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
1762 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
1763 			    F_FW_WR_EQUEQ);
1764 			eq->equeqidx = eq->pidx;
1765 		} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
1766 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
1767 			eq->equeqidx = eq->pidx;
1768 		}
1769 
1770 		dbdiff += n;
1771 		if (dbdiff >= 16) {
1772 			ring_eq_db(sc, eq, dbdiff);
1773 			dbdiff = 0;
1774 		}
1775 
1776 		STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
1777 		free_wrqe(wr);
1778 		MPASS(wrq->nwr_pending > 0);
1779 		wrq->nwr_pending--;
1780 		MPASS(wrq->ndesc_needed >= n);
1781 		wrq->ndesc_needed -= n;
1782 	} while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL);
1783 
1784 	if (dbdiff)
1785 		ring_eq_db(sc, eq, dbdiff);
1786 }
1787 
1788 /*
1789  * Doesn't fail.  Holds on to work requests it can't send right away.
1790  */
1791 void
1792 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
1793 {
1794 #ifdef INVARIANTS
1795 	struct sge_eq *eq = &wrq->eq;
1796 #endif
1797 
1798 	EQ_LOCK_ASSERT_OWNED(eq);
1799 	MPASS(wr != NULL);
1800 	MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN);
1801 	MPASS((wr->wr_len & 0x7) == 0);
1802 
1803 	STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
1804 	wrq->nwr_pending++;
1805 	wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE);
1806 
1807 	if (!TAILQ_EMPTY(&wrq->incomplete_wrs))
1808 		return;	/* commit_wrq_wr will drain wr_list as well. */
1809 
1810 	drain_wrq_wr_list(sc, wrq);
1811 
1812 	/* Doorbell must have caught up to the pidx. */
1813 	MPASS(eq->pidx == eq->dbidx);
1814 }
1815 
1816 void
1817 t4_update_fl_bufsize(struct ifnet *ifp)
1818 {
1819 	struct vi_info *vi = ifp->if_softc;
1820 	struct adapter *sc = vi->pi->adapter;
1821 	struct sge_rxq *rxq;
1822 #ifdef TCP_OFFLOAD
1823 	struct sge_ofld_rxq *ofld_rxq;
1824 #endif
1825 	struct sge_fl *fl;
1826 	int i, maxp, mtu = ifp->if_mtu;
1827 
1828 	maxp = mtu_to_max_payload(sc, mtu, 0);
1829 	for_each_rxq(vi, i, rxq) {
1830 		fl = &rxq->fl;
1831 
1832 		FL_LOCK(fl);
1833 		find_best_refill_source(sc, fl, maxp);
1834 		FL_UNLOCK(fl);
1835 	}
1836 #ifdef TCP_OFFLOAD
1837 	maxp = mtu_to_max_payload(sc, mtu, 1);
1838 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1839 		fl = &ofld_rxq->fl;
1840 
1841 		FL_LOCK(fl);
1842 		find_best_refill_source(sc, fl, maxp);
1843 		FL_UNLOCK(fl);
1844 	}
1845 #endif
1846 }
1847 
1848 static inline int
1849 mbuf_nsegs(struct mbuf *m)
1850 {
1851 
1852 	M_ASSERTPKTHDR(m);
1853 	KASSERT(m->m_pkthdr.l5hlen > 0,
1854 	    ("%s: mbuf %p missing information on # of segments.", __func__, m));
1855 
1856 	return (m->m_pkthdr.l5hlen);
1857 }
1858 
1859 static inline void
1860 set_mbuf_nsegs(struct mbuf *m, uint8_t nsegs)
1861 {
1862 
1863 	M_ASSERTPKTHDR(m);
1864 	m->m_pkthdr.l5hlen = nsegs;
1865 }
1866 
1867 static inline int
1868 mbuf_len16(struct mbuf *m)
1869 {
1870 	int n;
1871 
1872 	M_ASSERTPKTHDR(m);
1873 	n = m->m_pkthdr.PH_loc.eight[0];
1874 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
1875 
1876 	return (n);
1877 }
1878 
1879 static inline void
1880 set_mbuf_len16(struct mbuf *m, uint8_t len16)
1881 {
1882 
1883 	M_ASSERTPKTHDR(m);
1884 	m->m_pkthdr.PH_loc.eight[0] = len16;
1885 }
1886 
1887 static inline int
1888 needs_tso(struct mbuf *m)
1889 {
1890 
1891 	M_ASSERTPKTHDR(m);
1892 
1893 	if (m->m_pkthdr.csum_flags & CSUM_TSO) {
1894 		KASSERT(m->m_pkthdr.tso_segsz > 0,
1895 		    ("%s: TSO requested in mbuf %p but MSS not provided",
1896 		    __func__, m));
1897 		return (1);
1898 	}
1899 
1900 	return (0);
1901 }
1902 
1903 static inline int
1904 needs_l3_csum(struct mbuf *m)
1905 {
1906 
1907 	M_ASSERTPKTHDR(m);
1908 
1909 	if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO))
1910 		return (1);
1911 	return (0);
1912 }
1913 
1914 static inline int
1915 needs_l4_csum(struct mbuf *m)
1916 {
1917 
1918 	M_ASSERTPKTHDR(m);
1919 
1920 	if (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
1921 	    CSUM_TCP_IPV6 | CSUM_TSO))
1922 		return (1);
1923 	return (0);
1924 }
1925 
1926 static inline int
1927 needs_vlan_insertion(struct mbuf *m)
1928 {
1929 
1930 	M_ASSERTPKTHDR(m);
1931 
1932 	if (m->m_flags & M_VLANTAG) {
1933 		KASSERT(m->m_pkthdr.ether_vtag != 0,
1934 		    ("%s: HWVLAN requested in mbuf %p but tag not provided",
1935 		    __func__, m));
1936 		return (1);
1937 	}
1938 	return (0);
1939 }
1940 
1941 static void *
1942 m_advance(struct mbuf **pm, int *poffset, int len)
1943 {
1944 	struct mbuf *m = *pm;
1945 	int offset = *poffset;
1946 	uintptr_t p = 0;
1947 
1948 	MPASS(len > 0);
1949 
1950 	while (len) {
1951 		if (offset + len < m->m_len) {
1952 			offset += len;
1953 			p = mtod(m, uintptr_t) + offset;
1954 			break;
1955 		}
1956 		len -= m->m_len - offset;
1957 		m = m->m_next;
1958 		offset = 0;
1959 		MPASS(m != NULL);
1960 	}
1961 	*poffset = offset;
1962 	*pm = m;
1963 	return ((void *)p);
1964 }
1965 
1966 static inline int
1967 same_paddr(char *a, char *b)
1968 {
1969 
1970 	if (a == b)
1971 		return (1);
1972 	else if (a != NULL && b != NULL) {
1973 		vm_offset_t x = (vm_offset_t)a;
1974 		vm_offset_t y = (vm_offset_t)b;
1975 
1976 		if ((x & PAGE_MASK) == (y & PAGE_MASK) &&
1977 		    pmap_kextract(x) == pmap_kextract(y))
1978 			return (1);
1979 	}
1980 
1981 	return (0);
1982 }
1983 
1984 /*
1985  * Can deal with empty mbufs in the chain that have m_len = 0, but the chain
1986  * must have at least one mbuf that's not empty.
1987  */
1988 static inline int
1989 count_mbuf_nsegs(struct mbuf *m)
1990 {
1991 	char *prev_end, *start;
1992 	int len, nsegs;
1993 
1994 	MPASS(m != NULL);
1995 
1996 	nsegs = 0;
1997 	prev_end = NULL;
1998 	for (; m; m = m->m_next) {
1999 
2000 		len = m->m_len;
2001 		if (__predict_false(len == 0))
2002 			continue;
2003 		start = mtod(m, char *);
2004 
2005 		nsegs += sglist_count(start, len);
2006 		if (same_paddr(prev_end, start))
2007 			nsegs--;
2008 		prev_end = start + len;
2009 	}
2010 
2011 	MPASS(nsegs > 0);
2012 	return (nsegs);
2013 }
2014 
2015 /*
2016  * Analyze the mbuf to determine its tx needs.  The mbuf passed in may change:
2017  * a) caller can assume it's been freed if this function returns with an error.
2018  * b) it may get defragged up if the gather list is too long for the hardware.
2019  */
2020 int
2021 parse_pkt(struct mbuf **mp)
2022 {
2023 	struct mbuf *m0 = *mp, *m;
2024 	int rc, nsegs, defragged = 0, offset;
2025 	struct ether_header *eh;
2026 	void *l3hdr;
2027 #if defined(INET) || defined(INET6)
2028 	struct tcphdr *tcp;
2029 #endif
2030 	uint16_t eh_type;
2031 
2032 	M_ASSERTPKTHDR(m0);
2033 	if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) {
2034 		rc = EINVAL;
2035 fail:
2036 		m_freem(m0);
2037 		*mp = NULL;
2038 		return (rc);
2039 	}
2040 restart:
2041 	/*
2042 	 * First count the number of gather list segments in the payload.
2043 	 * Defrag the mbuf if nsegs exceeds the hardware limit.
2044 	 */
2045 	M_ASSERTPKTHDR(m0);
2046 	MPASS(m0->m_pkthdr.len > 0);
2047 	nsegs = count_mbuf_nsegs(m0);
2048 	if (nsegs > (needs_tso(m0) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS)) {
2049 		if (defragged++ > 0 || (m = m_defrag(m0, M_NOWAIT)) == NULL) {
2050 			rc = EFBIG;
2051 			goto fail;
2052 		}
2053 		*mp = m0 = m;	/* update caller's copy after defrag */
2054 		goto restart;
2055 	}
2056 
2057 	if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN)) {
2058 		m0 = m_pullup(m0, m0->m_pkthdr.len);
2059 		if (m0 == NULL) {
2060 			/* Should have left well enough alone. */
2061 			rc = EFBIG;
2062 			goto fail;
2063 		}
2064 		*mp = m0;	/* update caller's copy after pullup */
2065 		goto restart;
2066 	}
2067 	set_mbuf_nsegs(m0, nsegs);
2068 	set_mbuf_len16(m0, txpkt_len16(nsegs, needs_tso(m0)));
2069 
2070 	if (!needs_tso(m0))
2071 		return (0);
2072 
2073 	m = m0;
2074 	eh = mtod(m, struct ether_header *);
2075 	eh_type = ntohs(eh->ether_type);
2076 	if (eh_type == ETHERTYPE_VLAN) {
2077 		struct ether_vlan_header *evh = (void *)eh;
2078 
2079 		eh_type = ntohs(evh->evl_proto);
2080 		m0->m_pkthdr.l2hlen = sizeof(*evh);
2081 	} else
2082 		m0->m_pkthdr.l2hlen = sizeof(*eh);
2083 
2084 	offset = 0;
2085 	l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2086 
2087 	switch (eh_type) {
2088 #ifdef INET6
2089 	case ETHERTYPE_IPV6:
2090 	{
2091 		struct ip6_hdr *ip6 = l3hdr;
2092 
2093 		MPASS(ip6->ip6_nxt == IPPROTO_TCP);
2094 
2095 		m0->m_pkthdr.l3hlen = sizeof(*ip6);
2096 		break;
2097 	}
2098 #endif
2099 #ifdef INET
2100 	case ETHERTYPE_IP:
2101 	{
2102 		struct ip *ip = l3hdr;
2103 
2104 		m0->m_pkthdr.l3hlen = ip->ip_hl * 4;
2105 		break;
2106 	}
2107 #endif
2108 	default:
2109 		panic("%s: ethertype 0x%04x unknown.  if_cxgbe must be compiled"
2110 		    " with the same INET/INET6 options as the kernel.",
2111 		    __func__, eh_type);
2112 	}
2113 
2114 #if defined(INET) || defined(INET6)
2115 	tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen);
2116 	m0->m_pkthdr.l4hlen = tcp->th_off * 4;
2117 #endif
2118 	MPASS(m0 == *mp);
2119 	return (0);
2120 }
2121 
2122 void *
2123 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie)
2124 {
2125 	struct sge_eq *eq = &wrq->eq;
2126 	struct adapter *sc = wrq->adapter;
2127 	int ndesc, available;
2128 	struct wrqe *wr;
2129 	void *w;
2130 
2131 	MPASS(len16 > 0);
2132 	ndesc = howmany(len16, EQ_ESIZE / 16);
2133 	MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC);
2134 
2135 	EQ_LOCK(eq);
2136 
2137 	if (!STAILQ_EMPTY(&wrq->wr_list))
2138 		drain_wrq_wr_list(sc, wrq);
2139 
2140 	if (!STAILQ_EMPTY(&wrq->wr_list)) {
2141 slowpath:
2142 		EQ_UNLOCK(eq);
2143 		wr = alloc_wrqe(len16 * 16, wrq);
2144 		if (__predict_false(wr == NULL))
2145 			return (NULL);
2146 		cookie->pidx = -1;
2147 		cookie->ndesc = ndesc;
2148 		return (&wr->wr);
2149 	}
2150 
2151 	eq->cidx = read_hw_cidx(eq);
2152 	if (eq->pidx == eq->cidx)
2153 		available = eq->sidx - 1;
2154 	else
2155 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2156 	if (available < ndesc)
2157 		goto slowpath;
2158 
2159 	cookie->pidx = eq->pidx;
2160 	cookie->ndesc = ndesc;
2161 	TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link);
2162 
2163 	w = &eq->desc[eq->pidx];
2164 	IDXINCR(eq->pidx, ndesc, eq->sidx);
2165 	if (__predict_false(eq->pidx < ndesc - 1)) {
2166 		w = &wrq->ss[0];
2167 		wrq->ss_pidx = cookie->pidx;
2168 		wrq->ss_len = len16 * 16;
2169 	}
2170 
2171 	EQ_UNLOCK(eq);
2172 
2173 	return (w);
2174 }
2175 
2176 void
2177 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie)
2178 {
2179 	struct sge_eq *eq = &wrq->eq;
2180 	struct adapter *sc = wrq->adapter;
2181 	int ndesc, pidx;
2182 	struct wrq_cookie *prev, *next;
2183 
2184 	if (cookie->pidx == -1) {
2185 		struct wrqe *wr = __containerof(w, struct wrqe, wr);
2186 
2187 		t4_wrq_tx(sc, wr);
2188 		return;
2189 	}
2190 
2191 	ndesc = cookie->ndesc;	/* Can be more than SGE_MAX_WR_NDESC here. */
2192 	pidx = cookie->pidx;
2193 	MPASS(pidx >= 0 && pidx < eq->sidx);
2194 	if (__predict_false(w == &wrq->ss[0])) {
2195 		int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE;
2196 
2197 		MPASS(wrq->ss_len > n);	/* WR had better wrap around. */
2198 		bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n);
2199 		bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n);
2200 		wrq->tx_wrs_ss++;
2201 	} else
2202 		wrq->tx_wrs_direct++;
2203 
2204 	EQ_LOCK(eq);
2205 	prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link);
2206 	next = TAILQ_NEXT(cookie, link);
2207 	if (prev == NULL) {
2208 		MPASS(pidx == eq->dbidx);
2209 		if (next == NULL || ndesc >= 16)
2210 			ring_eq_db(wrq->adapter, eq, ndesc);
2211 		else {
2212 			MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc);
2213 			next->pidx = pidx;
2214 			next->ndesc += ndesc;
2215 		}
2216 	} else {
2217 		MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc);
2218 		prev->ndesc += ndesc;
2219 	}
2220 	TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link);
2221 
2222 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2223 		drain_wrq_wr_list(sc, wrq);
2224 
2225 #ifdef INVARIANTS
2226 	if (TAILQ_EMPTY(&wrq->incomplete_wrs)) {
2227 		/* Doorbell must have caught up to the pidx. */
2228 		MPASS(wrq->eq.pidx == wrq->eq.dbidx);
2229 	}
2230 #endif
2231 	EQ_UNLOCK(eq);
2232 }
2233 
2234 static u_int
2235 can_resume_eth_tx(struct mp_ring *r)
2236 {
2237 	struct sge_eq *eq = r->cookie;
2238 
2239 	return (total_available_tx_desc(eq) > eq->sidx / 8);
2240 }
2241 
2242 static inline int
2243 cannot_use_txpkts(struct mbuf *m)
2244 {
2245 	/* maybe put a GL limit too, to avoid silliness? */
2246 
2247 	return (needs_tso(m));
2248 }
2249 
2250 /*
2251  * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to
2252  * be consumed.  Return the actual number consumed.  0 indicates a stall.
2253  */
2254 static u_int
2255 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx)
2256 {
2257 	struct sge_txq *txq = r->cookie;
2258 	struct sge_eq *eq = &txq->eq;
2259 	struct ifnet *ifp = txq->ifp;
2260 	struct vi_info *vi = ifp->if_softc;
2261 	struct port_info *pi = vi->pi;
2262 	struct adapter *sc = pi->adapter;
2263 	u_int total, remaining;		/* # of packets */
2264 	u_int available, dbdiff;	/* # of hardware descriptors */
2265 	u_int n, next_cidx;
2266 	struct mbuf *m0, *tail;
2267 	struct txpkts txp;
2268 	struct fw_eth_tx_pkts_wr *wr;	/* any fw WR struct will do */
2269 
2270 	remaining = IDXDIFF(pidx, cidx, r->size);
2271 	MPASS(remaining > 0);	/* Must not be called without work to do. */
2272 	total = 0;
2273 
2274 	TXQ_LOCK(txq);
2275 	if (__predict_false((eq->flags & EQ_ENABLED) == 0)) {
2276 		while (cidx != pidx) {
2277 			m0 = r->items[cidx];
2278 			m_freem(m0);
2279 			if (++cidx == r->size)
2280 				cidx = 0;
2281 		}
2282 		reclaim_tx_descs(txq, 2048);
2283 		total = remaining;
2284 		goto done;
2285 	}
2286 
2287 	/* How many hardware descriptors do we have readily available. */
2288 	if (eq->pidx == eq->cidx)
2289 		available = eq->sidx - 1;
2290 	else
2291 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2292 	dbdiff = IDXDIFF(eq->pidx, eq->dbidx, eq->sidx);
2293 
2294 	while (remaining > 0) {
2295 
2296 		m0 = r->items[cidx];
2297 		M_ASSERTPKTHDR(m0);
2298 		MPASS(m0->m_nextpkt == NULL);
2299 
2300 		if (available < SGE_MAX_WR_NDESC) {
2301 			available += reclaim_tx_descs(txq, 64);
2302 			if (available < howmany(mbuf_len16(m0), EQ_ESIZE / 16))
2303 				break;	/* out of descriptors */
2304 		}
2305 
2306 		next_cidx = cidx + 1;
2307 		if (__predict_false(next_cidx == r->size))
2308 			next_cidx = 0;
2309 
2310 		wr = (void *)&eq->desc[eq->pidx];
2311 		if (remaining > 1 &&
2312 		    try_txpkts(m0, r->items[next_cidx], &txp, available) == 0) {
2313 
2314 			/* pkts at cidx, next_cidx should both be in txp. */
2315 			MPASS(txp.npkt == 2);
2316 			tail = r->items[next_cidx];
2317 			MPASS(tail->m_nextpkt == NULL);
2318 			ETHER_BPF_MTAP(ifp, m0);
2319 			ETHER_BPF_MTAP(ifp, tail);
2320 			m0->m_nextpkt = tail;
2321 
2322 			if (__predict_false(++next_cidx == r->size))
2323 				next_cidx = 0;
2324 
2325 			while (next_cidx != pidx) {
2326 				if (add_to_txpkts(r->items[next_cidx], &txp,
2327 				    available) != 0)
2328 					break;
2329 				tail->m_nextpkt = r->items[next_cidx];
2330 				tail = tail->m_nextpkt;
2331 				ETHER_BPF_MTAP(ifp, tail);
2332 				if (__predict_false(++next_cidx == r->size))
2333 					next_cidx = 0;
2334 			}
2335 
2336 			n = write_txpkts_wr(txq, wr, m0, &txp, available);
2337 			total += txp.npkt;
2338 			remaining -= txp.npkt;
2339 		} else {
2340 			total++;
2341 			remaining--;
2342 			ETHER_BPF_MTAP(ifp, m0);
2343 			n = write_txpkt_wr(txq, (void *)wr, m0, available);
2344 		}
2345 		MPASS(n >= 1 && n <= available && n <= SGE_MAX_WR_NDESC);
2346 
2347 		available -= n;
2348 		dbdiff += n;
2349 		IDXINCR(eq->pidx, n, eq->sidx);
2350 
2351 		if (total_available_tx_desc(eq) < eq->sidx / 4 &&
2352 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2353 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2354 			    F_FW_WR_EQUEQ);
2355 			eq->equeqidx = eq->pidx;
2356 		} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
2357 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
2358 			eq->equeqidx = eq->pidx;
2359 		}
2360 
2361 		if (dbdiff >= 16 && remaining >= 4) {
2362 			ring_eq_db(sc, eq, dbdiff);
2363 			available += reclaim_tx_descs(txq, 4 * dbdiff);
2364 			dbdiff = 0;
2365 		}
2366 
2367 		cidx = next_cidx;
2368 	}
2369 	if (dbdiff != 0) {
2370 		ring_eq_db(sc, eq, dbdiff);
2371 		reclaim_tx_descs(txq, 32);
2372 	}
2373 done:
2374 	TXQ_UNLOCK(txq);
2375 
2376 	return (total);
2377 }
2378 
2379 static inline void
2380 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
2381     int qsize)
2382 {
2383 
2384 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
2385 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
2386 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
2387 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
2388 
2389 	iq->flags = 0;
2390 	iq->adapter = sc;
2391 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
2392 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
2393 	if (pktc_idx >= 0) {
2394 		iq->intr_params |= F_QINTR_CNT_EN;
2395 		iq->intr_pktc_idx = pktc_idx;
2396 	}
2397 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
2398 	iq->sidx = iq->qsize - sc->params.sge.spg_len / IQ_ESIZE;
2399 }
2400 
2401 static inline void
2402 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name)
2403 {
2404 
2405 	fl->qsize = qsize;
2406 	fl->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2407 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
2408 	if (sc->flags & BUF_PACKING_OK &&
2409 	    ((!is_t4(sc) && buffer_packing) ||	/* T5+: enabled unless 0 */
2410 	    (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */
2411 		fl->flags |= FL_BUF_PACKING;
2412 	find_best_refill_source(sc, fl, maxp);
2413 	find_safe_refill_source(sc, fl);
2414 }
2415 
2416 static inline void
2417 init_eq(struct adapter *sc, struct sge_eq *eq, int eqtype, int qsize,
2418     uint8_t tx_chan, uint16_t iqid, char *name)
2419 {
2420 	KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
2421 
2422 	eq->flags = eqtype & EQ_TYPEMASK;
2423 	eq->tx_chan = tx_chan;
2424 	eq->iqid = iqid;
2425 	eq->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2426 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
2427 }
2428 
2429 static int
2430 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
2431     bus_dmamap_t *map, bus_addr_t *pa, void **va)
2432 {
2433 	int rc;
2434 
2435 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
2436 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
2437 	if (rc != 0) {
2438 		device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
2439 		goto done;
2440 	}
2441 
2442 	rc = bus_dmamem_alloc(*tag, va,
2443 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
2444 	if (rc != 0) {
2445 		device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
2446 		goto done;
2447 	}
2448 
2449 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
2450 	if (rc != 0) {
2451 		device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
2452 		goto done;
2453 	}
2454 done:
2455 	if (rc)
2456 		free_ring(sc, *tag, *map, *pa, *va);
2457 
2458 	return (rc);
2459 }
2460 
2461 static int
2462 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
2463     bus_addr_t pa, void *va)
2464 {
2465 	if (pa)
2466 		bus_dmamap_unload(tag, map);
2467 	if (va)
2468 		bus_dmamem_free(tag, va, map);
2469 	if (tag)
2470 		bus_dma_tag_destroy(tag);
2471 
2472 	return (0);
2473 }
2474 
2475 /*
2476  * Allocates the ring for an ingress queue and an optional freelist.  If the
2477  * freelist is specified it will be allocated and then associated with the
2478  * ingress queue.
2479  *
2480  * Returns errno on failure.  Resources allocated up to that point may still be
2481  * allocated.  Caller is responsible for cleanup in case this function fails.
2482  *
2483  * If the ingress queue will take interrupts directly (iq->flags & IQ_INTR) then
2484  * the intr_idx specifies the vector, starting from 0.  Otherwise it specifies
2485  * the abs_id of the ingress queue to which its interrupts should be forwarded.
2486  */
2487 static int
2488 alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl,
2489     int intr_idx, int cong)
2490 {
2491 	int rc, i, cntxt_id;
2492 	size_t len;
2493 	struct fw_iq_cmd c;
2494 	struct port_info *pi = vi->pi;
2495 	struct adapter *sc = iq->adapter;
2496 	struct sge_params *sp = &sc->params.sge;
2497 	__be32 v = 0;
2498 
2499 	len = iq->qsize * IQ_ESIZE;
2500 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
2501 	    (void **)&iq->desc);
2502 	if (rc != 0)
2503 		return (rc);
2504 
2505 	bzero(&c, sizeof(c));
2506 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
2507 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
2508 	    V_FW_IQ_CMD_VFN(0));
2509 
2510 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
2511 	    FW_LEN16(c));
2512 
2513 	/* Special handling for firmware event queue */
2514 	if (iq == &sc->sge.fwq)
2515 		v |= F_FW_IQ_CMD_IQASYNCH;
2516 
2517 	if (iq->flags & IQ_INTR) {
2518 		KASSERT(intr_idx < sc->intr_count,
2519 		    ("%s: invalid direct intr_idx %d", __func__, intr_idx));
2520 	} else
2521 		v |= F_FW_IQ_CMD_IQANDST;
2522 	v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
2523 
2524 	c.type_to_iqandstindex = htobe32(v |
2525 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
2526 	    V_FW_IQ_CMD_VIID(vi->viid) |
2527 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
2528 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
2529 	    F_FW_IQ_CMD_IQGTSMODE |
2530 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
2531 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
2532 	c.iqsize = htobe16(iq->qsize);
2533 	c.iqaddr = htobe64(iq->ba);
2534 	if (cong >= 0)
2535 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
2536 
2537 	if (fl) {
2538 		mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
2539 
2540 		len = fl->qsize * EQ_ESIZE;
2541 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
2542 		    &fl->ba, (void **)&fl->desc);
2543 		if (rc)
2544 			return (rc);
2545 
2546 		/* Allocate space for one software descriptor per buffer. */
2547 		rc = alloc_fl_sdesc(fl);
2548 		if (rc != 0) {
2549 			device_printf(sc->dev,
2550 			    "failed to setup fl software descriptors: %d\n",
2551 			    rc);
2552 			return (rc);
2553 		}
2554 
2555 		if (fl->flags & FL_BUF_PACKING) {
2556 			fl->lowat = roundup2(sp->fl_starve_threshold2, 8);
2557 			fl->buf_boundary = sp->pack_boundary;
2558 		} else {
2559 			fl->lowat = roundup2(sp->fl_starve_threshold, 8);
2560 			fl->buf_boundary = 16;
2561 		}
2562 		if (fl_pad && fl->buf_boundary < sp->pad_boundary)
2563 			fl->buf_boundary = sp->pad_boundary;
2564 
2565 		c.iqns_to_fl0congen |=
2566 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
2567 			F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
2568 			(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
2569 			(fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
2570 			    0));
2571 		if (cong >= 0) {
2572 			c.iqns_to_fl0congen |=
2573 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
2574 				    F_FW_IQ_CMD_FL0CONGCIF |
2575 				    F_FW_IQ_CMD_FL0CONGEN);
2576 		}
2577 		c.fl0dcaen_to_fl0cidxfthresh =
2578 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(X_FETCHBURSTMIN_128B) |
2579 			V_FW_IQ_CMD_FL0FBMAX(X_FETCHBURSTMAX_512B));
2580 		c.fl0size = htobe16(fl->qsize);
2581 		c.fl0addr = htobe64(fl->ba);
2582 	}
2583 
2584 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
2585 	if (rc != 0) {
2586 		device_printf(sc->dev,
2587 		    "failed to create ingress queue: %d\n", rc);
2588 		return (rc);
2589 	}
2590 
2591 	iq->cidx = 0;
2592 	iq->gen = F_RSPD_GEN;
2593 	iq->intr_next = iq->intr_params;
2594 	iq->cntxt_id = be16toh(c.iqid);
2595 	iq->abs_id = be16toh(c.physiqid);
2596 	iq->flags |= IQ_ALLOCATED;
2597 
2598 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
2599 	if (cntxt_id >= sc->sge.niq) {
2600 		panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
2601 		    cntxt_id, sc->sge.niq - 1);
2602 	}
2603 	sc->sge.iqmap[cntxt_id] = iq;
2604 
2605 	if (fl) {
2606 		u_int qid;
2607 
2608 		iq->flags |= IQ_HAS_FL;
2609 		fl->cntxt_id = be16toh(c.fl0id);
2610 		fl->pidx = fl->cidx = 0;
2611 
2612 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
2613 		if (cntxt_id >= sc->sge.neq) {
2614 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
2615 			    __func__, cntxt_id, sc->sge.neq - 1);
2616 		}
2617 		sc->sge.eqmap[cntxt_id] = (void *)fl;
2618 
2619 		qid = fl->cntxt_id;
2620 		if (isset(&sc->doorbells, DOORBELL_UDB)) {
2621 			uint32_t s_qpp = sc->params.sge.eq_s_qpp;
2622 			uint32_t mask = (1 << s_qpp) - 1;
2623 			volatile uint8_t *udb;
2624 
2625 			udb = sc->udbs_base + UDBS_DB_OFFSET;
2626 			udb += (qid >> s_qpp) << PAGE_SHIFT;
2627 			qid &= mask;
2628 			if (qid < PAGE_SIZE / UDBS_SEG_SIZE) {
2629 				udb += qid << UDBS_SEG_SHIFT;
2630 				qid = 0;
2631 			}
2632 			fl->udb = (volatile void *)udb;
2633 		}
2634 		fl->dbval = V_QID(qid) | sc->chip_params->sge_fl_db;
2635 
2636 		FL_LOCK(fl);
2637 		/* Enough to make sure the SGE doesn't think it's starved */
2638 		refill_fl(sc, fl, fl->lowat);
2639 		FL_UNLOCK(fl);
2640 	}
2641 
2642 	if (is_t5(sc) && cong >= 0) {
2643 		uint32_t param, val;
2644 
2645 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
2646 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
2647 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
2648 		if (cong == 0)
2649 			val = 1 << 19;
2650 		else {
2651 			val = 2 << 19;
2652 			for (i = 0; i < 4; i++) {
2653 				if (cong & (1 << i))
2654 					val |= 1 << (i << 2);
2655 			}
2656 		}
2657 
2658 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
2659 		if (rc != 0) {
2660 			/* report error but carry on */
2661 			device_printf(sc->dev,
2662 			    "failed to set congestion manager context for "
2663 			    "ingress queue %d: %d\n", iq->cntxt_id, rc);
2664 		}
2665 	}
2666 
2667 	/* Enable IQ interrupts */
2668 	atomic_store_rel_int(&iq->state, IQS_IDLE);
2669 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_SEINTARM(iq->intr_params) |
2670 	    V_INGRESSQID(iq->cntxt_id));
2671 
2672 	return (0);
2673 }
2674 
2675 static int
2676 free_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl)
2677 {
2678 	int rc;
2679 	struct adapter *sc = iq->adapter;
2680 	device_t dev;
2681 
2682 	if (sc == NULL)
2683 		return (0);	/* nothing to do */
2684 
2685 	dev = vi ? vi->dev : sc->dev;
2686 
2687 	if (iq->flags & IQ_ALLOCATED) {
2688 		rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
2689 		    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
2690 		    fl ? fl->cntxt_id : 0xffff, 0xffff);
2691 		if (rc != 0) {
2692 			device_printf(dev,
2693 			    "failed to free queue %p: %d\n", iq, rc);
2694 			return (rc);
2695 		}
2696 		iq->flags &= ~IQ_ALLOCATED;
2697 	}
2698 
2699 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
2700 
2701 	bzero(iq, sizeof(*iq));
2702 
2703 	if (fl) {
2704 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
2705 		    fl->desc);
2706 
2707 		if (fl->sdesc)
2708 			free_fl_sdesc(sc, fl);
2709 
2710 		if (mtx_initialized(&fl->fl_lock))
2711 			mtx_destroy(&fl->fl_lock);
2712 
2713 		bzero(fl, sizeof(*fl));
2714 	}
2715 
2716 	return (0);
2717 }
2718 
2719 static void
2720 add_fl_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
2721     struct sge_fl *fl)
2722 {
2723 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2724 
2725 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
2726 	    "freelist");
2727 	children = SYSCTL_CHILDREN(oid);
2728 
2729 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
2730 	    CTLTYPE_INT | CTLFLAG_RD, &fl->cntxt_id, 0, sysctl_uint16, "I",
2731 	    "SGE context id of the freelist");
2732 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL,
2733 	    fl_pad ? 1 : 0, "padding enabled");
2734 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL,
2735 	    fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled");
2736 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
2737 	    0, "consumer index");
2738 	if (fl->flags & FL_BUF_PACKING) {
2739 		SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
2740 		    CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
2741 	}
2742 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
2743 	    0, "producer index");
2744 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_allocated",
2745 	    CTLFLAG_RD, &fl->mbuf_allocated, "# of mbuf allocated");
2746 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_inlined",
2747 	    CTLFLAG_RD, &fl->mbuf_inlined, "# of mbuf inlined in clusters");
2748 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
2749 	    CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
2750 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
2751 	    CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
2752 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
2753 	    CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
2754 }
2755 
2756 static int
2757 alloc_fwq(struct adapter *sc)
2758 {
2759 	int rc, intr_idx;
2760 	struct sge_iq *fwq = &sc->sge.fwq;
2761 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
2762 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2763 
2764 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE);
2765 	fwq->flags |= IQ_INTR;	/* always */
2766 	intr_idx = sc->intr_count > 1 ? 1 : 0;
2767 	rc = alloc_iq_fl(&sc->port[0]->vi[0], fwq, NULL, intr_idx, -1);
2768 	if (rc != 0) {
2769 		device_printf(sc->dev,
2770 		    "failed to create firmware event queue: %d\n", rc);
2771 		return (rc);
2772 	}
2773 
2774 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD,
2775 	    NULL, "firmware event queue");
2776 	children = SYSCTL_CHILDREN(oid);
2777 
2778 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "abs_id",
2779 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->abs_id, 0, sysctl_uint16, "I",
2780 	    "absolute id of the queue");
2781 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cntxt_id",
2782 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->cntxt_id, 0, sysctl_uint16, "I",
2783 	    "SGE context id of the queue");
2784 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cidx",
2785 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->cidx, 0, sysctl_uint16, "I",
2786 	    "consumer index");
2787 
2788 	return (0);
2789 }
2790 
2791 static int
2792 free_fwq(struct adapter *sc)
2793 {
2794 	return free_iq_fl(NULL, &sc->sge.fwq, NULL);
2795 }
2796 
2797 static int
2798 alloc_mgmtq(struct adapter *sc)
2799 {
2800 	int rc;
2801 	struct sge_wrq *mgmtq = &sc->sge.mgmtq;
2802 	char name[16];
2803 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
2804 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2805 
2806 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "mgmtq", CTLFLAG_RD,
2807 	    NULL, "management queue");
2808 
2809 	snprintf(name, sizeof(name), "%s mgmtq", device_get_nameunit(sc->dev));
2810 	init_eq(sc, &mgmtq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[0]->tx_chan,
2811 	    sc->sge.fwq.cntxt_id, name);
2812 	rc = alloc_wrq(sc, NULL, mgmtq, oid);
2813 	if (rc != 0) {
2814 		device_printf(sc->dev,
2815 		    "failed to create management queue: %d\n", rc);
2816 		return (rc);
2817 	}
2818 
2819 	return (0);
2820 }
2821 
2822 static int
2823 free_mgmtq(struct adapter *sc)
2824 {
2825 
2826 	return free_wrq(sc, &sc->sge.mgmtq);
2827 }
2828 
2829 int
2830 tnl_cong(struct port_info *pi, int drop)
2831 {
2832 
2833 	if (drop == -1)
2834 		return (-1);
2835 	else if (drop == 1)
2836 		return (0);
2837 	else
2838 		return (pi->rx_chan_map);
2839 }
2840 
2841 static int
2842 alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int intr_idx, int idx,
2843     struct sysctl_oid *oid)
2844 {
2845 	int rc;
2846 	struct sysctl_oid_list *children;
2847 	char name[16];
2848 
2849 	rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, intr_idx,
2850 	    tnl_cong(vi->pi, cong_drop));
2851 	if (rc != 0)
2852 		return (rc);
2853 
2854 	/*
2855 	 * The freelist is just barely above the starvation threshold right now,
2856 	 * fill it up a bit more.
2857 	 */
2858 	FL_LOCK(&rxq->fl);
2859 	refill_fl(vi->pi->adapter, &rxq->fl, 128);
2860 	FL_UNLOCK(&rxq->fl);
2861 
2862 #if defined(INET) || defined(INET6)
2863 	rc = tcp_lro_init(&rxq->lro);
2864 	if (rc != 0)
2865 		return (rc);
2866 	rxq->lro.ifp = vi->ifp; /* also indicates LRO init'ed */
2867 
2868 	if (vi->ifp->if_capenable & IFCAP_LRO)
2869 		rxq->iq.flags |= IQ_LRO_ENABLED;
2870 #endif
2871 	rxq->ifp = vi->ifp;
2872 
2873 	children = SYSCTL_CHILDREN(oid);
2874 
2875 	snprintf(name, sizeof(name), "%d", idx);
2876 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
2877 	    NULL, "rx queue");
2878 	children = SYSCTL_CHILDREN(oid);
2879 
2880 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "abs_id",
2881 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.abs_id, 0, sysctl_uint16, "I",
2882 	    "absolute id of the queue");
2883 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cntxt_id",
2884 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cntxt_id, 0, sysctl_uint16, "I",
2885 	    "SGE context id of the queue");
2886 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
2887 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cidx, 0, sysctl_uint16, "I",
2888 	    "consumer index");
2889 #if defined(INET) || defined(INET6)
2890 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
2891 	    &rxq->lro.lro_queued, 0, NULL);
2892 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
2893 	    &rxq->lro.lro_flushed, 0, NULL);
2894 #endif
2895 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
2896 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
2897 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_extraction",
2898 	    CTLFLAG_RD, &rxq->vlan_extraction,
2899 	    "# of times hardware extracted 802.1Q tag");
2900 
2901 	add_fl_sysctls(&vi->ctx, oid, &rxq->fl);
2902 
2903 	return (rc);
2904 }
2905 
2906 static int
2907 free_rxq(struct vi_info *vi, struct sge_rxq *rxq)
2908 {
2909 	int rc;
2910 
2911 #if defined(INET) || defined(INET6)
2912 	if (rxq->lro.ifp) {
2913 		tcp_lro_free(&rxq->lro);
2914 		rxq->lro.ifp = NULL;
2915 	}
2916 #endif
2917 
2918 	rc = free_iq_fl(vi, &rxq->iq, &rxq->fl);
2919 	if (rc == 0)
2920 		bzero(rxq, sizeof(*rxq));
2921 
2922 	return (rc);
2923 }
2924 
2925 #ifdef TCP_OFFLOAD
2926 static int
2927 alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq,
2928     int intr_idx, int idx, struct sysctl_oid *oid)
2929 {
2930 	int rc;
2931 	struct sysctl_oid_list *children;
2932 	char name[16];
2933 
2934 	rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx,
2935 	    vi->pi->rx_chan_map);
2936 	if (rc != 0)
2937 		return (rc);
2938 
2939 	children = SYSCTL_CHILDREN(oid);
2940 
2941 	snprintf(name, sizeof(name), "%d", idx);
2942 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
2943 	    NULL, "rx queue");
2944 	children = SYSCTL_CHILDREN(oid);
2945 
2946 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "abs_id",
2947 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.abs_id, 0, sysctl_uint16,
2948 	    "I", "absolute id of the queue");
2949 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cntxt_id",
2950 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cntxt_id, 0, sysctl_uint16,
2951 	    "I", "SGE context id of the queue");
2952 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
2953 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cidx, 0, sysctl_uint16, "I",
2954 	    "consumer index");
2955 
2956 	add_fl_sysctls(&vi->ctx, oid, &ofld_rxq->fl);
2957 
2958 	return (rc);
2959 }
2960 
2961 static int
2962 free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq)
2963 {
2964 	int rc;
2965 
2966 	rc = free_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl);
2967 	if (rc == 0)
2968 		bzero(ofld_rxq, sizeof(*ofld_rxq));
2969 
2970 	return (rc);
2971 }
2972 #endif
2973 
2974 #ifdef DEV_NETMAP
2975 static int
2976 alloc_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int intr_idx,
2977     int idx, struct sysctl_oid *oid)
2978 {
2979 	int rc;
2980 	struct sysctl_oid_list *children;
2981 	struct sysctl_ctx_list *ctx;
2982 	char name[16];
2983 	size_t len;
2984 	struct adapter *sc = vi->pi->adapter;
2985 	struct netmap_adapter *na = NA(vi->ifp);
2986 
2987 	MPASS(na != NULL);
2988 
2989 	len = vi->qsize_rxq * IQ_ESIZE;
2990 	rc = alloc_ring(sc, len, &nm_rxq->iq_desc_tag, &nm_rxq->iq_desc_map,
2991 	    &nm_rxq->iq_ba, (void **)&nm_rxq->iq_desc);
2992 	if (rc != 0)
2993 		return (rc);
2994 
2995 	len = na->num_rx_desc * EQ_ESIZE + sc->params.sge.spg_len;
2996 	rc = alloc_ring(sc, len, &nm_rxq->fl_desc_tag, &nm_rxq->fl_desc_map,
2997 	    &nm_rxq->fl_ba, (void **)&nm_rxq->fl_desc);
2998 	if (rc != 0)
2999 		return (rc);
3000 
3001 	nm_rxq->vi = vi;
3002 	nm_rxq->nid = idx;
3003 	nm_rxq->iq_cidx = 0;
3004 	nm_rxq->iq_sidx = vi->qsize_rxq - sc->params.sge.spg_len / IQ_ESIZE;
3005 	nm_rxq->iq_gen = F_RSPD_GEN;
3006 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
3007 	nm_rxq->fl_sidx = na->num_rx_desc;
3008 	nm_rxq->intr_idx = intr_idx;
3009 
3010 	ctx = &vi->ctx;
3011 	children = SYSCTL_CHILDREN(oid);
3012 
3013 	snprintf(name, sizeof(name), "%d", idx);
3014 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL,
3015 	    "rx queue");
3016 	children = SYSCTL_CHILDREN(oid);
3017 
3018 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3019 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_abs_id, 0, sysctl_uint16,
3020 	    "I", "absolute id of the queue");
3021 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3022 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cntxt_id, 0, sysctl_uint16,
3023 	    "I", "SGE context id of the queue");
3024 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3025 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cidx, 0, sysctl_uint16, "I",
3026 	    "consumer index");
3027 
3028 	children = SYSCTL_CHILDREN(oid);
3029 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3030 	    "freelist");
3031 	children = SYSCTL_CHILDREN(oid);
3032 
3033 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3034 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->fl_cntxt_id, 0, sysctl_uint16,
3035 	    "I", "SGE context id of the freelist");
3036 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
3037 	    &nm_rxq->fl_cidx, 0, "consumer index");
3038 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
3039 	    &nm_rxq->fl_pidx, 0, "producer index");
3040 
3041 	return (rc);
3042 }
3043 
3044 
3045 static int
3046 free_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
3047 {
3048 	struct adapter *sc = vi->pi->adapter;
3049 
3050 	free_ring(sc, nm_rxq->iq_desc_tag, nm_rxq->iq_desc_map, nm_rxq->iq_ba,
3051 	    nm_rxq->iq_desc);
3052 	free_ring(sc, nm_rxq->fl_desc_tag, nm_rxq->fl_desc_map, nm_rxq->fl_ba,
3053 	    nm_rxq->fl_desc);
3054 
3055 	return (0);
3056 }
3057 
3058 static int
3059 alloc_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq, int iqidx, int idx,
3060     struct sysctl_oid *oid)
3061 {
3062 	int rc;
3063 	size_t len;
3064 	struct port_info *pi = vi->pi;
3065 	struct adapter *sc = pi->adapter;
3066 	struct netmap_adapter *na = NA(vi->ifp);
3067 	char name[16];
3068 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3069 
3070 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3071 	rc = alloc_ring(sc, len, &nm_txq->desc_tag, &nm_txq->desc_map,
3072 	    &nm_txq->ba, (void **)&nm_txq->desc);
3073 	if (rc)
3074 		return (rc);
3075 
3076 	nm_txq->pidx = nm_txq->cidx = 0;
3077 	nm_txq->sidx = na->num_tx_desc;
3078 	nm_txq->nid = idx;
3079 	nm_txq->iqidx = iqidx;
3080 	nm_txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3081 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_VF_VLD(1) |
3082 	    V_TXPKT_VF(vi->viid));
3083 
3084 	snprintf(name, sizeof(name), "%d", idx);
3085 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3086 	    NULL, "netmap tx queue");
3087 	children = SYSCTL_CHILDREN(oid);
3088 
3089 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3090 	    &nm_txq->cntxt_id, 0, "SGE context id of the queue");
3091 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3092 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->cidx, 0, sysctl_uint16, "I",
3093 	    "consumer index");
3094 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3095 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->pidx, 0, sysctl_uint16, "I",
3096 	    "producer index");
3097 
3098 	return (rc);
3099 }
3100 
3101 static int
3102 free_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
3103 {
3104 	struct adapter *sc = vi->pi->adapter;
3105 
3106 	free_ring(sc, nm_txq->desc_tag, nm_txq->desc_map, nm_txq->ba,
3107 	    nm_txq->desc);
3108 
3109 	return (0);
3110 }
3111 #endif
3112 
3113 static int
3114 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
3115 {
3116 	int rc, cntxt_id;
3117 	struct fw_eq_ctrl_cmd c;
3118 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3119 
3120 	bzero(&c, sizeof(c));
3121 
3122 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
3123 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
3124 	    V_FW_EQ_CTRL_CMD_VFN(0));
3125 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
3126 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
3127 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid));
3128 	c.physeqid_pkd = htobe32(0);
3129 	c.fetchszm_to_iqid =
3130 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3131 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
3132 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
3133 	c.dcaen_to_eqsize =
3134 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3135 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3136 		V_FW_EQ_CTRL_CMD_EQSIZE(qsize));
3137 	c.eqaddr = htobe64(eq->ba);
3138 
3139 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3140 	if (rc != 0) {
3141 		device_printf(sc->dev,
3142 		    "failed to create control queue %d: %d\n", eq->tx_chan, rc);
3143 		return (rc);
3144 	}
3145 	eq->flags |= EQ_ALLOCATED;
3146 
3147 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
3148 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3149 	if (cntxt_id >= sc->sge.neq)
3150 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3151 		cntxt_id, sc->sge.neq - 1);
3152 	sc->sge.eqmap[cntxt_id] = eq;
3153 
3154 	return (rc);
3155 }
3156 
3157 static int
3158 eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3159 {
3160 	int rc, cntxt_id;
3161 	struct fw_eq_eth_cmd c;
3162 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3163 
3164 	bzero(&c, sizeof(c));
3165 
3166 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
3167 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
3168 	    V_FW_EQ_ETH_CMD_VFN(0));
3169 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
3170 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
3171 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
3172 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
3173 	c.fetchszm_to_iqid =
3174 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3175 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
3176 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
3177 	c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3178 	    V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3179 	    V_FW_EQ_ETH_CMD_EQSIZE(qsize));
3180 	c.eqaddr = htobe64(eq->ba);
3181 
3182 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3183 	if (rc != 0) {
3184 		device_printf(vi->dev,
3185 		    "failed to create Ethernet egress queue: %d\n", rc);
3186 		return (rc);
3187 	}
3188 	eq->flags |= EQ_ALLOCATED;
3189 
3190 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
3191 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3192 	if (cntxt_id >= sc->sge.neq)
3193 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3194 		cntxt_id, sc->sge.neq - 1);
3195 	sc->sge.eqmap[cntxt_id] = eq;
3196 
3197 	return (rc);
3198 }
3199 
3200 #ifdef TCP_OFFLOAD
3201 static int
3202 ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3203 {
3204 	int rc, cntxt_id;
3205 	struct fw_eq_ofld_cmd c;
3206 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3207 
3208 	bzero(&c, sizeof(c));
3209 
3210 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
3211 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
3212 	    V_FW_EQ_OFLD_CMD_VFN(0));
3213 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
3214 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
3215 	c.fetchszm_to_iqid =
3216 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3217 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
3218 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
3219 	c.dcaen_to_eqsize =
3220 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3221 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3222 		V_FW_EQ_OFLD_CMD_EQSIZE(qsize));
3223 	c.eqaddr = htobe64(eq->ba);
3224 
3225 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3226 	if (rc != 0) {
3227 		device_printf(vi->dev,
3228 		    "failed to create egress queue for TCP offload: %d\n", rc);
3229 		return (rc);
3230 	}
3231 	eq->flags |= EQ_ALLOCATED;
3232 
3233 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
3234 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3235 	if (cntxt_id >= sc->sge.neq)
3236 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3237 		cntxt_id, sc->sge.neq - 1);
3238 	sc->sge.eqmap[cntxt_id] = eq;
3239 
3240 	return (rc);
3241 }
3242 #endif
3243 
3244 static int
3245 alloc_eq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3246 {
3247 	int rc, qsize;
3248 	size_t len;
3249 
3250 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
3251 
3252 	qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3253 	len = qsize * EQ_ESIZE;
3254 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
3255 	    &eq->ba, (void **)&eq->desc);
3256 	if (rc)
3257 		return (rc);
3258 
3259 	eq->pidx = eq->cidx = 0;
3260 	eq->equeqidx = eq->dbidx = 0;
3261 	eq->doorbells = sc->doorbells;
3262 
3263 	switch (eq->flags & EQ_TYPEMASK) {
3264 	case EQ_CTRL:
3265 		rc = ctrl_eq_alloc(sc, eq);
3266 		break;
3267 
3268 	case EQ_ETH:
3269 		rc = eth_eq_alloc(sc, vi, eq);
3270 		break;
3271 
3272 #ifdef TCP_OFFLOAD
3273 	case EQ_OFLD:
3274 		rc = ofld_eq_alloc(sc, vi, eq);
3275 		break;
3276 #endif
3277 
3278 	default:
3279 		panic("%s: invalid eq type %d.", __func__,
3280 		    eq->flags & EQ_TYPEMASK);
3281 	}
3282 	if (rc != 0) {
3283 		device_printf(sc->dev,
3284 		    "failed to allocate egress queue(%d): %d\n",
3285 		    eq->flags & EQ_TYPEMASK, rc);
3286 	}
3287 
3288 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
3289 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
3290 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
3291 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3292 		uint32_t mask = (1 << s_qpp) - 1;
3293 		volatile uint8_t *udb;
3294 
3295 		udb = sc->udbs_base + UDBS_DB_OFFSET;
3296 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
3297 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
3298 		if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
3299 	    		clrbit(&eq->doorbells, DOORBELL_WCWR);
3300 		else {
3301 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
3302 			eq->udb_qid = 0;
3303 		}
3304 		eq->udb = (volatile void *)udb;
3305 	}
3306 
3307 	return (rc);
3308 }
3309 
3310 static int
3311 free_eq(struct adapter *sc, struct sge_eq *eq)
3312 {
3313 	int rc;
3314 
3315 	if (eq->flags & EQ_ALLOCATED) {
3316 		switch (eq->flags & EQ_TYPEMASK) {
3317 		case EQ_CTRL:
3318 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
3319 			    eq->cntxt_id);
3320 			break;
3321 
3322 		case EQ_ETH:
3323 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
3324 			    eq->cntxt_id);
3325 			break;
3326 
3327 #ifdef TCP_OFFLOAD
3328 		case EQ_OFLD:
3329 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
3330 			    eq->cntxt_id);
3331 			break;
3332 #endif
3333 
3334 		default:
3335 			panic("%s: invalid eq type %d.", __func__,
3336 			    eq->flags & EQ_TYPEMASK);
3337 		}
3338 		if (rc != 0) {
3339 			device_printf(sc->dev,
3340 			    "failed to free egress queue (%d): %d\n",
3341 			    eq->flags & EQ_TYPEMASK, rc);
3342 			return (rc);
3343 		}
3344 		eq->flags &= ~EQ_ALLOCATED;
3345 	}
3346 
3347 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
3348 
3349 	if (mtx_initialized(&eq->eq_lock))
3350 		mtx_destroy(&eq->eq_lock);
3351 
3352 	bzero(eq, sizeof(*eq));
3353 	return (0);
3354 }
3355 
3356 static int
3357 alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq,
3358     struct sysctl_oid *oid)
3359 {
3360 	int rc;
3361 	struct sysctl_ctx_list *ctx = vi ? &vi->ctx : &sc->ctx;
3362 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3363 
3364 	rc = alloc_eq(sc, vi, &wrq->eq);
3365 	if (rc)
3366 		return (rc);
3367 
3368 	wrq->adapter = sc;
3369 	TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq);
3370 	TAILQ_INIT(&wrq->incomplete_wrs);
3371 	STAILQ_INIT(&wrq->wr_list);
3372 	wrq->nwr_pending = 0;
3373 	wrq->ndesc_needed = 0;
3374 
3375 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3376 	    &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
3377 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3378 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I",
3379 	    "consumer index");
3380 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx",
3381 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I",
3382 	    "producer index");
3383 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD,
3384 	    &wrq->tx_wrs_direct, "# of work requests (direct)");
3385 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD,
3386 	    &wrq->tx_wrs_copied, "# of work requests (copied)");
3387 
3388 	return (rc);
3389 }
3390 
3391 static int
3392 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
3393 {
3394 	int rc;
3395 
3396 	rc = free_eq(sc, &wrq->eq);
3397 	if (rc)
3398 		return (rc);
3399 
3400 	bzero(wrq, sizeof(*wrq));
3401 	return (0);
3402 }
3403 
3404 static int
3405 alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx,
3406     struct sysctl_oid *oid)
3407 {
3408 	int rc;
3409 	struct port_info *pi = vi->pi;
3410 	struct adapter *sc = pi->adapter;
3411 	struct sge_eq *eq = &txq->eq;
3412 	char name[16];
3413 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3414 
3415 	rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx, can_resume_eth_tx,
3416 	    M_CXGBE, M_WAITOK);
3417 	if (rc != 0) {
3418 		device_printf(sc->dev, "failed to allocate mp_ring: %d\n", rc);
3419 		return (rc);
3420 	}
3421 
3422 	rc = alloc_eq(sc, vi, eq);
3423 	if (rc != 0) {
3424 		mp_ring_free(txq->r);
3425 		txq->r = NULL;
3426 		return (rc);
3427 	}
3428 
3429 	/* Can't fail after this point. */
3430 
3431 	TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq);
3432 	txq->ifp = vi->ifp;
3433 	txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK);
3434 	txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3435 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_VF_VLD(1) |
3436 	    V_TXPKT_VF(vi->viid));
3437 	txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE,
3438 	    M_ZERO | M_WAITOK);
3439 
3440 	snprintf(name, sizeof(name), "%d", idx);
3441 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3442 	    NULL, "tx queue");
3443 	children = SYSCTL_CHILDREN(oid);
3444 
3445 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3446 	    &eq->cntxt_id, 0, "SGE context id of the queue");
3447 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3448 	    CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I",
3449 	    "consumer index");
3450 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3451 	    CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I",
3452 	    "producer index");
3453 
3454 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
3455 	    &txq->txcsum, "# of times hardware assisted with checksum");
3456 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_insertion",
3457 	    CTLFLAG_RD, &txq->vlan_insertion,
3458 	    "# of times hardware inserted 802.1Q tag");
3459 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
3460 	    &txq->tso_wrs, "# of TSO work requests");
3461 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
3462 	    &txq->imm_wrs, "# of work requests with immediate data");
3463 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
3464 	    &txq->sgl_wrs, "# of work requests with direct SGL");
3465 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
3466 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
3467 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_wrs",
3468 	    CTLFLAG_RD, &txq->txpkts0_wrs,
3469 	    "# of txpkts (type 0) work requests");
3470 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_wrs",
3471 	    CTLFLAG_RD, &txq->txpkts1_wrs,
3472 	    "# of txpkts (type 1) work requests");
3473 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_pkts",
3474 	    CTLFLAG_RD, &txq->txpkts0_pkts,
3475 	    "# of frames tx'd using type0 txpkts work requests");
3476 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_pkts",
3477 	    CTLFLAG_RD, &txq->txpkts1_pkts,
3478 	    "# of frames tx'd using type1 txpkts work requests");
3479 
3480 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_enqueues",
3481 	    CTLFLAG_RD, &txq->r->enqueues,
3482 	    "# of enqueues to the mp_ring for this queue");
3483 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_drops",
3484 	    CTLFLAG_RD, &txq->r->drops,
3485 	    "# of drops in the mp_ring for this queue");
3486 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_starts",
3487 	    CTLFLAG_RD, &txq->r->starts,
3488 	    "# of normal consumer starts in the mp_ring for this queue");
3489 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_stalls",
3490 	    CTLFLAG_RD, &txq->r->stalls,
3491 	    "# of consumer stalls in the mp_ring for this queue");
3492 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_restarts",
3493 	    CTLFLAG_RD, &txq->r->restarts,
3494 	    "# of consumer restarts in the mp_ring for this queue");
3495 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_abdications",
3496 	    CTLFLAG_RD, &txq->r->abdications,
3497 	    "# of consumer abdications in the mp_ring for this queue");
3498 
3499 	return (0);
3500 }
3501 
3502 static int
3503 free_txq(struct vi_info *vi, struct sge_txq *txq)
3504 {
3505 	int rc;
3506 	struct adapter *sc = vi->pi->adapter;
3507 	struct sge_eq *eq = &txq->eq;
3508 
3509 	rc = free_eq(sc, eq);
3510 	if (rc)
3511 		return (rc);
3512 
3513 	sglist_free(txq->gl);
3514 	free(txq->sdesc, M_CXGBE);
3515 	mp_ring_free(txq->r);
3516 
3517 	bzero(txq, sizeof(*txq));
3518 	return (0);
3519 }
3520 
3521 static void
3522 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
3523 {
3524 	bus_addr_t *ba = arg;
3525 
3526 	KASSERT(nseg == 1,
3527 	    ("%s meant for single segment mappings only.", __func__));
3528 
3529 	*ba = error ? 0 : segs->ds_addr;
3530 }
3531 
3532 static inline void
3533 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
3534 {
3535 	uint32_t n, v;
3536 
3537 	n = IDXDIFF(fl->pidx / 8, fl->dbidx, fl->sidx);
3538 	MPASS(n > 0);
3539 
3540 	wmb();
3541 	v = fl->dbval | V_PIDX(n);
3542 	if (fl->udb)
3543 		*fl->udb = htole32(v);
3544 	else
3545 		t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL), v);
3546 	IDXINCR(fl->dbidx, n, fl->sidx);
3547 }
3548 
3549 /*
3550  * Fills up the freelist by allocating upto 'n' buffers.  Buffers that are
3551  * recycled do not count towards this allocation budget.
3552  *
3553  * Returns non-zero to indicate that this freelist should be added to the list
3554  * of starving freelists.
3555  */
3556 static int
3557 refill_fl(struct adapter *sc, struct sge_fl *fl, int n)
3558 {
3559 	__be64 *d;
3560 	struct fl_sdesc *sd;
3561 	uintptr_t pa;
3562 	caddr_t cl;
3563 	struct cluster_layout *cll;
3564 	struct sw_zone_info *swz;
3565 	struct cluster_metadata *clm;
3566 	uint16_t max_pidx;
3567 	uint16_t hw_cidx = fl->hw_cidx;		/* stable snapshot */
3568 
3569 	FL_LOCK_ASSERT_OWNED(fl);
3570 
3571 	/*
3572 	 * We always stop at the begining of the hardware descriptor that's just
3573 	 * before the one with the hw cidx.  This is to avoid hw pidx = hw cidx,
3574 	 * which would mean an empty freelist to the chip.
3575 	 */
3576 	max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1;
3577 	if (fl->pidx == max_pidx * 8)
3578 		return (0);
3579 
3580 	d = &fl->desc[fl->pidx];
3581 	sd = &fl->sdesc[fl->pidx];
3582 	cll = &fl->cll_def;	/* default layout */
3583 	swz = &sc->sge.sw_zone_info[cll->zidx];
3584 
3585 	while (n > 0) {
3586 
3587 		if (sd->cl != NULL) {
3588 
3589 			if (sd->nmbuf == 0) {
3590 				/*
3591 				 * Fast recycle without involving any atomics on
3592 				 * the cluster's metadata (if the cluster has
3593 				 * metadata).  This happens when all frames
3594 				 * received in the cluster were small enough to
3595 				 * fit within a single mbuf each.
3596 				 */
3597 				fl->cl_fast_recycled++;
3598 #ifdef INVARIANTS
3599 				clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
3600 				if (clm != NULL)
3601 					MPASS(clm->refcount == 1);
3602 #endif
3603 				goto recycled_fast;
3604 			}
3605 
3606 			/*
3607 			 * Cluster is guaranteed to have metadata.  Clusters
3608 			 * without metadata always take the fast recycle path
3609 			 * when they're recycled.
3610 			 */
3611 			clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
3612 			MPASS(clm != NULL);
3613 
3614 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
3615 				fl->cl_recycled++;
3616 				counter_u64_add(extfree_rels, 1);
3617 				goto recycled;
3618 			}
3619 			sd->cl = NULL;	/* gave up my reference */
3620 		}
3621 		MPASS(sd->cl == NULL);
3622 alloc:
3623 		cl = uma_zalloc(swz->zone, M_NOWAIT);
3624 		if (__predict_false(cl == NULL)) {
3625 			if (cll == &fl->cll_alt || fl->cll_alt.zidx == -1 ||
3626 			    fl->cll_def.zidx == fl->cll_alt.zidx)
3627 				break;
3628 
3629 			/* fall back to the safe zone */
3630 			cll = &fl->cll_alt;
3631 			swz = &sc->sge.sw_zone_info[cll->zidx];
3632 			goto alloc;
3633 		}
3634 		fl->cl_allocated++;
3635 		n--;
3636 
3637 		pa = pmap_kextract((vm_offset_t)cl);
3638 		pa += cll->region1;
3639 		sd->cl = cl;
3640 		sd->cll = *cll;
3641 		*d = htobe64(pa | cll->hwidx);
3642 		clm = cl_metadata(sc, fl, cll, cl);
3643 		if (clm != NULL) {
3644 recycled:
3645 #ifdef INVARIANTS
3646 			clm->sd = sd;
3647 #endif
3648 			clm->refcount = 1;
3649 		}
3650 		sd->nmbuf = 0;
3651 recycled_fast:
3652 		d++;
3653 		sd++;
3654 		if (__predict_false(++fl->pidx % 8 == 0)) {
3655 			uint16_t pidx = fl->pidx / 8;
3656 
3657 			if (__predict_false(pidx == fl->sidx)) {
3658 				fl->pidx = 0;
3659 				pidx = 0;
3660 				sd = fl->sdesc;
3661 				d = fl->desc;
3662 			}
3663 			if (pidx == max_pidx)
3664 				break;
3665 
3666 			if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4)
3667 				ring_fl_db(sc, fl);
3668 		}
3669 	}
3670 
3671 	if (fl->pidx / 8 != fl->dbidx)
3672 		ring_fl_db(sc, fl);
3673 
3674 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
3675 }
3676 
3677 /*
3678  * Attempt to refill all starving freelists.
3679  */
3680 static void
3681 refill_sfl(void *arg)
3682 {
3683 	struct adapter *sc = arg;
3684 	struct sge_fl *fl, *fl_temp;
3685 
3686 	mtx_assert(&sc->sfl_lock, MA_OWNED);
3687 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
3688 		FL_LOCK(fl);
3689 		refill_fl(sc, fl, 64);
3690 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
3691 			TAILQ_REMOVE(&sc->sfl, fl, link);
3692 			fl->flags &= ~FL_STARVING;
3693 		}
3694 		FL_UNLOCK(fl);
3695 	}
3696 
3697 	if (!TAILQ_EMPTY(&sc->sfl))
3698 		callout_schedule(&sc->sfl_callout, hz / 5);
3699 }
3700 
3701 static int
3702 alloc_fl_sdesc(struct sge_fl *fl)
3703 {
3704 
3705 	fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc), M_CXGBE,
3706 	    M_ZERO | M_WAITOK);
3707 
3708 	return (0);
3709 }
3710 
3711 static void
3712 free_fl_sdesc(struct adapter *sc, struct sge_fl *fl)
3713 {
3714 	struct fl_sdesc *sd;
3715 	struct cluster_metadata *clm;
3716 	struct cluster_layout *cll;
3717 	int i;
3718 
3719 	sd = fl->sdesc;
3720 	for (i = 0; i < fl->sidx * 8; i++, sd++) {
3721 		if (sd->cl == NULL)
3722 			continue;
3723 
3724 		cll = &sd->cll;
3725 		clm = cl_metadata(sc, fl, cll, sd->cl);
3726 		if (sd->nmbuf == 0)
3727 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
3728 		else if (clm && atomic_fetchadd_int(&clm->refcount, -1) == 1) {
3729 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
3730 			counter_u64_add(extfree_rels, 1);
3731 		}
3732 		sd->cl = NULL;
3733 	}
3734 
3735 	free(fl->sdesc, M_CXGBE);
3736 	fl->sdesc = NULL;
3737 }
3738 
3739 static inline void
3740 get_pkt_gl(struct mbuf *m, struct sglist *gl)
3741 {
3742 	int rc;
3743 
3744 	M_ASSERTPKTHDR(m);
3745 
3746 	sglist_reset(gl);
3747 	rc = sglist_append_mbuf(gl, m);
3748 	if (__predict_false(rc != 0)) {
3749 		panic("%s: mbuf %p (%d segs) was vetted earlier but now fails "
3750 		    "with %d.", __func__, m, mbuf_nsegs(m), rc);
3751 	}
3752 
3753 	KASSERT(gl->sg_nseg == mbuf_nsegs(m),
3754 	    ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m,
3755 	    mbuf_nsegs(m), gl->sg_nseg));
3756 	KASSERT(gl->sg_nseg > 0 &&
3757 	    gl->sg_nseg <= (needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS),
3758 	    ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__,
3759 		gl->sg_nseg, needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS));
3760 }
3761 
3762 /*
3763  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
3764  */
3765 static inline u_int
3766 txpkt_len16(u_int nsegs, u_int tso)
3767 {
3768 	u_int n;
3769 
3770 	MPASS(nsegs > 0);
3771 
3772 	nsegs--; /* first segment is part of ulptx_sgl */
3773 	n = sizeof(struct fw_eth_tx_pkt_wr) + sizeof(struct cpl_tx_pkt_core) +
3774 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
3775 	if (tso)
3776 		n += sizeof(struct cpl_tx_pkt_lso_core);
3777 
3778 	return (howmany(n, 16));
3779 }
3780 
3781 /*
3782  * len16 for a txpkts type 0 WR with a GL.  Does not include the firmware work
3783  * request header.
3784  */
3785 static inline u_int
3786 txpkts0_len16(u_int nsegs)
3787 {
3788 	u_int n;
3789 
3790 	MPASS(nsegs > 0);
3791 
3792 	nsegs--; /* first segment is part of ulptx_sgl */
3793 	n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) +
3794 	    sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) +
3795 	    8 * ((3 * nsegs) / 2 + (nsegs & 1));
3796 
3797 	return (howmany(n, 16));
3798 }
3799 
3800 /*
3801  * len16 for a txpkts type 1 WR with a GL.  Does not include the firmware work
3802  * request header.
3803  */
3804 static inline u_int
3805 txpkts1_len16(void)
3806 {
3807 	u_int n;
3808 
3809 	n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl);
3810 
3811 	return (howmany(n, 16));
3812 }
3813 
3814 static inline u_int
3815 imm_payload(u_int ndesc)
3816 {
3817 	u_int n;
3818 
3819 	n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) -
3820 	    sizeof(struct cpl_tx_pkt_core);
3821 
3822 	return (n);
3823 }
3824 
3825 /*
3826  * Write a txpkt WR for this packet to the hardware descriptors, update the
3827  * software descriptor, and advance the pidx.  It is guaranteed that enough
3828  * descriptors are available.
3829  *
3830  * The return value is the # of hardware descriptors used.
3831  */
3832 static u_int
3833 write_txpkt_wr(struct sge_txq *txq, struct fw_eth_tx_pkt_wr *wr,
3834     struct mbuf *m0, u_int available)
3835 {
3836 	struct sge_eq *eq = &txq->eq;
3837 	struct tx_sdesc *txsd;
3838 	struct cpl_tx_pkt_core *cpl;
3839 	uint32_t ctrl;	/* used in many unrelated places */
3840 	uint64_t ctrl1;
3841 	int len16, ndesc, pktlen, nsegs;
3842 	caddr_t dst;
3843 
3844 	TXQ_LOCK_ASSERT_OWNED(txq);
3845 	M_ASSERTPKTHDR(m0);
3846 	MPASS(available > 0 && available < eq->sidx);
3847 
3848 	len16 = mbuf_len16(m0);
3849 	nsegs = mbuf_nsegs(m0);
3850 	pktlen = m0->m_pkthdr.len;
3851 	ctrl = sizeof(struct cpl_tx_pkt_core);
3852 	if (needs_tso(m0))
3853 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
3854 	else if (pktlen <= imm_payload(2) && available >= 2) {
3855 		/* Immediate data.  Recalculate len16 and set nsegs to 0. */
3856 		ctrl += pktlen;
3857 		len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) +
3858 		    sizeof(struct cpl_tx_pkt_core) + pktlen, 16);
3859 		nsegs = 0;
3860 	}
3861 	ndesc = howmany(len16, EQ_ESIZE / 16);
3862 	MPASS(ndesc <= available);
3863 
3864 	/* Firmware work request header */
3865 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
3866 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
3867 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
3868 
3869 	ctrl = V_FW_WR_LEN16(len16);
3870 	wr->equiq_to_len16 = htobe32(ctrl);
3871 	wr->r3 = 0;
3872 
3873 	if (needs_tso(m0)) {
3874 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
3875 
3876 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
3877 		    m0->m_pkthdr.l4hlen > 0,
3878 		    ("%s: mbuf %p needs TSO but missing header lengths",
3879 			__func__, m0));
3880 
3881 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
3882 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
3883 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
3884 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
3885 			ctrl |= V_LSO_ETHHDR_LEN(1);
3886 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
3887 			ctrl |= F_LSO_IPV6;
3888 
3889 		lso->lso_ctrl = htobe32(ctrl);
3890 		lso->ipid_ofst = htobe16(0);
3891 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
3892 		lso->seqno_offset = htobe32(0);
3893 		lso->len = htobe32(pktlen);
3894 
3895 		cpl = (void *)(lso + 1);
3896 
3897 		txq->tso_wrs++;
3898 	} else
3899 		cpl = (void *)(wr + 1);
3900 
3901 	/* Checksum offload */
3902 	ctrl1 = 0;
3903 	if (needs_l3_csum(m0) == 0)
3904 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
3905 	if (needs_l4_csum(m0) == 0)
3906 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
3907 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
3908 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
3909 		txq->txcsum++;	/* some hardware assistance provided */
3910 
3911 	/* VLAN tag insertion */
3912 	if (needs_vlan_insertion(m0)) {
3913 		ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
3914 		txq->vlan_insertion++;
3915 	}
3916 
3917 	/* CPL header */
3918 	cpl->ctrl0 = txq->cpl_ctrl0;
3919 	cpl->pack = 0;
3920 	cpl->len = htobe16(pktlen);
3921 	cpl->ctrl1 = htobe64(ctrl1);
3922 
3923 	/* SGL */
3924 	dst = (void *)(cpl + 1);
3925 	if (nsegs > 0) {
3926 
3927 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
3928 		txq->sgl_wrs++;
3929 	} else {
3930 		struct mbuf *m;
3931 
3932 		for (m = m0; m != NULL; m = m->m_next) {
3933 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
3934 #ifdef INVARIANTS
3935 			pktlen -= m->m_len;
3936 #endif
3937 		}
3938 #ifdef INVARIANTS
3939 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
3940 #endif
3941 		txq->imm_wrs++;
3942 	}
3943 
3944 	txq->txpkt_wrs++;
3945 
3946 	txsd = &txq->sdesc[eq->pidx];
3947 	txsd->m = m0;
3948 	txsd->desc_used = ndesc;
3949 
3950 	return (ndesc);
3951 }
3952 
3953 static int
3954 try_txpkts(struct mbuf *m, struct mbuf *n, struct txpkts *txp, u_int available)
3955 {
3956 	u_int needed, nsegs1, nsegs2, l1, l2;
3957 
3958 	if (cannot_use_txpkts(m) || cannot_use_txpkts(n))
3959 		return (1);
3960 
3961 	nsegs1 = mbuf_nsegs(m);
3962 	nsegs2 = mbuf_nsegs(n);
3963 	if (nsegs1 + nsegs2 == 2) {
3964 		txp->wr_type = 1;
3965 		l1 = l2 = txpkts1_len16();
3966 	} else {
3967 		txp->wr_type = 0;
3968 		l1 = txpkts0_len16(nsegs1);
3969 		l2 = txpkts0_len16(nsegs2);
3970 	}
3971 	txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) + l1 + l2;
3972 	needed = howmany(txp->len16, EQ_ESIZE / 16);
3973 	if (needed > SGE_MAX_WR_NDESC || needed > available)
3974 		return (1);
3975 
3976 	txp->plen = m->m_pkthdr.len + n->m_pkthdr.len;
3977 	if (txp->plen > 65535)
3978 		return (1);
3979 
3980 	txp->npkt = 2;
3981 	set_mbuf_len16(m, l1);
3982 	set_mbuf_len16(n, l2);
3983 
3984 	return (0);
3985 }
3986 
3987 static int
3988 add_to_txpkts(struct mbuf *m, struct txpkts *txp, u_int available)
3989 {
3990 	u_int plen, len16, needed, nsegs;
3991 
3992 	MPASS(txp->wr_type == 0 || txp->wr_type == 1);
3993 
3994 	nsegs = mbuf_nsegs(m);
3995 	if (needs_tso(m) || (txp->wr_type == 1 && nsegs != 1))
3996 		return (1);
3997 
3998 	plen = txp->plen + m->m_pkthdr.len;
3999 	if (plen > 65535)
4000 		return (1);
4001 
4002 	if (txp->wr_type == 0)
4003 		len16 = txpkts0_len16(nsegs);
4004 	else
4005 		len16 = txpkts1_len16();
4006 	needed = howmany(txp->len16 + len16, EQ_ESIZE / 16);
4007 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4008 		return (1);
4009 
4010 	txp->npkt++;
4011 	txp->plen = plen;
4012 	txp->len16 += len16;
4013 	set_mbuf_len16(m, len16);
4014 
4015 	return (0);
4016 }
4017 
4018 /*
4019  * Write a txpkts WR for the packets in txp to the hardware descriptors, update
4020  * the software descriptor, and advance the pidx.  It is guaranteed that enough
4021  * descriptors are available.
4022  *
4023  * The return value is the # of hardware descriptors used.
4024  */
4025 static u_int
4026 write_txpkts_wr(struct sge_txq *txq, struct fw_eth_tx_pkts_wr *wr,
4027     struct mbuf *m0, const struct txpkts *txp, u_int available)
4028 {
4029 	struct sge_eq *eq = &txq->eq;
4030 	struct tx_sdesc *txsd;
4031 	struct cpl_tx_pkt_core *cpl;
4032 	uint32_t ctrl;
4033 	uint64_t ctrl1;
4034 	int ndesc, checkwrap;
4035 	struct mbuf *m;
4036 	void *flitp;
4037 
4038 	TXQ_LOCK_ASSERT_OWNED(txq);
4039 	MPASS(txp->npkt > 0);
4040 	MPASS(txp->plen < 65536);
4041 	MPASS(m0 != NULL);
4042 	MPASS(m0->m_nextpkt != NULL);
4043 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
4044 	MPASS(available > 0 && available < eq->sidx);
4045 
4046 	ndesc = howmany(txp->len16, EQ_ESIZE / 16);
4047 	MPASS(ndesc <= available);
4048 
4049 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4050 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
4051 	ctrl = V_FW_WR_LEN16(txp->len16);
4052 	wr->equiq_to_len16 = htobe32(ctrl);
4053 	wr->plen = htobe16(txp->plen);
4054 	wr->npkt = txp->npkt;
4055 	wr->r3 = 0;
4056 	wr->type = txp->wr_type;
4057 	flitp = wr + 1;
4058 
4059 	/*
4060 	 * At this point we are 16B into a hardware descriptor.  If checkwrap is
4061 	 * set then we know the WR is going to wrap around somewhere.  We'll
4062 	 * check for that at appropriate points.
4063 	 */
4064 	checkwrap = eq->sidx - ndesc < eq->pidx;
4065 	for (m = m0; m != NULL; m = m->m_nextpkt) {
4066 		if (txp->wr_type == 0) {
4067 			struct ulp_txpkt *ulpmc;
4068 			struct ulptx_idata *ulpsc;
4069 
4070 			/* ULP master command */
4071 			ulpmc = flitp;
4072 			ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
4073 			    V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid));
4074 			ulpmc->len = htobe32(mbuf_len16(m));
4075 
4076 			/* ULP subcommand */
4077 			ulpsc = (void *)(ulpmc + 1);
4078 			ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
4079 			    F_ULP_TX_SC_MORE);
4080 			ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
4081 
4082 			cpl = (void *)(ulpsc + 1);
4083 			if (checkwrap &&
4084 			    (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx])
4085 				cpl = (void *)&eq->desc[0];
4086 			txq->txpkts0_pkts += txp->npkt;
4087 			txq->txpkts0_wrs++;
4088 		} else {
4089 			cpl = flitp;
4090 			txq->txpkts1_pkts += txp->npkt;
4091 			txq->txpkts1_wrs++;
4092 		}
4093 
4094 		/* Checksum offload */
4095 		ctrl1 = 0;
4096 		if (needs_l3_csum(m) == 0)
4097 			ctrl1 |= F_TXPKT_IPCSUM_DIS;
4098 		if (needs_l4_csum(m) == 0)
4099 			ctrl1 |= F_TXPKT_L4CSUM_DIS;
4100 		if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4101 		    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4102 			txq->txcsum++;	/* some hardware assistance provided */
4103 
4104 		/* VLAN tag insertion */
4105 		if (needs_vlan_insertion(m)) {
4106 			ctrl1 |= F_TXPKT_VLAN_VLD |
4107 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
4108 			txq->vlan_insertion++;
4109 		}
4110 
4111 		/* CPL header */
4112 		cpl->ctrl0 = txq->cpl_ctrl0;
4113 		cpl->pack = 0;
4114 		cpl->len = htobe16(m->m_pkthdr.len);
4115 		cpl->ctrl1 = htobe64(ctrl1);
4116 
4117 		flitp = cpl + 1;
4118 		if (checkwrap &&
4119 		    (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
4120 			flitp = (void *)&eq->desc[0];
4121 
4122 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap);
4123 
4124 	}
4125 
4126 	txsd = &txq->sdesc[eq->pidx];
4127 	txsd->m = m0;
4128 	txsd->desc_used = ndesc;
4129 
4130 	return (ndesc);
4131 }
4132 
4133 /*
4134  * If the SGL ends on an address that is not 16 byte aligned, this function will
4135  * add a 0 filled flit at the end.
4136  */
4137 static void
4138 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap)
4139 {
4140 	struct sge_eq *eq = &txq->eq;
4141 	struct sglist *gl = txq->gl;
4142 	struct sglist_seg *seg;
4143 	__be64 *flitp, *wrap;
4144 	struct ulptx_sgl *usgl;
4145 	int i, nflits, nsegs;
4146 
4147 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
4148 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
4149 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
4150 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
4151 
4152 	get_pkt_gl(m, gl);
4153 	nsegs = gl->sg_nseg;
4154 	MPASS(nsegs > 0);
4155 
4156 	nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
4157 	flitp = (__be64 *)(*to);
4158 	wrap = (__be64 *)(&eq->desc[eq->sidx]);
4159 	seg = &gl->sg_segs[0];
4160 	usgl = (void *)flitp;
4161 
4162 	/*
4163 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
4164 	 * ring, so we're at least 16 bytes away from the status page.  There is
4165 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
4166 	 */
4167 
4168 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
4169 	    V_ULPTX_NSGE(nsegs));
4170 	usgl->len0 = htobe32(seg->ss_len);
4171 	usgl->addr0 = htobe64(seg->ss_paddr);
4172 	seg++;
4173 
4174 	if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) {
4175 
4176 		/* Won't wrap around at all */
4177 
4178 		for (i = 0; i < nsegs - 1; i++, seg++) {
4179 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
4180 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
4181 		}
4182 		if (i & 1)
4183 			usgl->sge[i / 2].len[1] = htobe32(0);
4184 		flitp += nflits;
4185 	} else {
4186 
4187 		/* Will wrap somewhere in the rest of the SGL */
4188 
4189 		/* 2 flits already written, write the rest flit by flit */
4190 		flitp = (void *)(usgl + 1);
4191 		for (i = 0; i < nflits - 2; i++) {
4192 			if (flitp == wrap)
4193 				flitp = (void *)eq->desc;
4194 			*flitp++ = get_flit(seg, nsegs - 1, i);
4195 		}
4196 	}
4197 
4198 	if (nflits & 1) {
4199 		MPASS(((uintptr_t)flitp) & 0xf);
4200 		*flitp++ = 0;
4201 	}
4202 
4203 	MPASS((((uintptr_t)flitp) & 0xf) == 0);
4204 	if (__predict_false(flitp == wrap))
4205 		*to = (void *)eq->desc;
4206 	else
4207 		*to = (void *)flitp;
4208 }
4209 
4210 static inline void
4211 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
4212 {
4213 
4214 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
4215 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
4216 
4217 	if (__predict_true((uintptr_t)(*to) + len <=
4218 	    (uintptr_t)&eq->desc[eq->sidx])) {
4219 		bcopy(from, *to, len);
4220 		(*to) += len;
4221 	} else {
4222 		int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
4223 
4224 		bcopy(from, *to, portion);
4225 		from += portion;
4226 		portion = len - portion;	/* remaining */
4227 		bcopy(from, (void *)eq->desc, portion);
4228 		(*to) = (caddr_t)eq->desc + portion;
4229 	}
4230 }
4231 
4232 static inline void
4233 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n)
4234 {
4235 	u_int db;
4236 
4237 	MPASS(n > 0);
4238 
4239 	db = eq->doorbells;
4240 	if (n > 1)
4241 		clrbit(&db, DOORBELL_WCWR);
4242 	wmb();
4243 
4244 	switch (ffs(db) - 1) {
4245 	case DOORBELL_UDB:
4246 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
4247 		break;
4248 
4249 	case DOORBELL_WCWR: {
4250 		volatile uint64_t *dst, *src;
4251 		int i;
4252 
4253 		/*
4254 		 * Queues whose 128B doorbell segment fits in the page do not
4255 		 * use relative qid (udb_qid is always 0).  Only queues with
4256 		 * doorbell segments can do WCWR.
4257 		 */
4258 		KASSERT(eq->udb_qid == 0 && n == 1,
4259 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
4260 		    __func__, eq->doorbells, n, eq->dbidx, eq));
4261 
4262 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
4263 		    UDBS_DB_OFFSET);
4264 		i = eq->dbidx;
4265 		src = (void *)&eq->desc[i];
4266 		while (src != (void *)&eq->desc[i + 1])
4267 			*dst++ = *src++;
4268 		wmb();
4269 		break;
4270 	}
4271 
4272 	case DOORBELL_UDBWC:
4273 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
4274 		wmb();
4275 		break;
4276 
4277 	case DOORBELL_KDB:
4278 		t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
4279 		    V_QID(eq->cntxt_id) | V_PIDX(n));
4280 		break;
4281 	}
4282 
4283 	IDXINCR(eq->dbidx, n, eq->sidx);
4284 }
4285 
4286 static inline u_int
4287 reclaimable_tx_desc(struct sge_eq *eq)
4288 {
4289 	uint16_t hw_cidx;
4290 
4291 	hw_cidx = read_hw_cidx(eq);
4292 	return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx));
4293 }
4294 
4295 static inline u_int
4296 total_available_tx_desc(struct sge_eq *eq)
4297 {
4298 	uint16_t hw_cidx, pidx;
4299 
4300 	hw_cidx = read_hw_cidx(eq);
4301 	pidx = eq->pidx;
4302 
4303 	if (pidx == hw_cidx)
4304 		return (eq->sidx - 1);
4305 	else
4306 		return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1);
4307 }
4308 
4309 static inline uint16_t
4310 read_hw_cidx(struct sge_eq *eq)
4311 {
4312 	struct sge_qstat *spg = (void *)&eq->desc[eq->sidx];
4313 	uint16_t cidx = spg->cidx;	/* stable snapshot */
4314 
4315 	return (be16toh(cidx));
4316 }
4317 
4318 /*
4319  * Reclaim 'n' descriptors approximately.
4320  */
4321 static u_int
4322 reclaim_tx_descs(struct sge_txq *txq, u_int n)
4323 {
4324 	struct tx_sdesc *txsd;
4325 	struct sge_eq *eq = &txq->eq;
4326 	u_int can_reclaim, reclaimed;
4327 
4328 	TXQ_LOCK_ASSERT_OWNED(txq);
4329 	MPASS(n > 0);
4330 
4331 	reclaimed = 0;
4332 	can_reclaim = reclaimable_tx_desc(eq);
4333 	while (can_reclaim && reclaimed < n) {
4334 		int ndesc;
4335 		struct mbuf *m, *nextpkt;
4336 
4337 		txsd = &txq->sdesc[eq->cidx];
4338 		ndesc = txsd->desc_used;
4339 
4340 		/* Firmware doesn't return "partial" credits. */
4341 		KASSERT(can_reclaim >= ndesc,
4342 		    ("%s: unexpected number of credits: %d, %d",
4343 		    __func__, can_reclaim, ndesc));
4344 
4345 		for (m = txsd->m; m != NULL; m = nextpkt) {
4346 			nextpkt = m->m_nextpkt;
4347 			m->m_nextpkt = NULL;
4348 			m_freem(m);
4349 		}
4350 		reclaimed += ndesc;
4351 		can_reclaim -= ndesc;
4352 		IDXINCR(eq->cidx, ndesc, eq->sidx);
4353 	}
4354 
4355 	return (reclaimed);
4356 }
4357 
4358 static void
4359 tx_reclaim(void *arg, int n)
4360 {
4361 	struct sge_txq *txq = arg;
4362 	struct sge_eq *eq = &txq->eq;
4363 
4364 	do {
4365 		if (TXQ_TRYLOCK(txq) == 0)
4366 			break;
4367 		n = reclaim_tx_descs(txq, 32);
4368 		if (eq->cidx == eq->pidx)
4369 			eq->equeqidx = eq->pidx;
4370 		TXQ_UNLOCK(txq);
4371 	} while (n > 0);
4372 }
4373 
4374 static __be64
4375 get_flit(struct sglist_seg *segs, int nsegs, int idx)
4376 {
4377 	int i = (idx / 3) * 2;
4378 
4379 	switch (idx % 3) {
4380 	case 0: {
4381 		__be64 rc;
4382 
4383 		rc = htobe32(segs[i].ss_len);
4384 		if (i + 1 < nsegs)
4385 			rc |= (uint64_t)htobe32(segs[i + 1].ss_len) << 32;
4386 
4387 		return (rc);
4388 	}
4389 	case 1:
4390 		return (htobe64(segs[i].ss_paddr));
4391 	case 2:
4392 		return (htobe64(segs[i + 1].ss_paddr));
4393 	}
4394 
4395 	return (0);
4396 }
4397 
4398 static void
4399 find_best_refill_source(struct adapter *sc, struct sge_fl *fl, int maxp)
4400 {
4401 	int8_t zidx, hwidx, idx;
4402 	uint16_t region1, region3;
4403 	int spare, spare_needed, n;
4404 	struct sw_zone_info *swz;
4405 	struct hw_buf_info *hwb, *hwb_list = &sc->sge.hw_buf_info[0];
4406 
4407 	/*
4408 	 * Buffer Packing: Look for PAGE_SIZE or larger zone which has a bufsize
4409 	 * large enough for the max payload and cluster metadata.  Otherwise
4410 	 * settle for the largest bufsize that leaves enough room in the cluster
4411 	 * for metadata.
4412 	 *
4413 	 * Without buffer packing: Look for the smallest zone which has a
4414 	 * bufsize large enough for the max payload.  Settle for the largest
4415 	 * bufsize available if there's nothing big enough for max payload.
4416 	 */
4417 	spare_needed = fl->flags & FL_BUF_PACKING ? CL_METADATA_SIZE : 0;
4418 	swz = &sc->sge.sw_zone_info[0];
4419 	hwidx = -1;
4420 	for (zidx = 0; zidx < SW_ZONE_SIZES; zidx++, swz++) {
4421 		if (swz->size > largest_rx_cluster) {
4422 			if (__predict_true(hwidx != -1))
4423 				break;
4424 
4425 			/*
4426 			 * This is a misconfiguration.  largest_rx_cluster is
4427 			 * preventing us from finding a refill source.  See
4428 			 * dev.t5nex.<n>.buffer_sizes to figure out why.
4429 			 */
4430 			device_printf(sc->dev, "largest_rx_cluster=%u leaves no"
4431 			    " refill source for fl %p (dma %u).  Ignored.\n",
4432 			    largest_rx_cluster, fl, maxp);
4433 		}
4434 		for (idx = swz->head_hwidx; idx != -1; idx = hwb->next) {
4435 			hwb = &hwb_list[idx];
4436 			spare = swz->size - hwb->size;
4437 			if (spare < spare_needed)
4438 				continue;
4439 
4440 			hwidx = idx;		/* best option so far */
4441 			if (hwb->size >= maxp) {
4442 
4443 				if ((fl->flags & FL_BUF_PACKING) == 0)
4444 					goto done; /* stop looking (not packing) */
4445 
4446 				if (swz->size >= safest_rx_cluster)
4447 					goto done; /* stop looking (packing) */
4448 			}
4449 			break;		/* keep looking, next zone */
4450 		}
4451 	}
4452 done:
4453 	/* A usable hwidx has been located. */
4454 	MPASS(hwidx != -1);
4455 	hwb = &hwb_list[hwidx];
4456 	zidx = hwb->zidx;
4457 	swz = &sc->sge.sw_zone_info[zidx];
4458 	region1 = 0;
4459 	region3 = swz->size - hwb->size;
4460 
4461 	/*
4462 	 * Stay within this zone and see if there is a better match when mbuf
4463 	 * inlining is allowed.  Remember that the hwidx's are sorted in
4464 	 * decreasing order of size (so in increasing order of spare area).
4465 	 */
4466 	for (idx = hwidx; idx != -1; idx = hwb->next) {
4467 		hwb = &hwb_list[idx];
4468 		spare = swz->size - hwb->size;
4469 
4470 		if (allow_mbufs_in_cluster == 0 || hwb->size < maxp)
4471 			break;
4472 
4473 		/*
4474 		 * Do not inline mbufs if doing so would violate the pad/pack
4475 		 * boundary alignment requirement.
4476 		 */
4477 		if (fl_pad && (MSIZE % sc->params.sge.pad_boundary) != 0)
4478 			continue;
4479 		if (fl->flags & FL_BUF_PACKING &&
4480 		    (MSIZE % sc->params.sge.pack_boundary) != 0)
4481 			continue;
4482 
4483 		if (spare < CL_METADATA_SIZE + MSIZE)
4484 			continue;
4485 		n = (spare - CL_METADATA_SIZE) / MSIZE;
4486 		if (n > howmany(hwb->size, maxp))
4487 			break;
4488 
4489 		hwidx = idx;
4490 		if (fl->flags & FL_BUF_PACKING) {
4491 			region1 = n * MSIZE;
4492 			region3 = spare - region1;
4493 		} else {
4494 			region1 = MSIZE;
4495 			region3 = spare - region1;
4496 			break;
4497 		}
4498 	}
4499 
4500 	KASSERT(zidx >= 0 && zidx < SW_ZONE_SIZES,
4501 	    ("%s: bad zone %d for fl %p, maxp %d", __func__, zidx, fl, maxp));
4502 	KASSERT(hwidx >= 0 && hwidx <= SGE_FLBUF_SIZES,
4503 	    ("%s: bad hwidx %d for fl %p, maxp %d", __func__, hwidx, fl, maxp));
4504 	KASSERT(region1 + sc->sge.hw_buf_info[hwidx].size + region3 ==
4505 	    sc->sge.sw_zone_info[zidx].size,
4506 	    ("%s: bad buffer layout for fl %p, maxp %d. "
4507 		"cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
4508 		sc->sge.sw_zone_info[zidx].size, region1,
4509 		sc->sge.hw_buf_info[hwidx].size, region3));
4510 	if (fl->flags & FL_BUF_PACKING || region1 > 0) {
4511 		KASSERT(region3 >= CL_METADATA_SIZE,
4512 		    ("%s: no room for metadata.  fl %p, maxp %d; "
4513 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
4514 		    sc->sge.sw_zone_info[zidx].size, region1,
4515 		    sc->sge.hw_buf_info[hwidx].size, region3));
4516 		KASSERT(region1 % MSIZE == 0,
4517 		    ("%s: bad mbuf region for fl %p, maxp %d. "
4518 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
4519 		    sc->sge.sw_zone_info[zidx].size, region1,
4520 		    sc->sge.hw_buf_info[hwidx].size, region3));
4521 	}
4522 
4523 	fl->cll_def.zidx = zidx;
4524 	fl->cll_def.hwidx = hwidx;
4525 	fl->cll_def.region1 = region1;
4526 	fl->cll_def.region3 = region3;
4527 }
4528 
4529 static void
4530 find_safe_refill_source(struct adapter *sc, struct sge_fl *fl)
4531 {
4532 	struct sge *s = &sc->sge;
4533 	struct hw_buf_info *hwb;
4534 	struct sw_zone_info *swz;
4535 	int spare;
4536 	int8_t hwidx;
4537 
4538 	if (fl->flags & FL_BUF_PACKING)
4539 		hwidx = s->safe_hwidx2;	/* with room for metadata */
4540 	else if (allow_mbufs_in_cluster && s->safe_hwidx2 != -1) {
4541 		hwidx = s->safe_hwidx2;
4542 		hwb = &s->hw_buf_info[hwidx];
4543 		swz = &s->sw_zone_info[hwb->zidx];
4544 		spare = swz->size - hwb->size;
4545 
4546 		/* no good if there isn't room for an mbuf as well */
4547 		if (spare < CL_METADATA_SIZE + MSIZE)
4548 			hwidx = s->safe_hwidx1;
4549 	} else
4550 		hwidx = s->safe_hwidx1;
4551 
4552 	if (hwidx == -1) {
4553 		/* No fallback source */
4554 		fl->cll_alt.hwidx = -1;
4555 		fl->cll_alt.zidx = -1;
4556 
4557 		return;
4558 	}
4559 
4560 	hwb = &s->hw_buf_info[hwidx];
4561 	swz = &s->sw_zone_info[hwb->zidx];
4562 	spare = swz->size - hwb->size;
4563 	fl->cll_alt.hwidx = hwidx;
4564 	fl->cll_alt.zidx = hwb->zidx;
4565 	if (allow_mbufs_in_cluster &&
4566 	    (fl_pad == 0 || (MSIZE % sc->params.sge.pad_boundary) == 0))
4567 		fl->cll_alt.region1 = ((spare - CL_METADATA_SIZE) / MSIZE) * MSIZE;
4568 	else
4569 		fl->cll_alt.region1 = 0;
4570 	fl->cll_alt.region3 = spare - fl->cll_alt.region1;
4571 }
4572 
4573 static void
4574 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
4575 {
4576 	mtx_lock(&sc->sfl_lock);
4577 	FL_LOCK(fl);
4578 	if ((fl->flags & FL_DOOMED) == 0) {
4579 		fl->flags |= FL_STARVING;
4580 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
4581 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
4582 	}
4583 	FL_UNLOCK(fl);
4584 	mtx_unlock(&sc->sfl_lock);
4585 }
4586 
4587 static void
4588 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq)
4589 {
4590 	struct sge_wrq *wrq = (void *)eq;
4591 
4592 	atomic_readandclear_int(&eq->equiq);
4593 	taskqueue_enqueue(sc->tq[eq->tx_chan], &wrq->wrq_tx_task);
4594 }
4595 
4596 static void
4597 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq)
4598 {
4599 	struct sge_txq *txq = (void *)eq;
4600 
4601 	MPASS((eq->flags & EQ_TYPEMASK) == EQ_ETH);
4602 
4603 	atomic_readandclear_int(&eq->equiq);
4604 	mp_ring_check_drainage(txq->r, 0);
4605 	taskqueue_enqueue(sc->tq[eq->tx_chan], &txq->tx_reclaim_task);
4606 }
4607 
4608 static int
4609 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
4610     struct mbuf *m)
4611 {
4612 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
4613 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
4614 	struct adapter *sc = iq->adapter;
4615 	struct sge *s = &sc->sge;
4616 	struct sge_eq *eq;
4617 	static void (*h[])(struct adapter *, struct sge_eq *) = {NULL,
4618 		&handle_wrq_egr_update, &handle_eth_egr_update,
4619 		&handle_wrq_egr_update};
4620 
4621 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
4622 	    rss->opcode));
4623 
4624 	eq = s->eqmap[qid - s->eq_start];
4625 	(*h[eq->flags & EQ_TYPEMASK])(sc, eq);
4626 
4627 	return (0);
4628 }
4629 
4630 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
4631 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
4632     offsetof(struct cpl_fw6_msg, data));
4633 
4634 static int
4635 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
4636 {
4637 	struct adapter *sc = iq->adapter;
4638 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
4639 
4640 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
4641 	    rss->opcode));
4642 
4643 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
4644 		const struct rss_header *rss2;
4645 
4646 		rss2 = (const struct rss_header *)&cpl->data[0];
4647 		return (sc->cpl_handler[rss2->opcode](iq, rss2, m));
4648 	}
4649 
4650 	return (sc->fw_msg_handler[cpl->type](sc, &cpl->data[0]));
4651 }
4652 
4653 static int
4654 sysctl_uint16(SYSCTL_HANDLER_ARGS)
4655 {
4656 	uint16_t *id = arg1;
4657 	int i = *id;
4658 
4659 	return sysctl_handle_int(oidp, &i, 0, req);
4660 }
4661 
4662 static int
4663 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
4664 {
4665 	struct sge *s = arg1;
4666 	struct hw_buf_info *hwb = &s->hw_buf_info[0];
4667 	struct sw_zone_info *swz = &s->sw_zone_info[0];
4668 	int i, rc;
4669 	struct sbuf sb;
4670 	char c;
4671 
4672 	sbuf_new(&sb, NULL, 32, SBUF_AUTOEXTEND);
4673 	for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
4674 		if (hwb->zidx >= 0 && swz[hwb->zidx].size <= largest_rx_cluster)
4675 			c = '*';
4676 		else
4677 			c = '\0';
4678 
4679 		sbuf_printf(&sb, "%u%c ", hwb->size, c);
4680 	}
4681 	sbuf_trim(&sb);
4682 	sbuf_finish(&sb);
4683 	rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
4684 	sbuf_delete(&sb);
4685 	return (rc);
4686 }
4687