xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision fa3385055c6f82433ec4d571ae17fd8211584055)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2011 Chelsio Communications, Inc.
5  * All rights reserved.
6  * Written by: Navdeep Parhar <np@FreeBSD.org>
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include "opt_inet.h"
34 #include "opt_inet6.h"
35 
36 #include <sys/types.h>
37 #include <sys/eventhandler.h>
38 #include <sys/mbuf.h>
39 #include <sys/socket.h>
40 #include <sys/kernel.h>
41 #include <sys/malloc.h>
42 #include <sys/queue.h>
43 #include <sys/sbuf.h>
44 #include <sys/taskqueue.h>
45 #include <sys/time.h>
46 #include <sys/sglist.h>
47 #include <sys/sysctl.h>
48 #include <sys/smp.h>
49 #include <sys/counter.h>
50 #include <net/bpf.h>
51 #include <net/ethernet.h>
52 #include <net/if.h>
53 #include <net/if_vlan_var.h>
54 #include <netinet/in.h>
55 #include <netinet/ip.h>
56 #include <netinet/ip6.h>
57 #include <netinet/tcp.h>
58 #include <machine/in_cksum.h>
59 #include <machine/md_var.h>
60 #include <vm/vm.h>
61 #include <vm/pmap.h>
62 #ifdef DEV_NETMAP
63 #include <machine/bus.h>
64 #include <sys/selinfo.h>
65 #include <net/if_var.h>
66 #include <net/netmap.h>
67 #include <dev/netmap/netmap_kern.h>
68 #endif
69 
70 #include "common/common.h"
71 #include "common/t4_regs.h"
72 #include "common/t4_regs_values.h"
73 #include "common/t4_msg.h"
74 #include "t4_l2t.h"
75 #include "t4_mp_ring.h"
76 
77 #ifdef T4_PKT_TIMESTAMP
78 #define RX_COPY_THRESHOLD (MINCLSIZE - 8)
79 #else
80 #define RX_COPY_THRESHOLD MINCLSIZE
81 #endif
82 
83 /*
84  * Ethernet frames are DMA'd at this byte offset into the freelist buffer.
85  * 0-7 are valid values.
86  */
87 static int fl_pktshift = 2;
88 TUNABLE_INT("hw.cxgbe.fl_pktshift", &fl_pktshift);
89 
90 /*
91  * Pad ethernet payload up to this boundary.
92  * -1: driver should figure out a good value.
93  *  0: disable padding.
94  *  Any power of 2 from 32 to 4096 (both inclusive) is also a valid value.
95  */
96 int fl_pad = -1;
97 TUNABLE_INT("hw.cxgbe.fl_pad", &fl_pad);
98 
99 /*
100  * Status page length.
101  * -1: driver should figure out a good value.
102  *  64 or 128 are the only other valid values.
103  */
104 static int spg_len = -1;
105 TUNABLE_INT("hw.cxgbe.spg_len", &spg_len);
106 
107 /*
108  * Congestion drops.
109  * -1: no congestion feedback (not recommended).
110  *  0: backpressure the channel instead of dropping packets right away.
111  *  1: no backpressure, drop packets for the congested queue immediately.
112  */
113 static int cong_drop = 0;
114 TUNABLE_INT("hw.cxgbe.cong_drop", &cong_drop);
115 
116 /*
117  * Deliver multiple frames in the same free list buffer if they fit.
118  * -1: let the driver decide whether to enable buffer packing or not.
119  *  0: disable buffer packing.
120  *  1: enable buffer packing.
121  */
122 static int buffer_packing = -1;
123 TUNABLE_INT("hw.cxgbe.buffer_packing", &buffer_packing);
124 
125 /*
126  * Start next frame in a packed buffer at this boundary.
127  * -1: driver should figure out a good value.
128  * T4: driver will ignore this and use the same value as fl_pad above.
129  * T5: 16, or a power of 2 from 64 to 4096 (both inclusive) is a valid value.
130  */
131 static int fl_pack = -1;
132 TUNABLE_INT("hw.cxgbe.fl_pack", &fl_pack);
133 
134 /*
135  * Allow the driver to create mbuf(s) in a cluster allocated for rx.
136  * 0: never; always allocate mbufs from the zone_mbuf UMA zone.
137  * 1: ok to create mbuf(s) within a cluster if there is room.
138  */
139 static int allow_mbufs_in_cluster = 1;
140 TUNABLE_INT("hw.cxgbe.allow_mbufs_in_cluster", &allow_mbufs_in_cluster);
141 
142 /*
143  * Largest rx cluster size that the driver is allowed to allocate.
144  */
145 static int largest_rx_cluster = MJUM16BYTES;
146 TUNABLE_INT("hw.cxgbe.largest_rx_cluster", &largest_rx_cluster);
147 
148 /*
149  * Size of cluster allocation that's most likely to succeed.  The driver will
150  * fall back to this size if it fails to allocate clusters larger than this.
151  */
152 static int safest_rx_cluster = PAGE_SIZE;
153 TUNABLE_INT("hw.cxgbe.safest_rx_cluster", &safest_rx_cluster);
154 
155 /*
156  * The interrupt holdoff timers are multiplied by this value on T6+.
157  * 1 and 3-17 (both inclusive) are legal values.
158  */
159 static int tscale = 1;
160 TUNABLE_INT("hw.cxgbe.tscale", &tscale);
161 
162 /*
163  * Number of LRO entries in the lro_ctrl structure per rx queue.
164  */
165 static int lro_entries = TCP_LRO_ENTRIES;
166 TUNABLE_INT("hw.cxgbe.lro_entries", &lro_entries);
167 
168 /*
169  * This enables presorting of frames before they're fed into tcp_lro_rx.
170  */
171 static int lro_mbufs = 0;
172 TUNABLE_INT("hw.cxgbe.lro_mbufs", &lro_mbufs);
173 
174 struct txpkts {
175 	u_int wr_type;		/* type 0 or type 1 */
176 	u_int npkt;		/* # of packets in this work request */
177 	u_int plen;		/* total payload (sum of all packets) */
178 	u_int len16;		/* # of 16B pieces used by this work request */
179 };
180 
181 /* A packet's SGL.  This + m_pkthdr has all info needed for tx */
182 struct sgl {
183 	struct sglist sg;
184 	struct sglist_seg seg[TX_SGL_SEGS];
185 };
186 
187 static int service_iq(struct sge_iq *, int);
188 static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t);
189 static int t4_eth_rx(struct sge_iq *, const struct rss_header *, struct mbuf *);
190 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int);
191 static inline void init_fl(struct adapter *, struct sge_fl *, int, int, char *);
192 static inline void init_eq(struct adapter *, struct sge_eq *, int, int, uint8_t,
193     uint16_t, char *);
194 static int alloc_ring(struct adapter *, size_t, bus_dma_tag_t *, bus_dmamap_t *,
195     bus_addr_t *, void **);
196 static int free_ring(struct adapter *, bus_dma_tag_t, bus_dmamap_t, bus_addr_t,
197     void *);
198 static int alloc_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *,
199     int, int);
200 static int free_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *);
201 static void add_iq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
202     struct sge_iq *);
203 static void add_fl_sysctls(struct adapter *, struct sysctl_ctx_list *,
204     struct sysctl_oid *, struct sge_fl *);
205 static int alloc_fwq(struct adapter *);
206 static int free_fwq(struct adapter *);
207 static int alloc_mgmtq(struct adapter *);
208 static int free_mgmtq(struct adapter *);
209 static int alloc_rxq(struct vi_info *, struct sge_rxq *, int, int,
210     struct sysctl_oid *);
211 static int free_rxq(struct vi_info *, struct sge_rxq *);
212 #ifdef TCP_OFFLOAD
213 static int alloc_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *, int, int,
214     struct sysctl_oid *);
215 static int free_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *);
216 #endif
217 #ifdef DEV_NETMAP
218 static int alloc_nm_rxq(struct vi_info *, struct sge_nm_rxq *, int, int,
219     struct sysctl_oid *);
220 static int free_nm_rxq(struct vi_info *, struct sge_nm_rxq *);
221 static int alloc_nm_txq(struct vi_info *, struct sge_nm_txq *, int, int,
222     struct sysctl_oid *);
223 static int free_nm_txq(struct vi_info *, struct sge_nm_txq *);
224 #endif
225 static int ctrl_eq_alloc(struct adapter *, struct sge_eq *);
226 static int eth_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
227 #ifdef TCP_OFFLOAD
228 static int ofld_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
229 #endif
230 static int alloc_eq(struct adapter *, struct vi_info *, struct sge_eq *);
231 static int free_eq(struct adapter *, struct sge_eq *);
232 static int alloc_wrq(struct adapter *, struct vi_info *, struct sge_wrq *,
233     struct sysctl_oid *);
234 static int free_wrq(struct adapter *, struct sge_wrq *);
235 static int alloc_txq(struct vi_info *, struct sge_txq *, int,
236     struct sysctl_oid *);
237 static int free_txq(struct vi_info *, struct sge_txq *);
238 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int);
239 static inline void ring_fl_db(struct adapter *, struct sge_fl *);
240 static int refill_fl(struct adapter *, struct sge_fl *, int);
241 static void refill_sfl(void *);
242 static int alloc_fl_sdesc(struct sge_fl *);
243 static void free_fl_sdesc(struct adapter *, struct sge_fl *);
244 static void find_best_refill_source(struct adapter *, struct sge_fl *, int);
245 static void find_safe_refill_source(struct adapter *, struct sge_fl *);
246 static void add_fl_to_sfl(struct adapter *, struct sge_fl *);
247 
248 static inline void get_pkt_gl(struct mbuf *, struct sglist *);
249 static inline u_int txpkt_len16(u_int, u_int);
250 static inline u_int txpkt_vm_len16(u_int, u_int);
251 static inline u_int txpkts0_len16(u_int);
252 static inline u_int txpkts1_len16(void);
253 static u_int write_txpkt_wr(struct sge_txq *, struct fw_eth_tx_pkt_wr *,
254     struct mbuf *, u_int);
255 static u_int write_txpkt_vm_wr(struct adapter *, struct sge_txq *,
256     struct fw_eth_tx_pkt_vm_wr *, struct mbuf *, u_int);
257 static int try_txpkts(struct mbuf *, struct mbuf *, struct txpkts *, u_int);
258 static int add_to_txpkts(struct mbuf *, struct txpkts *, u_int);
259 static u_int write_txpkts_wr(struct sge_txq *, struct fw_eth_tx_pkts_wr *,
260     struct mbuf *, const struct txpkts *, u_int);
261 static void write_gl_to_txd(struct sge_txq *, struct mbuf *, caddr_t *, int);
262 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int);
263 static inline void ring_eq_db(struct adapter *, struct sge_eq *, u_int);
264 static inline uint16_t read_hw_cidx(struct sge_eq *);
265 static inline u_int reclaimable_tx_desc(struct sge_eq *);
266 static inline u_int total_available_tx_desc(struct sge_eq *);
267 static u_int reclaim_tx_descs(struct sge_txq *, u_int);
268 static void tx_reclaim(void *, int);
269 static __be64 get_flit(struct sglist_seg *, int, int);
270 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
271     struct mbuf *);
272 static int handle_fw_msg(struct sge_iq *, const struct rss_header *,
273     struct mbuf *);
274 static int t4_handle_wrerr_rpl(struct adapter *, const __be64 *);
275 static void wrq_tx_drain(void *, int);
276 static void drain_wrq_wr_list(struct adapter *, struct sge_wrq *);
277 
278 static int sysctl_uint16(SYSCTL_HANDLER_ARGS);
279 static int sysctl_bufsizes(SYSCTL_HANDLER_ARGS);
280 static int sysctl_tc(SYSCTL_HANDLER_ARGS);
281 
282 static counter_u64_t extfree_refs;
283 static counter_u64_t extfree_rels;
284 
285 an_handler_t t4_an_handler;
286 fw_msg_handler_t t4_fw_msg_handler[NUM_FW6_TYPES];
287 cpl_handler_t t4_cpl_handler[NUM_CPL_CMDS];
288 cpl_handler_t set_tcb_rpl_handlers[NUM_CPL_COOKIES];
289 cpl_handler_t l2t_write_rpl_handlers[NUM_CPL_COOKIES];
290 cpl_handler_t act_open_rpl_handlers[NUM_CPL_COOKIES];
291 
292 void
293 t4_register_an_handler(an_handler_t h)
294 {
295 	uintptr_t *loc;
296 
297 	MPASS(h == NULL || t4_an_handler == NULL);
298 
299 	loc = (uintptr_t *)&t4_an_handler;
300 	atomic_store_rel_ptr(loc, (uintptr_t)h);
301 }
302 
303 void
304 t4_register_fw_msg_handler(int type, fw_msg_handler_t h)
305 {
306 	uintptr_t *loc;
307 
308 	MPASS(type < nitems(t4_fw_msg_handler));
309 	MPASS(h == NULL || t4_fw_msg_handler[type] == NULL);
310 	/*
311 	 * These are dispatched by the handler for FW{4|6}_CPL_MSG using the CPL
312 	 * handler dispatch table.  Reject any attempt to install a handler for
313 	 * this subtype.
314 	 */
315 	MPASS(type != FW_TYPE_RSSCPL);
316 	MPASS(type != FW6_TYPE_RSSCPL);
317 
318 	loc = (uintptr_t *)&t4_fw_msg_handler[type];
319 	atomic_store_rel_ptr(loc, (uintptr_t)h);
320 }
321 
322 void
323 t4_register_cpl_handler(int opcode, cpl_handler_t h)
324 {
325 	uintptr_t *loc;
326 
327 	MPASS(opcode < nitems(t4_cpl_handler));
328 	MPASS(h == NULL || t4_cpl_handler[opcode] == NULL);
329 
330 	loc = (uintptr_t *)&t4_cpl_handler[opcode];
331 	atomic_store_rel_ptr(loc, (uintptr_t)h);
332 }
333 
334 static int
335 set_tcb_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
336     struct mbuf *m)
337 {
338 	const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
339 	u_int tid;
340 	int cookie;
341 
342 	MPASS(m == NULL);
343 
344 	tid = GET_TID(cpl);
345 	if (is_ftid(iq->adapter, tid)) {
346 		/*
347 		 * The return code for filter-write is put in the CPL cookie so
348 		 * we have to rely on the hardware tid (is_ftid) to determine
349 		 * that this is a response to a filter.
350 		 */
351 		cookie = CPL_COOKIE_FILTER;
352 	} else {
353 		cookie = G_COOKIE(cpl->cookie);
354 	}
355 	MPASS(cookie > CPL_COOKIE_RESERVED);
356 	MPASS(cookie < nitems(set_tcb_rpl_handlers));
357 
358 	return (set_tcb_rpl_handlers[cookie](iq, rss, m));
359 }
360 
361 static int
362 l2t_write_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
363     struct mbuf *m)
364 {
365 	const struct cpl_l2t_write_rpl *rpl = (const void *)(rss + 1);
366 	unsigned int cookie;
367 
368 	MPASS(m == NULL);
369 
370 	cookie = GET_TID(rpl) & F_SYNC_WR ? CPL_COOKIE_TOM : CPL_COOKIE_FILTER;
371 	return (l2t_write_rpl_handlers[cookie](iq, rss, m));
372 }
373 
374 static int
375 act_open_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
376     struct mbuf *m)
377 {
378 	const struct cpl_act_open_rpl *cpl = (const void *)(rss + 1);
379 	u_int cookie = G_TID_COOKIE(G_AOPEN_ATID(be32toh(cpl->atid_status)));
380 
381 	MPASS(m == NULL);
382 	MPASS(cookie != CPL_COOKIE_RESERVED);
383 
384 	return (act_open_rpl_handlers[cookie](iq, rss, m));
385 }
386 
387 static void
388 t4_init_shared_cpl_handlers(void)
389 {
390 
391 	t4_register_cpl_handler(CPL_SET_TCB_RPL, set_tcb_rpl_handler);
392 	t4_register_cpl_handler(CPL_L2T_WRITE_RPL, l2t_write_rpl_handler);
393 	t4_register_cpl_handler(CPL_ACT_OPEN_RPL, act_open_rpl_handler);
394 }
395 
396 void
397 t4_register_shared_cpl_handler(int opcode, cpl_handler_t h, int cookie)
398 {
399 	uintptr_t *loc;
400 
401 	MPASS(opcode < nitems(t4_cpl_handler));
402 	MPASS(cookie > CPL_COOKIE_RESERVED);
403 	MPASS(cookie < NUM_CPL_COOKIES);
404 	MPASS(t4_cpl_handler[opcode] != NULL);
405 
406 	switch (opcode) {
407 	case CPL_SET_TCB_RPL:
408 		loc = (uintptr_t *)&set_tcb_rpl_handlers[cookie];
409 		break;
410 	case CPL_L2T_WRITE_RPL:
411 		loc = (uintptr_t *)&l2t_write_rpl_handlers[cookie];
412 		break;
413 	case CPL_ACT_OPEN_RPL:
414 		loc = (uintptr_t *)&act_open_rpl_handlers[cookie];
415 		break;
416 	default:
417 		MPASS(0);
418 		return;
419 	}
420 	MPASS(h == NULL || *loc == (uintptr_t)NULL);
421 	atomic_store_rel_ptr(loc, (uintptr_t)h);
422 }
423 
424 /*
425  * Called on MOD_LOAD.  Validates and calculates the SGE tunables.
426  */
427 void
428 t4_sge_modload(void)
429 {
430 
431 	if (fl_pktshift < 0 || fl_pktshift > 7) {
432 		printf("Invalid hw.cxgbe.fl_pktshift value (%d),"
433 		    " using 2 instead.\n", fl_pktshift);
434 		fl_pktshift = 2;
435 	}
436 
437 	if (spg_len != 64 && spg_len != 128) {
438 		int len;
439 
440 #if defined(__i386__) || defined(__amd64__)
441 		len = cpu_clflush_line_size > 64 ? 128 : 64;
442 #else
443 		len = 64;
444 #endif
445 		if (spg_len != -1) {
446 			printf("Invalid hw.cxgbe.spg_len value (%d),"
447 			    " using %d instead.\n", spg_len, len);
448 		}
449 		spg_len = len;
450 	}
451 
452 	if (cong_drop < -1 || cong_drop > 1) {
453 		printf("Invalid hw.cxgbe.cong_drop value (%d),"
454 		    " using 0 instead.\n", cong_drop);
455 		cong_drop = 0;
456 	}
457 
458 	if (tscale != 1 && (tscale < 3 || tscale > 17)) {
459 		printf("Invalid hw.cxgbe.tscale value (%d),"
460 		    " using 1 instead.\n", tscale);
461 		tscale = 1;
462 	}
463 
464 	extfree_refs = counter_u64_alloc(M_WAITOK);
465 	extfree_rels = counter_u64_alloc(M_WAITOK);
466 	counter_u64_zero(extfree_refs);
467 	counter_u64_zero(extfree_rels);
468 
469 	t4_init_shared_cpl_handlers();
470 	t4_register_cpl_handler(CPL_FW4_MSG, handle_fw_msg);
471 	t4_register_cpl_handler(CPL_FW6_MSG, handle_fw_msg);
472 	t4_register_cpl_handler(CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
473 	t4_register_cpl_handler(CPL_RX_PKT, t4_eth_rx);
474 	t4_register_fw_msg_handler(FW6_TYPE_CMD_RPL, t4_handle_fw_rpl);
475 	t4_register_fw_msg_handler(FW6_TYPE_WRERR_RPL, t4_handle_wrerr_rpl);
476 }
477 
478 void
479 t4_sge_modunload(void)
480 {
481 
482 	counter_u64_free(extfree_refs);
483 	counter_u64_free(extfree_rels);
484 }
485 
486 uint64_t
487 t4_sge_extfree_refs(void)
488 {
489 	uint64_t refs, rels;
490 
491 	rels = counter_u64_fetch(extfree_rels);
492 	refs = counter_u64_fetch(extfree_refs);
493 
494 	return (refs - rels);
495 }
496 
497 static inline void
498 setup_pad_and_pack_boundaries(struct adapter *sc)
499 {
500 	uint32_t v, m;
501 	int pad, pack, pad_shift;
502 
503 	pad_shift = chip_id(sc) > CHELSIO_T5 ? X_T6_INGPADBOUNDARY_SHIFT :
504 	    X_INGPADBOUNDARY_SHIFT;
505 	pad = fl_pad;
506 	if (fl_pad < (1 << pad_shift) ||
507 	    fl_pad > (1 << (pad_shift + M_INGPADBOUNDARY)) ||
508 	    !powerof2(fl_pad)) {
509 		/*
510 		 * If there is any chance that we might use buffer packing and
511 		 * the chip is a T4, then pick 64 as the pad/pack boundary.  Set
512 		 * it to the minimum allowed in all other cases.
513 		 */
514 		pad = is_t4(sc) && buffer_packing ? 64 : 1 << pad_shift;
515 
516 		/*
517 		 * For fl_pad = 0 we'll still write a reasonable value to the
518 		 * register but all the freelists will opt out of padding.
519 		 * We'll complain here only if the user tried to set it to a
520 		 * value greater than 0 that was invalid.
521 		 */
522 		if (fl_pad > 0) {
523 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pad value"
524 			    " (%d), using %d instead.\n", fl_pad, pad);
525 		}
526 	}
527 	m = V_INGPADBOUNDARY(M_INGPADBOUNDARY);
528 	v = V_INGPADBOUNDARY(ilog2(pad) - pad_shift);
529 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
530 
531 	if (is_t4(sc)) {
532 		if (fl_pack != -1 && fl_pack != pad) {
533 			/* Complain but carry on. */
534 			device_printf(sc->dev, "hw.cxgbe.fl_pack (%d) ignored,"
535 			    " using %d instead.\n", fl_pack, pad);
536 		}
537 		return;
538 	}
539 
540 	pack = fl_pack;
541 	if (fl_pack < 16 || fl_pack == 32 || fl_pack > 4096 ||
542 	    !powerof2(fl_pack)) {
543 		pack = max(sc->params.pci.mps, CACHE_LINE_SIZE);
544 		MPASS(powerof2(pack));
545 		if (pack < 16)
546 			pack = 16;
547 		if (pack == 32)
548 			pack = 64;
549 		if (pack > 4096)
550 			pack = 4096;
551 		if (fl_pack != -1) {
552 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pack value"
553 			    " (%d), using %d instead.\n", fl_pack, pack);
554 		}
555 	}
556 	m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY);
557 	if (pack == 16)
558 		v = V_INGPACKBOUNDARY(0);
559 	else
560 		v = V_INGPACKBOUNDARY(ilog2(pack) - 5);
561 
562 	MPASS(!is_t4(sc));	/* T4 doesn't have SGE_CONTROL2 */
563 	t4_set_reg_field(sc, A_SGE_CONTROL2, m, v);
564 }
565 
566 /*
567  * adap->params.vpd.cclk must be set up before this is called.
568  */
569 void
570 t4_tweak_chip_settings(struct adapter *sc)
571 {
572 	int i;
573 	uint32_t v, m;
574 	int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200};
575 	int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk;
576 	int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */
577 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
578 	static int sge_flbuf_sizes[] = {
579 		MCLBYTES,
580 #if MJUMPAGESIZE != MCLBYTES
581 		MJUMPAGESIZE,
582 		MJUMPAGESIZE - CL_METADATA_SIZE,
583 		MJUMPAGESIZE - 2 * MSIZE - CL_METADATA_SIZE,
584 #endif
585 		MJUM9BYTES,
586 		MJUM16BYTES,
587 		MCLBYTES - MSIZE - CL_METADATA_SIZE,
588 		MJUM9BYTES - CL_METADATA_SIZE,
589 		MJUM16BYTES - CL_METADATA_SIZE,
590 	};
591 
592 	KASSERT(sc->flags & MASTER_PF,
593 	    ("%s: trying to change chip settings when not master.", __func__));
594 
595 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
596 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
597 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
598 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
599 
600 	setup_pad_and_pack_boundaries(sc);
601 
602 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
603 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
604 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
605 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
606 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
607 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
608 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
609 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
610 	t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v);
611 
612 	KASSERT(nitems(sge_flbuf_sizes) <= SGE_FLBUF_SIZES,
613 	    ("%s: hw buffer size table too big", __func__));
614 	for (i = 0; i < min(nitems(sge_flbuf_sizes), SGE_FLBUF_SIZES); i++) {
615 		t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i),
616 		    sge_flbuf_sizes[i]);
617 	}
618 
619 	v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) |
620 	    V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]);
621 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v);
622 
623 	KASSERT(intr_timer[0] <= timer_max,
624 	    ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0],
625 	    timer_max));
626 	for (i = 1; i < nitems(intr_timer); i++) {
627 		KASSERT(intr_timer[i] >= intr_timer[i - 1],
628 		    ("%s: timers not listed in increasing order (%d)",
629 		    __func__, i));
630 
631 		while (intr_timer[i] > timer_max) {
632 			if (i == nitems(intr_timer) - 1) {
633 				intr_timer[i] = timer_max;
634 				break;
635 			}
636 			intr_timer[i] += intr_timer[i - 1];
637 			intr_timer[i] /= 2;
638 		}
639 	}
640 
641 	v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) |
642 	    V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1]));
643 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v);
644 	v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) |
645 	    V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3]));
646 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v);
647 	v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) |
648 	    V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5]));
649 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v);
650 
651 	if (chip_id(sc) >= CHELSIO_T6) {
652 		m = V_TSCALE(M_TSCALE);
653 		if (tscale == 1)
654 			v = 0;
655 		else
656 			v = V_TSCALE(tscale - 2);
657 		t4_set_reg_field(sc, A_SGE_ITP_CONTROL, m, v);
658 
659 		if (sc->debug_flags & DF_DISABLE_TCB_CACHE) {
660 			m = V_RDTHRESHOLD(M_RDTHRESHOLD) | F_WRTHRTHRESHEN |
661 			    V_WRTHRTHRESH(M_WRTHRTHRESH);
662 			t4_tp_pio_read(sc, &v, 1, A_TP_CMM_CONFIG, 1);
663 			v &= ~m;
664 			v |= V_RDTHRESHOLD(1) | F_WRTHRTHRESHEN |
665 			    V_WRTHRTHRESH(16);
666 			t4_tp_pio_write(sc, &v, 1, A_TP_CMM_CONFIG, 1);
667 		}
668 	}
669 
670 	/* 4K, 16K, 64K, 256K DDP "page sizes" for TDDP */
671 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
672 	t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v);
673 
674 	/*
675 	 * 4K, 8K, 16K, 64K DDP "page sizes" for iSCSI DDP.  These have been
676 	 * chosen with MAXPHYS = 128K in mind.  The largest DDP buffer that we
677 	 * may have to deal with is MAXPHYS + 1 page.
678 	 */
679 	v = V_HPZ0(0) | V_HPZ1(1) | V_HPZ2(2) | V_HPZ3(4);
680 	t4_write_reg(sc, A_ULP_RX_ISCSI_PSZ, v);
681 
682 	/* We use multiple DDP page sizes both in plain-TOE and ISCSI modes. */
683 	m = v = F_TDDPTAGTCB | F_ISCSITAGTCB;
684 	t4_set_reg_field(sc, A_ULP_RX_CTL, m, v);
685 
686 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
687 	    F_RESETDDPOFFSET;
688 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
689 	t4_set_reg_field(sc, A_TP_PARA_REG5, m, v);
690 }
691 
692 /*
693  * SGE wants the buffer to be at least 64B and then a multiple of 16.  If
694  * padding is in use, the buffer's start and end need to be aligned to the pad
695  * boundary as well.  We'll just make sure that the size is a multiple of the
696  * boundary here, it is up to the buffer allocation code to make sure the start
697  * of the buffer is aligned as well.
698  */
699 static inline int
700 hwsz_ok(struct adapter *sc, int hwsz)
701 {
702 	int mask = fl_pad ? sc->params.sge.pad_boundary - 1 : 16 - 1;
703 
704 	return (hwsz >= 64 && (hwsz & mask) == 0);
705 }
706 
707 /*
708  * XXX: driver really should be able to deal with unexpected settings.
709  */
710 int
711 t4_read_chip_settings(struct adapter *sc)
712 {
713 	struct sge *s = &sc->sge;
714 	struct sge_params *sp = &sc->params.sge;
715 	int i, j, n, rc = 0;
716 	uint32_t m, v, r;
717 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
718 	static int sw_buf_sizes[] = {	/* Sorted by size */
719 		MCLBYTES,
720 #if MJUMPAGESIZE != MCLBYTES
721 		MJUMPAGESIZE,
722 #endif
723 		MJUM9BYTES,
724 		MJUM16BYTES
725 	};
726 	struct sw_zone_info *swz, *safe_swz;
727 	struct hw_buf_info *hwb;
728 
729 	m = F_RXPKTCPLMODE;
730 	v = F_RXPKTCPLMODE;
731 	r = sc->params.sge.sge_control;
732 	if ((r & m) != v) {
733 		device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r);
734 		rc = EINVAL;
735 	}
736 
737 	/*
738 	 * If this changes then every single use of PAGE_SHIFT in the driver
739 	 * needs to be carefully reviewed for PAGE_SHIFT vs sp->page_shift.
740 	 */
741 	if (sp->page_shift != PAGE_SHIFT) {
742 		device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r);
743 		rc = EINVAL;
744 	}
745 
746 	/* Filter out unusable hw buffer sizes entirely (mark with -2). */
747 	hwb = &s->hw_buf_info[0];
748 	for (i = 0; i < nitems(s->hw_buf_info); i++, hwb++) {
749 		r = sc->params.sge.sge_fl_buffer_size[i];
750 		hwb->size = r;
751 		hwb->zidx = hwsz_ok(sc, r) ? -1 : -2;
752 		hwb->next = -1;
753 	}
754 
755 	/*
756 	 * Create a sorted list in decreasing order of hw buffer sizes (and so
757 	 * increasing order of spare area) for each software zone.
758 	 *
759 	 * If padding is enabled then the start and end of the buffer must align
760 	 * to the pad boundary; if packing is enabled then they must align with
761 	 * the pack boundary as well.  Allocations from the cluster zones are
762 	 * aligned to min(size, 4K), so the buffer starts at that alignment and
763 	 * ends at hwb->size alignment.  If mbuf inlining is allowed the
764 	 * starting alignment will be reduced to MSIZE and the driver will
765 	 * exercise appropriate caution when deciding on the best buffer layout
766 	 * to use.
767 	 */
768 	n = 0;	/* no usable buffer size to begin with */
769 	swz = &s->sw_zone_info[0];
770 	safe_swz = NULL;
771 	for (i = 0; i < SW_ZONE_SIZES; i++, swz++) {
772 		int8_t head = -1, tail = -1;
773 
774 		swz->size = sw_buf_sizes[i];
775 		swz->zone = m_getzone(swz->size);
776 		swz->type = m_gettype(swz->size);
777 
778 		if (swz->size < PAGE_SIZE) {
779 			MPASS(powerof2(swz->size));
780 			if (fl_pad && (swz->size % sp->pad_boundary != 0))
781 				continue;
782 		}
783 
784 		if (swz->size == safest_rx_cluster)
785 			safe_swz = swz;
786 
787 		hwb = &s->hw_buf_info[0];
788 		for (j = 0; j < SGE_FLBUF_SIZES; j++, hwb++) {
789 			if (hwb->zidx != -1 || hwb->size > swz->size)
790 				continue;
791 #ifdef INVARIANTS
792 			if (fl_pad)
793 				MPASS(hwb->size % sp->pad_boundary == 0);
794 #endif
795 			hwb->zidx = i;
796 			if (head == -1)
797 				head = tail = j;
798 			else if (hwb->size < s->hw_buf_info[tail].size) {
799 				s->hw_buf_info[tail].next = j;
800 				tail = j;
801 			} else {
802 				int8_t *cur;
803 				struct hw_buf_info *t;
804 
805 				for (cur = &head; *cur != -1; cur = &t->next) {
806 					t = &s->hw_buf_info[*cur];
807 					if (hwb->size == t->size) {
808 						hwb->zidx = -2;
809 						break;
810 					}
811 					if (hwb->size > t->size) {
812 						hwb->next = *cur;
813 						*cur = j;
814 						break;
815 					}
816 				}
817 			}
818 		}
819 		swz->head_hwidx = head;
820 		swz->tail_hwidx = tail;
821 
822 		if (tail != -1) {
823 			n++;
824 			if (swz->size - s->hw_buf_info[tail].size >=
825 			    CL_METADATA_SIZE)
826 				sc->flags |= BUF_PACKING_OK;
827 		}
828 	}
829 	if (n == 0) {
830 		device_printf(sc->dev, "no usable SGE FL buffer size.\n");
831 		rc = EINVAL;
832 	}
833 
834 	s->safe_hwidx1 = -1;
835 	s->safe_hwidx2 = -1;
836 	if (safe_swz != NULL) {
837 		s->safe_hwidx1 = safe_swz->head_hwidx;
838 		for (i = safe_swz->head_hwidx; i != -1; i = hwb->next) {
839 			int spare;
840 
841 			hwb = &s->hw_buf_info[i];
842 #ifdef INVARIANTS
843 			if (fl_pad)
844 				MPASS(hwb->size % sp->pad_boundary == 0);
845 #endif
846 			spare = safe_swz->size - hwb->size;
847 			if (spare >= CL_METADATA_SIZE) {
848 				s->safe_hwidx2 = i;
849 				break;
850 			}
851 		}
852 	}
853 
854 	if (sc->flags & IS_VF)
855 		return (0);
856 
857 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
858 	r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ);
859 	if (r != v) {
860 		device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r);
861 		rc = EINVAL;
862 	}
863 
864 	m = v = F_TDDPTAGTCB;
865 	r = t4_read_reg(sc, A_ULP_RX_CTL);
866 	if ((r & m) != v) {
867 		device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r);
868 		rc = EINVAL;
869 	}
870 
871 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
872 	    F_RESETDDPOFFSET;
873 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
874 	r = t4_read_reg(sc, A_TP_PARA_REG5);
875 	if ((r & m) != v) {
876 		device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r);
877 		rc = EINVAL;
878 	}
879 
880 	t4_init_tp_params(sc, 1);
881 
882 	t4_read_mtu_tbl(sc, sc->params.mtus, NULL);
883 	t4_load_mtus(sc, sc->params.mtus, sc->params.a_wnd, sc->params.b_wnd);
884 
885 	return (rc);
886 }
887 
888 int
889 t4_create_dma_tag(struct adapter *sc)
890 {
891 	int rc;
892 
893 	rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
894 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
895 	    BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
896 	    NULL, &sc->dmat);
897 	if (rc != 0) {
898 		device_printf(sc->dev,
899 		    "failed to create main DMA tag: %d\n", rc);
900 	}
901 
902 	return (rc);
903 }
904 
905 void
906 t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
907     struct sysctl_oid_list *children)
908 {
909 	struct sge_params *sp = &sc->params.sge;
910 
911 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "buffer_sizes",
912 	    CTLTYPE_STRING | CTLFLAG_RD, &sc->sge, 0, sysctl_bufsizes, "A",
913 	    "freelist buffer sizes");
914 
915 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD,
916 	    NULL, sp->fl_pktshift, "payload DMA offset in rx buffer (bytes)");
917 
918 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD,
919 	    NULL, sp->pad_boundary, "payload pad boundary (bytes)");
920 
921 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD,
922 	    NULL, sp->spg_len, "status page size (bytes)");
923 
924 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD,
925 	    NULL, cong_drop, "congestion drop setting");
926 
927 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD,
928 	    NULL, sp->pack_boundary, "payload pack boundary (bytes)");
929 }
930 
931 int
932 t4_destroy_dma_tag(struct adapter *sc)
933 {
934 	if (sc->dmat)
935 		bus_dma_tag_destroy(sc->dmat);
936 
937 	return (0);
938 }
939 
940 /*
941  * Allocate and initialize the firmware event queue and the management queue.
942  *
943  * Returns errno on failure.  Resources allocated up to that point may still be
944  * allocated.  Caller is responsible for cleanup in case this function fails.
945  */
946 int
947 t4_setup_adapter_queues(struct adapter *sc)
948 {
949 	int rc;
950 
951 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
952 
953 	sysctl_ctx_init(&sc->ctx);
954 	sc->flags |= ADAP_SYSCTL_CTX;
955 
956 	/*
957 	 * Firmware event queue
958 	 */
959 	rc = alloc_fwq(sc);
960 	if (rc != 0)
961 		return (rc);
962 
963 	/*
964 	 * Management queue.  This is just a control queue that uses the fwq as
965 	 * its associated iq.
966 	 */
967 	if (!(sc->flags & IS_VF))
968 		rc = alloc_mgmtq(sc);
969 
970 	return (rc);
971 }
972 
973 /*
974  * Idempotent
975  */
976 int
977 t4_teardown_adapter_queues(struct adapter *sc)
978 {
979 
980 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
981 
982 	/* Do this before freeing the queue */
983 	if (sc->flags & ADAP_SYSCTL_CTX) {
984 		sysctl_ctx_free(&sc->ctx);
985 		sc->flags &= ~ADAP_SYSCTL_CTX;
986 	}
987 
988 	free_mgmtq(sc);
989 	free_fwq(sc);
990 
991 	return (0);
992 }
993 
994 /* Maximum payload that can be delivered with a single iq descriptor */
995 static inline int
996 mtu_to_max_payload(struct adapter *sc, int mtu, const int toe)
997 {
998 	int payload;
999 
1000 #ifdef TCP_OFFLOAD
1001 	if (toe) {
1002 		int rxcs = G_RXCOALESCESIZE(t4_read_reg(sc, A_TP_PARA_REG2));
1003 
1004 		/* Note that COP can set rx_coalesce on/off per connection. */
1005 		payload = max(mtu, rxcs);
1006 	} else {
1007 #endif
1008 		/* large enough even when hw VLAN extraction is disabled */
1009 		payload = sc->params.sge.fl_pktshift + ETHER_HDR_LEN +
1010 		    ETHER_VLAN_ENCAP_LEN + mtu;
1011 #ifdef TCP_OFFLOAD
1012 	}
1013 #endif
1014 
1015 	return (payload);
1016 }
1017 
1018 int
1019 t4_setup_vi_queues(struct vi_info *vi)
1020 {
1021 	int rc = 0, i, intr_idx, iqidx;
1022 	struct sge_rxq *rxq;
1023 	struct sge_txq *txq;
1024 	struct sge_wrq *ctrlq;
1025 #ifdef TCP_OFFLOAD
1026 	struct sge_ofld_rxq *ofld_rxq;
1027 	struct sge_wrq *ofld_txq;
1028 #endif
1029 #ifdef DEV_NETMAP
1030 	int saved_idx;
1031 	struct sge_nm_rxq *nm_rxq;
1032 	struct sge_nm_txq *nm_txq;
1033 #endif
1034 	char name[16];
1035 	struct port_info *pi = vi->pi;
1036 	struct adapter *sc = pi->adapter;
1037 	struct ifnet *ifp = vi->ifp;
1038 	struct sysctl_oid *oid = device_get_sysctl_tree(vi->dev);
1039 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
1040 	int maxp, mtu = ifp->if_mtu;
1041 
1042 	/* Interrupt vector to start from (when using multiple vectors) */
1043 	intr_idx = vi->first_intr;
1044 
1045 #ifdef DEV_NETMAP
1046 	saved_idx = intr_idx;
1047 	if (ifp->if_capabilities & IFCAP_NETMAP) {
1048 
1049 		/* netmap is supported with direct interrupts only. */
1050 		MPASS(!forwarding_intr_to_fwq(sc));
1051 
1052 		/*
1053 		 * We don't have buffers to back the netmap rx queues
1054 		 * right now so we create the queues in a way that
1055 		 * doesn't set off any congestion signal in the chip.
1056 		 */
1057 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_rxq",
1058 		    CTLFLAG_RD, NULL, "rx queues");
1059 		for_each_nm_rxq(vi, i, nm_rxq) {
1060 			rc = alloc_nm_rxq(vi, nm_rxq, intr_idx, i, oid);
1061 			if (rc != 0)
1062 				goto done;
1063 			intr_idx++;
1064 		}
1065 
1066 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_txq",
1067 		    CTLFLAG_RD, NULL, "tx queues");
1068 		for_each_nm_txq(vi, i, nm_txq) {
1069 			iqidx = vi->first_nm_rxq + (i % vi->nnmrxq);
1070 			rc = alloc_nm_txq(vi, nm_txq, iqidx, i, oid);
1071 			if (rc != 0)
1072 				goto done;
1073 		}
1074 	}
1075 
1076 	/* Normal rx queues and netmap rx queues share the same interrupts. */
1077 	intr_idx = saved_idx;
1078 #endif
1079 
1080 	/*
1081 	 * Allocate rx queues first because a default iqid is required when
1082 	 * creating a tx queue.
1083 	 */
1084 	maxp = mtu_to_max_payload(sc, mtu, 0);
1085 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
1086 	    CTLFLAG_RD, NULL, "rx queues");
1087 	for_each_rxq(vi, i, rxq) {
1088 
1089 		init_iq(&rxq->iq, sc, vi->tmr_idx, vi->pktc_idx, vi->qsize_rxq);
1090 
1091 		snprintf(name, sizeof(name), "%s rxq%d-fl",
1092 		    device_get_nameunit(vi->dev), i);
1093 		init_fl(sc, &rxq->fl, vi->qsize_rxq / 8, maxp, name);
1094 
1095 		rc = alloc_rxq(vi, rxq,
1096 		    forwarding_intr_to_fwq(sc) ? -1 : intr_idx, i, oid);
1097 		if (rc != 0)
1098 			goto done;
1099 		intr_idx++;
1100 	}
1101 #ifdef DEV_NETMAP
1102 	if (ifp->if_capabilities & IFCAP_NETMAP)
1103 		intr_idx = saved_idx + max(vi->nrxq, vi->nnmrxq);
1104 #endif
1105 #ifdef TCP_OFFLOAD
1106 	maxp = mtu_to_max_payload(sc, mtu, 1);
1107 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq",
1108 	    CTLFLAG_RD, NULL, "rx queues for offloaded TCP connections");
1109 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1110 
1111 		init_iq(&ofld_rxq->iq, sc, vi->ofld_tmr_idx, vi->ofld_pktc_idx,
1112 		    vi->qsize_rxq);
1113 
1114 		snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
1115 		    device_get_nameunit(vi->dev), i);
1116 		init_fl(sc, &ofld_rxq->fl, vi->qsize_rxq / 8, maxp, name);
1117 
1118 		rc = alloc_ofld_rxq(vi, ofld_rxq,
1119 		    forwarding_intr_to_fwq(sc) ? -1 : intr_idx, i, oid);
1120 		if (rc != 0)
1121 			goto done;
1122 		intr_idx++;
1123 	}
1124 #endif
1125 
1126 	/*
1127 	 * Now the tx queues.
1128 	 */
1129 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "txq", CTLFLAG_RD,
1130 	    NULL, "tx queues");
1131 	for_each_txq(vi, i, txq) {
1132 		iqidx = vi->first_rxq + (i % vi->nrxq);
1133 		snprintf(name, sizeof(name), "%s txq%d",
1134 		    device_get_nameunit(vi->dev), i);
1135 		init_eq(sc, &txq->eq, EQ_ETH, vi->qsize_txq, pi->tx_chan,
1136 		    sc->sge.rxq[iqidx].iq.cntxt_id, name);
1137 
1138 		rc = alloc_txq(vi, txq, i, oid);
1139 		if (rc != 0)
1140 			goto done;
1141 	}
1142 #ifdef TCP_OFFLOAD
1143 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_txq",
1144 	    CTLFLAG_RD, NULL, "tx queues for offloaded TCP connections");
1145 	for_each_ofld_txq(vi, i, ofld_txq) {
1146 		struct sysctl_oid *oid2;
1147 
1148 		iqidx = vi->first_ofld_rxq + (i % vi->nofldrxq);
1149 		snprintf(name, sizeof(name), "%s ofld_txq%d",
1150 		    device_get_nameunit(vi->dev), i);
1151 		init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq, pi->tx_chan,
1152 		    sc->sge.ofld_rxq[iqidx].iq.cntxt_id, name);
1153 
1154 		snprintf(name, sizeof(name), "%d", i);
1155 		oid2 = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
1156 		    name, CTLFLAG_RD, NULL, "offload tx queue");
1157 
1158 		rc = alloc_wrq(sc, vi, ofld_txq, oid2);
1159 		if (rc != 0)
1160 			goto done;
1161 	}
1162 #endif
1163 
1164 	/*
1165 	 * Finally, the control queue.
1166 	 */
1167 	if (!IS_MAIN_VI(vi) || sc->flags & IS_VF)
1168 		goto done;
1169 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ctrlq", CTLFLAG_RD,
1170 	    NULL, "ctrl queue");
1171 	ctrlq = &sc->sge.ctrlq[pi->port_id];
1172 	snprintf(name, sizeof(name), "%s ctrlq", device_get_nameunit(vi->dev));
1173 	init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, pi->tx_chan,
1174 	    sc->sge.rxq[vi->first_rxq].iq.cntxt_id, name);
1175 	rc = alloc_wrq(sc, vi, ctrlq, oid);
1176 
1177 done:
1178 	if (rc)
1179 		t4_teardown_vi_queues(vi);
1180 
1181 	return (rc);
1182 }
1183 
1184 /*
1185  * Idempotent
1186  */
1187 int
1188 t4_teardown_vi_queues(struct vi_info *vi)
1189 {
1190 	int i;
1191 	struct port_info *pi = vi->pi;
1192 	struct adapter *sc = pi->adapter;
1193 	struct sge_rxq *rxq;
1194 	struct sge_txq *txq;
1195 #ifdef TCP_OFFLOAD
1196 	struct sge_ofld_rxq *ofld_rxq;
1197 	struct sge_wrq *ofld_txq;
1198 #endif
1199 #ifdef DEV_NETMAP
1200 	struct sge_nm_rxq *nm_rxq;
1201 	struct sge_nm_txq *nm_txq;
1202 #endif
1203 
1204 	/* Do this before freeing the queues */
1205 	if (vi->flags & VI_SYSCTL_CTX) {
1206 		sysctl_ctx_free(&vi->ctx);
1207 		vi->flags &= ~VI_SYSCTL_CTX;
1208 	}
1209 
1210 #ifdef DEV_NETMAP
1211 	if (vi->ifp->if_capabilities & IFCAP_NETMAP) {
1212 		for_each_nm_txq(vi, i, nm_txq) {
1213 			free_nm_txq(vi, nm_txq);
1214 		}
1215 
1216 		for_each_nm_rxq(vi, i, nm_rxq) {
1217 			free_nm_rxq(vi, nm_rxq);
1218 		}
1219 	}
1220 #endif
1221 
1222 	/*
1223 	 * Take down all the tx queues first, as they reference the rx queues
1224 	 * (for egress updates, etc.).
1225 	 */
1226 
1227 	if (IS_MAIN_VI(vi) && !(sc->flags & IS_VF))
1228 		free_wrq(sc, &sc->sge.ctrlq[pi->port_id]);
1229 
1230 	for_each_txq(vi, i, txq) {
1231 		free_txq(vi, txq);
1232 	}
1233 #ifdef TCP_OFFLOAD
1234 	for_each_ofld_txq(vi, i, ofld_txq) {
1235 		free_wrq(sc, ofld_txq);
1236 	}
1237 #endif
1238 
1239 	/*
1240 	 * Then take down the rx queues.
1241 	 */
1242 
1243 	for_each_rxq(vi, i, rxq) {
1244 		free_rxq(vi, rxq);
1245 	}
1246 #ifdef TCP_OFFLOAD
1247 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1248 		free_ofld_rxq(vi, ofld_rxq);
1249 	}
1250 #endif
1251 
1252 	return (0);
1253 }
1254 
1255 /*
1256  * Deals with errors and the firmware event queue.  All data rx queues forward
1257  * their interrupt to the firmware event queue.
1258  */
1259 void
1260 t4_intr_all(void *arg)
1261 {
1262 	struct adapter *sc = arg;
1263 	struct sge_iq *fwq = &sc->sge.fwq;
1264 
1265 	t4_intr_err(arg);
1266 	if (atomic_cmpset_int(&fwq->state, IQS_IDLE, IQS_BUSY)) {
1267 		service_iq(fwq, 0);
1268 		atomic_cmpset_int(&fwq->state, IQS_BUSY, IQS_IDLE);
1269 	}
1270 }
1271 
1272 /* Deals with error interrupts */
1273 void
1274 t4_intr_err(void *arg)
1275 {
1276 	struct adapter *sc = arg;
1277 
1278 	t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
1279 	t4_slow_intr_handler(sc);
1280 }
1281 
1282 void
1283 t4_intr_evt(void *arg)
1284 {
1285 	struct sge_iq *iq = arg;
1286 
1287 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1288 		service_iq(iq, 0);
1289 		atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1290 	}
1291 }
1292 
1293 void
1294 t4_intr(void *arg)
1295 {
1296 	struct sge_iq *iq = arg;
1297 
1298 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1299 		service_iq(iq, 0);
1300 		atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1301 	}
1302 }
1303 
1304 void
1305 t4_vi_intr(void *arg)
1306 {
1307 	struct irq *irq = arg;
1308 
1309 #ifdef DEV_NETMAP
1310 	if (atomic_cmpset_int(&irq->nm_state, NM_ON, NM_BUSY)) {
1311 		t4_nm_intr(irq->nm_rxq);
1312 		atomic_cmpset_int(&irq->nm_state, NM_BUSY, NM_ON);
1313 	}
1314 #endif
1315 	if (irq->rxq != NULL)
1316 		t4_intr(irq->rxq);
1317 }
1318 
1319 static inline int
1320 sort_before_lro(struct lro_ctrl *lro)
1321 {
1322 
1323 	return (lro->lro_mbuf_max != 0);
1324 }
1325 
1326 /*
1327  * Deals with anything and everything on the given ingress queue.
1328  */
1329 static int
1330 service_iq(struct sge_iq *iq, int budget)
1331 {
1332 	struct sge_iq *q;
1333 	struct sge_rxq *rxq = iq_to_rxq(iq);	/* Use iff iq is part of rxq */
1334 	struct sge_fl *fl;			/* Use iff IQ_HAS_FL */
1335 	struct adapter *sc = iq->adapter;
1336 	struct iq_desc *d = &iq->desc[iq->cidx];
1337 	int ndescs = 0, limit;
1338 	int rsp_type, refill;
1339 	uint32_t lq;
1340 	uint16_t fl_hw_cidx;
1341 	struct mbuf *m0;
1342 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1343 #if defined(INET) || defined(INET6)
1344 	const struct timeval lro_timeout = {0, sc->lro_timeout};
1345 	struct lro_ctrl *lro = &rxq->lro;
1346 #endif
1347 
1348 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1349 
1350 	limit = budget ? budget : iq->qsize / 16;
1351 
1352 	if (iq->flags & IQ_HAS_FL) {
1353 		fl = &rxq->fl;
1354 		fl_hw_cidx = fl->hw_cidx;	/* stable snapshot */
1355 	} else {
1356 		fl = NULL;
1357 		fl_hw_cidx = 0;			/* to silence gcc warning */
1358 	}
1359 
1360 #if defined(INET) || defined(INET6)
1361 	if (iq->flags & IQ_ADJ_CREDIT) {
1362 		MPASS(sort_before_lro(lro));
1363 		iq->flags &= ~IQ_ADJ_CREDIT;
1364 		if ((d->rsp.u.type_gen & F_RSPD_GEN) != iq->gen) {
1365 			tcp_lro_flush_all(lro);
1366 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(1) |
1367 			    V_INGRESSQID((u32)iq->cntxt_id) |
1368 			    V_SEINTARM(iq->intr_params));
1369 			return (0);
1370 		}
1371 		ndescs = 1;
1372 	}
1373 #else
1374 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1375 #endif
1376 
1377 	/*
1378 	 * We always come back and check the descriptor ring for new indirect
1379 	 * interrupts and other responses after running a single handler.
1380 	 */
1381 	for (;;) {
1382 		while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1383 
1384 			rmb();
1385 
1386 			refill = 0;
1387 			m0 = NULL;
1388 			rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1389 			lq = be32toh(d->rsp.pldbuflen_qid);
1390 
1391 			switch (rsp_type) {
1392 			case X_RSPD_TYPE_FLBUF:
1393 
1394 				KASSERT(iq->flags & IQ_HAS_FL,
1395 				    ("%s: data for an iq (%p) with no freelist",
1396 				    __func__, iq));
1397 
1398 				m0 = get_fl_payload(sc, fl, lq);
1399 				if (__predict_false(m0 == NULL))
1400 					goto process_iql;
1401 				refill = IDXDIFF(fl->hw_cidx, fl_hw_cidx, fl->sidx) > 2;
1402 #ifdef T4_PKT_TIMESTAMP
1403 				/*
1404 				 * 60 bit timestamp for the payload is
1405 				 * *(uint64_t *)m0->m_pktdat.  Note that it is
1406 				 * in the leading free-space in the mbuf.  The
1407 				 * kernel can clobber it during a pullup,
1408 				 * m_copymdata, etc.  You need to make sure that
1409 				 * the mbuf reaches you unmolested if you care
1410 				 * about the timestamp.
1411 				 */
1412 				*(uint64_t *)m0->m_pktdat =
1413 				    be64toh(ctrl->u.last_flit) &
1414 				    0xfffffffffffffff;
1415 #endif
1416 
1417 				/* fall through */
1418 
1419 			case X_RSPD_TYPE_CPL:
1420 				KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1421 				    ("%s: bad opcode %02x.", __func__,
1422 				    d->rss.opcode));
1423 				t4_cpl_handler[d->rss.opcode](iq, &d->rss, m0);
1424 				break;
1425 
1426 			case X_RSPD_TYPE_INTR:
1427 
1428 				/*
1429 				 * Interrupts should be forwarded only to queues
1430 				 * that are not forwarding their interrupts.
1431 				 * This means service_iq can recurse but only 1
1432 				 * level deep.
1433 				 */
1434 				KASSERT(budget == 0,
1435 				    ("%s: budget %u, rsp_type %u", __func__,
1436 				    budget, rsp_type));
1437 
1438 				/*
1439 				 * There are 1K interrupt-capable queues (qids 0
1440 				 * through 1023).  A response type indicating a
1441 				 * forwarded interrupt with a qid >= 1K is an
1442 				 * iWARP async notification.
1443 				 */
1444 				if (lq >= 1024) {
1445                                         t4_an_handler(iq, &d->rsp);
1446                                         break;
1447                                 }
1448 
1449 				q = sc->sge.iqmap[lq - sc->sge.iq_start -
1450 				    sc->sge.iq_base];
1451 				if (atomic_cmpset_int(&q->state, IQS_IDLE,
1452 				    IQS_BUSY)) {
1453 					if (service_iq(q, q->qsize / 16) == 0) {
1454 						atomic_cmpset_int(&q->state,
1455 						    IQS_BUSY, IQS_IDLE);
1456 					} else {
1457 						STAILQ_INSERT_TAIL(&iql, q,
1458 						    link);
1459 					}
1460 				}
1461 				break;
1462 
1463 			default:
1464 				KASSERT(0,
1465 				    ("%s: illegal response type %d on iq %p",
1466 				    __func__, rsp_type, iq));
1467 				log(LOG_ERR,
1468 				    "%s: illegal response type %d on iq %p",
1469 				    device_get_nameunit(sc->dev), rsp_type, iq);
1470 				break;
1471 			}
1472 
1473 			d++;
1474 			if (__predict_false(++iq->cidx == iq->sidx)) {
1475 				iq->cidx = 0;
1476 				iq->gen ^= F_RSPD_GEN;
1477 				d = &iq->desc[0];
1478 			}
1479 			if (__predict_false(++ndescs == limit)) {
1480 				t4_write_reg(sc, sc->sge_gts_reg,
1481 				    V_CIDXINC(ndescs) |
1482 				    V_INGRESSQID(iq->cntxt_id) |
1483 				    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1484 				ndescs = 0;
1485 
1486 #if defined(INET) || defined(INET6)
1487 				if (iq->flags & IQ_LRO_ENABLED &&
1488 				    !sort_before_lro(lro) &&
1489 				    sc->lro_timeout != 0) {
1490 					tcp_lro_flush_inactive(lro,
1491 					    &lro_timeout);
1492 				}
1493 #endif
1494 
1495 				if (budget) {
1496 					if (iq->flags & IQ_HAS_FL) {
1497 						FL_LOCK(fl);
1498 						refill_fl(sc, fl, 32);
1499 						FL_UNLOCK(fl);
1500 					}
1501 					return (EINPROGRESS);
1502 				}
1503 			}
1504 			if (refill) {
1505 				FL_LOCK(fl);
1506 				refill_fl(sc, fl, 32);
1507 				FL_UNLOCK(fl);
1508 				fl_hw_cidx = fl->hw_cidx;
1509 			}
1510 		}
1511 
1512 process_iql:
1513 		if (STAILQ_EMPTY(&iql))
1514 			break;
1515 
1516 		/*
1517 		 * Process the head only, and send it to the back of the list if
1518 		 * it's still not done.
1519 		 */
1520 		q = STAILQ_FIRST(&iql);
1521 		STAILQ_REMOVE_HEAD(&iql, link);
1522 		if (service_iq(q, q->qsize / 8) == 0)
1523 			atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1524 		else
1525 			STAILQ_INSERT_TAIL(&iql, q, link);
1526 	}
1527 
1528 #if defined(INET) || defined(INET6)
1529 	if (iq->flags & IQ_LRO_ENABLED) {
1530 		if (ndescs > 0 && lro->lro_mbuf_count > 8) {
1531 			MPASS(sort_before_lro(lro));
1532 			/* hold back one credit and don't flush LRO state */
1533 			iq->flags |= IQ_ADJ_CREDIT;
1534 			ndescs--;
1535 		} else {
1536 			tcp_lro_flush_all(lro);
1537 		}
1538 	}
1539 #endif
1540 
1541 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1542 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1543 
1544 	if (iq->flags & IQ_HAS_FL) {
1545 		int starved;
1546 
1547 		FL_LOCK(fl);
1548 		starved = refill_fl(sc, fl, 64);
1549 		FL_UNLOCK(fl);
1550 		if (__predict_false(starved != 0))
1551 			add_fl_to_sfl(sc, fl);
1552 	}
1553 
1554 	return (0);
1555 }
1556 
1557 static inline int
1558 cl_has_metadata(struct sge_fl *fl, struct cluster_layout *cll)
1559 {
1560 	int rc = fl->flags & FL_BUF_PACKING || cll->region1 > 0;
1561 
1562 	if (rc)
1563 		MPASS(cll->region3 >= CL_METADATA_SIZE);
1564 
1565 	return (rc);
1566 }
1567 
1568 static inline struct cluster_metadata *
1569 cl_metadata(struct adapter *sc, struct sge_fl *fl, struct cluster_layout *cll,
1570     caddr_t cl)
1571 {
1572 
1573 	if (cl_has_metadata(fl, cll)) {
1574 		struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1575 
1576 		return ((struct cluster_metadata *)(cl + swz->size) - 1);
1577 	}
1578 	return (NULL);
1579 }
1580 
1581 static void
1582 rxb_free(struct mbuf *m)
1583 {
1584 	uma_zone_t zone = m->m_ext.ext_arg1;
1585 	void *cl = m->m_ext.ext_arg2;
1586 
1587 	uma_zfree(zone, cl);
1588 	counter_u64_add(extfree_rels, 1);
1589 }
1590 
1591 /*
1592  * The mbuf returned by this function could be allocated from zone_mbuf or
1593  * constructed in spare room in the cluster.
1594  *
1595  * The mbuf carries the payload in one of these ways
1596  * a) frame inside the mbuf (mbuf from zone_mbuf)
1597  * b) m_cljset (for clusters without metadata) zone_mbuf
1598  * c) m_extaddref (cluster with metadata) inline mbuf
1599  * d) m_extaddref (cluster with metadata) zone_mbuf
1600  */
1601 static struct mbuf *
1602 get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1603     int remaining)
1604 {
1605 	struct mbuf *m;
1606 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1607 	struct cluster_layout *cll = &sd->cll;
1608 	struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1609 	struct hw_buf_info *hwb = &sc->sge.hw_buf_info[cll->hwidx];
1610 	struct cluster_metadata *clm = cl_metadata(sc, fl, cll, sd->cl);
1611 	int len, blen;
1612 	caddr_t payload;
1613 
1614 	blen = hwb->size - fl->rx_offset;	/* max possible in this buf */
1615 	len = min(remaining, blen);
1616 	payload = sd->cl + cll->region1 + fl->rx_offset;
1617 	if (fl->flags & FL_BUF_PACKING) {
1618 		const u_int l = fr_offset + len;
1619 		const u_int pad = roundup2(l, fl->buf_boundary) - l;
1620 
1621 		if (fl->rx_offset + len + pad < hwb->size)
1622 			blen = len + pad;
1623 		MPASS(fl->rx_offset + blen <= hwb->size);
1624 	} else {
1625 		MPASS(fl->rx_offset == 0);	/* not packing */
1626 	}
1627 
1628 
1629 	if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) {
1630 
1631 		/*
1632 		 * Copy payload into a freshly allocated mbuf.
1633 		 */
1634 
1635 		m = fr_offset == 0 ?
1636 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1637 		if (m == NULL)
1638 			return (NULL);
1639 		fl->mbuf_allocated++;
1640 #ifdef T4_PKT_TIMESTAMP
1641 		/* Leave room for a timestamp */
1642 		m->m_data += 8;
1643 #endif
1644 		/* copy data to mbuf */
1645 		bcopy(payload, mtod(m, caddr_t), len);
1646 
1647 	} else if (sd->nmbuf * MSIZE < cll->region1) {
1648 
1649 		/*
1650 		 * There's spare room in the cluster for an mbuf.  Create one
1651 		 * and associate it with the payload that's in the cluster.
1652 		 */
1653 
1654 		MPASS(clm != NULL);
1655 		m = (struct mbuf *)(sd->cl + sd->nmbuf * MSIZE);
1656 		/* No bzero required */
1657 		if (m_init(m, M_NOWAIT, MT_DATA,
1658 		    fr_offset == 0 ? M_PKTHDR | M_NOFREE : M_NOFREE))
1659 			return (NULL);
1660 		fl->mbuf_inlined++;
1661 		m_extaddref(m, payload, blen, &clm->refcount, rxb_free,
1662 		    swz->zone, sd->cl);
1663 		if (sd->nmbuf++ == 0)
1664 			counter_u64_add(extfree_refs, 1);
1665 
1666 	} else {
1667 
1668 		/*
1669 		 * Grab an mbuf from zone_mbuf and associate it with the
1670 		 * payload in the cluster.
1671 		 */
1672 
1673 		m = fr_offset == 0 ?
1674 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1675 		if (m == NULL)
1676 			return (NULL);
1677 		fl->mbuf_allocated++;
1678 		if (clm != NULL) {
1679 			m_extaddref(m, payload, blen, &clm->refcount,
1680 			    rxb_free, swz->zone, sd->cl);
1681 			if (sd->nmbuf++ == 0)
1682 				counter_u64_add(extfree_refs, 1);
1683 		} else {
1684 			m_cljset(m, sd->cl, swz->type);
1685 			sd->cl = NULL;	/* consumed, not a recycle candidate */
1686 		}
1687 	}
1688 	if (fr_offset == 0)
1689 		m->m_pkthdr.len = remaining;
1690 	m->m_len = len;
1691 
1692 	if (fl->flags & FL_BUF_PACKING) {
1693 		fl->rx_offset += blen;
1694 		MPASS(fl->rx_offset <= hwb->size);
1695 		if (fl->rx_offset < hwb->size)
1696 			return (m);	/* without advancing the cidx */
1697 	}
1698 
1699 	if (__predict_false(++fl->cidx % 8 == 0)) {
1700 		uint16_t cidx = fl->cidx / 8;
1701 
1702 		if (__predict_false(cidx == fl->sidx))
1703 			fl->cidx = cidx = 0;
1704 		fl->hw_cidx = cidx;
1705 	}
1706 	fl->rx_offset = 0;
1707 
1708 	return (m);
1709 }
1710 
1711 static struct mbuf *
1712 get_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf)
1713 {
1714 	struct mbuf *m0, *m, **pnext;
1715 	u_int remaining;
1716 	const u_int total = G_RSPD_LEN(len_newbuf);
1717 
1718 	if (__predict_false(fl->flags & FL_BUF_RESUME)) {
1719 		M_ASSERTPKTHDR(fl->m0);
1720 		MPASS(fl->m0->m_pkthdr.len == total);
1721 		MPASS(fl->remaining < total);
1722 
1723 		m0 = fl->m0;
1724 		pnext = fl->pnext;
1725 		remaining = fl->remaining;
1726 		fl->flags &= ~FL_BUF_RESUME;
1727 		goto get_segment;
1728 	}
1729 
1730 	if (fl->rx_offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
1731 		fl->rx_offset = 0;
1732 		if (__predict_false(++fl->cidx % 8 == 0)) {
1733 			uint16_t cidx = fl->cidx / 8;
1734 
1735 			if (__predict_false(cidx == fl->sidx))
1736 				fl->cidx = cidx = 0;
1737 			fl->hw_cidx = cidx;
1738 		}
1739 	}
1740 
1741 	/*
1742 	 * Payload starts at rx_offset in the current hw buffer.  Its length is
1743 	 * 'len' and it may span multiple hw buffers.
1744 	 */
1745 
1746 	m0 = get_scatter_segment(sc, fl, 0, total);
1747 	if (m0 == NULL)
1748 		return (NULL);
1749 	remaining = total - m0->m_len;
1750 	pnext = &m0->m_next;
1751 	while (remaining > 0) {
1752 get_segment:
1753 		MPASS(fl->rx_offset == 0);
1754 		m = get_scatter_segment(sc, fl, total - remaining, remaining);
1755 		if (__predict_false(m == NULL)) {
1756 			fl->m0 = m0;
1757 			fl->pnext = pnext;
1758 			fl->remaining = remaining;
1759 			fl->flags |= FL_BUF_RESUME;
1760 			return (NULL);
1761 		}
1762 		*pnext = m;
1763 		pnext = &m->m_next;
1764 		remaining -= m->m_len;
1765 	}
1766 	*pnext = NULL;
1767 
1768 	M_ASSERTPKTHDR(m0);
1769 	return (m0);
1770 }
1771 
1772 static int
1773 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
1774 {
1775 	struct sge_rxq *rxq = iq_to_rxq(iq);
1776 	struct ifnet *ifp = rxq->ifp;
1777 	struct adapter *sc = iq->adapter;
1778 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
1779 #if defined(INET) || defined(INET6)
1780 	struct lro_ctrl *lro = &rxq->lro;
1781 #endif
1782 	static const int sw_hashtype[4][2] = {
1783 		{M_HASHTYPE_NONE, M_HASHTYPE_NONE},
1784 		{M_HASHTYPE_RSS_IPV4, M_HASHTYPE_RSS_IPV6},
1785 		{M_HASHTYPE_RSS_TCP_IPV4, M_HASHTYPE_RSS_TCP_IPV6},
1786 		{M_HASHTYPE_RSS_UDP_IPV4, M_HASHTYPE_RSS_UDP_IPV6},
1787 	};
1788 
1789 	KASSERT(m0 != NULL, ("%s: no payload with opcode %02x", __func__,
1790 	    rss->opcode));
1791 
1792 	m0->m_pkthdr.len -= sc->params.sge.fl_pktshift;
1793 	m0->m_len -= sc->params.sge.fl_pktshift;
1794 	m0->m_data += sc->params.sge.fl_pktshift;
1795 
1796 	m0->m_pkthdr.rcvif = ifp;
1797 	M_HASHTYPE_SET(m0, sw_hashtype[rss->hash_type][rss->ipv6]);
1798 	m0->m_pkthdr.flowid = be32toh(rss->hash_val);
1799 
1800 	if (cpl->csum_calc && !(cpl->err_vec & sc->params.tp.err_vec_mask)) {
1801 		if (ifp->if_capenable & IFCAP_RXCSUM &&
1802 		    cpl->l2info & htobe32(F_RXF_IP)) {
1803 			m0->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
1804 			    CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
1805 			rxq->rxcsum++;
1806 		} else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
1807 		    cpl->l2info & htobe32(F_RXF_IP6)) {
1808 			m0->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
1809 			    CSUM_PSEUDO_HDR);
1810 			rxq->rxcsum++;
1811 		}
1812 
1813 		if (__predict_false(cpl->ip_frag))
1814 			m0->m_pkthdr.csum_data = be16toh(cpl->csum);
1815 		else
1816 			m0->m_pkthdr.csum_data = 0xffff;
1817 	}
1818 
1819 	if (cpl->vlan_ex) {
1820 		m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
1821 		m0->m_flags |= M_VLANTAG;
1822 		rxq->vlan_extraction++;
1823 	}
1824 
1825 #if defined(INET) || defined(INET6)
1826 	if (iq->flags & IQ_LRO_ENABLED) {
1827 		if (sort_before_lro(lro)) {
1828 			tcp_lro_queue_mbuf(lro, m0);
1829 			return (0); /* queued for sort, then LRO */
1830 		}
1831 		if (tcp_lro_rx(lro, m0, 0) == 0)
1832 			return (0); /* queued for LRO */
1833 	}
1834 #endif
1835 	ifp->if_input(ifp, m0);
1836 
1837 	return (0);
1838 }
1839 
1840 /*
1841  * Must drain the wrq or make sure that someone else will.
1842  */
1843 static void
1844 wrq_tx_drain(void *arg, int n)
1845 {
1846 	struct sge_wrq *wrq = arg;
1847 	struct sge_eq *eq = &wrq->eq;
1848 
1849 	EQ_LOCK(eq);
1850 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
1851 		drain_wrq_wr_list(wrq->adapter, wrq);
1852 	EQ_UNLOCK(eq);
1853 }
1854 
1855 static void
1856 drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq)
1857 {
1858 	struct sge_eq *eq = &wrq->eq;
1859 	u_int available, dbdiff;	/* # of hardware descriptors */
1860 	u_int n;
1861 	struct wrqe *wr;
1862 	struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
1863 
1864 	EQ_LOCK_ASSERT_OWNED(eq);
1865 	MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
1866 	wr = STAILQ_FIRST(&wrq->wr_list);
1867 	MPASS(wr != NULL);	/* Must be called with something useful to do */
1868 	MPASS(eq->pidx == eq->dbidx);
1869 	dbdiff = 0;
1870 
1871 	do {
1872 		eq->cidx = read_hw_cidx(eq);
1873 		if (eq->pidx == eq->cidx)
1874 			available = eq->sidx - 1;
1875 		else
1876 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
1877 
1878 		MPASS(wr->wrq == wrq);
1879 		n = howmany(wr->wr_len, EQ_ESIZE);
1880 		if (available < n)
1881 			break;
1882 
1883 		dst = (void *)&eq->desc[eq->pidx];
1884 		if (__predict_true(eq->sidx - eq->pidx > n)) {
1885 			/* Won't wrap, won't end exactly at the status page. */
1886 			bcopy(&wr->wr[0], dst, wr->wr_len);
1887 			eq->pidx += n;
1888 		} else {
1889 			int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE;
1890 
1891 			bcopy(&wr->wr[0], dst, first_portion);
1892 			if (wr->wr_len > first_portion) {
1893 				bcopy(&wr->wr[first_portion], &eq->desc[0],
1894 				    wr->wr_len - first_portion);
1895 			}
1896 			eq->pidx = n - (eq->sidx - eq->pidx);
1897 		}
1898 		wrq->tx_wrs_copied++;
1899 
1900 		if (available < eq->sidx / 4 &&
1901 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
1902 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
1903 			    F_FW_WR_EQUEQ);
1904 			eq->equeqidx = eq->pidx;
1905 		} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
1906 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
1907 			eq->equeqidx = eq->pidx;
1908 		}
1909 
1910 		dbdiff += n;
1911 		if (dbdiff >= 16) {
1912 			ring_eq_db(sc, eq, dbdiff);
1913 			dbdiff = 0;
1914 		}
1915 
1916 		STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
1917 		free_wrqe(wr);
1918 		MPASS(wrq->nwr_pending > 0);
1919 		wrq->nwr_pending--;
1920 		MPASS(wrq->ndesc_needed >= n);
1921 		wrq->ndesc_needed -= n;
1922 	} while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL);
1923 
1924 	if (dbdiff)
1925 		ring_eq_db(sc, eq, dbdiff);
1926 }
1927 
1928 /*
1929  * Doesn't fail.  Holds on to work requests it can't send right away.
1930  */
1931 void
1932 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
1933 {
1934 #ifdef INVARIANTS
1935 	struct sge_eq *eq = &wrq->eq;
1936 #endif
1937 
1938 	EQ_LOCK_ASSERT_OWNED(eq);
1939 	MPASS(wr != NULL);
1940 	MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN);
1941 	MPASS((wr->wr_len & 0x7) == 0);
1942 
1943 	STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
1944 	wrq->nwr_pending++;
1945 	wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE);
1946 
1947 	if (!TAILQ_EMPTY(&wrq->incomplete_wrs))
1948 		return;	/* commit_wrq_wr will drain wr_list as well. */
1949 
1950 	drain_wrq_wr_list(sc, wrq);
1951 
1952 	/* Doorbell must have caught up to the pidx. */
1953 	MPASS(eq->pidx == eq->dbidx);
1954 }
1955 
1956 void
1957 t4_update_fl_bufsize(struct ifnet *ifp)
1958 {
1959 	struct vi_info *vi = ifp->if_softc;
1960 	struct adapter *sc = vi->pi->adapter;
1961 	struct sge_rxq *rxq;
1962 #ifdef TCP_OFFLOAD
1963 	struct sge_ofld_rxq *ofld_rxq;
1964 #endif
1965 	struct sge_fl *fl;
1966 	int i, maxp, mtu = ifp->if_mtu;
1967 
1968 	maxp = mtu_to_max_payload(sc, mtu, 0);
1969 	for_each_rxq(vi, i, rxq) {
1970 		fl = &rxq->fl;
1971 
1972 		FL_LOCK(fl);
1973 		find_best_refill_source(sc, fl, maxp);
1974 		FL_UNLOCK(fl);
1975 	}
1976 #ifdef TCP_OFFLOAD
1977 	maxp = mtu_to_max_payload(sc, mtu, 1);
1978 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1979 		fl = &ofld_rxq->fl;
1980 
1981 		FL_LOCK(fl);
1982 		find_best_refill_source(sc, fl, maxp);
1983 		FL_UNLOCK(fl);
1984 	}
1985 #endif
1986 }
1987 
1988 static inline int
1989 mbuf_nsegs(struct mbuf *m)
1990 {
1991 
1992 	M_ASSERTPKTHDR(m);
1993 	KASSERT(m->m_pkthdr.l5hlen > 0,
1994 	    ("%s: mbuf %p missing information on # of segments.", __func__, m));
1995 
1996 	return (m->m_pkthdr.l5hlen);
1997 }
1998 
1999 static inline void
2000 set_mbuf_nsegs(struct mbuf *m, uint8_t nsegs)
2001 {
2002 
2003 	M_ASSERTPKTHDR(m);
2004 	m->m_pkthdr.l5hlen = nsegs;
2005 }
2006 
2007 static inline int
2008 mbuf_len16(struct mbuf *m)
2009 {
2010 	int n;
2011 
2012 	M_ASSERTPKTHDR(m);
2013 	n = m->m_pkthdr.PH_loc.eight[0];
2014 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2015 
2016 	return (n);
2017 }
2018 
2019 static inline void
2020 set_mbuf_len16(struct mbuf *m, uint8_t len16)
2021 {
2022 
2023 	M_ASSERTPKTHDR(m);
2024 	m->m_pkthdr.PH_loc.eight[0] = len16;
2025 }
2026 
2027 static inline int
2028 needs_tso(struct mbuf *m)
2029 {
2030 
2031 	M_ASSERTPKTHDR(m);
2032 
2033 	if (m->m_pkthdr.csum_flags & CSUM_TSO) {
2034 		KASSERT(m->m_pkthdr.tso_segsz > 0,
2035 		    ("%s: TSO requested in mbuf %p but MSS not provided",
2036 		    __func__, m));
2037 		return (1);
2038 	}
2039 
2040 	return (0);
2041 }
2042 
2043 static inline int
2044 needs_l3_csum(struct mbuf *m)
2045 {
2046 
2047 	M_ASSERTPKTHDR(m);
2048 
2049 	if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO))
2050 		return (1);
2051 	return (0);
2052 }
2053 
2054 static inline int
2055 needs_l4_csum(struct mbuf *m)
2056 {
2057 
2058 	M_ASSERTPKTHDR(m);
2059 
2060 	if (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
2061 	    CSUM_TCP_IPV6 | CSUM_TSO))
2062 		return (1);
2063 	return (0);
2064 }
2065 
2066 static inline int
2067 needs_vlan_insertion(struct mbuf *m)
2068 {
2069 
2070 	M_ASSERTPKTHDR(m);
2071 
2072 	if (m->m_flags & M_VLANTAG) {
2073 		KASSERT(m->m_pkthdr.ether_vtag != 0,
2074 		    ("%s: HWVLAN requested in mbuf %p but tag not provided",
2075 		    __func__, m));
2076 		return (1);
2077 	}
2078 	return (0);
2079 }
2080 
2081 static void *
2082 m_advance(struct mbuf **pm, int *poffset, int len)
2083 {
2084 	struct mbuf *m = *pm;
2085 	int offset = *poffset;
2086 	uintptr_t p = 0;
2087 
2088 	MPASS(len > 0);
2089 
2090 	for (;;) {
2091 		if (offset + len < m->m_len) {
2092 			offset += len;
2093 			p = mtod(m, uintptr_t) + offset;
2094 			break;
2095 		}
2096 		len -= m->m_len - offset;
2097 		m = m->m_next;
2098 		offset = 0;
2099 		MPASS(m != NULL);
2100 	}
2101 	*poffset = offset;
2102 	*pm = m;
2103 	return ((void *)p);
2104 }
2105 
2106 /*
2107  * Can deal with empty mbufs in the chain that have m_len = 0, but the chain
2108  * must have at least one mbuf that's not empty.
2109  */
2110 static inline int
2111 count_mbuf_nsegs(struct mbuf *m)
2112 {
2113 	vm_paddr_t lastb, next;
2114 	vm_offset_t va;
2115 	int len, nsegs;
2116 
2117 	MPASS(m != NULL);
2118 
2119 	nsegs = 0;
2120 	lastb = 0;
2121 	for (; m; m = m->m_next) {
2122 
2123 		len = m->m_len;
2124 		if (__predict_false(len == 0))
2125 			continue;
2126 		va = mtod(m, vm_offset_t);
2127 		next = pmap_kextract(va);
2128 		nsegs += sglist_count(m->m_data, len);
2129 		if (lastb + 1 == next)
2130 			nsegs--;
2131 		lastb = pmap_kextract(va + len - 1);
2132 	}
2133 
2134 	MPASS(nsegs > 0);
2135 	return (nsegs);
2136 }
2137 
2138 /*
2139  * Analyze the mbuf to determine its tx needs.  The mbuf passed in may change:
2140  * a) caller can assume it's been freed if this function returns with an error.
2141  * b) it may get defragged up if the gather list is too long for the hardware.
2142  */
2143 int
2144 parse_pkt(struct adapter *sc, struct mbuf **mp)
2145 {
2146 	struct mbuf *m0 = *mp, *m;
2147 	int rc, nsegs, defragged = 0, offset;
2148 	struct ether_header *eh;
2149 	void *l3hdr;
2150 #if defined(INET) || defined(INET6)
2151 	struct tcphdr *tcp;
2152 #endif
2153 	uint16_t eh_type;
2154 
2155 	M_ASSERTPKTHDR(m0);
2156 	if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) {
2157 		rc = EINVAL;
2158 fail:
2159 		m_freem(m0);
2160 		*mp = NULL;
2161 		return (rc);
2162 	}
2163 restart:
2164 	/*
2165 	 * First count the number of gather list segments in the payload.
2166 	 * Defrag the mbuf if nsegs exceeds the hardware limit.
2167 	 */
2168 	M_ASSERTPKTHDR(m0);
2169 	MPASS(m0->m_pkthdr.len > 0);
2170 	nsegs = count_mbuf_nsegs(m0);
2171 	if (nsegs > (needs_tso(m0) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS)) {
2172 		if (defragged++ > 0 || (m = m_defrag(m0, M_NOWAIT)) == NULL) {
2173 			rc = EFBIG;
2174 			goto fail;
2175 		}
2176 		*mp = m0 = m;	/* update caller's copy after defrag */
2177 		goto restart;
2178 	}
2179 
2180 	if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN)) {
2181 		m0 = m_pullup(m0, m0->m_pkthdr.len);
2182 		if (m0 == NULL) {
2183 			/* Should have left well enough alone. */
2184 			rc = EFBIG;
2185 			goto fail;
2186 		}
2187 		*mp = m0;	/* update caller's copy after pullup */
2188 		goto restart;
2189 	}
2190 	set_mbuf_nsegs(m0, nsegs);
2191 	if (sc->flags & IS_VF)
2192 		set_mbuf_len16(m0, txpkt_vm_len16(nsegs, needs_tso(m0)));
2193 	else
2194 		set_mbuf_len16(m0, txpkt_len16(nsegs, needs_tso(m0)));
2195 
2196 	if (!needs_tso(m0) &&
2197 	    !(sc->flags & IS_VF && (needs_l3_csum(m0) || needs_l4_csum(m0))))
2198 		return (0);
2199 
2200 	m = m0;
2201 	eh = mtod(m, struct ether_header *);
2202 	eh_type = ntohs(eh->ether_type);
2203 	if (eh_type == ETHERTYPE_VLAN) {
2204 		struct ether_vlan_header *evh = (void *)eh;
2205 
2206 		eh_type = ntohs(evh->evl_proto);
2207 		m0->m_pkthdr.l2hlen = sizeof(*evh);
2208 	} else
2209 		m0->m_pkthdr.l2hlen = sizeof(*eh);
2210 
2211 	offset = 0;
2212 	l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2213 
2214 	switch (eh_type) {
2215 #ifdef INET6
2216 	case ETHERTYPE_IPV6:
2217 	{
2218 		struct ip6_hdr *ip6 = l3hdr;
2219 
2220 		MPASS(!needs_tso(m0) || ip6->ip6_nxt == IPPROTO_TCP);
2221 
2222 		m0->m_pkthdr.l3hlen = sizeof(*ip6);
2223 		break;
2224 	}
2225 #endif
2226 #ifdef INET
2227 	case ETHERTYPE_IP:
2228 	{
2229 		struct ip *ip = l3hdr;
2230 
2231 		m0->m_pkthdr.l3hlen = ip->ip_hl * 4;
2232 		break;
2233 	}
2234 #endif
2235 	default:
2236 		panic("%s: ethertype 0x%04x unknown.  if_cxgbe must be compiled"
2237 		    " with the same INET/INET6 options as the kernel.",
2238 		    __func__, eh_type);
2239 	}
2240 
2241 #if defined(INET) || defined(INET6)
2242 	if (needs_tso(m0)) {
2243 		tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen);
2244 		m0->m_pkthdr.l4hlen = tcp->th_off * 4;
2245 	}
2246 #endif
2247 	MPASS(m0 == *mp);
2248 	return (0);
2249 }
2250 
2251 void *
2252 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie)
2253 {
2254 	struct sge_eq *eq = &wrq->eq;
2255 	struct adapter *sc = wrq->adapter;
2256 	int ndesc, available;
2257 	struct wrqe *wr;
2258 	void *w;
2259 
2260 	MPASS(len16 > 0);
2261 	ndesc = howmany(len16, EQ_ESIZE / 16);
2262 	MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC);
2263 
2264 	EQ_LOCK(eq);
2265 
2266 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2267 		drain_wrq_wr_list(sc, wrq);
2268 
2269 	if (!STAILQ_EMPTY(&wrq->wr_list)) {
2270 slowpath:
2271 		EQ_UNLOCK(eq);
2272 		wr = alloc_wrqe(len16 * 16, wrq);
2273 		if (__predict_false(wr == NULL))
2274 			return (NULL);
2275 		cookie->pidx = -1;
2276 		cookie->ndesc = ndesc;
2277 		return (&wr->wr);
2278 	}
2279 
2280 	eq->cidx = read_hw_cidx(eq);
2281 	if (eq->pidx == eq->cidx)
2282 		available = eq->sidx - 1;
2283 	else
2284 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2285 	if (available < ndesc)
2286 		goto slowpath;
2287 
2288 	cookie->pidx = eq->pidx;
2289 	cookie->ndesc = ndesc;
2290 	TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link);
2291 
2292 	w = &eq->desc[eq->pidx];
2293 	IDXINCR(eq->pidx, ndesc, eq->sidx);
2294 	if (__predict_false(cookie->pidx + ndesc > eq->sidx)) {
2295 		w = &wrq->ss[0];
2296 		wrq->ss_pidx = cookie->pidx;
2297 		wrq->ss_len = len16 * 16;
2298 	}
2299 
2300 	EQ_UNLOCK(eq);
2301 
2302 	return (w);
2303 }
2304 
2305 void
2306 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie)
2307 {
2308 	struct sge_eq *eq = &wrq->eq;
2309 	struct adapter *sc = wrq->adapter;
2310 	int ndesc, pidx;
2311 	struct wrq_cookie *prev, *next;
2312 
2313 	if (cookie->pidx == -1) {
2314 		struct wrqe *wr = __containerof(w, struct wrqe, wr);
2315 
2316 		t4_wrq_tx(sc, wr);
2317 		return;
2318 	}
2319 
2320 	if (__predict_false(w == &wrq->ss[0])) {
2321 		int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE;
2322 
2323 		MPASS(wrq->ss_len > n);	/* WR had better wrap around. */
2324 		bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n);
2325 		bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n);
2326 		wrq->tx_wrs_ss++;
2327 	} else
2328 		wrq->tx_wrs_direct++;
2329 
2330 	EQ_LOCK(eq);
2331 	ndesc = cookie->ndesc;	/* Can be more than SGE_MAX_WR_NDESC here. */
2332 	pidx = cookie->pidx;
2333 	MPASS(pidx >= 0 && pidx < eq->sidx);
2334 	prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link);
2335 	next = TAILQ_NEXT(cookie, link);
2336 	if (prev == NULL) {
2337 		MPASS(pidx == eq->dbidx);
2338 		if (next == NULL || ndesc >= 16)
2339 			ring_eq_db(wrq->adapter, eq, ndesc);
2340 		else {
2341 			MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc);
2342 			next->pidx = pidx;
2343 			next->ndesc += ndesc;
2344 		}
2345 	} else {
2346 		MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc);
2347 		prev->ndesc += ndesc;
2348 	}
2349 	TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link);
2350 
2351 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2352 		drain_wrq_wr_list(sc, wrq);
2353 
2354 #ifdef INVARIANTS
2355 	if (TAILQ_EMPTY(&wrq->incomplete_wrs)) {
2356 		/* Doorbell must have caught up to the pidx. */
2357 		MPASS(wrq->eq.pidx == wrq->eq.dbidx);
2358 	}
2359 #endif
2360 	EQ_UNLOCK(eq);
2361 }
2362 
2363 static u_int
2364 can_resume_eth_tx(struct mp_ring *r)
2365 {
2366 	struct sge_eq *eq = r->cookie;
2367 
2368 	return (total_available_tx_desc(eq) > eq->sidx / 8);
2369 }
2370 
2371 static inline int
2372 cannot_use_txpkts(struct mbuf *m)
2373 {
2374 	/* maybe put a GL limit too, to avoid silliness? */
2375 
2376 	return (needs_tso(m));
2377 }
2378 
2379 static inline int
2380 discard_tx(struct sge_eq *eq)
2381 {
2382 
2383 	return ((eq->flags & (EQ_ENABLED | EQ_QFLUSH)) != EQ_ENABLED);
2384 }
2385 
2386 /*
2387  * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to
2388  * be consumed.  Return the actual number consumed.  0 indicates a stall.
2389  */
2390 static u_int
2391 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx)
2392 {
2393 	struct sge_txq *txq = r->cookie;
2394 	struct sge_eq *eq = &txq->eq;
2395 	struct ifnet *ifp = txq->ifp;
2396 	struct vi_info *vi = ifp->if_softc;
2397 	struct port_info *pi = vi->pi;
2398 	struct adapter *sc = pi->adapter;
2399 	u_int total, remaining;		/* # of packets */
2400 	u_int available, dbdiff;	/* # of hardware descriptors */
2401 	u_int n, next_cidx;
2402 	struct mbuf *m0, *tail;
2403 	struct txpkts txp;
2404 	struct fw_eth_tx_pkts_wr *wr;	/* any fw WR struct will do */
2405 
2406 	remaining = IDXDIFF(pidx, cidx, r->size);
2407 	MPASS(remaining > 0);	/* Must not be called without work to do. */
2408 	total = 0;
2409 
2410 	TXQ_LOCK(txq);
2411 	if (__predict_false(discard_tx(eq))) {
2412 		while (cidx != pidx) {
2413 			m0 = r->items[cidx];
2414 			m_freem(m0);
2415 			if (++cidx == r->size)
2416 				cidx = 0;
2417 		}
2418 		reclaim_tx_descs(txq, 2048);
2419 		total = remaining;
2420 		goto done;
2421 	}
2422 
2423 	/* How many hardware descriptors do we have readily available. */
2424 	if (eq->pidx == eq->cidx)
2425 		available = eq->sidx - 1;
2426 	else
2427 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2428 	dbdiff = IDXDIFF(eq->pidx, eq->dbidx, eq->sidx);
2429 
2430 	while (remaining > 0) {
2431 
2432 		m0 = r->items[cidx];
2433 		M_ASSERTPKTHDR(m0);
2434 		MPASS(m0->m_nextpkt == NULL);
2435 
2436 		if (available < SGE_MAX_WR_NDESC) {
2437 			available += reclaim_tx_descs(txq, 64);
2438 			if (available < howmany(mbuf_len16(m0), EQ_ESIZE / 16))
2439 				break;	/* out of descriptors */
2440 		}
2441 
2442 		next_cidx = cidx + 1;
2443 		if (__predict_false(next_cidx == r->size))
2444 			next_cidx = 0;
2445 
2446 		wr = (void *)&eq->desc[eq->pidx];
2447 		if (sc->flags & IS_VF) {
2448 			total++;
2449 			remaining--;
2450 			ETHER_BPF_MTAP(ifp, m0);
2451 			n = write_txpkt_vm_wr(sc, txq, (void *)wr, m0,
2452 			    available);
2453 		} else if (remaining > 1 &&
2454 		    try_txpkts(m0, r->items[next_cidx], &txp, available) == 0) {
2455 
2456 			/* pkts at cidx, next_cidx should both be in txp. */
2457 			MPASS(txp.npkt == 2);
2458 			tail = r->items[next_cidx];
2459 			MPASS(tail->m_nextpkt == NULL);
2460 			ETHER_BPF_MTAP(ifp, m0);
2461 			ETHER_BPF_MTAP(ifp, tail);
2462 			m0->m_nextpkt = tail;
2463 
2464 			if (__predict_false(++next_cidx == r->size))
2465 				next_cidx = 0;
2466 
2467 			while (next_cidx != pidx) {
2468 				if (add_to_txpkts(r->items[next_cidx], &txp,
2469 				    available) != 0)
2470 					break;
2471 				tail->m_nextpkt = r->items[next_cidx];
2472 				tail = tail->m_nextpkt;
2473 				ETHER_BPF_MTAP(ifp, tail);
2474 				if (__predict_false(++next_cidx == r->size))
2475 					next_cidx = 0;
2476 			}
2477 
2478 			n = write_txpkts_wr(txq, wr, m0, &txp, available);
2479 			total += txp.npkt;
2480 			remaining -= txp.npkt;
2481 		} else {
2482 			total++;
2483 			remaining--;
2484 			ETHER_BPF_MTAP(ifp, m0);
2485 			n = write_txpkt_wr(txq, (void *)wr, m0, available);
2486 		}
2487 		MPASS(n >= 1 && n <= available && n <= SGE_MAX_WR_NDESC);
2488 
2489 		available -= n;
2490 		dbdiff += n;
2491 		IDXINCR(eq->pidx, n, eq->sidx);
2492 
2493 		if (total_available_tx_desc(eq) < eq->sidx / 4 &&
2494 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2495 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2496 			    F_FW_WR_EQUEQ);
2497 			eq->equeqidx = eq->pidx;
2498 		} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
2499 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
2500 			eq->equeqidx = eq->pidx;
2501 		}
2502 
2503 		if (dbdiff >= 16 && remaining >= 4) {
2504 			ring_eq_db(sc, eq, dbdiff);
2505 			available += reclaim_tx_descs(txq, 4 * dbdiff);
2506 			dbdiff = 0;
2507 		}
2508 
2509 		cidx = next_cidx;
2510 	}
2511 	if (dbdiff != 0) {
2512 		ring_eq_db(sc, eq, dbdiff);
2513 		reclaim_tx_descs(txq, 32);
2514 	}
2515 done:
2516 	TXQ_UNLOCK(txq);
2517 
2518 	return (total);
2519 }
2520 
2521 static inline void
2522 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
2523     int qsize)
2524 {
2525 
2526 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
2527 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
2528 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
2529 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
2530 
2531 	iq->flags = 0;
2532 	iq->adapter = sc;
2533 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
2534 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
2535 	if (pktc_idx >= 0) {
2536 		iq->intr_params |= F_QINTR_CNT_EN;
2537 		iq->intr_pktc_idx = pktc_idx;
2538 	}
2539 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
2540 	iq->sidx = iq->qsize - sc->params.sge.spg_len / IQ_ESIZE;
2541 }
2542 
2543 static inline void
2544 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name)
2545 {
2546 
2547 	fl->qsize = qsize;
2548 	fl->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2549 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
2550 	if (sc->flags & BUF_PACKING_OK &&
2551 	    ((!is_t4(sc) && buffer_packing) ||	/* T5+: enabled unless 0 */
2552 	    (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */
2553 		fl->flags |= FL_BUF_PACKING;
2554 	find_best_refill_source(sc, fl, maxp);
2555 	find_safe_refill_source(sc, fl);
2556 }
2557 
2558 static inline void
2559 init_eq(struct adapter *sc, struct sge_eq *eq, int eqtype, int qsize,
2560     uint8_t tx_chan, uint16_t iqid, char *name)
2561 {
2562 	KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
2563 
2564 	eq->flags = eqtype & EQ_TYPEMASK;
2565 	eq->tx_chan = tx_chan;
2566 	eq->iqid = iqid;
2567 	eq->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2568 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
2569 }
2570 
2571 static int
2572 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
2573     bus_dmamap_t *map, bus_addr_t *pa, void **va)
2574 {
2575 	int rc;
2576 
2577 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
2578 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
2579 	if (rc != 0) {
2580 		device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
2581 		goto done;
2582 	}
2583 
2584 	rc = bus_dmamem_alloc(*tag, va,
2585 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
2586 	if (rc != 0) {
2587 		device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
2588 		goto done;
2589 	}
2590 
2591 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
2592 	if (rc != 0) {
2593 		device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
2594 		goto done;
2595 	}
2596 done:
2597 	if (rc)
2598 		free_ring(sc, *tag, *map, *pa, *va);
2599 
2600 	return (rc);
2601 }
2602 
2603 static int
2604 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
2605     bus_addr_t pa, void *va)
2606 {
2607 	if (pa)
2608 		bus_dmamap_unload(tag, map);
2609 	if (va)
2610 		bus_dmamem_free(tag, va, map);
2611 	if (tag)
2612 		bus_dma_tag_destroy(tag);
2613 
2614 	return (0);
2615 }
2616 
2617 /*
2618  * Allocates the ring for an ingress queue and an optional freelist.  If the
2619  * freelist is specified it will be allocated and then associated with the
2620  * ingress queue.
2621  *
2622  * Returns errno on failure.  Resources allocated up to that point may still be
2623  * allocated.  Caller is responsible for cleanup in case this function fails.
2624  *
2625  * If the ingress queue will take interrupts directly then the intr_idx
2626  * specifies the vector, starting from 0.  -1 means the interrupts for this
2627  * queue should be forwarded to the fwq.
2628  */
2629 static int
2630 alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl,
2631     int intr_idx, int cong)
2632 {
2633 	int rc, i, cntxt_id;
2634 	size_t len;
2635 	struct fw_iq_cmd c;
2636 	struct port_info *pi = vi->pi;
2637 	struct adapter *sc = iq->adapter;
2638 	struct sge_params *sp = &sc->params.sge;
2639 	__be32 v = 0;
2640 
2641 	len = iq->qsize * IQ_ESIZE;
2642 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
2643 	    (void **)&iq->desc);
2644 	if (rc != 0)
2645 		return (rc);
2646 
2647 	bzero(&c, sizeof(c));
2648 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
2649 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
2650 	    V_FW_IQ_CMD_VFN(0));
2651 
2652 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
2653 	    FW_LEN16(c));
2654 
2655 	/* Special handling for firmware event queue */
2656 	if (iq == &sc->sge.fwq)
2657 		v |= F_FW_IQ_CMD_IQASYNCH;
2658 
2659 	if (intr_idx < 0) {
2660 		/* Forwarded interrupts, all headed to fwq */
2661 		v |= F_FW_IQ_CMD_IQANDST;
2662 		v |= V_FW_IQ_CMD_IQANDSTINDEX(sc->sge.fwq.cntxt_id);
2663 	} else {
2664 		KASSERT(intr_idx < sc->intr_count,
2665 		    ("%s: invalid direct intr_idx %d", __func__, intr_idx));
2666 		v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
2667 	}
2668 
2669 	c.type_to_iqandstindex = htobe32(v |
2670 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
2671 	    V_FW_IQ_CMD_VIID(vi->viid) |
2672 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
2673 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
2674 	    F_FW_IQ_CMD_IQGTSMODE |
2675 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
2676 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
2677 	c.iqsize = htobe16(iq->qsize);
2678 	c.iqaddr = htobe64(iq->ba);
2679 	if (cong >= 0)
2680 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
2681 
2682 	if (fl) {
2683 		mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
2684 
2685 		len = fl->qsize * EQ_ESIZE;
2686 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
2687 		    &fl->ba, (void **)&fl->desc);
2688 		if (rc)
2689 			return (rc);
2690 
2691 		/* Allocate space for one software descriptor per buffer. */
2692 		rc = alloc_fl_sdesc(fl);
2693 		if (rc != 0) {
2694 			device_printf(sc->dev,
2695 			    "failed to setup fl software descriptors: %d\n",
2696 			    rc);
2697 			return (rc);
2698 		}
2699 
2700 		if (fl->flags & FL_BUF_PACKING) {
2701 			fl->lowat = roundup2(sp->fl_starve_threshold2, 8);
2702 			fl->buf_boundary = sp->pack_boundary;
2703 		} else {
2704 			fl->lowat = roundup2(sp->fl_starve_threshold, 8);
2705 			fl->buf_boundary = 16;
2706 		}
2707 		if (fl_pad && fl->buf_boundary < sp->pad_boundary)
2708 			fl->buf_boundary = sp->pad_boundary;
2709 
2710 		c.iqns_to_fl0congen |=
2711 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
2712 			F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
2713 			(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
2714 			(fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
2715 			    0));
2716 		if (cong >= 0) {
2717 			c.iqns_to_fl0congen |=
2718 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
2719 				    F_FW_IQ_CMD_FL0CONGCIF |
2720 				    F_FW_IQ_CMD_FL0CONGEN);
2721 		}
2722 		c.fl0dcaen_to_fl0cidxfthresh =
2723 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
2724 			X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B) |
2725 			V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
2726 			X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
2727 		c.fl0size = htobe16(fl->qsize);
2728 		c.fl0addr = htobe64(fl->ba);
2729 	}
2730 
2731 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
2732 	if (rc != 0) {
2733 		device_printf(sc->dev,
2734 		    "failed to create ingress queue: %d\n", rc);
2735 		return (rc);
2736 	}
2737 
2738 	iq->cidx = 0;
2739 	iq->gen = F_RSPD_GEN;
2740 	iq->intr_next = iq->intr_params;
2741 	iq->cntxt_id = be16toh(c.iqid);
2742 	iq->abs_id = be16toh(c.physiqid);
2743 	iq->flags |= IQ_ALLOCATED;
2744 
2745 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
2746 	if (cntxt_id >= sc->sge.niq) {
2747 		panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
2748 		    cntxt_id, sc->sge.niq - 1);
2749 	}
2750 	sc->sge.iqmap[cntxt_id] = iq;
2751 
2752 	if (fl) {
2753 		u_int qid;
2754 
2755 		iq->flags |= IQ_HAS_FL;
2756 		fl->cntxt_id = be16toh(c.fl0id);
2757 		fl->pidx = fl->cidx = 0;
2758 
2759 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
2760 		if (cntxt_id >= sc->sge.neq) {
2761 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
2762 			    __func__, cntxt_id, sc->sge.neq - 1);
2763 		}
2764 		sc->sge.eqmap[cntxt_id] = (void *)fl;
2765 
2766 		qid = fl->cntxt_id;
2767 		if (isset(&sc->doorbells, DOORBELL_UDB)) {
2768 			uint32_t s_qpp = sc->params.sge.eq_s_qpp;
2769 			uint32_t mask = (1 << s_qpp) - 1;
2770 			volatile uint8_t *udb;
2771 
2772 			udb = sc->udbs_base + UDBS_DB_OFFSET;
2773 			udb += (qid >> s_qpp) << PAGE_SHIFT;
2774 			qid &= mask;
2775 			if (qid < PAGE_SIZE / UDBS_SEG_SIZE) {
2776 				udb += qid << UDBS_SEG_SHIFT;
2777 				qid = 0;
2778 			}
2779 			fl->udb = (volatile void *)udb;
2780 		}
2781 		fl->dbval = V_QID(qid) | sc->chip_params->sge_fl_db;
2782 
2783 		FL_LOCK(fl);
2784 		/* Enough to make sure the SGE doesn't think it's starved */
2785 		refill_fl(sc, fl, fl->lowat);
2786 		FL_UNLOCK(fl);
2787 	}
2788 
2789 	if (chip_id(sc) >= CHELSIO_T5 && !(sc->flags & IS_VF) && cong >= 0) {
2790 		uint32_t param, val;
2791 
2792 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
2793 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
2794 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
2795 		if (cong == 0)
2796 			val = 1 << 19;
2797 		else {
2798 			val = 2 << 19;
2799 			for (i = 0; i < 4; i++) {
2800 				if (cong & (1 << i))
2801 					val |= 1 << (i << 2);
2802 			}
2803 		}
2804 
2805 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
2806 		if (rc != 0) {
2807 			/* report error but carry on */
2808 			device_printf(sc->dev,
2809 			    "failed to set congestion manager context for "
2810 			    "ingress queue %d: %d\n", iq->cntxt_id, rc);
2811 		}
2812 	}
2813 
2814 	/* Enable IQ interrupts */
2815 	atomic_store_rel_int(&iq->state, IQS_IDLE);
2816 	t4_write_reg(sc, sc->sge_gts_reg, V_SEINTARM(iq->intr_params) |
2817 	    V_INGRESSQID(iq->cntxt_id));
2818 
2819 	return (0);
2820 }
2821 
2822 static int
2823 free_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl)
2824 {
2825 	int rc;
2826 	struct adapter *sc = iq->adapter;
2827 	device_t dev;
2828 
2829 	if (sc == NULL)
2830 		return (0);	/* nothing to do */
2831 
2832 	dev = vi ? vi->dev : sc->dev;
2833 
2834 	if (iq->flags & IQ_ALLOCATED) {
2835 		rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
2836 		    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
2837 		    fl ? fl->cntxt_id : 0xffff, 0xffff);
2838 		if (rc != 0) {
2839 			device_printf(dev,
2840 			    "failed to free queue %p: %d\n", iq, rc);
2841 			return (rc);
2842 		}
2843 		iq->flags &= ~IQ_ALLOCATED;
2844 	}
2845 
2846 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
2847 
2848 	bzero(iq, sizeof(*iq));
2849 
2850 	if (fl) {
2851 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
2852 		    fl->desc);
2853 
2854 		if (fl->sdesc)
2855 			free_fl_sdesc(sc, fl);
2856 
2857 		if (mtx_initialized(&fl->fl_lock))
2858 			mtx_destroy(&fl->fl_lock);
2859 
2860 		bzero(fl, sizeof(*fl));
2861 	}
2862 
2863 	return (0);
2864 }
2865 
2866 static void
2867 add_iq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
2868     struct sge_iq *iq)
2869 {
2870 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2871 
2872 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD, &iq->ba,
2873 	    "bus address of descriptor ring");
2874 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
2875 	    iq->qsize * IQ_ESIZE, "descriptor ring size in bytes");
2876 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
2877 	    CTLTYPE_INT | CTLFLAG_RD, &iq->abs_id, 0, sysctl_uint16, "I",
2878 	    "absolute id of the queue");
2879 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
2880 	    CTLTYPE_INT | CTLFLAG_RD, &iq->cntxt_id, 0, sysctl_uint16, "I",
2881 	    "SGE context id of the queue");
2882 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
2883 	    CTLTYPE_INT | CTLFLAG_RD, &iq->cidx, 0, sysctl_uint16, "I",
2884 	    "consumer index");
2885 }
2886 
2887 static void
2888 add_fl_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
2889     struct sysctl_oid *oid, struct sge_fl *fl)
2890 {
2891 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2892 
2893 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
2894 	    "freelist");
2895 	children = SYSCTL_CHILDREN(oid);
2896 
2897 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
2898 	    &fl->ba, "bus address of descriptor ring");
2899 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
2900 	    fl->sidx * EQ_ESIZE + sc->params.sge.spg_len,
2901 	    "desc ring size in bytes");
2902 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
2903 	    CTLTYPE_INT | CTLFLAG_RD, &fl->cntxt_id, 0, sysctl_uint16, "I",
2904 	    "SGE context id of the freelist");
2905 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL,
2906 	    fl_pad ? 1 : 0, "padding enabled");
2907 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL,
2908 	    fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled");
2909 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
2910 	    0, "consumer index");
2911 	if (fl->flags & FL_BUF_PACKING) {
2912 		SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
2913 		    CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
2914 	}
2915 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
2916 	    0, "producer index");
2917 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_allocated",
2918 	    CTLFLAG_RD, &fl->mbuf_allocated, "# of mbuf allocated");
2919 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_inlined",
2920 	    CTLFLAG_RD, &fl->mbuf_inlined, "# of mbuf inlined in clusters");
2921 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
2922 	    CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
2923 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
2924 	    CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
2925 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
2926 	    CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
2927 }
2928 
2929 static int
2930 alloc_fwq(struct adapter *sc)
2931 {
2932 	int rc, intr_idx;
2933 	struct sge_iq *fwq = &sc->sge.fwq;
2934 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
2935 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2936 
2937 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE);
2938 	if (sc->flags & IS_VF)
2939 		intr_idx = 0;
2940 	else
2941 		intr_idx = sc->intr_count > 1 ? 1 : 0;
2942 	rc = alloc_iq_fl(&sc->port[0]->vi[0], fwq, NULL, intr_idx, -1);
2943 	if (rc != 0) {
2944 		device_printf(sc->dev,
2945 		    "failed to create firmware event queue: %d\n", rc);
2946 		return (rc);
2947 	}
2948 
2949 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD,
2950 	    NULL, "firmware event queue");
2951 	add_iq_sysctls(&sc->ctx, oid, fwq);
2952 
2953 	return (0);
2954 }
2955 
2956 static int
2957 free_fwq(struct adapter *sc)
2958 {
2959 	return free_iq_fl(NULL, &sc->sge.fwq, NULL);
2960 }
2961 
2962 static int
2963 alloc_mgmtq(struct adapter *sc)
2964 {
2965 	int rc;
2966 	struct sge_wrq *mgmtq = &sc->sge.mgmtq;
2967 	char name[16];
2968 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
2969 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2970 
2971 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "mgmtq", CTLFLAG_RD,
2972 	    NULL, "management queue");
2973 
2974 	snprintf(name, sizeof(name), "%s mgmtq", device_get_nameunit(sc->dev));
2975 	init_eq(sc, &mgmtq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[0]->tx_chan,
2976 	    sc->sge.fwq.cntxt_id, name);
2977 	rc = alloc_wrq(sc, NULL, mgmtq, oid);
2978 	if (rc != 0) {
2979 		device_printf(sc->dev,
2980 		    "failed to create management queue: %d\n", rc);
2981 		return (rc);
2982 	}
2983 
2984 	return (0);
2985 }
2986 
2987 static int
2988 free_mgmtq(struct adapter *sc)
2989 {
2990 
2991 	return free_wrq(sc, &sc->sge.mgmtq);
2992 }
2993 
2994 int
2995 tnl_cong(struct port_info *pi, int drop)
2996 {
2997 
2998 	if (drop == -1)
2999 		return (-1);
3000 	else if (drop == 1)
3001 		return (0);
3002 	else
3003 		return (pi->rx_e_chan_map);
3004 }
3005 
3006 static int
3007 alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int intr_idx, int idx,
3008     struct sysctl_oid *oid)
3009 {
3010 	int rc;
3011 	struct adapter *sc = vi->pi->adapter;
3012 	struct sysctl_oid_list *children;
3013 	char name[16];
3014 
3015 	rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, intr_idx,
3016 	    tnl_cong(vi->pi, cong_drop));
3017 	if (rc != 0)
3018 		return (rc);
3019 
3020 	if (idx == 0)
3021 		sc->sge.iq_base = rxq->iq.abs_id - rxq->iq.cntxt_id;
3022 	else
3023 		KASSERT(rxq->iq.cntxt_id + sc->sge.iq_base == rxq->iq.abs_id,
3024 		    ("iq_base mismatch"));
3025 	KASSERT(sc->sge.iq_base == 0 || sc->flags & IS_VF,
3026 	    ("PF with non-zero iq_base"));
3027 
3028 	/*
3029 	 * The freelist is just barely above the starvation threshold right now,
3030 	 * fill it up a bit more.
3031 	 */
3032 	FL_LOCK(&rxq->fl);
3033 	refill_fl(sc, &rxq->fl, 128);
3034 	FL_UNLOCK(&rxq->fl);
3035 
3036 #if defined(INET) || defined(INET6)
3037 	rc = tcp_lro_init_args(&rxq->lro, vi->ifp, lro_entries, lro_mbufs);
3038 	if (rc != 0)
3039 		return (rc);
3040 	MPASS(rxq->lro.ifp == vi->ifp);	/* also indicates LRO init'ed */
3041 
3042 	if (vi->ifp->if_capenable & IFCAP_LRO)
3043 		rxq->iq.flags |= IQ_LRO_ENABLED;
3044 #endif
3045 	rxq->ifp = vi->ifp;
3046 
3047 	children = SYSCTL_CHILDREN(oid);
3048 
3049 	snprintf(name, sizeof(name), "%d", idx);
3050 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3051 	    NULL, "rx queue");
3052 	children = SYSCTL_CHILDREN(oid);
3053 
3054 	add_iq_sysctls(&vi->ctx, oid, &rxq->iq);
3055 #if defined(INET) || defined(INET6)
3056 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
3057 	    &rxq->lro.lro_queued, 0, NULL);
3058 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
3059 	    &rxq->lro.lro_flushed, 0, NULL);
3060 #endif
3061 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
3062 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
3063 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_extraction",
3064 	    CTLFLAG_RD, &rxq->vlan_extraction,
3065 	    "# of times hardware extracted 802.1Q tag");
3066 
3067 	add_fl_sysctls(sc, &vi->ctx, oid, &rxq->fl);
3068 
3069 	return (rc);
3070 }
3071 
3072 static int
3073 free_rxq(struct vi_info *vi, struct sge_rxq *rxq)
3074 {
3075 	int rc;
3076 
3077 #if defined(INET) || defined(INET6)
3078 	if (rxq->lro.ifp) {
3079 		tcp_lro_free(&rxq->lro);
3080 		rxq->lro.ifp = NULL;
3081 	}
3082 #endif
3083 
3084 	rc = free_iq_fl(vi, &rxq->iq, &rxq->fl);
3085 	if (rc == 0)
3086 		bzero(rxq, sizeof(*rxq));
3087 
3088 	return (rc);
3089 }
3090 
3091 #ifdef TCP_OFFLOAD
3092 static int
3093 alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq,
3094     int intr_idx, int idx, struct sysctl_oid *oid)
3095 {
3096 	struct port_info *pi = vi->pi;
3097 	int rc;
3098 	struct sysctl_oid_list *children;
3099 	char name[16];
3100 
3101 	rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx, 0);
3102 	if (rc != 0)
3103 		return (rc);
3104 
3105 	children = SYSCTL_CHILDREN(oid);
3106 
3107 	snprintf(name, sizeof(name), "%d", idx);
3108 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3109 	    NULL, "rx queue");
3110 	add_iq_sysctls(&vi->ctx, oid, &ofld_rxq->iq);
3111 	add_fl_sysctls(pi->adapter, &vi->ctx, oid, &ofld_rxq->fl);
3112 
3113 	return (rc);
3114 }
3115 
3116 static int
3117 free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq)
3118 {
3119 	int rc;
3120 
3121 	rc = free_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl);
3122 	if (rc == 0)
3123 		bzero(ofld_rxq, sizeof(*ofld_rxq));
3124 
3125 	return (rc);
3126 }
3127 #endif
3128 
3129 #ifdef DEV_NETMAP
3130 static int
3131 alloc_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int intr_idx,
3132     int idx, struct sysctl_oid *oid)
3133 {
3134 	int rc;
3135 	struct sysctl_oid_list *children;
3136 	struct sysctl_ctx_list *ctx;
3137 	char name[16];
3138 	size_t len;
3139 	struct adapter *sc = vi->pi->adapter;
3140 	struct netmap_adapter *na = NA(vi->ifp);
3141 
3142 	MPASS(na != NULL);
3143 
3144 	len = vi->qsize_rxq * IQ_ESIZE;
3145 	rc = alloc_ring(sc, len, &nm_rxq->iq_desc_tag, &nm_rxq->iq_desc_map,
3146 	    &nm_rxq->iq_ba, (void **)&nm_rxq->iq_desc);
3147 	if (rc != 0)
3148 		return (rc);
3149 
3150 	len = na->num_rx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3151 	rc = alloc_ring(sc, len, &nm_rxq->fl_desc_tag, &nm_rxq->fl_desc_map,
3152 	    &nm_rxq->fl_ba, (void **)&nm_rxq->fl_desc);
3153 	if (rc != 0)
3154 		return (rc);
3155 
3156 	nm_rxq->vi = vi;
3157 	nm_rxq->nid = idx;
3158 	nm_rxq->iq_cidx = 0;
3159 	nm_rxq->iq_sidx = vi->qsize_rxq - sc->params.sge.spg_len / IQ_ESIZE;
3160 	nm_rxq->iq_gen = F_RSPD_GEN;
3161 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
3162 	nm_rxq->fl_sidx = na->num_rx_desc;
3163 	nm_rxq->intr_idx = intr_idx;
3164 	nm_rxq->iq_cntxt_id = INVALID_NM_RXQ_CNTXT_ID;
3165 
3166 	ctx = &vi->ctx;
3167 	children = SYSCTL_CHILDREN(oid);
3168 
3169 	snprintf(name, sizeof(name), "%d", idx);
3170 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL,
3171 	    "rx queue");
3172 	children = SYSCTL_CHILDREN(oid);
3173 
3174 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3175 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_abs_id, 0, sysctl_uint16,
3176 	    "I", "absolute id of the queue");
3177 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3178 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cntxt_id, 0, sysctl_uint16,
3179 	    "I", "SGE context id of the queue");
3180 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3181 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cidx, 0, sysctl_uint16, "I",
3182 	    "consumer index");
3183 
3184 	children = SYSCTL_CHILDREN(oid);
3185 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3186 	    "freelist");
3187 	children = SYSCTL_CHILDREN(oid);
3188 
3189 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3190 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->fl_cntxt_id, 0, sysctl_uint16,
3191 	    "I", "SGE context id of the freelist");
3192 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
3193 	    &nm_rxq->fl_cidx, 0, "consumer index");
3194 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
3195 	    &nm_rxq->fl_pidx, 0, "producer index");
3196 
3197 	return (rc);
3198 }
3199 
3200 
3201 static int
3202 free_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
3203 {
3204 	struct adapter *sc = vi->pi->adapter;
3205 
3206 	if (vi->flags & VI_INIT_DONE)
3207 		MPASS(nm_rxq->iq_cntxt_id == INVALID_NM_RXQ_CNTXT_ID);
3208 	else
3209 		MPASS(nm_rxq->iq_cntxt_id == 0);
3210 
3211 	free_ring(sc, nm_rxq->iq_desc_tag, nm_rxq->iq_desc_map, nm_rxq->iq_ba,
3212 	    nm_rxq->iq_desc);
3213 	free_ring(sc, nm_rxq->fl_desc_tag, nm_rxq->fl_desc_map, nm_rxq->fl_ba,
3214 	    nm_rxq->fl_desc);
3215 
3216 	return (0);
3217 }
3218 
3219 static int
3220 alloc_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq, int iqidx, int idx,
3221     struct sysctl_oid *oid)
3222 {
3223 	int rc;
3224 	size_t len;
3225 	struct port_info *pi = vi->pi;
3226 	struct adapter *sc = pi->adapter;
3227 	struct netmap_adapter *na = NA(vi->ifp);
3228 	char name[16];
3229 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3230 
3231 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3232 	rc = alloc_ring(sc, len, &nm_txq->desc_tag, &nm_txq->desc_map,
3233 	    &nm_txq->ba, (void **)&nm_txq->desc);
3234 	if (rc)
3235 		return (rc);
3236 
3237 	nm_txq->pidx = nm_txq->cidx = 0;
3238 	nm_txq->sidx = na->num_tx_desc;
3239 	nm_txq->nid = idx;
3240 	nm_txq->iqidx = iqidx;
3241 	nm_txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3242 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(G_FW_VIID_PFN(vi->viid)) |
3243 	    V_TXPKT_VF(G_FW_VIID_VIN(vi->viid)) |
3244 	    V_TXPKT_VF_VLD(G_FW_VIID_VIVLD(vi->viid)));
3245 	nm_txq->cntxt_id = INVALID_NM_TXQ_CNTXT_ID;
3246 
3247 	snprintf(name, sizeof(name), "%d", idx);
3248 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3249 	    NULL, "netmap tx queue");
3250 	children = SYSCTL_CHILDREN(oid);
3251 
3252 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3253 	    &nm_txq->cntxt_id, 0, "SGE context id of the queue");
3254 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3255 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->cidx, 0, sysctl_uint16, "I",
3256 	    "consumer index");
3257 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3258 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->pidx, 0, sysctl_uint16, "I",
3259 	    "producer index");
3260 
3261 	return (rc);
3262 }
3263 
3264 static int
3265 free_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
3266 {
3267 	struct adapter *sc = vi->pi->adapter;
3268 
3269 	if (vi->flags & VI_INIT_DONE)
3270 		MPASS(nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID);
3271 	else
3272 		MPASS(nm_txq->cntxt_id == 0);
3273 
3274 	free_ring(sc, nm_txq->desc_tag, nm_txq->desc_map, nm_txq->ba,
3275 	    nm_txq->desc);
3276 
3277 	return (0);
3278 }
3279 #endif
3280 
3281 static int
3282 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
3283 {
3284 	int rc, cntxt_id;
3285 	struct fw_eq_ctrl_cmd c;
3286 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3287 
3288 	bzero(&c, sizeof(c));
3289 
3290 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
3291 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
3292 	    V_FW_EQ_CTRL_CMD_VFN(0));
3293 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
3294 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
3295 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid));
3296 	c.physeqid_pkd = htobe32(0);
3297 	c.fetchszm_to_iqid =
3298 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
3299 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
3300 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
3301 	c.dcaen_to_eqsize =
3302 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3303 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3304 		V_FW_EQ_CTRL_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
3305 		V_FW_EQ_CTRL_CMD_EQSIZE(qsize));
3306 	c.eqaddr = htobe64(eq->ba);
3307 
3308 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3309 	if (rc != 0) {
3310 		device_printf(sc->dev,
3311 		    "failed to create control queue %d: %d\n", eq->tx_chan, rc);
3312 		return (rc);
3313 	}
3314 	eq->flags |= EQ_ALLOCATED;
3315 
3316 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
3317 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3318 	if (cntxt_id >= sc->sge.neq)
3319 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3320 		cntxt_id, sc->sge.neq - 1);
3321 	sc->sge.eqmap[cntxt_id] = eq;
3322 
3323 	return (rc);
3324 }
3325 
3326 static int
3327 eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3328 {
3329 	int rc, cntxt_id;
3330 	struct fw_eq_eth_cmd c;
3331 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3332 
3333 	bzero(&c, sizeof(c));
3334 
3335 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
3336 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
3337 	    V_FW_EQ_ETH_CMD_VFN(0));
3338 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
3339 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
3340 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
3341 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
3342 	c.fetchszm_to_iqid =
3343 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3344 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
3345 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
3346 	c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3347 	    V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3348 	    V_FW_EQ_ETH_CMD_EQSIZE(qsize));
3349 	c.eqaddr = htobe64(eq->ba);
3350 
3351 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3352 	if (rc != 0) {
3353 		device_printf(vi->dev,
3354 		    "failed to create Ethernet egress queue: %d\n", rc);
3355 		return (rc);
3356 	}
3357 	eq->flags |= EQ_ALLOCATED;
3358 
3359 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
3360 	eq->abs_id = G_FW_EQ_ETH_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
3361 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3362 	if (cntxt_id >= sc->sge.neq)
3363 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3364 		cntxt_id, sc->sge.neq - 1);
3365 	sc->sge.eqmap[cntxt_id] = eq;
3366 
3367 	return (rc);
3368 }
3369 
3370 #ifdef TCP_OFFLOAD
3371 static int
3372 ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3373 {
3374 	int rc, cntxt_id;
3375 	struct fw_eq_ofld_cmd c;
3376 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3377 
3378 	bzero(&c, sizeof(c));
3379 
3380 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
3381 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
3382 	    V_FW_EQ_OFLD_CMD_VFN(0));
3383 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
3384 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
3385 	c.fetchszm_to_iqid =
3386 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3387 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
3388 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
3389 	c.dcaen_to_eqsize =
3390 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3391 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3392 		V_FW_EQ_OFLD_CMD_EQSIZE(qsize));
3393 	c.eqaddr = htobe64(eq->ba);
3394 
3395 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3396 	if (rc != 0) {
3397 		device_printf(vi->dev,
3398 		    "failed to create egress queue for TCP offload: %d\n", rc);
3399 		return (rc);
3400 	}
3401 	eq->flags |= EQ_ALLOCATED;
3402 
3403 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
3404 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3405 	if (cntxt_id >= sc->sge.neq)
3406 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3407 		cntxt_id, sc->sge.neq - 1);
3408 	sc->sge.eqmap[cntxt_id] = eq;
3409 
3410 	return (rc);
3411 }
3412 #endif
3413 
3414 static int
3415 alloc_eq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3416 {
3417 	int rc, qsize;
3418 	size_t len;
3419 
3420 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
3421 
3422 	qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3423 	len = qsize * EQ_ESIZE;
3424 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
3425 	    &eq->ba, (void **)&eq->desc);
3426 	if (rc)
3427 		return (rc);
3428 
3429 	eq->pidx = eq->cidx = 0;
3430 	eq->equeqidx = eq->dbidx = 0;
3431 	eq->doorbells = sc->doorbells;
3432 
3433 	switch (eq->flags & EQ_TYPEMASK) {
3434 	case EQ_CTRL:
3435 		rc = ctrl_eq_alloc(sc, eq);
3436 		break;
3437 
3438 	case EQ_ETH:
3439 		rc = eth_eq_alloc(sc, vi, eq);
3440 		break;
3441 
3442 #ifdef TCP_OFFLOAD
3443 	case EQ_OFLD:
3444 		rc = ofld_eq_alloc(sc, vi, eq);
3445 		break;
3446 #endif
3447 
3448 	default:
3449 		panic("%s: invalid eq type %d.", __func__,
3450 		    eq->flags & EQ_TYPEMASK);
3451 	}
3452 	if (rc != 0) {
3453 		device_printf(sc->dev,
3454 		    "failed to allocate egress queue(%d): %d\n",
3455 		    eq->flags & EQ_TYPEMASK, rc);
3456 	}
3457 
3458 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
3459 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
3460 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
3461 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3462 		uint32_t mask = (1 << s_qpp) - 1;
3463 		volatile uint8_t *udb;
3464 
3465 		udb = sc->udbs_base + UDBS_DB_OFFSET;
3466 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
3467 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
3468 		if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
3469 	    		clrbit(&eq->doorbells, DOORBELL_WCWR);
3470 		else {
3471 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
3472 			eq->udb_qid = 0;
3473 		}
3474 		eq->udb = (volatile void *)udb;
3475 	}
3476 
3477 	return (rc);
3478 }
3479 
3480 static int
3481 free_eq(struct adapter *sc, struct sge_eq *eq)
3482 {
3483 	int rc;
3484 
3485 	if (eq->flags & EQ_ALLOCATED) {
3486 		switch (eq->flags & EQ_TYPEMASK) {
3487 		case EQ_CTRL:
3488 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
3489 			    eq->cntxt_id);
3490 			break;
3491 
3492 		case EQ_ETH:
3493 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
3494 			    eq->cntxt_id);
3495 			break;
3496 
3497 #ifdef TCP_OFFLOAD
3498 		case EQ_OFLD:
3499 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
3500 			    eq->cntxt_id);
3501 			break;
3502 #endif
3503 
3504 		default:
3505 			panic("%s: invalid eq type %d.", __func__,
3506 			    eq->flags & EQ_TYPEMASK);
3507 		}
3508 		if (rc != 0) {
3509 			device_printf(sc->dev,
3510 			    "failed to free egress queue (%d): %d\n",
3511 			    eq->flags & EQ_TYPEMASK, rc);
3512 			return (rc);
3513 		}
3514 		eq->flags &= ~EQ_ALLOCATED;
3515 	}
3516 
3517 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
3518 
3519 	if (mtx_initialized(&eq->eq_lock))
3520 		mtx_destroy(&eq->eq_lock);
3521 
3522 	bzero(eq, sizeof(*eq));
3523 	return (0);
3524 }
3525 
3526 static int
3527 alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq,
3528     struct sysctl_oid *oid)
3529 {
3530 	int rc;
3531 	struct sysctl_ctx_list *ctx = vi ? &vi->ctx : &sc->ctx;
3532 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3533 
3534 	rc = alloc_eq(sc, vi, &wrq->eq);
3535 	if (rc)
3536 		return (rc);
3537 
3538 	wrq->adapter = sc;
3539 	TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq);
3540 	TAILQ_INIT(&wrq->incomplete_wrs);
3541 	STAILQ_INIT(&wrq->wr_list);
3542 	wrq->nwr_pending = 0;
3543 	wrq->ndesc_needed = 0;
3544 
3545 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3546 	    &wrq->eq.ba, "bus address of descriptor ring");
3547 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3548 	    wrq->eq.sidx * EQ_ESIZE + sc->params.sge.spg_len,
3549 	    "desc ring size in bytes");
3550 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3551 	    &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
3552 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3553 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I",
3554 	    "consumer index");
3555 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx",
3556 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I",
3557 	    "producer index");
3558 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
3559 	    wrq->eq.sidx, "status page index");
3560 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD,
3561 	    &wrq->tx_wrs_direct, "# of work requests (direct)");
3562 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD,
3563 	    &wrq->tx_wrs_copied, "# of work requests (copied)");
3564 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_sspace", CTLFLAG_RD,
3565 	    &wrq->tx_wrs_ss, "# of work requests (copied from scratch space)");
3566 
3567 	return (rc);
3568 }
3569 
3570 static int
3571 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
3572 {
3573 	int rc;
3574 
3575 	rc = free_eq(sc, &wrq->eq);
3576 	if (rc)
3577 		return (rc);
3578 
3579 	bzero(wrq, sizeof(*wrq));
3580 	return (0);
3581 }
3582 
3583 static int
3584 alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx,
3585     struct sysctl_oid *oid)
3586 {
3587 	int rc;
3588 	struct port_info *pi = vi->pi;
3589 	struct adapter *sc = pi->adapter;
3590 	struct sge_eq *eq = &txq->eq;
3591 	char name[16];
3592 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3593 
3594 	rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx, can_resume_eth_tx,
3595 	    M_CXGBE, M_WAITOK);
3596 	if (rc != 0) {
3597 		device_printf(sc->dev, "failed to allocate mp_ring: %d\n", rc);
3598 		return (rc);
3599 	}
3600 
3601 	rc = alloc_eq(sc, vi, eq);
3602 	if (rc != 0) {
3603 		mp_ring_free(txq->r);
3604 		txq->r = NULL;
3605 		return (rc);
3606 	}
3607 
3608 	/* Can't fail after this point. */
3609 
3610 	if (idx == 0)
3611 		sc->sge.eq_base = eq->abs_id - eq->cntxt_id;
3612 	else
3613 		KASSERT(eq->cntxt_id + sc->sge.eq_base == eq->abs_id,
3614 		    ("eq_base mismatch"));
3615 	KASSERT(sc->sge.eq_base == 0 || sc->flags & IS_VF,
3616 	    ("PF with non-zero eq_base"));
3617 
3618 	TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq);
3619 	txq->ifp = vi->ifp;
3620 	txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK);
3621 	if (sc->flags & IS_VF)
3622 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
3623 		    V_TXPKT_INTF(pi->tx_chan));
3624 	else
3625 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3626 		    V_TXPKT_INTF(pi->tx_chan) |
3627 		    V_TXPKT_PF(G_FW_VIID_PFN(vi->viid)) |
3628 		    V_TXPKT_VF(G_FW_VIID_VIN(vi->viid)) |
3629 		    V_TXPKT_VF_VLD(G_FW_VIID_VIVLD(vi->viid)));
3630 	txq->tc_idx = -1;
3631 	txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE,
3632 	    M_ZERO | M_WAITOK);
3633 
3634 	snprintf(name, sizeof(name), "%d", idx);
3635 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3636 	    NULL, "tx queue");
3637 	children = SYSCTL_CHILDREN(oid);
3638 
3639 	SYSCTL_ADD_UAUTO(&vi->ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3640 	    &eq->ba, "bus address of descriptor ring");
3641 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3642 	    eq->sidx * EQ_ESIZE + sc->params.sge.spg_len,
3643 	    "desc ring size in bytes");
3644 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
3645 	    &eq->abs_id, 0, "absolute id of the queue");
3646 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3647 	    &eq->cntxt_id, 0, "SGE context id of the queue");
3648 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3649 	    CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I",
3650 	    "consumer index");
3651 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3652 	    CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I",
3653 	    "producer index");
3654 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
3655 	    eq->sidx, "status page index");
3656 
3657 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "tc",
3658 	    CTLTYPE_INT | CTLFLAG_RW, vi, idx, sysctl_tc, "I",
3659 	    "traffic class (-1 means none)");
3660 
3661 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
3662 	    &txq->txcsum, "# of times hardware assisted with checksum");
3663 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_insertion",
3664 	    CTLFLAG_RD, &txq->vlan_insertion,
3665 	    "# of times hardware inserted 802.1Q tag");
3666 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
3667 	    &txq->tso_wrs, "# of TSO work requests");
3668 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
3669 	    &txq->imm_wrs, "# of work requests with immediate data");
3670 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
3671 	    &txq->sgl_wrs, "# of work requests with direct SGL");
3672 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
3673 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
3674 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_wrs",
3675 	    CTLFLAG_RD, &txq->txpkts0_wrs,
3676 	    "# of txpkts (type 0) work requests");
3677 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_wrs",
3678 	    CTLFLAG_RD, &txq->txpkts1_wrs,
3679 	    "# of txpkts (type 1) work requests");
3680 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_pkts",
3681 	    CTLFLAG_RD, &txq->txpkts0_pkts,
3682 	    "# of frames tx'd using type0 txpkts work requests");
3683 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_pkts",
3684 	    CTLFLAG_RD, &txq->txpkts1_pkts,
3685 	    "# of frames tx'd using type1 txpkts work requests");
3686 
3687 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_enqueues",
3688 	    CTLFLAG_RD, &txq->r->enqueues,
3689 	    "# of enqueues to the mp_ring for this queue");
3690 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_drops",
3691 	    CTLFLAG_RD, &txq->r->drops,
3692 	    "# of drops in the mp_ring for this queue");
3693 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_starts",
3694 	    CTLFLAG_RD, &txq->r->starts,
3695 	    "# of normal consumer starts in the mp_ring for this queue");
3696 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_stalls",
3697 	    CTLFLAG_RD, &txq->r->stalls,
3698 	    "# of consumer stalls in the mp_ring for this queue");
3699 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_restarts",
3700 	    CTLFLAG_RD, &txq->r->restarts,
3701 	    "# of consumer restarts in the mp_ring for this queue");
3702 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_abdications",
3703 	    CTLFLAG_RD, &txq->r->abdications,
3704 	    "# of consumer abdications in the mp_ring for this queue");
3705 
3706 	return (0);
3707 }
3708 
3709 static int
3710 free_txq(struct vi_info *vi, struct sge_txq *txq)
3711 {
3712 	int rc;
3713 	struct adapter *sc = vi->pi->adapter;
3714 	struct sge_eq *eq = &txq->eq;
3715 
3716 	rc = free_eq(sc, eq);
3717 	if (rc)
3718 		return (rc);
3719 
3720 	sglist_free(txq->gl);
3721 	free(txq->sdesc, M_CXGBE);
3722 	mp_ring_free(txq->r);
3723 
3724 	bzero(txq, sizeof(*txq));
3725 	return (0);
3726 }
3727 
3728 static void
3729 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
3730 {
3731 	bus_addr_t *ba = arg;
3732 
3733 	KASSERT(nseg == 1,
3734 	    ("%s meant for single segment mappings only.", __func__));
3735 
3736 	*ba = error ? 0 : segs->ds_addr;
3737 }
3738 
3739 static inline void
3740 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
3741 {
3742 	uint32_t n, v;
3743 
3744 	n = IDXDIFF(fl->pidx / 8, fl->dbidx, fl->sidx);
3745 	MPASS(n > 0);
3746 
3747 	wmb();
3748 	v = fl->dbval | V_PIDX(n);
3749 	if (fl->udb)
3750 		*fl->udb = htole32(v);
3751 	else
3752 		t4_write_reg(sc, sc->sge_kdoorbell_reg, v);
3753 	IDXINCR(fl->dbidx, n, fl->sidx);
3754 }
3755 
3756 /*
3757  * Fills up the freelist by allocating up to 'n' buffers.  Buffers that are
3758  * recycled do not count towards this allocation budget.
3759  *
3760  * Returns non-zero to indicate that this freelist should be added to the list
3761  * of starving freelists.
3762  */
3763 static int
3764 refill_fl(struct adapter *sc, struct sge_fl *fl, int n)
3765 {
3766 	__be64 *d;
3767 	struct fl_sdesc *sd;
3768 	uintptr_t pa;
3769 	caddr_t cl;
3770 	struct cluster_layout *cll;
3771 	struct sw_zone_info *swz;
3772 	struct cluster_metadata *clm;
3773 	uint16_t max_pidx;
3774 	uint16_t hw_cidx = fl->hw_cidx;		/* stable snapshot */
3775 
3776 	FL_LOCK_ASSERT_OWNED(fl);
3777 
3778 	/*
3779 	 * We always stop at the beginning of the hardware descriptor that's just
3780 	 * before the one with the hw cidx.  This is to avoid hw pidx = hw cidx,
3781 	 * which would mean an empty freelist to the chip.
3782 	 */
3783 	max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1;
3784 	if (fl->pidx == max_pidx * 8)
3785 		return (0);
3786 
3787 	d = &fl->desc[fl->pidx];
3788 	sd = &fl->sdesc[fl->pidx];
3789 	cll = &fl->cll_def;	/* default layout */
3790 	swz = &sc->sge.sw_zone_info[cll->zidx];
3791 
3792 	while (n > 0) {
3793 
3794 		if (sd->cl != NULL) {
3795 
3796 			if (sd->nmbuf == 0) {
3797 				/*
3798 				 * Fast recycle without involving any atomics on
3799 				 * the cluster's metadata (if the cluster has
3800 				 * metadata).  This happens when all frames
3801 				 * received in the cluster were small enough to
3802 				 * fit within a single mbuf each.
3803 				 */
3804 				fl->cl_fast_recycled++;
3805 #ifdef INVARIANTS
3806 				clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
3807 				if (clm != NULL)
3808 					MPASS(clm->refcount == 1);
3809 #endif
3810 				goto recycled_fast;
3811 			}
3812 
3813 			/*
3814 			 * Cluster is guaranteed to have metadata.  Clusters
3815 			 * without metadata always take the fast recycle path
3816 			 * when they're recycled.
3817 			 */
3818 			clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
3819 			MPASS(clm != NULL);
3820 
3821 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
3822 				fl->cl_recycled++;
3823 				counter_u64_add(extfree_rels, 1);
3824 				goto recycled;
3825 			}
3826 			sd->cl = NULL;	/* gave up my reference */
3827 		}
3828 		MPASS(sd->cl == NULL);
3829 alloc:
3830 		cl = uma_zalloc(swz->zone, M_NOWAIT);
3831 		if (__predict_false(cl == NULL)) {
3832 			if (cll == &fl->cll_alt || fl->cll_alt.zidx == -1 ||
3833 			    fl->cll_def.zidx == fl->cll_alt.zidx)
3834 				break;
3835 
3836 			/* fall back to the safe zone */
3837 			cll = &fl->cll_alt;
3838 			swz = &sc->sge.sw_zone_info[cll->zidx];
3839 			goto alloc;
3840 		}
3841 		fl->cl_allocated++;
3842 		n--;
3843 
3844 		pa = pmap_kextract((vm_offset_t)cl);
3845 		pa += cll->region1;
3846 		sd->cl = cl;
3847 		sd->cll = *cll;
3848 		*d = htobe64(pa | cll->hwidx);
3849 		clm = cl_metadata(sc, fl, cll, cl);
3850 		if (clm != NULL) {
3851 recycled:
3852 #ifdef INVARIANTS
3853 			clm->sd = sd;
3854 #endif
3855 			clm->refcount = 1;
3856 		}
3857 		sd->nmbuf = 0;
3858 recycled_fast:
3859 		d++;
3860 		sd++;
3861 		if (__predict_false(++fl->pidx % 8 == 0)) {
3862 			uint16_t pidx = fl->pidx / 8;
3863 
3864 			if (__predict_false(pidx == fl->sidx)) {
3865 				fl->pidx = 0;
3866 				pidx = 0;
3867 				sd = fl->sdesc;
3868 				d = fl->desc;
3869 			}
3870 			if (pidx == max_pidx)
3871 				break;
3872 
3873 			if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4)
3874 				ring_fl_db(sc, fl);
3875 		}
3876 	}
3877 
3878 	if (fl->pidx / 8 != fl->dbidx)
3879 		ring_fl_db(sc, fl);
3880 
3881 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
3882 }
3883 
3884 /*
3885  * Attempt to refill all starving freelists.
3886  */
3887 static void
3888 refill_sfl(void *arg)
3889 {
3890 	struct adapter *sc = arg;
3891 	struct sge_fl *fl, *fl_temp;
3892 
3893 	mtx_assert(&sc->sfl_lock, MA_OWNED);
3894 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
3895 		FL_LOCK(fl);
3896 		refill_fl(sc, fl, 64);
3897 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
3898 			TAILQ_REMOVE(&sc->sfl, fl, link);
3899 			fl->flags &= ~FL_STARVING;
3900 		}
3901 		FL_UNLOCK(fl);
3902 	}
3903 
3904 	if (!TAILQ_EMPTY(&sc->sfl))
3905 		callout_schedule(&sc->sfl_callout, hz / 5);
3906 }
3907 
3908 static int
3909 alloc_fl_sdesc(struct sge_fl *fl)
3910 {
3911 
3912 	fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc), M_CXGBE,
3913 	    M_ZERO | M_WAITOK);
3914 
3915 	return (0);
3916 }
3917 
3918 static void
3919 free_fl_sdesc(struct adapter *sc, struct sge_fl *fl)
3920 {
3921 	struct fl_sdesc *sd;
3922 	struct cluster_metadata *clm;
3923 	struct cluster_layout *cll;
3924 	int i;
3925 
3926 	sd = fl->sdesc;
3927 	for (i = 0; i < fl->sidx * 8; i++, sd++) {
3928 		if (sd->cl == NULL)
3929 			continue;
3930 
3931 		cll = &sd->cll;
3932 		clm = cl_metadata(sc, fl, cll, sd->cl);
3933 		if (sd->nmbuf == 0)
3934 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
3935 		else if (clm && atomic_fetchadd_int(&clm->refcount, -1) == 1) {
3936 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
3937 			counter_u64_add(extfree_rels, 1);
3938 		}
3939 		sd->cl = NULL;
3940 	}
3941 
3942 	free(fl->sdesc, M_CXGBE);
3943 	fl->sdesc = NULL;
3944 }
3945 
3946 static inline void
3947 get_pkt_gl(struct mbuf *m, struct sglist *gl)
3948 {
3949 	int rc;
3950 
3951 	M_ASSERTPKTHDR(m);
3952 
3953 	sglist_reset(gl);
3954 	rc = sglist_append_mbuf(gl, m);
3955 	if (__predict_false(rc != 0)) {
3956 		panic("%s: mbuf %p (%d segs) was vetted earlier but now fails "
3957 		    "with %d.", __func__, m, mbuf_nsegs(m), rc);
3958 	}
3959 
3960 	KASSERT(gl->sg_nseg == mbuf_nsegs(m),
3961 	    ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m,
3962 	    mbuf_nsegs(m), gl->sg_nseg));
3963 	KASSERT(gl->sg_nseg > 0 &&
3964 	    gl->sg_nseg <= (needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS),
3965 	    ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__,
3966 		gl->sg_nseg, needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS));
3967 }
3968 
3969 /*
3970  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
3971  */
3972 static inline u_int
3973 txpkt_len16(u_int nsegs, u_int tso)
3974 {
3975 	u_int n;
3976 
3977 	MPASS(nsegs > 0);
3978 
3979 	nsegs--; /* first segment is part of ulptx_sgl */
3980 	n = sizeof(struct fw_eth_tx_pkt_wr) + sizeof(struct cpl_tx_pkt_core) +
3981 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
3982 	if (tso)
3983 		n += sizeof(struct cpl_tx_pkt_lso_core);
3984 
3985 	return (howmany(n, 16));
3986 }
3987 
3988 /*
3989  * len16 for a txpkt_vm WR with a GL.  Includes the firmware work
3990  * request header.
3991  */
3992 static inline u_int
3993 txpkt_vm_len16(u_int nsegs, u_int tso)
3994 {
3995 	u_int n;
3996 
3997 	MPASS(nsegs > 0);
3998 
3999 	nsegs--; /* first segment is part of ulptx_sgl */
4000 	n = sizeof(struct fw_eth_tx_pkt_vm_wr) +
4001 	    sizeof(struct cpl_tx_pkt_core) +
4002 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4003 	if (tso)
4004 		n += sizeof(struct cpl_tx_pkt_lso_core);
4005 
4006 	return (howmany(n, 16));
4007 }
4008 
4009 /*
4010  * len16 for a txpkts type 0 WR with a GL.  Does not include the firmware work
4011  * request header.
4012  */
4013 static inline u_int
4014 txpkts0_len16(u_int nsegs)
4015 {
4016 	u_int n;
4017 
4018 	MPASS(nsegs > 0);
4019 
4020 	nsegs--; /* first segment is part of ulptx_sgl */
4021 	n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) +
4022 	    sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) +
4023 	    8 * ((3 * nsegs) / 2 + (nsegs & 1));
4024 
4025 	return (howmany(n, 16));
4026 }
4027 
4028 /*
4029  * len16 for a txpkts type 1 WR with a GL.  Does not include the firmware work
4030  * request header.
4031  */
4032 static inline u_int
4033 txpkts1_len16(void)
4034 {
4035 	u_int n;
4036 
4037 	n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl);
4038 
4039 	return (howmany(n, 16));
4040 }
4041 
4042 static inline u_int
4043 imm_payload(u_int ndesc)
4044 {
4045 	u_int n;
4046 
4047 	n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) -
4048 	    sizeof(struct cpl_tx_pkt_core);
4049 
4050 	return (n);
4051 }
4052 
4053 /*
4054  * Write a VM txpkt WR for this packet to the hardware descriptors, update the
4055  * software descriptor, and advance the pidx.  It is guaranteed that enough
4056  * descriptors are available.
4057  *
4058  * The return value is the # of hardware descriptors used.
4059  */
4060 static u_int
4061 write_txpkt_vm_wr(struct adapter *sc, struct sge_txq *txq,
4062     struct fw_eth_tx_pkt_vm_wr *wr, struct mbuf *m0, u_int available)
4063 {
4064 	struct sge_eq *eq = &txq->eq;
4065 	struct tx_sdesc *txsd;
4066 	struct cpl_tx_pkt_core *cpl;
4067 	uint32_t ctrl;	/* used in many unrelated places */
4068 	uint64_t ctrl1;
4069 	int csum_type, len16, ndesc, pktlen, nsegs;
4070 	caddr_t dst;
4071 
4072 	TXQ_LOCK_ASSERT_OWNED(txq);
4073 	M_ASSERTPKTHDR(m0);
4074 	MPASS(available > 0 && available < eq->sidx);
4075 
4076 	len16 = mbuf_len16(m0);
4077 	nsegs = mbuf_nsegs(m0);
4078 	pktlen = m0->m_pkthdr.len;
4079 	ctrl = sizeof(struct cpl_tx_pkt_core);
4080 	if (needs_tso(m0))
4081 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4082 	ndesc = howmany(len16, EQ_ESIZE / 16);
4083 	MPASS(ndesc <= available);
4084 
4085 	/* Firmware work request header */
4086 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4087 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_VM_WR) |
4088 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4089 
4090 	ctrl = V_FW_WR_LEN16(len16);
4091 	wr->equiq_to_len16 = htobe32(ctrl);
4092 	wr->r3[0] = 0;
4093 	wr->r3[1] = 0;
4094 
4095 	/*
4096 	 * Copy over ethmacdst, ethmacsrc, ethtype, and vlantci.
4097 	 * vlantci is ignored unless the ethtype is 0x8100, so it's
4098 	 * simpler to always copy it rather than making it
4099 	 * conditional.  Also, it seems that we do not have to set
4100 	 * vlantci or fake the ethtype when doing VLAN tag insertion.
4101 	 */
4102 	m_copydata(m0, 0, sizeof(struct ether_header) + 2, wr->ethmacdst);
4103 
4104 	csum_type = -1;
4105 	if (needs_tso(m0)) {
4106 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4107 
4108 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4109 		    m0->m_pkthdr.l4hlen > 0,
4110 		    ("%s: mbuf %p needs TSO but missing header lengths",
4111 			__func__, m0));
4112 
4113 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4114 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4115 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4116 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4117 			ctrl |= V_LSO_ETHHDR_LEN(1);
4118 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4119 			ctrl |= F_LSO_IPV6;
4120 
4121 		lso->lso_ctrl = htobe32(ctrl);
4122 		lso->ipid_ofst = htobe16(0);
4123 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4124 		lso->seqno_offset = htobe32(0);
4125 		lso->len = htobe32(pktlen);
4126 
4127 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4128 			csum_type = TX_CSUM_TCPIP6;
4129 		else
4130 			csum_type = TX_CSUM_TCPIP;
4131 
4132 		cpl = (void *)(lso + 1);
4133 
4134 		txq->tso_wrs++;
4135 	} else {
4136 		if (m0->m_pkthdr.csum_flags & CSUM_IP_TCP)
4137 			csum_type = TX_CSUM_TCPIP;
4138 		else if (m0->m_pkthdr.csum_flags & CSUM_IP_UDP)
4139 			csum_type = TX_CSUM_UDPIP;
4140 		else if (m0->m_pkthdr.csum_flags & CSUM_IP6_TCP)
4141 			csum_type = TX_CSUM_TCPIP6;
4142 		else if (m0->m_pkthdr.csum_flags & CSUM_IP6_UDP)
4143 			csum_type = TX_CSUM_UDPIP6;
4144 #if defined(INET)
4145 		else if (m0->m_pkthdr.csum_flags & CSUM_IP) {
4146 			/*
4147 			 * XXX: The firmware appears to stomp on the
4148 			 * fragment/flags field of the IP header when
4149 			 * using TX_CSUM_IP.  Fall back to doing
4150 			 * software checksums.
4151 			 */
4152 			u_short *sump;
4153 			struct mbuf *m;
4154 			int offset;
4155 
4156 			m = m0;
4157 			offset = 0;
4158 			sump = m_advance(&m, &offset, m0->m_pkthdr.l2hlen +
4159 			    offsetof(struct ip, ip_sum));
4160 			*sump = in_cksum_skip(m0, m0->m_pkthdr.l2hlen +
4161 			    m0->m_pkthdr.l3hlen, m0->m_pkthdr.l2hlen);
4162 			m0->m_pkthdr.csum_flags &= ~CSUM_IP;
4163 		}
4164 #endif
4165 
4166 		cpl = (void *)(wr + 1);
4167 	}
4168 
4169 	/* Checksum offload */
4170 	ctrl1 = 0;
4171 	if (needs_l3_csum(m0) == 0)
4172 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4173 	if (csum_type >= 0) {
4174 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0,
4175 	    ("%s: mbuf %p needs checksum offload but missing header lengths",
4176 			__func__, m0));
4177 
4178 		if (chip_id(sc) <= CHELSIO_T5) {
4179 			ctrl1 |= V_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4180 			    ETHER_HDR_LEN);
4181 		} else {
4182 			ctrl1 |= V_T6_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4183 			    ETHER_HDR_LEN);
4184 		}
4185 		ctrl1 |= V_TXPKT_IPHDR_LEN(m0->m_pkthdr.l3hlen);
4186 		ctrl1 |= V_TXPKT_CSUM_TYPE(csum_type);
4187 	} else
4188 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4189 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4190 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4191 		txq->txcsum++;	/* some hardware assistance provided */
4192 
4193 	/* VLAN tag insertion */
4194 	if (needs_vlan_insertion(m0)) {
4195 		ctrl1 |= F_TXPKT_VLAN_VLD |
4196 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4197 		txq->vlan_insertion++;
4198 	}
4199 
4200 	/* CPL header */
4201 	cpl->ctrl0 = txq->cpl_ctrl0;
4202 	cpl->pack = 0;
4203 	cpl->len = htobe16(pktlen);
4204 	cpl->ctrl1 = htobe64(ctrl1);
4205 
4206 	/* SGL */
4207 	dst = (void *)(cpl + 1);
4208 
4209 	/*
4210 	 * A packet using TSO will use up an entire descriptor for the
4211 	 * firmware work request header, LSO CPL, and TX_PKT_XT CPL.
4212 	 * If this descriptor is the last descriptor in the ring, wrap
4213 	 * around to the front of the ring explicitly for the start of
4214 	 * the sgl.
4215 	 */
4216 	if (dst == (void *)&eq->desc[eq->sidx]) {
4217 		dst = (void *)&eq->desc[0];
4218 		write_gl_to_txd(txq, m0, &dst, 0);
4219 	} else
4220 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4221 	txq->sgl_wrs++;
4222 
4223 	txq->txpkt_wrs++;
4224 
4225 	txsd = &txq->sdesc[eq->pidx];
4226 	txsd->m = m0;
4227 	txsd->desc_used = ndesc;
4228 
4229 	return (ndesc);
4230 }
4231 
4232 /*
4233  * Write a txpkt WR for this packet to the hardware descriptors, update the
4234  * software descriptor, and advance the pidx.  It is guaranteed that enough
4235  * descriptors are available.
4236  *
4237  * The return value is the # of hardware descriptors used.
4238  */
4239 static u_int
4240 write_txpkt_wr(struct sge_txq *txq, struct fw_eth_tx_pkt_wr *wr,
4241     struct mbuf *m0, u_int available)
4242 {
4243 	struct sge_eq *eq = &txq->eq;
4244 	struct tx_sdesc *txsd;
4245 	struct cpl_tx_pkt_core *cpl;
4246 	uint32_t ctrl;	/* used in many unrelated places */
4247 	uint64_t ctrl1;
4248 	int len16, ndesc, pktlen, nsegs;
4249 	caddr_t dst;
4250 
4251 	TXQ_LOCK_ASSERT_OWNED(txq);
4252 	M_ASSERTPKTHDR(m0);
4253 	MPASS(available > 0 && available < eq->sidx);
4254 
4255 	len16 = mbuf_len16(m0);
4256 	nsegs = mbuf_nsegs(m0);
4257 	pktlen = m0->m_pkthdr.len;
4258 	ctrl = sizeof(struct cpl_tx_pkt_core);
4259 	if (needs_tso(m0))
4260 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4261 	else if (pktlen <= imm_payload(2) && available >= 2) {
4262 		/* Immediate data.  Recalculate len16 and set nsegs to 0. */
4263 		ctrl += pktlen;
4264 		len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) +
4265 		    sizeof(struct cpl_tx_pkt_core) + pktlen, 16);
4266 		nsegs = 0;
4267 	}
4268 	ndesc = howmany(len16, EQ_ESIZE / 16);
4269 	MPASS(ndesc <= available);
4270 
4271 	/* Firmware work request header */
4272 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4273 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
4274 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4275 
4276 	ctrl = V_FW_WR_LEN16(len16);
4277 	wr->equiq_to_len16 = htobe32(ctrl);
4278 	wr->r3 = 0;
4279 
4280 	if (needs_tso(m0)) {
4281 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4282 
4283 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4284 		    m0->m_pkthdr.l4hlen > 0,
4285 		    ("%s: mbuf %p needs TSO but missing header lengths",
4286 			__func__, m0));
4287 
4288 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4289 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4290 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4291 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4292 			ctrl |= V_LSO_ETHHDR_LEN(1);
4293 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4294 			ctrl |= F_LSO_IPV6;
4295 
4296 		lso->lso_ctrl = htobe32(ctrl);
4297 		lso->ipid_ofst = htobe16(0);
4298 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4299 		lso->seqno_offset = htobe32(0);
4300 		lso->len = htobe32(pktlen);
4301 
4302 		cpl = (void *)(lso + 1);
4303 
4304 		txq->tso_wrs++;
4305 	} else
4306 		cpl = (void *)(wr + 1);
4307 
4308 	/* Checksum offload */
4309 	ctrl1 = 0;
4310 	if (needs_l3_csum(m0) == 0)
4311 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4312 	if (needs_l4_csum(m0) == 0)
4313 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4314 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4315 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4316 		txq->txcsum++;	/* some hardware assistance provided */
4317 
4318 	/* VLAN tag insertion */
4319 	if (needs_vlan_insertion(m0)) {
4320 		ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4321 		txq->vlan_insertion++;
4322 	}
4323 
4324 	/* CPL header */
4325 	cpl->ctrl0 = txq->cpl_ctrl0;
4326 	cpl->pack = 0;
4327 	cpl->len = htobe16(pktlen);
4328 	cpl->ctrl1 = htobe64(ctrl1);
4329 
4330 	/* SGL */
4331 	dst = (void *)(cpl + 1);
4332 	if (nsegs > 0) {
4333 
4334 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4335 		txq->sgl_wrs++;
4336 	} else {
4337 		struct mbuf *m;
4338 
4339 		for (m = m0; m != NULL; m = m->m_next) {
4340 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4341 #ifdef INVARIANTS
4342 			pktlen -= m->m_len;
4343 #endif
4344 		}
4345 #ifdef INVARIANTS
4346 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
4347 #endif
4348 		txq->imm_wrs++;
4349 	}
4350 
4351 	txq->txpkt_wrs++;
4352 
4353 	txsd = &txq->sdesc[eq->pidx];
4354 	txsd->m = m0;
4355 	txsd->desc_used = ndesc;
4356 
4357 	return (ndesc);
4358 }
4359 
4360 static int
4361 try_txpkts(struct mbuf *m, struct mbuf *n, struct txpkts *txp, u_int available)
4362 {
4363 	u_int needed, nsegs1, nsegs2, l1, l2;
4364 
4365 	if (cannot_use_txpkts(m) || cannot_use_txpkts(n))
4366 		return (1);
4367 
4368 	nsegs1 = mbuf_nsegs(m);
4369 	nsegs2 = mbuf_nsegs(n);
4370 	if (nsegs1 + nsegs2 == 2) {
4371 		txp->wr_type = 1;
4372 		l1 = l2 = txpkts1_len16();
4373 	} else {
4374 		txp->wr_type = 0;
4375 		l1 = txpkts0_len16(nsegs1);
4376 		l2 = txpkts0_len16(nsegs2);
4377 	}
4378 	txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) + l1 + l2;
4379 	needed = howmany(txp->len16, EQ_ESIZE / 16);
4380 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4381 		return (1);
4382 
4383 	txp->plen = m->m_pkthdr.len + n->m_pkthdr.len;
4384 	if (txp->plen > 65535)
4385 		return (1);
4386 
4387 	txp->npkt = 2;
4388 	set_mbuf_len16(m, l1);
4389 	set_mbuf_len16(n, l2);
4390 
4391 	return (0);
4392 }
4393 
4394 static int
4395 add_to_txpkts(struct mbuf *m, struct txpkts *txp, u_int available)
4396 {
4397 	u_int plen, len16, needed, nsegs;
4398 
4399 	MPASS(txp->wr_type == 0 || txp->wr_type == 1);
4400 
4401 	nsegs = mbuf_nsegs(m);
4402 	if (needs_tso(m) || (txp->wr_type == 1 && nsegs != 1))
4403 		return (1);
4404 
4405 	plen = txp->plen + m->m_pkthdr.len;
4406 	if (plen > 65535)
4407 		return (1);
4408 
4409 	if (txp->wr_type == 0)
4410 		len16 = txpkts0_len16(nsegs);
4411 	else
4412 		len16 = txpkts1_len16();
4413 	needed = howmany(txp->len16 + len16, EQ_ESIZE / 16);
4414 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4415 		return (1);
4416 
4417 	txp->npkt++;
4418 	txp->plen = plen;
4419 	txp->len16 += len16;
4420 	set_mbuf_len16(m, len16);
4421 
4422 	return (0);
4423 }
4424 
4425 /*
4426  * Write a txpkts WR for the packets in txp to the hardware descriptors, update
4427  * the software descriptor, and advance the pidx.  It is guaranteed that enough
4428  * descriptors are available.
4429  *
4430  * The return value is the # of hardware descriptors used.
4431  */
4432 static u_int
4433 write_txpkts_wr(struct sge_txq *txq, struct fw_eth_tx_pkts_wr *wr,
4434     struct mbuf *m0, const struct txpkts *txp, u_int available)
4435 {
4436 	struct sge_eq *eq = &txq->eq;
4437 	struct tx_sdesc *txsd;
4438 	struct cpl_tx_pkt_core *cpl;
4439 	uint32_t ctrl;
4440 	uint64_t ctrl1;
4441 	int ndesc, checkwrap;
4442 	struct mbuf *m;
4443 	void *flitp;
4444 
4445 	TXQ_LOCK_ASSERT_OWNED(txq);
4446 	MPASS(txp->npkt > 0);
4447 	MPASS(txp->plen < 65536);
4448 	MPASS(m0 != NULL);
4449 	MPASS(m0->m_nextpkt != NULL);
4450 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
4451 	MPASS(available > 0 && available < eq->sidx);
4452 
4453 	ndesc = howmany(txp->len16, EQ_ESIZE / 16);
4454 	MPASS(ndesc <= available);
4455 
4456 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4457 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
4458 	ctrl = V_FW_WR_LEN16(txp->len16);
4459 	wr->equiq_to_len16 = htobe32(ctrl);
4460 	wr->plen = htobe16(txp->plen);
4461 	wr->npkt = txp->npkt;
4462 	wr->r3 = 0;
4463 	wr->type = txp->wr_type;
4464 	flitp = wr + 1;
4465 
4466 	/*
4467 	 * At this point we are 16B into a hardware descriptor.  If checkwrap is
4468 	 * set then we know the WR is going to wrap around somewhere.  We'll
4469 	 * check for that at appropriate points.
4470 	 */
4471 	checkwrap = eq->sidx - ndesc < eq->pidx;
4472 	for (m = m0; m != NULL; m = m->m_nextpkt) {
4473 		if (txp->wr_type == 0) {
4474 			struct ulp_txpkt *ulpmc;
4475 			struct ulptx_idata *ulpsc;
4476 
4477 			/* ULP master command */
4478 			ulpmc = flitp;
4479 			ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
4480 			    V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid));
4481 			ulpmc->len = htobe32(mbuf_len16(m));
4482 
4483 			/* ULP subcommand */
4484 			ulpsc = (void *)(ulpmc + 1);
4485 			ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
4486 			    F_ULP_TX_SC_MORE);
4487 			ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
4488 
4489 			cpl = (void *)(ulpsc + 1);
4490 			if (checkwrap &&
4491 			    (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx])
4492 				cpl = (void *)&eq->desc[0];
4493 		} else {
4494 			cpl = flitp;
4495 		}
4496 
4497 		/* Checksum offload */
4498 		ctrl1 = 0;
4499 		if (needs_l3_csum(m) == 0)
4500 			ctrl1 |= F_TXPKT_IPCSUM_DIS;
4501 		if (needs_l4_csum(m) == 0)
4502 			ctrl1 |= F_TXPKT_L4CSUM_DIS;
4503 		if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4504 		    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4505 			txq->txcsum++;	/* some hardware assistance provided */
4506 
4507 		/* VLAN tag insertion */
4508 		if (needs_vlan_insertion(m)) {
4509 			ctrl1 |= F_TXPKT_VLAN_VLD |
4510 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
4511 			txq->vlan_insertion++;
4512 		}
4513 
4514 		/* CPL header */
4515 		cpl->ctrl0 = txq->cpl_ctrl0;
4516 		cpl->pack = 0;
4517 		cpl->len = htobe16(m->m_pkthdr.len);
4518 		cpl->ctrl1 = htobe64(ctrl1);
4519 
4520 		flitp = cpl + 1;
4521 		if (checkwrap &&
4522 		    (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
4523 			flitp = (void *)&eq->desc[0];
4524 
4525 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap);
4526 
4527 	}
4528 
4529 	if (txp->wr_type == 0) {
4530 		txq->txpkts0_pkts += txp->npkt;
4531 		txq->txpkts0_wrs++;
4532 	} else {
4533 		txq->txpkts1_pkts += txp->npkt;
4534 		txq->txpkts1_wrs++;
4535 	}
4536 
4537 	txsd = &txq->sdesc[eq->pidx];
4538 	txsd->m = m0;
4539 	txsd->desc_used = ndesc;
4540 
4541 	return (ndesc);
4542 }
4543 
4544 /*
4545  * If the SGL ends on an address that is not 16 byte aligned, this function will
4546  * add a 0 filled flit at the end.
4547  */
4548 static void
4549 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap)
4550 {
4551 	struct sge_eq *eq = &txq->eq;
4552 	struct sglist *gl = txq->gl;
4553 	struct sglist_seg *seg;
4554 	__be64 *flitp, *wrap;
4555 	struct ulptx_sgl *usgl;
4556 	int i, nflits, nsegs;
4557 
4558 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
4559 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
4560 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
4561 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
4562 
4563 	get_pkt_gl(m, gl);
4564 	nsegs = gl->sg_nseg;
4565 	MPASS(nsegs > 0);
4566 
4567 	nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
4568 	flitp = (__be64 *)(*to);
4569 	wrap = (__be64 *)(&eq->desc[eq->sidx]);
4570 	seg = &gl->sg_segs[0];
4571 	usgl = (void *)flitp;
4572 
4573 	/*
4574 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
4575 	 * ring, so we're at least 16 bytes away from the status page.  There is
4576 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
4577 	 */
4578 
4579 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
4580 	    V_ULPTX_NSGE(nsegs));
4581 	usgl->len0 = htobe32(seg->ss_len);
4582 	usgl->addr0 = htobe64(seg->ss_paddr);
4583 	seg++;
4584 
4585 	if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) {
4586 
4587 		/* Won't wrap around at all */
4588 
4589 		for (i = 0; i < nsegs - 1; i++, seg++) {
4590 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
4591 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
4592 		}
4593 		if (i & 1)
4594 			usgl->sge[i / 2].len[1] = htobe32(0);
4595 		flitp += nflits;
4596 	} else {
4597 
4598 		/* Will wrap somewhere in the rest of the SGL */
4599 
4600 		/* 2 flits already written, write the rest flit by flit */
4601 		flitp = (void *)(usgl + 1);
4602 		for (i = 0; i < nflits - 2; i++) {
4603 			if (flitp == wrap)
4604 				flitp = (void *)eq->desc;
4605 			*flitp++ = get_flit(seg, nsegs - 1, i);
4606 		}
4607 	}
4608 
4609 	if (nflits & 1) {
4610 		MPASS(((uintptr_t)flitp) & 0xf);
4611 		*flitp++ = 0;
4612 	}
4613 
4614 	MPASS((((uintptr_t)flitp) & 0xf) == 0);
4615 	if (__predict_false(flitp == wrap))
4616 		*to = (void *)eq->desc;
4617 	else
4618 		*to = (void *)flitp;
4619 }
4620 
4621 static inline void
4622 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
4623 {
4624 
4625 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
4626 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
4627 
4628 	if (__predict_true((uintptr_t)(*to) + len <=
4629 	    (uintptr_t)&eq->desc[eq->sidx])) {
4630 		bcopy(from, *to, len);
4631 		(*to) += len;
4632 	} else {
4633 		int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
4634 
4635 		bcopy(from, *to, portion);
4636 		from += portion;
4637 		portion = len - portion;	/* remaining */
4638 		bcopy(from, (void *)eq->desc, portion);
4639 		(*to) = (caddr_t)eq->desc + portion;
4640 	}
4641 }
4642 
4643 static inline void
4644 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n)
4645 {
4646 	u_int db;
4647 
4648 	MPASS(n > 0);
4649 
4650 	db = eq->doorbells;
4651 	if (n > 1)
4652 		clrbit(&db, DOORBELL_WCWR);
4653 	wmb();
4654 
4655 	switch (ffs(db) - 1) {
4656 	case DOORBELL_UDB:
4657 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
4658 		break;
4659 
4660 	case DOORBELL_WCWR: {
4661 		volatile uint64_t *dst, *src;
4662 		int i;
4663 
4664 		/*
4665 		 * Queues whose 128B doorbell segment fits in the page do not
4666 		 * use relative qid (udb_qid is always 0).  Only queues with
4667 		 * doorbell segments can do WCWR.
4668 		 */
4669 		KASSERT(eq->udb_qid == 0 && n == 1,
4670 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
4671 		    __func__, eq->doorbells, n, eq->dbidx, eq));
4672 
4673 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
4674 		    UDBS_DB_OFFSET);
4675 		i = eq->dbidx;
4676 		src = (void *)&eq->desc[i];
4677 		while (src != (void *)&eq->desc[i + 1])
4678 			*dst++ = *src++;
4679 		wmb();
4680 		break;
4681 	}
4682 
4683 	case DOORBELL_UDBWC:
4684 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
4685 		wmb();
4686 		break;
4687 
4688 	case DOORBELL_KDB:
4689 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
4690 		    V_QID(eq->cntxt_id) | V_PIDX(n));
4691 		break;
4692 	}
4693 
4694 	IDXINCR(eq->dbidx, n, eq->sidx);
4695 }
4696 
4697 static inline u_int
4698 reclaimable_tx_desc(struct sge_eq *eq)
4699 {
4700 	uint16_t hw_cidx;
4701 
4702 	hw_cidx = read_hw_cidx(eq);
4703 	return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx));
4704 }
4705 
4706 static inline u_int
4707 total_available_tx_desc(struct sge_eq *eq)
4708 {
4709 	uint16_t hw_cidx, pidx;
4710 
4711 	hw_cidx = read_hw_cidx(eq);
4712 	pidx = eq->pidx;
4713 
4714 	if (pidx == hw_cidx)
4715 		return (eq->sidx - 1);
4716 	else
4717 		return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1);
4718 }
4719 
4720 static inline uint16_t
4721 read_hw_cidx(struct sge_eq *eq)
4722 {
4723 	struct sge_qstat *spg = (void *)&eq->desc[eq->sidx];
4724 	uint16_t cidx = spg->cidx;	/* stable snapshot */
4725 
4726 	return (be16toh(cidx));
4727 }
4728 
4729 /*
4730  * Reclaim 'n' descriptors approximately.
4731  */
4732 static u_int
4733 reclaim_tx_descs(struct sge_txq *txq, u_int n)
4734 {
4735 	struct tx_sdesc *txsd;
4736 	struct sge_eq *eq = &txq->eq;
4737 	u_int can_reclaim, reclaimed;
4738 
4739 	TXQ_LOCK_ASSERT_OWNED(txq);
4740 	MPASS(n > 0);
4741 
4742 	reclaimed = 0;
4743 	can_reclaim = reclaimable_tx_desc(eq);
4744 	while (can_reclaim && reclaimed < n) {
4745 		int ndesc;
4746 		struct mbuf *m, *nextpkt;
4747 
4748 		txsd = &txq->sdesc[eq->cidx];
4749 		ndesc = txsd->desc_used;
4750 
4751 		/* Firmware doesn't return "partial" credits. */
4752 		KASSERT(can_reclaim >= ndesc,
4753 		    ("%s: unexpected number of credits: %d, %d",
4754 		    __func__, can_reclaim, ndesc));
4755 
4756 		for (m = txsd->m; m != NULL; m = nextpkt) {
4757 			nextpkt = m->m_nextpkt;
4758 			m->m_nextpkt = NULL;
4759 			m_freem(m);
4760 		}
4761 		reclaimed += ndesc;
4762 		can_reclaim -= ndesc;
4763 		IDXINCR(eq->cidx, ndesc, eq->sidx);
4764 	}
4765 
4766 	return (reclaimed);
4767 }
4768 
4769 static void
4770 tx_reclaim(void *arg, int n)
4771 {
4772 	struct sge_txq *txq = arg;
4773 	struct sge_eq *eq = &txq->eq;
4774 
4775 	do {
4776 		if (TXQ_TRYLOCK(txq) == 0)
4777 			break;
4778 		n = reclaim_tx_descs(txq, 32);
4779 		if (eq->cidx == eq->pidx)
4780 			eq->equeqidx = eq->pidx;
4781 		TXQ_UNLOCK(txq);
4782 	} while (n > 0);
4783 }
4784 
4785 static __be64
4786 get_flit(struct sglist_seg *segs, int nsegs, int idx)
4787 {
4788 	int i = (idx / 3) * 2;
4789 
4790 	switch (idx % 3) {
4791 	case 0: {
4792 		uint64_t rc;
4793 
4794 		rc = (uint64_t)segs[i].ss_len << 32;
4795 		if (i + 1 < nsegs)
4796 			rc |= (uint64_t)(segs[i + 1].ss_len);
4797 
4798 		return (htobe64(rc));
4799 	}
4800 	case 1:
4801 		return (htobe64(segs[i].ss_paddr));
4802 	case 2:
4803 		return (htobe64(segs[i + 1].ss_paddr));
4804 	}
4805 
4806 	return (0);
4807 }
4808 
4809 static void
4810 find_best_refill_source(struct adapter *sc, struct sge_fl *fl, int maxp)
4811 {
4812 	int8_t zidx, hwidx, idx;
4813 	uint16_t region1, region3;
4814 	int spare, spare_needed, n;
4815 	struct sw_zone_info *swz;
4816 	struct hw_buf_info *hwb, *hwb_list = &sc->sge.hw_buf_info[0];
4817 
4818 	/*
4819 	 * Buffer Packing: Look for PAGE_SIZE or larger zone which has a bufsize
4820 	 * large enough for the max payload and cluster metadata.  Otherwise
4821 	 * settle for the largest bufsize that leaves enough room in the cluster
4822 	 * for metadata.
4823 	 *
4824 	 * Without buffer packing: Look for the smallest zone which has a
4825 	 * bufsize large enough for the max payload.  Settle for the largest
4826 	 * bufsize available if there's nothing big enough for max payload.
4827 	 */
4828 	spare_needed = fl->flags & FL_BUF_PACKING ? CL_METADATA_SIZE : 0;
4829 	swz = &sc->sge.sw_zone_info[0];
4830 	hwidx = -1;
4831 	for (zidx = 0; zidx < SW_ZONE_SIZES; zidx++, swz++) {
4832 		if (swz->size > largest_rx_cluster) {
4833 			if (__predict_true(hwidx != -1))
4834 				break;
4835 
4836 			/*
4837 			 * This is a misconfiguration.  largest_rx_cluster is
4838 			 * preventing us from finding a refill source.  See
4839 			 * dev.t5nex.<n>.buffer_sizes to figure out why.
4840 			 */
4841 			device_printf(sc->dev, "largest_rx_cluster=%u leaves no"
4842 			    " refill source for fl %p (dma %u).  Ignored.\n",
4843 			    largest_rx_cluster, fl, maxp);
4844 		}
4845 		for (idx = swz->head_hwidx; idx != -1; idx = hwb->next) {
4846 			hwb = &hwb_list[idx];
4847 			spare = swz->size - hwb->size;
4848 			if (spare < spare_needed)
4849 				continue;
4850 
4851 			hwidx = idx;		/* best option so far */
4852 			if (hwb->size >= maxp) {
4853 
4854 				if ((fl->flags & FL_BUF_PACKING) == 0)
4855 					goto done; /* stop looking (not packing) */
4856 
4857 				if (swz->size >= safest_rx_cluster)
4858 					goto done; /* stop looking (packing) */
4859 			}
4860 			break;		/* keep looking, next zone */
4861 		}
4862 	}
4863 done:
4864 	/* A usable hwidx has been located. */
4865 	MPASS(hwidx != -1);
4866 	hwb = &hwb_list[hwidx];
4867 	zidx = hwb->zidx;
4868 	swz = &sc->sge.sw_zone_info[zidx];
4869 	region1 = 0;
4870 	region3 = swz->size - hwb->size;
4871 
4872 	/*
4873 	 * Stay within this zone and see if there is a better match when mbuf
4874 	 * inlining is allowed.  Remember that the hwidx's are sorted in
4875 	 * decreasing order of size (so in increasing order of spare area).
4876 	 */
4877 	for (idx = hwidx; idx != -1; idx = hwb->next) {
4878 		hwb = &hwb_list[idx];
4879 		spare = swz->size - hwb->size;
4880 
4881 		if (allow_mbufs_in_cluster == 0 || hwb->size < maxp)
4882 			break;
4883 
4884 		/*
4885 		 * Do not inline mbufs if doing so would violate the pad/pack
4886 		 * boundary alignment requirement.
4887 		 */
4888 		if (fl_pad && (MSIZE % sc->params.sge.pad_boundary) != 0)
4889 			continue;
4890 		if (fl->flags & FL_BUF_PACKING &&
4891 		    (MSIZE % sc->params.sge.pack_boundary) != 0)
4892 			continue;
4893 
4894 		if (spare < CL_METADATA_SIZE + MSIZE)
4895 			continue;
4896 		n = (spare - CL_METADATA_SIZE) / MSIZE;
4897 		if (n > howmany(hwb->size, maxp))
4898 			break;
4899 
4900 		hwidx = idx;
4901 		if (fl->flags & FL_BUF_PACKING) {
4902 			region1 = n * MSIZE;
4903 			region3 = spare - region1;
4904 		} else {
4905 			region1 = MSIZE;
4906 			region3 = spare - region1;
4907 			break;
4908 		}
4909 	}
4910 
4911 	KASSERT(zidx >= 0 && zidx < SW_ZONE_SIZES,
4912 	    ("%s: bad zone %d for fl %p, maxp %d", __func__, zidx, fl, maxp));
4913 	KASSERT(hwidx >= 0 && hwidx <= SGE_FLBUF_SIZES,
4914 	    ("%s: bad hwidx %d for fl %p, maxp %d", __func__, hwidx, fl, maxp));
4915 	KASSERT(region1 + sc->sge.hw_buf_info[hwidx].size + region3 ==
4916 	    sc->sge.sw_zone_info[zidx].size,
4917 	    ("%s: bad buffer layout for fl %p, maxp %d. "
4918 		"cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
4919 		sc->sge.sw_zone_info[zidx].size, region1,
4920 		sc->sge.hw_buf_info[hwidx].size, region3));
4921 	if (fl->flags & FL_BUF_PACKING || region1 > 0) {
4922 		KASSERT(region3 >= CL_METADATA_SIZE,
4923 		    ("%s: no room for metadata.  fl %p, maxp %d; "
4924 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
4925 		    sc->sge.sw_zone_info[zidx].size, region1,
4926 		    sc->sge.hw_buf_info[hwidx].size, region3));
4927 		KASSERT(region1 % MSIZE == 0,
4928 		    ("%s: bad mbuf region for fl %p, maxp %d. "
4929 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
4930 		    sc->sge.sw_zone_info[zidx].size, region1,
4931 		    sc->sge.hw_buf_info[hwidx].size, region3));
4932 	}
4933 
4934 	fl->cll_def.zidx = zidx;
4935 	fl->cll_def.hwidx = hwidx;
4936 	fl->cll_def.region1 = region1;
4937 	fl->cll_def.region3 = region3;
4938 }
4939 
4940 static void
4941 find_safe_refill_source(struct adapter *sc, struct sge_fl *fl)
4942 {
4943 	struct sge *s = &sc->sge;
4944 	struct hw_buf_info *hwb;
4945 	struct sw_zone_info *swz;
4946 	int spare;
4947 	int8_t hwidx;
4948 
4949 	if (fl->flags & FL_BUF_PACKING)
4950 		hwidx = s->safe_hwidx2;	/* with room for metadata */
4951 	else if (allow_mbufs_in_cluster && s->safe_hwidx2 != -1) {
4952 		hwidx = s->safe_hwidx2;
4953 		hwb = &s->hw_buf_info[hwidx];
4954 		swz = &s->sw_zone_info[hwb->zidx];
4955 		spare = swz->size - hwb->size;
4956 
4957 		/* no good if there isn't room for an mbuf as well */
4958 		if (spare < CL_METADATA_SIZE + MSIZE)
4959 			hwidx = s->safe_hwidx1;
4960 	} else
4961 		hwidx = s->safe_hwidx1;
4962 
4963 	if (hwidx == -1) {
4964 		/* No fallback source */
4965 		fl->cll_alt.hwidx = -1;
4966 		fl->cll_alt.zidx = -1;
4967 
4968 		return;
4969 	}
4970 
4971 	hwb = &s->hw_buf_info[hwidx];
4972 	swz = &s->sw_zone_info[hwb->zidx];
4973 	spare = swz->size - hwb->size;
4974 	fl->cll_alt.hwidx = hwidx;
4975 	fl->cll_alt.zidx = hwb->zidx;
4976 	if (allow_mbufs_in_cluster &&
4977 	    (fl_pad == 0 || (MSIZE % sc->params.sge.pad_boundary) == 0))
4978 		fl->cll_alt.region1 = ((spare - CL_METADATA_SIZE) / MSIZE) * MSIZE;
4979 	else
4980 		fl->cll_alt.region1 = 0;
4981 	fl->cll_alt.region3 = spare - fl->cll_alt.region1;
4982 }
4983 
4984 static void
4985 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
4986 {
4987 	mtx_lock(&sc->sfl_lock);
4988 	FL_LOCK(fl);
4989 	if ((fl->flags & FL_DOOMED) == 0) {
4990 		fl->flags |= FL_STARVING;
4991 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
4992 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
4993 	}
4994 	FL_UNLOCK(fl);
4995 	mtx_unlock(&sc->sfl_lock);
4996 }
4997 
4998 static void
4999 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq)
5000 {
5001 	struct sge_wrq *wrq = (void *)eq;
5002 
5003 	atomic_readandclear_int(&eq->equiq);
5004 	taskqueue_enqueue(sc->tq[eq->tx_chan], &wrq->wrq_tx_task);
5005 }
5006 
5007 static void
5008 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq)
5009 {
5010 	struct sge_txq *txq = (void *)eq;
5011 
5012 	MPASS((eq->flags & EQ_TYPEMASK) == EQ_ETH);
5013 
5014 	atomic_readandclear_int(&eq->equiq);
5015 	mp_ring_check_drainage(txq->r, 0);
5016 	taskqueue_enqueue(sc->tq[eq->tx_chan], &txq->tx_reclaim_task);
5017 }
5018 
5019 static int
5020 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
5021     struct mbuf *m)
5022 {
5023 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
5024 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
5025 	struct adapter *sc = iq->adapter;
5026 	struct sge *s = &sc->sge;
5027 	struct sge_eq *eq;
5028 	static void (*h[])(struct adapter *, struct sge_eq *) = {NULL,
5029 		&handle_wrq_egr_update, &handle_eth_egr_update,
5030 		&handle_wrq_egr_update};
5031 
5032 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5033 	    rss->opcode));
5034 
5035 	eq = s->eqmap[qid - s->eq_start - s->eq_base];
5036 	(*h[eq->flags & EQ_TYPEMASK])(sc, eq);
5037 
5038 	return (0);
5039 }
5040 
5041 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
5042 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
5043     offsetof(struct cpl_fw6_msg, data));
5044 
5045 static int
5046 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
5047 {
5048 	struct adapter *sc = iq->adapter;
5049 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
5050 
5051 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5052 	    rss->opcode));
5053 
5054 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
5055 		const struct rss_header *rss2;
5056 
5057 		rss2 = (const struct rss_header *)&cpl->data[0];
5058 		return (t4_cpl_handler[rss2->opcode](iq, rss2, m));
5059 	}
5060 
5061 	return (t4_fw_msg_handler[cpl->type](sc, &cpl->data[0]));
5062 }
5063 
5064 /**
5065  *	t4_handle_wrerr_rpl - process a FW work request error message
5066  *	@adap: the adapter
5067  *	@rpl: start of the FW message
5068  */
5069 static int
5070 t4_handle_wrerr_rpl(struct adapter *adap, const __be64 *rpl)
5071 {
5072 	u8 opcode = *(const u8 *)rpl;
5073 	const struct fw_error_cmd *e = (const void *)rpl;
5074 	unsigned int i;
5075 
5076 	if (opcode != FW_ERROR_CMD) {
5077 		log(LOG_ERR,
5078 		    "%s: Received WRERR_RPL message with opcode %#x\n",
5079 		    device_get_nameunit(adap->dev), opcode);
5080 		return (EINVAL);
5081 	}
5082 	log(LOG_ERR, "%s: FW_ERROR (%s) ", device_get_nameunit(adap->dev),
5083 	    G_FW_ERROR_CMD_FATAL(be32toh(e->op_to_type)) ? "fatal" :
5084 	    "non-fatal");
5085 	switch (G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type))) {
5086 	case FW_ERROR_TYPE_EXCEPTION:
5087 		log(LOG_ERR, "exception info:\n");
5088 		for (i = 0; i < nitems(e->u.exception.info); i++)
5089 			log(LOG_ERR, "%s%08x", i == 0 ? "\t" : " ",
5090 			    be32toh(e->u.exception.info[i]));
5091 		log(LOG_ERR, "\n");
5092 		break;
5093 	case FW_ERROR_TYPE_HWMODULE:
5094 		log(LOG_ERR, "HW module regaddr %08x regval %08x\n",
5095 		    be32toh(e->u.hwmodule.regaddr),
5096 		    be32toh(e->u.hwmodule.regval));
5097 		break;
5098 	case FW_ERROR_TYPE_WR:
5099 		log(LOG_ERR, "WR cidx %d PF %d VF %d eqid %d hdr:\n",
5100 		    be16toh(e->u.wr.cidx),
5101 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.wr.pfn_vfn)),
5102 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.wr.pfn_vfn)),
5103 		    be32toh(e->u.wr.eqid));
5104 		for (i = 0; i < nitems(e->u.wr.wrhdr); i++)
5105 			log(LOG_ERR, "%s%02x", i == 0 ? "\t" : " ",
5106 			    e->u.wr.wrhdr[i]);
5107 		log(LOG_ERR, "\n");
5108 		break;
5109 	case FW_ERROR_TYPE_ACL:
5110 		log(LOG_ERR, "ACL cidx %d PF %d VF %d eqid %d %s",
5111 		    be16toh(e->u.acl.cidx),
5112 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.acl.pfn_vfn)),
5113 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.acl.pfn_vfn)),
5114 		    be32toh(e->u.acl.eqid),
5115 		    G_FW_ERROR_CMD_MV(be16toh(e->u.acl.mv_pkd)) ? "vlanid" :
5116 		    "MAC");
5117 		for (i = 0; i < nitems(e->u.acl.val); i++)
5118 			log(LOG_ERR, " %02x", e->u.acl.val[i]);
5119 		log(LOG_ERR, "\n");
5120 		break;
5121 	default:
5122 		log(LOG_ERR, "type %#x\n",
5123 		    G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type)));
5124 		return (EINVAL);
5125 	}
5126 	return (0);
5127 }
5128 
5129 static int
5130 sysctl_uint16(SYSCTL_HANDLER_ARGS)
5131 {
5132 	uint16_t *id = arg1;
5133 	int i = *id;
5134 
5135 	return sysctl_handle_int(oidp, &i, 0, req);
5136 }
5137 
5138 static int
5139 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
5140 {
5141 	struct sge *s = arg1;
5142 	struct hw_buf_info *hwb = &s->hw_buf_info[0];
5143 	struct sw_zone_info *swz = &s->sw_zone_info[0];
5144 	int i, rc;
5145 	struct sbuf sb;
5146 	char c;
5147 
5148 	sbuf_new(&sb, NULL, 32, SBUF_AUTOEXTEND);
5149 	for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
5150 		if (hwb->zidx >= 0 && swz[hwb->zidx].size <= largest_rx_cluster)
5151 			c = '*';
5152 		else
5153 			c = '\0';
5154 
5155 		sbuf_printf(&sb, "%u%c ", hwb->size, c);
5156 	}
5157 	sbuf_trim(&sb);
5158 	sbuf_finish(&sb);
5159 	rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
5160 	sbuf_delete(&sb);
5161 	return (rc);
5162 }
5163 
5164 static int
5165 sysctl_tc(SYSCTL_HANDLER_ARGS)
5166 {
5167 	struct vi_info *vi = arg1;
5168 	struct port_info *pi;
5169 	struct adapter *sc;
5170 	struct sge_txq *txq;
5171 	struct tx_cl_rl_params *tc;
5172 	int qidx = arg2, rc, tc_idx;
5173 	uint32_t fw_queue, fw_class;
5174 
5175 	MPASS(qidx >= 0 && qidx < vi->ntxq);
5176 	pi = vi->pi;
5177 	sc = pi->adapter;
5178 	txq = &sc->sge.txq[vi->first_txq + qidx];
5179 
5180 	tc_idx = txq->tc_idx;
5181 	rc = sysctl_handle_int(oidp, &tc_idx, 0, req);
5182 	if (rc != 0 || req->newptr == NULL)
5183 		return (rc);
5184 
5185 	if (sc->flags & IS_VF)
5186 		return (EPERM);
5187 
5188 	/* Note that -1 is legitimate input (it means unbind). */
5189 	if (tc_idx < -1 || tc_idx >= sc->chip_params->nsched_cls)
5190 		return (EINVAL);
5191 
5192 	mtx_lock(&sc->tc_lock);
5193 	if (tc_idx == txq->tc_idx) {
5194 		rc = 0;		/* No change, nothing to do. */
5195 		goto done;
5196 	}
5197 
5198 	fw_queue = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
5199 	    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_EQ_SCHEDCLASS_ETH) |
5200 	    V_FW_PARAMS_PARAM_YZ(txq->eq.cntxt_id);
5201 
5202 	if (tc_idx == -1)
5203 		fw_class = 0xffffffff;	/* Unbind. */
5204 	else {
5205 		/*
5206 		 * Bind to a different class.
5207 		 */
5208 		tc = &pi->sched_params->cl_rl[tc_idx];
5209 		if (tc->flags & TX_CLRL_ERROR) {
5210 			/* Previous attempt to set the cl-rl params failed. */
5211 			rc = EIO;
5212 			goto done;
5213 		} else {
5214 			/*
5215 			 * Ok to proceed.  Place a reference on the new class
5216 			 * while still holding on to the reference on the
5217 			 * previous class, if any.
5218 			 */
5219 			fw_class = tc_idx;
5220 			tc->refcount++;
5221 		}
5222 	}
5223 	mtx_unlock(&sc->tc_lock);
5224 
5225 	rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4stc");
5226 	if (rc)
5227 		return (rc);
5228 	rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &fw_queue, &fw_class);
5229 	end_synchronized_op(sc, 0);
5230 
5231 	mtx_lock(&sc->tc_lock);
5232 	if (rc == 0) {
5233 		if (txq->tc_idx != -1) {
5234 			tc = &pi->sched_params->cl_rl[txq->tc_idx];
5235 			MPASS(tc->refcount > 0);
5236 			tc->refcount--;
5237 		}
5238 		txq->tc_idx = tc_idx;
5239 	} else if (tc_idx != -1) {
5240 		tc = &pi->sched_params->cl_rl[tc_idx];
5241 		MPASS(tc->refcount > 0);
5242 		tc->refcount--;
5243 	}
5244 done:
5245 	mtx_unlock(&sc->tc_lock);
5246 	return (rc);
5247 }
5248