xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision cfd6422a5217410fbd66f7a7a8a64d9d85e61229)
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 #include "opt_kern_tls.h"
36 #include "opt_ratelimit.h"
37 
38 #include <sys/types.h>
39 #include <sys/eventhandler.h>
40 #include <sys/mbuf.h>
41 #include <sys/socket.h>
42 #include <sys/kernel.h>
43 #include <sys/ktls.h>
44 #include <sys/malloc.h>
45 #include <sys/queue.h>
46 #include <sys/sbuf.h>
47 #include <sys/taskqueue.h>
48 #include <sys/time.h>
49 #include <sys/sglist.h>
50 #include <sys/sysctl.h>
51 #include <sys/smp.h>
52 #include <sys/socketvar.h>
53 #include <sys/counter.h>
54 #include <net/bpf.h>
55 #include <net/ethernet.h>
56 #include <net/if.h>
57 #include <net/if_vlan_var.h>
58 #include <net/if_vxlan.h>
59 #include <netinet/in.h>
60 #include <netinet/ip.h>
61 #include <netinet/ip6.h>
62 #include <netinet/tcp.h>
63 #include <netinet/udp.h>
64 #include <machine/in_cksum.h>
65 #include <machine/md_var.h>
66 #include <vm/vm.h>
67 #include <vm/pmap.h>
68 #ifdef DEV_NETMAP
69 #include <machine/bus.h>
70 #include <sys/selinfo.h>
71 #include <net/if_var.h>
72 #include <net/netmap.h>
73 #include <dev/netmap/netmap_kern.h>
74 #endif
75 
76 #include "common/common.h"
77 #include "common/t4_regs.h"
78 #include "common/t4_regs_values.h"
79 #include "common/t4_msg.h"
80 #include "t4_l2t.h"
81 #include "t4_mp_ring.h"
82 
83 #ifdef T4_PKT_TIMESTAMP
84 #define RX_COPY_THRESHOLD (MINCLSIZE - 8)
85 #else
86 #define RX_COPY_THRESHOLD MINCLSIZE
87 #endif
88 
89 /* Internal mbuf flags stored in PH_loc.eight[1]. */
90 #define	MC_NOMAP		0x01
91 #define	MC_RAW_WR		0x02
92 #define	MC_TLS			0x04
93 
94 /*
95  * Ethernet frames are DMA'd at this byte offset into the freelist buffer.
96  * 0-7 are valid values.
97  */
98 static int fl_pktshift = 0;
99 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pktshift, CTLFLAG_RDTUN, &fl_pktshift, 0,
100     "payload DMA offset in rx buffer (bytes)");
101 
102 /*
103  * Pad ethernet payload up to this boundary.
104  * -1: driver should figure out a good value.
105  *  0: disable padding.
106  *  Any power of 2 from 32 to 4096 (both inclusive) is also a valid value.
107  */
108 int fl_pad = -1;
109 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pad, CTLFLAG_RDTUN, &fl_pad, 0,
110     "payload pad boundary (bytes)");
111 
112 /*
113  * Status page length.
114  * -1: driver should figure out a good value.
115  *  64 or 128 are the only other valid values.
116  */
117 static int spg_len = -1;
118 SYSCTL_INT(_hw_cxgbe, OID_AUTO, spg_len, CTLFLAG_RDTUN, &spg_len, 0,
119     "status page size (bytes)");
120 
121 /*
122  * Congestion drops.
123  * -1: no congestion feedback (not recommended).
124  *  0: backpressure the channel instead of dropping packets right away.
125  *  1: no backpressure, drop packets for the congested queue immediately.
126  */
127 static int cong_drop = 0;
128 SYSCTL_INT(_hw_cxgbe, OID_AUTO, cong_drop, CTLFLAG_RDTUN, &cong_drop, 0,
129     "Congestion control for RX queues (0 = backpressure, 1 = drop");
130 
131 /*
132  * Deliver multiple frames in the same free list buffer if they fit.
133  * -1: let the driver decide whether to enable buffer packing or not.
134  *  0: disable buffer packing.
135  *  1: enable buffer packing.
136  */
137 static int buffer_packing = -1;
138 SYSCTL_INT(_hw_cxgbe, OID_AUTO, buffer_packing, CTLFLAG_RDTUN, &buffer_packing,
139     0, "Enable buffer packing");
140 
141 /*
142  * Start next frame in a packed buffer at this boundary.
143  * -1: driver should figure out a good value.
144  * T4: driver will ignore this and use the same value as fl_pad above.
145  * T5: 16, or a power of 2 from 64 to 4096 (both inclusive) is a valid value.
146  */
147 static int fl_pack = -1;
148 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pack, CTLFLAG_RDTUN, &fl_pack, 0,
149     "payload pack boundary (bytes)");
150 
151 /*
152  * Largest rx cluster size that the driver is allowed to allocate.
153  */
154 static int largest_rx_cluster = MJUM16BYTES;
155 SYSCTL_INT(_hw_cxgbe, OID_AUTO, largest_rx_cluster, CTLFLAG_RDTUN,
156     &largest_rx_cluster, 0, "Largest rx cluster (bytes)");
157 
158 /*
159  * Size of cluster allocation that's most likely to succeed.  The driver will
160  * fall back to this size if it fails to allocate clusters larger than this.
161  */
162 static int safest_rx_cluster = PAGE_SIZE;
163 SYSCTL_INT(_hw_cxgbe, OID_AUTO, safest_rx_cluster, CTLFLAG_RDTUN,
164     &safest_rx_cluster, 0, "Safe rx cluster (bytes)");
165 
166 #ifdef RATELIMIT
167 /*
168  * Knob to control TCP timestamp rewriting, and the granularity of the tick used
169  * for rewriting.  -1 and 0-3 are all valid values.
170  * -1: hardware should leave the TCP timestamps alone.
171  * 0: 1ms
172  * 1: 100us
173  * 2: 10us
174  * 3: 1us
175  */
176 static int tsclk = -1;
177 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tsclk, CTLFLAG_RDTUN, &tsclk, 0,
178     "Control TCP timestamp rewriting when using pacing");
179 
180 static int eo_max_backlog = 1024 * 1024;
181 SYSCTL_INT(_hw_cxgbe, OID_AUTO, eo_max_backlog, CTLFLAG_RDTUN, &eo_max_backlog,
182     0, "Maximum backlog of ratelimited data per flow");
183 #endif
184 
185 /*
186  * The interrupt holdoff timers are multiplied by this value on T6+.
187  * 1 and 3-17 (both inclusive) are legal values.
188  */
189 static int tscale = 1;
190 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tscale, CTLFLAG_RDTUN, &tscale, 0,
191     "Interrupt holdoff timer scale on T6+");
192 
193 /*
194  * Number of LRO entries in the lro_ctrl structure per rx queue.
195  */
196 static int lro_entries = TCP_LRO_ENTRIES;
197 SYSCTL_INT(_hw_cxgbe, OID_AUTO, lro_entries, CTLFLAG_RDTUN, &lro_entries, 0,
198     "Number of LRO entries per RX queue");
199 
200 /*
201  * This enables presorting of frames before they're fed into tcp_lro_rx.
202  */
203 static int lro_mbufs = 0;
204 SYSCTL_INT(_hw_cxgbe, OID_AUTO, lro_mbufs, CTLFLAG_RDTUN, &lro_mbufs, 0,
205     "Enable presorting of LRO frames");
206 
207 static counter_u64_t pullups;
208 SYSCTL_COUNTER_U64(_hw_cxgbe, OID_AUTO, pullups, CTLFLAG_RD, &pullups,
209     "Number of mbuf pullups performed");
210 
211 static counter_u64_t defrags;
212 SYSCTL_COUNTER_U64(_hw_cxgbe, OID_AUTO, defrags, CTLFLAG_RD, &defrags,
213     "Number of mbuf defrags performed");
214 
215 
216 static int service_iq(struct sge_iq *, int);
217 static int service_iq_fl(struct sge_iq *, int);
218 static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t);
219 static int eth_rx(struct adapter *, struct sge_rxq *, const struct iq_desc *,
220     u_int);
221 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int);
222 static inline void init_fl(struct adapter *, struct sge_fl *, int, int, char *);
223 static inline void init_eq(struct adapter *, struct sge_eq *, int, int, uint8_t,
224     uint16_t, char *);
225 static int alloc_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *,
226     int, int);
227 static int free_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *);
228 static void add_iq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
229     struct sge_iq *);
230 static void add_fl_sysctls(struct adapter *, struct sysctl_ctx_list *,
231     struct sysctl_oid *, struct sge_fl *);
232 static int alloc_fwq(struct adapter *);
233 static int free_fwq(struct adapter *);
234 static int alloc_ctrlq(struct adapter *, struct sge_wrq *, int,
235     struct sysctl_oid *);
236 static int alloc_rxq(struct vi_info *, struct sge_rxq *, int, int,
237     struct sysctl_oid *);
238 static int free_rxq(struct vi_info *, struct sge_rxq *);
239 #ifdef TCP_OFFLOAD
240 static int alloc_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *, int, int,
241     struct sysctl_oid *);
242 static int free_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *);
243 #endif
244 static int ctrl_eq_alloc(struct adapter *, struct sge_eq *);
245 static int eth_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
246 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
247 static int ofld_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
248 #endif
249 static int alloc_eq(struct adapter *, struct vi_info *, struct sge_eq *);
250 static int free_eq(struct adapter *, struct sge_eq *);
251 static int alloc_wrq(struct adapter *, struct vi_info *, struct sge_wrq *,
252     struct sysctl_oid *);
253 static int free_wrq(struct adapter *, struct sge_wrq *);
254 static int alloc_txq(struct vi_info *, struct sge_txq *, int,
255     struct sysctl_oid *);
256 static int free_txq(struct vi_info *, struct sge_txq *);
257 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int);
258 static inline void ring_fl_db(struct adapter *, struct sge_fl *);
259 static int refill_fl(struct adapter *, struct sge_fl *, int);
260 static void refill_sfl(void *);
261 static int alloc_fl_sdesc(struct sge_fl *);
262 static void free_fl_sdesc(struct adapter *, struct sge_fl *);
263 static int find_refill_source(struct adapter *, int, bool);
264 static void add_fl_to_sfl(struct adapter *, struct sge_fl *);
265 
266 static inline void get_pkt_gl(struct mbuf *, struct sglist *);
267 static inline u_int txpkt_len16(u_int, const u_int);
268 static inline u_int txpkt_vm_len16(u_int, const u_int);
269 static inline void calculate_mbuf_len16(struct mbuf *, bool);
270 static inline u_int txpkts0_len16(u_int);
271 static inline u_int txpkts1_len16(void);
272 static u_int write_raw_wr(struct sge_txq *, void *, struct mbuf *, u_int);
273 static u_int write_txpkt_wr(struct adapter *, struct sge_txq *, struct mbuf *,
274     u_int);
275 static u_int write_txpkt_vm_wr(struct adapter *, struct sge_txq *,
276     struct mbuf *);
277 static int add_to_txpkts_vf(struct adapter *, struct sge_txq *, struct mbuf *,
278     int, bool *);
279 static int add_to_txpkts_pf(struct adapter *, struct sge_txq *, struct mbuf *,
280     int, bool *);
281 static u_int write_txpkts_wr(struct adapter *, struct sge_txq *);
282 static u_int write_txpkts_vm_wr(struct adapter *, struct sge_txq *);
283 static void write_gl_to_txd(struct sge_txq *, struct mbuf *, caddr_t *, int);
284 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int);
285 static inline void ring_eq_db(struct adapter *, struct sge_eq *, u_int);
286 static inline uint16_t read_hw_cidx(struct sge_eq *);
287 static inline u_int reclaimable_tx_desc(struct sge_eq *);
288 static inline u_int total_available_tx_desc(struct sge_eq *);
289 static u_int reclaim_tx_descs(struct sge_txq *, u_int);
290 static void tx_reclaim(void *, int);
291 static __be64 get_flit(struct sglist_seg *, int, int);
292 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
293     struct mbuf *);
294 static int handle_fw_msg(struct sge_iq *, const struct rss_header *,
295     struct mbuf *);
296 static int t4_handle_wrerr_rpl(struct adapter *, const __be64 *);
297 static void wrq_tx_drain(void *, int);
298 static void drain_wrq_wr_list(struct adapter *, struct sge_wrq *);
299 
300 static int sysctl_bufsizes(SYSCTL_HANDLER_ARGS);
301 #ifdef RATELIMIT
302 static inline u_int txpkt_eo_len16(u_int, u_int, u_int);
303 static int ethofld_fw4_ack(struct sge_iq *, const struct rss_header *,
304     struct mbuf *);
305 #endif
306 
307 static counter_u64_t extfree_refs;
308 static counter_u64_t extfree_rels;
309 
310 an_handler_t t4_an_handler;
311 fw_msg_handler_t t4_fw_msg_handler[NUM_FW6_TYPES];
312 cpl_handler_t t4_cpl_handler[NUM_CPL_CMDS];
313 cpl_handler_t set_tcb_rpl_handlers[NUM_CPL_COOKIES];
314 cpl_handler_t l2t_write_rpl_handlers[NUM_CPL_COOKIES];
315 cpl_handler_t act_open_rpl_handlers[NUM_CPL_COOKIES];
316 cpl_handler_t abort_rpl_rss_handlers[NUM_CPL_COOKIES];
317 cpl_handler_t fw4_ack_handlers[NUM_CPL_COOKIES];
318 
319 void
320 t4_register_an_handler(an_handler_t h)
321 {
322 	uintptr_t *loc;
323 
324 	MPASS(h == NULL || t4_an_handler == NULL);
325 
326 	loc = (uintptr_t *)&t4_an_handler;
327 	atomic_store_rel_ptr(loc, (uintptr_t)h);
328 }
329 
330 void
331 t4_register_fw_msg_handler(int type, fw_msg_handler_t h)
332 {
333 	uintptr_t *loc;
334 
335 	MPASS(type < nitems(t4_fw_msg_handler));
336 	MPASS(h == NULL || t4_fw_msg_handler[type] == NULL);
337 	/*
338 	 * These are dispatched by the handler for FW{4|6}_CPL_MSG using the CPL
339 	 * handler dispatch table.  Reject any attempt to install a handler for
340 	 * this subtype.
341 	 */
342 	MPASS(type != FW_TYPE_RSSCPL);
343 	MPASS(type != FW6_TYPE_RSSCPL);
344 
345 	loc = (uintptr_t *)&t4_fw_msg_handler[type];
346 	atomic_store_rel_ptr(loc, (uintptr_t)h);
347 }
348 
349 void
350 t4_register_cpl_handler(int opcode, cpl_handler_t h)
351 {
352 	uintptr_t *loc;
353 
354 	MPASS(opcode < nitems(t4_cpl_handler));
355 	MPASS(h == NULL || t4_cpl_handler[opcode] == NULL);
356 
357 	loc = (uintptr_t *)&t4_cpl_handler[opcode];
358 	atomic_store_rel_ptr(loc, (uintptr_t)h);
359 }
360 
361 static int
362 set_tcb_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
363     struct mbuf *m)
364 {
365 	const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
366 	u_int tid;
367 	int cookie;
368 
369 	MPASS(m == NULL);
370 
371 	tid = GET_TID(cpl);
372 	if (is_hpftid(iq->adapter, tid) || is_ftid(iq->adapter, tid)) {
373 		/*
374 		 * The return code for filter-write is put in the CPL cookie so
375 		 * we have to rely on the hardware tid (is_ftid) to determine
376 		 * that this is a response to a filter.
377 		 */
378 		cookie = CPL_COOKIE_FILTER;
379 	} else {
380 		cookie = G_COOKIE(cpl->cookie);
381 	}
382 	MPASS(cookie > CPL_COOKIE_RESERVED);
383 	MPASS(cookie < nitems(set_tcb_rpl_handlers));
384 
385 	return (set_tcb_rpl_handlers[cookie](iq, rss, m));
386 }
387 
388 static int
389 l2t_write_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
390     struct mbuf *m)
391 {
392 	const struct cpl_l2t_write_rpl *rpl = (const void *)(rss + 1);
393 	unsigned int cookie;
394 
395 	MPASS(m == NULL);
396 
397 	cookie = GET_TID(rpl) & F_SYNC_WR ? CPL_COOKIE_TOM : CPL_COOKIE_FILTER;
398 	return (l2t_write_rpl_handlers[cookie](iq, rss, m));
399 }
400 
401 static int
402 act_open_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
403     struct mbuf *m)
404 {
405 	const struct cpl_act_open_rpl *cpl = (const void *)(rss + 1);
406 	u_int cookie = G_TID_COOKIE(G_AOPEN_ATID(be32toh(cpl->atid_status)));
407 
408 	MPASS(m == NULL);
409 	MPASS(cookie != CPL_COOKIE_RESERVED);
410 
411 	return (act_open_rpl_handlers[cookie](iq, rss, m));
412 }
413 
414 static int
415 abort_rpl_rss_handler(struct sge_iq *iq, const struct rss_header *rss,
416     struct mbuf *m)
417 {
418 	struct adapter *sc = iq->adapter;
419 	u_int cookie;
420 
421 	MPASS(m == NULL);
422 	if (is_hashfilter(sc))
423 		cookie = CPL_COOKIE_HASHFILTER;
424 	else
425 		cookie = CPL_COOKIE_TOM;
426 
427 	return (abort_rpl_rss_handlers[cookie](iq, rss, m));
428 }
429 
430 static int
431 fw4_ack_handler(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
432 {
433 	struct adapter *sc = iq->adapter;
434 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
435 	unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
436 	u_int cookie;
437 
438 	MPASS(m == NULL);
439 	if (is_etid(sc, tid))
440 		cookie = CPL_COOKIE_ETHOFLD;
441 	else
442 		cookie = CPL_COOKIE_TOM;
443 
444 	return (fw4_ack_handlers[cookie](iq, rss, m));
445 }
446 
447 static void
448 t4_init_shared_cpl_handlers(void)
449 {
450 
451 	t4_register_cpl_handler(CPL_SET_TCB_RPL, set_tcb_rpl_handler);
452 	t4_register_cpl_handler(CPL_L2T_WRITE_RPL, l2t_write_rpl_handler);
453 	t4_register_cpl_handler(CPL_ACT_OPEN_RPL, act_open_rpl_handler);
454 	t4_register_cpl_handler(CPL_ABORT_RPL_RSS, abort_rpl_rss_handler);
455 	t4_register_cpl_handler(CPL_FW4_ACK, fw4_ack_handler);
456 }
457 
458 void
459 t4_register_shared_cpl_handler(int opcode, cpl_handler_t h, int cookie)
460 {
461 	uintptr_t *loc;
462 
463 	MPASS(opcode < nitems(t4_cpl_handler));
464 	MPASS(cookie > CPL_COOKIE_RESERVED);
465 	MPASS(cookie < NUM_CPL_COOKIES);
466 	MPASS(t4_cpl_handler[opcode] != NULL);
467 
468 	switch (opcode) {
469 	case CPL_SET_TCB_RPL:
470 		loc = (uintptr_t *)&set_tcb_rpl_handlers[cookie];
471 		break;
472 	case CPL_L2T_WRITE_RPL:
473 		loc = (uintptr_t *)&l2t_write_rpl_handlers[cookie];
474 		break;
475 	case CPL_ACT_OPEN_RPL:
476 		loc = (uintptr_t *)&act_open_rpl_handlers[cookie];
477 		break;
478 	case CPL_ABORT_RPL_RSS:
479 		loc = (uintptr_t *)&abort_rpl_rss_handlers[cookie];
480 		break;
481 	case CPL_FW4_ACK:
482 		loc = (uintptr_t *)&fw4_ack_handlers[cookie];
483 		break;
484 	default:
485 		MPASS(0);
486 		return;
487 	}
488 	MPASS(h == NULL || *loc == (uintptr_t)NULL);
489 	atomic_store_rel_ptr(loc, (uintptr_t)h);
490 }
491 
492 /*
493  * Called on MOD_LOAD.  Validates and calculates the SGE tunables.
494  */
495 void
496 t4_sge_modload(void)
497 {
498 
499 	if (fl_pktshift < 0 || fl_pktshift > 7) {
500 		printf("Invalid hw.cxgbe.fl_pktshift value (%d),"
501 		    " using 0 instead.\n", fl_pktshift);
502 		fl_pktshift = 0;
503 	}
504 
505 	if (spg_len != 64 && spg_len != 128) {
506 		int len;
507 
508 #if defined(__i386__) || defined(__amd64__)
509 		len = cpu_clflush_line_size > 64 ? 128 : 64;
510 #else
511 		len = 64;
512 #endif
513 		if (spg_len != -1) {
514 			printf("Invalid hw.cxgbe.spg_len value (%d),"
515 			    " using %d instead.\n", spg_len, len);
516 		}
517 		spg_len = len;
518 	}
519 
520 	if (cong_drop < -1 || cong_drop > 1) {
521 		printf("Invalid hw.cxgbe.cong_drop value (%d),"
522 		    " using 0 instead.\n", cong_drop);
523 		cong_drop = 0;
524 	}
525 
526 	if (tscale != 1 && (tscale < 3 || tscale > 17)) {
527 		printf("Invalid hw.cxgbe.tscale value (%d),"
528 		    " using 1 instead.\n", tscale);
529 		tscale = 1;
530 	}
531 
532 	if (largest_rx_cluster != MCLBYTES &&
533 #if MJUMPAGESIZE != MCLBYTES
534 	    largest_rx_cluster != MJUMPAGESIZE &&
535 #endif
536 	    largest_rx_cluster != MJUM9BYTES &&
537 	    largest_rx_cluster != MJUM16BYTES) {
538 		printf("Invalid hw.cxgbe.largest_rx_cluster value (%d),"
539 		    " using %d instead.\n", largest_rx_cluster, MJUM16BYTES);
540 		largest_rx_cluster = MJUM16BYTES;
541 	}
542 
543 	if (safest_rx_cluster != MCLBYTES &&
544 #if MJUMPAGESIZE != MCLBYTES
545 	    safest_rx_cluster != MJUMPAGESIZE &&
546 #endif
547 	    safest_rx_cluster != MJUM9BYTES &&
548 	    safest_rx_cluster != MJUM16BYTES) {
549 		printf("Invalid hw.cxgbe.safest_rx_cluster value (%d),"
550 		    " using %d instead.\n", safest_rx_cluster, MJUMPAGESIZE);
551 		safest_rx_cluster = MJUMPAGESIZE;
552 	}
553 
554 	extfree_refs = counter_u64_alloc(M_WAITOK);
555 	extfree_rels = counter_u64_alloc(M_WAITOK);
556 	pullups = counter_u64_alloc(M_WAITOK);
557 	defrags = counter_u64_alloc(M_WAITOK);
558 	counter_u64_zero(extfree_refs);
559 	counter_u64_zero(extfree_rels);
560 	counter_u64_zero(pullups);
561 	counter_u64_zero(defrags);
562 
563 	t4_init_shared_cpl_handlers();
564 	t4_register_cpl_handler(CPL_FW4_MSG, handle_fw_msg);
565 	t4_register_cpl_handler(CPL_FW6_MSG, handle_fw_msg);
566 	t4_register_cpl_handler(CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
567 #ifdef RATELIMIT
568 	t4_register_shared_cpl_handler(CPL_FW4_ACK, ethofld_fw4_ack,
569 	    CPL_COOKIE_ETHOFLD);
570 #endif
571 	t4_register_fw_msg_handler(FW6_TYPE_CMD_RPL, t4_handle_fw_rpl);
572 	t4_register_fw_msg_handler(FW6_TYPE_WRERR_RPL, t4_handle_wrerr_rpl);
573 }
574 
575 void
576 t4_sge_modunload(void)
577 {
578 
579 	counter_u64_free(extfree_refs);
580 	counter_u64_free(extfree_rels);
581 	counter_u64_free(pullups);
582 	counter_u64_free(defrags);
583 }
584 
585 uint64_t
586 t4_sge_extfree_refs(void)
587 {
588 	uint64_t refs, rels;
589 
590 	rels = counter_u64_fetch(extfree_rels);
591 	refs = counter_u64_fetch(extfree_refs);
592 
593 	return (refs - rels);
594 }
595 
596 /* max 4096 */
597 #define MAX_PACK_BOUNDARY 512
598 
599 static inline void
600 setup_pad_and_pack_boundaries(struct adapter *sc)
601 {
602 	uint32_t v, m;
603 	int pad, pack, pad_shift;
604 
605 	pad_shift = chip_id(sc) > CHELSIO_T5 ? X_T6_INGPADBOUNDARY_SHIFT :
606 	    X_INGPADBOUNDARY_SHIFT;
607 	pad = fl_pad;
608 	if (fl_pad < (1 << pad_shift) ||
609 	    fl_pad > (1 << (pad_shift + M_INGPADBOUNDARY)) ||
610 	    !powerof2(fl_pad)) {
611 		/*
612 		 * If there is any chance that we might use buffer packing and
613 		 * the chip is a T4, then pick 64 as the pad/pack boundary.  Set
614 		 * it to the minimum allowed in all other cases.
615 		 */
616 		pad = is_t4(sc) && buffer_packing ? 64 : 1 << pad_shift;
617 
618 		/*
619 		 * For fl_pad = 0 we'll still write a reasonable value to the
620 		 * register but all the freelists will opt out of padding.
621 		 * We'll complain here only if the user tried to set it to a
622 		 * value greater than 0 that was invalid.
623 		 */
624 		if (fl_pad > 0) {
625 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pad value"
626 			    " (%d), using %d instead.\n", fl_pad, pad);
627 		}
628 	}
629 	m = V_INGPADBOUNDARY(M_INGPADBOUNDARY);
630 	v = V_INGPADBOUNDARY(ilog2(pad) - pad_shift);
631 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
632 
633 	if (is_t4(sc)) {
634 		if (fl_pack != -1 && fl_pack != pad) {
635 			/* Complain but carry on. */
636 			device_printf(sc->dev, "hw.cxgbe.fl_pack (%d) ignored,"
637 			    " using %d instead.\n", fl_pack, pad);
638 		}
639 		return;
640 	}
641 
642 	pack = fl_pack;
643 	if (fl_pack < 16 || fl_pack == 32 || fl_pack > 4096 ||
644 	    !powerof2(fl_pack)) {
645 		if (sc->params.pci.mps > MAX_PACK_BOUNDARY)
646 			pack = MAX_PACK_BOUNDARY;
647 		else
648 			pack = max(sc->params.pci.mps, CACHE_LINE_SIZE);
649 		MPASS(powerof2(pack));
650 		if (pack < 16)
651 			pack = 16;
652 		if (pack == 32)
653 			pack = 64;
654 		if (pack > 4096)
655 			pack = 4096;
656 		if (fl_pack != -1) {
657 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pack value"
658 			    " (%d), using %d instead.\n", fl_pack, pack);
659 		}
660 	}
661 	m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY);
662 	if (pack == 16)
663 		v = V_INGPACKBOUNDARY(0);
664 	else
665 		v = V_INGPACKBOUNDARY(ilog2(pack) - 5);
666 
667 	MPASS(!is_t4(sc));	/* T4 doesn't have SGE_CONTROL2 */
668 	t4_set_reg_field(sc, A_SGE_CONTROL2, m, v);
669 }
670 
671 /*
672  * adap->params.vpd.cclk must be set up before this is called.
673  */
674 void
675 t4_tweak_chip_settings(struct adapter *sc)
676 {
677 	int i, reg;
678 	uint32_t v, m;
679 	int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200};
680 	int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk;
681 	int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */
682 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
683 	static int sw_buf_sizes[] = {
684 		MCLBYTES,
685 #if MJUMPAGESIZE != MCLBYTES
686 		MJUMPAGESIZE,
687 #endif
688 		MJUM9BYTES,
689 		MJUM16BYTES
690 	};
691 
692 	KASSERT(sc->flags & MASTER_PF,
693 	    ("%s: trying to change chip settings when not master.", __func__));
694 
695 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
696 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
697 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
698 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
699 
700 	setup_pad_and_pack_boundaries(sc);
701 
702 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
703 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
704 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
705 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
706 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
707 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
708 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
709 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
710 	t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v);
711 
712 	t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0, 4096);
713 	t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE1, 65536);
714 	reg = A_SGE_FL_BUFFER_SIZE2;
715 	for (i = 0; i < nitems(sw_buf_sizes); i++) {
716 		MPASS(reg <= A_SGE_FL_BUFFER_SIZE15);
717 		t4_write_reg(sc, reg, sw_buf_sizes[i]);
718 		reg += 4;
719 		MPASS(reg <= A_SGE_FL_BUFFER_SIZE15);
720 		t4_write_reg(sc, reg, sw_buf_sizes[i] - CL_METADATA_SIZE);
721 		reg += 4;
722 	}
723 
724 	v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) |
725 	    V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]);
726 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v);
727 
728 	KASSERT(intr_timer[0] <= timer_max,
729 	    ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0],
730 	    timer_max));
731 	for (i = 1; i < nitems(intr_timer); i++) {
732 		KASSERT(intr_timer[i] >= intr_timer[i - 1],
733 		    ("%s: timers not listed in increasing order (%d)",
734 		    __func__, i));
735 
736 		while (intr_timer[i] > timer_max) {
737 			if (i == nitems(intr_timer) - 1) {
738 				intr_timer[i] = timer_max;
739 				break;
740 			}
741 			intr_timer[i] += intr_timer[i - 1];
742 			intr_timer[i] /= 2;
743 		}
744 	}
745 
746 	v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) |
747 	    V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1]));
748 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v);
749 	v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) |
750 	    V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3]));
751 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v);
752 	v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) |
753 	    V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5]));
754 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v);
755 
756 	if (chip_id(sc) >= CHELSIO_T6) {
757 		m = V_TSCALE(M_TSCALE);
758 		if (tscale == 1)
759 			v = 0;
760 		else
761 			v = V_TSCALE(tscale - 2);
762 		t4_set_reg_field(sc, A_SGE_ITP_CONTROL, m, v);
763 
764 		if (sc->debug_flags & DF_DISABLE_TCB_CACHE) {
765 			m = V_RDTHRESHOLD(M_RDTHRESHOLD) | F_WRTHRTHRESHEN |
766 			    V_WRTHRTHRESH(M_WRTHRTHRESH);
767 			t4_tp_pio_read(sc, &v, 1, A_TP_CMM_CONFIG, 1);
768 			v &= ~m;
769 			v |= V_RDTHRESHOLD(1) | F_WRTHRTHRESHEN |
770 			    V_WRTHRTHRESH(16);
771 			t4_tp_pio_write(sc, &v, 1, A_TP_CMM_CONFIG, 1);
772 		}
773 	}
774 
775 	/* 4K, 16K, 64K, 256K DDP "page sizes" for TDDP */
776 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
777 	t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v);
778 
779 	/*
780 	 * 4K, 8K, 16K, 64K DDP "page sizes" for iSCSI DDP.  These have been
781 	 * chosen with MAXPHYS = 128K in mind.  The largest DDP buffer that we
782 	 * may have to deal with is MAXPHYS + 1 page.
783 	 */
784 	v = V_HPZ0(0) | V_HPZ1(1) | V_HPZ2(2) | V_HPZ3(4);
785 	t4_write_reg(sc, A_ULP_RX_ISCSI_PSZ, v);
786 
787 	/* We use multiple DDP page sizes both in plain-TOE and ISCSI modes. */
788 	m = v = F_TDDPTAGTCB | F_ISCSITAGTCB;
789 	t4_set_reg_field(sc, A_ULP_RX_CTL, m, v);
790 
791 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
792 	    F_RESETDDPOFFSET;
793 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
794 	t4_set_reg_field(sc, A_TP_PARA_REG5, m, v);
795 }
796 
797 /*
798  * SGE wants the buffer to be at least 64B and then a multiple of 16.  Its
799  * address mut be 16B aligned.  If padding is in use the buffer's start and end
800  * need to be aligned to the pad boundary as well.  We'll just make sure that
801  * the size is a multiple of the pad boundary here, it is up to the buffer
802  * allocation code to make sure the start of the buffer is aligned.
803  */
804 static inline int
805 hwsz_ok(struct adapter *sc, int hwsz)
806 {
807 	int mask = fl_pad ? sc->params.sge.pad_boundary - 1 : 16 - 1;
808 
809 	return (hwsz >= 64 && (hwsz & mask) == 0);
810 }
811 
812 /*
813  * XXX: driver really should be able to deal with unexpected settings.
814  */
815 int
816 t4_read_chip_settings(struct adapter *sc)
817 {
818 	struct sge *s = &sc->sge;
819 	struct sge_params *sp = &sc->params.sge;
820 	int i, j, n, rc = 0;
821 	uint32_t m, v, r;
822 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
823 	static int sw_buf_sizes[] = {	/* Sorted by size */
824 		MCLBYTES,
825 #if MJUMPAGESIZE != MCLBYTES
826 		MJUMPAGESIZE,
827 #endif
828 		MJUM9BYTES,
829 		MJUM16BYTES
830 	};
831 	struct rx_buf_info *rxb;
832 
833 	m = F_RXPKTCPLMODE;
834 	v = F_RXPKTCPLMODE;
835 	r = sc->params.sge.sge_control;
836 	if ((r & m) != v) {
837 		device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r);
838 		rc = EINVAL;
839 	}
840 
841 	/*
842 	 * If this changes then every single use of PAGE_SHIFT in the driver
843 	 * needs to be carefully reviewed for PAGE_SHIFT vs sp->page_shift.
844 	 */
845 	if (sp->page_shift != PAGE_SHIFT) {
846 		device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r);
847 		rc = EINVAL;
848 	}
849 
850 	s->safe_zidx = -1;
851 	rxb = &s->rx_buf_info[0];
852 	for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) {
853 		rxb->size1 = sw_buf_sizes[i];
854 		rxb->zone = m_getzone(rxb->size1);
855 		rxb->type = m_gettype(rxb->size1);
856 		rxb->size2 = 0;
857 		rxb->hwidx1 = -1;
858 		rxb->hwidx2 = -1;
859 		for (j = 0; j < SGE_FLBUF_SIZES; j++) {
860 			int hwsize = sp->sge_fl_buffer_size[j];
861 
862 			if (!hwsz_ok(sc, hwsize))
863 				continue;
864 
865 			/* hwidx for size1 */
866 			if (rxb->hwidx1 == -1 && rxb->size1 == hwsize)
867 				rxb->hwidx1 = j;
868 
869 			/* hwidx for size2 (buffer packing) */
870 			if (rxb->size1 - CL_METADATA_SIZE < hwsize)
871 				continue;
872 			n = rxb->size1 - hwsize - CL_METADATA_SIZE;
873 			if (n == 0) {
874 				rxb->hwidx2 = j;
875 				rxb->size2 = hwsize;
876 				break;	/* stop looking */
877 			}
878 			if (rxb->hwidx2 != -1) {
879 				if (n < sp->sge_fl_buffer_size[rxb->hwidx2] -
880 				    hwsize - CL_METADATA_SIZE) {
881 					rxb->hwidx2 = j;
882 					rxb->size2 = hwsize;
883 				}
884 			} else if (n <= 2 * CL_METADATA_SIZE) {
885 				rxb->hwidx2 = j;
886 				rxb->size2 = hwsize;
887 			}
888 		}
889 		if (rxb->hwidx2 != -1)
890 			sc->flags |= BUF_PACKING_OK;
891 		if (s->safe_zidx == -1 && rxb->size1 == safest_rx_cluster)
892 			s->safe_zidx = i;
893 	}
894 
895 	if (sc->flags & IS_VF)
896 		return (0);
897 
898 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
899 	r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ);
900 	if (r != v) {
901 		device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r);
902 		rc = EINVAL;
903 	}
904 
905 	m = v = F_TDDPTAGTCB;
906 	r = t4_read_reg(sc, A_ULP_RX_CTL);
907 	if ((r & m) != v) {
908 		device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r);
909 		rc = EINVAL;
910 	}
911 
912 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
913 	    F_RESETDDPOFFSET;
914 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
915 	r = t4_read_reg(sc, A_TP_PARA_REG5);
916 	if ((r & m) != v) {
917 		device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r);
918 		rc = EINVAL;
919 	}
920 
921 	t4_init_tp_params(sc, 1);
922 
923 	t4_read_mtu_tbl(sc, sc->params.mtus, NULL);
924 	t4_load_mtus(sc, sc->params.mtus, sc->params.a_wnd, sc->params.b_wnd);
925 
926 	return (rc);
927 }
928 
929 int
930 t4_create_dma_tag(struct adapter *sc)
931 {
932 	int rc;
933 
934 	rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
935 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
936 	    BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
937 	    NULL, &sc->dmat);
938 	if (rc != 0) {
939 		device_printf(sc->dev,
940 		    "failed to create main DMA tag: %d\n", rc);
941 	}
942 
943 	return (rc);
944 }
945 
946 void
947 t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
948     struct sysctl_oid_list *children)
949 {
950 	struct sge_params *sp = &sc->params.sge;
951 
952 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "buffer_sizes",
953 	    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, sc, 0,
954 	    sysctl_bufsizes, "A", "freelist buffer sizes");
955 
956 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD,
957 	    NULL, sp->fl_pktshift, "payload DMA offset in rx buffer (bytes)");
958 
959 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD,
960 	    NULL, sp->pad_boundary, "payload pad boundary (bytes)");
961 
962 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD,
963 	    NULL, sp->spg_len, "status page size (bytes)");
964 
965 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD,
966 	    NULL, cong_drop, "congestion drop setting");
967 
968 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD,
969 	    NULL, sp->pack_boundary, "payload pack boundary (bytes)");
970 }
971 
972 int
973 t4_destroy_dma_tag(struct adapter *sc)
974 {
975 	if (sc->dmat)
976 		bus_dma_tag_destroy(sc->dmat);
977 
978 	return (0);
979 }
980 
981 /*
982  * Allocate and initialize the firmware event queue, control queues, and special
983  * purpose rx queues owned by the adapter.
984  *
985  * Returns errno on failure.  Resources allocated up to that point may still be
986  * allocated.  Caller is responsible for cleanup in case this function fails.
987  */
988 int
989 t4_setup_adapter_queues(struct adapter *sc)
990 {
991 	struct sysctl_oid *oid;
992 	struct sysctl_oid_list *children;
993 	int rc, i;
994 
995 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
996 
997 	sysctl_ctx_init(&sc->ctx);
998 	sc->flags |= ADAP_SYSCTL_CTX;
999 
1000 	/*
1001 	 * Firmware event queue
1002 	 */
1003 	rc = alloc_fwq(sc);
1004 	if (rc != 0)
1005 		return (rc);
1006 
1007 	/*
1008 	 * That's all for the VF driver.
1009 	 */
1010 	if (sc->flags & IS_VF)
1011 		return (rc);
1012 
1013 	oid = device_get_sysctl_tree(sc->dev);
1014 	children = SYSCTL_CHILDREN(oid);
1015 
1016 	/*
1017 	 * XXX: General purpose rx queues, one per port.
1018 	 */
1019 
1020 	/*
1021 	 * Control queues, one per port.
1022 	 */
1023 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "ctrlq",
1024 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "control queues");
1025 	for_each_port(sc, i) {
1026 		struct sge_wrq *ctrlq = &sc->sge.ctrlq[i];
1027 
1028 		rc = alloc_ctrlq(sc, ctrlq, i, oid);
1029 		if (rc != 0)
1030 			return (rc);
1031 	}
1032 
1033 	return (rc);
1034 }
1035 
1036 /*
1037  * Idempotent
1038  */
1039 int
1040 t4_teardown_adapter_queues(struct adapter *sc)
1041 {
1042 	int i;
1043 
1044 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
1045 
1046 	/* Do this before freeing the queue */
1047 	if (sc->flags & ADAP_SYSCTL_CTX) {
1048 		sysctl_ctx_free(&sc->ctx);
1049 		sc->flags &= ~ADAP_SYSCTL_CTX;
1050 	}
1051 
1052 	if (!(sc->flags & IS_VF)) {
1053 		for_each_port(sc, i)
1054 			free_wrq(sc, &sc->sge.ctrlq[i]);
1055 	}
1056 	free_fwq(sc);
1057 
1058 	return (0);
1059 }
1060 
1061 /* Maximum payload that could arrive with a single iq descriptor. */
1062 static inline int
1063 max_rx_payload(struct adapter *sc, struct ifnet *ifp, const bool ofld)
1064 {
1065 	int maxp;
1066 
1067 	/* large enough even when hw VLAN extraction is disabled */
1068 	maxp = sc->params.sge.fl_pktshift + ETHER_HDR_LEN +
1069 	    ETHER_VLAN_ENCAP_LEN + ifp->if_mtu;
1070 	if (ofld && sc->tt.tls && sc->cryptocaps & FW_CAPS_CONFIG_TLSKEYS &&
1071 	    maxp < sc->params.tp.max_rx_pdu)
1072 		maxp = sc->params.tp.max_rx_pdu;
1073 	return (maxp);
1074 }
1075 
1076 int
1077 t4_setup_vi_queues(struct vi_info *vi)
1078 {
1079 	int rc = 0, i, intr_idx, iqidx;
1080 	struct sge_rxq *rxq;
1081 	struct sge_txq *txq;
1082 #ifdef TCP_OFFLOAD
1083 	struct sge_ofld_rxq *ofld_rxq;
1084 #endif
1085 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1086 	struct sge_wrq *ofld_txq;
1087 #endif
1088 #ifdef DEV_NETMAP
1089 	int saved_idx;
1090 	struct sge_nm_rxq *nm_rxq;
1091 	struct sge_nm_txq *nm_txq;
1092 #endif
1093 	char name[16];
1094 	struct port_info *pi = vi->pi;
1095 	struct adapter *sc = pi->adapter;
1096 	struct ifnet *ifp = vi->ifp;
1097 	struct sysctl_oid *oid = device_get_sysctl_tree(vi->dev);
1098 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
1099 	int maxp;
1100 
1101 	/* Interrupt vector to start from (when using multiple vectors) */
1102 	intr_idx = vi->first_intr;
1103 
1104 #ifdef DEV_NETMAP
1105 	saved_idx = intr_idx;
1106 	if (ifp->if_capabilities & IFCAP_NETMAP) {
1107 
1108 		/* netmap is supported with direct interrupts only. */
1109 		MPASS(!forwarding_intr_to_fwq(sc));
1110 
1111 		/*
1112 		 * We don't have buffers to back the netmap rx queues
1113 		 * right now so we create the queues in a way that
1114 		 * doesn't set off any congestion signal in the chip.
1115 		 */
1116 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_rxq",
1117 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "rx queues");
1118 		for_each_nm_rxq(vi, i, nm_rxq) {
1119 			rc = alloc_nm_rxq(vi, nm_rxq, intr_idx, i, oid);
1120 			if (rc != 0)
1121 				goto done;
1122 			intr_idx++;
1123 		}
1124 
1125 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_txq",
1126 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "tx queues");
1127 		for_each_nm_txq(vi, i, nm_txq) {
1128 			iqidx = vi->first_nm_rxq + (i % vi->nnmrxq);
1129 			rc = alloc_nm_txq(vi, nm_txq, iqidx, i, oid);
1130 			if (rc != 0)
1131 				goto done;
1132 		}
1133 	}
1134 
1135 	/* Normal rx queues and netmap rx queues share the same interrupts. */
1136 	intr_idx = saved_idx;
1137 #endif
1138 
1139 	/*
1140 	 * Allocate rx queues first because a default iqid is required when
1141 	 * creating a tx queue.
1142 	 */
1143 	maxp = max_rx_payload(sc, ifp, false);
1144 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
1145 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "rx queues");
1146 	for_each_rxq(vi, i, rxq) {
1147 
1148 		init_iq(&rxq->iq, sc, vi->tmr_idx, vi->pktc_idx, vi->qsize_rxq);
1149 
1150 		snprintf(name, sizeof(name), "%s rxq%d-fl",
1151 		    device_get_nameunit(vi->dev), i);
1152 		init_fl(sc, &rxq->fl, vi->qsize_rxq / 8, maxp, name);
1153 
1154 		rc = alloc_rxq(vi, rxq,
1155 		    forwarding_intr_to_fwq(sc) ? -1 : intr_idx, i, oid);
1156 		if (rc != 0)
1157 			goto done;
1158 		intr_idx++;
1159 	}
1160 #ifdef DEV_NETMAP
1161 	if (ifp->if_capabilities & IFCAP_NETMAP)
1162 		intr_idx = saved_idx + max(vi->nrxq, vi->nnmrxq);
1163 #endif
1164 #ifdef TCP_OFFLOAD
1165 	maxp = max_rx_payload(sc, ifp, true);
1166 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq",
1167 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "rx queues for offloaded TCP connections");
1168 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1169 
1170 		init_iq(&ofld_rxq->iq, sc, vi->ofld_tmr_idx, vi->ofld_pktc_idx,
1171 		    vi->qsize_rxq);
1172 
1173 		snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
1174 		    device_get_nameunit(vi->dev), i);
1175 		init_fl(sc, &ofld_rxq->fl, vi->qsize_rxq / 8, maxp, name);
1176 
1177 		rc = alloc_ofld_rxq(vi, ofld_rxq,
1178 		    forwarding_intr_to_fwq(sc) ? -1 : intr_idx, i, oid);
1179 		if (rc != 0)
1180 			goto done;
1181 		intr_idx++;
1182 	}
1183 #endif
1184 
1185 	/*
1186 	 * Now the tx queues.
1187 	 */
1188 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "txq",
1189 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "tx queues");
1190 	for_each_txq(vi, i, txq) {
1191 		iqidx = vi->first_rxq + (i % vi->nrxq);
1192 		snprintf(name, sizeof(name), "%s txq%d",
1193 		    device_get_nameunit(vi->dev), i);
1194 		init_eq(sc, &txq->eq, EQ_ETH, vi->qsize_txq, pi->tx_chan,
1195 		    sc->sge.rxq[iqidx].iq.cntxt_id, name);
1196 
1197 		rc = alloc_txq(vi, txq, i, oid);
1198 		if (rc != 0)
1199 			goto done;
1200 	}
1201 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1202 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_txq",
1203 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "tx queues for TOE/ETHOFLD");
1204 	for_each_ofld_txq(vi, i, ofld_txq) {
1205 		struct sysctl_oid *oid2;
1206 
1207 		snprintf(name, sizeof(name), "%s ofld_txq%d",
1208 		    device_get_nameunit(vi->dev), i);
1209 		if (vi->nofldrxq > 0) {
1210 			iqidx = vi->first_ofld_rxq + (i % vi->nofldrxq);
1211 			init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq,
1212 			    pi->tx_chan, sc->sge.ofld_rxq[iqidx].iq.cntxt_id,
1213 			    name);
1214 		} else {
1215 			iqidx = vi->first_rxq + (i % vi->nrxq);
1216 			init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq,
1217 			    pi->tx_chan, sc->sge.rxq[iqidx].iq.cntxt_id, name);
1218 		}
1219 
1220 		snprintf(name, sizeof(name), "%d", i);
1221 		oid2 = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
1222 		    name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "offload tx queue");
1223 
1224 		rc = alloc_wrq(sc, vi, ofld_txq, oid2);
1225 		if (rc != 0)
1226 			goto done;
1227 	}
1228 #endif
1229 done:
1230 	if (rc)
1231 		t4_teardown_vi_queues(vi);
1232 
1233 	return (rc);
1234 }
1235 
1236 /*
1237  * Idempotent
1238  */
1239 int
1240 t4_teardown_vi_queues(struct vi_info *vi)
1241 {
1242 	int i;
1243 	struct sge_rxq *rxq;
1244 	struct sge_txq *txq;
1245 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1246 	struct port_info *pi = vi->pi;
1247 	struct adapter *sc = pi->adapter;
1248 	struct sge_wrq *ofld_txq;
1249 #endif
1250 #ifdef TCP_OFFLOAD
1251 	struct sge_ofld_rxq *ofld_rxq;
1252 #endif
1253 #ifdef DEV_NETMAP
1254 	struct sge_nm_rxq *nm_rxq;
1255 	struct sge_nm_txq *nm_txq;
1256 #endif
1257 
1258 	/* Do this before freeing the queues */
1259 	if (vi->flags & VI_SYSCTL_CTX) {
1260 		sysctl_ctx_free(&vi->ctx);
1261 		vi->flags &= ~VI_SYSCTL_CTX;
1262 	}
1263 
1264 #ifdef DEV_NETMAP
1265 	if (vi->ifp->if_capabilities & IFCAP_NETMAP) {
1266 		for_each_nm_txq(vi, i, nm_txq) {
1267 			free_nm_txq(vi, nm_txq);
1268 		}
1269 
1270 		for_each_nm_rxq(vi, i, nm_rxq) {
1271 			free_nm_rxq(vi, nm_rxq);
1272 		}
1273 	}
1274 #endif
1275 
1276 	/*
1277 	 * Take down all the tx queues first, as they reference the rx queues
1278 	 * (for egress updates, etc.).
1279 	 */
1280 
1281 	for_each_txq(vi, i, txq) {
1282 		free_txq(vi, txq);
1283 	}
1284 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1285 	for_each_ofld_txq(vi, i, ofld_txq) {
1286 		free_wrq(sc, ofld_txq);
1287 	}
1288 #endif
1289 
1290 	/*
1291 	 * Then take down the rx queues.
1292 	 */
1293 
1294 	for_each_rxq(vi, i, rxq) {
1295 		free_rxq(vi, rxq);
1296 	}
1297 #ifdef TCP_OFFLOAD
1298 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1299 		free_ofld_rxq(vi, ofld_rxq);
1300 	}
1301 #endif
1302 
1303 	return (0);
1304 }
1305 
1306 /*
1307  * Interrupt handler when the driver is using only 1 interrupt.  This is a very
1308  * unusual scenario.
1309  *
1310  * a) Deals with errors, if any.
1311  * b) Services firmware event queue, which is taking interrupts for all other
1312  *    queues.
1313  */
1314 void
1315 t4_intr_all(void *arg)
1316 {
1317 	struct adapter *sc = arg;
1318 	struct sge_iq *fwq = &sc->sge.fwq;
1319 
1320 	MPASS(sc->intr_count == 1);
1321 
1322 	if (sc->intr_type == INTR_INTX)
1323 		t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
1324 
1325 	t4_intr_err(arg);
1326 	t4_intr_evt(fwq);
1327 }
1328 
1329 /*
1330  * Interrupt handler for errors (installed directly when multiple interrupts are
1331  * being used, or called by t4_intr_all).
1332  */
1333 void
1334 t4_intr_err(void *arg)
1335 {
1336 	struct adapter *sc = arg;
1337 	uint32_t v;
1338 	const bool verbose = (sc->debug_flags & DF_VERBOSE_SLOWINTR) != 0;
1339 
1340 	if (sc->flags & ADAP_ERR)
1341 		return;
1342 
1343 	v = t4_read_reg(sc, MYPF_REG(A_PL_PF_INT_CAUSE));
1344 	if (v & F_PFSW) {
1345 		sc->swintr++;
1346 		t4_write_reg(sc, MYPF_REG(A_PL_PF_INT_CAUSE), v);
1347 	}
1348 
1349 	t4_slow_intr_handler(sc, verbose);
1350 }
1351 
1352 /*
1353  * Interrupt handler for iq-only queues.  The firmware event queue is the only
1354  * such queue right now.
1355  */
1356 void
1357 t4_intr_evt(void *arg)
1358 {
1359 	struct sge_iq *iq = arg;
1360 
1361 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1362 		service_iq(iq, 0);
1363 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1364 	}
1365 }
1366 
1367 /*
1368  * Interrupt handler for iq+fl queues.
1369  */
1370 void
1371 t4_intr(void *arg)
1372 {
1373 	struct sge_iq *iq = arg;
1374 
1375 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1376 		service_iq_fl(iq, 0);
1377 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1378 	}
1379 }
1380 
1381 #ifdef DEV_NETMAP
1382 /*
1383  * Interrupt handler for netmap rx queues.
1384  */
1385 void
1386 t4_nm_intr(void *arg)
1387 {
1388 	struct sge_nm_rxq *nm_rxq = arg;
1389 
1390 	if (atomic_cmpset_int(&nm_rxq->nm_state, NM_ON, NM_BUSY)) {
1391 		service_nm_rxq(nm_rxq);
1392 		(void) atomic_cmpset_int(&nm_rxq->nm_state, NM_BUSY, NM_ON);
1393 	}
1394 }
1395 
1396 /*
1397  * Interrupt handler for vectors shared between NIC and netmap rx queues.
1398  */
1399 void
1400 t4_vi_intr(void *arg)
1401 {
1402 	struct irq *irq = arg;
1403 
1404 	MPASS(irq->nm_rxq != NULL);
1405 	t4_nm_intr(irq->nm_rxq);
1406 
1407 	MPASS(irq->rxq != NULL);
1408 	t4_intr(irq->rxq);
1409 }
1410 #endif
1411 
1412 /*
1413  * Deals with interrupts on an iq-only (no freelist) queue.
1414  */
1415 static int
1416 service_iq(struct sge_iq *iq, int budget)
1417 {
1418 	struct sge_iq *q;
1419 	struct adapter *sc = iq->adapter;
1420 	struct iq_desc *d = &iq->desc[iq->cidx];
1421 	int ndescs = 0, limit;
1422 	int rsp_type;
1423 	uint32_t lq;
1424 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1425 
1426 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1427 	KASSERT((iq->flags & IQ_HAS_FL) == 0,
1428 	    ("%s: called for iq %p with fl (iq->flags 0x%x)", __func__, iq,
1429 	    iq->flags));
1430 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1431 	MPASS((iq->flags & IQ_LRO_ENABLED) == 0);
1432 
1433 	limit = budget ? budget : iq->qsize / 16;
1434 
1435 	/*
1436 	 * We always come back and check the descriptor ring for new indirect
1437 	 * interrupts and other responses after running a single handler.
1438 	 */
1439 	for (;;) {
1440 		while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1441 
1442 			rmb();
1443 
1444 			rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1445 			lq = be32toh(d->rsp.pldbuflen_qid);
1446 
1447 			switch (rsp_type) {
1448 			case X_RSPD_TYPE_FLBUF:
1449 				panic("%s: data for an iq (%p) with no freelist",
1450 				    __func__, iq);
1451 
1452 				/* NOTREACHED */
1453 
1454 			case X_RSPD_TYPE_CPL:
1455 				KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1456 				    ("%s: bad opcode %02x.", __func__,
1457 				    d->rss.opcode));
1458 				t4_cpl_handler[d->rss.opcode](iq, &d->rss, NULL);
1459 				break;
1460 
1461 			case X_RSPD_TYPE_INTR:
1462 				/*
1463 				 * There are 1K interrupt-capable queues (qids 0
1464 				 * through 1023).  A response type indicating a
1465 				 * forwarded interrupt with a qid >= 1K is an
1466 				 * iWARP async notification.
1467 				 */
1468 				if (__predict_true(lq >= 1024)) {
1469 					t4_an_handler(iq, &d->rsp);
1470 					break;
1471 				}
1472 
1473 				q = sc->sge.iqmap[lq - sc->sge.iq_start -
1474 				    sc->sge.iq_base];
1475 				if (atomic_cmpset_int(&q->state, IQS_IDLE,
1476 				    IQS_BUSY)) {
1477 					if (service_iq_fl(q, q->qsize / 16) == 0) {
1478 						(void) atomic_cmpset_int(&q->state,
1479 						    IQS_BUSY, IQS_IDLE);
1480 					} else {
1481 						STAILQ_INSERT_TAIL(&iql, q,
1482 						    link);
1483 					}
1484 				}
1485 				break;
1486 
1487 			default:
1488 				KASSERT(0,
1489 				    ("%s: illegal response type %d on iq %p",
1490 				    __func__, rsp_type, iq));
1491 				log(LOG_ERR,
1492 				    "%s: illegal response type %d on iq %p",
1493 				    device_get_nameunit(sc->dev), rsp_type, iq);
1494 				break;
1495 			}
1496 
1497 			d++;
1498 			if (__predict_false(++iq->cidx == iq->sidx)) {
1499 				iq->cidx = 0;
1500 				iq->gen ^= F_RSPD_GEN;
1501 				d = &iq->desc[0];
1502 			}
1503 			if (__predict_false(++ndescs == limit)) {
1504 				t4_write_reg(sc, sc->sge_gts_reg,
1505 				    V_CIDXINC(ndescs) |
1506 				    V_INGRESSQID(iq->cntxt_id) |
1507 				    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1508 				ndescs = 0;
1509 
1510 				if (budget) {
1511 					return (EINPROGRESS);
1512 				}
1513 			}
1514 		}
1515 
1516 		if (STAILQ_EMPTY(&iql))
1517 			break;
1518 
1519 		/*
1520 		 * Process the head only, and send it to the back of the list if
1521 		 * it's still not done.
1522 		 */
1523 		q = STAILQ_FIRST(&iql);
1524 		STAILQ_REMOVE_HEAD(&iql, link);
1525 		if (service_iq_fl(q, q->qsize / 8) == 0)
1526 			(void) atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1527 		else
1528 			STAILQ_INSERT_TAIL(&iql, q, link);
1529 	}
1530 
1531 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1532 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1533 
1534 	return (0);
1535 }
1536 
1537 static inline int
1538 sort_before_lro(struct lro_ctrl *lro)
1539 {
1540 
1541 	return (lro->lro_mbuf_max != 0);
1542 }
1543 
1544 static inline uint64_t
1545 last_flit_to_ns(struct adapter *sc, uint64_t lf)
1546 {
1547 	uint64_t n = be64toh(lf) & 0xfffffffffffffff;	/* 60b, not 64b. */
1548 
1549 	if (n > UINT64_MAX / 1000000)
1550 		return (n / sc->params.vpd.cclk * 1000000);
1551 	else
1552 		return (n * 1000000 / sc->params.vpd.cclk);
1553 }
1554 
1555 static inline void
1556 move_to_next_rxbuf(struct sge_fl *fl)
1557 {
1558 
1559 	fl->rx_offset = 0;
1560 	if (__predict_false((++fl->cidx & 7) == 0)) {
1561 		uint16_t cidx = fl->cidx >> 3;
1562 
1563 		if (__predict_false(cidx == fl->sidx))
1564 			fl->cidx = cidx = 0;
1565 		fl->hw_cidx = cidx;
1566 	}
1567 }
1568 
1569 /*
1570  * Deals with interrupts on an iq+fl queue.
1571  */
1572 static int
1573 service_iq_fl(struct sge_iq *iq, int budget)
1574 {
1575 	struct sge_rxq *rxq = iq_to_rxq(iq);
1576 	struct sge_fl *fl;
1577 	struct adapter *sc = iq->adapter;
1578 	struct iq_desc *d = &iq->desc[iq->cidx];
1579 	int ndescs, limit;
1580 	int rsp_type, starved;
1581 	uint32_t lq;
1582 	uint16_t fl_hw_cidx;
1583 	struct mbuf *m0;
1584 #if defined(INET) || defined(INET6)
1585 	const struct timeval lro_timeout = {0, sc->lro_timeout};
1586 	struct lro_ctrl *lro = &rxq->lro;
1587 #endif
1588 
1589 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1590 	MPASS(iq->flags & IQ_HAS_FL);
1591 
1592 	ndescs = 0;
1593 #if defined(INET) || defined(INET6)
1594 	if (iq->flags & IQ_ADJ_CREDIT) {
1595 		MPASS(sort_before_lro(lro));
1596 		iq->flags &= ~IQ_ADJ_CREDIT;
1597 		if ((d->rsp.u.type_gen & F_RSPD_GEN) != iq->gen) {
1598 			tcp_lro_flush_all(lro);
1599 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(1) |
1600 			    V_INGRESSQID((u32)iq->cntxt_id) |
1601 			    V_SEINTARM(iq->intr_params));
1602 			return (0);
1603 		}
1604 		ndescs = 1;
1605 	}
1606 #else
1607 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1608 #endif
1609 
1610 	limit = budget ? budget : iq->qsize / 16;
1611 	fl = &rxq->fl;
1612 	fl_hw_cidx = fl->hw_cidx;	/* stable snapshot */
1613 	while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1614 
1615 		rmb();
1616 
1617 		m0 = NULL;
1618 		rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1619 		lq = be32toh(d->rsp.pldbuflen_qid);
1620 
1621 		switch (rsp_type) {
1622 		case X_RSPD_TYPE_FLBUF:
1623 			if (lq & F_RSPD_NEWBUF) {
1624 				if (fl->rx_offset > 0)
1625 					move_to_next_rxbuf(fl);
1626 				lq = G_RSPD_LEN(lq);
1627 			}
1628 			if (IDXDIFF(fl->hw_cidx, fl_hw_cidx, fl->sidx) > 4) {
1629 				FL_LOCK(fl);
1630 				refill_fl(sc, fl, 64);
1631 				FL_UNLOCK(fl);
1632 				fl_hw_cidx = fl->hw_cidx;
1633 			}
1634 
1635 			if (d->rss.opcode == CPL_RX_PKT) {
1636 				if (__predict_true(eth_rx(sc, rxq, d, lq) == 0))
1637 					break;
1638 				goto out;
1639 			}
1640 			m0 = get_fl_payload(sc, fl, lq);
1641 			if (__predict_false(m0 == NULL))
1642 				goto out;
1643 
1644 			/* fall through */
1645 
1646 		case X_RSPD_TYPE_CPL:
1647 			KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1648 			    ("%s: bad opcode %02x.", __func__, d->rss.opcode));
1649 			t4_cpl_handler[d->rss.opcode](iq, &d->rss, m0);
1650 			break;
1651 
1652 		case X_RSPD_TYPE_INTR:
1653 
1654 			/*
1655 			 * There are 1K interrupt-capable queues (qids 0
1656 			 * through 1023).  A response type indicating a
1657 			 * forwarded interrupt with a qid >= 1K is an
1658 			 * iWARP async notification.  That is the only
1659 			 * acceptable indirect interrupt on this queue.
1660 			 */
1661 			if (__predict_false(lq < 1024)) {
1662 				panic("%s: indirect interrupt on iq_fl %p "
1663 				    "with qid %u", __func__, iq, lq);
1664 			}
1665 
1666 			t4_an_handler(iq, &d->rsp);
1667 			break;
1668 
1669 		default:
1670 			KASSERT(0, ("%s: illegal response type %d on iq %p",
1671 			    __func__, rsp_type, iq));
1672 			log(LOG_ERR, "%s: illegal response type %d on iq %p",
1673 			    device_get_nameunit(sc->dev), rsp_type, iq);
1674 			break;
1675 		}
1676 
1677 		d++;
1678 		if (__predict_false(++iq->cidx == iq->sidx)) {
1679 			iq->cidx = 0;
1680 			iq->gen ^= F_RSPD_GEN;
1681 			d = &iq->desc[0];
1682 		}
1683 		if (__predict_false(++ndescs == limit)) {
1684 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1685 			    V_INGRESSQID(iq->cntxt_id) |
1686 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1687 
1688 #if defined(INET) || defined(INET6)
1689 			if (iq->flags & IQ_LRO_ENABLED &&
1690 			    !sort_before_lro(lro) &&
1691 			    sc->lro_timeout != 0) {
1692 				tcp_lro_flush_inactive(lro, &lro_timeout);
1693 			}
1694 #endif
1695 			if (budget)
1696 				return (EINPROGRESS);
1697 			ndescs = 0;
1698 		}
1699 	}
1700 out:
1701 #if defined(INET) || defined(INET6)
1702 	if (iq->flags & IQ_LRO_ENABLED) {
1703 		if (ndescs > 0 && lro->lro_mbuf_count > 8) {
1704 			MPASS(sort_before_lro(lro));
1705 			/* hold back one credit and don't flush LRO state */
1706 			iq->flags |= IQ_ADJ_CREDIT;
1707 			ndescs--;
1708 		} else {
1709 			tcp_lro_flush_all(lro);
1710 		}
1711 	}
1712 #endif
1713 
1714 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1715 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1716 
1717 	FL_LOCK(fl);
1718 	starved = refill_fl(sc, fl, 64);
1719 	FL_UNLOCK(fl);
1720 	if (__predict_false(starved != 0))
1721 		add_fl_to_sfl(sc, fl);
1722 
1723 	return (0);
1724 }
1725 
1726 static inline struct cluster_metadata *
1727 cl_metadata(struct fl_sdesc *sd)
1728 {
1729 
1730 	return ((void *)(sd->cl + sd->moff));
1731 }
1732 
1733 static void
1734 rxb_free(struct mbuf *m)
1735 {
1736 	struct cluster_metadata *clm = m->m_ext.ext_arg1;
1737 
1738 	uma_zfree(clm->zone, clm->cl);
1739 	counter_u64_add(extfree_rels, 1);
1740 }
1741 
1742 /*
1743  * The mbuf returned comes from zone_muf and carries the payload in one of these
1744  * ways
1745  * a) complete frame inside the mbuf
1746  * b) m_cljset (for clusters without metadata)
1747  * d) m_extaddref (cluster with metadata)
1748  */
1749 static struct mbuf *
1750 get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1751     int remaining)
1752 {
1753 	struct mbuf *m;
1754 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1755 	struct rx_buf_info *rxb = &sc->sge.rx_buf_info[sd->zidx];
1756 	struct cluster_metadata *clm;
1757 	int len, blen;
1758 	caddr_t payload;
1759 
1760 	if (fl->flags & FL_BUF_PACKING) {
1761 		u_int l, pad;
1762 
1763 		blen = rxb->size2 - fl->rx_offset;	/* max possible in this buf */
1764 		len = min(remaining, blen);
1765 		payload = sd->cl + fl->rx_offset;
1766 
1767 		l = fr_offset + len;
1768 		pad = roundup2(l, fl->buf_boundary) - l;
1769 		if (fl->rx_offset + len + pad < rxb->size2)
1770 			blen = len + pad;
1771 		MPASS(fl->rx_offset + blen <= rxb->size2);
1772 	} else {
1773 		MPASS(fl->rx_offset == 0);	/* not packing */
1774 		blen = rxb->size1;
1775 		len = min(remaining, blen);
1776 		payload = sd->cl;
1777 	}
1778 
1779 	if (fr_offset == 0) {
1780 		m = m_gethdr(M_NOWAIT, MT_DATA);
1781 		if (__predict_false(m == NULL))
1782 			return (NULL);
1783 		m->m_pkthdr.len = remaining;
1784 	} else {
1785 		m = m_get(M_NOWAIT, MT_DATA);
1786 		if (__predict_false(m == NULL))
1787 			return (NULL);
1788 	}
1789 	m->m_len = len;
1790 
1791 	if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) {
1792 		/* copy data to mbuf */
1793 		bcopy(payload, mtod(m, caddr_t), len);
1794 		if (fl->flags & FL_BUF_PACKING) {
1795 			fl->rx_offset += blen;
1796 			MPASS(fl->rx_offset <= rxb->size2);
1797 			if (fl->rx_offset < rxb->size2)
1798 				return (m);	/* without advancing the cidx */
1799 		}
1800 	} else if (fl->flags & FL_BUF_PACKING) {
1801 		clm = cl_metadata(sd);
1802 		if (sd->nmbuf++ == 0) {
1803 			clm->refcount = 1;
1804 			clm->zone = rxb->zone;
1805 			clm->cl = sd->cl;
1806 			counter_u64_add(extfree_refs, 1);
1807 		}
1808 		m_extaddref(m, payload, blen, &clm->refcount, rxb_free, clm,
1809 		    NULL);
1810 
1811 		fl->rx_offset += blen;
1812 		MPASS(fl->rx_offset <= rxb->size2);
1813 		if (fl->rx_offset < rxb->size2)
1814 			return (m);	/* without advancing the cidx */
1815 	} else {
1816 		m_cljset(m, sd->cl, rxb->type);
1817 		sd->cl = NULL;	/* consumed, not a recycle candidate */
1818 	}
1819 
1820 	move_to_next_rxbuf(fl);
1821 
1822 	return (m);
1823 }
1824 
1825 static struct mbuf *
1826 get_fl_payload(struct adapter *sc, struct sge_fl *fl, const u_int plen)
1827 {
1828 	struct mbuf *m0, *m, **pnext;
1829 	u_int remaining;
1830 
1831 	if (__predict_false(fl->flags & FL_BUF_RESUME)) {
1832 		M_ASSERTPKTHDR(fl->m0);
1833 		MPASS(fl->m0->m_pkthdr.len == plen);
1834 		MPASS(fl->remaining < plen);
1835 
1836 		m0 = fl->m0;
1837 		pnext = fl->pnext;
1838 		remaining = fl->remaining;
1839 		fl->flags &= ~FL_BUF_RESUME;
1840 		goto get_segment;
1841 	}
1842 
1843 	/*
1844 	 * Payload starts at rx_offset in the current hw buffer.  Its length is
1845 	 * 'len' and it may span multiple hw buffers.
1846 	 */
1847 
1848 	m0 = get_scatter_segment(sc, fl, 0, plen);
1849 	if (m0 == NULL)
1850 		return (NULL);
1851 	remaining = plen - m0->m_len;
1852 	pnext = &m0->m_next;
1853 	while (remaining > 0) {
1854 get_segment:
1855 		MPASS(fl->rx_offset == 0);
1856 		m = get_scatter_segment(sc, fl, plen - remaining, remaining);
1857 		if (__predict_false(m == NULL)) {
1858 			fl->m0 = m0;
1859 			fl->pnext = pnext;
1860 			fl->remaining = remaining;
1861 			fl->flags |= FL_BUF_RESUME;
1862 			return (NULL);
1863 		}
1864 		*pnext = m;
1865 		pnext = &m->m_next;
1866 		remaining -= m->m_len;
1867 	}
1868 	*pnext = NULL;
1869 
1870 	M_ASSERTPKTHDR(m0);
1871 	return (m0);
1872 }
1873 
1874 static int
1875 skip_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1876     int remaining)
1877 {
1878 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1879 	struct rx_buf_info *rxb = &sc->sge.rx_buf_info[sd->zidx];
1880 	int len, blen;
1881 
1882 	if (fl->flags & FL_BUF_PACKING) {
1883 		u_int l, pad;
1884 
1885 		blen = rxb->size2 - fl->rx_offset;	/* max possible in this buf */
1886 		len = min(remaining, blen);
1887 
1888 		l = fr_offset + len;
1889 		pad = roundup2(l, fl->buf_boundary) - l;
1890 		if (fl->rx_offset + len + pad < rxb->size2)
1891 			blen = len + pad;
1892 		fl->rx_offset += blen;
1893 		MPASS(fl->rx_offset <= rxb->size2);
1894 		if (fl->rx_offset < rxb->size2)
1895 			return (len);	/* without advancing the cidx */
1896 	} else {
1897 		MPASS(fl->rx_offset == 0);	/* not packing */
1898 		blen = rxb->size1;
1899 		len = min(remaining, blen);
1900 	}
1901 	move_to_next_rxbuf(fl);
1902 	return (len);
1903 }
1904 
1905 static inline void
1906 skip_fl_payload(struct adapter *sc, struct sge_fl *fl, int plen)
1907 {
1908 	int remaining, fr_offset, len;
1909 
1910 	fr_offset = 0;
1911 	remaining = plen;
1912 	while (remaining > 0) {
1913 		len = skip_scatter_segment(sc, fl, fr_offset, remaining);
1914 		fr_offset += len;
1915 		remaining -= len;
1916 	}
1917 }
1918 
1919 static inline int
1920 get_segment_len(struct adapter *sc, struct sge_fl *fl, int plen)
1921 {
1922 	int len;
1923 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1924 	struct rx_buf_info *rxb = &sc->sge.rx_buf_info[sd->zidx];
1925 
1926 	if (fl->flags & FL_BUF_PACKING)
1927 		len = rxb->size2 - fl->rx_offset;
1928 	else
1929 		len = rxb->size1;
1930 
1931 	return (min(plen, len));
1932 }
1933 
1934 static int
1935 eth_rx(struct adapter *sc, struct sge_rxq *rxq, const struct iq_desc *d,
1936     u_int plen)
1937 {
1938 	struct mbuf *m0;
1939 	struct ifnet *ifp = rxq->ifp;
1940 	struct sge_fl *fl = &rxq->fl;
1941 	struct vi_info *vi = ifp->if_softc;
1942 	const struct cpl_rx_pkt *cpl;
1943 #if defined(INET) || defined(INET6)
1944 	struct lro_ctrl *lro = &rxq->lro;
1945 #endif
1946 	uint16_t err_vec, tnl_type, tnlhdr_len;
1947 	static const int sw_hashtype[4][2] = {
1948 		{M_HASHTYPE_NONE, M_HASHTYPE_NONE},
1949 		{M_HASHTYPE_RSS_IPV4, M_HASHTYPE_RSS_IPV6},
1950 		{M_HASHTYPE_RSS_TCP_IPV4, M_HASHTYPE_RSS_TCP_IPV6},
1951 		{M_HASHTYPE_RSS_UDP_IPV4, M_HASHTYPE_RSS_UDP_IPV6},
1952 	};
1953 	static const int sw_csum_flags[2][2] = {
1954 		{
1955 			/* IP, inner IP */
1956 			CSUM_ENCAP_VXLAN |
1957 			    CSUM_L3_CALC | CSUM_L3_VALID |
1958 			    CSUM_L4_CALC | CSUM_L4_VALID |
1959 			    CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID |
1960 			    CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID,
1961 
1962 			/* IP, inner IP6 */
1963 			CSUM_ENCAP_VXLAN |
1964 			    CSUM_L3_CALC | CSUM_L3_VALID |
1965 			    CSUM_L4_CALC | CSUM_L4_VALID |
1966 			    CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID,
1967 		},
1968 		{
1969 			/* IP6, inner IP */
1970 			CSUM_ENCAP_VXLAN |
1971 			    CSUM_L4_CALC | CSUM_L4_VALID |
1972 			    CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID |
1973 			    CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID,
1974 
1975 			/* IP6, inner IP6 */
1976 			CSUM_ENCAP_VXLAN |
1977 			    CSUM_L4_CALC | CSUM_L4_VALID |
1978 			    CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID,
1979 		},
1980 	};
1981 
1982 	MPASS(plen > sc->params.sge.fl_pktshift);
1983 	if (vi->pfil != NULL && PFIL_HOOKED_IN(vi->pfil) &&
1984 	    __predict_true((fl->flags & FL_BUF_RESUME) == 0)) {
1985 		struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1986 		caddr_t frame;
1987 		int rc, slen;
1988 
1989 		slen = get_segment_len(sc, fl, plen) -
1990 		    sc->params.sge.fl_pktshift;
1991 		frame = sd->cl + fl->rx_offset + sc->params.sge.fl_pktshift;
1992 		CURVNET_SET_QUIET(ifp->if_vnet);
1993 		rc = pfil_run_hooks(vi->pfil, frame, ifp,
1994 		    slen | PFIL_MEMPTR | PFIL_IN, NULL);
1995 		CURVNET_RESTORE();
1996 		if (rc == PFIL_DROPPED || rc == PFIL_CONSUMED) {
1997 			skip_fl_payload(sc, fl, plen);
1998 			return (0);
1999 		}
2000 		if (rc == PFIL_REALLOCED) {
2001 			skip_fl_payload(sc, fl, plen);
2002 			m0 = pfil_mem2mbuf(frame);
2003 			goto have_mbuf;
2004 		}
2005 	}
2006 
2007 	m0 = get_fl_payload(sc, fl, plen);
2008 	if (__predict_false(m0 == NULL))
2009 		return (ENOMEM);
2010 
2011 	m0->m_pkthdr.len -= sc->params.sge.fl_pktshift;
2012 	m0->m_len -= sc->params.sge.fl_pktshift;
2013 	m0->m_data += sc->params.sge.fl_pktshift;
2014 
2015 have_mbuf:
2016 	m0->m_pkthdr.rcvif = ifp;
2017 	M_HASHTYPE_SET(m0, sw_hashtype[d->rss.hash_type][d->rss.ipv6]);
2018 	m0->m_pkthdr.flowid = be32toh(d->rss.hash_val);
2019 
2020 	cpl = (const void *)(&d->rss + 1);
2021 	if (sc->params.tp.rx_pkt_encap) {
2022 		const uint16_t ev = be16toh(cpl->err_vec);
2023 
2024 		err_vec = G_T6_COMPR_RXERR_VEC(ev);
2025 		tnl_type = G_T6_RX_TNL_TYPE(ev);
2026 		tnlhdr_len = G_T6_RX_TNLHDR_LEN(ev);
2027 	} else {
2028 		err_vec = be16toh(cpl->err_vec);
2029 		tnl_type = 0;
2030 		tnlhdr_len = 0;
2031 	}
2032 	if (cpl->csum_calc && err_vec == 0) {
2033 		int ipv6 = !!(cpl->l2info & htobe32(F_RXF_IP6));
2034 
2035 		/* checksum(s) calculated and found to be correct. */
2036 
2037 		MPASS((cpl->l2info & htobe32(F_RXF_IP)) ^
2038 		    (cpl->l2info & htobe32(F_RXF_IP6)));
2039 		m0->m_pkthdr.csum_data = be16toh(cpl->csum);
2040 		if (tnl_type == 0) {
2041 	    		if (!ipv6 && ifp->if_capenable & IFCAP_RXCSUM) {
2042 				m0->m_pkthdr.csum_flags = CSUM_L3_CALC |
2043 				    CSUM_L3_VALID | CSUM_L4_CALC |
2044 				    CSUM_L4_VALID;
2045 			} else if (ipv6 && ifp->if_capenable & IFCAP_RXCSUM_IPV6) {
2046 				m0->m_pkthdr.csum_flags = CSUM_L4_CALC |
2047 				    CSUM_L4_VALID;
2048 			}
2049 			rxq->rxcsum++;
2050 		} else {
2051 			MPASS(tnl_type == RX_PKT_TNL_TYPE_VXLAN);
2052 			if (__predict_false(cpl->ip_frag)) {
2053 				/*
2054 				 * csum_data is for the inner frame (which is an
2055 				 * IP fragment) and is not 0xffff.  There is no
2056 				 * way to pass the inner csum_data to the stack.
2057 				 * We don't want the stack to use the inner
2058 				 * csum_data to validate the outer frame or it
2059 				 * will get rejected.  So we fix csum_data here
2060 				 * and let sw do the checksum of inner IP
2061 				 * fragments.
2062 				 *
2063 				 * XXX: Need 32b for csum_data2 in an rx mbuf.
2064 				 * Maybe stuff it into rcv_tstmp?
2065 				 */
2066 				m0->m_pkthdr.csum_data = 0xffff;
2067 				if (ipv6) {
2068 					m0->m_pkthdr.csum_flags = CSUM_L4_CALC |
2069 					    CSUM_L4_VALID;
2070 				} else {
2071 					m0->m_pkthdr.csum_flags = CSUM_L3_CALC |
2072 					    CSUM_L3_VALID | CSUM_L4_CALC |
2073 					    CSUM_L4_VALID;
2074 				}
2075 			} else {
2076 				int outer_ipv6;
2077 
2078 				MPASS(m0->m_pkthdr.csum_data == 0xffff);
2079 
2080 				outer_ipv6 = tnlhdr_len >=
2081 				    sizeof(struct ether_header) +
2082 				    sizeof(struct ip6_hdr);
2083 				m0->m_pkthdr.csum_flags =
2084 				    sw_csum_flags[outer_ipv6][ipv6];
2085 			}
2086 			rxq->vxlan_rxcsum++;
2087 		}
2088 	}
2089 
2090 	if (cpl->vlan_ex) {
2091 		m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
2092 		m0->m_flags |= M_VLANTAG;
2093 		rxq->vlan_extraction++;
2094 	}
2095 
2096 	if (rxq->iq.flags & IQ_RX_TIMESTAMP) {
2097 		/*
2098 		 * Fill up rcv_tstmp but do not set M_TSTMP.
2099 		 * rcv_tstmp is not in the format that the
2100 		 * kernel expects and we don't want to mislead
2101 		 * it.  For now this is only for custom code
2102 		 * that knows how to interpret cxgbe's stamp.
2103 		 */
2104 		m0->m_pkthdr.rcv_tstmp =
2105 		    last_flit_to_ns(sc, d->rsp.u.last_flit);
2106 #ifdef notyet
2107 		m0->m_flags |= M_TSTMP;
2108 #endif
2109 	}
2110 
2111 #ifdef NUMA
2112 	m0->m_pkthdr.numa_domain = ifp->if_numa_domain;
2113 #endif
2114 #if defined(INET) || defined(INET6)
2115 	if (rxq->iq.flags & IQ_LRO_ENABLED && tnl_type == 0 &&
2116 	    (M_HASHTYPE_GET(m0) == M_HASHTYPE_RSS_TCP_IPV4 ||
2117 	    M_HASHTYPE_GET(m0) == M_HASHTYPE_RSS_TCP_IPV6)) {
2118 		if (sort_before_lro(lro)) {
2119 			tcp_lro_queue_mbuf(lro, m0);
2120 			return (0); /* queued for sort, then LRO */
2121 		}
2122 		if (tcp_lro_rx(lro, m0, 0) == 0)
2123 			return (0); /* queued for LRO */
2124 	}
2125 #endif
2126 	ifp->if_input(ifp, m0);
2127 
2128 	return (0);
2129 }
2130 
2131 /*
2132  * Must drain the wrq or make sure that someone else will.
2133  */
2134 static void
2135 wrq_tx_drain(void *arg, int n)
2136 {
2137 	struct sge_wrq *wrq = arg;
2138 	struct sge_eq *eq = &wrq->eq;
2139 
2140 	EQ_LOCK(eq);
2141 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2142 		drain_wrq_wr_list(wrq->adapter, wrq);
2143 	EQ_UNLOCK(eq);
2144 }
2145 
2146 static void
2147 drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq)
2148 {
2149 	struct sge_eq *eq = &wrq->eq;
2150 	u_int available, dbdiff;	/* # of hardware descriptors */
2151 	u_int n;
2152 	struct wrqe *wr;
2153 	struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
2154 
2155 	EQ_LOCK_ASSERT_OWNED(eq);
2156 	MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
2157 	wr = STAILQ_FIRST(&wrq->wr_list);
2158 	MPASS(wr != NULL);	/* Must be called with something useful to do */
2159 	MPASS(eq->pidx == eq->dbidx);
2160 	dbdiff = 0;
2161 
2162 	do {
2163 		eq->cidx = read_hw_cidx(eq);
2164 		if (eq->pidx == eq->cidx)
2165 			available = eq->sidx - 1;
2166 		else
2167 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2168 
2169 		MPASS(wr->wrq == wrq);
2170 		n = howmany(wr->wr_len, EQ_ESIZE);
2171 		if (available < n)
2172 			break;
2173 
2174 		dst = (void *)&eq->desc[eq->pidx];
2175 		if (__predict_true(eq->sidx - eq->pidx > n)) {
2176 			/* Won't wrap, won't end exactly at the status page. */
2177 			bcopy(&wr->wr[0], dst, wr->wr_len);
2178 			eq->pidx += n;
2179 		} else {
2180 			int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE;
2181 
2182 			bcopy(&wr->wr[0], dst, first_portion);
2183 			if (wr->wr_len > first_portion) {
2184 				bcopy(&wr->wr[first_portion], &eq->desc[0],
2185 				    wr->wr_len - first_portion);
2186 			}
2187 			eq->pidx = n - (eq->sidx - eq->pidx);
2188 		}
2189 		wrq->tx_wrs_copied++;
2190 
2191 		if (available < eq->sidx / 4 &&
2192 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2193 				/*
2194 				 * XXX: This is not 100% reliable with some
2195 				 * types of WRs.  But this is a very unusual
2196 				 * situation for an ofld/ctrl queue anyway.
2197 				 */
2198 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2199 			    F_FW_WR_EQUEQ);
2200 		}
2201 
2202 		dbdiff += n;
2203 		if (dbdiff >= 16) {
2204 			ring_eq_db(sc, eq, dbdiff);
2205 			dbdiff = 0;
2206 		}
2207 
2208 		STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
2209 		free_wrqe(wr);
2210 		MPASS(wrq->nwr_pending > 0);
2211 		wrq->nwr_pending--;
2212 		MPASS(wrq->ndesc_needed >= n);
2213 		wrq->ndesc_needed -= n;
2214 	} while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL);
2215 
2216 	if (dbdiff)
2217 		ring_eq_db(sc, eq, dbdiff);
2218 }
2219 
2220 /*
2221  * Doesn't fail.  Holds on to work requests it can't send right away.
2222  */
2223 void
2224 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
2225 {
2226 #ifdef INVARIANTS
2227 	struct sge_eq *eq = &wrq->eq;
2228 #endif
2229 
2230 	EQ_LOCK_ASSERT_OWNED(eq);
2231 	MPASS(wr != NULL);
2232 	MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN);
2233 	MPASS((wr->wr_len & 0x7) == 0);
2234 
2235 	STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
2236 	wrq->nwr_pending++;
2237 	wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE);
2238 
2239 	if (!TAILQ_EMPTY(&wrq->incomplete_wrs))
2240 		return;	/* commit_wrq_wr will drain wr_list as well. */
2241 
2242 	drain_wrq_wr_list(sc, wrq);
2243 
2244 	/* Doorbell must have caught up to the pidx. */
2245 	MPASS(eq->pidx == eq->dbidx);
2246 }
2247 
2248 void
2249 t4_update_fl_bufsize(struct ifnet *ifp)
2250 {
2251 	struct vi_info *vi = ifp->if_softc;
2252 	struct adapter *sc = vi->adapter;
2253 	struct sge_rxq *rxq;
2254 #ifdef TCP_OFFLOAD
2255 	struct sge_ofld_rxq *ofld_rxq;
2256 #endif
2257 	struct sge_fl *fl;
2258 	int i, maxp;
2259 
2260 	maxp = max_rx_payload(sc, ifp, false);
2261 	for_each_rxq(vi, i, rxq) {
2262 		fl = &rxq->fl;
2263 
2264 		FL_LOCK(fl);
2265 		fl->zidx = find_refill_source(sc, maxp,
2266 		    fl->flags & FL_BUF_PACKING);
2267 		FL_UNLOCK(fl);
2268 	}
2269 #ifdef TCP_OFFLOAD
2270 	maxp = max_rx_payload(sc, ifp, true);
2271 	for_each_ofld_rxq(vi, i, ofld_rxq) {
2272 		fl = &ofld_rxq->fl;
2273 
2274 		FL_LOCK(fl);
2275 		fl->zidx = find_refill_source(sc, maxp,
2276 		    fl->flags & FL_BUF_PACKING);
2277 		FL_UNLOCK(fl);
2278 	}
2279 #endif
2280 }
2281 
2282 static inline int
2283 mbuf_nsegs(struct mbuf *m)
2284 {
2285 
2286 	M_ASSERTPKTHDR(m);
2287 	KASSERT(m->m_pkthdr.inner_l5hlen > 0,
2288 	    ("%s: mbuf %p missing information on # of segments.", __func__, m));
2289 
2290 	return (m->m_pkthdr.inner_l5hlen);
2291 }
2292 
2293 static inline void
2294 set_mbuf_nsegs(struct mbuf *m, uint8_t nsegs)
2295 {
2296 
2297 	M_ASSERTPKTHDR(m);
2298 	m->m_pkthdr.inner_l5hlen = nsegs;
2299 }
2300 
2301 static inline int
2302 mbuf_cflags(struct mbuf *m)
2303 {
2304 
2305 	M_ASSERTPKTHDR(m);
2306 	return (m->m_pkthdr.PH_loc.eight[4]);
2307 }
2308 
2309 static inline void
2310 set_mbuf_cflags(struct mbuf *m, uint8_t flags)
2311 {
2312 
2313 	M_ASSERTPKTHDR(m);
2314 	m->m_pkthdr.PH_loc.eight[4] = flags;
2315 }
2316 
2317 static inline int
2318 mbuf_len16(struct mbuf *m)
2319 {
2320 	int n;
2321 
2322 	M_ASSERTPKTHDR(m);
2323 	n = m->m_pkthdr.PH_loc.eight[0];
2324 	if (!(mbuf_cflags(m) & MC_TLS))
2325 		MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2326 
2327 	return (n);
2328 }
2329 
2330 static inline void
2331 set_mbuf_len16(struct mbuf *m, uint8_t len16)
2332 {
2333 
2334 	M_ASSERTPKTHDR(m);
2335 	if (!(mbuf_cflags(m) & MC_TLS))
2336 		MPASS(len16 > 0 && len16 <= SGE_MAX_WR_LEN / 16);
2337 	m->m_pkthdr.PH_loc.eight[0] = len16;
2338 }
2339 
2340 #ifdef RATELIMIT
2341 static inline int
2342 mbuf_eo_nsegs(struct mbuf *m)
2343 {
2344 
2345 	M_ASSERTPKTHDR(m);
2346 	return (m->m_pkthdr.PH_loc.eight[1]);
2347 }
2348 
2349 static inline void
2350 set_mbuf_eo_nsegs(struct mbuf *m, uint8_t nsegs)
2351 {
2352 
2353 	M_ASSERTPKTHDR(m);
2354 	m->m_pkthdr.PH_loc.eight[1] = nsegs;
2355 }
2356 
2357 static inline int
2358 mbuf_eo_len16(struct mbuf *m)
2359 {
2360 	int n;
2361 
2362 	M_ASSERTPKTHDR(m);
2363 	n = m->m_pkthdr.PH_loc.eight[2];
2364 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2365 
2366 	return (n);
2367 }
2368 
2369 static inline void
2370 set_mbuf_eo_len16(struct mbuf *m, uint8_t len16)
2371 {
2372 
2373 	M_ASSERTPKTHDR(m);
2374 	m->m_pkthdr.PH_loc.eight[2] = len16;
2375 }
2376 
2377 static inline int
2378 mbuf_eo_tsclk_tsoff(struct mbuf *m)
2379 {
2380 
2381 	M_ASSERTPKTHDR(m);
2382 	return (m->m_pkthdr.PH_loc.eight[3]);
2383 }
2384 
2385 static inline void
2386 set_mbuf_eo_tsclk_tsoff(struct mbuf *m, uint8_t tsclk_tsoff)
2387 {
2388 
2389 	M_ASSERTPKTHDR(m);
2390 	m->m_pkthdr.PH_loc.eight[3] = tsclk_tsoff;
2391 }
2392 
2393 static inline int
2394 needs_eo(struct m_snd_tag *mst)
2395 {
2396 
2397 	return (mst != NULL && mst->type == IF_SND_TAG_TYPE_RATE_LIMIT);
2398 }
2399 #endif
2400 
2401 /*
2402  * Try to allocate an mbuf to contain a raw work request.  To make it
2403  * easy to construct the work request, don't allocate a chain but a
2404  * single mbuf.
2405  */
2406 struct mbuf *
2407 alloc_wr_mbuf(int len, int how)
2408 {
2409 	struct mbuf *m;
2410 
2411 	if (len <= MHLEN)
2412 		m = m_gethdr(how, MT_DATA);
2413 	else if (len <= MCLBYTES)
2414 		m = m_getcl(how, MT_DATA, M_PKTHDR);
2415 	else
2416 		m = NULL;
2417 	if (m == NULL)
2418 		return (NULL);
2419 	m->m_pkthdr.len = len;
2420 	m->m_len = len;
2421 	set_mbuf_cflags(m, MC_RAW_WR);
2422 	set_mbuf_len16(m, howmany(len, 16));
2423 	return (m);
2424 }
2425 
2426 static inline bool
2427 needs_hwcsum(struct mbuf *m)
2428 {
2429 	const uint32_t csum_flags = CSUM_IP | CSUM_IP_UDP | CSUM_IP_TCP |
2430 	    CSUM_IP_TSO | CSUM_INNER_IP | CSUM_INNER_IP_UDP |
2431 	    CSUM_INNER_IP_TCP | CSUM_INNER_IP_TSO | CSUM_IP6_UDP |
2432 	    CSUM_IP6_TCP | CSUM_IP6_TSO | CSUM_INNER_IP6_UDP |
2433 	    CSUM_INNER_IP6_TCP | CSUM_INNER_IP6_TSO;
2434 
2435 	M_ASSERTPKTHDR(m);
2436 
2437 	return (m->m_pkthdr.csum_flags & csum_flags);
2438 }
2439 
2440 static inline bool
2441 needs_tso(struct mbuf *m)
2442 {
2443 	const uint32_t csum_flags = CSUM_IP_TSO | CSUM_IP6_TSO |
2444 	    CSUM_INNER_IP_TSO | CSUM_INNER_IP6_TSO;
2445 
2446 	M_ASSERTPKTHDR(m);
2447 
2448 	return (m->m_pkthdr.csum_flags & csum_flags);
2449 }
2450 
2451 static inline bool
2452 needs_vxlan_csum(struct mbuf *m)
2453 {
2454 
2455 	M_ASSERTPKTHDR(m);
2456 
2457 	return (m->m_pkthdr.csum_flags & CSUM_ENCAP_VXLAN);
2458 }
2459 
2460 static inline bool
2461 needs_vxlan_tso(struct mbuf *m)
2462 {
2463 	const uint32_t csum_flags = CSUM_ENCAP_VXLAN | CSUM_INNER_IP_TSO |
2464 	    CSUM_INNER_IP6_TSO;
2465 
2466 	M_ASSERTPKTHDR(m);
2467 
2468 	return ((m->m_pkthdr.csum_flags & csum_flags) != 0 &&
2469 	    (m->m_pkthdr.csum_flags & csum_flags) != CSUM_ENCAP_VXLAN);
2470 }
2471 
2472 static inline bool
2473 needs_inner_tcp_csum(struct mbuf *m)
2474 {
2475 	const uint32_t csum_flags = CSUM_INNER_IP_TSO | CSUM_INNER_IP6_TSO;
2476 
2477 	M_ASSERTPKTHDR(m);
2478 
2479 	return (m->m_pkthdr.csum_flags & csum_flags);
2480 }
2481 
2482 static inline bool
2483 needs_l3_csum(struct mbuf *m)
2484 {
2485 	const uint32_t csum_flags = CSUM_IP | CSUM_IP_TSO | CSUM_INNER_IP |
2486 	    CSUM_INNER_IP_TSO;
2487 
2488 	M_ASSERTPKTHDR(m);
2489 
2490 	return (m->m_pkthdr.csum_flags & csum_flags);
2491 }
2492 
2493 static inline bool
2494 needs_outer_tcp_csum(struct mbuf *m)
2495 {
2496 	const uint32_t csum_flags = CSUM_IP_TCP | CSUM_IP_TSO | CSUM_IP6_TCP |
2497 	    CSUM_IP6_TSO;
2498 
2499 	M_ASSERTPKTHDR(m);
2500 
2501 	return (m->m_pkthdr.csum_flags & csum_flags);
2502 }
2503 
2504 #ifdef RATELIMIT
2505 static inline bool
2506 needs_outer_l4_csum(struct mbuf *m)
2507 {
2508 	const uint32_t csum_flags = CSUM_IP_UDP | CSUM_IP_TCP | CSUM_IP_TSO |
2509 	    CSUM_IP6_UDP | CSUM_IP6_TCP | CSUM_IP6_TSO;
2510 
2511 	M_ASSERTPKTHDR(m);
2512 
2513 	return (m->m_pkthdr.csum_flags & csum_flags);
2514 }
2515 
2516 static inline bool
2517 needs_outer_udp_csum(struct mbuf *m)
2518 {
2519 	const uint32_t csum_flags = CSUM_IP_UDP | CSUM_IP6_UDP;
2520 
2521 	M_ASSERTPKTHDR(m);
2522 
2523 	return (m->m_pkthdr.csum_flags & csum_flags);
2524 }
2525 #endif
2526 
2527 static inline bool
2528 needs_vlan_insertion(struct mbuf *m)
2529 {
2530 
2531 	M_ASSERTPKTHDR(m);
2532 
2533 	return (m->m_flags & M_VLANTAG);
2534 }
2535 
2536 static void *
2537 m_advance(struct mbuf **pm, int *poffset, int len)
2538 {
2539 	struct mbuf *m = *pm;
2540 	int offset = *poffset;
2541 	uintptr_t p = 0;
2542 
2543 	MPASS(len > 0);
2544 
2545 	for (;;) {
2546 		if (offset + len < m->m_len) {
2547 			offset += len;
2548 			p = mtod(m, uintptr_t) + offset;
2549 			break;
2550 		}
2551 		len -= m->m_len - offset;
2552 		m = m->m_next;
2553 		offset = 0;
2554 		MPASS(m != NULL);
2555 	}
2556 	*poffset = offset;
2557 	*pm = m;
2558 	return ((void *)p);
2559 }
2560 
2561 static inline int
2562 count_mbuf_ext_pgs(struct mbuf *m, int skip, vm_paddr_t *nextaddr)
2563 {
2564 	vm_paddr_t paddr;
2565 	int i, len, off, pglen, pgoff, seglen, segoff;
2566 	int nsegs = 0;
2567 
2568 	M_ASSERTEXTPG(m);
2569 	off = mtod(m, vm_offset_t);
2570 	len = m->m_len;
2571 	off += skip;
2572 	len -= skip;
2573 
2574 	if (m->m_epg_hdrlen != 0) {
2575 		if (off >= m->m_epg_hdrlen) {
2576 			off -= m->m_epg_hdrlen;
2577 		} else {
2578 			seglen = m->m_epg_hdrlen - off;
2579 			segoff = off;
2580 			seglen = min(seglen, len);
2581 			off = 0;
2582 			len -= seglen;
2583 			paddr = pmap_kextract(
2584 			    (vm_offset_t)&m->m_epg_hdr[segoff]);
2585 			if (*nextaddr != paddr)
2586 				nsegs++;
2587 			*nextaddr = paddr + seglen;
2588 		}
2589 	}
2590 	pgoff = m->m_epg_1st_off;
2591 	for (i = 0; i < m->m_epg_npgs && len > 0; i++) {
2592 		pglen = m_epg_pagelen(m, i, pgoff);
2593 		if (off >= pglen) {
2594 			off -= pglen;
2595 			pgoff = 0;
2596 			continue;
2597 		}
2598 		seglen = pglen - off;
2599 		segoff = pgoff + off;
2600 		off = 0;
2601 		seglen = min(seglen, len);
2602 		len -= seglen;
2603 		paddr = m->m_epg_pa[i] + segoff;
2604 		if (*nextaddr != paddr)
2605 			nsegs++;
2606 		*nextaddr = paddr + seglen;
2607 		pgoff = 0;
2608 	};
2609 	if (len != 0) {
2610 		seglen = min(len, m->m_epg_trllen - off);
2611 		len -= seglen;
2612 		paddr = pmap_kextract((vm_offset_t)&m->m_epg_trail[off]);
2613 		if (*nextaddr != paddr)
2614 			nsegs++;
2615 		*nextaddr = paddr + seglen;
2616 	}
2617 
2618 	return (nsegs);
2619 }
2620 
2621 
2622 /*
2623  * Can deal with empty mbufs in the chain that have m_len = 0, but the chain
2624  * must have at least one mbuf that's not empty.  It is possible for this
2625  * routine to return 0 if skip accounts for all the contents of the mbuf chain.
2626  */
2627 static inline int
2628 count_mbuf_nsegs(struct mbuf *m, int skip, uint8_t *cflags)
2629 {
2630 	vm_paddr_t nextaddr, paddr;
2631 	vm_offset_t va;
2632 	int len, nsegs;
2633 
2634 	M_ASSERTPKTHDR(m);
2635 	MPASS(m->m_pkthdr.len > 0);
2636 	MPASS(m->m_pkthdr.len >= skip);
2637 
2638 	nsegs = 0;
2639 	nextaddr = 0;
2640 	for (; m; m = m->m_next) {
2641 		len = m->m_len;
2642 		if (__predict_false(len == 0))
2643 			continue;
2644 		if (skip >= len) {
2645 			skip -= len;
2646 			continue;
2647 		}
2648 		if ((m->m_flags & M_EXTPG) != 0) {
2649 			*cflags |= MC_NOMAP;
2650 			nsegs += count_mbuf_ext_pgs(m, skip, &nextaddr);
2651 			skip = 0;
2652 			continue;
2653 		}
2654 		va = mtod(m, vm_offset_t) + skip;
2655 		len -= skip;
2656 		skip = 0;
2657 		paddr = pmap_kextract(va);
2658 		nsegs += sglist_count((void *)(uintptr_t)va, len);
2659 		if (paddr == nextaddr)
2660 			nsegs--;
2661 		nextaddr = pmap_kextract(va + len - 1) + 1;
2662 	}
2663 
2664 	return (nsegs);
2665 }
2666 
2667 /*
2668  * The maximum number of segments that can fit in a WR.
2669  */
2670 static int
2671 max_nsegs_allowed(struct mbuf *m, bool vm_wr)
2672 {
2673 
2674 	if (vm_wr) {
2675 		if (needs_tso(m))
2676 			return (TX_SGL_SEGS_VM_TSO);
2677 		return (TX_SGL_SEGS_VM);
2678 	}
2679 
2680 	if (needs_tso(m)) {
2681 		if (needs_vxlan_tso(m))
2682 			return (TX_SGL_SEGS_VXLAN_TSO);
2683 		else
2684 			return (TX_SGL_SEGS_TSO);
2685 	}
2686 
2687 	return (TX_SGL_SEGS);
2688 }
2689 
2690 /*
2691  * Analyze the mbuf to determine its tx needs.  The mbuf passed in may change:
2692  * a) caller can assume it's been freed if this function returns with an error.
2693  * b) it may get defragged up if the gather list is too long for the hardware.
2694  */
2695 int
2696 parse_pkt(struct mbuf **mp, bool vm_wr)
2697 {
2698 	struct mbuf *m0 = *mp, *m;
2699 	int rc, nsegs, defragged = 0, offset;
2700 	struct ether_header *eh;
2701 	void *l3hdr;
2702 #if defined(INET) || defined(INET6)
2703 	struct tcphdr *tcp;
2704 #endif
2705 #if defined(KERN_TLS) || defined(RATELIMIT)
2706 	struct m_snd_tag *mst;
2707 #endif
2708 	uint16_t eh_type;
2709 	uint8_t cflags;
2710 
2711 	cflags = 0;
2712 	M_ASSERTPKTHDR(m0);
2713 	if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) {
2714 		rc = EINVAL;
2715 fail:
2716 		m_freem(m0);
2717 		*mp = NULL;
2718 		return (rc);
2719 	}
2720 restart:
2721 	/*
2722 	 * First count the number of gather list segments in the payload.
2723 	 * Defrag the mbuf if nsegs exceeds the hardware limit.
2724 	 */
2725 	M_ASSERTPKTHDR(m0);
2726 	MPASS(m0->m_pkthdr.len > 0);
2727 	nsegs = count_mbuf_nsegs(m0, 0, &cflags);
2728 #if defined(KERN_TLS) || defined(RATELIMIT)
2729 	if (m0->m_pkthdr.csum_flags & CSUM_SND_TAG)
2730 		mst = m0->m_pkthdr.snd_tag;
2731 	else
2732 		mst = NULL;
2733 #endif
2734 #ifdef KERN_TLS
2735 	if (mst != NULL && mst->type == IF_SND_TAG_TYPE_TLS) {
2736 		int len16;
2737 
2738 		cflags |= MC_TLS;
2739 		set_mbuf_cflags(m0, cflags);
2740 		rc = t6_ktls_parse_pkt(m0, &nsegs, &len16);
2741 		if (rc != 0)
2742 			goto fail;
2743 		set_mbuf_nsegs(m0, nsegs);
2744 		set_mbuf_len16(m0, len16);
2745 		return (0);
2746 	}
2747 #endif
2748 	if (nsegs > max_nsegs_allowed(m0, vm_wr)) {
2749 		if (defragged++ > 0) {
2750 			rc = EFBIG;
2751 			goto fail;
2752 		}
2753 		counter_u64_add(defrags, 1);
2754 		if ((m = m_defrag(m0, M_NOWAIT)) == NULL) {
2755 			rc = ENOMEM;
2756 			goto fail;
2757 		}
2758 		*mp = m0 = m;	/* update caller's copy after defrag */
2759 		goto restart;
2760 	}
2761 
2762 	if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN &&
2763 	    !(cflags & MC_NOMAP))) {
2764 		counter_u64_add(pullups, 1);
2765 		m0 = m_pullup(m0, m0->m_pkthdr.len);
2766 		if (m0 == NULL) {
2767 			/* Should have left well enough alone. */
2768 			rc = EFBIG;
2769 			goto fail;
2770 		}
2771 		*mp = m0;	/* update caller's copy after pullup */
2772 		goto restart;
2773 	}
2774 	set_mbuf_nsegs(m0, nsegs);
2775 	set_mbuf_cflags(m0, cflags);
2776 	calculate_mbuf_len16(m0, vm_wr);
2777 
2778 #ifdef RATELIMIT
2779 	/*
2780 	 * Ethofld is limited to TCP and UDP for now, and only when L4 hw
2781 	 * checksumming is enabled.  needs_outer_l4_csum happens to check for
2782 	 * all the right things.
2783 	 */
2784 	if (__predict_false(needs_eo(mst) && !needs_outer_l4_csum(m0))) {
2785 		m_snd_tag_rele(m0->m_pkthdr.snd_tag);
2786 		m0->m_pkthdr.snd_tag = NULL;
2787 		m0->m_pkthdr.csum_flags &= ~CSUM_SND_TAG;
2788 		mst = NULL;
2789 	}
2790 #endif
2791 
2792 	if (!needs_hwcsum(m0)
2793 #ifdef RATELIMIT
2794    		 && !needs_eo(mst)
2795 #endif
2796 	)
2797 		return (0);
2798 
2799 	m = m0;
2800 	eh = mtod(m, struct ether_header *);
2801 	eh_type = ntohs(eh->ether_type);
2802 	if (eh_type == ETHERTYPE_VLAN) {
2803 		struct ether_vlan_header *evh = (void *)eh;
2804 
2805 		eh_type = ntohs(evh->evl_proto);
2806 		m0->m_pkthdr.l2hlen = sizeof(*evh);
2807 	} else
2808 		m0->m_pkthdr.l2hlen = sizeof(*eh);
2809 
2810 	offset = 0;
2811 	l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2812 
2813 	switch (eh_type) {
2814 #ifdef INET6
2815 	case ETHERTYPE_IPV6:
2816 		m0->m_pkthdr.l3hlen = sizeof(struct ip6_hdr);
2817 		break;
2818 #endif
2819 #ifdef INET
2820 	case ETHERTYPE_IP:
2821 	{
2822 		struct ip *ip = l3hdr;
2823 
2824 		if (needs_vxlan_csum(m0)) {
2825 			/* Driver will do the outer IP hdr checksum. */
2826 			ip->ip_sum = 0;
2827 			if (needs_vxlan_tso(m0)) {
2828 				const uint16_t ipl = ip->ip_len;
2829 
2830 				ip->ip_len = 0;
2831 				ip->ip_sum = ~in_cksum_hdr(ip);
2832 				ip->ip_len = ipl;
2833 			} else
2834 				ip->ip_sum = in_cksum_hdr(ip);
2835 		}
2836 		m0->m_pkthdr.l3hlen = ip->ip_hl << 2;
2837 		break;
2838 	}
2839 #endif
2840 	default:
2841 		panic("%s: ethertype 0x%04x unknown.  if_cxgbe must be compiled"
2842 		    " with the same INET/INET6 options as the kernel.",
2843 		    __func__, eh_type);
2844 	}
2845 
2846 	if (needs_vxlan_csum(m0)) {
2847 		m0->m_pkthdr.l4hlen = sizeof(struct udphdr);
2848 		m0->m_pkthdr.l5hlen = sizeof(struct vxlan_header);
2849 
2850 		/* Inner headers. */
2851 		eh = m_advance(&m, &offset, m0->m_pkthdr.l3hlen +
2852 		    sizeof(struct udphdr) + sizeof(struct vxlan_header));
2853 		eh_type = ntohs(eh->ether_type);
2854 		if (eh_type == ETHERTYPE_VLAN) {
2855 			struct ether_vlan_header *evh = (void *)eh;
2856 
2857 			eh_type = ntohs(evh->evl_proto);
2858 			m0->m_pkthdr.inner_l2hlen = sizeof(*evh);
2859 		} else
2860 			m0->m_pkthdr.inner_l2hlen = sizeof(*eh);
2861 		l3hdr = m_advance(&m, &offset, m0->m_pkthdr.inner_l2hlen);
2862 
2863 		switch (eh_type) {
2864 #ifdef INET6
2865 		case ETHERTYPE_IPV6:
2866 			m0->m_pkthdr.inner_l3hlen = sizeof(struct ip6_hdr);
2867 			break;
2868 #endif
2869 #ifdef INET
2870 		case ETHERTYPE_IP:
2871 		{
2872 			struct ip *ip = l3hdr;
2873 
2874 			m0->m_pkthdr.inner_l3hlen = ip->ip_hl << 2;
2875 			break;
2876 		}
2877 #endif
2878 		default:
2879 			panic("%s: VXLAN hw offload requested with unknown "
2880 			    "ethertype 0x%04x.  if_cxgbe must be compiled"
2881 			    " with the same INET/INET6 options as the kernel.",
2882 			    __func__, eh_type);
2883 		}
2884 #if defined(INET) || defined(INET6)
2885 		if (needs_inner_tcp_csum(m0)) {
2886 			tcp = m_advance(&m, &offset, m0->m_pkthdr.inner_l3hlen);
2887 			m0->m_pkthdr.inner_l4hlen = tcp->th_off * 4;
2888 		}
2889 #endif
2890 		MPASS((m0->m_pkthdr.csum_flags & CSUM_SND_TAG) == 0);
2891 		m0->m_pkthdr.csum_flags &= CSUM_INNER_IP6_UDP |
2892 		    CSUM_INNER_IP6_TCP | CSUM_INNER_IP6_TSO | CSUM_INNER_IP |
2893 		    CSUM_INNER_IP_UDP | CSUM_INNER_IP_TCP | CSUM_INNER_IP_TSO |
2894 		    CSUM_ENCAP_VXLAN;
2895 	}
2896 
2897 #if defined(INET) || defined(INET6)
2898 	if (needs_outer_tcp_csum(m0)) {
2899 		tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen);
2900 		m0->m_pkthdr.l4hlen = tcp->th_off * 4;
2901 #ifdef RATELIMIT
2902 		if (tsclk >= 0 && *(uint32_t *)(tcp + 1) == ntohl(0x0101080a)) {
2903 			set_mbuf_eo_tsclk_tsoff(m0,
2904 			    V_FW_ETH_TX_EO_WR_TSCLK(tsclk) |
2905 			    V_FW_ETH_TX_EO_WR_TSOFF(sizeof(*tcp) / 2 + 1));
2906 		} else
2907 			set_mbuf_eo_tsclk_tsoff(m0, 0);
2908 	} else if (needs_outer_udp_csum(m0)) {
2909 		m0->m_pkthdr.l4hlen = sizeof(struct udphdr);
2910 #endif
2911 	}
2912 #ifdef RATELIMIT
2913 	if (needs_eo(mst)) {
2914 		u_int immhdrs;
2915 
2916 		/* EO WRs have the headers in the WR and not the GL. */
2917 		immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen +
2918 		    m0->m_pkthdr.l4hlen;
2919 		cflags = 0;
2920 		nsegs = count_mbuf_nsegs(m0, immhdrs, &cflags);
2921 		MPASS(cflags == mbuf_cflags(m0));
2922 		set_mbuf_eo_nsegs(m0, nsegs);
2923 		set_mbuf_eo_len16(m0,
2924 		    txpkt_eo_len16(nsegs, immhdrs, needs_tso(m0)));
2925 	}
2926 #endif
2927 #endif
2928 	MPASS(m0 == *mp);
2929 	return (0);
2930 }
2931 
2932 void *
2933 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie)
2934 {
2935 	struct sge_eq *eq = &wrq->eq;
2936 	struct adapter *sc = wrq->adapter;
2937 	int ndesc, available;
2938 	struct wrqe *wr;
2939 	void *w;
2940 
2941 	MPASS(len16 > 0);
2942 	ndesc = tx_len16_to_desc(len16);
2943 	MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC);
2944 
2945 	EQ_LOCK(eq);
2946 
2947 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2948 		drain_wrq_wr_list(sc, wrq);
2949 
2950 	if (!STAILQ_EMPTY(&wrq->wr_list)) {
2951 slowpath:
2952 		EQ_UNLOCK(eq);
2953 		wr = alloc_wrqe(len16 * 16, wrq);
2954 		if (__predict_false(wr == NULL))
2955 			return (NULL);
2956 		cookie->pidx = -1;
2957 		cookie->ndesc = ndesc;
2958 		return (&wr->wr);
2959 	}
2960 
2961 	eq->cidx = read_hw_cidx(eq);
2962 	if (eq->pidx == eq->cidx)
2963 		available = eq->sidx - 1;
2964 	else
2965 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2966 	if (available < ndesc)
2967 		goto slowpath;
2968 
2969 	cookie->pidx = eq->pidx;
2970 	cookie->ndesc = ndesc;
2971 	TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link);
2972 
2973 	w = &eq->desc[eq->pidx];
2974 	IDXINCR(eq->pidx, ndesc, eq->sidx);
2975 	if (__predict_false(cookie->pidx + ndesc > eq->sidx)) {
2976 		w = &wrq->ss[0];
2977 		wrq->ss_pidx = cookie->pidx;
2978 		wrq->ss_len = len16 * 16;
2979 	}
2980 
2981 	EQ_UNLOCK(eq);
2982 
2983 	return (w);
2984 }
2985 
2986 void
2987 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie)
2988 {
2989 	struct sge_eq *eq = &wrq->eq;
2990 	struct adapter *sc = wrq->adapter;
2991 	int ndesc, pidx;
2992 	struct wrq_cookie *prev, *next;
2993 
2994 	if (cookie->pidx == -1) {
2995 		struct wrqe *wr = __containerof(w, struct wrqe, wr);
2996 
2997 		t4_wrq_tx(sc, wr);
2998 		return;
2999 	}
3000 
3001 	if (__predict_false(w == &wrq->ss[0])) {
3002 		int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE;
3003 
3004 		MPASS(wrq->ss_len > n);	/* WR had better wrap around. */
3005 		bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n);
3006 		bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n);
3007 		wrq->tx_wrs_ss++;
3008 	} else
3009 		wrq->tx_wrs_direct++;
3010 
3011 	EQ_LOCK(eq);
3012 	ndesc = cookie->ndesc;	/* Can be more than SGE_MAX_WR_NDESC here. */
3013 	pidx = cookie->pidx;
3014 	MPASS(pidx >= 0 && pidx < eq->sidx);
3015 	prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link);
3016 	next = TAILQ_NEXT(cookie, link);
3017 	if (prev == NULL) {
3018 		MPASS(pidx == eq->dbidx);
3019 		if (next == NULL || ndesc >= 16) {
3020 			int available;
3021 			struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
3022 
3023 			/*
3024 			 * Note that the WR via which we'll request tx updates
3025 			 * is at pidx and not eq->pidx, which has moved on
3026 			 * already.
3027 			 */
3028 			dst = (void *)&eq->desc[pidx];
3029 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
3030 			if (available < eq->sidx / 4 &&
3031 			    atomic_cmpset_int(&eq->equiq, 0, 1)) {
3032 				/*
3033 				 * XXX: This is not 100% reliable with some
3034 				 * types of WRs.  But this is a very unusual
3035 				 * situation for an ofld/ctrl queue anyway.
3036 				 */
3037 				dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
3038 				    F_FW_WR_EQUEQ);
3039 			}
3040 
3041 			ring_eq_db(wrq->adapter, eq, ndesc);
3042 		} else {
3043 			MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc);
3044 			next->pidx = pidx;
3045 			next->ndesc += ndesc;
3046 		}
3047 	} else {
3048 		MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc);
3049 		prev->ndesc += ndesc;
3050 	}
3051 	TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link);
3052 
3053 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
3054 		drain_wrq_wr_list(sc, wrq);
3055 
3056 #ifdef INVARIANTS
3057 	if (TAILQ_EMPTY(&wrq->incomplete_wrs)) {
3058 		/* Doorbell must have caught up to the pidx. */
3059 		MPASS(wrq->eq.pidx == wrq->eq.dbidx);
3060 	}
3061 #endif
3062 	EQ_UNLOCK(eq);
3063 }
3064 
3065 static u_int
3066 can_resume_eth_tx(struct mp_ring *r)
3067 {
3068 	struct sge_eq *eq = r->cookie;
3069 
3070 	return (total_available_tx_desc(eq) > eq->sidx / 8);
3071 }
3072 
3073 static inline bool
3074 cannot_use_txpkts(struct mbuf *m)
3075 {
3076 	/* maybe put a GL limit too, to avoid silliness? */
3077 
3078 	return (needs_tso(m) || (mbuf_cflags(m) & (MC_RAW_WR | MC_TLS)) != 0);
3079 }
3080 
3081 static inline int
3082 discard_tx(struct sge_eq *eq)
3083 {
3084 
3085 	return ((eq->flags & (EQ_ENABLED | EQ_QFLUSH)) != EQ_ENABLED);
3086 }
3087 
3088 static inline int
3089 wr_can_update_eq(void *p)
3090 {
3091 	struct fw_eth_tx_pkts_wr *wr = p;
3092 
3093 	switch (G_FW_WR_OP(be32toh(wr->op_pkd))) {
3094 	case FW_ULPTX_WR:
3095 	case FW_ETH_TX_PKT_WR:
3096 	case FW_ETH_TX_PKTS_WR:
3097 	case FW_ETH_TX_PKTS2_WR:
3098 	case FW_ETH_TX_PKT_VM_WR:
3099 	case FW_ETH_TX_PKTS_VM_WR:
3100 		return (1);
3101 	default:
3102 		return (0);
3103 	}
3104 }
3105 
3106 static inline void
3107 set_txupdate_flags(struct sge_txq *txq, u_int avail,
3108     struct fw_eth_tx_pkt_wr *wr)
3109 {
3110 	struct sge_eq *eq = &txq->eq;
3111 	struct txpkts *txp = &txq->txp;
3112 
3113 	if ((txp->npkt > 0 || avail < eq->sidx / 2) &&
3114 	    atomic_cmpset_int(&eq->equiq, 0, 1)) {
3115 		wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ | F_FW_WR_EQUIQ);
3116 		eq->equeqidx = eq->pidx;
3117 	} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
3118 		wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
3119 		eq->equeqidx = eq->pidx;
3120 	}
3121 }
3122 
3123 /*
3124  * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to
3125  * be consumed.  Return the actual number consumed.  0 indicates a stall.
3126  */
3127 static u_int
3128 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx, bool *coalescing)
3129 {
3130 	struct sge_txq *txq = r->cookie;
3131 	struct ifnet *ifp = txq->ifp;
3132 	struct sge_eq *eq = &txq->eq;
3133 	struct txpkts *txp = &txq->txp;
3134 	struct vi_info *vi = ifp->if_softc;
3135 	struct adapter *sc = vi->adapter;
3136 	u_int total, remaining;		/* # of packets */
3137 	u_int n, avail, dbdiff;		/* # of hardware descriptors */
3138 	int i, rc;
3139 	struct mbuf *m0;
3140 	bool snd;
3141 	void *wr;	/* start of the last WR written to the ring */
3142 
3143 	TXQ_LOCK_ASSERT_OWNED(txq);
3144 
3145 	remaining = IDXDIFF(pidx, cidx, r->size);
3146 	if (__predict_false(discard_tx(eq))) {
3147 		for (i = 0; i < txp->npkt; i++)
3148 			m_freem(txp->mb[i]);
3149 		txp->npkt = 0;
3150 		while (cidx != pidx) {
3151 			m0 = r->items[cidx];
3152 			m_freem(m0);
3153 			if (++cidx == r->size)
3154 				cidx = 0;
3155 		}
3156 		reclaim_tx_descs(txq, eq->sidx);
3157 		*coalescing = false;
3158 		return (remaining);	/* emptied */
3159 	}
3160 
3161 	/* How many hardware descriptors do we have readily available. */
3162 	if (eq->pidx == eq->cidx) {
3163 		avail = eq->sidx - 1;
3164 		if (txp->score++ >= 5)
3165 			txp->score = 5;	/* tx is completely idle, reset. */
3166 	} else
3167 		avail = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
3168 
3169 	total = 0;
3170 	if (remaining == 0) {
3171 		if (txp->score-- == 1)	/* egr_update had to drain txpkts */
3172 			txp->score = 1;
3173 		goto send_txpkts;
3174 	}
3175 
3176 	dbdiff = 0;
3177 	MPASS(remaining > 0);
3178 	while (remaining > 0) {
3179 		m0 = r->items[cidx];
3180 		M_ASSERTPKTHDR(m0);
3181 		MPASS(m0->m_nextpkt == NULL);
3182 
3183 		if (avail < 2 * SGE_MAX_WR_NDESC)
3184 			avail += reclaim_tx_descs(txq, 64);
3185 
3186 		if (txp->npkt > 0 || remaining > 1 || txp->score > 3 ||
3187 		    atomic_load_int(&txq->eq.equiq) != 0) {
3188 			if (vi->flags & TX_USES_VM_WR)
3189 				rc = add_to_txpkts_vf(sc, txq, m0, avail, &snd);
3190 			else
3191 				rc = add_to_txpkts_pf(sc, txq, m0, avail, &snd);
3192 		} else {
3193 			snd = false;
3194 			rc = EINVAL;
3195 		}
3196 		if (snd) {
3197 			MPASS(txp->npkt > 0);
3198 			for (i = 0; i < txp->npkt; i++)
3199 				ETHER_BPF_MTAP(ifp, txp->mb[i]);
3200 			if (txp->npkt > 1) {
3201 				if (txp->score++ >= 10)
3202 					txp->score = 10;
3203 				MPASS(avail >= tx_len16_to_desc(txp->len16));
3204 				if (vi->flags & TX_USES_VM_WR)
3205 					n = write_txpkts_vm_wr(sc, txq);
3206 				else
3207 					n = write_txpkts_wr(sc, txq);
3208 			} else {
3209 				MPASS(avail >=
3210 				    tx_len16_to_desc(mbuf_len16(txp->mb[0])));
3211 				if (vi->flags & TX_USES_VM_WR)
3212 					n = write_txpkt_vm_wr(sc, txq,
3213 					    txp->mb[0]);
3214 				else
3215 					n = write_txpkt_wr(sc, txq, txp->mb[0],
3216 					    avail);
3217 			}
3218 			MPASS(n <= SGE_MAX_WR_NDESC);
3219 			avail -= n;
3220 			dbdiff += n;
3221 			wr = &eq->desc[eq->pidx];
3222 			IDXINCR(eq->pidx, n, eq->sidx);
3223 			txp->npkt = 0;	/* emptied */
3224 		}
3225 		if (rc == 0) {
3226 			/* m0 was coalesced into txq->txpkts. */
3227 			goto next_mbuf;
3228 		}
3229 		if (rc == EAGAIN) {
3230 			/*
3231 			 * m0 is suitable for tx coalescing but could not be
3232 			 * combined with the existing txq->txpkts, which has now
3233 			 * been transmitted.  Start a new txpkts with m0.
3234 			 */
3235 			MPASS(snd);
3236 			MPASS(txp->npkt == 0);
3237 			continue;
3238 		}
3239 
3240 		MPASS(rc != 0 && rc != EAGAIN);
3241 		MPASS(txp->npkt == 0);
3242 
3243 		n = tx_len16_to_desc(mbuf_len16(m0));
3244 		if (__predict_false(avail < n)) {
3245 			avail += reclaim_tx_descs(txq, min(n, 32));
3246 			if (avail < n)
3247 				break;	/* out of descriptors */
3248 		}
3249 
3250 		wr = &eq->desc[eq->pidx];
3251 		if (mbuf_cflags(m0) & MC_RAW_WR) {
3252 			n = write_raw_wr(txq, wr, m0, avail);
3253 #ifdef KERN_TLS
3254 		} else if (mbuf_cflags(m0) & MC_TLS) {
3255 			ETHER_BPF_MTAP(ifp, m0);
3256 			n = t6_ktls_write_wr(txq, wr, m0, mbuf_nsegs(m0),
3257 			    avail);
3258 #endif
3259 		} else {
3260 			ETHER_BPF_MTAP(ifp, m0);
3261 			if (vi->flags & TX_USES_VM_WR)
3262 				n = write_txpkt_vm_wr(sc, txq, m0);
3263 			else
3264 				n = write_txpkt_wr(sc, txq, m0, avail);
3265 		}
3266 		MPASS(n >= 1 && n <= avail);
3267 		if (!(mbuf_cflags(m0) & MC_TLS))
3268 			MPASS(n <= SGE_MAX_WR_NDESC);
3269 
3270 		avail -= n;
3271 		dbdiff += n;
3272 		IDXINCR(eq->pidx, n, eq->sidx);
3273 
3274 		if (dbdiff >= 512 / EQ_ESIZE) {	/* X_FETCHBURSTMAX_512B */
3275 			if (wr_can_update_eq(wr))
3276 				set_txupdate_flags(txq, avail, wr);
3277 			ring_eq_db(sc, eq, dbdiff);
3278 			avail += reclaim_tx_descs(txq, 32);
3279 			dbdiff = 0;
3280 		}
3281 next_mbuf:
3282 		total++;
3283 		remaining--;
3284 		if (__predict_false(++cidx == r->size))
3285 			cidx = 0;
3286 	}
3287 	if (dbdiff != 0) {
3288 		if (wr_can_update_eq(wr))
3289 			set_txupdate_flags(txq, avail, wr);
3290 		ring_eq_db(sc, eq, dbdiff);
3291 		reclaim_tx_descs(txq, 32);
3292 	} else if (eq->pidx == eq->cidx && txp->npkt > 0 &&
3293 	    atomic_load_int(&txq->eq.equiq) == 0) {
3294 		/*
3295 		 * If nothing was submitted to the chip for tx (it was coalesced
3296 		 * into txpkts instead) and there is no tx update outstanding
3297 		 * then we need to send txpkts now.
3298 		 */
3299 send_txpkts:
3300 		MPASS(txp->npkt > 0);
3301 		for (i = 0; i < txp->npkt; i++)
3302 			ETHER_BPF_MTAP(ifp, txp->mb[i]);
3303 		if (txp->npkt > 1) {
3304 			MPASS(avail >= tx_len16_to_desc(txp->len16));
3305 			if (vi->flags & TX_USES_VM_WR)
3306 				n = write_txpkts_vm_wr(sc, txq);
3307 			else
3308 				n = write_txpkts_wr(sc, txq);
3309 		} else {
3310 			MPASS(avail >=
3311 			    tx_len16_to_desc(mbuf_len16(txp->mb[0])));
3312 			if (vi->flags & TX_USES_VM_WR)
3313 				n = write_txpkt_vm_wr(sc, txq, txp->mb[0]);
3314 			else
3315 				n = write_txpkt_wr(sc, txq, txp->mb[0], avail);
3316 		}
3317 		MPASS(n <= SGE_MAX_WR_NDESC);
3318 		wr = &eq->desc[eq->pidx];
3319 		IDXINCR(eq->pidx, n, eq->sidx);
3320 		txp->npkt = 0;	/* emptied */
3321 
3322 		MPASS(wr_can_update_eq(wr));
3323 		set_txupdate_flags(txq, avail - n, wr);
3324 		ring_eq_db(sc, eq, n);
3325 		reclaim_tx_descs(txq, 32);
3326 	}
3327 	*coalescing = txp->npkt > 0;
3328 
3329 	return (total);
3330 }
3331 
3332 static inline void
3333 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
3334     int qsize)
3335 {
3336 
3337 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
3338 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
3339 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
3340 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
3341 
3342 	iq->flags = 0;
3343 	iq->adapter = sc;
3344 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
3345 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
3346 	if (pktc_idx >= 0) {
3347 		iq->intr_params |= F_QINTR_CNT_EN;
3348 		iq->intr_pktc_idx = pktc_idx;
3349 	}
3350 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
3351 	iq->sidx = iq->qsize - sc->params.sge.spg_len / IQ_ESIZE;
3352 }
3353 
3354 static inline void
3355 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name)
3356 {
3357 
3358 	fl->qsize = qsize;
3359 	fl->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
3360 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
3361 	if (sc->flags & BUF_PACKING_OK &&
3362 	    ((!is_t4(sc) && buffer_packing) ||	/* T5+: enabled unless 0 */
3363 	    (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */
3364 		fl->flags |= FL_BUF_PACKING;
3365 	fl->zidx = find_refill_source(sc, maxp, fl->flags & FL_BUF_PACKING);
3366 	fl->safe_zidx = sc->sge.safe_zidx;
3367 }
3368 
3369 static inline void
3370 init_eq(struct adapter *sc, struct sge_eq *eq, int eqtype, int qsize,
3371     uint8_t tx_chan, uint16_t iqid, char *name)
3372 {
3373 	KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
3374 
3375 	eq->flags = eqtype & EQ_TYPEMASK;
3376 	eq->tx_chan = tx_chan;
3377 	eq->iqid = iqid;
3378 	eq->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
3379 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
3380 }
3381 
3382 int
3383 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
3384     bus_dmamap_t *map, bus_addr_t *pa, void **va)
3385 {
3386 	int rc;
3387 
3388 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
3389 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
3390 	if (rc != 0) {
3391 		device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
3392 		goto done;
3393 	}
3394 
3395 	rc = bus_dmamem_alloc(*tag, va,
3396 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
3397 	if (rc != 0) {
3398 		device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
3399 		goto done;
3400 	}
3401 
3402 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
3403 	if (rc != 0) {
3404 		device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
3405 		goto done;
3406 	}
3407 done:
3408 	if (rc)
3409 		free_ring(sc, *tag, *map, *pa, *va);
3410 
3411 	return (rc);
3412 }
3413 
3414 int
3415 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
3416     bus_addr_t pa, void *va)
3417 {
3418 	if (pa)
3419 		bus_dmamap_unload(tag, map);
3420 	if (va)
3421 		bus_dmamem_free(tag, va, map);
3422 	if (tag)
3423 		bus_dma_tag_destroy(tag);
3424 
3425 	return (0);
3426 }
3427 
3428 /*
3429  * Allocates the ring for an ingress queue and an optional freelist.  If the
3430  * freelist is specified it will be allocated and then associated with the
3431  * ingress queue.
3432  *
3433  * Returns errno on failure.  Resources allocated up to that point may still be
3434  * allocated.  Caller is responsible for cleanup in case this function fails.
3435  *
3436  * If the ingress queue will take interrupts directly then the intr_idx
3437  * specifies the vector, starting from 0.  -1 means the interrupts for this
3438  * queue should be forwarded to the fwq.
3439  */
3440 static int
3441 alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl,
3442     int intr_idx, int cong)
3443 {
3444 	int rc, i, cntxt_id;
3445 	size_t len;
3446 	struct fw_iq_cmd c;
3447 	struct port_info *pi = vi->pi;
3448 	struct adapter *sc = iq->adapter;
3449 	struct sge_params *sp = &sc->params.sge;
3450 	__be32 v = 0;
3451 
3452 	len = iq->qsize * IQ_ESIZE;
3453 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
3454 	    (void **)&iq->desc);
3455 	if (rc != 0)
3456 		return (rc);
3457 
3458 	bzero(&c, sizeof(c));
3459 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
3460 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
3461 	    V_FW_IQ_CMD_VFN(0));
3462 
3463 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
3464 	    FW_LEN16(c));
3465 
3466 	/* Special handling for firmware event queue */
3467 	if (iq == &sc->sge.fwq)
3468 		v |= F_FW_IQ_CMD_IQASYNCH;
3469 
3470 	if (intr_idx < 0) {
3471 		/* Forwarded interrupts, all headed to fwq */
3472 		v |= F_FW_IQ_CMD_IQANDST;
3473 		v |= V_FW_IQ_CMD_IQANDSTINDEX(sc->sge.fwq.cntxt_id);
3474 	} else {
3475 		KASSERT(intr_idx < sc->intr_count,
3476 		    ("%s: invalid direct intr_idx %d", __func__, intr_idx));
3477 		v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
3478 	}
3479 
3480 	c.type_to_iqandstindex = htobe32(v |
3481 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
3482 	    V_FW_IQ_CMD_VIID(vi->viid) |
3483 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
3484 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
3485 	    F_FW_IQ_CMD_IQGTSMODE |
3486 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
3487 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
3488 	c.iqsize = htobe16(iq->qsize);
3489 	c.iqaddr = htobe64(iq->ba);
3490 	if (cong >= 0)
3491 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
3492 
3493 	if (fl) {
3494 		mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
3495 
3496 		len = fl->qsize * EQ_ESIZE;
3497 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
3498 		    &fl->ba, (void **)&fl->desc);
3499 		if (rc)
3500 			return (rc);
3501 
3502 		/* Allocate space for one software descriptor per buffer. */
3503 		rc = alloc_fl_sdesc(fl);
3504 		if (rc != 0) {
3505 			device_printf(sc->dev,
3506 			    "failed to setup fl software descriptors: %d\n",
3507 			    rc);
3508 			return (rc);
3509 		}
3510 
3511 		if (fl->flags & FL_BUF_PACKING) {
3512 			fl->lowat = roundup2(sp->fl_starve_threshold2, 8);
3513 			fl->buf_boundary = sp->pack_boundary;
3514 		} else {
3515 			fl->lowat = roundup2(sp->fl_starve_threshold, 8);
3516 			fl->buf_boundary = 16;
3517 		}
3518 		if (fl_pad && fl->buf_boundary < sp->pad_boundary)
3519 			fl->buf_boundary = sp->pad_boundary;
3520 
3521 		c.iqns_to_fl0congen |=
3522 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
3523 			F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
3524 			(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
3525 			(fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
3526 			    0));
3527 		if (cong >= 0) {
3528 			c.iqns_to_fl0congen |=
3529 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
3530 				    F_FW_IQ_CMD_FL0CONGCIF |
3531 				    F_FW_IQ_CMD_FL0CONGEN);
3532 		}
3533 		c.fl0dcaen_to_fl0cidxfthresh =
3534 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
3535 			X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B_T6) |
3536 			V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
3537 			X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
3538 		c.fl0size = htobe16(fl->qsize);
3539 		c.fl0addr = htobe64(fl->ba);
3540 	}
3541 
3542 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3543 	if (rc != 0) {
3544 		device_printf(sc->dev,
3545 		    "failed to create ingress queue: %d\n", rc);
3546 		return (rc);
3547 	}
3548 
3549 	iq->cidx = 0;
3550 	iq->gen = F_RSPD_GEN;
3551 	iq->intr_next = iq->intr_params;
3552 	iq->cntxt_id = be16toh(c.iqid);
3553 	iq->abs_id = be16toh(c.physiqid);
3554 	iq->flags |= IQ_ALLOCATED;
3555 
3556 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
3557 	if (cntxt_id >= sc->sge.iqmap_sz) {
3558 		panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
3559 		    cntxt_id, sc->sge.iqmap_sz - 1);
3560 	}
3561 	sc->sge.iqmap[cntxt_id] = iq;
3562 
3563 	if (fl) {
3564 		u_int qid;
3565 
3566 		iq->flags |= IQ_HAS_FL;
3567 		fl->cntxt_id = be16toh(c.fl0id);
3568 		fl->pidx = fl->cidx = 0;
3569 
3570 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
3571 		if (cntxt_id >= sc->sge.eqmap_sz) {
3572 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
3573 			    __func__, cntxt_id, sc->sge.eqmap_sz - 1);
3574 		}
3575 		sc->sge.eqmap[cntxt_id] = (void *)fl;
3576 
3577 		qid = fl->cntxt_id;
3578 		if (isset(&sc->doorbells, DOORBELL_UDB)) {
3579 			uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3580 			uint32_t mask = (1 << s_qpp) - 1;
3581 			volatile uint8_t *udb;
3582 
3583 			udb = sc->udbs_base + UDBS_DB_OFFSET;
3584 			udb += (qid >> s_qpp) << PAGE_SHIFT;
3585 			qid &= mask;
3586 			if (qid < PAGE_SIZE / UDBS_SEG_SIZE) {
3587 				udb += qid << UDBS_SEG_SHIFT;
3588 				qid = 0;
3589 			}
3590 			fl->udb = (volatile void *)udb;
3591 		}
3592 		fl->dbval = V_QID(qid) | sc->chip_params->sge_fl_db;
3593 
3594 		FL_LOCK(fl);
3595 		/* Enough to make sure the SGE doesn't think it's starved */
3596 		refill_fl(sc, fl, fl->lowat);
3597 		FL_UNLOCK(fl);
3598 	}
3599 
3600 	if (chip_id(sc) >= CHELSIO_T5 && !(sc->flags & IS_VF) && cong >= 0) {
3601 		uint32_t param, val;
3602 
3603 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
3604 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
3605 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
3606 		if (cong == 0)
3607 			val = 1 << 19;
3608 		else {
3609 			val = 2 << 19;
3610 			for (i = 0; i < 4; i++) {
3611 				if (cong & (1 << i))
3612 					val |= 1 << (i << 2);
3613 			}
3614 		}
3615 
3616 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
3617 		if (rc != 0) {
3618 			/* report error but carry on */
3619 			device_printf(sc->dev,
3620 			    "failed to set congestion manager context for "
3621 			    "ingress queue %d: %d\n", iq->cntxt_id, rc);
3622 		}
3623 	}
3624 
3625 	/* Enable IQ interrupts */
3626 	atomic_store_rel_int(&iq->state, IQS_IDLE);
3627 	t4_write_reg(sc, sc->sge_gts_reg, V_SEINTARM(iq->intr_params) |
3628 	    V_INGRESSQID(iq->cntxt_id));
3629 
3630 	return (0);
3631 }
3632 
3633 static int
3634 free_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl)
3635 {
3636 	int rc;
3637 	struct adapter *sc = iq->adapter;
3638 	device_t dev;
3639 
3640 	if (sc == NULL)
3641 		return (0);	/* nothing to do */
3642 
3643 	dev = vi ? vi->dev : sc->dev;
3644 
3645 	if (iq->flags & IQ_ALLOCATED) {
3646 		rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
3647 		    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
3648 		    fl ? fl->cntxt_id : 0xffff, 0xffff);
3649 		if (rc != 0) {
3650 			device_printf(dev,
3651 			    "failed to free queue %p: %d\n", iq, rc);
3652 			return (rc);
3653 		}
3654 		iq->flags &= ~IQ_ALLOCATED;
3655 	}
3656 
3657 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
3658 
3659 	bzero(iq, sizeof(*iq));
3660 
3661 	if (fl) {
3662 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
3663 		    fl->desc);
3664 
3665 		if (fl->sdesc)
3666 			free_fl_sdesc(sc, fl);
3667 
3668 		if (mtx_initialized(&fl->fl_lock))
3669 			mtx_destroy(&fl->fl_lock);
3670 
3671 		bzero(fl, sizeof(*fl));
3672 	}
3673 
3674 	return (0);
3675 }
3676 
3677 static void
3678 add_iq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
3679     struct sge_iq *iq)
3680 {
3681 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3682 
3683 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD, &iq->ba,
3684 	    "bus address of descriptor ring");
3685 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3686 	    iq->qsize * IQ_ESIZE, "descriptor ring size in bytes");
3687 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3688 	    CTLTYPE_INT | CTLFLAG_RD | CTLFLAG_MPSAFE, &iq->abs_id, 0,
3689 	    sysctl_uint16, "I", "absolute id of the queue");
3690 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3691 	    CTLTYPE_INT | CTLFLAG_RD | CTLFLAG_MPSAFE, &iq->cntxt_id, 0,
3692 	    sysctl_uint16, "I", "SGE context id of the queue");
3693 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3694 	    CTLTYPE_INT | CTLFLAG_RD | CTLFLAG_MPSAFE, &iq->cidx, 0,
3695 	    sysctl_uint16, "I", "consumer index");
3696 }
3697 
3698 static void
3699 add_fl_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
3700     struct sysctl_oid *oid, struct sge_fl *fl)
3701 {
3702 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3703 
3704 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl",
3705 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist");
3706 	children = SYSCTL_CHILDREN(oid);
3707 
3708 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3709 	    &fl->ba, "bus address of descriptor ring");
3710 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3711 	    fl->sidx * EQ_ESIZE + sc->params.sge.spg_len,
3712 	    "desc ring size in bytes");
3713 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3714 	    CTLTYPE_INT | CTLFLAG_RD | CTLFLAG_MPSAFE, &fl->cntxt_id, 0,
3715 	    sysctl_uint16, "I", "SGE context id of the freelist");
3716 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL,
3717 	    fl_pad ? 1 : 0, "padding enabled");
3718 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL,
3719 	    fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled");
3720 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
3721 	    0, "consumer index");
3722 	if (fl->flags & FL_BUF_PACKING) {
3723 		SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
3724 		    CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
3725 	}
3726 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
3727 	    0, "producer index");
3728 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
3729 	    CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
3730 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
3731 	    CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
3732 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
3733 	    CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
3734 }
3735 
3736 static int
3737 alloc_fwq(struct adapter *sc)
3738 {
3739 	int rc, intr_idx;
3740 	struct sge_iq *fwq = &sc->sge.fwq;
3741 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
3742 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3743 
3744 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE);
3745 	if (sc->flags & IS_VF)
3746 		intr_idx = 0;
3747 	else
3748 		intr_idx = sc->intr_count > 1 ? 1 : 0;
3749 	rc = alloc_iq_fl(&sc->port[0]->vi[0], fwq, NULL, intr_idx, -1);
3750 	if (rc != 0) {
3751 		device_printf(sc->dev,
3752 		    "failed to create firmware event queue: %d\n", rc);
3753 		return (rc);
3754 	}
3755 
3756 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq",
3757 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "firmware event queue");
3758 	add_iq_sysctls(&sc->ctx, oid, fwq);
3759 
3760 	return (0);
3761 }
3762 
3763 static int
3764 free_fwq(struct adapter *sc)
3765 {
3766 	return free_iq_fl(NULL, &sc->sge.fwq, NULL);
3767 }
3768 
3769 static int
3770 alloc_ctrlq(struct adapter *sc, struct sge_wrq *ctrlq, int idx,
3771     struct sysctl_oid *oid)
3772 {
3773 	int rc;
3774 	char name[16];
3775 	struct sysctl_oid_list *children;
3776 
3777 	snprintf(name, sizeof(name), "%s ctrlq%d", device_get_nameunit(sc->dev),
3778 	    idx);
3779 	init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[idx]->tx_chan,
3780 	    sc->sge.fwq.cntxt_id, name);
3781 
3782 	children = SYSCTL_CHILDREN(oid);
3783 	snprintf(name, sizeof(name), "%d", idx);
3784 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, name,
3785 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "ctrl queue");
3786 	rc = alloc_wrq(sc, NULL, ctrlq, oid);
3787 
3788 	return (rc);
3789 }
3790 
3791 int
3792 tnl_cong(struct port_info *pi, int drop)
3793 {
3794 
3795 	if (drop == -1)
3796 		return (-1);
3797 	else if (drop == 1)
3798 		return (0);
3799 	else
3800 		return (pi->rx_e_chan_map);
3801 }
3802 
3803 static int
3804 alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int intr_idx, int idx,
3805     struct sysctl_oid *oid)
3806 {
3807 	int rc;
3808 	struct adapter *sc = vi->adapter;
3809 	struct sysctl_oid_list *children;
3810 	char name[16];
3811 
3812 	rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, intr_idx,
3813 	    tnl_cong(vi->pi, cong_drop));
3814 	if (rc != 0)
3815 		return (rc);
3816 
3817 	if (idx == 0)
3818 		sc->sge.iq_base = rxq->iq.abs_id - rxq->iq.cntxt_id;
3819 	else
3820 		KASSERT(rxq->iq.cntxt_id + sc->sge.iq_base == rxq->iq.abs_id,
3821 		    ("iq_base mismatch"));
3822 	KASSERT(sc->sge.iq_base == 0 || sc->flags & IS_VF,
3823 	    ("PF with non-zero iq_base"));
3824 
3825 	/*
3826 	 * The freelist is just barely above the starvation threshold right now,
3827 	 * fill it up a bit more.
3828 	 */
3829 	FL_LOCK(&rxq->fl);
3830 	refill_fl(sc, &rxq->fl, 128);
3831 	FL_UNLOCK(&rxq->fl);
3832 
3833 #if defined(INET) || defined(INET6)
3834 	rc = tcp_lro_init_args(&rxq->lro, vi->ifp, lro_entries, lro_mbufs);
3835 	if (rc != 0)
3836 		return (rc);
3837 	MPASS(rxq->lro.ifp == vi->ifp);	/* also indicates LRO init'ed */
3838 
3839 	if (vi->ifp->if_capenable & IFCAP_LRO)
3840 		rxq->iq.flags |= IQ_LRO_ENABLED;
3841 #endif
3842 	if (vi->ifp->if_capenable & IFCAP_HWRXTSTMP)
3843 		rxq->iq.flags |= IQ_RX_TIMESTAMP;
3844 	rxq->ifp = vi->ifp;
3845 
3846 	children = SYSCTL_CHILDREN(oid);
3847 
3848 	snprintf(name, sizeof(name), "%d", idx);
3849 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name,
3850 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "rx queue");
3851 	children = SYSCTL_CHILDREN(oid);
3852 
3853 	add_iq_sysctls(&vi->ctx, oid, &rxq->iq);
3854 #if defined(INET) || defined(INET6)
3855 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
3856 	    &rxq->lro.lro_queued, 0, NULL);
3857 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
3858 	    &rxq->lro.lro_flushed, 0, NULL);
3859 #endif
3860 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
3861 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
3862 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_extraction",
3863 	    CTLFLAG_RD, &rxq->vlan_extraction,
3864 	    "# of times hardware extracted 802.1Q tag");
3865 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vxlan_rxcsum",
3866 	    CTLFLAG_RD, &rxq->vxlan_rxcsum,
3867 	    "# of times hardware assisted with inner checksum (VXLAN) ");
3868 
3869 	add_fl_sysctls(sc, &vi->ctx, oid, &rxq->fl);
3870 
3871 	return (rc);
3872 }
3873 
3874 static int
3875 free_rxq(struct vi_info *vi, struct sge_rxq *rxq)
3876 {
3877 	int rc;
3878 
3879 #if defined(INET) || defined(INET6)
3880 	if (rxq->lro.ifp) {
3881 		tcp_lro_free(&rxq->lro);
3882 		rxq->lro.ifp = NULL;
3883 	}
3884 #endif
3885 
3886 	rc = free_iq_fl(vi, &rxq->iq, &rxq->fl);
3887 	if (rc == 0)
3888 		bzero(rxq, sizeof(*rxq));
3889 
3890 	return (rc);
3891 }
3892 
3893 #ifdef TCP_OFFLOAD
3894 static int
3895 alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq,
3896     int intr_idx, int idx, struct sysctl_oid *oid)
3897 {
3898 	struct port_info *pi = vi->pi;
3899 	int rc;
3900 	struct sysctl_oid_list *children;
3901 	char name[16];
3902 
3903 	rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx, 0);
3904 	if (rc != 0)
3905 		return (rc);
3906 
3907 	children = SYSCTL_CHILDREN(oid);
3908 
3909 	snprintf(name, sizeof(name), "%d", idx);
3910 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name,
3911 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "rx queue");
3912 	add_iq_sysctls(&vi->ctx, oid, &ofld_rxq->iq);
3913 	add_fl_sysctls(pi->adapter, &vi->ctx, oid, &ofld_rxq->fl);
3914 
3915 	return (rc);
3916 }
3917 
3918 static int
3919 free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq)
3920 {
3921 	int rc;
3922 
3923 	rc = free_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl);
3924 	if (rc == 0)
3925 		bzero(ofld_rxq, sizeof(*ofld_rxq));
3926 
3927 	return (rc);
3928 }
3929 #endif
3930 
3931 /*
3932  * Returns a reasonable automatic cidx flush threshold for a given queue size.
3933  */
3934 static u_int
3935 qsize_to_fthresh(int qsize)
3936 {
3937 	u_int fthresh;
3938 
3939 	while (!powerof2(qsize))
3940 		qsize++;
3941 	fthresh = ilog2(qsize);
3942 	if (fthresh > X_CIDXFLUSHTHRESH_128)
3943 		fthresh = X_CIDXFLUSHTHRESH_128;
3944 
3945 	return (fthresh);
3946 }
3947 
3948 static int
3949 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
3950 {
3951 	int rc, cntxt_id;
3952 	struct fw_eq_ctrl_cmd c;
3953 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3954 
3955 	bzero(&c, sizeof(c));
3956 
3957 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
3958 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
3959 	    V_FW_EQ_CTRL_CMD_VFN(0));
3960 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
3961 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
3962 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid));
3963 	c.physeqid_pkd = htobe32(0);
3964 	c.fetchszm_to_iqid =
3965 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
3966 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
3967 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
3968 	c.dcaen_to_eqsize =
3969 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
3970 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
3971 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3972 		V_FW_EQ_CTRL_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
3973 		V_FW_EQ_CTRL_CMD_EQSIZE(qsize));
3974 	c.eqaddr = htobe64(eq->ba);
3975 
3976 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3977 	if (rc != 0) {
3978 		device_printf(sc->dev,
3979 		    "failed to create control queue %d: %d\n", eq->tx_chan, rc);
3980 		return (rc);
3981 	}
3982 	eq->flags |= EQ_ALLOCATED;
3983 
3984 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
3985 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3986 	if (cntxt_id >= sc->sge.eqmap_sz)
3987 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3988 		cntxt_id, sc->sge.eqmap_sz - 1);
3989 	sc->sge.eqmap[cntxt_id] = eq;
3990 
3991 	return (rc);
3992 }
3993 
3994 static int
3995 eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3996 {
3997 	int rc, cntxt_id;
3998 	struct fw_eq_eth_cmd c;
3999 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
4000 
4001 	bzero(&c, sizeof(c));
4002 
4003 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
4004 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
4005 	    V_FW_EQ_ETH_CMD_VFN(0));
4006 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
4007 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
4008 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
4009 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
4010 	c.fetchszm_to_iqid =
4011 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
4012 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
4013 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
4014 	c.dcaen_to_eqsize =
4015 	    htobe32(V_FW_EQ_ETH_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
4016 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
4017 		V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
4018 		V_FW_EQ_ETH_CMD_EQSIZE(qsize));
4019 	c.eqaddr = htobe64(eq->ba);
4020 
4021 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
4022 	if (rc != 0) {
4023 		device_printf(vi->dev,
4024 		    "failed to create Ethernet egress queue: %d\n", rc);
4025 		return (rc);
4026 	}
4027 	eq->flags |= EQ_ALLOCATED;
4028 
4029 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
4030 	eq->abs_id = G_FW_EQ_ETH_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
4031 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
4032 	if (cntxt_id >= sc->sge.eqmap_sz)
4033 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
4034 		cntxt_id, sc->sge.eqmap_sz - 1);
4035 	sc->sge.eqmap[cntxt_id] = eq;
4036 
4037 	return (rc);
4038 }
4039 
4040 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
4041 static int
4042 ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
4043 {
4044 	int rc, cntxt_id;
4045 	struct fw_eq_ofld_cmd c;
4046 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
4047 
4048 	bzero(&c, sizeof(c));
4049 
4050 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
4051 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
4052 	    V_FW_EQ_OFLD_CMD_VFN(0));
4053 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
4054 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
4055 	c.fetchszm_to_iqid =
4056 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
4057 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
4058 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
4059 	c.dcaen_to_eqsize =
4060 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
4061 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
4062 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
4063 		V_FW_EQ_OFLD_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
4064 		V_FW_EQ_OFLD_CMD_EQSIZE(qsize));
4065 	c.eqaddr = htobe64(eq->ba);
4066 
4067 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
4068 	if (rc != 0) {
4069 		device_printf(vi->dev,
4070 		    "failed to create egress queue for TCP offload: %d\n", rc);
4071 		return (rc);
4072 	}
4073 	eq->flags |= EQ_ALLOCATED;
4074 
4075 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
4076 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
4077 	if (cntxt_id >= sc->sge.eqmap_sz)
4078 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
4079 		cntxt_id, sc->sge.eqmap_sz - 1);
4080 	sc->sge.eqmap[cntxt_id] = eq;
4081 
4082 	return (rc);
4083 }
4084 #endif
4085 
4086 static int
4087 alloc_eq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
4088 {
4089 	int rc, qsize;
4090 	size_t len;
4091 
4092 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
4093 
4094 	qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
4095 	len = qsize * EQ_ESIZE;
4096 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
4097 	    &eq->ba, (void **)&eq->desc);
4098 	if (rc)
4099 		return (rc);
4100 
4101 	eq->pidx = eq->cidx = eq->dbidx = 0;
4102 	/* Note that equeqidx is not used with sge_wrq (OFLD/CTRL) queues. */
4103 	eq->equeqidx = 0;
4104 	eq->doorbells = sc->doorbells;
4105 
4106 	switch (eq->flags & EQ_TYPEMASK) {
4107 	case EQ_CTRL:
4108 		rc = ctrl_eq_alloc(sc, eq);
4109 		break;
4110 
4111 	case EQ_ETH:
4112 		rc = eth_eq_alloc(sc, vi, eq);
4113 		break;
4114 
4115 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
4116 	case EQ_OFLD:
4117 		rc = ofld_eq_alloc(sc, vi, eq);
4118 		break;
4119 #endif
4120 
4121 	default:
4122 		panic("%s: invalid eq type %d.", __func__,
4123 		    eq->flags & EQ_TYPEMASK);
4124 	}
4125 	if (rc != 0) {
4126 		device_printf(sc->dev,
4127 		    "failed to allocate egress queue(%d): %d\n",
4128 		    eq->flags & EQ_TYPEMASK, rc);
4129 	}
4130 
4131 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
4132 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
4133 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
4134 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
4135 		uint32_t mask = (1 << s_qpp) - 1;
4136 		volatile uint8_t *udb;
4137 
4138 		udb = sc->udbs_base + UDBS_DB_OFFSET;
4139 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
4140 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
4141 		if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
4142 	    		clrbit(&eq->doorbells, DOORBELL_WCWR);
4143 		else {
4144 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
4145 			eq->udb_qid = 0;
4146 		}
4147 		eq->udb = (volatile void *)udb;
4148 	}
4149 
4150 	return (rc);
4151 }
4152 
4153 static int
4154 free_eq(struct adapter *sc, struct sge_eq *eq)
4155 {
4156 	int rc;
4157 
4158 	if (eq->flags & EQ_ALLOCATED) {
4159 		switch (eq->flags & EQ_TYPEMASK) {
4160 		case EQ_CTRL:
4161 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
4162 			    eq->cntxt_id);
4163 			break;
4164 
4165 		case EQ_ETH:
4166 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
4167 			    eq->cntxt_id);
4168 			break;
4169 
4170 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
4171 		case EQ_OFLD:
4172 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
4173 			    eq->cntxt_id);
4174 			break;
4175 #endif
4176 
4177 		default:
4178 			panic("%s: invalid eq type %d.", __func__,
4179 			    eq->flags & EQ_TYPEMASK);
4180 		}
4181 		if (rc != 0) {
4182 			device_printf(sc->dev,
4183 			    "failed to free egress queue (%d): %d\n",
4184 			    eq->flags & EQ_TYPEMASK, rc);
4185 			return (rc);
4186 		}
4187 		eq->flags &= ~EQ_ALLOCATED;
4188 	}
4189 
4190 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
4191 
4192 	if (mtx_initialized(&eq->eq_lock))
4193 		mtx_destroy(&eq->eq_lock);
4194 
4195 	bzero(eq, sizeof(*eq));
4196 	return (0);
4197 }
4198 
4199 static int
4200 alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq,
4201     struct sysctl_oid *oid)
4202 {
4203 	int rc;
4204 	struct sysctl_ctx_list *ctx = vi ? &vi->ctx : &sc->ctx;
4205 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
4206 
4207 	rc = alloc_eq(sc, vi, &wrq->eq);
4208 	if (rc)
4209 		return (rc);
4210 
4211 	wrq->adapter = sc;
4212 	TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq);
4213 	TAILQ_INIT(&wrq->incomplete_wrs);
4214 	STAILQ_INIT(&wrq->wr_list);
4215 	wrq->nwr_pending = 0;
4216 	wrq->ndesc_needed = 0;
4217 
4218 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
4219 	    &wrq->eq.ba, "bus address of descriptor ring");
4220 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
4221 	    wrq->eq.sidx * EQ_ESIZE + sc->params.sge.spg_len,
4222 	    "desc ring size in bytes");
4223 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
4224 	    &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
4225 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
4226 	    CTLTYPE_INT | CTLFLAG_RD | CTLFLAG_MPSAFE, &wrq->eq.cidx, 0,
4227 	    sysctl_uint16, "I", "consumer index");
4228 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx",
4229 	    CTLTYPE_INT | CTLFLAG_RD | CTLFLAG_MPSAFE, &wrq->eq.pidx, 0,
4230 	    sysctl_uint16, "I", "producer index");
4231 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
4232 	    wrq->eq.sidx, "status page index");
4233 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD,
4234 	    &wrq->tx_wrs_direct, "# of work requests (direct)");
4235 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD,
4236 	    &wrq->tx_wrs_copied, "# of work requests (copied)");
4237 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_sspace", CTLFLAG_RD,
4238 	    &wrq->tx_wrs_ss, "# of work requests (copied from scratch space)");
4239 
4240 	return (rc);
4241 }
4242 
4243 static int
4244 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
4245 {
4246 	int rc;
4247 
4248 	rc = free_eq(sc, &wrq->eq);
4249 	if (rc)
4250 		return (rc);
4251 
4252 	bzero(wrq, sizeof(*wrq));
4253 	return (0);
4254 }
4255 
4256 static int
4257 alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx,
4258     struct sysctl_oid *oid)
4259 {
4260 	int rc;
4261 	struct port_info *pi = vi->pi;
4262 	struct adapter *sc = pi->adapter;
4263 	struct sge_eq *eq = &txq->eq;
4264 	struct txpkts *txp;
4265 	char name[16];
4266 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
4267 
4268 	rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx, can_resume_eth_tx,
4269 	    M_CXGBE, &eq->eq_lock, M_WAITOK);
4270 	if (rc != 0) {
4271 		device_printf(sc->dev, "failed to allocate mp_ring: %d\n", rc);
4272 		return (rc);
4273 	}
4274 
4275 	rc = alloc_eq(sc, vi, eq);
4276 	if (rc != 0) {
4277 		mp_ring_free(txq->r);
4278 		txq->r = NULL;
4279 		return (rc);
4280 	}
4281 
4282 	/* Can't fail after this point. */
4283 
4284 	if (idx == 0)
4285 		sc->sge.eq_base = eq->abs_id - eq->cntxt_id;
4286 	else
4287 		KASSERT(eq->cntxt_id + sc->sge.eq_base == eq->abs_id,
4288 		    ("eq_base mismatch"));
4289 	KASSERT(sc->sge.eq_base == 0 || sc->flags & IS_VF,
4290 	    ("PF with non-zero eq_base"));
4291 
4292 	TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq);
4293 	txq->ifp = vi->ifp;
4294 	txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK);
4295 	if (vi->flags & TX_USES_VM_WR)
4296 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
4297 		    V_TXPKT_INTF(pi->tx_chan));
4298 	else
4299 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
4300 		    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(sc->pf) |
4301 		    V_TXPKT_VF(vi->vin) | V_TXPKT_VF_VLD(vi->vfvld));
4302 	txq->tc_idx = -1;
4303 	txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE,
4304 	    M_ZERO | M_WAITOK);
4305 
4306 	txp = &txq->txp;
4307 	txp->score = 5;
4308 	MPASS(nitems(txp->mb) >= sc->params.max_pkts_per_eth_tx_pkts_wr);
4309 	txq->txp.max_npkt = min(nitems(txp->mb),
4310 	    sc->params.max_pkts_per_eth_tx_pkts_wr);
4311 	if (vi->flags & TX_USES_VM_WR && !(sc->flags & IS_VF))
4312 		txq->txp.max_npkt--;
4313 
4314 	snprintf(name, sizeof(name), "%d", idx);
4315 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name,
4316 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "tx queue");
4317 	children = SYSCTL_CHILDREN(oid);
4318 
4319 	SYSCTL_ADD_UAUTO(&vi->ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
4320 	    &eq->ba, "bus address of descriptor ring");
4321 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
4322 	    eq->sidx * EQ_ESIZE + sc->params.sge.spg_len,
4323 	    "desc ring size in bytes");
4324 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
4325 	    &eq->abs_id, 0, "absolute id of the queue");
4326 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
4327 	    &eq->cntxt_id, 0, "SGE context id of the queue");
4328 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
4329 	    CTLTYPE_INT | CTLFLAG_RD | CTLFLAG_MPSAFE, &eq->cidx, 0,
4330 	    sysctl_uint16, "I", "consumer index");
4331 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
4332 	    CTLTYPE_INT | CTLFLAG_RD | CTLFLAG_MPSAFE, &eq->pidx, 0,
4333 	    sysctl_uint16, "I", "producer index");
4334 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
4335 	    eq->sidx, "status page index");
4336 
4337 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "tc",
4338 	    CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, vi, idx, sysctl_tc,
4339 	    "I", "traffic class (-1 means none)");
4340 
4341 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
4342 	    &txq->txcsum, "# of times hardware assisted with checksum");
4343 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_insertion",
4344 	    CTLFLAG_RD, &txq->vlan_insertion,
4345 	    "# of times hardware inserted 802.1Q tag");
4346 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
4347 	    &txq->tso_wrs, "# of TSO work requests");
4348 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
4349 	    &txq->imm_wrs, "# of work requests with immediate data");
4350 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
4351 	    &txq->sgl_wrs, "# of work requests with direct SGL");
4352 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
4353 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
4354 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_wrs",
4355 	    CTLFLAG_RD, &txq->txpkts0_wrs,
4356 	    "# of txpkts (type 0) work requests");
4357 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_wrs",
4358 	    CTLFLAG_RD, &txq->txpkts1_wrs,
4359 	    "# of txpkts (type 1) work requests");
4360 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_pkts",
4361 	    CTLFLAG_RD, &txq->txpkts0_pkts,
4362 	    "# of frames tx'd using type0 txpkts work requests");
4363 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_pkts",
4364 	    CTLFLAG_RD, &txq->txpkts1_pkts,
4365 	    "# of frames tx'd using type1 txpkts work requests");
4366 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "raw_wrs", CTLFLAG_RD,
4367 	    &txq->raw_wrs, "# of raw work requests (non-packets)");
4368 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vxlan_tso_wrs",
4369 	    CTLFLAG_RD, &txq->vxlan_tso_wrs, "# of VXLAN TSO work requests");
4370 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vxlan_txcsum",
4371 	    CTLFLAG_RD, &txq->vxlan_txcsum,
4372 	    "# of times hardware assisted with inner checksums (VXLAN)");
4373 
4374 #ifdef KERN_TLS
4375 	if (sc->flags & KERN_TLS_OK) {
4376 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4377 		    "kern_tls_records", CTLFLAG_RD, &txq->kern_tls_records,
4378 		    "# of NIC TLS records transmitted");
4379 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4380 		    "kern_tls_short", CTLFLAG_RD, &txq->kern_tls_short,
4381 		    "# of short NIC TLS records transmitted");
4382 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4383 		    "kern_tls_partial", CTLFLAG_RD, &txq->kern_tls_partial,
4384 		    "# of partial NIC TLS records transmitted");
4385 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4386 		    "kern_tls_full", CTLFLAG_RD, &txq->kern_tls_full,
4387 		    "# of full NIC TLS records transmitted");
4388 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4389 		    "kern_tls_octets", CTLFLAG_RD, &txq->kern_tls_octets,
4390 		    "# of payload octets in transmitted NIC TLS records");
4391 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4392 		    "kern_tls_waste", CTLFLAG_RD, &txq->kern_tls_waste,
4393 		    "# of octets DMAd but not transmitted in NIC TLS records");
4394 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4395 		    "kern_tls_options", CTLFLAG_RD, &txq->kern_tls_options,
4396 		    "# of NIC TLS options-only packets transmitted");
4397 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4398 		    "kern_tls_header", CTLFLAG_RD, &txq->kern_tls_header,
4399 		    "# of NIC TLS header-only packets transmitted");
4400 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4401 		    "kern_tls_fin", CTLFLAG_RD, &txq->kern_tls_fin,
4402 		    "# of NIC TLS FIN-only packets transmitted");
4403 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4404 		    "kern_tls_fin_short", CTLFLAG_RD, &txq->kern_tls_fin_short,
4405 		    "# of NIC TLS padded FIN packets on short TLS records");
4406 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4407 		    "kern_tls_cbc", CTLFLAG_RD, &txq->kern_tls_cbc,
4408 		    "# of NIC TLS sessions using AES-CBC");
4409 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4410 		    "kern_tls_gcm", CTLFLAG_RD, &txq->kern_tls_gcm,
4411 		    "# of NIC TLS sessions using AES-GCM");
4412 	}
4413 #endif
4414 	mp_ring_sysctls(txq->r, &vi->ctx, children);
4415 
4416 	return (0);
4417 }
4418 
4419 static int
4420 free_txq(struct vi_info *vi, struct sge_txq *txq)
4421 {
4422 	int rc;
4423 	struct adapter *sc = vi->adapter;
4424 	struct sge_eq *eq = &txq->eq;
4425 
4426 	rc = free_eq(sc, eq);
4427 	if (rc)
4428 		return (rc);
4429 
4430 	sglist_free(txq->gl);
4431 	free(txq->sdesc, M_CXGBE);
4432 	mp_ring_free(txq->r);
4433 
4434 	bzero(txq, sizeof(*txq));
4435 	return (0);
4436 }
4437 
4438 static void
4439 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
4440 {
4441 	bus_addr_t *ba = arg;
4442 
4443 	KASSERT(nseg == 1,
4444 	    ("%s meant for single segment mappings only.", __func__));
4445 
4446 	*ba = error ? 0 : segs->ds_addr;
4447 }
4448 
4449 static inline void
4450 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
4451 {
4452 	uint32_t n, v;
4453 
4454 	n = IDXDIFF(fl->pidx >> 3, fl->dbidx, fl->sidx);
4455 	MPASS(n > 0);
4456 
4457 	wmb();
4458 	v = fl->dbval | V_PIDX(n);
4459 	if (fl->udb)
4460 		*fl->udb = htole32(v);
4461 	else
4462 		t4_write_reg(sc, sc->sge_kdoorbell_reg, v);
4463 	IDXINCR(fl->dbidx, n, fl->sidx);
4464 }
4465 
4466 /*
4467  * Fills up the freelist by allocating up to 'n' buffers.  Buffers that are
4468  * recycled do not count towards this allocation budget.
4469  *
4470  * Returns non-zero to indicate that this freelist should be added to the list
4471  * of starving freelists.
4472  */
4473 static int
4474 refill_fl(struct adapter *sc, struct sge_fl *fl, int n)
4475 {
4476 	__be64 *d;
4477 	struct fl_sdesc *sd;
4478 	uintptr_t pa;
4479 	caddr_t cl;
4480 	struct rx_buf_info *rxb;
4481 	struct cluster_metadata *clm;
4482 	uint16_t max_pidx;
4483 	uint16_t hw_cidx = fl->hw_cidx;		/* stable snapshot */
4484 
4485 	FL_LOCK_ASSERT_OWNED(fl);
4486 
4487 	/*
4488 	 * We always stop at the beginning of the hardware descriptor that's just
4489 	 * before the one with the hw cidx.  This is to avoid hw pidx = hw cidx,
4490 	 * which would mean an empty freelist to the chip.
4491 	 */
4492 	max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1;
4493 	if (fl->pidx == max_pidx * 8)
4494 		return (0);
4495 
4496 	d = &fl->desc[fl->pidx];
4497 	sd = &fl->sdesc[fl->pidx];
4498 
4499 	while (n > 0) {
4500 
4501 		if (sd->cl != NULL) {
4502 
4503 			if (sd->nmbuf == 0) {
4504 				/*
4505 				 * Fast recycle without involving any atomics on
4506 				 * the cluster's metadata (if the cluster has
4507 				 * metadata).  This happens when all frames
4508 				 * received in the cluster were small enough to
4509 				 * fit within a single mbuf each.
4510 				 */
4511 				fl->cl_fast_recycled++;
4512 				goto recycled;
4513 			}
4514 
4515 			/*
4516 			 * Cluster is guaranteed to have metadata.  Clusters
4517 			 * without metadata always take the fast recycle path
4518 			 * when they're recycled.
4519 			 */
4520 			clm = cl_metadata(sd);
4521 			MPASS(clm != NULL);
4522 
4523 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4524 				fl->cl_recycled++;
4525 				counter_u64_add(extfree_rels, 1);
4526 				goto recycled;
4527 			}
4528 			sd->cl = NULL;	/* gave up my reference */
4529 		}
4530 		MPASS(sd->cl == NULL);
4531 		rxb = &sc->sge.rx_buf_info[fl->zidx];
4532 		cl = uma_zalloc(rxb->zone, M_NOWAIT);
4533 		if (__predict_false(cl == NULL)) {
4534 			if (fl->zidx != fl->safe_zidx) {
4535 				rxb = &sc->sge.rx_buf_info[fl->safe_zidx];
4536 				cl = uma_zalloc(rxb->zone, M_NOWAIT);
4537 			}
4538 			if (cl == NULL)
4539 				break;
4540 		}
4541 		fl->cl_allocated++;
4542 		n--;
4543 
4544 		pa = pmap_kextract((vm_offset_t)cl);
4545 		sd->cl = cl;
4546 		sd->zidx = fl->zidx;
4547 
4548 		if (fl->flags & FL_BUF_PACKING) {
4549 			*d = htobe64(pa | rxb->hwidx2);
4550 			sd->moff = rxb->size2;
4551 		} else {
4552 			*d = htobe64(pa | rxb->hwidx1);
4553 			sd->moff = 0;
4554 		}
4555 recycled:
4556 		sd->nmbuf = 0;
4557 		d++;
4558 		sd++;
4559 		if (__predict_false((++fl->pidx & 7) == 0)) {
4560 			uint16_t pidx = fl->pidx >> 3;
4561 
4562 			if (__predict_false(pidx == fl->sidx)) {
4563 				fl->pidx = 0;
4564 				pidx = 0;
4565 				sd = fl->sdesc;
4566 				d = fl->desc;
4567 			}
4568 			if (n < 8 || pidx == max_pidx)
4569 				break;
4570 
4571 			if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4)
4572 				ring_fl_db(sc, fl);
4573 		}
4574 	}
4575 
4576 	if ((fl->pidx >> 3) != fl->dbidx)
4577 		ring_fl_db(sc, fl);
4578 
4579 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
4580 }
4581 
4582 /*
4583  * Attempt to refill all starving freelists.
4584  */
4585 static void
4586 refill_sfl(void *arg)
4587 {
4588 	struct adapter *sc = arg;
4589 	struct sge_fl *fl, *fl_temp;
4590 
4591 	mtx_assert(&sc->sfl_lock, MA_OWNED);
4592 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
4593 		FL_LOCK(fl);
4594 		refill_fl(sc, fl, 64);
4595 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
4596 			TAILQ_REMOVE(&sc->sfl, fl, link);
4597 			fl->flags &= ~FL_STARVING;
4598 		}
4599 		FL_UNLOCK(fl);
4600 	}
4601 
4602 	if (!TAILQ_EMPTY(&sc->sfl))
4603 		callout_schedule(&sc->sfl_callout, hz / 5);
4604 }
4605 
4606 static int
4607 alloc_fl_sdesc(struct sge_fl *fl)
4608 {
4609 
4610 	fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc), M_CXGBE,
4611 	    M_ZERO | M_WAITOK);
4612 
4613 	return (0);
4614 }
4615 
4616 static void
4617 free_fl_sdesc(struct adapter *sc, struct sge_fl *fl)
4618 {
4619 	struct fl_sdesc *sd;
4620 	struct cluster_metadata *clm;
4621 	int i;
4622 
4623 	sd = fl->sdesc;
4624 	for (i = 0; i < fl->sidx * 8; i++, sd++) {
4625 		if (sd->cl == NULL)
4626 			continue;
4627 
4628 		if (sd->nmbuf == 0)
4629 			uma_zfree(sc->sge.rx_buf_info[sd->zidx].zone, sd->cl);
4630 		else if (fl->flags & FL_BUF_PACKING) {
4631 			clm = cl_metadata(sd);
4632 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4633 				uma_zfree(sc->sge.rx_buf_info[sd->zidx].zone,
4634 				    sd->cl);
4635 				counter_u64_add(extfree_rels, 1);
4636 			}
4637 		}
4638 		sd->cl = NULL;
4639 	}
4640 
4641 	free(fl->sdesc, M_CXGBE);
4642 	fl->sdesc = NULL;
4643 }
4644 
4645 static inline void
4646 get_pkt_gl(struct mbuf *m, struct sglist *gl)
4647 {
4648 	int rc;
4649 
4650 	M_ASSERTPKTHDR(m);
4651 
4652 	sglist_reset(gl);
4653 	rc = sglist_append_mbuf(gl, m);
4654 	if (__predict_false(rc != 0)) {
4655 		panic("%s: mbuf %p (%d segs) was vetted earlier but now fails "
4656 		    "with %d.", __func__, m, mbuf_nsegs(m), rc);
4657 	}
4658 
4659 	KASSERT(gl->sg_nseg == mbuf_nsegs(m),
4660 	    ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m,
4661 	    mbuf_nsegs(m), gl->sg_nseg));
4662 #if 0	/* vm_wr not readily available here. */
4663 	KASSERT(gl->sg_nseg > 0 && gl->sg_nseg <= max_nsegs_allowed(m, vm_wr),
4664 	    ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__,
4665 		gl->sg_nseg, max_nsegs_allowed(m, vm_wr)));
4666 #endif
4667 }
4668 
4669 /*
4670  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
4671  */
4672 static inline u_int
4673 txpkt_len16(u_int nsegs, const u_int extra)
4674 {
4675 	u_int n;
4676 
4677 	MPASS(nsegs > 0);
4678 
4679 	nsegs--; /* first segment is part of ulptx_sgl */
4680 	n = extra + sizeof(struct fw_eth_tx_pkt_wr) +
4681 	    sizeof(struct cpl_tx_pkt_core) +
4682 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4683 
4684 	return (howmany(n, 16));
4685 }
4686 
4687 /*
4688  * len16 for a txpkt_vm WR with a GL.  Includes the firmware work
4689  * request header.
4690  */
4691 static inline u_int
4692 txpkt_vm_len16(u_int nsegs, const u_int extra)
4693 {
4694 	u_int n;
4695 
4696 	MPASS(nsegs > 0);
4697 
4698 	nsegs--; /* first segment is part of ulptx_sgl */
4699 	n = extra + sizeof(struct fw_eth_tx_pkt_vm_wr) +
4700 	    sizeof(struct cpl_tx_pkt_core) +
4701 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4702 
4703 	return (howmany(n, 16));
4704 }
4705 
4706 static inline void
4707 calculate_mbuf_len16(struct mbuf *m, bool vm_wr)
4708 {
4709 	const int lso = sizeof(struct cpl_tx_pkt_lso_core);
4710 	const int tnl_lso = sizeof(struct cpl_tx_tnl_lso);
4711 
4712 	if (vm_wr) {
4713 		if (needs_tso(m))
4714 			set_mbuf_len16(m, txpkt_vm_len16(mbuf_nsegs(m), lso));
4715 		else
4716 			set_mbuf_len16(m, txpkt_vm_len16(mbuf_nsegs(m), 0));
4717 		return;
4718 	}
4719 
4720 	if (needs_tso(m)) {
4721 		if (needs_vxlan_tso(m))
4722 			set_mbuf_len16(m, txpkt_len16(mbuf_nsegs(m), tnl_lso));
4723 		else
4724 			set_mbuf_len16(m, txpkt_len16(mbuf_nsegs(m), lso));
4725 	} else
4726 		set_mbuf_len16(m, txpkt_len16(mbuf_nsegs(m), 0));
4727 }
4728 
4729 /*
4730  * len16 for a txpkts type 0 WR with a GL.  Does not include the firmware work
4731  * request header.
4732  */
4733 static inline u_int
4734 txpkts0_len16(u_int nsegs)
4735 {
4736 	u_int n;
4737 
4738 	MPASS(nsegs > 0);
4739 
4740 	nsegs--; /* first segment is part of ulptx_sgl */
4741 	n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) +
4742 	    sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) +
4743 	    8 * ((3 * nsegs) / 2 + (nsegs & 1));
4744 
4745 	return (howmany(n, 16));
4746 }
4747 
4748 /*
4749  * len16 for a txpkts type 1 WR with a GL.  Does not include the firmware work
4750  * request header.
4751  */
4752 static inline u_int
4753 txpkts1_len16(void)
4754 {
4755 	u_int n;
4756 
4757 	n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl);
4758 
4759 	return (howmany(n, 16));
4760 }
4761 
4762 static inline u_int
4763 imm_payload(u_int ndesc)
4764 {
4765 	u_int n;
4766 
4767 	n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) -
4768 	    sizeof(struct cpl_tx_pkt_core);
4769 
4770 	return (n);
4771 }
4772 
4773 static inline uint64_t
4774 csum_to_ctrl(struct adapter *sc, struct mbuf *m)
4775 {
4776 	uint64_t ctrl;
4777 	int csum_type, l2hlen, l3hlen;
4778 	int x, y;
4779 	static const int csum_types[3][2] = {
4780 		{TX_CSUM_TCPIP, TX_CSUM_TCPIP6},
4781 		{TX_CSUM_UDPIP, TX_CSUM_UDPIP6},
4782 		{TX_CSUM_IP, 0}
4783 	};
4784 
4785 	M_ASSERTPKTHDR(m);
4786 
4787 	if (!needs_hwcsum(m))
4788 		return (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS);
4789 
4790 	MPASS(m->m_pkthdr.l2hlen >= ETHER_HDR_LEN);
4791 	MPASS(m->m_pkthdr.l3hlen >= sizeof(struct ip));
4792 
4793 	if (needs_vxlan_csum(m)) {
4794 		MPASS(m->m_pkthdr.l4hlen > 0);
4795 		MPASS(m->m_pkthdr.l5hlen > 0);
4796 		MPASS(m->m_pkthdr.inner_l2hlen >= ETHER_HDR_LEN);
4797 		MPASS(m->m_pkthdr.inner_l3hlen >= sizeof(struct ip));
4798 
4799 		l2hlen = m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen +
4800 		    m->m_pkthdr.l4hlen + m->m_pkthdr.l5hlen +
4801 		    m->m_pkthdr.inner_l2hlen - ETHER_HDR_LEN;
4802 		l3hlen = m->m_pkthdr.inner_l3hlen;
4803 	} else {
4804 		l2hlen = m->m_pkthdr.l2hlen - ETHER_HDR_LEN;
4805 		l3hlen = m->m_pkthdr.l3hlen;
4806 	}
4807 
4808 	ctrl = 0;
4809 	if (!needs_l3_csum(m))
4810 		ctrl |= F_TXPKT_IPCSUM_DIS;
4811 
4812 	if (m->m_pkthdr.csum_flags & (CSUM_IP_TCP | CSUM_INNER_IP_TCP |
4813 	    CSUM_IP6_TCP | CSUM_INNER_IP6_TCP))
4814 		x = 0;	/* TCP */
4815 	else if (m->m_pkthdr.csum_flags & (CSUM_IP_UDP | CSUM_INNER_IP_UDP |
4816 	    CSUM_IP6_UDP | CSUM_INNER_IP6_UDP))
4817 		x = 1;	/* UDP */
4818 	else
4819 		x = 2;
4820 
4821 	if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_IP_TCP | CSUM_IP_UDP |
4822 	    CSUM_INNER_IP | CSUM_INNER_IP_TCP | CSUM_INNER_IP_UDP))
4823 		y = 0;	/* IPv4 */
4824 	else {
4825 		MPASS(m->m_pkthdr.csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP |
4826 		    CSUM_INNER_IP6_TCP | CSUM_INNER_IP6_UDP));
4827 		y = 1;	/* IPv6 */
4828 	}
4829 	/*
4830 	 * needs_hwcsum returned true earlier so there must be some kind of
4831 	 * checksum to calculate.
4832 	 */
4833 	csum_type = csum_types[x][y];
4834 	MPASS(csum_type != 0);
4835 	if (csum_type == TX_CSUM_IP)
4836 		ctrl |= F_TXPKT_L4CSUM_DIS;
4837 	ctrl |= V_TXPKT_CSUM_TYPE(csum_type) | V_TXPKT_IPHDR_LEN(l3hlen);
4838 	if (chip_id(sc) <= CHELSIO_T5)
4839 		ctrl |= V_TXPKT_ETHHDR_LEN(l2hlen);
4840 	else
4841 		ctrl |= V_T6_TXPKT_ETHHDR_LEN(l2hlen);
4842 
4843 	return (ctrl);
4844 }
4845 
4846 static inline void *
4847 write_lso_cpl(void *cpl, struct mbuf *m0)
4848 {
4849 	struct cpl_tx_pkt_lso_core *lso;
4850 	uint32_t ctrl;
4851 
4852 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4853 	    m0->m_pkthdr.l4hlen > 0,
4854 	    ("%s: mbuf %p needs TSO but missing header lengths",
4855 		__func__, m0));
4856 
4857 	ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) |
4858 	    F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE |
4859 	    V_LSO_ETHHDR_LEN((m0->m_pkthdr.l2hlen - ETHER_HDR_LEN) >> 2) |
4860 	    V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) |
4861 	    V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4862 	if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4863 		ctrl |= F_LSO_IPV6;
4864 
4865 	lso = cpl;
4866 	lso->lso_ctrl = htobe32(ctrl);
4867 	lso->ipid_ofst = htobe16(0);
4868 	lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4869 	lso->seqno_offset = htobe32(0);
4870 	lso->len = htobe32(m0->m_pkthdr.len);
4871 
4872 	return (lso + 1);
4873 }
4874 
4875 static void *
4876 write_tnl_lso_cpl(void *cpl, struct mbuf *m0)
4877 {
4878 	struct cpl_tx_tnl_lso *tnl_lso = cpl;
4879 	uint32_t ctrl;
4880 
4881 	KASSERT(m0->m_pkthdr.inner_l2hlen > 0 &&
4882 	    m0->m_pkthdr.inner_l3hlen > 0 && m0->m_pkthdr.inner_l4hlen > 0 &&
4883 	    m0->m_pkthdr.inner_l5hlen > 0,
4884 	    ("%s: mbuf %p needs VXLAN_TSO but missing inner header lengths",
4885 		__func__, m0));
4886 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4887 	    m0->m_pkthdr.l4hlen > 0 && m0->m_pkthdr.l5hlen > 0,
4888 	    ("%s: mbuf %p needs VXLAN_TSO but missing outer header lengths",
4889 		__func__, m0));
4890 
4891 	/* Outer headers. */
4892 	ctrl = V_CPL_TX_TNL_LSO_OPCODE(CPL_TX_TNL_LSO) |
4893 	    F_CPL_TX_TNL_LSO_FIRST | F_CPL_TX_TNL_LSO_LAST |
4894 	    V_CPL_TX_TNL_LSO_ETHHDRLENOUT(
4895 		(m0->m_pkthdr.l2hlen - ETHER_HDR_LEN) >> 2) |
4896 	    V_CPL_TX_TNL_LSO_IPHDRLENOUT(m0->m_pkthdr.l3hlen >> 2) |
4897 	    F_CPL_TX_TNL_LSO_IPLENSETOUT;
4898 	if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4899 		ctrl |= F_CPL_TX_TNL_LSO_IPV6OUT;
4900 	else {
4901 		ctrl |= F_CPL_TX_TNL_LSO_IPHDRCHKOUT |
4902 		    F_CPL_TX_TNL_LSO_IPIDINCOUT;
4903 	}
4904 	tnl_lso->op_to_IpIdSplitOut = htobe32(ctrl);
4905 	tnl_lso->IpIdOffsetOut = 0;
4906 	tnl_lso->UdpLenSetOut_to_TnlHdrLen =
4907 		htobe16(F_CPL_TX_TNL_LSO_UDPCHKCLROUT |
4908 		    F_CPL_TX_TNL_LSO_UDPLENSETOUT |
4909 		    V_CPL_TX_TNL_LSO_TNLHDRLEN(m0->m_pkthdr.l2hlen +
4910 			m0->m_pkthdr.l3hlen + m0->m_pkthdr.l4hlen +
4911 			m0->m_pkthdr.l5hlen) |
4912 		    V_CPL_TX_TNL_LSO_TNLTYPE(TX_TNL_TYPE_VXLAN));
4913 	tnl_lso->r1 = 0;
4914 
4915 	/* Inner headers. */
4916 	ctrl = V_CPL_TX_TNL_LSO_ETHHDRLEN(
4917 	    (m0->m_pkthdr.inner_l2hlen - ETHER_HDR_LEN) >> 2) |
4918 	    V_CPL_TX_TNL_LSO_IPHDRLEN(m0->m_pkthdr.inner_l3hlen >> 2) |
4919 	    V_CPL_TX_TNL_LSO_TCPHDRLEN(m0->m_pkthdr.inner_l4hlen >> 2);
4920 	if (m0->m_pkthdr.inner_l3hlen == sizeof(struct ip6_hdr))
4921 		ctrl |= F_CPL_TX_TNL_LSO_IPV6;
4922 	tnl_lso->Flow_to_TcpHdrLen = htobe32(ctrl);
4923 	tnl_lso->IpIdOffset = 0;
4924 	tnl_lso->IpIdSplit_to_Mss =
4925 	    htobe16(V_CPL_TX_TNL_LSO_MSS(m0->m_pkthdr.tso_segsz));
4926 	tnl_lso->TCPSeqOffset = 0;
4927 	tnl_lso->EthLenOffset_Size =
4928 	    htobe32(V_CPL_TX_TNL_LSO_SIZE(m0->m_pkthdr.len));
4929 
4930 	return (tnl_lso + 1);
4931 }
4932 
4933 #define VM_TX_L2HDR_LEN	16	/* ethmacdst to vlantci */
4934 
4935 /*
4936  * Write a VM txpkt WR for this packet to the hardware descriptors, update the
4937  * software descriptor, and advance the pidx.  It is guaranteed that enough
4938  * descriptors are available.
4939  *
4940  * The return value is the # of hardware descriptors used.
4941  */
4942 static u_int
4943 write_txpkt_vm_wr(struct adapter *sc, struct sge_txq *txq, struct mbuf *m0)
4944 {
4945 	struct sge_eq *eq;
4946 	struct fw_eth_tx_pkt_vm_wr *wr;
4947 	struct tx_sdesc *txsd;
4948 	struct cpl_tx_pkt_core *cpl;
4949 	uint32_t ctrl;	/* used in many unrelated places */
4950 	uint64_t ctrl1;
4951 	int len16, ndesc, pktlen, nsegs;
4952 	caddr_t dst;
4953 
4954 	TXQ_LOCK_ASSERT_OWNED(txq);
4955 	M_ASSERTPKTHDR(m0);
4956 
4957 	len16 = mbuf_len16(m0);
4958 	nsegs = mbuf_nsegs(m0);
4959 	pktlen = m0->m_pkthdr.len;
4960 	ctrl = sizeof(struct cpl_tx_pkt_core);
4961 	if (needs_tso(m0))
4962 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4963 	ndesc = tx_len16_to_desc(len16);
4964 
4965 	/* Firmware work request header */
4966 	eq = &txq->eq;
4967 	wr = (void *)&eq->desc[eq->pidx];
4968 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_VM_WR) |
4969 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4970 
4971 	ctrl = V_FW_WR_LEN16(len16);
4972 	wr->equiq_to_len16 = htobe32(ctrl);
4973 	wr->r3[0] = 0;
4974 	wr->r3[1] = 0;
4975 
4976 	/*
4977 	 * Copy over ethmacdst, ethmacsrc, ethtype, and vlantci.
4978 	 * vlantci is ignored unless the ethtype is 0x8100, so it's
4979 	 * simpler to always copy it rather than making it
4980 	 * conditional.  Also, it seems that we do not have to set
4981 	 * vlantci or fake the ethtype when doing VLAN tag insertion.
4982 	 */
4983 	m_copydata(m0, 0, VM_TX_L2HDR_LEN, wr->ethmacdst);
4984 
4985 	if (needs_tso(m0)) {
4986 		cpl = write_lso_cpl(wr + 1, m0);
4987 		txq->tso_wrs++;
4988 	} else
4989 		cpl = (void *)(wr + 1);
4990 
4991 	/* Checksum offload */
4992 	ctrl1 = csum_to_ctrl(sc, m0);
4993 	if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS))
4994 		txq->txcsum++;	/* some hardware assistance provided */
4995 
4996 	/* VLAN tag insertion */
4997 	if (needs_vlan_insertion(m0)) {
4998 		ctrl1 |= F_TXPKT_VLAN_VLD |
4999 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
5000 		txq->vlan_insertion++;
5001 	}
5002 
5003 	/* CPL header */
5004 	cpl->ctrl0 = txq->cpl_ctrl0;
5005 	cpl->pack = 0;
5006 	cpl->len = htobe16(pktlen);
5007 	cpl->ctrl1 = htobe64(ctrl1);
5008 
5009 	/* SGL */
5010 	dst = (void *)(cpl + 1);
5011 
5012 	/*
5013 	 * A packet using TSO will use up an entire descriptor for the
5014 	 * firmware work request header, LSO CPL, and TX_PKT_XT CPL.
5015 	 * If this descriptor is the last descriptor in the ring, wrap
5016 	 * around to the front of the ring explicitly for the start of
5017 	 * the sgl.
5018 	 */
5019 	if (dst == (void *)&eq->desc[eq->sidx]) {
5020 		dst = (void *)&eq->desc[0];
5021 		write_gl_to_txd(txq, m0, &dst, 0);
5022 	} else
5023 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
5024 	txq->sgl_wrs++;
5025 	txq->txpkt_wrs++;
5026 
5027 	txsd = &txq->sdesc[eq->pidx];
5028 	txsd->m = m0;
5029 	txsd->desc_used = ndesc;
5030 
5031 	return (ndesc);
5032 }
5033 
5034 /*
5035  * Write a raw WR to the hardware descriptors, update the software
5036  * descriptor, and advance the pidx.  It is guaranteed that enough
5037  * descriptors are available.
5038  *
5039  * The return value is the # of hardware descriptors used.
5040  */
5041 static u_int
5042 write_raw_wr(struct sge_txq *txq, void *wr, struct mbuf *m0, u_int available)
5043 {
5044 	struct sge_eq *eq = &txq->eq;
5045 	struct tx_sdesc *txsd;
5046 	struct mbuf *m;
5047 	caddr_t dst;
5048 	int len16, ndesc;
5049 
5050 	len16 = mbuf_len16(m0);
5051 	ndesc = tx_len16_to_desc(len16);
5052 	MPASS(ndesc <= available);
5053 
5054 	dst = wr;
5055 	for (m = m0; m != NULL; m = m->m_next)
5056 		copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
5057 
5058 	txq->raw_wrs++;
5059 
5060 	txsd = &txq->sdesc[eq->pidx];
5061 	txsd->m = m0;
5062 	txsd->desc_used = ndesc;
5063 
5064 	return (ndesc);
5065 }
5066 
5067 /*
5068  * Write a txpkt WR for this packet to the hardware descriptors, update the
5069  * software descriptor, and advance the pidx.  It is guaranteed that enough
5070  * descriptors are available.
5071  *
5072  * The return value is the # of hardware descriptors used.
5073  */
5074 static u_int
5075 write_txpkt_wr(struct adapter *sc, struct sge_txq *txq, struct mbuf *m0,
5076     u_int available)
5077 {
5078 	struct sge_eq *eq;
5079 	struct fw_eth_tx_pkt_wr *wr;
5080 	struct tx_sdesc *txsd;
5081 	struct cpl_tx_pkt_core *cpl;
5082 	uint32_t ctrl;	/* used in many unrelated places */
5083 	uint64_t ctrl1;
5084 	int len16, ndesc, pktlen, nsegs;
5085 	caddr_t dst;
5086 
5087 	TXQ_LOCK_ASSERT_OWNED(txq);
5088 	M_ASSERTPKTHDR(m0);
5089 
5090 	len16 = mbuf_len16(m0);
5091 	nsegs = mbuf_nsegs(m0);
5092 	pktlen = m0->m_pkthdr.len;
5093 	ctrl = sizeof(struct cpl_tx_pkt_core);
5094 	if (needs_tso(m0)) {
5095 		if (needs_vxlan_tso(m0))
5096 			ctrl += sizeof(struct cpl_tx_tnl_lso);
5097 		else
5098 			ctrl += sizeof(struct cpl_tx_pkt_lso_core);
5099 	} else if (!(mbuf_cflags(m0) & MC_NOMAP) && pktlen <= imm_payload(2) &&
5100 	    available >= 2) {
5101 		/* Immediate data.  Recalculate len16 and set nsegs to 0. */
5102 		ctrl += pktlen;
5103 		len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) +
5104 		    sizeof(struct cpl_tx_pkt_core) + pktlen, 16);
5105 		nsegs = 0;
5106 	}
5107 	ndesc = tx_len16_to_desc(len16);
5108 	MPASS(ndesc <= available);
5109 
5110 	/* Firmware work request header */
5111 	eq = &txq->eq;
5112 	wr = (void *)&eq->desc[eq->pidx];
5113 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
5114 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
5115 
5116 	ctrl = V_FW_WR_LEN16(len16);
5117 	wr->equiq_to_len16 = htobe32(ctrl);
5118 	wr->r3 = 0;
5119 
5120 	if (needs_tso(m0)) {
5121 		if (needs_vxlan_tso(m0)) {
5122 			cpl = write_tnl_lso_cpl(wr + 1, m0);
5123 			txq->vxlan_tso_wrs++;
5124 		} else {
5125 			cpl = write_lso_cpl(wr + 1, m0);
5126 			txq->tso_wrs++;
5127 		}
5128 	} else
5129 		cpl = (void *)(wr + 1);
5130 
5131 	/* Checksum offload */
5132 	ctrl1 = csum_to_ctrl(sc, m0);
5133 	if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS)) {
5134 		/* some hardware assistance provided */
5135 		if (needs_vxlan_csum(m0))
5136 			txq->vxlan_txcsum++;
5137 		else
5138 			txq->txcsum++;
5139 	}
5140 
5141 	/* VLAN tag insertion */
5142 	if (needs_vlan_insertion(m0)) {
5143 		ctrl1 |= F_TXPKT_VLAN_VLD |
5144 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
5145 		txq->vlan_insertion++;
5146 	}
5147 
5148 	/* CPL header */
5149 	cpl->ctrl0 = txq->cpl_ctrl0;
5150 	cpl->pack = 0;
5151 	cpl->len = htobe16(pktlen);
5152 	cpl->ctrl1 = htobe64(ctrl1);
5153 
5154 	/* SGL */
5155 	dst = (void *)(cpl + 1);
5156 	if (__predict_false((uintptr_t)dst == (uintptr_t)&eq->desc[eq->sidx]))
5157 		dst = (caddr_t)&eq->desc[0];
5158 	if (nsegs > 0) {
5159 
5160 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
5161 		txq->sgl_wrs++;
5162 	} else {
5163 		struct mbuf *m;
5164 
5165 		for (m = m0; m != NULL; m = m->m_next) {
5166 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
5167 #ifdef INVARIANTS
5168 			pktlen -= m->m_len;
5169 #endif
5170 		}
5171 #ifdef INVARIANTS
5172 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
5173 #endif
5174 		txq->imm_wrs++;
5175 	}
5176 
5177 	txq->txpkt_wrs++;
5178 
5179 	txsd = &txq->sdesc[eq->pidx];
5180 	txsd->m = m0;
5181 	txsd->desc_used = ndesc;
5182 
5183 	return (ndesc);
5184 }
5185 
5186 static inline bool
5187 cmp_l2hdr(struct txpkts *txp, struct mbuf *m)
5188 {
5189 	int len;
5190 
5191 	MPASS(txp->npkt > 0);
5192 	MPASS(m->m_len >= VM_TX_L2HDR_LEN);
5193 
5194 	if (txp->ethtype == be16toh(ETHERTYPE_VLAN))
5195 		len = VM_TX_L2HDR_LEN;
5196 	else
5197 		len = sizeof(struct ether_header);
5198 
5199 	return (memcmp(m->m_data, &txp->ethmacdst[0], len) != 0);
5200 }
5201 
5202 static inline void
5203 save_l2hdr(struct txpkts *txp, struct mbuf *m)
5204 {
5205 	MPASS(m->m_len >= VM_TX_L2HDR_LEN);
5206 
5207 	memcpy(&txp->ethmacdst[0], mtod(m, const void *), VM_TX_L2HDR_LEN);
5208 }
5209 
5210 static int
5211 add_to_txpkts_vf(struct adapter *sc, struct sge_txq *txq, struct mbuf *m,
5212     int avail, bool *send)
5213 {
5214 	struct txpkts *txp = &txq->txp;
5215 
5216 	/* Cannot have TSO and coalesce at the same time. */
5217 	if (cannot_use_txpkts(m)) {
5218 cannot_coalesce:
5219 		*send = txp->npkt > 0;
5220 		return (EINVAL);
5221 	}
5222 
5223 	/* VF allows coalescing of type 1 (1 GL) only */
5224 	if (mbuf_nsegs(m) > 1)
5225 		goto cannot_coalesce;
5226 
5227 	*send = false;
5228 	if (txp->npkt > 0) {
5229 		MPASS(tx_len16_to_desc(txp->len16) <= avail);
5230 		MPASS(txp->npkt < txp->max_npkt);
5231 		MPASS(txp->wr_type == 1);	/* VF supports type 1 only */
5232 
5233 		if (tx_len16_to_desc(txp->len16 + txpkts1_len16()) > avail) {
5234 retry_after_send:
5235 			*send = true;
5236 			return (EAGAIN);
5237 		}
5238 		if (m->m_pkthdr.len + txp->plen > 65535)
5239 			goto retry_after_send;
5240 		if (cmp_l2hdr(txp, m))
5241 			goto retry_after_send;
5242 
5243 		txp->len16 += txpkts1_len16();
5244 		txp->plen += m->m_pkthdr.len;
5245 		txp->mb[txp->npkt++] = m;
5246 		if (txp->npkt == txp->max_npkt)
5247 			*send = true;
5248 	} else {
5249 		txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_vm_wr), 16) +
5250 		    txpkts1_len16();
5251 		if (tx_len16_to_desc(txp->len16) > avail)
5252 			goto cannot_coalesce;
5253 		txp->npkt = 1;
5254 		txp->wr_type = 1;
5255 		txp->plen = m->m_pkthdr.len;
5256 		txp->mb[0] = m;
5257 		save_l2hdr(txp, m);
5258 	}
5259 	return (0);
5260 }
5261 
5262 static int
5263 add_to_txpkts_pf(struct adapter *sc, struct sge_txq *txq, struct mbuf *m,
5264     int avail, bool *send)
5265 {
5266 	struct txpkts *txp = &txq->txp;
5267 	int nsegs;
5268 
5269 	MPASS(!(sc->flags & IS_VF));
5270 
5271 	/* Cannot have TSO and coalesce at the same time. */
5272 	if (cannot_use_txpkts(m)) {
5273 cannot_coalesce:
5274 		*send = txp->npkt > 0;
5275 		return (EINVAL);
5276 	}
5277 
5278 	*send = false;
5279 	nsegs = mbuf_nsegs(m);
5280 	if (txp->npkt == 0) {
5281 		if (m->m_pkthdr.len > 65535)
5282 			goto cannot_coalesce;
5283 		if (nsegs > 1) {
5284 			txp->wr_type = 0;
5285 			txp->len16 =
5286 			    howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) +
5287 			    txpkts0_len16(nsegs);
5288 		} else {
5289 			txp->wr_type = 1;
5290 			txp->len16 =
5291 			    howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) +
5292 			    txpkts1_len16();
5293 		}
5294 		if (tx_len16_to_desc(txp->len16) > avail)
5295 			goto cannot_coalesce;
5296 		txp->npkt = 1;
5297 		txp->plen = m->m_pkthdr.len;
5298 		txp->mb[0] = m;
5299 	} else {
5300 		MPASS(tx_len16_to_desc(txp->len16) <= avail);
5301 		MPASS(txp->npkt < txp->max_npkt);
5302 
5303 		if (m->m_pkthdr.len + txp->plen > 65535) {
5304 retry_after_send:
5305 			*send = true;
5306 			return (EAGAIN);
5307 		}
5308 
5309 		MPASS(txp->wr_type == 0 || txp->wr_type == 1);
5310 		if (txp->wr_type == 0) {
5311 			if (tx_len16_to_desc(txp->len16 +
5312 			    txpkts0_len16(nsegs)) > min(avail, SGE_MAX_WR_NDESC))
5313 				goto retry_after_send;
5314 			txp->len16 += txpkts0_len16(nsegs);
5315 		} else {
5316 			if (nsegs != 1)
5317 				goto retry_after_send;
5318 			if (tx_len16_to_desc(txp->len16 + txpkts1_len16()) >
5319 			    avail)
5320 				goto retry_after_send;
5321 			txp->len16 += txpkts1_len16();
5322 		}
5323 
5324 		txp->plen += m->m_pkthdr.len;
5325 		txp->mb[txp->npkt++] = m;
5326 		if (txp->npkt == txp->max_npkt)
5327 			*send = true;
5328 	}
5329 	return (0);
5330 }
5331 
5332 /*
5333  * Write a txpkts WR for the packets in txp to the hardware descriptors, update
5334  * the software descriptor, and advance the pidx.  It is guaranteed that enough
5335  * descriptors are available.
5336  *
5337  * The return value is the # of hardware descriptors used.
5338  */
5339 static u_int
5340 write_txpkts_wr(struct adapter *sc, struct sge_txq *txq)
5341 {
5342 	const struct txpkts *txp = &txq->txp;
5343 	struct sge_eq *eq = &txq->eq;
5344 	struct fw_eth_tx_pkts_wr *wr;
5345 	struct tx_sdesc *txsd;
5346 	struct cpl_tx_pkt_core *cpl;
5347 	uint64_t ctrl1;
5348 	int ndesc, i, checkwrap;
5349 	struct mbuf *m, *last;
5350 	void *flitp;
5351 
5352 	TXQ_LOCK_ASSERT_OWNED(txq);
5353 	MPASS(txp->npkt > 0);
5354 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
5355 
5356 	wr = (void *)&eq->desc[eq->pidx];
5357 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
5358 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(txp->len16));
5359 	wr->plen = htobe16(txp->plen);
5360 	wr->npkt = txp->npkt;
5361 	wr->r3 = 0;
5362 	wr->type = txp->wr_type;
5363 	flitp = wr + 1;
5364 
5365 	/*
5366 	 * At this point we are 16B into a hardware descriptor.  If checkwrap is
5367 	 * set then we know the WR is going to wrap around somewhere.  We'll
5368 	 * check for that at appropriate points.
5369 	 */
5370 	ndesc = tx_len16_to_desc(txp->len16);
5371 	last = NULL;
5372 	checkwrap = eq->sidx - ndesc < eq->pidx;
5373 	for (i = 0; i < txp->npkt; i++) {
5374 		m = txp->mb[i];
5375 		if (txp->wr_type == 0) {
5376 			struct ulp_txpkt *ulpmc;
5377 			struct ulptx_idata *ulpsc;
5378 
5379 			/* ULP master command */
5380 			ulpmc = flitp;
5381 			ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
5382 			    V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid));
5383 			ulpmc->len = htobe32(txpkts0_len16(mbuf_nsegs(m)));
5384 
5385 			/* ULP subcommand */
5386 			ulpsc = (void *)(ulpmc + 1);
5387 			ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
5388 			    F_ULP_TX_SC_MORE);
5389 			ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
5390 
5391 			cpl = (void *)(ulpsc + 1);
5392 			if (checkwrap &&
5393 			    (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx])
5394 				cpl = (void *)&eq->desc[0];
5395 		} else {
5396 			cpl = flitp;
5397 		}
5398 
5399 		/* Checksum offload */
5400 		ctrl1 = csum_to_ctrl(sc, m);
5401 		if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS)) {
5402 			/* some hardware assistance provided */
5403 			if (needs_vxlan_csum(m))
5404 				txq->vxlan_txcsum++;
5405 			else
5406 				txq->txcsum++;
5407 		}
5408 
5409 		/* VLAN tag insertion */
5410 		if (needs_vlan_insertion(m)) {
5411 			ctrl1 |= F_TXPKT_VLAN_VLD |
5412 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
5413 			txq->vlan_insertion++;
5414 		}
5415 
5416 		/* CPL header */
5417 		cpl->ctrl0 = txq->cpl_ctrl0;
5418 		cpl->pack = 0;
5419 		cpl->len = htobe16(m->m_pkthdr.len);
5420 		cpl->ctrl1 = htobe64(ctrl1);
5421 
5422 		flitp = cpl + 1;
5423 		if (checkwrap &&
5424 		    (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
5425 			flitp = (void *)&eq->desc[0];
5426 
5427 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap);
5428 
5429 		if (last != NULL)
5430 			last->m_nextpkt = m;
5431 		last = m;
5432 	}
5433 
5434 	txq->sgl_wrs++;
5435 	if (txp->wr_type == 0) {
5436 		txq->txpkts0_pkts += txp->npkt;
5437 		txq->txpkts0_wrs++;
5438 	} else {
5439 		txq->txpkts1_pkts += txp->npkt;
5440 		txq->txpkts1_wrs++;
5441 	}
5442 
5443 	txsd = &txq->sdesc[eq->pidx];
5444 	txsd->m = txp->mb[0];
5445 	txsd->desc_used = ndesc;
5446 
5447 	return (ndesc);
5448 }
5449 
5450 static u_int
5451 write_txpkts_vm_wr(struct adapter *sc, struct sge_txq *txq)
5452 {
5453 	const struct txpkts *txp = &txq->txp;
5454 	struct sge_eq *eq = &txq->eq;
5455 	struct fw_eth_tx_pkts_vm_wr *wr;
5456 	struct tx_sdesc *txsd;
5457 	struct cpl_tx_pkt_core *cpl;
5458 	uint64_t ctrl1;
5459 	int ndesc, i;
5460 	struct mbuf *m, *last;
5461 	void *flitp;
5462 
5463 	TXQ_LOCK_ASSERT_OWNED(txq);
5464 	MPASS(txp->npkt > 0);
5465 	MPASS(txp->wr_type == 1);	/* VF supports type 1 only */
5466 	MPASS(txp->mb[0] != NULL);
5467 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
5468 
5469 	wr = (void *)&eq->desc[eq->pidx];
5470 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_VM_WR));
5471 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(txp->len16));
5472 	wr->r3 = 0;
5473 	wr->plen = htobe16(txp->plen);
5474 	wr->npkt = txp->npkt;
5475 	wr->r4 = 0;
5476 	memcpy(&wr->ethmacdst[0], &txp->ethmacdst[0], 16);
5477 	flitp = wr + 1;
5478 
5479 	/*
5480 	 * At this point we are 32B into a hardware descriptor.  Each mbuf in
5481 	 * the WR will take 32B so we check for the end of the descriptor ring
5482 	 * before writing odd mbufs (mb[1], 3, 5, ..)
5483 	 */
5484 	ndesc = tx_len16_to_desc(txp->len16);
5485 	last = NULL;
5486 	for (i = 0; i < txp->npkt; i++) {
5487 		m = txp->mb[i];
5488 		if (i & 1 && (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
5489 			flitp = &eq->desc[0];
5490 		cpl = flitp;
5491 
5492 		/* Checksum offload */
5493 		ctrl1 = csum_to_ctrl(sc, m);
5494 		if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS))
5495 			txq->txcsum++;	/* some hardware assistance provided */
5496 
5497 		/* VLAN tag insertion */
5498 		if (needs_vlan_insertion(m)) {
5499 			ctrl1 |= F_TXPKT_VLAN_VLD |
5500 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
5501 			txq->vlan_insertion++;
5502 		}
5503 
5504 		/* CPL header */
5505 		cpl->ctrl0 = txq->cpl_ctrl0;
5506 		cpl->pack = 0;
5507 		cpl->len = htobe16(m->m_pkthdr.len);
5508 		cpl->ctrl1 = htobe64(ctrl1);
5509 
5510 		flitp = cpl + 1;
5511 		MPASS(mbuf_nsegs(m) == 1);
5512 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), 0);
5513 
5514 		if (last != NULL)
5515 			last->m_nextpkt = m;
5516 		last = m;
5517 	}
5518 
5519 	txq->sgl_wrs++;
5520 	txq->txpkts1_pkts += txp->npkt;
5521 	txq->txpkts1_wrs++;
5522 
5523 	txsd = &txq->sdesc[eq->pidx];
5524 	txsd->m = txp->mb[0];
5525 	txsd->desc_used = ndesc;
5526 
5527 	return (ndesc);
5528 }
5529 
5530 /*
5531  * If the SGL ends on an address that is not 16 byte aligned, this function will
5532  * add a 0 filled flit at the end.
5533  */
5534 static void
5535 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap)
5536 {
5537 	struct sge_eq *eq = &txq->eq;
5538 	struct sglist *gl = txq->gl;
5539 	struct sglist_seg *seg;
5540 	__be64 *flitp, *wrap;
5541 	struct ulptx_sgl *usgl;
5542 	int i, nflits, nsegs;
5543 
5544 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
5545 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
5546 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
5547 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
5548 
5549 	get_pkt_gl(m, gl);
5550 	nsegs = gl->sg_nseg;
5551 	MPASS(nsegs > 0);
5552 
5553 	nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
5554 	flitp = (__be64 *)(*to);
5555 	wrap = (__be64 *)(&eq->desc[eq->sidx]);
5556 	seg = &gl->sg_segs[0];
5557 	usgl = (void *)flitp;
5558 
5559 	/*
5560 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
5561 	 * ring, so we're at least 16 bytes away from the status page.  There is
5562 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
5563 	 */
5564 
5565 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
5566 	    V_ULPTX_NSGE(nsegs));
5567 	usgl->len0 = htobe32(seg->ss_len);
5568 	usgl->addr0 = htobe64(seg->ss_paddr);
5569 	seg++;
5570 
5571 	if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) {
5572 
5573 		/* Won't wrap around at all */
5574 
5575 		for (i = 0; i < nsegs - 1; i++, seg++) {
5576 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
5577 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
5578 		}
5579 		if (i & 1)
5580 			usgl->sge[i / 2].len[1] = htobe32(0);
5581 		flitp += nflits;
5582 	} else {
5583 
5584 		/* Will wrap somewhere in the rest of the SGL */
5585 
5586 		/* 2 flits already written, write the rest flit by flit */
5587 		flitp = (void *)(usgl + 1);
5588 		for (i = 0; i < nflits - 2; i++) {
5589 			if (flitp == wrap)
5590 				flitp = (void *)eq->desc;
5591 			*flitp++ = get_flit(seg, nsegs - 1, i);
5592 		}
5593 	}
5594 
5595 	if (nflits & 1) {
5596 		MPASS(((uintptr_t)flitp) & 0xf);
5597 		*flitp++ = 0;
5598 	}
5599 
5600 	MPASS((((uintptr_t)flitp) & 0xf) == 0);
5601 	if (__predict_false(flitp == wrap))
5602 		*to = (void *)eq->desc;
5603 	else
5604 		*to = (void *)flitp;
5605 }
5606 
5607 static inline void
5608 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
5609 {
5610 
5611 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
5612 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
5613 
5614 	if (__predict_true((uintptr_t)(*to) + len <=
5615 	    (uintptr_t)&eq->desc[eq->sidx])) {
5616 		bcopy(from, *to, len);
5617 		(*to) += len;
5618 	} else {
5619 		int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
5620 
5621 		bcopy(from, *to, portion);
5622 		from += portion;
5623 		portion = len - portion;	/* remaining */
5624 		bcopy(from, (void *)eq->desc, portion);
5625 		(*to) = (caddr_t)eq->desc + portion;
5626 	}
5627 }
5628 
5629 static inline void
5630 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n)
5631 {
5632 	u_int db;
5633 
5634 	MPASS(n > 0);
5635 
5636 	db = eq->doorbells;
5637 	if (n > 1)
5638 		clrbit(&db, DOORBELL_WCWR);
5639 	wmb();
5640 
5641 	switch (ffs(db) - 1) {
5642 	case DOORBELL_UDB:
5643 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
5644 		break;
5645 
5646 	case DOORBELL_WCWR: {
5647 		volatile uint64_t *dst, *src;
5648 		int i;
5649 
5650 		/*
5651 		 * Queues whose 128B doorbell segment fits in the page do not
5652 		 * use relative qid (udb_qid is always 0).  Only queues with
5653 		 * doorbell segments can do WCWR.
5654 		 */
5655 		KASSERT(eq->udb_qid == 0 && n == 1,
5656 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
5657 		    __func__, eq->doorbells, n, eq->dbidx, eq));
5658 
5659 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
5660 		    UDBS_DB_OFFSET);
5661 		i = eq->dbidx;
5662 		src = (void *)&eq->desc[i];
5663 		while (src != (void *)&eq->desc[i + 1])
5664 			*dst++ = *src++;
5665 		wmb();
5666 		break;
5667 	}
5668 
5669 	case DOORBELL_UDBWC:
5670 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
5671 		wmb();
5672 		break;
5673 
5674 	case DOORBELL_KDB:
5675 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
5676 		    V_QID(eq->cntxt_id) | V_PIDX(n));
5677 		break;
5678 	}
5679 
5680 	IDXINCR(eq->dbidx, n, eq->sidx);
5681 }
5682 
5683 static inline u_int
5684 reclaimable_tx_desc(struct sge_eq *eq)
5685 {
5686 	uint16_t hw_cidx;
5687 
5688 	hw_cidx = read_hw_cidx(eq);
5689 	return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx));
5690 }
5691 
5692 static inline u_int
5693 total_available_tx_desc(struct sge_eq *eq)
5694 {
5695 	uint16_t hw_cidx, pidx;
5696 
5697 	hw_cidx = read_hw_cidx(eq);
5698 	pidx = eq->pidx;
5699 
5700 	if (pidx == hw_cidx)
5701 		return (eq->sidx - 1);
5702 	else
5703 		return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1);
5704 }
5705 
5706 static inline uint16_t
5707 read_hw_cidx(struct sge_eq *eq)
5708 {
5709 	struct sge_qstat *spg = (void *)&eq->desc[eq->sidx];
5710 	uint16_t cidx = spg->cidx;	/* stable snapshot */
5711 
5712 	return (be16toh(cidx));
5713 }
5714 
5715 /*
5716  * Reclaim 'n' descriptors approximately.
5717  */
5718 static u_int
5719 reclaim_tx_descs(struct sge_txq *txq, u_int n)
5720 {
5721 	struct tx_sdesc *txsd;
5722 	struct sge_eq *eq = &txq->eq;
5723 	u_int can_reclaim, reclaimed;
5724 
5725 	TXQ_LOCK_ASSERT_OWNED(txq);
5726 	MPASS(n > 0);
5727 
5728 	reclaimed = 0;
5729 	can_reclaim = reclaimable_tx_desc(eq);
5730 	while (can_reclaim && reclaimed < n) {
5731 		int ndesc;
5732 		struct mbuf *m, *nextpkt;
5733 
5734 		txsd = &txq->sdesc[eq->cidx];
5735 		ndesc = txsd->desc_used;
5736 
5737 		/* Firmware doesn't return "partial" credits. */
5738 		KASSERT(can_reclaim >= ndesc,
5739 		    ("%s: unexpected number of credits: %d, %d",
5740 		    __func__, can_reclaim, ndesc));
5741 		KASSERT(ndesc != 0,
5742 		    ("%s: descriptor with no credits: cidx %d",
5743 		    __func__, eq->cidx));
5744 
5745 		for (m = txsd->m; m != NULL; m = nextpkt) {
5746 			nextpkt = m->m_nextpkt;
5747 			m->m_nextpkt = NULL;
5748 			m_freem(m);
5749 		}
5750 		reclaimed += ndesc;
5751 		can_reclaim -= ndesc;
5752 		IDXINCR(eq->cidx, ndesc, eq->sidx);
5753 	}
5754 
5755 	return (reclaimed);
5756 }
5757 
5758 static void
5759 tx_reclaim(void *arg, int n)
5760 {
5761 	struct sge_txq *txq = arg;
5762 	struct sge_eq *eq = &txq->eq;
5763 
5764 	do {
5765 		if (TXQ_TRYLOCK(txq) == 0)
5766 			break;
5767 		n = reclaim_tx_descs(txq, 32);
5768 		if (eq->cidx == eq->pidx)
5769 			eq->equeqidx = eq->pidx;
5770 		TXQ_UNLOCK(txq);
5771 	} while (n > 0);
5772 }
5773 
5774 static __be64
5775 get_flit(struct sglist_seg *segs, int nsegs, int idx)
5776 {
5777 	int i = (idx / 3) * 2;
5778 
5779 	switch (idx % 3) {
5780 	case 0: {
5781 		uint64_t rc;
5782 
5783 		rc = (uint64_t)segs[i].ss_len << 32;
5784 		if (i + 1 < nsegs)
5785 			rc |= (uint64_t)(segs[i + 1].ss_len);
5786 
5787 		return (htobe64(rc));
5788 	}
5789 	case 1:
5790 		return (htobe64(segs[i].ss_paddr));
5791 	case 2:
5792 		return (htobe64(segs[i + 1].ss_paddr));
5793 	}
5794 
5795 	return (0);
5796 }
5797 
5798 static int
5799 find_refill_source(struct adapter *sc, int maxp, bool packing)
5800 {
5801 	int i, zidx = -1;
5802 	struct rx_buf_info *rxb = &sc->sge.rx_buf_info[0];
5803 
5804 	if (packing) {
5805 		for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) {
5806 			if (rxb->hwidx2 == -1)
5807 				continue;
5808 			if (rxb->size1 < PAGE_SIZE &&
5809 			    rxb->size1 < largest_rx_cluster)
5810 				continue;
5811 			if (rxb->size1 > largest_rx_cluster)
5812 				break;
5813 			MPASS(rxb->size1 - rxb->size2 >= CL_METADATA_SIZE);
5814 			if (rxb->size2 >= maxp)
5815 				return (i);
5816 			zidx = i;
5817 		}
5818 	} else {
5819 		for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) {
5820 			if (rxb->hwidx1 == -1)
5821 				continue;
5822 			if (rxb->size1 > largest_rx_cluster)
5823 				break;
5824 			if (rxb->size1 >= maxp)
5825 				return (i);
5826 			zidx = i;
5827 		}
5828 	}
5829 
5830 	return (zidx);
5831 }
5832 
5833 static void
5834 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
5835 {
5836 	mtx_lock(&sc->sfl_lock);
5837 	FL_LOCK(fl);
5838 	if ((fl->flags & FL_DOOMED) == 0) {
5839 		fl->flags |= FL_STARVING;
5840 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
5841 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
5842 	}
5843 	FL_UNLOCK(fl);
5844 	mtx_unlock(&sc->sfl_lock);
5845 }
5846 
5847 static void
5848 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq)
5849 {
5850 	struct sge_wrq *wrq = (void *)eq;
5851 
5852 	atomic_readandclear_int(&eq->equiq);
5853 	taskqueue_enqueue(sc->tq[eq->tx_chan], &wrq->wrq_tx_task);
5854 }
5855 
5856 static void
5857 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq)
5858 {
5859 	struct sge_txq *txq = (void *)eq;
5860 
5861 	MPASS((eq->flags & EQ_TYPEMASK) == EQ_ETH);
5862 
5863 	atomic_readandclear_int(&eq->equiq);
5864 	if (mp_ring_is_idle(txq->r))
5865 		taskqueue_enqueue(sc->tq[eq->tx_chan], &txq->tx_reclaim_task);
5866 	else
5867 		mp_ring_check_drainage(txq->r, 64);
5868 }
5869 
5870 static int
5871 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
5872     struct mbuf *m)
5873 {
5874 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
5875 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
5876 	struct adapter *sc = iq->adapter;
5877 	struct sge *s = &sc->sge;
5878 	struct sge_eq *eq;
5879 	static void (*h[])(struct adapter *, struct sge_eq *) = {NULL,
5880 		&handle_wrq_egr_update, &handle_eth_egr_update,
5881 		&handle_wrq_egr_update};
5882 
5883 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5884 	    rss->opcode));
5885 
5886 	eq = s->eqmap[qid - s->eq_start - s->eq_base];
5887 	(*h[eq->flags & EQ_TYPEMASK])(sc, eq);
5888 
5889 	return (0);
5890 }
5891 
5892 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
5893 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
5894     offsetof(struct cpl_fw6_msg, data));
5895 
5896 static int
5897 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
5898 {
5899 	struct adapter *sc = iq->adapter;
5900 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
5901 
5902 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5903 	    rss->opcode));
5904 
5905 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
5906 		const struct rss_header *rss2;
5907 
5908 		rss2 = (const struct rss_header *)&cpl->data[0];
5909 		return (t4_cpl_handler[rss2->opcode](iq, rss2, m));
5910 	}
5911 
5912 	return (t4_fw_msg_handler[cpl->type](sc, &cpl->data[0]));
5913 }
5914 
5915 /**
5916  *	t4_handle_wrerr_rpl - process a FW work request error message
5917  *	@adap: the adapter
5918  *	@rpl: start of the FW message
5919  */
5920 static int
5921 t4_handle_wrerr_rpl(struct adapter *adap, const __be64 *rpl)
5922 {
5923 	u8 opcode = *(const u8 *)rpl;
5924 	const struct fw_error_cmd *e = (const void *)rpl;
5925 	unsigned int i;
5926 
5927 	if (opcode != FW_ERROR_CMD) {
5928 		log(LOG_ERR,
5929 		    "%s: Received WRERR_RPL message with opcode %#x\n",
5930 		    device_get_nameunit(adap->dev), opcode);
5931 		return (EINVAL);
5932 	}
5933 	log(LOG_ERR, "%s: FW_ERROR (%s) ", device_get_nameunit(adap->dev),
5934 	    G_FW_ERROR_CMD_FATAL(be32toh(e->op_to_type)) ? "fatal" :
5935 	    "non-fatal");
5936 	switch (G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type))) {
5937 	case FW_ERROR_TYPE_EXCEPTION:
5938 		log(LOG_ERR, "exception info:\n");
5939 		for (i = 0; i < nitems(e->u.exception.info); i++)
5940 			log(LOG_ERR, "%s%08x", i == 0 ? "\t" : " ",
5941 			    be32toh(e->u.exception.info[i]));
5942 		log(LOG_ERR, "\n");
5943 		break;
5944 	case FW_ERROR_TYPE_HWMODULE:
5945 		log(LOG_ERR, "HW module regaddr %08x regval %08x\n",
5946 		    be32toh(e->u.hwmodule.regaddr),
5947 		    be32toh(e->u.hwmodule.regval));
5948 		break;
5949 	case FW_ERROR_TYPE_WR:
5950 		log(LOG_ERR, "WR cidx %d PF %d VF %d eqid %d hdr:\n",
5951 		    be16toh(e->u.wr.cidx),
5952 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.wr.pfn_vfn)),
5953 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.wr.pfn_vfn)),
5954 		    be32toh(e->u.wr.eqid));
5955 		for (i = 0; i < nitems(e->u.wr.wrhdr); i++)
5956 			log(LOG_ERR, "%s%02x", i == 0 ? "\t" : " ",
5957 			    e->u.wr.wrhdr[i]);
5958 		log(LOG_ERR, "\n");
5959 		break;
5960 	case FW_ERROR_TYPE_ACL:
5961 		log(LOG_ERR, "ACL cidx %d PF %d VF %d eqid %d %s",
5962 		    be16toh(e->u.acl.cidx),
5963 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.acl.pfn_vfn)),
5964 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.acl.pfn_vfn)),
5965 		    be32toh(e->u.acl.eqid),
5966 		    G_FW_ERROR_CMD_MV(be16toh(e->u.acl.mv_pkd)) ? "vlanid" :
5967 		    "MAC");
5968 		for (i = 0; i < nitems(e->u.acl.val); i++)
5969 			log(LOG_ERR, " %02x", e->u.acl.val[i]);
5970 		log(LOG_ERR, "\n");
5971 		break;
5972 	default:
5973 		log(LOG_ERR, "type %#x\n",
5974 		    G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type)));
5975 		return (EINVAL);
5976 	}
5977 	return (0);
5978 }
5979 
5980 int
5981 sysctl_uint16(SYSCTL_HANDLER_ARGS)
5982 {
5983 	uint16_t *id = arg1;
5984 	int i = *id;
5985 
5986 	return sysctl_handle_int(oidp, &i, 0, req);
5987 }
5988 
5989 static inline bool
5990 bufidx_used(struct adapter *sc, int idx)
5991 {
5992 	struct rx_buf_info *rxb = &sc->sge.rx_buf_info[0];
5993 	int i;
5994 
5995 	for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) {
5996 		if (rxb->size1 > largest_rx_cluster)
5997 			continue;
5998 		if (rxb->hwidx1 == idx || rxb->hwidx2 == idx)
5999 			return (true);
6000 	}
6001 
6002 	return (false);
6003 }
6004 
6005 static int
6006 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
6007 {
6008 	struct adapter *sc = arg1;
6009 	struct sge_params *sp = &sc->params.sge;
6010 	int i, rc;
6011 	struct sbuf sb;
6012 	char c;
6013 
6014 	sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
6015 	for (i = 0; i < SGE_FLBUF_SIZES; i++) {
6016 		if (bufidx_used(sc, i))
6017 			c = '*';
6018 		else
6019 			c = '\0';
6020 
6021 		sbuf_printf(&sb, "%u%c ", sp->sge_fl_buffer_size[i], c);
6022 	}
6023 	sbuf_trim(&sb);
6024 	sbuf_finish(&sb);
6025 	rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
6026 	sbuf_delete(&sb);
6027 	return (rc);
6028 }
6029 
6030 #ifdef RATELIMIT
6031 /*
6032  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
6033  */
6034 static inline u_int
6035 txpkt_eo_len16(u_int nsegs, u_int immhdrs, u_int tso)
6036 {
6037 	u_int n;
6038 
6039 	MPASS(immhdrs > 0);
6040 
6041 	n = roundup2(sizeof(struct fw_eth_tx_eo_wr) +
6042 	    sizeof(struct cpl_tx_pkt_core) + immhdrs, 16);
6043 	if (__predict_false(nsegs == 0))
6044 		goto done;
6045 
6046 	nsegs--; /* first segment is part of ulptx_sgl */
6047 	n += sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
6048 	if (tso)
6049 		n += sizeof(struct cpl_tx_pkt_lso_core);
6050 
6051 done:
6052 	return (howmany(n, 16));
6053 }
6054 
6055 #define ETID_FLOWC_NPARAMS 6
6056 #define ETID_FLOWC_LEN (roundup2((sizeof(struct fw_flowc_wr) + \
6057     ETID_FLOWC_NPARAMS * sizeof(struct fw_flowc_mnemval)), 16))
6058 #define ETID_FLOWC_LEN16 (howmany(ETID_FLOWC_LEN, 16))
6059 
6060 static int
6061 send_etid_flowc_wr(struct cxgbe_rate_tag *cst, struct port_info *pi,
6062     struct vi_info *vi)
6063 {
6064 	struct wrq_cookie cookie;
6065 	u_int pfvf = pi->adapter->pf << S_FW_VIID_PFN;
6066 	struct fw_flowc_wr *flowc;
6067 
6068 	mtx_assert(&cst->lock, MA_OWNED);
6069 	MPASS((cst->flags & (EO_FLOWC_PENDING | EO_FLOWC_RPL_PENDING)) ==
6070 	    EO_FLOWC_PENDING);
6071 
6072 	flowc = start_wrq_wr(cst->eo_txq, ETID_FLOWC_LEN16, &cookie);
6073 	if (__predict_false(flowc == NULL))
6074 		return (ENOMEM);
6075 
6076 	bzero(flowc, ETID_FLOWC_LEN);
6077 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
6078 	    V_FW_FLOWC_WR_NPARAMS(ETID_FLOWC_NPARAMS) | V_FW_WR_COMPL(0));
6079 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(ETID_FLOWC_LEN16) |
6080 	    V_FW_WR_FLOWID(cst->etid));
6081 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN;
6082 	flowc->mnemval[0].val = htobe32(pfvf);
6083 	flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH;
6084 	flowc->mnemval[1].val = htobe32(pi->tx_chan);
6085 	flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT;
6086 	flowc->mnemval[2].val = htobe32(pi->tx_chan);
6087 	flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID;
6088 	flowc->mnemval[3].val = htobe32(cst->iqid);
6089 	flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_EOSTATE;
6090 	flowc->mnemval[4].val = htobe32(FW_FLOWC_MNEM_EOSTATE_ESTABLISHED);
6091 	flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS;
6092 	flowc->mnemval[5].val = htobe32(cst->schedcl);
6093 
6094 	commit_wrq_wr(cst->eo_txq, flowc, &cookie);
6095 
6096 	cst->flags &= ~EO_FLOWC_PENDING;
6097 	cst->flags |= EO_FLOWC_RPL_PENDING;
6098 	MPASS(cst->tx_credits >= ETID_FLOWC_LEN16);	/* flowc is first WR. */
6099 	cst->tx_credits -= ETID_FLOWC_LEN16;
6100 
6101 	return (0);
6102 }
6103 
6104 #define ETID_FLUSH_LEN16 (howmany(sizeof (struct fw_flowc_wr), 16))
6105 
6106 void
6107 send_etid_flush_wr(struct cxgbe_rate_tag *cst)
6108 {
6109 	struct fw_flowc_wr *flowc;
6110 	struct wrq_cookie cookie;
6111 
6112 	mtx_assert(&cst->lock, MA_OWNED);
6113 
6114 	flowc = start_wrq_wr(cst->eo_txq, ETID_FLUSH_LEN16, &cookie);
6115 	if (__predict_false(flowc == NULL))
6116 		CXGBE_UNIMPLEMENTED(__func__);
6117 
6118 	bzero(flowc, ETID_FLUSH_LEN16 * 16);
6119 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
6120 	    V_FW_FLOWC_WR_NPARAMS(0) | F_FW_WR_COMPL);
6121 	flowc->flowid_len16 = htobe32(V_FW_WR_LEN16(ETID_FLUSH_LEN16) |
6122 	    V_FW_WR_FLOWID(cst->etid));
6123 
6124 	commit_wrq_wr(cst->eo_txq, flowc, &cookie);
6125 
6126 	cst->flags |= EO_FLUSH_RPL_PENDING;
6127 	MPASS(cst->tx_credits >= ETID_FLUSH_LEN16);
6128 	cst->tx_credits -= ETID_FLUSH_LEN16;
6129 	cst->ncompl++;
6130 }
6131 
6132 static void
6133 write_ethofld_wr(struct cxgbe_rate_tag *cst, struct fw_eth_tx_eo_wr *wr,
6134     struct mbuf *m0, int compl)
6135 {
6136 	struct cpl_tx_pkt_core *cpl;
6137 	uint64_t ctrl1;
6138 	uint32_t ctrl;	/* used in many unrelated places */
6139 	int len16, pktlen, nsegs, immhdrs;
6140 	caddr_t dst;
6141 	uintptr_t p;
6142 	struct ulptx_sgl *usgl;
6143 	struct sglist sg;
6144 	struct sglist_seg segs[38];	/* XXX: find real limit.  XXX: get off the stack */
6145 
6146 	mtx_assert(&cst->lock, MA_OWNED);
6147 	M_ASSERTPKTHDR(m0);
6148 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
6149 	    m0->m_pkthdr.l4hlen > 0,
6150 	    ("%s: ethofld mbuf %p is missing header lengths", __func__, m0));
6151 
6152 	len16 = mbuf_eo_len16(m0);
6153 	nsegs = mbuf_eo_nsegs(m0);
6154 	pktlen = m0->m_pkthdr.len;
6155 	ctrl = sizeof(struct cpl_tx_pkt_core);
6156 	if (needs_tso(m0))
6157 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
6158 	immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen + m0->m_pkthdr.l4hlen;
6159 	ctrl += immhdrs;
6160 
6161 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_EO_WR) |
6162 	    V_FW_ETH_TX_EO_WR_IMMDLEN(ctrl) | V_FW_WR_COMPL(!!compl));
6163 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(len16) |
6164 	    V_FW_WR_FLOWID(cst->etid));
6165 	wr->r3 = 0;
6166 	if (needs_outer_udp_csum(m0)) {
6167 		wr->u.udpseg.type = FW_ETH_TX_EO_TYPE_UDPSEG;
6168 		wr->u.udpseg.ethlen = m0->m_pkthdr.l2hlen;
6169 		wr->u.udpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
6170 		wr->u.udpseg.udplen = m0->m_pkthdr.l4hlen;
6171 		wr->u.udpseg.rtplen = 0;
6172 		wr->u.udpseg.r4 = 0;
6173 		wr->u.udpseg.mss = htobe16(pktlen - immhdrs);
6174 		wr->u.udpseg.schedpktsize = wr->u.udpseg.mss;
6175 		wr->u.udpseg.plen = htobe32(pktlen - immhdrs);
6176 		cpl = (void *)(wr + 1);
6177 	} else {
6178 		MPASS(needs_outer_tcp_csum(m0));
6179 		wr->u.tcpseg.type = FW_ETH_TX_EO_TYPE_TCPSEG;
6180 		wr->u.tcpseg.ethlen = m0->m_pkthdr.l2hlen;
6181 		wr->u.tcpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
6182 		wr->u.tcpseg.tcplen = m0->m_pkthdr.l4hlen;
6183 		wr->u.tcpseg.tsclk_tsoff = mbuf_eo_tsclk_tsoff(m0);
6184 		wr->u.tcpseg.r4 = 0;
6185 		wr->u.tcpseg.r5 = 0;
6186 		wr->u.tcpseg.plen = htobe32(pktlen - immhdrs);
6187 
6188 		if (needs_tso(m0)) {
6189 			struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
6190 
6191 			wr->u.tcpseg.mss = htobe16(m0->m_pkthdr.tso_segsz);
6192 
6193 			ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) |
6194 			    F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE |
6195 			    V_LSO_ETHHDR_LEN((m0->m_pkthdr.l2hlen -
6196 				ETHER_HDR_LEN) >> 2) |
6197 			    V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) |
6198 			    V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
6199 			if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
6200 				ctrl |= F_LSO_IPV6;
6201 			lso->lso_ctrl = htobe32(ctrl);
6202 			lso->ipid_ofst = htobe16(0);
6203 			lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
6204 			lso->seqno_offset = htobe32(0);
6205 			lso->len = htobe32(pktlen);
6206 
6207 			cpl = (void *)(lso + 1);
6208 		} else {
6209 			wr->u.tcpseg.mss = htobe16(0xffff);
6210 			cpl = (void *)(wr + 1);
6211 		}
6212 	}
6213 
6214 	/* Checksum offload must be requested for ethofld. */
6215 	MPASS(needs_outer_l4_csum(m0));
6216 	ctrl1 = csum_to_ctrl(cst->adapter, m0);
6217 
6218 	/* VLAN tag insertion */
6219 	if (needs_vlan_insertion(m0)) {
6220 		ctrl1 |= F_TXPKT_VLAN_VLD |
6221 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
6222 	}
6223 
6224 	/* CPL header */
6225 	cpl->ctrl0 = cst->ctrl0;
6226 	cpl->pack = 0;
6227 	cpl->len = htobe16(pktlen);
6228 	cpl->ctrl1 = htobe64(ctrl1);
6229 
6230 	/* Copy Ethernet, IP & TCP/UDP hdrs as immediate data */
6231 	p = (uintptr_t)(cpl + 1);
6232 	m_copydata(m0, 0, immhdrs, (void *)p);
6233 
6234 	/* SGL */
6235 	dst = (void *)(cpl + 1);
6236 	if (nsegs > 0) {
6237 		int i, pad;
6238 
6239 		/* zero-pad upto next 16Byte boundary, if not 16Byte aligned */
6240 		p += immhdrs;
6241 		pad = 16 - (immhdrs & 0xf);
6242 		bzero((void *)p, pad);
6243 
6244 		usgl = (void *)(p + pad);
6245 		usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
6246 		    V_ULPTX_NSGE(nsegs));
6247 
6248 		sglist_init(&sg, nitems(segs), segs);
6249 		for (; m0 != NULL; m0 = m0->m_next) {
6250 			if (__predict_false(m0->m_len == 0))
6251 				continue;
6252 			if (immhdrs >= m0->m_len) {
6253 				immhdrs -= m0->m_len;
6254 				continue;
6255 			}
6256 			if (m0->m_flags & M_EXTPG)
6257 				sglist_append_mbuf_epg(&sg, m0,
6258 				    mtod(m0, vm_offset_t), m0->m_len);
6259                         else
6260 				sglist_append(&sg, mtod(m0, char *) + immhdrs,
6261 				    m0->m_len - immhdrs);
6262 			immhdrs = 0;
6263 		}
6264 		MPASS(sg.sg_nseg == nsegs);
6265 
6266 		/*
6267 		 * Zero pad last 8B in case the WR doesn't end on a 16B
6268 		 * boundary.
6269 		 */
6270 		*(uint64_t *)((char *)wr + len16 * 16 - 8) = 0;
6271 
6272 		usgl->len0 = htobe32(segs[0].ss_len);
6273 		usgl->addr0 = htobe64(segs[0].ss_paddr);
6274 		for (i = 0; i < nsegs - 1; i++) {
6275 			usgl->sge[i / 2].len[i & 1] = htobe32(segs[i + 1].ss_len);
6276 			usgl->sge[i / 2].addr[i & 1] = htobe64(segs[i + 1].ss_paddr);
6277 		}
6278 		if (i & 1)
6279 			usgl->sge[i / 2].len[1] = htobe32(0);
6280 	}
6281 
6282 }
6283 
6284 static void
6285 ethofld_tx(struct cxgbe_rate_tag *cst)
6286 {
6287 	struct mbuf *m;
6288 	struct wrq_cookie cookie;
6289 	int next_credits, compl;
6290 	struct fw_eth_tx_eo_wr *wr;
6291 
6292 	mtx_assert(&cst->lock, MA_OWNED);
6293 
6294 	while ((m = mbufq_first(&cst->pending_tx)) != NULL) {
6295 		M_ASSERTPKTHDR(m);
6296 
6297 		/* How many len16 credits do we need to send this mbuf. */
6298 		next_credits = mbuf_eo_len16(m);
6299 		MPASS(next_credits > 0);
6300 		if (next_credits > cst->tx_credits) {
6301 			/*
6302 			 * Tx will make progress eventually because there is at
6303 			 * least one outstanding fw4_ack that will return
6304 			 * credits and kick the tx.
6305 			 */
6306 			MPASS(cst->ncompl > 0);
6307 			return;
6308 		}
6309 		wr = start_wrq_wr(cst->eo_txq, next_credits, &cookie);
6310 		if (__predict_false(wr == NULL)) {
6311 			/* XXX: wishful thinking, not a real assertion. */
6312 			MPASS(cst->ncompl > 0);
6313 			return;
6314 		}
6315 		cst->tx_credits -= next_credits;
6316 		cst->tx_nocompl += next_credits;
6317 		compl = cst->ncompl == 0 || cst->tx_nocompl >= cst->tx_total / 2;
6318 		ETHER_BPF_MTAP(cst->com.ifp, m);
6319 		write_ethofld_wr(cst, wr, m, compl);
6320 		commit_wrq_wr(cst->eo_txq, wr, &cookie);
6321 		if (compl) {
6322 			cst->ncompl++;
6323 			cst->tx_nocompl	= 0;
6324 		}
6325 		(void) mbufq_dequeue(&cst->pending_tx);
6326 
6327 		/*
6328 		 * Drop the mbuf's reference on the tag now rather
6329 		 * than waiting until m_freem().  This ensures that
6330 		 * cxgbe_rate_tag_free gets called when the inp drops
6331 		 * its reference on the tag and there are no more
6332 		 * mbufs in the pending_tx queue and can flush any
6333 		 * pending requests.  Otherwise if the last mbuf
6334 		 * doesn't request a completion the etid will never be
6335 		 * released.
6336 		 */
6337 		m->m_pkthdr.snd_tag = NULL;
6338 		m->m_pkthdr.csum_flags &= ~CSUM_SND_TAG;
6339 		m_snd_tag_rele(&cst->com);
6340 
6341 		mbufq_enqueue(&cst->pending_fwack, m);
6342 	}
6343 }
6344 
6345 int
6346 ethofld_transmit(struct ifnet *ifp, struct mbuf *m0)
6347 {
6348 	struct cxgbe_rate_tag *cst;
6349 	int rc;
6350 
6351 	MPASS(m0->m_nextpkt == NULL);
6352 	MPASS(m0->m_pkthdr.csum_flags & CSUM_SND_TAG);
6353 	MPASS(m0->m_pkthdr.snd_tag != NULL);
6354 	cst = mst_to_crt(m0->m_pkthdr.snd_tag);
6355 
6356 	mtx_lock(&cst->lock);
6357 	MPASS(cst->flags & EO_SND_TAG_REF);
6358 
6359 	if (__predict_false(cst->flags & EO_FLOWC_PENDING)) {
6360 		struct vi_info *vi = ifp->if_softc;
6361 		struct port_info *pi = vi->pi;
6362 		struct adapter *sc = pi->adapter;
6363 		const uint32_t rss_mask = vi->rss_size - 1;
6364 		uint32_t rss_hash;
6365 
6366 		cst->eo_txq = &sc->sge.ofld_txq[vi->first_ofld_txq];
6367 		if (M_HASHTYPE_ISHASH(m0))
6368 			rss_hash = m0->m_pkthdr.flowid;
6369 		else
6370 			rss_hash = arc4random();
6371 		/* We assume RSS hashing */
6372 		cst->iqid = vi->rss[rss_hash & rss_mask];
6373 		cst->eo_txq += rss_hash % vi->nofldtxq;
6374 		rc = send_etid_flowc_wr(cst, pi, vi);
6375 		if (rc != 0)
6376 			goto done;
6377 	}
6378 
6379 	if (__predict_false(cst->plen + m0->m_pkthdr.len > eo_max_backlog)) {
6380 		rc = ENOBUFS;
6381 		goto done;
6382 	}
6383 
6384 	mbufq_enqueue(&cst->pending_tx, m0);
6385 	cst->plen += m0->m_pkthdr.len;
6386 
6387 	/*
6388 	 * Hold an extra reference on the tag while generating work
6389 	 * requests to ensure that we don't try to free the tag during
6390 	 * ethofld_tx() in case we are sending the final mbuf after
6391 	 * the inp was freed.
6392 	 */
6393 	m_snd_tag_ref(&cst->com);
6394 	ethofld_tx(cst);
6395 	mtx_unlock(&cst->lock);
6396 	m_snd_tag_rele(&cst->com);
6397 	return (0);
6398 
6399 done:
6400 	mtx_unlock(&cst->lock);
6401 	if (__predict_false(rc != 0))
6402 		m_freem(m0);
6403 	return (rc);
6404 }
6405 
6406 static int
6407 ethofld_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
6408 {
6409 	struct adapter *sc = iq->adapter;
6410 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
6411 	struct mbuf *m;
6412 	u_int etid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
6413 	struct cxgbe_rate_tag *cst;
6414 	uint8_t credits = cpl->credits;
6415 
6416 	cst = lookup_etid(sc, etid);
6417 	mtx_lock(&cst->lock);
6418 	if (__predict_false(cst->flags & EO_FLOWC_RPL_PENDING)) {
6419 		MPASS(credits >= ETID_FLOWC_LEN16);
6420 		credits -= ETID_FLOWC_LEN16;
6421 		cst->flags &= ~EO_FLOWC_RPL_PENDING;
6422 	}
6423 
6424 	KASSERT(cst->ncompl > 0,
6425 	    ("%s: etid %u (%p) wasn't expecting completion.",
6426 	    __func__, etid, cst));
6427 	cst->ncompl--;
6428 
6429 	while (credits > 0) {
6430 		m = mbufq_dequeue(&cst->pending_fwack);
6431 		if (__predict_false(m == NULL)) {
6432 			/*
6433 			 * The remaining credits are for the final flush that
6434 			 * was issued when the tag was freed by the kernel.
6435 			 */
6436 			MPASS((cst->flags &
6437 			    (EO_FLUSH_RPL_PENDING | EO_SND_TAG_REF)) ==
6438 			    EO_FLUSH_RPL_PENDING);
6439 			MPASS(credits == ETID_FLUSH_LEN16);
6440 			MPASS(cst->tx_credits + cpl->credits == cst->tx_total);
6441 			MPASS(cst->ncompl == 0);
6442 
6443 			cst->flags &= ~EO_FLUSH_RPL_PENDING;
6444 			cst->tx_credits += cpl->credits;
6445 			cxgbe_rate_tag_free_locked(cst);
6446 			return (0);	/* cst is gone. */
6447 		}
6448 		KASSERT(m != NULL,
6449 		    ("%s: too many credits (%u, %u)", __func__, cpl->credits,
6450 		    credits));
6451 		KASSERT(credits >= mbuf_eo_len16(m),
6452 		    ("%s: too few credits (%u, %u, %u)", __func__,
6453 		    cpl->credits, credits, mbuf_eo_len16(m)));
6454 		credits -= mbuf_eo_len16(m);
6455 		cst->plen -= m->m_pkthdr.len;
6456 		m_freem(m);
6457 	}
6458 
6459 	cst->tx_credits += cpl->credits;
6460 	MPASS(cst->tx_credits <= cst->tx_total);
6461 
6462 	if (cst->flags & EO_SND_TAG_REF) {
6463 		/*
6464 		 * As with ethofld_transmit(), hold an extra reference
6465 		 * so that the tag is stable across ethold_tx().
6466 		 */
6467 		m_snd_tag_ref(&cst->com);
6468 		m = mbufq_first(&cst->pending_tx);
6469 		if (m != NULL && cst->tx_credits >= mbuf_eo_len16(m))
6470 			ethofld_tx(cst);
6471 		mtx_unlock(&cst->lock);
6472 		m_snd_tag_rele(&cst->com);
6473 	} else {
6474 		/*
6475 		 * There shouldn't be any pending packets if the tag
6476 		 * was freed by the kernel since any pending packet
6477 		 * should hold a reference to the tag.
6478 		 */
6479 		MPASS(mbufq_first(&cst->pending_tx) == NULL);
6480 		mtx_unlock(&cst->lock);
6481 	}
6482 
6483 	return (0);
6484 }
6485 #endif
6486