xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision bbd421cdf6d8c6102e6fd3979c5bec21ace3c2e3)
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 static int t4_tx_coalesce = 1;
216 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tx_coalesce, CTLFLAG_RWTUN, &t4_tx_coalesce, 0,
217     "tx coalescing allowed");
218 
219 /*
220  * The driver will make aggressive attempts at tx coalescing if it sees these
221  * many packets eligible for coalescing in quick succession, with no more than
222  * the specified gap in between the eth_tx calls that delivered the packets.
223  */
224 static int t4_tx_coalesce_pkts = 32;
225 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tx_coalesce_pkts, CTLFLAG_RWTUN,
226     &t4_tx_coalesce_pkts, 0,
227     "# of consecutive packets (1 - 255) that will trigger tx coalescing");
228 static int t4_tx_coalesce_gap = 5;
229 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tx_coalesce_gap, CTLFLAG_RWTUN,
230     &t4_tx_coalesce_gap, 0, "tx gap (in microseconds)");
231 
232 static int service_iq(struct sge_iq *, int);
233 static int service_iq_fl(struct sge_iq *, int);
234 static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t);
235 static int eth_rx(struct adapter *, struct sge_rxq *, const struct iq_desc *,
236     u_int);
237 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int);
238 static inline void init_fl(struct adapter *, struct sge_fl *, int, int, char *);
239 static inline void init_eq(struct adapter *, struct sge_eq *, int, int, uint8_t,
240     uint16_t, char *);
241 static int alloc_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *,
242     int, int);
243 static int free_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *);
244 static void add_iq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
245     struct sge_iq *);
246 static void add_fl_sysctls(struct adapter *, struct sysctl_ctx_list *,
247     struct sysctl_oid *, struct sge_fl *);
248 static int alloc_fwq(struct adapter *);
249 static int free_fwq(struct adapter *);
250 static int alloc_ctrlq(struct adapter *, struct sge_wrq *, int,
251     struct sysctl_oid *);
252 static int alloc_rxq(struct vi_info *, struct sge_rxq *, int, int,
253     struct sysctl_oid *);
254 static int free_rxq(struct vi_info *, struct sge_rxq *);
255 #ifdef TCP_OFFLOAD
256 static int alloc_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *, int, int,
257     struct sysctl_oid *);
258 static int free_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *);
259 #endif
260 static int ctrl_eq_alloc(struct adapter *, struct sge_eq *);
261 static int eth_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
262 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
263 static int ofld_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
264 #endif
265 static int alloc_eq(struct adapter *, struct vi_info *, struct sge_eq *);
266 static int free_eq(struct adapter *, struct sge_eq *);
267 static int alloc_wrq(struct adapter *, struct vi_info *, struct sge_wrq *,
268     struct sysctl_oid *);
269 static int free_wrq(struct adapter *, struct sge_wrq *);
270 static int alloc_txq(struct vi_info *, struct sge_txq *, int,
271     struct sysctl_oid *);
272 static int free_txq(struct vi_info *, struct sge_txq *);
273 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
274 static int alloc_ofld_txq(struct vi_info *, struct sge_ofld_txq *, int,
275     struct sysctl_oid *);
276 static int free_ofld_txq(struct vi_info *, struct sge_ofld_txq *);
277 #endif
278 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int);
279 static inline void ring_fl_db(struct adapter *, struct sge_fl *);
280 static int refill_fl(struct adapter *, struct sge_fl *, int);
281 static void refill_sfl(void *);
282 static int alloc_fl_sdesc(struct sge_fl *);
283 static void free_fl_sdesc(struct adapter *, struct sge_fl *);
284 static int find_refill_source(struct adapter *, int, bool);
285 static void add_fl_to_sfl(struct adapter *, struct sge_fl *);
286 
287 static inline void get_pkt_gl(struct mbuf *, struct sglist *);
288 static inline u_int txpkt_len16(u_int, const u_int);
289 static inline u_int txpkt_vm_len16(u_int, const u_int);
290 static inline void calculate_mbuf_len16(struct mbuf *, bool);
291 static inline u_int txpkts0_len16(u_int);
292 static inline u_int txpkts1_len16(void);
293 static u_int write_raw_wr(struct sge_txq *, void *, struct mbuf *, u_int);
294 static u_int write_txpkt_wr(struct adapter *, struct sge_txq *, struct mbuf *,
295     u_int);
296 static u_int write_txpkt_vm_wr(struct adapter *, struct sge_txq *,
297     struct mbuf *);
298 static int add_to_txpkts_vf(struct adapter *, struct sge_txq *, struct mbuf *,
299     int, bool *);
300 static int add_to_txpkts_pf(struct adapter *, struct sge_txq *, struct mbuf *,
301     int, bool *);
302 static u_int write_txpkts_wr(struct adapter *, struct sge_txq *);
303 static u_int write_txpkts_vm_wr(struct adapter *, struct sge_txq *);
304 static void write_gl_to_txd(struct sge_txq *, struct mbuf *, caddr_t *, int);
305 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int);
306 static inline void ring_eq_db(struct adapter *, struct sge_eq *, u_int);
307 static inline uint16_t read_hw_cidx(struct sge_eq *);
308 static inline u_int reclaimable_tx_desc(struct sge_eq *);
309 static inline u_int total_available_tx_desc(struct sge_eq *);
310 static u_int reclaim_tx_descs(struct sge_txq *, u_int);
311 static void tx_reclaim(void *, int);
312 static __be64 get_flit(struct sglist_seg *, int, int);
313 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
314     struct mbuf *);
315 static int handle_fw_msg(struct sge_iq *, const struct rss_header *,
316     struct mbuf *);
317 static int t4_handle_wrerr_rpl(struct adapter *, const __be64 *);
318 static void wrq_tx_drain(void *, int);
319 static void drain_wrq_wr_list(struct adapter *, struct sge_wrq *);
320 
321 static int sysctl_bufsizes(SYSCTL_HANDLER_ARGS);
322 #ifdef RATELIMIT
323 static inline u_int txpkt_eo_len16(u_int, u_int, u_int);
324 static int ethofld_fw4_ack(struct sge_iq *, const struct rss_header *,
325     struct mbuf *);
326 #endif
327 
328 static counter_u64_t extfree_refs;
329 static counter_u64_t extfree_rels;
330 
331 an_handler_t t4_an_handler;
332 fw_msg_handler_t t4_fw_msg_handler[NUM_FW6_TYPES];
333 cpl_handler_t t4_cpl_handler[NUM_CPL_CMDS];
334 cpl_handler_t set_tcb_rpl_handlers[NUM_CPL_COOKIES];
335 cpl_handler_t l2t_write_rpl_handlers[NUM_CPL_COOKIES];
336 cpl_handler_t act_open_rpl_handlers[NUM_CPL_COOKIES];
337 cpl_handler_t abort_rpl_rss_handlers[NUM_CPL_COOKIES];
338 cpl_handler_t fw4_ack_handlers[NUM_CPL_COOKIES];
339 
340 void
341 t4_register_an_handler(an_handler_t h)
342 {
343 	uintptr_t *loc;
344 
345 	MPASS(h == NULL || t4_an_handler == NULL);
346 
347 	loc = (uintptr_t *)&t4_an_handler;
348 	atomic_store_rel_ptr(loc, (uintptr_t)h);
349 }
350 
351 void
352 t4_register_fw_msg_handler(int type, fw_msg_handler_t h)
353 {
354 	uintptr_t *loc;
355 
356 	MPASS(type < nitems(t4_fw_msg_handler));
357 	MPASS(h == NULL || t4_fw_msg_handler[type] == NULL);
358 	/*
359 	 * These are dispatched by the handler for FW{4|6}_CPL_MSG using the CPL
360 	 * handler dispatch table.  Reject any attempt to install a handler for
361 	 * this subtype.
362 	 */
363 	MPASS(type != FW_TYPE_RSSCPL);
364 	MPASS(type != FW6_TYPE_RSSCPL);
365 
366 	loc = (uintptr_t *)&t4_fw_msg_handler[type];
367 	atomic_store_rel_ptr(loc, (uintptr_t)h);
368 }
369 
370 void
371 t4_register_cpl_handler(int opcode, cpl_handler_t h)
372 {
373 	uintptr_t *loc;
374 
375 	MPASS(opcode < nitems(t4_cpl_handler));
376 	MPASS(h == NULL || t4_cpl_handler[opcode] == NULL);
377 
378 	loc = (uintptr_t *)&t4_cpl_handler[opcode];
379 	atomic_store_rel_ptr(loc, (uintptr_t)h);
380 }
381 
382 static int
383 set_tcb_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
384     struct mbuf *m)
385 {
386 	const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
387 	u_int tid;
388 	int cookie;
389 
390 	MPASS(m == NULL);
391 
392 	tid = GET_TID(cpl);
393 	if (is_hpftid(iq->adapter, tid) || is_ftid(iq->adapter, tid)) {
394 		/*
395 		 * The return code for filter-write is put in the CPL cookie so
396 		 * we have to rely on the hardware tid (is_ftid) to determine
397 		 * that this is a response to a filter.
398 		 */
399 		cookie = CPL_COOKIE_FILTER;
400 	} else {
401 		cookie = G_COOKIE(cpl->cookie);
402 	}
403 	MPASS(cookie > CPL_COOKIE_RESERVED);
404 	MPASS(cookie < nitems(set_tcb_rpl_handlers));
405 
406 	return (set_tcb_rpl_handlers[cookie](iq, rss, m));
407 }
408 
409 static int
410 l2t_write_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
411     struct mbuf *m)
412 {
413 	const struct cpl_l2t_write_rpl *rpl = (const void *)(rss + 1);
414 	unsigned int cookie;
415 
416 	MPASS(m == NULL);
417 
418 	cookie = GET_TID(rpl) & F_SYNC_WR ? CPL_COOKIE_TOM : CPL_COOKIE_FILTER;
419 	return (l2t_write_rpl_handlers[cookie](iq, rss, m));
420 }
421 
422 static int
423 act_open_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
424     struct mbuf *m)
425 {
426 	const struct cpl_act_open_rpl *cpl = (const void *)(rss + 1);
427 	u_int cookie = G_TID_COOKIE(G_AOPEN_ATID(be32toh(cpl->atid_status)));
428 
429 	MPASS(m == NULL);
430 	MPASS(cookie != CPL_COOKIE_RESERVED);
431 
432 	return (act_open_rpl_handlers[cookie](iq, rss, m));
433 }
434 
435 static int
436 abort_rpl_rss_handler(struct sge_iq *iq, const struct rss_header *rss,
437     struct mbuf *m)
438 {
439 	struct adapter *sc = iq->adapter;
440 	u_int cookie;
441 
442 	MPASS(m == NULL);
443 	if (is_hashfilter(sc))
444 		cookie = CPL_COOKIE_HASHFILTER;
445 	else
446 		cookie = CPL_COOKIE_TOM;
447 
448 	return (abort_rpl_rss_handlers[cookie](iq, rss, m));
449 }
450 
451 static int
452 fw4_ack_handler(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
453 {
454 	struct adapter *sc = iq->adapter;
455 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
456 	unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
457 	u_int cookie;
458 
459 	MPASS(m == NULL);
460 	if (is_etid(sc, tid))
461 		cookie = CPL_COOKIE_ETHOFLD;
462 	else
463 		cookie = CPL_COOKIE_TOM;
464 
465 	return (fw4_ack_handlers[cookie](iq, rss, m));
466 }
467 
468 static void
469 t4_init_shared_cpl_handlers(void)
470 {
471 
472 	t4_register_cpl_handler(CPL_SET_TCB_RPL, set_tcb_rpl_handler);
473 	t4_register_cpl_handler(CPL_L2T_WRITE_RPL, l2t_write_rpl_handler);
474 	t4_register_cpl_handler(CPL_ACT_OPEN_RPL, act_open_rpl_handler);
475 	t4_register_cpl_handler(CPL_ABORT_RPL_RSS, abort_rpl_rss_handler);
476 	t4_register_cpl_handler(CPL_FW4_ACK, fw4_ack_handler);
477 }
478 
479 void
480 t4_register_shared_cpl_handler(int opcode, cpl_handler_t h, int cookie)
481 {
482 	uintptr_t *loc;
483 
484 	MPASS(opcode < nitems(t4_cpl_handler));
485 	MPASS(cookie > CPL_COOKIE_RESERVED);
486 	MPASS(cookie < NUM_CPL_COOKIES);
487 	MPASS(t4_cpl_handler[opcode] != NULL);
488 
489 	switch (opcode) {
490 	case CPL_SET_TCB_RPL:
491 		loc = (uintptr_t *)&set_tcb_rpl_handlers[cookie];
492 		break;
493 	case CPL_L2T_WRITE_RPL:
494 		loc = (uintptr_t *)&l2t_write_rpl_handlers[cookie];
495 		break;
496 	case CPL_ACT_OPEN_RPL:
497 		loc = (uintptr_t *)&act_open_rpl_handlers[cookie];
498 		break;
499 	case CPL_ABORT_RPL_RSS:
500 		loc = (uintptr_t *)&abort_rpl_rss_handlers[cookie];
501 		break;
502 	case CPL_FW4_ACK:
503 		loc = (uintptr_t *)&fw4_ack_handlers[cookie];
504 		break;
505 	default:
506 		MPASS(0);
507 		return;
508 	}
509 	MPASS(h == NULL || *loc == (uintptr_t)NULL);
510 	atomic_store_rel_ptr(loc, (uintptr_t)h);
511 }
512 
513 /*
514  * Called on MOD_LOAD.  Validates and calculates the SGE tunables.
515  */
516 void
517 t4_sge_modload(void)
518 {
519 
520 	if (fl_pktshift < 0 || fl_pktshift > 7) {
521 		printf("Invalid hw.cxgbe.fl_pktshift value (%d),"
522 		    " using 0 instead.\n", fl_pktshift);
523 		fl_pktshift = 0;
524 	}
525 
526 	if (spg_len != 64 && spg_len != 128) {
527 		int len;
528 
529 #if defined(__i386__) || defined(__amd64__)
530 		len = cpu_clflush_line_size > 64 ? 128 : 64;
531 #else
532 		len = 64;
533 #endif
534 		if (spg_len != -1) {
535 			printf("Invalid hw.cxgbe.spg_len value (%d),"
536 			    " using %d instead.\n", spg_len, len);
537 		}
538 		spg_len = len;
539 	}
540 
541 	if (cong_drop < -1 || cong_drop > 1) {
542 		printf("Invalid hw.cxgbe.cong_drop value (%d),"
543 		    " using 0 instead.\n", cong_drop);
544 		cong_drop = 0;
545 	}
546 
547 	if (tscale != 1 && (tscale < 3 || tscale > 17)) {
548 		printf("Invalid hw.cxgbe.tscale value (%d),"
549 		    " using 1 instead.\n", tscale);
550 		tscale = 1;
551 	}
552 
553 	if (largest_rx_cluster != MCLBYTES &&
554 #if MJUMPAGESIZE != MCLBYTES
555 	    largest_rx_cluster != MJUMPAGESIZE &&
556 #endif
557 	    largest_rx_cluster != MJUM9BYTES &&
558 	    largest_rx_cluster != MJUM16BYTES) {
559 		printf("Invalid hw.cxgbe.largest_rx_cluster value (%d),"
560 		    " using %d instead.\n", largest_rx_cluster, MJUM16BYTES);
561 		largest_rx_cluster = MJUM16BYTES;
562 	}
563 
564 	if (safest_rx_cluster != MCLBYTES &&
565 #if MJUMPAGESIZE != MCLBYTES
566 	    safest_rx_cluster != MJUMPAGESIZE &&
567 #endif
568 	    safest_rx_cluster != MJUM9BYTES &&
569 	    safest_rx_cluster != MJUM16BYTES) {
570 		printf("Invalid hw.cxgbe.safest_rx_cluster value (%d),"
571 		    " using %d instead.\n", safest_rx_cluster, MJUMPAGESIZE);
572 		safest_rx_cluster = MJUMPAGESIZE;
573 	}
574 
575 	extfree_refs = counter_u64_alloc(M_WAITOK);
576 	extfree_rels = counter_u64_alloc(M_WAITOK);
577 	pullups = counter_u64_alloc(M_WAITOK);
578 	defrags = counter_u64_alloc(M_WAITOK);
579 	counter_u64_zero(extfree_refs);
580 	counter_u64_zero(extfree_rels);
581 	counter_u64_zero(pullups);
582 	counter_u64_zero(defrags);
583 
584 	t4_init_shared_cpl_handlers();
585 	t4_register_cpl_handler(CPL_FW4_MSG, handle_fw_msg);
586 	t4_register_cpl_handler(CPL_FW6_MSG, handle_fw_msg);
587 	t4_register_cpl_handler(CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
588 #ifdef RATELIMIT
589 	t4_register_shared_cpl_handler(CPL_FW4_ACK, ethofld_fw4_ack,
590 	    CPL_COOKIE_ETHOFLD);
591 #endif
592 	t4_register_fw_msg_handler(FW6_TYPE_CMD_RPL, t4_handle_fw_rpl);
593 	t4_register_fw_msg_handler(FW6_TYPE_WRERR_RPL, t4_handle_wrerr_rpl);
594 }
595 
596 void
597 t4_sge_modunload(void)
598 {
599 
600 	counter_u64_free(extfree_refs);
601 	counter_u64_free(extfree_rels);
602 	counter_u64_free(pullups);
603 	counter_u64_free(defrags);
604 }
605 
606 uint64_t
607 t4_sge_extfree_refs(void)
608 {
609 	uint64_t refs, rels;
610 
611 	rels = counter_u64_fetch(extfree_rels);
612 	refs = counter_u64_fetch(extfree_refs);
613 
614 	return (refs - rels);
615 }
616 
617 /* max 4096 */
618 #define MAX_PACK_BOUNDARY 512
619 
620 static inline void
621 setup_pad_and_pack_boundaries(struct adapter *sc)
622 {
623 	uint32_t v, m;
624 	int pad, pack, pad_shift;
625 
626 	pad_shift = chip_id(sc) > CHELSIO_T5 ? X_T6_INGPADBOUNDARY_SHIFT :
627 	    X_INGPADBOUNDARY_SHIFT;
628 	pad = fl_pad;
629 	if (fl_pad < (1 << pad_shift) ||
630 	    fl_pad > (1 << (pad_shift + M_INGPADBOUNDARY)) ||
631 	    !powerof2(fl_pad)) {
632 		/*
633 		 * If there is any chance that we might use buffer packing and
634 		 * the chip is a T4, then pick 64 as the pad/pack boundary.  Set
635 		 * it to the minimum allowed in all other cases.
636 		 */
637 		pad = is_t4(sc) && buffer_packing ? 64 : 1 << pad_shift;
638 
639 		/*
640 		 * For fl_pad = 0 we'll still write a reasonable value to the
641 		 * register but all the freelists will opt out of padding.
642 		 * We'll complain here only if the user tried to set it to a
643 		 * value greater than 0 that was invalid.
644 		 */
645 		if (fl_pad > 0) {
646 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pad value"
647 			    " (%d), using %d instead.\n", fl_pad, pad);
648 		}
649 	}
650 	m = V_INGPADBOUNDARY(M_INGPADBOUNDARY);
651 	v = V_INGPADBOUNDARY(ilog2(pad) - pad_shift);
652 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
653 
654 	if (is_t4(sc)) {
655 		if (fl_pack != -1 && fl_pack != pad) {
656 			/* Complain but carry on. */
657 			device_printf(sc->dev, "hw.cxgbe.fl_pack (%d) ignored,"
658 			    " using %d instead.\n", fl_pack, pad);
659 		}
660 		return;
661 	}
662 
663 	pack = fl_pack;
664 	if (fl_pack < 16 || fl_pack == 32 || fl_pack > 4096 ||
665 	    !powerof2(fl_pack)) {
666 		if (sc->params.pci.mps > MAX_PACK_BOUNDARY)
667 			pack = MAX_PACK_BOUNDARY;
668 		else
669 			pack = max(sc->params.pci.mps, CACHE_LINE_SIZE);
670 		MPASS(powerof2(pack));
671 		if (pack < 16)
672 			pack = 16;
673 		if (pack == 32)
674 			pack = 64;
675 		if (pack > 4096)
676 			pack = 4096;
677 		if (fl_pack != -1) {
678 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pack value"
679 			    " (%d), using %d instead.\n", fl_pack, pack);
680 		}
681 	}
682 	m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY);
683 	if (pack == 16)
684 		v = V_INGPACKBOUNDARY(0);
685 	else
686 		v = V_INGPACKBOUNDARY(ilog2(pack) - 5);
687 
688 	MPASS(!is_t4(sc));	/* T4 doesn't have SGE_CONTROL2 */
689 	t4_set_reg_field(sc, A_SGE_CONTROL2, m, v);
690 }
691 
692 /*
693  * adap->params.vpd.cclk must be set up before this is called.
694  */
695 void
696 t4_tweak_chip_settings(struct adapter *sc)
697 {
698 	int i, reg;
699 	uint32_t v, m;
700 	int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200};
701 	int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk;
702 	int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */
703 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
704 	static int sw_buf_sizes[] = {
705 		MCLBYTES,
706 #if MJUMPAGESIZE != MCLBYTES
707 		MJUMPAGESIZE,
708 #endif
709 		MJUM9BYTES,
710 		MJUM16BYTES
711 	};
712 
713 	KASSERT(sc->flags & MASTER_PF,
714 	    ("%s: trying to change chip settings when not master.", __func__));
715 
716 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
717 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
718 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
719 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
720 
721 	setup_pad_and_pack_boundaries(sc);
722 
723 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
724 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
725 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
726 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
727 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
728 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
729 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
730 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
731 	t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v);
732 
733 	t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0, 4096);
734 	t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE1, 65536);
735 	reg = A_SGE_FL_BUFFER_SIZE2;
736 	for (i = 0; i < nitems(sw_buf_sizes); i++) {
737 		MPASS(reg <= A_SGE_FL_BUFFER_SIZE15);
738 		t4_write_reg(sc, reg, sw_buf_sizes[i]);
739 		reg += 4;
740 		MPASS(reg <= A_SGE_FL_BUFFER_SIZE15);
741 		t4_write_reg(sc, reg, sw_buf_sizes[i] - CL_METADATA_SIZE);
742 		reg += 4;
743 	}
744 
745 	v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) |
746 	    V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]);
747 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v);
748 
749 	KASSERT(intr_timer[0] <= timer_max,
750 	    ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0],
751 	    timer_max));
752 	for (i = 1; i < nitems(intr_timer); i++) {
753 		KASSERT(intr_timer[i] >= intr_timer[i - 1],
754 		    ("%s: timers not listed in increasing order (%d)",
755 		    __func__, i));
756 
757 		while (intr_timer[i] > timer_max) {
758 			if (i == nitems(intr_timer) - 1) {
759 				intr_timer[i] = timer_max;
760 				break;
761 			}
762 			intr_timer[i] += intr_timer[i - 1];
763 			intr_timer[i] /= 2;
764 		}
765 	}
766 
767 	v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) |
768 	    V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1]));
769 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v);
770 	v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) |
771 	    V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3]));
772 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v);
773 	v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) |
774 	    V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5]));
775 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v);
776 
777 	if (chip_id(sc) >= CHELSIO_T6) {
778 		m = V_TSCALE(M_TSCALE);
779 		if (tscale == 1)
780 			v = 0;
781 		else
782 			v = V_TSCALE(tscale - 2);
783 		t4_set_reg_field(sc, A_SGE_ITP_CONTROL, m, v);
784 
785 		if (sc->debug_flags & DF_DISABLE_TCB_CACHE) {
786 			m = V_RDTHRESHOLD(M_RDTHRESHOLD) | F_WRTHRTHRESHEN |
787 			    V_WRTHRTHRESH(M_WRTHRTHRESH);
788 			t4_tp_pio_read(sc, &v, 1, A_TP_CMM_CONFIG, 1);
789 			v &= ~m;
790 			v |= V_RDTHRESHOLD(1) | F_WRTHRTHRESHEN |
791 			    V_WRTHRTHRESH(16);
792 			t4_tp_pio_write(sc, &v, 1, A_TP_CMM_CONFIG, 1);
793 		}
794 	}
795 
796 	/* 4K, 16K, 64K, 256K DDP "page sizes" for TDDP */
797 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
798 	t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v);
799 
800 	/*
801 	 * 4K, 8K, 16K, 64K DDP "page sizes" for iSCSI DDP.  These have been
802 	 * chosen with MAXPHYS = 128K in mind.  The largest DDP buffer that we
803 	 * may have to deal with is MAXPHYS + 1 page.
804 	 */
805 	v = V_HPZ0(0) | V_HPZ1(1) | V_HPZ2(2) | V_HPZ3(4);
806 	t4_write_reg(sc, A_ULP_RX_ISCSI_PSZ, v);
807 
808 	/* We use multiple DDP page sizes both in plain-TOE and ISCSI modes. */
809 	m = v = F_TDDPTAGTCB | F_ISCSITAGTCB;
810 	t4_set_reg_field(sc, A_ULP_RX_CTL, m, v);
811 
812 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
813 	    F_RESETDDPOFFSET;
814 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
815 	t4_set_reg_field(sc, A_TP_PARA_REG5, m, v);
816 }
817 
818 /*
819  * SGE wants the buffer to be at least 64B and then a multiple of 16.  Its
820  * address mut be 16B aligned.  If padding is in use the buffer's start and end
821  * need to be aligned to the pad boundary as well.  We'll just make sure that
822  * the size is a multiple of the pad boundary here, it is up to the buffer
823  * allocation code to make sure the start of the buffer is aligned.
824  */
825 static inline int
826 hwsz_ok(struct adapter *sc, int hwsz)
827 {
828 	int mask = fl_pad ? sc->params.sge.pad_boundary - 1 : 16 - 1;
829 
830 	return (hwsz >= 64 && (hwsz & mask) == 0);
831 }
832 
833 /*
834  * Initialize the rx buffer sizes and figure out which zones the buffers will
835  * be allocated from.
836  */
837 void
838 t4_init_rx_buf_info(struct adapter *sc)
839 {
840 	struct sge *s = &sc->sge;
841 	struct sge_params *sp = &sc->params.sge;
842 	int i, j, n;
843 	static int sw_buf_sizes[] = {	/* Sorted by size */
844 		MCLBYTES,
845 #if MJUMPAGESIZE != MCLBYTES
846 		MJUMPAGESIZE,
847 #endif
848 		MJUM9BYTES,
849 		MJUM16BYTES
850 	};
851 	struct rx_buf_info *rxb;
852 
853 	s->safe_zidx = -1;
854 	rxb = &s->rx_buf_info[0];
855 	for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) {
856 		rxb->size1 = sw_buf_sizes[i];
857 		rxb->zone = m_getzone(rxb->size1);
858 		rxb->type = m_gettype(rxb->size1);
859 		rxb->size2 = 0;
860 		rxb->hwidx1 = -1;
861 		rxb->hwidx2 = -1;
862 		for (j = 0; j < SGE_FLBUF_SIZES; j++) {
863 			int hwsize = sp->sge_fl_buffer_size[j];
864 
865 			if (!hwsz_ok(sc, hwsize))
866 				continue;
867 
868 			/* hwidx for size1 */
869 			if (rxb->hwidx1 == -1 && rxb->size1 == hwsize)
870 				rxb->hwidx1 = j;
871 
872 			/* hwidx for size2 (buffer packing) */
873 			if (rxb->size1 - CL_METADATA_SIZE < hwsize)
874 				continue;
875 			n = rxb->size1 - hwsize - CL_METADATA_SIZE;
876 			if (n == 0) {
877 				rxb->hwidx2 = j;
878 				rxb->size2 = hwsize;
879 				break;	/* stop looking */
880 			}
881 			if (rxb->hwidx2 != -1) {
882 				if (n < sp->sge_fl_buffer_size[rxb->hwidx2] -
883 				    hwsize - CL_METADATA_SIZE) {
884 					rxb->hwidx2 = j;
885 					rxb->size2 = hwsize;
886 				}
887 			} else if (n <= 2 * CL_METADATA_SIZE) {
888 				rxb->hwidx2 = j;
889 				rxb->size2 = hwsize;
890 			}
891 		}
892 		if (rxb->hwidx2 != -1)
893 			sc->flags |= BUF_PACKING_OK;
894 		if (s->safe_zidx == -1 && rxb->size1 == safest_rx_cluster)
895 			s->safe_zidx = i;
896 	}
897 }
898 
899 /*
900  * Verify some basic SGE settings for the PF and VF driver, and other
901  * miscellaneous settings for the PF driver.
902  */
903 int
904 t4_verify_chip_settings(struct adapter *sc)
905 {
906 	struct sge_params *sp = &sc->params.sge;
907 	uint32_t m, v, r;
908 	int rc = 0;
909 	const uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
910 
911 	m = F_RXPKTCPLMODE;
912 	v = F_RXPKTCPLMODE;
913 	r = sp->sge_control;
914 	if ((r & m) != v) {
915 		device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r);
916 		rc = EINVAL;
917 	}
918 
919 	/*
920 	 * If this changes then every single use of PAGE_SHIFT in the driver
921 	 * needs to be carefully reviewed for PAGE_SHIFT vs sp->page_shift.
922 	 */
923 	if (sp->page_shift != PAGE_SHIFT) {
924 		device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r);
925 		rc = EINVAL;
926 	}
927 
928 	if (sc->flags & IS_VF)
929 		return (0);
930 
931 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
932 	r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ);
933 	if (r != v) {
934 		device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r);
935 		if (sc->vres.ddp.size != 0)
936 			rc = EINVAL;
937 	}
938 
939 	m = v = F_TDDPTAGTCB;
940 	r = t4_read_reg(sc, A_ULP_RX_CTL);
941 	if ((r & m) != v) {
942 		device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r);
943 		if (sc->vres.ddp.size != 0)
944 			rc = EINVAL;
945 	}
946 
947 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
948 	    F_RESETDDPOFFSET;
949 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
950 	r = t4_read_reg(sc, A_TP_PARA_REG5);
951 	if ((r & m) != v) {
952 		device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r);
953 		if (sc->vres.ddp.size != 0)
954 			rc = EINVAL;
955 	}
956 
957 	return (rc);
958 }
959 
960 int
961 t4_create_dma_tag(struct adapter *sc)
962 {
963 	int rc;
964 
965 	rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
966 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
967 	    BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
968 	    NULL, &sc->dmat);
969 	if (rc != 0) {
970 		device_printf(sc->dev,
971 		    "failed to create main DMA tag: %d\n", rc);
972 	}
973 
974 	return (rc);
975 }
976 
977 void
978 t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
979     struct sysctl_oid_list *children)
980 {
981 	struct sge_params *sp = &sc->params.sge;
982 
983 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "buffer_sizes",
984 	    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, sc, 0,
985 	    sysctl_bufsizes, "A", "freelist buffer sizes");
986 
987 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD,
988 	    NULL, sp->fl_pktshift, "payload DMA offset in rx buffer (bytes)");
989 
990 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD,
991 	    NULL, sp->pad_boundary, "payload pad boundary (bytes)");
992 
993 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD,
994 	    NULL, sp->spg_len, "status page size (bytes)");
995 
996 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD,
997 	    NULL, cong_drop, "congestion drop setting");
998 
999 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD,
1000 	    NULL, sp->pack_boundary, "payload pack boundary (bytes)");
1001 }
1002 
1003 int
1004 t4_destroy_dma_tag(struct adapter *sc)
1005 {
1006 	if (sc->dmat)
1007 		bus_dma_tag_destroy(sc->dmat);
1008 
1009 	return (0);
1010 }
1011 
1012 /*
1013  * Allocate and initialize the firmware event queue, control queues, and special
1014  * purpose rx queues owned by the adapter.
1015  *
1016  * Returns errno on failure.  Resources allocated up to that point may still be
1017  * allocated.  Caller is responsible for cleanup in case this function fails.
1018  */
1019 int
1020 t4_setup_adapter_queues(struct adapter *sc)
1021 {
1022 	struct sysctl_oid *oid;
1023 	struct sysctl_oid_list *children;
1024 	int rc, i;
1025 
1026 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
1027 
1028 	sysctl_ctx_init(&sc->ctx);
1029 	sc->flags |= ADAP_SYSCTL_CTX;
1030 
1031 	/*
1032 	 * Firmware event queue
1033 	 */
1034 	rc = alloc_fwq(sc);
1035 	if (rc != 0)
1036 		return (rc);
1037 
1038 	/*
1039 	 * That's all for the VF driver.
1040 	 */
1041 	if (sc->flags & IS_VF)
1042 		return (rc);
1043 
1044 	oid = device_get_sysctl_tree(sc->dev);
1045 	children = SYSCTL_CHILDREN(oid);
1046 
1047 	/*
1048 	 * XXX: General purpose rx queues, one per port.
1049 	 */
1050 
1051 	/*
1052 	 * Control queues, one per port.
1053 	 */
1054 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "ctrlq",
1055 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "control queues");
1056 	for_each_port(sc, i) {
1057 		struct sge_wrq *ctrlq = &sc->sge.ctrlq[i];
1058 
1059 		rc = alloc_ctrlq(sc, ctrlq, i, oid);
1060 		if (rc != 0)
1061 			return (rc);
1062 	}
1063 
1064 	return (rc);
1065 }
1066 
1067 /*
1068  * Idempotent
1069  */
1070 int
1071 t4_teardown_adapter_queues(struct adapter *sc)
1072 {
1073 	int i;
1074 
1075 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
1076 
1077 	/* Do this before freeing the queue */
1078 	if (sc->flags & ADAP_SYSCTL_CTX) {
1079 		sysctl_ctx_free(&sc->ctx);
1080 		sc->flags &= ~ADAP_SYSCTL_CTX;
1081 	}
1082 
1083 	if (!(sc->flags & IS_VF)) {
1084 		for_each_port(sc, i)
1085 			free_wrq(sc, &sc->sge.ctrlq[i]);
1086 	}
1087 	free_fwq(sc);
1088 
1089 	return (0);
1090 }
1091 
1092 /* Maximum payload that could arrive with a single iq descriptor. */
1093 static inline int
1094 max_rx_payload(struct adapter *sc, struct ifnet *ifp, const bool ofld)
1095 {
1096 	int maxp;
1097 
1098 	/* large enough even when hw VLAN extraction is disabled */
1099 	maxp = sc->params.sge.fl_pktshift + ETHER_HDR_LEN +
1100 	    ETHER_VLAN_ENCAP_LEN + ifp->if_mtu;
1101 	if (ofld && sc->tt.tls && sc->cryptocaps & FW_CAPS_CONFIG_TLSKEYS &&
1102 	    maxp < sc->params.tp.max_rx_pdu)
1103 		maxp = sc->params.tp.max_rx_pdu;
1104 	return (maxp);
1105 }
1106 
1107 int
1108 t4_setup_vi_queues(struct vi_info *vi)
1109 {
1110 	int rc = 0, i, intr_idx, iqidx;
1111 	struct sge_rxq *rxq;
1112 	struct sge_txq *txq;
1113 #ifdef TCP_OFFLOAD
1114 	struct sge_ofld_rxq *ofld_rxq;
1115 #endif
1116 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1117 	struct sge_ofld_txq *ofld_txq;
1118 #endif
1119 #ifdef DEV_NETMAP
1120 	int saved_idx;
1121 	struct sge_nm_rxq *nm_rxq;
1122 	struct sge_nm_txq *nm_txq;
1123 #endif
1124 	char name[16];
1125 	struct port_info *pi = vi->pi;
1126 	struct adapter *sc = pi->adapter;
1127 	struct ifnet *ifp = vi->ifp;
1128 	struct sysctl_oid *oid = device_get_sysctl_tree(vi->dev);
1129 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
1130 	int maxp;
1131 
1132 	/* Interrupt vector to start from (when using multiple vectors) */
1133 	intr_idx = vi->first_intr;
1134 
1135 #ifdef DEV_NETMAP
1136 	saved_idx = intr_idx;
1137 	if (ifp->if_capabilities & IFCAP_NETMAP) {
1138 
1139 		/* netmap is supported with direct interrupts only. */
1140 		MPASS(!forwarding_intr_to_fwq(sc));
1141 
1142 		/*
1143 		 * We don't have buffers to back the netmap rx queues
1144 		 * right now so we create the queues in a way that
1145 		 * doesn't set off any congestion signal in the chip.
1146 		 */
1147 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_rxq",
1148 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "rx queues");
1149 		for_each_nm_rxq(vi, i, nm_rxq) {
1150 			rc = alloc_nm_rxq(vi, nm_rxq, intr_idx, i, oid);
1151 			if (rc != 0)
1152 				goto done;
1153 			intr_idx++;
1154 		}
1155 
1156 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_txq",
1157 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "tx queues");
1158 		for_each_nm_txq(vi, i, nm_txq) {
1159 			iqidx = vi->first_nm_rxq + (i % vi->nnmrxq);
1160 			rc = alloc_nm_txq(vi, nm_txq, iqidx, i, oid);
1161 			if (rc != 0)
1162 				goto done;
1163 		}
1164 	}
1165 
1166 	/* Normal rx queues and netmap rx queues share the same interrupts. */
1167 	intr_idx = saved_idx;
1168 #endif
1169 
1170 	/*
1171 	 * Allocate rx queues first because a default iqid is required when
1172 	 * creating a tx queue.
1173 	 */
1174 	maxp = max_rx_payload(sc, ifp, false);
1175 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
1176 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "rx queues");
1177 	for_each_rxq(vi, i, rxq) {
1178 
1179 		init_iq(&rxq->iq, sc, vi->tmr_idx, vi->pktc_idx, vi->qsize_rxq);
1180 
1181 		snprintf(name, sizeof(name), "%s rxq%d-fl",
1182 		    device_get_nameunit(vi->dev), i);
1183 		init_fl(sc, &rxq->fl, vi->qsize_rxq / 8, maxp, name);
1184 
1185 		rc = alloc_rxq(vi, rxq,
1186 		    forwarding_intr_to_fwq(sc) ? -1 : intr_idx, i, oid);
1187 		if (rc != 0)
1188 			goto done;
1189 		intr_idx++;
1190 	}
1191 #ifdef DEV_NETMAP
1192 	if (ifp->if_capabilities & IFCAP_NETMAP)
1193 		intr_idx = saved_idx + max(vi->nrxq, vi->nnmrxq);
1194 #endif
1195 #ifdef TCP_OFFLOAD
1196 	maxp = max_rx_payload(sc, ifp, true);
1197 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq",
1198 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "rx queues for offloaded TCP connections");
1199 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1200 
1201 		init_iq(&ofld_rxq->iq, sc, vi->ofld_tmr_idx, vi->ofld_pktc_idx,
1202 		    vi->qsize_rxq);
1203 
1204 		snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
1205 		    device_get_nameunit(vi->dev), i);
1206 		init_fl(sc, &ofld_rxq->fl, vi->qsize_rxq / 8, maxp, name);
1207 
1208 		rc = alloc_ofld_rxq(vi, ofld_rxq,
1209 		    forwarding_intr_to_fwq(sc) ? -1 : intr_idx, i, oid);
1210 		if (rc != 0)
1211 			goto done;
1212 		intr_idx++;
1213 	}
1214 #endif
1215 
1216 	/*
1217 	 * Now the tx queues.
1218 	 */
1219 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "txq",
1220 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "tx queues");
1221 	for_each_txq(vi, i, txq) {
1222 		iqidx = vi->first_rxq + (i % vi->nrxq);
1223 		snprintf(name, sizeof(name), "%s txq%d",
1224 		    device_get_nameunit(vi->dev), i);
1225 		init_eq(sc, &txq->eq, EQ_ETH, vi->qsize_txq, pi->tx_chan,
1226 		    sc->sge.rxq[iqidx].iq.cntxt_id, name);
1227 
1228 		rc = alloc_txq(vi, txq, i, oid);
1229 		if (rc != 0)
1230 			goto done;
1231 	}
1232 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1233 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_txq",
1234 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "tx queues for TOE/ETHOFLD");
1235 	for_each_ofld_txq(vi, i, ofld_txq) {
1236 		snprintf(name, sizeof(name), "%s ofld_txq%d",
1237 		    device_get_nameunit(vi->dev), i);
1238 		if (vi->nofldrxq > 0) {
1239 			iqidx = vi->first_ofld_rxq + (i % vi->nofldrxq);
1240 			init_eq(sc, &ofld_txq->wrq.eq, EQ_OFLD, vi->qsize_txq,
1241 			    pi->tx_chan, sc->sge.ofld_rxq[iqidx].iq.cntxt_id,
1242 			    name);
1243 		} else {
1244 			iqidx = vi->first_rxq + (i % vi->nrxq);
1245 			init_eq(sc, &ofld_txq->wrq.eq, EQ_OFLD, vi->qsize_txq,
1246 			    pi->tx_chan, sc->sge.rxq[iqidx].iq.cntxt_id, name);
1247 		}
1248 
1249 		rc = alloc_ofld_txq(vi, ofld_txq, i, oid);
1250 		if (rc != 0)
1251 			goto done;
1252 	}
1253 #endif
1254 done:
1255 	if (rc)
1256 		t4_teardown_vi_queues(vi);
1257 
1258 	return (rc);
1259 }
1260 
1261 /*
1262  * Idempotent
1263  */
1264 int
1265 t4_teardown_vi_queues(struct vi_info *vi)
1266 {
1267 	int i;
1268 	struct sge_rxq *rxq;
1269 	struct sge_txq *txq;
1270 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1271 	struct sge_ofld_txq *ofld_txq;
1272 #endif
1273 #ifdef TCP_OFFLOAD
1274 	struct sge_ofld_rxq *ofld_rxq;
1275 #endif
1276 #ifdef DEV_NETMAP
1277 	struct sge_nm_rxq *nm_rxq;
1278 	struct sge_nm_txq *nm_txq;
1279 #endif
1280 
1281 	/* Do this before freeing the queues */
1282 	if (vi->flags & VI_SYSCTL_CTX) {
1283 		sysctl_ctx_free(&vi->ctx);
1284 		vi->flags &= ~VI_SYSCTL_CTX;
1285 	}
1286 
1287 #ifdef DEV_NETMAP
1288 	if (vi->ifp->if_capabilities & IFCAP_NETMAP) {
1289 		for_each_nm_txq(vi, i, nm_txq) {
1290 			free_nm_txq(vi, nm_txq);
1291 		}
1292 
1293 		for_each_nm_rxq(vi, i, nm_rxq) {
1294 			free_nm_rxq(vi, nm_rxq);
1295 		}
1296 	}
1297 #endif
1298 
1299 	/*
1300 	 * Take down all the tx queues first, as they reference the rx queues
1301 	 * (for egress updates, etc.).
1302 	 */
1303 
1304 	for_each_txq(vi, i, txq) {
1305 		free_txq(vi, txq);
1306 	}
1307 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1308 	for_each_ofld_txq(vi, i, ofld_txq) {
1309 		free_ofld_txq(vi, ofld_txq);
1310 	}
1311 #endif
1312 
1313 	/*
1314 	 * Then take down the rx queues.
1315 	 */
1316 
1317 	for_each_rxq(vi, i, rxq) {
1318 		free_rxq(vi, rxq);
1319 	}
1320 #ifdef TCP_OFFLOAD
1321 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1322 		free_ofld_rxq(vi, ofld_rxq);
1323 	}
1324 #endif
1325 
1326 	return (0);
1327 }
1328 
1329 /*
1330  * Interrupt handler when the driver is using only 1 interrupt.  This is a very
1331  * unusual scenario.
1332  *
1333  * a) Deals with errors, if any.
1334  * b) Services firmware event queue, which is taking interrupts for all other
1335  *    queues.
1336  */
1337 void
1338 t4_intr_all(void *arg)
1339 {
1340 	struct adapter *sc = arg;
1341 	struct sge_iq *fwq = &sc->sge.fwq;
1342 
1343 	MPASS(sc->intr_count == 1);
1344 
1345 	if (sc->intr_type == INTR_INTX)
1346 		t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
1347 
1348 	t4_intr_err(arg);
1349 	t4_intr_evt(fwq);
1350 }
1351 
1352 /*
1353  * Interrupt handler for errors (installed directly when multiple interrupts are
1354  * being used, or called by t4_intr_all).
1355  */
1356 void
1357 t4_intr_err(void *arg)
1358 {
1359 	struct adapter *sc = arg;
1360 	uint32_t v;
1361 	const bool verbose = (sc->debug_flags & DF_VERBOSE_SLOWINTR) != 0;
1362 
1363 	if (sc->flags & ADAP_ERR)
1364 		return;
1365 
1366 	v = t4_read_reg(sc, MYPF_REG(A_PL_PF_INT_CAUSE));
1367 	if (v & F_PFSW) {
1368 		sc->swintr++;
1369 		t4_write_reg(sc, MYPF_REG(A_PL_PF_INT_CAUSE), v);
1370 	}
1371 
1372 	t4_slow_intr_handler(sc, verbose);
1373 }
1374 
1375 /*
1376  * Interrupt handler for iq-only queues.  The firmware event queue is the only
1377  * such queue right now.
1378  */
1379 void
1380 t4_intr_evt(void *arg)
1381 {
1382 	struct sge_iq *iq = arg;
1383 
1384 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1385 		service_iq(iq, 0);
1386 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1387 	}
1388 }
1389 
1390 /*
1391  * Interrupt handler for iq+fl queues.
1392  */
1393 void
1394 t4_intr(void *arg)
1395 {
1396 	struct sge_iq *iq = arg;
1397 
1398 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1399 		service_iq_fl(iq, 0);
1400 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1401 	}
1402 }
1403 
1404 #ifdef DEV_NETMAP
1405 /*
1406  * Interrupt handler for netmap rx queues.
1407  */
1408 void
1409 t4_nm_intr(void *arg)
1410 {
1411 	struct sge_nm_rxq *nm_rxq = arg;
1412 
1413 	if (atomic_cmpset_int(&nm_rxq->nm_state, NM_ON, NM_BUSY)) {
1414 		service_nm_rxq(nm_rxq);
1415 		(void) atomic_cmpset_int(&nm_rxq->nm_state, NM_BUSY, NM_ON);
1416 	}
1417 }
1418 
1419 /*
1420  * Interrupt handler for vectors shared between NIC and netmap rx queues.
1421  */
1422 void
1423 t4_vi_intr(void *arg)
1424 {
1425 	struct irq *irq = arg;
1426 
1427 	MPASS(irq->nm_rxq != NULL);
1428 	t4_nm_intr(irq->nm_rxq);
1429 
1430 	MPASS(irq->rxq != NULL);
1431 	t4_intr(irq->rxq);
1432 }
1433 #endif
1434 
1435 /*
1436  * Deals with interrupts on an iq-only (no freelist) queue.
1437  */
1438 static int
1439 service_iq(struct sge_iq *iq, int budget)
1440 {
1441 	struct sge_iq *q;
1442 	struct adapter *sc = iq->adapter;
1443 	struct iq_desc *d = &iq->desc[iq->cidx];
1444 	int ndescs = 0, limit;
1445 	int rsp_type;
1446 	uint32_t lq;
1447 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1448 
1449 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1450 	KASSERT((iq->flags & IQ_HAS_FL) == 0,
1451 	    ("%s: called for iq %p with fl (iq->flags 0x%x)", __func__, iq,
1452 	    iq->flags));
1453 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1454 	MPASS((iq->flags & IQ_LRO_ENABLED) == 0);
1455 
1456 	limit = budget ? budget : iq->qsize / 16;
1457 
1458 	/*
1459 	 * We always come back and check the descriptor ring for new indirect
1460 	 * interrupts and other responses after running a single handler.
1461 	 */
1462 	for (;;) {
1463 		while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1464 
1465 			rmb();
1466 
1467 			rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1468 			lq = be32toh(d->rsp.pldbuflen_qid);
1469 
1470 			switch (rsp_type) {
1471 			case X_RSPD_TYPE_FLBUF:
1472 				panic("%s: data for an iq (%p) with no freelist",
1473 				    __func__, iq);
1474 
1475 				/* NOTREACHED */
1476 
1477 			case X_RSPD_TYPE_CPL:
1478 				KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1479 				    ("%s: bad opcode %02x.", __func__,
1480 				    d->rss.opcode));
1481 				t4_cpl_handler[d->rss.opcode](iq, &d->rss, NULL);
1482 				break;
1483 
1484 			case X_RSPD_TYPE_INTR:
1485 				/*
1486 				 * There are 1K interrupt-capable queues (qids 0
1487 				 * through 1023).  A response type indicating a
1488 				 * forwarded interrupt with a qid >= 1K is an
1489 				 * iWARP async notification.
1490 				 */
1491 				if (__predict_true(lq >= 1024)) {
1492 					t4_an_handler(iq, &d->rsp);
1493 					break;
1494 				}
1495 
1496 				q = sc->sge.iqmap[lq - sc->sge.iq_start -
1497 				    sc->sge.iq_base];
1498 				if (atomic_cmpset_int(&q->state, IQS_IDLE,
1499 				    IQS_BUSY)) {
1500 					if (service_iq_fl(q, q->qsize / 16) == 0) {
1501 						(void) atomic_cmpset_int(&q->state,
1502 						    IQS_BUSY, IQS_IDLE);
1503 					} else {
1504 						STAILQ_INSERT_TAIL(&iql, q,
1505 						    link);
1506 					}
1507 				}
1508 				break;
1509 
1510 			default:
1511 				KASSERT(0,
1512 				    ("%s: illegal response type %d on iq %p",
1513 				    __func__, rsp_type, iq));
1514 				log(LOG_ERR,
1515 				    "%s: illegal response type %d on iq %p",
1516 				    device_get_nameunit(sc->dev), rsp_type, iq);
1517 				break;
1518 			}
1519 
1520 			d++;
1521 			if (__predict_false(++iq->cidx == iq->sidx)) {
1522 				iq->cidx = 0;
1523 				iq->gen ^= F_RSPD_GEN;
1524 				d = &iq->desc[0];
1525 			}
1526 			if (__predict_false(++ndescs == limit)) {
1527 				t4_write_reg(sc, sc->sge_gts_reg,
1528 				    V_CIDXINC(ndescs) |
1529 				    V_INGRESSQID(iq->cntxt_id) |
1530 				    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1531 				ndescs = 0;
1532 
1533 				if (budget) {
1534 					return (EINPROGRESS);
1535 				}
1536 			}
1537 		}
1538 
1539 		if (STAILQ_EMPTY(&iql))
1540 			break;
1541 
1542 		/*
1543 		 * Process the head only, and send it to the back of the list if
1544 		 * it's still not done.
1545 		 */
1546 		q = STAILQ_FIRST(&iql);
1547 		STAILQ_REMOVE_HEAD(&iql, link);
1548 		if (service_iq_fl(q, q->qsize / 8) == 0)
1549 			(void) atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1550 		else
1551 			STAILQ_INSERT_TAIL(&iql, q, link);
1552 	}
1553 
1554 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1555 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1556 
1557 	return (0);
1558 }
1559 
1560 static inline int
1561 sort_before_lro(struct lro_ctrl *lro)
1562 {
1563 
1564 	return (lro->lro_mbuf_max != 0);
1565 }
1566 
1567 static inline uint64_t
1568 last_flit_to_ns(struct adapter *sc, uint64_t lf)
1569 {
1570 	uint64_t n = be64toh(lf) & 0xfffffffffffffff;	/* 60b, not 64b. */
1571 
1572 	if (n > UINT64_MAX / 1000000)
1573 		return (n / sc->params.vpd.cclk * 1000000);
1574 	else
1575 		return (n * 1000000 / sc->params.vpd.cclk);
1576 }
1577 
1578 static inline void
1579 move_to_next_rxbuf(struct sge_fl *fl)
1580 {
1581 
1582 	fl->rx_offset = 0;
1583 	if (__predict_false((++fl->cidx & 7) == 0)) {
1584 		uint16_t cidx = fl->cidx >> 3;
1585 
1586 		if (__predict_false(cidx == fl->sidx))
1587 			fl->cidx = cidx = 0;
1588 		fl->hw_cidx = cidx;
1589 	}
1590 }
1591 
1592 /*
1593  * Deals with interrupts on an iq+fl queue.
1594  */
1595 static int
1596 service_iq_fl(struct sge_iq *iq, int budget)
1597 {
1598 	struct sge_rxq *rxq = iq_to_rxq(iq);
1599 	struct sge_fl *fl;
1600 	struct adapter *sc = iq->adapter;
1601 	struct iq_desc *d = &iq->desc[iq->cidx];
1602 	int ndescs, limit;
1603 	int rsp_type, starved;
1604 	uint32_t lq;
1605 	uint16_t fl_hw_cidx;
1606 	struct mbuf *m0;
1607 #if defined(INET) || defined(INET6)
1608 	const struct timeval lro_timeout = {0, sc->lro_timeout};
1609 	struct lro_ctrl *lro = &rxq->lro;
1610 #endif
1611 
1612 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1613 	MPASS(iq->flags & IQ_HAS_FL);
1614 
1615 	ndescs = 0;
1616 #if defined(INET) || defined(INET6)
1617 	if (iq->flags & IQ_ADJ_CREDIT) {
1618 		MPASS(sort_before_lro(lro));
1619 		iq->flags &= ~IQ_ADJ_CREDIT;
1620 		if ((d->rsp.u.type_gen & F_RSPD_GEN) != iq->gen) {
1621 			tcp_lro_flush_all(lro);
1622 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(1) |
1623 			    V_INGRESSQID((u32)iq->cntxt_id) |
1624 			    V_SEINTARM(iq->intr_params));
1625 			return (0);
1626 		}
1627 		ndescs = 1;
1628 	}
1629 #else
1630 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1631 #endif
1632 
1633 	limit = budget ? budget : iq->qsize / 16;
1634 	fl = &rxq->fl;
1635 	fl_hw_cidx = fl->hw_cidx;	/* stable snapshot */
1636 	while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1637 
1638 		rmb();
1639 
1640 		m0 = NULL;
1641 		rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1642 		lq = be32toh(d->rsp.pldbuflen_qid);
1643 
1644 		switch (rsp_type) {
1645 		case X_RSPD_TYPE_FLBUF:
1646 			if (lq & F_RSPD_NEWBUF) {
1647 				if (fl->rx_offset > 0)
1648 					move_to_next_rxbuf(fl);
1649 				lq = G_RSPD_LEN(lq);
1650 			}
1651 			if (IDXDIFF(fl->hw_cidx, fl_hw_cidx, fl->sidx) > 4) {
1652 				FL_LOCK(fl);
1653 				refill_fl(sc, fl, 64);
1654 				FL_UNLOCK(fl);
1655 				fl_hw_cidx = fl->hw_cidx;
1656 			}
1657 
1658 			if (d->rss.opcode == CPL_RX_PKT) {
1659 				if (__predict_true(eth_rx(sc, rxq, d, lq) == 0))
1660 					break;
1661 				goto out;
1662 			}
1663 			m0 = get_fl_payload(sc, fl, lq);
1664 			if (__predict_false(m0 == NULL))
1665 				goto out;
1666 
1667 			/* fall through */
1668 
1669 		case X_RSPD_TYPE_CPL:
1670 			KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1671 			    ("%s: bad opcode %02x.", __func__, d->rss.opcode));
1672 			t4_cpl_handler[d->rss.opcode](iq, &d->rss, m0);
1673 			break;
1674 
1675 		case X_RSPD_TYPE_INTR:
1676 
1677 			/*
1678 			 * There are 1K interrupt-capable queues (qids 0
1679 			 * through 1023).  A response type indicating a
1680 			 * forwarded interrupt with a qid >= 1K is an
1681 			 * iWARP async notification.  That is the only
1682 			 * acceptable indirect interrupt on this queue.
1683 			 */
1684 			if (__predict_false(lq < 1024)) {
1685 				panic("%s: indirect interrupt on iq_fl %p "
1686 				    "with qid %u", __func__, iq, lq);
1687 			}
1688 
1689 			t4_an_handler(iq, &d->rsp);
1690 			break;
1691 
1692 		default:
1693 			KASSERT(0, ("%s: illegal response type %d on iq %p",
1694 			    __func__, rsp_type, iq));
1695 			log(LOG_ERR, "%s: illegal response type %d on iq %p",
1696 			    device_get_nameunit(sc->dev), rsp_type, iq);
1697 			break;
1698 		}
1699 
1700 		d++;
1701 		if (__predict_false(++iq->cidx == iq->sidx)) {
1702 			iq->cidx = 0;
1703 			iq->gen ^= F_RSPD_GEN;
1704 			d = &iq->desc[0];
1705 		}
1706 		if (__predict_false(++ndescs == limit)) {
1707 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1708 			    V_INGRESSQID(iq->cntxt_id) |
1709 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1710 
1711 #if defined(INET) || defined(INET6)
1712 			if (iq->flags & IQ_LRO_ENABLED &&
1713 			    !sort_before_lro(lro) &&
1714 			    sc->lro_timeout != 0) {
1715 				tcp_lro_flush_inactive(lro, &lro_timeout);
1716 			}
1717 #endif
1718 			if (budget)
1719 				return (EINPROGRESS);
1720 			ndescs = 0;
1721 		}
1722 	}
1723 out:
1724 #if defined(INET) || defined(INET6)
1725 	if (iq->flags & IQ_LRO_ENABLED) {
1726 		if (ndescs > 0 && lro->lro_mbuf_count > 8) {
1727 			MPASS(sort_before_lro(lro));
1728 			/* hold back one credit and don't flush LRO state */
1729 			iq->flags |= IQ_ADJ_CREDIT;
1730 			ndescs--;
1731 		} else {
1732 			tcp_lro_flush_all(lro);
1733 		}
1734 	}
1735 #endif
1736 
1737 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1738 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1739 
1740 	FL_LOCK(fl);
1741 	starved = refill_fl(sc, fl, 64);
1742 	FL_UNLOCK(fl);
1743 	if (__predict_false(starved != 0))
1744 		add_fl_to_sfl(sc, fl);
1745 
1746 	return (0);
1747 }
1748 
1749 static inline struct cluster_metadata *
1750 cl_metadata(struct fl_sdesc *sd)
1751 {
1752 
1753 	return ((void *)(sd->cl + sd->moff));
1754 }
1755 
1756 static void
1757 rxb_free(struct mbuf *m)
1758 {
1759 	struct cluster_metadata *clm = m->m_ext.ext_arg1;
1760 
1761 	uma_zfree(clm->zone, clm->cl);
1762 	counter_u64_add(extfree_rels, 1);
1763 }
1764 
1765 /*
1766  * The mbuf returned comes from zone_muf and carries the payload in one of these
1767  * ways
1768  * a) complete frame inside the mbuf
1769  * b) m_cljset (for clusters without metadata)
1770  * d) m_extaddref (cluster with metadata)
1771  */
1772 static struct mbuf *
1773 get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1774     int remaining)
1775 {
1776 	struct mbuf *m;
1777 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1778 	struct rx_buf_info *rxb = &sc->sge.rx_buf_info[sd->zidx];
1779 	struct cluster_metadata *clm;
1780 	int len, blen;
1781 	caddr_t payload;
1782 
1783 	if (fl->flags & FL_BUF_PACKING) {
1784 		u_int l, pad;
1785 
1786 		blen = rxb->size2 - fl->rx_offset;	/* max possible in this buf */
1787 		len = min(remaining, blen);
1788 		payload = sd->cl + fl->rx_offset;
1789 
1790 		l = fr_offset + len;
1791 		pad = roundup2(l, fl->buf_boundary) - l;
1792 		if (fl->rx_offset + len + pad < rxb->size2)
1793 			blen = len + pad;
1794 		MPASS(fl->rx_offset + blen <= rxb->size2);
1795 	} else {
1796 		MPASS(fl->rx_offset == 0);	/* not packing */
1797 		blen = rxb->size1;
1798 		len = min(remaining, blen);
1799 		payload = sd->cl;
1800 	}
1801 
1802 	if (fr_offset == 0) {
1803 		m = m_gethdr(M_NOWAIT, MT_DATA);
1804 		if (__predict_false(m == NULL))
1805 			return (NULL);
1806 		m->m_pkthdr.len = remaining;
1807 	} else {
1808 		m = m_get(M_NOWAIT, MT_DATA);
1809 		if (__predict_false(m == NULL))
1810 			return (NULL);
1811 	}
1812 	m->m_len = len;
1813 
1814 	if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) {
1815 		/* copy data to mbuf */
1816 		bcopy(payload, mtod(m, caddr_t), len);
1817 		if (fl->flags & FL_BUF_PACKING) {
1818 			fl->rx_offset += blen;
1819 			MPASS(fl->rx_offset <= rxb->size2);
1820 			if (fl->rx_offset < rxb->size2)
1821 				return (m);	/* without advancing the cidx */
1822 		}
1823 	} else if (fl->flags & FL_BUF_PACKING) {
1824 		clm = cl_metadata(sd);
1825 		if (sd->nmbuf++ == 0) {
1826 			clm->refcount = 1;
1827 			clm->zone = rxb->zone;
1828 			clm->cl = sd->cl;
1829 			counter_u64_add(extfree_refs, 1);
1830 		}
1831 		m_extaddref(m, payload, blen, &clm->refcount, rxb_free, clm,
1832 		    NULL);
1833 
1834 		fl->rx_offset += blen;
1835 		MPASS(fl->rx_offset <= rxb->size2);
1836 		if (fl->rx_offset < rxb->size2)
1837 			return (m);	/* without advancing the cidx */
1838 	} else {
1839 		m_cljset(m, sd->cl, rxb->type);
1840 		sd->cl = NULL;	/* consumed, not a recycle candidate */
1841 	}
1842 
1843 	move_to_next_rxbuf(fl);
1844 
1845 	return (m);
1846 }
1847 
1848 static struct mbuf *
1849 get_fl_payload(struct adapter *sc, struct sge_fl *fl, const u_int plen)
1850 {
1851 	struct mbuf *m0, *m, **pnext;
1852 	u_int remaining;
1853 
1854 	if (__predict_false(fl->flags & FL_BUF_RESUME)) {
1855 		M_ASSERTPKTHDR(fl->m0);
1856 		MPASS(fl->m0->m_pkthdr.len == plen);
1857 		MPASS(fl->remaining < plen);
1858 
1859 		m0 = fl->m0;
1860 		pnext = fl->pnext;
1861 		remaining = fl->remaining;
1862 		fl->flags &= ~FL_BUF_RESUME;
1863 		goto get_segment;
1864 	}
1865 
1866 	/*
1867 	 * Payload starts at rx_offset in the current hw buffer.  Its length is
1868 	 * 'len' and it may span multiple hw buffers.
1869 	 */
1870 
1871 	m0 = get_scatter_segment(sc, fl, 0, plen);
1872 	if (m0 == NULL)
1873 		return (NULL);
1874 	remaining = plen - m0->m_len;
1875 	pnext = &m0->m_next;
1876 	while (remaining > 0) {
1877 get_segment:
1878 		MPASS(fl->rx_offset == 0);
1879 		m = get_scatter_segment(sc, fl, plen - remaining, remaining);
1880 		if (__predict_false(m == NULL)) {
1881 			fl->m0 = m0;
1882 			fl->pnext = pnext;
1883 			fl->remaining = remaining;
1884 			fl->flags |= FL_BUF_RESUME;
1885 			return (NULL);
1886 		}
1887 		*pnext = m;
1888 		pnext = &m->m_next;
1889 		remaining -= m->m_len;
1890 	}
1891 	*pnext = NULL;
1892 
1893 	M_ASSERTPKTHDR(m0);
1894 	return (m0);
1895 }
1896 
1897 static int
1898 skip_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1899     int remaining)
1900 {
1901 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1902 	struct rx_buf_info *rxb = &sc->sge.rx_buf_info[sd->zidx];
1903 	int len, blen;
1904 
1905 	if (fl->flags & FL_BUF_PACKING) {
1906 		u_int l, pad;
1907 
1908 		blen = rxb->size2 - fl->rx_offset;	/* max possible in this buf */
1909 		len = min(remaining, blen);
1910 
1911 		l = fr_offset + len;
1912 		pad = roundup2(l, fl->buf_boundary) - l;
1913 		if (fl->rx_offset + len + pad < rxb->size2)
1914 			blen = len + pad;
1915 		fl->rx_offset += blen;
1916 		MPASS(fl->rx_offset <= rxb->size2);
1917 		if (fl->rx_offset < rxb->size2)
1918 			return (len);	/* without advancing the cidx */
1919 	} else {
1920 		MPASS(fl->rx_offset == 0);	/* not packing */
1921 		blen = rxb->size1;
1922 		len = min(remaining, blen);
1923 	}
1924 	move_to_next_rxbuf(fl);
1925 	return (len);
1926 }
1927 
1928 static inline void
1929 skip_fl_payload(struct adapter *sc, struct sge_fl *fl, int plen)
1930 {
1931 	int remaining, fr_offset, len;
1932 
1933 	fr_offset = 0;
1934 	remaining = plen;
1935 	while (remaining > 0) {
1936 		len = skip_scatter_segment(sc, fl, fr_offset, remaining);
1937 		fr_offset += len;
1938 		remaining -= len;
1939 	}
1940 }
1941 
1942 static inline int
1943 get_segment_len(struct adapter *sc, struct sge_fl *fl, int plen)
1944 {
1945 	int len;
1946 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1947 	struct rx_buf_info *rxb = &sc->sge.rx_buf_info[sd->zidx];
1948 
1949 	if (fl->flags & FL_BUF_PACKING)
1950 		len = rxb->size2 - fl->rx_offset;
1951 	else
1952 		len = rxb->size1;
1953 
1954 	return (min(plen, len));
1955 }
1956 
1957 static int
1958 eth_rx(struct adapter *sc, struct sge_rxq *rxq, const struct iq_desc *d,
1959     u_int plen)
1960 {
1961 	struct mbuf *m0;
1962 	struct ifnet *ifp = rxq->ifp;
1963 	struct sge_fl *fl = &rxq->fl;
1964 	struct vi_info *vi = ifp->if_softc;
1965 	const struct cpl_rx_pkt *cpl;
1966 #if defined(INET) || defined(INET6)
1967 	struct lro_ctrl *lro = &rxq->lro;
1968 #endif
1969 	uint16_t err_vec, tnl_type, tnlhdr_len;
1970 	static const int sw_hashtype[4][2] = {
1971 		{M_HASHTYPE_NONE, M_HASHTYPE_NONE},
1972 		{M_HASHTYPE_RSS_IPV4, M_HASHTYPE_RSS_IPV6},
1973 		{M_HASHTYPE_RSS_TCP_IPV4, M_HASHTYPE_RSS_TCP_IPV6},
1974 		{M_HASHTYPE_RSS_UDP_IPV4, M_HASHTYPE_RSS_UDP_IPV6},
1975 	};
1976 	static const int sw_csum_flags[2][2] = {
1977 		{
1978 			/* IP, inner IP */
1979 			CSUM_ENCAP_VXLAN |
1980 			    CSUM_L3_CALC | CSUM_L3_VALID |
1981 			    CSUM_L4_CALC | CSUM_L4_VALID |
1982 			    CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID |
1983 			    CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID,
1984 
1985 			/* IP, inner IP6 */
1986 			CSUM_ENCAP_VXLAN |
1987 			    CSUM_L3_CALC | CSUM_L3_VALID |
1988 			    CSUM_L4_CALC | CSUM_L4_VALID |
1989 			    CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID,
1990 		},
1991 		{
1992 			/* IP6, inner IP */
1993 			CSUM_ENCAP_VXLAN |
1994 			    CSUM_L4_CALC | CSUM_L4_VALID |
1995 			    CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID |
1996 			    CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID,
1997 
1998 			/* IP6, inner IP6 */
1999 			CSUM_ENCAP_VXLAN |
2000 			    CSUM_L4_CALC | CSUM_L4_VALID |
2001 			    CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID,
2002 		},
2003 	};
2004 
2005 	MPASS(plen > sc->params.sge.fl_pktshift);
2006 	if (vi->pfil != NULL && PFIL_HOOKED_IN(vi->pfil) &&
2007 	    __predict_true((fl->flags & FL_BUF_RESUME) == 0)) {
2008 		struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
2009 		caddr_t frame;
2010 		int rc, slen;
2011 
2012 		slen = get_segment_len(sc, fl, plen) -
2013 		    sc->params.sge.fl_pktshift;
2014 		frame = sd->cl + fl->rx_offset + sc->params.sge.fl_pktshift;
2015 		CURVNET_SET_QUIET(ifp->if_vnet);
2016 		rc = pfil_run_hooks(vi->pfil, frame, ifp,
2017 		    slen | PFIL_MEMPTR | PFIL_IN, NULL);
2018 		CURVNET_RESTORE();
2019 		if (rc == PFIL_DROPPED || rc == PFIL_CONSUMED) {
2020 			skip_fl_payload(sc, fl, plen);
2021 			return (0);
2022 		}
2023 		if (rc == PFIL_REALLOCED) {
2024 			skip_fl_payload(sc, fl, plen);
2025 			m0 = pfil_mem2mbuf(frame);
2026 			goto have_mbuf;
2027 		}
2028 	}
2029 
2030 	m0 = get_fl_payload(sc, fl, plen);
2031 	if (__predict_false(m0 == NULL))
2032 		return (ENOMEM);
2033 
2034 	m0->m_pkthdr.len -= sc->params.sge.fl_pktshift;
2035 	m0->m_len -= sc->params.sge.fl_pktshift;
2036 	m0->m_data += sc->params.sge.fl_pktshift;
2037 
2038 have_mbuf:
2039 	m0->m_pkthdr.rcvif = ifp;
2040 	M_HASHTYPE_SET(m0, sw_hashtype[d->rss.hash_type][d->rss.ipv6]);
2041 	m0->m_pkthdr.flowid = be32toh(d->rss.hash_val);
2042 
2043 	cpl = (const void *)(&d->rss + 1);
2044 	if (sc->params.tp.rx_pkt_encap) {
2045 		const uint16_t ev = be16toh(cpl->err_vec);
2046 
2047 		err_vec = G_T6_COMPR_RXERR_VEC(ev);
2048 		tnl_type = G_T6_RX_TNL_TYPE(ev);
2049 		tnlhdr_len = G_T6_RX_TNLHDR_LEN(ev);
2050 	} else {
2051 		err_vec = be16toh(cpl->err_vec);
2052 		tnl_type = 0;
2053 		tnlhdr_len = 0;
2054 	}
2055 	if (cpl->csum_calc && err_vec == 0) {
2056 		int ipv6 = !!(cpl->l2info & htobe32(F_RXF_IP6));
2057 
2058 		/* checksum(s) calculated and found to be correct. */
2059 
2060 		MPASS((cpl->l2info & htobe32(F_RXF_IP)) ^
2061 		    (cpl->l2info & htobe32(F_RXF_IP6)));
2062 		m0->m_pkthdr.csum_data = be16toh(cpl->csum);
2063 		if (tnl_type == 0) {
2064 	    		if (!ipv6 && ifp->if_capenable & IFCAP_RXCSUM) {
2065 				m0->m_pkthdr.csum_flags = CSUM_L3_CALC |
2066 				    CSUM_L3_VALID | CSUM_L4_CALC |
2067 				    CSUM_L4_VALID;
2068 			} else if (ipv6 && ifp->if_capenable & IFCAP_RXCSUM_IPV6) {
2069 				m0->m_pkthdr.csum_flags = CSUM_L4_CALC |
2070 				    CSUM_L4_VALID;
2071 			}
2072 			rxq->rxcsum++;
2073 		} else {
2074 			MPASS(tnl_type == RX_PKT_TNL_TYPE_VXLAN);
2075 
2076 			M_HASHTYPE_SETINNER(m0);
2077 			if (__predict_false(cpl->ip_frag)) {
2078 				/*
2079 				 * csum_data is for the inner frame (which is an
2080 				 * IP fragment) and is not 0xffff.  There is no
2081 				 * way to pass the inner csum_data to the stack.
2082 				 * We don't want the stack to use the inner
2083 				 * csum_data to validate the outer frame or it
2084 				 * will get rejected.  So we fix csum_data here
2085 				 * and let sw do the checksum of inner IP
2086 				 * fragments.
2087 				 *
2088 				 * XXX: Need 32b for csum_data2 in an rx mbuf.
2089 				 * Maybe stuff it into rcv_tstmp?
2090 				 */
2091 				m0->m_pkthdr.csum_data = 0xffff;
2092 				if (ipv6) {
2093 					m0->m_pkthdr.csum_flags = CSUM_L4_CALC |
2094 					    CSUM_L4_VALID;
2095 				} else {
2096 					m0->m_pkthdr.csum_flags = CSUM_L3_CALC |
2097 					    CSUM_L3_VALID | CSUM_L4_CALC |
2098 					    CSUM_L4_VALID;
2099 				}
2100 			} else {
2101 				int outer_ipv6;
2102 
2103 				MPASS(m0->m_pkthdr.csum_data == 0xffff);
2104 
2105 				outer_ipv6 = tnlhdr_len >=
2106 				    sizeof(struct ether_header) +
2107 				    sizeof(struct ip6_hdr);
2108 				m0->m_pkthdr.csum_flags =
2109 				    sw_csum_flags[outer_ipv6][ipv6];
2110 			}
2111 			rxq->vxlan_rxcsum++;
2112 		}
2113 	}
2114 
2115 	if (cpl->vlan_ex) {
2116 		m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
2117 		m0->m_flags |= M_VLANTAG;
2118 		rxq->vlan_extraction++;
2119 	}
2120 
2121 	if (rxq->iq.flags & IQ_RX_TIMESTAMP) {
2122 		/*
2123 		 * Fill up rcv_tstmp but do not set M_TSTMP.
2124 		 * rcv_tstmp is not in the format that the
2125 		 * kernel expects and we don't want to mislead
2126 		 * it.  For now this is only for custom code
2127 		 * that knows how to interpret cxgbe's stamp.
2128 		 */
2129 		m0->m_pkthdr.rcv_tstmp =
2130 		    last_flit_to_ns(sc, d->rsp.u.last_flit);
2131 #ifdef notyet
2132 		m0->m_flags |= M_TSTMP;
2133 #endif
2134 	}
2135 
2136 #ifdef NUMA
2137 	m0->m_pkthdr.numa_domain = ifp->if_numa_domain;
2138 #endif
2139 #if defined(INET) || defined(INET6)
2140 	if (rxq->iq.flags & IQ_LRO_ENABLED && tnl_type == 0 &&
2141 	    (M_HASHTYPE_GET(m0) == M_HASHTYPE_RSS_TCP_IPV4 ||
2142 	    M_HASHTYPE_GET(m0) == M_HASHTYPE_RSS_TCP_IPV6)) {
2143 		if (sort_before_lro(lro)) {
2144 			tcp_lro_queue_mbuf(lro, m0);
2145 			return (0); /* queued for sort, then LRO */
2146 		}
2147 		if (tcp_lro_rx(lro, m0, 0) == 0)
2148 			return (0); /* queued for LRO */
2149 	}
2150 #endif
2151 	ifp->if_input(ifp, m0);
2152 
2153 	return (0);
2154 }
2155 
2156 /*
2157  * Must drain the wrq or make sure that someone else will.
2158  */
2159 static void
2160 wrq_tx_drain(void *arg, int n)
2161 {
2162 	struct sge_wrq *wrq = arg;
2163 	struct sge_eq *eq = &wrq->eq;
2164 
2165 	EQ_LOCK(eq);
2166 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2167 		drain_wrq_wr_list(wrq->adapter, wrq);
2168 	EQ_UNLOCK(eq);
2169 }
2170 
2171 static void
2172 drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq)
2173 {
2174 	struct sge_eq *eq = &wrq->eq;
2175 	u_int available, dbdiff;	/* # of hardware descriptors */
2176 	u_int n;
2177 	struct wrqe *wr;
2178 	struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
2179 
2180 	EQ_LOCK_ASSERT_OWNED(eq);
2181 	MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
2182 	wr = STAILQ_FIRST(&wrq->wr_list);
2183 	MPASS(wr != NULL);	/* Must be called with something useful to do */
2184 	MPASS(eq->pidx == eq->dbidx);
2185 	dbdiff = 0;
2186 
2187 	do {
2188 		eq->cidx = read_hw_cidx(eq);
2189 		if (eq->pidx == eq->cidx)
2190 			available = eq->sidx - 1;
2191 		else
2192 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2193 
2194 		MPASS(wr->wrq == wrq);
2195 		n = howmany(wr->wr_len, EQ_ESIZE);
2196 		if (available < n)
2197 			break;
2198 
2199 		dst = (void *)&eq->desc[eq->pidx];
2200 		if (__predict_true(eq->sidx - eq->pidx > n)) {
2201 			/* Won't wrap, won't end exactly at the status page. */
2202 			bcopy(&wr->wr[0], dst, wr->wr_len);
2203 			eq->pidx += n;
2204 		} else {
2205 			int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE;
2206 
2207 			bcopy(&wr->wr[0], dst, first_portion);
2208 			if (wr->wr_len > first_portion) {
2209 				bcopy(&wr->wr[first_portion], &eq->desc[0],
2210 				    wr->wr_len - first_portion);
2211 			}
2212 			eq->pidx = n - (eq->sidx - eq->pidx);
2213 		}
2214 		wrq->tx_wrs_copied++;
2215 
2216 		if (available < eq->sidx / 4 &&
2217 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2218 				/*
2219 				 * XXX: This is not 100% reliable with some
2220 				 * types of WRs.  But this is a very unusual
2221 				 * situation for an ofld/ctrl queue anyway.
2222 				 */
2223 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2224 			    F_FW_WR_EQUEQ);
2225 		}
2226 
2227 		dbdiff += n;
2228 		if (dbdiff >= 16) {
2229 			ring_eq_db(sc, eq, dbdiff);
2230 			dbdiff = 0;
2231 		}
2232 
2233 		STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
2234 		free_wrqe(wr);
2235 		MPASS(wrq->nwr_pending > 0);
2236 		wrq->nwr_pending--;
2237 		MPASS(wrq->ndesc_needed >= n);
2238 		wrq->ndesc_needed -= n;
2239 	} while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL);
2240 
2241 	if (dbdiff)
2242 		ring_eq_db(sc, eq, dbdiff);
2243 }
2244 
2245 /*
2246  * Doesn't fail.  Holds on to work requests it can't send right away.
2247  */
2248 void
2249 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
2250 {
2251 #ifdef INVARIANTS
2252 	struct sge_eq *eq = &wrq->eq;
2253 #endif
2254 
2255 	EQ_LOCK_ASSERT_OWNED(eq);
2256 	MPASS(wr != NULL);
2257 	MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN);
2258 	MPASS((wr->wr_len & 0x7) == 0);
2259 
2260 	STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
2261 	wrq->nwr_pending++;
2262 	wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE);
2263 
2264 	if (!TAILQ_EMPTY(&wrq->incomplete_wrs))
2265 		return;	/* commit_wrq_wr will drain wr_list as well. */
2266 
2267 	drain_wrq_wr_list(sc, wrq);
2268 
2269 	/* Doorbell must have caught up to the pidx. */
2270 	MPASS(eq->pidx == eq->dbidx);
2271 }
2272 
2273 void
2274 t4_update_fl_bufsize(struct ifnet *ifp)
2275 {
2276 	struct vi_info *vi = ifp->if_softc;
2277 	struct adapter *sc = vi->adapter;
2278 	struct sge_rxq *rxq;
2279 #ifdef TCP_OFFLOAD
2280 	struct sge_ofld_rxq *ofld_rxq;
2281 #endif
2282 	struct sge_fl *fl;
2283 	int i, maxp;
2284 
2285 	maxp = max_rx_payload(sc, ifp, false);
2286 	for_each_rxq(vi, i, rxq) {
2287 		fl = &rxq->fl;
2288 
2289 		FL_LOCK(fl);
2290 		fl->zidx = find_refill_source(sc, maxp,
2291 		    fl->flags & FL_BUF_PACKING);
2292 		FL_UNLOCK(fl);
2293 	}
2294 #ifdef TCP_OFFLOAD
2295 	maxp = max_rx_payload(sc, ifp, true);
2296 	for_each_ofld_rxq(vi, i, ofld_rxq) {
2297 		fl = &ofld_rxq->fl;
2298 
2299 		FL_LOCK(fl);
2300 		fl->zidx = find_refill_source(sc, maxp,
2301 		    fl->flags & FL_BUF_PACKING);
2302 		FL_UNLOCK(fl);
2303 	}
2304 #endif
2305 }
2306 
2307 static inline int
2308 mbuf_nsegs(struct mbuf *m)
2309 {
2310 
2311 	M_ASSERTPKTHDR(m);
2312 	KASSERT(m->m_pkthdr.inner_l5hlen > 0,
2313 	    ("%s: mbuf %p missing information on # of segments.", __func__, m));
2314 
2315 	return (m->m_pkthdr.inner_l5hlen);
2316 }
2317 
2318 static inline void
2319 set_mbuf_nsegs(struct mbuf *m, uint8_t nsegs)
2320 {
2321 
2322 	M_ASSERTPKTHDR(m);
2323 	m->m_pkthdr.inner_l5hlen = nsegs;
2324 }
2325 
2326 static inline int
2327 mbuf_cflags(struct mbuf *m)
2328 {
2329 
2330 	M_ASSERTPKTHDR(m);
2331 	return (m->m_pkthdr.PH_loc.eight[4]);
2332 }
2333 
2334 static inline void
2335 set_mbuf_cflags(struct mbuf *m, uint8_t flags)
2336 {
2337 
2338 	M_ASSERTPKTHDR(m);
2339 	m->m_pkthdr.PH_loc.eight[4] = flags;
2340 }
2341 
2342 static inline int
2343 mbuf_len16(struct mbuf *m)
2344 {
2345 	int n;
2346 
2347 	M_ASSERTPKTHDR(m);
2348 	n = m->m_pkthdr.PH_loc.eight[0];
2349 	if (!(mbuf_cflags(m) & MC_TLS))
2350 		MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2351 
2352 	return (n);
2353 }
2354 
2355 static inline void
2356 set_mbuf_len16(struct mbuf *m, uint8_t len16)
2357 {
2358 
2359 	M_ASSERTPKTHDR(m);
2360 	if (!(mbuf_cflags(m) & MC_TLS))
2361 		MPASS(len16 > 0 && len16 <= SGE_MAX_WR_LEN / 16);
2362 	m->m_pkthdr.PH_loc.eight[0] = len16;
2363 }
2364 
2365 #ifdef RATELIMIT
2366 static inline int
2367 mbuf_eo_nsegs(struct mbuf *m)
2368 {
2369 
2370 	M_ASSERTPKTHDR(m);
2371 	return (m->m_pkthdr.PH_loc.eight[1]);
2372 }
2373 
2374 static inline void
2375 set_mbuf_eo_nsegs(struct mbuf *m, uint8_t nsegs)
2376 {
2377 
2378 	M_ASSERTPKTHDR(m);
2379 	m->m_pkthdr.PH_loc.eight[1] = nsegs;
2380 }
2381 
2382 static inline int
2383 mbuf_eo_len16(struct mbuf *m)
2384 {
2385 	int n;
2386 
2387 	M_ASSERTPKTHDR(m);
2388 	n = m->m_pkthdr.PH_loc.eight[2];
2389 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2390 
2391 	return (n);
2392 }
2393 
2394 static inline void
2395 set_mbuf_eo_len16(struct mbuf *m, uint8_t len16)
2396 {
2397 
2398 	M_ASSERTPKTHDR(m);
2399 	m->m_pkthdr.PH_loc.eight[2] = len16;
2400 }
2401 
2402 static inline int
2403 mbuf_eo_tsclk_tsoff(struct mbuf *m)
2404 {
2405 
2406 	M_ASSERTPKTHDR(m);
2407 	return (m->m_pkthdr.PH_loc.eight[3]);
2408 }
2409 
2410 static inline void
2411 set_mbuf_eo_tsclk_tsoff(struct mbuf *m, uint8_t tsclk_tsoff)
2412 {
2413 
2414 	M_ASSERTPKTHDR(m);
2415 	m->m_pkthdr.PH_loc.eight[3] = tsclk_tsoff;
2416 }
2417 
2418 static inline int
2419 needs_eo(struct m_snd_tag *mst)
2420 {
2421 
2422 	return (mst != NULL && mst->type == IF_SND_TAG_TYPE_RATE_LIMIT);
2423 }
2424 #endif
2425 
2426 /*
2427  * Try to allocate an mbuf to contain a raw work request.  To make it
2428  * easy to construct the work request, don't allocate a chain but a
2429  * single mbuf.
2430  */
2431 struct mbuf *
2432 alloc_wr_mbuf(int len, int how)
2433 {
2434 	struct mbuf *m;
2435 
2436 	if (len <= MHLEN)
2437 		m = m_gethdr(how, MT_DATA);
2438 	else if (len <= MCLBYTES)
2439 		m = m_getcl(how, MT_DATA, M_PKTHDR);
2440 	else
2441 		m = NULL;
2442 	if (m == NULL)
2443 		return (NULL);
2444 	m->m_pkthdr.len = len;
2445 	m->m_len = len;
2446 	set_mbuf_cflags(m, MC_RAW_WR);
2447 	set_mbuf_len16(m, howmany(len, 16));
2448 	return (m);
2449 }
2450 
2451 static inline bool
2452 needs_hwcsum(struct mbuf *m)
2453 {
2454 	const uint32_t csum_flags = CSUM_IP | CSUM_IP_UDP | CSUM_IP_TCP |
2455 	    CSUM_IP_TSO | CSUM_INNER_IP | CSUM_INNER_IP_UDP |
2456 	    CSUM_INNER_IP_TCP | CSUM_INNER_IP_TSO | CSUM_IP6_UDP |
2457 	    CSUM_IP6_TCP | CSUM_IP6_TSO | CSUM_INNER_IP6_UDP |
2458 	    CSUM_INNER_IP6_TCP | CSUM_INNER_IP6_TSO;
2459 
2460 	M_ASSERTPKTHDR(m);
2461 
2462 	return (m->m_pkthdr.csum_flags & csum_flags);
2463 }
2464 
2465 static inline bool
2466 needs_tso(struct mbuf *m)
2467 {
2468 	const uint32_t csum_flags = CSUM_IP_TSO | CSUM_IP6_TSO |
2469 	    CSUM_INNER_IP_TSO | CSUM_INNER_IP6_TSO;
2470 
2471 	M_ASSERTPKTHDR(m);
2472 
2473 	return (m->m_pkthdr.csum_flags & csum_flags);
2474 }
2475 
2476 static inline bool
2477 needs_vxlan_csum(struct mbuf *m)
2478 {
2479 
2480 	M_ASSERTPKTHDR(m);
2481 
2482 	return (m->m_pkthdr.csum_flags & CSUM_ENCAP_VXLAN);
2483 }
2484 
2485 static inline bool
2486 needs_vxlan_tso(struct mbuf *m)
2487 {
2488 	const uint32_t csum_flags = CSUM_ENCAP_VXLAN | CSUM_INNER_IP_TSO |
2489 	    CSUM_INNER_IP6_TSO;
2490 
2491 	M_ASSERTPKTHDR(m);
2492 
2493 	return ((m->m_pkthdr.csum_flags & csum_flags) != 0 &&
2494 	    (m->m_pkthdr.csum_flags & csum_flags) != CSUM_ENCAP_VXLAN);
2495 }
2496 
2497 static inline bool
2498 needs_inner_tcp_csum(struct mbuf *m)
2499 {
2500 	const uint32_t csum_flags = CSUM_INNER_IP_TSO | CSUM_INNER_IP6_TSO;
2501 
2502 	M_ASSERTPKTHDR(m);
2503 
2504 	return (m->m_pkthdr.csum_flags & csum_flags);
2505 }
2506 
2507 static inline bool
2508 needs_l3_csum(struct mbuf *m)
2509 {
2510 	const uint32_t csum_flags = CSUM_IP | CSUM_IP_TSO | CSUM_INNER_IP |
2511 	    CSUM_INNER_IP_TSO;
2512 
2513 	M_ASSERTPKTHDR(m);
2514 
2515 	return (m->m_pkthdr.csum_flags & csum_flags);
2516 }
2517 
2518 static inline bool
2519 needs_outer_tcp_csum(struct mbuf *m)
2520 {
2521 	const uint32_t csum_flags = CSUM_IP_TCP | CSUM_IP_TSO | CSUM_IP6_TCP |
2522 	    CSUM_IP6_TSO;
2523 
2524 	M_ASSERTPKTHDR(m);
2525 
2526 	return (m->m_pkthdr.csum_flags & csum_flags);
2527 }
2528 
2529 #ifdef RATELIMIT
2530 static inline bool
2531 needs_outer_l4_csum(struct mbuf *m)
2532 {
2533 	const uint32_t csum_flags = CSUM_IP_UDP | CSUM_IP_TCP | CSUM_IP_TSO |
2534 	    CSUM_IP6_UDP | CSUM_IP6_TCP | CSUM_IP6_TSO;
2535 
2536 	M_ASSERTPKTHDR(m);
2537 
2538 	return (m->m_pkthdr.csum_flags & csum_flags);
2539 }
2540 
2541 static inline bool
2542 needs_outer_udp_csum(struct mbuf *m)
2543 {
2544 	const uint32_t csum_flags = CSUM_IP_UDP | CSUM_IP6_UDP;
2545 
2546 	M_ASSERTPKTHDR(m);
2547 
2548 	return (m->m_pkthdr.csum_flags & csum_flags);
2549 }
2550 #endif
2551 
2552 static inline bool
2553 needs_vlan_insertion(struct mbuf *m)
2554 {
2555 
2556 	M_ASSERTPKTHDR(m);
2557 
2558 	return (m->m_flags & M_VLANTAG);
2559 }
2560 
2561 static void *
2562 m_advance(struct mbuf **pm, int *poffset, int len)
2563 {
2564 	struct mbuf *m = *pm;
2565 	int offset = *poffset;
2566 	uintptr_t p = 0;
2567 
2568 	MPASS(len > 0);
2569 
2570 	for (;;) {
2571 		if (offset + len < m->m_len) {
2572 			offset += len;
2573 			p = mtod(m, uintptr_t) + offset;
2574 			break;
2575 		}
2576 		len -= m->m_len - offset;
2577 		m = m->m_next;
2578 		offset = 0;
2579 		MPASS(m != NULL);
2580 	}
2581 	*poffset = offset;
2582 	*pm = m;
2583 	return ((void *)p);
2584 }
2585 
2586 static inline int
2587 count_mbuf_ext_pgs(struct mbuf *m, int skip, vm_paddr_t *nextaddr)
2588 {
2589 	vm_paddr_t paddr;
2590 	int i, len, off, pglen, pgoff, seglen, segoff;
2591 	int nsegs = 0;
2592 
2593 	M_ASSERTEXTPG(m);
2594 	off = mtod(m, vm_offset_t);
2595 	len = m->m_len;
2596 	off += skip;
2597 	len -= skip;
2598 
2599 	if (m->m_epg_hdrlen != 0) {
2600 		if (off >= m->m_epg_hdrlen) {
2601 			off -= m->m_epg_hdrlen;
2602 		} else {
2603 			seglen = m->m_epg_hdrlen - off;
2604 			segoff = off;
2605 			seglen = min(seglen, len);
2606 			off = 0;
2607 			len -= seglen;
2608 			paddr = pmap_kextract(
2609 			    (vm_offset_t)&m->m_epg_hdr[segoff]);
2610 			if (*nextaddr != paddr)
2611 				nsegs++;
2612 			*nextaddr = paddr + seglen;
2613 		}
2614 	}
2615 	pgoff = m->m_epg_1st_off;
2616 	for (i = 0; i < m->m_epg_npgs && len > 0; i++) {
2617 		pglen = m_epg_pagelen(m, i, pgoff);
2618 		if (off >= pglen) {
2619 			off -= pglen;
2620 			pgoff = 0;
2621 			continue;
2622 		}
2623 		seglen = pglen - off;
2624 		segoff = pgoff + off;
2625 		off = 0;
2626 		seglen = min(seglen, len);
2627 		len -= seglen;
2628 		paddr = m->m_epg_pa[i] + segoff;
2629 		if (*nextaddr != paddr)
2630 			nsegs++;
2631 		*nextaddr = paddr + seglen;
2632 		pgoff = 0;
2633 	};
2634 	if (len != 0) {
2635 		seglen = min(len, m->m_epg_trllen - off);
2636 		len -= seglen;
2637 		paddr = pmap_kextract((vm_offset_t)&m->m_epg_trail[off]);
2638 		if (*nextaddr != paddr)
2639 			nsegs++;
2640 		*nextaddr = paddr + seglen;
2641 	}
2642 
2643 	return (nsegs);
2644 }
2645 
2646 
2647 /*
2648  * Can deal with empty mbufs in the chain that have m_len = 0, but the chain
2649  * must have at least one mbuf that's not empty.  It is possible for this
2650  * routine to return 0 if skip accounts for all the contents of the mbuf chain.
2651  */
2652 static inline int
2653 count_mbuf_nsegs(struct mbuf *m, int skip, uint8_t *cflags)
2654 {
2655 	vm_paddr_t nextaddr, paddr;
2656 	vm_offset_t va;
2657 	int len, nsegs;
2658 
2659 	M_ASSERTPKTHDR(m);
2660 	MPASS(m->m_pkthdr.len > 0);
2661 	MPASS(m->m_pkthdr.len >= skip);
2662 
2663 	nsegs = 0;
2664 	nextaddr = 0;
2665 	for (; m; m = m->m_next) {
2666 		len = m->m_len;
2667 		if (__predict_false(len == 0))
2668 			continue;
2669 		if (skip >= len) {
2670 			skip -= len;
2671 			continue;
2672 		}
2673 		if ((m->m_flags & M_EXTPG) != 0) {
2674 			*cflags |= MC_NOMAP;
2675 			nsegs += count_mbuf_ext_pgs(m, skip, &nextaddr);
2676 			skip = 0;
2677 			continue;
2678 		}
2679 		va = mtod(m, vm_offset_t) + skip;
2680 		len -= skip;
2681 		skip = 0;
2682 		paddr = pmap_kextract(va);
2683 		nsegs += sglist_count((void *)(uintptr_t)va, len);
2684 		if (paddr == nextaddr)
2685 			nsegs--;
2686 		nextaddr = pmap_kextract(va + len - 1) + 1;
2687 	}
2688 
2689 	return (nsegs);
2690 }
2691 
2692 /*
2693  * The maximum number of segments that can fit in a WR.
2694  */
2695 static int
2696 max_nsegs_allowed(struct mbuf *m, bool vm_wr)
2697 {
2698 
2699 	if (vm_wr) {
2700 		if (needs_tso(m))
2701 			return (TX_SGL_SEGS_VM_TSO);
2702 		return (TX_SGL_SEGS_VM);
2703 	}
2704 
2705 	if (needs_tso(m)) {
2706 		if (needs_vxlan_tso(m))
2707 			return (TX_SGL_SEGS_VXLAN_TSO);
2708 		else
2709 			return (TX_SGL_SEGS_TSO);
2710 	}
2711 
2712 	return (TX_SGL_SEGS);
2713 }
2714 
2715 /*
2716  * Analyze the mbuf to determine its tx needs.  The mbuf passed in may change:
2717  * a) caller can assume it's been freed if this function returns with an error.
2718  * b) it may get defragged up if the gather list is too long for the hardware.
2719  */
2720 int
2721 parse_pkt(struct mbuf **mp, bool vm_wr)
2722 {
2723 	struct mbuf *m0 = *mp, *m;
2724 	int rc, nsegs, defragged = 0, offset;
2725 	struct ether_header *eh;
2726 	void *l3hdr;
2727 #if defined(INET) || defined(INET6)
2728 	struct tcphdr *tcp;
2729 #endif
2730 #if defined(KERN_TLS) || defined(RATELIMIT)
2731 	struct m_snd_tag *mst;
2732 #endif
2733 	uint16_t eh_type;
2734 	uint8_t cflags;
2735 
2736 	cflags = 0;
2737 	M_ASSERTPKTHDR(m0);
2738 	if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) {
2739 		rc = EINVAL;
2740 fail:
2741 		m_freem(m0);
2742 		*mp = NULL;
2743 		return (rc);
2744 	}
2745 restart:
2746 	/*
2747 	 * First count the number of gather list segments in the payload.
2748 	 * Defrag the mbuf if nsegs exceeds the hardware limit.
2749 	 */
2750 	M_ASSERTPKTHDR(m0);
2751 	MPASS(m0->m_pkthdr.len > 0);
2752 	nsegs = count_mbuf_nsegs(m0, 0, &cflags);
2753 #if defined(KERN_TLS) || defined(RATELIMIT)
2754 	if (m0->m_pkthdr.csum_flags & CSUM_SND_TAG)
2755 		mst = m0->m_pkthdr.snd_tag;
2756 	else
2757 		mst = NULL;
2758 #endif
2759 #ifdef KERN_TLS
2760 	if (mst != NULL && mst->type == IF_SND_TAG_TYPE_TLS) {
2761 		int len16;
2762 
2763 		cflags |= MC_TLS;
2764 		set_mbuf_cflags(m0, cflags);
2765 		rc = t6_ktls_parse_pkt(m0, &nsegs, &len16);
2766 		if (rc != 0)
2767 			goto fail;
2768 		set_mbuf_nsegs(m0, nsegs);
2769 		set_mbuf_len16(m0, len16);
2770 		return (0);
2771 	}
2772 #endif
2773 	if (nsegs > max_nsegs_allowed(m0, vm_wr)) {
2774 		if (defragged++ > 0) {
2775 			rc = EFBIG;
2776 			goto fail;
2777 		}
2778 		counter_u64_add(defrags, 1);
2779 		if ((m = m_defrag(m0, M_NOWAIT)) == NULL) {
2780 			rc = ENOMEM;
2781 			goto fail;
2782 		}
2783 		*mp = m0 = m;	/* update caller's copy after defrag */
2784 		goto restart;
2785 	}
2786 
2787 	if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN &&
2788 	    !(cflags & MC_NOMAP))) {
2789 		counter_u64_add(pullups, 1);
2790 		m0 = m_pullup(m0, m0->m_pkthdr.len);
2791 		if (m0 == NULL) {
2792 			/* Should have left well enough alone. */
2793 			rc = EFBIG;
2794 			goto fail;
2795 		}
2796 		*mp = m0;	/* update caller's copy after pullup */
2797 		goto restart;
2798 	}
2799 	set_mbuf_nsegs(m0, nsegs);
2800 	set_mbuf_cflags(m0, cflags);
2801 	calculate_mbuf_len16(m0, vm_wr);
2802 
2803 #ifdef RATELIMIT
2804 	/*
2805 	 * Ethofld is limited to TCP and UDP for now, and only when L4 hw
2806 	 * checksumming is enabled.  needs_outer_l4_csum happens to check for
2807 	 * all the right things.
2808 	 */
2809 	if (__predict_false(needs_eo(mst) && !needs_outer_l4_csum(m0))) {
2810 		m_snd_tag_rele(m0->m_pkthdr.snd_tag);
2811 		m0->m_pkthdr.snd_tag = NULL;
2812 		m0->m_pkthdr.csum_flags &= ~CSUM_SND_TAG;
2813 		mst = NULL;
2814 	}
2815 #endif
2816 
2817 	if (!needs_hwcsum(m0)
2818 #ifdef RATELIMIT
2819    		 && !needs_eo(mst)
2820 #endif
2821 	)
2822 		return (0);
2823 
2824 	m = m0;
2825 	eh = mtod(m, struct ether_header *);
2826 	eh_type = ntohs(eh->ether_type);
2827 	if (eh_type == ETHERTYPE_VLAN) {
2828 		struct ether_vlan_header *evh = (void *)eh;
2829 
2830 		eh_type = ntohs(evh->evl_proto);
2831 		m0->m_pkthdr.l2hlen = sizeof(*evh);
2832 	} else
2833 		m0->m_pkthdr.l2hlen = sizeof(*eh);
2834 
2835 	offset = 0;
2836 	l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2837 
2838 	switch (eh_type) {
2839 #ifdef INET6
2840 	case ETHERTYPE_IPV6:
2841 		m0->m_pkthdr.l3hlen = sizeof(struct ip6_hdr);
2842 		break;
2843 #endif
2844 #ifdef INET
2845 	case ETHERTYPE_IP:
2846 	{
2847 		struct ip *ip = l3hdr;
2848 
2849 		if (needs_vxlan_csum(m0)) {
2850 			/* Driver will do the outer IP hdr checksum. */
2851 			ip->ip_sum = 0;
2852 			if (needs_vxlan_tso(m0)) {
2853 				const uint16_t ipl = ip->ip_len;
2854 
2855 				ip->ip_len = 0;
2856 				ip->ip_sum = ~in_cksum_hdr(ip);
2857 				ip->ip_len = ipl;
2858 			} else
2859 				ip->ip_sum = in_cksum_hdr(ip);
2860 		}
2861 		m0->m_pkthdr.l3hlen = ip->ip_hl << 2;
2862 		break;
2863 	}
2864 #endif
2865 	default:
2866 		panic("%s: ethertype 0x%04x unknown.  if_cxgbe must be compiled"
2867 		    " with the same INET/INET6 options as the kernel.",
2868 		    __func__, eh_type);
2869 	}
2870 
2871 	if (needs_vxlan_csum(m0)) {
2872 		m0->m_pkthdr.l4hlen = sizeof(struct udphdr);
2873 		m0->m_pkthdr.l5hlen = sizeof(struct vxlan_header);
2874 
2875 		/* Inner headers. */
2876 		eh = m_advance(&m, &offset, m0->m_pkthdr.l3hlen +
2877 		    sizeof(struct udphdr) + sizeof(struct vxlan_header));
2878 		eh_type = ntohs(eh->ether_type);
2879 		if (eh_type == ETHERTYPE_VLAN) {
2880 			struct ether_vlan_header *evh = (void *)eh;
2881 
2882 			eh_type = ntohs(evh->evl_proto);
2883 			m0->m_pkthdr.inner_l2hlen = sizeof(*evh);
2884 		} else
2885 			m0->m_pkthdr.inner_l2hlen = sizeof(*eh);
2886 		l3hdr = m_advance(&m, &offset, m0->m_pkthdr.inner_l2hlen);
2887 
2888 		switch (eh_type) {
2889 #ifdef INET6
2890 		case ETHERTYPE_IPV6:
2891 			m0->m_pkthdr.inner_l3hlen = sizeof(struct ip6_hdr);
2892 			break;
2893 #endif
2894 #ifdef INET
2895 		case ETHERTYPE_IP:
2896 		{
2897 			struct ip *ip = l3hdr;
2898 
2899 			m0->m_pkthdr.inner_l3hlen = ip->ip_hl << 2;
2900 			break;
2901 		}
2902 #endif
2903 		default:
2904 			panic("%s: VXLAN hw offload requested with unknown "
2905 			    "ethertype 0x%04x.  if_cxgbe must be compiled"
2906 			    " with the same INET/INET6 options as the kernel.",
2907 			    __func__, eh_type);
2908 		}
2909 #if defined(INET) || defined(INET6)
2910 		if (needs_inner_tcp_csum(m0)) {
2911 			tcp = m_advance(&m, &offset, m0->m_pkthdr.inner_l3hlen);
2912 			m0->m_pkthdr.inner_l4hlen = tcp->th_off * 4;
2913 		}
2914 #endif
2915 		MPASS((m0->m_pkthdr.csum_flags & CSUM_SND_TAG) == 0);
2916 		m0->m_pkthdr.csum_flags &= CSUM_INNER_IP6_UDP |
2917 		    CSUM_INNER_IP6_TCP | CSUM_INNER_IP6_TSO | CSUM_INNER_IP |
2918 		    CSUM_INNER_IP_UDP | CSUM_INNER_IP_TCP | CSUM_INNER_IP_TSO |
2919 		    CSUM_ENCAP_VXLAN;
2920 	}
2921 
2922 #if defined(INET) || defined(INET6)
2923 	if (needs_outer_tcp_csum(m0)) {
2924 		tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen);
2925 		m0->m_pkthdr.l4hlen = tcp->th_off * 4;
2926 #ifdef RATELIMIT
2927 		if (tsclk >= 0 && *(uint32_t *)(tcp + 1) == ntohl(0x0101080a)) {
2928 			set_mbuf_eo_tsclk_tsoff(m0,
2929 			    V_FW_ETH_TX_EO_WR_TSCLK(tsclk) |
2930 			    V_FW_ETH_TX_EO_WR_TSOFF(sizeof(*tcp) / 2 + 1));
2931 		} else
2932 			set_mbuf_eo_tsclk_tsoff(m0, 0);
2933 	} else if (needs_outer_udp_csum(m0)) {
2934 		m0->m_pkthdr.l4hlen = sizeof(struct udphdr);
2935 #endif
2936 	}
2937 #ifdef RATELIMIT
2938 	if (needs_eo(mst)) {
2939 		u_int immhdrs;
2940 
2941 		/* EO WRs have the headers in the WR and not the GL. */
2942 		immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen +
2943 		    m0->m_pkthdr.l4hlen;
2944 		cflags = 0;
2945 		nsegs = count_mbuf_nsegs(m0, immhdrs, &cflags);
2946 		MPASS(cflags == mbuf_cflags(m0));
2947 		set_mbuf_eo_nsegs(m0, nsegs);
2948 		set_mbuf_eo_len16(m0,
2949 		    txpkt_eo_len16(nsegs, immhdrs, needs_tso(m0)));
2950 	}
2951 #endif
2952 #endif
2953 	MPASS(m0 == *mp);
2954 	return (0);
2955 }
2956 
2957 void *
2958 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie)
2959 {
2960 	struct sge_eq *eq = &wrq->eq;
2961 	struct adapter *sc = wrq->adapter;
2962 	int ndesc, available;
2963 	struct wrqe *wr;
2964 	void *w;
2965 
2966 	MPASS(len16 > 0);
2967 	ndesc = tx_len16_to_desc(len16);
2968 	MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC);
2969 
2970 	EQ_LOCK(eq);
2971 
2972 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2973 		drain_wrq_wr_list(sc, wrq);
2974 
2975 	if (!STAILQ_EMPTY(&wrq->wr_list)) {
2976 slowpath:
2977 		EQ_UNLOCK(eq);
2978 		wr = alloc_wrqe(len16 * 16, wrq);
2979 		if (__predict_false(wr == NULL))
2980 			return (NULL);
2981 		cookie->pidx = -1;
2982 		cookie->ndesc = ndesc;
2983 		return (&wr->wr);
2984 	}
2985 
2986 	eq->cidx = read_hw_cidx(eq);
2987 	if (eq->pidx == eq->cidx)
2988 		available = eq->sidx - 1;
2989 	else
2990 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2991 	if (available < ndesc)
2992 		goto slowpath;
2993 
2994 	cookie->pidx = eq->pidx;
2995 	cookie->ndesc = ndesc;
2996 	TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link);
2997 
2998 	w = &eq->desc[eq->pidx];
2999 	IDXINCR(eq->pidx, ndesc, eq->sidx);
3000 	if (__predict_false(cookie->pidx + ndesc > eq->sidx)) {
3001 		w = &wrq->ss[0];
3002 		wrq->ss_pidx = cookie->pidx;
3003 		wrq->ss_len = len16 * 16;
3004 	}
3005 
3006 	EQ_UNLOCK(eq);
3007 
3008 	return (w);
3009 }
3010 
3011 void
3012 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie)
3013 {
3014 	struct sge_eq *eq = &wrq->eq;
3015 	struct adapter *sc = wrq->adapter;
3016 	int ndesc, pidx;
3017 	struct wrq_cookie *prev, *next;
3018 
3019 	if (cookie->pidx == -1) {
3020 		struct wrqe *wr = __containerof(w, struct wrqe, wr);
3021 
3022 		t4_wrq_tx(sc, wr);
3023 		return;
3024 	}
3025 
3026 	if (__predict_false(w == &wrq->ss[0])) {
3027 		int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE;
3028 
3029 		MPASS(wrq->ss_len > n);	/* WR had better wrap around. */
3030 		bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n);
3031 		bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n);
3032 		wrq->tx_wrs_ss++;
3033 	} else
3034 		wrq->tx_wrs_direct++;
3035 
3036 	EQ_LOCK(eq);
3037 	ndesc = cookie->ndesc;	/* Can be more than SGE_MAX_WR_NDESC here. */
3038 	pidx = cookie->pidx;
3039 	MPASS(pidx >= 0 && pidx < eq->sidx);
3040 	prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link);
3041 	next = TAILQ_NEXT(cookie, link);
3042 	if (prev == NULL) {
3043 		MPASS(pidx == eq->dbidx);
3044 		if (next == NULL || ndesc >= 16) {
3045 			int available;
3046 			struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
3047 
3048 			/*
3049 			 * Note that the WR via which we'll request tx updates
3050 			 * is at pidx and not eq->pidx, which has moved on
3051 			 * already.
3052 			 */
3053 			dst = (void *)&eq->desc[pidx];
3054 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
3055 			if (available < eq->sidx / 4 &&
3056 			    atomic_cmpset_int(&eq->equiq, 0, 1)) {
3057 				/*
3058 				 * XXX: This is not 100% reliable with some
3059 				 * types of WRs.  But this is a very unusual
3060 				 * situation for an ofld/ctrl queue anyway.
3061 				 */
3062 				dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
3063 				    F_FW_WR_EQUEQ);
3064 			}
3065 
3066 			ring_eq_db(wrq->adapter, eq, ndesc);
3067 		} else {
3068 			MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc);
3069 			next->pidx = pidx;
3070 			next->ndesc += ndesc;
3071 		}
3072 	} else {
3073 		MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc);
3074 		prev->ndesc += ndesc;
3075 	}
3076 	TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link);
3077 
3078 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
3079 		drain_wrq_wr_list(sc, wrq);
3080 
3081 #ifdef INVARIANTS
3082 	if (TAILQ_EMPTY(&wrq->incomplete_wrs)) {
3083 		/* Doorbell must have caught up to the pidx. */
3084 		MPASS(wrq->eq.pidx == wrq->eq.dbidx);
3085 	}
3086 #endif
3087 	EQ_UNLOCK(eq);
3088 }
3089 
3090 static u_int
3091 can_resume_eth_tx(struct mp_ring *r)
3092 {
3093 	struct sge_eq *eq = r->cookie;
3094 
3095 	return (total_available_tx_desc(eq) > eq->sidx / 8);
3096 }
3097 
3098 static inline bool
3099 cannot_use_txpkts(struct mbuf *m)
3100 {
3101 	/* maybe put a GL limit too, to avoid silliness? */
3102 
3103 	return (needs_tso(m) || (mbuf_cflags(m) & (MC_RAW_WR | MC_TLS)) != 0);
3104 }
3105 
3106 static inline int
3107 discard_tx(struct sge_eq *eq)
3108 {
3109 
3110 	return ((eq->flags & (EQ_ENABLED | EQ_QFLUSH)) != EQ_ENABLED);
3111 }
3112 
3113 static inline int
3114 wr_can_update_eq(void *p)
3115 {
3116 	struct fw_eth_tx_pkts_wr *wr = p;
3117 
3118 	switch (G_FW_WR_OP(be32toh(wr->op_pkd))) {
3119 	case FW_ULPTX_WR:
3120 	case FW_ETH_TX_PKT_WR:
3121 	case FW_ETH_TX_PKTS_WR:
3122 	case FW_ETH_TX_PKTS2_WR:
3123 	case FW_ETH_TX_PKT_VM_WR:
3124 	case FW_ETH_TX_PKTS_VM_WR:
3125 		return (1);
3126 	default:
3127 		return (0);
3128 	}
3129 }
3130 
3131 static inline void
3132 set_txupdate_flags(struct sge_txq *txq, u_int avail,
3133     struct fw_eth_tx_pkt_wr *wr)
3134 {
3135 	struct sge_eq *eq = &txq->eq;
3136 	struct txpkts *txp = &txq->txp;
3137 
3138 	if ((txp->npkt > 0 || avail < eq->sidx / 2) &&
3139 	    atomic_cmpset_int(&eq->equiq, 0, 1)) {
3140 		wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ | F_FW_WR_EQUIQ);
3141 		eq->equeqidx = eq->pidx;
3142 	} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
3143 		wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
3144 		eq->equeqidx = eq->pidx;
3145 	}
3146 }
3147 
3148 #if defined(__i386__) || defined(__amd64__)
3149 extern uint64_t tsc_freq;
3150 #endif
3151 
3152 static inline bool
3153 record_eth_tx_time(struct sge_txq *txq)
3154 {
3155 	const uint64_t cycles = get_cyclecount();
3156 	const uint64_t last_tx = txq->last_tx;
3157 #if defined(__i386__) || defined(__amd64__)
3158 	const uint64_t itg = tsc_freq * t4_tx_coalesce_gap / 1000000;
3159 #else
3160 	const uint64_t itg = 0;
3161 #endif
3162 
3163 	MPASS(cycles >= last_tx);
3164 	txq->last_tx = cycles;
3165 	return (cycles - last_tx < itg);
3166 }
3167 
3168 /*
3169  * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to
3170  * be consumed.  Return the actual number consumed.  0 indicates a stall.
3171  */
3172 static u_int
3173 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx, bool *coalescing)
3174 {
3175 	struct sge_txq *txq = r->cookie;
3176 	struct ifnet *ifp = txq->ifp;
3177 	struct sge_eq *eq = &txq->eq;
3178 	struct txpkts *txp = &txq->txp;
3179 	struct vi_info *vi = ifp->if_softc;
3180 	struct adapter *sc = vi->adapter;
3181 	u_int total, remaining;		/* # of packets */
3182 	u_int n, avail, dbdiff;		/* # of hardware descriptors */
3183 	int i, rc;
3184 	struct mbuf *m0;
3185 	bool snd, recent_tx;
3186 	void *wr;	/* start of the last WR written to the ring */
3187 
3188 	TXQ_LOCK_ASSERT_OWNED(txq);
3189 	recent_tx = record_eth_tx_time(txq);
3190 
3191 	remaining = IDXDIFF(pidx, cidx, r->size);
3192 	if (__predict_false(discard_tx(eq))) {
3193 		for (i = 0; i < txp->npkt; i++)
3194 			m_freem(txp->mb[i]);
3195 		txp->npkt = 0;
3196 		while (cidx != pidx) {
3197 			m0 = r->items[cidx];
3198 			m_freem(m0);
3199 			if (++cidx == r->size)
3200 				cidx = 0;
3201 		}
3202 		reclaim_tx_descs(txq, eq->sidx);
3203 		*coalescing = false;
3204 		return (remaining);	/* emptied */
3205 	}
3206 
3207 	/* How many hardware descriptors do we have readily available. */
3208 	if (eq->pidx == eq->cidx)
3209 		avail = eq->sidx - 1;
3210 	else
3211 		avail = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
3212 
3213 	total = 0;
3214 	if (remaining == 0) {
3215 		txp->score = 0;
3216 		txq->txpkts_flush++;
3217 		goto send_txpkts;
3218 	}
3219 
3220 	dbdiff = 0;
3221 	MPASS(remaining > 0);
3222 	while (remaining > 0) {
3223 		m0 = r->items[cidx];
3224 		M_ASSERTPKTHDR(m0);
3225 		MPASS(m0->m_nextpkt == NULL);
3226 
3227 		if (avail < 2 * SGE_MAX_WR_NDESC)
3228 			avail += reclaim_tx_descs(txq, 64);
3229 
3230 		if (t4_tx_coalesce == 0 && txp->npkt == 0)
3231 			goto skip_coalescing;
3232 		if (cannot_use_txpkts(m0))
3233 			txp->score = 0;
3234 		else if (recent_tx) {
3235 			if (++txp->score == 0)
3236 				txp->score = UINT8_MAX;
3237 		} else
3238 			txp->score = 1;
3239 		if (txp->npkt > 0 || remaining > 1 ||
3240 		    txp->score >= t4_tx_coalesce_pkts ||
3241 		    atomic_load_int(&txq->eq.equiq) != 0) {
3242 			if (vi->flags & TX_USES_VM_WR)
3243 				rc = add_to_txpkts_vf(sc, txq, m0, avail, &snd);
3244 			else
3245 				rc = add_to_txpkts_pf(sc, txq, m0, avail, &snd);
3246 		} else {
3247 			snd = false;
3248 			rc = EINVAL;
3249 		}
3250 		if (snd) {
3251 			MPASS(txp->npkt > 0);
3252 			for (i = 0; i < txp->npkt; i++)
3253 				ETHER_BPF_MTAP(ifp, txp->mb[i]);
3254 			if (txp->npkt > 1) {
3255 				MPASS(avail >= tx_len16_to_desc(txp->len16));
3256 				if (vi->flags & TX_USES_VM_WR)
3257 					n = write_txpkts_vm_wr(sc, txq);
3258 				else
3259 					n = write_txpkts_wr(sc, txq);
3260 			} else {
3261 				MPASS(avail >=
3262 				    tx_len16_to_desc(mbuf_len16(txp->mb[0])));
3263 				if (vi->flags & TX_USES_VM_WR)
3264 					n = write_txpkt_vm_wr(sc, txq,
3265 					    txp->mb[0]);
3266 				else
3267 					n = write_txpkt_wr(sc, txq, txp->mb[0],
3268 					    avail);
3269 			}
3270 			MPASS(n <= SGE_MAX_WR_NDESC);
3271 			avail -= n;
3272 			dbdiff += n;
3273 			wr = &eq->desc[eq->pidx];
3274 			IDXINCR(eq->pidx, n, eq->sidx);
3275 			txp->npkt = 0;	/* emptied */
3276 		}
3277 		if (rc == 0) {
3278 			/* m0 was coalesced into txq->txpkts. */
3279 			goto next_mbuf;
3280 		}
3281 		if (rc == EAGAIN) {
3282 			/*
3283 			 * m0 is suitable for tx coalescing but could not be
3284 			 * combined with the existing txq->txpkts, which has now
3285 			 * been transmitted.  Start a new txpkts with m0.
3286 			 */
3287 			MPASS(snd);
3288 			MPASS(txp->npkt == 0);
3289 			continue;
3290 		}
3291 
3292 		MPASS(rc != 0 && rc != EAGAIN);
3293 		MPASS(txp->npkt == 0);
3294 skip_coalescing:
3295 		n = tx_len16_to_desc(mbuf_len16(m0));
3296 		if (__predict_false(avail < n)) {
3297 			avail += reclaim_tx_descs(txq, min(n, 32));
3298 			if (avail < n)
3299 				break;	/* out of descriptors */
3300 		}
3301 
3302 		wr = &eq->desc[eq->pidx];
3303 		if (mbuf_cflags(m0) & MC_RAW_WR) {
3304 			n = write_raw_wr(txq, wr, m0, avail);
3305 #ifdef KERN_TLS
3306 		} else if (mbuf_cflags(m0) & MC_TLS) {
3307 			ETHER_BPF_MTAP(ifp, m0);
3308 			n = t6_ktls_write_wr(txq, wr, m0, mbuf_nsegs(m0),
3309 			    avail);
3310 #endif
3311 		} else {
3312 			ETHER_BPF_MTAP(ifp, m0);
3313 			if (vi->flags & TX_USES_VM_WR)
3314 				n = write_txpkt_vm_wr(sc, txq, m0);
3315 			else
3316 				n = write_txpkt_wr(sc, txq, m0, avail);
3317 		}
3318 		MPASS(n >= 1 && n <= avail);
3319 		if (!(mbuf_cflags(m0) & MC_TLS))
3320 			MPASS(n <= SGE_MAX_WR_NDESC);
3321 
3322 		avail -= n;
3323 		dbdiff += n;
3324 		IDXINCR(eq->pidx, n, eq->sidx);
3325 
3326 		if (dbdiff >= 512 / EQ_ESIZE) {	/* X_FETCHBURSTMAX_512B */
3327 			if (wr_can_update_eq(wr))
3328 				set_txupdate_flags(txq, avail, wr);
3329 			ring_eq_db(sc, eq, dbdiff);
3330 			avail += reclaim_tx_descs(txq, 32);
3331 			dbdiff = 0;
3332 		}
3333 next_mbuf:
3334 		total++;
3335 		remaining--;
3336 		if (__predict_false(++cidx == r->size))
3337 			cidx = 0;
3338 	}
3339 	if (dbdiff != 0) {
3340 		if (wr_can_update_eq(wr))
3341 			set_txupdate_flags(txq, avail, wr);
3342 		ring_eq_db(sc, eq, dbdiff);
3343 		reclaim_tx_descs(txq, 32);
3344 	} else if (eq->pidx == eq->cidx && txp->npkt > 0 &&
3345 	    atomic_load_int(&txq->eq.equiq) == 0) {
3346 		/*
3347 		 * If nothing was submitted to the chip for tx (it was coalesced
3348 		 * into txpkts instead) and there is no tx update outstanding
3349 		 * then we need to send txpkts now.
3350 		 */
3351 send_txpkts:
3352 		MPASS(txp->npkt > 0);
3353 		for (i = 0; i < txp->npkt; i++)
3354 			ETHER_BPF_MTAP(ifp, txp->mb[i]);
3355 		if (txp->npkt > 1) {
3356 			MPASS(avail >= tx_len16_to_desc(txp->len16));
3357 			if (vi->flags & TX_USES_VM_WR)
3358 				n = write_txpkts_vm_wr(sc, txq);
3359 			else
3360 				n = write_txpkts_wr(sc, txq);
3361 		} else {
3362 			MPASS(avail >=
3363 			    tx_len16_to_desc(mbuf_len16(txp->mb[0])));
3364 			if (vi->flags & TX_USES_VM_WR)
3365 				n = write_txpkt_vm_wr(sc, txq, txp->mb[0]);
3366 			else
3367 				n = write_txpkt_wr(sc, txq, txp->mb[0], avail);
3368 		}
3369 		MPASS(n <= SGE_MAX_WR_NDESC);
3370 		wr = &eq->desc[eq->pidx];
3371 		IDXINCR(eq->pidx, n, eq->sidx);
3372 		txp->npkt = 0;	/* emptied */
3373 
3374 		MPASS(wr_can_update_eq(wr));
3375 		set_txupdate_flags(txq, avail - n, wr);
3376 		ring_eq_db(sc, eq, n);
3377 		reclaim_tx_descs(txq, 32);
3378 	}
3379 	*coalescing = txp->npkt > 0;
3380 
3381 	return (total);
3382 }
3383 
3384 static inline void
3385 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
3386     int qsize)
3387 {
3388 
3389 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
3390 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
3391 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
3392 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
3393 
3394 	iq->flags = 0;
3395 	iq->adapter = sc;
3396 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
3397 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
3398 	if (pktc_idx >= 0) {
3399 		iq->intr_params |= F_QINTR_CNT_EN;
3400 		iq->intr_pktc_idx = pktc_idx;
3401 	}
3402 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
3403 	iq->sidx = iq->qsize - sc->params.sge.spg_len / IQ_ESIZE;
3404 }
3405 
3406 static inline void
3407 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name)
3408 {
3409 
3410 	fl->qsize = qsize;
3411 	fl->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
3412 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
3413 	if (sc->flags & BUF_PACKING_OK &&
3414 	    ((!is_t4(sc) && buffer_packing) ||	/* T5+: enabled unless 0 */
3415 	    (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */
3416 		fl->flags |= FL_BUF_PACKING;
3417 	fl->zidx = find_refill_source(sc, maxp, fl->flags & FL_BUF_PACKING);
3418 	fl->safe_zidx = sc->sge.safe_zidx;
3419 }
3420 
3421 static inline void
3422 init_eq(struct adapter *sc, struct sge_eq *eq, int eqtype, int qsize,
3423     uint8_t tx_chan, uint16_t iqid, char *name)
3424 {
3425 	KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
3426 
3427 	eq->flags = eqtype & EQ_TYPEMASK;
3428 	eq->tx_chan = tx_chan;
3429 	eq->iqid = iqid;
3430 	eq->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
3431 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
3432 }
3433 
3434 int
3435 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
3436     bus_dmamap_t *map, bus_addr_t *pa, void **va)
3437 {
3438 	int rc;
3439 
3440 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
3441 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
3442 	if (rc != 0) {
3443 		device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
3444 		goto done;
3445 	}
3446 
3447 	rc = bus_dmamem_alloc(*tag, va,
3448 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
3449 	if (rc != 0) {
3450 		device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
3451 		goto done;
3452 	}
3453 
3454 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
3455 	if (rc != 0) {
3456 		device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
3457 		goto done;
3458 	}
3459 done:
3460 	if (rc)
3461 		free_ring(sc, *tag, *map, *pa, *va);
3462 
3463 	return (rc);
3464 }
3465 
3466 int
3467 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
3468     bus_addr_t pa, void *va)
3469 {
3470 	if (pa)
3471 		bus_dmamap_unload(tag, map);
3472 	if (va)
3473 		bus_dmamem_free(tag, va, map);
3474 	if (tag)
3475 		bus_dma_tag_destroy(tag);
3476 
3477 	return (0);
3478 }
3479 
3480 /*
3481  * Allocates the ring for an ingress queue and an optional freelist.  If the
3482  * freelist is specified it will be allocated and then associated with the
3483  * ingress queue.
3484  *
3485  * Returns errno on failure.  Resources allocated up to that point may still be
3486  * allocated.  Caller is responsible for cleanup in case this function fails.
3487  *
3488  * If the ingress queue will take interrupts directly then the intr_idx
3489  * specifies the vector, starting from 0.  -1 means the interrupts for this
3490  * queue should be forwarded to the fwq.
3491  */
3492 static int
3493 alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl,
3494     int intr_idx, int cong)
3495 {
3496 	int rc, i, cntxt_id;
3497 	size_t len;
3498 	struct fw_iq_cmd c;
3499 	struct port_info *pi = vi->pi;
3500 	struct adapter *sc = iq->adapter;
3501 	struct sge_params *sp = &sc->params.sge;
3502 	__be32 v = 0;
3503 
3504 	len = iq->qsize * IQ_ESIZE;
3505 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
3506 	    (void **)&iq->desc);
3507 	if (rc != 0)
3508 		return (rc);
3509 
3510 	bzero(&c, sizeof(c));
3511 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
3512 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
3513 	    V_FW_IQ_CMD_VFN(0));
3514 
3515 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
3516 	    FW_LEN16(c));
3517 
3518 	/* Special handling for firmware event queue */
3519 	if (iq == &sc->sge.fwq)
3520 		v |= F_FW_IQ_CMD_IQASYNCH;
3521 
3522 	if (intr_idx < 0) {
3523 		/* Forwarded interrupts, all headed to fwq */
3524 		v |= F_FW_IQ_CMD_IQANDST;
3525 		v |= V_FW_IQ_CMD_IQANDSTINDEX(sc->sge.fwq.cntxt_id);
3526 	} else {
3527 		KASSERT(intr_idx < sc->intr_count,
3528 		    ("%s: invalid direct intr_idx %d", __func__, intr_idx));
3529 		v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
3530 	}
3531 
3532 	c.type_to_iqandstindex = htobe32(v |
3533 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
3534 	    V_FW_IQ_CMD_VIID(vi->viid) |
3535 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
3536 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
3537 	    F_FW_IQ_CMD_IQGTSMODE |
3538 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
3539 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
3540 	c.iqsize = htobe16(iq->qsize);
3541 	c.iqaddr = htobe64(iq->ba);
3542 	if (cong >= 0)
3543 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
3544 
3545 	if (fl) {
3546 		mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
3547 
3548 		len = fl->qsize * EQ_ESIZE;
3549 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
3550 		    &fl->ba, (void **)&fl->desc);
3551 		if (rc)
3552 			return (rc);
3553 
3554 		/* Allocate space for one software descriptor per buffer. */
3555 		rc = alloc_fl_sdesc(fl);
3556 		if (rc != 0) {
3557 			device_printf(sc->dev,
3558 			    "failed to setup fl software descriptors: %d\n",
3559 			    rc);
3560 			return (rc);
3561 		}
3562 
3563 		if (fl->flags & FL_BUF_PACKING) {
3564 			fl->lowat = roundup2(sp->fl_starve_threshold2, 8);
3565 			fl->buf_boundary = sp->pack_boundary;
3566 		} else {
3567 			fl->lowat = roundup2(sp->fl_starve_threshold, 8);
3568 			fl->buf_boundary = 16;
3569 		}
3570 		if (fl_pad && fl->buf_boundary < sp->pad_boundary)
3571 			fl->buf_boundary = sp->pad_boundary;
3572 
3573 		c.iqns_to_fl0congen |=
3574 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
3575 			F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
3576 			(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
3577 			(fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
3578 			    0));
3579 		if (cong >= 0) {
3580 			c.iqns_to_fl0congen |=
3581 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
3582 				    F_FW_IQ_CMD_FL0CONGCIF |
3583 				    F_FW_IQ_CMD_FL0CONGEN);
3584 		}
3585 		c.fl0dcaen_to_fl0cidxfthresh =
3586 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
3587 			X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B_T6) |
3588 			V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
3589 			X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
3590 		c.fl0size = htobe16(fl->qsize);
3591 		c.fl0addr = htobe64(fl->ba);
3592 	}
3593 
3594 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3595 	if (rc != 0) {
3596 		device_printf(sc->dev,
3597 		    "failed to create ingress queue: %d\n", rc);
3598 		return (rc);
3599 	}
3600 
3601 	iq->cidx = 0;
3602 	iq->gen = F_RSPD_GEN;
3603 	iq->intr_next = iq->intr_params;
3604 	iq->cntxt_id = be16toh(c.iqid);
3605 	iq->abs_id = be16toh(c.physiqid);
3606 	iq->flags |= IQ_ALLOCATED;
3607 
3608 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
3609 	if (cntxt_id >= sc->sge.iqmap_sz) {
3610 		panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
3611 		    cntxt_id, sc->sge.iqmap_sz - 1);
3612 	}
3613 	sc->sge.iqmap[cntxt_id] = iq;
3614 
3615 	if (fl) {
3616 		u_int qid;
3617 
3618 		iq->flags |= IQ_HAS_FL;
3619 		fl->cntxt_id = be16toh(c.fl0id);
3620 		fl->pidx = fl->cidx = 0;
3621 
3622 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
3623 		if (cntxt_id >= sc->sge.eqmap_sz) {
3624 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
3625 			    __func__, cntxt_id, sc->sge.eqmap_sz - 1);
3626 		}
3627 		sc->sge.eqmap[cntxt_id] = (void *)fl;
3628 
3629 		qid = fl->cntxt_id;
3630 		if (isset(&sc->doorbells, DOORBELL_UDB)) {
3631 			uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3632 			uint32_t mask = (1 << s_qpp) - 1;
3633 			volatile uint8_t *udb;
3634 
3635 			udb = sc->udbs_base + UDBS_DB_OFFSET;
3636 			udb += (qid >> s_qpp) << PAGE_SHIFT;
3637 			qid &= mask;
3638 			if (qid < PAGE_SIZE / UDBS_SEG_SIZE) {
3639 				udb += qid << UDBS_SEG_SHIFT;
3640 				qid = 0;
3641 			}
3642 			fl->udb = (volatile void *)udb;
3643 		}
3644 		fl->dbval = V_QID(qid) | sc->chip_params->sge_fl_db;
3645 
3646 		FL_LOCK(fl);
3647 		/* Enough to make sure the SGE doesn't think it's starved */
3648 		refill_fl(sc, fl, fl->lowat);
3649 		FL_UNLOCK(fl);
3650 	}
3651 
3652 	if (chip_id(sc) >= CHELSIO_T5 && !(sc->flags & IS_VF) && cong >= 0) {
3653 		uint32_t param, val;
3654 
3655 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
3656 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
3657 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
3658 		if (cong == 0)
3659 			val = 1 << 19;
3660 		else {
3661 			val = 2 << 19;
3662 			for (i = 0; i < 4; i++) {
3663 				if (cong & (1 << i))
3664 					val |= 1 << (i << 2);
3665 			}
3666 		}
3667 
3668 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
3669 		if (rc != 0) {
3670 			/* report error but carry on */
3671 			device_printf(sc->dev,
3672 			    "failed to set congestion manager context for "
3673 			    "ingress queue %d: %d\n", iq->cntxt_id, rc);
3674 		}
3675 	}
3676 
3677 	/* Enable IQ interrupts */
3678 	atomic_store_rel_int(&iq->state, IQS_IDLE);
3679 	t4_write_reg(sc, sc->sge_gts_reg, V_SEINTARM(iq->intr_params) |
3680 	    V_INGRESSQID(iq->cntxt_id));
3681 
3682 	return (0);
3683 }
3684 
3685 static int
3686 free_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl)
3687 {
3688 	int rc;
3689 	struct adapter *sc = iq->adapter;
3690 	device_t dev;
3691 
3692 	if (sc == NULL)
3693 		return (0);	/* nothing to do */
3694 
3695 	dev = vi ? vi->dev : sc->dev;
3696 
3697 	if (iq->flags & IQ_ALLOCATED) {
3698 		rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
3699 		    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
3700 		    fl ? fl->cntxt_id : 0xffff, 0xffff);
3701 		if (rc != 0) {
3702 			device_printf(dev,
3703 			    "failed to free queue %p: %d\n", iq, rc);
3704 			return (rc);
3705 		}
3706 		iq->flags &= ~IQ_ALLOCATED;
3707 	}
3708 
3709 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
3710 
3711 	bzero(iq, sizeof(*iq));
3712 
3713 	if (fl) {
3714 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
3715 		    fl->desc);
3716 
3717 		if (fl->sdesc)
3718 			free_fl_sdesc(sc, fl);
3719 
3720 		if (mtx_initialized(&fl->fl_lock))
3721 			mtx_destroy(&fl->fl_lock);
3722 
3723 		bzero(fl, sizeof(*fl));
3724 	}
3725 
3726 	return (0);
3727 }
3728 
3729 static void
3730 add_iq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
3731     struct sge_iq *iq)
3732 {
3733 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3734 
3735 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD, &iq->ba,
3736 	    "bus address of descriptor ring");
3737 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3738 	    iq->qsize * IQ_ESIZE, "descriptor ring size in bytes");
3739 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
3740 	    &iq->abs_id, 0, "absolute id of the queue");
3741 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3742 	    &iq->cntxt_id, 0, "SGE context id of the queue");
3743 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &iq->cidx,
3744 	    0, "consumer index");
3745 }
3746 
3747 static void
3748 add_fl_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
3749     struct sysctl_oid *oid, struct sge_fl *fl)
3750 {
3751 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3752 
3753 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl",
3754 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist");
3755 	children = SYSCTL_CHILDREN(oid);
3756 
3757 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3758 	    &fl->ba, "bus address of descriptor ring");
3759 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3760 	    fl->sidx * EQ_ESIZE + sc->params.sge.spg_len,
3761 	    "desc ring size in bytes");
3762 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3763 	    &fl->cntxt_id, 0, "SGE context id of the freelist");
3764 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL,
3765 	    fl_pad ? 1 : 0, "padding enabled");
3766 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL,
3767 	    fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled");
3768 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
3769 	    0, "consumer index");
3770 	if (fl->flags & FL_BUF_PACKING) {
3771 		SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
3772 		    CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
3773 	}
3774 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
3775 	    0, "producer index");
3776 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
3777 	    CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
3778 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
3779 	    CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
3780 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
3781 	    CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
3782 }
3783 
3784 static int
3785 alloc_fwq(struct adapter *sc)
3786 {
3787 	int rc, intr_idx;
3788 	struct sge_iq *fwq = &sc->sge.fwq;
3789 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
3790 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3791 
3792 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE);
3793 	if (sc->flags & IS_VF)
3794 		intr_idx = 0;
3795 	else
3796 		intr_idx = sc->intr_count > 1 ? 1 : 0;
3797 	rc = alloc_iq_fl(&sc->port[0]->vi[0], fwq, NULL, intr_idx, -1);
3798 	if (rc != 0) {
3799 		device_printf(sc->dev,
3800 		    "failed to create firmware event queue: %d\n", rc);
3801 		return (rc);
3802 	}
3803 
3804 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq",
3805 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "firmware event queue");
3806 	add_iq_sysctls(&sc->ctx, oid, fwq);
3807 
3808 	return (0);
3809 }
3810 
3811 static int
3812 free_fwq(struct adapter *sc)
3813 {
3814 	return free_iq_fl(NULL, &sc->sge.fwq, NULL);
3815 }
3816 
3817 static int
3818 alloc_ctrlq(struct adapter *sc, struct sge_wrq *ctrlq, int idx,
3819     struct sysctl_oid *oid)
3820 {
3821 	int rc;
3822 	char name[16];
3823 	struct sysctl_oid_list *children;
3824 
3825 	snprintf(name, sizeof(name), "%s ctrlq%d", device_get_nameunit(sc->dev),
3826 	    idx);
3827 	init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[idx]->tx_chan,
3828 	    sc->sge.fwq.cntxt_id, name);
3829 
3830 	children = SYSCTL_CHILDREN(oid);
3831 	snprintf(name, sizeof(name), "%d", idx);
3832 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, name,
3833 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "ctrl queue");
3834 	rc = alloc_wrq(sc, NULL, ctrlq, oid);
3835 
3836 	return (rc);
3837 }
3838 
3839 int
3840 tnl_cong(struct port_info *pi, int drop)
3841 {
3842 
3843 	if (drop == -1)
3844 		return (-1);
3845 	else if (drop == 1)
3846 		return (0);
3847 	else
3848 		return (pi->rx_e_chan_map);
3849 }
3850 
3851 static int
3852 alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int intr_idx, int idx,
3853     struct sysctl_oid *oid)
3854 {
3855 	int rc;
3856 	struct adapter *sc = vi->adapter;
3857 	struct sysctl_oid_list *children;
3858 	char name[16];
3859 
3860 	rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, intr_idx,
3861 	    tnl_cong(vi->pi, cong_drop));
3862 	if (rc != 0)
3863 		return (rc);
3864 
3865 	if (idx == 0)
3866 		sc->sge.iq_base = rxq->iq.abs_id - rxq->iq.cntxt_id;
3867 	else
3868 		KASSERT(rxq->iq.cntxt_id + sc->sge.iq_base == rxq->iq.abs_id,
3869 		    ("iq_base mismatch"));
3870 	KASSERT(sc->sge.iq_base == 0 || sc->flags & IS_VF,
3871 	    ("PF with non-zero iq_base"));
3872 
3873 	/*
3874 	 * The freelist is just barely above the starvation threshold right now,
3875 	 * fill it up a bit more.
3876 	 */
3877 	FL_LOCK(&rxq->fl);
3878 	refill_fl(sc, &rxq->fl, 128);
3879 	FL_UNLOCK(&rxq->fl);
3880 
3881 #if defined(INET) || defined(INET6)
3882 	rc = tcp_lro_init_args(&rxq->lro, vi->ifp, lro_entries, lro_mbufs);
3883 	if (rc != 0)
3884 		return (rc);
3885 	MPASS(rxq->lro.ifp == vi->ifp);	/* also indicates LRO init'ed */
3886 
3887 	if (vi->ifp->if_capenable & IFCAP_LRO)
3888 		rxq->iq.flags |= IQ_LRO_ENABLED;
3889 #endif
3890 	if (vi->ifp->if_capenable & IFCAP_HWRXTSTMP)
3891 		rxq->iq.flags |= IQ_RX_TIMESTAMP;
3892 	rxq->ifp = vi->ifp;
3893 
3894 	children = SYSCTL_CHILDREN(oid);
3895 
3896 	snprintf(name, sizeof(name), "%d", idx);
3897 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name,
3898 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "rx queue");
3899 	children = SYSCTL_CHILDREN(oid);
3900 
3901 	add_iq_sysctls(&vi->ctx, oid, &rxq->iq);
3902 #if defined(INET) || defined(INET6)
3903 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
3904 	    &rxq->lro.lro_queued, 0, NULL);
3905 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
3906 	    &rxq->lro.lro_flushed, 0, NULL);
3907 #endif
3908 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
3909 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
3910 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_extraction",
3911 	    CTLFLAG_RD, &rxq->vlan_extraction,
3912 	    "# of times hardware extracted 802.1Q tag");
3913 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vxlan_rxcsum",
3914 	    CTLFLAG_RD, &rxq->vxlan_rxcsum,
3915 	    "# of times hardware assisted with inner checksum (VXLAN) ");
3916 
3917 	add_fl_sysctls(sc, &vi->ctx, oid, &rxq->fl);
3918 
3919 	return (rc);
3920 }
3921 
3922 static int
3923 free_rxq(struct vi_info *vi, struct sge_rxq *rxq)
3924 {
3925 	int rc;
3926 
3927 #if defined(INET) || defined(INET6)
3928 	if (rxq->lro.ifp) {
3929 		tcp_lro_free(&rxq->lro);
3930 		rxq->lro.ifp = NULL;
3931 	}
3932 #endif
3933 
3934 	rc = free_iq_fl(vi, &rxq->iq, &rxq->fl);
3935 	if (rc == 0)
3936 		bzero(rxq, sizeof(*rxq));
3937 
3938 	return (rc);
3939 }
3940 
3941 #ifdef TCP_OFFLOAD
3942 static int
3943 alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq,
3944     int intr_idx, int idx, struct sysctl_oid *oid)
3945 {
3946 	struct port_info *pi = vi->pi;
3947 	int rc;
3948 	struct sysctl_oid_list *children;
3949 	char name[16];
3950 
3951 	rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx, 0);
3952 	if (rc != 0)
3953 		return (rc);
3954 
3955 	children = SYSCTL_CHILDREN(oid);
3956 
3957 	snprintf(name, sizeof(name), "%d", idx);
3958 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name,
3959 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "rx queue");
3960 	add_iq_sysctls(&vi->ctx, oid, &ofld_rxq->iq);
3961 	add_fl_sysctls(pi->adapter, &vi->ctx, oid, &ofld_rxq->fl);
3962 
3963 	SYSCTL_ADD_ULONG(&vi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
3964 	    "rx_toe_tls_records", CTLFLAG_RD, &ofld_rxq->rx_toe_tls_records,
3965 	    "# of TOE TLS records received");
3966 	SYSCTL_ADD_ULONG(&vi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
3967 	    "rx_toe_tls_octets", CTLFLAG_RD, &ofld_rxq->rx_toe_tls_octets,
3968 	    "# of payload octets in received TOE TLS records");
3969 
3970 	return (rc);
3971 }
3972 
3973 static int
3974 free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq)
3975 {
3976 	int rc;
3977 
3978 	rc = free_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl);
3979 	if (rc == 0)
3980 		bzero(ofld_rxq, sizeof(*ofld_rxq));
3981 
3982 	return (rc);
3983 }
3984 #endif
3985 
3986 /*
3987  * Returns a reasonable automatic cidx flush threshold for a given queue size.
3988  */
3989 static u_int
3990 qsize_to_fthresh(int qsize)
3991 {
3992 	u_int fthresh;
3993 
3994 	while (!powerof2(qsize))
3995 		qsize++;
3996 	fthresh = ilog2(qsize);
3997 	if (fthresh > X_CIDXFLUSHTHRESH_128)
3998 		fthresh = X_CIDXFLUSHTHRESH_128;
3999 
4000 	return (fthresh);
4001 }
4002 
4003 static int
4004 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
4005 {
4006 	int rc, cntxt_id;
4007 	struct fw_eq_ctrl_cmd c;
4008 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
4009 
4010 	bzero(&c, sizeof(c));
4011 
4012 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
4013 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
4014 	    V_FW_EQ_CTRL_CMD_VFN(0));
4015 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
4016 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
4017 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid));
4018 	c.physeqid_pkd = htobe32(0);
4019 	c.fetchszm_to_iqid =
4020 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
4021 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
4022 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
4023 	c.dcaen_to_eqsize =
4024 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
4025 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
4026 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
4027 		V_FW_EQ_CTRL_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
4028 		V_FW_EQ_CTRL_CMD_EQSIZE(qsize));
4029 	c.eqaddr = htobe64(eq->ba);
4030 
4031 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
4032 	if (rc != 0) {
4033 		device_printf(sc->dev,
4034 		    "failed to create control queue %d: %d\n", eq->tx_chan, rc);
4035 		return (rc);
4036 	}
4037 	eq->flags |= EQ_ALLOCATED;
4038 
4039 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
4040 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
4041 	if (cntxt_id >= sc->sge.eqmap_sz)
4042 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
4043 		cntxt_id, sc->sge.eqmap_sz - 1);
4044 	sc->sge.eqmap[cntxt_id] = eq;
4045 
4046 	return (rc);
4047 }
4048 
4049 static int
4050 eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
4051 {
4052 	int rc, cntxt_id;
4053 	struct fw_eq_eth_cmd c;
4054 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
4055 
4056 	bzero(&c, sizeof(c));
4057 
4058 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
4059 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
4060 	    V_FW_EQ_ETH_CMD_VFN(0));
4061 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
4062 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
4063 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
4064 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
4065 	c.fetchszm_to_iqid =
4066 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
4067 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
4068 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
4069 	c.dcaen_to_eqsize =
4070 	    htobe32(V_FW_EQ_ETH_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
4071 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
4072 		V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
4073 		V_FW_EQ_ETH_CMD_EQSIZE(qsize));
4074 	c.eqaddr = htobe64(eq->ba);
4075 
4076 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
4077 	if (rc != 0) {
4078 		device_printf(vi->dev,
4079 		    "failed to create Ethernet egress queue: %d\n", rc);
4080 		return (rc);
4081 	}
4082 	eq->flags |= EQ_ALLOCATED;
4083 
4084 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
4085 	eq->abs_id = G_FW_EQ_ETH_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
4086 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
4087 	if (cntxt_id >= sc->sge.eqmap_sz)
4088 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
4089 		cntxt_id, sc->sge.eqmap_sz - 1);
4090 	sc->sge.eqmap[cntxt_id] = eq;
4091 
4092 	return (rc);
4093 }
4094 
4095 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
4096 static int
4097 ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
4098 {
4099 	int rc, cntxt_id;
4100 	struct fw_eq_ofld_cmd c;
4101 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
4102 
4103 	bzero(&c, sizeof(c));
4104 
4105 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
4106 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
4107 	    V_FW_EQ_OFLD_CMD_VFN(0));
4108 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
4109 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
4110 	c.fetchszm_to_iqid =
4111 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
4112 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
4113 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
4114 	c.dcaen_to_eqsize =
4115 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
4116 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
4117 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
4118 		V_FW_EQ_OFLD_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
4119 		V_FW_EQ_OFLD_CMD_EQSIZE(qsize));
4120 	c.eqaddr = htobe64(eq->ba);
4121 
4122 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
4123 	if (rc != 0) {
4124 		device_printf(vi->dev,
4125 		    "failed to create egress queue for TCP offload: %d\n", rc);
4126 		return (rc);
4127 	}
4128 	eq->flags |= EQ_ALLOCATED;
4129 
4130 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
4131 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
4132 	if (cntxt_id >= sc->sge.eqmap_sz)
4133 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
4134 		cntxt_id, sc->sge.eqmap_sz - 1);
4135 	sc->sge.eqmap[cntxt_id] = eq;
4136 
4137 	return (rc);
4138 }
4139 #endif
4140 
4141 static int
4142 alloc_eq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
4143 {
4144 	int rc, qsize;
4145 	size_t len;
4146 
4147 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
4148 
4149 	qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
4150 	len = qsize * EQ_ESIZE;
4151 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
4152 	    &eq->ba, (void **)&eq->desc);
4153 	if (rc)
4154 		return (rc);
4155 
4156 	eq->pidx = eq->cidx = eq->dbidx = 0;
4157 	/* Note that equeqidx is not used with sge_wrq (OFLD/CTRL) queues. */
4158 	eq->equeqidx = 0;
4159 	eq->doorbells = sc->doorbells;
4160 
4161 	switch (eq->flags & EQ_TYPEMASK) {
4162 	case EQ_CTRL:
4163 		rc = ctrl_eq_alloc(sc, eq);
4164 		break;
4165 
4166 	case EQ_ETH:
4167 		rc = eth_eq_alloc(sc, vi, eq);
4168 		break;
4169 
4170 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
4171 	case EQ_OFLD:
4172 		rc = ofld_eq_alloc(sc, vi, eq);
4173 		break;
4174 #endif
4175 
4176 	default:
4177 		panic("%s: invalid eq type %d.", __func__,
4178 		    eq->flags & EQ_TYPEMASK);
4179 	}
4180 	if (rc != 0) {
4181 		device_printf(sc->dev,
4182 		    "failed to allocate egress queue(%d): %d\n",
4183 		    eq->flags & EQ_TYPEMASK, rc);
4184 	}
4185 
4186 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
4187 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
4188 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
4189 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
4190 		uint32_t mask = (1 << s_qpp) - 1;
4191 		volatile uint8_t *udb;
4192 
4193 		udb = sc->udbs_base + UDBS_DB_OFFSET;
4194 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
4195 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
4196 		if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
4197 	    		clrbit(&eq->doorbells, DOORBELL_WCWR);
4198 		else {
4199 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
4200 			eq->udb_qid = 0;
4201 		}
4202 		eq->udb = (volatile void *)udb;
4203 	}
4204 
4205 	return (rc);
4206 }
4207 
4208 static int
4209 free_eq(struct adapter *sc, struct sge_eq *eq)
4210 {
4211 	int rc;
4212 
4213 	if (eq->flags & EQ_ALLOCATED) {
4214 		switch (eq->flags & EQ_TYPEMASK) {
4215 		case EQ_CTRL:
4216 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
4217 			    eq->cntxt_id);
4218 			break;
4219 
4220 		case EQ_ETH:
4221 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
4222 			    eq->cntxt_id);
4223 			break;
4224 
4225 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
4226 		case EQ_OFLD:
4227 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
4228 			    eq->cntxt_id);
4229 			break;
4230 #endif
4231 
4232 		default:
4233 			panic("%s: invalid eq type %d.", __func__,
4234 			    eq->flags & EQ_TYPEMASK);
4235 		}
4236 		if (rc != 0) {
4237 			device_printf(sc->dev,
4238 			    "failed to free egress queue (%d): %d\n",
4239 			    eq->flags & EQ_TYPEMASK, rc);
4240 			return (rc);
4241 		}
4242 		eq->flags &= ~EQ_ALLOCATED;
4243 	}
4244 
4245 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
4246 
4247 	if (mtx_initialized(&eq->eq_lock))
4248 		mtx_destroy(&eq->eq_lock);
4249 
4250 	bzero(eq, sizeof(*eq));
4251 	return (0);
4252 }
4253 
4254 static int
4255 alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq,
4256     struct sysctl_oid *oid)
4257 {
4258 	int rc;
4259 	struct sysctl_ctx_list *ctx = vi ? &vi->ctx : &sc->ctx;
4260 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
4261 
4262 	rc = alloc_eq(sc, vi, &wrq->eq);
4263 	if (rc)
4264 		return (rc);
4265 
4266 	wrq->adapter = sc;
4267 	TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq);
4268 	TAILQ_INIT(&wrq->incomplete_wrs);
4269 	STAILQ_INIT(&wrq->wr_list);
4270 	wrq->nwr_pending = 0;
4271 	wrq->ndesc_needed = 0;
4272 
4273 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
4274 	    &wrq->eq.ba, "bus address of descriptor ring");
4275 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
4276 	    wrq->eq.sidx * EQ_ESIZE + sc->params.sge.spg_len,
4277 	    "desc ring size in bytes");
4278 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
4279 	    &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
4280 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
4281 	    &wrq->eq.cidx, 0, "consumer index");
4282 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
4283 	    &wrq->eq.pidx, 0, "producer index");
4284 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
4285 	    wrq->eq.sidx, "status page index");
4286 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD,
4287 	    &wrq->tx_wrs_direct, "# of work requests (direct)");
4288 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD,
4289 	    &wrq->tx_wrs_copied, "# of work requests (copied)");
4290 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_sspace", CTLFLAG_RD,
4291 	    &wrq->tx_wrs_ss, "# of work requests (copied from scratch space)");
4292 
4293 	return (rc);
4294 }
4295 
4296 static int
4297 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
4298 {
4299 	int rc;
4300 
4301 	rc = free_eq(sc, &wrq->eq);
4302 	if (rc)
4303 		return (rc);
4304 
4305 	bzero(wrq, sizeof(*wrq));
4306 	return (0);
4307 }
4308 
4309 static int
4310 alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx,
4311     struct sysctl_oid *oid)
4312 {
4313 	int rc;
4314 	struct port_info *pi = vi->pi;
4315 	struct adapter *sc = pi->adapter;
4316 	struct sge_eq *eq = &txq->eq;
4317 	struct txpkts *txp;
4318 	char name[16];
4319 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
4320 
4321 	rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx, can_resume_eth_tx,
4322 	    M_CXGBE, &eq->eq_lock, M_WAITOK);
4323 	if (rc != 0) {
4324 		device_printf(sc->dev, "failed to allocate mp_ring: %d\n", rc);
4325 		return (rc);
4326 	}
4327 
4328 	rc = alloc_eq(sc, vi, eq);
4329 	if (rc != 0) {
4330 		mp_ring_free(txq->r);
4331 		txq->r = NULL;
4332 		return (rc);
4333 	}
4334 
4335 	/* Can't fail after this point. */
4336 
4337 	if (idx == 0)
4338 		sc->sge.eq_base = eq->abs_id - eq->cntxt_id;
4339 	else
4340 		KASSERT(eq->cntxt_id + sc->sge.eq_base == eq->abs_id,
4341 		    ("eq_base mismatch"));
4342 	KASSERT(sc->sge.eq_base == 0 || sc->flags & IS_VF,
4343 	    ("PF with non-zero eq_base"));
4344 
4345 	TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq);
4346 	txq->ifp = vi->ifp;
4347 	txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK);
4348 	if (vi->flags & TX_USES_VM_WR)
4349 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
4350 		    V_TXPKT_INTF(pi->tx_chan));
4351 	else
4352 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
4353 		    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(sc->pf) |
4354 		    V_TXPKT_VF(vi->vin) | V_TXPKT_VF_VLD(vi->vfvld));
4355 	txq->tc_idx = -1;
4356 	txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE,
4357 	    M_ZERO | M_WAITOK);
4358 
4359 	txp = &txq->txp;
4360 	MPASS(nitems(txp->mb) >= sc->params.max_pkts_per_eth_tx_pkts_wr);
4361 	txq->txp.max_npkt = min(nitems(txp->mb),
4362 	    sc->params.max_pkts_per_eth_tx_pkts_wr);
4363 	if (vi->flags & TX_USES_VM_WR && !(sc->flags & IS_VF))
4364 		txq->txp.max_npkt--;
4365 
4366 	snprintf(name, sizeof(name), "%d", idx);
4367 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name,
4368 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "tx queue");
4369 	children = SYSCTL_CHILDREN(oid);
4370 
4371 	SYSCTL_ADD_UAUTO(&vi->ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
4372 	    &eq->ba, "bus address of descriptor ring");
4373 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
4374 	    eq->sidx * EQ_ESIZE + sc->params.sge.spg_len,
4375 	    "desc ring size in bytes");
4376 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
4377 	    &eq->abs_id, 0, "absolute id of the queue");
4378 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
4379 	    &eq->cntxt_id, 0, "SGE context id of the queue");
4380 	SYSCTL_ADD_U16(&vi->ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
4381 	    &eq->cidx, 0, "consumer index");
4382 	SYSCTL_ADD_U16(&vi->ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
4383 	    &eq->pidx, 0, "producer index");
4384 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
4385 	    eq->sidx, "status page index");
4386 
4387 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "tc",
4388 	    CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, vi, idx, sysctl_tc,
4389 	    "I", "traffic class (-1 means none)");
4390 
4391 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
4392 	    &txq->txcsum, "# of times hardware assisted with checksum");
4393 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_insertion",
4394 	    CTLFLAG_RD, &txq->vlan_insertion,
4395 	    "# of times hardware inserted 802.1Q tag");
4396 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
4397 	    &txq->tso_wrs, "# of TSO work requests");
4398 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
4399 	    &txq->imm_wrs, "# of work requests with immediate data");
4400 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
4401 	    &txq->sgl_wrs, "# of work requests with direct SGL");
4402 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
4403 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
4404 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_wrs",
4405 	    CTLFLAG_RD, &txq->txpkts0_wrs,
4406 	    "# of txpkts (type 0) work requests");
4407 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_wrs",
4408 	    CTLFLAG_RD, &txq->txpkts1_wrs,
4409 	    "# of txpkts (type 1) work requests");
4410 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_pkts",
4411 	    CTLFLAG_RD, &txq->txpkts0_pkts,
4412 	    "# of frames tx'd using type0 txpkts work requests");
4413 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_pkts",
4414 	    CTLFLAG_RD, &txq->txpkts1_pkts,
4415 	    "# of frames tx'd using type1 txpkts work requests");
4416 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts_flush",
4417 	    CTLFLAG_RD, &txq->txpkts_flush,
4418 	    "# of times txpkts had to be flushed out by an egress-update");
4419 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "raw_wrs", CTLFLAG_RD,
4420 	    &txq->raw_wrs, "# of raw work requests (non-packets)");
4421 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vxlan_tso_wrs",
4422 	    CTLFLAG_RD, &txq->vxlan_tso_wrs, "# of VXLAN TSO work requests");
4423 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vxlan_txcsum",
4424 	    CTLFLAG_RD, &txq->vxlan_txcsum,
4425 	    "# of times hardware assisted with inner checksums (VXLAN)");
4426 
4427 #ifdef KERN_TLS
4428 	if (is_ktls(sc)) {
4429 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4430 		    "kern_tls_records", CTLFLAG_RD, &txq->kern_tls_records,
4431 		    "# of NIC TLS records transmitted");
4432 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4433 		    "kern_tls_short", CTLFLAG_RD, &txq->kern_tls_short,
4434 		    "# of short NIC TLS records transmitted");
4435 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4436 		    "kern_tls_partial", CTLFLAG_RD, &txq->kern_tls_partial,
4437 		    "# of partial NIC TLS records transmitted");
4438 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4439 		    "kern_tls_full", CTLFLAG_RD, &txq->kern_tls_full,
4440 		    "# of full NIC TLS records transmitted");
4441 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4442 		    "kern_tls_octets", CTLFLAG_RD, &txq->kern_tls_octets,
4443 		    "# of payload octets in transmitted NIC TLS records");
4444 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4445 		    "kern_tls_waste", CTLFLAG_RD, &txq->kern_tls_waste,
4446 		    "# of octets DMAd but not transmitted in NIC TLS records");
4447 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4448 		    "kern_tls_options", CTLFLAG_RD, &txq->kern_tls_options,
4449 		    "# of NIC TLS options-only packets transmitted");
4450 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4451 		    "kern_tls_header", CTLFLAG_RD, &txq->kern_tls_header,
4452 		    "# of NIC TLS header-only packets transmitted");
4453 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4454 		    "kern_tls_fin", CTLFLAG_RD, &txq->kern_tls_fin,
4455 		    "# of NIC TLS FIN-only packets transmitted");
4456 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4457 		    "kern_tls_fin_short", CTLFLAG_RD, &txq->kern_tls_fin_short,
4458 		    "# of NIC TLS padded FIN packets on short TLS records");
4459 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4460 		    "kern_tls_cbc", CTLFLAG_RD, &txq->kern_tls_cbc,
4461 		    "# of NIC TLS sessions using AES-CBC");
4462 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4463 		    "kern_tls_gcm", CTLFLAG_RD, &txq->kern_tls_gcm,
4464 		    "# of NIC TLS sessions using AES-GCM");
4465 	}
4466 #endif
4467 	mp_ring_sysctls(txq->r, &vi->ctx, children);
4468 
4469 	return (0);
4470 }
4471 
4472 static int
4473 free_txq(struct vi_info *vi, struct sge_txq *txq)
4474 {
4475 	int rc;
4476 	struct adapter *sc = vi->adapter;
4477 	struct sge_eq *eq = &txq->eq;
4478 
4479 	rc = free_eq(sc, eq);
4480 	if (rc)
4481 		return (rc);
4482 
4483 	sglist_free(txq->gl);
4484 	free(txq->sdesc, M_CXGBE);
4485 	mp_ring_free(txq->r);
4486 
4487 	bzero(txq, sizeof(*txq));
4488 	return (0);
4489 }
4490 
4491 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
4492 static int
4493 alloc_ofld_txq(struct vi_info *vi, struct sge_ofld_txq *ofld_txq, int idx,
4494     struct sysctl_oid *oid)
4495 {
4496 	struct adapter *sc = vi->adapter;
4497 	struct sysctl_oid_list *children;
4498 	char name[16];
4499 	int rc;
4500 
4501 	children = SYSCTL_CHILDREN(oid);
4502 
4503 	snprintf(name, sizeof(name), "%d", idx);
4504 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name,
4505 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "offload tx queue");
4506 	children = SYSCTL_CHILDREN(oid);
4507 
4508 	rc = alloc_wrq(sc, vi, &ofld_txq->wrq, oid);
4509 	if (rc != 0)
4510 		return (rc);
4511 
4512 	ofld_txq->tx_iscsi_pdus = counter_u64_alloc(M_WAITOK);
4513 	ofld_txq->tx_iscsi_octets = counter_u64_alloc(M_WAITOK);
4514 	ofld_txq->tx_toe_tls_records = counter_u64_alloc(M_WAITOK);
4515 	ofld_txq->tx_toe_tls_octets = counter_u64_alloc(M_WAITOK);
4516 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO,
4517 	    "tx_iscsi_pdus", CTLFLAG_RD, &ofld_txq->tx_iscsi_pdus,
4518 	    "# of iSCSI PDUs transmitted");
4519 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO,
4520 	    "tx_iscsi_octets", CTLFLAG_RD, &ofld_txq->tx_iscsi_octets,
4521 	    "# of payload octets in transmitted iSCSI PDUs");
4522 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO,
4523 	    "tx_toe_tls_records", CTLFLAG_RD, &ofld_txq->tx_toe_tls_records,
4524 	    "# of TOE TLS records transmitted");
4525 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO,
4526 	    "tx_toe_tls_octets", CTLFLAG_RD, &ofld_txq->tx_toe_tls_octets,
4527 	    "# of payload octets in transmitted TOE TLS records");
4528 
4529 	return (rc);
4530 }
4531 
4532 static int
4533 free_ofld_txq(struct vi_info *vi, struct sge_ofld_txq *ofld_txq)
4534 {
4535 	struct adapter *sc = vi->adapter;
4536 	int rc;
4537 
4538 	rc = free_wrq(sc, &ofld_txq->wrq);
4539 	if (rc != 0)
4540 		return (rc);
4541 
4542 	counter_u64_free(ofld_txq->tx_iscsi_pdus);
4543 	counter_u64_free(ofld_txq->tx_iscsi_octets);
4544 	counter_u64_free(ofld_txq->tx_toe_tls_records);
4545 	counter_u64_free(ofld_txq->tx_toe_tls_octets);
4546 
4547 	bzero(ofld_txq, sizeof(*ofld_txq));
4548 	return (0);
4549 }
4550 #endif
4551 
4552 static void
4553 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
4554 {
4555 	bus_addr_t *ba = arg;
4556 
4557 	KASSERT(nseg == 1,
4558 	    ("%s meant for single segment mappings only.", __func__));
4559 
4560 	*ba = error ? 0 : segs->ds_addr;
4561 }
4562 
4563 static inline void
4564 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
4565 {
4566 	uint32_t n, v;
4567 
4568 	n = IDXDIFF(fl->pidx >> 3, fl->dbidx, fl->sidx);
4569 	MPASS(n > 0);
4570 
4571 	wmb();
4572 	v = fl->dbval | V_PIDX(n);
4573 	if (fl->udb)
4574 		*fl->udb = htole32(v);
4575 	else
4576 		t4_write_reg(sc, sc->sge_kdoorbell_reg, v);
4577 	IDXINCR(fl->dbidx, n, fl->sidx);
4578 }
4579 
4580 /*
4581  * Fills up the freelist by allocating up to 'n' buffers.  Buffers that are
4582  * recycled do not count towards this allocation budget.
4583  *
4584  * Returns non-zero to indicate that this freelist should be added to the list
4585  * of starving freelists.
4586  */
4587 static int
4588 refill_fl(struct adapter *sc, struct sge_fl *fl, int n)
4589 {
4590 	__be64 *d;
4591 	struct fl_sdesc *sd;
4592 	uintptr_t pa;
4593 	caddr_t cl;
4594 	struct rx_buf_info *rxb;
4595 	struct cluster_metadata *clm;
4596 	uint16_t max_pidx, zidx = fl->zidx;
4597 	uint16_t hw_cidx = fl->hw_cidx;		/* stable snapshot */
4598 
4599 	FL_LOCK_ASSERT_OWNED(fl);
4600 
4601 	/*
4602 	 * We always stop at the beginning of the hardware descriptor that's just
4603 	 * before the one with the hw cidx.  This is to avoid hw pidx = hw cidx,
4604 	 * which would mean an empty freelist to the chip.
4605 	 */
4606 	max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1;
4607 	if (fl->pidx == max_pidx * 8)
4608 		return (0);
4609 
4610 	d = &fl->desc[fl->pidx];
4611 	sd = &fl->sdesc[fl->pidx];
4612 	rxb = &sc->sge.rx_buf_info[zidx];
4613 
4614 	while (n > 0) {
4615 
4616 		if (sd->cl != NULL) {
4617 
4618 			if (sd->nmbuf == 0) {
4619 				/*
4620 				 * Fast recycle without involving any atomics on
4621 				 * the cluster's metadata (if the cluster has
4622 				 * metadata).  This happens when all frames
4623 				 * received in the cluster were small enough to
4624 				 * fit within a single mbuf each.
4625 				 */
4626 				fl->cl_fast_recycled++;
4627 				goto recycled;
4628 			}
4629 
4630 			/*
4631 			 * Cluster is guaranteed to have metadata.  Clusters
4632 			 * without metadata always take the fast recycle path
4633 			 * when they're recycled.
4634 			 */
4635 			clm = cl_metadata(sd);
4636 			MPASS(clm != NULL);
4637 
4638 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4639 				fl->cl_recycled++;
4640 				counter_u64_add(extfree_rels, 1);
4641 				goto recycled;
4642 			}
4643 			sd->cl = NULL;	/* gave up my reference */
4644 		}
4645 		MPASS(sd->cl == NULL);
4646 		cl = uma_zalloc(rxb->zone, M_NOWAIT);
4647 		if (__predict_false(cl == NULL)) {
4648 			if (zidx != fl->safe_zidx) {
4649 				zidx = fl->safe_zidx;
4650 				rxb = &sc->sge.rx_buf_info[zidx];
4651 				cl = uma_zalloc(rxb->zone, M_NOWAIT);
4652 			}
4653 			if (cl == NULL)
4654 				break;
4655 		}
4656 		fl->cl_allocated++;
4657 		n--;
4658 
4659 		pa = pmap_kextract((vm_offset_t)cl);
4660 		sd->cl = cl;
4661 		sd->zidx = zidx;
4662 
4663 		if (fl->flags & FL_BUF_PACKING) {
4664 			*d = htobe64(pa | rxb->hwidx2);
4665 			sd->moff = rxb->size2;
4666 		} else {
4667 			*d = htobe64(pa | rxb->hwidx1);
4668 			sd->moff = 0;
4669 		}
4670 recycled:
4671 		sd->nmbuf = 0;
4672 		d++;
4673 		sd++;
4674 		if (__predict_false((++fl->pidx & 7) == 0)) {
4675 			uint16_t pidx = fl->pidx >> 3;
4676 
4677 			if (__predict_false(pidx == fl->sidx)) {
4678 				fl->pidx = 0;
4679 				pidx = 0;
4680 				sd = fl->sdesc;
4681 				d = fl->desc;
4682 			}
4683 			if (n < 8 || pidx == max_pidx)
4684 				break;
4685 
4686 			if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4)
4687 				ring_fl_db(sc, fl);
4688 		}
4689 	}
4690 
4691 	if ((fl->pidx >> 3) != fl->dbidx)
4692 		ring_fl_db(sc, fl);
4693 
4694 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
4695 }
4696 
4697 /*
4698  * Attempt to refill all starving freelists.
4699  */
4700 static void
4701 refill_sfl(void *arg)
4702 {
4703 	struct adapter *sc = arg;
4704 	struct sge_fl *fl, *fl_temp;
4705 
4706 	mtx_assert(&sc->sfl_lock, MA_OWNED);
4707 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
4708 		FL_LOCK(fl);
4709 		refill_fl(sc, fl, 64);
4710 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
4711 			TAILQ_REMOVE(&sc->sfl, fl, link);
4712 			fl->flags &= ~FL_STARVING;
4713 		}
4714 		FL_UNLOCK(fl);
4715 	}
4716 
4717 	if (!TAILQ_EMPTY(&sc->sfl))
4718 		callout_schedule(&sc->sfl_callout, hz / 5);
4719 }
4720 
4721 static int
4722 alloc_fl_sdesc(struct sge_fl *fl)
4723 {
4724 
4725 	fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc), M_CXGBE,
4726 	    M_ZERO | M_WAITOK);
4727 
4728 	return (0);
4729 }
4730 
4731 static void
4732 free_fl_sdesc(struct adapter *sc, struct sge_fl *fl)
4733 {
4734 	struct fl_sdesc *sd;
4735 	struct cluster_metadata *clm;
4736 	int i;
4737 
4738 	sd = fl->sdesc;
4739 	for (i = 0; i < fl->sidx * 8; i++, sd++) {
4740 		if (sd->cl == NULL)
4741 			continue;
4742 
4743 		if (sd->nmbuf == 0)
4744 			uma_zfree(sc->sge.rx_buf_info[sd->zidx].zone, sd->cl);
4745 		else if (fl->flags & FL_BUF_PACKING) {
4746 			clm = cl_metadata(sd);
4747 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4748 				uma_zfree(sc->sge.rx_buf_info[sd->zidx].zone,
4749 				    sd->cl);
4750 				counter_u64_add(extfree_rels, 1);
4751 			}
4752 		}
4753 		sd->cl = NULL;
4754 	}
4755 
4756 	free(fl->sdesc, M_CXGBE);
4757 	fl->sdesc = NULL;
4758 }
4759 
4760 static inline void
4761 get_pkt_gl(struct mbuf *m, struct sglist *gl)
4762 {
4763 	int rc;
4764 
4765 	M_ASSERTPKTHDR(m);
4766 
4767 	sglist_reset(gl);
4768 	rc = sglist_append_mbuf(gl, m);
4769 	if (__predict_false(rc != 0)) {
4770 		panic("%s: mbuf %p (%d segs) was vetted earlier but now fails "
4771 		    "with %d.", __func__, m, mbuf_nsegs(m), rc);
4772 	}
4773 
4774 	KASSERT(gl->sg_nseg == mbuf_nsegs(m),
4775 	    ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m,
4776 	    mbuf_nsegs(m), gl->sg_nseg));
4777 #if 0	/* vm_wr not readily available here. */
4778 	KASSERT(gl->sg_nseg > 0 && gl->sg_nseg <= max_nsegs_allowed(m, vm_wr),
4779 	    ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__,
4780 		gl->sg_nseg, max_nsegs_allowed(m, vm_wr)));
4781 #endif
4782 }
4783 
4784 /*
4785  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
4786  */
4787 static inline u_int
4788 txpkt_len16(u_int nsegs, const u_int extra)
4789 {
4790 	u_int n;
4791 
4792 	MPASS(nsegs > 0);
4793 
4794 	nsegs--; /* first segment is part of ulptx_sgl */
4795 	n = extra + sizeof(struct fw_eth_tx_pkt_wr) +
4796 	    sizeof(struct cpl_tx_pkt_core) +
4797 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4798 
4799 	return (howmany(n, 16));
4800 }
4801 
4802 /*
4803  * len16 for a txpkt_vm WR with a GL.  Includes the firmware work
4804  * request header.
4805  */
4806 static inline u_int
4807 txpkt_vm_len16(u_int nsegs, const u_int extra)
4808 {
4809 	u_int n;
4810 
4811 	MPASS(nsegs > 0);
4812 
4813 	nsegs--; /* first segment is part of ulptx_sgl */
4814 	n = extra + sizeof(struct fw_eth_tx_pkt_vm_wr) +
4815 	    sizeof(struct cpl_tx_pkt_core) +
4816 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4817 
4818 	return (howmany(n, 16));
4819 }
4820 
4821 static inline void
4822 calculate_mbuf_len16(struct mbuf *m, bool vm_wr)
4823 {
4824 	const int lso = sizeof(struct cpl_tx_pkt_lso_core);
4825 	const int tnl_lso = sizeof(struct cpl_tx_tnl_lso);
4826 
4827 	if (vm_wr) {
4828 		if (needs_tso(m))
4829 			set_mbuf_len16(m, txpkt_vm_len16(mbuf_nsegs(m), lso));
4830 		else
4831 			set_mbuf_len16(m, txpkt_vm_len16(mbuf_nsegs(m), 0));
4832 		return;
4833 	}
4834 
4835 	if (needs_tso(m)) {
4836 		if (needs_vxlan_tso(m))
4837 			set_mbuf_len16(m, txpkt_len16(mbuf_nsegs(m), tnl_lso));
4838 		else
4839 			set_mbuf_len16(m, txpkt_len16(mbuf_nsegs(m), lso));
4840 	} else
4841 		set_mbuf_len16(m, txpkt_len16(mbuf_nsegs(m), 0));
4842 }
4843 
4844 /*
4845  * len16 for a txpkts type 0 WR with a GL.  Does not include the firmware work
4846  * request header.
4847  */
4848 static inline u_int
4849 txpkts0_len16(u_int nsegs)
4850 {
4851 	u_int n;
4852 
4853 	MPASS(nsegs > 0);
4854 
4855 	nsegs--; /* first segment is part of ulptx_sgl */
4856 	n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) +
4857 	    sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) +
4858 	    8 * ((3 * nsegs) / 2 + (nsegs & 1));
4859 
4860 	return (howmany(n, 16));
4861 }
4862 
4863 /*
4864  * len16 for a txpkts type 1 WR with a GL.  Does not include the firmware work
4865  * request header.
4866  */
4867 static inline u_int
4868 txpkts1_len16(void)
4869 {
4870 	u_int n;
4871 
4872 	n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl);
4873 
4874 	return (howmany(n, 16));
4875 }
4876 
4877 static inline u_int
4878 imm_payload(u_int ndesc)
4879 {
4880 	u_int n;
4881 
4882 	n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) -
4883 	    sizeof(struct cpl_tx_pkt_core);
4884 
4885 	return (n);
4886 }
4887 
4888 static inline uint64_t
4889 csum_to_ctrl(struct adapter *sc, struct mbuf *m)
4890 {
4891 	uint64_t ctrl;
4892 	int csum_type, l2hlen, l3hlen;
4893 	int x, y;
4894 	static const int csum_types[3][2] = {
4895 		{TX_CSUM_TCPIP, TX_CSUM_TCPIP6},
4896 		{TX_CSUM_UDPIP, TX_CSUM_UDPIP6},
4897 		{TX_CSUM_IP, 0}
4898 	};
4899 
4900 	M_ASSERTPKTHDR(m);
4901 
4902 	if (!needs_hwcsum(m))
4903 		return (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS);
4904 
4905 	MPASS(m->m_pkthdr.l2hlen >= ETHER_HDR_LEN);
4906 	MPASS(m->m_pkthdr.l3hlen >= sizeof(struct ip));
4907 
4908 	if (needs_vxlan_csum(m)) {
4909 		MPASS(m->m_pkthdr.l4hlen > 0);
4910 		MPASS(m->m_pkthdr.l5hlen > 0);
4911 		MPASS(m->m_pkthdr.inner_l2hlen >= ETHER_HDR_LEN);
4912 		MPASS(m->m_pkthdr.inner_l3hlen >= sizeof(struct ip));
4913 
4914 		l2hlen = m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen +
4915 		    m->m_pkthdr.l4hlen + m->m_pkthdr.l5hlen +
4916 		    m->m_pkthdr.inner_l2hlen - ETHER_HDR_LEN;
4917 		l3hlen = m->m_pkthdr.inner_l3hlen;
4918 	} else {
4919 		l2hlen = m->m_pkthdr.l2hlen - ETHER_HDR_LEN;
4920 		l3hlen = m->m_pkthdr.l3hlen;
4921 	}
4922 
4923 	ctrl = 0;
4924 	if (!needs_l3_csum(m))
4925 		ctrl |= F_TXPKT_IPCSUM_DIS;
4926 
4927 	if (m->m_pkthdr.csum_flags & (CSUM_IP_TCP | CSUM_INNER_IP_TCP |
4928 	    CSUM_IP6_TCP | CSUM_INNER_IP6_TCP))
4929 		x = 0;	/* TCP */
4930 	else if (m->m_pkthdr.csum_flags & (CSUM_IP_UDP | CSUM_INNER_IP_UDP |
4931 	    CSUM_IP6_UDP | CSUM_INNER_IP6_UDP))
4932 		x = 1;	/* UDP */
4933 	else
4934 		x = 2;
4935 
4936 	if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_IP_TCP | CSUM_IP_UDP |
4937 	    CSUM_INNER_IP | CSUM_INNER_IP_TCP | CSUM_INNER_IP_UDP))
4938 		y = 0;	/* IPv4 */
4939 	else {
4940 		MPASS(m->m_pkthdr.csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP |
4941 		    CSUM_INNER_IP6_TCP | CSUM_INNER_IP6_UDP));
4942 		y = 1;	/* IPv6 */
4943 	}
4944 	/*
4945 	 * needs_hwcsum returned true earlier so there must be some kind of
4946 	 * checksum to calculate.
4947 	 */
4948 	csum_type = csum_types[x][y];
4949 	MPASS(csum_type != 0);
4950 	if (csum_type == TX_CSUM_IP)
4951 		ctrl |= F_TXPKT_L4CSUM_DIS;
4952 	ctrl |= V_TXPKT_CSUM_TYPE(csum_type) | V_TXPKT_IPHDR_LEN(l3hlen);
4953 	if (chip_id(sc) <= CHELSIO_T5)
4954 		ctrl |= V_TXPKT_ETHHDR_LEN(l2hlen);
4955 	else
4956 		ctrl |= V_T6_TXPKT_ETHHDR_LEN(l2hlen);
4957 
4958 	return (ctrl);
4959 }
4960 
4961 static inline void *
4962 write_lso_cpl(void *cpl, struct mbuf *m0)
4963 {
4964 	struct cpl_tx_pkt_lso_core *lso;
4965 	uint32_t ctrl;
4966 
4967 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4968 	    m0->m_pkthdr.l4hlen > 0,
4969 	    ("%s: mbuf %p needs TSO but missing header lengths",
4970 		__func__, m0));
4971 
4972 	ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) |
4973 	    F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE |
4974 	    V_LSO_ETHHDR_LEN((m0->m_pkthdr.l2hlen - ETHER_HDR_LEN) >> 2) |
4975 	    V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) |
4976 	    V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4977 	if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4978 		ctrl |= F_LSO_IPV6;
4979 
4980 	lso = cpl;
4981 	lso->lso_ctrl = htobe32(ctrl);
4982 	lso->ipid_ofst = htobe16(0);
4983 	lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4984 	lso->seqno_offset = htobe32(0);
4985 	lso->len = htobe32(m0->m_pkthdr.len);
4986 
4987 	return (lso + 1);
4988 }
4989 
4990 static void *
4991 write_tnl_lso_cpl(void *cpl, struct mbuf *m0)
4992 {
4993 	struct cpl_tx_tnl_lso *tnl_lso = cpl;
4994 	uint32_t ctrl;
4995 
4996 	KASSERT(m0->m_pkthdr.inner_l2hlen > 0 &&
4997 	    m0->m_pkthdr.inner_l3hlen > 0 && m0->m_pkthdr.inner_l4hlen > 0 &&
4998 	    m0->m_pkthdr.inner_l5hlen > 0,
4999 	    ("%s: mbuf %p needs VXLAN_TSO but missing inner header lengths",
5000 		__func__, m0));
5001 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
5002 	    m0->m_pkthdr.l4hlen > 0 && m0->m_pkthdr.l5hlen > 0,
5003 	    ("%s: mbuf %p needs VXLAN_TSO but missing outer header lengths",
5004 		__func__, m0));
5005 
5006 	/* Outer headers. */
5007 	ctrl = V_CPL_TX_TNL_LSO_OPCODE(CPL_TX_TNL_LSO) |
5008 	    F_CPL_TX_TNL_LSO_FIRST | F_CPL_TX_TNL_LSO_LAST |
5009 	    V_CPL_TX_TNL_LSO_ETHHDRLENOUT(
5010 		(m0->m_pkthdr.l2hlen - ETHER_HDR_LEN) >> 2) |
5011 	    V_CPL_TX_TNL_LSO_IPHDRLENOUT(m0->m_pkthdr.l3hlen >> 2) |
5012 	    F_CPL_TX_TNL_LSO_IPLENSETOUT;
5013 	if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
5014 		ctrl |= F_CPL_TX_TNL_LSO_IPV6OUT;
5015 	else {
5016 		ctrl |= F_CPL_TX_TNL_LSO_IPHDRCHKOUT |
5017 		    F_CPL_TX_TNL_LSO_IPIDINCOUT;
5018 	}
5019 	tnl_lso->op_to_IpIdSplitOut = htobe32(ctrl);
5020 	tnl_lso->IpIdOffsetOut = 0;
5021 	tnl_lso->UdpLenSetOut_to_TnlHdrLen =
5022 		htobe16(F_CPL_TX_TNL_LSO_UDPCHKCLROUT |
5023 		    F_CPL_TX_TNL_LSO_UDPLENSETOUT |
5024 		    V_CPL_TX_TNL_LSO_TNLHDRLEN(m0->m_pkthdr.l2hlen +
5025 			m0->m_pkthdr.l3hlen + m0->m_pkthdr.l4hlen +
5026 			m0->m_pkthdr.l5hlen) |
5027 		    V_CPL_TX_TNL_LSO_TNLTYPE(TX_TNL_TYPE_VXLAN));
5028 	tnl_lso->r1 = 0;
5029 
5030 	/* Inner headers. */
5031 	ctrl = V_CPL_TX_TNL_LSO_ETHHDRLEN(
5032 	    (m0->m_pkthdr.inner_l2hlen - ETHER_HDR_LEN) >> 2) |
5033 	    V_CPL_TX_TNL_LSO_IPHDRLEN(m0->m_pkthdr.inner_l3hlen >> 2) |
5034 	    V_CPL_TX_TNL_LSO_TCPHDRLEN(m0->m_pkthdr.inner_l4hlen >> 2);
5035 	if (m0->m_pkthdr.inner_l3hlen == sizeof(struct ip6_hdr))
5036 		ctrl |= F_CPL_TX_TNL_LSO_IPV6;
5037 	tnl_lso->Flow_to_TcpHdrLen = htobe32(ctrl);
5038 	tnl_lso->IpIdOffset = 0;
5039 	tnl_lso->IpIdSplit_to_Mss =
5040 	    htobe16(V_CPL_TX_TNL_LSO_MSS(m0->m_pkthdr.tso_segsz));
5041 	tnl_lso->TCPSeqOffset = 0;
5042 	tnl_lso->EthLenOffset_Size =
5043 	    htobe32(V_CPL_TX_TNL_LSO_SIZE(m0->m_pkthdr.len));
5044 
5045 	return (tnl_lso + 1);
5046 }
5047 
5048 #define VM_TX_L2HDR_LEN	16	/* ethmacdst to vlantci */
5049 
5050 /*
5051  * Write a VM txpkt WR for this packet to the hardware descriptors, update the
5052  * software descriptor, and advance the pidx.  It is guaranteed that enough
5053  * descriptors are available.
5054  *
5055  * The return value is the # of hardware descriptors used.
5056  */
5057 static u_int
5058 write_txpkt_vm_wr(struct adapter *sc, struct sge_txq *txq, struct mbuf *m0)
5059 {
5060 	struct sge_eq *eq;
5061 	struct fw_eth_tx_pkt_vm_wr *wr;
5062 	struct tx_sdesc *txsd;
5063 	struct cpl_tx_pkt_core *cpl;
5064 	uint32_t ctrl;	/* used in many unrelated places */
5065 	uint64_t ctrl1;
5066 	int len16, ndesc, pktlen, nsegs;
5067 	caddr_t dst;
5068 
5069 	TXQ_LOCK_ASSERT_OWNED(txq);
5070 	M_ASSERTPKTHDR(m0);
5071 
5072 	len16 = mbuf_len16(m0);
5073 	nsegs = mbuf_nsegs(m0);
5074 	pktlen = m0->m_pkthdr.len;
5075 	ctrl = sizeof(struct cpl_tx_pkt_core);
5076 	if (needs_tso(m0))
5077 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
5078 	ndesc = tx_len16_to_desc(len16);
5079 
5080 	/* Firmware work request header */
5081 	eq = &txq->eq;
5082 	wr = (void *)&eq->desc[eq->pidx];
5083 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_VM_WR) |
5084 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
5085 
5086 	ctrl = V_FW_WR_LEN16(len16);
5087 	wr->equiq_to_len16 = htobe32(ctrl);
5088 	wr->r3[0] = 0;
5089 	wr->r3[1] = 0;
5090 
5091 	/*
5092 	 * Copy over ethmacdst, ethmacsrc, ethtype, and vlantci.
5093 	 * vlantci is ignored unless the ethtype is 0x8100, so it's
5094 	 * simpler to always copy it rather than making it
5095 	 * conditional.  Also, it seems that we do not have to set
5096 	 * vlantci or fake the ethtype when doing VLAN tag insertion.
5097 	 */
5098 	m_copydata(m0, 0, VM_TX_L2HDR_LEN, wr->ethmacdst);
5099 
5100 	if (needs_tso(m0)) {
5101 		cpl = write_lso_cpl(wr + 1, m0);
5102 		txq->tso_wrs++;
5103 	} else
5104 		cpl = (void *)(wr + 1);
5105 
5106 	/* Checksum offload */
5107 	ctrl1 = csum_to_ctrl(sc, m0);
5108 	if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS))
5109 		txq->txcsum++;	/* some hardware assistance provided */
5110 
5111 	/* VLAN tag insertion */
5112 	if (needs_vlan_insertion(m0)) {
5113 		ctrl1 |= F_TXPKT_VLAN_VLD |
5114 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
5115 		txq->vlan_insertion++;
5116 	}
5117 
5118 	/* CPL header */
5119 	cpl->ctrl0 = txq->cpl_ctrl0;
5120 	cpl->pack = 0;
5121 	cpl->len = htobe16(pktlen);
5122 	cpl->ctrl1 = htobe64(ctrl1);
5123 
5124 	/* SGL */
5125 	dst = (void *)(cpl + 1);
5126 
5127 	/*
5128 	 * A packet using TSO will use up an entire descriptor for the
5129 	 * firmware work request header, LSO CPL, and TX_PKT_XT CPL.
5130 	 * If this descriptor is the last descriptor in the ring, wrap
5131 	 * around to the front of the ring explicitly for the start of
5132 	 * the sgl.
5133 	 */
5134 	if (dst == (void *)&eq->desc[eq->sidx]) {
5135 		dst = (void *)&eq->desc[0];
5136 		write_gl_to_txd(txq, m0, &dst, 0);
5137 	} else
5138 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
5139 	txq->sgl_wrs++;
5140 	txq->txpkt_wrs++;
5141 
5142 	txsd = &txq->sdesc[eq->pidx];
5143 	txsd->m = m0;
5144 	txsd->desc_used = ndesc;
5145 
5146 	return (ndesc);
5147 }
5148 
5149 /*
5150  * Write a raw WR to the hardware descriptors, update the software
5151  * descriptor, and advance the pidx.  It is guaranteed that enough
5152  * descriptors are available.
5153  *
5154  * The return value is the # of hardware descriptors used.
5155  */
5156 static u_int
5157 write_raw_wr(struct sge_txq *txq, void *wr, struct mbuf *m0, u_int available)
5158 {
5159 	struct sge_eq *eq = &txq->eq;
5160 	struct tx_sdesc *txsd;
5161 	struct mbuf *m;
5162 	caddr_t dst;
5163 	int len16, ndesc;
5164 
5165 	len16 = mbuf_len16(m0);
5166 	ndesc = tx_len16_to_desc(len16);
5167 	MPASS(ndesc <= available);
5168 
5169 	dst = wr;
5170 	for (m = m0; m != NULL; m = m->m_next)
5171 		copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
5172 
5173 	txq->raw_wrs++;
5174 
5175 	txsd = &txq->sdesc[eq->pidx];
5176 	txsd->m = m0;
5177 	txsd->desc_used = ndesc;
5178 
5179 	return (ndesc);
5180 }
5181 
5182 /*
5183  * Write a txpkt WR for this packet to the hardware descriptors, update the
5184  * software descriptor, and advance the pidx.  It is guaranteed that enough
5185  * descriptors are available.
5186  *
5187  * The return value is the # of hardware descriptors used.
5188  */
5189 static u_int
5190 write_txpkt_wr(struct adapter *sc, struct sge_txq *txq, struct mbuf *m0,
5191     u_int available)
5192 {
5193 	struct sge_eq *eq;
5194 	struct fw_eth_tx_pkt_wr *wr;
5195 	struct tx_sdesc *txsd;
5196 	struct cpl_tx_pkt_core *cpl;
5197 	uint32_t ctrl;	/* used in many unrelated places */
5198 	uint64_t ctrl1;
5199 	int len16, ndesc, pktlen, nsegs;
5200 	caddr_t dst;
5201 
5202 	TXQ_LOCK_ASSERT_OWNED(txq);
5203 	M_ASSERTPKTHDR(m0);
5204 
5205 	len16 = mbuf_len16(m0);
5206 	nsegs = mbuf_nsegs(m0);
5207 	pktlen = m0->m_pkthdr.len;
5208 	ctrl = sizeof(struct cpl_tx_pkt_core);
5209 	if (needs_tso(m0)) {
5210 		if (needs_vxlan_tso(m0))
5211 			ctrl += sizeof(struct cpl_tx_tnl_lso);
5212 		else
5213 			ctrl += sizeof(struct cpl_tx_pkt_lso_core);
5214 	} else if (!(mbuf_cflags(m0) & MC_NOMAP) && pktlen <= imm_payload(2) &&
5215 	    available >= 2) {
5216 		/* Immediate data.  Recalculate len16 and set nsegs to 0. */
5217 		ctrl += pktlen;
5218 		len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) +
5219 		    sizeof(struct cpl_tx_pkt_core) + pktlen, 16);
5220 		nsegs = 0;
5221 	}
5222 	ndesc = tx_len16_to_desc(len16);
5223 	MPASS(ndesc <= available);
5224 
5225 	/* Firmware work request header */
5226 	eq = &txq->eq;
5227 	wr = (void *)&eq->desc[eq->pidx];
5228 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
5229 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
5230 
5231 	ctrl = V_FW_WR_LEN16(len16);
5232 	wr->equiq_to_len16 = htobe32(ctrl);
5233 	wr->r3 = 0;
5234 
5235 	if (needs_tso(m0)) {
5236 		if (needs_vxlan_tso(m0)) {
5237 			cpl = write_tnl_lso_cpl(wr + 1, m0);
5238 			txq->vxlan_tso_wrs++;
5239 		} else {
5240 			cpl = write_lso_cpl(wr + 1, m0);
5241 			txq->tso_wrs++;
5242 		}
5243 	} else
5244 		cpl = (void *)(wr + 1);
5245 
5246 	/* Checksum offload */
5247 	ctrl1 = csum_to_ctrl(sc, m0);
5248 	if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS)) {
5249 		/* some hardware assistance provided */
5250 		if (needs_vxlan_csum(m0))
5251 			txq->vxlan_txcsum++;
5252 		else
5253 			txq->txcsum++;
5254 	}
5255 
5256 	/* VLAN tag insertion */
5257 	if (needs_vlan_insertion(m0)) {
5258 		ctrl1 |= F_TXPKT_VLAN_VLD |
5259 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
5260 		txq->vlan_insertion++;
5261 	}
5262 
5263 	/* CPL header */
5264 	cpl->ctrl0 = txq->cpl_ctrl0;
5265 	cpl->pack = 0;
5266 	cpl->len = htobe16(pktlen);
5267 	cpl->ctrl1 = htobe64(ctrl1);
5268 
5269 	/* SGL */
5270 	dst = (void *)(cpl + 1);
5271 	if (__predict_false((uintptr_t)dst == (uintptr_t)&eq->desc[eq->sidx]))
5272 		dst = (caddr_t)&eq->desc[0];
5273 	if (nsegs > 0) {
5274 
5275 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
5276 		txq->sgl_wrs++;
5277 	} else {
5278 		struct mbuf *m;
5279 
5280 		for (m = m0; m != NULL; m = m->m_next) {
5281 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
5282 #ifdef INVARIANTS
5283 			pktlen -= m->m_len;
5284 #endif
5285 		}
5286 #ifdef INVARIANTS
5287 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
5288 #endif
5289 		txq->imm_wrs++;
5290 	}
5291 
5292 	txq->txpkt_wrs++;
5293 
5294 	txsd = &txq->sdesc[eq->pidx];
5295 	txsd->m = m0;
5296 	txsd->desc_used = ndesc;
5297 
5298 	return (ndesc);
5299 }
5300 
5301 static inline bool
5302 cmp_l2hdr(struct txpkts *txp, struct mbuf *m)
5303 {
5304 	int len;
5305 
5306 	MPASS(txp->npkt > 0);
5307 	MPASS(m->m_len >= VM_TX_L2HDR_LEN);
5308 
5309 	if (txp->ethtype == be16toh(ETHERTYPE_VLAN))
5310 		len = VM_TX_L2HDR_LEN;
5311 	else
5312 		len = sizeof(struct ether_header);
5313 
5314 	return (memcmp(m->m_data, &txp->ethmacdst[0], len) != 0);
5315 }
5316 
5317 static inline void
5318 save_l2hdr(struct txpkts *txp, struct mbuf *m)
5319 {
5320 	MPASS(m->m_len >= VM_TX_L2HDR_LEN);
5321 
5322 	memcpy(&txp->ethmacdst[0], mtod(m, const void *), VM_TX_L2HDR_LEN);
5323 }
5324 
5325 static int
5326 add_to_txpkts_vf(struct adapter *sc, struct sge_txq *txq, struct mbuf *m,
5327     int avail, bool *send)
5328 {
5329 	struct txpkts *txp = &txq->txp;
5330 
5331 	/* Cannot have TSO and coalesce at the same time. */
5332 	if (cannot_use_txpkts(m)) {
5333 cannot_coalesce:
5334 		*send = txp->npkt > 0;
5335 		return (EINVAL);
5336 	}
5337 
5338 	/* VF allows coalescing of type 1 (1 GL) only */
5339 	if (mbuf_nsegs(m) > 1)
5340 		goto cannot_coalesce;
5341 
5342 	*send = false;
5343 	if (txp->npkt > 0) {
5344 		MPASS(tx_len16_to_desc(txp->len16) <= avail);
5345 		MPASS(txp->npkt < txp->max_npkt);
5346 		MPASS(txp->wr_type == 1);	/* VF supports type 1 only */
5347 
5348 		if (tx_len16_to_desc(txp->len16 + txpkts1_len16()) > avail) {
5349 retry_after_send:
5350 			*send = true;
5351 			return (EAGAIN);
5352 		}
5353 		if (m->m_pkthdr.len + txp->plen > 65535)
5354 			goto retry_after_send;
5355 		if (cmp_l2hdr(txp, m))
5356 			goto retry_after_send;
5357 
5358 		txp->len16 += txpkts1_len16();
5359 		txp->plen += m->m_pkthdr.len;
5360 		txp->mb[txp->npkt++] = m;
5361 		if (txp->npkt == txp->max_npkt)
5362 			*send = true;
5363 	} else {
5364 		txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_vm_wr), 16) +
5365 		    txpkts1_len16();
5366 		if (tx_len16_to_desc(txp->len16) > avail)
5367 			goto cannot_coalesce;
5368 		txp->npkt = 1;
5369 		txp->wr_type = 1;
5370 		txp->plen = m->m_pkthdr.len;
5371 		txp->mb[0] = m;
5372 		save_l2hdr(txp, m);
5373 	}
5374 	return (0);
5375 }
5376 
5377 static int
5378 add_to_txpkts_pf(struct adapter *sc, struct sge_txq *txq, struct mbuf *m,
5379     int avail, bool *send)
5380 {
5381 	struct txpkts *txp = &txq->txp;
5382 	int nsegs;
5383 
5384 	MPASS(!(sc->flags & IS_VF));
5385 
5386 	/* Cannot have TSO and coalesce at the same time. */
5387 	if (cannot_use_txpkts(m)) {
5388 cannot_coalesce:
5389 		*send = txp->npkt > 0;
5390 		return (EINVAL);
5391 	}
5392 
5393 	*send = false;
5394 	nsegs = mbuf_nsegs(m);
5395 	if (txp->npkt == 0) {
5396 		if (m->m_pkthdr.len > 65535)
5397 			goto cannot_coalesce;
5398 		if (nsegs > 1) {
5399 			txp->wr_type = 0;
5400 			txp->len16 =
5401 			    howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) +
5402 			    txpkts0_len16(nsegs);
5403 		} else {
5404 			txp->wr_type = 1;
5405 			txp->len16 =
5406 			    howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) +
5407 			    txpkts1_len16();
5408 		}
5409 		if (tx_len16_to_desc(txp->len16) > avail)
5410 			goto cannot_coalesce;
5411 		txp->npkt = 1;
5412 		txp->plen = m->m_pkthdr.len;
5413 		txp->mb[0] = m;
5414 	} else {
5415 		MPASS(tx_len16_to_desc(txp->len16) <= avail);
5416 		MPASS(txp->npkt < txp->max_npkt);
5417 
5418 		if (m->m_pkthdr.len + txp->plen > 65535) {
5419 retry_after_send:
5420 			*send = true;
5421 			return (EAGAIN);
5422 		}
5423 
5424 		MPASS(txp->wr_type == 0 || txp->wr_type == 1);
5425 		if (txp->wr_type == 0) {
5426 			if (tx_len16_to_desc(txp->len16 +
5427 			    txpkts0_len16(nsegs)) > min(avail, SGE_MAX_WR_NDESC))
5428 				goto retry_after_send;
5429 			txp->len16 += txpkts0_len16(nsegs);
5430 		} else {
5431 			if (nsegs != 1)
5432 				goto retry_after_send;
5433 			if (tx_len16_to_desc(txp->len16 + txpkts1_len16()) >
5434 			    avail)
5435 				goto retry_after_send;
5436 			txp->len16 += txpkts1_len16();
5437 		}
5438 
5439 		txp->plen += m->m_pkthdr.len;
5440 		txp->mb[txp->npkt++] = m;
5441 		if (txp->npkt == txp->max_npkt)
5442 			*send = true;
5443 	}
5444 	return (0);
5445 }
5446 
5447 /*
5448  * Write a txpkts WR for the packets in txp to the hardware descriptors, update
5449  * the software descriptor, and advance the pidx.  It is guaranteed that enough
5450  * descriptors are available.
5451  *
5452  * The return value is the # of hardware descriptors used.
5453  */
5454 static u_int
5455 write_txpkts_wr(struct adapter *sc, struct sge_txq *txq)
5456 {
5457 	const struct txpkts *txp = &txq->txp;
5458 	struct sge_eq *eq = &txq->eq;
5459 	struct fw_eth_tx_pkts_wr *wr;
5460 	struct tx_sdesc *txsd;
5461 	struct cpl_tx_pkt_core *cpl;
5462 	uint64_t ctrl1;
5463 	int ndesc, i, checkwrap;
5464 	struct mbuf *m, *last;
5465 	void *flitp;
5466 
5467 	TXQ_LOCK_ASSERT_OWNED(txq);
5468 	MPASS(txp->npkt > 0);
5469 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
5470 
5471 	wr = (void *)&eq->desc[eq->pidx];
5472 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
5473 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(txp->len16));
5474 	wr->plen = htobe16(txp->plen);
5475 	wr->npkt = txp->npkt;
5476 	wr->r3 = 0;
5477 	wr->type = txp->wr_type;
5478 	flitp = wr + 1;
5479 
5480 	/*
5481 	 * At this point we are 16B into a hardware descriptor.  If checkwrap is
5482 	 * set then we know the WR is going to wrap around somewhere.  We'll
5483 	 * check for that at appropriate points.
5484 	 */
5485 	ndesc = tx_len16_to_desc(txp->len16);
5486 	last = NULL;
5487 	checkwrap = eq->sidx - ndesc < eq->pidx;
5488 	for (i = 0; i < txp->npkt; i++) {
5489 		m = txp->mb[i];
5490 		if (txp->wr_type == 0) {
5491 			struct ulp_txpkt *ulpmc;
5492 			struct ulptx_idata *ulpsc;
5493 
5494 			/* ULP master command */
5495 			ulpmc = flitp;
5496 			ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
5497 			    V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid));
5498 			ulpmc->len = htobe32(txpkts0_len16(mbuf_nsegs(m)));
5499 
5500 			/* ULP subcommand */
5501 			ulpsc = (void *)(ulpmc + 1);
5502 			ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
5503 			    F_ULP_TX_SC_MORE);
5504 			ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
5505 
5506 			cpl = (void *)(ulpsc + 1);
5507 			if (checkwrap &&
5508 			    (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx])
5509 				cpl = (void *)&eq->desc[0];
5510 		} else {
5511 			cpl = flitp;
5512 		}
5513 
5514 		/* Checksum offload */
5515 		ctrl1 = csum_to_ctrl(sc, m);
5516 		if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS)) {
5517 			/* some hardware assistance provided */
5518 			if (needs_vxlan_csum(m))
5519 				txq->vxlan_txcsum++;
5520 			else
5521 				txq->txcsum++;
5522 		}
5523 
5524 		/* VLAN tag insertion */
5525 		if (needs_vlan_insertion(m)) {
5526 			ctrl1 |= F_TXPKT_VLAN_VLD |
5527 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
5528 			txq->vlan_insertion++;
5529 		}
5530 
5531 		/* CPL header */
5532 		cpl->ctrl0 = txq->cpl_ctrl0;
5533 		cpl->pack = 0;
5534 		cpl->len = htobe16(m->m_pkthdr.len);
5535 		cpl->ctrl1 = htobe64(ctrl1);
5536 
5537 		flitp = cpl + 1;
5538 		if (checkwrap &&
5539 		    (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
5540 			flitp = (void *)&eq->desc[0];
5541 
5542 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap);
5543 
5544 		if (last != NULL)
5545 			last->m_nextpkt = m;
5546 		last = m;
5547 	}
5548 
5549 	txq->sgl_wrs++;
5550 	if (txp->wr_type == 0) {
5551 		txq->txpkts0_pkts += txp->npkt;
5552 		txq->txpkts0_wrs++;
5553 	} else {
5554 		txq->txpkts1_pkts += txp->npkt;
5555 		txq->txpkts1_wrs++;
5556 	}
5557 
5558 	txsd = &txq->sdesc[eq->pidx];
5559 	txsd->m = txp->mb[0];
5560 	txsd->desc_used = ndesc;
5561 
5562 	return (ndesc);
5563 }
5564 
5565 static u_int
5566 write_txpkts_vm_wr(struct adapter *sc, struct sge_txq *txq)
5567 {
5568 	const struct txpkts *txp = &txq->txp;
5569 	struct sge_eq *eq = &txq->eq;
5570 	struct fw_eth_tx_pkts_vm_wr *wr;
5571 	struct tx_sdesc *txsd;
5572 	struct cpl_tx_pkt_core *cpl;
5573 	uint64_t ctrl1;
5574 	int ndesc, i;
5575 	struct mbuf *m, *last;
5576 	void *flitp;
5577 
5578 	TXQ_LOCK_ASSERT_OWNED(txq);
5579 	MPASS(txp->npkt > 0);
5580 	MPASS(txp->wr_type == 1);	/* VF supports type 1 only */
5581 	MPASS(txp->mb[0] != NULL);
5582 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
5583 
5584 	wr = (void *)&eq->desc[eq->pidx];
5585 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_VM_WR));
5586 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(txp->len16));
5587 	wr->r3 = 0;
5588 	wr->plen = htobe16(txp->plen);
5589 	wr->npkt = txp->npkt;
5590 	wr->r4 = 0;
5591 	memcpy(&wr->ethmacdst[0], &txp->ethmacdst[0], 16);
5592 	flitp = wr + 1;
5593 
5594 	/*
5595 	 * At this point we are 32B into a hardware descriptor.  Each mbuf in
5596 	 * the WR will take 32B so we check for the end of the descriptor ring
5597 	 * before writing odd mbufs (mb[1], 3, 5, ..)
5598 	 */
5599 	ndesc = tx_len16_to_desc(txp->len16);
5600 	last = NULL;
5601 	for (i = 0; i < txp->npkt; i++) {
5602 		m = txp->mb[i];
5603 		if (i & 1 && (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
5604 			flitp = &eq->desc[0];
5605 		cpl = flitp;
5606 
5607 		/* Checksum offload */
5608 		ctrl1 = csum_to_ctrl(sc, m);
5609 		if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS))
5610 			txq->txcsum++;	/* some hardware assistance provided */
5611 
5612 		/* VLAN tag insertion */
5613 		if (needs_vlan_insertion(m)) {
5614 			ctrl1 |= F_TXPKT_VLAN_VLD |
5615 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
5616 			txq->vlan_insertion++;
5617 		}
5618 
5619 		/* CPL header */
5620 		cpl->ctrl0 = txq->cpl_ctrl0;
5621 		cpl->pack = 0;
5622 		cpl->len = htobe16(m->m_pkthdr.len);
5623 		cpl->ctrl1 = htobe64(ctrl1);
5624 
5625 		flitp = cpl + 1;
5626 		MPASS(mbuf_nsegs(m) == 1);
5627 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), 0);
5628 
5629 		if (last != NULL)
5630 			last->m_nextpkt = m;
5631 		last = m;
5632 	}
5633 
5634 	txq->sgl_wrs++;
5635 	txq->txpkts1_pkts += txp->npkt;
5636 	txq->txpkts1_wrs++;
5637 
5638 	txsd = &txq->sdesc[eq->pidx];
5639 	txsd->m = txp->mb[0];
5640 	txsd->desc_used = ndesc;
5641 
5642 	return (ndesc);
5643 }
5644 
5645 /*
5646  * If the SGL ends on an address that is not 16 byte aligned, this function will
5647  * add a 0 filled flit at the end.
5648  */
5649 static void
5650 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap)
5651 {
5652 	struct sge_eq *eq = &txq->eq;
5653 	struct sglist *gl = txq->gl;
5654 	struct sglist_seg *seg;
5655 	__be64 *flitp, *wrap;
5656 	struct ulptx_sgl *usgl;
5657 	int i, nflits, nsegs;
5658 
5659 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
5660 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
5661 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
5662 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
5663 
5664 	get_pkt_gl(m, gl);
5665 	nsegs = gl->sg_nseg;
5666 	MPASS(nsegs > 0);
5667 
5668 	nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
5669 	flitp = (__be64 *)(*to);
5670 	wrap = (__be64 *)(&eq->desc[eq->sidx]);
5671 	seg = &gl->sg_segs[0];
5672 	usgl = (void *)flitp;
5673 
5674 	/*
5675 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
5676 	 * ring, so we're at least 16 bytes away from the status page.  There is
5677 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
5678 	 */
5679 
5680 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
5681 	    V_ULPTX_NSGE(nsegs));
5682 	usgl->len0 = htobe32(seg->ss_len);
5683 	usgl->addr0 = htobe64(seg->ss_paddr);
5684 	seg++;
5685 
5686 	if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) {
5687 
5688 		/* Won't wrap around at all */
5689 
5690 		for (i = 0; i < nsegs - 1; i++, seg++) {
5691 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
5692 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
5693 		}
5694 		if (i & 1)
5695 			usgl->sge[i / 2].len[1] = htobe32(0);
5696 		flitp += nflits;
5697 	} else {
5698 
5699 		/* Will wrap somewhere in the rest of the SGL */
5700 
5701 		/* 2 flits already written, write the rest flit by flit */
5702 		flitp = (void *)(usgl + 1);
5703 		for (i = 0; i < nflits - 2; i++) {
5704 			if (flitp == wrap)
5705 				flitp = (void *)eq->desc;
5706 			*flitp++ = get_flit(seg, nsegs - 1, i);
5707 		}
5708 	}
5709 
5710 	if (nflits & 1) {
5711 		MPASS(((uintptr_t)flitp) & 0xf);
5712 		*flitp++ = 0;
5713 	}
5714 
5715 	MPASS((((uintptr_t)flitp) & 0xf) == 0);
5716 	if (__predict_false(flitp == wrap))
5717 		*to = (void *)eq->desc;
5718 	else
5719 		*to = (void *)flitp;
5720 }
5721 
5722 static inline void
5723 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
5724 {
5725 
5726 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
5727 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
5728 
5729 	if (__predict_true((uintptr_t)(*to) + len <=
5730 	    (uintptr_t)&eq->desc[eq->sidx])) {
5731 		bcopy(from, *to, len);
5732 		(*to) += len;
5733 	} else {
5734 		int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
5735 
5736 		bcopy(from, *to, portion);
5737 		from += portion;
5738 		portion = len - portion;	/* remaining */
5739 		bcopy(from, (void *)eq->desc, portion);
5740 		(*to) = (caddr_t)eq->desc + portion;
5741 	}
5742 }
5743 
5744 static inline void
5745 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n)
5746 {
5747 	u_int db;
5748 
5749 	MPASS(n > 0);
5750 
5751 	db = eq->doorbells;
5752 	if (n > 1)
5753 		clrbit(&db, DOORBELL_WCWR);
5754 	wmb();
5755 
5756 	switch (ffs(db) - 1) {
5757 	case DOORBELL_UDB:
5758 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
5759 		break;
5760 
5761 	case DOORBELL_WCWR: {
5762 		volatile uint64_t *dst, *src;
5763 		int i;
5764 
5765 		/*
5766 		 * Queues whose 128B doorbell segment fits in the page do not
5767 		 * use relative qid (udb_qid is always 0).  Only queues with
5768 		 * doorbell segments can do WCWR.
5769 		 */
5770 		KASSERT(eq->udb_qid == 0 && n == 1,
5771 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
5772 		    __func__, eq->doorbells, n, eq->dbidx, eq));
5773 
5774 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
5775 		    UDBS_DB_OFFSET);
5776 		i = eq->dbidx;
5777 		src = (void *)&eq->desc[i];
5778 		while (src != (void *)&eq->desc[i + 1])
5779 			*dst++ = *src++;
5780 		wmb();
5781 		break;
5782 	}
5783 
5784 	case DOORBELL_UDBWC:
5785 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
5786 		wmb();
5787 		break;
5788 
5789 	case DOORBELL_KDB:
5790 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
5791 		    V_QID(eq->cntxt_id) | V_PIDX(n));
5792 		break;
5793 	}
5794 
5795 	IDXINCR(eq->dbidx, n, eq->sidx);
5796 }
5797 
5798 static inline u_int
5799 reclaimable_tx_desc(struct sge_eq *eq)
5800 {
5801 	uint16_t hw_cidx;
5802 
5803 	hw_cidx = read_hw_cidx(eq);
5804 	return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx));
5805 }
5806 
5807 static inline u_int
5808 total_available_tx_desc(struct sge_eq *eq)
5809 {
5810 	uint16_t hw_cidx, pidx;
5811 
5812 	hw_cidx = read_hw_cidx(eq);
5813 	pidx = eq->pidx;
5814 
5815 	if (pidx == hw_cidx)
5816 		return (eq->sidx - 1);
5817 	else
5818 		return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1);
5819 }
5820 
5821 static inline uint16_t
5822 read_hw_cidx(struct sge_eq *eq)
5823 {
5824 	struct sge_qstat *spg = (void *)&eq->desc[eq->sidx];
5825 	uint16_t cidx = spg->cidx;	/* stable snapshot */
5826 
5827 	return (be16toh(cidx));
5828 }
5829 
5830 /*
5831  * Reclaim 'n' descriptors approximately.
5832  */
5833 static u_int
5834 reclaim_tx_descs(struct sge_txq *txq, u_int n)
5835 {
5836 	struct tx_sdesc *txsd;
5837 	struct sge_eq *eq = &txq->eq;
5838 	u_int can_reclaim, reclaimed;
5839 
5840 	TXQ_LOCK_ASSERT_OWNED(txq);
5841 	MPASS(n > 0);
5842 
5843 	reclaimed = 0;
5844 	can_reclaim = reclaimable_tx_desc(eq);
5845 	while (can_reclaim && reclaimed < n) {
5846 		int ndesc;
5847 		struct mbuf *m, *nextpkt;
5848 
5849 		txsd = &txq->sdesc[eq->cidx];
5850 		ndesc = txsd->desc_used;
5851 
5852 		/* Firmware doesn't return "partial" credits. */
5853 		KASSERT(can_reclaim >= ndesc,
5854 		    ("%s: unexpected number of credits: %d, %d",
5855 		    __func__, can_reclaim, ndesc));
5856 		KASSERT(ndesc != 0,
5857 		    ("%s: descriptor with no credits: cidx %d",
5858 		    __func__, eq->cidx));
5859 
5860 		for (m = txsd->m; m != NULL; m = nextpkt) {
5861 			nextpkt = m->m_nextpkt;
5862 			m->m_nextpkt = NULL;
5863 			m_freem(m);
5864 		}
5865 		reclaimed += ndesc;
5866 		can_reclaim -= ndesc;
5867 		IDXINCR(eq->cidx, ndesc, eq->sidx);
5868 	}
5869 
5870 	return (reclaimed);
5871 }
5872 
5873 static void
5874 tx_reclaim(void *arg, int n)
5875 {
5876 	struct sge_txq *txq = arg;
5877 	struct sge_eq *eq = &txq->eq;
5878 
5879 	do {
5880 		if (TXQ_TRYLOCK(txq) == 0)
5881 			break;
5882 		n = reclaim_tx_descs(txq, 32);
5883 		if (eq->cidx == eq->pidx)
5884 			eq->equeqidx = eq->pidx;
5885 		TXQ_UNLOCK(txq);
5886 	} while (n > 0);
5887 }
5888 
5889 static __be64
5890 get_flit(struct sglist_seg *segs, int nsegs, int idx)
5891 {
5892 	int i = (idx / 3) * 2;
5893 
5894 	switch (idx % 3) {
5895 	case 0: {
5896 		uint64_t rc;
5897 
5898 		rc = (uint64_t)segs[i].ss_len << 32;
5899 		if (i + 1 < nsegs)
5900 			rc |= (uint64_t)(segs[i + 1].ss_len);
5901 
5902 		return (htobe64(rc));
5903 	}
5904 	case 1:
5905 		return (htobe64(segs[i].ss_paddr));
5906 	case 2:
5907 		return (htobe64(segs[i + 1].ss_paddr));
5908 	}
5909 
5910 	return (0);
5911 }
5912 
5913 static int
5914 find_refill_source(struct adapter *sc, int maxp, bool packing)
5915 {
5916 	int i, zidx = -1;
5917 	struct rx_buf_info *rxb = &sc->sge.rx_buf_info[0];
5918 
5919 	if (packing) {
5920 		for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) {
5921 			if (rxb->hwidx2 == -1)
5922 				continue;
5923 			if (rxb->size1 < PAGE_SIZE &&
5924 			    rxb->size1 < largest_rx_cluster)
5925 				continue;
5926 			if (rxb->size1 > largest_rx_cluster)
5927 				break;
5928 			MPASS(rxb->size1 - rxb->size2 >= CL_METADATA_SIZE);
5929 			if (rxb->size2 >= maxp)
5930 				return (i);
5931 			zidx = i;
5932 		}
5933 	} else {
5934 		for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) {
5935 			if (rxb->hwidx1 == -1)
5936 				continue;
5937 			if (rxb->size1 > largest_rx_cluster)
5938 				break;
5939 			if (rxb->size1 >= maxp)
5940 				return (i);
5941 			zidx = i;
5942 		}
5943 	}
5944 
5945 	return (zidx);
5946 }
5947 
5948 static void
5949 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
5950 {
5951 	mtx_lock(&sc->sfl_lock);
5952 	FL_LOCK(fl);
5953 	if ((fl->flags & FL_DOOMED) == 0) {
5954 		fl->flags |= FL_STARVING;
5955 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
5956 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
5957 	}
5958 	FL_UNLOCK(fl);
5959 	mtx_unlock(&sc->sfl_lock);
5960 }
5961 
5962 static void
5963 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq)
5964 {
5965 	struct sge_wrq *wrq = (void *)eq;
5966 
5967 	atomic_readandclear_int(&eq->equiq);
5968 	taskqueue_enqueue(sc->tq[eq->tx_chan], &wrq->wrq_tx_task);
5969 }
5970 
5971 static void
5972 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq)
5973 {
5974 	struct sge_txq *txq = (void *)eq;
5975 
5976 	MPASS((eq->flags & EQ_TYPEMASK) == EQ_ETH);
5977 
5978 	atomic_readandclear_int(&eq->equiq);
5979 	if (mp_ring_is_idle(txq->r))
5980 		taskqueue_enqueue(sc->tq[eq->tx_chan], &txq->tx_reclaim_task);
5981 	else
5982 		mp_ring_check_drainage(txq->r, 64);
5983 }
5984 
5985 static int
5986 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
5987     struct mbuf *m)
5988 {
5989 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
5990 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
5991 	struct adapter *sc = iq->adapter;
5992 	struct sge *s = &sc->sge;
5993 	struct sge_eq *eq;
5994 	static void (*h[])(struct adapter *, struct sge_eq *) = {NULL,
5995 		&handle_wrq_egr_update, &handle_eth_egr_update,
5996 		&handle_wrq_egr_update};
5997 
5998 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5999 	    rss->opcode));
6000 
6001 	eq = s->eqmap[qid - s->eq_start - s->eq_base];
6002 	(*h[eq->flags & EQ_TYPEMASK])(sc, eq);
6003 
6004 	return (0);
6005 }
6006 
6007 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
6008 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
6009     offsetof(struct cpl_fw6_msg, data));
6010 
6011 static int
6012 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
6013 {
6014 	struct adapter *sc = iq->adapter;
6015 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
6016 
6017 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
6018 	    rss->opcode));
6019 
6020 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
6021 		const struct rss_header *rss2;
6022 
6023 		rss2 = (const struct rss_header *)&cpl->data[0];
6024 		return (t4_cpl_handler[rss2->opcode](iq, rss2, m));
6025 	}
6026 
6027 	return (t4_fw_msg_handler[cpl->type](sc, &cpl->data[0]));
6028 }
6029 
6030 /**
6031  *	t4_handle_wrerr_rpl - process a FW work request error message
6032  *	@adap: the adapter
6033  *	@rpl: start of the FW message
6034  */
6035 static int
6036 t4_handle_wrerr_rpl(struct adapter *adap, const __be64 *rpl)
6037 {
6038 	u8 opcode = *(const u8 *)rpl;
6039 	const struct fw_error_cmd *e = (const void *)rpl;
6040 	unsigned int i;
6041 
6042 	if (opcode != FW_ERROR_CMD) {
6043 		log(LOG_ERR,
6044 		    "%s: Received WRERR_RPL message with opcode %#x\n",
6045 		    device_get_nameunit(adap->dev), opcode);
6046 		return (EINVAL);
6047 	}
6048 	log(LOG_ERR, "%s: FW_ERROR (%s) ", device_get_nameunit(adap->dev),
6049 	    G_FW_ERROR_CMD_FATAL(be32toh(e->op_to_type)) ? "fatal" :
6050 	    "non-fatal");
6051 	switch (G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type))) {
6052 	case FW_ERROR_TYPE_EXCEPTION:
6053 		log(LOG_ERR, "exception info:\n");
6054 		for (i = 0; i < nitems(e->u.exception.info); i++)
6055 			log(LOG_ERR, "%s%08x", i == 0 ? "\t" : " ",
6056 			    be32toh(e->u.exception.info[i]));
6057 		log(LOG_ERR, "\n");
6058 		break;
6059 	case FW_ERROR_TYPE_HWMODULE:
6060 		log(LOG_ERR, "HW module regaddr %08x regval %08x\n",
6061 		    be32toh(e->u.hwmodule.regaddr),
6062 		    be32toh(e->u.hwmodule.regval));
6063 		break;
6064 	case FW_ERROR_TYPE_WR:
6065 		log(LOG_ERR, "WR cidx %d PF %d VF %d eqid %d hdr:\n",
6066 		    be16toh(e->u.wr.cidx),
6067 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.wr.pfn_vfn)),
6068 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.wr.pfn_vfn)),
6069 		    be32toh(e->u.wr.eqid));
6070 		for (i = 0; i < nitems(e->u.wr.wrhdr); i++)
6071 			log(LOG_ERR, "%s%02x", i == 0 ? "\t" : " ",
6072 			    e->u.wr.wrhdr[i]);
6073 		log(LOG_ERR, "\n");
6074 		break;
6075 	case FW_ERROR_TYPE_ACL:
6076 		log(LOG_ERR, "ACL cidx %d PF %d VF %d eqid %d %s",
6077 		    be16toh(e->u.acl.cidx),
6078 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.acl.pfn_vfn)),
6079 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.acl.pfn_vfn)),
6080 		    be32toh(e->u.acl.eqid),
6081 		    G_FW_ERROR_CMD_MV(be16toh(e->u.acl.mv_pkd)) ? "vlanid" :
6082 		    "MAC");
6083 		for (i = 0; i < nitems(e->u.acl.val); i++)
6084 			log(LOG_ERR, " %02x", e->u.acl.val[i]);
6085 		log(LOG_ERR, "\n");
6086 		break;
6087 	default:
6088 		log(LOG_ERR, "type %#x\n",
6089 		    G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type)));
6090 		return (EINVAL);
6091 	}
6092 	return (0);
6093 }
6094 
6095 static inline bool
6096 bufidx_used(struct adapter *sc, int idx)
6097 {
6098 	struct rx_buf_info *rxb = &sc->sge.rx_buf_info[0];
6099 	int i;
6100 
6101 	for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) {
6102 		if (rxb->size1 > largest_rx_cluster)
6103 			continue;
6104 		if (rxb->hwidx1 == idx || rxb->hwidx2 == idx)
6105 			return (true);
6106 	}
6107 
6108 	return (false);
6109 }
6110 
6111 static int
6112 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
6113 {
6114 	struct adapter *sc = arg1;
6115 	struct sge_params *sp = &sc->params.sge;
6116 	int i, rc;
6117 	struct sbuf sb;
6118 	char c;
6119 
6120 	sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
6121 	for (i = 0; i < SGE_FLBUF_SIZES; i++) {
6122 		if (bufidx_used(sc, i))
6123 			c = '*';
6124 		else
6125 			c = '\0';
6126 
6127 		sbuf_printf(&sb, "%u%c ", sp->sge_fl_buffer_size[i], c);
6128 	}
6129 	sbuf_trim(&sb);
6130 	sbuf_finish(&sb);
6131 	rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
6132 	sbuf_delete(&sb);
6133 	return (rc);
6134 }
6135 
6136 #ifdef RATELIMIT
6137 /*
6138  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
6139  */
6140 static inline u_int
6141 txpkt_eo_len16(u_int nsegs, u_int immhdrs, u_int tso)
6142 {
6143 	u_int n;
6144 
6145 	MPASS(immhdrs > 0);
6146 
6147 	n = roundup2(sizeof(struct fw_eth_tx_eo_wr) +
6148 	    sizeof(struct cpl_tx_pkt_core) + immhdrs, 16);
6149 	if (__predict_false(nsegs == 0))
6150 		goto done;
6151 
6152 	nsegs--; /* first segment is part of ulptx_sgl */
6153 	n += sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
6154 	if (tso)
6155 		n += sizeof(struct cpl_tx_pkt_lso_core);
6156 
6157 done:
6158 	return (howmany(n, 16));
6159 }
6160 
6161 #define ETID_FLOWC_NPARAMS 6
6162 #define ETID_FLOWC_LEN (roundup2((sizeof(struct fw_flowc_wr) + \
6163     ETID_FLOWC_NPARAMS * sizeof(struct fw_flowc_mnemval)), 16))
6164 #define ETID_FLOWC_LEN16 (howmany(ETID_FLOWC_LEN, 16))
6165 
6166 static int
6167 send_etid_flowc_wr(struct cxgbe_rate_tag *cst, struct port_info *pi,
6168     struct vi_info *vi)
6169 {
6170 	struct wrq_cookie cookie;
6171 	u_int pfvf = pi->adapter->pf << S_FW_VIID_PFN;
6172 	struct fw_flowc_wr *flowc;
6173 
6174 	mtx_assert(&cst->lock, MA_OWNED);
6175 	MPASS((cst->flags & (EO_FLOWC_PENDING | EO_FLOWC_RPL_PENDING)) ==
6176 	    EO_FLOWC_PENDING);
6177 
6178 	flowc = start_wrq_wr(&cst->eo_txq->wrq, ETID_FLOWC_LEN16, &cookie);
6179 	if (__predict_false(flowc == NULL))
6180 		return (ENOMEM);
6181 
6182 	bzero(flowc, ETID_FLOWC_LEN);
6183 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
6184 	    V_FW_FLOWC_WR_NPARAMS(ETID_FLOWC_NPARAMS) | V_FW_WR_COMPL(0));
6185 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(ETID_FLOWC_LEN16) |
6186 	    V_FW_WR_FLOWID(cst->etid));
6187 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN;
6188 	flowc->mnemval[0].val = htobe32(pfvf);
6189 	flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH;
6190 	flowc->mnemval[1].val = htobe32(pi->tx_chan);
6191 	flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT;
6192 	flowc->mnemval[2].val = htobe32(pi->tx_chan);
6193 	flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID;
6194 	flowc->mnemval[3].val = htobe32(cst->iqid);
6195 	flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_EOSTATE;
6196 	flowc->mnemval[4].val = htobe32(FW_FLOWC_MNEM_EOSTATE_ESTABLISHED);
6197 	flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS;
6198 	flowc->mnemval[5].val = htobe32(cst->schedcl);
6199 
6200 	commit_wrq_wr(&cst->eo_txq->wrq, flowc, &cookie);
6201 
6202 	cst->flags &= ~EO_FLOWC_PENDING;
6203 	cst->flags |= EO_FLOWC_RPL_PENDING;
6204 	MPASS(cst->tx_credits >= ETID_FLOWC_LEN16);	/* flowc is first WR. */
6205 	cst->tx_credits -= ETID_FLOWC_LEN16;
6206 
6207 	return (0);
6208 }
6209 
6210 #define ETID_FLUSH_LEN16 (howmany(sizeof (struct fw_flowc_wr), 16))
6211 
6212 void
6213 send_etid_flush_wr(struct cxgbe_rate_tag *cst)
6214 {
6215 	struct fw_flowc_wr *flowc;
6216 	struct wrq_cookie cookie;
6217 
6218 	mtx_assert(&cst->lock, MA_OWNED);
6219 
6220 	flowc = start_wrq_wr(&cst->eo_txq->wrq, ETID_FLUSH_LEN16, &cookie);
6221 	if (__predict_false(flowc == NULL))
6222 		CXGBE_UNIMPLEMENTED(__func__);
6223 
6224 	bzero(flowc, ETID_FLUSH_LEN16 * 16);
6225 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
6226 	    V_FW_FLOWC_WR_NPARAMS(0) | F_FW_WR_COMPL);
6227 	flowc->flowid_len16 = htobe32(V_FW_WR_LEN16(ETID_FLUSH_LEN16) |
6228 	    V_FW_WR_FLOWID(cst->etid));
6229 
6230 	commit_wrq_wr(&cst->eo_txq->wrq, flowc, &cookie);
6231 
6232 	cst->flags |= EO_FLUSH_RPL_PENDING;
6233 	MPASS(cst->tx_credits >= ETID_FLUSH_LEN16);
6234 	cst->tx_credits -= ETID_FLUSH_LEN16;
6235 	cst->ncompl++;
6236 }
6237 
6238 static void
6239 write_ethofld_wr(struct cxgbe_rate_tag *cst, struct fw_eth_tx_eo_wr *wr,
6240     struct mbuf *m0, int compl)
6241 {
6242 	struct cpl_tx_pkt_core *cpl;
6243 	uint64_t ctrl1;
6244 	uint32_t ctrl;	/* used in many unrelated places */
6245 	int len16, pktlen, nsegs, immhdrs;
6246 	caddr_t dst;
6247 	uintptr_t p;
6248 	struct ulptx_sgl *usgl;
6249 	struct sglist sg;
6250 	struct sglist_seg segs[38];	/* XXX: find real limit.  XXX: get off the stack */
6251 
6252 	mtx_assert(&cst->lock, MA_OWNED);
6253 	M_ASSERTPKTHDR(m0);
6254 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
6255 	    m0->m_pkthdr.l4hlen > 0,
6256 	    ("%s: ethofld mbuf %p is missing header lengths", __func__, m0));
6257 
6258 	len16 = mbuf_eo_len16(m0);
6259 	nsegs = mbuf_eo_nsegs(m0);
6260 	pktlen = m0->m_pkthdr.len;
6261 	ctrl = sizeof(struct cpl_tx_pkt_core);
6262 	if (needs_tso(m0))
6263 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
6264 	immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen + m0->m_pkthdr.l4hlen;
6265 	ctrl += immhdrs;
6266 
6267 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_EO_WR) |
6268 	    V_FW_ETH_TX_EO_WR_IMMDLEN(ctrl) | V_FW_WR_COMPL(!!compl));
6269 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(len16) |
6270 	    V_FW_WR_FLOWID(cst->etid));
6271 	wr->r3 = 0;
6272 	if (needs_outer_udp_csum(m0)) {
6273 		wr->u.udpseg.type = FW_ETH_TX_EO_TYPE_UDPSEG;
6274 		wr->u.udpseg.ethlen = m0->m_pkthdr.l2hlen;
6275 		wr->u.udpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
6276 		wr->u.udpseg.udplen = m0->m_pkthdr.l4hlen;
6277 		wr->u.udpseg.rtplen = 0;
6278 		wr->u.udpseg.r4 = 0;
6279 		wr->u.udpseg.mss = htobe16(pktlen - immhdrs);
6280 		wr->u.udpseg.schedpktsize = wr->u.udpseg.mss;
6281 		wr->u.udpseg.plen = htobe32(pktlen - immhdrs);
6282 		cpl = (void *)(wr + 1);
6283 	} else {
6284 		MPASS(needs_outer_tcp_csum(m0));
6285 		wr->u.tcpseg.type = FW_ETH_TX_EO_TYPE_TCPSEG;
6286 		wr->u.tcpseg.ethlen = m0->m_pkthdr.l2hlen;
6287 		wr->u.tcpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
6288 		wr->u.tcpseg.tcplen = m0->m_pkthdr.l4hlen;
6289 		wr->u.tcpseg.tsclk_tsoff = mbuf_eo_tsclk_tsoff(m0);
6290 		wr->u.tcpseg.r4 = 0;
6291 		wr->u.tcpseg.r5 = 0;
6292 		wr->u.tcpseg.plen = htobe32(pktlen - immhdrs);
6293 
6294 		if (needs_tso(m0)) {
6295 			struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
6296 
6297 			wr->u.tcpseg.mss = htobe16(m0->m_pkthdr.tso_segsz);
6298 
6299 			ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) |
6300 			    F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE |
6301 			    V_LSO_ETHHDR_LEN((m0->m_pkthdr.l2hlen -
6302 				ETHER_HDR_LEN) >> 2) |
6303 			    V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) |
6304 			    V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
6305 			if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
6306 				ctrl |= F_LSO_IPV6;
6307 			lso->lso_ctrl = htobe32(ctrl);
6308 			lso->ipid_ofst = htobe16(0);
6309 			lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
6310 			lso->seqno_offset = htobe32(0);
6311 			lso->len = htobe32(pktlen);
6312 
6313 			cpl = (void *)(lso + 1);
6314 		} else {
6315 			wr->u.tcpseg.mss = htobe16(0xffff);
6316 			cpl = (void *)(wr + 1);
6317 		}
6318 	}
6319 
6320 	/* Checksum offload must be requested for ethofld. */
6321 	MPASS(needs_outer_l4_csum(m0));
6322 	ctrl1 = csum_to_ctrl(cst->adapter, m0);
6323 
6324 	/* VLAN tag insertion */
6325 	if (needs_vlan_insertion(m0)) {
6326 		ctrl1 |= F_TXPKT_VLAN_VLD |
6327 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
6328 	}
6329 
6330 	/* CPL header */
6331 	cpl->ctrl0 = cst->ctrl0;
6332 	cpl->pack = 0;
6333 	cpl->len = htobe16(pktlen);
6334 	cpl->ctrl1 = htobe64(ctrl1);
6335 
6336 	/* Copy Ethernet, IP & TCP/UDP hdrs as immediate data */
6337 	p = (uintptr_t)(cpl + 1);
6338 	m_copydata(m0, 0, immhdrs, (void *)p);
6339 
6340 	/* SGL */
6341 	dst = (void *)(cpl + 1);
6342 	if (nsegs > 0) {
6343 		int i, pad;
6344 
6345 		/* zero-pad upto next 16Byte boundary, if not 16Byte aligned */
6346 		p += immhdrs;
6347 		pad = 16 - (immhdrs & 0xf);
6348 		bzero((void *)p, pad);
6349 
6350 		usgl = (void *)(p + pad);
6351 		usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
6352 		    V_ULPTX_NSGE(nsegs));
6353 
6354 		sglist_init(&sg, nitems(segs), segs);
6355 		for (; m0 != NULL; m0 = m0->m_next) {
6356 			if (__predict_false(m0->m_len == 0))
6357 				continue;
6358 			if (immhdrs >= m0->m_len) {
6359 				immhdrs -= m0->m_len;
6360 				continue;
6361 			}
6362 			if (m0->m_flags & M_EXTPG)
6363 				sglist_append_mbuf_epg(&sg, m0,
6364 				    mtod(m0, vm_offset_t), m0->m_len);
6365                         else
6366 				sglist_append(&sg, mtod(m0, char *) + immhdrs,
6367 				    m0->m_len - immhdrs);
6368 			immhdrs = 0;
6369 		}
6370 		MPASS(sg.sg_nseg == nsegs);
6371 
6372 		/*
6373 		 * Zero pad last 8B in case the WR doesn't end on a 16B
6374 		 * boundary.
6375 		 */
6376 		*(uint64_t *)((char *)wr + len16 * 16 - 8) = 0;
6377 
6378 		usgl->len0 = htobe32(segs[0].ss_len);
6379 		usgl->addr0 = htobe64(segs[0].ss_paddr);
6380 		for (i = 0; i < nsegs - 1; i++) {
6381 			usgl->sge[i / 2].len[i & 1] = htobe32(segs[i + 1].ss_len);
6382 			usgl->sge[i / 2].addr[i & 1] = htobe64(segs[i + 1].ss_paddr);
6383 		}
6384 		if (i & 1)
6385 			usgl->sge[i / 2].len[1] = htobe32(0);
6386 	}
6387 
6388 }
6389 
6390 static void
6391 ethofld_tx(struct cxgbe_rate_tag *cst)
6392 {
6393 	struct mbuf *m;
6394 	struct wrq_cookie cookie;
6395 	int next_credits, compl;
6396 	struct fw_eth_tx_eo_wr *wr;
6397 
6398 	mtx_assert(&cst->lock, MA_OWNED);
6399 
6400 	while ((m = mbufq_first(&cst->pending_tx)) != NULL) {
6401 		M_ASSERTPKTHDR(m);
6402 
6403 		/* How many len16 credits do we need to send this mbuf. */
6404 		next_credits = mbuf_eo_len16(m);
6405 		MPASS(next_credits > 0);
6406 		if (next_credits > cst->tx_credits) {
6407 			/*
6408 			 * Tx will make progress eventually because there is at
6409 			 * least one outstanding fw4_ack that will return
6410 			 * credits and kick the tx.
6411 			 */
6412 			MPASS(cst->ncompl > 0);
6413 			return;
6414 		}
6415 		wr = start_wrq_wr(&cst->eo_txq->wrq, next_credits, &cookie);
6416 		if (__predict_false(wr == NULL)) {
6417 			/* XXX: wishful thinking, not a real assertion. */
6418 			MPASS(cst->ncompl > 0);
6419 			return;
6420 		}
6421 		cst->tx_credits -= next_credits;
6422 		cst->tx_nocompl += next_credits;
6423 		compl = cst->ncompl == 0 || cst->tx_nocompl >= cst->tx_total / 2;
6424 		ETHER_BPF_MTAP(cst->com.ifp, m);
6425 		write_ethofld_wr(cst, wr, m, compl);
6426 		commit_wrq_wr(&cst->eo_txq->wrq, wr, &cookie);
6427 		if (compl) {
6428 			cst->ncompl++;
6429 			cst->tx_nocompl	= 0;
6430 		}
6431 		(void) mbufq_dequeue(&cst->pending_tx);
6432 
6433 		/*
6434 		 * Drop the mbuf's reference on the tag now rather
6435 		 * than waiting until m_freem().  This ensures that
6436 		 * cxgbe_rate_tag_free gets called when the inp drops
6437 		 * its reference on the tag and there are no more
6438 		 * mbufs in the pending_tx queue and can flush any
6439 		 * pending requests.  Otherwise if the last mbuf
6440 		 * doesn't request a completion the etid will never be
6441 		 * released.
6442 		 */
6443 		m->m_pkthdr.snd_tag = NULL;
6444 		m->m_pkthdr.csum_flags &= ~CSUM_SND_TAG;
6445 		m_snd_tag_rele(&cst->com);
6446 
6447 		mbufq_enqueue(&cst->pending_fwack, m);
6448 	}
6449 }
6450 
6451 int
6452 ethofld_transmit(struct ifnet *ifp, struct mbuf *m0)
6453 {
6454 	struct cxgbe_rate_tag *cst;
6455 	int rc;
6456 
6457 	MPASS(m0->m_nextpkt == NULL);
6458 	MPASS(m0->m_pkthdr.csum_flags & CSUM_SND_TAG);
6459 	MPASS(m0->m_pkthdr.snd_tag != NULL);
6460 	cst = mst_to_crt(m0->m_pkthdr.snd_tag);
6461 
6462 	mtx_lock(&cst->lock);
6463 	MPASS(cst->flags & EO_SND_TAG_REF);
6464 
6465 	if (__predict_false(cst->flags & EO_FLOWC_PENDING)) {
6466 		struct vi_info *vi = ifp->if_softc;
6467 		struct port_info *pi = vi->pi;
6468 		struct adapter *sc = pi->adapter;
6469 		const uint32_t rss_mask = vi->rss_size - 1;
6470 		uint32_t rss_hash;
6471 
6472 		cst->eo_txq = &sc->sge.ofld_txq[vi->first_ofld_txq];
6473 		if (M_HASHTYPE_ISHASH(m0))
6474 			rss_hash = m0->m_pkthdr.flowid;
6475 		else
6476 			rss_hash = arc4random();
6477 		/* We assume RSS hashing */
6478 		cst->iqid = vi->rss[rss_hash & rss_mask];
6479 		cst->eo_txq += rss_hash % vi->nofldtxq;
6480 		rc = send_etid_flowc_wr(cst, pi, vi);
6481 		if (rc != 0)
6482 			goto done;
6483 	}
6484 
6485 	if (__predict_false(cst->plen + m0->m_pkthdr.len > eo_max_backlog)) {
6486 		rc = ENOBUFS;
6487 		goto done;
6488 	}
6489 
6490 	mbufq_enqueue(&cst->pending_tx, m0);
6491 	cst->plen += m0->m_pkthdr.len;
6492 
6493 	/*
6494 	 * Hold an extra reference on the tag while generating work
6495 	 * requests to ensure that we don't try to free the tag during
6496 	 * ethofld_tx() in case we are sending the final mbuf after
6497 	 * the inp was freed.
6498 	 */
6499 	m_snd_tag_ref(&cst->com);
6500 	ethofld_tx(cst);
6501 	mtx_unlock(&cst->lock);
6502 	m_snd_tag_rele(&cst->com);
6503 	return (0);
6504 
6505 done:
6506 	mtx_unlock(&cst->lock);
6507 	if (__predict_false(rc != 0))
6508 		m_freem(m0);
6509 	return (rc);
6510 }
6511 
6512 static int
6513 ethofld_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
6514 {
6515 	struct adapter *sc = iq->adapter;
6516 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
6517 	struct mbuf *m;
6518 	u_int etid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
6519 	struct cxgbe_rate_tag *cst;
6520 	uint8_t credits = cpl->credits;
6521 
6522 	cst = lookup_etid(sc, etid);
6523 	mtx_lock(&cst->lock);
6524 	if (__predict_false(cst->flags & EO_FLOWC_RPL_PENDING)) {
6525 		MPASS(credits >= ETID_FLOWC_LEN16);
6526 		credits -= ETID_FLOWC_LEN16;
6527 		cst->flags &= ~EO_FLOWC_RPL_PENDING;
6528 	}
6529 
6530 	KASSERT(cst->ncompl > 0,
6531 	    ("%s: etid %u (%p) wasn't expecting completion.",
6532 	    __func__, etid, cst));
6533 	cst->ncompl--;
6534 
6535 	while (credits > 0) {
6536 		m = mbufq_dequeue(&cst->pending_fwack);
6537 		if (__predict_false(m == NULL)) {
6538 			/*
6539 			 * The remaining credits are for the final flush that
6540 			 * was issued when the tag was freed by the kernel.
6541 			 */
6542 			MPASS((cst->flags &
6543 			    (EO_FLUSH_RPL_PENDING | EO_SND_TAG_REF)) ==
6544 			    EO_FLUSH_RPL_PENDING);
6545 			MPASS(credits == ETID_FLUSH_LEN16);
6546 			MPASS(cst->tx_credits + cpl->credits == cst->tx_total);
6547 			MPASS(cst->ncompl == 0);
6548 
6549 			cst->flags &= ~EO_FLUSH_RPL_PENDING;
6550 			cst->tx_credits += cpl->credits;
6551 			cxgbe_rate_tag_free_locked(cst);
6552 			return (0);	/* cst is gone. */
6553 		}
6554 		KASSERT(m != NULL,
6555 		    ("%s: too many credits (%u, %u)", __func__, cpl->credits,
6556 		    credits));
6557 		KASSERT(credits >= mbuf_eo_len16(m),
6558 		    ("%s: too few credits (%u, %u, %u)", __func__,
6559 		    cpl->credits, credits, mbuf_eo_len16(m)));
6560 		credits -= mbuf_eo_len16(m);
6561 		cst->plen -= m->m_pkthdr.len;
6562 		m_freem(m);
6563 	}
6564 
6565 	cst->tx_credits += cpl->credits;
6566 	MPASS(cst->tx_credits <= cst->tx_total);
6567 
6568 	if (cst->flags & EO_SND_TAG_REF) {
6569 		/*
6570 		 * As with ethofld_transmit(), hold an extra reference
6571 		 * so that the tag is stable across ethold_tx().
6572 		 */
6573 		m_snd_tag_ref(&cst->com);
6574 		m = mbufq_first(&cst->pending_tx);
6575 		if (m != NULL && cst->tx_credits >= mbuf_eo_len16(m))
6576 			ethofld_tx(cst);
6577 		mtx_unlock(&cst->lock);
6578 		m_snd_tag_rele(&cst->com);
6579 	} else {
6580 		/*
6581 		 * There shouldn't be any pending packets if the tag
6582 		 * was freed by the kernel since any pending packet
6583 		 * should hold a reference to the tag.
6584 		 */
6585 		MPASS(mbufq_first(&cst->pending_tx) == NULL);
6586 		mtx_unlock(&cst->lock);
6587 	}
6588 
6589 	return (0);
6590 }
6591 #endif
6592