xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision 5ca8e32633c4ffbbcd6762e5888b6a4ba0708c6c)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
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 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_kern_tls.h"
34 #include "opt_ratelimit.h"
35 
36 #include <sys/types.h>
37 #include <sys/eventhandler.h>
38 #include <sys/mbuf.h>
39 #include <sys/socket.h>
40 #include <sys/kernel.h>
41 #include <sys/ktls.h>
42 #include <sys/malloc.h>
43 #include <sys/msan.h>
44 #include <sys/queue.h>
45 #include <sys/sbuf.h>
46 #include <sys/taskqueue.h>
47 #include <sys/time.h>
48 #include <sys/sglist.h>
49 #include <sys/sysctl.h>
50 #include <sys/smp.h>
51 #include <sys/socketvar.h>
52 #include <sys/counter.h>
53 #include <net/bpf.h>
54 #include <net/ethernet.h>
55 #include <net/if.h>
56 #include <net/if_vlan_var.h>
57 #include <net/if_vxlan.h>
58 #include <netinet/in.h>
59 #include <netinet/ip.h>
60 #include <netinet/ip6.h>
61 #include <netinet/tcp.h>
62 #include <netinet/udp.h>
63 #include <machine/in_cksum.h>
64 #include <machine/md_var.h>
65 #include <vm/vm.h>
66 #include <vm/pmap.h>
67 #ifdef DEV_NETMAP
68 #include <machine/bus.h>
69 #include <sys/selinfo.h>
70 #include <net/if_var.h>
71 #include <net/netmap.h>
72 #include <dev/netmap/netmap_kern.h>
73 #endif
74 
75 #include "common/common.h"
76 #include "common/t4_regs.h"
77 #include "common/t4_regs_values.h"
78 #include "common/t4_msg.h"
79 #include "t4_l2t.h"
80 #include "t4_mp_ring.h"
81 
82 #define RX_COPY_THRESHOLD MINCLSIZE
83 
84 /*
85  * Ethernet frames are DMA'd at this byte offset into the freelist buffer.
86  * 0-7 are valid values.
87  */
88 static int fl_pktshift = 0;
89 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pktshift, CTLFLAG_RDTUN, &fl_pktshift, 0,
90     "payload DMA offset in rx buffer (bytes)");
91 
92 /*
93  * Pad ethernet payload up to this boundary.
94  * -1: driver should figure out a good value.
95  *  0: disable padding.
96  *  Any power of 2 from 32 to 4096 (both inclusive) is also a valid value.
97  */
98 int fl_pad = -1;
99 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pad, CTLFLAG_RDTUN, &fl_pad, 0,
100     "payload pad boundary (bytes)");
101 
102 /*
103  * Status page length.
104  * -1: driver should figure out a good value.
105  *  64 or 128 are the only other valid values.
106  */
107 static int spg_len = -1;
108 SYSCTL_INT(_hw_cxgbe, OID_AUTO, spg_len, CTLFLAG_RDTUN, &spg_len, 0,
109     "status page size (bytes)");
110 
111 /*
112  * Congestion drops.
113  * -1: no congestion feedback (not recommended).
114  *  0: backpressure the channel instead of dropping packets right away.
115  *  1: no backpressure, drop packets for the congested queue immediately.
116  *  2: both backpressure and drop.
117  */
118 static int cong_drop = 0;
119 SYSCTL_INT(_hw_cxgbe, OID_AUTO, cong_drop, CTLFLAG_RDTUN, &cong_drop, 0,
120     "Congestion control for NIC RX queues (0 = backpressure, 1 = drop, 2 = both");
121 #ifdef TCP_OFFLOAD
122 static int ofld_cong_drop = 0;
123 SYSCTL_INT(_hw_cxgbe, OID_AUTO, ofld_cong_drop, CTLFLAG_RDTUN, &ofld_cong_drop, 0,
124     "Congestion control for TOE RX queues (0 = backpressure, 1 = drop, 2 = both");
125 #endif
126 
127 /*
128  * Deliver multiple frames in the same free list buffer if they fit.
129  * -1: let the driver decide whether to enable buffer packing or not.
130  *  0: disable buffer packing.
131  *  1: enable buffer packing.
132  */
133 static int buffer_packing = -1;
134 SYSCTL_INT(_hw_cxgbe, OID_AUTO, buffer_packing, CTLFLAG_RDTUN, &buffer_packing,
135     0, "Enable buffer packing");
136 
137 /*
138  * Start next frame in a packed buffer at this boundary.
139  * -1: driver should figure out a good value.
140  * T4: driver will ignore this and use the same value as fl_pad above.
141  * T5: 16, or a power of 2 from 64 to 4096 (both inclusive) is a valid value.
142  */
143 static int fl_pack = -1;
144 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pack, CTLFLAG_RDTUN, &fl_pack, 0,
145     "payload pack boundary (bytes)");
146 
147 /*
148  * Largest rx cluster size that the driver is allowed to allocate.
149  */
150 static int largest_rx_cluster = MJUM16BYTES;
151 SYSCTL_INT(_hw_cxgbe, OID_AUTO, largest_rx_cluster, CTLFLAG_RDTUN,
152     &largest_rx_cluster, 0, "Largest rx cluster (bytes)");
153 
154 /*
155  * Size of cluster allocation that's most likely to succeed.  The driver will
156  * fall back to this size if it fails to allocate clusters larger than this.
157  */
158 static int safest_rx_cluster = PAGE_SIZE;
159 SYSCTL_INT(_hw_cxgbe, OID_AUTO, safest_rx_cluster, CTLFLAG_RDTUN,
160     &safest_rx_cluster, 0, "Safe rx cluster (bytes)");
161 
162 #ifdef RATELIMIT
163 /*
164  * Knob to control TCP timestamp rewriting, and the granularity of the tick used
165  * for rewriting.  -1 and 0-3 are all valid values.
166  * -1: hardware should leave the TCP timestamps alone.
167  * 0: 1ms
168  * 1: 100us
169  * 2: 10us
170  * 3: 1us
171  */
172 static int tsclk = -1;
173 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tsclk, CTLFLAG_RDTUN, &tsclk, 0,
174     "Control TCP timestamp rewriting when using pacing");
175 
176 static int eo_max_backlog = 1024 * 1024;
177 SYSCTL_INT(_hw_cxgbe, OID_AUTO, eo_max_backlog, CTLFLAG_RDTUN, &eo_max_backlog,
178     0, "Maximum backlog of ratelimited data per flow");
179 #endif
180 
181 /*
182  * The interrupt holdoff timers are multiplied by this value on T6+.
183  * 1 and 3-17 (both inclusive) are legal values.
184  */
185 static int tscale = 1;
186 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tscale, CTLFLAG_RDTUN, &tscale, 0,
187     "Interrupt holdoff timer scale on T6+");
188 
189 /*
190  * Number of LRO entries in the lro_ctrl structure per rx queue.
191  */
192 static int lro_entries = TCP_LRO_ENTRIES;
193 SYSCTL_INT(_hw_cxgbe, OID_AUTO, lro_entries, CTLFLAG_RDTUN, &lro_entries, 0,
194     "Number of LRO entries per RX queue");
195 
196 /*
197  * This enables presorting of frames before they're fed into tcp_lro_rx.
198  */
199 static int lro_mbufs = 0;
200 SYSCTL_INT(_hw_cxgbe, OID_AUTO, lro_mbufs, CTLFLAG_RDTUN, &lro_mbufs, 0,
201     "Enable presorting of LRO frames");
202 
203 static counter_u64_t pullups;
204 SYSCTL_COUNTER_U64(_hw_cxgbe, OID_AUTO, pullups, CTLFLAG_RD, &pullups,
205     "Number of mbuf pullups performed");
206 
207 static counter_u64_t defrags;
208 SYSCTL_COUNTER_U64(_hw_cxgbe, OID_AUTO, defrags, CTLFLAG_RD, &defrags,
209     "Number of mbuf defrags performed");
210 
211 static int t4_tx_coalesce = 1;
212 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tx_coalesce, CTLFLAG_RWTUN, &t4_tx_coalesce, 0,
213     "tx coalescing allowed");
214 
215 /*
216  * The driver will make aggressive attempts at tx coalescing if it sees these
217  * many packets eligible for coalescing in quick succession, with no more than
218  * the specified gap in between the eth_tx calls that delivered the packets.
219  */
220 static int t4_tx_coalesce_pkts = 32;
221 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tx_coalesce_pkts, CTLFLAG_RWTUN,
222     &t4_tx_coalesce_pkts, 0,
223     "# of consecutive packets (1 - 255) that will trigger tx coalescing");
224 static int t4_tx_coalesce_gap = 5;
225 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tx_coalesce_gap, CTLFLAG_RWTUN,
226     &t4_tx_coalesce_gap, 0, "tx gap (in microseconds)");
227 
228 static int service_iq(struct sge_iq *, int);
229 static int service_iq_fl(struct sge_iq *, int);
230 static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t);
231 static int eth_rx(struct adapter *, struct sge_rxq *, const struct iq_desc *,
232     u_int);
233 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int,
234     int, int, int);
235 static inline void init_fl(struct adapter *, struct sge_fl *, int, int, char *);
236 static inline void init_eq(struct adapter *, struct sge_eq *, int, int, uint8_t,
237     struct sge_iq *, char *);
238 static int alloc_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *,
239     struct sysctl_ctx_list *, struct sysctl_oid *);
240 static void free_iq_fl(struct adapter *, struct sge_iq *, struct sge_fl *);
241 static void add_iq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
242     struct sge_iq *);
243 static void add_fl_sysctls(struct adapter *, struct sysctl_ctx_list *,
244     struct sysctl_oid *, struct sge_fl *);
245 static int alloc_iq_fl_hwq(struct vi_info *, struct sge_iq *, struct sge_fl *);
246 static int free_iq_fl_hwq(struct adapter *, struct sge_iq *, struct sge_fl *);
247 static int alloc_fwq(struct adapter *);
248 static void free_fwq(struct adapter *);
249 static int alloc_ctrlq(struct adapter *, int);
250 static void free_ctrlq(struct adapter *, int);
251 static int alloc_rxq(struct vi_info *, struct sge_rxq *, int, int, int);
252 static void free_rxq(struct vi_info *, struct sge_rxq *);
253 static void add_rxq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
254     struct sge_rxq *);
255 #ifdef TCP_OFFLOAD
256 static int alloc_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *, int, int,
257     int);
258 static void free_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *);
259 static void add_ofld_rxq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
260     struct sge_ofld_rxq *);
261 #endif
262 static int ctrl_eq_alloc(struct adapter *, struct sge_eq *);
263 static int eth_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
264 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
265 static int ofld_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
266 #endif
267 static int alloc_eq(struct adapter *, struct sge_eq *, struct sysctl_ctx_list *,
268     struct sysctl_oid *);
269 static void free_eq(struct adapter *, struct sge_eq *);
270 static void add_eq_sysctls(struct adapter *, struct sysctl_ctx_list *,
271     struct sysctl_oid *, struct sge_eq *);
272 static int alloc_eq_hwq(struct adapter *, struct vi_info *, struct sge_eq *);
273 static int free_eq_hwq(struct adapter *, struct vi_info *, struct sge_eq *);
274 static int alloc_wrq(struct adapter *, struct vi_info *, struct sge_wrq *,
275     struct sysctl_ctx_list *, struct sysctl_oid *);
276 static void free_wrq(struct adapter *, struct sge_wrq *);
277 static void add_wrq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
278     struct sge_wrq *);
279 static int alloc_txq(struct vi_info *, struct sge_txq *, int);
280 static void free_txq(struct vi_info *, struct sge_txq *);
281 static void add_txq_sysctls(struct vi_info *, struct sysctl_ctx_list *,
282     struct sysctl_oid *, struct sge_txq *);
283 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
284 static int alloc_ofld_txq(struct vi_info *, struct sge_ofld_txq *, int);
285 static void free_ofld_txq(struct vi_info *, struct sge_ofld_txq *);
286 static void add_ofld_txq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
287     struct sge_ofld_txq *);
288 #endif
289 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int);
290 static inline void ring_fl_db(struct adapter *, struct sge_fl *);
291 static int refill_fl(struct adapter *, struct sge_fl *, int);
292 static void refill_sfl(void *);
293 static int find_refill_source(struct adapter *, int, bool);
294 static void add_fl_to_sfl(struct adapter *, struct sge_fl *);
295 
296 static inline void get_pkt_gl(struct mbuf *, struct sglist *);
297 static inline u_int txpkt_len16(u_int, const u_int);
298 static inline u_int txpkt_vm_len16(u_int, const u_int);
299 static inline void calculate_mbuf_len16(struct mbuf *, bool);
300 static inline u_int txpkts0_len16(u_int);
301 static inline u_int txpkts1_len16(void);
302 static u_int write_raw_wr(struct sge_txq *, void *, struct mbuf *, u_int);
303 static u_int write_txpkt_wr(struct adapter *, struct sge_txq *, struct mbuf *,
304     u_int);
305 static u_int write_txpkt_vm_wr(struct adapter *, struct sge_txq *,
306     struct mbuf *);
307 static int add_to_txpkts_vf(struct adapter *, struct sge_txq *, struct mbuf *,
308     int, bool *);
309 static int add_to_txpkts_pf(struct adapter *, struct sge_txq *, struct mbuf *,
310     int, bool *);
311 static u_int write_txpkts_wr(struct adapter *, struct sge_txq *);
312 static u_int write_txpkts_vm_wr(struct adapter *, struct sge_txq *);
313 static void write_gl_to_txd(struct sge_txq *, struct mbuf *, caddr_t *, int);
314 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int);
315 static inline void ring_eq_db(struct adapter *, struct sge_eq *, u_int);
316 static inline uint16_t read_hw_cidx(struct sge_eq *);
317 static inline u_int reclaimable_tx_desc(struct sge_eq *);
318 static inline u_int total_available_tx_desc(struct sge_eq *);
319 static u_int reclaim_tx_descs(struct sge_txq *, u_int);
320 static void tx_reclaim(void *, int);
321 static __be64 get_flit(struct sglist_seg *, int, int);
322 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
323     struct mbuf *);
324 static int handle_fw_msg(struct sge_iq *, const struct rss_header *,
325     struct mbuf *);
326 static int t4_handle_wrerr_rpl(struct adapter *, const __be64 *);
327 static void wrq_tx_drain(void *, int);
328 static void drain_wrq_wr_list(struct adapter *, struct sge_wrq *);
329 
330 static int sysctl_bufsizes(SYSCTL_HANDLER_ARGS);
331 #ifdef RATELIMIT
332 static int ethofld_fw4_ack(struct sge_iq *, const struct rss_header *,
333     struct mbuf *);
334 #if defined(INET) || defined(INET6)
335 static inline u_int txpkt_eo_len16(u_int, u_int, u_int);
336 static int ethofld_transmit(if_t, struct mbuf *);
337 #endif
338 #endif
339 
340 static counter_u64_t extfree_refs;
341 static counter_u64_t extfree_rels;
342 
343 an_handler_t t4_an_handler;
344 fw_msg_handler_t t4_fw_msg_handler[NUM_FW6_TYPES];
345 cpl_handler_t t4_cpl_handler[NUM_CPL_CMDS];
346 cpl_handler_t set_tcb_rpl_handlers[NUM_CPL_COOKIES];
347 cpl_handler_t l2t_write_rpl_handlers[NUM_CPL_COOKIES];
348 cpl_handler_t act_open_rpl_handlers[NUM_CPL_COOKIES];
349 cpl_handler_t abort_rpl_rss_handlers[NUM_CPL_COOKIES];
350 cpl_handler_t fw4_ack_handlers[NUM_CPL_COOKIES];
351 
352 void
353 t4_register_an_handler(an_handler_t h)
354 {
355 	uintptr_t *loc;
356 
357 	MPASS(h == NULL || t4_an_handler == NULL);
358 
359 	loc = (uintptr_t *)&t4_an_handler;
360 	atomic_store_rel_ptr(loc, (uintptr_t)h);
361 }
362 
363 void
364 t4_register_fw_msg_handler(int type, fw_msg_handler_t h)
365 {
366 	uintptr_t *loc;
367 
368 	MPASS(type < nitems(t4_fw_msg_handler));
369 	MPASS(h == NULL || t4_fw_msg_handler[type] == NULL);
370 	/*
371 	 * These are dispatched by the handler for FW{4|6}_CPL_MSG using the CPL
372 	 * handler dispatch table.  Reject any attempt to install a handler for
373 	 * this subtype.
374 	 */
375 	MPASS(type != FW_TYPE_RSSCPL);
376 	MPASS(type != FW6_TYPE_RSSCPL);
377 
378 	loc = (uintptr_t *)&t4_fw_msg_handler[type];
379 	atomic_store_rel_ptr(loc, (uintptr_t)h);
380 }
381 
382 void
383 t4_register_cpl_handler(int opcode, cpl_handler_t h)
384 {
385 	uintptr_t *loc;
386 
387 	MPASS(opcode < nitems(t4_cpl_handler));
388 	MPASS(h == NULL || t4_cpl_handler[opcode] == NULL);
389 
390 	loc = (uintptr_t *)&t4_cpl_handler[opcode];
391 	atomic_store_rel_ptr(loc, (uintptr_t)h);
392 }
393 
394 static int
395 set_tcb_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
396     struct mbuf *m)
397 {
398 	const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
399 	u_int tid;
400 	int cookie;
401 
402 	MPASS(m == NULL);
403 
404 	tid = GET_TID(cpl);
405 	if (is_hpftid(iq->adapter, tid) || is_ftid(iq->adapter, tid)) {
406 		/*
407 		 * The return code for filter-write is put in the CPL cookie so
408 		 * we have to rely on the hardware tid (is_ftid) to determine
409 		 * that this is a response to a filter.
410 		 */
411 		cookie = CPL_COOKIE_FILTER;
412 	} else {
413 		cookie = G_COOKIE(cpl->cookie);
414 	}
415 	MPASS(cookie > CPL_COOKIE_RESERVED);
416 	MPASS(cookie < nitems(set_tcb_rpl_handlers));
417 
418 	return (set_tcb_rpl_handlers[cookie](iq, rss, m));
419 }
420 
421 static int
422 l2t_write_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
423     struct mbuf *m)
424 {
425 	const struct cpl_l2t_write_rpl *rpl = (const void *)(rss + 1);
426 	unsigned int cookie;
427 
428 	MPASS(m == NULL);
429 
430 	cookie = GET_TID(rpl) & F_SYNC_WR ? CPL_COOKIE_TOM : CPL_COOKIE_FILTER;
431 	return (l2t_write_rpl_handlers[cookie](iq, rss, m));
432 }
433 
434 static int
435 act_open_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
436     struct mbuf *m)
437 {
438 	const struct cpl_act_open_rpl *cpl = (const void *)(rss + 1);
439 	u_int cookie = G_TID_COOKIE(G_AOPEN_ATID(be32toh(cpl->atid_status)));
440 
441 	MPASS(m == NULL);
442 	MPASS(cookie != CPL_COOKIE_RESERVED);
443 
444 	return (act_open_rpl_handlers[cookie](iq, rss, m));
445 }
446 
447 static int
448 abort_rpl_rss_handler(struct sge_iq *iq, const struct rss_header *rss,
449     struct mbuf *m)
450 {
451 	struct adapter *sc = iq->adapter;
452 	u_int cookie;
453 
454 	MPASS(m == NULL);
455 	if (is_hashfilter(sc))
456 		cookie = CPL_COOKIE_HASHFILTER;
457 	else
458 		cookie = CPL_COOKIE_TOM;
459 
460 	return (abort_rpl_rss_handlers[cookie](iq, rss, m));
461 }
462 
463 static int
464 fw4_ack_handler(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
465 {
466 	struct adapter *sc = iq->adapter;
467 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
468 	unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
469 	u_int cookie;
470 
471 	MPASS(m == NULL);
472 	if (is_etid(sc, tid))
473 		cookie = CPL_COOKIE_ETHOFLD;
474 	else
475 		cookie = CPL_COOKIE_TOM;
476 
477 	return (fw4_ack_handlers[cookie](iq, rss, m));
478 }
479 
480 static void
481 t4_init_shared_cpl_handlers(void)
482 {
483 
484 	t4_register_cpl_handler(CPL_SET_TCB_RPL, set_tcb_rpl_handler);
485 	t4_register_cpl_handler(CPL_L2T_WRITE_RPL, l2t_write_rpl_handler);
486 	t4_register_cpl_handler(CPL_ACT_OPEN_RPL, act_open_rpl_handler);
487 	t4_register_cpl_handler(CPL_ABORT_RPL_RSS, abort_rpl_rss_handler);
488 	t4_register_cpl_handler(CPL_FW4_ACK, fw4_ack_handler);
489 }
490 
491 void
492 t4_register_shared_cpl_handler(int opcode, cpl_handler_t h, int cookie)
493 {
494 	uintptr_t *loc;
495 
496 	MPASS(opcode < nitems(t4_cpl_handler));
497 	MPASS(cookie > CPL_COOKIE_RESERVED);
498 	MPASS(cookie < NUM_CPL_COOKIES);
499 	MPASS(t4_cpl_handler[opcode] != NULL);
500 
501 	switch (opcode) {
502 	case CPL_SET_TCB_RPL:
503 		loc = (uintptr_t *)&set_tcb_rpl_handlers[cookie];
504 		break;
505 	case CPL_L2T_WRITE_RPL:
506 		loc = (uintptr_t *)&l2t_write_rpl_handlers[cookie];
507 		break;
508 	case CPL_ACT_OPEN_RPL:
509 		loc = (uintptr_t *)&act_open_rpl_handlers[cookie];
510 		break;
511 	case CPL_ABORT_RPL_RSS:
512 		loc = (uintptr_t *)&abort_rpl_rss_handlers[cookie];
513 		break;
514 	case CPL_FW4_ACK:
515 		loc = (uintptr_t *)&fw4_ack_handlers[cookie];
516 		break;
517 	default:
518 		MPASS(0);
519 		return;
520 	}
521 	MPASS(h == NULL || *loc == (uintptr_t)NULL);
522 	atomic_store_rel_ptr(loc, (uintptr_t)h);
523 }
524 
525 /*
526  * Called on MOD_LOAD.  Validates and calculates the SGE tunables.
527  */
528 void
529 t4_sge_modload(void)
530 {
531 
532 	if (fl_pktshift < 0 || fl_pktshift > 7) {
533 		printf("Invalid hw.cxgbe.fl_pktshift value (%d),"
534 		    " using 0 instead.\n", fl_pktshift);
535 		fl_pktshift = 0;
536 	}
537 
538 	if (spg_len != 64 && spg_len != 128) {
539 		int len;
540 
541 #if defined(__i386__) || defined(__amd64__)
542 		len = cpu_clflush_line_size > 64 ? 128 : 64;
543 #else
544 		len = 64;
545 #endif
546 		if (spg_len != -1) {
547 			printf("Invalid hw.cxgbe.spg_len value (%d),"
548 			    " using %d instead.\n", spg_len, len);
549 		}
550 		spg_len = len;
551 	}
552 
553 	if (cong_drop < -1 || cong_drop > 2) {
554 		printf("Invalid hw.cxgbe.cong_drop value (%d),"
555 		    " using 0 instead.\n", cong_drop);
556 		cong_drop = 0;
557 	}
558 #ifdef TCP_OFFLOAD
559 	if (ofld_cong_drop < -1 || ofld_cong_drop > 2) {
560 		printf("Invalid hw.cxgbe.ofld_cong_drop value (%d),"
561 		    " using 0 instead.\n", ofld_cong_drop);
562 		ofld_cong_drop = 0;
563 	}
564 #endif
565 
566 	if (tscale != 1 && (tscale < 3 || tscale > 17)) {
567 		printf("Invalid hw.cxgbe.tscale value (%d),"
568 		    " using 1 instead.\n", tscale);
569 		tscale = 1;
570 	}
571 
572 	if (largest_rx_cluster != MCLBYTES &&
573 	    largest_rx_cluster != MJUMPAGESIZE &&
574 	    largest_rx_cluster != MJUM9BYTES &&
575 	    largest_rx_cluster != MJUM16BYTES) {
576 		printf("Invalid hw.cxgbe.largest_rx_cluster value (%d),"
577 		    " using %d instead.\n", largest_rx_cluster, MJUM16BYTES);
578 		largest_rx_cluster = MJUM16BYTES;
579 	}
580 
581 	if (safest_rx_cluster != MCLBYTES &&
582 	    safest_rx_cluster != MJUMPAGESIZE &&
583 	    safest_rx_cluster != MJUM9BYTES &&
584 	    safest_rx_cluster != MJUM16BYTES) {
585 		printf("Invalid hw.cxgbe.safest_rx_cluster value (%d),"
586 		    " using %d instead.\n", safest_rx_cluster, MJUMPAGESIZE);
587 		safest_rx_cluster = MJUMPAGESIZE;
588 	}
589 
590 	extfree_refs = counter_u64_alloc(M_WAITOK);
591 	extfree_rels = counter_u64_alloc(M_WAITOK);
592 	pullups = counter_u64_alloc(M_WAITOK);
593 	defrags = counter_u64_alloc(M_WAITOK);
594 	counter_u64_zero(extfree_refs);
595 	counter_u64_zero(extfree_rels);
596 	counter_u64_zero(pullups);
597 	counter_u64_zero(defrags);
598 
599 	t4_init_shared_cpl_handlers();
600 	t4_register_cpl_handler(CPL_FW4_MSG, handle_fw_msg);
601 	t4_register_cpl_handler(CPL_FW6_MSG, handle_fw_msg);
602 	t4_register_cpl_handler(CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
603 #ifdef RATELIMIT
604 	t4_register_shared_cpl_handler(CPL_FW4_ACK, ethofld_fw4_ack,
605 	    CPL_COOKIE_ETHOFLD);
606 #endif
607 	t4_register_fw_msg_handler(FW6_TYPE_CMD_RPL, t4_handle_fw_rpl);
608 	t4_register_fw_msg_handler(FW6_TYPE_WRERR_RPL, t4_handle_wrerr_rpl);
609 }
610 
611 void
612 t4_sge_modunload(void)
613 {
614 
615 	counter_u64_free(extfree_refs);
616 	counter_u64_free(extfree_rels);
617 	counter_u64_free(pullups);
618 	counter_u64_free(defrags);
619 }
620 
621 uint64_t
622 t4_sge_extfree_refs(void)
623 {
624 	uint64_t refs, rels;
625 
626 	rels = counter_u64_fetch(extfree_rels);
627 	refs = counter_u64_fetch(extfree_refs);
628 
629 	return (refs - rels);
630 }
631 
632 /* max 4096 */
633 #define MAX_PACK_BOUNDARY 512
634 
635 static inline void
636 setup_pad_and_pack_boundaries(struct adapter *sc)
637 {
638 	uint32_t v, m;
639 	int pad, pack, pad_shift;
640 
641 	pad_shift = chip_id(sc) > CHELSIO_T5 ? X_T6_INGPADBOUNDARY_SHIFT :
642 	    X_INGPADBOUNDARY_SHIFT;
643 	pad = fl_pad;
644 	if (fl_pad < (1 << pad_shift) ||
645 	    fl_pad > (1 << (pad_shift + M_INGPADBOUNDARY)) ||
646 	    !powerof2(fl_pad)) {
647 		/*
648 		 * If there is any chance that we might use buffer packing and
649 		 * the chip is a T4, then pick 64 as the pad/pack boundary.  Set
650 		 * it to the minimum allowed in all other cases.
651 		 */
652 		pad = is_t4(sc) && buffer_packing ? 64 : 1 << pad_shift;
653 
654 		/*
655 		 * For fl_pad = 0 we'll still write a reasonable value to the
656 		 * register but all the freelists will opt out of padding.
657 		 * We'll complain here only if the user tried to set it to a
658 		 * value greater than 0 that was invalid.
659 		 */
660 		if (fl_pad > 0) {
661 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pad value"
662 			    " (%d), using %d instead.\n", fl_pad, pad);
663 		}
664 	}
665 	m = V_INGPADBOUNDARY(M_INGPADBOUNDARY);
666 	v = V_INGPADBOUNDARY(ilog2(pad) - pad_shift);
667 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
668 
669 	if (is_t4(sc)) {
670 		if (fl_pack != -1 && fl_pack != pad) {
671 			/* Complain but carry on. */
672 			device_printf(sc->dev, "hw.cxgbe.fl_pack (%d) ignored,"
673 			    " using %d instead.\n", fl_pack, pad);
674 		}
675 		return;
676 	}
677 
678 	pack = fl_pack;
679 	if (fl_pack < 16 || fl_pack == 32 || fl_pack > 4096 ||
680 	    !powerof2(fl_pack)) {
681 		if (sc->params.pci.mps > MAX_PACK_BOUNDARY)
682 			pack = MAX_PACK_BOUNDARY;
683 		else
684 			pack = max(sc->params.pci.mps, CACHE_LINE_SIZE);
685 		MPASS(powerof2(pack));
686 		if (pack < 16)
687 			pack = 16;
688 		if (pack == 32)
689 			pack = 64;
690 		if (pack > 4096)
691 			pack = 4096;
692 		if (fl_pack != -1) {
693 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pack value"
694 			    " (%d), using %d instead.\n", fl_pack, pack);
695 		}
696 	}
697 	m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY);
698 	if (pack == 16)
699 		v = V_INGPACKBOUNDARY(0);
700 	else
701 		v = V_INGPACKBOUNDARY(ilog2(pack) - 5);
702 
703 	MPASS(!is_t4(sc));	/* T4 doesn't have SGE_CONTROL2 */
704 	t4_set_reg_field(sc, A_SGE_CONTROL2, m, v);
705 }
706 
707 /*
708  * adap->params.vpd.cclk must be set up before this is called.
709  */
710 void
711 t4_tweak_chip_settings(struct adapter *sc)
712 {
713 	int i, reg;
714 	uint32_t v, m;
715 	int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200};
716 	int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk;
717 	int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */
718 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
719 	static int sw_buf_sizes[] = {
720 		MCLBYTES,
721 		MJUMPAGESIZE,
722 		MJUM9BYTES,
723 		MJUM16BYTES
724 	};
725 
726 	KASSERT(sc->flags & MASTER_PF,
727 	    ("%s: trying to change chip settings when not master.", __func__));
728 
729 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
730 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
731 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
732 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
733 
734 	setup_pad_and_pack_boundaries(sc);
735 
736 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
737 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
738 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
739 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
740 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
741 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
742 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
743 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
744 	t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v);
745 
746 	t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0, 4096);
747 	t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE1, 65536);
748 	reg = A_SGE_FL_BUFFER_SIZE2;
749 	for (i = 0; i < nitems(sw_buf_sizes); i++) {
750 		MPASS(reg <= A_SGE_FL_BUFFER_SIZE15);
751 		t4_write_reg(sc, reg, sw_buf_sizes[i]);
752 		reg += 4;
753 		MPASS(reg <= A_SGE_FL_BUFFER_SIZE15);
754 		t4_write_reg(sc, reg, sw_buf_sizes[i] - CL_METADATA_SIZE);
755 		reg += 4;
756 	}
757 
758 	v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) |
759 	    V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]);
760 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v);
761 
762 	KASSERT(intr_timer[0] <= timer_max,
763 	    ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0],
764 	    timer_max));
765 	for (i = 1; i < nitems(intr_timer); i++) {
766 		KASSERT(intr_timer[i] >= intr_timer[i - 1],
767 		    ("%s: timers not listed in increasing order (%d)",
768 		    __func__, i));
769 
770 		while (intr_timer[i] > timer_max) {
771 			if (i == nitems(intr_timer) - 1) {
772 				intr_timer[i] = timer_max;
773 				break;
774 			}
775 			intr_timer[i] += intr_timer[i - 1];
776 			intr_timer[i] /= 2;
777 		}
778 	}
779 
780 	v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) |
781 	    V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1]));
782 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v);
783 	v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) |
784 	    V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3]));
785 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v);
786 	v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) |
787 	    V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5]));
788 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v);
789 
790 	if (chip_id(sc) >= CHELSIO_T6) {
791 		m = V_TSCALE(M_TSCALE);
792 		if (tscale == 1)
793 			v = 0;
794 		else
795 			v = V_TSCALE(tscale - 2);
796 		t4_set_reg_field(sc, A_SGE_ITP_CONTROL, m, v);
797 
798 		if (sc->debug_flags & DF_DISABLE_TCB_CACHE) {
799 			m = V_RDTHRESHOLD(M_RDTHRESHOLD) | F_WRTHRTHRESHEN |
800 			    V_WRTHRTHRESH(M_WRTHRTHRESH);
801 			t4_tp_pio_read(sc, &v, 1, A_TP_CMM_CONFIG, 1);
802 			v &= ~m;
803 			v |= V_RDTHRESHOLD(1) | F_WRTHRTHRESHEN |
804 			    V_WRTHRTHRESH(16);
805 			t4_tp_pio_write(sc, &v, 1, A_TP_CMM_CONFIG, 1);
806 		}
807 	}
808 
809 	/* 4K, 16K, 64K, 256K DDP "page sizes" for TDDP */
810 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
811 	t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v);
812 
813 	/*
814 	 * 4K, 8K, 16K, 64K DDP "page sizes" for iSCSI DDP.  These have been
815 	 * chosen with MAXPHYS = 128K in mind.  The largest DDP buffer that we
816 	 * may have to deal with is MAXPHYS + 1 page.
817 	 */
818 	v = V_HPZ0(0) | V_HPZ1(1) | V_HPZ2(2) | V_HPZ3(4);
819 	t4_write_reg(sc, A_ULP_RX_ISCSI_PSZ, v);
820 
821 	/* We use multiple DDP page sizes both in plain-TOE and ISCSI modes. */
822 	m = v = F_TDDPTAGTCB | F_ISCSITAGTCB;
823 	t4_set_reg_field(sc, A_ULP_RX_CTL, m, v);
824 
825 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
826 	    F_RESETDDPOFFSET;
827 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
828 	t4_set_reg_field(sc, A_TP_PARA_REG5, m, v);
829 }
830 
831 /*
832  * SGE wants the buffer to be at least 64B and then a multiple of 16.  Its
833  * address mut be 16B aligned.  If padding is in use the buffer's start and end
834  * need to be aligned to the pad boundary as well.  We'll just make sure that
835  * the size is a multiple of the pad boundary here, it is up to the buffer
836  * allocation code to make sure the start of the buffer is aligned.
837  */
838 static inline int
839 hwsz_ok(struct adapter *sc, int hwsz)
840 {
841 	int mask = fl_pad ? sc->params.sge.pad_boundary - 1 : 16 - 1;
842 
843 	return (hwsz >= 64 && (hwsz & mask) == 0);
844 }
845 
846 /*
847  * Initialize the rx buffer sizes and figure out which zones the buffers will
848  * be allocated from.
849  */
850 void
851 t4_init_rx_buf_info(struct adapter *sc)
852 {
853 	struct sge *s = &sc->sge;
854 	struct sge_params *sp = &sc->params.sge;
855 	int i, j, n;
856 	static int sw_buf_sizes[] = {	/* Sorted by size */
857 		MCLBYTES,
858 		MJUMPAGESIZE,
859 		MJUM9BYTES,
860 		MJUM16BYTES
861 	};
862 	struct rx_buf_info *rxb;
863 
864 	s->safe_zidx = -1;
865 	rxb = &s->rx_buf_info[0];
866 	for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) {
867 		rxb->size1 = sw_buf_sizes[i];
868 		rxb->zone = m_getzone(rxb->size1);
869 		rxb->type = m_gettype(rxb->size1);
870 		rxb->size2 = 0;
871 		rxb->hwidx1 = -1;
872 		rxb->hwidx2 = -1;
873 		for (j = 0; j < SGE_FLBUF_SIZES; j++) {
874 			int hwsize = sp->sge_fl_buffer_size[j];
875 
876 			if (!hwsz_ok(sc, hwsize))
877 				continue;
878 
879 			/* hwidx for size1 */
880 			if (rxb->hwidx1 == -1 && rxb->size1 == hwsize)
881 				rxb->hwidx1 = j;
882 
883 			/* hwidx for size2 (buffer packing) */
884 			if (rxb->size1 - CL_METADATA_SIZE < hwsize)
885 				continue;
886 			n = rxb->size1 - hwsize - CL_METADATA_SIZE;
887 			if (n == 0) {
888 				rxb->hwidx2 = j;
889 				rxb->size2 = hwsize;
890 				break;	/* stop looking */
891 			}
892 			if (rxb->hwidx2 != -1) {
893 				if (n < sp->sge_fl_buffer_size[rxb->hwidx2] -
894 				    hwsize - CL_METADATA_SIZE) {
895 					rxb->hwidx2 = j;
896 					rxb->size2 = hwsize;
897 				}
898 			} else if (n <= 2 * CL_METADATA_SIZE) {
899 				rxb->hwidx2 = j;
900 				rxb->size2 = hwsize;
901 			}
902 		}
903 		if (rxb->hwidx2 != -1)
904 			sc->flags |= BUF_PACKING_OK;
905 		if (s->safe_zidx == -1 && rxb->size1 == safest_rx_cluster)
906 			s->safe_zidx = i;
907 	}
908 }
909 
910 /*
911  * Verify some basic SGE settings for the PF and VF driver, and other
912  * miscellaneous settings for the PF driver.
913  */
914 int
915 t4_verify_chip_settings(struct adapter *sc)
916 {
917 	struct sge_params *sp = &sc->params.sge;
918 	uint32_t m, v, r;
919 	int rc = 0;
920 	const uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
921 
922 	m = F_RXPKTCPLMODE;
923 	v = F_RXPKTCPLMODE;
924 	r = sp->sge_control;
925 	if ((r & m) != v) {
926 		device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r);
927 		rc = EINVAL;
928 	}
929 
930 	/*
931 	 * If this changes then every single use of PAGE_SHIFT in the driver
932 	 * needs to be carefully reviewed for PAGE_SHIFT vs sp->page_shift.
933 	 */
934 	if (sp->page_shift != PAGE_SHIFT) {
935 		device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r);
936 		rc = EINVAL;
937 	}
938 
939 	if (sc->flags & IS_VF)
940 		return (0);
941 
942 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
943 	r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ);
944 	if (r != v) {
945 		device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r);
946 		if (sc->vres.ddp.size != 0)
947 			rc = EINVAL;
948 	}
949 
950 	m = v = F_TDDPTAGTCB;
951 	r = t4_read_reg(sc, A_ULP_RX_CTL);
952 	if ((r & m) != v) {
953 		device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r);
954 		if (sc->vres.ddp.size != 0)
955 			rc = EINVAL;
956 	}
957 
958 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
959 	    F_RESETDDPOFFSET;
960 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
961 	r = t4_read_reg(sc, A_TP_PARA_REG5);
962 	if ((r & m) != v) {
963 		device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r);
964 		if (sc->vres.ddp.size != 0)
965 			rc = EINVAL;
966 	}
967 
968 	return (rc);
969 }
970 
971 int
972 t4_create_dma_tag(struct adapter *sc)
973 {
974 	int rc;
975 
976 	rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
977 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
978 	    BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
979 	    NULL, &sc->dmat);
980 	if (rc != 0) {
981 		device_printf(sc->dev,
982 		    "failed to create main DMA tag: %d\n", rc);
983 	}
984 
985 	return (rc);
986 }
987 
988 void
989 t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
990     struct sysctl_oid_list *children)
991 {
992 	struct sge_params *sp = &sc->params.sge;
993 
994 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "buffer_sizes",
995 	    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, sc, 0,
996 	    sysctl_bufsizes, "A", "freelist buffer sizes");
997 
998 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD,
999 	    NULL, sp->fl_pktshift, "payload DMA offset in rx buffer (bytes)");
1000 
1001 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD,
1002 	    NULL, sp->pad_boundary, "payload pad boundary (bytes)");
1003 
1004 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD,
1005 	    NULL, sp->spg_len, "status page size (bytes)");
1006 
1007 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD,
1008 	    NULL, cong_drop, "congestion drop setting");
1009 #ifdef TCP_OFFLOAD
1010 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "ofld_cong_drop", CTLFLAG_RD,
1011 	    NULL, ofld_cong_drop, "congestion drop setting");
1012 #endif
1013 
1014 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD,
1015 	    NULL, sp->pack_boundary, "payload pack boundary (bytes)");
1016 }
1017 
1018 int
1019 t4_destroy_dma_tag(struct adapter *sc)
1020 {
1021 	if (sc->dmat)
1022 		bus_dma_tag_destroy(sc->dmat);
1023 
1024 	return (0);
1025 }
1026 
1027 /*
1028  * Allocate and initialize the firmware event queue, control queues, and special
1029  * purpose rx queues owned by the adapter.
1030  *
1031  * Returns errno on failure.  Resources allocated up to that point may still be
1032  * allocated.  Caller is responsible for cleanup in case this function fails.
1033  */
1034 int
1035 t4_setup_adapter_queues(struct adapter *sc)
1036 {
1037 	int rc, i;
1038 
1039 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
1040 
1041 	/*
1042 	 * Firmware event queue
1043 	 */
1044 	rc = alloc_fwq(sc);
1045 	if (rc != 0)
1046 		return (rc);
1047 
1048 	/*
1049 	 * That's all for the VF driver.
1050 	 */
1051 	if (sc->flags & IS_VF)
1052 		return (rc);
1053 
1054 	/*
1055 	 * XXX: General purpose rx queues, one per port.
1056 	 */
1057 
1058 	/*
1059 	 * Control queues, one per port.
1060 	 */
1061 	for_each_port(sc, i) {
1062 		rc = alloc_ctrlq(sc, i);
1063 		if (rc != 0)
1064 			return (rc);
1065 	}
1066 
1067 	return (rc);
1068 }
1069 
1070 /*
1071  * Idempotent
1072  */
1073 int
1074 t4_teardown_adapter_queues(struct adapter *sc)
1075 {
1076 	int i;
1077 
1078 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
1079 
1080 	if (sc->sge.ctrlq != NULL) {
1081 		MPASS(!(sc->flags & IS_VF));	/* VFs don't allocate ctrlq. */
1082 		for_each_port(sc, i)
1083 			free_ctrlq(sc, i);
1084 	}
1085 	free_fwq(sc);
1086 
1087 	return (0);
1088 }
1089 
1090 /* Maximum payload that could arrive with a single iq descriptor. */
1091 static inline int
1092 max_rx_payload(struct adapter *sc, if_t ifp, const bool ofld)
1093 {
1094 	int maxp;
1095 
1096 	/* large enough even when hw VLAN extraction is disabled */
1097 	maxp = sc->params.sge.fl_pktshift + ETHER_HDR_LEN +
1098 	    ETHER_VLAN_ENCAP_LEN + if_getmtu(ifp);
1099 	if (ofld && sc->tt.tls && sc->cryptocaps & FW_CAPS_CONFIG_TLSKEYS &&
1100 	    maxp < sc->params.tp.max_rx_pdu)
1101 		maxp = sc->params.tp.max_rx_pdu;
1102 	return (maxp);
1103 }
1104 
1105 int
1106 t4_setup_vi_queues(struct vi_info *vi)
1107 {
1108 	int rc = 0, i, intr_idx;
1109 	struct sge_rxq *rxq;
1110 	struct sge_txq *txq;
1111 #ifdef TCP_OFFLOAD
1112 	struct sge_ofld_rxq *ofld_rxq;
1113 #endif
1114 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1115 	struct sge_ofld_txq *ofld_txq;
1116 #endif
1117 #ifdef DEV_NETMAP
1118 	int saved_idx, iqidx;
1119 	struct sge_nm_rxq *nm_rxq;
1120 	struct sge_nm_txq *nm_txq;
1121 #endif
1122 	struct adapter *sc = vi->adapter;
1123 	if_t ifp = vi->ifp;
1124 	int maxp;
1125 
1126 	/* Interrupt vector to start from (when using multiple vectors) */
1127 	intr_idx = vi->first_intr;
1128 
1129 #ifdef DEV_NETMAP
1130 	saved_idx = intr_idx;
1131 	if (if_getcapabilities(ifp) & IFCAP_NETMAP) {
1132 
1133 		/* netmap is supported with direct interrupts only. */
1134 		MPASS(!forwarding_intr_to_fwq(sc));
1135 		MPASS(vi->first_intr >= 0);
1136 
1137 		/*
1138 		 * We don't have buffers to back the netmap rx queues
1139 		 * right now so we create the queues in a way that
1140 		 * doesn't set off any congestion signal in the chip.
1141 		 */
1142 		for_each_nm_rxq(vi, i, nm_rxq) {
1143 			rc = alloc_nm_rxq(vi, nm_rxq, intr_idx, i);
1144 			if (rc != 0)
1145 				goto done;
1146 			intr_idx++;
1147 		}
1148 
1149 		for_each_nm_txq(vi, i, nm_txq) {
1150 			iqidx = vi->first_nm_rxq + (i % vi->nnmrxq);
1151 			rc = alloc_nm_txq(vi, nm_txq, iqidx, i);
1152 			if (rc != 0)
1153 				goto done;
1154 		}
1155 	}
1156 
1157 	/* Normal rx queues and netmap rx queues share the same interrupts. */
1158 	intr_idx = saved_idx;
1159 #endif
1160 
1161 	/*
1162 	 * Allocate rx queues first because a default iqid is required when
1163 	 * creating a tx queue.
1164 	 */
1165 	maxp = max_rx_payload(sc, ifp, false);
1166 	for_each_rxq(vi, i, rxq) {
1167 		rc = alloc_rxq(vi, rxq, i, intr_idx, maxp);
1168 		if (rc != 0)
1169 			goto done;
1170 		if (!forwarding_intr_to_fwq(sc))
1171 			intr_idx++;
1172 	}
1173 #ifdef DEV_NETMAP
1174 	if (if_getcapabilities(ifp) & IFCAP_NETMAP)
1175 		intr_idx = saved_idx + max(vi->nrxq, vi->nnmrxq);
1176 #endif
1177 #ifdef TCP_OFFLOAD
1178 	maxp = max_rx_payload(sc, ifp, true);
1179 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1180 		rc = alloc_ofld_rxq(vi, ofld_rxq, i, intr_idx, maxp);
1181 		if (rc != 0)
1182 			goto done;
1183 		if (!forwarding_intr_to_fwq(sc))
1184 			intr_idx++;
1185 	}
1186 #endif
1187 
1188 	/*
1189 	 * Now the tx queues.
1190 	 */
1191 	for_each_txq(vi, i, txq) {
1192 		rc = alloc_txq(vi, txq, i);
1193 		if (rc != 0)
1194 			goto done;
1195 	}
1196 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1197 	for_each_ofld_txq(vi, i, ofld_txq) {
1198 		rc = alloc_ofld_txq(vi, ofld_txq, i);
1199 		if (rc != 0)
1200 			goto done;
1201 	}
1202 #endif
1203 done:
1204 	if (rc)
1205 		t4_teardown_vi_queues(vi);
1206 
1207 	return (rc);
1208 }
1209 
1210 /*
1211  * Idempotent
1212  */
1213 int
1214 t4_teardown_vi_queues(struct vi_info *vi)
1215 {
1216 	int i;
1217 	struct sge_rxq *rxq;
1218 	struct sge_txq *txq;
1219 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1220 	struct sge_ofld_txq *ofld_txq;
1221 #endif
1222 #ifdef TCP_OFFLOAD
1223 	struct sge_ofld_rxq *ofld_rxq;
1224 #endif
1225 #ifdef DEV_NETMAP
1226 	struct sge_nm_rxq *nm_rxq;
1227 	struct sge_nm_txq *nm_txq;
1228 #endif
1229 
1230 #ifdef DEV_NETMAP
1231 	if (if_getcapabilities(vi->ifp) & IFCAP_NETMAP) {
1232 		for_each_nm_txq(vi, i, nm_txq) {
1233 			free_nm_txq(vi, nm_txq);
1234 		}
1235 
1236 		for_each_nm_rxq(vi, i, nm_rxq) {
1237 			free_nm_rxq(vi, nm_rxq);
1238 		}
1239 	}
1240 #endif
1241 
1242 	/*
1243 	 * Take down all the tx queues first, as they reference the rx queues
1244 	 * (for egress updates, etc.).
1245 	 */
1246 
1247 	for_each_txq(vi, i, txq) {
1248 		free_txq(vi, txq);
1249 	}
1250 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1251 	for_each_ofld_txq(vi, i, ofld_txq) {
1252 		free_ofld_txq(vi, ofld_txq);
1253 	}
1254 #endif
1255 
1256 	/*
1257 	 * Then take down the rx queues.
1258 	 */
1259 
1260 	for_each_rxq(vi, i, rxq) {
1261 		free_rxq(vi, rxq);
1262 	}
1263 #ifdef TCP_OFFLOAD
1264 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1265 		free_ofld_rxq(vi, ofld_rxq);
1266 	}
1267 #endif
1268 
1269 	return (0);
1270 }
1271 
1272 /*
1273  * Interrupt handler when the driver is using only 1 interrupt.  This is a very
1274  * unusual scenario.
1275  *
1276  * a) Deals with errors, if any.
1277  * b) Services firmware event queue, which is taking interrupts for all other
1278  *    queues.
1279  */
1280 void
1281 t4_intr_all(void *arg)
1282 {
1283 	struct adapter *sc = arg;
1284 	struct sge_iq *fwq = &sc->sge.fwq;
1285 
1286 	MPASS(sc->intr_count == 1);
1287 
1288 	if (sc->intr_type == INTR_INTX)
1289 		t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
1290 
1291 	t4_intr_err(arg);
1292 	t4_intr_evt(fwq);
1293 }
1294 
1295 /*
1296  * Interrupt handler for errors (installed directly when multiple interrupts are
1297  * being used, or called by t4_intr_all).
1298  */
1299 void
1300 t4_intr_err(void *arg)
1301 {
1302 	struct adapter *sc = arg;
1303 	uint32_t v;
1304 	const bool verbose = (sc->debug_flags & DF_VERBOSE_SLOWINTR) != 0;
1305 
1306 	if (atomic_load_int(&sc->error_flags) & ADAP_FATAL_ERR)
1307 		return;
1308 
1309 	v = t4_read_reg(sc, MYPF_REG(A_PL_PF_INT_CAUSE));
1310 	if (v & F_PFSW) {
1311 		sc->swintr++;
1312 		t4_write_reg(sc, MYPF_REG(A_PL_PF_INT_CAUSE), v);
1313 	}
1314 
1315 	if (t4_slow_intr_handler(sc, verbose))
1316 		t4_fatal_err(sc, false);
1317 }
1318 
1319 /*
1320  * Interrupt handler for iq-only queues.  The firmware event queue is the only
1321  * such queue right now.
1322  */
1323 void
1324 t4_intr_evt(void *arg)
1325 {
1326 	struct sge_iq *iq = arg;
1327 
1328 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1329 		service_iq(iq, 0);
1330 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1331 	}
1332 }
1333 
1334 /*
1335  * Interrupt handler for iq+fl queues.
1336  */
1337 void
1338 t4_intr(void *arg)
1339 {
1340 	struct sge_iq *iq = arg;
1341 
1342 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1343 		service_iq_fl(iq, 0);
1344 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1345 	}
1346 }
1347 
1348 #ifdef DEV_NETMAP
1349 /*
1350  * Interrupt handler for netmap rx queues.
1351  */
1352 void
1353 t4_nm_intr(void *arg)
1354 {
1355 	struct sge_nm_rxq *nm_rxq = arg;
1356 
1357 	if (atomic_cmpset_int(&nm_rxq->nm_state, NM_ON, NM_BUSY)) {
1358 		service_nm_rxq(nm_rxq);
1359 		(void) atomic_cmpset_int(&nm_rxq->nm_state, NM_BUSY, NM_ON);
1360 	}
1361 }
1362 
1363 /*
1364  * Interrupt handler for vectors shared between NIC and netmap rx queues.
1365  */
1366 void
1367 t4_vi_intr(void *arg)
1368 {
1369 	struct irq *irq = arg;
1370 
1371 	MPASS(irq->nm_rxq != NULL);
1372 	t4_nm_intr(irq->nm_rxq);
1373 
1374 	MPASS(irq->rxq != NULL);
1375 	t4_intr(irq->rxq);
1376 }
1377 #endif
1378 
1379 /*
1380  * Deals with interrupts on an iq-only (no freelist) queue.
1381  */
1382 static int
1383 service_iq(struct sge_iq *iq, int budget)
1384 {
1385 	struct sge_iq *q;
1386 	struct adapter *sc = iq->adapter;
1387 	struct iq_desc *d = &iq->desc[iq->cidx];
1388 	int ndescs = 0, limit;
1389 	int rsp_type;
1390 	uint32_t lq;
1391 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1392 
1393 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1394 	KASSERT((iq->flags & IQ_HAS_FL) == 0,
1395 	    ("%s: called for iq %p with fl (iq->flags 0x%x)", __func__, iq,
1396 	    iq->flags));
1397 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1398 	MPASS((iq->flags & IQ_LRO_ENABLED) == 0);
1399 
1400 	limit = budget ? budget : iq->qsize / 16;
1401 
1402 	/*
1403 	 * We always come back and check the descriptor ring for new indirect
1404 	 * interrupts and other responses after running a single handler.
1405 	 */
1406 	for (;;) {
1407 		while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1408 
1409 			rmb();
1410 
1411 			rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1412 			lq = be32toh(d->rsp.pldbuflen_qid);
1413 
1414 			switch (rsp_type) {
1415 			case X_RSPD_TYPE_FLBUF:
1416 				panic("%s: data for an iq (%p) with no freelist",
1417 				    __func__, iq);
1418 
1419 				/* NOTREACHED */
1420 
1421 			case X_RSPD_TYPE_CPL:
1422 				KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1423 				    ("%s: bad opcode %02x.", __func__,
1424 				    d->rss.opcode));
1425 				t4_cpl_handler[d->rss.opcode](iq, &d->rss, NULL);
1426 				break;
1427 
1428 			case X_RSPD_TYPE_INTR:
1429 				/*
1430 				 * There are 1K interrupt-capable queues (qids 0
1431 				 * through 1023).  A response type indicating a
1432 				 * forwarded interrupt with a qid >= 1K is an
1433 				 * iWARP async notification.
1434 				 */
1435 				if (__predict_true(lq >= 1024)) {
1436 					t4_an_handler(iq, &d->rsp);
1437 					break;
1438 				}
1439 
1440 				q = sc->sge.iqmap[lq - sc->sge.iq_start -
1441 				    sc->sge.iq_base];
1442 				if (atomic_cmpset_int(&q->state, IQS_IDLE,
1443 				    IQS_BUSY)) {
1444 					if (service_iq_fl(q, q->qsize / 16) == 0) {
1445 						(void) atomic_cmpset_int(&q->state,
1446 						    IQS_BUSY, IQS_IDLE);
1447 					} else {
1448 						STAILQ_INSERT_TAIL(&iql, q,
1449 						    link);
1450 					}
1451 				}
1452 				break;
1453 
1454 			default:
1455 				KASSERT(0,
1456 				    ("%s: illegal response type %d on iq %p",
1457 				    __func__, rsp_type, iq));
1458 				log(LOG_ERR,
1459 				    "%s: illegal response type %d on iq %p",
1460 				    device_get_nameunit(sc->dev), rsp_type, iq);
1461 				break;
1462 			}
1463 
1464 			d++;
1465 			if (__predict_false(++iq->cidx == iq->sidx)) {
1466 				iq->cidx = 0;
1467 				iq->gen ^= F_RSPD_GEN;
1468 				d = &iq->desc[0];
1469 			}
1470 			if (__predict_false(++ndescs == limit)) {
1471 				t4_write_reg(sc, sc->sge_gts_reg,
1472 				    V_CIDXINC(ndescs) |
1473 				    V_INGRESSQID(iq->cntxt_id) |
1474 				    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1475 				ndescs = 0;
1476 
1477 				if (budget) {
1478 					return (EINPROGRESS);
1479 				}
1480 			}
1481 		}
1482 
1483 		if (STAILQ_EMPTY(&iql))
1484 			break;
1485 
1486 		/*
1487 		 * Process the head only, and send it to the back of the list if
1488 		 * it's still not done.
1489 		 */
1490 		q = STAILQ_FIRST(&iql);
1491 		STAILQ_REMOVE_HEAD(&iql, link);
1492 		if (service_iq_fl(q, q->qsize / 8) == 0)
1493 			(void) atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1494 		else
1495 			STAILQ_INSERT_TAIL(&iql, q, link);
1496 	}
1497 
1498 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1499 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1500 
1501 	return (0);
1502 }
1503 
1504 #if defined(INET) || defined(INET6)
1505 static inline int
1506 sort_before_lro(struct lro_ctrl *lro)
1507 {
1508 
1509 	return (lro->lro_mbuf_max != 0);
1510 }
1511 #endif
1512 
1513 #define CGBE_SHIFT_SCALE 10
1514 
1515 static inline uint64_t
1516 t4_tstmp_to_ns(struct adapter *sc, uint64_t lf)
1517 {
1518 	struct clock_sync *cur, dcur;
1519 	uint64_t hw_clocks;
1520 	uint64_t hw_clk_div;
1521 	sbintime_t sbt_cur_to_prev, sbt;
1522 	uint64_t hw_tstmp = lf & 0xfffffffffffffffULL;	/* 60b, not 64b. */
1523 	seqc_t gen;
1524 
1525 	for (;;) {
1526 		cur = &sc->cal_info[sc->cal_current];
1527 		gen = seqc_read(&cur->gen);
1528 		if (gen == 0)
1529 			return (0);
1530 		dcur = *cur;
1531 		if (seqc_consistent(&cur->gen, gen))
1532 			break;
1533 	}
1534 
1535 	/*
1536 	 * Our goal here is to have a result that is:
1537 	 *
1538 	 * (                             (cur_time - prev_time)   )
1539 	 * ((hw_tstmp - hw_prev) *  ----------------------------- ) + prev_time
1540 	 * (                             (hw_cur - hw_prev)       )
1541 	 *
1542 	 * With the constraints that we cannot use float and we
1543 	 * don't want to overflow the uint64_t numbers we are using.
1544 	 */
1545 	hw_clocks = hw_tstmp - dcur.hw_prev;
1546 	sbt_cur_to_prev = (dcur.sbt_cur - dcur.sbt_prev);
1547 	hw_clk_div = dcur.hw_cur - dcur.hw_prev;
1548 	sbt = hw_clocks * sbt_cur_to_prev / hw_clk_div + dcur.sbt_prev;
1549 	return (sbttons(sbt));
1550 }
1551 
1552 static inline void
1553 move_to_next_rxbuf(struct sge_fl *fl)
1554 {
1555 
1556 	fl->rx_offset = 0;
1557 	if (__predict_false((++fl->cidx & 7) == 0)) {
1558 		uint16_t cidx = fl->cidx >> 3;
1559 
1560 		if (__predict_false(cidx == fl->sidx))
1561 			fl->cidx = cidx = 0;
1562 		fl->hw_cidx = cidx;
1563 	}
1564 }
1565 
1566 /*
1567  * Deals with interrupts on an iq+fl queue.
1568  */
1569 static int
1570 service_iq_fl(struct sge_iq *iq, int budget)
1571 {
1572 	struct sge_rxq *rxq = iq_to_rxq(iq);
1573 	struct sge_fl *fl;
1574 	struct adapter *sc = iq->adapter;
1575 	struct iq_desc *d = &iq->desc[iq->cidx];
1576 	int ndescs, limit;
1577 	int rsp_type, starved;
1578 	uint32_t lq;
1579 	uint16_t fl_hw_cidx;
1580 	struct mbuf *m0;
1581 #if defined(INET) || defined(INET6)
1582 	const struct timeval lro_timeout = {0, sc->lro_timeout};
1583 	struct lro_ctrl *lro = &rxq->lro;
1584 #endif
1585 
1586 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1587 	MPASS(iq->flags & IQ_HAS_FL);
1588 
1589 	ndescs = 0;
1590 #if defined(INET) || defined(INET6)
1591 	if (iq->flags & IQ_ADJ_CREDIT) {
1592 		MPASS(sort_before_lro(lro));
1593 		iq->flags &= ~IQ_ADJ_CREDIT;
1594 		if ((d->rsp.u.type_gen & F_RSPD_GEN) != iq->gen) {
1595 			tcp_lro_flush_all(lro);
1596 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(1) |
1597 			    V_INGRESSQID((u32)iq->cntxt_id) |
1598 			    V_SEINTARM(iq->intr_params));
1599 			return (0);
1600 		}
1601 		ndescs = 1;
1602 	}
1603 #else
1604 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1605 #endif
1606 
1607 	limit = budget ? budget : iq->qsize / 16;
1608 	fl = &rxq->fl;
1609 	fl_hw_cidx = fl->hw_cidx;	/* stable snapshot */
1610 	while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1611 
1612 		rmb();
1613 
1614 		m0 = NULL;
1615 		rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1616 		lq = be32toh(d->rsp.pldbuflen_qid);
1617 
1618 		switch (rsp_type) {
1619 		case X_RSPD_TYPE_FLBUF:
1620 			if (lq & F_RSPD_NEWBUF) {
1621 				if (fl->rx_offset > 0)
1622 					move_to_next_rxbuf(fl);
1623 				lq = G_RSPD_LEN(lq);
1624 			}
1625 			if (IDXDIFF(fl->hw_cidx, fl_hw_cidx, fl->sidx) > 4) {
1626 				FL_LOCK(fl);
1627 				refill_fl(sc, fl, 64);
1628 				FL_UNLOCK(fl);
1629 				fl_hw_cidx = fl->hw_cidx;
1630 			}
1631 
1632 			if (d->rss.opcode == CPL_RX_PKT) {
1633 				if (__predict_true(eth_rx(sc, rxq, d, lq) == 0))
1634 					break;
1635 				goto out;
1636 			}
1637 			m0 = get_fl_payload(sc, fl, lq);
1638 			if (__predict_false(m0 == NULL))
1639 				goto out;
1640 
1641 			/* fall through */
1642 
1643 		case X_RSPD_TYPE_CPL:
1644 			KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1645 			    ("%s: bad opcode %02x.", __func__, d->rss.opcode));
1646 			t4_cpl_handler[d->rss.opcode](iq, &d->rss, m0);
1647 			break;
1648 
1649 		case X_RSPD_TYPE_INTR:
1650 
1651 			/*
1652 			 * There are 1K interrupt-capable queues (qids 0
1653 			 * through 1023).  A response type indicating a
1654 			 * forwarded interrupt with a qid >= 1K is an
1655 			 * iWARP async notification.  That is the only
1656 			 * acceptable indirect interrupt on this queue.
1657 			 */
1658 			if (__predict_false(lq < 1024)) {
1659 				panic("%s: indirect interrupt on iq_fl %p "
1660 				    "with qid %u", __func__, iq, lq);
1661 			}
1662 
1663 			t4_an_handler(iq, &d->rsp);
1664 			break;
1665 
1666 		default:
1667 			KASSERT(0, ("%s: illegal response type %d on iq %p",
1668 			    __func__, rsp_type, iq));
1669 			log(LOG_ERR, "%s: illegal response type %d on iq %p",
1670 			    device_get_nameunit(sc->dev), rsp_type, iq);
1671 			break;
1672 		}
1673 
1674 		d++;
1675 		if (__predict_false(++iq->cidx == iq->sidx)) {
1676 			iq->cidx = 0;
1677 			iq->gen ^= F_RSPD_GEN;
1678 			d = &iq->desc[0];
1679 		}
1680 		if (__predict_false(++ndescs == limit)) {
1681 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1682 			    V_INGRESSQID(iq->cntxt_id) |
1683 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1684 
1685 #if defined(INET) || defined(INET6)
1686 			if (iq->flags & IQ_LRO_ENABLED &&
1687 			    !sort_before_lro(lro) &&
1688 			    sc->lro_timeout != 0) {
1689 				tcp_lro_flush_inactive(lro, &lro_timeout);
1690 			}
1691 #endif
1692 			if (budget)
1693 				return (EINPROGRESS);
1694 			ndescs = 0;
1695 		}
1696 	}
1697 out:
1698 #if defined(INET) || defined(INET6)
1699 	if (iq->flags & IQ_LRO_ENABLED) {
1700 		if (ndescs > 0 && lro->lro_mbuf_count > 8) {
1701 			MPASS(sort_before_lro(lro));
1702 			/* hold back one credit and don't flush LRO state */
1703 			iq->flags |= IQ_ADJ_CREDIT;
1704 			ndescs--;
1705 		} else {
1706 			tcp_lro_flush_all(lro);
1707 		}
1708 	}
1709 #endif
1710 
1711 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1712 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1713 
1714 	FL_LOCK(fl);
1715 	starved = refill_fl(sc, fl, 64);
1716 	FL_UNLOCK(fl);
1717 	if (__predict_false(starved != 0))
1718 		add_fl_to_sfl(sc, fl);
1719 
1720 	return (0);
1721 }
1722 
1723 static inline struct cluster_metadata *
1724 cl_metadata(struct fl_sdesc *sd)
1725 {
1726 
1727 	return ((void *)(sd->cl + sd->moff));
1728 }
1729 
1730 static void
1731 rxb_free(struct mbuf *m)
1732 {
1733 	struct cluster_metadata *clm = m->m_ext.ext_arg1;
1734 
1735 	uma_zfree(clm->zone, clm->cl);
1736 	counter_u64_add(extfree_rels, 1);
1737 }
1738 
1739 /*
1740  * The mbuf returned comes from zone_muf and carries the payload in one of these
1741  * ways
1742  * a) complete frame inside the mbuf
1743  * b) m_cljset (for clusters without metadata)
1744  * d) m_extaddref (cluster with metadata)
1745  */
1746 static struct mbuf *
1747 get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1748     int remaining)
1749 {
1750 	struct mbuf *m;
1751 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1752 	struct rx_buf_info *rxb = &sc->sge.rx_buf_info[sd->zidx];
1753 	struct cluster_metadata *clm;
1754 	int len, blen;
1755 	caddr_t payload;
1756 
1757 	if (fl->flags & FL_BUF_PACKING) {
1758 		u_int l, pad;
1759 
1760 		blen = rxb->size2 - fl->rx_offset;	/* max possible in this buf */
1761 		len = min(remaining, blen);
1762 		payload = sd->cl + fl->rx_offset;
1763 
1764 		l = fr_offset + len;
1765 		pad = roundup2(l, fl->buf_boundary) - l;
1766 		if (fl->rx_offset + len + pad < rxb->size2)
1767 			blen = len + pad;
1768 		MPASS(fl->rx_offset + blen <= rxb->size2);
1769 	} else {
1770 		MPASS(fl->rx_offset == 0);	/* not packing */
1771 		blen = rxb->size1;
1772 		len = min(remaining, blen);
1773 		payload = sd->cl;
1774 	}
1775 
1776 	if (fr_offset == 0) {
1777 		m = m_gethdr(M_NOWAIT, MT_DATA);
1778 		if (__predict_false(m == NULL))
1779 			return (NULL);
1780 		m->m_pkthdr.len = remaining;
1781 	} else {
1782 		m = m_get(M_NOWAIT, MT_DATA);
1783 		if (__predict_false(m == NULL))
1784 			return (NULL);
1785 	}
1786 	m->m_len = len;
1787 	kmsan_mark(payload, len, KMSAN_STATE_INITED);
1788 
1789 	if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) {
1790 		/* copy data to mbuf */
1791 		bcopy(payload, mtod(m, caddr_t), len);
1792 		if (fl->flags & FL_BUF_PACKING) {
1793 			fl->rx_offset += blen;
1794 			MPASS(fl->rx_offset <= rxb->size2);
1795 			if (fl->rx_offset < rxb->size2)
1796 				return (m);	/* without advancing the cidx */
1797 		}
1798 	} else if (fl->flags & FL_BUF_PACKING) {
1799 		clm = cl_metadata(sd);
1800 		if (sd->nmbuf++ == 0) {
1801 			clm->refcount = 1;
1802 			clm->zone = rxb->zone;
1803 			clm->cl = sd->cl;
1804 			counter_u64_add(extfree_refs, 1);
1805 		}
1806 		m_extaddref(m, payload, blen, &clm->refcount, rxb_free, clm,
1807 		    NULL);
1808 
1809 		fl->rx_offset += blen;
1810 		MPASS(fl->rx_offset <= rxb->size2);
1811 		if (fl->rx_offset < rxb->size2)
1812 			return (m);	/* without advancing the cidx */
1813 	} else {
1814 		m_cljset(m, sd->cl, rxb->type);
1815 		sd->cl = NULL;	/* consumed, not a recycle candidate */
1816 	}
1817 
1818 	move_to_next_rxbuf(fl);
1819 
1820 	return (m);
1821 }
1822 
1823 static struct mbuf *
1824 get_fl_payload(struct adapter *sc, struct sge_fl *fl, const u_int plen)
1825 {
1826 	struct mbuf *m0, *m, **pnext;
1827 	u_int remaining;
1828 
1829 	if (__predict_false(fl->flags & FL_BUF_RESUME)) {
1830 		M_ASSERTPKTHDR(fl->m0);
1831 		MPASS(fl->m0->m_pkthdr.len == plen);
1832 		MPASS(fl->remaining < plen);
1833 
1834 		m0 = fl->m0;
1835 		pnext = fl->pnext;
1836 		remaining = fl->remaining;
1837 		fl->flags &= ~FL_BUF_RESUME;
1838 		goto get_segment;
1839 	}
1840 
1841 	/*
1842 	 * Payload starts at rx_offset in the current hw buffer.  Its length is
1843 	 * 'len' and it may span multiple hw buffers.
1844 	 */
1845 
1846 	m0 = get_scatter_segment(sc, fl, 0, plen);
1847 	if (m0 == NULL)
1848 		return (NULL);
1849 	remaining = plen - m0->m_len;
1850 	pnext = &m0->m_next;
1851 	while (remaining > 0) {
1852 get_segment:
1853 		MPASS(fl->rx_offset == 0);
1854 		m = get_scatter_segment(sc, fl, plen - remaining, remaining);
1855 		if (__predict_false(m == NULL)) {
1856 			fl->m0 = m0;
1857 			fl->pnext = pnext;
1858 			fl->remaining = remaining;
1859 			fl->flags |= FL_BUF_RESUME;
1860 			return (NULL);
1861 		}
1862 		*pnext = m;
1863 		pnext = &m->m_next;
1864 		remaining -= m->m_len;
1865 	}
1866 	*pnext = NULL;
1867 
1868 	M_ASSERTPKTHDR(m0);
1869 	return (m0);
1870 }
1871 
1872 static int
1873 skip_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1874     int remaining)
1875 {
1876 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1877 	struct rx_buf_info *rxb = &sc->sge.rx_buf_info[sd->zidx];
1878 	int len, blen;
1879 
1880 	if (fl->flags & FL_BUF_PACKING) {
1881 		u_int l, pad;
1882 
1883 		blen = rxb->size2 - fl->rx_offset;	/* max possible in this buf */
1884 		len = min(remaining, blen);
1885 
1886 		l = fr_offset + len;
1887 		pad = roundup2(l, fl->buf_boundary) - l;
1888 		if (fl->rx_offset + len + pad < rxb->size2)
1889 			blen = len + pad;
1890 		fl->rx_offset += blen;
1891 		MPASS(fl->rx_offset <= rxb->size2);
1892 		if (fl->rx_offset < rxb->size2)
1893 			return (len);	/* without advancing the cidx */
1894 	} else {
1895 		MPASS(fl->rx_offset == 0);	/* not packing */
1896 		blen = rxb->size1;
1897 		len = min(remaining, blen);
1898 	}
1899 	move_to_next_rxbuf(fl);
1900 	return (len);
1901 }
1902 
1903 static inline void
1904 skip_fl_payload(struct adapter *sc, struct sge_fl *fl, int plen)
1905 {
1906 	int remaining, fr_offset, len;
1907 
1908 	fr_offset = 0;
1909 	remaining = plen;
1910 	while (remaining > 0) {
1911 		len = skip_scatter_segment(sc, fl, fr_offset, remaining);
1912 		fr_offset += len;
1913 		remaining -= len;
1914 	}
1915 }
1916 
1917 static inline int
1918 get_segment_len(struct adapter *sc, struct sge_fl *fl, int plen)
1919 {
1920 	int len;
1921 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1922 	struct rx_buf_info *rxb = &sc->sge.rx_buf_info[sd->zidx];
1923 
1924 	if (fl->flags & FL_BUF_PACKING)
1925 		len = rxb->size2 - fl->rx_offset;
1926 	else
1927 		len = rxb->size1;
1928 
1929 	return (min(plen, len));
1930 }
1931 
1932 static int
1933 eth_rx(struct adapter *sc, struct sge_rxq *rxq, const struct iq_desc *d,
1934     u_int plen)
1935 {
1936 	struct mbuf *m0;
1937 	if_t ifp = rxq->ifp;
1938 	struct sge_fl *fl = &rxq->fl;
1939 	struct vi_info *vi = if_getsoftc(ifp);
1940 	const struct cpl_rx_pkt *cpl;
1941 #if defined(INET) || defined(INET6)
1942 	struct lro_ctrl *lro = &rxq->lro;
1943 #endif
1944 	uint16_t err_vec, tnl_type, tnlhdr_len;
1945 	static const int sw_hashtype[4][2] = {
1946 		{M_HASHTYPE_NONE, M_HASHTYPE_NONE},
1947 		{M_HASHTYPE_RSS_IPV4, M_HASHTYPE_RSS_IPV6},
1948 		{M_HASHTYPE_RSS_TCP_IPV4, M_HASHTYPE_RSS_TCP_IPV6},
1949 		{M_HASHTYPE_RSS_UDP_IPV4, M_HASHTYPE_RSS_UDP_IPV6},
1950 	};
1951 	static const int sw_csum_flags[2][2] = {
1952 		{
1953 			/* IP, inner IP */
1954 			CSUM_ENCAP_VXLAN |
1955 			    CSUM_L3_CALC | CSUM_L3_VALID |
1956 			    CSUM_L4_CALC | CSUM_L4_VALID |
1957 			    CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID |
1958 			    CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID,
1959 
1960 			/* IP, inner IP6 */
1961 			CSUM_ENCAP_VXLAN |
1962 			    CSUM_L3_CALC | CSUM_L3_VALID |
1963 			    CSUM_L4_CALC | CSUM_L4_VALID |
1964 			    CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID,
1965 		},
1966 		{
1967 			/* IP6, inner IP */
1968 			CSUM_ENCAP_VXLAN |
1969 			    CSUM_L4_CALC | CSUM_L4_VALID |
1970 			    CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID |
1971 			    CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID,
1972 
1973 			/* IP6, inner IP6 */
1974 			CSUM_ENCAP_VXLAN |
1975 			    CSUM_L4_CALC | CSUM_L4_VALID |
1976 			    CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID,
1977 		},
1978 	};
1979 
1980 	MPASS(plen > sc->params.sge.fl_pktshift);
1981 	if (vi->pfil != NULL && PFIL_HOOKED_IN(vi->pfil) &&
1982 	    __predict_true((fl->flags & FL_BUF_RESUME) == 0)) {
1983 		struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1984 		caddr_t frame;
1985 		int rc, slen;
1986 
1987 		slen = get_segment_len(sc, fl, plen) -
1988 		    sc->params.sge.fl_pktshift;
1989 		frame = sd->cl + fl->rx_offset + sc->params.sge.fl_pktshift;
1990 		CURVNET_SET_QUIET(if_getvnet(ifp));
1991 		rc = pfil_mem_in(vi->pfil, frame, slen, ifp, &m0);
1992 		CURVNET_RESTORE();
1993 		if (rc == PFIL_DROPPED || rc == PFIL_CONSUMED) {
1994 			skip_fl_payload(sc, fl, plen);
1995 			return (0);
1996 		}
1997 		if (rc == PFIL_REALLOCED) {
1998 			skip_fl_payload(sc, fl, plen);
1999 			goto have_mbuf;
2000 		}
2001 	}
2002 
2003 	m0 = get_fl_payload(sc, fl, plen);
2004 	if (__predict_false(m0 == NULL))
2005 		return (ENOMEM);
2006 
2007 	m0->m_pkthdr.len -= sc->params.sge.fl_pktshift;
2008 	m0->m_len -= sc->params.sge.fl_pktshift;
2009 	m0->m_data += sc->params.sge.fl_pktshift;
2010 
2011 have_mbuf:
2012 	m0->m_pkthdr.rcvif = ifp;
2013 	M_HASHTYPE_SET(m0, sw_hashtype[d->rss.hash_type][d->rss.ipv6]);
2014 	m0->m_pkthdr.flowid = be32toh(d->rss.hash_val);
2015 
2016 	cpl = (const void *)(&d->rss + 1);
2017 	if (sc->params.tp.rx_pkt_encap) {
2018 		const uint16_t ev = be16toh(cpl->err_vec);
2019 
2020 		err_vec = G_T6_COMPR_RXERR_VEC(ev);
2021 		tnl_type = G_T6_RX_TNL_TYPE(ev);
2022 		tnlhdr_len = G_T6_RX_TNLHDR_LEN(ev);
2023 	} else {
2024 		err_vec = be16toh(cpl->err_vec);
2025 		tnl_type = 0;
2026 		tnlhdr_len = 0;
2027 	}
2028 	if (cpl->csum_calc && err_vec == 0) {
2029 		int ipv6 = !!(cpl->l2info & htobe32(F_RXF_IP6));
2030 
2031 		/* checksum(s) calculated and found to be correct. */
2032 
2033 		MPASS((cpl->l2info & htobe32(F_RXF_IP)) ^
2034 		    (cpl->l2info & htobe32(F_RXF_IP6)));
2035 		m0->m_pkthdr.csum_data = be16toh(cpl->csum);
2036 		if (tnl_type == 0) {
2037 			if (!ipv6 && if_getcapenable(ifp) & IFCAP_RXCSUM) {
2038 				m0->m_pkthdr.csum_flags = CSUM_L3_CALC |
2039 				    CSUM_L3_VALID | CSUM_L4_CALC |
2040 				    CSUM_L4_VALID;
2041 			} else if (ipv6 && if_getcapenable(ifp) & IFCAP_RXCSUM_IPV6) {
2042 				m0->m_pkthdr.csum_flags = CSUM_L4_CALC |
2043 				    CSUM_L4_VALID;
2044 			}
2045 			rxq->rxcsum++;
2046 		} else {
2047 			MPASS(tnl_type == RX_PKT_TNL_TYPE_VXLAN);
2048 
2049 			M_HASHTYPE_SETINNER(m0);
2050 			if (__predict_false(cpl->ip_frag)) {
2051 				/*
2052 				 * csum_data is for the inner frame (which is an
2053 				 * IP fragment) and is not 0xffff.  There is no
2054 				 * way to pass the inner csum_data to the stack.
2055 				 * We don't want the stack to use the inner
2056 				 * csum_data to validate the outer frame or it
2057 				 * will get rejected.  So we fix csum_data here
2058 				 * and let sw do the checksum of inner IP
2059 				 * fragments.
2060 				 *
2061 				 * XXX: Need 32b for csum_data2 in an rx mbuf.
2062 				 * Maybe stuff it into rcv_tstmp?
2063 				 */
2064 				m0->m_pkthdr.csum_data = 0xffff;
2065 				if (ipv6) {
2066 					m0->m_pkthdr.csum_flags = CSUM_L4_CALC |
2067 					    CSUM_L4_VALID;
2068 				} else {
2069 					m0->m_pkthdr.csum_flags = CSUM_L3_CALC |
2070 					    CSUM_L3_VALID | CSUM_L4_CALC |
2071 					    CSUM_L4_VALID;
2072 				}
2073 			} else {
2074 				int outer_ipv6;
2075 
2076 				MPASS(m0->m_pkthdr.csum_data == 0xffff);
2077 
2078 				outer_ipv6 = tnlhdr_len >=
2079 				    sizeof(struct ether_header) +
2080 				    sizeof(struct ip6_hdr);
2081 				m0->m_pkthdr.csum_flags =
2082 				    sw_csum_flags[outer_ipv6][ipv6];
2083 			}
2084 			rxq->vxlan_rxcsum++;
2085 		}
2086 	}
2087 
2088 	if (cpl->vlan_ex) {
2089 		m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
2090 		m0->m_flags |= M_VLANTAG;
2091 		rxq->vlan_extraction++;
2092 	}
2093 
2094 	if (rxq->iq.flags & IQ_RX_TIMESTAMP) {
2095 		/*
2096 		 * Fill up rcv_tstmp but do not set M_TSTMP as
2097 		 * long as we get a non-zero back from t4_tstmp_to_ns().
2098 		 */
2099 		m0->m_pkthdr.rcv_tstmp = t4_tstmp_to_ns(sc,
2100 		    be64toh(d->rsp.u.last_flit));
2101 		if (m0->m_pkthdr.rcv_tstmp != 0)
2102 			m0->m_flags |= M_TSTMP;
2103 	}
2104 
2105 #ifdef NUMA
2106 	m0->m_pkthdr.numa_domain = if_getnumadomain(ifp);
2107 #endif
2108 #if defined(INET) || defined(INET6)
2109 	if (rxq->iq.flags & IQ_LRO_ENABLED && tnl_type == 0 &&
2110 	    (M_HASHTYPE_GET(m0) == M_HASHTYPE_RSS_TCP_IPV4 ||
2111 	    M_HASHTYPE_GET(m0) == M_HASHTYPE_RSS_TCP_IPV6)) {
2112 		if (sort_before_lro(lro)) {
2113 			tcp_lro_queue_mbuf(lro, m0);
2114 			return (0); /* queued for sort, then LRO */
2115 		}
2116 		if (tcp_lro_rx(lro, m0, 0) == 0)
2117 			return (0); /* queued for LRO */
2118 	}
2119 #endif
2120 	if_input(ifp, m0);
2121 
2122 	return (0);
2123 }
2124 
2125 /*
2126  * Must drain the wrq or make sure that someone else will.
2127  */
2128 static void
2129 wrq_tx_drain(void *arg, int n)
2130 {
2131 	struct sge_wrq *wrq = arg;
2132 	struct sge_eq *eq = &wrq->eq;
2133 
2134 	EQ_LOCK(eq);
2135 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2136 		drain_wrq_wr_list(wrq->adapter, wrq);
2137 	EQ_UNLOCK(eq);
2138 }
2139 
2140 static void
2141 drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq)
2142 {
2143 	struct sge_eq *eq = &wrq->eq;
2144 	u_int available, dbdiff;	/* # of hardware descriptors */
2145 	u_int n;
2146 	struct wrqe *wr;
2147 	struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
2148 
2149 	EQ_LOCK_ASSERT_OWNED(eq);
2150 	MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
2151 	wr = STAILQ_FIRST(&wrq->wr_list);
2152 	MPASS(wr != NULL);	/* Must be called with something useful to do */
2153 	MPASS(eq->pidx == eq->dbidx);
2154 	dbdiff = 0;
2155 
2156 	do {
2157 		eq->cidx = read_hw_cidx(eq);
2158 		if (eq->pidx == eq->cidx)
2159 			available = eq->sidx - 1;
2160 		else
2161 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2162 
2163 		MPASS(wr->wrq == wrq);
2164 		n = howmany(wr->wr_len, EQ_ESIZE);
2165 		if (available < n)
2166 			break;
2167 
2168 		dst = (void *)&eq->desc[eq->pidx];
2169 		if (__predict_true(eq->sidx - eq->pidx > n)) {
2170 			/* Won't wrap, won't end exactly at the status page. */
2171 			bcopy(&wr->wr[0], dst, wr->wr_len);
2172 			eq->pidx += n;
2173 		} else {
2174 			int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE;
2175 
2176 			bcopy(&wr->wr[0], dst, first_portion);
2177 			if (wr->wr_len > first_portion) {
2178 				bcopy(&wr->wr[first_portion], &eq->desc[0],
2179 				    wr->wr_len - first_portion);
2180 			}
2181 			eq->pidx = n - (eq->sidx - eq->pidx);
2182 		}
2183 		wrq->tx_wrs_copied++;
2184 
2185 		if (available < eq->sidx / 4 &&
2186 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2187 				/*
2188 				 * XXX: This is not 100% reliable with some
2189 				 * types of WRs.  But this is a very unusual
2190 				 * situation for an ofld/ctrl queue anyway.
2191 				 */
2192 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2193 			    F_FW_WR_EQUEQ);
2194 		}
2195 
2196 		dbdiff += n;
2197 		if (dbdiff >= 16) {
2198 			ring_eq_db(sc, eq, dbdiff);
2199 			dbdiff = 0;
2200 		}
2201 
2202 		STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
2203 		free_wrqe(wr);
2204 		MPASS(wrq->nwr_pending > 0);
2205 		wrq->nwr_pending--;
2206 		MPASS(wrq->ndesc_needed >= n);
2207 		wrq->ndesc_needed -= n;
2208 	} while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL);
2209 
2210 	if (dbdiff)
2211 		ring_eq_db(sc, eq, dbdiff);
2212 }
2213 
2214 /*
2215  * Doesn't fail.  Holds on to work requests it can't send right away.
2216  */
2217 void
2218 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
2219 {
2220 #ifdef INVARIANTS
2221 	struct sge_eq *eq = &wrq->eq;
2222 #endif
2223 
2224 	EQ_LOCK_ASSERT_OWNED(eq);
2225 	MPASS(wr != NULL);
2226 	MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN);
2227 	MPASS((wr->wr_len & 0x7) == 0);
2228 
2229 	STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
2230 	wrq->nwr_pending++;
2231 	wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE);
2232 
2233 	if (!TAILQ_EMPTY(&wrq->incomplete_wrs))
2234 		return;	/* commit_wrq_wr will drain wr_list as well. */
2235 
2236 	drain_wrq_wr_list(sc, wrq);
2237 
2238 	/* Doorbell must have caught up to the pidx. */
2239 	MPASS(eq->pidx == eq->dbidx);
2240 }
2241 
2242 void
2243 t4_update_fl_bufsize(if_t ifp)
2244 {
2245 	struct vi_info *vi = if_getsoftc(ifp);
2246 	struct adapter *sc = vi->adapter;
2247 	struct sge_rxq *rxq;
2248 #ifdef TCP_OFFLOAD
2249 	struct sge_ofld_rxq *ofld_rxq;
2250 #endif
2251 	struct sge_fl *fl;
2252 	int i, maxp;
2253 
2254 	maxp = max_rx_payload(sc, ifp, false);
2255 	for_each_rxq(vi, i, rxq) {
2256 		fl = &rxq->fl;
2257 
2258 		FL_LOCK(fl);
2259 		fl->zidx = find_refill_source(sc, maxp,
2260 		    fl->flags & FL_BUF_PACKING);
2261 		FL_UNLOCK(fl);
2262 	}
2263 #ifdef TCP_OFFLOAD
2264 	maxp = max_rx_payload(sc, ifp, true);
2265 	for_each_ofld_rxq(vi, i, ofld_rxq) {
2266 		fl = &ofld_rxq->fl;
2267 
2268 		FL_LOCK(fl);
2269 		fl->zidx = find_refill_source(sc, maxp,
2270 		    fl->flags & FL_BUF_PACKING);
2271 		FL_UNLOCK(fl);
2272 	}
2273 #endif
2274 }
2275 
2276 #ifdef RATELIMIT
2277 static inline int
2278 mbuf_eo_nsegs(struct mbuf *m)
2279 {
2280 
2281 	M_ASSERTPKTHDR(m);
2282 	return (m->m_pkthdr.PH_loc.eight[1]);
2283 }
2284 
2285 #if defined(INET) || defined(INET6)
2286 static inline void
2287 set_mbuf_eo_nsegs(struct mbuf *m, uint8_t nsegs)
2288 {
2289 
2290 	M_ASSERTPKTHDR(m);
2291 	m->m_pkthdr.PH_loc.eight[1] = nsegs;
2292 }
2293 #endif
2294 
2295 static inline int
2296 mbuf_eo_len16(struct mbuf *m)
2297 {
2298 	int n;
2299 
2300 	M_ASSERTPKTHDR(m);
2301 	n = m->m_pkthdr.PH_loc.eight[2];
2302 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2303 
2304 	return (n);
2305 }
2306 
2307 #if defined(INET) || defined(INET6)
2308 static inline void
2309 set_mbuf_eo_len16(struct mbuf *m, uint8_t len16)
2310 {
2311 
2312 	M_ASSERTPKTHDR(m);
2313 	m->m_pkthdr.PH_loc.eight[2] = len16;
2314 }
2315 #endif
2316 
2317 static inline int
2318 mbuf_eo_tsclk_tsoff(struct mbuf *m)
2319 {
2320 
2321 	M_ASSERTPKTHDR(m);
2322 	return (m->m_pkthdr.PH_loc.eight[3]);
2323 }
2324 
2325 #if defined(INET) || defined(INET6)
2326 static inline void
2327 set_mbuf_eo_tsclk_tsoff(struct mbuf *m, uint8_t tsclk_tsoff)
2328 {
2329 
2330 	M_ASSERTPKTHDR(m);
2331 	m->m_pkthdr.PH_loc.eight[3] = tsclk_tsoff;
2332 }
2333 #endif
2334 
2335 static inline int
2336 needs_eo(struct m_snd_tag *mst)
2337 {
2338 
2339 	return (mst != NULL && mst->sw->type == IF_SND_TAG_TYPE_RATE_LIMIT);
2340 }
2341 #endif
2342 
2343 /*
2344  * Try to allocate an mbuf to contain a raw work request.  To make it
2345  * easy to construct the work request, don't allocate a chain but a
2346  * single mbuf.
2347  */
2348 struct mbuf *
2349 alloc_wr_mbuf(int len, int how)
2350 {
2351 	struct mbuf *m;
2352 
2353 	if (len <= MHLEN)
2354 		m = m_gethdr(how, MT_DATA);
2355 	else if (len <= MCLBYTES)
2356 		m = m_getcl(how, MT_DATA, M_PKTHDR);
2357 	else
2358 		m = NULL;
2359 	if (m == NULL)
2360 		return (NULL);
2361 	m->m_pkthdr.len = len;
2362 	m->m_len = len;
2363 	set_mbuf_cflags(m, MC_RAW_WR);
2364 	set_mbuf_len16(m, howmany(len, 16));
2365 	return (m);
2366 }
2367 
2368 static inline bool
2369 needs_hwcsum(struct mbuf *m)
2370 {
2371 	const uint32_t csum_flags = CSUM_IP | CSUM_IP_UDP | CSUM_IP_TCP |
2372 	    CSUM_IP_TSO | CSUM_INNER_IP | CSUM_INNER_IP_UDP |
2373 	    CSUM_INNER_IP_TCP | CSUM_INNER_IP_TSO | CSUM_IP6_UDP |
2374 	    CSUM_IP6_TCP | CSUM_IP6_TSO | CSUM_INNER_IP6_UDP |
2375 	    CSUM_INNER_IP6_TCP | CSUM_INNER_IP6_TSO;
2376 
2377 	M_ASSERTPKTHDR(m);
2378 
2379 	return (m->m_pkthdr.csum_flags & csum_flags);
2380 }
2381 
2382 static inline bool
2383 needs_tso(struct mbuf *m)
2384 {
2385 	const uint32_t csum_flags = CSUM_IP_TSO | CSUM_IP6_TSO |
2386 	    CSUM_INNER_IP_TSO | CSUM_INNER_IP6_TSO;
2387 
2388 	M_ASSERTPKTHDR(m);
2389 
2390 	return (m->m_pkthdr.csum_flags & csum_flags);
2391 }
2392 
2393 static inline bool
2394 needs_vxlan_csum(struct mbuf *m)
2395 {
2396 
2397 	M_ASSERTPKTHDR(m);
2398 
2399 	return (m->m_pkthdr.csum_flags & CSUM_ENCAP_VXLAN);
2400 }
2401 
2402 static inline bool
2403 needs_vxlan_tso(struct mbuf *m)
2404 {
2405 	const uint32_t csum_flags = CSUM_ENCAP_VXLAN | CSUM_INNER_IP_TSO |
2406 	    CSUM_INNER_IP6_TSO;
2407 
2408 	M_ASSERTPKTHDR(m);
2409 
2410 	return ((m->m_pkthdr.csum_flags & csum_flags) != 0 &&
2411 	    (m->m_pkthdr.csum_flags & csum_flags) != CSUM_ENCAP_VXLAN);
2412 }
2413 
2414 #if defined(INET) || defined(INET6)
2415 static inline bool
2416 needs_inner_tcp_csum(struct mbuf *m)
2417 {
2418 	const uint32_t csum_flags = CSUM_INNER_IP_TSO | CSUM_INNER_IP6_TSO;
2419 
2420 	M_ASSERTPKTHDR(m);
2421 
2422 	return (m->m_pkthdr.csum_flags & csum_flags);
2423 }
2424 #endif
2425 
2426 static inline bool
2427 needs_l3_csum(struct mbuf *m)
2428 {
2429 	const uint32_t csum_flags = CSUM_IP | CSUM_IP_TSO | CSUM_INNER_IP |
2430 	    CSUM_INNER_IP_TSO;
2431 
2432 	M_ASSERTPKTHDR(m);
2433 
2434 	return (m->m_pkthdr.csum_flags & csum_flags);
2435 }
2436 
2437 static inline bool
2438 needs_outer_tcp_csum(struct mbuf *m)
2439 {
2440 	const uint32_t csum_flags = CSUM_IP_TCP | CSUM_IP_TSO | CSUM_IP6_TCP |
2441 	    CSUM_IP6_TSO;
2442 
2443 	M_ASSERTPKTHDR(m);
2444 
2445 	return (m->m_pkthdr.csum_flags & csum_flags);
2446 }
2447 
2448 #ifdef RATELIMIT
2449 static inline bool
2450 needs_outer_l4_csum(struct mbuf *m)
2451 {
2452 	const uint32_t csum_flags = CSUM_IP_UDP | CSUM_IP_TCP | CSUM_IP_TSO |
2453 	    CSUM_IP6_UDP | CSUM_IP6_TCP | CSUM_IP6_TSO;
2454 
2455 	M_ASSERTPKTHDR(m);
2456 
2457 	return (m->m_pkthdr.csum_flags & csum_flags);
2458 }
2459 
2460 static inline bool
2461 needs_outer_udp_csum(struct mbuf *m)
2462 {
2463 	const uint32_t csum_flags = CSUM_IP_UDP | CSUM_IP6_UDP;
2464 
2465 	M_ASSERTPKTHDR(m);
2466 
2467 	return (m->m_pkthdr.csum_flags & csum_flags);
2468 }
2469 #endif
2470 
2471 static inline bool
2472 needs_vlan_insertion(struct mbuf *m)
2473 {
2474 
2475 	M_ASSERTPKTHDR(m);
2476 
2477 	return (m->m_flags & M_VLANTAG);
2478 }
2479 
2480 #if defined(INET) || defined(INET6)
2481 static void *
2482 m_advance(struct mbuf **pm, int *poffset, int len)
2483 {
2484 	struct mbuf *m = *pm;
2485 	int offset = *poffset;
2486 	uintptr_t p = 0;
2487 
2488 	MPASS(len > 0);
2489 
2490 	for (;;) {
2491 		if (offset + len < m->m_len) {
2492 			offset += len;
2493 			p = mtod(m, uintptr_t) + offset;
2494 			break;
2495 		}
2496 		len -= m->m_len - offset;
2497 		m = m->m_next;
2498 		offset = 0;
2499 		MPASS(m != NULL);
2500 	}
2501 	*poffset = offset;
2502 	*pm = m;
2503 	return ((void *)p);
2504 }
2505 #endif
2506 
2507 static inline int
2508 count_mbuf_ext_pgs(struct mbuf *m, int skip, vm_paddr_t *nextaddr)
2509 {
2510 	vm_paddr_t paddr;
2511 	int i, len, off, pglen, pgoff, seglen, segoff;
2512 	int nsegs = 0;
2513 
2514 	M_ASSERTEXTPG(m);
2515 	off = mtod(m, vm_offset_t);
2516 	len = m->m_len;
2517 	off += skip;
2518 	len -= skip;
2519 
2520 	if (m->m_epg_hdrlen != 0) {
2521 		if (off >= m->m_epg_hdrlen) {
2522 			off -= m->m_epg_hdrlen;
2523 		} else {
2524 			seglen = m->m_epg_hdrlen - off;
2525 			segoff = off;
2526 			seglen = min(seglen, len);
2527 			off = 0;
2528 			len -= seglen;
2529 			paddr = pmap_kextract(
2530 			    (vm_offset_t)&m->m_epg_hdr[segoff]);
2531 			if (*nextaddr != paddr)
2532 				nsegs++;
2533 			*nextaddr = paddr + seglen;
2534 		}
2535 	}
2536 	pgoff = m->m_epg_1st_off;
2537 	for (i = 0; i < m->m_epg_npgs && len > 0; i++) {
2538 		pglen = m_epg_pagelen(m, i, pgoff);
2539 		if (off >= pglen) {
2540 			off -= pglen;
2541 			pgoff = 0;
2542 			continue;
2543 		}
2544 		seglen = pglen - off;
2545 		segoff = pgoff + off;
2546 		off = 0;
2547 		seglen = min(seglen, len);
2548 		len -= seglen;
2549 		paddr = m->m_epg_pa[i] + segoff;
2550 		if (*nextaddr != paddr)
2551 			nsegs++;
2552 		*nextaddr = paddr + seglen;
2553 		pgoff = 0;
2554 	};
2555 	if (len != 0) {
2556 		seglen = min(len, m->m_epg_trllen - off);
2557 		len -= seglen;
2558 		paddr = pmap_kextract((vm_offset_t)&m->m_epg_trail[off]);
2559 		if (*nextaddr != paddr)
2560 			nsegs++;
2561 		*nextaddr = paddr + seglen;
2562 	}
2563 
2564 	return (nsegs);
2565 }
2566 
2567 
2568 /*
2569  * Can deal with empty mbufs in the chain that have m_len = 0, but the chain
2570  * must have at least one mbuf that's not empty.  It is possible for this
2571  * routine to return 0 if skip accounts for all the contents of the mbuf chain.
2572  */
2573 static inline int
2574 count_mbuf_nsegs(struct mbuf *m, int skip, uint8_t *cflags)
2575 {
2576 	vm_paddr_t nextaddr, paddr;
2577 	vm_offset_t va;
2578 	int len, nsegs;
2579 
2580 	M_ASSERTPKTHDR(m);
2581 	MPASS(m->m_pkthdr.len > 0);
2582 	MPASS(m->m_pkthdr.len >= skip);
2583 
2584 	nsegs = 0;
2585 	nextaddr = 0;
2586 	for (; m; m = m->m_next) {
2587 		len = m->m_len;
2588 		if (__predict_false(len == 0))
2589 			continue;
2590 		if (skip >= len) {
2591 			skip -= len;
2592 			continue;
2593 		}
2594 		if ((m->m_flags & M_EXTPG) != 0) {
2595 			*cflags |= MC_NOMAP;
2596 			nsegs += count_mbuf_ext_pgs(m, skip, &nextaddr);
2597 			skip = 0;
2598 			continue;
2599 		}
2600 		va = mtod(m, vm_offset_t) + skip;
2601 		len -= skip;
2602 		skip = 0;
2603 		paddr = pmap_kextract(va);
2604 		nsegs += sglist_count((void *)(uintptr_t)va, len);
2605 		if (paddr == nextaddr)
2606 			nsegs--;
2607 		nextaddr = pmap_kextract(va + len - 1) + 1;
2608 	}
2609 
2610 	return (nsegs);
2611 }
2612 
2613 /*
2614  * The maximum number of segments that can fit in a WR.
2615  */
2616 static int
2617 max_nsegs_allowed(struct mbuf *m, bool vm_wr)
2618 {
2619 
2620 	if (vm_wr) {
2621 		if (needs_tso(m))
2622 			return (TX_SGL_SEGS_VM_TSO);
2623 		return (TX_SGL_SEGS_VM);
2624 	}
2625 
2626 	if (needs_tso(m)) {
2627 		if (needs_vxlan_tso(m))
2628 			return (TX_SGL_SEGS_VXLAN_TSO);
2629 		else
2630 			return (TX_SGL_SEGS_TSO);
2631 	}
2632 
2633 	return (TX_SGL_SEGS);
2634 }
2635 
2636 static struct timeval txerr_ratecheck = {0};
2637 static const struct timeval txerr_interval = {3, 0};
2638 
2639 /*
2640  * Analyze the mbuf to determine its tx needs.  The mbuf passed in may change:
2641  * a) caller can assume it's been freed if this function returns with an error.
2642  * b) it may get defragged up if the gather list is too long for the hardware.
2643  */
2644 int
2645 parse_pkt(struct mbuf **mp, bool vm_wr)
2646 {
2647 	struct mbuf *m0 = *mp, *m;
2648 	int rc, nsegs, defragged = 0;
2649 	struct ether_header *eh;
2650 #ifdef INET
2651 	void *l3hdr;
2652 #endif
2653 #if defined(INET) || defined(INET6)
2654 	int offset;
2655 	struct tcphdr *tcp;
2656 #endif
2657 #if defined(KERN_TLS) || defined(RATELIMIT)
2658 	struct m_snd_tag *mst;
2659 #endif
2660 	uint16_t eh_type;
2661 	uint8_t cflags;
2662 
2663 	cflags = 0;
2664 	M_ASSERTPKTHDR(m0);
2665 	if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) {
2666 		rc = EINVAL;
2667 fail:
2668 		m_freem(m0);
2669 		*mp = NULL;
2670 		return (rc);
2671 	}
2672 restart:
2673 	/*
2674 	 * First count the number of gather list segments in the payload.
2675 	 * Defrag the mbuf if nsegs exceeds the hardware limit.
2676 	 */
2677 	M_ASSERTPKTHDR(m0);
2678 	MPASS(m0->m_pkthdr.len > 0);
2679 	nsegs = count_mbuf_nsegs(m0, 0, &cflags);
2680 #if defined(KERN_TLS) || defined(RATELIMIT)
2681 	if (m0->m_pkthdr.csum_flags & CSUM_SND_TAG)
2682 		mst = m0->m_pkthdr.snd_tag;
2683 	else
2684 		mst = NULL;
2685 #endif
2686 #ifdef KERN_TLS
2687 	if (mst != NULL && mst->sw->type == IF_SND_TAG_TYPE_TLS) {
2688 		cflags |= MC_TLS;
2689 		set_mbuf_cflags(m0, cflags);
2690 		rc = t6_ktls_parse_pkt(m0);
2691 		if (rc != 0)
2692 			goto fail;
2693 		return (EINPROGRESS);
2694 	}
2695 #endif
2696 	if (nsegs > max_nsegs_allowed(m0, vm_wr)) {
2697 		if (defragged++ > 0) {
2698 			rc = EFBIG;
2699 			goto fail;
2700 		}
2701 		counter_u64_add(defrags, 1);
2702 		if ((m = m_defrag(m0, M_NOWAIT)) == NULL) {
2703 			rc = ENOMEM;
2704 			goto fail;
2705 		}
2706 		*mp = m0 = m;	/* update caller's copy after defrag */
2707 		goto restart;
2708 	}
2709 
2710 	if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN &&
2711 	    !(cflags & MC_NOMAP))) {
2712 		counter_u64_add(pullups, 1);
2713 		m0 = m_pullup(m0, m0->m_pkthdr.len);
2714 		if (m0 == NULL) {
2715 			/* Should have left well enough alone. */
2716 			rc = EFBIG;
2717 			goto fail;
2718 		}
2719 		*mp = m0;	/* update caller's copy after pullup */
2720 		goto restart;
2721 	}
2722 	set_mbuf_nsegs(m0, nsegs);
2723 	set_mbuf_cflags(m0, cflags);
2724 	calculate_mbuf_len16(m0, vm_wr);
2725 
2726 #ifdef RATELIMIT
2727 	/*
2728 	 * Ethofld is limited to TCP and UDP for now, and only when L4 hw
2729 	 * checksumming is enabled.  needs_outer_l4_csum happens to check for
2730 	 * all the right things.
2731 	 */
2732 	if (__predict_false(needs_eo(mst) && !needs_outer_l4_csum(m0))) {
2733 		m_snd_tag_rele(m0->m_pkthdr.snd_tag);
2734 		m0->m_pkthdr.snd_tag = NULL;
2735 		m0->m_pkthdr.csum_flags &= ~CSUM_SND_TAG;
2736 		mst = NULL;
2737 	}
2738 #endif
2739 
2740 	if (!needs_hwcsum(m0)
2741 #ifdef RATELIMIT
2742 		 && !needs_eo(mst)
2743 #endif
2744 	)
2745 		return (0);
2746 
2747 	m = m0;
2748 	eh = mtod(m, struct ether_header *);
2749 	eh_type = ntohs(eh->ether_type);
2750 	if (eh_type == ETHERTYPE_VLAN) {
2751 		struct ether_vlan_header *evh = (void *)eh;
2752 
2753 		eh_type = ntohs(evh->evl_proto);
2754 		m0->m_pkthdr.l2hlen = sizeof(*evh);
2755 	} else
2756 		m0->m_pkthdr.l2hlen = sizeof(*eh);
2757 
2758 #if defined(INET) || defined(INET6)
2759 	offset = 0;
2760 #ifdef INET
2761 	l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2762 #else
2763 	m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2764 #endif
2765 #endif
2766 
2767 	switch (eh_type) {
2768 #ifdef INET6
2769 	case ETHERTYPE_IPV6:
2770 		m0->m_pkthdr.l3hlen = sizeof(struct ip6_hdr);
2771 		break;
2772 #endif
2773 #ifdef INET
2774 	case ETHERTYPE_IP:
2775 	{
2776 		struct ip *ip = l3hdr;
2777 
2778 		if (needs_vxlan_csum(m0)) {
2779 			/* Driver will do the outer IP hdr checksum. */
2780 			ip->ip_sum = 0;
2781 			if (needs_vxlan_tso(m0)) {
2782 				const uint16_t ipl = ip->ip_len;
2783 
2784 				ip->ip_len = 0;
2785 				ip->ip_sum = ~in_cksum_hdr(ip);
2786 				ip->ip_len = ipl;
2787 			} else
2788 				ip->ip_sum = in_cksum_hdr(ip);
2789 		}
2790 		m0->m_pkthdr.l3hlen = ip->ip_hl << 2;
2791 		break;
2792 	}
2793 #endif
2794 	default:
2795 		if (ratecheck(&txerr_ratecheck, &txerr_interval)) {
2796 			log(LOG_ERR, "%s: ethertype 0x%04x unknown.  "
2797 			    "if_cxgbe must be compiled with the same "
2798 			    "INET/INET6 options as the kernel.\n", __func__,
2799 			    eh_type);
2800 		}
2801 		rc = EINVAL;
2802 		goto fail;
2803 	}
2804 
2805 #if defined(INET) || defined(INET6)
2806 	if (needs_vxlan_csum(m0)) {
2807 		m0->m_pkthdr.l4hlen = sizeof(struct udphdr);
2808 		m0->m_pkthdr.l5hlen = sizeof(struct vxlan_header);
2809 
2810 		/* Inner headers. */
2811 		eh = m_advance(&m, &offset, m0->m_pkthdr.l3hlen +
2812 		    sizeof(struct udphdr) + sizeof(struct vxlan_header));
2813 		eh_type = ntohs(eh->ether_type);
2814 		if (eh_type == ETHERTYPE_VLAN) {
2815 			struct ether_vlan_header *evh = (void *)eh;
2816 
2817 			eh_type = ntohs(evh->evl_proto);
2818 			m0->m_pkthdr.inner_l2hlen = sizeof(*evh);
2819 		} else
2820 			m0->m_pkthdr.inner_l2hlen = sizeof(*eh);
2821 #ifdef INET
2822 		l3hdr = m_advance(&m, &offset, m0->m_pkthdr.inner_l2hlen);
2823 #else
2824 		m_advance(&m, &offset, m0->m_pkthdr.inner_l2hlen);
2825 #endif
2826 
2827 		switch (eh_type) {
2828 #ifdef INET6
2829 		case ETHERTYPE_IPV6:
2830 			m0->m_pkthdr.inner_l3hlen = sizeof(struct ip6_hdr);
2831 			break;
2832 #endif
2833 #ifdef INET
2834 		case ETHERTYPE_IP:
2835 		{
2836 			struct ip *ip = l3hdr;
2837 
2838 			m0->m_pkthdr.inner_l3hlen = ip->ip_hl << 2;
2839 			break;
2840 		}
2841 #endif
2842 		default:
2843 			if (ratecheck(&txerr_ratecheck, &txerr_interval)) {
2844 				log(LOG_ERR, "%s: VXLAN hw offload requested"
2845 				    "with unknown ethertype 0x%04x.  if_cxgbe "
2846 				    "must be compiled with the same INET/INET6 "
2847 				    "options as the kernel.\n", __func__,
2848 				    eh_type);
2849 			}
2850 			rc = EINVAL;
2851 			goto fail;
2852 		}
2853 		if (needs_inner_tcp_csum(m0)) {
2854 			tcp = m_advance(&m, &offset, m0->m_pkthdr.inner_l3hlen);
2855 			m0->m_pkthdr.inner_l4hlen = tcp->th_off * 4;
2856 		}
2857 		MPASS((m0->m_pkthdr.csum_flags & CSUM_SND_TAG) == 0);
2858 		m0->m_pkthdr.csum_flags &= CSUM_INNER_IP6_UDP |
2859 		    CSUM_INNER_IP6_TCP | CSUM_INNER_IP6_TSO | CSUM_INNER_IP |
2860 		    CSUM_INNER_IP_UDP | CSUM_INNER_IP_TCP | CSUM_INNER_IP_TSO |
2861 		    CSUM_ENCAP_VXLAN;
2862 	}
2863 
2864 	if (needs_outer_tcp_csum(m0)) {
2865 		tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen);
2866 		m0->m_pkthdr.l4hlen = tcp->th_off * 4;
2867 #ifdef RATELIMIT
2868 		if (tsclk >= 0 && *(uint32_t *)(tcp + 1) == ntohl(0x0101080a)) {
2869 			set_mbuf_eo_tsclk_tsoff(m0,
2870 			    V_FW_ETH_TX_EO_WR_TSCLK(tsclk) |
2871 			    V_FW_ETH_TX_EO_WR_TSOFF(sizeof(*tcp) / 2 + 1));
2872 		} else
2873 			set_mbuf_eo_tsclk_tsoff(m0, 0);
2874 	} else if (needs_outer_udp_csum(m0)) {
2875 		m0->m_pkthdr.l4hlen = sizeof(struct udphdr);
2876 #endif
2877 	}
2878 #ifdef RATELIMIT
2879 	if (needs_eo(mst)) {
2880 		u_int immhdrs;
2881 
2882 		/* EO WRs have the headers in the WR and not the GL. */
2883 		immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen +
2884 		    m0->m_pkthdr.l4hlen;
2885 		cflags = 0;
2886 		nsegs = count_mbuf_nsegs(m0, immhdrs, &cflags);
2887 		MPASS(cflags == mbuf_cflags(m0));
2888 		set_mbuf_eo_nsegs(m0, nsegs);
2889 		set_mbuf_eo_len16(m0,
2890 		    txpkt_eo_len16(nsegs, immhdrs, needs_tso(m0)));
2891 		rc = ethofld_transmit(mst->ifp, m0);
2892 		if (rc != 0)
2893 			goto fail;
2894 		return (EINPROGRESS);
2895 	}
2896 #endif
2897 #endif
2898 	MPASS(m0 == *mp);
2899 	return (0);
2900 }
2901 
2902 void *
2903 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie)
2904 {
2905 	struct sge_eq *eq = &wrq->eq;
2906 	struct adapter *sc = wrq->adapter;
2907 	int ndesc, available;
2908 	struct wrqe *wr;
2909 	void *w;
2910 
2911 	MPASS(len16 > 0);
2912 	ndesc = tx_len16_to_desc(len16);
2913 	MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC);
2914 
2915 	EQ_LOCK(eq);
2916 
2917 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2918 		drain_wrq_wr_list(sc, wrq);
2919 
2920 	if (!STAILQ_EMPTY(&wrq->wr_list)) {
2921 slowpath:
2922 		EQ_UNLOCK(eq);
2923 		wr = alloc_wrqe(len16 * 16, wrq);
2924 		if (__predict_false(wr == NULL))
2925 			return (NULL);
2926 		cookie->pidx = -1;
2927 		cookie->ndesc = ndesc;
2928 		return (&wr->wr);
2929 	}
2930 
2931 	eq->cidx = read_hw_cidx(eq);
2932 	if (eq->pidx == eq->cidx)
2933 		available = eq->sidx - 1;
2934 	else
2935 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2936 	if (available < ndesc)
2937 		goto slowpath;
2938 
2939 	cookie->pidx = eq->pidx;
2940 	cookie->ndesc = ndesc;
2941 	TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link);
2942 
2943 	w = &eq->desc[eq->pidx];
2944 	IDXINCR(eq->pidx, ndesc, eq->sidx);
2945 	if (__predict_false(cookie->pidx + ndesc > eq->sidx)) {
2946 		w = &wrq->ss[0];
2947 		wrq->ss_pidx = cookie->pidx;
2948 		wrq->ss_len = len16 * 16;
2949 	}
2950 
2951 	EQ_UNLOCK(eq);
2952 
2953 	return (w);
2954 }
2955 
2956 void
2957 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie)
2958 {
2959 	struct sge_eq *eq = &wrq->eq;
2960 	struct adapter *sc = wrq->adapter;
2961 	int ndesc, pidx;
2962 	struct wrq_cookie *prev, *next;
2963 
2964 	if (cookie->pidx == -1) {
2965 		struct wrqe *wr = __containerof(w, struct wrqe, wr);
2966 
2967 		t4_wrq_tx(sc, wr);
2968 		return;
2969 	}
2970 
2971 	if (__predict_false(w == &wrq->ss[0])) {
2972 		int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE;
2973 
2974 		MPASS(wrq->ss_len > n);	/* WR had better wrap around. */
2975 		bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n);
2976 		bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n);
2977 		wrq->tx_wrs_ss++;
2978 	} else
2979 		wrq->tx_wrs_direct++;
2980 
2981 	EQ_LOCK(eq);
2982 	ndesc = cookie->ndesc;	/* Can be more than SGE_MAX_WR_NDESC here. */
2983 	pidx = cookie->pidx;
2984 	MPASS(pidx >= 0 && pidx < eq->sidx);
2985 	prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link);
2986 	next = TAILQ_NEXT(cookie, link);
2987 	if (prev == NULL) {
2988 		MPASS(pidx == eq->dbidx);
2989 		if (next == NULL || ndesc >= 16) {
2990 			int available;
2991 			struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
2992 
2993 			/*
2994 			 * Note that the WR via which we'll request tx updates
2995 			 * is at pidx and not eq->pidx, which has moved on
2996 			 * already.
2997 			 */
2998 			dst = (void *)&eq->desc[pidx];
2999 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
3000 			if (available < eq->sidx / 4 &&
3001 			    atomic_cmpset_int(&eq->equiq, 0, 1)) {
3002 				/*
3003 				 * XXX: This is not 100% reliable with some
3004 				 * types of WRs.  But this is a very unusual
3005 				 * situation for an ofld/ctrl queue anyway.
3006 				 */
3007 				dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
3008 				    F_FW_WR_EQUEQ);
3009 			}
3010 
3011 			ring_eq_db(wrq->adapter, eq, ndesc);
3012 		} else {
3013 			MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc);
3014 			next->pidx = pidx;
3015 			next->ndesc += ndesc;
3016 		}
3017 	} else {
3018 		MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc);
3019 		prev->ndesc += ndesc;
3020 	}
3021 	TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link);
3022 
3023 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
3024 		drain_wrq_wr_list(sc, wrq);
3025 
3026 #ifdef INVARIANTS
3027 	if (TAILQ_EMPTY(&wrq->incomplete_wrs)) {
3028 		/* Doorbell must have caught up to the pidx. */
3029 		MPASS(wrq->eq.pidx == wrq->eq.dbidx);
3030 	}
3031 #endif
3032 	EQ_UNLOCK(eq);
3033 }
3034 
3035 static u_int
3036 can_resume_eth_tx(struct mp_ring *r)
3037 {
3038 	struct sge_eq *eq = r->cookie;
3039 
3040 	return (total_available_tx_desc(eq) > eq->sidx / 8);
3041 }
3042 
3043 static inline bool
3044 cannot_use_txpkts(struct mbuf *m)
3045 {
3046 	/* maybe put a GL limit too, to avoid silliness? */
3047 
3048 	return (needs_tso(m) || (mbuf_cflags(m) & (MC_RAW_WR | MC_TLS)) != 0);
3049 }
3050 
3051 static inline int
3052 discard_tx(struct sge_eq *eq)
3053 {
3054 
3055 	return ((eq->flags & (EQ_ENABLED | EQ_QFLUSH)) != EQ_ENABLED);
3056 }
3057 
3058 static inline int
3059 wr_can_update_eq(void *p)
3060 {
3061 	struct fw_eth_tx_pkts_wr *wr = p;
3062 
3063 	switch (G_FW_WR_OP(be32toh(wr->op_pkd))) {
3064 	case FW_ULPTX_WR:
3065 	case FW_ETH_TX_PKT_WR:
3066 	case FW_ETH_TX_PKTS_WR:
3067 	case FW_ETH_TX_PKTS2_WR:
3068 	case FW_ETH_TX_PKT_VM_WR:
3069 	case FW_ETH_TX_PKTS_VM_WR:
3070 		return (1);
3071 	default:
3072 		return (0);
3073 	}
3074 }
3075 
3076 static inline void
3077 set_txupdate_flags(struct sge_txq *txq, u_int avail,
3078     struct fw_eth_tx_pkt_wr *wr)
3079 {
3080 	struct sge_eq *eq = &txq->eq;
3081 	struct txpkts *txp = &txq->txp;
3082 
3083 	if ((txp->npkt > 0 || avail < eq->sidx / 2) &&
3084 	    atomic_cmpset_int(&eq->equiq, 0, 1)) {
3085 		wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ | F_FW_WR_EQUIQ);
3086 		eq->equeqidx = eq->pidx;
3087 	} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
3088 		wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
3089 		eq->equeqidx = eq->pidx;
3090 	}
3091 }
3092 
3093 #if defined(__i386__) || defined(__amd64__)
3094 extern uint64_t tsc_freq;
3095 #endif
3096 
3097 static inline bool
3098 record_eth_tx_time(struct sge_txq *txq)
3099 {
3100 	const uint64_t cycles = get_cyclecount();
3101 	const uint64_t last_tx = txq->last_tx;
3102 #if defined(__i386__) || defined(__amd64__)
3103 	const uint64_t itg = tsc_freq * t4_tx_coalesce_gap / 1000000;
3104 #else
3105 	const uint64_t itg = 0;
3106 #endif
3107 
3108 	MPASS(cycles >= last_tx);
3109 	txq->last_tx = cycles;
3110 	return (cycles - last_tx < itg);
3111 }
3112 
3113 /*
3114  * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to
3115  * be consumed.  Return the actual number consumed.  0 indicates a stall.
3116  */
3117 static u_int
3118 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx, bool *coalescing)
3119 {
3120 	struct sge_txq *txq = r->cookie;
3121 	if_t ifp = txq->ifp;
3122 	struct sge_eq *eq = &txq->eq;
3123 	struct txpkts *txp = &txq->txp;
3124 	struct vi_info *vi = if_getsoftc(ifp);
3125 	struct adapter *sc = vi->adapter;
3126 	u_int total, remaining;		/* # of packets */
3127 	u_int n, avail, dbdiff;		/* # of hardware descriptors */
3128 	int i, rc;
3129 	struct mbuf *m0;
3130 	bool snd, recent_tx;
3131 	void *wr;	/* start of the last WR written to the ring */
3132 
3133 	TXQ_LOCK_ASSERT_OWNED(txq);
3134 	recent_tx = record_eth_tx_time(txq);
3135 
3136 	remaining = IDXDIFF(pidx, cidx, r->size);
3137 	if (__predict_false(discard_tx(eq))) {
3138 		for (i = 0; i < txp->npkt; i++)
3139 			m_freem(txp->mb[i]);
3140 		txp->npkt = 0;
3141 		while (cidx != pidx) {
3142 			m0 = r->items[cidx];
3143 			m_freem(m0);
3144 			if (++cidx == r->size)
3145 				cidx = 0;
3146 		}
3147 		reclaim_tx_descs(txq, eq->sidx);
3148 		*coalescing = false;
3149 		return (remaining);	/* emptied */
3150 	}
3151 
3152 	/* How many hardware descriptors do we have readily available. */
3153 	if (eq->pidx == eq->cidx)
3154 		avail = eq->sidx - 1;
3155 	else
3156 		avail = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
3157 
3158 	total = 0;
3159 	if (remaining == 0) {
3160 		txp->score = 0;
3161 		txq->txpkts_flush++;
3162 		goto send_txpkts;
3163 	}
3164 
3165 	dbdiff = 0;
3166 	MPASS(remaining > 0);
3167 	while (remaining > 0) {
3168 		m0 = r->items[cidx];
3169 		M_ASSERTPKTHDR(m0);
3170 		MPASS(m0->m_nextpkt == NULL);
3171 
3172 		if (avail < 2 * SGE_MAX_WR_NDESC)
3173 			avail += reclaim_tx_descs(txq, 64);
3174 
3175 		if (t4_tx_coalesce == 0 && txp->npkt == 0)
3176 			goto skip_coalescing;
3177 		if (cannot_use_txpkts(m0))
3178 			txp->score = 0;
3179 		else if (recent_tx) {
3180 			if (++txp->score == 0)
3181 				txp->score = UINT8_MAX;
3182 		} else
3183 			txp->score = 1;
3184 		if (txp->npkt > 0 || remaining > 1 ||
3185 		    txp->score >= t4_tx_coalesce_pkts ||
3186 		    atomic_load_int(&txq->eq.equiq) != 0) {
3187 			if (vi->flags & TX_USES_VM_WR)
3188 				rc = add_to_txpkts_vf(sc, txq, m0, avail, &snd);
3189 			else
3190 				rc = add_to_txpkts_pf(sc, txq, m0, avail, &snd);
3191 		} else {
3192 			snd = false;
3193 			rc = EINVAL;
3194 		}
3195 		if (snd) {
3196 			MPASS(txp->npkt > 0);
3197 			for (i = 0; i < txp->npkt; i++)
3198 				ETHER_BPF_MTAP(ifp, txp->mb[i]);
3199 			if (txp->npkt > 1) {
3200 				MPASS(avail >= tx_len16_to_desc(txp->len16));
3201 				if (vi->flags & TX_USES_VM_WR)
3202 					n = write_txpkts_vm_wr(sc, txq);
3203 				else
3204 					n = write_txpkts_wr(sc, txq);
3205 			} else {
3206 				MPASS(avail >=
3207 				    tx_len16_to_desc(mbuf_len16(txp->mb[0])));
3208 				if (vi->flags & TX_USES_VM_WR)
3209 					n = write_txpkt_vm_wr(sc, txq,
3210 					    txp->mb[0]);
3211 				else
3212 					n = write_txpkt_wr(sc, txq, txp->mb[0],
3213 					    avail);
3214 			}
3215 			MPASS(n <= SGE_MAX_WR_NDESC);
3216 			avail -= n;
3217 			dbdiff += n;
3218 			wr = &eq->desc[eq->pidx];
3219 			IDXINCR(eq->pidx, n, eq->sidx);
3220 			txp->npkt = 0;	/* emptied */
3221 		}
3222 		if (rc == 0) {
3223 			/* m0 was coalesced into txq->txpkts. */
3224 			goto next_mbuf;
3225 		}
3226 		if (rc == EAGAIN) {
3227 			/*
3228 			 * m0 is suitable for tx coalescing but could not be
3229 			 * combined with the existing txq->txpkts, which has now
3230 			 * been transmitted.  Start a new txpkts with m0.
3231 			 */
3232 			MPASS(snd);
3233 			MPASS(txp->npkt == 0);
3234 			continue;
3235 		}
3236 
3237 		MPASS(rc != 0 && rc != EAGAIN);
3238 		MPASS(txp->npkt == 0);
3239 skip_coalescing:
3240 		n = tx_len16_to_desc(mbuf_len16(m0));
3241 		if (__predict_false(avail < n)) {
3242 			avail += reclaim_tx_descs(txq, min(n, 32));
3243 			if (avail < n)
3244 				break;	/* out of descriptors */
3245 		}
3246 
3247 		wr = &eq->desc[eq->pidx];
3248 		if (mbuf_cflags(m0) & MC_RAW_WR) {
3249 			n = write_raw_wr(txq, wr, m0, avail);
3250 #ifdef KERN_TLS
3251 		} else if (mbuf_cflags(m0) & MC_TLS) {
3252 			ETHER_BPF_MTAP(ifp, m0);
3253 			n = t6_ktls_write_wr(txq, wr, m0, avail);
3254 #endif
3255 		} else {
3256 			ETHER_BPF_MTAP(ifp, m0);
3257 			if (vi->flags & TX_USES_VM_WR)
3258 				n = write_txpkt_vm_wr(sc, txq, m0);
3259 			else
3260 				n = write_txpkt_wr(sc, txq, m0, avail);
3261 		}
3262 		MPASS(n >= 1 && n <= avail);
3263 		if (!(mbuf_cflags(m0) & MC_TLS))
3264 			MPASS(n <= SGE_MAX_WR_NDESC);
3265 
3266 		avail -= n;
3267 		dbdiff += n;
3268 		IDXINCR(eq->pidx, n, eq->sidx);
3269 
3270 		if (dbdiff >= 512 / EQ_ESIZE) {	/* X_FETCHBURSTMAX_512B */
3271 			if (wr_can_update_eq(wr))
3272 				set_txupdate_flags(txq, avail, wr);
3273 			ring_eq_db(sc, eq, dbdiff);
3274 			avail += reclaim_tx_descs(txq, 32);
3275 			dbdiff = 0;
3276 		}
3277 next_mbuf:
3278 		total++;
3279 		remaining--;
3280 		if (__predict_false(++cidx == r->size))
3281 			cidx = 0;
3282 	}
3283 	if (dbdiff != 0) {
3284 		if (wr_can_update_eq(wr))
3285 			set_txupdate_flags(txq, avail, wr);
3286 		ring_eq_db(sc, eq, dbdiff);
3287 		reclaim_tx_descs(txq, 32);
3288 	} else if (eq->pidx == eq->cidx && txp->npkt > 0 &&
3289 	    atomic_load_int(&txq->eq.equiq) == 0) {
3290 		/*
3291 		 * If nothing was submitted to the chip for tx (it was coalesced
3292 		 * into txpkts instead) and there is no tx update outstanding
3293 		 * then we need to send txpkts now.
3294 		 */
3295 send_txpkts:
3296 		MPASS(txp->npkt > 0);
3297 		for (i = 0; i < txp->npkt; i++)
3298 			ETHER_BPF_MTAP(ifp, txp->mb[i]);
3299 		if (txp->npkt > 1) {
3300 			MPASS(avail >= tx_len16_to_desc(txp->len16));
3301 			if (vi->flags & TX_USES_VM_WR)
3302 				n = write_txpkts_vm_wr(sc, txq);
3303 			else
3304 				n = write_txpkts_wr(sc, txq);
3305 		} else {
3306 			MPASS(avail >=
3307 			    tx_len16_to_desc(mbuf_len16(txp->mb[0])));
3308 			if (vi->flags & TX_USES_VM_WR)
3309 				n = write_txpkt_vm_wr(sc, txq, txp->mb[0]);
3310 			else
3311 				n = write_txpkt_wr(sc, txq, txp->mb[0], avail);
3312 		}
3313 		MPASS(n <= SGE_MAX_WR_NDESC);
3314 		wr = &eq->desc[eq->pidx];
3315 		IDXINCR(eq->pidx, n, eq->sidx);
3316 		txp->npkt = 0;	/* emptied */
3317 
3318 		MPASS(wr_can_update_eq(wr));
3319 		set_txupdate_flags(txq, avail - n, wr);
3320 		ring_eq_db(sc, eq, n);
3321 		reclaim_tx_descs(txq, 32);
3322 	}
3323 	*coalescing = txp->npkt > 0;
3324 
3325 	return (total);
3326 }
3327 
3328 static inline void
3329 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
3330     int qsize, int intr_idx, int cong, int qtype)
3331 {
3332 
3333 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
3334 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
3335 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
3336 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
3337 	KASSERT(intr_idx >= -1 && intr_idx < sc->intr_count,
3338 	    ("%s: bad intr_idx %d", __func__, intr_idx));
3339 	KASSERT(qtype == FW_IQ_IQTYPE_OTHER || qtype == FW_IQ_IQTYPE_NIC ||
3340 	    qtype == FW_IQ_IQTYPE_OFLD, ("%s: bad qtype %d", __func__, qtype));
3341 
3342 	iq->flags = 0;
3343 	iq->state = IQS_DISABLED;
3344 	iq->adapter = sc;
3345 	iq->qtype = qtype;
3346 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
3347 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
3348 	if (pktc_idx >= 0) {
3349 		iq->intr_params |= F_QINTR_CNT_EN;
3350 		iq->intr_pktc_idx = pktc_idx;
3351 	}
3352 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
3353 	iq->sidx = iq->qsize - sc->params.sge.spg_len / IQ_ESIZE;
3354 	iq->intr_idx = intr_idx;
3355 	iq->cong_drop = cong;
3356 }
3357 
3358 static inline void
3359 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name)
3360 {
3361 	struct sge_params *sp = &sc->params.sge;
3362 
3363 	fl->qsize = qsize;
3364 	fl->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
3365 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
3366 	mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
3367 	if (sc->flags & BUF_PACKING_OK &&
3368 	    ((!is_t4(sc) && buffer_packing) ||	/* T5+: enabled unless 0 */
3369 	    (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */
3370 		fl->flags |= FL_BUF_PACKING;
3371 	fl->zidx = find_refill_source(sc, maxp, fl->flags & FL_BUF_PACKING);
3372 	fl->safe_zidx = sc->sge.safe_zidx;
3373 	if (fl->flags & FL_BUF_PACKING) {
3374 		fl->lowat = roundup2(sp->fl_starve_threshold2, 8);
3375 		fl->buf_boundary = sp->pack_boundary;
3376 	} else {
3377 		fl->lowat = roundup2(sp->fl_starve_threshold, 8);
3378 		fl->buf_boundary = 16;
3379 	}
3380 	if (fl_pad && fl->buf_boundary < sp->pad_boundary)
3381 		fl->buf_boundary = sp->pad_boundary;
3382 }
3383 
3384 static inline void
3385 init_eq(struct adapter *sc, struct sge_eq *eq, int eqtype, int qsize,
3386     uint8_t tx_chan, struct sge_iq *iq, char *name)
3387 {
3388 	KASSERT(eqtype >= EQ_CTRL && eqtype <= EQ_OFLD,
3389 	    ("%s: bad qtype %d", __func__, eqtype));
3390 
3391 	eq->type = eqtype;
3392 	eq->tx_chan = tx_chan;
3393 	eq->iq = iq;
3394 	eq->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
3395 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
3396 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
3397 }
3398 
3399 int
3400 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
3401     bus_dmamap_t *map, bus_addr_t *pa, void **va)
3402 {
3403 	int rc;
3404 
3405 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
3406 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
3407 	if (rc != 0) {
3408 		CH_ERR(sc, "cannot allocate DMA tag: %d\n", rc);
3409 		goto done;
3410 	}
3411 
3412 	rc = bus_dmamem_alloc(*tag, va,
3413 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
3414 	if (rc != 0) {
3415 		CH_ERR(sc, "cannot allocate DMA memory: %d\n", rc);
3416 		goto done;
3417 	}
3418 
3419 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
3420 	if (rc != 0) {
3421 		CH_ERR(sc, "cannot load DMA map: %d\n", rc);
3422 		goto done;
3423 	}
3424 done:
3425 	if (rc)
3426 		free_ring(sc, *tag, *map, *pa, *va);
3427 
3428 	return (rc);
3429 }
3430 
3431 int
3432 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
3433     bus_addr_t pa, void *va)
3434 {
3435 	if (pa)
3436 		bus_dmamap_unload(tag, map);
3437 	if (va)
3438 		bus_dmamem_free(tag, va, map);
3439 	if (tag)
3440 		bus_dma_tag_destroy(tag);
3441 
3442 	return (0);
3443 }
3444 
3445 /*
3446  * Allocates the software resources (mainly memory and sysctl nodes) for an
3447  * ingress queue and an optional freelist.
3448  *
3449  * Sets IQ_SW_ALLOCATED and returns 0 on success.
3450  */
3451 static int
3452 alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl,
3453     struct sysctl_ctx_list *ctx, struct sysctl_oid *oid)
3454 {
3455 	int rc;
3456 	size_t len;
3457 	struct adapter *sc = vi->adapter;
3458 
3459 	MPASS(!(iq->flags & IQ_SW_ALLOCATED));
3460 
3461 	len = iq->qsize * IQ_ESIZE;
3462 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
3463 	    (void **)&iq->desc);
3464 	if (rc != 0)
3465 		return (rc);
3466 
3467 	if (fl) {
3468 		len = fl->qsize * EQ_ESIZE;
3469 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
3470 		    &fl->ba, (void **)&fl->desc);
3471 		if (rc) {
3472 			free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba,
3473 			    iq->desc);
3474 			return (rc);
3475 		}
3476 
3477 		/* Allocate space for one software descriptor per buffer. */
3478 		fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc),
3479 		    M_CXGBE, M_ZERO | M_WAITOK);
3480 
3481 		add_fl_sysctls(sc, ctx, oid, fl);
3482 		iq->flags |= IQ_HAS_FL;
3483 	}
3484 	add_iq_sysctls(ctx, oid, iq);
3485 	iq->flags |= IQ_SW_ALLOCATED;
3486 
3487 	return (0);
3488 }
3489 
3490 /*
3491  * Frees all software resources (memory and locks) associated with an ingress
3492  * queue and an optional freelist.
3493  */
3494 static void
3495 free_iq_fl(struct adapter *sc, struct sge_iq *iq, struct sge_fl *fl)
3496 {
3497 	MPASS(iq->flags & IQ_SW_ALLOCATED);
3498 
3499 	if (fl) {
3500 		MPASS(iq->flags & IQ_HAS_FL);
3501 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba, fl->desc);
3502 		free_fl_buffers(sc, fl);
3503 		free(fl->sdesc, M_CXGBE);
3504 		mtx_destroy(&fl->fl_lock);
3505 		bzero(fl, sizeof(*fl));
3506 	}
3507 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
3508 	bzero(iq, sizeof(*iq));
3509 }
3510 
3511 /*
3512  * Allocates a hardware ingress queue and an optional freelist that will be
3513  * associated with it.
3514  *
3515  * Returns errno on failure.  Resources allocated up to that point may still be
3516  * allocated.  Caller is responsible for cleanup in case this function fails.
3517  */
3518 static int
3519 alloc_iq_fl_hwq(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl)
3520 {
3521 	int rc, cntxt_id, cong_map;
3522 	struct fw_iq_cmd c;
3523 	struct adapter *sc = vi->adapter;
3524 	struct port_info *pi = vi->pi;
3525 	__be32 v = 0;
3526 
3527 	MPASS (!(iq->flags & IQ_HW_ALLOCATED));
3528 
3529 	bzero(&c, sizeof(c));
3530 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
3531 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
3532 	    V_FW_IQ_CMD_VFN(0));
3533 
3534 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
3535 	    FW_LEN16(c));
3536 
3537 	/* Special handling for firmware event queue */
3538 	if (iq == &sc->sge.fwq)
3539 		v |= F_FW_IQ_CMD_IQASYNCH;
3540 
3541 	if (iq->intr_idx < 0) {
3542 		/* Forwarded interrupts, all headed to fwq */
3543 		v |= F_FW_IQ_CMD_IQANDST;
3544 		v |= V_FW_IQ_CMD_IQANDSTINDEX(sc->sge.fwq.cntxt_id);
3545 	} else {
3546 		KASSERT(iq->intr_idx < sc->intr_count,
3547 		    ("%s: invalid direct intr_idx %d", __func__, iq->intr_idx));
3548 		v |= V_FW_IQ_CMD_IQANDSTINDEX(iq->intr_idx);
3549 	}
3550 
3551 	bzero(iq->desc, iq->qsize * IQ_ESIZE);
3552 	c.type_to_iqandstindex = htobe32(v |
3553 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
3554 	    V_FW_IQ_CMD_VIID(vi->viid) |
3555 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
3556 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
3557 	    F_FW_IQ_CMD_IQGTSMODE |
3558 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
3559 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
3560 	c.iqsize = htobe16(iq->qsize);
3561 	c.iqaddr = htobe64(iq->ba);
3562 	c.iqns_to_fl0congen = htobe32(V_FW_IQ_CMD_IQTYPE(iq->qtype));
3563 	if (iq->cong_drop != -1) {
3564 		cong_map = iq->qtype == IQ_ETH ? pi->rx_e_chan_map : 0;
3565 		c.iqns_to_fl0congen |= htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
3566 	}
3567 
3568 	if (fl) {
3569 		bzero(fl->desc, fl->sidx * EQ_ESIZE + sc->params.sge.spg_len);
3570 		c.iqns_to_fl0congen |=
3571 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
3572 			F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
3573 			(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
3574 			(fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
3575 			    0));
3576 		if (iq->cong_drop != -1) {
3577 			c.iqns_to_fl0congen |=
3578 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong_map) |
3579 				    F_FW_IQ_CMD_FL0CONGCIF |
3580 				    F_FW_IQ_CMD_FL0CONGEN);
3581 		}
3582 		c.fl0dcaen_to_fl0cidxfthresh =
3583 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
3584 			X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B_T6) |
3585 			V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
3586 			X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
3587 		c.fl0size = htobe16(fl->qsize);
3588 		c.fl0addr = htobe64(fl->ba);
3589 	}
3590 
3591 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3592 	if (rc != 0) {
3593 		CH_ERR(sc, "failed to create hw ingress queue: %d\n", rc);
3594 		return (rc);
3595 	}
3596 
3597 	iq->cidx = 0;
3598 	iq->gen = F_RSPD_GEN;
3599 	iq->cntxt_id = be16toh(c.iqid);
3600 	iq->abs_id = be16toh(c.physiqid);
3601 
3602 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
3603 	if (cntxt_id >= sc->sge.iqmap_sz) {
3604 		panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
3605 		    cntxt_id, sc->sge.iqmap_sz - 1);
3606 	}
3607 	sc->sge.iqmap[cntxt_id] = iq;
3608 
3609 	if (fl) {
3610 		u_int qid;
3611 #ifdef INVARIANTS
3612 		int i;
3613 
3614 		MPASS(!(fl->flags & FL_BUF_RESUME));
3615 		for (i = 0; i < fl->sidx * 8; i++)
3616 			MPASS(fl->sdesc[i].cl == NULL);
3617 #endif
3618 		fl->cntxt_id = be16toh(c.fl0id);
3619 		fl->pidx = fl->cidx = fl->hw_cidx = fl->dbidx = 0;
3620 		fl->rx_offset = 0;
3621 		fl->flags &= ~(FL_STARVING | FL_DOOMED);
3622 
3623 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
3624 		if (cntxt_id >= sc->sge.eqmap_sz) {
3625 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
3626 			    __func__, cntxt_id, sc->sge.eqmap_sz - 1);
3627 		}
3628 		sc->sge.eqmap[cntxt_id] = (void *)fl;
3629 
3630 		qid = fl->cntxt_id;
3631 		if (isset(&sc->doorbells, DOORBELL_UDB)) {
3632 			uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3633 			uint32_t mask = (1 << s_qpp) - 1;
3634 			volatile uint8_t *udb;
3635 
3636 			udb = sc->udbs_base + UDBS_DB_OFFSET;
3637 			udb += (qid >> s_qpp) << PAGE_SHIFT;
3638 			qid &= mask;
3639 			if (qid < PAGE_SIZE / UDBS_SEG_SIZE) {
3640 				udb += qid << UDBS_SEG_SHIFT;
3641 				qid = 0;
3642 			}
3643 			fl->udb = (volatile void *)udb;
3644 		}
3645 		fl->dbval = V_QID(qid) | sc->chip_params->sge_fl_db;
3646 
3647 		FL_LOCK(fl);
3648 		/* Enough to make sure the SGE doesn't think it's starved */
3649 		refill_fl(sc, fl, fl->lowat);
3650 		FL_UNLOCK(fl);
3651 	}
3652 
3653 	if (chip_id(sc) >= CHELSIO_T5 && !(sc->flags & IS_VF) &&
3654 	    iq->cong_drop != -1) {
3655 		t4_sge_set_conm_context(sc, iq->cntxt_id, iq->cong_drop,
3656 		    cong_map);
3657 	}
3658 
3659 	/* Enable IQ interrupts */
3660 	atomic_store_rel_int(&iq->state, IQS_IDLE);
3661 	t4_write_reg(sc, sc->sge_gts_reg, V_SEINTARM(iq->intr_params) |
3662 	    V_INGRESSQID(iq->cntxt_id));
3663 
3664 	iq->flags |= IQ_HW_ALLOCATED;
3665 
3666 	return (0);
3667 }
3668 
3669 static int
3670 free_iq_fl_hwq(struct adapter *sc, struct sge_iq *iq, struct sge_fl *fl)
3671 {
3672 	int rc;
3673 
3674 	MPASS(iq->flags & IQ_HW_ALLOCATED);
3675 	rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0, FW_IQ_TYPE_FL_INT_CAP,
3676 	    iq->cntxt_id, fl ? fl->cntxt_id : 0xffff, 0xffff);
3677 	if (rc != 0) {
3678 		CH_ERR(sc, "failed to free iq %p: %d\n", iq, rc);
3679 		return (rc);
3680 	}
3681 	iq->flags &= ~IQ_HW_ALLOCATED;
3682 
3683 	return (0);
3684 }
3685 
3686 static void
3687 add_iq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
3688     struct sge_iq *iq)
3689 {
3690 	struct sysctl_oid_list *children;
3691 
3692 	if (ctx == NULL || oid == NULL)
3693 		return;
3694 
3695 	children = SYSCTL_CHILDREN(oid);
3696 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD, &iq->ba,
3697 	    "bus address of descriptor ring");
3698 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3699 	    iq->qsize * IQ_ESIZE, "descriptor ring size in bytes");
3700 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
3701 	    &iq->abs_id, 0, "absolute id of the queue");
3702 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3703 	    &iq->cntxt_id, 0, "SGE context id of the queue");
3704 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &iq->cidx,
3705 	    0, "consumer index");
3706 }
3707 
3708 static void
3709 add_fl_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
3710     struct sysctl_oid *oid, struct sge_fl *fl)
3711 {
3712 	struct sysctl_oid_list *children;
3713 
3714 	if (ctx == NULL || oid == NULL)
3715 		return;
3716 
3717 	children = SYSCTL_CHILDREN(oid);
3718 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl",
3719 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist");
3720 	children = SYSCTL_CHILDREN(oid);
3721 
3722 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3723 	    &fl->ba, "bus address of descriptor ring");
3724 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3725 	    fl->sidx * EQ_ESIZE + sc->params.sge.spg_len,
3726 	    "desc ring size in bytes");
3727 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3728 	    &fl->cntxt_id, 0, "SGE context id of the freelist");
3729 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL,
3730 	    fl_pad ? 1 : 0, "padding enabled");
3731 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL,
3732 	    fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled");
3733 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
3734 	    0, "consumer index");
3735 	if (fl->flags & FL_BUF_PACKING) {
3736 		SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
3737 		    CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
3738 	}
3739 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
3740 	    0, "producer index");
3741 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
3742 	    CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
3743 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
3744 	    CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
3745 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
3746 	    CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
3747 }
3748 
3749 /*
3750  * Idempotent.
3751  */
3752 static int
3753 alloc_fwq(struct adapter *sc)
3754 {
3755 	int rc, intr_idx;
3756 	struct sge_iq *fwq = &sc->sge.fwq;
3757 	struct vi_info *vi = &sc->port[0]->vi[0];
3758 
3759 	if (!(fwq->flags & IQ_SW_ALLOCATED)) {
3760 		MPASS(!(fwq->flags & IQ_HW_ALLOCATED));
3761 
3762 		if (sc->flags & IS_VF)
3763 			intr_idx = 0;
3764 		else
3765 			intr_idx = sc->intr_count > 1 ? 1 : 0;
3766 		init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE, intr_idx, -1, IQ_OTHER);
3767 		rc = alloc_iq_fl(vi, fwq, NULL, &sc->ctx, sc->fwq_oid);
3768 		if (rc != 0) {
3769 			CH_ERR(sc, "failed to allocate fwq: %d\n", rc);
3770 			return (rc);
3771 		}
3772 		MPASS(fwq->flags & IQ_SW_ALLOCATED);
3773 	}
3774 
3775 	if (!(fwq->flags & IQ_HW_ALLOCATED)) {
3776 		MPASS(fwq->flags & IQ_SW_ALLOCATED);
3777 
3778 		rc = alloc_iq_fl_hwq(vi, fwq, NULL);
3779 		if (rc != 0) {
3780 			CH_ERR(sc, "failed to create hw fwq: %d\n", rc);
3781 			return (rc);
3782 		}
3783 		MPASS(fwq->flags & IQ_HW_ALLOCATED);
3784 	}
3785 
3786 	return (0);
3787 }
3788 
3789 /*
3790  * Idempotent.
3791  */
3792 static void
3793 free_fwq(struct adapter *sc)
3794 {
3795 	struct sge_iq *fwq = &sc->sge.fwq;
3796 
3797 	if (fwq->flags & IQ_HW_ALLOCATED) {
3798 		MPASS(fwq->flags & IQ_SW_ALLOCATED);
3799 		free_iq_fl_hwq(sc, fwq, NULL);
3800 		MPASS(!(fwq->flags & IQ_HW_ALLOCATED));
3801 	}
3802 
3803 	if (fwq->flags & IQ_SW_ALLOCATED) {
3804 		MPASS(!(fwq->flags & IQ_HW_ALLOCATED));
3805 		free_iq_fl(sc, fwq, NULL);
3806 		MPASS(!(fwq->flags & IQ_SW_ALLOCATED));
3807 	}
3808 }
3809 
3810 /*
3811  * Idempotent.
3812  */
3813 static int
3814 alloc_ctrlq(struct adapter *sc, int idx)
3815 {
3816 	int rc;
3817 	char name[16];
3818 	struct sysctl_oid *oid;
3819 	struct sge_wrq *ctrlq = &sc->sge.ctrlq[idx];
3820 
3821 	MPASS(idx < sc->params.nports);
3822 
3823 	if (!(ctrlq->eq.flags & EQ_SW_ALLOCATED)) {
3824 		MPASS(!(ctrlq->eq.flags & EQ_HW_ALLOCATED));
3825 
3826 		snprintf(name, sizeof(name), "%d", idx);
3827 		oid = SYSCTL_ADD_NODE(&sc->ctx, SYSCTL_CHILDREN(sc->ctrlq_oid),
3828 		    OID_AUTO, name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
3829 		    "ctrl queue");
3830 
3831 		snprintf(name, sizeof(name), "%s ctrlq%d",
3832 		    device_get_nameunit(sc->dev), idx);
3833 		init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE,
3834 		    sc->port[idx]->tx_chan, &sc->sge.fwq, name);
3835 		rc = alloc_wrq(sc, NULL, ctrlq, &sc->ctx, oid);
3836 		if (rc != 0) {
3837 			CH_ERR(sc, "failed to allocate ctrlq%d: %d\n", idx, rc);
3838 			sysctl_remove_oid(oid, 1, 1);
3839 			return (rc);
3840 		}
3841 		MPASS(ctrlq->eq.flags & EQ_SW_ALLOCATED);
3842 	}
3843 
3844 	if (!(ctrlq->eq.flags & EQ_HW_ALLOCATED)) {
3845 		MPASS(ctrlq->eq.flags & EQ_SW_ALLOCATED);
3846 
3847 		rc = alloc_eq_hwq(sc, NULL, &ctrlq->eq);
3848 		if (rc != 0) {
3849 			CH_ERR(sc, "failed to create hw ctrlq%d: %d\n", idx, rc);
3850 			return (rc);
3851 		}
3852 		MPASS(ctrlq->eq.flags & EQ_HW_ALLOCATED);
3853 	}
3854 
3855 	return (0);
3856 }
3857 
3858 /*
3859  * Idempotent.
3860  */
3861 static void
3862 free_ctrlq(struct adapter *sc, int idx)
3863 {
3864 	struct sge_wrq *ctrlq = &sc->sge.ctrlq[idx];
3865 
3866 	if (ctrlq->eq.flags & EQ_HW_ALLOCATED) {
3867 		MPASS(ctrlq->eq.flags & EQ_SW_ALLOCATED);
3868 		free_eq_hwq(sc, NULL, &ctrlq->eq);
3869 		MPASS(!(ctrlq->eq.flags & EQ_HW_ALLOCATED));
3870 	}
3871 
3872 	if (ctrlq->eq.flags & EQ_SW_ALLOCATED) {
3873 		MPASS(!(ctrlq->eq.flags & EQ_HW_ALLOCATED));
3874 		free_wrq(sc, ctrlq);
3875 		MPASS(!(ctrlq->eq.flags & EQ_SW_ALLOCATED));
3876 	}
3877 }
3878 
3879 int
3880 t4_sge_set_conm_context(struct adapter *sc, int cntxt_id, int cong_drop,
3881     int cong_map)
3882 {
3883 	const int cng_ch_bits_log = sc->chip_params->cng_ch_bits_log;
3884 	uint32_t param, val;
3885 	uint16_t ch_map;
3886 	int cong_mode, rc, i;
3887 
3888 	if (chip_id(sc) < CHELSIO_T5)
3889 		return (ENOTSUP);
3890 
3891 	/* Convert the driver knob to the mode understood by the firmware. */
3892 	switch (cong_drop) {
3893 	case -1:
3894 		cong_mode = X_CONMCTXT_CNGTPMODE_DISABLE;
3895 		break;
3896 	case 0:
3897 		cong_mode = X_CONMCTXT_CNGTPMODE_CHANNEL;
3898 		break;
3899 	case 1:
3900 		cong_mode = X_CONMCTXT_CNGTPMODE_QUEUE;
3901 		break;
3902 	case 2:
3903 		cong_mode = X_CONMCTXT_CNGTPMODE_BOTH;
3904 		break;
3905 	default:
3906 		MPASS(0);
3907 		CH_ERR(sc, "cong_drop = %d is invalid (ingress queue %d).\n",
3908 		    cong_drop, cntxt_id);
3909 		return (EINVAL);
3910 	}
3911 
3912 	param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
3913 	    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
3914 	    V_FW_PARAMS_PARAM_YZ(cntxt_id);
3915 	val = V_CONMCTXT_CNGTPMODE(cong_mode);
3916 	if (cong_mode == X_CONMCTXT_CNGTPMODE_CHANNEL ||
3917 	    cong_mode == X_CONMCTXT_CNGTPMODE_BOTH) {
3918 		for (i = 0, ch_map = 0; i < 4; i++) {
3919 			if (cong_map & (1 << i))
3920 				ch_map |= 1 << (i << cng_ch_bits_log);
3921 		}
3922 		val |= V_CONMCTXT_CNGCHMAP(ch_map);
3923 	}
3924 	rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
3925 	if (rc != 0) {
3926 		CH_ERR(sc, "failed to set congestion manager context "
3927 		    "for ingress queue %d: %d\n", cntxt_id, rc);
3928 	}
3929 
3930 	return (rc);
3931 }
3932 
3933 /*
3934  * Idempotent.
3935  */
3936 static int
3937 alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int idx, int intr_idx,
3938     int maxp)
3939 {
3940 	int rc;
3941 	struct adapter *sc = vi->adapter;
3942 	if_t ifp = vi->ifp;
3943 	struct sysctl_oid *oid;
3944 	char name[16];
3945 
3946 	if (!(rxq->iq.flags & IQ_SW_ALLOCATED)) {
3947 		MPASS(!(rxq->iq.flags & IQ_HW_ALLOCATED));
3948 #if defined(INET) || defined(INET6)
3949 		rc = tcp_lro_init_args(&rxq->lro, ifp, lro_entries, lro_mbufs);
3950 		if (rc != 0)
3951 			return (rc);
3952 		MPASS(rxq->lro.ifp == ifp);	/* also indicates LRO init'ed */
3953 #endif
3954 		rxq->ifp = ifp;
3955 
3956 		snprintf(name, sizeof(name), "%d", idx);
3957 		oid = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(vi->rxq_oid),
3958 		    OID_AUTO, name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
3959 		    "rx queue");
3960 
3961 		init_iq(&rxq->iq, sc, vi->tmr_idx, vi->pktc_idx, vi->qsize_rxq,
3962 		    intr_idx, cong_drop, IQ_ETH);
3963 #if defined(INET) || defined(INET6)
3964 		if (if_getcapenable(ifp) & IFCAP_LRO)
3965 			rxq->iq.flags |= IQ_LRO_ENABLED;
3966 #endif
3967 		if (if_getcapenable(ifp) & IFCAP_HWRXTSTMP)
3968 			rxq->iq.flags |= IQ_RX_TIMESTAMP;
3969 		snprintf(name, sizeof(name), "%s rxq%d-fl",
3970 		    device_get_nameunit(vi->dev), idx);
3971 		init_fl(sc, &rxq->fl, vi->qsize_rxq / 8, maxp, name);
3972 		rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, &vi->ctx, oid);
3973 		if (rc != 0) {
3974 			CH_ERR(vi, "failed to allocate rxq%d: %d\n", idx, rc);
3975 			sysctl_remove_oid(oid, 1, 1);
3976 #if defined(INET) || defined(INET6)
3977 			tcp_lro_free(&rxq->lro);
3978 			rxq->lro.ifp = NULL;
3979 #endif
3980 			return (rc);
3981 		}
3982 		MPASS(rxq->iq.flags & IQ_SW_ALLOCATED);
3983 		add_rxq_sysctls(&vi->ctx, oid, rxq);
3984 	}
3985 
3986 	if (!(rxq->iq.flags & IQ_HW_ALLOCATED)) {
3987 		MPASS(rxq->iq.flags & IQ_SW_ALLOCATED);
3988 		rc = alloc_iq_fl_hwq(vi, &rxq->iq, &rxq->fl);
3989 		if (rc != 0) {
3990 			CH_ERR(vi, "failed to create hw rxq%d: %d\n", idx, rc);
3991 			return (rc);
3992 		}
3993 		MPASS(rxq->iq.flags & IQ_HW_ALLOCATED);
3994 
3995 		if (idx == 0)
3996 			sc->sge.iq_base = rxq->iq.abs_id - rxq->iq.cntxt_id;
3997 		else
3998 			KASSERT(rxq->iq.cntxt_id + sc->sge.iq_base == rxq->iq.abs_id,
3999 			    ("iq_base mismatch"));
4000 		KASSERT(sc->sge.iq_base == 0 || sc->flags & IS_VF,
4001 		    ("PF with non-zero iq_base"));
4002 
4003 		/*
4004 		 * The freelist is just barely above the starvation threshold
4005 		 * right now, fill it up a bit more.
4006 		 */
4007 		FL_LOCK(&rxq->fl);
4008 		refill_fl(sc, &rxq->fl, 128);
4009 		FL_UNLOCK(&rxq->fl);
4010 	}
4011 
4012 	return (0);
4013 }
4014 
4015 /*
4016  * Idempotent.
4017  */
4018 static void
4019 free_rxq(struct vi_info *vi, struct sge_rxq *rxq)
4020 {
4021 	if (rxq->iq.flags & IQ_HW_ALLOCATED) {
4022 		MPASS(rxq->iq.flags & IQ_SW_ALLOCATED);
4023 		free_iq_fl_hwq(vi->adapter, &rxq->iq, &rxq->fl);
4024 		MPASS(!(rxq->iq.flags & IQ_HW_ALLOCATED));
4025 	}
4026 
4027 	if (rxq->iq.flags & IQ_SW_ALLOCATED) {
4028 		MPASS(!(rxq->iq.flags & IQ_HW_ALLOCATED));
4029 #if defined(INET) || defined(INET6)
4030 		tcp_lro_free(&rxq->lro);
4031 #endif
4032 		free_iq_fl(vi->adapter, &rxq->iq, &rxq->fl);
4033 		MPASS(!(rxq->iq.flags & IQ_SW_ALLOCATED));
4034 		bzero(rxq, sizeof(*rxq));
4035 	}
4036 }
4037 
4038 static void
4039 add_rxq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
4040     struct sge_rxq *rxq)
4041 {
4042 	struct sysctl_oid_list *children;
4043 
4044 	if (ctx == NULL || oid == NULL)
4045 		return;
4046 
4047 	children = SYSCTL_CHILDREN(oid);
4048 #if defined(INET) || defined(INET6)
4049 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
4050 	    &rxq->lro.lro_queued, 0, NULL);
4051 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
4052 	    &rxq->lro.lro_flushed, 0, NULL);
4053 #endif
4054 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
4055 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
4056 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "vlan_extraction", CTLFLAG_RD,
4057 	    &rxq->vlan_extraction, "# of times hardware extracted 802.1Q tag");
4058 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "vxlan_rxcsum", CTLFLAG_RD,
4059 	    &rxq->vxlan_rxcsum,
4060 	    "# of times hardware assisted with inner checksum (VXLAN)");
4061 }
4062 
4063 #ifdef TCP_OFFLOAD
4064 /*
4065  * Idempotent.
4066  */
4067 static int
4068 alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq, int idx,
4069     int intr_idx, int maxp)
4070 {
4071 	int rc;
4072 	struct adapter *sc = vi->adapter;
4073 	struct sysctl_oid *oid;
4074 	char name[16];
4075 
4076 	if (!(ofld_rxq->iq.flags & IQ_SW_ALLOCATED)) {
4077 		MPASS(!(ofld_rxq->iq.flags & IQ_HW_ALLOCATED));
4078 
4079 		snprintf(name, sizeof(name), "%d", idx);
4080 		oid = SYSCTL_ADD_NODE(&vi->ctx,
4081 		    SYSCTL_CHILDREN(vi->ofld_rxq_oid), OID_AUTO, name,
4082 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "offload rx queue");
4083 
4084 		init_iq(&ofld_rxq->iq, sc, vi->ofld_tmr_idx, vi->ofld_pktc_idx,
4085 		    vi->qsize_rxq, intr_idx, ofld_cong_drop, IQ_OFLD);
4086 		snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
4087 		    device_get_nameunit(vi->dev), idx);
4088 		init_fl(sc, &ofld_rxq->fl, vi->qsize_rxq / 8, maxp, name);
4089 		rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, &vi->ctx,
4090 		    oid);
4091 		if (rc != 0) {
4092 			CH_ERR(vi, "failed to allocate ofld_rxq%d: %d\n", idx,
4093 			    rc);
4094 			sysctl_remove_oid(oid, 1, 1);
4095 			return (rc);
4096 		}
4097 		MPASS(ofld_rxq->iq.flags & IQ_SW_ALLOCATED);
4098 		ofld_rxq->rx_iscsi_ddp_setup_ok = counter_u64_alloc(M_WAITOK);
4099 		ofld_rxq->rx_iscsi_ddp_setup_error =
4100 		    counter_u64_alloc(M_WAITOK);
4101 		ofld_rxq->ddp_buffer_alloc = counter_u64_alloc(M_WAITOK);
4102 		ofld_rxq->ddp_buffer_reuse = counter_u64_alloc(M_WAITOK);
4103 		ofld_rxq->ddp_buffer_free = counter_u64_alloc(M_WAITOK);
4104 		add_ofld_rxq_sysctls(&vi->ctx, oid, ofld_rxq);
4105 	}
4106 
4107 	if (!(ofld_rxq->iq.flags & IQ_HW_ALLOCATED)) {
4108 		MPASS(ofld_rxq->iq.flags & IQ_SW_ALLOCATED);
4109 		rc = alloc_iq_fl_hwq(vi, &ofld_rxq->iq, &ofld_rxq->fl);
4110 		if (rc != 0) {
4111 			CH_ERR(vi, "failed to create hw ofld_rxq%d: %d\n", idx,
4112 			    rc);
4113 			return (rc);
4114 		}
4115 		MPASS(ofld_rxq->iq.flags & IQ_HW_ALLOCATED);
4116 	}
4117 	return (rc);
4118 }
4119 
4120 /*
4121  * Idempotent.
4122  */
4123 static void
4124 free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq)
4125 {
4126 	if (ofld_rxq->iq.flags & IQ_HW_ALLOCATED) {
4127 		MPASS(ofld_rxq->iq.flags & IQ_SW_ALLOCATED);
4128 		free_iq_fl_hwq(vi->adapter, &ofld_rxq->iq, &ofld_rxq->fl);
4129 		MPASS(!(ofld_rxq->iq.flags & IQ_HW_ALLOCATED));
4130 	}
4131 
4132 	if (ofld_rxq->iq.flags & IQ_SW_ALLOCATED) {
4133 		MPASS(!(ofld_rxq->iq.flags & IQ_HW_ALLOCATED));
4134 		free_iq_fl(vi->adapter, &ofld_rxq->iq, &ofld_rxq->fl);
4135 		MPASS(!(ofld_rxq->iq.flags & IQ_SW_ALLOCATED));
4136 		counter_u64_free(ofld_rxq->rx_iscsi_ddp_setup_ok);
4137 		counter_u64_free(ofld_rxq->rx_iscsi_ddp_setup_error);
4138 		counter_u64_free(ofld_rxq->ddp_buffer_alloc);
4139 		counter_u64_free(ofld_rxq->ddp_buffer_reuse);
4140 		counter_u64_free(ofld_rxq->ddp_buffer_free);
4141 		bzero(ofld_rxq, sizeof(*ofld_rxq));
4142 	}
4143 }
4144 
4145 static void
4146 add_ofld_rxq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
4147     struct sge_ofld_rxq *ofld_rxq)
4148 {
4149 	struct sysctl_oid_list *children;
4150 
4151 	if (ctx == NULL || oid == NULL)
4152 		return;
4153 
4154 	children = SYSCTL_CHILDREN(oid);
4155 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "rx_aio_ddp_jobs",
4156 	    CTLFLAG_RD, &ofld_rxq->rx_aio_ddp_jobs, 0,
4157 	    "# of aio_read(2) jobs completed via DDP");
4158 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "rx_aio_ddp_octets",
4159 	    CTLFLAG_RD, &ofld_rxq->rx_aio_ddp_octets, 0,
4160 	    "# of octets placed directly for aio_read(2) jobs");
4161 	SYSCTL_ADD_ULONG(ctx, children, OID_AUTO,
4162 	    "rx_toe_tls_records", CTLFLAG_RD, &ofld_rxq->rx_toe_tls_records,
4163 	    "# of TOE TLS records received");
4164 	SYSCTL_ADD_ULONG(ctx, children, OID_AUTO,
4165 	    "rx_toe_tls_octets", CTLFLAG_RD, &ofld_rxq->rx_toe_tls_octets,
4166 	    "# of payload octets in received TOE TLS records");
4167 	SYSCTL_ADD_ULONG(ctx, children, OID_AUTO,
4168 	    "rx_toe_ddp_octets", CTLFLAG_RD, &ofld_rxq->rx_toe_ddp_octets,
4169 	    "# of payload octets received via TCP DDP");
4170 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO,
4171 	    "ddp_buffer_alloc", CTLFLAG_RD, &ofld_rxq->ddp_buffer_alloc,
4172 	    "# of DDP RCV buffers allocated");
4173 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO,
4174 	    "ddp_buffer_reuse", CTLFLAG_RD, &ofld_rxq->ddp_buffer_reuse,
4175 	    "# of DDP RCV buffers reused");
4176 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO,
4177 	    "ddp_buffer_free", CTLFLAG_RD, &ofld_rxq->ddp_buffer_free,
4178 	    "# of DDP RCV buffers freed");
4179 
4180 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "iscsi",
4181 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "TOE iSCSI statistics");
4182 	children = SYSCTL_CHILDREN(oid);
4183 
4184 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "ddp_setup_ok",
4185 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_ddp_setup_ok,
4186 	    "# of times DDP buffer was setup successfully.");
4187 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "ddp_setup_error",
4188 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_ddp_setup_error,
4189 	    "# of times DDP buffer setup failed.");
4190 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "ddp_octets",
4191 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_ddp_octets, 0,
4192 	    "# of octets placed directly");
4193 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "ddp_pdus",
4194 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_ddp_pdus, 0,
4195 	    "# of PDUs with data placed directly.");
4196 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "fl_octets",
4197 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_fl_octets, 0,
4198 	    "# of data octets delivered in freelist");
4199 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "fl_pdus",
4200 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_fl_pdus, 0,
4201 	    "# of PDUs with data delivered in freelist");
4202 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "padding_errors",
4203 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_padding_errors, 0,
4204 	    "# of PDUs with invalid padding");
4205 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "header_digest_errors",
4206 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_header_digest_errors, 0,
4207 	    "# of PDUs with invalid header digests");
4208 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "data_digest_errors",
4209 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_data_digest_errors, 0,
4210 	    "# of PDUs with invalid data digests");
4211 }
4212 #endif
4213 
4214 /*
4215  * Returns a reasonable automatic cidx flush threshold for a given queue size.
4216  */
4217 static u_int
4218 qsize_to_fthresh(int qsize)
4219 {
4220 	u_int fthresh;
4221 
4222 	while (!powerof2(qsize))
4223 		qsize++;
4224 	fthresh = ilog2(qsize);
4225 	if (fthresh > X_CIDXFLUSHTHRESH_128)
4226 		fthresh = X_CIDXFLUSHTHRESH_128;
4227 
4228 	return (fthresh);
4229 }
4230 
4231 static int
4232 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
4233 {
4234 	int rc, cntxt_id;
4235 	struct fw_eq_ctrl_cmd c;
4236 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
4237 
4238 	bzero(&c, sizeof(c));
4239 
4240 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
4241 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
4242 	    V_FW_EQ_CTRL_CMD_VFN(0));
4243 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
4244 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
4245 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid));
4246 	c.physeqid_pkd = htobe32(0);
4247 	c.fetchszm_to_iqid =
4248 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
4249 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
4250 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
4251 	c.dcaen_to_eqsize =
4252 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
4253 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
4254 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
4255 		V_FW_EQ_CTRL_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
4256 		V_FW_EQ_CTRL_CMD_EQSIZE(qsize));
4257 	c.eqaddr = htobe64(eq->ba);
4258 
4259 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
4260 	if (rc != 0) {
4261 		CH_ERR(sc, "failed to create hw ctrlq for tx_chan %d: %d\n",
4262 		    eq->tx_chan, rc);
4263 		return (rc);
4264 	}
4265 
4266 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
4267 	eq->abs_id = G_FW_EQ_CTRL_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
4268 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
4269 	if (cntxt_id >= sc->sge.eqmap_sz)
4270 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
4271 		cntxt_id, sc->sge.eqmap_sz - 1);
4272 	sc->sge.eqmap[cntxt_id] = eq;
4273 
4274 	return (rc);
4275 }
4276 
4277 static int
4278 eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
4279 {
4280 	int rc, cntxt_id;
4281 	struct fw_eq_eth_cmd c;
4282 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
4283 
4284 	bzero(&c, sizeof(c));
4285 
4286 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
4287 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
4288 	    V_FW_EQ_ETH_CMD_VFN(0));
4289 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
4290 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
4291 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
4292 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
4293 	c.fetchszm_to_iqid =
4294 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
4295 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
4296 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
4297 	c.dcaen_to_eqsize =
4298 	    htobe32(V_FW_EQ_ETH_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
4299 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
4300 		V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
4301 		V_FW_EQ_ETH_CMD_EQSIZE(qsize));
4302 	c.eqaddr = htobe64(eq->ba);
4303 
4304 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
4305 	if (rc != 0) {
4306 		device_printf(vi->dev,
4307 		    "failed to create Ethernet egress queue: %d\n", rc);
4308 		return (rc);
4309 	}
4310 
4311 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
4312 	eq->abs_id = G_FW_EQ_ETH_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
4313 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
4314 	if (cntxt_id >= sc->sge.eqmap_sz)
4315 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
4316 		cntxt_id, sc->sge.eqmap_sz - 1);
4317 	sc->sge.eqmap[cntxt_id] = eq;
4318 
4319 	return (rc);
4320 }
4321 
4322 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
4323 static int
4324 ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
4325 {
4326 	int rc, cntxt_id;
4327 	struct fw_eq_ofld_cmd c;
4328 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
4329 
4330 	bzero(&c, sizeof(c));
4331 
4332 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
4333 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
4334 	    V_FW_EQ_OFLD_CMD_VFN(0));
4335 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
4336 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
4337 	c.fetchszm_to_iqid =
4338 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
4339 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
4340 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
4341 	c.dcaen_to_eqsize =
4342 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
4343 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
4344 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
4345 		V_FW_EQ_OFLD_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
4346 		V_FW_EQ_OFLD_CMD_EQSIZE(qsize));
4347 	c.eqaddr = htobe64(eq->ba);
4348 
4349 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
4350 	if (rc != 0) {
4351 		device_printf(vi->dev,
4352 		    "failed to create egress queue for TCP offload: %d\n", rc);
4353 		return (rc);
4354 	}
4355 
4356 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
4357 	eq->abs_id = G_FW_EQ_OFLD_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
4358 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
4359 	if (cntxt_id >= sc->sge.eqmap_sz)
4360 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
4361 		cntxt_id, sc->sge.eqmap_sz - 1);
4362 	sc->sge.eqmap[cntxt_id] = eq;
4363 
4364 	return (rc);
4365 }
4366 #endif
4367 
4368 /* SW only */
4369 static int
4370 alloc_eq(struct adapter *sc, struct sge_eq *eq, struct sysctl_ctx_list *ctx,
4371     struct sysctl_oid *oid)
4372 {
4373 	int rc, qsize;
4374 	size_t len;
4375 
4376 	MPASS(!(eq->flags & EQ_SW_ALLOCATED));
4377 
4378 	qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
4379 	len = qsize * EQ_ESIZE;
4380 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map, &eq->ba,
4381 	    (void **)&eq->desc);
4382 	if (rc)
4383 		return (rc);
4384 	if (ctx != NULL && oid != NULL)
4385 		add_eq_sysctls(sc, ctx, oid, eq);
4386 	eq->flags |= EQ_SW_ALLOCATED;
4387 
4388 	return (0);
4389 }
4390 
4391 /* SW only */
4392 static void
4393 free_eq(struct adapter *sc, struct sge_eq *eq)
4394 {
4395 	MPASS(eq->flags & EQ_SW_ALLOCATED);
4396 	if (eq->type == EQ_ETH)
4397 		MPASS(eq->pidx == eq->cidx);
4398 
4399 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
4400 	mtx_destroy(&eq->eq_lock);
4401 	bzero(eq, sizeof(*eq));
4402 }
4403 
4404 static void
4405 add_eq_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
4406     struct sysctl_oid *oid, struct sge_eq *eq)
4407 {
4408 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
4409 
4410 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD, &eq->ba,
4411 	    "bus address of descriptor ring");
4412 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
4413 	    eq->sidx * EQ_ESIZE + sc->params.sge.spg_len,
4414 	    "desc ring size in bytes");
4415 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
4416 	    &eq->abs_id, 0, "absolute id of the queue");
4417 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
4418 	    &eq->cntxt_id, 0, "SGE context id of the queue");
4419 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &eq->cidx,
4420 	    0, "consumer index");
4421 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &eq->pidx,
4422 	    0, "producer index");
4423 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
4424 	    eq->sidx, "status page index");
4425 }
4426 
4427 static int
4428 alloc_eq_hwq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
4429 {
4430 	int rc;
4431 
4432 	MPASS(!(eq->flags & EQ_HW_ALLOCATED));
4433 
4434 	eq->iqid = eq->iq->cntxt_id;
4435 	eq->pidx = eq->cidx = eq->dbidx = 0;
4436 	/* Note that equeqidx is not used with sge_wrq (OFLD/CTRL) queues. */
4437 	eq->equeqidx = 0;
4438 	eq->doorbells = sc->doorbells;
4439 	bzero(eq->desc, eq->sidx * EQ_ESIZE + sc->params.sge.spg_len);
4440 
4441 	switch (eq->type) {
4442 	case EQ_CTRL:
4443 		rc = ctrl_eq_alloc(sc, eq);
4444 		break;
4445 
4446 	case EQ_ETH:
4447 		rc = eth_eq_alloc(sc, vi, eq);
4448 		break;
4449 
4450 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
4451 	case EQ_OFLD:
4452 		rc = ofld_eq_alloc(sc, vi, eq);
4453 		break;
4454 #endif
4455 
4456 	default:
4457 		panic("%s: invalid eq type %d.", __func__, eq->type);
4458 	}
4459 	if (rc != 0) {
4460 		CH_ERR(sc, "failed to allocate egress queue(%d): %d\n",
4461 		    eq->type, rc);
4462 		return (rc);
4463 	}
4464 
4465 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
4466 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
4467 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
4468 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
4469 		uint32_t mask = (1 << s_qpp) - 1;
4470 		volatile uint8_t *udb;
4471 
4472 		udb = sc->udbs_base + UDBS_DB_OFFSET;
4473 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
4474 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
4475 		if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
4476 			clrbit(&eq->doorbells, DOORBELL_WCWR);
4477 		else {
4478 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
4479 			eq->udb_qid = 0;
4480 		}
4481 		eq->udb = (volatile void *)udb;
4482 	}
4483 
4484 	eq->flags |= EQ_HW_ALLOCATED;
4485 	return (0);
4486 }
4487 
4488 static int
4489 free_eq_hwq(struct adapter *sc, struct vi_info *vi __unused, struct sge_eq *eq)
4490 {
4491 	int rc;
4492 
4493 	MPASS(eq->flags & EQ_HW_ALLOCATED);
4494 
4495 	switch (eq->type) {
4496 	case EQ_CTRL:
4497 		rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0, eq->cntxt_id);
4498 		break;
4499 	case EQ_ETH:
4500 		rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0, eq->cntxt_id);
4501 		break;
4502 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
4503 	case EQ_OFLD:
4504 		rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0, eq->cntxt_id);
4505 		break;
4506 #endif
4507 	default:
4508 		panic("%s: invalid eq type %d.", __func__, eq->type);
4509 	}
4510 	if (rc != 0) {
4511 		CH_ERR(sc, "failed to free eq (type %d): %d\n", eq->type, rc);
4512 		return (rc);
4513 	}
4514 	eq->flags &= ~EQ_HW_ALLOCATED;
4515 
4516 	return (0);
4517 }
4518 
4519 static int
4520 alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq,
4521     struct sysctl_ctx_list *ctx, struct sysctl_oid *oid)
4522 {
4523 	struct sge_eq *eq = &wrq->eq;
4524 	int rc;
4525 
4526 	MPASS(!(eq->flags & EQ_SW_ALLOCATED));
4527 
4528 	rc = alloc_eq(sc, eq, ctx, oid);
4529 	if (rc)
4530 		return (rc);
4531 	MPASS(eq->flags & EQ_SW_ALLOCATED);
4532 	/* Can't fail after this. */
4533 
4534 	wrq->adapter = sc;
4535 	TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq);
4536 	TAILQ_INIT(&wrq->incomplete_wrs);
4537 	STAILQ_INIT(&wrq->wr_list);
4538 	wrq->nwr_pending = 0;
4539 	wrq->ndesc_needed = 0;
4540 	add_wrq_sysctls(ctx, oid, wrq);
4541 
4542 	return (0);
4543 }
4544 
4545 static void
4546 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
4547 {
4548 	free_eq(sc, &wrq->eq);
4549 	MPASS(wrq->nwr_pending == 0);
4550 	MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
4551 	MPASS(STAILQ_EMPTY(&wrq->wr_list));
4552 	bzero(wrq, sizeof(*wrq));
4553 }
4554 
4555 static void
4556 add_wrq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
4557     struct sge_wrq *wrq)
4558 {
4559 	struct sysctl_oid_list *children;
4560 
4561 	if (ctx == NULL || oid == NULL)
4562 		return;
4563 
4564 	children = SYSCTL_CHILDREN(oid);
4565 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD,
4566 	    &wrq->tx_wrs_direct, "# of work requests (direct)");
4567 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD,
4568 	    &wrq->tx_wrs_copied, "# of work requests (copied)");
4569 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_sspace", CTLFLAG_RD,
4570 	    &wrq->tx_wrs_ss, "# of work requests (copied from scratch space)");
4571 }
4572 
4573 /*
4574  * Idempotent.
4575  */
4576 static int
4577 alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx)
4578 {
4579 	int rc, iqidx;
4580 	struct port_info *pi = vi->pi;
4581 	struct adapter *sc = vi->adapter;
4582 	struct sge_eq *eq = &txq->eq;
4583 	struct txpkts *txp;
4584 	char name[16];
4585 	struct sysctl_oid *oid;
4586 
4587 	if (!(eq->flags & EQ_SW_ALLOCATED)) {
4588 		MPASS(!(eq->flags & EQ_HW_ALLOCATED));
4589 
4590 		snprintf(name, sizeof(name), "%d", idx);
4591 		oid = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(vi->txq_oid),
4592 		    OID_AUTO, name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
4593 		    "tx queue");
4594 
4595 		iqidx = vi->first_rxq + (idx % vi->nrxq);
4596 		snprintf(name, sizeof(name), "%s txq%d",
4597 		    device_get_nameunit(vi->dev), idx);
4598 		init_eq(sc, &txq->eq, EQ_ETH, vi->qsize_txq, pi->tx_chan,
4599 		    &sc->sge.rxq[iqidx].iq, name);
4600 
4601 		rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx,
4602 		    can_resume_eth_tx, M_CXGBE, &eq->eq_lock, M_WAITOK);
4603 		if (rc != 0) {
4604 			CH_ERR(vi, "failed to allocate mp_ring for txq%d: %d\n",
4605 			    idx, rc);
4606 failed:
4607 			sysctl_remove_oid(oid, 1, 1);
4608 			return (rc);
4609 		}
4610 
4611 		rc = alloc_eq(sc, eq, &vi->ctx, oid);
4612 		if (rc) {
4613 			CH_ERR(vi, "failed to allocate txq%d: %d\n", idx, rc);
4614 			mp_ring_free(txq->r);
4615 			goto failed;
4616 		}
4617 		MPASS(eq->flags & EQ_SW_ALLOCATED);
4618 		/* Can't fail after this point. */
4619 
4620 		TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq);
4621 		txq->ifp = vi->ifp;
4622 		txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK);
4623 		txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE,
4624 		    M_ZERO | M_WAITOK);
4625 
4626 		add_txq_sysctls(vi, &vi->ctx, oid, txq);
4627 	}
4628 
4629 	if (!(eq->flags & EQ_HW_ALLOCATED)) {
4630 		MPASS(eq->flags & EQ_SW_ALLOCATED);
4631 		rc = alloc_eq_hwq(sc, vi, eq);
4632 		if (rc != 0) {
4633 			CH_ERR(vi, "failed to create hw txq%d: %d\n", idx, rc);
4634 			return (rc);
4635 		}
4636 		MPASS(eq->flags & EQ_HW_ALLOCATED);
4637 		/* Can't fail after this point. */
4638 
4639 		if (idx == 0)
4640 			sc->sge.eq_base = eq->abs_id - eq->cntxt_id;
4641 		else
4642 			KASSERT(eq->cntxt_id + sc->sge.eq_base == eq->abs_id,
4643 			    ("eq_base mismatch"));
4644 		KASSERT(sc->sge.eq_base == 0 || sc->flags & IS_VF,
4645 		    ("PF with non-zero eq_base"));
4646 
4647 		txp = &txq->txp;
4648 		MPASS(nitems(txp->mb) >= sc->params.max_pkts_per_eth_tx_pkts_wr);
4649 		txq->txp.max_npkt = min(nitems(txp->mb),
4650 		    sc->params.max_pkts_per_eth_tx_pkts_wr);
4651 		if (vi->flags & TX_USES_VM_WR && !(sc->flags & IS_VF))
4652 			txq->txp.max_npkt--;
4653 
4654 		if (vi->flags & TX_USES_VM_WR)
4655 			txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
4656 			    V_TXPKT_INTF(pi->tx_chan));
4657 		else
4658 			txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
4659 			    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(sc->pf) |
4660 			    V_TXPKT_VF(vi->vin) | V_TXPKT_VF_VLD(vi->vfvld));
4661 
4662 		txq->tc_idx = -1;
4663 	}
4664 
4665 	return (0);
4666 }
4667 
4668 /*
4669  * Idempotent.
4670  */
4671 static void
4672 free_txq(struct vi_info *vi, struct sge_txq *txq)
4673 {
4674 	struct adapter *sc = vi->adapter;
4675 	struct sge_eq *eq = &txq->eq;
4676 
4677 	if (eq->flags & EQ_HW_ALLOCATED) {
4678 		MPASS(eq->flags & EQ_SW_ALLOCATED);
4679 		free_eq_hwq(sc, NULL, eq);
4680 		MPASS(!(eq->flags & EQ_HW_ALLOCATED));
4681 	}
4682 
4683 	if (eq->flags & EQ_SW_ALLOCATED) {
4684 		MPASS(!(eq->flags & EQ_HW_ALLOCATED));
4685 		sglist_free(txq->gl);
4686 		free(txq->sdesc, M_CXGBE);
4687 		mp_ring_free(txq->r);
4688 		free_eq(sc, eq);
4689 		MPASS(!(eq->flags & EQ_SW_ALLOCATED));
4690 		bzero(txq, sizeof(*txq));
4691 	}
4692 }
4693 
4694 static void
4695 add_txq_sysctls(struct vi_info *vi, struct sysctl_ctx_list *ctx,
4696     struct sysctl_oid *oid, struct sge_txq *txq)
4697 {
4698 	struct adapter *sc;
4699 	struct sysctl_oid_list *children;
4700 
4701 	if (ctx == NULL || oid == NULL)
4702 		return;
4703 
4704 	sc = vi->adapter;
4705 	children = SYSCTL_CHILDREN(oid);
4706 
4707 	mp_ring_sysctls(txq->r, ctx, children);
4708 
4709 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "tc",
4710 	    CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, vi, txq - sc->sge.txq,
4711 	    sysctl_tc, "I", "traffic class (-1 means none)");
4712 
4713 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
4714 	    &txq->txcsum, "# of times hardware assisted with checksum");
4715 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "vlan_insertion", CTLFLAG_RD,
4716 	    &txq->vlan_insertion, "# of times hardware inserted 802.1Q tag");
4717 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
4718 	    &txq->tso_wrs, "# of TSO work requests");
4719 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
4720 	    &txq->imm_wrs, "# of work requests with immediate data");
4721 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
4722 	    &txq->sgl_wrs, "# of work requests with direct SGL");
4723 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
4724 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
4725 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txpkts0_wrs", CTLFLAG_RD,
4726 	    &txq->txpkts0_wrs, "# of txpkts (type 0) work requests");
4727 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txpkts1_wrs", CTLFLAG_RD,
4728 	    &txq->txpkts1_wrs, "# of txpkts (type 1) work requests");
4729 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txpkts0_pkts", CTLFLAG_RD,
4730 	    &txq->txpkts0_pkts,
4731 	    "# of frames tx'd using type0 txpkts work requests");
4732 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txpkts1_pkts", CTLFLAG_RD,
4733 	    &txq->txpkts1_pkts,
4734 	    "# of frames tx'd using type1 txpkts work requests");
4735 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txpkts_flush", CTLFLAG_RD,
4736 	    &txq->txpkts_flush,
4737 	    "# of times txpkts had to be flushed out by an egress-update");
4738 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "raw_wrs", CTLFLAG_RD,
4739 	    &txq->raw_wrs, "# of raw work requests (non-packets)");
4740 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "vxlan_tso_wrs", CTLFLAG_RD,
4741 	    &txq->vxlan_tso_wrs, "# of VXLAN TSO work requests");
4742 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "vxlan_txcsum", CTLFLAG_RD,
4743 	    &txq->vxlan_txcsum,
4744 	    "# of times hardware assisted with inner checksums (VXLAN)");
4745 
4746 #ifdef KERN_TLS
4747 	if (is_ktls(sc)) {
4748 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_records",
4749 		    CTLFLAG_RD, &txq->kern_tls_records,
4750 		    "# of NIC TLS records transmitted");
4751 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_short",
4752 		    CTLFLAG_RD, &txq->kern_tls_short,
4753 		    "# of short NIC TLS records transmitted");
4754 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_partial",
4755 		    CTLFLAG_RD, &txq->kern_tls_partial,
4756 		    "# of partial NIC TLS records transmitted");
4757 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_full",
4758 		    CTLFLAG_RD, &txq->kern_tls_full,
4759 		    "# of full NIC TLS records transmitted");
4760 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_octets",
4761 		    CTLFLAG_RD, &txq->kern_tls_octets,
4762 		    "# of payload octets in transmitted NIC TLS records");
4763 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_waste",
4764 		    CTLFLAG_RD, &txq->kern_tls_waste,
4765 		    "# of octets DMAd but not transmitted in NIC TLS records");
4766 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_options",
4767 		    CTLFLAG_RD, &txq->kern_tls_options,
4768 		    "# of NIC TLS options-only packets transmitted");
4769 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_header",
4770 		    CTLFLAG_RD, &txq->kern_tls_header,
4771 		    "# of NIC TLS header-only packets transmitted");
4772 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_fin",
4773 		    CTLFLAG_RD, &txq->kern_tls_fin,
4774 		    "# of NIC TLS FIN-only packets transmitted");
4775 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_fin_short",
4776 		    CTLFLAG_RD, &txq->kern_tls_fin_short,
4777 		    "# of NIC TLS padded FIN packets on short TLS records");
4778 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_cbc",
4779 		    CTLFLAG_RD, &txq->kern_tls_cbc,
4780 		    "# of NIC TLS sessions using AES-CBC");
4781 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_gcm",
4782 		    CTLFLAG_RD, &txq->kern_tls_gcm,
4783 		    "# of NIC TLS sessions using AES-GCM");
4784 	}
4785 #endif
4786 }
4787 
4788 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
4789 /*
4790  * Idempotent.
4791  */
4792 static int
4793 alloc_ofld_txq(struct vi_info *vi, struct sge_ofld_txq *ofld_txq, int idx)
4794 {
4795 	struct sysctl_oid *oid;
4796 	struct port_info *pi = vi->pi;
4797 	struct adapter *sc = vi->adapter;
4798 	struct sge_eq *eq = &ofld_txq->wrq.eq;
4799 	int rc, iqidx;
4800 	char name[16];
4801 
4802 	MPASS(idx >= 0);
4803 	MPASS(idx < vi->nofldtxq);
4804 
4805 	if (!(eq->flags & EQ_SW_ALLOCATED)) {
4806 		snprintf(name, sizeof(name), "%d", idx);
4807 		oid = SYSCTL_ADD_NODE(&vi->ctx,
4808 		    SYSCTL_CHILDREN(vi->ofld_txq_oid), OID_AUTO, name,
4809 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "offload tx queue");
4810 
4811 		snprintf(name, sizeof(name), "%s ofld_txq%d",
4812 		    device_get_nameunit(vi->dev), idx);
4813 		if (vi->nofldrxq > 0) {
4814 			iqidx = vi->first_ofld_rxq + (idx % vi->nofldrxq);
4815 			init_eq(sc, eq, EQ_OFLD, vi->qsize_txq, pi->tx_chan,
4816 			    &sc->sge.ofld_rxq[iqidx].iq, name);
4817 		} else {
4818 			iqidx = vi->first_rxq + (idx % vi->nrxq);
4819 			init_eq(sc, eq, EQ_OFLD, vi->qsize_txq, pi->tx_chan,
4820 			    &sc->sge.rxq[iqidx].iq, name);
4821 		}
4822 
4823 		rc = alloc_wrq(sc, vi, &ofld_txq->wrq, &vi->ctx, oid);
4824 		if (rc != 0) {
4825 			CH_ERR(vi, "failed to allocate ofld_txq%d: %d\n", idx,
4826 			    rc);
4827 			sysctl_remove_oid(oid, 1, 1);
4828 			return (rc);
4829 		}
4830 		MPASS(eq->flags & EQ_SW_ALLOCATED);
4831 		/* Can't fail after this point. */
4832 
4833 		ofld_txq->tx_iscsi_pdus = counter_u64_alloc(M_WAITOK);
4834 		ofld_txq->tx_iscsi_octets = counter_u64_alloc(M_WAITOK);
4835 		ofld_txq->tx_iscsi_iso_wrs = counter_u64_alloc(M_WAITOK);
4836 		ofld_txq->tx_aio_jobs = counter_u64_alloc(M_WAITOK);
4837 		ofld_txq->tx_aio_octets = counter_u64_alloc(M_WAITOK);
4838 		ofld_txq->tx_toe_tls_records = counter_u64_alloc(M_WAITOK);
4839 		ofld_txq->tx_toe_tls_octets = counter_u64_alloc(M_WAITOK);
4840 		add_ofld_txq_sysctls(&vi->ctx, oid, ofld_txq);
4841 	}
4842 
4843 	if (!(eq->flags & EQ_HW_ALLOCATED)) {
4844 		rc = alloc_eq_hwq(sc, vi, eq);
4845 		if (rc != 0) {
4846 			CH_ERR(vi, "failed to create hw ofld_txq%d: %d\n", idx,
4847 			    rc);
4848 			return (rc);
4849 		}
4850 		MPASS(eq->flags & EQ_HW_ALLOCATED);
4851 	}
4852 
4853 	return (0);
4854 }
4855 
4856 /*
4857  * Idempotent.
4858  */
4859 static void
4860 free_ofld_txq(struct vi_info *vi, struct sge_ofld_txq *ofld_txq)
4861 {
4862 	struct adapter *sc = vi->adapter;
4863 	struct sge_eq *eq = &ofld_txq->wrq.eq;
4864 
4865 	if (eq->flags & EQ_HW_ALLOCATED) {
4866 		MPASS(eq->flags & EQ_SW_ALLOCATED);
4867 		free_eq_hwq(sc, NULL, eq);
4868 		MPASS(!(eq->flags & EQ_HW_ALLOCATED));
4869 	}
4870 
4871 	if (eq->flags & EQ_SW_ALLOCATED) {
4872 		MPASS(!(eq->flags & EQ_HW_ALLOCATED));
4873 		counter_u64_free(ofld_txq->tx_iscsi_pdus);
4874 		counter_u64_free(ofld_txq->tx_iscsi_octets);
4875 		counter_u64_free(ofld_txq->tx_iscsi_iso_wrs);
4876 		counter_u64_free(ofld_txq->tx_aio_jobs);
4877 		counter_u64_free(ofld_txq->tx_aio_octets);
4878 		counter_u64_free(ofld_txq->tx_toe_tls_records);
4879 		counter_u64_free(ofld_txq->tx_toe_tls_octets);
4880 		free_wrq(sc, &ofld_txq->wrq);
4881 		MPASS(!(eq->flags & EQ_SW_ALLOCATED));
4882 		bzero(ofld_txq, sizeof(*ofld_txq));
4883 	}
4884 }
4885 
4886 static void
4887 add_ofld_txq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
4888     struct sge_ofld_txq *ofld_txq)
4889 {
4890 	struct sysctl_oid_list *children;
4891 
4892 	if (ctx == NULL || oid == NULL)
4893 		return;
4894 
4895 	children = SYSCTL_CHILDREN(oid);
4896 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_iscsi_pdus",
4897 	    CTLFLAG_RD, &ofld_txq->tx_iscsi_pdus,
4898 	    "# of iSCSI PDUs transmitted");
4899 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_iscsi_octets",
4900 	    CTLFLAG_RD, &ofld_txq->tx_iscsi_octets,
4901 	    "# of payload octets in transmitted iSCSI PDUs");
4902 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_iscsi_iso_wrs",
4903 	    CTLFLAG_RD, &ofld_txq->tx_iscsi_iso_wrs,
4904 	    "# of iSCSI segmentation offload work requests");
4905 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_aio_jobs",
4906 	    CTLFLAG_RD, &ofld_txq->tx_aio_jobs,
4907 	    "# of zero-copy aio_write(2) jobs transmitted");
4908 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_aio_octets",
4909 	    CTLFLAG_RD, &ofld_txq->tx_aio_octets,
4910 	    "# of payload octets in transmitted zero-copy aio_write(2) jobs");
4911 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_toe_tls_records",
4912 	    CTLFLAG_RD, &ofld_txq->tx_toe_tls_records,
4913 	    "# of TOE TLS records transmitted");
4914 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_toe_tls_octets",
4915 	    CTLFLAG_RD, &ofld_txq->tx_toe_tls_octets,
4916 	    "# of payload octets in transmitted TOE TLS records");
4917 }
4918 #endif
4919 
4920 static void
4921 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
4922 {
4923 	bus_addr_t *ba = arg;
4924 
4925 	KASSERT(nseg == 1,
4926 	    ("%s meant for single segment mappings only.", __func__));
4927 
4928 	*ba = error ? 0 : segs->ds_addr;
4929 }
4930 
4931 static inline void
4932 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
4933 {
4934 	uint32_t n, v;
4935 
4936 	n = IDXDIFF(fl->pidx >> 3, fl->dbidx, fl->sidx);
4937 	MPASS(n > 0);
4938 
4939 	wmb();
4940 	v = fl->dbval | V_PIDX(n);
4941 	if (fl->udb)
4942 		*fl->udb = htole32(v);
4943 	else
4944 		t4_write_reg(sc, sc->sge_kdoorbell_reg, v);
4945 	IDXINCR(fl->dbidx, n, fl->sidx);
4946 }
4947 
4948 /*
4949  * Fills up the freelist by allocating up to 'n' buffers.  Buffers that are
4950  * recycled do not count towards this allocation budget.
4951  *
4952  * Returns non-zero to indicate that this freelist should be added to the list
4953  * of starving freelists.
4954  */
4955 static int
4956 refill_fl(struct adapter *sc, struct sge_fl *fl, int n)
4957 {
4958 	__be64 *d;
4959 	struct fl_sdesc *sd;
4960 	uintptr_t pa;
4961 	caddr_t cl;
4962 	struct rx_buf_info *rxb;
4963 	struct cluster_metadata *clm;
4964 	uint16_t max_pidx, zidx = fl->zidx;
4965 	uint16_t hw_cidx = fl->hw_cidx;		/* stable snapshot */
4966 
4967 	FL_LOCK_ASSERT_OWNED(fl);
4968 
4969 	/*
4970 	 * We always stop at the beginning of the hardware descriptor that's just
4971 	 * before the one with the hw cidx.  This is to avoid hw pidx = hw cidx,
4972 	 * which would mean an empty freelist to the chip.
4973 	 */
4974 	max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1;
4975 	if (fl->pidx == max_pidx * 8)
4976 		return (0);
4977 
4978 	d = &fl->desc[fl->pidx];
4979 	sd = &fl->sdesc[fl->pidx];
4980 	rxb = &sc->sge.rx_buf_info[zidx];
4981 
4982 	while (n > 0) {
4983 
4984 		if (sd->cl != NULL) {
4985 
4986 			if (sd->nmbuf == 0) {
4987 				/*
4988 				 * Fast recycle without involving any atomics on
4989 				 * the cluster's metadata (if the cluster has
4990 				 * metadata).  This happens when all frames
4991 				 * received in the cluster were small enough to
4992 				 * fit within a single mbuf each.
4993 				 */
4994 				fl->cl_fast_recycled++;
4995 				goto recycled;
4996 			}
4997 
4998 			/*
4999 			 * Cluster is guaranteed to have metadata.  Clusters
5000 			 * without metadata always take the fast recycle path
5001 			 * when they're recycled.
5002 			 */
5003 			clm = cl_metadata(sd);
5004 			MPASS(clm != NULL);
5005 
5006 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
5007 				fl->cl_recycled++;
5008 				counter_u64_add(extfree_rels, 1);
5009 				goto recycled;
5010 			}
5011 			sd->cl = NULL;	/* gave up my reference */
5012 		}
5013 		MPASS(sd->cl == NULL);
5014 		cl = uma_zalloc(rxb->zone, M_NOWAIT);
5015 		if (__predict_false(cl == NULL)) {
5016 			if (zidx != fl->safe_zidx) {
5017 				zidx = fl->safe_zidx;
5018 				rxb = &sc->sge.rx_buf_info[zidx];
5019 				cl = uma_zalloc(rxb->zone, M_NOWAIT);
5020 			}
5021 			if (cl == NULL)
5022 				break;
5023 		}
5024 		fl->cl_allocated++;
5025 		n--;
5026 
5027 		pa = pmap_kextract((vm_offset_t)cl);
5028 		sd->cl = cl;
5029 		sd->zidx = zidx;
5030 
5031 		if (fl->flags & FL_BUF_PACKING) {
5032 			*d = htobe64(pa | rxb->hwidx2);
5033 			sd->moff = rxb->size2;
5034 		} else {
5035 			*d = htobe64(pa | rxb->hwidx1);
5036 			sd->moff = 0;
5037 		}
5038 recycled:
5039 		sd->nmbuf = 0;
5040 		d++;
5041 		sd++;
5042 		if (__predict_false((++fl->pidx & 7) == 0)) {
5043 			uint16_t pidx = fl->pidx >> 3;
5044 
5045 			if (__predict_false(pidx == fl->sidx)) {
5046 				fl->pidx = 0;
5047 				pidx = 0;
5048 				sd = fl->sdesc;
5049 				d = fl->desc;
5050 			}
5051 			if (n < 8 || pidx == max_pidx)
5052 				break;
5053 
5054 			if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4)
5055 				ring_fl_db(sc, fl);
5056 		}
5057 	}
5058 
5059 	if ((fl->pidx >> 3) != fl->dbidx)
5060 		ring_fl_db(sc, fl);
5061 
5062 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
5063 }
5064 
5065 /*
5066  * Attempt to refill all starving freelists.
5067  */
5068 static void
5069 refill_sfl(void *arg)
5070 {
5071 	struct adapter *sc = arg;
5072 	struct sge_fl *fl, *fl_temp;
5073 
5074 	mtx_assert(&sc->sfl_lock, MA_OWNED);
5075 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
5076 		FL_LOCK(fl);
5077 		refill_fl(sc, fl, 64);
5078 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
5079 			TAILQ_REMOVE(&sc->sfl, fl, link);
5080 			fl->flags &= ~FL_STARVING;
5081 		}
5082 		FL_UNLOCK(fl);
5083 	}
5084 
5085 	if (!TAILQ_EMPTY(&sc->sfl))
5086 		callout_schedule(&sc->sfl_callout, hz / 5);
5087 }
5088 
5089 /*
5090  * Release the driver's reference on all buffers in the given freelist.  Buffers
5091  * with kernel references cannot be freed and will prevent the driver from being
5092  * unloaded safely.
5093  */
5094 void
5095 free_fl_buffers(struct adapter *sc, struct sge_fl *fl)
5096 {
5097 	struct fl_sdesc *sd;
5098 	struct cluster_metadata *clm;
5099 	int i;
5100 
5101 	sd = fl->sdesc;
5102 	for (i = 0; i < fl->sidx * 8; i++, sd++) {
5103 		if (sd->cl == NULL)
5104 			continue;
5105 
5106 		if (sd->nmbuf == 0)
5107 			uma_zfree(sc->sge.rx_buf_info[sd->zidx].zone, sd->cl);
5108 		else if (fl->flags & FL_BUF_PACKING) {
5109 			clm = cl_metadata(sd);
5110 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
5111 				uma_zfree(sc->sge.rx_buf_info[sd->zidx].zone,
5112 				    sd->cl);
5113 				counter_u64_add(extfree_rels, 1);
5114 			}
5115 		}
5116 		sd->cl = NULL;
5117 	}
5118 
5119 	if (fl->flags & FL_BUF_RESUME) {
5120 		m_freem(fl->m0);
5121 		fl->flags &= ~FL_BUF_RESUME;
5122 	}
5123 }
5124 
5125 static inline void
5126 get_pkt_gl(struct mbuf *m, struct sglist *gl)
5127 {
5128 	int rc;
5129 
5130 	M_ASSERTPKTHDR(m);
5131 
5132 	sglist_reset(gl);
5133 	rc = sglist_append_mbuf(gl, m);
5134 	if (__predict_false(rc != 0)) {
5135 		panic("%s: mbuf %p (%d segs) was vetted earlier but now fails "
5136 		    "with %d.", __func__, m, mbuf_nsegs(m), rc);
5137 	}
5138 
5139 	KASSERT(gl->sg_nseg == mbuf_nsegs(m),
5140 	    ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m,
5141 	    mbuf_nsegs(m), gl->sg_nseg));
5142 #if 0	/* vm_wr not readily available here. */
5143 	KASSERT(gl->sg_nseg > 0 && gl->sg_nseg <= max_nsegs_allowed(m, vm_wr),
5144 	    ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__,
5145 		gl->sg_nseg, max_nsegs_allowed(m, vm_wr)));
5146 #endif
5147 }
5148 
5149 /*
5150  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
5151  */
5152 static inline u_int
5153 txpkt_len16(u_int nsegs, const u_int extra)
5154 {
5155 	u_int n;
5156 
5157 	MPASS(nsegs > 0);
5158 
5159 	nsegs--; /* first segment is part of ulptx_sgl */
5160 	n = extra + sizeof(struct fw_eth_tx_pkt_wr) +
5161 	    sizeof(struct cpl_tx_pkt_core) +
5162 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
5163 
5164 	return (howmany(n, 16));
5165 }
5166 
5167 /*
5168  * len16 for a txpkt_vm WR with a GL.  Includes the firmware work
5169  * request header.
5170  */
5171 static inline u_int
5172 txpkt_vm_len16(u_int nsegs, const u_int extra)
5173 {
5174 	u_int n;
5175 
5176 	MPASS(nsegs > 0);
5177 
5178 	nsegs--; /* first segment is part of ulptx_sgl */
5179 	n = extra + sizeof(struct fw_eth_tx_pkt_vm_wr) +
5180 	    sizeof(struct cpl_tx_pkt_core) +
5181 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
5182 
5183 	return (howmany(n, 16));
5184 }
5185 
5186 static inline void
5187 calculate_mbuf_len16(struct mbuf *m, bool vm_wr)
5188 {
5189 	const int lso = sizeof(struct cpl_tx_pkt_lso_core);
5190 	const int tnl_lso = sizeof(struct cpl_tx_tnl_lso);
5191 
5192 	if (vm_wr) {
5193 		if (needs_tso(m))
5194 			set_mbuf_len16(m, txpkt_vm_len16(mbuf_nsegs(m), lso));
5195 		else
5196 			set_mbuf_len16(m, txpkt_vm_len16(mbuf_nsegs(m), 0));
5197 		return;
5198 	}
5199 
5200 	if (needs_tso(m)) {
5201 		if (needs_vxlan_tso(m))
5202 			set_mbuf_len16(m, txpkt_len16(mbuf_nsegs(m), tnl_lso));
5203 		else
5204 			set_mbuf_len16(m, txpkt_len16(mbuf_nsegs(m), lso));
5205 	} else
5206 		set_mbuf_len16(m, txpkt_len16(mbuf_nsegs(m), 0));
5207 }
5208 
5209 /*
5210  * len16 for a txpkts type 0 WR with a GL.  Does not include the firmware work
5211  * request header.
5212  */
5213 static inline u_int
5214 txpkts0_len16(u_int nsegs)
5215 {
5216 	u_int n;
5217 
5218 	MPASS(nsegs > 0);
5219 
5220 	nsegs--; /* first segment is part of ulptx_sgl */
5221 	n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) +
5222 	    sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) +
5223 	    8 * ((3 * nsegs) / 2 + (nsegs & 1));
5224 
5225 	return (howmany(n, 16));
5226 }
5227 
5228 /*
5229  * len16 for a txpkts type 1 WR with a GL.  Does not include the firmware work
5230  * request header.
5231  */
5232 static inline u_int
5233 txpkts1_len16(void)
5234 {
5235 	u_int n;
5236 
5237 	n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl);
5238 
5239 	return (howmany(n, 16));
5240 }
5241 
5242 static inline u_int
5243 imm_payload(u_int ndesc)
5244 {
5245 	u_int n;
5246 
5247 	n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) -
5248 	    sizeof(struct cpl_tx_pkt_core);
5249 
5250 	return (n);
5251 }
5252 
5253 static inline uint64_t
5254 csum_to_ctrl(struct adapter *sc, struct mbuf *m)
5255 {
5256 	uint64_t ctrl;
5257 	int csum_type, l2hlen, l3hlen;
5258 	int x, y;
5259 	static const int csum_types[3][2] = {
5260 		{TX_CSUM_TCPIP, TX_CSUM_TCPIP6},
5261 		{TX_CSUM_UDPIP, TX_CSUM_UDPIP6},
5262 		{TX_CSUM_IP, 0}
5263 	};
5264 
5265 	M_ASSERTPKTHDR(m);
5266 
5267 	if (!needs_hwcsum(m))
5268 		return (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS);
5269 
5270 	MPASS(m->m_pkthdr.l2hlen >= ETHER_HDR_LEN);
5271 	MPASS(m->m_pkthdr.l3hlen >= sizeof(struct ip));
5272 
5273 	if (needs_vxlan_csum(m)) {
5274 		MPASS(m->m_pkthdr.l4hlen > 0);
5275 		MPASS(m->m_pkthdr.l5hlen > 0);
5276 		MPASS(m->m_pkthdr.inner_l2hlen >= ETHER_HDR_LEN);
5277 		MPASS(m->m_pkthdr.inner_l3hlen >= sizeof(struct ip));
5278 
5279 		l2hlen = m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen +
5280 		    m->m_pkthdr.l4hlen + m->m_pkthdr.l5hlen +
5281 		    m->m_pkthdr.inner_l2hlen - ETHER_HDR_LEN;
5282 		l3hlen = m->m_pkthdr.inner_l3hlen;
5283 	} else {
5284 		l2hlen = m->m_pkthdr.l2hlen - ETHER_HDR_LEN;
5285 		l3hlen = m->m_pkthdr.l3hlen;
5286 	}
5287 
5288 	ctrl = 0;
5289 	if (!needs_l3_csum(m))
5290 		ctrl |= F_TXPKT_IPCSUM_DIS;
5291 
5292 	if (m->m_pkthdr.csum_flags & (CSUM_IP_TCP | CSUM_INNER_IP_TCP |
5293 	    CSUM_IP6_TCP | CSUM_INNER_IP6_TCP))
5294 		x = 0;	/* TCP */
5295 	else if (m->m_pkthdr.csum_flags & (CSUM_IP_UDP | CSUM_INNER_IP_UDP |
5296 	    CSUM_IP6_UDP | CSUM_INNER_IP6_UDP))
5297 		x = 1;	/* UDP */
5298 	else
5299 		x = 2;
5300 
5301 	if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_IP_TCP | CSUM_IP_UDP |
5302 	    CSUM_INNER_IP | CSUM_INNER_IP_TCP | CSUM_INNER_IP_UDP))
5303 		y = 0;	/* IPv4 */
5304 	else {
5305 		MPASS(m->m_pkthdr.csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP |
5306 		    CSUM_INNER_IP6_TCP | CSUM_INNER_IP6_UDP));
5307 		y = 1;	/* IPv6 */
5308 	}
5309 	/*
5310 	 * needs_hwcsum returned true earlier so there must be some kind of
5311 	 * checksum to calculate.
5312 	 */
5313 	csum_type = csum_types[x][y];
5314 	MPASS(csum_type != 0);
5315 	if (csum_type == TX_CSUM_IP)
5316 		ctrl |= F_TXPKT_L4CSUM_DIS;
5317 	ctrl |= V_TXPKT_CSUM_TYPE(csum_type) | V_TXPKT_IPHDR_LEN(l3hlen);
5318 	if (chip_id(sc) <= CHELSIO_T5)
5319 		ctrl |= V_TXPKT_ETHHDR_LEN(l2hlen);
5320 	else
5321 		ctrl |= V_T6_TXPKT_ETHHDR_LEN(l2hlen);
5322 
5323 	return (ctrl);
5324 }
5325 
5326 static inline void *
5327 write_lso_cpl(void *cpl, struct mbuf *m0)
5328 {
5329 	struct cpl_tx_pkt_lso_core *lso;
5330 	uint32_t ctrl;
5331 
5332 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
5333 	    m0->m_pkthdr.l4hlen > 0,
5334 	    ("%s: mbuf %p needs TSO but missing header lengths",
5335 		__func__, m0));
5336 
5337 	ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) |
5338 	    F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE |
5339 	    V_LSO_ETHHDR_LEN((m0->m_pkthdr.l2hlen - ETHER_HDR_LEN) >> 2) |
5340 	    V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) |
5341 	    V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
5342 	if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
5343 		ctrl |= F_LSO_IPV6;
5344 
5345 	lso = cpl;
5346 	lso->lso_ctrl = htobe32(ctrl);
5347 	lso->ipid_ofst = htobe16(0);
5348 	lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
5349 	lso->seqno_offset = htobe32(0);
5350 	lso->len = htobe32(m0->m_pkthdr.len);
5351 
5352 	return (lso + 1);
5353 }
5354 
5355 static void *
5356 write_tnl_lso_cpl(void *cpl, struct mbuf *m0)
5357 {
5358 	struct cpl_tx_tnl_lso *tnl_lso = cpl;
5359 	uint32_t ctrl;
5360 
5361 	KASSERT(m0->m_pkthdr.inner_l2hlen > 0 &&
5362 	    m0->m_pkthdr.inner_l3hlen > 0 && m0->m_pkthdr.inner_l4hlen > 0 &&
5363 	    m0->m_pkthdr.inner_l5hlen > 0,
5364 	    ("%s: mbuf %p needs VXLAN_TSO but missing inner header lengths",
5365 		__func__, m0));
5366 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
5367 	    m0->m_pkthdr.l4hlen > 0 && m0->m_pkthdr.l5hlen > 0,
5368 	    ("%s: mbuf %p needs VXLAN_TSO but missing outer header lengths",
5369 		__func__, m0));
5370 
5371 	/* Outer headers. */
5372 	ctrl = V_CPL_TX_TNL_LSO_OPCODE(CPL_TX_TNL_LSO) |
5373 	    F_CPL_TX_TNL_LSO_FIRST | F_CPL_TX_TNL_LSO_LAST |
5374 	    V_CPL_TX_TNL_LSO_ETHHDRLENOUT(
5375 		(m0->m_pkthdr.l2hlen - ETHER_HDR_LEN) >> 2) |
5376 	    V_CPL_TX_TNL_LSO_IPHDRLENOUT(m0->m_pkthdr.l3hlen >> 2) |
5377 	    F_CPL_TX_TNL_LSO_IPLENSETOUT;
5378 	if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
5379 		ctrl |= F_CPL_TX_TNL_LSO_IPV6OUT;
5380 	else {
5381 		ctrl |= F_CPL_TX_TNL_LSO_IPHDRCHKOUT |
5382 		    F_CPL_TX_TNL_LSO_IPIDINCOUT;
5383 	}
5384 	tnl_lso->op_to_IpIdSplitOut = htobe32(ctrl);
5385 	tnl_lso->IpIdOffsetOut = 0;
5386 	tnl_lso->UdpLenSetOut_to_TnlHdrLen =
5387 		htobe16(F_CPL_TX_TNL_LSO_UDPCHKCLROUT |
5388 		    F_CPL_TX_TNL_LSO_UDPLENSETOUT |
5389 		    V_CPL_TX_TNL_LSO_TNLHDRLEN(m0->m_pkthdr.l2hlen +
5390 			m0->m_pkthdr.l3hlen + m0->m_pkthdr.l4hlen +
5391 			m0->m_pkthdr.l5hlen) |
5392 		    V_CPL_TX_TNL_LSO_TNLTYPE(TX_TNL_TYPE_VXLAN));
5393 	tnl_lso->r1 = 0;
5394 
5395 	/* Inner headers. */
5396 	ctrl = V_CPL_TX_TNL_LSO_ETHHDRLEN(
5397 	    (m0->m_pkthdr.inner_l2hlen - ETHER_HDR_LEN) >> 2) |
5398 	    V_CPL_TX_TNL_LSO_IPHDRLEN(m0->m_pkthdr.inner_l3hlen >> 2) |
5399 	    V_CPL_TX_TNL_LSO_TCPHDRLEN(m0->m_pkthdr.inner_l4hlen >> 2);
5400 	if (m0->m_pkthdr.inner_l3hlen == sizeof(struct ip6_hdr))
5401 		ctrl |= F_CPL_TX_TNL_LSO_IPV6;
5402 	tnl_lso->Flow_to_TcpHdrLen = htobe32(ctrl);
5403 	tnl_lso->IpIdOffset = 0;
5404 	tnl_lso->IpIdSplit_to_Mss =
5405 	    htobe16(V_CPL_TX_TNL_LSO_MSS(m0->m_pkthdr.tso_segsz));
5406 	tnl_lso->TCPSeqOffset = 0;
5407 	tnl_lso->EthLenOffset_Size =
5408 	    htobe32(V_CPL_TX_TNL_LSO_SIZE(m0->m_pkthdr.len));
5409 
5410 	return (tnl_lso + 1);
5411 }
5412 
5413 #define VM_TX_L2HDR_LEN	16	/* ethmacdst to vlantci */
5414 
5415 /*
5416  * Write a VM txpkt WR for this packet to the hardware descriptors, update the
5417  * software descriptor, and advance the pidx.  It is guaranteed that enough
5418  * descriptors are available.
5419  *
5420  * The return value is the # of hardware descriptors used.
5421  */
5422 static u_int
5423 write_txpkt_vm_wr(struct adapter *sc, struct sge_txq *txq, struct mbuf *m0)
5424 {
5425 	struct sge_eq *eq;
5426 	struct fw_eth_tx_pkt_vm_wr *wr;
5427 	struct tx_sdesc *txsd;
5428 	struct cpl_tx_pkt_core *cpl;
5429 	uint32_t ctrl;	/* used in many unrelated places */
5430 	uint64_t ctrl1;
5431 	int len16, ndesc, pktlen;
5432 	caddr_t dst;
5433 
5434 	TXQ_LOCK_ASSERT_OWNED(txq);
5435 	M_ASSERTPKTHDR(m0);
5436 
5437 	len16 = mbuf_len16(m0);
5438 	pktlen = m0->m_pkthdr.len;
5439 	ctrl = sizeof(struct cpl_tx_pkt_core);
5440 	if (needs_tso(m0))
5441 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
5442 	ndesc = tx_len16_to_desc(len16);
5443 
5444 	/* Firmware work request header */
5445 	eq = &txq->eq;
5446 	wr = (void *)&eq->desc[eq->pidx];
5447 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_VM_WR) |
5448 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
5449 
5450 	ctrl = V_FW_WR_LEN16(len16);
5451 	wr->equiq_to_len16 = htobe32(ctrl);
5452 	wr->r3[0] = 0;
5453 	wr->r3[1] = 0;
5454 
5455 	/*
5456 	 * Copy over ethmacdst, ethmacsrc, ethtype, and vlantci.
5457 	 * vlantci is ignored unless the ethtype is 0x8100, so it's
5458 	 * simpler to always copy it rather than making it
5459 	 * conditional.  Also, it seems that we do not have to set
5460 	 * vlantci or fake the ethtype when doing VLAN tag insertion.
5461 	 */
5462 	m_copydata(m0, 0, VM_TX_L2HDR_LEN, wr->ethmacdst);
5463 
5464 	if (needs_tso(m0)) {
5465 		cpl = write_lso_cpl(wr + 1, m0);
5466 		txq->tso_wrs++;
5467 	} else
5468 		cpl = (void *)(wr + 1);
5469 
5470 	/* Checksum offload */
5471 	ctrl1 = csum_to_ctrl(sc, m0);
5472 	if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS))
5473 		txq->txcsum++;	/* some hardware assistance provided */
5474 
5475 	/* VLAN tag insertion */
5476 	if (needs_vlan_insertion(m0)) {
5477 		ctrl1 |= F_TXPKT_VLAN_VLD |
5478 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
5479 		txq->vlan_insertion++;
5480 	}
5481 
5482 	/* CPL header */
5483 	cpl->ctrl0 = txq->cpl_ctrl0;
5484 	cpl->pack = 0;
5485 	cpl->len = htobe16(pktlen);
5486 	cpl->ctrl1 = htobe64(ctrl1);
5487 
5488 	/* SGL */
5489 	dst = (void *)(cpl + 1);
5490 
5491 	/*
5492 	 * A packet using TSO will use up an entire descriptor for the
5493 	 * firmware work request header, LSO CPL, and TX_PKT_XT CPL.
5494 	 * If this descriptor is the last descriptor in the ring, wrap
5495 	 * around to the front of the ring explicitly for the start of
5496 	 * the sgl.
5497 	 */
5498 	if (dst == (void *)&eq->desc[eq->sidx]) {
5499 		dst = (void *)&eq->desc[0];
5500 		write_gl_to_txd(txq, m0, &dst, 0);
5501 	} else
5502 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
5503 	txq->sgl_wrs++;
5504 	txq->txpkt_wrs++;
5505 
5506 	txsd = &txq->sdesc[eq->pidx];
5507 	txsd->m = m0;
5508 	txsd->desc_used = ndesc;
5509 
5510 	return (ndesc);
5511 }
5512 
5513 /*
5514  * Write a raw WR to the hardware descriptors, update the software
5515  * descriptor, and advance the pidx.  It is guaranteed that enough
5516  * descriptors are available.
5517  *
5518  * The return value is the # of hardware descriptors used.
5519  */
5520 static u_int
5521 write_raw_wr(struct sge_txq *txq, void *wr, struct mbuf *m0, u_int available)
5522 {
5523 	struct sge_eq *eq = &txq->eq;
5524 	struct tx_sdesc *txsd;
5525 	struct mbuf *m;
5526 	caddr_t dst;
5527 	int len16, ndesc;
5528 
5529 	len16 = mbuf_len16(m0);
5530 	ndesc = tx_len16_to_desc(len16);
5531 	MPASS(ndesc <= available);
5532 
5533 	dst = wr;
5534 	for (m = m0; m != NULL; m = m->m_next)
5535 		copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
5536 
5537 	txq->raw_wrs++;
5538 
5539 	txsd = &txq->sdesc[eq->pidx];
5540 	txsd->m = m0;
5541 	txsd->desc_used = ndesc;
5542 
5543 	return (ndesc);
5544 }
5545 
5546 /*
5547  * Write a txpkt WR for this packet to the hardware descriptors, update the
5548  * software descriptor, and advance the pidx.  It is guaranteed that enough
5549  * descriptors are available.
5550  *
5551  * The return value is the # of hardware descriptors used.
5552  */
5553 static u_int
5554 write_txpkt_wr(struct adapter *sc, struct sge_txq *txq, struct mbuf *m0,
5555     u_int available)
5556 {
5557 	struct sge_eq *eq;
5558 	struct fw_eth_tx_pkt_wr *wr;
5559 	struct tx_sdesc *txsd;
5560 	struct cpl_tx_pkt_core *cpl;
5561 	uint32_t ctrl;	/* used in many unrelated places */
5562 	uint64_t ctrl1;
5563 	int len16, ndesc, pktlen, nsegs;
5564 	caddr_t dst;
5565 
5566 	TXQ_LOCK_ASSERT_OWNED(txq);
5567 	M_ASSERTPKTHDR(m0);
5568 
5569 	len16 = mbuf_len16(m0);
5570 	nsegs = mbuf_nsegs(m0);
5571 	pktlen = m0->m_pkthdr.len;
5572 	ctrl = sizeof(struct cpl_tx_pkt_core);
5573 	if (needs_tso(m0)) {
5574 		if (needs_vxlan_tso(m0))
5575 			ctrl += sizeof(struct cpl_tx_tnl_lso);
5576 		else
5577 			ctrl += sizeof(struct cpl_tx_pkt_lso_core);
5578 	} else if (!(mbuf_cflags(m0) & MC_NOMAP) && pktlen <= imm_payload(2) &&
5579 	    available >= 2) {
5580 		/* Immediate data.  Recalculate len16 and set nsegs to 0. */
5581 		ctrl += pktlen;
5582 		len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) +
5583 		    sizeof(struct cpl_tx_pkt_core) + pktlen, 16);
5584 		nsegs = 0;
5585 	}
5586 	ndesc = tx_len16_to_desc(len16);
5587 	MPASS(ndesc <= available);
5588 
5589 	/* Firmware work request header */
5590 	eq = &txq->eq;
5591 	wr = (void *)&eq->desc[eq->pidx];
5592 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
5593 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
5594 
5595 	ctrl = V_FW_WR_LEN16(len16);
5596 	wr->equiq_to_len16 = htobe32(ctrl);
5597 	wr->r3 = 0;
5598 
5599 	if (needs_tso(m0)) {
5600 		if (needs_vxlan_tso(m0)) {
5601 			cpl = write_tnl_lso_cpl(wr + 1, m0);
5602 			txq->vxlan_tso_wrs++;
5603 		} else {
5604 			cpl = write_lso_cpl(wr + 1, m0);
5605 			txq->tso_wrs++;
5606 		}
5607 	} else
5608 		cpl = (void *)(wr + 1);
5609 
5610 	/* Checksum offload */
5611 	ctrl1 = csum_to_ctrl(sc, m0);
5612 	if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS)) {
5613 		/* some hardware assistance provided */
5614 		if (needs_vxlan_csum(m0))
5615 			txq->vxlan_txcsum++;
5616 		else
5617 			txq->txcsum++;
5618 	}
5619 
5620 	/* VLAN tag insertion */
5621 	if (needs_vlan_insertion(m0)) {
5622 		ctrl1 |= F_TXPKT_VLAN_VLD |
5623 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
5624 		txq->vlan_insertion++;
5625 	}
5626 
5627 	/* CPL header */
5628 	cpl->ctrl0 = txq->cpl_ctrl0;
5629 	cpl->pack = 0;
5630 	cpl->len = htobe16(pktlen);
5631 	cpl->ctrl1 = htobe64(ctrl1);
5632 
5633 	/* SGL */
5634 	dst = (void *)(cpl + 1);
5635 	if (__predict_false((uintptr_t)dst == (uintptr_t)&eq->desc[eq->sidx]))
5636 		dst = (caddr_t)&eq->desc[0];
5637 	if (nsegs > 0) {
5638 
5639 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
5640 		txq->sgl_wrs++;
5641 	} else {
5642 		struct mbuf *m;
5643 
5644 		for (m = m0; m != NULL; m = m->m_next) {
5645 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
5646 #ifdef INVARIANTS
5647 			pktlen -= m->m_len;
5648 #endif
5649 		}
5650 #ifdef INVARIANTS
5651 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
5652 #endif
5653 		txq->imm_wrs++;
5654 	}
5655 
5656 	txq->txpkt_wrs++;
5657 
5658 	txsd = &txq->sdesc[eq->pidx];
5659 	txsd->m = m0;
5660 	txsd->desc_used = ndesc;
5661 
5662 	return (ndesc);
5663 }
5664 
5665 static inline bool
5666 cmp_l2hdr(struct txpkts *txp, struct mbuf *m)
5667 {
5668 	int len;
5669 
5670 	MPASS(txp->npkt > 0);
5671 	MPASS(m->m_len >= VM_TX_L2HDR_LEN);
5672 
5673 	if (txp->ethtype == be16toh(ETHERTYPE_VLAN))
5674 		len = VM_TX_L2HDR_LEN;
5675 	else
5676 		len = sizeof(struct ether_header);
5677 
5678 	return (memcmp(m->m_data, &txp->ethmacdst[0], len) != 0);
5679 }
5680 
5681 static inline void
5682 save_l2hdr(struct txpkts *txp, struct mbuf *m)
5683 {
5684 	MPASS(m->m_len >= VM_TX_L2HDR_LEN);
5685 
5686 	memcpy(&txp->ethmacdst[0], mtod(m, const void *), VM_TX_L2HDR_LEN);
5687 }
5688 
5689 static int
5690 add_to_txpkts_vf(struct adapter *sc, struct sge_txq *txq, struct mbuf *m,
5691     int avail, bool *send)
5692 {
5693 	struct txpkts *txp = &txq->txp;
5694 
5695 	/* Cannot have TSO and coalesce at the same time. */
5696 	if (cannot_use_txpkts(m)) {
5697 cannot_coalesce:
5698 		*send = txp->npkt > 0;
5699 		return (EINVAL);
5700 	}
5701 
5702 	/* VF allows coalescing of type 1 (1 GL) only */
5703 	if (mbuf_nsegs(m) > 1)
5704 		goto cannot_coalesce;
5705 
5706 	*send = false;
5707 	if (txp->npkt > 0) {
5708 		MPASS(tx_len16_to_desc(txp->len16) <= avail);
5709 		MPASS(txp->npkt < txp->max_npkt);
5710 		MPASS(txp->wr_type == 1);	/* VF supports type 1 only */
5711 
5712 		if (tx_len16_to_desc(txp->len16 + txpkts1_len16()) > avail) {
5713 retry_after_send:
5714 			*send = true;
5715 			return (EAGAIN);
5716 		}
5717 		if (m->m_pkthdr.len + txp->plen > 65535)
5718 			goto retry_after_send;
5719 		if (cmp_l2hdr(txp, m))
5720 			goto retry_after_send;
5721 
5722 		txp->len16 += txpkts1_len16();
5723 		txp->plen += m->m_pkthdr.len;
5724 		txp->mb[txp->npkt++] = m;
5725 		if (txp->npkt == txp->max_npkt)
5726 			*send = true;
5727 	} else {
5728 		txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_vm_wr), 16) +
5729 		    txpkts1_len16();
5730 		if (tx_len16_to_desc(txp->len16) > avail)
5731 			goto cannot_coalesce;
5732 		txp->npkt = 1;
5733 		txp->wr_type = 1;
5734 		txp->plen = m->m_pkthdr.len;
5735 		txp->mb[0] = m;
5736 		save_l2hdr(txp, m);
5737 	}
5738 	return (0);
5739 }
5740 
5741 static int
5742 add_to_txpkts_pf(struct adapter *sc, struct sge_txq *txq, struct mbuf *m,
5743     int avail, bool *send)
5744 {
5745 	struct txpkts *txp = &txq->txp;
5746 	int nsegs;
5747 
5748 	MPASS(!(sc->flags & IS_VF));
5749 
5750 	/* Cannot have TSO and coalesce at the same time. */
5751 	if (cannot_use_txpkts(m)) {
5752 cannot_coalesce:
5753 		*send = txp->npkt > 0;
5754 		return (EINVAL);
5755 	}
5756 
5757 	*send = false;
5758 	nsegs = mbuf_nsegs(m);
5759 	if (txp->npkt == 0) {
5760 		if (m->m_pkthdr.len > 65535)
5761 			goto cannot_coalesce;
5762 		if (nsegs > 1) {
5763 			txp->wr_type = 0;
5764 			txp->len16 =
5765 			    howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) +
5766 			    txpkts0_len16(nsegs);
5767 		} else {
5768 			txp->wr_type = 1;
5769 			txp->len16 =
5770 			    howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) +
5771 			    txpkts1_len16();
5772 		}
5773 		if (tx_len16_to_desc(txp->len16) > avail)
5774 			goto cannot_coalesce;
5775 		txp->npkt = 1;
5776 		txp->plen = m->m_pkthdr.len;
5777 		txp->mb[0] = m;
5778 	} else {
5779 		MPASS(tx_len16_to_desc(txp->len16) <= avail);
5780 		MPASS(txp->npkt < txp->max_npkt);
5781 
5782 		if (m->m_pkthdr.len + txp->plen > 65535) {
5783 retry_after_send:
5784 			*send = true;
5785 			return (EAGAIN);
5786 		}
5787 
5788 		MPASS(txp->wr_type == 0 || txp->wr_type == 1);
5789 		if (txp->wr_type == 0) {
5790 			if (tx_len16_to_desc(txp->len16 +
5791 			    txpkts0_len16(nsegs)) > min(avail, SGE_MAX_WR_NDESC))
5792 				goto retry_after_send;
5793 			txp->len16 += txpkts0_len16(nsegs);
5794 		} else {
5795 			if (nsegs != 1)
5796 				goto retry_after_send;
5797 			if (tx_len16_to_desc(txp->len16 + txpkts1_len16()) >
5798 			    avail)
5799 				goto retry_after_send;
5800 			txp->len16 += txpkts1_len16();
5801 		}
5802 
5803 		txp->plen += m->m_pkthdr.len;
5804 		txp->mb[txp->npkt++] = m;
5805 		if (txp->npkt == txp->max_npkt)
5806 			*send = true;
5807 	}
5808 	return (0);
5809 }
5810 
5811 /*
5812  * Write a txpkts WR for the packets in txp to the hardware descriptors, update
5813  * the software descriptor, and advance the pidx.  It is guaranteed that enough
5814  * descriptors are available.
5815  *
5816  * The return value is the # of hardware descriptors used.
5817  */
5818 static u_int
5819 write_txpkts_wr(struct adapter *sc, struct sge_txq *txq)
5820 {
5821 	const struct txpkts *txp = &txq->txp;
5822 	struct sge_eq *eq = &txq->eq;
5823 	struct fw_eth_tx_pkts_wr *wr;
5824 	struct tx_sdesc *txsd;
5825 	struct cpl_tx_pkt_core *cpl;
5826 	uint64_t ctrl1;
5827 	int ndesc, i, checkwrap;
5828 	struct mbuf *m, *last;
5829 	void *flitp;
5830 
5831 	TXQ_LOCK_ASSERT_OWNED(txq);
5832 	MPASS(txp->npkt > 0);
5833 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
5834 
5835 	wr = (void *)&eq->desc[eq->pidx];
5836 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
5837 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(txp->len16));
5838 	wr->plen = htobe16(txp->plen);
5839 	wr->npkt = txp->npkt;
5840 	wr->r3 = 0;
5841 	wr->type = txp->wr_type;
5842 	flitp = wr + 1;
5843 
5844 	/*
5845 	 * At this point we are 16B into a hardware descriptor.  If checkwrap is
5846 	 * set then we know the WR is going to wrap around somewhere.  We'll
5847 	 * check for that at appropriate points.
5848 	 */
5849 	ndesc = tx_len16_to_desc(txp->len16);
5850 	last = NULL;
5851 	checkwrap = eq->sidx - ndesc < eq->pidx;
5852 	for (i = 0; i < txp->npkt; i++) {
5853 		m = txp->mb[i];
5854 		if (txp->wr_type == 0) {
5855 			struct ulp_txpkt *ulpmc;
5856 			struct ulptx_idata *ulpsc;
5857 
5858 			/* ULP master command */
5859 			ulpmc = flitp;
5860 			ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
5861 			    V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid));
5862 			ulpmc->len = htobe32(txpkts0_len16(mbuf_nsegs(m)));
5863 
5864 			/* ULP subcommand */
5865 			ulpsc = (void *)(ulpmc + 1);
5866 			ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
5867 			    F_ULP_TX_SC_MORE);
5868 			ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
5869 
5870 			cpl = (void *)(ulpsc + 1);
5871 			if (checkwrap &&
5872 			    (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx])
5873 				cpl = (void *)&eq->desc[0];
5874 		} else {
5875 			cpl = flitp;
5876 		}
5877 
5878 		/* Checksum offload */
5879 		ctrl1 = csum_to_ctrl(sc, m);
5880 		if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS)) {
5881 			/* some hardware assistance provided */
5882 			if (needs_vxlan_csum(m))
5883 				txq->vxlan_txcsum++;
5884 			else
5885 				txq->txcsum++;
5886 		}
5887 
5888 		/* VLAN tag insertion */
5889 		if (needs_vlan_insertion(m)) {
5890 			ctrl1 |= F_TXPKT_VLAN_VLD |
5891 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
5892 			txq->vlan_insertion++;
5893 		}
5894 
5895 		/* CPL header */
5896 		cpl->ctrl0 = txq->cpl_ctrl0;
5897 		cpl->pack = 0;
5898 		cpl->len = htobe16(m->m_pkthdr.len);
5899 		cpl->ctrl1 = htobe64(ctrl1);
5900 
5901 		flitp = cpl + 1;
5902 		if (checkwrap &&
5903 		    (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
5904 			flitp = (void *)&eq->desc[0];
5905 
5906 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap);
5907 
5908 		if (last != NULL)
5909 			last->m_nextpkt = m;
5910 		last = m;
5911 	}
5912 
5913 	txq->sgl_wrs++;
5914 	if (txp->wr_type == 0) {
5915 		txq->txpkts0_pkts += txp->npkt;
5916 		txq->txpkts0_wrs++;
5917 	} else {
5918 		txq->txpkts1_pkts += txp->npkt;
5919 		txq->txpkts1_wrs++;
5920 	}
5921 
5922 	txsd = &txq->sdesc[eq->pidx];
5923 	txsd->m = txp->mb[0];
5924 	txsd->desc_used = ndesc;
5925 
5926 	return (ndesc);
5927 }
5928 
5929 static u_int
5930 write_txpkts_vm_wr(struct adapter *sc, struct sge_txq *txq)
5931 {
5932 	const struct txpkts *txp = &txq->txp;
5933 	struct sge_eq *eq = &txq->eq;
5934 	struct fw_eth_tx_pkts_vm_wr *wr;
5935 	struct tx_sdesc *txsd;
5936 	struct cpl_tx_pkt_core *cpl;
5937 	uint64_t ctrl1;
5938 	int ndesc, i;
5939 	struct mbuf *m, *last;
5940 	void *flitp;
5941 
5942 	TXQ_LOCK_ASSERT_OWNED(txq);
5943 	MPASS(txp->npkt > 0);
5944 	MPASS(txp->wr_type == 1);	/* VF supports type 1 only */
5945 	MPASS(txp->mb[0] != NULL);
5946 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
5947 
5948 	wr = (void *)&eq->desc[eq->pidx];
5949 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_VM_WR));
5950 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(txp->len16));
5951 	wr->r3 = 0;
5952 	wr->plen = htobe16(txp->plen);
5953 	wr->npkt = txp->npkt;
5954 	wr->r4 = 0;
5955 	memcpy(&wr->ethmacdst[0], &txp->ethmacdst[0], 16);
5956 	flitp = wr + 1;
5957 
5958 	/*
5959 	 * At this point we are 32B into a hardware descriptor.  Each mbuf in
5960 	 * the WR will take 32B so we check for the end of the descriptor ring
5961 	 * before writing odd mbufs (mb[1], 3, 5, ..)
5962 	 */
5963 	ndesc = tx_len16_to_desc(txp->len16);
5964 	last = NULL;
5965 	for (i = 0; i < txp->npkt; i++) {
5966 		m = txp->mb[i];
5967 		if (i & 1 && (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
5968 			flitp = &eq->desc[0];
5969 		cpl = flitp;
5970 
5971 		/* Checksum offload */
5972 		ctrl1 = csum_to_ctrl(sc, m);
5973 		if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS))
5974 			txq->txcsum++;	/* some hardware assistance provided */
5975 
5976 		/* VLAN tag insertion */
5977 		if (needs_vlan_insertion(m)) {
5978 			ctrl1 |= F_TXPKT_VLAN_VLD |
5979 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
5980 			txq->vlan_insertion++;
5981 		}
5982 
5983 		/* CPL header */
5984 		cpl->ctrl0 = txq->cpl_ctrl0;
5985 		cpl->pack = 0;
5986 		cpl->len = htobe16(m->m_pkthdr.len);
5987 		cpl->ctrl1 = htobe64(ctrl1);
5988 
5989 		flitp = cpl + 1;
5990 		MPASS(mbuf_nsegs(m) == 1);
5991 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), 0);
5992 
5993 		if (last != NULL)
5994 			last->m_nextpkt = m;
5995 		last = m;
5996 	}
5997 
5998 	txq->sgl_wrs++;
5999 	txq->txpkts1_pkts += txp->npkt;
6000 	txq->txpkts1_wrs++;
6001 
6002 	txsd = &txq->sdesc[eq->pidx];
6003 	txsd->m = txp->mb[0];
6004 	txsd->desc_used = ndesc;
6005 
6006 	return (ndesc);
6007 }
6008 
6009 /*
6010  * If the SGL ends on an address that is not 16 byte aligned, this function will
6011  * add a 0 filled flit at the end.
6012  */
6013 static void
6014 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap)
6015 {
6016 	struct sge_eq *eq = &txq->eq;
6017 	struct sglist *gl = txq->gl;
6018 	struct sglist_seg *seg;
6019 	__be64 *flitp, *wrap;
6020 	struct ulptx_sgl *usgl;
6021 	int i, nflits, nsegs;
6022 
6023 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
6024 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
6025 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
6026 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
6027 
6028 	get_pkt_gl(m, gl);
6029 	nsegs = gl->sg_nseg;
6030 	MPASS(nsegs > 0);
6031 
6032 	nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
6033 	flitp = (__be64 *)(*to);
6034 	wrap = (__be64 *)(&eq->desc[eq->sidx]);
6035 	seg = &gl->sg_segs[0];
6036 	usgl = (void *)flitp;
6037 
6038 	/*
6039 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
6040 	 * ring, so we're at least 16 bytes away from the status page.  There is
6041 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
6042 	 */
6043 
6044 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
6045 	    V_ULPTX_NSGE(nsegs));
6046 	usgl->len0 = htobe32(seg->ss_len);
6047 	usgl->addr0 = htobe64(seg->ss_paddr);
6048 	seg++;
6049 
6050 	if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) {
6051 
6052 		/* Won't wrap around at all */
6053 
6054 		for (i = 0; i < nsegs - 1; i++, seg++) {
6055 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
6056 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
6057 		}
6058 		if (i & 1)
6059 			usgl->sge[i / 2].len[1] = htobe32(0);
6060 		flitp += nflits;
6061 	} else {
6062 
6063 		/* Will wrap somewhere in the rest of the SGL */
6064 
6065 		/* 2 flits already written, write the rest flit by flit */
6066 		flitp = (void *)(usgl + 1);
6067 		for (i = 0; i < nflits - 2; i++) {
6068 			if (flitp == wrap)
6069 				flitp = (void *)eq->desc;
6070 			*flitp++ = get_flit(seg, nsegs - 1, i);
6071 		}
6072 	}
6073 
6074 	if (nflits & 1) {
6075 		MPASS(((uintptr_t)flitp) & 0xf);
6076 		*flitp++ = 0;
6077 	}
6078 
6079 	MPASS((((uintptr_t)flitp) & 0xf) == 0);
6080 	if (__predict_false(flitp == wrap))
6081 		*to = (void *)eq->desc;
6082 	else
6083 		*to = (void *)flitp;
6084 }
6085 
6086 static inline void
6087 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
6088 {
6089 
6090 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
6091 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
6092 
6093 	if (__predict_true((uintptr_t)(*to) + len <=
6094 	    (uintptr_t)&eq->desc[eq->sidx])) {
6095 		bcopy(from, *to, len);
6096 		(*to) += len;
6097 	} else {
6098 		int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
6099 
6100 		bcopy(from, *to, portion);
6101 		from += portion;
6102 		portion = len - portion;	/* remaining */
6103 		bcopy(from, (void *)eq->desc, portion);
6104 		(*to) = (caddr_t)eq->desc + portion;
6105 	}
6106 }
6107 
6108 static inline void
6109 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n)
6110 {
6111 	u_int db;
6112 
6113 	MPASS(n > 0);
6114 
6115 	db = eq->doorbells;
6116 	if (n > 1)
6117 		clrbit(&db, DOORBELL_WCWR);
6118 	wmb();
6119 
6120 	switch (ffs(db) - 1) {
6121 	case DOORBELL_UDB:
6122 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
6123 		break;
6124 
6125 	case DOORBELL_WCWR: {
6126 		volatile uint64_t *dst, *src;
6127 		int i;
6128 
6129 		/*
6130 		 * Queues whose 128B doorbell segment fits in the page do not
6131 		 * use relative qid (udb_qid is always 0).  Only queues with
6132 		 * doorbell segments can do WCWR.
6133 		 */
6134 		KASSERT(eq->udb_qid == 0 && n == 1,
6135 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
6136 		    __func__, eq->doorbells, n, eq->dbidx, eq));
6137 
6138 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
6139 		    UDBS_DB_OFFSET);
6140 		i = eq->dbidx;
6141 		src = (void *)&eq->desc[i];
6142 		while (src != (void *)&eq->desc[i + 1])
6143 			*dst++ = *src++;
6144 		wmb();
6145 		break;
6146 	}
6147 
6148 	case DOORBELL_UDBWC:
6149 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
6150 		wmb();
6151 		break;
6152 
6153 	case DOORBELL_KDB:
6154 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
6155 		    V_QID(eq->cntxt_id) | V_PIDX(n));
6156 		break;
6157 	}
6158 
6159 	IDXINCR(eq->dbidx, n, eq->sidx);
6160 }
6161 
6162 static inline u_int
6163 reclaimable_tx_desc(struct sge_eq *eq)
6164 {
6165 	uint16_t hw_cidx;
6166 
6167 	hw_cidx = read_hw_cidx(eq);
6168 	return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx));
6169 }
6170 
6171 static inline u_int
6172 total_available_tx_desc(struct sge_eq *eq)
6173 {
6174 	uint16_t hw_cidx, pidx;
6175 
6176 	hw_cidx = read_hw_cidx(eq);
6177 	pidx = eq->pidx;
6178 
6179 	if (pidx == hw_cidx)
6180 		return (eq->sidx - 1);
6181 	else
6182 		return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1);
6183 }
6184 
6185 static inline uint16_t
6186 read_hw_cidx(struct sge_eq *eq)
6187 {
6188 	struct sge_qstat *spg = (void *)&eq->desc[eq->sidx];
6189 	uint16_t cidx = spg->cidx;	/* stable snapshot */
6190 
6191 	return (be16toh(cidx));
6192 }
6193 
6194 /*
6195  * Reclaim 'n' descriptors approximately.
6196  */
6197 static u_int
6198 reclaim_tx_descs(struct sge_txq *txq, u_int n)
6199 {
6200 	struct tx_sdesc *txsd;
6201 	struct sge_eq *eq = &txq->eq;
6202 	u_int can_reclaim, reclaimed;
6203 
6204 	TXQ_LOCK_ASSERT_OWNED(txq);
6205 	MPASS(n > 0);
6206 
6207 	reclaimed = 0;
6208 	can_reclaim = reclaimable_tx_desc(eq);
6209 	while (can_reclaim && reclaimed < n) {
6210 		int ndesc;
6211 		struct mbuf *m, *nextpkt;
6212 
6213 		txsd = &txq->sdesc[eq->cidx];
6214 		ndesc = txsd->desc_used;
6215 
6216 		/* Firmware doesn't return "partial" credits. */
6217 		KASSERT(can_reclaim >= ndesc,
6218 		    ("%s: unexpected number of credits: %d, %d",
6219 		    __func__, can_reclaim, ndesc));
6220 		KASSERT(ndesc != 0,
6221 		    ("%s: descriptor with no credits: cidx %d",
6222 		    __func__, eq->cidx));
6223 
6224 		for (m = txsd->m; m != NULL; m = nextpkt) {
6225 			nextpkt = m->m_nextpkt;
6226 			m->m_nextpkt = NULL;
6227 			m_freem(m);
6228 		}
6229 		reclaimed += ndesc;
6230 		can_reclaim -= ndesc;
6231 		IDXINCR(eq->cidx, ndesc, eq->sidx);
6232 	}
6233 
6234 	return (reclaimed);
6235 }
6236 
6237 static void
6238 tx_reclaim(void *arg, int n)
6239 {
6240 	struct sge_txq *txq = arg;
6241 	struct sge_eq *eq = &txq->eq;
6242 
6243 	do {
6244 		if (TXQ_TRYLOCK(txq) == 0)
6245 			break;
6246 		n = reclaim_tx_descs(txq, 32);
6247 		if (eq->cidx == eq->pidx)
6248 			eq->equeqidx = eq->pidx;
6249 		TXQ_UNLOCK(txq);
6250 	} while (n > 0);
6251 }
6252 
6253 static __be64
6254 get_flit(struct sglist_seg *segs, int nsegs, int idx)
6255 {
6256 	int i = (idx / 3) * 2;
6257 
6258 	switch (idx % 3) {
6259 	case 0: {
6260 		uint64_t rc;
6261 
6262 		rc = (uint64_t)segs[i].ss_len << 32;
6263 		if (i + 1 < nsegs)
6264 			rc |= (uint64_t)(segs[i + 1].ss_len);
6265 
6266 		return (htobe64(rc));
6267 	}
6268 	case 1:
6269 		return (htobe64(segs[i].ss_paddr));
6270 	case 2:
6271 		return (htobe64(segs[i + 1].ss_paddr));
6272 	}
6273 
6274 	return (0);
6275 }
6276 
6277 static int
6278 find_refill_source(struct adapter *sc, int maxp, bool packing)
6279 {
6280 	int i, zidx = -1;
6281 	struct rx_buf_info *rxb = &sc->sge.rx_buf_info[0];
6282 
6283 	if (packing) {
6284 		for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) {
6285 			if (rxb->hwidx2 == -1)
6286 				continue;
6287 			if (rxb->size1 < PAGE_SIZE &&
6288 			    rxb->size1 < largest_rx_cluster)
6289 				continue;
6290 			if (rxb->size1 > largest_rx_cluster)
6291 				break;
6292 			MPASS(rxb->size1 - rxb->size2 >= CL_METADATA_SIZE);
6293 			if (rxb->size2 >= maxp)
6294 				return (i);
6295 			zidx = i;
6296 		}
6297 	} else {
6298 		for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) {
6299 			if (rxb->hwidx1 == -1)
6300 				continue;
6301 			if (rxb->size1 > largest_rx_cluster)
6302 				break;
6303 			if (rxb->size1 >= maxp)
6304 				return (i);
6305 			zidx = i;
6306 		}
6307 	}
6308 
6309 	return (zidx);
6310 }
6311 
6312 static void
6313 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
6314 {
6315 	mtx_lock(&sc->sfl_lock);
6316 	FL_LOCK(fl);
6317 	if ((fl->flags & FL_DOOMED) == 0) {
6318 		fl->flags |= FL_STARVING;
6319 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
6320 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
6321 	}
6322 	FL_UNLOCK(fl);
6323 	mtx_unlock(&sc->sfl_lock);
6324 }
6325 
6326 static void
6327 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq)
6328 {
6329 	struct sge_wrq *wrq = (void *)eq;
6330 
6331 	atomic_readandclear_int(&eq->equiq);
6332 	taskqueue_enqueue(sc->tq[eq->tx_chan], &wrq->wrq_tx_task);
6333 }
6334 
6335 static void
6336 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq)
6337 {
6338 	struct sge_txq *txq = (void *)eq;
6339 
6340 	MPASS(eq->type == EQ_ETH);
6341 
6342 	atomic_readandclear_int(&eq->equiq);
6343 	if (mp_ring_is_idle(txq->r))
6344 		taskqueue_enqueue(sc->tq[eq->tx_chan], &txq->tx_reclaim_task);
6345 	else
6346 		mp_ring_check_drainage(txq->r, 64);
6347 }
6348 
6349 static int
6350 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
6351     struct mbuf *m)
6352 {
6353 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
6354 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
6355 	struct adapter *sc = iq->adapter;
6356 	struct sge *s = &sc->sge;
6357 	struct sge_eq *eq;
6358 	static void (*h[])(struct adapter *, struct sge_eq *) = {NULL,
6359 		&handle_wrq_egr_update, &handle_eth_egr_update,
6360 		&handle_wrq_egr_update};
6361 
6362 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
6363 	    rss->opcode));
6364 
6365 	eq = s->eqmap[qid - s->eq_start - s->eq_base];
6366 	(*h[eq->type])(sc, eq);
6367 
6368 	return (0);
6369 }
6370 
6371 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
6372 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
6373     offsetof(struct cpl_fw6_msg, data));
6374 
6375 static int
6376 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
6377 {
6378 	struct adapter *sc = iq->adapter;
6379 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
6380 
6381 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
6382 	    rss->opcode));
6383 
6384 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
6385 		const struct rss_header *rss2;
6386 
6387 		rss2 = (const struct rss_header *)&cpl->data[0];
6388 		return (t4_cpl_handler[rss2->opcode](iq, rss2, m));
6389 	}
6390 
6391 	return (t4_fw_msg_handler[cpl->type](sc, &cpl->data[0]));
6392 }
6393 
6394 /**
6395  *	t4_handle_wrerr_rpl - process a FW work request error message
6396  *	@adap: the adapter
6397  *	@rpl: start of the FW message
6398  */
6399 static int
6400 t4_handle_wrerr_rpl(struct adapter *adap, const __be64 *rpl)
6401 {
6402 	u8 opcode = *(const u8 *)rpl;
6403 	const struct fw_error_cmd *e = (const void *)rpl;
6404 	unsigned int i;
6405 
6406 	if (opcode != FW_ERROR_CMD) {
6407 		log(LOG_ERR,
6408 		    "%s: Received WRERR_RPL message with opcode %#x\n",
6409 		    device_get_nameunit(adap->dev), opcode);
6410 		return (EINVAL);
6411 	}
6412 	log(LOG_ERR, "%s: FW_ERROR (%s) ", device_get_nameunit(adap->dev),
6413 	    G_FW_ERROR_CMD_FATAL(be32toh(e->op_to_type)) ? "fatal" :
6414 	    "non-fatal");
6415 	switch (G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type))) {
6416 	case FW_ERROR_TYPE_EXCEPTION:
6417 		log(LOG_ERR, "exception info:\n");
6418 		for (i = 0; i < nitems(e->u.exception.info); i++)
6419 			log(LOG_ERR, "%s%08x", i == 0 ? "\t" : " ",
6420 			    be32toh(e->u.exception.info[i]));
6421 		log(LOG_ERR, "\n");
6422 		break;
6423 	case FW_ERROR_TYPE_HWMODULE:
6424 		log(LOG_ERR, "HW module regaddr %08x regval %08x\n",
6425 		    be32toh(e->u.hwmodule.regaddr),
6426 		    be32toh(e->u.hwmodule.regval));
6427 		break;
6428 	case FW_ERROR_TYPE_WR:
6429 		log(LOG_ERR, "WR cidx %d PF %d VF %d eqid %d hdr:\n",
6430 		    be16toh(e->u.wr.cidx),
6431 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.wr.pfn_vfn)),
6432 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.wr.pfn_vfn)),
6433 		    be32toh(e->u.wr.eqid));
6434 		for (i = 0; i < nitems(e->u.wr.wrhdr); i++)
6435 			log(LOG_ERR, "%s%02x", i == 0 ? "\t" : " ",
6436 			    e->u.wr.wrhdr[i]);
6437 		log(LOG_ERR, "\n");
6438 		break;
6439 	case FW_ERROR_TYPE_ACL:
6440 		log(LOG_ERR, "ACL cidx %d PF %d VF %d eqid %d %s",
6441 		    be16toh(e->u.acl.cidx),
6442 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.acl.pfn_vfn)),
6443 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.acl.pfn_vfn)),
6444 		    be32toh(e->u.acl.eqid),
6445 		    G_FW_ERROR_CMD_MV(be16toh(e->u.acl.mv_pkd)) ? "vlanid" :
6446 		    "MAC");
6447 		for (i = 0; i < nitems(e->u.acl.val); i++)
6448 			log(LOG_ERR, " %02x", e->u.acl.val[i]);
6449 		log(LOG_ERR, "\n");
6450 		break;
6451 	default:
6452 		log(LOG_ERR, "type %#x\n",
6453 		    G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type)));
6454 		return (EINVAL);
6455 	}
6456 	return (0);
6457 }
6458 
6459 static inline bool
6460 bufidx_used(struct adapter *sc, int idx)
6461 {
6462 	struct rx_buf_info *rxb = &sc->sge.rx_buf_info[0];
6463 	int i;
6464 
6465 	for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) {
6466 		if (rxb->size1 > largest_rx_cluster)
6467 			continue;
6468 		if (rxb->hwidx1 == idx || rxb->hwidx2 == idx)
6469 			return (true);
6470 	}
6471 
6472 	return (false);
6473 }
6474 
6475 static int
6476 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
6477 {
6478 	struct adapter *sc = arg1;
6479 	struct sge_params *sp = &sc->params.sge;
6480 	int i, rc;
6481 	struct sbuf sb;
6482 	char c;
6483 
6484 	sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
6485 	for (i = 0; i < SGE_FLBUF_SIZES; i++) {
6486 		if (bufidx_used(sc, i))
6487 			c = '*';
6488 		else
6489 			c = '\0';
6490 
6491 		sbuf_printf(&sb, "%u%c ", sp->sge_fl_buffer_size[i], c);
6492 	}
6493 	sbuf_trim(&sb);
6494 	sbuf_finish(&sb);
6495 	rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
6496 	sbuf_delete(&sb);
6497 	return (rc);
6498 }
6499 
6500 #ifdef RATELIMIT
6501 #if defined(INET) || defined(INET6)
6502 /*
6503  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
6504  */
6505 static inline u_int
6506 txpkt_eo_len16(u_int nsegs, u_int immhdrs, u_int tso)
6507 {
6508 	u_int n;
6509 
6510 	MPASS(immhdrs > 0);
6511 
6512 	n = roundup2(sizeof(struct fw_eth_tx_eo_wr) +
6513 	    sizeof(struct cpl_tx_pkt_core) + immhdrs, 16);
6514 	if (__predict_false(nsegs == 0))
6515 		goto done;
6516 
6517 	nsegs--; /* first segment is part of ulptx_sgl */
6518 	n += sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
6519 	if (tso)
6520 		n += sizeof(struct cpl_tx_pkt_lso_core);
6521 
6522 done:
6523 	return (howmany(n, 16));
6524 }
6525 #endif
6526 
6527 #define ETID_FLOWC_NPARAMS 6
6528 #define ETID_FLOWC_LEN (roundup2((sizeof(struct fw_flowc_wr) + \
6529     ETID_FLOWC_NPARAMS * sizeof(struct fw_flowc_mnemval)), 16))
6530 #define ETID_FLOWC_LEN16 (howmany(ETID_FLOWC_LEN, 16))
6531 
6532 #if defined(INET) || defined(INET6)
6533 static int
6534 send_etid_flowc_wr(struct cxgbe_rate_tag *cst, struct port_info *pi,
6535     struct vi_info *vi)
6536 {
6537 	struct wrq_cookie cookie;
6538 	u_int pfvf = pi->adapter->pf << S_FW_VIID_PFN;
6539 	struct fw_flowc_wr *flowc;
6540 
6541 	mtx_assert(&cst->lock, MA_OWNED);
6542 	MPASS((cst->flags & (EO_FLOWC_PENDING | EO_FLOWC_RPL_PENDING)) ==
6543 	    EO_FLOWC_PENDING);
6544 
6545 	flowc = start_wrq_wr(&cst->eo_txq->wrq, ETID_FLOWC_LEN16, &cookie);
6546 	if (__predict_false(flowc == NULL))
6547 		return (ENOMEM);
6548 
6549 	bzero(flowc, ETID_FLOWC_LEN);
6550 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
6551 	    V_FW_FLOWC_WR_NPARAMS(ETID_FLOWC_NPARAMS) | V_FW_WR_COMPL(0));
6552 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(ETID_FLOWC_LEN16) |
6553 	    V_FW_WR_FLOWID(cst->etid));
6554 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN;
6555 	flowc->mnemval[0].val = htobe32(pfvf);
6556 	flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH;
6557 	flowc->mnemval[1].val = htobe32(pi->tx_chan);
6558 	flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT;
6559 	flowc->mnemval[2].val = htobe32(pi->tx_chan);
6560 	flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID;
6561 	flowc->mnemval[3].val = htobe32(cst->iqid);
6562 	flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_EOSTATE;
6563 	flowc->mnemval[4].val = htobe32(FW_FLOWC_MNEM_EOSTATE_ESTABLISHED);
6564 	flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS;
6565 	flowc->mnemval[5].val = htobe32(cst->schedcl);
6566 
6567 	commit_wrq_wr(&cst->eo_txq->wrq, flowc, &cookie);
6568 
6569 	cst->flags &= ~EO_FLOWC_PENDING;
6570 	cst->flags |= EO_FLOWC_RPL_PENDING;
6571 	MPASS(cst->tx_credits >= ETID_FLOWC_LEN16);	/* flowc is first WR. */
6572 	cst->tx_credits -= ETID_FLOWC_LEN16;
6573 
6574 	return (0);
6575 }
6576 #endif
6577 
6578 #define ETID_FLUSH_LEN16 (howmany(sizeof (struct fw_flowc_wr), 16))
6579 
6580 void
6581 send_etid_flush_wr(struct cxgbe_rate_tag *cst)
6582 {
6583 	struct fw_flowc_wr *flowc;
6584 	struct wrq_cookie cookie;
6585 
6586 	mtx_assert(&cst->lock, MA_OWNED);
6587 
6588 	flowc = start_wrq_wr(&cst->eo_txq->wrq, ETID_FLUSH_LEN16, &cookie);
6589 	if (__predict_false(flowc == NULL))
6590 		CXGBE_UNIMPLEMENTED(__func__);
6591 
6592 	bzero(flowc, ETID_FLUSH_LEN16 * 16);
6593 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
6594 	    V_FW_FLOWC_WR_NPARAMS(0) | F_FW_WR_COMPL);
6595 	flowc->flowid_len16 = htobe32(V_FW_WR_LEN16(ETID_FLUSH_LEN16) |
6596 	    V_FW_WR_FLOWID(cst->etid));
6597 
6598 	commit_wrq_wr(&cst->eo_txq->wrq, flowc, &cookie);
6599 
6600 	cst->flags |= EO_FLUSH_RPL_PENDING;
6601 	MPASS(cst->tx_credits >= ETID_FLUSH_LEN16);
6602 	cst->tx_credits -= ETID_FLUSH_LEN16;
6603 	cst->ncompl++;
6604 }
6605 
6606 static void
6607 write_ethofld_wr(struct cxgbe_rate_tag *cst, struct fw_eth_tx_eo_wr *wr,
6608     struct mbuf *m0, int compl)
6609 {
6610 	struct cpl_tx_pkt_core *cpl;
6611 	uint64_t ctrl1;
6612 	uint32_t ctrl;	/* used in many unrelated places */
6613 	int len16, pktlen, nsegs, immhdrs;
6614 	uintptr_t p;
6615 	struct ulptx_sgl *usgl;
6616 	struct sglist sg;
6617 	struct sglist_seg segs[38];	/* XXX: find real limit.  XXX: get off the stack */
6618 
6619 	mtx_assert(&cst->lock, MA_OWNED);
6620 	M_ASSERTPKTHDR(m0);
6621 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
6622 	    m0->m_pkthdr.l4hlen > 0,
6623 	    ("%s: ethofld mbuf %p is missing header lengths", __func__, m0));
6624 
6625 	len16 = mbuf_eo_len16(m0);
6626 	nsegs = mbuf_eo_nsegs(m0);
6627 	pktlen = m0->m_pkthdr.len;
6628 	ctrl = sizeof(struct cpl_tx_pkt_core);
6629 	if (needs_tso(m0))
6630 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
6631 	immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen + m0->m_pkthdr.l4hlen;
6632 	ctrl += immhdrs;
6633 
6634 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_EO_WR) |
6635 	    V_FW_ETH_TX_EO_WR_IMMDLEN(ctrl) | V_FW_WR_COMPL(!!compl));
6636 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(len16) |
6637 	    V_FW_WR_FLOWID(cst->etid));
6638 	wr->r3 = 0;
6639 	if (needs_outer_udp_csum(m0)) {
6640 		wr->u.udpseg.type = FW_ETH_TX_EO_TYPE_UDPSEG;
6641 		wr->u.udpseg.ethlen = m0->m_pkthdr.l2hlen;
6642 		wr->u.udpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
6643 		wr->u.udpseg.udplen = m0->m_pkthdr.l4hlen;
6644 		wr->u.udpseg.rtplen = 0;
6645 		wr->u.udpseg.r4 = 0;
6646 		wr->u.udpseg.mss = htobe16(pktlen - immhdrs);
6647 		wr->u.udpseg.schedpktsize = wr->u.udpseg.mss;
6648 		wr->u.udpseg.plen = htobe32(pktlen - immhdrs);
6649 		cpl = (void *)(wr + 1);
6650 	} else {
6651 		MPASS(needs_outer_tcp_csum(m0));
6652 		wr->u.tcpseg.type = FW_ETH_TX_EO_TYPE_TCPSEG;
6653 		wr->u.tcpseg.ethlen = m0->m_pkthdr.l2hlen;
6654 		wr->u.tcpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
6655 		wr->u.tcpseg.tcplen = m0->m_pkthdr.l4hlen;
6656 		wr->u.tcpseg.tsclk_tsoff = mbuf_eo_tsclk_tsoff(m0);
6657 		wr->u.tcpseg.r4 = 0;
6658 		wr->u.tcpseg.r5 = 0;
6659 		wr->u.tcpseg.plen = htobe32(pktlen - immhdrs);
6660 
6661 		if (needs_tso(m0)) {
6662 			struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
6663 
6664 			wr->u.tcpseg.mss = htobe16(m0->m_pkthdr.tso_segsz);
6665 
6666 			ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) |
6667 			    F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE |
6668 			    V_LSO_ETHHDR_LEN((m0->m_pkthdr.l2hlen -
6669 				ETHER_HDR_LEN) >> 2) |
6670 			    V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) |
6671 			    V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
6672 			if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
6673 				ctrl |= F_LSO_IPV6;
6674 			lso->lso_ctrl = htobe32(ctrl);
6675 			lso->ipid_ofst = htobe16(0);
6676 			lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
6677 			lso->seqno_offset = htobe32(0);
6678 			lso->len = htobe32(pktlen);
6679 
6680 			cpl = (void *)(lso + 1);
6681 		} else {
6682 			wr->u.tcpseg.mss = htobe16(0xffff);
6683 			cpl = (void *)(wr + 1);
6684 		}
6685 	}
6686 
6687 	/* Checksum offload must be requested for ethofld. */
6688 	MPASS(needs_outer_l4_csum(m0));
6689 	ctrl1 = csum_to_ctrl(cst->adapter, m0);
6690 
6691 	/* VLAN tag insertion */
6692 	if (needs_vlan_insertion(m0)) {
6693 		ctrl1 |= F_TXPKT_VLAN_VLD |
6694 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
6695 	}
6696 
6697 	/* CPL header */
6698 	cpl->ctrl0 = cst->ctrl0;
6699 	cpl->pack = 0;
6700 	cpl->len = htobe16(pktlen);
6701 	cpl->ctrl1 = htobe64(ctrl1);
6702 
6703 	/* Copy Ethernet, IP & TCP/UDP hdrs as immediate data */
6704 	p = (uintptr_t)(cpl + 1);
6705 	m_copydata(m0, 0, immhdrs, (void *)p);
6706 
6707 	/* SGL */
6708 	if (nsegs > 0) {
6709 		int i, pad;
6710 
6711 		/* zero-pad upto next 16Byte boundary, if not 16Byte aligned */
6712 		p += immhdrs;
6713 		pad = 16 - (immhdrs & 0xf);
6714 		bzero((void *)p, pad);
6715 
6716 		usgl = (void *)(p + pad);
6717 		usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
6718 		    V_ULPTX_NSGE(nsegs));
6719 
6720 		sglist_init(&sg, nitems(segs), segs);
6721 		for (; m0 != NULL; m0 = m0->m_next) {
6722 			if (__predict_false(m0->m_len == 0))
6723 				continue;
6724 			if (immhdrs >= m0->m_len) {
6725 				immhdrs -= m0->m_len;
6726 				continue;
6727 			}
6728 			if (m0->m_flags & M_EXTPG)
6729 				sglist_append_mbuf_epg(&sg, m0,
6730 				    mtod(m0, vm_offset_t), m0->m_len);
6731                         else
6732 				sglist_append(&sg, mtod(m0, char *) + immhdrs,
6733 				    m0->m_len - immhdrs);
6734 			immhdrs = 0;
6735 		}
6736 		MPASS(sg.sg_nseg == nsegs);
6737 
6738 		/*
6739 		 * Zero pad last 8B in case the WR doesn't end on a 16B
6740 		 * boundary.
6741 		 */
6742 		*(uint64_t *)((char *)wr + len16 * 16 - 8) = 0;
6743 
6744 		usgl->len0 = htobe32(segs[0].ss_len);
6745 		usgl->addr0 = htobe64(segs[0].ss_paddr);
6746 		for (i = 0; i < nsegs - 1; i++) {
6747 			usgl->sge[i / 2].len[i & 1] = htobe32(segs[i + 1].ss_len);
6748 			usgl->sge[i / 2].addr[i & 1] = htobe64(segs[i + 1].ss_paddr);
6749 		}
6750 		if (i & 1)
6751 			usgl->sge[i / 2].len[1] = htobe32(0);
6752 	}
6753 
6754 }
6755 
6756 static void
6757 ethofld_tx(struct cxgbe_rate_tag *cst)
6758 {
6759 	struct mbuf *m;
6760 	struct wrq_cookie cookie;
6761 	int next_credits, compl;
6762 	struct fw_eth_tx_eo_wr *wr;
6763 
6764 	mtx_assert(&cst->lock, MA_OWNED);
6765 
6766 	while ((m = mbufq_first(&cst->pending_tx)) != NULL) {
6767 		M_ASSERTPKTHDR(m);
6768 
6769 		/* How many len16 credits do we need to send this mbuf. */
6770 		next_credits = mbuf_eo_len16(m);
6771 		MPASS(next_credits > 0);
6772 		if (next_credits > cst->tx_credits) {
6773 			/*
6774 			 * Tx will make progress eventually because there is at
6775 			 * least one outstanding fw4_ack that will return
6776 			 * credits and kick the tx.
6777 			 */
6778 			MPASS(cst->ncompl > 0);
6779 			return;
6780 		}
6781 		wr = start_wrq_wr(&cst->eo_txq->wrq, next_credits, &cookie);
6782 		if (__predict_false(wr == NULL)) {
6783 			/* XXX: wishful thinking, not a real assertion. */
6784 			MPASS(cst->ncompl > 0);
6785 			return;
6786 		}
6787 		cst->tx_credits -= next_credits;
6788 		cst->tx_nocompl += next_credits;
6789 		compl = cst->ncompl == 0 || cst->tx_nocompl >= cst->tx_total / 2;
6790 		ETHER_BPF_MTAP(cst->com.ifp, m);
6791 		write_ethofld_wr(cst, wr, m, compl);
6792 		commit_wrq_wr(&cst->eo_txq->wrq, wr, &cookie);
6793 		if (compl) {
6794 			cst->ncompl++;
6795 			cst->tx_nocompl	= 0;
6796 		}
6797 		(void) mbufq_dequeue(&cst->pending_tx);
6798 
6799 		/*
6800 		 * Drop the mbuf's reference on the tag now rather
6801 		 * than waiting until m_freem().  This ensures that
6802 		 * cxgbe_rate_tag_free gets called when the inp drops
6803 		 * its reference on the tag and there are no more
6804 		 * mbufs in the pending_tx queue and can flush any
6805 		 * pending requests.  Otherwise if the last mbuf
6806 		 * doesn't request a completion the etid will never be
6807 		 * released.
6808 		 */
6809 		m->m_pkthdr.snd_tag = NULL;
6810 		m->m_pkthdr.csum_flags &= ~CSUM_SND_TAG;
6811 		m_snd_tag_rele(&cst->com);
6812 
6813 		mbufq_enqueue(&cst->pending_fwack, m);
6814 	}
6815 }
6816 
6817 #if defined(INET) || defined(INET6)
6818 static int
6819 ethofld_transmit(if_t ifp, struct mbuf *m0)
6820 {
6821 	struct cxgbe_rate_tag *cst;
6822 	int rc;
6823 
6824 	MPASS(m0->m_nextpkt == NULL);
6825 	MPASS(m0->m_pkthdr.csum_flags & CSUM_SND_TAG);
6826 	MPASS(m0->m_pkthdr.snd_tag != NULL);
6827 	cst = mst_to_crt(m0->m_pkthdr.snd_tag);
6828 
6829 	mtx_lock(&cst->lock);
6830 	MPASS(cst->flags & EO_SND_TAG_REF);
6831 
6832 	if (__predict_false(cst->flags & EO_FLOWC_PENDING)) {
6833 		struct vi_info *vi = if_getsoftc(ifp);
6834 		struct port_info *pi = vi->pi;
6835 		struct adapter *sc = pi->adapter;
6836 		const uint32_t rss_mask = vi->rss_size - 1;
6837 		uint32_t rss_hash;
6838 
6839 		cst->eo_txq = &sc->sge.ofld_txq[vi->first_ofld_txq];
6840 		if (M_HASHTYPE_ISHASH(m0))
6841 			rss_hash = m0->m_pkthdr.flowid;
6842 		else
6843 			rss_hash = arc4random();
6844 		/* We assume RSS hashing */
6845 		cst->iqid = vi->rss[rss_hash & rss_mask];
6846 		cst->eo_txq += rss_hash % vi->nofldtxq;
6847 		rc = send_etid_flowc_wr(cst, pi, vi);
6848 		if (rc != 0)
6849 			goto done;
6850 	}
6851 
6852 	if (__predict_false(cst->plen + m0->m_pkthdr.len > eo_max_backlog)) {
6853 		rc = ENOBUFS;
6854 		goto done;
6855 	}
6856 
6857 	mbufq_enqueue(&cst->pending_tx, m0);
6858 	cst->plen += m0->m_pkthdr.len;
6859 
6860 	/*
6861 	 * Hold an extra reference on the tag while generating work
6862 	 * requests to ensure that we don't try to free the tag during
6863 	 * ethofld_tx() in case we are sending the final mbuf after
6864 	 * the inp was freed.
6865 	 */
6866 	m_snd_tag_ref(&cst->com);
6867 	ethofld_tx(cst);
6868 	mtx_unlock(&cst->lock);
6869 	m_snd_tag_rele(&cst->com);
6870 	return (0);
6871 
6872 done:
6873 	mtx_unlock(&cst->lock);
6874 	return (rc);
6875 }
6876 #endif
6877 
6878 static int
6879 ethofld_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
6880 {
6881 	struct adapter *sc = iq->adapter;
6882 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
6883 	struct mbuf *m;
6884 	u_int etid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
6885 	struct cxgbe_rate_tag *cst;
6886 	uint8_t credits = cpl->credits;
6887 
6888 	cst = lookup_etid(sc, etid);
6889 	mtx_lock(&cst->lock);
6890 	if (__predict_false(cst->flags & EO_FLOWC_RPL_PENDING)) {
6891 		MPASS(credits >= ETID_FLOWC_LEN16);
6892 		credits -= ETID_FLOWC_LEN16;
6893 		cst->flags &= ~EO_FLOWC_RPL_PENDING;
6894 	}
6895 
6896 	KASSERT(cst->ncompl > 0,
6897 	    ("%s: etid %u (%p) wasn't expecting completion.",
6898 	    __func__, etid, cst));
6899 	cst->ncompl--;
6900 
6901 	while (credits > 0) {
6902 		m = mbufq_dequeue(&cst->pending_fwack);
6903 		if (__predict_false(m == NULL)) {
6904 			/*
6905 			 * The remaining credits are for the final flush that
6906 			 * was issued when the tag was freed by the kernel.
6907 			 */
6908 			MPASS((cst->flags &
6909 			    (EO_FLUSH_RPL_PENDING | EO_SND_TAG_REF)) ==
6910 			    EO_FLUSH_RPL_PENDING);
6911 			MPASS(credits == ETID_FLUSH_LEN16);
6912 			MPASS(cst->tx_credits + cpl->credits == cst->tx_total);
6913 			MPASS(cst->ncompl == 0);
6914 
6915 			cst->flags &= ~EO_FLUSH_RPL_PENDING;
6916 			cst->tx_credits += cpl->credits;
6917 			cxgbe_rate_tag_free_locked(cst);
6918 			return (0);	/* cst is gone. */
6919 		}
6920 		KASSERT(m != NULL,
6921 		    ("%s: too many credits (%u, %u)", __func__, cpl->credits,
6922 		    credits));
6923 		KASSERT(credits >= mbuf_eo_len16(m),
6924 		    ("%s: too few credits (%u, %u, %u)", __func__,
6925 		    cpl->credits, credits, mbuf_eo_len16(m)));
6926 		credits -= mbuf_eo_len16(m);
6927 		cst->plen -= m->m_pkthdr.len;
6928 		m_freem(m);
6929 	}
6930 
6931 	cst->tx_credits += cpl->credits;
6932 	MPASS(cst->tx_credits <= cst->tx_total);
6933 
6934 	if (cst->flags & EO_SND_TAG_REF) {
6935 		/*
6936 		 * As with ethofld_transmit(), hold an extra reference
6937 		 * so that the tag is stable across ethold_tx().
6938 		 */
6939 		m_snd_tag_ref(&cst->com);
6940 		m = mbufq_first(&cst->pending_tx);
6941 		if (m != NULL && cst->tx_credits >= mbuf_eo_len16(m))
6942 			ethofld_tx(cst);
6943 		mtx_unlock(&cst->lock);
6944 		m_snd_tag_rele(&cst->com);
6945 	} else {
6946 		/*
6947 		 * There shouldn't be any pending packets if the tag
6948 		 * was freed by the kernel since any pending packet
6949 		 * should hold a reference to the tag.
6950 		 */
6951 		MPASS(mbufq_first(&cst->pending_tx) == NULL);
6952 		mtx_unlock(&cst->lock);
6953 	}
6954 
6955 	return (0);
6956 }
6957 #endif
6958