xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision 87b759f0fa1f7554d50ce640c40138512bbded44)
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 		if (sc->flags & IS_VF && sc->vlan_id) {
2090 			/*
2091 			 * HW is not setup correctly if extracted vlan_id does
2092 			 * not match the VF's setting.
2093 			 */
2094 			MPASS(be16toh(cpl->vlan) == sc->vlan_id);
2095 		} else {
2096 			m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
2097 			m0->m_flags |= M_VLANTAG;
2098 			rxq->vlan_extraction++;
2099 		}
2100 	}
2101 
2102 	if (rxq->iq.flags & IQ_RX_TIMESTAMP) {
2103 		/*
2104 		 * Fill up rcv_tstmp but do not set M_TSTMP as
2105 		 * long as we get a non-zero back from t4_tstmp_to_ns().
2106 		 */
2107 		m0->m_pkthdr.rcv_tstmp = t4_tstmp_to_ns(sc,
2108 		    be64toh(d->rsp.u.last_flit));
2109 		if (m0->m_pkthdr.rcv_tstmp != 0)
2110 			m0->m_flags |= M_TSTMP;
2111 	}
2112 
2113 #ifdef NUMA
2114 	m0->m_pkthdr.numa_domain = if_getnumadomain(ifp);
2115 #endif
2116 #if defined(INET) || defined(INET6)
2117 	if (rxq->iq.flags & IQ_LRO_ENABLED && tnl_type == 0 &&
2118 	    (M_HASHTYPE_GET(m0) == M_HASHTYPE_RSS_TCP_IPV4 ||
2119 	    M_HASHTYPE_GET(m0) == M_HASHTYPE_RSS_TCP_IPV6)) {
2120 		if (sort_before_lro(lro)) {
2121 			tcp_lro_queue_mbuf(lro, m0);
2122 			return (0); /* queued for sort, then LRO */
2123 		}
2124 		if (tcp_lro_rx(lro, m0, 0) == 0)
2125 			return (0); /* queued for LRO */
2126 	}
2127 #endif
2128 	if_input(ifp, m0);
2129 
2130 	return (0);
2131 }
2132 
2133 /*
2134  * Must drain the wrq or make sure that someone else will.
2135  */
2136 static void
2137 wrq_tx_drain(void *arg, int n)
2138 {
2139 	struct sge_wrq *wrq = arg;
2140 	struct sge_eq *eq = &wrq->eq;
2141 
2142 	EQ_LOCK(eq);
2143 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2144 		drain_wrq_wr_list(wrq->adapter, wrq);
2145 	EQ_UNLOCK(eq);
2146 }
2147 
2148 static void
2149 drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq)
2150 {
2151 	struct sge_eq *eq = &wrq->eq;
2152 	u_int available, dbdiff;	/* # of hardware descriptors */
2153 	u_int n;
2154 	struct wrqe *wr;
2155 	struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
2156 
2157 	EQ_LOCK_ASSERT_OWNED(eq);
2158 	MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
2159 	wr = STAILQ_FIRST(&wrq->wr_list);
2160 	MPASS(wr != NULL);	/* Must be called with something useful to do */
2161 	MPASS(eq->pidx == eq->dbidx);
2162 	dbdiff = 0;
2163 
2164 	do {
2165 		eq->cidx = read_hw_cidx(eq);
2166 		if (eq->pidx == eq->cidx)
2167 			available = eq->sidx - 1;
2168 		else
2169 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2170 
2171 		MPASS(wr->wrq == wrq);
2172 		n = howmany(wr->wr_len, EQ_ESIZE);
2173 		if (available < n)
2174 			break;
2175 
2176 		dst = (void *)&eq->desc[eq->pidx];
2177 		if (__predict_true(eq->sidx - eq->pidx > n)) {
2178 			/* Won't wrap, won't end exactly at the status page. */
2179 			bcopy(&wr->wr[0], dst, wr->wr_len);
2180 			eq->pidx += n;
2181 		} else {
2182 			int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE;
2183 
2184 			bcopy(&wr->wr[0], dst, first_portion);
2185 			if (wr->wr_len > first_portion) {
2186 				bcopy(&wr->wr[first_portion], &eq->desc[0],
2187 				    wr->wr_len - first_portion);
2188 			}
2189 			eq->pidx = n - (eq->sidx - eq->pidx);
2190 		}
2191 		wrq->tx_wrs_copied++;
2192 
2193 		if (available < eq->sidx / 4 &&
2194 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2195 				/*
2196 				 * XXX: This is not 100% reliable with some
2197 				 * types of WRs.  But this is a very unusual
2198 				 * situation for an ofld/ctrl queue anyway.
2199 				 */
2200 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2201 			    F_FW_WR_EQUEQ);
2202 		}
2203 
2204 		dbdiff += n;
2205 		if (dbdiff >= 16) {
2206 			ring_eq_db(sc, eq, dbdiff);
2207 			dbdiff = 0;
2208 		}
2209 
2210 		STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
2211 		free_wrqe(wr);
2212 		MPASS(wrq->nwr_pending > 0);
2213 		wrq->nwr_pending--;
2214 		MPASS(wrq->ndesc_needed >= n);
2215 		wrq->ndesc_needed -= n;
2216 	} while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL);
2217 
2218 	if (dbdiff)
2219 		ring_eq_db(sc, eq, dbdiff);
2220 }
2221 
2222 /*
2223  * Doesn't fail.  Holds on to work requests it can't send right away.
2224  */
2225 void
2226 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
2227 {
2228 #ifdef INVARIANTS
2229 	struct sge_eq *eq = &wrq->eq;
2230 #endif
2231 
2232 	EQ_LOCK_ASSERT_OWNED(eq);
2233 	MPASS(wr != NULL);
2234 	MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN);
2235 	MPASS((wr->wr_len & 0x7) == 0);
2236 
2237 	STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
2238 	wrq->nwr_pending++;
2239 	wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE);
2240 
2241 	if (!TAILQ_EMPTY(&wrq->incomplete_wrs))
2242 		return;	/* commit_wrq_wr will drain wr_list as well. */
2243 
2244 	drain_wrq_wr_list(sc, wrq);
2245 
2246 	/* Doorbell must have caught up to the pidx. */
2247 	MPASS(eq->pidx == eq->dbidx);
2248 }
2249 
2250 void
2251 t4_update_fl_bufsize(if_t ifp)
2252 {
2253 	struct vi_info *vi = if_getsoftc(ifp);
2254 	struct adapter *sc = vi->adapter;
2255 	struct sge_rxq *rxq;
2256 #ifdef TCP_OFFLOAD
2257 	struct sge_ofld_rxq *ofld_rxq;
2258 #endif
2259 	struct sge_fl *fl;
2260 	int i, maxp;
2261 
2262 	maxp = max_rx_payload(sc, ifp, false);
2263 	for_each_rxq(vi, i, rxq) {
2264 		fl = &rxq->fl;
2265 
2266 		FL_LOCK(fl);
2267 		fl->zidx = find_refill_source(sc, maxp,
2268 		    fl->flags & FL_BUF_PACKING);
2269 		FL_UNLOCK(fl);
2270 	}
2271 #ifdef TCP_OFFLOAD
2272 	maxp = max_rx_payload(sc, ifp, true);
2273 	for_each_ofld_rxq(vi, i, ofld_rxq) {
2274 		fl = &ofld_rxq->fl;
2275 
2276 		FL_LOCK(fl);
2277 		fl->zidx = find_refill_source(sc, maxp,
2278 		    fl->flags & FL_BUF_PACKING);
2279 		FL_UNLOCK(fl);
2280 	}
2281 #endif
2282 }
2283 
2284 #ifdef RATELIMIT
2285 static inline int
2286 mbuf_eo_nsegs(struct mbuf *m)
2287 {
2288 
2289 	M_ASSERTPKTHDR(m);
2290 	return (m->m_pkthdr.PH_loc.eight[1]);
2291 }
2292 
2293 #if defined(INET) || defined(INET6)
2294 static inline void
2295 set_mbuf_eo_nsegs(struct mbuf *m, uint8_t nsegs)
2296 {
2297 
2298 	M_ASSERTPKTHDR(m);
2299 	m->m_pkthdr.PH_loc.eight[1] = nsegs;
2300 }
2301 #endif
2302 
2303 static inline int
2304 mbuf_eo_len16(struct mbuf *m)
2305 {
2306 	int n;
2307 
2308 	M_ASSERTPKTHDR(m);
2309 	n = m->m_pkthdr.PH_loc.eight[2];
2310 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2311 
2312 	return (n);
2313 }
2314 
2315 #if defined(INET) || defined(INET6)
2316 static inline void
2317 set_mbuf_eo_len16(struct mbuf *m, uint8_t len16)
2318 {
2319 
2320 	M_ASSERTPKTHDR(m);
2321 	m->m_pkthdr.PH_loc.eight[2] = len16;
2322 }
2323 #endif
2324 
2325 static inline int
2326 mbuf_eo_tsclk_tsoff(struct mbuf *m)
2327 {
2328 
2329 	M_ASSERTPKTHDR(m);
2330 	return (m->m_pkthdr.PH_loc.eight[3]);
2331 }
2332 
2333 #if defined(INET) || defined(INET6)
2334 static inline void
2335 set_mbuf_eo_tsclk_tsoff(struct mbuf *m, uint8_t tsclk_tsoff)
2336 {
2337 
2338 	M_ASSERTPKTHDR(m);
2339 	m->m_pkthdr.PH_loc.eight[3] = tsclk_tsoff;
2340 }
2341 #endif
2342 
2343 static inline int
2344 needs_eo(struct m_snd_tag *mst)
2345 {
2346 
2347 	return (mst != NULL && mst->sw->type == IF_SND_TAG_TYPE_RATE_LIMIT);
2348 }
2349 #endif
2350 
2351 /*
2352  * Try to allocate an mbuf to contain a raw work request.  To make it
2353  * easy to construct the work request, don't allocate a chain but a
2354  * single mbuf.
2355  */
2356 struct mbuf *
2357 alloc_wr_mbuf(int len, int how)
2358 {
2359 	struct mbuf *m;
2360 
2361 	if (len <= MHLEN)
2362 		m = m_gethdr(how, MT_DATA);
2363 	else if (len <= MCLBYTES)
2364 		m = m_getcl(how, MT_DATA, M_PKTHDR);
2365 	else
2366 		m = NULL;
2367 	if (m == NULL)
2368 		return (NULL);
2369 	m->m_pkthdr.len = len;
2370 	m->m_len = len;
2371 	set_mbuf_cflags(m, MC_RAW_WR);
2372 	set_mbuf_len16(m, howmany(len, 16));
2373 	return (m);
2374 }
2375 
2376 static inline bool
2377 needs_hwcsum(struct mbuf *m)
2378 {
2379 	const uint32_t csum_flags = CSUM_IP | CSUM_IP_UDP | CSUM_IP_TCP |
2380 	    CSUM_IP_TSO | CSUM_INNER_IP | CSUM_INNER_IP_UDP |
2381 	    CSUM_INNER_IP_TCP | CSUM_INNER_IP_TSO | CSUM_IP6_UDP |
2382 	    CSUM_IP6_TCP | CSUM_IP6_TSO | CSUM_INNER_IP6_UDP |
2383 	    CSUM_INNER_IP6_TCP | CSUM_INNER_IP6_TSO;
2384 
2385 	M_ASSERTPKTHDR(m);
2386 
2387 	return (m->m_pkthdr.csum_flags & csum_flags);
2388 }
2389 
2390 static inline bool
2391 needs_tso(struct mbuf *m)
2392 {
2393 	const uint32_t csum_flags = CSUM_IP_TSO | CSUM_IP6_TSO |
2394 	    CSUM_INNER_IP_TSO | CSUM_INNER_IP6_TSO;
2395 
2396 	M_ASSERTPKTHDR(m);
2397 
2398 	return (m->m_pkthdr.csum_flags & csum_flags);
2399 }
2400 
2401 static inline bool
2402 needs_vxlan_csum(struct mbuf *m)
2403 {
2404 
2405 	M_ASSERTPKTHDR(m);
2406 
2407 	return (m->m_pkthdr.csum_flags & CSUM_ENCAP_VXLAN);
2408 }
2409 
2410 static inline bool
2411 needs_vxlan_tso(struct mbuf *m)
2412 {
2413 	const uint32_t csum_flags = CSUM_ENCAP_VXLAN | CSUM_INNER_IP_TSO |
2414 	    CSUM_INNER_IP6_TSO;
2415 
2416 	M_ASSERTPKTHDR(m);
2417 
2418 	return ((m->m_pkthdr.csum_flags & csum_flags) != 0 &&
2419 	    (m->m_pkthdr.csum_flags & csum_flags) != CSUM_ENCAP_VXLAN);
2420 }
2421 
2422 #if defined(INET) || defined(INET6)
2423 static inline bool
2424 needs_inner_tcp_csum(struct mbuf *m)
2425 {
2426 	const uint32_t csum_flags = CSUM_INNER_IP_TSO | CSUM_INNER_IP6_TSO;
2427 
2428 	M_ASSERTPKTHDR(m);
2429 
2430 	return (m->m_pkthdr.csum_flags & csum_flags);
2431 }
2432 #endif
2433 
2434 static inline bool
2435 needs_l3_csum(struct mbuf *m)
2436 {
2437 	const uint32_t csum_flags = CSUM_IP | CSUM_IP_TSO | CSUM_INNER_IP |
2438 	    CSUM_INNER_IP_TSO;
2439 
2440 	M_ASSERTPKTHDR(m);
2441 
2442 	return (m->m_pkthdr.csum_flags & csum_flags);
2443 }
2444 
2445 static inline bool
2446 needs_outer_tcp_csum(struct mbuf *m)
2447 {
2448 	const uint32_t csum_flags = CSUM_IP_TCP | CSUM_IP_TSO | CSUM_IP6_TCP |
2449 	    CSUM_IP6_TSO;
2450 
2451 	M_ASSERTPKTHDR(m);
2452 
2453 	return (m->m_pkthdr.csum_flags & csum_flags);
2454 }
2455 
2456 #ifdef RATELIMIT
2457 static inline bool
2458 needs_outer_l4_csum(struct mbuf *m)
2459 {
2460 	const uint32_t csum_flags = CSUM_IP_UDP | CSUM_IP_TCP | CSUM_IP_TSO |
2461 	    CSUM_IP6_UDP | CSUM_IP6_TCP | CSUM_IP6_TSO;
2462 
2463 	M_ASSERTPKTHDR(m);
2464 
2465 	return (m->m_pkthdr.csum_flags & csum_flags);
2466 }
2467 
2468 static inline bool
2469 needs_outer_udp_csum(struct mbuf *m)
2470 {
2471 	const uint32_t csum_flags = CSUM_IP_UDP | CSUM_IP6_UDP;
2472 
2473 	M_ASSERTPKTHDR(m);
2474 
2475 	return (m->m_pkthdr.csum_flags & csum_flags);
2476 }
2477 #endif
2478 
2479 static inline bool
2480 needs_vlan_insertion(struct mbuf *m)
2481 {
2482 
2483 	M_ASSERTPKTHDR(m);
2484 
2485 	return (m->m_flags & M_VLANTAG);
2486 }
2487 
2488 #if defined(INET) || defined(INET6)
2489 static void *
2490 m_advance(struct mbuf **pm, int *poffset, int len)
2491 {
2492 	struct mbuf *m = *pm;
2493 	int offset = *poffset;
2494 	uintptr_t p = 0;
2495 
2496 	MPASS(len > 0);
2497 
2498 	for (;;) {
2499 		if (offset + len < m->m_len) {
2500 			offset += len;
2501 			p = mtod(m, uintptr_t) + offset;
2502 			break;
2503 		}
2504 		len -= m->m_len - offset;
2505 		m = m->m_next;
2506 		offset = 0;
2507 		MPASS(m != NULL);
2508 	}
2509 	*poffset = offset;
2510 	*pm = m;
2511 	return ((void *)p);
2512 }
2513 #endif
2514 
2515 static inline int
2516 count_mbuf_ext_pgs(struct mbuf *m, int skip, vm_paddr_t *nextaddr)
2517 {
2518 	vm_paddr_t paddr;
2519 	int i, len, off, pglen, pgoff, seglen, segoff;
2520 	int nsegs = 0;
2521 
2522 	M_ASSERTEXTPG(m);
2523 	off = mtod(m, vm_offset_t);
2524 	len = m->m_len;
2525 	off += skip;
2526 	len -= skip;
2527 
2528 	if (m->m_epg_hdrlen != 0) {
2529 		if (off >= m->m_epg_hdrlen) {
2530 			off -= m->m_epg_hdrlen;
2531 		} else {
2532 			seglen = m->m_epg_hdrlen - off;
2533 			segoff = off;
2534 			seglen = min(seglen, len);
2535 			off = 0;
2536 			len -= seglen;
2537 			paddr = pmap_kextract(
2538 			    (vm_offset_t)&m->m_epg_hdr[segoff]);
2539 			if (*nextaddr != paddr)
2540 				nsegs++;
2541 			*nextaddr = paddr + seglen;
2542 		}
2543 	}
2544 	pgoff = m->m_epg_1st_off;
2545 	for (i = 0; i < m->m_epg_npgs && len > 0; i++) {
2546 		pglen = m_epg_pagelen(m, i, pgoff);
2547 		if (off >= pglen) {
2548 			off -= pglen;
2549 			pgoff = 0;
2550 			continue;
2551 		}
2552 		seglen = pglen - off;
2553 		segoff = pgoff + off;
2554 		off = 0;
2555 		seglen = min(seglen, len);
2556 		len -= seglen;
2557 		paddr = m->m_epg_pa[i] + segoff;
2558 		if (*nextaddr != paddr)
2559 			nsegs++;
2560 		*nextaddr = paddr + seglen;
2561 		pgoff = 0;
2562 	};
2563 	if (len != 0) {
2564 		seglen = min(len, m->m_epg_trllen - off);
2565 		len -= seglen;
2566 		paddr = pmap_kextract((vm_offset_t)&m->m_epg_trail[off]);
2567 		if (*nextaddr != paddr)
2568 			nsegs++;
2569 		*nextaddr = paddr + seglen;
2570 	}
2571 
2572 	return (nsegs);
2573 }
2574 
2575 
2576 /*
2577  * Can deal with empty mbufs in the chain that have m_len = 0, but the chain
2578  * must have at least one mbuf that's not empty.  It is possible for this
2579  * routine to return 0 if skip accounts for all the contents of the mbuf chain.
2580  */
2581 static inline int
2582 count_mbuf_nsegs(struct mbuf *m, int skip, uint8_t *cflags)
2583 {
2584 	vm_paddr_t nextaddr, paddr;
2585 	vm_offset_t va;
2586 	int len, nsegs;
2587 
2588 	M_ASSERTPKTHDR(m);
2589 	MPASS(m->m_pkthdr.len > 0);
2590 	MPASS(m->m_pkthdr.len >= skip);
2591 
2592 	nsegs = 0;
2593 	nextaddr = 0;
2594 	for (; m; m = m->m_next) {
2595 		len = m->m_len;
2596 		if (__predict_false(len == 0))
2597 			continue;
2598 		if (skip >= len) {
2599 			skip -= len;
2600 			continue;
2601 		}
2602 		if ((m->m_flags & M_EXTPG) != 0) {
2603 			*cflags |= MC_NOMAP;
2604 			nsegs += count_mbuf_ext_pgs(m, skip, &nextaddr);
2605 			skip = 0;
2606 			continue;
2607 		}
2608 		va = mtod(m, vm_offset_t) + skip;
2609 		len -= skip;
2610 		skip = 0;
2611 		paddr = pmap_kextract(va);
2612 		nsegs += sglist_count((void *)(uintptr_t)va, len);
2613 		if (paddr == nextaddr)
2614 			nsegs--;
2615 		nextaddr = pmap_kextract(va + len - 1) + 1;
2616 	}
2617 
2618 	return (nsegs);
2619 }
2620 
2621 /*
2622  * The maximum number of segments that can fit in a WR.
2623  */
2624 static int
2625 max_nsegs_allowed(struct mbuf *m, bool vm_wr)
2626 {
2627 
2628 	if (vm_wr) {
2629 		if (needs_tso(m))
2630 			return (TX_SGL_SEGS_VM_TSO);
2631 		return (TX_SGL_SEGS_VM);
2632 	}
2633 
2634 	if (needs_tso(m)) {
2635 		if (needs_vxlan_tso(m))
2636 			return (TX_SGL_SEGS_VXLAN_TSO);
2637 		else
2638 			return (TX_SGL_SEGS_TSO);
2639 	}
2640 
2641 	return (TX_SGL_SEGS);
2642 }
2643 
2644 static struct timeval txerr_ratecheck = {0};
2645 static const struct timeval txerr_interval = {3, 0};
2646 
2647 /*
2648  * Analyze the mbuf to determine its tx needs.  The mbuf passed in may change:
2649  * a) caller can assume it's been freed if this function returns with an error.
2650  * b) it may get defragged up if the gather list is too long for the hardware.
2651  */
2652 int
2653 parse_pkt(struct mbuf **mp, bool vm_wr)
2654 {
2655 	struct mbuf *m0 = *mp, *m;
2656 	int rc, nsegs, defragged = 0;
2657 	struct ether_header *eh;
2658 #ifdef INET
2659 	void *l3hdr;
2660 #endif
2661 #if defined(INET) || defined(INET6)
2662 	int offset;
2663 	struct tcphdr *tcp;
2664 #endif
2665 #if defined(KERN_TLS) || defined(RATELIMIT)
2666 	struct m_snd_tag *mst;
2667 #endif
2668 	uint16_t eh_type;
2669 	uint8_t cflags;
2670 
2671 	cflags = 0;
2672 	M_ASSERTPKTHDR(m0);
2673 	if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) {
2674 		rc = EINVAL;
2675 fail:
2676 		m_freem(m0);
2677 		*mp = NULL;
2678 		return (rc);
2679 	}
2680 restart:
2681 	/*
2682 	 * First count the number of gather list segments in the payload.
2683 	 * Defrag the mbuf if nsegs exceeds the hardware limit.
2684 	 */
2685 	M_ASSERTPKTHDR(m0);
2686 	MPASS(m0->m_pkthdr.len > 0);
2687 	nsegs = count_mbuf_nsegs(m0, 0, &cflags);
2688 #if defined(KERN_TLS) || defined(RATELIMIT)
2689 	if (m0->m_pkthdr.csum_flags & CSUM_SND_TAG)
2690 		mst = m0->m_pkthdr.snd_tag;
2691 	else
2692 		mst = NULL;
2693 #endif
2694 #ifdef KERN_TLS
2695 	if (mst != NULL && mst->sw->type == IF_SND_TAG_TYPE_TLS) {
2696 		cflags |= MC_TLS;
2697 		set_mbuf_cflags(m0, cflags);
2698 		rc = t6_ktls_parse_pkt(m0);
2699 		if (rc != 0)
2700 			goto fail;
2701 		return (EINPROGRESS);
2702 	}
2703 #endif
2704 	if (nsegs > max_nsegs_allowed(m0, vm_wr)) {
2705 		if (defragged++ > 0) {
2706 			rc = EFBIG;
2707 			goto fail;
2708 		}
2709 		counter_u64_add(defrags, 1);
2710 		if ((m = m_defrag(m0, M_NOWAIT)) == NULL) {
2711 			rc = ENOMEM;
2712 			goto fail;
2713 		}
2714 		*mp = m0 = m;	/* update caller's copy after defrag */
2715 		goto restart;
2716 	}
2717 
2718 	if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN &&
2719 	    !(cflags & MC_NOMAP))) {
2720 		counter_u64_add(pullups, 1);
2721 		m0 = m_pullup(m0, m0->m_pkthdr.len);
2722 		if (m0 == NULL) {
2723 			/* Should have left well enough alone. */
2724 			rc = EFBIG;
2725 			goto fail;
2726 		}
2727 		*mp = m0;	/* update caller's copy after pullup */
2728 		goto restart;
2729 	}
2730 	set_mbuf_nsegs(m0, nsegs);
2731 	set_mbuf_cflags(m0, cflags);
2732 	calculate_mbuf_len16(m0, vm_wr);
2733 
2734 #ifdef RATELIMIT
2735 	/*
2736 	 * Ethofld is limited to TCP and UDP for now, and only when L4 hw
2737 	 * checksumming is enabled.  needs_outer_l4_csum happens to check for
2738 	 * all the right things.
2739 	 */
2740 	if (__predict_false(needs_eo(mst) && !needs_outer_l4_csum(m0))) {
2741 		m_snd_tag_rele(m0->m_pkthdr.snd_tag);
2742 		m0->m_pkthdr.snd_tag = NULL;
2743 		m0->m_pkthdr.csum_flags &= ~CSUM_SND_TAG;
2744 		mst = NULL;
2745 	}
2746 #endif
2747 
2748 	if (!needs_hwcsum(m0)
2749 #ifdef RATELIMIT
2750 		 && !needs_eo(mst)
2751 #endif
2752 	)
2753 		return (0);
2754 
2755 	m = m0;
2756 	eh = mtod(m, struct ether_header *);
2757 	eh_type = ntohs(eh->ether_type);
2758 	if (eh_type == ETHERTYPE_VLAN) {
2759 		struct ether_vlan_header *evh = (void *)eh;
2760 
2761 		eh_type = ntohs(evh->evl_proto);
2762 		m0->m_pkthdr.l2hlen = sizeof(*evh);
2763 	} else
2764 		m0->m_pkthdr.l2hlen = sizeof(*eh);
2765 
2766 #if defined(INET) || defined(INET6)
2767 	offset = 0;
2768 #ifdef INET
2769 	l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2770 #else
2771 	m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2772 #endif
2773 #endif
2774 
2775 	switch (eh_type) {
2776 #ifdef INET6
2777 	case ETHERTYPE_IPV6:
2778 		m0->m_pkthdr.l3hlen = sizeof(struct ip6_hdr);
2779 		break;
2780 #endif
2781 #ifdef INET
2782 	case ETHERTYPE_IP:
2783 	{
2784 		struct ip *ip = l3hdr;
2785 
2786 		if (needs_vxlan_csum(m0)) {
2787 			/* Driver will do the outer IP hdr checksum. */
2788 			ip->ip_sum = 0;
2789 			if (needs_vxlan_tso(m0)) {
2790 				const uint16_t ipl = ip->ip_len;
2791 
2792 				ip->ip_len = 0;
2793 				ip->ip_sum = ~in_cksum_hdr(ip);
2794 				ip->ip_len = ipl;
2795 			} else
2796 				ip->ip_sum = in_cksum_hdr(ip);
2797 		}
2798 		m0->m_pkthdr.l3hlen = ip->ip_hl << 2;
2799 		break;
2800 	}
2801 #endif
2802 	default:
2803 		if (ratecheck(&txerr_ratecheck, &txerr_interval)) {
2804 			log(LOG_ERR, "%s: ethertype 0x%04x unknown.  "
2805 			    "if_cxgbe must be compiled with the same "
2806 			    "INET/INET6 options as the kernel.\n", __func__,
2807 			    eh_type);
2808 		}
2809 		rc = EINVAL;
2810 		goto fail;
2811 	}
2812 
2813 #if defined(INET) || defined(INET6)
2814 	if (needs_vxlan_csum(m0)) {
2815 		m0->m_pkthdr.l4hlen = sizeof(struct udphdr);
2816 		m0->m_pkthdr.l5hlen = sizeof(struct vxlan_header);
2817 
2818 		/* Inner headers. */
2819 		eh = m_advance(&m, &offset, m0->m_pkthdr.l3hlen +
2820 		    sizeof(struct udphdr) + sizeof(struct vxlan_header));
2821 		eh_type = ntohs(eh->ether_type);
2822 		if (eh_type == ETHERTYPE_VLAN) {
2823 			struct ether_vlan_header *evh = (void *)eh;
2824 
2825 			eh_type = ntohs(evh->evl_proto);
2826 			m0->m_pkthdr.inner_l2hlen = sizeof(*evh);
2827 		} else
2828 			m0->m_pkthdr.inner_l2hlen = sizeof(*eh);
2829 #ifdef INET
2830 		l3hdr = m_advance(&m, &offset, m0->m_pkthdr.inner_l2hlen);
2831 #else
2832 		m_advance(&m, &offset, m0->m_pkthdr.inner_l2hlen);
2833 #endif
2834 
2835 		switch (eh_type) {
2836 #ifdef INET6
2837 		case ETHERTYPE_IPV6:
2838 			m0->m_pkthdr.inner_l3hlen = sizeof(struct ip6_hdr);
2839 			break;
2840 #endif
2841 #ifdef INET
2842 		case ETHERTYPE_IP:
2843 		{
2844 			struct ip *ip = l3hdr;
2845 
2846 			m0->m_pkthdr.inner_l3hlen = ip->ip_hl << 2;
2847 			break;
2848 		}
2849 #endif
2850 		default:
2851 			if (ratecheck(&txerr_ratecheck, &txerr_interval)) {
2852 				log(LOG_ERR, "%s: VXLAN hw offload requested"
2853 				    "with unknown ethertype 0x%04x.  if_cxgbe "
2854 				    "must be compiled with the same INET/INET6 "
2855 				    "options as the kernel.\n", __func__,
2856 				    eh_type);
2857 			}
2858 			rc = EINVAL;
2859 			goto fail;
2860 		}
2861 		if (needs_inner_tcp_csum(m0)) {
2862 			tcp = m_advance(&m, &offset, m0->m_pkthdr.inner_l3hlen);
2863 			m0->m_pkthdr.inner_l4hlen = tcp->th_off * 4;
2864 		}
2865 		MPASS((m0->m_pkthdr.csum_flags & CSUM_SND_TAG) == 0);
2866 		m0->m_pkthdr.csum_flags &= CSUM_INNER_IP6_UDP |
2867 		    CSUM_INNER_IP6_TCP | CSUM_INNER_IP6_TSO | CSUM_INNER_IP |
2868 		    CSUM_INNER_IP_UDP | CSUM_INNER_IP_TCP | CSUM_INNER_IP_TSO |
2869 		    CSUM_ENCAP_VXLAN;
2870 	}
2871 
2872 	if (needs_outer_tcp_csum(m0)) {
2873 		tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen);
2874 		m0->m_pkthdr.l4hlen = tcp->th_off * 4;
2875 #ifdef RATELIMIT
2876 		if (tsclk >= 0 && *(uint32_t *)(tcp + 1) == ntohl(0x0101080a)) {
2877 			set_mbuf_eo_tsclk_tsoff(m0,
2878 			    V_FW_ETH_TX_EO_WR_TSCLK(tsclk) |
2879 			    V_FW_ETH_TX_EO_WR_TSOFF(sizeof(*tcp) / 2 + 1));
2880 		} else
2881 			set_mbuf_eo_tsclk_tsoff(m0, 0);
2882 	} else if (needs_outer_udp_csum(m0)) {
2883 		m0->m_pkthdr.l4hlen = sizeof(struct udphdr);
2884 #endif
2885 	}
2886 #ifdef RATELIMIT
2887 	if (needs_eo(mst)) {
2888 		u_int immhdrs;
2889 
2890 		/* EO WRs have the headers in the WR and not the GL. */
2891 		immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen +
2892 		    m0->m_pkthdr.l4hlen;
2893 		cflags = 0;
2894 		nsegs = count_mbuf_nsegs(m0, immhdrs, &cflags);
2895 		MPASS(cflags == mbuf_cflags(m0));
2896 		set_mbuf_eo_nsegs(m0, nsegs);
2897 		set_mbuf_eo_len16(m0,
2898 		    txpkt_eo_len16(nsegs, immhdrs, needs_tso(m0)));
2899 		rc = ethofld_transmit(mst->ifp, m0);
2900 		if (rc != 0)
2901 			goto fail;
2902 		return (EINPROGRESS);
2903 	}
2904 #endif
2905 #endif
2906 	MPASS(m0 == *mp);
2907 	return (0);
2908 }
2909 
2910 void *
2911 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie)
2912 {
2913 	struct sge_eq *eq = &wrq->eq;
2914 	struct adapter *sc = wrq->adapter;
2915 	int ndesc, available;
2916 	struct wrqe *wr;
2917 	void *w;
2918 
2919 	MPASS(len16 > 0);
2920 	ndesc = tx_len16_to_desc(len16);
2921 	MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC);
2922 
2923 	EQ_LOCK(eq);
2924 	if (__predict_false((eq->flags & EQ_HW_ALLOCATED) == 0)) {
2925 		EQ_UNLOCK(eq);
2926 		return (NULL);
2927 	}
2928 
2929 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2930 		drain_wrq_wr_list(sc, wrq);
2931 
2932 	if (!STAILQ_EMPTY(&wrq->wr_list)) {
2933 slowpath:
2934 		EQ_UNLOCK(eq);
2935 		wr = alloc_wrqe(len16 * 16, wrq);
2936 		if (__predict_false(wr == NULL))
2937 			return (NULL);
2938 		cookie->pidx = -1;
2939 		cookie->ndesc = ndesc;
2940 		return (&wr->wr);
2941 	}
2942 
2943 	eq->cidx = read_hw_cidx(eq);
2944 	if (eq->pidx == eq->cidx)
2945 		available = eq->sidx - 1;
2946 	else
2947 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2948 	if (available < ndesc)
2949 		goto slowpath;
2950 
2951 	cookie->pidx = eq->pidx;
2952 	cookie->ndesc = ndesc;
2953 	TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link);
2954 
2955 	w = &eq->desc[eq->pidx];
2956 	IDXINCR(eq->pidx, ndesc, eq->sidx);
2957 	if (__predict_false(cookie->pidx + ndesc > eq->sidx)) {
2958 		w = &wrq->ss[0];
2959 		wrq->ss_pidx = cookie->pidx;
2960 		wrq->ss_len = len16 * 16;
2961 	}
2962 
2963 	EQ_UNLOCK(eq);
2964 
2965 	return (w);
2966 }
2967 
2968 void
2969 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie)
2970 {
2971 	struct sge_eq *eq = &wrq->eq;
2972 	struct adapter *sc = wrq->adapter;
2973 	int ndesc, pidx;
2974 	struct wrq_cookie *prev, *next;
2975 
2976 	if (cookie->pidx == -1) {
2977 		struct wrqe *wr = __containerof(w, struct wrqe, wr);
2978 
2979 		t4_wrq_tx(sc, wr);
2980 		return;
2981 	}
2982 
2983 	if (__predict_false(w == &wrq->ss[0])) {
2984 		int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE;
2985 
2986 		MPASS(wrq->ss_len > n);	/* WR had better wrap around. */
2987 		bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n);
2988 		bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n);
2989 		wrq->tx_wrs_ss++;
2990 	} else
2991 		wrq->tx_wrs_direct++;
2992 
2993 	EQ_LOCK(eq);
2994 	ndesc = cookie->ndesc;	/* Can be more than SGE_MAX_WR_NDESC here. */
2995 	pidx = cookie->pidx;
2996 	MPASS(pidx >= 0 && pidx < eq->sidx);
2997 	prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link);
2998 	next = TAILQ_NEXT(cookie, link);
2999 	if (prev == NULL) {
3000 		MPASS(pidx == eq->dbidx);
3001 		if (next == NULL || ndesc >= 16) {
3002 			int available;
3003 			struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
3004 
3005 			/*
3006 			 * Note that the WR via which we'll request tx updates
3007 			 * is at pidx and not eq->pidx, which has moved on
3008 			 * already.
3009 			 */
3010 			dst = (void *)&eq->desc[pidx];
3011 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
3012 			if (available < eq->sidx / 4 &&
3013 			    atomic_cmpset_int(&eq->equiq, 0, 1)) {
3014 				/*
3015 				 * XXX: This is not 100% reliable with some
3016 				 * types of WRs.  But this is a very unusual
3017 				 * situation for an ofld/ctrl queue anyway.
3018 				 */
3019 				dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
3020 				    F_FW_WR_EQUEQ);
3021 			}
3022 
3023 			if (__predict_true(eq->flags & EQ_HW_ALLOCATED))
3024 				ring_eq_db(wrq->adapter, eq, ndesc);
3025 			else
3026 				IDXINCR(eq->dbidx, ndesc, eq->sidx);
3027 		} else {
3028 			MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc);
3029 			next->pidx = pidx;
3030 			next->ndesc += ndesc;
3031 		}
3032 	} else {
3033 		MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc);
3034 		prev->ndesc += ndesc;
3035 	}
3036 	TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link);
3037 
3038 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
3039 		drain_wrq_wr_list(sc, wrq);
3040 
3041 #ifdef INVARIANTS
3042 	if (TAILQ_EMPTY(&wrq->incomplete_wrs)) {
3043 		/* Doorbell must have caught up to the pidx. */
3044 		MPASS(wrq->eq.pidx == wrq->eq.dbidx);
3045 	}
3046 #endif
3047 	EQ_UNLOCK(eq);
3048 }
3049 
3050 static u_int
3051 can_resume_eth_tx(struct mp_ring *r)
3052 {
3053 	struct sge_eq *eq = r->cookie;
3054 
3055 	return (total_available_tx_desc(eq) > eq->sidx / 8);
3056 }
3057 
3058 static inline bool
3059 cannot_use_txpkts(struct mbuf *m)
3060 {
3061 	/* maybe put a GL limit too, to avoid silliness? */
3062 
3063 	return (needs_tso(m) || (mbuf_cflags(m) & (MC_RAW_WR | MC_TLS)) != 0);
3064 }
3065 
3066 static inline int
3067 discard_tx(struct sge_eq *eq)
3068 {
3069 
3070 	return ((eq->flags & (EQ_ENABLED | EQ_QFLUSH)) != EQ_ENABLED);
3071 }
3072 
3073 static inline int
3074 wr_can_update_eq(void *p)
3075 {
3076 	struct fw_eth_tx_pkts_wr *wr = p;
3077 
3078 	switch (G_FW_WR_OP(be32toh(wr->op_pkd))) {
3079 	case FW_ULPTX_WR:
3080 	case FW_ETH_TX_PKT_WR:
3081 	case FW_ETH_TX_PKTS_WR:
3082 	case FW_ETH_TX_PKTS2_WR:
3083 	case FW_ETH_TX_PKT_VM_WR:
3084 	case FW_ETH_TX_PKTS_VM_WR:
3085 		return (1);
3086 	default:
3087 		return (0);
3088 	}
3089 }
3090 
3091 static inline void
3092 set_txupdate_flags(struct sge_txq *txq, u_int avail,
3093     struct fw_eth_tx_pkt_wr *wr)
3094 {
3095 	struct sge_eq *eq = &txq->eq;
3096 	struct txpkts *txp = &txq->txp;
3097 
3098 	if ((txp->npkt > 0 || avail < eq->sidx / 2) &&
3099 	    atomic_cmpset_int(&eq->equiq, 0, 1)) {
3100 		wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ | F_FW_WR_EQUIQ);
3101 		eq->equeqidx = eq->pidx;
3102 	} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
3103 		wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
3104 		eq->equeqidx = eq->pidx;
3105 	}
3106 }
3107 
3108 #if defined(__i386__) || defined(__amd64__)
3109 extern uint64_t tsc_freq;
3110 #endif
3111 
3112 static inline bool
3113 record_eth_tx_time(struct sge_txq *txq)
3114 {
3115 	const uint64_t cycles = get_cyclecount();
3116 	const uint64_t last_tx = txq->last_tx;
3117 #if defined(__i386__) || defined(__amd64__)
3118 	const uint64_t itg = tsc_freq * t4_tx_coalesce_gap / 1000000;
3119 #else
3120 	const uint64_t itg = 0;
3121 #endif
3122 
3123 	MPASS(cycles >= last_tx);
3124 	txq->last_tx = cycles;
3125 	return (cycles - last_tx < itg);
3126 }
3127 
3128 /*
3129  * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to
3130  * be consumed.  Return the actual number consumed.  0 indicates a stall.
3131  */
3132 static u_int
3133 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx, bool *coalescing)
3134 {
3135 	struct sge_txq *txq = r->cookie;
3136 	if_t ifp = txq->ifp;
3137 	struct sge_eq *eq = &txq->eq;
3138 	struct txpkts *txp = &txq->txp;
3139 	struct vi_info *vi = if_getsoftc(ifp);
3140 	struct adapter *sc = vi->adapter;
3141 	u_int total, remaining;		/* # of packets */
3142 	u_int n, avail, dbdiff;		/* # of hardware descriptors */
3143 	int i, rc;
3144 	struct mbuf *m0;
3145 	bool snd, recent_tx;
3146 	void *wr;	/* start of the last WR written to the ring */
3147 
3148 	TXQ_LOCK_ASSERT_OWNED(txq);
3149 	recent_tx = record_eth_tx_time(txq);
3150 
3151 	remaining = IDXDIFF(pidx, cidx, r->size);
3152 	if (__predict_false(discard_tx(eq))) {
3153 		for (i = 0; i < txp->npkt; i++)
3154 			m_freem(txp->mb[i]);
3155 		txp->npkt = 0;
3156 		while (cidx != pidx) {
3157 			m0 = r->items[cidx];
3158 			m_freem(m0);
3159 			if (++cidx == r->size)
3160 				cidx = 0;
3161 		}
3162 		reclaim_tx_descs(txq, eq->sidx);
3163 		*coalescing = false;
3164 		return (remaining);	/* emptied */
3165 	}
3166 
3167 	/* How many hardware descriptors do we have readily available. */
3168 	if (eq->pidx == eq->cidx)
3169 		avail = eq->sidx - 1;
3170 	else
3171 		avail = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
3172 
3173 	total = 0;
3174 	if (remaining == 0) {
3175 		txp->score = 0;
3176 		txq->txpkts_flush++;
3177 		goto send_txpkts;
3178 	}
3179 
3180 	dbdiff = 0;
3181 	MPASS(remaining > 0);
3182 	while (remaining > 0) {
3183 		m0 = r->items[cidx];
3184 		M_ASSERTPKTHDR(m0);
3185 		MPASS(m0->m_nextpkt == NULL);
3186 
3187 		if (avail < 2 * SGE_MAX_WR_NDESC)
3188 			avail += reclaim_tx_descs(txq, 64);
3189 
3190 		if (t4_tx_coalesce == 0 && txp->npkt == 0)
3191 			goto skip_coalescing;
3192 		if (cannot_use_txpkts(m0))
3193 			txp->score = 0;
3194 		else if (recent_tx) {
3195 			if (++txp->score == 0)
3196 				txp->score = UINT8_MAX;
3197 		} else
3198 			txp->score = 1;
3199 		if (txp->npkt > 0 || remaining > 1 ||
3200 		    txp->score >= t4_tx_coalesce_pkts ||
3201 		    atomic_load_int(&txq->eq.equiq) != 0) {
3202 			if (vi->flags & TX_USES_VM_WR)
3203 				rc = add_to_txpkts_vf(sc, txq, m0, avail, &snd);
3204 			else
3205 				rc = add_to_txpkts_pf(sc, txq, m0, avail, &snd);
3206 		} else {
3207 			snd = false;
3208 			rc = EINVAL;
3209 		}
3210 		if (snd) {
3211 			MPASS(txp->npkt > 0);
3212 			for (i = 0; i < txp->npkt; i++)
3213 				ETHER_BPF_MTAP(ifp, txp->mb[i]);
3214 			if (txp->npkt > 1) {
3215 				MPASS(avail >= tx_len16_to_desc(txp->len16));
3216 				if (vi->flags & TX_USES_VM_WR)
3217 					n = write_txpkts_vm_wr(sc, txq);
3218 				else
3219 					n = write_txpkts_wr(sc, txq);
3220 			} else {
3221 				MPASS(avail >=
3222 				    tx_len16_to_desc(mbuf_len16(txp->mb[0])));
3223 				if (vi->flags & TX_USES_VM_WR)
3224 					n = write_txpkt_vm_wr(sc, txq,
3225 					    txp->mb[0]);
3226 				else
3227 					n = write_txpkt_wr(sc, txq, txp->mb[0],
3228 					    avail);
3229 			}
3230 			MPASS(n <= SGE_MAX_WR_NDESC);
3231 			avail -= n;
3232 			dbdiff += n;
3233 			wr = &eq->desc[eq->pidx];
3234 			IDXINCR(eq->pidx, n, eq->sidx);
3235 			txp->npkt = 0;	/* emptied */
3236 		}
3237 		if (rc == 0) {
3238 			/* m0 was coalesced into txq->txpkts. */
3239 			goto next_mbuf;
3240 		}
3241 		if (rc == EAGAIN) {
3242 			/*
3243 			 * m0 is suitable for tx coalescing but could not be
3244 			 * combined with the existing txq->txpkts, which has now
3245 			 * been transmitted.  Start a new txpkts with m0.
3246 			 */
3247 			MPASS(snd);
3248 			MPASS(txp->npkt == 0);
3249 			continue;
3250 		}
3251 
3252 		MPASS(rc != 0 && rc != EAGAIN);
3253 		MPASS(txp->npkt == 0);
3254 skip_coalescing:
3255 		n = tx_len16_to_desc(mbuf_len16(m0));
3256 		if (__predict_false(avail < n)) {
3257 			avail += reclaim_tx_descs(txq, min(n, 32));
3258 			if (avail < n)
3259 				break;	/* out of descriptors */
3260 		}
3261 
3262 		wr = &eq->desc[eq->pidx];
3263 		if (mbuf_cflags(m0) & MC_RAW_WR) {
3264 			n = write_raw_wr(txq, wr, m0, avail);
3265 #ifdef KERN_TLS
3266 		} else if (mbuf_cflags(m0) & MC_TLS) {
3267 			ETHER_BPF_MTAP(ifp, m0);
3268 			n = t6_ktls_write_wr(txq, wr, m0, avail);
3269 #endif
3270 		} else {
3271 			ETHER_BPF_MTAP(ifp, m0);
3272 			if (vi->flags & TX_USES_VM_WR)
3273 				n = write_txpkt_vm_wr(sc, txq, m0);
3274 			else
3275 				n = write_txpkt_wr(sc, txq, m0, avail);
3276 		}
3277 		MPASS(n >= 1 && n <= avail);
3278 		if (!(mbuf_cflags(m0) & MC_TLS))
3279 			MPASS(n <= SGE_MAX_WR_NDESC);
3280 
3281 		avail -= n;
3282 		dbdiff += n;
3283 		IDXINCR(eq->pidx, n, eq->sidx);
3284 
3285 		if (dbdiff >= 512 / EQ_ESIZE) {	/* X_FETCHBURSTMAX_512B */
3286 			if (wr_can_update_eq(wr))
3287 				set_txupdate_flags(txq, avail, wr);
3288 			ring_eq_db(sc, eq, dbdiff);
3289 			avail += reclaim_tx_descs(txq, 32);
3290 			dbdiff = 0;
3291 		}
3292 next_mbuf:
3293 		total++;
3294 		remaining--;
3295 		if (__predict_false(++cidx == r->size))
3296 			cidx = 0;
3297 	}
3298 	if (dbdiff != 0) {
3299 		if (wr_can_update_eq(wr))
3300 			set_txupdate_flags(txq, avail, wr);
3301 		ring_eq_db(sc, eq, dbdiff);
3302 		reclaim_tx_descs(txq, 32);
3303 	} else if (eq->pidx == eq->cidx && txp->npkt > 0 &&
3304 	    atomic_load_int(&txq->eq.equiq) == 0) {
3305 		/*
3306 		 * If nothing was submitted to the chip for tx (it was coalesced
3307 		 * into txpkts instead) and there is no tx update outstanding
3308 		 * then we need to send txpkts now.
3309 		 */
3310 send_txpkts:
3311 		MPASS(txp->npkt > 0);
3312 		for (i = 0; i < txp->npkt; i++)
3313 			ETHER_BPF_MTAP(ifp, txp->mb[i]);
3314 		if (txp->npkt > 1) {
3315 			MPASS(avail >= tx_len16_to_desc(txp->len16));
3316 			if (vi->flags & TX_USES_VM_WR)
3317 				n = write_txpkts_vm_wr(sc, txq);
3318 			else
3319 				n = write_txpkts_wr(sc, txq);
3320 		} else {
3321 			MPASS(avail >=
3322 			    tx_len16_to_desc(mbuf_len16(txp->mb[0])));
3323 			if (vi->flags & TX_USES_VM_WR)
3324 				n = write_txpkt_vm_wr(sc, txq, txp->mb[0]);
3325 			else
3326 				n = write_txpkt_wr(sc, txq, txp->mb[0], avail);
3327 		}
3328 		MPASS(n <= SGE_MAX_WR_NDESC);
3329 		wr = &eq->desc[eq->pidx];
3330 		IDXINCR(eq->pidx, n, eq->sidx);
3331 		txp->npkt = 0;	/* emptied */
3332 
3333 		MPASS(wr_can_update_eq(wr));
3334 		set_txupdate_flags(txq, avail - n, wr);
3335 		ring_eq_db(sc, eq, n);
3336 		reclaim_tx_descs(txq, 32);
3337 	}
3338 	*coalescing = txp->npkt > 0;
3339 
3340 	return (total);
3341 }
3342 
3343 static inline void
3344 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
3345     int qsize, int intr_idx, int cong, int qtype)
3346 {
3347 
3348 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
3349 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
3350 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
3351 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
3352 	KASSERT(intr_idx >= -1 && intr_idx < sc->intr_count,
3353 	    ("%s: bad intr_idx %d", __func__, intr_idx));
3354 	KASSERT(qtype == FW_IQ_IQTYPE_OTHER || qtype == FW_IQ_IQTYPE_NIC ||
3355 	    qtype == FW_IQ_IQTYPE_OFLD, ("%s: bad qtype %d", __func__, qtype));
3356 
3357 	iq->flags = 0;
3358 	iq->state = IQS_DISABLED;
3359 	iq->adapter = sc;
3360 	iq->qtype = qtype;
3361 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
3362 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
3363 	if (pktc_idx >= 0) {
3364 		iq->intr_params |= F_QINTR_CNT_EN;
3365 		iq->intr_pktc_idx = pktc_idx;
3366 	}
3367 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
3368 	iq->sidx = iq->qsize - sc->params.sge.spg_len / IQ_ESIZE;
3369 	iq->intr_idx = intr_idx;
3370 	iq->cong_drop = cong;
3371 }
3372 
3373 static inline void
3374 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name)
3375 {
3376 	struct sge_params *sp = &sc->params.sge;
3377 
3378 	fl->qsize = qsize;
3379 	fl->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
3380 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
3381 	mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
3382 	if (sc->flags & BUF_PACKING_OK &&
3383 	    ((!is_t4(sc) && buffer_packing) ||	/* T5+: enabled unless 0 */
3384 	    (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */
3385 		fl->flags |= FL_BUF_PACKING;
3386 	fl->zidx = find_refill_source(sc, maxp, fl->flags & FL_BUF_PACKING);
3387 	fl->safe_zidx = sc->sge.safe_zidx;
3388 	if (fl->flags & FL_BUF_PACKING) {
3389 		fl->lowat = roundup2(sp->fl_starve_threshold2, 8);
3390 		fl->buf_boundary = sp->pack_boundary;
3391 	} else {
3392 		fl->lowat = roundup2(sp->fl_starve_threshold, 8);
3393 		fl->buf_boundary = 16;
3394 	}
3395 	if (fl_pad && fl->buf_boundary < sp->pad_boundary)
3396 		fl->buf_boundary = sp->pad_boundary;
3397 }
3398 
3399 static inline void
3400 init_eq(struct adapter *sc, struct sge_eq *eq, int eqtype, int qsize,
3401     uint8_t port_id, struct sge_iq *iq, char *name)
3402 {
3403 	KASSERT(eqtype >= EQ_CTRL && eqtype <= EQ_OFLD,
3404 	    ("%s: bad qtype %d", __func__, eqtype));
3405 
3406 	eq->type = eqtype;
3407 	eq->port_id = port_id;
3408 	eq->tx_chan = sc->port[port_id]->tx_chan;
3409 	eq->iq = iq;
3410 	eq->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
3411 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
3412 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
3413 }
3414 
3415 int
3416 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
3417     bus_dmamap_t *map, bus_addr_t *pa, void **va)
3418 {
3419 	int rc;
3420 
3421 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
3422 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
3423 	if (rc != 0) {
3424 		CH_ERR(sc, "cannot allocate DMA tag: %d\n", rc);
3425 		goto done;
3426 	}
3427 
3428 	rc = bus_dmamem_alloc(*tag, va,
3429 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
3430 	if (rc != 0) {
3431 		CH_ERR(sc, "cannot allocate DMA memory: %d\n", rc);
3432 		goto done;
3433 	}
3434 
3435 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
3436 	if (rc != 0) {
3437 		CH_ERR(sc, "cannot load DMA map: %d\n", rc);
3438 		goto done;
3439 	}
3440 done:
3441 	if (rc)
3442 		free_ring(sc, *tag, *map, *pa, *va);
3443 
3444 	return (rc);
3445 }
3446 
3447 int
3448 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
3449     bus_addr_t pa, void *va)
3450 {
3451 	if (pa)
3452 		bus_dmamap_unload(tag, map);
3453 	if (va)
3454 		bus_dmamem_free(tag, va, map);
3455 	if (tag)
3456 		bus_dma_tag_destroy(tag);
3457 
3458 	return (0);
3459 }
3460 
3461 /*
3462  * Allocates the software resources (mainly memory and sysctl nodes) for an
3463  * ingress queue and an optional freelist.
3464  *
3465  * Sets IQ_SW_ALLOCATED and returns 0 on success.
3466  */
3467 static int
3468 alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl,
3469     struct sysctl_ctx_list *ctx, struct sysctl_oid *oid)
3470 {
3471 	int rc;
3472 	size_t len;
3473 	struct adapter *sc = vi->adapter;
3474 
3475 	MPASS(!(iq->flags & IQ_SW_ALLOCATED));
3476 
3477 	len = iq->qsize * IQ_ESIZE;
3478 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
3479 	    (void **)&iq->desc);
3480 	if (rc != 0)
3481 		return (rc);
3482 
3483 	if (fl) {
3484 		len = fl->qsize * EQ_ESIZE;
3485 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
3486 		    &fl->ba, (void **)&fl->desc);
3487 		if (rc) {
3488 			free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba,
3489 			    iq->desc);
3490 			return (rc);
3491 		}
3492 
3493 		/* Allocate space for one software descriptor per buffer. */
3494 		fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc),
3495 		    M_CXGBE, M_ZERO | M_WAITOK);
3496 
3497 		add_fl_sysctls(sc, ctx, oid, fl);
3498 		iq->flags |= IQ_HAS_FL;
3499 	}
3500 	add_iq_sysctls(ctx, oid, iq);
3501 	iq->flags |= IQ_SW_ALLOCATED;
3502 
3503 	return (0);
3504 }
3505 
3506 /*
3507  * Frees all software resources (memory and locks) associated with an ingress
3508  * queue and an optional freelist.
3509  */
3510 static void
3511 free_iq_fl(struct adapter *sc, struct sge_iq *iq, struct sge_fl *fl)
3512 {
3513 	MPASS(iq->flags & IQ_SW_ALLOCATED);
3514 
3515 	if (fl) {
3516 		MPASS(iq->flags & IQ_HAS_FL);
3517 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba, fl->desc);
3518 		free_fl_buffers(sc, fl);
3519 		free(fl->sdesc, M_CXGBE);
3520 		mtx_destroy(&fl->fl_lock);
3521 		bzero(fl, sizeof(*fl));
3522 	}
3523 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
3524 	bzero(iq, sizeof(*iq));
3525 }
3526 
3527 /*
3528  * Allocates a hardware ingress queue and an optional freelist that will be
3529  * associated with it.
3530  *
3531  * Returns errno on failure.  Resources allocated up to that point may still be
3532  * allocated.  Caller is responsible for cleanup in case this function fails.
3533  */
3534 static int
3535 alloc_iq_fl_hwq(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl)
3536 {
3537 	int rc, cntxt_id, cong_map;
3538 	struct fw_iq_cmd c;
3539 	struct adapter *sc = vi->adapter;
3540 	struct port_info *pi = vi->pi;
3541 	__be32 v = 0;
3542 
3543 	MPASS (!(iq->flags & IQ_HW_ALLOCATED));
3544 
3545 	bzero(&c, sizeof(c));
3546 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
3547 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
3548 	    V_FW_IQ_CMD_VFN(0));
3549 
3550 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
3551 	    FW_LEN16(c));
3552 
3553 	/* Special handling for firmware event queue */
3554 	if (iq == &sc->sge.fwq)
3555 		v |= F_FW_IQ_CMD_IQASYNCH;
3556 
3557 	if (iq->intr_idx < 0) {
3558 		/* Forwarded interrupts, all headed to fwq */
3559 		v |= F_FW_IQ_CMD_IQANDST;
3560 		v |= V_FW_IQ_CMD_IQANDSTINDEX(sc->sge.fwq.cntxt_id);
3561 	} else {
3562 		KASSERT(iq->intr_idx < sc->intr_count,
3563 		    ("%s: invalid direct intr_idx %d", __func__, iq->intr_idx));
3564 		v |= V_FW_IQ_CMD_IQANDSTINDEX(iq->intr_idx);
3565 	}
3566 
3567 	bzero(iq->desc, iq->qsize * IQ_ESIZE);
3568 	c.type_to_iqandstindex = htobe32(v |
3569 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
3570 	    V_FW_IQ_CMD_VIID(vi->viid) |
3571 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
3572 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
3573 	    F_FW_IQ_CMD_IQGTSMODE |
3574 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
3575 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
3576 	c.iqsize = htobe16(iq->qsize);
3577 	c.iqaddr = htobe64(iq->ba);
3578 	c.iqns_to_fl0congen = htobe32(V_FW_IQ_CMD_IQTYPE(iq->qtype));
3579 	if (iq->cong_drop != -1) {
3580 		cong_map = iq->qtype == IQ_ETH ? pi->rx_e_chan_map : 0;
3581 		c.iqns_to_fl0congen |= htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
3582 	}
3583 
3584 	if (fl) {
3585 		bzero(fl->desc, fl->sidx * EQ_ESIZE + sc->params.sge.spg_len);
3586 		c.iqns_to_fl0congen |=
3587 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
3588 			F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
3589 			(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
3590 			(fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
3591 			    0));
3592 		if (iq->cong_drop != -1) {
3593 			c.iqns_to_fl0congen |=
3594 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong_map) |
3595 				    F_FW_IQ_CMD_FL0CONGCIF |
3596 				    F_FW_IQ_CMD_FL0CONGEN);
3597 		}
3598 		c.fl0dcaen_to_fl0cidxfthresh =
3599 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
3600 			X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B_T6) |
3601 			V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
3602 			X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
3603 		c.fl0size = htobe16(fl->qsize);
3604 		c.fl0addr = htobe64(fl->ba);
3605 	}
3606 
3607 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3608 	if (rc != 0) {
3609 		CH_ERR(sc, "failed to create hw ingress queue: %d\n", rc);
3610 		return (rc);
3611 	}
3612 
3613 	iq->cidx = 0;
3614 	iq->gen = F_RSPD_GEN;
3615 	iq->cntxt_id = be16toh(c.iqid);
3616 	iq->abs_id = be16toh(c.physiqid);
3617 
3618 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
3619 	if (cntxt_id >= sc->sge.iqmap_sz) {
3620 		panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
3621 		    cntxt_id, sc->sge.iqmap_sz - 1);
3622 	}
3623 	sc->sge.iqmap[cntxt_id] = iq;
3624 
3625 	if (fl) {
3626 		u_int qid;
3627 #ifdef INVARIANTS
3628 		int i;
3629 
3630 		MPASS(!(fl->flags & FL_BUF_RESUME));
3631 		for (i = 0; i < fl->sidx * 8; i++)
3632 			MPASS(fl->sdesc[i].cl == NULL);
3633 #endif
3634 		fl->cntxt_id = be16toh(c.fl0id);
3635 		fl->pidx = fl->cidx = fl->hw_cidx = fl->dbidx = 0;
3636 		fl->rx_offset = 0;
3637 		fl->flags &= ~(FL_STARVING | FL_DOOMED);
3638 
3639 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
3640 		if (cntxt_id >= sc->sge.eqmap_sz) {
3641 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
3642 			    __func__, cntxt_id, sc->sge.eqmap_sz - 1);
3643 		}
3644 		sc->sge.eqmap[cntxt_id] = (void *)fl;
3645 
3646 		qid = fl->cntxt_id;
3647 		if (isset(&sc->doorbells, DOORBELL_UDB)) {
3648 			uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3649 			uint32_t mask = (1 << s_qpp) - 1;
3650 			volatile uint8_t *udb;
3651 
3652 			udb = sc->udbs_base + UDBS_DB_OFFSET;
3653 			udb += (qid >> s_qpp) << PAGE_SHIFT;
3654 			qid &= mask;
3655 			if (qid < PAGE_SIZE / UDBS_SEG_SIZE) {
3656 				udb += qid << UDBS_SEG_SHIFT;
3657 				qid = 0;
3658 			}
3659 			fl->udb = (volatile void *)udb;
3660 		}
3661 		fl->dbval = V_QID(qid) | sc->chip_params->sge_fl_db;
3662 
3663 		FL_LOCK(fl);
3664 		/* Enough to make sure the SGE doesn't think it's starved */
3665 		refill_fl(sc, fl, fl->lowat);
3666 		FL_UNLOCK(fl);
3667 	}
3668 
3669 	if (chip_id(sc) >= CHELSIO_T5 && !(sc->flags & IS_VF) &&
3670 	    iq->cong_drop != -1) {
3671 		t4_sge_set_conm_context(sc, iq->cntxt_id, iq->cong_drop,
3672 		    cong_map);
3673 	}
3674 
3675 	/* Enable IQ interrupts */
3676 	atomic_store_rel_int(&iq->state, IQS_IDLE);
3677 	t4_write_reg(sc, sc->sge_gts_reg, V_SEINTARM(iq->intr_params) |
3678 	    V_INGRESSQID(iq->cntxt_id));
3679 
3680 	iq->flags |= IQ_HW_ALLOCATED;
3681 
3682 	return (0);
3683 }
3684 
3685 static int
3686 free_iq_fl_hwq(struct adapter *sc, struct sge_iq *iq, struct sge_fl *fl)
3687 {
3688 	int rc;
3689 
3690 	MPASS(iq->flags & IQ_HW_ALLOCATED);
3691 	rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0, FW_IQ_TYPE_FL_INT_CAP,
3692 	    iq->cntxt_id, fl ? fl->cntxt_id : 0xffff, 0xffff);
3693 	if (rc != 0) {
3694 		CH_ERR(sc, "failed to free iq %p: %d\n", iq, rc);
3695 		return (rc);
3696 	}
3697 	iq->flags &= ~IQ_HW_ALLOCATED;
3698 
3699 	return (0);
3700 }
3701 
3702 static void
3703 add_iq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
3704     struct sge_iq *iq)
3705 {
3706 	struct sysctl_oid_list *children;
3707 
3708 	if (ctx == NULL || oid == NULL)
3709 		return;
3710 
3711 	children = SYSCTL_CHILDREN(oid);
3712 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD, &iq->ba,
3713 	    "bus address of descriptor ring");
3714 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3715 	    iq->qsize * IQ_ESIZE, "descriptor ring size in bytes");
3716 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
3717 	    &iq->abs_id, 0, "absolute id of the queue");
3718 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3719 	    &iq->cntxt_id, 0, "SGE context id of the queue");
3720 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &iq->cidx,
3721 	    0, "consumer index");
3722 }
3723 
3724 static void
3725 add_fl_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
3726     struct sysctl_oid *oid, struct sge_fl *fl)
3727 {
3728 	struct sysctl_oid_list *children;
3729 
3730 	if (ctx == NULL || oid == NULL)
3731 		return;
3732 
3733 	children = SYSCTL_CHILDREN(oid);
3734 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl",
3735 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist");
3736 	children = SYSCTL_CHILDREN(oid);
3737 
3738 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3739 	    &fl->ba, "bus address of descriptor ring");
3740 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3741 	    fl->sidx * EQ_ESIZE + sc->params.sge.spg_len,
3742 	    "desc ring size in bytes");
3743 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3744 	    &fl->cntxt_id, 0, "SGE context id of the freelist");
3745 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL,
3746 	    fl_pad ? 1 : 0, "padding enabled");
3747 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL,
3748 	    fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled");
3749 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
3750 	    0, "consumer index");
3751 	if (fl->flags & FL_BUF_PACKING) {
3752 		SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
3753 		    CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
3754 	}
3755 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
3756 	    0, "producer index");
3757 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
3758 	    CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
3759 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
3760 	    CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
3761 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
3762 	    CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
3763 }
3764 
3765 /*
3766  * Idempotent.
3767  */
3768 static int
3769 alloc_fwq(struct adapter *sc)
3770 {
3771 	int rc, intr_idx;
3772 	struct sge_iq *fwq = &sc->sge.fwq;
3773 	struct vi_info *vi = &sc->port[0]->vi[0];
3774 
3775 	if (!(fwq->flags & IQ_SW_ALLOCATED)) {
3776 		MPASS(!(fwq->flags & IQ_HW_ALLOCATED));
3777 
3778 		if (sc->flags & IS_VF)
3779 			intr_idx = 0;
3780 		else
3781 			intr_idx = sc->intr_count > 1 ? 1 : 0;
3782 		init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE, intr_idx, -1, IQ_OTHER);
3783 		rc = alloc_iq_fl(vi, fwq, NULL, &sc->ctx, sc->fwq_oid);
3784 		if (rc != 0) {
3785 			CH_ERR(sc, "failed to allocate fwq: %d\n", rc);
3786 			return (rc);
3787 		}
3788 		MPASS(fwq->flags & IQ_SW_ALLOCATED);
3789 	}
3790 
3791 	if (!(fwq->flags & IQ_HW_ALLOCATED)) {
3792 		MPASS(fwq->flags & IQ_SW_ALLOCATED);
3793 
3794 		rc = alloc_iq_fl_hwq(vi, fwq, NULL);
3795 		if (rc != 0) {
3796 			CH_ERR(sc, "failed to create hw fwq: %d\n", rc);
3797 			return (rc);
3798 		}
3799 		MPASS(fwq->flags & IQ_HW_ALLOCATED);
3800 	}
3801 
3802 	return (0);
3803 }
3804 
3805 /*
3806  * Idempotent.
3807  */
3808 static void
3809 free_fwq(struct adapter *sc)
3810 {
3811 	struct sge_iq *fwq = &sc->sge.fwq;
3812 
3813 	if (fwq->flags & IQ_HW_ALLOCATED) {
3814 		MPASS(fwq->flags & IQ_SW_ALLOCATED);
3815 		free_iq_fl_hwq(sc, fwq, NULL);
3816 		MPASS(!(fwq->flags & IQ_HW_ALLOCATED));
3817 	}
3818 
3819 	if (fwq->flags & IQ_SW_ALLOCATED) {
3820 		MPASS(!(fwq->flags & IQ_HW_ALLOCATED));
3821 		free_iq_fl(sc, fwq, NULL);
3822 		MPASS(!(fwq->flags & IQ_SW_ALLOCATED));
3823 	}
3824 }
3825 
3826 /*
3827  * Idempotent.
3828  */
3829 static int
3830 alloc_ctrlq(struct adapter *sc, int idx)
3831 {
3832 	int rc;
3833 	char name[16];
3834 	struct sysctl_oid *oid;
3835 	struct sge_wrq *ctrlq = &sc->sge.ctrlq[idx];
3836 
3837 	MPASS(idx < sc->params.nports);
3838 
3839 	if (!(ctrlq->eq.flags & EQ_SW_ALLOCATED)) {
3840 		MPASS(!(ctrlq->eq.flags & EQ_HW_ALLOCATED));
3841 
3842 		snprintf(name, sizeof(name), "%d", idx);
3843 		oid = SYSCTL_ADD_NODE(&sc->ctx, SYSCTL_CHILDREN(sc->ctrlq_oid),
3844 		    OID_AUTO, name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
3845 		    "ctrl queue");
3846 
3847 		snprintf(name, sizeof(name), "%s ctrlq%d",
3848 		    device_get_nameunit(sc->dev), idx);
3849 		init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, idx,
3850 		    &sc->sge.fwq, name);
3851 		rc = alloc_wrq(sc, NULL, ctrlq, &sc->ctx, oid);
3852 		if (rc != 0) {
3853 			CH_ERR(sc, "failed to allocate ctrlq%d: %d\n", idx, rc);
3854 			sysctl_remove_oid(oid, 1, 1);
3855 			return (rc);
3856 		}
3857 		MPASS(ctrlq->eq.flags & EQ_SW_ALLOCATED);
3858 	}
3859 
3860 	if (!(ctrlq->eq.flags & EQ_HW_ALLOCATED)) {
3861 		MPASS(ctrlq->eq.flags & EQ_SW_ALLOCATED);
3862 		MPASS(ctrlq->nwr_pending == 0);
3863 		MPASS(ctrlq->ndesc_needed == 0);
3864 
3865 		rc = alloc_eq_hwq(sc, NULL, &ctrlq->eq);
3866 		if (rc != 0) {
3867 			CH_ERR(sc, "failed to create hw ctrlq%d: %d\n", idx, rc);
3868 			return (rc);
3869 		}
3870 		MPASS(ctrlq->eq.flags & EQ_HW_ALLOCATED);
3871 	}
3872 
3873 	return (0);
3874 }
3875 
3876 /*
3877  * Idempotent.
3878  */
3879 static void
3880 free_ctrlq(struct adapter *sc, int idx)
3881 {
3882 	struct sge_wrq *ctrlq = &sc->sge.ctrlq[idx];
3883 
3884 	if (ctrlq->eq.flags & EQ_HW_ALLOCATED) {
3885 		MPASS(ctrlq->eq.flags & EQ_SW_ALLOCATED);
3886 		free_eq_hwq(sc, NULL, &ctrlq->eq);
3887 		MPASS(!(ctrlq->eq.flags & EQ_HW_ALLOCATED));
3888 	}
3889 
3890 	if (ctrlq->eq.flags & EQ_SW_ALLOCATED) {
3891 		MPASS(!(ctrlq->eq.flags & EQ_HW_ALLOCATED));
3892 		free_wrq(sc, ctrlq);
3893 		MPASS(!(ctrlq->eq.flags & EQ_SW_ALLOCATED));
3894 	}
3895 }
3896 
3897 int
3898 t4_sge_set_conm_context(struct adapter *sc, int cntxt_id, int cong_drop,
3899     int cong_map)
3900 {
3901 	const int cng_ch_bits_log = sc->chip_params->cng_ch_bits_log;
3902 	uint32_t param, val;
3903 	uint16_t ch_map;
3904 	int cong_mode, rc, i;
3905 
3906 	if (chip_id(sc) < CHELSIO_T5)
3907 		return (ENOTSUP);
3908 
3909 	/* Convert the driver knob to the mode understood by the firmware. */
3910 	switch (cong_drop) {
3911 	case -1:
3912 		cong_mode = X_CONMCTXT_CNGTPMODE_DISABLE;
3913 		break;
3914 	case 0:
3915 		cong_mode = X_CONMCTXT_CNGTPMODE_CHANNEL;
3916 		break;
3917 	case 1:
3918 		cong_mode = X_CONMCTXT_CNGTPMODE_QUEUE;
3919 		break;
3920 	case 2:
3921 		cong_mode = X_CONMCTXT_CNGTPMODE_BOTH;
3922 		break;
3923 	default:
3924 		MPASS(0);
3925 		CH_ERR(sc, "cong_drop = %d is invalid (ingress queue %d).\n",
3926 		    cong_drop, cntxt_id);
3927 		return (EINVAL);
3928 	}
3929 
3930 	param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
3931 	    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
3932 	    V_FW_PARAMS_PARAM_YZ(cntxt_id);
3933 	val = V_CONMCTXT_CNGTPMODE(cong_mode);
3934 	if (cong_mode == X_CONMCTXT_CNGTPMODE_CHANNEL ||
3935 	    cong_mode == X_CONMCTXT_CNGTPMODE_BOTH) {
3936 		for (i = 0, ch_map = 0; i < 4; i++) {
3937 			if (cong_map & (1 << i))
3938 				ch_map |= 1 << (i << cng_ch_bits_log);
3939 		}
3940 		val |= V_CONMCTXT_CNGCHMAP(ch_map);
3941 	}
3942 	rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
3943 	if (rc != 0) {
3944 		CH_ERR(sc, "failed to set congestion manager context "
3945 		    "for ingress queue %d: %d\n", cntxt_id, rc);
3946 	}
3947 
3948 	return (rc);
3949 }
3950 
3951 /*
3952  * Idempotent.
3953  */
3954 static int
3955 alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int idx, int intr_idx,
3956     int maxp)
3957 {
3958 	int rc;
3959 	struct adapter *sc = vi->adapter;
3960 	if_t ifp = vi->ifp;
3961 	struct sysctl_oid *oid;
3962 	char name[16];
3963 
3964 	if (!(rxq->iq.flags & IQ_SW_ALLOCATED)) {
3965 		MPASS(!(rxq->iq.flags & IQ_HW_ALLOCATED));
3966 #if defined(INET) || defined(INET6)
3967 		rc = tcp_lro_init_args(&rxq->lro, ifp, lro_entries, lro_mbufs);
3968 		if (rc != 0)
3969 			return (rc);
3970 		MPASS(rxq->lro.ifp == ifp);	/* also indicates LRO init'ed */
3971 #endif
3972 		rxq->ifp = ifp;
3973 
3974 		snprintf(name, sizeof(name), "%d", idx);
3975 		oid = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(vi->rxq_oid),
3976 		    OID_AUTO, name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
3977 		    "rx queue");
3978 
3979 		init_iq(&rxq->iq, sc, vi->tmr_idx, vi->pktc_idx, vi->qsize_rxq,
3980 		    intr_idx, cong_drop, IQ_ETH);
3981 #if defined(INET) || defined(INET6)
3982 		if (if_getcapenable(ifp) & IFCAP_LRO)
3983 			rxq->iq.flags |= IQ_LRO_ENABLED;
3984 #endif
3985 		if (if_getcapenable(ifp) & IFCAP_HWRXTSTMP)
3986 			rxq->iq.flags |= IQ_RX_TIMESTAMP;
3987 		snprintf(name, sizeof(name), "%s rxq%d-fl",
3988 		    device_get_nameunit(vi->dev), idx);
3989 		init_fl(sc, &rxq->fl, vi->qsize_rxq / 8, maxp, name);
3990 		rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, &vi->ctx, oid);
3991 		if (rc != 0) {
3992 			CH_ERR(vi, "failed to allocate rxq%d: %d\n", idx, rc);
3993 			sysctl_remove_oid(oid, 1, 1);
3994 #if defined(INET) || defined(INET6)
3995 			tcp_lro_free(&rxq->lro);
3996 			rxq->lro.ifp = NULL;
3997 #endif
3998 			return (rc);
3999 		}
4000 		MPASS(rxq->iq.flags & IQ_SW_ALLOCATED);
4001 		add_rxq_sysctls(&vi->ctx, oid, rxq);
4002 	}
4003 
4004 	if (!(rxq->iq.flags & IQ_HW_ALLOCATED)) {
4005 		MPASS(rxq->iq.flags & IQ_SW_ALLOCATED);
4006 		rc = alloc_iq_fl_hwq(vi, &rxq->iq, &rxq->fl);
4007 		if (rc != 0) {
4008 			CH_ERR(vi, "failed to create hw rxq%d: %d\n", idx, rc);
4009 			return (rc);
4010 		}
4011 		MPASS(rxq->iq.flags & IQ_HW_ALLOCATED);
4012 
4013 		if (idx == 0)
4014 			sc->sge.iq_base = rxq->iq.abs_id - rxq->iq.cntxt_id;
4015 		else
4016 			KASSERT(rxq->iq.cntxt_id + sc->sge.iq_base == rxq->iq.abs_id,
4017 			    ("iq_base mismatch"));
4018 		KASSERT(sc->sge.iq_base == 0 || sc->flags & IS_VF,
4019 		    ("PF with non-zero iq_base"));
4020 
4021 		/*
4022 		 * The freelist is just barely above the starvation threshold
4023 		 * right now, fill it up a bit more.
4024 		 */
4025 		FL_LOCK(&rxq->fl);
4026 		refill_fl(sc, &rxq->fl, 128);
4027 		FL_UNLOCK(&rxq->fl);
4028 	}
4029 
4030 	return (0);
4031 }
4032 
4033 /*
4034  * Idempotent.
4035  */
4036 static void
4037 free_rxq(struct vi_info *vi, struct sge_rxq *rxq)
4038 {
4039 	if (rxq->iq.flags & IQ_HW_ALLOCATED) {
4040 		MPASS(rxq->iq.flags & IQ_SW_ALLOCATED);
4041 		free_iq_fl_hwq(vi->adapter, &rxq->iq, &rxq->fl);
4042 		MPASS(!(rxq->iq.flags & IQ_HW_ALLOCATED));
4043 	}
4044 
4045 	if (rxq->iq.flags & IQ_SW_ALLOCATED) {
4046 		MPASS(!(rxq->iq.flags & IQ_HW_ALLOCATED));
4047 #if defined(INET) || defined(INET6)
4048 		tcp_lro_free(&rxq->lro);
4049 #endif
4050 		free_iq_fl(vi->adapter, &rxq->iq, &rxq->fl);
4051 		MPASS(!(rxq->iq.flags & IQ_SW_ALLOCATED));
4052 		bzero(rxq, sizeof(*rxq));
4053 	}
4054 }
4055 
4056 static void
4057 add_rxq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
4058     struct sge_rxq *rxq)
4059 {
4060 	struct sysctl_oid_list *children;
4061 
4062 	if (ctx == NULL || oid == NULL)
4063 		return;
4064 
4065 	children = SYSCTL_CHILDREN(oid);
4066 #if defined(INET) || defined(INET6)
4067 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
4068 	    &rxq->lro.lro_queued, 0, NULL);
4069 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
4070 	    &rxq->lro.lro_flushed, 0, NULL);
4071 #endif
4072 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
4073 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
4074 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "vlan_extraction", CTLFLAG_RD,
4075 	    &rxq->vlan_extraction, "# of times hardware extracted 802.1Q tag");
4076 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "vxlan_rxcsum", CTLFLAG_RD,
4077 	    &rxq->vxlan_rxcsum,
4078 	    "# of times hardware assisted with inner checksum (VXLAN)");
4079 }
4080 
4081 #ifdef TCP_OFFLOAD
4082 /*
4083  * Idempotent.
4084  */
4085 static int
4086 alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq, int idx,
4087     int intr_idx, int maxp)
4088 {
4089 	int rc;
4090 	struct adapter *sc = vi->adapter;
4091 	struct sysctl_oid *oid;
4092 	char name[16];
4093 
4094 	if (!(ofld_rxq->iq.flags & IQ_SW_ALLOCATED)) {
4095 		MPASS(!(ofld_rxq->iq.flags & IQ_HW_ALLOCATED));
4096 
4097 		snprintf(name, sizeof(name), "%d", idx);
4098 		oid = SYSCTL_ADD_NODE(&vi->ctx,
4099 		    SYSCTL_CHILDREN(vi->ofld_rxq_oid), OID_AUTO, name,
4100 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "offload rx queue");
4101 
4102 		init_iq(&ofld_rxq->iq, sc, vi->ofld_tmr_idx, vi->ofld_pktc_idx,
4103 		    vi->qsize_rxq, intr_idx, ofld_cong_drop, IQ_OFLD);
4104 		snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
4105 		    device_get_nameunit(vi->dev), idx);
4106 		init_fl(sc, &ofld_rxq->fl, vi->qsize_rxq / 8, maxp, name);
4107 		rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, &vi->ctx,
4108 		    oid);
4109 		if (rc != 0) {
4110 			CH_ERR(vi, "failed to allocate ofld_rxq%d: %d\n", idx,
4111 			    rc);
4112 			sysctl_remove_oid(oid, 1, 1);
4113 			return (rc);
4114 		}
4115 		MPASS(ofld_rxq->iq.flags & IQ_SW_ALLOCATED);
4116 		ofld_rxq->rx_iscsi_ddp_setup_ok = counter_u64_alloc(M_WAITOK);
4117 		ofld_rxq->rx_iscsi_ddp_setup_error =
4118 		    counter_u64_alloc(M_WAITOK);
4119 		ofld_rxq->ddp_buffer_alloc = counter_u64_alloc(M_WAITOK);
4120 		ofld_rxq->ddp_buffer_reuse = counter_u64_alloc(M_WAITOK);
4121 		ofld_rxq->ddp_buffer_free = counter_u64_alloc(M_WAITOK);
4122 		add_ofld_rxq_sysctls(&vi->ctx, oid, ofld_rxq);
4123 	}
4124 
4125 	if (!(ofld_rxq->iq.flags & IQ_HW_ALLOCATED)) {
4126 		MPASS(ofld_rxq->iq.flags & IQ_SW_ALLOCATED);
4127 		rc = alloc_iq_fl_hwq(vi, &ofld_rxq->iq, &ofld_rxq->fl);
4128 		if (rc != 0) {
4129 			CH_ERR(vi, "failed to create hw ofld_rxq%d: %d\n", idx,
4130 			    rc);
4131 			return (rc);
4132 		}
4133 		MPASS(ofld_rxq->iq.flags & IQ_HW_ALLOCATED);
4134 	}
4135 	return (rc);
4136 }
4137 
4138 /*
4139  * Idempotent.
4140  */
4141 static void
4142 free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq)
4143 {
4144 	if (ofld_rxq->iq.flags & IQ_HW_ALLOCATED) {
4145 		MPASS(ofld_rxq->iq.flags & IQ_SW_ALLOCATED);
4146 		free_iq_fl_hwq(vi->adapter, &ofld_rxq->iq, &ofld_rxq->fl);
4147 		MPASS(!(ofld_rxq->iq.flags & IQ_HW_ALLOCATED));
4148 	}
4149 
4150 	if (ofld_rxq->iq.flags & IQ_SW_ALLOCATED) {
4151 		MPASS(!(ofld_rxq->iq.flags & IQ_HW_ALLOCATED));
4152 		free_iq_fl(vi->adapter, &ofld_rxq->iq, &ofld_rxq->fl);
4153 		MPASS(!(ofld_rxq->iq.flags & IQ_SW_ALLOCATED));
4154 		counter_u64_free(ofld_rxq->rx_iscsi_ddp_setup_ok);
4155 		counter_u64_free(ofld_rxq->rx_iscsi_ddp_setup_error);
4156 		counter_u64_free(ofld_rxq->ddp_buffer_alloc);
4157 		counter_u64_free(ofld_rxq->ddp_buffer_reuse);
4158 		counter_u64_free(ofld_rxq->ddp_buffer_free);
4159 		bzero(ofld_rxq, sizeof(*ofld_rxq));
4160 	}
4161 }
4162 
4163 static void
4164 add_ofld_rxq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
4165     struct sge_ofld_rxq *ofld_rxq)
4166 {
4167 	struct sysctl_oid_list *children;
4168 
4169 	if (ctx == NULL || oid == NULL)
4170 		return;
4171 
4172 	children = SYSCTL_CHILDREN(oid);
4173 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "rx_aio_ddp_jobs",
4174 	    CTLFLAG_RD, &ofld_rxq->rx_aio_ddp_jobs, 0,
4175 	    "# of aio_read(2) jobs completed via DDP");
4176 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "rx_aio_ddp_octets",
4177 	    CTLFLAG_RD, &ofld_rxq->rx_aio_ddp_octets, 0,
4178 	    "# of octets placed directly for aio_read(2) jobs");
4179 	SYSCTL_ADD_ULONG(ctx, children, OID_AUTO,
4180 	    "rx_toe_tls_records", CTLFLAG_RD, &ofld_rxq->rx_toe_tls_records,
4181 	    "# of TOE TLS records received");
4182 	SYSCTL_ADD_ULONG(ctx, children, OID_AUTO,
4183 	    "rx_toe_tls_octets", CTLFLAG_RD, &ofld_rxq->rx_toe_tls_octets,
4184 	    "# of payload octets in received TOE TLS records");
4185 	SYSCTL_ADD_ULONG(ctx, children, OID_AUTO,
4186 	    "rx_toe_ddp_octets", CTLFLAG_RD, &ofld_rxq->rx_toe_ddp_octets,
4187 	    "# of payload octets received via TCP DDP");
4188 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO,
4189 	    "ddp_buffer_alloc", CTLFLAG_RD, &ofld_rxq->ddp_buffer_alloc,
4190 	    "# of DDP RCV buffers allocated");
4191 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO,
4192 	    "ddp_buffer_reuse", CTLFLAG_RD, &ofld_rxq->ddp_buffer_reuse,
4193 	    "# of DDP RCV buffers reused");
4194 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO,
4195 	    "ddp_buffer_free", CTLFLAG_RD, &ofld_rxq->ddp_buffer_free,
4196 	    "# of DDP RCV buffers freed");
4197 
4198 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "iscsi",
4199 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "TOE iSCSI statistics");
4200 	children = SYSCTL_CHILDREN(oid);
4201 
4202 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "ddp_setup_ok",
4203 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_ddp_setup_ok,
4204 	    "# of times DDP buffer was setup successfully.");
4205 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "ddp_setup_error",
4206 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_ddp_setup_error,
4207 	    "# of times DDP buffer setup failed.");
4208 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "ddp_octets",
4209 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_ddp_octets, 0,
4210 	    "# of octets placed directly");
4211 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "ddp_pdus",
4212 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_ddp_pdus, 0,
4213 	    "# of PDUs with data placed directly.");
4214 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "fl_octets",
4215 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_fl_octets, 0,
4216 	    "# of data octets delivered in freelist");
4217 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "fl_pdus",
4218 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_fl_pdus, 0,
4219 	    "# of PDUs with data delivered in freelist");
4220 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "padding_errors",
4221 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_padding_errors, 0,
4222 	    "# of PDUs with invalid padding");
4223 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "header_digest_errors",
4224 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_header_digest_errors, 0,
4225 	    "# of PDUs with invalid header digests");
4226 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "data_digest_errors",
4227 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_data_digest_errors, 0,
4228 	    "# of PDUs with invalid data digests");
4229 }
4230 #endif
4231 
4232 /*
4233  * Returns a reasonable automatic cidx flush threshold for a given queue size.
4234  */
4235 static u_int
4236 qsize_to_fthresh(int qsize)
4237 {
4238 	u_int fthresh;
4239 
4240 	fthresh = qsize == 0 ? 0 : order_base_2(qsize);
4241 	if (fthresh > X_CIDXFLUSHTHRESH_128)
4242 		fthresh = X_CIDXFLUSHTHRESH_128;
4243 
4244 	return (fthresh);
4245 }
4246 
4247 static int
4248 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
4249 {
4250 	int rc, cntxt_id;
4251 	struct fw_eq_ctrl_cmd c;
4252 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
4253 
4254 	bzero(&c, sizeof(c));
4255 
4256 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
4257 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
4258 	    V_FW_EQ_CTRL_CMD_VFN(0));
4259 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
4260 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
4261 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid));
4262 	c.physeqid_pkd = htobe32(0);
4263 	c.fetchszm_to_iqid =
4264 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
4265 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
4266 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
4267 	c.dcaen_to_eqsize =
4268 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
4269 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
4270 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
4271 		V_FW_EQ_CTRL_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
4272 		V_FW_EQ_CTRL_CMD_EQSIZE(qsize));
4273 	c.eqaddr = htobe64(eq->ba);
4274 
4275 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
4276 	if (rc != 0) {
4277 		CH_ERR(sc, "failed to create hw ctrlq for tx_chan %d: %d\n",
4278 		    eq->tx_chan, rc);
4279 		return (rc);
4280 	}
4281 
4282 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
4283 	eq->abs_id = G_FW_EQ_CTRL_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
4284 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
4285 	if (cntxt_id >= sc->sge.eqmap_sz)
4286 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
4287 		cntxt_id, sc->sge.eqmap_sz - 1);
4288 	sc->sge.eqmap[cntxt_id] = eq;
4289 
4290 	return (rc);
4291 }
4292 
4293 static int
4294 eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
4295 {
4296 	int rc, cntxt_id;
4297 	struct fw_eq_eth_cmd c;
4298 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
4299 
4300 	bzero(&c, sizeof(c));
4301 
4302 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
4303 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
4304 	    V_FW_EQ_ETH_CMD_VFN(0));
4305 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
4306 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
4307 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
4308 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
4309 	c.fetchszm_to_iqid =
4310 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
4311 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
4312 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
4313 	c.dcaen_to_eqsize =
4314 	    htobe32(V_FW_EQ_ETH_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
4315 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
4316 		V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
4317 		V_FW_EQ_ETH_CMD_EQSIZE(qsize));
4318 	c.eqaddr = htobe64(eq->ba);
4319 
4320 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
4321 	if (rc != 0) {
4322 		device_printf(vi->dev,
4323 		    "failed to create Ethernet egress queue: %d\n", rc);
4324 		return (rc);
4325 	}
4326 
4327 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
4328 	eq->abs_id = G_FW_EQ_ETH_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
4329 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
4330 	if (cntxt_id >= sc->sge.eqmap_sz)
4331 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
4332 		cntxt_id, sc->sge.eqmap_sz - 1);
4333 	sc->sge.eqmap[cntxt_id] = eq;
4334 
4335 	return (rc);
4336 }
4337 
4338 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
4339 static int
4340 ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
4341 {
4342 	int rc, cntxt_id;
4343 	struct fw_eq_ofld_cmd c;
4344 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
4345 
4346 	bzero(&c, sizeof(c));
4347 
4348 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
4349 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
4350 	    V_FW_EQ_OFLD_CMD_VFN(0));
4351 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
4352 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
4353 	c.fetchszm_to_iqid =
4354 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
4355 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
4356 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
4357 	c.dcaen_to_eqsize =
4358 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
4359 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
4360 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
4361 		V_FW_EQ_OFLD_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
4362 		V_FW_EQ_OFLD_CMD_EQSIZE(qsize));
4363 	c.eqaddr = htobe64(eq->ba);
4364 
4365 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
4366 	if (rc != 0) {
4367 		device_printf(vi->dev,
4368 		    "failed to create egress queue for TCP offload: %d\n", rc);
4369 		return (rc);
4370 	}
4371 
4372 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
4373 	eq->abs_id = G_FW_EQ_OFLD_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
4374 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
4375 	if (cntxt_id >= sc->sge.eqmap_sz)
4376 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
4377 		cntxt_id, sc->sge.eqmap_sz - 1);
4378 	sc->sge.eqmap[cntxt_id] = eq;
4379 
4380 	return (rc);
4381 }
4382 #endif
4383 
4384 /* SW only */
4385 static int
4386 alloc_eq(struct adapter *sc, struct sge_eq *eq, struct sysctl_ctx_list *ctx,
4387     struct sysctl_oid *oid)
4388 {
4389 	int rc, qsize;
4390 	size_t len;
4391 
4392 	MPASS(!(eq->flags & EQ_SW_ALLOCATED));
4393 
4394 	qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
4395 	len = qsize * EQ_ESIZE;
4396 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map, &eq->ba,
4397 	    (void **)&eq->desc);
4398 	if (rc)
4399 		return (rc);
4400 	if (ctx != NULL && oid != NULL)
4401 		add_eq_sysctls(sc, ctx, oid, eq);
4402 	eq->flags |= EQ_SW_ALLOCATED;
4403 
4404 	return (0);
4405 }
4406 
4407 /* SW only */
4408 static void
4409 free_eq(struct adapter *sc, struct sge_eq *eq)
4410 {
4411 	MPASS(eq->flags & EQ_SW_ALLOCATED);
4412 	if (eq->type == EQ_ETH)
4413 		MPASS(eq->pidx == eq->cidx);
4414 
4415 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
4416 	mtx_destroy(&eq->eq_lock);
4417 	bzero(eq, sizeof(*eq));
4418 }
4419 
4420 static void
4421 add_eq_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
4422     struct sysctl_oid *oid, struct sge_eq *eq)
4423 {
4424 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
4425 
4426 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD, &eq->ba,
4427 	    "bus address of descriptor ring");
4428 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
4429 	    eq->sidx * EQ_ESIZE + sc->params.sge.spg_len,
4430 	    "desc ring size in bytes");
4431 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
4432 	    &eq->abs_id, 0, "absolute id of the queue");
4433 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
4434 	    &eq->cntxt_id, 0, "SGE context id of the queue");
4435 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &eq->cidx,
4436 	    0, "consumer index");
4437 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &eq->pidx,
4438 	    0, "producer index");
4439 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
4440 	    eq->sidx, "status page index");
4441 }
4442 
4443 static int
4444 alloc_eq_hwq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
4445 {
4446 	int rc;
4447 
4448 	MPASS(!(eq->flags & EQ_HW_ALLOCATED));
4449 
4450 	eq->iqid = eq->iq->cntxt_id;
4451 	eq->pidx = eq->cidx = eq->dbidx = 0;
4452 	/* Note that equeqidx is not used with sge_wrq (OFLD/CTRL) queues. */
4453 	eq->equeqidx = 0;
4454 	eq->doorbells = sc->doorbells;
4455 	bzero(eq->desc, eq->sidx * EQ_ESIZE + sc->params.sge.spg_len);
4456 
4457 	switch (eq->type) {
4458 	case EQ_CTRL:
4459 		rc = ctrl_eq_alloc(sc, eq);
4460 		break;
4461 
4462 	case EQ_ETH:
4463 		rc = eth_eq_alloc(sc, vi, eq);
4464 		break;
4465 
4466 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
4467 	case EQ_OFLD:
4468 		rc = ofld_eq_alloc(sc, vi, eq);
4469 		break;
4470 #endif
4471 
4472 	default:
4473 		panic("%s: invalid eq type %d.", __func__, eq->type);
4474 	}
4475 	if (rc != 0) {
4476 		CH_ERR(sc, "failed to allocate egress queue(%d): %d\n",
4477 		    eq->type, rc);
4478 		return (rc);
4479 	}
4480 
4481 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
4482 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
4483 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
4484 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
4485 		uint32_t mask = (1 << s_qpp) - 1;
4486 		volatile uint8_t *udb;
4487 
4488 		udb = sc->udbs_base + UDBS_DB_OFFSET;
4489 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
4490 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
4491 		if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
4492 			clrbit(&eq->doorbells, DOORBELL_WCWR);
4493 		else {
4494 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
4495 			eq->udb_qid = 0;
4496 		}
4497 		eq->udb = (volatile void *)udb;
4498 	}
4499 
4500 	eq->flags |= EQ_HW_ALLOCATED;
4501 	return (0);
4502 }
4503 
4504 static int
4505 free_eq_hwq(struct adapter *sc, struct vi_info *vi __unused, struct sge_eq *eq)
4506 {
4507 	int rc;
4508 
4509 	MPASS(eq->flags & EQ_HW_ALLOCATED);
4510 
4511 	switch (eq->type) {
4512 	case EQ_CTRL:
4513 		rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0, eq->cntxt_id);
4514 		break;
4515 	case EQ_ETH:
4516 		rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0, eq->cntxt_id);
4517 		break;
4518 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
4519 	case EQ_OFLD:
4520 		rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0, eq->cntxt_id);
4521 		break;
4522 #endif
4523 	default:
4524 		panic("%s: invalid eq type %d.", __func__, eq->type);
4525 	}
4526 	if (rc != 0) {
4527 		CH_ERR(sc, "failed to free eq (type %d): %d\n", eq->type, rc);
4528 		return (rc);
4529 	}
4530 	eq->flags &= ~EQ_HW_ALLOCATED;
4531 
4532 	return (0);
4533 }
4534 
4535 static int
4536 alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq,
4537     struct sysctl_ctx_list *ctx, struct sysctl_oid *oid)
4538 {
4539 	struct sge_eq *eq = &wrq->eq;
4540 	int rc;
4541 
4542 	MPASS(!(eq->flags & EQ_SW_ALLOCATED));
4543 
4544 	rc = alloc_eq(sc, eq, ctx, oid);
4545 	if (rc)
4546 		return (rc);
4547 	MPASS(eq->flags & EQ_SW_ALLOCATED);
4548 	/* Can't fail after this. */
4549 
4550 	wrq->adapter = sc;
4551 	TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq);
4552 	TAILQ_INIT(&wrq->incomplete_wrs);
4553 	STAILQ_INIT(&wrq->wr_list);
4554 	wrq->nwr_pending = 0;
4555 	wrq->ndesc_needed = 0;
4556 	add_wrq_sysctls(ctx, oid, wrq);
4557 
4558 	return (0);
4559 }
4560 
4561 static void
4562 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
4563 {
4564 	free_eq(sc, &wrq->eq);
4565 	MPASS(wrq->nwr_pending == 0);
4566 	MPASS(wrq->ndesc_needed == 0);
4567 	MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
4568 	MPASS(STAILQ_EMPTY(&wrq->wr_list));
4569 	bzero(wrq, sizeof(*wrq));
4570 }
4571 
4572 static void
4573 add_wrq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
4574     struct sge_wrq *wrq)
4575 {
4576 	struct sysctl_oid_list *children;
4577 
4578 	if (ctx == NULL || oid == NULL)
4579 		return;
4580 
4581 	children = SYSCTL_CHILDREN(oid);
4582 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD,
4583 	    &wrq->tx_wrs_direct, "# of work requests (direct)");
4584 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD,
4585 	    &wrq->tx_wrs_copied, "# of work requests (copied)");
4586 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_sspace", CTLFLAG_RD,
4587 	    &wrq->tx_wrs_ss, "# of work requests (copied from scratch space)");
4588 }
4589 
4590 /*
4591  * Idempotent.
4592  */
4593 static int
4594 alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx)
4595 {
4596 	int rc, iqidx;
4597 	struct port_info *pi = vi->pi;
4598 	struct adapter *sc = vi->adapter;
4599 	struct sge_eq *eq = &txq->eq;
4600 	struct txpkts *txp;
4601 	char name[16];
4602 	struct sysctl_oid *oid;
4603 
4604 	if (!(eq->flags & EQ_SW_ALLOCATED)) {
4605 		MPASS(!(eq->flags & EQ_HW_ALLOCATED));
4606 
4607 		snprintf(name, sizeof(name), "%d", idx);
4608 		oid = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(vi->txq_oid),
4609 		    OID_AUTO, name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
4610 		    "tx queue");
4611 
4612 		iqidx = vi->first_rxq + (idx % vi->nrxq);
4613 		snprintf(name, sizeof(name), "%s txq%d",
4614 		    device_get_nameunit(vi->dev), idx);
4615 		init_eq(sc, &txq->eq, EQ_ETH, vi->qsize_txq, pi->port_id,
4616 		    &sc->sge.rxq[iqidx].iq, name);
4617 
4618 		rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx,
4619 		    can_resume_eth_tx, M_CXGBE, &eq->eq_lock, M_WAITOK);
4620 		if (rc != 0) {
4621 			CH_ERR(vi, "failed to allocate mp_ring for txq%d: %d\n",
4622 			    idx, rc);
4623 failed:
4624 			sysctl_remove_oid(oid, 1, 1);
4625 			return (rc);
4626 		}
4627 
4628 		rc = alloc_eq(sc, eq, &vi->ctx, oid);
4629 		if (rc) {
4630 			CH_ERR(vi, "failed to allocate txq%d: %d\n", idx, rc);
4631 			mp_ring_free(txq->r);
4632 			goto failed;
4633 		}
4634 		MPASS(eq->flags & EQ_SW_ALLOCATED);
4635 		/* Can't fail after this point. */
4636 
4637 		TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq);
4638 		txq->ifp = vi->ifp;
4639 		txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK);
4640 		txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE,
4641 		    M_ZERO | M_WAITOK);
4642 
4643 		add_txq_sysctls(vi, &vi->ctx, oid, txq);
4644 	}
4645 
4646 	if (!(eq->flags & EQ_HW_ALLOCATED)) {
4647 		MPASS(eq->flags & EQ_SW_ALLOCATED);
4648 		rc = alloc_eq_hwq(sc, vi, eq);
4649 		if (rc != 0) {
4650 			CH_ERR(vi, "failed to create hw txq%d: %d\n", idx, rc);
4651 			return (rc);
4652 		}
4653 		MPASS(eq->flags & EQ_HW_ALLOCATED);
4654 		/* Can't fail after this point. */
4655 
4656 		if (idx == 0)
4657 			sc->sge.eq_base = eq->abs_id - eq->cntxt_id;
4658 		else
4659 			KASSERT(eq->cntxt_id + sc->sge.eq_base == eq->abs_id,
4660 			    ("eq_base mismatch"));
4661 		KASSERT(sc->sge.eq_base == 0 || sc->flags & IS_VF,
4662 		    ("PF with non-zero eq_base"));
4663 
4664 		txp = &txq->txp;
4665 		MPASS(nitems(txp->mb) >= sc->params.max_pkts_per_eth_tx_pkts_wr);
4666 		txq->txp.max_npkt = min(nitems(txp->mb),
4667 		    sc->params.max_pkts_per_eth_tx_pkts_wr);
4668 		if (vi->flags & TX_USES_VM_WR && !(sc->flags & IS_VF))
4669 			txq->txp.max_npkt--;
4670 
4671 		if (vi->flags & TX_USES_VM_WR)
4672 			txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
4673 			    V_TXPKT_INTF(pi->tx_chan));
4674 		else
4675 			txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
4676 			    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(sc->pf) |
4677 			    V_TXPKT_VF(vi->vin) | V_TXPKT_VF_VLD(vi->vfvld));
4678 
4679 		txq->tc_idx = -1;
4680 	}
4681 
4682 	return (0);
4683 }
4684 
4685 /*
4686  * Idempotent.
4687  */
4688 static void
4689 free_txq(struct vi_info *vi, struct sge_txq *txq)
4690 {
4691 	struct adapter *sc = vi->adapter;
4692 	struct sge_eq *eq = &txq->eq;
4693 
4694 	if (eq->flags & EQ_HW_ALLOCATED) {
4695 		MPASS(eq->flags & EQ_SW_ALLOCATED);
4696 		free_eq_hwq(sc, NULL, eq);
4697 		MPASS(!(eq->flags & EQ_HW_ALLOCATED));
4698 	}
4699 
4700 	if (eq->flags & EQ_SW_ALLOCATED) {
4701 		MPASS(!(eq->flags & EQ_HW_ALLOCATED));
4702 		sglist_free(txq->gl);
4703 		free(txq->sdesc, M_CXGBE);
4704 		mp_ring_free(txq->r);
4705 		free_eq(sc, eq);
4706 		MPASS(!(eq->flags & EQ_SW_ALLOCATED));
4707 		bzero(txq, sizeof(*txq));
4708 	}
4709 }
4710 
4711 static void
4712 add_txq_sysctls(struct vi_info *vi, struct sysctl_ctx_list *ctx,
4713     struct sysctl_oid *oid, struct sge_txq *txq)
4714 {
4715 	struct adapter *sc;
4716 	struct sysctl_oid_list *children;
4717 
4718 	if (ctx == NULL || oid == NULL)
4719 		return;
4720 
4721 	sc = vi->adapter;
4722 	children = SYSCTL_CHILDREN(oid);
4723 
4724 	mp_ring_sysctls(txq->r, ctx, children);
4725 
4726 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "tc",
4727 	    CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, vi, txq - sc->sge.txq,
4728 	    sysctl_tc, "I", "traffic class (-1 means none)");
4729 
4730 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
4731 	    &txq->txcsum, "# of times hardware assisted with checksum");
4732 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "vlan_insertion", CTLFLAG_RD,
4733 	    &txq->vlan_insertion, "# of times hardware inserted 802.1Q tag");
4734 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
4735 	    &txq->tso_wrs, "# of TSO work requests");
4736 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
4737 	    &txq->imm_wrs, "# of work requests with immediate data");
4738 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
4739 	    &txq->sgl_wrs, "# of work requests with direct SGL");
4740 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
4741 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
4742 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txpkts0_wrs", CTLFLAG_RD,
4743 	    &txq->txpkts0_wrs, "# of txpkts (type 0) work requests");
4744 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txpkts1_wrs", CTLFLAG_RD,
4745 	    &txq->txpkts1_wrs, "# of txpkts (type 1) work requests");
4746 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txpkts0_pkts", CTLFLAG_RD,
4747 	    &txq->txpkts0_pkts,
4748 	    "# of frames tx'd using type0 txpkts work requests");
4749 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txpkts1_pkts", CTLFLAG_RD,
4750 	    &txq->txpkts1_pkts,
4751 	    "# of frames tx'd using type1 txpkts work requests");
4752 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txpkts_flush", CTLFLAG_RD,
4753 	    &txq->txpkts_flush,
4754 	    "# of times txpkts had to be flushed out by an egress-update");
4755 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "raw_wrs", CTLFLAG_RD,
4756 	    &txq->raw_wrs, "# of raw work requests (non-packets)");
4757 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "vxlan_tso_wrs", CTLFLAG_RD,
4758 	    &txq->vxlan_tso_wrs, "# of VXLAN TSO work requests");
4759 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "vxlan_txcsum", CTLFLAG_RD,
4760 	    &txq->vxlan_txcsum,
4761 	    "# of times hardware assisted with inner checksums (VXLAN)");
4762 
4763 #ifdef KERN_TLS
4764 	if (is_ktls(sc)) {
4765 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_records",
4766 		    CTLFLAG_RD, &txq->kern_tls_records,
4767 		    "# of NIC TLS records transmitted");
4768 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_short",
4769 		    CTLFLAG_RD, &txq->kern_tls_short,
4770 		    "# of short NIC TLS records transmitted");
4771 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_partial",
4772 		    CTLFLAG_RD, &txq->kern_tls_partial,
4773 		    "# of partial NIC TLS records transmitted");
4774 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_full",
4775 		    CTLFLAG_RD, &txq->kern_tls_full,
4776 		    "# of full NIC TLS records transmitted");
4777 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_octets",
4778 		    CTLFLAG_RD, &txq->kern_tls_octets,
4779 		    "# of payload octets in transmitted NIC TLS records");
4780 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_waste",
4781 		    CTLFLAG_RD, &txq->kern_tls_waste,
4782 		    "# of octets DMAd but not transmitted in NIC TLS records");
4783 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_options",
4784 		    CTLFLAG_RD, &txq->kern_tls_options,
4785 		    "# of NIC TLS options-only packets transmitted");
4786 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_header",
4787 		    CTLFLAG_RD, &txq->kern_tls_header,
4788 		    "# of NIC TLS header-only packets transmitted");
4789 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_fin",
4790 		    CTLFLAG_RD, &txq->kern_tls_fin,
4791 		    "# of NIC TLS FIN-only packets transmitted");
4792 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_fin_short",
4793 		    CTLFLAG_RD, &txq->kern_tls_fin_short,
4794 		    "# of NIC TLS padded FIN packets on short TLS records");
4795 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_cbc",
4796 		    CTLFLAG_RD, &txq->kern_tls_cbc,
4797 		    "# of NIC TLS sessions using AES-CBC");
4798 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_gcm",
4799 		    CTLFLAG_RD, &txq->kern_tls_gcm,
4800 		    "# of NIC TLS sessions using AES-GCM");
4801 	}
4802 #endif
4803 }
4804 
4805 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
4806 /*
4807  * Idempotent.
4808  */
4809 static int
4810 alloc_ofld_txq(struct vi_info *vi, struct sge_ofld_txq *ofld_txq, int idx)
4811 {
4812 	struct sysctl_oid *oid;
4813 	struct port_info *pi = vi->pi;
4814 	struct adapter *sc = vi->adapter;
4815 	struct sge_eq *eq = &ofld_txq->wrq.eq;
4816 	int rc, iqidx;
4817 	char name[16];
4818 
4819 	MPASS(idx >= 0);
4820 	MPASS(idx < vi->nofldtxq);
4821 
4822 	if (!(eq->flags & EQ_SW_ALLOCATED)) {
4823 		snprintf(name, sizeof(name), "%d", idx);
4824 		oid = SYSCTL_ADD_NODE(&vi->ctx,
4825 		    SYSCTL_CHILDREN(vi->ofld_txq_oid), OID_AUTO, name,
4826 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "offload tx queue");
4827 
4828 		snprintf(name, sizeof(name), "%s ofld_txq%d",
4829 		    device_get_nameunit(vi->dev), idx);
4830 		if (vi->nofldrxq > 0) {
4831 			iqidx = vi->first_ofld_rxq + (idx % vi->nofldrxq);
4832 			init_eq(sc, eq, EQ_OFLD, vi->qsize_txq, pi->port_id,
4833 			    &sc->sge.ofld_rxq[iqidx].iq, name);
4834 		} else {
4835 			iqidx = vi->first_rxq + (idx % vi->nrxq);
4836 			init_eq(sc, eq, EQ_OFLD, vi->qsize_txq, pi->port_id,
4837 			    &sc->sge.rxq[iqidx].iq, name);
4838 		}
4839 
4840 		rc = alloc_wrq(sc, vi, &ofld_txq->wrq, &vi->ctx, oid);
4841 		if (rc != 0) {
4842 			CH_ERR(vi, "failed to allocate ofld_txq%d: %d\n", idx,
4843 			    rc);
4844 			sysctl_remove_oid(oid, 1, 1);
4845 			return (rc);
4846 		}
4847 		MPASS(eq->flags & EQ_SW_ALLOCATED);
4848 		/* Can't fail after this point. */
4849 
4850 		ofld_txq->tx_iscsi_pdus = counter_u64_alloc(M_WAITOK);
4851 		ofld_txq->tx_iscsi_octets = counter_u64_alloc(M_WAITOK);
4852 		ofld_txq->tx_iscsi_iso_wrs = counter_u64_alloc(M_WAITOK);
4853 		ofld_txq->tx_aio_jobs = counter_u64_alloc(M_WAITOK);
4854 		ofld_txq->tx_aio_octets = counter_u64_alloc(M_WAITOK);
4855 		ofld_txq->tx_toe_tls_records = counter_u64_alloc(M_WAITOK);
4856 		ofld_txq->tx_toe_tls_octets = counter_u64_alloc(M_WAITOK);
4857 		add_ofld_txq_sysctls(&vi->ctx, oid, ofld_txq);
4858 	}
4859 
4860 	if (!(eq->flags & EQ_HW_ALLOCATED)) {
4861 		MPASS(eq->flags & EQ_SW_ALLOCATED);
4862 		MPASS(ofld_txq->wrq.nwr_pending == 0);
4863 		MPASS(ofld_txq->wrq.ndesc_needed == 0);
4864 		rc = alloc_eq_hwq(sc, vi, eq);
4865 		if (rc != 0) {
4866 			CH_ERR(vi, "failed to create hw ofld_txq%d: %d\n", idx,
4867 			    rc);
4868 			return (rc);
4869 		}
4870 		MPASS(eq->flags & EQ_HW_ALLOCATED);
4871 	}
4872 
4873 	return (0);
4874 }
4875 
4876 /*
4877  * Idempotent.
4878  */
4879 static void
4880 free_ofld_txq(struct vi_info *vi, struct sge_ofld_txq *ofld_txq)
4881 {
4882 	struct adapter *sc = vi->adapter;
4883 	struct sge_eq *eq = &ofld_txq->wrq.eq;
4884 
4885 	if (eq->flags & EQ_HW_ALLOCATED) {
4886 		MPASS(eq->flags & EQ_SW_ALLOCATED);
4887 		free_eq_hwq(sc, NULL, eq);
4888 		MPASS(!(eq->flags & EQ_HW_ALLOCATED));
4889 	}
4890 
4891 	if (eq->flags & EQ_SW_ALLOCATED) {
4892 		MPASS(!(eq->flags & EQ_HW_ALLOCATED));
4893 		counter_u64_free(ofld_txq->tx_iscsi_pdus);
4894 		counter_u64_free(ofld_txq->tx_iscsi_octets);
4895 		counter_u64_free(ofld_txq->tx_iscsi_iso_wrs);
4896 		counter_u64_free(ofld_txq->tx_aio_jobs);
4897 		counter_u64_free(ofld_txq->tx_aio_octets);
4898 		counter_u64_free(ofld_txq->tx_toe_tls_records);
4899 		counter_u64_free(ofld_txq->tx_toe_tls_octets);
4900 		free_wrq(sc, &ofld_txq->wrq);
4901 		MPASS(!(eq->flags & EQ_SW_ALLOCATED));
4902 		bzero(ofld_txq, sizeof(*ofld_txq));
4903 	}
4904 }
4905 
4906 static void
4907 add_ofld_txq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
4908     struct sge_ofld_txq *ofld_txq)
4909 {
4910 	struct sysctl_oid_list *children;
4911 
4912 	if (ctx == NULL || oid == NULL)
4913 		return;
4914 
4915 	children = SYSCTL_CHILDREN(oid);
4916 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_iscsi_pdus",
4917 	    CTLFLAG_RD, &ofld_txq->tx_iscsi_pdus,
4918 	    "# of iSCSI PDUs transmitted");
4919 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_iscsi_octets",
4920 	    CTLFLAG_RD, &ofld_txq->tx_iscsi_octets,
4921 	    "# of payload octets in transmitted iSCSI PDUs");
4922 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_iscsi_iso_wrs",
4923 	    CTLFLAG_RD, &ofld_txq->tx_iscsi_iso_wrs,
4924 	    "# of iSCSI segmentation offload work requests");
4925 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_aio_jobs",
4926 	    CTLFLAG_RD, &ofld_txq->tx_aio_jobs,
4927 	    "# of zero-copy aio_write(2) jobs transmitted");
4928 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_aio_octets",
4929 	    CTLFLAG_RD, &ofld_txq->tx_aio_octets,
4930 	    "# of payload octets in transmitted zero-copy aio_write(2) jobs");
4931 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_toe_tls_records",
4932 	    CTLFLAG_RD, &ofld_txq->tx_toe_tls_records,
4933 	    "# of TOE TLS records transmitted");
4934 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_toe_tls_octets",
4935 	    CTLFLAG_RD, &ofld_txq->tx_toe_tls_octets,
4936 	    "# of payload octets in transmitted TOE TLS records");
4937 }
4938 #endif
4939 
4940 static void
4941 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
4942 {
4943 	bus_addr_t *ba = arg;
4944 
4945 	KASSERT(nseg == 1,
4946 	    ("%s meant for single segment mappings only.", __func__));
4947 
4948 	*ba = error ? 0 : segs->ds_addr;
4949 }
4950 
4951 static inline void
4952 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
4953 {
4954 	uint32_t n, v;
4955 
4956 	n = IDXDIFF(fl->pidx >> 3, fl->dbidx, fl->sidx);
4957 	MPASS(n > 0);
4958 
4959 	wmb();
4960 	v = fl->dbval | V_PIDX(n);
4961 	if (fl->udb)
4962 		*fl->udb = htole32(v);
4963 	else
4964 		t4_write_reg(sc, sc->sge_kdoorbell_reg, v);
4965 	IDXINCR(fl->dbidx, n, fl->sidx);
4966 }
4967 
4968 /*
4969  * Fills up the freelist by allocating up to 'n' buffers.  Buffers that are
4970  * recycled do not count towards this allocation budget.
4971  *
4972  * Returns non-zero to indicate that this freelist should be added to the list
4973  * of starving freelists.
4974  */
4975 static int
4976 refill_fl(struct adapter *sc, struct sge_fl *fl, int n)
4977 {
4978 	__be64 *d;
4979 	struct fl_sdesc *sd;
4980 	uintptr_t pa;
4981 	caddr_t cl;
4982 	struct rx_buf_info *rxb;
4983 	struct cluster_metadata *clm;
4984 	uint16_t max_pidx, zidx = fl->zidx;
4985 	uint16_t hw_cidx = fl->hw_cidx;		/* stable snapshot */
4986 
4987 	FL_LOCK_ASSERT_OWNED(fl);
4988 
4989 	/*
4990 	 * We always stop at the beginning of the hardware descriptor that's just
4991 	 * before the one with the hw cidx.  This is to avoid hw pidx = hw cidx,
4992 	 * which would mean an empty freelist to the chip.
4993 	 */
4994 	max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1;
4995 	if (fl->pidx == max_pidx * 8)
4996 		return (0);
4997 
4998 	d = &fl->desc[fl->pidx];
4999 	sd = &fl->sdesc[fl->pidx];
5000 	rxb = &sc->sge.rx_buf_info[zidx];
5001 
5002 	while (n > 0) {
5003 
5004 		if (sd->cl != NULL) {
5005 
5006 			if (sd->nmbuf == 0) {
5007 				/*
5008 				 * Fast recycle without involving any atomics on
5009 				 * the cluster's metadata (if the cluster has
5010 				 * metadata).  This happens when all frames
5011 				 * received in the cluster were small enough to
5012 				 * fit within a single mbuf each.
5013 				 */
5014 				fl->cl_fast_recycled++;
5015 				goto recycled;
5016 			}
5017 
5018 			/*
5019 			 * Cluster is guaranteed to have metadata.  Clusters
5020 			 * without metadata always take the fast recycle path
5021 			 * when they're recycled.
5022 			 */
5023 			clm = cl_metadata(sd);
5024 			MPASS(clm != NULL);
5025 
5026 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
5027 				fl->cl_recycled++;
5028 				counter_u64_add(extfree_rels, 1);
5029 				goto recycled;
5030 			}
5031 			sd->cl = NULL;	/* gave up my reference */
5032 		}
5033 		MPASS(sd->cl == NULL);
5034 		cl = uma_zalloc(rxb->zone, M_NOWAIT);
5035 		if (__predict_false(cl == NULL)) {
5036 			if (zidx != fl->safe_zidx) {
5037 				zidx = fl->safe_zidx;
5038 				rxb = &sc->sge.rx_buf_info[zidx];
5039 				cl = uma_zalloc(rxb->zone, M_NOWAIT);
5040 			}
5041 			if (cl == NULL)
5042 				break;
5043 		}
5044 		fl->cl_allocated++;
5045 		n--;
5046 
5047 		pa = pmap_kextract((vm_offset_t)cl);
5048 		sd->cl = cl;
5049 		sd->zidx = zidx;
5050 
5051 		if (fl->flags & FL_BUF_PACKING) {
5052 			*d = htobe64(pa | rxb->hwidx2);
5053 			sd->moff = rxb->size2;
5054 		} else {
5055 			*d = htobe64(pa | rxb->hwidx1);
5056 			sd->moff = 0;
5057 		}
5058 recycled:
5059 		sd->nmbuf = 0;
5060 		d++;
5061 		sd++;
5062 		if (__predict_false((++fl->pidx & 7) == 0)) {
5063 			uint16_t pidx = fl->pidx >> 3;
5064 
5065 			if (__predict_false(pidx == fl->sidx)) {
5066 				fl->pidx = 0;
5067 				pidx = 0;
5068 				sd = fl->sdesc;
5069 				d = fl->desc;
5070 			}
5071 			if (n < 8 || pidx == max_pidx)
5072 				break;
5073 
5074 			if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4)
5075 				ring_fl_db(sc, fl);
5076 		}
5077 	}
5078 
5079 	if ((fl->pidx >> 3) != fl->dbidx)
5080 		ring_fl_db(sc, fl);
5081 
5082 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
5083 }
5084 
5085 /*
5086  * Attempt to refill all starving freelists.
5087  */
5088 static void
5089 refill_sfl(void *arg)
5090 {
5091 	struct adapter *sc = arg;
5092 	struct sge_fl *fl, *fl_temp;
5093 
5094 	mtx_assert(&sc->sfl_lock, MA_OWNED);
5095 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
5096 		FL_LOCK(fl);
5097 		refill_fl(sc, fl, 64);
5098 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
5099 			TAILQ_REMOVE(&sc->sfl, fl, link);
5100 			fl->flags &= ~FL_STARVING;
5101 		}
5102 		FL_UNLOCK(fl);
5103 	}
5104 
5105 	if (!TAILQ_EMPTY(&sc->sfl))
5106 		callout_schedule(&sc->sfl_callout, hz / 5);
5107 }
5108 
5109 /*
5110  * Release the driver's reference on all buffers in the given freelist.  Buffers
5111  * with kernel references cannot be freed and will prevent the driver from being
5112  * unloaded safely.
5113  */
5114 void
5115 free_fl_buffers(struct adapter *sc, struct sge_fl *fl)
5116 {
5117 	struct fl_sdesc *sd;
5118 	struct cluster_metadata *clm;
5119 	int i;
5120 
5121 	sd = fl->sdesc;
5122 	for (i = 0; i < fl->sidx * 8; i++, sd++) {
5123 		if (sd->cl == NULL)
5124 			continue;
5125 
5126 		if (sd->nmbuf == 0)
5127 			uma_zfree(sc->sge.rx_buf_info[sd->zidx].zone, sd->cl);
5128 		else if (fl->flags & FL_BUF_PACKING) {
5129 			clm = cl_metadata(sd);
5130 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
5131 				uma_zfree(sc->sge.rx_buf_info[sd->zidx].zone,
5132 				    sd->cl);
5133 				counter_u64_add(extfree_rels, 1);
5134 			}
5135 		}
5136 		sd->cl = NULL;
5137 	}
5138 
5139 	if (fl->flags & FL_BUF_RESUME) {
5140 		m_freem(fl->m0);
5141 		fl->flags &= ~FL_BUF_RESUME;
5142 	}
5143 }
5144 
5145 static inline void
5146 get_pkt_gl(struct mbuf *m, struct sglist *gl)
5147 {
5148 	int rc;
5149 
5150 	M_ASSERTPKTHDR(m);
5151 
5152 	sglist_reset(gl);
5153 	rc = sglist_append_mbuf(gl, m);
5154 	if (__predict_false(rc != 0)) {
5155 		panic("%s: mbuf %p (%d segs) was vetted earlier but now fails "
5156 		    "with %d.", __func__, m, mbuf_nsegs(m), rc);
5157 	}
5158 
5159 	KASSERT(gl->sg_nseg == mbuf_nsegs(m),
5160 	    ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m,
5161 	    mbuf_nsegs(m), gl->sg_nseg));
5162 #if 0	/* vm_wr not readily available here. */
5163 	KASSERT(gl->sg_nseg > 0 && gl->sg_nseg <= max_nsegs_allowed(m, vm_wr),
5164 	    ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__,
5165 		gl->sg_nseg, max_nsegs_allowed(m, vm_wr)));
5166 #endif
5167 }
5168 
5169 /*
5170  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
5171  */
5172 static inline u_int
5173 txpkt_len16(u_int nsegs, const u_int extra)
5174 {
5175 	u_int n;
5176 
5177 	MPASS(nsegs > 0);
5178 
5179 	nsegs--; /* first segment is part of ulptx_sgl */
5180 	n = extra + sizeof(struct fw_eth_tx_pkt_wr) +
5181 	    sizeof(struct cpl_tx_pkt_core) +
5182 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
5183 
5184 	return (howmany(n, 16));
5185 }
5186 
5187 /*
5188  * len16 for a txpkt_vm WR with a GL.  Includes the firmware work
5189  * request header.
5190  */
5191 static inline u_int
5192 txpkt_vm_len16(u_int nsegs, const u_int extra)
5193 {
5194 	u_int n;
5195 
5196 	MPASS(nsegs > 0);
5197 
5198 	nsegs--; /* first segment is part of ulptx_sgl */
5199 	n = extra + sizeof(struct fw_eth_tx_pkt_vm_wr) +
5200 	    sizeof(struct cpl_tx_pkt_core) +
5201 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
5202 
5203 	return (howmany(n, 16));
5204 }
5205 
5206 static inline void
5207 calculate_mbuf_len16(struct mbuf *m, bool vm_wr)
5208 {
5209 	const int lso = sizeof(struct cpl_tx_pkt_lso_core);
5210 	const int tnl_lso = sizeof(struct cpl_tx_tnl_lso);
5211 
5212 	if (vm_wr) {
5213 		if (needs_tso(m))
5214 			set_mbuf_len16(m, txpkt_vm_len16(mbuf_nsegs(m), lso));
5215 		else
5216 			set_mbuf_len16(m, txpkt_vm_len16(mbuf_nsegs(m), 0));
5217 		return;
5218 	}
5219 
5220 	if (needs_tso(m)) {
5221 		if (needs_vxlan_tso(m))
5222 			set_mbuf_len16(m, txpkt_len16(mbuf_nsegs(m), tnl_lso));
5223 		else
5224 			set_mbuf_len16(m, txpkt_len16(mbuf_nsegs(m), lso));
5225 	} else
5226 		set_mbuf_len16(m, txpkt_len16(mbuf_nsegs(m), 0));
5227 }
5228 
5229 /*
5230  * len16 for a txpkts type 0 WR with a GL.  Does not include the firmware work
5231  * request header.
5232  */
5233 static inline u_int
5234 txpkts0_len16(u_int nsegs)
5235 {
5236 	u_int n;
5237 
5238 	MPASS(nsegs > 0);
5239 
5240 	nsegs--; /* first segment is part of ulptx_sgl */
5241 	n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) +
5242 	    sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) +
5243 	    8 * ((3 * nsegs) / 2 + (nsegs & 1));
5244 
5245 	return (howmany(n, 16));
5246 }
5247 
5248 /*
5249  * len16 for a txpkts type 1 WR with a GL.  Does not include the firmware work
5250  * request header.
5251  */
5252 static inline u_int
5253 txpkts1_len16(void)
5254 {
5255 	u_int n;
5256 
5257 	n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl);
5258 
5259 	return (howmany(n, 16));
5260 }
5261 
5262 static inline u_int
5263 imm_payload(u_int ndesc)
5264 {
5265 	u_int n;
5266 
5267 	n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) -
5268 	    sizeof(struct cpl_tx_pkt_core);
5269 
5270 	return (n);
5271 }
5272 
5273 static inline uint64_t
5274 csum_to_ctrl(struct adapter *sc, struct mbuf *m)
5275 {
5276 	uint64_t ctrl;
5277 	int csum_type, l2hlen, l3hlen;
5278 	int x, y;
5279 	static const int csum_types[3][2] = {
5280 		{TX_CSUM_TCPIP, TX_CSUM_TCPIP6},
5281 		{TX_CSUM_UDPIP, TX_CSUM_UDPIP6},
5282 		{TX_CSUM_IP, 0}
5283 	};
5284 
5285 	M_ASSERTPKTHDR(m);
5286 
5287 	if (!needs_hwcsum(m))
5288 		return (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS);
5289 
5290 	MPASS(m->m_pkthdr.l2hlen >= ETHER_HDR_LEN);
5291 	MPASS(m->m_pkthdr.l3hlen >= sizeof(struct ip));
5292 
5293 	if (needs_vxlan_csum(m)) {
5294 		MPASS(m->m_pkthdr.l4hlen > 0);
5295 		MPASS(m->m_pkthdr.l5hlen > 0);
5296 		MPASS(m->m_pkthdr.inner_l2hlen >= ETHER_HDR_LEN);
5297 		MPASS(m->m_pkthdr.inner_l3hlen >= sizeof(struct ip));
5298 
5299 		l2hlen = m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen +
5300 		    m->m_pkthdr.l4hlen + m->m_pkthdr.l5hlen +
5301 		    m->m_pkthdr.inner_l2hlen - ETHER_HDR_LEN;
5302 		l3hlen = m->m_pkthdr.inner_l3hlen;
5303 	} else {
5304 		l2hlen = m->m_pkthdr.l2hlen - ETHER_HDR_LEN;
5305 		l3hlen = m->m_pkthdr.l3hlen;
5306 	}
5307 
5308 	ctrl = 0;
5309 	if (!needs_l3_csum(m))
5310 		ctrl |= F_TXPKT_IPCSUM_DIS;
5311 
5312 	if (m->m_pkthdr.csum_flags & (CSUM_IP_TCP | CSUM_INNER_IP_TCP |
5313 	    CSUM_IP6_TCP | CSUM_INNER_IP6_TCP))
5314 		x = 0;	/* TCP */
5315 	else if (m->m_pkthdr.csum_flags & (CSUM_IP_UDP | CSUM_INNER_IP_UDP |
5316 	    CSUM_IP6_UDP | CSUM_INNER_IP6_UDP))
5317 		x = 1;	/* UDP */
5318 	else
5319 		x = 2;
5320 
5321 	if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_IP_TCP | CSUM_IP_UDP |
5322 	    CSUM_INNER_IP | CSUM_INNER_IP_TCP | CSUM_INNER_IP_UDP))
5323 		y = 0;	/* IPv4 */
5324 	else {
5325 		MPASS(m->m_pkthdr.csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP |
5326 		    CSUM_INNER_IP6_TCP | CSUM_INNER_IP6_UDP));
5327 		y = 1;	/* IPv6 */
5328 	}
5329 	/*
5330 	 * needs_hwcsum returned true earlier so there must be some kind of
5331 	 * checksum to calculate.
5332 	 */
5333 	csum_type = csum_types[x][y];
5334 	MPASS(csum_type != 0);
5335 	if (csum_type == TX_CSUM_IP)
5336 		ctrl |= F_TXPKT_L4CSUM_DIS;
5337 	ctrl |= V_TXPKT_CSUM_TYPE(csum_type) | V_TXPKT_IPHDR_LEN(l3hlen);
5338 	if (chip_id(sc) <= CHELSIO_T5)
5339 		ctrl |= V_TXPKT_ETHHDR_LEN(l2hlen);
5340 	else
5341 		ctrl |= V_T6_TXPKT_ETHHDR_LEN(l2hlen);
5342 
5343 	return (ctrl);
5344 }
5345 
5346 static inline void *
5347 write_lso_cpl(void *cpl, struct mbuf *m0)
5348 {
5349 	struct cpl_tx_pkt_lso_core *lso;
5350 	uint32_t ctrl;
5351 
5352 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
5353 	    m0->m_pkthdr.l4hlen > 0,
5354 	    ("%s: mbuf %p needs TSO but missing header lengths",
5355 		__func__, m0));
5356 
5357 	ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) |
5358 	    F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE |
5359 	    V_LSO_ETHHDR_LEN((m0->m_pkthdr.l2hlen - ETHER_HDR_LEN) >> 2) |
5360 	    V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) |
5361 	    V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
5362 	if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
5363 		ctrl |= F_LSO_IPV6;
5364 
5365 	lso = cpl;
5366 	lso->lso_ctrl = htobe32(ctrl);
5367 	lso->ipid_ofst = htobe16(0);
5368 	lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
5369 	lso->seqno_offset = htobe32(0);
5370 	lso->len = htobe32(m0->m_pkthdr.len);
5371 
5372 	return (lso + 1);
5373 }
5374 
5375 static void *
5376 write_tnl_lso_cpl(void *cpl, struct mbuf *m0)
5377 {
5378 	struct cpl_tx_tnl_lso *tnl_lso = cpl;
5379 	uint32_t ctrl;
5380 
5381 	KASSERT(m0->m_pkthdr.inner_l2hlen > 0 &&
5382 	    m0->m_pkthdr.inner_l3hlen > 0 && m0->m_pkthdr.inner_l4hlen > 0 &&
5383 	    m0->m_pkthdr.inner_l5hlen > 0,
5384 	    ("%s: mbuf %p needs VXLAN_TSO but missing inner header lengths",
5385 		__func__, m0));
5386 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
5387 	    m0->m_pkthdr.l4hlen > 0 && m0->m_pkthdr.l5hlen > 0,
5388 	    ("%s: mbuf %p needs VXLAN_TSO but missing outer header lengths",
5389 		__func__, m0));
5390 
5391 	/* Outer headers. */
5392 	ctrl = V_CPL_TX_TNL_LSO_OPCODE(CPL_TX_TNL_LSO) |
5393 	    F_CPL_TX_TNL_LSO_FIRST | F_CPL_TX_TNL_LSO_LAST |
5394 	    V_CPL_TX_TNL_LSO_ETHHDRLENOUT(
5395 		(m0->m_pkthdr.l2hlen - ETHER_HDR_LEN) >> 2) |
5396 	    V_CPL_TX_TNL_LSO_IPHDRLENOUT(m0->m_pkthdr.l3hlen >> 2) |
5397 	    F_CPL_TX_TNL_LSO_IPLENSETOUT;
5398 	if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
5399 		ctrl |= F_CPL_TX_TNL_LSO_IPV6OUT;
5400 	else {
5401 		ctrl |= F_CPL_TX_TNL_LSO_IPHDRCHKOUT |
5402 		    F_CPL_TX_TNL_LSO_IPIDINCOUT;
5403 	}
5404 	tnl_lso->op_to_IpIdSplitOut = htobe32(ctrl);
5405 	tnl_lso->IpIdOffsetOut = 0;
5406 	tnl_lso->UdpLenSetOut_to_TnlHdrLen =
5407 		htobe16(F_CPL_TX_TNL_LSO_UDPCHKCLROUT |
5408 		    F_CPL_TX_TNL_LSO_UDPLENSETOUT |
5409 		    V_CPL_TX_TNL_LSO_TNLHDRLEN(m0->m_pkthdr.l2hlen +
5410 			m0->m_pkthdr.l3hlen + m0->m_pkthdr.l4hlen +
5411 			m0->m_pkthdr.l5hlen) |
5412 		    V_CPL_TX_TNL_LSO_TNLTYPE(TX_TNL_TYPE_VXLAN));
5413 	tnl_lso->r1 = 0;
5414 
5415 	/* Inner headers. */
5416 	ctrl = V_CPL_TX_TNL_LSO_ETHHDRLEN(
5417 	    (m0->m_pkthdr.inner_l2hlen - ETHER_HDR_LEN) >> 2) |
5418 	    V_CPL_TX_TNL_LSO_IPHDRLEN(m0->m_pkthdr.inner_l3hlen >> 2) |
5419 	    V_CPL_TX_TNL_LSO_TCPHDRLEN(m0->m_pkthdr.inner_l4hlen >> 2);
5420 	if (m0->m_pkthdr.inner_l3hlen == sizeof(struct ip6_hdr))
5421 		ctrl |= F_CPL_TX_TNL_LSO_IPV6;
5422 	tnl_lso->Flow_to_TcpHdrLen = htobe32(ctrl);
5423 	tnl_lso->IpIdOffset = 0;
5424 	tnl_lso->IpIdSplit_to_Mss =
5425 	    htobe16(V_CPL_TX_TNL_LSO_MSS(m0->m_pkthdr.tso_segsz));
5426 	tnl_lso->TCPSeqOffset = 0;
5427 	tnl_lso->EthLenOffset_Size =
5428 	    htobe32(V_CPL_TX_TNL_LSO_SIZE(m0->m_pkthdr.len));
5429 
5430 	return (tnl_lso + 1);
5431 }
5432 
5433 #define VM_TX_L2HDR_LEN	16	/* ethmacdst to vlantci */
5434 
5435 /*
5436  * Write a VM txpkt WR for this packet to the hardware descriptors, update the
5437  * software descriptor, and advance the pidx.  It is guaranteed that enough
5438  * descriptors are available.
5439  *
5440  * The return value is the # of hardware descriptors used.
5441  */
5442 static u_int
5443 write_txpkt_vm_wr(struct adapter *sc, struct sge_txq *txq, struct mbuf *m0)
5444 {
5445 	struct sge_eq *eq;
5446 	struct fw_eth_tx_pkt_vm_wr *wr;
5447 	struct tx_sdesc *txsd;
5448 	struct cpl_tx_pkt_core *cpl;
5449 	uint32_t ctrl;	/* used in many unrelated places */
5450 	uint64_t ctrl1;
5451 	int len16, ndesc, pktlen;
5452 	caddr_t dst;
5453 
5454 	TXQ_LOCK_ASSERT_OWNED(txq);
5455 	M_ASSERTPKTHDR(m0);
5456 
5457 	len16 = mbuf_len16(m0);
5458 	pktlen = m0->m_pkthdr.len;
5459 	ctrl = sizeof(struct cpl_tx_pkt_core);
5460 	if (needs_tso(m0))
5461 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
5462 	ndesc = tx_len16_to_desc(len16);
5463 
5464 	/* Firmware work request header */
5465 	eq = &txq->eq;
5466 	wr = (void *)&eq->desc[eq->pidx];
5467 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_VM_WR) |
5468 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
5469 
5470 	ctrl = V_FW_WR_LEN16(len16);
5471 	wr->equiq_to_len16 = htobe32(ctrl);
5472 	wr->r3[0] = 0;
5473 	wr->r3[1] = 0;
5474 
5475 	/*
5476 	 * Copy over ethmacdst, ethmacsrc, ethtype, and vlantci.
5477 	 * vlantci is ignored unless the ethtype is 0x8100, so it's
5478 	 * simpler to always copy it rather than making it
5479 	 * conditional.  Also, it seems that we do not have to set
5480 	 * vlantci or fake the ethtype when doing VLAN tag insertion.
5481 	 */
5482 	m_copydata(m0, 0, VM_TX_L2HDR_LEN, wr->ethmacdst);
5483 
5484 	if (needs_tso(m0)) {
5485 		cpl = write_lso_cpl(wr + 1, m0);
5486 		txq->tso_wrs++;
5487 	} else
5488 		cpl = (void *)(wr + 1);
5489 
5490 	/* Checksum offload */
5491 	ctrl1 = csum_to_ctrl(sc, m0);
5492 	if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS))
5493 		txq->txcsum++;	/* some hardware assistance provided */
5494 
5495 	/* VLAN tag insertion */
5496 	if (needs_vlan_insertion(m0)) {
5497 		ctrl1 |= F_TXPKT_VLAN_VLD |
5498 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
5499 		txq->vlan_insertion++;
5500 	} else if (sc->vlan_id)
5501 		ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(sc->vlan_id);
5502 
5503 	/* CPL header */
5504 	cpl->ctrl0 = txq->cpl_ctrl0;
5505 	cpl->pack = 0;
5506 	cpl->len = htobe16(pktlen);
5507 	cpl->ctrl1 = htobe64(ctrl1);
5508 
5509 	/* SGL */
5510 	dst = (void *)(cpl + 1);
5511 
5512 	/*
5513 	 * A packet using TSO will use up an entire descriptor for the
5514 	 * firmware work request header, LSO CPL, and TX_PKT_XT CPL.
5515 	 * If this descriptor is the last descriptor in the ring, wrap
5516 	 * around to the front of the ring explicitly for the start of
5517 	 * the sgl.
5518 	 */
5519 	if (dst == (void *)&eq->desc[eq->sidx]) {
5520 		dst = (void *)&eq->desc[0];
5521 		write_gl_to_txd(txq, m0, &dst, 0);
5522 	} else
5523 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
5524 	txq->sgl_wrs++;
5525 	txq->txpkt_wrs++;
5526 
5527 	txsd = &txq->sdesc[eq->pidx];
5528 	txsd->m = m0;
5529 	txsd->desc_used = ndesc;
5530 
5531 	return (ndesc);
5532 }
5533 
5534 /*
5535  * Write a raw WR to the hardware descriptors, update the software
5536  * descriptor, and advance the pidx.  It is guaranteed that enough
5537  * descriptors are available.
5538  *
5539  * The return value is the # of hardware descriptors used.
5540  */
5541 static u_int
5542 write_raw_wr(struct sge_txq *txq, void *wr, struct mbuf *m0, u_int available)
5543 {
5544 	struct sge_eq *eq = &txq->eq;
5545 	struct tx_sdesc *txsd;
5546 	struct mbuf *m;
5547 	caddr_t dst;
5548 	int len16, ndesc;
5549 
5550 	len16 = mbuf_len16(m0);
5551 	ndesc = tx_len16_to_desc(len16);
5552 	MPASS(ndesc <= available);
5553 
5554 	dst = wr;
5555 	for (m = m0; m != NULL; m = m->m_next)
5556 		copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
5557 
5558 	txq->raw_wrs++;
5559 
5560 	txsd = &txq->sdesc[eq->pidx];
5561 	txsd->m = m0;
5562 	txsd->desc_used = ndesc;
5563 
5564 	return (ndesc);
5565 }
5566 
5567 /*
5568  * Write a txpkt WR for this packet to the hardware descriptors, update the
5569  * software descriptor, and advance the pidx.  It is guaranteed that enough
5570  * descriptors are available.
5571  *
5572  * The return value is the # of hardware descriptors used.
5573  */
5574 static u_int
5575 write_txpkt_wr(struct adapter *sc, struct sge_txq *txq, struct mbuf *m0,
5576     u_int available)
5577 {
5578 	struct sge_eq *eq;
5579 	struct fw_eth_tx_pkt_wr *wr;
5580 	struct tx_sdesc *txsd;
5581 	struct cpl_tx_pkt_core *cpl;
5582 	uint32_t ctrl;	/* used in many unrelated places */
5583 	uint64_t ctrl1;
5584 	int len16, ndesc, pktlen, nsegs;
5585 	caddr_t dst;
5586 
5587 	TXQ_LOCK_ASSERT_OWNED(txq);
5588 	M_ASSERTPKTHDR(m0);
5589 
5590 	len16 = mbuf_len16(m0);
5591 	nsegs = mbuf_nsegs(m0);
5592 	pktlen = m0->m_pkthdr.len;
5593 	ctrl = sizeof(struct cpl_tx_pkt_core);
5594 	if (needs_tso(m0)) {
5595 		if (needs_vxlan_tso(m0))
5596 			ctrl += sizeof(struct cpl_tx_tnl_lso);
5597 		else
5598 			ctrl += sizeof(struct cpl_tx_pkt_lso_core);
5599 	} else if (!(mbuf_cflags(m0) & MC_NOMAP) && pktlen <= imm_payload(2) &&
5600 	    available >= 2) {
5601 		/* Immediate data.  Recalculate len16 and set nsegs to 0. */
5602 		ctrl += pktlen;
5603 		len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) +
5604 		    sizeof(struct cpl_tx_pkt_core) + pktlen, 16);
5605 		nsegs = 0;
5606 	}
5607 	ndesc = tx_len16_to_desc(len16);
5608 	MPASS(ndesc <= available);
5609 
5610 	/* Firmware work request header */
5611 	eq = &txq->eq;
5612 	wr = (void *)&eq->desc[eq->pidx];
5613 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
5614 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
5615 
5616 	ctrl = V_FW_WR_LEN16(len16);
5617 	wr->equiq_to_len16 = htobe32(ctrl);
5618 	wr->r3 = 0;
5619 
5620 	if (needs_tso(m0)) {
5621 		if (needs_vxlan_tso(m0)) {
5622 			cpl = write_tnl_lso_cpl(wr + 1, m0);
5623 			txq->vxlan_tso_wrs++;
5624 		} else {
5625 			cpl = write_lso_cpl(wr + 1, m0);
5626 			txq->tso_wrs++;
5627 		}
5628 	} else
5629 		cpl = (void *)(wr + 1);
5630 
5631 	/* Checksum offload */
5632 	ctrl1 = csum_to_ctrl(sc, m0);
5633 	if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS)) {
5634 		/* some hardware assistance provided */
5635 		if (needs_vxlan_csum(m0))
5636 			txq->vxlan_txcsum++;
5637 		else
5638 			txq->txcsum++;
5639 	}
5640 
5641 	/* VLAN tag insertion */
5642 	if (needs_vlan_insertion(m0)) {
5643 		ctrl1 |= F_TXPKT_VLAN_VLD |
5644 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
5645 		txq->vlan_insertion++;
5646 	}
5647 
5648 	/* CPL header */
5649 	cpl->ctrl0 = txq->cpl_ctrl0;
5650 	cpl->pack = 0;
5651 	cpl->len = htobe16(pktlen);
5652 	cpl->ctrl1 = htobe64(ctrl1);
5653 
5654 	/* SGL */
5655 	dst = (void *)(cpl + 1);
5656 	if (__predict_false((uintptr_t)dst == (uintptr_t)&eq->desc[eq->sidx]))
5657 		dst = (caddr_t)&eq->desc[0];
5658 	if (nsegs > 0) {
5659 
5660 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
5661 		txq->sgl_wrs++;
5662 	} else {
5663 		struct mbuf *m;
5664 
5665 		for (m = m0; m != NULL; m = m->m_next) {
5666 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
5667 #ifdef INVARIANTS
5668 			pktlen -= m->m_len;
5669 #endif
5670 		}
5671 #ifdef INVARIANTS
5672 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
5673 #endif
5674 		txq->imm_wrs++;
5675 	}
5676 
5677 	txq->txpkt_wrs++;
5678 
5679 	txsd = &txq->sdesc[eq->pidx];
5680 	txsd->m = m0;
5681 	txsd->desc_used = ndesc;
5682 
5683 	return (ndesc);
5684 }
5685 
5686 static inline bool
5687 cmp_l2hdr(struct txpkts *txp, struct mbuf *m)
5688 {
5689 	int len;
5690 
5691 	MPASS(txp->npkt > 0);
5692 	MPASS(m->m_len >= VM_TX_L2HDR_LEN);
5693 
5694 	if (txp->ethtype == be16toh(ETHERTYPE_VLAN))
5695 		len = VM_TX_L2HDR_LEN;
5696 	else
5697 		len = sizeof(struct ether_header);
5698 
5699 	return (memcmp(m->m_data, &txp->ethmacdst[0], len) != 0);
5700 }
5701 
5702 static inline void
5703 save_l2hdr(struct txpkts *txp, struct mbuf *m)
5704 {
5705 	MPASS(m->m_len >= VM_TX_L2HDR_LEN);
5706 
5707 	memcpy(&txp->ethmacdst[0], mtod(m, const void *), VM_TX_L2HDR_LEN);
5708 }
5709 
5710 static int
5711 add_to_txpkts_vf(struct adapter *sc, struct sge_txq *txq, struct mbuf *m,
5712     int avail, bool *send)
5713 {
5714 	struct txpkts *txp = &txq->txp;
5715 
5716 	/* Cannot have TSO and coalesce at the same time. */
5717 	if (cannot_use_txpkts(m)) {
5718 cannot_coalesce:
5719 		*send = txp->npkt > 0;
5720 		return (EINVAL);
5721 	}
5722 
5723 	/* VF allows coalescing of type 1 (1 GL) only */
5724 	if (mbuf_nsegs(m) > 1)
5725 		goto cannot_coalesce;
5726 
5727 	*send = false;
5728 	if (txp->npkt > 0) {
5729 		MPASS(tx_len16_to_desc(txp->len16) <= avail);
5730 		MPASS(txp->npkt < txp->max_npkt);
5731 		MPASS(txp->wr_type == 1);	/* VF supports type 1 only */
5732 
5733 		if (tx_len16_to_desc(txp->len16 + txpkts1_len16()) > avail) {
5734 retry_after_send:
5735 			*send = true;
5736 			return (EAGAIN);
5737 		}
5738 		if (m->m_pkthdr.len + txp->plen > 65535)
5739 			goto retry_after_send;
5740 		if (cmp_l2hdr(txp, m))
5741 			goto retry_after_send;
5742 
5743 		txp->len16 += txpkts1_len16();
5744 		txp->plen += m->m_pkthdr.len;
5745 		txp->mb[txp->npkt++] = m;
5746 		if (txp->npkt == txp->max_npkt)
5747 			*send = true;
5748 	} else {
5749 		txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_vm_wr), 16) +
5750 		    txpkts1_len16();
5751 		if (tx_len16_to_desc(txp->len16) > avail)
5752 			goto cannot_coalesce;
5753 		txp->npkt = 1;
5754 		txp->wr_type = 1;
5755 		txp->plen = m->m_pkthdr.len;
5756 		txp->mb[0] = m;
5757 		save_l2hdr(txp, m);
5758 	}
5759 	return (0);
5760 }
5761 
5762 static int
5763 add_to_txpkts_pf(struct adapter *sc, struct sge_txq *txq, struct mbuf *m,
5764     int avail, bool *send)
5765 {
5766 	struct txpkts *txp = &txq->txp;
5767 	int nsegs;
5768 
5769 	MPASS(!(sc->flags & IS_VF));
5770 
5771 	/* Cannot have TSO and coalesce at the same time. */
5772 	if (cannot_use_txpkts(m)) {
5773 cannot_coalesce:
5774 		*send = txp->npkt > 0;
5775 		return (EINVAL);
5776 	}
5777 
5778 	*send = false;
5779 	nsegs = mbuf_nsegs(m);
5780 	if (txp->npkt == 0) {
5781 		if (m->m_pkthdr.len > 65535)
5782 			goto cannot_coalesce;
5783 		if (nsegs > 1) {
5784 			txp->wr_type = 0;
5785 			txp->len16 =
5786 			    howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) +
5787 			    txpkts0_len16(nsegs);
5788 		} else {
5789 			txp->wr_type = 1;
5790 			txp->len16 =
5791 			    howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) +
5792 			    txpkts1_len16();
5793 		}
5794 		if (tx_len16_to_desc(txp->len16) > avail)
5795 			goto cannot_coalesce;
5796 		txp->npkt = 1;
5797 		txp->plen = m->m_pkthdr.len;
5798 		txp->mb[0] = m;
5799 	} else {
5800 		MPASS(tx_len16_to_desc(txp->len16) <= avail);
5801 		MPASS(txp->npkt < txp->max_npkt);
5802 
5803 		if (m->m_pkthdr.len + txp->plen > 65535) {
5804 retry_after_send:
5805 			*send = true;
5806 			return (EAGAIN);
5807 		}
5808 
5809 		MPASS(txp->wr_type == 0 || txp->wr_type == 1);
5810 		if (txp->wr_type == 0) {
5811 			if (tx_len16_to_desc(txp->len16 +
5812 			    txpkts0_len16(nsegs)) > min(avail, SGE_MAX_WR_NDESC))
5813 				goto retry_after_send;
5814 			txp->len16 += txpkts0_len16(nsegs);
5815 		} else {
5816 			if (nsegs != 1)
5817 				goto retry_after_send;
5818 			if (tx_len16_to_desc(txp->len16 + txpkts1_len16()) >
5819 			    avail)
5820 				goto retry_after_send;
5821 			txp->len16 += txpkts1_len16();
5822 		}
5823 
5824 		txp->plen += m->m_pkthdr.len;
5825 		txp->mb[txp->npkt++] = m;
5826 		if (txp->npkt == txp->max_npkt)
5827 			*send = true;
5828 	}
5829 	return (0);
5830 }
5831 
5832 /*
5833  * Write a txpkts WR for the packets in txp to the hardware descriptors, update
5834  * the software descriptor, and advance the pidx.  It is guaranteed that enough
5835  * descriptors are available.
5836  *
5837  * The return value is the # of hardware descriptors used.
5838  */
5839 static u_int
5840 write_txpkts_wr(struct adapter *sc, struct sge_txq *txq)
5841 {
5842 	const struct txpkts *txp = &txq->txp;
5843 	struct sge_eq *eq = &txq->eq;
5844 	struct fw_eth_tx_pkts_wr *wr;
5845 	struct tx_sdesc *txsd;
5846 	struct cpl_tx_pkt_core *cpl;
5847 	uint64_t ctrl1;
5848 	int ndesc, i, checkwrap;
5849 	struct mbuf *m, *last;
5850 	void *flitp;
5851 
5852 	TXQ_LOCK_ASSERT_OWNED(txq);
5853 	MPASS(txp->npkt > 0);
5854 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
5855 
5856 	wr = (void *)&eq->desc[eq->pidx];
5857 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
5858 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(txp->len16));
5859 	wr->plen = htobe16(txp->plen);
5860 	wr->npkt = txp->npkt;
5861 	wr->r3 = 0;
5862 	wr->type = txp->wr_type;
5863 	flitp = wr + 1;
5864 
5865 	/*
5866 	 * At this point we are 16B into a hardware descriptor.  If checkwrap is
5867 	 * set then we know the WR is going to wrap around somewhere.  We'll
5868 	 * check for that at appropriate points.
5869 	 */
5870 	ndesc = tx_len16_to_desc(txp->len16);
5871 	last = NULL;
5872 	checkwrap = eq->sidx - ndesc < eq->pidx;
5873 	for (i = 0; i < txp->npkt; i++) {
5874 		m = txp->mb[i];
5875 		if (txp->wr_type == 0) {
5876 			struct ulp_txpkt *ulpmc;
5877 			struct ulptx_idata *ulpsc;
5878 
5879 			/* ULP master command */
5880 			ulpmc = flitp;
5881 			ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
5882 			    V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid));
5883 			ulpmc->len = htobe32(txpkts0_len16(mbuf_nsegs(m)));
5884 
5885 			/* ULP subcommand */
5886 			ulpsc = (void *)(ulpmc + 1);
5887 			ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
5888 			    F_ULP_TX_SC_MORE);
5889 			ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
5890 
5891 			cpl = (void *)(ulpsc + 1);
5892 			if (checkwrap &&
5893 			    (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx])
5894 				cpl = (void *)&eq->desc[0];
5895 		} else {
5896 			cpl = flitp;
5897 		}
5898 
5899 		/* Checksum offload */
5900 		ctrl1 = csum_to_ctrl(sc, m);
5901 		if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS)) {
5902 			/* some hardware assistance provided */
5903 			if (needs_vxlan_csum(m))
5904 				txq->vxlan_txcsum++;
5905 			else
5906 				txq->txcsum++;
5907 		}
5908 
5909 		/* VLAN tag insertion */
5910 		if (needs_vlan_insertion(m)) {
5911 			ctrl1 |= F_TXPKT_VLAN_VLD |
5912 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
5913 			txq->vlan_insertion++;
5914 		}
5915 
5916 		/* CPL header */
5917 		cpl->ctrl0 = txq->cpl_ctrl0;
5918 		cpl->pack = 0;
5919 		cpl->len = htobe16(m->m_pkthdr.len);
5920 		cpl->ctrl1 = htobe64(ctrl1);
5921 
5922 		flitp = cpl + 1;
5923 		if (checkwrap &&
5924 		    (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
5925 			flitp = (void *)&eq->desc[0];
5926 
5927 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap);
5928 
5929 		if (last != NULL)
5930 			last->m_nextpkt = m;
5931 		last = m;
5932 	}
5933 
5934 	txq->sgl_wrs++;
5935 	if (txp->wr_type == 0) {
5936 		txq->txpkts0_pkts += txp->npkt;
5937 		txq->txpkts0_wrs++;
5938 	} else {
5939 		txq->txpkts1_pkts += txp->npkt;
5940 		txq->txpkts1_wrs++;
5941 	}
5942 
5943 	txsd = &txq->sdesc[eq->pidx];
5944 	txsd->m = txp->mb[0];
5945 	txsd->desc_used = ndesc;
5946 
5947 	return (ndesc);
5948 }
5949 
5950 static u_int
5951 write_txpkts_vm_wr(struct adapter *sc, struct sge_txq *txq)
5952 {
5953 	const struct txpkts *txp = &txq->txp;
5954 	struct sge_eq *eq = &txq->eq;
5955 	struct fw_eth_tx_pkts_vm_wr *wr;
5956 	struct tx_sdesc *txsd;
5957 	struct cpl_tx_pkt_core *cpl;
5958 	uint64_t ctrl1;
5959 	int ndesc, i;
5960 	struct mbuf *m, *last;
5961 	void *flitp;
5962 
5963 	TXQ_LOCK_ASSERT_OWNED(txq);
5964 	MPASS(txp->npkt > 0);
5965 	MPASS(txp->wr_type == 1);	/* VF supports type 1 only */
5966 	MPASS(txp->mb[0] != NULL);
5967 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
5968 
5969 	wr = (void *)&eq->desc[eq->pidx];
5970 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_VM_WR));
5971 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(txp->len16));
5972 	wr->r3 = 0;
5973 	wr->plen = htobe16(txp->plen);
5974 	wr->npkt = txp->npkt;
5975 	wr->r4 = 0;
5976 	memcpy(&wr->ethmacdst[0], &txp->ethmacdst[0], 16);
5977 	flitp = wr + 1;
5978 
5979 	/*
5980 	 * At this point we are 32B into a hardware descriptor.  Each mbuf in
5981 	 * the WR will take 32B so we check for the end of the descriptor ring
5982 	 * before writing odd mbufs (mb[1], 3, 5, ..)
5983 	 */
5984 	ndesc = tx_len16_to_desc(txp->len16);
5985 	last = NULL;
5986 	for (i = 0; i < txp->npkt; i++) {
5987 		m = txp->mb[i];
5988 		if (i & 1 && (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
5989 			flitp = &eq->desc[0];
5990 		cpl = flitp;
5991 
5992 		/* Checksum offload */
5993 		ctrl1 = csum_to_ctrl(sc, m);
5994 		if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS))
5995 			txq->txcsum++;	/* some hardware assistance provided */
5996 
5997 		/* VLAN tag insertion */
5998 		if (needs_vlan_insertion(m)) {
5999 			ctrl1 |= F_TXPKT_VLAN_VLD |
6000 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
6001 			txq->vlan_insertion++;
6002 		} else if (sc->vlan_id)
6003 			ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(sc->vlan_id);
6004 
6005 		/* CPL header */
6006 		cpl->ctrl0 = txq->cpl_ctrl0;
6007 		cpl->pack = 0;
6008 		cpl->len = htobe16(m->m_pkthdr.len);
6009 		cpl->ctrl1 = htobe64(ctrl1);
6010 
6011 		flitp = cpl + 1;
6012 		MPASS(mbuf_nsegs(m) == 1);
6013 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), 0);
6014 
6015 		if (last != NULL)
6016 			last->m_nextpkt = m;
6017 		last = m;
6018 	}
6019 
6020 	txq->sgl_wrs++;
6021 	txq->txpkts1_pkts += txp->npkt;
6022 	txq->txpkts1_wrs++;
6023 
6024 	txsd = &txq->sdesc[eq->pidx];
6025 	txsd->m = txp->mb[0];
6026 	txsd->desc_used = ndesc;
6027 
6028 	return (ndesc);
6029 }
6030 
6031 /*
6032  * If the SGL ends on an address that is not 16 byte aligned, this function will
6033  * add a 0 filled flit at the end.
6034  */
6035 static void
6036 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap)
6037 {
6038 	struct sge_eq *eq = &txq->eq;
6039 	struct sglist *gl = txq->gl;
6040 	struct sglist_seg *seg;
6041 	__be64 *flitp, *wrap;
6042 	struct ulptx_sgl *usgl;
6043 	int i, nflits, nsegs;
6044 
6045 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
6046 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
6047 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
6048 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
6049 
6050 	get_pkt_gl(m, gl);
6051 	nsegs = gl->sg_nseg;
6052 	MPASS(nsegs > 0);
6053 
6054 	nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
6055 	flitp = (__be64 *)(*to);
6056 	wrap = (__be64 *)(&eq->desc[eq->sidx]);
6057 	seg = &gl->sg_segs[0];
6058 	usgl = (void *)flitp;
6059 
6060 	/*
6061 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
6062 	 * ring, so we're at least 16 bytes away from the status page.  There is
6063 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
6064 	 */
6065 
6066 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
6067 	    V_ULPTX_NSGE(nsegs));
6068 	usgl->len0 = htobe32(seg->ss_len);
6069 	usgl->addr0 = htobe64(seg->ss_paddr);
6070 	seg++;
6071 
6072 	if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) {
6073 
6074 		/* Won't wrap around at all */
6075 
6076 		for (i = 0; i < nsegs - 1; i++, seg++) {
6077 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
6078 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
6079 		}
6080 		if (i & 1)
6081 			usgl->sge[i / 2].len[1] = htobe32(0);
6082 		flitp += nflits;
6083 	} else {
6084 
6085 		/* Will wrap somewhere in the rest of the SGL */
6086 
6087 		/* 2 flits already written, write the rest flit by flit */
6088 		flitp = (void *)(usgl + 1);
6089 		for (i = 0; i < nflits - 2; i++) {
6090 			if (flitp == wrap)
6091 				flitp = (void *)eq->desc;
6092 			*flitp++ = get_flit(seg, nsegs - 1, i);
6093 		}
6094 	}
6095 
6096 	if (nflits & 1) {
6097 		MPASS(((uintptr_t)flitp) & 0xf);
6098 		*flitp++ = 0;
6099 	}
6100 
6101 	MPASS((((uintptr_t)flitp) & 0xf) == 0);
6102 	if (__predict_false(flitp == wrap))
6103 		*to = (void *)eq->desc;
6104 	else
6105 		*to = (void *)flitp;
6106 }
6107 
6108 static inline void
6109 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
6110 {
6111 
6112 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
6113 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
6114 
6115 	if (__predict_true((uintptr_t)(*to) + len <=
6116 	    (uintptr_t)&eq->desc[eq->sidx])) {
6117 		bcopy(from, *to, len);
6118 		(*to) += len;
6119 	} else {
6120 		int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
6121 
6122 		bcopy(from, *to, portion);
6123 		from += portion;
6124 		portion = len - portion;	/* remaining */
6125 		bcopy(from, (void *)eq->desc, portion);
6126 		(*to) = (caddr_t)eq->desc + portion;
6127 	}
6128 }
6129 
6130 static inline void
6131 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n)
6132 {
6133 	u_int db;
6134 
6135 	MPASS(n > 0);
6136 
6137 	db = eq->doorbells;
6138 	if (n > 1)
6139 		clrbit(&db, DOORBELL_WCWR);
6140 	wmb();
6141 
6142 	switch (ffs(db) - 1) {
6143 	case DOORBELL_UDB:
6144 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
6145 		break;
6146 
6147 	case DOORBELL_WCWR: {
6148 		volatile uint64_t *dst, *src;
6149 		int i;
6150 
6151 		/*
6152 		 * Queues whose 128B doorbell segment fits in the page do not
6153 		 * use relative qid (udb_qid is always 0).  Only queues with
6154 		 * doorbell segments can do WCWR.
6155 		 */
6156 		KASSERT(eq->udb_qid == 0 && n == 1,
6157 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
6158 		    __func__, eq->doorbells, n, eq->dbidx, eq));
6159 
6160 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
6161 		    UDBS_DB_OFFSET);
6162 		i = eq->dbidx;
6163 		src = (void *)&eq->desc[i];
6164 		while (src != (void *)&eq->desc[i + 1])
6165 			*dst++ = *src++;
6166 		wmb();
6167 		break;
6168 	}
6169 
6170 	case DOORBELL_UDBWC:
6171 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
6172 		wmb();
6173 		break;
6174 
6175 	case DOORBELL_KDB:
6176 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
6177 		    V_QID(eq->cntxt_id) | V_PIDX(n));
6178 		break;
6179 	}
6180 
6181 	IDXINCR(eq->dbidx, n, eq->sidx);
6182 }
6183 
6184 static inline u_int
6185 reclaimable_tx_desc(struct sge_eq *eq)
6186 {
6187 	uint16_t hw_cidx;
6188 
6189 	hw_cidx = read_hw_cidx(eq);
6190 	return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx));
6191 }
6192 
6193 static inline u_int
6194 total_available_tx_desc(struct sge_eq *eq)
6195 {
6196 	uint16_t hw_cidx, pidx;
6197 
6198 	hw_cidx = read_hw_cidx(eq);
6199 	pidx = eq->pidx;
6200 
6201 	if (pidx == hw_cidx)
6202 		return (eq->sidx - 1);
6203 	else
6204 		return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1);
6205 }
6206 
6207 static inline uint16_t
6208 read_hw_cidx(struct sge_eq *eq)
6209 {
6210 	struct sge_qstat *spg = (void *)&eq->desc[eq->sidx];
6211 	uint16_t cidx = spg->cidx;	/* stable snapshot */
6212 
6213 	return (be16toh(cidx));
6214 }
6215 
6216 /*
6217  * Reclaim 'n' descriptors approximately.
6218  */
6219 static u_int
6220 reclaim_tx_descs(struct sge_txq *txq, u_int n)
6221 {
6222 	struct tx_sdesc *txsd;
6223 	struct sge_eq *eq = &txq->eq;
6224 	u_int can_reclaim, reclaimed;
6225 
6226 	TXQ_LOCK_ASSERT_OWNED(txq);
6227 	MPASS(n > 0);
6228 
6229 	reclaimed = 0;
6230 	can_reclaim = reclaimable_tx_desc(eq);
6231 	while (can_reclaim && reclaimed < n) {
6232 		int ndesc;
6233 		struct mbuf *m, *nextpkt;
6234 
6235 		txsd = &txq->sdesc[eq->cidx];
6236 		ndesc = txsd->desc_used;
6237 
6238 		/* Firmware doesn't return "partial" credits. */
6239 		KASSERT(can_reclaim >= ndesc,
6240 		    ("%s: unexpected number of credits: %d, %d",
6241 		    __func__, can_reclaim, ndesc));
6242 		KASSERT(ndesc != 0,
6243 		    ("%s: descriptor with no credits: cidx %d",
6244 		    __func__, eq->cidx));
6245 
6246 		for (m = txsd->m; m != NULL; m = nextpkt) {
6247 			nextpkt = m->m_nextpkt;
6248 			m->m_nextpkt = NULL;
6249 			m_freem(m);
6250 		}
6251 		reclaimed += ndesc;
6252 		can_reclaim -= ndesc;
6253 		IDXINCR(eq->cidx, ndesc, eq->sidx);
6254 	}
6255 
6256 	return (reclaimed);
6257 }
6258 
6259 static void
6260 tx_reclaim(void *arg, int n)
6261 {
6262 	struct sge_txq *txq = arg;
6263 	struct sge_eq *eq = &txq->eq;
6264 
6265 	do {
6266 		if (TXQ_TRYLOCK(txq) == 0)
6267 			break;
6268 		n = reclaim_tx_descs(txq, 32);
6269 		if (eq->cidx == eq->pidx)
6270 			eq->equeqidx = eq->pidx;
6271 		TXQ_UNLOCK(txq);
6272 	} while (n > 0);
6273 }
6274 
6275 static __be64
6276 get_flit(struct sglist_seg *segs, int nsegs, int idx)
6277 {
6278 	int i = (idx / 3) * 2;
6279 
6280 	switch (idx % 3) {
6281 	case 0: {
6282 		uint64_t rc;
6283 
6284 		rc = (uint64_t)segs[i].ss_len << 32;
6285 		if (i + 1 < nsegs)
6286 			rc |= (uint64_t)(segs[i + 1].ss_len);
6287 
6288 		return (htobe64(rc));
6289 	}
6290 	case 1:
6291 		return (htobe64(segs[i].ss_paddr));
6292 	case 2:
6293 		return (htobe64(segs[i + 1].ss_paddr));
6294 	}
6295 
6296 	return (0);
6297 }
6298 
6299 static int
6300 find_refill_source(struct adapter *sc, int maxp, bool packing)
6301 {
6302 	int i, zidx = -1;
6303 	struct rx_buf_info *rxb = &sc->sge.rx_buf_info[0];
6304 
6305 	if (packing) {
6306 		for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) {
6307 			if (rxb->hwidx2 == -1)
6308 				continue;
6309 			if (rxb->size1 < PAGE_SIZE &&
6310 			    rxb->size1 < largest_rx_cluster)
6311 				continue;
6312 			if (rxb->size1 > largest_rx_cluster)
6313 				break;
6314 			MPASS(rxb->size1 - rxb->size2 >= CL_METADATA_SIZE);
6315 			if (rxb->size2 >= maxp)
6316 				return (i);
6317 			zidx = i;
6318 		}
6319 	} else {
6320 		for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) {
6321 			if (rxb->hwidx1 == -1)
6322 				continue;
6323 			if (rxb->size1 > largest_rx_cluster)
6324 				break;
6325 			if (rxb->size1 >= maxp)
6326 				return (i);
6327 			zidx = i;
6328 		}
6329 	}
6330 
6331 	return (zidx);
6332 }
6333 
6334 static void
6335 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
6336 {
6337 	mtx_lock(&sc->sfl_lock);
6338 	FL_LOCK(fl);
6339 	if ((fl->flags & FL_DOOMED) == 0) {
6340 		fl->flags |= FL_STARVING;
6341 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
6342 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
6343 	}
6344 	FL_UNLOCK(fl);
6345 	mtx_unlock(&sc->sfl_lock);
6346 }
6347 
6348 static void
6349 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq)
6350 {
6351 	struct sge_wrq *wrq = (void *)eq;
6352 
6353 	atomic_readandclear_int(&eq->equiq);
6354 	taskqueue_enqueue(sc->tq[eq->port_id], &wrq->wrq_tx_task);
6355 }
6356 
6357 static void
6358 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq)
6359 {
6360 	struct sge_txq *txq = (void *)eq;
6361 
6362 	MPASS(eq->type == EQ_ETH);
6363 
6364 	atomic_readandclear_int(&eq->equiq);
6365 	if (mp_ring_is_idle(txq->r))
6366 		taskqueue_enqueue(sc->tq[eq->port_id], &txq->tx_reclaim_task);
6367 	else
6368 		mp_ring_check_drainage(txq->r, 64);
6369 }
6370 
6371 static int
6372 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
6373     struct mbuf *m)
6374 {
6375 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
6376 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
6377 	struct adapter *sc = iq->adapter;
6378 	struct sge *s = &sc->sge;
6379 	struct sge_eq *eq;
6380 	static void (*h[])(struct adapter *, struct sge_eq *) = {NULL,
6381 		&handle_wrq_egr_update, &handle_eth_egr_update,
6382 		&handle_wrq_egr_update};
6383 
6384 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
6385 	    rss->opcode));
6386 
6387 	eq = s->eqmap[qid - s->eq_start - s->eq_base];
6388 	(*h[eq->type])(sc, eq);
6389 
6390 	return (0);
6391 }
6392 
6393 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
6394 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
6395     offsetof(struct cpl_fw6_msg, data));
6396 
6397 static int
6398 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
6399 {
6400 	struct adapter *sc = iq->adapter;
6401 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
6402 
6403 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
6404 	    rss->opcode));
6405 
6406 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
6407 		const struct rss_header *rss2;
6408 
6409 		rss2 = (const struct rss_header *)&cpl->data[0];
6410 		return (t4_cpl_handler[rss2->opcode](iq, rss2, m));
6411 	}
6412 
6413 	return (t4_fw_msg_handler[cpl->type](sc, &cpl->data[0]));
6414 }
6415 
6416 /**
6417  *	t4_handle_wrerr_rpl - process a FW work request error message
6418  *	@adap: the adapter
6419  *	@rpl: start of the FW message
6420  */
6421 static int
6422 t4_handle_wrerr_rpl(struct adapter *adap, const __be64 *rpl)
6423 {
6424 	u8 opcode = *(const u8 *)rpl;
6425 	const struct fw_error_cmd *e = (const void *)rpl;
6426 	unsigned int i;
6427 
6428 	if (opcode != FW_ERROR_CMD) {
6429 		log(LOG_ERR,
6430 		    "%s: Received WRERR_RPL message with opcode %#x\n",
6431 		    device_get_nameunit(adap->dev), opcode);
6432 		return (EINVAL);
6433 	}
6434 	log(LOG_ERR, "%s: FW_ERROR (%s) ", device_get_nameunit(adap->dev),
6435 	    G_FW_ERROR_CMD_FATAL(be32toh(e->op_to_type)) ? "fatal" :
6436 	    "non-fatal");
6437 	switch (G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type))) {
6438 	case FW_ERROR_TYPE_EXCEPTION:
6439 		log(LOG_ERR, "exception info:\n");
6440 		for (i = 0; i < nitems(e->u.exception.info); i++)
6441 			log(LOG_ERR, "%s%08x", i == 0 ? "\t" : " ",
6442 			    be32toh(e->u.exception.info[i]));
6443 		log(LOG_ERR, "\n");
6444 		break;
6445 	case FW_ERROR_TYPE_HWMODULE:
6446 		log(LOG_ERR, "HW module regaddr %08x regval %08x\n",
6447 		    be32toh(e->u.hwmodule.regaddr),
6448 		    be32toh(e->u.hwmodule.regval));
6449 		break;
6450 	case FW_ERROR_TYPE_WR:
6451 		log(LOG_ERR, "WR cidx %d PF %d VF %d eqid %d hdr:\n",
6452 		    be16toh(e->u.wr.cidx),
6453 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.wr.pfn_vfn)),
6454 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.wr.pfn_vfn)),
6455 		    be32toh(e->u.wr.eqid));
6456 		for (i = 0; i < nitems(e->u.wr.wrhdr); i++)
6457 			log(LOG_ERR, "%s%02x", i == 0 ? "\t" : " ",
6458 			    e->u.wr.wrhdr[i]);
6459 		log(LOG_ERR, "\n");
6460 		break;
6461 	case FW_ERROR_TYPE_ACL:
6462 		log(LOG_ERR, "ACL cidx %d PF %d VF %d eqid %d %s",
6463 		    be16toh(e->u.acl.cidx),
6464 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.acl.pfn_vfn)),
6465 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.acl.pfn_vfn)),
6466 		    be32toh(e->u.acl.eqid),
6467 		    G_FW_ERROR_CMD_MV(be16toh(e->u.acl.mv_pkd)) ? "vlanid" :
6468 		    "MAC");
6469 		for (i = 0; i < nitems(e->u.acl.val); i++)
6470 			log(LOG_ERR, " %02x", e->u.acl.val[i]);
6471 		log(LOG_ERR, "\n");
6472 		break;
6473 	default:
6474 		log(LOG_ERR, "type %#x\n",
6475 		    G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type)));
6476 		return (EINVAL);
6477 	}
6478 	return (0);
6479 }
6480 
6481 static inline bool
6482 bufidx_used(struct adapter *sc, int idx)
6483 {
6484 	struct rx_buf_info *rxb = &sc->sge.rx_buf_info[0];
6485 	int i;
6486 
6487 	for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) {
6488 		if (rxb->size1 > largest_rx_cluster)
6489 			continue;
6490 		if (rxb->hwidx1 == idx || rxb->hwidx2 == idx)
6491 			return (true);
6492 	}
6493 
6494 	return (false);
6495 }
6496 
6497 static int
6498 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
6499 {
6500 	struct adapter *sc = arg1;
6501 	struct sge_params *sp = &sc->params.sge;
6502 	int i, rc;
6503 	struct sbuf sb;
6504 	char c;
6505 
6506 	sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
6507 	for (i = 0; i < SGE_FLBUF_SIZES; i++) {
6508 		if (bufidx_used(sc, i))
6509 			c = '*';
6510 		else
6511 			c = '\0';
6512 
6513 		sbuf_printf(&sb, "%u%c ", sp->sge_fl_buffer_size[i], c);
6514 	}
6515 	sbuf_trim(&sb);
6516 	sbuf_finish(&sb);
6517 	rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
6518 	sbuf_delete(&sb);
6519 	return (rc);
6520 }
6521 
6522 #ifdef RATELIMIT
6523 #if defined(INET) || defined(INET6)
6524 /*
6525  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
6526  */
6527 static inline u_int
6528 txpkt_eo_len16(u_int nsegs, u_int immhdrs, u_int tso)
6529 {
6530 	u_int n;
6531 
6532 	MPASS(immhdrs > 0);
6533 
6534 	n = roundup2(sizeof(struct fw_eth_tx_eo_wr) +
6535 	    sizeof(struct cpl_tx_pkt_core) + immhdrs, 16);
6536 	if (__predict_false(nsegs == 0))
6537 		goto done;
6538 
6539 	nsegs--; /* first segment is part of ulptx_sgl */
6540 	n += sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
6541 	if (tso)
6542 		n += sizeof(struct cpl_tx_pkt_lso_core);
6543 
6544 done:
6545 	return (howmany(n, 16));
6546 }
6547 #endif
6548 
6549 #define ETID_FLOWC_NPARAMS 6
6550 #define ETID_FLOWC_LEN (roundup2((sizeof(struct fw_flowc_wr) + \
6551     ETID_FLOWC_NPARAMS * sizeof(struct fw_flowc_mnemval)), 16))
6552 #define ETID_FLOWC_LEN16 (howmany(ETID_FLOWC_LEN, 16))
6553 
6554 #if defined(INET) || defined(INET6)
6555 static int
6556 send_etid_flowc_wr(struct cxgbe_rate_tag *cst, struct port_info *pi,
6557     struct vi_info *vi)
6558 {
6559 	struct wrq_cookie cookie;
6560 	u_int pfvf = pi->adapter->pf << S_FW_VIID_PFN;
6561 	struct fw_flowc_wr *flowc;
6562 
6563 	mtx_assert(&cst->lock, MA_OWNED);
6564 	MPASS((cst->flags & (EO_FLOWC_PENDING | EO_FLOWC_RPL_PENDING)) ==
6565 	    EO_FLOWC_PENDING);
6566 
6567 	flowc = start_wrq_wr(&cst->eo_txq->wrq, ETID_FLOWC_LEN16, &cookie);
6568 	if (__predict_false(flowc == NULL))
6569 		return (ENOMEM);
6570 
6571 	bzero(flowc, ETID_FLOWC_LEN);
6572 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
6573 	    V_FW_FLOWC_WR_NPARAMS(ETID_FLOWC_NPARAMS) | V_FW_WR_COMPL(0));
6574 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(ETID_FLOWC_LEN16) |
6575 	    V_FW_WR_FLOWID(cst->etid));
6576 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN;
6577 	flowc->mnemval[0].val = htobe32(pfvf);
6578 	flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH;
6579 	flowc->mnemval[1].val = htobe32(pi->tx_chan);
6580 	flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT;
6581 	flowc->mnemval[2].val = htobe32(pi->tx_chan);
6582 	flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID;
6583 	flowc->mnemval[3].val = htobe32(cst->iqid);
6584 	flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_EOSTATE;
6585 	flowc->mnemval[4].val = htobe32(FW_FLOWC_MNEM_EOSTATE_ESTABLISHED);
6586 	flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS;
6587 	flowc->mnemval[5].val = htobe32(cst->schedcl);
6588 
6589 	commit_wrq_wr(&cst->eo_txq->wrq, flowc, &cookie);
6590 
6591 	cst->flags &= ~EO_FLOWC_PENDING;
6592 	cst->flags |= EO_FLOWC_RPL_PENDING;
6593 	MPASS(cst->tx_credits >= ETID_FLOWC_LEN16);	/* flowc is first WR. */
6594 	cst->tx_credits -= ETID_FLOWC_LEN16;
6595 
6596 	return (0);
6597 }
6598 #endif
6599 
6600 #define ETID_FLUSH_LEN16 (howmany(sizeof (struct fw_flowc_wr), 16))
6601 
6602 void
6603 send_etid_flush_wr(struct cxgbe_rate_tag *cst)
6604 {
6605 	struct fw_flowc_wr *flowc;
6606 	struct wrq_cookie cookie;
6607 
6608 	mtx_assert(&cst->lock, MA_OWNED);
6609 
6610 	flowc = start_wrq_wr(&cst->eo_txq->wrq, ETID_FLUSH_LEN16, &cookie);
6611 	if (__predict_false(flowc == NULL))
6612 		CXGBE_UNIMPLEMENTED(__func__);
6613 
6614 	bzero(flowc, ETID_FLUSH_LEN16 * 16);
6615 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
6616 	    V_FW_FLOWC_WR_NPARAMS(0) | F_FW_WR_COMPL);
6617 	flowc->flowid_len16 = htobe32(V_FW_WR_LEN16(ETID_FLUSH_LEN16) |
6618 	    V_FW_WR_FLOWID(cst->etid));
6619 
6620 	commit_wrq_wr(&cst->eo_txq->wrq, flowc, &cookie);
6621 
6622 	cst->flags |= EO_FLUSH_RPL_PENDING;
6623 	MPASS(cst->tx_credits >= ETID_FLUSH_LEN16);
6624 	cst->tx_credits -= ETID_FLUSH_LEN16;
6625 	cst->ncompl++;
6626 }
6627 
6628 static void
6629 write_ethofld_wr(struct cxgbe_rate_tag *cst, struct fw_eth_tx_eo_wr *wr,
6630     struct mbuf *m0, int compl)
6631 {
6632 	struct cpl_tx_pkt_core *cpl;
6633 	uint64_t ctrl1;
6634 	uint32_t ctrl;	/* used in many unrelated places */
6635 	int len16, pktlen, nsegs, immhdrs;
6636 	uintptr_t p;
6637 	struct ulptx_sgl *usgl;
6638 	struct sglist sg;
6639 	struct sglist_seg segs[38];	/* XXX: find real limit.  XXX: get off the stack */
6640 
6641 	mtx_assert(&cst->lock, MA_OWNED);
6642 	M_ASSERTPKTHDR(m0);
6643 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
6644 	    m0->m_pkthdr.l4hlen > 0,
6645 	    ("%s: ethofld mbuf %p is missing header lengths", __func__, m0));
6646 
6647 	len16 = mbuf_eo_len16(m0);
6648 	nsegs = mbuf_eo_nsegs(m0);
6649 	pktlen = m0->m_pkthdr.len;
6650 	ctrl = sizeof(struct cpl_tx_pkt_core);
6651 	if (needs_tso(m0))
6652 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
6653 	immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen + m0->m_pkthdr.l4hlen;
6654 	ctrl += immhdrs;
6655 
6656 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_EO_WR) |
6657 	    V_FW_ETH_TX_EO_WR_IMMDLEN(ctrl) | V_FW_WR_COMPL(!!compl));
6658 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(len16) |
6659 	    V_FW_WR_FLOWID(cst->etid));
6660 	wr->r3 = 0;
6661 	if (needs_outer_udp_csum(m0)) {
6662 		wr->u.udpseg.type = FW_ETH_TX_EO_TYPE_UDPSEG;
6663 		wr->u.udpseg.ethlen = m0->m_pkthdr.l2hlen;
6664 		wr->u.udpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
6665 		wr->u.udpseg.udplen = m0->m_pkthdr.l4hlen;
6666 		wr->u.udpseg.rtplen = 0;
6667 		wr->u.udpseg.r4 = 0;
6668 		wr->u.udpseg.mss = htobe16(pktlen - immhdrs);
6669 		wr->u.udpseg.schedpktsize = wr->u.udpseg.mss;
6670 		wr->u.udpseg.plen = htobe32(pktlen - immhdrs);
6671 		cpl = (void *)(wr + 1);
6672 	} else {
6673 		MPASS(needs_outer_tcp_csum(m0));
6674 		wr->u.tcpseg.type = FW_ETH_TX_EO_TYPE_TCPSEG;
6675 		wr->u.tcpseg.ethlen = m0->m_pkthdr.l2hlen;
6676 		wr->u.tcpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
6677 		wr->u.tcpseg.tcplen = m0->m_pkthdr.l4hlen;
6678 		wr->u.tcpseg.tsclk_tsoff = mbuf_eo_tsclk_tsoff(m0);
6679 		wr->u.tcpseg.r4 = 0;
6680 		wr->u.tcpseg.r5 = 0;
6681 		wr->u.tcpseg.plen = htobe32(pktlen - immhdrs);
6682 
6683 		if (needs_tso(m0)) {
6684 			struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
6685 
6686 			wr->u.tcpseg.mss = htobe16(m0->m_pkthdr.tso_segsz);
6687 
6688 			ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) |
6689 			    F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE |
6690 			    V_LSO_ETHHDR_LEN((m0->m_pkthdr.l2hlen -
6691 				ETHER_HDR_LEN) >> 2) |
6692 			    V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) |
6693 			    V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
6694 			if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
6695 				ctrl |= F_LSO_IPV6;
6696 			lso->lso_ctrl = htobe32(ctrl);
6697 			lso->ipid_ofst = htobe16(0);
6698 			lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
6699 			lso->seqno_offset = htobe32(0);
6700 			lso->len = htobe32(pktlen);
6701 
6702 			cpl = (void *)(lso + 1);
6703 		} else {
6704 			wr->u.tcpseg.mss = htobe16(0xffff);
6705 			cpl = (void *)(wr + 1);
6706 		}
6707 	}
6708 
6709 	/* Checksum offload must be requested for ethofld. */
6710 	MPASS(needs_outer_l4_csum(m0));
6711 	ctrl1 = csum_to_ctrl(cst->adapter, m0);
6712 
6713 	/* VLAN tag insertion */
6714 	if (needs_vlan_insertion(m0)) {
6715 		ctrl1 |= F_TXPKT_VLAN_VLD |
6716 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
6717 	}
6718 
6719 	/* CPL header */
6720 	cpl->ctrl0 = cst->ctrl0;
6721 	cpl->pack = 0;
6722 	cpl->len = htobe16(pktlen);
6723 	cpl->ctrl1 = htobe64(ctrl1);
6724 
6725 	/* Copy Ethernet, IP & TCP/UDP hdrs as immediate data */
6726 	p = (uintptr_t)(cpl + 1);
6727 	m_copydata(m0, 0, immhdrs, (void *)p);
6728 
6729 	/* SGL */
6730 	if (nsegs > 0) {
6731 		int i, pad;
6732 
6733 		/* zero-pad upto next 16Byte boundary, if not 16Byte aligned */
6734 		p += immhdrs;
6735 		pad = 16 - (immhdrs & 0xf);
6736 		bzero((void *)p, pad);
6737 
6738 		usgl = (void *)(p + pad);
6739 		usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
6740 		    V_ULPTX_NSGE(nsegs));
6741 
6742 		sglist_init(&sg, nitems(segs), segs);
6743 		for (; m0 != NULL; m0 = m0->m_next) {
6744 			if (__predict_false(m0->m_len == 0))
6745 				continue;
6746 			if (immhdrs >= m0->m_len) {
6747 				immhdrs -= m0->m_len;
6748 				continue;
6749 			}
6750 			if (m0->m_flags & M_EXTPG)
6751 				sglist_append_mbuf_epg(&sg, m0,
6752 				    mtod(m0, vm_offset_t), m0->m_len);
6753                         else
6754 				sglist_append(&sg, mtod(m0, char *) + immhdrs,
6755 				    m0->m_len - immhdrs);
6756 			immhdrs = 0;
6757 		}
6758 		MPASS(sg.sg_nseg == nsegs);
6759 
6760 		/*
6761 		 * Zero pad last 8B in case the WR doesn't end on a 16B
6762 		 * boundary.
6763 		 */
6764 		*(uint64_t *)((char *)wr + len16 * 16 - 8) = 0;
6765 
6766 		usgl->len0 = htobe32(segs[0].ss_len);
6767 		usgl->addr0 = htobe64(segs[0].ss_paddr);
6768 		for (i = 0; i < nsegs - 1; i++) {
6769 			usgl->sge[i / 2].len[i & 1] = htobe32(segs[i + 1].ss_len);
6770 			usgl->sge[i / 2].addr[i & 1] = htobe64(segs[i + 1].ss_paddr);
6771 		}
6772 		if (i & 1)
6773 			usgl->sge[i / 2].len[1] = htobe32(0);
6774 	}
6775 
6776 }
6777 
6778 static void
6779 ethofld_tx(struct cxgbe_rate_tag *cst)
6780 {
6781 	struct mbuf *m;
6782 	struct wrq_cookie cookie;
6783 	int next_credits, compl;
6784 	struct fw_eth_tx_eo_wr *wr;
6785 
6786 	mtx_assert(&cst->lock, MA_OWNED);
6787 
6788 	while ((m = mbufq_first(&cst->pending_tx)) != NULL) {
6789 		M_ASSERTPKTHDR(m);
6790 
6791 		/* How many len16 credits do we need to send this mbuf. */
6792 		next_credits = mbuf_eo_len16(m);
6793 		MPASS(next_credits > 0);
6794 		if (next_credits > cst->tx_credits) {
6795 			/*
6796 			 * Tx will make progress eventually because there is at
6797 			 * least one outstanding fw4_ack that will return
6798 			 * credits and kick the tx.
6799 			 */
6800 			MPASS(cst->ncompl > 0);
6801 			return;
6802 		}
6803 		wr = start_wrq_wr(&cst->eo_txq->wrq, next_credits, &cookie);
6804 		if (__predict_false(wr == NULL)) {
6805 			/* XXX: wishful thinking, not a real assertion. */
6806 			MPASS(cst->ncompl > 0);
6807 			return;
6808 		}
6809 		cst->tx_credits -= next_credits;
6810 		cst->tx_nocompl += next_credits;
6811 		compl = cst->ncompl == 0 || cst->tx_nocompl >= cst->tx_total / 2;
6812 		ETHER_BPF_MTAP(cst->com.ifp, m);
6813 		write_ethofld_wr(cst, wr, m, compl);
6814 		commit_wrq_wr(&cst->eo_txq->wrq, wr, &cookie);
6815 		if (compl) {
6816 			cst->ncompl++;
6817 			cst->tx_nocompl	= 0;
6818 		}
6819 		(void) mbufq_dequeue(&cst->pending_tx);
6820 
6821 		/*
6822 		 * Drop the mbuf's reference on the tag now rather
6823 		 * than waiting until m_freem().  This ensures that
6824 		 * cxgbe_rate_tag_free gets called when the inp drops
6825 		 * its reference on the tag and there are no more
6826 		 * mbufs in the pending_tx queue and can flush any
6827 		 * pending requests.  Otherwise if the last mbuf
6828 		 * doesn't request a completion the etid will never be
6829 		 * released.
6830 		 */
6831 		m->m_pkthdr.snd_tag = NULL;
6832 		m->m_pkthdr.csum_flags &= ~CSUM_SND_TAG;
6833 		m_snd_tag_rele(&cst->com);
6834 
6835 		mbufq_enqueue(&cst->pending_fwack, m);
6836 	}
6837 }
6838 
6839 #if defined(INET) || defined(INET6)
6840 static int
6841 ethofld_transmit(if_t ifp, struct mbuf *m0)
6842 {
6843 	struct cxgbe_rate_tag *cst;
6844 	int rc;
6845 
6846 	MPASS(m0->m_nextpkt == NULL);
6847 	MPASS(m0->m_pkthdr.csum_flags & CSUM_SND_TAG);
6848 	MPASS(m0->m_pkthdr.snd_tag != NULL);
6849 	cst = mst_to_crt(m0->m_pkthdr.snd_tag);
6850 
6851 	mtx_lock(&cst->lock);
6852 	MPASS(cst->flags & EO_SND_TAG_REF);
6853 
6854 	if (__predict_false(cst->flags & EO_FLOWC_PENDING)) {
6855 		struct vi_info *vi = if_getsoftc(ifp);
6856 		struct port_info *pi = vi->pi;
6857 		struct adapter *sc = pi->adapter;
6858 		const uint32_t rss_mask = vi->rss_size - 1;
6859 		uint32_t rss_hash;
6860 
6861 		cst->eo_txq = &sc->sge.ofld_txq[vi->first_ofld_txq];
6862 		if (M_HASHTYPE_ISHASH(m0))
6863 			rss_hash = m0->m_pkthdr.flowid;
6864 		else
6865 			rss_hash = arc4random();
6866 		/* We assume RSS hashing */
6867 		cst->iqid = vi->rss[rss_hash & rss_mask];
6868 		cst->eo_txq += rss_hash % vi->nofldtxq;
6869 		rc = send_etid_flowc_wr(cst, pi, vi);
6870 		if (rc != 0)
6871 			goto done;
6872 	}
6873 
6874 	if (__predict_false(cst->plen + m0->m_pkthdr.len > eo_max_backlog)) {
6875 		rc = ENOBUFS;
6876 		goto done;
6877 	}
6878 
6879 	mbufq_enqueue(&cst->pending_tx, m0);
6880 	cst->plen += m0->m_pkthdr.len;
6881 
6882 	/*
6883 	 * Hold an extra reference on the tag while generating work
6884 	 * requests to ensure that we don't try to free the tag during
6885 	 * ethofld_tx() in case we are sending the final mbuf after
6886 	 * the inp was freed.
6887 	 */
6888 	m_snd_tag_ref(&cst->com);
6889 	ethofld_tx(cst);
6890 	mtx_unlock(&cst->lock);
6891 	m_snd_tag_rele(&cst->com);
6892 	return (0);
6893 
6894 done:
6895 	mtx_unlock(&cst->lock);
6896 	return (rc);
6897 }
6898 #endif
6899 
6900 static int
6901 ethofld_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
6902 {
6903 	struct adapter *sc = iq->adapter;
6904 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
6905 	struct mbuf *m;
6906 	u_int etid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
6907 	struct cxgbe_rate_tag *cst;
6908 	uint8_t credits = cpl->credits;
6909 
6910 	cst = lookup_etid(sc, etid);
6911 	mtx_lock(&cst->lock);
6912 	if (__predict_false(cst->flags & EO_FLOWC_RPL_PENDING)) {
6913 		MPASS(credits >= ETID_FLOWC_LEN16);
6914 		credits -= ETID_FLOWC_LEN16;
6915 		cst->flags &= ~EO_FLOWC_RPL_PENDING;
6916 	}
6917 
6918 	KASSERT(cst->ncompl > 0,
6919 	    ("%s: etid %u (%p) wasn't expecting completion.",
6920 	    __func__, etid, cst));
6921 	cst->ncompl--;
6922 
6923 	while (credits > 0) {
6924 		m = mbufq_dequeue(&cst->pending_fwack);
6925 		if (__predict_false(m == NULL)) {
6926 			/*
6927 			 * The remaining credits are for the final flush that
6928 			 * was issued when the tag was freed by the kernel.
6929 			 */
6930 			MPASS((cst->flags &
6931 			    (EO_FLUSH_RPL_PENDING | EO_SND_TAG_REF)) ==
6932 			    EO_FLUSH_RPL_PENDING);
6933 			MPASS(credits == ETID_FLUSH_LEN16);
6934 			MPASS(cst->tx_credits + cpl->credits == cst->tx_total);
6935 			MPASS(cst->ncompl == 0);
6936 
6937 			cst->flags &= ~EO_FLUSH_RPL_PENDING;
6938 			cst->tx_credits += cpl->credits;
6939 			cxgbe_rate_tag_free_locked(cst);
6940 			return (0);	/* cst is gone. */
6941 		}
6942 		KASSERT(m != NULL,
6943 		    ("%s: too many credits (%u, %u)", __func__, cpl->credits,
6944 		    credits));
6945 		KASSERT(credits >= mbuf_eo_len16(m),
6946 		    ("%s: too few credits (%u, %u, %u)", __func__,
6947 		    cpl->credits, credits, mbuf_eo_len16(m)));
6948 		credits -= mbuf_eo_len16(m);
6949 		cst->plen -= m->m_pkthdr.len;
6950 		m_freem(m);
6951 	}
6952 
6953 	cst->tx_credits += cpl->credits;
6954 	MPASS(cst->tx_credits <= cst->tx_total);
6955 
6956 	if (cst->flags & EO_SND_TAG_REF) {
6957 		/*
6958 		 * As with ethofld_transmit(), hold an extra reference
6959 		 * so that the tag is stable across ethold_tx().
6960 		 */
6961 		m_snd_tag_ref(&cst->com);
6962 		m = mbufq_first(&cst->pending_tx);
6963 		if (m != NULL && cst->tx_credits >= mbuf_eo_len16(m))
6964 			ethofld_tx(cst);
6965 		mtx_unlock(&cst->lock);
6966 		m_snd_tag_rele(&cst->com);
6967 	} else {
6968 		/*
6969 		 * There shouldn't be any pending packets if the tag
6970 		 * was freed by the kernel since any pending packet
6971 		 * should hold a reference to the tag.
6972 		 */
6973 		MPASS(mbufq_first(&cst->pending_tx) == NULL);
6974 		mtx_unlock(&cst->lock);
6975 	}
6976 
6977 	return (0);
6978 }
6979 #endif
6980