xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision f7c32ed617858bcd22f8d1b03199099d50125721)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2011 Chelsio Communications, Inc.
5  * All rights reserved.
6  * Written by: Navdeep Parhar <np@FreeBSD.org>
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include "opt_inet.h"
34 #include "opt_inet6.h"
35 #include "opt_kern_tls.h"
36 #include "opt_ratelimit.h"
37 
38 #include <sys/types.h>
39 #include <sys/eventhandler.h>
40 #include <sys/mbuf.h>
41 #include <sys/socket.h>
42 #include <sys/kernel.h>
43 #include <sys/ktls.h>
44 #include <sys/malloc.h>
45 #include <sys/msan.h>
46 #include <sys/queue.h>
47 #include <sys/sbuf.h>
48 #include <sys/taskqueue.h>
49 #include <sys/time.h>
50 #include <sys/sglist.h>
51 #include <sys/sysctl.h>
52 #include <sys/smp.h>
53 #include <sys/socketvar.h>
54 #include <sys/counter.h>
55 #include <net/bpf.h>
56 #include <net/ethernet.h>
57 #include <net/if.h>
58 #include <net/if_vlan_var.h>
59 #include <net/if_vxlan.h>
60 #include <netinet/in.h>
61 #include <netinet/ip.h>
62 #include <netinet/ip6.h>
63 #include <netinet/tcp.h>
64 #include <netinet/udp.h>
65 #include <machine/in_cksum.h>
66 #include <machine/md_var.h>
67 #include <vm/vm.h>
68 #include <vm/pmap.h>
69 #ifdef DEV_NETMAP
70 #include <machine/bus.h>
71 #include <sys/selinfo.h>
72 #include <net/if_var.h>
73 #include <net/netmap.h>
74 #include <dev/netmap/netmap_kern.h>
75 #endif
76 
77 #include "common/common.h"
78 #include "common/t4_regs.h"
79 #include "common/t4_regs_values.h"
80 #include "common/t4_msg.h"
81 #include "t4_l2t.h"
82 #include "t4_mp_ring.h"
83 
84 #ifdef T4_PKT_TIMESTAMP
85 #define RX_COPY_THRESHOLD (MINCLSIZE - 8)
86 #else
87 #define RX_COPY_THRESHOLD MINCLSIZE
88 #endif
89 
90 /* Internal mbuf flags stored in PH_loc.eight[1]. */
91 #define	MC_NOMAP		0x01
92 #define	MC_RAW_WR		0x02
93 #define	MC_TLS			0x04
94 
95 /*
96  * Ethernet frames are DMA'd at this byte offset into the freelist buffer.
97  * 0-7 are valid values.
98  */
99 static int fl_pktshift = 0;
100 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pktshift, CTLFLAG_RDTUN, &fl_pktshift, 0,
101     "payload DMA offset in rx buffer (bytes)");
102 
103 /*
104  * Pad ethernet payload up to this boundary.
105  * -1: driver should figure out a good value.
106  *  0: disable padding.
107  *  Any power of 2 from 32 to 4096 (both inclusive) is also a valid value.
108  */
109 int fl_pad = -1;
110 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pad, CTLFLAG_RDTUN, &fl_pad, 0,
111     "payload pad boundary (bytes)");
112 
113 /*
114  * Status page length.
115  * -1: driver should figure out a good value.
116  *  64 or 128 are the only other valid values.
117  */
118 static int spg_len = -1;
119 SYSCTL_INT(_hw_cxgbe, OID_AUTO, spg_len, CTLFLAG_RDTUN, &spg_len, 0,
120     "status page size (bytes)");
121 
122 /*
123  * Congestion drops.
124  * -1: no congestion feedback (not recommended).
125  *  0: backpressure the channel instead of dropping packets right away.
126  *  1: no backpressure, drop packets for the congested queue immediately.
127  */
128 static int cong_drop = 0;
129 SYSCTL_INT(_hw_cxgbe, OID_AUTO, cong_drop, CTLFLAG_RDTUN, &cong_drop, 0,
130     "Congestion control for RX queues (0 = backpressure, 1 = drop");
131 
132 /*
133  * Deliver multiple frames in the same free list buffer if they fit.
134  * -1: let the driver decide whether to enable buffer packing or not.
135  *  0: disable buffer packing.
136  *  1: enable buffer packing.
137  */
138 static int buffer_packing = -1;
139 SYSCTL_INT(_hw_cxgbe, OID_AUTO, buffer_packing, CTLFLAG_RDTUN, &buffer_packing,
140     0, "Enable buffer packing");
141 
142 /*
143  * Start next frame in a packed buffer at this boundary.
144  * -1: driver should figure out a good value.
145  * T4: driver will ignore this and use the same value as fl_pad above.
146  * T5: 16, or a power of 2 from 64 to 4096 (both inclusive) is a valid value.
147  */
148 static int fl_pack = -1;
149 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pack, CTLFLAG_RDTUN, &fl_pack, 0,
150     "payload pack boundary (bytes)");
151 
152 /*
153  * Largest rx cluster size that the driver is allowed to allocate.
154  */
155 static int largest_rx_cluster = MJUM16BYTES;
156 SYSCTL_INT(_hw_cxgbe, OID_AUTO, largest_rx_cluster, CTLFLAG_RDTUN,
157     &largest_rx_cluster, 0, "Largest rx cluster (bytes)");
158 
159 /*
160  * Size of cluster allocation that's most likely to succeed.  The driver will
161  * fall back to this size if it fails to allocate clusters larger than this.
162  */
163 static int safest_rx_cluster = PAGE_SIZE;
164 SYSCTL_INT(_hw_cxgbe, OID_AUTO, safest_rx_cluster, CTLFLAG_RDTUN,
165     &safest_rx_cluster, 0, "Safe rx cluster (bytes)");
166 
167 #ifdef RATELIMIT
168 /*
169  * Knob to control TCP timestamp rewriting, and the granularity of the tick used
170  * for rewriting.  -1 and 0-3 are all valid values.
171  * -1: hardware should leave the TCP timestamps alone.
172  * 0: 1ms
173  * 1: 100us
174  * 2: 10us
175  * 3: 1us
176  */
177 static int tsclk = -1;
178 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tsclk, CTLFLAG_RDTUN, &tsclk, 0,
179     "Control TCP timestamp rewriting when using pacing");
180 
181 static int eo_max_backlog = 1024 * 1024;
182 SYSCTL_INT(_hw_cxgbe, OID_AUTO, eo_max_backlog, CTLFLAG_RDTUN, &eo_max_backlog,
183     0, "Maximum backlog of ratelimited data per flow");
184 #endif
185 
186 /*
187  * The interrupt holdoff timers are multiplied by this value on T6+.
188  * 1 and 3-17 (both inclusive) are legal values.
189  */
190 static int tscale = 1;
191 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tscale, CTLFLAG_RDTUN, &tscale, 0,
192     "Interrupt holdoff timer scale on T6+");
193 
194 /*
195  * Number of LRO entries in the lro_ctrl structure per rx queue.
196  */
197 static int lro_entries = TCP_LRO_ENTRIES;
198 SYSCTL_INT(_hw_cxgbe, OID_AUTO, lro_entries, CTLFLAG_RDTUN, &lro_entries, 0,
199     "Number of LRO entries per RX queue");
200 
201 /*
202  * This enables presorting of frames before they're fed into tcp_lro_rx.
203  */
204 static int lro_mbufs = 0;
205 SYSCTL_INT(_hw_cxgbe, OID_AUTO, lro_mbufs, CTLFLAG_RDTUN, &lro_mbufs, 0,
206     "Enable presorting of LRO frames");
207 
208 static counter_u64_t pullups;
209 SYSCTL_COUNTER_U64(_hw_cxgbe, OID_AUTO, pullups, CTLFLAG_RD, &pullups,
210     "Number of mbuf pullups performed");
211 
212 static counter_u64_t defrags;
213 SYSCTL_COUNTER_U64(_hw_cxgbe, OID_AUTO, defrags, CTLFLAG_RD, &defrags,
214     "Number of mbuf defrags performed");
215 
216 static int t4_tx_coalesce = 1;
217 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tx_coalesce, CTLFLAG_RWTUN, &t4_tx_coalesce, 0,
218     "tx coalescing allowed");
219 
220 /*
221  * The driver will make aggressive attempts at tx coalescing if it sees these
222  * many packets eligible for coalescing in quick succession, with no more than
223  * the specified gap in between the eth_tx calls that delivered the packets.
224  */
225 static int t4_tx_coalesce_pkts = 32;
226 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tx_coalesce_pkts, CTLFLAG_RWTUN,
227     &t4_tx_coalesce_pkts, 0,
228     "# of consecutive packets (1 - 255) that will trigger tx coalescing");
229 static int t4_tx_coalesce_gap = 5;
230 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tx_coalesce_gap, CTLFLAG_RWTUN,
231     &t4_tx_coalesce_gap, 0, "tx gap (in microseconds)");
232 
233 static int service_iq(struct sge_iq *, int);
234 static int service_iq_fl(struct sge_iq *, int);
235 static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t);
236 static int eth_rx(struct adapter *, struct sge_rxq *, const struct iq_desc *,
237     u_int);
238 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int,
239     int, int);
240 static inline void init_fl(struct adapter *, struct sge_fl *, int, int, char *);
241 static inline void init_eq(struct adapter *, struct sge_eq *, int, int, uint8_t,
242     struct sge_iq *, char *);
243 static int alloc_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *,
244     struct sysctl_ctx_list *, struct sysctl_oid *);
245 static void free_iq_fl(struct adapter *, struct sge_iq *, struct sge_fl *);
246 static void add_iq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
247     struct sge_iq *);
248 static void add_fl_sysctls(struct adapter *, struct sysctl_ctx_list *,
249     struct sysctl_oid *, struct sge_fl *);
250 static int alloc_iq_fl_hwq(struct vi_info *, struct sge_iq *, struct sge_fl *);
251 static int free_iq_fl_hwq(struct adapter *, struct sge_iq *, struct sge_fl *);
252 static int alloc_fwq(struct adapter *);
253 static void free_fwq(struct adapter *);
254 static int alloc_ctrlq(struct adapter *, int);
255 static void free_ctrlq(struct adapter *, int);
256 static int alloc_rxq(struct vi_info *, struct sge_rxq *, int, int, int);
257 static void free_rxq(struct vi_info *, struct sge_rxq *);
258 static void add_rxq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
259     struct sge_rxq *);
260 #ifdef TCP_OFFLOAD
261 static int alloc_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *, int, int,
262     int);
263 static void free_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *);
264 static void add_ofld_rxq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
265     struct sge_ofld_rxq *);
266 #endif
267 static int ctrl_eq_alloc(struct adapter *, struct sge_eq *);
268 static int eth_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
269 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
270 static int ofld_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
271 #endif
272 static int alloc_eq(struct adapter *, struct sge_eq *, struct sysctl_ctx_list *,
273     struct sysctl_oid *);
274 static void free_eq(struct adapter *, struct sge_eq *);
275 static void add_eq_sysctls(struct adapter *, struct sysctl_ctx_list *,
276     struct sysctl_oid *, struct sge_eq *);
277 static int alloc_eq_hwq(struct adapter *, struct vi_info *, struct sge_eq *);
278 static int free_eq_hwq(struct adapter *, struct vi_info *, struct sge_eq *);
279 static int alloc_wrq(struct adapter *, struct vi_info *, struct sge_wrq *,
280     struct sysctl_ctx_list *, struct sysctl_oid *);
281 static void free_wrq(struct adapter *, struct sge_wrq *);
282 static void add_wrq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
283     struct sge_wrq *);
284 static int alloc_txq(struct vi_info *, struct sge_txq *, int);
285 static void free_txq(struct vi_info *, struct sge_txq *);
286 static void add_txq_sysctls(struct vi_info *, struct sysctl_ctx_list *,
287     struct sysctl_oid *, struct sge_txq *);
288 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
289 static int alloc_ofld_txq(struct vi_info *, struct sge_ofld_txq *, int);
290 static void free_ofld_txq(struct vi_info *, struct sge_ofld_txq *);
291 static void add_ofld_txq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
292     struct sge_ofld_txq *);
293 #endif
294 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int);
295 static inline void ring_fl_db(struct adapter *, struct sge_fl *);
296 static int refill_fl(struct adapter *, struct sge_fl *, int);
297 static void refill_sfl(void *);
298 static int find_refill_source(struct adapter *, int, bool);
299 static void add_fl_to_sfl(struct adapter *, struct sge_fl *);
300 
301 static inline void get_pkt_gl(struct mbuf *, struct sglist *);
302 static inline u_int txpkt_len16(u_int, const u_int);
303 static inline u_int txpkt_vm_len16(u_int, const u_int);
304 static inline void calculate_mbuf_len16(struct mbuf *, bool);
305 static inline u_int txpkts0_len16(u_int);
306 static inline u_int txpkts1_len16(void);
307 static u_int write_raw_wr(struct sge_txq *, void *, struct mbuf *, u_int);
308 static u_int write_txpkt_wr(struct adapter *, struct sge_txq *, struct mbuf *,
309     u_int);
310 static u_int write_txpkt_vm_wr(struct adapter *, struct sge_txq *,
311     struct mbuf *);
312 static int add_to_txpkts_vf(struct adapter *, struct sge_txq *, struct mbuf *,
313     int, bool *);
314 static int add_to_txpkts_pf(struct adapter *, struct sge_txq *, struct mbuf *,
315     int, bool *);
316 static u_int write_txpkts_wr(struct adapter *, struct sge_txq *);
317 static u_int write_txpkts_vm_wr(struct adapter *, struct sge_txq *);
318 static void write_gl_to_txd(struct sge_txq *, struct mbuf *, caddr_t *, int);
319 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int);
320 static inline void ring_eq_db(struct adapter *, struct sge_eq *, u_int);
321 static inline uint16_t read_hw_cidx(struct sge_eq *);
322 static inline u_int reclaimable_tx_desc(struct sge_eq *);
323 static inline u_int total_available_tx_desc(struct sge_eq *);
324 static u_int reclaim_tx_descs(struct sge_txq *, u_int);
325 static void tx_reclaim(void *, int);
326 static __be64 get_flit(struct sglist_seg *, int, int);
327 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
328     struct mbuf *);
329 static int handle_fw_msg(struct sge_iq *, const struct rss_header *,
330     struct mbuf *);
331 static int t4_handle_wrerr_rpl(struct adapter *, const __be64 *);
332 static void wrq_tx_drain(void *, int);
333 static void drain_wrq_wr_list(struct adapter *, struct sge_wrq *);
334 
335 static int sysctl_bufsizes(SYSCTL_HANDLER_ARGS);
336 #ifdef RATELIMIT
337 #if defined(INET) || defined(INET6)
338 static inline u_int txpkt_eo_len16(u_int, u_int, u_int);
339 #endif
340 static int ethofld_fw4_ack(struct sge_iq *, const struct rss_header *,
341     struct mbuf *);
342 #endif
343 
344 static counter_u64_t extfree_refs;
345 static counter_u64_t extfree_rels;
346 
347 an_handler_t t4_an_handler;
348 fw_msg_handler_t t4_fw_msg_handler[NUM_FW6_TYPES];
349 cpl_handler_t t4_cpl_handler[NUM_CPL_CMDS];
350 cpl_handler_t set_tcb_rpl_handlers[NUM_CPL_COOKIES];
351 cpl_handler_t l2t_write_rpl_handlers[NUM_CPL_COOKIES];
352 cpl_handler_t act_open_rpl_handlers[NUM_CPL_COOKIES];
353 cpl_handler_t abort_rpl_rss_handlers[NUM_CPL_COOKIES];
354 cpl_handler_t fw4_ack_handlers[NUM_CPL_COOKIES];
355 
356 void
357 t4_register_an_handler(an_handler_t h)
358 {
359 	uintptr_t *loc;
360 
361 	MPASS(h == NULL || t4_an_handler == NULL);
362 
363 	loc = (uintptr_t *)&t4_an_handler;
364 	atomic_store_rel_ptr(loc, (uintptr_t)h);
365 }
366 
367 void
368 t4_register_fw_msg_handler(int type, fw_msg_handler_t h)
369 {
370 	uintptr_t *loc;
371 
372 	MPASS(type < nitems(t4_fw_msg_handler));
373 	MPASS(h == NULL || t4_fw_msg_handler[type] == NULL);
374 	/*
375 	 * These are dispatched by the handler for FW{4|6}_CPL_MSG using the CPL
376 	 * handler dispatch table.  Reject any attempt to install a handler for
377 	 * this subtype.
378 	 */
379 	MPASS(type != FW_TYPE_RSSCPL);
380 	MPASS(type != FW6_TYPE_RSSCPL);
381 
382 	loc = (uintptr_t *)&t4_fw_msg_handler[type];
383 	atomic_store_rel_ptr(loc, (uintptr_t)h);
384 }
385 
386 void
387 t4_register_cpl_handler(int opcode, cpl_handler_t h)
388 {
389 	uintptr_t *loc;
390 
391 	MPASS(opcode < nitems(t4_cpl_handler));
392 	MPASS(h == NULL || t4_cpl_handler[opcode] == NULL);
393 
394 	loc = (uintptr_t *)&t4_cpl_handler[opcode];
395 	atomic_store_rel_ptr(loc, (uintptr_t)h);
396 }
397 
398 static int
399 set_tcb_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
400     struct mbuf *m)
401 {
402 	const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
403 	u_int tid;
404 	int cookie;
405 
406 	MPASS(m == NULL);
407 
408 	tid = GET_TID(cpl);
409 	if (is_hpftid(iq->adapter, tid) || is_ftid(iq->adapter, tid)) {
410 		/*
411 		 * The return code for filter-write is put in the CPL cookie so
412 		 * we have to rely on the hardware tid (is_ftid) to determine
413 		 * that this is a response to a filter.
414 		 */
415 		cookie = CPL_COOKIE_FILTER;
416 	} else {
417 		cookie = G_COOKIE(cpl->cookie);
418 	}
419 	MPASS(cookie > CPL_COOKIE_RESERVED);
420 	MPASS(cookie < nitems(set_tcb_rpl_handlers));
421 
422 	return (set_tcb_rpl_handlers[cookie](iq, rss, m));
423 }
424 
425 static int
426 l2t_write_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
427     struct mbuf *m)
428 {
429 	const struct cpl_l2t_write_rpl *rpl = (const void *)(rss + 1);
430 	unsigned int cookie;
431 
432 	MPASS(m == NULL);
433 
434 	cookie = GET_TID(rpl) & F_SYNC_WR ? CPL_COOKIE_TOM : CPL_COOKIE_FILTER;
435 	return (l2t_write_rpl_handlers[cookie](iq, rss, m));
436 }
437 
438 static int
439 act_open_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
440     struct mbuf *m)
441 {
442 	const struct cpl_act_open_rpl *cpl = (const void *)(rss + 1);
443 	u_int cookie = G_TID_COOKIE(G_AOPEN_ATID(be32toh(cpl->atid_status)));
444 
445 	MPASS(m == NULL);
446 	MPASS(cookie != CPL_COOKIE_RESERVED);
447 
448 	return (act_open_rpl_handlers[cookie](iq, rss, m));
449 }
450 
451 static int
452 abort_rpl_rss_handler(struct sge_iq *iq, const struct rss_header *rss,
453     struct mbuf *m)
454 {
455 	struct adapter *sc = iq->adapter;
456 	u_int cookie;
457 
458 	MPASS(m == NULL);
459 	if (is_hashfilter(sc))
460 		cookie = CPL_COOKIE_HASHFILTER;
461 	else
462 		cookie = CPL_COOKIE_TOM;
463 
464 	return (abort_rpl_rss_handlers[cookie](iq, rss, m));
465 }
466 
467 static int
468 fw4_ack_handler(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
469 {
470 	struct adapter *sc = iq->adapter;
471 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
472 	unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
473 	u_int cookie;
474 
475 	MPASS(m == NULL);
476 	if (is_etid(sc, tid))
477 		cookie = CPL_COOKIE_ETHOFLD;
478 	else
479 		cookie = CPL_COOKIE_TOM;
480 
481 	return (fw4_ack_handlers[cookie](iq, rss, m));
482 }
483 
484 static void
485 t4_init_shared_cpl_handlers(void)
486 {
487 
488 	t4_register_cpl_handler(CPL_SET_TCB_RPL, set_tcb_rpl_handler);
489 	t4_register_cpl_handler(CPL_L2T_WRITE_RPL, l2t_write_rpl_handler);
490 	t4_register_cpl_handler(CPL_ACT_OPEN_RPL, act_open_rpl_handler);
491 	t4_register_cpl_handler(CPL_ABORT_RPL_RSS, abort_rpl_rss_handler);
492 	t4_register_cpl_handler(CPL_FW4_ACK, fw4_ack_handler);
493 }
494 
495 void
496 t4_register_shared_cpl_handler(int opcode, cpl_handler_t h, int cookie)
497 {
498 	uintptr_t *loc;
499 
500 	MPASS(opcode < nitems(t4_cpl_handler));
501 	MPASS(cookie > CPL_COOKIE_RESERVED);
502 	MPASS(cookie < NUM_CPL_COOKIES);
503 	MPASS(t4_cpl_handler[opcode] != NULL);
504 
505 	switch (opcode) {
506 	case CPL_SET_TCB_RPL:
507 		loc = (uintptr_t *)&set_tcb_rpl_handlers[cookie];
508 		break;
509 	case CPL_L2T_WRITE_RPL:
510 		loc = (uintptr_t *)&l2t_write_rpl_handlers[cookie];
511 		break;
512 	case CPL_ACT_OPEN_RPL:
513 		loc = (uintptr_t *)&act_open_rpl_handlers[cookie];
514 		break;
515 	case CPL_ABORT_RPL_RSS:
516 		loc = (uintptr_t *)&abort_rpl_rss_handlers[cookie];
517 		break;
518 	case CPL_FW4_ACK:
519 		loc = (uintptr_t *)&fw4_ack_handlers[cookie];
520 		break;
521 	default:
522 		MPASS(0);
523 		return;
524 	}
525 	MPASS(h == NULL || *loc == (uintptr_t)NULL);
526 	atomic_store_rel_ptr(loc, (uintptr_t)h);
527 }
528 
529 /*
530  * Called on MOD_LOAD.  Validates and calculates the SGE tunables.
531  */
532 void
533 t4_sge_modload(void)
534 {
535 
536 	if (fl_pktshift < 0 || fl_pktshift > 7) {
537 		printf("Invalid hw.cxgbe.fl_pktshift value (%d),"
538 		    " using 0 instead.\n", fl_pktshift);
539 		fl_pktshift = 0;
540 	}
541 
542 	if (spg_len != 64 && spg_len != 128) {
543 		int len;
544 
545 #if defined(__i386__) || defined(__amd64__)
546 		len = cpu_clflush_line_size > 64 ? 128 : 64;
547 #else
548 		len = 64;
549 #endif
550 		if (spg_len != -1) {
551 			printf("Invalid hw.cxgbe.spg_len value (%d),"
552 			    " using %d instead.\n", spg_len, len);
553 		}
554 		spg_len = len;
555 	}
556 
557 	if (cong_drop < -1 || cong_drop > 1) {
558 		printf("Invalid hw.cxgbe.cong_drop value (%d),"
559 		    " using 0 instead.\n", cong_drop);
560 		cong_drop = 0;
561 	}
562 
563 	if (tscale != 1 && (tscale < 3 || tscale > 17)) {
564 		printf("Invalid hw.cxgbe.tscale value (%d),"
565 		    " using 1 instead.\n", tscale);
566 		tscale = 1;
567 	}
568 
569 	if (largest_rx_cluster != MCLBYTES &&
570 #if MJUMPAGESIZE != MCLBYTES
571 	    largest_rx_cluster != MJUMPAGESIZE &&
572 #endif
573 	    largest_rx_cluster != MJUM9BYTES &&
574 	    largest_rx_cluster != MJUM16BYTES) {
575 		printf("Invalid hw.cxgbe.largest_rx_cluster value (%d),"
576 		    " using %d instead.\n", largest_rx_cluster, MJUM16BYTES);
577 		largest_rx_cluster = MJUM16BYTES;
578 	}
579 
580 	if (safest_rx_cluster != MCLBYTES &&
581 #if MJUMPAGESIZE != MCLBYTES
582 	    safest_rx_cluster != MJUMPAGESIZE &&
583 #endif
584 	    safest_rx_cluster != MJUM9BYTES &&
585 	    safest_rx_cluster != MJUM16BYTES) {
586 		printf("Invalid hw.cxgbe.safest_rx_cluster value (%d),"
587 		    " using %d instead.\n", safest_rx_cluster, MJUMPAGESIZE);
588 		safest_rx_cluster = MJUMPAGESIZE;
589 	}
590 
591 	extfree_refs = counter_u64_alloc(M_WAITOK);
592 	extfree_rels = counter_u64_alloc(M_WAITOK);
593 	pullups = counter_u64_alloc(M_WAITOK);
594 	defrags = counter_u64_alloc(M_WAITOK);
595 	counter_u64_zero(extfree_refs);
596 	counter_u64_zero(extfree_rels);
597 	counter_u64_zero(pullups);
598 	counter_u64_zero(defrags);
599 
600 	t4_init_shared_cpl_handlers();
601 	t4_register_cpl_handler(CPL_FW4_MSG, handle_fw_msg);
602 	t4_register_cpl_handler(CPL_FW6_MSG, handle_fw_msg);
603 	t4_register_cpl_handler(CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
604 #ifdef RATELIMIT
605 	t4_register_shared_cpl_handler(CPL_FW4_ACK, ethofld_fw4_ack,
606 	    CPL_COOKIE_ETHOFLD);
607 #endif
608 	t4_register_fw_msg_handler(FW6_TYPE_CMD_RPL, t4_handle_fw_rpl);
609 	t4_register_fw_msg_handler(FW6_TYPE_WRERR_RPL, t4_handle_wrerr_rpl);
610 }
611 
612 void
613 t4_sge_modunload(void)
614 {
615 
616 	counter_u64_free(extfree_refs);
617 	counter_u64_free(extfree_rels);
618 	counter_u64_free(pullups);
619 	counter_u64_free(defrags);
620 }
621 
622 uint64_t
623 t4_sge_extfree_refs(void)
624 {
625 	uint64_t refs, rels;
626 
627 	rels = counter_u64_fetch(extfree_rels);
628 	refs = counter_u64_fetch(extfree_refs);
629 
630 	return (refs - rels);
631 }
632 
633 /* max 4096 */
634 #define MAX_PACK_BOUNDARY 512
635 
636 static inline void
637 setup_pad_and_pack_boundaries(struct adapter *sc)
638 {
639 	uint32_t v, m;
640 	int pad, pack, pad_shift;
641 
642 	pad_shift = chip_id(sc) > CHELSIO_T5 ? X_T6_INGPADBOUNDARY_SHIFT :
643 	    X_INGPADBOUNDARY_SHIFT;
644 	pad = fl_pad;
645 	if (fl_pad < (1 << pad_shift) ||
646 	    fl_pad > (1 << (pad_shift + M_INGPADBOUNDARY)) ||
647 	    !powerof2(fl_pad)) {
648 		/*
649 		 * If there is any chance that we might use buffer packing and
650 		 * the chip is a T4, then pick 64 as the pad/pack boundary.  Set
651 		 * it to the minimum allowed in all other cases.
652 		 */
653 		pad = is_t4(sc) && buffer_packing ? 64 : 1 << pad_shift;
654 
655 		/*
656 		 * For fl_pad = 0 we'll still write a reasonable value to the
657 		 * register but all the freelists will opt out of padding.
658 		 * We'll complain here only if the user tried to set it to a
659 		 * value greater than 0 that was invalid.
660 		 */
661 		if (fl_pad > 0) {
662 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pad value"
663 			    " (%d), using %d instead.\n", fl_pad, pad);
664 		}
665 	}
666 	m = V_INGPADBOUNDARY(M_INGPADBOUNDARY);
667 	v = V_INGPADBOUNDARY(ilog2(pad) - pad_shift);
668 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
669 
670 	if (is_t4(sc)) {
671 		if (fl_pack != -1 && fl_pack != pad) {
672 			/* Complain but carry on. */
673 			device_printf(sc->dev, "hw.cxgbe.fl_pack (%d) ignored,"
674 			    " using %d instead.\n", fl_pack, pad);
675 		}
676 		return;
677 	}
678 
679 	pack = fl_pack;
680 	if (fl_pack < 16 || fl_pack == 32 || fl_pack > 4096 ||
681 	    !powerof2(fl_pack)) {
682 		if (sc->params.pci.mps > MAX_PACK_BOUNDARY)
683 			pack = MAX_PACK_BOUNDARY;
684 		else
685 			pack = max(sc->params.pci.mps, CACHE_LINE_SIZE);
686 		MPASS(powerof2(pack));
687 		if (pack < 16)
688 			pack = 16;
689 		if (pack == 32)
690 			pack = 64;
691 		if (pack > 4096)
692 			pack = 4096;
693 		if (fl_pack != -1) {
694 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pack value"
695 			    " (%d), using %d instead.\n", fl_pack, pack);
696 		}
697 	}
698 	m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY);
699 	if (pack == 16)
700 		v = V_INGPACKBOUNDARY(0);
701 	else
702 		v = V_INGPACKBOUNDARY(ilog2(pack) - 5);
703 
704 	MPASS(!is_t4(sc));	/* T4 doesn't have SGE_CONTROL2 */
705 	t4_set_reg_field(sc, A_SGE_CONTROL2, m, v);
706 }
707 
708 /*
709  * adap->params.vpd.cclk must be set up before this is called.
710  */
711 void
712 t4_tweak_chip_settings(struct adapter *sc)
713 {
714 	int i, reg;
715 	uint32_t v, m;
716 	int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200};
717 	int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk;
718 	int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */
719 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
720 	static int sw_buf_sizes[] = {
721 		MCLBYTES,
722 #if MJUMPAGESIZE != MCLBYTES
723 		MJUMPAGESIZE,
724 #endif
725 		MJUM9BYTES,
726 		MJUM16BYTES
727 	};
728 
729 	KASSERT(sc->flags & MASTER_PF,
730 	    ("%s: trying to change chip settings when not master.", __func__));
731 
732 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
733 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
734 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
735 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
736 
737 	setup_pad_and_pack_boundaries(sc);
738 
739 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
740 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
741 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
742 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
743 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
744 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
745 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
746 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
747 	t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v);
748 
749 	t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0, 4096);
750 	t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE1, 65536);
751 	reg = A_SGE_FL_BUFFER_SIZE2;
752 	for (i = 0; i < nitems(sw_buf_sizes); i++) {
753 		MPASS(reg <= A_SGE_FL_BUFFER_SIZE15);
754 		t4_write_reg(sc, reg, sw_buf_sizes[i]);
755 		reg += 4;
756 		MPASS(reg <= A_SGE_FL_BUFFER_SIZE15);
757 		t4_write_reg(sc, reg, sw_buf_sizes[i] - CL_METADATA_SIZE);
758 		reg += 4;
759 	}
760 
761 	v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) |
762 	    V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]);
763 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v);
764 
765 	KASSERT(intr_timer[0] <= timer_max,
766 	    ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0],
767 	    timer_max));
768 	for (i = 1; i < nitems(intr_timer); i++) {
769 		KASSERT(intr_timer[i] >= intr_timer[i - 1],
770 		    ("%s: timers not listed in increasing order (%d)",
771 		    __func__, i));
772 
773 		while (intr_timer[i] > timer_max) {
774 			if (i == nitems(intr_timer) - 1) {
775 				intr_timer[i] = timer_max;
776 				break;
777 			}
778 			intr_timer[i] += intr_timer[i - 1];
779 			intr_timer[i] /= 2;
780 		}
781 	}
782 
783 	v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) |
784 	    V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1]));
785 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v);
786 	v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) |
787 	    V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3]));
788 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v);
789 	v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) |
790 	    V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5]));
791 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v);
792 
793 	if (chip_id(sc) >= CHELSIO_T6) {
794 		m = V_TSCALE(M_TSCALE);
795 		if (tscale == 1)
796 			v = 0;
797 		else
798 			v = V_TSCALE(tscale - 2);
799 		t4_set_reg_field(sc, A_SGE_ITP_CONTROL, m, v);
800 
801 		if (sc->debug_flags & DF_DISABLE_TCB_CACHE) {
802 			m = V_RDTHRESHOLD(M_RDTHRESHOLD) | F_WRTHRTHRESHEN |
803 			    V_WRTHRTHRESH(M_WRTHRTHRESH);
804 			t4_tp_pio_read(sc, &v, 1, A_TP_CMM_CONFIG, 1);
805 			v &= ~m;
806 			v |= V_RDTHRESHOLD(1) | F_WRTHRTHRESHEN |
807 			    V_WRTHRTHRESH(16);
808 			t4_tp_pio_write(sc, &v, 1, A_TP_CMM_CONFIG, 1);
809 		}
810 	}
811 
812 	/* 4K, 16K, 64K, 256K DDP "page sizes" for TDDP */
813 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
814 	t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v);
815 
816 	/*
817 	 * 4K, 8K, 16K, 64K DDP "page sizes" for iSCSI DDP.  These have been
818 	 * chosen with MAXPHYS = 128K in mind.  The largest DDP buffer that we
819 	 * may have to deal with is MAXPHYS + 1 page.
820 	 */
821 	v = V_HPZ0(0) | V_HPZ1(1) | V_HPZ2(2) | V_HPZ3(4);
822 	t4_write_reg(sc, A_ULP_RX_ISCSI_PSZ, v);
823 
824 	/* We use multiple DDP page sizes both in plain-TOE and ISCSI modes. */
825 	m = v = F_TDDPTAGTCB | F_ISCSITAGTCB;
826 	t4_set_reg_field(sc, A_ULP_RX_CTL, m, v);
827 
828 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
829 	    F_RESETDDPOFFSET;
830 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
831 	t4_set_reg_field(sc, A_TP_PARA_REG5, m, v);
832 }
833 
834 /*
835  * SGE wants the buffer to be at least 64B and then a multiple of 16.  Its
836  * address mut be 16B aligned.  If padding is in use the buffer's start and end
837  * need to be aligned to the pad boundary as well.  We'll just make sure that
838  * the size is a multiple of the pad boundary here, it is up to the buffer
839  * allocation code to make sure the start of the buffer is aligned.
840  */
841 static inline int
842 hwsz_ok(struct adapter *sc, int hwsz)
843 {
844 	int mask = fl_pad ? sc->params.sge.pad_boundary - 1 : 16 - 1;
845 
846 	return (hwsz >= 64 && (hwsz & mask) == 0);
847 }
848 
849 /*
850  * Initialize the rx buffer sizes and figure out which zones the buffers will
851  * be allocated from.
852  */
853 void
854 t4_init_rx_buf_info(struct adapter *sc)
855 {
856 	struct sge *s = &sc->sge;
857 	struct sge_params *sp = &sc->params.sge;
858 	int i, j, n;
859 	static int sw_buf_sizes[] = {	/* Sorted by size */
860 		MCLBYTES,
861 #if MJUMPAGESIZE != MCLBYTES
862 		MJUMPAGESIZE,
863 #endif
864 		MJUM9BYTES,
865 		MJUM16BYTES
866 	};
867 	struct rx_buf_info *rxb;
868 
869 	s->safe_zidx = -1;
870 	rxb = &s->rx_buf_info[0];
871 	for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) {
872 		rxb->size1 = sw_buf_sizes[i];
873 		rxb->zone = m_getzone(rxb->size1);
874 		rxb->type = m_gettype(rxb->size1);
875 		rxb->size2 = 0;
876 		rxb->hwidx1 = -1;
877 		rxb->hwidx2 = -1;
878 		for (j = 0; j < SGE_FLBUF_SIZES; j++) {
879 			int hwsize = sp->sge_fl_buffer_size[j];
880 
881 			if (!hwsz_ok(sc, hwsize))
882 				continue;
883 
884 			/* hwidx for size1 */
885 			if (rxb->hwidx1 == -1 && rxb->size1 == hwsize)
886 				rxb->hwidx1 = j;
887 
888 			/* hwidx for size2 (buffer packing) */
889 			if (rxb->size1 - CL_METADATA_SIZE < hwsize)
890 				continue;
891 			n = rxb->size1 - hwsize - CL_METADATA_SIZE;
892 			if (n == 0) {
893 				rxb->hwidx2 = j;
894 				rxb->size2 = hwsize;
895 				break;	/* stop looking */
896 			}
897 			if (rxb->hwidx2 != -1) {
898 				if (n < sp->sge_fl_buffer_size[rxb->hwidx2] -
899 				    hwsize - CL_METADATA_SIZE) {
900 					rxb->hwidx2 = j;
901 					rxb->size2 = hwsize;
902 				}
903 			} else if (n <= 2 * CL_METADATA_SIZE) {
904 				rxb->hwidx2 = j;
905 				rxb->size2 = hwsize;
906 			}
907 		}
908 		if (rxb->hwidx2 != -1)
909 			sc->flags |= BUF_PACKING_OK;
910 		if (s->safe_zidx == -1 && rxb->size1 == safest_rx_cluster)
911 			s->safe_zidx = i;
912 	}
913 }
914 
915 /*
916  * Verify some basic SGE settings for the PF and VF driver, and other
917  * miscellaneous settings for the PF driver.
918  */
919 int
920 t4_verify_chip_settings(struct adapter *sc)
921 {
922 	struct sge_params *sp = &sc->params.sge;
923 	uint32_t m, v, r;
924 	int rc = 0;
925 	const uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
926 
927 	m = F_RXPKTCPLMODE;
928 	v = F_RXPKTCPLMODE;
929 	r = sp->sge_control;
930 	if ((r & m) != v) {
931 		device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r);
932 		rc = EINVAL;
933 	}
934 
935 	/*
936 	 * If this changes then every single use of PAGE_SHIFT in the driver
937 	 * needs to be carefully reviewed for PAGE_SHIFT vs sp->page_shift.
938 	 */
939 	if (sp->page_shift != PAGE_SHIFT) {
940 		device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r);
941 		rc = EINVAL;
942 	}
943 
944 	if (sc->flags & IS_VF)
945 		return (0);
946 
947 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
948 	r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ);
949 	if (r != v) {
950 		device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r);
951 		if (sc->vres.ddp.size != 0)
952 			rc = EINVAL;
953 	}
954 
955 	m = v = F_TDDPTAGTCB;
956 	r = t4_read_reg(sc, A_ULP_RX_CTL);
957 	if ((r & m) != v) {
958 		device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r);
959 		if (sc->vres.ddp.size != 0)
960 			rc = EINVAL;
961 	}
962 
963 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
964 	    F_RESETDDPOFFSET;
965 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
966 	r = t4_read_reg(sc, A_TP_PARA_REG5);
967 	if ((r & m) != v) {
968 		device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r);
969 		if (sc->vres.ddp.size != 0)
970 			rc = EINVAL;
971 	}
972 
973 	return (rc);
974 }
975 
976 int
977 t4_create_dma_tag(struct adapter *sc)
978 {
979 	int rc;
980 
981 	rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
982 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
983 	    BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
984 	    NULL, &sc->dmat);
985 	if (rc != 0) {
986 		device_printf(sc->dev,
987 		    "failed to create main DMA tag: %d\n", rc);
988 	}
989 
990 	return (rc);
991 }
992 
993 void
994 t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
995     struct sysctl_oid_list *children)
996 {
997 	struct sge_params *sp = &sc->params.sge;
998 
999 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "buffer_sizes",
1000 	    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, sc, 0,
1001 	    sysctl_bufsizes, "A", "freelist buffer sizes");
1002 
1003 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD,
1004 	    NULL, sp->fl_pktshift, "payload DMA offset in rx buffer (bytes)");
1005 
1006 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD,
1007 	    NULL, sp->pad_boundary, "payload pad boundary (bytes)");
1008 
1009 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD,
1010 	    NULL, sp->spg_len, "status page size (bytes)");
1011 
1012 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD,
1013 	    NULL, cong_drop, "congestion drop setting");
1014 
1015 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD,
1016 	    NULL, sp->pack_boundary, "payload pack boundary (bytes)");
1017 }
1018 
1019 int
1020 t4_destroy_dma_tag(struct adapter *sc)
1021 {
1022 	if (sc->dmat)
1023 		bus_dma_tag_destroy(sc->dmat);
1024 
1025 	return (0);
1026 }
1027 
1028 /*
1029  * Allocate and initialize the firmware event queue, control queues, and special
1030  * purpose rx queues owned by the adapter.
1031  *
1032  * Returns errno on failure.  Resources allocated up to that point may still be
1033  * allocated.  Caller is responsible for cleanup in case this function fails.
1034  */
1035 int
1036 t4_setup_adapter_queues(struct adapter *sc)
1037 {
1038 	int rc, i;
1039 
1040 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
1041 
1042 	/*
1043 	 * Firmware event queue
1044 	 */
1045 	rc = alloc_fwq(sc);
1046 	if (rc != 0)
1047 		return (rc);
1048 
1049 	/*
1050 	 * That's all for the VF driver.
1051 	 */
1052 	if (sc->flags & IS_VF)
1053 		return (rc);
1054 
1055 	/*
1056 	 * XXX: General purpose rx queues, one per port.
1057 	 */
1058 
1059 	/*
1060 	 * Control queues, one per port.
1061 	 */
1062 	for_each_port(sc, i) {
1063 		rc = alloc_ctrlq(sc, i);
1064 		if (rc != 0)
1065 			return (rc);
1066 	}
1067 
1068 	return (rc);
1069 }
1070 
1071 /*
1072  * Idempotent
1073  */
1074 int
1075 t4_teardown_adapter_queues(struct adapter *sc)
1076 {
1077 	int i;
1078 
1079 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
1080 
1081 	if (!(sc->flags & IS_VF)) {
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, struct ifnet *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 + ifp->if_mtu;
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 	struct ifnet *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 (ifp->if_capabilities & 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 (ifp->if_capabilities & 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 (vi->ifp->if_capabilities & 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 (sc->flags & ADAP_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 	t4_slow_intr_handler(sc, verbose);
1316 }
1317 
1318 /*
1319  * Interrupt handler for iq-only queues.  The firmware event queue is the only
1320  * such queue right now.
1321  */
1322 void
1323 t4_intr_evt(void *arg)
1324 {
1325 	struct sge_iq *iq = arg;
1326 
1327 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1328 		service_iq(iq, 0);
1329 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1330 	}
1331 }
1332 
1333 /*
1334  * Interrupt handler for iq+fl queues.
1335  */
1336 void
1337 t4_intr(void *arg)
1338 {
1339 	struct sge_iq *iq = arg;
1340 
1341 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1342 		service_iq_fl(iq, 0);
1343 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1344 	}
1345 }
1346 
1347 #ifdef DEV_NETMAP
1348 /*
1349  * Interrupt handler for netmap rx queues.
1350  */
1351 void
1352 t4_nm_intr(void *arg)
1353 {
1354 	struct sge_nm_rxq *nm_rxq = arg;
1355 
1356 	if (atomic_cmpset_int(&nm_rxq->nm_state, NM_ON, NM_BUSY)) {
1357 		service_nm_rxq(nm_rxq);
1358 		(void) atomic_cmpset_int(&nm_rxq->nm_state, NM_BUSY, NM_ON);
1359 	}
1360 }
1361 
1362 /*
1363  * Interrupt handler for vectors shared between NIC and netmap rx queues.
1364  */
1365 void
1366 t4_vi_intr(void *arg)
1367 {
1368 	struct irq *irq = arg;
1369 
1370 	MPASS(irq->nm_rxq != NULL);
1371 	t4_nm_intr(irq->nm_rxq);
1372 
1373 	MPASS(irq->rxq != NULL);
1374 	t4_intr(irq->rxq);
1375 }
1376 #endif
1377 
1378 /*
1379  * Deals with interrupts on an iq-only (no freelist) queue.
1380  */
1381 static int
1382 service_iq(struct sge_iq *iq, int budget)
1383 {
1384 	struct sge_iq *q;
1385 	struct adapter *sc = iq->adapter;
1386 	struct iq_desc *d = &iq->desc[iq->cidx];
1387 	int ndescs = 0, limit;
1388 	int rsp_type;
1389 	uint32_t lq;
1390 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1391 
1392 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1393 	KASSERT((iq->flags & IQ_HAS_FL) == 0,
1394 	    ("%s: called for iq %p with fl (iq->flags 0x%x)", __func__, iq,
1395 	    iq->flags));
1396 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1397 	MPASS((iq->flags & IQ_LRO_ENABLED) == 0);
1398 
1399 	limit = budget ? budget : iq->qsize / 16;
1400 
1401 	/*
1402 	 * We always come back and check the descriptor ring for new indirect
1403 	 * interrupts and other responses after running a single handler.
1404 	 */
1405 	for (;;) {
1406 		while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1407 
1408 			rmb();
1409 
1410 			rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1411 			lq = be32toh(d->rsp.pldbuflen_qid);
1412 
1413 			switch (rsp_type) {
1414 			case X_RSPD_TYPE_FLBUF:
1415 				panic("%s: data for an iq (%p) with no freelist",
1416 				    __func__, iq);
1417 
1418 				/* NOTREACHED */
1419 
1420 			case X_RSPD_TYPE_CPL:
1421 				KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1422 				    ("%s: bad opcode %02x.", __func__,
1423 				    d->rss.opcode));
1424 				t4_cpl_handler[d->rss.opcode](iq, &d->rss, NULL);
1425 				break;
1426 
1427 			case X_RSPD_TYPE_INTR:
1428 				/*
1429 				 * There are 1K interrupt-capable queues (qids 0
1430 				 * through 1023).  A response type indicating a
1431 				 * forwarded interrupt with a qid >= 1K is an
1432 				 * iWARP async notification.
1433 				 */
1434 				if (__predict_true(lq >= 1024)) {
1435 					t4_an_handler(iq, &d->rsp);
1436 					break;
1437 				}
1438 
1439 				q = sc->sge.iqmap[lq - sc->sge.iq_start -
1440 				    sc->sge.iq_base];
1441 				if (atomic_cmpset_int(&q->state, IQS_IDLE,
1442 				    IQS_BUSY)) {
1443 					if (service_iq_fl(q, q->qsize / 16) == 0) {
1444 						(void) atomic_cmpset_int(&q->state,
1445 						    IQS_BUSY, IQS_IDLE);
1446 					} else {
1447 						STAILQ_INSERT_TAIL(&iql, q,
1448 						    link);
1449 					}
1450 				}
1451 				break;
1452 
1453 			default:
1454 				KASSERT(0,
1455 				    ("%s: illegal response type %d on iq %p",
1456 				    __func__, rsp_type, iq));
1457 				log(LOG_ERR,
1458 				    "%s: illegal response type %d on iq %p",
1459 				    device_get_nameunit(sc->dev), rsp_type, iq);
1460 				break;
1461 			}
1462 
1463 			d++;
1464 			if (__predict_false(++iq->cidx == iq->sidx)) {
1465 				iq->cidx = 0;
1466 				iq->gen ^= F_RSPD_GEN;
1467 				d = &iq->desc[0];
1468 			}
1469 			if (__predict_false(++ndescs == limit)) {
1470 				t4_write_reg(sc, sc->sge_gts_reg,
1471 				    V_CIDXINC(ndescs) |
1472 				    V_INGRESSQID(iq->cntxt_id) |
1473 				    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1474 				ndescs = 0;
1475 
1476 				if (budget) {
1477 					return (EINPROGRESS);
1478 				}
1479 			}
1480 		}
1481 
1482 		if (STAILQ_EMPTY(&iql))
1483 			break;
1484 
1485 		/*
1486 		 * Process the head only, and send it to the back of the list if
1487 		 * it's still not done.
1488 		 */
1489 		q = STAILQ_FIRST(&iql);
1490 		STAILQ_REMOVE_HEAD(&iql, link);
1491 		if (service_iq_fl(q, q->qsize / 8) == 0)
1492 			(void) atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1493 		else
1494 			STAILQ_INSERT_TAIL(&iql, q, link);
1495 	}
1496 
1497 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1498 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1499 
1500 	return (0);
1501 }
1502 
1503 #if defined(INET) || defined(INET6)
1504 static inline int
1505 sort_before_lro(struct lro_ctrl *lro)
1506 {
1507 
1508 	return (lro->lro_mbuf_max != 0);
1509 }
1510 #endif
1511 
1512 static inline uint64_t
1513 last_flit_to_ns(struct adapter *sc, uint64_t lf)
1514 {
1515 	uint64_t n = be64toh(lf) & 0xfffffffffffffff;	/* 60b, not 64b. */
1516 
1517 	if (n > UINT64_MAX / 1000000)
1518 		return (n / sc->params.vpd.cclk * 1000000);
1519 	else
1520 		return (n * 1000000 / sc->params.vpd.cclk);
1521 }
1522 
1523 static inline void
1524 move_to_next_rxbuf(struct sge_fl *fl)
1525 {
1526 
1527 	fl->rx_offset = 0;
1528 	if (__predict_false((++fl->cidx & 7) == 0)) {
1529 		uint16_t cidx = fl->cidx >> 3;
1530 
1531 		if (__predict_false(cidx == fl->sidx))
1532 			fl->cidx = cidx = 0;
1533 		fl->hw_cidx = cidx;
1534 	}
1535 }
1536 
1537 /*
1538  * Deals with interrupts on an iq+fl queue.
1539  */
1540 static int
1541 service_iq_fl(struct sge_iq *iq, int budget)
1542 {
1543 	struct sge_rxq *rxq = iq_to_rxq(iq);
1544 	struct sge_fl *fl;
1545 	struct adapter *sc = iq->adapter;
1546 	struct iq_desc *d = &iq->desc[iq->cidx];
1547 	int ndescs, limit;
1548 	int rsp_type, starved;
1549 	uint32_t lq;
1550 	uint16_t fl_hw_cidx;
1551 	struct mbuf *m0;
1552 #if defined(INET) || defined(INET6)
1553 	const struct timeval lro_timeout = {0, sc->lro_timeout};
1554 	struct lro_ctrl *lro = &rxq->lro;
1555 #endif
1556 
1557 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1558 	MPASS(iq->flags & IQ_HAS_FL);
1559 
1560 	ndescs = 0;
1561 #if defined(INET) || defined(INET6)
1562 	if (iq->flags & IQ_ADJ_CREDIT) {
1563 		MPASS(sort_before_lro(lro));
1564 		iq->flags &= ~IQ_ADJ_CREDIT;
1565 		if ((d->rsp.u.type_gen & F_RSPD_GEN) != iq->gen) {
1566 			tcp_lro_flush_all(lro);
1567 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(1) |
1568 			    V_INGRESSQID((u32)iq->cntxt_id) |
1569 			    V_SEINTARM(iq->intr_params));
1570 			return (0);
1571 		}
1572 		ndescs = 1;
1573 	}
1574 #else
1575 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1576 #endif
1577 
1578 	limit = budget ? budget : iq->qsize / 16;
1579 	fl = &rxq->fl;
1580 	fl_hw_cidx = fl->hw_cidx;	/* stable snapshot */
1581 	while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1582 
1583 		rmb();
1584 
1585 		m0 = NULL;
1586 		rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1587 		lq = be32toh(d->rsp.pldbuflen_qid);
1588 
1589 		switch (rsp_type) {
1590 		case X_RSPD_TYPE_FLBUF:
1591 			if (lq & F_RSPD_NEWBUF) {
1592 				if (fl->rx_offset > 0)
1593 					move_to_next_rxbuf(fl);
1594 				lq = G_RSPD_LEN(lq);
1595 			}
1596 			if (IDXDIFF(fl->hw_cidx, fl_hw_cidx, fl->sidx) > 4) {
1597 				FL_LOCK(fl);
1598 				refill_fl(sc, fl, 64);
1599 				FL_UNLOCK(fl);
1600 				fl_hw_cidx = fl->hw_cidx;
1601 			}
1602 
1603 			if (d->rss.opcode == CPL_RX_PKT) {
1604 				if (__predict_true(eth_rx(sc, rxq, d, lq) == 0))
1605 					break;
1606 				goto out;
1607 			}
1608 			m0 = get_fl_payload(sc, fl, lq);
1609 			if (__predict_false(m0 == NULL))
1610 				goto out;
1611 
1612 			/* fall through */
1613 
1614 		case X_RSPD_TYPE_CPL:
1615 			KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1616 			    ("%s: bad opcode %02x.", __func__, d->rss.opcode));
1617 			t4_cpl_handler[d->rss.opcode](iq, &d->rss, m0);
1618 			break;
1619 
1620 		case X_RSPD_TYPE_INTR:
1621 
1622 			/*
1623 			 * There are 1K interrupt-capable queues (qids 0
1624 			 * through 1023).  A response type indicating a
1625 			 * forwarded interrupt with a qid >= 1K is an
1626 			 * iWARP async notification.  That is the only
1627 			 * acceptable indirect interrupt on this queue.
1628 			 */
1629 			if (__predict_false(lq < 1024)) {
1630 				panic("%s: indirect interrupt on iq_fl %p "
1631 				    "with qid %u", __func__, iq, lq);
1632 			}
1633 
1634 			t4_an_handler(iq, &d->rsp);
1635 			break;
1636 
1637 		default:
1638 			KASSERT(0, ("%s: illegal response type %d on iq %p",
1639 			    __func__, rsp_type, iq));
1640 			log(LOG_ERR, "%s: illegal response type %d on iq %p",
1641 			    device_get_nameunit(sc->dev), rsp_type, iq);
1642 			break;
1643 		}
1644 
1645 		d++;
1646 		if (__predict_false(++iq->cidx == iq->sidx)) {
1647 			iq->cidx = 0;
1648 			iq->gen ^= F_RSPD_GEN;
1649 			d = &iq->desc[0];
1650 		}
1651 		if (__predict_false(++ndescs == limit)) {
1652 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1653 			    V_INGRESSQID(iq->cntxt_id) |
1654 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1655 
1656 #if defined(INET) || defined(INET6)
1657 			if (iq->flags & IQ_LRO_ENABLED &&
1658 			    !sort_before_lro(lro) &&
1659 			    sc->lro_timeout != 0) {
1660 				tcp_lro_flush_inactive(lro, &lro_timeout);
1661 			}
1662 #endif
1663 			if (budget)
1664 				return (EINPROGRESS);
1665 			ndescs = 0;
1666 		}
1667 	}
1668 out:
1669 #if defined(INET) || defined(INET6)
1670 	if (iq->flags & IQ_LRO_ENABLED) {
1671 		if (ndescs > 0 && lro->lro_mbuf_count > 8) {
1672 			MPASS(sort_before_lro(lro));
1673 			/* hold back one credit and don't flush LRO state */
1674 			iq->flags |= IQ_ADJ_CREDIT;
1675 			ndescs--;
1676 		} else {
1677 			tcp_lro_flush_all(lro);
1678 		}
1679 	}
1680 #endif
1681 
1682 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1683 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1684 
1685 	FL_LOCK(fl);
1686 	starved = refill_fl(sc, fl, 64);
1687 	FL_UNLOCK(fl);
1688 	if (__predict_false(starved != 0))
1689 		add_fl_to_sfl(sc, fl);
1690 
1691 	return (0);
1692 }
1693 
1694 static inline struct cluster_metadata *
1695 cl_metadata(struct fl_sdesc *sd)
1696 {
1697 
1698 	return ((void *)(sd->cl + sd->moff));
1699 }
1700 
1701 static void
1702 rxb_free(struct mbuf *m)
1703 {
1704 	struct cluster_metadata *clm = m->m_ext.ext_arg1;
1705 
1706 	uma_zfree(clm->zone, clm->cl);
1707 	counter_u64_add(extfree_rels, 1);
1708 }
1709 
1710 /*
1711  * The mbuf returned comes from zone_muf and carries the payload in one of these
1712  * ways
1713  * a) complete frame inside the mbuf
1714  * b) m_cljset (for clusters without metadata)
1715  * d) m_extaddref (cluster with metadata)
1716  */
1717 static struct mbuf *
1718 get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1719     int remaining)
1720 {
1721 	struct mbuf *m;
1722 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1723 	struct rx_buf_info *rxb = &sc->sge.rx_buf_info[sd->zidx];
1724 	struct cluster_metadata *clm;
1725 	int len, blen;
1726 	caddr_t payload;
1727 
1728 	if (fl->flags & FL_BUF_PACKING) {
1729 		u_int l, pad;
1730 
1731 		blen = rxb->size2 - fl->rx_offset;	/* max possible in this buf */
1732 		len = min(remaining, blen);
1733 		payload = sd->cl + fl->rx_offset;
1734 
1735 		l = fr_offset + len;
1736 		pad = roundup2(l, fl->buf_boundary) - l;
1737 		if (fl->rx_offset + len + pad < rxb->size2)
1738 			blen = len + pad;
1739 		MPASS(fl->rx_offset + blen <= rxb->size2);
1740 	} else {
1741 		MPASS(fl->rx_offset == 0);	/* not packing */
1742 		blen = rxb->size1;
1743 		len = min(remaining, blen);
1744 		payload = sd->cl;
1745 	}
1746 
1747 	if (fr_offset == 0) {
1748 		m = m_gethdr(M_NOWAIT, MT_DATA);
1749 		if (__predict_false(m == NULL))
1750 			return (NULL);
1751 		m->m_pkthdr.len = remaining;
1752 	} else {
1753 		m = m_get(M_NOWAIT, MT_DATA);
1754 		if (__predict_false(m == NULL))
1755 			return (NULL);
1756 	}
1757 	m->m_len = len;
1758 	kmsan_mark(payload, len, KMSAN_STATE_INITED);
1759 
1760 	if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) {
1761 		/* copy data to mbuf */
1762 		bcopy(payload, mtod(m, caddr_t), len);
1763 		if (fl->flags & FL_BUF_PACKING) {
1764 			fl->rx_offset += blen;
1765 			MPASS(fl->rx_offset <= rxb->size2);
1766 			if (fl->rx_offset < rxb->size2)
1767 				return (m);	/* without advancing the cidx */
1768 		}
1769 	} else if (fl->flags & FL_BUF_PACKING) {
1770 		clm = cl_metadata(sd);
1771 		if (sd->nmbuf++ == 0) {
1772 			clm->refcount = 1;
1773 			clm->zone = rxb->zone;
1774 			clm->cl = sd->cl;
1775 			counter_u64_add(extfree_refs, 1);
1776 		}
1777 		m_extaddref(m, payload, blen, &clm->refcount, rxb_free, clm,
1778 		    NULL);
1779 
1780 		fl->rx_offset += blen;
1781 		MPASS(fl->rx_offset <= rxb->size2);
1782 		if (fl->rx_offset < rxb->size2)
1783 			return (m);	/* without advancing the cidx */
1784 	} else {
1785 		m_cljset(m, sd->cl, rxb->type);
1786 		sd->cl = NULL;	/* consumed, not a recycle candidate */
1787 	}
1788 
1789 	move_to_next_rxbuf(fl);
1790 
1791 	return (m);
1792 }
1793 
1794 static struct mbuf *
1795 get_fl_payload(struct adapter *sc, struct sge_fl *fl, const u_int plen)
1796 {
1797 	struct mbuf *m0, *m, **pnext;
1798 	u_int remaining;
1799 
1800 	if (__predict_false(fl->flags & FL_BUF_RESUME)) {
1801 		M_ASSERTPKTHDR(fl->m0);
1802 		MPASS(fl->m0->m_pkthdr.len == plen);
1803 		MPASS(fl->remaining < plen);
1804 
1805 		m0 = fl->m0;
1806 		pnext = fl->pnext;
1807 		remaining = fl->remaining;
1808 		fl->flags &= ~FL_BUF_RESUME;
1809 		goto get_segment;
1810 	}
1811 
1812 	/*
1813 	 * Payload starts at rx_offset in the current hw buffer.  Its length is
1814 	 * 'len' and it may span multiple hw buffers.
1815 	 */
1816 
1817 	m0 = get_scatter_segment(sc, fl, 0, plen);
1818 	if (m0 == NULL)
1819 		return (NULL);
1820 	remaining = plen - m0->m_len;
1821 	pnext = &m0->m_next;
1822 	while (remaining > 0) {
1823 get_segment:
1824 		MPASS(fl->rx_offset == 0);
1825 		m = get_scatter_segment(sc, fl, plen - remaining, remaining);
1826 		if (__predict_false(m == NULL)) {
1827 			fl->m0 = m0;
1828 			fl->pnext = pnext;
1829 			fl->remaining = remaining;
1830 			fl->flags |= FL_BUF_RESUME;
1831 			return (NULL);
1832 		}
1833 		*pnext = m;
1834 		pnext = &m->m_next;
1835 		remaining -= m->m_len;
1836 	}
1837 	*pnext = NULL;
1838 
1839 	M_ASSERTPKTHDR(m0);
1840 	return (m0);
1841 }
1842 
1843 static int
1844 skip_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1845     int remaining)
1846 {
1847 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1848 	struct rx_buf_info *rxb = &sc->sge.rx_buf_info[sd->zidx];
1849 	int len, blen;
1850 
1851 	if (fl->flags & FL_BUF_PACKING) {
1852 		u_int l, pad;
1853 
1854 		blen = rxb->size2 - fl->rx_offset;	/* max possible in this buf */
1855 		len = min(remaining, blen);
1856 
1857 		l = fr_offset + len;
1858 		pad = roundup2(l, fl->buf_boundary) - l;
1859 		if (fl->rx_offset + len + pad < rxb->size2)
1860 			blen = len + pad;
1861 		fl->rx_offset += blen;
1862 		MPASS(fl->rx_offset <= rxb->size2);
1863 		if (fl->rx_offset < rxb->size2)
1864 			return (len);	/* without advancing the cidx */
1865 	} else {
1866 		MPASS(fl->rx_offset == 0);	/* not packing */
1867 		blen = rxb->size1;
1868 		len = min(remaining, blen);
1869 	}
1870 	move_to_next_rxbuf(fl);
1871 	return (len);
1872 }
1873 
1874 static inline void
1875 skip_fl_payload(struct adapter *sc, struct sge_fl *fl, int plen)
1876 {
1877 	int remaining, fr_offset, len;
1878 
1879 	fr_offset = 0;
1880 	remaining = plen;
1881 	while (remaining > 0) {
1882 		len = skip_scatter_segment(sc, fl, fr_offset, remaining);
1883 		fr_offset += len;
1884 		remaining -= len;
1885 	}
1886 }
1887 
1888 static inline int
1889 get_segment_len(struct adapter *sc, struct sge_fl *fl, int plen)
1890 {
1891 	int len;
1892 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1893 	struct rx_buf_info *rxb = &sc->sge.rx_buf_info[sd->zidx];
1894 
1895 	if (fl->flags & FL_BUF_PACKING)
1896 		len = rxb->size2 - fl->rx_offset;
1897 	else
1898 		len = rxb->size1;
1899 
1900 	return (min(plen, len));
1901 }
1902 
1903 static int
1904 eth_rx(struct adapter *sc, struct sge_rxq *rxq, const struct iq_desc *d,
1905     u_int plen)
1906 {
1907 	struct mbuf *m0;
1908 	struct ifnet *ifp = rxq->ifp;
1909 	struct sge_fl *fl = &rxq->fl;
1910 	struct vi_info *vi = ifp->if_softc;
1911 	const struct cpl_rx_pkt *cpl;
1912 #if defined(INET) || defined(INET6)
1913 	struct lro_ctrl *lro = &rxq->lro;
1914 #endif
1915 	uint16_t err_vec, tnl_type, tnlhdr_len;
1916 	static const int sw_hashtype[4][2] = {
1917 		{M_HASHTYPE_NONE, M_HASHTYPE_NONE},
1918 		{M_HASHTYPE_RSS_IPV4, M_HASHTYPE_RSS_IPV6},
1919 		{M_HASHTYPE_RSS_TCP_IPV4, M_HASHTYPE_RSS_TCP_IPV6},
1920 		{M_HASHTYPE_RSS_UDP_IPV4, M_HASHTYPE_RSS_UDP_IPV6},
1921 	};
1922 	static const int sw_csum_flags[2][2] = {
1923 		{
1924 			/* IP, inner IP */
1925 			CSUM_ENCAP_VXLAN |
1926 			    CSUM_L3_CALC | CSUM_L3_VALID |
1927 			    CSUM_L4_CALC | CSUM_L4_VALID |
1928 			    CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID |
1929 			    CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID,
1930 
1931 			/* IP, inner IP6 */
1932 			CSUM_ENCAP_VXLAN |
1933 			    CSUM_L3_CALC | CSUM_L3_VALID |
1934 			    CSUM_L4_CALC | CSUM_L4_VALID |
1935 			    CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID,
1936 		},
1937 		{
1938 			/* IP6, inner IP */
1939 			CSUM_ENCAP_VXLAN |
1940 			    CSUM_L4_CALC | CSUM_L4_VALID |
1941 			    CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID |
1942 			    CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID,
1943 
1944 			/* IP6, inner IP6 */
1945 			CSUM_ENCAP_VXLAN |
1946 			    CSUM_L4_CALC | CSUM_L4_VALID |
1947 			    CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID,
1948 		},
1949 	};
1950 
1951 	MPASS(plen > sc->params.sge.fl_pktshift);
1952 	if (vi->pfil != NULL && PFIL_HOOKED_IN(vi->pfil) &&
1953 	    __predict_true((fl->flags & FL_BUF_RESUME) == 0)) {
1954 		struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1955 		caddr_t frame;
1956 		int rc, slen;
1957 
1958 		slen = get_segment_len(sc, fl, plen) -
1959 		    sc->params.sge.fl_pktshift;
1960 		frame = sd->cl + fl->rx_offset + sc->params.sge.fl_pktshift;
1961 		CURVNET_SET_QUIET(ifp->if_vnet);
1962 		rc = pfil_run_hooks(vi->pfil, frame, ifp,
1963 		    slen | PFIL_MEMPTR | PFIL_IN, NULL);
1964 		CURVNET_RESTORE();
1965 		if (rc == PFIL_DROPPED || rc == PFIL_CONSUMED) {
1966 			skip_fl_payload(sc, fl, plen);
1967 			return (0);
1968 		}
1969 		if (rc == PFIL_REALLOCED) {
1970 			skip_fl_payload(sc, fl, plen);
1971 			m0 = pfil_mem2mbuf(frame);
1972 			goto have_mbuf;
1973 		}
1974 	}
1975 
1976 	m0 = get_fl_payload(sc, fl, plen);
1977 	if (__predict_false(m0 == NULL))
1978 		return (ENOMEM);
1979 
1980 	m0->m_pkthdr.len -= sc->params.sge.fl_pktshift;
1981 	m0->m_len -= sc->params.sge.fl_pktshift;
1982 	m0->m_data += sc->params.sge.fl_pktshift;
1983 
1984 have_mbuf:
1985 	m0->m_pkthdr.rcvif = ifp;
1986 	M_HASHTYPE_SET(m0, sw_hashtype[d->rss.hash_type][d->rss.ipv6]);
1987 	m0->m_pkthdr.flowid = be32toh(d->rss.hash_val);
1988 
1989 	cpl = (const void *)(&d->rss + 1);
1990 	if (sc->params.tp.rx_pkt_encap) {
1991 		const uint16_t ev = be16toh(cpl->err_vec);
1992 
1993 		err_vec = G_T6_COMPR_RXERR_VEC(ev);
1994 		tnl_type = G_T6_RX_TNL_TYPE(ev);
1995 		tnlhdr_len = G_T6_RX_TNLHDR_LEN(ev);
1996 	} else {
1997 		err_vec = be16toh(cpl->err_vec);
1998 		tnl_type = 0;
1999 		tnlhdr_len = 0;
2000 	}
2001 	if (cpl->csum_calc && err_vec == 0) {
2002 		int ipv6 = !!(cpl->l2info & htobe32(F_RXF_IP6));
2003 
2004 		/* checksum(s) calculated and found to be correct. */
2005 
2006 		MPASS((cpl->l2info & htobe32(F_RXF_IP)) ^
2007 		    (cpl->l2info & htobe32(F_RXF_IP6)));
2008 		m0->m_pkthdr.csum_data = be16toh(cpl->csum);
2009 		if (tnl_type == 0) {
2010 	    		if (!ipv6 && ifp->if_capenable & IFCAP_RXCSUM) {
2011 				m0->m_pkthdr.csum_flags = CSUM_L3_CALC |
2012 				    CSUM_L3_VALID | CSUM_L4_CALC |
2013 				    CSUM_L4_VALID;
2014 			} else if (ipv6 && ifp->if_capenable & IFCAP_RXCSUM_IPV6) {
2015 				m0->m_pkthdr.csum_flags = CSUM_L4_CALC |
2016 				    CSUM_L4_VALID;
2017 			}
2018 			rxq->rxcsum++;
2019 		} else {
2020 			MPASS(tnl_type == RX_PKT_TNL_TYPE_VXLAN);
2021 
2022 			M_HASHTYPE_SETINNER(m0);
2023 			if (__predict_false(cpl->ip_frag)) {
2024 				/*
2025 				 * csum_data is for the inner frame (which is an
2026 				 * IP fragment) and is not 0xffff.  There is no
2027 				 * way to pass the inner csum_data to the stack.
2028 				 * We don't want the stack to use the inner
2029 				 * csum_data to validate the outer frame or it
2030 				 * will get rejected.  So we fix csum_data here
2031 				 * and let sw do the checksum of inner IP
2032 				 * fragments.
2033 				 *
2034 				 * XXX: Need 32b for csum_data2 in an rx mbuf.
2035 				 * Maybe stuff it into rcv_tstmp?
2036 				 */
2037 				m0->m_pkthdr.csum_data = 0xffff;
2038 				if (ipv6) {
2039 					m0->m_pkthdr.csum_flags = CSUM_L4_CALC |
2040 					    CSUM_L4_VALID;
2041 				} else {
2042 					m0->m_pkthdr.csum_flags = CSUM_L3_CALC |
2043 					    CSUM_L3_VALID | CSUM_L4_CALC |
2044 					    CSUM_L4_VALID;
2045 				}
2046 			} else {
2047 				int outer_ipv6;
2048 
2049 				MPASS(m0->m_pkthdr.csum_data == 0xffff);
2050 
2051 				outer_ipv6 = tnlhdr_len >=
2052 				    sizeof(struct ether_header) +
2053 				    sizeof(struct ip6_hdr);
2054 				m0->m_pkthdr.csum_flags =
2055 				    sw_csum_flags[outer_ipv6][ipv6];
2056 			}
2057 			rxq->vxlan_rxcsum++;
2058 		}
2059 	}
2060 
2061 	if (cpl->vlan_ex) {
2062 		m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
2063 		m0->m_flags |= M_VLANTAG;
2064 		rxq->vlan_extraction++;
2065 	}
2066 
2067 	if (rxq->iq.flags & IQ_RX_TIMESTAMP) {
2068 		/*
2069 		 * Fill up rcv_tstmp but do not set M_TSTMP.
2070 		 * rcv_tstmp is not in the format that the
2071 		 * kernel expects and we don't want to mislead
2072 		 * it.  For now this is only for custom code
2073 		 * that knows how to interpret cxgbe's stamp.
2074 		 */
2075 		m0->m_pkthdr.rcv_tstmp =
2076 		    last_flit_to_ns(sc, d->rsp.u.last_flit);
2077 #ifdef notyet
2078 		m0->m_flags |= M_TSTMP;
2079 #endif
2080 	}
2081 
2082 #ifdef NUMA
2083 	m0->m_pkthdr.numa_domain = ifp->if_numa_domain;
2084 #endif
2085 #if defined(INET) || defined(INET6)
2086 	if (rxq->iq.flags & IQ_LRO_ENABLED && tnl_type == 0 &&
2087 	    (M_HASHTYPE_GET(m0) == M_HASHTYPE_RSS_TCP_IPV4 ||
2088 	    M_HASHTYPE_GET(m0) == M_HASHTYPE_RSS_TCP_IPV6)) {
2089 		if (sort_before_lro(lro)) {
2090 			tcp_lro_queue_mbuf(lro, m0);
2091 			return (0); /* queued for sort, then LRO */
2092 		}
2093 		if (tcp_lro_rx(lro, m0, 0) == 0)
2094 			return (0); /* queued for LRO */
2095 	}
2096 #endif
2097 	ifp->if_input(ifp, m0);
2098 
2099 	return (0);
2100 }
2101 
2102 /*
2103  * Must drain the wrq or make sure that someone else will.
2104  */
2105 static void
2106 wrq_tx_drain(void *arg, int n)
2107 {
2108 	struct sge_wrq *wrq = arg;
2109 	struct sge_eq *eq = &wrq->eq;
2110 
2111 	EQ_LOCK(eq);
2112 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2113 		drain_wrq_wr_list(wrq->adapter, wrq);
2114 	EQ_UNLOCK(eq);
2115 }
2116 
2117 static void
2118 drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq)
2119 {
2120 	struct sge_eq *eq = &wrq->eq;
2121 	u_int available, dbdiff;	/* # of hardware descriptors */
2122 	u_int n;
2123 	struct wrqe *wr;
2124 	struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
2125 
2126 	EQ_LOCK_ASSERT_OWNED(eq);
2127 	MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
2128 	wr = STAILQ_FIRST(&wrq->wr_list);
2129 	MPASS(wr != NULL);	/* Must be called with something useful to do */
2130 	MPASS(eq->pidx == eq->dbidx);
2131 	dbdiff = 0;
2132 
2133 	do {
2134 		eq->cidx = read_hw_cidx(eq);
2135 		if (eq->pidx == eq->cidx)
2136 			available = eq->sidx - 1;
2137 		else
2138 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2139 
2140 		MPASS(wr->wrq == wrq);
2141 		n = howmany(wr->wr_len, EQ_ESIZE);
2142 		if (available < n)
2143 			break;
2144 
2145 		dst = (void *)&eq->desc[eq->pidx];
2146 		if (__predict_true(eq->sidx - eq->pidx > n)) {
2147 			/* Won't wrap, won't end exactly at the status page. */
2148 			bcopy(&wr->wr[0], dst, wr->wr_len);
2149 			eq->pidx += n;
2150 		} else {
2151 			int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE;
2152 
2153 			bcopy(&wr->wr[0], dst, first_portion);
2154 			if (wr->wr_len > first_portion) {
2155 				bcopy(&wr->wr[first_portion], &eq->desc[0],
2156 				    wr->wr_len - first_portion);
2157 			}
2158 			eq->pidx = n - (eq->sidx - eq->pidx);
2159 		}
2160 		wrq->tx_wrs_copied++;
2161 
2162 		if (available < eq->sidx / 4 &&
2163 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2164 				/*
2165 				 * XXX: This is not 100% reliable with some
2166 				 * types of WRs.  But this is a very unusual
2167 				 * situation for an ofld/ctrl queue anyway.
2168 				 */
2169 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2170 			    F_FW_WR_EQUEQ);
2171 		}
2172 
2173 		dbdiff += n;
2174 		if (dbdiff >= 16) {
2175 			ring_eq_db(sc, eq, dbdiff);
2176 			dbdiff = 0;
2177 		}
2178 
2179 		STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
2180 		free_wrqe(wr);
2181 		MPASS(wrq->nwr_pending > 0);
2182 		wrq->nwr_pending--;
2183 		MPASS(wrq->ndesc_needed >= n);
2184 		wrq->ndesc_needed -= n;
2185 	} while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL);
2186 
2187 	if (dbdiff)
2188 		ring_eq_db(sc, eq, dbdiff);
2189 }
2190 
2191 /*
2192  * Doesn't fail.  Holds on to work requests it can't send right away.
2193  */
2194 void
2195 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
2196 {
2197 #ifdef INVARIANTS
2198 	struct sge_eq *eq = &wrq->eq;
2199 #endif
2200 
2201 	EQ_LOCK_ASSERT_OWNED(eq);
2202 	MPASS(wr != NULL);
2203 	MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN);
2204 	MPASS((wr->wr_len & 0x7) == 0);
2205 
2206 	STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
2207 	wrq->nwr_pending++;
2208 	wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE);
2209 
2210 	if (!TAILQ_EMPTY(&wrq->incomplete_wrs))
2211 		return;	/* commit_wrq_wr will drain wr_list as well. */
2212 
2213 	drain_wrq_wr_list(sc, wrq);
2214 
2215 	/* Doorbell must have caught up to the pidx. */
2216 	MPASS(eq->pidx == eq->dbidx);
2217 }
2218 
2219 void
2220 t4_update_fl_bufsize(struct ifnet *ifp)
2221 {
2222 	struct vi_info *vi = ifp->if_softc;
2223 	struct adapter *sc = vi->adapter;
2224 	struct sge_rxq *rxq;
2225 #ifdef TCP_OFFLOAD
2226 	struct sge_ofld_rxq *ofld_rxq;
2227 #endif
2228 	struct sge_fl *fl;
2229 	int i, maxp;
2230 
2231 	maxp = max_rx_payload(sc, ifp, false);
2232 	for_each_rxq(vi, i, rxq) {
2233 		fl = &rxq->fl;
2234 
2235 		FL_LOCK(fl);
2236 		fl->zidx = find_refill_source(sc, maxp,
2237 		    fl->flags & FL_BUF_PACKING);
2238 		FL_UNLOCK(fl);
2239 	}
2240 #ifdef TCP_OFFLOAD
2241 	maxp = max_rx_payload(sc, ifp, true);
2242 	for_each_ofld_rxq(vi, i, ofld_rxq) {
2243 		fl = &ofld_rxq->fl;
2244 
2245 		FL_LOCK(fl);
2246 		fl->zidx = find_refill_source(sc, maxp,
2247 		    fl->flags & FL_BUF_PACKING);
2248 		FL_UNLOCK(fl);
2249 	}
2250 #endif
2251 }
2252 
2253 static inline int
2254 mbuf_nsegs(struct mbuf *m)
2255 {
2256 
2257 	M_ASSERTPKTHDR(m);
2258 	KASSERT(m->m_pkthdr.inner_l5hlen > 0,
2259 	    ("%s: mbuf %p missing information on # of segments.", __func__, m));
2260 
2261 	return (m->m_pkthdr.inner_l5hlen);
2262 }
2263 
2264 static inline void
2265 set_mbuf_nsegs(struct mbuf *m, uint8_t nsegs)
2266 {
2267 
2268 	M_ASSERTPKTHDR(m);
2269 	m->m_pkthdr.inner_l5hlen = nsegs;
2270 }
2271 
2272 static inline int
2273 mbuf_cflags(struct mbuf *m)
2274 {
2275 
2276 	M_ASSERTPKTHDR(m);
2277 	return (m->m_pkthdr.PH_loc.eight[4]);
2278 }
2279 
2280 static inline void
2281 set_mbuf_cflags(struct mbuf *m, uint8_t flags)
2282 {
2283 
2284 	M_ASSERTPKTHDR(m);
2285 	m->m_pkthdr.PH_loc.eight[4] = flags;
2286 }
2287 
2288 static inline int
2289 mbuf_len16(struct mbuf *m)
2290 {
2291 	int n;
2292 
2293 	M_ASSERTPKTHDR(m);
2294 	n = m->m_pkthdr.PH_loc.eight[0];
2295 	if (!(mbuf_cflags(m) & MC_TLS))
2296 		MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2297 
2298 	return (n);
2299 }
2300 
2301 static inline void
2302 set_mbuf_len16(struct mbuf *m, uint8_t len16)
2303 {
2304 
2305 	M_ASSERTPKTHDR(m);
2306 	if (!(mbuf_cflags(m) & MC_TLS))
2307 		MPASS(len16 > 0 && len16 <= SGE_MAX_WR_LEN / 16);
2308 	m->m_pkthdr.PH_loc.eight[0] = len16;
2309 }
2310 
2311 #ifdef RATELIMIT
2312 static inline int
2313 mbuf_eo_nsegs(struct mbuf *m)
2314 {
2315 
2316 	M_ASSERTPKTHDR(m);
2317 	return (m->m_pkthdr.PH_loc.eight[1]);
2318 }
2319 
2320 #if defined(INET) || defined(INET6)
2321 static inline void
2322 set_mbuf_eo_nsegs(struct mbuf *m, uint8_t nsegs)
2323 {
2324 
2325 	M_ASSERTPKTHDR(m);
2326 	m->m_pkthdr.PH_loc.eight[1] = nsegs;
2327 }
2328 #endif
2329 
2330 static inline int
2331 mbuf_eo_len16(struct mbuf *m)
2332 {
2333 	int n;
2334 
2335 	M_ASSERTPKTHDR(m);
2336 	n = m->m_pkthdr.PH_loc.eight[2];
2337 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2338 
2339 	return (n);
2340 }
2341 
2342 #if defined(INET) || defined(INET6)
2343 static inline void
2344 set_mbuf_eo_len16(struct mbuf *m, uint8_t len16)
2345 {
2346 
2347 	M_ASSERTPKTHDR(m);
2348 	m->m_pkthdr.PH_loc.eight[2] = len16;
2349 }
2350 #endif
2351 
2352 static inline int
2353 mbuf_eo_tsclk_tsoff(struct mbuf *m)
2354 {
2355 
2356 	M_ASSERTPKTHDR(m);
2357 	return (m->m_pkthdr.PH_loc.eight[3]);
2358 }
2359 
2360 #if defined(INET) || defined(INET6)
2361 static inline void
2362 set_mbuf_eo_tsclk_tsoff(struct mbuf *m, uint8_t tsclk_tsoff)
2363 {
2364 
2365 	M_ASSERTPKTHDR(m);
2366 	m->m_pkthdr.PH_loc.eight[3] = tsclk_tsoff;
2367 }
2368 #endif
2369 
2370 static inline int
2371 needs_eo(struct m_snd_tag *mst)
2372 {
2373 
2374 	return (mst != NULL && mst->sw->type == IF_SND_TAG_TYPE_RATE_LIMIT);
2375 }
2376 #endif
2377 
2378 /*
2379  * Try to allocate an mbuf to contain a raw work request.  To make it
2380  * easy to construct the work request, don't allocate a chain but a
2381  * single mbuf.
2382  */
2383 struct mbuf *
2384 alloc_wr_mbuf(int len, int how)
2385 {
2386 	struct mbuf *m;
2387 
2388 	if (len <= MHLEN)
2389 		m = m_gethdr(how, MT_DATA);
2390 	else if (len <= MCLBYTES)
2391 		m = m_getcl(how, MT_DATA, M_PKTHDR);
2392 	else
2393 		m = NULL;
2394 	if (m == NULL)
2395 		return (NULL);
2396 	m->m_pkthdr.len = len;
2397 	m->m_len = len;
2398 	set_mbuf_cflags(m, MC_RAW_WR);
2399 	set_mbuf_len16(m, howmany(len, 16));
2400 	return (m);
2401 }
2402 
2403 static inline bool
2404 needs_hwcsum(struct mbuf *m)
2405 {
2406 	const uint32_t csum_flags = CSUM_IP | CSUM_IP_UDP | CSUM_IP_TCP |
2407 	    CSUM_IP_TSO | CSUM_INNER_IP | CSUM_INNER_IP_UDP |
2408 	    CSUM_INNER_IP_TCP | CSUM_INNER_IP_TSO | CSUM_IP6_UDP |
2409 	    CSUM_IP6_TCP | CSUM_IP6_TSO | CSUM_INNER_IP6_UDP |
2410 	    CSUM_INNER_IP6_TCP | CSUM_INNER_IP6_TSO;
2411 
2412 	M_ASSERTPKTHDR(m);
2413 
2414 	return (m->m_pkthdr.csum_flags & csum_flags);
2415 }
2416 
2417 static inline bool
2418 needs_tso(struct mbuf *m)
2419 {
2420 	const uint32_t csum_flags = CSUM_IP_TSO | CSUM_IP6_TSO |
2421 	    CSUM_INNER_IP_TSO | CSUM_INNER_IP6_TSO;
2422 
2423 	M_ASSERTPKTHDR(m);
2424 
2425 	return (m->m_pkthdr.csum_flags & csum_flags);
2426 }
2427 
2428 static inline bool
2429 needs_vxlan_csum(struct mbuf *m)
2430 {
2431 
2432 	M_ASSERTPKTHDR(m);
2433 
2434 	return (m->m_pkthdr.csum_flags & CSUM_ENCAP_VXLAN);
2435 }
2436 
2437 static inline bool
2438 needs_vxlan_tso(struct mbuf *m)
2439 {
2440 	const uint32_t csum_flags = CSUM_ENCAP_VXLAN | CSUM_INNER_IP_TSO |
2441 	    CSUM_INNER_IP6_TSO;
2442 
2443 	M_ASSERTPKTHDR(m);
2444 
2445 	return ((m->m_pkthdr.csum_flags & csum_flags) != 0 &&
2446 	    (m->m_pkthdr.csum_flags & csum_flags) != CSUM_ENCAP_VXLAN);
2447 }
2448 
2449 #if defined(INET) || defined(INET6)
2450 static inline bool
2451 needs_inner_tcp_csum(struct mbuf *m)
2452 {
2453 	const uint32_t csum_flags = CSUM_INNER_IP_TSO | CSUM_INNER_IP6_TSO;
2454 
2455 	M_ASSERTPKTHDR(m);
2456 
2457 	return (m->m_pkthdr.csum_flags & csum_flags);
2458 }
2459 #endif
2460 
2461 static inline bool
2462 needs_l3_csum(struct mbuf *m)
2463 {
2464 	const uint32_t csum_flags = CSUM_IP | CSUM_IP_TSO | CSUM_INNER_IP |
2465 	    CSUM_INNER_IP_TSO;
2466 
2467 	M_ASSERTPKTHDR(m);
2468 
2469 	return (m->m_pkthdr.csum_flags & csum_flags);
2470 }
2471 
2472 static inline bool
2473 needs_outer_tcp_csum(struct mbuf *m)
2474 {
2475 	const uint32_t csum_flags = CSUM_IP_TCP | CSUM_IP_TSO | CSUM_IP6_TCP |
2476 	    CSUM_IP6_TSO;
2477 
2478 	M_ASSERTPKTHDR(m);
2479 
2480 	return (m->m_pkthdr.csum_flags & csum_flags);
2481 }
2482 
2483 #ifdef RATELIMIT
2484 static inline bool
2485 needs_outer_l4_csum(struct mbuf *m)
2486 {
2487 	const uint32_t csum_flags = CSUM_IP_UDP | CSUM_IP_TCP | CSUM_IP_TSO |
2488 	    CSUM_IP6_UDP | CSUM_IP6_TCP | CSUM_IP6_TSO;
2489 
2490 	M_ASSERTPKTHDR(m);
2491 
2492 	return (m->m_pkthdr.csum_flags & csum_flags);
2493 }
2494 
2495 static inline bool
2496 needs_outer_udp_csum(struct mbuf *m)
2497 {
2498 	const uint32_t csum_flags = CSUM_IP_UDP | CSUM_IP6_UDP;
2499 
2500 	M_ASSERTPKTHDR(m);
2501 
2502 	return (m->m_pkthdr.csum_flags & csum_flags);
2503 }
2504 #endif
2505 
2506 static inline bool
2507 needs_vlan_insertion(struct mbuf *m)
2508 {
2509 
2510 	M_ASSERTPKTHDR(m);
2511 
2512 	return (m->m_flags & M_VLANTAG);
2513 }
2514 
2515 static void *
2516 m_advance(struct mbuf **pm, int *poffset, int len)
2517 {
2518 	struct mbuf *m = *pm;
2519 	int offset = *poffset;
2520 	uintptr_t p = 0;
2521 
2522 	MPASS(len > 0);
2523 
2524 	for (;;) {
2525 		if (offset + len < m->m_len) {
2526 			offset += len;
2527 			p = mtod(m, uintptr_t) + offset;
2528 			break;
2529 		}
2530 		len -= m->m_len - offset;
2531 		m = m->m_next;
2532 		offset = 0;
2533 		MPASS(m != NULL);
2534 	}
2535 	*poffset = offset;
2536 	*pm = m;
2537 	return ((void *)p);
2538 }
2539 
2540 static inline int
2541 count_mbuf_ext_pgs(struct mbuf *m, int skip, vm_paddr_t *nextaddr)
2542 {
2543 	vm_paddr_t paddr;
2544 	int i, len, off, pglen, pgoff, seglen, segoff;
2545 	int nsegs = 0;
2546 
2547 	M_ASSERTEXTPG(m);
2548 	off = mtod(m, vm_offset_t);
2549 	len = m->m_len;
2550 	off += skip;
2551 	len -= skip;
2552 
2553 	if (m->m_epg_hdrlen != 0) {
2554 		if (off >= m->m_epg_hdrlen) {
2555 			off -= m->m_epg_hdrlen;
2556 		} else {
2557 			seglen = m->m_epg_hdrlen - off;
2558 			segoff = off;
2559 			seglen = min(seglen, len);
2560 			off = 0;
2561 			len -= seglen;
2562 			paddr = pmap_kextract(
2563 			    (vm_offset_t)&m->m_epg_hdr[segoff]);
2564 			if (*nextaddr != paddr)
2565 				nsegs++;
2566 			*nextaddr = paddr + seglen;
2567 		}
2568 	}
2569 	pgoff = m->m_epg_1st_off;
2570 	for (i = 0; i < m->m_epg_npgs && len > 0; i++) {
2571 		pglen = m_epg_pagelen(m, i, pgoff);
2572 		if (off >= pglen) {
2573 			off -= pglen;
2574 			pgoff = 0;
2575 			continue;
2576 		}
2577 		seglen = pglen - off;
2578 		segoff = pgoff + off;
2579 		off = 0;
2580 		seglen = min(seglen, len);
2581 		len -= seglen;
2582 		paddr = m->m_epg_pa[i] + segoff;
2583 		if (*nextaddr != paddr)
2584 			nsegs++;
2585 		*nextaddr = paddr + seglen;
2586 		pgoff = 0;
2587 	};
2588 	if (len != 0) {
2589 		seglen = min(len, m->m_epg_trllen - off);
2590 		len -= seglen;
2591 		paddr = pmap_kextract((vm_offset_t)&m->m_epg_trail[off]);
2592 		if (*nextaddr != paddr)
2593 			nsegs++;
2594 		*nextaddr = paddr + seglen;
2595 	}
2596 
2597 	return (nsegs);
2598 }
2599 
2600 
2601 /*
2602  * Can deal with empty mbufs in the chain that have m_len = 0, but the chain
2603  * must have at least one mbuf that's not empty.  It is possible for this
2604  * routine to return 0 if skip accounts for all the contents of the mbuf chain.
2605  */
2606 static inline int
2607 count_mbuf_nsegs(struct mbuf *m, int skip, uint8_t *cflags)
2608 {
2609 	vm_paddr_t nextaddr, paddr;
2610 	vm_offset_t va;
2611 	int len, nsegs;
2612 
2613 	M_ASSERTPKTHDR(m);
2614 	MPASS(m->m_pkthdr.len > 0);
2615 	MPASS(m->m_pkthdr.len >= skip);
2616 
2617 	nsegs = 0;
2618 	nextaddr = 0;
2619 	for (; m; m = m->m_next) {
2620 		len = m->m_len;
2621 		if (__predict_false(len == 0))
2622 			continue;
2623 		if (skip >= len) {
2624 			skip -= len;
2625 			continue;
2626 		}
2627 		if ((m->m_flags & M_EXTPG) != 0) {
2628 			*cflags |= MC_NOMAP;
2629 			nsegs += count_mbuf_ext_pgs(m, skip, &nextaddr);
2630 			skip = 0;
2631 			continue;
2632 		}
2633 		va = mtod(m, vm_offset_t) + skip;
2634 		len -= skip;
2635 		skip = 0;
2636 		paddr = pmap_kextract(va);
2637 		nsegs += sglist_count((void *)(uintptr_t)va, len);
2638 		if (paddr == nextaddr)
2639 			nsegs--;
2640 		nextaddr = pmap_kextract(va + len - 1) + 1;
2641 	}
2642 
2643 	return (nsegs);
2644 }
2645 
2646 /*
2647  * The maximum number of segments that can fit in a WR.
2648  */
2649 static int
2650 max_nsegs_allowed(struct mbuf *m, bool vm_wr)
2651 {
2652 
2653 	if (vm_wr) {
2654 		if (needs_tso(m))
2655 			return (TX_SGL_SEGS_VM_TSO);
2656 		return (TX_SGL_SEGS_VM);
2657 	}
2658 
2659 	if (needs_tso(m)) {
2660 		if (needs_vxlan_tso(m))
2661 			return (TX_SGL_SEGS_VXLAN_TSO);
2662 		else
2663 			return (TX_SGL_SEGS_TSO);
2664 	}
2665 
2666 	return (TX_SGL_SEGS);
2667 }
2668 
2669 static struct timeval txerr_ratecheck = {0};
2670 static const struct timeval txerr_interval = {3, 0};
2671 
2672 /*
2673  * Analyze the mbuf to determine its tx needs.  The mbuf passed in may change:
2674  * a) caller can assume it's been freed if this function returns with an error.
2675  * b) it may get defragged up if the gather list is too long for the hardware.
2676  */
2677 int
2678 parse_pkt(struct mbuf **mp, bool vm_wr)
2679 {
2680 	struct mbuf *m0 = *mp, *m;
2681 	int rc, nsegs, defragged = 0, offset;
2682 	struct ether_header *eh;
2683 	void *l3hdr;
2684 #if defined(INET) || defined(INET6)
2685 	struct tcphdr *tcp;
2686 #endif
2687 #if defined(KERN_TLS) || defined(RATELIMIT)
2688 	struct m_snd_tag *mst;
2689 #endif
2690 	uint16_t eh_type;
2691 	uint8_t cflags;
2692 
2693 	cflags = 0;
2694 	M_ASSERTPKTHDR(m0);
2695 	if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) {
2696 		rc = EINVAL;
2697 fail:
2698 		m_freem(m0);
2699 		*mp = NULL;
2700 		return (rc);
2701 	}
2702 restart:
2703 	/*
2704 	 * First count the number of gather list segments in the payload.
2705 	 * Defrag the mbuf if nsegs exceeds the hardware limit.
2706 	 */
2707 	M_ASSERTPKTHDR(m0);
2708 	MPASS(m0->m_pkthdr.len > 0);
2709 	nsegs = count_mbuf_nsegs(m0, 0, &cflags);
2710 #if defined(KERN_TLS) || defined(RATELIMIT)
2711 	if (m0->m_pkthdr.csum_flags & CSUM_SND_TAG)
2712 		mst = m0->m_pkthdr.snd_tag;
2713 	else
2714 		mst = NULL;
2715 #endif
2716 #ifdef KERN_TLS
2717 	if (mst != NULL && mst->sw->type == IF_SND_TAG_TYPE_TLS) {
2718 		int len16;
2719 
2720 		cflags |= MC_TLS;
2721 		set_mbuf_cflags(m0, cflags);
2722 		rc = t6_ktls_parse_pkt(m0, &nsegs, &len16);
2723 		if (rc != 0)
2724 			goto fail;
2725 		set_mbuf_nsegs(m0, nsegs);
2726 		set_mbuf_len16(m0, len16);
2727 		return (0);
2728 	}
2729 #endif
2730 	if (nsegs > max_nsegs_allowed(m0, vm_wr)) {
2731 		if (defragged++ > 0) {
2732 			rc = EFBIG;
2733 			goto fail;
2734 		}
2735 		counter_u64_add(defrags, 1);
2736 		if ((m = m_defrag(m0, M_NOWAIT)) == NULL) {
2737 			rc = ENOMEM;
2738 			goto fail;
2739 		}
2740 		*mp = m0 = m;	/* update caller's copy after defrag */
2741 		goto restart;
2742 	}
2743 
2744 	if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN &&
2745 	    !(cflags & MC_NOMAP))) {
2746 		counter_u64_add(pullups, 1);
2747 		m0 = m_pullup(m0, m0->m_pkthdr.len);
2748 		if (m0 == NULL) {
2749 			/* Should have left well enough alone. */
2750 			rc = EFBIG;
2751 			goto fail;
2752 		}
2753 		*mp = m0;	/* update caller's copy after pullup */
2754 		goto restart;
2755 	}
2756 	set_mbuf_nsegs(m0, nsegs);
2757 	set_mbuf_cflags(m0, cflags);
2758 	calculate_mbuf_len16(m0, vm_wr);
2759 
2760 #ifdef RATELIMIT
2761 	/*
2762 	 * Ethofld is limited to TCP and UDP for now, and only when L4 hw
2763 	 * checksumming is enabled.  needs_outer_l4_csum happens to check for
2764 	 * all the right things.
2765 	 */
2766 	if (__predict_false(needs_eo(mst) && !needs_outer_l4_csum(m0))) {
2767 		m_snd_tag_rele(m0->m_pkthdr.snd_tag);
2768 		m0->m_pkthdr.snd_tag = NULL;
2769 		m0->m_pkthdr.csum_flags &= ~CSUM_SND_TAG;
2770 		mst = NULL;
2771 	}
2772 #endif
2773 
2774 	if (!needs_hwcsum(m0)
2775 #ifdef RATELIMIT
2776    		 && !needs_eo(mst)
2777 #endif
2778 	)
2779 		return (0);
2780 
2781 	m = m0;
2782 	eh = mtod(m, struct ether_header *);
2783 	eh_type = ntohs(eh->ether_type);
2784 	if (eh_type == ETHERTYPE_VLAN) {
2785 		struct ether_vlan_header *evh = (void *)eh;
2786 
2787 		eh_type = ntohs(evh->evl_proto);
2788 		m0->m_pkthdr.l2hlen = sizeof(*evh);
2789 	} else
2790 		m0->m_pkthdr.l2hlen = sizeof(*eh);
2791 
2792 	offset = 0;
2793 	l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2794 
2795 	switch (eh_type) {
2796 #ifdef INET6
2797 	case ETHERTYPE_IPV6:
2798 		m0->m_pkthdr.l3hlen = sizeof(struct ip6_hdr);
2799 		break;
2800 #endif
2801 #ifdef INET
2802 	case ETHERTYPE_IP:
2803 	{
2804 		struct ip *ip = l3hdr;
2805 
2806 		if (needs_vxlan_csum(m0)) {
2807 			/* Driver will do the outer IP hdr checksum. */
2808 			ip->ip_sum = 0;
2809 			if (needs_vxlan_tso(m0)) {
2810 				const uint16_t ipl = ip->ip_len;
2811 
2812 				ip->ip_len = 0;
2813 				ip->ip_sum = ~in_cksum_hdr(ip);
2814 				ip->ip_len = ipl;
2815 			} else
2816 				ip->ip_sum = in_cksum_hdr(ip);
2817 		}
2818 		m0->m_pkthdr.l3hlen = ip->ip_hl << 2;
2819 		break;
2820 	}
2821 #endif
2822 	default:
2823 		if (ratecheck(&txerr_ratecheck, &txerr_interval)) {
2824 			log(LOG_ERR, "%s: ethertype 0x%04x unknown.  "
2825 			    "if_cxgbe must be compiled with the same "
2826 			    "INET/INET6 options as the kernel.\n", __func__,
2827 			    eh_type);
2828 		}
2829 		rc = EINVAL;
2830 		goto fail;
2831 	}
2832 
2833 	if (needs_vxlan_csum(m0)) {
2834 		m0->m_pkthdr.l4hlen = sizeof(struct udphdr);
2835 		m0->m_pkthdr.l5hlen = sizeof(struct vxlan_header);
2836 
2837 		/* Inner headers. */
2838 		eh = m_advance(&m, &offset, m0->m_pkthdr.l3hlen +
2839 		    sizeof(struct udphdr) + sizeof(struct vxlan_header));
2840 		eh_type = ntohs(eh->ether_type);
2841 		if (eh_type == ETHERTYPE_VLAN) {
2842 			struct ether_vlan_header *evh = (void *)eh;
2843 
2844 			eh_type = ntohs(evh->evl_proto);
2845 			m0->m_pkthdr.inner_l2hlen = sizeof(*evh);
2846 		} else
2847 			m0->m_pkthdr.inner_l2hlen = sizeof(*eh);
2848 		l3hdr = m_advance(&m, &offset, m0->m_pkthdr.inner_l2hlen);
2849 
2850 		switch (eh_type) {
2851 #ifdef INET6
2852 		case ETHERTYPE_IPV6:
2853 			m0->m_pkthdr.inner_l3hlen = sizeof(struct ip6_hdr);
2854 			break;
2855 #endif
2856 #ifdef INET
2857 		case ETHERTYPE_IP:
2858 		{
2859 			struct ip *ip = l3hdr;
2860 
2861 			m0->m_pkthdr.inner_l3hlen = ip->ip_hl << 2;
2862 			break;
2863 		}
2864 #endif
2865 		default:
2866 			if (ratecheck(&txerr_ratecheck, &txerr_interval)) {
2867 				log(LOG_ERR, "%s: VXLAN hw offload requested"
2868 				    "with unknown ethertype 0x%04x.  if_cxgbe "
2869 				    "must be compiled with the same INET/INET6 "
2870 				    "options as the kernel.\n", __func__,
2871 				    eh_type);
2872 			}
2873 			rc = EINVAL;
2874 			goto fail;
2875 		}
2876 #if defined(INET) || defined(INET6)
2877 		if (needs_inner_tcp_csum(m0)) {
2878 			tcp = m_advance(&m, &offset, m0->m_pkthdr.inner_l3hlen);
2879 			m0->m_pkthdr.inner_l4hlen = tcp->th_off * 4;
2880 		}
2881 #endif
2882 		MPASS((m0->m_pkthdr.csum_flags & CSUM_SND_TAG) == 0);
2883 		m0->m_pkthdr.csum_flags &= CSUM_INNER_IP6_UDP |
2884 		    CSUM_INNER_IP6_TCP | CSUM_INNER_IP6_TSO | CSUM_INNER_IP |
2885 		    CSUM_INNER_IP_UDP | CSUM_INNER_IP_TCP | CSUM_INNER_IP_TSO |
2886 		    CSUM_ENCAP_VXLAN;
2887 	}
2888 
2889 #if defined(INET) || defined(INET6)
2890 	if (needs_outer_tcp_csum(m0)) {
2891 		tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen);
2892 		m0->m_pkthdr.l4hlen = tcp->th_off * 4;
2893 #ifdef RATELIMIT
2894 		if (tsclk >= 0 && *(uint32_t *)(tcp + 1) == ntohl(0x0101080a)) {
2895 			set_mbuf_eo_tsclk_tsoff(m0,
2896 			    V_FW_ETH_TX_EO_WR_TSCLK(tsclk) |
2897 			    V_FW_ETH_TX_EO_WR_TSOFF(sizeof(*tcp) / 2 + 1));
2898 		} else
2899 			set_mbuf_eo_tsclk_tsoff(m0, 0);
2900 	} else if (needs_outer_udp_csum(m0)) {
2901 		m0->m_pkthdr.l4hlen = sizeof(struct udphdr);
2902 #endif
2903 	}
2904 #ifdef RATELIMIT
2905 	if (needs_eo(mst)) {
2906 		u_int immhdrs;
2907 
2908 		/* EO WRs have the headers in the WR and not the GL. */
2909 		immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen +
2910 		    m0->m_pkthdr.l4hlen;
2911 		cflags = 0;
2912 		nsegs = count_mbuf_nsegs(m0, immhdrs, &cflags);
2913 		MPASS(cflags == mbuf_cflags(m0));
2914 		set_mbuf_eo_nsegs(m0, nsegs);
2915 		set_mbuf_eo_len16(m0,
2916 		    txpkt_eo_len16(nsegs, immhdrs, needs_tso(m0)));
2917 	}
2918 #endif
2919 #endif
2920 	MPASS(m0 == *mp);
2921 	return (0);
2922 }
2923 
2924 void *
2925 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie)
2926 {
2927 	struct sge_eq *eq = &wrq->eq;
2928 	struct adapter *sc = wrq->adapter;
2929 	int ndesc, available;
2930 	struct wrqe *wr;
2931 	void *w;
2932 
2933 	MPASS(len16 > 0);
2934 	ndesc = tx_len16_to_desc(len16);
2935 	MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC);
2936 
2937 	EQ_LOCK(eq);
2938 
2939 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2940 		drain_wrq_wr_list(sc, wrq);
2941 
2942 	if (!STAILQ_EMPTY(&wrq->wr_list)) {
2943 slowpath:
2944 		EQ_UNLOCK(eq);
2945 		wr = alloc_wrqe(len16 * 16, wrq);
2946 		if (__predict_false(wr == NULL))
2947 			return (NULL);
2948 		cookie->pidx = -1;
2949 		cookie->ndesc = ndesc;
2950 		return (&wr->wr);
2951 	}
2952 
2953 	eq->cidx = read_hw_cidx(eq);
2954 	if (eq->pidx == eq->cidx)
2955 		available = eq->sidx - 1;
2956 	else
2957 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2958 	if (available < ndesc)
2959 		goto slowpath;
2960 
2961 	cookie->pidx = eq->pidx;
2962 	cookie->ndesc = ndesc;
2963 	TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link);
2964 
2965 	w = &eq->desc[eq->pidx];
2966 	IDXINCR(eq->pidx, ndesc, eq->sidx);
2967 	if (__predict_false(cookie->pidx + ndesc > eq->sidx)) {
2968 		w = &wrq->ss[0];
2969 		wrq->ss_pidx = cookie->pidx;
2970 		wrq->ss_len = len16 * 16;
2971 	}
2972 
2973 	EQ_UNLOCK(eq);
2974 
2975 	return (w);
2976 }
2977 
2978 void
2979 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie)
2980 {
2981 	struct sge_eq *eq = &wrq->eq;
2982 	struct adapter *sc = wrq->adapter;
2983 	int ndesc, pidx;
2984 	struct wrq_cookie *prev, *next;
2985 
2986 	if (cookie->pidx == -1) {
2987 		struct wrqe *wr = __containerof(w, struct wrqe, wr);
2988 
2989 		t4_wrq_tx(sc, wr);
2990 		return;
2991 	}
2992 
2993 	if (__predict_false(w == &wrq->ss[0])) {
2994 		int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE;
2995 
2996 		MPASS(wrq->ss_len > n);	/* WR had better wrap around. */
2997 		bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n);
2998 		bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n);
2999 		wrq->tx_wrs_ss++;
3000 	} else
3001 		wrq->tx_wrs_direct++;
3002 
3003 	EQ_LOCK(eq);
3004 	ndesc = cookie->ndesc;	/* Can be more than SGE_MAX_WR_NDESC here. */
3005 	pidx = cookie->pidx;
3006 	MPASS(pidx >= 0 && pidx < eq->sidx);
3007 	prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link);
3008 	next = TAILQ_NEXT(cookie, link);
3009 	if (prev == NULL) {
3010 		MPASS(pidx == eq->dbidx);
3011 		if (next == NULL || ndesc >= 16) {
3012 			int available;
3013 			struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
3014 
3015 			/*
3016 			 * Note that the WR via which we'll request tx updates
3017 			 * is at pidx and not eq->pidx, which has moved on
3018 			 * already.
3019 			 */
3020 			dst = (void *)&eq->desc[pidx];
3021 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
3022 			if (available < eq->sidx / 4 &&
3023 			    atomic_cmpset_int(&eq->equiq, 0, 1)) {
3024 				/*
3025 				 * XXX: This is not 100% reliable with some
3026 				 * types of WRs.  But this is a very unusual
3027 				 * situation for an ofld/ctrl queue anyway.
3028 				 */
3029 				dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
3030 				    F_FW_WR_EQUEQ);
3031 			}
3032 
3033 			ring_eq_db(wrq->adapter, eq, ndesc);
3034 		} else {
3035 			MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc);
3036 			next->pidx = pidx;
3037 			next->ndesc += ndesc;
3038 		}
3039 	} else {
3040 		MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc);
3041 		prev->ndesc += ndesc;
3042 	}
3043 	TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link);
3044 
3045 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
3046 		drain_wrq_wr_list(sc, wrq);
3047 
3048 #ifdef INVARIANTS
3049 	if (TAILQ_EMPTY(&wrq->incomplete_wrs)) {
3050 		/* Doorbell must have caught up to the pidx. */
3051 		MPASS(wrq->eq.pidx == wrq->eq.dbidx);
3052 	}
3053 #endif
3054 	EQ_UNLOCK(eq);
3055 }
3056 
3057 static u_int
3058 can_resume_eth_tx(struct mp_ring *r)
3059 {
3060 	struct sge_eq *eq = r->cookie;
3061 
3062 	return (total_available_tx_desc(eq) > eq->sidx / 8);
3063 }
3064 
3065 static inline bool
3066 cannot_use_txpkts(struct mbuf *m)
3067 {
3068 	/* maybe put a GL limit too, to avoid silliness? */
3069 
3070 	return (needs_tso(m) || (mbuf_cflags(m) & (MC_RAW_WR | MC_TLS)) != 0);
3071 }
3072 
3073 static inline int
3074 discard_tx(struct sge_eq *eq)
3075 {
3076 
3077 	return ((eq->flags & (EQ_ENABLED | EQ_QFLUSH)) != EQ_ENABLED);
3078 }
3079 
3080 static inline int
3081 wr_can_update_eq(void *p)
3082 {
3083 	struct fw_eth_tx_pkts_wr *wr = p;
3084 
3085 	switch (G_FW_WR_OP(be32toh(wr->op_pkd))) {
3086 	case FW_ULPTX_WR:
3087 	case FW_ETH_TX_PKT_WR:
3088 	case FW_ETH_TX_PKTS_WR:
3089 	case FW_ETH_TX_PKTS2_WR:
3090 	case FW_ETH_TX_PKT_VM_WR:
3091 	case FW_ETH_TX_PKTS_VM_WR:
3092 		return (1);
3093 	default:
3094 		return (0);
3095 	}
3096 }
3097 
3098 static inline void
3099 set_txupdate_flags(struct sge_txq *txq, u_int avail,
3100     struct fw_eth_tx_pkt_wr *wr)
3101 {
3102 	struct sge_eq *eq = &txq->eq;
3103 	struct txpkts *txp = &txq->txp;
3104 
3105 	if ((txp->npkt > 0 || avail < eq->sidx / 2) &&
3106 	    atomic_cmpset_int(&eq->equiq, 0, 1)) {
3107 		wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ | F_FW_WR_EQUIQ);
3108 		eq->equeqidx = eq->pidx;
3109 	} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
3110 		wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
3111 		eq->equeqidx = eq->pidx;
3112 	}
3113 }
3114 
3115 #if defined(__i386__) || defined(__amd64__)
3116 extern uint64_t tsc_freq;
3117 #endif
3118 
3119 static inline bool
3120 record_eth_tx_time(struct sge_txq *txq)
3121 {
3122 	const uint64_t cycles = get_cyclecount();
3123 	const uint64_t last_tx = txq->last_tx;
3124 #if defined(__i386__) || defined(__amd64__)
3125 	const uint64_t itg = tsc_freq * t4_tx_coalesce_gap / 1000000;
3126 #else
3127 	const uint64_t itg = 0;
3128 #endif
3129 
3130 	MPASS(cycles >= last_tx);
3131 	txq->last_tx = cycles;
3132 	return (cycles - last_tx < itg);
3133 }
3134 
3135 /*
3136  * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to
3137  * be consumed.  Return the actual number consumed.  0 indicates a stall.
3138  */
3139 static u_int
3140 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx, bool *coalescing)
3141 {
3142 	struct sge_txq *txq = r->cookie;
3143 	struct ifnet *ifp = txq->ifp;
3144 	struct sge_eq *eq = &txq->eq;
3145 	struct txpkts *txp = &txq->txp;
3146 	struct vi_info *vi = ifp->if_softc;
3147 	struct adapter *sc = vi->adapter;
3148 	u_int total, remaining;		/* # of packets */
3149 	u_int n, avail, dbdiff;		/* # of hardware descriptors */
3150 	int i, rc;
3151 	struct mbuf *m0;
3152 	bool snd, recent_tx;
3153 	void *wr;	/* start of the last WR written to the ring */
3154 
3155 	TXQ_LOCK_ASSERT_OWNED(txq);
3156 	recent_tx = record_eth_tx_time(txq);
3157 
3158 	remaining = IDXDIFF(pidx, cidx, r->size);
3159 	if (__predict_false(discard_tx(eq))) {
3160 		for (i = 0; i < txp->npkt; i++)
3161 			m_freem(txp->mb[i]);
3162 		txp->npkt = 0;
3163 		while (cidx != pidx) {
3164 			m0 = r->items[cidx];
3165 			m_freem(m0);
3166 			if (++cidx == r->size)
3167 				cidx = 0;
3168 		}
3169 		reclaim_tx_descs(txq, eq->sidx);
3170 		*coalescing = false;
3171 		return (remaining);	/* emptied */
3172 	}
3173 
3174 	/* How many hardware descriptors do we have readily available. */
3175 	if (eq->pidx == eq->cidx)
3176 		avail = eq->sidx - 1;
3177 	else
3178 		avail = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
3179 
3180 	total = 0;
3181 	if (remaining == 0) {
3182 		txp->score = 0;
3183 		txq->txpkts_flush++;
3184 		goto send_txpkts;
3185 	}
3186 
3187 	dbdiff = 0;
3188 	MPASS(remaining > 0);
3189 	while (remaining > 0) {
3190 		m0 = r->items[cidx];
3191 		M_ASSERTPKTHDR(m0);
3192 		MPASS(m0->m_nextpkt == NULL);
3193 
3194 		if (avail < 2 * SGE_MAX_WR_NDESC)
3195 			avail += reclaim_tx_descs(txq, 64);
3196 
3197 		if (t4_tx_coalesce == 0 && txp->npkt == 0)
3198 			goto skip_coalescing;
3199 		if (cannot_use_txpkts(m0))
3200 			txp->score = 0;
3201 		else if (recent_tx) {
3202 			if (++txp->score == 0)
3203 				txp->score = UINT8_MAX;
3204 		} else
3205 			txp->score = 1;
3206 		if (txp->npkt > 0 || remaining > 1 ||
3207 		    txp->score >= t4_tx_coalesce_pkts ||
3208 		    atomic_load_int(&txq->eq.equiq) != 0) {
3209 			if (vi->flags & TX_USES_VM_WR)
3210 				rc = add_to_txpkts_vf(sc, txq, m0, avail, &snd);
3211 			else
3212 				rc = add_to_txpkts_pf(sc, txq, m0, avail, &snd);
3213 		} else {
3214 			snd = false;
3215 			rc = EINVAL;
3216 		}
3217 		if (snd) {
3218 			MPASS(txp->npkt > 0);
3219 			for (i = 0; i < txp->npkt; i++)
3220 				ETHER_BPF_MTAP(ifp, txp->mb[i]);
3221 			if (txp->npkt > 1) {
3222 				MPASS(avail >= tx_len16_to_desc(txp->len16));
3223 				if (vi->flags & TX_USES_VM_WR)
3224 					n = write_txpkts_vm_wr(sc, txq);
3225 				else
3226 					n = write_txpkts_wr(sc, txq);
3227 			} else {
3228 				MPASS(avail >=
3229 				    tx_len16_to_desc(mbuf_len16(txp->mb[0])));
3230 				if (vi->flags & TX_USES_VM_WR)
3231 					n = write_txpkt_vm_wr(sc, txq,
3232 					    txp->mb[0]);
3233 				else
3234 					n = write_txpkt_wr(sc, txq, txp->mb[0],
3235 					    avail);
3236 			}
3237 			MPASS(n <= SGE_MAX_WR_NDESC);
3238 			avail -= n;
3239 			dbdiff += n;
3240 			wr = &eq->desc[eq->pidx];
3241 			IDXINCR(eq->pidx, n, eq->sidx);
3242 			txp->npkt = 0;	/* emptied */
3243 		}
3244 		if (rc == 0) {
3245 			/* m0 was coalesced into txq->txpkts. */
3246 			goto next_mbuf;
3247 		}
3248 		if (rc == EAGAIN) {
3249 			/*
3250 			 * m0 is suitable for tx coalescing but could not be
3251 			 * combined with the existing txq->txpkts, which has now
3252 			 * been transmitted.  Start a new txpkts with m0.
3253 			 */
3254 			MPASS(snd);
3255 			MPASS(txp->npkt == 0);
3256 			continue;
3257 		}
3258 
3259 		MPASS(rc != 0 && rc != EAGAIN);
3260 		MPASS(txp->npkt == 0);
3261 skip_coalescing:
3262 		n = tx_len16_to_desc(mbuf_len16(m0));
3263 		if (__predict_false(avail < n)) {
3264 			avail += reclaim_tx_descs(txq, min(n, 32));
3265 			if (avail < n)
3266 				break;	/* out of descriptors */
3267 		}
3268 
3269 		wr = &eq->desc[eq->pidx];
3270 		if (mbuf_cflags(m0) & MC_RAW_WR) {
3271 			n = write_raw_wr(txq, wr, m0, avail);
3272 #ifdef KERN_TLS
3273 		} else if (mbuf_cflags(m0) & MC_TLS) {
3274 			ETHER_BPF_MTAP(ifp, m0);
3275 			n = t6_ktls_write_wr(txq, wr, m0, mbuf_nsegs(m0),
3276 			    avail);
3277 #endif
3278 		} else {
3279 			ETHER_BPF_MTAP(ifp, m0);
3280 			if (vi->flags & TX_USES_VM_WR)
3281 				n = write_txpkt_vm_wr(sc, txq, m0);
3282 			else
3283 				n = write_txpkt_wr(sc, txq, m0, avail);
3284 		}
3285 		MPASS(n >= 1 && n <= avail);
3286 		if (!(mbuf_cflags(m0) & MC_TLS))
3287 			MPASS(n <= SGE_MAX_WR_NDESC);
3288 
3289 		avail -= n;
3290 		dbdiff += n;
3291 		IDXINCR(eq->pidx, n, eq->sidx);
3292 
3293 		if (dbdiff >= 512 / EQ_ESIZE) {	/* X_FETCHBURSTMAX_512B */
3294 			if (wr_can_update_eq(wr))
3295 				set_txupdate_flags(txq, avail, wr);
3296 			ring_eq_db(sc, eq, dbdiff);
3297 			avail += reclaim_tx_descs(txq, 32);
3298 			dbdiff = 0;
3299 		}
3300 next_mbuf:
3301 		total++;
3302 		remaining--;
3303 		if (__predict_false(++cidx == r->size))
3304 			cidx = 0;
3305 	}
3306 	if (dbdiff != 0) {
3307 		if (wr_can_update_eq(wr))
3308 			set_txupdate_flags(txq, avail, wr);
3309 		ring_eq_db(sc, eq, dbdiff);
3310 		reclaim_tx_descs(txq, 32);
3311 	} else if (eq->pidx == eq->cidx && txp->npkt > 0 &&
3312 	    atomic_load_int(&txq->eq.equiq) == 0) {
3313 		/*
3314 		 * If nothing was submitted to the chip for tx (it was coalesced
3315 		 * into txpkts instead) and there is no tx update outstanding
3316 		 * then we need to send txpkts now.
3317 		 */
3318 send_txpkts:
3319 		MPASS(txp->npkt > 0);
3320 		for (i = 0; i < txp->npkt; i++)
3321 			ETHER_BPF_MTAP(ifp, txp->mb[i]);
3322 		if (txp->npkt > 1) {
3323 			MPASS(avail >= tx_len16_to_desc(txp->len16));
3324 			if (vi->flags & TX_USES_VM_WR)
3325 				n = write_txpkts_vm_wr(sc, txq);
3326 			else
3327 				n = write_txpkts_wr(sc, txq);
3328 		} else {
3329 			MPASS(avail >=
3330 			    tx_len16_to_desc(mbuf_len16(txp->mb[0])));
3331 			if (vi->flags & TX_USES_VM_WR)
3332 				n = write_txpkt_vm_wr(sc, txq, txp->mb[0]);
3333 			else
3334 				n = write_txpkt_wr(sc, txq, txp->mb[0], avail);
3335 		}
3336 		MPASS(n <= SGE_MAX_WR_NDESC);
3337 		wr = &eq->desc[eq->pidx];
3338 		IDXINCR(eq->pidx, n, eq->sidx);
3339 		txp->npkt = 0;	/* emptied */
3340 
3341 		MPASS(wr_can_update_eq(wr));
3342 		set_txupdate_flags(txq, avail - n, wr);
3343 		ring_eq_db(sc, eq, n);
3344 		reclaim_tx_descs(txq, 32);
3345 	}
3346 	*coalescing = txp->npkt > 0;
3347 
3348 	return (total);
3349 }
3350 
3351 static inline void
3352 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
3353     int qsize, int intr_idx, int cong)
3354 {
3355 
3356 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
3357 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
3358 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
3359 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
3360 	KASSERT(intr_idx >= -1 && intr_idx < sc->intr_count,
3361 	    ("%s: bad intr_idx %d", __func__, intr_idx));
3362 
3363 	iq->flags = 0;
3364 	iq->state = IQS_DISABLED;
3365 	iq->adapter = sc;
3366 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
3367 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
3368 	if (pktc_idx >= 0) {
3369 		iq->intr_params |= F_QINTR_CNT_EN;
3370 		iq->intr_pktc_idx = pktc_idx;
3371 	}
3372 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
3373 	iq->sidx = iq->qsize - sc->params.sge.spg_len / IQ_ESIZE;
3374 	iq->intr_idx = intr_idx;
3375 	iq->cong = cong;
3376 }
3377 
3378 static inline void
3379 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name)
3380 {
3381 	struct sge_params *sp = &sc->params.sge;
3382 
3383 	fl->qsize = qsize;
3384 	fl->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
3385 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
3386 	mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
3387 	if (sc->flags & BUF_PACKING_OK &&
3388 	    ((!is_t4(sc) && buffer_packing) ||	/* T5+: enabled unless 0 */
3389 	    (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */
3390 		fl->flags |= FL_BUF_PACKING;
3391 	fl->zidx = find_refill_source(sc, maxp, fl->flags & FL_BUF_PACKING);
3392 	fl->safe_zidx = sc->sge.safe_zidx;
3393 	if (fl->flags & FL_BUF_PACKING) {
3394 		fl->lowat = roundup2(sp->fl_starve_threshold2, 8);
3395 		fl->buf_boundary = sp->pack_boundary;
3396 	} else {
3397 		fl->lowat = roundup2(sp->fl_starve_threshold, 8);
3398 		fl->buf_boundary = 16;
3399 	}
3400 	if (fl_pad && fl->buf_boundary < sp->pad_boundary)
3401 		fl->buf_boundary = sp->pad_boundary;
3402 }
3403 
3404 static inline void
3405 init_eq(struct adapter *sc, struct sge_eq *eq, int eqtype, int qsize,
3406     uint8_t tx_chan, struct sge_iq *iq, char *name)
3407 {
3408 	KASSERT(eqtype >= EQ_CTRL && eqtype <= EQ_OFLD,
3409 	    ("%s: bad qtype %d", __func__, eqtype));
3410 
3411 	eq->type = eqtype;
3412 	eq->tx_chan = tx_chan;
3413 	eq->iq = iq;
3414 	eq->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
3415 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
3416 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
3417 }
3418 
3419 int
3420 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
3421     bus_dmamap_t *map, bus_addr_t *pa, void **va)
3422 {
3423 	int rc;
3424 
3425 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
3426 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
3427 	if (rc != 0) {
3428 		CH_ERR(sc, "cannot allocate DMA tag: %d\n", rc);
3429 		goto done;
3430 	}
3431 
3432 	rc = bus_dmamem_alloc(*tag, va,
3433 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
3434 	if (rc != 0) {
3435 		CH_ERR(sc, "cannot allocate DMA memory: %d\n", rc);
3436 		goto done;
3437 	}
3438 
3439 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
3440 	if (rc != 0) {
3441 		CH_ERR(sc, "cannot load DMA map: %d\n", rc);
3442 		goto done;
3443 	}
3444 done:
3445 	if (rc)
3446 		free_ring(sc, *tag, *map, *pa, *va);
3447 
3448 	return (rc);
3449 }
3450 
3451 int
3452 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
3453     bus_addr_t pa, void *va)
3454 {
3455 	if (pa)
3456 		bus_dmamap_unload(tag, map);
3457 	if (va)
3458 		bus_dmamem_free(tag, va, map);
3459 	if (tag)
3460 		bus_dma_tag_destroy(tag);
3461 
3462 	return (0);
3463 }
3464 
3465 /*
3466  * Allocates the software resources (mainly memory and sysctl nodes) for an
3467  * ingress queue and an optional freelist.
3468  *
3469  * Sets IQ_SW_ALLOCATED and returns 0 on success.
3470  */
3471 static int
3472 alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl,
3473     struct sysctl_ctx_list *ctx, struct sysctl_oid *oid)
3474 {
3475 	int rc;
3476 	size_t len;
3477 	struct adapter *sc = vi->adapter;
3478 
3479 	MPASS(!(iq->flags & IQ_SW_ALLOCATED));
3480 
3481 	len = iq->qsize * IQ_ESIZE;
3482 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
3483 	    (void **)&iq->desc);
3484 	if (rc != 0)
3485 		return (rc);
3486 
3487 	if (fl) {
3488 		len = fl->qsize * EQ_ESIZE;
3489 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
3490 		    &fl->ba, (void **)&fl->desc);
3491 		if (rc) {
3492 			free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba,
3493 			    iq->desc);
3494 			return (rc);
3495 		}
3496 
3497 		/* Allocate space for one software descriptor per buffer. */
3498 		fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc),
3499 		    M_CXGBE, M_ZERO | M_WAITOK);
3500 
3501 		add_fl_sysctls(sc, ctx, oid, fl);
3502 		iq->flags |= IQ_HAS_FL;
3503 	}
3504 	add_iq_sysctls(ctx, oid, iq);
3505 	iq->flags |= IQ_SW_ALLOCATED;
3506 
3507 	return (0);
3508 }
3509 
3510 /*
3511  * Frees all software resources (memory and locks) associated with an ingress
3512  * queue and an optional freelist.
3513  */
3514 static void
3515 free_iq_fl(struct adapter *sc, struct sge_iq *iq, struct sge_fl *fl)
3516 {
3517 	MPASS(iq->flags & IQ_SW_ALLOCATED);
3518 
3519 	if (fl) {
3520 		MPASS(iq->flags & IQ_HAS_FL);
3521 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba, fl->desc);
3522 		free_fl_buffers(sc, fl);
3523 		free(fl->sdesc, M_CXGBE);
3524 		mtx_destroy(&fl->fl_lock);
3525 		bzero(fl, sizeof(*fl));
3526 	}
3527 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
3528 	bzero(iq, sizeof(*iq));
3529 }
3530 
3531 /*
3532  * Allocates a hardware ingress queue and an optional freelist that will be
3533  * associated with it.
3534  *
3535  * Returns errno on failure.  Resources allocated up to that point may still be
3536  * allocated.  Caller is responsible for cleanup in case this function fails.
3537  */
3538 static int
3539 alloc_iq_fl_hwq(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl)
3540 {
3541 	int rc, i, cntxt_id;
3542 	struct fw_iq_cmd c;
3543 	struct adapter *sc = vi->adapter;
3544 	__be32 v = 0;
3545 
3546 	MPASS (!(iq->flags & IQ_HW_ALLOCATED));
3547 
3548 	bzero(&c, sizeof(c));
3549 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
3550 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
3551 	    V_FW_IQ_CMD_VFN(0));
3552 
3553 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
3554 	    FW_LEN16(c));
3555 
3556 	/* Special handling for firmware event queue */
3557 	if (iq == &sc->sge.fwq)
3558 		v |= F_FW_IQ_CMD_IQASYNCH;
3559 
3560 	if (iq->intr_idx < 0) {
3561 		/* Forwarded interrupts, all headed to fwq */
3562 		v |= F_FW_IQ_CMD_IQANDST;
3563 		v |= V_FW_IQ_CMD_IQANDSTINDEX(sc->sge.fwq.cntxt_id);
3564 	} else {
3565 		KASSERT(iq->intr_idx < sc->intr_count,
3566 		    ("%s: invalid direct intr_idx %d", __func__, iq->intr_idx));
3567 		v |= V_FW_IQ_CMD_IQANDSTINDEX(iq->intr_idx);
3568 	}
3569 
3570 	bzero(iq->desc, iq->qsize * IQ_ESIZE);
3571 	c.type_to_iqandstindex = htobe32(v |
3572 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
3573 	    V_FW_IQ_CMD_VIID(vi->viid) |
3574 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
3575 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(vi->pi->tx_chan) |
3576 	    F_FW_IQ_CMD_IQGTSMODE |
3577 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
3578 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
3579 	c.iqsize = htobe16(iq->qsize);
3580 	c.iqaddr = htobe64(iq->ba);
3581 	if (iq->cong >= 0)
3582 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
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 >= 0) {
3593 			c.iqns_to_fl0congen |=
3594 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(iq->cong) |
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 		MPASS(!(fl->flags & FL_BUF_RESUME));
3629 		for (i = 0; i < fl->sidx * 8; i++)
3630 			MPASS(fl->sdesc[i].cl == NULL);
3631 #endif
3632 		fl->cntxt_id = be16toh(c.fl0id);
3633 		fl->pidx = fl->cidx = fl->hw_cidx = fl->dbidx = 0;
3634 		fl->rx_offset = 0;
3635 		fl->flags &= ~(FL_STARVING | FL_DOOMED);
3636 
3637 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
3638 		if (cntxt_id >= sc->sge.eqmap_sz) {
3639 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
3640 			    __func__, cntxt_id, sc->sge.eqmap_sz - 1);
3641 		}
3642 		sc->sge.eqmap[cntxt_id] = (void *)fl;
3643 
3644 		qid = fl->cntxt_id;
3645 		if (isset(&sc->doorbells, DOORBELL_UDB)) {
3646 			uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3647 			uint32_t mask = (1 << s_qpp) - 1;
3648 			volatile uint8_t *udb;
3649 
3650 			udb = sc->udbs_base + UDBS_DB_OFFSET;
3651 			udb += (qid >> s_qpp) << PAGE_SHIFT;
3652 			qid &= mask;
3653 			if (qid < PAGE_SIZE / UDBS_SEG_SIZE) {
3654 				udb += qid << UDBS_SEG_SHIFT;
3655 				qid = 0;
3656 			}
3657 			fl->udb = (volatile void *)udb;
3658 		}
3659 		fl->dbval = V_QID(qid) | sc->chip_params->sge_fl_db;
3660 
3661 		FL_LOCK(fl);
3662 		/* Enough to make sure the SGE doesn't think it's starved */
3663 		refill_fl(sc, fl, fl->lowat);
3664 		FL_UNLOCK(fl);
3665 	}
3666 
3667 	if (chip_id(sc) >= CHELSIO_T5 && !(sc->flags & IS_VF) && iq->cong >= 0) {
3668 		uint32_t param, val;
3669 
3670 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
3671 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
3672 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
3673 		if (iq->cong == 0)
3674 			val = 1 << 19;
3675 		else {
3676 			val = 2 << 19;
3677 			for (i = 0; i < 4; i++) {
3678 				if (iq->cong & (1 << i))
3679 					val |= 1 << (i << 2);
3680 			}
3681 		}
3682 
3683 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
3684 		if (rc != 0) {
3685 			/* report error but carry on */
3686 			CH_ERR(sc, "failed to set congestion manager context "
3687 			    "for ingress queue %d: %d\n", iq->cntxt_id, rc);
3688 		}
3689 	}
3690 
3691 	/* Enable IQ interrupts */
3692 	atomic_store_rel_int(&iq->state, IQS_IDLE);
3693 	t4_write_reg(sc, sc->sge_gts_reg, V_SEINTARM(iq->intr_params) |
3694 	    V_INGRESSQID(iq->cntxt_id));
3695 
3696 	iq->flags |= IQ_HW_ALLOCATED;
3697 
3698 	return (0);
3699 }
3700 
3701 static int
3702 free_iq_fl_hwq(struct adapter *sc, struct sge_iq *iq, struct sge_fl *fl)
3703 {
3704 	int rc;
3705 
3706 	MPASS(iq->flags & IQ_HW_ALLOCATED);
3707 	rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0, FW_IQ_TYPE_FL_INT_CAP,
3708 	    iq->cntxt_id, fl ? fl->cntxt_id : 0xffff, 0xffff);
3709 	if (rc != 0) {
3710 		CH_ERR(sc, "failed to free iq %p: %d\n", iq, rc);
3711 		return (rc);
3712 	}
3713 	iq->flags &= ~IQ_HW_ALLOCATED;
3714 
3715 	return (0);
3716 }
3717 
3718 static void
3719 add_iq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
3720     struct sge_iq *iq)
3721 {
3722 	struct sysctl_oid_list *children;
3723 
3724 	if (ctx == NULL || oid == NULL)
3725 		return;
3726 
3727 	children = SYSCTL_CHILDREN(oid);
3728 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD, &iq->ba,
3729 	    "bus address of descriptor ring");
3730 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3731 	    iq->qsize * IQ_ESIZE, "descriptor ring size in bytes");
3732 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
3733 	    &iq->abs_id, 0, "absolute id of the queue");
3734 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3735 	    &iq->cntxt_id, 0, "SGE context id of the queue");
3736 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &iq->cidx,
3737 	    0, "consumer index");
3738 }
3739 
3740 static void
3741 add_fl_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
3742     struct sysctl_oid *oid, struct sge_fl *fl)
3743 {
3744 	struct sysctl_oid_list *children;
3745 
3746 	if (ctx == NULL || oid == NULL)
3747 		return;
3748 
3749 	children = SYSCTL_CHILDREN(oid);
3750 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl",
3751 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist");
3752 	children = SYSCTL_CHILDREN(oid);
3753 
3754 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3755 	    &fl->ba, "bus address of descriptor ring");
3756 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3757 	    fl->sidx * EQ_ESIZE + sc->params.sge.spg_len,
3758 	    "desc ring size in bytes");
3759 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3760 	    &fl->cntxt_id, 0, "SGE context id of the freelist");
3761 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL,
3762 	    fl_pad ? 1 : 0, "padding enabled");
3763 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL,
3764 	    fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled");
3765 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
3766 	    0, "consumer index");
3767 	if (fl->flags & FL_BUF_PACKING) {
3768 		SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
3769 		    CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
3770 	}
3771 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
3772 	    0, "producer index");
3773 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
3774 	    CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
3775 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
3776 	    CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
3777 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
3778 	    CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
3779 }
3780 
3781 /*
3782  * Idempotent.
3783  */
3784 static int
3785 alloc_fwq(struct adapter *sc)
3786 {
3787 	int rc, intr_idx;
3788 	struct sge_iq *fwq = &sc->sge.fwq;
3789 	struct vi_info *vi = &sc->port[0]->vi[0];
3790 
3791 	if (!(fwq->flags & IQ_SW_ALLOCATED)) {
3792 		MPASS(!(fwq->flags & IQ_HW_ALLOCATED));
3793 
3794 		if (sc->flags & IS_VF)
3795 			intr_idx = 0;
3796 		else
3797 			intr_idx = sc->intr_count > 1 ? 1 : 0;
3798 		init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE, intr_idx, -1);
3799 		rc = alloc_iq_fl(vi, fwq, NULL, &sc->ctx, sc->fwq_oid);
3800 		if (rc != 0) {
3801 			CH_ERR(sc, "failed to allocate fwq: %d\n", rc);
3802 			return (rc);
3803 		}
3804 		MPASS(fwq->flags & IQ_SW_ALLOCATED);
3805 	}
3806 
3807 	if (!(fwq->flags & IQ_HW_ALLOCATED)) {
3808 		MPASS(fwq->flags & IQ_SW_ALLOCATED);
3809 
3810 		rc = alloc_iq_fl_hwq(vi, fwq, NULL);
3811 		if (rc != 0) {
3812 			CH_ERR(sc, "failed to create hw fwq: %d\n", rc);
3813 			return (rc);
3814 		}
3815 		MPASS(fwq->flags & IQ_HW_ALLOCATED);
3816 	}
3817 
3818 	return (0);
3819 }
3820 
3821 /*
3822  * Idempotent.
3823  */
3824 static void
3825 free_fwq(struct adapter *sc)
3826 {
3827 	struct sge_iq *fwq = &sc->sge.fwq;
3828 
3829 	if (fwq->flags & IQ_HW_ALLOCATED) {
3830 		MPASS(fwq->flags & IQ_SW_ALLOCATED);
3831 		free_iq_fl_hwq(sc, fwq, NULL);
3832 		MPASS(!(fwq->flags & IQ_HW_ALLOCATED));
3833 	}
3834 
3835 	if (fwq->flags & IQ_SW_ALLOCATED) {
3836 		MPASS(!(fwq->flags & IQ_HW_ALLOCATED));
3837 		free_iq_fl(sc, fwq, NULL);
3838 		MPASS(!(fwq->flags & IQ_SW_ALLOCATED));
3839 	}
3840 }
3841 
3842 /*
3843  * Idempotent.
3844  */
3845 static int
3846 alloc_ctrlq(struct adapter *sc, int idx)
3847 {
3848 	int rc;
3849 	char name[16];
3850 	struct sysctl_oid *oid;
3851 	struct sge_wrq *ctrlq = &sc->sge.ctrlq[idx];
3852 
3853 	MPASS(idx < sc->params.nports);
3854 
3855 	if (!(ctrlq->eq.flags & EQ_SW_ALLOCATED)) {
3856 		MPASS(!(ctrlq->eq.flags & EQ_HW_ALLOCATED));
3857 
3858 		snprintf(name, sizeof(name), "%d", idx);
3859 		oid = SYSCTL_ADD_NODE(&sc->ctx, SYSCTL_CHILDREN(sc->ctrlq_oid),
3860 		    OID_AUTO, name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
3861 		    "ctrl queue");
3862 
3863 		snprintf(name, sizeof(name), "%s ctrlq%d",
3864 		    device_get_nameunit(sc->dev), idx);
3865 		init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE,
3866 		    sc->port[idx]->tx_chan, &sc->sge.fwq, name);
3867 		rc = alloc_wrq(sc, NULL, ctrlq, &sc->ctx, oid);
3868 		if (rc != 0) {
3869 			CH_ERR(sc, "failed to allocate ctrlq%d: %d\n", idx, rc);
3870 			sysctl_remove_oid(oid, 1, 1);
3871 			return (rc);
3872 		}
3873 		MPASS(ctrlq->eq.flags & EQ_SW_ALLOCATED);
3874 	}
3875 
3876 	if (!(ctrlq->eq.flags & EQ_HW_ALLOCATED)) {
3877 		MPASS(ctrlq->eq.flags & EQ_SW_ALLOCATED);
3878 
3879 		rc = alloc_eq_hwq(sc, NULL, &ctrlq->eq);
3880 		if (rc != 0) {
3881 			CH_ERR(sc, "failed to create hw ctrlq%d: %d\n", idx, rc);
3882 			return (rc);
3883 		}
3884 		MPASS(ctrlq->eq.flags & EQ_HW_ALLOCATED);
3885 	}
3886 
3887 	return (0);
3888 }
3889 
3890 /*
3891  * Idempotent.
3892  */
3893 static void
3894 free_ctrlq(struct adapter *sc, int idx)
3895 {
3896 	struct sge_wrq *ctrlq = &sc->sge.ctrlq[idx];
3897 
3898 	if (ctrlq->eq.flags & EQ_HW_ALLOCATED) {
3899 		MPASS(ctrlq->eq.flags & EQ_SW_ALLOCATED);
3900 		free_eq_hwq(sc, NULL, &ctrlq->eq);
3901 		MPASS(!(ctrlq->eq.flags & EQ_HW_ALLOCATED));
3902 	}
3903 
3904 	if (ctrlq->eq.flags & EQ_SW_ALLOCATED) {
3905 		MPASS(!(ctrlq->eq.flags & EQ_HW_ALLOCATED));
3906 		free_wrq(sc, ctrlq);
3907 		MPASS(!(ctrlq->eq.flags & EQ_SW_ALLOCATED));
3908 	}
3909 }
3910 
3911 int
3912 tnl_cong(struct port_info *pi, int drop)
3913 {
3914 
3915 	if (drop == -1)
3916 		return (-1);
3917 	else if (drop == 1)
3918 		return (0);
3919 	else
3920 		return (pi->rx_e_chan_map);
3921 }
3922 
3923 /*
3924  * Idempotent.
3925  */
3926 static int
3927 alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int idx, int intr_idx,
3928     int maxp)
3929 {
3930 	int rc;
3931 	struct adapter *sc = vi->adapter;
3932 	struct ifnet *ifp = vi->ifp;
3933 	struct sysctl_oid *oid;
3934 	char name[16];
3935 
3936 	if (!(rxq->iq.flags & IQ_SW_ALLOCATED)) {
3937 		MPASS(!(rxq->iq.flags & IQ_HW_ALLOCATED));
3938 #if defined(INET) || defined(INET6)
3939 		rc = tcp_lro_init_args(&rxq->lro, ifp, lro_entries, lro_mbufs);
3940 		if (rc != 0)
3941 			return (rc);
3942 		MPASS(rxq->lro.ifp == ifp);	/* also indicates LRO init'ed */
3943 #endif
3944 		rxq->ifp = ifp;
3945 
3946 		snprintf(name, sizeof(name), "%d", idx);
3947 		oid = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(vi->rxq_oid),
3948 		    OID_AUTO, name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
3949 		    "rx queue");
3950 
3951 		init_iq(&rxq->iq, sc, vi->tmr_idx, vi->pktc_idx, vi->qsize_rxq,
3952 		    intr_idx, tnl_cong(vi->pi, cong_drop));
3953 #if defined(INET) || defined(INET6)
3954 		if (ifp->if_capenable & IFCAP_LRO)
3955 			rxq->iq.flags |= IQ_LRO_ENABLED;
3956 #endif
3957 		if (ifp->if_capenable & IFCAP_HWRXTSTMP)
3958 			rxq->iq.flags |= IQ_RX_TIMESTAMP;
3959 		snprintf(name, sizeof(name), "%s rxq%d-fl",
3960 		    device_get_nameunit(vi->dev), idx);
3961 		init_fl(sc, &rxq->fl, vi->qsize_rxq / 8, maxp, name);
3962 		rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, &vi->ctx, oid);
3963 		if (rc != 0) {
3964 			CH_ERR(vi, "failed to allocate rxq%d: %d\n", idx, rc);
3965 			sysctl_remove_oid(oid, 1, 1);
3966 #if defined(INET) || defined(INET6)
3967 			tcp_lro_free(&rxq->lro);
3968 			rxq->lro.ifp = NULL;
3969 #endif
3970 			return (rc);
3971 		}
3972 		MPASS(rxq->iq.flags & IQ_SW_ALLOCATED);
3973 		add_rxq_sysctls(&vi->ctx, oid, rxq);
3974 	}
3975 
3976 	if (!(rxq->iq.flags & IQ_HW_ALLOCATED)) {
3977 		MPASS(rxq->iq.flags & IQ_SW_ALLOCATED);
3978 		rc = alloc_iq_fl_hwq(vi, &rxq->iq, &rxq->fl);
3979 		if (rc != 0) {
3980 			CH_ERR(vi, "failed to create hw rxq%d: %d\n", idx, rc);
3981 			return (rc);
3982 		}
3983 		MPASS(rxq->iq.flags & IQ_HW_ALLOCATED);
3984 
3985 		if (idx == 0)
3986 			sc->sge.iq_base = rxq->iq.abs_id - rxq->iq.cntxt_id;
3987 		else
3988 			KASSERT(rxq->iq.cntxt_id + sc->sge.iq_base == rxq->iq.abs_id,
3989 			    ("iq_base mismatch"));
3990 		KASSERT(sc->sge.iq_base == 0 || sc->flags & IS_VF,
3991 		    ("PF with non-zero iq_base"));
3992 
3993 		/*
3994 		 * The freelist is just barely above the starvation threshold
3995 		 * right now, fill it up a bit more.
3996 		 */
3997 		FL_LOCK(&rxq->fl);
3998 		refill_fl(sc, &rxq->fl, 128);
3999 		FL_UNLOCK(&rxq->fl);
4000 	}
4001 
4002 	return (0);
4003 }
4004 
4005 /*
4006  * Idempotent.
4007  */
4008 static void
4009 free_rxq(struct vi_info *vi, struct sge_rxq *rxq)
4010 {
4011 	if (rxq->iq.flags & IQ_HW_ALLOCATED) {
4012 		MPASS(rxq->iq.flags & IQ_SW_ALLOCATED);
4013 		free_iq_fl_hwq(vi->adapter, &rxq->iq, &rxq->fl);
4014 		MPASS(!(rxq->iq.flags & IQ_HW_ALLOCATED));
4015 	}
4016 
4017 	if (rxq->iq.flags & IQ_SW_ALLOCATED) {
4018 		MPASS(!(rxq->iq.flags & IQ_HW_ALLOCATED));
4019 #if defined(INET) || defined(INET6)
4020 		tcp_lro_free(&rxq->lro);
4021 #endif
4022 		free_iq_fl(vi->adapter, &rxq->iq, &rxq->fl);
4023 		MPASS(!(rxq->iq.flags & IQ_SW_ALLOCATED));
4024 		bzero(rxq, sizeof(*rxq));
4025 	}
4026 }
4027 
4028 static void
4029 add_rxq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
4030     struct sge_rxq *rxq)
4031 {
4032 	struct sysctl_oid_list *children;
4033 
4034 	if (ctx == NULL || oid == NULL)
4035 		return;
4036 
4037 	children = SYSCTL_CHILDREN(oid);
4038 #if defined(INET) || defined(INET6)
4039 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
4040 	    &rxq->lro.lro_queued, 0, NULL);
4041 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
4042 	    &rxq->lro.lro_flushed, 0, NULL);
4043 #endif
4044 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
4045 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
4046 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "vlan_extraction", CTLFLAG_RD,
4047 	    &rxq->vlan_extraction, "# of times hardware extracted 802.1Q tag");
4048 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "vxlan_rxcsum", CTLFLAG_RD,
4049 	    &rxq->vxlan_rxcsum,
4050 	    "# of times hardware assisted with inner checksum (VXLAN)");
4051 }
4052 
4053 #ifdef TCP_OFFLOAD
4054 /*
4055  * Idempotent.
4056  */
4057 static int
4058 alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq, int idx,
4059     int intr_idx, int maxp)
4060 {
4061 	int rc;
4062 	struct adapter *sc = vi->adapter;
4063 	struct sysctl_oid *oid;
4064 	char name[16];
4065 
4066 	if (!(ofld_rxq->iq.flags & IQ_SW_ALLOCATED)) {
4067 		MPASS(!(ofld_rxq->iq.flags & IQ_HW_ALLOCATED));
4068 
4069 		snprintf(name, sizeof(name), "%d", idx);
4070 		oid = SYSCTL_ADD_NODE(&vi->ctx,
4071 		    SYSCTL_CHILDREN(vi->ofld_rxq_oid), OID_AUTO, name,
4072 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "offload rx queue");
4073 
4074 		init_iq(&ofld_rxq->iq, sc, vi->ofld_tmr_idx, vi->ofld_pktc_idx,
4075 		    vi->qsize_rxq, intr_idx, 0);
4076 		snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
4077 		    device_get_nameunit(vi->dev), idx);
4078 		init_fl(sc, &ofld_rxq->fl, vi->qsize_rxq / 8, maxp, name);
4079 		rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, &vi->ctx,
4080 		    oid);
4081 		if (rc != 0) {
4082 			CH_ERR(vi, "failed to allocate ofld_rxq%d: %d\n", idx,
4083 			    rc);
4084 			sysctl_remove_oid(oid, 1, 1);
4085 			return (rc);
4086 		}
4087 		MPASS(ofld_rxq->iq.flags & IQ_SW_ALLOCATED);
4088 		ofld_rxq->rx_iscsi_ddp_setup_ok = counter_u64_alloc(M_WAITOK);
4089 		ofld_rxq->rx_iscsi_ddp_setup_error =
4090 		    counter_u64_alloc(M_WAITOK);
4091 		add_ofld_rxq_sysctls(&vi->ctx, oid, ofld_rxq);
4092 	}
4093 
4094 	if (!(ofld_rxq->iq.flags & IQ_HW_ALLOCATED)) {
4095 		MPASS(ofld_rxq->iq.flags & IQ_SW_ALLOCATED);
4096 		rc = alloc_iq_fl_hwq(vi, &ofld_rxq->iq, &ofld_rxq->fl);
4097 		if (rc != 0) {
4098 			CH_ERR(vi, "failed to create hw ofld_rxq%d: %d\n", idx,
4099 			    rc);
4100 			return (rc);
4101 		}
4102 		MPASS(ofld_rxq->iq.flags & IQ_HW_ALLOCATED);
4103 	}
4104 	return (rc);
4105 }
4106 
4107 /*
4108  * Idempotent.
4109  */
4110 static void
4111 free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq)
4112 {
4113 	if (ofld_rxq->iq.flags & IQ_HW_ALLOCATED) {
4114 		MPASS(ofld_rxq->iq.flags & IQ_SW_ALLOCATED);
4115 		free_iq_fl_hwq(vi->adapter, &ofld_rxq->iq, &ofld_rxq->fl);
4116 		MPASS(!(ofld_rxq->iq.flags & IQ_HW_ALLOCATED));
4117 	}
4118 
4119 	if (ofld_rxq->iq.flags & IQ_SW_ALLOCATED) {
4120 		MPASS(!(ofld_rxq->iq.flags & IQ_HW_ALLOCATED));
4121 		free_iq_fl(vi->adapter, &ofld_rxq->iq, &ofld_rxq->fl);
4122 		MPASS(!(ofld_rxq->iq.flags & IQ_SW_ALLOCATED));
4123 		counter_u64_free(ofld_rxq->rx_iscsi_ddp_setup_ok);
4124 		counter_u64_free(ofld_rxq->rx_iscsi_ddp_setup_error);
4125 		bzero(ofld_rxq, sizeof(*ofld_rxq));
4126 	}
4127 }
4128 
4129 static void
4130 add_ofld_rxq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
4131     struct sge_ofld_rxq *ofld_rxq)
4132 {
4133 	struct sysctl_oid_list *children;
4134 
4135 	if (ctx == NULL || oid == NULL)
4136 		return;
4137 
4138 	children = SYSCTL_CHILDREN(oid);
4139 	SYSCTL_ADD_ULONG(ctx, children, OID_AUTO,
4140 	    "rx_toe_tls_records", CTLFLAG_RD, &ofld_rxq->rx_toe_tls_records,
4141 	    "# of TOE TLS records received");
4142 	SYSCTL_ADD_ULONG(ctx, children, OID_AUTO,
4143 	    "rx_toe_tls_octets", CTLFLAG_RD, &ofld_rxq->rx_toe_tls_octets,
4144 	    "# of payload octets in received TOE TLS records");
4145 
4146 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "iscsi",
4147 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "TOE iSCSI statistics");
4148 	children = SYSCTL_CHILDREN(oid);
4149 
4150 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "ddp_setup_ok",
4151 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_ddp_setup_ok,
4152 	    "# of times DDP buffer was setup successfully.");
4153 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "ddp_setup_error",
4154 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_ddp_setup_error,
4155 	    "# of times DDP buffer setup failed.");
4156 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "ddp_octets",
4157 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_ddp_octets, 0,
4158 	    "# of octets placed directly");
4159 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "ddp_pdus",
4160 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_ddp_pdus, 0,
4161 	    "# of PDUs with data placed directly.");
4162 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "fl_octets",
4163 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_fl_octets, 0,
4164 	    "# of data octets delivered in freelist");
4165 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "fl_pdus",
4166 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_fl_pdus, 0,
4167 	    "# of PDUs with data delivered in freelist");
4168 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "padding_errors",
4169 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_padding_errors, 0,
4170 	    "# of PDUs with invalid padding");
4171 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "header_digest_errors",
4172 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_header_digest_errors, 0,
4173 	    "# of PDUs with invalid header digests");
4174 	SYSCTL_ADD_U64(ctx, children, OID_AUTO, "data_digest_errors",
4175 	    CTLFLAG_RD, &ofld_rxq->rx_iscsi_data_digest_errors, 0,
4176 	    "# of PDUs with invalid data digests");
4177 }
4178 #endif
4179 
4180 /*
4181  * Returns a reasonable automatic cidx flush threshold for a given queue size.
4182  */
4183 static u_int
4184 qsize_to_fthresh(int qsize)
4185 {
4186 	u_int fthresh;
4187 
4188 	while (!powerof2(qsize))
4189 		qsize++;
4190 	fthresh = ilog2(qsize);
4191 	if (fthresh > X_CIDXFLUSHTHRESH_128)
4192 		fthresh = X_CIDXFLUSHTHRESH_128;
4193 
4194 	return (fthresh);
4195 }
4196 
4197 static int
4198 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
4199 {
4200 	int rc, cntxt_id;
4201 	struct fw_eq_ctrl_cmd c;
4202 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
4203 
4204 	bzero(&c, sizeof(c));
4205 
4206 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
4207 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
4208 	    V_FW_EQ_CTRL_CMD_VFN(0));
4209 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
4210 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
4211 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid));
4212 	c.physeqid_pkd = htobe32(0);
4213 	c.fetchszm_to_iqid =
4214 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
4215 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
4216 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
4217 	c.dcaen_to_eqsize =
4218 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
4219 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
4220 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
4221 		V_FW_EQ_CTRL_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
4222 		V_FW_EQ_CTRL_CMD_EQSIZE(qsize));
4223 	c.eqaddr = htobe64(eq->ba);
4224 
4225 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
4226 	if (rc != 0) {
4227 		CH_ERR(sc, "failed to create hw ctrlq for tx_chan %d: %d\n",
4228 		    eq->tx_chan, rc);
4229 		return (rc);
4230 	}
4231 
4232 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
4233 	eq->abs_id = G_FW_EQ_CTRL_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
4234 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
4235 	if (cntxt_id >= sc->sge.eqmap_sz)
4236 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
4237 		cntxt_id, sc->sge.eqmap_sz - 1);
4238 	sc->sge.eqmap[cntxt_id] = eq;
4239 
4240 	return (rc);
4241 }
4242 
4243 static int
4244 eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
4245 {
4246 	int rc, cntxt_id;
4247 	struct fw_eq_eth_cmd c;
4248 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
4249 
4250 	bzero(&c, sizeof(c));
4251 
4252 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
4253 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
4254 	    V_FW_EQ_ETH_CMD_VFN(0));
4255 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
4256 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
4257 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
4258 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
4259 	c.fetchszm_to_iqid =
4260 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
4261 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
4262 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
4263 	c.dcaen_to_eqsize =
4264 	    htobe32(V_FW_EQ_ETH_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
4265 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
4266 		V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
4267 		V_FW_EQ_ETH_CMD_EQSIZE(qsize));
4268 	c.eqaddr = htobe64(eq->ba);
4269 
4270 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
4271 	if (rc != 0) {
4272 		device_printf(vi->dev,
4273 		    "failed to create Ethernet egress queue: %d\n", rc);
4274 		return (rc);
4275 	}
4276 
4277 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
4278 	eq->abs_id = G_FW_EQ_ETH_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
4279 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
4280 	if (cntxt_id >= sc->sge.eqmap_sz)
4281 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
4282 		cntxt_id, sc->sge.eqmap_sz - 1);
4283 	sc->sge.eqmap[cntxt_id] = eq;
4284 
4285 	return (rc);
4286 }
4287 
4288 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
4289 static int
4290 ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
4291 {
4292 	int rc, cntxt_id;
4293 	struct fw_eq_ofld_cmd c;
4294 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
4295 
4296 	bzero(&c, sizeof(c));
4297 
4298 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
4299 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
4300 	    V_FW_EQ_OFLD_CMD_VFN(0));
4301 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
4302 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
4303 	c.fetchszm_to_iqid =
4304 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
4305 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
4306 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
4307 	c.dcaen_to_eqsize =
4308 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
4309 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
4310 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
4311 		V_FW_EQ_OFLD_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
4312 		V_FW_EQ_OFLD_CMD_EQSIZE(qsize));
4313 	c.eqaddr = htobe64(eq->ba);
4314 
4315 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
4316 	if (rc != 0) {
4317 		device_printf(vi->dev,
4318 		    "failed to create egress queue for TCP offload: %d\n", rc);
4319 		return (rc);
4320 	}
4321 
4322 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
4323 	eq->abs_id = G_FW_EQ_OFLD_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
4324 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
4325 	if (cntxt_id >= sc->sge.eqmap_sz)
4326 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
4327 		cntxt_id, sc->sge.eqmap_sz - 1);
4328 	sc->sge.eqmap[cntxt_id] = eq;
4329 
4330 	return (rc);
4331 }
4332 #endif
4333 
4334 /* SW only */
4335 static int
4336 alloc_eq(struct adapter *sc, struct sge_eq *eq, struct sysctl_ctx_list *ctx,
4337     struct sysctl_oid *oid)
4338 {
4339 	int rc, qsize;
4340 	size_t len;
4341 
4342 	MPASS(!(eq->flags & EQ_SW_ALLOCATED));
4343 
4344 	qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
4345 	len = qsize * EQ_ESIZE;
4346 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map, &eq->ba,
4347 	    (void **)&eq->desc);
4348 	if (rc)
4349 		return (rc);
4350 	if (ctx != NULL && oid != NULL)
4351 		add_eq_sysctls(sc, ctx, oid, eq);
4352 	eq->flags |= EQ_SW_ALLOCATED;
4353 
4354 	return (0);
4355 }
4356 
4357 /* SW only */
4358 static void
4359 free_eq(struct adapter *sc, struct sge_eq *eq)
4360 {
4361 	MPASS(eq->flags & EQ_SW_ALLOCATED);
4362 	if (eq->type == EQ_ETH)
4363 		MPASS(eq->pidx == eq->cidx);
4364 
4365 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
4366 	mtx_destroy(&eq->eq_lock);
4367 	bzero(eq, sizeof(*eq));
4368 }
4369 
4370 static void
4371 add_eq_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
4372     struct sysctl_oid *oid, struct sge_eq *eq)
4373 {
4374 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
4375 
4376 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD, &eq->ba,
4377 	    "bus address of descriptor ring");
4378 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
4379 	    eq->sidx * EQ_ESIZE + sc->params.sge.spg_len,
4380 	    "desc ring size in bytes");
4381 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
4382 	    &eq->abs_id, 0, "absolute id of the queue");
4383 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
4384 	    &eq->cntxt_id, 0, "SGE context id of the queue");
4385 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &eq->cidx,
4386 	    0, "consumer index");
4387 	SYSCTL_ADD_U16(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &eq->pidx,
4388 	    0, "producer index");
4389 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
4390 	    eq->sidx, "status page index");
4391 }
4392 
4393 static int
4394 alloc_eq_hwq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
4395 {
4396 	int rc;
4397 
4398 	MPASS(!(eq->flags & EQ_HW_ALLOCATED));
4399 
4400 	eq->iqid = eq->iq->cntxt_id;
4401 	eq->pidx = eq->cidx = eq->dbidx = 0;
4402 	/* Note that equeqidx is not used with sge_wrq (OFLD/CTRL) queues. */
4403 	eq->equeqidx = 0;
4404 	eq->doorbells = sc->doorbells;
4405 	bzero(eq->desc, eq->sidx * EQ_ESIZE + sc->params.sge.spg_len);
4406 
4407 	switch (eq->type) {
4408 	case EQ_CTRL:
4409 		rc = ctrl_eq_alloc(sc, eq);
4410 		break;
4411 
4412 	case EQ_ETH:
4413 		rc = eth_eq_alloc(sc, vi, eq);
4414 		break;
4415 
4416 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
4417 	case EQ_OFLD:
4418 		rc = ofld_eq_alloc(sc, vi, eq);
4419 		break;
4420 #endif
4421 
4422 	default:
4423 		panic("%s: invalid eq type %d.", __func__, eq->type);
4424 	}
4425 	if (rc != 0) {
4426 		CH_ERR(sc, "failed to allocate egress queue(%d): %d\n",
4427 		    eq->type, rc);
4428 		return (rc);
4429 	}
4430 
4431 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
4432 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
4433 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
4434 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
4435 		uint32_t mask = (1 << s_qpp) - 1;
4436 		volatile uint8_t *udb;
4437 
4438 		udb = sc->udbs_base + UDBS_DB_OFFSET;
4439 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
4440 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
4441 		if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
4442 	    		clrbit(&eq->doorbells, DOORBELL_WCWR);
4443 		else {
4444 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
4445 			eq->udb_qid = 0;
4446 		}
4447 		eq->udb = (volatile void *)udb;
4448 	}
4449 
4450 	eq->flags |= EQ_HW_ALLOCATED;
4451 	return (0);
4452 }
4453 
4454 static int
4455 free_eq_hwq(struct adapter *sc, struct vi_info *vi __unused, struct sge_eq *eq)
4456 {
4457 	int rc;
4458 
4459 	MPASS(eq->flags & EQ_HW_ALLOCATED);
4460 
4461 	switch (eq->type) {
4462 	case EQ_CTRL:
4463 		rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0, eq->cntxt_id);
4464 		break;
4465 	case EQ_ETH:
4466 		rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0, eq->cntxt_id);
4467 		break;
4468 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
4469 	case EQ_OFLD:
4470 		rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0, eq->cntxt_id);
4471 		break;
4472 #endif
4473 	default:
4474 		panic("%s: invalid eq type %d.", __func__, eq->type);
4475 	}
4476 	if (rc != 0) {
4477 		CH_ERR(sc, "failed to free eq (type %d): %d\n", eq->type, rc);
4478 		return (rc);
4479 	}
4480 	eq->flags &= ~EQ_HW_ALLOCATED;
4481 
4482 	return (0);
4483 }
4484 
4485 static int
4486 alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq,
4487     struct sysctl_ctx_list *ctx, struct sysctl_oid *oid)
4488 {
4489 	struct sge_eq *eq = &wrq->eq;
4490 	int rc;
4491 
4492 	MPASS(!(eq->flags & EQ_SW_ALLOCATED));
4493 
4494 	rc = alloc_eq(sc, eq, ctx, oid);
4495 	if (rc)
4496 		return (rc);
4497 	MPASS(eq->flags & EQ_SW_ALLOCATED);
4498 	/* Can't fail after this. */
4499 
4500 	wrq->adapter = sc;
4501 	TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq);
4502 	TAILQ_INIT(&wrq->incomplete_wrs);
4503 	STAILQ_INIT(&wrq->wr_list);
4504 	wrq->nwr_pending = 0;
4505 	wrq->ndesc_needed = 0;
4506 	add_wrq_sysctls(ctx, oid, wrq);
4507 
4508 	return (0);
4509 }
4510 
4511 static void
4512 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
4513 {
4514 	free_eq(sc, &wrq->eq);
4515 	MPASS(wrq->nwr_pending == 0);
4516 	MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
4517 	MPASS(STAILQ_EMPTY(&wrq->wr_list));
4518 	bzero(wrq, sizeof(*wrq));
4519 }
4520 
4521 static void
4522 add_wrq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
4523     struct sge_wrq *wrq)
4524 {
4525 	struct sysctl_oid_list *children;
4526 
4527 	if (ctx == NULL || oid == NULL)
4528 		return;
4529 
4530 	children = SYSCTL_CHILDREN(oid);
4531 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD,
4532 	    &wrq->tx_wrs_direct, "# of work requests (direct)");
4533 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD,
4534 	    &wrq->tx_wrs_copied, "# of work requests (copied)");
4535 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_sspace", CTLFLAG_RD,
4536 	    &wrq->tx_wrs_ss, "# of work requests (copied from scratch space)");
4537 }
4538 
4539 /*
4540  * Idempotent.
4541  */
4542 static int
4543 alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx)
4544 {
4545 	int rc, iqidx;
4546 	struct port_info *pi = vi->pi;
4547 	struct adapter *sc = vi->adapter;
4548 	struct sge_eq *eq = &txq->eq;
4549 	struct txpkts *txp;
4550 	char name[16];
4551 	struct sysctl_oid *oid;
4552 
4553 	if (!(eq->flags & EQ_SW_ALLOCATED)) {
4554 		MPASS(!(eq->flags & EQ_HW_ALLOCATED));
4555 
4556 		snprintf(name, sizeof(name), "%d", idx);
4557 		oid = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(vi->txq_oid),
4558 		    OID_AUTO, name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
4559 		    "tx queue");
4560 
4561 		iqidx = vi->first_rxq + (idx % vi->nrxq);
4562 		snprintf(name, sizeof(name), "%s txq%d",
4563 		    device_get_nameunit(vi->dev), idx);
4564 		init_eq(sc, &txq->eq, EQ_ETH, vi->qsize_txq, pi->tx_chan,
4565 		    &sc->sge.rxq[iqidx].iq, name);
4566 
4567 		rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx,
4568 		    can_resume_eth_tx, M_CXGBE, &eq->eq_lock, M_WAITOK);
4569 		if (rc != 0) {
4570 			CH_ERR(vi, "failed to allocate mp_ring for txq%d: %d\n",
4571 			    idx, rc);
4572 failed:
4573 			sysctl_remove_oid(oid, 1, 1);
4574 			return (rc);
4575 		}
4576 
4577 		rc = alloc_eq(sc, eq, &vi->ctx, oid);
4578 		if (rc) {
4579 			CH_ERR(vi, "failed to allocate txq%d: %d\n", idx, rc);
4580 			mp_ring_free(txq->r);
4581 			goto failed;
4582 		}
4583 		MPASS(eq->flags & EQ_SW_ALLOCATED);
4584 		/* Can't fail after this point. */
4585 
4586 		TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq);
4587 		txq->ifp = vi->ifp;
4588 		txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK);
4589 		txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE,
4590 		    M_ZERO | M_WAITOK);
4591 
4592 		add_txq_sysctls(vi, &vi->ctx, oid, txq);
4593 	}
4594 
4595 	if (!(eq->flags & EQ_HW_ALLOCATED)) {
4596 		MPASS(eq->flags & EQ_SW_ALLOCATED);
4597 		rc = alloc_eq_hwq(sc, vi, eq);
4598 		if (rc != 0) {
4599 			CH_ERR(vi, "failed to create hw txq%d: %d\n", idx, rc);
4600 			return (rc);
4601 		}
4602 		MPASS(eq->flags & EQ_HW_ALLOCATED);
4603 		/* Can't fail after this point. */
4604 
4605 		if (idx == 0)
4606 			sc->sge.eq_base = eq->abs_id - eq->cntxt_id;
4607 		else
4608 			KASSERT(eq->cntxt_id + sc->sge.eq_base == eq->abs_id,
4609 			    ("eq_base mismatch"));
4610 		KASSERT(sc->sge.eq_base == 0 || sc->flags & IS_VF,
4611 		    ("PF with non-zero eq_base"));
4612 
4613 		txp = &txq->txp;
4614 		MPASS(nitems(txp->mb) >= sc->params.max_pkts_per_eth_tx_pkts_wr);
4615 		txq->txp.max_npkt = min(nitems(txp->mb),
4616 		    sc->params.max_pkts_per_eth_tx_pkts_wr);
4617 		if (vi->flags & TX_USES_VM_WR && !(sc->flags & IS_VF))
4618 			txq->txp.max_npkt--;
4619 
4620 		if (vi->flags & TX_USES_VM_WR)
4621 			txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
4622 			    V_TXPKT_INTF(pi->tx_chan));
4623 		else
4624 			txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
4625 			    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(sc->pf) |
4626 			    V_TXPKT_VF(vi->vin) | V_TXPKT_VF_VLD(vi->vfvld));
4627 
4628 		txq->tc_idx = -1;
4629 	}
4630 
4631 	return (0);
4632 }
4633 
4634 /*
4635  * Idempotent.
4636  */
4637 static void
4638 free_txq(struct vi_info *vi, struct sge_txq *txq)
4639 {
4640 	struct adapter *sc = vi->adapter;
4641 	struct sge_eq *eq = &txq->eq;
4642 
4643 	if (eq->flags & EQ_HW_ALLOCATED) {
4644 		MPASS(eq->flags & EQ_SW_ALLOCATED);
4645 		free_eq_hwq(sc, NULL, eq);
4646 		MPASS(!(eq->flags & EQ_HW_ALLOCATED));
4647 	}
4648 
4649 	if (eq->flags & EQ_SW_ALLOCATED) {
4650 		MPASS(!(eq->flags & EQ_HW_ALLOCATED));
4651 		sglist_free(txq->gl);
4652 		free(txq->sdesc, M_CXGBE);
4653 		mp_ring_free(txq->r);
4654 		free_eq(sc, eq);
4655 		MPASS(!(eq->flags & EQ_SW_ALLOCATED));
4656 		bzero(txq, sizeof(*txq));
4657 	}
4658 }
4659 
4660 static void
4661 add_txq_sysctls(struct vi_info *vi, struct sysctl_ctx_list *ctx,
4662     struct sysctl_oid *oid, struct sge_txq *txq)
4663 {
4664 	struct adapter *sc;
4665 	struct sysctl_oid_list *children;
4666 
4667 	if (ctx == NULL || oid == NULL)
4668 		return;
4669 
4670 	sc = vi->adapter;
4671 	children = SYSCTL_CHILDREN(oid);
4672 
4673 	mp_ring_sysctls(txq->r, ctx, children);
4674 
4675 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "tc",
4676 	    CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, vi, txq - sc->sge.txq,
4677 	    sysctl_tc, "I", "traffic class (-1 means none)");
4678 
4679 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
4680 	    &txq->txcsum, "# of times hardware assisted with checksum");
4681 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "vlan_insertion", CTLFLAG_RD,
4682 	    &txq->vlan_insertion, "# of times hardware inserted 802.1Q tag");
4683 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
4684 	    &txq->tso_wrs, "# of TSO work requests");
4685 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
4686 	    &txq->imm_wrs, "# of work requests with immediate data");
4687 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
4688 	    &txq->sgl_wrs, "# of work requests with direct SGL");
4689 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
4690 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
4691 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txpkts0_wrs", CTLFLAG_RD,
4692 	    &txq->txpkts0_wrs, "# of txpkts (type 0) work requests");
4693 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txpkts1_wrs", CTLFLAG_RD,
4694 	    &txq->txpkts1_wrs, "# of txpkts (type 1) work requests");
4695 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txpkts0_pkts", CTLFLAG_RD,
4696 	    &txq->txpkts0_pkts,
4697 	    "# of frames tx'd using type0 txpkts work requests");
4698 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txpkts1_pkts", CTLFLAG_RD,
4699 	    &txq->txpkts1_pkts,
4700 	    "# of frames tx'd using type1 txpkts work requests");
4701 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "txpkts_flush", CTLFLAG_RD,
4702 	    &txq->txpkts_flush,
4703 	    "# of times txpkts had to be flushed out by an egress-update");
4704 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "raw_wrs", CTLFLAG_RD,
4705 	    &txq->raw_wrs, "# of raw work requests (non-packets)");
4706 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "vxlan_tso_wrs", CTLFLAG_RD,
4707 	    &txq->vxlan_tso_wrs, "# of VXLAN TSO work requests");
4708 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "vxlan_txcsum", CTLFLAG_RD,
4709 	    &txq->vxlan_txcsum,
4710 	    "# of times hardware assisted with inner checksums (VXLAN)");
4711 
4712 #ifdef KERN_TLS
4713 	if (is_ktls(sc)) {
4714 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_records",
4715 		    CTLFLAG_RD, &txq->kern_tls_records,
4716 		    "# of NIC TLS records transmitted");
4717 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_short",
4718 		    CTLFLAG_RD, &txq->kern_tls_short,
4719 		    "# of short NIC TLS records transmitted");
4720 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_partial",
4721 		    CTLFLAG_RD, &txq->kern_tls_partial,
4722 		    "# of partial NIC TLS records transmitted");
4723 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_full",
4724 		    CTLFLAG_RD, &txq->kern_tls_full,
4725 		    "# of full NIC TLS records transmitted");
4726 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_octets",
4727 		    CTLFLAG_RD, &txq->kern_tls_octets,
4728 		    "# of payload octets in transmitted NIC TLS records");
4729 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_waste",
4730 		    CTLFLAG_RD, &txq->kern_tls_waste,
4731 		    "# of octets DMAd but not transmitted in NIC TLS records");
4732 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_options",
4733 		    CTLFLAG_RD, &txq->kern_tls_options,
4734 		    "# of NIC TLS options-only packets transmitted");
4735 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_header",
4736 		    CTLFLAG_RD, &txq->kern_tls_header,
4737 		    "# of NIC TLS header-only packets transmitted");
4738 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_fin",
4739 		    CTLFLAG_RD, &txq->kern_tls_fin,
4740 		    "# of NIC TLS FIN-only packets transmitted");
4741 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_fin_short",
4742 		    CTLFLAG_RD, &txq->kern_tls_fin_short,
4743 		    "# of NIC TLS padded FIN packets on short TLS records");
4744 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_cbc",
4745 		    CTLFLAG_RD, &txq->kern_tls_cbc,
4746 		    "# of NIC TLS sessions using AES-CBC");
4747 		SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "kern_tls_gcm",
4748 		    CTLFLAG_RD, &txq->kern_tls_gcm,
4749 		    "# of NIC TLS sessions using AES-GCM");
4750 	}
4751 #endif
4752 }
4753 
4754 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
4755 /*
4756  * Idempotent.
4757  */
4758 static int
4759 alloc_ofld_txq(struct vi_info *vi, struct sge_ofld_txq *ofld_txq, int idx)
4760 {
4761 	struct sysctl_oid *oid;
4762 	struct port_info *pi = vi->pi;
4763 	struct adapter *sc = vi->adapter;
4764 	struct sge_eq *eq = &ofld_txq->wrq.eq;
4765 	int rc, iqidx;
4766 	char name[16];
4767 
4768 	MPASS(idx >= 0);
4769 	MPASS(idx < vi->nofldtxq);
4770 
4771 	if (!(eq->flags & EQ_SW_ALLOCATED)) {
4772 		snprintf(name, sizeof(name), "%d", idx);
4773 		oid = SYSCTL_ADD_NODE(&vi->ctx,
4774 		    SYSCTL_CHILDREN(vi->ofld_txq_oid), OID_AUTO, name,
4775 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "offload tx queue");
4776 
4777 		snprintf(name, sizeof(name), "%s ofld_txq%d",
4778 		    device_get_nameunit(vi->dev), idx);
4779 		if (vi->nofldrxq > 0) {
4780 			iqidx = vi->first_ofld_rxq + (idx % vi->nofldrxq);
4781 			init_eq(sc, eq, EQ_OFLD, vi->qsize_txq, pi->tx_chan,
4782 			    &sc->sge.ofld_rxq[iqidx].iq, name);
4783 		} else {
4784 			iqidx = vi->first_rxq + (idx % vi->nrxq);
4785 			init_eq(sc, eq, EQ_OFLD, vi->qsize_txq, pi->tx_chan,
4786 			    &sc->sge.rxq[iqidx].iq, name);
4787 		}
4788 
4789 		rc = alloc_wrq(sc, vi, &ofld_txq->wrq, &vi->ctx, oid);
4790 		if (rc != 0) {
4791 			CH_ERR(vi, "failed to allocate ofld_txq%d: %d\n", idx,
4792 			    rc);
4793 			sysctl_remove_oid(oid, 1, 1);
4794 			return (rc);
4795 		}
4796 		MPASS(eq->flags & EQ_SW_ALLOCATED);
4797 		/* Can't fail after this point. */
4798 
4799 		ofld_txq->tx_iscsi_pdus = counter_u64_alloc(M_WAITOK);
4800 		ofld_txq->tx_iscsi_octets = counter_u64_alloc(M_WAITOK);
4801 		ofld_txq->tx_iscsi_iso_wrs = counter_u64_alloc(M_WAITOK);
4802 		ofld_txq->tx_toe_tls_records = counter_u64_alloc(M_WAITOK);
4803 		ofld_txq->tx_toe_tls_octets = counter_u64_alloc(M_WAITOK);
4804 		add_ofld_txq_sysctls(&vi->ctx, oid, ofld_txq);
4805 	}
4806 
4807 	if (!(eq->flags & EQ_HW_ALLOCATED)) {
4808 		rc = alloc_eq_hwq(sc, vi, eq);
4809 		if (rc != 0) {
4810 			CH_ERR(vi, "failed to create hw ofld_txq%d: %d\n", idx,
4811 			    rc);
4812 			return (rc);
4813 		}
4814 		MPASS(eq->flags & EQ_HW_ALLOCATED);
4815 	}
4816 
4817 	return (0);
4818 }
4819 
4820 /*
4821  * Idempotent.
4822  */
4823 static void
4824 free_ofld_txq(struct vi_info *vi, struct sge_ofld_txq *ofld_txq)
4825 {
4826 	struct adapter *sc = vi->adapter;
4827 	struct sge_eq *eq = &ofld_txq->wrq.eq;
4828 
4829 	if (eq->flags & EQ_HW_ALLOCATED) {
4830 		MPASS(eq->flags & EQ_SW_ALLOCATED);
4831 		free_eq_hwq(sc, NULL, eq);
4832 		MPASS(!(eq->flags & EQ_HW_ALLOCATED));
4833 	}
4834 
4835 	if (eq->flags & EQ_SW_ALLOCATED) {
4836 		MPASS(!(eq->flags & EQ_HW_ALLOCATED));
4837 		counter_u64_free(ofld_txq->tx_iscsi_pdus);
4838 		counter_u64_free(ofld_txq->tx_iscsi_octets);
4839 		counter_u64_free(ofld_txq->tx_iscsi_iso_wrs);
4840 		counter_u64_free(ofld_txq->tx_toe_tls_records);
4841 		counter_u64_free(ofld_txq->tx_toe_tls_octets);
4842 		free_wrq(sc, &ofld_txq->wrq);
4843 		MPASS(!(eq->flags & EQ_SW_ALLOCATED));
4844 		bzero(ofld_txq, sizeof(*ofld_txq));
4845 	}
4846 }
4847 
4848 static void
4849 add_ofld_txq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
4850     struct sge_ofld_txq *ofld_txq)
4851 {
4852 	struct sysctl_oid_list *children;
4853 
4854 	if (ctx == NULL || oid == NULL)
4855 		return;
4856 
4857 	children = SYSCTL_CHILDREN(oid);
4858 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_iscsi_pdus",
4859 	    CTLFLAG_RD, &ofld_txq->tx_iscsi_pdus,
4860 	    "# of iSCSI PDUs transmitted");
4861 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_iscsi_octets",
4862 	    CTLFLAG_RD, &ofld_txq->tx_iscsi_octets,
4863 	    "# of payload octets in transmitted iSCSI PDUs");
4864 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_iscsi_iso_wrs",
4865 	    CTLFLAG_RD, &ofld_txq->tx_iscsi_iso_wrs,
4866 	    "# of iSCSI segmentation offload work requests");
4867 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_toe_tls_records",
4868 	    CTLFLAG_RD, &ofld_txq->tx_toe_tls_records,
4869 	    "# of TOE TLS records transmitted");
4870 	SYSCTL_ADD_COUNTER_U64(ctx, children, OID_AUTO, "tx_toe_tls_octets",
4871 	    CTLFLAG_RD, &ofld_txq->tx_toe_tls_octets,
4872 	    "# of payload octets in transmitted TOE TLS records");
4873 }
4874 #endif
4875 
4876 static void
4877 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
4878 {
4879 	bus_addr_t *ba = arg;
4880 
4881 	KASSERT(nseg == 1,
4882 	    ("%s meant for single segment mappings only.", __func__));
4883 
4884 	*ba = error ? 0 : segs->ds_addr;
4885 }
4886 
4887 static inline void
4888 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
4889 {
4890 	uint32_t n, v;
4891 
4892 	n = IDXDIFF(fl->pidx >> 3, fl->dbidx, fl->sidx);
4893 	MPASS(n > 0);
4894 
4895 	wmb();
4896 	v = fl->dbval | V_PIDX(n);
4897 	if (fl->udb)
4898 		*fl->udb = htole32(v);
4899 	else
4900 		t4_write_reg(sc, sc->sge_kdoorbell_reg, v);
4901 	IDXINCR(fl->dbidx, n, fl->sidx);
4902 }
4903 
4904 /*
4905  * Fills up the freelist by allocating up to 'n' buffers.  Buffers that are
4906  * recycled do not count towards this allocation budget.
4907  *
4908  * Returns non-zero to indicate that this freelist should be added to the list
4909  * of starving freelists.
4910  */
4911 static int
4912 refill_fl(struct adapter *sc, struct sge_fl *fl, int n)
4913 {
4914 	__be64 *d;
4915 	struct fl_sdesc *sd;
4916 	uintptr_t pa;
4917 	caddr_t cl;
4918 	struct rx_buf_info *rxb;
4919 	struct cluster_metadata *clm;
4920 	uint16_t max_pidx, zidx = fl->zidx;
4921 	uint16_t hw_cidx = fl->hw_cidx;		/* stable snapshot */
4922 
4923 	FL_LOCK_ASSERT_OWNED(fl);
4924 
4925 	/*
4926 	 * We always stop at the beginning of the hardware descriptor that's just
4927 	 * before the one with the hw cidx.  This is to avoid hw pidx = hw cidx,
4928 	 * which would mean an empty freelist to the chip.
4929 	 */
4930 	max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1;
4931 	if (fl->pidx == max_pidx * 8)
4932 		return (0);
4933 
4934 	d = &fl->desc[fl->pidx];
4935 	sd = &fl->sdesc[fl->pidx];
4936 	rxb = &sc->sge.rx_buf_info[zidx];
4937 
4938 	while (n > 0) {
4939 
4940 		if (sd->cl != NULL) {
4941 
4942 			if (sd->nmbuf == 0) {
4943 				/*
4944 				 * Fast recycle without involving any atomics on
4945 				 * the cluster's metadata (if the cluster has
4946 				 * metadata).  This happens when all frames
4947 				 * received in the cluster were small enough to
4948 				 * fit within a single mbuf each.
4949 				 */
4950 				fl->cl_fast_recycled++;
4951 				goto recycled;
4952 			}
4953 
4954 			/*
4955 			 * Cluster is guaranteed to have metadata.  Clusters
4956 			 * without metadata always take the fast recycle path
4957 			 * when they're recycled.
4958 			 */
4959 			clm = cl_metadata(sd);
4960 			MPASS(clm != NULL);
4961 
4962 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4963 				fl->cl_recycled++;
4964 				counter_u64_add(extfree_rels, 1);
4965 				goto recycled;
4966 			}
4967 			sd->cl = NULL;	/* gave up my reference */
4968 		}
4969 		MPASS(sd->cl == NULL);
4970 		cl = uma_zalloc(rxb->zone, M_NOWAIT);
4971 		if (__predict_false(cl == NULL)) {
4972 			if (zidx != fl->safe_zidx) {
4973 				zidx = fl->safe_zidx;
4974 				rxb = &sc->sge.rx_buf_info[zidx];
4975 				cl = uma_zalloc(rxb->zone, M_NOWAIT);
4976 			}
4977 			if (cl == NULL)
4978 				break;
4979 		}
4980 		fl->cl_allocated++;
4981 		n--;
4982 
4983 		pa = pmap_kextract((vm_offset_t)cl);
4984 		sd->cl = cl;
4985 		sd->zidx = zidx;
4986 
4987 		if (fl->flags & FL_BUF_PACKING) {
4988 			*d = htobe64(pa | rxb->hwidx2);
4989 			sd->moff = rxb->size2;
4990 		} else {
4991 			*d = htobe64(pa | rxb->hwidx1);
4992 			sd->moff = 0;
4993 		}
4994 recycled:
4995 		sd->nmbuf = 0;
4996 		d++;
4997 		sd++;
4998 		if (__predict_false((++fl->pidx & 7) == 0)) {
4999 			uint16_t pidx = fl->pidx >> 3;
5000 
5001 			if (__predict_false(pidx == fl->sidx)) {
5002 				fl->pidx = 0;
5003 				pidx = 0;
5004 				sd = fl->sdesc;
5005 				d = fl->desc;
5006 			}
5007 			if (n < 8 || pidx == max_pidx)
5008 				break;
5009 
5010 			if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4)
5011 				ring_fl_db(sc, fl);
5012 		}
5013 	}
5014 
5015 	if ((fl->pidx >> 3) != fl->dbidx)
5016 		ring_fl_db(sc, fl);
5017 
5018 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
5019 }
5020 
5021 /*
5022  * Attempt to refill all starving freelists.
5023  */
5024 static void
5025 refill_sfl(void *arg)
5026 {
5027 	struct adapter *sc = arg;
5028 	struct sge_fl *fl, *fl_temp;
5029 
5030 	mtx_assert(&sc->sfl_lock, MA_OWNED);
5031 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
5032 		FL_LOCK(fl);
5033 		refill_fl(sc, fl, 64);
5034 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
5035 			TAILQ_REMOVE(&sc->sfl, fl, link);
5036 			fl->flags &= ~FL_STARVING;
5037 		}
5038 		FL_UNLOCK(fl);
5039 	}
5040 
5041 	if (!TAILQ_EMPTY(&sc->sfl))
5042 		callout_schedule(&sc->sfl_callout, hz / 5);
5043 }
5044 
5045 /*
5046  * Release the driver's reference on all buffers in the given freelist.  Buffers
5047  * with kernel references cannot be freed and will prevent the driver from being
5048  * unloaded safely.
5049  */
5050 void
5051 free_fl_buffers(struct adapter *sc, struct sge_fl *fl)
5052 {
5053 	struct fl_sdesc *sd;
5054 	struct cluster_metadata *clm;
5055 	int i;
5056 
5057 	sd = fl->sdesc;
5058 	for (i = 0; i < fl->sidx * 8; i++, sd++) {
5059 		if (sd->cl == NULL)
5060 			continue;
5061 
5062 		if (sd->nmbuf == 0)
5063 			uma_zfree(sc->sge.rx_buf_info[sd->zidx].zone, sd->cl);
5064 		else if (fl->flags & FL_BUF_PACKING) {
5065 			clm = cl_metadata(sd);
5066 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
5067 				uma_zfree(sc->sge.rx_buf_info[sd->zidx].zone,
5068 				    sd->cl);
5069 				counter_u64_add(extfree_rels, 1);
5070 			}
5071 		}
5072 		sd->cl = NULL;
5073 	}
5074 
5075 	if (fl->flags & FL_BUF_RESUME) {
5076 		m_freem(fl->m0);
5077 		fl->flags &= ~FL_BUF_RESUME;
5078 	}
5079 }
5080 
5081 static inline void
5082 get_pkt_gl(struct mbuf *m, struct sglist *gl)
5083 {
5084 	int rc;
5085 
5086 	M_ASSERTPKTHDR(m);
5087 
5088 	sglist_reset(gl);
5089 	rc = sglist_append_mbuf(gl, m);
5090 	if (__predict_false(rc != 0)) {
5091 		panic("%s: mbuf %p (%d segs) was vetted earlier but now fails "
5092 		    "with %d.", __func__, m, mbuf_nsegs(m), rc);
5093 	}
5094 
5095 	KASSERT(gl->sg_nseg == mbuf_nsegs(m),
5096 	    ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m,
5097 	    mbuf_nsegs(m), gl->sg_nseg));
5098 #if 0	/* vm_wr not readily available here. */
5099 	KASSERT(gl->sg_nseg > 0 && gl->sg_nseg <= max_nsegs_allowed(m, vm_wr),
5100 	    ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__,
5101 		gl->sg_nseg, max_nsegs_allowed(m, vm_wr)));
5102 #endif
5103 }
5104 
5105 /*
5106  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
5107  */
5108 static inline u_int
5109 txpkt_len16(u_int nsegs, const u_int extra)
5110 {
5111 	u_int n;
5112 
5113 	MPASS(nsegs > 0);
5114 
5115 	nsegs--; /* first segment is part of ulptx_sgl */
5116 	n = extra + sizeof(struct fw_eth_tx_pkt_wr) +
5117 	    sizeof(struct cpl_tx_pkt_core) +
5118 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
5119 
5120 	return (howmany(n, 16));
5121 }
5122 
5123 /*
5124  * len16 for a txpkt_vm WR with a GL.  Includes the firmware work
5125  * request header.
5126  */
5127 static inline u_int
5128 txpkt_vm_len16(u_int nsegs, const u_int extra)
5129 {
5130 	u_int n;
5131 
5132 	MPASS(nsegs > 0);
5133 
5134 	nsegs--; /* first segment is part of ulptx_sgl */
5135 	n = extra + sizeof(struct fw_eth_tx_pkt_vm_wr) +
5136 	    sizeof(struct cpl_tx_pkt_core) +
5137 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
5138 
5139 	return (howmany(n, 16));
5140 }
5141 
5142 static inline void
5143 calculate_mbuf_len16(struct mbuf *m, bool vm_wr)
5144 {
5145 	const int lso = sizeof(struct cpl_tx_pkt_lso_core);
5146 	const int tnl_lso = sizeof(struct cpl_tx_tnl_lso);
5147 
5148 	if (vm_wr) {
5149 		if (needs_tso(m))
5150 			set_mbuf_len16(m, txpkt_vm_len16(mbuf_nsegs(m), lso));
5151 		else
5152 			set_mbuf_len16(m, txpkt_vm_len16(mbuf_nsegs(m), 0));
5153 		return;
5154 	}
5155 
5156 	if (needs_tso(m)) {
5157 		if (needs_vxlan_tso(m))
5158 			set_mbuf_len16(m, txpkt_len16(mbuf_nsegs(m), tnl_lso));
5159 		else
5160 			set_mbuf_len16(m, txpkt_len16(mbuf_nsegs(m), lso));
5161 	} else
5162 		set_mbuf_len16(m, txpkt_len16(mbuf_nsegs(m), 0));
5163 }
5164 
5165 /*
5166  * len16 for a txpkts type 0 WR with a GL.  Does not include the firmware work
5167  * request header.
5168  */
5169 static inline u_int
5170 txpkts0_len16(u_int nsegs)
5171 {
5172 	u_int n;
5173 
5174 	MPASS(nsegs > 0);
5175 
5176 	nsegs--; /* first segment is part of ulptx_sgl */
5177 	n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) +
5178 	    sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) +
5179 	    8 * ((3 * nsegs) / 2 + (nsegs & 1));
5180 
5181 	return (howmany(n, 16));
5182 }
5183 
5184 /*
5185  * len16 for a txpkts type 1 WR with a GL.  Does not include the firmware work
5186  * request header.
5187  */
5188 static inline u_int
5189 txpkts1_len16(void)
5190 {
5191 	u_int n;
5192 
5193 	n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl);
5194 
5195 	return (howmany(n, 16));
5196 }
5197 
5198 static inline u_int
5199 imm_payload(u_int ndesc)
5200 {
5201 	u_int n;
5202 
5203 	n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) -
5204 	    sizeof(struct cpl_tx_pkt_core);
5205 
5206 	return (n);
5207 }
5208 
5209 static inline uint64_t
5210 csum_to_ctrl(struct adapter *sc, struct mbuf *m)
5211 {
5212 	uint64_t ctrl;
5213 	int csum_type, l2hlen, l3hlen;
5214 	int x, y;
5215 	static const int csum_types[3][2] = {
5216 		{TX_CSUM_TCPIP, TX_CSUM_TCPIP6},
5217 		{TX_CSUM_UDPIP, TX_CSUM_UDPIP6},
5218 		{TX_CSUM_IP, 0}
5219 	};
5220 
5221 	M_ASSERTPKTHDR(m);
5222 
5223 	if (!needs_hwcsum(m))
5224 		return (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS);
5225 
5226 	MPASS(m->m_pkthdr.l2hlen >= ETHER_HDR_LEN);
5227 	MPASS(m->m_pkthdr.l3hlen >= sizeof(struct ip));
5228 
5229 	if (needs_vxlan_csum(m)) {
5230 		MPASS(m->m_pkthdr.l4hlen > 0);
5231 		MPASS(m->m_pkthdr.l5hlen > 0);
5232 		MPASS(m->m_pkthdr.inner_l2hlen >= ETHER_HDR_LEN);
5233 		MPASS(m->m_pkthdr.inner_l3hlen >= sizeof(struct ip));
5234 
5235 		l2hlen = m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen +
5236 		    m->m_pkthdr.l4hlen + m->m_pkthdr.l5hlen +
5237 		    m->m_pkthdr.inner_l2hlen - ETHER_HDR_LEN;
5238 		l3hlen = m->m_pkthdr.inner_l3hlen;
5239 	} else {
5240 		l2hlen = m->m_pkthdr.l2hlen - ETHER_HDR_LEN;
5241 		l3hlen = m->m_pkthdr.l3hlen;
5242 	}
5243 
5244 	ctrl = 0;
5245 	if (!needs_l3_csum(m))
5246 		ctrl |= F_TXPKT_IPCSUM_DIS;
5247 
5248 	if (m->m_pkthdr.csum_flags & (CSUM_IP_TCP | CSUM_INNER_IP_TCP |
5249 	    CSUM_IP6_TCP | CSUM_INNER_IP6_TCP))
5250 		x = 0;	/* TCP */
5251 	else if (m->m_pkthdr.csum_flags & (CSUM_IP_UDP | CSUM_INNER_IP_UDP |
5252 	    CSUM_IP6_UDP | CSUM_INNER_IP6_UDP))
5253 		x = 1;	/* UDP */
5254 	else
5255 		x = 2;
5256 
5257 	if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_IP_TCP | CSUM_IP_UDP |
5258 	    CSUM_INNER_IP | CSUM_INNER_IP_TCP | CSUM_INNER_IP_UDP))
5259 		y = 0;	/* IPv4 */
5260 	else {
5261 		MPASS(m->m_pkthdr.csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP |
5262 		    CSUM_INNER_IP6_TCP | CSUM_INNER_IP6_UDP));
5263 		y = 1;	/* IPv6 */
5264 	}
5265 	/*
5266 	 * needs_hwcsum returned true earlier so there must be some kind of
5267 	 * checksum to calculate.
5268 	 */
5269 	csum_type = csum_types[x][y];
5270 	MPASS(csum_type != 0);
5271 	if (csum_type == TX_CSUM_IP)
5272 		ctrl |= F_TXPKT_L4CSUM_DIS;
5273 	ctrl |= V_TXPKT_CSUM_TYPE(csum_type) | V_TXPKT_IPHDR_LEN(l3hlen);
5274 	if (chip_id(sc) <= CHELSIO_T5)
5275 		ctrl |= V_TXPKT_ETHHDR_LEN(l2hlen);
5276 	else
5277 		ctrl |= V_T6_TXPKT_ETHHDR_LEN(l2hlen);
5278 
5279 	return (ctrl);
5280 }
5281 
5282 static inline void *
5283 write_lso_cpl(void *cpl, struct mbuf *m0)
5284 {
5285 	struct cpl_tx_pkt_lso_core *lso;
5286 	uint32_t ctrl;
5287 
5288 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
5289 	    m0->m_pkthdr.l4hlen > 0,
5290 	    ("%s: mbuf %p needs TSO but missing header lengths",
5291 		__func__, m0));
5292 
5293 	ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) |
5294 	    F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE |
5295 	    V_LSO_ETHHDR_LEN((m0->m_pkthdr.l2hlen - ETHER_HDR_LEN) >> 2) |
5296 	    V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) |
5297 	    V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
5298 	if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
5299 		ctrl |= F_LSO_IPV6;
5300 
5301 	lso = cpl;
5302 	lso->lso_ctrl = htobe32(ctrl);
5303 	lso->ipid_ofst = htobe16(0);
5304 	lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
5305 	lso->seqno_offset = htobe32(0);
5306 	lso->len = htobe32(m0->m_pkthdr.len);
5307 
5308 	return (lso + 1);
5309 }
5310 
5311 static void *
5312 write_tnl_lso_cpl(void *cpl, struct mbuf *m0)
5313 {
5314 	struct cpl_tx_tnl_lso *tnl_lso = cpl;
5315 	uint32_t ctrl;
5316 
5317 	KASSERT(m0->m_pkthdr.inner_l2hlen > 0 &&
5318 	    m0->m_pkthdr.inner_l3hlen > 0 && m0->m_pkthdr.inner_l4hlen > 0 &&
5319 	    m0->m_pkthdr.inner_l5hlen > 0,
5320 	    ("%s: mbuf %p needs VXLAN_TSO but missing inner header lengths",
5321 		__func__, m0));
5322 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
5323 	    m0->m_pkthdr.l4hlen > 0 && m0->m_pkthdr.l5hlen > 0,
5324 	    ("%s: mbuf %p needs VXLAN_TSO but missing outer header lengths",
5325 		__func__, m0));
5326 
5327 	/* Outer headers. */
5328 	ctrl = V_CPL_TX_TNL_LSO_OPCODE(CPL_TX_TNL_LSO) |
5329 	    F_CPL_TX_TNL_LSO_FIRST | F_CPL_TX_TNL_LSO_LAST |
5330 	    V_CPL_TX_TNL_LSO_ETHHDRLENOUT(
5331 		(m0->m_pkthdr.l2hlen - ETHER_HDR_LEN) >> 2) |
5332 	    V_CPL_TX_TNL_LSO_IPHDRLENOUT(m0->m_pkthdr.l3hlen >> 2) |
5333 	    F_CPL_TX_TNL_LSO_IPLENSETOUT;
5334 	if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
5335 		ctrl |= F_CPL_TX_TNL_LSO_IPV6OUT;
5336 	else {
5337 		ctrl |= F_CPL_TX_TNL_LSO_IPHDRCHKOUT |
5338 		    F_CPL_TX_TNL_LSO_IPIDINCOUT;
5339 	}
5340 	tnl_lso->op_to_IpIdSplitOut = htobe32(ctrl);
5341 	tnl_lso->IpIdOffsetOut = 0;
5342 	tnl_lso->UdpLenSetOut_to_TnlHdrLen =
5343 		htobe16(F_CPL_TX_TNL_LSO_UDPCHKCLROUT |
5344 		    F_CPL_TX_TNL_LSO_UDPLENSETOUT |
5345 		    V_CPL_TX_TNL_LSO_TNLHDRLEN(m0->m_pkthdr.l2hlen +
5346 			m0->m_pkthdr.l3hlen + m0->m_pkthdr.l4hlen +
5347 			m0->m_pkthdr.l5hlen) |
5348 		    V_CPL_TX_TNL_LSO_TNLTYPE(TX_TNL_TYPE_VXLAN));
5349 	tnl_lso->r1 = 0;
5350 
5351 	/* Inner headers. */
5352 	ctrl = V_CPL_TX_TNL_LSO_ETHHDRLEN(
5353 	    (m0->m_pkthdr.inner_l2hlen - ETHER_HDR_LEN) >> 2) |
5354 	    V_CPL_TX_TNL_LSO_IPHDRLEN(m0->m_pkthdr.inner_l3hlen >> 2) |
5355 	    V_CPL_TX_TNL_LSO_TCPHDRLEN(m0->m_pkthdr.inner_l4hlen >> 2);
5356 	if (m0->m_pkthdr.inner_l3hlen == sizeof(struct ip6_hdr))
5357 		ctrl |= F_CPL_TX_TNL_LSO_IPV6;
5358 	tnl_lso->Flow_to_TcpHdrLen = htobe32(ctrl);
5359 	tnl_lso->IpIdOffset = 0;
5360 	tnl_lso->IpIdSplit_to_Mss =
5361 	    htobe16(V_CPL_TX_TNL_LSO_MSS(m0->m_pkthdr.tso_segsz));
5362 	tnl_lso->TCPSeqOffset = 0;
5363 	tnl_lso->EthLenOffset_Size =
5364 	    htobe32(V_CPL_TX_TNL_LSO_SIZE(m0->m_pkthdr.len));
5365 
5366 	return (tnl_lso + 1);
5367 }
5368 
5369 #define VM_TX_L2HDR_LEN	16	/* ethmacdst to vlantci */
5370 
5371 /*
5372  * Write a VM txpkt WR for this packet to the hardware descriptors, update the
5373  * software descriptor, and advance the pidx.  It is guaranteed that enough
5374  * descriptors are available.
5375  *
5376  * The return value is the # of hardware descriptors used.
5377  */
5378 static u_int
5379 write_txpkt_vm_wr(struct adapter *sc, struct sge_txq *txq, struct mbuf *m0)
5380 {
5381 	struct sge_eq *eq;
5382 	struct fw_eth_tx_pkt_vm_wr *wr;
5383 	struct tx_sdesc *txsd;
5384 	struct cpl_tx_pkt_core *cpl;
5385 	uint32_t ctrl;	/* used in many unrelated places */
5386 	uint64_t ctrl1;
5387 	int len16, ndesc, pktlen, nsegs;
5388 	caddr_t dst;
5389 
5390 	TXQ_LOCK_ASSERT_OWNED(txq);
5391 	M_ASSERTPKTHDR(m0);
5392 
5393 	len16 = mbuf_len16(m0);
5394 	nsegs = mbuf_nsegs(m0);
5395 	pktlen = m0->m_pkthdr.len;
5396 	ctrl = sizeof(struct cpl_tx_pkt_core);
5397 	if (needs_tso(m0))
5398 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
5399 	ndesc = tx_len16_to_desc(len16);
5400 
5401 	/* Firmware work request header */
5402 	eq = &txq->eq;
5403 	wr = (void *)&eq->desc[eq->pidx];
5404 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_VM_WR) |
5405 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
5406 
5407 	ctrl = V_FW_WR_LEN16(len16);
5408 	wr->equiq_to_len16 = htobe32(ctrl);
5409 	wr->r3[0] = 0;
5410 	wr->r3[1] = 0;
5411 
5412 	/*
5413 	 * Copy over ethmacdst, ethmacsrc, ethtype, and vlantci.
5414 	 * vlantci is ignored unless the ethtype is 0x8100, so it's
5415 	 * simpler to always copy it rather than making it
5416 	 * conditional.  Also, it seems that we do not have to set
5417 	 * vlantci or fake the ethtype when doing VLAN tag insertion.
5418 	 */
5419 	m_copydata(m0, 0, VM_TX_L2HDR_LEN, wr->ethmacdst);
5420 
5421 	if (needs_tso(m0)) {
5422 		cpl = write_lso_cpl(wr + 1, m0);
5423 		txq->tso_wrs++;
5424 	} else
5425 		cpl = (void *)(wr + 1);
5426 
5427 	/* Checksum offload */
5428 	ctrl1 = csum_to_ctrl(sc, m0);
5429 	if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS))
5430 		txq->txcsum++;	/* some hardware assistance provided */
5431 
5432 	/* VLAN tag insertion */
5433 	if (needs_vlan_insertion(m0)) {
5434 		ctrl1 |= F_TXPKT_VLAN_VLD |
5435 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
5436 		txq->vlan_insertion++;
5437 	}
5438 
5439 	/* CPL header */
5440 	cpl->ctrl0 = txq->cpl_ctrl0;
5441 	cpl->pack = 0;
5442 	cpl->len = htobe16(pktlen);
5443 	cpl->ctrl1 = htobe64(ctrl1);
5444 
5445 	/* SGL */
5446 	dst = (void *)(cpl + 1);
5447 
5448 	/*
5449 	 * A packet using TSO will use up an entire descriptor for the
5450 	 * firmware work request header, LSO CPL, and TX_PKT_XT CPL.
5451 	 * If this descriptor is the last descriptor in the ring, wrap
5452 	 * around to the front of the ring explicitly for the start of
5453 	 * the sgl.
5454 	 */
5455 	if (dst == (void *)&eq->desc[eq->sidx]) {
5456 		dst = (void *)&eq->desc[0];
5457 		write_gl_to_txd(txq, m0, &dst, 0);
5458 	} else
5459 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
5460 	txq->sgl_wrs++;
5461 	txq->txpkt_wrs++;
5462 
5463 	txsd = &txq->sdesc[eq->pidx];
5464 	txsd->m = m0;
5465 	txsd->desc_used = ndesc;
5466 
5467 	return (ndesc);
5468 }
5469 
5470 /*
5471  * Write a raw WR to the hardware descriptors, update the software
5472  * descriptor, and advance the pidx.  It is guaranteed that enough
5473  * descriptors are available.
5474  *
5475  * The return value is the # of hardware descriptors used.
5476  */
5477 static u_int
5478 write_raw_wr(struct sge_txq *txq, void *wr, struct mbuf *m0, u_int available)
5479 {
5480 	struct sge_eq *eq = &txq->eq;
5481 	struct tx_sdesc *txsd;
5482 	struct mbuf *m;
5483 	caddr_t dst;
5484 	int len16, ndesc;
5485 
5486 	len16 = mbuf_len16(m0);
5487 	ndesc = tx_len16_to_desc(len16);
5488 	MPASS(ndesc <= available);
5489 
5490 	dst = wr;
5491 	for (m = m0; m != NULL; m = m->m_next)
5492 		copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
5493 
5494 	txq->raw_wrs++;
5495 
5496 	txsd = &txq->sdesc[eq->pidx];
5497 	txsd->m = m0;
5498 	txsd->desc_used = ndesc;
5499 
5500 	return (ndesc);
5501 }
5502 
5503 /*
5504  * Write a txpkt WR for this packet to the hardware descriptors, update the
5505  * software descriptor, and advance the pidx.  It is guaranteed that enough
5506  * descriptors are available.
5507  *
5508  * The return value is the # of hardware descriptors used.
5509  */
5510 static u_int
5511 write_txpkt_wr(struct adapter *sc, struct sge_txq *txq, struct mbuf *m0,
5512     u_int available)
5513 {
5514 	struct sge_eq *eq;
5515 	struct fw_eth_tx_pkt_wr *wr;
5516 	struct tx_sdesc *txsd;
5517 	struct cpl_tx_pkt_core *cpl;
5518 	uint32_t ctrl;	/* used in many unrelated places */
5519 	uint64_t ctrl1;
5520 	int len16, ndesc, pktlen, nsegs;
5521 	caddr_t dst;
5522 
5523 	TXQ_LOCK_ASSERT_OWNED(txq);
5524 	M_ASSERTPKTHDR(m0);
5525 
5526 	len16 = mbuf_len16(m0);
5527 	nsegs = mbuf_nsegs(m0);
5528 	pktlen = m0->m_pkthdr.len;
5529 	ctrl = sizeof(struct cpl_tx_pkt_core);
5530 	if (needs_tso(m0)) {
5531 		if (needs_vxlan_tso(m0))
5532 			ctrl += sizeof(struct cpl_tx_tnl_lso);
5533 		else
5534 			ctrl += sizeof(struct cpl_tx_pkt_lso_core);
5535 	} else if (!(mbuf_cflags(m0) & MC_NOMAP) && pktlen <= imm_payload(2) &&
5536 	    available >= 2) {
5537 		/* Immediate data.  Recalculate len16 and set nsegs to 0. */
5538 		ctrl += pktlen;
5539 		len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) +
5540 		    sizeof(struct cpl_tx_pkt_core) + pktlen, 16);
5541 		nsegs = 0;
5542 	}
5543 	ndesc = tx_len16_to_desc(len16);
5544 	MPASS(ndesc <= available);
5545 
5546 	/* Firmware work request header */
5547 	eq = &txq->eq;
5548 	wr = (void *)&eq->desc[eq->pidx];
5549 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
5550 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
5551 
5552 	ctrl = V_FW_WR_LEN16(len16);
5553 	wr->equiq_to_len16 = htobe32(ctrl);
5554 	wr->r3 = 0;
5555 
5556 	if (needs_tso(m0)) {
5557 		if (needs_vxlan_tso(m0)) {
5558 			cpl = write_tnl_lso_cpl(wr + 1, m0);
5559 			txq->vxlan_tso_wrs++;
5560 		} else {
5561 			cpl = write_lso_cpl(wr + 1, m0);
5562 			txq->tso_wrs++;
5563 		}
5564 	} else
5565 		cpl = (void *)(wr + 1);
5566 
5567 	/* Checksum offload */
5568 	ctrl1 = csum_to_ctrl(sc, m0);
5569 	if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS)) {
5570 		/* some hardware assistance provided */
5571 		if (needs_vxlan_csum(m0))
5572 			txq->vxlan_txcsum++;
5573 		else
5574 			txq->txcsum++;
5575 	}
5576 
5577 	/* VLAN tag insertion */
5578 	if (needs_vlan_insertion(m0)) {
5579 		ctrl1 |= F_TXPKT_VLAN_VLD |
5580 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
5581 		txq->vlan_insertion++;
5582 	}
5583 
5584 	/* CPL header */
5585 	cpl->ctrl0 = txq->cpl_ctrl0;
5586 	cpl->pack = 0;
5587 	cpl->len = htobe16(pktlen);
5588 	cpl->ctrl1 = htobe64(ctrl1);
5589 
5590 	/* SGL */
5591 	dst = (void *)(cpl + 1);
5592 	if (__predict_false((uintptr_t)dst == (uintptr_t)&eq->desc[eq->sidx]))
5593 		dst = (caddr_t)&eq->desc[0];
5594 	if (nsegs > 0) {
5595 
5596 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
5597 		txq->sgl_wrs++;
5598 	} else {
5599 		struct mbuf *m;
5600 
5601 		for (m = m0; m != NULL; m = m->m_next) {
5602 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
5603 #ifdef INVARIANTS
5604 			pktlen -= m->m_len;
5605 #endif
5606 		}
5607 #ifdef INVARIANTS
5608 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
5609 #endif
5610 		txq->imm_wrs++;
5611 	}
5612 
5613 	txq->txpkt_wrs++;
5614 
5615 	txsd = &txq->sdesc[eq->pidx];
5616 	txsd->m = m0;
5617 	txsd->desc_used = ndesc;
5618 
5619 	return (ndesc);
5620 }
5621 
5622 static inline bool
5623 cmp_l2hdr(struct txpkts *txp, struct mbuf *m)
5624 {
5625 	int len;
5626 
5627 	MPASS(txp->npkt > 0);
5628 	MPASS(m->m_len >= VM_TX_L2HDR_LEN);
5629 
5630 	if (txp->ethtype == be16toh(ETHERTYPE_VLAN))
5631 		len = VM_TX_L2HDR_LEN;
5632 	else
5633 		len = sizeof(struct ether_header);
5634 
5635 	return (memcmp(m->m_data, &txp->ethmacdst[0], len) != 0);
5636 }
5637 
5638 static inline void
5639 save_l2hdr(struct txpkts *txp, struct mbuf *m)
5640 {
5641 	MPASS(m->m_len >= VM_TX_L2HDR_LEN);
5642 
5643 	memcpy(&txp->ethmacdst[0], mtod(m, const void *), VM_TX_L2HDR_LEN);
5644 }
5645 
5646 static int
5647 add_to_txpkts_vf(struct adapter *sc, struct sge_txq *txq, struct mbuf *m,
5648     int avail, bool *send)
5649 {
5650 	struct txpkts *txp = &txq->txp;
5651 
5652 	/* Cannot have TSO and coalesce at the same time. */
5653 	if (cannot_use_txpkts(m)) {
5654 cannot_coalesce:
5655 		*send = txp->npkt > 0;
5656 		return (EINVAL);
5657 	}
5658 
5659 	/* VF allows coalescing of type 1 (1 GL) only */
5660 	if (mbuf_nsegs(m) > 1)
5661 		goto cannot_coalesce;
5662 
5663 	*send = false;
5664 	if (txp->npkt > 0) {
5665 		MPASS(tx_len16_to_desc(txp->len16) <= avail);
5666 		MPASS(txp->npkt < txp->max_npkt);
5667 		MPASS(txp->wr_type == 1);	/* VF supports type 1 only */
5668 
5669 		if (tx_len16_to_desc(txp->len16 + txpkts1_len16()) > avail) {
5670 retry_after_send:
5671 			*send = true;
5672 			return (EAGAIN);
5673 		}
5674 		if (m->m_pkthdr.len + txp->plen > 65535)
5675 			goto retry_after_send;
5676 		if (cmp_l2hdr(txp, m))
5677 			goto retry_after_send;
5678 
5679 		txp->len16 += txpkts1_len16();
5680 		txp->plen += m->m_pkthdr.len;
5681 		txp->mb[txp->npkt++] = m;
5682 		if (txp->npkt == txp->max_npkt)
5683 			*send = true;
5684 	} else {
5685 		txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_vm_wr), 16) +
5686 		    txpkts1_len16();
5687 		if (tx_len16_to_desc(txp->len16) > avail)
5688 			goto cannot_coalesce;
5689 		txp->npkt = 1;
5690 		txp->wr_type = 1;
5691 		txp->plen = m->m_pkthdr.len;
5692 		txp->mb[0] = m;
5693 		save_l2hdr(txp, m);
5694 	}
5695 	return (0);
5696 }
5697 
5698 static int
5699 add_to_txpkts_pf(struct adapter *sc, struct sge_txq *txq, struct mbuf *m,
5700     int avail, bool *send)
5701 {
5702 	struct txpkts *txp = &txq->txp;
5703 	int nsegs;
5704 
5705 	MPASS(!(sc->flags & IS_VF));
5706 
5707 	/* Cannot have TSO and coalesce at the same time. */
5708 	if (cannot_use_txpkts(m)) {
5709 cannot_coalesce:
5710 		*send = txp->npkt > 0;
5711 		return (EINVAL);
5712 	}
5713 
5714 	*send = false;
5715 	nsegs = mbuf_nsegs(m);
5716 	if (txp->npkt == 0) {
5717 		if (m->m_pkthdr.len > 65535)
5718 			goto cannot_coalesce;
5719 		if (nsegs > 1) {
5720 			txp->wr_type = 0;
5721 			txp->len16 =
5722 			    howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) +
5723 			    txpkts0_len16(nsegs);
5724 		} else {
5725 			txp->wr_type = 1;
5726 			txp->len16 =
5727 			    howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) +
5728 			    txpkts1_len16();
5729 		}
5730 		if (tx_len16_to_desc(txp->len16) > avail)
5731 			goto cannot_coalesce;
5732 		txp->npkt = 1;
5733 		txp->plen = m->m_pkthdr.len;
5734 		txp->mb[0] = m;
5735 	} else {
5736 		MPASS(tx_len16_to_desc(txp->len16) <= avail);
5737 		MPASS(txp->npkt < txp->max_npkt);
5738 
5739 		if (m->m_pkthdr.len + txp->plen > 65535) {
5740 retry_after_send:
5741 			*send = true;
5742 			return (EAGAIN);
5743 		}
5744 
5745 		MPASS(txp->wr_type == 0 || txp->wr_type == 1);
5746 		if (txp->wr_type == 0) {
5747 			if (tx_len16_to_desc(txp->len16 +
5748 			    txpkts0_len16(nsegs)) > min(avail, SGE_MAX_WR_NDESC))
5749 				goto retry_after_send;
5750 			txp->len16 += txpkts0_len16(nsegs);
5751 		} else {
5752 			if (nsegs != 1)
5753 				goto retry_after_send;
5754 			if (tx_len16_to_desc(txp->len16 + txpkts1_len16()) >
5755 			    avail)
5756 				goto retry_after_send;
5757 			txp->len16 += txpkts1_len16();
5758 		}
5759 
5760 		txp->plen += m->m_pkthdr.len;
5761 		txp->mb[txp->npkt++] = m;
5762 		if (txp->npkt == txp->max_npkt)
5763 			*send = true;
5764 	}
5765 	return (0);
5766 }
5767 
5768 /*
5769  * Write a txpkts WR for the packets in txp to the hardware descriptors, update
5770  * the software descriptor, and advance the pidx.  It is guaranteed that enough
5771  * descriptors are available.
5772  *
5773  * The return value is the # of hardware descriptors used.
5774  */
5775 static u_int
5776 write_txpkts_wr(struct adapter *sc, struct sge_txq *txq)
5777 {
5778 	const struct txpkts *txp = &txq->txp;
5779 	struct sge_eq *eq = &txq->eq;
5780 	struct fw_eth_tx_pkts_wr *wr;
5781 	struct tx_sdesc *txsd;
5782 	struct cpl_tx_pkt_core *cpl;
5783 	uint64_t ctrl1;
5784 	int ndesc, i, checkwrap;
5785 	struct mbuf *m, *last;
5786 	void *flitp;
5787 
5788 	TXQ_LOCK_ASSERT_OWNED(txq);
5789 	MPASS(txp->npkt > 0);
5790 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
5791 
5792 	wr = (void *)&eq->desc[eq->pidx];
5793 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
5794 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(txp->len16));
5795 	wr->plen = htobe16(txp->plen);
5796 	wr->npkt = txp->npkt;
5797 	wr->r3 = 0;
5798 	wr->type = txp->wr_type;
5799 	flitp = wr + 1;
5800 
5801 	/*
5802 	 * At this point we are 16B into a hardware descriptor.  If checkwrap is
5803 	 * set then we know the WR is going to wrap around somewhere.  We'll
5804 	 * check for that at appropriate points.
5805 	 */
5806 	ndesc = tx_len16_to_desc(txp->len16);
5807 	last = NULL;
5808 	checkwrap = eq->sidx - ndesc < eq->pidx;
5809 	for (i = 0; i < txp->npkt; i++) {
5810 		m = txp->mb[i];
5811 		if (txp->wr_type == 0) {
5812 			struct ulp_txpkt *ulpmc;
5813 			struct ulptx_idata *ulpsc;
5814 
5815 			/* ULP master command */
5816 			ulpmc = flitp;
5817 			ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
5818 			    V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid));
5819 			ulpmc->len = htobe32(txpkts0_len16(mbuf_nsegs(m)));
5820 
5821 			/* ULP subcommand */
5822 			ulpsc = (void *)(ulpmc + 1);
5823 			ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
5824 			    F_ULP_TX_SC_MORE);
5825 			ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
5826 
5827 			cpl = (void *)(ulpsc + 1);
5828 			if (checkwrap &&
5829 			    (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx])
5830 				cpl = (void *)&eq->desc[0];
5831 		} else {
5832 			cpl = flitp;
5833 		}
5834 
5835 		/* Checksum offload */
5836 		ctrl1 = csum_to_ctrl(sc, m);
5837 		if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS)) {
5838 			/* some hardware assistance provided */
5839 			if (needs_vxlan_csum(m))
5840 				txq->vxlan_txcsum++;
5841 			else
5842 				txq->txcsum++;
5843 		}
5844 
5845 		/* VLAN tag insertion */
5846 		if (needs_vlan_insertion(m)) {
5847 			ctrl1 |= F_TXPKT_VLAN_VLD |
5848 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
5849 			txq->vlan_insertion++;
5850 		}
5851 
5852 		/* CPL header */
5853 		cpl->ctrl0 = txq->cpl_ctrl0;
5854 		cpl->pack = 0;
5855 		cpl->len = htobe16(m->m_pkthdr.len);
5856 		cpl->ctrl1 = htobe64(ctrl1);
5857 
5858 		flitp = cpl + 1;
5859 		if (checkwrap &&
5860 		    (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
5861 			flitp = (void *)&eq->desc[0];
5862 
5863 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap);
5864 
5865 		if (last != NULL)
5866 			last->m_nextpkt = m;
5867 		last = m;
5868 	}
5869 
5870 	txq->sgl_wrs++;
5871 	if (txp->wr_type == 0) {
5872 		txq->txpkts0_pkts += txp->npkt;
5873 		txq->txpkts0_wrs++;
5874 	} else {
5875 		txq->txpkts1_pkts += txp->npkt;
5876 		txq->txpkts1_wrs++;
5877 	}
5878 
5879 	txsd = &txq->sdesc[eq->pidx];
5880 	txsd->m = txp->mb[0];
5881 	txsd->desc_used = ndesc;
5882 
5883 	return (ndesc);
5884 }
5885 
5886 static u_int
5887 write_txpkts_vm_wr(struct adapter *sc, struct sge_txq *txq)
5888 {
5889 	const struct txpkts *txp = &txq->txp;
5890 	struct sge_eq *eq = &txq->eq;
5891 	struct fw_eth_tx_pkts_vm_wr *wr;
5892 	struct tx_sdesc *txsd;
5893 	struct cpl_tx_pkt_core *cpl;
5894 	uint64_t ctrl1;
5895 	int ndesc, i;
5896 	struct mbuf *m, *last;
5897 	void *flitp;
5898 
5899 	TXQ_LOCK_ASSERT_OWNED(txq);
5900 	MPASS(txp->npkt > 0);
5901 	MPASS(txp->wr_type == 1);	/* VF supports type 1 only */
5902 	MPASS(txp->mb[0] != NULL);
5903 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
5904 
5905 	wr = (void *)&eq->desc[eq->pidx];
5906 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_VM_WR));
5907 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(txp->len16));
5908 	wr->r3 = 0;
5909 	wr->plen = htobe16(txp->plen);
5910 	wr->npkt = txp->npkt;
5911 	wr->r4 = 0;
5912 	memcpy(&wr->ethmacdst[0], &txp->ethmacdst[0], 16);
5913 	flitp = wr + 1;
5914 
5915 	/*
5916 	 * At this point we are 32B into a hardware descriptor.  Each mbuf in
5917 	 * the WR will take 32B so we check for the end of the descriptor ring
5918 	 * before writing odd mbufs (mb[1], 3, 5, ..)
5919 	 */
5920 	ndesc = tx_len16_to_desc(txp->len16);
5921 	last = NULL;
5922 	for (i = 0; i < txp->npkt; i++) {
5923 		m = txp->mb[i];
5924 		if (i & 1 && (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
5925 			flitp = &eq->desc[0];
5926 		cpl = flitp;
5927 
5928 		/* Checksum offload */
5929 		ctrl1 = csum_to_ctrl(sc, m);
5930 		if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS))
5931 			txq->txcsum++;	/* some hardware assistance provided */
5932 
5933 		/* VLAN tag insertion */
5934 		if (needs_vlan_insertion(m)) {
5935 			ctrl1 |= F_TXPKT_VLAN_VLD |
5936 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
5937 			txq->vlan_insertion++;
5938 		}
5939 
5940 		/* CPL header */
5941 		cpl->ctrl0 = txq->cpl_ctrl0;
5942 		cpl->pack = 0;
5943 		cpl->len = htobe16(m->m_pkthdr.len);
5944 		cpl->ctrl1 = htobe64(ctrl1);
5945 
5946 		flitp = cpl + 1;
5947 		MPASS(mbuf_nsegs(m) == 1);
5948 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), 0);
5949 
5950 		if (last != NULL)
5951 			last->m_nextpkt = m;
5952 		last = m;
5953 	}
5954 
5955 	txq->sgl_wrs++;
5956 	txq->txpkts1_pkts += txp->npkt;
5957 	txq->txpkts1_wrs++;
5958 
5959 	txsd = &txq->sdesc[eq->pidx];
5960 	txsd->m = txp->mb[0];
5961 	txsd->desc_used = ndesc;
5962 
5963 	return (ndesc);
5964 }
5965 
5966 /*
5967  * If the SGL ends on an address that is not 16 byte aligned, this function will
5968  * add a 0 filled flit at the end.
5969  */
5970 static void
5971 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap)
5972 {
5973 	struct sge_eq *eq = &txq->eq;
5974 	struct sglist *gl = txq->gl;
5975 	struct sglist_seg *seg;
5976 	__be64 *flitp, *wrap;
5977 	struct ulptx_sgl *usgl;
5978 	int i, nflits, nsegs;
5979 
5980 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
5981 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
5982 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
5983 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
5984 
5985 	get_pkt_gl(m, gl);
5986 	nsegs = gl->sg_nseg;
5987 	MPASS(nsegs > 0);
5988 
5989 	nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
5990 	flitp = (__be64 *)(*to);
5991 	wrap = (__be64 *)(&eq->desc[eq->sidx]);
5992 	seg = &gl->sg_segs[0];
5993 	usgl = (void *)flitp;
5994 
5995 	/*
5996 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
5997 	 * ring, so we're at least 16 bytes away from the status page.  There is
5998 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
5999 	 */
6000 
6001 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
6002 	    V_ULPTX_NSGE(nsegs));
6003 	usgl->len0 = htobe32(seg->ss_len);
6004 	usgl->addr0 = htobe64(seg->ss_paddr);
6005 	seg++;
6006 
6007 	if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) {
6008 
6009 		/* Won't wrap around at all */
6010 
6011 		for (i = 0; i < nsegs - 1; i++, seg++) {
6012 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
6013 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
6014 		}
6015 		if (i & 1)
6016 			usgl->sge[i / 2].len[1] = htobe32(0);
6017 		flitp += nflits;
6018 	} else {
6019 
6020 		/* Will wrap somewhere in the rest of the SGL */
6021 
6022 		/* 2 flits already written, write the rest flit by flit */
6023 		flitp = (void *)(usgl + 1);
6024 		for (i = 0; i < nflits - 2; i++) {
6025 			if (flitp == wrap)
6026 				flitp = (void *)eq->desc;
6027 			*flitp++ = get_flit(seg, nsegs - 1, i);
6028 		}
6029 	}
6030 
6031 	if (nflits & 1) {
6032 		MPASS(((uintptr_t)flitp) & 0xf);
6033 		*flitp++ = 0;
6034 	}
6035 
6036 	MPASS((((uintptr_t)flitp) & 0xf) == 0);
6037 	if (__predict_false(flitp == wrap))
6038 		*to = (void *)eq->desc;
6039 	else
6040 		*to = (void *)flitp;
6041 }
6042 
6043 static inline void
6044 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
6045 {
6046 
6047 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
6048 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
6049 
6050 	if (__predict_true((uintptr_t)(*to) + len <=
6051 	    (uintptr_t)&eq->desc[eq->sidx])) {
6052 		bcopy(from, *to, len);
6053 		(*to) += len;
6054 	} else {
6055 		int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
6056 
6057 		bcopy(from, *to, portion);
6058 		from += portion;
6059 		portion = len - portion;	/* remaining */
6060 		bcopy(from, (void *)eq->desc, portion);
6061 		(*to) = (caddr_t)eq->desc + portion;
6062 	}
6063 }
6064 
6065 static inline void
6066 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n)
6067 {
6068 	u_int db;
6069 
6070 	MPASS(n > 0);
6071 
6072 	db = eq->doorbells;
6073 	if (n > 1)
6074 		clrbit(&db, DOORBELL_WCWR);
6075 	wmb();
6076 
6077 	switch (ffs(db) - 1) {
6078 	case DOORBELL_UDB:
6079 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
6080 		break;
6081 
6082 	case DOORBELL_WCWR: {
6083 		volatile uint64_t *dst, *src;
6084 		int i;
6085 
6086 		/*
6087 		 * Queues whose 128B doorbell segment fits in the page do not
6088 		 * use relative qid (udb_qid is always 0).  Only queues with
6089 		 * doorbell segments can do WCWR.
6090 		 */
6091 		KASSERT(eq->udb_qid == 0 && n == 1,
6092 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
6093 		    __func__, eq->doorbells, n, eq->dbidx, eq));
6094 
6095 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
6096 		    UDBS_DB_OFFSET);
6097 		i = eq->dbidx;
6098 		src = (void *)&eq->desc[i];
6099 		while (src != (void *)&eq->desc[i + 1])
6100 			*dst++ = *src++;
6101 		wmb();
6102 		break;
6103 	}
6104 
6105 	case DOORBELL_UDBWC:
6106 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
6107 		wmb();
6108 		break;
6109 
6110 	case DOORBELL_KDB:
6111 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
6112 		    V_QID(eq->cntxt_id) | V_PIDX(n));
6113 		break;
6114 	}
6115 
6116 	IDXINCR(eq->dbidx, n, eq->sidx);
6117 }
6118 
6119 static inline u_int
6120 reclaimable_tx_desc(struct sge_eq *eq)
6121 {
6122 	uint16_t hw_cidx;
6123 
6124 	hw_cidx = read_hw_cidx(eq);
6125 	return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx));
6126 }
6127 
6128 static inline u_int
6129 total_available_tx_desc(struct sge_eq *eq)
6130 {
6131 	uint16_t hw_cidx, pidx;
6132 
6133 	hw_cidx = read_hw_cidx(eq);
6134 	pidx = eq->pidx;
6135 
6136 	if (pidx == hw_cidx)
6137 		return (eq->sidx - 1);
6138 	else
6139 		return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1);
6140 }
6141 
6142 static inline uint16_t
6143 read_hw_cidx(struct sge_eq *eq)
6144 {
6145 	struct sge_qstat *spg = (void *)&eq->desc[eq->sidx];
6146 	uint16_t cidx = spg->cidx;	/* stable snapshot */
6147 
6148 	return (be16toh(cidx));
6149 }
6150 
6151 /*
6152  * Reclaim 'n' descriptors approximately.
6153  */
6154 static u_int
6155 reclaim_tx_descs(struct sge_txq *txq, u_int n)
6156 {
6157 	struct tx_sdesc *txsd;
6158 	struct sge_eq *eq = &txq->eq;
6159 	u_int can_reclaim, reclaimed;
6160 
6161 	TXQ_LOCK_ASSERT_OWNED(txq);
6162 	MPASS(n > 0);
6163 
6164 	reclaimed = 0;
6165 	can_reclaim = reclaimable_tx_desc(eq);
6166 	while (can_reclaim && reclaimed < n) {
6167 		int ndesc;
6168 		struct mbuf *m, *nextpkt;
6169 
6170 		txsd = &txq->sdesc[eq->cidx];
6171 		ndesc = txsd->desc_used;
6172 
6173 		/* Firmware doesn't return "partial" credits. */
6174 		KASSERT(can_reclaim >= ndesc,
6175 		    ("%s: unexpected number of credits: %d, %d",
6176 		    __func__, can_reclaim, ndesc));
6177 		KASSERT(ndesc != 0,
6178 		    ("%s: descriptor with no credits: cidx %d",
6179 		    __func__, eq->cidx));
6180 
6181 		for (m = txsd->m; m != NULL; m = nextpkt) {
6182 			nextpkt = m->m_nextpkt;
6183 			m->m_nextpkt = NULL;
6184 			m_freem(m);
6185 		}
6186 		reclaimed += ndesc;
6187 		can_reclaim -= ndesc;
6188 		IDXINCR(eq->cidx, ndesc, eq->sidx);
6189 	}
6190 
6191 	return (reclaimed);
6192 }
6193 
6194 static void
6195 tx_reclaim(void *arg, int n)
6196 {
6197 	struct sge_txq *txq = arg;
6198 	struct sge_eq *eq = &txq->eq;
6199 
6200 	do {
6201 		if (TXQ_TRYLOCK(txq) == 0)
6202 			break;
6203 		n = reclaim_tx_descs(txq, 32);
6204 		if (eq->cidx == eq->pidx)
6205 			eq->equeqidx = eq->pidx;
6206 		TXQ_UNLOCK(txq);
6207 	} while (n > 0);
6208 }
6209 
6210 static __be64
6211 get_flit(struct sglist_seg *segs, int nsegs, int idx)
6212 {
6213 	int i = (idx / 3) * 2;
6214 
6215 	switch (idx % 3) {
6216 	case 0: {
6217 		uint64_t rc;
6218 
6219 		rc = (uint64_t)segs[i].ss_len << 32;
6220 		if (i + 1 < nsegs)
6221 			rc |= (uint64_t)(segs[i + 1].ss_len);
6222 
6223 		return (htobe64(rc));
6224 	}
6225 	case 1:
6226 		return (htobe64(segs[i].ss_paddr));
6227 	case 2:
6228 		return (htobe64(segs[i + 1].ss_paddr));
6229 	}
6230 
6231 	return (0);
6232 }
6233 
6234 static int
6235 find_refill_source(struct adapter *sc, int maxp, bool packing)
6236 {
6237 	int i, zidx = -1;
6238 	struct rx_buf_info *rxb = &sc->sge.rx_buf_info[0];
6239 
6240 	if (packing) {
6241 		for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) {
6242 			if (rxb->hwidx2 == -1)
6243 				continue;
6244 			if (rxb->size1 < PAGE_SIZE &&
6245 			    rxb->size1 < largest_rx_cluster)
6246 				continue;
6247 			if (rxb->size1 > largest_rx_cluster)
6248 				break;
6249 			MPASS(rxb->size1 - rxb->size2 >= CL_METADATA_SIZE);
6250 			if (rxb->size2 >= maxp)
6251 				return (i);
6252 			zidx = i;
6253 		}
6254 	} else {
6255 		for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) {
6256 			if (rxb->hwidx1 == -1)
6257 				continue;
6258 			if (rxb->size1 > largest_rx_cluster)
6259 				break;
6260 			if (rxb->size1 >= maxp)
6261 				return (i);
6262 			zidx = i;
6263 		}
6264 	}
6265 
6266 	return (zidx);
6267 }
6268 
6269 static void
6270 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
6271 {
6272 	mtx_lock(&sc->sfl_lock);
6273 	FL_LOCK(fl);
6274 	if ((fl->flags & FL_DOOMED) == 0) {
6275 		fl->flags |= FL_STARVING;
6276 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
6277 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
6278 	}
6279 	FL_UNLOCK(fl);
6280 	mtx_unlock(&sc->sfl_lock);
6281 }
6282 
6283 static void
6284 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq)
6285 {
6286 	struct sge_wrq *wrq = (void *)eq;
6287 
6288 	atomic_readandclear_int(&eq->equiq);
6289 	taskqueue_enqueue(sc->tq[eq->tx_chan], &wrq->wrq_tx_task);
6290 }
6291 
6292 static void
6293 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq)
6294 {
6295 	struct sge_txq *txq = (void *)eq;
6296 
6297 	MPASS(eq->type == EQ_ETH);
6298 
6299 	atomic_readandclear_int(&eq->equiq);
6300 	if (mp_ring_is_idle(txq->r))
6301 		taskqueue_enqueue(sc->tq[eq->tx_chan], &txq->tx_reclaim_task);
6302 	else
6303 		mp_ring_check_drainage(txq->r, 64);
6304 }
6305 
6306 static int
6307 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
6308     struct mbuf *m)
6309 {
6310 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
6311 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
6312 	struct adapter *sc = iq->adapter;
6313 	struct sge *s = &sc->sge;
6314 	struct sge_eq *eq;
6315 	static void (*h[])(struct adapter *, struct sge_eq *) = {NULL,
6316 		&handle_wrq_egr_update, &handle_eth_egr_update,
6317 		&handle_wrq_egr_update};
6318 
6319 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
6320 	    rss->opcode));
6321 
6322 	eq = s->eqmap[qid - s->eq_start - s->eq_base];
6323 	(*h[eq->type])(sc, eq);
6324 
6325 	return (0);
6326 }
6327 
6328 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
6329 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
6330     offsetof(struct cpl_fw6_msg, data));
6331 
6332 static int
6333 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
6334 {
6335 	struct adapter *sc = iq->adapter;
6336 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
6337 
6338 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
6339 	    rss->opcode));
6340 
6341 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
6342 		const struct rss_header *rss2;
6343 
6344 		rss2 = (const struct rss_header *)&cpl->data[0];
6345 		return (t4_cpl_handler[rss2->opcode](iq, rss2, m));
6346 	}
6347 
6348 	return (t4_fw_msg_handler[cpl->type](sc, &cpl->data[0]));
6349 }
6350 
6351 /**
6352  *	t4_handle_wrerr_rpl - process a FW work request error message
6353  *	@adap: the adapter
6354  *	@rpl: start of the FW message
6355  */
6356 static int
6357 t4_handle_wrerr_rpl(struct adapter *adap, const __be64 *rpl)
6358 {
6359 	u8 opcode = *(const u8 *)rpl;
6360 	const struct fw_error_cmd *e = (const void *)rpl;
6361 	unsigned int i;
6362 
6363 	if (opcode != FW_ERROR_CMD) {
6364 		log(LOG_ERR,
6365 		    "%s: Received WRERR_RPL message with opcode %#x\n",
6366 		    device_get_nameunit(adap->dev), opcode);
6367 		return (EINVAL);
6368 	}
6369 	log(LOG_ERR, "%s: FW_ERROR (%s) ", device_get_nameunit(adap->dev),
6370 	    G_FW_ERROR_CMD_FATAL(be32toh(e->op_to_type)) ? "fatal" :
6371 	    "non-fatal");
6372 	switch (G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type))) {
6373 	case FW_ERROR_TYPE_EXCEPTION:
6374 		log(LOG_ERR, "exception info:\n");
6375 		for (i = 0; i < nitems(e->u.exception.info); i++)
6376 			log(LOG_ERR, "%s%08x", i == 0 ? "\t" : " ",
6377 			    be32toh(e->u.exception.info[i]));
6378 		log(LOG_ERR, "\n");
6379 		break;
6380 	case FW_ERROR_TYPE_HWMODULE:
6381 		log(LOG_ERR, "HW module regaddr %08x regval %08x\n",
6382 		    be32toh(e->u.hwmodule.regaddr),
6383 		    be32toh(e->u.hwmodule.regval));
6384 		break;
6385 	case FW_ERROR_TYPE_WR:
6386 		log(LOG_ERR, "WR cidx %d PF %d VF %d eqid %d hdr:\n",
6387 		    be16toh(e->u.wr.cidx),
6388 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.wr.pfn_vfn)),
6389 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.wr.pfn_vfn)),
6390 		    be32toh(e->u.wr.eqid));
6391 		for (i = 0; i < nitems(e->u.wr.wrhdr); i++)
6392 			log(LOG_ERR, "%s%02x", i == 0 ? "\t" : " ",
6393 			    e->u.wr.wrhdr[i]);
6394 		log(LOG_ERR, "\n");
6395 		break;
6396 	case FW_ERROR_TYPE_ACL:
6397 		log(LOG_ERR, "ACL cidx %d PF %d VF %d eqid %d %s",
6398 		    be16toh(e->u.acl.cidx),
6399 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.acl.pfn_vfn)),
6400 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.acl.pfn_vfn)),
6401 		    be32toh(e->u.acl.eqid),
6402 		    G_FW_ERROR_CMD_MV(be16toh(e->u.acl.mv_pkd)) ? "vlanid" :
6403 		    "MAC");
6404 		for (i = 0; i < nitems(e->u.acl.val); i++)
6405 			log(LOG_ERR, " %02x", e->u.acl.val[i]);
6406 		log(LOG_ERR, "\n");
6407 		break;
6408 	default:
6409 		log(LOG_ERR, "type %#x\n",
6410 		    G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type)));
6411 		return (EINVAL);
6412 	}
6413 	return (0);
6414 }
6415 
6416 static inline bool
6417 bufidx_used(struct adapter *sc, int idx)
6418 {
6419 	struct rx_buf_info *rxb = &sc->sge.rx_buf_info[0];
6420 	int i;
6421 
6422 	for (i = 0; i < SW_ZONE_SIZES; i++, rxb++) {
6423 		if (rxb->size1 > largest_rx_cluster)
6424 			continue;
6425 		if (rxb->hwidx1 == idx || rxb->hwidx2 == idx)
6426 			return (true);
6427 	}
6428 
6429 	return (false);
6430 }
6431 
6432 static int
6433 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
6434 {
6435 	struct adapter *sc = arg1;
6436 	struct sge_params *sp = &sc->params.sge;
6437 	int i, rc;
6438 	struct sbuf sb;
6439 	char c;
6440 
6441 	sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND);
6442 	for (i = 0; i < SGE_FLBUF_SIZES; i++) {
6443 		if (bufidx_used(sc, i))
6444 			c = '*';
6445 		else
6446 			c = '\0';
6447 
6448 		sbuf_printf(&sb, "%u%c ", sp->sge_fl_buffer_size[i], c);
6449 	}
6450 	sbuf_trim(&sb);
6451 	sbuf_finish(&sb);
6452 	rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
6453 	sbuf_delete(&sb);
6454 	return (rc);
6455 }
6456 
6457 #ifdef RATELIMIT
6458 #if defined(INET) || defined(INET6)
6459 /*
6460  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
6461  */
6462 static inline u_int
6463 txpkt_eo_len16(u_int nsegs, u_int immhdrs, u_int tso)
6464 {
6465 	u_int n;
6466 
6467 	MPASS(immhdrs > 0);
6468 
6469 	n = roundup2(sizeof(struct fw_eth_tx_eo_wr) +
6470 	    sizeof(struct cpl_tx_pkt_core) + immhdrs, 16);
6471 	if (__predict_false(nsegs == 0))
6472 		goto done;
6473 
6474 	nsegs--; /* first segment is part of ulptx_sgl */
6475 	n += sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
6476 	if (tso)
6477 		n += sizeof(struct cpl_tx_pkt_lso_core);
6478 
6479 done:
6480 	return (howmany(n, 16));
6481 }
6482 #endif
6483 
6484 #define ETID_FLOWC_NPARAMS 6
6485 #define ETID_FLOWC_LEN (roundup2((sizeof(struct fw_flowc_wr) + \
6486     ETID_FLOWC_NPARAMS * sizeof(struct fw_flowc_mnemval)), 16))
6487 #define ETID_FLOWC_LEN16 (howmany(ETID_FLOWC_LEN, 16))
6488 
6489 static int
6490 send_etid_flowc_wr(struct cxgbe_rate_tag *cst, struct port_info *pi,
6491     struct vi_info *vi)
6492 {
6493 	struct wrq_cookie cookie;
6494 	u_int pfvf = pi->adapter->pf << S_FW_VIID_PFN;
6495 	struct fw_flowc_wr *flowc;
6496 
6497 	mtx_assert(&cst->lock, MA_OWNED);
6498 	MPASS((cst->flags & (EO_FLOWC_PENDING | EO_FLOWC_RPL_PENDING)) ==
6499 	    EO_FLOWC_PENDING);
6500 
6501 	flowc = start_wrq_wr(&cst->eo_txq->wrq, ETID_FLOWC_LEN16, &cookie);
6502 	if (__predict_false(flowc == NULL))
6503 		return (ENOMEM);
6504 
6505 	bzero(flowc, ETID_FLOWC_LEN);
6506 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
6507 	    V_FW_FLOWC_WR_NPARAMS(ETID_FLOWC_NPARAMS) | V_FW_WR_COMPL(0));
6508 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(ETID_FLOWC_LEN16) |
6509 	    V_FW_WR_FLOWID(cst->etid));
6510 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN;
6511 	flowc->mnemval[0].val = htobe32(pfvf);
6512 	flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH;
6513 	flowc->mnemval[1].val = htobe32(pi->tx_chan);
6514 	flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT;
6515 	flowc->mnemval[2].val = htobe32(pi->tx_chan);
6516 	flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID;
6517 	flowc->mnemval[3].val = htobe32(cst->iqid);
6518 	flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_EOSTATE;
6519 	flowc->mnemval[4].val = htobe32(FW_FLOWC_MNEM_EOSTATE_ESTABLISHED);
6520 	flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS;
6521 	flowc->mnemval[5].val = htobe32(cst->schedcl);
6522 
6523 	commit_wrq_wr(&cst->eo_txq->wrq, flowc, &cookie);
6524 
6525 	cst->flags &= ~EO_FLOWC_PENDING;
6526 	cst->flags |= EO_FLOWC_RPL_PENDING;
6527 	MPASS(cst->tx_credits >= ETID_FLOWC_LEN16);	/* flowc is first WR. */
6528 	cst->tx_credits -= ETID_FLOWC_LEN16;
6529 
6530 	return (0);
6531 }
6532 
6533 #define ETID_FLUSH_LEN16 (howmany(sizeof (struct fw_flowc_wr), 16))
6534 
6535 void
6536 send_etid_flush_wr(struct cxgbe_rate_tag *cst)
6537 {
6538 	struct fw_flowc_wr *flowc;
6539 	struct wrq_cookie cookie;
6540 
6541 	mtx_assert(&cst->lock, MA_OWNED);
6542 
6543 	flowc = start_wrq_wr(&cst->eo_txq->wrq, ETID_FLUSH_LEN16, &cookie);
6544 	if (__predict_false(flowc == NULL))
6545 		CXGBE_UNIMPLEMENTED(__func__);
6546 
6547 	bzero(flowc, ETID_FLUSH_LEN16 * 16);
6548 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
6549 	    V_FW_FLOWC_WR_NPARAMS(0) | F_FW_WR_COMPL);
6550 	flowc->flowid_len16 = htobe32(V_FW_WR_LEN16(ETID_FLUSH_LEN16) |
6551 	    V_FW_WR_FLOWID(cst->etid));
6552 
6553 	commit_wrq_wr(&cst->eo_txq->wrq, flowc, &cookie);
6554 
6555 	cst->flags |= EO_FLUSH_RPL_PENDING;
6556 	MPASS(cst->tx_credits >= ETID_FLUSH_LEN16);
6557 	cst->tx_credits -= ETID_FLUSH_LEN16;
6558 	cst->ncompl++;
6559 }
6560 
6561 static void
6562 write_ethofld_wr(struct cxgbe_rate_tag *cst, struct fw_eth_tx_eo_wr *wr,
6563     struct mbuf *m0, int compl)
6564 {
6565 	struct cpl_tx_pkt_core *cpl;
6566 	uint64_t ctrl1;
6567 	uint32_t ctrl;	/* used in many unrelated places */
6568 	int len16, pktlen, nsegs, immhdrs;
6569 	caddr_t dst;
6570 	uintptr_t p;
6571 	struct ulptx_sgl *usgl;
6572 	struct sglist sg;
6573 	struct sglist_seg segs[38];	/* XXX: find real limit.  XXX: get off the stack */
6574 
6575 	mtx_assert(&cst->lock, MA_OWNED);
6576 	M_ASSERTPKTHDR(m0);
6577 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
6578 	    m0->m_pkthdr.l4hlen > 0,
6579 	    ("%s: ethofld mbuf %p is missing header lengths", __func__, m0));
6580 
6581 	len16 = mbuf_eo_len16(m0);
6582 	nsegs = mbuf_eo_nsegs(m0);
6583 	pktlen = m0->m_pkthdr.len;
6584 	ctrl = sizeof(struct cpl_tx_pkt_core);
6585 	if (needs_tso(m0))
6586 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
6587 	immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen + m0->m_pkthdr.l4hlen;
6588 	ctrl += immhdrs;
6589 
6590 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_EO_WR) |
6591 	    V_FW_ETH_TX_EO_WR_IMMDLEN(ctrl) | V_FW_WR_COMPL(!!compl));
6592 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(len16) |
6593 	    V_FW_WR_FLOWID(cst->etid));
6594 	wr->r3 = 0;
6595 	if (needs_outer_udp_csum(m0)) {
6596 		wr->u.udpseg.type = FW_ETH_TX_EO_TYPE_UDPSEG;
6597 		wr->u.udpseg.ethlen = m0->m_pkthdr.l2hlen;
6598 		wr->u.udpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
6599 		wr->u.udpseg.udplen = m0->m_pkthdr.l4hlen;
6600 		wr->u.udpseg.rtplen = 0;
6601 		wr->u.udpseg.r4 = 0;
6602 		wr->u.udpseg.mss = htobe16(pktlen - immhdrs);
6603 		wr->u.udpseg.schedpktsize = wr->u.udpseg.mss;
6604 		wr->u.udpseg.plen = htobe32(pktlen - immhdrs);
6605 		cpl = (void *)(wr + 1);
6606 	} else {
6607 		MPASS(needs_outer_tcp_csum(m0));
6608 		wr->u.tcpseg.type = FW_ETH_TX_EO_TYPE_TCPSEG;
6609 		wr->u.tcpseg.ethlen = m0->m_pkthdr.l2hlen;
6610 		wr->u.tcpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
6611 		wr->u.tcpseg.tcplen = m0->m_pkthdr.l4hlen;
6612 		wr->u.tcpseg.tsclk_tsoff = mbuf_eo_tsclk_tsoff(m0);
6613 		wr->u.tcpseg.r4 = 0;
6614 		wr->u.tcpseg.r5 = 0;
6615 		wr->u.tcpseg.plen = htobe32(pktlen - immhdrs);
6616 
6617 		if (needs_tso(m0)) {
6618 			struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
6619 
6620 			wr->u.tcpseg.mss = htobe16(m0->m_pkthdr.tso_segsz);
6621 
6622 			ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) |
6623 			    F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE |
6624 			    V_LSO_ETHHDR_LEN((m0->m_pkthdr.l2hlen -
6625 				ETHER_HDR_LEN) >> 2) |
6626 			    V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) |
6627 			    V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
6628 			if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
6629 				ctrl |= F_LSO_IPV6;
6630 			lso->lso_ctrl = htobe32(ctrl);
6631 			lso->ipid_ofst = htobe16(0);
6632 			lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
6633 			lso->seqno_offset = htobe32(0);
6634 			lso->len = htobe32(pktlen);
6635 
6636 			cpl = (void *)(lso + 1);
6637 		} else {
6638 			wr->u.tcpseg.mss = htobe16(0xffff);
6639 			cpl = (void *)(wr + 1);
6640 		}
6641 	}
6642 
6643 	/* Checksum offload must be requested for ethofld. */
6644 	MPASS(needs_outer_l4_csum(m0));
6645 	ctrl1 = csum_to_ctrl(cst->adapter, m0);
6646 
6647 	/* VLAN tag insertion */
6648 	if (needs_vlan_insertion(m0)) {
6649 		ctrl1 |= F_TXPKT_VLAN_VLD |
6650 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
6651 	}
6652 
6653 	/* CPL header */
6654 	cpl->ctrl0 = cst->ctrl0;
6655 	cpl->pack = 0;
6656 	cpl->len = htobe16(pktlen);
6657 	cpl->ctrl1 = htobe64(ctrl1);
6658 
6659 	/* Copy Ethernet, IP & TCP/UDP hdrs as immediate data */
6660 	p = (uintptr_t)(cpl + 1);
6661 	m_copydata(m0, 0, immhdrs, (void *)p);
6662 
6663 	/* SGL */
6664 	dst = (void *)(cpl + 1);
6665 	if (nsegs > 0) {
6666 		int i, pad;
6667 
6668 		/* zero-pad upto next 16Byte boundary, if not 16Byte aligned */
6669 		p += immhdrs;
6670 		pad = 16 - (immhdrs & 0xf);
6671 		bzero((void *)p, pad);
6672 
6673 		usgl = (void *)(p + pad);
6674 		usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
6675 		    V_ULPTX_NSGE(nsegs));
6676 
6677 		sglist_init(&sg, nitems(segs), segs);
6678 		for (; m0 != NULL; m0 = m0->m_next) {
6679 			if (__predict_false(m0->m_len == 0))
6680 				continue;
6681 			if (immhdrs >= m0->m_len) {
6682 				immhdrs -= m0->m_len;
6683 				continue;
6684 			}
6685 			if (m0->m_flags & M_EXTPG)
6686 				sglist_append_mbuf_epg(&sg, m0,
6687 				    mtod(m0, vm_offset_t), m0->m_len);
6688                         else
6689 				sglist_append(&sg, mtod(m0, char *) + immhdrs,
6690 				    m0->m_len - immhdrs);
6691 			immhdrs = 0;
6692 		}
6693 		MPASS(sg.sg_nseg == nsegs);
6694 
6695 		/*
6696 		 * Zero pad last 8B in case the WR doesn't end on a 16B
6697 		 * boundary.
6698 		 */
6699 		*(uint64_t *)((char *)wr + len16 * 16 - 8) = 0;
6700 
6701 		usgl->len0 = htobe32(segs[0].ss_len);
6702 		usgl->addr0 = htobe64(segs[0].ss_paddr);
6703 		for (i = 0; i < nsegs - 1; i++) {
6704 			usgl->sge[i / 2].len[i & 1] = htobe32(segs[i + 1].ss_len);
6705 			usgl->sge[i / 2].addr[i & 1] = htobe64(segs[i + 1].ss_paddr);
6706 		}
6707 		if (i & 1)
6708 			usgl->sge[i / 2].len[1] = htobe32(0);
6709 	}
6710 
6711 }
6712 
6713 static void
6714 ethofld_tx(struct cxgbe_rate_tag *cst)
6715 {
6716 	struct mbuf *m;
6717 	struct wrq_cookie cookie;
6718 	int next_credits, compl;
6719 	struct fw_eth_tx_eo_wr *wr;
6720 
6721 	mtx_assert(&cst->lock, MA_OWNED);
6722 
6723 	while ((m = mbufq_first(&cst->pending_tx)) != NULL) {
6724 		M_ASSERTPKTHDR(m);
6725 
6726 		/* How many len16 credits do we need to send this mbuf. */
6727 		next_credits = mbuf_eo_len16(m);
6728 		MPASS(next_credits > 0);
6729 		if (next_credits > cst->tx_credits) {
6730 			/*
6731 			 * Tx will make progress eventually because there is at
6732 			 * least one outstanding fw4_ack that will return
6733 			 * credits and kick the tx.
6734 			 */
6735 			MPASS(cst->ncompl > 0);
6736 			return;
6737 		}
6738 		wr = start_wrq_wr(&cst->eo_txq->wrq, next_credits, &cookie);
6739 		if (__predict_false(wr == NULL)) {
6740 			/* XXX: wishful thinking, not a real assertion. */
6741 			MPASS(cst->ncompl > 0);
6742 			return;
6743 		}
6744 		cst->tx_credits -= next_credits;
6745 		cst->tx_nocompl += next_credits;
6746 		compl = cst->ncompl == 0 || cst->tx_nocompl >= cst->tx_total / 2;
6747 		ETHER_BPF_MTAP(cst->com.ifp, m);
6748 		write_ethofld_wr(cst, wr, m, compl);
6749 		commit_wrq_wr(&cst->eo_txq->wrq, wr, &cookie);
6750 		if (compl) {
6751 			cst->ncompl++;
6752 			cst->tx_nocompl	= 0;
6753 		}
6754 		(void) mbufq_dequeue(&cst->pending_tx);
6755 
6756 		/*
6757 		 * Drop the mbuf's reference on the tag now rather
6758 		 * than waiting until m_freem().  This ensures that
6759 		 * cxgbe_rate_tag_free gets called when the inp drops
6760 		 * its reference on the tag and there are no more
6761 		 * mbufs in the pending_tx queue and can flush any
6762 		 * pending requests.  Otherwise if the last mbuf
6763 		 * doesn't request a completion the etid will never be
6764 		 * released.
6765 		 */
6766 		m->m_pkthdr.snd_tag = NULL;
6767 		m->m_pkthdr.csum_flags &= ~CSUM_SND_TAG;
6768 		m_snd_tag_rele(&cst->com);
6769 
6770 		mbufq_enqueue(&cst->pending_fwack, m);
6771 	}
6772 }
6773 
6774 int
6775 ethofld_transmit(struct ifnet *ifp, struct mbuf *m0)
6776 {
6777 	struct cxgbe_rate_tag *cst;
6778 	int rc;
6779 
6780 	MPASS(m0->m_nextpkt == NULL);
6781 	MPASS(m0->m_pkthdr.csum_flags & CSUM_SND_TAG);
6782 	MPASS(m0->m_pkthdr.snd_tag != NULL);
6783 	cst = mst_to_crt(m0->m_pkthdr.snd_tag);
6784 
6785 	mtx_lock(&cst->lock);
6786 	MPASS(cst->flags & EO_SND_TAG_REF);
6787 
6788 	if (__predict_false(cst->flags & EO_FLOWC_PENDING)) {
6789 		struct vi_info *vi = ifp->if_softc;
6790 		struct port_info *pi = vi->pi;
6791 		struct adapter *sc = pi->adapter;
6792 		const uint32_t rss_mask = vi->rss_size - 1;
6793 		uint32_t rss_hash;
6794 
6795 		cst->eo_txq = &sc->sge.ofld_txq[vi->first_ofld_txq];
6796 		if (M_HASHTYPE_ISHASH(m0))
6797 			rss_hash = m0->m_pkthdr.flowid;
6798 		else
6799 			rss_hash = arc4random();
6800 		/* We assume RSS hashing */
6801 		cst->iqid = vi->rss[rss_hash & rss_mask];
6802 		cst->eo_txq += rss_hash % vi->nofldtxq;
6803 		rc = send_etid_flowc_wr(cst, pi, vi);
6804 		if (rc != 0)
6805 			goto done;
6806 	}
6807 
6808 	if (__predict_false(cst->plen + m0->m_pkthdr.len > eo_max_backlog)) {
6809 		rc = ENOBUFS;
6810 		goto done;
6811 	}
6812 
6813 	mbufq_enqueue(&cst->pending_tx, m0);
6814 	cst->plen += m0->m_pkthdr.len;
6815 
6816 	/*
6817 	 * Hold an extra reference on the tag while generating work
6818 	 * requests to ensure that we don't try to free the tag during
6819 	 * ethofld_tx() in case we are sending the final mbuf after
6820 	 * the inp was freed.
6821 	 */
6822 	m_snd_tag_ref(&cst->com);
6823 	ethofld_tx(cst);
6824 	mtx_unlock(&cst->lock);
6825 	m_snd_tag_rele(&cst->com);
6826 	return (0);
6827 
6828 done:
6829 	mtx_unlock(&cst->lock);
6830 	if (__predict_false(rc != 0))
6831 		m_freem(m0);
6832 	return (rc);
6833 }
6834 
6835 static int
6836 ethofld_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
6837 {
6838 	struct adapter *sc = iq->adapter;
6839 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
6840 	struct mbuf *m;
6841 	u_int etid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
6842 	struct cxgbe_rate_tag *cst;
6843 	uint8_t credits = cpl->credits;
6844 
6845 	cst = lookup_etid(sc, etid);
6846 	mtx_lock(&cst->lock);
6847 	if (__predict_false(cst->flags & EO_FLOWC_RPL_PENDING)) {
6848 		MPASS(credits >= ETID_FLOWC_LEN16);
6849 		credits -= ETID_FLOWC_LEN16;
6850 		cst->flags &= ~EO_FLOWC_RPL_PENDING;
6851 	}
6852 
6853 	KASSERT(cst->ncompl > 0,
6854 	    ("%s: etid %u (%p) wasn't expecting completion.",
6855 	    __func__, etid, cst));
6856 	cst->ncompl--;
6857 
6858 	while (credits > 0) {
6859 		m = mbufq_dequeue(&cst->pending_fwack);
6860 		if (__predict_false(m == NULL)) {
6861 			/*
6862 			 * The remaining credits are for the final flush that
6863 			 * was issued when the tag was freed by the kernel.
6864 			 */
6865 			MPASS((cst->flags &
6866 			    (EO_FLUSH_RPL_PENDING | EO_SND_TAG_REF)) ==
6867 			    EO_FLUSH_RPL_PENDING);
6868 			MPASS(credits == ETID_FLUSH_LEN16);
6869 			MPASS(cst->tx_credits + cpl->credits == cst->tx_total);
6870 			MPASS(cst->ncompl == 0);
6871 
6872 			cst->flags &= ~EO_FLUSH_RPL_PENDING;
6873 			cst->tx_credits += cpl->credits;
6874 			cxgbe_rate_tag_free_locked(cst);
6875 			return (0);	/* cst is gone. */
6876 		}
6877 		KASSERT(m != NULL,
6878 		    ("%s: too many credits (%u, %u)", __func__, cpl->credits,
6879 		    credits));
6880 		KASSERT(credits >= mbuf_eo_len16(m),
6881 		    ("%s: too few credits (%u, %u, %u)", __func__,
6882 		    cpl->credits, credits, mbuf_eo_len16(m)));
6883 		credits -= mbuf_eo_len16(m);
6884 		cst->plen -= m->m_pkthdr.len;
6885 		m_freem(m);
6886 	}
6887 
6888 	cst->tx_credits += cpl->credits;
6889 	MPASS(cst->tx_credits <= cst->tx_total);
6890 
6891 	if (cst->flags & EO_SND_TAG_REF) {
6892 		/*
6893 		 * As with ethofld_transmit(), hold an extra reference
6894 		 * so that the tag is stable across ethold_tx().
6895 		 */
6896 		m_snd_tag_ref(&cst->com);
6897 		m = mbufq_first(&cst->pending_tx);
6898 		if (m != NULL && cst->tx_credits >= mbuf_eo_len16(m))
6899 			ethofld_tx(cst);
6900 		mtx_unlock(&cst->lock);
6901 		m_snd_tag_rele(&cst->com);
6902 	} else {
6903 		/*
6904 		 * There shouldn't be any pending packets if the tag
6905 		 * was freed by the kernel since any pending packet
6906 		 * should hold a reference to the tag.
6907 		 */
6908 		MPASS(mbufq_first(&cst->pending_tx) == NULL);
6909 		mtx_unlock(&cst->lock);
6910 	}
6911 
6912 	return (0);
6913 }
6914 #endif
6915