xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision c1cdf6a42f0d951ba720688dfc6ce07608b02f6e)
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_ratelimit.h"
36 
37 #include <sys/types.h>
38 #include <sys/eventhandler.h>
39 #include <sys/mbuf.h>
40 #include <sys/socket.h>
41 #include <sys/kernel.h>
42 #include <sys/malloc.h>
43 #include <sys/queue.h>
44 #include <sys/sbuf.h>
45 #include <sys/taskqueue.h>
46 #include <sys/time.h>
47 #include <sys/sglist.h>
48 #include <sys/sysctl.h>
49 #include <sys/smp.h>
50 #include <sys/counter.h>
51 #include <net/bpf.h>
52 #include <net/ethernet.h>
53 #include <net/if.h>
54 #include <net/if_vlan_var.h>
55 #include <netinet/in.h>
56 #include <netinet/ip.h>
57 #include <netinet/ip6.h>
58 #include <netinet/tcp.h>
59 #include <netinet/udp.h>
60 #include <machine/in_cksum.h>
61 #include <machine/md_var.h>
62 #include <vm/vm.h>
63 #include <vm/pmap.h>
64 #ifdef DEV_NETMAP
65 #include <machine/bus.h>
66 #include <sys/selinfo.h>
67 #include <net/if_var.h>
68 #include <net/netmap.h>
69 #include <dev/netmap/netmap_kern.h>
70 #endif
71 
72 #include "common/common.h"
73 #include "common/t4_regs.h"
74 #include "common/t4_regs_values.h"
75 #include "common/t4_msg.h"
76 #include "t4_l2t.h"
77 #include "t4_mp_ring.h"
78 
79 #ifdef T4_PKT_TIMESTAMP
80 #define RX_COPY_THRESHOLD (MINCLSIZE - 8)
81 #else
82 #define RX_COPY_THRESHOLD MINCLSIZE
83 #endif
84 
85 /*
86  * Ethernet frames are DMA'd at this byte offset into the freelist buffer.
87  * 0-7 are valid values.
88  */
89 static int fl_pktshift = 0;
90 TUNABLE_INT("hw.cxgbe.fl_pktshift", &fl_pktshift);
91 
92 /*
93  * Pad ethernet payload up to this boundary.
94  * -1: driver should figure out a good value.
95  *  0: disable padding.
96  *  Any power of 2 from 32 to 4096 (both inclusive) is also a valid value.
97  */
98 int fl_pad = -1;
99 TUNABLE_INT("hw.cxgbe.fl_pad", &fl_pad);
100 
101 /*
102  * Status page length.
103  * -1: driver should figure out a good value.
104  *  64 or 128 are the only other valid values.
105  */
106 static int spg_len = -1;
107 TUNABLE_INT("hw.cxgbe.spg_len", &spg_len);
108 
109 /*
110  * Congestion drops.
111  * -1: no congestion feedback (not recommended).
112  *  0: backpressure the channel instead of dropping packets right away.
113  *  1: no backpressure, drop packets for the congested queue immediately.
114  */
115 static int cong_drop = 0;
116 TUNABLE_INT("hw.cxgbe.cong_drop", &cong_drop);
117 
118 /*
119  * Deliver multiple frames in the same free list buffer if they fit.
120  * -1: let the driver decide whether to enable buffer packing or not.
121  *  0: disable buffer packing.
122  *  1: enable buffer packing.
123  */
124 static int buffer_packing = -1;
125 TUNABLE_INT("hw.cxgbe.buffer_packing", &buffer_packing);
126 
127 /*
128  * Start next frame in a packed buffer at this boundary.
129  * -1: driver should figure out a good value.
130  * T4: driver will ignore this and use the same value as fl_pad above.
131  * T5: 16, or a power of 2 from 64 to 4096 (both inclusive) is a valid value.
132  */
133 static int fl_pack = -1;
134 TUNABLE_INT("hw.cxgbe.fl_pack", &fl_pack);
135 
136 /*
137  * Allow the driver to create mbuf(s) in a cluster allocated for rx.
138  * 0: never; always allocate mbufs from the zone_mbuf UMA zone.
139  * 1: ok to create mbuf(s) within a cluster if there is room.
140  */
141 static int allow_mbufs_in_cluster = 1;
142 TUNABLE_INT("hw.cxgbe.allow_mbufs_in_cluster", &allow_mbufs_in_cluster);
143 
144 /*
145  * Largest rx cluster size that the driver is allowed to allocate.
146  */
147 static int largest_rx_cluster = MJUM16BYTES;
148 TUNABLE_INT("hw.cxgbe.largest_rx_cluster", &largest_rx_cluster);
149 
150 /*
151  * Size of cluster allocation that's most likely to succeed.  The driver will
152  * fall back to this size if it fails to allocate clusters larger than this.
153  */
154 static int safest_rx_cluster = PAGE_SIZE;
155 TUNABLE_INT("hw.cxgbe.safest_rx_cluster", &safest_rx_cluster);
156 
157 #ifdef RATELIMIT
158 /*
159  * Knob to control TCP timestamp rewriting, and the granularity of the tick used
160  * for rewriting.  -1 and 0-3 are all valid values.
161  * -1: hardware should leave the TCP timestamps alone.
162  * 0: 1ms
163  * 1: 100us
164  * 2: 10us
165  * 3: 1us
166  */
167 static int tsclk = -1;
168 TUNABLE_INT("hw.cxgbe.tsclk", &tsclk);
169 
170 static int eo_max_backlog = 1024 * 1024;
171 TUNABLE_INT("hw.cxgbe.eo_max_backlog", &eo_max_backlog);
172 #endif
173 
174 /*
175  * The interrupt holdoff timers are multiplied by this value on T6+.
176  * 1 and 3-17 (both inclusive) are legal values.
177  */
178 static int tscale = 1;
179 TUNABLE_INT("hw.cxgbe.tscale", &tscale);
180 
181 /*
182  * Number of LRO entries in the lro_ctrl structure per rx queue.
183  */
184 static int lro_entries = TCP_LRO_ENTRIES;
185 TUNABLE_INT("hw.cxgbe.lro_entries", &lro_entries);
186 
187 /*
188  * This enables presorting of frames before they're fed into tcp_lro_rx.
189  */
190 static int lro_mbufs = 0;
191 TUNABLE_INT("hw.cxgbe.lro_mbufs", &lro_mbufs);
192 
193 struct txpkts {
194 	u_int wr_type;		/* type 0 or type 1 */
195 	u_int npkt;		/* # of packets in this work request */
196 	u_int plen;		/* total payload (sum of all packets) */
197 	u_int len16;		/* # of 16B pieces used by this work request */
198 };
199 
200 /* A packet's SGL.  This + m_pkthdr has all info needed for tx */
201 struct sgl {
202 	struct sglist sg;
203 	struct sglist_seg seg[TX_SGL_SEGS];
204 };
205 
206 static int service_iq(struct sge_iq *, int);
207 static int service_iq_fl(struct sge_iq *, int);
208 static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t);
209 static int t4_eth_rx(struct sge_iq *, const struct rss_header *, struct mbuf *);
210 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int);
211 static inline void init_fl(struct adapter *, struct sge_fl *, int, int, char *);
212 static inline void init_eq(struct adapter *, struct sge_eq *, int, int, uint8_t,
213     uint16_t, char *);
214 static int alloc_ring(struct adapter *, size_t, bus_dma_tag_t *, bus_dmamap_t *,
215     bus_addr_t *, void **);
216 static int free_ring(struct adapter *, bus_dma_tag_t, bus_dmamap_t, bus_addr_t,
217     void *);
218 static int alloc_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *,
219     int, int);
220 static int free_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *);
221 static void add_iq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
222     struct sge_iq *);
223 static void add_fl_sysctls(struct adapter *, struct sysctl_ctx_list *,
224     struct sysctl_oid *, struct sge_fl *);
225 static int alloc_fwq(struct adapter *);
226 static int free_fwq(struct adapter *);
227 static int alloc_ctrlq(struct adapter *, struct sge_wrq *, int,
228     struct sysctl_oid *);
229 static int alloc_rxq(struct vi_info *, struct sge_rxq *, int, int,
230     struct sysctl_oid *);
231 static int free_rxq(struct vi_info *, struct sge_rxq *);
232 #ifdef TCP_OFFLOAD
233 static int alloc_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *, int, int,
234     struct sysctl_oid *);
235 static int free_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *);
236 #endif
237 #ifdef DEV_NETMAP
238 static int alloc_nm_rxq(struct vi_info *, struct sge_nm_rxq *, int, int,
239     struct sysctl_oid *);
240 static int free_nm_rxq(struct vi_info *, struct sge_nm_rxq *);
241 static int alloc_nm_txq(struct vi_info *, struct sge_nm_txq *, int, int,
242     struct sysctl_oid *);
243 static int free_nm_txq(struct vi_info *, struct sge_nm_txq *);
244 #endif
245 static int ctrl_eq_alloc(struct adapter *, struct sge_eq *);
246 static int eth_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
247 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
248 static int ofld_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
249 #endif
250 static int alloc_eq(struct adapter *, struct vi_info *, struct sge_eq *);
251 static int free_eq(struct adapter *, struct sge_eq *);
252 static int alloc_wrq(struct adapter *, struct vi_info *, struct sge_wrq *,
253     struct sysctl_oid *);
254 static int free_wrq(struct adapter *, struct sge_wrq *);
255 static int alloc_txq(struct vi_info *, struct sge_txq *, int,
256     struct sysctl_oid *);
257 static int free_txq(struct vi_info *, struct sge_txq *);
258 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int);
259 static inline void ring_fl_db(struct adapter *, struct sge_fl *);
260 static int refill_fl(struct adapter *, struct sge_fl *, int);
261 static void refill_sfl(void *);
262 static int alloc_fl_sdesc(struct sge_fl *);
263 static void free_fl_sdesc(struct adapter *, struct sge_fl *);
264 static void find_best_refill_source(struct adapter *, struct sge_fl *, int);
265 static void find_safe_refill_source(struct adapter *, struct sge_fl *);
266 static void add_fl_to_sfl(struct adapter *, struct sge_fl *);
267 
268 static inline void get_pkt_gl(struct mbuf *, struct sglist *);
269 static inline u_int txpkt_len16(u_int, u_int);
270 static inline u_int txpkt_vm_len16(u_int, u_int);
271 static inline u_int txpkts0_len16(u_int);
272 static inline u_int txpkts1_len16(void);
273 static u_int write_txpkt_wr(struct sge_txq *, struct fw_eth_tx_pkt_wr *,
274     struct mbuf *, u_int);
275 static u_int write_txpkt_vm_wr(struct adapter *, struct sge_txq *,
276     struct fw_eth_tx_pkt_vm_wr *, struct mbuf *, u_int);
277 static int try_txpkts(struct mbuf *, struct mbuf *, struct txpkts *, u_int);
278 static int add_to_txpkts(struct mbuf *, struct txpkts *, u_int);
279 static u_int write_txpkts_wr(struct sge_txq *, struct fw_eth_tx_pkts_wr *,
280     struct mbuf *, const struct txpkts *, u_int);
281 static void write_gl_to_txd(struct sge_txq *, struct mbuf *, caddr_t *, int);
282 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int);
283 static inline void ring_eq_db(struct adapter *, struct sge_eq *, u_int);
284 static inline uint16_t read_hw_cidx(struct sge_eq *);
285 static inline u_int reclaimable_tx_desc(struct sge_eq *);
286 static inline u_int total_available_tx_desc(struct sge_eq *);
287 static u_int reclaim_tx_descs(struct sge_txq *, u_int);
288 static void tx_reclaim(void *, int);
289 static __be64 get_flit(struct sglist_seg *, int, int);
290 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
291     struct mbuf *);
292 static int handle_fw_msg(struct sge_iq *, const struct rss_header *,
293     struct mbuf *);
294 static int t4_handle_wrerr_rpl(struct adapter *, const __be64 *);
295 static void wrq_tx_drain(void *, int);
296 static void drain_wrq_wr_list(struct adapter *, struct sge_wrq *);
297 
298 static int sysctl_uint16(SYSCTL_HANDLER_ARGS);
299 static int sysctl_bufsizes(SYSCTL_HANDLER_ARGS);
300 #ifdef RATELIMIT
301 static inline u_int txpkt_eo_len16(u_int, u_int, u_int);
302 static int ethofld_fw4_ack(struct sge_iq *, const struct rss_header *,
303     struct mbuf *);
304 #endif
305 
306 static counter_u64_t extfree_refs;
307 static counter_u64_t extfree_rels;
308 
309 an_handler_t t4_an_handler;
310 fw_msg_handler_t t4_fw_msg_handler[NUM_FW6_TYPES];
311 cpl_handler_t t4_cpl_handler[NUM_CPL_CMDS];
312 cpl_handler_t set_tcb_rpl_handlers[NUM_CPL_COOKIES];
313 cpl_handler_t l2t_write_rpl_handlers[NUM_CPL_COOKIES];
314 cpl_handler_t act_open_rpl_handlers[NUM_CPL_COOKIES];
315 cpl_handler_t abort_rpl_rss_handlers[NUM_CPL_COOKIES];
316 cpl_handler_t fw4_ack_handlers[NUM_CPL_COOKIES];
317 
318 void
319 t4_register_an_handler(an_handler_t h)
320 {
321 	uintptr_t *loc;
322 
323 	MPASS(h == NULL || t4_an_handler == NULL);
324 
325 	loc = (uintptr_t *)&t4_an_handler;
326 	atomic_store_rel_ptr(loc, (uintptr_t)h);
327 }
328 
329 void
330 t4_register_fw_msg_handler(int type, fw_msg_handler_t h)
331 {
332 	uintptr_t *loc;
333 
334 	MPASS(type < nitems(t4_fw_msg_handler));
335 	MPASS(h == NULL || t4_fw_msg_handler[type] == NULL);
336 	/*
337 	 * These are dispatched by the handler for FW{4|6}_CPL_MSG using the CPL
338 	 * handler dispatch table.  Reject any attempt to install a handler for
339 	 * this subtype.
340 	 */
341 	MPASS(type != FW_TYPE_RSSCPL);
342 	MPASS(type != FW6_TYPE_RSSCPL);
343 
344 	loc = (uintptr_t *)&t4_fw_msg_handler[type];
345 	atomic_store_rel_ptr(loc, (uintptr_t)h);
346 }
347 
348 void
349 t4_register_cpl_handler(int opcode, cpl_handler_t h)
350 {
351 	uintptr_t *loc;
352 
353 	MPASS(opcode < nitems(t4_cpl_handler));
354 	MPASS(h == NULL || t4_cpl_handler[opcode] == NULL);
355 
356 	loc = (uintptr_t *)&t4_cpl_handler[opcode];
357 	atomic_store_rel_ptr(loc, (uintptr_t)h);
358 }
359 
360 static int
361 set_tcb_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
362     struct mbuf *m)
363 {
364 	const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
365 	u_int tid;
366 	int cookie;
367 
368 	MPASS(m == NULL);
369 
370 	tid = GET_TID(cpl);
371 	if (is_hpftid(iq->adapter, tid) || is_ftid(iq->adapter, tid)) {
372 		/*
373 		 * The return code for filter-write is put in the CPL cookie so
374 		 * we have to rely on the hardware tid (is_ftid) to determine
375 		 * that this is a response to a filter.
376 		 */
377 		cookie = CPL_COOKIE_FILTER;
378 	} else {
379 		cookie = G_COOKIE(cpl->cookie);
380 	}
381 	MPASS(cookie > CPL_COOKIE_RESERVED);
382 	MPASS(cookie < nitems(set_tcb_rpl_handlers));
383 
384 	return (set_tcb_rpl_handlers[cookie](iq, rss, m));
385 }
386 
387 static int
388 l2t_write_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
389     struct mbuf *m)
390 {
391 	const struct cpl_l2t_write_rpl *rpl = (const void *)(rss + 1);
392 	unsigned int cookie;
393 
394 	MPASS(m == NULL);
395 
396 	cookie = GET_TID(rpl) & F_SYNC_WR ? CPL_COOKIE_TOM : CPL_COOKIE_FILTER;
397 	return (l2t_write_rpl_handlers[cookie](iq, rss, m));
398 }
399 
400 static int
401 act_open_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
402     struct mbuf *m)
403 {
404 	const struct cpl_act_open_rpl *cpl = (const void *)(rss + 1);
405 	u_int cookie = G_TID_COOKIE(G_AOPEN_ATID(be32toh(cpl->atid_status)));
406 
407 	MPASS(m == NULL);
408 	MPASS(cookie != CPL_COOKIE_RESERVED);
409 
410 	return (act_open_rpl_handlers[cookie](iq, rss, m));
411 }
412 
413 static int
414 abort_rpl_rss_handler(struct sge_iq *iq, const struct rss_header *rss,
415     struct mbuf *m)
416 {
417 	struct adapter *sc = iq->adapter;
418 	u_int cookie;
419 
420 	MPASS(m == NULL);
421 	if (is_hashfilter(sc))
422 		cookie = CPL_COOKIE_HASHFILTER;
423 	else
424 		cookie = CPL_COOKIE_TOM;
425 
426 	return (abort_rpl_rss_handlers[cookie](iq, rss, m));
427 }
428 
429 static int
430 fw4_ack_handler(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
431 {
432 	struct adapter *sc = iq->adapter;
433 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
434 	unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
435 	u_int cookie;
436 
437 	MPASS(m == NULL);
438 	if (is_etid(sc, tid))
439 		cookie = CPL_COOKIE_ETHOFLD;
440 	else
441 		cookie = CPL_COOKIE_TOM;
442 
443 	return (fw4_ack_handlers[cookie](iq, rss, m));
444 }
445 
446 static void
447 t4_init_shared_cpl_handlers(void)
448 {
449 
450 	t4_register_cpl_handler(CPL_SET_TCB_RPL, set_tcb_rpl_handler);
451 	t4_register_cpl_handler(CPL_L2T_WRITE_RPL, l2t_write_rpl_handler);
452 	t4_register_cpl_handler(CPL_ACT_OPEN_RPL, act_open_rpl_handler);
453 	t4_register_cpl_handler(CPL_ABORT_RPL_RSS, abort_rpl_rss_handler);
454 	t4_register_cpl_handler(CPL_FW4_ACK, fw4_ack_handler);
455 }
456 
457 void
458 t4_register_shared_cpl_handler(int opcode, cpl_handler_t h, int cookie)
459 {
460 	uintptr_t *loc;
461 
462 	MPASS(opcode < nitems(t4_cpl_handler));
463 	MPASS(cookie > CPL_COOKIE_RESERVED);
464 	MPASS(cookie < NUM_CPL_COOKIES);
465 	MPASS(t4_cpl_handler[opcode] != NULL);
466 
467 	switch (opcode) {
468 	case CPL_SET_TCB_RPL:
469 		loc = (uintptr_t *)&set_tcb_rpl_handlers[cookie];
470 		break;
471 	case CPL_L2T_WRITE_RPL:
472 		loc = (uintptr_t *)&l2t_write_rpl_handlers[cookie];
473 		break;
474 	case CPL_ACT_OPEN_RPL:
475 		loc = (uintptr_t *)&act_open_rpl_handlers[cookie];
476 		break;
477 	case CPL_ABORT_RPL_RSS:
478 		loc = (uintptr_t *)&abort_rpl_rss_handlers[cookie];
479 		break;
480 	case CPL_FW4_ACK:
481 		loc = (uintptr_t *)&fw4_ack_handlers[cookie];
482 		break;
483 	default:
484 		MPASS(0);
485 		return;
486 	}
487 	MPASS(h == NULL || *loc == (uintptr_t)NULL);
488 	atomic_store_rel_ptr(loc, (uintptr_t)h);
489 }
490 
491 /*
492  * Called on MOD_LOAD.  Validates and calculates the SGE tunables.
493  */
494 void
495 t4_sge_modload(void)
496 {
497 
498 	if (fl_pktshift < 0 || fl_pktshift > 7) {
499 		printf("Invalid hw.cxgbe.fl_pktshift value (%d),"
500 		    " using 0 instead.\n", fl_pktshift);
501 		fl_pktshift = 0;
502 	}
503 
504 	if (spg_len != 64 && spg_len != 128) {
505 		int len;
506 
507 #if defined(__i386__) || defined(__amd64__)
508 		len = cpu_clflush_line_size > 64 ? 128 : 64;
509 #else
510 		len = 64;
511 #endif
512 		if (spg_len != -1) {
513 			printf("Invalid hw.cxgbe.spg_len value (%d),"
514 			    " using %d instead.\n", spg_len, len);
515 		}
516 		spg_len = len;
517 	}
518 
519 	if (cong_drop < -1 || cong_drop > 1) {
520 		printf("Invalid hw.cxgbe.cong_drop value (%d),"
521 		    " using 0 instead.\n", cong_drop);
522 		cong_drop = 0;
523 	}
524 
525 	if (tscale != 1 && (tscale < 3 || tscale > 17)) {
526 		printf("Invalid hw.cxgbe.tscale value (%d),"
527 		    " using 1 instead.\n", tscale);
528 		tscale = 1;
529 	}
530 
531 	extfree_refs = counter_u64_alloc(M_WAITOK);
532 	extfree_rels = counter_u64_alloc(M_WAITOK);
533 	counter_u64_zero(extfree_refs);
534 	counter_u64_zero(extfree_rels);
535 
536 	t4_init_shared_cpl_handlers();
537 	t4_register_cpl_handler(CPL_FW4_MSG, handle_fw_msg);
538 	t4_register_cpl_handler(CPL_FW6_MSG, handle_fw_msg);
539 	t4_register_cpl_handler(CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
540 	t4_register_cpl_handler(CPL_RX_PKT, t4_eth_rx);
541 #ifdef RATELIMIT
542 	t4_register_shared_cpl_handler(CPL_FW4_ACK, ethofld_fw4_ack,
543 	    CPL_COOKIE_ETHOFLD);
544 #endif
545 	t4_register_fw_msg_handler(FW6_TYPE_CMD_RPL, t4_handle_fw_rpl);
546 	t4_register_fw_msg_handler(FW6_TYPE_WRERR_RPL, t4_handle_wrerr_rpl);
547 }
548 
549 void
550 t4_sge_modunload(void)
551 {
552 
553 	counter_u64_free(extfree_refs);
554 	counter_u64_free(extfree_rels);
555 }
556 
557 uint64_t
558 t4_sge_extfree_refs(void)
559 {
560 	uint64_t refs, rels;
561 
562 	rels = counter_u64_fetch(extfree_rels);
563 	refs = counter_u64_fetch(extfree_refs);
564 
565 	return (refs - rels);
566 }
567 
568 static inline void
569 setup_pad_and_pack_boundaries(struct adapter *sc)
570 {
571 	uint32_t v, m;
572 	int pad, pack, pad_shift;
573 
574 	pad_shift = chip_id(sc) > CHELSIO_T5 ? X_T6_INGPADBOUNDARY_SHIFT :
575 	    X_INGPADBOUNDARY_SHIFT;
576 	pad = fl_pad;
577 	if (fl_pad < (1 << pad_shift) ||
578 	    fl_pad > (1 << (pad_shift + M_INGPADBOUNDARY)) ||
579 	    !powerof2(fl_pad)) {
580 		/*
581 		 * If there is any chance that we might use buffer packing and
582 		 * the chip is a T4, then pick 64 as the pad/pack boundary.  Set
583 		 * it to the minimum allowed in all other cases.
584 		 */
585 		pad = is_t4(sc) && buffer_packing ? 64 : 1 << pad_shift;
586 
587 		/*
588 		 * For fl_pad = 0 we'll still write a reasonable value to the
589 		 * register but all the freelists will opt out of padding.
590 		 * We'll complain here only if the user tried to set it to a
591 		 * value greater than 0 that was invalid.
592 		 */
593 		if (fl_pad > 0) {
594 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pad value"
595 			    " (%d), using %d instead.\n", fl_pad, pad);
596 		}
597 	}
598 	m = V_INGPADBOUNDARY(M_INGPADBOUNDARY);
599 	v = V_INGPADBOUNDARY(ilog2(pad) - pad_shift);
600 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
601 
602 	if (is_t4(sc)) {
603 		if (fl_pack != -1 && fl_pack != pad) {
604 			/* Complain but carry on. */
605 			device_printf(sc->dev, "hw.cxgbe.fl_pack (%d) ignored,"
606 			    " using %d instead.\n", fl_pack, pad);
607 		}
608 		return;
609 	}
610 
611 	pack = fl_pack;
612 	if (fl_pack < 16 || fl_pack == 32 || fl_pack > 4096 ||
613 	    !powerof2(fl_pack)) {
614 		pack = max(sc->params.pci.mps, CACHE_LINE_SIZE);
615 		MPASS(powerof2(pack));
616 		if (pack < 16)
617 			pack = 16;
618 		if (pack == 32)
619 			pack = 64;
620 		if (pack > 4096)
621 			pack = 4096;
622 		if (fl_pack != -1) {
623 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pack value"
624 			    " (%d), using %d instead.\n", fl_pack, pack);
625 		}
626 	}
627 	m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY);
628 	if (pack == 16)
629 		v = V_INGPACKBOUNDARY(0);
630 	else
631 		v = V_INGPACKBOUNDARY(ilog2(pack) - 5);
632 
633 	MPASS(!is_t4(sc));	/* T4 doesn't have SGE_CONTROL2 */
634 	t4_set_reg_field(sc, A_SGE_CONTROL2, m, v);
635 }
636 
637 /*
638  * adap->params.vpd.cclk must be set up before this is called.
639  */
640 void
641 t4_tweak_chip_settings(struct adapter *sc)
642 {
643 	int i;
644 	uint32_t v, m;
645 	int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200};
646 	int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk;
647 	int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */
648 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
649 	static int sge_flbuf_sizes[] = {
650 		MCLBYTES,
651 #if MJUMPAGESIZE != MCLBYTES
652 		MJUMPAGESIZE,
653 		MJUMPAGESIZE - CL_METADATA_SIZE,
654 		MJUMPAGESIZE - 2 * MSIZE - CL_METADATA_SIZE,
655 #endif
656 		MJUM9BYTES,
657 		MJUM16BYTES,
658 		MCLBYTES - MSIZE - CL_METADATA_SIZE,
659 		MJUM9BYTES - CL_METADATA_SIZE,
660 		MJUM16BYTES - CL_METADATA_SIZE,
661 	};
662 
663 	KASSERT(sc->flags & MASTER_PF,
664 	    ("%s: trying to change chip settings when not master.", __func__));
665 
666 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
667 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
668 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
669 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
670 
671 	setup_pad_and_pack_boundaries(sc);
672 
673 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
674 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
675 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
676 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
677 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
678 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
679 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
680 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
681 	t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v);
682 
683 	KASSERT(nitems(sge_flbuf_sizes) <= SGE_FLBUF_SIZES,
684 	    ("%s: hw buffer size table too big", __func__));
685 	for (i = 0; i < min(nitems(sge_flbuf_sizes), SGE_FLBUF_SIZES); i++) {
686 		t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i),
687 		    sge_flbuf_sizes[i]);
688 	}
689 
690 	v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) |
691 	    V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]);
692 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v);
693 
694 	KASSERT(intr_timer[0] <= timer_max,
695 	    ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0],
696 	    timer_max));
697 	for (i = 1; i < nitems(intr_timer); i++) {
698 		KASSERT(intr_timer[i] >= intr_timer[i - 1],
699 		    ("%s: timers not listed in increasing order (%d)",
700 		    __func__, i));
701 
702 		while (intr_timer[i] > timer_max) {
703 			if (i == nitems(intr_timer) - 1) {
704 				intr_timer[i] = timer_max;
705 				break;
706 			}
707 			intr_timer[i] += intr_timer[i - 1];
708 			intr_timer[i] /= 2;
709 		}
710 	}
711 
712 	v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) |
713 	    V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1]));
714 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v);
715 	v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) |
716 	    V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3]));
717 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v);
718 	v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) |
719 	    V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5]));
720 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v);
721 
722 	if (chip_id(sc) >= CHELSIO_T6) {
723 		m = V_TSCALE(M_TSCALE);
724 		if (tscale == 1)
725 			v = 0;
726 		else
727 			v = V_TSCALE(tscale - 2);
728 		t4_set_reg_field(sc, A_SGE_ITP_CONTROL, m, v);
729 
730 		if (sc->debug_flags & DF_DISABLE_TCB_CACHE) {
731 			m = V_RDTHRESHOLD(M_RDTHRESHOLD) | F_WRTHRTHRESHEN |
732 			    V_WRTHRTHRESH(M_WRTHRTHRESH);
733 			t4_tp_pio_read(sc, &v, 1, A_TP_CMM_CONFIG, 1);
734 			v &= ~m;
735 			v |= V_RDTHRESHOLD(1) | F_WRTHRTHRESHEN |
736 			    V_WRTHRTHRESH(16);
737 			t4_tp_pio_write(sc, &v, 1, A_TP_CMM_CONFIG, 1);
738 		}
739 	}
740 
741 	/* 4K, 16K, 64K, 256K DDP "page sizes" for TDDP */
742 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
743 	t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v);
744 
745 	/*
746 	 * 4K, 8K, 16K, 64K DDP "page sizes" for iSCSI DDP.  These have been
747 	 * chosen with MAXPHYS = 128K in mind.  The largest DDP buffer that we
748 	 * may have to deal with is MAXPHYS + 1 page.
749 	 */
750 	v = V_HPZ0(0) | V_HPZ1(1) | V_HPZ2(2) | V_HPZ3(4);
751 	t4_write_reg(sc, A_ULP_RX_ISCSI_PSZ, v);
752 
753 	/* We use multiple DDP page sizes both in plain-TOE and ISCSI modes. */
754 	m = v = F_TDDPTAGTCB | F_ISCSITAGTCB;
755 	t4_set_reg_field(sc, A_ULP_RX_CTL, m, v);
756 
757 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
758 	    F_RESETDDPOFFSET;
759 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
760 	t4_set_reg_field(sc, A_TP_PARA_REG5, m, v);
761 }
762 
763 /*
764  * SGE wants the buffer to be at least 64B and then a multiple of 16.  If
765  * padding is in use, the buffer's start and end need to be aligned to the pad
766  * boundary as well.  We'll just make sure that the size is a multiple of the
767  * boundary here, it is up to the buffer allocation code to make sure the start
768  * of the buffer is aligned as well.
769  */
770 static inline int
771 hwsz_ok(struct adapter *sc, int hwsz)
772 {
773 	int mask = fl_pad ? sc->params.sge.pad_boundary - 1 : 16 - 1;
774 
775 	return (hwsz >= 64 && (hwsz & mask) == 0);
776 }
777 
778 /*
779  * XXX: driver really should be able to deal with unexpected settings.
780  */
781 int
782 t4_read_chip_settings(struct adapter *sc)
783 {
784 	struct sge *s = &sc->sge;
785 	struct sge_params *sp = &sc->params.sge;
786 	int i, j, n, rc = 0;
787 	uint32_t m, v, r;
788 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
789 	static int sw_buf_sizes[] = {	/* Sorted by size */
790 		MCLBYTES,
791 #if MJUMPAGESIZE != MCLBYTES
792 		MJUMPAGESIZE,
793 #endif
794 		MJUM9BYTES,
795 		MJUM16BYTES
796 	};
797 	struct sw_zone_info *swz, *safe_swz;
798 	struct hw_buf_info *hwb;
799 
800 	m = F_RXPKTCPLMODE;
801 	v = F_RXPKTCPLMODE;
802 	r = sc->params.sge.sge_control;
803 	if ((r & m) != v) {
804 		device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r);
805 		rc = EINVAL;
806 	}
807 
808 	/*
809 	 * If this changes then every single use of PAGE_SHIFT in the driver
810 	 * needs to be carefully reviewed for PAGE_SHIFT vs sp->page_shift.
811 	 */
812 	if (sp->page_shift != PAGE_SHIFT) {
813 		device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r);
814 		rc = EINVAL;
815 	}
816 
817 	/* Filter out unusable hw buffer sizes entirely (mark with -2). */
818 	hwb = &s->hw_buf_info[0];
819 	for (i = 0; i < nitems(s->hw_buf_info); i++, hwb++) {
820 		r = sc->params.sge.sge_fl_buffer_size[i];
821 		hwb->size = r;
822 		hwb->zidx = hwsz_ok(sc, r) ? -1 : -2;
823 		hwb->next = -1;
824 	}
825 
826 	/*
827 	 * Create a sorted list in decreasing order of hw buffer sizes (and so
828 	 * increasing order of spare area) for each software zone.
829 	 *
830 	 * If padding is enabled then the start and end of the buffer must align
831 	 * to the pad boundary; if packing is enabled then they must align with
832 	 * the pack boundary as well.  Allocations from the cluster zones are
833 	 * aligned to min(size, 4K), so the buffer starts at that alignment and
834 	 * ends at hwb->size alignment.  If mbuf inlining is allowed the
835 	 * starting alignment will be reduced to MSIZE and the driver will
836 	 * exercise appropriate caution when deciding on the best buffer layout
837 	 * to use.
838 	 */
839 	n = 0;	/* no usable buffer size to begin with */
840 	swz = &s->sw_zone_info[0];
841 	safe_swz = NULL;
842 	for (i = 0; i < SW_ZONE_SIZES; i++, swz++) {
843 		int8_t head = -1, tail = -1;
844 
845 		swz->size = sw_buf_sizes[i];
846 		swz->zone = m_getzone(swz->size);
847 		swz->type = m_gettype(swz->size);
848 
849 		if (swz->size < PAGE_SIZE) {
850 			MPASS(powerof2(swz->size));
851 			if (fl_pad && (swz->size % sp->pad_boundary != 0))
852 				continue;
853 		}
854 
855 		if (swz->size == safest_rx_cluster)
856 			safe_swz = swz;
857 
858 		hwb = &s->hw_buf_info[0];
859 		for (j = 0; j < SGE_FLBUF_SIZES; j++, hwb++) {
860 			if (hwb->zidx != -1 || hwb->size > swz->size)
861 				continue;
862 #ifdef INVARIANTS
863 			if (fl_pad)
864 				MPASS(hwb->size % sp->pad_boundary == 0);
865 #endif
866 			hwb->zidx = i;
867 			if (head == -1)
868 				head = tail = j;
869 			else if (hwb->size < s->hw_buf_info[tail].size) {
870 				s->hw_buf_info[tail].next = j;
871 				tail = j;
872 			} else {
873 				int8_t *cur;
874 				struct hw_buf_info *t;
875 
876 				for (cur = &head; *cur != -1; cur = &t->next) {
877 					t = &s->hw_buf_info[*cur];
878 					if (hwb->size == t->size) {
879 						hwb->zidx = -2;
880 						break;
881 					}
882 					if (hwb->size > t->size) {
883 						hwb->next = *cur;
884 						*cur = j;
885 						break;
886 					}
887 				}
888 			}
889 		}
890 		swz->head_hwidx = head;
891 		swz->tail_hwidx = tail;
892 
893 		if (tail != -1) {
894 			n++;
895 			if (swz->size - s->hw_buf_info[tail].size >=
896 			    CL_METADATA_SIZE)
897 				sc->flags |= BUF_PACKING_OK;
898 		}
899 	}
900 	if (n == 0) {
901 		device_printf(sc->dev, "no usable SGE FL buffer size.\n");
902 		rc = EINVAL;
903 	}
904 
905 	s->safe_hwidx1 = -1;
906 	s->safe_hwidx2 = -1;
907 	if (safe_swz != NULL) {
908 		s->safe_hwidx1 = safe_swz->head_hwidx;
909 		for (i = safe_swz->head_hwidx; i != -1; i = hwb->next) {
910 			int spare;
911 
912 			hwb = &s->hw_buf_info[i];
913 #ifdef INVARIANTS
914 			if (fl_pad)
915 				MPASS(hwb->size % sp->pad_boundary == 0);
916 #endif
917 			spare = safe_swz->size - hwb->size;
918 			if (spare >= CL_METADATA_SIZE) {
919 				s->safe_hwidx2 = i;
920 				break;
921 			}
922 		}
923 	}
924 
925 	if (sc->flags & IS_VF)
926 		return (0);
927 
928 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
929 	r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ);
930 	if (r != v) {
931 		device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r);
932 		rc = EINVAL;
933 	}
934 
935 	m = v = F_TDDPTAGTCB;
936 	r = t4_read_reg(sc, A_ULP_RX_CTL);
937 	if ((r & m) != v) {
938 		device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r);
939 		rc = EINVAL;
940 	}
941 
942 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
943 	    F_RESETDDPOFFSET;
944 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
945 	r = t4_read_reg(sc, A_TP_PARA_REG5);
946 	if ((r & m) != v) {
947 		device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r);
948 		rc = EINVAL;
949 	}
950 
951 	t4_init_tp_params(sc, 1);
952 
953 	t4_read_mtu_tbl(sc, sc->params.mtus, NULL);
954 	t4_load_mtus(sc, sc->params.mtus, sc->params.a_wnd, sc->params.b_wnd);
955 
956 	return (rc);
957 }
958 
959 int
960 t4_create_dma_tag(struct adapter *sc)
961 {
962 	int rc;
963 
964 	rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
965 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
966 	    BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
967 	    NULL, &sc->dmat);
968 	if (rc != 0) {
969 		device_printf(sc->dev,
970 		    "failed to create main DMA tag: %d\n", rc);
971 	}
972 
973 	return (rc);
974 }
975 
976 void
977 t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
978     struct sysctl_oid_list *children)
979 {
980 	struct sge_params *sp = &sc->params.sge;
981 
982 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "buffer_sizes",
983 	    CTLTYPE_STRING | CTLFLAG_RD, &sc->sge, 0, sysctl_bufsizes, "A",
984 	    "freelist buffer sizes");
985 
986 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD,
987 	    NULL, sp->fl_pktshift, "payload DMA offset in rx buffer (bytes)");
988 
989 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD,
990 	    NULL, sp->pad_boundary, "payload pad boundary (bytes)");
991 
992 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD,
993 	    NULL, sp->spg_len, "status page size (bytes)");
994 
995 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD,
996 	    NULL, cong_drop, "congestion drop setting");
997 
998 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD,
999 	    NULL, sp->pack_boundary, "payload pack boundary (bytes)");
1000 }
1001 
1002 int
1003 t4_destroy_dma_tag(struct adapter *sc)
1004 {
1005 	if (sc->dmat)
1006 		bus_dma_tag_destroy(sc->dmat);
1007 
1008 	return (0);
1009 }
1010 
1011 /*
1012  * Allocate and initialize the firmware event queue, control queues, and special
1013  * purpose rx queues owned by the adapter.
1014  *
1015  * Returns errno on failure.  Resources allocated up to that point may still be
1016  * allocated.  Caller is responsible for cleanup in case this function fails.
1017  */
1018 int
1019 t4_setup_adapter_queues(struct adapter *sc)
1020 {
1021 	struct sysctl_oid *oid;
1022 	struct sysctl_oid_list *children;
1023 	int rc, i;
1024 
1025 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
1026 
1027 	sysctl_ctx_init(&sc->ctx);
1028 	sc->flags |= ADAP_SYSCTL_CTX;
1029 
1030 	/*
1031 	 * Firmware event queue
1032 	 */
1033 	rc = alloc_fwq(sc);
1034 	if (rc != 0)
1035 		return (rc);
1036 
1037 	/*
1038 	 * That's all for the VF driver.
1039 	 */
1040 	if (sc->flags & IS_VF)
1041 		return (rc);
1042 
1043 	oid = device_get_sysctl_tree(sc->dev);
1044 	children = SYSCTL_CHILDREN(oid);
1045 
1046 	/*
1047 	 * XXX: General purpose rx queues, one per port.
1048 	 */
1049 
1050 	/*
1051 	 * Control queues, one per port.
1052 	 */
1053 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "ctrlq",
1054 	    CTLFLAG_RD, NULL, "control queues");
1055 	for_each_port(sc, i) {
1056 		struct sge_wrq *ctrlq = &sc->sge.ctrlq[i];
1057 
1058 		rc = alloc_ctrlq(sc, ctrlq, i, oid);
1059 		if (rc != 0)
1060 			return (rc);
1061 	}
1062 
1063 	return (rc);
1064 }
1065 
1066 /*
1067  * Idempotent
1068  */
1069 int
1070 t4_teardown_adapter_queues(struct adapter *sc)
1071 {
1072 	int i;
1073 
1074 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
1075 
1076 	/* Do this before freeing the queue */
1077 	if (sc->flags & ADAP_SYSCTL_CTX) {
1078 		sysctl_ctx_free(&sc->ctx);
1079 		sc->flags &= ~ADAP_SYSCTL_CTX;
1080 	}
1081 
1082 	if (!(sc->flags & IS_VF)) {
1083 		for_each_port(sc, i)
1084 			free_wrq(sc, &sc->sge.ctrlq[i]);
1085 	}
1086 	free_fwq(sc);
1087 
1088 	return (0);
1089 }
1090 
1091 /* Maximum payload that can be delivered with a single iq descriptor */
1092 static inline int
1093 mtu_to_max_payload(struct adapter *sc, int mtu, const int toe)
1094 {
1095 	int payload;
1096 
1097 #ifdef TCP_OFFLOAD
1098 	if (toe) {
1099 		int rxcs = G_RXCOALESCESIZE(t4_read_reg(sc, A_TP_PARA_REG2));
1100 
1101 		/* Note that COP can set rx_coalesce on/off per connection. */
1102 		payload = max(mtu, rxcs);
1103 	} else {
1104 #endif
1105 		/* large enough even when hw VLAN extraction is disabled */
1106 		payload = sc->params.sge.fl_pktshift + ETHER_HDR_LEN +
1107 		    ETHER_VLAN_ENCAP_LEN + mtu;
1108 #ifdef TCP_OFFLOAD
1109 	}
1110 #endif
1111 
1112 	return (payload);
1113 }
1114 
1115 int
1116 t4_setup_vi_queues(struct vi_info *vi)
1117 {
1118 	int rc = 0, i, intr_idx, iqidx;
1119 	struct sge_rxq *rxq;
1120 	struct sge_txq *txq;
1121 #ifdef TCP_OFFLOAD
1122 	struct sge_ofld_rxq *ofld_rxq;
1123 #endif
1124 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1125 	struct sge_wrq *ofld_txq;
1126 #endif
1127 #ifdef DEV_NETMAP
1128 	int saved_idx;
1129 	struct sge_nm_rxq *nm_rxq;
1130 	struct sge_nm_txq *nm_txq;
1131 #endif
1132 	char name[16];
1133 	struct port_info *pi = vi->pi;
1134 	struct adapter *sc = pi->adapter;
1135 	struct ifnet *ifp = vi->ifp;
1136 	struct sysctl_oid *oid = device_get_sysctl_tree(vi->dev);
1137 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
1138 	int maxp, mtu = ifp->if_mtu;
1139 
1140 	/* Interrupt vector to start from (when using multiple vectors) */
1141 	intr_idx = vi->first_intr;
1142 
1143 #ifdef DEV_NETMAP
1144 	saved_idx = intr_idx;
1145 	if (ifp->if_capabilities & IFCAP_NETMAP) {
1146 
1147 		/* netmap is supported with direct interrupts only. */
1148 		MPASS(!forwarding_intr_to_fwq(sc));
1149 
1150 		/*
1151 		 * We don't have buffers to back the netmap rx queues
1152 		 * right now so we create the queues in a way that
1153 		 * doesn't set off any congestion signal in the chip.
1154 		 */
1155 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_rxq",
1156 		    CTLFLAG_RD, NULL, "rx queues");
1157 		for_each_nm_rxq(vi, i, nm_rxq) {
1158 			rc = alloc_nm_rxq(vi, nm_rxq, intr_idx, i, oid);
1159 			if (rc != 0)
1160 				goto done;
1161 			intr_idx++;
1162 		}
1163 
1164 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_txq",
1165 		    CTLFLAG_RD, NULL, "tx queues");
1166 		for_each_nm_txq(vi, i, nm_txq) {
1167 			iqidx = vi->first_nm_rxq + (i % vi->nnmrxq);
1168 			rc = alloc_nm_txq(vi, nm_txq, iqidx, i, oid);
1169 			if (rc != 0)
1170 				goto done;
1171 		}
1172 	}
1173 
1174 	/* Normal rx queues and netmap rx queues share the same interrupts. */
1175 	intr_idx = saved_idx;
1176 #endif
1177 
1178 	/*
1179 	 * Allocate rx queues first because a default iqid is required when
1180 	 * creating a tx queue.
1181 	 */
1182 	maxp = mtu_to_max_payload(sc, mtu, 0);
1183 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
1184 	    CTLFLAG_RD, NULL, "rx queues");
1185 	for_each_rxq(vi, i, rxq) {
1186 
1187 		init_iq(&rxq->iq, sc, vi->tmr_idx, vi->pktc_idx, vi->qsize_rxq);
1188 
1189 		snprintf(name, sizeof(name), "%s rxq%d-fl",
1190 		    device_get_nameunit(vi->dev), i);
1191 		init_fl(sc, &rxq->fl, vi->qsize_rxq / 8, maxp, name);
1192 
1193 		rc = alloc_rxq(vi, rxq,
1194 		    forwarding_intr_to_fwq(sc) ? -1 : intr_idx, i, oid);
1195 		if (rc != 0)
1196 			goto done;
1197 		intr_idx++;
1198 	}
1199 #ifdef DEV_NETMAP
1200 	if (ifp->if_capabilities & IFCAP_NETMAP)
1201 		intr_idx = saved_idx + max(vi->nrxq, vi->nnmrxq);
1202 #endif
1203 #ifdef TCP_OFFLOAD
1204 	maxp = mtu_to_max_payload(sc, mtu, 1);
1205 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq",
1206 	    CTLFLAG_RD, NULL, "rx queues for offloaded TCP connections");
1207 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1208 
1209 		init_iq(&ofld_rxq->iq, sc, vi->ofld_tmr_idx, vi->ofld_pktc_idx,
1210 		    vi->qsize_rxq);
1211 
1212 		snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
1213 		    device_get_nameunit(vi->dev), i);
1214 		init_fl(sc, &ofld_rxq->fl, vi->qsize_rxq / 8, maxp, name);
1215 
1216 		rc = alloc_ofld_rxq(vi, ofld_rxq,
1217 		    forwarding_intr_to_fwq(sc) ? -1 : intr_idx, i, oid);
1218 		if (rc != 0)
1219 			goto done;
1220 		intr_idx++;
1221 	}
1222 #endif
1223 
1224 	/*
1225 	 * Now the tx queues.
1226 	 */
1227 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "txq", CTLFLAG_RD,
1228 	    NULL, "tx queues");
1229 	for_each_txq(vi, i, txq) {
1230 		iqidx = vi->first_rxq + (i % vi->nrxq);
1231 		snprintf(name, sizeof(name), "%s txq%d",
1232 		    device_get_nameunit(vi->dev), i);
1233 		init_eq(sc, &txq->eq, EQ_ETH, vi->qsize_txq, pi->tx_chan,
1234 		    sc->sge.rxq[iqidx].iq.cntxt_id, name);
1235 
1236 		rc = alloc_txq(vi, txq, i, oid);
1237 		if (rc != 0)
1238 			goto done;
1239 	}
1240 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1241 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_txq",
1242 	    CTLFLAG_RD, NULL, "tx queues for TOE/ETHOFLD");
1243 	for_each_ofld_txq(vi, i, ofld_txq) {
1244 		struct sysctl_oid *oid2;
1245 
1246 		snprintf(name, sizeof(name), "%s ofld_txq%d",
1247 		    device_get_nameunit(vi->dev), i);
1248 		if (vi->nofldrxq > 0) {
1249 			iqidx = vi->first_ofld_rxq + (i % vi->nofldrxq);
1250 			init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq,
1251 			    pi->tx_chan, sc->sge.ofld_rxq[iqidx].iq.cntxt_id,
1252 			    name);
1253 		} else {
1254 			iqidx = vi->first_rxq + (i % vi->nrxq);
1255 			init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq,
1256 			    pi->tx_chan, sc->sge.rxq[iqidx].iq.cntxt_id, name);
1257 		}
1258 
1259 		snprintf(name, sizeof(name), "%d", i);
1260 		oid2 = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
1261 		    name, CTLFLAG_RD, NULL, "offload tx queue");
1262 
1263 		rc = alloc_wrq(sc, vi, ofld_txq, oid2);
1264 		if (rc != 0)
1265 			goto done;
1266 	}
1267 #endif
1268 done:
1269 	if (rc)
1270 		t4_teardown_vi_queues(vi);
1271 
1272 	return (rc);
1273 }
1274 
1275 /*
1276  * Idempotent
1277  */
1278 int
1279 t4_teardown_vi_queues(struct vi_info *vi)
1280 {
1281 	int i;
1282 	struct sge_rxq *rxq;
1283 	struct sge_txq *txq;
1284 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1285 	struct port_info *pi = vi->pi;
1286 	struct adapter *sc = pi->adapter;
1287 	struct sge_wrq *ofld_txq;
1288 #endif
1289 #ifdef TCP_OFFLOAD
1290 	struct sge_ofld_rxq *ofld_rxq;
1291 #endif
1292 #ifdef DEV_NETMAP
1293 	struct sge_nm_rxq *nm_rxq;
1294 	struct sge_nm_txq *nm_txq;
1295 #endif
1296 
1297 	/* Do this before freeing the queues */
1298 	if (vi->flags & VI_SYSCTL_CTX) {
1299 		sysctl_ctx_free(&vi->ctx);
1300 		vi->flags &= ~VI_SYSCTL_CTX;
1301 	}
1302 
1303 #ifdef DEV_NETMAP
1304 	if (vi->ifp->if_capabilities & IFCAP_NETMAP) {
1305 		for_each_nm_txq(vi, i, nm_txq) {
1306 			free_nm_txq(vi, nm_txq);
1307 		}
1308 
1309 		for_each_nm_rxq(vi, i, nm_rxq) {
1310 			free_nm_rxq(vi, nm_rxq);
1311 		}
1312 	}
1313 #endif
1314 
1315 	/*
1316 	 * Take down all the tx queues first, as they reference the rx queues
1317 	 * (for egress updates, etc.).
1318 	 */
1319 
1320 	for_each_txq(vi, i, txq) {
1321 		free_txq(vi, txq);
1322 	}
1323 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1324 	for_each_ofld_txq(vi, i, ofld_txq) {
1325 		free_wrq(sc, ofld_txq);
1326 	}
1327 #endif
1328 
1329 	/*
1330 	 * Then take down the rx queues.
1331 	 */
1332 
1333 	for_each_rxq(vi, i, rxq) {
1334 		free_rxq(vi, rxq);
1335 	}
1336 #ifdef TCP_OFFLOAD
1337 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1338 		free_ofld_rxq(vi, ofld_rxq);
1339 	}
1340 #endif
1341 
1342 	return (0);
1343 }
1344 
1345 /*
1346  * Interrupt handler when the driver is using only 1 interrupt.  This is a very
1347  * unusual scenario.
1348  *
1349  * a) Deals with errors, if any.
1350  * b) Services firmware event queue, which is taking interrupts for all other
1351  *    queues.
1352  */
1353 void
1354 t4_intr_all(void *arg)
1355 {
1356 	struct adapter *sc = arg;
1357 	struct sge_iq *fwq = &sc->sge.fwq;
1358 
1359 	MPASS(sc->intr_count == 1);
1360 
1361 	t4_intr_err(arg);
1362 	t4_intr_evt(fwq);
1363 }
1364 
1365 /*
1366  * Interrupt handler for errors (installed directly when multiple interrupts are
1367  * being used, or called by t4_intr_all).
1368  */
1369 void
1370 t4_intr_err(void *arg)
1371 {
1372 	struct adapter *sc = arg;
1373 
1374 	t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
1375 	t4_slow_intr_handler(sc);
1376 }
1377 
1378 /*
1379  * Interrupt handler for iq-only queues.  The firmware event queue is the only
1380  * such queue right now.
1381  */
1382 void
1383 t4_intr_evt(void *arg)
1384 {
1385 	struct sge_iq *iq = arg;
1386 
1387 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1388 		service_iq(iq, 0);
1389 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1390 	}
1391 }
1392 
1393 /*
1394  * Interrupt handler for iq+fl queues.
1395  */
1396 void
1397 t4_intr(void *arg)
1398 {
1399 	struct sge_iq *iq = arg;
1400 
1401 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1402 		service_iq_fl(iq, 0);
1403 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1404 	}
1405 }
1406 
1407 #ifdef DEV_NETMAP
1408 /*
1409  * Interrupt handler for netmap rx queues.
1410  */
1411 void
1412 t4_nm_intr(void *arg)
1413 {
1414 	struct sge_nm_rxq *nm_rxq = arg;
1415 
1416 	if (atomic_cmpset_int(&nm_rxq->nm_state, NM_ON, NM_BUSY)) {
1417 		service_nm_rxq(nm_rxq);
1418 		(void) atomic_cmpset_int(&nm_rxq->nm_state, NM_BUSY, NM_ON);
1419 	}
1420 }
1421 
1422 /*
1423  * Interrupt handler for vectors shared between NIC and netmap rx queues.
1424  */
1425 void
1426 t4_vi_intr(void *arg)
1427 {
1428 	struct irq *irq = arg;
1429 
1430 	MPASS(irq->nm_rxq != NULL);
1431 	t4_nm_intr(irq->nm_rxq);
1432 
1433 	MPASS(irq->rxq != NULL);
1434 	t4_intr(irq->rxq);
1435 }
1436 #endif
1437 
1438 /*
1439  * Deals with interrupts on an iq-only (no freelist) queue.
1440  */
1441 static int
1442 service_iq(struct sge_iq *iq, int budget)
1443 {
1444 	struct sge_iq *q;
1445 	struct adapter *sc = iq->adapter;
1446 	struct iq_desc *d = &iq->desc[iq->cidx];
1447 	int ndescs = 0, limit;
1448 	int rsp_type;
1449 	uint32_t lq;
1450 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1451 
1452 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1453 	KASSERT((iq->flags & IQ_HAS_FL) == 0,
1454 	    ("%s: called for iq %p with fl (iq->flags 0x%x)", __func__, iq,
1455 	    iq->flags));
1456 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1457 	MPASS((iq->flags & IQ_LRO_ENABLED) == 0);
1458 
1459 	limit = budget ? budget : iq->qsize / 16;
1460 
1461 	/*
1462 	 * We always come back and check the descriptor ring for new indirect
1463 	 * interrupts and other responses after running a single handler.
1464 	 */
1465 	for (;;) {
1466 		while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1467 
1468 			rmb();
1469 
1470 			rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1471 			lq = be32toh(d->rsp.pldbuflen_qid);
1472 
1473 			switch (rsp_type) {
1474 			case X_RSPD_TYPE_FLBUF:
1475 				panic("%s: data for an iq (%p) with no freelist",
1476 				    __func__, iq);
1477 
1478 				/* NOTREACHED */
1479 
1480 			case X_RSPD_TYPE_CPL:
1481 				KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1482 				    ("%s: bad opcode %02x.", __func__,
1483 				    d->rss.opcode));
1484 				t4_cpl_handler[d->rss.opcode](iq, &d->rss, NULL);
1485 				break;
1486 
1487 			case X_RSPD_TYPE_INTR:
1488 				/*
1489 				 * There are 1K interrupt-capable queues (qids 0
1490 				 * through 1023).  A response type indicating a
1491 				 * forwarded interrupt with a qid >= 1K is an
1492 				 * iWARP async notification.
1493 				 */
1494 				if (__predict_true(lq >= 1024)) {
1495 					t4_an_handler(iq, &d->rsp);
1496 					break;
1497 				}
1498 
1499 				q = sc->sge.iqmap[lq - sc->sge.iq_start -
1500 				    sc->sge.iq_base];
1501 				if (atomic_cmpset_int(&q->state, IQS_IDLE,
1502 				    IQS_BUSY)) {
1503 					if (service_iq_fl(q, q->qsize / 16) == 0) {
1504 						(void) atomic_cmpset_int(&q->state,
1505 						    IQS_BUSY, IQS_IDLE);
1506 					} else {
1507 						STAILQ_INSERT_TAIL(&iql, q,
1508 						    link);
1509 					}
1510 				}
1511 				break;
1512 
1513 			default:
1514 				KASSERT(0,
1515 				    ("%s: illegal response type %d on iq %p",
1516 				    __func__, rsp_type, iq));
1517 				log(LOG_ERR,
1518 				    "%s: illegal response type %d on iq %p",
1519 				    device_get_nameunit(sc->dev), rsp_type, iq);
1520 				break;
1521 			}
1522 
1523 			d++;
1524 			if (__predict_false(++iq->cidx == iq->sidx)) {
1525 				iq->cidx = 0;
1526 				iq->gen ^= F_RSPD_GEN;
1527 				d = &iq->desc[0];
1528 			}
1529 			if (__predict_false(++ndescs == limit)) {
1530 				t4_write_reg(sc, sc->sge_gts_reg,
1531 				    V_CIDXINC(ndescs) |
1532 				    V_INGRESSQID(iq->cntxt_id) |
1533 				    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1534 				ndescs = 0;
1535 
1536 				if (budget) {
1537 					return (EINPROGRESS);
1538 				}
1539 			}
1540 		}
1541 
1542 		if (STAILQ_EMPTY(&iql))
1543 			break;
1544 
1545 		/*
1546 		 * Process the head only, and send it to the back of the list if
1547 		 * it's still not done.
1548 		 */
1549 		q = STAILQ_FIRST(&iql);
1550 		STAILQ_REMOVE_HEAD(&iql, link);
1551 		if (service_iq_fl(q, q->qsize / 8) == 0)
1552 			(void) atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1553 		else
1554 			STAILQ_INSERT_TAIL(&iql, q, link);
1555 	}
1556 
1557 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1558 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1559 
1560 	return (0);
1561 }
1562 
1563 static inline int
1564 sort_before_lro(struct lro_ctrl *lro)
1565 {
1566 
1567 	return (lro->lro_mbuf_max != 0);
1568 }
1569 
1570 static inline uint64_t
1571 last_flit_to_ns(struct adapter *sc, uint64_t lf)
1572 {
1573 	uint64_t n = be64toh(lf) & 0xfffffffffffffff;	/* 60b, not 64b. */
1574 
1575 	if (n > UINT64_MAX / 1000000)
1576 		return (n / sc->params.vpd.cclk * 1000000);
1577 	else
1578 		return (n * 1000000 / sc->params.vpd.cclk);
1579 }
1580 
1581 /*
1582  * Deals with interrupts on an iq+fl queue.
1583  */
1584 static int
1585 service_iq_fl(struct sge_iq *iq, int budget)
1586 {
1587 	struct sge_rxq *rxq = iq_to_rxq(iq);
1588 	struct sge_fl *fl;
1589 	struct adapter *sc = iq->adapter;
1590 	struct iq_desc *d = &iq->desc[iq->cidx];
1591 	int ndescs = 0, limit;
1592 	int rsp_type, refill, starved;
1593 	uint32_t lq;
1594 	uint16_t fl_hw_cidx;
1595 	struct mbuf *m0;
1596 #if defined(INET) || defined(INET6)
1597 	const struct timeval lro_timeout = {0, sc->lro_timeout};
1598 	struct lro_ctrl *lro = &rxq->lro;
1599 #endif
1600 
1601 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1602 	MPASS(iq->flags & IQ_HAS_FL);
1603 
1604 	limit = budget ? budget : iq->qsize / 16;
1605 	fl = &rxq->fl;
1606 	fl_hw_cidx = fl->hw_cidx;	/* stable snapshot */
1607 
1608 #if defined(INET) || defined(INET6)
1609 	if (iq->flags & IQ_ADJ_CREDIT) {
1610 		MPASS(sort_before_lro(lro));
1611 		iq->flags &= ~IQ_ADJ_CREDIT;
1612 		if ((d->rsp.u.type_gen & F_RSPD_GEN) != iq->gen) {
1613 			tcp_lro_flush_all(lro);
1614 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(1) |
1615 			    V_INGRESSQID((u32)iq->cntxt_id) |
1616 			    V_SEINTARM(iq->intr_params));
1617 			return (0);
1618 		}
1619 		ndescs = 1;
1620 	}
1621 #else
1622 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1623 #endif
1624 
1625 	while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1626 
1627 		rmb();
1628 
1629 		refill = 0;
1630 		m0 = NULL;
1631 		rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1632 		lq = be32toh(d->rsp.pldbuflen_qid);
1633 
1634 		switch (rsp_type) {
1635 		case X_RSPD_TYPE_FLBUF:
1636 
1637 			m0 = get_fl_payload(sc, fl, lq);
1638 			if (__predict_false(m0 == NULL))
1639 				goto out;
1640 			refill = IDXDIFF(fl->hw_cidx, fl_hw_cidx, fl->sidx) > 2;
1641 
1642 			if (iq->flags & IQ_RX_TIMESTAMP) {
1643 				/*
1644 				 * Fill up rcv_tstmp but do not set M_TSTMP.
1645 				 * rcv_tstmp is not in the format that the
1646 				 * kernel expects and we don't want to mislead
1647 				 * it.  For now this is only for custom code
1648 				 * that knows how to interpret cxgbe's stamp.
1649 				 */
1650 				m0->m_pkthdr.rcv_tstmp =
1651 				    last_flit_to_ns(sc, d->rsp.u.last_flit);
1652 #ifdef notyet
1653 				m0->m_flags |= M_TSTMP;
1654 #endif
1655 			}
1656 
1657 			/* fall through */
1658 
1659 		case X_RSPD_TYPE_CPL:
1660 			KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1661 			    ("%s: bad opcode %02x.", __func__, d->rss.opcode));
1662 			t4_cpl_handler[d->rss.opcode](iq, &d->rss, m0);
1663 			break;
1664 
1665 		case X_RSPD_TYPE_INTR:
1666 
1667 			/*
1668 			 * There are 1K interrupt-capable queues (qids 0
1669 			 * through 1023).  A response type indicating a
1670 			 * forwarded interrupt with a qid >= 1K is an
1671 			 * iWARP async notification.  That is the only
1672 			 * acceptable indirect interrupt on this queue.
1673 			 */
1674 			if (__predict_false(lq < 1024)) {
1675 				panic("%s: indirect interrupt on iq_fl %p "
1676 				    "with qid %u", __func__, iq, lq);
1677 			}
1678 
1679 			t4_an_handler(iq, &d->rsp);
1680 			break;
1681 
1682 		default:
1683 			KASSERT(0, ("%s: illegal response type %d on iq %p",
1684 			    __func__, rsp_type, iq));
1685 			log(LOG_ERR, "%s: illegal response type %d on iq %p",
1686 			    device_get_nameunit(sc->dev), rsp_type, iq);
1687 			break;
1688 		}
1689 
1690 		d++;
1691 		if (__predict_false(++iq->cidx == iq->sidx)) {
1692 			iq->cidx = 0;
1693 			iq->gen ^= F_RSPD_GEN;
1694 			d = &iq->desc[0];
1695 		}
1696 		if (__predict_false(++ndescs == limit)) {
1697 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1698 			    V_INGRESSQID(iq->cntxt_id) |
1699 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1700 			ndescs = 0;
1701 
1702 #if defined(INET) || defined(INET6)
1703 			if (iq->flags & IQ_LRO_ENABLED &&
1704 			    !sort_before_lro(lro) &&
1705 			    sc->lro_timeout != 0) {
1706 				tcp_lro_flush_inactive(lro, &lro_timeout);
1707 			}
1708 #endif
1709 			if (budget) {
1710 				FL_LOCK(fl);
1711 				refill_fl(sc, fl, 32);
1712 				FL_UNLOCK(fl);
1713 
1714 				return (EINPROGRESS);
1715 			}
1716 		}
1717 		if (refill) {
1718 			FL_LOCK(fl);
1719 			refill_fl(sc, fl, 32);
1720 			FL_UNLOCK(fl);
1721 			fl_hw_cidx = fl->hw_cidx;
1722 		}
1723 	}
1724 out:
1725 #if defined(INET) || defined(INET6)
1726 	if (iq->flags & IQ_LRO_ENABLED) {
1727 		if (ndescs > 0 && lro->lro_mbuf_count > 8) {
1728 			MPASS(sort_before_lro(lro));
1729 			/* hold back one credit and don't flush LRO state */
1730 			iq->flags |= IQ_ADJ_CREDIT;
1731 			ndescs--;
1732 		} else {
1733 			tcp_lro_flush_all(lro);
1734 		}
1735 	}
1736 #endif
1737 
1738 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1739 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1740 
1741 	FL_LOCK(fl);
1742 	starved = refill_fl(sc, fl, 64);
1743 	FL_UNLOCK(fl);
1744 	if (__predict_false(starved != 0))
1745 		add_fl_to_sfl(sc, fl);
1746 
1747 	return (0);
1748 }
1749 
1750 static inline int
1751 cl_has_metadata(struct sge_fl *fl, struct cluster_layout *cll)
1752 {
1753 	int rc = fl->flags & FL_BUF_PACKING || cll->region1 > 0;
1754 
1755 	if (rc)
1756 		MPASS(cll->region3 >= CL_METADATA_SIZE);
1757 
1758 	return (rc);
1759 }
1760 
1761 static inline struct cluster_metadata *
1762 cl_metadata(struct adapter *sc, struct sge_fl *fl, struct cluster_layout *cll,
1763     caddr_t cl)
1764 {
1765 
1766 	if (cl_has_metadata(fl, cll)) {
1767 		struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1768 
1769 		return ((struct cluster_metadata *)(cl + swz->size) - 1);
1770 	}
1771 	return (NULL);
1772 }
1773 
1774 static void
1775 rxb_free(struct mbuf *m)
1776 {
1777 	uma_zone_t zone = m->m_ext.ext_arg1;
1778 	void *cl = m->m_ext.ext_arg2;
1779 
1780 	uma_zfree(zone, cl);
1781 	counter_u64_add(extfree_rels, 1);
1782 }
1783 
1784 /*
1785  * The mbuf returned by this function could be allocated from zone_mbuf or
1786  * constructed in spare room in the cluster.
1787  *
1788  * The mbuf carries the payload in one of these ways
1789  * a) frame inside the mbuf (mbuf from zone_mbuf)
1790  * b) m_cljset (for clusters without metadata) zone_mbuf
1791  * c) m_extaddref (cluster with metadata) inline mbuf
1792  * d) m_extaddref (cluster with metadata) zone_mbuf
1793  */
1794 static struct mbuf *
1795 get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1796     int remaining)
1797 {
1798 	struct mbuf *m;
1799 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1800 	struct cluster_layout *cll = &sd->cll;
1801 	struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1802 	struct hw_buf_info *hwb = &sc->sge.hw_buf_info[cll->hwidx];
1803 	struct cluster_metadata *clm = cl_metadata(sc, fl, cll, sd->cl);
1804 	int len, blen;
1805 	caddr_t payload;
1806 
1807 	blen = hwb->size - fl->rx_offset;	/* max possible in this buf */
1808 	len = min(remaining, blen);
1809 	payload = sd->cl + cll->region1 + fl->rx_offset;
1810 	if (fl->flags & FL_BUF_PACKING) {
1811 		const u_int l = fr_offset + len;
1812 		const u_int pad = roundup2(l, fl->buf_boundary) - l;
1813 
1814 		if (fl->rx_offset + len + pad < hwb->size)
1815 			blen = len + pad;
1816 		MPASS(fl->rx_offset + blen <= hwb->size);
1817 	} else {
1818 		MPASS(fl->rx_offset == 0);	/* not packing */
1819 	}
1820 
1821 
1822 	if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) {
1823 
1824 		/*
1825 		 * Copy payload into a freshly allocated mbuf.
1826 		 */
1827 
1828 		m = fr_offset == 0 ?
1829 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1830 		if (m == NULL)
1831 			return (NULL);
1832 		fl->mbuf_allocated++;
1833 
1834 		/* copy data to mbuf */
1835 		bcopy(payload, mtod(m, caddr_t), len);
1836 
1837 	} else if (sd->nmbuf * MSIZE < cll->region1) {
1838 
1839 		/*
1840 		 * There's spare room in the cluster for an mbuf.  Create one
1841 		 * and associate it with the payload that's in the cluster.
1842 		 */
1843 
1844 		MPASS(clm != NULL);
1845 		m = (struct mbuf *)(sd->cl + sd->nmbuf * MSIZE);
1846 		/* No bzero required */
1847 		if (m_init(m, M_NOWAIT, MT_DATA,
1848 		    fr_offset == 0 ? M_PKTHDR | M_NOFREE : M_NOFREE))
1849 			return (NULL);
1850 		fl->mbuf_inlined++;
1851 		m_extaddref(m, payload, blen, &clm->refcount, rxb_free,
1852 		    swz->zone, sd->cl);
1853 		if (sd->nmbuf++ == 0)
1854 			counter_u64_add(extfree_refs, 1);
1855 
1856 	} else {
1857 
1858 		/*
1859 		 * Grab an mbuf from zone_mbuf and associate it with the
1860 		 * payload in the cluster.
1861 		 */
1862 
1863 		m = fr_offset == 0 ?
1864 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1865 		if (m == NULL)
1866 			return (NULL);
1867 		fl->mbuf_allocated++;
1868 		if (clm != NULL) {
1869 			m_extaddref(m, payload, blen, &clm->refcount,
1870 			    rxb_free, swz->zone, sd->cl);
1871 			if (sd->nmbuf++ == 0)
1872 				counter_u64_add(extfree_refs, 1);
1873 		} else {
1874 			m_cljset(m, sd->cl, swz->type);
1875 			sd->cl = NULL;	/* consumed, not a recycle candidate */
1876 		}
1877 	}
1878 	if (fr_offset == 0)
1879 		m->m_pkthdr.len = remaining;
1880 	m->m_len = len;
1881 
1882 	if (fl->flags & FL_BUF_PACKING) {
1883 		fl->rx_offset += blen;
1884 		MPASS(fl->rx_offset <= hwb->size);
1885 		if (fl->rx_offset < hwb->size)
1886 			return (m);	/* without advancing the cidx */
1887 	}
1888 
1889 	if (__predict_false(++fl->cidx % 8 == 0)) {
1890 		uint16_t cidx = fl->cidx / 8;
1891 
1892 		if (__predict_false(cidx == fl->sidx))
1893 			fl->cidx = cidx = 0;
1894 		fl->hw_cidx = cidx;
1895 	}
1896 	fl->rx_offset = 0;
1897 
1898 	return (m);
1899 }
1900 
1901 static struct mbuf *
1902 get_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf)
1903 {
1904 	struct mbuf *m0, *m, **pnext;
1905 	u_int remaining;
1906 	const u_int total = G_RSPD_LEN(len_newbuf);
1907 
1908 	if (__predict_false(fl->flags & FL_BUF_RESUME)) {
1909 		M_ASSERTPKTHDR(fl->m0);
1910 		MPASS(fl->m0->m_pkthdr.len == total);
1911 		MPASS(fl->remaining < total);
1912 
1913 		m0 = fl->m0;
1914 		pnext = fl->pnext;
1915 		remaining = fl->remaining;
1916 		fl->flags &= ~FL_BUF_RESUME;
1917 		goto get_segment;
1918 	}
1919 
1920 	if (fl->rx_offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
1921 		fl->rx_offset = 0;
1922 		if (__predict_false(++fl->cidx % 8 == 0)) {
1923 			uint16_t cidx = fl->cidx / 8;
1924 
1925 			if (__predict_false(cidx == fl->sidx))
1926 				fl->cidx = cidx = 0;
1927 			fl->hw_cidx = cidx;
1928 		}
1929 	}
1930 
1931 	/*
1932 	 * Payload starts at rx_offset in the current hw buffer.  Its length is
1933 	 * 'len' and it may span multiple hw buffers.
1934 	 */
1935 
1936 	m0 = get_scatter_segment(sc, fl, 0, total);
1937 	if (m0 == NULL)
1938 		return (NULL);
1939 	remaining = total - m0->m_len;
1940 	pnext = &m0->m_next;
1941 	while (remaining > 0) {
1942 get_segment:
1943 		MPASS(fl->rx_offset == 0);
1944 		m = get_scatter_segment(sc, fl, total - remaining, remaining);
1945 		if (__predict_false(m == NULL)) {
1946 			fl->m0 = m0;
1947 			fl->pnext = pnext;
1948 			fl->remaining = remaining;
1949 			fl->flags |= FL_BUF_RESUME;
1950 			return (NULL);
1951 		}
1952 		*pnext = m;
1953 		pnext = &m->m_next;
1954 		remaining -= m->m_len;
1955 	}
1956 	*pnext = NULL;
1957 
1958 	M_ASSERTPKTHDR(m0);
1959 	return (m0);
1960 }
1961 
1962 static int
1963 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
1964 {
1965 	struct sge_rxq *rxq = iq_to_rxq(iq);
1966 	struct ifnet *ifp = rxq->ifp;
1967 	struct adapter *sc = iq->adapter;
1968 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
1969 #if defined(INET) || defined(INET6)
1970 	struct lro_ctrl *lro = &rxq->lro;
1971 #endif
1972 	static const int sw_hashtype[4][2] = {
1973 		{M_HASHTYPE_NONE, M_HASHTYPE_NONE},
1974 		{M_HASHTYPE_RSS_IPV4, M_HASHTYPE_RSS_IPV6},
1975 		{M_HASHTYPE_RSS_TCP_IPV4, M_HASHTYPE_RSS_TCP_IPV6},
1976 		{M_HASHTYPE_RSS_UDP_IPV4, M_HASHTYPE_RSS_UDP_IPV6},
1977 	};
1978 
1979 	KASSERT(m0 != NULL, ("%s: no payload with opcode %02x", __func__,
1980 	    rss->opcode));
1981 
1982 	m0->m_pkthdr.len -= sc->params.sge.fl_pktshift;
1983 	m0->m_len -= sc->params.sge.fl_pktshift;
1984 	m0->m_data += sc->params.sge.fl_pktshift;
1985 
1986 	m0->m_pkthdr.rcvif = ifp;
1987 	M_HASHTYPE_SET(m0, sw_hashtype[rss->hash_type][rss->ipv6]);
1988 	m0->m_pkthdr.flowid = be32toh(rss->hash_val);
1989 
1990 	if (cpl->csum_calc && !(cpl->err_vec & sc->params.tp.err_vec_mask)) {
1991 		if (ifp->if_capenable & IFCAP_RXCSUM &&
1992 		    cpl->l2info & htobe32(F_RXF_IP)) {
1993 			m0->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
1994 			    CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
1995 			rxq->rxcsum++;
1996 		} else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
1997 		    cpl->l2info & htobe32(F_RXF_IP6)) {
1998 			m0->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
1999 			    CSUM_PSEUDO_HDR);
2000 			rxq->rxcsum++;
2001 		}
2002 
2003 		if (__predict_false(cpl->ip_frag))
2004 			m0->m_pkthdr.csum_data = be16toh(cpl->csum);
2005 		else
2006 			m0->m_pkthdr.csum_data = 0xffff;
2007 	}
2008 
2009 	if (cpl->vlan_ex) {
2010 		m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
2011 		m0->m_flags |= M_VLANTAG;
2012 		rxq->vlan_extraction++;
2013 	}
2014 
2015 #if defined(INET) || defined(INET6)
2016 	if (iq->flags & IQ_LRO_ENABLED) {
2017 		if (sort_before_lro(lro)) {
2018 			tcp_lro_queue_mbuf(lro, m0);
2019 			return (0); /* queued for sort, then LRO */
2020 		}
2021 		if (tcp_lro_rx(lro, m0, 0) == 0)
2022 			return (0); /* queued for LRO */
2023 	}
2024 #endif
2025 	ifp->if_input(ifp, m0);
2026 
2027 	return (0);
2028 }
2029 
2030 /*
2031  * Must drain the wrq or make sure that someone else will.
2032  */
2033 static void
2034 wrq_tx_drain(void *arg, int n)
2035 {
2036 	struct sge_wrq *wrq = arg;
2037 	struct sge_eq *eq = &wrq->eq;
2038 
2039 	EQ_LOCK(eq);
2040 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2041 		drain_wrq_wr_list(wrq->adapter, wrq);
2042 	EQ_UNLOCK(eq);
2043 }
2044 
2045 static void
2046 drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq)
2047 {
2048 	struct sge_eq *eq = &wrq->eq;
2049 	u_int available, dbdiff;	/* # of hardware descriptors */
2050 	u_int n;
2051 	struct wrqe *wr;
2052 	struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
2053 
2054 	EQ_LOCK_ASSERT_OWNED(eq);
2055 	MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
2056 	wr = STAILQ_FIRST(&wrq->wr_list);
2057 	MPASS(wr != NULL);	/* Must be called with something useful to do */
2058 	MPASS(eq->pidx == eq->dbidx);
2059 	dbdiff = 0;
2060 
2061 	do {
2062 		eq->cidx = read_hw_cidx(eq);
2063 		if (eq->pidx == eq->cidx)
2064 			available = eq->sidx - 1;
2065 		else
2066 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2067 
2068 		MPASS(wr->wrq == wrq);
2069 		n = howmany(wr->wr_len, EQ_ESIZE);
2070 		if (available < n)
2071 			break;
2072 
2073 		dst = (void *)&eq->desc[eq->pidx];
2074 		if (__predict_true(eq->sidx - eq->pidx > n)) {
2075 			/* Won't wrap, won't end exactly at the status page. */
2076 			bcopy(&wr->wr[0], dst, wr->wr_len);
2077 			eq->pidx += n;
2078 		} else {
2079 			int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE;
2080 
2081 			bcopy(&wr->wr[0], dst, first_portion);
2082 			if (wr->wr_len > first_portion) {
2083 				bcopy(&wr->wr[first_portion], &eq->desc[0],
2084 				    wr->wr_len - first_portion);
2085 			}
2086 			eq->pidx = n - (eq->sidx - eq->pidx);
2087 		}
2088 		wrq->tx_wrs_copied++;
2089 
2090 		if (available < eq->sidx / 4 &&
2091 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2092 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2093 			    F_FW_WR_EQUEQ);
2094 			eq->equeqidx = eq->pidx;
2095 		} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
2096 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
2097 			eq->equeqidx = eq->pidx;
2098 		}
2099 
2100 		dbdiff += n;
2101 		if (dbdiff >= 16) {
2102 			ring_eq_db(sc, eq, dbdiff);
2103 			dbdiff = 0;
2104 		}
2105 
2106 		STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
2107 		free_wrqe(wr);
2108 		MPASS(wrq->nwr_pending > 0);
2109 		wrq->nwr_pending--;
2110 		MPASS(wrq->ndesc_needed >= n);
2111 		wrq->ndesc_needed -= n;
2112 	} while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL);
2113 
2114 	if (dbdiff)
2115 		ring_eq_db(sc, eq, dbdiff);
2116 }
2117 
2118 /*
2119  * Doesn't fail.  Holds on to work requests it can't send right away.
2120  */
2121 void
2122 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
2123 {
2124 #ifdef INVARIANTS
2125 	struct sge_eq *eq = &wrq->eq;
2126 #endif
2127 
2128 	EQ_LOCK_ASSERT_OWNED(eq);
2129 	MPASS(wr != NULL);
2130 	MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN);
2131 	MPASS((wr->wr_len & 0x7) == 0);
2132 
2133 	STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
2134 	wrq->nwr_pending++;
2135 	wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE);
2136 
2137 	if (!TAILQ_EMPTY(&wrq->incomplete_wrs))
2138 		return;	/* commit_wrq_wr will drain wr_list as well. */
2139 
2140 	drain_wrq_wr_list(sc, wrq);
2141 
2142 	/* Doorbell must have caught up to the pidx. */
2143 	MPASS(eq->pidx == eq->dbidx);
2144 }
2145 
2146 void
2147 t4_update_fl_bufsize(struct ifnet *ifp)
2148 {
2149 	struct vi_info *vi = ifp->if_softc;
2150 	struct adapter *sc = vi->pi->adapter;
2151 	struct sge_rxq *rxq;
2152 #ifdef TCP_OFFLOAD
2153 	struct sge_ofld_rxq *ofld_rxq;
2154 #endif
2155 	struct sge_fl *fl;
2156 	int i, maxp, mtu = ifp->if_mtu;
2157 
2158 	maxp = mtu_to_max_payload(sc, mtu, 0);
2159 	for_each_rxq(vi, i, rxq) {
2160 		fl = &rxq->fl;
2161 
2162 		FL_LOCK(fl);
2163 		find_best_refill_source(sc, fl, maxp);
2164 		FL_UNLOCK(fl);
2165 	}
2166 #ifdef TCP_OFFLOAD
2167 	maxp = mtu_to_max_payload(sc, mtu, 1);
2168 	for_each_ofld_rxq(vi, i, ofld_rxq) {
2169 		fl = &ofld_rxq->fl;
2170 
2171 		FL_LOCK(fl);
2172 		find_best_refill_source(sc, fl, maxp);
2173 		FL_UNLOCK(fl);
2174 	}
2175 #endif
2176 }
2177 
2178 static inline int
2179 mbuf_nsegs(struct mbuf *m)
2180 {
2181 
2182 	M_ASSERTPKTHDR(m);
2183 	KASSERT(m->m_pkthdr.l5hlen > 0,
2184 	    ("%s: mbuf %p missing information on # of segments.", __func__, m));
2185 
2186 	return (m->m_pkthdr.l5hlen);
2187 }
2188 
2189 static inline void
2190 set_mbuf_nsegs(struct mbuf *m, uint8_t nsegs)
2191 {
2192 
2193 	M_ASSERTPKTHDR(m);
2194 	m->m_pkthdr.l5hlen = nsegs;
2195 }
2196 
2197 static inline int
2198 mbuf_len16(struct mbuf *m)
2199 {
2200 	int n;
2201 
2202 	M_ASSERTPKTHDR(m);
2203 	n = m->m_pkthdr.PH_loc.eight[0];
2204 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2205 
2206 	return (n);
2207 }
2208 
2209 static inline void
2210 set_mbuf_len16(struct mbuf *m, uint8_t len16)
2211 {
2212 
2213 	M_ASSERTPKTHDR(m);
2214 	m->m_pkthdr.PH_loc.eight[0] = len16;
2215 }
2216 
2217 #ifdef RATELIMIT
2218 static inline int
2219 mbuf_eo_nsegs(struct mbuf *m)
2220 {
2221 
2222 	M_ASSERTPKTHDR(m);
2223 	return (m->m_pkthdr.PH_loc.eight[1]);
2224 }
2225 
2226 static inline void
2227 set_mbuf_eo_nsegs(struct mbuf *m, uint8_t nsegs)
2228 {
2229 
2230 	M_ASSERTPKTHDR(m);
2231 	m->m_pkthdr.PH_loc.eight[1] = nsegs;
2232 }
2233 
2234 static inline int
2235 mbuf_eo_len16(struct mbuf *m)
2236 {
2237 	int n;
2238 
2239 	M_ASSERTPKTHDR(m);
2240 	n = m->m_pkthdr.PH_loc.eight[2];
2241 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2242 
2243 	return (n);
2244 }
2245 
2246 static inline void
2247 set_mbuf_eo_len16(struct mbuf *m, uint8_t len16)
2248 {
2249 
2250 	M_ASSERTPKTHDR(m);
2251 	m->m_pkthdr.PH_loc.eight[2] = len16;
2252 }
2253 
2254 static inline int
2255 mbuf_eo_tsclk_tsoff(struct mbuf *m)
2256 {
2257 
2258 	M_ASSERTPKTHDR(m);
2259 	return (m->m_pkthdr.PH_loc.eight[3]);
2260 }
2261 
2262 static inline void
2263 set_mbuf_eo_tsclk_tsoff(struct mbuf *m, uint8_t tsclk_tsoff)
2264 {
2265 
2266 	M_ASSERTPKTHDR(m);
2267 	m->m_pkthdr.PH_loc.eight[3] = tsclk_tsoff;
2268 }
2269 
2270 static inline int
2271 needs_eo(struct mbuf *m)
2272 {
2273 
2274 	return (m->m_pkthdr.snd_tag != NULL);
2275 }
2276 #endif
2277 
2278 static inline int
2279 needs_tso(struct mbuf *m)
2280 {
2281 
2282 	M_ASSERTPKTHDR(m);
2283 
2284 	return (m->m_pkthdr.csum_flags & CSUM_TSO);
2285 }
2286 
2287 static inline int
2288 needs_l3_csum(struct mbuf *m)
2289 {
2290 
2291 	M_ASSERTPKTHDR(m);
2292 
2293 	return (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO));
2294 }
2295 
2296 static inline int
2297 needs_l4_csum(struct mbuf *m)
2298 {
2299 
2300 	M_ASSERTPKTHDR(m);
2301 
2302 	return (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
2303 	    CSUM_TCP_IPV6 | CSUM_TSO));
2304 }
2305 
2306 static inline int
2307 needs_tcp_csum(struct mbuf *m)
2308 {
2309 
2310 	M_ASSERTPKTHDR(m);
2311 	return (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_TCP_IPV6 | CSUM_TSO));
2312 }
2313 
2314 #ifdef RATELIMIT
2315 static inline int
2316 needs_udp_csum(struct mbuf *m)
2317 {
2318 
2319 	M_ASSERTPKTHDR(m);
2320 	return (m->m_pkthdr.csum_flags & (CSUM_UDP | CSUM_UDP_IPV6));
2321 }
2322 #endif
2323 
2324 static inline int
2325 needs_vlan_insertion(struct mbuf *m)
2326 {
2327 
2328 	M_ASSERTPKTHDR(m);
2329 
2330 	return (m->m_flags & M_VLANTAG);
2331 }
2332 
2333 static void *
2334 m_advance(struct mbuf **pm, int *poffset, int len)
2335 {
2336 	struct mbuf *m = *pm;
2337 	int offset = *poffset;
2338 	uintptr_t p = 0;
2339 
2340 	MPASS(len > 0);
2341 
2342 	for (;;) {
2343 		if (offset + len < m->m_len) {
2344 			offset += len;
2345 			p = mtod(m, uintptr_t) + offset;
2346 			break;
2347 		}
2348 		len -= m->m_len - offset;
2349 		m = m->m_next;
2350 		offset = 0;
2351 		MPASS(m != NULL);
2352 	}
2353 	*poffset = offset;
2354 	*pm = m;
2355 	return ((void *)p);
2356 }
2357 
2358 /*
2359  * Can deal with empty mbufs in the chain that have m_len = 0, but the chain
2360  * must have at least one mbuf that's not empty.  It is possible for this
2361  * routine to return 0 if skip accounts for all the contents of the mbuf chain.
2362  */
2363 static inline int
2364 count_mbuf_nsegs(struct mbuf *m, int skip)
2365 {
2366 	vm_paddr_t lastb, next;
2367 	vm_offset_t va;
2368 	int len, nsegs;
2369 
2370 	M_ASSERTPKTHDR(m);
2371 	MPASS(m->m_pkthdr.len > 0);
2372 	MPASS(m->m_pkthdr.len >= skip);
2373 
2374 	nsegs = 0;
2375 	lastb = 0;
2376 	for (; m; m = m->m_next) {
2377 
2378 		len = m->m_len;
2379 		if (__predict_false(len == 0))
2380 			continue;
2381 		if (skip >= len) {
2382 			skip -= len;
2383 			continue;
2384 		}
2385 		va = mtod(m, vm_offset_t) + skip;
2386 		len -= skip;
2387 		skip = 0;
2388 		next = pmap_kextract(va);
2389 		nsegs += sglist_count((void *)(uintptr_t)va, len);
2390 		if (lastb + 1 == next)
2391 			nsegs--;
2392 		lastb = pmap_kextract(va + len - 1);
2393 	}
2394 
2395 	return (nsegs);
2396 }
2397 
2398 /*
2399  * Analyze the mbuf to determine its tx needs.  The mbuf passed in may change:
2400  * a) caller can assume it's been freed if this function returns with an error.
2401  * b) it may get defragged up if the gather list is too long for the hardware.
2402  */
2403 int
2404 parse_pkt(struct adapter *sc, struct mbuf **mp)
2405 {
2406 	struct mbuf *m0 = *mp, *m;
2407 	int rc, nsegs, defragged = 0, offset;
2408 	struct ether_header *eh;
2409 	void *l3hdr;
2410 #if defined(INET) || defined(INET6)
2411 	struct tcphdr *tcp;
2412 #endif
2413 	uint16_t eh_type;
2414 
2415 	M_ASSERTPKTHDR(m0);
2416 	if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) {
2417 		rc = EINVAL;
2418 fail:
2419 		m_freem(m0);
2420 		*mp = NULL;
2421 		return (rc);
2422 	}
2423 restart:
2424 	/*
2425 	 * First count the number of gather list segments in the payload.
2426 	 * Defrag the mbuf if nsegs exceeds the hardware limit.
2427 	 */
2428 	M_ASSERTPKTHDR(m0);
2429 	MPASS(m0->m_pkthdr.len > 0);
2430 	nsegs = count_mbuf_nsegs(m0, 0);
2431 	if (nsegs > (needs_tso(m0) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS)) {
2432 		if (defragged++ > 0 || (m = m_defrag(m0, M_NOWAIT)) == NULL) {
2433 			rc = EFBIG;
2434 			goto fail;
2435 		}
2436 		*mp = m0 = m;	/* update caller's copy after defrag */
2437 		goto restart;
2438 	}
2439 
2440 	if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN)) {
2441 		m0 = m_pullup(m0, m0->m_pkthdr.len);
2442 		if (m0 == NULL) {
2443 			/* Should have left well enough alone. */
2444 			rc = EFBIG;
2445 			goto fail;
2446 		}
2447 		*mp = m0;	/* update caller's copy after pullup */
2448 		goto restart;
2449 	}
2450 	set_mbuf_nsegs(m0, nsegs);
2451 	if (sc->flags & IS_VF)
2452 		set_mbuf_len16(m0, txpkt_vm_len16(nsegs, needs_tso(m0)));
2453 	else
2454 		set_mbuf_len16(m0, txpkt_len16(nsegs, needs_tso(m0)));
2455 
2456 #ifdef RATELIMIT
2457 	/*
2458 	 * Ethofld is limited to TCP and UDP for now, and only when L4 hw
2459 	 * checksumming is enabled.  needs_l4_csum happens to check for all the
2460 	 * right things.
2461 	 */
2462 	if (__predict_false(needs_eo(m0) && !needs_l4_csum(m0)))
2463 		m0->m_pkthdr.snd_tag = NULL;
2464 #endif
2465 
2466 	if (!needs_tso(m0) &&
2467 #ifdef RATELIMIT
2468 	    !needs_eo(m0) &&
2469 #endif
2470 	    !(sc->flags & IS_VF && (needs_l3_csum(m0) || needs_l4_csum(m0))))
2471 		return (0);
2472 
2473 	m = m0;
2474 	eh = mtod(m, struct ether_header *);
2475 	eh_type = ntohs(eh->ether_type);
2476 	if (eh_type == ETHERTYPE_VLAN) {
2477 		struct ether_vlan_header *evh = (void *)eh;
2478 
2479 		eh_type = ntohs(evh->evl_proto);
2480 		m0->m_pkthdr.l2hlen = sizeof(*evh);
2481 	} else
2482 		m0->m_pkthdr.l2hlen = sizeof(*eh);
2483 
2484 	offset = 0;
2485 	l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2486 
2487 	switch (eh_type) {
2488 #ifdef INET6
2489 	case ETHERTYPE_IPV6:
2490 	{
2491 		struct ip6_hdr *ip6 = l3hdr;
2492 
2493 		MPASS(!needs_tso(m0) || ip6->ip6_nxt == IPPROTO_TCP);
2494 
2495 		m0->m_pkthdr.l3hlen = sizeof(*ip6);
2496 		break;
2497 	}
2498 #endif
2499 #ifdef INET
2500 	case ETHERTYPE_IP:
2501 	{
2502 		struct ip *ip = l3hdr;
2503 
2504 		m0->m_pkthdr.l3hlen = ip->ip_hl * 4;
2505 		break;
2506 	}
2507 #endif
2508 	default:
2509 		panic("%s: ethertype 0x%04x unknown.  if_cxgbe must be compiled"
2510 		    " with the same INET/INET6 options as the kernel.",
2511 		    __func__, eh_type);
2512 	}
2513 
2514 #if defined(INET) || defined(INET6)
2515 	if (needs_tcp_csum(m0)) {
2516 		tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen);
2517 		m0->m_pkthdr.l4hlen = tcp->th_off * 4;
2518 #ifdef RATELIMIT
2519 		if (tsclk >= 0 && *(uint32_t *)(tcp + 1) == ntohl(0x0101080a)) {
2520 			set_mbuf_eo_tsclk_tsoff(m0,
2521 			    V_FW_ETH_TX_EO_WR_TSCLK(tsclk) |
2522 			    V_FW_ETH_TX_EO_WR_TSOFF(sizeof(*tcp) / 2 + 1));
2523 		} else
2524 			set_mbuf_eo_tsclk_tsoff(m0, 0);
2525 	} else if (needs_udp_csum(m)) {
2526 		m0->m_pkthdr.l4hlen = sizeof(struct udphdr);
2527 #endif
2528 	}
2529 #ifdef RATELIMIT
2530 	if (needs_eo(m0)) {
2531 		u_int immhdrs;
2532 
2533 		/* EO WRs have the headers in the WR and not the GL. */
2534 		immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen +
2535 		    m0->m_pkthdr.l4hlen;
2536 		nsegs = count_mbuf_nsegs(m0, immhdrs);
2537 		set_mbuf_eo_nsegs(m0, nsegs);
2538 		set_mbuf_eo_len16(m0,
2539 		    txpkt_eo_len16(nsegs, immhdrs, needs_tso(m0)));
2540 	}
2541 #endif
2542 #endif
2543 	MPASS(m0 == *mp);
2544 	return (0);
2545 }
2546 
2547 void *
2548 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie)
2549 {
2550 	struct sge_eq *eq = &wrq->eq;
2551 	struct adapter *sc = wrq->adapter;
2552 	int ndesc, available;
2553 	struct wrqe *wr;
2554 	void *w;
2555 
2556 	MPASS(len16 > 0);
2557 	ndesc = howmany(len16, EQ_ESIZE / 16);
2558 	MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC);
2559 
2560 	EQ_LOCK(eq);
2561 
2562 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2563 		drain_wrq_wr_list(sc, wrq);
2564 
2565 	if (!STAILQ_EMPTY(&wrq->wr_list)) {
2566 slowpath:
2567 		EQ_UNLOCK(eq);
2568 		wr = alloc_wrqe(len16 * 16, wrq);
2569 		if (__predict_false(wr == NULL))
2570 			return (NULL);
2571 		cookie->pidx = -1;
2572 		cookie->ndesc = ndesc;
2573 		return (&wr->wr);
2574 	}
2575 
2576 	eq->cidx = read_hw_cidx(eq);
2577 	if (eq->pidx == eq->cidx)
2578 		available = eq->sidx - 1;
2579 	else
2580 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2581 	if (available < ndesc)
2582 		goto slowpath;
2583 
2584 	cookie->pidx = eq->pidx;
2585 	cookie->ndesc = ndesc;
2586 	TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link);
2587 
2588 	w = &eq->desc[eq->pidx];
2589 	IDXINCR(eq->pidx, ndesc, eq->sidx);
2590 	if (__predict_false(cookie->pidx + ndesc > eq->sidx)) {
2591 		w = &wrq->ss[0];
2592 		wrq->ss_pidx = cookie->pidx;
2593 		wrq->ss_len = len16 * 16;
2594 	}
2595 
2596 	EQ_UNLOCK(eq);
2597 
2598 	return (w);
2599 }
2600 
2601 void
2602 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie)
2603 {
2604 	struct sge_eq *eq = &wrq->eq;
2605 	struct adapter *sc = wrq->adapter;
2606 	int ndesc, pidx;
2607 	struct wrq_cookie *prev, *next;
2608 
2609 	if (cookie->pidx == -1) {
2610 		struct wrqe *wr = __containerof(w, struct wrqe, wr);
2611 
2612 		t4_wrq_tx(sc, wr);
2613 		return;
2614 	}
2615 
2616 	if (__predict_false(w == &wrq->ss[0])) {
2617 		int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE;
2618 
2619 		MPASS(wrq->ss_len > n);	/* WR had better wrap around. */
2620 		bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n);
2621 		bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n);
2622 		wrq->tx_wrs_ss++;
2623 	} else
2624 		wrq->tx_wrs_direct++;
2625 
2626 	EQ_LOCK(eq);
2627 	ndesc = cookie->ndesc;	/* Can be more than SGE_MAX_WR_NDESC here. */
2628 	pidx = cookie->pidx;
2629 	MPASS(pidx >= 0 && pidx < eq->sidx);
2630 	prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link);
2631 	next = TAILQ_NEXT(cookie, link);
2632 	if (prev == NULL) {
2633 		MPASS(pidx == eq->dbidx);
2634 		if (next == NULL || ndesc >= 16) {
2635 			int available;
2636 			struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
2637 
2638 			/*
2639 			 * Note that the WR via which we'll request tx updates
2640 			 * is at pidx and not eq->pidx, which has moved on
2641 			 * already.
2642 			 */
2643 			dst = (void *)&eq->desc[pidx];
2644 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2645 			if (available < eq->sidx / 4 &&
2646 			    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2647 				dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2648 				    F_FW_WR_EQUEQ);
2649 				eq->equeqidx = pidx;
2650 			} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
2651 				dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
2652 				eq->equeqidx = pidx;
2653 			}
2654 
2655 			ring_eq_db(wrq->adapter, eq, ndesc);
2656 		} else {
2657 			MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc);
2658 			next->pidx = pidx;
2659 			next->ndesc += ndesc;
2660 		}
2661 	} else {
2662 		MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc);
2663 		prev->ndesc += ndesc;
2664 	}
2665 	TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link);
2666 
2667 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2668 		drain_wrq_wr_list(sc, wrq);
2669 
2670 #ifdef INVARIANTS
2671 	if (TAILQ_EMPTY(&wrq->incomplete_wrs)) {
2672 		/* Doorbell must have caught up to the pidx. */
2673 		MPASS(wrq->eq.pidx == wrq->eq.dbidx);
2674 	}
2675 #endif
2676 	EQ_UNLOCK(eq);
2677 }
2678 
2679 static u_int
2680 can_resume_eth_tx(struct mp_ring *r)
2681 {
2682 	struct sge_eq *eq = r->cookie;
2683 
2684 	return (total_available_tx_desc(eq) > eq->sidx / 8);
2685 }
2686 
2687 static inline int
2688 cannot_use_txpkts(struct mbuf *m)
2689 {
2690 	/* maybe put a GL limit too, to avoid silliness? */
2691 
2692 	return (needs_tso(m));
2693 }
2694 
2695 static inline int
2696 discard_tx(struct sge_eq *eq)
2697 {
2698 
2699 	return ((eq->flags & (EQ_ENABLED | EQ_QFLUSH)) != EQ_ENABLED);
2700 }
2701 
2702 /*
2703  * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to
2704  * be consumed.  Return the actual number consumed.  0 indicates a stall.
2705  */
2706 static u_int
2707 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx)
2708 {
2709 	struct sge_txq *txq = r->cookie;
2710 	struct sge_eq *eq = &txq->eq;
2711 	struct ifnet *ifp = txq->ifp;
2712 	struct vi_info *vi = ifp->if_softc;
2713 	struct port_info *pi = vi->pi;
2714 	struct adapter *sc = pi->adapter;
2715 	u_int total, remaining;		/* # of packets */
2716 	u_int available, dbdiff;	/* # of hardware descriptors */
2717 	u_int n, next_cidx;
2718 	struct mbuf *m0, *tail;
2719 	struct txpkts txp;
2720 	struct fw_eth_tx_pkts_wr *wr;	/* any fw WR struct will do */
2721 
2722 	remaining = IDXDIFF(pidx, cidx, r->size);
2723 	MPASS(remaining > 0);	/* Must not be called without work to do. */
2724 	total = 0;
2725 
2726 	TXQ_LOCK(txq);
2727 	if (__predict_false(discard_tx(eq))) {
2728 		while (cidx != pidx) {
2729 			m0 = r->items[cidx];
2730 			m_freem(m0);
2731 			if (++cidx == r->size)
2732 				cidx = 0;
2733 		}
2734 		reclaim_tx_descs(txq, 2048);
2735 		total = remaining;
2736 		goto done;
2737 	}
2738 
2739 	/* How many hardware descriptors do we have readily available. */
2740 	if (eq->pidx == eq->cidx)
2741 		available = eq->sidx - 1;
2742 	else
2743 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2744 	dbdiff = IDXDIFF(eq->pidx, eq->dbidx, eq->sidx);
2745 
2746 	while (remaining > 0) {
2747 
2748 		m0 = r->items[cidx];
2749 		M_ASSERTPKTHDR(m0);
2750 		MPASS(m0->m_nextpkt == NULL);
2751 
2752 		if (available < SGE_MAX_WR_NDESC) {
2753 			available += reclaim_tx_descs(txq, 64);
2754 			if (available < howmany(mbuf_len16(m0), EQ_ESIZE / 16))
2755 				break;	/* out of descriptors */
2756 		}
2757 
2758 		next_cidx = cidx + 1;
2759 		if (__predict_false(next_cidx == r->size))
2760 			next_cidx = 0;
2761 
2762 		wr = (void *)&eq->desc[eq->pidx];
2763 		if (sc->flags & IS_VF) {
2764 			total++;
2765 			remaining--;
2766 			ETHER_BPF_MTAP(ifp, m0);
2767 			n = write_txpkt_vm_wr(sc, txq, (void *)wr, m0,
2768 			    available);
2769 		} else if (remaining > 1 &&
2770 		    try_txpkts(m0, r->items[next_cidx], &txp, available) == 0) {
2771 
2772 			/* pkts at cidx, next_cidx should both be in txp. */
2773 			MPASS(txp.npkt == 2);
2774 			tail = r->items[next_cidx];
2775 			MPASS(tail->m_nextpkt == NULL);
2776 			ETHER_BPF_MTAP(ifp, m0);
2777 			ETHER_BPF_MTAP(ifp, tail);
2778 			m0->m_nextpkt = tail;
2779 
2780 			if (__predict_false(++next_cidx == r->size))
2781 				next_cidx = 0;
2782 
2783 			while (next_cidx != pidx) {
2784 				if (add_to_txpkts(r->items[next_cidx], &txp,
2785 				    available) != 0)
2786 					break;
2787 				tail->m_nextpkt = r->items[next_cidx];
2788 				tail = tail->m_nextpkt;
2789 				ETHER_BPF_MTAP(ifp, tail);
2790 				if (__predict_false(++next_cidx == r->size))
2791 					next_cidx = 0;
2792 			}
2793 
2794 			n = write_txpkts_wr(txq, wr, m0, &txp, available);
2795 			total += txp.npkt;
2796 			remaining -= txp.npkt;
2797 		} else {
2798 			total++;
2799 			remaining--;
2800 			ETHER_BPF_MTAP(ifp, m0);
2801 			n = write_txpkt_wr(txq, (void *)wr, m0, available);
2802 		}
2803 		MPASS(n >= 1 && n <= available && n <= SGE_MAX_WR_NDESC);
2804 
2805 		available -= n;
2806 		dbdiff += n;
2807 		IDXINCR(eq->pidx, n, eq->sidx);
2808 
2809 		if (total_available_tx_desc(eq) < eq->sidx / 4 &&
2810 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2811 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2812 			    F_FW_WR_EQUEQ);
2813 			eq->equeqidx = eq->pidx;
2814 		} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
2815 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
2816 			eq->equeqidx = eq->pidx;
2817 		}
2818 
2819 		if (dbdiff >= 16 && remaining >= 4) {
2820 			ring_eq_db(sc, eq, dbdiff);
2821 			available += reclaim_tx_descs(txq, 4 * dbdiff);
2822 			dbdiff = 0;
2823 		}
2824 
2825 		cidx = next_cidx;
2826 	}
2827 	if (dbdiff != 0) {
2828 		ring_eq_db(sc, eq, dbdiff);
2829 		reclaim_tx_descs(txq, 32);
2830 	}
2831 done:
2832 	TXQ_UNLOCK(txq);
2833 
2834 	return (total);
2835 }
2836 
2837 static inline void
2838 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
2839     int qsize)
2840 {
2841 
2842 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
2843 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
2844 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
2845 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
2846 
2847 	iq->flags = 0;
2848 	iq->adapter = sc;
2849 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
2850 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
2851 	if (pktc_idx >= 0) {
2852 		iq->intr_params |= F_QINTR_CNT_EN;
2853 		iq->intr_pktc_idx = pktc_idx;
2854 	}
2855 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
2856 	iq->sidx = iq->qsize - sc->params.sge.spg_len / IQ_ESIZE;
2857 }
2858 
2859 static inline void
2860 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name)
2861 {
2862 
2863 	fl->qsize = qsize;
2864 	fl->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2865 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
2866 	if (sc->flags & BUF_PACKING_OK &&
2867 	    ((!is_t4(sc) && buffer_packing) ||	/* T5+: enabled unless 0 */
2868 	    (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */
2869 		fl->flags |= FL_BUF_PACKING;
2870 	find_best_refill_source(sc, fl, maxp);
2871 	find_safe_refill_source(sc, fl);
2872 }
2873 
2874 static inline void
2875 init_eq(struct adapter *sc, struct sge_eq *eq, int eqtype, int qsize,
2876     uint8_t tx_chan, uint16_t iqid, char *name)
2877 {
2878 	KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
2879 
2880 	eq->flags = eqtype & EQ_TYPEMASK;
2881 	eq->tx_chan = tx_chan;
2882 	eq->iqid = iqid;
2883 	eq->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2884 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
2885 }
2886 
2887 static int
2888 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
2889     bus_dmamap_t *map, bus_addr_t *pa, void **va)
2890 {
2891 	int rc;
2892 
2893 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
2894 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
2895 	if (rc != 0) {
2896 		device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
2897 		goto done;
2898 	}
2899 
2900 	rc = bus_dmamem_alloc(*tag, va,
2901 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
2902 	if (rc != 0) {
2903 		device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
2904 		goto done;
2905 	}
2906 
2907 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
2908 	if (rc != 0) {
2909 		device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
2910 		goto done;
2911 	}
2912 done:
2913 	if (rc)
2914 		free_ring(sc, *tag, *map, *pa, *va);
2915 
2916 	return (rc);
2917 }
2918 
2919 static int
2920 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
2921     bus_addr_t pa, void *va)
2922 {
2923 	if (pa)
2924 		bus_dmamap_unload(tag, map);
2925 	if (va)
2926 		bus_dmamem_free(tag, va, map);
2927 	if (tag)
2928 		bus_dma_tag_destroy(tag);
2929 
2930 	return (0);
2931 }
2932 
2933 /*
2934  * Allocates the ring for an ingress queue and an optional freelist.  If the
2935  * freelist is specified it will be allocated and then associated with the
2936  * ingress queue.
2937  *
2938  * Returns errno on failure.  Resources allocated up to that point may still be
2939  * allocated.  Caller is responsible for cleanup in case this function fails.
2940  *
2941  * If the ingress queue will take interrupts directly then the intr_idx
2942  * specifies the vector, starting from 0.  -1 means the interrupts for this
2943  * queue should be forwarded to the fwq.
2944  */
2945 static int
2946 alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl,
2947     int intr_idx, int cong)
2948 {
2949 	int rc, i, cntxt_id;
2950 	size_t len;
2951 	struct fw_iq_cmd c;
2952 	struct port_info *pi = vi->pi;
2953 	struct adapter *sc = iq->adapter;
2954 	struct sge_params *sp = &sc->params.sge;
2955 	__be32 v = 0;
2956 
2957 	len = iq->qsize * IQ_ESIZE;
2958 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
2959 	    (void **)&iq->desc);
2960 	if (rc != 0)
2961 		return (rc);
2962 
2963 	bzero(&c, sizeof(c));
2964 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
2965 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
2966 	    V_FW_IQ_CMD_VFN(0));
2967 
2968 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
2969 	    FW_LEN16(c));
2970 
2971 	/* Special handling for firmware event queue */
2972 	if (iq == &sc->sge.fwq)
2973 		v |= F_FW_IQ_CMD_IQASYNCH;
2974 
2975 	if (intr_idx < 0) {
2976 		/* Forwarded interrupts, all headed to fwq */
2977 		v |= F_FW_IQ_CMD_IQANDST;
2978 		v |= V_FW_IQ_CMD_IQANDSTINDEX(sc->sge.fwq.cntxt_id);
2979 	} else {
2980 		KASSERT(intr_idx < sc->intr_count,
2981 		    ("%s: invalid direct intr_idx %d", __func__, intr_idx));
2982 		v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
2983 	}
2984 
2985 	c.type_to_iqandstindex = htobe32(v |
2986 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
2987 	    V_FW_IQ_CMD_VIID(vi->viid) |
2988 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
2989 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
2990 	    F_FW_IQ_CMD_IQGTSMODE |
2991 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
2992 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
2993 	c.iqsize = htobe16(iq->qsize);
2994 	c.iqaddr = htobe64(iq->ba);
2995 	if (cong >= 0)
2996 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
2997 
2998 	if (fl) {
2999 		mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
3000 
3001 		len = fl->qsize * EQ_ESIZE;
3002 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
3003 		    &fl->ba, (void **)&fl->desc);
3004 		if (rc)
3005 			return (rc);
3006 
3007 		/* Allocate space for one software descriptor per buffer. */
3008 		rc = alloc_fl_sdesc(fl);
3009 		if (rc != 0) {
3010 			device_printf(sc->dev,
3011 			    "failed to setup fl software descriptors: %d\n",
3012 			    rc);
3013 			return (rc);
3014 		}
3015 
3016 		if (fl->flags & FL_BUF_PACKING) {
3017 			fl->lowat = roundup2(sp->fl_starve_threshold2, 8);
3018 			fl->buf_boundary = sp->pack_boundary;
3019 		} else {
3020 			fl->lowat = roundup2(sp->fl_starve_threshold, 8);
3021 			fl->buf_boundary = 16;
3022 		}
3023 		if (fl_pad && fl->buf_boundary < sp->pad_boundary)
3024 			fl->buf_boundary = sp->pad_boundary;
3025 
3026 		c.iqns_to_fl0congen |=
3027 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
3028 			F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
3029 			(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
3030 			(fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
3031 			    0));
3032 		if (cong >= 0) {
3033 			c.iqns_to_fl0congen |=
3034 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
3035 				    F_FW_IQ_CMD_FL0CONGCIF |
3036 				    F_FW_IQ_CMD_FL0CONGEN);
3037 		}
3038 		c.fl0dcaen_to_fl0cidxfthresh =
3039 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
3040 			X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B) |
3041 			V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
3042 			X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
3043 		c.fl0size = htobe16(fl->qsize);
3044 		c.fl0addr = htobe64(fl->ba);
3045 	}
3046 
3047 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3048 	if (rc != 0) {
3049 		device_printf(sc->dev,
3050 		    "failed to create ingress queue: %d\n", rc);
3051 		return (rc);
3052 	}
3053 
3054 	iq->cidx = 0;
3055 	iq->gen = F_RSPD_GEN;
3056 	iq->intr_next = iq->intr_params;
3057 	iq->cntxt_id = be16toh(c.iqid);
3058 	iq->abs_id = be16toh(c.physiqid);
3059 	iq->flags |= IQ_ALLOCATED;
3060 
3061 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
3062 	if (cntxt_id >= sc->sge.niq) {
3063 		panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
3064 		    cntxt_id, sc->sge.niq - 1);
3065 	}
3066 	sc->sge.iqmap[cntxt_id] = iq;
3067 
3068 	if (fl) {
3069 		u_int qid;
3070 
3071 		iq->flags |= IQ_HAS_FL;
3072 		fl->cntxt_id = be16toh(c.fl0id);
3073 		fl->pidx = fl->cidx = 0;
3074 
3075 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
3076 		if (cntxt_id >= sc->sge.neq) {
3077 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
3078 			    __func__, cntxt_id, sc->sge.neq - 1);
3079 		}
3080 		sc->sge.eqmap[cntxt_id] = (void *)fl;
3081 
3082 		qid = fl->cntxt_id;
3083 		if (isset(&sc->doorbells, DOORBELL_UDB)) {
3084 			uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3085 			uint32_t mask = (1 << s_qpp) - 1;
3086 			volatile uint8_t *udb;
3087 
3088 			udb = sc->udbs_base + UDBS_DB_OFFSET;
3089 			udb += (qid >> s_qpp) << PAGE_SHIFT;
3090 			qid &= mask;
3091 			if (qid < PAGE_SIZE / UDBS_SEG_SIZE) {
3092 				udb += qid << UDBS_SEG_SHIFT;
3093 				qid = 0;
3094 			}
3095 			fl->udb = (volatile void *)udb;
3096 		}
3097 		fl->dbval = V_QID(qid) | sc->chip_params->sge_fl_db;
3098 
3099 		FL_LOCK(fl);
3100 		/* Enough to make sure the SGE doesn't think it's starved */
3101 		refill_fl(sc, fl, fl->lowat);
3102 		FL_UNLOCK(fl);
3103 	}
3104 
3105 	if (chip_id(sc) >= CHELSIO_T5 && !(sc->flags & IS_VF) && cong >= 0) {
3106 		uint32_t param, val;
3107 
3108 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
3109 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
3110 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
3111 		if (cong == 0)
3112 			val = 1 << 19;
3113 		else {
3114 			val = 2 << 19;
3115 			for (i = 0; i < 4; i++) {
3116 				if (cong & (1 << i))
3117 					val |= 1 << (i << 2);
3118 			}
3119 		}
3120 
3121 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
3122 		if (rc != 0) {
3123 			/* report error but carry on */
3124 			device_printf(sc->dev,
3125 			    "failed to set congestion manager context for "
3126 			    "ingress queue %d: %d\n", iq->cntxt_id, rc);
3127 		}
3128 	}
3129 
3130 	/* Enable IQ interrupts */
3131 	atomic_store_rel_int(&iq->state, IQS_IDLE);
3132 	t4_write_reg(sc, sc->sge_gts_reg, V_SEINTARM(iq->intr_params) |
3133 	    V_INGRESSQID(iq->cntxt_id));
3134 
3135 	return (0);
3136 }
3137 
3138 static int
3139 free_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl)
3140 {
3141 	int rc;
3142 	struct adapter *sc = iq->adapter;
3143 	device_t dev;
3144 
3145 	if (sc == NULL)
3146 		return (0);	/* nothing to do */
3147 
3148 	dev = vi ? vi->dev : sc->dev;
3149 
3150 	if (iq->flags & IQ_ALLOCATED) {
3151 		rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
3152 		    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
3153 		    fl ? fl->cntxt_id : 0xffff, 0xffff);
3154 		if (rc != 0) {
3155 			device_printf(dev,
3156 			    "failed to free queue %p: %d\n", iq, rc);
3157 			return (rc);
3158 		}
3159 		iq->flags &= ~IQ_ALLOCATED;
3160 	}
3161 
3162 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
3163 
3164 	bzero(iq, sizeof(*iq));
3165 
3166 	if (fl) {
3167 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
3168 		    fl->desc);
3169 
3170 		if (fl->sdesc)
3171 			free_fl_sdesc(sc, fl);
3172 
3173 		if (mtx_initialized(&fl->fl_lock))
3174 			mtx_destroy(&fl->fl_lock);
3175 
3176 		bzero(fl, sizeof(*fl));
3177 	}
3178 
3179 	return (0);
3180 }
3181 
3182 static void
3183 add_iq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
3184     struct sge_iq *iq)
3185 {
3186 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3187 
3188 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD, &iq->ba,
3189 	    "bus address of descriptor ring");
3190 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3191 	    iq->qsize * IQ_ESIZE, "descriptor ring size in bytes");
3192 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3193 	    CTLTYPE_INT | CTLFLAG_RD, &iq->abs_id, 0, sysctl_uint16, "I",
3194 	    "absolute id of the queue");
3195 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3196 	    CTLTYPE_INT | CTLFLAG_RD, &iq->cntxt_id, 0, sysctl_uint16, "I",
3197 	    "SGE context id of the queue");
3198 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3199 	    CTLTYPE_INT | CTLFLAG_RD, &iq->cidx, 0, sysctl_uint16, "I",
3200 	    "consumer index");
3201 }
3202 
3203 static void
3204 add_fl_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
3205     struct sysctl_oid *oid, struct sge_fl *fl)
3206 {
3207 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3208 
3209 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3210 	    "freelist");
3211 	children = SYSCTL_CHILDREN(oid);
3212 
3213 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3214 	    &fl->ba, "bus address of descriptor ring");
3215 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3216 	    fl->sidx * EQ_ESIZE + sc->params.sge.spg_len,
3217 	    "desc ring size in bytes");
3218 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3219 	    CTLTYPE_INT | CTLFLAG_RD, &fl->cntxt_id, 0, sysctl_uint16, "I",
3220 	    "SGE context id of the freelist");
3221 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL,
3222 	    fl_pad ? 1 : 0, "padding enabled");
3223 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL,
3224 	    fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled");
3225 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
3226 	    0, "consumer index");
3227 	if (fl->flags & FL_BUF_PACKING) {
3228 		SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
3229 		    CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
3230 	}
3231 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
3232 	    0, "producer index");
3233 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_allocated",
3234 	    CTLFLAG_RD, &fl->mbuf_allocated, "# of mbuf allocated");
3235 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_inlined",
3236 	    CTLFLAG_RD, &fl->mbuf_inlined, "# of mbuf inlined in clusters");
3237 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
3238 	    CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
3239 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
3240 	    CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
3241 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
3242 	    CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
3243 }
3244 
3245 static int
3246 alloc_fwq(struct adapter *sc)
3247 {
3248 	int rc, intr_idx;
3249 	struct sge_iq *fwq = &sc->sge.fwq;
3250 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
3251 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3252 
3253 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE);
3254 	if (sc->flags & IS_VF)
3255 		intr_idx = 0;
3256 	else
3257 		intr_idx = sc->intr_count > 1 ? 1 : 0;
3258 	rc = alloc_iq_fl(&sc->port[0]->vi[0], fwq, NULL, intr_idx, -1);
3259 	if (rc != 0) {
3260 		device_printf(sc->dev,
3261 		    "failed to create firmware event queue: %d\n", rc);
3262 		return (rc);
3263 	}
3264 
3265 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD,
3266 	    NULL, "firmware event queue");
3267 	add_iq_sysctls(&sc->ctx, oid, fwq);
3268 
3269 	return (0);
3270 }
3271 
3272 static int
3273 free_fwq(struct adapter *sc)
3274 {
3275 	return free_iq_fl(NULL, &sc->sge.fwq, NULL);
3276 }
3277 
3278 static int
3279 alloc_ctrlq(struct adapter *sc, struct sge_wrq *ctrlq, int idx,
3280     struct sysctl_oid *oid)
3281 {
3282 	int rc;
3283 	char name[16];
3284 	struct sysctl_oid_list *children;
3285 
3286 	snprintf(name, sizeof(name), "%s ctrlq%d", device_get_nameunit(sc->dev),
3287 	    idx);
3288 	init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[idx]->tx_chan,
3289 	    sc->sge.fwq.cntxt_id, name);
3290 
3291 	children = SYSCTL_CHILDREN(oid);
3292 	snprintf(name, sizeof(name), "%d", idx);
3293 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3294 	    NULL, "ctrl queue");
3295 	rc = alloc_wrq(sc, NULL, ctrlq, oid);
3296 
3297 	return (rc);
3298 }
3299 
3300 int
3301 tnl_cong(struct port_info *pi, int drop)
3302 {
3303 
3304 	if (drop == -1)
3305 		return (-1);
3306 	else if (drop == 1)
3307 		return (0);
3308 	else
3309 		return (pi->rx_e_chan_map);
3310 }
3311 
3312 static int
3313 alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int intr_idx, int idx,
3314     struct sysctl_oid *oid)
3315 {
3316 	int rc;
3317 	struct adapter *sc = vi->pi->adapter;
3318 	struct sysctl_oid_list *children;
3319 	char name[16];
3320 
3321 	rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, intr_idx,
3322 	    tnl_cong(vi->pi, cong_drop));
3323 	if (rc != 0)
3324 		return (rc);
3325 
3326 	if (idx == 0)
3327 		sc->sge.iq_base = rxq->iq.abs_id - rxq->iq.cntxt_id;
3328 	else
3329 		KASSERT(rxq->iq.cntxt_id + sc->sge.iq_base == rxq->iq.abs_id,
3330 		    ("iq_base mismatch"));
3331 	KASSERT(sc->sge.iq_base == 0 || sc->flags & IS_VF,
3332 	    ("PF with non-zero iq_base"));
3333 
3334 	/*
3335 	 * The freelist is just barely above the starvation threshold right now,
3336 	 * fill it up a bit more.
3337 	 */
3338 	FL_LOCK(&rxq->fl);
3339 	refill_fl(sc, &rxq->fl, 128);
3340 	FL_UNLOCK(&rxq->fl);
3341 
3342 #if defined(INET) || defined(INET6)
3343 	rc = tcp_lro_init_args(&rxq->lro, vi->ifp, lro_entries, lro_mbufs);
3344 	if (rc != 0)
3345 		return (rc);
3346 	MPASS(rxq->lro.ifp == vi->ifp);	/* also indicates LRO init'ed */
3347 
3348 	if (vi->ifp->if_capenable & IFCAP_LRO)
3349 		rxq->iq.flags |= IQ_LRO_ENABLED;
3350 #endif
3351 	rxq->ifp = vi->ifp;
3352 
3353 	children = SYSCTL_CHILDREN(oid);
3354 
3355 	snprintf(name, sizeof(name), "%d", idx);
3356 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3357 	    NULL, "rx queue");
3358 	children = SYSCTL_CHILDREN(oid);
3359 
3360 	add_iq_sysctls(&vi->ctx, oid, &rxq->iq);
3361 #if defined(INET) || defined(INET6)
3362 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
3363 	    &rxq->lro.lro_queued, 0, NULL);
3364 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
3365 	    &rxq->lro.lro_flushed, 0, NULL);
3366 #endif
3367 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
3368 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
3369 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_extraction",
3370 	    CTLFLAG_RD, &rxq->vlan_extraction,
3371 	    "# of times hardware extracted 802.1Q tag");
3372 
3373 	add_fl_sysctls(sc, &vi->ctx, oid, &rxq->fl);
3374 
3375 	return (rc);
3376 }
3377 
3378 static int
3379 free_rxq(struct vi_info *vi, struct sge_rxq *rxq)
3380 {
3381 	int rc;
3382 
3383 #if defined(INET) || defined(INET6)
3384 	if (rxq->lro.ifp) {
3385 		tcp_lro_free(&rxq->lro);
3386 		rxq->lro.ifp = NULL;
3387 	}
3388 #endif
3389 
3390 	rc = free_iq_fl(vi, &rxq->iq, &rxq->fl);
3391 	if (rc == 0)
3392 		bzero(rxq, sizeof(*rxq));
3393 
3394 	return (rc);
3395 }
3396 
3397 #ifdef TCP_OFFLOAD
3398 static int
3399 alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq,
3400     int intr_idx, int idx, struct sysctl_oid *oid)
3401 {
3402 	struct port_info *pi = vi->pi;
3403 	int rc;
3404 	struct sysctl_oid_list *children;
3405 	char name[16];
3406 
3407 	rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx, 0);
3408 	if (rc != 0)
3409 		return (rc);
3410 
3411 	children = SYSCTL_CHILDREN(oid);
3412 
3413 	snprintf(name, sizeof(name), "%d", idx);
3414 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3415 	    NULL, "rx queue");
3416 	add_iq_sysctls(&vi->ctx, oid, &ofld_rxq->iq);
3417 	add_fl_sysctls(pi->adapter, &vi->ctx, oid, &ofld_rxq->fl);
3418 
3419 	return (rc);
3420 }
3421 
3422 static int
3423 free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq)
3424 {
3425 	int rc;
3426 
3427 	rc = free_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl);
3428 	if (rc == 0)
3429 		bzero(ofld_rxq, sizeof(*ofld_rxq));
3430 
3431 	return (rc);
3432 }
3433 #endif
3434 
3435 #ifdef DEV_NETMAP
3436 static int
3437 alloc_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int intr_idx,
3438     int idx, struct sysctl_oid *oid)
3439 {
3440 	int rc;
3441 	struct sysctl_oid_list *children;
3442 	struct sysctl_ctx_list *ctx;
3443 	char name[16];
3444 	size_t len;
3445 	struct adapter *sc = vi->pi->adapter;
3446 	struct netmap_adapter *na = NA(vi->ifp);
3447 
3448 	MPASS(na != NULL);
3449 
3450 	len = vi->qsize_rxq * IQ_ESIZE;
3451 	rc = alloc_ring(sc, len, &nm_rxq->iq_desc_tag, &nm_rxq->iq_desc_map,
3452 	    &nm_rxq->iq_ba, (void **)&nm_rxq->iq_desc);
3453 	if (rc != 0)
3454 		return (rc);
3455 
3456 	len = na->num_rx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3457 	rc = alloc_ring(sc, len, &nm_rxq->fl_desc_tag, &nm_rxq->fl_desc_map,
3458 	    &nm_rxq->fl_ba, (void **)&nm_rxq->fl_desc);
3459 	if (rc != 0)
3460 		return (rc);
3461 
3462 	nm_rxq->vi = vi;
3463 	nm_rxq->nid = idx;
3464 	nm_rxq->iq_cidx = 0;
3465 	nm_rxq->iq_sidx = vi->qsize_rxq - sc->params.sge.spg_len / IQ_ESIZE;
3466 	nm_rxq->iq_gen = F_RSPD_GEN;
3467 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
3468 	nm_rxq->fl_sidx = na->num_rx_desc;
3469 	nm_rxq->intr_idx = intr_idx;
3470 	nm_rxq->iq_cntxt_id = INVALID_NM_RXQ_CNTXT_ID;
3471 
3472 	ctx = &vi->ctx;
3473 	children = SYSCTL_CHILDREN(oid);
3474 
3475 	snprintf(name, sizeof(name), "%d", idx);
3476 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL,
3477 	    "rx queue");
3478 	children = SYSCTL_CHILDREN(oid);
3479 
3480 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3481 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_abs_id, 0, sysctl_uint16,
3482 	    "I", "absolute id of the queue");
3483 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3484 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cntxt_id, 0, sysctl_uint16,
3485 	    "I", "SGE context id of the queue");
3486 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3487 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cidx, 0, sysctl_uint16, "I",
3488 	    "consumer index");
3489 
3490 	children = SYSCTL_CHILDREN(oid);
3491 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3492 	    "freelist");
3493 	children = SYSCTL_CHILDREN(oid);
3494 
3495 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3496 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->fl_cntxt_id, 0, sysctl_uint16,
3497 	    "I", "SGE context id of the freelist");
3498 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
3499 	    &nm_rxq->fl_cidx, 0, "consumer index");
3500 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
3501 	    &nm_rxq->fl_pidx, 0, "producer index");
3502 
3503 	return (rc);
3504 }
3505 
3506 
3507 static int
3508 free_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
3509 {
3510 	struct adapter *sc = vi->pi->adapter;
3511 
3512 	if (vi->flags & VI_INIT_DONE)
3513 		MPASS(nm_rxq->iq_cntxt_id == INVALID_NM_RXQ_CNTXT_ID);
3514 	else
3515 		MPASS(nm_rxq->iq_cntxt_id == 0);
3516 
3517 	free_ring(sc, nm_rxq->iq_desc_tag, nm_rxq->iq_desc_map, nm_rxq->iq_ba,
3518 	    nm_rxq->iq_desc);
3519 	free_ring(sc, nm_rxq->fl_desc_tag, nm_rxq->fl_desc_map, nm_rxq->fl_ba,
3520 	    nm_rxq->fl_desc);
3521 
3522 	return (0);
3523 }
3524 
3525 static int
3526 alloc_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq, int iqidx, int idx,
3527     struct sysctl_oid *oid)
3528 {
3529 	int rc;
3530 	size_t len;
3531 	struct port_info *pi = vi->pi;
3532 	struct adapter *sc = pi->adapter;
3533 	struct netmap_adapter *na = NA(vi->ifp);
3534 	char name[16];
3535 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3536 
3537 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3538 	rc = alloc_ring(sc, len, &nm_txq->desc_tag, &nm_txq->desc_map,
3539 	    &nm_txq->ba, (void **)&nm_txq->desc);
3540 	if (rc)
3541 		return (rc);
3542 
3543 	nm_txq->pidx = nm_txq->cidx = 0;
3544 	nm_txq->sidx = na->num_tx_desc;
3545 	nm_txq->nid = idx;
3546 	nm_txq->iqidx = iqidx;
3547 	nm_txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3548 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(G_FW_VIID_PFN(vi->viid)) |
3549 	    V_TXPKT_VF(G_FW_VIID_VIN(vi->viid)) |
3550 	    V_TXPKT_VF_VLD(G_FW_VIID_VIVLD(vi->viid)));
3551 	nm_txq->cntxt_id = INVALID_NM_TXQ_CNTXT_ID;
3552 
3553 	snprintf(name, sizeof(name), "%d", idx);
3554 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3555 	    NULL, "netmap tx queue");
3556 	children = SYSCTL_CHILDREN(oid);
3557 
3558 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3559 	    &nm_txq->cntxt_id, 0, "SGE context id of the queue");
3560 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3561 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->cidx, 0, sysctl_uint16, "I",
3562 	    "consumer index");
3563 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3564 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->pidx, 0, sysctl_uint16, "I",
3565 	    "producer index");
3566 
3567 	return (rc);
3568 }
3569 
3570 static int
3571 free_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
3572 {
3573 	struct adapter *sc = vi->pi->adapter;
3574 
3575 	if (vi->flags & VI_INIT_DONE)
3576 		MPASS(nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID);
3577 	else
3578 		MPASS(nm_txq->cntxt_id == 0);
3579 
3580 	free_ring(sc, nm_txq->desc_tag, nm_txq->desc_map, nm_txq->ba,
3581 	    nm_txq->desc);
3582 
3583 	return (0);
3584 }
3585 #endif
3586 
3587 static int
3588 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
3589 {
3590 	int rc, cntxt_id;
3591 	struct fw_eq_ctrl_cmd c;
3592 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3593 
3594 	bzero(&c, sizeof(c));
3595 
3596 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
3597 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
3598 	    V_FW_EQ_CTRL_CMD_VFN(0));
3599 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
3600 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
3601 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid));
3602 	c.physeqid_pkd = htobe32(0);
3603 	c.fetchszm_to_iqid =
3604 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
3605 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
3606 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
3607 	c.dcaen_to_eqsize =
3608 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3609 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3610 		V_FW_EQ_CTRL_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
3611 		V_FW_EQ_CTRL_CMD_EQSIZE(qsize));
3612 	c.eqaddr = htobe64(eq->ba);
3613 
3614 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3615 	if (rc != 0) {
3616 		device_printf(sc->dev,
3617 		    "failed to create control queue %d: %d\n", eq->tx_chan, rc);
3618 		return (rc);
3619 	}
3620 	eq->flags |= EQ_ALLOCATED;
3621 
3622 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
3623 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3624 	if (cntxt_id >= sc->sge.neq)
3625 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3626 		cntxt_id, sc->sge.neq - 1);
3627 	sc->sge.eqmap[cntxt_id] = eq;
3628 
3629 	return (rc);
3630 }
3631 
3632 static int
3633 eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3634 {
3635 	int rc, cntxt_id;
3636 	struct fw_eq_eth_cmd c;
3637 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3638 
3639 	bzero(&c, sizeof(c));
3640 
3641 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
3642 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
3643 	    V_FW_EQ_ETH_CMD_VFN(0));
3644 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
3645 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
3646 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
3647 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
3648 	c.fetchszm_to_iqid =
3649 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3650 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
3651 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
3652 	c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3653 	    V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3654 	    V_FW_EQ_ETH_CMD_EQSIZE(qsize));
3655 	c.eqaddr = htobe64(eq->ba);
3656 
3657 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3658 	if (rc != 0) {
3659 		device_printf(vi->dev,
3660 		    "failed to create Ethernet egress queue: %d\n", rc);
3661 		return (rc);
3662 	}
3663 	eq->flags |= EQ_ALLOCATED;
3664 
3665 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
3666 	eq->abs_id = G_FW_EQ_ETH_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
3667 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3668 	if (cntxt_id >= sc->sge.neq)
3669 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3670 		cntxt_id, sc->sge.neq - 1);
3671 	sc->sge.eqmap[cntxt_id] = eq;
3672 
3673 	return (rc);
3674 }
3675 
3676 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3677 static int
3678 ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3679 {
3680 	int rc, cntxt_id;
3681 	struct fw_eq_ofld_cmd c;
3682 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3683 
3684 	bzero(&c, sizeof(c));
3685 
3686 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
3687 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
3688 	    V_FW_EQ_OFLD_CMD_VFN(0));
3689 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
3690 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
3691 	c.fetchszm_to_iqid =
3692 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3693 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
3694 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
3695 	c.dcaen_to_eqsize =
3696 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3697 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3698 		V_FW_EQ_OFLD_CMD_EQSIZE(qsize));
3699 	c.eqaddr = htobe64(eq->ba);
3700 
3701 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3702 	if (rc != 0) {
3703 		device_printf(vi->dev,
3704 		    "failed to create egress queue for TCP offload: %d\n", rc);
3705 		return (rc);
3706 	}
3707 	eq->flags |= EQ_ALLOCATED;
3708 
3709 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
3710 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3711 	if (cntxt_id >= sc->sge.neq)
3712 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3713 		cntxt_id, sc->sge.neq - 1);
3714 	sc->sge.eqmap[cntxt_id] = eq;
3715 
3716 	return (rc);
3717 }
3718 #endif
3719 
3720 static int
3721 alloc_eq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3722 {
3723 	int rc, qsize;
3724 	size_t len;
3725 
3726 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
3727 
3728 	qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3729 	len = qsize * EQ_ESIZE;
3730 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
3731 	    &eq->ba, (void **)&eq->desc);
3732 	if (rc)
3733 		return (rc);
3734 
3735 	eq->pidx = eq->cidx = 0;
3736 	eq->equeqidx = eq->dbidx = 0;
3737 	eq->doorbells = sc->doorbells;
3738 
3739 	switch (eq->flags & EQ_TYPEMASK) {
3740 	case EQ_CTRL:
3741 		rc = ctrl_eq_alloc(sc, eq);
3742 		break;
3743 
3744 	case EQ_ETH:
3745 		rc = eth_eq_alloc(sc, vi, eq);
3746 		break;
3747 
3748 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3749 	case EQ_OFLD:
3750 		rc = ofld_eq_alloc(sc, vi, eq);
3751 		break;
3752 #endif
3753 
3754 	default:
3755 		panic("%s: invalid eq type %d.", __func__,
3756 		    eq->flags & EQ_TYPEMASK);
3757 	}
3758 	if (rc != 0) {
3759 		device_printf(sc->dev,
3760 		    "failed to allocate egress queue(%d): %d\n",
3761 		    eq->flags & EQ_TYPEMASK, rc);
3762 	}
3763 
3764 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
3765 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
3766 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
3767 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3768 		uint32_t mask = (1 << s_qpp) - 1;
3769 		volatile uint8_t *udb;
3770 
3771 		udb = sc->udbs_base + UDBS_DB_OFFSET;
3772 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
3773 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
3774 		if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
3775 	    		clrbit(&eq->doorbells, DOORBELL_WCWR);
3776 		else {
3777 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
3778 			eq->udb_qid = 0;
3779 		}
3780 		eq->udb = (volatile void *)udb;
3781 	}
3782 
3783 	return (rc);
3784 }
3785 
3786 static int
3787 free_eq(struct adapter *sc, struct sge_eq *eq)
3788 {
3789 	int rc;
3790 
3791 	if (eq->flags & EQ_ALLOCATED) {
3792 		switch (eq->flags & EQ_TYPEMASK) {
3793 		case EQ_CTRL:
3794 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
3795 			    eq->cntxt_id);
3796 			break;
3797 
3798 		case EQ_ETH:
3799 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
3800 			    eq->cntxt_id);
3801 			break;
3802 
3803 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3804 		case EQ_OFLD:
3805 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
3806 			    eq->cntxt_id);
3807 			break;
3808 #endif
3809 
3810 		default:
3811 			panic("%s: invalid eq type %d.", __func__,
3812 			    eq->flags & EQ_TYPEMASK);
3813 		}
3814 		if (rc != 0) {
3815 			device_printf(sc->dev,
3816 			    "failed to free egress queue (%d): %d\n",
3817 			    eq->flags & EQ_TYPEMASK, rc);
3818 			return (rc);
3819 		}
3820 		eq->flags &= ~EQ_ALLOCATED;
3821 	}
3822 
3823 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
3824 
3825 	if (mtx_initialized(&eq->eq_lock))
3826 		mtx_destroy(&eq->eq_lock);
3827 
3828 	bzero(eq, sizeof(*eq));
3829 	return (0);
3830 }
3831 
3832 static int
3833 alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq,
3834     struct sysctl_oid *oid)
3835 {
3836 	int rc;
3837 	struct sysctl_ctx_list *ctx = vi ? &vi->ctx : &sc->ctx;
3838 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3839 
3840 	rc = alloc_eq(sc, vi, &wrq->eq);
3841 	if (rc)
3842 		return (rc);
3843 
3844 	wrq->adapter = sc;
3845 	TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq);
3846 	TAILQ_INIT(&wrq->incomplete_wrs);
3847 	STAILQ_INIT(&wrq->wr_list);
3848 	wrq->nwr_pending = 0;
3849 	wrq->ndesc_needed = 0;
3850 
3851 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3852 	    &wrq->eq.ba, "bus address of descriptor ring");
3853 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3854 	    wrq->eq.sidx * EQ_ESIZE + sc->params.sge.spg_len,
3855 	    "desc ring size in bytes");
3856 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3857 	    &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
3858 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3859 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I",
3860 	    "consumer index");
3861 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx",
3862 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I",
3863 	    "producer index");
3864 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
3865 	    wrq->eq.sidx, "status page index");
3866 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD,
3867 	    &wrq->tx_wrs_direct, "# of work requests (direct)");
3868 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD,
3869 	    &wrq->tx_wrs_copied, "# of work requests (copied)");
3870 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_sspace", CTLFLAG_RD,
3871 	    &wrq->tx_wrs_ss, "# of work requests (copied from scratch space)");
3872 
3873 	return (rc);
3874 }
3875 
3876 static int
3877 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
3878 {
3879 	int rc;
3880 
3881 	rc = free_eq(sc, &wrq->eq);
3882 	if (rc)
3883 		return (rc);
3884 
3885 	bzero(wrq, sizeof(*wrq));
3886 	return (0);
3887 }
3888 
3889 static int
3890 alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx,
3891     struct sysctl_oid *oid)
3892 {
3893 	int rc;
3894 	struct port_info *pi = vi->pi;
3895 	struct adapter *sc = pi->adapter;
3896 	struct sge_eq *eq = &txq->eq;
3897 	char name[16];
3898 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3899 
3900 	rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx, can_resume_eth_tx,
3901 	    M_CXGBE, M_WAITOK);
3902 	if (rc != 0) {
3903 		device_printf(sc->dev, "failed to allocate mp_ring: %d\n", rc);
3904 		return (rc);
3905 	}
3906 
3907 	rc = alloc_eq(sc, vi, eq);
3908 	if (rc != 0) {
3909 		mp_ring_free(txq->r);
3910 		txq->r = NULL;
3911 		return (rc);
3912 	}
3913 
3914 	/* Can't fail after this point. */
3915 
3916 	if (idx == 0)
3917 		sc->sge.eq_base = eq->abs_id - eq->cntxt_id;
3918 	else
3919 		KASSERT(eq->cntxt_id + sc->sge.eq_base == eq->abs_id,
3920 		    ("eq_base mismatch"));
3921 	KASSERT(sc->sge.eq_base == 0 || sc->flags & IS_VF,
3922 	    ("PF with non-zero eq_base"));
3923 
3924 	TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq);
3925 	txq->ifp = vi->ifp;
3926 	txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK);
3927 	if (sc->flags & IS_VF)
3928 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
3929 		    V_TXPKT_INTF(pi->tx_chan));
3930 	else
3931 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3932 		    V_TXPKT_INTF(pi->tx_chan) |
3933 		    V_TXPKT_PF(G_FW_VIID_PFN(vi->viid)) |
3934 		    V_TXPKT_VF(G_FW_VIID_VIN(vi->viid)) |
3935 		    V_TXPKT_VF_VLD(G_FW_VIID_VIVLD(vi->viid)));
3936 	txq->tc_idx = -1;
3937 	txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE,
3938 	    M_ZERO | M_WAITOK);
3939 
3940 	snprintf(name, sizeof(name), "%d", idx);
3941 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3942 	    NULL, "tx queue");
3943 	children = SYSCTL_CHILDREN(oid);
3944 
3945 	SYSCTL_ADD_UAUTO(&vi->ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3946 	    &eq->ba, "bus address of descriptor ring");
3947 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3948 	    eq->sidx * EQ_ESIZE + sc->params.sge.spg_len,
3949 	    "desc ring size in bytes");
3950 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
3951 	    &eq->abs_id, 0, "absolute id of the queue");
3952 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3953 	    &eq->cntxt_id, 0, "SGE context id of the queue");
3954 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3955 	    CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I",
3956 	    "consumer index");
3957 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3958 	    CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I",
3959 	    "producer index");
3960 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
3961 	    eq->sidx, "status page index");
3962 
3963 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "tc",
3964 	    CTLTYPE_INT | CTLFLAG_RW, vi, idx, sysctl_tc, "I",
3965 	    "traffic class (-1 means none)");
3966 
3967 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
3968 	    &txq->txcsum, "# of times hardware assisted with checksum");
3969 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_insertion",
3970 	    CTLFLAG_RD, &txq->vlan_insertion,
3971 	    "# of times hardware inserted 802.1Q tag");
3972 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
3973 	    &txq->tso_wrs, "# of TSO work requests");
3974 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
3975 	    &txq->imm_wrs, "# of work requests with immediate data");
3976 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
3977 	    &txq->sgl_wrs, "# of work requests with direct SGL");
3978 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
3979 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
3980 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_wrs",
3981 	    CTLFLAG_RD, &txq->txpkts0_wrs,
3982 	    "# of txpkts (type 0) work requests");
3983 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_wrs",
3984 	    CTLFLAG_RD, &txq->txpkts1_wrs,
3985 	    "# of txpkts (type 1) work requests");
3986 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_pkts",
3987 	    CTLFLAG_RD, &txq->txpkts0_pkts,
3988 	    "# of frames tx'd using type0 txpkts work requests");
3989 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_pkts",
3990 	    CTLFLAG_RD, &txq->txpkts1_pkts,
3991 	    "# of frames tx'd using type1 txpkts work requests");
3992 
3993 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_enqueues",
3994 	    CTLFLAG_RD, &txq->r->enqueues,
3995 	    "# of enqueues to the mp_ring for this queue");
3996 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_drops",
3997 	    CTLFLAG_RD, &txq->r->drops,
3998 	    "# of drops in the mp_ring for this queue");
3999 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_starts",
4000 	    CTLFLAG_RD, &txq->r->starts,
4001 	    "# of normal consumer starts in the mp_ring for this queue");
4002 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_stalls",
4003 	    CTLFLAG_RD, &txq->r->stalls,
4004 	    "# of consumer stalls in the mp_ring for this queue");
4005 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_restarts",
4006 	    CTLFLAG_RD, &txq->r->restarts,
4007 	    "# of consumer restarts in the mp_ring for this queue");
4008 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_abdications",
4009 	    CTLFLAG_RD, &txq->r->abdications,
4010 	    "# of consumer abdications in the mp_ring for this queue");
4011 
4012 	return (0);
4013 }
4014 
4015 static int
4016 free_txq(struct vi_info *vi, struct sge_txq *txq)
4017 {
4018 	int rc;
4019 	struct adapter *sc = vi->pi->adapter;
4020 	struct sge_eq *eq = &txq->eq;
4021 
4022 	rc = free_eq(sc, eq);
4023 	if (rc)
4024 		return (rc);
4025 
4026 	sglist_free(txq->gl);
4027 	free(txq->sdesc, M_CXGBE);
4028 	mp_ring_free(txq->r);
4029 
4030 	bzero(txq, sizeof(*txq));
4031 	return (0);
4032 }
4033 
4034 static void
4035 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
4036 {
4037 	bus_addr_t *ba = arg;
4038 
4039 	KASSERT(nseg == 1,
4040 	    ("%s meant for single segment mappings only.", __func__));
4041 
4042 	*ba = error ? 0 : segs->ds_addr;
4043 }
4044 
4045 static inline void
4046 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
4047 {
4048 	uint32_t n, v;
4049 
4050 	n = IDXDIFF(fl->pidx / 8, fl->dbidx, fl->sidx);
4051 	MPASS(n > 0);
4052 
4053 	wmb();
4054 	v = fl->dbval | V_PIDX(n);
4055 	if (fl->udb)
4056 		*fl->udb = htole32(v);
4057 	else
4058 		t4_write_reg(sc, sc->sge_kdoorbell_reg, v);
4059 	IDXINCR(fl->dbidx, n, fl->sidx);
4060 }
4061 
4062 /*
4063  * Fills up the freelist by allocating up to 'n' buffers.  Buffers that are
4064  * recycled do not count towards this allocation budget.
4065  *
4066  * Returns non-zero to indicate that this freelist should be added to the list
4067  * of starving freelists.
4068  */
4069 static int
4070 refill_fl(struct adapter *sc, struct sge_fl *fl, int n)
4071 {
4072 	__be64 *d;
4073 	struct fl_sdesc *sd;
4074 	uintptr_t pa;
4075 	caddr_t cl;
4076 	struct cluster_layout *cll;
4077 	struct sw_zone_info *swz;
4078 	struct cluster_metadata *clm;
4079 	uint16_t max_pidx;
4080 	uint16_t hw_cidx = fl->hw_cidx;		/* stable snapshot */
4081 
4082 	FL_LOCK_ASSERT_OWNED(fl);
4083 
4084 	/*
4085 	 * We always stop at the beginning of the hardware descriptor that's just
4086 	 * before the one with the hw cidx.  This is to avoid hw pidx = hw cidx,
4087 	 * which would mean an empty freelist to the chip.
4088 	 */
4089 	max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1;
4090 	if (fl->pidx == max_pidx * 8)
4091 		return (0);
4092 
4093 	d = &fl->desc[fl->pidx];
4094 	sd = &fl->sdesc[fl->pidx];
4095 	cll = &fl->cll_def;	/* default layout */
4096 	swz = &sc->sge.sw_zone_info[cll->zidx];
4097 
4098 	while (n > 0) {
4099 
4100 		if (sd->cl != NULL) {
4101 
4102 			if (sd->nmbuf == 0) {
4103 				/*
4104 				 * Fast recycle without involving any atomics on
4105 				 * the cluster's metadata (if the cluster has
4106 				 * metadata).  This happens when all frames
4107 				 * received in the cluster were small enough to
4108 				 * fit within a single mbuf each.
4109 				 */
4110 				fl->cl_fast_recycled++;
4111 #ifdef INVARIANTS
4112 				clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
4113 				if (clm != NULL)
4114 					MPASS(clm->refcount == 1);
4115 #endif
4116 				goto recycled_fast;
4117 			}
4118 
4119 			/*
4120 			 * Cluster is guaranteed to have metadata.  Clusters
4121 			 * without metadata always take the fast recycle path
4122 			 * when they're recycled.
4123 			 */
4124 			clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
4125 			MPASS(clm != NULL);
4126 
4127 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4128 				fl->cl_recycled++;
4129 				counter_u64_add(extfree_rels, 1);
4130 				goto recycled;
4131 			}
4132 			sd->cl = NULL;	/* gave up my reference */
4133 		}
4134 		MPASS(sd->cl == NULL);
4135 alloc:
4136 		cl = uma_zalloc(swz->zone, M_NOWAIT);
4137 		if (__predict_false(cl == NULL)) {
4138 			if (cll == &fl->cll_alt || fl->cll_alt.zidx == -1 ||
4139 			    fl->cll_def.zidx == fl->cll_alt.zidx)
4140 				break;
4141 
4142 			/* fall back to the safe zone */
4143 			cll = &fl->cll_alt;
4144 			swz = &sc->sge.sw_zone_info[cll->zidx];
4145 			goto alloc;
4146 		}
4147 		fl->cl_allocated++;
4148 		n--;
4149 
4150 		pa = pmap_kextract((vm_offset_t)cl);
4151 		pa += cll->region1;
4152 		sd->cl = cl;
4153 		sd->cll = *cll;
4154 		*d = htobe64(pa | cll->hwidx);
4155 		clm = cl_metadata(sc, fl, cll, cl);
4156 		if (clm != NULL) {
4157 recycled:
4158 #ifdef INVARIANTS
4159 			clm->sd = sd;
4160 #endif
4161 			clm->refcount = 1;
4162 		}
4163 		sd->nmbuf = 0;
4164 recycled_fast:
4165 		d++;
4166 		sd++;
4167 		if (__predict_false(++fl->pidx % 8 == 0)) {
4168 			uint16_t pidx = fl->pidx / 8;
4169 
4170 			if (__predict_false(pidx == fl->sidx)) {
4171 				fl->pidx = 0;
4172 				pidx = 0;
4173 				sd = fl->sdesc;
4174 				d = fl->desc;
4175 			}
4176 			if (pidx == max_pidx)
4177 				break;
4178 
4179 			if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4)
4180 				ring_fl_db(sc, fl);
4181 		}
4182 	}
4183 
4184 	if (fl->pidx / 8 != fl->dbidx)
4185 		ring_fl_db(sc, fl);
4186 
4187 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
4188 }
4189 
4190 /*
4191  * Attempt to refill all starving freelists.
4192  */
4193 static void
4194 refill_sfl(void *arg)
4195 {
4196 	struct adapter *sc = arg;
4197 	struct sge_fl *fl, *fl_temp;
4198 
4199 	mtx_assert(&sc->sfl_lock, MA_OWNED);
4200 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
4201 		FL_LOCK(fl);
4202 		refill_fl(sc, fl, 64);
4203 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
4204 			TAILQ_REMOVE(&sc->sfl, fl, link);
4205 			fl->flags &= ~FL_STARVING;
4206 		}
4207 		FL_UNLOCK(fl);
4208 	}
4209 
4210 	if (!TAILQ_EMPTY(&sc->sfl))
4211 		callout_schedule(&sc->sfl_callout, hz / 5);
4212 }
4213 
4214 static int
4215 alloc_fl_sdesc(struct sge_fl *fl)
4216 {
4217 
4218 	fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc), M_CXGBE,
4219 	    M_ZERO | M_WAITOK);
4220 
4221 	return (0);
4222 }
4223 
4224 static void
4225 free_fl_sdesc(struct adapter *sc, struct sge_fl *fl)
4226 {
4227 	struct fl_sdesc *sd;
4228 	struct cluster_metadata *clm;
4229 	struct cluster_layout *cll;
4230 	int i;
4231 
4232 	sd = fl->sdesc;
4233 	for (i = 0; i < fl->sidx * 8; i++, sd++) {
4234 		if (sd->cl == NULL)
4235 			continue;
4236 
4237 		cll = &sd->cll;
4238 		clm = cl_metadata(sc, fl, cll, sd->cl);
4239 		if (sd->nmbuf == 0)
4240 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4241 		else if (clm && atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4242 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4243 			counter_u64_add(extfree_rels, 1);
4244 		}
4245 		sd->cl = NULL;
4246 	}
4247 
4248 	free(fl->sdesc, M_CXGBE);
4249 	fl->sdesc = NULL;
4250 }
4251 
4252 static inline void
4253 get_pkt_gl(struct mbuf *m, struct sglist *gl)
4254 {
4255 	int rc;
4256 
4257 	M_ASSERTPKTHDR(m);
4258 
4259 	sglist_reset(gl);
4260 	rc = sglist_append_mbuf(gl, m);
4261 	if (__predict_false(rc != 0)) {
4262 		panic("%s: mbuf %p (%d segs) was vetted earlier but now fails "
4263 		    "with %d.", __func__, m, mbuf_nsegs(m), rc);
4264 	}
4265 
4266 	KASSERT(gl->sg_nseg == mbuf_nsegs(m),
4267 	    ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m,
4268 	    mbuf_nsegs(m), gl->sg_nseg));
4269 	KASSERT(gl->sg_nseg > 0 &&
4270 	    gl->sg_nseg <= (needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS),
4271 	    ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__,
4272 		gl->sg_nseg, needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS));
4273 }
4274 
4275 /*
4276  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
4277  */
4278 static inline u_int
4279 txpkt_len16(u_int nsegs, u_int tso)
4280 {
4281 	u_int n;
4282 
4283 	MPASS(nsegs > 0);
4284 
4285 	nsegs--; /* first segment is part of ulptx_sgl */
4286 	n = sizeof(struct fw_eth_tx_pkt_wr) + sizeof(struct cpl_tx_pkt_core) +
4287 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4288 	if (tso)
4289 		n += sizeof(struct cpl_tx_pkt_lso_core);
4290 
4291 	return (howmany(n, 16));
4292 }
4293 
4294 /*
4295  * len16 for a txpkt_vm WR with a GL.  Includes the firmware work
4296  * request header.
4297  */
4298 static inline u_int
4299 txpkt_vm_len16(u_int nsegs, u_int tso)
4300 {
4301 	u_int n;
4302 
4303 	MPASS(nsegs > 0);
4304 
4305 	nsegs--; /* first segment is part of ulptx_sgl */
4306 	n = sizeof(struct fw_eth_tx_pkt_vm_wr) +
4307 	    sizeof(struct cpl_tx_pkt_core) +
4308 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4309 	if (tso)
4310 		n += sizeof(struct cpl_tx_pkt_lso_core);
4311 
4312 	return (howmany(n, 16));
4313 }
4314 
4315 /*
4316  * len16 for a txpkts type 0 WR with a GL.  Does not include the firmware work
4317  * request header.
4318  */
4319 static inline u_int
4320 txpkts0_len16(u_int nsegs)
4321 {
4322 	u_int n;
4323 
4324 	MPASS(nsegs > 0);
4325 
4326 	nsegs--; /* first segment is part of ulptx_sgl */
4327 	n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) +
4328 	    sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) +
4329 	    8 * ((3 * nsegs) / 2 + (nsegs & 1));
4330 
4331 	return (howmany(n, 16));
4332 }
4333 
4334 /*
4335  * len16 for a txpkts type 1 WR with a GL.  Does not include the firmware work
4336  * request header.
4337  */
4338 static inline u_int
4339 txpkts1_len16(void)
4340 {
4341 	u_int n;
4342 
4343 	n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl);
4344 
4345 	return (howmany(n, 16));
4346 }
4347 
4348 static inline u_int
4349 imm_payload(u_int ndesc)
4350 {
4351 	u_int n;
4352 
4353 	n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) -
4354 	    sizeof(struct cpl_tx_pkt_core);
4355 
4356 	return (n);
4357 }
4358 
4359 /*
4360  * Write a VM txpkt WR for this packet to the hardware descriptors, update the
4361  * software descriptor, and advance the pidx.  It is guaranteed that enough
4362  * descriptors are available.
4363  *
4364  * The return value is the # of hardware descriptors used.
4365  */
4366 static u_int
4367 write_txpkt_vm_wr(struct adapter *sc, struct sge_txq *txq,
4368     struct fw_eth_tx_pkt_vm_wr *wr, struct mbuf *m0, u_int available)
4369 {
4370 	struct sge_eq *eq = &txq->eq;
4371 	struct tx_sdesc *txsd;
4372 	struct cpl_tx_pkt_core *cpl;
4373 	uint32_t ctrl;	/* used in many unrelated places */
4374 	uint64_t ctrl1;
4375 	int csum_type, len16, ndesc, pktlen, nsegs;
4376 	caddr_t dst;
4377 
4378 	TXQ_LOCK_ASSERT_OWNED(txq);
4379 	M_ASSERTPKTHDR(m0);
4380 	MPASS(available > 0 && available < eq->sidx);
4381 
4382 	len16 = mbuf_len16(m0);
4383 	nsegs = mbuf_nsegs(m0);
4384 	pktlen = m0->m_pkthdr.len;
4385 	ctrl = sizeof(struct cpl_tx_pkt_core);
4386 	if (needs_tso(m0))
4387 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4388 	ndesc = howmany(len16, EQ_ESIZE / 16);
4389 	MPASS(ndesc <= available);
4390 
4391 	/* Firmware work request header */
4392 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4393 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_VM_WR) |
4394 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4395 
4396 	ctrl = V_FW_WR_LEN16(len16);
4397 	wr->equiq_to_len16 = htobe32(ctrl);
4398 	wr->r3[0] = 0;
4399 	wr->r3[1] = 0;
4400 
4401 	/*
4402 	 * Copy over ethmacdst, ethmacsrc, ethtype, and vlantci.
4403 	 * vlantci is ignored unless the ethtype is 0x8100, so it's
4404 	 * simpler to always copy it rather than making it
4405 	 * conditional.  Also, it seems that we do not have to set
4406 	 * vlantci or fake the ethtype when doing VLAN tag insertion.
4407 	 */
4408 	m_copydata(m0, 0, sizeof(struct ether_header) + 2, wr->ethmacdst);
4409 
4410 	csum_type = -1;
4411 	if (needs_tso(m0)) {
4412 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4413 
4414 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4415 		    m0->m_pkthdr.l4hlen > 0,
4416 		    ("%s: mbuf %p needs TSO but missing header lengths",
4417 			__func__, m0));
4418 
4419 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4420 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4421 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4422 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4423 			ctrl |= V_LSO_ETHHDR_LEN(1);
4424 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4425 			ctrl |= F_LSO_IPV6;
4426 
4427 		lso->lso_ctrl = htobe32(ctrl);
4428 		lso->ipid_ofst = htobe16(0);
4429 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4430 		lso->seqno_offset = htobe32(0);
4431 		lso->len = htobe32(pktlen);
4432 
4433 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4434 			csum_type = TX_CSUM_TCPIP6;
4435 		else
4436 			csum_type = TX_CSUM_TCPIP;
4437 
4438 		cpl = (void *)(lso + 1);
4439 
4440 		txq->tso_wrs++;
4441 	} else {
4442 		if (m0->m_pkthdr.csum_flags & CSUM_IP_TCP)
4443 			csum_type = TX_CSUM_TCPIP;
4444 		else if (m0->m_pkthdr.csum_flags & CSUM_IP_UDP)
4445 			csum_type = TX_CSUM_UDPIP;
4446 		else if (m0->m_pkthdr.csum_flags & CSUM_IP6_TCP)
4447 			csum_type = TX_CSUM_TCPIP6;
4448 		else if (m0->m_pkthdr.csum_flags & CSUM_IP6_UDP)
4449 			csum_type = TX_CSUM_UDPIP6;
4450 #if defined(INET)
4451 		else if (m0->m_pkthdr.csum_flags & CSUM_IP) {
4452 			/*
4453 			 * XXX: The firmware appears to stomp on the
4454 			 * fragment/flags field of the IP header when
4455 			 * using TX_CSUM_IP.  Fall back to doing
4456 			 * software checksums.
4457 			 */
4458 			u_short *sump;
4459 			struct mbuf *m;
4460 			int offset;
4461 
4462 			m = m0;
4463 			offset = 0;
4464 			sump = m_advance(&m, &offset, m0->m_pkthdr.l2hlen +
4465 			    offsetof(struct ip, ip_sum));
4466 			*sump = in_cksum_skip(m0, m0->m_pkthdr.l2hlen +
4467 			    m0->m_pkthdr.l3hlen, m0->m_pkthdr.l2hlen);
4468 			m0->m_pkthdr.csum_flags &= ~CSUM_IP;
4469 		}
4470 #endif
4471 
4472 		cpl = (void *)(wr + 1);
4473 	}
4474 
4475 	/* Checksum offload */
4476 	ctrl1 = 0;
4477 	if (needs_l3_csum(m0) == 0)
4478 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4479 	if (csum_type >= 0) {
4480 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0,
4481 	    ("%s: mbuf %p needs checksum offload but missing header lengths",
4482 			__func__, m0));
4483 
4484 		if (chip_id(sc) <= CHELSIO_T5) {
4485 			ctrl1 |= V_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4486 			    ETHER_HDR_LEN);
4487 		} else {
4488 			ctrl1 |= V_T6_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4489 			    ETHER_HDR_LEN);
4490 		}
4491 		ctrl1 |= V_TXPKT_IPHDR_LEN(m0->m_pkthdr.l3hlen);
4492 		ctrl1 |= V_TXPKT_CSUM_TYPE(csum_type);
4493 	} else
4494 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4495 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4496 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4497 		txq->txcsum++;	/* some hardware assistance provided */
4498 
4499 	/* VLAN tag insertion */
4500 	if (needs_vlan_insertion(m0)) {
4501 		ctrl1 |= F_TXPKT_VLAN_VLD |
4502 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4503 		txq->vlan_insertion++;
4504 	}
4505 
4506 	/* CPL header */
4507 	cpl->ctrl0 = txq->cpl_ctrl0;
4508 	cpl->pack = 0;
4509 	cpl->len = htobe16(pktlen);
4510 	cpl->ctrl1 = htobe64(ctrl1);
4511 
4512 	/* SGL */
4513 	dst = (void *)(cpl + 1);
4514 
4515 	/*
4516 	 * A packet using TSO will use up an entire descriptor for the
4517 	 * firmware work request header, LSO CPL, and TX_PKT_XT CPL.
4518 	 * If this descriptor is the last descriptor in the ring, wrap
4519 	 * around to the front of the ring explicitly for the start of
4520 	 * the sgl.
4521 	 */
4522 	if (dst == (void *)&eq->desc[eq->sidx]) {
4523 		dst = (void *)&eq->desc[0];
4524 		write_gl_to_txd(txq, m0, &dst, 0);
4525 	} else
4526 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4527 	txq->sgl_wrs++;
4528 
4529 	txq->txpkt_wrs++;
4530 
4531 	txsd = &txq->sdesc[eq->pidx];
4532 	txsd->m = m0;
4533 	txsd->desc_used = ndesc;
4534 
4535 	return (ndesc);
4536 }
4537 
4538 /*
4539  * Write a txpkt WR for this packet to the hardware descriptors, update the
4540  * software descriptor, and advance the pidx.  It is guaranteed that enough
4541  * descriptors are available.
4542  *
4543  * The return value is the # of hardware descriptors used.
4544  */
4545 static u_int
4546 write_txpkt_wr(struct sge_txq *txq, struct fw_eth_tx_pkt_wr *wr,
4547     struct mbuf *m0, u_int available)
4548 {
4549 	struct sge_eq *eq = &txq->eq;
4550 	struct tx_sdesc *txsd;
4551 	struct cpl_tx_pkt_core *cpl;
4552 	uint32_t ctrl;	/* used in many unrelated places */
4553 	uint64_t ctrl1;
4554 	int len16, ndesc, pktlen, nsegs;
4555 	caddr_t dst;
4556 
4557 	TXQ_LOCK_ASSERT_OWNED(txq);
4558 	M_ASSERTPKTHDR(m0);
4559 	MPASS(available > 0 && available < eq->sidx);
4560 
4561 	len16 = mbuf_len16(m0);
4562 	nsegs = mbuf_nsegs(m0);
4563 	pktlen = m0->m_pkthdr.len;
4564 	ctrl = sizeof(struct cpl_tx_pkt_core);
4565 	if (needs_tso(m0))
4566 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4567 	else if (pktlen <= imm_payload(2) && available >= 2) {
4568 		/* Immediate data.  Recalculate len16 and set nsegs to 0. */
4569 		ctrl += pktlen;
4570 		len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) +
4571 		    sizeof(struct cpl_tx_pkt_core) + pktlen, 16);
4572 		nsegs = 0;
4573 	}
4574 	ndesc = howmany(len16, EQ_ESIZE / 16);
4575 	MPASS(ndesc <= available);
4576 
4577 	/* Firmware work request header */
4578 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4579 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
4580 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4581 
4582 	ctrl = V_FW_WR_LEN16(len16);
4583 	wr->equiq_to_len16 = htobe32(ctrl);
4584 	wr->r3 = 0;
4585 
4586 	if (needs_tso(m0)) {
4587 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4588 
4589 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4590 		    m0->m_pkthdr.l4hlen > 0,
4591 		    ("%s: mbuf %p needs TSO but missing header lengths",
4592 			__func__, m0));
4593 
4594 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4595 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4596 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4597 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4598 			ctrl |= V_LSO_ETHHDR_LEN(1);
4599 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4600 			ctrl |= F_LSO_IPV6;
4601 
4602 		lso->lso_ctrl = htobe32(ctrl);
4603 		lso->ipid_ofst = htobe16(0);
4604 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4605 		lso->seqno_offset = htobe32(0);
4606 		lso->len = htobe32(pktlen);
4607 
4608 		cpl = (void *)(lso + 1);
4609 
4610 		txq->tso_wrs++;
4611 	} else
4612 		cpl = (void *)(wr + 1);
4613 
4614 	/* Checksum offload */
4615 	ctrl1 = 0;
4616 	if (needs_l3_csum(m0) == 0)
4617 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4618 	if (needs_l4_csum(m0) == 0)
4619 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4620 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4621 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4622 		txq->txcsum++;	/* some hardware assistance provided */
4623 
4624 	/* VLAN tag insertion */
4625 	if (needs_vlan_insertion(m0)) {
4626 		ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4627 		txq->vlan_insertion++;
4628 	}
4629 
4630 	/* CPL header */
4631 	cpl->ctrl0 = txq->cpl_ctrl0;
4632 	cpl->pack = 0;
4633 	cpl->len = htobe16(pktlen);
4634 	cpl->ctrl1 = htobe64(ctrl1);
4635 
4636 	/* SGL */
4637 	dst = (void *)(cpl + 1);
4638 	if (nsegs > 0) {
4639 
4640 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4641 		txq->sgl_wrs++;
4642 	} else {
4643 		struct mbuf *m;
4644 
4645 		for (m = m0; m != NULL; m = m->m_next) {
4646 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4647 #ifdef INVARIANTS
4648 			pktlen -= m->m_len;
4649 #endif
4650 		}
4651 #ifdef INVARIANTS
4652 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
4653 #endif
4654 		txq->imm_wrs++;
4655 	}
4656 
4657 	txq->txpkt_wrs++;
4658 
4659 	txsd = &txq->sdesc[eq->pidx];
4660 	txsd->m = m0;
4661 	txsd->desc_used = ndesc;
4662 
4663 	return (ndesc);
4664 }
4665 
4666 static int
4667 try_txpkts(struct mbuf *m, struct mbuf *n, struct txpkts *txp, u_int available)
4668 {
4669 	u_int needed, nsegs1, nsegs2, l1, l2;
4670 
4671 	if (cannot_use_txpkts(m) || cannot_use_txpkts(n))
4672 		return (1);
4673 
4674 	nsegs1 = mbuf_nsegs(m);
4675 	nsegs2 = mbuf_nsegs(n);
4676 	if (nsegs1 + nsegs2 == 2) {
4677 		txp->wr_type = 1;
4678 		l1 = l2 = txpkts1_len16();
4679 	} else {
4680 		txp->wr_type = 0;
4681 		l1 = txpkts0_len16(nsegs1);
4682 		l2 = txpkts0_len16(nsegs2);
4683 	}
4684 	txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) + l1 + l2;
4685 	needed = howmany(txp->len16, EQ_ESIZE / 16);
4686 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4687 		return (1);
4688 
4689 	txp->plen = m->m_pkthdr.len + n->m_pkthdr.len;
4690 	if (txp->plen > 65535)
4691 		return (1);
4692 
4693 	txp->npkt = 2;
4694 	set_mbuf_len16(m, l1);
4695 	set_mbuf_len16(n, l2);
4696 
4697 	return (0);
4698 }
4699 
4700 static int
4701 add_to_txpkts(struct mbuf *m, struct txpkts *txp, u_int available)
4702 {
4703 	u_int plen, len16, needed, nsegs;
4704 
4705 	MPASS(txp->wr_type == 0 || txp->wr_type == 1);
4706 
4707 	nsegs = mbuf_nsegs(m);
4708 	if (needs_tso(m) || (txp->wr_type == 1 && nsegs != 1))
4709 		return (1);
4710 
4711 	plen = txp->plen + m->m_pkthdr.len;
4712 	if (plen > 65535)
4713 		return (1);
4714 
4715 	if (txp->wr_type == 0)
4716 		len16 = txpkts0_len16(nsegs);
4717 	else
4718 		len16 = txpkts1_len16();
4719 	needed = howmany(txp->len16 + len16, EQ_ESIZE / 16);
4720 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4721 		return (1);
4722 
4723 	txp->npkt++;
4724 	txp->plen = plen;
4725 	txp->len16 += len16;
4726 	set_mbuf_len16(m, len16);
4727 
4728 	return (0);
4729 }
4730 
4731 /*
4732  * Write a txpkts WR for the packets in txp to the hardware descriptors, update
4733  * the software descriptor, and advance the pidx.  It is guaranteed that enough
4734  * descriptors are available.
4735  *
4736  * The return value is the # of hardware descriptors used.
4737  */
4738 static u_int
4739 write_txpkts_wr(struct sge_txq *txq, struct fw_eth_tx_pkts_wr *wr,
4740     struct mbuf *m0, const struct txpkts *txp, u_int available)
4741 {
4742 	struct sge_eq *eq = &txq->eq;
4743 	struct tx_sdesc *txsd;
4744 	struct cpl_tx_pkt_core *cpl;
4745 	uint32_t ctrl;
4746 	uint64_t ctrl1;
4747 	int ndesc, checkwrap;
4748 	struct mbuf *m;
4749 	void *flitp;
4750 
4751 	TXQ_LOCK_ASSERT_OWNED(txq);
4752 	MPASS(txp->npkt > 0);
4753 	MPASS(txp->plen < 65536);
4754 	MPASS(m0 != NULL);
4755 	MPASS(m0->m_nextpkt != NULL);
4756 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
4757 	MPASS(available > 0 && available < eq->sidx);
4758 
4759 	ndesc = howmany(txp->len16, EQ_ESIZE / 16);
4760 	MPASS(ndesc <= available);
4761 
4762 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4763 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
4764 	ctrl = V_FW_WR_LEN16(txp->len16);
4765 	wr->equiq_to_len16 = htobe32(ctrl);
4766 	wr->plen = htobe16(txp->plen);
4767 	wr->npkt = txp->npkt;
4768 	wr->r3 = 0;
4769 	wr->type = txp->wr_type;
4770 	flitp = wr + 1;
4771 
4772 	/*
4773 	 * At this point we are 16B into a hardware descriptor.  If checkwrap is
4774 	 * set then we know the WR is going to wrap around somewhere.  We'll
4775 	 * check for that at appropriate points.
4776 	 */
4777 	checkwrap = eq->sidx - ndesc < eq->pidx;
4778 	for (m = m0; m != NULL; m = m->m_nextpkt) {
4779 		if (txp->wr_type == 0) {
4780 			struct ulp_txpkt *ulpmc;
4781 			struct ulptx_idata *ulpsc;
4782 
4783 			/* ULP master command */
4784 			ulpmc = flitp;
4785 			ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
4786 			    V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid));
4787 			ulpmc->len = htobe32(mbuf_len16(m));
4788 
4789 			/* ULP subcommand */
4790 			ulpsc = (void *)(ulpmc + 1);
4791 			ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
4792 			    F_ULP_TX_SC_MORE);
4793 			ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
4794 
4795 			cpl = (void *)(ulpsc + 1);
4796 			if (checkwrap &&
4797 			    (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx])
4798 				cpl = (void *)&eq->desc[0];
4799 		} else {
4800 			cpl = flitp;
4801 		}
4802 
4803 		/* Checksum offload */
4804 		ctrl1 = 0;
4805 		if (needs_l3_csum(m) == 0)
4806 			ctrl1 |= F_TXPKT_IPCSUM_DIS;
4807 		if (needs_l4_csum(m) == 0)
4808 			ctrl1 |= F_TXPKT_L4CSUM_DIS;
4809 		if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4810 		    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4811 			txq->txcsum++;	/* some hardware assistance provided */
4812 
4813 		/* VLAN tag insertion */
4814 		if (needs_vlan_insertion(m)) {
4815 			ctrl1 |= F_TXPKT_VLAN_VLD |
4816 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
4817 			txq->vlan_insertion++;
4818 		}
4819 
4820 		/* CPL header */
4821 		cpl->ctrl0 = txq->cpl_ctrl0;
4822 		cpl->pack = 0;
4823 		cpl->len = htobe16(m->m_pkthdr.len);
4824 		cpl->ctrl1 = htobe64(ctrl1);
4825 
4826 		flitp = cpl + 1;
4827 		if (checkwrap &&
4828 		    (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
4829 			flitp = (void *)&eq->desc[0];
4830 
4831 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap);
4832 
4833 	}
4834 
4835 	if (txp->wr_type == 0) {
4836 		txq->txpkts0_pkts += txp->npkt;
4837 		txq->txpkts0_wrs++;
4838 	} else {
4839 		txq->txpkts1_pkts += txp->npkt;
4840 		txq->txpkts1_wrs++;
4841 	}
4842 
4843 	txsd = &txq->sdesc[eq->pidx];
4844 	txsd->m = m0;
4845 	txsd->desc_used = ndesc;
4846 
4847 	return (ndesc);
4848 }
4849 
4850 /*
4851  * If the SGL ends on an address that is not 16 byte aligned, this function will
4852  * add a 0 filled flit at the end.
4853  */
4854 static void
4855 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap)
4856 {
4857 	struct sge_eq *eq = &txq->eq;
4858 	struct sglist *gl = txq->gl;
4859 	struct sglist_seg *seg;
4860 	__be64 *flitp, *wrap;
4861 	struct ulptx_sgl *usgl;
4862 	int i, nflits, nsegs;
4863 
4864 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
4865 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
4866 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
4867 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
4868 
4869 	get_pkt_gl(m, gl);
4870 	nsegs = gl->sg_nseg;
4871 	MPASS(nsegs > 0);
4872 
4873 	nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
4874 	flitp = (__be64 *)(*to);
4875 	wrap = (__be64 *)(&eq->desc[eq->sidx]);
4876 	seg = &gl->sg_segs[0];
4877 	usgl = (void *)flitp;
4878 
4879 	/*
4880 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
4881 	 * ring, so we're at least 16 bytes away from the status page.  There is
4882 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
4883 	 */
4884 
4885 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
4886 	    V_ULPTX_NSGE(nsegs));
4887 	usgl->len0 = htobe32(seg->ss_len);
4888 	usgl->addr0 = htobe64(seg->ss_paddr);
4889 	seg++;
4890 
4891 	if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) {
4892 
4893 		/* Won't wrap around at all */
4894 
4895 		for (i = 0; i < nsegs - 1; i++, seg++) {
4896 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
4897 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
4898 		}
4899 		if (i & 1)
4900 			usgl->sge[i / 2].len[1] = htobe32(0);
4901 		flitp += nflits;
4902 	} else {
4903 
4904 		/* Will wrap somewhere in the rest of the SGL */
4905 
4906 		/* 2 flits already written, write the rest flit by flit */
4907 		flitp = (void *)(usgl + 1);
4908 		for (i = 0; i < nflits - 2; i++) {
4909 			if (flitp == wrap)
4910 				flitp = (void *)eq->desc;
4911 			*flitp++ = get_flit(seg, nsegs - 1, i);
4912 		}
4913 	}
4914 
4915 	if (nflits & 1) {
4916 		MPASS(((uintptr_t)flitp) & 0xf);
4917 		*flitp++ = 0;
4918 	}
4919 
4920 	MPASS((((uintptr_t)flitp) & 0xf) == 0);
4921 	if (__predict_false(flitp == wrap))
4922 		*to = (void *)eq->desc;
4923 	else
4924 		*to = (void *)flitp;
4925 }
4926 
4927 static inline void
4928 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
4929 {
4930 
4931 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
4932 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
4933 
4934 	if (__predict_true((uintptr_t)(*to) + len <=
4935 	    (uintptr_t)&eq->desc[eq->sidx])) {
4936 		bcopy(from, *to, len);
4937 		(*to) += len;
4938 	} else {
4939 		int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
4940 
4941 		bcopy(from, *to, portion);
4942 		from += portion;
4943 		portion = len - portion;	/* remaining */
4944 		bcopy(from, (void *)eq->desc, portion);
4945 		(*to) = (caddr_t)eq->desc + portion;
4946 	}
4947 }
4948 
4949 static inline void
4950 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n)
4951 {
4952 	u_int db;
4953 
4954 	MPASS(n > 0);
4955 
4956 	db = eq->doorbells;
4957 	if (n > 1)
4958 		clrbit(&db, DOORBELL_WCWR);
4959 	wmb();
4960 
4961 	switch (ffs(db) - 1) {
4962 	case DOORBELL_UDB:
4963 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
4964 		break;
4965 
4966 	case DOORBELL_WCWR: {
4967 		volatile uint64_t *dst, *src;
4968 		int i;
4969 
4970 		/*
4971 		 * Queues whose 128B doorbell segment fits in the page do not
4972 		 * use relative qid (udb_qid is always 0).  Only queues with
4973 		 * doorbell segments can do WCWR.
4974 		 */
4975 		KASSERT(eq->udb_qid == 0 && n == 1,
4976 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
4977 		    __func__, eq->doorbells, n, eq->dbidx, eq));
4978 
4979 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
4980 		    UDBS_DB_OFFSET);
4981 		i = eq->dbidx;
4982 		src = (void *)&eq->desc[i];
4983 		while (src != (void *)&eq->desc[i + 1])
4984 			*dst++ = *src++;
4985 		wmb();
4986 		break;
4987 	}
4988 
4989 	case DOORBELL_UDBWC:
4990 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
4991 		wmb();
4992 		break;
4993 
4994 	case DOORBELL_KDB:
4995 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
4996 		    V_QID(eq->cntxt_id) | V_PIDX(n));
4997 		break;
4998 	}
4999 
5000 	IDXINCR(eq->dbidx, n, eq->sidx);
5001 }
5002 
5003 static inline u_int
5004 reclaimable_tx_desc(struct sge_eq *eq)
5005 {
5006 	uint16_t hw_cidx;
5007 
5008 	hw_cidx = read_hw_cidx(eq);
5009 	return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx));
5010 }
5011 
5012 static inline u_int
5013 total_available_tx_desc(struct sge_eq *eq)
5014 {
5015 	uint16_t hw_cidx, pidx;
5016 
5017 	hw_cidx = read_hw_cidx(eq);
5018 	pidx = eq->pidx;
5019 
5020 	if (pidx == hw_cidx)
5021 		return (eq->sidx - 1);
5022 	else
5023 		return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1);
5024 }
5025 
5026 static inline uint16_t
5027 read_hw_cidx(struct sge_eq *eq)
5028 {
5029 	struct sge_qstat *spg = (void *)&eq->desc[eq->sidx];
5030 	uint16_t cidx = spg->cidx;	/* stable snapshot */
5031 
5032 	return (be16toh(cidx));
5033 }
5034 
5035 /*
5036  * Reclaim 'n' descriptors approximately.
5037  */
5038 static u_int
5039 reclaim_tx_descs(struct sge_txq *txq, u_int n)
5040 {
5041 	struct tx_sdesc *txsd;
5042 	struct sge_eq *eq = &txq->eq;
5043 	u_int can_reclaim, reclaimed;
5044 
5045 	TXQ_LOCK_ASSERT_OWNED(txq);
5046 	MPASS(n > 0);
5047 
5048 	reclaimed = 0;
5049 	can_reclaim = reclaimable_tx_desc(eq);
5050 	while (can_reclaim && reclaimed < n) {
5051 		int ndesc;
5052 		struct mbuf *m, *nextpkt;
5053 
5054 		txsd = &txq->sdesc[eq->cidx];
5055 		ndesc = txsd->desc_used;
5056 
5057 		/* Firmware doesn't return "partial" credits. */
5058 		KASSERT(can_reclaim >= ndesc,
5059 		    ("%s: unexpected number of credits: %d, %d",
5060 		    __func__, can_reclaim, ndesc));
5061 
5062 		for (m = txsd->m; m != NULL; m = nextpkt) {
5063 			nextpkt = m->m_nextpkt;
5064 			m->m_nextpkt = NULL;
5065 			m_freem(m);
5066 		}
5067 		reclaimed += ndesc;
5068 		can_reclaim -= ndesc;
5069 		IDXINCR(eq->cidx, ndesc, eq->sidx);
5070 	}
5071 
5072 	return (reclaimed);
5073 }
5074 
5075 static void
5076 tx_reclaim(void *arg, int n)
5077 {
5078 	struct sge_txq *txq = arg;
5079 	struct sge_eq *eq = &txq->eq;
5080 
5081 	do {
5082 		if (TXQ_TRYLOCK(txq) == 0)
5083 			break;
5084 		n = reclaim_tx_descs(txq, 32);
5085 		if (eq->cidx == eq->pidx)
5086 			eq->equeqidx = eq->pidx;
5087 		TXQ_UNLOCK(txq);
5088 	} while (n > 0);
5089 }
5090 
5091 static __be64
5092 get_flit(struct sglist_seg *segs, int nsegs, int idx)
5093 {
5094 	int i = (idx / 3) * 2;
5095 
5096 	switch (idx % 3) {
5097 	case 0: {
5098 		uint64_t rc;
5099 
5100 		rc = (uint64_t)segs[i].ss_len << 32;
5101 		if (i + 1 < nsegs)
5102 			rc |= (uint64_t)(segs[i + 1].ss_len);
5103 
5104 		return (htobe64(rc));
5105 	}
5106 	case 1:
5107 		return (htobe64(segs[i].ss_paddr));
5108 	case 2:
5109 		return (htobe64(segs[i + 1].ss_paddr));
5110 	}
5111 
5112 	return (0);
5113 }
5114 
5115 static void
5116 find_best_refill_source(struct adapter *sc, struct sge_fl *fl, int maxp)
5117 {
5118 	int8_t zidx, hwidx, idx;
5119 	uint16_t region1, region3;
5120 	int spare, spare_needed, n;
5121 	struct sw_zone_info *swz;
5122 	struct hw_buf_info *hwb, *hwb_list = &sc->sge.hw_buf_info[0];
5123 
5124 	/*
5125 	 * Buffer Packing: Look for PAGE_SIZE or larger zone which has a bufsize
5126 	 * large enough for the max payload and cluster metadata.  Otherwise
5127 	 * settle for the largest bufsize that leaves enough room in the cluster
5128 	 * for metadata.
5129 	 *
5130 	 * Without buffer packing: Look for the smallest zone which has a
5131 	 * bufsize large enough for the max payload.  Settle for the largest
5132 	 * bufsize available if there's nothing big enough for max payload.
5133 	 */
5134 	spare_needed = fl->flags & FL_BUF_PACKING ? CL_METADATA_SIZE : 0;
5135 	swz = &sc->sge.sw_zone_info[0];
5136 	hwidx = -1;
5137 	for (zidx = 0; zidx < SW_ZONE_SIZES; zidx++, swz++) {
5138 		if (swz->size > largest_rx_cluster) {
5139 			if (__predict_true(hwidx != -1))
5140 				break;
5141 
5142 			/*
5143 			 * This is a misconfiguration.  largest_rx_cluster is
5144 			 * preventing us from finding a refill source.  See
5145 			 * dev.t5nex.<n>.buffer_sizes to figure out why.
5146 			 */
5147 			device_printf(sc->dev, "largest_rx_cluster=%u leaves no"
5148 			    " refill source for fl %p (dma %u).  Ignored.\n",
5149 			    largest_rx_cluster, fl, maxp);
5150 		}
5151 		for (idx = swz->head_hwidx; idx != -1; idx = hwb->next) {
5152 			hwb = &hwb_list[idx];
5153 			spare = swz->size - hwb->size;
5154 			if (spare < spare_needed)
5155 				continue;
5156 
5157 			hwidx = idx;		/* best option so far */
5158 			if (hwb->size >= maxp) {
5159 
5160 				if ((fl->flags & FL_BUF_PACKING) == 0)
5161 					goto done; /* stop looking (not packing) */
5162 
5163 				if (swz->size >= safest_rx_cluster)
5164 					goto done; /* stop looking (packing) */
5165 			}
5166 			break;		/* keep looking, next zone */
5167 		}
5168 	}
5169 done:
5170 	/* A usable hwidx has been located. */
5171 	MPASS(hwidx != -1);
5172 	hwb = &hwb_list[hwidx];
5173 	zidx = hwb->zidx;
5174 	swz = &sc->sge.sw_zone_info[zidx];
5175 	region1 = 0;
5176 	region3 = swz->size - hwb->size;
5177 
5178 	/*
5179 	 * Stay within this zone and see if there is a better match when mbuf
5180 	 * inlining is allowed.  Remember that the hwidx's are sorted in
5181 	 * decreasing order of size (so in increasing order of spare area).
5182 	 */
5183 	for (idx = hwidx; idx != -1; idx = hwb->next) {
5184 		hwb = &hwb_list[idx];
5185 		spare = swz->size - hwb->size;
5186 
5187 		if (allow_mbufs_in_cluster == 0 || hwb->size < maxp)
5188 			break;
5189 
5190 		/*
5191 		 * Do not inline mbufs if doing so would violate the pad/pack
5192 		 * boundary alignment requirement.
5193 		 */
5194 		if (fl_pad && (MSIZE % sc->params.sge.pad_boundary) != 0)
5195 			continue;
5196 		if (fl->flags & FL_BUF_PACKING &&
5197 		    (MSIZE % sc->params.sge.pack_boundary) != 0)
5198 			continue;
5199 
5200 		if (spare < CL_METADATA_SIZE + MSIZE)
5201 			continue;
5202 		n = (spare - CL_METADATA_SIZE) / MSIZE;
5203 		if (n > howmany(hwb->size, maxp))
5204 			break;
5205 
5206 		hwidx = idx;
5207 		if (fl->flags & FL_BUF_PACKING) {
5208 			region1 = n * MSIZE;
5209 			region3 = spare - region1;
5210 		} else {
5211 			region1 = MSIZE;
5212 			region3 = spare - region1;
5213 			break;
5214 		}
5215 	}
5216 
5217 	KASSERT(zidx >= 0 && zidx < SW_ZONE_SIZES,
5218 	    ("%s: bad zone %d for fl %p, maxp %d", __func__, zidx, fl, maxp));
5219 	KASSERT(hwidx >= 0 && hwidx <= SGE_FLBUF_SIZES,
5220 	    ("%s: bad hwidx %d for fl %p, maxp %d", __func__, hwidx, fl, maxp));
5221 	KASSERT(region1 + sc->sge.hw_buf_info[hwidx].size + region3 ==
5222 	    sc->sge.sw_zone_info[zidx].size,
5223 	    ("%s: bad buffer layout for fl %p, maxp %d. "
5224 		"cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5225 		sc->sge.sw_zone_info[zidx].size, region1,
5226 		sc->sge.hw_buf_info[hwidx].size, region3));
5227 	if (fl->flags & FL_BUF_PACKING || region1 > 0) {
5228 		KASSERT(region3 >= CL_METADATA_SIZE,
5229 		    ("%s: no room for metadata.  fl %p, maxp %d; "
5230 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5231 		    sc->sge.sw_zone_info[zidx].size, region1,
5232 		    sc->sge.hw_buf_info[hwidx].size, region3));
5233 		KASSERT(region1 % MSIZE == 0,
5234 		    ("%s: bad mbuf region for fl %p, maxp %d. "
5235 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5236 		    sc->sge.sw_zone_info[zidx].size, region1,
5237 		    sc->sge.hw_buf_info[hwidx].size, region3));
5238 	}
5239 
5240 	fl->cll_def.zidx = zidx;
5241 	fl->cll_def.hwidx = hwidx;
5242 	fl->cll_def.region1 = region1;
5243 	fl->cll_def.region3 = region3;
5244 }
5245 
5246 static void
5247 find_safe_refill_source(struct adapter *sc, struct sge_fl *fl)
5248 {
5249 	struct sge *s = &sc->sge;
5250 	struct hw_buf_info *hwb;
5251 	struct sw_zone_info *swz;
5252 	int spare;
5253 	int8_t hwidx;
5254 
5255 	if (fl->flags & FL_BUF_PACKING)
5256 		hwidx = s->safe_hwidx2;	/* with room for metadata */
5257 	else if (allow_mbufs_in_cluster && s->safe_hwidx2 != -1) {
5258 		hwidx = s->safe_hwidx2;
5259 		hwb = &s->hw_buf_info[hwidx];
5260 		swz = &s->sw_zone_info[hwb->zidx];
5261 		spare = swz->size - hwb->size;
5262 
5263 		/* no good if there isn't room for an mbuf as well */
5264 		if (spare < CL_METADATA_SIZE + MSIZE)
5265 			hwidx = s->safe_hwidx1;
5266 	} else
5267 		hwidx = s->safe_hwidx1;
5268 
5269 	if (hwidx == -1) {
5270 		/* No fallback source */
5271 		fl->cll_alt.hwidx = -1;
5272 		fl->cll_alt.zidx = -1;
5273 
5274 		return;
5275 	}
5276 
5277 	hwb = &s->hw_buf_info[hwidx];
5278 	swz = &s->sw_zone_info[hwb->zidx];
5279 	spare = swz->size - hwb->size;
5280 	fl->cll_alt.hwidx = hwidx;
5281 	fl->cll_alt.zidx = hwb->zidx;
5282 	if (allow_mbufs_in_cluster &&
5283 	    (fl_pad == 0 || (MSIZE % sc->params.sge.pad_boundary) == 0))
5284 		fl->cll_alt.region1 = ((spare - CL_METADATA_SIZE) / MSIZE) * MSIZE;
5285 	else
5286 		fl->cll_alt.region1 = 0;
5287 	fl->cll_alt.region3 = spare - fl->cll_alt.region1;
5288 }
5289 
5290 static void
5291 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
5292 {
5293 	mtx_lock(&sc->sfl_lock);
5294 	FL_LOCK(fl);
5295 	if ((fl->flags & FL_DOOMED) == 0) {
5296 		fl->flags |= FL_STARVING;
5297 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
5298 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
5299 	}
5300 	FL_UNLOCK(fl);
5301 	mtx_unlock(&sc->sfl_lock);
5302 }
5303 
5304 static void
5305 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq)
5306 {
5307 	struct sge_wrq *wrq = (void *)eq;
5308 
5309 	atomic_readandclear_int(&eq->equiq);
5310 	taskqueue_enqueue(sc->tq[eq->tx_chan], &wrq->wrq_tx_task);
5311 }
5312 
5313 static void
5314 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq)
5315 {
5316 	struct sge_txq *txq = (void *)eq;
5317 
5318 	MPASS((eq->flags & EQ_TYPEMASK) == EQ_ETH);
5319 
5320 	atomic_readandclear_int(&eq->equiq);
5321 	mp_ring_check_drainage(txq->r, 0);
5322 	taskqueue_enqueue(sc->tq[eq->tx_chan], &txq->tx_reclaim_task);
5323 }
5324 
5325 static int
5326 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
5327     struct mbuf *m)
5328 {
5329 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
5330 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
5331 	struct adapter *sc = iq->adapter;
5332 	struct sge *s = &sc->sge;
5333 	struct sge_eq *eq;
5334 	static void (*h[])(struct adapter *, struct sge_eq *) = {NULL,
5335 		&handle_wrq_egr_update, &handle_eth_egr_update,
5336 		&handle_wrq_egr_update};
5337 
5338 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5339 	    rss->opcode));
5340 
5341 	eq = s->eqmap[qid - s->eq_start - s->eq_base];
5342 	(*h[eq->flags & EQ_TYPEMASK])(sc, eq);
5343 
5344 	return (0);
5345 }
5346 
5347 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
5348 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
5349     offsetof(struct cpl_fw6_msg, data));
5350 
5351 static int
5352 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
5353 {
5354 	struct adapter *sc = iq->adapter;
5355 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
5356 
5357 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5358 	    rss->opcode));
5359 
5360 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
5361 		const struct rss_header *rss2;
5362 
5363 		rss2 = (const struct rss_header *)&cpl->data[0];
5364 		return (t4_cpl_handler[rss2->opcode](iq, rss2, m));
5365 	}
5366 
5367 	return (t4_fw_msg_handler[cpl->type](sc, &cpl->data[0]));
5368 }
5369 
5370 /**
5371  *	t4_handle_wrerr_rpl - process a FW work request error message
5372  *	@adap: the adapter
5373  *	@rpl: start of the FW message
5374  */
5375 static int
5376 t4_handle_wrerr_rpl(struct adapter *adap, const __be64 *rpl)
5377 {
5378 	u8 opcode = *(const u8 *)rpl;
5379 	const struct fw_error_cmd *e = (const void *)rpl;
5380 	unsigned int i;
5381 
5382 	if (opcode != FW_ERROR_CMD) {
5383 		log(LOG_ERR,
5384 		    "%s: Received WRERR_RPL message with opcode %#x\n",
5385 		    device_get_nameunit(adap->dev), opcode);
5386 		return (EINVAL);
5387 	}
5388 	log(LOG_ERR, "%s: FW_ERROR (%s) ", device_get_nameunit(adap->dev),
5389 	    G_FW_ERROR_CMD_FATAL(be32toh(e->op_to_type)) ? "fatal" :
5390 	    "non-fatal");
5391 	switch (G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type))) {
5392 	case FW_ERROR_TYPE_EXCEPTION:
5393 		log(LOG_ERR, "exception info:\n");
5394 		for (i = 0; i < nitems(e->u.exception.info); i++)
5395 			log(LOG_ERR, "%s%08x", i == 0 ? "\t" : " ",
5396 			    be32toh(e->u.exception.info[i]));
5397 		log(LOG_ERR, "\n");
5398 		break;
5399 	case FW_ERROR_TYPE_HWMODULE:
5400 		log(LOG_ERR, "HW module regaddr %08x regval %08x\n",
5401 		    be32toh(e->u.hwmodule.regaddr),
5402 		    be32toh(e->u.hwmodule.regval));
5403 		break;
5404 	case FW_ERROR_TYPE_WR:
5405 		log(LOG_ERR, "WR cidx %d PF %d VF %d eqid %d hdr:\n",
5406 		    be16toh(e->u.wr.cidx),
5407 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.wr.pfn_vfn)),
5408 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.wr.pfn_vfn)),
5409 		    be32toh(e->u.wr.eqid));
5410 		for (i = 0; i < nitems(e->u.wr.wrhdr); i++)
5411 			log(LOG_ERR, "%s%02x", i == 0 ? "\t" : " ",
5412 			    e->u.wr.wrhdr[i]);
5413 		log(LOG_ERR, "\n");
5414 		break;
5415 	case FW_ERROR_TYPE_ACL:
5416 		log(LOG_ERR, "ACL cidx %d PF %d VF %d eqid %d %s",
5417 		    be16toh(e->u.acl.cidx),
5418 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.acl.pfn_vfn)),
5419 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.acl.pfn_vfn)),
5420 		    be32toh(e->u.acl.eqid),
5421 		    G_FW_ERROR_CMD_MV(be16toh(e->u.acl.mv_pkd)) ? "vlanid" :
5422 		    "MAC");
5423 		for (i = 0; i < nitems(e->u.acl.val); i++)
5424 			log(LOG_ERR, " %02x", e->u.acl.val[i]);
5425 		log(LOG_ERR, "\n");
5426 		break;
5427 	default:
5428 		log(LOG_ERR, "type %#x\n",
5429 		    G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type)));
5430 		return (EINVAL);
5431 	}
5432 	return (0);
5433 }
5434 
5435 static int
5436 sysctl_uint16(SYSCTL_HANDLER_ARGS)
5437 {
5438 	uint16_t *id = arg1;
5439 	int i = *id;
5440 
5441 	return sysctl_handle_int(oidp, &i, 0, req);
5442 }
5443 
5444 static int
5445 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
5446 {
5447 	struct sge *s = arg1;
5448 	struct hw_buf_info *hwb = &s->hw_buf_info[0];
5449 	struct sw_zone_info *swz = &s->sw_zone_info[0];
5450 	int i, rc;
5451 	struct sbuf sb;
5452 	char c;
5453 
5454 	sbuf_new(&sb, NULL, 32, SBUF_AUTOEXTEND);
5455 	for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
5456 		if (hwb->zidx >= 0 && swz[hwb->zidx].size <= largest_rx_cluster)
5457 			c = '*';
5458 		else
5459 			c = '\0';
5460 
5461 		sbuf_printf(&sb, "%u%c ", hwb->size, c);
5462 	}
5463 	sbuf_trim(&sb);
5464 	sbuf_finish(&sb);
5465 	rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
5466 	sbuf_delete(&sb);
5467 	return (rc);
5468 }
5469 
5470 #ifdef RATELIMIT
5471 /*
5472  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
5473  */
5474 static inline u_int
5475 txpkt_eo_len16(u_int nsegs, u_int immhdrs, u_int tso)
5476 {
5477 	u_int n;
5478 
5479 	MPASS(immhdrs > 0);
5480 
5481 	n = roundup2(sizeof(struct fw_eth_tx_eo_wr) +
5482 	    sizeof(struct cpl_tx_pkt_core) + immhdrs, 16);
5483 	if (__predict_false(nsegs == 0))
5484 		goto done;
5485 
5486 	nsegs--; /* first segment is part of ulptx_sgl */
5487 	n += sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
5488 	if (tso)
5489 		n += sizeof(struct cpl_tx_pkt_lso_core);
5490 
5491 done:
5492 	return (howmany(n, 16));
5493 }
5494 
5495 #define ETID_FLOWC_NPARAMS 6
5496 #define ETID_FLOWC_LEN (roundup2((sizeof(struct fw_flowc_wr) + \
5497     ETID_FLOWC_NPARAMS * sizeof(struct fw_flowc_mnemval)), 16))
5498 #define ETID_FLOWC_LEN16 (howmany(ETID_FLOWC_LEN, 16))
5499 
5500 static int
5501 send_etid_flowc_wr(struct cxgbe_snd_tag *cst, struct port_info *pi,
5502     struct vi_info *vi)
5503 {
5504 	struct wrq_cookie cookie;
5505 	u_int pfvf = G_FW_VIID_PFN(vi->viid) << S_FW_VIID_PFN;
5506 	struct fw_flowc_wr *flowc;
5507 
5508 	mtx_assert(&cst->lock, MA_OWNED);
5509 	MPASS((cst->flags & (EO_FLOWC_PENDING | EO_FLOWC_RPL_PENDING)) ==
5510 	    EO_FLOWC_PENDING);
5511 
5512 	flowc = start_wrq_wr(cst->eo_txq, ETID_FLOWC_LEN16, &cookie);
5513 	if (__predict_false(flowc == NULL))
5514 		return (ENOMEM);
5515 
5516 	bzero(flowc, ETID_FLOWC_LEN);
5517 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
5518 	    V_FW_FLOWC_WR_NPARAMS(ETID_FLOWC_NPARAMS) | V_FW_WR_COMPL(0));
5519 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(ETID_FLOWC_LEN16) |
5520 	    V_FW_WR_FLOWID(cst->etid));
5521 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN;
5522 	flowc->mnemval[0].val = htobe32(pfvf);
5523 	flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH;
5524 	flowc->mnemval[1].val = htobe32(pi->tx_chan);
5525 	flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT;
5526 	flowc->mnemval[2].val = htobe32(pi->tx_chan);
5527 	flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID;
5528 	flowc->mnemval[3].val = htobe32(cst->iqid);
5529 	flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_EOSTATE;
5530 	flowc->mnemval[4].val = htobe32(FW_FLOWC_MNEM_EOSTATE_ESTABLISHED);
5531 	flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS;
5532 	flowc->mnemval[5].val = htobe32(cst->schedcl);
5533 
5534 	commit_wrq_wr(cst->eo_txq, flowc, &cookie);
5535 
5536 	cst->flags &= ~EO_FLOWC_PENDING;
5537 	cst->flags |= EO_FLOWC_RPL_PENDING;
5538 	MPASS(cst->tx_credits >= ETID_FLOWC_LEN16);	/* flowc is first WR. */
5539 	cst->tx_credits -= ETID_FLOWC_LEN16;
5540 
5541 	return (0);
5542 }
5543 
5544 #define ETID_FLUSH_LEN16 (howmany(sizeof (struct fw_flowc_wr), 16))
5545 
5546 void
5547 send_etid_flush_wr(struct cxgbe_snd_tag *cst)
5548 {
5549 	struct fw_flowc_wr *flowc;
5550 	struct wrq_cookie cookie;
5551 
5552 	mtx_assert(&cst->lock, MA_OWNED);
5553 
5554 	flowc = start_wrq_wr(cst->eo_txq, ETID_FLUSH_LEN16, &cookie);
5555 	if (__predict_false(flowc == NULL))
5556 		CXGBE_UNIMPLEMENTED(__func__);
5557 
5558 	bzero(flowc, ETID_FLUSH_LEN16 * 16);
5559 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
5560 	    V_FW_FLOWC_WR_NPARAMS(0) | F_FW_WR_COMPL);
5561 	flowc->flowid_len16 = htobe32(V_FW_WR_LEN16(ETID_FLUSH_LEN16) |
5562 	    V_FW_WR_FLOWID(cst->etid));
5563 
5564 	commit_wrq_wr(cst->eo_txq, flowc, &cookie);
5565 
5566 	cst->flags |= EO_FLUSH_RPL_PENDING;
5567 	MPASS(cst->tx_credits >= ETID_FLUSH_LEN16);
5568 	cst->tx_credits -= ETID_FLUSH_LEN16;
5569 	cst->ncompl++;
5570 }
5571 
5572 static void
5573 write_ethofld_wr(struct cxgbe_snd_tag *cst, struct fw_eth_tx_eo_wr *wr,
5574     struct mbuf *m0, int compl)
5575 {
5576 	struct cpl_tx_pkt_core *cpl;
5577 	uint64_t ctrl1;
5578 	uint32_t ctrl;	/* used in many unrelated places */
5579 	int len16, pktlen, nsegs, immhdrs;
5580 	caddr_t dst;
5581 	uintptr_t p;
5582 	struct ulptx_sgl *usgl;
5583 	struct sglist sg;
5584 	struct sglist_seg segs[38];	/* XXX: find real limit.  XXX: get off the stack */
5585 
5586 	mtx_assert(&cst->lock, MA_OWNED);
5587 	M_ASSERTPKTHDR(m0);
5588 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
5589 	    m0->m_pkthdr.l4hlen > 0,
5590 	    ("%s: ethofld mbuf %p is missing header lengths", __func__, m0));
5591 
5592 	if (needs_udp_csum(m0)) {
5593 		CXGBE_UNIMPLEMENTED("UDP ethofld");
5594 	}
5595 
5596 	len16 = mbuf_eo_len16(m0);
5597 	nsegs = mbuf_eo_nsegs(m0);
5598 	pktlen = m0->m_pkthdr.len;
5599 	ctrl = sizeof(struct cpl_tx_pkt_core);
5600 	if (needs_tso(m0))
5601 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
5602 	immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen + m0->m_pkthdr.l4hlen;
5603 	ctrl += immhdrs;
5604 
5605 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_EO_WR) |
5606 	    V_FW_ETH_TX_EO_WR_IMMDLEN(ctrl) | V_FW_WR_COMPL(!!compl));
5607 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(len16) |
5608 	    V_FW_WR_FLOWID(cst->etid));
5609 	wr->r3 = 0;
5610 	wr->u.tcpseg.type = FW_ETH_TX_EO_TYPE_TCPSEG;
5611 	wr->u.tcpseg.ethlen = m0->m_pkthdr.l2hlen;
5612 	wr->u.tcpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
5613 	wr->u.tcpseg.tcplen = m0->m_pkthdr.l4hlen;
5614 	wr->u.tcpseg.tsclk_tsoff = mbuf_eo_tsclk_tsoff(m0);
5615 	wr->u.tcpseg.r4 = 0;
5616 	wr->u.tcpseg.r5 = 0;
5617 	wr->u.tcpseg.plen = htobe32(pktlen - immhdrs);
5618 
5619 	if (needs_tso(m0)) {
5620 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
5621 
5622 		wr->u.tcpseg.mss = htobe16(m0->m_pkthdr.tso_segsz);
5623 
5624 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
5625 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
5626 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
5627 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
5628 			ctrl |= V_LSO_ETHHDR_LEN(1);
5629 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
5630 			ctrl |= F_LSO_IPV6;
5631 		lso->lso_ctrl = htobe32(ctrl);
5632 		lso->ipid_ofst = htobe16(0);
5633 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
5634 		lso->seqno_offset = htobe32(0);
5635 		lso->len = htobe32(pktlen);
5636 
5637 		cpl = (void *)(lso + 1);
5638 	} else {
5639 		wr->u.tcpseg.mss = htobe16(0xffff);
5640 		cpl = (void *)(wr + 1);
5641 	}
5642 
5643 	/* Checksum offload must be requested for ethofld. */
5644 	ctrl1 = 0;
5645 	MPASS(needs_l4_csum(m0));
5646 
5647 	/* VLAN tag insertion */
5648 	if (needs_vlan_insertion(m0)) {
5649 		ctrl1 |= F_TXPKT_VLAN_VLD |
5650 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
5651 	}
5652 
5653 	/* CPL header */
5654 	cpl->ctrl0 = cst->ctrl0;
5655 	cpl->pack = 0;
5656 	cpl->len = htobe16(pktlen);
5657 	cpl->ctrl1 = htobe64(ctrl1);
5658 
5659 	/* Copy Ethernet, IP & TCP hdrs as immediate data */
5660 	p = (uintptr_t)(cpl + 1);
5661 	m_copydata(m0, 0, immhdrs, (void *)p);
5662 
5663 	/* SGL */
5664 	dst = (void *)(cpl + 1);
5665 	if (nsegs > 0) {
5666 		int i, pad;
5667 
5668 		/* zero-pad upto next 16Byte boundary, if not 16Byte aligned */
5669 		p += immhdrs;
5670 		pad = 16 - (immhdrs & 0xf);
5671 		bzero((void *)p, pad);
5672 
5673 		usgl = (void *)(p + pad);
5674 		usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
5675 		    V_ULPTX_NSGE(nsegs));
5676 
5677 		sglist_init(&sg, nitems(segs), segs);
5678 		for (; m0 != NULL; m0 = m0->m_next) {
5679 			if (__predict_false(m0->m_len == 0))
5680 				continue;
5681 			if (immhdrs >= m0->m_len) {
5682 				immhdrs -= m0->m_len;
5683 				continue;
5684 			}
5685 
5686 			sglist_append(&sg, mtod(m0, char *) + immhdrs,
5687 			    m0->m_len - immhdrs);
5688 			immhdrs = 0;
5689 		}
5690 		MPASS(sg.sg_nseg == nsegs);
5691 
5692 		/*
5693 		 * Zero pad last 8B in case the WR doesn't end on a 16B
5694 		 * boundary.
5695 		 */
5696 		*(uint64_t *)((char *)wr + len16 * 16 - 8) = 0;
5697 
5698 		usgl->len0 = htobe32(segs[0].ss_len);
5699 		usgl->addr0 = htobe64(segs[0].ss_paddr);
5700 		for (i = 0; i < nsegs - 1; i++) {
5701 			usgl->sge[i / 2].len[i & 1] = htobe32(segs[i + 1].ss_len);
5702 			usgl->sge[i / 2].addr[i & 1] = htobe64(segs[i + 1].ss_paddr);
5703 		}
5704 		if (i & 1)
5705 			usgl->sge[i / 2].len[1] = htobe32(0);
5706 	}
5707 
5708 }
5709 
5710 static void
5711 ethofld_tx(struct cxgbe_snd_tag *cst)
5712 {
5713 	struct mbuf *m;
5714 	struct wrq_cookie cookie;
5715 	int next_credits, compl;
5716 	struct fw_eth_tx_eo_wr *wr;
5717 
5718 	mtx_assert(&cst->lock, MA_OWNED);
5719 
5720 	while ((m = mbufq_first(&cst->pending_tx)) != NULL) {
5721 		M_ASSERTPKTHDR(m);
5722 
5723 		/* How many len16 credits do we need to send this mbuf. */
5724 		next_credits = mbuf_eo_len16(m);
5725 		MPASS(next_credits > 0);
5726 		if (next_credits > cst->tx_credits) {
5727 			/*
5728 			 * Tx will make progress eventually because there is at
5729 			 * least one outstanding fw4_ack that will return
5730 			 * credits and kick the tx.
5731 			 */
5732 			MPASS(cst->ncompl > 0);
5733 			return;
5734 		}
5735 		wr = start_wrq_wr(cst->eo_txq, next_credits, &cookie);
5736 		if (__predict_false(wr == NULL)) {
5737 			/* XXX: wishful thinking, not a real assertion. */
5738 			MPASS(cst->ncompl > 0);
5739 			return;
5740 		}
5741 		cst->tx_credits -= next_credits;
5742 		cst->tx_nocompl += next_credits;
5743 		compl = cst->ncompl == 0 || cst->tx_nocompl >= cst->tx_total / 2;
5744 		ETHER_BPF_MTAP(cst->com.ifp, m);
5745 		write_ethofld_wr(cst, wr, m, compl);
5746 		commit_wrq_wr(cst->eo_txq, wr, &cookie);
5747 		if (compl) {
5748 			cst->ncompl++;
5749 			cst->tx_nocompl	= 0;
5750 		}
5751 		(void) mbufq_dequeue(&cst->pending_tx);
5752 		mbufq_enqueue(&cst->pending_fwack, m);
5753 	}
5754 }
5755 
5756 int
5757 ethofld_transmit(struct ifnet *ifp, struct mbuf *m0)
5758 {
5759 	struct cxgbe_snd_tag *cst;
5760 	int rc;
5761 
5762 	MPASS(m0->m_nextpkt == NULL);
5763 	MPASS(m0->m_pkthdr.snd_tag != NULL);
5764 	cst = mst_to_cst(m0->m_pkthdr.snd_tag);
5765 
5766 	mtx_lock(&cst->lock);
5767 	MPASS(cst->flags & EO_SND_TAG_REF);
5768 
5769 	if (__predict_false(cst->flags & EO_FLOWC_PENDING)) {
5770 		struct vi_info *vi = ifp->if_softc;
5771 		struct port_info *pi = vi->pi;
5772 		struct adapter *sc = pi->adapter;
5773 		const uint32_t rss_mask = vi->rss_size - 1;
5774 		uint32_t rss_hash;
5775 
5776 		cst->eo_txq = &sc->sge.ofld_txq[vi->first_ofld_txq];
5777 		if (M_HASHTYPE_ISHASH(m0))
5778 			rss_hash = m0->m_pkthdr.flowid;
5779 		else
5780 			rss_hash = arc4random();
5781 		/* We assume RSS hashing */
5782 		cst->iqid = vi->rss[rss_hash & rss_mask];
5783 		cst->eo_txq += rss_hash % vi->nofldtxq;
5784 		rc = send_etid_flowc_wr(cst, pi, vi);
5785 		if (rc != 0)
5786 			goto done;
5787 	}
5788 
5789 	if (__predict_false(cst->plen + m0->m_pkthdr.len > eo_max_backlog)) {
5790 		rc = ENOBUFS;
5791 		goto done;
5792 	}
5793 
5794 	mbufq_enqueue(&cst->pending_tx, m0);
5795 	cst->plen += m0->m_pkthdr.len;
5796 
5797 	ethofld_tx(cst);
5798 	rc = 0;
5799 done:
5800 	mtx_unlock(&cst->lock);
5801 	if (__predict_false(rc != 0))
5802 		m_freem(m0);
5803 	return (rc);
5804 }
5805 
5806 static int
5807 ethofld_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
5808 {
5809 	struct adapter *sc = iq->adapter;
5810 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
5811 	struct mbuf *m;
5812 	u_int etid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
5813 	struct cxgbe_snd_tag *cst;
5814 	uint8_t credits = cpl->credits;
5815 
5816 	cst = lookup_etid(sc, etid);
5817 	mtx_lock(&cst->lock);
5818 	if (__predict_false(cst->flags & EO_FLOWC_RPL_PENDING)) {
5819 		MPASS(credits >= ETID_FLOWC_LEN16);
5820 		credits -= ETID_FLOWC_LEN16;
5821 		cst->flags &= ~EO_FLOWC_RPL_PENDING;
5822 	}
5823 
5824 	KASSERT(cst->ncompl > 0,
5825 	    ("%s: etid %u (%p) wasn't expecting completion.",
5826 	    __func__, etid, cst));
5827 	cst->ncompl--;
5828 
5829 	while (credits > 0) {
5830 		m = mbufq_dequeue(&cst->pending_fwack);
5831 		if (__predict_false(m == NULL)) {
5832 			/*
5833 			 * The remaining credits are for the final flush that
5834 			 * was issued when the tag was freed by the kernel.
5835 			 */
5836 			MPASS((cst->flags &
5837 			    (EO_FLUSH_RPL_PENDING | EO_SND_TAG_REF)) ==
5838 			    EO_FLUSH_RPL_PENDING);
5839 			MPASS(credits == ETID_FLUSH_LEN16);
5840 			MPASS(cst->tx_credits + cpl->credits == cst->tx_total);
5841 			MPASS(cst->ncompl == 0);
5842 
5843 			cst->flags &= ~EO_FLUSH_RPL_PENDING;
5844 			cst->tx_credits += cpl->credits;
5845 freetag:
5846 			cxgbe_snd_tag_free_locked(cst);
5847 			return (0);	/* cst is gone. */
5848 		}
5849 		KASSERT(m != NULL,
5850 		    ("%s: too many credits (%u, %u)", __func__, cpl->credits,
5851 		    credits));
5852 		KASSERT(credits >= mbuf_eo_len16(m),
5853 		    ("%s: too few credits (%u, %u, %u)", __func__,
5854 		    cpl->credits, credits, mbuf_eo_len16(m)));
5855 		credits -= mbuf_eo_len16(m);
5856 		cst->plen -= m->m_pkthdr.len;
5857 		m_freem(m);
5858 	}
5859 
5860 	cst->tx_credits += cpl->credits;
5861 	MPASS(cst->tx_credits <= cst->tx_total);
5862 
5863 	m = mbufq_first(&cst->pending_tx);
5864 	if (m != NULL && cst->tx_credits >= mbuf_eo_len16(m))
5865 		ethofld_tx(cst);
5866 
5867 	if (__predict_false((cst->flags & EO_SND_TAG_REF) == 0) &&
5868 	    cst->ncompl == 0) {
5869 		if (cst->tx_credits == cst->tx_total)
5870 			goto freetag;
5871 		else {
5872 			MPASS((cst->flags & EO_FLUSH_RPL_PENDING) == 0);
5873 			send_etid_flush_wr(cst);
5874 		}
5875 	}
5876 
5877 	mtx_unlock(&cst->lock);
5878 
5879 	return (0);
5880 }
5881 #endif
5882