xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision 752d135e0dacd9a463d24ffb89779b67ce0a7ea0)
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 /* Internal mbuf flags stored in PH_loc.eight[1]. */
86 #define	MC_RAW_WR		0x02
87 
88 /*
89  * Ethernet frames are DMA'd at this byte offset into the freelist buffer.
90  * 0-7 are valid values.
91  */
92 static int fl_pktshift = 0;
93 TUNABLE_INT("hw.cxgbe.fl_pktshift", &fl_pktshift);
94 
95 /*
96  * Pad ethernet payload up to this boundary.
97  * -1: driver should figure out a good value.
98  *  0: disable padding.
99  *  Any power of 2 from 32 to 4096 (both inclusive) is also a valid value.
100  */
101 int fl_pad = -1;
102 TUNABLE_INT("hw.cxgbe.fl_pad", &fl_pad);
103 
104 /*
105  * Status page length.
106  * -1: driver should figure out a good value.
107  *  64 or 128 are the only other valid values.
108  */
109 static int spg_len = -1;
110 TUNABLE_INT("hw.cxgbe.spg_len", &spg_len);
111 
112 /*
113  * Congestion drops.
114  * -1: no congestion feedback (not recommended).
115  *  0: backpressure the channel instead of dropping packets right away.
116  *  1: no backpressure, drop packets for the congested queue immediately.
117  */
118 static int cong_drop = 0;
119 TUNABLE_INT("hw.cxgbe.cong_drop", &cong_drop);
120 
121 /*
122  * Deliver multiple frames in the same free list buffer if they fit.
123  * -1: let the driver decide whether to enable buffer packing or not.
124  *  0: disable buffer packing.
125  *  1: enable buffer packing.
126  */
127 static int buffer_packing = -1;
128 TUNABLE_INT("hw.cxgbe.buffer_packing", &buffer_packing);
129 
130 /*
131  * Start next frame in a packed buffer at this boundary.
132  * -1: driver should figure out a good value.
133  * T4: driver will ignore this and use the same value as fl_pad above.
134  * T5: 16, or a power of 2 from 64 to 4096 (both inclusive) is a valid value.
135  */
136 static int fl_pack = -1;
137 TUNABLE_INT("hw.cxgbe.fl_pack", &fl_pack);
138 
139 /*
140  * Allow the driver to create mbuf(s) in a cluster allocated for rx.
141  * 0: never; always allocate mbufs from the zone_mbuf UMA zone.
142  * 1: ok to create mbuf(s) within a cluster if there is room.
143  */
144 static int allow_mbufs_in_cluster = 1;
145 TUNABLE_INT("hw.cxgbe.allow_mbufs_in_cluster", &allow_mbufs_in_cluster);
146 
147 /*
148  * Largest rx cluster size that the driver is allowed to allocate.
149  */
150 static int largest_rx_cluster = MJUM16BYTES;
151 TUNABLE_INT("hw.cxgbe.largest_rx_cluster", &largest_rx_cluster);
152 
153 /*
154  * Size of cluster allocation that's most likely to succeed.  The driver will
155  * fall back to this size if it fails to allocate clusters larger than this.
156  */
157 static int safest_rx_cluster = PAGE_SIZE;
158 TUNABLE_INT("hw.cxgbe.safest_rx_cluster", &safest_rx_cluster);
159 
160 #ifdef RATELIMIT
161 /*
162  * Knob to control TCP timestamp rewriting, and the granularity of the tick used
163  * for rewriting.  -1 and 0-3 are all valid values.
164  * -1: hardware should leave the TCP timestamps alone.
165  * 0: 1ms
166  * 1: 100us
167  * 2: 10us
168  * 3: 1us
169  */
170 static int tsclk = -1;
171 TUNABLE_INT("hw.cxgbe.tsclk", &tsclk);
172 
173 static int eo_max_backlog = 1024 * 1024;
174 TUNABLE_INT("hw.cxgbe.eo_max_backlog", &eo_max_backlog);
175 #endif
176 
177 /*
178  * The interrupt holdoff timers are multiplied by this value on T6+.
179  * 1 and 3-17 (both inclusive) are legal values.
180  */
181 static int tscale = 1;
182 TUNABLE_INT("hw.cxgbe.tscale", &tscale);
183 
184 /*
185  * Number of LRO entries in the lro_ctrl structure per rx queue.
186  */
187 static int lro_entries = TCP_LRO_ENTRIES;
188 TUNABLE_INT("hw.cxgbe.lro_entries", &lro_entries);
189 
190 /*
191  * This enables presorting of frames before they're fed into tcp_lro_rx.
192  */
193 static int lro_mbufs = 0;
194 TUNABLE_INT("hw.cxgbe.lro_mbufs", &lro_mbufs);
195 
196 struct txpkts {
197 	u_int wr_type;		/* type 0 or type 1 */
198 	u_int npkt;		/* # of packets in this work request */
199 	u_int plen;		/* total payload (sum of all packets) */
200 	u_int len16;		/* # of 16B pieces used by this work request */
201 };
202 
203 /* A packet's SGL.  This + m_pkthdr has all info needed for tx */
204 struct sgl {
205 	struct sglist sg;
206 	struct sglist_seg seg[TX_SGL_SEGS];
207 };
208 
209 static int service_iq(struct sge_iq *, int);
210 static int service_iq_fl(struct sge_iq *, int);
211 static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t);
212 static int t4_eth_rx(struct sge_iq *, const struct rss_header *, struct mbuf *);
213 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int);
214 static inline void init_fl(struct adapter *, struct sge_fl *, int, int, char *);
215 static inline void init_eq(struct adapter *, struct sge_eq *, int, int, uint8_t,
216     uint16_t, char *);
217 static int alloc_ring(struct adapter *, size_t, bus_dma_tag_t *, bus_dmamap_t *,
218     bus_addr_t *, void **);
219 static int free_ring(struct adapter *, bus_dma_tag_t, bus_dmamap_t, bus_addr_t,
220     void *);
221 static int alloc_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *,
222     int, int);
223 static int free_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *);
224 static void add_iq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
225     struct sge_iq *);
226 static void add_fl_sysctls(struct adapter *, struct sysctl_ctx_list *,
227     struct sysctl_oid *, struct sge_fl *);
228 static int alloc_fwq(struct adapter *);
229 static int free_fwq(struct adapter *);
230 static int alloc_ctrlq(struct adapter *, struct sge_wrq *, int,
231     struct sysctl_oid *);
232 static int alloc_rxq(struct vi_info *, struct sge_rxq *, int, int,
233     struct sysctl_oid *);
234 static int free_rxq(struct vi_info *, struct sge_rxq *);
235 #ifdef TCP_OFFLOAD
236 static int alloc_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *, int, int,
237     struct sysctl_oid *);
238 static int free_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *);
239 #endif
240 #ifdef DEV_NETMAP
241 static int alloc_nm_rxq(struct vi_info *, struct sge_nm_rxq *, int, int,
242     struct sysctl_oid *);
243 static int free_nm_rxq(struct vi_info *, struct sge_nm_rxq *);
244 static int alloc_nm_txq(struct vi_info *, struct sge_nm_txq *, int, int,
245     struct sysctl_oid *);
246 static int free_nm_txq(struct vi_info *, struct sge_nm_txq *);
247 #endif
248 static int ctrl_eq_alloc(struct adapter *, struct sge_eq *);
249 static int eth_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
250 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
251 static int ofld_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
252 #endif
253 static int alloc_eq(struct adapter *, struct vi_info *, struct sge_eq *);
254 static int free_eq(struct adapter *, struct sge_eq *);
255 static int alloc_wrq(struct adapter *, struct vi_info *, struct sge_wrq *,
256     struct sysctl_oid *);
257 static int free_wrq(struct adapter *, struct sge_wrq *);
258 static int alloc_txq(struct vi_info *, struct sge_txq *, int,
259     struct sysctl_oid *);
260 static int free_txq(struct vi_info *, struct sge_txq *);
261 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int);
262 static inline void ring_fl_db(struct adapter *, struct sge_fl *);
263 static int refill_fl(struct adapter *, struct sge_fl *, int);
264 static void refill_sfl(void *);
265 static int alloc_fl_sdesc(struct sge_fl *);
266 static void free_fl_sdesc(struct adapter *, struct sge_fl *);
267 static void find_best_refill_source(struct adapter *, struct sge_fl *, int);
268 static void find_safe_refill_source(struct adapter *, struct sge_fl *);
269 static void add_fl_to_sfl(struct adapter *, struct sge_fl *);
270 
271 static inline void get_pkt_gl(struct mbuf *, struct sglist *);
272 static inline u_int txpkt_len16(u_int, u_int);
273 static inline u_int txpkt_vm_len16(u_int, u_int);
274 static inline u_int txpkts0_len16(u_int);
275 static inline u_int txpkts1_len16(void);
276 static u_int write_raw_wr(struct sge_txq *, void *, struct mbuf *, u_int);
277 static u_int write_txpkt_wr(struct sge_txq *, struct fw_eth_tx_pkt_wr *,
278     struct mbuf *, u_int);
279 static u_int write_txpkt_vm_wr(struct adapter *, struct sge_txq *,
280     struct fw_eth_tx_pkt_vm_wr *, struct mbuf *, u_int);
281 static int try_txpkts(struct mbuf *, struct mbuf *, struct txpkts *, u_int);
282 static int add_to_txpkts(struct mbuf *, struct txpkts *, u_int);
283 static u_int write_txpkts_wr(struct sge_txq *, struct fw_eth_tx_pkts_wr *,
284     struct mbuf *, const struct txpkts *, u_int);
285 static void write_gl_to_txd(struct sge_txq *, struct mbuf *, caddr_t *, int);
286 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int);
287 static inline void ring_eq_db(struct adapter *, struct sge_eq *, u_int);
288 static inline uint16_t read_hw_cidx(struct sge_eq *);
289 static inline u_int reclaimable_tx_desc(struct sge_eq *);
290 static inline u_int total_available_tx_desc(struct sge_eq *);
291 static u_int reclaim_tx_descs(struct sge_txq *, u_int);
292 static void tx_reclaim(void *, int);
293 static __be64 get_flit(struct sglist_seg *, int, int);
294 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
295     struct mbuf *);
296 static int handle_fw_msg(struct sge_iq *, const struct rss_header *,
297     struct mbuf *);
298 static int t4_handle_wrerr_rpl(struct adapter *, const __be64 *);
299 static void wrq_tx_drain(void *, int);
300 static void drain_wrq_wr_list(struct adapter *, struct sge_wrq *);
301 
302 static int sysctl_uint16(SYSCTL_HANDLER_ARGS);
303 static int sysctl_bufsizes(SYSCTL_HANDLER_ARGS);
304 #ifdef RATELIMIT
305 static inline u_int txpkt_eo_len16(u_int, u_int, u_int);
306 static int ethofld_fw4_ack(struct sge_iq *, const struct rss_header *,
307     struct mbuf *);
308 #endif
309 
310 static counter_u64_t extfree_refs;
311 static counter_u64_t extfree_rels;
312 
313 an_handler_t t4_an_handler;
314 fw_msg_handler_t t4_fw_msg_handler[NUM_FW6_TYPES];
315 cpl_handler_t t4_cpl_handler[NUM_CPL_CMDS];
316 cpl_handler_t set_tcb_rpl_handlers[NUM_CPL_COOKIES];
317 cpl_handler_t l2t_write_rpl_handlers[NUM_CPL_COOKIES];
318 cpl_handler_t act_open_rpl_handlers[NUM_CPL_COOKIES];
319 cpl_handler_t abort_rpl_rss_handlers[NUM_CPL_COOKIES];
320 cpl_handler_t fw4_ack_handlers[NUM_CPL_COOKIES];
321 
322 void
323 t4_register_an_handler(an_handler_t h)
324 {
325 	uintptr_t *loc;
326 
327 	MPASS(h == NULL || t4_an_handler == NULL);
328 
329 	loc = (uintptr_t *)&t4_an_handler;
330 	atomic_store_rel_ptr(loc, (uintptr_t)h);
331 }
332 
333 void
334 t4_register_fw_msg_handler(int type, fw_msg_handler_t h)
335 {
336 	uintptr_t *loc;
337 
338 	MPASS(type < nitems(t4_fw_msg_handler));
339 	MPASS(h == NULL || t4_fw_msg_handler[type] == NULL);
340 	/*
341 	 * These are dispatched by the handler for FW{4|6}_CPL_MSG using the CPL
342 	 * handler dispatch table.  Reject any attempt to install a handler for
343 	 * this subtype.
344 	 */
345 	MPASS(type != FW_TYPE_RSSCPL);
346 	MPASS(type != FW6_TYPE_RSSCPL);
347 
348 	loc = (uintptr_t *)&t4_fw_msg_handler[type];
349 	atomic_store_rel_ptr(loc, (uintptr_t)h);
350 }
351 
352 void
353 t4_register_cpl_handler(int opcode, cpl_handler_t h)
354 {
355 	uintptr_t *loc;
356 
357 	MPASS(opcode < nitems(t4_cpl_handler));
358 	MPASS(h == NULL || t4_cpl_handler[opcode] == NULL);
359 
360 	loc = (uintptr_t *)&t4_cpl_handler[opcode];
361 	atomic_store_rel_ptr(loc, (uintptr_t)h);
362 }
363 
364 static int
365 set_tcb_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
366     struct mbuf *m)
367 {
368 	const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
369 	u_int tid;
370 	int cookie;
371 
372 	MPASS(m == NULL);
373 
374 	tid = GET_TID(cpl);
375 	if (is_hpftid(iq->adapter, tid) || is_ftid(iq->adapter, tid)) {
376 		/*
377 		 * The return code for filter-write is put in the CPL cookie so
378 		 * we have to rely on the hardware tid (is_ftid) to determine
379 		 * that this is a response to a filter.
380 		 */
381 		cookie = CPL_COOKIE_FILTER;
382 	} else {
383 		cookie = G_COOKIE(cpl->cookie);
384 	}
385 	MPASS(cookie > CPL_COOKIE_RESERVED);
386 	MPASS(cookie < nitems(set_tcb_rpl_handlers));
387 
388 	return (set_tcb_rpl_handlers[cookie](iq, rss, m));
389 }
390 
391 static int
392 l2t_write_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
393     struct mbuf *m)
394 {
395 	const struct cpl_l2t_write_rpl *rpl = (const void *)(rss + 1);
396 	unsigned int cookie;
397 
398 	MPASS(m == NULL);
399 
400 	cookie = GET_TID(rpl) & F_SYNC_WR ? CPL_COOKIE_TOM : CPL_COOKIE_FILTER;
401 	return (l2t_write_rpl_handlers[cookie](iq, rss, m));
402 }
403 
404 static int
405 act_open_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
406     struct mbuf *m)
407 {
408 	const struct cpl_act_open_rpl *cpl = (const void *)(rss + 1);
409 	u_int cookie = G_TID_COOKIE(G_AOPEN_ATID(be32toh(cpl->atid_status)));
410 
411 	MPASS(m == NULL);
412 	MPASS(cookie != CPL_COOKIE_RESERVED);
413 
414 	return (act_open_rpl_handlers[cookie](iq, rss, m));
415 }
416 
417 static int
418 abort_rpl_rss_handler(struct sge_iq *iq, const struct rss_header *rss,
419     struct mbuf *m)
420 {
421 	struct adapter *sc = iq->adapter;
422 	u_int cookie;
423 
424 	MPASS(m == NULL);
425 	if (is_hashfilter(sc))
426 		cookie = CPL_COOKIE_HASHFILTER;
427 	else
428 		cookie = CPL_COOKIE_TOM;
429 
430 	return (abort_rpl_rss_handlers[cookie](iq, rss, m));
431 }
432 
433 static int
434 fw4_ack_handler(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
435 {
436 	struct adapter *sc = iq->adapter;
437 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
438 	unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
439 	u_int cookie;
440 
441 	MPASS(m == NULL);
442 	if (is_etid(sc, tid))
443 		cookie = CPL_COOKIE_ETHOFLD;
444 	else
445 		cookie = CPL_COOKIE_TOM;
446 
447 	return (fw4_ack_handlers[cookie](iq, rss, m));
448 }
449 
450 static void
451 t4_init_shared_cpl_handlers(void)
452 {
453 
454 	t4_register_cpl_handler(CPL_SET_TCB_RPL, set_tcb_rpl_handler);
455 	t4_register_cpl_handler(CPL_L2T_WRITE_RPL, l2t_write_rpl_handler);
456 	t4_register_cpl_handler(CPL_ACT_OPEN_RPL, act_open_rpl_handler);
457 	t4_register_cpl_handler(CPL_ABORT_RPL_RSS, abort_rpl_rss_handler);
458 	t4_register_cpl_handler(CPL_FW4_ACK, fw4_ack_handler);
459 }
460 
461 void
462 t4_register_shared_cpl_handler(int opcode, cpl_handler_t h, int cookie)
463 {
464 	uintptr_t *loc;
465 
466 	MPASS(opcode < nitems(t4_cpl_handler));
467 	MPASS(cookie > CPL_COOKIE_RESERVED);
468 	MPASS(cookie < NUM_CPL_COOKIES);
469 	MPASS(t4_cpl_handler[opcode] != NULL);
470 
471 	switch (opcode) {
472 	case CPL_SET_TCB_RPL:
473 		loc = (uintptr_t *)&set_tcb_rpl_handlers[cookie];
474 		break;
475 	case CPL_L2T_WRITE_RPL:
476 		loc = (uintptr_t *)&l2t_write_rpl_handlers[cookie];
477 		break;
478 	case CPL_ACT_OPEN_RPL:
479 		loc = (uintptr_t *)&act_open_rpl_handlers[cookie];
480 		break;
481 	case CPL_ABORT_RPL_RSS:
482 		loc = (uintptr_t *)&abort_rpl_rss_handlers[cookie];
483 		break;
484 	case CPL_FW4_ACK:
485 		loc = (uintptr_t *)&fw4_ack_handlers[cookie];
486 		break;
487 	default:
488 		MPASS(0);
489 		return;
490 	}
491 	MPASS(h == NULL || *loc == (uintptr_t)NULL);
492 	atomic_store_rel_ptr(loc, (uintptr_t)h);
493 }
494 
495 /*
496  * Called on MOD_LOAD.  Validates and calculates the SGE tunables.
497  */
498 void
499 t4_sge_modload(void)
500 {
501 
502 	if (fl_pktshift < 0 || fl_pktshift > 7) {
503 		printf("Invalid hw.cxgbe.fl_pktshift value (%d),"
504 		    " using 0 instead.\n", fl_pktshift);
505 		fl_pktshift = 0;
506 	}
507 
508 	if (spg_len != 64 && spg_len != 128) {
509 		int len;
510 
511 #if defined(__i386__) || defined(__amd64__)
512 		len = cpu_clflush_line_size > 64 ? 128 : 64;
513 #else
514 		len = 64;
515 #endif
516 		if (spg_len != -1) {
517 			printf("Invalid hw.cxgbe.spg_len value (%d),"
518 			    " using %d instead.\n", spg_len, len);
519 		}
520 		spg_len = len;
521 	}
522 
523 	if (cong_drop < -1 || cong_drop > 1) {
524 		printf("Invalid hw.cxgbe.cong_drop value (%d),"
525 		    " using 0 instead.\n", cong_drop);
526 		cong_drop = 0;
527 	}
528 
529 	if (tscale != 1 && (tscale < 3 || tscale > 17)) {
530 		printf("Invalid hw.cxgbe.tscale value (%d),"
531 		    " using 1 instead.\n", tscale);
532 		tscale = 1;
533 	}
534 
535 	extfree_refs = counter_u64_alloc(M_WAITOK);
536 	extfree_rels = counter_u64_alloc(M_WAITOK);
537 	counter_u64_zero(extfree_refs);
538 	counter_u64_zero(extfree_rels);
539 
540 	t4_init_shared_cpl_handlers();
541 	t4_register_cpl_handler(CPL_FW4_MSG, handle_fw_msg);
542 	t4_register_cpl_handler(CPL_FW6_MSG, handle_fw_msg);
543 	t4_register_cpl_handler(CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
544 	t4_register_cpl_handler(CPL_RX_PKT, t4_eth_rx);
545 #ifdef RATELIMIT
546 	t4_register_shared_cpl_handler(CPL_FW4_ACK, ethofld_fw4_ack,
547 	    CPL_COOKIE_ETHOFLD);
548 #endif
549 	t4_register_fw_msg_handler(FW6_TYPE_CMD_RPL, t4_handle_fw_rpl);
550 	t4_register_fw_msg_handler(FW6_TYPE_WRERR_RPL, t4_handle_wrerr_rpl);
551 }
552 
553 void
554 t4_sge_modunload(void)
555 {
556 
557 	counter_u64_free(extfree_refs);
558 	counter_u64_free(extfree_rels);
559 }
560 
561 uint64_t
562 t4_sge_extfree_refs(void)
563 {
564 	uint64_t refs, rels;
565 
566 	rels = counter_u64_fetch(extfree_rels);
567 	refs = counter_u64_fetch(extfree_refs);
568 
569 	return (refs - rels);
570 }
571 
572 static inline void
573 setup_pad_and_pack_boundaries(struct adapter *sc)
574 {
575 	uint32_t v, m;
576 	int pad, pack, pad_shift;
577 
578 	pad_shift = chip_id(sc) > CHELSIO_T5 ? X_T6_INGPADBOUNDARY_SHIFT :
579 	    X_INGPADBOUNDARY_SHIFT;
580 	pad = fl_pad;
581 	if (fl_pad < (1 << pad_shift) ||
582 	    fl_pad > (1 << (pad_shift + M_INGPADBOUNDARY)) ||
583 	    !powerof2(fl_pad)) {
584 		/*
585 		 * If there is any chance that we might use buffer packing and
586 		 * the chip is a T4, then pick 64 as the pad/pack boundary.  Set
587 		 * it to the minimum allowed in all other cases.
588 		 */
589 		pad = is_t4(sc) && buffer_packing ? 64 : 1 << pad_shift;
590 
591 		/*
592 		 * For fl_pad = 0 we'll still write a reasonable value to the
593 		 * register but all the freelists will opt out of padding.
594 		 * We'll complain here only if the user tried to set it to a
595 		 * value greater than 0 that was invalid.
596 		 */
597 		if (fl_pad > 0) {
598 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pad value"
599 			    " (%d), using %d instead.\n", fl_pad, pad);
600 		}
601 	}
602 	m = V_INGPADBOUNDARY(M_INGPADBOUNDARY);
603 	v = V_INGPADBOUNDARY(ilog2(pad) - pad_shift);
604 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
605 
606 	if (is_t4(sc)) {
607 		if (fl_pack != -1 && fl_pack != pad) {
608 			/* Complain but carry on. */
609 			device_printf(sc->dev, "hw.cxgbe.fl_pack (%d) ignored,"
610 			    " using %d instead.\n", fl_pack, pad);
611 		}
612 		return;
613 	}
614 
615 	pack = fl_pack;
616 	if (fl_pack < 16 || fl_pack == 32 || fl_pack > 4096 ||
617 	    !powerof2(fl_pack)) {
618 		pack = max(sc->params.pci.mps, CACHE_LINE_SIZE);
619 		MPASS(powerof2(pack));
620 		if (pack < 16)
621 			pack = 16;
622 		if (pack == 32)
623 			pack = 64;
624 		if (pack > 4096)
625 			pack = 4096;
626 		if (fl_pack != -1) {
627 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pack value"
628 			    " (%d), using %d instead.\n", fl_pack, pack);
629 		}
630 	}
631 	m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY);
632 	if (pack == 16)
633 		v = V_INGPACKBOUNDARY(0);
634 	else
635 		v = V_INGPACKBOUNDARY(ilog2(pack) - 5);
636 
637 	MPASS(!is_t4(sc));	/* T4 doesn't have SGE_CONTROL2 */
638 	t4_set_reg_field(sc, A_SGE_CONTROL2, m, v);
639 }
640 
641 /*
642  * adap->params.vpd.cclk must be set up before this is called.
643  */
644 void
645 t4_tweak_chip_settings(struct adapter *sc)
646 {
647 	int i;
648 	uint32_t v, m;
649 	int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200};
650 	int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk;
651 	int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */
652 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
653 	static int sge_flbuf_sizes[] = {
654 		MCLBYTES,
655 #if MJUMPAGESIZE != MCLBYTES
656 		MJUMPAGESIZE,
657 		MJUMPAGESIZE - CL_METADATA_SIZE,
658 		MJUMPAGESIZE - 2 * MSIZE - CL_METADATA_SIZE,
659 #endif
660 		MJUM9BYTES,
661 		MJUM16BYTES,
662 		MCLBYTES - MSIZE - CL_METADATA_SIZE,
663 		MJUM9BYTES - CL_METADATA_SIZE,
664 		MJUM16BYTES - CL_METADATA_SIZE,
665 	};
666 
667 	KASSERT(sc->flags & MASTER_PF,
668 	    ("%s: trying to change chip settings when not master.", __func__));
669 
670 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
671 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
672 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
673 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
674 
675 	setup_pad_and_pack_boundaries(sc);
676 
677 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
678 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
679 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
680 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
681 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
682 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
683 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
684 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
685 	t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v);
686 
687 	KASSERT(nitems(sge_flbuf_sizes) <= SGE_FLBUF_SIZES,
688 	    ("%s: hw buffer size table too big", __func__));
689 	for (i = 0; i < min(nitems(sge_flbuf_sizes), SGE_FLBUF_SIZES); i++) {
690 		t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i),
691 		    sge_flbuf_sizes[i]);
692 	}
693 
694 	v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) |
695 	    V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]);
696 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v);
697 
698 	KASSERT(intr_timer[0] <= timer_max,
699 	    ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0],
700 	    timer_max));
701 	for (i = 1; i < nitems(intr_timer); i++) {
702 		KASSERT(intr_timer[i] >= intr_timer[i - 1],
703 		    ("%s: timers not listed in increasing order (%d)",
704 		    __func__, i));
705 
706 		while (intr_timer[i] > timer_max) {
707 			if (i == nitems(intr_timer) - 1) {
708 				intr_timer[i] = timer_max;
709 				break;
710 			}
711 			intr_timer[i] += intr_timer[i - 1];
712 			intr_timer[i] /= 2;
713 		}
714 	}
715 
716 	v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) |
717 	    V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1]));
718 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v);
719 	v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) |
720 	    V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3]));
721 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v);
722 	v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) |
723 	    V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5]));
724 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v);
725 
726 	if (chip_id(sc) >= CHELSIO_T6) {
727 		m = V_TSCALE(M_TSCALE);
728 		if (tscale == 1)
729 			v = 0;
730 		else
731 			v = V_TSCALE(tscale - 2);
732 		t4_set_reg_field(sc, A_SGE_ITP_CONTROL, m, v);
733 
734 		if (sc->debug_flags & DF_DISABLE_TCB_CACHE) {
735 			m = V_RDTHRESHOLD(M_RDTHRESHOLD) | F_WRTHRTHRESHEN |
736 			    V_WRTHRTHRESH(M_WRTHRTHRESH);
737 			t4_tp_pio_read(sc, &v, 1, A_TP_CMM_CONFIG, 1);
738 			v &= ~m;
739 			v |= V_RDTHRESHOLD(1) | F_WRTHRTHRESHEN |
740 			    V_WRTHRTHRESH(16);
741 			t4_tp_pio_write(sc, &v, 1, A_TP_CMM_CONFIG, 1);
742 		}
743 	}
744 
745 	/* 4K, 16K, 64K, 256K DDP "page sizes" for TDDP */
746 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
747 	t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v);
748 
749 	/*
750 	 * 4K, 8K, 16K, 64K DDP "page sizes" for iSCSI DDP.  These have been
751 	 * chosen with MAXPHYS = 128K in mind.  The largest DDP buffer that we
752 	 * may have to deal with is MAXPHYS + 1 page.
753 	 */
754 	v = V_HPZ0(0) | V_HPZ1(1) | V_HPZ2(2) | V_HPZ3(4);
755 	t4_write_reg(sc, A_ULP_RX_ISCSI_PSZ, v);
756 
757 	/* We use multiple DDP page sizes both in plain-TOE and ISCSI modes. */
758 	m = v = F_TDDPTAGTCB | F_ISCSITAGTCB;
759 	t4_set_reg_field(sc, A_ULP_RX_CTL, m, v);
760 
761 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
762 	    F_RESETDDPOFFSET;
763 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
764 	t4_set_reg_field(sc, A_TP_PARA_REG5, m, v);
765 }
766 
767 /*
768  * SGE wants the buffer to be at least 64B and then a multiple of 16.  If
769  * padding is in use, the buffer's start and end need to be aligned to the pad
770  * boundary as well.  We'll just make sure that the size is a multiple of the
771  * boundary here, it is up to the buffer allocation code to make sure the start
772  * of the buffer is aligned as well.
773  */
774 static inline int
775 hwsz_ok(struct adapter *sc, int hwsz)
776 {
777 	int mask = fl_pad ? sc->params.sge.pad_boundary - 1 : 16 - 1;
778 
779 	return (hwsz >= 64 && (hwsz & mask) == 0);
780 }
781 
782 /*
783  * XXX: driver really should be able to deal with unexpected settings.
784  */
785 int
786 t4_read_chip_settings(struct adapter *sc)
787 {
788 	struct sge *s = &sc->sge;
789 	struct sge_params *sp = &sc->params.sge;
790 	int i, j, n, rc = 0;
791 	uint32_t m, v, r;
792 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
793 	static int sw_buf_sizes[] = {	/* Sorted by size */
794 		MCLBYTES,
795 #if MJUMPAGESIZE != MCLBYTES
796 		MJUMPAGESIZE,
797 #endif
798 		MJUM9BYTES,
799 		MJUM16BYTES
800 	};
801 	struct sw_zone_info *swz, *safe_swz;
802 	struct hw_buf_info *hwb;
803 
804 	m = F_RXPKTCPLMODE;
805 	v = F_RXPKTCPLMODE;
806 	r = sc->params.sge.sge_control;
807 	if ((r & m) != v) {
808 		device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r);
809 		rc = EINVAL;
810 	}
811 
812 	/*
813 	 * If this changes then every single use of PAGE_SHIFT in the driver
814 	 * needs to be carefully reviewed for PAGE_SHIFT vs sp->page_shift.
815 	 */
816 	if (sp->page_shift != PAGE_SHIFT) {
817 		device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r);
818 		rc = EINVAL;
819 	}
820 
821 	/* Filter out unusable hw buffer sizes entirely (mark with -2). */
822 	hwb = &s->hw_buf_info[0];
823 	for (i = 0; i < nitems(s->hw_buf_info); i++, hwb++) {
824 		r = sc->params.sge.sge_fl_buffer_size[i];
825 		hwb->size = r;
826 		hwb->zidx = hwsz_ok(sc, r) ? -1 : -2;
827 		hwb->next = -1;
828 	}
829 
830 	/*
831 	 * Create a sorted list in decreasing order of hw buffer sizes (and so
832 	 * increasing order of spare area) for each software zone.
833 	 *
834 	 * If padding is enabled then the start and end of the buffer must align
835 	 * to the pad boundary; if packing is enabled then they must align with
836 	 * the pack boundary as well.  Allocations from the cluster zones are
837 	 * aligned to min(size, 4K), so the buffer starts at that alignment and
838 	 * ends at hwb->size alignment.  If mbuf inlining is allowed the
839 	 * starting alignment will be reduced to MSIZE and the driver will
840 	 * exercise appropriate caution when deciding on the best buffer layout
841 	 * to use.
842 	 */
843 	n = 0;	/* no usable buffer size to begin with */
844 	swz = &s->sw_zone_info[0];
845 	safe_swz = NULL;
846 	for (i = 0; i < SW_ZONE_SIZES; i++, swz++) {
847 		int8_t head = -1, tail = -1;
848 
849 		swz->size = sw_buf_sizes[i];
850 		swz->zone = m_getzone(swz->size);
851 		swz->type = m_gettype(swz->size);
852 
853 		if (swz->size < PAGE_SIZE) {
854 			MPASS(powerof2(swz->size));
855 			if (fl_pad && (swz->size % sp->pad_boundary != 0))
856 				continue;
857 		}
858 
859 		if (swz->size == safest_rx_cluster)
860 			safe_swz = swz;
861 
862 		hwb = &s->hw_buf_info[0];
863 		for (j = 0; j < SGE_FLBUF_SIZES; j++, hwb++) {
864 			if (hwb->zidx != -1 || hwb->size > swz->size)
865 				continue;
866 #ifdef INVARIANTS
867 			if (fl_pad)
868 				MPASS(hwb->size % sp->pad_boundary == 0);
869 #endif
870 			hwb->zidx = i;
871 			if (head == -1)
872 				head = tail = j;
873 			else if (hwb->size < s->hw_buf_info[tail].size) {
874 				s->hw_buf_info[tail].next = j;
875 				tail = j;
876 			} else {
877 				int8_t *cur;
878 				struct hw_buf_info *t;
879 
880 				for (cur = &head; *cur != -1; cur = &t->next) {
881 					t = &s->hw_buf_info[*cur];
882 					if (hwb->size == t->size) {
883 						hwb->zidx = -2;
884 						break;
885 					}
886 					if (hwb->size > t->size) {
887 						hwb->next = *cur;
888 						*cur = j;
889 						break;
890 					}
891 				}
892 			}
893 		}
894 		swz->head_hwidx = head;
895 		swz->tail_hwidx = tail;
896 
897 		if (tail != -1) {
898 			n++;
899 			if (swz->size - s->hw_buf_info[tail].size >=
900 			    CL_METADATA_SIZE)
901 				sc->flags |= BUF_PACKING_OK;
902 		}
903 	}
904 	if (n == 0) {
905 		device_printf(sc->dev, "no usable SGE FL buffer size.\n");
906 		rc = EINVAL;
907 	}
908 
909 	s->safe_hwidx1 = -1;
910 	s->safe_hwidx2 = -1;
911 	if (safe_swz != NULL) {
912 		s->safe_hwidx1 = safe_swz->head_hwidx;
913 		for (i = safe_swz->head_hwidx; i != -1; i = hwb->next) {
914 			int spare;
915 
916 			hwb = &s->hw_buf_info[i];
917 #ifdef INVARIANTS
918 			if (fl_pad)
919 				MPASS(hwb->size % sp->pad_boundary == 0);
920 #endif
921 			spare = safe_swz->size - hwb->size;
922 			if (spare >= CL_METADATA_SIZE) {
923 				s->safe_hwidx2 = i;
924 				break;
925 			}
926 		}
927 	}
928 
929 	if (sc->flags & IS_VF)
930 		return (0);
931 
932 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
933 	r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ);
934 	if (r != v) {
935 		device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r);
936 		rc = EINVAL;
937 	}
938 
939 	m = v = F_TDDPTAGTCB;
940 	r = t4_read_reg(sc, A_ULP_RX_CTL);
941 	if ((r & m) != v) {
942 		device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r);
943 		rc = EINVAL;
944 	}
945 
946 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
947 	    F_RESETDDPOFFSET;
948 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
949 	r = t4_read_reg(sc, A_TP_PARA_REG5);
950 	if ((r & m) != v) {
951 		device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r);
952 		rc = EINVAL;
953 	}
954 
955 	t4_init_tp_params(sc, 1);
956 
957 	t4_read_mtu_tbl(sc, sc->params.mtus, NULL);
958 	t4_load_mtus(sc, sc->params.mtus, sc->params.a_wnd, sc->params.b_wnd);
959 
960 	return (rc);
961 }
962 
963 int
964 t4_create_dma_tag(struct adapter *sc)
965 {
966 	int rc;
967 
968 	rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
969 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
970 	    BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
971 	    NULL, &sc->dmat);
972 	if (rc != 0) {
973 		device_printf(sc->dev,
974 		    "failed to create main DMA tag: %d\n", rc);
975 	}
976 
977 	return (rc);
978 }
979 
980 void
981 t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
982     struct sysctl_oid_list *children)
983 {
984 	struct sge_params *sp = &sc->params.sge;
985 
986 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "buffer_sizes",
987 	    CTLTYPE_STRING | CTLFLAG_RD, &sc->sge, 0, sysctl_bufsizes, "A",
988 	    "freelist buffer sizes");
989 
990 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD,
991 	    NULL, sp->fl_pktshift, "payload DMA offset in rx buffer (bytes)");
992 
993 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD,
994 	    NULL, sp->pad_boundary, "payload pad boundary (bytes)");
995 
996 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD,
997 	    NULL, sp->spg_len, "status page size (bytes)");
998 
999 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD,
1000 	    NULL, cong_drop, "congestion drop setting");
1001 
1002 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD,
1003 	    NULL, sp->pack_boundary, "payload pack boundary (bytes)");
1004 }
1005 
1006 int
1007 t4_destroy_dma_tag(struct adapter *sc)
1008 {
1009 	if (sc->dmat)
1010 		bus_dma_tag_destroy(sc->dmat);
1011 
1012 	return (0);
1013 }
1014 
1015 /*
1016  * Allocate and initialize the firmware event queue, control queues, and special
1017  * purpose rx queues owned by the adapter.
1018  *
1019  * Returns errno on failure.  Resources allocated up to that point may still be
1020  * allocated.  Caller is responsible for cleanup in case this function fails.
1021  */
1022 int
1023 t4_setup_adapter_queues(struct adapter *sc)
1024 {
1025 	struct sysctl_oid *oid;
1026 	struct sysctl_oid_list *children;
1027 	int rc, i;
1028 
1029 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
1030 
1031 	sysctl_ctx_init(&sc->ctx);
1032 	sc->flags |= ADAP_SYSCTL_CTX;
1033 
1034 	/*
1035 	 * Firmware event queue
1036 	 */
1037 	rc = alloc_fwq(sc);
1038 	if (rc != 0)
1039 		return (rc);
1040 
1041 	/*
1042 	 * That's all for the VF driver.
1043 	 */
1044 	if (sc->flags & IS_VF)
1045 		return (rc);
1046 
1047 	oid = device_get_sysctl_tree(sc->dev);
1048 	children = SYSCTL_CHILDREN(oid);
1049 
1050 	/*
1051 	 * XXX: General purpose rx queues, one per port.
1052 	 */
1053 
1054 	/*
1055 	 * Control queues, one per port.
1056 	 */
1057 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "ctrlq",
1058 	    CTLFLAG_RD, NULL, "control queues");
1059 	for_each_port(sc, i) {
1060 		struct sge_wrq *ctrlq = &sc->sge.ctrlq[i];
1061 
1062 		rc = alloc_ctrlq(sc, ctrlq, i, oid);
1063 		if (rc != 0)
1064 			return (rc);
1065 	}
1066 
1067 	return (rc);
1068 }
1069 
1070 /*
1071  * Idempotent
1072  */
1073 int
1074 t4_teardown_adapter_queues(struct adapter *sc)
1075 {
1076 	int i;
1077 
1078 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
1079 
1080 	/* Do this before freeing the queue */
1081 	if (sc->flags & ADAP_SYSCTL_CTX) {
1082 		sysctl_ctx_free(&sc->ctx);
1083 		sc->flags &= ~ADAP_SYSCTL_CTX;
1084 	}
1085 
1086 	if (!(sc->flags & IS_VF)) {
1087 		for_each_port(sc, i)
1088 			free_wrq(sc, &sc->sge.ctrlq[i]);
1089 	}
1090 	free_fwq(sc);
1091 
1092 	return (0);
1093 }
1094 
1095 /* Maximum payload that can be delivered with a single iq descriptor */
1096 static inline int
1097 mtu_to_max_payload(struct adapter *sc, int mtu, const int toe)
1098 {
1099 	int payload;
1100 
1101 #ifdef TCP_OFFLOAD
1102 	if (toe) {
1103 		int rxcs = G_RXCOALESCESIZE(t4_read_reg(sc, A_TP_PARA_REG2));
1104 
1105 		/* Note that COP can set rx_coalesce on/off per connection. */
1106 		payload = max(mtu, rxcs);
1107 	} else {
1108 #endif
1109 		/* large enough even when hw VLAN extraction is disabled */
1110 		payload = sc->params.sge.fl_pktshift + ETHER_HDR_LEN +
1111 		    ETHER_VLAN_ENCAP_LEN + mtu;
1112 #ifdef TCP_OFFLOAD
1113 	}
1114 #endif
1115 
1116 	return (payload);
1117 }
1118 
1119 int
1120 t4_setup_vi_queues(struct vi_info *vi)
1121 {
1122 	int rc = 0, i, intr_idx, iqidx;
1123 	struct sge_rxq *rxq;
1124 	struct sge_txq *txq;
1125 #ifdef TCP_OFFLOAD
1126 	struct sge_ofld_rxq *ofld_rxq;
1127 #endif
1128 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1129 	struct sge_wrq *ofld_txq;
1130 #endif
1131 #ifdef DEV_NETMAP
1132 	int saved_idx;
1133 	struct sge_nm_rxq *nm_rxq;
1134 	struct sge_nm_txq *nm_txq;
1135 #endif
1136 	char name[16];
1137 	struct port_info *pi = vi->pi;
1138 	struct adapter *sc = pi->adapter;
1139 	struct ifnet *ifp = vi->ifp;
1140 	struct sysctl_oid *oid = device_get_sysctl_tree(vi->dev);
1141 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
1142 	int maxp, mtu = ifp->if_mtu;
1143 
1144 	/* Interrupt vector to start from (when using multiple vectors) */
1145 	intr_idx = vi->first_intr;
1146 
1147 #ifdef DEV_NETMAP
1148 	saved_idx = intr_idx;
1149 	if (ifp->if_capabilities & IFCAP_NETMAP) {
1150 
1151 		/* netmap is supported with direct interrupts only. */
1152 		MPASS(!forwarding_intr_to_fwq(sc));
1153 
1154 		/*
1155 		 * We don't have buffers to back the netmap rx queues
1156 		 * right now so we create the queues in a way that
1157 		 * doesn't set off any congestion signal in the chip.
1158 		 */
1159 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_rxq",
1160 		    CTLFLAG_RD, NULL, "rx queues");
1161 		for_each_nm_rxq(vi, i, nm_rxq) {
1162 			rc = alloc_nm_rxq(vi, nm_rxq, intr_idx, i, oid);
1163 			if (rc != 0)
1164 				goto done;
1165 			intr_idx++;
1166 		}
1167 
1168 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_txq",
1169 		    CTLFLAG_RD, NULL, "tx queues");
1170 		for_each_nm_txq(vi, i, nm_txq) {
1171 			iqidx = vi->first_nm_rxq + (i % vi->nnmrxq);
1172 			rc = alloc_nm_txq(vi, nm_txq, iqidx, i, oid);
1173 			if (rc != 0)
1174 				goto done;
1175 		}
1176 	}
1177 
1178 	/* Normal rx queues and netmap rx queues share the same interrupts. */
1179 	intr_idx = saved_idx;
1180 #endif
1181 
1182 	/*
1183 	 * Allocate rx queues first because a default iqid is required when
1184 	 * creating a tx queue.
1185 	 */
1186 	maxp = mtu_to_max_payload(sc, mtu, 0);
1187 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
1188 	    CTLFLAG_RD, NULL, "rx queues");
1189 	for_each_rxq(vi, i, rxq) {
1190 
1191 		init_iq(&rxq->iq, sc, vi->tmr_idx, vi->pktc_idx, vi->qsize_rxq);
1192 
1193 		snprintf(name, sizeof(name), "%s rxq%d-fl",
1194 		    device_get_nameunit(vi->dev), i);
1195 		init_fl(sc, &rxq->fl, vi->qsize_rxq / 8, maxp, name);
1196 
1197 		rc = alloc_rxq(vi, rxq,
1198 		    forwarding_intr_to_fwq(sc) ? -1 : intr_idx, i, oid);
1199 		if (rc != 0)
1200 			goto done;
1201 		intr_idx++;
1202 	}
1203 #ifdef DEV_NETMAP
1204 	if (ifp->if_capabilities & IFCAP_NETMAP)
1205 		intr_idx = saved_idx + max(vi->nrxq, vi->nnmrxq);
1206 #endif
1207 #ifdef TCP_OFFLOAD
1208 	maxp = mtu_to_max_payload(sc, mtu, 1);
1209 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq",
1210 	    CTLFLAG_RD, NULL, "rx queues for offloaded TCP connections");
1211 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1212 
1213 		init_iq(&ofld_rxq->iq, sc, vi->ofld_tmr_idx, vi->ofld_pktc_idx,
1214 		    vi->qsize_rxq);
1215 
1216 		snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
1217 		    device_get_nameunit(vi->dev), i);
1218 		init_fl(sc, &ofld_rxq->fl, vi->qsize_rxq / 8, maxp, name);
1219 
1220 		rc = alloc_ofld_rxq(vi, ofld_rxq,
1221 		    forwarding_intr_to_fwq(sc) ? -1 : intr_idx, i, oid);
1222 		if (rc != 0)
1223 			goto done;
1224 		intr_idx++;
1225 	}
1226 #endif
1227 
1228 	/*
1229 	 * Now the tx queues.
1230 	 */
1231 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "txq", CTLFLAG_RD,
1232 	    NULL, "tx queues");
1233 	for_each_txq(vi, i, txq) {
1234 		iqidx = vi->first_rxq + (i % vi->nrxq);
1235 		snprintf(name, sizeof(name), "%s txq%d",
1236 		    device_get_nameunit(vi->dev), i);
1237 		init_eq(sc, &txq->eq, EQ_ETH, vi->qsize_txq, pi->tx_chan,
1238 		    sc->sge.rxq[iqidx].iq.cntxt_id, name);
1239 
1240 		rc = alloc_txq(vi, txq, i, oid);
1241 		if (rc != 0)
1242 			goto done;
1243 	}
1244 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1245 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_txq",
1246 	    CTLFLAG_RD, NULL, "tx queues for TOE/ETHOFLD");
1247 	for_each_ofld_txq(vi, i, ofld_txq) {
1248 		struct sysctl_oid *oid2;
1249 
1250 		snprintf(name, sizeof(name), "%s ofld_txq%d",
1251 		    device_get_nameunit(vi->dev), i);
1252 		if (vi->nofldrxq > 0) {
1253 			iqidx = vi->first_ofld_rxq + (i % vi->nofldrxq);
1254 			init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq,
1255 			    pi->tx_chan, sc->sge.ofld_rxq[iqidx].iq.cntxt_id,
1256 			    name);
1257 		} else {
1258 			iqidx = vi->first_rxq + (i % vi->nrxq);
1259 			init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq,
1260 			    pi->tx_chan, sc->sge.rxq[iqidx].iq.cntxt_id, name);
1261 		}
1262 
1263 		snprintf(name, sizeof(name), "%d", i);
1264 		oid2 = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
1265 		    name, CTLFLAG_RD, NULL, "offload tx queue");
1266 
1267 		rc = alloc_wrq(sc, vi, ofld_txq, oid2);
1268 		if (rc != 0)
1269 			goto done;
1270 	}
1271 #endif
1272 done:
1273 	if (rc)
1274 		t4_teardown_vi_queues(vi);
1275 
1276 	return (rc);
1277 }
1278 
1279 /*
1280  * Idempotent
1281  */
1282 int
1283 t4_teardown_vi_queues(struct vi_info *vi)
1284 {
1285 	int i;
1286 	struct sge_rxq *rxq;
1287 	struct sge_txq *txq;
1288 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1289 	struct port_info *pi = vi->pi;
1290 	struct adapter *sc = pi->adapter;
1291 	struct sge_wrq *ofld_txq;
1292 #endif
1293 #ifdef TCP_OFFLOAD
1294 	struct sge_ofld_rxq *ofld_rxq;
1295 #endif
1296 #ifdef DEV_NETMAP
1297 	struct sge_nm_rxq *nm_rxq;
1298 	struct sge_nm_txq *nm_txq;
1299 #endif
1300 
1301 	/* Do this before freeing the queues */
1302 	if (vi->flags & VI_SYSCTL_CTX) {
1303 		sysctl_ctx_free(&vi->ctx);
1304 		vi->flags &= ~VI_SYSCTL_CTX;
1305 	}
1306 
1307 #ifdef DEV_NETMAP
1308 	if (vi->ifp->if_capabilities & IFCAP_NETMAP) {
1309 		for_each_nm_txq(vi, i, nm_txq) {
1310 			free_nm_txq(vi, nm_txq);
1311 		}
1312 
1313 		for_each_nm_rxq(vi, i, nm_rxq) {
1314 			free_nm_rxq(vi, nm_rxq);
1315 		}
1316 	}
1317 #endif
1318 
1319 	/*
1320 	 * Take down all the tx queues first, as they reference the rx queues
1321 	 * (for egress updates, etc.).
1322 	 */
1323 
1324 	for_each_txq(vi, i, txq) {
1325 		free_txq(vi, txq);
1326 	}
1327 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1328 	for_each_ofld_txq(vi, i, ofld_txq) {
1329 		free_wrq(sc, ofld_txq);
1330 	}
1331 #endif
1332 
1333 	/*
1334 	 * Then take down the rx queues.
1335 	 */
1336 
1337 	for_each_rxq(vi, i, rxq) {
1338 		free_rxq(vi, rxq);
1339 	}
1340 #ifdef TCP_OFFLOAD
1341 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1342 		free_ofld_rxq(vi, ofld_rxq);
1343 	}
1344 #endif
1345 
1346 	return (0);
1347 }
1348 
1349 /*
1350  * Interrupt handler when the driver is using only 1 interrupt.  This is a very
1351  * unusual scenario.
1352  *
1353  * a) Deals with errors, if any.
1354  * b) Services firmware event queue, which is taking interrupts for all other
1355  *    queues.
1356  */
1357 void
1358 t4_intr_all(void *arg)
1359 {
1360 	struct adapter *sc = arg;
1361 	struct sge_iq *fwq = &sc->sge.fwq;
1362 
1363 	MPASS(sc->intr_count == 1);
1364 
1365 	t4_intr_err(arg);
1366 	t4_intr_evt(fwq);
1367 }
1368 
1369 /*
1370  * Interrupt handler for errors (installed directly when multiple interrupts are
1371  * being used, or called by t4_intr_all).
1372  */
1373 void
1374 t4_intr_err(void *arg)
1375 {
1376 	struct adapter *sc = arg;
1377 
1378 	t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
1379 	t4_slow_intr_handler(sc);
1380 }
1381 
1382 /*
1383  * Interrupt handler for iq-only queues.  The firmware event queue is the only
1384  * such queue right now.
1385  */
1386 void
1387 t4_intr_evt(void *arg)
1388 {
1389 	struct sge_iq *iq = arg;
1390 
1391 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1392 		service_iq(iq, 0);
1393 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1394 	}
1395 }
1396 
1397 /*
1398  * Interrupt handler for iq+fl queues.
1399  */
1400 void
1401 t4_intr(void *arg)
1402 {
1403 	struct sge_iq *iq = arg;
1404 
1405 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1406 		service_iq_fl(iq, 0);
1407 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1408 	}
1409 }
1410 
1411 #ifdef DEV_NETMAP
1412 /*
1413  * Interrupt handler for netmap rx queues.
1414  */
1415 void
1416 t4_nm_intr(void *arg)
1417 {
1418 	struct sge_nm_rxq *nm_rxq = arg;
1419 
1420 	if (atomic_cmpset_int(&nm_rxq->nm_state, NM_ON, NM_BUSY)) {
1421 		service_nm_rxq(nm_rxq);
1422 		(void) atomic_cmpset_int(&nm_rxq->nm_state, NM_BUSY, NM_ON);
1423 	}
1424 }
1425 
1426 /*
1427  * Interrupt handler for vectors shared between NIC and netmap rx queues.
1428  */
1429 void
1430 t4_vi_intr(void *arg)
1431 {
1432 	struct irq *irq = arg;
1433 
1434 	MPASS(irq->nm_rxq != NULL);
1435 	t4_nm_intr(irq->nm_rxq);
1436 
1437 	MPASS(irq->rxq != NULL);
1438 	t4_intr(irq->rxq);
1439 }
1440 #endif
1441 
1442 /*
1443  * Deals with interrupts on an iq-only (no freelist) queue.
1444  */
1445 static int
1446 service_iq(struct sge_iq *iq, int budget)
1447 {
1448 	struct sge_iq *q;
1449 	struct adapter *sc = iq->adapter;
1450 	struct iq_desc *d = &iq->desc[iq->cidx];
1451 	int ndescs = 0, limit;
1452 	int rsp_type;
1453 	uint32_t lq;
1454 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1455 
1456 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1457 	KASSERT((iq->flags & IQ_HAS_FL) == 0,
1458 	    ("%s: called for iq %p with fl (iq->flags 0x%x)", __func__, iq,
1459 	    iq->flags));
1460 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1461 	MPASS((iq->flags & IQ_LRO_ENABLED) == 0);
1462 
1463 	limit = budget ? budget : iq->qsize / 16;
1464 
1465 	/*
1466 	 * We always come back and check the descriptor ring for new indirect
1467 	 * interrupts and other responses after running a single handler.
1468 	 */
1469 	for (;;) {
1470 		while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1471 
1472 			rmb();
1473 
1474 			rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1475 			lq = be32toh(d->rsp.pldbuflen_qid);
1476 
1477 			switch (rsp_type) {
1478 			case X_RSPD_TYPE_FLBUF:
1479 				panic("%s: data for an iq (%p) with no freelist",
1480 				    __func__, iq);
1481 
1482 				/* NOTREACHED */
1483 
1484 			case X_RSPD_TYPE_CPL:
1485 				KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1486 				    ("%s: bad opcode %02x.", __func__,
1487 				    d->rss.opcode));
1488 				t4_cpl_handler[d->rss.opcode](iq, &d->rss, NULL);
1489 				break;
1490 
1491 			case X_RSPD_TYPE_INTR:
1492 				/*
1493 				 * There are 1K interrupt-capable queues (qids 0
1494 				 * through 1023).  A response type indicating a
1495 				 * forwarded interrupt with a qid >= 1K is an
1496 				 * iWARP async notification.
1497 				 */
1498 				if (__predict_true(lq >= 1024)) {
1499 					t4_an_handler(iq, &d->rsp);
1500 					break;
1501 				}
1502 
1503 				q = sc->sge.iqmap[lq - sc->sge.iq_start -
1504 				    sc->sge.iq_base];
1505 				if (atomic_cmpset_int(&q->state, IQS_IDLE,
1506 				    IQS_BUSY)) {
1507 					if (service_iq_fl(q, q->qsize / 16) == 0) {
1508 						(void) atomic_cmpset_int(&q->state,
1509 						    IQS_BUSY, IQS_IDLE);
1510 					} else {
1511 						STAILQ_INSERT_TAIL(&iql, q,
1512 						    link);
1513 					}
1514 				}
1515 				break;
1516 
1517 			default:
1518 				KASSERT(0,
1519 				    ("%s: illegal response type %d on iq %p",
1520 				    __func__, rsp_type, iq));
1521 				log(LOG_ERR,
1522 				    "%s: illegal response type %d on iq %p",
1523 				    device_get_nameunit(sc->dev), rsp_type, iq);
1524 				break;
1525 			}
1526 
1527 			d++;
1528 			if (__predict_false(++iq->cidx == iq->sidx)) {
1529 				iq->cidx = 0;
1530 				iq->gen ^= F_RSPD_GEN;
1531 				d = &iq->desc[0];
1532 			}
1533 			if (__predict_false(++ndescs == limit)) {
1534 				t4_write_reg(sc, sc->sge_gts_reg,
1535 				    V_CIDXINC(ndescs) |
1536 				    V_INGRESSQID(iq->cntxt_id) |
1537 				    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1538 				ndescs = 0;
1539 
1540 				if (budget) {
1541 					return (EINPROGRESS);
1542 				}
1543 			}
1544 		}
1545 
1546 		if (STAILQ_EMPTY(&iql))
1547 			break;
1548 
1549 		/*
1550 		 * Process the head only, and send it to the back of the list if
1551 		 * it's still not done.
1552 		 */
1553 		q = STAILQ_FIRST(&iql);
1554 		STAILQ_REMOVE_HEAD(&iql, link);
1555 		if (service_iq_fl(q, q->qsize / 8) == 0)
1556 			(void) atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1557 		else
1558 			STAILQ_INSERT_TAIL(&iql, q, link);
1559 	}
1560 
1561 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1562 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1563 
1564 	return (0);
1565 }
1566 
1567 static inline int
1568 sort_before_lro(struct lro_ctrl *lro)
1569 {
1570 
1571 	return (lro->lro_mbuf_max != 0);
1572 }
1573 
1574 static inline uint64_t
1575 last_flit_to_ns(struct adapter *sc, uint64_t lf)
1576 {
1577 	uint64_t n = be64toh(lf) & 0xfffffffffffffff;	/* 60b, not 64b. */
1578 
1579 	if (n > UINT64_MAX / 1000000)
1580 		return (n / sc->params.vpd.cclk * 1000000);
1581 	else
1582 		return (n * 1000000 / sc->params.vpd.cclk);
1583 }
1584 
1585 /*
1586  * Deals with interrupts on an iq+fl queue.
1587  */
1588 static int
1589 service_iq_fl(struct sge_iq *iq, int budget)
1590 {
1591 	struct sge_rxq *rxq = iq_to_rxq(iq);
1592 	struct sge_fl *fl;
1593 	struct adapter *sc = iq->adapter;
1594 	struct iq_desc *d = &iq->desc[iq->cidx];
1595 	int ndescs = 0, limit;
1596 	int rsp_type, refill, starved;
1597 	uint32_t lq;
1598 	uint16_t fl_hw_cidx;
1599 	struct mbuf *m0;
1600 #if defined(INET) || defined(INET6)
1601 	const struct timeval lro_timeout = {0, sc->lro_timeout};
1602 	struct lro_ctrl *lro = &rxq->lro;
1603 #endif
1604 
1605 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1606 	MPASS(iq->flags & IQ_HAS_FL);
1607 
1608 	limit = budget ? budget : iq->qsize / 16;
1609 	fl = &rxq->fl;
1610 	fl_hw_cidx = fl->hw_cidx;	/* stable snapshot */
1611 
1612 #if defined(INET) || defined(INET6)
1613 	if (iq->flags & IQ_ADJ_CREDIT) {
1614 		MPASS(sort_before_lro(lro));
1615 		iq->flags &= ~IQ_ADJ_CREDIT;
1616 		if ((d->rsp.u.type_gen & F_RSPD_GEN) != iq->gen) {
1617 			tcp_lro_flush_all(lro);
1618 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(1) |
1619 			    V_INGRESSQID((u32)iq->cntxt_id) |
1620 			    V_SEINTARM(iq->intr_params));
1621 			return (0);
1622 		}
1623 		ndescs = 1;
1624 	}
1625 #else
1626 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1627 #endif
1628 
1629 	while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1630 
1631 		rmb();
1632 
1633 		refill = 0;
1634 		m0 = NULL;
1635 		rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1636 		lq = be32toh(d->rsp.pldbuflen_qid);
1637 
1638 		switch (rsp_type) {
1639 		case X_RSPD_TYPE_FLBUF:
1640 
1641 			m0 = get_fl_payload(sc, fl, lq);
1642 			if (__predict_false(m0 == NULL))
1643 				goto out;
1644 			refill = IDXDIFF(fl->hw_cidx, fl_hw_cidx, fl->sidx) > 2;
1645 
1646 			if (iq->flags & IQ_RX_TIMESTAMP) {
1647 				/*
1648 				 * Fill up rcv_tstmp but do not set M_TSTMP.
1649 				 * rcv_tstmp is not in the format that the
1650 				 * kernel expects and we don't want to mislead
1651 				 * it.  For now this is only for custom code
1652 				 * that knows how to interpret cxgbe's stamp.
1653 				 */
1654 				m0->m_pkthdr.rcv_tstmp =
1655 				    last_flit_to_ns(sc, d->rsp.u.last_flit);
1656 #ifdef notyet
1657 				m0->m_flags |= M_TSTMP;
1658 #endif
1659 			}
1660 
1661 			/* fall through */
1662 
1663 		case X_RSPD_TYPE_CPL:
1664 			KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1665 			    ("%s: bad opcode %02x.", __func__, d->rss.opcode));
1666 			t4_cpl_handler[d->rss.opcode](iq, &d->rss, m0);
1667 			break;
1668 
1669 		case X_RSPD_TYPE_INTR:
1670 
1671 			/*
1672 			 * There are 1K interrupt-capable queues (qids 0
1673 			 * through 1023).  A response type indicating a
1674 			 * forwarded interrupt with a qid >= 1K is an
1675 			 * iWARP async notification.  That is the only
1676 			 * acceptable indirect interrupt on this queue.
1677 			 */
1678 			if (__predict_false(lq < 1024)) {
1679 				panic("%s: indirect interrupt on iq_fl %p "
1680 				    "with qid %u", __func__, iq, lq);
1681 			}
1682 
1683 			t4_an_handler(iq, &d->rsp);
1684 			break;
1685 
1686 		default:
1687 			KASSERT(0, ("%s: illegal response type %d on iq %p",
1688 			    __func__, rsp_type, iq));
1689 			log(LOG_ERR, "%s: illegal response type %d on iq %p",
1690 			    device_get_nameunit(sc->dev), rsp_type, iq);
1691 			break;
1692 		}
1693 
1694 		d++;
1695 		if (__predict_false(++iq->cidx == iq->sidx)) {
1696 			iq->cidx = 0;
1697 			iq->gen ^= F_RSPD_GEN;
1698 			d = &iq->desc[0];
1699 		}
1700 		if (__predict_false(++ndescs == limit)) {
1701 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1702 			    V_INGRESSQID(iq->cntxt_id) |
1703 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1704 			ndescs = 0;
1705 
1706 #if defined(INET) || defined(INET6)
1707 			if (iq->flags & IQ_LRO_ENABLED &&
1708 			    !sort_before_lro(lro) &&
1709 			    sc->lro_timeout != 0) {
1710 				tcp_lro_flush_inactive(lro, &lro_timeout);
1711 			}
1712 #endif
1713 			if (budget) {
1714 				FL_LOCK(fl);
1715 				refill_fl(sc, fl, 32);
1716 				FL_UNLOCK(fl);
1717 
1718 				return (EINPROGRESS);
1719 			}
1720 		}
1721 		if (refill) {
1722 			FL_LOCK(fl);
1723 			refill_fl(sc, fl, 32);
1724 			FL_UNLOCK(fl);
1725 			fl_hw_cidx = fl->hw_cidx;
1726 		}
1727 	}
1728 out:
1729 #if defined(INET) || defined(INET6)
1730 	if (iq->flags & IQ_LRO_ENABLED) {
1731 		if (ndescs > 0 && lro->lro_mbuf_count > 8) {
1732 			MPASS(sort_before_lro(lro));
1733 			/* hold back one credit and don't flush LRO state */
1734 			iq->flags |= IQ_ADJ_CREDIT;
1735 			ndescs--;
1736 		} else {
1737 			tcp_lro_flush_all(lro);
1738 		}
1739 	}
1740 #endif
1741 
1742 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1743 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1744 
1745 	FL_LOCK(fl);
1746 	starved = refill_fl(sc, fl, 64);
1747 	FL_UNLOCK(fl);
1748 	if (__predict_false(starved != 0))
1749 		add_fl_to_sfl(sc, fl);
1750 
1751 	return (0);
1752 }
1753 
1754 static inline int
1755 cl_has_metadata(struct sge_fl *fl, struct cluster_layout *cll)
1756 {
1757 	int rc = fl->flags & FL_BUF_PACKING || cll->region1 > 0;
1758 
1759 	if (rc)
1760 		MPASS(cll->region3 >= CL_METADATA_SIZE);
1761 
1762 	return (rc);
1763 }
1764 
1765 static inline struct cluster_metadata *
1766 cl_metadata(struct adapter *sc, struct sge_fl *fl, struct cluster_layout *cll,
1767     caddr_t cl)
1768 {
1769 
1770 	if (cl_has_metadata(fl, cll)) {
1771 		struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1772 
1773 		return ((struct cluster_metadata *)(cl + swz->size) - 1);
1774 	}
1775 	return (NULL);
1776 }
1777 
1778 static void
1779 rxb_free(struct mbuf *m)
1780 {
1781 	uma_zone_t zone = m->m_ext.ext_arg1;
1782 	void *cl = m->m_ext.ext_arg2;
1783 
1784 	uma_zfree(zone, cl);
1785 	counter_u64_add(extfree_rels, 1);
1786 }
1787 
1788 /*
1789  * The mbuf returned by this function could be allocated from zone_mbuf or
1790  * constructed in spare room in the cluster.
1791  *
1792  * The mbuf carries the payload in one of these ways
1793  * a) frame inside the mbuf (mbuf from zone_mbuf)
1794  * b) m_cljset (for clusters without metadata) zone_mbuf
1795  * c) m_extaddref (cluster with metadata) inline mbuf
1796  * d) m_extaddref (cluster with metadata) zone_mbuf
1797  */
1798 static struct mbuf *
1799 get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1800     int remaining)
1801 {
1802 	struct mbuf *m;
1803 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1804 	struct cluster_layout *cll = &sd->cll;
1805 	struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1806 	struct hw_buf_info *hwb = &sc->sge.hw_buf_info[cll->hwidx];
1807 	struct cluster_metadata *clm = cl_metadata(sc, fl, cll, sd->cl);
1808 	int len, blen;
1809 	caddr_t payload;
1810 
1811 	blen = hwb->size - fl->rx_offset;	/* max possible in this buf */
1812 	len = min(remaining, blen);
1813 	payload = sd->cl + cll->region1 + fl->rx_offset;
1814 	if (fl->flags & FL_BUF_PACKING) {
1815 		const u_int l = fr_offset + len;
1816 		const u_int pad = roundup2(l, fl->buf_boundary) - l;
1817 
1818 		if (fl->rx_offset + len + pad < hwb->size)
1819 			blen = len + pad;
1820 		MPASS(fl->rx_offset + blen <= hwb->size);
1821 	} else {
1822 		MPASS(fl->rx_offset == 0);	/* not packing */
1823 	}
1824 
1825 
1826 	if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) {
1827 
1828 		/*
1829 		 * Copy payload into a freshly allocated mbuf.
1830 		 */
1831 
1832 		m = fr_offset == 0 ?
1833 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1834 		if (m == NULL)
1835 			return (NULL);
1836 		fl->mbuf_allocated++;
1837 
1838 		/* copy data to mbuf */
1839 		bcopy(payload, mtod(m, caddr_t), len);
1840 
1841 	} else if (sd->nmbuf * MSIZE < cll->region1) {
1842 
1843 		/*
1844 		 * There's spare room in the cluster for an mbuf.  Create one
1845 		 * and associate it with the payload that's in the cluster.
1846 		 */
1847 
1848 		MPASS(clm != NULL);
1849 		m = (struct mbuf *)(sd->cl + sd->nmbuf * MSIZE);
1850 		/* No bzero required */
1851 		if (m_init(m, M_NOWAIT, MT_DATA,
1852 		    fr_offset == 0 ? M_PKTHDR | M_NOFREE : M_NOFREE))
1853 			return (NULL);
1854 		fl->mbuf_inlined++;
1855 		m_extaddref(m, payload, blen, &clm->refcount, rxb_free,
1856 		    swz->zone, sd->cl);
1857 		if (sd->nmbuf++ == 0)
1858 			counter_u64_add(extfree_refs, 1);
1859 
1860 	} else {
1861 
1862 		/*
1863 		 * Grab an mbuf from zone_mbuf and associate it with the
1864 		 * payload in the cluster.
1865 		 */
1866 
1867 		m = fr_offset == 0 ?
1868 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1869 		if (m == NULL)
1870 			return (NULL);
1871 		fl->mbuf_allocated++;
1872 		if (clm != NULL) {
1873 			m_extaddref(m, payload, blen, &clm->refcount,
1874 			    rxb_free, swz->zone, sd->cl);
1875 			if (sd->nmbuf++ == 0)
1876 				counter_u64_add(extfree_refs, 1);
1877 		} else {
1878 			m_cljset(m, sd->cl, swz->type);
1879 			sd->cl = NULL;	/* consumed, not a recycle candidate */
1880 		}
1881 	}
1882 	if (fr_offset == 0)
1883 		m->m_pkthdr.len = remaining;
1884 	m->m_len = len;
1885 
1886 	if (fl->flags & FL_BUF_PACKING) {
1887 		fl->rx_offset += blen;
1888 		MPASS(fl->rx_offset <= hwb->size);
1889 		if (fl->rx_offset < hwb->size)
1890 			return (m);	/* without advancing the cidx */
1891 	}
1892 
1893 	if (__predict_false(++fl->cidx % 8 == 0)) {
1894 		uint16_t cidx = fl->cidx / 8;
1895 
1896 		if (__predict_false(cidx == fl->sidx))
1897 			fl->cidx = cidx = 0;
1898 		fl->hw_cidx = cidx;
1899 	}
1900 	fl->rx_offset = 0;
1901 
1902 	return (m);
1903 }
1904 
1905 static struct mbuf *
1906 get_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf)
1907 {
1908 	struct mbuf *m0, *m, **pnext;
1909 	u_int remaining;
1910 	const u_int total = G_RSPD_LEN(len_newbuf);
1911 
1912 	if (__predict_false(fl->flags & FL_BUF_RESUME)) {
1913 		M_ASSERTPKTHDR(fl->m0);
1914 		MPASS(fl->m0->m_pkthdr.len == total);
1915 		MPASS(fl->remaining < total);
1916 
1917 		m0 = fl->m0;
1918 		pnext = fl->pnext;
1919 		remaining = fl->remaining;
1920 		fl->flags &= ~FL_BUF_RESUME;
1921 		goto get_segment;
1922 	}
1923 
1924 	if (fl->rx_offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
1925 		fl->rx_offset = 0;
1926 		if (__predict_false(++fl->cidx % 8 == 0)) {
1927 			uint16_t cidx = fl->cidx / 8;
1928 
1929 			if (__predict_false(cidx == fl->sidx))
1930 				fl->cidx = cidx = 0;
1931 			fl->hw_cidx = cidx;
1932 		}
1933 	}
1934 
1935 	/*
1936 	 * Payload starts at rx_offset in the current hw buffer.  Its length is
1937 	 * 'len' and it may span multiple hw buffers.
1938 	 */
1939 
1940 	m0 = get_scatter_segment(sc, fl, 0, total);
1941 	if (m0 == NULL)
1942 		return (NULL);
1943 	remaining = total - m0->m_len;
1944 	pnext = &m0->m_next;
1945 	while (remaining > 0) {
1946 get_segment:
1947 		MPASS(fl->rx_offset == 0);
1948 		m = get_scatter_segment(sc, fl, total - remaining, remaining);
1949 		if (__predict_false(m == NULL)) {
1950 			fl->m0 = m0;
1951 			fl->pnext = pnext;
1952 			fl->remaining = remaining;
1953 			fl->flags |= FL_BUF_RESUME;
1954 			return (NULL);
1955 		}
1956 		*pnext = m;
1957 		pnext = &m->m_next;
1958 		remaining -= m->m_len;
1959 	}
1960 	*pnext = NULL;
1961 
1962 	M_ASSERTPKTHDR(m0);
1963 	return (m0);
1964 }
1965 
1966 static int
1967 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
1968 {
1969 	struct sge_rxq *rxq = iq_to_rxq(iq);
1970 	struct ifnet *ifp = rxq->ifp;
1971 	struct adapter *sc = iq->adapter;
1972 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
1973 #if defined(INET) || defined(INET6)
1974 	struct lro_ctrl *lro = &rxq->lro;
1975 #endif
1976 	static const int sw_hashtype[4][2] = {
1977 		{M_HASHTYPE_NONE, M_HASHTYPE_NONE},
1978 		{M_HASHTYPE_RSS_IPV4, M_HASHTYPE_RSS_IPV6},
1979 		{M_HASHTYPE_RSS_TCP_IPV4, M_HASHTYPE_RSS_TCP_IPV6},
1980 		{M_HASHTYPE_RSS_UDP_IPV4, M_HASHTYPE_RSS_UDP_IPV6},
1981 	};
1982 
1983 	KASSERT(m0 != NULL, ("%s: no payload with opcode %02x", __func__,
1984 	    rss->opcode));
1985 
1986 	m0->m_pkthdr.len -= sc->params.sge.fl_pktshift;
1987 	m0->m_len -= sc->params.sge.fl_pktshift;
1988 	m0->m_data += sc->params.sge.fl_pktshift;
1989 
1990 	m0->m_pkthdr.rcvif = ifp;
1991 	M_HASHTYPE_SET(m0, sw_hashtype[rss->hash_type][rss->ipv6]);
1992 	m0->m_pkthdr.flowid = be32toh(rss->hash_val);
1993 
1994 	if (cpl->csum_calc && !(cpl->err_vec & sc->params.tp.err_vec_mask)) {
1995 		if (ifp->if_capenable & IFCAP_RXCSUM &&
1996 		    cpl->l2info & htobe32(F_RXF_IP)) {
1997 			m0->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
1998 			    CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
1999 			rxq->rxcsum++;
2000 		} else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
2001 		    cpl->l2info & htobe32(F_RXF_IP6)) {
2002 			m0->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
2003 			    CSUM_PSEUDO_HDR);
2004 			rxq->rxcsum++;
2005 		}
2006 
2007 		if (__predict_false(cpl->ip_frag))
2008 			m0->m_pkthdr.csum_data = be16toh(cpl->csum);
2009 		else
2010 			m0->m_pkthdr.csum_data = 0xffff;
2011 	}
2012 
2013 	if (cpl->vlan_ex) {
2014 		m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
2015 		m0->m_flags |= M_VLANTAG;
2016 		rxq->vlan_extraction++;
2017 	}
2018 
2019 #if defined(INET) || defined(INET6)
2020 	if (iq->flags & IQ_LRO_ENABLED) {
2021 		if (sort_before_lro(lro)) {
2022 			tcp_lro_queue_mbuf(lro, m0);
2023 			return (0); /* queued for sort, then LRO */
2024 		}
2025 		if (tcp_lro_rx(lro, m0, 0) == 0)
2026 			return (0); /* queued for LRO */
2027 	}
2028 #endif
2029 	ifp->if_input(ifp, m0);
2030 
2031 	return (0);
2032 }
2033 
2034 /*
2035  * Must drain the wrq or make sure that someone else will.
2036  */
2037 static void
2038 wrq_tx_drain(void *arg, int n)
2039 {
2040 	struct sge_wrq *wrq = arg;
2041 	struct sge_eq *eq = &wrq->eq;
2042 
2043 	EQ_LOCK(eq);
2044 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2045 		drain_wrq_wr_list(wrq->adapter, wrq);
2046 	EQ_UNLOCK(eq);
2047 }
2048 
2049 static void
2050 drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq)
2051 {
2052 	struct sge_eq *eq = &wrq->eq;
2053 	u_int available, dbdiff;	/* # of hardware descriptors */
2054 	u_int n;
2055 	struct wrqe *wr;
2056 	struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
2057 
2058 	EQ_LOCK_ASSERT_OWNED(eq);
2059 	MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
2060 	wr = STAILQ_FIRST(&wrq->wr_list);
2061 	MPASS(wr != NULL);	/* Must be called with something useful to do */
2062 	MPASS(eq->pidx == eq->dbidx);
2063 	dbdiff = 0;
2064 
2065 	do {
2066 		eq->cidx = read_hw_cidx(eq);
2067 		if (eq->pidx == eq->cidx)
2068 			available = eq->sidx - 1;
2069 		else
2070 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2071 
2072 		MPASS(wr->wrq == wrq);
2073 		n = howmany(wr->wr_len, EQ_ESIZE);
2074 		if (available < n)
2075 			break;
2076 
2077 		dst = (void *)&eq->desc[eq->pidx];
2078 		if (__predict_true(eq->sidx - eq->pidx > n)) {
2079 			/* Won't wrap, won't end exactly at the status page. */
2080 			bcopy(&wr->wr[0], dst, wr->wr_len);
2081 			eq->pidx += n;
2082 		} else {
2083 			int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE;
2084 
2085 			bcopy(&wr->wr[0], dst, first_portion);
2086 			if (wr->wr_len > first_portion) {
2087 				bcopy(&wr->wr[first_portion], &eq->desc[0],
2088 				    wr->wr_len - first_portion);
2089 			}
2090 			eq->pidx = n - (eq->sidx - eq->pidx);
2091 		}
2092 		wrq->tx_wrs_copied++;
2093 
2094 		if (available < eq->sidx / 4 &&
2095 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2096 				/*
2097 				 * XXX: This is not 100% reliable with some
2098 				 * types of WRs.  But this is a very unusual
2099 				 * situation for an ofld/ctrl queue anyway.
2100 				 */
2101 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2102 			    F_FW_WR_EQUEQ);
2103 		}
2104 
2105 		dbdiff += n;
2106 		if (dbdiff >= 16) {
2107 			ring_eq_db(sc, eq, dbdiff);
2108 			dbdiff = 0;
2109 		}
2110 
2111 		STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
2112 		free_wrqe(wr);
2113 		MPASS(wrq->nwr_pending > 0);
2114 		wrq->nwr_pending--;
2115 		MPASS(wrq->ndesc_needed >= n);
2116 		wrq->ndesc_needed -= n;
2117 	} while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL);
2118 
2119 	if (dbdiff)
2120 		ring_eq_db(sc, eq, dbdiff);
2121 }
2122 
2123 /*
2124  * Doesn't fail.  Holds on to work requests it can't send right away.
2125  */
2126 void
2127 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
2128 {
2129 #ifdef INVARIANTS
2130 	struct sge_eq *eq = &wrq->eq;
2131 #endif
2132 
2133 	EQ_LOCK_ASSERT_OWNED(eq);
2134 	MPASS(wr != NULL);
2135 	MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN);
2136 	MPASS((wr->wr_len & 0x7) == 0);
2137 
2138 	STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
2139 	wrq->nwr_pending++;
2140 	wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE);
2141 
2142 	if (!TAILQ_EMPTY(&wrq->incomplete_wrs))
2143 		return;	/* commit_wrq_wr will drain wr_list as well. */
2144 
2145 	drain_wrq_wr_list(sc, wrq);
2146 
2147 	/* Doorbell must have caught up to the pidx. */
2148 	MPASS(eq->pidx == eq->dbidx);
2149 }
2150 
2151 void
2152 t4_update_fl_bufsize(struct ifnet *ifp)
2153 {
2154 	struct vi_info *vi = ifp->if_softc;
2155 	struct adapter *sc = vi->pi->adapter;
2156 	struct sge_rxq *rxq;
2157 #ifdef TCP_OFFLOAD
2158 	struct sge_ofld_rxq *ofld_rxq;
2159 #endif
2160 	struct sge_fl *fl;
2161 	int i, maxp, mtu = ifp->if_mtu;
2162 
2163 	maxp = mtu_to_max_payload(sc, mtu, 0);
2164 	for_each_rxq(vi, i, rxq) {
2165 		fl = &rxq->fl;
2166 
2167 		FL_LOCK(fl);
2168 		find_best_refill_source(sc, fl, maxp);
2169 		FL_UNLOCK(fl);
2170 	}
2171 #ifdef TCP_OFFLOAD
2172 	maxp = mtu_to_max_payload(sc, mtu, 1);
2173 	for_each_ofld_rxq(vi, i, ofld_rxq) {
2174 		fl = &ofld_rxq->fl;
2175 
2176 		FL_LOCK(fl);
2177 		find_best_refill_source(sc, fl, maxp);
2178 		FL_UNLOCK(fl);
2179 	}
2180 #endif
2181 }
2182 
2183 static inline int
2184 mbuf_nsegs(struct mbuf *m)
2185 {
2186 
2187 	M_ASSERTPKTHDR(m);
2188 	KASSERT(m->m_pkthdr.l5hlen > 0,
2189 	    ("%s: mbuf %p missing information on # of segments.", __func__, m));
2190 
2191 	return (m->m_pkthdr.l5hlen);
2192 }
2193 
2194 static inline void
2195 set_mbuf_nsegs(struct mbuf *m, uint8_t nsegs)
2196 {
2197 
2198 	M_ASSERTPKTHDR(m);
2199 	m->m_pkthdr.l5hlen = nsegs;
2200 }
2201 
2202 static inline int
2203 mbuf_cflags(struct mbuf *m)
2204 {
2205 
2206 	M_ASSERTPKTHDR(m);
2207 	return (m->m_pkthdr.PH_loc.eight[4]);
2208 }
2209 
2210 static inline void
2211 set_mbuf_cflags(struct mbuf *m, uint8_t flags)
2212 {
2213 
2214 	M_ASSERTPKTHDR(m);
2215 	m->m_pkthdr.PH_loc.eight[4] = flags;
2216 }
2217 
2218 static inline int
2219 mbuf_len16(struct mbuf *m)
2220 {
2221 	int n;
2222 
2223 	M_ASSERTPKTHDR(m);
2224 	n = m->m_pkthdr.PH_loc.eight[0];
2225 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2226 
2227 	return (n);
2228 }
2229 
2230 static inline void
2231 set_mbuf_len16(struct mbuf *m, uint8_t len16)
2232 {
2233 
2234 	M_ASSERTPKTHDR(m);
2235 	m->m_pkthdr.PH_loc.eight[0] = len16;
2236 }
2237 
2238 #ifdef RATELIMIT
2239 static inline int
2240 mbuf_eo_nsegs(struct mbuf *m)
2241 {
2242 
2243 	M_ASSERTPKTHDR(m);
2244 	return (m->m_pkthdr.PH_loc.eight[1]);
2245 }
2246 
2247 static inline void
2248 set_mbuf_eo_nsegs(struct mbuf *m, uint8_t nsegs)
2249 {
2250 
2251 	M_ASSERTPKTHDR(m);
2252 	m->m_pkthdr.PH_loc.eight[1] = nsegs;
2253 }
2254 
2255 static inline int
2256 mbuf_eo_len16(struct mbuf *m)
2257 {
2258 	int n;
2259 
2260 	M_ASSERTPKTHDR(m);
2261 	n = m->m_pkthdr.PH_loc.eight[2];
2262 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2263 
2264 	return (n);
2265 }
2266 
2267 static inline void
2268 set_mbuf_eo_len16(struct mbuf *m, uint8_t len16)
2269 {
2270 
2271 	M_ASSERTPKTHDR(m);
2272 	m->m_pkthdr.PH_loc.eight[2] = len16;
2273 }
2274 
2275 static inline int
2276 mbuf_eo_tsclk_tsoff(struct mbuf *m)
2277 {
2278 
2279 	M_ASSERTPKTHDR(m);
2280 	return (m->m_pkthdr.PH_loc.eight[3]);
2281 }
2282 
2283 static inline void
2284 set_mbuf_eo_tsclk_tsoff(struct mbuf *m, uint8_t tsclk_tsoff)
2285 {
2286 
2287 	M_ASSERTPKTHDR(m);
2288 	m->m_pkthdr.PH_loc.eight[3] = tsclk_tsoff;
2289 }
2290 
2291 static inline int
2292 needs_eo(struct mbuf *m)
2293 {
2294 
2295 	return (m->m_pkthdr.snd_tag != NULL);
2296 }
2297 #endif
2298 
2299 /*
2300  * Try to allocate an mbuf to contain a raw work request.  To make it
2301  * easy to construct the work request, don't allocate a chain but a
2302  * single mbuf.
2303  */
2304 struct mbuf *
2305 alloc_wr_mbuf(int len, int how)
2306 {
2307 	struct mbuf *m;
2308 
2309 	if (len <= MHLEN)
2310 		m = m_gethdr(how, MT_DATA);
2311 	else if (len <= MCLBYTES)
2312 		m = m_getcl(how, MT_DATA, M_PKTHDR);
2313 	else
2314 		m = NULL;
2315 	if (m == NULL)
2316 		return (NULL);
2317 	m->m_pkthdr.len = len;
2318 	m->m_len = len;
2319 	set_mbuf_cflags(m, MC_RAW_WR);
2320 	set_mbuf_len16(m, howmany(len, 16));
2321 	return (m);
2322 }
2323 
2324 static inline int
2325 needs_tso(struct mbuf *m)
2326 {
2327 
2328 	M_ASSERTPKTHDR(m);
2329 
2330 	return (m->m_pkthdr.csum_flags & CSUM_TSO);
2331 }
2332 
2333 static inline int
2334 needs_l3_csum(struct mbuf *m)
2335 {
2336 
2337 	M_ASSERTPKTHDR(m);
2338 
2339 	return (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO));
2340 }
2341 
2342 static inline int
2343 needs_l4_csum(struct mbuf *m)
2344 {
2345 
2346 	M_ASSERTPKTHDR(m);
2347 
2348 	return (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
2349 	    CSUM_TCP_IPV6 | CSUM_TSO));
2350 }
2351 
2352 static inline int
2353 needs_tcp_csum(struct mbuf *m)
2354 {
2355 
2356 	M_ASSERTPKTHDR(m);
2357 	return (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_TCP_IPV6 | CSUM_TSO));
2358 }
2359 
2360 #ifdef RATELIMIT
2361 static inline int
2362 needs_udp_csum(struct mbuf *m)
2363 {
2364 
2365 	M_ASSERTPKTHDR(m);
2366 	return (m->m_pkthdr.csum_flags & (CSUM_UDP | CSUM_UDP_IPV6));
2367 }
2368 #endif
2369 
2370 static inline int
2371 needs_vlan_insertion(struct mbuf *m)
2372 {
2373 
2374 	M_ASSERTPKTHDR(m);
2375 
2376 	return (m->m_flags & M_VLANTAG);
2377 }
2378 
2379 static void *
2380 m_advance(struct mbuf **pm, int *poffset, int len)
2381 {
2382 	struct mbuf *m = *pm;
2383 	int offset = *poffset;
2384 	uintptr_t p = 0;
2385 
2386 	MPASS(len > 0);
2387 
2388 	for (;;) {
2389 		if (offset + len < m->m_len) {
2390 			offset += len;
2391 			p = mtod(m, uintptr_t) + offset;
2392 			break;
2393 		}
2394 		len -= m->m_len - offset;
2395 		m = m->m_next;
2396 		offset = 0;
2397 		MPASS(m != NULL);
2398 	}
2399 	*poffset = offset;
2400 	*pm = m;
2401 	return ((void *)p);
2402 }
2403 
2404 /*
2405  * Can deal with empty mbufs in the chain that have m_len = 0, but the chain
2406  * must have at least one mbuf that's not empty.  It is possible for this
2407  * routine to return 0 if skip accounts for all the contents of the mbuf chain.
2408  */
2409 static inline int
2410 count_mbuf_nsegs(struct mbuf *m, int skip)
2411 {
2412 	vm_paddr_t lastb, next;
2413 	vm_offset_t va;
2414 	int len, nsegs;
2415 
2416 	M_ASSERTPKTHDR(m);
2417 	MPASS(m->m_pkthdr.len > 0);
2418 	MPASS(m->m_pkthdr.len >= skip);
2419 
2420 	nsegs = 0;
2421 	lastb = 0;
2422 	for (; m; m = m->m_next) {
2423 
2424 		len = m->m_len;
2425 		if (__predict_false(len == 0))
2426 			continue;
2427 		if (skip >= len) {
2428 			skip -= len;
2429 			continue;
2430 		}
2431 		va = mtod(m, vm_offset_t) + skip;
2432 		len -= skip;
2433 		skip = 0;
2434 		next = pmap_kextract(va);
2435 		nsegs += sglist_count((void *)(uintptr_t)va, len);
2436 		if (lastb + 1 == next)
2437 			nsegs--;
2438 		lastb = pmap_kextract(va + len - 1);
2439 	}
2440 
2441 	return (nsegs);
2442 }
2443 
2444 /*
2445  * Analyze the mbuf to determine its tx needs.  The mbuf passed in may change:
2446  * a) caller can assume it's been freed if this function returns with an error.
2447  * b) it may get defragged up if the gather list is too long for the hardware.
2448  */
2449 int
2450 parse_pkt(struct adapter *sc, struct mbuf **mp)
2451 {
2452 	struct mbuf *m0 = *mp, *m;
2453 	int rc, nsegs, defragged = 0, offset;
2454 	struct ether_header *eh;
2455 	void *l3hdr;
2456 #if defined(INET) || defined(INET6)
2457 	struct tcphdr *tcp;
2458 #endif
2459 	uint16_t eh_type;
2460 
2461 	M_ASSERTPKTHDR(m0);
2462 	if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) {
2463 		rc = EINVAL;
2464 fail:
2465 		m_freem(m0);
2466 		*mp = NULL;
2467 		return (rc);
2468 	}
2469 restart:
2470 	/*
2471 	 * First count the number of gather list segments in the payload.
2472 	 * Defrag the mbuf if nsegs exceeds the hardware limit.
2473 	 */
2474 	M_ASSERTPKTHDR(m0);
2475 	MPASS(m0->m_pkthdr.len > 0);
2476 	nsegs = count_mbuf_nsegs(m0, 0);
2477 	if (nsegs > (needs_tso(m0) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS)) {
2478 		if (defragged++ > 0 || (m = m_defrag(m0, M_NOWAIT)) == NULL) {
2479 			rc = EFBIG;
2480 			goto fail;
2481 		}
2482 		*mp = m0 = m;	/* update caller's copy after defrag */
2483 		goto restart;
2484 	}
2485 
2486 	if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN)) {
2487 		m0 = m_pullup(m0, m0->m_pkthdr.len);
2488 		if (m0 == NULL) {
2489 			/* Should have left well enough alone. */
2490 			rc = EFBIG;
2491 			goto fail;
2492 		}
2493 		*mp = m0;	/* update caller's copy after pullup */
2494 		goto restart;
2495 	}
2496 	set_mbuf_nsegs(m0, nsegs);
2497 	set_mbuf_cflags(m0, 0);
2498 	if (sc->flags & IS_VF)
2499 		set_mbuf_len16(m0, txpkt_vm_len16(nsegs, needs_tso(m0)));
2500 	else
2501 		set_mbuf_len16(m0, txpkt_len16(nsegs, needs_tso(m0)));
2502 
2503 #ifdef RATELIMIT
2504 	/*
2505 	 * Ethofld is limited to TCP and UDP for now, and only when L4 hw
2506 	 * checksumming is enabled.  needs_l4_csum happens to check for all the
2507 	 * right things.
2508 	 */
2509 	if (__predict_false(needs_eo(m0) && !needs_l4_csum(m0)))
2510 		m0->m_pkthdr.snd_tag = NULL;
2511 #endif
2512 
2513 	if (!needs_tso(m0) &&
2514 #ifdef RATELIMIT
2515 	    !needs_eo(m0) &&
2516 #endif
2517 	    !(sc->flags & IS_VF && (needs_l3_csum(m0) || needs_l4_csum(m0))))
2518 		return (0);
2519 
2520 	m = m0;
2521 	eh = mtod(m, struct ether_header *);
2522 	eh_type = ntohs(eh->ether_type);
2523 	if (eh_type == ETHERTYPE_VLAN) {
2524 		struct ether_vlan_header *evh = (void *)eh;
2525 
2526 		eh_type = ntohs(evh->evl_proto);
2527 		m0->m_pkthdr.l2hlen = sizeof(*evh);
2528 	} else
2529 		m0->m_pkthdr.l2hlen = sizeof(*eh);
2530 
2531 	offset = 0;
2532 	l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2533 
2534 	switch (eh_type) {
2535 #ifdef INET6
2536 	case ETHERTYPE_IPV6:
2537 	{
2538 		struct ip6_hdr *ip6 = l3hdr;
2539 
2540 		MPASS(!needs_tso(m0) || ip6->ip6_nxt == IPPROTO_TCP);
2541 
2542 		m0->m_pkthdr.l3hlen = sizeof(*ip6);
2543 		break;
2544 	}
2545 #endif
2546 #ifdef INET
2547 	case ETHERTYPE_IP:
2548 	{
2549 		struct ip *ip = l3hdr;
2550 
2551 		m0->m_pkthdr.l3hlen = ip->ip_hl * 4;
2552 		break;
2553 	}
2554 #endif
2555 	default:
2556 		panic("%s: ethertype 0x%04x unknown.  if_cxgbe must be compiled"
2557 		    " with the same INET/INET6 options as the kernel.",
2558 		    __func__, eh_type);
2559 	}
2560 
2561 #if defined(INET) || defined(INET6)
2562 	if (needs_tcp_csum(m0)) {
2563 		tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen);
2564 		m0->m_pkthdr.l4hlen = tcp->th_off * 4;
2565 #ifdef RATELIMIT
2566 		if (tsclk >= 0 && *(uint32_t *)(tcp + 1) == ntohl(0x0101080a)) {
2567 			set_mbuf_eo_tsclk_tsoff(m0,
2568 			    V_FW_ETH_TX_EO_WR_TSCLK(tsclk) |
2569 			    V_FW_ETH_TX_EO_WR_TSOFF(sizeof(*tcp) / 2 + 1));
2570 		} else
2571 			set_mbuf_eo_tsclk_tsoff(m0, 0);
2572 	} else if (needs_udp_csum(m)) {
2573 		m0->m_pkthdr.l4hlen = sizeof(struct udphdr);
2574 #endif
2575 	}
2576 #ifdef RATELIMIT
2577 	if (needs_eo(m0)) {
2578 		u_int immhdrs;
2579 
2580 		/* EO WRs have the headers in the WR and not the GL. */
2581 		immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen +
2582 		    m0->m_pkthdr.l4hlen;
2583 		nsegs = count_mbuf_nsegs(m0, immhdrs);
2584 		set_mbuf_eo_nsegs(m0, nsegs);
2585 		set_mbuf_eo_len16(m0,
2586 		    txpkt_eo_len16(nsegs, immhdrs, needs_tso(m0)));
2587 	}
2588 #endif
2589 #endif
2590 	MPASS(m0 == *mp);
2591 	return (0);
2592 }
2593 
2594 void *
2595 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie)
2596 {
2597 	struct sge_eq *eq = &wrq->eq;
2598 	struct adapter *sc = wrq->adapter;
2599 	int ndesc, available;
2600 	struct wrqe *wr;
2601 	void *w;
2602 
2603 	MPASS(len16 > 0);
2604 	ndesc = howmany(len16, EQ_ESIZE / 16);
2605 	MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC);
2606 
2607 	EQ_LOCK(eq);
2608 
2609 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2610 		drain_wrq_wr_list(sc, wrq);
2611 
2612 	if (!STAILQ_EMPTY(&wrq->wr_list)) {
2613 slowpath:
2614 		EQ_UNLOCK(eq);
2615 		wr = alloc_wrqe(len16 * 16, wrq);
2616 		if (__predict_false(wr == NULL))
2617 			return (NULL);
2618 		cookie->pidx = -1;
2619 		cookie->ndesc = ndesc;
2620 		return (&wr->wr);
2621 	}
2622 
2623 	eq->cidx = read_hw_cidx(eq);
2624 	if (eq->pidx == eq->cidx)
2625 		available = eq->sidx - 1;
2626 	else
2627 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2628 	if (available < ndesc)
2629 		goto slowpath;
2630 
2631 	cookie->pidx = eq->pidx;
2632 	cookie->ndesc = ndesc;
2633 	TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link);
2634 
2635 	w = &eq->desc[eq->pidx];
2636 	IDXINCR(eq->pidx, ndesc, eq->sidx);
2637 	if (__predict_false(cookie->pidx + ndesc > eq->sidx)) {
2638 		w = &wrq->ss[0];
2639 		wrq->ss_pidx = cookie->pidx;
2640 		wrq->ss_len = len16 * 16;
2641 	}
2642 
2643 	EQ_UNLOCK(eq);
2644 
2645 	return (w);
2646 }
2647 
2648 void
2649 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie)
2650 {
2651 	struct sge_eq *eq = &wrq->eq;
2652 	struct adapter *sc = wrq->adapter;
2653 	int ndesc, pidx;
2654 	struct wrq_cookie *prev, *next;
2655 
2656 	if (cookie->pidx == -1) {
2657 		struct wrqe *wr = __containerof(w, struct wrqe, wr);
2658 
2659 		t4_wrq_tx(sc, wr);
2660 		return;
2661 	}
2662 
2663 	if (__predict_false(w == &wrq->ss[0])) {
2664 		int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE;
2665 
2666 		MPASS(wrq->ss_len > n);	/* WR had better wrap around. */
2667 		bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n);
2668 		bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n);
2669 		wrq->tx_wrs_ss++;
2670 	} else
2671 		wrq->tx_wrs_direct++;
2672 
2673 	EQ_LOCK(eq);
2674 	ndesc = cookie->ndesc;	/* Can be more than SGE_MAX_WR_NDESC here. */
2675 	pidx = cookie->pidx;
2676 	MPASS(pidx >= 0 && pidx < eq->sidx);
2677 	prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link);
2678 	next = TAILQ_NEXT(cookie, link);
2679 	if (prev == NULL) {
2680 		MPASS(pidx == eq->dbidx);
2681 		if (next == NULL || ndesc >= 16) {
2682 			int available;
2683 			struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
2684 
2685 			/*
2686 			 * Note that the WR via which we'll request tx updates
2687 			 * is at pidx and not eq->pidx, which has moved on
2688 			 * already.
2689 			 */
2690 			dst = (void *)&eq->desc[pidx];
2691 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2692 			if (available < eq->sidx / 4 &&
2693 			    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2694 				/*
2695 				 * XXX: This is not 100% reliable with some
2696 				 * types of WRs.  But this is a very unusual
2697 				 * situation for an ofld/ctrl queue anyway.
2698 				 */
2699 				dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2700 				    F_FW_WR_EQUEQ);
2701 			}
2702 
2703 			ring_eq_db(wrq->adapter, eq, ndesc);
2704 		} else {
2705 			MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc);
2706 			next->pidx = pidx;
2707 			next->ndesc += ndesc;
2708 		}
2709 	} else {
2710 		MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc);
2711 		prev->ndesc += ndesc;
2712 	}
2713 	TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link);
2714 
2715 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2716 		drain_wrq_wr_list(sc, wrq);
2717 
2718 #ifdef INVARIANTS
2719 	if (TAILQ_EMPTY(&wrq->incomplete_wrs)) {
2720 		/* Doorbell must have caught up to the pidx. */
2721 		MPASS(wrq->eq.pidx == wrq->eq.dbidx);
2722 	}
2723 #endif
2724 	EQ_UNLOCK(eq);
2725 }
2726 
2727 static u_int
2728 can_resume_eth_tx(struct mp_ring *r)
2729 {
2730 	struct sge_eq *eq = r->cookie;
2731 
2732 	return (total_available_tx_desc(eq) > eq->sidx / 8);
2733 }
2734 
2735 static inline int
2736 cannot_use_txpkts(struct mbuf *m)
2737 {
2738 	/* maybe put a GL limit too, to avoid silliness? */
2739 
2740 	return (needs_tso(m) || (mbuf_cflags(m) & MC_RAW_WR) != 0);
2741 }
2742 
2743 static inline int
2744 discard_tx(struct sge_eq *eq)
2745 {
2746 
2747 	return ((eq->flags & (EQ_ENABLED | EQ_QFLUSH)) != EQ_ENABLED);
2748 }
2749 
2750 static inline int
2751 wr_can_update_eq(struct fw_eth_tx_pkts_wr *wr)
2752 {
2753 
2754 	switch (G_FW_WR_OP(be32toh(wr->op_pkd))) {
2755 	case FW_ULPTX_WR:
2756 	case FW_ETH_TX_PKT_WR:
2757 	case FW_ETH_TX_PKTS_WR:
2758 	case FW_ETH_TX_PKT_VM_WR:
2759 		return (1);
2760 	default:
2761 		return (0);
2762 	}
2763 }
2764 
2765 /*
2766  * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to
2767  * be consumed.  Return the actual number consumed.  0 indicates a stall.
2768  */
2769 static u_int
2770 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx)
2771 {
2772 	struct sge_txq *txq = r->cookie;
2773 	struct sge_eq *eq = &txq->eq;
2774 	struct ifnet *ifp = txq->ifp;
2775 	struct vi_info *vi = ifp->if_softc;
2776 	struct port_info *pi = vi->pi;
2777 	struct adapter *sc = pi->adapter;
2778 	u_int total, remaining;		/* # of packets */
2779 	u_int available, dbdiff;	/* # of hardware descriptors */
2780 	u_int n, next_cidx;
2781 	struct mbuf *m0, *tail;
2782 	struct txpkts txp;
2783 	struct fw_eth_tx_pkts_wr *wr;	/* any fw WR struct will do */
2784 
2785 	remaining = IDXDIFF(pidx, cidx, r->size);
2786 	MPASS(remaining > 0);	/* Must not be called without work to do. */
2787 	total = 0;
2788 
2789 	TXQ_LOCK(txq);
2790 	if (__predict_false(discard_tx(eq))) {
2791 		while (cidx != pidx) {
2792 			m0 = r->items[cidx];
2793 			m_freem(m0);
2794 			if (++cidx == r->size)
2795 				cidx = 0;
2796 		}
2797 		reclaim_tx_descs(txq, 2048);
2798 		total = remaining;
2799 		goto done;
2800 	}
2801 
2802 	/* How many hardware descriptors do we have readily available. */
2803 	if (eq->pidx == eq->cidx)
2804 		available = eq->sidx - 1;
2805 	else
2806 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2807 	dbdiff = IDXDIFF(eq->pidx, eq->dbidx, eq->sidx);
2808 
2809 	while (remaining > 0) {
2810 
2811 		m0 = r->items[cidx];
2812 		M_ASSERTPKTHDR(m0);
2813 		MPASS(m0->m_nextpkt == NULL);
2814 
2815 		if (available < SGE_MAX_WR_NDESC) {
2816 			available += reclaim_tx_descs(txq, 64);
2817 			if (available < howmany(mbuf_len16(m0), EQ_ESIZE / 16))
2818 				break;	/* out of descriptors */
2819 		}
2820 
2821 		next_cidx = cidx + 1;
2822 		if (__predict_false(next_cidx == r->size))
2823 			next_cidx = 0;
2824 
2825 		wr = (void *)&eq->desc[eq->pidx];
2826 		if (sc->flags & IS_VF) {
2827 			total++;
2828 			remaining--;
2829 			ETHER_BPF_MTAP(ifp, m0);
2830 			n = write_txpkt_vm_wr(sc, txq, (void *)wr, m0,
2831 			    available);
2832 		} else if (remaining > 1 &&
2833 		    try_txpkts(m0, r->items[next_cidx], &txp, available) == 0) {
2834 
2835 			/* pkts at cidx, next_cidx should both be in txp. */
2836 			MPASS(txp.npkt == 2);
2837 			tail = r->items[next_cidx];
2838 			MPASS(tail->m_nextpkt == NULL);
2839 			ETHER_BPF_MTAP(ifp, m0);
2840 			ETHER_BPF_MTAP(ifp, tail);
2841 			m0->m_nextpkt = tail;
2842 
2843 			if (__predict_false(++next_cidx == r->size))
2844 				next_cidx = 0;
2845 
2846 			while (next_cidx != pidx) {
2847 				if (add_to_txpkts(r->items[next_cidx], &txp,
2848 				    available) != 0)
2849 					break;
2850 				tail->m_nextpkt = r->items[next_cidx];
2851 				tail = tail->m_nextpkt;
2852 				ETHER_BPF_MTAP(ifp, tail);
2853 				if (__predict_false(++next_cidx == r->size))
2854 					next_cidx = 0;
2855 			}
2856 
2857 			n = write_txpkts_wr(txq, wr, m0, &txp, available);
2858 			total += txp.npkt;
2859 			remaining -= txp.npkt;
2860 		} else if (mbuf_cflags(m0) & MC_RAW_WR) {
2861 			total++;
2862 			remaining--;
2863 			n = write_raw_wr(txq, (void *)wr, m0, available);
2864 		} else {
2865 			total++;
2866 			remaining--;
2867 			ETHER_BPF_MTAP(ifp, m0);
2868 			n = write_txpkt_wr(txq, (void *)wr, m0, available);
2869 		}
2870 		MPASS(n >= 1 && n <= available && n <= SGE_MAX_WR_NDESC);
2871 
2872 		available -= n;
2873 		dbdiff += n;
2874 		IDXINCR(eq->pidx, n, eq->sidx);
2875 
2876 		if (wr_can_update_eq(wr)) {
2877 			if (total_available_tx_desc(eq) < eq->sidx / 4 &&
2878 			    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2879 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2880 				    F_FW_WR_EQUEQ);
2881 				eq->equeqidx = eq->pidx;
2882 			} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >=
2883 			    32) {
2884 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
2885 				eq->equeqidx = eq->pidx;
2886 			}
2887 		}
2888 
2889 		if (dbdiff >= 16 && remaining >= 4) {
2890 			ring_eq_db(sc, eq, dbdiff);
2891 			available += reclaim_tx_descs(txq, 4 * dbdiff);
2892 			dbdiff = 0;
2893 		}
2894 
2895 		cidx = next_cidx;
2896 	}
2897 	if (dbdiff != 0) {
2898 		ring_eq_db(sc, eq, dbdiff);
2899 		reclaim_tx_descs(txq, 32);
2900 	}
2901 done:
2902 	TXQ_UNLOCK(txq);
2903 
2904 	return (total);
2905 }
2906 
2907 static inline void
2908 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
2909     int qsize)
2910 {
2911 
2912 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
2913 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
2914 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
2915 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
2916 
2917 	iq->flags = 0;
2918 	iq->adapter = sc;
2919 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
2920 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
2921 	if (pktc_idx >= 0) {
2922 		iq->intr_params |= F_QINTR_CNT_EN;
2923 		iq->intr_pktc_idx = pktc_idx;
2924 	}
2925 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
2926 	iq->sidx = iq->qsize - sc->params.sge.spg_len / IQ_ESIZE;
2927 }
2928 
2929 static inline void
2930 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name)
2931 {
2932 
2933 	fl->qsize = qsize;
2934 	fl->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2935 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
2936 	if (sc->flags & BUF_PACKING_OK &&
2937 	    ((!is_t4(sc) && buffer_packing) ||	/* T5+: enabled unless 0 */
2938 	    (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */
2939 		fl->flags |= FL_BUF_PACKING;
2940 	find_best_refill_source(sc, fl, maxp);
2941 	find_safe_refill_source(sc, fl);
2942 }
2943 
2944 static inline void
2945 init_eq(struct adapter *sc, struct sge_eq *eq, int eqtype, int qsize,
2946     uint8_t tx_chan, uint16_t iqid, char *name)
2947 {
2948 	KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
2949 
2950 	eq->flags = eqtype & EQ_TYPEMASK;
2951 	eq->tx_chan = tx_chan;
2952 	eq->iqid = iqid;
2953 	eq->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2954 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
2955 }
2956 
2957 static int
2958 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
2959     bus_dmamap_t *map, bus_addr_t *pa, void **va)
2960 {
2961 	int rc;
2962 
2963 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
2964 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
2965 	if (rc != 0) {
2966 		device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
2967 		goto done;
2968 	}
2969 
2970 	rc = bus_dmamem_alloc(*tag, va,
2971 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
2972 	if (rc != 0) {
2973 		device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
2974 		goto done;
2975 	}
2976 
2977 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
2978 	if (rc != 0) {
2979 		device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
2980 		goto done;
2981 	}
2982 done:
2983 	if (rc)
2984 		free_ring(sc, *tag, *map, *pa, *va);
2985 
2986 	return (rc);
2987 }
2988 
2989 static int
2990 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
2991     bus_addr_t pa, void *va)
2992 {
2993 	if (pa)
2994 		bus_dmamap_unload(tag, map);
2995 	if (va)
2996 		bus_dmamem_free(tag, va, map);
2997 	if (tag)
2998 		bus_dma_tag_destroy(tag);
2999 
3000 	return (0);
3001 }
3002 
3003 /*
3004  * Allocates the ring for an ingress queue and an optional freelist.  If the
3005  * freelist is specified it will be allocated and then associated with the
3006  * ingress queue.
3007  *
3008  * Returns errno on failure.  Resources allocated up to that point may still be
3009  * allocated.  Caller is responsible for cleanup in case this function fails.
3010  *
3011  * If the ingress queue will take interrupts directly then the intr_idx
3012  * specifies the vector, starting from 0.  -1 means the interrupts for this
3013  * queue should be forwarded to the fwq.
3014  */
3015 static int
3016 alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl,
3017     int intr_idx, int cong)
3018 {
3019 	int rc, i, cntxt_id;
3020 	size_t len;
3021 	struct fw_iq_cmd c;
3022 	struct port_info *pi = vi->pi;
3023 	struct adapter *sc = iq->adapter;
3024 	struct sge_params *sp = &sc->params.sge;
3025 	__be32 v = 0;
3026 
3027 	len = iq->qsize * IQ_ESIZE;
3028 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
3029 	    (void **)&iq->desc);
3030 	if (rc != 0)
3031 		return (rc);
3032 
3033 	bzero(&c, sizeof(c));
3034 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
3035 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
3036 	    V_FW_IQ_CMD_VFN(0));
3037 
3038 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
3039 	    FW_LEN16(c));
3040 
3041 	/* Special handling for firmware event queue */
3042 	if (iq == &sc->sge.fwq)
3043 		v |= F_FW_IQ_CMD_IQASYNCH;
3044 
3045 	if (intr_idx < 0) {
3046 		/* Forwarded interrupts, all headed to fwq */
3047 		v |= F_FW_IQ_CMD_IQANDST;
3048 		v |= V_FW_IQ_CMD_IQANDSTINDEX(sc->sge.fwq.cntxt_id);
3049 	} else {
3050 		KASSERT(intr_idx < sc->intr_count,
3051 		    ("%s: invalid direct intr_idx %d", __func__, intr_idx));
3052 		v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
3053 	}
3054 
3055 	c.type_to_iqandstindex = htobe32(v |
3056 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
3057 	    V_FW_IQ_CMD_VIID(vi->viid) |
3058 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
3059 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
3060 	    F_FW_IQ_CMD_IQGTSMODE |
3061 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
3062 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
3063 	c.iqsize = htobe16(iq->qsize);
3064 	c.iqaddr = htobe64(iq->ba);
3065 	if (cong >= 0)
3066 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
3067 
3068 	if (fl) {
3069 		mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
3070 
3071 		len = fl->qsize * EQ_ESIZE;
3072 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
3073 		    &fl->ba, (void **)&fl->desc);
3074 		if (rc)
3075 			return (rc);
3076 
3077 		/* Allocate space for one software descriptor per buffer. */
3078 		rc = alloc_fl_sdesc(fl);
3079 		if (rc != 0) {
3080 			device_printf(sc->dev,
3081 			    "failed to setup fl software descriptors: %d\n",
3082 			    rc);
3083 			return (rc);
3084 		}
3085 
3086 		if (fl->flags & FL_BUF_PACKING) {
3087 			fl->lowat = roundup2(sp->fl_starve_threshold2, 8);
3088 			fl->buf_boundary = sp->pack_boundary;
3089 		} else {
3090 			fl->lowat = roundup2(sp->fl_starve_threshold, 8);
3091 			fl->buf_boundary = 16;
3092 		}
3093 		if (fl_pad && fl->buf_boundary < sp->pad_boundary)
3094 			fl->buf_boundary = sp->pad_boundary;
3095 
3096 		c.iqns_to_fl0congen |=
3097 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
3098 			F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
3099 			(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
3100 			(fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
3101 			    0));
3102 		if (cong >= 0) {
3103 			c.iqns_to_fl0congen |=
3104 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
3105 				    F_FW_IQ_CMD_FL0CONGCIF |
3106 				    F_FW_IQ_CMD_FL0CONGEN);
3107 		}
3108 		c.fl0dcaen_to_fl0cidxfthresh =
3109 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
3110 			X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B) |
3111 			V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
3112 			X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
3113 		c.fl0size = htobe16(fl->qsize);
3114 		c.fl0addr = htobe64(fl->ba);
3115 	}
3116 
3117 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3118 	if (rc != 0) {
3119 		device_printf(sc->dev,
3120 		    "failed to create ingress queue: %d\n", rc);
3121 		return (rc);
3122 	}
3123 
3124 	iq->cidx = 0;
3125 	iq->gen = F_RSPD_GEN;
3126 	iq->intr_next = iq->intr_params;
3127 	iq->cntxt_id = be16toh(c.iqid);
3128 	iq->abs_id = be16toh(c.physiqid);
3129 	iq->flags |= IQ_ALLOCATED;
3130 
3131 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
3132 	if (cntxt_id >= sc->sge.niq) {
3133 		panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
3134 		    cntxt_id, sc->sge.niq - 1);
3135 	}
3136 	sc->sge.iqmap[cntxt_id] = iq;
3137 
3138 	if (fl) {
3139 		u_int qid;
3140 
3141 		iq->flags |= IQ_HAS_FL;
3142 		fl->cntxt_id = be16toh(c.fl0id);
3143 		fl->pidx = fl->cidx = 0;
3144 
3145 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
3146 		if (cntxt_id >= sc->sge.neq) {
3147 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
3148 			    __func__, cntxt_id, sc->sge.neq - 1);
3149 		}
3150 		sc->sge.eqmap[cntxt_id] = (void *)fl;
3151 
3152 		qid = fl->cntxt_id;
3153 		if (isset(&sc->doorbells, DOORBELL_UDB)) {
3154 			uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3155 			uint32_t mask = (1 << s_qpp) - 1;
3156 			volatile uint8_t *udb;
3157 
3158 			udb = sc->udbs_base + UDBS_DB_OFFSET;
3159 			udb += (qid >> s_qpp) << PAGE_SHIFT;
3160 			qid &= mask;
3161 			if (qid < PAGE_SIZE / UDBS_SEG_SIZE) {
3162 				udb += qid << UDBS_SEG_SHIFT;
3163 				qid = 0;
3164 			}
3165 			fl->udb = (volatile void *)udb;
3166 		}
3167 		fl->dbval = V_QID(qid) | sc->chip_params->sge_fl_db;
3168 
3169 		FL_LOCK(fl);
3170 		/* Enough to make sure the SGE doesn't think it's starved */
3171 		refill_fl(sc, fl, fl->lowat);
3172 		FL_UNLOCK(fl);
3173 	}
3174 
3175 	if (chip_id(sc) >= CHELSIO_T5 && !(sc->flags & IS_VF) && cong >= 0) {
3176 		uint32_t param, val;
3177 
3178 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
3179 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
3180 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
3181 		if (cong == 0)
3182 			val = 1 << 19;
3183 		else {
3184 			val = 2 << 19;
3185 			for (i = 0; i < 4; i++) {
3186 				if (cong & (1 << i))
3187 					val |= 1 << (i << 2);
3188 			}
3189 		}
3190 
3191 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
3192 		if (rc != 0) {
3193 			/* report error but carry on */
3194 			device_printf(sc->dev,
3195 			    "failed to set congestion manager context for "
3196 			    "ingress queue %d: %d\n", iq->cntxt_id, rc);
3197 		}
3198 	}
3199 
3200 	/* Enable IQ interrupts */
3201 	atomic_store_rel_int(&iq->state, IQS_IDLE);
3202 	t4_write_reg(sc, sc->sge_gts_reg, V_SEINTARM(iq->intr_params) |
3203 	    V_INGRESSQID(iq->cntxt_id));
3204 
3205 	return (0);
3206 }
3207 
3208 static int
3209 free_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl)
3210 {
3211 	int rc;
3212 	struct adapter *sc = iq->adapter;
3213 	device_t dev;
3214 
3215 	if (sc == NULL)
3216 		return (0);	/* nothing to do */
3217 
3218 	dev = vi ? vi->dev : sc->dev;
3219 
3220 	if (iq->flags & IQ_ALLOCATED) {
3221 		rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
3222 		    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
3223 		    fl ? fl->cntxt_id : 0xffff, 0xffff);
3224 		if (rc != 0) {
3225 			device_printf(dev,
3226 			    "failed to free queue %p: %d\n", iq, rc);
3227 			return (rc);
3228 		}
3229 		iq->flags &= ~IQ_ALLOCATED;
3230 	}
3231 
3232 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
3233 
3234 	bzero(iq, sizeof(*iq));
3235 
3236 	if (fl) {
3237 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
3238 		    fl->desc);
3239 
3240 		if (fl->sdesc)
3241 			free_fl_sdesc(sc, fl);
3242 
3243 		if (mtx_initialized(&fl->fl_lock))
3244 			mtx_destroy(&fl->fl_lock);
3245 
3246 		bzero(fl, sizeof(*fl));
3247 	}
3248 
3249 	return (0);
3250 }
3251 
3252 static void
3253 add_iq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
3254     struct sge_iq *iq)
3255 {
3256 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3257 
3258 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD, &iq->ba,
3259 	    "bus address of descriptor ring");
3260 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3261 	    iq->qsize * IQ_ESIZE, "descriptor ring size in bytes");
3262 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3263 	    CTLTYPE_INT | CTLFLAG_RD, &iq->abs_id, 0, sysctl_uint16, "I",
3264 	    "absolute id of the queue");
3265 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3266 	    CTLTYPE_INT | CTLFLAG_RD, &iq->cntxt_id, 0, sysctl_uint16, "I",
3267 	    "SGE context id of the queue");
3268 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3269 	    CTLTYPE_INT | CTLFLAG_RD, &iq->cidx, 0, sysctl_uint16, "I",
3270 	    "consumer index");
3271 }
3272 
3273 static void
3274 add_fl_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
3275     struct sysctl_oid *oid, struct sge_fl *fl)
3276 {
3277 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3278 
3279 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3280 	    "freelist");
3281 	children = SYSCTL_CHILDREN(oid);
3282 
3283 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3284 	    &fl->ba, "bus address of descriptor ring");
3285 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3286 	    fl->sidx * EQ_ESIZE + sc->params.sge.spg_len,
3287 	    "desc ring size in bytes");
3288 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3289 	    CTLTYPE_INT | CTLFLAG_RD, &fl->cntxt_id, 0, sysctl_uint16, "I",
3290 	    "SGE context id of the freelist");
3291 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL,
3292 	    fl_pad ? 1 : 0, "padding enabled");
3293 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL,
3294 	    fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled");
3295 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
3296 	    0, "consumer index");
3297 	if (fl->flags & FL_BUF_PACKING) {
3298 		SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
3299 		    CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
3300 	}
3301 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
3302 	    0, "producer index");
3303 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_allocated",
3304 	    CTLFLAG_RD, &fl->mbuf_allocated, "# of mbuf allocated");
3305 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_inlined",
3306 	    CTLFLAG_RD, &fl->mbuf_inlined, "# of mbuf inlined in clusters");
3307 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
3308 	    CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
3309 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
3310 	    CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
3311 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
3312 	    CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
3313 }
3314 
3315 static int
3316 alloc_fwq(struct adapter *sc)
3317 {
3318 	int rc, intr_idx;
3319 	struct sge_iq *fwq = &sc->sge.fwq;
3320 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
3321 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3322 
3323 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE);
3324 	if (sc->flags & IS_VF)
3325 		intr_idx = 0;
3326 	else
3327 		intr_idx = sc->intr_count > 1 ? 1 : 0;
3328 	rc = alloc_iq_fl(&sc->port[0]->vi[0], fwq, NULL, intr_idx, -1);
3329 	if (rc != 0) {
3330 		device_printf(sc->dev,
3331 		    "failed to create firmware event queue: %d\n", rc);
3332 		return (rc);
3333 	}
3334 
3335 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD,
3336 	    NULL, "firmware event queue");
3337 	add_iq_sysctls(&sc->ctx, oid, fwq);
3338 
3339 	return (0);
3340 }
3341 
3342 static int
3343 free_fwq(struct adapter *sc)
3344 {
3345 	return free_iq_fl(NULL, &sc->sge.fwq, NULL);
3346 }
3347 
3348 static int
3349 alloc_ctrlq(struct adapter *sc, struct sge_wrq *ctrlq, int idx,
3350     struct sysctl_oid *oid)
3351 {
3352 	int rc;
3353 	char name[16];
3354 	struct sysctl_oid_list *children;
3355 
3356 	snprintf(name, sizeof(name), "%s ctrlq%d", device_get_nameunit(sc->dev),
3357 	    idx);
3358 	init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[idx]->tx_chan,
3359 	    sc->sge.fwq.cntxt_id, name);
3360 
3361 	children = SYSCTL_CHILDREN(oid);
3362 	snprintf(name, sizeof(name), "%d", idx);
3363 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3364 	    NULL, "ctrl queue");
3365 	rc = alloc_wrq(sc, NULL, ctrlq, oid);
3366 
3367 	return (rc);
3368 }
3369 
3370 int
3371 tnl_cong(struct port_info *pi, int drop)
3372 {
3373 
3374 	if (drop == -1)
3375 		return (-1);
3376 	else if (drop == 1)
3377 		return (0);
3378 	else
3379 		return (pi->rx_e_chan_map);
3380 }
3381 
3382 static int
3383 alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int intr_idx, int idx,
3384     struct sysctl_oid *oid)
3385 {
3386 	int rc;
3387 	struct adapter *sc = vi->pi->adapter;
3388 	struct sysctl_oid_list *children;
3389 	char name[16];
3390 
3391 	rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, intr_idx,
3392 	    tnl_cong(vi->pi, cong_drop));
3393 	if (rc != 0)
3394 		return (rc);
3395 
3396 	if (idx == 0)
3397 		sc->sge.iq_base = rxq->iq.abs_id - rxq->iq.cntxt_id;
3398 	else
3399 		KASSERT(rxq->iq.cntxt_id + sc->sge.iq_base == rxq->iq.abs_id,
3400 		    ("iq_base mismatch"));
3401 	KASSERT(sc->sge.iq_base == 0 || sc->flags & IS_VF,
3402 	    ("PF with non-zero iq_base"));
3403 
3404 	/*
3405 	 * The freelist is just barely above the starvation threshold right now,
3406 	 * fill it up a bit more.
3407 	 */
3408 	FL_LOCK(&rxq->fl);
3409 	refill_fl(sc, &rxq->fl, 128);
3410 	FL_UNLOCK(&rxq->fl);
3411 
3412 #if defined(INET) || defined(INET6)
3413 	rc = tcp_lro_init_args(&rxq->lro, vi->ifp, lro_entries, lro_mbufs);
3414 	if (rc != 0)
3415 		return (rc);
3416 	MPASS(rxq->lro.ifp == vi->ifp);	/* also indicates LRO init'ed */
3417 
3418 	if (vi->ifp->if_capenable & IFCAP_LRO)
3419 		rxq->iq.flags |= IQ_LRO_ENABLED;
3420 #endif
3421 	rxq->ifp = vi->ifp;
3422 
3423 	children = SYSCTL_CHILDREN(oid);
3424 
3425 	snprintf(name, sizeof(name), "%d", idx);
3426 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3427 	    NULL, "rx queue");
3428 	children = SYSCTL_CHILDREN(oid);
3429 
3430 	add_iq_sysctls(&vi->ctx, oid, &rxq->iq);
3431 #if defined(INET) || defined(INET6)
3432 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
3433 	    &rxq->lro.lro_queued, 0, NULL);
3434 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
3435 	    &rxq->lro.lro_flushed, 0, NULL);
3436 #endif
3437 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
3438 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
3439 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_extraction",
3440 	    CTLFLAG_RD, &rxq->vlan_extraction,
3441 	    "# of times hardware extracted 802.1Q tag");
3442 
3443 	add_fl_sysctls(sc, &vi->ctx, oid, &rxq->fl);
3444 
3445 	return (rc);
3446 }
3447 
3448 static int
3449 free_rxq(struct vi_info *vi, struct sge_rxq *rxq)
3450 {
3451 	int rc;
3452 
3453 #if defined(INET) || defined(INET6)
3454 	if (rxq->lro.ifp) {
3455 		tcp_lro_free(&rxq->lro);
3456 		rxq->lro.ifp = NULL;
3457 	}
3458 #endif
3459 
3460 	rc = free_iq_fl(vi, &rxq->iq, &rxq->fl);
3461 	if (rc == 0)
3462 		bzero(rxq, sizeof(*rxq));
3463 
3464 	return (rc);
3465 }
3466 
3467 #ifdef TCP_OFFLOAD
3468 static int
3469 alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq,
3470     int intr_idx, int idx, struct sysctl_oid *oid)
3471 {
3472 	struct port_info *pi = vi->pi;
3473 	int rc;
3474 	struct sysctl_oid_list *children;
3475 	char name[16];
3476 
3477 	rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx, 0);
3478 	if (rc != 0)
3479 		return (rc);
3480 
3481 	children = SYSCTL_CHILDREN(oid);
3482 
3483 	snprintf(name, sizeof(name), "%d", idx);
3484 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3485 	    NULL, "rx queue");
3486 	add_iq_sysctls(&vi->ctx, oid, &ofld_rxq->iq);
3487 	add_fl_sysctls(pi->adapter, &vi->ctx, oid, &ofld_rxq->fl);
3488 
3489 	return (rc);
3490 }
3491 
3492 static int
3493 free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq)
3494 {
3495 	int rc;
3496 
3497 	rc = free_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl);
3498 	if (rc == 0)
3499 		bzero(ofld_rxq, sizeof(*ofld_rxq));
3500 
3501 	return (rc);
3502 }
3503 #endif
3504 
3505 #ifdef DEV_NETMAP
3506 static int
3507 alloc_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int intr_idx,
3508     int idx, struct sysctl_oid *oid)
3509 {
3510 	int rc;
3511 	struct sysctl_oid_list *children;
3512 	struct sysctl_ctx_list *ctx;
3513 	char name[16];
3514 	size_t len;
3515 	struct adapter *sc = vi->pi->adapter;
3516 	struct netmap_adapter *na = NA(vi->ifp);
3517 
3518 	MPASS(na != NULL);
3519 
3520 	len = vi->qsize_rxq * IQ_ESIZE;
3521 	rc = alloc_ring(sc, len, &nm_rxq->iq_desc_tag, &nm_rxq->iq_desc_map,
3522 	    &nm_rxq->iq_ba, (void **)&nm_rxq->iq_desc);
3523 	if (rc != 0)
3524 		return (rc);
3525 
3526 	len = na->num_rx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3527 	rc = alloc_ring(sc, len, &nm_rxq->fl_desc_tag, &nm_rxq->fl_desc_map,
3528 	    &nm_rxq->fl_ba, (void **)&nm_rxq->fl_desc);
3529 	if (rc != 0)
3530 		return (rc);
3531 
3532 	nm_rxq->vi = vi;
3533 	nm_rxq->nid = idx;
3534 	nm_rxq->iq_cidx = 0;
3535 	nm_rxq->iq_sidx = vi->qsize_rxq - sc->params.sge.spg_len / IQ_ESIZE;
3536 	nm_rxq->iq_gen = F_RSPD_GEN;
3537 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
3538 	nm_rxq->fl_sidx = na->num_rx_desc;
3539 	nm_rxq->intr_idx = intr_idx;
3540 	nm_rxq->iq_cntxt_id = INVALID_NM_RXQ_CNTXT_ID;
3541 
3542 	ctx = &vi->ctx;
3543 	children = SYSCTL_CHILDREN(oid);
3544 
3545 	snprintf(name, sizeof(name), "%d", idx);
3546 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL,
3547 	    "rx queue");
3548 	children = SYSCTL_CHILDREN(oid);
3549 
3550 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3551 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_abs_id, 0, sysctl_uint16,
3552 	    "I", "absolute id of the queue");
3553 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3554 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cntxt_id, 0, sysctl_uint16,
3555 	    "I", "SGE context id of the queue");
3556 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3557 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cidx, 0, sysctl_uint16, "I",
3558 	    "consumer index");
3559 
3560 	children = SYSCTL_CHILDREN(oid);
3561 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3562 	    "freelist");
3563 	children = SYSCTL_CHILDREN(oid);
3564 
3565 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3566 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->fl_cntxt_id, 0, sysctl_uint16,
3567 	    "I", "SGE context id of the freelist");
3568 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
3569 	    &nm_rxq->fl_cidx, 0, "consumer index");
3570 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
3571 	    &nm_rxq->fl_pidx, 0, "producer index");
3572 
3573 	return (rc);
3574 }
3575 
3576 
3577 static int
3578 free_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
3579 {
3580 	struct adapter *sc = vi->pi->adapter;
3581 
3582 	if (vi->flags & VI_INIT_DONE)
3583 		MPASS(nm_rxq->iq_cntxt_id == INVALID_NM_RXQ_CNTXT_ID);
3584 	else
3585 		MPASS(nm_rxq->iq_cntxt_id == 0);
3586 
3587 	free_ring(sc, nm_rxq->iq_desc_tag, nm_rxq->iq_desc_map, nm_rxq->iq_ba,
3588 	    nm_rxq->iq_desc);
3589 	free_ring(sc, nm_rxq->fl_desc_tag, nm_rxq->fl_desc_map, nm_rxq->fl_ba,
3590 	    nm_rxq->fl_desc);
3591 
3592 	return (0);
3593 }
3594 
3595 static int
3596 alloc_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq, int iqidx, int idx,
3597     struct sysctl_oid *oid)
3598 {
3599 	int rc;
3600 	size_t len;
3601 	struct port_info *pi = vi->pi;
3602 	struct adapter *sc = pi->adapter;
3603 	struct netmap_adapter *na = NA(vi->ifp);
3604 	char name[16];
3605 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3606 
3607 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3608 	rc = alloc_ring(sc, len, &nm_txq->desc_tag, &nm_txq->desc_map,
3609 	    &nm_txq->ba, (void **)&nm_txq->desc);
3610 	if (rc)
3611 		return (rc);
3612 
3613 	nm_txq->pidx = nm_txq->cidx = 0;
3614 	nm_txq->sidx = na->num_tx_desc;
3615 	nm_txq->nid = idx;
3616 	nm_txq->iqidx = iqidx;
3617 	nm_txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3618 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(G_FW_VIID_PFN(vi->viid)) |
3619 	    V_TXPKT_VF(G_FW_VIID_VIN(vi->viid)) |
3620 	    V_TXPKT_VF_VLD(G_FW_VIID_VIVLD(vi->viid)));
3621 	nm_txq->cntxt_id = INVALID_NM_TXQ_CNTXT_ID;
3622 
3623 	snprintf(name, sizeof(name), "%d", idx);
3624 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3625 	    NULL, "netmap tx queue");
3626 	children = SYSCTL_CHILDREN(oid);
3627 
3628 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3629 	    &nm_txq->cntxt_id, 0, "SGE context id of the queue");
3630 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3631 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->cidx, 0, sysctl_uint16, "I",
3632 	    "consumer index");
3633 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3634 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->pidx, 0, sysctl_uint16, "I",
3635 	    "producer index");
3636 
3637 	return (rc);
3638 }
3639 
3640 static int
3641 free_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
3642 {
3643 	struct adapter *sc = vi->pi->adapter;
3644 
3645 	if (vi->flags & VI_INIT_DONE)
3646 		MPASS(nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID);
3647 	else
3648 		MPASS(nm_txq->cntxt_id == 0);
3649 
3650 	free_ring(sc, nm_txq->desc_tag, nm_txq->desc_map, nm_txq->ba,
3651 	    nm_txq->desc);
3652 
3653 	return (0);
3654 }
3655 #endif
3656 
3657 /*
3658  * Returns a reasonable automatic cidx flush threshold for a given queue size.
3659  */
3660 static u_int
3661 qsize_to_fthresh(int qsize)
3662 {
3663 	u_int fthresh;
3664 
3665 	while (!powerof2(qsize))
3666 		qsize++;
3667 	fthresh = ilog2(qsize);
3668 	if (fthresh > X_CIDXFLUSHTHRESH_128)
3669 		fthresh = X_CIDXFLUSHTHRESH_128;
3670 
3671 	return (fthresh);
3672 }
3673 
3674 static int
3675 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
3676 {
3677 	int rc, cntxt_id;
3678 	struct fw_eq_ctrl_cmd c;
3679 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3680 
3681 	bzero(&c, sizeof(c));
3682 
3683 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
3684 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
3685 	    V_FW_EQ_CTRL_CMD_VFN(0));
3686 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
3687 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
3688 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid));
3689 	c.physeqid_pkd = htobe32(0);
3690 	c.fetchszm_to_iqid =
3691 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
3692 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
3693 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
3694 	c.dcaen_to_eqsize =
3695 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3696 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3697 		V_FW_EQ_CTRL_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
3698 		V_FW_EQ_CTRL_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(sc->dev,
3704 		    "failed to create control queue %d: %d\n", eq->tx_chan, rc);
3705 		return (rc);
3706 	}
3707 	eq->flags |= EQ_ALLOCATED;
3708 
3709 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
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 
3719 static int
3720 eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3721 {
3722 	int rc, cntxt_id;
3723 	struct fw_eq_eth_cmd c;
3724 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3725 
3726 	bzero(&c, sizeof(c));
3727 
3728 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
3729 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
3730 	    V_FW_EQ_ETH_CMD_VFN(0));
3731 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
3732 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
3733 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
3734 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
3735 	c.fetchszm_to_iqid =
3736 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3737 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
3738 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
3739 	c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3740 	    V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3741 	    V_FW_EQ_ETH_CMD_EQSIZE(qsize));
3742 	c.eqaddr = htobe64(eq->ba);
3743 
3744 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3745 	if (rc != 0) {
3746 		device_printf(vi->dev,
3747 		    "failed to create Ethernet egress queue: %d\n", rc);
3748 		return (rc);
3749 	}
3750 	eq->flags |= EQ_ALLOCATED;
3751 
3752 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
3753 	eq->abs_id = G_FW_EQ_ETH_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
3754 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3755 	if (cntxt_id >= sc->sge.neq)
3756 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3757 		cntxt_id, sc->sge.neq - 1);
3758 	sc->sge.eqmap[cntxt_id] = eq;
3759 
3760 	return (rc);
3761 }
3762 
3763 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3764 static int
3765 ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3766 {
3767 	int rc, cntxt_id;
3768 	struct fw_eq_ofld_cmd c;
3769 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3770 
3771 	bzero(&c, sizeof(c));
3772 
3773 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
3774 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
3775 	    V_FW_EQ_OFLD_CMD_VFN(0));
3776 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
3777 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
3778 	c.fetchszm_to_iqid =
3779 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
3780 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
3781 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
3782 	c.dcaen_to_eqsize =
3783 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3784 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3785 		V_FW_EQ_OFLD_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
3786 		V_FW_EQ_OFLD_CMD_EQSIZE(qsize));
3787 	c.eqaddr = htobe64(eq->ba);
3788 
3789 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3790 	if (rc != 0) {
3791 		device_printf(vi->dev,
3792 		    "failed to create egress queue for TCP offload: %d\n", rc);
3793 		return (rc);
3794 	}
3795 	eq->flags |= EQ_ALLOCATED;
3796 
3797 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
3798 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3799 	if (cntxt_id >= sc->sge.neq)
3800 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3801 		cntxt_id, sc->sge.neq - 1);
3802 	sc->sge.eqmap[cntxt_id] = eq;
3803 
3804 	return (rc);
3805 }
3806 #endif
3807 
3808 static int
3809 alloc_eq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3810 {
3811 	int rc, qsize;
3812 	size_t len;
3813 
3814 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
3815 
3816 	qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3817 	len = qsize * EQ_ESIZE;
3818 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
3819 	    &eq->ba, (void **)&eq->desc);
3820 	if (rc)
3821 		return (rc);
3822 
3823 	eq->pidx = eq->cidx = eq->dbidx = 0;
3824 	/* Note that equeqidx is not used with sge_wrq (OFLD/CTRL) queues. */
3825 	eq->equeqidx = 0;
3826 	eq->doorbells = sc->doorbells;
3827 
3828 	switch (eq->flags & EQ_TYPEMASK) {
3829 	case EQ_CTRL:
3830 		rc = ctrl_eq_alloc(sc, eq);
3831 		break;
3832 
3833 	case EQ_ETH:
3834 		rc = eth_eq_alloc(sc, vi, eq);
3835 		break;
3836 
3837 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3838 	case EQ_OFLD:
3839 		rc = ofld_eq_alloc(sc, vi, eq);
3840 		break;
3841 #endif
3842 
3843 	default:
3844 		panic("%s: invalid eq type %d.", __func__,
3845 		    eq->flags & EQ_TYPEMASK);
3846 	}
3847 	if (rc != 0) {
3848 		device_printf(sc->dev,
3849 		    "failed to allocate egress queue(%d): %d\n",
3850 		    eq->flags & EQ_TYPEMASK, rc);
3851 	}
3852 
3853 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
3854 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
3855 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
3856 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3857 		uint32_t mask = (1 << s_qpp) - 1;
3858 		volatile uint8_t *udb;
3859 
3860 		udb = sc->udbs_base + UDBS_DB_OFFSET;
3861 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
3862 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
3863 		if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
3864 	    		clrbit(&eq->doorbells, DOORBELL_WCWR);
3865 		else {
3866 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
3867 			eq->udb_qid = 0;
3868 		}
3869 		eq->udb = (volatile void *)udb;
3870 	}
3871 
3872 	return (rc);
3873 }
3874 
3875 static int
3876 free_eq(struct adapter *sc, struct sge_eq *eq)
3877 {
3878 	int rc;
3879 
3880 	if (eq->flags & EQ_ALLOCATED) {
3881 		switch (eq->flags & EQ_TYPEMASK) {
3882 		case EQ_CTRL:
3883 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
3884 			    eq->cntxt_id);
3885 			break;
3886 
3887 		case EQ_ETH:
3888 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
3889 			    eq->cntxt_id);
3890 			break;
3891 
3892 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3893 		case EQ_OFLD:
3894 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
3895 			    eq->cntxt_id);
3896 			break;
3897 #endif
3898 
3899 		default:
3900 			panic("%s: invalid eq type %d.", __func__,
3901 			    eq->flags & EQ_TYPEMASK);
3902 		}
3903 		if (rc != 0) {
3904 			device_printf(sc->dev,
3905 			    "failed to free egress queue (%d): %d\n",
3906 			    eq->flags & EQ_TYPEMASK, rc);
3907 			return (rc);
3908 		}
3909 		eq->flags &= ~EQ_ALLOCATED;
3910 	}
3911 
3912 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
3913 
3914 	if (mtx_initialized(&eq->eq_lock))
3915 		mtx_destroy(&eq->eq_lock);
3916 
3917 	bzero(eq, sizeof(*eq));
3918 	return (0);
3919 }
3920 
3921 static int
3922 alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq,
3923     struct sysctl_oid *oid)
3924 {
3925 	int rc;
3926 	struct sysctl_ctx_list *ctx = vi ? &vi->ctx : &sc->ctx;
3927 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3928 
3929 	rc = alloc_eq(sc, vi, &wrq->eq);
3930 	if (rc)
3931 		return (rc);
3932 
3933 	wrq->adapter = sc;
3934 	TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq);
3935 	TAILQ_INIT(&wrq->incomplete_wrs);
3936 	STAILQ_INIT(&wrq->wr_list);
3937 	wrq->nwr_pending = 0;
3938 	wrq->ndesc_needed = 0;
3939 
3940 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3941 	    &wrq->eq.ba, "bus address of descriptor ring");
3942 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3943 	    wrq->eq.sidx * EQ_ESIZE + sc->params.sge.spg_len,
3944 	    "desc ring size in bytes");
3945 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3946 	    &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
3947 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3948 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I",
3949 	    "consumer index");
3950 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx",
3951 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I",
3952 	    "producer index");
3953 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
3954 	    wrq->eq.sidx, "status page index");
3955 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD,
3956 	    &wrq->tx_wrs_direct, "# of work requests (direct)");
3957 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD,
3958 	    &wrq->tx_wrs_copied, "# of work requests (copied)");
3959 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_sspace", CTLFLAG_RD,
3960 	    &wrq->tx_wrs_ss, "# of work requests (copied from scratch space)");
3961 
3962 	return (rc);
3963 }
3964 
3965 static int
3966 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
3967 {
3968 	int rc;
3969 
3970 	rc = free_eq(sc, &wrq->eq);
3971 	if (rc)
3972 		return (rc);
3973 
3974 	bzero(wrq, sizeof(*wrq));
3975 	return (0);
3976 }
3977 
3978 static int
3979 alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx,
3980     struct sysctl_oid *oid)
3981 {
3982 	int rc;
3983 	struct port_info *pi = vi->pi;
3984 	struct adapter *sc = pi->adapter;
3985 	struct sge_eq *eq = &txq->eq;
3986 	char name[16];
3987 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3988 
3989 	rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx, can_resume_eth_tx,
3990 	    M_CXGBE, M_WAITOK);
3991 	if (rc != 0) {
3992 		device_printf(sc->dev, "failed to allocate mp_ring: %d\n", rc);
3993 		return (rc);
3994 	}
3995 
3996 	rc = alloc_eq(sc, vi, eq);
3997 	if (rc != 0) {
3998 		mp_ring_free(txq->r);
3999 		txq->r = NULL;
4000 		return (rc);
4001 	}
4002 
4003 	/* Can't fail after this point. */
4004 
4005 	if (idx == 0)
4006 		sc->sge.eq_base = eq->abs_id - eq->cntxt_id;
4007 	else
4008 		KASSERT(eq->cntxt_id + sc->sge.eq_base == eq->abs_id,
4009 		    ("eq_base mismatch"));
4010 	KASSERT(sc->sge.eq_base == 0 || sc->flags & IS_VF,
4011 	    ("PF with non-zero eq_base"));
4012 
4013 	TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq);
4014 	txq->ifp = vi->ifp;
4015 	txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK);
4016 	if (sc->flags & IS_VF)
4017 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
4018 		    V_TXPKT_INTF(pi->tx_chan));
4019 	else
4020 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
4021 		    V_TXPKT_INTF(pi->tx_chan) |
4022 		    V_TXPKT_PF(G_FW_VIID_PFN(vi->viid)) |
4023 		    V_TXPKT_VF(G_FW_VIID_VIN(vi->viid)) |
4024 		    V_TXPKT_VF_VLD(G_FW_VIID_VIVLD(vi->viid)));
4025 	txq->tc_idx = -1;
4026 	txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE,
4027 	    M_ZERO | M_WAITOK);
4028 
4029 	snprintf(name, sizeof(name), "%d", idx);
4030 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
4031 	    NULL, "tx queue");
4032 	children = SYSCTL_CHILDREN(oid);
4033 
4034 	SYSCTL_ADD_UAUTO(&vi->ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
4035 	    &eq->ba, "bus address of descriptor ring");
4036 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
4037 	    eq->sidx * EQ_ESIZE + sc->params.sge.spg_len,
4038 	    "desc ring size in bytes");
4039 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
4040 	    &eq->abs_id, 0, "absolute id of the queue");
4041 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
4042 	    &eq->cntxt_id, 0, "SGE context id of the queue");
4043 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
4044 	    CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I",
4045 	    "consumer index");
4046 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
4047 	    CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I",
4048 	    "producer index");
4049 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
4050 	    eq->sidx, "status page index");
4051 
4052 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "tc",
4053 	    CTLTYPE_INT | CTLFLAG_RW, vi, idx, sysctl_tc, "I",
4054 	    "traffic class (-1 means none)");
4055 
4056 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
4057 	    &txq->txcsum, "# of times hardware assisted with checksum");
4058 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_insertion",
4059 	    CTLFLAG_RD, &txq->vlan_insertion,
4060 	    "# of times hardware inserted 802.1Q tag");
4061 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
4062 	    &txq->tso_wrs, "# of TSO work requests");
4063 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
4064 	    &txq->imm_wrs, "# of work requests with immediate data");
4065 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
4066 	    &txq->sgl_wrs, "# of work requests with direct SGL");
4067 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
4068 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
4069 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_wrs",
4070 	    CTLFLAG_RD, &txq->txpkts0_wrs,
4071 	    "# of txpkts (type 0) work requests");
4072 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_wrs",
4073 	    CTLFLAG_RD, &txq->txpkts1_wrs,
4074 	    "# of txpkts (type 1) work requests");
4075 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_pkts",
4076 	    CTLFLAG_RD, &txq->txpkts0_pkts,
4077 	    "# of frames tx'd using type0 txpkts work requests");
4078 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_pkts",
4079 	    CTLFLAG_RD, &txq->txpkts1_pkts,
4080 	    "# of frames tx'd using type1 txpkts work requests");
4081 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "raw_wrs", CTLFLAG_RD,
4082 	    &txq->raw_wrs, "# of raw work requests (non-packets)");
4083 
4084 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_enqueues",
4085 	    CTLFLAG_RD, &txq->r->enqueues,
4086 	    "# of enqueues to the mp_ring for this queue");
4087 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_drops",
4088 	    CTLFLAG_RD, &txq->r->drops,
4089 	    "# of drops in the mp_ring for this queue");
4090 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_starts",
4091 	    CTLFLAG_RD, &txq->r->starts,
4092 	    "# of normal consumer starts in the mp_ring for this queue");
4093 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_stalls",
4094 	    CTLFLAG_RD, &txq->r->stalls,
4095 	    "# of consumer stalls in the mp_ring for this queue");
4096 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_restarts",
4097 	    CTLFLAG_RD, &txq->r->restarts,
4098 	    "# of consumer restarts in the mp_ring for this queue");
4099 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_abdications",
4100 	    CTLFLAG_RD, &txq->r->abdications,
4101 	    "# of consumer abdications in the mp_ring for this queue");
4102 
4103 	return (0);
4104 }
4105 
4106 static int
4107 free_txq(struct vi_info *vi, struct sge_txq *txq)
4108 {
4109 	int rc;
4110 	struct adapter *sc = vi->pi->adapter;
4111 	struct sge_eq *eq = &txq->eq;
4112 
4113 	rc = free_eq(sc, eq);
4114 	if (rc)
4115 		return (rc);
4116 
4117 	sglist_free(txq->gl);
4118 	free(txq->sdesc, M_CXGBE);
4119 	mp_ring_free(txq->r);
4120 
4121 	bzero(txq, sizeof(*txq));
4122 	return (0);
4123 }
4124 
4125 static void
4126 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
4127 {
4128 	bus_addr_t *ba = arg;
4129 
4130 	KASSERT(nseg == 1,
4131 	    ("%s meant for single segment mappings only.", __func__));
4132 
4133 	*ba = error ? 0 : segs->ds_addr;
4134 }
4135 
4136 static inline void
4137 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
4138 {
4139 	uint32_t n, v;
4140 
4141 	n = IDXDIFF(fl->pidx / 8, fl->dbidx, fl->sidx);
4142 	MPASS(n > 0);
4143 
4144 	wmb();
4145 	v = fl->dbval | V_PIDX(n);
4146 	if (fl->udb)
4147 		*fl->udb = htole32(v);
4148 	else
4149 		t4_write_reg(sc, sc->sge_kdoorbell_reg, v);
4150 	IDXINCR(fl->dbidx, n, fl->sidx);
4151 }
4152 
4153 /*
4154  * Fills up the freelist by allocating up to 'n' buffers.  Buffers that are
4155  * recycled do not count towards this allocation budget.
4156  *
4157  * Returns non-zero to indicate that this freelist should be added to the list
4158  * of starving freelists.
4159  */
4160 static int
4161 refill_fl(struct adapter *sc, struct sge_fl *fl, int n)
4162 {
4163 	__be64 *d;
4164 	struct fl_sdesc *sd;
4165 	uintptr_t pa;
4166 	caddr_t cl;
4167 	struct cluster_layout *cll;
4168 	struct sw_zone_info *swz;
4169 	struct cluster_metadata *clm;
4170 	uint16_t max_pidx;
4171 	uint16_t hw_cidx = fl->hw_cidx;		/* stable snapshot */
4172 
4173 	FL_LOCK_ASSERT_OWNED(fl);
4174 
4175 	/*
4176 	 * We always stop at the beginning of the hardware descriptor that's just
4177 	 * before the one with the hw cidx.  This is to avoid hw pidx = hw cidx,
4178 	 * which would mean an empty freelist to the chip.
4179 	 */
4180 	max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1;
4181 	if (fl->pidx == max_pidx * 8)
4182 		return (0);
4183 
4184 	d = &fl->desc[fl->pidx];
4185 	sd = &fl->sdesc[fl->pidx];
4186 	cll = &fl->cll_def;	/* default layout */
4187 	swz = &sc->sge.sw_zone_info[cll->zidx];
4188 
4189 	while (n > 0) {
4190 
4191 		if (sd->cl != NULL) {
4192 
4193 			if (sd->nmbuf == 0) {
4194 				/*
4195 				 * Fast recycle without involving any atomics on
4196 				 * the cluster's metadata (if the cluster has
4197 				 * metadata).  This happens when all frames
4198 				 * received in the cluster were small enough to
4199 				 * fit within a single mbuf each.
4200 				 */
4201 				fl->cl_fast_recycled++;
4202 #ifdef INVARIANTS
4203 				clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
4204 				if (clm != NULL)
4205 					MPASS(clm->refcount == 1);
4206 #endif
4207 				goto recycled_fast;
4208 			}
4209 
4210 			/*
4211 			 * Cluster is guaranteed to have metadata.  Clusters
4212 			 * without metadata always take the fast recycle path
4213 			 * when they're recycled.
4214 			 */
4215 			clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
4216 			MPASS(clm != NULL);
4217 
4218 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4219 				fl->cl_recycled++;
4220 				counter_u64_add(extfree_rels, 1);
4221 				goto recycled;
4222 			}
4223 			sd->cl = NULL;	/* gave up my reference */
4224 		}
4225 		MPASS(sd->cl == NULL);
4226 alloc:
4227 		cl = uma_zalloc(swz->zone, M_NOWAIT);
4228 		if (__predict_false(cl == NULL)) {
4229 			if (cll == &fl->cll_alt || fl->cll_alt.zidx == -1 ||
4230 			    fl->cll_def.zidx == fl->cll_alt.zidx)
4231 				break;
4232 
4233 			/* fall back to the safe zone */
4234 			cll = &fl->cll_alt;
4235 			swz = &sc->sge.sw_zone_info[cll->zidx];
4236 			goto alloc;
4237 		}
4238 		fl->cl_allocated++;
4239 		n--;
4240 
4241 		pa = pmap_kextract((vm_offset_t)cl);
4242 		pa += cll->region1;
4243 		sd->cl = cl;
4244 		sd->cll = *cll;
4245 		*d = htobe64(pa | cll->hwidx);
4246 		clm = cl_metadata(sc, fl, cll, cl);
4247 		if (clm != NULL) {
4248 recycled:
4249 #ifdef INVARIANTS
4250 			clm->sd = sd;
4251 #endif
4252 			clm->refcount = 1;
4253 		}
4254 		sd->nmbuf = 0;
4255 recycled_fast:
4256 		d++;
4257 		sd++;
4258 		if (__predict_false(++fl->pidx % 8 == 0)) {
4259 			uint16_t pidx = fl->pidx / 8;
4260 
4261 			if (__predict_false(pidx == fl->sidx)) {
4262 				fl->pidx = 0;
4263 				pidx = 0;
4264 				sd = fl->sdesc;
4265 				d = fl->desc;
4266 			}
4267 			if (pidx == max_pidx)
4268 				break;
4269 
4270 			if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4)
4271 				ring_fl_db(sc, fl);
4272 		}
4273 	}
4274 
4275 	if (fl->pidx / 8 != fl->dbidx)
4276 		ring_fl_db(sc, fl);
4277 
4278 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
4279 }
4280 
4281 /*
4282  * Attempt to refill all starving freelists.
4283  */
4284 static void
4285 refill_sfl(void *arg)
4286 {
4287 	struct adapter *sc = arg;
4288 	struct sge_fl *fl, *fl_temp;
4289 
4290 	mtx_assert(&sc->sfl_lock, MA_OWNED);
4291 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
4292 		FL_LOCK(fl);
4293 		refill_fl(sc, fl, 64);
4294 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
4295 			TAILQ_REMOVE(&sc->sfl, fl, link);
4296 			fl->flags &= ~FL_STARVING;
4297 		}
4298 		FL_UNLOCK(fl);
4299 	}
4300 
4301 	if (!TAILQ_EMPTY(&sc->sfl))
4302 		callout_schedule(&sc->sfl_callout, hz / 5);
4303 }
4304 
4305 static int
4306 alloc_fl_sdesc(struct sge_fl *fl)
4307 {
4308 
4309 	fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc), M_CXGBE,
4310 	    M_ZERO | M_WAITOK);
4311 
4312 	return (0);
4313 }
4314 
4315 static void
4316 free_fl_sdesc(struct adapter *sc, struct sge_fl *fl)
4317 {
4318 	struct fl_sdesc *sd;
4319 	struct cluster_metadata *clm;
4320 	struct cluster_layout *cll;
4321 	int i;
4322 
4323 	sd = fl->sdesc;
4324 	for (i = 0; i < fl->sidx * 8; i++, sd++) {
4325 		if (sd->cl == NULL)
4326 			continue;
4327 
4328 		cll = &sd->cll;
4329 		clm = cl_metadata(sc, fl, cll, sd->cl);
4330 		if (sd->nmbuf == 0)
4331 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4332 		else if (clm && atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4333 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4334 			counter_u64_add(extfree_rels, 1);
4335 		}
4336 		sd->cl = NULL;
4337 	}
4338 
4339 	free(fl->sdesc, M_CXGBE);
4340 	fl->sdesc = NULL;
4341 }
4342 
4343 static inline void
4344 get_pkt_gl(struct mbuf *m, struct sglist *gl)
4345 {
4346 	int rc;
4347 
4348 	M_ASSERTPKTHDR(m);
4349 
4350 	sglist_reset(gl);
4351 	rc = sglist_append_mbuf(gl, m);
4352 	if (__predict_false(rc != 0)) {
4353 		panic("%s: mbuf %p (%d segs) was vetted earlier but now fails "
4354 		    "with %d.", __func__, m, mbuf_nsegs(m), rc);
4355 	}
4356 
4357 	KASSERT(gl->sg_nseg == mbuf_nsegs(m),
4358 	    ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m,
4359 	    mbuf_nsegs(m), gl->sg_nseg));
4360 	KASSERT(gl->sg_nseg > 0 &&
4361 	    gl->sg_nseg <= (needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS),
4362 	    ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__,
4363 		gl->sg_nseg, needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS));
4364 }
4365 
4366 /*
4367  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
4368  */
4369 static inline u_int
4370 txpkt_len16(u_int nsegs, u_int tso)
4371 {
4372 	u_int n;
4373 
4374 	MPASS(nsegs > 0);
4375 
4376 	nsegs--; /* first segment is part of ulptx_sgl */
4377 	n = sizeof(struct fw_eth_tx_pkt_wr) + sizeof(struct cpl_tx_pkt_core) +
4378 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4379 	if (tso)
4380 		n += sizeof(struct cpl_tx_pkt_lso_core);
4381 
4382 	return (howmany(n, 16));
4383 }
4384 
4385 /*
4386  * len16 for a txpkt_vm WR with a GL.  Includes the firmware work
4387  * request header.
4388  */
4389 static inline u_int
4390 txpkt_vm_len16(u_int nsegs, u_int tso)
4391 {
4392 	u_int n;
4393 
4394 	MPASS(nsegs > 0);
4395 
4396 	nsegs--; /* first segment is part of ulptx_sgl */
4397 	n = sizeof(struct fw_eth_tx_pkt_vm_wr) +
4398 	    sizeof(struct cpl_tx_pkt_core) +
4399 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4400 	if (tso)
4401 		n += sizeof(struct cpl_tx_pkt_lso_core);
4402 
4403 	return (howmany(n, 16));
4404 }
4405 
4406 /*
4407  * len16 for a txpkts type 0 WR with a GL.  Does not include the firmware work
4408  * request header.
4409  */
4410 static inline u_int
4411 txpkts0_len16(u_int nsegs)
4412 {
4413 	u_int n;
4414 
4415 	MPASS(nsegs > 0);
4416 
4417 	nsegs--; /* first segment is part of ulptx_sgl */
4418 	n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) +
4419 	    sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) +
4420 	    8 * ((3 * nsegs) / 2 + (nsegs & 1));
4421 
4422 	return (howmany(n, 16));
4423 }
4424 
4425 /*
4426  * len16 for a txpkts type 1 WR with a GL.  Does not include the firmware work
4427  * request header.
4428  */
4429 static inline u_int
4430 txpkts1_len16(void)
4431 {
4432 	u_int n;
4433 
4434 	n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl);
4435 
4436 	return (howmany(n, 16));
4437 }
4438 
4439 static inline u_int
4440 imm_payload(u_int ndesc)
4441 {
4442 	u_int n;
4443 
4444 	n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) -
4445 	    sizeof(struct cpl_tx_pkt_core);
4446 
4447 	return (n);
4448 }
4449 
4450 /*
4451  * Write a VM txpkt WR for this packet to the hardware descriptors, update the
4452  * software descriptor, and advance the pidx.  It is guaranteed that enough
4453  * descriptors are available.
4454  *
4455  * The return value is the # of hardware descriptors used.
4456  */
4457 static u_int
4458 write_txpkt_vm_wr(struct adapter *sc, struct sge_txq *txq,
4459     struct fw_eth_tx_pkt_vm_wr *wr, struct mbuf *m0, u_int available)
4460 {
4461 	struct sge_eq *eq = &txq->eq;
4462 	struct tx_sdesc *txsd;
4463 	struct cpl_tx_pkt_core *cpl;
4464 	uint32_t ctrl;	/* used in many unrelated places */
4465 	uint64_t ctrl1;
4466 	int csum_type, len16, ndesc, pktlen, nsegs;
4467 	caddr_t dst;
4468 
4469 	TXQ_LOCK_ASSERT_OWNED(txq);
4470 	M_ASSERTPKTHDR(m0);
4471 	MPASS(available > 0 && available < eq->sidx);
4472 
4473 	len16 = mbuf_len16(m0);
4474 	nsegs = mbuf_nsegs(m0);
4475 	pktlen = m0->m_pkthdr.len;
4476 	ctrl = sizeof(struct cpl_tx_pkt_core);
4477 	if (needs_tso(m0))
4478 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4479 	ndesc = howmany(len16, EQ_ESIZE / 16);
4480 	MPASS(ndesc <= available);
4481 
4482 	/* Firmware work request header */
4483 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4484 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_VM_WR) |
4485 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4486 
4487 	ctrl = V_FW_WR_LEN16(len16);
4488 	wr->equiq_to_len16 = htobe32(ctrl);
4489 	wr->r3[0] = 0;
4490 	wr->r3[1] = 0;
4491 
4492 	/*
4493 	 * Copy over ethmacdst, ethmacsrc, ethtype, and vlantci.
4494 	 * vlantci is ignored unless the ethtype is 0x8100, so it's
4495 	 * simpler to always copy it rather than making it
4496 	 * conditional.  Also, it seems that we do not have to set
4497 	 * vlantci or fake the ethtype when doing VLAN tag insertion.
4498 	 */
4499 	m_copydata(m0, 0, sizeof(struct ether_header) + 2, wr->ethmacdst);
4500 
4501 	csum_type = -1;
4502 	if (needs_tso(m0)) {
4503 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4504 
4505 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4506 		    m0->m_pkthdr.l4hlen > 0,
4507 		    ("%s: mbuf %p needs TSO but missing header lengths",
4508 			__func__, m0));
4509 
4510 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4511 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4512 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4513 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4514 			ctrl |= V_LSO_ETHHDR_LEN(1);
4515 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4516 			ctrl |= F_LSO_IPV6;
4517 
4518 		lso->lso_ctrl = htobe32(ctrl);
4519 		lso->ipid_ofst = htobe16(0);
4520 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4521 		lso->seqno_offset = htobe32(0);
4522 		lso->len = htobe32(pktlen);
4523 
4524 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4525 			csum_type = TX_CSUM_TCPIP6;
4526 		else
4527 			csum_type = TX_CSUM_TCPIP;
4528 
4529 		cpl = (void *)(lso + 1);
4530 
4531 		txq->tso_wrs++;
4532 	} else {
4533 		if (m0->m_pkthdr.csum_flags & CSUM_IP_TCP)
4534 			csum_type = TX_CSUM_TCPIP;
4535 		else if (m0->m_pkthdr.csum_flags & CSUM_IP_UDP)
4536 			csum_type = TX_CSUM_UDPIP;
4537 		else if (m0->m_pkthdr.csum_flags & CSUM_IP6_TCP)
4538 			csum_type = TX_CSUM_TCPIP6;
4539 		else if (m0->m_pkthdr.csum_flags & CSUM_IP6_UDP)
4540 			csum_type = TX_CSUM_UDPIP6;
4541 #if defined(INET)
4542 		else if (m0->m_pkthdr.csum_flags & CSUM_IP) {
4543 			/*
4544 			 * XXX: The firmware appears to stomp on the
4545 			 * fragment/flags field of the IP header when
4546 			 * using TX_CSUM_IP.  Fall back to doing
4547 			 * software checksums.
4548 			 */
4549 			u_short *sump;
4550 			struct mbuf *m;
4551 			int offset;
4552 
4553 			m = m0;
4554 			offset = 0;
4555 			sump = m_advance(&m, &offset, m0->m_pkthdr.l2hlen +
4556 			    offsetof(struct ip, ip_sum));
4557 			*sump = in_cksum_skip(m0, m0->m_pkthdr.l2hlen +
4558 			    m0->m_pkthdr.l3hlen, m0->m_pkthdr.l2hlen);
4559 			m0->m_pkthdr.csum_flags &= ~CSUM_IP;
4560 		}
4561 #endif
4562 
4563 		cpl = (void *)(wr + 1);
4564 	}
4565 
4566 	/* Checksum offload */
4567 	ctrl1 = 0;
4568 	if (needs_l3_csum(m0) == 0)
4569 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4570 	if (csum_type >= 0) {
4571 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0,
4572 	    ("%s: mbuf %p needs checksum offload but missing header lengths",
4573 			__func__, m0));
4574 
4575 		if (chip_id(sc) <= CHELSIO_T5) {
4576 			ctrl1 |= V_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4577 			    ETHER_HDR_LEN);
4578 		} else {
4579 			ctrl1 |= V_T6_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4580 			    ETHER_HDR_LEN);
4581 		}
4582 		ctrl1 |= V_TXPKT_IPHDR_LEN(m0->m_pkthdr.l3hlen);
4583 		ctrl1 |= V_TXPKT_CSUM_TYPE(csum_type);
4584 	} else
4585 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4586 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4587 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4588 		txq->txcsum++;	/* some hardware assistance provided */
4589 
4590 	/* VLAN tag insertion */
4591 	if (needs_vlan_insertion(m0)) {
4592 		ctrl1 |= F_TXPKT_VLAN_VLD |
4593 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4594 		txq->vlan_insertion++;
4595 	}
4596 
4597 	/* CPL header */
4598 	cpl->ctrl0 = txq->cpl_ctrl0;
4599 	cpl->pack = 0;
4600 	cpl->len = htobe16(pktlen);
4601 	cpl->ctrl1 = htobe64(ctrl1);
4602 
4603 	/* SGL */
4604 	dst = (void *)(cpl + 1);
4605 
4606 	/*
4607 	 * A packet using TSO will use up an entire descriptor for the
4608 	 * firmware work request header, LSO CPL, and TX_PKT_XT CPL.
4609 	 * If this descriptor is the last descriptor in the ring, wrap
4610 	 * around to the front of the ring explicitly for the start of
4611 	 * the sgl.
4612 	 */
4613 	if (dst == (void *)&eq->desc[eq->sidx]) {
4614 		dst = (void *)&eq->desc[0];
4615 		write_gl_to_txd(txq, m0, &dst, 0);
4616 	} else
4617 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4618 	txq->sgl_wrs++;
4619 
4620 	txq->txpkt_wrs++;
4621 
4622 	txsd = &txq->sdesc[eq->pidx];
4623 	txsd->m = m0;
4624 	txsd->desc_used = ndesc;
4625 
4626 	return (ndesc);
4627 }
4628 
4629 /*
4630  * Write a raw WR to the hardware descriptors, update the software
4631  * descriptor, and advance the pidx.  It is guaranteed that enough
4632  * descriptors are available.
4633  *
4634  * The return value is the # of hardware descriptors used.
4635  */
4636 static u_int
4637 write_raw_wr(struct sge_txq *txq, void *wr, struct mbuf *m0, u_int available)
4638 {
4639 	struct sge_eq *eq = &txq->eq;
4640 	struct tx_sdesc *txsd;
4641 	struct mbuf *m;
4642 	caddr_t dst;
4643 	int len16, ndesc;
4644 
4645 	len16 = mbuf_len16(m0);
4646 	ndesc = howmany(len16, EQ_ESIZE / 16);
4647 	MPASS(ndesc <= available);
4648 
4649 	dst = wr;
4650 	for (m = m0; m != NULL; m = m->m_next)
4651 		copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4652 
4653 	txq->raw_wrs++;
4654 
4655 	txsd = &txq->sdesc[eq->pidx];
4656 	txsd->m = m0;
4657 	txsd->desc_used = ndesc;
4658 
4659 	return (ndesc);
4660 }
4661 
4662 /*
4663  * Write a txpkt WR for this packet to the hardware descriptors, update the
4664  * software descriptor, and advance the pidx.  It is guaranteed that enough
4665  * descriptors are available.
4666  *
4667  * The return value is the # of hardware descriptors used.
4668  */
4669 static u_int
4670 write_txpkt_wr(struct sge_txq *txq, struct fw_eth_tx_pkt_wr *wr,
4671     struct mbuf *m0, u_int available)
4672 {
4673 	struct sge_eq *eq = &txq->eq;
4674 	struct tx_sdesc *txsd;
4675 	struct cpl_tx_pkt_core *cpl;
4676 	uint32_t ctrl;	/* used in many unrelated places */
4677 	uint64_t ctrl1;
4678 	int len16, ndesc, pktlen, nsegs;
4679 	caddr_t dst;
4680 
4681 	TXQ_LOCK_ASSERT_OWNED(txq);
4682 	M_ASSERTPKTHDR(m0);
4683 	MPASS(available > 0 && available < eq->sidx);
4684 
4685 	len16 = mbuf_len16(m0);
4686 	nsegs = mbuf_nsegs(m0);
4687 	pktlen = m0->m_pkthdr.len;
4688 	ctrl = sizeof(struct cpl_tx_pkt_core);
4689 	if (needs_tso(m0))
4690 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4691 	else if (pktlen <= imm_payload(2) && available >= 2) {
4692 		/* Immediate data.  Recalculate len16 and set nsegs to 0. */
4693 		ctrl += pktlen;
4694 		len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) +
4695 		    sizeof(struct cpl_tx_pkt_core) + pktlen, 16);
4696 		nsegs = 0;
4697 	}
4698 	ndesc = howmany(len16, EQ_ESIZE / 16);
4699 	MPASS(ndesc <= available);
4700 
4701 	/* Firmware work request header */
4702 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4703 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
4704 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4705 
4706 	ctrl = V_FW_WR_LEN16(len16);
4707 	wr->equiq_to_len16 = htobe32(ctrl);
4708 	wr->r3 = 0;
4709 
4710 	if (needs_tso(m0)) {
4711 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4712 
4713 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4714 		    m0->m_pkthdr.l4hlen > 0,
4715 		    ("%s: mbuf %p needs TSO but missing header lengths",
4716 			__func__, m0));
4717 
4718 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4719 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4720 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4721 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4722 			ctrl |= V_LSO_ETHHDR_LEN(1);
4723 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4724 			ctrl |= F_LSO_IPV6;
4725 
4726 		lso->lso_ctrl = htobe32(ctrl);
4727 		lso->ipid_ofst = htobe16(0);
4728 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4729 		lso->seqno_offset = htobe32(0);
4730 		lso->len = htobe32(pktlen);
4731 
4732 		cpl = (void *)(lso + 1);
4733 
4734 		txq->tso_wrs++;
4735 	} else
4736 		cpl = (void *)(wr + 1);
4737 
4738 	/* Checksum offload */
4739 	ctrl1 = 0;
4740 	if (needs_l3_csum(m0) == 0)
4741 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4742 	if (needs_l4_csum(m0) == 0)
4743 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4744 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4745 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4746 		txq->txcsum++;	/* some hardware assistance provided */
4747 
4748 	/* VLAN tag insertion */
4749 	if (needs_vlan_insertion(m0)) {
4750 		ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4751 		txq->vlan_insertion++;
4752 	}
4753 
4754 	/* CPL header */
4755 	cpl->ctrl0 = txq->cpl_ctrl0;
4756 	cpl->pack = 0;
4757 	cpl->len = htobe16(pktlen);
4758 	cpl->ctrl1 = htobe64(ctrl1);
4759 
4760 	/* SGL */
4761 	dst = (void *)(cpl + 1);
4762 	if (nsegs > 0) {
4763 
4764 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4765 		txq->sgl_wrs++;
4766 	} else {
4767 		struct mbuf *m;
4768 
4769 		for (m = m0; m != NULL; m = m->m_next) {
4770 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4771 #ifdef INVARIANTS
4772 			pktlen -= m->m_len;
4773 #endif
4774 		}
4775 #ifdef INVARIANTS
4776 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
4777 #endif
4778 		txq->imm_wrs++;
4779 	}
4780 
4781 	txq->txpkt_wrs++;
4782 
4783 	txsd = &txq->sdesc[eq->pidx];
4784 	txsd->m = m0;
4785 	txsd->desc_used = ndesc;
4786 
4787 	return (ndesc);
4788 }
4789 
4790 static int
4791 try_txpkts(struct mbuf *m, struct mbuf *n, struct txpkts *txp, u_int available)
4792 {
4793 	u_int needed, nsegs1, nsegs2, l1, l2;
4794 
4795 	if (cannot_use_txpkts(m) || cannot_use_txpkts(n))
4796 		return (1);
4797 
4798 	nsegs1 = mbuf_nsegs(m);
4799 	nsegs2 = mbuf_nsegs(n);
4800 	if (nsegs1 + nsegs2 == 2) {
4801 		txp->wr_type = 1;
4802 		l1 = l2 = txpkts1_len16();
4803 	} else {
4804 		txp->wr_type = 0;
4805 		l1 = txpkts0_len16(nsegs1);
4806 		l2 = txpkts0_len16(nsegs2);
4807 	}
4808 	txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) + l1 + l2;
4809 	needed = howmany(txp->len16, EQ_ESIZE / 16);
4810 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4811 		return (1);
4812 
4813 	txp->plen = m->m_pkthdr.len + n->m_pkthdr.len;
4814 	if (txp->plen > 65535)
4815 		return (1);
4816 
4817 	txp->npkt = 2;
4818 	set_mbuf_len16(m, l1);
4819 	set_mbuf_len16(n, l2);
4820 
4821 	return (0);
4822 }
4823 
4824 static int
4825 add_to_txpkts(struct mbuf *m, struct txpkts *txp, u_int available)
4826 {
4827 	u_int plen, len16, needed, nsegs;
4828 
4829 	MPASS(txp->wr_type == 0 || txp->wr_type == 1);
4830 
4831 	if (cannot_use_txpkts(m))
4832 		return (1);
4833 
4834 	nsegs = mbuf_nsegs(m);
4835 	if (txp->wr_type == 1 && nsegs != 1)
4836 		return (1);
4837 
4838 	plen = txp->plen + m->m_pkthdr.len;
4839 	if (plen > 65535)
4840 		return (1);
4841 
4842 	if (txp->wr_type == 0)
4843 		len16 = txpkts0_len16(nsegs);
4844 	else
4845 		len16 = txpkts1_len16();
4846 	needed = howmany(txp->len16 + len16, EQ_ESIZE / 16);
4847 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4848 		return (1);
4849 
4850 	txp->npkt++;
4851 	txp->plen = plen;
4852 	txp->len16 += len16;
4853 	set_mbuf_len16(m, len16);
4854 
4855 	return (0);
4856 }
4857 
4858 /*
4859  * Write a txpkts WR for the packets in txp to the hardware descriptors, update
4860  * the software descriptor, and advance the pidx.  It is guaranteed that enough
4861  * descriptors are available.
4862  *
4863  * The return value is the # of hardware descriptors used.
4864  */
4865 static u_int
4866 write_txpkts_wr(struct sge_txq *txq, struct fw_eth_tx_pkts_wr *wr,
4867     struct mbuf *m0, const struct txpkts *txp, u_int available)
4868 {
4869 	struct sge_eq *eq = &txq->eq;
4870 	struct tx_sdesc *txsd;
4871 	struct cpl_tx_pkt_core *cpl;
4872 	uint32_t ctrl;
4873 	uint64_t ctrl1;
4874 	int ndesc, checkwrap;
4875 	struct mbuf *m;
4876 	void *flitp;
4877 
4878 	TXQ_LOCK_ASSERT_OWNED(txq);
4879 	MPASS(txp->npkt > 0);
4880 	MPASS(txp->plen < 65536);
4881 	MPASS(m0 != NULL);
4882 	MPASS(m0->m_nextpkt != NULL);
4883 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
4884 	MPASS(available > 0 && available < eq->sidx);
4885 
4886 	ndesc = howmany(txp->len16, EQ_ESIZE / 16);
4887 	MPASS(ndesc <= available);
4888 
4889 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4890 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
4891 	ctrl = V_FW_WR_LEN16(txp->len16);
4892 	wr->equiq_to_len16 = htobe32(ctrl);
4893 	wr->plen = htobe16(txp->plen);
4894 	wr->npkt = txp->npkt;
4895 	wr->r3 = 0;
4896 	wr->type = txp->wr_type;
4897 	flitp = wr + 1;
4898 
4899 	/*
4900 	 * At this point we are 16B into a hardware descriptor.  If checkwrap is
4901 	 * set then we know the WR is going to wrap around somewhere.  We'll
4902 	 * check for that at appropriate points.
4903 	 */
4904 	checkwrap = eq->sidx - ndesc < eq->pidx;
4905 	for (m = m0; m != NULL; m = m->m_nextpkt) {
4906 		if (txp->wr_type == 0) {
4907 			struct ulp_txpkt *ulpmc;
4908 			struct ulptx_idata *ulpsc;
4909 
4910 			/* ULP master command */
4911 			ulpmc = flitp;
4912 			ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
4913 			    V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid));
4914 			ulpmc->len = htobe32(mbuf_len16(m));
4915 
4916 			/* ULP subcommand */
4917 			ulpsc = (void *)(ulpmc + 1);
4918 			ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
4919 			    F_ULP_TX_SC_MORE);
4920 			ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
4921 
4922 			cpl = (void *)(ulpsc + 1);
4923 			if (checkwrap &&
4924 			    (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx])
4925 				cpl = (void *)&eq->desc[0];
4926 		} else {
4927 			cpl = flitp;
4928 		}
4929 
4930 		/* Checksum offload */
4931 		ctrl1 = 0;
4932 		if (needs_l3_csum(m) == 0)
4933 			ctrl1 |= F_TXPKT_IPCSUM_DIS;
4934 		if (needs_l4_csum(m) == 0)
4935 			ctrl1 |= F_TXPKT_L4CSUM_DIS;
4936 		if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4937 		    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4938 			txq->txcsum++;	/* some hardware assistance provided */
4939 
4940 		/* VLAN tag insertion */
4941 		if (needs_vlan_insertion(m)) {
4942 			ctrl1 |= F_TXPKT_VLAN_VLD |
4943 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
4944 			txq->vlan_insertion++;
4945 		}
4946 
4947 		/* CPL header */
4948 		cpl->ctrl0 = txq->cpl_ctrl0;
4949 		cpl->pack = 0;
4950 		cpl->len = htobe16(m->m_pkthdr.len);
4951 		cpl->ctrl1 = htobe64(ctrl1);
4952 
4953 		flitp = cpl + 1;
4954 		if (checkwrap &&
4955 		    (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
4956 			flitp = (void *)&eq->desc[0];
4957 
4958 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap);
4959 
4960 	}
4961 
4962 	if (txp->wr_type == 0) {
4963 		txq->txpkts0_pkts += txp->npkt;
4964 		txq->txpkts0_wrs++;
4965 	} else {
4966 		txq->txpkts1_pkts += txp->npkt;
4967 		txq->txpkts1_wrs++;
4968 	}
4969 
4970 	txsd = &txq->sdesc[eq->pidx];
4971 	txsd->m = m0;
4972 	txsd->desc_used = ndesc;
4973 
4974 	return (ndesc);
4975 }
4976 
4977 /*
4978  * If the SGL ends on an address that is not 16 byte aligned, this function will
4979  * add a 0 filled flit at the end.
4980  */
4981 static void
4982 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap)
4983 {
4984 	struct sge_eq *eq = &txq->eq;
4985 	struct sglist *gl = txq->gl;
4986 	struct sglist_seg *seg;
4987 	__be64 *flitp, *wrap;
4988 	struct ulptx_sgl *usgl;
4989 	int i, nflits, nsegs;
4990 
4991 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
4992 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
4993 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
4994 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
4995 
4996 	get_pkt_gl(m, gl);
4997 	nsegs = gl->sg_nseg;
4998 	MPASS(nsegs > 0);
4999 
5000 	nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
5001 	flitp = (__be64 *)(*to);
5002 	wrap = (__be64 *)(&eq->desc[eq->sidx]);
5003 	seg = &gl->sg_segs[0];
5004 	usgl = (void *)flitp;
5005 
5006 	/*
5007 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
5008 	 * ring, so we're at least 16 bytes away from the status page.  There is
5009 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
5010 	 */
5011 
5012 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
5013 	    V_ULPTX_NSGE(nsegs));
5014 	usgl->len0 = htobe32(seg->ss_len);
5015 	usgl->addr0 = htobe64(seg->ss_paddr);
5016 	seg++;
5017 
5018 	if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) {
5019 
5020 		/* Won't wrap around at all */
5021 
5022 		for (i = 0; i < nsegs - 1; i++, seg++) {
5023 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
5024 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
5025 		}
5026 		if (i & 1)
5027 			usgl->sge[i / 2].len[1] = htobe32(0);
5028 		flitp += nflits;
5029 	} else {
5030 
5031 		/* Will wrap somewhere in the rest of the SGL */
5032 
5033 		/* 2 flits already written, write the rest flit by flit */
5034 		flitp = (void *)(usgl + 1);
5035 		for (i = 0; i < nflits - 2; i++) {
5036 			if (flitp == wrap)
5037 				flitp = (void *)eq->desc;
5038 			*flitp++ = get_flit(seg, nsegs - 1, i);
5039 		}
5040 	}
5041 
5042 	if (nflits & 1) {
5043 		MPASS(((uintptr_t)flitp) & 0xf);
5044 		*flitp++ = 0;
5045 	}
5046 
5047 	MPASS((((uintptr_t)flitp) & 0xf) == 0);
5048 	if (__predict_false(flitp == wrap))
5049 		*to = (void *)eq->desc;
5050 	else
5051 		*to = (void *)flitp;
5052 }
5053 
5054 static inline void
5055 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
5056 {
5057 
5058 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
5059 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
5060 
5061 	if (__predict_true((uintptr_t)(*to) + len <=
5062 	    (uintptr_t)&eq->desc[eq->sidx])) {
5063 		bcopy(from, *to, len);
5064 		(*to) += len;
5065 	} else {
5066 		int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
5067 
5068 		bcopy(from, *to, portion);
5069 		from += portion;
5070 		portion = len - portion;	/* remaining */
5071 		bcopy(from, (void *)eq->desc, portion);
5072 		(*to) = (caddr_t)eq->desc + portion;
5073 	}
5074 }
5075 
5076 static inline void
5077 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n)
5078 {
5079 	u_int db;
5080 
5081 	MPASS(n > 0);
5082 
5083 	db = eq->doorbells;
5084 	if (n > 1)
5085 		clrbit(&db, DOORBELL_WCWR);
5086 	wmb();
5087 
5088 	switch (ffs(db) - 1) {
5089 	case DOORBELL_UDB:
5090 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
5091 		break;
5092 
5093 	case DOORBELL_WCWR: {
5094 		volatile uint64_t *dst, *src;
5095 		int i;
5096 
5097 		/*
5098 		 * Queues whose 128B doorbell segment fits in the page do not
5099 		 * use relative qid (udb_qid is always 0).  Only queues with
5100 		 * doorbell segments can do WCWR.
5101 		 */
5102 		KASSERT(eq->udb_qid == 0 && n == 1,
5103 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
5104 		    __func__, eq->doorbells, n, eq->dbidx, eq));
5105 
5106 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
5107 		    UDBS_DB_OFFSET);
5108 		i = eq->dbidx;
5109 		src = (void *)&eq->desc[i];
5110 		while (src != (void *)&eq->desc[i + 1])
5111 			*dst++ = *src++;
5112 		wmb();
5113 		break;
5114 	}
5115 
5116 	case DOORBELL_UDBWC:
5117 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
5118 		wmb();
5119 		break;
5120 
5121 	case DOORBELL_KDB:
5122 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
5123 		    V_QID(eq->cntxt_id) | V_PIDX(n));
5124 		break;
5125 	}
5126 
5127 	IDXINCR(eq->dbidx, n, eq->sidx);
5128 }
5129 
5130 static inline u_int
5131 reclaimable_tx_desc(struct sge_eq *eq)
5132 {
5133 	uint16_t hw_cidx;
5134 
5135 	hw_cidx = read_hw_cidx(eq);
5136 	return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx));
5137 }
5138 
5139 static inline u_int
5140 total_available_tx_desc(struct sge_eq *eq)
5141 {
5142 	uint16_t hw_cidx, pidx;
5143 
5144 	hw_cidx = read_hw_cidx(eq);
5145 	pidx = eq->pidx;
5146 
5147 	if (pidx == hw_cidx)
5148 		return (eq->sidx - 1);
5149 	else
5150 		return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1);
5151 }
5152 
5153 static inline uint16_t
5154 read_hw_cidx(struct sge_eq *eq)
5155 {
5156 	struct sge_qstat *spg = (void *)&eq->desc[eq->sidx];
5157 	uint16_t cidx = spg->cidx;	/* stable snapshot */
5158 
5159 	return (be16toh(cidx));
5160 }
5161 
5162 /*
5163  * Reclaim 'n' descriptors approximately.
5164  */
5165 static u_int
5166 reclaim_tx_descs(struct sge_txq *txq, u_int n)
5167 {
5168 	struct tx_sdesc *txsd;
5169 	struct sge_eq *eq = &txq->eq;
5170 	u_int can_reclaim, reclaimed;
5171 
5172 	TXQ_LOCK_ASSERT_OWNED(txq);
5173 	MPASS(n > 0);
5174 
5175 	reclaimed = 0;
5176 	can_reclaim = reclaimable_tx_desc(eq);
5177 	while (can_reclaim && reclaimed < n) {
5178 		int ndesc;
5179 		struct mbuf *m, *nextpkt;
5180 
5181 		txsd = &txq->sdesc[eq->cidx];
5182 		ndesc = txsd->desc_used;
5183 
5184 		/* Firmware doesn't return "partial" credits. */
5185 		KASSERT(can_reclaim >= ndesc,
5186 		    ("%s: unexpected number of credits: %d, %d",
5187 		    __func__, can_reclaim, ndesc));
5188 		KASSERT(ndesc != 0,
5189 		    ("%s: descriptor with no credits: cidx %d",
5190 		    __func__, eq->cidx));
5191 
5192 		for (m = txsd->m; m != NULL; m = nextpkt) {
5193 			nextpkt = m->m_nextpkt;
5194 			m->m_nextpkt = NULL;
5195 			m_freem(m);
5196 		}
5197 		reclaimed += ndesc;
5198 		can_reclaim -= ndesc;
5199 		IDXINCR(eq->cidx, ndesc, eq->sidx);
5200 	}
5201 
5202 	return (reclaimed);
5203 }
5204 
5205 static void
5206 tx_reclaim(void *arg, int n)
5207 {
5208 	struct sge_txq *txq = arg;
5209 	struct sge_eq *eq = &txq->eq;
5210 
5211 	do {
5212 		if (TXQ_TRYLOCK(txq) == 0)
5213 			break;
5214 		n = reclaim_tx_descs(txq, 32);
5215 		if (eq->cidx == eq->pidx)
5216 			eq->equeqidx = eq->pidx;
5217 		TXQ_UNLOCK(txq);
5218 	} while (n > 0);
5219 }
5220 
5221 static __be64
5222 get_flit(struct sglist_seg *segs, int nsegs, int idx)
5223 {
5224 	int i = (idx / 3) * 2;
5225 
5226 	switch (idx % 3) {
5227 	case 0: {
5228 		uint64_t rc;
5229 
5230 		rc = (uint64_t)segs[i].ss_len << 32;
5231 		if (i + 1 < nsegs)
5232 			rc |= (uint64_t)(segs[i + 1].ss_len);
5233 
5234 		return (htobe64(rc));
5235 	}
5236 	case 1:
5237 		return (htobe64(segs[i].ss_paddr));
5238 	case 2:
5239 		return (htobe64(segs[i + 1].ss_paddr));
5240 	}
5241 
5242 	return (0);
5243 }
5244 
5245 static void
5246 find_best_refill_source(struct adapter *sc, struct sge_fl *fl, int maxp)
5247 {
5248 	int8_t zidx, hwidx, idx;
5249 	uint16_t region1, region3;
5250 	int spare, spare_needed, n;
5251 	struct sw_zone_info *swz;
5252 	struct hw_buf_info *hwb, *hwb_list = &sc->sge.hw_buf_info[0];
5253 
5254 	/*
5255 	 * Buffer Packing: Look for PAGE_SIZE or larger zone which has a bufsize
5256 	 * large enough for the max payload and cluster metadata.  Otherwise
5257 	 * settle for the largest bufsize that leaves enough room in the cluster
5258 	 * for metadata.
5259 	 *
5260 	 * Without buffer packing: Look for the smallest zone which has a
5261 	 * bufsize large enough for the max payload.  Settle for the largest
5262 	 * bufsize available if there's nothing big enough for max payload.
5263 	 */
5264 	spare_needed = fl->flags & FL_BUF_PACKING ? CL_METADATA_SIZE : 0;
5265 	swz = &sc->sge.sw_zone_info[0];
5266 	hwidx = -1;
5267 	for (zidx = 0; zidx < SW_ZONE_SIZES; zidx++, swz++) {
5268 		if (swz->size > largest_rx_cluster) {
5269 			if (__predict_true(hwidx != -1))
5270 				break;
5271 
5272 			/*
5273 			 * This is a misconfiguration.  largest_rx_cluster is
5274 			 * preventing us from finding a refill source.  See
5275 			 * dev.t5nex.<n>.buffer_sizes to figure out why.
5276 			 */
5277 			device_printf(sc->dev, "largest_rx_cluster=%u leaves no"
5278 			    " refill source for fl %p (dma %u).  Ignored.\n",
5279 			    largest_rx_cluster, fl, maxp);
5280 		}
5281 		for (idx = swz->head_hwidx; idx != -1; idx = hwb->next) {
5282 			hwb = &hwb_list[idx];
5283 			spare = swz->size - hwb->size;
5284 			if (spare < spare_needed)
5285 				continue;
5286 
5287 			hwidx = idx;		/* best option so far */
5288 			if (hwb->size >= maxp) {
5289 
5290 				if ((fl->flags & FL_BUF_PACKING) == 0)
5291 					goto done; /* stop looking (not packing) */
5292 
5293 				if (swz->size >= safest_rx_cluster)
5294 					goto done; /* stop looking (packing) */
5295 			}
5296 			break;		/* keep looking, next zone */
5297 		}
5298 	}
5299 done:
5300 	/* A usable hwidx has been located. */
5301 	MPASS(hwidx != -1);
5302 	hwb = &hwb_list[hwidx];
5303 	zidx = hwb->zidx;
5304 	swz = &sc->sge.sw_zone_info[zidx];
5305 	region1 = 0;
5306 	region3 = swz->size - hwb->size;
5307 
5308 	/*
5309 	 * Stay within this zone and see if there is a better match when mbuf
5310 	 * inlining is allowed.  Remember that the hwidx's are sorted in
5311 	 * decreasing order of size (so in increasing order of spare area).
5312 	 */
5313 	for (idx = hwidx; idx != -1; idx = hwb->next) {
5314 		hwb = &hwb_list[idx];
5315 		spare = swz->size - hwb->size;
5316 
5317 		if (allow_mbufs_in_cluster == 0 || hwb->size < maxp)
5318 			break;
5319 
5320 		/*
5321 		 * Do not inline mbufs if doing so would violate the pad/pack
5322 		 * boundary alignment requirement.
5323 		 */
5324 		if (fl_pad && (MSIZE % sc->params.sge.pad_boundary) != 0)
5325 			continue;
5326 		if (fl->flags & FL_BUF_PACKING &&
5327 		    (MSIZE % sc->params.sge.pack_boundary) != 0)
5328 			continue;
5329 
5330 		if (spare < CL_METADATA_SIZE + MSIZE)
5331 			continue;
5332 		n = (spare - CL_METADATA_SIZE) / MSIZE;
5333 		if (n > howmany(hwb->size, maxp))
5334 			break;
5335 
5336 		hwidx = idx;
5337 		if (fl->flags & FL_BUF_PACKING) {
5338 			region1 = n * MSIZE;
5339 			region3 = spare - region1;
5340 		} else {
5341 			region1 = MSIZE;
5342 			region3 = spare - region1;
5343 			break;
5344 		}
5345 	}
5346 
5347 	KASSERT(zidx >= 0 && zidx < SW_ZONE_SIZES,
5348 	    ("%s: bad zone %d for fl %p, maxp %d", __func__, zidx, fl, maxp));
5349 	KASSERT(hwidx >= 0 && hwidx <= SGE_FLBUF_SIZES,
5350 	    ("%s: bad hwidx %d for fl %p, maxp %d", __func__, hwidx, fl, maxp));
5351 	KASSERT(region1 + sc->sge.hw_buf_info[hwidx].size + region3 ==
5352 	    sc->sge.sw_zone_info[zidx].size,
5353 	    ("%s: bad buffer layout for fl %p, maxp %d. "
5354 		"cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5355 		sc->sge.sw_zone_info[zidx].size, region1,
5356 		sc->sge.hw_buf_info[hwidx].size, region3));
5357 	if (fl->flags & FL_BUF_PACKING || region1 > 0) {
5358 		KASSERT(region3 >= CL_METADATA_SIZE,
5359 		    ("%s: no room for metadata.  fl %p, maxp %d; "
5360 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5361 		    sc->sge.sw_zone_info[zidx].size, region1,
5362 		    sc->sge.hw_buf_info[hwidx].size, region3));
5363 		KASSERT(region1 % MSIZE == 0,
5364 		    ("%s: bad mbuf region for fl %p, maxp %d. "
5365 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5366 		    sc->sge.sw_zone_info[zidx].size, region1,
5367 		    sc->sge.hw_buf_info[hwidx].size, region3));
5368 	}
5369 
5370 	fl->cll_def.zidx = zidx;
5371 	fl->cll_def.hwidx = hwidx;
5372 	fl->cll_def.region1 = region1;
5373 	fl->cll_def.region3 = region3;
5374 }
5375 
5376 static void
5377 find_safe_refill_source(struct adapter *sc, struct sge_fl *fl)
5378 {
5379 	struct sge *s = &sc->sge;
5380 	struct hw_buf_info *hwb;
5381 	struct sw_zone_info *swz;
5382 	int spare;
5383 	int8_t hwidx;
5384 
5385 	if (fl->flags & FL_BUF_PACKING)
5386 		hwidx = s->safe_hwidx2;	/* with room for metadata */
5387 	else if (allow_mbufs_in_cluster && s->safe_hwidx2 != -1) {
5388 		hwidx = s->safe_hwidx2;
5389 		hwb = &s->hw_buf_info[hwidx];
5390 		swz = &s->sw_zone_info[hwb->zidx];
5391 		spare = swz->size - hwb->size;
5392 
5393 		/* no good if there isn't room for an mbuf as well */
5394 		if (spare < CL_METADATA_SIZE + MSIZE)
5395 			hwidx = s->safe_hwidx1;
5396 	} else
5397 		hwidx = s->safe_hwidx1;
5398 
5399 	if (hwidx == -1) {
5400 		/* No fallback source */
5401 		fl->cll_alt.hwidx = -1;
5402 		fl->cll_alt.zidx = -1;
5403 
5404 		return;
5405 	}
5406 
5407 	hwb = &s->hw_buf_info[hwidx];
5408 	swz = &s->sw_zone_info[hwb->zidx];
5409 	spare = swz->size - hwb->size;
5410 	fl->cll_alt.hwidx = hwidx;
5411 	fl->cll_alt.zidx = hwb->zidx;
5412 	if (allow_mbufs_in_cluster &&
5413 	    (fl_pad == 0 || (MSIZE % sc->params.sge.pad_boundary) == 0))
5414 		fl->cll_alt.region1 = ((spare - CL_METADATA_SIZE) / MSIZE) * MSIZE;
5415 	else
5416 		fl->cll_alt.region1 = 0;
5417 	fl->cll_alt.region3 = spare - fl->cll_alt.region1;
5418 }
5419 
5420 static void
5421 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
5422 {
5423 	mtx_lock(&sc->sfl_lock);
5424 	FL_LOCK(fl);
5425 	if ((fl->flags & FL_DOOMED) == 0) {
5426 		fl->flags |= FL_STARVING;
5427 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
5428 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
5429 	}
5430 	FL_UNLOCK(fl);
5431 	mtx_unlock(&sc->sfl_lock);
5432 }
5433 
5434 static void
5435 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq)
5436 {
5437 	struct sge_wrq *wrq = (void *)eq;
5438 
5439 	atomic_readandclear_int(&eq->equiq);
5440 	taskqueue_enqueue(sc->tq[eq->tx_chan], &wrq->wrq_tx_task);
5441 }
5442 
5443 static void
5444 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq)
5445 {
5446 	struct sge_txq *txq = (void *)eq;
5447 
5448 	MPASS((eq->flags & EQ_TYPEMASK) == EQ_ETH);
5449 
5450 	atomic_readandclear_int(&eq->equiq);
5451 	mp_ring_check_drainage(txq->r, 0);
5452 	taskqueue_enqueue(sc->tq[eq->tx_chan], &txq->tx_reclaim_task);
5453 }
5454 
5455 static int
5456 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
5457     struct mbuf *m)
5458 {
5459 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
5460 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
5461 	struct adapter *sc = iq->adapter;
5462 	struct sge *s = &sc->sge;
5463 	struct sge_eq *eq;
5464 	static void (*h[])(struct adapter *, struct sge_eq *) = {NULL,
5465 		&handle_wrq_egr_update, &handle_eth_egr_update,
5466 		&handle_wrq_egr_update};
5467 
5468 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5469 	    rss->opcode));
5470 
5471 	eq = s->eqmap[qid - s->eq_start - s->eq_base];
5472 	(*h[eq->flags & EQ_TYPEMASK])(sc, eq);
5473 
5474 	return (0);
5475 }
5476 
5477 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
5478 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
5479     offsetof(struct cpl_fw6_msg, data));
5480 
5481 static int
5482 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
5483 {
5484 	struct adapter *sc = iq->adapter;
5485 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
5486 
5487 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5488 	    rss->opcode));
5489 
5490 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
5491 		const struct rss_header *rss2;
5492 
5493 		rss2 = (const struct rss_header *)&cpl->data[0];
5494 		return (t4_cpl_handler[rss2->opcode](iq, rss2, m));
5495 	}
5496 
5497 	return (t4_fw_msg_handler[cpl->type](sc, &cpl->data[0]));
5498 }
5499 
5500 /**
5501  *	t4_handle_wrerr_rpl - process a FW work request error message
5502  *	@adap: the adapter
5503  *	@rpl: start of the FW message
5504  */
5505 static int
5506 t4_handle_wrerr_rpl(struct adapter *adap, const __be64 *rpl)
5507 {
5508 	u8 opcode = *(const u8 *)rpl;
5509 	const struct fw_error_cmd *e = (const void *)rpl;
5510 	unsigned int i;
5511 
5512 	if (opcode != FW_ERROR_CMD) {
5513 		log(LOG_ERR,
5514 		    "%s: Received WRERR_RPL message with opcode %#x\n",
5515 		    device_get_nameunit(adap->dev), opcode);
5516 		return (EINVAL);
5517 	}
5518 	log(LOG_ERR, "%s: FW_ERROR (%s) ", device_get_nameunit(adap->dev),
5519 	    G_FW_ERROR_CMD_FATAL(be32toh(e->op_to_type)) ? "fatal" :
5520 	    "non-fatal");
5521 	switch (G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type))) {
5522 	case FW_ERROR_TYPE_EXCEPTION:
5523 		log(LOG_ERR, "exception info:\n");
5524 		for (i = 0; i < nitems(e->u.exception.info); i++)
5525 			log(LOG_ERR, "%s%08x", i == 0 ? "\t" : " ",
5526 			    be32toh(e->u.exception.info[i]));
5527 		log(LOG_ERR, "\n");
5528 		break;
5529 	case FW_ERROR_TYPE_HWMODULE:
5530 		log(LOG_ERR, "HW module regaddr %08x regval %08x\n",
5531 		    be32toh(e->u.hwmodule.regaddr),
5532 		    be32toh(e->u.hwmodule.regval));
5533 		break;
5534 	case FW_ERROR_TYPE_WR:
5535 		log(LOG_ERR, "WR cidx %d PF %d VF %d eqid %d hdr:\n",
5536 		    be16toh(e->u.wr.cidx),
5537 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.wr.pfn_vfn)),
5538 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.wr.pfn_vfn)),
5539 		    be32toh(e->u.wr.eqid));
5540 		for (i = 0; i < nitems(e->u.wr.wrhdr); i++)
5541 			log(LOG_ERR, "%s%02x", i == 0 ? "\t" : " ",
5542 			    e->u.wr.wrhdr[i]);
5543 		log(LOG_ERR, "\n");
5544 		break;
5545 	case FW_ERROR_TYPE_ACL:
5546 		log(LOG_ERR, "ACL cidx %d PF %d VF %d eqid %d %s",
5547 		    be16toh(e->u.acl.cidx),
5548 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.acl.pfn_vfn)),
5549 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.acl.pfn_vfn)),
5550 		    be32toh(e->u.acl.eqid),
5551 		    G_FW_ERROR_CMD_MV(be16toh(e->u.acl.mv_pkd)) ? "vlanid" :
5552 		    "MAC");
5553 		for (i = 0; i < nitems(e->u.acl.val); i++)
5554 			log(LOG_ERR, " %02x", e->u.acl.val[i]);
5555 		log(LOG_ERR, "\n");
5556 		break;
5557 	default:
5558 		log(LOG_ERR, "type %#x\n",
5559 		    G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type)));
5560 		return (EINVAL);
5561 	}
5562 	return (0);
5563 }
5564 
5565 static int
5566 sysctl_uint16(SYSCTL_HANDLER_ARGS)
5567 {
5568 	uint16_t *id = arg1;
5569 	int i = *id;
5570 
5571 	return sysctl_handle_int(oidp, &i, 0, req);
5572 }
5573 
5574 static int
5575 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
5576 {
5577 	struct sge *s = arg1;
5578 	struct hw_buf_info *hwb = &s->hw_buf_info[0];
5579 	struct sw_zone_info *swz = &s->sw_zone_info[0];
5580 	int i, rc;
5581 	struct sbuf sb;
5582 	char c;
5583 
5584 	sbuf_new(&sb, NULL, 32, SBUF_AUTOEXTEND);
5585 	for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
5586 		if (hwb->zidx >= 0 && swz[hwb->zidx].size <= largest_rx_cluster)
5587 			c = '*';
5588 		else
5589 			c = '\0';
5590 
5591 		sbuf_printf(&sb, "%u%c ", hwb->size, c);
5592 	}
5593 	sbuf_trim(&sb);
5594 	sbuf_finish(&sb);
5595 	rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
5596 	sbuf_delete(&sb);
5597 	return (rc);
5598 }
5599 
5600 #ifdef RATELIMIT
5601 /*
5602  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
5603  */
5604 static inline u_int
5605 txpkt_eo_len16(u_int nsegs, u_int immhdrs, u_int tso)
5606 {
5607 	u_int n;
5608 
5609 	MPASS(immhdrs > 0);
5610 
5611 	n = roundup2(sizeof(struct fw_eth_tx_eo_wr) +
5612 	    sizeof(struct cpl_tx_pkt_core) + immhdrs, 16);
5613 	if (__predict_false(nsegs == 0))
5614 		goto done;
5615 
5616 	nsegs--; /* first segment is part of ulptx_sgl */
5617 	n += sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
5618 	if (tso)
5619 		n += sizeof(struct cpl_tx_pkt_lso_core);
5620 
5621 done:
5622 	return (howmany(n, 16));
5623 }
5624 
5625 #define ETID_FLOWC_NPARAMS 6
5626 #define ETID_FLOWC_LEN (roundup2((sizeof(struct fw_flowc_wr) + \
5627     ETID_FLOWC_NPARAMS * sizeof(struct fw_flowc_mnemval)), 16))
5628 #define ETID_FLOWC_LEN16 (howmany(ETID_FLOWC_LEN, 16))
5629 
5630 static int
5631 send_etid_flowc_wr(struct cxgbe_snd_tag *cst, struct port_info *pi,
5632     struct vi_info *vi)
5633 {
5634 	struct wrq_cookie cookie;
5635 	u_int pfvf = G_FW_VIID_PFN(vi->viid) << S_FW_VIID_PFN;
5636 	struct fw_flowc_wr *flowc;
5637 
5638 	mtx_assert(&cst->lock, MA_OWNED);
5639 	MPASS((cst->flags & (EO_FLOWC_PENDING | EO_FLOWC_RPL_PENDING)) ==
5640 	    EO_FLOWC_PENDING);
5641 
5642 	flowc = start_wrq_wr(cst->eo_txq, ETID_FLOWC_LEN16, &cookie);
5643 	if (__predict_false(flowc == NULL))
5644 		return (ENOMEM);
5645 
5646 	bzero(flowc, ETID_FLOWC_LEN);
5647 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
5648 	    V_FW_FLOWC_WR_NPARAMS(ETID_FLOWC_NPARAMS) | V_FW_WR_COMPL(0));
5649 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(ETID_FLOWC_LEN16) |
5650 	    V_FW_WR_FLOWID(cst->etid));
5651 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN;
5652 	flowc->mnemval[0].val = htobe32(pfvf);
5653 	flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH;
5654 	flowc->mnemval[1].val = htobe32(pi->tx_chan);
5655 	flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT;
5656 	flowc->mnemval[2].val = htobe32(pi->tx_chan);
5657 	flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID;
5658 	flowc->mnemval[3].val = htobe32(cst->iqid);
5659 	flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_EOSTATE;
5660 	flowc->mnemval[4].val = htobe32(FW_FLOWC_MNEM_EOSTATE_ESTABLISHED);
5661 	flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS;
5662 	flowc->mnemval[5].val = htobe32(cst->schedcl);
5663 
5664 	commit_wrq_wr(cst->eo_txq, flowc, &cookie);
5665 
5666 	cst->flags &= ~EO_FLOWC_PENDING;
5667 	cst->flags |= EO_FLOWC_RPL_PENDING;
5668 	MPASS(cst->tx_credits >= ETID_FLOWC_LEN16);	/* flowc is first WR. */
5669 	cst->tx_credits -= ETID_FLOWC_LEN16;
5670 
5671 	return (0);
5672 }
5673 
5674 #define ETID_FLUSH_LEN16 (howmany(sizeof (struct fw_flowc_wr), 16))
5675 
5676 void
5677 send_etid_flush_wr(struct cxgbe_snd_tag *cst)
5678 {
5679 	struct fw_flowc_wr *flowc;
5680 	struct wrq_cookie cookie;
5681 
5682 	mtx_assert(&cst->lock, MA_OWNED);
5683 
5684 	flowc = start_wrq_wr(cst->eo_txq, ETID_FLUSH_LEN16, &cookie);
5685 	if (__predict_false(flowc == NULL))
5686 		CXGBE_UNIMPLEMENTED(__func__);
5687 
5688 	bzero(flowc, ETID_FLUSH_LEN16 * 16);
5689 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
5690 	    V_FW_FLOWC_WR_NPARAMS(0) | F_FW_WR_COMPL);
5691 	flowc->flowid_len16 = htobe32(V_FW_WR_LEN16(ETID_FLUSH_LEN16) |
5692 	    V_FW_WR_FLOWID(cst->etid));
5693 
5694 	commit_wrq_wr(cst->eo_txq, flowc, &cookie);
5695 
5696 	cst->flags |= EO_FLUSH_RPL_PENDING;
5697 	MPASS(cst->tx_credits >= ETID_FLUSH_LEN16);
5698 	cst->tx_credits -= ETID_FLUSH_LEN16;
5699 	cst->ncompl++;
5700 }
5701 
5702 static void
5703 write_ethofld_wr(struct cxgbe_snd_tag *cst, struct fw_eth_tx_eo_wr *wr,
5704     struct mbuf *m0, int compl)
5705 {
5706 	struct cpl_tx_pkt_core *cpl;
5707 	uint64_t ctrl1;
5708 	uint32_t ctrl;	/* used in many unrelated places */
5709 	int len16, pktlen, nsegs, immhdrs;
5710 	caddr_t dst;
5711 	uintptr_t p;
5712 	struct ulptx_sgl *usgl;
5713 	struct sglist sg;
5714 	struct sglist_seg segs[38];	/* XXX: find real limit.  XXX: get off the stack */
5715 
5716 	mtx_assert(&cst->lock, MA_OWNED);
5717 	M_ASSERTPKTHDR(m0);
5718 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
5719 	    m0->m_pkthdr.l4hlen > 0,
5720 	    ("%s: ethofld mbuf %p is missing header lengths", __func__, m0));
5721 
5722 	len16 = mbuf_eo_len16(m0);
5723 	nsegs = mbuf_eo_nsegs(m0);
5724 	pktlen = m0->m_pkthdr.len;
5725 	ctrl = sizeof(struct cpl_tx_pkt_core);
5726 	if (needs_tso(m0))
5727 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
5728 	immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen + m0->m_pkthdr.l4hlen;
5729 	ctrl += immhdrs;
5730 
5731 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_EO_WR) |
5732 	    V_FW_ETH_TX_EO_WR_IMMDLEN(ctrl) | V_FW_WR_COMPL(!!compl));
5733 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(len16) |
5734 	    V_FW_WR_FLOWID(cst->etid));
5735 	wr->r3 = 0;
5736 	if (needs_udp_csum(m0)) {
5737 		wr->u.udpseg.type = FW_ETH_TX_EO_TYPE_UDPSEG;
5738 		wr->u.udpseg.ethlen = m0->m_pkthdr.l2hlen;
5739 		wr->u.udpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
5740 		wr->u.udpseg.udplen = m0->m_pkthdr.l4hlen;
5741 		wr->u.udpseg.rtplen = 0;
5742 		wr->u.udpseg.r4 = 0;
5743 		wr->u.udpseg.mss = htobe16(pktlen - immhdrs);
5744 		wr->u.udpseg.schedpktsize = wr->u.udpseg.mss;
5745 		wr->u.udpseg.plen = htobe32(pktlen - immhdrs);
5746 		cpl = (void *)(wr + 1);
5747 	} else {
5748 		MPASS(needs_tcp_csum(m0));
5749 		wr->u.tcpseg.type = FW_ETH_TX_EO_TYPE_TCPSEG;
5750 		wr->u.tcpseg.ethlen = m0->m_pkthdr.l2hlen;
5751 		wr->u.tcpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
5752 		wr->u.tcpseg.tcplen = m0->m_pkthdr.l4hlen;
5753 		wr->u.tcpseg.tsclk_tsoff = mbuf_eo_tsclk_tsoff(m0);
5754 		wr->u.tcpseg.r4 = 0;
5755 		wr->u.tcpseg.r5 = 0;
5756 		wr->u.tcpseg.plen = htobe32(pktlen - immhdrs);
5757 
5758 		if (needs_tso(m0)) {
5759 			struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
5760 
5761 			wr->u.tcpseg.mss = htobe16(m0->m_pkthdr.tso_segsz);
5762 
5763 			ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) |
5764 			    F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE |
5765 			    V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) |
5766 			    V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
5767 			if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
5768 				ctrl |= V_LSO_ETHHDR_LEN(1);
5769 			if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
5770 				ctrl |= F_LSO_IPV6;
5771 			lso->lso_ctrl = htobe32(ctrl);
5772 			lso->ipid_ofst = htobe16(0);
5773 			lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
5774 			lso->seqno_offset = htobe32(0);
5775 			lso->len = htobe32(pktlen);
5776 
5777 			cpl = (void *)(lso + 1);
5778 		} else {
5779 			wr->u.tcpseg.mss = htobe16(0xffff);
5780 			cpl = (void *)(wr + 1);
5781 		}
5782 	}
5783 
5784 	/* Checksum offload must be requested for ethofld. */
5785 	ctrl1 = 0;
5786 	MPASS(needs_l4_csum(m0));
5787 
5788 	/* VLAN tag insertion */
5789 	if (needs_vlan_insertion(m0)) {
5790 		ctrl1 |= F_TXPKT_VLAN_VLD |
5791 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
5792 	}
5793 
5794 	/* CPL header */
5795 	cpl->ctrl0 = cst->ctrl0;
5796 	cpl->pack = 0;
5797 	cpl->len = htobe16(pktlen);
5798 	cpl->ctrl1 = htobe64(ctrl1);
5799 
5800 	/* Copy Ethernet, IP & TCP/UDP hdrs as immediate data */
5801 	p = (uintptr_t)(cpl + 1);
5802 	m_copydata(m0, 0, immhdrs, (void *)p);
5803 
5804 	/* SGL */
5805 	dst = (void *)(cpl + 1);
5806 	if (nsegs > 0) {
5807 		int i, pad;
5808 
5809 		/* zero-pad upto next 16Byte boundary, if not 16Byte aligned */
5810 		p += immhdrs;
5811 		pad = 16 - (immhdrs & 0xf);
5812 		bzero((void *)p, pad);
5813 
5814 		usgl = (void *)(p + pad);
5815 		usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
5816 		    V_ULPTX_NSGE(nsegs));
5817 
5818 		sglist_init(&sg, nitems(segs), segs);
5819 		for (; m0 != NULL; m0 = m0->m_next) {
5820 			if (__predict_false(m0->m_len == 0))
5821 				continue;
5822 			if (immhdrs >= m0->m_len) {
5823 				immhdrs -= m0->m_len;
5824 				continue;
5825 			}
5826 
5827 			sglist_append(&sg, mtod(m0, char *) + immhdrs,
5828 			    m0->m_len - immhdrs);
5829 			immhdrs = 0;
5830 		}
5831 		MPASS(sg.sg_nseg == nsegs);
5832 
5833 		/*
5834 		 * Zero pad last 8B in case the WR doesn't end on a 16B
5835 		 * boundary.
5836 		 */
5837 		*(uint64_t *)((char *)wr + len16 * 16 - 8) = 0;
5838 
5839 		usgl->len0 = htobe32(segs[0].ss_len);
5840 		usgl->addr0 = htobe64(segs[0].ss_paddr);
5841 		for (i = 0; i < nsegs - 1; i++) {
5842 			usgl->sge[i / 2].len[i & 1] = htobe32(segs[i + 1].ss_len);
5843 			usgl->sge[i / 2].addr[i & 1] = htobe64(segs[i + 1].ss_paddr);
5844 		}
5845 		if (i & 1)
5846 			usgl->sge[i / 2].len[1] = htobe32(0);
5847 	}
5848 
5849 }
5850 
5851 static void
5852 ethofld_tx(struct cxgbe_snd_tag *cst)
5853 {
5854 	struct mbuf *m;
5855 	struct wrq_cookie cookie;
5856 	int next_credits, compl;
5857 	struct fw_eth_tx_eo_wr *wr;
5858 
5859 	mtx_assert(&cst->lock, MA_OWNED);
5860 
5861 	while ((m = mbufq_first(&cst->pending_tx)) != NULL) {
5862 		M_ASSERTPKTHDR(m);
5863 
5864 		/* How many len16 credits do we need to send this mbuf. */
5865 		next_credits = mbuf_eo_len16(m);
5866 		MPASS(next_credits > 0);
5867 		if (next_credits > cst->tx_credits) {
5868 			/*
5869 			 * Tx will make progress eventually because there is at
5870 			 * least one outstanding fw4_ack that will return
5871 			 * credits and kick the tx.
5872 			 */
5873 			MPASS(cst->ncompl > 0);
5874 			return;
5875 		}
5876 		wr = start_wrq_wr(cst->eo_txq, next_credits, &cookie);
5877 		if (__predict_false(wr == NULL)) {
5878 			/* XXX: wishful thinking, not a real assertion. */
5879 			MPASS(cst->ncompl > 0);
5880 			return;
5881 		}
5882 		cst->tx_credits -= next_credits;
5883 		cst->tx_nocompl += next_credits;
5884 		compl = cst->ncompl == 0 || cst->tx_nocompl >= cst->tx_total / 2;
5885 		ETHER_BPF_MTAP(cst->com.ifp, m);
5886 		write_ethofld_wr(cst, wr, m, compl);
5887 		commit_wrq_wr(cst->eo_txq, wr, &cookie);
5888 		if (compl) {
5889 			cst->ncompl++;
5890 			cst->tx_nocompl	= 0;
5891 		}
5892 		(void) mbufq_dequeue(&cst->pending_tx);
5893 		mbufq_enqueue(&cst->pending_fwack, m);
5894 	}
5895 }
5896 
5897 int
5898 ethofld_transmit(struct ifnet *ifp, struct mbuf *m0)
5899 {
5900 	struct cxgbe_snd_tag *cst;
5901 	int rc;
5902 
5903 	MPASS(m0->m_nextpkt == NULL);
5904 	MPASS(m0->m_pkthdr.snd_tag != NULL);
5905 	cst = mst_to_cst(m0->m_pkthdr.snd_tag);
5906 
5907 	mtx_lock(&cst->lock);
5908 	MPASS(cst->flags & EO_SND_TAG_REF);
5909 
5910 	if (__predict_false(cst->flags & EO_FLOWC_PENDING)) {
5911 		struct vi_info *vi = ifp->if_softc;
5912 		struct port_info *pi = vi->pi;
5913 		struct adapter *sc = pi->adapter;
5914 		const uint32_t rss_mask = vi->rss_size - 1;
5915 		uint32_t rss_hash;
5916 
5917 		cst->eo_txq = &sc->sge.ofld_txq[vi->first_ofld_txq];
5918 		if (M_HASHTYPE_ISHASH(m0))
5919 			rss_hash = m0->m_pkthdr.flowid;
5920 		else
5921 			rss_hash = arc4random();
5922 		/* We assume RSS hashing */
5923 		cst->iqid = vi->rss[rss_hash & rss_mask];
5924 		cst->eo_txq += rss_hash % vi->nofldtxq;
5925 		rc = send_etid_flowc_wr(cst, pi, vi);
5926 		if (rc != 0)
5927 			goto done;
5928 	}
5929 
5930 	if (__predict_false(cst->plen + m0->m_pkthdr.len > eo_max_backlog)) {
5931 		rc = ENOBUFS;
5932 		goto done;
5933 	}
5934 
5935 	mbufq_enqueue(&cst->pending_tx, m0);
5936 	cst->plen += m0->m_pkthdr.len;
5937 
5938 	ethofld_tx(cst);
5939 	rc = 0;
5940 done:
5941 	mtx_unlock(&cst->lock);
5942 	if (__predict_false(rc != 0))
5943 		m_freem(m0);
5944 	return (rc);
5945 }
5946 
5947 static int
5948 ethofld_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
5949 {
5950 	struct adapter *sc = iq->adapter;
5951 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
5952 	struct mbuf *m;
5953 	u_int etid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
5954 	struct cxgbe_snd_tag *cst;
5955 	uint8_t credits = cpl->credits;
5956 
5957 	cst = lookup_etid(sc, etid);
5958 	mtx_lock(&cst->lock);
5959 	if (__predict_false(cst->flags & EO_FLOWC_RPL_PENDING)) {
5960 		MPASS(credits >= ETID_FLOWC_LEN16);
5961 		credits -= ETID_FLOWC_LEN16;
5962 		cst->flags &= ~EO_FLOWC_RPL_PENDING;
5963 	}
5964 
5965 	KASSERT(cst->ncompl > 0,
5966 	    ("%s: etid %u (%p) wasn't expecting completion.",
5967 	    __func__, etid, cst));
5968 	cst->ncompl--;
5969 
5970 	while (credits > 0) {
5971 		m = mbufq_dequeue(&cst->pending_fwack);
5972 		if (__predict_false(m == NULL)) {
5973 			/*
5974 			 * The remaining credits are for the final flush that
5975 			 * was issued when the tag was freed by the kernel.
5976 			 */
5977 			MPASS((cst->flags &
5978 			    (EO_FLUSH_RPL_PENDING | EO_SND_TAG_REF)) ==
5979 			    EO_FLUSH_RPL_PENDING);
5980 			MPASS(credits == ETID_FLUSH_LEN16);
5981 			MPASS(cst->tx_credits + cpl->credits == cst->tx_total);
5982 			MPASS(cst->ncompl == 0);
5983 
5984 			cst->flags &= ~EO_FLUSH_RPL_PENDING;
5985 			cst->tx_credits += cpl->credits;
5986 freetag:
5987 			cxgbe_snd_tag_free_locked(cst);
5988 			return (0);	/* cst is gone. */
5989 		}
5990 		KASSERT(m != NULL,
5991 		    ("%s: too many credits (%u, %u)", __func__, cpl->credits,
5992 		    credits));
5993 		KASSERT(credits >= mbuf_eo_len16(m),
5994 		    ("%s: too few credits (%u, %u, %u)", __func__,
5995 		    cpl->credits, credits, mbuf_eo_len16(m)));
5996 		credits -= mbuf_eo_len16(m);
5997 		cst->plen -= m->m_pkthdr.len;
5998 		m_freem(m);
5999 	}
6000 
6001 	cst->tx_credits += cpl->credits;
6002 	MPASS(cst->tx_credits <= cst->tx_total);
6003 
6004 	m = mbufq_first(&cst->pending_tx);
6005 	if (m != NULL && cst->tx_credits >= mbuf_eo_len16(m))
6006 		ethofld_tx(cst);
6007 
6008 	if (__predict_false((cst->flags & EO_SND_TAG_REF) == 0) &&
6009 	    cst->ncompl == 0) {
6010 		if (cst->tx_credits == cst->tx_total)
6011 			goto freetag;
6012 		else {
6013 			MPASS((cst->flags & EO_FLUSH_RPL_PENDING) == 0);
6014 			send_etid_flush_wr(cst);
6015 		}
6016 	}
6017 
6018 	mtx_unlock(&cst->lock);
6019 
6020 	return (0);
6021 }
6022 #endif
6023