xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision 78b9f0095b4af3aca6c931b2c7b009ddb8a05125)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2011 Chelsio Communications, Inc.
5  * All rights reserved.
6  * Written by: Navdeep Parhar <np@FreeBSD.org>
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include "opt_inet.h"
34 #include "opt_inet6.h"
35 #include "opt_ratelimit.h"
36 
37 #include <sys/types.h>
38 #include <sys/eventhandler.h>
39 #include <sys/mbuf.h>
40 #include <sys/socket.h>
41 #include <sys/kernel.h>
42 #include <sys/malloc.h>
43 #include <sys/queue.h>
44 #include <sys/sbuf.h>
45 #include <sys/taskqueue.h>
46 #include <sys/time.h>
47 #include <sys/sglist.h>
48 #include <sys/sysctl.h>
49 #include <sys/smp.h>
50 #include <sys/counter.h>
51 #include <net/bpf.h>
52 #include <net/ethernet.h>
53 #include <net/if.h>
54 #include <net/if_vlan_var.h>
55 #include <netinet/in.h>
56 #include <netinet/ip.h>
57 #include <netinet/ip6.h>
58 #include <netinet/tcp.h>
59 #include <netinet/udp.h>
60 #include <machine/in_cksum.h>
61 #include <machine/md_var.h>
62 #include <vm/vm.h>
63 #include <vm/pmap.h>
64 #ifdef DEV_NETMAP
65 #include <machine/bus.h>
66 #include <sys/selinfo.h>
67 #include <net/if_var.h>
68 #include <net/netmap.h>
69 #include <dev/netmap/netmap_kern.h>
70 #endif
71 
72 #include "common/common.h"
73 #include "common/t4_regs.h"
74 #include "common/t4_regs_values.h"
75 #include "common/t4_msg.h"
76 #include "t4_l2t.h"
77 #include "t4_mp_ring.h"
78 
79 #ifdef T4_PKT_TIMESTAMP
80 #define RX_COPY_THRESHOLD (MINCLSIZE - 8)
81 #else
82 #define RX_COPY_THRESHOLD MINCLSIZE
83 #endif
84 
85 /*
86  * Ethernet frames are DMA'd at this byte offset into the freelist buffer.
87  * 0-7 are valid values.
88  */
89 static int fl_pktshift = 0;
90 TUNABLE_INT("hw.cxgbe.fl_pktshift", &fl_pktshift);
91 
92 /*
93  * Pad ethernet payload up to this boundary.
94  * -1: driver should figure out a good value.
95  *  0: disable padding.
96  *  Any power of 2 from 32 to 4096 (both inclusive) is also a valid value.
97  */
98 int fl_pad = -1;
99 TUNABLE_INT("hw.cxgbe.fl_pad", &fl_pad);
100 
101 /*
102  * Status page length.
103  * -1: driver should figure out a good value.
104  *  64 or 128 are the only other valid values.
105  */
106 static int spg_len = -1;
107 TUNABLE_INT("hw.cxgbe.spg_len", &spg_len);
108 
109 /*
110  * Congestion drops.
111  * -1: no congestion feedback (not recommended).
112  *  0: backpressure the channel instead of dropping packets right away.
113  *  1: no backpressure, drop packets for the congested queue immediately.
114  */
115 static int cong_drop = 0;
116 TUNABLE_INT("hw.cxgbe.cong_drop", &cong_drop);
117 
118 /*
119  * Deliver multiple frames in the same free list buffer if they fit.
120  * -1: let the driver decide whether to enable buffer packing or not.
121  *  0: disable buffer packing.
122  *  1: enable buffer packing.
123  */
124 static int buffer_packing = -1;
125 TUNABLE_INT("hw.cxgbe.buffer_packing", &buffer_packing);
126 
127 /*
128  * Start next frame in a packed buffer at this boundary.
129  * -1: driver should figure out a good value.
130  * T4: driver will ignore this and use the same value as fl_pad above.
131  * T5: 16, or a power of 2 from 64 to 4096 (both inclusive) is a valid value.
132  */
133 static int fl_pack = -1;
134 TUNABLE_INT("hw.cxgbe.fl_pack", &fl_pack);
135 
136 /*
137  * Allow the driver to create mbuf(s) in a cluster allocated for rx.
138  * 0: never; always allocate mbufs from the zone_mbuf UMA zone.
139  * 1: ok to create mbuf(s) within a cluster if there is room.
140  */
141 static int allow_mbufs_in_cluster = 1;
142 TUNABLE_INT("hw.cxgbe.allow_mbufs_in_cluster", &allow_mbufs_in_cluster);
143 
144 /*
145  * Largest rx cluster size that the driver is allowed to allocate.
146  */
147 static int largest_rx_cluster = MJUM16BYTES;
148 TUNABLE_INT("hw.cxgbe.largest_rx_cluster", &largest_rx_cluster);
149 
150 /*
151  * Size of cluster allocation that's most likely to succeed.  The driver will
152  * fall back to this size if it fails to allocate clusters larger than this.
153  */
154 static int safest_rx_cluster = PAGE_SIZE;
155 TUNABLE_INT("hw.cxgbe.safest_rx_cluster", &safest_rx_cluster);
156 
157 #ifdef RATELIMIT
158 /*
159  * Knob to control TCP timestamp rewriting, and the granularity of the tick used
160  * for rewriting.  -1 and 0-3 are all valid values.
161  * -1: hardware should leave the TCP timestamps alone.
162  * 0: 1ms
163  * 1: 100us
164  * 2: 10us
165  * 3: 1us
166  */
167 static int tsclk = -1;
168 TUNABLE_INT("hw.cxgbe.tsclk", &tsclk);
169 
170 static int eo_max_backlog = 1024 * 1024;
171 TUNABLE_INT("hw.cxgbe.eo_max_backlog", &eo_max_backlog);
172 #endif
173 
174 /*
175  * The interrupt holdoff timers are multiplied by this value on T6+.
176  * 1 and 3-17 (both inclusive) are legal values.
177  */
178 static int tscale = 1;
179 TUNABLE_INT("hw.cxgbe.tscale", &tscale);
180 
181 /*
182  * Number of LRO entries in the lro_ctrl structure per rx queue.
183  */
184 static int lro_entries = TCP_LRO_ENTRIES;
185 TUNABLE_INT("hw.cxgbe.lro_entries", &lro_entries);
186 
187 /*
188  * This enables presorting of frames before they're fed into tcp_lro_rx.
189  */
190 static int lro_mbufs = 0;
191 TUNABLE_INT("hw.cxgbe.lro_mbufs", &lro_mbufs);
192 
193 struct txpkts {
194 	u_int wr_type;		/* type 0 or type 1 */
195 	u_int npkt;		/* # of packets in this work request */
196 	u_int plen;		/* total payload (sum of all packets) */
197 	u_int len16;		/* # of 16B pieces used by this work request */
198 };
199 
200 /* A packet's SGL.  This + m_pkthdr has all info needed for tx */
201 struct sgl {
202 	struct sglist sg;
203 	struct sglist_seg seg[TX_SGL_SEGS];
204 };
205 
206 static int service_iq(struct sge_iq *, int);
207 static int service_iq_fl(struct sge_iq *, int);
208 static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t);
209 static int t4_eth_rx(struct sge_iq *, const struct rss_header *, struct mbuf *);
210 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int);
211 static inline void init_fl(struct adapter *, struct sge_fl *, int, int, char *);
212 static inline void init_eq(struct adapter *, struct sge_eq *, int, int, uint8_t,
213     uint16_t, char *);
214 static int alloc_ring(struct adapter *, size_t, bus_dma_tag_t *, bus_dmamap_t *,
215     bus_addr_t *, void **);
216 static int free_ring(struct adapter *, bus_dma_tag_t, bus_dmamap_t, bus_addr_t,
217     void *);
218 static int alloc_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *,
219     int, int);
220 static int free_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *);
221 static void add_iq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
222     struct sge_iq *);
223 static void add_fl_sysctls(struct adapter *, struct sysctl_ctx_list *,
224     struct sysctl_oid *, struct sge_fl *);
225 static int alloc_fwq(struct adapter *);
226 static int free_fwq(struct adapter *);
227 static int alloc_ctrlq(struct adapter *, struct sge_wrq *, int,
228     struct sysctl_oid *);
229 static int alloc_rxq(struct vi_info *, struct sge_rxq *, int, int,
230     struct sysctl_oid *);
231 static int free_rxq(struct vi_info *, struct sge_rxq *);
232 #ifdef TCP_OFFLOAD
233 static int alloc_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *, int, int,
234     struct sysctl_oid *);
235 static int free_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *);
236 #endif
237 #ifdef DEV_NETMAP
238 static int alloc_nm_rxq(struct vi_info *, struct sge_nm_rxq *, int, int,
239     struct sysctl_oid *);
240 static int free_nm_rxq(struct vi_info *, struct sge_nm_rxq *);
241 static int alloc_nm_txq(struct vi_info *, struct sge_nm_txq *, int, int,
242     struct sysctl_oid *);
243 static int free_nm_txq(struct vi_info *, struct sge_nm_txq *);
244 #endif
245 static int ctrl_eq_alloc(struct adapter *, struct sge_eq *);
246 static int eth_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
247 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
248 static int ofld_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
249 #endif
250 static int alloc_eq(struct adapter *, struct vi_info *, struct sge_eq *);
251 static int free_eq(struct adapter *, struct sge_eq *);
252 static int alloc_wrq(struct adapter *, struct vi_info *, struct sge_wrq *,
253     struct sysctl_oid *);
254 static int free_wrq(struct adapter *, struct sge_wrq *);
255 static int alloc_txq(struct vi_info *, struct sge_txq *, int,
256     struct sysctl_oid *);
257 static int free_txq(struct vi_info *, struct sge_txq *);
258 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int);
259 static inline void ring_fl_db(struct adapter *, struct sge_fl *);
260 static int refill_fl(struct adapter *, struct sge_fl *, int);
261 static void refill_sfl(void *);
262 static int alloc_fl_sdesc(struct sge_fl *);
263 static void free_fl_sdesc(struct adapter *, struct sge_fl *);
264 static void find_best_refill_source(struct adapter *, struct sge_fl *, int);
265 static void find_safe_refill_source(struct adapter *, struct sge_fl *);
266 static void add_fl_to_sfl(struct adapter *, struct sge_fl *);
267 
268 static inline void get_pkt_gl(struct mbuf *, struct sglist *);
269 static inline u_int txpkt_len16(u_int, u_int);
270 static inline u_int txpkt_vm_len16(u_int, u_int);
271 static inline u_int txpkts0_len16(u_int);
272 static inline u_int txpkts1_len16(void);
273 static u_int write_txpkt_wr(struct sge_txq *, struct fw_eth_tx_pkt_wr *,
274     struct mbuf *, u_int);
275 static u_int write_txpkt_vm_wr(struct adapter *, struct sge_txq *,
276     struct fw_eth_tx_pkt_vm_wr *, struct mbuf *, u_int);
277 static int try_txpkts(struct mbuf *, struct mbuf *, struct txpkts *, u_int);
278 static int add_to_txpkts(struct mbuf *, struct txpkts *, u_int);
279 static u_int write_txpkts_wr(struct sge_txq *, struct fw_eth_tx_pkts_wr *,
280     struct mbuf *, const struct txpkts *, u_int);
281 static void write_gl_to_txd(struct sge_txq *, struct mbuf *, caddr_t *, int);
282 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int);
283 static inline void ring_eq_db(struct adapter *, struct sge_eq *, u_int);
284 static inline uint16_t read_hw_cidx(struct sge_eq *);
285 static inline u_int reclaimable_tx_desc(struct sge_eq *);
286 static inline u_int total_available_tx_desc(struct sge_eq *);
287 static u_int reclaim_tx_descs(struct sge_txq *, u_int);
288 static void tx_reclaim(void *, int);
289 static __be64 get_flit(struct sglist_seg *, int, int);
290 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
291     struct mbuf *);
292 static int handle_fw_msg(struct sge_iq *, const struct rss_header *,
293     struct mbuf *);
294 static int t4_handle_wrerr_rpl(struct adapter *, const __be64 *);
295 static void wrq_tx_drain(void *, int);
296 static void drain_wrq_wr_list(struct adapter *, struct sge_wrq *);
297 
298 static int sysctl_uint16(SYSCTL_HANDLER_ARGS);
299 static int sysctl_bufsizes(SYSCTL_HANDLER_ARGS);
300 #ifdef RATELIMIT
301 static inline u_int txpkt_eo_len16(u_int, u_int, u_int);
302 static int ethofld_fw4_ack(struct sge_iq *, const struct rss_header *,
303     struct mbuf *);
304 #endif
305 
306 static counter_u64_t extfree_refs;
307 static counter_u64_t extfree_rels;
308 
309 an_handler_t t4_an_handler;
310 fw_msg_handler_t t4_fw_msg_handler[NUM_FW6_TYPES];
311 cpl_handler_t t4_cpl_handler[NUM_CPL_CMDS];
312 cpl_handler_t set_tcb_rpl_handlers[NUM_CPL_COOKIES];
313 cpl_handler_t l2t_write_rpl_handlers[NUM_CPL_COOKIES];
314 cpl_handler_t act_open_rpl_handlers[NUM_CPL_COOKIES];
315 cpl_handler_t abort_rpl_rss_handlers[NUM_CPL_COOKIES];
316 cpl_handler_t fw4_ack_handlers[NUM_CPL_COOKIES];
317 
318 void
319 t4_register_an_handler(an_handler_t h)
320 {
321 	uintptr_t *loc;
322 
323 	MPASS(h == NULL || t4_an_handler == NULL);
324 
325 	loc = (uintptr_t *)&t4_an_handler;
326 	atomic_store_rel_ptr(loc, (uintptr_t)h);
327 }
328 
329 void
330 t4_register_fw_msg_handler(int type, fw_msg_handler_t h)
331 {
332 	uintptr_t *loc;
333 
334 	MPASS(type < nitems(t4_fw_msg_handler));
335 	MPASS(h == NULL || t4_fw_msg_handler[type] == NULL);
336 	/*
337 	 * These are dispatched by the handler for FW{4|6}_CPL_MSG using the CPL
338 	 * handler dispatch table.  Reject any attempt to install a handler for
339 	 * this subtype.
340 	 */
341 	MPASS(type != FW_TYPE_RSSCPL);
342 	MPASS(type != FW6_TYPE_RSSCPL);
343 
344 	loc = (uintptr_t *)&t4_fw_msg_handler[type];
345 	atomic_store_rel_ptr(loc, (uintptr_t)h);
346 }
347 
348 void
349 t4_register_cpl_handler(int opcode, cpl_handler_t h)
350 {
351 	uintptr_t *loc;
352 
353 	MPASS(opcode < nitems(t4_cpl_handler));
354 	MPASS(h == NULL || t4_cpl_handler[opcode] == NULL);
355 
356 	loc = (uintptr_t *)&t4_cpl_handler[opcode];
357 	atomic_store_rel_ptr(loc, (uintptr_t)h);
358 }
359 
360 static int
361 set_tcb_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
362     struct mbuf *m)
363 {
364 	const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
365 	u_int tid;
366 	int cookie;
367 
368 	MPASS(m == NULL);
369 
370 	tid = GET_TID(cpl);
371 	if (is_hpftid(iq->adapter, tid) || is_ftid(iq->adapter, tid)) {
372 		/*
373 		 * The return code for filter-write is put in the CPL cookie so
374 		 * we have to rely on the hardware tid (is_ftid) to determine
375 		 * that this is a response to a filter.
376 		 */
377 		cookie = CPL_COOKIE_FILTER;
378 	} else {
379 		cookie = G_COOKIE(cpl->cookie);
380 	}
381 	MPASS(cookie > CPL_COOKIE_RESERVED);
382 	MPASS(cookie < nitems(set_tcb_rpl_handlers));
383 
384 	return (set_tcb_rpl_handlers[cookie](iq, rss, m));
385 }
386 
387 static int
388 l2t_write_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
389     struct mbuf *m)
390 {
391 	const struct cpl_l2t_write_rpl *rpl = (const void *)(rss + 1);
392 	unsigned int cookie;
393 
394 	MPASS(m == NULL);
395 
396 	cookie = GET_TID(rpl) & F_SYNC_WR ? CPL_COOKIE_TOM : CPL_COOKIE_FILTER;
397 	return (l2t_write_rpl_handlers[cookie](iq, rss, m));
398 }
399 
400 static int
401 act_open_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
402     struct mbuf *m)
403 {
404 	const struct cpl_act_open_rpl *cpl = (const void *)(rss + 1);
405 	u_int cookie = G_TID_COOKIE(G_AOPEN_ATID(be32toh(cpl->atid_status)));
406 
407 	MPASS(m == NULL);
408 	MPASS(cookie != CPL_COOKIE_RESERVED);
409 
410 	return (act_open_rpl_handlers[cookie](iq, rss, m));
411 }
412 
413 static int
414 abort_rpl_rss_handler(struct sge_iq *iq, const struct rss_header *rss,
415     struct mbuf *m)
416 {
417 	struct adapter *sc = iq->adapter;
418 	u_int cookie;
419 
420 	MPASS(m == NULL);
421 	if (is_hashfilter(sc))
422 		cookie = CPL_COOKIE_HASHFILTER;
423 	else
424 		cookie = CPL_COOKIE_TOM;
425 
426 	return (abort_rpl_rss_handlers[cookie](iq, rss, m));
427 }
428 
429 static int
430 fw4_ack_handler(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
431 {
432 	struct adapter *sc = iq->adapter;
433 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
434 	unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
435 	u_int cookie;
436 
437 	MPASS(m == NULL);
438 	if (is_etid(sc, tid))
439 		cookie = CPL_COOKIE_ETHOFLD;
440 	else
441 		cookie = CPL_COOKIE_TOM;
442 
443 	return (fw4_ack_handlers[cookie](iq, rss, m));
444 }
445 
446 static void
447 t4_init_shared_cpl_handlers(void)
448 {
449 
450 	t4_register_cpl_handler(CPL_SET_TCB_RPL, set_tcb_rpl_handler);
451 	t4_register_cpl_handler(CPL_L2T_WRITE_RPL, l2t_write_rpl_handler);
452 	t4_register_cpl_handler(CPL_ACT_OPEN_RPL, act_open_rpl_handler);
453 	t4_register_cpl_handler(CPL_ABORT_RPL_RSS, abort_rpl_rss_handler);
454 	t4_register_cpl_handler(CPL_FW4_ACK, fw4_ack_handler);
455 }
456 
457 void
458 t4_register_shared_cpl_handler(int opcode, cpl_handler_t h, int cookie)
459 {
460 	uintptr_t *loc;
461 
462 	MPASS(opcode < nitems(t4_cpl_handler));
463 	MPASS(cookie > CPL_COOKIE_RESERVED);
464 	MPASS(cookie < NUM_CPL_COOKIES);
465 	MPASS(t4_cpl_handler[opcode] != NULL);
466 
467 	switch (opcode) {
468 	case CPL_SET_TCB_RPL:
469 		loc = (uintptr_t *)&set_tcb_rpl_handlers[cookie];
470 		break;
471 	case CPL_L2T_WRITE_RPL:
472 		loc = (uintptr_t *)&l2t_write_rpl_handlers[cookie];
473 		break;
474 	case CPL_ACT_OPEN_RPL:
475 		loc = (uintptr_t *)&act_open_rpl_handlers[cookie];
476 		break;
477 	case CPL_ABORT_RPL_RSS:
478 		loc = (uintptr_t *)&abort_rpl_rss_handlers[cookie];
479 		break;
480 	case CPL_FW4_ACK:
481 		loc = (uintptr_t *)&fw4_ack_handlers[cookie];
482 		break;
483 	default:
484 		MPASS(0);
485 		return;
486 	}
487 	MPASS(h == NULL || *loc == (uintptr_t)NULL);
488 	atomic_store_rel_ptr(loc, (uintptr_t)h);
489 }
490 
491 /*
492  * Called on MOD_LOAD.  Validates and calculates the SGE tunables.
493  */
494 void
495 t4_sge_modload(void)
496 {
497 
498 	if (fl_pktshift < 0 || fl_pktshift > 7) {
499 		printf("Invalid hw.cxgbe.fl_pktshift value (%d),"
500 		    " using 0 instead.\n", fl_pktshift);
501 		fl_pktshift = 0;
502 	}
503 
504 	if (spg_len != 64 && spg_len != 128) {
505 		int len;
506 
507 #if defined(__i386__) || defined(__amd64__)
508 		len = cpu_clflush_line_size > 64 ? 128 : 64;
509 #else
510 		len = 64;
511 #endif
512 		if (spg_len != -1) {
513 			printf("Invalid hw.cxgbe.spg_len value (%d),"
514 			    " using %d instead.\n", spg_len, len);
515 		}
516 		spg_len = len;
517 	}
518 
519 	if (cong_drop < -1 || cong_drop > 1) {
520 		printf("Invalid hw.cxgbe.cong_drop value (%d),"
521 		    " using 0 instead.\n", cong_drop);
522 		cong_drop = 0;
523 	}
524 
525 	if (tscale != 1 && (tscale < 3 || tscale > 17)) {
526 		printf("Invalid hw.cxgbe.tscale value (%d),"
527 		    " using 1 instead.\n", tscale);
528 		tscale = 1;
529 	}
530 
531 	extfree_refs = counter_u64_alloc(M_WAITOK);
532 	extfree_rels = counter_u64_alloc(M_WAITOK);
533 	counter_u64_zero(extfree_refs);
534 	counter_u64_zero(extfree_rels);
535 
536 	t4_init_shared_cpl_handlers();
537 	t4_register_cpl_handler(CPL_FW4_MSG, handle_fw_msg);
538 	t4_register_cpl_handler(CPL_FW6_MSG, handle_fw_msg);
539 	t4_register_cpl_handler(CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
540 	t4_register_cpl_handler(CPL_RX_PKT, t4_eth_rx);
541 #ifdef RATELIMIT
542 	t4_register_shared_cpl_handler(CPL_FW4_ACK, ethofld_fw4_ack,
543 	    CPL_COOKIE_ETHOFLD);
544 #endif
545 	t4_register_fw_msg_handler(FW6_TYPE_CMD_RPL, t4_handle_fw_rpl);
546 	t4_register_fw_msg_handler(FW6_TYPE_WRERR_RPL, t4_handle_wrerr_rpl);
547 }
548 
549 void
550 t4_sge_modunload(void)
551 {
552 
553 	counter_u64_free(extfree_refs);
554 	counter_u64_free(extfree_rels);
555 }
556 
557 uint64_t
558 t4_sge_extfree_refs(void)
559 {
560 	uint64_t refs, rels;
561 
562 	rels = counter_u64_fetch(extfree_rels);
563 	refs = counter_u64_fetch(extfree_refs);
564 
565 	return (refs - rels);
566 }
567 
568 static inline void
569 setup_pad_and_pack_boundaries(struct adapter *sc)
570 {
571 	uint32_t v, m;
572 	int pad, pack, pad_shift;
573 
574 	pad_shift = chip_id(sc) > CHELSIO_T5 ? X_T6_INGPADBOUNDARY_SHIFT :
575 	    X_INGPADBOUNDARY_SHIFT;
576 	pad = fl_pad;
577 	if (fl_pad < (1 << pad_shift) ||
578 	    fl_pad > (1 << (pad_shift + M_INGPADBOUNDARY)) ||
579 	    !powerof2(fl_pad)) {
580 		/*
581 		 * If there is any chance that we might use buffer packing and
582 		 * the chip is a T4, then pick 64 as the pad/pack boundary.  Set
583 		 * it to the minimum allowed in all other cases.
584 		 */
585 		pad = is_t4(sc) && buffer_packing ? 64 : 1 << pad_shift;
586 
587 		/*
588 		 * For fl_pad = 0 we'll still write a reasonable value to the
589 		 * register but all the freelists will opt out of padding.
590 		 * We'll complain here only if the user tried to set it to a
591 		 * value greater than 0 that was invalid.
592 		 */
593 		if (fl_pad > 0) {
594 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pad value"
595 			    " (%d), using %d instead.\n", fl_pad, pad);
596 		}
597 	}
598 	m = V_INGPADBOUNDARY(M_INGPADBOUNDARY);
599 	v = V_INGPADBOUNDARY(ilog2(pad) - pad_shift);
600 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
601 
602 	if (is_t4(sc)) {
603 		if (fl_pack != -1 && fl_pack != pad) {
604 			/* Complain but carry on. */
605 			device_printf(sc->dev, "hw.cxgbe.fl_pack (%d) ignored,"
606 			    " using %d instead.\n", fl_pack, pad);
607 		}
608 		return;
609 	}
610 
611 	pack = fl_pack;
612 	if (fl_pack < 16 || fl_pack == 32 || fl_pack > 4096 ||
613 	    !powerof2(fl_pack)) {
614 		pack = max(sc->params.pci.mps, CACHE_LINE_SIZE);
615 		MPASS(powerof2(pack));
616 		if (pack < 16)
617 			pack = 16;
618 		if (pack == 32)
619 			pack = 64;
620 		if (pack > 4096)
621 			pack = 4096;
622 		if (fl_pack != -1) {
623 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pack value"
624 			    " (%d), using %d instead.\n", fl_pack, pack);
625 		}
626 	}
627 	m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY);
628 	if (pack == 16)
629 		v = V_INGPACKBOUNDARY(0);
630 	else
631 		v = V_INGPACKBOUNDARY(ilog2(pack) - 5);
632 
633 	MPASS(!is_t4(sc));	/* T4 doesn't have SGE_CONTROL2 */
634 	t4_set_reg_field(sc, A_SGE_CONTROL2, m, v);
635 }
636 
637 /*
638  * adap->params.vpd.cclk must be set up before this is called.
639  */
640 void
641 t4_tweak_chip_settings(struct adapter *sc)
642 {
643 	int i;
644 	uint32_t v, m;
645 	int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200};
646 	int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk;
647 	int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */
648 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
649 	static int sge_flbuf_sizes[] = {
650 		MCLBYTES,
651 #if MJUMPAGESIZE != MCLBYTES
652 		MJUMPAGESIZE,
653 		MJUMPAGESIZE - CL_METADATA_SIZE,
654 		MJUMPAGESIZE - 2 * MSIZE - CL_METADATA_SIZE,
655 #endif
656 		MJUM9BYTES,
657 		MJUM16BYTES,
658 		MCLBYTES - MSIZE - CL_METADATA_SIZE,
659 		MJUM9BYTES - CL_METADATA_SIZE,
660 		MJUM16BYTES - CL_METADATA_SIZE,
661 	};
662 
663 	KASSERT(sc->flags & MASTER_PF,
664 	    ("%s: trying to change chip settings when not master.", __func__));
665 
666 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
667 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
668 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
669 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
670 
671 	setup_pad_and_pack_boundaries(sc);
672 
673 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
674 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
675 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
676 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
677 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
678 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
679 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
680 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
681 	t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v);
682 
683 	KASSERT(nitems(sge_flbuf_sizes) <= SGE_FLBUF_SIZES,
684 	    ("%s: hw buffer size table too big", __func__));
685 	for (i = 0; i < min(nitems(sge_flbuf_sizes), SGE_FLBUF_SIZES); i++) {
686 		t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i),
687 		    sge_flbuf_sizes[i]);
688 	}
689 
690 	v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) |
691 	    V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]);
692 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v);
693 
694 	KASSERT(intr_timer[0] <= timer_max,
695 	    ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0],
696 	    timer_max));
697 	for (i = 1; i < nitems(intr_timer); i++) {
698 		KASSERT(intr_timer[i] >= intr_timer[i - 1],
699 		    ("%s: timers not listed in increasing order (%d)",
700 		    __func__, i));
701 
702 		while (intr_timer[i] > timer_max) {
703 			if (i == nitems(intr_timer) - 1) {
704 				intr_timer[i] = timer_max;
705 				break;
706 			}
707 			intr_timer[i] += intr_timer[i - 1];
708 			intr_timer[i] /= 2;
709 		}
710 	}
711 
712 	v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) |
713 	    V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1]));
714 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v);
715 	v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) |
716 	    V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3]));
717 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v);
718 	v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) |
719 	    V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5]));
720 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v);
721 
722 	if (chip_id(sc) >= CHELSIO_T6) {
723 		m = V_TSCALE(M_TSCALE);
724 		if (tscale == 1)
725 			v = 0;
726 		else
727 			v = V_TSCALE(tscale - 2);
728 		t4_set_reg_field(sc, A_SGE_ITP_CONTROL, m, v);
729 
730 		if (sc->debug_flags & DF_DISABLE_TCB_CACHE) {
731 			m = V_RDTHRESHOLD(M_RDTHRESHOLD) | F_WRTHRTHRESHEN |
732 			    V_WRTHRTHRESH(M_WRTHRTHRESH);
733 			t4_tp_pio_read(sc, &v, 1, A_TP_CMM_CONFIG, 1);
734 			v &= ~m;
735 			v |= V_RDTHRESHOLD(1) | F_WRTHRTHRESHEN |
736 			    V_WRTHRTHRESH(16);
737 			t4_tp_pio_write(sc, &v, 1, A_TP_CMM_CONFIG, 1);
738 		}
739 	}
740 
741 	/* 4K, 16K, 64K, 256K DDP "page sizes" for TDDP */
742 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
743 	t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v);
744 
745 	/*
746 	 * 4K, 8K, 16K, 64K DDP "page sizes" for iSCSI DDP.  These have been
747 	 * chosen with MAXPHYS = 128K in mind.  The largest DDP buffer that we
748 	 * may have to deal with is MAXPHYS + 1 page.
749 	 */
750 	v = V_HPZ0(0) | V_HPZ1(1) | V_HPZ2(2) | V_HPZ3(4);
751 	t4_write_reg(sc, A_ULP_RX_ISCSI_PSZ, v);
752 
753 	/* We use multiple DDP page sizes both in plain-TOE and ISCSI modes. */
754 	m = v = F_TDDPTAGTCB | F_ISCSITAGTCB;
755 	t4_set_reg_field(sc, A_ULP_RX_CTL, m, v);
756 
757 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
758 	    F_RESETDDPOFFSET;
759 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
760 	t4_set_reg_field(sc, A_TP_PARA_REG5, m, v);
761 }
762 
763 /*
764  * SGE wants the buffer to be at least 64B and then a multiple of 16.  If
765  * padding is in use, the buffer's start and end need to be aligned to the pad
766  * boundary as well.  We'll just make sure that the size is a multiple of the
767  * boundary here, it is up to the buffer allocation code to make sure the start
768  * of the buffer is aligned as well.
769  */
770 static inline int
771 hwsz_ok(struct adapter *sc, int hwsz)
772 {
773 	int mask = fl_pad ? sc->params.sge.pad_boundary - 1 : 16 - 1;
774 
775 	return (hwsz >= 64 && (hwsz & mask) == 0);
776 }
777 
778 /*
779  * XXX: driver really should be able to deal with unexpected settings.
780  */
781 int
782 t4_read_chip_settings(struct adapter *sc)
783 {
784 	struct sge *s = &sc->sge;
785 	struct sge_params *sp = &sc->params.sge;
786 	int i, j, n, rc = 0;
787 	uint32_t m, v, r;
788 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
789 	static int sw_buf_sizes[] = {	/* Sorted by size */
790 		MCLBYTES,
791 #if MJUMPAGESIZE != MCLBYTES
792 		MJUMPAGESIZE,
793 #endif
794 		MJUM9BYTES,
795 		MJUM16BYTES
796 	};
797 	struct sw_zone_info *swz, *safe_swz;
798 	struct hw_buf_info *hwb;
799 
800 	m = F_RXPKTCPLMODE;
801 	v = F_RXPKTCPLMODE;
802 	r = sc->params.sge.sge_control;
803 	if ((r & m) != v) {
804 		device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r);
805 		rc = EINVAL;
806 	}
807 
808 	/*
809 	 * If this changes then every single use of PAGE_SHIFT in the driver
810 	 * needs to be carefully reviewed for PAGE_SHIFT vs sp->page_shift.
811 	 */
812 	if (sp->page_shift != PAGE_SHIFT) {
813 		device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r);
814 		rc = EINVAL;
815 	}
816 
817 	/* Filter out unusable hw buffer sizes entirely (mark with -2). */
818 	hwb = &s->hw_buf_info[0];
819 	for (i = 0; i < nitems(s->hw_buf_info); i++, hwb++) {
820 		r = sc->params.sge.sge_fl_buffer_size[i];
821 		hwb->size = r;
822 		hwb->zidx = hwsz_ok(sc, r) ? -1 : -2;
823 		hwb->next = -1;
824 	}
825 
826 	/*
827 	 * Create a sorted list in decreasing order of hw buffer sizes (and so
828 	 * increasing order of spare area) for each software zone.
829 	 *
830 	 * If padding is enabled then the start and end of the buffer must align
831 	 * to the pad boundary; if packing is enabled then they must align with
832 	 * the pack boundary as well.  Allocations from the cluster zones are
833 	 * aligned to min(size, 4K), so the buffer starts at that alignment and
834 	 * ends at hwb->size alignment.  If mbuf inlining is allowed the
835 	 * starting alignment will be reduced to MSIZE and the driver will
836 	 * exercise appropriate caution when deciding on the best buffer layout
837 	 * to use.
838 	 */
839 	n = 0;	/* no usable buffer size to begin with */
840 	swz = &s->sw_zone_info[0];
841 	safe_swz = NULL;
842 	for (i = 0; i < SW_ZONE_SIZES; i++, swz++) {
843 		int8_t head = -1, tail = -1;
844 
845 		swz->size = sw_buf_sizes[i];
846 		swz->zone = m_getzone(swz->size);
847 		swz->type = m_gettype(swz->size);
848 
849 		if (swz->size < PAGE_SIZE) {
850 			MPASS(powerof2(swz->size));
851 			if (fl_pad && (swz->size % sp->pad_boundary != 0))
852 				continue;
853 		}
854 
855 		if (swz->size == safest_rx_cluster)
856 			safe_swz = swz;
857 
858 		hwb = &s->hw_buf_info[0];
859 		for (j = 0; j < SGE_FLBUF_SIZES; j++, hwb++) {
860 			if (hwb->zidx != -1 || hwb->size > swz->size)
861 				continue;
862 #ifdef INVARIANTS
863 			if (fl_pad)
864 				MPASS(hwb->size % sp->pad_boundary == 0);
865 #endif
866 			hwb->zidx = i;
867 			if (head == -1)
868 				head = tail = j;
869 			else if (hwb->size < s->hw_buf_info[tail].size) {
870 				s->hw_buf_info[tail].next = j;
871 				tail = j;
872 			} else {
873 				int8_t *cur;
874 				struct hw_buf_info *t;
875 
876 				for (cur = &head; *cur != -1; cur = &t->next) {
877 					t = &s->hw_buf_info[*cur];
878 					if (hwb->size == t->size) {
879 						hwb->zidx = -2;
880 						break;
881 					}
882 					if (hwb->size > t->size) {
883 						hwb->next = *cur;
884 						*cur = j;
885 						break;
886 					}
887 				}
888 			}
889 		}
890 		swz->head_hwidx = head;
891 		swz->tail_hwidx = tail;
892 
893 		if (tail != -1) {
894 			n++;
895 			if (swz->size - s->hw_buf_info[tail].size >=
896 			    CL_METADATA_SIZE)
897 				sc->flags |= BUF_PACKING_OK;
898 		}
899 	}
900 	if (n == 0) {
901 		device_printf(sc->dev, "no usable SGE FL buffer size.\n");
902 		rc = EINVAL;
903 	}
904 
905 	s->safe_hwidx1 = -1;
906 	s->safe_hwidx2 = -1;
907 	if (safe_swz != NULL) {
908 		s->safe_hwidx1 = safe_swz->head_hwidx;
909 		for (i = safe_swz->head_hwidx; i != -1; i = hwb->next) {
910 			int spare;
911 
912 			hwb = &s->hw_buf_info[i];
913 #ifdef INVARIANTS
914 			if (fl_pad)
915 				MPASS(hwb->size % sp->pad_boundary == 0);
916 #endif
917 			spare = safe_swz->size - hwb->size;
918 			if (spare >= CL_METADATA_SIZE) {
919 				s->safe_hwidx2 = i;
920 				break;
921 			}
922 		}
923 	}
924 
925 	if (sc->flags & IS_VF)
926 		return (0);
927 
928 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
929 	r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ);
930 	if (r != v) {
931 		device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r);
932 		rc = EINVAL;
933 	}
934 
935 	m = v = F_TDDPTAGTCB;
936 	r = t4_read_reg(sc, A_ULP_RX_CTL);
937 	if ((r & m) != v) {
938 		device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r);
939 		rc = EINVAL;
940 	}
941 
942 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
943 	    F_RESETDDPOFFSET;
944 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
945 	r = t4_read_reg(sc, A_TP_PARA_REG5);
946 	if ((r & m) != v) {
947 		device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r);
948 		rc = EINVAL;
949 	}
950 
951 	t4_init_tp_params(sc, 1);
952 
953 	t4_read_mtu_tbl(sc, sc->params.mtus, NULL);
954 	t4_load_mtus(sc, sc->params.mtus, sc->params.a_wnd, sc->params.b_wnd);
955 
956 	return (rc);
957 }
958 
959 int
960 t4_create_dma_tag(struct adapter *sc)
961 {
962 	int rc;
963 
964 	rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
965 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
966 	    BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
967 	    NULL, &sc->dmat);
968 	if (rc != 0) {
969 		device_printf(sc->dev,
970 		    "failed to create main DMA tag: %d\n", rc);
971 	}
972 
973 	return (rc);
974 }
975 
976 void
977 t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
978     struct sysctl_oid_list *children)
979 {
980 	struct sge_params *sp = &sc->params.sge;
981 
982 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "buffer_sizes",
983 	    CTLTYPE_STRING | CTLFLAG_RD, &sc->sge, 0, sysctl_bufsizes, "A",
984 	    "freelist buffer sizes");
985 
986 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD,
987 	    NULL, sp->fl_pktshift, "payload DMA offset in rx buffer (bytes)");
988 
989 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD,
990 	    NULL, sp->pad_boundary, "payload pad boundary (bytes)");
991 
992 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD,
993 	    NULL, sp->spg_len, "status page size (bytes)");
994 
995 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD,
996 	    NULL, cong_drop, "congestion drop setting");
997 
998 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD,
999 	    NULL, sp->pack_boundary, "payload pack boundary (bytes)");
1000 }
1001 
1002 int
1003 t4_destroy_dma_tag(struct adapter *sc)
1004 {
1005 	if (sc->dmat)
1006 		bus_dma_tag_destroy(sc->dmat);
1007 
1008 	return (0);
1009 }
1010 
1011 /*
1012  * Allocate and initialize the firmware event queue, control queues, and special
1013  * purpose rx queues owned by the adapter.
1014  *
1015  * Returns errno on failure.  Resources allocated up to that point may still be
1016  * allocated.  Caller is responsible for cleanup in case this function fails.
1017  */
1018 int
1019 t4_setup_adapter_queues(struct adapter *sc)
1020 {
1021 	struct sysctl_oid *oid;
1022 	struct sysctl_oid_list *children;
1023 	int rc, i;
1024 
1025 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
1026 
1027 	sysctl_ctx_init(&sc->ctx);
1028 	sc->flags |= ADAP_SYSCTL_CTX;
1029 
1030 	/*
1031 	 * Firmware event queue
1032 	 */
1033 	rc = alloc_fwq(sc);
1034 	if (rc != 0)
1035 		return (rc);
1036 
1037 	/*
1038 	 * That's all for the VF driver.
1039 	 */
1040 	if (sc->flags & IS_VF)
1041 		return (rc);
1042 
1043 	oid = device_get_sysctl_tree(sc->dev);
1044 	children = SYSCTL_CHILDREN(oid);
1045 
1046 	/*
1047 	 * XXX: General purpose rx queues, one per port.
1048 	 */
1049 
1050 	/*
1051 	 * Control queues, one per port.
1052 	 */
1053 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "ctrlq",
1054 	    CTLFLAG_RD, NULL, "control queues");
1055 	for_each_port(sc, i) {
1056 		struct sge_wrq *ctrlq = &sc->sge.ctrlq[i];
1057 
1058 		rc = alloc_ctrlq(sc, ctrlq, i, oid);
1059 		if (rc != 0)
1060 			return (rc);
1061 	}
1062 
1063 	return (rc);
1064 }
1065 
1066 /*
1067  * Idempotent
1068  */
1069 int
1070 t4_teardown_adapter_queues(struct adapter *sc)
1071 {
1072 	int i;
1073 
1074 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
1075 
1076 	/* Do this before freeing the queue */
1077 	if (sc->flags & ADAP_SYSCTL_CTX) {
1078 		sysctl_ctx_free(&sc->ctx);
1079 		sc->flags &= ~ADAP_SYSCTL_CTX;
1080 	}
1081 
1082 	for_each_port(sc, i)
1083 		free_wrq(sc, &sc->sge.ctrlq[i]);
1084 	free_fwq(sc);
1085 
1086 	return (0);
1087 }
1088 
1089 /* Maximum payload that can be delivered with a single iq descriptor */
1090 static inline int
1091 mtu_to_max_payload(struct adapter *sc, int mtu, const int toe)
1092 {
1093 	int payload;
1094 
1095 #ifdef TCP_OFFLOAD
1096 	if (toe) {
1097 		int rxcs = G_RXCOALESCESIZE(t4_read_reg(sc, A_TP_PARA_REG2));
1098 
1099 		/* Note that COP can set rx_coalesce on/off per connection. */
1100 		payload = max(mtu, rxcs);
1101 	} else {
1102 #endif
1103 		/* large enough even when hw VLAN extraction is disabled */
1104 		payload = sc->params.sge.fl_pktshift + ETHER_HDR_LEN +
1105 		    ETHER_VLAN_ENCAP_LEN + mtu;
1106 #ifdef TCP_OFFLOAD
1107 	}
1108 #endif
1109 
1110 	return (payload);
1111 }
1112 
1113 int
1114 t4_setup_vi_queues(struct vi_info *vi)
1115 {
1116 	int rc = 0, i, intr_idx, iqidx;
1117 	struct sge_rxq *rxq;
1118 	struct sge_txq *txq;
1119 #ifdef TCP_OFFLOAD
1120 	struct sge_ofld_rxq *ofld_rxq;
1121 #endif
1122 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1123 	struct sge_wrq *ofld_txq;
1124 #endif
1125 #ifdef DEV_NETMAP
1126 	int saved_idx;
1127 	struct sge_nm_rxq *nm_rxq;
1128 	struct sge_nm_txq *nm_txq;
1129 #endif
1130 	char name[16];
1131 	struct port_info *pi = vi->pi;
1132 	struct adapter *sc = pi->adapter;
1133 	struct ifnet *ifp = vi->ifp;
1134 	struct sysctl_oid *oid = device_get_sysctl_tree(vi->dev);
1135 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
1136 	int maxp, mtu = ifp->if_mtu;
1137 
1138 	/* Interrupt vector to start from (when using multiple vectors) */
1139 	intr_idx = vi->first_intr;
1140 
1141 #ifdef DEV_NETMAP
1142 	saved_idx = intr_idx;
1143 	if (ifp->if_capabilities & IFCAP_NETMAP) {
1144 
1145 		/* netmap is supported with direct interrupts only. */
1146 		MPASS(!forwarding_intr_to_fwq(sc));
1147 
1148 		/*
1149 		 * We don't have buffers to back the netmap rx queues
1150 		 * right now so we create the queues in a way that
1151 		 * doesn't set off any congestion signal in the chip.
1152 		 */
1153 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_rxq",
1154 		    CTLFLAG_RD, NULL, "rx queues");
1155 		for_each_nm_rxq(vi, i, nm_rxq) {
1156 			rc = alloc_nm_rxq(vi, nm_rxq, intr_idx, i, oid);
1157 			if (rc != 0)
1158 				goto done;
1159 			intr_idx++;
1160 		}
1161 
1162 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_txq",
1163 		    CTLFLAG_RD, NULL, "tx queues");
1164 		for_each_nm_txq(vi, i, nm_txq) {
1165 			iqidx = vi->first_nm_rxq + (i % vi->nnmrxq);
1166 			rc = alloc_nm_txq(vi, nm_txq, iqidx, i, oid);
1167 			if (rc != 0)
1168 				goto done;
1169 		}
1170 	}
1171 
1172 	/* Normal rx queues and netmap rx queues share the same interrupts. */
1173 	intr_idx = saved_idx;
1174 #endif
1175 
1176 	/*
1177 	 * Allocate rx queues first because a default iqid is required when
1178 	 * creating a tx queue.
1179 	 */
1180 	maxp = mtu_to_max_payload(sc, mtu, 0);
1181 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
1182 	    CTLFLAG_RD, NULL, "rx queues");
1183 	for_each_rxq(vi, i, rxq) {
1184 
1185 		init_iq(&rxq->iq, sc, vi->tmr_idx, vi->pktc_idx, vi->qsize_rxq);
1186 
1187 		snprintf(name, sizeof(name), "%s rxq%d-fl",
1188 		    device_get_nameunit(vi->dev), i);
1189 		init_fl(sc, &rxq->fl, vi->qsize_rxq / 8, maxp, name);
1190 
1191 		rc = alloc_rxq(vi, rxq,
1192 		    forwarding_intr_to_fwq(sc) ? -1 : intr_idx, i, oid);
1193 		if (rc != 0)
1194 			goto done;
1195 		intr_idx++;
1196 	}
1197 #ifdef DEV_NETMAP
1198 	if (ifp->if_capabilities & IFCAP_NETMAP)
1199 		intr_idx = saved_idx + max(vi->nrxq, vi->nnmrxq);
1200 #endif
1201 #ifdef TCP_OFFLOAD
1202 	maxp = mtu_to_max_payload(sc, mtu, 1);
1203 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq",
1204 	    CTLFLAG_RD, NULL, "rx queues for offloaded TCP connections");
1205 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1206 
1207 		init_iq(&ofld_rxq->iq, sc, vi->ofld_tmr_idx, vi->ofld_pktc_idx,
1208 		    vi->qsize_rxq);
1209 
1210 		snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
1211 		    device_get_nameunit(vi->dev), i);
1212 		init_fl(sc, &ofld_rxq->fl, vi->qsize_rxq / 8, maxp, name);
1213 
1214 		rc = alloc_ofld_rxq(vi, ofld_rxq,
1215 		    forwarding_intr_to_fwq(sc) ? -1 : intr_idx, i, oid);
1216 		if (rc != 0)
1217 			goto done;
1218 		intr_idx++;
1219 	}
1220 #endif
1221 
1222 	/*
1223 	 * Now the tx queues.
1224 	 */
1225 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "txq", CTLFLAG_RD,
1226 	    NULL, "tx queues");
1227 	for_each_txq(vi, i, txq) {
1228 		iqidx = vi->first_rxq + (i % vi->nrxq);
1229 		snprintf(name, sizeof(name), "%s txq%d",
1230 		    device_get_nameunit(vi->dev), i);
1231 		init_eq(sc, &txq->eq, EQ_ETH, vi->qsize_txq, pi->tx_chan,
1232 		    sc->sge.rxq[iqidx].iq.cntxt_id, name);
1233 
1234 		rc = alloc_txq(vi, txq, i, oid);
1235 		if (rc != 0)
1236 			goto done;
1237 	}
1238 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1239 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_txq",
1240 	    CTLFLAG_RD, NULL, "tx queues for TOE/ETHOFLD");
1241 	for_each_ofld_txq(vi, i, ofld_txq) {
1242 		struct sysctl_oid *oid2;
1243 
1244 		snprintf(name, sizeof(name), "%s ofld_txq%d",
1245 		    device_get_nameunit(vi->dev), i);
1246 #ifdef TCP_OFFLOAD
1247 		iqidx = vi->first_ofld_rxq + (i % vi->nofldrxq);
1248 		init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq, pi->tx_chan,
1249 		    sc->sge.ofld_rxq[iqidx].iq.cntxt_id, name);
1250 #else
1251 		iqidx = vi->first_rxq + (i % vi->nrxq);
1252 		init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq, pi->tx_chan,
1253 		    sc->sge.rxq[iqidx].iq.cntxt_id, name);
1254 #endif
1255 
1256 		snprintf(name, sizeof(name), "%d", i);
1257 		oid2 = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
1258 		    name, CTLFLAG_RD, NULL, "offload tx queue");
1259 
1260 		rc = alloc_wrq(sc, vi, ofld_txq, oid2);
1261 		if (rc != 0)
1262 			goto done;
1263 	}
1264 #endif
1265 done:
1266 	if (rc)
1267 		t4_teardown_vi_queues(vi);
1268 
1269 	return (rc);
1270 }
1271 
1272 /*
1273  * Idempotent
1274  */
1275 int
1276 t4_teardown_vi_queues(struct vi_info *vi)
1277 {
1278 	int i;
1279 	struct sge_rxq *rxq;
1280 	struct sge_txq *txq;
1281 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1282 	struct port_info *pi = vi->pi;
1283 	struct adapter *sc = pi->adapter;
1284 	struct sge_wrq *ofld_txq;
1285 #endif
1286 #ifdef TCP_OFFLOAD
1287 	struct sge_ofld_rxq *ofld_rxq;
1288 #endif
1289 #ifdef DEV_NETMAP
1290 	struct sge_nm_rxq *nm_rxq;
1291 	struct sge_nm_txq *nm_txq;
1292 #endif
1293 
1294 	/* Do this before freeing the queues */
1295 	if (vi->flags & VI_SYSCTL_CTX) {
1296 		sysctl_ctx_free(&vi->ctx);
1297 		vi->flags &= ~VI_SYSCTL_CTX;
1298 	}
1299 
1300 #ifdef DEV_NETMAP
1301 	if (vi->ifp->if_capabilities & IFCAP_NETMAP) {
1302 		for_each_nm_txq(vi, i, nm_txq) {
1303 			free_nm_txq(vi, nm_txq);
1304 		}
1305 
1306 		for_each_nm_rxq(vi, i, nm_rxq) {
1307 			free_nm_rxq(vi, nm_rxq);
1308 		}
1309 	}
1310 #endif
1311 
1312 	/*
1313 	 * Take down all the tx queues first, as they reference the rx queues
1314 	 * (for egress updates, etc.).
1315 	 */
1316 
1317 	for_each_txq(vi, i, txq) {
1318 		free_txq(vi, txq);
1319 	}
1320 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1321 	for_each_ofld_txq(vi, i, ofld_txq) {
1322 		free_wrq(sc, ofld_txq);
1323 	}
1324 #endif
1325 
1326 	/*
1327 	 * Then take down the rx queues.
1328 	 */
1329 
1330 	for_each_rxq(vi, i, rxq) {
1331 		free_rxq(vi, rxq);
1332 	}
1333 #ifdef TCP_OFFLOAD
1334 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1335 		free_ofld_rxq(vi, ofld_rxq);
1336 	}
1337 #endif
1338 
1339 	return (0);
1340 }
1341 
1342 /*
1343  * Interrupt handler when the driver is using only 1 interrupt.  This is a very
1344  * unusual scenario.
1345  *
1346  * a) Deals with errors, if any.
1347  * b) Services firmware event queue, which is taking interrupts for all other
1348  *    queues.
1349  */
1350 void
1351 t4_intr_all(void *arg)
1352 {
1353 	struct adapter *sc = arg;
1354 	struct sge_iq *fwq = &sc->sge.fwq;
1355 
1356 	MPASS(sc->intr_count == 1);
1357 
1358 	t4_intr_err(arg);
1359 	t4_intr_evt(fwq);
1360 }
1361 
1362 /*
1363  * Interrupt handler for errors (installed directly when multiple interrupts are
1364  * being used, or called by t4_intr_all).
1365  */
1366 void
1367 t4_intr_err(void *arg)
1368 {
1369 	struct adapter *sc = arg;
1370 
1371 	t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
1372 	t4_slow_intr_handler(sc);
1373 }
1374 
1375 /*
1376  * Interrupt handler for iq-only queues.  The firmware event queue is the only
1377  * such queue right now.
1378  */
1379 void
1380 t4_intr_evt(void *arg)
1381 {
1382 	struct sge_iq *iq = arg;
1383 
1384 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1385 		service_iq(iq, 0);
1386 		atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1387 	}
1388 }
1389 
1390 /*
1391  * Interrupt handler for iq+fl queues.
1392  */
1393 void
1394 t4_intr(void *arg)
1395 {
1396 	struct sge_iq *iq = arg;
1397 
1398 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1399 		service_iq_fl(iq, 0);
1400 		atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1401 	}
1402 }
1403 
1404 #ifdef DEV_NETMAP
1405 /*
1406  * Interrupt handler for netmap rx queues.
1407  */
1408 void
1409 t4_nm_intr(void *arg)
1410 {
1411 	struct sge_nm_rxq *nm_rxq = arg;
1412 
1413 	if (atomic_cmpset_int(&nm_rxq->nm_state, NM_ON, NM_BUSY)) {
1414 		service_nm_rxq(nm_rxq);
1415 		atomic_cmpset_int(&nm_rxq->nm_state, NM_BUSY, NM_ON);
1416 	}
1417 }
1418 
1419 /*
1420  * Interrupt handler for vectors shared between NIC and netmap rx queues.
1421  */
1422 void
1423 t4_vi_intr(void *arg)
1424 {
1425 	struct irq *irq = arg;
1426 
1427 	MPASS(irq->nm_rxq != NULL);
1428 	t4_nm_intr(irq->nm_rxq);
1429 
1430 	MPASS(irq->rxq != NULL);
1431 	t4_intr(irq->rxq);
1432 }
1433 #endif
1434 
1435 /*
1436  * Deals with interrupts on an iq-only (no freelist) queue.
1437  */
1438 static int
1439 service_iq(struct sge_iq *iq, int budget)
1440 {
1441 	struct sge_iq *q;
1442 	struct adapter *sc = iq->adapter;
1443 	struct iq_desc *d = &iq->desc[iq->cidx];
1444 	int ndescs = 0, limit;
1445 	int rsp_type;
1446 	uint32_t lq;
1447 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1448 
1449 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1450 	KASSERT((iq->flags & IQ_HAS_FL) == 0,
1451 	    ("%s: called for iq %p with fl (iq->flags 0x%x)", __func__, iq,
1452 	    iq->flags));
1453 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1454 	MPASS((iq->flags & IQ_LRO_ENABLED) == 0);
1455 
1456 	limit = budget ? budget : iq->qsize / 16;
1457 
1458 	/*
1459 	 * We always come back and check the descriptor ring for new indirect
1460 	 * interrupts and other responses after running a single handler.
1461 	 */
1462 	for (;;) {
1463 		while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1464 
1465 			rmb();
1466 
1467 			rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1468 			lq = be32toh(d->rsp.pldbuflen_qid);
1469 
1470 			switch (rsp_type) {
1471 			case X_RSPD_TYPE_FLBUF:
1472 				panic("%s: data for an iq (%p) with no freelist",
1473 				    __func__, iq);
1474 
1475 				/* NOTREACHED */
1476 
1477 			case X_RSPD_TYPE_CPL:
1478 				KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1479 				    ("%s: bad opcode %02x.", __func__,
1480 				    d->rss.opcode));
1481 				t4_cpl_handler[d->rss.opcode](iq, &d->rss, NULL);
1482 				break;
1483 
1484 			case X_RSPD_TYPE_INTR:
1485 				/*
1486 				 * There are 1K interrupt-capable queues (qids 0
1487 				 * through 1023).  A response type indicating a
1488 				 * forwarded interrupt with a qid >= 1K is an
1489 				 * iWARP async notification.
1490 				 */
1491 				if (__predict_true(lq >= 1024)) {
1492 					t4_an_handler(iq, &d->rsp);
1493 					break;
1494 				}
1495 
1496 				q = sc->sge.iqmap[lq - sc->sge.iq_start -
1497 				    sc->sge.iq_base];
1498 				if (atomic_cmpset_int(&q->state, IQS_IDLE,
1499 				    IQS_BUSY)) {
1500 					if (service_iq_fl(q, q->qsize / 16) == 0) {
1501 						atomic_cmpset_int(&q->state,
1502 						    IQS_BUSY, IQS_IDLE);
1503 					} else {
1504 						STAILQ_INSERT_TAIL(&iql, q,
1505 						    link);
1506 					}
1507 				}
1508 				break;
1509 
1510 			default:
1511 				KASSERT(0,
1512 				    ("%s: illegal response type %d on iq %p",
1513 				    __func__, rsp_type, iq));
1514 				log(LOG_ERR,
1515 				    "%s: illegal response type %d on iq %p",
1516 				    device_get_nameunit(sc->dev), rsp_type, iq);
1517 				break;
1518 			}
1519 
1520 			d++;
1521 			if (__predict_false(++iq->cidx == iq->sidx)) {
1522 				iq->cidx = 0;
1523 				iq->gen ^= F_RSPD_GEN;
1524 				d = &iq->desc[0];
1525 			}
1526 			if (__predict_false(++ndescs == limit)) {
1527 				t4_write_reg(sc, sc->sge_gts_reg,
1528 				    V_CIDXINC(ndescs) |
1529 				    V_INGRESSQID(iq->cntxt_id) |
1530 				    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1531 				ndescs = 0;
1532 
1533 				if (budget) {
1534 					return (EINPROGRESS);
1535 				}
1536 			}
1537 		}
1538 
1539 		if (STAILQ_EMPTY(&iql))
1540 			break;
1541 
1542 		/*
1543 		 * Process the head only, and send it to the back of the list if
1544 		 * it's still not done.
1545 		 */
1546 		q = STAILQ_FIRST(&iql);
1547 		STAILQ_REMOVE_HEAD(&iql, link);
1548 		if (service_iq_fl(q, q->qsize / 8) == 0)
1549 			atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1550 		else
1551 			STAILQ_INSERT_TAIL(&iql, q, link);
1552 	}
1553 
1554 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1555 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1556 
1557 	return (0);
1558 }
1559 
1560 static inline int
1561 sort_before_lro(struct lro_ctrl *lro)
1562 {
1563 
1564 	return (lro->lro_mbuf_max != 0);
1565 }
1566 
1567 /*
1568  * Deals with interrupts on an iq+fl queue.
1569  */
1570 static int
1571 service_iq_fl(struct sge_iq *iq, int budget)
1572 {
1573 	struct sge_rxq *rxq = iq_to_rxq(iq);
1574 	struct sge_fl *fl;
1575 	struct adapter *sc = iq->adapter;
1576 	struct iq_desc *d = &iq->desc[iq->cidx];
1577 	int ndescs = 0, limit;
1578 	int rsp_type, refill, starved;
1579 	uint32_t lq;
1580 	uint16_t fl_hw_cidx;
1581 	struct mbuf *m0;
1582 #if defined(INET) || defined(INET6)
1583 	const struct timeval lro_timeout = {0, sc->lro_timeout};
1584 	struct lro_ctrl *lro = &rxq->lro;
1585 #endif
1586 
1587 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1588 	MPASS(iq->flags & IQ_HAS_FL);
1589 
1590 	limit = budget ? budget : iq->qsize / 16;
1591 	fl = &rxq->fl;
1592 	fl_hw_cidx = fl->hw_cidx;	/* stable snapshot */
1593 
1594 #if defined(INET) || defined(INET6)
1595 	if (iq->flags & IQ_ADJ_CREDIT) {
1596 		MPASS(sort_before_lro(lro));
1597 		iq->flags &= ~IQ_ADJ_CREDIT;
1598 		if ((d->rsp.u.type_gen & F_RSPD_GEN) != iq->gen) {
1599 			tcp_lro_flush_all(lro);
1600 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(1) |
1601 			    V_INGRESSQID((u32)iq->cntxt_id) |
1602 			    V_SEINTARM(iq->intr_params));
1603 			return (0);
1604 		}
1605 		ndescs = 1;
1606 	}
1607 #else
1608 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1609 #endif
1610 
1611 	while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1612 
1613 		rmb();
1614 
1615 		refill = 0;
1616 		m0 = NULL;
1617 		rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1618 		lq = be32toh(d->rsp.pldbuflen_qid);
1619 
1620 		switch (rsp_type) {
1621 		case X_RSPD_TYPE_FLBUF:
1622 
1623 			m0 = get_fl_payload(sc, fl, lq);
1624 			if (__predict_false(m0 == NULL))
1625 				goto out;
1626 			refill = IDXDIFF(fl->hw_cidx, fl_hw_cidx, fl->sidx) > 2;
1627 #ifdef T4_PKT_TIMESTAMP
1628 			/*
1629 			 * 60 bit timestamp for the payload is
1630 			 * *(uint64_t *)m0->m_pktdat.  Note that it is
1631 			 * in the leading free-space in the mbuf.  The
1632 			 * kernel can clobber it during a pullup,
1633 			 * m_copymdata, etc.  You need to make sure that
1634 			 * the mbuf reaches you unmolested if you care
1635 			 * about the timestamp.
1636 			 */
1637 			*(uint64_t *)m0->m_pktdat =
1638 			    be64toh(ctrl->u.last_flit) & 0xfffffffffffffff;
1639 #endif
1640 
1641 			/* fall through */
1642 
1643 		case X_RSPD_TYPE_CPL:
1644 			KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1645 			    ("%s: bad opcode %02x.", __func__, d->rss.opcode));
1646 			t4_cpl_handler[d->rss.opcode](iq, &d->rss, m0);
1647 			break;
1648 
1649 		case X_RSPD_TYPE_INTR:
1650 
1651 			/*
1652 			 * There are 1K interrupt-capable queues (qids 0
1653 			 * through 1023).  A response type indicating a
1654 			 * forwarded interrupt with a qid >= 1K is an
1655 			 * iWARP async notification.  That is the only
1656 			 * acceptable indirect interrupt on this queue.
1657 			 */
1658 			if (__predict_false(lq < 1024)) {
1659 				panic("%s: indirect interrupt on iq_fl %p "
1660 				    "with qid %u", __func__, iq, lq);
1661 			}
1662 
1663 			t4_an_handler(iq, &d->rsp);
1664 			break;
1665 
1666 		default:
1667 			KASSERT(0, ("%s: illegal response type %d on iq %p",
1668 			    __func__, rsp_type, iq));
1669 			log(LOG_ERR, "%s: illegal response type %d on iq %p",
1670 			    device_get_nameunit(sc->dev), rsp_type, iq);
1671 			break;
1672 		}
1673 
1674 		d++;
1675 		if (__predict_false(++iq->cidx == iq->sidx)) {
1676 			iq->cidx = 0;
1677 			iq->gen ^= F_RSPD_GEN;
1678 			d = &iq->desc[0];
1679 		}
1680 		if (__predict_false(++ndescs == limit)) {
1681 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1682 			    V_INGRESSQID(iq->cntxt_id) |
1683 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1684 			ndescs = 0;
1685 
1686 #if defined(INET) || defined(INET6)
1687 			if (iq->flags & IQ_LRO_ENABLED &&
1688 			    !sort_before_lro(lro) &&
1689 			    sc->lro_timeout != 0) {
1690 				tcp_lro_flush_inactive(lro, &lro_timeout);
1691 			}
1692 #endif
1693 			if (budget) {
1694 				FL_LOCK(fl);
1695 				refill_fl(sc, fl, 32);
1696 				FL_UNLOCK(fl);
1697 
1698 				return (EINPROGRESS);
1699 			}
1700 		}
1701 		if (refill) {
1702 			FL_LOCK(fl);
1703 			refill_fl(sc, fl, 32);
1704 			FL_UNLOCK(fl);
1705 			fl_hw_cidx = fl->hw_cidx;
1706 		}
1707 	}
1708 out:
1709 #if defined(INET) || defined(INET6)
1710 	if (iq->flags & IQ_LRO_ENABLED) {
1711 		if (ndescs > 0 && lro->lro_mbuf_count > 8) {
1712 			MPASS(sort_before_lro(lro));
1713 			/* hold back one credit and don't flush LRO state */
1714 			iq->flags |= IQ_ADJ_CREDIT;
1715 			ndescs--;
1716 		} else {
1717 			tcp_lro_flush_all(lro);
1718 		}
1719 	}
1720 #endif
1721 
1722 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1723 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1724 
1725 	FL_LOCK(fl);
1726 	starved = refill_fl(sc, fl, 64);
1727 	FL_UNLOCK(fl);
1728 	if (__predict_false(starved != 0))
1729 		add_fl_to_sfl(sc, fl);
1730 
1731 	return (0);
1732 }
1733 
1734 static inline int
1735 cl_has_metadata(struct sge_fl *fl, struct cluster_layout *cll)
1736 {
1737 	int rc = fl->flags & FL_BUF_PACKING || cll->region1 > 0;
1738 
1739 	if (rc)
1740 		MPASS(cll->region3 >= CL_METADATA_SIZE);
1741 
1742 	return (rc);
1743 }
1744 
1745 static inline struct cluster_metadata *
1746 cl_metadata(struct adapter *sc, struct sge_fl *fl, struct cluster_layout *cll,
1747     caddr_t cl)
1748 {
1749 
1750 	if (cl_has_metadata(fl, cll)) {
1751 		struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1752 
1753 		return ((struct cluster_metadata *)(cl + swz->size) - 1);
1754 	}
1755 	return (NULL);
1756 }
1757 
1758 static void
1759 rxb_free(struct mbuf *m)
1760 {
1761 	uma_zone_t zone = m->m_ext.ext_arg1;
1762 	void *cl = m->m_ext.ext_arg2;
1763 
1764 	uma_zfree(zone, cl);
1765 	counter_u64_add(extfree_rels, 1);
1766 }
1767 
1768 /*
1769  * The mbuf returned by this function could be allocated from zone_mbuf or
1770  * constructed in spare room in the cluster.
1771  *
1772  * The mbuf carries the payload in one of these ways
1773  * a) frame inside the mbuf (mbuf from zone_mbuf)
1774  * b) m_cljset (for clusters without metadata) zone_mbuf
1775  * c) m_extaddref (cluster with metadata) inline mbuf
1776  * d) m_extaddref (cluster with metadata) zone_mbuf
1777  */
1778 static struct mbuf *
1779 get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1780     int remaining)
1781 {
1782 	struct mbuf *m;
1783 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1784 	struct cluster_layout *cll = &sd->cll;
1785 	struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1786 	struct hw_buf_info *hwb = &sc->sge.hw_buf_info[cll->hwidx];
1787 	struct cluster_metadata *clm = cl_metadata(sc, fl, cll, sd->cl);
1788 	int len, blen;
1789 	caddr_t payload;
1790 
1791 	blen = hwb->size - fl->rx_offset;	/* max possible in this buf */
1792 	len = min(remaining, blen);
1793 	payload = sd->cl + cll->region1 + fl->rx_offset;
1794 	if (fl->flags & FL_BUF_PACKING) {
1795 		const u_int l = fr_offset + len;
1796 		const u_int pad = roundup2(l, fl->buf_boundary) - l;
1797 
1798 		if (fl->rx_offset + len + pad < hwb->size)
1799 			blen = len + pad;
1800 		MPASS(fl->rx_offset + blen <= hwb->size);
1801 	} else {
1802 		MPASS(fl->rx_offset == 0);	/* not packing */
1803 	}
1804 
1805 
1806 	if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) {
1807 
1808 		/*
1809 		 * Copy payload into a freshly allocated mbuf.
1810 		 */
1811 
1812 		m = fr_offset == 0 ?
1813 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1814 		if (m == NULL)
1815 			return (NULL);
1816 		fl->mbuf_allocated++;
1817 #ifdef T4_PKT_TIMESTAMP
1818 		/* Leave room for a timestamp */
1819 		m->m_data += 8;
1820 #endif
1821 		/* copy data to mbuf */
1822 		bcopy(payload, mtod(m, caddr_t), len);
1823 
1824 	} else if (sd->nmbuf * MSIZE < cll->region1) {
1825 
1826 		/*
1827 		 * There's spare room in the cluster for an mbuf.  Create one
1828 		 * and associate it with the payload that's in the cluster.
1829 		 */
1830 
1831 		MPASS(clm != NULL);
1832 		m = (struct mbuf *)(sd->cl + sd->nmbuf * MSIZE);
1833 		/* No bzero required */
1834 		if (m_init(m, M_NOWAIT, MT_DATA,
1835 		    fr_offset == 0 ? M_PKTHDR | M_NOFREE : M_NOFREE))
1836 			return (NULL);
1837 		fl->mbuf_inlined++;
1838 		m_extaddref(m, payload, blen, &clm->refcount, rxb_free,
1839 		    swz->zone, sd->cl);
1840 		if (sd->nmbuf++ == 0)
1841 			counter_u64_add(extfree_refs, 1);
1842 
1843 	} else {
1844 
1845 		/*
1846 		 * Grab an mbuf from zone_mbuf and associate it with the
1847 		 * payload in the cluster.
1848 		 */
1849 
1850 		m = fr_offset == 0 ?
1851 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1852 		if (m == NULL)
1853 			return (NULL);
1854 		fl->mbuf_allocated++;
1855 		if (clm != NULL) {
1856 			m_extaddref(m, payload, blen, &clm->refcount,
1857 			    rxb_free, swz->zone, sd->cl);
1858 			if (sd->nmbuf++ == 0)
1859 				counter_u64_add(extfree_refs, 1);
1860 		} else {
1861 			m_cljset(m, sd->cl, swz->type);
1862 			sd->cl = NULL;	/* consumed, not a recycle candidate */
1863 		}
1864 	}
1865 	if (fr_offset == 0)
1866 		m->m_pkthdr.len = remaining;
1867 	m->m_len = len;
1868 
1869 	if (fl->flags & FL_BUF_PACKING) {
1870 		fl->rx_offset += blen;
1871 		MPASS(fl->rx_offset <= hwb->size);
1872 		if (fl->rx_offset < hwb->size)
1873 			return (m);	/* without advancing the cidx */
1874 	}
1875 
1876 	if (__predict_false(++fl->cidx % 8 == 0)) {
1877 		uint16_t cidx = fl->cidx / 8;
1878 
1879 		if (__predict_false(cidx == fl->sidx))
1880 			fl->cidx = cidx = 0;
1881 		fl->hw_cidx = cidx;
1882 	}
1883 	fl->rx_offset = 0;
1884 
1885 	return (m);
1886 }
1887 
1888 static struct mbuf *
1889 get_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf)
1890 {
1891 	struct mbuf *m0, *m, **pnext;
1892 	u_int remaining;
1893 	const u_int total = G_RSPD_LEN(len_newbuf);
1894 
1895 	if (__predict_false(fl->flags & FL_BUF_RESUME)) {
1896 		M_ASSERTPKTHDR(fl->m0);
1897 		MPASS(fl->m0->m_pkthdr.len == total);
1898 		MPASS(fl->remaining < total);
1899 
1900 		m0 = fl->m0;
1901 		pnext = fl->pnext;
1902 		remaining = fl->remaining;
1903 		fl->flags &= ~FL_BUF_RESUME;
1904 		goto get_segment;
1905 	}
1906 
1907 	if (fl->rx_offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
1908 		fl->rx_offset = 0;
1909 		if (__predict_false(++fl->cidx % 8 == 0)) {
1910 			uint16_t cidx = fl->cidx / 8;
1911 
1912 			if (__predict_false(cidx == fl->sidx))
1913 				fl->cidx = cidx = 0;
1914 			fl->hw_cidx = cidx;
1915 		}
1916 	}
1917 
1918 	/*
1919 	 * Payload starts at rx_offset in the current hw buffer.  Its length is
1920 	 * 'len' and it may span multiple hw buffers.
1921 	 */
1922 
1923 	m0 = get_scatter_segment(sc, fl, 0, total);
1924 	if (m0 == NULL)
1925 		return (NULL);
1926 	remaining = total - m0->m_len;
1927 	pnext = &m0->m_next;
1928 	while (remaining > 0) {
1929 get_segment:
1930 		MPASS(fl->rx_offset == 0);
1931 		m = get_scatter_segment(sc, fl, total - remaining, remaining);
1932 		if (__predict_false(m == NULL)) {
1933 			fl->m0 = m0;
1934 			fl->pnext = pnext;
1935 			fl->remaining = remaining;
1936 			fl->flags |= FL_BUF_RESUME;
1937 			return (NULL);
1938 		}
1939 		*pnext = m;
1940 		pnext = &m->m_next;
1941 		remaining -= m->m_len;
1942 	}
1943 	*pnext = NULL;
1944 
1945 	M_ASSERTPKTHDR(m0);
1946 	return (m0);
1947 }
1948 
1949 static int
1950 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
1951 {
1952 	struct sge_rxq *rxq = iq_to_rxq(iq);
1953 	struct ifnet *ifp = rxq->ifp;
1954 	struct adapter *sc = iq->adapter;
1955 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
1956 #if defined(INET) || defined(INET6)
1957 	struct lro_ctrl *lro = &rxq->lro;
1958 #endif
1959 	static const int sw_hashtype[4][2] = {
1960 		{M_HASHTYPE_NONE, M_HASHTYPE_NONE},
1961 		{M_HASHTYPE_RSS_IPV4, M_HASHTYPE_RSS_IPV6},
1962 		{M_HASHTYPE_RSS_TCP_IPV4, M_HASHTYPE_RSS_TCP_IPV6},
1963 		{M_HASHTYPE_RSS_UDP_IPV4, M_HASHTYPE_RSS_UDP_IPV6},
1964 	};
1965 
1966 	KASSERT(m0 != NULL, ("%s: no payload with opcode %02x", __func__,
1967 	    rss->opcode));
1968 
1969 	m0->m_pkthdr.len -= sc->params.sge.fl_pktshift;
1970 	m0->m_len -= sc->params.sge.fl_pktshift;
1971 	m0->m_data += sc->params.sge.fl_pktshift;
1972 
1973 	m0->m_pkthdr.rcvif = ifp;
1974 	M_HASHTYPE_SET(m0, sw_hashtype[rss->hash_type][rss->ipv6]);
1975 	m0->m_pkthdr.flowid = be32toh(rss->hash_val);
1976 
1977 	if (cpl->csum_calc && !(cpl->err_vec & sc->params.tp.err_vec_mask)) {
1978 		if (ifp->if_capenable & IFCAP_RXCSUM &&
1979 		    cpl->l2info & htobe32(F_RXF_IP)) {
1980 			m0->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
1981 			    CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
1982 			rxq->rxcsum++;
1983 		} else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
1984 		    cpl->l2info & htobe32(F_RXF_IP6)) {
1985 			m0->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
1986 			    CSUM_PSEUDO_HDR);
1987 			rxq->rxcsum++;
1988 		}
1989 
1990 		if (__predict_false(cpl->ip_frag))
1991 			m0->m_pkthdr.csum_data = be16toh(cpl->csum);
1992 		else
1993 			m0->m_pkthdr.csum_data = 0xffff;
1994 	}
1995 
1996 	if (cpl->vlan_ex) {
1997 		m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
1998 		m0->m_flags |= M_VLANTAG;
1999 		rxq->vlan_extraction++;
2000 	}
2001 
2002 #if defined(INET) || defined(INET6)
2003 	if (iq->flags & IQ_LRO_ENABLED) {
2004 		if (sort_before_lro(lro)) {
2005 			tcp_lro_queue_mbuf(lro, m0);
2006 			return (0); /* queued for sort, then LRO */
2007 		}
2008 		if (tcp_lro_rx(lro, m0, 0) == 0)
2009 			return (0); /* queued for LRO */
2010 	}
2011 #endif
2012 	ifp->if_input(ifp, m0);
2013 
2014 	return (0);
2015 }
2016 
2017 /*
2018  * Must drain the wrq or make sure that someone else will.
2019  */
2020 static void
2021 wrq_tx_drain(void *arg, int n)
2022 {
2023 	struct sge_wrq *wrq = arg;
2024 	struct sge_eq *eq = &wrq->eq;
2025 
2026 	EQ_LOCK(eq);
2027 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2028 		drain_wrq_wr_list(wrq->adapter, wrq);
2029 	EQ_UNLOCK(eq);
2030 }
2031 
2032 static void
2033 drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq)
2034 {
2035 	struct sge_eq *eq = &wrq->eq;
2036 	u_int available, dbdiff;	/* # of hardware descriptors */
2037 	u_int n;
2038 	struct wrqe *wr;
2039 	struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
2040 
2041 	EQ_LOCK_ASSERT_OWNED(eq);
2042 	MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
2043 	wr = STAILQ_FIRST(&wrq->wr_list);
2044 	MPASS(wr != NULL);	/* Must be called with something useful to do */
2045 	MPASS(eq->pidx == eq->dbidx);
2046 	dbdiff = 0;
2047 
2048 	do {
2049 		eq->cidx = read_hw_cidx(eq);
2050 		if (eq->pidx == eq->cidx)
2051 			available = eq->sidx - 1;
2052 		else
2053 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2054 
2055 		MPASS(wr->wrq == wrq);
2056 		n = howmany(wr->wr_len, EQ_ESIZE);
2057 		if (available < n)
2058 			break;
2059 
2060 		dst = (void *)&eq->desc[eq->pidx];
2061 		if (__predict_true(eq->sidx - eq->pidx > n)) {
2062 			/* Won't wrap, won't end exactly at the status page. */
2063 			bcopy(&wr->wr[0], dst, wr->wr_len);
2064 			eq->pidx += n;
2065 		} else {
2066 			int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE;
2067 
2068 			bcopy(&wr->wr[0], dst, first_portion);
2069 			if (wr->wr_len > first_portion) {
2070 				bcopy(&wr->wr[first_portion], &eq->desc[0],
2071 				    wr->wr_len - first_portion);
2072 			}
2073 			eq->pidx = n - (eq->sidx - eq->pidx);
2074 		}
2075 		wrq->tx_wrs_copied++;
2076 
2077 		if (available < eq->sidx / 4 &&
2078 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2079 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2080 			    F_FW_WR_EQUEQ);
2081 			eq->equeqidx = eq->pidx;
2082 		} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
2083 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
2084 			eq->equeqidx = eq->pidx;
2085 		}
2086 
2087 		dbdiff += n;
2088 		if (dbdiff >= 16) {
2089 			ring_eq_db(sc, eq, dbdiff);
2090 			dbdiff = 0;
2091 		}
2092 
2093 		STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
2094 		free_wrqe(wr);
2095 		MPASS(wrq->nwr_pending > 0);
2096 		wrq->nwr_pending--;
2097 		MPASS(wrq->ndesc_needed >= n);
2098 		wrq->ndesc_needed -= n;
2099 	} while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL);
2100 
2101 	if (dbdiff)
2102 		ring_eq_db(sc, eq, dbdiff);
2103 }
2104 
2105 /*
2106  * Doesn't fail.  Holds on to work requests it can't send right away.
2107  */
2108 void
2109 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
2110 {
2111 #ifdef INVARIANTS
2112 	struct sge_eq *eq = &wrq->eq;
2113 #endif
2114 
2115 	EQ_LOCK_ASSERT_OWNED(eq);
2116 	MPASS(wr != NULL);
2117 	MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN);
2118 	MPASS((wr->wr_len & 0x7) == 0);
2119 
2120 	STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
2121 	wrq->nwr_pending++;
2122 	wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE);
2123 
2124 	if (!TAILQ_EMPTY(&wrq->incomplete_wrs))
2125 		return;	/* commit_wrq_wr will drain wr_list as well. */
2126 
2127 	drain_wrq_wr_list(sc, wrq);
2128 
2129 	/* Doorbell must have caught up to the pidx. */
2130 	MPASS(eq->pidx == eq->dbidx);
2131 }
2132 
2133 void
2134 t4_update_fl_bufsize(struct ifnet *ifp)
2135 {
2136 	struct vi_info *vi = ifp->if_softc;
2137 	struct adapter *sc = vi->pi->adapter;
2138 	struct sge_rxq *rxq;
2139 #ifdef TCP_OFFLOAD
2140 	struct sge_ofld_rxq *ofld_rxq;
2141 #endif
2142 	struct sge_fl *fl;
2143 	int i, maxp, mtu = ifp->if_mtu;
2144 
2145 	maxp = mtu_to_max_payload(sc, mtu, 0);
2146 	for_each_rxq(vi, i, rxq) {
2147 		fl = &rxq->fl;
2148 
2149 		FL_LOCK(fl);
2150 		find_best_refill_source(sc, fl, maxp);
2151 		FL_UNLOCK(fl);
2152 	}
2153 #ifdef TCP_OFFLOAD
2154 	maxp = mtu_to_max_payload(sc, mtu, 1);
2155 	for_each_ofld_rxq(vi, i, ofld_rxq) {
2156 		fl = &ofld_rxq->fl;
2157 
2158 		FL_LOCK(fl);
2159 		find_best_refill_source(sc, fl, maxp);
2160 		FL_UNLOCK(fl);
2161 	}
2162 #endif
2163 }
2164 
2165 static inline int
2166 mbuf_nsegs(struct mbuf *m)
2167 {
2168 
2169 	M_ASSERTPKTHDR(m);
2170 	KASSERT(m->m_pkthdr.l5hlen > 0,
2171 	    ("%s: mbuf %p missing information on # of segments.", __func__, m));
2172 
2173 	return (m->m_pkthdr.l5hlen);
2174 }
2175 
2176 static inline void
2177 set_mbuf_nsegs(struct mbuf *m, uint8_t nsegs)
2178 {
2179 
2180 	M_ASSERTPKTHDR(m);
2181 	m->m_pkthdr.l5hlen = nsegs;
2182 }
2183 
2184 static inline int
2185 mbuf_len16(struct mbuf *m)
2186 {
2187 	int n;
2188 
2189 	M_ASSERTPKTHDR(m);
2190 	n = m->m_pkthdr.PH_loc.eight[0];
2191 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2192 
2193 	return (n);
2194 }
2195 
2196 static inline void
2197 set_mbuf_len16(struct mbuf *m, uint8_t len16)
2198 {
2199 
2200 	M_ASSERTPKTHDR(m);
2201 	m->m_pkthdr.PH_loc.eight[0] = len16;
2202 }
2203 
2204 #ifdef RATELIMIT
2205 static inline int
2206 mbuf_eo_nsegs(struct mbuf *m)
2207 {
2208 
2209 	M_ASSERTPKTHDR(m);
2210 	return (m->m_pkthdr.PH_loc.eight[1]);
2211 }
2212 
2213 static inline void
2214 set_mbuf_eo_nsegs(struct mbuf *m, uint8_t nsegs)
2215 {
2216 
2217 	M_ASSERTPKTHDR(m);
2218 	m->m_pkthdr.PH_loc.eight[1] = nsegs;
2219 }
2220 
2221 static inline int
2222 mbuf_eo_len16(struct mbuf *m)
2223 {
2224 	int n;
2225 
2226 	M_ASSERTPKTHDR(m);
2227 	n = m->m_pkthdr.PH_loc.eight[2];
2228 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2229 
2230 	return (n);
2231 }
2232 
2233 static inline void
2234 set_mbuf_eo_len16(struct mbuf *m, uint8_t len16)
2235 {
2236 
2237 	M_ASSERTPKTHDR(m);
2238 	m->m_pkthdr.PH_loc.eight[2] = len16;
2239 }
2240 
2241 static inline int
2242 mbuf_eo_tsclk_tsoff(struct mbuf *m)
2243 {
2244 
2245 	M_ASSERTPKTHDR(m);
2246 	return (m->m_pkthdr.PH_loc.eight[3]);
2247 }
2248 
2249 static inline void
2250 set_mbuf_eo_tsclk_tsoff(struct mbuf *m, uint8_t tsclk_tsoff)
2251 {
2252 
2253 	M_ASSERTPKTHDR(m);
2254 	m->m_pkthdr.PH_loc.eight[3] = tsclk_tsoff;
2255 }
2256 
2257 static inline int
2258 needs_eo(struct mbuf *m)
2259 {
2260 
2261 	return (m->m_pkthdr.snd_tag != NULL);
2262 }
2263 #endif
2264 
2265 static inline int
2266 needs_tso(struct mbuf *m)
2267 {
2268 
2269 	M_ASSERTPKTHDR(m);
2270 
2271 	return (m->m_pkthdr.csum_flags & CSUM_TSO);
2272 }
2273 
2274 static inline int
2275 needs_l3_csum(struct mbuf *m)
2276 {
2277 
2278 	M_ASSERTPKTHDR(m);
2279 
2280 	return (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO));
2281 }
2282 
2283 static inline int
2284 needs_l4_csum(struct mbuf *m)
2285 {
2286 
2287 	M_ASSERTPKTHDR(m);
2288 
2289 	return (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
2290 	    CSUM_TCP_IPV6 | CSUM_TSO));
2291 }
2292 
2293 static inline int
2294 needs_tcp_csum(struct mbuf *m)
2295 {
2296 
2297 	M_ASSERTPKTHDR(m);
2298 	return (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_TCP_IPV6 | CSUM_TSO));
2299 }
2300 
2301 #ifdef RATELIMIT
2302 static inline int
2303 needs_udp_csum(struct mbuf *m)
2304 {
2305 
2306 	M_ASSERTPKTHDR(m);
2307 	return (m->m_pkthdr.csum_flags & (CSUM_UDP | CSUM_UDP_IPV6));
2308 }
2309 #endif
2310 
2311 static inline int
2312 needs_vlan_insertion(struct mbuf *m)
2313 {
2314 
2315 	M_ASSERTPKTHDR(m);
2316 
2317 	return (m->m_flags & M_VLANTAG);
2318 }
2319 
2320 static void *
2321 m_advance(struct mbuf **pm, int *poffset, int len)
2322 {
2323 	struct mbuf *m = *pm;
2324 	int offset = *poffset;
2325 	uintptr_t p = 0;
2326 
2327 	MPASS(len > 0);
2328 
2329 	for (;;) {
2330 		if (offset + len < m->m_len) {
2331 			offset += len;
2332 			p = mtod(m, uintptr_t) + offset;
2333 			break;
2334 		}
2335 		len -= m->m_len - offset;
2336 		m = m->m_next;
2337 		offset = 0;
2338 		MPASS(m != NULL);
2339 	}
2340 	*poffset = offset;
2341 	*pm = m;
2342 	return ((void *)p);
2343 }
2344 
2345 /*
2346  * Can deal with empty mbufs in the chain that have m_len = 0, but the chain
2347  * must have at least one mbuf that's not empty.  It is possible for this
2348  * routine to return 0 if skip accounts for all the contents of the mbuf chain.
2349  */
2350 static inline int
2351 count_mbuf_nsegs(struct mbuf *m, int skip)
2352 {
2353 	vm_paddr_t lastb, next;
2354 	vm_offset_t va;
2355 	int len, nsegs;
2356 
2357 	M_ASSERTPKTHDR(m);
2358 	MPASS(m->m_pkthdr.len > 0);
2359 	MPASS(m->m_pkthdr.len >= skip);
2360 
2361 	nsegs = 0;
2362 	lastb = 0;
2363 	for (; m; m = m->m_next) {
2364 
2365 		len = m->m_len;
2366 		if (__predict_false(len == 0))
2367 			continue;
2368 		if (skip >= len) {
2369 			skip -= len;
2370 			continue;
2371 		}
2372 		va = mtod(m, vm_offset_t) + skip;
2373 		len -= skip;
2374 		skip = 0;
2375 		next = pmap_kextract(va);
2376 		nsegs += sglist_count((void *)(uintptr_t)va, len);
2377 		if (lastb + 1 == next)
2378 			nsegs--;
2379 		lastb = pmap_kextract(va + len - 1);
2380 	}
2381 
2382 	return (nsegs);
2383 }
2384 
2385 /*
2386  * Analyze the mbuf to determine its tx needs.  The mbuf passed in may change:
2387  * a) caller can assume it's been freed if this function returns with an error.
2388  * b) it may get defragged up if the gather list is too long for the hardware.
2389  */
2390 int
2391 parse_pkt(struct adapter *sc, struct mbuf **mp)
2392 {
2393 	struct mbuf *m0 = *mp, *m;
2394 	int rc, nsegs, defragged = 0, offset;
2395 	struct ether_header *eh;
2396 	void *l3hdr;
2397 #if defined(INET) || defined(INET6)
2398 	struct tcphdr *tcp;
2399 #endif
2400 	uint16_t eh_type;
2401 
2402 	M_ASSERTPKTHDR(m0);
2403 	if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) {
2404 		rc = EINVAL;
2405 fail:
2406 		m_freem(m0);
2407 		*mp = NULL;
2408 		return (rc);
2409 	}
2410 restart:
2411 	/*
2412 	 * First count the number of gather list segments in the payload.
2413 	 * Defrag the mbuf if nsegs exceeds the hardware limit.
2414 	 */
2415 	M_ASSERTPKTHDR(m0);
2416 	MPASS(m0->m_pkthdr.len > 0);
2417 	nsegs = count_mbuf_nsegs(m0, 0);
2418 	if (nsegs > (needs_tso(m0) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS)) {
2419 		if (defragged++ > 0 || (m = m_defrag(m0, M_NOWAIT)) == NULL) {
2420 			rc = EFBIG;
2421 			goto fail;
2422 		}
2423 		*mp = m0 = m;	/* update caller's copy after defrag */
2424 		goto restart;
2425 	}
2426 
2427 	if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN)) {
2428 		m0 = m_pullup(m0, m0->m_pkthdr.len);
2429 		if (m0 == NULL) {
2430 			/* Should have left well enough alone. */
2431 			rc = EFBIG;
2432 			goto fail;
2433 		}
2434 		*mp = m0;	/* update caller's copy after pullup */
2435 		goto restart;
2436 	}
2437 	set_mbuf_nsegs(m0, nsegs);
2438 	if (sc->flags & IS_VF)
2439 		set_mbuf_len16(m0, txpkt_vm_len16(nsegs, needs_tso(m0)));
2440 	else
2441 		set_mbuf_len16(m0, txpkt_len16(nsegs, needs_tso(m0)));
2442 
2443 #ifdef RATELIMIT
2444 	/*
2445 	 * Ethofld is limited to TCP and UDP for now, and only when L4 hw
2446 	 * checksumming is enabled.  needs_l4_csum happens to check for all the
2447 	 * right things.
2448 	 */
2449 	if (__predict_false(needs_eo(m0) && !needs_l4_csum(m0)))
2450 		m0->m_pkthdr.snd_tag = NULL;
2451 #endif
2452 
2453 	if (!needs_tso(m0) &&
2454 #ifdef RATELIMIT
2455 	    !needs_eo(m0) &&
2456 #endif
2457 	    !(sc->flags & IS_VF && (needs_l3_csum(m0) || needs_l4_csum(m0))))
2458 		return (0);
2459 
2460 	m = m0;
2461 	eh = mtod(m, struct ether_header *);
2462 	eh_type = ntohs(eh->ether_type);
2463 	if (eh_type == ETHERTYPE_VLAN) {
2464 		struct ether_vlan_header *evh = (void *)eh;
2465 
2466 		eh_type = ntohs(evh->evl_proto);
2467 		m0->m_pkthdr.l2hlen = sizeof(*evh);
2468 	} else
2469 		m0->m_pkthdr.l2hlen = sizeof(*eh);
2470 
2471 	offset = 0;
2472 	l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2473 
2474 	switch (eh_type) {
2475 #ifdef INET6
2476 	case ETHERTYPE_IPV6:
2477 	{
2478 		struct ip6_hdr *ip6 = l3hdr;
2479 
2480 		MPASS(!needs_tso(m0) || ip6->ip6_nxt == IPPROTO_TCP);
2481 
2482 		m0->m_pkthdr.l3hlen = sizeof(*ip6);
2483 		break;
2484 	}
2485 #endif
2486 #ifdef INET
2487 	case ETHERTYPE_IP:
2488 	{
2489 		struct ip *ip = l3hdr;
2490 
2491 		m0->m_pkthdr.l3hlen = ip->ip_hl * 4;
2492 		break;
2493 	}
2494 #endif
2495 	default:
2496 		panic("%s: ethertype 0x%04x unknown.  if_cxgbe must be compiled"
2497 		    " with the same INET/INET6 options as the kernel.",
2498 		    __func__, eh_type);
2499 	}
2500 
2501 #if defined(INET) || defined(INET6)
2502 	if (needs_tcp_csum(m0)) {
2503 		tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen);
2504 		m0->m_pkthdr.l4hlen = tcp->th_off * 4;
2505 #ifdef RATELIMIT
2506 		if (tsclk >= 0 && *(uint32_t *)(tcp + 1) == ntohl(0x0101080a)) {
2507 			set_mbuf_eo_tsclk_tsoff(m0,
2508 			    V_FW_ETH_TX_EO_WR_TSCLK(tsclk) |
2509 			    V_FW_ETH_TX_EO_WR_TSOFF(sizeof(*tcp) / 2 + 1));
2510 		} else
2511 			set_mbuf_eo_tsclk_tsoff(m0, 0);
2512 	} else if (needs_udp_csum(m)) {
2513 		m0->m_pkthdr.l4hlen = sizeof(struct udphdr);
2514 #endif
2515 	}
2516 #ifdef RATELIMIT
2517 	if (needs_eo(m0)) {
2518 		u_int immhdrs;
2519 
2520 		/* EO WRs have the headers in the WR and not the GL. */
2521 		immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen +
2522 		    m0->m_pkthdr.l4hlen;
2523 		nsegs = count_mbuf_nsegs(m0, immhdrs);
2524 		set_mbuf_eo_nsegs(m0, nsegs);
2525 		set_mbuf_eo_len16(m0,
2526 		    txpkt_eo_len16(nsegs, immhdrs, needs_tso(m0)));
2527 	}
2528 #endif
2529 #endif
2530 	MPASS(m0 == *mp);
2531 	return (0);
2532 }
2533 
2534 void *
2535 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie)
2536 {
2537 	struct sge_eq *eq = &wrq->eq;
2538 	struct adapter *sc = wrq->adapter;
2539 	int ndesc, available;
2540 	struct wrqe *wr;
2541 	void *w;
2542 
2543 	MPASS(len16 > 0);
2544 	ndesc = howmany(len16, EQ_ESIZE / 16);
2545 	MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC);
2546 
2547 	EQ_LOCK(eq);
2548 
2549 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2550 		drain_wrq_wr_list(sc, wrq);
2551 
2552 	if (!STAILQ_EMPTY(&wrq->wr_list)) {
2553 slowpath:
2554 		EQ_UNLOCK(eq);
2555 		wr = alloc_wrqe(len16 * 16, wrq);
2556 		if (__predict_false(wr == NULL))
2557 			return (NULL);
2558 		cookie->pidx = -1;
2559 		cookie->ndesc = ndesc;
2560 		return (&wr->wr);
2561 	}
2562 
2563 	eq->cidx = read_hw_cidx(eq);
2564 	if (eq->pidx == eq->cidx)
2565 		available = eq->sidx - 1;
2566 	else
2567 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2568 	if (available < ndesc)
2569 		goto slowpath;
2570 
2571 	cookie->pidx = eq->pidx;
2572 	cookie->ndesc = ndesc;
2573 	TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link);
2574 
2575 	w = &eq->desc[eq->pidx];
2576 	IDXINCR(eq->pidx, ndesc, eq->sidx);
2577 	if (__predict_false(cookie->pidx + ndesc > eq->sidx)) {
2578 		w = &wrq->ss[0];
2579 		wrq->ss_pidx = cookie->pidx;
2580 		wrq->ss_len = len16 * 16;
2581 	}
2582 
2583 	EQ_UNLOCK(eq);
2584 
2585 	return (w);
2586 }
2587 
2588 void
2589 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie)
2590 {
2591 	struct sge_eq *eq = &wrq->eq;
2592 	struct adapter *sc = wrq->adapter;
2593 	int ndesc, pidx;
2594 	struct wrq_cookie *prev, *next;
2595 
2596 	if (cookie->pidx == -1) {
2597 		struct wrqe *wr = __containerof(w, struct wrqe, wr);
2598 
2599 		t4_wrq_tx(sc, wr);
2600 		return;
2601 	}
2602 
2603 	if (__predict_false(w == &wrq->ss[0])) {
2604 		int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE;
2605 
2606 		MPASS(wrq->ss_len > n);	/* WR had better wrap around. */
2607 		bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n);
2608 		bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n);
2609 		wrq->tx_wrs_ss++;
2610 	} else
2611 		wrq->tx_wrs_direct++;
2612 
2613 	EQ_LOCK(eq);
2614 	ndesc = cookie->ndesc;	/* Can be more than SGE_MAX_WR_NDESC here. */
2615 	pidx = cookie->pidx;
2616 	MPASS(pidx >= 0 && pidx < eq->sidx);
2617 	prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link);
2618 	next = TAILQ_NEXT(cookie, link);
2619 	if (prev == NULL) {
2620 		MPASS(pidx == eq->dbidx);
2621 		if (next == NULL || ndesc >= 16) {
2622 			int available;
2623 			struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
2624 
2625 			/*
2626 			 * Note that the WR via which we'll request tx updates
2627 			 * is at pidx and not eq->pidx, which has moved on
2628 			 * already.
2629 			 */
2630 			dst = (void *)&eq->desc[pidx];
2631 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2632 			if (available < eq->sidx / 4 &&
2633 			    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2634 				dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2635 				    F_FW_WR_EQUEQ);
2636 				eq->equeqidx = pidx;
2637 			} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
2638 				dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
2639 				eq->equeqidx = pidx;
2640 			}
2641 
2642 			ring_eq_db(wrq->adapter, eq, ndesc);
2643 		} else {
2644 			MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc);
2645 			next->pidx = pidx;
2646 			next->ndesc += ndesc;
2647 		}
2648 	} else {
2649 		MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc);
2650 		prev->ndesc += ndesc;
2651 	}
2652 	TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link);
2653 
2654 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2655 		drain_wrq_wr_list(sc, wrq);
2656 
2657 #ifdef INVARIANTS
2658 	if (TAILQ_EMPTY(&wrq->incomplete_wrs)) {
2659 		/* Doorbell must have caught up to the pidx. */
2660 		MPASS(wrq->eq.pidx == wrq->eq.dbidx);
2661 	}
2662 #endif
2663 	EQ_UNLOCK(eq);
2664 }
2665 
2666 static u_int
2667 can_resume_eth_tx(struct mp_ring *r)
2668 {
2669 	struct sge_eq *eq = r->cookie;
2670 
2671 	return (total_available_tx_desc(eq) > eq->sidx / 8);
2672 }
2673 
2674 static inline int
2675 cannot_use_txpkts(struct mbuf *m)
2676 {
2677 	/* maybe put a GL limit too, to avoid silliness? */
2678 
2679 	return (needs_tso(m));
2680 }
2681 
2682 static inline int
2683 discard_tx(struct sge_eq *eq)
2684 {
2685 
2686 	return ((eq->flags & (EQ_ENABLED | EQ_QFLUSH)) != EQ_ENABLED);
2687 }
2688 
2689 /*
2690  * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to
2691  * be consumed.  Return the actual number consumed.  0 indicates a stall.
2692  */
2693 static u_int
2694 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx)
2695 {
2696 	struct sge_txq *txq = r->cookie;
2697 	struct sge_eq *eq = &txq->eq;
2698 	struct ifnet *ifp = txq->ifp;
2699 	struct vi_info *vi = ifp->if_softc;
2700 	struct port_info *pi = vi->pi;
2701 	struct adapter *sc = pi->adapter;
2702 	u_int total, remaining;		/* # of packets */
2703 	u_int available, dbdiff;	/* # of hardware descriptors */
2704 	u_int n, next_cidx;
2705 	struct mbuf *m0, *tail;
2706 	struct txpkts txp;
2707 	struct fw_eth_tx_pkts_wr *wr;	/* any fw WR struct will do */
2708 
2709 	remaining = IDXDIFF(pidx, cidx, r->size);
2710 	MPASS(remaining > 0);	/* Must not be called without work to do. */
2711 	total = 0;
2712 
2713 	TXQ_LOCK(txq);
2714 	if (__predict_false(discard_tx(eq))) {
2715 		while (cidx != pidx) {
2716 			m0 = r->items[cidx];
2717 			m_freem(m0);
2718 			if (++cidx == r->size)
2719 				cidx = 0;
2720 		}
2721 		reclaim_tx_descs(txq, 2048);
2722 		total = remaining;
2723 		goto done;
2724 	}
2725 
2726 	/* How many hardware descriptors do we have readily available. */
2727 	if (eq->pidx == eq->cidx)
2728 		available = eq->sidx - 1;
2729 	else
2730 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2731 	dbdiff = IDXDIFF(eq->pidx, eq->dbidx, eq->sidx);
2732 
2733 	while (remaining > 0) {
2734 
2735 		m0 = r->items[cidx];
2736 		M_ASSERTPKTHDR(m0);
2737 		MPASS(m0->m_nextpkt == NULL);
2738 
2739 		if (available < SGE_MAX_WR_NDESC) {
2740 			available += reclaim_tx_descs(txq, 64);
2741 			if (available < howmany(mbuf_len16(m0), EQ_ESIZE / 16))
2742 				break;	/* out of descriptors */
2743 		}
2744 
2745 		next_cidx = cidx + 1;
2746 		if (__predict_false(next_cidx == r->size))
2747 			next_cidx = 0;
2748 
2749 		wr = (void *)&eq->desc[eq->pidx];
2750 		if (sc->flags & IS_VF) {
2751 			total++;
2752 			remaining--;
2753 			ETHER_BPF_MTAP(ifp, m0);
2754 			n = write_txpkt_vm_wr(sc, txq, (void *)wr, m0,
2755 			    available);
2756 		} else if (remaining > 1 &&
2757 		    try_txpkts(m0, r->items[next_cidx], &txp, available) == 0) {
2758 
2759 			/* pkts at cidx, next_cidx should both be in txp. */
2760 			MPASS(txp.npkt == 2);
2761 			tail = r->items[next_cidx];
2762 			MPASS(tail->m_nextpkt == NULL);
2763 			ETHER_BPF_MTAP(ifp, m0);
2764 			ETHER_BPF_MTAP(ifp, tail);
2765 			m0->m_nextpkt = tail;
2766 
2767 			if (__predict_false(++next_cidx == r->size))
2768 				next_cidx = 0;
2769 
2770 			while (next_cidx != pidx) {
2771 				if (add_to_txpkts(r->items[next_cidx], &txp,
2772 				    available) != 0)
2773 					break;
2774 				tail->m_nextpkt = r->items[next_cidx];
2775 				tail = tail->m_nextpkt;
2776 				ETHER_BPF_MTAP(ifp, tail);
2777 				if (__predict_false(++next_cidx == r->size))
2778 					next_cidx = 0;
2779 			}
2780 
2781 			n = write_txpkts_wr(txq, wr, m0, &txp, available);
2782 			total += txp.npkt;
2783 			remaining -= txp.npkt;
2784 		} else {
2785 			total++;
2786 			remaining--;
2787 			ETHER_BPF_MTAP(ifp, m0);
2788 			n = write_txpkt_wr(txq, (void *)wr, m0, available);
2789 		}
2790 		MPASS(n >= 1 && n <= available && n <= SGE_MAX_WR_NDESC);
2791 
2792 		available -= n;
2793 		dbdiff += n;
2794 		IDXINCR(eq->pidx, n, eq->sidx);
2795 
2796 		if (total_available_tx_desc(eq) < eq->sidx / 4 &&
2797 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2798 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2799 			    F_FW_WR_EQUEQ);
2800 			eq->equeqidx = eq->pidx;
2801 		} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
2802 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
2803 			eq->equeqidx = eq->pidx;
2804 		}
2805 
2806 		if (dbdiff >= 16 && remaining >= 4) {
2807 			ring_eq_db(sc, eq, dbdiff);
2808 			available += reclaim_tx_descs(txq, 4 * dbdiff);
2809 			dbdiff = 0;
2810 		}
2811 
2812 		cidx = next_cidx;
2813 	}
2814 	if (dbdiff != 0) {
2815 		ring_eq_db(sc, eq, dbdiff);
2816 		reclaim_tx_descs(txq, 32);
2817 	}
2818 done:
2819 	TXQ_UNLOCK(txq);
2820 
2821 	return (total);
2822 }
2823 
2824 static inline void
2825 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
2826     int qsize)
2827 {
2828 
2829 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
2830 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
2831 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
2832 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
2833 
2834 	iq->flags = 0;
2835 	iq->adapter = sc;
2836 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
2837 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
2838 	if (pktc_idx >= 0) {
2839 		iq->intr_params |= F_QINTR_CNT_EN;
2840 		iq->intr_pktc_idx = pktc_idx;
2841 	}
2842 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
2843 	iq->sidx = iq->qsize - sc->params.sge.spg_len / IQ_ESIZE;
2844 }
2845 
2846 static inline void
2847 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name)
2848 {
2849 
2850 	fl->qsize = qsize;
2851 	fl->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2852 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
2853 	if (sc->flags & BUF_PACKING_OK &&
2854 	    ((!is_t4(sc) && buffer_packing) ||	/* T5+: enabled unless 0 */
2855 	    (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */
2856 		fl->flags |= FL_BUF_PACKING;
2857 	find_best_refill_source(sc, fl, maxp);
2858 	find_safe_refill_source(sc, fl);
2859 }
2860 
2861 static inline void
2862 init_eq(struct adapter *sc, struct sge_eq *eq, int eqtype, int qsize,
2863     uint8_t tx_chan, uint16_t iqid, char *name)
2864 {
2865 	KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
2866 
2867 	eq->flags = eqtype & EQ_TYPEMASK;
2868 	eq->tx_chan = tx_chan;
2869 	eq->iqid = iqid;
2870 	eq->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2871 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
2872 }
2873 
2874 static int
2875 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
2876     bus_dmamap_t *map, bus_addr_t *pa, void **va)
2877 {
2878 	int rc;
2879 
2880 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
2881 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
2882 	if (rc != 0) {
2883 		device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
2884 		goto done;
2885 	}
2886 
2887 	rc = bus_dmamem_alloc(*tag, va,
2888 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
2889 	if (rc != 0) {
2890 		device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
2891 		goto done;
2892 	}
2893 
2894 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
2895 	if (rc != 0) {
2896 		device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
2897 		goto done;
2898 	}
2899 done:
2900 	if (rc)
2901 		free_ring(sc, *tag, *map, *pa, *va);
2902 
2903 	return (rc);
2904 }
2905 
2906 static int
2907 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
2908     bus_addr_t pa, void *va)
2909 {
2910 	if (pa)
2911 		bus_dmamap_unload(tag, map);
2912 	if (va)
2913 		bus_dmamem_free(tag, va, map);
2914 	if (tag)
2915 		bus_dma_tag_destroy(tag);
2916 
2917 	return (0);
2918 }
2919 
2920 /*
2921  * Allocates the ring for an ingress queue and an optional freelist.  If the
2922  * freelist is specified it will be allocated and then associated with the
2923  * ingress queue.
2924  *
2925  * Returns errno on failure.  Resources allocated up to that point may still be
2926  * allocated.  Caller is responsible for cleanup in case this function fails.
2927  *
2928  * If the ingress queue will take interrupts directly then the intr_idx
2929  * specifies the vector, starting from 0.  -1 means the interrupts for this
2930  * queue should be forwarded to the fwq.
2931  */
2932 static int
2933 alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl,
2934     int intr_idx, int cong)
2935 {
2936 	int rc, i, cntxt_id;
2937 	size_t len;
2938 	struct fw_iq_cmd c;
2939 	struct port_info *pi = vi->pi;
2940 	struct adapter *sc = iq->adapter;
2941 	struct sge_params *sp = &sc->params.sge;
2942 	__be32 v = 0;
2943 
2944 	len = iq->qsize * IQ_ESIZE;
2945 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
2946 	    (void **)&iq->desc);
2947 	if (rc != 0)
2948 		return (rc);
2949 
2950 	bzero(&c, sizeof(c));
2951 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
2952 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
2953 	    V_FW_IQ_CMD_VFN(0));
2954 
2955 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
2956 	    FW_LEN16(c));
2957 
2958 	/* Special handling for firmware event queue */
2959 	if (iq == &sc->sge.fwq)
2960 		v |= F_FW_IQ_CMD_IQASYNCH;
2961 
2962 	if (intr_idx < 0) {
2963 		/* Forwarded interrupts, all headed to fwq */
2964 		v |= F_FW_IQ_CMD_IQANDST;
2965 		v |= V_FW_IQ_CMD_IQANDSTINDEX(sc->sge.fwq.cntxt_id);
2966 	} else {
2967 		KASSERT(intr_idx < sc->intr_count,
2968 		    ("%s: invalid direct intr_idx %d", __func__, intr_idx));
2969 		v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
2970 	}
2971 
2972 	c.type_to_iqandstindex = htobe32(v |
2973 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
2974 	    V_FW_IQ_CMD_VIID(vi->viid) |
2975 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
2976 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
2977 	    F_FW_IQ_CMD_IQGTSMODE |
2978 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
2979 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
2980 	c.iqsize = htobe16(iq->qsize);
2981 	c.iqaddr = htobe64(iq->ba);
2982 	if (cong >= 0)
2983 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
2984 
2985 	if (fl) {
2986 		mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
2987 
2988 		len = fl->qsize * EQ_ESIZE;
2989 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
2990 		    &fl->ba, (void **)&fl->desc);
2991 		if (rc)
2992 			return (rc);
2993 
2994 		/* Allocate space for one software descriptor per buffer. */
2995 		rc = alloc_fl_sdesc(fl);
2996 		if (rc != 0) {
2997 			device_printf(sc->dev,
2998 			    "failed to setup fl software descriptors: %d\n",
2999 			    rc);
3000 			return (rc);
3001 		}
3002 
3003 		if (fl->flags & FL_BUF_PACKING) {
3004 			fl->lowat = roundup2(sp->fl_starve_threshold2, 8);
3005 			fl->buf_boundary = sp->pack_boundary;
3006 		} else {
3007 			fl->lowat = roundup2(sp->fl_starve_threshold, 8);
3008 			fl->buf_boundary = 16;
3009 		}
3010 		if (fl_pad && fl->buf_boundary < sp->pad_boundary)
3011 			fl->buf_boundary = sp->pad_boundary;
3012 
3013 		c.iqns_to_fl0congen |=
3014 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
3015 			F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
3016 			(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
3017 			(fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
3018 			    0));
3019 		if (cong >= 0) {
3020 			c.iqns_to_fl0congen |=
3021 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
3022 				    F_FW_IQ_CMD_FL0CONGCIF |
3023 				    F_FW_IQ_CMD_FL0CONGEN);
3024 		}
3025 		c.fl0dcaen_to_fl0cidxfthresh =
3026 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
3027 			X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B) |
3028 			V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
3029 			X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
3030 		c.fl0size = htobe16(fl->qsize);
3031 		c.fl0addr = htobe64(fl->ba);
3032 	}
3033 
3034 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3035 	if (rc != 0) {
3036 		device_printf(sc->dev,
3037 		    "failed to create ingress queue: %d\n", rc);
3038 		return (rc);
3039 	}
3040 
3041 	iq->cidx = 0;
3042 	iq->gen = F_RSPD_GEN;
3043 	iq->intr_next = iq->intr_params;
3044 	iq->cntxt_id = be16toh(c.iqid);
3045 	iq->abs_id = be16toh(c.physiqid);
3046 	iq->flags |= IQ_ALLOCATED;
3047 
3048 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
3049 	if (cntxt_id >= sc->sge.niq) {
3050 		panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
3051 		    cntxt_id, sc->sge.niq - 1);
3052 	}
3053 	sc->sge.iqmap[cntxt_id] = iq;
3054 
3055 	if (fl) {
3056 		u_int qid;
3057 
3058 		iq->flags |= IQ_HAS_FL;
3059 		fl->cntxt_id = be16toh(c.fl0id);
3060 		fl->pidx = fl->cidx = 0;
3061 
3062 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
3063 		if (cntxt_id >= sc->sge.neq) {
3064 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
3065 			    __func__, cntxt_id, sc->sge.neq - 1);
3066 		}
3067 		sc->sge.eqmap[cntxt_id] = (void *)fl;
3068 
3069 		qid = fl->cntxt_id;
3070 		if (isset(&sc->doorbells, DOORBELL_UDB)) {
3071 			uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3072 			uint32_t mask = (1 << s_qpp) - 1;
3073 			volatile uint8_t *udb;
3074 
3075 			udb = sc->udbs_base + UDBS_DB_OFFSET;
3076 			udb += (qid >> s_qpp) << PAGE_SHIFT;
3077 			qid &= mask;
3078 			if (qid < PAGE_SIZE / UDBS_SEG_SIZE) {
3079 				udb += qid << UDBS_SEG_SHIFT;
3080 				qid = 0;
3081 			}
3082 			fl->udb = (volatile void *)udb;
3083 		}
3084 		fl->dbval = V_QID(qid) | sc->chip_params->sge_fl_db;
3085 
3086 		FL_LOCK(fl);
3087 		/* Enough to make sure the SGE doesn't think it's starved */
3088 		refill_fl(sc, fl, fl->lowat);
3089 		FL_UNLOCK(fl);
3090 	}
3091 
3092 	if (chip_id(sc) >= CHELSIO_T5 && !(sc->flags & IS_VF) && cong >= 0) {
3093 		uint32_t param, val;
3094 
3095 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
3096 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
3097 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
3098 		if (cong == 0)
3099 			val = 1 << 19;
3100 		else {
3101 			val = 2 << 19;
3102 			for (i = 0; i < 4; i++) {
3103 				if (cong & (1 << i))
3104 					val |= 1 << (i << 2);
3105 			}
3106 		}
3107 
3108 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
3109 		if (rc != 0) {
3110 			/* report error but carry on */
3111 			device_printf(sc->dev,
3112 			    "failed to set congestion manager context for "
3113 			    "ingress queue %d: %d\n", iq->cntxt_id, rc);
3114 		}
3115 	}
3116 
3117 	/* Enable IQ interrupts */
3118 	atomic_store_rel_int(&iq->state, IQS_IDLE);
3119 	t4_write_reg(sc, sc->sge_gts_reg, V_SEINTARM(iq->intr_params) |
3120 	    V_INGRESSQID(iq->cntxt_id));
3121 
3122 	return (0);
3123 }
3124 
3125 static int
3126 free_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl)
3127 {
3128 	int rc;
3129 	struct adapter *sc = iq->adapter;
3130 	device_t dev;
3131 
3132 	if (sc == NULL)
3133 		return (0);	/* nothing to do */
3134 
3135 	dev = vi ? vi->dev : sc->dev;
3136 
3137 	if (iq->flags & IQ_ALLOCATED) {
3138 		rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
3139 		    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
3140 		    fl ? fl->cntxt_id : 0xffff, 0xffff);
3141 		if (rc != 0) {
3142 			device_printf(dev,
3143 			    "failed to free queue %p: %d\n", iq, rc);
3144 			return (rc);
3145 		}
3146 		iq->flags &= ~IQ_ALLOCATED;
3147 	}
3148 
3149 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
3150 
3151 	bzero(iq, sizeof(*iq));
3152 
3153 	if (fl) {
3154 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
3155 		    fl->desc);
3156 
3157 		if (fl->sdesc)
3158 			free_fl_sdesc(sc, fl);
3159 
3160 		if (mtx_initialized(&fl->fl_lock))
3161 			mtx_destroy(&fl->fl_lock);
3162 
3163 		bzero(fl, sizeof(*fl));
3164 	}
3165 
3166 	return (0);
3167 }
3168 
3169 static void
3170 add_iq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
3171     struct sge_iq *iq)
3172 {
3173 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3174 
3175 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD, &iq->ba,
3176 	    "bus address of descriptor ring");
3177 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3178 	    iq->qsize * IQ_ESIZE, "descriptor ring size in bytes");
3179 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3180 	    CTLTYPE_INT | CTLFLAG_RD, &iq->abs_id, 0, sysctl_uint16, "I",
3181 	    "absolute id of the queue");
3182 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3183 	    CTLTYPE_INT | CTLFLAG_RD, &iq->cntxt_id, 0, sysctl_uint16, "I",
3184 	    "SGE context id of the queue");
3185 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3186 	    CTLTYPE_INT | CTLFLAG_RD, &iq->cidx, 0, sysctl_uint16, "I",
3187 	    "consumer index");
3188 }
3189 
3190 static void
3191 add_fl_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
3192     struct sysctl_oid *oid, struct sge_fl *fl)
3193 {
3194 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3195 
3196 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3197 	    "freelist");
3198 	children = SYSCTL_CHILDREN(oid);
3199 
3200 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3201 	    &fl->ba, "bus address of descriptor ring");
3202 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3203 	    fl->sidx * EQ_ESIZE + sc->params.sge.spg_len,
3204 	    "desc ring size in bytes");
3205 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3206 	    CTLTYPE_INT | CTLFLAG_RD, &fl->cntxt_id, 0, sysctl_uint16, "I",
3207 	    "SGE context id of the freelist");
3208 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL,
3209 	    fl_pad ? 1 : 0, "padding enabled");
3210 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL,
3211 	    fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled");
3212 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
3213 	    0, "consumer index");
3214 	if (fl->flags & FL_BUF_PACKING) {
3215 		SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
3216 		    CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
3217 	}
3218 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
3219 	    0, "producer index");
3220 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_allocated",
3221 	    CTLFLAG_RD, &fl->mbuf_allocated, "# of mbuf allocated");
3222 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_inlined",
3223 	    CTLFLAG_RD, &fl->mbuf_inlined, "# of mbuf inlined in clusters");
3224 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
3225 	    CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
3226 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
3227 	    CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
3228 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
3229 	    CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
3230 }
3231 
3232 static int
3233 alloc_fwq(struct adapter *sc)
3234 {
3235 	int rc, intr_idx;
3236 	struct sge_iq *fwq = &sc->sge.fwq;
3237 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
3238 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3239 
3240 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE);
3241 	if (sc->flags & IS_VF)
3242 		intr_idx = 0;
3243 	else
3244 		intr_idx = sc->intr_count > 1 ? 1 : 0;
3245 	rc = alloc_iq_fl(&sc->port[0]->vi[0], fwq, NULL, intr_idx, -1);
3246 	if (rc != 0) {
3247 		device_printf(sc->dev,
3248 		    "failed to create firmware event queue: %d\n", rc);
3249 		return (rc);
3250 	}
3251 
3252 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD,
3253 	    NULL, "firmware event queue");
3254 	add_iq_sysctls(&sc->ctx, oid, fwq);
3255 
3256 	return (0);
3257 }
3258 
3259 static int
3260 free_fwq(struct adapter *sc)
3261 {
3262 	return free_iq_fl(NULL, &sc->sge.fwq, NULL);
3263 }
3264 
3265 static int
3266 alloc_ctrlq(struct adapter *sc, struct sge_wrq *ctrlq, int idx,
3267     struct sysctl_oid *oid)
3268 {
3269 	int rc;
3270 	char name[16];
3271 	struct sysctl_oid_list *children;
3272 
3273 	snprintf(name, sizeof(name), "%s ctrlq%d", device_get_nameunit(sc->dev),
3274 	    idx);
3275 	init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[idx]->tx_chan,
3276 	    sc->sge.fwq.cntxt_id, name);
3277 
3278 	children = SYSCTL_CHILDREN(oid);
3279 	snprintf(name, sizeof(name), "%d", idx);
3280 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3281 	    NULL, "ctrl queue");
3282 	rc = alloc_wrq(sc, NULL, ctrlq, oid);
3283 
3284 	return (rc);
3285 }
3286 
3287 int
3288 tnl_cong(struct port_info *pi, int drop)
3289 {
3290 
3291 	if (drop == -1)
3292 		return (-1);
3293 	else if (drop == 1)
3294 		return (0);
3295 	else
3296 		return (pi->rx_e_chan_map);
3297 }
3298 
3299 static int
3300 alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int intr_idx, int idx,
3301     struct sysctl_oid *oid)
3302 {
3303 	int rc;
3304 	struct adapter *sc = vi->pi->adapter;
3305 	struct sysctl_oid_list *children;
3306 	char name[16];
3307 
3308 	rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, intr_idx,
3309 	    tnl_cong(vi->pi, cong_drop));
3310 	if (rc != 0)
3311 		return (rc);
3312 
3313 	if (idx == 0)
3314 		sc->sge.iq_base = rxq->iq.abs_id - rxq->iq.cntxt_id;
3315 	else
3316 		KASSERT(rxq->iq.cntxt_id + sc->sge.iq_base == rxq->iq.abs_id,
3317 		    ("iq_base mismatch"));
3318 	KASSERT(sc->sge.iq_base == 0 || sc->flags & IS_VF,
3319 	    ("PF with non-zero iq_base"));
3320 
3321 	/*
3322 	 * The freelist is just barely above the starvation threshold right now,
3323 	 * fill it up a bit more.
3324 	 */
3325 	FL_LOCK(&rxq->fl);
3326 	refill_fl(sc, &rxq->fl, 128);
3327 	FL_UNLOCK(&rxq->fl);
3328 
3329 #if defined(INET) || defined(INET6)
3330 	rc = tcp_lro_init_args(&rxq->lro, vi->ifp, lro_entries, lro_mbufs);
3331 	if (rc != 0)
3332 		return (rc);
3333 	MPASS(rxq->lro.ifp == vi->ifp);	/* also indicates LRO init'ed */
3334 
3335 	if (vi->ifp->if_capenable & IFCAP_LRO)
3336 		rxq->iq.flags |= IQ_LRO_ENABLED;
3337 #endif
3338 	rxq->ifp = vi->ifp;
3339 
3340 	children = SYSCTL_CHILDREN(oid);
3341 
3342 	snprintf(name, sizeof(name), "%d", idx);
3343 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3344 	    NULL, "rx queue");
3345 	children = SYSCTL_CHILDREN(oid);
3346 
3347 	add_iq_sysctls(&vi->ctx, oid, &rxq->iq);
3348 #if defined(INET) || defined(INET6)
3349 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
3350 	    &rxq->lro.lro_queued, 0, NULL);
3351 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
3352 	    &rxq->lro.lro_flushed, 0, NULL);
3353 #endif
3354 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
3355 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
3356 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_extraction",
3357 	    CTLFLAG_RD, &rxq->vlan_extraction,
3358 	    "# of times hardware extracted 802.1Q tag");
3359 
3360 	add_fl_sysctls(sc, &vi->ctx, oid, &rxq->fl);
3361 
3362 	return (rc);
3363 }
3364 
3365 static int
3366 free_rxq(struct vi_info *vi, struct sge_rxq *rxq)
3367 {
3368 	int rc;
3369 
3370 #if defined(INET) || defined(INET6)
3371 	if (rxq->lro.ifp) {
3372 		tcp_lro_free(&rxq->lro);
3373 		rxq->lro.ifp = NULL;
3374 	}
3375 #endif
3376 
3377 	rc = free_iq_fl(vi, &rxq->iq, &rxq->fl);
3378 	if (rc == 0)
3379 		bzero(rxq, sizeof(*rxq));
3380 
3381 	return (rc);
3382 }
3383 
3384 #ifdef TCP_OFFLOAD
3385 static int
3386 alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq,
3387     int intr_idx, int idx, struct sysctl_oid *oid)
3388 {
3389 	struct port_info *pi = vi->pi;
3390 	int rc;
3391 	struct sysctl_oid_list *children;
3392 	char name[16];
3393 
3394 	rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx, 0);
3395 	if (rc != 0)
3396 		return (rc);
3397 
3398 	children = SYSCTL_CHILDREN(oid);
3399 
3400 	snprintf(name, sizeof(name), "%d", idx);
3401 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3402 	    NULL, "rx queue");
3403 	add_iq_sysctls(&vi->ctx, oid, &ofld_rxq->iq);
3404 	add_fl_sysctls(pi->adapter, &vi->ctx, oid, &ofld_rxq->fl);
3405 
3406 	return (rc);
3407 }
3408 
3409 static int
3410 free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq)
3411 {
3412 	int rc;
3413 
3414 	rc = free_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl);
3415 	if (rc == 0)
3416 		bzero(ofld_rxq, sizeof(*ofld_rxq));
3417 
3418 	return (rc);
3419 }
3420 #endif
3421 
3422 #ifdef DEV_NETMAP
3423 static int
3424 alloc_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int intr_idx,
3425     int idx, struct sysctl_oid *oid)
3426 {
3427 	int rc;
3428 	struct sysctl_oid_list *children;
3429 	struct sysctl_ctx_list *ctx;
3430 	char name[16];
3431 	size_t len;
3432 	struct adapter *sc = vi->pi->adapter;
3433 	struct netmap_adapter *na = NA(vi->ifp);
3434 
3435 	MPASS(na != NULL);
3436 
3437 	len = vi->qsize_rxq * IQ_ESIZE;
3438 	rc = alloc_ring(sc, len, &nm_rxq->iq_desc_tag, &nm_rxq->iq_desc_map,
3439 	    &nm_rxq->iq_ba, (void **)&nm_rxq->iq_desc);
3440 	if (rc != 0)
3441 		return (rc);
3442 
3443 	len = na->num_rx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3444 	rc = alloc_ring(sc, len, &nm_rxq->fl_desc_tag, &nm_rxq->fl_desc_map,
3445 	    &nm_rxq->fl_ba, (void **)&nm_rxq->fl_desc);
3446 	if (rc != 0)
3447 		return (rc);
3448 
3449 	nm_rxq->vi = vi;
3450 	nm_rxq->nid = idx;
3451 	nm_rxq->iq_cidx = 0;
3452 	nm_rxq->iq_sidx = vi->qsize_rxq - sc->params.sge.spg_len / IQ_ESIZE;
3453 	nm_rxq->iq_gen = F_RSPD_GEN;
3454 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
3455 	nm_rxq->fl_sidx = na->num_rx_desc;
3456 	nm_rxq->intr_idx = intr_idx;
3457 	nm_rxq->iq_cntxt_id = INVALID_NM_RXQ_CNTXT_ID;
3458 
3459 	ctx = &vi->ctx;
3460 	children = SYSCTL_CHILDREN(oid);
3461 
3462 	snprintf(name, sizeof(name), "%d", idx);
3463 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL,
3464 	    "rx queue");
3465 	children = SYSCTL_CHILDREN(oid);
3466 
3467 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3468 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_abs_id, 0, sysctl_uint16,
3469 	    "I", "absolute id of the queue");
3470 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3471 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cntxt_id, 0, sysctl_uint16,
3472 	    "I", "SGE context id of the queue");
3473 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3474 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cidx, 0, sysctl_uint16, "I",
3475 	    "consumer index");
3476 
3477 	children = SYSCTL_CHILDREN(oid);
3478 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3479 	    "freelist");
3480 	children = SYSCTL_CHILDREN(oid);
3481 
3482 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3483 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->fl_cntxt_id, 0, sysctl_uint16,
3484 	    "I", "SGE context id of the freelist");
3485 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
3486 	    &nm_rxq->fl_cidx, 0, "consumer index");
3487 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
3488 	    &nm_rxq->fl_pidx, 0, "producer index");
3489 
3490 	return (rc);
3491 }
3492 
3493 
3494 static int
3495 free_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
3496 {
3497 	struct adapter *sc = vi->pi->adapter;
3498 
3499 	if (vi->flags & VI_INIT_DONE)
3500 		MPASS(nm_rxq->iq_cntxt_id == INVALID_NM_RXQ_CNTXT_ID);
3501 	else
3502 		MPASS(nm_rxq->iq_cntxt_id == 0);
3503 
3504 	free_ring(sc, nm_rxq->iq_desc_tag, nm_rxq->iq_desc_map, nm_rxq->iq_ba,
3505 	    nm_rxq->iq_desc);
3506 	free_ring(sc, nm_rxq->fl_desc_tag, nm_rxq->fl_desc_map, nm_rxq->fl_ba,
3507 	    nm_rxq->fl_desc);
3508 
3509 	return (0);
3510 }
3511 
3512 static int
3513 alloc_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq, int iqidx, int idx,
3514     struct sysctl_oid *oid)
3515 {
3516 	int rc;
3517 	size_t len;
3518 	struct port_info *pi = vi->pi;
3519 	struct adapter *sc = pi->adapter;
3520 	struct netmap_adapter *na = NA(vi->ifp);
3521 	char name[16];
3522 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3523 
3524 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3525 	rc = alloc_ring(sc, len, &nm_txq->desc_tag, &nm_txq->desc_map,
3526 	    &nm_txq->ba, (void **)&nm_txq->desc);
3527 	if (rc)
3528 		return (rc);
3529 
3530 	nm_txq->pidx = nm_txq->cidx = 0;
3531 	nm_txq->sidx = na->num_tx_desc;
3532 	nm_txq->nid = idx;
3533 	nm_txq->iqidx = iqidx;
3534 	nm_txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3535 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(G_FW_VIID_PFN(vi->viid)) |
3536 	    V_TXPKT_VF(G_FW_VIID_VIN(vi->viid)) |
3537 	    V_TXPKT_VF_VLD(G_FW_VIID_VIVLD(vi->viid)));
3538 	nm_txq->cntxt_id = INVALID_NM_TXQ_CNTXT_ID;
3539 
3540 	snprintf(name, sizeof(name), "%d", idx);
3541 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3542 	    NULL, "netmap tx queue");
3543 	children = SYSCTL_CHILDREN(oid);
3544 
3545 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3546 	    &nm_txq->cntxt_id, 0, "SGE context id of the queue");
3547 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3548 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->cidx, 0, sysctl_uint16, "I",
3549 	    "consumer index");
3550 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3551 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->pidx, 0, sysctl_uint16, "I",
3552 	    "producer index");
3553 
3554 	return (rc);
3555 }
3556 
3557 static int
3558 free_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
3559 {
3560 	struct adapter *sc = vi->pi->adapter;
3561 
3562 	if (vi->flags & VI_INIT_DONE)
3563 		MPASS(nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID);
3564 	else
3565 		MPASS(nm_txq->cntxt_id == 0);
3566 
3567 	free_ring(sc, nm_txq->desc_tag, nm_txq->desc_map, nm_txq->ba,
3568 	    nm_txq->desc);
3569 
3570 	return (0);
3571 }
3572 #endif
3573 
3574 static int
3575 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
3576 {
3577 	int rc, cntxt_id;
3578 	struct fw_eq_ctrl_cmd c;
3579 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3580 
3581 	bzero(&c, sizeof(c));
3582 
3583 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
3584 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
3585 	    V_FW_EQ_CTRL_CMD_VFN(0));
3586 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
3587 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
3588 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid));
3589 	c.physeqid_pkd = htobe32(0);
3590 	c.fetchszm_to_iqid =
3591 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
3592 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
3593 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
3594 	c.dcaen_to_eqsize =
3595 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3596 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3597 		V_FW_EQ_CTRL_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
3598 		V_FW_EQ_CTRL_CMD_EQSIZE(qsize));
3599 	c.eqaddr = htobe64(eq->ba);
3600 
3601 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3602 	if (rc != 0) {
3603 		device_printf(sc->dev,
3604 		    "failed to create control queue %d: %d\n", eq->tx_chan, rc);
3605 		return (rc);
3606 	}
3607 	eq->flags |= EQ_ALLOCATED;
3608 
3609 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
3610 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3611 	if (cntxt_id >= sc->sge.neq)
3612 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3613 		cntxt_id, sc->sge.neq - 1);
3614 	sc->sge.eqmap[cntxt_id] = eq;
3615 
3616 	return (rc);
3617 }
3618 
3619 static int
3620 eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3621 {
3622 	int rc, cntxt_id;
3623 	struct fw_eq_eth_cmd c;
3624 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3625 
3626 	bzero(&c, sizeof(c));
3627 
3628 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
3629 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
3630 	    V_FW_EQ_ETH_CMD_VFN(0));
3631 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
3632 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
3633 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
3634 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
3635 	c.fetchszm_to_iqid =
3636 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3637 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
3638 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
3639 	c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3640 	    V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3641 	    V_FW_EQ_ETH_CMD_EQSIZE(qsize));
3642 	c.eqaddr = htobe64(eq->ba);
3643 
3644 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3645 	if (rc != 0) {
3646 		device_printf(vi->dev,
3647 		    "failed to create Ethernet egress queue: %d\n", rc);
3648 		return (rc);
3649 	}
3650 	eq->flags |= EQ_ALLOCATED;
3651 
3652 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
3653 	eq->abs_id = G_FW_EQ_ETH_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
3654 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3655 	if (cntxt_id >= sc->sge.neq)
3656 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3657 		cntxt_id, sc->sge.neq - 1);
3658 	sc->sge.eqmap[cntxt_id] = eq;
3659 
3660 	return (rc);
3661 }
3662 
3663 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3664 static int
3665 ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3666 {
3667 	int rc, cntxt_id;
3668 	struct fw_eq_ofld_cmd c;
3669 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3670 
3671 	bzero(&c, sizeof(c));
3672 
3673 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
3674 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
3675 	    V_FW_EQ_OFLD_CMD_VFN(0));
3676 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
3677 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
3678 	c.fetchszm_to_iqid =
3679 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3680 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
3681 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
3682 	c.dcaen_to_eqsize =
3683 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3684 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3685 		V_FW_EQ_OFLD_CMD_EQSIZE(qsize));
3686 	c.eqaddr = htobe64(eq->ba);
3687 
3688 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3689 	if (rc != 0) {
3690 		device_printf(vi->dev,
3691 		    "failed to create egress queue for TCP offload: %d\n", rc);
3692 		return (rc);
3693 	}
3694 	eq->flags |= EQ_ALLOCATED;
3695 
3696 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
3697 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3698 	if (cntxt_id >= sc->sge.neq)
3699 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3700 		cntxt_id, sc->sge.neq - 1);
3701 	sc->sge.eqmap[cntxt_id] = eq;
3702 
3703 	return (rc);
3704 }
3705 #endif
3706 
3707 static int
3708 alloc_eq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3709 {
3710 	int rc, qsize;
3711 	size_t len;
3712 
3713 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
3714 
3715 	qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3716 	len = qsize * EQ_ESIZE;
3717 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
3718 	    &eq->ba, (void **)&eq->desc);
3719 	if (rc)
3720 		return (rc);
3721 
3722 	eq->pidx = eq->cidx = 0;
3723 	eq->equeqidx = eq->dbidx = 0;
3724 	eq->doorbells = sc->doorbells;
3725 
3726 	switch (eq->flags & EQ_TYPEMASK) {
3727 	case EQ_CTRL:
3728 		rc = ctrl_eq_alloc(sc, eq);
3729 		break;
3730 
3731 	case EQ_ETH:
3732 		rc = eth_eq_alloc(sc, vi, eq);
3733 		break;
3734 
3735 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3736 	case EQ_OFLD:
3737 		rc = ofld_eq_alloc(sc, vi, eq);
3738 		break;
3739 #endif
3740 
3741 	default:
3742 		panic("%s: invalid eq type %d.", __func__,
3743 		    eq->flags & EQ_TYPEMASK);
3744 	}
3745 	if (rc != 0) {
3746 		device_printf(sc->dev,
3747 		    "failed to allocate egress queue(%d): %d\n",
3748 		    eq->flags & EQ_TYPEMASK, rc);
3749 	}
3750 
3751 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
3752 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
3753 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
3754 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3755 		uint32_t mask = (1 << s_qpp) - 1;
3756 		volatile uint8_t *udb;
3757 
3758 		udb = sc->udbs_base + UDBS_DB_OFFSET;
3759 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
3760 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
3761 		if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
3762 	    		clrbit(&eq->doorbells, DOORBELL_WCWR);
3763 		else {
3764 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
3765 			eq->udb_qid = 0;
3766 		}
3767 		eq->udb = (volatile void *)udb;
3768 	}
3769 
3770 	return (rc);
3771 }
3772 
3773 static int
3774 free_eq(struct adapter *sc, struct sge_eq *eq)
3775 {
3776 	int rc;
3777 
3778 	if (eq->flags & EQ_ALLOCATED) {
3779 		switch (eq->flags & EQ_TYPEMASK) {
3780 		case EQ_CTRL:
3781 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
3782 			    eq->cntxt_id);
3783 			break;
3784 
3785 		case EQ_ETH:
3786 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
3787 			    eq->cntxt_id);
3788 			break;
3789 
3790 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3791 		case EQ_OFLD:
3792 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
3793 			    eq->cntxt_id);
3794 			break;
3795 #endif
3796 
3797 		default:
3798 			panic("%s: invalid eq type %d.", __func__,
3799 			    eq->flags & EQ_TYPEMASK);
3800 		}
3801 		if (rc != 0) {
3802 			device_printf(sc->dev,
3803 			    "failed to free egress queue (%d): %d\n",
3804 			    eq->flags & EQ_TYPEMASK, rc);
3805 			return (rc);
3806 		}
3807 		eq->flags &= ~EQ_ALLOCATED;
3808 	}
3809 
3810 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
3811 
3812 	if (mtx_initialized(&eq->eq_lock))
3813 		mtx_destroy(&eq->eq_lock);
3814 
3815 	bzero(eq, sizeof(*eq));
3816 	return (0);
3817 }
3818 
3819 static int
3820 alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq,
3821     struct sysctl_oid *oid)
3822 {
3823 	int rc;
3824 	struct sysctl_ctx_list *ctx = vi ? &vi->ctx : &sc->ctx;
3825 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3826 
3827 	rc = alloc_eq(sc, vi, &wrq->eq);
3828 	if (rc)
3829 		return (rc);
3830 
3831 	wrq->adapter = sc;
3832 	TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq);
3833 	TAILQ_INIT(&wrq->incomplete_wrs);
3834 	STAILQ_INIT(&wrq->wr_list);
3835 	wrq->nwr_pending = 0;
3836 	wrq->ndesc_needed = 0;
3837 
3838 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3839 	    &wrq->eq.ba, "bus address of descriptor ring");
3840 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3841 	    wrq->eq.sidx * EQ_ESIZE + sc->params.sge.spg_len,
3842 	    "desc ring size in bytes");
3843 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3844 	    &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
3845 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3846 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I",
3847 	    "consumer index");
3848 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx",
3849 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I",
3850 	    "producer index");
3851 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
3852 	    wrq->eq.sidx, "status page index");
3853 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD,
3854 	    &wrq->tx_wrs_direct, "# of work requests (direct)");
3855 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD,
3856 	    &wrq->tx_wrs_copied, "# of work requests (copied)");
3857 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_sspace", CTLFLAG_RD,
3858 	    &wrq->tx_wrs_ss, "# of work requests (copied from scratch space)");
3859 
3860 	return (rc);
3861 }
3862 
3863 static int
3864 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
3865 {
3866 	int rc;
3867 
3868 	rc = free_eq(sc, &wrq->eq);
3869 	if (rc)
3870 		return (rc);
3871 
3872 	bzero(wrq, sizeof(*wrq));
3873 	return (0);
3874 }
3875 
3876 static int
3877 alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx,
3878     struct sysctl_oid *oid)
3879 {
3880 	int rc;
3881 	struct port_info *pi = vi->pi;
3882 	struct adapter *sc = pi->adapter;
3883 	struct sge_eq *eq = &txq->eq;
3884 	char name[16];
3885 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3886 
3887 	rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx, can_resume_eth_tx,
3888 	    M_CXGBE, M_WAITOK);
3889 	if (rc != 0) {
3890 		device_printf(sc->dev, "failed to allocate mp_ring: %d\n", rc);
3891 		return (rc);
3892 	}
3893 
3894 	rc = alloc_eq(sc, vi, eq);
3895 	if (rc != 0) {
3896 		mp_ring_free(txq->r);
3897 		txq->r = NULL;
3898 		return (rc);
3899 	}
3900 
3901 	/* Can't fail after this point. */
3902 
3903 	if (idx == 0)
3904 		sc->sge.eq_base = eq->abs_id - eq->cntxt_id;
3905 	else
3906 		KASSERT(eq->cntxt_id + sc->sge.eq_base == eq->abs_id,
3907 		    ("eq_base mismatch"));
3908 	KASSERT(sc->sge.eq_base == 0 || sc->flags & IS_VF,
3909 	    ("PF with non-zero eq_base"));
3910 
3911 	TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq);
3912 	txq->ifp = vi->ifp;
3913 	txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK);
3914 	if (sc->flags & IS_VF)
3915 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
3916 		    V_TXPKT_INTF(pi->tx_chan));
3917 	else
3918 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3919 		    V_TXPKT_INTF(pi->tx_chan) |
3920 		    V_TXPKT_PF(G_FW_VIID_PFN(vi->viid)) |
3921 		    V_TXPKT_VF(G_FW_VIID_VIN(vi->viid)) |
3922 		    V_TXPKT_VF_VLD(G_FW_VIID_VIVLD(vi->viid)));
3923 	txq->tc_idx = -1;
3924 	txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE,
3925 	    M_ZERO | M_WAITOK);
3926 
3927 	snprintf(name, sizeof(name), "%d", idx);
3928 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3929 	    NULL, "tx queue");
3930 	children = SYSCTL_CHILDREN(oid);
3931 
3932 	SYSCTL_ADD_UAUTO(&vi->ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3933 	    &eq->ba, "bus address of descriptor ring");
3934 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3935 	    eq->sidx * EQ_ESIZE + sc->params.sge.spg_len,
3936 	    "desc ring size in bytes");
3937 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
3938 	    &eq->abs_id, 0, "absolute id of the queue");
3939 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3940 	    &eq->cntxt_id, 0, "SGE context id of the queue");
3941 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3942 	    CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I",
3943 	    "consumer index");
3944 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3945 	    CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I",
3946 	    "producer index");
3947 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
3948 	    eq->sidx, "status page index");
3949 
3950 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "tc",
3951 	    CTLTYPE_INT | CTLFLAG_RW, vi, idx, sysctl_tc, "I",
3952 	    "traffic class (-1 means none)");
3953 
3954 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
3955 	    &txq->txcsum, "# of times hardware assisted with checksum");
3956 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_insertion",
3957 	    CTLFLAG_RD, &txq->vlan_insertion,
3958 	    "# of times hardware inserted 802.1Q tag");
3959 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
3960 	    &txq->tso_wrs, "# of TSO work requests");
3961 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
3962 	    &txq->imm_wrs, "# of work requests with immediate data");
3963 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
3964 	    &txq->sgl_wrs, "# of work requests with direct SGL");
3965 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
3966 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
3967 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_wrs",
3968 	    CTLFLAG_RD, &txq->txpkts0_wrs,
3969 	    "# of txpkts (type 0) work requests");
3970 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_wrs",
3971 	    CTLFLAG_RD, &txq->txpkts1_wrs,
3972 	    "# of txpkts (type 1) work requests");
3973 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_pkts",
3974 	    CTLFLAG_RD, &txq->txpkts0_pkts,
3975 	    "# of frames tx'd using type0 txpkts work requests");
3976 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_pkts",
3977 	    CTLFLAG_RD, &txq->txpkts1_pkts,
3978 	    "# of frames tx'd using type1 txpkts work requests");
3979 
3980 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_enqueues",
3981 	    CTLFLAG_RD, &txq->r->enqueues,
3982 	    "# of enqueues to the mp_ring for this queue");
3983 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_drops",
3984 	    CTLFLAG_RD, &txq->r->drops,
3985 	    "# of drops in the mp_ring for this queue");
3986 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_starts",
3987 	    CTLFLAG_RD, &txq->r->starts,
3988 	    "# of normal consumer starts in the mp_ring for this queue");
3989 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_stalls",
3990 	    CTLFLAG_RD, &txq->r->stalls,
3991 	    "# of consumer stalls in the mp_ring for this queue");
3992 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_restarts",
3993 	    CTLFLAG_RD, &txq->r->restarts,
3994 	    "# of consumer restarts in the mp_ring for this queue");
3995 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_abdications",
3996 	    CTLFLAG_RD, &txq->r->abdications,
3997 	    "# of consumer abdications in the mp_ring for this queue");
3998 
3999 	return (0);
4000 }
4001 
4002 static int
4003 free_txq(struct vi_info *vi, struct sge_txq *txq)
4004 {
4005 	int rc;
4006 	struct adapter *sc = vi->pi->adapter;
4007 	struct sge_eq *eq = &txq->eq;
4008 
4009 	rc = free_eq(sc, eq);
4010 	if (rc)
4011 		return (rc);
4012 
4013 	sglist_free(txq->gl);
4014 	free(txq->sdesc, M_CXGBE);
4015 	mp_ring_free(txq->r);
4016 
4017 	bzero(txq, sizeof(*txq));
4018 	return (0);
4019 }
4020 
4021 static void
4022 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
4023 {
4024 	bus_addr_t *ba = arg;
4025 
4026 	KASSERT(nseg == 1,
4027 	    ("%s meant for single segment mappings only.", __func__));
4028 
4029 	*ba = error ? 0 : segs->ds_addr;
4030 }
4031 
4032 static inline void
4033 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
4034 {
4035 	uint32_t n, v;
4036 
4037 	n = IDXDIFF(fl->pidx / 8, fl->dbidx, fl->sidx);
4038 	MPASS(n > 0);
4039 
4040 	wmb();
4041 	v = fl->dbval | V_PIDX(n);
4042 	if (fl->udb)
4043 		*fl->udb = htole32(v);
4044 	else
4045 		t4_write_reg(sc, sc->sge_kdoorbell_reg, v);
4046 	IDXINCR(fl->dbidx, n, fl->sidx);
4047 }
4048 
4049 /*
4050  * Fills up the freelist by allocating up to 'n' buffers.  Buffers that are
4051  * recycled do not count towards this allocation budget.
4052  *
4053  * Returns non-zero to indicate that this freelist should be added to the list
4054  * of starving freelists.
4055  */
4056 static int
4057 refill_fl(struct adapter *sc, struct sge_fl *fl, int n)
4058 {
4059 	__be64 *d;
4060 	struct fl_sdesc *sd;
4061 	uintptr_t pa;
4062 	caddr_t cl;
4063 	struct cluster_layout *cll;
4064 	struct sw_zone_info *swz;
4065 	struct cluster_metadata *clm;
4066 	uint16_t max_pidx;
4067 	uint16_t hw_cidx = fl->hw_cidx;		/* stable snapshot */
4068 
4069 	FL_LOCK_ASSERT_OWNED(fl);
4070 
4071 	/*
4072 	 * We always stop at the beginning of the hardware descriptor that's just
4073 	 * before the one with the hw cidx.  This is to avoid hw pidx = hw cidx,
4074 	 * which would mean an empty freelist to the chip.
4075 	 */
4076 	max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1;
4077 	if (fl->pidx == max_pidx * 8)
4078 		return (0);
4079 
4080 	d = &fl->desc[fl->pidx];
4081 	sd = &fl->sdesc[fl->pidx];
4082 	cll = &fl->cll_def;	/* default layout */
4083 	swz = &sc->sge.sw_zone_info[cll->zidx];
4084 
4085 	while (n > 0) {
4086 
4087 		if (sd->cl != NULL) {
4088 
4089 			if (sd->nmbuf == 0) {
4090 				/*
4091 				 * Fast recycle without involving any atomics on
4092 				 * the cluster's metadata (if the cluster has
4093 				 * metadata).  This happens when all frames
4094 				 * received in the cluster were small enough to
4095 				 * fit within a single mbuf each.
4096 				 */
4097 				fl->cl_fast_recycled++;
4098 #ifdef INVARIANTS
4099 				clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
4100 				if (clm != NULL)
4101 					MPASS(clm->refcount == 1);
4102 #endif
4103 				goto recycled_fast;
4104 			}
4105 
4106 			/*
4107 			 * Cluster is guaranteed to have metadata.  Clusters
4108 			 * without metadata always take the fast recycle path
4109 			 * when they're recycled.
4110 			 */
4111 			clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
4112 			MPASS(clm != NULL);
4113 
4114 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4115 				fl->cl_recycled++;
4116 				counter_u64_add(extfree_rels, 1);
4117 				goto recycled;
4118 			}
4119 			sd->cl = NULL;	/* gave up my reference */
4120 		}
4121 		MPASS(sd->cl == NULL);
4122 alloc:
4123 		cl = uma_zalloc(swz->zone, M_NOWAIT);
4124 		if (__predict_false(cl == NULL)) {
4125 			if (cll == &fl->cll_alt || fl->cll_alt.zidx == -1 ||
4126 			    fl->cll_def.zidx == fl->cll_alt.zidx)
4127 				break;
4128 
4129 			/* fall back to the safe zone */
4130 			cll = &fl->cll_alt;
4131 			swz = &sc->sge.sw_zone_info[cll->zidx];
4132 			goto alloc;
4133 		}
4134 		fl->cl_allocated++;
4135 		n--;
4136 
4137 		pa = pmap_kextract((vm_offset_t)cl);
4138 		pa += cll->region1;
4139 		sd->cl = cl;
4140 		sd->cll = *cll;
4141 		*d = htobe64(pa | cll->hwidx);
4142 		clm = cl_metadata(sc, fl, cll, cl);
4143 		if (clm != NULL) {
4144 recycled:
4145 #ifdef INVARIANTS
4146 			clm->sd = sd;
4147 #endif
4148 			clm->refcount = 1;
4149 		}
4150 		sd->nmbuf = 0;
4151 recycled_fast:
4152 		d++;
4153 		sd++;
4154 		if (__predict_false(++fl->pidx % 8 == 0)) {
4155 			uint16_t pidx = fl->pidx / 8;
4156 
4157 			if (__predict_false(pidx == fl->sidx)) {
4158 				fl->pidx = 0;
4159 				pidx = 0;
4160 				sd = fl->sdesc;
4161 				d = fl->desc;
4162 			}
4163 			if (pidx == max_pidx)
4164 				break;
4165 
4166 			if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4)
4167 				ring_fl_db(sc, fl);
4168 		}
4169 	}
4170 
4171 	if (fl->pidx / 8 != fl->dbidx)
4172 		ring_fl_db(sc, fl);
4173 
4174 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
4175 }
4176 
4177 /*
4178  * Attempt to refill all starving freelists.
4179  */
4180 static void
4181 refill_sfl(void *arg)
4182 {
4183 	struct adapter *sc = arg;
4184 	struct sge_fl *fl, *fl_temp;
4185 
4186 	mtx_assert(&sc->sfl_lock, MA_OWNED);
4187 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
4188 		FL_LOCK(fl);
4189 		refill_fl(sc, fl, 64);
4190 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
4191 			TAILQ_REMOVE(&sc->sfl, fl, link);
4192 			fl->flags &= ~FL_STARVING;
4193 		}
4194 		FL_UNLOCK(fl);
4195 	}
4196 
4197 	if (!TAILQ_EMPTY(&sc->sfl))
4198 		callout_schedule(&sc->sfl_callout, hz / 5);
4199 }
4200 
4201 static int
4202 alloc_fl_sdesc(struct sge_fl *fl)
4203 {
4204 
4205 	fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc), M_CXGBE,
4206 	    M_ZERO | M_WAITOK);
4207 
4208 	return (0);
4209 }
4210 
4211 static void
4212 free_fl_sdesc(struct adapter *sc, struct sge_fl *fl)
4213 {
4214 	struct fl_sdesc *sd;
4215 	struct cluster_metadata *clm;
4216 	struct cluster_layout *cll;
4217 	int i;
4218 
4219 	sd = fl->sdesc;
4220 	for (i = 0; i < fl->sidx * 8; i++, sd++) {
4221 		if (sd->cl == NULL)
4222 			continue;
4223 
4224 		cll = &sd->cll;
4225 		clm = cl_metadata(sc, fl, cll, sd->cl);
4226 		if (sd->nmbuf == 0)
4227 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4228 		else if (clm && atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4229 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4230 			counter_u64_add(extfree_rels, 1);
4231 		}
4232 		sd->cl = NULL;
4233 	}
4234 
4235 	free(fl->sdesc, M_CXGBE);
4236 	fl->sdesc = NULL;
4237 }
4238 
4239 static inline void
4240 get_pkt_gl(struct mbuf *m, struct sglist *gl)
4241 {
4242 	int rc;
4243 
4244 	M_ASSERTPKTHDR(m);
4245 
4246 	sglist_reset(gl);
4247 	rc = sglist_append_mbuf(gl, m);
4248 	if (__predict_false(rc != 0)) {
4249 		panic("%s: mbuf %p (%d segs) was vetted earlier but now fails "
4250 		    "with %d.", __func__, m, mbuf_nsegs(m), rc);
4251 	}
4252 
4253 	KASSERT(gl->sg_nseg == mbuf_nsegs(m),
4254 	    ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m,
4255 	    mbuf_nsegs(m), gl->sg_nseg));
4256 	KASSERT(gl->sg_nseg > 0 &&
4257 	    gl->sg_nseg <= (needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS),
4258 	    ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__,
4259 		gl->sg_nseg, needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS));
4260 }
4261 
4262 /*
4263  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
4264  */
4265 static inline u_int
4266 txpkt_len16(u_int nsegs, u_int tso)
4267 {
4268 	u_int n;
4269 
4270 	MPASS(nsegs > 0);
4271 
4272 	nsegs--; /* first segment is part of ulptx_sgl */
4273 	n = sizeof(struct fw_eth_tx_pkt_wr) + sizeof(struct cpl_tx_pkt_core) +
4274 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4275 	if (tso)
4276 		n += sizeof(struct cpl_tx_pkt_lso_core);
4277 
4278 	return (howmany(n, 16));
4279 }
4280 
4281 /*
4282  * len16 for a txpkt_vm WR with a GL.  Includes the firmware work
4283  * request header.
4284  */
4285 static inline u_int
4286 txpkt_vm_len16(u_int nsegs, u_int tso)
4287 {
4288 	u_int n;
4289 
4290 	MPASS(nsegs > 0);
4291 
4292 	nsegs--; /* first segment is part of ulptx_sgl */
4293 	n = sizeof(struct fw_eth_tx_pkt_vm_wr) +
4294 	    sizeof(struct cpl_tx_pkt_core) +
4295 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4296 	if (tso)
4297 		n += sizeof(struct cpl_tx_pkt_lso_core);
4298 
4299 	return (howmany(n, 16));
4300 }
4301 
4302 /*
4303  * len16 for a txpkts type 0 WR with a GL.  Does not include the firmware work
4304  * request header.
4305  */
4306 static inline u_int
4307 txpkts0_len16(u_int nsegs)
4308 {
4309 	u_int n;
4310 
4311 	MPASS(nsegs > 0);
4312 
4313 	nsegs--; /* first segment is part of ulptx_sgl */
4314 	n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) +
4315 	    sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) +
4316 	    8 * ((3 * nsegs) / 2 + (nsegs & 1));
4317 
4318 	return (howmany(n, 16));
4319 }
4320 
4321 /*
4322  * len16 for a txpkts type 1 WR with a GL.  Does not include the firmware work
4323  * request header.
4324  */
4325 static inline u_int
4326 txpkts1_len16(void)
4327 {
4328 	u_int n;
4329 
4330 	n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl);
4331 
4332 	return (howmany(n, 16));
4333 }
4334 
4335 static inline u_int
4336 imm_payload(u_int ndesc)
4337 {
4338 	u_int n;
4339 
4340 	n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) -
4341 	    sizeof(struct cpl_tx_pkt_core);
4342 
4343 	return (n);
4344 }
4345 
4346 /*
4347  * Write a VM txpkt WR for this packet to the hardware descriptors, update the
4348  * software descriptor, and advance the pidx.  It is guaranteed that enough
4349  * descriptors are available.
4350  *
4351  * The return value is the # of hardware descriptors used.
4352  */
4353 static u_int
4354 write_txpkt_vm_wr(struct adapter *sc, struct sge_txq *txq,
4355     struct fw_eth_tx_pkt_vm_wr *wr, struct mbuf *m0, u_int available)
4356 {
4357 	struct sge_eq *eq = &txq->eq;
4358 	struct tx_sdesc *txsd;
4359 	struct cpl_tx_pkt_core *cpl;
4360 	uint32_t ctrl;	/* used in many unrelated places */
4361 	uint64_t ctrl1;
4362 	int csum_type, len16, ndesc, pktlen, nsegs;
4363 	caddr_t dst;
4364 
4365 	TXQ_LOCK_ASSERT_OWNED(txq);
4366 	M_ASSERTPKTHDR(m0);
4367 	MPASS(available > 0 && available < eq->sidx);
4368 
4369 	len16 = mbuf_len16(m0);
4370 	nsegs = mbuf_nsegs(m0);
4371 	pktlen = m0->m_pkthdr.len;
4372 	ctrl = sizeof(struct cpl_tx_pkt_core);
4373 	if (needs_tso(m0))
4374 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4375 	ndesc = howmany(len16, EQ_ESIZE / 16);
4376 	MPASS(ndesc <= available);
4377 
4378 	/* Firmware work request header */
4379 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4380 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_VM_WR) |
4381 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4382 
4383 	ctrl = V_FW_WR_LEN16(len16);
4384 	wr->equiq_to_len16 = htobe32(ctrl);
4385 	wr->r3[0] = 0;
4386 	wr->r3[1] = 0;
4387 
4388 	/*
4389 	 * Copy over ethmacdst, ethmacsrc, ethtype, and vlantci.
4390 	 * vlantci is ignored unless the ethtype is 0x8100, so it's
4391 	 * simpler to always copy it rather than making it
4392 	 * conditional.  Also, it seems that we do not have to set
4393 	 * vlantci or fake the ethtype when doing VLAN tag insertion.
4394 	 */
4395 	m_copydata(m0, 0, sizeof(struct ether_header) + 2, wr->ethmacdst);
4396 
4397 	csum_type = -1;
4398 	if (needs_tso(m0)) {
4399 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4400 
4401 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4402 		    m0->m_pkthdr.l4hlen > 0,
4403 		    ("%s: mbuf %p needs TSO but missing header lengths",
4404 			__func__, m0));
4405 
4406 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4407 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4408 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4409 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4410 			ctrl |= V_LSO_ETHHDR_LEN(1);
4411 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4412 			ctrl |= F_LSO_IPV6;
4413 
4414 		lso->lso_ctrl = htobe32(ctrl);
4415 		lso->ipid_ofst = htobe16(0);
4416 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4417 		lso->seqno_offset = htobe32(0);
4418 		lso->len = htobe32(pktlen);
4419 
4420 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4421 			csum_type = TX_CSUM_TCPIP6;
4422 		else
4423 			csum_type = TX_CSUM_TCPIP;
4424 
4425 		cpl = (void *)(lso + 1);
4426 
4427 		txq->tso_wrs++;
4428 	} else {
4429 		if (m0->m_pkthdr.csum_flags & CSUM_IP_TCP)
4430 			csum_type = TX_CSUM_TCPIP;
4431 		else if (m0->m_pkthdr.csum_flags & CSUM_IP_UDP)
4432 			csum_type = TX_CSUM_UDPIP;
4433 		else if (m0->m_pkthdr.csum_flags & CSUM_IP6_TCP)
4434 			csum_type = TX_CSUM_TCPIP6;
4435 		else if (m0->m_pkthdr.csum_flags & CSUM_IP6_UDP)
4436 			csum_type = TX_CSUM_UDPIP6;
4437 #if defined(INET)
4438 		else if (m0->m_pkthdr.csum_flags & CSUM_IP) {
4439 			/*
4440 			 * XXX: The firmware appears to stomp on the
4441 			 * fragment/flags field of the IP header when
4442 			 * using TX_CSUM_IP.  Fall back to doing
4443 			 * software checksums.
4444 			 */
4445 			u_short *sump;
4446 			struct mbuf *m;
4447 			int offset;
4448 
4449 			m = m0;
4450 			offset = 0;
4451 			sump = m_advance(&m, &offset, m0->m_pkthdr.l2hlen +
4452 			    offsetof(struct ip, ip_sum));
4453 			*sump = in_cksum_skip(m0, m0->m_pkthdr.l2hlen +
4454 			    m0->m_pkthdr.l3hlen, m0->m_pkthdr.l2hlen);
4455 			m0->m_pkthdr.csum_flags &= ~CSUM_IP;
4456 		}
4457 #endif
4458 
4459 		cpl = (void *)(wr + 1);
4460 	}
4461 
4462 	/* Checksum offload */
4463 	ctrl1 = 0;
4464 	if (needs_l3_csum(m0) == 0)
4465 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4466 	if (csum_type >= 0) {
4467 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0,
4468 	    ("%s: mbuf %p needs checksum offload but missing header lengths",
4469 			__func__, m0));
4470 
4471 		if (chip_id(sc) <= CHELSIO_T5) {
4472 			ctrl1 |= V_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4473 			    ETHER_HDR_LEN);
4474 		} else {
4475 			ctrl1 |= V_T6_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4476 			    ETHER_HDR_LEN);
4477 		}
4478 		ctrl1 |= V_TXPKT_IPHDR_LEN(m0->m_pkthdr.l3hlen);
4479 		ctrl1 |= V_TXPKT_CSUM_TYPE(csum_type);
4480 	} else
4481 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4482 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4483 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4484 		txq->txcsum++;	/* some hardware assistance provided */
4485 
4486 	/* VLAN tag insertion */
4487 	if (needs_vlan_insertion(m0)) {
4488 		ctrl1 |= F_TXPKT_VLAN_VLD |
4489 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4490 		txq->vlan_insertion++;
4491 	}
4492 
4493 	/* CPL header */
4494 	cpl->ctrl0 = txq->cpl_ctrl0;
4495 	cpl->pack = 0;
4496 	cpl->len = htobe16(pktlen);
4497 	cpl->ctrl1 = htobe64(ctrl1);
4498 
4499 	/* SGL */
4500 	dst = (void *)(cpl + 1);
4501 
4502 	/*
4503 	 * A packet using TSO will use up an entire descriptor for the
4504 	 * firmware work request header, LSO CPL, and TX_PKT_XT CPL.
4505 	 * If this descriptor is the last descriptor in the ring, wrap
4506 	 * around to the front of the ring explicitly for the start of
4507 	 * the sgl.
4508 	 */
4509 	if (dst == (void *)&eq->desc[eq->sidx]) {
4510 		dst = (void *)&eq->desc[0];
4511 		write_gl_to_txd(txq, m0, &dst, 0);
4512 	} else
4513 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4514 	txq->sgl_wrs++;
4515 
4516 	txq->txpkt_wrs++;
4517 
4518 	txsd = &txq->sdesc[eq->pidx];
4519 	txsd->m = m0;
4520 	txsd->desc_used = ndesc;
4521 
4522 	return (ndesc);
4523 }
4524 
4525 /*
4526  * Write a txpkt WR for this packet to the hardware descriptors, update the
4527  * software descriptor, and advance the pidx.  It is guaranteed that enough
4528  * descriptors are available.
4529  *
4530  * The return value is the # of hardware descriptors used.
4531  */
4532 static u_int
4533 write_txpkt_wr(struct sge_txq *txq, struct fw_eth_tx_pkt_wr *wr,
4534     struct mbuf *m0, u_int available)
4535 {
4536 	struct sge_eq *eq = &txq->eq;
4537 	struct tx_sdesc *txsd;
4538 	struct cpl_tx_pkt_core *cpl;
4539 	uint32_t ctrl;	/* used in many unrelated places */
4540 	uint64_t ctrl1;
4541 	int len16, ndesc, pktlen, nsegs;
4542 	caddr_t dst;
4543 
4544 	TXQ_LOCK_ASSERT_OWNED(txq);
4545 	M_ASSERTPKTHDR(m0);
4546 	MPASS(available > 0 && available < eq->sidx);
4547 
4548 	len16 = mbuf_len16(m0);
4549 	nsegs = mbuf_nsegs(m0);
4550 	pktlen = m0->m_pkthdr.len;
4551 	ctrl = sizeof(struct cpl_tx_pkt_core);
4552 	if (needs_tso(m0))
4553 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4554 	else if (pktlen <= imm_payload(2) && available >= 2) {
4555 		/* Immediate data.  Recalculate len16 and set nsegs to 0. */
4556 		ctrl += pktlen;
4557 		len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) +
4558 		    sizeof(struct cpl_tx_pkt_core) + pktlen, 16);
4559 		nsegs = 0;
4560 	}
4561 	ndesc = howmany(len16, EQ_ESIZE / 16);
4562 	MPASS(ndesc <= available);
4563 
4564 	/* Firmware work request header */
4565 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4566 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
4567 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4568 
4569 	ctrl = V_FW_WR_LEN16(len16);
4570 	wr->equiq_to_len16 = htobe32(ctrl);
4571 	wr->r3 = 0;
4572 
4573 	if (needs_tso(m0)) {
4574 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4575 
4576 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4577 		    m0->m_pkthdr.l4hlen > 0,
4578 		    ("%s: mbuf %p needs TSO but missing header lengths",
4579 			__func__, m0));
4580 
4581 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4582 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4583 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4584 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4585 			ctrl |= V_LSO_ETHHDR_LEN(1);
4586 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4587 			ctrl |= F_LSO_IPV6;
4588 
4589 		lso->lso_ctrl = htobe32(ctrl);
4590 		lso->ipid_ofst = htobe16(0);
4591 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4592 		lso->seqno_offset = htobe32(0);
4593 		lso->len = htobe32(pktlen);
4594 
4595 		cpl = (void *)(lso + 1);
4596 
4597 		txq->tso_wrs++;
4598 	} else
4599 		cpl = (void *)(wr + 1);
4600 
4601 	/* Checksum offload */
4602 	ctrl1 = 0;
4603 	if (needs_l3_csum(m0) == 0)
4604 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4605 	if (needs_l4_csum(m0) == 0)
4606 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4607 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4608 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4609 		txq->txcsum++;	/* some hardware assistance provided */
4610 
4611 	/* VLAN tag insertion */
4612 	if (needs_vlan_insertion(m0)) {
4613 		ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4614 		txq->vlan_insertion++;
4615 	}
4616 
4617 	/* CPL header */
4618 	cpl->ctrl0 = txq->cpl_ctrl0;
4619 	cpl->pack = 0;
4620 	cpl->len = htobe16(pktlen);
4621 	cpl->ctrl1 = htobe64(ctrl1);
4622 
4623 	/* SGL */
4624 	dst = (void *)(cpl + 1);
4625 	if (nsegs > 0) {
4626 
4627 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4628 		txq->sgl_wrs++;
4629 	} else {
4630 		struct mbuf *m;
4631 
4632 		for (m = m0; m != NULL; m = m->m_next) {
4633 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4634 #ifdef INVARIANTS
4635 			pktlen -= m->m_len;
4636 #endif
4637 		}
4638 #ifdef INVARIANTS
4639 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
4640 #endif
4641 		txq->imm_wrs++;
4642 	}
4643 
4644 	txq->txpkt_wrs++;
4645 
4646 	txsd = &txq->sdesc[eq->pidx];
4647 	txsd->m = m0;
4648 	txsd->desc_used = ndesc;
4649 
4650 	return (ndesc);
4651 }
4652 
4653 static int
4654 try_txpkts(struct mbuf *m, struct mbuf *n, struct txpkts *txp, u_int available)
4655 {
4656 	u_int needed, nsegs1, nsegs2, l1, l2;
4657 
4658 	if (cannot_use_txpkts(m) || cannot_use_txpkts(n))
4659 		return (1);
4660 
4661 	nsegs1 = mbuf_nsegs(m);
4662 	nsegs2 = mbuf_nsegs(n);
4663 	if (nsegs1 + nsegs2 == 2) {
4664 		txp->wr_type = 1;
4665 		l1 = l2 = txpkts1_len16();
4666 	} else {
4667 		txp->wr_type = 0;
4668 		l1 = txpkts0_len16(nsegs1);
4669 		l2 = txpkts0_len16(nsegs2);
4670 	}
4671 	txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) + l1 + l2;
4672 	needed = howmany(txp->len16, EQ_ESIZE / 16);
4673 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4674 		return (1);
4675 
4676 	txp->plen = m->m_pkthdr.len + n->m_pkthdr.len;
4677 	if (txp->plen > 65535)
4678 		return (1);
4679 
4680 	txp->npkt = 2;
4681 	set_mbuf_len16(m, l1);
4682 	set_mbuf_len16(n, l2);
4683 
4684 	return (0);
4685 }
4686 
4687 static int
4688 add_to_txpkts(struct mbuf *m, struct txpkts *txp, u_int available)
4689 {
4690 	u_int plen, len16, needed, nsegs;
4691 
4692 	MPASS(txp->wr_type == 0 || txp->wr_type == 1);
4693 
4694 	nsegs = mbuf_nsegs(m);
4695 	if (needs_tso(m) || (txp->wr_type == 1 && nsegs != 1))
4696 		return (1);
4697 
4698 	plen = txp->plen + m->m_pkthdr.len;
4699 	if (plen > 65535)
4700 		return (1);
4701 
4702 	if (txp->wr_type == 0)
4703 		len16 = txpkts0_len16(nsegs);
4704 	else
4705 		len16 = txpkts1_len16();
4706 	needed = howmany(txp->len16 + len16, EQ_ESIZE / 16);
4707 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4708 		return (1);
4709 
4710 	txp->npkt++;
4711 	txp->plen = plen;
4712 	txp->len16 += len16;
4713 	set_mbuf_len16(m, len16);
4714 
4715 	return (0);
4716 }
4717 
4718 /*
4719  * Write a txpkts WR for the packets in txp to the hardware descriptors, update
4720  * the software descriptor, and advance the pidx.  It is guaranteed that enough
4721  * descriptors are available.
4722  *
4723  * The return value is the # of hardware descriptors used.
4724  */
4725 static u_int
4726 write_txpkts_wr(struct sge_txq *txq, struct fw_eth_tx_pkts_wr *wr,
4727     struct mbuf *m0, const struct txpkts *txp, u_int available)
4728 {
4729 	struct sge_eq *eq = &txq->eq;
4730 	struct tx_sdesc *txsd;
4731 	struct cpl_tx_pkt_core *cpl;
4732 	uint32_t ctrl;
4733 	uint64_t ctrl1;
4734 	int ndesc, checkwrap;
4735 	struct mbuf *m;
4736 	void *flitp;
4737 
4738 	TXQ_LOCK_ASSERT_OWNED(txq);
4739 	MPASS(txp->npkt > 0);
4740 	MPASS(txp->plen < 65536);
4741 	MPASS(m0 != NULL);
4742 	MPASS(m0->m_nextpkt != NULL);
4743 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
4744 	MPASS(available > 0 && available < eq->sidx);
4745 
4746 	ndesc = howmany(txp->len16, EQ_ESIZE / 16);
4747 	MPASS(ndesc <= available);
4748 
4749 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4750 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
4751 	ctrl = V_FW_WR_LEN16(txp->len16);
4752 	wr->equiq_to_len16 = htobe32(ctrl);
4753 	wr->plen = htobe16(txp->plen);
4754 	wr->npkt = txp->npkt;
4755 	wr->r3 = 0;
4756 	wr->type = txp->wr_type;
4757 	flitp = wr + 1;
4758 
4759 	/*
4760 	 * At this point we are 16B into a hardware descriptor.  If checkwrap is
4761 	 * set then we know the WR is going to wrap around somewhere.  We'll
4762 	 * check for that at appropriate points.
4763 	 */
4764 	checkwrap = eq->sidx - ndesc < eq->pidx;
4765 	for (m = m0; m != NULL; m = m->m_nextpkt) {
4766 		if (txp->wr_type == 0) {
4767 			struct ulp_txpkt *ulpmc;
4768 			struct ulptx_idata *ulpsc;
4769 
4770 			/* ULP master command */
4771 			ulpmc = flitp;
4772 			ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
4773 			    V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid));
4774 			ulpmc->len = htobe32(mbuf_len16(m));
4775 
4776 			/* ULP subcommand */
4777 			ulpsc = (void *)(ulpmc + 1);
4778 			ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
4779 			    F_ULP_TX_SC_MORE);
4780 			ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
4781 
4782 			cpl = (void *)(ulpsc + 1);
4783 			if (checkwrap &&
4784 			    (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx])
4785 				cpl = (void *)&eq->desc[0];
4786 		} else {
4787 			cpl = flitp;
4788 		}
4789 
4790 		/* Checksum offload */
4791 		ctrl1 = 0;
4792 		if (needs_l3_csum(m) == 0)
4793 			ctrl1 |= F_TXPKT_IPCSUM_DIS;
4794 		if (needs_l4_csum(m) == 0)
4795 			ctrl1 |= F_TXPKT_L4CSUM_DIS;
4796 		if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4797 		    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4798 			txq->txcsum++;	/* some hardware assistance provided */
4799 
4800 		/* VLAN tag insertion */
4801 		if (needs_vlan_insertion(m)) {
4802 			ctrl1 |= F_TXPKT_VLAN_VLD |
4803 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
4804 			txq->vlan_insertion++;
4805 		}
4806 
4807 		/* CPL header */
4808 		cpl->ctrl0 = txq->cpl_ctrl0;
4809 		cpl->pack = 0;
4810 		cpl->len = htobe16(m->m_pkthdr.len);
4811 		cpl->ctrl1 = htobe64(ctrl1);
4812 
4813 		flitp = cpl + 1;
4814 		if (checkwrap &&
4815 		    (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
4816 			flitp = (void *)&eq->desc[0];
4817 
4818 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap);
4819 
4820 	}
4821 
4822 	if (txp->wr_type == 0) {
4823 		txq->txpkts0_pkts += txp->npkt;
4824 		txq->txpkts0_wrs++;
4825 	} else {
4826 		txq->txpkts1_pkts += txp->npkt;
4827 		txq->txpkts1_wrs++;
4828 	}
4829 
4830 	txsd = &txq->sdesc[eq->pidx];
4831 	txsd->m = m0;
4832 	txsd->desc_used = ndesc;
4833 
4834 	return (ndesc);
4835 }
4836 
4837 /*
4838  * If the SGL ends on an address that is not 16 byte aligned, this function will
4839  * add a 0 filled flit at the end.
4840  */
4841 static void
4842 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap)
4843 {
4844 	struct sge_eq *eq = &txq->eq;
4845 	struct sglist *gl = txq->gl;
4846 	struct sglist_seg *seg;
4847 	__be64 *flitp, *wrap;
4848 	struct ulptx_sgl *usgl;
4849 	int i, nflits, nsegs;
4850 
4851 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
4852 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
4853 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
4854 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
4855 
4856 	get_pkt_gl(m, gl);
4857 	nsegs = gl->sg_nseg;
4858 	MPASS(nsegs > 0);
4859 
4860 	nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
4861 	flitp = (__be64 *)(*to);
4862 	wrap = (__be64 *)(&eq->desc[eq->sidx]);
4863 	seg = &gl->sg_segs[0];
4864 	usgl = (void *)flitp;
4865 
4866 	/*
4867 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
4868 	 * ring, so we're at least 16 bytes away from the status page.  There is
4869 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
4870 	 */
4871 
4872 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
4873 	    V_ULPTX_NSGE(nsegs));
4874 	usgl->len0 = htobe32(seg->ss_len);
4875 	usgl->addr0 = htobe64(seg->ss_paddr);
4876 	seg++;
4877 
4878 	if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) {
4879 
4880 		/* Won't wrap around at all */
4881 
4882 		for (i = 0; i < nsegs - 1; i++, seg++) {
4883 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
4884 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
4885 		}
4886 		if (i & 1)
4887 			usgl->sge[i / 2].len[1] = htobe32(0);
4888 		flitp += nflits;
4889 	} else {
4890 
4891 		/* Will wrap somewhere in the rest of the SGL */
4892 
4893 		/* 2 flits already written, write the rest flit by flit */
4894 		flitp = (void *)(usgl + 1);
4895 		for (i = 0; i < nflits - 2; i++) {
4896 			if (flitp == wrap)
4897 				flitp = (void *)eq->desc;
4898 			*flitp++ = get_flit(seg, nsegs - 1, i);
4899 		}
4900 	}
4901 
4902 	if (nflits & 1) {
4903 		MPASS(((uintptr_t)flitp) & 0xf);
4904 		*flitp++ = 0;
4905 	}
4906 
4907 	MPASS((((uintptr_t)flitp) & 0xf) == 0);
4908 	if (__predict_false(flitp == wrap))
4909 		*to = (void *)eq->desc;
4910 	else
4911 		*to = (void *)flitp;
4912 }
4913 
4914 static inline void
4915 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
4916 {
4917 
4918 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
4919 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
4920 
4921 	if (__predict_true((uintptr_t)(*to) + len <=
4922 	    (uintptr_t)&eq->desc[eq->sidx])) {
4923 		bcopy(from, *to, len);
4924 		(*to) += len;
4925 	} else {
4926 		int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
4927 
4928 		bcopy(from, *to, portion);
4929 		from += portion;
4930 		portion = len - portion;	/* remaining */
4931 		bcopy(from, (void *)eq->desc, portion);
4932 		(*to) = (caddr_t)eq->desc + portion;
4933 	}
4934 }
4935 
4936 static inline void
4937 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n)
4938 {
4939 	u_int db;
4940 
4941 	MPASS(n > 0);
4942 
4943 	db = eq->doorbells;
4944 	if (n > 1)
4945 		clrbit(&db, DOORBELL_WCWR);
4946 	wmb();
4947 
4948 	switch (ffs(db) - 1) {
4949 	case DOORBELL_UDB:
4950 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
4951 		break;
4952 
4953 	case DOORBELL_WCWR: {
4954 		volatile uint64_t *dst, *src;
4955 		int i;
4956 
4957 		/*
4958 		 * Queues whose 128B doorbell segment fits in the page do not
4959 		 * use relative qid (udb_qid is always 0).  Only queues with
4960 		 * doorbell segments can do WCWR.
4961 		 */
4962 		KASSERT(eq->udb_qid == 0 && n == 1,
4963 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
4964 		    __func__, eq->doorbells, n, eq->dbidx, eq));
4965 
4966 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
4967 		    UDBS_DB_OFFSET);
4968 		i = eq->dbidx;
4969 		src = (void *)&eq->desc[i];
4970 		while (src != (void *)&eq->desc[i + 1])
4971 			*dst++ = *src++;
4972 		wmb();
4973 		break;
4974 	}
4975 
4976 	case DOORBELL_UDBWC:
4977 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
4978 		wmb();
4979 		break;
4980 
4981 	case DOORBELL_KDB:
4982 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
4983 		    V_QID(eq->cntxt_id) | V_PIDX(n));
4984 		break;
4985 	}
4986 
4987 	IDXINCR(eq->dbidx, n, eq->sidx);
4988 }
4989 
4990 static inline u_int
4991 reclaimable_tx_desc(struct sge_eq *eq)
4992 {
4993 	uint16_t hw_cidx;
4994 
4995 	hw_cidx = read_hw_cidx(eq);
4996 	return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx));
4997 }
4998 
4999 static inline u_int
5000 total_available_tx_desc(struct sge_eq *eq)
5001 {
5002 	uint16_t hw_cidx, pidx;
5003 
5004 	hw_cidx = read_hw_cidx(eq);
5005 	pidx = eq->pidx;
5006 
5007 	if (pidx == hw_cidx)
5008 		return (eq->sidx - 1);
5009 	else
5010 		return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1);
5011 }
5012 
5013 static inline uint16_t
5014 read_hw_cidx(struct sge_eq *eq)
5015 {
5016 	struct sge_qstat *spg = (void *)&eq->desc[eq->sidx];
5017 	uint16_t cidx = spg->cidx;	/* stable snapshot */
5018 
5019 	return (be16toh(cidx));
5020 }
5021 
5022 /*
5023  * Reclaim 'n' descriptors approximately.
5024  */
5025 static u_int
5026 reclaim_tx_descs(struct sge_txq *txq, u_int n)
5027 {
5028 	struct tx_sdesc *txsd;
5029 	struct sge_eq *eq = &txq->eq;
5030 	u_int can_reclaim, reclaimed;
5031 
5032 	TXQ_LOCK_ASSERT_OWNED(txq);
5033 	MPASS(n > 0);
5034 
5035 	reclaimed = 0;
5036 	can_reclaim = reclaimable_tx_desc(eq);
5037 	while (can_reclaim && reclaimed < n) {
5038 		int ndesc;
5039 		struct mbuf *m, *nextpkt;
5040 
5041 		txsd = &txq->sdesc[eq->cidx];
5042 		ndesc = txsd->desc_used;
5043 
5044 		/* Firmware doesn't return "partial" credits. */
5045 		KASSERT(can_reclaim >= ndesc,
5046 		    ("%s: unexpected number of credits: %d, %d",
5047 		    __func__, can_reclaim, ndesc));
5048 
5049 		for (m = txsd->m; m != NULL; m = nextpkt) {
5050 			nextpkt = m->m_nextpkt;
5051 			m->m_nextpkt = NULL;
5052 			m_freem(m);
5053 		}
5054 		reclaimed += ndesc;
5055 		can_reclaim -= ndesc;
5056 		IDXINCR(eq->cidx, ndesc, eq->sidx);
5057 	}
5058 
5059 	return (reclaimed);
5060 }
5061 
5062 static void
5063 tx_reclaim(void *arg, int n)
5064 {
5065 	struct sge_txq *txq = arg;
5066 	struct sge_eq *eq = &txq->eq;
5067 
5068 	do {
5069 		if (TXQ_TRYLOCK(txq) == 0)
5070 			break;
5071 		n = reclaim_tx_descs(txq, 32);
5072 		if (eq->cidx == eq->pidx)
5073 			eq->equeqidx = eq->pidx;
5074 		TXQ_UNLOCK(txq);
5075 	} while (n > 0);
5076 }
5077 
5078 static __be64
5079 get_flit(struct sglist_seg *segs, int nsegs, int idx)
5080 {
5081 	int i = (idx / 3) * 2;
5082 
5083 	switch (idx % 3) {
5084 	case 0: {
5085 		uint64_t rc;
5086 
5087 		rc = (uint64_t)segs[i].ss_len << 32;
5088 		if (i + 1 < nsegs)
5089 			rc |= (uint64_t)(segs[i + 1].ss_len);
5090 
5091 		return (htobe64(rc));
5092 	}
5093 	case 1:
5094 		return (htobe64(segs[i].ss_paddr));
5095 	case 2:
5096 		return (htobe64(segs[i + 1].ss_paddr));
5097 	}
5098 
5099 	return (0);
5100 }
5101 
5102 static void
5103 find_best_refill_source(struct adapter *sc, struct sge_fl *fl, int maxp)
5104 {
5105 	int8_t zidx, hwidx, idx;
5106 	uint16_t region1, region3;
5107 	int spare, spare_needed, n;
5108 	struct sw_zone_info *swz;
5109 	struct hw_buf_info *hwb, *hwb_list = &sc->sge.hw_buf_info[0];
5110 
5111 	/*
5112 	 * Buffer Packing: Look for PAGE_SIZE or larger zone which has a bufsize
5113 	 * large enough for the max payload and cluster metadata.  Otherwise
5114 	 * settle for the largest bufsize that leaves enough room in the cluster
5115 	 * for metadata.
5116 	 *
5117 	 * Without buffer packing: Look for the smallest zone which has a
5118 	 * bufsize large enough for the max payload.  Settle for the largest
5119 	 * bufsize available if there's nothing big enough for max payload.
5120 	 */
5121 	spare_needed = fl->flags & FL_BUF_PACKING ? CL_METADATA_SIZE : 0;
5122 	swz = &sc->sge.sw_zone_info[0];
5123 	hwidx = -1;
5124 	for (zidx = 0; zidx < SW_ZONE_SIZES; zidx++, swz++) {
5125 		if (swz->size > largest_rx_cluster) {
5126 			if (__predict_true(hwidx != -1))
5127 				break;
5128 
5129 			/*
5130 			 * This is a misconfiguration.  largest_rx_cluster is
5131 			 * preventing us from finding a refill source.  See
5132 			 * dev.t5nex.<n>.buffer_sizes to figure out why.
5133 			 */
5134 			device_printf(sc->dev, "largest_rx_cluster=%u leaves no"
5135 			    " refill source for fl %p (dma %u).  Ignored.\n",
5136 			    largest_rx_cluster, fl, maxp);
5137 		}
5138 		for (idx = swz->head_hwidx; idx != -1; idx = hwb->next) {
5139 			hwb = &hwb_list[idx];
5140 			spare = swz->size - hwb->size;
5141 			if (spare < spare_needed)
5142 				continue;
5143 
5144 			hwidx = idx;		/* best option so far */
5145 			if (hwb->size >= maxp) {
5146 
5147 				if ((fl->flags & FL_BUF_PACKING) == 0)
5148 					goto done; /* stop looking (not packing) */
5149 
5150 				if (swz->size >= safest_rx_cluster)
5151 					goto done; /* stop looking (packing) */
5152 			}
5153 			break;		/* keep looking, next zone */
5154 		}
5155 	}
5156 done:
5157 	/* A usable hwidx has been located. */
5158 	MPASS(hwidx != -1);
5159 	hwb = &hwb_list[hwidx];
5160 	zidx = hwb->zidx;
5161 	swz = &sc->sge.sw_zone_info[zidx];
5162 	region1 = 0;
5163 	region3 = swz->size - hwb->size;
5164 
5165 	/*
5166 	 * Stay within this zone and see if there is a better match when mbuf
5167 	 * inlining is allowed.  Remember that the hwidx's are sorted in
5168 	 * decreasing order of size (so in increasing order of spare area).
5169 	 */
5170 	for (idx = hwidx; idx != -1; idx = hwb->next) {
5171 		hwb = &hwb_list[idx];
5172 		spare = swz->size - hwb->size;
5173 
5174 		if (allow_mbufs_in_cluster == 0 || hwb->size < maxp)
5175 			break;
5176 
5177 		/*
5178 		 * Do not inline mbufs if doing so would violate the pad/pack
5179 		 * boundary alignment requirement.
5180 		 */
5181 		if (fl_pad && (MSIZE % sc->params.sge.pad_boundary) != 0)
5182 			continue;
5183 		if (fl->flags & FL_BUF_PACKING &&
5184 		    (MSIZE % sc->params.sge.pack_boundary) != 0)
5185 			continue;
5186 
5187 		if (spare < CL_METADATA_SIZE + MSIZE)
5188 			continue;
5189 		n = (spare - CL_METADATA_SIZE) / MSIZE;
5190 		if (n > howmany(hwb->size, maxp))
5191 			break;
5192 
5193 		hwidx = idx;
5194 		if (fl->flags & FL_BUF_PACKING) {
5195 			region1 = n * MSIZE;
5196 			region3 = spare - region1;
5197 		} else {
5198 			region1 = MSIZE;
5199 			region3 = spare - region1;
5200 			break;
5201 		}
5202 	}
5203 
5204 	KASSERT(zidx >= 0 && zidx < SW_ZONE_SIZES,
5205 	    ("%s: bad zone %d for fl %p, maxp %d", __func__, zidx, fl, maxp));
5206 	KASSERT(hwidx >= 0 && hwidx <= SGE_FLBUF_SIZES,
5207 	    ("%s: bad hwidx %d for fl %p, maxp %d", __func__, hwidx, fl, maxp));
5208 	KASSERT(region1 + sc->sge.hw_buf_info[hwidx].size + region3 ==
5209 	    sc->sge.sw_zone_info[zidx].size,
5210 	    ("%s: bad buffer layout for fl %p, maxp %d. "
5211 		"cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5212 		sc->sge.sw_zone_info[zidx].size, region1,
5213 		sc->sge.hw_buf_info[hwidx].size, region3));
5214 	if (fl->flags & FL_BUF_PACKING || region1 > 0) {
5215 		KASSERT(region3 >= CL_METADATA_SIZE,
5216 		    ("%s: no room for metadata.  fl %p, maxp %d; "
5217 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5218 		    sc->sge.sw_zone_info[zidx].size, region1,
5219 		    sc->sge.hw_buf_info[hwidx].size, region3));
5220 		KASSERT(region1 % MSIZE == 0,
5221 		    ("%s: bad mbuf region for fl %p, maxp %d. "
5222 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5223 		    sc->sge.sw_zone_info[zidx].size, region1,
5224 		    sc->sge.hw_buf_info[hwidx].size, region3));
5225 	}
5226 
5227 	fl->cll_def.zidx = zidx;
5228 	fl->cll_def.hwidx = hwidx;
5229 	fl->cll_def.region1 = region1;
5230 	fl->cll_def.region3 = region3;
5231 }
5232 
5233 static void
5234 find_safe_refill_source(struct adapter *sc, struct sge_fl *fl)
5235 {
5236 	struct sge *s = &sc->sge;
5237 	struct hw_buf_info *hwb;
5238 	struct sw_zone_info *swz;
5239 	int spare;
5240 	int8_t hwidx;
5241 
5242 	if (fl->flags & FL_BUF_PACKING)
5243 		hwidx = s->safe_hwidx2;	/* with room for metadata */
5244 	else if (allow_mbufs_in_cluster && s->safe_hwidx2 != -1) {
5245 		hwidx = s->safe_hwidx2;
5246 		hwb = &s->hw_buf_info[hwidx];
5247 		swz = &s->sw_zone_info[hwb->zidx];
5248 		spare = swz->size - hwb->size;
5249 
5250 		/* no good if there isn't room for an mbuf as well */
5251 		if (spare < CL_METADATA_SIZE + MSIZE)
5252 			hwidx = s->safe_hwidx1;
5253 	} else
5254 		hwidx = s->safe_hwidx1;
5255 
5256 	if (hwidx == -1) {
5257 		/* No fallback source */
5258 		fl->cll_alt.hwidx = -1;
5259 		fl->cll_alt.zidx = -1;
5260 
5261 		return;
5262 	}
5263 
5264 	hwb = &s->hw_buf_info[hwidx];
5265 	swz = &s->sw_zone_info[hwb->zidx];
5266 	spare = swz->size - hwb->size;
5267 	fl->cll_alt.hwidx = hwidx;
5268 	fl->cll_alt.zidx = hwb->zidx;
5269 	if (allow_mbufs_in_cluster &&
5270 	    (fl_pad == 0 || (MSIZE % sc->params.sge.pad_boundary) == 0))
5271 		fl->cll_alt.region1 = ((spare - CL_METADATA_SIZE) / MSIZE) * MSIZE;
5272 	else
5273 		fl->cll_alt.region1 = 0;
5274 	fl->cll_alt.region3 = spare - fl->cll_alt.region1;
5275 }
5276 
5277 static void
5278 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
5279 {
5280 	mtx_lock(&sc->sfl_lock);
5281 	FL_LOCK(fl);
5282 	if ((fl->flags & FL_DOOMED) == 0) {
5283 		fl->flags |= FL_STARVING;
5284 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
5285 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
5286 	}
5287 	FL_UNLOCK(fl);
5288 	mtx_unlock(&sc->sfl_lock);
5289 }
5290 
5291 static void
5292 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq)
5293 {
5294 	struct sge_wrq *wrq = (void *)eq;
5295 
5296 	atomic_readandclear_int(&eq->equiq);
5297 	taskqueue_enqueue(sc->tq[eq->tx_chan], &wrq->wrq_tx_task);
5298 }
5299 
5300 static void
5301 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq)
5302 {
5303 	struct sge_txq *txq = (void *)eq;
5304 
5305 	MPASS((eq->flags & EQ_TYPEMASK) == EQ_ETH);
5306 
5307 	atomic_readandclear_int(&eq->equiq);
5308 	mp_ring_check_drainage(txq->r, 0);
5309 	taskqueue_enqueue(sc->tq[eq->tx_chan], &txq->tx_reclaim_task);
5310 }
5311 
5312 static int
5313 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
5314     struct mbuf *m)
5315 {
5316 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
5317 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
5318 	struct adapter *sc = iq->adapter;
5319 	struct sge *s = &sc->sge;
5320 	struct sge_eq *eq;
5321 	static void (*h[])(struct adapter *, struct sge_eq *) = {NULL,
5322 		&handle_wrq_egr_update, &handle_eth_egr_update,
5323 		&handle_wrq_egr_update};
5324 
5325 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5326 	    rss->opcode));
5327 
5328 	eq = s->eqmap[qid - s->eq_start - s->eq_base];
5329 	(*h[eq->flags & EQ_TYPEMASK])(sc, eq);
5330 
5331 	return (0);
5332 }
5333 
5334 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
5335 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
5336     offsetof(struct cpl_fw6_msg, data));
5337 
5338 static int
5339 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
5340 {
5341 	struct adapter *sc = iq->adapter;
5342 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
5343 
5344 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5345 	    rss->opcode));
5346 
5347 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
5348 		const struct rss_header *rss2;
5349 
5350 		rss2 = (const struct rss_header *)&cpl->data[0];
5351 		return (t4_cpl_handler[rss2->opcode](iq, rss2, m));
5352 	}
5353 
5354 	return (t4_fw_msg_handler[cpl->type](sc, &cpl->data[0]));
5355 }
5356 
5357 /**
5358  *	t4_handle_wrerr_rpl - process a FW work request error message
5359  *	@adap: the adapter
5360  *	@rpl: start of the FW message
5361  */
5362 static int
5363 t4_handle_wrerr_rpl(struct adapter *adap, const __be64 *rpl)
5364 {
5365 	u8 opcode = *(const u8 *)rpl;
5366 	const struct fw_error_cmd *e = (const void *)rpl;
5367 	unsigned int i;
5368 
5369 	if (opcode != FW_ERROR_CMD) {
5370 		log(LOG_ERR,
5371 		    "%s: Received WRERR_RPL message with opcode %#x\n",
5372 		    device_get_nameunit(adap->dev), opcode);
5373 		return (EINVAL);
5374 	}
5375 	log(LOG_ERR, "%s: FW_ERROR (%s) ", device_get_nameunit(adap->dev),
5376 	    G_FW_ERROR_CMD_FATAL(be32toh(e->op_to_type)) ? "fatal" :
5377 	    "non-fatal");
5378 	switch (G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type))) {
5379 	case FW_ERROR_TYPE_EXCEPTION:
5380 		log(LOG_ERR, "exception info:\n");
5381 		for (i = 0; i < nitems(e->u.exception.info); i++)
5382 			log(LOG_ERR, "%s%08x", i == 0 ? "\t" : " ",
5383 			    be32toh(e->u.exception.info[i]));
5384 		log(LOG_ERR, "\n");
5385 		break;
5386 	case FW_ERROR_TYPE_HWMODULE:
5387 		log(LOG_ERR, "HW module regaddr %08x regval %08x\n",
5388 		    be32toh(e->u.hwmodule.regaddr),
5389 		    be32toh(e->u.hwmodule.regval));
5390 		break;
5391 	case FW_ERROR_TYPE_WR:
5392 		log(LOG_ERR, "WR cidx %d PF %d VF %d eqid %d hdr:\n",
5393 		    be16toh(e->u.wr.cidx),
5394 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.wr.pfn_vfn)),
5395 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.wr.pfn_vfn)),
5396 		    be32toh(e->u.wr.eqid));
5397 		for (i = 0; i < nitems(e->u.wr.wrhdr); i++)
5398 			log(LOG_ERR, "%s%02x", i == 0 ? "\t" : " ",
5399 			    e->u.wr.wrhdr[i]);
5400 		log(LOG_ERR, "\n");
5401 		break;
5402 	case FW_ERROR_TYPE_ACL:
5403 		log(LOG_ERR, "ACL cidx %d PF %d VF %d eqid %d %s",
5404 		    be16toh(e->u.acl.cidx),
5405 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.acl.pfn_vfn)),
5406 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.acl.pfn_vfn)),
5407 		    be32toh(e->u.acl.eqid),
5408 		    G_FW_ERROR_CMD_MV(be16toh(e->u.acl.mv_pkd)) ? "vlanid" :
5409 		    "MAC");
5410 		for (i = 0; i < nitems(e->u.acl.val); i++)
5411 			log(LOG_ERR, " %02x", e->u.acl.val[i]);
5412 		log(LOG_ERR, "\n");
5413 		break;
5414 	default:
5415 		log(LOG_ERR, "type %#x\n",
5416 		    G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type)));
5417 		return (EINVAL);
5418 	}
5419 	return (0);
5420 }
5421 
5422 static int
5423 sysctl_uint16(SYSCTL_HANDLER_ARGS)
5424 {
5425 	uint16_t *id = arg1;
5426 	int i = *id;
5427 
5428 	return sysctl_handle_int(oidp, &i, 0, req);
5429 }
5430 
5431 static int
5432 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
5433 {
5434 	struct sge *s = arg1;
5435 	struct hw_buf_info *hwb = &s->hw_buf_info[0];
5436 	struct sw_zone_info *swz = &s->sw_zone_info[0];
5437 	int i, rc;
5438 	struct sbuf sb;
5439 	char c;
5440 
5441 	sbuf_new(&sb, NULL, 32, SBUF_AUTOEXTEND);
5442 	for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
5443 		if (hwb->zidx >= 0 && swz[hwb->zidx].size <= largest_rx_cluster)
5444 			c = '*';
5445 		else
5446 			c = '\0';
5447 
5448 		sbuf_printf(&sb, "%u%c ", hwb->size, c);
5449 	}
5450 	sbuf_trim(&sb);
5451 	sbuf_finish(&sb);
5452 	rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
5453 	sbuf_delete(&sb);
5454 	return (rc);
5455 }
5456 
5457 #ifdef RATELIMIT
5458 /*
5459  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
5460  */
5461 static inline u_int
5462 txpkt_eo_len16(u_int nsegs, u_int immhdrs, u_int tso)
5463 {
5464 	u_int n;
5465 
5466 	MPASS(immhdrs > 0);
5467 
5468 	n = roundup2(sizeof(struct fw_eth_tx_eo_wr) +
5469 	    sizeof(struct cpl_tx_pkt_core) + immhdrs, 16);
5470 	if (__predict_false(nsegs == 0))
5471 		goto done;
5472 
5473 	nsegs--; /* first segment is part of ulptx_sgl */
5474 	n += sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
5475 	if (tso)
5476 		n += sizeof(struct cpl_tx_pkt_lso_core);
5477 
5478 done:
5479 	return (howmany(n, 16));
5480 }
5481 
5482 #define ETID_FLOWC_NPARAMS 6
5483 #define ETID_FLOWC_LEN (roundup2((sizeof(struct fw_flowc_wr) + \
5484     ETID_FLOWC_NPARAMS * sizeof(struct fw_flowc_mnemval)), 16))
5485 #define ETID_FLOWC_LEN16 (howmany(ETID_FLOWC_LEN, 16))
5486 
5487 static int
5488 send_etid_flowc_wr(struct cxgbe_snd_tag *cst, struct port_info *pi,
5489     struct vi_info *vi)
5490 {
5491 	struct wrq_cookie cookie;
5492 	u_int pfvf = G_FW_VIID_PFN(vi->viid) << S_FW_VIID_PFN;
5493 	struct fw_flowc_wr *flowc;
5494 
5495 	mtx_assert(&cst->lock, MA_OWNED);
5496 	MPASS((cst->flags & (EO_FLOWC_PENDING | EO_FLOWC_RPL_PENDING)) ==
5497 	    EO_FLOWC_PENDING);
5498 
5499 	flowc = start_wrq_wr(cst->eo_txq, ETID_FLOWC_LEN16, &cookie);
5500 	if (__predict_false(flowc == NULL))
5501 		return (ENOMEM);
5502 
5503 	bzero(flowc, ETID_FLOWC_LEN);
5504 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
5505 	    V_FW_FLOWC_WR_NPARAMS(ETID_FLOWC_NPARAMS) | V_FW_WR_COMPL(0));
5506 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(ETID_FLOWC_LEN16) |
5507 	    V_FW_WR_FLOWID(cst->etid));
5508 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN;
5509 	flowc->mnemval[0].val = htobe32(pfvf);
5510 	flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH;
5511 	flowc->mnemval[1].val = htobe32(pi->tx_chan);
5512 	flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT;
5513 	flowc->mnemval[2].val = htobe32(pi->tx_chan);
5514 	flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID;
5515 	flowc->mnemval[3].val = htobe32(cst->iqid);
5516 	flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_EOSTATE;
5517 	flowc->mnemval[4].val = htobe32(FW_FLOWC_MNEM_EOSTATE_ESTABLISHED);
5518 	flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS;
5519 	flowc->mnemval[5].val = htobe32(cst->schedcl);
5520 
5521 	commit_wrq_wr(cst->eo_txq, flowc, &cookie);
5522 
5523 	cst->flags &= ~EO_FLOWC_PENDING;
5524 	cst->flags |= EO_FLOWC_RPL_PENDING;
5525 	MPASS(cst->tx_credits >= ETID_FLOWC_LEN16);	/* flowc is first WR. */
5526 	cst->tx_credits -= ETID_FLOWC_LEN16;
5527 
5528 	return (0);
5529 }
5530 
5531 #define ETID_FLUSH_LEN16 (howmany(sizeof (struct fw_flowc_wr), 16))
5532 
5533 void
5534 send_etid_flush_wr(struct cxgbe_snd_tag *cst)
5535 {
5536 	struct fw_flowc_wr *flowc;
5537 	struct wrq_cookie cookie;
5538 
5539 	mtx_assert(&cst->lock, MA_OWNED);
5540 
5541 	flowc = start_wrq_wr(cst->eo_txq, ETID_FLUSH_LEN16, &cookie);
5542 	if (__predict_false(flowc == NULL))
5543 		CXGBE_UNIMPLEMENTED(__func__);
5544 
5545 	bzero(flowc, ETID_FLUSH_LEN16 * 16);
5546 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
5547 	    V_FW_FLOWC_WR_NPARAMS(0) | F_FW_WR_COMPL);
5548 	flowc->flowid_len16 = htobe32(V_FW_WR_LEN16(ETID_FLUSH_LEN16) |
5549 	    V_FW_WR_FLOWID(cst->etid));
5550 
5551 	commit_wrq_wr(cst->eo_txq, flowc, &cookie);
5552 
5553 	cst->flags |= EO_FLUSH_RPL_PENDING;
5554 	MPASS(cst->tx_credits >= ETID_FLUSH_LEN16);
5555 	cst->tx_credits -= ETID_FLUSH_LEN16;
5556 	cst->ncompl++;
5557 }
5558 
5559 static void
5560 write_ethofld_wr(struct cxgbe_snd_tag *cst, struct fw_eth_tx_eo_wr *wr,
5561     struct mbuf *m0, int compl)
5562 {
5563 	struct cpl_tx_pkt_core *cpl;
5564 	uint64_t ctrl1;
5565 	uint32_t ctrl;	/* used in many unrelated places */
5566 	int len16, pktlen, nsegs, immhdrs;
5567 	caddr_t dst;
5568 	uintptr_t p;
5569 	struct ulptx_sgl *usgl;
5570 	struct sglist sg;
5571 	struct sglist_seg segs[38];	/* XXX: find real limit.  XXX: get off the stack */
5572 
5573 	mtx_assert(&cst->lock, MA_OWNED);
5574 	M_ASSERTPKTHDR(m0);
5575 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
5576 	    m0->m_pkthdr.l4hlen > 0,
5577 	    ("%s: ethofld mbuf %p is missing header lengths", __func__, m0));
5578 
5579 	if (needs_udp_csum(m0)) {
5580 		CXGBE_UNIMPLEMENTED("UDP ethofld");
5581 	}
5582 
5583 	len16 = mbuf_eo_len16(m0);
5584 	nsegs = mbuf_eo_nsegs(m0);
5585 	pktlen = m0->m_pkthdr.len;
5586 	ctrl = sizeof(struct cpl_tx_pkt_core);
5587 	if (needs_tso(m0))
5588 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
5589 	immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen + m0->m_pkthdr.l4hlen;
5590 	ctrl += immhdrs;
5591 
5592 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_EO_WR) |
5593 	    V_FW_ETH_TX_EO_WR_IMMDLEN(ctrl) | V_FW_WR_COMPL(!!compl));
5594 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(len16) |
5595 	    V_FW_WR_FLOWID(cst->etid));
5596 	wr->r3 = 0;
5597 	wr->u.tcpseg.type = FW_ETH_TX_EO_TYPE_TCPSEG;
5598 	wr->u.tcpseg.ethlen = m0->m_pkthdr.l2hlen;
5599 	wr->u.tcpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
5600 	wr->u.tcpseg.tcplen = m0->m_pkthdr.l4hlen;
5601 	wr->u.tcpseg.tsclk_tsoff = mbuf_eo_tsclk_tsoff(m0);
5602 	wr->u.tcpseg.r4 = 0;
5603 	wr->u.tcpseg.r5 = 0;
5604 	wr->u.tcpseg.plen = htobe32(pktlen - immhdrs);
5605 
5606 	if (needs_tso(m0)) {
5607 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
5608 
5609 		wr->u.tcpseg.mss = htobe16(m0->m_pkthdr.tso_segsz);
5610 
5611 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
5612 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
5613 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
5614 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
5615 			ctrl |= V_LSO_ETHHDR_LEN(1);
5616 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
5617 			ctrl |= F_LSO_IPV6;
5618 		lso->lso_ctrl = htobe32(ctrl);
5619 		lso->ipid_ofst = htobe16(0);
5620 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
5621 		lso->seqno_offset = htobe32(0);
5622 		lso->len = htobe32(pktlen);
5623 
5624 		cpl = (void *)(lso + 1);
5625 	} else {
5626 		wr->u.tcpseg.mss = htobe16(0xffff);
5627 		cpl = (void *)(wr + 1);
5628 	}
5629 
5630 	/* Checksum offload must be requested for ethofld. */
5631 	ctrl1 = 0;
5632 	MPASS(needs_l4_csum(m0));
5633 
5634 	/* VLAN tag insertion */
5635 	if (needs_vlan_insertion(m0)) {
5636 		ctrl1 |= F_TXPKT_VLAN_VLD |
5637 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
5638 	}
5639 
5640 	/* CPL header */
5641 	cpl->ctrl0 = cst->ctrl0;
5642 	cpl->pack = 0;
5643 	cpl->len = htobe16(pktlen);
5644 	cpl->ctrl1 = htobe64(ctrl1);
5645 
5646 	/* Copy Ethernet, IP & TCP hdrs as immediate data */
5647 	p = (uintptr_t)(cpl + 1);
5648 	m_copydata(m0, 0, immhdrs, (void *)p);
5649 
5650 	/* SGL */
5651 	dst = (void *)(cpl + 1);
5652 	if (nsegs > 0) {
5653 		int i, pad;
5654 
5655 		/* zero-pad upto next 16Byte boundary, if not 16Byte aligned */
5656 		p += immhdrs;
5657 		pad = 16 - (immhdrs & 0xf);
5658 		bzero((void *)p, pad);
5659 
5660 		usgl = (void *)(p + pad);
5661 		usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
5662 		    V_ULPTX_NSGE(nsegs));
5663 
5664 		sglist_init(&sg, nitems(segs), segs);
5665 		for (; m0 != NULL; m0 = m0->m_next) {
5666 			if (__predict_false(m0->m_len == 0))
5667 				continue;
5668 			if (immhdrs >= m0->m_len) {
5669 				immhdrs -= m0->m_len;
5670 				continue;
5671 			}
5672 
5673 			sglist_append(&sg, mtod(m0, char *) + immhdrs,
5674 			    m0->m_len - immhdrs);
5675 			immhdrs = 0;
5676 		}
5677 		MPASS(sg.sg_nseg == nsegs);
5678 
5679 		/*
5680 		 * Zero pad last 8B in case the WR doesn't end on a 16B
5681 		 * boundary.
5682 		 */
5683 		*(uint64_t *)((char *)wr + len16 * 16 - 8) = 0;
5684 
5685 		usgl->len0 = htobe32(segs[0].ss_len);
5686 		usgl->addr0 = htobe64(segs[0].ss_paddr);
5687 		for (i = 0; i < nsegs - 1; i++) {
5688 			usgl->sge[i / 2].len[i & 1] = htobe32(segs[i + 1].ss_len);
5689 			usgl->sge[i / 2].addr[i & 1] = htobe64(segs[i + 1].ss_paddr);
5690 		}
5691 		if (i & 1)
5692 			usgl->sge[i / 2].len[1] = htobe32(0);
5693 	}
5694 
5695 }
5696 
5697 static void
5698 ethofld_tx(struct cxgbe_snd_tag *cst)
5699 {
5700 	struct mbuf *m;
5701 	struct wrq_cookie cookie;
5702 	int next_credits, compl;
5703 	struct fw_eth_tx_eo_wr *wr;
5704 
5705 	mtx_assert(&cst->lock, MA_OWNED);
5706 
5707 	while ((m = mbufq_first(&cst->pending_tx)) != NULL) {
5708 		M_ASSERTPKTHDR(m);
5709 
5710 		/* How many len16 credits do we need to send this mbuf. */
5711 		next_credits = mbuf_eo_len16(m);
5712 		MPASS(next_credits > 0);
5713 		if (next_credits > cst->tx_credits) {
5714 			/*
5715 			 * Tx will make progress eventually because there is at
5716 			 * least one outstanding fw4_ack that will return
5717 			 * credits and kick the tx.
5718 			 */
5719 			MPASS(cst->ncompl > 0);
5720 			return;
5721 		}
5722 		wr = start_wrq_wr(cst->eo_txq, next_credits, &cookie);
5723 		if (__predict_false(wr == NULL)) {
5724 			/* XXX: wishful thinking, not a real assertion. */
5725 			MPASS(cst->ncompl > 0);
5726 			return;
5727 		}
5728 		cst->tx_credits -= next_credits;
5729 		cst->tx_nocompl += next_credits;
5730 		compl = cst->ncompl == 0 || cst->tx_nocompl >= cst->tx_total / 2;
5731 		ETHER_BPF_MTAP(cst->com.ifp, m);
5732 		write_ethofld_wr(cst, wr, m, compl);
5733 		commit_wrq_wr(cst->eo_txq, wr, &cookie);
5734 		if (compl) {
5735 			cst->ncompl++;
5736 			cst->tx_nocompl	= 0;
5737 		}
5738 		(void) mbufq_dequeue(&cst->pending_tx);
5739 		mbufq_enqueue(&cst->pending_fwack, m);
5740 	}
5741 }
5742 
5743 int
5744 ethofld_transmit(struct ifnet *ifp, struct mbuf *m0)
5745 {
5746 	struct cxgbe_snd_tag *cst;
5747 	int rc;
5748 
5749 	MPASS(m0->m_nextpkt == NULL);
5750 	MPASS(m0->m_pkthdr.snd_tag != NULL);
5751 	cst = mst_to_cst(m0->m_pkthdr.snd_tag);
5752 
5753 	mtx_lock(&cst->lock);
5754 	MPASS(cst->flags & EO_SND_TAG_REF);
5755 
5756 	if (__predict_false(cst->flags & EO_FLOWC_PENDING)) {
5757 		struct vi_info *vi = ifp->if_softc;
5758 		struct port_info *pi = vi->pi;
5759 		struct adapter *sc = pi->adapter;
5760 		const uint32_t rss_mask = vi->rss_size - 1;
5761 		uint32_t rss_hash;
5762 
5763 		cst->eo_txq = &sc->sge.ofld_txq[vi->first_ofld_txq];
5764 		if (M_HASHTYPE_ISHASH(m0))
5765 			rss_hash = m0->m_pkthdr.flowid;
5766 		else
5767 			rss_hash = arc4random();
5768 		/* We assume RSS hashing */
5769 		cst->iqid = vi->rss[rss_hash & rss_mask];
5770 		cst->eo_txq += rss_hash % vi->nofldtxq;
5771 		rc = send_etid_flowc_wr(cst, pi, vi);
5772 		if (rc != 0)
5773 			goto done;
5774 	}
5775 
5776 	if (__predict_false(cst->plen + m0->m_pkthdr.len > eo_max_backlog)) {
5777 		rc = ENOBUFS;
5778 		goto done;
5779 	}
5780 
5781 	mbufq_enqueue(&cst->pending_tx, m0);
5782 	cst->plen += m0->m_pkthdr.len;
5783 
5784 	ethofld_tx(cst);
5785 	rc = 0;
5786 done:
5787 	mtx_unlock(&cst->lock);
5788 	if (__predict_false(rc != 0))
5789 		m_freem(m0);
5790 	return (rc);
5791 }
5792 
5793 static int
5794 ethofld_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
5795 {
5796 	struct adapter *sc = iq->adapter;
5797 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
5798 	struct mbuf *m;
5799 	u_int etid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
5800 	struct cxgbe_snd_tag *cst;
5801 	uint8_t credits = cpl->credits;
5802 
5803 	cst = lookup_etid(sc, etid);
5804 	mtx_lock(&cst->lock);
5805 	if (__predict_false(cst->flags & EO_FLOWC_RPL_PENDING)) {
5806 		MPASS(credits >= ETID_FLOWC_LEN16);
5807 		credits -= ETID_FLOWC_LEN16;
5808 		cst->flags &= ~EO_FLOWC_RPL_PENDING;
5809 	}
5810 
5811 	KASSERT(cst->ncompl > 0,
5812 	    ("%s: etid %u (%p) wasn't expecting completion.",
5813 	    __func__, etid, cst));
5814 	cst->ncompl--;
5815 
5816 	while (credits > 0) {
5817 		m = mbufq_dequeue(&cst->pending_fwack);
5818 		if (__predict_false(m == NULL)) {
5819 			/*
5820 			 * The remaining credits are for the final flush that
5821 			 * was issued when the tag was freed by the kernel.
5822 			 */
5823 			MPASS((cst->flags &
5824 			    (EO_FLUSH_RPL_PENDING | EO_SND_TAG_REF)) ==
5825 			    EO_FLUSH_RPL_PENDING);
5826 			MPASS(credits == ETID_FLUSH_LEN16);
5827 			MPASS(cst->tx_credits + cpl->credits == cst->tx_total);
5828 			MPASS(cst->ncompl == 0);
5829 
5830 			cst->flags &= ~EO_FLUSH_RPL_PENDING;
5831 			cst->tx_credits += cpl->credits;
5832 freetag:
5833 			cxgbe_snd_tag_free_locked(cst);
5834 			return (0);	/* cst is gone. */
5835 		}
5836 		KASSERT(m != NULL,
5837 		    ("%s: too many credits (%u, %u)", __func__, cpl->credits,
5838 		    credits));
5839 		KASSERT(credits >= mbuf_eo_len16(m),
5840 		    ("%s: too few credits (%u, %u, %u)", __func__,
5841 		    cpl->credits, credits, mbuf_eo_len16(m)));
5842 		credits -= mbuf_eo_len16(m);
5843 		cst->plen -= m->m_pkthdr.len;
5844 		m_freem(m);
5845 	}
5846 
5847 	cst->tx_credits += cpl->credits;
5848 	MPASS(cst->tx_credits <= cst->tx_total);
5849 
5850 	m = mbufq_first(&cst->pending_tx);
5851 	if (m != NULL && cst->tx_credits >= mbuf_eo_len16(m))
5852 		ethofld_tx(cst);
5853 
5854 	if (__predict_false((cst->flags & EO_SND_TAG_REF) == 0) &&
5855 	    cst->ncompl == 0) {
5856 		if (cst->tx_credits == cst->tx_total)
5857 			goto freetag;
5858 		else {
5859 			MPASS((cst->flags & EO_FLUSH_RPL_PENDING) == 0);
5860 			send_etid_flush_wr(cst);
5861 		}
5862 	}
5863 
5864 	mtx_unlock(&cst->lock);
5865 
5866 	return (0);
5867 }
5868 #endif
5869