xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision 0070b575f47d9fa116f7c4367faed3e1b5f88555)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2011 Chelsio Communications, Inc.
5  * All rights reserved.
6  * Written by: Navdeep Parhar <np@FreeBSD.org>
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include "opt_inet.h"
34 #include "opt_inet6.h"
35 #include "opt_ratelimit.h"
36 
37 #include <sys/types.h>
38 #include <sys/eventhandler.h>
39 #include <sys/mbuf.h>
40 #include <sys/socket.h>
41 #include <sys/kernel.h>
42 #include <sys/malloc.h>
43 #include <sys/queue.h>
44 #include <sys/sbuf.h>
45 #include <sys/taskqueue.h>
46 #include <sys/time.h>
47 #include <sys/sglist.h>
48 #include <sys/sysctl.h>
49 #include <sys/smp.h>
50 #include <sys/counter.h>
51 #include <net/bpf.h>
52 #include <net/ethernet.h>
53 #include <net/if.h>
54 #include <net/if_vlan_var.h>
55 #include <netinet/in.h>
56 #include <netinet/ip.h>
57 #include <netinet/ip6.h>
58 #include <netinet/tcp.h>
59 #include <netinet/udp.h>
60 #include <machine/in_cksum.h>
61 #include <machine/md_var.h>
62 #include <vm/vm.h>
63 #include <vm/pmap.h>
64 #ifdef DEV_NETMAP
65 #include <machine/bus.h>
66 #include <sys/selinfo.h>
67 #include <net/if_var.h>
68 #include <net/netmap.h>
69 #include <dev/netmap/netmap_kern.h>
70 #endif
71 
72 #include "common/common.h"
73 #include "common/t4_regs.h"
74 #include "common/t4_regs_values.h"
75 #include "common/t4_msg.h"
76 #include "t4_l2t.h"
77 #include "t4_mp_ring.h"
78 
79 #ifdef T4_PKT_TIMESTAMP
80 #define RX_COPY_THRESHOLD (MINCLSIZE - 8)
81 #else
82 #define RX_COPY_THRESHOLD MINCLSIZE
83 #endif
84 
85 /* Internal mbuf flags stored in PH_loc.eight[1]. */
86 #define	MC_NOMAP		0x01
87 #define	MC_RAW_WR		0x02
88 
89 /*
90  * Ethernet frames are DMA'd at this byte offset into the freelist buffer.
91  * 0-7 are valid values.
92  */
93 static int fl_pktshift = 0;
94 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pktshift, CTLFLAG_RDTUN, &fl_pktshift, 0,
95     "payload DMA offset in rx buffer (bytes)");
96 
97 /*
98  * Pad ethernet payload up to this boundary.
99  * -1: driver should figure out a good value.
100  *  0: disable padding.
101  *  Any power of 2 from 32 to 4096 (both inclusive) is also a valid value.
102  */
103 int fl_pad = -1;
104 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pad, CTLFLAG_RDTUN, &fl_pad, 0,
105     "payload pad boundary (bytes)");
106 
107 /*
108  * Status page length.
109  * -1: driver should figure out a good value.
110  *  64 or 128 are the only other valid values.
111  */
112 static int spg_len = -1;
113 SYSCTL_INT(_hw_cxgbe, OID_AUTO, spg_len, CTLFLAG_RDTUN, &spg_len, 0,
114     "status page size (bytes)");
115 
116 /*
117  * Congestion drops.
118  * -1: no congestion feedback (not recommended).
119  *  0: backpressure the channel instead of dropping packets right away.
120  *  1: no backpressure, drop packets for the congested queue immediately.
121  */
122 static int cong_drop = 0;
123 SYSCTL_INT(_hw_cxgbe, OID_AUTO, cong_drop, CTLFLAG_RDTUN, &cong_drop, 0,
124     "Congestion control for RX queues (0 = backpressure, 1 = drop");
125 
126 /*
127  * Deliver multiple frames in the same free list buffer if they fit.
128  * -1: let the driver decide whether to enable buffer packing or not.
129  *  0: disable buffer packing.
130  *  1: enable buffer packing.
131  */
132 static int buffer_packing = -1;
133 SYSCTL_INT(_hw_cxgbe, OID_AUTO, buffer_packing, CTLFLAG_RDTUN, &buffer_packing,
134     0, "Enable buffer packing");
135 
136 /*
137  * Start next frame in a packed buffer at this boundary.
138  * -1: driver should figure out a good value.
139  * T4: driver will ignore this and use the same value as fl_pad above.
140  * T5: 16, or a power of 2 from 64 to 4096 (both inclusive) is a valid value.
141  */
142 static int fl_pack = -1;
143 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pack, CTLFLAG_RDTUN, &fl_pack, 0,
144     "payload pack boundary (bytes)");
145 
146 /*
147  * Allow the driver to create mbuf(s) in a cluster allocated for rx.
148  * 0: never; always allocate mbufs from the zone_mbuf UMA zone.
149  * 1: ok to create mbuf(s) within a cluster if there is room.
150  */
151 static int allow_mbufs_in_cluster = 1;
152 SYSCTL_INT(_hw_cxgbe, OID_AUTO, allow_mbufs_in_cluster, CTLFLAG_RDTUN,
153     &allow_mbufs_in_cluster, 0,
154     "Allow driver to create mbufs within a rx cluster");
155 
156 /*
157  * Largest rx cluster size that the driver is allowed to allocate.
158  */
159 static int largest_rx_cluster = MJUM16BYTES;
160 SYSCTL_INT(_hw_cxgbe, OID_AUTO, largest_rx_cluster, CTLFLAG_RDTUN,
161     &largest_rx_cluster, 0, "Largest rx cluster (bytes)");
162 
163 /*
164  * Size of cluster allocation that's most likely to succeed.  The driver will
165  * fall back to this size if it fails to allocate clusters larger than this.
166  */
167 static int safest_rx_cluster = PAGE_SIZE;
168 SYSCTL_INT(_hw_cxgbe, OID_AUTO, safest_rx_cluster, CTLFLAG_RDTUN,
169     &safest_rx_cluster, 0, "Safe rx cluster (bytes)");
170 
171 #ifdef RATELIMIT
172 /*
173  * Knob to control TCP timestamp rewriting, and the granularity of the tick used
174  * for rewriting.  -1 and 0-3 are all valid values.
175  * -1: hardware should leave the TCP timestamps alone.
176  * 0: 1ms
177  * 1: 100us
178  * 2: 10us
179  * 3: 1us
180  */
181 static int tsclk = -1;
182 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tsclk, CTLFLAG_RDTUN, &tsclk, 0,
183     "Control TCP timestamp rewriting when using pacing");
184 
185 static int eo_max_backlog = 1024 * 1024;
186 SYSCTL_INT(_hw_cxgbe, OID_AUTO, eo_max_backlog, CTLFLAG_RDTUN, &eo_max_backlog,
187     0, "Maximum backlog of ratelimited data per flow");
188 #endif
189 
190 /*
191  * The interrupt holdoff timers are multiplied by this value on T6+.
192  * 1 and 3-17 (both inclusive) are legal values.
193  */
194 static int tscale = 1;
195 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tscale, CTLFLAG_RDTUN, &tscale, 0,
196     "Interrupt holdoff timer scale on T6+");
197 
198 /*
199  * Number of LRO entries in the lro_ctrl structure per rx queue.
200  */
201 static int lro_entries = TCP_LRO_ENTRIES;
202 SYSCTL_INT(_hw_cxgbe, OID_AUTO, lro_entries, CTLFLAG_RDTUN, &lro_entries, 0,
203     "Number of LRO entries per RX queue");
204 
205 /*
206  * This enables presorting of frames before they're fed into tcp_lro_rx.
207  */
208 static int lro_mbufs = 0;
209 SYSCTL_INT(_hw_cxgbe, OID_AUTO, lro_mbufs, CTLFLAG_RDTUN, &lro_mbufs, 0,
210     "Enable presorting of LRO frames");
211 
212 struct txpkts {
213 	u_int wr_type;		/* type 0 or type 1 */
214 	u_int npkt;		/* # of packets in this work request */
215 	u_int plen;		/* total payload (sum of all packets) */
216 	u_int len16;		/* # of 16B pieces used by this work request */
217 };
218 
219 /* A packet's SGL.  This + m_pkthdr has all info needed for tx */
220 struct sgl {
221 	struct sglist sg;
222 	struct sglist_seg seg[TX_SGL_SEGS];
223 };
224 
225 static int service_iq(struct sge_iq *, int);
226 static int service_iq_fl(struct sge_iq *, int);
227 static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t);
228 static int t4_eth_rx(struct sge_iq *, const struct rss_header *, struct mbuf *);
229 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int);
230 static inline void init_fl(struct adapter *, struct sge_fl *, int, int, char *);
231 static inline void init_eq(struct adapter *, struct sge_eq *, int, int, uint8_t,
232     uint16_t, char *);
233 static int alloc_ring(struct adapter *, size_t, bus_dma_tag_t *, bus_dmamap_t *,
234     bus_addr_t *, void **);
235 static int free_ring(struct adapter *, bus_dma_tag_t, bus_dmamap_t, bus_addr_t,
236     void *);
237 static int alloc_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *,
238     int, int);
239 static int free_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *);
240 static void add_iq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
241     struct sge_iq *);
242 static void add_fl_sysctls(struct adapter *, struct sysctl_ctx_list *,
243     struct sysctl_oid *, struct sge_fl *);
244 static int alloc_fwq(struct adapter *);
245 static int free_fwq(struct adapter *);
246 static int alloc_ctrlq(struct adapter *, struct sge_wrq *, int,
247     struct sysctl_oid *);
248 static int alloc_rxq(struct vi_info *, struct sge_rxq *, int, int,
249     struct sysctl_oid *);
250 static int free_rxq(struct vi_info *, struct sge_rxq *);
251 #ifdef TCP_OFFLOAD
252 static int alloc_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *, int, int,
253     struct sysctl_oid *);
254 static int free_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *);
255 #endif
256 #ifdef DEV_NETMAP
257 static int alloc_nm_rxq(struct vi_info *, struct sge_nm_rxq *, int, int,
258     struct sysctl_oid *);
259 static int free_nm_rxq(struct vi_info *, struct sge_nm_rxq *);
260 static int alloc_nm_txq(struct vi_info *, struct sge_nm_txq *, int, int,
261     struct sysctl_oid *);
262 static int free_nm_txq(struct vi_info *, struct sge_nm_txq *);
263 #endif
264 static int ctrl_eq_alloc(struct adapter *, struct sge_eq *);
265 static int eth_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
266 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
267 static int ofld_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
268 #endif
269 static int alloc_eq(struct adapter *, struct vi_info *, struct sge_eq *);
270 static int free_eq(struct adapter *, struct sge_eq *);
271 static int alloc_wrq(struct adapter *, struct vi_info *, struct sge_wrq *,
272     struct sysctl_oid *);
273 static int free_wrq(struct adapter *, struct sge_wrq *);
274 static int alloc_txq(struct vi_info *, struct sge_txq *, int,
275     struct sysctl_oid *);
276 static int free_txq(struct vi_info *, struct sge_txq *);
277 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int);
278 static inline void ring_fl_db(struct adapter *, struct sge_fl *);
279 static int refill_fl(struct adapter *, struct sge_fl *, int);
280 static void refill_sfl(void *);
281 static int alloc_fl_sdesc(struct sge_fl *);
282 static void free_fl_sdesc(struct adapter *, struct sge_fl *);
283 static void find_best_refill_source(struct adapter *, struct sge_fl *, int);
284 static void find_safe_refill_source(struct adapter *, struct sge_fl *);
285 static void add_fl_to_sfl(struct adapter *, struct sge_fl *);
286 
287 static inline void get_pkt_gl(struct mbuf *, struct sglist *);
288 static inline u_int txpkt_len16(u_int, u_int);
289 static inline u_int txpkt_vm_len16(u_int, u_int);
290 static inline u_int txpkts0_len16(u_int);
291 static inline u_int txpkts1_len16(void);
292 static u_int write_raw_wr(struct sge_txq *, void *, struct mbuf *, u_int);
293 static u_int write_txpkt_wr(struct sge_txq *, struct fw_eth_tx_pkt_wr *,
294     struct mbuf *, u_int);
295 static u_int write_txpkt_vm_wr(struct adapter *, struct sge_txq *,
296     struct fw_eth_tx_pkt_vm_wr *, struct mbuf *, u_int);
297 static int try_txpkts(struct mbuf *, struct mbuf *, struct txpkts *, u_int);
298 static int add_to_txpkts(struct mbuf *, struct txpkts *, u_int);
299 static u_int write_txpkts_wr(struct sge_txq *, struct fw_eth_tx_pkts_wr *,
300     struct mbuf *, const struct txpkts *, u_int);
301 static void write_gl_to_txd(struct sge_txq *, struct mbuf *, caddr_t *, int);
302 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int);
303 static inline void ring_eq_db(struct adapter *, struct sge_eq *, u_int);
304 static inline uint16_t read_hw_cidx(struct sge_eq *);
305 static inline u_int reclaimable_tx_desc(struct sge_eq *);
306 static inline u_int total_available_tx_desc(struct sge_eq *);
307 static u_int reclaim_tx_descs(struct sge_txq *, u_int);
308 static void tx_reclaim(void *, int);
309 static __be64 get_flit(struct sglist_seg *, int, int);
310 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
311     struct mbuf *);
312 static int handle_fw_msg(struct sge_iq *, const struct rss_header *,
313     struct mbuf *);
314 static int t4_handle_wrerr_rpl(struct adapter *, const __be64 *);
315 static void wrq_tx_drain(void *, int);
316 static void drain_wrq_wr_list(struct adapter *, struct sge_wrq *);
317 
318 static int sysctl_uint16(SYSCTL_HANDLER_ARGS);
319 static int sysctl_bufsizes(SYSCTL_HANDLER_ARGS);
320 #ifdef RATELIMIT
321 static inline u_int txpkt_eo_len16(u_int, u_int, u_int);
322 static int ethofld_fw4_ack(struct sge_iq *, const struct rss_header *,
323     struct mbuf *);
324 #endif
325 
326 static counter_u64_t extfree_refs;
327 static counter_u64_t extfree_rels;
328 
329 an_handler_t t4_an_handler;
330 fw_msg_handler_t t4_fw_msg_handler[NUM_FW6_TYPES];
331 cpl_handler_t t4_cpl_handler[NUM_CPL_CMDS];
332 cpl_handler_t set_tcb_rpl_handlers[NUM_CPL_COOKIES];
333 cpl_handler_t l2t_write_rpl_handlers[NUM_CPL_COOKIES];
334 cpl_handler_t act_open_rpl_handlers[NUM_CPL_COOKIES];
335 cpl_handler_t abort_rpl_rss_handlers[NUM_CPL_COOKIES];
336 cpl_handler_t fw4_ack_handlers[NUM_CPL_COOKIES];
337 
338 void
339 t4_register_an_handler(an_handler_t h)
340 {
341 	uintptr_t *loc;
342 
343 	MPASS(h == NULL || t4_an_handler == NULL);
344 
345 	loc = (uintptr_t *)&t4_an_handler;
346 	atomic_store_rel_ptr(loc, (uintptr_t)h);
347 }
348 
349 void
350 t4_register_fw_msg_handler(int type, fw_msg_handler_t h)
351 {
352 	uintptr_t *loc;
353 
354 	MPASS(type < nitems(t4_fw_msg_handler));
355 	MPASS(h == NULL || t4_fw_msg_handler[type] == NULL);
356 	/*
357 	 * These are dispatched by the handler for FW{4|6}_CPL_MSG using the CPL
358 	 * handler dispatch table.  Reject any attempt to install a handler for
359 	 * this subtype.
360 	 */
361 	MPASS(type != FW_TYPE_RSSCPL);
362 	MPASS(type != FW6_TYPE_RSSCPL);
363 
364 	loc = (uintptr_t *)&t4_fw_msg_handler[type];
365 	atomic_store_rel_ptr(loc, (uintptr_t)h);
366 }
367 
368 void
369 t4_register_cpl_handler(int opcode, cpl_handler_t h)
370 {
371 	uintptr_t *loc;
372 
373 	MPASS(opcode < nitems(t4_cpl_handler));
374 	MPASS(h == NULL || t4_cpl_handler[opcode] == NULL);
375 
376 	loc = (uintptr_t *)&t4_cpl_handler[opcode];
377 	atomic_store_rel_ptr(loc, (uintptr_t)h);
378 }
379 
380 static int
381 set_tcb_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
382     struct mbuf *m)
383 {
384 	const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
385 	u_int tid;
386 	int cookie;
387 
388 	MPASS(m == NULL);
389 
390 	tid = GET_TID(cpl);
391 	if (is_hpftid(iq->adapter, tid) || is_ftid(iq->adapter, tid)) {
392 		/*
393 		 * The return code for filter-write is put in the CPL cookie so
394 		 * we have to rely on the hardware tid (is_ftid) to determine
395 		 * that this is a response to a filter.
396 		 */
397 		cookie = CPL_COOKIE_FILTER;
398 	} else {
399 		cookie = G_COOKIE(cpl->cookie);
400 	}
401 	MPASS(cookie > CPL_COOKIE_RESERVED);
402 	MPASS(cookie < nitems(set_tcb_rpl_handlers));
403 
404 	return (set_tcb_rpl_handlers[cookie](iq, rss, m));
405 }
406 
407 static int
408 l2t_write_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
409     struct mbuf *m)
410 {
411 	const struct cpl_l2t_write_rpl *rpl = (const void *)(rss + 1);
412 	unsigned int cookie;
413 
414 	MPASS(m == NULL);
415 
416 	cookie = GET_TID(rpl) & F_SYNC_WR ? CPL_COOKIE_TOM : CPL_COOKIE_FILTER;
417 	return (l2t_write_rpl_handlers[cookie](iq, rss, m));
418 }
419 
420 static int
421 act_open_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
422     struct mbuf *m)
423 {
424 	const struct cpl_act_open_rpl *cpl = (const void *)(rss + 1);
425 	u_int cookie = G_TID_COOKIE(G_AOPEN_ATID(be32toh(cpl->atid_status)));
426 
427 	MPASS(m == NULL);
428 	MPASS(cookie != CPL_COOKIE_RESERVED);
429 
430 	return (act_open_rpl_handlers[cookie](iq, rss, m));
431 }
432 
433 static int
434 abort_rpl_rss_handler(struct sge_iq *iq, const struct rss_header *rss,
435     struct mbuf *m)
436 {
437 	struct adapter *sc = iq->adapter;
438 	u_int cookie;
439 
440 	MPASS(m == NULL);
441 	if (is_hashfilter(sc))
442 		cookie = CPL_COOKIE_HASHFILTER;
443 	else
444 		cookie = CPL_COOKIE_TOM;
445 
446 	return (abort_rpl_rss_handlers[cookie](iq, rss, m));
447 }
448 
449 static int
450 fw4_ack_handler(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
451 {
452 	struct adapter *sc = iq->adapter;
453 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
454 	unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
455 	u_int cookie;
456 
457 	MPASS(m == NULL);
458 	if (is_etid(sc, tid))
459 		cookie = CPL_COOKIE_ETHOFLD;
460 	else
461 		cookie = CPL_COOKIE_TOM;
462 
463 	return (fw4_ack_handlers[cookie](iq, rss, m));
464 }
465 
466 static void
467 t4_init_shared_cpl_handlers(void)
468 {
469 
470 	t4_register_cpl_handler(CPL_SET_TCB_RPL, set_tcb_rpl_handler);
471 	t4_register_cpl_handler(CPL_L2T_WRITE_RPL, l2t_write_rpl_handler);
472 	t4_register_cpl_handler(CPL_ACT_OPEN_RPL, act_open_rpl_handler);
473 	t4_register_cpl_handler(CPL_ABORT_RPL_RSS, abort_rpl_rss_handler);
474 	t4_register_cpl_handler(CPL_FW4_ACK, fw4_ack_handler);
475 }
476 
477 void
478 t4_register_shared_cpl_handler(int opcode, cpl_handler_t h, int cookie)
479 {
480 	uintptr_t *loc;
481 
482 	MPASS(opcode < nitems(t4_cpl_handler));
483 	MPASS(cookie > CPL_COOKIE_RESERVED);
484 	MPASS(cookie < NUM_CPL_COOKIES);
485 	MPASS(t4_cpl_handler[opcode] != NULL);
486 
487 	switch (opcode) {
488 	case CPL_SET_TCB_RPL:
489 		loc = (uintptr_t *)&set_tcb_rpl_handlers[cookie];
490 		break;
491 	case CPL_L2T_WRITE_RPL:
492 		loc = (uintptr_t *)&l2t_write_rpl_handlers[cookie];
493 		break;
494 	case CPL_ACT_OPEN_RPL:
495 		loc = (uintptr_t *)&act_open_rpl_handlers[cookie];
496 		break;
497 	case CPL_ABORT_RPL_RSS:
498 		loc = (uintptr_t *)&abort_rpl_rss_handlers[cookie];
499 		break;
500 	case CPL_FW4_ACK:
501 		loc = (uintptr_t *)&fw4_ack_handlers[cookie];
502 		break;
503 	default:
504 		MPASS(0);
505 		return;
506 	}
507 	MPASS(h == NULL || *loc == (uintptr_t)NULL);
508 	atomic_store_rel_ptr(loc, (uintptr_t)h);
509 }
510 
511 /*
512  * Called on MOD_LOAD.  Validates and calculates the SGE tunables.
513  */
514 void
515 t4_sge_modload(void)
516 {
517 
518 	if (fl_pktshift < 0 || fl_pktshift > 7) {
519 		printf("Invalid hw.cxgbe.fl_pktshift value (%d),"
520 		    " using 0 instead.\n", fl_pktshift);
521 		fl_pktshift = 0;
522 	}
523 
524 	if (spg_len != 64 && spg_len != 128) {
525 		int len;
526 
527 #if defined(__i386__) || defined(__amd64__)
528 		len = cpu_clflush_line_size > 64 ? 128 : 64;
529 #else
530 		len = 64;
531 #endif
532 		if (spg_len != -1) {
533 			printf("Invalid hw.cxgbe.spg_len value (%d),"
534 			    " using %d instead.\n", spg_len, len);
535 		}
536 		spg_len = len;
537 	}
538 
539 	if (cong_drop < -1 || cong_drop > 1) {
540 		printf("Invalid hw.cxgbe.cong_drop value (%d),"
541 		    " using 0 instead.\n", cong_drop);
542 		cong_drop = 0;
543 	}
544 
545 	if (tscale != 1 && (tscale < 3 || tscale > 17)) {
546 		printf("Invalid hw.cxgbe.tscale value (%d),"
547 		    " using 1 instead.\n", tscale);
548 		tscale = 1;
549 	}
550 
551 	extfree_refs = counter_u64_alloc(M_WAITOK);
552 	extfree_rels = counter_u64_alloc(M_WAITOK);
553 	counter_u64_zero(extfree_refs);
554 	counter_u64_zero(extfree_rels);
555 
556 	t4_init_shared_cpl_handlers();
557 	t4_register_cpl_handler(CPL_FW4_MSG, handle_fw_msg);
558 	t4_register_cpl_handler(CPL_FW6_MSG, handle_fw_msg);
559 	t4_register_cpl_handler(CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
560 	t4_register_cpl_handler(CPL_RX_PKT, t4_eth_rx);
561 #ifdef RATELIMIT
562 	t4_register_shared_cpl_handler(CPL_FW4_ACK, ethofld_fw4_ack,
563 	    CPL_COOKIE_ETHOFLD);
564 #endif
565 	t4_register_fw_msg_handler(FW6_TYPE_CMD_RPL, t4_handle_fw_rpl);
566 	t4_register_fw_msg_handler(FW6_TYPE_WRERR_RPL, t4_handle_wrerr_rpl);
567 }
568 
569 void
570 t4_sge_modunload(void)
571 {
572 
573 	counter_u64_free(extfree_refs);
574 	counter_u64_free(extfree_rels);
575 }
576 
577 uint64_t
578 t4_sge_extfree_refs(void)
579 {
580 	uint64_t refs, rels;
581 
582 	rels = counter_u64_fetch(extfree_rels);
583 	refs = counter_u64_fetch(extfree_refs);
584 
585 	return (refs - rels);
586 }
587 
588 static inline void
589 setup_pad_and_pack_boundaries(struct adapter *sc)
590 {
591 	uint32_t v, m;
592 	int pad, pack, pad_shift;
593 
594 	pad_shift = chip_id(sc) > CHELSIO_T5 ? X_T6_INGPADBOUNDARY_SHIFT :
595 	    X_INGPADBOUNDARY_SHIFT;
596 	pad = fl_pad;
597 	if (fl_pad < (1 << pad_shift) ||
598 	    fl_pad > (1 << (pad_shift + M_INGPADBOUNDARY)) ||
599 	    !powerof2(fl_pad)) {
600 		/*
601 		 * If there is any chance that we might use buffer packing and
602 		 * the chip is a T4, then pick 64 as the pad/pack boundary.  Set
603 		 * it to the minimum allowed in all other cases.
604 		 */
605 		pad = is_t4(sc) && buffer_packing ? 64 : 1 << pad_shift;
606 
607 		/*
608 		 * For fl_pad = 0 we'll still write a reasonable value to the
609 		 * register but all the freelists will opt out of padding.
610 		 * We'll complain here only if the user tried to set it to a
611 		 * value greater than 0 that was invalid.
612 		 */
613 		if (fl_pad > 0) {
614 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pad value"
615 			    " (%d), using %d instead.\n", fl_pad, pad);
616 		}
617 	}
618 	m = V_INGPADBOUNDARY(M_INGPADBOUNDARY);
619 	v = V_INGPADBOUNDARY(ilog2(pad) - pad_shift);
620 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
621 
622 	if (is_t4(sc)) {
623 		if (fl_pack != -1 && fl_pack != pad) {
624 			/* Complain but carry on. */
625 			device_printf(sc->dev, "hw.cxgbe.fl_pack (%d) ignored,"
626 			    " using %d instead.\n", fl_pack, pad);
627 		}
628 		return;
629 	}
630 
631 	pack = fl_pack;
632 	if (fl_pack < 16 || fl_pack == 32 || fl_pack > 4096 ||
633 	    !powerof2(fl_pack)) {
634 		pack = max(sc->params.pci.mps, CACHE_LINE_SIZE);
635 		MPASS(powerof2(pack));
636 		if (pack < 16)
637 			pack = 16;
638 		if (pack == 32)
639 			pack = 64;
640 		if (pack > 4096)
641 			pack = 4096;
642 		if (fl_pack != -1) {
643 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pack value"
644 			    " (%d), using %d instead.\n", fl_pack, pack);
645 		}
646 	}
647 	m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY);
648 	if (pack == 16)
649 		v = V_INGPACKBOUNDARY(0);
650 	else
651 		v = V_INGPACKBOUNDARY(ilog2(pack) - 5);
652 
653 	MPASS(!is_t4(sc));	/* T4 doesn't have SGE_CONTROL2 */
654 	t4_set_reg_field(sc, A_SGE_CONTROL2, m, v);
655 }
656 
657 /*
658  * adap->params.vpd.cclk must be set up before this is called.
659  */
660 void
661 t4_tweak_chip_settings(struct adapter *sc)
662 {
663 	int i;
664 	uint32_t v, m;
665 	int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200};
666 	int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk;
667 	int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */
668 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
669 	static int sge_flbuf_sizes[] = {
670 		MCLBYTES,
671 #if MJUMPAGESIZE != MCLBYTES
672 		MJUMPAGESIZE,
673 		MJUMPAGESIZE - CL_METADATA_SIZE,
674 		MJUMPAGESIZE - 2 * MSIZE - CL_METADATA_SIZE,
675 #endif
676 		MJUM9BYTES,
677 		MJUM16BYTES,
678 		MCLBYTES - MSIZE - CL_METADATA_SIZE,
679 		MJUM9BYTES - CL_METADATA_SIZE,
680 		MJUM16BYTES - CL_METADATA_SIZE,
681 	};
682 
683 	KASSERT(sc->flags & MASTER_PF,
684 	    ("%s: trying to change chip settings when not master.", __func__));
685 
686 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
687 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
688 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
689 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
690 
691 	setup_pad_and_pack_boundaries(sc);
692 
693 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
694 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
695 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
696 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
697 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
698 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
699 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
700 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
701 	t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v);
702 
703 	KASSERT(nitems(sge_flbuf_sizes) <= SGE_FLBUF_SIZES,
704 	    ("%s: hw buffer size table too big", __func__));
705 	t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0, 4096);
706 	t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE1, 65536);
707 	for (i = 0; i < min(nitems(sge_flbuf_sizes), SGE_FLBUF_SIZES); i++) {
708 		t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE15 - (4 * i),
709 		    sge_flbuf_sizes[i]);
710 	}
711 
712 	v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) |
713 	    V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]);
714 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v);
715 
716 	KASSERT(intr_timer[0] <= timer_max,
717 	    ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0],
718 	    timer_max));
719 	for (i = 1; i < nitems(intr_timer); i++) {
720 		KASSERT(intr_timer[i] >= intr_timer[i - 1],
721 		    ("%s: timers not listed in increasing order (%d)",
722 		    __func__, i));
723 
724 		while (intr_timer[i] > timer_max) {
725 			if (i == nitems(intr_timer) - 1) {
726 				intr_timer[i] = timer_max;
727 				break;
728 			}
729 			intr_timer[i] += intr_timer[i - 1];
730 			intr_timer[i] /= 2;
731 		}
732 	}
733 
734 	v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) |
735 	    V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1]));
736 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v);
737 	v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) |
738 	    V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3]));
739 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v);
740 	v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) |
741 	    V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5]));
742 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v);
743 
744 	if (chip_id(sc) >= CHELSIO_T6) {
745 		m = V_TSCALE(M_TSCALE);
746 		if (tscale == 1)
747 			v = 0;
748 		else
749 			v = V_TSCALE(tscale - 2);
750 		t4_set_reg_field(sc, A_SGE_ITP_CONTROL, m, v);
751 
752 		if (sc->debug_flags & DF_DISABLE_TCB_CACHE) {
753 			m = V_RDTHRESHOLD(M_RDTHRESHOLD) | F_WRTHRTHRESHEN |
754 			    V_WRTHRTHRESH(M_WRTHRTHRESH);
755 			t4_tp_pio_read(sc, &v, 1, A_TP_CMM_CONFIG, 1);
756 			v &= ~m;
757 			v |= V_RDTHRESHOLD(1) | F_WRTHRTHRESHEN |
758 			    V_WRTHRTHRESH(16);
759 			t4_tp_pio_write(sc, &v, 1, A_TP_CMM_CONFIG, 1);
760 		}
761 	}
762 
763 	/* 4K, 16K, 64K, 256K DDP "page sizes" for TDDP */
764 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
765 	t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v);
766 
767 	/*
768 	 * 4K, 8K, 16K, 64K DDP "page sizes" for iSCSI DDP.  These have been
769 	 * chosen with MAXPHYS = 128K in mind.  The largest DDP buffer that we
770 	 * may have to deal with is MAXPHYS + 1 page.
771 	 */
772 	v = V_HPZ0(0) | V_HPZ1(1) | V_HPZ2(2) | V_HPZ3(4);
773 	t4_write_reg(sc, A_ULP_RX_ISCSI_PSZ, v);
774 
775 	/* We use multiple DDP page sizes both in plain-TOE and ISCSI modes. */
776 	m = v = F_TDDPTAGTCB | F_ISCSITAGTCB;
777 	t4_set_reg_field(sc, A_ULP_RX_CTL, m, v);
778 
779 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
780 	    F_RESETDDPOFFSET;
781 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
782 	t4_set_reg_field(sc, A_TP_PARA_REG5, m, v);
783 }
784 
785 /*
786  * SGE wants the buffer to be at least 64B and then a multiple of 16.  If
787  * padding is in use, the buffer's start and end need to be aligned to the pad
788  * boundary as well.  We'll just make sure that the size is a multiple of the
789  * boundary here, it is up to the buffer allocation code to make sure the start
790  * of the buffer is aligned as well.
791  */
792 static inline int
793 hwsz_ok(struct adapter *sc, int hwsz)
794 {
795 	int mask = fl_pad ? sc->params.sge.pad_boundary - 1 : 16 - 1;
796 
797 	return (hwsz >= 64 && (hwsz & mask) == 0);
798 }
799 
800 /*
801  * XXX: driver really should be able to deal with unexpected settings.
802  */
803 int
804 t4_read_chip_settings(struct adapter *sc)
805 {
806 	struct sge *s = &sc->sge;
807 	struct sge_params *sp = &sc->params.sge;
808 	int i, j, n, rc = 0;
809 	uint32_t m, v, r;
810 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
811 	static int sw_buf_sizes[] = {	/* Sorted by size */
812 		MCLBYTES,
813 #if MJUMPAGESIZE != MCLBYTES
814 		MJUMPAGESIZE,
815 #endif
816 		MJUM9BYTES,
817 		MJUM16BYTES
818 	};
819 	struct sw_zone_info *swz, *safe_swz;
820 	struct hw_buf_info *hwb;
821 
822 	m = F_RXPKTCPLMODE;
823 	v = F_RXPKTCPLMODE;
824 	r = sc->params.sge.sge_control;
825 	if ((r & m) != v) {
826 		device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r);
827 		rc = EINVAL;
828 	}
829 
830 	/*
831 	 * If this changes then every single use of PAGE_SHIFT in the driver
832 	 * needs to be carefully reviewed for PAGE_SHIFT vs sp->page_shift.
833 	 */
834 	if (sp->page_shift != PAGE_SHIFT) {
835 		device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r);
836 		rc = EINVAL;
837 	}
838 
839 	/* Filter out unusable hw buffer sizes entirely (mark with -2). */
840 	hwb = &s->hw_buf_info[0];
841 	for (i = 0; i < nitems(s->hw_buf_info); i++, hwb++) {
842 		r = sc->params.sge.sge_fl_buffer_size[i];
843 		hwb->size = r;
844 		hwb->zidx = hwsz_ok(sc, r) ? -1 : -2;
845 		hwb->next = -1;
846 	}
847 
848 	/*
849 	 * Create a sorted list in decreasing order of hw buffer sizes (and so
850 	 * increasing order of spare area) for each software zone.
851 	 *
852 	 * If padding is enabled then the start and end of the buffer must align
853 	 * to the pad boundary; if packing is enabled then they must align with
854 	 * the pack boundary as well.  Allocations from the cluster zones are
855 	 * aligned to min(size, 4K), so the buffer starts at that alignment and
856 	 * ends at hwb->size alignment.  If mbuf inlining is allowed the
857 	 * starting alignment will be reduced to MSIZE and the driver will
858 	 * exercise appropriate caution when deciding on the best buffer layout
859 	 * to use.
860 	 */
861 	n = 0;	/* no usable buffer size to begin with */
862 	swz = &s->sw_zone_info[0];
863 	safe_swz = NULL;
864 	for (i = 0; i < SW_ZONE_SIZES; i++, swz++) {
865 		int8_t head = -1, tail = -1;
866 
867 		swz->size = sw_buf_sizes[i];
868 		swz->zone = m_getzone(swz->size);
869 		swz->type = m_gettype(swz->size);
870 
871 		if (swz->size < PAGE_SIZE) {
872 			MPASS(powerof2(swz->size));
873 			if (fl_pad && (swz->size % sp->pad_boundary != 0))
874 				continue;
875 		}
876 
877 		if (swz->size == safest_rx_cluster)
878 			safe_swz = swz;
879 
880 		hwb = &s->hw_buf_info[0];
881 		for (j = 0; j < SGE_FLBUF_SIZES; j++, hwb++) {
882 			if (hwb->zidx != -1 || hwb->size > swz->size)
883 				continue;
884 #ifdef INVARIANTS
885 			if (fl_pad)
886 				MPASS(hwb->size % sp->pad_boundary == 0);
887 #endif
888 			hwb->zidx = i;
889 			if (head == -1)
890 				head = tail = j;
891 			else if (hwb->size < s->hw_buf_info[tail].size) {
892 				s->hw_buf_info[tail].next = j;
893 				tail = j;
894 			} else {
895 				int8_t *cur;
896 				struct hw_buf_info *t;
897 
898 				for (cur = &head; *cur != -1; cur = &t->next) {
899 					t = &s->hw_buf_info[*cur];
900 					if (hwb->size == t->size) {
901 						hwb->zidx = -2;
902 						break;
903 					}
904 					if (hwb->size > t->size) {
905 						hwb->next = *cur;
906 						*cur = j;
907 						break;
908 					}
909 				}
910 			}
911 		}
912 		swz->head_hwidx = head;
913 		swz->tail_hwidx = tail;
914 
915 		if (tail != -1) {
916 			n++;
917 			if (swz->size - s->hw_buf_info[tail].size >=
918 			    CL_METADATA_SIZE)
919 				sc->flags |= BUF_PACKING_OK;
920 		}
921 	}
922 	if (n == 0) {
923 		device_printf(sc->dev, "no usable SGE FL buffer size.\n");
924 		rc = EINVAL;
925 	}
926 
927 	s->safe_hwidx1 = -1;
928 	s->safe_hwidx2 = -1;
929 	if (safe_swz != NULL) {
930 		s->safe_hwidx1 = safe_swz->head_hwidx;
931 		for (i = safe_swz->head_hwidx; i != -1; i = hwb->next) {
932 			int spare;
933 
934 			hwb = &s->hw_buf_info[i];
935 #ifdef INVARIANTS
936 			if (fl_pad)
937 				MPASS(hwb->size % sp->pad_boundary == 0);
938 #endif
939 			spare = safe_swz->size - hwb->size;
940 			if (spare >= CL_METADATA_SIZE) {
941 				s->safe_hwidx2 = i;
942 				break;
943 			}
944 		}
945 	}
946 
947 	if (sc->flags & IS_VF)
948 		return (0);
949 
950 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
951 	r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ);
952 	if (r != v) {
953 		device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r);
954 		rc = EINVAL;
955 	}
956 
957 	m = v = F_TDDPTAGTCB;
958 	r = t4_read_reg(sc, A_ULP_RX_CTL);
959 	if ((r & m) != v) {
960 		device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r);
961 		rc = EINVAL;
962 	}
963 
964 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
965 	    F_RESETDDPOFFSET;
966 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
967 	r = t4_read_reg(sc, A_TP_PARA_REG5);
968 	if ((r & m) != v) {
969 		device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r);
970 		rc = EINVAL;
971 	}
972 
973 	t4_init_tp_params(sc, 1);
974 
975 	t4_read_mtu_tbl(sc, sc->params.mtus, NULL);
976 	t4_load_mtus(sc, sc->params.mtus, sc->params.a_wnd, sc->params.b_wnd);
977 
978 	return (rc);
979 }
980 
981 int
982 t4_create_dma_tag(struct adapter *sc)
983 {
984 	int rc;
985 
986 	rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
987 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
988 	    BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
989 	    NULL, &sc->dmat);
990 	if (rc != 0) {
991 		device_printf(sc->dev,
992 		    "failed to create main DMA tag: %d\n", rc);
993 	}
994 
995 	return (rc);
996 }
997 
998 void
999 t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
1000     struct sysctl_oid_list *children)
1001 {
1002 	struct sge_params *sp = &sc->params.sge;
1003 
1004 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "buffer_sizes",
1005 	    CTLTYPE_STRING | CTLFLAG_RD, &sc->sge, 0, sysctl_bufsizes, "A",
1006 	    "freelist buffer sizes");
1007 
1008 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD,
1009 	    NULL, sp->fl_pktshift, "payload DMA offset in rx buffer (bytes)");
1010 
1011 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD,
1012 	    NULL, sp->pad_boundary, "payload pad boundary (bytes)");
1013 
1014 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD,
1015 	    NULL, sp->spg_len, "status page size (bytes)");
1016 
1017 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD,
1018 	    NULL, cong_drop, "congestion drop setting");
1019 
1020 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD,
1021 	    NULL, sp->pack_boundary, "payload pack boundary (bytes)");
1022 }
1023 
1024 int
1025 t4_destroy_dma_tag(struct adapter *sc)
1026 {
1027 	if (sc->dmat)
1028 		bus_dma_tag_destroy(sc->dmat);
1029 
1030 	return (0);
1031 }
1032 
1033 /*
1034  * Allocate and initialize the firmware event queue, control queues, and special
1035  * purpose rx queues owned by the adapter.
1036  *
1037  * Returns errno on failure.  Resources allocated up to that point may still be
1038  * allocated.  Caller is responsible for cleanup in case this function fails.
1039  */
1040 int
1041 t4_setup_adapter_queues(struct adapter *sc)
1042 {
1043 	struct sysctl_oid *oid;
1044 	struct sysctl_oid_list *children;
1045 	int rc, i;
1046 
1047 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
1048 
1049 	sysctl_ctx_init(&sc->ctx);
1050 	sc->flags |= ADAP_SYSCTL_CTX;
1051 
1052 	/*
1053 	 * Firmware event queue
1054 	 */
1055 	rc = alloc_fwq(sc);
1056 	if (rc != 0)
1057 		return (rc);
1058 
1059 	/*
1060 	 * That's all for the VF driver.
1061 	 */
1062 	if (sc->flags & IS_VF)
1063 		return (rc);
1064 
1065 	oid = device_get_sysctl_tree(sc->dev);
1066 	children = SYSCTL_CHILDREN(oid);
1067 
1068 	/*
1069 	 * XXX: General purpose rx queues, one per port.
1070 	 */
1071 
1072 	/*
1073 	 * Control queues, one per port.
1074 	 */
1075 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "ctrlq",
1076 	    CTLFLAG_RD, NULL, "control queues");
1077 	for_each_port(sc, i) {
1078 		struct sge_wrq *ctrlq = &sc->sge.ctrlq[i];
1079 
1080 		rc = alloc_ctrlq(sc, ctrlq, i, oid);
1081 		if (rc != 0)
1082 			return (rc);
1083 	}
1084 
1085 	return (rc);
1086 }
1087 
1088 /*
1089  * Idempotent
1090  */
1091 int
1092 t4_teardown_adapter_queues(struct adapter *sc)
1093 {
1094 	int i;
1095 
1096 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
1097 
1098 	/* Do this before freeing the queue */
1099 	if (sc->flags & ADAP_SYSCTL_CTX) {
1100 		sysctl_ctx_free(&sc->ctx);
1101 		sc->flags &= ~ADAP_SYSCTL_CTX;
1102 	}
1103 
1104 	if (!(sc->flags & IS_VF)) {
1105 		for_each_port(sc, i)
1106 			free_wrq(sc, &sc->sge.ctrlq[i]);
1107 	}
1108 	free_fwq(sc);
1109 
1110 	return (0);
1111 }
1112 
1113 /* Maximum payload that can be delivered with a single iq descriptor */
1114 static inline int
1115 mtu_to_max_payload(struct adapter *sc, int mtu, const int toe)
1116 {
1117 	int payload;
1118 
1119 #ifdef TCP_OFFLOAD
1120 	if (toe) {
1121 		int rxcs = G_RXCOALESCESIZE(t4_read_reg(sc, A_TP_PARA_REG2));
1122 
1123 		/* Note that COP can set rx_coalesce on/off per connection. */
1124 		payload = max(mtu, rxcs);
1125 	} else {
1126 #endif
1127 		/* large enough even when hw VLAN extraction is disabled */
1128 		payload = sc->params.sge.fl_pktshift + ETHER_HDR_LEN +
1129 		    ETHER_VLAN_ENCAP_LEN + mtu;
1130 #ifdef TCP_OFFLOAD
1131 	}
1132 #endif
1133 
1134 	return (payload);
1135 }
1136 
1137 int
1138 t4_setup_vi_queues(struct vi_info *vi)
1139 {
1140 	int rc = 0, i, intr_idx, iqidx;
1141 	struct sge_rxq *rxq;
1142 	struct sge_txq *txq;
1143 #ifdef TCP_OFFLOAD
1144 	struct sge_ofld_rxq *ofld_rxq;
1145 #endif
1146 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1147 	struct sge_wrq *ofld_txq;
1148 #endif
1149 #ifdef DEV_NETMAP
1150 	int saved_idx;
1151 	struct sge_nm_rxq *nm_rxq;
1152 	struct sge_nm_txq *nm_txq;
1153 #endif
1154 	char name[16];
1155 	struct port_info *pi = vi->pi;
1156 	struct adapter *sc = pi->adapter;
1157 	struct ifnet *ifp = vi->ifp;
1158 	struct sysctl_oid *oid = device_get_sysctl_tree(vi->dev);
1159 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
1160 	int maxp, mtu = ifp->if_mtu;
1161 
1162 	/* Interrupt vector to start from (when using multiple vectors) */
1163 	intr_idx = vi->first_intr;
1164 
1165 #ifdef DEV_NETMAP
1166 	saved_idx = intr_idx;
1167 	if (ifp->if_capabilities & IFCAP_NETMAP) {
1168 
1169 		/* netmap is supported with direct interrupts only. */
1170 		MPASS(!forwarding_intr_to_fwq(sc));
1171 
1172 		/*
1173 		 * We don't have buffers to back the netmap rx queues
1174 		 * right now so we create the queues in a way that
1175 		 * doesn't set off any congestion signal in the chip.
1176 		 */
1177 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_rxq",
1178 		    CTLFLAG_RD, NULL, "rx queues");
1179 		for_each_nm_rxq(vi, i, nm_rxq) {
1180 			rc = alloc_nm_rxq(vi, nm_rxq, intr_idx, i, oid);
1181 			if (rc != 0)
1182 				goto done;
1183 			intr_idx++;
1184 		}
1185 
1186 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_txq",
1187 		    CTLFLAG_RD, NULL, "tx queues");
1188 		for_each_nm_txq(vi, i, nm_txq) {
1189 			iqidx = vi->first_nm_rxq + (i % vi->nnmrxq);
1190 			rc = alloc_nm_txq(vi, nm_txq, iqidx, i, oid);
1191 			if (rc != 0)
1192 				goto done;
1193 		}
1194 	}
1195 
1196 	/* Normal rx queues and netmap rx queues share the same interrupts. */
1197 	intr_idx = saved_idx;
1198 #endif
1199 
1200 	/*
1201 	 * Allocate rx queues first because a default iqid is required when
1202 	 * creating a tx queue.
1203 	 */
1204 	maxp = mtu_to_max_payload(sc, mtu, 0);
1205 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
1206 	    CTLFLAG_RD, NULL, "rx queues");
1207 	for_each_rxq(vi, i, rxq) {
1208 
1209 		init_iq(&rxq->iq, sc, vi->tmr_idx, vi->pktc_idx, vi->qsize_rxq);
1210 
1211 		snprintf(name, sizeof(name), "%s rxq%d-fl",
1212 		    device_get_nameunit(vi->dev), i);
1213 		init_fl(sc, &rxq->fl, vi->qsize_rxq / 8, maxp, name);
1214 
1215 		rc = alloc_rxq(vi, rxq,
1216 		    forwarding_intr_to_fwq(sc) ? -1 : intr_idx, i, oid);
1217 		if (rc != 0)
1218 			goto done;
1219 		intr_idx++;
1220 	}
1221 #ifdef DEV_NETMAP
1222 	if (ifp->if_capabilities & IFCAP_NETMAP)
1223 		intr_idx = saved_idx + max(vi->nrxq, vi->nnmrxq);
1224 #endif
1225 #ifdef TCP_OFFLOAD
1226 	maxp = mtu_to_max_payload(sc, mtu, 1);
1227 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq",
1228 	    CTLFLAG_RD, NULL, "rx queues for offloaded TCP connections");
1229 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1230 
1231 		init_iq(&ofld_rxq->iq, sc, vi->ofld_tmr_idx, vi->ofld_pktc_idx,
1232 		    vi->qsize_rxq);
1233 
1234 		snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
1235 		    device_get_nameunit(vi->dev), i);
1236 		init_fl(sc, &ofld_rxq->fl, vi->qsize_rxq / 8, maxp, name);
1237 
1238 		rc = alloc_ofld_rxq(vi, ofld_rxq,
1239 		    forwarding_intr_to_fwq(sc) ? -1 : intr_idx, i, oid);
1240 		if (rc != 0)
1241 			goto done;
1242 		intr_idx++;
1243 	}
1244 #endif
1245 
1246 	/*
1247 	 * Now the tx queues.
1248 	 */
1249 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "txq", CTLFLAG_RD,
1250 	    NULL, "tx queues");
1251 	for_each_txq(vi, i, txq) {
1252 		iqidx = vi->first_rxq + (i % vi->nrxq);
1253 		snprintf(name, sizeof(name), "%s txq%d",
1254 		    device_get_nameunit(vi->dev), i);
1255 		init_eq(sc, &txq->eq, EQ_ETH, vi->qsize_txq, pi->tx_chan,
1256 		    sc->sge.rxq[iqidx].iq.cntxt_id, name);
1257 
1258 		rc = alloc_txq(vi, txq, i, oid);
1259 		if (rc != 0)
1260 			goto done;
1261 	}
1262 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1263 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_txq",
1264 	    CTLFLAG_RD, NULL, "tx queues for TOE/ETHOFLD");
1265 	for_each_ofld_txq(vi, i, ofld_txq) {
1266 		struct sysctl_oid *oid2;
1267 
1268 		snprintf(name, sizeof(name), "%s ofld_txq%d",
1269 		    device_get_nameunit(vi->dev), i);
1270 		if (vi->nofldrxq > 0) {
1271 			iqidx = vi->first_ofld_rxq + (i % vi->nofldrxq);
1272 			init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq,
1273 			    pi->tx_chan, sc->sge.ofld_rxq[iqidx].iq.cntxt_id,
1274 			    name);
1275 		} else {
1276 			iqidx = vi->first_rxq + (i % vi->nrxq);
1277 			init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq,
1278 			    pi->tx_chan, sc->sge.rxq[iqidx].iq.cntxt_id, name);
1279 		}
1280 
1281 		snprintf(name, sizeof(name), "%d", i);
1282 		oid2 = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
1283 		    name, CTLFLAG_RD, NULL, "offload tx queue");
1284 
1285 		rc = alloc_wrq(sc, vi, ofld_txq, oid2);
1286 		if (rc != 0)
1287 			goto done;
1288 	}
1289 #endif
1290 done:
1291 	if (rc)
1292 		t4_teardown_vi_queues(vi);
1293 
1294 	return (rc);
1295 }
1296 
1297 /*
1298  * Idempotent
1299  */
1300 int
1301 t4_teardown_vi_queues(struct vi_info *vi)
1302 {
1303 	int i;
1304 	struct sge_rxq *rxq;
1305 	struct sge_txq *txq;
1306 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1307 	struct port_info *pi = vi->pi;
1308 	struct adapter *sc = pi->adapter;
1309 	struct sge_wrq *ofld_txq;
1310 #endif
1311 #ifdef TCP_OFFLOAD
1312 	struct sge_ofld_rxq *ofld_rxq;
1313 #endif
1314 #ifdef DEV_NETMAP
1315 	struct sge_nm_rxq *nm_rxq;
1316 	struct sge_nm_txq *nm_txq;
1317 #endif
1318 
1319 	/* Do this before freeing the queues */
1320 	if (vi->flags & VI_SYSCTL_CTX) {
1321 		sysctl_ctx_free(&vi->ctx);
1322 		vi->flags &= ~VI_SYSCTL_CTX;
1323 	}
1324 
1325 #ifdef DEV_NETMAP
1326 	if (vi->ifp->if_capabilities & IFCAP_NETMAP) {
1327 		for_each_nm_txq(vi, i, nm_txq) {
1328 			free_nm_txq(vi, nm_txq);
1329 		}
1330 
1331 		for_each_nm_rxq(vi, i, nm_rxq) {
1332 			free_nm_rxq(vi, nm_rxq);
1333 		}
1334 	}
1335 #endif
1336 
1337 	/*
1338 	 * Take down all the tx queues first, as they reference the rx queues
1339 	 * (for egress updates, etc.).
1340 	 */
1341 
1342 	for_each_txq(vi, i, txq) {
1343 		free_txq(vi, txq);
1344 	}
1345 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1346 	for_each_ofld_txq(vi, i, ofld_txq) {
1347 		free_wrq(sc, ofld_txq);
1348 	}
1349 #endif
1350 
1351 	/*
1352 	 * Then take down the rx queues.
1353 	 */
1354 
1355 	for_each_rxq(vi, i, rxq) {
1356 		free_rxq(vi, rxq);
1357 	}
1358 #ifdef TCP_OFFLOAD
1359 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1360 		free_ofld_rxq(vi, ofld_rxq);
1361 	}
1362 #endif
1363 
1364 	return (0);
1365 }
1366 
1367 /*
1368  * Interrupt handler when the driver is using only 1 interrupt.  This is a very
1369  * unusual scenario.
1370  *
1371  * a) Deals with errors, if any.
1372  * b) Services firmware event queue, which is taking interrupts for all other
1373  *    queues.
1374  */
1375 void
1376 t4_intr_all(void *arg)
1377 {
1378 	struct adapter *sc = arg;
1379 	struct sge_iq *fwq = &sc->sge.fwq;
1380 
1381 	MPASS(sc->intr_count == 1);
1382 
1383 	if (sc->intr_type == INTR_INTX)
1384 		t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
1385 
1386 	t4_intr_err(arg);
1387 	t4_intr_evt(fwq);
1388 }
1389 
1390 /*
1391  * Interrupt handler for errors (installed directly when multiple interrupts are
1392  * being used, or called by t4_intr_all).
1393  */
1394 void
1395 t4_intr_err(void *arg)
1396 {
1397 	struct adapter *sc = arg;
1398 	uint32_t v;
1399 	const bool verbose = (sc->debug_flags & DF_VERBOSE_SLOWINTR) != 0;
1400 
1401 	if (sc->flags & ADAP_ERR)
1402 		return;
1403 
1404 	v = t4_read_reg(sc, MYPF_REG(A_PL_PF_INT_CAUSE));
1405 	if (v & F_PFSW) {
1406 		sc->swintr++;
1407 		t4_write_reg(sc, MYPF_REG(A_PL_PF_INT_CAUSE), v);
1408 	}
1409 
1410 	t4_slow_intr_handler(sc, verbose);
1411 }
1412 
1413 /*
1414  * Interrupt handler for iq-only queues.  The firmware event queue is the only
1415  * such queue right now.
1416  */
1417 void
1418 t4_intr_evt(void *arg)
1419 {
1420 	struct sge_iq *iq = arg;
1421 
1422 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1423 		service_iq(iq, 0);
1424 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1425 	}
1426 }
1427 
1428 /*
1429  * Interrupt handler for iq+fl queues.
1430  */
1431 void
1432 t4_intr(void *arg)
1433 {
1434 	struct sge_iq *iq = arg;
1435 
1436 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1437 		service_iq_fl(iq, 0);
1438 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1439 	}
1440 }
1441 
1442 #ifdef DEV_NETMAP
1443 /*
1444  * Interrupt handler for netmap rx queues.
1445  */
1446 void
1447 t4_nm_intr(void *arg)
1448 {
1449 	struct sge_nm_rxq *nm_rxq = arg;
1450 
1451 	if (atomic_cmpset_int(&nm_rxq->nm_state, NM_ON, NM_BUSY)) {
1452 		service_nm_rxq(nm_rxq);
1453 		(void) atomic_cmpset_int(&nm_rxq->nm_state, NM_BUSY, NM_ON);
1454 	}
1455 }
1456 
1457 /*
1458  * Interrupt handler for vectors shared between NIC and netmap rx queues.
1459  */
1460 void
1461 t4_vi_intr(void *arg)
1462 {
1463 	struct irq *irq = arg;
1464 
1465 	MPASS(irq->nm_rxq != NULL);
1466 	t4_nm_intr(irq->nm_rxq);
1467 
1468 	MPASS(irq->rxq != NULL);
1469 	t4_intr(irq->rxq);
1470 }
1471 #endif
1472 
1473 /*
1474  * Deals with interrupts on an iq-only (no freelist) queue.
1475  */
1476 static int
1477 service_iq(struct sge_iq *iq, int budget)
1478 {
1479 	struct sge_iq *q;
1480 	struct adapter *sc = iq->adapter;
1481 	struct iq_desc *d = &iq->desc[iq->cidx];
1482 	int ndescs = 0, limit;
1483 	int rsp_type;
1484 	uint32_t lq;
1485 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1486 
1487 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1488 	KASSERT((iq->flags & IQ_HAS_FL) == 0,
1489 	    ("%s: called for iq %p with fl (iq->flags 0x%x)", __func__, iq,
1490 	    iq->flags));
1491 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1492 	MPASS((iq->flags & IQ_LRO_ENABLED) == 0);
1493 
1494 	limit = budget ? budget : iq->qsize / 16;
1495 
1496 	/*
1497 	 * We always come back and check the descriptor ring for new indirect
1498 	 * interrupts and other responses after running a single handler.
1499 	 */
1500 	for (;;) {
1501 		while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1502 
1503 			rmb();
1504 
1505 			rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1506 			lq = be32toh(d->rsp.pldbuflen_qid);
1507 
1508 			switch (rsp_type) {
1509 			case X_RSPD_TYPE_FLBUF:
1510 				panic("%s: data for an iq (%p) with no freelist",
1511 				    __func__, iq);
1512 
1513 				/* NOTREACHED */
1514 
1515 			case X_RSPD_TYPE_CPL:
1516 				KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1517 				    ("%s: bad opcode %02x.", __func__,
1518 				    d->rss.opcode));
1519 				t4_cpl_handler[d->rss.opcode](iq, &d->rss, NULL);
1520 				break;
1521 
1522 			case X_RSPD_TYPE_INTR:
1523 				/*
1524 				 * There are 1K interrupt-capable queues (qids 0
1525 				 * through 1023).  A response type indicating a
1526 				 * forwarded interrupt with a qid >= 1K is an
1527 				 * iWARP async notification.
1528 				 */
1529 				if (__predict_true(lq >= 1024)) {
1530 					t4_an_handler(iq, &d->rsp);
1531 					break;
1532 				}
1533 
1534 				q = sc->sge.iqmap[lq - sc->sge.iq_start -
1535 				    sc->sge.iq_base];
1536 				if (atomic_cmpset_int(&q->state, IQS_IDLE,
1537 				    IQS_BUSY)) {
1538 					if (service_iq_fl(q, q->qsize / 16) == 0) {
1539 						(void) atomic_cmpset_int(&q->state,
1540 						    IQS_BUSY, IQS_IDLE);
1541 					} else {
1542 						STAILQ_INSERT_TAIL(&iql, q,
1543 						    link);
1544 					}
1545 				}
1546 				break;
1547 
1548 			default:
1549 				KASSERT(0,
1550 				    ("%s: illegal response type %d on iq %p",
1551 				    __func__, rsp_type, iq));
1552 				log(LOG_ERR,
1553 				    "%s: illegal response type %d on iq %p",
1554 				    device_get_nameunit(sc->dev), rsp_type, iq);
1555 				break;
1556 			}
1557 
1558 			d++;
1559 			if (__predict_false(++iq->cidx == iq->sidx)) {
1560 				iq->cidx = 0;
1561 				iq->gen ^= F_RSPD_GEN;
1562 				d = &iq->desc[0];
1563 			}
1564 			if (__predict_false(++ndescs == limit)) {
1565 				t4_write_reg(sc, sc->sge_gts_reg,
1566 				    V_CIDXINC(ndescs) |
1567 				    V_INGRESSQID(iq->cntxt_id) |
1568 				    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1569 				ndescs = 0;
1570 
1571 				if (budget) {
1572 					return (EINPROGRESS);
1573 				}
1574 			}
1575 		}
1576 
1577 		if (STAILQ_EMPTY(&iql))
1578 			break;
1579 
1580 		/*
1581 		 * Process the head only, and send it to the back of the list if
1582 		 * it's still not done.
1583 		 */
1584 		q = STAILQ_FIRST(&iql);
1585 		STAILQ_REMOVE_HEAD(&iql, link);
1586 		if (service_iq_fl(q, q->qsize / 8) == 0)
1587 			(void) atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1588 		else
1589 			STAILQ_INSERT_TAIL(&iql, q, link);
1590 	}
1591 
1592 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1593 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1594 
1595 	return (0);
1596 }
1597 
1598 static inline int
1599 sort_before_lro(struct lro_ctrl *lro)
1600 {
1601 
1602 	return (lro->lro_mbuf_max != 0);
1603 }
1604 
1605 static inline uint64_t
1606 last_flit_to_ns(struct adapter *sc, uint64_t lf)
1607 {
1608 	uint64_t n = be64toh(lf) & 0xfffffffffffffff;	/* 60b, not 64b. */
1609 
1610 	if (n > UINT64_MAX / 1000000)
1611 		return (n / sc->params.vpd.cclk * 1000000);
1612 	else
1613 		return (n * 1000000 / sc->params.vpd.cclk);
1614 }
1615 
1616 /*
1617  * Deals with interrupts on an iq+fl queue.
1618  */
1619 static int
1620 service_iq_fl(struct sge_iq *iq, int budget)
1621 {
1622 	struct sge_rxq *rxq = iq_to_rxq(iq);
1623 	struct sge_fl *fl;
1624 	struct adapter *sc = iq->adapter;
1625 	struct iq_desc *d = &iq->desc[iq->cidx];
1626 	int ndescs = 0, limit;
1627 	int rsp_type, refill, starved;
1628 	uint32_t lq;
1629 	uint16_t fl_hw_cidx;
1630 	struct mbuf *m0;
1631 #if defined(INET) || defined(INET6)
1632 	const struct timeval lro_timeout = {0, sc->lro_timeout};
1633 	struct lro_ctrl *lro = &rxq->lro;
1634 #endif
1635 
1636 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1637 	MPASS(iq->flags & IQ_HAS_FL);
1638 
1639 	limit = budget ? budget : iq->qsize / 16;
1640 	fl = &rxq->fl;
1641 	fl_hw_cidx = fl->hw_cidx;	/* stable snapshot */
1642 
1643 #if defined(INET) || defined(INET6)
1644 	if (iq->flags & IQ_ADJ_CREDIT) {
1645 		MPASS(sort_before_lro(lro));
1646 		iq->flags &= ~IQ_ADJ_CREDIT;
1647 		if ((d->rsp.u.type_gen & F_RSPD_GEN) != iq->gen) {
1648 			tcp_lro_flush_all(lro);
1649 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(1) |
1650 			    V_INGRESSQID((u32)iq->cntxt_id) |
1651 			    V_SEINTARM(iq->intr_params));
1652 			return (0);
1653 		}
1654 		ndescs = 1;
1655 	}
1656 #else
1657 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1658 #endif
1659 
1660 	while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1661 
1662 		rmb();
1663 
1664 		refill = 0;
1665 		m0 = NULL;
1666 		rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1667 		lq = be32toh(d->rsp.pldbuflen_qid);
1668 
1669 		switch (rsp_type) {
1670 		case X_RSPD_TYPE_FLBUF:
1671 
1672 			m0 = get_fl_payload(sc, fl, lq);
1673 			if (__predict_false(m0 == NULL))
1674 				goto out;
1675 			refill = IDXDIFF(fl->hw_cidx, fl_hw_cidx, fl->sidx) > 2;
1676 
1677 			if (iq->flags & IQ_RX_TIMESTAMP) {
1678 				/*
1679 				 * Fill up rcv_tstmp but do not set M_TSTMP.
1680 				 * rcv_tstmp is not in the format that the
1681 				 * kernel expects and we don't want to mislead
1682 				 * it.  For now this is only for custom code
1683 				 * that knows how to interpret cxgbe's stamp.
1684 				 */
1685 				m0->m_pkthdr.rcv_tstmp =
1686 				    last_flit_to_ns(sc, d->rsp.u.last_flit);
1687 #ifdef notyet
1688 				m0->m_flags |= M_TSTMP;
1689 #endif
1690 			}
1691 
1692 			/* fall through */
1693 
1694 		case X_RSPD_TYPE_CPL:
1695 			KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1696 			    ("%s: bad opcode %02x.", __func__, d->rss.opcode));
1697 			t4_cpl_handler[d->rss.opcode](iq, &d->rss, m0);
1698 			break;
1699 
1700 		case X_RSPD_TYPE_INTR:
1701 
1702 			/*
1703 			 * There are 1K interrupt-capable queues (qids 0
1704 			 * through 1023).  A response type indicating a
1705 			 * forwarded interrupt with a qid >= 1K is an
1706 			 * iWARP async notification.  That is the only
1707 			 * acceptable indirect interrupt on this queue.
1708 			 */
1709 			if (__predict_false(lq < 1024)) {
1710 				panic("%s: indirect interrupt on iq_fl %p "
1711 				    "with qid %u", __func__, iq, lq);
1712 			}
1713 
1714 			t4_an_handler(iq, &d->rsp);
1715 			break;
1716 
1717 		default:
1718 			KASSERT(0, ("%s: illegal response type %d on iq %p",
1719 			    __func__, rsp_type, iq));
1720 			log(LOG_ERR, "%s: illegal response type %d on iq %p",
1721 			    device_get_nameunit(sc->dev), rsp_type, iq);
1722 			break;
1723 		}
1724 
1725 		d++;
1726 		if (__predict_false(++iq->cidx == iq->sidx)) {
1727 			iq->cidx = 0;
1728 			iq->gen ^= F_RSPD_GEN;
1729 			d = &iq->desc[0];
1730 		}
1731 		if (__predict_false(++ndescs == limit)) {
1732 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1733 			    V_INGRESSQID(iq->cntxt_id) |
1734 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1735 			ndescs = 0;
1736 
1737 #if defined(INET) || defined(INET6)
1738 			if (iq->flags & IQ_LRO_ENABLED &&
1739 			    !sort_before_lro(lro) &&
1740 			    sc->lro_timeout != 0) {
1741 				tcp_lro_flush_inactive(lro, &lro_timeout);
1742 			}
1743 #endif
1744 			if (budget) {
1745 				FL_LOCK(fl);
1746 				refill_fl(sc, fl, 32);
1747 				FL_UNLOCK(fl);
1748 
1749 				return (EINPROGRESS);
1750 			}
1751 		}
1752 		if (refill) {
1753 			FL_LOCK(fl);
1754 			refill_fl(sc, fl, 32);
1755 			FL_UNLOCK(fl);
1756 			fl_hw_cidx = fl->hw_cidx;
1757 		}
1758 	}
1759 out:
1760 #if defined(INET) || defined(INET6)
1761 	if (iq->flags & IQ_LRO_ENABLED) {
1762 		if (ndescs > 0 && lro->lro_mbuf_count > 8) {
1763 			MPASS(sort_before_lro(lro));
1764 			/* hold back one credit and don't flush LRO state */
1765 			iq->flags |= IQ_ADJ_CREDIT;
1766 			ndescs--;
1767 		} else {
1768 			tcp_lro_flush_all(lro);
1769 		}
1770 	}
1771 #endif
1772 
1773 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1774 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1775 
1776 	FL_LOCK(fl);
1777 	starved = refill_fl(sc, fl, 64);
1778 	FL_UNLOCK(fl);
1779 	if (__predict_false(starved != 0))
1780 		add_fl_to_sfl(sc, fl);
1781 
1782 	return (0);
1783 }
1784 
1785 static inline int
1786 cl_has_metadata(struct sge_fl *fl, struct cluster_layout *cll)
1787 {
1788 	int rc = fl->flags & FL_BUF_PACKING || cll->region1 > 0;
1789 
1790 	if (rc)
1791 		MPASS(cll->region3 >= CL_METADATA_SIZE);
1792 
1793 	return (rc);
1794 }
1795 
1796 static inline struct cluster_metadata *
1797 cl_metadata(struct adapter *sc, struct sge_fl *fl, struct cluster_layout *cll,
1798     caddr_t cl)
1799 {
1800 
1801 	if (cl_has_metadata(fl, cll)) {
1802 		struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1803 
1804 		return ((struct cluster_metadata *)(cl + swz->size) - 1);
1805 	}
1806 	return (NULL);
1807 }
1808 
1809 static void
1810 rxb_free(struct mbuf *m)
1811 {
1812 	uma_zone_t zone = m->m_ext.ext_arg1;
1813 	void *cl = m->m_ext.ext_arg2;
1814 
1815 	uma_zfree(zone, cl);
1816 	counter_u64_add(extfree_rels, 1);
1817 }
1818 
1819 /*
1820  * The mbuf returned by this function could be allocated from zone_mbuf or
1821  * constructed in spare room in the cluster.
1822  *
1823  * The mbuf carries the payload in one of these ways
1824  * a) frame inside the mbuf (mbuf from zone_mbuf)
1825  * b) m_cljset (for clusters without metadata) zone_mbuf
1826  * c) m_extaddref (cluster with metadata) inline mbuf
1827  * d) m_extaddref (cluster with metadata) zone_mbuf
1828  */
1829 static struct mbuf *
1830 get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1831     int remaining)
1832 {
1833 	struct mbuf *m;
1834 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1835 	struct cluster_layout *cll = &sd->cll;
1836 	struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1837 	struct hw_buf_info *hwb = &sc->sge.hw_buf_info[cll->hwidx];
1838 	struct cluster_metadata *clm = cl_metadata(sc, fl, cll, sd->cl);
1839 	int len, blen;
1840 	caddr_t payload;
1841 
1842 	blen = hwb->size - fl->rx_offset;	/* max possible in this buf */
1843 	len = min(remaining, blen);
1844 	payload = sd->cl + cll->region1 + fl->rx_offset;
1845 	if (fl->flags & FL_BUF_PACKING) {
1846 		const u_int l = fr_offset + len;
1847 		const u_int pad = roundup2(l, fl->buf_boundary) - l;
1848 
1849 		if (fl->rx_offset + len + pad < hwb->size)
1850 			blen = len + pad;
1851 		MPASS(fl->rx_offset + blen <= hwb->size);
1852 	} else {
1853 		MPASS(fl->rx_offset == 0);	/* not packing */
1854 	}
1855 
1856 
1857 	if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) {
1858 
1859 		/*
1860 		 * Copy payload into a freshly allocated mbuf.
1861 		 */
1862 
1863 		m = fr_offset == 0 ?
1864 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1865 		if (m == NULL)
1866 			return (NULL);
1867 		fl->mbuf_allocated++;
1868 
1869 		/* copy data to mbuf */
1870 		bcopy(payload, mtod(m, caddr_t), len);
1871 
1872 	} else if (sd->nmbuf * MSIZE < cll->region1) {
1873 
1874 		/*
1875 		 * There's spare room in the cluster for an mbuf.  Create one
1876 		 * and associate it with the payload that's in the cluster.
1877 		 */
1878 
1879 		MPASS(clm != NULL);
1880 		m = (struct mbuf *)(sd->cl + sd->nmbuf * MSIZE);
1881 		/* No bzero required */
1882 		if (m_init(m, M_NOWAIT, MT_DATA,
1883 		    fr_offset == 0 ? M_PKTHDR | M_NOFREE : M_NOFREE))
1884 			return (NULL);
1885 		fl->mbuf_inlined++;
1886 		m_extaddref(m, payload, blen, &clm->refcount, rxb_free,
1887 		    swz->zone, sd->cl);
1888 		if (sd->nmbuf++ == 0)
1889 			counter_u64_add(extfree_refs, 1);
1890 
1891 	} else {
1892 
1893 		/*
1894 		 * Grab an mbuf from zone_mbuf and associate it with the
1895 		 * payload in the cluster.
1896 		 */
1897 
1898 		m = fr_offset == 0 ?
1899 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1900 		if (m == NULL)
1901 			return (NULL);
1902 		fl->mbuf_allocated++;
1903 		if (clm != NULL) {
1904 			m_extaddref(m, payload, blen, &clm->refcount,
1905 			    rxb_free, swz->zone, sd->cl);
1906 			if (sd->nmbuf++ == 0)
1907 				counter_u64_add(extfree_refs, 1);
1908 		} else {
1909 			m_cljset(m, sd->cl, swz->type);
1910 			sd->cl = NULL;	/* consumed, not a recycle candidate */
1911 		}
1912 	}
1913 	if (fr_offset == 0)
1914 		m->m_pkthdr.len = remaining;
1915 	m->m_len = len;
1916 
1917 	if (fl->flags & FL_BUF_PACKING) {
1918 		fl->rx_offset += blen;
1919 		MPASS(fl->rx_offset <= hwb->size);
1920 		if (fl->rx_offset < hwb->size)
1921 			return (m);	/* without advancing the cidx */
1922 	}
1923 
1924 	if (__predict_false(++fl->cidx % 8 == 0)) {
1925 		uint16_t cidx = fl->cidx / 8;
1926 
1927 		if (__predict_false(cidx == fl->sidx))
1928 			fl->cidx = cidx = 0;
1929 		fl->hw_cidx = cidx;
1930 	}
1931 	fl->rx_offset = 0;
1932 
1933 	return (m);
1934 }
1935 
1936 static struct mbuf *
1937 get_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf)
1938 {
1939 	struct mbuf *m0, *m, **pnext;
1940 	u_int remaining;
1941 	const u_int total = G_RSPD_LEN(len_newbuf);
1942 
1943 	if (__predict_false(fl->flags & FL_BUF_RESUME)) {
1944 		M_ASSERTPKTHDR(fl->m0);
1945 		MPASS(fl->m0->m_pkthdr.len == total);
1946 		MPASS(fl->remaining < total);
1947 
1948 		m0 = fl->m0;
1949 		pnext = fl->pnext;
1950 		remaining = fl->remaining;
1951 		fl->flags &= ~FL_BUF_RESUME;
1952 		goto get_segment;
1953 	}
1954 
1955 	if (fl->rx_offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
1956 		fl->rx_offset = 0;
1957 		if (__predict_false(++fl->cidx % 8 == 0)) {
1958 			uint16_t cidx = fl->cidx / 8;
1959 
1960 			if (__predict_false(cidx == fl->sidx))
1961 				fl->cidx = cidx = 0;
1962 			fl->hw_cidx = cidx;
1963 		}
1964 	}
1965 
1966 	/*
1967 	 * Payload starts at rx_offset in the current hw buffer.  Its length is
1968 	 * 'len' and it may span multiple hw buffers.
1969 	 */
1970 
1971 	m0 = get_scatter_segment(sc, fl, 0, total);
1972 	if (m0 == NULL)
1973 		return (NULL);
1974 	remaining = total - m0->m_len;
1975 	pnext = &m0->m_next;
1976 	while (remaining > 0) {
1977 get_segment:
1978 		MPASS(fl->rx_offset == 0);
1979 		m = get_scatter_segment(sc, fl, total - remaining, remaining);
1980 		if (__predict_false(m == NULL)) {
1981 			fl->m0 = m0;
1982 			fl->pnext = pnext;
1983 			fl->remaining = remaining;
1984 			fl->flags |= FL_BUF_RESUME;
1985 			return (NULL);
1986 		}
1987 		*pnext = m;
1988 		pnext = &m->m_next;
1989 		remaining -= m->m_len;
1990 	}
1991 	*pnext = NULL;
1992 
1993 	M_ASSERTPKTHDR(m0);
1994 	return (m0);
1995 }
1996 
1997 static int
1998 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
1999 {
2000 	struct sge_rxq *rxq = iq_to_rxq(iq);
2001 	struct ifnet *ifp = rxq->ifp;
2002 	struct adapter *sc = iq->adapter;
2003 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
2004 #if defined(INET) || defined(INET6)
2005 	struct lro_ctrl *lro = &rxq->lro;
2006 #endif
2007 	static const int sw_hashtype[4][2] = {
2008 		{M_HASHTYPE_NONE, M_HASHTYPE_NONE},
2009 		{M_HASHTYPE_RSS_IPV4, M_HASHTYPE_RSS_IPV6},
2010 		{M_HASHTYPE_RSS_TCP_IPV4, M_HASHTYPE_RSS_TCP_IPV6},
2011 		{M_HASHTYPE_RSS_UDP_IPV4, M_HASHTYPE_RSS_UDP_IPV6},
2012 	};
2013 
2014 	KASSERT(m0 != NULL, ("%s: no payload with opcode %02x", __func__,
2015 	    rss->opcode));
2016 
2017 	m0->m_pkthdr.len -= sc->params.sge.fl_pktshift;
2018 	m0->m_len -= sc->params.sge.fl_pktshift;
2019 	m0->m_data += sc->params.sge.fl_pktshift;
2020 
2021 	m0->m_pkthdr.rcvif = ifp;
2022 	M_HASHTYPE_SET(m0, sw_hashtype[rss->hash_type][rss->ipv6]);
2023 	m0->m_pkthdr.flowid = be32toh(rss->hash_val);
2024 
2025 	if (cpl->csum_calc && !(cpl->err_vec & sc->params.tp.err_vec_mask)) {
2026 		if (ifp->if_capenable & IFCAP_RXCSUM &&
2027 		    cpl->l2info & htobe32(F_RXF_IP)) {
2028 			m0->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
2029 			    CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
2030 			rxq->rxcsum++;
2031 		} else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
2032 		    cpl->l2info & htobe32(F_RXF_IP6)) {
2033 			m0->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
2034 			    CSUM_PSEUDO_HDR);
2035 			rxq->rxcsum++;
2036 		}
2037 
2038 		if (__predict_false(cpl->ip_frag))
2039 			m0->m_pkthdr.csum_data = be16toh(cpl->csum);
2040 		else
2041 			m0->m_pkthdr.csum_data = 0xffff;
2042 	}
2043 
2044 	if (cpl->vlan_ex) {
2045 		m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
2046 		m0->m_flags |= M_VLANTAG;
2047 		rxq->vlan_extraction++;
2048 	}
2049 
2050 #ifdef NUMA
2051 	m0->m_pkthdr.numa_domain = ifp->if_numa_domain;
2052 #endif
2053 #if defined(INET) || defined(INET6)
2054 	if (iq->flags & IQ_LRO_ENABLED) {
2055 		if (sort_before_lro(lro)) {
2056 			tcp_lro_queue_mbuf(lro, m0);
2057 			return (0); /* queued for sort, then LRO */
2058 		}
2059 		if (tcp_lro_rx(lro, m0, 0) == 0)
2060 			return (0); /* queued for LRO */
2061 	}
2062 #endif
2063 	ifp->if_input(ifp, m0);
2064 
2065 	return (0);
2066 }
2067 
2068 /*
2069  * Must drain the wrq or make sure that someone else will.
2070  */
2071 static void
2072 wrq_tx_drain(void *arg, int n)
2073 {
2074 	struct sge_wrq *wrq = arg;
2075 	struct sge_eq *eq = &wrq->eq;
2076 
2077 	EQ_LOCK(eq);
2078 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2079 		drain_wrq_wr_list(wrq->adapter, wrq);
2080 	EQ_UNLOCK(eq);
2081 }
2082 
2083 static void
2084 drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq)
2085 {
2086 	struct sge_eq *eq = &wrq->eq;
2087 	u_int available, dbdiff;	/* # of hardware descriptors */
2088 	u_int n;
2089 	struct wrqe *wr;
2090 	struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
2091 
2092 	EQ_LOCK_ASSERT_OWNED(eq);
2093 	MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
2094 	wr = STAILQ_FIRST(&wrq->wr_list);
2095 	MPASS(wr != NULL);	/* Must be called with something useful to do */
2096 	MPASS(eq->pidx == eq->dbidx);
2097 	dbdiff = 0;
2098 
2099 	do {
2100 		eq->cidx = read_hw_cidx(eq);
2101 		if (eq->pidx == eq->cidx)
2102 			available = eq->sidx - 1;
2103 		else
2104 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2105 
2106 		MPASS(wr->wrq == wrq);
2107 		n = howmany(wr->wr_len, EQ_ESIZE);
2108 		if (available < n)
2109 			break;
2110 
2111 		dst = (void *)&eq->desc[eq->pidx];
2112 		if (__predict_true(eq->sidx - eq->pidx > n)) {
2113 			/* Won't wrap, won't end exactly at the status page. */
2114 			bcopy(&wr->wr[0], dst, wr->wr_len);
2115 			eq->pidx += n;
2116 		} else {
2117 			int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE;
2118 
2119 			bcopy(&wr->wr[0], dst, first_portion);
2120 			if (wr->wr_len > first_portion) {
2121 				bcopy(&wr->wr[first_portion], &eq->desc[0],
2122 				    wr->wr_len - first_portion);
2123 			}
2124 			eq->pidx = n - (eq->sidx - eq->pidx);
2125 		}
2126 		wrq->tx_wrs_copied++;
2127 
2128 		if (available < eq->sidx / 4 &&
2129 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2130 				/*
2131 				 * XXX: This is not 100% reliable with some
2132 				 * types of WRs.  But this is a very unusual
2133 				 * situation for an ofld/ctrl queue anyway.
2134 				 */
2135 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2136 			    F_FW_WR_EQUEQ);
2137 		}
2138 
2139 		dbdiff += n;
2140 		if (dbdiff >= 16) {
2141 			ring_eq_db(sc, eq, dbdiff);
2142 			dbdiff = 0;
2143 		}
2144 
2145 		STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
2146 		free_wrqe(wr);
2147 		MPASS(wrq->nwr_pending > 0);
2148 		wrq->nwr_pending--;
2149 		MPASS(wrq->ndesc_needed >= n);
2150 		wrq->ndesc_needed -= n;
2151 	} while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL);
2152 
2153 	if (dbdiff)
2154 		ring_eq_db(sc, eq, dbdiff);
2155 }
2156 
2157 /*
2158  * Doesn't fail.  Holds on to work requests it can't send right away.
2159  */
2160 void
2161 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
2162 {
2163 #ifdef INVARIANTS
2164 	struct sge_eq *eq = &wrq->eq;
2165 #endif
2166 
2167 	EQ_LOCK_ASSERT_OWNED(eq);
2168 	MPASS(wr != NULL);
2169 	MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN);
2170 	MPASS((wr->wr_len & 0x7) == 0);
2171 
2172 	STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
2173 	wrq->nwr_pending++;
2174 	wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE);
2175 
2176 	if (!TAILQ_EMPTY(&wrq->incomplete_wrs))
2177 		return;	/* commit_wrq_wr will drain wr_list as well. */
2178 
2179 	drain_wrq_wr_list(sc, wrq);
2180 
2181 	/* Doorbell must have caught up to the pidx. */
2182 	MPASS(eq->pidx == eq->dbidx);
2183 }
2184 
2185 void
2186 t4_update_fl_bufsize(struct ifnet *ifp)
2187 {
2188 	struct vi_info *vi = ifp->if_softc;
2189 	struct adapter *sc = vi->pi->adapter;
2190 	struct sge_rxq *rxq;
2191 #ifdef TCP_OFFLOAD
2192 	struct sge_ofld_rxq *ofld_rxq;
2193 #endif
2194 	struct sge_fl *fl;
2195 	int i, maxp, mtu = ifp->if_mtu;
2196 
2197 	maxp = mtu_to_max_payload(sc, mtu, 0);
2198 	for_each_rxq(vi, i, rxq) {
2199 		fl = &rxq->fl;
2200 
2201 		FL_LOCK(fl);
2202 		find_best_refill_source(sc, fl, maxp);
2203 		FL_UNLOCK(fl);
2204 	}
2205 #ifdef TCP_OFFLOAD
2206 	maxp = mtu_to_max_payload(sc, mtu, 1);
2207 	for_each_ofld_rxq(vi, i, ofld_rxq) {
2208 		fl = &ofld_rxq->fl;
2209 
2210 		FL_LOCK(fl);
2211 		find_best_refill_source(sc, fl, maxp);
2212 		FL_UNLOCK(fl);
2213 	}
2214 #endif
2215 }
2216 
2217 static inline int
2218 mbuf_nsegs(struct mbuf *m)
2219 {
2220 
2221 	M_ASSERTPKTHDR(m);
2222 	KASSERT(m->m_pkthdr.l5hlen > 0,
2223 	    ("%s: mbuf %p missing information on # of segments.", __func__, m));
2224 
2225 	return (m->m_pkthdr.l5hlen);
2226 }
2227 
2228 static inline void
2229 set_mbuf_nsegs(struct mbuf *m, uint8_t nsegs)
2230 {
2231 
2232 	M_ASSERTPKTHDR(m);
2233 	m->m_pkthdr.l5hlen = nsegs;
2234 }
2235 
2236 static inline int
2237 mbuf_cflags(struct mbuf *m)
2238 {
2239 
2240 	M_ASSERTPKTHDR(m);
2241 	return (m->m_pkthdr.PH_loc.eight[4]);
2242 }
2243 
2244 static inline void
2245 set_mbuf_cflags(struct mbuf *m, uint8_t flags)
2246 {
2247 
2248 	M_ASSERTPKTHDR(m);
2249 	m->m_pkthdr.PH_loc.eight[4] = flags;
2250 }
2251 
2252 static inline int
2253 mbuf_len16(struct mbuf *m)
2254 {
2255 	int n;
2256 
2257 	M_ASSERTPKTHDR(m);
2258 	n = m->m_pkthdr.PH_loc.eight[0];
2259 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2260 
2261 	return (n);
2262 }
2263 
2264 static inline void
2265 set_mbuf_len16(struct mbuf *m, uint8_t len16)
2266 {
2267 
2268 	M_ASSERTPKTHDR(m);
2269 	m->m_pkthdr.PH_loc.eight[0] = len16;
2270 }
2271 
2272 #ifdef RATELIMIT
2273 static inline int
2274 mbuf_eo_nsegs(struct mbuf *m)
2275 {
2276 
2277 	M_ASSERTPKTHDR(m);
2278 	return (m->m_pkthdr.PH_loc.eight[1]);
2279 }
2280 
2281 static inline void
2282 set_mbuf_eo_nsegs(struct mbuf *m, uint8_t nsegs)
2283 {
2284 
2285 	M_ASSERTPKTHDR(m);
2286 	m->m_pkthdr.PH_loc.eight[1] = nsegs;
2287 }
2288 
2289 static inline int
2290 mbuf_eo_len16(struct mbuf *m)
2291 {
2292 	int n;
2293 
2294 	M_ASSERTPKTHDR(m);
2295 	n = m->m_pkthdr.PH_loc.eight[2];
2296 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2297 
2298 	return (n);
2299 }
2300 
2301 static inline void
2302 set_mbuf_eo_len16(struct mbuf *m, uint8_t len16)
2303 {
2304 
2305 	M_ASSERTPKTHDR(m);
2306 	m->m_pkthdr.PH_loc.eight[2] = len16;
2307 }
2308 
2309 static inline int
2310 mbuf_eo_tsclk_tsoff(struct mbuf *m)
2311 {
2312 
2313 	M_ASSERTPKTHDR(m);
2314 	return (m->m_pkthdr.PH_loc.eight[3]);
2315 }
2316 
2317 static inline void
2318 set_mbuf_eo_tsclk_tsoff(struct mbuf *m, uint8_t tsclk_tsoff)
2319 {
2320 
2321 	M_ASSERTPKTHDR(m);
2322 	m->m_pkthdr.PH_loc.eight[3] = tsclk_tsoff;
2323 }
2324 
2325 static inline int
2326 needs_eo(struct mbuf *m)
2327 {
2328 
2329 	return (m->m_pkthdr.csum_flags & CSUM_SND_TAG);
2330 }
2331 #endif
2332 
2333 /*
2334  * Try to allocate an mbuf to contain a raw work request.  To make it
2335  * easy to construct the work request, don't allocate a chain but a
2336  * single mbuf.
2337  */
2338 struct mbuf *
2339 alloc_wr_mbuf(int len, int how)
2340 {
2341 	struct mbuf *m;
2342 
2343 	if (len <= MHLEN)
2344 		m = m_gethdr(how, MT_DATA);
2345 	else if (len <= MCLBYTES)
2346 		m = m_getcl(how, MT_DATA, M_PKTHDR);
2347 	else
2348 		m = NULL;
2349 	if (m == NULL)
2350 		return (NULL);
2351 	m->m_pkthdr.len = len;
2352 	m->m_len = len;
2353 	set_mbuf_cflags(m, MC_RAW_WR);
2354 	set_mbuf_len16(m, howmany(len, 16));
2355 	return (m);
2356 }
2357 
2358 static inline int
2359 needs_tso(struct mbuf *m)
2360 {
2361 
2362 	M_ASSERTPKTHDR(m);
2363 
2364 	return (m->m_pkthdr.csum_flags & CSUM_TSO);
2365 }
2366 
2367 static inline int
2368 needs_l3_csum(struct mbuf *m)
2369 {
2370 
2371 	M_ASSERTPKTHDR(m);
2372 
2373 	return (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO));
2374 }
2375 
2376 static inline int
2377 needs_l4_csum(struct mbuf *m)
2378 {
2379 
2380 	M_ASSERTPKTHDR(m);
2381 
2382 	return (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
2383 	    CSUM_TCP_IPV6 | CSUM_TSO));
2384 }
2385 
2386 static inline int
2387 needs_tcp_csum(struct mbuf *m)
2388 {
2389 
2390 	M_ASSERTPKTHDR(m);
2391 	return (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_TCP_IPV6 | CSUM_TSO));
2392 }
2393 
2394 #ifdef RATELIMIT
2395 static inline int
2396 needs_udp_csum(struct mbuf *m)
2397 {
2398 
2399 	M_ASSERTPKTHDR(m);
2400 	return (m->m_pkthdr.csum_flags & (CSUM_UDP | CSUM_UDP_IPV6));
2401 }
2402 #endif
2403 
2404 static inline int
2405 needs_vlan_insertion(struct mbuf *m)
2406 {
2407 
2408 	M_ASSERTPKTHDR(m);
2409 
2410 	return (m->m_flags & M_VLANTAG);
2411 }
2412 
2413 static void *
2414 m_advance(struct mbuf **pm, int *poffset, int len)
2415 {
2416 	struct mbuf *m = *pm;
2417 	int offset = *poffset;
2418 	uintptr_t p = 0;
2419 
2420 	MPASS(len > 0);
2421 
2422 	for (;;) {
2423 		if (offset + len < m->m_len) {
2424 			offset += len;
2425 			p = mtod(m, uintptr_t) + offset;
2426 			break;
2427 		}
2428 		len -= m->m_len - offset;
2429 		m = m->m_next;
2430 		offset = 0;
2431 		MPASS(m != NULL);
2432 	}
2433 	*poffset = offset;
2434 	*pm = m;
2435 	return ((void *)p);
2436 }
2437 
2438 static inline int
2439 count_mbuf_ext_pgs(struct mbuf *m, int skip, vm_paddr_t *nextaddr)
2440 {
2441 	struct mbuf_ext_pgs *ext_pgs;
2442 	vm_paddr_t paddr;
2443 	int i, len, off, pglen, pgoff, seglen, segoff;
2444 	int nsegs = 0;
2445 
2446 	MBUF_EXT_PGS_ASSERT(m);
2447 	ext_pgs = m->m_ext.ext_pgs;
2448 	off = mtod(m, vm_offset_t);
2449 	len = m->m_len;
2450 	off += skip;
2451 	len -= skip;
2452 
2453 	if (ext_pgs->hdr_len != 0) {
2454 		if (off >= ext_pgs->hdr_len) {
2455 			off -= ext_pgs->hdr_len;
2456 		} else {
2457 			seglen = ext_pgs->hdr_len - off;
2458 			segoff = off;
2459 			seglen = min(seglen, len);
2460 			off = 0;
2461 			len -= seglen;
2462 			paddr = pmap_kextract(
2463 			    (vm_offset_t)&ext_pgs->hdr[segoff]);
2464 			if (*nextaddr != paddr)
2465 				nsegs++;
2466 			*nextaddr = paddr + seglen;
2467 		}
2468 	}
2469 	pgoff = ext_pgs->first_pg_off;
2470 	for (i = 0; i < ext_pgs->npgs && len > 0; i++) {
2471 		pglen = mbuf_ext_pg_len(ext_pgs, i, pgoff);
2472 		if (off >= pglen) {
2473 			off -= pglen;
2474 			pgoff = 0;
2475 			continue;
2476 		}
2477 		seglen = pglen - off;
2478 		segoff = pgoff + off;
2479 		off = 0;
2480 		seglen = min(seglen, len);
2481 		len -= seglen;
2482 		paddr = ext_pgs->pa[i] + segoff;
2483 		if (*nextaddr != paddr)
2484 			nsegs++;
2485 		*nextaddr = paddr + seglen;
2486 		pgoff = 0;
2487 	};
2488 	if (len != 0) {
2489 		seglen = min(len, ext_pgs->trail_len - off);
2490 		len -= seglen;
2491 		paddr = pmap_kextract((vm_offset_t)&ext_pgs->trail[off]);
2492 		if (*nextaddr != paddr)
2493 			nsegs++;
2494 		*nextaddr = paddr + seglen;
2495 	}
2496 
2497 	return (nsegs);
2498 }
2499 
2500 
2501 /*
2502  * Can deal with empty mbufs in the chain that have m_len = 0, but the chain
2503  * must have at least one mbuf that's not empty.  It is possible for this
2504  * routine to return 0 if skip accounts for all the contents of the mbuf chain.
2505  */
2506 static inline int
2507 count_mbuf_nsegs(struct mbuf *m, int skip, uint8_t *cflags)
2508 {
2509 	vm_paddr_t nextaddr, paddr;
2510 	vm_offset_t va;
2511 	int len, nsegs;
2512 
2513 	M_ASSERTPKTHDR(m);
2514 	MPASS(m->m_pkthdr.len > 0);
2515 	MPASS(m->m_pkthdr.len >= skip);
2516 
2517 	nsegs = 0;
2518 	nextaddr = 0;
2519 	for (; m; m = m->m_next) {
2520 		len = m->m_len;
2521 		if (__predict_false(len == 0))
2522 			continue;
2523 		if (skip >= len) {
2524 			skip -= len;
2525 			continue;
2526 		}
2527 		if ((m->m_flags & M_NOMAP) != 0) {
2528 			*cflags |= MC_NOMAP;
2529 			nsegs += count_mbuf_ext_pgs(m, skip, &nextaddr);
2530 			skip = 0;
2531 			continue;
2532 		}
2533 		va = mtod(m, vm_offset_t) + skip;
2534 		len -= skip;
2535 		skip = 0;
2536 		paddr = pmap_kextract(va);
2537 		nsegs += sglist_count((void *)(uintptr_t)va, len);
2538 		if (paddr == nextaddr)
2539 			nsegs--;
2540 		nextaddr = pmap_kextract(va + len - 1) + 1;
2541 	}
2542 
2543 	return (nsegs);
2544 }
2545 
2546 /*
2547  * Analyze the mbuf to determine its tx needs.  The mbuf passed in may change:
2548  * a) caller can assume it's been freed if this function returns with an error.
2549  * b) it may get defragged up if the gather list is too long for the hardware.
2550  */
2551 int
2552 parse_pkt(struct adapter *sc, struct mbuf **mp)
2553 {
2554 	struct mbuf *m0 = *mp, *m;
2555 	int rc, nsegs, defragged = 0, offset;
2556 	struct ether_header *eh;
2557 	void *l3hdr;
2558 #if defined(INET) || defined(INET6)
2559 	struct tcphdr *tcp;
2560 #endif
2561 	uint16_t eh_type;
2562 	uint8_t cflags;
2563 
2564 	cflags = 0;
2565 	M_ASSERTPKTHDR(m0);
2566 	if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) {
2567 		rc = EINVAL;
2568 fail:
2569 		m_freem(m0);
2570 		*mp = NULL;
2571 		return (rc);
2572 	}
2573 restart:
2574 	/*
2575 	 * First count the number of gather list segments in the payload.
2576 	 * Defrag the mbuf if nsegs exceeds the hardware limit.
2577 	 */
2578 	M_ASSERTPKTHDR(m0);
2579 	MPASS(m0->m_pkthdr.len > 0);
2580 	nsegs = count_mbuf_nsegs(m0, 0, &cflags);
2581 	if (nsegs > (needs_tso(m0) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS)) {
2582 		if (defragged++ > 0 || (m = m_defrag(m0, M_NOWAIT)) == NULL) {
2583 			rc = EFBIG;
2584 			goto fail;
2585 		}
2586 		*mp = m0 = m;	/* update caller's copy after defrag */
2587 		goto restart;
2588 	}
2589 
2590 	if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN &&
2591 	    !(cflags & MC_NOMAP))) {
2592 		m0 = m_pullup(m0, m0->m_pkthdr.len);
2593 		if (m0 == NULL) {
2594 			/* Should have left well enough alone. */
2595 			rc = EFBIG;
2596 			goto fail;
2597 		}
2598 		*mp = m0;	/* update caller's copy after pullup */
2599 		goto restart;
2600 	}
2601 	set_mbuf_nsegs(m0, nsegs);
2602 	set_mbuf_cflags(m0, cflags);
2603 	if (sc->flags & IS_VF)
2604 		set_mbuf_len16(m0, txpkt_vm_len16(nsegs, needs_tso(m0)));
2605 	else
2606 		set_mbuf_len16(m0, txpkt_len16(nsegs, needs_tso(m0)));
2607 
2608 #ifdef RATELIMIT
2609 	/*
2610 	 * Ethofld is limited to TCP and UDP for now, and only when L4 hw
2611 	 * checksumming is enabled.  needs_l4_csum happens to check for all the
2612 	 * right things.
2613 	 */
2614 	if (__predict_false(needs_eo(m0) && !needs_l4_csum(m0))) {
2615 		m_snd_tag_rele(m0->m_pkthdr.snd_tag);
2616 		m0->m_pkthdr.snd_tag = NULL;
2617 		m0->m_pkthdr.csum_flags &= ~CSUM_SND_TAG;
2618 	}
2619 #endif
2620 
2621 	if (!needs_tso(m0) &&
2622 #ifdef RATELIMIT
2623 	    !needs_eo(m0) &&
2624 #endif
2625 	    !(sc->flags & IS_VF && (needs_l3_csum(m0) || needs_l4_csum(m0))))
2626 		return (0);
2627 
2628 	m = m0;
2629 	eh = mtod(m, struct ether_header *);
2630 	eh_type = ntohs(eh->ether_type);
2631 	if (eh_type == ETHERTYPE_VLAN) {
2632 		struct ether_vlan_header *evh = (void *)eh;
2633 
2634 		eh_type = ntohs(evh->evl_proto);
2635 		m0->m_pkthdr.l2hlen = sizeof(*evh);
2636 	} else
2637 		m0->m_pkthdr.l2hlen = sizeof(*eh);
2638 
2639 	offset = 0;
2640 	l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2641 
2642 	switch (eh_type) {
2643 #ifdef INET6
2644 	case ETHERTYPE_IPV6:
2645 	{
2646 		struct ip6_hdr *ip6 = l3hdr;
2647 
2648 		MPASS(!needs_tso(m0) || ip6->ip6_nxt == IPPROTO_TCP);
2649 
2650 		m0->m_pkthdr.l3hlen = sizeof(*ip6);
2651 		break;
2652 	}
2653 #endif
2654 #ifdef INET
2655 	case ETHERTYPE_IP:
2656 	{
2657 		struct ip *ip = l3hdr;
2658 
2659 		m0->m_pkthdr.l3hlen = ip->ip_hl * 4;
2660 		break;
2661 	}
2662 #endif
2663 	default:
2664 		panic("%s: ethertype 0x%04x unknown.  if_cxgbe must be compiled"
2665 		    " with the same INET/INET6 options as the kernel.",
2666 		    __func__, eh_type);
2667 	}
2668 
2669 #if defined(INET) || defined(INET6)
2670 	if (needs_tcp_csum(m0)) {
2671 		tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen);
2672 		m0->m_pkthdr.l4hlen = tcp->th_off * 4;
2673 #ifdef RATELIMIT
2674 		if (tsclk >= 0 && *(uint32_t *)(tcp + 1) == ntohl(0x0101080a)) {
2675 			set_mbuf_eo_tsclk_tsoff(m0,
2676 			    V_FW_ETH_TX_EO_WR_TSCLK(tsclk) |
2677 			    V_FW_ETH_TX_EO_WR_TSOFF(sizeof(*tcp) / 2 + 1));
2678 		} else
2679 			set_mbuf_eo_tsclk_tsoff(m0, 0);
2680 	} else if (needs_udp_csum(m)) {
2681 		m0->m_pkthdr.l4hlen = sizeof(struct udphdr);
2682 #endif
2683 	}
2684 #ifdef RATELIMIT
2685 	if (needs_eo(m0)) {
2686 		u_int immhdrs;
2687 
2688 		/* EO WRs have the headers in the WR and not the GL. */
2689 		immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen +
2690 		    m0->m_pkthdr.l4hlen;
2691 		cflags = 0;
2692 		nsegs = count_mbuf_nsegs(m0, immhdrs, &cflags);
2693 		MPASS(cflags == mbuf_cflags(m0));
2694 		set_mbuf_eo_nsegs(m0, nsegs);
2695 		set_mbuf_eo_len16(m0,
2696 		    txpkt_eo_len16(nsegs, immhdrs, needs_tso(m0)));
2697 	}
2698 #endif
2699 #endif
2700 	MPASS(m0 == *mp);
2701 	return (0);
2702 }
2703 
2704 void *
2705 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie)
2706 {
2707 	struct sge_eq *eq = &wrq->eq;
2708 	struct adapter *sc = wrq->adapter;
2709 	int ndesc, available;
2710 	struct wrqe *wr;
2711 	void *w;
2712 
2713 	MPASS(len16 > 0);
2714 	ndesc = howmany(len16, EQ_ESIZE / 16);
2715 	MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC);
2716 
2717 	EQ_LOCK(eq);
2718 
2719 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2720 		drain_wrq_wr_list(sc, wrq);
2721 
2722 	if (!STAILQ_EMPTY(&wrq->wr_list)) {
2723 slowpath:
2724 		EQ_UNLOCK(eq);
2725 		wr = alloc_wrqe(len16 * 16, wrq);
2726 		if (__predict_false(wr == NULL))
2727 			return (NULL);
2728 		cookie->pidx = -1;
2729 		cookie->ndesc = ndesc;
2730 		return (&wr->wr);
2731 	}
2732 
2733 	eq->cidx = read_hw_cidx(eq);
2734 	if (eq->pidx == eq->cidx)
2735 		available = eq->sidx - 1;
2736 	else
2737 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2738 	if (available < ndesc)
2739 		goto slowpath;
2740 
2741 	cookie->pidx = eq->pidx;
2742 	cookie->ndesc = ndesc;
2743 	TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link);
2744 
2745 	w = &eq->desc[eq->pidx];
2746 	IDXINCR(eq->pidx, ndesc, eq->sidx);
2747 	if (__predict_false(cookie->pidx + ndesc > eq->sidx)) {
2748 		w = &wrq->ss[0];
2749 		wrq->ss_pidx = cookie->pidx;
2750 		wrq->ss_len = len16 * 16;
2751 	}
2752 
2753 	EQ_UNLOCK(eq);
2754 
2755 	return (w);
2756 }
2757 
2758 void
2759 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie)
2760 {
2761 	struct sge_eq *eq = &wrq->eq;
2762 	struct adapter *sc = wrq->adapter;
2763 	int ndesc, pidx;
2764 	struct wrq_cookie *prev, *next;
2765 
2766 	if (cookie->pidx == -1) {
2767 		struct wrqe *wr = __containerof(w, struct wrqe, wr);
2768 
2769 		t4_wrq_tx(sc, wr);
2770 		return;
2771 	}
2772 
2773 	if (__predict_false(w == &wrq->ss[0])) {
2774 		int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE;
2775 
2776 		MPASS(wrq->ss_len > n);	/* WR had better wrap around. */
2777 		bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n);
2778 		bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n);
2779 		wrq->tx_wrs_ss++;
2780 	} else
2781 		wrq->tx_wrs_direct++;
2782 
2783 	EQ_LOCK(eq);
2784 	ndesc = cookie->ndesc;	/* Can be more than SGE_MAX_WR_NDESC here. */
2785 	pidx = cookie->pidx;
2786 	MPASS(pidx >= 0 && pidx < eq->sidx);
2787 	prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link);
2788 	next = TAILQ_NEXT(cookie, link);
2789 	if (prev == NULL) {
2790 		MPASS(pidx == eq->dbidx);
2791 		if (next == NULL || ndesc >= 16) {
2792 			int available;
2793 			struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
2794 
2795 			/*
2796 			 * Note that the WR via which we'll request tx updates
2797 			 * is at pidx and not eq->pidx, which has moved on
2798 			 * already.
2799 			 */
2800 			dst = (void *)&eq->desc[pidx];
2801 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2802 			if (available < eq->sidx / 4 &&
2803 			    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2804 				/*
2805 				 * XXX: This is not 100% reliable with some
2806 				 * types of WRs.  But this is a very unusual
2807 				 * situation for an ofld/ctrl queue anyway.
2808 				 */
2809 				dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2810 				    F_FW_WR_EQUEQ);
2811 			}
2812 
2813 			ring_eq_db(wrq->adapter, eq, ndesc);
2814 		} else {
2815 			MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc);
2816 			next->pidx = pidx;
2817 			next->ndesc += ndesc;
2818 		}
2819 	} else {
2820 		MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc);
2821 		prev->ndesc += ndesc;
2822 	}
2823 	TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link);
2824 
2825 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2826 		drain_wrq_wr_list(sc, wrq);
2827 
2828 #ifdef INVARIANTS
2829 	if (TAILQ_EMPTY(&wrq->incomplete_wrs)) {
2830 		/* Doorbell must have caught up to the pidx. */
2831 		MPASS(wrq->eq.pidx == wrq->eq.dbidx);
2832 	}
2833 #endif
2834 	EQ_UNLOCK(eq);
2835 }
2836 
2837 static u_int
2838 can_resume_eth_tx(struct mp_ring *r)
2839 {
2840 	struct sge_eq *eq = r->cookie;
2841 
2842 	return (total_available_tx_desc(eq) > eq->sidx / 8);
2843 }
2844 
2845 static inline int
2846 cannot_use_txpkts(struct mbuf *m)
2847 {
2848 	/* maybe put a GL limit too, to avoid silliness? */
2849 
2850 	return (needs_tso(m) || (mbuf_cflags(m) & MC_RAW_WR) != 0);
2851 }
2852 
2853 static inline int
2854 discard_tx(struct sge_eq *eq)
2855 {
2856 
2857 	return ((eq->flags & (EQ_ENABLED | EQ_QFLUSH)) != EQ_ENABLED);
2858 }
2859 
2860 static inline int
2861 wr_can_update_eq(struct fw_eth_tx_pkts_wr *wr)
2862 {
2863 
2864 	switch (G_FW_WR_OP(be32toh(wr->op_pkd))) {
2865 	case FW_ULPTX_WR:
2866 	case FW_ETH_TX_PKT_WR:
2867 	case FW_ETH_TX_PKTS_WR:
2868 	case FW_ETH_TX_PKT_VM_WR:
2869 		return (1);
2870 	default:
2871 		return (0);
2872 	}
2873 }
2874 
2875 /*
2876  * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to
2877  * be consumed.  Return the actual number consumed.  0 indicates a stall.
2878  */
2879 static u_int
2880 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx)
2881 {
2882 	struct sge_txq *txq = r->cookie;
2883 	struct sge_eq *eq = &txq->eq;
2884 	struct ifnet *ifp = txq->ifp;
2885 	struct vi_info *vi = ifp->if_softc;
2886 	struct port_info *pi = vi->pi;
2887 	struct adapter *sc = pi->adapter;
2888 	u_int total, remaining;		/* # of packets */
2889 	u_int available, dbdiff;	/* # of hardware descriptors */
2890 	u_int n, next_cidx;
2891 	struct mbuf *m0, *tail;
2892 	struct txpkts txp;
2893 	struct fw_eth_tx_pkts_wr *wr;	/* any fw WR struct will do */
2894 
2895 	remaining = IDXDIFF(pidx, cidx, r->size);
2896 	MPASS(remaining > 0);	/* Must not be called without work to do. */
2897 	total = 0;
2898 
2899 	TXQ_LOCK(txq);
2900 	if (__predict_false(discard_tx(eq))) {
2901 		while (cidx != pidx) {
2902 			m0 = r->items[cidx];
2903 			m_freem(m0);
2904 			if (++cidx == r->size)
2905 				cidx = 0;
2906 		}
2907 		reclaim_tx_descs(txq, 2048);
2908 		total = remaining;
2909 		goto done;
2910 	}
2911 
2912 	/* How many hardware descriptors do we have readily available. */
2913 	if (eq->pidx == eq->cidx)
2914 		available = eq->sidx - 1;
2915 	else
2916 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2917 	dbdiff = IDXDIFF(eq->pidx, eq->dbidx, eq->sidx);
2918 
2919 	while (remaining > 0) {
2920 
2921 		m0 = r->items[cidx];
2922 		M_ASSERTPKTHDR(m0);
2923 		MPASS(m0->m_nextpkt == NULL);
2924 
2925 		if (available < SGE_MAX_WR_NDESC) {
2926 			available += reclaim_tx_descs(txq, 64);
2927 			if (available < howmany(mbuf_len16(m0), EQ_ESIZE / 16))
2928 				break;	/* out of descriptors */
2929 		}
2930 
2931 		next_cidx = cidx + 1;
2932 		if (__predict_false(next_cidx == r->size))
2933 			next_cidx = 0;
2934 
2935 		wr = (void *)&eq->desc[eq->pidx];
2936 		if (sc->flags & IS_VF) {
2937 			total++;
2938 			remaining--;
2939 			ETHER_BPF_MTAP(ifp, m0);
2940 			n = write_txpkt_vm_wr(sc, txq, (void *)wr, m0,
2941 			    available);
2942 		} else if (remaining > 1 &&
2943 		    try_txpkts(m0, r->items[next_cidx], &txp, available) == 0) {
2944 
2945 			/* pkts at cidx, next_cidx should both be in txp. */
2946 			MPASS(txp.npkt == 2);
2947 			tail = r->items[next_cidx];
2948 			MPASS(tail->m_nextpkt == NULL);
2949 			ETHER_BPF_MTAP(ifp, m0);
2950 			ETHER_BPF_MTAP(ifp, tail);
2951 			m0->m_nextpkt = tail;
2952 
2953 			if (__predict_false(++next_cidx == r->size))
2954 				next_cidx = 0;
2955 
2956 			while (next_cidx != pidx) {
2957 				if (add_to_txpkts(r->items[next_cidx], &txp,
2958 				    available) != 0)
2959 					break;
2960 				tail->m_nextpkt = r->items[next_cidx];
2961 				tail = tail->m_nextpkt;
2962 				ETHER_BPF_MTAP(ifp, tail);
2963 				if (__predict_false(++next_cidx == r->size))
2964 					next_cidx = 0;
2965 			}
2966 
2967 			n = write_txpkts_wr(txq, wr, m0, &txp, available);
2968 			total += txp.npkt;
2969 			remaining -= txp.npkt;
2970 		} else if (mbuf_cflags(m0) & MC_RAW_WR) {
2971 			total++;
2972 			remaining--;
2973 			n = write_raw_wr(txq, (void *)wr, m0, available);
2974 		} else {
2975 			total++;
2976 			remaining--;
2977 			ETHER_BPF_MTAP(ifp, m0);
2978 			n = write_txpkt_wr(txq, (void *)wr, m0, available);
2979 		}
2980 		MPASS(n >= 1 && n <= available && n <= SGE_MAX_WR_NDESC);
2981 
2982 		available -= n;
2983 		dbdiff += n;
2984 		IDXINCR(eq->pidx, n, eq->sidx);
2985 
2986 		if (wr_can_update_eq(wr)) {
2987 			if (total_available_tx_desc(eq) < eq->sidx / 4 &&
2988 			    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2989 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2990 				    F_FW_WR_EQUEQ);
2991 				eq->equeqidx = eq->pidx;
2992 			} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >=
2993 			    32) {
2994 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
2995 				eq->equeqidx = eq->pidx;
2996 			}
2997 		}
2998 
2999 		if (dbdiff >= 16 && remaining >= 4) {
3000 			ring_eq_db(sc, eq, dbdiff);
3001 			available += reclaim_tx_descs(txq, 4 * dbdiff);
3002 			dbdiff = 0;
3003 		}
3004 
3005 		cidx = next_cidx;
3006 	}
3007 	if (dbdiff != 0) {
3008 		ring_eq_db(sc, eq, dbdiff);
3009 		reclaim_tx_descs(txq, 32);
3010 	}
3011 done:
3012 	TXQ_UNLOCK(txq);
3013 
3014 	return (total);
3015 }
3016 
3017 static inline void
3018 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
3019     int qsize)
3020 {
3021 
3022 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
3023 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
3024 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
3025 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
3026 
3027 	iq->flags = 0;
3028 	iq->adapter = sc;
3029 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
3030 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
3031 	if (pktc_idx >= 0) {
3032 		iq->intr_params |= F_QINTR_CNT_EN;
3033 		iq->intr_pktc_idx = pktc_idx;
3034 	}
3035 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
3036 	iq->sidx = iq->qsize - sc->params.sge.spg_len / IQ_ESIZE;
3037 }
3038 
3039 static inline void
3040 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name)
3041 {
3042 
3043 	fl->qsize = qsize;
3044 	fl->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
3045 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
3046 	if (sc->flags & BUF_PACKING_OK &&
3047 	    ((!is_t4(sc) && buffer_packing) ||	/* T5+: enabled unless 0 */
3048 	    (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */
3049 		fl->flags |= FL_BUF_PACKING;
3050 	find_best_refill_source(sc, fl, maxp);
3051 	find_safe_refill_source(sc, fl);
3052 }
3053 
3054 static inline void
3055 init_eq(struct adapter *sc, struct sge_eq *eq, int eqtype, int qsize,
3056     uint8_t tx_chan, uint16_t iqid, char *name)
3057 {
3058 	KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
3059 
3060 	eq->flags = eqtype & EQ_TYPEMASK;
3061 	eq->tx_chan = tx_chan;
3062 	eq->iqid = iqid;
3063 	eq->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
3064 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
3065 }
3066 
3067 static int
3068 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
3069     bus_dmamap_t *map, bus_addr_t *pa, void **va)
3070 {
3071 	int rc;
3072 
3073 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
3074 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
3075 	if (rc != 0) {
3076 		device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
3077 		goto done;
3078 	}
3079 
3080 	rc = bus_dmamem_alloc(*tag, va,
3081 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
3082 	if (rc != 0) {
3083 		device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
3084 		goto done;
3085 	}
3086 
3087 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
3088 	if (rc != 0) {
3089 		device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
3090 		goto done;
3091 	}
3092 done:
3093 	if (rc)
3094 		free_ring(sc, *tag, *map, *pa, *va);
3095 
3096 	return (rc);
3097 }
3098 
3099 static int
3100 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
3101     bus_addr_t pa, void *va)
3102 {
3103 	if (pa)
3104 		bus_dmamap_unload(tag, map);
3105 	if (va)
3106 		bus_dmamem_free(tag, va, map);
3107 	if (tag)
3108 		bus_dma_tag_destroy(tag);
3109 
3110 	return (0);
3111 }
3112 
3113 /*
3114  * Allocates the ring for an ingress queue and an optional freelist.  If the
3115  * freelist is specified it will be allocated and then associated with the
3116  * ingress queue.
3117  *
3118  * Returns errno on failure.  Resources allocated up to that point may still be
3119  * allocated.  Caller is responsible for cleanup in case this function fails.
3120  *
3121  * If the ingress queue will take interrupts directly then the intr_idx
3122  * specifies the vector, starting from 0.  -1 means the interrupts for this
3123  * queue should be forwarded to the fwq.
3124  */
3125 static int
3126 alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl,
3127     int intr_idx, int cong)
3128 {
3129 	int rc, i, cntxt_id;
3130 	size_t len;
3131 	struct fw_iq_cmd c;
3132 	struct port_info *pi = vi->pi;
3133 	struct adapter *sc = iq->adapter;
3134 	struct sge_params *sp = &sc->params.sge;
3135 	__be32 v = 0;
3136 
3137 	len = iq->qsize * IQ_ESIZE;
3138 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
3139 	    (void **)&iq->desc);
3140 	if (rc != 0)
3141 		return (rc);
3142 
3143 	bzero(&c, sizeof(c));
3144 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
3145 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
3146 	    V_FW_IQ_CMD_VFN(0));
3147 
3148 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
3149 	    FW_LEN16(c));
3150 
3151 	/* Special handling for firmware event queue */
3152 	if (iq == &sc->sge.fwq)
3153 		v |= F_FW_IQ_CMD_IQASYNCH;
3154 
3155 	if (intr_idx < 0) {
3156 		/* Forwarded interrupts, all headed to fwq */
3157 		v |= F_FW_IQ_CMD_IQANDST;
3158 		v |= V_FW_IQ_CMD_IQANDSTINDEX(sc->sge.fwq.cntxt_id);
3159 	} else {
3160 		KASSERT(intr_idx < sc->intr_count,
3161 		    ("%s: invalid direct intr_idx %d", __func__, intr_idx));
3162 		v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
3163 	}
3164 
3165 	c.type_to_iqandstindex = htobe32(v |
3166 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
3167 	    V_FW_IQ_CMD_VIID(vi->viid) |
3168 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
3169 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
3170 	    F_FW_IQ_CMD_IQGTSMODE |
3171 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
3172 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
3173 	c.iqsize = htobe16(iq->qsize);
3174 	c.iqaddr = htobe64(iq->ba);
3175 	if (cong >= 0)
3176 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
3177 
3178 	if (fl) {
3179 		mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
3180 
3181 		len = fl->qsize * EQ_ESIZE;
3182 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
3183 		    &fl->ba, (void **)&fl->desc);
3184 		if (rc)
3185 			return (rc);
3186 
3187 		/* Allocate space for one software descriptor per buffer. */
3188 		rc = alloc_fl_sdesc(fl);
3189 		if (rc != 0) {
3190 			device_printf(sc->dev,
3191 			    "failed to setup fl software descriptors: %d\n",
3192 			    rc);
3193 			return (rc);
3194 		}
3195 
3196 		if (fl->flags & FL_BUF_PACKING) {
3197 			fl->lowat = roundup2(sp->fl_starve_threshold2, 8);
3198 			fl->buf_boundary = sp->pack_boundary;
3199 		} else {
3200 			fl->lowat = roundup2(sp->fl_starve_threshold, 8);
3201 			fl->buf_boundary = 16;
3202 		}
3203 		if (fl_pad && fl->buf_boundary < sp->pad_boundary)
3204 			fl->buf_boundary = sp->pad_boundary;
3205 
3206 		c.iqns_to_fl0congen |=
3207 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
3208 			F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
3209 			(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
3210 			(fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
3211 			    0));
3212 		if (cong >= 0) {
3213 			c.iqns_to_fl0congen |=
3214 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
3215 				    F_FW_IQ_CMD_FL0CONGCIF |
3216 				    F_FW_IQ_CMD_FL0CONGEN);
3217 		}
3218 		c.fl0dcaen_to_fl0cidxfthresh =
3219 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
3220 			X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B) |
3221 			V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
3222 			X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
3223 		c.fl0size = htobe16(fl->qsize);
3224 		c.fl0addr = htobe64(fl->ba);
3225 	}
3226 
3227 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3228 	if (rc != 0) {
3229 		device_printf(sc->dev,
3230 		    "failed to create ingress queue: %d\n", rc);
3231 		return (rc);
3232 	}
3233 
3234 	iq->cidx = 0;
3235 	iq->gen = F_RSPD_GEN;
3236 	iq->intr_next = iq->intr_params;
3237 	iq->cntxt_id = be16toh(c.iqid);
3238 	iq->abs_id = be16toh(c.physiqid);
3239 	iq->flags |= IQ_ALLOCATED;
3240 
3241 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
3242 	if (cntxt_id >= sc->sge.niq) {
3243 		panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
3244 		    cntxt_id, sc->sge.niq - 1);
3245 	}
3246 	sc->sge.iqmap[cntxt_id] = iq;
3247 
3248 	if (fl) {
3249 		u_int qid;
3250 
3251 		iq->flags |= IQ_HAS_FL;
3252 		fl->cntxt_id = be16toh(c.fl0id);
3253 		fl->pidx = fl->cidx = 0;
3254 
3255 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
3256 		if (cntxt_id >= sc->sge.neq) {
3257 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
3258 			    __func__, cntxt_id, sc->sge.neq - 1);
3259 		}
3260 		sc->sge.eqmap[cntxt_id] = (void *)fl;
3261 
3262 		qid = fl->cntxt_id;
3263 		if (isset(&sc->doorbells, DOORBELL_UDB)) {
3264 			uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3265 			uint32_t mask = (1 << s_qpp) - 1;
3266 			volatile uint8_t *udb;
3267 
3268 			udb = sc->udbs_base + UDBS_DB_OFFSET;
3269 			udb += (qid >> s_qpp) << PAGE_SHIFT;
3270 			qid &= mask;
3271 			if (qid < PAGE_SIZE / UDBS_SEG_SIZE) {
3272 				udb += qid << UDBS_SEG_SHIFT;
3273 				qid = 0;
3274 			}
3275 			fl->udb = (volatile void *)udb;
3276 		}
3277 		fl->dbval = V_QID(qid) | sc->chip_params->sge_fl_db;
3278 
3279 		FL_LOCK(fl);
3280 		/* Enough to make sure the SGE doesn't think it's starved */
3281 		refill_fl(sc, fl, fl->lowat);
3282 		FL_UNLOCK(fl);
3283 	}
3284 
3285 	if (chip_id(sc) >= CHELSIO_T5 && !(sc->flags & IS_VF) && cong >= 0) {
3286 		uint32_t param, val;
3287 
3288 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
3289 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
3290 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
3291 		if (cong == 0)
3292 			val = 1 << 19;
3293 		else {
3294 			val = 2 << 19;
3295 			for (i = 0; i < 4; i++) {
3296 				if (cong & (1 << i))
3297 					val |= 1 << (i << 2);
3298 			}
3299 		}
3300 
3301 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
3302 		if (rc != 0) {
3303 			/* report error but carry on */
3304 			device_printf(sc->dev,
3305 			    "failed to set congestion manager context for "
3306 			    "ingress queue %d: %d\n", iq->cntxt_id, rc);
3307 		}
3308 	}
3309 
3310 	/* Enable IQ interrupts */
3311 	atomic_store_rel_int(&iq->state, IQS_IDLE);
3312 	t4_write_reg(sc, sc->sge_gts_reg, V_SEINTARM(iq->intr_params) |
3313 	    V_INGRESSQID(iq->cntxt_id));
3314 
3315 	return (0);
3316 }
3317 
3318 static int
3319 free_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl)
3320 {
3321 	int rc;
3322 	struct adapter *sc = iq->adapter;
3323 	device_t dev;
3324 
3325 	if (sc == NULL)
3326 		return (0);	/* nothing to do */
3327 
3328 	dev = vi ? vi->dev : sc->dev;
3329 
3330 	if (iq->flags & IQ_ALLOCATED) {
3331 		rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
3332 		    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
3333 		    fl ? fl->cntxt_id : 0xffff, 0xffff);
3334 		if (rc != 0) {
3335 			device_printf(dev,
3336 			    "failed to free queue %p: %d\n", iq, rc);
3337 			return (rc);
3338 		}
3339 		iq->flags &= ~IQ_ALLOCATED;
3340 	}
3341 
3342 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
3343 
3344 	bzero(iq, sizeof(*iq));
3345 
3346 	if (fl) {
3347 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
3348 		    fl->desc);
3349 
3350 		if (fl->sdesc)
3351 			free_fl_sdesc(sc, fl);
3352 
3353 		if (mtx_initialized(&fl->fl_lock))
3354 			mtx_destroy(&fl->fl_lock);
3355 
3356 		bzero(fl, sizeof(*fl));
3357 	}
3358 
3359 	return (0);
3360 }
3361 
3362 static void
3363 add_iq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
3364     struct sge_iq *iq)
3365 {
3366 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3367 
3368 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD, &iq->ba,
3369 	    "bus address of descriptor ring");
3370 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3371 	    iq->qsize * IQ_ESIZE, "descriptor ring size in bytes");
3372 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3373 	    CTLTYPE_INT | CTLFLAG_RD, &iq->abs_id, 0, sysctl_uint16, "I",
3374 	    "absolute id of the queue");
3375 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3376 	    CTLTYPE_INT | CTLFLAG_RD, &iq->cntxt_id, 0, sysctl_uint16, "I",
3377 	    "SGE context id of the queue");
3378 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3379 	    CTLTYPE_INT | CTLFLAG_RD, &iq->cidx, 0, sysctl_uint16, "I",
3380 	    "consumer index");
3381 }
3382 
3383 static void
3384 add_fl_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
3385     struct sysctl_oid *oid, struct sge_fl *fl)
3386 {
3387 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3388 
3389 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3390 	    "freelist");
3391 	children = SYSCTL_CHILDREN(oid);
3392 
3393 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3394 	    &fl->ba, "bus address of descriptor ring");
3395 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3396 	    fl->sidx * EQ_ESIZE + sc->params.sge.spg_len,
3397 	    "desc ring size in bytes");
3398 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3399 	    CTLTYPE_INT | CTLFLAG_RD, &fl->cntxt_id, 0, sysctl_uint16, "I",
3400 	    "SGE context id of the freelist");
3401 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL,
3402 	    fl_pad ? 1 : 0, "padding enabled");
3403 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL,
3404 	    fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled");
3405 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
3406 	    0, "consumer index");
3407 	if (fl->flags & FL_BUF_PACKING) {
3408 		SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
3409 		    CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
3410 	}
3411 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
3412 	    0, "producer index");
3413 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_allocated",
3414 	    CTLFLAG_RD, &fl->mbuf_allocated, "# of mbuf allocated");
3415 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_inlined",
3416 	    CTLFLAG_RD, &fl->mbuf_inlined, "# of mbuf inlined in clusters");
3417 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
3418 	    CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
3419 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
3420 	    CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
3421 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
3422 	    CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
3423 }
3424 
3425 static int
3426 alloc_fwq(struct adapter *sc)
3427 {
3428 	int rc, intr_idx;
3429 	struct sge_iq *fwq = &sc->sge.fwq;
3430 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
3431 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3432 
3433 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE);
3434 	if (sc->flags & IS_VF)
3435 		intr_idx = 0;
3436 	else
3437 		intr_idx = sc->intr_count > 1 ? 1 : 0;
3438 	rc = alloc_iq_fl(&sc->port[0]->vi[0], fwq, NULL, intr_idx, -1);
3439 	if (rc != 0) {
3440 		device_printf(sc->dev,
3441 		    "failed to create firmware event queue: %d\n", rc);
3442 		return (rc);
3443 	}
3444 
3445 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD,
3446 	    NULL, "firmware event queue");
3447 	add_iq_sysctls(&sc->ctx, oid, fwq);
3448 
3449 	return (0);
3450 }
3451 
3452 static int
3453 free_fwq(struct adapter *sc)
3454 {
3455 	return free_iq_fl(NULL, &sc->sge.fwq, NULL);
3456 }
3457 
3458 static int
3459 alloc_ctrlq(struct adapter *sc, struct sge_wrq *ctrlq, int idx,
3460     struct sysctl_oid *oid)
3461 {
3462 	int rc;
3463 	char name[16];
3464 	struct sysctl_oid_list *children;
3465 
3466 	snprintf(name, sizeof(name), "%s ctrlq%d", device_get_nameunit(sc->dev),
3467 	    idx);
3468 	init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[idx]->tx_chan,
3469 	    sc->sge.fwq.cntxt_id, name);
3470 
3471 	children = SYSCTL_CHILDREN(oid);
3472 	snprintf(name, sizeof(name), "%d", idx);
3473 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3474 	    NULL, "ctrl queue");
3475 	rc = alloc_wrq(sc, NULL, ctrlq, oid);
3476 
3477 	return (rc);
3478 }
3479 
3480 int
3481 tnl_cong(struct port_info *pi, int drop)
3482 {
3483 
3484 	if (drop == -1)
3485 		return (-1);
3486 	else if (drop == 1)
3487 		return (0);
3488 	else
3489 		return (pi->rx_e_chan_map);
3490 }
3491 
3492 static int
3493 alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int intr_idx, int idx,
3494     struct sysctl_oid *oid)
3495 {
3496 	int rc;
3497 	struct adapter *sc = vi->pi->adapter;
3498 	struct sysctl_oid_list *children;
3499 	char name[16];
3500 
3501 	rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, intr_idx,
3502 	    tnl_cong(vi->pi, cong_drop));
3503 	if (rc != 0)
3504 		return (rc);
3505 
3506 	if (idx == 0)
3507 		sc->sge.iq_base = rxq->iq.abs_id - rxq->iq.cntxt_id;
3508 	else
3509 		KASSERT(rxq->iq.cntxt_id + sc->sge.iq_base == rxq->iq.abs_id,
3510 		    ("iq_base mismatch"));
3511 	KASSERT(sc->sge.iq_base == 0 || sc->flags & IS_VF,
3512 	    ("PF with non-zero iq_base"));
3513 
3514 	/*
3515 	 * The freelist is just barely above the starvation threshold right now,
3516 	 * fill it up a bit more.
3517 	 */
3518 	FL_LOCK(&rxq->fl);
3519 	refill_fl(sc, &rxq->fl, 128);
3520 	FL_UNLOCK(&rxq->fl);
3521 
3522 #if defined(INET) || defined(INET6)
3523 	rc = tcp_lro_init_args(&rxq->lro, vi->ifp, lro_entries, lro_mbufs);
3524 	if (rc != 0)
3525 		return (rc);
3526 	MPASS(rxq->lro.ifp == vi->ifp);	/* also indicates LRO init'ed */
3527 
3528 	if (vi->ifp->if_capenable & IFCAP_LRO)
3529 		rxq->iq.flags |= IQ_LRO_ENABLED;
3530 #endif
3531 	if (vi->ifp->if_capenable & IFCAP_HWRXTSTMP)
3532 		rxq->iq.flags |= IQ_RX_TIMESTAMP;
3533 	rxq->ifp = vi->ifp;
3534 
3535 	children = SYSCTL_CHILDREN(oid);
3536 
3537 	snprintf(name, sizeof(name), "%d", idx);
3538 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3539 	    NULL, "rx queue");
3540 	children = SYSCTL_CHILDREN(oid);
3541 
3542 	add_iq_sysctls(&vi->ctx, oid, &rxq->iq);
3543 #if defined(INET) || defined(INET6)
3544 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
3545 	    &rxq->lro.lro_queued, 0, NULL);
3546 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
3547 	    &rxq->lro.lro_flushed, 0, NULL);
3548 #endif
3549 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
3550 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
3551 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_extraction",
3552 	    CTLFLAG_RD, &rxq->vlan_extraction,
3553 	    "# of times hardware extracted 802.1Q tag");
3554 
3555 	add_fl_sysctls(sc, &vi->ctx, oid, &rxq->fl);
3556 
3557 	return (rc);
3558 }
3559 
3560 static int
3561 free_rxq(struct vi_info *vi, struct sge_rxq *rxq)
3562 {
3563 	int rc;
3564 
3565 #if defined(INET) || defined(INET6)
3566 	if (rxq->lro.ifp) {
3567 		tcp_lro_free(&rxq->lro);
3568 		rxq->lro.ifp = NULL;
3569 	}
3570 #endif
3571 
3572 	rc = free_iq_fl(vi, &rxq->iq, &rxq->fl);
3573 	if (rc == 0)
3574 		bzero(rxq, sizeof(*rxq));
3575 
3576 	return (rc);
3577 }
3578 
3579 #ifdef TCP_OFFLOAD
3580 static int
3581 alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq,
3582     int intr_idx, int idx, struct sysctl_oid *oid)
3583 {
3584 	struct port_info *pi = vi->pi;
3585 	int rc;
3586 	struct sysctl_oid_list *children;
3587 	char name[16];
3588 
3589 	rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx, 0);
3590 	if (rc != 0)
3591 		return (rc);
3592 
3593 	children = SYSCTL_CHILDREN(oid);
3594 
3595 	snprintf(name, sizeof(name), "%d", idx);
3596 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3597 	    NULL, "rx queue");
3598 	add_iq_sysctls(&vi->ctx, oid, &ofld_rxq->iq);
3599 	add_fl_sysctls(pi->adapter, &vi->ctx, oid, &ofld_rxq->fl);
3600 
3601 	return (rc);
3602 }
3603 
3604 static int
3605 free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq)
3606 {
3607 	int rc;
3608 
3609 	rc = free_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl);
3610 	if (rc == 0)
3611 		bzero(ofld_rxq, sizeof(*ofld_rxq));
3612 
3613 	return (rc);
3614 }
3615 #endif
3616 
3617 #ifdef DEV_NETMAP
3618 static int
3619 alloc_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int intr_idx,
3620     int idx, struct sysctl_oid *oid)
3621 {
3622 	int rc;
3623 	struct sysctl_oid_list *children;
3624 	struct sysctl_ctx_list *ctx;
3625 	char name[16];
3626 	size_t len;
3627 	struct adapter *sc = vi->pi->adapter;
3628 	struct netmap_adapter *na = NA(vi->ifp);
3629 
3630 	MPASS(na != NULL);
3631 
3632 	len = vi->qsize_rxq * IQ_ESIZE;
3633 	rc = alloc_ring(sc, len, &nm_rxq->iq_desc_tag, &nm_rxq->iq_desc_map,
3634 	    &nm_rxq->iq_ba, (void **)&nm_rxq->iq_desc);
3635 	if (rc != 0)
3636 		return (rc);
3637 
3638 	len = na->num_rx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3639 	rc = alloc_ring(sc, len, &nm_rxq->fl_desc_tag, &nm_rxq->fl_desc_map,
3640 	    &nm_rxq->fl_ba, (void **)&nm_rxq->fl_desc);
3641 	if (rc != 0)
3642 		return (rc);
3643 
3644 	nm_rxq->vi = vi;
3645 	nm_rxq->nid = idx;
3646 	nm_rxq->iq_cidx = 0;
3647 	nm_rxq->iq_sidx = vi->qsize_rxq - sc->params.sge.spg_len / IQ_ESIZE;
3648 	nm_rxq->iq_gen = F_RSPD_GEN;
3649 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
3650 	nm_rxq->fl_sidx = na->num_rx_desc;
3651 	nm_rxq->intr_idx = intr_idx;
3652 	nm_rxq->iq_cntxt_id = INVALID_NM_RXQ_CNTXT_ID;
3653 
3654 	ctx = &vi->ctx;
3655 	children = SYSCTL_CHILDREN(oid);
3656 
3657 	snprintf(name, sizeof(name), "%d", idx);
3658 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL,
3659 	    "rx queue");
3660 	children = SYSCTL_CHILDREN(oid);
3661 
3662 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3663 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_abs_id, 0, sysctl_uint16,
3664 	    "I", "absolute id of the queue");
3665 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3666 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cntxt_id, 0, sysctl_uint16,
3667 	    "I", "SGE context id of the queue");
3668 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3669 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cidx, 0, sysctl_uint16, "I",
3670 	    "consumer index");
3671 
3672 	children = SYSCTL_CHILDREN(oid);
3673 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3674 	    "freelist");
3675 	children = SYSCTL_CHILDREN(oid);
3676 
3677 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3678 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->fl_cntxt_id, 0, sysctl_uint16,
3679 	    "I", "SGE context id of the freelist");
3680 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
3681 	    &nm_rxq->fl_cidx, 0, "consumer index");
3682 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
3683 	    &nm_rxq->fl_pidx, 0, "producer index");
3684 
3685 	return (rc);
3686 }
3687 
3688 
3689 static int
3690 free_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
3691 {
3692 	struct adapter *sc = vi->pi->adapter;
3693 
3694 	if (vi->flags & VI_INIT_DONE)
3695 		MPASS(nm_rxq->iq_cntxt_id == INVALID_NM_RXQ_CNTXT_ID);
3696 	else
3697 		MPASS(nm_rxq->iq_cntxt_id == 0);
3698 
3699 	free_ring(sc, nm_rxq->iq_desc_tag, nm_rxq->iq_desc_map, nm_rxq->iq_ba,
3700 	    nm_rxq->iq_desc);
3701 	free_ring(sc, nm_rxq->fl_desc_tag, nm_rxq->fl_desc_map, nm_rxq->fl_ba,
3702 	    nm_rxq->fl_desc);
3703 
3704 	return (0);
3705 }
3706 
3707 static int
3708 alloc_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq, int iqidx, int idx,
3709     struct sysctl_oid *oid)
3710 {
3711 	int rc;
3712 	size_t len;
3713 	struct port_info *pi = vi->pi;
3714 	struct adapter *sc = pi->adapter;
3715 	struct netmap_adapter *na = NA(vi->ifp);
3716 	char name[16];
3717 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3718 
3719 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3720 	rc = alloc_ring(sc, len, &nm_txq->desc_tag, &nm_txq->desc_map,
3721 	    &nm_txq->ba, (void **)&nm_txq->desc);
3722 	if (rc)
3723 		return (rc);
3724 
3725 	nm_txq->pidx = nm_txq->cidx = 0;
3726 	nm_txq->sidx = na->num_tx_desc;
3727 	nm_txq->nid = idx;
3728 	nm_txq->iqidx = iqidx;
3729 	nm_txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3730 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(sc->pf) |
3731 	    V_TXPKT_VF(vi->vin) | V_TXPKT_VF_VLD(vi->vfvld));
3732 	nm_txq->cntxt_id = INVALID_NM_TXQ_CNTXT_ID;
3733 
3734 	snprintf(name, sizeof(name), "%d", idx);
3735 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3736 	    NULL, "netmap tx queue");
3737 	children = SYSCTL_CHILDREN(oid);
3738 
3739 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3740 	    &nm_txq->cntxt_id, 0, "SGE context id of the queue");
3741 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3742 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->cidx, 0, sysctl_uint16, "I",
3743 	    "consumer index");
3744 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3745 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->pidx, 0, sysctl_uint16, "I",
3746 	    "producer index");
3747 
3748 	return (rc);
3749 }
3750 
3751 static int
3752 free_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
3753 {
3754 	struct adapter *sc = vi->pi->adapter;
3755 
3756 	if (vi->flags & VI_INIT_DONE)
3757 		MPASS(nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID);
3758 	else
3759 		MPASS(nm_txq->cntxt_id == 0);
3760 
3761 	free_ring(sc, nm_txq->desc_tag, nm_txq->desc_map, nm_txq->ba,
3762 	    nm_txq->desc);
3763 
3764 	return (0);
3765 }
3766 #endif
3767 
3768 /*
3769  * Returns a reasonable automatic cidx flush threshold for a given queue size.
3770  */
3771 static u_int
3772 qsize_to_fthresh(int qsize)
3773 {
3774 	u_int fthresh;
3775 
3776 	while (!powerof2(qsize))
3777 		qsize++;
3778 	fthresh = ilog2(qsize);
3779 	if (fthresh > X_CIDXFLUSHTHRESH_128)
3780 		fthresh = X_CIDXFLUSHTHRESH_128;
3781 
3782 	return (fthresh);
3783 }
3784 
3785 static int
3786 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
3787 {
3788 	int rc, cntxt_id;
3789 	struct fw_eq_ctrl_cmd c;
3790 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3791 
3792 	bzero(&c, sizeof(c));
3793 
3794 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
3795 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
3796 	    V_FW_EQ_CTRL_CMD_VFN(0));
3797 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
3798 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
3799 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid));
3800 	c.physeqid_pkd = htobe32(0);
3801 	c.fetchszm_to_iqid =
3802 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
3803 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
3804 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
3805 	c.dcaen_to_eqsize =
3806 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3807 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3808 		V_FW_EQ_CTRL_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
3809 		V_FW_EQ_CTRL_CMD_EQSIZE(qsize));
3810 	c.eqaddr = htobe64(eq->ba);
3811 
3812 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3813 	if (rc != 0) {
3814 		device_printf(sc->dev,
3815 		    "failed to create control queue %d: %d\n", eq->tx_chan, rc);
3816 		return (rc);
3817 	}
3818 	eq->flags |= EQ_ALLOCATED;
3819 
3820 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
3821 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3822 	if (cntxt_id >= sc->sge.neq)
3823 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3824 		cntxt_id, sc->sge.neq - 1);
3825 	sc->sge.eqmap[cntxt_id] = eq;
3826 
3827 	return (rc);
3828 }
3829 
3830 static int
3831 eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3832 {
3833 	int rc, cntxt_id;
3834 	struct fw_eq_eth_cmd c;
3835 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3836 
3837 	bzero(&c, sizeof(c));
3838 
3839 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
3840 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
3841 	    V_FW_EQ_ETH_CMD_VFN(0));
3842 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
3843 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
3844 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
3845 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
3846 	c.fetchszm_to_iqid =
3847 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3848 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
3849 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
3850 	c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3851 	    V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3852 	    V_FW_EQ_ETH_CMD_EQSIZE(qsize));
3853 	c.eqaddr = htobe64(eq->ba);
3854 
3855 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3856 	if (rc != 0) {
3857 		device_printf(vi->dev,
3858 		    "failed to create Ethernet egress queue: %d\n", rc);
3859 		return (rc);
3860 	}
3861 	eq->flags |= EQ_ALLOCATED;
3862 
3863 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
3864 	eq->abs_id = G_FW_EQ_ETH_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
3865 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3866 	if (cntxt_id >= sc->sge.neq)
3867 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3868 		cntxt_id, sc->sge.neq - 1);
3869 	sc->sge.eqmap[cntxt_id] = eq;
3870 
3871 	return (rc);
3872 }
3873 
3874 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3875 static int
3876 ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3877 {
3878 	int rc, cntxt_id;
3879 	struct fw_eq_ofld_cmd c;
3880 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3881 
3882 	bzero(&c, sizeof(c));
3883 
3884 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
3885 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
3886 	    V_FW_EQ_OFLD_CMD_VFN(0));
3887 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
3888 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
3889 	c.fetchszm_to_iqid =
3890 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
3891 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
3892 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
3893 	c.dcaen_to_eqsize =
3894 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3895 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3896 		V_FW_EQ_OFLD_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
3897 		V_FW_EQ_OFLD_CMD_EQSIZE(qsize));
3898 	c.eqaddr = htobe64(eq->ba);
3899 
3900 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3901 	if (rc != 0) {
3902 		device_printf(vi->dev,
3903 		    "failed to create egress queue for TCP offload: %d\n", rc);
3904 		return (rc);
3905 	}
3906 	eq->flags |= EQ_ALLOCATED;
3907 
3908 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
3909 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3910 	if (cntxt_id >= sc->sge.neq)
3911 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3912 		cntxt_id, sc->sge.neq - 1);
3913 	sc->sge.eqmap[cntxt_id] = eq;
3914 
3915 	return (rc);
3916 }
3917 #endif
3918 
3919 static int
3920 alloc_eq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3921 {
3922 	int rc, qsize;
3923 	size_t len;
3924 
3925 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
3926 
3927 	qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3928 	len = qsize * EQ_ESIZE;
3929 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
3930 	    &eq->ba, (void **)&eq->desc);
3931 	if (rc)
3932 		return (rc);
3933 
3934 	eq->pidx = eq->cidx = eq->dbidx = 0;
3935 	/* Note that equeqidx is not used with sge_wrq (OFLD/CTRL) queues. */
3936 	eq->equeqidx = 0;
3937 	eq->doorbells = sc->doorbells;
3938 
3939 	switch (eq->flags & EQ_TYPEMASK) {
3940 	case EQ_CTRL:
3941 		rc = ctrl_eq_alloc(sc, eq);
3942 		break;
3943 
3944 	case EQ_ETH:
3945 		rc = eth_eq_alloc(sc, vi, eq);
3946 		break;
3947 
3948 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3949 	case EQ_OFLD:
3950 		rc = ofld_eq_alloc(sc, vi, eq);
3951 		break;
3952 #endif
3953 
3954 	default:
3955 		panic("%s: invalid eq type %d.", __func__,
3956 		    eq->flags & EQ_TYPEMASK);
3957 	}
3958 	if (rc != 0) {
3959 		device_printf(sc->dev,
3960 		    "failed to allocate egress queue(%d): %d\n",
3961 		    eq->flags & EQ_TYPEMASK, rc);
3962 	}
3963 
3964 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
3965 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
3966 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
3967 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3968 		uint32_t mask = (1 << s_qpp) - 1;
3969 		volatile uint8_t *udb;
3970 
3971 		udb = sc->udbs_base + UDBS_DB_OFFSET;
3972 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
3973 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
3974 		if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
3975 	    		clrbit(&eq->doorbells, DOORBELL_WCWR);
3976 		else {
3977 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
3978 			eq->udb_qid = 0;
3979 		}
3980 		eq->udb = (volatile void *)udb;
3981 	}
3982 
3983 	return (rc);
3984 }
3985 
3986 static int
3987 free_eq(struct adapter *sc, struct sge_eq *eq)
3988 {
3989 	int rc;
3990 
3991 	if (eq->flags & EQ_ALLOCATED) {
3992 		switch (eq->flags & EQ_TYPEMASK) {
3993 		case EQ_CTRL:
3994 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
3995 			    eq->cntxt_id);
3996 			break;
3997 
3998 		case EQ_ETH:
3999 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
4000 			    eq->cntxt_id);
4001 			break;
4002 
4003 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
4004 		case EQ_OFLD:
4005 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
4006 			    eq->cntxt_id);
4007 			break;
4008 #endif
4009 
4010 		default:
4011 			panic("%s: invalid eq type %d.", __func__,
4012 			    eq->flags & EQ_TYPEMASK);
4013 		}
4014 		if (rc != 0) {
4015 			device_printf(sc->dev,
4016 			    "failed to free egress queue (%d): %d\n",
4017 			    eq->flags & EQ_TYPEMASK, rc);
4018 			return (rc);
4019 		}
4020 		eq->flags &= ~EQ_ALLOCATED;
4021 	}
4022 
4023 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
4024 
4025 	if (mtx_initialized(&eq->eq_lock))
4026 		mtx_destroy(&eq->eq_lock);
4027 
4028 	bzero(eq, sizeof(*eq));
4029 	return (0);
4030 }
4031 
4032 static int
4033 alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq,
4034     struct sysctl_oid *oid)
4035 {
4036 	int rc;
4037 	struct sysctl_ctx_list *ctx = vi ? &vi->ctx : &sc->ctx;
4038 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
4039 
4040 	rc = alloc_eq(sc, vi, &wrq->eq);
4041 	if (rc)
4042 		return (rc);
4043 
4044 	wrq->adapter = sc;
4045 	TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq);
4046 	TAILQ_INIT(&wrq->incomplete_wrs);
4047 	STAILQ_INIT(&wrq->wr_list);
4048 	wrq->nwr_pending = 0;
4049 	wrq->ndesc_needed = 0;
4050 
4051 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
4052 	    &wrq->eq.ba, "bus address of descriptor ring");
4053 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
4054 	    wrq->eq.sidx * EQ_ESIZE + sc->params.sge.spg_len,
4055 	    "desc ring size in bytes");
4056 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
4057 	    &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
4058 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
4059 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I",
4060 	    "consumer index");
4061 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx",
4062 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I",
4063 	    "producer index");
4064 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
4065 	    wrq->eq.sidx, "status page index");
4066 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD,
4067 	    &wrq->tx_wrs_direct, "# of work requests (direct)");
4068 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD,
4069 	    &wrq->tx_wrs_copied, "# of work requests (copied)");
4070 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_sspace", CTLFLAG_RD,
4071 	    &wrq->tx_wrs_ss, "# of work requests (copied from scratch space)");
4072 
4073 	return (rc);
4074 }
4075 
4076 static int
4077 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
4078 {
4079 	int rc;
4080 
4081 	rc = free_eq(sc, &wrq->eq);
4082 	if (rc)
4083 		return (rc);
4084 
4085 	bzero(wrq, sizeof(*wrq));
4086 	return (0);
4087 }
4088 
4089 static int
4090 alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx,
4091     struct sysctl_oid *oid)
4092 {
4093 	int rc;
4094 	struct port_info *pi = vi->pi;
4095 	struct adapter *sc = pi->adapter;
4096 	struct sge_eq *eq = &txq->eq;
4097 	char name[16];
4098 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
4099 
4100 	rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx, can_resume_eth_tx,
4101 	    M_CXGBE, M_WAITOK);
4102 	if (rc != 0) {
4103 		device_printf(sc->dev, "failed to allocate mp_ring: %d\n", rc);
4104 		return (rc);
4105 	}
4106 
4107 	rc = alloc_eq(sc, vi, eq);
4108 	if (rc != 0) {
4109 		mp_ring_free(txq->r);
4110 		txq->r = NULL;
4111 		return (rc);
4112 	}
4113 
4114 	/* Can't fail after this point. */
4115 
4116 	if (idx == 0)
4117 		sc->sge.eq_base = eq->abs_id - eq->cntxt_id;
4118 	else
4119 		KASSERT(eq->cntxt_id + sc->sge.eq_base == eq->abs_id,
4120 		    ("eq_base mismatch"));
4121 	KASSERT(sc->sge.eq_base == 0 || sc->flags & IS_VF,
4122 	    ("PF with non-zero eq_base"));
4123 
4124 	TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq);
4125 	txq->ifp = vi->ifp;
4126 	txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK);
4127 	if (sc->flags & IS_VF)
4128 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
4129 		    V_TXPKT_INTF(pi->tx_chan));
4130 	else
4131 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
4132 		    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(sc->pf) |
4133 		    V_TXPKT_VF(vi->vin) | V_TXPKT_VF_VLD(vi->vfvld));
4134 	txq->tc_idx = -1;
4135 	txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE,
4136 	    M_ZERO | M_WAITOK);
4137 
4138 	snprintf(name, sizeof(name), "%d", idx);
4139 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
4140 	    NULL, "tx queue");
4141 	children = SYSCTL_CHILDREN(oid);
4142 
4143 	SYSCTL_ADD_UAUTO(&vi->ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
4144 	    &eq->ba, "bus address of descriptor ring");
4145 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
4146 	    eq->sidx * EQ_ESIZE + sc->params.sge.spg_len,
4147 	    "desc ring size in bytes");
4148 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
4149 	    &eq->abs_id, 0, "absolute id of the queue");
4150 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
4151 	    &eq->cntxt_id, 0, "SGE context id of the queue");
4152 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
4153 	    CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I",
4154 	    "consumer index");
4155 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
4156 	    CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I",
4157 	    "producer index");
4158 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
4159 	    eq->sidx, "status page index");
4160 
4161 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "tc",
4162 	    CTLTYPE_INT | CTLFLAG_RW, vi, idx, sysctl_tc, "I",
4163 	    "traffic class (-1 means none)");
4164 
4165 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
4166 	    &txq->txcsum, "# of times hardware assisted with checksum");
4167 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_insertion",
4168 	    CTLFLAG_RD, &txq->vlan_insertion,
4169 	    "# of times hardware inserted 802.1Q tag");
4170 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
4171 	    &txq->tso_wrs, "# of TSO work requests");
4172 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
4173 	    &txq->imm_wrs, "# of work requests with immediate data");
4174 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
4175 	    &txq->sgl_wrs, "# of work requests with direct SGL");
4176 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
4177 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
4178 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_wrs",
4179 	    CTLFLAG_RD, &txq->txpkts0_wrs,
4180 	    "# of txpkts (type 0) work requests");
4181 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_wrs",
4182 	    CTLFLAG_RD, &txq->txpkts1_wrs,
4183 	    "# of txpkts (type 1) work requests");
4184 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_pkts",
4185 	    CTLFLAG_RD, &txq->txpkts0_pkts,
4186 	    "# of frames tx'd using type0 txpkts work requests");
4187 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_pkts",
4188 	    CTLFLAG_RD, &txq->txpkts1_pkts,
4189 	    "# of frames tx'd using type1 txpkts work requests");
4190 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "raw_wrs", CTLFLAG_RD,
4191 	    &txq->raw_wrs, "# of raw work requests (non-packets)");
4192 
4193 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_enqueues",
4194 	    CTLFLAG_RD, &txq->r->enqueues,
4195 	    "# of enqueues to the mp_ring for this queue");
4196 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_drops",
4197 	    CTLFLAG_RD, &txq->r->drops,
4198 	    "# of drops in the mp_ring for this queue");
4199 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_starts",
4200 	    CTLFLAG_RD, &txq->r->starts,
4201 	    "# of normal consumer starts in the mp_ring for this queue");
4202 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_stalls",
4203 	    CTLFLAG_RD, &txq->r->stalls,
4204 	    "# of consumer stalls in the mp_ring for this queue");
4205 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_restarts",
4206 	    CTLFLAG_RD, &txq->r->restarts,
4207 	    "# of consumer restarts in the mp_ring for this queue");
4208 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_abdications",
4209 	    CTLFLAG_RD, &txq->r->abdications,
4210 	    "# of consumer abdications in the mp_ring for this queue");
4211 
4212 	return (0);
4213 }
4214 
4215 static int
4216 free_txq(struct vi_info *vi, struct sge_txq *txq)
4217 {
4218 	int rc;
4219 	struct adapter *sc = vi->pi->adapter;
4220 	struct sge_eq *eq = &txq->eq;
4221 
4222 	rc = free_eq(sc, eq);
4223 	if (rc)
4224 		return (rc);
4225 
4226 	sglist_free(txq->gl);
4227 	free(txq->sdesc, M_CXGBE);
4228 	mp_ring_free(txq->r);
4229 
4230 	bzero(txq, sizeof(*txq));
4231 	return (0);
4232 }
4233 
4234 static void
4235 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
4236 {
4237 	bus_addr_t *ba = arg;
4238 
4239 	KASSERT(nseg == 1,
4240 	    ("%s meant for single segment mappings only.", __func__));
4241 
4242 	*ba = error ? 0 : segs->ds_addr;
4243 }
4244 
4245 static inline void
4246 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
4247 {
4248 	uint32_t n, v;
4249 
4250 	n = IDXDIFF(fl->pidx / 8, fl->dbidx, fl->sidx);
4251 	MPASS(n > 0);
4252 
4253 	wmb();
4254 	v = fl->dbval | V_PIDX(n);
4255 	if (fl->udb)
4256 		*fl->udb = htole32(v);
4257 	else
4258 		t4_write_reg(sc, sc->sge_kdoorbell_reg, v);
4259 	IDXINCR(fl->dbidx, n, fl->sidx);
4260 }
4261 
4262 /*
4263  * Fills up the freelist by allocating up to 'n' buffers.  Buffers that are
4264  * recycled do not count towards this allocation budget.
4265  *
4266  * Returns non-zero to indicate that this freelist should be added to the list
4267  * of starving freelists.
4268  */
4269 static int
4270 refill_fl(struct adapter *sc, struct sge_fl *fl, int n)
4271 {
4272 	__be64 *d;
4273 	struct fl_sdesc *sd;
4274 	uintptr_t pa;
4275 	caddr_t cl;
4276 	struct cluster_layout *cll;
4277 	struct sw_zone_info *swz;
4278 	struct cluster_metadata *clm;
4279 	uint16_t max_pidx;
4280 	uint16_t hw_cidx = fl->hw_cidx;		/* stable snapshot */
4281 
4282 	FL_LOCK_ASSERT_OWNED(fl);
4283 
4284 	/*
4285 	 * We always stop at the beginning of the hardware descriptor that's just
4286 	 * before the one with the hw cidx.  This is to avoid hw pidx = hw cidx,
4287 	 * which would mean an empty freelist to the chip.
4288 	 */
4289 	max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1;
4290 	if (fl->pidx == max_pidx * 8)
4291 		return (0);
4292 
4293 	d = &fl->desc[fl->pidx];
4294 	sd = &fl->sdesc[fl->pidx];
4295 	cll = &fl->cll_def;	/* default layout */
4296 	swz = &sc->sge.sw_zone_info[cll->zidx];
4297 
4298 	while (n > 0) {
4299 
4300 		if (sd->cl != NULL) {
4301 
4302 			if (sd->nmbuf == 0) {
4303 				/*
4304 				 * Fast recycle without involving any atomics on
4305 				 * the cluster's metadata (if the cluster has
4306 				 * metadata).  This happens when all frames
4307 				 * received in the cluster were small enough to
4308 				 * fit within a single mbuf each.
4309 				 */
4310 				fl->cl_fast_recycled++;
4311 #ifdef INVARIANTS
4312 				clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
4313 				if (clm != NULL)
4314 					MPASS(clm->refcount == 1);
4315 #endif
4316 				goto recycled_fast;
4317 			}
4318 
4319 			/*
4320 			 * Cluster is guaranteed to have metadata.  Clusters
4321 			 * without metadata always take the fast recycle path
4322 			 * when they're recycled.
4323 			 */
4324 			clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
4325 			MPASS(clm != NULL);
4326 
4327 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4328 				fl->cl_recycled++;
4329 				counter_u64_add(extfree_rels, 1);
4330 				goto recycled;
4331 			}
4332 			sd->cl = NULL;	/* gave up my reference */
4333 		}
4334 		MPASS(sd->cl == NULL);
4335 alloc:
4336 		cl = uma_zalloc(swz->zone, M_NOWAIT);
4337 		if (__predict_false(cl == NULL)) {
4338 			if (cll == &fl->cll_alt || fl->cll_alt.zidx == -1 ||
4339 			    fl->cll_def.zidx == fl->cll_alt.zidx)
4340 				break;
4341 
4342 			/* fall back to the safe zone */
4343 			cll = &fl->cll_alt;
4344 			swz = &sc->sge.sw_zone_info[cll->zidx];
4345 			goto alloc;
4346 		}
4347 		fl->cl_allocated++;
4348 		n--;
4349 
4350 		pa = pmap_kextract((vm_offset_t)cl);
4351 		pa += cll->region1;
4352 		sd->cl = cl;
4353 		sd->cll = *cll;
4354 		*d = htobe64(pa | cll->hwidx);
4355 		clm = cl_metadata(sc, fl, cll, cl);
4356 		if (clm != NULL) {
4357 recycled:
4358 #ifdef INVARIANTS
4359 			clm->sd = sd;
4360 #endif
4361 			clm->refcount = 1;
4362 		}
4363 		sd->nmbuf = 0;
4364 recycled_fast:
4365 		d++;
4366 		sd++;
4367 		if (__predict_false(++fl->pidx % 8 == 0)) {
4368 			uint16_t pidx = fl->pidx / 8;
4369 
4370 			if (__predict_false(pidx == fl->sidx)) {
4371 				fl->pidx = 0;
4372 				pidx = 0;
4373 				sd = fl->sdesc;
4374 				d = fl->desc;
4375 			}
4376 			if (pidx == max_pidx)
4377 				break;
4378 
4379 			if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4)
4380 				ring_fl_db(sc, fl);
4381 		}
4382 	}
4383 
4384 	if (fl->pidx / 8 != fl->dbidx)
4385 		ring_fl_db(sc, fl);
4386 
4387 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
4388 }
4389 
4390 /*
4391  * Attempt to refill all starving freelists.
4392  */
4393 static void
4394 refill_sfl(void *arg)
4395 {
4396 	struct adapter *sc = arg;
4397 	struct sge_fl *fl, *fl_temp;
4398 
4399 	mtx_assert(&sc->sfl_lock, MA_OWNED);
4400 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
4401 		FL_LOCK(fl);
4402 		refill_fl(sc, fl, 64);
4403 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
4404 			TAILQ_REMOVE(&sc->sfl, fl, link);
4405 			fl->flags &= ~FL_STARVING;
4406 		}
4407 		FL_UNLOCK(fl);
4408 	}
4409 
4410 	if (!TAILQ_EMPTY(&sc->sfl))
4411 		callout_schedule(&sc->sfl_callout, hz / 5);
4412 }
4413 
4414 static int
4415 alloc_fl_sdesc(struct sge_fl *fl)
4416 {
4417 
4418 	fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc), M_CXGBE,
4419 	    M_ZERO | M_WAITOK);
4420 
4421 	return (0);
4422 }
4423 
4424 static void
4425 free_fl_sdesc(struct adapter *sc, struct sge_fl *fl)
4426 {
4427 	struct fl_sdesc *sd;
4428 	struct cluster_metadata *clm;
4429 	struct cluster_layout *cll;
4430 	int i;
4431 
4432 	sd = fl->sdesc;
4433 	for (i = 0; i < fl->sidx * 8; i++, sd++) {
4434 		if (sd->cl == NULL)
4435 			continue;
4436 
4437 		cll = &sd->cll;
4438 		clm = cl_metadata(sc, fl, cll, sd->cl);
4439 		if (sd->nmbuf == 0)
4440 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4441 		else if (clm && atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4442 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4443 			counter_u64_add(extfree_rels, 1);
4444 		}
4445 		sd->cl = NULL;
4446 	}
4447 
4448 	free(fl->sdesc, M_CXGBE);
4449 	fl->sdesc = NULL;
4450 }
4451 
4452 static inline void
4453 get_pkt_gl(struct mbuf *m, struct sglist *gl)
4454 {
4455 	int rc;
4456 
4457 	M_ASSERTPKTHDR(m);
4458 
4459 	sglist_reset(gl);
4460 	rc = sglist_append_mbuf(gl, m);
4461 	if (__predict_false(rc != 0)) {
4462 		panic("%s: mbuf %p (%d segs) was vetted earlier but now fails "
4463 		    "with %d.", __func__, m, mbuf_nsegs(m), rc);
4464 	}
4465 
4466 	KASSERT(gl->sg_nseg == mbuf_nsegs(m),
4467 	    ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m,
4468 	    mbuf_nsegs(m), gl->sg_nseg));
4469 	KASSERT(gl->sg_nseg > 0 &&
4470 	    gl->sg_nseg <= (needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS),
4471 	    ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__,
4472 		gl->sg_nseg, needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS));
4473 }
4474 
4475 /*
4476  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
4477  */
4478 static inline u_int
4479 txpkt_len16(u_int nsegs, u_int tso)
4480 {
4481 	u_int n;
4482 
4483 	MPASS(nsegs > 0);
4484 
4485 	nsegs--; /* first segment is part of ulptx_sgl */
4486 	n = sizeof(struct fw_eth_tx_pkt_wr) + sizeof(struct cpl_tx_pkt_core) +
4487 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4488 	if (tso)
4489 		n += sizeof(struct cpl_tx_pkt_lso_core);
4490 
4491 	return (howmany(n, 16));
4492 }
4493 
4494 /*
4495  * len16 for a txpkt_vm WR with a GL.  Includes the firmware work
4496  * request header.
4497  */
4498 static inline u_int
4499 txpkt_vm_len16(u_int nsegs, u_int tso)
4500 {
4501 	u_int n;
4502 
4503 	MPASS(nsegs > 0);
4504 
4505 	nsegs--; /* first segment is part of ulptx_sgl */
4506 	n = sizeof(struct fw_eth_tx_pkt_vm_wr) +
4507 	    sizeof(struct cpl_tx_pkt_core) +
4508 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4509 	if (tso)
4510 		n += sizeof(struct cpl_tx_pkt_lso_core);
4511 
4512 	return (howmany(n, 16));
4513 }
4514 
4515 /*
4516  * len16 for a txpkts type 0 WR with a GL.  Does not include the firmware work
4517  * request header.
4518  */
4519 static inline u_int
4520 txpkts0_len16(u_int nsegs)
4521 {
4522 	u_int n;
4523 
4524 	MPASS(nsegs > 0);
4525 
4526 	nsegs--; /* first segment is part of ulptx_sgl */
4527 	n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) +
4528 	    sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) +
4529 	    8 * ((3 * nsegs) / 2 + (nsegs & 1));
4530 
4531 	return (howmany(n, 16));
4532 }
4533 
4534 /*
4535  * len16 for a txpkts type 1 WR with a GL.  Does not include the firmware work
4536  * request header.
4537  */
4538 static inline u_int
4539 txpkts1_len16(void)
4540 {
4541 	u_int n;
4542 
4543 	n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl);
4544 
4545 	return (howmany(n, 16));
4546 }
4547 
4548 static inline u_int
4549 imm_payload(u_int ndesc)
4550 {
4551 	u_int n;
4552 
4553 	n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) -
4554 	    sizeof(struct cpl_tx_pkt_core);
4555 
4556 	return (n);
4557 }
4558 
4559 /*
4560  * Write a VM txpkt WR for this packet to the hardware descriptors, update the
4561  * software descriptor, and advance the pidx.  It is guaranteed that enough
4562  * descriptors are available.
4563  *
4564  * The return value is the # of hardware descriptors used.
4565  */
4566 static u_int
4567 write_txpkt_vm_wr(struct adapter *sc, struct sge_txq *txq,
4568     struct fw_eth_tx_pkt_vm_wr *wr, struct mbuf *m0, u_int available)
4569 {
4570 	struct sge_eq *eq = &txq->eq;
4571 	struct tx_sdesc *txsd;
4572 	struct cpl_tx_pkt_core *cpl;
4573 	uint32_t ctrl;	/* used in many unrelated places */
4574 	uint64_t ctrl1;
4575 	int csum_type, len16, ndesc, pktlen, nsegs;
4576 	caddr_t dst;
4577 
4578 	TXQ_LOCK_ASSERT_OWNED(txq);
4579 	M_ASSERTPKTHDR(m0);
4580 	MPASS(available > 0 && available < eq->sidx);
4581 
4582 	len16 = mbuf_len16(m0);
4583 	nsegs = mbuf_nsegs(m0);
4584 	pktlen = m0->m_pkthdr.len;
4585 	ctrl = sizeof(struct cpl_tx_pkt_core);
4586 	if (needs_tso(m0))
4587 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4588 	ndesc = howmany(len16, EQ_ESIZE / 16);
4589 	MPASS(ndesc <= available);
4590 
4591 	/* Firmware work request header */
4592 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4593 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_VM_WR) |
4594 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4595 
4596 	ctrl = V_FW_WR_LEN16(len16);
4597 	wr->equiq_to_len16 = htobe32(ctrl);
4598 	wr->r3[0] = 0;
4599 	wr->r3[1] = 0;
4600 
4601 	/*
4602 	 * Copy over ethmacdst, ethmacsrc, ethtype, and vlantci.
4603 	 * vlantci is ignored unless the ethtype is 0x8100, so it's
4604 	 * simpler to always copy it rather than making it
4605 	 * conditional.  Also, it seems that we do not have to set
4606 	 * vlantci or fake the ethtype when doing VLAN tag insertion.
4607 	 */
4608 	m_copydata(m0, 0, sizeof(struct ether_header) + 2, wr->ethmacdst);
4609 
4610 	csum_type = -1;
4611 	if (needs_tso(m0)) {
4612 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4613 
4614 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4615 		    m0->m_pkthdr.l4hlen > 0,
4616 		    ("%s: mbuf %p needs TSO but missing header lengths",
4617 			__func__, m0));
4618 
4619 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4620 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4621 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4622 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4623 			ctrl |= V_LSO_ETHHDR_LEN(1);
4624 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4625 			ctrl |= F_LSO_IPV6;
4626 
4627 		lso->lso_ctrl = htobe32(ctrl);
4628 		lso->ipid_ofst = htobe16(0);
4629 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4630 		lso->seqno_offset = htobe32(0);
4631 		lso->len = htobe32(pktlen);
4632 
4633 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4634 			csum_type = TX_CSUM_TCPIP6;
4635 		else
4636 			csum_type = TX_CSUM_TCPIP;
4637 
4638 		cpl = (void *)(lso + 1);
4639 
4640 		txq->tso_wrs++;
4641 	} else {
4642 		if (m0->m_pkthdr.csum_flags & CSUM_IP_TCP)
4643 			csum_type = TX_CSUM_TCPIP;
4644 		else if (m0->m_pkthdr.csum_flags & CSUM_IP_UDP)
4645 			csum_type = TX_CSUM_UDPIP;
4646 		else if (m0->m_pkthdr.csum_flags & CSUM_IP6_TCP)
4647 			csum_type = TX_CSUM_TCPIP6;
4648 		else if (m0->m_pkthdr.csum_flags & CSUM_IP6_UDP)
4649 			csum_type = TX_CSUM_UDPIP6;
4650 #if defined(INET)
4651 		else if (m0->m_pkthdr.csum_flags & CSUM_IP) {
4652 			/*
4653 			 * XXX: The firmware appears to stomp on the
4654 			 * fragment/flags field of the IP header when
4655 			 * using TX_CSUM_IP.  Fall back to doing
4656 			 * software checksums.
4657 			 */
4658 			u_short *sump;
4659 			struct mbuf *m;
4660 			int offset;
4661 
4662 			m = m0;
4663 			offset = 0;
4664 			sump = m_advance(&m, &offset, m0->m_pkthdr.l2hlen +
4665 			    offsetof(struct ip, ip_sum));
4666 			*sump = in_cksum_skip(m0, m0->m_pkthdr.l2hlen +
4667 			    m0->m_pkthdr.l3hlen, m0->m_pkthdr.l2hlen);
4668 			m0->m_pkthdr.csum_flags &= ~CSUM_IP;
4669 		}
4670 #endif
4671 
4672 		cpl = (void *)(wr + 1);
4673 	}
4674 
4675 	/* Checksum offload */
4676 	ctrl1 = 0;
4677 	if (needs_l3_csum(m0) == 0)
4678 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4679 	if (csum_type >= 0) {
4680 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0,
4681 	    ("%s: mbuf %p needs checksum offload but missing header lengths",
4682 			__func__, m0));
4683 
4684 		if (chip_id(sc) <= CHELSIO_T5) {
4685 			ctrl1 |= V_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4686 			    ETHER_HDR_LEN);
4687 		} else {
4688 			ctrl1 |= V_T6_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4689 			    ETHER_HDR_LEN);
4690 		}
4691 		ctrl1 |= V_TXPKT_IPHDR_LEN(m0->m_pkthdr.l3hlen);
4692 		ctrl1 |= V_TXPKT_CSUM_TYPE(csum_type);
4693 	} else
4694 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4695 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4696 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4697 		txq->txcsum++;	/* some hardware assistance provided */
4698 
4699 	/* VLAN tag insertion */
4700 	if (needs_vlan_insertion(m0)) {
4701 		ctrl1 |= F_TXPKT_VLAN_VLD |
4702 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4703 		txq->vlan_insertion++;
4704 	}
4705 
4706 	/* CPL header */
4707 	cpl->ctrl0 = txq->cpl_ctrl0;
4708 	cpl->pack = 0;
4709 	cpl->len = htobe16(pktlen);
4710 	cpl->ctrl1 = htobe64(ctrl1);
4711 
4712 	/* SGL */
4713 	dst = (void *)(cpl + 1);
4714 
4715 	/*
4716 	 * A packet using TSO will use up an entire descriptor for the
4717 	 * firmware work request header, LSO CPL, and TX_PKT_XT CPL.
4718 	 * If this descriptor is the last descriptor in the ring, wrap
4719 	 * around to the front of the ring explicitly for the start of
4720 	 * the sgl.
4721 	 */
4722 	if (dst == (void *)&eq->desc[eq->sidx]) {
4723 		dst = (void *)&eq->desc[0];
4724 		write_gl_to_txd(txq, m0, &dst, 0);
4725 	} else
4726 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4727 	txq->sgl_wrs++;
4728 
4729 	txq->txpkt_wrs++;
4730 
4731 	txsd = &txq->sdesc[eq->pidx];
4732 	txsd->m = m0;
4733 	txsd->desc_used = ndesc;
4734 
4735 	return (ndesc);
4736 }
4737 
4738 /*
4739  * Write a raw WR to the hardware descriptors, update the software
4740  * descriptor, and advance the pidx.  It is guaranteed that enough
4741  * descriptors are available.
4742  *
4743  * The return value is the # of hardware descriptors used.
4744  */
4745 static u_int
4746 write_raw_wr(struct sge_txq *txq, void *wr, struct mbuf *m0, u_int available)
4747 {
4748 	struct sge_eq *eq = &txq->eq;
4749 	struct tx_sdesc *txsd;
4750 	struct mbuf *m;
4751 	caddr_t dst;
4752 	int len16, ndesc;
4753 
4754 	len16 = mbuf_len16(m0);
4755 	ndesc = howmany(len16, EQ_ESIZE / 16);
4756 	MPASS(ndesc <= available);
4757 
4758 	dst = wr;
4759 	for (m = m0; m != NULL; m = m->m_next)
4760 		copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4761 
4762 	txq->raw_wrs++;
4763 
4764 	txsd = &txq->sdesc[eq->pidx];
4765 	txsd->m = m0;
4766 	txsd->desc_used = ndesc;
4767 
4768 	return (ndesc);
4769 }
4770 
4771 /*
4772  * Write a txpkt WR for this packet to the hardware descriptors, update the
4773  * software descriptor, and advance the pidx.  It is guaranteed that enough
4774  * descriptors are available.
4775  *
4776  * The return value is the # of hardware descriptors used.
4777  */
4778 static u_int
4779 write_txpkt_wr(struct sge_txq *txq, struct fw_eth_tx_pkt_wr *wr,
4780     struct mbuf *m0, u_int available)
4781 {
4782 	struct sge_eq *eq = &txq->eq;
4783 	struct tx_sdesc *txsd;
4784 	struct cpl_tx_pkt_core *cpl;
4785 	uint32_t ctrl;	/* used in many unrelated places */
4786 	uint64_t ctrl1;
4787 	int len16, ndesc, pktlen, nsegs;
4788 	caddr_t dst;
4789 
4790 	TXQ_LOCK_ASSERT_OWNED(txq);
4791 	M_ASSERTPKTHDR(m0);
4792 	MPASS(available > 0 && available < eq->sidx);
4793 
4794 	len16 = mbuf_len16(m0);
4795 	nsegs = mbuf_nsegs(m0);
4796 	pktlen = m0->m_pkthdr.len;
4797 	ctrl = sizeof(struct cpl_tx_pkt_core);
4798 	if (needs_tso(m0))
4799 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4800 	else if (!(mbuf_cflags(m0) & MC_NOMAP) && pktlen <= imm_payload(2) &&
4801 	    available >= 2) {
4802 		/* Immediate data.  Recalculate len16 and set nsegs to 0. */
4803 		ctrl += pktlen;
4804 		len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) +
4805 		    sizeof(struct cpl_tx_pkt_core) + pktlen, 16);
4806 		nsegs = 0;
4807 	}
4808 	ndesc = howmany(len16, EQ_ESIZE / 16);
4809 	MPASS(ndesc <= available);
4810 
4811 	/* Firmware work request header */
4812 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4813 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
4814 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4815 
4816 	ctrl = V_FW_WR_LEN16(len16);
4817 	wr->equiq_to_len16 = htobe32(ctrl);
4818 	wr->r3 = 0;
4819 
4820 	if (needs_tso(m0)) {
4821 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4822 
4823 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4824 		    m0->m_pkthdr.l4hlen > 0,
4825 		    ("%s: mbuf %p needs TSO but missing header lengths",
4826 			__func__, m0));
4827 
4828 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4829 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4830 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4831 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4832 			ctrl |= V_LSO_ETHHDR_LEN(1);
4833 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4834 			ctrl |= F_LSO_IPV6;
4835 
4836 		lso->lso_ctrl = htobe32(ctrl);
4837 		lso->ipid_ofst = htobe16(0);
4838 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4839 		lso->seqno_offset = htobe32(0);
4840 		lso->len = htobe32(pktlen);
4841 
4842 		cpl = (void *)(lso + 1);
4843 
4844 		txq->tso_wrs++;
4845 	} else
4846 		cpl = (void *)(wr + 1);
4847 
4848 	/* Checksum offload */
4849 	ctrl1 = 0;
4850 	if (needs_l3_csum(m0) == 0)
4851 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4852 	if (needs_l4_csum(m0) == 0)
4853 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4854 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4855 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4856 		txq->txcsum++;	/* some hardware assistance provided */
4857 
4858 	/* VLAN tag insertion */
4859 	if (needs_vlan_insertion(m0)) {
4860 		ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4861 		txq->vlan_insertion++;
4862 	}
4863 
4864 	/* CPL header */
4865 	cpl->ctrl0 = txq->cpl_ctrl0;
4866 	cpl->pack = 0;
4867 	cpl->len = htobe16(pktlen);
4868 	cpl->ctrl1 = htobe64(ctrl1);
4869 
4870 	/* SGL */
4871 	dst = (void *)(cpl + 1);
4872 	if (nsegs > 0) {
4873 
4874 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4875 		txq->sgl_wrs++;
4876 	} else {
4877 		struct mbuf *m;
4878 
4879 		for (m = m0; m != NULL; m = m->m_next) {
4880 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4881 #ifdef INVARIANTS
4882 			pktlen -= m->m_len;
4883 #endif
4884 		}
4885 #ifdef INVARIANTS
4886 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
4887 #endif
4888 		txq->imm_wrs++;
4889 	}
4890 
4891 	txq->txpkt_wrs++;
4892 
4893 	txsd = &txq->sdesc[eq->pidx];
4894 	txsd->m = m0;
4895 	txsd->desc_used = ndesc;
4896 
4897 	return (ndesc);
4898 }
4899 
4900 static int
4901 try_txpkts(struct mbuf *m, struct mbuf *n, struct txpkts *txp, u_int available)
4902 {
4903 	u_int needed, nsegs1, nsegs2, l1, l2;
4904 
4905 	if (cannot_use_txpkts(m) || cannot_use_txpkts(n))
4906 		return (1);
4907 
4908 	nsegs1 = mbuf_nsegs(m);
4909 	nsegs2 = mbuf_nsegs(n);
4910 	if (nsegs1 + nsegs2 == 2) {
4911 		txp->wr_type = 1;
4912 		l1 = l2 = txpkts1_len16();
4913 	} else {
4914 		txp->wr_type = 0;
4915 		l1 = txpkts0_len16(nsegs1);
4916 		l2 = txpkts0_len16(nsegs2);
4917 	}
4918 	txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) + l1 + l2;
4919 	needed = howmany(txp->len16, EQ_ESIZE / 16);
4920 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4921 		return (1);
4922 
4923 	txp->plen = m->m_pkthdr.len + n->m_pkthdr.len;
4924 	if (txp->plen > 65535)
4925 		return (1);
4926 
4927 	txp->npkt = 2;
4928 	set_mbuf_len16(m, l1);
4929 	set_mbuf_len16(n, l2);
4930 
4931 	return (0);
4932 }
4933 
4934 static int
4935 add_to_txpkts(struct mbuf *m, struct txpkts *txp, u_int available)
4936 {
4937 	u_int plen, len16, needed, nsegs;
4938 
4939 	MPASS(txp->wr_type == 0 || txp->wr_type == 1);
4940 
4941 	if (cannot_use_txpkts(m))
4942 		return (1);
4943 
4944 	nsegs = mbuf_nsegs(m);
4945 	if (txp->wr_type == 1 && nsegs != 1)
4946 		return (1);
4947 
4948 	plen = txp->plen + m->m_pkthdr.len;
4949 	if (plen > 65535)
4950 		return (1);
4951 
4952 	if (txp->wr_type == 0)
4953 		len16 = txpkts0_len16(nsegs);
4954 	else
4955 		len16 = txpkts1_len16();
4956 	needed = howmany(txp->len16 + len16, EQ_ESIZE / 16);
4957 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4958 		return (1);
4959 
4960 	txp->npkt++;
4961 	txp->plen = plen;
4962 	txp->len16 += len16;
4963 	set_mbuf_len16(m, len16);
4964 
4965 	return (0);
4966 }
4967 
4968 /*
4969  * Write a txpkts WR for the packets in txp to the hardware descriptors, update
4970  * the software descriptor, and advance the pidx.  It is guaranteed that enough
4971  * descriptors are available.
4972  *
4973  * The return value is the # of hardware descriptors used.
4974  */
4975 static u_int
4976 write_txpkts_wr(struct sge_txq *txq, struct fw_eth_tx_pkts_wr *wr,
4977     struct mbuf *m0, const struct txpkts *txp, u_int available)
4978 {
4979 	struct sge_eq *eq = &txq->eq;
4980 	struct tx_sdesc *txsd;
4981 	struct cpl_tx_pkt_core *cpl;
4982 	uint32_t ctrl;
4983 	uint64_t ctrl1;
4984 	int ndesc, checkwrap;
4985 	struct mbuf *m;
4986 	void *flitp;
4987 
4988 	TXQ_LOCK_ASSERT_OWNED(txq);
4989 	MPASS(txp->npkt > 0);
4990 	MPASS(txp->plen < 65536);
4991 	MPASS(m0 != NULL);
4992 	MPASS(m0->m_nextpkt != NULL);
4993 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
4994 	MPASS(available > 0 && available < eq->sidx);
4995 
4996 	ndesc = howmany(txp->len16, EQ_ESIZE / 16);
4997 	MPASS(ndesc <= available);
4998 
4999 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
5000 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
5001 	ctrl = V_FW_WR_LEN16(txp->len16);
5002 	wr->equiq_to_len16 = htobe32(ctrl);
5003 	wr->plen = htobe16(txp->plen);
5004 	wr->npkt = txp->npkt;
5005 	wr->r3 = 0;
5006 	wr->type = txp->wr_type;
5007 	flitp = wr + 1;
5008 
5009 	/*
5010 	 * At this point we are 16B into a hardware descriptor.  If checkwrap is
5011 	 * set then we know the WR is going to wrap around somewhere.  We'll
5012 	 * check for that at appropriate points.
5013 	 */
5014 	checkwrap = eq->sidx - ndesc < eq->pidx;
5015 	for (m = m0; m != NULL; m = m->m_nextpkt) {
5016 		if (txp->wr_type == 0) {
5017 			struct ulp_txpkt *ulpmc;
5018 			struct ulptx_idata *ulpsc;
5019 
5020 			/* ULP master command */
5021 			ulpmc = flitp;
5022 			ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
5023 			    V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid));
5024 			ulpmc->len = htobe32(mbuf_len16(m));
5025 
5026 			/* ULP subcommand */
5027 			ulpsc = (void *)(ulpmc + 1);
5028 			ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
5029 			    F_ULP_TX_SC_MORE);
5030 			ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
5031 
5032 			cpl = (void *)(ulpsc + 1);
5033 			if (checkwrap &&
5034 			    (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx])
5035 				cpl = (void *)&eq->desc[0];
5036 		} else {
5037 			cpl = flitp;
5038 		}
5039 
5040 		/* Checksum offload */
5041 		ctrl1 = 0;
5042 		if (needs_l3_csum(m) == 0)
5043 			ctrl1 |= F_TXPKT_IPCSUM_DIS;
5044 		if (needs_l4_csum(m) == 0)
5045 			ctrl1 |= F_TXPKT_L4CSUM_DIS;
5046 		if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
5047 		    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
5048 			txq->txcsum++;	/* some hardware assistance provided */
5049 
5050 		/* VLAN tag insertion */
5051 		if (needs_vlan_insertion(m)) {
5052 			ctrl1 |= F_TXPKT_VLAN_VLD |
5053 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
5054 			txq->vlan_insertion++;
5055 		}
5056 
5057 		/* CPL header */
5058 		cpl->ctrl0 = txq->cpl_ctrl0;
5059 		cpl->pack = 0;
5060 		cpl->len = htobe16(m->m_pkthdr.len);
5061 		cpl->ctrl1 = htobe64(ctrl1);
5062 
5063 		flitp = cpl + 1;
5064 		if (checkwrap &&
5065 		    (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
5066 			flitp = (void *)&eq->desc[0];
5067 
5068 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap);
5069 
5070 	}
5071 
5072 	if (txp->wr_type == 0) {
5073 		txq->txpkts0_pkts += txp->npkt;
5074 		txq->txpkts0_wrs++;
5075 	} else {
5076 		txq->txpkts1_pkts += txp->npkt;
5077 		txq->txpkts1_wrs++;
5078 	}
5079 
5080 	txsd = &txq->sdesc[eq->pidx];
5081 	txsd->m = m0;
5082 	txsd->desc_used = ndesc;
5083 
5084 	return (ndesc);
5085 }
5086 
5087 /*
5088  * If the SGL ends on an address that is not 16 byte aligned, this function will
5089  * add a 0 filled flit at the end.
5090  */
5091 static void
5092 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap)
5093 {
5094 	struct sge_eq *eq = &txq->eq;
5095 	struct sglist *gl = txq->gl;
5096 	struct sglist_seg *seg;
5097 	__be64 *flitp, *wrap;
5098 	struct ulptx_sgl *usgl;
5099 	int i, nflits, nsegs;
5100 
5101 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
5102 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
5103 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
5104 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
5105 
5106 	get_pkt_gl(m, gl);
5107 	nsegs = gl->sg_nseg;
5108 	MPASS(nsegs > 0);
5109 
5110 	nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
5111 	flitp = (__be64 *)(*to);
5112 	wrap = (__be64 *)(&eq->desc[eq->sidx]);
5113 	seg = &gl->sg_segs[0];
5114 	usgl = (void *)flitp;
5115 
5116 	/*
5117 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
5118 	 * ring, so we're at least 16 bytes away from the status page.  There is
5119 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
5120 	 */
5121 
5122 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
5123 	    V_ULPTX_NSGE(nsegs));
5124 	usgl->len0 = htobe32(seg->ss_len);
5125 	usgl->addr0 = htobe64(seg->ss_paddr);
5126 	seg++;
5127 
5128 	if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) {
5129 
5130 		/* Won't wrap around at all */
5131 
5132 		for (i = 0; i < nsegs - 1; i++, seg++) {
5133 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
5134 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
5135 		}
5136 		if (i & 1)
5137 			usgl->sge[i / 2].len[1] = htobe32(0);
5138 		flitp += nflits;
5139 	} else {
5140 
5141 		/* Will wrap somewhere in the rest of the SGL */
5142 
5143 		/* 2 flits already written, write the rest flit by flit */
5144 		flitp = (void *)(usgl + 1);
5145 		for (i = 0; i < nflits - 2; i++) {
5146 			if (flitp == wrap)
5147 				flitp = (void *)eq->desc;
5148 			*flitp++ = get_flit(seg, nsegs - 1, i);
5149 		}
5150 	}
5151 
5152 	if (nflits & 1) {
5153 		MPASS(((uintptr_t)flitp) & 0xf);
5154 		*flitp++ = 0;
5155 	}
5156 
5157 	MPASS((((uintptr_t)flitp) & 0xf) == 0);
5158 	if (__predict_false(flitp == wrap))
5159 		*to = (void *)eq->desc;
5160 	else
5161 		*to = (void *)flitp;
5162 }
5163 
5164 static inline void
5165 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
5166 {
5167 
5168 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
5169 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
5170 
5171 	if (__predict_true((uintptr_t)(*to) + len <=
5172 	    (uintptr_t)&eq->desc[eq->sidx])) {
5173 		bcopy(from, *to, len);
5174 		(*to) += len;
5175 	} else {
5176 		int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
5177 
5178 		bcopy(from, *to, portion);
5179 		from += portion;
5180 		portion = len - portion;	/* remaining */
5181 		bcopy(from, (void *)eq->desc, portion);
5182 		(*to) = (caddr_t)eq->desc + portion;
5183 	}
5184 }
5185 
5186 static inline void
5187 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n)
5188 {
5189 	u_int db;
5190 
5191 	MPASS(n > 0);
5192 
5193 	db = eq->doorbells;
5194 	if (n > 1)
5195 		clrbit(&db, DOORBELL_WCWR);
5196 	wmb();
5197 
5198 	switch (ffs(db) - 1) {
5199 	case DOORBELL_UDB:
5200 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
5201 		break;
5202 
5203 	case DOORBELL_WCWR: {
5204 		volatile uint64_t *dst, *src;
5205 		int i;
5206 
5207 		/*
5208 		 * Queues whose 128B doorbell segment fits in the page do not
5209 		 * use relative qid (udb_qid is always 0).  Only queues with
5210 		 * doorbell segments can do WCWR.
5211 		 */
5212 		KASSERT(eq->udb_qid == 0 && n == 1,
5213 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
5214 		    __func__, eq->doorbells, n, eq->dbidx, eq));
5215 
5216 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
5217 		    UDBS_DB_OFFSET);
5218 		i = eq->dbidx;
5219 		src = (void *)&eq->desc[i];
5220 		while (src != (void *)&eq->desc[i + 1])
5221 			*dst++ = *src++;
5222 		wmb();
5223 		break;
5224 	}
5225 
5226 	case DOORBELL_UDBWC:
5227 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
5228 		wmb();
5229 		break;
5230 
5231 	case DOORBELL_KDB:
5232 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
5233 		    V_QID(eq->cntxt_id) | V_PIDX(n));
5234 		break;
5235 	}
5236 
5237 	IDXINCR(eq->dbidx, n, eq->sidx);
5238 }
5239 
5240 static inline u_int
5241 reclaimable_tx_desc(struct sge_eq *eq)
5242 {
5243 	uint16_t hw_cidx;
5244 
5245 	hw_cidx = read_hw_cidx(eq);
5246 	return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx));
5247 }
5248 
5249 static inline u_int
5250 total_available_tx_desc(struct sge_eq *eq)
5251 {
5252 	uint16_t hw_cidx, pidx;
5253 
5254 	hw_cidx = read_hw_cidx(eq);
5255 	pidx = eq->pidx;
5256 
5257 	if (pidx == hw_cidx)
5258 		return (eq->sidx - 1);
5259 	else
5260 		return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1);
5261 }
5262 
5263 static inline uint16_t
5264 read_hw_cidx(struct sge_eq *eq)
5265 {
5266 	struct sge_qstat *spg = (void *)&eq->desc[eq->sidx];
5267 	uint16_t cidx = spg->cidx;	/* stable snapshot */
5268 
5269 	return (be16toh(cidx));
5270 }
5271 
5272 /*
5273  * Reclaim 'n' descriptors approximately.
5274  */
5275 static u_int
5276 reclaim_tx_descs(struct sge_txq *txq, u_int n)
5277 {
5278 	struct tx_sdesc *txsd;
5279 	struct sge_eq *eq = &txq->eq;
5280 	u_int can_reclaim, reclaimed;
5281 
5282 	TXQ_LOCK_ASSERT_OWNED(txq);
5283 	MPASS(n > 0);
5284 
5285 	reclaimed = 0;
5286 	can_reclaim = reclaimable_tx_desc(eq);
5287 	while (can_reclaim && reclaimed < n) {
5288 		int ndesc;
5289 		struct mbuf *m, *nextpkt;
5290 
5291 		txsd = &txq->sdesc[eq->cidx];
5292 		ndesc = txsd->desc_used;
5293 
5294 		/* Firmware doesn't return "partial" credits. */
5295 		KASSERT(can_reclaim >= ndesc,
5296 		    ("%s: unexpected number of credits: %d, %d",
5297 		    __func__, can_reclaim, ndesc));
5298 		KASSERT(ndesc != 0,
5299 		    ("%s: descriptor with no credits: cidx %d",
5300 		    __func__, eq->cidx));
5301 
5302 		for (m = txsd->m; m != NULL; m = nextpkt) {
5303 			nextpkt = m->m_nextpkt;
5304 			m->m_nextpkt = NULL;
5305 			m_freem(m);
5306 		}
5307 		reclaimed += ndesc;
5308 		can_reclaim -= ndesc;
5309 		IDXINCR(eq->cidx, ndesc, eq->sidx);
5310 	}
5311 
5312 	return (reclaimed);
5313 }
5314 
5315 static void
5316 tx_reclaim(void *arg, int n)
5317 {
5318 	struct sge_txq *txq = arg;
5319 	struct sge_eq *eq = &txq->eq;
5320 
5321 	do {
5322 		if (TXQ_TRYLOCK(txq) == 0)
5323 			break;
5324 		n = reclaim_tx_descs(txq, 32);
5325 		if (eq->cidx == eq->pidx)
5326 			eq->equeqidx = eq->pidx;
5327 		TXQ_UNLOCK(txq);
5328 	} while (n > 0);
5329 }
5330 
5331 static __be64
5332 get_flit(struct sglist_seg *segs, int nsegs, int idx)
5333 {
5334 	int i = (idx / 3) * 2;
5335 
5336 	switch (idx % 3) {
5337 	case 0: {
5338 		uint64_t rc;
5339 
5340 		rc = (uint64_t)segs[i].ss_len << 32;
5341 		if (i + 1 < nsegs)
5342 			rc |= (uint64_t)(segs[i + 1].ss_len);
5343 
5344 		return (htobe64(rc));
5345 	}
5346 	case 1:
5347 		return (htobe64(segs[i].ss_paddr));
5348 	case 2:
5349 		return (htobe64(segs[i + 1].ss_paddr));
5350 	}
5351 
5352 	return (0);
5353 }
5354 
5355 static void
5356 find_best_refill_source(struct adapter *sc, struct sge_fl *fl, int maxp)
5357 {
5358 	int8_t zidx, hwidx, idx;
5359 	uint16_t region1, region3;
5360 	int spare, spare_needed, n;
5361 	struct sw_zone_info *swz;
5362 	struct hw_buf_info *hwb, *hwb_list = &sc->sge.hw_buf_info[0];
5363 
5364 	/*
5365 	 * Buffer Packing: Look for PAGE_SIZE or larger zone which has a bufsize
5366 	 * large enough for the max payload and cluster metadata.  Otherwise
5367 	 * settle for the largest bufsize that leaves enough room in the cluster
5368 	 * for metadata.
5369 	 *
5370 	 * Without buffer packing: Look for the smallest zone which has a
5371 	 * bufsize large enough for the max payload.  Settle for the largest
5372 	 * bufsize available if there's nothing big enough for max payload.
5373 	 */
5374 	spare_needed = fl->flags & FL_BUF_PACKING ? CL_METADATA_SIZE : 0;
5375 	swz = &sc->sge.sw_zone_info[0];
5376 	hwidx = -1;
5377 	for (zidx = 0; zidx < SW_ZONE_SIZES; zidx++, swz++) {
5378 		if (swz->size > largest_rx_cluster) {
5379 			if (__predict_true(hwidx != -1))
5380 				break;
5381 
5382 			/*
5383 			 * This is a misconfiguration.  largest_rx_cluster is
5384 			 * preventing us from finding a refill source.  See
5385 			 * dev.t5nex.<n>.buffer_sizes to figure out why.
5386 			 */
5387 			device_printf(sc->dev, "largest_rx_cluster=%u leaves no"
5388 			    " refill source for fl %p (dma %u).  Ignored.\n",
5389 			    largest_rx_cluster, fl, maxp);
5390 		}
5391 		for (idx = swz->head_hwidx; idx != -1; idx = hwb->next) {
5392 			hwb = &hwb_list[idx];
5393 			spare = swz->size - hwb->size;
5394 			if (spare < spare_needed)
5395 				continue;
5396 
5397 			hwidx = idx;		/* best option so far */
5398 			if (hwb->size >= maxp) {
5399 
5400 				if ((fl->flags & FL_BUF_PACKING) == 0)
5401 					goto done; /* stop looking (not packing) */
5402 
5403 				if (swz->size >= safest_rx_cluster)
5404 					goto done; /* stop looking (packing) */
5405 			}
5406 			break;		/* keep looking, next zone */
5407 		}
5408 	}
5409 done:
5410 	/* A usable hwidx has been located. */
5411 	MPASS(hwidx != -1);
5412 	hwb = &hwb_list[hwidx];
5413 	zidx = hwb->zidx;
5414 	swz = &sc->sge.sw_zone_info[zidx];
5415 	region1 = 0;
5416 	region3 = swz->size - hwb->size;
5417 
5418 	/*
5419 	 * Stay within this zone and see if there is a better match when mbuf
5420 	 * inlining is allowed.  Remember that the hwidx's are sorted in
5421 	 * decreasing order of size (so in increasing order of spare area).
5422 	 */
5423 	for (idx = hwidx; idx != -1; idx = hwb->next) {
5424 		hwb = &hwb_list[idx];
5425 		spare = swz->size - hwb->size;
5426 
5427 		if (allow_mbufs_in_cluster == 0 || hwb->size < maxp)
5428 			break;
5429 
5430 		/*
5431 		 * Do not inline mbufs if doing so would violate the pad/pack
5432 		 * boundary alignment requirement.
5433 		 */
5434 		if (fl_pad && (MSIZE % sc->params.sge.pad_boundary) != 0)
5435 			continue;
5436 		if (fl->flags & FL_BUF_PACKING &&
5437 		    (MSIZE % sc->params.sge.pack_boundary) != 0)
5438 			continue;
5439 
5440 		if (spare < CL_METADATA_SIZE + MSIZE)
5441 			continue;
5442 		n = (spare - CL_METADATA_SIZE) / MSIZE;
5443 		if (n > howmany(hwb->size, maxp))
5444 			break;
5445 
5446 		hwidx = idx;
5447 		if (fl->flags & FL_BUF_PACKING) {
5448 			region1 = n * MSIZE;
5449 			region3 = spare - region1;
5450 		} else {
5451 			region1 = MSIZE;
5452 			region3 = spare - region1;
5453 			break;
5454 		}
5455 	}
5456 
5457 	KASSERT(zidx >= 0 && zidx < SW_ZONE_SIZES,
5458 	    ("%s: bad zone %d for fl %p, maxp %d", __func__, zidx, fl, maxp));
5459 	KASSERT(hwidx >= 0 && hwidx <= SGE_FLBUF_SIZES,
5460 	    ("%s: bad hwidx %d for fl %p, maxp %d", __func__, hwidx, fl, maxp));
5461 	KASSERT(region1 + sc->sge.hw_buf_info[hwidx].size + region3 ==
5462 	    sc->sge.sw_zone_info[zidx].size,
5463 	    ("%s: bad buffer layout for fl %p, maxp %d. "
5464 		"cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5465 		sc->sge.sw_zone_info[zidx].size, region1,
5466 		sc->sge.hw_buf_info[hwidx].size, region3));
5467 	if (fl->flags & FL_BUF_PACKING || region1 > 0) {
5468 		KASSERT(region3 >= CL_METADATA_SIZE,
5469 		    ("%s: no room for metadata.  fl %p, maxp %d; "
5470 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5471 		    sc->sge.sw_zone_info[zidx].size, region1,
5472 		    sc->sge.hw_buf_info[hwidx].size, region3));
5473 		KASSERT(region1 % MSIZE == 0,
5474 		    ("%s: bad mbuf region for fl %p, maxp %d. "
5475 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5476 		    sc->sge.sw_zone_info[zidx].size, region1,
5477 		    sc->sge.hw_buf_info[hwidx].size, region3));
5478 	}
5479 
5480 	fl->cll_def.zidx = zidx;
5481 	fl->cll_def.hwidx = hwidx;
5482 	fl->cll_def.region1 = region1;
5483 	fl->cll_def.region3 = region3;
5484 }
5485 
5486 static void
5487 find_safe_refill_source(struct adapter *sc, struct sge_fl *fl)
5488 {
5489 	struct sge *s = &sc->sge;
5490 	struct hw_buf_info *hwb;
5491 	struct sw_zone_info *swz;
5492 	int spare;
5493 	int8_t hwidx;
5494 
5495 	if (fl->flags & FL_BUF_PACKING)
5496 		hwidx = s->safe_hwidx2;	/* with room for metadata */
5497 	else if (allow_mbufs_in_cluster && s->safe_hwidx2 != -1) {
5498 		hwidx = s->safe_hwidx2;
5499 		hwb = &s->hw_buf_info[hwidx];
5500 		swz = &s->sw_zone_info[hwb->zidx];
5501 		spare = swz->size - hwb->size;
5502 
5503 		/* no good if there isn't room for an mbuf as well */
5504 		if (spare < CL_METADATA_SIZE + MSIZE)
5505 			hwidx = s->safe_hwidx1;
5506 	} else
5507 		hwidx = s->safe_hwidx1;
5508 
5509 	if (hwidx == -1) {
5510 		/* No fallback source */
5511 		fl->cll_alt.hwidx = -1;
5512 		fl->cll_alt.zidx = -1;
5513 
5514 		return;
5515 	}
5516 
5517 	hwb = &s->hw_buf_info[hwidx];
5518 	swz = &s->sw_zone_info[hwb->zidx];
5519 	spare = swz->size - hwb->size;
5520 	fl->cll_alt.hwidx = hwidx;
5521 	fl->cll_alt.zidx = hwb->zidx;
5522 	if (allow_mbufs_in_cluster &&
5523 	    (fl_pad == 0 || (MSIZE % sc->params.sge.pad_boundary) == 0))
5524 		fl->cll_alt.region1 = ((spare - CL_METADATA_SIZE) / MSIZE) * MSIZE;
5525 	else
5526 		fl->cll_alt.region1 = 0;
5527 	fl->cll_alt.region3 = spare - fl->cll_alt.region1;
5528 }
5529 
5530 static void
5531 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
5532 {
5533 	mtx_lock(&sc->sfl_lock);
5534 	FL_LOCK(fl);
5535 	if ((fl->flags & FL_DOOMED) == 0) {
5536 		fl->flags |= FL_STARVING;
5537 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
5538 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
5539 	}
5540 	FL_UNLOCK(fl);
5541 	mtx_unlock(&sc->sfl_lock);
5542 }
5543 
5544 static void
5545 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq)
5546 {
5547 	struct sge_wrq *wrq = (void *)eq;
5548 
5549 	atomic_readandclear_int(&eq->equiq);
5550 	taskqueue_enqueue(sc->tq[eq->tx_chan], &wrq->wrq_tx_task);
5551 }
5552 
5553 static void
5554 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq)
5555 {
5556 	struct sge_txq *txq = (void *)eq;
5557 
5558 	MPASS((eq->flags & EQ_TYPEMASK) == EQ_ETH);
5559 
5560 	atomic_readandclear_int(&eq->equiq);
5561 	mp_ring_check_drainage(txq->r, 0);
5562 	taskqueue_enqueue(sc->tq[eq->tx_chan], &txq->tx_reclaim_task);
5563 }
5564 
5565 static int
5566 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
5567     struct mbuf *m)
5568 {
5569 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
5570 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
5571 	struct adapter *sc = iq->adapter;
5572 	struct sge *s = &sc->sge;
5573 	struct sge_eq *eq;
5574 	static void (*h[])(struct adapter *, struct sge_eq *) = {NULL,
5575 		&handle_wrq_egr_update, &handle_eth_egr_update,
5576 		&handle_wrq_egr_update};
5577 
5578 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5579 	    rss->opcode));
5580 
5581 	eq = s->eqmap[qid - s->eq_start - s->eq_base];
5582 	(*h[eq->flags & EQ_TYPEMASK])(sc, eq);
5583 
5584 	return (0);
5585 }
5586 
5587 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
5588 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
5589     offsetof(struct cpl_fw6_msg, data));
5590 
5591 static int
5592 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
5593 {
5594 	struct adapter *sc = iq->adapter;
5595 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
5596 
5597 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5598 	    rss->opcode));
5599 
5600 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
5601 		const struct rss_header *rss2;
5602 
5603 		rss2 = (const struct rss_header *)&cpl->data[0];
5604 		return (t4_cpl_handler[rss2->opcode](iq, rss2, m));
5605 	}
5606 
5607 	return (t4_fw_msg_handler[cpl->type](sc, &cpl->data[0]));
5608 }
5609 
5610 /**
5611  *	t4_handle_wrerr_rpl - process a FW work request error message
5612  *	@adap: the adapter
5613  *	@rpl: start of the FW message
5614  */
5615 static int
5616 t4_handle_wrerr_rpl(struct adapter *adap, const __be64 *rpl)
5617 {
5618 	u8 opcode = *(const u8 *)rpl;
5619 	const struct fw_error_cmd *e = (const void *)rpl;
5620 	unsigned int i;
5621 
5622 	if (opcode != FW_ERROR_CMD) {
5623 		log(LOG_ERR,
5624 		    "%s: Received WRERR_RPL message with opcode %#x\n",
5625 		    device_get_nameunit(adap->dev), opcode);
5626 		return (EINVAL);
5627 	}
5628 	log(LOG_ERR, "%s: FW_ERROR (%s) ", device_get_nameunit(adap->dev),
5629 	    G_FW_ERROR_CMD_FATAL(be32toh(e->op_to_type)) ? "fatal" :
5630 	    "non-fatal");
5631 	switch (G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type))) {
5632 	case FW_ERROR_TYPE_EXCEPTION:
5633 		log(LOG_ERR, "exception info:\n");
5634 		for (i = 0; i < nitems(e->u.exception.info); i++)
5635 			log(LOG_ERR, "%s%08x", i == 0 ? "\t" : " ",
5636 			    be32toh(e->u.exception.info[i]));
5637 		log(LOG_ERR, "\n");
5638 		break;
5639 	case FW_ERROR_TYPE_HWMODULE:
5640 		log(LOG_ERR, "HW module regaddr %08x regval %08x\n",
5641 		    be32toh(e->u.hwmodule.regaddr),
5642 		    be32toh(e->u.hwmodule.regval));
5643 		break;
5644 	case FW_ERROR_TYPE_WR:
5645 		log(LOG_ERR, "WR cidx %d PF %d VF %d eqid %d hdr:\n",
5646 		    be16toh(e->u.wr.cidx),
5647 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.wr.pfn_vfn)),
5648 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.wr.pfn_vfn)),
5649 		    be32toh(e->u.wr.eqid));
5650 		for (i = 0; i < nitems(e->u.wr.wrhdr); i++)
5651 			log(LOG_ERR, "%s%02x", i == 0 ? "\t" : " ",
5652 			    e->u.wr.wrhdr[i]);
5653 		log(LOG_ERR, "\n");
5654 		break;
5655 	case FW_ERROR_TYPE_ACL:
5656 		log(LOG_ERR, "ACL cidx %d PF %d VF %d eqid %d %s",
5657 		    be16toh(e->u.acl.cidx),
5658 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.acl.pfn_vfn)),
5659 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.acl.pfn_vfn)),
5660 		    be32toh(e->u.acl.eqid),
5661 		    G_FW_ERROR_CMD_MV(be16toh(e->u.acl.mv_pkd)) ? "vlanid" :
5662 		    "MAC");
5663 		for (i = 0; i < nitems(e->u.acl.val); i++)
5664 			log(LOG_ERR, " %02x", e->u.acl.val[i]);
5665 		log(LOG_ERR, "\n");
5666 		break;
5667 	default:
5668 		log(LOG_ERR, "type %#x\n",
5669 		    G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type)));
5670 		return (EINVAL);
5671 	}
5672 	return (0);
5673 }
5674 
5675 static int
5676 sysctl_uint16(SYSCTL_HANDLER_ARGS)
5677 {
5678 	uint16_t *id = arg1;
5679 	int i = *id;
5680 
5681 	return sysctl_handle_int(oidp, &i, 0, req);
5682 }
5683 
5684 static int
5685 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
5686 {
5687 	struct sge *s = arg1;
5688 	struct hw_buf_info *hwb = &s->hw_buf_info[0];
5689 	struct sw_zone_info *swz = &s->sw_zone_info[0];
5690 	int i, rc;
5691 	struct sbuf sb;
5692 	char c;
5693 
5694 	sbuf_new(&sb, NULL, 32, SBUF_AUTOEXTEND);
5695 	for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
5696 		if (hwb->zidx >= 0 && swz[hwb->zidx].size <= largest_rx_cluster)
5697 			c = '*';
5698 		else
5699 			c = '\0';
5700 
5701 		sbuf_printf(&sb, "%u%c ", hwb->size, c);
5702 	}
5703 	sbuf_trim(&sb);
5704 	sbuf_finish(&sb);
5705 	rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
5706 	sbuf_delete(&sb);
5707 	return (rc);
5708 }
5709 
5710 #ifdef RATELIMIT
5711 /*
5712  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
5713  */
5714 static inline u_int
5715 txpkt_eo_len16(u_int nsegs, u_int immhdrs, u_int tso)
5716 {
5717 	u_int n;
5718 
5719 	MPASS(immhdrs > 0);
5720 
5721 	n = roundup2(sizeof(struct fw_eth_tx_eo_wr) +
5722 	    sizeof(struct cpl_tx_pkt_core) + immhdrs, 16);
5723 	if (__predict_false(nsegs == 0))
5724 		goto done;
5725 
5726 	nsegs--; /* first segment is part of ulptx_sgl */
5727 	n += sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
5728 	if (tso)
5729 		n += sizeof(struct cpl_tx_pkt_lso_core);
5730 
5731 done:
5732 	return (howmany(n, 16));
5733 }
5734 
5735 #define ETID_FLOWC_NPARAMS 6
5736 #define ETID_FLOWC_LEN (roundup2((sizeof(struct fw_flowc_wr) + \
5737     ETID_FLOWC_NPARAMS * sizeof(struct fw_flowc_mnemval)), 16))
5738 #define ETID_FLOWC_LEN16 (howmany(ETID_FLOWC_LEN, 16))
5739 
5740 static int
5741 send_etid_flowc_wr(struct cxgbe_snd_tag *cst, struct port_info *pi,
5742     struct vi_info *vi)
5743 {
5744 	struct wrq_cookie cookie;
5745 	u_int pfvf = pi->adapter->pf << S_FW_VIID_PFN;
5746 	struct fw_flowc_wr *flowc;
5747 
5748 	mtx_assert(&cst->lock, MA_OWNED);
5749 	MPASS((cst->flags & (EO_FLOWC_PENDING | EO_FLOWC_RPL_PENDING)) ==
5750 	    EO_FLOWC_PENDING);
5751 
5752 	flowc = start_wrq_wr(cst->eo_txq, ETID_FLOWC_LEN16, &cookie);
5753 	if (__predict_false(flowc == NULL))
5754 		return (ENOMEM);
5755 
5756 	bzero(flowc, ETID_FLOWC_LEN);
5757 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
5758 	    V_FW_FLOWC_WR_NPARAMS(ETID_FLOWC_NPARAMS) | V_FW_WR_COMPL(0));
5759 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(ETID_FLOWC_LEN16) |
5760 	    V_FW_WR_FLOWID(cst->etid));
5761 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN;
5762 	flowc->mnemval[0].val = htobe32(pfvf);
5763 	flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH;
5764 	flowc->mnemval[1].val = htobe32(pi->tx_chan);
5765 	flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT;
5766 	flowc->mnemval[2].val = htobe32(pi->tx_chan);
5767 	flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID;
5768 	flowc->mnemval[3].val = htobe32(cst->iqid);
5769 	flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_EOSTATE;
5770 	flowc->mnemval[4].val = htobe32(FW_FLOWC_MNEM_EOSTATE_ESTABLISHED);
5771 	flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS;
5772 	flowc->mnemval[5].val = htobe32(cst->schedcl);
5773 
5774 	commit_wrq_wr(cst->eo_txq, flowc, &cookie);
5775 
5776 	cst->flags &= ~EO_FLOWC_PENDING;
5777 	cst->flags |= EO_FLOWC_RPL_PENDING;
5778 	MPASS(cst->tx_credits >= ETID_FLOWC_LEN16);	/* flowc is first WR. */
5779 	cst->tx_credits -= ETID_FLOWC_LEN16;
5780 
5781 	return (0);
5782 }
5783 
5784 #define ETID_FLUSH_LEN16 (howmany(sizeof (struct fw_flowc_wr), 16))
5785 
5786 void
5787 send_etid_flush_wr(struct cxgbe_snd_tag *cst)
5788 {
5789 	struct fw_flowc_wr *flowc;
5790 	struct wrq_cookie cookie;
5791 
5792 	mtx_assert(&cst->lock, MA_OWNED);
5793 
5794 	flowc = start_wrq_wr(cst->eo_txq, ETID_FLUSH_LEN16, &cookie);
5795 	if (__predict_false(flowc == NULL))
5796 		CXGBE_UNIMPLEMENTED(__func__);
5797 
5798 	bzero(flowc, ETID_FLUSH_LEN16 * 16);
5799 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
5800 	    V_FW_FLOWC_WR_NPARAMS(0) | F_FW_WR_COMPL);
5801 	flowc->flowid_len16 = htobe32(V_FW_WR_LEN16(ETID_FLUSH_LEN16) |
5802 	    V_FW_WR_FLOWID(cst->etid));
5803 
5804 	commit_wrq_wr(cst->eo_txq, flowc, &cookie);
5805 
5806 	cst->flags |= EO_FLUSH_RPL_PENDING;
5807 	MPASS(cst->tx_credits >= ETID_FLUSH_LEN16);
5808 	cst->tx_credits -= ETID_FLUSH_LEN16;
5809 	cst->ncompl++;
5810 }
5811 
5812 static void
5813 write_ethofld_wr(struct cxgbe_snd_tag *cst, struct fw_eth_tx_eo_wr *wr,
5814     struct mbuf *m0, int compl)
5815 {
5816 	struct cpl_tx_pkt_core *cpl;
5817 	uint64_t ctrl1;
5818 	uint32_t ctrl;	/* used in many unrelated places */
5819 	int len16, pktlen, nsegs, immhdrs;
5820 	caddr_t dst;
5821 	uintptr_t p;
5822 	struct ulptx_sgl *usgl;
5823 	struct sglist sg;
5824 	struct sglist_seg segs[38];	/* XXX: find real limit.  XXX: get off the stack */
5825 
5826 	mtx_assert(&cst->lock, MA_OWNED);
5827 	M_ASSERTPKTHDR(m0);
5828 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
5829 	    m0->m_pkthdr.l4hlen > 0,
5830 	    ("%s: ethofld mbuf %p is missing header lengths", __func__, m0));
5831 
5832 	len16 = mbuf_eo_len16(m0);
5833 	nsegs = mbuf_eo_nsegs(m0);
5834 	pktlen = m0->m_pkthdr.len;
5835 	ctrl = sizeof(struct cpl_tx_pkt_core);
5836 	if (needs_tso(m0))
5837 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
5838 	immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen + m0->m_pkthdr.l4hlen;
5839 	ctrl += immhdrs;
5840 
5841 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_EO_WR) |
5842 	    V_FW_ETH_TX_EO_WR_IMMDLEN(ctrl) | V_FW_WR_COMPL(!!compl));
5843 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(len16) |
5844 	    V_FW_WR_FLOWID(cst->etid));
5845 	wr->r3 = 0;
5846 	if (needs_udp_csum(m0)) {
5847 		wr->u.udpseg.type = FW_ETH_TX_EO_TYPE_UDPSEG;
5848 		wr->u.udpseg.ethlen = m0->m_pkthdr.l2hlen;
5849 		wr->u.udpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
5850 		wr->u.udpseg.udplen = m0->m_pkthdr.l4hlen;
5851 		wr->u.udpseg.rtplen = 0;
5852 		wr->u.udpseg.r4 = 0;
5853 		wr->u.udpseg.mss = htobe16(pktlen - immhdrs);
5854 		wr->u.udpseg.schedpktsize = wr->u.udpseg.mss;
5855 		wr->u.udpseg.plen = htobe32(pktlen - immhdrs);
5856 		cpl = (void *)(wr + 1);
5857 	} else {
5858 		MPASS(needs_tcp_csum(m0));
5859 		wr->u.tcpseg.type = FW_ETH_TX_EO_TYPE_TCPSEG;
5860 		wr->u.tcpseg.ethlen = m0->m_pkthdr.l2hlen;
5861 		wr->u.tcpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
5862 		wr->u.tcpseg.tcplen = m0->m_pkthdr.l4hlen;
5863 		wr->u.tcpseg.tsclk_tsoff = mbuf_eo_tsclk_tsoff(m0);
5864 		wr->u.tcpseg.r4 = 0;
5865 		wr->u.tcpseg.r5 = 0;
5866 		wr->u.tcpseg.plen = htobe32(pktlen - immhdrs);
5867 
5868 		if (needs_tso(m0)) {
5869 			struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
5870 
5871 			wr->u.tcpseg.mss = htobe16(m0->m_pkthdr.tso_segsz);
5872 
5873 			ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) |
5874 			    F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE |
5875 			    V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) |
5876 			    V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
5877 			if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
5878 				ctrl |= V_LSO_ETHHDR_LEN(1);
5879 			if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
5880 				ctrl |= F_LSO_IPV6;
5881 			lso->lso_ctrl = htobe32(ctrl);
5882 			lso->ipid_ofst = htobe16(0);
5883 			lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
5884 			lso->seqno_offset = htobe32(0);
5885 			lso->len = htobe32(pktlen);
5886 
5887 			cpl = (void *)(lso + 1);
5888 		} else {
5889 			wr->u.tcpseg.mss = htobe16(0xffff);
5890 			cpl = (void *)(wr + 1);
5891 		}
5892 	}
5893 
5894 	/* Checksum offload must be requested for ethofld. */
5895 	ctrl1 = 0;
5896 	MPASS(needs_l4_csum(m0));
5897 
5898 	/* VLAN tag insertion */
5899 	if (needs_vlan_insertion(m0)) {
5900 		ctrl1 |= F_TXPKT_VLAN_VLD |
5901 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
5902 	}
5903 
5904 	/* CPL header */
5905 	cpl->ctrl0 = cst->ctrl0;
5906 	cpl->pack = 0;
5907 	cpl->len = htobe16(pktlen);
5908 	cpl->ctrl1 = htobe64(ctrl1);
5909 
5910 	/* Copy Ethernet, IP & TCP/UDP hdrs as immediate data */
5911 	p = (uintptr_t)(cpl + 1);
5912 	m_copydata(m0, 0, immhdrs, (void *)p);
5913 
5914 	/* SGL */
5915 	dst = (void *)(cpl + 1);
5916 	if (nsegs > 0) {
5917 		int i, pad;
5918 
5919 		/* zero-pad upto next 16Byte boundary, if not 16Byte aligned */
5920 		p += immhdrs;
5921 		pad = 16 - (immhdrs & 0xf);
5922 		bzero((void *)p, pad);
5923 
5924 		usgl = (void *)(p + pad);
5925 		usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
5926 		    V_ULPTX_NSGE(nsegs));
5927 
5928 		sglist_init(&sg, nitems(segs), segs);
5929 		for (; m0 != NULL; m0 = m0->m_next) {
5930 			if (__predict_false(m0->m_len == 0))
5931 				continue;
5932 			if (immhdrs >= m0->m_len) {
5933 				immhdrs -= m0->m_len;
5934 				continue;
5935 			}
5936 
5937 			sglist_append(&sg, mtod(m0, char *) + immhdrs,
5938 			    m0->m_len - immhdrs);
5939 			immhdrs = 0;
5940 		}
5941 		MPASS(sg.sg_nseg == nsegs);
5942 
5943 		/*
5944 		 * Zero pad last 8B in case the WR doesn't end on a 16B
5945 		 * boundary.
5946 		 */
5947 		*(uint64_t *)((char *)wr + len16 * 16 - 8) = 0;
5948 
5949 		usgl->len0 = htobe32(segs[0].ss_len);
5950 		usgl->addr0 = htobe64(segs[0].ss_paddr);
5951 		for (i = 0; i < nsegs - 1; i++) {
5952 			usgl->sge[i / 2].len[i & 1] = htobe32(segs[i + 1].ss_len);
5953 			usgl->sge[i / 2].addr[i & 1] = htobe64(segs[i + 1].ss_paddr);
5954 		}
5955 		if (i & 1)
5956 			usgl->sge[i / 2].len[1] = htobe32(0);
5957 	}
5958 
5959 }
5960 
5961 static void
5962 ethofld_tx(struct cxgbe_snd_tag *cst)
5963 {
5964 	struct mbuf *m;
5965 	struct wrq_cookie cookie;
5966 	int next_credits, compl;
5967 	struct fw_eth_tx_eo_wr *wr;
5968 
5969 	mtx_assert(&cst->lock, MA_OWNED);
5970 
5971 	while ((m = mbufq_first(&cst->pending_tx)) != NULL) {
5972 		M_ASSERTPKTHDR(m);
5973 
5974 		/* How many len16 credits do we need to send this mbuf. */
5975 		next_credits = mbuf_eo_len16(m);
5976 		MPASS(next_credits > 0);
5977 		if (next_credits > cst->tx_credits) {
5978 			/*
5979 			 * Tx will make progress eventually because there is at
5980 			 * least one outstanding fw4_ack that will return
5981 			 * credits and kick the tx.
5982 			 */
5983 			MPASS(cst->ncompl > 0);
5984 			return;
5985 		}
5986 		wr = start_wrq_wr(cst->eo_txq, next_credits, &cookie);
5987 		if (__predict_false(wr == NULL)) {
5988 			/* XXX: wishful thinking, not a real assertion. */
5989 			MPASS(cst->ncompl > 0);
5990 			return;
5991 		}
5992 		cst->tx_credits -= next_credits;
5993 		cst->tx_nocompl += next_credits;
5994 		compl = cst->ncompl == 0 || cst->tx_nocompl >= cst->tx_total / 2;
5995 		ETHER_BPF_MTAP(cst->com.ifp, m);
5996 		write_ethofld_wr(cst, wr, m, compl);
5997 		commit_wrq_wr(cst->eo_txq, wr, &cookie);
5998 		if (compl) {
5999 			cst->ncompl++;
6000 			cst->tx_nocompl	= 0;
6001 		}
6002 		(void) mbufq_dequeue(&cst->pending_tx);
6003 
6004 		/*
6005 		 * Drop the mbuf's reference on the tag now rather
6006 		 * than waiting until m_freem().  This ensures that
6007 		 * cxgbe_snd_tag_free gets called when the inp drops
6008 		 * its reference on the tag and there are no more
6009 		 * mbufs in the pending_tx queue and can flush any
6010 		 * pending requests.  Otherwise if the last mbuf
6011 		 * doesn't request a completion the etid will never be
6012 		 * released.
6013 		 */
6014 		m->m_pkthdr.snd_tag = NULL;
6015 		m->m_pkthdr.csum_flags &= ~CSUM_SND_TAG;
6016 		m_snd_tag_rele(&cst->com);
6017 
6018 		mbufq_enqueue(&cst->pending_fwack, m);
6019 	}
6020 }
6021 
6022 int
6023 ethofld_transmit(struct ifnet *ifp, struct mbuf *m0)
6024 {
6025 	struct cxgbe_snd_tag *cst;
6026 	int rc;
6027 
6028 	MPASS(m0->m_nextpkt == NULL);
6029 	MPASS(m0->m_pkthdr.csum_flags & CSUM_SND_TAG);
6030 	MPASS(m0->m_pkthdr.snd_tag != NULL);
6031 	cst = mst_to_cst(m0->m_pkthdr.snd_tag);
6032 
6033 	mtx_lock(&cst->lock);
6034 	MPASS(cst->flags & EO_SND_TAG_REF);
6035 
6036 	if (__predict_false(cst->flags & EO_FLOWC_PENDING)) {
6037 		struct vi_info *vi = ifp->if_softc;
6038 		struct port_info *pi = vi->pi;
6039 		struct adapter *sc = pi->adapter;
6040 		const uint32_t rss_mask = vi->rss_size - 1;
6041 		uint32_t rss_hash;
6042 
6043 		cst->eo_txq = &sc->sge.ofld_txq[vi->first_ofld_txq];
6044 		if (M_HASHTYPE_ISHASH(m0))
6045 			rss_hash = m0->m_pkthdr.flowid;
6046 		else
6047 			rss_hash = arc4random();
6048 		/* We assume RSS hashing */
6049 		cst->iqid = vi->rss[rss_hash & rss_mask];
6050 		cst->eo_txq += rss_hash % vi->nofldtxq;
6051 		rc = send_etid_flowc_wr(cst, pi, vi);
6052 		if (rc != 0)
6053 			goto done;
6054 	}
6055 
6056 	if (__predict_false(cst->plen + m0->m_pkthdr.len > eo_max_backlog)) {
6057 		rc = ENOBUFS;
6058 		goto done;
6059 	}
6060 
6061 	mbufq_enqueue(&cst->pending_tx, m0);
6062 	cst->plen += m0->m_pkthdr.len;
6063 
6064 	/*
6065 	 * Hold an extra reference on the tag while generating work
6066 	 * requests to ensure that we don't try to free the tag during
6067 	 * ethofld_tx() in case we are sending the final mbuf after
6068 	 * the inp was freed.
6069 	 */
6070 	m_snd_tag_ref(&cst->com);
6071 	ethofld_tx(cst);
6072 	mtx_unlock(&cst->lock);
6073 	m_snd_tag_rele(&cst->com);
6074 	return (0);
6075 
6076 done:
6077 	mtx_unlock(&cst->lock);
6078 	if (__predict_false(rc != 0))
6079 		m_freem(m0);
6080 	return (rc);
6081 }
6082 
6083 static int
6084 ethofld_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
6085 {
6086 	struct adapter *sc = iq->adapter;
6087 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
6088 	struct mbuf *m;
6089 	u_int etid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
6090 	struct cxgbe_snd_tag *cst;
6091 	uint8_t credits = cpl->credits;
6092 
6093 	cst = lookup_etid(sc, etid);
6094 	mtx_lock(&cst->lock);
6095 	if (__predict_false(cst->flags & EO_FLOWC_RPL_PENDING)) {
6096 		MPASS(credits >= ETID_FLOWC_LEN16);
6097 		credits -= ETID_FLOWC_LEN16;
6098 		cst->flags &= ~EO_FLOWC_RPL_PENDING;
6099 	}
6100 
6101 	KASSERT(cst->ncompl > 0,
6102 	    ("%s: etid %u (%p) wasn't expecting completion.",
6103 	    __func__, etid, cst));
6104 	cst->ncompl--;
6105 
6106 	while (credits > 0) {
6107 		m = mbufq_dequeue(&cst->pending_fwack);
6108 		if (__predict_false(m == NULL)) {
6109 			/*
6110 			 * The remaining credits are for the final flush that
6111 			 * was issued when the tag was freed by the kernel.
6112 			 */
6113 			MPASS((cst->flags &
6114 			    (EO_FLUSH_RPL_PENDING | EO_SND_TAG_REF)) ==
6115 			    EO_FLUSH_RPL_PENDING);
6116 			MPASS(credits == ETID_FLUSH_LEN16);
6117 			MPASS(cst->tx_credits + cpl->credits == cst->tx_total);
6118 			MPASS(cst->ncompl == 0);
6119 
6120 			cst->flags &= ~EO_FLUSH_RPL_PENDING;
6121 			cst->tx_credits += cpl->credits;
6122 			cxgbe_snd_tag_free_locked(cst);
6123 			return (0);	/* cst is gone. */
6124 		}
6125 		KASSERT(m != NULL,
6126 		    ("%s: too many credits (%u, %u)", __func__, cpl->credits,
6127 		    credits));
6128 		KASSERT(credits >= mbuf_eo_len16(m),
6129 		    ("%s: too few credits (%u, %u, %u)", __func__,
6130 		    cpl->credits, credits, mbuf_eo_len16(m)));
6131 		credits -= mbuf_eo_len16(m);
6132 		cst->plen -= m->m_pkthdr.len;
6133 		m_freem(m);
6134 	}
6135 
6136 	cst->tx_credits += cpl->credits;
6137 	MPASS(cst->tx_credits <= cst->tx_total);
6138 
6139 	if (cst->flags & EO_SND_TAG_REF) {
6140 		/*
6141 		 * As with ethofld_transmit(), hold an extra reference
6142 		 * so that the tag is stable across ethold_tx().
6143 		 */
6144 		m_snd_tag_ref(&cst->com);
6145 		m = mbufq_first(&cst->pending_tx);
6146 		if (m != NULL && cst->tx_credits >= mbuf_eo_len16(m))
6147 			ethofld_tx(cst);
6148 		mtx_unlock(&cst->lock);
6149 		m_snd_tag_rele(&cst->com);
6150 	} else {
6151 		/*
6152 		 * There shouldn't be any pending packets if the tag
6153 		 * was freed by the kernel since any pending packet
6154 		 * should hold a reference to the tag.
6155 		 */
6156 		MPASS(mbufq_first(&cst->pending_tx) == NULL);
6157 		mtx_unlock(&cst->lock);
6158 	}
6159 
6160 	return (0);
6161 }
6162 #endif
6163