xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision 6829dae12bb055451fa467da4589c43bd03b1e64)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2011 Chelsio Communications, Inc.
5  * All rights reserved.
6  * Written by: Navdeep Parhar <np@FreeBSD.org>
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include "opt_inet.h"
34 #include "opt_inet6.h"
35 #include "opt_ratelimit.h"
36 
37 #include <sys/types.h>
38 #include <sys/eventhandler.h>
39 #include <sys/mbuf.h>
40 #include <sys/socket.h>
41 #include <sys/kernel.h>
42 #include <sys/malloc.h>
43 #include <sys/queue.h>
44 #include <sys/sbuf.h>
45 #include <sys/taskqueue.h>
46 #include <sys/time.h>
47 #include <sys/sglist.h>
48 #include <sys/sysctl.h>
49 #include <sys/smp.h>
50 #include <sys/counter.h>
51 #include <net/bpf.h>
52 #include <net/ethernet.h>
53 #include <net/if.h>
54 #include <net/if_vlan_var.h>
55 #include <netinet/in.h>
56 #include <netinet/ip.h>
57 #include <netinet/ip6.h>
58 #include <netinet/tcp.h>
59 #include <netinet/udp.h>
60 #include <machine/in_cksum.h>
61 #include <machine/md_var.h>
62 #include <vm/vm.h>
63 #include <vm/pmap.h>
64 #ifdef DEV_NETMAP
65 #include <machine/bus.h>
66 #include <sys/selinfo.h>
67 #include <net/if_var.h>
68 #include <net/netmap.h>
69 #include <dev/netmap/netmap_kern.h>
70 #endif
71 
72 #include "common/common.h"
73 #include "common/t4_regs.h"
74 #include "common/t4_regs_values.h"
75 #include "common/t4_msg.h"
76 #include "t4_l2t.h"
77 #include "t4_mp_ring.h"
78 
79 #ifdef T4_PKT_TIMESTAMP
80 #define RX_COPY_THRESHOLD (MINCLSIZE - 8)
81 #else
82 #define RX_COPY_THRESHOLD MINCLSIZE
83 #endif
84 
85 /* Internal mbuf flags stored in PH_loc.eight[1]. */
86 #define	MC_RAW_WR		0x02
87 
88 /*
89  * Ethernet frames are DMA'd at this byte offset into the freelist buffer.
90  * 0-7 are valid values.
91  */
92 static int fl_pktshift = 0;
93 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pktshift, CTLFLAG_RDTUN, &fl_pktshift, 0,
94     "payload DMA offset in rx buffer (bytes)");
95 
96 /*
97  * Pad ethernet payload up to this boundary.
98  * -1: driver should figure out a good value.
99  *  0: disable padding.
100  *  Any power of 2 from 32 to 4096 (both inclusive) is also a valid value.
101  */
102 int fl_pad = -1;
103 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pad, CTLFLAG_RDTUN, &fl_pad, 0,
104     "payload pad boundary (bytes)");
105 
106 /*
107  * Status page length.
108  * -1: driver should figure out a good value.
109  *  64 or 128 are the only other valid values.
110  */
111 static int spg_len = -1;
112 SYSCTL_INT(_hw_cxgbe, OID_AUTO, spg_len, CTLFLAG_RDTUN, &spg_len, 0,
113     "status page size (bytes)");
114 
115 /*
116  * Congestion drops.
117  * -1: no congestion feedback (not recommended).
118  *  0: backpressure the channel instead of dropping packets right away.
119  *  1: no backpressure, drop packets for the congested queue immediately.
120  */
121 static int cong_drop = 0;
122 SYSCTL_INT(_hw_cxgbe, OID_AUTO, cong_drop, CTLFLAG_RDTUN, &cong_drop, 0,
123     "Congestion control for RX queues (0 = backpressure, 1 = drop");
124 
125 /*
126  * Deliver multiple frames in the same free list buffer if they fit.
127  * -1: let the driver decide whether to enable buffer packing or not.
128  *  0: disable buffer packing.
129  *  1: enable buffer packing.
130  */
131 static int buffer_packing = -1;
132 SYSCTL_INT(_hw_cxgbe, OID_AUTO, buffer_packing, CTLFLAG_RDTUN, &buffer_packing,
133     0, "Enable buffer packing");
134 
135 /*
136  * Start next frame in a packed buffer at this boundary.
137  * -1: driver should figure out a good value.
138  * T4: driver will ignore this and use the same value as fl_pad above.
139  * T5: 16, or a power of 2 from 64 to 4096 (both inclusive) is a valid value.
140  */
141 static int fl_pack = -1;
142 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pack, CTLFLAG_RDTUN, &fl_pack, 0,
143     "payload pack boundary (bytes)");
144 
145 /*
146  * Allow the driver to create mbuf(s) in a cluster allocated for rx.
147  * 0: never; always allocate mbufs from the zone_mbuf UMA zone.
148  * 1: ok to create mbuf(s) within a cluster if there is room.
149  */
150 static int allow_mbufs_in_cluster = 1;
151 SYSCTL_INT(_hw_cxgbe, OID_AUTO, allow_mbufs_in_cluster, CTLFLAG_RDTUN,
152     &allow_mbufs_in_cluster, 0,
153     "Allow driver to create mbufs within a rx cluster");
154 
155 /*
156  * Largest rx cluster size that the driver is allowed to allocate.
157  */
158 static int largest_rx_cluster = MJUM16BYTES;
159 SYSCTL_INT(_hw_cxgbe, OID_AUTO, largest_rx_cluster, CTLFLAG_RDTUN,
160     &largest_rx_cluster, 0, "Largest rx cluster (bytes)");
161 
162 /*
163  * Size of cluster allocation that's most likely to succeed.  The driver will
164  * fall back to this size if it fails to allocate clusters larger than this.
165  */
166 static int safest_rx_cluster = PAGE_SIZE;
167 SYSCTL_INT(_hw_cxgbe, OID_AUTO, safest_rx_cluster, CTLFLAG_RDTUN,
168     &safest_rx_cluster, 0, "Safe rx cluster (bytes)");
169 
170 #ifdef RATELIMIT
171 /*
172  * Knob to control TCP timestamp rewriting, and the granularity of the tick used
173  * for rewriting.  -1 and 0-3 are all valid values.
174  * -1: hardware should leave the TCP timestamps alone.
175  * 0: 1ms
176  * 1: 100us
177  * 2: 10us
178  * 3: 1us
179  */
180 static int tsclk = -1;
181 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tsclk, CTLFLAG_RDTUN, &tsclk, 0,
182     "Control TCP timestamp rewriting when using pacing");
183 
184 static int eo_max_backlog = 1024 * 1024;
185 SYSCTL_INT(_hw_cxgbe, OID_AUTO, eo_max_backlog, CTLFLAG_RDTUN, &eo_max_backlog,
186     0, "Maximum backlog of ratelimited data per flow");
187 #endif
188 
189 /*
190  * The interrupt holdoff timers are multiplied by this value on T6+.
191  * 1 and 3-17 (both inclusive) are legal values.
192  */
193 static int tscale = 1;
194 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tscale, CTLFLAG_RDTUN, &tscale, 0,
195     "Interrupt holdoff timer scale on T6+");
196 
197 /*
198  * Number of LRO entries in the lro_ctrl structure per rx queue.
199  */
200 static int lro_entries = TCP_LRO_ENTRIES;
201 SYSCTL_INT(_hw_cxgbe, OID_AUTO, lro_entries, CTLFLAG_RDTUN, &lro_entries, 0,
202     "Number of LRO entries per RX queue");
203 
204 /*
205  * This enables presorting of frames before they're fed into tcp_lro_rx.
206  */
207 static int lro_mbufs = 0;
208 SYSCTL_INT(_hw_cxgbe, OID_AUTO, lro_mbufs, CTLFLAG_RDTUN, &lro_mbufs, 0,
209     "Enable presorting of LRO frames");
210 
211 struct txpkts {
212 	u_int wr_type;		/* type 0 or type 1 */
213 	u_int npkt;		/* # of packets in this work request */
214 	u_int plen;		/* total payload (sum of all packets) */
215 	u_int len16;		/* # of 16B pieces used by this work request */
216 };
217 
218 /* A packet's SGL.  This + m_pkthdr has all info needed for tx */
219 struct sgl {
220 	struct sglist sg;
221 	struct sglist_seg seg[TX_SGL_SEGS];
222 };
223 
224 static int service_iq(struct sge_iq *, int);
225 static int service_iq_fl(struct sge_iq *, int);
226 static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t);
227 static int t4_eth_rx(struct sge_iq *, const struct rss_header *, struct mbuf *);
228 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int);
229 static inline void init_fl(struct adapter *, struct sge_fl *, int, int, char *);
230 static inline void init_eq(struct adapter *, struct sge_eq *, int, int, uint8_t,
231     uint16_t, char *);
232 static int alloc_ring(struct adapter *, size_t, bus_dma_tag_t *, bus_dmamap_t *,
233     bus_addr_t *, void **);
234 static int free_ring(struct adapter *, bus_dma_tag_t, bus_dmamap_t, bus_addr_t,
235     void *);
236 static int alloc_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *,
237     int, int);
238 static int free_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *);
239 static void add_iq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
240     struct sge_iq *);
241 static void add_fl_sysctls(struct adapter *, struct sysctl_ctx_list *,
242     struct sysctl_oid *, struct sge_fl *);
243 static int alloc_fwq(struct adapter *);
244 static int free_fwq(struct adapter *);
245 static int alloc_ctrlq(struct adapter *, struct sge_wrq *, int,
246     struct sysctl_oid *);
247 static int alloc_rxq(struct vi_info *, struct sge_rxq *, int, int,
248     struct sysctl_oid *);
249 static int free_rxq(struct vi_info *, struct sge_rxq *);
250 #ifdef TCP_OFFLOAD
251 static int alloc_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *, int, int,
252     struct sysctl_oid *);
253 static int free_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *);
254 #endif
255 #ifdef DEV_NETMAP
256 static int alloc_nm_rxq(struct vi_info *, struct sge_nm_rxq *, int, int,
257     struct sysctl_oid *);
258 static int free_nm_rxq(struct vi_info *, struct sge_nm_rxq *);
259 static int alloc_nm_txq(struct vi_info *, struct sge_nm_txq *, int, int,
260     struct sysctl_oid *);
261 static int free_nm_txq(struct vi_info *, struct sge_nm_txq *);
262 #endif
263 static int ctrl_eq_alloc(struct adapter *, struct sge_eq *);
264 static int eth_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
265 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
266 static int ofld_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
267 #endif
268 static int alloc_eq(struct adapter *, struct vi_info *, struct sge_eq *);
269 static int free_eq(struct adapter *, struct sge_eq *);
270 static int alloc_wrq(struct adapter *, struct vi_info *, struct sge_wrq *,
271     struct sysctl_oid *);
272 static int free_wrq(struct adapter *, struct sge_wrq *);
273 static int alloc_txq(struct vi_info *, struct sge_txq *, int,
274     struct sysctl_oid *);
275 static int free_txq(struct vi_info *, struct sge_txq *);
276 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int);
277 static inline void ring_fl_db(struct adapter *, struct sge_fl *);
278 static int refill_fl(struct adapter *, struct sge_fl *, int);
279 static void refill_sfl(void *);
280 static int alloc_fl_sdesc(struct sge_fl *);
281 static void free_fl_sdesc(struct adapter *, struct sge_fl *);
282 static void find_best_refill_source(struct adapter *, struct sge_fl *, int);
283 static void find_safe_refill_source(struct adapter *, struct sge_fl *);
284 static void add_fl_to_sfl(struct adapter *, struct sge_fl *);
285 
286 static inline void get_pkt_gl(struct mbuf *, struct sglist *);
287 static inline u_int txpkt_len16(u_int, u_int);
288 static inline u_int txpkt_vm_len16(u_int, u_int);
289 static inline u_int txpkts0_len16(u_int);
290 static inline u_int txpkts1_len16(void);
291 static u_int write_raw_wr(struct sge_txq *, void *, struct mbuf *, u_int);
292 static u_int write_txpkt_wr(struct sge_txq *, struct fw_eth_tx_pkt_wr *,
293     struct mbuf *, u_int);
294 static u_int write_txpkt_vm_wr(struct adapter *, struct sge_txq *,
295     struct fw_eth_tx_pkt_vm_wr *, struct mbuf *, u_int);
296 static int try_txpkts(struct mbuf *, struct mbuf *, struct txpkts *, u_int);
297 static int add_to_txpkts(struct mbuf *, struct txpkts *, u_int);
298 static u_int write_txpkts_wr(struct sge_txq *, struct fw_eth_tx_pkts_wr *,
299     struct mbuf *, const struct txpkts *, u_int);
300 static void write_gl_to_txd(struct sge_txq *, struct mbuf *, caddr_t *, int);
301 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int);
302 static inline void ring_eq_db(struct adapter *, struct sge_eq *, u_int);
303 static inline uint16_t read_hw_cidx(struct sge_eq *);
304 static inline u_int reclaimable_tx_desc(struct sge_eq *);
305 static inline u_int total_available_tx_desc(struct sge_eq *);
306 static u_int reclaim_tx_descs(struct sge_txq *, u_int);
307 static void tx_reclaim(void *, int);
308 static __be64 get_flit(struct sglist_seg *, int, int);
309 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
310     struct mbuf *);
311 static int handle_fw_msg(struct sge_iq *, const struct rss_header *,
312     struct mbuf *);
313 static int t4_handle_wrerr_rpl(struct adapter *, const __be64 *);
314 static void wrq_tx_drain(void *, int);
315 static void drain_wrq_wr_list(struct adapter *, struct sge_wrq *);
316 
317 static int sysctl_uint16(SYSCTL_HANDLER_ARGS);
318 static int sysctl_bufsizes(SYSCTL_HANDLER_ARGS);
319 #ifdef RATELIMIT
320 static inline u_int txpkt_eo_len16(u_int, u_int, u_int);
321 static int ethofld_fw4_ack(struct sge_iq *, const struct rss_header *,
322     struct mbuf *);
323 #endif
324 
325 static counter_u64_t extfree_refs;
326 static counter_u64_t extfree_rels;
327 
328 an_handler_t t4_an_handler;
329 fw_msg_handler_t t4_fw_msg_handler[NUM_FW6_TYPES];
330 cpl_handler_t t4_cpl_handler[NUM_CPL_CMDS];
331 cpl_handler_t set_tcb_rpl_handlers[NUM_CPL_COOKIES];
332 cpl_handler_t l2t_write_rpl_handlers[NUM_CPL_COOKIES];
333 cpl_handler_t act_open_rpl_handlers[NUM_CPL_COOKIES];
334 cpl_handler_t abort_rpl_rss_handlers[NUM_CPL_COOKIES];
335 cpl_handler_t fw4_ack_handlers[NUM_CPL_COOKIES];
336 
337 void
338 t4_register_an_handler(an_handler_t h)
339 {
340 	uintptr_t *loc;
341 
342 	MPASS(h == NULL || t4_an_handler == NULL);
343 
344 	loc = (uintptr_t *)&t4_an_handler;
345 	atomic_store_rel_ptr(loc, (uintptr_t)h);
346 }
347 
348 void
349 t4_register_fw_msg_handler(int type, fw_msg_handler_t h)
350 {
351 	uintptr_t *loc;
352 
353 	MPASS(type < nitems(t4_fw_msg_handler));
354 	MPASS(h == NULL || t4_fw_msg_handler[type] == NULL);
355 	/*
356 	 * These are dispatched by the handler for FW{4|6}_CPL_MSG using the CPL
357 	 * handler dispatch table.  Reject any attempt to install a handler for
358 	 * this subtype.
359 	 */
360 	MPASS(type != FW_TYPE_RSSCPL);
361 	MPASS(type != FW6_TYPE_RSSCPL);
362 
363 	loc = (uintptr_t *)&t4_fw_msg_handler[type];
364 	atomic_store_rel_ptr(loc, (uintptr_t)h);
365 }
366 
367 void
368 t4_register_cpl_handler(int opcode, cpl_handler_t h)
369 {
370 	uintptr_t *loc;
371 
372 	MPASS(opcode < nitems(t4_cpl_handler));
373 	MPASS(h == NULL || t4_cpl_handler[opcode] == NULL);
374 
375 	loc = (uintptr_t *)&t4_cpl_handler[opcode];
376 	atomic_store_rel_ptr(loc, (uintptr_t)h);
377 }
378 
379 static int
380 set_tcb_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
381     struct mbuf *m)
382 {
383 	const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
384 	u_int tid;
385 	int cookie;
386 
387 	MPASS(m == NULL);
388 
389 	tid = GET_TID(cpl);
390 	if (is_hpftid(iq->adapter, tid) || is_ftid(iq->adapter, tid)) {
391 		/*
392 		 * The return code for filter-write is put in the CPL cookie so
393 		 * we have to rely on the hardware tid (is_ftid) to determine
394 		 * that this is a response to a filter.
395 		 */
396 		cookie = CPL_COOKIE_FILTER;
397 	} else {
398 		cookie = G_COOKIE(cpl->cookie);
399 	}
400 	MPASS(cookie > CPL_COOKIE_RESERVED);
401 	MPASS(cookie < nitems(set_tcb_rpl_handlers));
402 
403 	return (set_tcb_rpl_handlers[cookie](iq, rss, m));
404 }
405 
406 static int
407 l2t_write_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
408     struct mbuf *m)
409 {
410 	const struct cpl_l2t_write_rpl *rpl = (const void *)(rss + 1);
411 	unsigned int cookie;
412 
413 	MPASS(m == NULL);
414 
415 	cookie = GET_TID(rpl) & F_SYNC_WR ? CPL_COOKIE_TOM : CPL_COOKIE_FILTER;
416 	return (l2t_write_rpl_handlers[cookie](iq, rss, m));
417 }
418 
419 static int
420 act_open_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
421     struct mbuf *m)
422 {
423 	const struct cpl_act_open_rpl *cpl = (const void *)(rss + 1);
424 	u_int cookie = G_TID_COOKIE(G_AOPEN_ATID(be32toh(cpl->atid_status)));
425 
426 	MPASS(m == NULL);
427 	MPASS(cookie != CPL_COOKIE_RESERVED);
428 
429 	return (act_open_rpl_handlers[cookie](iq, rss, m));
430 }
431 
432 static int
433 abort_rpl_rss_handler(struct sge_iq *iq, const struct rss_header *rss,
434     struct mbuf *m)
435 {
436 	struct adapter *sc = iq->adapter;
437 	u_int cookie;
438 
439 	MPASS(m == NULL);
440 	if (is_hashfilter(sc))
441 		cookie = CPL_COOKIE_HASHFILTER;
442 	else
443 		cookie = CPL_COOKIE_TOM;
444 
445 	return (abort_rpl_rss_handlers[cookie](iq, rss, m));
446 }
447 
448 static int
449 fw4_ack_handler(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
450 {
451 	struct adapter *sc = iq->adapter;
452 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
453 	unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
454 	u_int cookie;
455 
456 	MPASS(m == NULL);
457 	if (is_etid(sc, tid))
458 		cookie = CPL_COOKIE_ETHOFLD;
459 	else
460 		cookie = CPL_COOKIE_TOM;
461 
462 	return (fw4_ack_handlers[cookie](iq, rss, m));
463 }
464 
465 static void
466 t4_init_shared_cpl_handlers(void)
467 {
468 
469 	t4_register_cpl_handler(CPL_SET_TCB_RPL, set_tcb_rpl_handler);
470 	t4_register_cpl_handler(CPL_L2T_WRITE_RPL, l2t_write_rpl_handler);
471 	t4_register_cpl_handler(CPL_ACT_OPEN_RPL, act_open_rpl_handler);
472 	t4_register_cpl_handler(CPL_ABORT_RPL_RSS, abort_rpl_rss_handler);
473 	t4_register_cpl_handler(CPL_FW4_ACK, fw4_ack_handler);
474 }
475 
476 void
477 t4_register_shared_cpl_handler(int opcode, cpl_handler_t h, int cookie)
478 {
479 	uintptr_t *loc;
480 
481 	MPASS(opcode < nitems(t4_cpl_handler));
482 	MPASS(cookie > CPL_COOKIE_RESERVED);
483 	MPASS(cookie < NUM_CPL_COOKIES);
484 	MPASS(t4_cpl_handler[opcode] != NULL);
485 
486 	switch (opcode) {
487 	case CPL_SET_TCB_RPL:
488 		loc = (uintptr_t *)&set_tcb_rpl_handlers[cookie];
489 		break;
490 	case CPL_L2T_WRITE_RPL:
491 		loc = (uintptr_t *)&l2t_write_rpl_handlers[cookie];
492 		break;
493 	case CPL_ACT_OPEN_RPL:
494 		loc = (uintptr_t *)&act_open_rpl_handlers[cookie];
495 		break;
496 	case CPL_ABORT_RPL_RSS:
497 		loc = (uintptr_t *)&abort_rpl_rss_handlers[cookie];
498 		break;
499 	case CPL_FW4_ACK:
500 		loc = (uintptr_t *)&fw4_ack_handlers[cookie];
501 		break;
502 	default:
503 		MPASS(0);
504 		return;
505 	}
506 	MPASS(h == NULL || *loc == (uintptr_t)NULL);
507 	atomic_store_rel_ptr(loc, (uintptr_t)h);
508 }
509 
510 /*
511  * Called on MOD_LOAD.  Validates and calculates the SGE tunables.
512  */
513 void
514 t4_sge_modload(void)
515 {
516 
517 	if (fl_pktshift < 0 || fl_pktshift > 7) {
518 		printf("Invalid hw.cxgbe.fl_pktshift value (%d),"
519 		    " using 0 instead.\n", fl_pktshift);
520 		fl_pktshift = 0;
521 	}
522 
523 	if (spg_len != 64 && spg_len != 128) {
524 		int len;
525 
526 #if defined(__i386__) || defined(__amd64__)
527 		len = cpu_clflush_line_size > 64 ? 128 : 64;
528 #else
529 		len = 64;
530 #endif
531 		if (spg_len != -1) {
532 			printf("Invalid hw.cxgbe.spg_len value (%d),"
533 			    " using %d instead.\n", spg_len, len);
534 		}
535 		spg_len = len;
536 	}
537 
538 	if (cong_drop < -1 || cong_drop > 1) {
539 		printf("Invalid hw.cxgbe.cong_drop value (%d),"
540 		    " using 0 instead.\n", cong_drop);
541 		cong_drop = 0;
542 	}
543 
544 	if (tscale != 1 && (tscale < 3 || tscale > 17)) {
545 		printf("Invalid hw.cxgbe.tscale value (%d),"
546 		    " using 1 instead.\n", tscale);
547 		tscale = 1;
548 	}
549 
550 	extfree_refs = counter_u64_alloc(M_WAITOK);
551 	extfree_rels = counter_u64_alloc(M_WAITOK);
552 	counter_u64_zero(extfree_refs);
553 	counter_u64_zero(extfree_rels);
554 
555 	t4_init_shared_cpl_handlers();
556 	t4_register_cpl_handler(CPL_FW4_MSG, handle_fw_msg);
557 	t4_register_cpl_handler(CPL_FW6_MSG, handle_fw_msg);
558 	t4_register_cpl_handler(CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
559 	t4_register_cpl_handler(CPL_RX_PKT, t4_eth_rx);
560 #ifdef RATELIMIT
561 	t4_register_shared_cpl_handler(CPL_FW4_ACK, ethofld_fw4_ack,
562 	    CPL_COOKIE_ETHOFLD);
563 #endif
564 	t4_register_fw_msg_handler(FW6_TYPE_CMD_RPL, t4_handle_fw_rpl);
565 	t4_register_fw_msg_handler(FW6_TYPE_WRERR_RPL, t4_handle_wrerr_rpl);
566 }
567 
568 void
569 t4_sge_modunload(void)
570 {
571 
572 	counter_u64_free(extfree_refs);
573 	counter_u64_free(extfree_rels);
574 }
575 
576 uint64_t
577 t4_sge_extfree_refs(void)
578 {
579 	uint64_t refs, rels;
580 
581 	rels = counter_u64_fetch(extfree_rels);
582 	refs = counter_u64_fetch(extfree_refs);
583 
584 	return (refs - rels);
585 }
586 
587 static inline void
588 setup_pad_and_pack_boundaries(struct adapter *sc)
589 {
590 	uint32_t v, m;
591 	int pad, pack, pad_shift;
592 
593 	pad_shift = chip_id(sc) > CHELSIO_T5 ? X_T6_INGPADBOUNDARY_SHIFT :
594 	    X_INGPADBOUNDARY_SHIFT;
595 	pad = fl_pad;
596 	if (fl_pad < (1 << pad_shift) ||
597 	    fl_pad > (1 << (pad_shift + M_INGPADBOUNDARY)) ||
598 	    !powerof2(fl_pad)) {
599 		/*
600 		 * If there is any chance that we might use buffer packing and
601 		 * the chip is a T4, then pick 64 as the pad/pack boundary.  Set
602 		 * it to the minimum allowed in all other cases.
603 		 */
604 		pad = is_t4(sc) && buffer_packing ? 64 : 1 << pad_shift;
605 
606 		/*
607 		 * For fl_pad = 0 we'll still write a reasonable value to the
608 		 * register but all the freelists will opt out of padding.
609 		 * We'll complain here only if the user tried to set it to a
610 		 * value greater than 0 that was invalid.
611 		 */
612 		if (fl_pad > 0) {
613 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pad value"
614 			    " (%d), using %d instead.\n", fl_pad, pad);
615 		}
616 	}
617 	m = V_INGPADBOUNDARY(M_INGPADBOUNDARY);
618 	v = V_INGPADBOUNDARY(ilog2(pad) - pad_shift);
619 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
620 
621 	if (is_t4(sc)) {
622 		if (fl_pack != -1 && fl_pack != pad) {
623 			/* Complain but carry on. */
624 			device_printf(sc->dev, "hw.cxgbe.fl_pack (%d) ignored,"
625 			    " using %d instead.\n", fl_pack, pad);
626 		}
627 		return;
628 	}
629 
630 	pack = fl_pack;
631 	if (fl_pack < 16 || fl_pack == 32 || fl_pack > 4096 ||
632 	    !powerof2(fl_pack)) {
633 		pack = max(sc->params.pci.mps, CACHE_LINE_SIZE);
634 		MPASS(powerof2(pack));
635 		if (pack < 16)
636 			pack = 16;
637 		if (pack == 32)
638 			pack = 64;
639 		if (pack > 4096)
640 			pack = 4096;
641 		if (fl_pack != -1) {
642 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pack value"
643 			    " (%d), using %d instead.\n", fl_pack, pack);
644 		}
645 	}
646 	m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY);
647 	if (pack == 16)
648 		v = V_INGPACKBOUNDARY(0);
649 	else
650 		v = V_INGPACKBOUNDARY(ilog2(pack) - 5);
651 
652 	MPASS(!is_t4(sc));	/* T4 doesn't have SGE_CONTROL2 */
653 	t4_set_reg_field(sc, A_SGE_CONTROL2, m, v);
654 }
655 
656 /*
657  * adap->params.vpd.cclk must be set up before this is called.
658  */
659 void
660 t4_tweak_chip_settings(struct adapter *sc)
661 {
662 	int i;
663 	uint32_t v, m;
664 	int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200};
665 	int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk;
666 	int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */
667 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
668 	static int sge_flbuf_sizes[] = {
669 		MCLBYTES,
670 #if MJUMPAGESIZE != MCLBYTES
671 		MJUMPAGESIZE,
672 		MJUMPAGESIZE - CL_METADATA_SIZE,
673 		MJUMPAGESIZE - 2 * MSIZE - CL_METADATA_SIZE,
674 #endif
675 		MJUM9BYTES,
676 		MJUM16BYTES,
677 		MCLBYTES - MSIZE - CL_METADATA_SIZE,
678 		MJUM9BYTES - CL_METADATA_SIZE,
679 		MJUM16BYTES - CL_METADATA_SIZE,
680 	};
681 
682 	KASSERT(sc->flags & MASTER_PF,
683 	    ("%s: trying to change chip settings when not master.", __func__));
684 
685 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
686 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
687 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
688 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
689 
690 	setup_pad_and_pack_boundaries(sc);
691 
692 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
693 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
694 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
695 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
696 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
697 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
698 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
699 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
700 	t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v);
701 
702 	KASSERT(nitems(sge_flbuf_sizes) <= SGE_FLBUF_SIZES,
703 	    ("%s: hw buffer size table too big", __func__));
704 	t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0, 4096);
705 	t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE1, 65536);
706 	for (i = 0; i < min(nitems(sge_flbuf_sizes), SGE_FLBUF_SIZES); i++) {
707 		t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE15 - (4 * i),
708 		    sge_flbuf_sizes[i]);
709 	}
710 
711 	v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) |
712 	    V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]);
713 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v);
714 
715 	KASSERT(intr_timer[0] <= timer_max,
716 	    ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0],
717 	    timer_max));
718 	for (i = 1; i < nitems(intr_timer); i++) {
719 		KASSERT(intr_timer[i] >= intr_timer[i - 1],
720 		    ("%s: timers not listed in increasing order (%d)",
721 		    __func__, i));
722 
723 		while (intr_timer[i] > timer_max) {
724 			if (i == nitems(intr_timer) - 1) {
725 				intr_timer[i] = timer_max;
726 				break;
727 			}
728 			intr_timer[i] += intr_timer[i - 1];
729 			intr_timer[i] /= 2;
730 		}
731 	}
732 
733 	v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) |
734 	    V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1]));
735 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v);
736 	v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) |
737 	    V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3]));
738 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v);
739 	v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) |
740 	    V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5]));
741 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v);
742 
743 	if (chip_id(sc) >= CHELSIO_T6) {
744 		m = V_TSCALE(M_TSCALE);
745 		if (tscale == 1)
746 			v = 0;
747 		else
748 			v = V_TSCALE(tscale - 2);
749 		t4_set_reg_field(sc, A_SGE_ITP_CONTROL, m, v);
750 
751 		if (sc->debug_flags & DF_DISABLE_TCB_CACHE) {
752 			m = V_RDTHRESHOLD(M_RDTHRESHOLD) | F_WRTHRTHRESHEN |
753 			    V_WRTHRTHRESH(M_WRTHRTHRESH);
754 			t4_tp_pio_read(sc, &v, 1, A_TP_CMM_CONFIG, 1);
755 			v &= ~m;
756 			v |= V_RDTHRESHOLD(1) | F_WRTHRTHRESHEN |
757 			    V_WRTHRTHRESH(16);
758 			t4_tp_pio_write(sc, &v, 1, A_TP_CMM_CONFIG, 1);
759 		}
760 	}
761 
762 	/* 4K, 16K, 64K, 256K DDP "page sizes" for TDDP */
763 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
764 	t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v);
765 
766 	/*
767 	 * 4K, 8K, 16K, 64K DDP "page sizes" for iSCSI DDP.  These have been
768 	 * chosen with MAXPHYS = 128K in mind.  The largest DDP buffer that we
769 	 * may have to deal with is MAXPHYS + 1 page.
770 	 */
771 	v = V_HPZ0(0) | V_HPZ1(1) | V_HPZ2(2) | V_HPZ3(4);
772 	t4_write_reg(sc, A_ULP_RX_ISCSI_PSZ, v);
773 
774 	/* We use multiple DDP page sizes both in plain-TOE and ISCSI modes. */
775 	m = v = F_TDDPTAGTCB | F_ISCSITAGTCB;
776 	t4_set_reg_field(sc, A_ULP_RX_CTL, m, v);
777 
778 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
779 	    F_RESETDDPOFFSET;
780 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
781 	t4_set_reg_field(sc, A_TP_PARA_REG5, m, v);
782 }
783 
784 /*
785  * SGE wants the buffer to be at least 64B and then a multiple of 16.  If
786  * padding is in use, the buffer's start and end need to be aligned to the pad
787  * boundary as well.  We'll just make sure that the size is a multiple of the
788  * boundary here, it is up to the buffer allocation code to make sure the start
789  * of the buffer is aligned as well.
790  */
791 static inline int
792 hwsz_ok(struct adapter *sc, int hwsz)
793 {
794 	int mask = fl_pad ? sc->params.sge.pad_boundary - 1 : 16 - 1;
795 
796 	return (hwsz >= 64 && (hwsz & mask) == 0);
797 }
798 
799 /*
800  * XXX: driver really should be able to deal with unexpected settings.
801  */
802 int
803 t4_read_chip_settings(struct adapter *sc)
804 {
805 	struct sge *s = &sc->sge;
806 	struct sge_params *sp = &sc->params.sge;
807 	int i, j, n, rc = 0;
808 	uint32_t m, v, r;
809 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
810 	static int sw_buf_sizes[] = {	/* Sorted by size */
811 		MCLBYTES,
812 #if MJUMPAGESIZE != MCLBYTES
813 		MJUMPAGESIZE,
814 #endif
815 		MJUM9BYTES,
816 		MJUM16BYTES
817 	};
818 	struct sw_zone_info *swz, *safe_swz;
819 	struct hw_buf_info *hwb;
820 
821 	m = F_RXPKTCPLMODE;
822 	v = F_RXPKTCPLMODE;
823 	r = sc->params.sge.sge_control;
824 	if ((r & m) != v) {
825 		device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r);
826 		rc = EINVAL;
827 	}
828 
829 	/*
830 	 * If this changes then every single use of PAGE_SHIFT in the driver
831 	 * needs to be carefully reviewed for PAGE_SHIFT vs sp->page_shift.
832 	 */
833 	if (sp->page_shift != PAGE_SHIFT) {
834 		device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r);
835 		rc = EINVAL;
836 	}
837 
838 	/* Filter out unusable hw buffer sizes entirely (mark with -2). */
839 	hwb = &s->hw_buf_info[0];
840 	for (i = 0; i < nitems(s->hw_buf_info); i++, hwb++) {
841 		r = sc->params.sge.sge_fl_buffer_size[i];
842 		hwb->size = r;
843 		hwb->zidx = hwsz_ok(sc, r) ? -1 : -2;
844 		hwb->next = -1;
845 	}
846 
847 	/*
848 	 * Create a sorted list in decreasing order of hw buffer sizes (and so
849 	 * increasing order of spare area) for each software zone.
850 	 *
851 	 * If padding is enabled then the start and end of the buffer must align
852 	 * to the pad boundary; if packing is enabled then they must align with
853 	 * the pack boundary as well.  Allocations from the cluster zones are
854 	 * aligned to min(size, 4K), so the buffer starts at that alignment and
855 	 * ends at hwb->size alignment.  If mbuf inlining is allowed the
856 	 * starting alignment will be reduced to MSIZE and the driver will
857 	 * exercise appropriate caution when deciding on the best buffer layout
858 	 * to use.
859 	 */
860 	n = 0;	/* no usable buffer size to begin with */
861 	swz = &s->sw_zone_info[0];
862 	safe_swz = NULL;
863 	for (i = 0; i < SW_ZONE_SIZES; i++, swz++) {
864 		int8_t head = -1, tail = -1;
865 
866 		swz->size = sw_buf_sizes[i];
867 		swz->zone = m_getzone(swz->size);
868 		swz->type = m_gettype(swz->size);
869 
870 		if (swz->size < PAGE_SIZE) {
871 			MPASS(powerof2(swz->size));
872 			if (fl_pad && (swz->size % sp->pad_boundary != 0))
873 				continue;
874 		}
875 
876 		if (swz->size == safest_rx_cluster)
877 			safe_swz = swz;
878 
879 		hwb = &s->hw_buf_info[0];
880 		for (j = 0; j < SGE_FLBUF_SIZES; j++, hwb++) {
881 			if (hwb->zidx != -1 || hwb->size > swz->size)
882 				continue;
883 #ifdef INVARIANTS
884 			if (fl_pad)
885 				MPASS(hwb->size % sp->pad_boundary == 0);
886 #endif
887 			hwb->zidx = i;
888 			if (head == -1)
889 				head = tail = j;
890 			else if (hwb->size < s->hw_buf_info[tail].size) {
891 				s->hw_buf_info[tail].next = j;
892 				tail = j;
893 			} else {
894 				int8_t *cur;
895 				struct hw_buf_info *t;
896 
897 				for (cur = &head; *cur != -1; cur = &t->next) {
898 					t = &s->hw_buf_info[*cur];
899 					if (hwb->size == t->size) {
900 						hwb->zidx = -2;
901 						break;
902 					}
903 					if (hwb->size > t->size) {
904 						hwb->next = *cur;
905 						*cur = j;
906 						break;
907 					}
908 				}
909 			}
910 		}
911 		swz->head_hwidx = head;
912 		swz->tail_hwidx = tail;
913 
914 		if (tail != -1) {
915 			n++;
916 			if (swz->size - s->hw_buf_info[tail].size >=
917 			    CL_METADATA_SIZE)
918 				sc->flags |= BUF_PACKING_OK;
919 		}
920 	}
921 	if (n == 0) {
922 		device_printf(sc->dev, "no usable SGE FL buffer size.\n");
923 		rc = EINVAL;
924 	}
925 
926 	s->safe_hwidx1 = -1;
927 	s->safe_hwidx2 = -1;
928 	if (safe_swz != NULL) {
929 		s->safe_hwidx1 = safe_swz->head_hwidx;
930 		for (i = safe_swz->head_hwidx; i != -1; i = hwb->next) {
931 			int spare;
932 
933 			hwb = &s->hw_buf_info[i];
934 #ifdef INVARIANTS
935 			if (fl_pad)
936 				MPASS(hwb->size % sp->pad_boundary == 0);
937 #endif
938 			spare = safe_swz->size - hwb->size;
939 			if (spare >= CL_METADATA_SIZE) {
940 				s->safe_hwidx2 = i;
941 				break;
942 			}
943 		}
944 	}
945 
946 	if (sc->flags & IS_VF)
947 		return (0);
948 
949 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
950 	r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ);
951 	if (r != v) {
952 		device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r);
953 		rc = EINVAL;
954 	}
955 
956 	m = v = F_TDDPTAGTCB;
957 	r = t4_read_reg(sc, A_ULP_RX_CTL);
958 	if ((r & m) != v) {
959 		device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r);
960 		rc = EINVAL;
961 	}
962 
963 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
964 	    F_RESETDDPOFFSET;
965 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
966 	r = t4_read_reg(sc, A_TP_PARA_REG5);
967 	if ((r & m) != v) {
968 		device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r);
969 		rc = EINVAL;
970 	}
971 
972 	t4_init_tp_params(sc, 1);
973 
974 	t4_read_mtu_tbl(sc, sc->params.mtus, NULL);
975 	t4_load_mtus(sc, sc->params.mtus, sc->params.a_wnd, sc->params.b_wnd);
976 
977 	return (rc);
978 }
979 
980 int
981 t4_create_dma_tag(struct adapter *sc)
982 {
983 	int rc;
984 
985 	rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
986 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
987 	    BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
988 	    NULL, &sc->dmat);
989 	if (rc != 0) {
990 		device_printf(sc->dev,
991 		    "failed to create main DMA tag: %d\n", rc);
992 	}
993 
994 	return (rc);
995 }
996 
997 void
998 t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
999     struct sysctl_oid_list *children)
1000 {
1001 	struct sge_params *sp = &sc->params.sge;
1002 
1003 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "buffer_sizes",
1004 	    CTLTYPE_STRING | CTLFLAG_RD, &sc->sge, 0, sysctl_bufsizes, "A",
1005 	    "freelist buffer sizes");
1006 
1007 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD,
1008 	    NULL, sp->fl_pktshift, "payload DMA offset in rx buffer (bytes)");
1009 
1010 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD,
1011 	    NULL, sp->pad_boundary, "payload pad boundary (bytes)");
1012 
1013 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD,
1014 	    NULL, sp->spg_len, "status page size (bytes)");
1015 
1016 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD,
1017 	    NULL, cong_drop, "congestion drop setting");
1018 
1019 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD,
1020 	    NULL, sp->pack_boundary, "payload pack boundary (bytes)");
1021 }
1022 
1023 int
1024 t4_destroy_dma_tag(struct adapter *sc)
1025 {
1026 	if (sc->dmat)
1027 		bus_dma_tag_destroy(sc->dmat);
1028 
1029 	return (0);
1030 }
1031 
1032 /*
1033  * Allocate and initialize the firmware event queue, control queues, and special
1034  * purpose rx queues owned by the adapter.
1035  *
1036  * Returns errno on failure.  Resources allocated up to that point may still be
1037  * allocated.  Caller is responsible for cleanup in case this function fails.
1038  */
1039 int
1040 t4_setup_adapter_queues(struct adapter *sc)
1041 {
1042 	struct sysctl_oid *oid;
1043 	struct sysctl_oid_list *children;
1044 	int rc, i;
1045 
1046 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
1047 
1048 	sysctl_ctx_init(&sc->ctx);
1049 	sc->flags |= ADAP_SYSCTL_CTX;
1050 
1051 	/*
1052 	 * Firmware event queue
1053 	 */
1054 	rc = alloc_fwq(sc);
1055 	if (rc != 0)
1056 		return (rc);
1057 
1058 	/*
1059 	 * That's all for the VF driver.
1060 	 */
1061 	if (sc->flags & IS_VF)
1062 		return (rc);
1063 
1064 	oid = device_get_sysctl_tree(sc->dev);
1065 	children = SYSCTL_CHILDREN(oid);
1066 
1067 	/*
1068 	 * XXX: General purpose rx queues, one per port.
1069 	 */
1070 
1071 	/*
1072 	 * Control queues, one per port.
1073 	 */
1074 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "ctrlq",
1075 	    CTLFLAG_RD, NULL, "control queues");
1076 	for_each_port(sc, i) {
1077 		struct sge_wrq *ctrlq = &sc->sge.ctrlq[i];
1078 
1079 		rc = alloc_ctrlq(sc, ctrlq, i, oid);
1080 		if (rc != 0)
1081 			return (rc);
1082 	}
1083 
1084 	return (rc);
1085 }
1086 
1087 /*
1088  * Idempotent
1089  */
1090 int
1091 t4_teardown_adapter_queues(struct adapter *sc)
1092 {
1093 	int i;
1094 
1095 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
1096 
1097 	/* Do this before freeing the queue */
1098 	if (sc->flags & ADAP_SYSCTL_CTX) {
1099 		sysctl_ctx_free(&sc->ctx);
1100 		sc->flags &= ~ADAP_SYSCTL_CTX;
1101 	}
1102 
1103 	if (!(sc->flags & IS_VF)) {
1104 		for_each_port(sc, i)
1105 			free_wrq(sc, &sc->sge.ctrlq[i]);
1106 	}
1107 	free_fwq(sc);
1108 
1109 	return (0);
1110 }
1111 
1112 /* Maximum payload that can be delivered with a single iq descriptor */
1113 static inline int
1114 mtu_to_max_payload(struct adapter *sc, int mtu, const int toe)
1115 {
1116 	int payload;
1117 
1118 #ifdef TCP_OFFLOAD
1119 	if (toe) {
1120 		int rxcs = G_RXCOALESCESIZE(t4_read_reg(sc, A_TP_PARA_REG2));
1121 
1122 		/* Note that COP can set rx_coalesce on/off per connection. */
1123 		payload = max(mtu, rxcs);
1124 	} else {
1125 #endif
1126 		/* large enough even when hw VLAN extraction is disabled */
1127 		payload = sc->params.sge.fl_pktshift + ETHER_HDR_LEN +
1128 		    ETHER_VLAN_ENCAP_LEN + mtu;
1129 #ifdef TCP_OFFLOAD
1130 	}
1131 #endif
1132 
1133 	return (payload);
1134 }
1135 
1136 int
1137 t4_setup_vi_queues(struct vi_info *vi)
1138 {
1139 	int rc = 0, i, intr_idx, iqidx;
1140 	struct sge_rxq *rxq;
1141 	struct sge_txq *txq;
1142 #ifdef TCP_OFFLOAD
1143 	struct sge_ofld_rxq *ofld_rxq;
1144 #endif
1145 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1146 	struct sge_wrq *ofld_txq;
1147 #endif
1148 #ifdef DEV_NETMAP
1149 	int saved_idx;
1150 	struct sge_nm_rxq *nm_rxq;
1151 	struct sge_nm_txq *nm_txq;
1152 #endif
1153 	char name[16];
1154 	struct port_info *pi = vi->pi;
1155 	struct adapter *sc = pi->adapter;
1156 	struct ifnet *ifp = vi->ifp;
1157 	struct sysctl_oid *oid = device_get_sysctl_tree(vi->dev);
1158 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
1159 	int maxp, mtu = ifp->if_mtu;
1160 
1161 	/* Interrupt vector to start from (when using multiple vectors) */
1162 	intr_idx = vi->first_intr;
1163 
1164 #ifdef DEV_NETMAP
1165 	saved_idx = intr_idx;
1166 	if (ifp->if_capabilities & IFCAP_NETMAP) {
1167 
1168 		/* netmap is supported with direct interrupts only. */
1169 		MPASS(!forwarding_intr_to_fwq(sc));
1170 
1171 		/*
1172 		 * We don't have buffers to back the netmap rx queues
1173 		 * right now so we create the queues in a way that
1174 		 * doesn't set off any congestion signal in the chip.
1175 		 */
1176 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_rxq",
1177 		    CTLFLAG_RD, NULL, "rx queues");
1178 		for_each_nm_rxq(vi, i, nm_rxq) {
1179 			rc = alloc_nm_rxq(vi, nm_rxq, intr_idx, i, oid);
1180 			if (rc != 0)
1181 				goto done;
1182 			intr_idx++;
1183 		}
1184 
1185 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_txq",
1186 		    CTLFLAG_RD, NULL, "tx queues");
1187 		for_each_nm_txq(vi, i, nm_txq) {
1188 			iqidx = vi->first_nm_rxq + (i % vi->nnmrxq);
1189 			rc = alloc_nm_txq(vi, nm_txq, iqidx, i, oid);
1190 			if (rc != 0)
1191 				goto done;
1192 		}
1193 	}
1194 
1195 	/* Normal rx queues and netmap rx queues share the same interrupts. */
1196 	intr_idx = saved_idx;
1197 #endif
1198 
1199 	/*
1200 	 * Allocate rx queues first because a default iqid is required when
1201 	 * creating a tx queue.
1202 	 */
1203 	maxp = mtu_to_max_payload(sc, mtu, 0);
1204 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
1205 	    CTLFLAG_RD, NULL, "rx queues");
1206 	for_each_rxq(vi, i, rxq) {
1207 
1208 		init_iq(&rxq->iq, sc, vi->tmr_idx, vi->pktc_idx, vi->qsize_rxq);
1209 
1210 		snprintf(name, sizeof(name), "%s rxq%d-fl",
1211 		    device_get_nameunit(vi->dev), i);
1212 		init_fl(sc, &rxq->fl, vi->qsize_rxq / 8, maxp, name);
1213 
1214 		rc = alloc_rxq(vi, rxq,
1215 		    forwarding_intr_to_fwq(sc) ? -1 : intr_idx, i, oid);
1216 		if (rc != 0)
1217 			goto done;
1218 		intr_idx++;
1219 	}
1220 #ifdef DEV_NETMAP
1221 	if (ifp->if_capabilities & IFCAP_NETMAP)
1222 		intr_idx = saved_idx + max(vi->nrxq, vi->nnmrxq);
1223 #endif
1224 #ifdef TCP_OFFLOAD
1225 	maxp = mtu_to_max_payload(sc, mtu, 1);
1226 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq",
1227 	    CTLFLAG_RD, NULL, "rx queues for offloaded TCP connections");
1228 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1229 
1230 		init_iq(&ofld_rxq->iq, sc, vi->ofld_tmr_idx, vi->ofld_pktc_idx,
1231 		    vi->qsize_rxq);
1232 
1233 		snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
1234 		    device_get_nameunit(vi->dev), i);
1235 		init_fl(sc, &ofld_rxq->fl, vi->qsize_rxq / 8, maxp, name);
1236 
1237 		rc = alloc_ofld_rxq(vi, ofld_rxq,
1238 		    forwarding_intr_to_fwq(sc) ? -1 : intr_idx, i, oid);
1239 		if (rc != 0)
1240 			goto done;
1241 		intr_idx++;
1242 	}
1243 #endif
1244 
1245 	/*
1246 	 * Now the tx queues.
1247 	 */
1248 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "txq", CTLFLAG_RD,
1249 	    NULL, "tx queues");
1250 	for_each_txq(vi, i, txq) {
1251 		iqidx = vi->first_rxq + (i % vi->nrxq);
1252 		snprintf(name, sizeof(name), "%s txq%d",
1253 		    device_get_nameunit(vi->dev), i);
1254 		init_eq(sc, &txq->eq, EQ_ETH, vi->qsize_txq, pi->tx_chan,
1255 		    sc->sge.rxq[iqidx].iq.cntxt_id, name);
1256 
1257 		rc = alloc_txq(vi, txq, i, oid);
1258 		if (rc != 0)
1259 			goto done;
1260 	}
1261 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1262 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_txq",
1263 	    CTLFLAG_RD, NULL, "tx queues for TOE/ETHOFLD");
1264 	for_each_ofld_txq(vi, i, ofld_txq) {
1265 		struct sysctl_oid *oid2;
1266 
1267 		snprintf(name, sizeof(name), "%s ofld_txq%d",
1268 		    device_get_nameunit(vi->dev), i);
1269 		if (vi->nofldrxq > 0) {
1270 			iqidx = vi->first_ofld_rxq + (i % vi->nofldrxq);
1271 			init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq,
1272 			    pi->tx_chan, sc->sge.ofld_rxq[iqidx].iq.cntxt_id,
1273 			    name);
1274 		} else {
1275 			iqidx = vi->first_rxq + (i % vi->nrxq);
1276 			init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq,
1277 			    pi->tx_chan, sc->sge.rxq[iqidx].iq.cntxt_id, name);
1278 		}
1279 
1280 		snprintf(name, sizeof(name), "%d", i);
1281 		oid2 = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
1282 		    name, CTLFLAG_RD, NULL, "offload tx queue");
1283 
1284 		rc = alloc_wrq(sc, vi, ofld_txq, oid2);
1285 		if (rc != 0)
1286 			goto done;
1287 	}
1288 #endif
1289 done:
1290 	if (rc)
1291 		t4_teardown_vi_queues(vi);
1292 
1293 	return (rc);
1294 }
1295 
1296 /*
1297  * Idempotent
1298  */
1299 int
1300 t4_teardown_vi_queues(struct vi_info *vi)
1301 {
1302 	int i;
1303 	struct sge_rxq *rxq;
1304 	struct sge_txq *txq;
1305 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1306 	struct port_info *pi = vi->pi;
1307 	struct adapter *sc = pi->adapter;
1308 	struct sge_wrq *ofld_txq;
1309 #endif
1310 #ifdef TCP_OFFLOAD
1311 	struct sge_ofld_rxq *ofld_rxq;
1312 #endif
1313 #ifdef DEV_NETMAP
1314 	struct sge_nm_rxq *nm_rxq;
1315 	struct sge_nm_txq *nm_txq;
1316 #endif
1317 
1318 	/* Do this before freeing the queues */
1319 	if (vi->flags & VI_SYSCTL_CTX) {
1320 		sysctl_ctx_free(&vi->ctx);
1321 		vi->flags &= ~VI_SYSCTL_CTX;
1322 	}
1323 
1324 #ifdef DEV_NETMAP
1325 	if (vi->ifp->if_capabilities & IFCAP_NETMAP) {
1326 		for_each_nm_txq(vi, i, nm_txq) {
1327 			free_nm_txq(vi, nm_txq);
1328 		}
1329 
1330 		for_each_nm_rxq(vi, i, nm_rxq) {
1331 			free_nm_rxq(vi, nm_rxq);
1332 		}
1333 	}
1334 #endif
1335 
1336 	/*
1337 	 * Take down all the tx queues first, as they reference the rx queues
1338 	 * (for egress updates, etc.).
1339 	 */
1340 
1341 	for_each_txq(vi, i, txq) {
1342 		free_txq(vi, txq);
1343 	}
1344 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1345 	for_each_ofld_txq(vi, i, ofld_txq) {
1346 		free_wrq(sc, ofld_txq);
1347 	}
1348 #endif
1349 
1350 	/*
1351 	 * Then take down the rx queues.
1352 	 */
1353 
1354 	for_each_rxq(vi, i, rxq) {
1355 		free_rxq(vi, rxq);
1356 	}
1357 #ifdef TCP_OFFLOAD
1358 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1359 		free_ofld_rxq(vi, ofld_rxq);
1360 	}
1361 #endif
1362 
1363 	return (0);
1364 }
1365 
1366 /*
1367  * Interrupt handler when the driver is using only 1 interrupt.  This is a very
1368  * unusual scenario.
1369  *
1370  * a) Deals with errors, if any.
1371  * b) Services firmware event queue, which is taking interrupts for all other
1372  *    queues.
1373  */
1374 void
1375 t4_intr_all(void *arg)
1376 {
1377 	struct adapter *sc = arg;
1378 	struct sge_iq *fwq = &sc->sge.fwq;
1379 
1380 	MPASS(sc->intr_count == 1);
1381 
1382 	if (sc->intr_type == INTR_INTX)
1383 		t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
1384 
1385 	t4_intr_err(arg);
1386 	t4_intr_evt(fwq);
1387 }
1388 
1389 /*
1390  * Interrupt handler for errors (installed directly when multiple interrupts are
1391  * being used, or called by t4_intr_all).
1392  */
1393 void
1394 t4_intr_err(void *arg)
1395 {
1396 	struct adapter *sc = arg;
1397 	const bool verbose = (sc->debug_flags & DF_VERBOSE_SLOWINTR) != 0;
1398 
1399 	if (sc->flags & ADAP_ERR)
1400 		return;
1401 
1402 	t4_slow_intr_handler(sc, verbose);
1403 }
1404 
1405 /*
1406  * Interrupt handler for iq-only queues.  The firmware event queue is the only
1407  * such queue right now.
1408  */
1409 void
1410 t4_intr_evt(void *arg)
1411 {
1412 	struct sge_iq *iq = arg;
1413 
1414 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1415 		service_iq(iq, 0);
1416 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1417 	}
1418 }
1419 
1420 /*
1421  * Interrupt handler for iq+fl queues.
1422  */
1423 void
1424 t4_intr(void *arg)
1425 {
1426 	struct sge_iq *iq = arg;
1427 
1428 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1429 		service_iq_fl(iq, 0);
1430 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1431 	}
1432 }
1433 
1434 #ifdef DEV_NETMAP
1435 /*
1436  * Interrupt handler for netmap rx queues.
1437  */
1438 void
1439 t4_nm_intr(void *arg)
1440 {
1441 	struct sge_nm_rxq *nm_rxq = arg;
1442 
1443 	if (atomic_cmpset_int(&nm_rxq->nm_state, NM_ON, NM_BUSY)) {
1444 		service_nm_rxq(nm_rxq);
1445 		(void) atomic_cmpset_int(&nm_rxq->nm_state, NM_BUSY, NM_ON);
1446 	}
1447 }
1448 
1449 /*
1450  * Interrupt handler for vectors shared between NIC and netmap rx queues.
1451  */
1452 void
1453 t4_vi_intr(void *arg)
1454 {
1455 	struct irq *irq = arg;
1456 
1457 	MPASS(irq->nm_rxq != NULL);
1458 	t4_nm_intr(irq->nm_rxq);
1459 
1460 	MPASS(irq->rxq != NULL);
1461 	t4_intr(irq->rxq);
1462 }
1463 #endif
1464 
1465 /*
1466  * Deals with interrupts on an iq-only (no freelist) queue.
1467  */
1468 static int
1469 service_iq(struct sge_iq *iq, int budget)
1470 {
1471 	struct sge_iq *q;
1472 	struct adapter *sc = iq->adapter;
1473 	struct iq_desc *d = &iq->desc[iq->cidx];
1474 	int ndescs = 0, limit;
1475 	int rsp_type;
1476 	uint32_t lq;
1477 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1478 
1479 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1480 	KASSERT((iq->flags & IQ_HAS_FL) == 0,
1481 	    ("%s: called for iq %p with fl (iq->flags 0x%x)", __func__, iq,
1482 	    iq->flags));
1483 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1484 	MPASS((iq->flags & IQ_LRO_ENABLED) == 0);
1485 
1486 	limit = budget ? budget : iq->qsize / 16;
1487 
1488 	/*
1489 	 * We always come back and check the descriptor ring for new indirect
1490 	 * interrupts and other responses after running a single handler.
1491 	 */
1492 	for (;;) {
1493 		while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1494 
1495 			rmb();
1496 
1497 			rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1498 			lq = be32toh(d->rsp.pldbuflen_qid);
1499 
1500 			switch (rsp_type) {
1501 			case X_RSPD_TYPE_FLBUF:
1502 				panic("%s: data for an iq (%p) with no freelist",
1503 				    __func__, iq);
1504 
1505 				/* NOTREACHED */
1506 
1507 			case X_RSPD_TYPE_CPL:
1508 				KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1509 				    ("%s: bad opcode %02x.", __func__,
1510 				    d->rss.opcode));
1511 				t4_cpl_handler[d->rss.opcode](iq, &d->rss, NULL);
1512 				break;
1513 
1514 			case X_RSPD_TYPE_INTR:
1515 				/*
1516 				 * There are 1K interrupt-capable queues (qids 0
1517 				 * through 1023).  A response type indicating a
1518 				 * forwarded interrupt with a qid >= 1K is an
1519 				 * iWARP async notification.
1520 				 */
1521 				if (__predict_true(lq >= 1024)) {
1522 					t4_an_handler(iq, &d->rsp);
1523 					break;
1524 				}
1525 
1526 				q = sc->sge.iqmap[lq - sc->sge.iq_start -
1527 				    sc->sge.iq_base];
1528 				if (atomic_cmpset_int(&q->state, IQS_IDLE,
1529 				    IQS_BUSY)) {
1530 					if (service_iq_fl(q, q->qsize / 16) == 0) {
1531 						(void) atomic_cmpset_int(&q->state,
1532 						    IQS_BUSY, IQS_IDLE);
1533 					} else {
1534 						STAILQ_INSERT_TAIL(&iql, q,
1535 						    link);
1536 					}
1537 				}
1538 				break;
1539 
1540 			default:
1541 				KASSERT(0,
1542 				    ("%s: illegal response type %d on iq %p",
1543 				    __func__, rsp_type, iq));
1544 				log(LOG_ERR,
1545 				    "%s: illegal response type %d on iq %p",
1546 				    device_get_nameunit(sc->dev), rsp_type, iq);
1547 				break;
1548 			}
1549 
1550 			d++;
1551 			if (__predict_false(++iq->cidx == iq->sidx)) {
1552 				iq->cidx = 0;
1553 				iq->gen ^= F_RSPD_GEN;
1554 				d = &iq->desc[0];
1555 			}
1556 			if (__predict_false(++ndescs == limit)) {
1557 				t4_write_reg(sc, sc->sge_gts_reg,
1558 				    V_CIDXINC(ndescs) |
1559 				    V_INGRESSQID(iq->cntxt_id) |
1560 				    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1561 				ndescs = 0;
1562 
1563 				if (budget) {
1564 					return (EINPROGRESS);
1565 				}
1566 			}
1567 		}
1568 
1569 		if (STAILQ_EMPTY(&iql))
1570 			break;
1571 
1572 		/*
1573 		 * Process the head only, and send it to the back of the list if
1574 		 * it's still not done.
1575 		 */
1576 		q = STAILQ_FIRST(&iql);
1577 		STAILQ_REMOVE_HEAD(&iql, link);
1578 		if (service_iq_fl(q, q->qsize / 8) == 0)
1579 			(void) atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1580 		else
1581 			STAILQ_INSERT_TAIL(&iql, q, link);
1582 	}
1583 
1584 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1585 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1586 
1587 	return (0);
1588 }
1589 
1590 static inline int
1591 sort_before_lro(struct lro_ctrl *lro)
1592 {
1593 
1594 	return (lro->lro_mbuf_max != 0);
1595 }
1596 
1597 static inline uint64_t
1598 last_flit_to_ns(struct adapter *sc, uint64_t lf)
1599 {
1600 	uint64_t n = be64toh(lf) & 0xfffffffffffffff;	/* 60b, not 64b. */
1601 
1602 	if (n > UINT64_MAX / 1000000)
1603 		return (n / sc->params.vpd.cclk * 1000000);
1604 	else
1605 		return (n * 1000000 / sc->params.vpd.cclk);
1606 }
1607 
1608 /*
1609  * Deals with interrupts on an iq+fl queue.
1610  */
1611 static int
1612 service_iq_fl(struct sge_iq *iq, int budget)
1613 {
1614 	struct sge_rxq *rxq = iq_to_rxq(iq);
1615 	struct sge_fl *fl;
1616 	struct adapter *sc = iq->adapter;
1617 	struct iq_desc *d = &iq->desc[iq->cidx];
1618 	int ndescs = 0, limit;
1619 	int rsp_type, refill, starved;
1620 	uint32_t lq;
1621 	uint16_t fl_hw_cidx;
1622 	struct mbuf *m0;
1623 #if defined(INET) || defined(INET6)
1624 	const struct timeval lro_timeout = {0, sc->lro_timeout};
1625 	struct lro_ctrl *lro = &rxq->lro;
1626 #endif
1627 
1628 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1629 	MPASS(iq->flags & IQ_HAS_FL);
1630 
1631 	limit = budget ? budget : iq->qsize / 16;
1632 	fl = &rxq->fl;
1633 	fl_hw_cidx = fl->hw_cidx;	/* stable snapshot */
1634 
1635 #if defined(INET) || defined(INET6)
1636 	if (iq->flags & IQ_ADJ_CREDIT) {
1637 		MPASS(sort_before_lro(lro));
1638 		iq->flags &= ~IQ_ADJ_CREDIT;
1639 		if ((d->rsp.u.type_gen & F_RSPD_GEN) != iq->gen) {
1640 			tcp_lro_flush_all(lro);
1641 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(1) |
1642 			    V_INGRESSQID((u32)iq->cntxt_id) |
1643 			    V_SEINTARM(iq->intr_params));
1644 			return (0);
1645 		}
1646 		ndescs = 1;
1647 	}
1648 #else
1649 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1650 #endif
1651 
1652 	while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1653 
1654 		rmb();
1655 
1656 		refill = 0;
1657 		m0 = NULL;
1658 		rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1659 		lq = be32toh(d->rsp.pldbuflen_qid);
1660 
1661 		switch (rsp_type) {
1662 		case X_RSPD_TYPE_FLBUF:
1663 
1664 			m0 = get_fl_payload(sc, fl, lq);
1665 			if (__predict_false(m0 == NULL))
1666 				goto out;
1667 			refill = IDXDIFF(fl->hw_cidx, fl_hw_cidx, fl->sidx) > 2;
1668 
1669 			if (iq->flags & IQ_RX_TIMESTAMP) {
1670 				/*
1671 				 * Fill up rcv_tstmp but do not set M_TSTMP.
1672 				 * rcv_tstmp is not in the format that the
1673 				 * kernel expects and we don't want to mislead
1674 				 * it.  For now this is only for custom code
1675 				 * that knows how to interpret cxgbe's stamp.
1676 				 */
1677 				m0->m_pkthdr.rcv_tstmp =
1678 				    last_flit_to_ns(sc, d->rsp.u.last_flit);
1679 #ifdef notyet
1680 				m0->m_flags |= M_TSTMP;
1681 #endif
1682 			}
1683 
1684 			/* fall through */
1685 
1686 		case X_RSPD_TYPE_CPL:
1687 			KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1688 			    ("%s: bad opcode %02x.", __func__, d->rss.opcode));
1689 			t4_cpl_handler[d->rss.opcode](iq, &d->rss, m0);
1690 			break;
1691 
1692 		case X_RSPD_TYPE_INTR:
1693 
1694 			/*
1695 			 * There are 1K interrupt-capable queues (qids 0
1696 			 * through 1023).  A response type indicating a
1697 			 * forwarded interrupt with a qid >= 1K is an
1698 			 * iWARP async notification.  That is the only
1699 			 * acceptable indirect interrupt on this queue.
1700 			 */
1701 			if (__predict_false(lq < 1024)) {
1702 				panic("%s: indirect interrupt on iq_fl %p "
1703 				    "with qid %u", __func__, iq, lq);
1704 			}
1705 
1706 			t4_an_handler(iq, &d->rsp);
1707 			break;
1708 
1709 		default:
1710 			KASSERT(0, ("%s: illegal response type %d on iq %p",
1711 			    __func__, rsp_type, iq));
1712 			log(LOG_ERR, "%s: illegal response type %d on iq %p",
1713 			    device_get_nameunit(sc->dev), rsp_type, iq);
1714 			break;
1715 		}
1716 
1717 		d++;
1718 		if (__predict_false(++iq->cidx == iq->sidx)) {
1719 			iq->cidx = 0;
1720 			iq->gen ^= F_RSPD_GEN;
1721 			d = &iq->desc[0];
1722 		}
1723 		if (__predict_false(++ndescs == limit)) {
1724 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1725 			    V_INGRESSQID(iq->cntxt_id) |
1726 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1727 			ndescs = 0;
1728 
1729 #if defined(INET) || defined(INET6)
1730 			if (iq->flags & IQ_LRO_ENABLED &&
1731 			    !sort_before_lro(lro) &&
1732 			    sc->lro_timeout != 0) {
1733 				tcp_lro_flush_inactive(lro, &lro_timeout);
1734 			}
1735 #endif
1736 			if (budget) {
1737 				FL_LOCK(fl);
1738 				refill_fl(sc, fl, 32);
1739 				FL_UNLOCK(fl);
1740 
1741 				return (EINPROGRESS);
1742 			}
1743 		}
1744 		if (refill) {
1745 			FL_LOCK(fl);
1746 			refill_fl(sc, fl, 32);
1747 			FL_UNLOCK(fl);
1748 			fl_hw_cidx = fl->hw_cidx;
1749 		}
1750 	}
1751 out:
1752 #if defined(INET) || defined(INET6)
1753 	if (iq->flags & IQ_LRO_ENABLED) {
1754 		if (ndescs > 0 && lro->lro_mbuf_count > 8) {
1755 			MPASS(sort_before_lro(lro));
1756 			/* hold back one credit and don't flush LRO state */
1757 			iq->flags |= IQ_ADJ_CREDIT;
1758 			ndescs--;
1759 		} else {
1760 			tcp_lro_flush_all(lro);
1761 		}
1762 	}
1763 #endif
1764 
1765 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1766 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1767 
1768 	FL_LOCK(fl);
1769 	starved = refill_fl(sc, fl, 64);
1770 	FL_UNLOCK(fl);
1771 	if (__predict_false(starved != 0))
1772 		add_fl_to_sfl(sc, fl);
1773 
1774 	return (0);
1775 }
1776 
1777 static inline int
1778 cl_has_metadata(struct sge_fl *fl, struct cluster_layout *cll)
1779 {
1780 	int rc = fl->flags & FL_BUF_PACKING || cll->region1 > 0;
1781 
1782 	if (rc)
1783 		MPASS(cll->region3 >= CL_METADATA_SIZE);
1784 
1785 	return (rc);
1786 }
1787 
1788 static inline struct cluster_metadata *
1789 cl_metadata(struct adapter *sc, struct sge_fl *fl, struct cluster_layout *cll,
1790     caddr_t cl)
1791 {
1792 
1793 	if (cl_has_metadata(fl, cll)) {
1794 		struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1795 
1796 		return ((struct cluster_metadata *)(cl + swz->size) - 1);
1797 	}
1798 	return (NULL);
1799 }
1800 
1801 static void
1802 rxb_free(struct mbuf *m)
1803 {
1804 	uma_zone_t zone = m->m_ext.ext_arg1;
1805 	void *cl = m->m_ext.ext_arg2;
1806 
1807 	uma_zfree(zone, cl);
1808 	counter_u64_add(extfree_rels, 1);
1809 }
1810 
1811 /*
1812  * The mbuf returned by this function could be allocated from zone_mbuf or
1813  * constructed in spare room in the cluster.
1814  *
1815  * The mbuf carries the payload in one of these ways
1816  * a) frame inside the mbuf (mbuf from zone_mbuf)
1817  * b) m_cljset (for clusters without metadata) zone_mbuf
1818  * c) m_extaddref (cluster with metadata) inline mbuf
1819  * d) m_extaddref (cluster with metadata) zone_mbuf
1820  */
1821 static struct mbuf *
1822 get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1823     int remaining)
1824 {
1825 	struct mbuf *m;
1826 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1827 	struct cluster_layout *cll = &sd->cll;
1828 	struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1829 	struct hw_buf_info *hwb = &sc->sge.hw_buf_info[cll->hwidx];
1830 	struct cluster_metadata *clm = cl_metadata(sc, fl, cll, sd->cl);
1831 	int len, blen;
1832 	caddr_t payload;
1833 
1834 	blen = hwb->size - fl->rx_offset;	/* max possible in this buf */
1835 	len = min(remaining, blen);
1836 	payload = sd->cl + cll->region1 + fl->rx_offset;
1837 	if (fl->flags & FL_BUF_PACKING) {
1838 		const u_int l = fr_offset + len;
1839 		const u_int pad = roundup2(l, fl->buf_boundary) - l;
1840 
1841 		if (fl->rx_offset + len + pad < hwb->size)
1842 			blen = len + pad;
1843 		MPASS(fl->rx_offset + blen <= hwb->size);
1844 	} else {
1845 		MPASS(fl->rx_offset == 0);	/* not packing */
1846 	}
1847 
1848 
1849 	if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) {
1850 
1851 		/*
1852 		 * Copy payload into a freshly allocated mbuf.
1853 		 */
1854 
1855 		m = fr_offset == 0 ?
1856 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1857 		if (m == NULL)
1858 			return (NULL);
1859 		fl->mbuf_allocated++;
1860 
1861 		/* copy data to mbuf */
1862 		bcopy(payload, mtod(m, caddr_t), len);
1863 
1864 	} else if (sd->nmbuf * MSIZE < cll->region1) {
1865 
1866 		/*
1867 		 * There's spare room in the cluster for an mbuf.  Create one
1868 		 * and associate it with the payload that's in the cluster.
1869 		 */
1870 
1871 		MPASS(clm != NULL);
1872 		m = (struct mbuf *)(sd->cl + sd->nmbuf * MSIZE);
1873 		/* No bzero required */
1874 		if (m_init(m, M_NOWAIT, MT_DATA,
1875 		    fr_offset == 0 ? M_PKTHDR | M_NOFREE : M_NOFREE))
1876 			return (NULL);
1877 		fl->mbuf_inlined++;
1878 		m_extaddref(m, payload, blen, &clm->refcount, rxb_free,
1879 		    swz->zone, sd->cl);
1880 		if (sd->nmbuf++ == 0)
1881 			counter_u64_add(extfree_refs, 1);
1882 
1883 	} else {
1884 
1885 		/*
1886 		 * Grab an mbuf from zone_mbuf and associate it with the
1887 		 * payload in the cluster.
1888 		 */
1889 
1890 		m = fr_offset == 0 ?
1891 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1892 		if (m == NULL)
1893 			return (NULL);
1894 		fl->mbuf_allocated++;
1895 		if (clm != NULL) {
1896 			m_extaddref(m, payload, blen, &clm->refcount,
1897 			    rxb_free, swz->zone, sd->cl);
1898 			if (sd->nmbuf++ == 0)
1899 				counter_u64_add(extfree_refs, 1);
1900 		} else {
1901 			m_cljset(m, sd->cl, swz->type);
1902 			sd->cl = NULL;	/* consumed, not a recycle candidate */
1903 		}
1904 	}
1905 	if (fr_offset == 0)
1906 		m->m_pkthdr.len = remaining;
1907 	m->m_len = len;
1908 
1909 	if (fl->flags & FL_BUF_PACKING) {
1910 		fl->rx_offset += blen;
1911 		MPASS(fl->rx_offset <= hwb->size);
1912 		if (fl->rx_offset < hwb->size)
1913 			return (m);	/* without advancing the cidx */
1914 	}
1915 
1916 	if (__predict_false(++fl->cidx % 8 == 0)) {
1917 		uint16_t cidx = fl->cidx / 8;
1918 
1919 		if (__predict_false(cidx == fl->sidx))
1920 			fl->cidx = cidx = 0;
1921 		fl->hw_cidx = cidx;
1922 	}
1923 	fl->rx_offset = 0;
1924 
1925 	return (m);
1926 }
1927 
1928 static struct mbuf *
1929 get_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf)
1930 {
1931 	struct mbuf *m0, *m, **pnext;
1932 	u_int remaining;
1933 	const u_int total = G_RSPD_LEN(len_newbuf);
1934 
1935 	if (__predict_false(fl->flags & FL_BUF_RESUME)) {
1936 		M_ASSERTPKTHDR(fl->m0);
1937 		MPASS(fl->m0->m_pkthdr.len == total);
1938 		MPASS(fl->remaining < total);
1939 
1940 		m0 = fl->m0;
1941 		pnext = fl->pnext;
1942 		remaining = fl->remaining;
1943 		fl->flags &= ~FL_BUF_RESUME;
1944 		goto get_segment;
1945 	}
1946 
1947 	if (fl->rx_offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
1948 		fl->rx_offset = 0;
1949 		if (__predict_false(++fl->cidx % 8 == 0)) {
1950 			uint16_t cidx = fl->cidx / 8;
1951 
1952 			if (__predict_false(cidx == fl->sidx))
1953 				fl->cidx = cidx = 0;
1954 			fl->hw_cidx = cidx;
1955 		}
1956 	}
1957 
1958 	/*
1959 	 * Payload starts at rx_offset in the current hw buffer.  Its length is
1960 	 * 'len' and it may span multiple hw buffers.
1961 	 */
1962 
1963 	m0 = get_scatter_segment(sc, fl, 0, total);
1964 	if (m0 == NULL)
1965 		return (NULL);
1966 	remaining = total - m0->m_len;
1967 	pnext = &m0->m_next;
1968 	while (remaining > 0) {
1969 get_segment:
1970 		MPASS(fl->rx_offset == 0);
1971 		m = get_scatter_segment(sc, fl, total - remaining, remaining);
1972 		if (__predict_false(m == NULL)) {
1973 			fl->m0 = m0;
1974 			fl->pnext = pnext;
1975 			fl->remaining = remaining;
1976 			fl->flags |= FL_BUF_RESUME;
1977 			return (NULL);
1978 		}
1979 		*pnext = m;
1980 		pnext = &m->m_next;
1981 		remaining -= m->m_len;
1982 	}
1983 	*pnext = NULL;
1984 
1985 	M_ASSERTPKTHDR(m0);
1986 	return (m0);
1987 }
1988 
1989 static int
1990 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
1991 {
1992 	struct sge_rxq *rxq = iq_to_rxq(iq);
1993 	struct ifnet *ifp = rxq->ifp;
1994 	struct adapter *sc = iq->adapter;
1995 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
1996 #if defined(INET) || defined(INET6)
1997 	struct lro_ctrl *lro = &rxq->lro;
1998 #endif
1999 	static const int sw_hashtype[4][2] = {
2000 		{M_HASHTYPE_NONE, M_HASHTYPE_NONE},
2001 		{M_HASHTYPE_RSS_IPV4, M_HASHTYPE_RSS_IPV6},
2002 		{M_HASHTYPE_RSS_TCP_IPV4, M_HASHTYPE_RSS_TCP_IPV6},
2003 		{M_HASHTYPE_RSS_UDP_IPV4, M_HASHTYPE_RSS_UDP_IPV6},
2004 	};
2005 
2006 	KASSERT(m0 != NULL, ("%s: no payload with opcode %02x", __func__,
2007 	    rss->opcode));
2008 
2009 	m0->m_pkthdr.len -= sc->params.sge.fl_pktshift;
2010 	m0->m_len -= sc->params.sge.fl_pktshift;
2011 	m0->m_data += sc->params.sge.fl_pktshift;
2012 
2013 	m0->m_pkthdr.rcvif = ifp;
2014 	M_HASHTYPE_SET(m0, sw_hashtype[rss->hash_type][rss->ipv6]);
2015 	m0->m_pkthdr.flowid = be32toh(rss->hash_val);
2016 
2017 	if (cpl->csum_calc && !(cpl->err_vec & sc->params.tp.err_vec_mask)) {
2018 		if (ifp->if_capenable & IFCAP_RXCSUM &&
2019 		    cpl->l2info & htobe32(F_RXF_IP)) {
2020 			m0->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
2021 			    CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
2022 			rxq->rxcsum++;
2023 		} else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
2024 		    cpl->l2info & htobe32(F_RXF_IP6)) {
2025 			m0->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
2026 			    CSUM_PSEUDO_HDR);
2027 			rxq->rxcsum++;
2028 		}
2029 
2030 		if (__predict_false(cpl->ip_frag))
2031 			m0->m_pkthdr.csum_data = be16toh(cpl->csum);
2032 		else
2033 			m0->m_pkthdr.csum_data = 0xffff;
2034 	}
2035 
2036 	if (cpl->vlan_ex) {
2037 		m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
2038 		m0->m_flags |= M_VLANTAG;
2039 		rxq->vlan_extraction++;
2040 	}
2041 
2042 #if defined(INET) || defined(INET6)
2043 	if (iq->flags & IQ_LRO_ENABLED) {
2044 		if (sort_before_lro(lro)) {
2045 			tcp_lro_queue_mbuf(lro, m0);
2046 			return (0); /* queued for sort, then LRO */
2047 		}
2048 		if (tcp_lro_rx(lro, m0, 0) == 0)
2049 			return (0); /* queued for LRO */
2050 	}
2051 #endif
2052 	ifp->if_input(ifp, m0);
2053 
2054 	return (0);
2055 }
2056 
2057 /*
2058  * Must drain the wrq or make sure that someone else will.
2059  */
2060 static void
2061 wrq_tx_drain(void *arg, int n)
2062 {
2063 	struct sge_wrq *wrq = arg;
2064 	struct sge_eq *eq = &wrq->eq;
2065 
2066 	EQ_LOCK(eq);
2067 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2068 		drain_wrq_wr_list(wrq->adapter, wrq);
2069 	EQ_UNLOCK(eq);
2070 }
2071 
2072 static void
2073 drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq)
2074 {
2075 	struct sge_eq *eq = &wrq->eq;
2076 	u_int available, dbdiff;	/* # of hardware descriptors */
2077 	u_int n;
2078 	struct wrqe *wr;
2079 	struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
2080 
2081 	EQ_LOCK_ASSERT_OWNED(eq);
2082 	MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
2083 	wr = STAILQ_FIRST(&wrq->wr_list);
2084 	MPASS(wr != NULL);	/* Must be called with something useful to do */
2085 	MPASS(eq->pidx == eq->dbidx);
2086 	dbdiff = 0;
2087 
2088 	do {
2089 		eq->cidx = read_hw_cidx(eq);
2090 		if (eq->pidx == eq->cidx)
2091 			available = eq->sidx - 1;
2092 		else
2093 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2094 
2095 		MPASS(wr->wrq == wrq);
2096 		n = howmany(wr->wr_len, EQ_ESIZE);
2097 		if (available < n)
2098 			break;
2099 
2100 		dst = (void *)&eq->desc[eq->pidx];
2101 		if (__predict_true(eq->sidx - eq->pidx > n)) {
2102 			/* Won't wrap, won't end exactly at the status page. */
2103 			bcopy(&wr->wr[0], dst, wr->wr_len);
2104 			eq->pidx += n;
2105 		} else {
2106 			int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE;
2107 
2108 			bcopy(&wr->wr[0], dst, first_portion);
2109 			if (wr->wr_len > first_portion) {
2110 				bcopy(&wr->wr[first_portion], &eq->desc[0],
2111 				    wr->wr_len - first_portion);
2112 			}
2113 			eq->pidx = n - (eq->sidx - eq->pidx);
2114 		}
2115 		wrq->tx_wrs_copied++;
2116 
2117 		if (available < eq->sidx / 4 &&
2118 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2119 				/*
2120 				 * XXX: This is not 100% reliable with some
2121 				 * types of WRs.  But this is a very unusual
2122 				 * situation for an ofld/ctrl queue anyway.
2123 				 */
2124 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2125 			    F_FW_WR_EQUEQ);
2126 		}
2127 
2128 		dbdiff += n;
2129 		if (dbdiff >= 16) {
2130 			ring_eq_db(sc, eq, dbdiff);
2131 			dbdiff = 0;
2132 		}
2133 
2134 		STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
2135 		free_wrqe(wr);
2136 		MPASS(wrq->nwr_pending > 0);
2137 		wrq->nwr_pending--;
2138 		MPASS(wrq->ndesc_needed >= n);
2139 		wrq->ndesc_needed -= n;
2140 	} while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL);
2141 
2142 	if (dbdiff)
2143 		ring_eq_db(sc, eq, dbdiff);
2144 }
2145 
2146 /*
2147  * Doesn't fail.  Holds on to work requests it can't send right away.
2148  */
2149 void
2150 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
2151 {
2152 #ifdef INVARIANTS
2153 	struct sge_eq *eq = &wrq->eq;
2154 #endif
2155 
2156 	EQ_LOCK_ASSERT_OWNED(eq);
2157 	MPASS(wr != NULL);
2158 	MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN);
2159 	MPASS((wr->wr_len & 0x7) == 0);
2160 
2161 	STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
2162 	wrq->nwr_pending++;
2163 	wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE);
2164 
2165 	if (!TAILQ_EMPTY(&wrq->incomplete_wrs))
2166 		return;	/* commit_wrq_wr will drain wr_list as well. */
2167 
2168 	drain_wrq_wr_list(sc, wrq);
2169 
2170 	/* Doorbell must have caught up to the pidx. */
2171 	MPASS(eq->pidx == eq->dbidx);
2172 }
2173 
2174 void
2175 t4_update_fl_bufsize(struct ifnet *ifp)
2176 {
2177 	struct vi_info *vi = ifp->if_softc;
2178 	struct adapter *sc = vi->pi->adapter;
2179 	struct sge_rxq *rxq;
2180 #ifdef TCP_OFFLOAD
2181 	struct sge_ofld_rxq *ofld_rxq;
2182 #endif
2183 	struct sge_fl *fl;
2184 	int i, maxp, mtu = ifp->if_mtu;
2185 
2186 	maxp = mtu_to_max_payload(sc, mtu, 0);
2187 	for_each_rxq(vi, i, rxq) {
2188 		fl = &rxq->fl;
2189 
2190 		FL_LOCK(fl);
2191 		find_best_refill_source(sc, fl, maxp);
2192 		FL_UNLOCK(fl);
2193 	}
2194 #ifdef TCP_OFFLOAD
2195 	maxp = mtu_to_max_payload(sc, mtu, 1);
2196 	for_each_ofld_rxq(vi, i, ofld_rxq) {
2197 		fl = &ofld_rxq->fl;
2198 
2199 		FL_LOCK(fl);
2200 		find_best_refill_source(sc, fl, maxp);
2201 		FL_UNLOCK(fl);
2202 	}
2203 #endif
2204 }
2205 
2206 static inline int
2207 mbuf_nsegs(struct mbuf *m)
2208 {
2209 
2210 	M_ASSERTPKTHDR(m);
2211 	KASSERT(m->m_pkthdr.l5hlen > 0,
2212 	    ("%s: mbuf %p missing information on # of segments.", __func__, m));
2213 
2214 	return (m->m_pkthdr.l5hlen);
2215 }
2216 
2217 static inline void
2218 set_mbuf_nsegs(struct mbuf *m, uint8_t nsegs)
2219 {
2220 
2221 	M_ASSERTPKTHDR(m);
2222 	m->m_pkthdr.l5hlen = nsegs;
2223 }
2224 
2225 static inline int
2226 mbuf_cflags(struct mbuf *m)
2227 {
2228 
2229 	M_ASSERTPKTHDR(m);
2230 	return (m->m_pkthdr.PH_loc.eight[4]);
2231 }
2232 
2233 static inline void
2234 set_mbuf_cflags(struct mbuf *m, uint8_t flags)
2235 {
2236 
2237 	M_ASSERTPKTHDR(m);
2238 	m->m_pkthdr.PH_loc.eight[4] = flags;
2239 }
2240 
2241 static inline int
2242 mbuf_len16(struct mbuf *m)
2243 {
2244 	int n;
2245 
2246 	M_ASSERTPKTHDR(m);
2247 	n = m->m_pkthdr.PH_loc.eight[0];
2248 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2249 
2250 	return (n);
2251 }
2252 
2253 static inline void
2254 set_mbuf_len16(struct mbuf *m, uint8_t len16)
2255 {
2256 
2257 	M_ASSERTPKTHDR(m);
2258 	m->m_pkthdr.PH_loc.eight[0] = len16;
2259 }
2260 
2261 #ifdef RATELIMIT
2262 static inline int
2263 mbuf_eo_nsegs(struct mbuf *m)
2264 {
2265 
2266 	M_ASSERTPKTHDR(m);
2267 	return (m->m_pkthdr.PH_loc.eight[1]);
2268 }
2269 
2270 static inline void
2271 set_mbuf_eo_nsegs(struct mbuf *m, uint8_t nsegs)
2272 {
2273 
2274 	M_ASSERTPKTHDR(m);
2275 	m->m_pkthdr.PH_loc.eight[1] = nsegs;
2276 }
2277 
2278 static inline int
2279 mbuf_eo_len16(struct mbuf *m)
2280 {
2281 	int n;
2282 
2283 	M_ASSERTPKTHDR(m);
2284 	n = m->m_pkthdr.PH_loc.eight[2];
2285 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2286 
2287 	return (n);
2288 }
2289 
2290 static inline void
2291 set_mbuf_eo_len16(struct mbuf *m, uint8_t len16)
2292 {
2293 
2294 	M_ASSERTPKTHDR(m);
2295 	m->m_pkthdr.PH_loc.eight[2] = len16;
2296 }
2297 
2298 static inline int
2299 mbuf_eo_tsclk_tsoff(struct mbuf *m)
2300 {
2301 
2302 	M_ASSERTPKTHDR(m);
2303 	return (m->m_pkthdr.PH_loc.eight[3]);
2304 }
2305 
2306 static inline void
2307 set_mbuf_eo_tsclk_tsoff(struct mbuf *m, uint8_t tsclk_tsoff)
2308 {
2309 
2310 	M_ASSERTPKTHDR(m);
2311 	m->m_pkthdr.PH_loc.eight[3] = tsclk_tsoff;
2312 }
2313 
2314 static inline int
2315 needs_eo(struct mbuf *m)
2316 {
2317 
2318 	return (m->m_pkthdr.snd_tag != NULL);
2319 }
2320 #endif
2321 
2322 /*
2323  * Try to allocate an mbuf to contain a raw work request.  To make it
2324  * easy to construct the work request, don't allocate a chain but a
2325  * single mbuf.
2326  */
2327 struct mbuf *
2328 alloc_wr_mbuf(int len, int how)
2329 {
2330 	struct mbuf *m;
2331 
2332 	if (len <= MHLEN)
2333 		m = m_gethdr(how, MT_DATA);
2334 	else if (len <= MCLBYTES)
2335 		m = m_getcl(how, MT_DATA, M_PKTHDR);
2336 	else
2337 		m = NULL;
2338 	if (m == NULL)
2339 		return (NULL);
2340 	m->m_pkthdr.len = len;
2341 	m->m_len = len;
2342 	set_mbuf_cflags(m, MC_RAW_WR);
2343 	set_mbuf_len16(m, howmany(len, 16));
2344 	return (m);
2345 }
2346 
2347 static inline int
2348 needs_tso(struct mbuf *m)
2349 {
2350 
2351 	M_ASSERTPKTHDR(m);
2352 
2353 	return (m->m_pkthdr.csum_flags & CSUM_TSO);
2354 }
2355 
2356 static inline int
2357 needs_l3_csum(struct mbuf *m)
2358 {
2359 
2360 	M_ASSERTPKTHDR(m);
2361 
2362 	return (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO));
2363 }
2364 
2365 static inline int
2366 needs_l4_csum(struct mbuf *m)
2367 {
2368 
2369 	M_ASSERTPKTHDR(m);
2370 
2371 	return (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
2372 	    CSUM_TCP_IPV6 | CSUM_TSO));
2373 }
2374 
2375 static inline int
2376 needs_tcp_csum(struct mbuf *m)
2377 {
2378 
2379 	M_ASSERTPKTHDR(m);
2380 	return (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_TCP_IPV6 | CSUM_TSO));
2381 }
2382 
2383 #ifdef RATELIMIT
2384 static inline int
2385 needs_udp_csum(struct mbuf *m)
2386 {
2387 
2388 	M_ASSERTPKTHDR(m);
2389 	return (m->m_pkthdr.csum_flags & (CSUM_UDP | CSUM_UDP_IPV6));
2390 }
2391 #endif
2392 
2393 static inline int
2394 needs_vlan_insertion(struct mbuf *m)
2395 {
2396 
2397 	M_ASSERTPKTHDR(m);
2398 
2399 	return (m->m_flags & M_VLANTAG);
2400 }
2401 
2402 static void *
2403 m_advance(struct mbuf **pm, int *poffset, int len)
2404 {
2405 	struct mbuf *m = *pm;
2406 	int offset = *poffset;
2407 	uintptr_t p = 0;
2408 
2409 	MPASS(len > 0);
2410 
2411 	for (;;) {
2412 		if (offset + len < m->m_len) {
2413 			offset += len;
2414 			p = mtod(m, uintptr_t) + offset;
2415 			break;
2416 		}
2417 		len -= m->m_len - offset;
2418 		m = m->m_next;
2419 		offset = 0;
2420 		MPASS(m != NULL);
2421 	}
2422 	*poffset = offset;
2423 	*pm = m;
2424 	return ((void *)p);
2425 }
2426 
2427 /*
2428  * Can deal with empty mbufs in the chain that have m_len = 0, but the chain
2429  * must have at least one mbuf that's not empty.  It is possible for this
2430  * routine to return 0 if skip accounts for all the contents of the mbuf chain.
2431  */
2432 static inline int
2433 count_mbuf_nsegs(struct mbuf *m, int skip)
2434 {
2435 	vm_paddr_t lastb, next;
2436 	vm_offset_t va;
2437 	int len, nsegs;
2438 
2439 	M_ASSERTPKTHDR(m);
2440 	MPASS(m->m_pkthdr.len > 0);
2441 	MPASS(m->m_pkthdr.len >= skip);
2442 
2443 	nsegs = 0;
2444 	lastb = 0;
2445 	for (; m; m = m->m_next) {
2446 
2447 		len = m->m_len;
2448 		if (__predict_false(len == 0))
2449 			continue;
2450 		if (skip >= len) {
2451 			skip -= len;
2452 			continue;
2453 		}
2454 		va = mtod(m, vm_offset_t) + skip;
2455 		len -= skip;
2456 		skip = 0;
2457 		next = pmap_kextract(va);
2458 		nsegs += sglist_count((void *)(uintptr_t)va, len);
2459 		if (lastb + 1 == next)
2460 			nsegs--;
2461 		lastb = pmap_kextract(va + len - 1);
2462 	}
2463 
2464 	return (nsegs);
2465 }
2466 
2467 /*
2468  * Analyze the mbuf to determine its tx needs.  The mbuf passed in may change:
2469  * a) caller can assume it's been freed if this function returns with an error.
2470  * b) it may get defragged up if the gather list is too long for the hardware.
2471  */
2472 int
2473 parse_pkt(struct adapter *sc, struct mbuf **mp)
2474 {
2475 	struct mbuf *m0 = *mp, *m;
2476 	int rc, nsegs, defragged = 0, offset;
2477 	struct ether_header *eh;
2478 	void *l3hdr;
2479 #if defined(INET) || defined(INET6)
2480 	struct tcphdr *tcp;
2481 #endif
2482 	uint16_t eh_type;
2483 
2484 	M_ASSERTPKTHDR(m0);
2485 	if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) {
2486 		rc = EINVAL;
2487 fail:
2488 		m_freem(m0);
2489 		*mp = NULL;
2490 		return (rc);
2491 	}
2492 restart:
2493 	/*
2494 	 * First count the number of gather list segments in the payload.
2495 	 * Defrag the mbuf if nsegs exceeds the hardware limit.
2496 	 */
2497 	M_ASSERTPKTHDR(m0);
2498 	MPASS(m0->m_pkthdr.len > 0);
2499 	nsegs = count_mbuf_nsegs(m0, 0);
2500 	if (nsegs > (needs_tso(m0) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS)) {
2501 		if (defragged++ > 0 || (m = m_defrag(m0, M_NOWAIT)) == NULL) {
2502 			rc = EFBIG;
2503 			goto fail;
2504 		}
2505 		*mp = m0 = m;	/* update caller's copy after defrag */
2506 		goto restart;
2507 	}
2508 
2509 	if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN)) {
2510 		m0 = m_pullup(m0, m0->m_pkthdr.len);
2511 		if (m0 == NULL) {
2512 			/* Should have left well enough alone. */
2513 			rc = EFBIG;
2514 			goto fail;
2515 		}
2516 		*mp = m0;	/* update caller's copy after pullup */
2517 		goto restart;
2518 	}
2519 	set_mbuf_nsegs(m0, nsegs);
2520 	set_mbuf_cflags(m0, 0);
2521 	if (sc->flags & IS_VF)
2522 		set_mbuf_len16(m0, txpkt_vm_len16(nsegs, needs_tso(m0)));
2523 	else
2524 		set_mbuf_len16(m0, txpkt_len16(nsegs, needs_tso(m0)));
2525 
2526 #ifdef RATELIMIT
2527 	/*
2528 	 * Ethofld is limited to TCP and UDP for now, and only when L4 hw
2529 	 * checksumming is enabled.  needs_l4_csum happens to check for all the
2530 	 * right things.
2531 	 */
2532 	if (__predict_false(needs_eo(m0) && !needs_l4_csum(m0)))
2533 		m0->m_pkthdr.snd_tag = NULL;
2534 #endif
2535 
2536 	if (!needs_tso(m0) &&
2537 #ifdef RATELIMIT
2538 	    !needs_eo(m0) &&
2539 #endif
2540 	    !(sc->flags & IS_VF && (needs_l3_csum(m0) || needs_l4_csum(m0))))
2541 		return (0);
2542 
2543 	m = m0;
2544 	eh = mtod(m, struct ether_header *);
2545 	eh_type = ntohs(eh->ether_type);
2546 	if (eh_type == ETHERTYPE_VLAN) {
2547 		struct ether_vlan_header *evh = (void *)eh;
2548 
2549 		eh_type = ntohs(evh->evl_proto);
2550 		m0->m_pkthdr.l2hlen = sizeof(*evh);
2551 	} else
2552 		m0->m_pkthdr.l2hlen = sizeof(*eh);
2553 
2554 	offset = 0;
2555 	l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2556 
2557 	switch (eh_type) {
2558 #ifdef INET6
2559 	case ETHERTYPE_IPV6:
2560 	{
2561 		struct ip6_hdr *ip6 = l3hdr;
2562 
2563 		MPASS(!needs_tso(m0) || ip6->ip6_nxt == IPPROTO_TCP);
2564 
2565 		m0->m_pkthdr.l3hlen = sizeof(*ip6);
2566 		break;
2567 	}
2568 #endif
2569 #ifdef INET
2570 	case ETHERTYPE_IP:
2571 	{
2572 		struct ip *ip = l3hdr;
2573 
2574 		m0->m_pkthdr.l3hlen = ip->ip_hl * 4;
2575 		break;
2576 	}
2577 #endif
2578 	default:
2579 		panic("%s: ethertype 0x%04x unknown.  if_cxgbe must be compiled"
2580 		    " with the same INET/INET6 options as the kernel.",
2581 		    __func__, eh_type);
2582 	}
2583 
2584 #if defined(INET) || defined(INET6)
2585 	if (needs_tcp_csum(m0)) {
2586 		tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen);
2587 		m0->m_pkthdr.l4hlen = tcp->th_off * 4;
2588 #ifdef RATELIMIT
2589 		if (tsclk >= 0 && *(uint32_t *)(tcp + 1) == ntohl(0x0101080a)) {
2590 			set_mbuf_eo_tsclk_tsoff(m0,
2591 			    V_FW_ETH_TX_EO_WR_TSCLK(tsclk) |
2592 			    V_FW_ETH_TX_EO_WR_TSOFF(sizeof(*tcp) / 2 + 1));
2593 		} else
2594 			set_mbuf_eo_tsclk_tsoff(m0, 0);
2595 	} else if (needs_udp_csum(m)) {
2596 		m0->m_pkthdr.l4hlen = sizeof(struct udphdr);
2597 #endif
2598 	}
2599 #ifdef RATELIMIT
2600 	if (needs_eo(m0)) {
2601 		u_int immhdrs;
2602 
2603 		/* EO WRs have the headers in the WR and not the GL. */
2604 		immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen +
2605 		    m0->m_pkthdr.l4hlen;
2606 		nsegs = count_mbuf_nsegs(m0, immhdrs);
2607 		set_mbuf_eo_nsegs(m0, nsegs);
2608 		set_mbuf_eo_len16(m0,
2609 		    txpkt_eo_len16(nsegs, immhdrs, needs_tso(m0)));
2610 	}
2611 #endif
2612 #endif
2613 	MPASS(m0 == *mp);
2614 	return (0);
2615 }
2616 
2617 void *
2618 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie)
2619 {
2620 	struct sge_eq *eq = &wrq->eq;
2621 	struct adapter *sc = wrq->adapter;
2622 	int ndesc, available;
2623 	struct wrqe *wr;
2624 	void *w;
2625 
2626 	MPASS(len16 > 0);
2627 	ndesc = howmany(len16, EQ_ESIZE / 16);
2628 	MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC);
2629 
2630 	EQ_LOCK(eq);
2631 
2632 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2633 		drain_wrq_wr_list(sc, wrq);
2634 
2635 	if (!STAILQ_EMPTY(&wrq->wr_list)) {
2636 slowpath:
2637 		EQ_UNLOCK(eq);
2638 		wr = alloc_wrqe(len16 * 16, wrq);
2639 		if (__predict_false(wr == NULL))
2640 			return (NULL);
2641 		cookie->pidx = -1;
2642 		cookie->ndesc = ndesc;
2643 		return (&wr->wr);
2644 	}
2645 
2646 	eq->cidx = read_hw_cidx(eq);
2647 	if (eq->pidx == eq->cidx)
2648 		available = eq->sidx - 1;
2649 	else
2650 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2651 	if (available < ndesc)
2652 		goto slowpath;
2653 
2654 	cookie->pidx = eq->pidx;
2655 	cookie->ndesc = ndesc;
2656 	TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link);
2657 
2658 	w = &eq->desc[eq->pidx];
2659 	IDXINCR(eq->pidx, ndesc, eq->sidx);
2660 	if (__predict_false(cookie->pidx + ndesc > eq->sidx)) {
2661 		w = &wrq->ss[0];
2662 		wrq->ss_pidx = cookie->pidx;
2663 		wrq->ss_len = len16 * 16;
2664 	}
2665 
2666 	EQ_UNLOCK(eq);
2667 
2668 	return (w);
2669 }
2670 
2671 void
2672 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie)
2673 {
2674 	struct sge_eq *eq = &wrq->eq;
2675 	struct adapter *sc = wrq->adapter;
2676 	int ndesc, pidx;
2677 	struct wrq_cookie *prev, *next;
2678 
2679 	if (cookie->pidx == -1) {
2680 		struct wrqe *wr = __containerof(w, struct wrqe, wr);
2681 
2682 		t4_wrq_tx(sc, wr);
2683 		return;
2684 	}
2685 
2686 	if (__predict_false(w == &wrq->ss[0])) {
2687 		int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE;
2688 
2689 		MPASS(wrq->ss_len > n);	/* WR had better wrap around. */
2690 		bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n);
2691 		bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n);
2692 		wrq->tx_wrs_ss++;
2693 	} else
2694 		wrq->tx_wrs_direct++;
2695 
2696 	EQ_LOCK(eq);
2697 	ndesc = cookie->ndesc;	/* Can be more than SGE_MAX_WR_NDESC here. */
2698 	pidx = cookie->pidx;
2699 	MPASS(pidx >= 0 && pidx < eq->sidx);
2700 	prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link);
2701 	next = TAILQ_NEXT(cookie, link);
2702 	if (prev == NULL) {
2703 		MPASS(pidx == eq->dbidx);
2704 		if (next == NULL || ndesc >= 16) {
2705 			int available;
2706 			struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
2707 
2708 			/*
2709 			 * Note that the WR via which we'll request tx updates
2710 			 * is at pidx and not eq->pidx, which has moved on
2711 			 * already.
2712 			 */
2713 			dst = (void *)&eq->desc[pidx];
2714 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2715 			if (available < eq->sidx / 4 &&
2716 			    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2717 				/*
2718 				 * XXX: This is not 100% reliable with some
2719 				 * types of WRs.  But this is a very unusual
2720 				 * situation for an ofld/ctrl queue anyway.
2721 				 */
2722 				dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2723 				    F_FW_WR_EQUEQ);
2724 			}
2725 
2726 			ring_eq_db(wrq->adapter, eq, ndesc);
2727 		} else {
2728 			MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc);
2729 			next->pidx = pidx;
2730 			next->ndesc += ndesc;
2731 		}
2732 	} else {
2733 		MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc);
2734 		prev->ndesc += ndesc;
2735 	}
2736 	TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link);
2737 
2738 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2739 		drain_wrq_wr_list(sc, wrq);
2740 
2741 #ifdef INVARIANTS
2742 	if (TAILQ_EMPTY(&wrq->incomplete_wrs)) {
2743 		/* Doorbell must have caught up to the pidx. */
2744 		MPASS(wrq->eq.pidx == wrq->eq.dbidx);
2745 	}
2746 #endif
2747 	EQ_UNLOCK(eq);
2748 }
2749 
2750 static u_int
2751 can_resume_eth_tx(struct mp_ring *r)
2752 {
2753 	struct sge_eq *eq = r->cookie;
2754 
2755 	return (total_available_tx_desc(eq) > eq->sidx / 8);
2756 }
2757 
2758 static inline int
2759 cannot_use_txpkts(struct mbuf *m)
2760 {
2761 	/* maybe put a GL limit too, to avoid silliness? */
2762 
2763 	return (needs_tso(m) || (mbuf_cflags(m) & MC_RAW_WR) != 0);
2764 }
2765 
2766 static inline int
2767 discard_tx(struct sge_eq *eq)
2768 {
2769 
2770 	return ((eq->flags & (EQ_ENABLED | EQ_QFLUSH)) != EQ_ENABLED);
2771 }
2772 
2773 static inline int
2774 wr_can_update_eq(struct fw_eth_tx_pkts_wr *wr)
2775 {
2776 
2777 	switch (G_FW_WR_OP(be32toh(wr->op_pkd))) {
2778 	case FW_ULPTX_WR:
2779 	case FW_ETH_TX_PKT_WR:
2780 	case FW_ETH_TX_PKTS_WR:
2781 	case FW_ETH_TX_PKT_VM_WR:
2782 		return (1);
2783 	default:
2784 		return (0);
2785 	}
2786 }
2787 
2788 /*
2789  * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to
2790  * be consumed.  Return the actual number consumed.  0 indicates a stall.
2791  */
2792 static u_int
2793 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx)
2794 {
2795 	struct sge_txq *txq = r->cookie;
2796 	struct sge_eq *eq = &txq->eq;
2797 	struct ifnet *ifp = txq->ifp;
2798 	struct vi_info *vi = ifp->if_softc;
2799 	struct port_info *pi = vi->pi;
2800 	struct adapter *sc = pi->adapter;
2801 	u_int total, remaining;		/* # of packets */
2802 	u_int available, dbdiff;	/* # of hardware descriptors */
2803 	u_int n, next_cidx;
2804 	struct mbuf *m0, *tail;
2805 	struct txpkts txp;
2806 	struct fw_eth_tx_pkts_wr *wr;	/* any fw WR struct will do */
2807 
2808 	remaining = IDXDIFF(pidx, cidx, r->size);
2809 	MPASS(remaining > 0);	/* Must not be called without work to do. */
2810 	total = 0;
2811 
2812 	TXQ_LOCK(txq);
2813 	if (__predict_false(discard_tx(eq))) {
2814 		while (cidx != pidx) {
2815 			m0 = r->items[cidx];
2816 			m_freem(m0);
2817 			if (++cidx == r->size)
2818 				cidx = 0;
2819 		}
2820 		reclaim_tx_descs(txq, 2048);
2821 		total = remaining;
2822 		goto done;
2823 	}
2824 
2825 	/* How many hardware descriptors do we have readily available. */
2826 	if (eq->pidx == eq->cidx)
2827 		available = eq->sidx - 1;
2828 	else
2829 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2830 	dbdiff = IDXDIFF(eq->pidx, eq->dbidx, eq->sidx);
2831 
2832 	while (remaining > 0) {
2833 
2834 		m0 = r->items[cidx];
2835 		M_ASSERTPKTHDR(m0);
2836 		MPASS(m0->m_nextpkt == NULL);
2837 
2838 		if (available < SGE_MAX_WR_NDESC) {
2839 			available += reclaim_tx_descs(txq, 64);
2840 			if (available < howmany(mbuf_len16(m0), EQ_ESIZE / 16))
2841 				break;	/* out of descriptors */
2842 		}
2843 
2844 		next_cidx = cidx + 1;
2845 		if (__predict_false(next_cidx == r->size))
2846 			next_cidx = 0;
2847 
2848 		wr = (void *)&eq->desc[eq->pidx];
2849 		if (sc->flags & IS_VF) {
2850 			total++;
2851 			remaining--;
2852 			ETHER_BPF_MTAP(ifp, m0);
2853 			n = write_txpkt_vm_wr(sc, txq, (void *)wr, m0,
2854 			    available);
2855 		} else if (remaining > 1 &&
2856 		    try_txpkts(m0, r->items[next_cidx], &txp, available) == 0) {
2857 
2858 			/* pkts at cidx, next_cidx should both be in txp. */
2859 			MPASS(txp.npkt == 2);
2860 			tail = r->items[next_cidx];
2861 			MPASS(tail->m_nextpkt == NULL);
2862 			ETHER_BPF_MTAP(ifp, m0);
2863 			ETHER_BPF_MTAP(ifp, tail);
2864 			m0->m_nextpkt = tail;
2865 
2866 			if (__predict_false(++next_cidx == r->size))
2867 				next_cidx = 0;
2868 
2869 			while (next_cidx != pidx) {
2870 				if (add_to_txpkts(r->items[next_cidx], &txp,
2871 				    available) != 0)
2872 					break;
2873 				tail->m_nextpkt = r->items[next_cidx];
2874 				tail = tail->m_nextpkt;
2875 				ETHER_BPF_MTAP(ifp, tail);
2876 				if (__predict_false(++next_cidx == r->size))
2877 					next_cidx = 0;
2878 			}
2879 
2880 			n = write_txpkts_wr(txq, wr, m0, &txp, available);
2881 			total += txp.npkt;
2882 			remaining -= txp.npkt;
2883 		} else if (mbuf_cflags(m0) & MC_RAW_WR) {
2884 			total++;
2885 			remaining--;
2886 			n = write_raw_wr(txq, (void *)wr, m0, available);
2887 		} else {
2888 			total++;
2889 			remaining--;
2890 			ETHER_BPF_MTAP(ifp, m0);
2891 			n = write_txpkt_wr(txq, (void *)wr, m0, available);
2892 		}
2893 		MPASS(n >= 1 && n <= available && n <= SGE_MAX_WR_NDESC);
2894 
2895 		available -= n;
2896 		dbdiff += n;
2897 		IDXINCR(eq->pidx, n, eq->sidx);
2898 
2899 		if (wr_can_update_eq(wr)) {
2900 			if (total_available_tx_desc(eq) < eq->sidx / 4 &&
2901 			    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2902 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2903 				    F_FW_WR_EQUEQ);
2904 				eq->equeqidx = eq->pidx;
2905 			} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >=
2906 			    32) {
2907 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
2908 				eq->equeqidx = eq->pidx;
2909 			}
2910 		}
2911 
2912 		if (dbdiff >= 16 && remaining >= 4) {
2913 			ring_eq_db(sc, eq, dbdiff);
2914 			available += reclaim_tx_descs(txq, 4 * dbdiff);
2915 			dbdiff = 0;
2916 		}
2917 
2918 		cidx = next_cidx;
2919 	}
2920 	if (dbdiff != 0) {
2921 		ring_eq_db(sc, eq, dbdiff);
2922 		reclaim_tx_descs(txq, 32);
2923 	}
2924 done:
2925 	TXQ_UNLOCK(txq);
2926 
2927 	return (total);
2928 }
2929 
2930 static inline void
2931 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
2932     int qsize)
2933 {
2934 
2935 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
2936 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
2937 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
2938 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
2939 
2940 	iq->flags = 0;
2941 	iq->adapter = sc;
2942 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
2943 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
2944 	if (pktc_idx >= 0) {
2945 		iq->intr_params |= F_QINTR_CNT_EN;
2946 		iq->intr_pktc_idx = pktc_idx;
2947 	}
2948 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
2949 	iq->sidx = iq->qsize - sc->params.sge.spg_len / IQ_ESIZE;
2950 }
2951 
2952 static inline void
2953 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name)
2954 {
2955 
2956 	fl->qsize = qsize;
2957 	fl->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2958 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
2959 	if (sc->flags & BUF_PACKING_OK &&
2960 	    ((!is_t4(sc) && buffer_packing) ||	/* T5+: enabled unless 0 */
2961 	    (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */
2962 		fl->flags |= FL_BUF_PACKING;
2963 	find_best_refill_source(sc, fl, maxp);
2964 	find_safe_refill_source(sc, fl);
2965 }
2966 
2967 static inline void
2968 init_eq(struct adapter *sc, struct sge_eq *eq, int eqtype, int qsize,
2969     uint8_t tx_chan, uint16_t iqid, char *name)
2970 {
2971 	KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
2972 
2973 	eq->flags = eqtype & EQ_TYPEMASK;
2974 	eq->tx_chan = tx_chan;
2975 	eq->iqid = iqid;
2976 	eq->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2977 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
2978 }
2979 
2980 static int
2981 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
2982     bus_dmamap_t *map, bus_addr_t *pa, void **va)
2983 {
2984 	int rc;
2985 
2986 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
2987 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
2988 	if (rc != 0) {
2989 		device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
2990 		goto done;
2991 	}
2992 
2993 	rc = bus_dmamem_alloc(*tag, va,
2994 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
2995 	if (rc != 0) {
2996 		device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
2997 		goto done;
2998 	}
2999 
3000 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
3001 	if (rc != 0) {
3002 		device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
3003 		goto done;
3004 	}
3005 done:
3006 	if (rc)
3007 		free_ring(sc, *tag, *map, *pa, *va);
3008 
3009 	return (rc);
3010 }
3011 
3012 static int
3013 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
3014     bus_addr_t pa, void *va)
3015 {
3016 	if (pa)
3017 		bus_dmamap_unload(tag, map);
3018 	if (va)
3019 		bus_dmamem_free(tag, va, map);
3020 	if (tag)
3021 		bus_dma_tag_destroy(tag);
3022 
3023 	return (0);
3024 }
3025 
3026 /*
3027  * Allocates the ring for an ingress queue and an optional freelist.  If the
3028  * freelist is specified it will be allocated and then associated with the
3029  * ingress queue.
3030  *
3031  * Returns errno on failure.  Resources allocated up to that point may still be
3032  * allocated.  Caller is responsible for cleanup in case this function fails.
3033  *
3034  * If the ingress queue will take interrupts directly then the intr_idx
3035  * specifies the vector, starting from 0.  -1 means the interrupts for this
3036  * queue should be forwarded to the fwq.
3037  */
3038 static int
3039 alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl,
3040     int intr_idx, int cong)
3041 {
3042 	int rc, i, cntxt_id;
3043 	size_t len;
3044 	struct fw_iq_cmd c;
3045 	struct port_info *pi = vi->pi;
3046 	struct adapter *sc = iq->adapter;
3047 	struct sge_params *sp = &sc->params.sge;
3048 	__be32 v = 0;
3049 
3050 	len = iq->qsize * IQ_ESIZE;
3051 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
3052 	    (void **)&iq->desc);
3053 	if (rc != 0)
3054 		return (rc);
3055 
3056 	bzero(&c, sizeof(c));
3057 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
3058 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
3059 	    V_FW_IQ_CMD_VFN(0));
3060 
3061 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
3062 	    FW_LEN16(c));
3063 
3064 	/* Special handling for firmware event queue */
3065 	if (iq == &sc->sge.fwq)
3066 		v |= F_FW_IQ_CMD_IQASYNCH;
3067 
3068 	if (intr_idx < 0) {
3069 		/* Forwarded interrupts, all headed to fwq */
3070 		v |= F_FW_IQ_CMD_IQANDST;
3071 		v |= V_FW_IQ_CMD_IQANDSTINDEX(sc->sge.fwq.cntxt_id);
3072 	} else {
3073 		KASSERT(intr_idx < sc->intr_count,
3074 		    ("%s: invalid direct intr_idx %d", __func__, intr_idx));
3075 		v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
3076 	}
3077 
3078 	c.type_to_iqandstindex = htobe32(v |
3079 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
3080 	    V_FW_IQ_CMD_VIID(vi->viid) |
3081 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
3082 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
3083 	    F_FW_IQ_CMD_IQGTSMODE |
3084 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
3085 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
3086 	c.iqsize = htobe16(iq->qsize);
3087 	c.iqaddr = htobe64(iq->ba);
3088 	if (cong >= 0)
3089 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
3090 
3091 	if (fl) {
3092 		mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
3093 
3094 		len = fl->qsize * EQ_ESIZE;
3095 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
3096 		    &fl->ba, (void **)&fl->desc);
3097 		if (rc)
3098 			return (rc);
3099 
3100 		/* Allocate space for one software descriptor per buffer. */
3101 		rc = alloc_fl_sdesc(fl);
3102 		if (rc != 0) {
3103 			device_printf(sc->dev,
3104 			    "failed to setup fl software descriptors: %d\n",
3105 			    rc);
3106 			return (rc);
3107 		}
3108 
3109 		if (fl->flags & FL_BUF_PACKING) {
3110 			fl->lowat = roundup2(sp->fl_starve_threshold2, 8);
3111 			fl->buf_boundary = sp->pack_boundary;
3112 		} else {
3113 			fl->lowat = roundup2(sp->fl_starve_threshold, 8);
3114 			fl->buf_boundary = 16;
3115 		}
3116 		if (fl_pad && fl->buf_boundary < sp->pad_boundary)
3117 			fl->buf_boundary = sp->pad_boundary;
3118 
3119 		c.iqns_to_fl0congen |=
3120 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
3121 			F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
3122 			(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
3123 			(fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
3124 			    0));
3125 		if (cong >= 0) {
3126 			c.iqns_to_fl0congen |=
3127 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
3128 				    F_FW_IQ_CMD_FL0CONGCIF |
3129 				    F_FW_IQ_CMD_FL0CONGEN);
3130 		}
3131 		c.fl0dcaen_to_fl0cidxfthresh =
3132 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
3133 			X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B) |
3134 			V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
3135 			X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
3136 		c.fl0size = htobe16(fl->qsize);
3137 		c.fl0addr = htobe64(fl->ba);
3138 	}
3139 
3140 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3141 	if (rc != 0) {
3142 		device_printf(sc->dev,
3143 		    "failed to create ingress queue: %d\n", rc);
3144 		return (rc);
3145 	}
3146 
3147 	iq->cidx = 0;
3148 	iq->gen = F_RSPD_GEN;
3149 	iq->intr_next = iq->intr_params;
3150 	iq->cntxt_id = be16toh(c.iqid);
3151 	iq->abs_id = be16toh(c.physiqid);
3152 	iq->flags |= IQ_ALLOCATED;
3153 
3154 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
3155 	if (cntxt_id >= sc->sge.niq) {
3156 		panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
3157 		    cntxt_id, sc->sge.niq - 1);
3158 	}
3159 	sc->sge.iqmap[cntxt_id] = iq;
3160 
3161 	if (fl) {
3162 		u_int qid;
3163 
3164 		iq->flags |= IQ_HAS_FL;
3165 		fl->cntxt_id = be16toh(c.fl0id);
3166 		fl->pidx = fl->cidx = 0;
3167 
3168 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
3169 		if (cntxt_id >= sc->sge.neq) {
3170 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
3171 			    __func__, cntxt_id, sc->sge.neq - 1);
3172 		}
3173 		sc->sge.eqmap[cntxt_id] = (void *)fl;
3174 
3175 		qid = fl->cntxt_id;
3176 		if (isset(&sc->doorbells, DOORBELL_UDB)) {
3177 			uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3178 			uint32_t mask = (1 << s_qpp) - 1;
3179 			volatile uint8_t *udb;
3180 
3181 			udb = sc->udbs_base + UDBS_DB_OFFSET;
3182 			udb += (qid >> s_qpp) << PAGE_SHIFT;
3183 			qid &= mask;
3184 			if (qid < PAGE_SIZE / UDBS_SEG_SIZE) {
3185 				udb += qid << UDBS_SEG_SHIFT;
3186 				qid = 0;
3187 			}
3188 			fl->udb = (volatile void *)udb;
3189 		}
3190 		fl->dbval = V_QID(qid) | sc->chip_params->sge_fl_db;
3191 
3192 		FL_LOCK(fl);
3193 		/* Enough to make sure the SGE doesn't think it's starved */
3194 		refill_fl(sc, fl, fl->lowat);
3195 		FL_UNLOCK(fl);
3196 	}
3197 
3198 	if (chip_id(sc) >= CHELSIO_T5 && !(sc->flags & IS_VF) && cong >= 0) {
3199 		uint32_t param, val;
3200 
3201 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
3202 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
3203 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
3204 		if (cong == 0)
3205 			val = 1 << 19;
3206 		else {
3207 			val = 2 << 19;
3208 			for (i = 0; i < 4; i++) {
3209 				if (cong & (1 << i))
3210 					val |= 1 << (i << 2);
3211 			}
3212 		}
3213 
3214 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
3215 		if (rc != 0) {
3216 			/* report error but carry on */
3217 			device_printf(sc->dev,
3218 			    "failed to set congestion manager context for "
3219 			    "ingress queue %d: %d\n", iq->cntxt_id, rc);
3220 		}
3221 	}
3222 
3223 	/* Enable IQ interrupts */
3224 	atomic_store_rel_int(&iq->state, IQS_IDLE);
3225 	t4_write_reg(sc, sc->sge_gts_reg, V_SEINTARM(iq->intr_params) |
3226 	    V_INGRESSQID(iq->cntxt_id));
3227 
3228 	return (0);
3229 }
3230 
3231 static int
3232 free_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl)
3233 {
3234 	int rc;
3235 	struct adapter *sc = iq->adapter;
3236 	device_t dev;
3237 
3238 	if (sc == NULL)
3239 		return (0);	/* nothing to do */
3240 
3241 	dev = vi ? vi->dev : sc->dev;
3242 
3243 	if (iq->flags & IQ_ALLOCATED) {
3244 		rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
3245 		    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
3246 		    fl ? fl->cntxt_id : 0xffff, 0xffff);
3247 		if (rc != 0) {
3248 			device_printf(dev,
3249 			    "failed to free queue %p: %d\n", iq, rc);
3250 			return (rc);
3251 		}
3252 		iq->flags &= ~IQ_ALLOCATED;
3253 	}
3254 
3255 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
3256 
3257 	bzero(iq, sizeof(*iq));
3258 
3259 	if (fl) {
3260 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
3261 		    fl->desc);
3262 
3263 		if (fl->sdesc)
3264 			free_fl_sdesc(sc, fl);
3265 
3266 		if (mtx_initialized(&fl->fl_lock))
3267 			mtx_destroy(&fl->fl_lock);
3268 
3269 		bzero(fl, sizeof(*fl));
3270 	}
3271 
3272 	return (0);
3273 }
3274 
3275 static void
3276 add_iq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
3277     struct sge_iq *iq)
3278 {
3279 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3280 
3281 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD, &iq->ba,
3282 	    "bus address of descriptor ring");
3283 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3284 	    iq->qsize * IQ_ESIZE, "descriptor ring size in bytes");
3285 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3286 	    CTLTYPE_INT | CTLFLAG_RD, &iq->abs_id, 0, sysctl_uint16, "I",
3287 	    "absolute id of the queue");
3288 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3289 	    CTLTYPE_INT | CTLFLAG_RD, &iq->cntxt_id, 0, sysctl_uint16, "I",
3290 	    "SGE context id of the queue");
3291 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3292 	    CTLTYPE_INT | CTLFLAG_RD, &iq->cidx, 0, sysctl_uint16, "I",
3293 	    "consumer index");
3294 }
3295 
3296 static void
3297 add_fl_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
3298     struct sysctl_oid *oid, struct sge_fl *fl)
3299 {
3300 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3301 
3302 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3303 	    "freelist");
3304 	children = SYSCTL_CHILDREN(oid);
3305 
3306 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3307 	    &fl->ba, "bus address of descriptor ring");
3308 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3309 	    fl->sidx * EQ_ESIZE + sc->params.sge.spg_len,
3310 	    "desc ring size in bytes");
3311 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3312 	    CTLTYPE_INT | CTLFLAG_RD, &fl->cntxt_id, 0, sysctl_uint16, "I",
3313 	    "SGE context id of the freelist");
3314 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL,
3315 	    fl_pad ? 1 : 0, "padding enabled");
3316 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL,
3317 	    fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled");
3318 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
3319 	    0, "consumer index");
3320 	if (fl->flags & FL_BUF_PACKING) {
3321 		SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
3322 		    CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
3323 	}
3324 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
3325 	    0, "producer index");
3326 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_allocated",
3327 	    CTLFLAG_RD, &fl->mbuf_allocated, "# of mbuf allocated");
3328 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_inlined",
3329 	    CTLFLAG_RD, &fl->mbuf_inlined, "# of mbuf inlined in clusters");
3330 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
3331 	    CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
3332 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
3333 	    CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
3334 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
3335 	    CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
3336 }
3337 
3338 static int
3339 alloc_fwq(struct adapter *sc)
3340 {
3341 	int rc, intr_idx;
3342 	struct sge_iq *fwq = &sc->sge.fwq;
3343 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
3344 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3345 
3346 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE);
3347 	if (sc->flags & IS_VF)
3348 		intr_idx = 0;
3349 	else
3350 		intr_idx = sc->intr_count > 1 ? 1 : 0;
3351 	rc = alloc_iq_fl(&sc->port[0]->vi[0], fwq, NULL, intr_idx, -1);
3352 	if (rc != 0) {
3353 		device_printf(sc->dev,
3354 		    "failed to create firmware event queue: %d\n", rc);
3355 		return (rc);
3356 	}
3357 
3358 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD,
3359 	    NULL, "firmware event queue");
3360 	add_iq_sysctls(&sc->ctx, oid, fwq);
3361 
3362 	return (0);
3363 }
3364 
3365 static int
3366 free_fwq(struct adapter *sc)
3367 {
3368 	return free_iq_fl(NULL, &sc->sge.fwq, NULL);
3369 }
3370 
3371 static int
3372 alloc_ctrlq(struct adapter *sc, struct sge_wrq *ctrlq, int idx,
3373     struct sysctl_oid *oid)
3374 {
3375 	int rc;
3376 	char name[16];
3377 	struct sysctl_oid_list *children;
3378 
3379 	snprintf(name, sizeof(name), "%s ctrlq%d", device_get_nameunit(sc->dev),
3380 	    idx);
3381 	init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[idx]->tx_chan,
3382 	    sc->sge.fwq.cntxt_id, name);
3383 
3384 	children = SYSCTL_CHILDREN(oid);
3385 	snprintf(name, sizeof(name), "%d", idx);
3386 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3387 	    NULL, "ctrl queue");
3388 	rc = alloc_wrq(sc, NULL, ctrlq, oid);
3389 
3390 	return (rc);
3391 }
3392 
3393 int
3394 tnl_cong(struct port_info *pi, int drop)
3395 {
3396 
3397 	if (drop == -1)
3398 		return (-1);
3399 	else if (drop == 1)
3400 		return (0);
3401 	else
3402 		return (pi->rx_e_chan_map);
3403 }
3404 
3405 static int
3406 alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int intr_idx, int idx,
3407     struct sysctl_oid *oid)
3408 {
3409 	int rc;
3410 	struct adapter *sc = vi->pi->adapter;
3411 	struct sysctl_oid_list *children;
3412 	char name[16];
3413 
3414 	rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, intr_idx,
3415 	    tnl_cong(vi->pi, cong_drop));
3416 	if (rc != 0)
3417 		return (rc);
3418 
3419 	if (idx == 0)
3420 		sc->sge.iq_base = rxq->iq.abs_id - rxq->iq.cntxt_id;
3421 	else
3422 		KASSERT(rxq->iq.cntxt_id + sc->sge.iq_base == rxq->iq.abs_id,
3423 		    ("iq_base mismatch"));
3424 	KASSERT(sc->sge.iq_base == 0 || sc->flags & IS_VF,
3425 	    ("PF with non-zero iq_base"));
3426 
3427 	/*
3428 	 * The freelist is just barely above the starvation threshold right now,
3429 	 * fill it up a bit more.
3430 	 */
3431 	FL_LOCK(&rxq->fl);
3432 	refill_fl(sc, &rxq->fl, 128);
3433 	FL_UNLOCK(&rxq->fl);
3434 
3435 #if defined(INET) || defined(INET6)
3436 	rc = tcp_lro_init_args(&rxq->lro, vi->ifp, lro_entries, lro_mbufs);
3437 	if (rc != 0)
3438 		return (rc);
3439 	MPASS(rxq->lro.ifp == vi->ifp);	/* also indicates LRO init'ed */
3440 
3441 	if (vi->ifp->if_capenable & IFCAP_LRO)
3442 		rxq->iq.flags |= IQ_LRO_ENABLED;
3443 #endif
3444 	if (vi->ifp->if_capenable & IFCAP_HWRXTSTMP)
3445 		rxq->iq.flags |= IQ_RX_TIMESTAMP;
3446 	rxq->ifp = vi->ifp;
3447 
3448 	children = SYSCTL_CHILDREN(oid);
3449 
3450 	snprintf(name, sizeof(name), "%d", idx);
3451 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3452 	    NULL, "rx queue");
3453 	children = SYSCTL_CHILDREN(oid);
3454 
3455 	add_iq_sysctls(&vi->ctx, oid, &rxq->iq);
3456 #if defined(INET) || defined(INET6)
3457 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
3458 	    &rxq->lro.lro_queued, 0, NULL);
3459 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
3460 	    &rxq->lro.lro_flushed, 0, NULL);
3461 #endif
3462 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
3463 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
3464 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_extraction",
3465 	    CTLFLAG_RD, &rxq->vlan_extraction,
3466 	    "# of times hardware extracted 802.1Q tag");
3467 
3468 	add_fl_sysctls(sc, &vi->ctx, oid, &rxq->fl);
3469 
3470 	return (rc);
3471 }
3472 
3473 static int
3474 free_rxq(struct vi_info *vi, struct sge_rxq *rxq)
3475 {
3476 	int rc;
3477 
3478 #if defined(INET) || defined(INET6)
3479 	if (rxq->lro.ifp) {
3480 		tcp_lro_free(&rxq->lro);
3481 		rxq->lro.ifp = NULL;
3482 	}
3483 #endif
3484 
3485 	rc = free_iq_fl(vi, &rxq->iq, &rxq->fl);
3486 	if (rc == 0)
3487 		bzero(rxq, sizeof(*rxq));
3488 
3489 	return (rc);
3490 }
3491 
3492 #ifdef TCP_OFFLOAD
3493 static int
3494 alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq,
3495     int intr_idx, int idx, struct sysctl_oid *oid)
3496 {
3497 	struct port_info *pi = vi->pi;
3498 	int rc;
3499 	struct sysctl_oid_list *children;
3500 	char name[16];
3501 
3502 	rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx, 0);
3503 	if (rc != 0)
3504 		return (rc);
3505 
3506 	children = SYSCTL_CHILDREN(oid);
3507 
3508 	snprintf(name, sizeof(name), "%d", idx);
3509 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3510 	    NULL, "rx queue");
3511 	add_iq_sysctls(&vi->ctx, oid, &ofld_rxq->iq);
3512 	add_fl_sysctls(pi->adapter, &vi->ctx, oid, &ofld_rxq->fl);
3513 
3514 	return (rc);
3515 }
3516 
3517 static int
3518 free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq)
3519 {
3520 	int rc;
3521 
3522 	rc = free_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl);
3523 	if (rc == 0)
3524 		bzero(ofld_rxq, sizeof(*ofld_rxq));
3525 
3526 	return (rc);
3527 }
3528 #endif
3529 
3530 #ifdef DEV_NETMAP
3531 static int
3532 alloc_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int intr_idx,
3533     int idx, struct sysctl_oid *oid)
3534 {
3535 	int rc;
3536 	struct sysctl_oid_list *children;
3537 	struct sysctl_ctx_list *ctx;
3538 	char name[16];
3539 	size_t len;
3540 	struct adapter *sc = vi->pi->adapter;
3541 	struct netmap_adapter *na = NA(vi->ifp);
3542 
3543 	MPASS(na != NULL);
3544 
3545 	len = vi->qsize_rxq * IQ_ESIZE;
3546 	rc = alloc_ring(sc, len, &nm_rxq->iq_desc_tag, &nm_rxq->iq_desc_map,
3547 	    &nm_rxq->iq_ba, (void **)&nm_rxq->iq_desc);
3548 	if (rc != 0)
3549 		return (rc);
3550 
3551 	len = na->num_rx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3552 	rc = alloc_ring(sc, len, &nm_rxq->fl_desc_tag, &nm_rxq->fl_desc_map,
3553 	    &nm_rxq->fl_ba, (void **)&nm_rxq->fl_desc);
3554 	if (rc != 0)
3555 		return (rc);
3556 
3557 	nm_rxq->vi = vi;
3558 	nm_rxq->nid = idx;
3559 	nm_rxq->iq_cidx = 0;
3560 	nm_rxq->iq_sidx = vi->qsize_rxq - sc->params.sge.spg_len / IQ_ESIZE;
3561 	nm_rxq->iq_gen = F_RSPD_GEN;
3562 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
3563 	nm_rxq->fl_sidx = na->num_rx_desc;
3564 	nm_rxq->intr_idx = intr_idx;
3565 	nm_rxq->iq_cntxt_id = INVALID_NM_RXQ_CNTXT_ID;
3566 
3567 	ctx = &vi->ctx;
3568 	children = SYSCTL_CHILDREN(oid);
3569 
3570 	snprintf(name, sizeof(name), "%d", idx);
3571 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL,
3572 	    "rx queue");
3573 	children = SYSCTL_CHILDREN(oid);
3574 
3575 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3576 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_abs_id, 0, sysctl_uint16,
3577 	    "I", "absolute id of the queue");
3578 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3579 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cntxt_id, 0, sysctl_uint16,
3580 	    "I", "SGE context id of the queue");
3581 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3582 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cidx, 0, sysctl_uint16, "I",
3583 	    "consumer index");
3584 
3585 	children = SYSCTL_CHILDREN(oid);
3586 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3587 	    "freelist");
3588 	children = SYSCTL_CHILDREN(oid);
3589 
3590 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3591 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->fl_cntxt_id, 0, sysctl_uint16,
3592 	    "I", "SGE context id of the freelist");
3593 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
3594 	    &nm_rxq->fl_cidx, 0, "consumer index");
3595 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
3596 	    &nm_rxq->fl_pidx, 0, "producer index");
3597 
3598 	return (rc);
3599 }
3600 
3601 
3602 static int
3603 free_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
3604 {
3605 	struct adapter *sc = vi->pi->adapter;
3606 
3607 	if (vi->flags & VI_INIT_DONE)
3608 		MPASS(nm_rxq->iq_cntxt_id == INVALID_NM_RXQ_CNTXT_ID);
3609 	else
3610 		MPASS(nm_rxq->iq_cntxt_id == 0);
3611 
3612 	free_ring(sc, nm_rxq->iq_desc_tag, nm_rxq->iq_desc_map, nm_rxq->iq_ba,
3613 	    nm_rxq->iq_desc);
3614 	free_ring(sc, nm_rxq->fl_desc_tag, nm_rxq->fl_desc_map, nm_rxq->fl_ba,
3615 	    nm_rxq->fl_desc);
3616 
3617 	return (0);
3618 }
3619 
3620 static int
3621 alloc_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq, int iqidx, int idx,
3622     struct sysctl_oid *oid)
3623 {
3624 	int rc;
3625 	size_t len;
3626 	struct port_info *pi = vi->pi;
3627 	struct adapter *sc = pi->adapter;
3628 	struct netmap_adapter *na = NA(vi->ifp);
3629 	char name[16];
3630 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3631 
3632 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3633 	rc = alloc_ring(sc, len, &nm_txq->desc_tag, &nm_txq->desc_map,
3634 	    &nm_txq->ba, (void **)&nm_txq->desc);
3635 	if (rc)
3636 		return (rc);
3637 
3638 	nm_txq->pidx = nm_txq->cidx = 0;
3639 	nm_txq->sidx = na->num_tx_desc;
3640 	nm_txq->nid = idx;
3641 	nm_txq->iqidx = iqidx;
3642 	nm_txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3643 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(G_FW_VIID_PFN(vi->viid)) |
3644 	    V_TXPKT_VF(G_FW_VIID_VIN(vi->viid)) |
3645 	    V_TXPKT_VF_VLD(G_FW_VIID_VIVLD(vi->viid)));
3646 	nm_txq->cntxt_id = INVALID_NM_TXQ_CNTXT_ID;
3647 
3648 	snprintf(name, sizeof(name), "%d", idx);
3649 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3650 	    NULL, "netmap tx queue");
3651 	children = SYSCTL_CHILDREN(oid);
3652 
3653 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3654 	    &nm_txq->cntxt_id, 0, "SGE context id of the queue");
3655 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3656 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->cidx, 0, sysctl_uint16, "I",
3657 	    "consumer index");
3658 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3659 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->pidx, 0, sysctl_uint16, "I",
3660 	    "producer index");
3661 
3662 	return (rc);
3663 }
3664 
3665 static int
3666 free_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
3667 {
3668 	struct adapter *sc = vi->pi->adapter;
3669 
3670 	if (vi->flags & VI_INIT_DONE)
3671 		MPASS(nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID);
3672 	else
3673 		MPASS(nm_txq->cntxt_id == 0);
3674 
3675 	free_ring(sc, nm_txq->desc_tag, nm_txq->desc_map, nm_txq->ba,
3676 	    nm_txq->desc);
3677 
3678 	return (0);
3679 }
3680 #endif
3681 
3682 /*
3683  * Returns a reasonable automatic cidx flush threshold for a given queue size.
3684  */
3685 static u_int
3686 qsize_to_fthresh(int qsize)
3687 {
3688 	u_int fthresh;
3689 
3690 	while (!powerof2(qsize))
3691 		qsize++;
3692 	fthresh = ilog2(qsize);
3693 	if (fthresh > X_CIDXFLUSHTHRESH_128)
3694 		fthresh = X_CIDXFLUSHTHRESH_128;
3695 
3696 	return (fthresh);
3697 }
3698 
3699 static int
3700 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
3701 {
3702 	int rc, cntxt_id;
3703 	struct fw_eq_ctrl_cmd c;
3704 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3705 
3706 	bzero(&c, sizeof(c));
3707 
3708 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
3709 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
3710 	    V_FW_EQ_CTRL_CMD_VFN(0));
3711 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
3712 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
3713 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid));
3714 	c.physeqid_pkd = htobe32(0);
3715 	c.fetchszm_to_iqid =
3716 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
3717 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
3718 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
3719 	c.dcaen_to_eqsize =
3720 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3721 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3722 		V_FW_EQ_CTRL_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
3723 		V_FW_EQ_CTRL_CMD_EQSIZE(qsize));
3724 	c.eqaddr = htobe64(eq->ba);
3725 
3726 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3727 	if (rc != 0) {
3728 		device_printf(sc->dev,
3729 		    "failed to create control queue %d: %d\n", eq->tx_chan, rc);
3730 		return (rc);
3731 	}
3732 	eq->flags |= EQ_ALLOCATED;
3733 
3734 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
3735 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3736 	if (cntxt_id >= sc->sge.neq)
3737 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3738 		cntxt_id, sc->sge.neq - 1);
3739 	sc->sge.eqmap[cntxt_id] = eq;
3740 
3741 	return (rc);
3742 }
3743 
3744 static int
3745 eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3746 {
3747 	int rc, cntxt_id;
3748 	struct fw_eq_eth_cmd c;
3749 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3750 
3751 	bzero(&c, sizeof(c));
3752 
3753 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
3754 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
3755 	    V_FW_EQ_ETH_CMD_VFN(0));
3756 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
3757 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
3758 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
3759 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
3760 	c.fetchszm_to_iqid =
3761 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3762 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
3763 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
3764 	c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3765 	    V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3766 	    V_FW_EQ_ETH_CMD_EQSIZE(qsize));
3767 	c.eqaddr = htobe64(eq->ba);
3768 
3769 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3770 	if (rc != 0) {
3771 		device_printf(vi->dev,
3772 		    "failed to create Ethernet egress queue: %d\n", rc);
3773 		return (rc);
3774 	}
3775 	eq->flags |= EQ_ALLOCATED;
3776 
3777 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
3778 	eq->abs_id = G_FW_EQ_ETH_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
3779 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3780 	if (cntxt_id >= sc->sge.neq)
3781 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3782 		cntxt_id, sc->sge.neq - 1);
3783 	sc->sge.eqmap[cntxt_id] = eq;
3784 
3785 	return (rc);
3786 }
3787 
3788 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3789 static int
3790 ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3791 {
3792 	int rc, cntxt_id;
3793 	struct fw_eq_ofld_cmd c;
3794 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3795 
3796 	bzero(&c, sizeof(c));
3797 
3798 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
3799 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
3800 	    V_FW_EQ_OFLD_CMD_VFN(0));
3801 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
3802 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
3803 	c.fetchszm_to_iqid =
3804 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
3805 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
3806 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
3807 	c.dcaen_to_eqsize =
3808 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3809 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3810 		V_FW_EQ_OFLD_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
3811 		V_FW_EQ_OFLD_CMD_EQSIZE(qsize));
3812 	c.eqaddr = htobe64(eq->ba);
3813 
3814 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3815 	if (rc != 0) {
3816 		device_printf(vi->dev,
3817 		    "failed to create egress queue for TCP offload: %d\n", rc);
3818 		return (rc);
3819 	}
3820 	eq->flags |= EQ_ALLOCATED;
3821 
3822 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
3823 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3824 	if (cntxt_id >= sc->sge.neq)
3825 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3826 		cntxt_id, sc->sge.neq - 1);
3827 	sc->sge.eqmap[cntxt_id] = eq;
3828 
3829 	return (rc);
3830 }
3831 #endif
3832 
3833 static int
3834 alloc_eq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3835 {
3836 	int rc, qsize;
3837 	size_t len;
3838 
3839 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
3840 
3841 	qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3842 	len = qsize * EQ_ESIZE;
3843 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
3844 	    &eq->ba, (void **)&eq->desc);
3845 	if (rc)
3846 		return (rc);
3847 
3848 	eq->pidx = eq->cidx = eq->dbidx = 0;
3849 	/* Note that equeqidx is not used with sge_wrq (OFLD/CTRL) queues. */
3850 	eq->equeqidx = 0;
3851 	eq->doorbells = sc->doorbells;
3852 
3853 	switch (eq->flags & EQ_TYPEMASK) {
3854 	case EQ_CTRL:
3855 		rc = ctrl_eq_alloc(sc, eq);
3856 		break;
3857 
3858 	case EQ_ETH:
3859 		rc = eth_eq_alloc(sc, vi, eq);
3860 		break;
3861 
3862 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3863 	case EQ_OFLD:
3864 		rc = ofld_eq_alloc(sc, vi, eq);
3865 		break;
3866 #endif
3867 
3868 	default:
3869 		panic("%s: invalid eq type %d.", __func__,
3870 		    eq->flags & EQ_TYPEMASK);
3871 	}
3872 	if (rc != 0) {
3873 		device_printf(sc->dev,
3874 		    "failed to allocate egress queue(%d): %d\n",
3875 		    eq->flags & EQ_TYPEMASK, rc);
3876 	}
3877 
3878 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
3879 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
3880 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
3881 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3882 		uint32_t mask = (1 << s_qpp) - 1;
3883 		volatile uint8_t *udb;
3884 
3885 		udb = sc->udbs_base + UDBS_DB_OFFSET;
3886 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
3887 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
3888 		if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
3889 	    		clrbit(&eq->doorbells, DOORBELL_WCWR);
3890 		else {
3891 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
3892 			eq->udb_qid = 0;
3893 		}
3894 		eq->udb = (volatile void *)udb;
3895 	}
3896 
3897 	return (rc);
3898 }
3899 
3900 static int
3901 free_eq(struct adapter *sc, struct sge_eq *eq)
3902 {
3903 	int rc;
3904 
3905 	if (eq->flags & EQ_ALLOCATED) {
3906 		switch (eq->flags & EQ_TYPEMASK) {
3907 		case EQ_CTRL:
3908 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
3909 			    eq->cntxt_id);
3910 			break;
3911 
3912 		case EQ_ETH:
3913 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
3914 			    eq->cntxt_id);
3915 			break;
3916 
3917 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3918 		case EQ_OFLD:
3919 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
3920 			    eq->cntxt_id);
3921 			break;
3922 #endif
3923 
3924 		default:
3925 			panic("%s: invalid eq type %d.", __func__,
3926 			    eq->flags & EQ_TYPEMASK);
3927 		}
3928 		if (rc != 0) {
3929 			device_printf(sc->dev,
3930 			    "failed to free egress queue (%d): %d\n",
3931 			    eq->flags & EQ_TYPEMASK, rc);
3932 			return (rc);
3933 		}
3934 		eq->flags &= ~EQ_ALLOCATED;
3935 	}
3936 
3937 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
3938 
3939 	if (mtx_initialized(&eq->eq_lock))
3940 		mtx_destroy(&eq->eq_lock);
3941 
3942 	bzero(eq, sizeof(*eq));
3943 	return (0);
3944 }
3945 
3946 static int
3947 alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq,
3948     struct sysctl_oid *oid)
3949 {
3950 	int rc;
3951 	struct sysctl_ctx_list *ctx = vi ? &vi->ctx : &sc->ctx;
3952 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3953 
3954 	rc = alloc_eq(sc, vi, &wrq->eq);
3955 	if (rc)
3956 		return (rc);
3957 
3958 	wrq->adapter = sc;
3959 	TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq);
3960 	TAILQ_INIT(&wrq->incomplete_wrs);
3961 	STAILQ_INIT(&wrq->wr_list);
3962 	wrq->nwr_pending = 0;
3963 	wrq->ndesc_needed = 0;
3964 
3965 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3966 	    &wrq->eq.ba, "bus address of descriptor ring");
3967 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3968 	    wrq->eq.sidx * EQ_ESIZE + sc->params.sge.spg_len,
3969 	    "desc ring size in bytes");
3970 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3971 	    &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
3972 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3973 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I",
3974 	    "consumer index");
3975 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx",
3976 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I",
3977 	    "producer index");
3978 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
3979 	    wrq->eq.sidx, "status page index");
3980 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD,
3981 	    &wrq->tx_wrs_direct, "# of work requests (direct)");
3982 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD,
3983 	    &wrq->tx_wrs_copied, "# of work requests (copied)");
3984 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_sspace", CTLFLAG_RD,
3985 	    &wrq->tx_wrs_ss, "# of work requests (copied from scratch space)");
3986 
3987 	return (rc);
3988 }
3989 
3990 static int
3991 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
3992 {
3993 	int rc;
3994 
3995 	rc = free_eq(sc, &wrq->eq);
3996 	if (rc)
3997 		return (rc);
3998 
3999 	bzero(wrq, sizeof(*wrq));
4000 	return (0);
4001 }
4002 
4003 static int
4004 alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx,
4005     struct sysctl_oid *oid)
4006 {
4007 	int rc;
4008 	struct port_info *pi = vi->pi;
4009 	struct adapter *sc = pi->adapter;
4010 	struct sge_eq *eq = &txq->eq;
4011 	char name[16];
4012 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
4013 
4014 	rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx, can_resume_eth_tx,
4015 	    M_CXGBE, M_WAITOK);
4016 	if (rc != 0) {
4017 		device_printf(sc->dev, "failed to allocate mp_ring: %d\n", rc);
4018 		return (rc);
4019 	}
4020 
4021 	rc = alloc_eq(sc, vi, eq);
4022 	if (rc != 0) {
4023 		mp_ring_free(txq->r);
4024 		txq->r = NULL;
4025 		return (rc);
4026 	}
4027 
4028 	/* Can't fail after this point. */
4029 
4030 	if (idx == 0)
4031 		sc->sge.eq_base = eq->abs_id - eq->cntxt_id;
4032 	else
4033 		KASSERT(eq->cntxt_id + sc->sge.eq_base == eq->abs_id,
4034 		    ("eq_base mismatch"));
4035 	KASSERT(sc->sge.eq_base == 0 || sc->flags & IS_VF,
4036 	    ("PF with non-zero eq_base"));
4037 
4038 	TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq);
4039 	txq->ifp = vi->ifp;
4040 	txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK);
4041 	if (sc->flags & IS_VF)
4042 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
4043 		    V_TXPKT_INTF(pi->tx_chan));
4044 	else
4045 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
4046 		    V_TXPKT_INTF(pi->tx_chan) |
4047 		    V_TXPKT_PF(G_FW_VIID_PFN(vi->viid)) |
4048 		    V_TXPKT_VF(G_FW_VIID_VIN(vi->viid)) |
4049 		    V_TXPKT_VF_VLD(G_FW_VIID_VIVLD(vi->viid)));
4050 	txq->tc_idx = -1;
4051 	txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE,
4052 	    M_ZERO | M_WAITOK);
4053 
4054 	snprintf(name, sizeof(name), "%d", idx);
4055 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
4056 	    NULL, "tx queue");
4057 	children = SYSCTL_CHILDREN(oid);
4058 
4059 	SYSCTL_ADD_UAUTO(&vi->ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
4060 	    &eq->ba, "bus address of descriptor ring");
4061 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
4062 	    eq->sidx * EQ_ESIZE + sc->params.sge.spg_len,
4063 	    "desc ring size in bytes");
4064 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
4065 	    &eq->abs_id, 0, "absolute id of the queue");
4066 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
4067 	    &eq->cntxt_id, 0, "SGE context id of the queue");
4068 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
4069 	    CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I",
4070 	    "consumer index");
4071 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
4072 	    CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I",
4073 	    "producer index");
4074 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
4075 	    eq->sidx, "status page index");
4076 
4077 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "tc",
4078 	    CTLTYPE_INT | CTLFLAG_RW, vi, idx, sysctl_tc, "I",
4079 	    "traffic class (-1 means none)");
4080 
4081 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
4082 	    &txq->txcsum, "# of times hardware assisted with checksum");
4083 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_insertion",
4084 	    CTLFLAG_RD, &txq->vlan_insertion,
4085 	    "# of times hardware inserted 802.1Q tag");
4086 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
4087 	    &txq->tso_wrs, "# of TSO work requests");
4088 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
4089 	    &txq->imm_wrs, "# of work requests with immediate data");
4090 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
4091 	    &txq->sgl_wrs, "# of work requests with direct SGL");
4092 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
4093 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
4094 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_wrs",
4095 	    CTLFLAG_RD, &txq->txpkts0_wrs,
4096 	    "# of txpkts (type 0) work requests");
4097 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_wrs",
4098 	    CTLFLAG_RD, &txq->txpkts1_wrs,
4099 	    "# of txpkts (type 1) work requests");
4100 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_pkts",
4101 	    CTLFLAG_RD, &txq->txpkts0_pkts,
4102 	    "# of frames tx'd using type0 txpkts work requests");
4103 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_pkts",
4104 	    CTLFLAG_RD, &txq->txpkts1_pkts,
4105 	    "# of frames tx'd using type1 txpkts work requests");
4106 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "raw_wrs", CTLFLAG_RD,
4107 	    &txq->raw_wrs, "# of raw work requests (non-packets)");
4108 
4109 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_enqueues",
4110 	    CTLFLAG_RD, &txq->r->enqueues,
4111 	    "# of enqueues to the mp_ring for this queue");
4112 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_drops",
4113 	    CTLFLAG_RD, &txq->r->drops,
4114 	    "# of drops in the mp_ring for this queue");
4115 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_starts",
4116 	    CTLFLAG_RD, &txq->r->starts,
4117 	    "# of normal consumer starts in the mp_ring for this queue");
4118 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_stalls",
4119 	    CTLFLAG_RD, &txq->r->stalls,
4120 	    "# of consumer stalls in the mp_ring for this queue");
4121 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_restarts",
4122 	    CTLFLAG_RD, &txq->r->restarts,
4123 	    "# of consumer restarts in the mp_ring for this queue");
4124 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_abdications",
4125 	    CTLFLAG_RD, &txq->r->abdications,
4126 	    "# of consumer abdications in the mp_ring for this queue");
4127 
4128 	return (0);
4129 }
4130 
4131 static int
4132 free_txq(struct vi_info *vi, struct sge_txq *txq)
4133 {
4134 	int rc;
4135 	struct adapter *sc = vi->pi->adapter;
4136 	struct sge_eq *eq = &txq->eq;
4137 
4138 	rc = free_eq(sc, eq);
4139 	if (rc)
4140 		return (rc);
4141 
4142 	sglist_free(txq->gl);
4143 	free(txq->sdesc, M_CXGBE);
4144 	mp_ring_free(txq->r);
4145 
4146 	bzero(txq, sizeof(*txq));
4147 	return (0);
4148 }
4149 
4150 static void
4151 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
4152 {
4153 	bus_addr_t *ba = arg;
4154 
4155 	KASSERT(nseg == 1,
4156 	    ("%s meant for single segment mappings only.", __func__));
4157 
4158 	*ba = error ? 0 : segs->ds_addr;
4159 }
4160 
4161 static inline void
4162 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
4163 {
4164 	uint32_t n, v;
4165 
4166 	n = IDXDIFF(fl->pidx / 8, fl->dbidx, fl->sidx);
4167 	MPASS(n > 0);
4168 
4169 	wmb();
4170 	v = fl->dbval | V_PIDX(n);
4171 	if (fl->udb)
4172 		*fl->udb = htole32(v);
4173 	else
4174 		t4_write_reg(sc, sc->sge_kdoorbell_reg, v);
4175 	IDXINCR(fl->dbidx, n, fl->sidx);
4176 }
4177 
4178 /*
4179  * Fills up the freelist by allocating up to 'n' buffers.  Buffers that are
4180  * recycled do not count towards this allocation budget.
4181  *
4182  * Returns non-zero to indicate that this freelist should be added to the list
4183  * of starving freelists.
4184  */
4185 static int
4186 refill_fl(struct adapter *sc, struct sge_fl *fl, int n)
4187 {
4188 	__be64 *d;
4189 	struct fl_sdesc *sd;
4190 	uintptr_t pa;
4191 	caddr_t cl;
4192 	struct cluster_layout *cll;
4193 	struct sw_zone_info *swz;
4194 	struct cluster_metadata *clm;
4195 	uint16_t max_pidx;
4196 	uint16_t hw_cidx = fl->hw_cidx;		/* stable snapshot */
4197 
4198 	FL_LOCK_ASSERT_OWNED(fl);
4199 
4200 	/*
4201 	 * We always stop at the beginning of the hardware descriptor that's just
4202 	 * before the one with the hw cidx.  This is to avoid hw pidx = hw cidx,
4203 	 * which would mean an empty freelist to the chip.
4204 	 */
4205 	max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1;
4206 	if (fl->pidx == max_pidx * 8)
4207 		return (0);
4208 
4209 	d = &fl->desc[fl->pidx];
4210 	sd = &fl->sdesc[fl->pidx];
4211 	cll = &fl->cll_def;	/* default layout */
4212 	swz = &sc->sge.sw_zone_info[cll->zidx];
4213 
4214 	while (n > 0) {
4215 
4216 		if (sd->cl != NULL) {
4217 
4218 			if (sd->nmbuf == 0) {
4219 				/*
4220 				 * Fast recycle without involving any atomics on
4221 				 * the cluster's metadata (if the cluster has
4222 				 * metadata).  This happens when all frames
4223 				 * received in the cluster were small enough to
4224 				 * fit within a single mbuf each.
4225 				 */
4226 				fl->cl_fast_recycled++;
4227 #ifdef INVARIANTS
4228 				clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
4229 				if (clm != NULL)
4230 					MPASS(clm->refcount == 1);
4231 #endif
4232 				goto recycled_fast;
4233 			}
4234 
4235 			/*
4236 			 * Cluster is guaranteed to have metadata.  Clusters
4237 			 * without metadata always take the fast recycle path
4238 			 * when they're recycled.
4239 			 */
4240 			clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
4241 			MPASS(clm != NULL);
4242 
4243 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4244 				fl->cl_recycled++;
4245 				counter_u64_add(extfree_rels, 1);
4246 				goto recycled;
4247 			}
4248 			sd->cl = NULL;	/* gave up my reference */
4249 		}
4250 		MPASS(sd->cl == NULL);
4251 alloc:
4252 		cl = uma_zalloc(swz->zone, M_NOWAIT);
4253 		if (__predict_false(cl == NULL)) {
4254 			if (cll == &fl->cll_alt || fl->cll_alt.zidx == -1 ||
4255 			    fl->cll_def.zidx == fl->cll_alt.zidx)
4256 				break;
4257 
4258 			/* fall back to the safe zone */
4259 			cll = &fl->cll_alt;
4260 			swz = &sc->sge.sw_zone_info[cll->zidx];
4261 			goto alloc;
4262 		}
4263 		fl->cl_allocated++;
4264 		n--;
4265 
4266 		pa = pmap_kextract((vm_offset_t)cl);
4267 		pa += cll->region1;
4268 		sd->cl = cl;
4269 		sd->cll = *cll;
4270 		*d = htobe64(pa | cll->hwidx);
4271 		clm = cl_metadata(sc, fl, cll, cl);
4272 		if (clm != NULL) {
4273 recycled:
4274 #ifdef INVARIANTS
4275 			clm->sd = sd;
4276 #endif
4277 			clm->refcount = 1;
4278 		}
4279 		sd->nmbuf = 0;
4280 recycled_fast:
4281 		d++;
4282 		sd++;
4283 		if (__predict_false(++fl->pidx % 8 == 0)) {
4284 			uint16_t pidx = fl->pidx / 8;
4285 
4286 			if (__predict_false(pidx == fl->sidx)) {
4287 				fl->pidx = 0;
4288 				pidx = 0;
4289 				sd = fl->sdesc;
4290 				d = fl->desc;
4291 			}
4292 			if (pidx == max_pidx)
4293 				break;
4294 
4295 			if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4)
4296 				ring_fl_db(sc, fl);
4297 		}
4298 	}
4299 
4300 	if (fl->pidx / 8 != fl->dbidx)
4301 		ring_fl_db(sc, fl);
4302 
4303 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
4304 }
4305 
4306 /*
4307  * Attempt to refill all starving freelists.
4308  */
4309 static void
4310 refill_sfl(void *arg)
4311 {
4312 	struct adapter *sc = arg;
4313 	struct sge_fl *fl, *fl_temp;
4314 
4315 	mtx_assert(&sc->sfl_lock, MA_OWNED);
4316 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
4317 		FL_LOCK(fl);
4318 		refill_fl(sc, fl, 64);
4319 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
4320 			TAILQ_REMOVE(&sc->sfl, fl, link);
4321 			fl->flags &= ~FL_STARVING;
4322 		}
4323 		FL_UNLOCK(fl);
4324 	}
4325 
4326 	if (!TAILQ_EMPTY(&sc->sfl))
4327 		callout_schedule(&sc->sfl_callout, hz / 5);
4328 }
4329 
4330 static int
4331 alloc_fl_sdesc(struct sge_fl *fl)
4332 {
4333 
4334 	fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc), M_CXGBE,
4335 	    M_ZERO | M_WAITOK);
4336 
4337 	return (0);
4338 }
4339 
4340 static void
4341 free_fl_sdesc(struct adapter *sc, struct sge_fl *fl)
4342 {
4343 	struct fl_sdesc *sd;
4344 	struct cluster_metadata *clm;
4345 	struct cluster_layout *cll;
4346 	int i;
4347 
4348 	sd = fl->sdesc;
4349 	for (i = 0; i < fl->sidx * 8; i++, sd++) {
4350 		if (sd->cl == NULL)
4351 			continue;
4352 
4353 		cll = &sd->cll;
4354 		clm = cl_metadata(sc, fl, cll, sd->cl);
4355 		if (sd->nmbuf == 0)
4356 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4357 		else if (clm && atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4358 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4359 			counter_u64_add(extfree_rels, 1);
4360 		}
4361 		sd->cl = NULL;
4362 	}
4363 
4364 	free(fl->sdesc, M_CXGBE);
4365 	fl->sdesc = NULL;
4366 }
4367 
4368 static inline void
4369 get_pkt_gl(struct mbuf *m, struct sglist *gl)
4370 {
4371 	int rc;
4372 
4373 	M_ASSERTPKTHDR(m);
4374 
4375 	sglist_reset(gl);
4376 	rc = sglist_append_mbuf(gl, m);
4377 	if (__predict_false(rc != 0)) {
4378 		panic("%s: mbuf %p (%d segs) was vetted earlier but now fails "
4379 		    "with %d.", __func__, m, mbuf_nsegs(m), rc);
4380 	}
4381 
4382 	KASSERT(gl->sg_nseg == mbuf_nsegs(m),
4383 	    ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m,
4384 	    mbuf_nsegs(m), gl->sg_nseg));
4385 	KASSERT(gl->sg_nseg > 0 &&
4386 	    gl->sg_nseg <= (needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS),
4387 	    ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__,
4388 		gl->sg_nseg, needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS));
4389 }
4390 
4391 /*
4392  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
4393  */
4394 static inline u_int
4395 txpkt_len16(u_int nsegs, u_int tso)
4396 {
4397 	u_int n;
4398 
4399 	MPASS(nsegs > 0);
4400 
4401 	nsegs--; /* first segment is part of ulptx_sgl */
4402 	n = sizeof(struct fw_eth_tx_pkt_wr) + sizeof(struct cpl_tx_pkt_core) +
4403 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4404 	if (tso)
4405 		n += sizeof(struct cpl_tx_pkt_lso_core);
4406 
4407 	return (howmany(n, 16));
4408 }
4409 
4410 /*
4411  * len16 for a txpkt_vm WR with a GL.  Includes the firmware work
4412  * request header.
4413  */
4414 static inline u_int
4415 txpkt_vm_len16(u_int nsegs, u_int tso)
4416 {
4417 	u_int n;
4418 
4419 	MPASS(nsegs > 0);
4420 
4421 	nsegs--; /* first segment is part of ulptx_sgl */
4422 	n = sizeof(struct fw_eth_tx_pkt_vm_wr) +
4423 	    sizeof(struct cpl_tx_pkt_core) +
4424 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4425 	if (tso)
4426 		n += sizeof(struct cpl_tx_pkt_lso_core);
4427 
4428 	return (howmany(n, 16));
4429 }
4430 
4431 /*
4432  * len16 for a txpkts type 0 WR with a GL.  Does not include the firmware work
4433  * request header.
4434  */
4435 static inline u_int
4436 txpkts0_len16(u_int nsegs)
4437 {
4438 	u_int n;
4439 
4440 	MPASS(nsegs > 0);
4441 
4442 	nsegs--; /* first segment is part of ulptx_sgl */
4443 	n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) +
4444 	    sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) +
4445 	    8 * ((3 * nsegs) / 2 + (nsegs & 1));
4446 
4447 	return (howmany(n, 16));
4448 }
4449 
4450 /*
4451  * len16 for a txpkts type 1 WR with a GL.  Does not include the firmware work
4452  * request header.
4453  */
4454 static inline u_int
4455 txpkts1_len16(void)
4456 {
4457 	u_int n;
4458 
4459 	n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl);
4460 
4461 	return (howmany(n, 16));
4462 }
4463 
4464 static inline u_int
4465 imm_payload(u_int ndesc)
4466 {
4467 	u_int n;
4468 
4469 	n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) -
4470 	    sizeof(struct cpl_tx_pkt_core);
4471 
4472 	return (n);
4473 }
4474 
4475 /*
4476  * Write a VM txpkt WR for this packet to the hardware descriptors, update the
4477  * software descriptor, and advance the pidx.  It is guaranteed that enough
4478  * descriptors are available.
4479  *
4480  * The return value is the # of hardware descriptors used.
4481  */
4482 static u_int
4483 write_txpkt_vm_wr(struct adapter *sc, struct sge_txq *txq,
4484     struct fw_eth_tx_pkt_vm_wr *wr, struct mbuf *m0, u_int available)
4485 {
4486 	struct sge_eq *eq = &txq->eq;
4487 	struct tx_sdesc *txsd;
4488 	struct cpl_tx_pkt_core *cpl;
4489 	uint32_t ctrl;	/* used in many unrelated places */
4490 	uint64_t ctrl1;
4491 	int csum_type, len16, ndesc, pktlen, nsegs;
4492 	caddr_t dst;
4493 
4494 	TXQ_LOCK_ASSERT_OWNED(txq);
4495 	M_ASSERTPKTHDR(m0);
4496 	MPASS(available > 0 && available < eq->sidx);
4497 
4498 	len16 = mbuf_len16(m0);
4499 	nsegs = mbuf_nsegs(m0);
4500 	pktlen = m0->m_pkthdr.len;
4501 	ctrl = sizeof(struct cpl_tx_pkt_core);
4502 	if (needs_tso(m0))
4503 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4504 	ndesc = howmany(len16, EQ_ESIZE / 16);
4505 	MPASS(ndesc <= available);
4506 
4507 	/* Firmware work request header */
4508 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4509 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_VM_WR) |
4510 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4511 
4512 	ctrl = V_FW_WR_LEN16(len16);
4513 	wr->equiq_to_len16 = htobe32(ctrl);
4514 	wr->r3[0] = 0;
4515 	wr->r3[1] = 0;
4516 
4517 	/*
4518 	 * Copy over ethmacdst, ethmacsrc, ethtype, and vlantci.
4519 	 * vlantci is ignored unless the ethtype is 0x8100, so it's
4520 	 * simpler to always copy it rather than making it
4521 	 * conditional.  Also, it seems that we do not have to set
4522 	 * vlantci or fake the ethtype when doing VLAN tag insertion.
4523 	 */
4524 	m_copydata(m0, 0, sizeof(struct ether_header) + 2, wr->ethmacdst);
4525 
4526 	csum_type = -1;
4527 	if (needs_tso(m0)) {
4528 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4529 
4530 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4531 		    m0->m_pkthdr.l4hlen > 0,
4532 		    ("%s: mbuf %p needs TSO but missing header lengths",
4533 			__func__, m0));
4534 
4535 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4536 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4537 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4538 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4539 			ctrl |= V_LSO_ETHHDR_LEN(1);
4540 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4541 			ctrl |= F_LSO_IPV6;
4542 
4543 		lso->lso_ctrl = htobe32(ctrl);
4544 		lso->ipid_ofst = htobe16(0);
4545 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4546 		lso->seqno_offset = htobe32(0);
4547 		lso->len = htobe32(pktlen);
4548 
4549 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4550 			csum_type = TX_CSUM_TCPIP6;
4551 		else
4552 			csum_type = TX_CSUM_TCPIP;
4553 
4554 		cpl = (void *)(lso + 1);
4555 
4556 		txq->tso_wrs++;
4557 	} else {
4558 		if (m0->m_pkthdr.csum_flags & CSUM_IP_TCP)
4559 			csum_type = TX_CSUM_TCPIP;
4560 		else if (m0->m_pkthdr.csum_flags & CSUM_IP_UDP)
4561 			csum_type = TX_CSUM_UDPIP;
4562 		else if (m0->m_pkthdr.csum_flags & CSUM_IP6_TCP)
4563 			csum_type = TX_CSUM_TCPIP6;
4564 		else if (m0->m_pkthdr.csum_flags & CSUM_IP6_UDP)
4565 			csum_type = TX_CSUM_UDPIP6;
4566 #if defined(INET)
4567 		else if (m0->m_pkthdr.csum_flags & CSUM_IP) {
4568 			/*
4569 			 * XXX: The firmware appears to stomp on the
4570 			 * fragment/flags field of the IP header when
4571 			 * using TX_CSUM_IP.  Fall back to doing
4572 			 * software checksums.
4573 			 */
4574 			u_short *sump;
4575 			struct mbuf *m;
4576 			int offset;
4577 
4578 			m = m0;
4579 			offset = 0;
4580 			sump = m_advance(&m, &offset, m0->m_pkthdr.l2hlen +
4581 			    offsetof(struct ip, ip_sum));
4582 			*sump = in_cksum_skip(m0, m0->m_pkthdr.l2hlen +
4583 			    m0->m_pkthdr.l3hlen, m0->m_pkthdr.l2hlen);
4584 			m0->m_pkthdr.csum_flags &= ~CSUM_IP;
4585 		}
4586 #endif
4587 
4588 		cpl = (void *)(wr + 1);
4589 	}
4590 
4591 	/* Checksum offload */
4592 	ctrl1 = 0;
4593 	if (needs_l3_csum(m0) == 0)
4594 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4595 	if (csum_type >= 0) {
4596 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0,
4597 	    ("%s: mbuf %p needs checksum offload but missing header lengths",
4598 			__func__, m0));
4599 
4600 		if (chip_id(sc) <= CHELSIO_T5) {
4601 			ctrl1 |= V_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4602 			    ETHER_HDR_LEN);
4603 		} else {
4604 			ctrl1 |= V_T6_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4605 			    ETHER_HDR_LEN);
4606 		}
4607 		ctrl1 |= V_TXPKT_IPHDR_LEN(m0->m_pkthdr.l3hlen);
4608 		ctrl1 |= V_TXPKT_CSUM_TYPE(csum_type);
4609 	} else
4610 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4611 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4612 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4613 		txq->txcsum++;	/* some hardware assistance provided */
4614 
4615 	/* VLAN tag insertion */
4616 	if (needs_vlan_insertion(m0)) {
4617 		ctrl1 |= F_TXPKT_VLAN_VLD |
4618 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4619 		txq->vlan_insertion++;
4620 	}
4621 
4622 	/* CPL header */
4623 	cpl->ctrl0 = txq->cpl_ctrl0;
4624 	cpl->pack = 0;
4625 	cpl->len = htobe16(pktlen);
4626 	cpl->ctrl1 = htobe64(ctrl1);
4627 
4628 	/* SGL */
4629 	dst = (void *)(cpl + 1);
4630 
4631 	/*
4632 	 * A packet using TSO will use up an entire descriptor for the
4633 	 * firmware work request header, LSO CPL, and TX_PKT_XT CPL.
4634 	 * If this descriptor is the last descriptor in the ring, wrap
4635 	 * around to the front of the ring explicitly for the start of
4636 	 * the sgl.
4637 	 */
4638 	if (dst == (void *)&eq->desc[eq->sidx]) {
4639 		dst = (void *)&eq->desc[0];
4640 		write_gl_to_txd(txq, m0, &dst, 0);
4641 	} else
4642 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4643 	txq->sgl_wrs++;
4644 
4645 	txq->txpkt_wrs++;
4646 
4647 	txsd = &txq->sdesc[eq->pidx];
4648 	txsd->m = m0;
4649 	txsd->desc_used = ndesc;
4650 
4651 	return (ndesc);
4652 }
4653 
4654 /*
4655  * Write a raw WR to the hardware descriptors, update the software
4656  * descriptor, and advance the pidx.  It is guaranteed that enough
4657  * descriptors are available.
4658  *
4659  * The return value is the # of hardware descriptors used.
4660  */
4661 static u_int
4662 write_raw_wr(struct sge_txq *txq, void *wr, struct mbuf *m0, u_int available)
4663 {
4664 	struct sge_eq *eq = &txq->eq;
4665 	struct tx_sdesc *txsd;
4666 	struct mbuf *m;
4667 	caddr_t dst;
4668 	int len16, ndesc;
4669 
4670 	len16 = mbuf_len16(m0);
4671 	ndesc = howmany(len16, EQ_ESIZE / 16);
4672 	MPASS(ndesc <= available);
4673 
4674 	dst = wr;
4675 	for (m = m0; m != NULL; m = m->m_next)
4676 		copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4677 
4678 	txq->raw_wrs++;
4679 
4680 	txsd = &txq->sdesc[eq->pidx];
4681 	txsd->m = m0;
4682 	txsd->desc_used = ndesc;
4683 
4684 	return (ndesc);
4685 }
4686 
4687 /*
4688  * Write a txpkt WR for this packet to the hardware descriptors, update the
4689  * software descriptor, and advance the pidx.  It is guaranteed that enough
4690  * descriptors are available.
4691  *
4692  * The return value is the # of hardware descriptors used.
4693  */
4694 static u_int
4695 write_txpkt_wr(struct sge_txq *txq, struct fw_eth_tx_pkt_wr *wr,
4696     struct mbuf *m0, u_int available)
4697 {
4698 	struct sge_eq *eq = &txq->eq;
4699 	struct tx_sdesc *txsd;
4700 	struct cpl_tx_pkt_core *cpl;
4701 	uint32_t ctrl;	/* used in many unrelated places */
4702 	uint64_t ctrl1;
4703 	int len16, ndesc, pktlen, nsegs;
4704 	caddr_t dst;
4705 
4706 	TXQ_LOCK_ASSERT_OWNED(txq);
4707 	M_ASSERTPKTHDR(m0);
4708 	MPASS(available > 0 && available < eq->sidx);
4709 
4710 	len16 = mbuf_len16(m0);
4711 	nsegs = mbuf_nsegs(m0);
4712 	pktlen = m0->m_pkthdr.len;
4713 	ctrl = sizeof(struct cpl_tx_pkt_core);
4714 	if (needs_tso(m0))
4715 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4716 	else if (pktlen <= imm_payload(2) && available >= 2) {
4717 		/* Immediate data.  Recalculate len16 and set nsegs to 0. */
4718 		ctrl += pktlen;
4719 		len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) +
4720 		    sizeof(struct cpl_tx_pkt_core) + pktlen, 16);
4721 		nsegs = 0;
4722 	}
4723 	ndesc = howmany(len16, EQ_ESIZE / 16);
4724 	MPASS(ndesc <= available);
4725 
4726 	/* Firmware work request header */
4727 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4728 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
4729 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4730 
4731 	ctrl = V_FW_WR_LEN16(len16);
4732 	wr->equiq_to_len16 = htobe32(ctrl);
4733 	wr->r3 = 0;
4734 
4735 	if (needs_tso(m0)) {
4736 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4737 
4738 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4739 		    m0->m_pkthdr.l4hlen > 0,
4740 		    ("%s: mbuf %p needs TSO but missing header lengths",
4741 			__func__, m0));
4742 
4743 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4744 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4745 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4746 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4747 			ctrl |= V_LSO_ETHHDR_LEN(1);
4748 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4749 			ctrl |= F_LSO_IPV6;
4750 
4751 		lso->lso_ctrl = htobe32(ctrl);
4752 		lso->ipid_ofst = htobe16(0);
4753 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4754 		lso->seqno_offset = htobe32(0);
4755 		lso->len = htobe32(pktlen);
4756 
4757 		cpl = (void *)(lso + 1);
4758 
4759 		txq->tso_wrs++;
4760 	} else
4761 		cpl = (void *)(wr + 1);
4762 
4763 	/* Checksum offload */
4764 	ctrl1 = 0;
4765 	if (needs_l3_csum(m0) == 0)
4766 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4767 	if (needs_l4_csum(m0) == 0)
4768 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4769 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4770 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4771 		txq->txcsum++;	/* some hardware assistance provided */
4772 
4773 	/* VLAN tag insertion */
4774 	if (needs_vlan_insertion(m0)) {
4775 		ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4776 		txq->vlan_insertion++;
4777 	}
4778 
4779 	/* CPL header */
4780 	cpl->ctrl0 = txq->cpl_ctrl0;
4781 	cpl->pack = 0;
4782 	cpl->len = htobe16(pktlen);
4783 	cpl->ctrl1 = htobe64(ctrl1);
4784 
4785 	/* SGL */
4786 	dst = (void *)(cpl + 1);
4787 	if (nsegs > 0) {
4788 
4789 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4790 		txq->sgl_wrs++;
4791 	} else {
4792 		struct mbuf *m;
4793 
4794 		for (m = m0; m != NULL; m = m->m_next) {
4795 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4796 #ifdef INVARIANTS
4797 			pktlen -= m->m_len;
4798 #endif
4799 		}
4800 #ifdef INVARIANTS
4801 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
4802 #endif
4803 		txq->imm_wrs++;
4804 	}
4805 
4806 	txq->txpkt_wrs++;
4807 
4808 	txsd = &txq->sdesc[eq->pidx];
4809 	txsd->m = m0;
4810 	txsd->desc_used = ndesc;
4811 
4812 	return (ndesc);
4813 }
4814 
4815 static int
4816 try_txpkts(struct mbuf *m, struct mbuf *n, struct txpkts *txp, u_int available)
4817 {
4818 	u_int needed, nsegs1, nsegs2, l1, l2;
4819 
4820 	if (cannot_use_txpkts(m) || cannot_use_txpkts(n))
4821 		return (1);
4822 
4823 	nsegs1 = mbuf_nsegs(m);
4824 	nsegs2 = mbuf_nsegs(n);
4825 	if (nsegs1 + nsegs2 == 2) {
4826 		txp->wr_type = 1;
4827 		l1 = l2 = txpkts1_len16();
4828 	} else {
4829 		txp->wr_type = 0;
4830 		l1 = txpkts0_len16(nsegs1);
4831 		l2 = txpkts0_len16(nsegs2);
4832 	}
4833 	txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) + l1 + l2;
4834 	needed = howmany(txp->len16, EQ_ESIZE / 16);
4835 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4836 		return (1);
4837 
4838 	txp->plen = m->m_pkthdr.len + n->m_pkthdr.len;
4839 	if (txp->plen > 65535)
4840 		return (1);
4841 
4842 	txp->npkt = 2;
4843 	set_mbuf_len16(m, l1);
4844 	set_mbuf_len16(n, l2);
4845 
4846 	return (0);
4847 }
4848 
4849 static int
4850 add_to_txpkts(struct mbuf *m, struct txpkts *txp, u_int available)
4851 {
4852 	u_int plen, len16, needed, nsegs;
4853 
4854 	MPASS(txp->wr_type == 0 || txp->wr_type == 1);
4855 
4856 	if (cannot_use_txpkts(m))
4857 		return (1);
4858 
4859 	nsegs = mbuf_nsegs(m);
4860 	if (txp->wr_type == 1 && nsegs != 1)
4861 		return (1);
4862 
4863 	plen = txp->plen + m->m_pkthdr.len;
4864 	if (plen > 65535)
4865 		return (1);
4866 
4867 	if (txp->wr_type == 0)
4868 		len16 = txpkts0_len16(nsegs);
4869 	else
4870 		len16 = txpkts1_len16();
4871 	needed = howmany(txp->len16 + len16, EQ_ESIZE / 16);
4872 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4873 		return (1);
4874 
4875 	txp->npkt++;
4876 	txp->plen = plen;
4877 	txp->len16 += len16;
4878 	set_mbuf_len16(m, len16);
4879 
4880 	return (0);
4881 }
4882 
4883 /*
4884  * Write a txpkts WR for the packets in txp to the hardware descriptors, update
4885  * the software descriptor, and advance the pidx.  It is guaranteed that enough
4886  * descriptors are available.
4887  *
4888  * The return value is the # of hardware descriptors used.
4889  */
4890 static u_int
4891 write_txpkts_wr(struct sge_txq *txq, struct fw_eth_tx_pkts_wr *wr,
4892     struct mbuf *m0, const struct txpkts *txp, u_int available)
4893 {
4894 	struct sge_eq *eq = &txq->eq;
4895 	struct tx_sdesc *txsd;
4896 	struct cpl_tx_pkt_core *cpl;
4897 	uint32_t ctrl;
4898 	uint64_t ctrl1;
4899 	int ndesc, checkwrap;
4900 	struct mbuf *m;
4901 	void *flitp;
4902 
4903 	TXQ_LOCK_ASSERT_OWNED(txq);
4904 	MPASS(txp->npkt > 0);
4905 	MPASS(txp->plen < 65536);
4906 	MPASS(m0 != NULL);
4907 	MPASS(m0->m_nextpkt != NULL);
4908 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
4909 	MPASS(available > 0 && available < eq->sidx);
4910 
4911 	ndesc = howmany(txp->len16, EQ_ESIZE / 16);
4912 	MPASS(ndesc <= available);
4913 
4914 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4915 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
4916 	ctrl = V_FW_WR_LEN16(txp->len16);
4917 	wr->equiq_to_len16 = htobe32(ctrl);
4918 	wr->plen = htobe16(txp->plen);
4919 	wr->npkt = txp->npkt;
4920 	wr->r3 = 0;
4921 	wr->type = txp->wr_type;
4922 	flitp = wr + 1;
4923 
4924 	/*
4925 	 * At this point we are 16B into a hardware descriptor.  If checkwrap is
4926 	 * set then we know the WR is going to wrap around somewhere.  We'll
4927 	 * check for that at appropriate points.
4928 	 */
4929 	checkwrap = eq->sidx - ndesc < eq->pidx;
4930 	for (m = m0; m != NULL; m = m->m_nextpkt) {
4931 		if (txp->wr_type == 0) {
4932 			struct ulp_txpkt *ulpmc;
4933 			struct ulptx_idata *ulpsc;
4934 
4935 			/* ULP master command */
4936 			ulpmc = flitp;
4937 			ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
4938 			    V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid));
4939 			ulpmc->len = htobe32(mbuf_len16(m));
4940 
4941 			/* ULP subcommand */
4942 			ulpsc = (void *)(ulpmc + 1);
4943 			ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
4944 			    F_ULP_TX_SC_MORE);
4945 			ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
4946 
4947 			cpl = (void *)(ulpsc + 1);
4948 			if (checkwrap &&
4949 			    (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx])
4950 				cpl = (void *)&eq->desc[0];
4951 		} else {
4952 			cpl = flitp;
4953 		}
4954 
4955 		/* Checksum offload */
4956 		ctrl1 = 0;
4957 		if (needs_l3_csum(m) == 0)
4958 			ctrl1 |= F_TXPKT_IPCSUM_DIS;
4959 		if (needs_l4_csum(m) == 0)
4960 			ctrl1 |= F_TXPKT_L4CSUM_DIS;
4961 		if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4962 		    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4963 			txq->txcsum++;	/* some hardware assistance provided */
4964 
4965 		/* VLAN tag insertion */
4966 		if (needs_vlan_insertion(m)) {
4967 			ctrl1 |= F_TXPKT_VLAN_VLD |
4968 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
4969 			txq->vlan_insertion++;
4970 		}
4971 
4972 		/* CPL header */
4973 		cpl->ctrl0 = txq->cpl_ctrl0;
4974 		cpl->pack = 0;
4975 		cpl->len = htobe16(m->m_pkthdr.len);
4976 		cpl->ctrl1 = htobe64(ctrl1);
4977 
4978 		flitp = cpl + 1;
4979 		if (checkwrap &&
4980 		    (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
4981 			flitp = (void *)&eq->desc[0];
4982 
4983 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap);
4984 
4985 	}
4986 
4987 	if (txp->wr_type == 0) {
4988 		txq->txpkts0_pkts += txp->npkt;
4989 		txq->txpkts0_wrs++;
4990 	} else {
4991 		txq->txpkts1_pkts += txp->npkt;
4992 		txq->txpkts1_wrs++;
4993 	}
4994 
4995 	txsd = &txq->sdesc[eq->pidx];
4996 	txsd->m = m0;
4997 	txsd->desc_used = ndesc;
4998 
4999 	return (ndesc);
5000 }
5001 
5002 /*
5003  * If the SGL ends on an address that is not 16 byte aligned, this function will
5004  * add a 0 filled flit at the end.
5005  */
5006 static void
5007 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap)
5008 {
5009 	struct sge_eq *eq = &txq->eq;
5010 	struct sglist *gl = txq->gl;
5011 	struct sglist_seg *seg;
5012 	__be64 *flitp, *wrap;
5013 	struct ulptx_sgl *usgl;
5014 	int i, nflits, nsegs;
5015 
5016 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
5017 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
5018 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
5019 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
5020 
5021 	get_pkt_gl(m, gl);
5022 	nsegs = gl->sg_nseg;
5023 	MPASS(nsegs > 0);
5024 
5025 	nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
5026 	flitp = (__be64 *)(*to);
5027 	wrap = (__be64 *)(&eq->desc[eq->sidx]);
5028 	seg = &gl->sg_segs[0];
5029 	usgl = (void *)flitp;
5030 
5031 	/*
5032 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
5033 	 * ring, so we're at least 16 bytes away from the status page.  There is
5034 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
5035 	 */
5036 
5037 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
5038 	    V_ULPTX_NSGE(nsegs));
5039 	usgl->len0 = htobe32(seg->ss_len);
5040 	usgl->addr0 = htobe64(seg->ss_paddr);
5041 	seg++;
5042 
5043 	if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) {
5044 
5045 		/* Won't wrap around at all */
5046 
5047 		for (i = 0; i < nsegs - 1; i++, seg++) {
5048 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
5049 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
5050 		}
5051 		if (i & 1)
5052 			usgl->sge[i / 2].len[1] = htobe32(0);
5053 		flitp += nflits;
5054 	} else {
5055 
5056 		/* Will wrap somewhere in the rest of the SGL */
5057 
5058 		/* 2 flits already written, write the rest flit by flit */
5059 		flitp = (void *)(usgl + 1);
5060 		for (i = 0; i < nflits - 2; i++) {
5061 			if (flitp == wrap)
5062 				flitp = (void *)eq->desc;
5063 			*flitp++ = get_flit(seg, nsegs - 1, i);
5064 		}
5065 	}
5066 
5067 	if (nflits & 1) {
5068 		MPASS(((uintptr_t)flitp) & 0xf);
5069 		*flitp++ = 0;
5070 	}
5071 
5072 	MPASS((((uintptr_t)flitp) & 0xf) == 0);
5073 	if (__predict_false(flitp == wrap))
5074 		*to = (void *)eq->desc;
5075 	else
5076 		*to = (void *)flitp;
5077 }
5078 
5079 static inline void
5080 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
5081 {
5082 
5083 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
5084 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
5085 
5086 	if (__predict_true((uintptr_t)(*to) + len <=
5087 	    (uintptr_t)&eq->desc[eq->sidx])) {
5088 		bcopy(from, *to, len);
5089 		(*to) += len;
5090 	} else {
5091 		int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
5092 
5093 		bcopy(from, *to, portion);
5094 		from += portion;
5095 		portion = len - portion;	/* remaining */
5096 		bcopy(from, (void *)eq->desc, portion);
5097 		(*to) = (caddr_t)eq->desc + portion;
5098 	}
5099 }
5100 
5101 static inline void
5102 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n)
5103 {
5104 	u_int db;
5105 
5106 	MPASS(n > 0);
5107 
5108 	db = eq->doorbells;
5109 	if (n > 1)
5110 		clrbit(&db, DOORBELL_WCWR);
5111 	wmb();
5112 
5113 	switch (ffs(db) - 1) {
5114 	case DOORBELL_UDB:
5115 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
5116 		break;
5117 
5118 	case DOORBELL_WCWR: {
5119 		volatile uint64_t *dst, *src;
5120 		int i;
5121 
5122 		/*
5123 		 * Queues whose 128B doorbell segment fits in the page do not
5124 		 * use relative qid (udb_qid is always 0).  Only queues with
5125 		 * doorbell segments can do WCWR.
5126 		 */
5127 		KASSERT(eq->udb_qid == 0 && n == 1,
5128 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
5129 		    __func__, eq->doorbells, n, eq->dbidx, eq));
5130 
5131 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
5132 		    UDBS_DB_OFFSET);
5133 		i = eq->dbidx;
5134 		src = (void *)&eq->desc[i];
5135 		while (src != (void *)&eq->desc[i + 1])
5136 			*dst++ = *src++;
5137 		wmb();
5138 		break;
5139 	}
5140 
5141 	case DOORBELL_UDBWC:
5142 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
5143 		wmb();
5144 		break;
5145 
5146 	case DOORBELL_KDB:
5147 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
5148 		    V_QID(eq->cntxt_id) | V_PIDX(n));
5149 		break;
5150 	}
5151 
5152 	IDXINCR(eq->dbidx, n, eq->sidx);
5153 }
5154 
5155 static inline u_int
5156 reclaimable_tx_desc(struct sge_eq *eq)
5157 {
5158 	uint16_t hw_cidx;
5159 
5160 	hw_cidx = read_hw_cidx(eq);
5161 	return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx));
5162 }
5163 
5164 static inline u_int
5165 total_available_tx_desc(struct sge_eq *eq)
5166 {
5167 	uint16_t hw_cidx, pidx;
5168 
5169 	hw_cidx = read_hw_cidx(eq);
5170 	pidx = eq->pidx;
5171 
5172 	if (pidx == hw_cidx)
5173 		return (eq->sidx - 1);
5174 	else
5175 		return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1);
5176 }
5177 
5178 static inline uint16_t
5179 read_hw_cidx(struct sge_eq *eq)
5180 {
5181 	struct sge_qstat *spg = (void *)&eq->desc[eq->sidx];
5182 	uint16_t cidx = spg->cidx;	/* stable snapshot */
5183 
5184 	return (be16toh(cidx));
5185 }
5186 
5187 /*
5188  * Reclaim 'n' descriptors approximately.
5189  */
5190 static u_int
5191 reclaim_tx_descs(struct sge_txq *txq, u_int n)
5192 {
5193 	struct tx_sdesc *txsd;
5194 	struct sge_eq *eq = &txq->eq;
5195 	u_int can_reclaim, reclaimed;
5196 
5197 	TXQ_LOCK_ASSERT_OWNED(txq);
5198 	MPASS(n > 0);
5199 
5200 	reclaimed = 0;
5201 	can_reclaim = reclaimable_tx_desc(eq);
5202 	while (can_reclaim && reclaimed < n) {
5203 		int ndesc;
5204 		struct mbuf *m, *nextpkt;
5205 
5206 		txsd = &txq->sdesc[eq->cidx];
5207 		ndesc = txsd->desc_used;
5208 
5209 		/* Firmware doesn't return "partial" credits. */
5210 		KASSERT(can_reclaim >= ndesc,
5211 		    ("%s: unexpected number of credits: %d, %d",
5212 		    __func__, can_reclaim, ndesc));
5213 		KASSERT(ndesc != 0,
5214 		    ("%s: descriptor with no credits: cidx %d",
5215 		    __func__, eq->cidx));
5216 
5217 		for (m = txsd->m; m != NULL; m = nextpkt) {
5218 			nextpkt = m->m_nextpkt;
5219 			m->m_nextpkt = NULL;
5220 			m_freem(m);
5221 		}
5222 		reclaimed += ndesc;
5223 		can_reclaim -= ndesc;
5224 		IDXINCR(eq->cidx, ndesc, eq->sidx);
5225 	}
5226 
5227 	return (reclaimed);
5228 }
5229 
5230 static void
5231 tx_reclaim(void *arg, int n)
5232 {
5233 	struct sge_txq *txq = arg;
5234 	struct sge_eq *eq = &txq->eq;
5235 
5236 	do {
5237 		if (TXQ_TRYLOCK(txq) == 0)
5238 			break;
5239 		n = reclaim_tx_descs(txq, 32);
5240 		if (eq->cidx == eq->pidx)
5241 			eq->equeqidx = eq->pidx;
5242 		TXQ_UNLOCK(txq);
5243 	} while (n > 0);
5244 }
5245 
5246 static __be64
5247 get_flit(struct sglist_seg *segs, int nsegs, int idx)
5248 {
5249 	int i = (idx / 3) * 2;
5250 
5251 	switch (idx % 3) {
5252 	case 0: {
5253 		uint64_t rc;
5254 
5255 		rc = (uint64_t)segs[i].ss_len << 32;
5256 		if (i + 1 < nsegs)
5257 			rc |= (uint64_t)(segs[i + 1].ss_len);
5258 
5259 		return (htobe64(rc));
5260 	}
5261 	case 1:
5262 		return (htobe64(segs[i].ss_paddr));
5263 	case 2:
5264 		return (htobe64(segs[i + 1].ss_paddr));
5265 	}
5266 
5267 	return (0);
5268 }
5269 
5270 static void
5271 find_best_refill_source(struct adapter *sc, struct sge_fl *fl, int maxp)
5272 {
5273 	int8_t zidx, hwidx, idx;
5274 	uint16_t region1, region3;
5275 	int spare, spare_needed, n;
5276 	struct sw_zone_info *swz;
5277 	struct hw_buf_info *hwb, *hwb_list = &sc->sge.hw_buf_info[0];
5278 
5279 	/*
5280 	 * Buffer Packing: Look for PAGE_SIZE or larger zone which has a bufsize
5281 	 * large enough for the max payload and cluster metadata.  Otherwise
5282 	 * settle for the largest bufsize that leaves enough room in the cluster
5283 	 * for metadata.
5284 	 *
5285 	 * Without buffer packing: Look for the smallest zone which has a
5286 	 * bufsize large enough for the max payload.  Settle for the largest
5287 	 * bufsize available if there's nothing big enough for max payload.
5288 	 */
5289 	spare_needed = fl->flags & FL_BUF_PACKING ? CL_METADATA_SIZE : 0;
5290 	swz = &sc->sge.sw_zone_info[0];
5291 	hwidx = -1;
5292 	for (zidx = 0; zidx < SW_ZONE_SIZES; zidx++, swz++) {
5293 		if (swz->size > largest_rx_cluster) {
5294 			if (__predict_true(hwidx != -1))
5295 				break;
5296 
5297 			/*
5298 			 * This is a misconfiguration.  largest_rx_cluster is
5299 			 * preventing us from finding a refill source.  See
5300 			 * dev.t5nex.<n>.buffer_sizes to figure out why.
5301 			 */
5302 			device_printf(sc->dev, "largest_rx_cluster=%u leaves no"
5303 			    " refill source for fl %p (dma %u).  Ignored.\n",
5304 			    largest_rx_cluster, fl, maxp);
5305 		}
5306 		for (idx = swz->head_hwidx; idx != -1; idx = hwb->next) {
5307 			hwb = &hwb_list[idx];
5308 			spare = swz->size - hwb->size;
5309 			if (spare < spare_needed)
5310 				continue;
5311 
5312 			hwidx = idx;		/* best option so far */
5313 			if (hwb->size >= maxp) {
5314 
5315 				if ((fl->flags & FL_BUF_PACKING) == 0)
5316 					goto done; /* stop looking (not packing) */
5317 
5318 				if (swz->size >= safest_rx_cluster)
5319 					goto done; /* stop looking (packing) */
5320 			}
5321 			break;		/* keep looking, next zone */
5322 		}
5323 	}
5324 done:
5325 	/* A usable hwidx has been located. */
5326 	MPASS(hwidx != -1);
5327 	hwb = &hwb_list[hwidx];
5328 	zidx = hwb->zidx;
5329 	swz = &sc->sge.sw_zone_info[zidx];
5330 	region1 = 0;
5331 	region3 = swz->size - hwb->size;
5332 
5333 	/*
5334 	 * Stay within this zone and see if there is a better match when mbuf
5335 	 * inlining is allowed.  Remember that the hwidx's are sorted in
5336 	 * decreasing order of size (so in increasing order of spare area).
5337 	 */
5338 	for (idx = hwidx; idx != -1; idx = hwb->next) {
5339 		hwb = &hwb_list[idx];
5340 		spare = swz->size - hwb->size;
5341 
5342 		if (allow_mbufs_in_cluster == 0 || hwb->size < maxp)
5343 			break;
5344 
5345 		/*
5346 		 * Do not inline mbufs if doing so would violate the pad/pack
5347 		 * boundary alignment requirement.
5348 		 */
5349 		if (fl_pad && (MSIZE % sc->params.sge.pad_boundary) != 0)
5350 			continue;
5351 		if (fl->flags & FL_BUF_PACKING &&
5352 		    (MSIZE % sc->params.sge.pack_boundary) != 0)
5353 			continue;
5354 
5355 		if (spare < CL_METADATA_SIZE + MSIZE)
5356 			continue;
5357 		n = (spare - CL_METADATA_SIZE) / MSIZE;
5358 		if (n > howmany(hwb->size, maxp))
5359 			break;
5360 
5361 		hwidx = idx;
5362 		if (fl->flags & FL_BUF_PACKING) {
5363 			region1 = n * MSIZE;
5364 			region3 = spare - region1;
5365 		} else {
5366 			region1 = MSIZE;
5367 			region3 = spare - region1;
5368 			break;
5369 		}
5370 	}
5371 
5372 	KASSERT(zidx >= 0 && zidx < SW_ZONE_SIZES,
5373 	    ("%s: bad zone %d for fl %p, maxp %d", __func__, zidx, fl, maxp));
5374 	KASSERT(hwidx >= 0 && hwidx <= SGE_FLBUF_SIZES,
5375 	    ("%s: bad hwidx %d for fl %p, maxp %d", __func__, hwidx, fl, maxp));
5376 	KASSERT(region1 + sc->sge.hw_buf_info[hwidx].size + region3 ==
5377 	    sc->sge.sw_zone_info[zidx].size,
5378 	    ("%s: bad buffer layout for fl %p, maxp %d. "
5379 		"cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5380 		sc->sge.sw_zone_info[zidx].size, region1,
5381 		sc->sge.hw_buf_info[hwidx].size, region3));
5382 	if (fl->flags & FL_BUF_PACKING || region1 > 0) {
5383 		KASSERT(region3 >= CL_METADATA_SIZE,
5384 		    ("%s: no room for metadata.  fl %p, maxp %d; "
5385 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5386 		    sc->sge.sw_zone_info[zidx].size, region1,
5387 		    sc->sge.hw_buf_info[hwidx].size, region3));
5388 		KASSERT(region1 % MSIZE == 0,
5389 		    ("%s: bad mbuf region for fl %p, maxp %d. "
5390 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5391 		    sc->sge.sw_zone_info[zidx].size, region1,
5392 		    sc->sge.hw_buf_info[hwidx].size, region3));
5393 	}
5394 
5395 	fl->cll_def.zidx = zidx;
5396 	fl->cll_def.hwidx = hwidx;
5397 	fl->cll_def.region1 = region1;
5398 	fl->cll_def.region3 = region3;
5399 }
5400 
5401 static void
5402 find_safe_refill_source(struct adapter *sc, struct sge_fl *fl)
5403 {
5404 	struct sge *s = &sc->sge;
5405 	struct hw_buf_info *hwb;
5406 	struct sw_zone_info *swz;
5407 	int spare;
5408 	int8_t hwidx;
5409 
5410 	if (fl->flags & FL_BUF_PACKING)
5411 		hwidx = s->safe_hwidx2;	/* with room for metadata */
5412 	else if (allow_mbufs_in_cluster && s->safe_hwidx2 != -1) {
5413 		hwidx = s->safe_hwidx2;
5414 		hwb = &s->hw_buf_info[hwidx];
5415 		swz = &s->sw_zone_info[hwb->zidx];
5416 		spare = swz->size - hwb->size;
5417 
5418 		/* no good if there isn't room for an mbuf as well */
5419 		if (spare < CL_METADATA_SIZE + MSIZE)
5420 			hwidx = s->safe_hwidx1;
5421 	} else
5422 		hwidx = s->safe_hwidx1;
5423 
5424 	if (hwidx == -1) {
5425 		/* No fallback source */
5426 		fl->cll_alt.hwidx = -1;
5427 		fl->cll_alt.zidx = -1;
5428 
5429 		return;
5430 	}
5431 
5432 	hwb = &s->hw_buf_info[hwidx];
5433 	swz = &s->sw_zone_info[hwb->zidx];
5434 	spare = swz->size - hwb->size;
5435 	fl->cll_alt.hwidx = hwidx;
5436 	fl->cll_alt.zidx = hwb->zidx;
5437 	if (allow_mbufs_in_cluster &&
5438 	    (fl_pad == 0 || (MSIZE % sc->params.sge.pad_boundary) == 0))
5439 		fl->cll_alt.region1 = ((spare - CL_METADATA_SIZE) / MSIZE) * MSIZE;
5440 	else
5441 		fl->cll_alt.region1 = 0;
5442 	fl->cll_alt.region3 = spare - fl->cll_alt.region1;
5443 }
5444 
5445 static void
5446 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
5447 {
5448 	mtx_lock(&sc->sfl_lock);
5449 	FL_LOCK(fl);
5450 	if ((fl->flags & FL_DOOMED) == 0) {
5451 		fl->flags |= FL_STARVING;
5452 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
5453 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
5454 	}
5455 	FL_UNLOCK(fl);
5456 	mtx_unlock(&sc->sfl_lock);
5457 }
5458 
5459 static void
5460 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq)
5461 {
5462 	struct sge_wrq *wrq = (void *)eq;
5463 
5464 	atomic_readandclear_int(&eq->equiq);
5465 	taskqueue_enqueue(sc->tq[eq->tx_chan], &wrq->wrq_tx_task);
5466 }
5467 
5468 static void
5469 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq)
5470 {
5471 	struct sge_txq *txq = (void *)eq;
5472 
5473 	MPASS((eq->flags & EQ_TYPEMASK) == EQ_ETH);
5474 
5475 	atomic_readandclear_int(&eq->equiq);
5476 	mp_ring_check_drainage(txq->r, 0);
5477 	taskqueue_enqueue(sc->tq[eq->tx_chan], &txq->tx_reclaim_task);
5478 }
5479 
5480 static int
5481 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
5482     struct mbuf *m)
5483 {
5484 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
5485 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
5486 	struct adapter *sc = iq->adapter;
5487 	struct sge *s = &sc->sge;
5488 	struct sge_eq *eq;
5489 	static void (*h[])(struct adapter *, struct sge_eq *) = {NULL,
5490 		&handle_wrq_egr_update, &handle_eth_egr_update,
5491 		&handle_wrq_egr_update};
5492 
5493 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5494 	    rss->opcode));
5495 
5496 	eq = s->eqmap[qid - s->eq_start - s->eq_base];
5497 	(*h[eq->flags & EQ_TYPEMASK])(sc, eq);
5498 
5499 	return (0);
5500 }
5501 
5502 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
5503 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
5504     offsetof(struct cpl_fw6_msg, data));
5505 
5506 static int
5507 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
5508 {
5509 	struct adapter *sc = iq->adapter;
5510 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
5511 
5512 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5513 	    rss->opcode));
5514 
5515 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
5516 		const struct rss_header *rss2;
5517 
5518 		rss2 = (const struct rss_header *)&cpl->data[0];
5519 		return (t4_cpl_handler[rss2->opcode](iq, rss2, m));
5520 	}
5521 
5522 	return (t4_fw_msg_handler[cpl->type](sc, &cpl->data[0]));
5523 }
5524 
5525 /**
5526  *	t4_handle_wrerr_rpl - process a FW work request error message
5527  *	@adap: the adapter
5528  *	@rpl: start of the FW message
5529  */
5530 static int
5531 t4_handle_wrerr_rpl(struct adapter *adap, const __be64 *rpl)
5532 {
5533 	u8 opcode = *(const u8 *)rpl;
5534 	const struct fw_error_cmd *e = (const void *)rpl;
5535 	unsigned int i;
5536 
5537 	if (opcode != FW_ERROR_CMD) {
5538 		log(LOG_ERR,
5539 		    "%s: Received WRERR_RPL message with opcode %#x\n",
5540 		    device_get_nameunit(adap->dev), opcode);
5541 		return (EINVAL);
5542 	}
5543 	log(LOG_ERR, "%s: FW_ERROR (%s) ", device_get_nameunit(adap->dev),
5544 	    G_FW_ERROR_CMD_FATAL(be32toh(e->op_to_type)) ? "fatal" :
5545 	    "non-fatal");
5546 	switch (G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type))) {
5547 	case FW_ERROR_TYPE_EXCEPTION:
5548 		log(LOG_ERR, "exception info:\n");
5549 		for (i = 0; i < nitems(e->u.exception.info); i++)
5550 			log(LOG_ERR, "%s%08x", i == 0 ? "\t" : " ",
5551 			    be32toh(e->u.exception.info[i]));
5552 		log(LOG_ERR, "\n");
5553 		break;
5554 	case FW_ERROR_TYPE_HWMODULE:
5555 		log(LOG_ERR, "HW module regaddr %08x regval %08x\n",
5556 		    be32toh(e->u.hwmodule.regaddr),
5557 		    be32toh(e->u.hwmodule.regval));
5558 		break;
5559 	case FW_ERROR_TYPE_WR:
5560 		log(LOG_ERR, "WR cidx %d PF %d VF %d eqid %d hdr:\n",
5561 		    be16toh(e->u.wr.cidx),
5562 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.wr.pfn_vfn)),
5563 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.wr.pfn_vfn)),
5564 		    be32toh(e->u.wr.eqid));
5565 		for (i = 0; i < nitems(e->u.wr.wrhdr); i++)
5566 			log(LOG_ERR, "%s%02x", i == 0 ? "\t" : " ",
5567 			    e->u.wr.wrhdr[i]);
5568 		log(LOG_ERR, "\n");
5569 		break;
5570 	case FW_ERROR_TYPE_ACL:
5571 		log(LOG_ERR, "ACL cidx %d PF %d VF %d eqid %d %s",
5572 		    be16toh(e->u.acl.cidx),
5573 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.acl.pfn_vfn)),
5574 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.acl.pfn_vfn)),
5575 		    be32toh(e->u.acl.eqid),
5576 		    G_FW_ERROR_CMD_MV(be16toh(e->u.acl.mv_pkd)) ? "vlanid" :
5577 		    "MAC");
5578 		for (i = 0; i < nitems(e->u.acl.val); i++)
5579 			log(LOG_ERR, " %02x", e->u.acl.val[i]);
5580 		log(LOG_ERR, "\n");
5581 		break;
5582 	default:
5583 		log(LOG_ERR, "type %#x\n",
5584 		    G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type)));
5585 		return (EINVAL);
5586 	}
5587 	return (0);
5588 }
5589 
5590 static int
5591 sysctl_uint16(SYSCTL_HANDLER_ARGS)
5592 {
5593 	uint16_t *id = arg1;
5594 	int i = *id;
5595 
5596 	return sysctl_handle_int(oidp, &i, 0, req);
5597 }
5598 
5599 static int
5600 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
5601 {
5602 	struct sge *s = arg1;
5603 	struct hw_buf_info *hwb = &s->hw_buf_info[0];
5604 	struct sw_zone_info *swz = &s->sw_zone_info[0];
5605 	int i, rc;
5606 	struct sbuf sb;
5607 	char c;
5608 
5609 	sbuf_new(&sb, NULL, 32, SBUF_AUTOEXTEND);
5610 	for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
5611 		if (hwb->zidx >= 0 && swz[hwb->zidx].size <= largest_rx_cluster)
5612 			c = '*';
5613 		else
5614 			c = '\0';
5615 
5616 		sbuf_printf(&sb, "%u%c ", hwb->size, c);
5617 	}
5618 	sbuf_trim(&sb);
5619 	sbuf_finish(&sb);
5620 	rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
5621 	sbuf_delete(&sb);
5622 	return (rc);
5623 }
5624 
5625 #ifdef RATELIMIT
5626 /*
5627  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
5628  */
5629 static inline u_int
5630 txpkt_eo_len16(u_int nsegs, u_int immhdrs, u_int tso)
5631 {
5632 	u_int n;
5633 
5634 	MPASS(immhdrs > 0);
5635 
5636 	n = roundup2(sizeof(struct fw_eth_tx_eo_wr) +
5637 	    sizeof(struct cpl_tx_pkt_core) + immhdrs, 16);
5638 	if (__predict_false(nsegs == 0))
5639 		goto done;
5640 
5641 	nsegs--; /* first segment is part of ulptx_sgl */
5642 	n += sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
5643 	if (tso)
5644 		n += sizeof(struct cpl_tx_pkt_lso_core);
5645 
5646 done:
5647 	return (howmany(n, 16));
5648 }
5649 
5650 #define ETID_FLOWC_NPARAMS 6
5651 #define ETID_FLOWC_LEN (roundup2((sizeof(struct fw_flowc_wr) + \
5652     ETID_FLOWC_NPARAMS * sizeof(struct fw_flowc_mnemval)), 16))
5653 #define ETID_FLOWC_LEN16 (howmany(ETID_FLOWC_LEN, 16))
5654 
5655 static int
5656 send_etid_flowc_wr(struct cxgbe_snd_tag *cst, struct port_info *pi,
5657     struct vi_info *vi)
5658 {
5659 	struct wrq_cookie cookie;
5660 	u_int pfvf = G_FW_VIID_PFN(vi->viid) << S_FW_VIID_PFN;
5661 	struct fw_flowc_wr *flowc;
5662 
5663 	mtx_assert(&cst->lock, MA_OWNED);
5664 	MPASS((cst->flags & (EO_FLOWC_PENDING | EO_FLOWC_RPL_PENDING)) ==
5665 	    EO_FLOWC_PENDING);
5666 
5667 	flowc = start_wrq_wr(cst->eo_txq, ETID_FLOWC_LEN16, &cookie);
5668 	if (__predict_false(flowc == NULL))
5669 		return (ENOMEM);
5670 
5671 	bzero(flowc, ETID_FLOWC_LEN);
5672 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
5673 	    V_FW_FLOWC_WR_NPARAMS(ETID_FLOWC_NPARAMS) | V_FW_WR_COMPL(0));
5674 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(ETID_FLOWC_LEN16) |
5675 	    V_FW_WR_FLOWID(cst->etid));
5676 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN;
5677 	flowc->mnemval[0].val = htobe32(pfvf);
5678 	flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH;
5679 	flowc->mnemval[1].val = htobe32(pi->tx_chan);
5680 	flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT;
5681 	flowc->mnemval[2].val = htobe32(pi->tx_chan);
5682 	flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID;
5683 	flowc->mnemval[3].val = htobe32(cst->iqid);
5684 	flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_EOSTATE;
5685 	flowc->mnemval[4].val = htobe32(FW_FLOWC_MNEM_EOSTATE_ESTABLISHED);
5686 	flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS;
5687 	flowc->mnemval[5].val = htobe32(cst->schedcl);
5688 
5689 	commit_wrq_wr(cst->eo_txq, flowc, &cookie);
5690 
5691 	cst->flags &= ~EO_FLOWC_PENDING;
5692 	cst->flags |= EO_FLOWC_RPL_PENDING;
5693 	MPASS(cst->tx_credits >= ETID_FLOWC_LEN16);	/* flowc is first WR. */
5694 	cst->tx_credits -= ETID_FLOWC_LEN16;
5695 
5696 	return (0);
5697 }
5698 
5699 #define ETID_FLUSH_LEN16 (howmany(sizeof (struct fw_flowc_wr), 16))
5700 
5701 void
5702 send_etid_flush_wr(struct cxgbe_snd_tag *cst)
5703 {
5704 	struct fw_flowc_wr *flowc;
5705 	struct wrq_cookie cookie;
5706 
5707 	mtx_assert(&cst->lock, MA_OWNED);
5708 
5709 	flowc = start_wrq_wr(cst->eo_txq, ETID_FLUSH_LEN16, &cookie);
5710 	if (__predict_false(flowc == NULL))
5711 		CXGBE_UNIMPLEMENTED(__func__);
5712 
5713 	bzero(flowc, ETID_FLUSH_LEN16 * 16);
5714 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
5715 	    V_FW_FLOWC_WR_NPARAMS(0) | F_FW_WR_COMPL);
5716 	flowc->flowid_len16 = htobe32(V_FW_WR_LEN16(ETID_FLUSH_LEN16) |
5717 	    V_FW_WR_FLOWID(cst->etid));
5718 
5719 	commit_wrq_wr(cst->eo_txq, flowc, &cookie);
5720 
5721 	cst->flags |= EO_FLUSH_RPL_PENDING;
5722 	MPASS(cst->tx_credits >= ETID_FLUSH_LEN16);
5723 	cst->tx_credits -= ETID_FLUSH_LEN16;
5724 	cst->ncompl++;
5725 }
5726 
5727 static void
5728 write_ethofld_wr(struct cxgbe_snd_tag *cst, struct fw_eth_tx_eo_wr *wr,
5729     struct mbuf *m0, int compl)
5730 {
5731 	struct cpl_tx_pkt_core *cpl;
5732 	uint64_t ctrl1;
5733 	uint32_t ctrl;	/* used in many unrelated places */
5734 	int len16, pktlen, nsegs, immhdrs;
5735 	caddr_t dst;
5736 	uintptr_t p;
5737 	struct ulptx_sgl *usgl;
5738 	struct sglist sg;
5739 	struct sglist_seg segs[38];	/* XXX: find real limit.  XXX: get off the stack */
5740 
5741 	mtx_assert(&cst->lock, MA_OWNED);
5742 	M_ASSERTPKTHDR(m0);
5743 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
5744 	    m0->m_pkthdr.l4hlen > 0,
5745 	    ("%s: ethofld mbuf %p is missing header lengths", __func__, m0));
5746 
5747 	len16 = mbuf_eo_len16(m0);
5748 	nsegs = mbuf_eo_nsegs(m0);
5749 	pktlen = m0->m_pkthdr.len;
5750 	ctrl = sizeof(struct cpl_tx_pkt_core);
5751 	if (needs_tso(m0))
5752 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
5753 	immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen + m0->m_pkthdr.l4hlen;
5754 	ctrl += immhdrs;
5755 
5756 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_EO_WR) |
5757 	    V_FW_ETH_TX_EO_WR_IMMDLEN(ctrl) | V_FW_WR_COMPL(!!compl));
5758 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(len16) |
5759 	    V_FW_WR_FLOWID(cst->etid));
5760 	wr->r3 = 0;
5761 	if (needs_udp_csum(m0)) {
5762 		wr->u.udpseg.type = FW_ETH_TX_EO_TYPE_UDPSEG;
5763 		wr->u.udpseg.ethlen = m0->m_pkthdr.l2hlen;
5764 		wr->u.udpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
5765 		wr->u.udpseg.udplen = m0->m_pkthdr.l4hlen;
5766 		wr->u.udpseg.rtplen = 0;
5767 		wr->u.udpseg.r4 = 0;
5768 		wr->u.udpseg.mss = htobe16(pktlen - immhdrs);
5769 		wr->u.udpseg.schedpktsize = wr->u.udpseg.mss;
5770 		wr->u.udpseg.plen = htobe32(pktlen - immhdrs);
5771 		cpl = (void *)(wr + 1);
5772 	} else {
5773 		MPASS(needs_tcp_csum(m0));
5774 		wr->u.tcpseg.type = FW_ETH_TX_EO_TYPE_TCPSEG;
5775 		wr->u.tcpseg.ethlen = m0->m_pkthdr.l2hlen;
5776 		wr->u.tcpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
5777 		wr->u.tcpseg.tcplen = m0->m_pkthdr.l4hlen;
5778 		wr->u.tcpseg.tsclk_tsoff = mbuf_eo_tsclk_tsoff(m0);
5779 		wr->u.tcpseg.r4 = 0;
5780 		wr->u.tcpseg.r5 = 0;
5781 		wr->u.tcpseg.plen = htobe32(pktlen - immhdrs);
5782 
5783 		if (needs_tso(m0)) {
5784 			struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
5785 
5786 			wr->u.tcpseg.mss = htobe16(m0->m_pkthdr.tso_segsz);
5787 
5788 			ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) |
5789 			    F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE |
5790 			    V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) |
5791 			    V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
5792 			if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
5793 				ctrl |= V_LSO_ETHHDR_LEN(1);
5794 			if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
5795 				ctrl |= F_LSO_IPV6;
5796 			lso->lso_ctrl = htobe32(ctrl);
5797 			lso->ipid_ofst = htobe16(0);
5798 			lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
5799 			lso->seqno_offset = htobe32(0);
5800 			lso->len = htobe32(pktlen);
5801 
5802 			cpl = (void *)(lso + 1);
5803 		} else {
5804 			wr->u.tcpseg.mss = htobe16(0xffff);
5805 			cpl = (void *)(wr + 1);
5806 		}
5807 	}
5808 
5809 	/* Checksum offload must be requested for ethofld. */
5810 	ctrl1 = 0;
5811 	MPASS(needs_l4_csum(m0));
5812 
5813 	/* VLAN tag insertion */
5814 	if (needs_vlan_insertion(m0)) {
5815 		ctrl1 |= F_TXPKT_VLAN_VLD |
5816 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
5817 	}
5818 
5819 	/* CPL header */
5820 	cpl->ctrl0 = cst->ctrl0;
5821 	cpl->pack = 0;
5822 	cpl->len = htobe16(pktlen);
5823 	cpl->ctrl1 = htobe64(ctrl1);
5824 
5825 	/* Copy Ethernet, IP & TCP/UDP hdrs as immediate data */
5826 	p = (uintptr_t)(cpl + 1);
5827 	m_copydata(m0, 0, immhdrs, (void *)p);
5828 
5829 	/* SGL */
5830 	dst = (void *)(cpl + 1);
5831 	if (nsegs > 0) {
5832 		int i, pad;
5833 
5834 		/* zero-pad upto next 16Byte boundary, if not 16Byte aligned */
5835 		p += immhdrs;
5836 		pad = 16 - (immhdrs & 0xf);
5837 		bzero((void *)p, pad);
5838 
5839 		usgl = (void *)(p + pad);
5840 		usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
5841 		    V_ULPTX_NSGE(nsegs));
5842 
5843 		sglist_init(&sg, nitems(segs), segs);
5844 		for (; m0 != NULL; m0 = m0->m_next) {
5845 			if (__predict_false(m0->m_len == 0))
5846 				continue;
5847 			if (immhdrs >= m0->m_len) {
5848 				immhdrs -= m0->m_len;
5849 				continue;
5850 			}
5851 
5852 			sglist_append(&sg, mtod(m0, char *) + immhdrs,
5853 			    m0->m_len - immhdrs);
5854 			immhdrs = 0;
5855 		}
5856 		MPASS(sg.sg_nseg == nsegs);
5857 
5858 		/*
5859 		 * Zero pad last 8B in case the WR doesn't end on a 16B
5860 		 * boundary.
5861 		 */
5862 		*(uint64_t *)((char *)wr + len16 * 16 - 8) = 0;
5863 
5864 		usgl->len0 = htobe32(segs[0].ss_len);
5865 		usgl->addr0 = htobe64(segs[0].ss_paddr);
5866 		for (i = 0; i < nsegs - 1; i++) {
5867 			usgl->sge[i / 2].len[i & 1] = htobe32(segs[i + 1].ss_len);
5868 			usgl->sge[i / 2].addr[i & 1] = htobe64(segs[i + 1].ss_paddr);
5869 		}
5870 		if (i & 1)
5871 			usgl->sge[i / 2].len[1] = htobe32(0);
5872 	}
5873 
5874 }
5875 
5876 static void
5877 ethofld_tx(struct cxgbe_snd_tag *cst)
5878 {
5879 	struct mbuf *m;
5880 	struct wrq_cookie cookie;
5881 	int next_credits, compl;
5882 	struct fw_eth_tx_eo_wr *wr;
5883 
5884 	mtx_assert(&cst->lock, MA_OWNED);
5885 
5886 	while ((m = mbufq_first(&cst->pending_tx)) != NULL) {
5887 		M_ASSERTPKTHDR(m);
5888 
5889 		/* How many len16 credits do we need to send this mbuf. */
5890 		next_credits = mbuf_eo_len16(m);
5891 		MPASS(next_credits > 0);
5892 		if (next_credits > cst->tx_credits) {
5893 			/*
5894 			 * Tx will make progress eventually because there is at
5895 			 * least one outstanding fw4_ack that will return
5896 			 * credits and kick the tx.
5897 			 */
5898 			MPASS(cst->ncompl > 0);
5899 			return;
5900 		}
5901 		wr = start_wrq_wr(cst->eo_txq, next_credits, &cookie);
5902 		if (__predict_false(wr == NULL)) {
5903 			/* XXX: wishful thinking, not a real assertion. */
5904 			MPASS(cst->ncompl > 0);
5905 			return;
5906 		}
5907 		cst->tx_credits -= next_credits;
5908 		cst->tx_nocompl += next_credits;
5909 		compl = cst->ncompl == 0 || cst->tx_nocompl >= cst->tx_total / 2;
5910 		ETHER_BPF_MTAP(cst->com.ifp, m);
5911 		write_ethofld_wr(cst, wr, m, compl);
5912 		commit_wrq_wr(cst->eo_txq, wr, &cookie);
5913 		if (compl) {
5914 			cst->ncompl++;
5915 			cst->tx_nocompl	= 0;
5916 		}
5917 		(void) mbufq_dequeue(&cst->pending_tx);
5918 		mbufq_enqueue(&cst->pending_fwack, m);
5919 	}
5920 }
5921 
5922 int
5923 ethofld_transmit(struct ifnet *ifp, struct mbuf *m0)
5924 {
5925 	struct cxgbe_snd_tag *cst;
5926 	int rc;
5927 
5928 	MPASS(m0->m_nextpkt == NULL);
5929 	MPASS(m0->m_pkthdr.snd_tag != NULL);
5930 	cst = mst_to_cst(m0->m_pkthdr.snd_tag);
5931 
5932 	mtx_lock(&cst->lock);
5933 	MPASS(cst->flags & EO_SND_TAG_REF);
5934 
5935 	if (__predict_false(cst->flags & EO_FLOWC_PENDING)) {
5936 		struct vi_info *vi = ifp->if_softc;
5937 		struct port_info *pi = vi->pi;
5938 		struct adapter *sc = pi->adapter;
5939 		const uint32_t rss_mask = vi->rss_size - 1;
5940 		uint32_t rss_hash;
5941 
5942 		cst->eo_txq = &sc->sge.ofld_txq[vi->first_ofld_txq];
5943 		if (M_HASHTYPE_ISHASH(m0))
5944 			rss_hash = m0->m_pkthdr.flowid;
5945 		else
5946 			rss_hash = arc4random();
5947 		/* We assume RSS hashing */
5948 		cst->iqid = vi->rss[rss_hash & rss_mask];
5949 		cst->eo_txq += rss_hash % vi->nofldtxq;
5950 		rc = send_etid_flowc_wr(cst, pi, vi);
5951 		if (rc != 0)
5952 			goto done;
5953 	}
5954 
5955 	if (__predict_false(cst->plen + m0->m_pkthdr.len > eo_max_backlog)) {
5956 		rc = ENOBUFS;
5957 		goto done;
5958 	}
5959 
5960 	mbufq_enqueue(&cst->pending_tx, m0);
5961 	cst->plen += m0->m_pkthdr.len;
5962 
5963 	ethofld_tx(cst);
5964 	rc = 0;
5965 done:
5966 	mtx_unlock(&cst->lock);
5967 	if (__predict_false(rc != 0))
5968 		m_freem(m0);
5969 	return (rc);
5970 }
5971 
5972 static int
5973 ethofld_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
5974 {
5975 	struct adapter *sc = iq->adapter;
5976 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
5977 	struct mbuf *m;
5978 	u_int etid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
5979 	struct cxgbe_snd_tag *cst;
5980 	uint8_t credits = cpl->credits;
5981 
5982 	cst = lookup_etid(sc, etid);
5983 	mtx_lock(&cst->lock);
5984 	if (__predict_false(cst->flags & EO_FLOWC_RPL_PENDING)) {
5985 		MPASS(credits >= ETID_FLOWC_LEN16);
5986 		credits -= ETID_FLOWC_LEN16;
5987 		cst->flags &= ~EO_FLOWC_RPL_PENDING;
5988 	}
5989 
5990 	KASSERT(cst->ncompl > 0,
5991 	    ("%s: etid %u (%p) wasn't expecting completion.",
5992 	    __func__, etid, cst));
5993 	cst->ncompl--;
5994 
5995 	while (credits > 0) {
5996 		m = mbufq_dequeue(&cst->pending_fwack);
5997 		if (__predict_false(m == NULL)) {
5998 			/*
5999 			 * The remaining credits are for the final flush that
6000 			 * was issued when the tag was freed by the kernel.
6001 			 */
6002 			MPASS((cst->flags &
6003 			    (EO_FLUSH_RPL_PENDING | EO_SND_TAG_REF)) ==
6004 			    EO_FLUSH_RPL_PENDING);
6005 			MPASS(credits == ETID_FLUSH_LEN16);
6006 			MPASS(cst->tx_credits + cpl->credits == cst->tx_total);
6007 			MPASS(cst->ncompl == 0);
6008 
6009 			cst->flags &= ~EO_FLUSH_RPL_PENDING;
6010 			cst->tx_credits += cpl->credits;
6011 freetag:
6012 			cxgbe_snd_tag_free_locked(cst);
6013 			return (0);	/* cst is gone. */
6014 		}
6015 		KASSERT(m != NULL,
6016 		    ("%s: too many credits (%u, %u)", __func__, cpl->credits,
6017 		    credits));
6018 		KASSERT(credits >= mbuf_eo_len16(m),
6019 		    ("%s: too few credits (%u, %u, %u)", __func__,
6020 		    cpl->credits, credits, mbuf_eo_len16(m)));
6021 		credits -= mbuf_eo_len16(m);
6022 		cst->plen -= m->m_pkthdr.len;
6023 		m_freem(m);
6024 	}
6025 
6026 	cst->tx_credits += cpl->credits;
6027 	MPASS(cst->tx_credits <= cst->tx_total);
6028 
6029 	m = mbufq_first(&cst->pending_tx);
6030 	if (m != NULL && cst->tx_credits >= mbuf_eo_len16(m))
6031 		ethofld_tx(cst);
6032 
6033 	if (__predict_false((cst->flags & EO_SND_TAG_REF) == 0) &&
6034 	    cst->ncompl == 0) {
6035 		if (cst->tx_credits == cst->tx_total)
6036 			goto freetag;
6037 		else {
6038 			MPASS((cst->flags & EO_FLUSH_RPL_PENDING) == 0);
6039 			send_etid_flush_wr(cst);
6040 		}
6041 	}
6042 
6043 	mtx_unlock(&cst->lock);
6044 
6045 	return (0);
6046 }
6047 #endif
6048