xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision 08c4a937a6685f05667996228898521fc453f8f3)
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 	t4_intr_err(arg);
1383 	t4_intr_evt(fwq);
1384 }
1385 
1386 /*
1387  * Interrupt handler for errors (installed directly when multiple interrupts are
1388  * being used, or called by t4_intr_all).
1389  */
1390 void
1391 t4_intr_err(void *arg)
1392 {
1393 	struct adapter *sc = arg;
1394 
1395 	t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
1396 	t4_slow_intr_handler(sc);
1397 }
1398 
1399 /*
1400  * Interrupt handler for iq-only queues.  The firmware event queue is the only
1401  * such queue right now.
1402  */
1403 void
1404 t4_intr_evt(void *arg)
1405 {
1406 	struct sge_iq *iq = arg;
1407 
1408 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1409 		service_iq(iq, 0);
1410 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1411 	}
1412 }
1413 
1414 /*
1415  * Interrupt handler for iq+fl queues.
1416  */
1417 void
1418 t4_intr(void *arg)
1419 {
1420 	struct sge_iq *iq = arg;
1421 
1422 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1423 		service_iq_fl(iq, 0);
1424 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1425 	}
1426 }
1427 
1428 #ifdef DEV_NETMAP
1429 /*
1430  * Interrupt handler for netmap rx queues.
1431  */
1432 void
1433 t4_nm_intr(void *arg)
1434 {
1435 	struct sge_nm_rxq *nm_rxq = arg;
1436 
1437 	if (atomic_cmpset_int(&nm_rxq->nm_state, NM_ON, NM_BUSY)) {
1438 		service_nm_rxq(nm_rxq);
1439 		(void) atomic_cmpset_int(&nm_rxq->nm_state, NM_BUSY, NM_ON);
1440 	}
1441 }
1442 
1443 /*
1444  * Interrupt handler for vectors shared between NIC and netmap rx queues.
1445  */
1446 void
1447 t4_vi_intr(void *arg)
1448 {
1449 	struct irq *irq = arg;
1450 
1451 	MPASS(irq->nm_rxq != NULL);
1452 	t4_nm_intr(irq->nm_rxq);
1453 
1454 	MPASS(irq->rxq != NULL);
1455 	t4_intr(irq->rxq);
1456 }
1457 #endif
1458 
1459 /*
1460  * Deals with interrupts on an iq-only (no freelist) queue.
1461  */
1462 static int
1463 service_iq(struct sge_iq *iq, int budget)
1464 {
1465 	struct sge_iq *q;
1466 	struct adapter *sc = iq->adapter;
1467 	struct iq_desc *d = &iq->desc[iq->cidx];
1468 	int ndescs = 0, limit;
1469 	int rsp_type;
1470 	uint32_t lq;
1471 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1472 
1473 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1474 	KASSERT((iq->flags & IQ_HAS_FL) == 0,
1475 	    ("%s: called for iq %p with fl (iq->flags 0x%x)", __func__, iq,
1476 	    iq->flags));
1477 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1478 	MPASS((iq->flags & IQ_LRO_ENABLED) == 0);
1479 
1480 	limit = budget ? budget : iq->qsize / 16;
1481 
1482 	/*
1483 	 * We always come back and check the descriptor ring for new indirect
1484 	 * interrupts and other responses after running a single handler.
1485 	 */
1486 	for (;;) {
1487 		while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1488 
1489 			rmb();
1490 
1491 			rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1492 			lq = be32toh(d->rsp.pldbuflen_qid);
1493 
1494 			switch (rsp_type) {
1495 			case X_RSPD_TYPE_FLBUF:
1496 				panic("%s: data for an iq (%p) with no freelist",
1497 				    __func__, iq);
1498 
1499 				/* NOTREACHED */
1500 
1501 			case X_RSPD_TYPE_CPL:
1502 				KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1503 				    ("%s: bad opcode %02x.", __func__,
1504 				    d->rss.opcode));
1505 				t4_cpl_handler[d->rss.opcode](iq, &d->rss, NULL);
1506 				break;
1507 
1508 			case X_RSPD_TYPE_INTR:
1509 				/*
1510 				 * There are 1K interrupt-capable queues (qids 0
1511 				 * through 1023).  A response type indicating a
1512 				 * forwarded interrupt with a qid >= 1K is an
1513 				 * iWARP async notification.
1514 				 */
1515 				if (__predict_true(lq >= 1024)) {
1516 					t4_an_handler(iq, &d->rsp);
1517 					break;
1518 				}
1519 
1520 				q = sc->sge.iqmap[lq - sc->sge.iq_start -
1521 				    sc->sge.iq_base];
1522 				if (atomic_cmpset_int(&q->state, IQS_IDLE,
1523 				    IQS_BUSY)) {
1524 					if (service_iq_fl(q, q->qsize / 16) == 0) {
1525 						(void) atomic_cmpset_int(&q->state,
1526 						    IQS_BUSY, IQS_IDLE);
1527 					} else {
1528 						STAILQ_INSERT_TAIL(&iql, q,
1529 						    link);
1530 					}
1531 				}
1532 				break;
1533 
1534 			default:
1535 				KASSERT(0,
1536 				    ("%s: illegal response type %d on iq %p",
1537 				    __func__, rsp_type, iq));
1538 				log(LOG_ERR,
1539 				    "%s: illegal response type %d on iq %p",
1540 				    device_get_nameunit(sc->dev), rsp_type, iq);
1541 				break;
1542 			}
1543 
1544 			d++;
1545 			if (__predict_false(++iq->cidx == iq->sidx)) {
1546 				iq->cidx = 0;
1547 				iq->gen ^= F_RSPD_GEN;
1548 				d = &iq->desc[0];
1549 			}
1550 			if (__predict_false(++ndescs == limit)) {
1551 				t4_write_reg(sc, sc->sge_gts_reg,
1552 				    V_CIDXINC(ndescs) |
1553 				    V_INGRESSQID(iq->cntxt_id) |
1554 				    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1555 				ndescs = 0;
1556 
1557 				if (budget) {
1558 					return (EINPROGRESS);
1559 				}
1560 			}
1561 		}
1562 
1563 		if (STAILQ_EMPTY(&iql))
1564 			break;
1565 
1566 		/*
1567 		 * Process the head only, and send it to the back of the list if
1568 		 * it's still not done.
1569 		 */
1570 		q = STAILQ_FIRST(&iql);
1571 		STAILQ_REMOVE_HEAD(&iql, link);
1572 		if (service_iq_fl(q, q->qsize / 8) == 0)
1573 			(void) atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1574 		else
1575 			STAILQ_INSERT_TAIL(&iql, q, link);
1576 	}
1577 
1578 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1579 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1580 
1581 	return (0);
1582 }
1583 
1584 static inline int
1585 sort_before_lro(struct lro_ctrl *lro)
1586 {
1587 
1588 	return (lro->lro_mbuf_max != 0);
1589 }
1590 
1591 static inline uint64_t
1592 last_flit_to_ns(struct adapter *sc, uint64_t lf)
1593 {
1594 	uint64_t n = be64toh(lf) & 0xfffffffffffffff;	/* 60b, not 64b. */
1595 
1596 	if (n > UINT64_MAX / 1000000)
1597 		return (n / sc->params.vpd.cclk * 1000000);
1598 	else
1599 		return (n * 1000000 / sc->params.vpd.cclk);
1600 }
1601 
1602 /*
1603  * Deals with interrupts on an iq+fl queue.
1604  */
1605 static int
1606 service_iq_fl(struct sge_iq *iq, int budget)
1607 {
1608 	struct sge_rxq *rxq = iq_to_rxq(iq);
1609 	struct sge_fl *fl;
1610 	struct adapter *sc = iq->adapter;
1611 	struct iq_desc *d = &iq->desc[iq->cidx];
1612 	int ndescs = 0, limit;
1613 	int rsp_type, refill, starved;
1614 	uint32_t lq;
1615 	uint16_t fl_hw_cidx;
1616 	struct mbuf *m0;
1617 #if defined(INET) || defined(INET6)
1618 	const struct timeval lro_timeout = {0, sc->lro_timeout};
1619 	struct lro_ctrl *lro = &rxq->lro;
1620 #endif
1621 
1622 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1623 	MPASS(iq->flags & IQ_HAS_FL);
1624 
1625 	limit = budget ? budget : iq->qsize / 16;
1626 	fl = &rxq->fl;
1627 	fl_hw_cidx = fl->hw_cidx;	/* stable snapshot */
1628 
1629 #if defined(INET) || defined(INET6)
1630 	if (iq->flags & IQ_ADJ_CREDIT) {
1631 		MPASS(sort_before_lro(lro));
1632 		iq->flags &= ~IQ_ADJ_CREDIT;
1633 		if ((d->rsp.u.type_gen & F_RSPD_GEN) != iq->gen) {
1634 			tcp_lro_flush_all(lro);
1635 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(1) |
1636 			    V_INGRESSQID((u32)iq->cntxt_id) |
1637 			    V_SEINTARM(iq->intr_params));
1638 			return (0);
1639 		}
1640 		ndescs = 1;
1641 	}
1642 #else
1643 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1644 #endif
1645 
1646 	while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1647 
1648 		rmb();
1649 
1650 		refill = 0;
1651 		m0 = NULL;
1652 		rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1653 		lq = be32toh(d->rsp.pldbuflen_qid);
1654 
1655 		switch (rsp_type) {
1656 		case X_RSPD_TYPE_FLBUF:
1657 
1658 			m0 = get_fl_payload(sc, fl, lq);
1659 			if (__predict_false(m0 == NULL))
1660 				goto out;
1661 			refill = IDXDIFF(fl->hw_cidx, fl_hw_cidx, fl->sidx) > 2;
1662 
1663 			if (iq->flags & IQ_RX_TIMESTAMP) {
1664 				/*
1665 				 * Fill up rcv_tstmp but do not set M_TSTMP.
1666 				 * rcv_tstmp is not in the format that the
1667 				 * kernel expects and we don't want to mislead
1668 				 * it.  For now this is only for custom code
1669 				 * that knows how to interpret cxgbe's stamp.
1670 				 */
1671 				m0->m_pkthdr.rcv_tstmp =
1672 				    last_flit_to_ns(sc, d->rsp.u.last_flit);
1673 #ifdef notyet
1674 				m0->m_flags |= M_TSTMP;
1675 #endif
1676 			}
1677 
1678 			/* fall through */
1679 
1680 		case X_RSPD_TYPE_CPL:
1681 			KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1682 			    ("%s: bad opcode %02x.", __func__, d->rss.opcode));
1683 			t4_cpl_handler[d->rss.opcode](iq, &d->rss, m0);
1684 			break;
1685 
1686 		case X_RSPD_TYPE_INTR:
1687 
1688 			/*
1689 			 * There are 1K interrupt-capable queues (qids 0
1690 			 * through 1023).  A response type indicating a
1691 			 * forwarded interrupt with a qid >= 1K is an
1692 			 * iWARP async notification.  That is the only
1693 			 * acceptable indirect interrupt on this queue.
1694 			 */
1695 			if (__predict_false(lq < 1024)) {
1696 				panic("%s: indirect interrupt on iq_fl %p "
1697 				    "with qid %u", __func__, iq, lq);
1698 			}
1699 
1700 			t4_an_handler(iq, &d->rsp);
1701 			break;
1702 
1703 		default:
1704 			KASSERT(0, ("%s: illegal response type %d on iq %p",
1705 			    __func__, rsp_type, iq));
1706 			log(LOG_ERR, "%s: illegal response type %d on iq %p",
1707 			    device_get_nameunit(sc->dev), rsp_type, iq);
1708 			break;
1709 		}
1710 
1711 		d++;
1712 		if (__predict_false(++iq->cidx == iq->sidx)) {
1713 			iq->cidx = 0;
1714 			iq->gen ^= F_RSPD_GEN;
1715 			d = &iq->desc[0];
1716 		}
1717 		if (__predict_false(++ndescs == limit)) {
1718 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1719 			    V_INGRESSQID(iq->cntxt_id) |
1720 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1721 			ndescs = 0;
1722 
1723 #if defined(INET) || defined(INET6)
1724 			if (iq->flags & IQ_LRO_ENABLED &&
1725 			    !sort_before_lro(lro) &&
1726 			    sc->lro_timeout != 0) {
1727 				tcp_lro_flush_inactive(lro, &lro_timeout);
1728 			}
1729 #endif
1730 			if (budget) {
1731 				FL_LOCK(fl);
1732 				refill_fl(sc, fl, 32);
1733 				FL_UNLOCK(fl);
1734 
1735 				return (EINPROGRESS);
1736 			}
1737 		}
1738 		if (refill) {
1739 			FL_LOCK(fl);
1740 			refill_fl(sc, fl, 32);
1741 			FL_UNLOCK(fl);
1742 			fl_hw_cidx = fl->hw_cidx;
1743 		}
1744 	}
1745 out:
1746 #if defined(INET) || defined(INET6)
1747 	if (iq->flags & IQ_LRO_ENABLED) {
1748 		if (ndescs > 0 && lro->lro_mbuf_count > 8) {
1749 			MPASS(sort_before_lro(lro));
1750 			/* hold back one credit and don't flush LRO state */
1751 			iq->flags |= IQ_ADJ_CREDIT;
1752 			ndescs--;
1753 		} else {
1754 			tcp_lro_flush_all(lro);
1755 		}
1756 	}
1757 #endif
1758 
1759 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1760 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1761 
1762 	FL_LOCK(fl);
1763 	starved = refill_fl(sc, fl, 64);
1764 	FL_UNLOCK(fl);
1765 	if (__predict_false(starved != 0))
1766 		add_fl_to_sfl(sc, fl);
1767 
1768 	return (0);
1769 }
1770 
1771 static inline int
1772 cl_has_metadata(struct sge_fl *fl, struct cluster_layout *cll)
1773 {
1774 	int rc = fl->flags & FL_BUF_PACKING || cll->region1 > 0;
1775 
1776 	if (rc)
1777 		MPASS(cll->region3 >= CL_METADATA_SIZE);
1778 
1779 	return (rc);
1780 }
1781 
1782 static inline struct cluster_metadata *
1783 cl_metadata(struct adapter *sc, struct sge_fl *fl, struct cluster_layout *cll,
1784     caddr_t cl)
1785 {
1786 
1787 	if (cl_has_metadata(fl, cll)) {
1788 		struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1789 
1790 		return ((struct cluster_metadata *)(cl + swz->size) - 1);
1791 	}
1792 	return (NULL);
1793 }
1794 
1795 static void
1796 rxb_free(struct mbuf *m)
1797 {
1798 	uma_zone_t zone = m->m_ext.ext_arg1;
1799 	void *cl = m->m_ext.ext_arg2;
1800 
1801 	uma_zfree(zone, cl);
1802 	counter_u64_add(extfree_rels, 1);
1803 }
1804 
1805 /*
1806  * The mbuf returned by this function could be allocated from zone_mbuf or
1807  * constructed in spare room in the cluster.
1808  *
1809  * The mbuf carries the payload in one of these ways
1810  * a) frame inside the mbuf (mbuf from zone_mbuf)
1811  * b) m_cljset (for clusters without metadata) zone_mbuf
1812  * c) m_extaddref (cluster with metadata) inline mbuf
1813  * d) m_extaddref (cluster with metadata) zone_mbuf
1814  */
1815 static struct mbuf *
1816 get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1817     int remaining)
1818 {
1819 	struct mbuf *m;
1820 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1821 	struct cluster_layout *cll = &sd->cll;
1822 	struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1823 	struct hw_buf_info *hwb = &sc->sge.hw_buf_info[cll->hwidx];
1824 	struct cluster_metadata *clm = cl_metadata(sc, fl, cll, sd->cl);
1825 	int len, blen;
1826 	caddr_t payload;
1827 
1828 	blen = hwb->size - fl->rx_offset;	/* max possible in this buf */
1829 	len = min(remaining, blen);
1830 	payload = sd->cl + cll->region1 + fl->rx_offset;
1831 	if (fl->flags & FL_BUF_PACKING) {
1832 		const u_int l = fr_offset + len;
1833 		const u_int pad = roundup2(l, fl->buf_boundary) - l;
1834 
1835 		if (fl->rx_offset + len + pad < hwb->size)
1836 			blen = len + pad;
1837 		MPASS(fl->rx_offset + blen <= hwb->size);
1838 	} else {
1839 		MPASS(fl->rx_offset == 0);	/* not packing */
1840 	}
1841 
1842 
1843 	if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) {
1844 
1845 		/*
1846 		 * Copy payload into a freshly allocated mbuf.
1847 		 */
1848 
1849 		m = fr_offset == 0 ?
1850 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1851 		if (m == NULL)
1852 			return (NULL);
1853 		fl->mbuf_allocated++;
1854 
1855 		/* copy data to mbuf */
1856 		bcopy(payload, mtod(m, caddr_t), len);
1857 
1858 	} else if (sd->nmbuf * MSIZE < cll->region1) {
1859 
1860 		/*
1861 		 * There's spare room in the cluster for an mbuf.  Create one
1862 		 * and associate it with the payload that's in the cluster.
1863 		 */
1864 
1865 		MPASS(clm != NULL);
1866 		m = (struct mbuf *)(sd->cl + sd->nmbuf * MSIZE);
1867 		/* No bzero required */
1868 		if (m_init(m, M_NOWAIT, MT_DATA,
1869 		    fr_offset == 0 ? M_PKTHDR | M_NOFREE : M_NOFREE))
1870 			return (NULL);
1871 		fl->mbuf_inlined++;
1872 		m_extaddref(m, payload, blen, &clm->refcount, rxb_free,
1873 		    swz->zone, sd->cl);
1874 		if (sd->nmbuf++ == 0)
1875 			counter_u64_add(extfree_refs, 1);
1876 
1877 	} else {
1878 
1879 		/*
1880 		 * Grab an mbuf from zone_mbuf and associate it with the
1881 		 * payload in the cluster.
1882 		 */
1883 
1884 		m = fr_offset == 0 ?
1885 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1886 		if (m == NULL)
1887 			return (NULL);
1888 		fl->mbuf_allocated++;
1889 		if (clm != NULL) {
1890 			m_extaddref(m, payload, blen, &clm->refcount,
1891 			    rxb_free, swz->zone, sd->cl);
1892 			if (sd->nmbuf++ == 0)
1893 				counter_u64_add(extfree_refs, 1);
1894 		} else {
1895 			m_cljset(m, sd->cl, swz->type);
1896 			sd->cl = NULL;	/* consumed, not a recycle candidate */
1897 		}
1898 	}
1899 	if (fr_offset == 0)
1900 		m->m_pkthdr.len = remaining;
1901 	m->m_len = len;
1902 
1903 	if (fl->flags & FL_BUF_PACKING) {
1904 		fl->rx_offset += blen;
1905 		MPASS(fl->rx_offset <= hwb->size);
1906 		if (fl->rx_offset < hwb->size)
1907 			return (m);	/* without advancing the cidx */
1908 	}
1909 
1910 	if (__predict_false(++fl->cidx % 8 == 0)) {
1911 		uint16_t cidx = fl->cidx / 8;
1912 
1913 		if (__predict_false(cidx == fl->sidx))
1914 			fl->cidx = cidx = 0;
1915 		fl->hw_cidx = cidx;
1916 	}
1917 	fl->rx_offset = 0;
1918 
1919 	return (m);
1920 }
1921 
1922 static struct mbuf *
1923 get_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf)
1924 {
1925 	struct mbuf *m0, *m, **pnext;
1926 	u_int remaining;
1927 	const u_int total = G_RSPD_LEN(len_newbuf);
1928 
1929 	if (__predict_false(fl->flags & FL_BUF_RESUME)) {
1930 		M_ASSERTPKTHDR(fl->m0);
1931 		MPASS(fl->m0->m_pkthdr.len == total);
1932 		MPASS(fl->remaining < total);
1933 
1934 		m0 = fl->m0;
1935 		pnext = fl->pnext;
1936 		remaining = fl->remaining;
1937 		fl->flags &= ~FL_BUF_RESUME;
1938 		goto get_segment;
1939 	}
1940 
1941 	if (fl->rx_offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
1942 		fl->rx_offset = 0;
1943 		if (__predict_false(++fl->cidx % 8 == 0)) {
1944 			uint16_t cidx = fl->cidx / 8;
1945 
1946 			if (__predict_false(cidx == fl->sidx))
1947 				fl->cidx = cidx = 0;
1948 			fl->hw_cidx = cidx;
1949 		}
1950 	}
1951 
1952 	/*
1953 	 * Payload starts at rx_offset in the current hw buffer.  Its length is
1954 	 * 'len' and it may span multiple hw buffers.
1955 	 */
1956 
1957 	m0 = get_scatter_segment(sc, fl, 0, total);
1958 	if (m0 == NULL)
1959 		return (NULL);
1960 	remaining = total - m0->m_len;
1961 	pnext = &m0->m_next;
1962 	while (remaining > 0) {
1963 get_segment:
1964 		MPASS(fl->rx_offset == 0);
1965 		m = get_scatter_segment(sc, fl, total - remaining, remaining);
1966 		if (__predict_false(m == NULL)) {
1967 			fl->m0 = m0;
1968 			fl->pnext = pnext;
1969 			fl->remaining = remaining;
1970 			fl->flags |= FL_BUF_RESUME;
1971 			return (NULL);
1972 		}
1973 		*pnext = m;
1974 		pnext = &m->m_next;
1975 		remaining -= m->m_len;
1976 	}
1977 	*pnext = NULL;
1978 
1979 	M_ASSERTPKTHDR(m0);
1980 	return (m0);
1981 }
1982 
1983 static int
1984 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
1985 {
1986 	struct sge_rxq *rxq = iq_to_rxq(iq);
1987 	struct ifnet *ifp = rxq->ifp;
1988 	struct adapter *sc = iq->adapter;
1989 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
1990 #if defined(INET) || defined(INET6)
1991 	struct lro_ctrl *lro = &rxq->lro;
1992 #endif
1993 	static const int sw_hashtype[4][2] = {
1994 		{M_HASHTYPE_NONE, M_HASHTYPE_NONE},
1995 		{M_HASHTYPE_RSS_IPV4, M_HASHTYPE_RSS_IPV6},
1996 		{M_HASHTYPE_RSS_TCP_IPV4, M_HASHTYPE_RSS_TCP_IPV6},
1997 		{M_HASHTYPE_RSS_UDP_IPV4, M_HASHTYPE_RSS_UDP_IPV6},
1998 	};
1999 
2000 	KASSERT(m0 != NULL, ("%s: no payload with opcode %02x", __func__,
2001 	    rss->opcode));
2002 
2003 	m0->m_pkthdr.len -= sc->params.sge.fl_pktshift;
2004 	m0->m_len -= sc->params.sge.fl_pktshift;
2005 	m0->m_data += sc->params.sge.fl_pktshift;
2006 
2007 	m0->m_pkthdr.rcvif = ifp;
2008 	M_HASHTYPE_SET(m0, sw_hashtype[rss->hash_type][rss->ipv6]);
2009 	m0->m_pkthdr.flowid = be32toh(rss->hash_val);
2010 
2011 	if (cpl->csum_calc && !(cpl->err_vec & sc->params.tp.err_vec_mask)) {
2012 		if (ifp->if_capenable & IFCAP_RXCSUM &&
2013 		    cpl->l2info & htobe32(F_RXF_IP)) {
2014 			m0->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
2015 			    CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
2016 			rxq->rxcsum++;
2017 		} else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
2018 		    cpl->l2info & htobe32(F_RXF_IP6)) {
2019 			m0->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
2020 			    CSUM_PSEUDO_HDR);
2021 			rxq->rxcsum++;
2022 		}
2023 
2024 		if (__predict_false(cpl->ip_frag))
2025 			m0->m_pkthdr.csum_data = be16toh(cpl->csum);
2026 		else
2027 			m0->m_pkthdr.csum_data = 0xffff;
2028 	}
2029 
2030 	if (cpl->vlan_ex) {
2031 		m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
2032 		m0->m_flags |= M_VLANTAG;
2033 		rxq->vlan_extraction++;
2034 	}
2035 
2036 #if defined(INET) || defined(INET6)
2037 	if (iq->flags & IQ_LRO_ENABLED) {
2038 		if (sort_before_lro(lro)) {
2039 			tcp_lro_queue_mbuf(lro, m0);
2040 			return (0); /* queued for sort, then LRO */
2041 		}
2042 		if (tcp_lro_rx(lro, m0, 0) == 0)
2043 			return (0); /* queued for LRO */
2044 	}
2045 #endif
2046 	ifp->if_input(ifp, m0);
2047 
2048 	return (0);
2049 }
2050 
2051 /*
2052  * Must drain the wrq or make sure that someone else will.
2053  */
2054 static void
2055 wrq_tx_drain(void *arg, int n)
2056 {
2057 	struct sge_wrq *wrq = arg;
2058 	struct sge_eq *eq = &wrq->eq;
2059 
2060 	EQ_LOCK(eq);
2061 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2062 		drain_wrq_wr_list(wrq->adapter, wrq);
2063 	EQ_UNLOCK(eq);
2064 }
2065 
2066 static void
2067 drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq)
2068 {
2069 	struct sge_eq *eq = &wrq->eq;
2070 	u_int available, dbdiff;	/* # of hardware descriptors */
2071 	u_int n;
2072 	struct wrqe *wr;
2073 	struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
2074 
2075 	EQ_LOCK_ASSERT_OWNED(eq);
2076 	MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
2077 	wr = STAILQ_FIRST(&wrq->wr_list);
2078 	MPASS(wr != NULL);	/* Must be called with something useful to do */
2079 	MPASS(eq->pidx == eq->dbidx);
2080 	dbdiff = 0;
2081 
2082 	do {
2083 		eq->cidx = read_hw_cidx(eq);
2084 		if (eq->pidx == eq->cidx)
2085 			available = eq->sidx - 1;
2086 		else
2087 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2088 
2089 		MPASS(wr->wrq == wrq);
2090 		n = howmany(wr->wr_len, EQ_ESIZE);
2091 		if (available < n)
2092 			break;
2093 
2094 		dst = (void *)&eq->desc[eq->pidx];
2095 		if (__predict_true(eq->sidx - eq->pidx > n)) {
2096 			/* Won't wrap, won't end exactly at the status page. */
2097 			bcopy(&wr->wr[0], dst, wr->wr_len);
2098 			eq->pidx += n;
2099 		} else {
2100 			int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE;
2101 
2102 			bcopy(&wr->wr[0], dst, first_portion);
2103 			if (wr->wr_len > first_portion) {
2104 				bcopy(&wr->wr[first_portion], &eq->desc[0],
2105 				    wr->wr_len - first_portion);
2106 			}
2107 			eq->pidx = n - (eq->sidx - eq->pidx);
2108 		}
2109 		wrq->tx_wrs_copied++;
2110 
2111 		if (available < eq->sidx / 4 &&
2112 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2113 				/*
2114 				 * XXX: This is not 100% reliable with some
2115 				 * types of WRs.  But this is a very unusual
2116 				 * situation for an ofld/ctrl queue anyway.
2117 				 */
2118 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2119 			    F_FW_WR_EQUEQ);
2120 		}
2121 
2122 		dbdiff += n;
2123 		if (dbdiff >= 16) {
2124 			ring_eq_db(sc, eq, dbdiff);
2125 			dbdiff = 0;
2126 		}
2127 
2128 		STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
2129 		free_wrqe(wr);
2130 		MPASS(wrq->nwr_pending > 0);
2131 		wrq->nwr_pending--;
2132 		MPASS(wrq->ndesc_needed >= n);
2133 		wrq->ndesc_needed -= n;
2134 	} while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL);
2135 
2136 	if (dbdiff)
2137 		ring_eq_db(sc, eq, dbdiff);
2138 }
2139 
2140 /*
2141  * Doesn't fail.  Holds on to work requests it can't send right away.
2142  */
2143 void
2144 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
2145 {
2146 #ifdef INVARIANTS
2147 	struct sge_eq *eq = &wrq->eq;
2148 #endif
2149 
2150 	EQ_LOCK_ASSERT_OWNED(eq);
2151 	MPASS(wr != NULL);
2152 	MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN);
2153 	MPASS((wr->wr_len & 0x7) == 0);
2154 
2155 	STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
2156 	wrq->nwr_pending++;
2157 	wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE);
2158 
2159 	if (!TAILQ_EMPTY(&wrq->incomplete_wrs))
2160 		return;	/* commit_wrq_wr will drain wr_list as well. */
2161 
2162 	drain_wrq_wr_list(sc, wrq);
2163 
2164 	/* Doorbell must have caught up to the pidx. */
2165 	MPASS(eq->pidx == eq->dbidx);
2166 }
2167 
2168 void
2169 t4_update_fl_bufsize(struct ifnet *ifp)
2170 {
2171 	struct vi_info *vi = ifp->if_softc;
2172 	struct adapter *sc = vi->pi->adapter;
2173 	struct sge_rxq *rxq;
2174 #ifdef TCP_OFFLOAD
2175 	struct sge_ofld_rxq *ofld_rxq;
2176 #endif
2177 	struct sge_fl *fl;
2178 	int i, maxp, mtu = ifp->if_mtu;
2179 
2180 	maxp = mtu_to_max_payload(sc, mtu, 0);
2181 	for_each_rxq(vi, i, rxq) {
2182 		fl = &rxq->fl;
2183 
2184 		FL_LOCK(fl);
2185 		find_best_refill_source(sc, fl, maxp);
2186 		FL_UNLOCK(fl);
2187 	}
2188 #ifdef TCP_OFFLOAD
2189 	maxp = mtu_to_max_payload(sc, mtu, 1);
2190 	for_each_ofld_rxq(vi, i, ofld_rxq) {
2191 		fl = &ofld_rxq->fl;
2192 
2193 		FL_LOCK(fl);
2194 		find_best_refill_source(sc, fl, maxp);
2195 		FL_UNLOCK(fl);
2196 	}
2197 #endif
2198 }
2199 
2200 static inline int
2201 mbuf_nsegs(struct mbuf *m)
2202 {
2203 
2204 	M_ASSERTPKTHDR(m);
2205 	KASSERT(m->m_pkthdr.l5hlen > 0,
2206 	    ("%s: mbuf %p missing information on # of segments.", __func__, m));
2207 
2208 	return (m->m_pkthdr.l5hlen);
2209 }
2210 
2211 static inline void
2212 set_mbuf_nsegs(struct mbuf *m, uint8_t nsegs)
2213 {
2214 
2215 	M_ASSERTPKTHDR(m);
2216 	m->m_pkthdr.l5hlen = nsegs;
2217 }
2218 
2219 static inline int
2220 mbuf_cflags(struct mbuf *m)
2221 {
2222 
2223 	M_ASSERTPKTHDR(m);
2224 	return (m->m_pkthdr.PH_loc.eight[4]);
2225 }
2226 
2227 static inline void
2228 set_mbuf_cflags(struct mbuf *m, uint8_t flags)
2229 {
2230 
2231 	M_ASSERTPKTHDR(m);
2232 	m->m_pkthdr.PH_loc.eight[4] = flags;
2233 }
2234 
2235 static inline int
2236 mbuf_len16(struct mbuf *m)
2237 {
2238 	int n;
2239 
2240 	M_ASSERTPKTHDR(m);
2241 	n = m->m_pkthdr.PH_loc.eight[0];
2242 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2243 
2244 	return (n);
2245 }
2246 
2247 static inline void
2248 set_mbuf_len16(struct mbuf *m, uint8_t len16)
2249 {
2250 
2251 	M_ASSERTPKTHDR(m);
2252 	m->m_pkthdr.PH_loc.eight[0] = len16;
2253 }
2254 
2255 #ifdef RATELIMIT
2256 static inline int
2257 mbuf_eo_nsegs(struct mbuf *m)
2258 {
2259 
2260 	M_ASSERTPKTHDR(m);
2261 	return (m->m_pkthdr.PH_loc.eight[1]);
2262 }
2263 
2264 static inline void
2265 set_mbuf_eo_nsegs(struct mbuf *m, uint8_t nsegs)
2266 {
2267 
2268 	M_ASSERTPKTHDR(m);
2269 	m->m_pkthdr.PH_loc.eight[1] = nsegs;
2270 }
2271 
2272 static inline int
2273 mbuf_eo_len16(struct mbuf *m)
2274 {
2275 	int n;
2276 
2277 	M_ASSERTPKTHDR(m);
2278 	n = m->m_pkthdr.PH_loc.eight[2];
2279 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2280 
2281 	return (n);
2282 }
2283 
2284 static inline void
2285 set_mbuf_eo_len16(struct mbuf *m, uint8_t len16)
2286 {
2287 
2288 	M_ASSERTPKTHDR(m);
2289 	m->m_pkthdr.PH_loc.eight[2] = len16;
2290 }
2291 
2292 static inline int
2293 mbuf_eo_tsclk_tsoff(struct mbuf *m)
2294 {
2295 
2296 	M_ASSERTPKTHDR(m);
2297 	return (m->m_pkthdr.PH_loc.eight[3]);
2298 }
2299 
2300 static inline void
2301 set_mbuf_eo_tsclk_tsoff(struct mbuf *m, uint8_t tsclk_tsoff)
2302 {
2303 
2304 	M_ASSERTPKTHDR(m);
2305 	m->m_pkthdr.PH_loc.eight[3] = tsclk_tsoff;
2306 }
2307 
2308 static inline int
2309 needs_eo(struct mbuf *m)
2310 {
2311 
2312 	return (m->m_pkthdr.snd_tag != NULL);
2313 }
2314 #endif
2315 
2316 /*
2317  * Try to allocate an mbuf to contain a raw work request.  To make it
2318  * easy to construct the work request, don't allocate a chain but a
2319  * single mbuf.
2320  */
2321 struct mbuf *
2322 alloc_wr_mbuf(int len, int how)
2323 {
2324 	struct mbuf *m;
2325 
2326 	if (len <= MHLEN)
2327 		m = m_gethdr(how, MT_DATA);
2328 	else if (len <= MCLBYTES)
2329 		m = m_getcl(how, MT_DATA, M_PKTHDR);
2330 	else
2331 		m = NULL;
2332 	if (m == NULL)
2333 		return (NULL);
2334 	m->m_pkthdr.len = len;
2335 	m->m_len = len;
2336 	set_mbuf_cflags(m, MC_RAW_WR);
2337 	set_mbuf_len16(m, howmany(len, 16));
2338 	return (m);
2339 }
2340 
2341 static inline int
2342 needs_tso(struct mbuf *m)
2343 {
2344 
2345 	M_ASSERTPKTHDR(m);
2346 
2347 	return (m->m_pkthdr.csum_flags & CSUM_TSO);
2348 }
2349 
2350 static inline int
2351 needs_l3_csum(struct mbuf *m)
2352 {
2353 
2354 	M_ASSERTPKTHDR(m);
2355 
2356 	return (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO));
2357 }
2358 
2359 static inline int
2360 needs_l4_csum(struct mbuf *m)
2361 {
2362 
2363 	M_ASSERTPKTHDR(m);
2364 
2365 	return (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
2366 	    CSUM_TCP_IPV6 | CSUM_TSO));
2367 }
2368 
2369 static inline int
2370 needs_tcp_csum(struct mbuf *m)
2371 {
2372 
2373 	M_ASSERTPKTHDR(m);
2374 	return (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_TCP_IPV6 | CSUM_TSO));
2375 }
2376 
2377 #ifdef RATELIMIT
2378 static inline int
2379 needs_udp_csum(struct mbuf *m)
2380 {
2381 
2382 	M_ASSERTPKTHDR(m);
2383 	return (m->m_pkthdr.csum_flags & (CSUM_UDP | CSUM_UDP_IPV6));
2384 }
2385 #endif
2386 
2387 static inline int
2388 needs_vlan_insertion(struct mbuf *m)
2389 {
2390 
2391 	M_ASSERTPKTHDR(m);
2392 
2393 	return (m->m_flags & M_VLANTAG);
2394 }
2395 
2396 static void *
2397 m_advance(struct mbuf **pm, int *poffset, int len)
2398 {
2399 	struct mbuf *m = *pm;
2400 	int offset = *poffset;
2401 	uintptr_t p = 0;
2402 
2403 	MPASS(len > 0);
2404 
2405 	for (;;) {
2406 		if (offset + len < m->m_len) {
2407 			offset += len;
2408 			p = mtod(m, uintptr_t) + offset;
2409 			break;
2410 		}
2411 		len -= m->m_len - offset;
2412 		m = m->m_next;
2413 		offset = 0;
2414 		MPASS(m != NULL);
2415 	}
2416 	*poffset = offset;
2417 	*pm = m;
2418 	return ((void *)p);
2419 }
2420 
2421 /*
2422  * Can deal with empty mbufs in the chain that have m_len = 0, but the chain
2423  * must have at least one mbuf that's not empty.  It is possible for this
2424  * routine to return 0 if skip accounts for all the contents of the mbuf chain.
2425  */
2426 static inline int
2427 count_mbuf_nsegs(struct mbuf *m, int skip)
2428 {
2429 	vm_paddr_t lastb, next;
2430 	vm_offset_t va;
2431 	int len, nsegs;
2432 
2433 	M_ASSERTPKTHDR(m);
2434 	MPASS(m->m_pkthdr.len > 0);
2435 	MPASS(m->m_pkthdr.len >= skip);
2436 
2437 	nsegs = 0;
2438 	lastb = 0;
2439 	for (; m; m = m->m_next) {
2440 
2441 		len = m->m_len;
2442 		if (__predict_false(len == 0))
2443 			continue;
2444 		if (skip >= len) {
2445 			skip -= len;
2446 			continue;
2447 		}
2448 		va = mtod(m, vm_offset_t) + skip;
2449 		len -= skip;
2450 		skip = 0;
2451 		next = pmap_kextract(va);
2452 		nsegs += sglist_count((void *)(uintptr_t)va, len);
2453 		if (lastb + 1 == next)
2454 			nsegs--;
2455 		lastb = pmap_kextract(va + len - 1);
2456 	}
2457 
2458 	return (nsegs);
2459 }
2460 
2461 /*
2462  * Analyze the mbuf to determine its tx needs.  The mbuf passed in may change:
2463  * a) caller can assume it's been freed if this function returns with an error.
2464  * b) it may get defragged up if the gather list is too long for the hardware.
2465  */
2466 int
2467 parse_pkt(struct adapter *sc, struct mbuf **mp)
2468 {
2469 	struct mbuf *m0 = *mp, *m;
2470 	int rc, nsegs, defragged = 0, offset;
2471 	struct ether_header *eh;
2472 	void *l3hdr;
2473 #if defined(INET) || defined(INET6)
2474 	struct tcphdr *tcp;
2475 #endif
2476 	uint16_t eh_type;
2477 
2478 	M_ASSERTPKTHDR(m0);
2479 	if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) {
2480 		rc = EINVAL;
2481 fail:
2482 		m_freem(m0);
2483 		*mp = NULL;
2484 		return (rc);
2485 	}
2486 restart:
2487 	/*
2488 	 * First count the number of gather list segments in the payload.
2489 	 * Defrag the mbuf if nsegs exceeds the hardware limit.
2490 	 */
2491 	M_ASSERTPKTHDR(m0);
2492 	MPASS(m0->m_pkthdr.len > 0);
2493 	nsegs = count_mbuf_nsegs(m0, 0);
2494 	if (nsegs > (needs_tso(m0) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS)) {
2495 		if (defragged++ > 0 || (m = m_defrag(m0, M_NOWAIT)) == NULL) {
2496 			rc = EFBIG;
2497 			goto fail;
2498 		}
2499 		*mp = m0 = m;	/* update caller's copy after defrag */
2500 		goto restart;
2501 	}
2502 
2503 	if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN)) {
2504 		m0 = m_pullup(m0, m0->m_pkthdr.len);
2505 		if (m0 == NULL) {
2506 			/* Should have left well enough alone. */
2507 			rc = EFBIG;
2508 			goto fail;
2509 		}
2510 		*mp = m0;	/* update caller's copy after pullup */
2511 		goto restart;
2512 	}
2513 	set_mbuf_nsegs(m0, nsegs);
2514 	set_mbuf_cflags(m0, 0);
2515 	if (sc->flags & IS_VF)
2516 		set_mbuf_len16(m0, txpkt_vm_len16(nsegs, needs_tso(m0)));
2517 	else
2518 		set_mbuf_len16(m0, txpkt_len16(nsegs, needs_tso(m0)));
2519 
2520 #ifdef RATELIMIT
2521 	/*
2522 	 * Ethofld is limited to TCP and UDP for now, and only when L4 hw
2523 	 * checksumming is enabled.  needs_l4_csum happens to check for all the
2524 	 * right things.
2525 	 */
2526 	if (__predict_false(needs_eo(m0) && !needs_l4_csum(m0)))
2527 		m0->m_pkthdr.snd_tag = NULL;
2528 #endif
2529 
2530 	if (!needs_tso(m0) &&
2531 #ifdef RATELIMIT
2532 	    !needs_eo(m0) &&
2533 #endif
2534 	    !(sc->flags & IS_VF && (needs_l3_csum(m0) || needs_l4_csum(m0))))
2535 		return (0);
2536 
2537 	m = m0;
2538 	eh = mtod(m, struct ether_header *);
2539 	eh_type = ntohs(eh->ether_type);
2540 	if (eh_type == ETHERTYPE_VLAN) {
2541 		struct ether_vlan_header *evh = (void *)eh;
2542 
2543 		eh_type = ntohs(evh->evl_proto);
2544 		m0->m_pkthdr.l2hlen = sizeof(*evh);
2545 	} else
2546 		m0->m_pkthdr.l2hlen = sizeof(*eh);
2547 
2548 	offset = 0;
2549 	l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2550 
2551 	switch (eh_type) {
2552 #ifdef INET6
2553 	case ETHERTYPE_IPV6:
2554 	{
2555 		struct ip6_hdr *ip6 = l3hdr;
2556 
2557 		MPASS(!needs_tso(m0) || ip6->ip6_nxt == IPPROTO_TCP);
2558 
2559 		m0->m_pkthdr.l3hlen = sizeof(*ip6);
2560 		break;
2561 	}
2562 #endif
2563 #ifdef INET
2564 	case ETHERTYPE_IP:
2565 	{
2566 		struct ip *ip = l3hdr;
2567 
2568 		m0->m_pkthdr.l3hlen = ip->ip_hl * 4;
2569 		break;
2570 	}
2571 #endif
2572 	default:
2573 		panic("%s: ethertype 0x%04x unknown.  if_cxgbe must be compiled"
2574 		    " with the same INET/INET6 options as the kernel.",
2575 		    __func__, eh_type);
2576 	}
2577 
2578 #if defined(INET) || defined(INET6)
2579 	if (needs_tcp_csum(m0)) {
2580 		tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen);
2581 		m0->m_pkthdr.l4hlen = tcp->th_off * 4;
2582 #ifdef RATELIMIT
2583 		if (tsclk >= 0 && *(uint32_t *)(tcp + 1) == ntohl(0x0101080a)) {
2584 			set_mbuf_eo_tsclk_tsoff(m0,
2585 			    V_FW_ETH_TX_EO_WR_TSCLK(tsclk) |
2586 			    V_FW_ETH_TX_EO_WR_TSOFF(sizeof(*tcp) / 2 + 1));
2587 		} else
2588 			set_mbuf_eo_tsclk_tsoff(m0, 0);
2589 	} else if (needs_udp_csum(m)) {
2590 		m0->m_pkthdr.l4hlen = sizeof(struct udphdr);
2591 #endif
2592 	}
2593 #ifdef RATELIMIT
2594 	if (needs_eo(m0)) {
2595 		u_int immhdrs;
2596 
2597 		/* EO WRs have the headers in the WR and not the GL. */
2598 		immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen +
2599 		    m0->m_pkthdr.l4hlen;
2600 		nsegs = count_mbuf_nsegs(m0, immhdrs);
2601 		set_mbuf_eo_nsegs(m0, nsegs);
2602 		set_mbuf_eo_len16(m0,
2603 		    txpkt_eo_len16(nsegs, immhdrs, needs_tso(m0)));
2604 	}
2605 #endif
2606 #endif
2607 	MPASS(m0 == *mp);
2608 	return (0);
2609 }
2610 
2611 void *
2612 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie)
2613 {
2614 	struct sge_eq *eq = &wrq->eq;
2615 	struct adapter *sc = wrq->adapter;
2616 	int ndesc, available;
2617 	struct wrqe *wr;
2618 	void *w;
2619 
2620 	MPASS(len16 > 0);
2621 	ndesc = howmany(len16, EQ_ESIZE / 16);
2622 	MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC);
2623 
2624 	EQ_LOCK(eq);
2625 
2626 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2627 		drain_wrq_wr_list(sc, wrq);
2628 
2629 	if (!STAILQ_EMPTY(&wrq->wr_list)) {
2630 slowpath:
2631 		EQ_UNLOCK(eq);
2632 		wr = alloc_wrqe(len16 * 16, wrq);
2633 		if (__predict_false(wr == NULL))
2634 			return (NULL);
2635 		cookie->pidx = -1;
2636 		cookie->ndesc = ndesc;
2637 		return (&wr->wr);
2638 	}
2639 
2640 	eq->cidx = read_hw_cidx(eq);
2641 	if (eq->pidx == eq->cidx)
2642 		available = eq->sidx - 1;
2643 	else
2644 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2645 	if (available < ndesc)
2646 		goto slowpath;
2647 
2648 	cookie->pidx = eq->pidx;
2649 	cookie->ndesc = ndesc;
2650 	TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link);
2651 
2652 	w = &eq->desc[eq->pidx];
2653 	IDXINCR(eq->pidx, ndesc, eq->sidx);
2654 	if (__predict_false(cookie->pidx + ndesc > eq->sidx)) {
2655 		w = &wrq->ss[0];
2656 		wrq->ss_pidx = cookie->pidx;
2657 		wrq->ss_len = len16 * 16;
2658 	}
2659 
2660 	EQ_UNLOCK(eq);
2661 
2662 	return (w);
2663 }
2664 
2665 void
2666 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie)
2667 {
2668 	struct sge_eq *eq = &wrq->eq;
2669 	struct adapter *sc = wrq->adapter;
2670 	int ndesc, pidx;
2671 	struct wrq_cookie *prev, *next;
2672 
2673 	if (cookie->pidx == -1) {
2674 		struct wrqe *wr = __containerof(w, struct wrqe, wr);
2675 
2676 		t4_wrq_tx(sc, wr);
2677 		return;
2678 	}
2679 
2680 	if (__predict_false(w == &wrq->ss[0])) {
2681 		int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE;
2682 
2683 		MPASS(wrq->ss_len > n);	/* WR had better wrap around. */
2684 		bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n);
2685 		bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n);
2686 		wrq->tx_wrs_ss++;
2687 	} else
2688 		wrq->tx_wrs_direct++;
2689 
2690 	EQ_LOCK(eq);
2691 	ndesc = cookie->ndesc;	/* Can be more than SGE_MAX_WR_NDESC here. */
2692 	pidx = cookie->pidx;
2693 	MPASS(pidx >= 0 && pidx < eq->sidx);
2694 	prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link);
2695 	next = TAILQ_NEXT(cookie, link);
2696 	if (prev == NULL) {
2697 		MPASS(pidx == eq->dbidx);
2698 		if (next == NULL || ndesc >= 16) {
2699 			int available;
2700 			struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
2701 
2702 			/*
2703 			 * Note that the WR via which we'll request tx updates
2704 			 * is at pidx and not eq->pidx, which has moved on
2705 			 * already.
2706 			 */
2707 			dst = (void *)&eq->desc[pidx];
2708 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2709 			if (available < eq->sidx / 4 &&
2710 			    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2711 				/*
2712 				 * XXX: This is not 100% reliable with some
2713 				 * types of WRs.  But this is a very unusual
2714 				 * situation for an ofld/ctrl queue anyway.
2715 				 */
2716 				dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2717 				    F_FW_WR_EQUEQ);
2718 			}
2719 
2720 			ring_eq_db(wrq->adapter, eq, ndesc);
2721 		} else {
2722 			MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc);
2723 			next->pidx = pidx;
2724 			next->ndesc += ndesc;
2725 		}
2726 	} else {
2727 		MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc);
2728 		prev->ndesc += ndesc;
2729 	}
2730 	TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link);
2731 
2732 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2733 		drain_wrq_wr_list(sc, wrq);
2734 
2735 #ifdef INVARIANTS
2736 	if (TAILQ_EMPTY(&wrq->incomplete_wrs)) {
2737 		/* Doorbell must have caught up to the pidx. */
2738 		MPASS(wrq->eq.pidx == wrq->eq.dbidx);
2739 	}
2740 #endif
2741 	EQ_UNLOCK(eq);
2742 }
2743 
2744 static u_int
2745 can_resume_eth_tx(struct mp_ring *r)
2746 {
2747 	struct sge_eq *eq = r->cookie;
2748 
2749 	return (total_available_tx_desc(eq) > eq->sidx / 8);
2750 }
2751 
2752 static inline int
2753 cannot_use_txpkts(struct mbuf *m)
2754 {
2755 	/* maybe put a GL limit too, to avoid silliness? */
2756 
2757 	return (needs_tso(m) || (mbuf_cflags(m) & MC_RAW_WR) != 0);
2758 }
2759 
2760 static inline int
2761 discard_tx(struct sge_eq *eq)
2762 {
2763 
2764 	return ((eq->flags & (EQ_ENABLED | EQ_QFLUSH)) != EQ_ENABLED);
2765 }
2766 
2767 static inline int
2768 wr_can_update_eq(struct fw_eth_tx_pkts_wr *wr)
2769 {
2770 
2771 	switch (G_FW_WR_OP(be32toh(wr->op_pkd))) {
2772 	case FW_ULPTX_WR:
2773 	case FW_ETH_TX_PKT_WR:
2774 	case FW_ETH_TX_PKTS_WR:
2775 	case FW_ETH_TX_PKT_VM_WR:
2776 		return (1);
2777 	default:
2778 		return (0);
2779 	}
2780 }
2781 
2782 /*
2783  * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to
2784  * be consumed.  Return the actual number consumed.  0 indicates a stall.
2785  */
2786 static u_int
2787 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx)
2788 {
2789 	struct sge_txq *txq = r->cookie;
2790 	struct sge_eq *eq = &txq->eq;
2791 	struct ifnet *ifp = txq->ifp;
2792 	struct vi_info *vi = ifp->if_softc;
2793 	struct port_info *pi = vi->pi;
2794 	struct adapter *sc = pi->adapter;
2795 	u_int total, remaining;		/* # of packets */
2796 	u_int available, dbdiff;	/* # of hardware descriptors */
2797 	u_int n, next_cidx;
2798 	struct mbuf *m0, *tail;
2799 	struct txpkts txp;
2800 	struct fw_eth_tx_pkts_wr *wr;	/* any fw WR struct will do */
2801 
2802 	remaining = IDXDIFF(pidx, cidx, r->size);
2803 	MPASS(remaining > 0);	/* Must not be called without work to do. */
2804 	total = 0;
2805 
2806 	TXQ_LOCK(txq);
2807 	if (__predict_false(discard_tx(eq))) {
2808 		while (cidx != pidx) {
2809 			m0 = r->items[cidx];
2810 			m_freem(m0);
2811 			if (++cidx == r->size)
2812 				cidx = 0;
2813 		}
2814 		reclaim_tx_descs(txq, 2048);
2815 		total = remaining;
2816 		goto done;
2817 	}
2818 
2819 	/* How many hardware descriptors do we have readily available. */
2820 	if (eq->pidx == eq->cidx)
2821 		available = eq->sidx - 1;
2822 	else
2823 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2824 	dbdiff = IDXDIFF(eq->pidx, eq->dbidx, eq->sidx);
2825 
2826 	while (remaining > 0) {
2827 
2828 		m0 = r->items[cidx];
2829 		M_ASSERTPKTHDR(m0);
2830 		MPASS(m0->m_nextpkt == NULL);
2831 
2832 		if (available < SGE_MAX_WR_NDESC) {
2833 			available += reclaim_tx_descs(txq, 64);
2834 			if (available < howmany(mbuf_len16(m0), EQ_ESIZE / 16))
2835 				break;	/* out of descriptors */
2836 		}
2837 
2838 		next_cidx = cidx + 1;
2839 		if (__predict_false(next_cidx == r->size))
2840 			next_cidx = 0;
2841 
2842 		wr = (void *)&eq->desc[eq->pidx];
2843 		if (sc->flags & IS_VF) {
2844 			total++;
2845 			remaining--;
2846 			ETHER_BPF_MTAP(ifp, m0);
2847 			n = write_txpkt_vm_wr(sc, txq, (void *)wr, m0,
2848 			    available);
2849 		} else if (remaining > 1 &&
2850 		    try_txpkts(m0, r->items[next_cidx], &txp, available) == 0) {
2851 
2852 			/* pkts at cidx, next_cidx should both be in txp. */
2853 			MPASS(txp.npkt == 2);
2854 			tail = r->items[next_cidx];
2855 			MPASS(tail->m_nextpkt == NULL);
2856 			ETHER_BPF_MTAP(ifp, m0);
2857 			ETHER_BPF_MTAP(ifp, tail);
2858 			m0->m_nextpkt = tail;
2859 
2860 			if (__predict_false(++next_cidx == r->size))
2861 				next_cidx = 0;
2862 
2863 			while (next_cidx != pidx) {
2864 				if (add_to_txpkts(r->items[next_cidx], &txp,
2865 				    available) != 0)
2866 					break;
2867 				tail->m_nextpkt = r->items[next_cidx];
2868 				tail = tail->m_nextpkt;
2869 				ETHER_BPF_MTAP(ifp, tail);
2870 				if (__predict_false(++next_cidx == r->size))
2871 					next_cidx = 0;
2872 			}
2873 
2874 			n = write_txpkts_wr(txq, wr, m0, &txp, available);
2875 			total += txp.npkt;
2876 			remaining -= txp.npkt;
2877 		} else if (mbuf_cflags(m0) & MC_RAW_WR) {
2878 			total++;
2879 			remaining--;
2880 			n = write_raw_wr(txq, (void *)wr, m0, available);
2881 		} else {
2882 			total++;
2883 			remaining--;
2884 			ETHER_BPF_MTAP(ifp, m0);
2885 			n = write_txpkt_wr(txq, (void *)wr, m0, available);
2886 		}
2887 		MPASS(n >= 1 && n <= available && n <= SGE_MAX_WR_NDESC);
2888 
2889 		available -= n;
2890 		dbdiff += n;
2891 		IDXINCR(eq->pidx, n, eq->sidx);
2892 
2893 		if (wr_can_update_eq(wr)) {
2894 			if (total_available_tx_desc(eq) < eq->sidx / 4 &&
2895 			    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2896 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2897 				    F_FW_WR_EQUEQ);
2898 				eq->equeqidx = eq->pidx;
2899 			} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >=
2900 			    32) {
2901 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
2902 				eq->equeqidx = eq->pidx;
2903 			}
2904 		}
2905 
2906 		if (dbdiff >= 16 && remaining >= 4) {
2907 			ring_eq_db(sc, eq, dbdiff);
2908 			available += reclaim_tx_descs(txq, 4 * dbdiff);
2909 			dbdiff = 0;
2910 		}
2911 
2912 		cidx = next_cidx;
2913 	}
2914 	if (dbdiff != 0) {
2915 		ring_eq_db(sc, eq, dbdiff);
2916 		reclaim_tx_descs(txq, 32);
2917 	}
2918 done:
2919 	TXQ_UNLOCK(txq);
2920 
2921 	return (total);
2922 }
2923 
2924 static inline void
2925 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
2926     int qsize)
2927 {
2928 
2929 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
2930 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
2931 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
2932 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
2933 
2934 	iq->flags = 0;
2935 	iq->adapter = sc;
2936 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
2937 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
2938 	if (pktc_idx >= 0) {
2939 		iq->intr_params |= F_QINTR_CNT_EN;
2940 		iq->intr_pktc_idx = pktc_idx;
2941 	}
2942 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
2943 	iq->sidx = iq->qsize - sc->params.sge.spg_len / IQ_ESIZE;
2944 }
2945 
2946 static inline void
2947 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name)
2948 {
2949 
2950 	fl->qsize = qsize;
2951 	fl->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2952 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
2953 	if (sc->flags & BUF_PACKING_OK &&
2954 	    ((!is_t4(sc) && buffer_packing) ||	/* T5+: enabled unless 0 */
2955 	    (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */
2956 		fl->flags |= FL_BUF_PACKING;
2957 	find_best_refill_source(sc, fl, maxp);
2958 	find_safe_refill_source(sc, fl);
2959 }
2960 
2961 static inline void
2962 init_eq(struct adapter *sc, struct sge_eq *eq, int eqtype, int qsize,
2963     uint8_t tx_chan, uint16_t iqid, char *name)
2964 {
2965 	KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
2966 
2967 	eq->flags = eqtype & EQ_TYPEMASK;
2968 	eq->tx_chan = tx_chan;
2969 	eq->iqid = iqid;
2970 	eq->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2971 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
2972 }
2973 
2974 static int
2975 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
2976     bus_dmamap_t *map, bus_addr_t *pa, void **va)
2977 {
2978 	int rc;
2979 
2980 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
2981 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
2982 	if (rc != 0) {
2983 		device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
2984 		goto done;
2985 	}
2986 
2987 	rc = bus_dmamem_alloc(*tag, va,
2988 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
2989 	if (rc != 0) {
2990 		device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
2991 		goto done;
2992 	}
2993 
2994 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
2995 	if (rc != 0) {
2996 		device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
2997 		goto done;
2998 	}
2999 done:
3000 	if (rc)
3001 		free_ring(sc, *tag, *map, *pa, *va);
3002 
3003 	return (rc);
3004 }
3005 
3006 static int
3007 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
3008     bus_addr_t pa, void *va)
3009 {
3010 	if (pa)
3011 		bus_dmamap_unload(tag, map);
3012 	if (va)
3013 		bus_dmamem_free(tag, va, map);
3014 	if (tag)
3015 		bus_dma_tag_destroy(tag);
3016 
3017 	return (0);
3018 }
3019 
3020 /*
3021  * Allocates the ring for an ingress queue and an optional freelist.  If the
3022  * freelist is specified it will be allocated and then associated with the
3023  * ingress queue.
3024  *
3025  * Returns errno on failure.  Resources allocated up to that point may still be
3026  * allocated.  Caller is responsible for cleanup in case this function fails.
3027  *
3028  * If the ingress queue will take interrupts directly then the intr_idx
3029  * specifies the vector, starting from 0.  -1 means the interrupts for this
3030  * queue should be forwarded to the fwq.
3031  */
3032 static int
3033 alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl,
3034     int intr_idx, int cong)
3035 {
3036 	int rc, i, cntxt_id;
3037 	size_t len;
3038 	struct fw_iq_cmd c;
3039 	struct port_info *pi = vi->pi;
3040 	struct adapter *sc = iq->adapter;
3041 	struct sge_params *sp = &sc->params.sge;
3042 	__be32 v = 0;
3043 
3044 	len = iq->qsize * IQ_ESIZE;
3045 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
3046 	    (void **)&iq->desc);
3047 	if (rc != 0)
3048 		return (rc);
3049 
3050 	bzero(&c, sizeof(c));
3051 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
3052 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
3053 	    V_FW_IQ_CMD_VFN(0));
3054 
3055 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
3056 	    FW_LEN16(c));
3057 
3058 	/* Special handling for firmware event queue */
3059 	if (iq == &sc->sge.fwq)
3060 		v |= F_FW_IQ_CMD_IQASYNCH;
3061 
3062 	if (intr_idx < 0) {
3063 		/* Forwarded interrupts, all headed to fwq */
3064 		v |= F_FW_IQ_CMD_IQANDST;
3065 		v |= V_FW_IQ_CMD_IQANDSTINDEX(sc->sge.fwq.cntxt_id);
3066 	} else {
3067 		KASSERT(intr_idx < sc->intr_count,
3068 		    ("%s: invalid direct intr_idx %d", __func__, intr_idx));
3069 		v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
3070 	}
3071 
3072 	c.type_to_iqandstindex = htobe32(v |
3073 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
3074 	    V_FW_IQ_CMD_VIID(vi->viid) |
3075 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
3076 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
3077 	    F_FW_IQ_CMD_IQGTSMODE |
3078 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
3079 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
3080 	c.iqsize = htobe16(iq->qsize);
3081 	c.iqaddr = htobe64(iq->ba);
3082 	if (cong >= 0)
3083 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
3084 
3085 	if (fl) {
3086 		mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
3087 
3088 		len = fl->qsize * EQ_ESIZE;
3089 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
3090 		    &fl->ba, (void **)&fl->desc);
3091 		if (rc)
3092 			return (rc);
3093 
3094 		/* Allocate space for one software descriptor per buffer. */
3095 		rc = alloc_fl_sdesc(fl);
3096 		if (rc != 0) {
3097 			device_printf(sc->dev,
3098 			    "failed to setup fl software descriptors: %d\n",
3099 			    rc);
3100 			return (rc);
3101 		}
3102 
3103 		if (fl->flags & FL_BUF_PACKING) {
3104 			fl->lowat = roundup2(sp->fl_starve_threshold2, 8);
3105 			fl->buf_boundary = sp->pack_boundary;
3106 		} else {
3107 			fl->lowat = roundup2(sp->fl_starve_threshold, 8);
3108 			fl->buf_boundary = 16;
3109 		}
3110 		if (fl_pad && fl->buf_boundary < sp->pad_boundary)
3111 			fl->buf_boundary = sp->pad_boundary;
3112 
3113 		c.iqns_to_fl0congen |=
3114 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
3115 			F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
3116 			(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
3117 			(fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
3118 			    0));
3119 		if (cong >= 0) {
3120 			c.iqns_to_fl0congen |=
3121 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
3122 				    F_FW_IQ_CMD_FL0CONGCIF |
3123 				    F_FW_IQ_CMD_FL0CONGEN);
3124 		}
3125 		c.fl0dcaen_to_fl0cidxfthresh =
3126 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
3127 			X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B) |
3128 			V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
3129 			X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
3130 		c.fl0size = htobe16(fl->qsize);
3131 		c.fl0addr = htobe64(fl->ba);
3132 	}
3133 
3134 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3135 	if (rc != 0) {
3136 		device_printf(sc->dev,
3137 		    "failed to create ingress queue: %d\n", rc);
3138 		return (rc);
3139 	}
3140 
3141 	iq->cidx = 0;
3142 	iq->gen = F_RSPD_GEN;
3143 	iq->intr_next = iq->intr_params;
3144 	iq->cntxt_id = be16toh(c.iqid);
3145 	iq->abs_id = be16toh(c.physiqid);
3146 	iq->flags |= IQ_ALLOCATED;
3147 
3148 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
3149 	if (cntxt_id >= sc->sge.niq) {
3150 		panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
3151 		    cntxt_id, sc->sge.niq - 1);
3152 	}
3153 	sc->sge.iqmap[cntxt_id] = iq;
3154 
3155 	if (fl) {
3156 		u_int qid;
3157 
3158 		iq->flags |= IQ_HAS_FL;
3159 		fl->cntxt_id = be16toh(c.fl0id);
3160 		fl->pidx = fl->cidx = 0;
3161 
3162 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
3163 		if (cntxt_id >= sc->sge.neq) {
3164 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
3165 			    __func__, cntxt_id, sc->sge.neq - 1);
3166 		}
3167 		sc->sge.eqmap[cntxt_id] = (void *)fl;
3168 
3169 		qid = fl->cntxt_id;
3170 		if (isset(&sc->doorbells, DOORBELL_UDB)) {
3171 			uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3172 			uint32_t mask = (1 << s_qpp) - 1;
3173 			volatile uint8_t *udb;
3174 
3175 			udb = sc->udbs_base + UDBS_DB_OFFSET;
3176 			udb += (qid >> s_qpp) << PAGE_SHIFT;
3177 			qid &= mask;
3178 			if (qid < PAGE_SIZE / UDBS_SEG_SIZE) {
3179 				udb += qid << UDBS_SEG_SHIFT;
3180 				qid = 0;
3181 			}
3182 			fl->udb = (volatile void *)udb;
3183 		}
3184 		fl->dbval = V_QID(qid) | sc->chip_params->sge_fl_db;
3185 
3186 		FL_LOCK(fl);
3187 		/* Enough to make sure the SGE doesn't think it's starved */
3188 		refill_fl(sc, fl, fl->lowat);
3189 		FL_UNLOCK(fl);
3190 	}
3191 
3192 	if (chip_id(sc) >= CHELSIO_T5 && !(sc->flags & IS_VF) && cong >= 0) {
3193 		uint32_t param, val;
3194 
3195 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
3196 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
3197 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
3198 		if (cong == 0)
3199 			val = 1 << 19;
3200 		else {
3201 			val = 2 << 19;
3202 			for (i = 0; i < 4; i++) {
3203 				if (cong & (1 << i))
3204 					val |= 1 << (i << 2);
3205 			}
3206 		}
3207 
3208 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
3209 		if (rc != 0) {
3210 			/* report error but carry on */
3211 			device_printf(sc->dev,
3212 			    "failed to set congestion manager context for "
3213 			    "ingress queue %d: %d\n", iq->cntxt_id, rc);
3214 		}
3215 	}
3216 
3217 	/* Enable IQ interrupts */
3218 	atomic_store_rel_int(&iq->state, IQS_IDLE);
3219 	t4_write_reg(sc, sc->sge_gts_reg, V_SEINTARM(iq->intr_params) |
3220 	    V_INGRESSQID(iq->cntxt_id));
3221 
3222 	return (0);
3223 }
3224 
3225 static int
3226 free_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl)
3227 {
3228 	int rc;
3229 	struct adapter *sc = iq->adapter;
3230 	device_t dev;
3231 
3232 	if (sc == NULL)
3233 		return (0);	/* nothing to do */
3234 
3235 	dev = vi ? vi->dev : sc->dev;
3236 
3237 	if (iq->flags & IQ_ALLOCATED) {
3238 		rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
3239 		    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
3240 		    fl ? fl->cntxt_id : 0xffff, 0xffff);
3241 		if (rc != 0) {
3242 			device_printf(dev,
3243 			    "failed to free queue %p: %d\n", iq, rc);
3244 			return (rc);
3245 		}
3246 		iq->flags &= ~IQ_ALLOCATED;
3247 	}
3248 
3249 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
3250 
3251 	bzero(iq, sizeof(*iq));
3252 
3253 	if (fl) {
3254 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
3255 		    fl->desc);
3256 
3257 		if (fl->sdesc)
3258 			free_fl_sdesc(sc, fl);
3259 
3260 		if (mtx_initialized(&fl->fl_lock))
3261 			mtx_destroy(&fl->fl_lock);
3262 
3263 		bzero(fl, sizeof(*fl));
3264 	}
3265 
3266 	return (0);
3267 }
3268 
3269 static void
3270 add_iq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
3271     struct sge_iq *iq)
3272 {
3273 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3274 
3275 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD, &iq->ba,
3276 	    "bus address of descriptor ring");
3277 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3278 	    iq->qsize * IQ_ESIZE, "descriptor ring size in bytes");
3279 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3280 	    CTLTYPE_INT | CTLFLAG_RD, &iq->abs_id, 0, sysctl_uint16, "I",
3281 	    "absolute id of the queue");
3282 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3283 	    CTLTYPE_INT | CTLFLAG_RD, &iq->cntxt_id, 0, sysctl_uint16, "I",
3284 	    "SGE context id of the queue");
3285 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3286 	    CTLTYPE_INT | CTLFLAG_RD, &iq->cidx, 0, sysctl_uint16, "I",
3287 	    "consumer index");
3288 }
3289 
3290 static void
3291 add_fl_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
3292     struct sysctl_oid *oid, struct sge_fl *fl)
3293 {
3294 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3295 
3296 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3297 	    "freelist");
3298 	children = SYSCTL_CHILDREN(oid);
3299 
3300 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3301 	    &fl->ba, "bus address of descriptor ring");
3302 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3303 	    fl->sidx * EQ_ESIZE + sc->params.sge.spg_len,
3304 	    "desc ring size in bytes");
3305 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3306 	    CTLTYPE_INT | CTLFLAG_RD, &fl->cntxt_id, 0, sysctl_uint16, "I",
3307 	    "SGE context id of the freelist");
3308 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL,
3309 	    fl_pad ? 1 : 0, "padding enabled");
3310 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL,
3311 	    fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled");
3312 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
3313 	    0, "consumer index");
3314 	if (fl->flags & FL_BUF_PACKING) {
3315 		SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
3316 		    CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
3317 	}
3318 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
3319 	    0, "producer index");
3320 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_allocated",
3321 	    CTLFLAG_RD, &fl->mbuf_allocated, "# of mbuf allocated");
3322 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_inlined",
3323 	    CTLFLAG_RD, &fl->mbuf_inlined, "# of mbuf inlined in clusters");
3324 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
3325 	    CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
3326 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
3327 	    CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
3328 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
3329 	    CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
3330 }
3331 
3332 static int
3333 alloc_fwq(struct adapter *sc)
3334 {
3335 	int rc, intr_idx;
3336 	struct sge_iq *fwq = &sc->sge.fwq;
3337 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
3338 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3339 
3340 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE);
3341 	if (sc->flags & IS_VF)
3342 		intr_idx = 0;
3343 	else
3344 		intr_idx = sc->intr_count > 1 ? 1 : 0;
3345 	rc = alloc_iq_fl(&sc->port[0]->vi[0], fwq, NULL, intr_idx, -1);
3346 	if (rc != 0) {
3347 		device_printf(sc->dev,
3348 		    "failed to create firmware event queue: %d\n", rc);
3349 		return (rc);
3350 	}
3351 
3352 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD,
3353 	    NULL, "firmware event queue");
3354 	add_iq_sysctls(&sc->ctx, oid, fwq);
3355 
3356 	return (0);
3357 }
3358 
3359 static int
3360 free_fwq(struct adapter *sc)
3361 {
3362 	return free_iq_fl(NULL, &sc->sge.fwq, NULL);
3363 }
3364 
3365 static int
3366 alloc_ctrlq(struct adapter *sc, struct sge_wrq *ctrlq, int idx,
3367     struct sysctl_oid *oid)
3368 {
3369 	int rc;
3370 	char name[16];
3371 	struct sysctl_oid_list *children;
3372 
3373 	snprintf(name, sizeof(name), "%s ctrlq%d", device_get_nameunit(sc->dev),
3374 	    idx);
3375 	init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[idx]->tx_chan,
3376 	    sc->sge.fwq.cntxt_id, name);
3377 
3378 	children = SYSCTL_CHILDREN(oid);
3379 	snprintf(name, sizeof(name), "%d", idx);
3380 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3381 	    NULL, "ctrl queue");
3382 	rc = alloc_wrq(sc, NULL, ctrlq, oid);
3383 
3384 	return (rc);
3385 }
3386 
3387 int
3388 tnl_cong(struct port_info *pi, int drop)
3389 {
3390 
3391 	if (drop == -1)
3392 		return (-1);
3393 	else if (drop == 1)
3394 		return (0);
3395 	else
3396 		return (pi->rx_e_chan_map);
3397 }
3398 
3399 static int
3400 alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int intr_idx, int idx,
3401     struct sysctl_oid *oid)
3402 {
3403 	int rc;
3404 	struct adapter *sc = vi->pi->adapter;
3405 	struct sysctl_oid_list *children;
3406 	char name[16];
3407 
3408 	rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, intr_idx,
3409 	    tnl_cong(vi->pi, cong_drop));
3410 	if (rc != 0)
3411 		return (rc);
3412 
3413 	if (idx == 0)
3414 		sc->sge.iq_base = rxq->iq.abs_id - rxq->iq.cntxt_id;
3415 	else
3416 		KASSERT(rxq->iq.cntxt_id + sc->sge.iq_base == rxq->iq.abs_id,
3417 		    ("iq_base mismatch"));
3418 	KASSERT(sc->sge.iq_base == 0 || sc->flags & IS_VF,
3419 	    ("PF with non-zero iq_base"));
3420 
3421 	/*
3422 	 * The freelist is just barely above the starvation threshold right now,
3423 	 * fill it up a bit more.
3424 	 */
3425 	FL_LOCK(&rxq->fl);
3426 	refill_fl(sc, &rxq->fl, 128);
3427 	FL_UNLOCK(&rxq->fl);
3428 
3429 #if defined(INET) || defined(INET6)
3430 	rc = tcp_lro_init_args(&rxq->lro, vi->ifp, lro_entries, lro_mbufs);
3431 	if (rc != 0)
3432 		return (rc);
3433 	MPASS(rxq->lro.ifp == vi->ifp);	/* also indicates LRO init'ed */
3434 
3435 	if (vi->ifp->if_capenable & IFCAP_LRO)
3436 		rxq->iq.flags |= IQ_LRO_ENABLED;
3437 #endif
3438 	rxq->ifp = vi->ifp;
3439 
3440 	children = SYSCTL_CHILDREN(oid);
3441 
3442 	snprintf(name, sizeof(name), "%d", idx);
3443 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3444 	    NULL, "rx queue");
3445 	children = SYSCTL_CHILDREN(oid);
3446 
3447 	add_iq_sysctls(&vi->ctx, oid, &rxq->iq);
3448 #if defined(INET) || defined(INET6)
3449 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
3450 	    &rxq->lro.lro_queued, 0, NULL);
3451 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
3452 	    &rxq->lro.lro_flushed, 0, NULL);
3453 #endif
3454 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
3455 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
3456 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_extraction",
3457 	    CTLFLAG_RD, &rxq->vlan_extraction,
3458 	    "# of times hardware extracted 802.1Q tag");
3459 
3460 	add_fl_sysctls(sc, &vi->ctx, oid, &rxq->fl);
3461 
3462 	return (rc);
3463 }
3464 
3465 static int
3466 free_rxq(struct vi_info *vi, struct sge_rxq *rxq)
3467 {
3468 	int rc;
3469 
3470 #if defined(INET) || defined(INET6)
3471 	if (rxq->lro.ifp) {
3472 		tcp_lro_free(&rxq->lro);
3473 		rxq->lro.ifp = NULL;
3474 	}
3475 #endif
3476 
3477 	rc = free_iq_fl(vi, &rxq->iq, &rxq->fl);
3478 	if (rc == 0)
3479 		bzero(rxq, sizeof(*rxq));
3480 
3481 	return (rc);
3482 }
3483 
3484 #ifdef TCP_OFFLOAD
3485 static int
3486 alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq,
3487     int intr_idx, int idx, struct sysctl_oid *oid)
3488 {
3489 	struct port_info *pi = vi->pi;
3490 	int rc;
3491 	struct sysctl_oid_list *children;
3492 	char name[16];
3493 
3494 	rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx, 0);
3495 	if (rc != 0)
3496 		return (rc);
3497 
3498 	children = SYSCTL_CHILDREN(oid);
3499 
3500 	snprintf(name, sizeof(name), "%d", idx);
3501 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3502 	    NULL, "rx queue");
3503 	add_iq_sysctls(&vi->ctx, oid, &ofld_rxq->iq);
3504 	add_fl_sysctls(pi->adapter, &vi->ctx, oid, &ofld_rxq->fl);
3505 
3506 	return (rc);
3507 }
3508 
3509 static int
3510 free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq)
3511 {
3512 	int rc;
3513 
3514 	rc = free_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl);
3515 	if (rc == 0)
3516 		bzero(ofld_rxq, sizeof(*ofld_rxq));
3517 
3518 	return (rc);
3519 }
3520 #endif
3521 
3522 #ifdef DEV_NETMAP
3523 static int
3524 alloc_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int intr_idx,
3525     int idx, struct sysctl_oid *oid)
3526 {
3527 	int rc;
3528 	struct sysctl_oid_list *children;
3529 	struct sysctl_ctx_list *ctx;
3530 	char name[16];
3531 	size_t len;
3532 	struct adapter *sc = vi->pi->adapter;
3533 	struct netmap_adapter *na = NA(vi->ifp);
3534 
3535 	MPASS(na != NULL);
3536 
3537 	len = vi->qsize_rxq * IQ_ESIZE;
3538 	rc = alloc_ring(sc, len, &nm_rxq->iq_desc_tag, &nm_rxq->iq_desc_map,
3539 	    &nm_rxq->iq_ba, (void **)&nm_rxq->iq_desc);
3540 	if (rc != 0)
3541 		return (rc);
3542 
3543 	len = na->num_rx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3544 	rc = alloc_ring(sc, len, &nm_rxq->fl_desc_tag, &nm_rxq->fl_desc_map,
3545 	    &nm_rxq->fl_ba, (void **)&nm_rxq->fl_desc);
3546 	if (rc != 0)
3547 		return (rc);
3548 
3549 	nm_rxq->vi = vi;
3550 	nm_rxq->nid = idx;
3551 	nm_rxq->iq_cidx = 0;
3552 	nm_rxq->iq_sidx = vi->qsize_rxq - sc->params.sge.spg_len / IQ_ESIZE;
3553 	nm_rxq->iq_gen = F_RSPD_GEN;
3554 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
3555 	nm_rxq->fl_sidx = na->num_rx_desc;
3556 	nm_rxq->intr_idx = intr_idx;
3557 	nm_rxq->iq_cntxt_id = INVALID_NM_RXQ_CNTXT_ID;
3558 
3559 	ctx = &vi->ctx;
3560 	children = SYSCTL_CHILDREN(oid);
3561 
3562 	snprintf(name, sizeof(name), "%d", idx);
3563 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL,
3564 	    "rx queue");
3565 	children = SYSCTL_CHILDREN(oid);
3566 
3567 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3568 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_abs_id, 0, sysctl_uint16,
3569 	    "I", "absolute id of the queue");
3570 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3571 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cntxt_id, 0, sysctl_uint16,
3572 	    "I", "SGE context id of the queue");
3573 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3574 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cidx, 0, sysctl_uint16, "I",
3575 	    "consumer index");
3576 
3577 	children = SYSCTL_CHILDREN(oid);
3578 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3579 	    "freelist");
3580 	children = SYSCTL_CHILDREN(oid);
3581 
3582 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3583 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->fl_cntxt_id, 0, sysctl_uint16,
3584 	    "I", "SGE context id of the freelist");
3585 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
3586 	    &nm_rxq->fl_cidx, 0, "consumer index");
3587 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
3588 	    &nm_rxq->fl_pidx, 0, "producer index");
3589 
3590 	return (rc);
3591 }
3592 
3593 
3594 static int
3595 free_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
3596 {
3597 	struct adapter *sc = vi->pi->adapter;
3598 
3599 	if (vi->flags & VI_INIT_DONE)
3600 		MPASS(nm_rxq->iq_cntxt_id == INVALID_NM_RXQ_CNTXT_ID);
3601 	else
3602 		MPASS(nm_rxq->iq_cntxt_id == 0);
3603 
3604 	free_ring(sc, nm_rxq->iq_desc_tag, nm_rxq->iq_desc_map, nm_rxq->iq_ba,
3605 	    nm_rxq->iq_desc);
3606 	free_ring(sc, nm_rxq->fl_desc_tag, nm_rxq->fl_desc_map, nm_rxq->fl_ba,
3607 	    nm_rxq->fl_desc);
3608 
3609 	return (0);
3610 }
3611 
3612 static int
3613 alloc_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq, int iqidx, int idx,
3614     struct sysctl_oid *oid)
3615 {
3616 	int rc;
3617 	size_t len;
3618 	struct port_info *pi = vi->pi;
3619 	struct adapter *sc = pi->adapter;
3620 	struct netmap_adapter *na = NA(vi->ifp);
3621 	char name[16];
3622 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3623 
3624 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3625 	rc = alloc_ring(sc, len, &nm_txq->desc_tag, &nm_txq->desc_map,
3626 	    &nm_txq->ba, (void **)&nm_txq->desc);
3627 	if (rc)
3628 		return (rc);
3629 
3630 	nm_txq->pidx = nm_txq->cidx = 0;
3631 	nm_txq->sidx = na->num_tx_desc;
3632 	nm_txq->nid = idx;
3633 	nm_txq->iqidx = iqidx;
3634 	nm_txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3635 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(G_FW_VIID_PFN(vi->viid)) |
3636 	    V_TXPKT_VF(G_FW_VIID_VIN(vi->viid)) |
3637 	    V_TXPKT_VF_VLD(G_FW_VIID_VIVLD(vi->viid)));
3638 	nm_txq->cntxt_id = INVALID_NM_TXQ_CNTXT_ID;
3639 
3640 	snprintf(name, sizeof(name), "%d", idx);
3641 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3642 	    NULL, "netmap tx queue");
3643 	children = SYSCTL_CHILDREN(oid);
3644 
3645 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3646 	    &nm_txq->cntxt_id, 0, "SGE context id of the queue");
3647 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3648 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->cidx, 0, sysctl_uint16, "I",
3649 	    "consumer index");
3650 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3651 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->pidx, 0, sysctl_uint16, "I",
3652 	    "producer index");
3653 
3654 	return (rc);
3655 }
3656 
3657 static int
3658 free_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
3659 {
3660 	struct adapter *sc = vi->pi->adapter;
3661 
3662 	if (vi->flags & VI_INIT_DONE)
3663 		MPASS(nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID);
3664 	else
3665 		MPASS(nm_txq->cntxt_id == 0);
3666 
3667 	free_ring(sc, nm_txq->desc_tag, nm_txq->desc_map, nm_txq->ba,
3668 	    nm_txq->desc);
3669 
3670 	return (0);
3671 }
3672 #endif
3673 
3674 /*
3675  * Returns a reasonable automatic cidx flush threshold for a given queue size.
3676  */
3677 static u_int
3678 qsize_to_fthresh(int qsize)
3679 {
3680 	u_int fthresh;
3681 
3682 	while (!powerof2(qsize))
3683 		qsize++;
3684 	fthresh = ilog2(qsize);
3685 	if (fthresh > X_CIDXFLUSHTHRESH_128)
3686 		fthresh = X_CIDXFLUSHTHRESH_128;
3687 
3688 	return (fthresh);
3689 }
3690 
3691 static int
3692 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
3693 {
3694 	int rc, cntxt_id;
3695 	struct fw_eq_ctrl_cmd c;
3696 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3697 
3698 	bzero(&c, sizeof(c));
3699 
3700 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
3701 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
3702 	    V_FW_EQ_CTRL_CMD_VFN(0));
3703 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
3704 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
3705 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid));
3706 	c.physeqid_pkd = htobe32(0);
3707 	c.fetchszm_to_iqid =
3708 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
3709 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
3710 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
3711 	c.dcaen_to_eqsize =
3712 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3713 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3714 		V_FW_EQ_CTRL_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
3715 		V_FW_EQ_CTRL_CMD_EQSIZE(qsize));
3716 	c.eqaddr = htobe64(eq->ba);
3717 
3718 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3719 	if (rc != 0) {
3720 		device_printf(sc->dev,
3721 		    "failed to create control queue %d: %d\n", eq->tx_chan, rc);
3722 		return (rc);
3723 	}
3724 	eq->flags |= EQ_ALLOCATED;
3725 
3726 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
3727 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3728 	if (cntxt_id >= sc->sge.neq)
3729 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3730 		cntxt_id, sc->sge.neq - 1);
3731 	sc->sge.eqmap[cntxt_id] = eq;
3732 
3733 	return (rc);
3734 }
3735 
3736 static int
3737 eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3738 {
3739 	int rc, cntxt_id;
3740 	struct fw_eq_eth_cmd c;
3741 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3742 
3743 	bzero(&c, sizeof(c));
3744 
3745 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
3746 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
3747 	    V_FW_EQ_ETH_CMD_VFN(0));
3748 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
3749 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
3750 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
3751 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
3752 	c.fetchszm_to_iqid =
3753 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3754 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
3755 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
3756 	c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3757 	    V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3758 	    V_FW_EQ_ETH_CMD_EQSIZE(qsize));
3759 	c.eqaddr = htobe64(eq->ba);
3760 
3761 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3762 	if (rc != 0) {
3763 		device_printf(vi->dev,
3764 		    "failed to create Ethernet egress queue: %d\n", rc);
3765 		return (rc);
3766 	}
3767 	eq->flags |= EQ_ALLOCATED;
3768 
3769 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
3770 	eq->abs_id = G_FW_EQ_ETH_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
3771 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3772 	if (cntxt_id >= sc->sge.neq)
3773 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3774 		cntxt_id, sc->sge.neq - 1);
3775 	sc->sge.eqmap[cntxt_id] = eq;
3776 
3777 	return (rc);
3778 }
3779 
3780 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3781 static int
3782 ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3783 {
3784 	int rc, cntxt_id;
3785 	struct fw_eq_ofld_cmd c;
3786 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3787 
3788 	bzero(&c, sizeof(c));
3789 
3790 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
3791 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
3792 	    V_FW_EQ_OFLD_CMD_VFN(0));
3793 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
3794 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
3795 	c.fetchszm_to_iqid =
3796 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
3797 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
3798 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
3799 	c.dcaen_to_eqsize =
3800 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3801 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3802 		V_FW_EQ_OFLD_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
3803 		V_FW_EQ_OFLD_CMD_EQSIZE(qsize));
3804 	c.eqaddr = htobe64(eq->ba);
3805 
3806 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3807 	if (rc != 0) {
3808 		device_printf(vi->dev,
3809 		    "failed to create egress queue for TCP offload: %d\n", rc);
3810 		return (rc);
3811 	}
3812 	eq->flags |= EQ_ALLOCATED;
3813 
3814 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
3815 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3816 	if (cntxt_id >= sc->sge.neq)
3817 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3818 		cntxt_id, sc->sge.neq - 1);
3819 	sc->sge.eqmap[cntxt_id] = eq;
3820 
3821 	return (rc);
3822 }
3823 #endif
3824 
3825 static int
3826 alloc_eq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3827 {
3828 	int rc, qsize;
3829 	size_t len;
3830 
3831 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
3832 
3833 	qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3834 	len = qsize * EQ_ESIZE;
3835 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
3836 	    &eq->ba, (void **)&eq->desc);
3837 	if (rc)
3838 		return (rc);
3839 
3840 	eq->pidx = eq->cidx = eq->dbidx = 0;
3841 	/* Note that equeqidx is not used with sge_wrq (OFLD/CTRL) queues. */
3842 	eq->equeqidx = 0;
3843 	eq->doorbells = sc->doorbells;
3844 
3845 	switch (eq->flags & EQ_TYPEMASK) {
3846 	case EQ_CTRL:
3847 		rc = ctrl_eq_alloc(sc, eq);
3848 		break;
3849 
3850 	case EQ_ETH:
3851 		rc = eth_eq_alloc(sc, vi, eq);
3852 		break;
3853 
3854 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3855 	case EQ_OFLD:
3856 		rc = ofld_eq_alloc(sc, vi, eq);
3857 		break;
3858 #endif
3859 
3860 	default:
3861 		panic("%s: invalid eq type %d.", __func__,
3862 		    eq->flags & EQ_TYPEMASK);
3863 	}
3864 	if (rc != 0) {
3865 		device_printf(sc->dev,
3866 		    "failed to allocate egress queue(%d): %d\n",
3867 		    eq->flags & EQ_TYPEMASK, rc);
3868 	}
3869 
3870 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
3871 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
3872 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
3873 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3874 		uint32_t mask = (1 << s_qpp) - 1;
3875 		volatile uint8_t *udb;
3876 
3877 		udb = sc->udbs_base + UDBS_DB_OFFSET;
3878 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
3879 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
3880 		if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
3881 	    		clrbit(&eq->doorbells, DOORBELL_WCWR);
3882 		else {
3883 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
3884 			eq->udb_qid = 0;
3885 		}
3886 		eq->udb = (volatile void *)udb;
3887 	}
3888 
3889 	return (rc);
3890 }
3891 
3892 static int
3893 free_eq(struct adapter *sc, struct sge_eq *eq)
3894 {
3895 	int rc;
3896 
3897 	if (eq->flags & EQ_ALLOCATED) {
3898 		switch (eq->flags & EQ_TYPEMASK) {
3899 		case EQ_CTRL:
3900 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
3901 			    eq->cntxt_id);
3902 			break;
3903 
3904 		case EQ_ETH:
3905 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
3906 			    eq->cntxt_id);
3907 			break;
3908 
3909 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3910 		case EQ_OFLD:
3911 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
3912 			    eq->cntxt_id);
3913 			break;
3914 #endif
3915 
3916 		default:
3917 			panic("%s: invalid eq type %d.", __func__,
3918 			    eq->flags & EQ_TYPEMASK);
3919 		}
3920 		if (rc != 0) {
3921 			device_printf(sc->dev,
3922 			    "failed to free egress queue (%d): %d\n",
3923 			    eq->flags & EQ_TYPEMASK, rc);
3924 			return (rc);
3925 		}
3926 		eq->flags &= ~EQ_ALLOCATED;
3927 	}
3928 
3929 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
3930 
3931 	if (mtx_initialized(&eq->eq_lock))
3932 		mtx_destroy(&eq->eq_lock);
3933 
3934 	bzero(eq, sizeof(*eq));
3935 	return (0);
3936 }
3937 
3938 static int
3939 alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq,
3940     struct sysctl_oid *oid)
3941 {
3942 	int rc;
3943 	struct sysctl_ctx_list *ctx = vi ? &vi->ctx : &sc->ctx;
3944 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3945 
3946 	rc = alloc_eq(sc, vi, &wrq->eq);
3947 	if (rc)
3948 		return (rc);
3949 
3950 	wrq->adapter = sc;
3951 	TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq);
3952 	TAILQ_INIT(&wrq->incomplete_wrs);
3953 	STAILQ_INIT(&wrq->wr_list);
3954 	wrq->nwr_pending = 0;
3955 	wrq->ndesc_needed = 0;
3956 
3957 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3958 	    &wrq->eq.ba, "bus address of descriptor ring");
3959 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3960 	    wrq->eq.sidx * EQ_ESIZE + sc->params.sge.spg_len,
3961 	    "desc ring size in bytes");
3962 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3963 	    &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
3964 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3965 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I",
3966 	    "consumer index");
3967 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx",
3968 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I",
3969 	    "producer index");
3970 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
3971 	    wrq->eq.sidx, "status page index");
3972 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD,
3973 	    &wrq->tx_wrs_direct, "# of work requests (direct)");
3974 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD,
3975 	    &wrq->tx_wrs_copied, "# of work requests (copied)");
3976 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_sspace", CTLFLAG_RD,
3977 	    &wrq->tx_wrs_ss, "# of work requests (copied from scratch space)");
3978 
3979 	return (rc);
3980 }
3981 
3982 static int
3983 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
3984 {
3985 	int rc;
3986 
3987 	rc = free_eq(sc, &wrq->eq);
3988 	if (rc)
3989 		return (rc);
3990 
3991 	bzero(wrq, sizeof(*wrq));
3992 	return (0);
3993 }
3994 
3995 static int
3996 alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx,
3997     struct sysctl_oid *oid)
3998 {
3999 	int rc;
4000 	struct port_info *pi = vi->pi;
4001 	struct adapter *sc = pi->adapter;
4002 	struct sge_eq *eq = &txq->eq;
4003 	char name[16];
4004 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
4005 
4006 	rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx, can_resume_eth_tx,
4007 	    M_CXGBE, M_WAITOK);
4008 	if (rc != 0) {
4009 		device_printf(sc->dev, "failed to allocate mp_ring: %d\n", rc);
4010 		return (rc);
4011 	}
4012 
4013 	rc = alloc_eq(sc, vi, eq);
4014 	if (rc != 0) {
4015 		mp_ring_free(txq->r);
4016 		txq->r = NULL;
4017 		return (rc);
4018 	}
4019 
4020 	/* Can't fail after this point. */
4021 
4022 	if (idx == 0)
4023 		sc->sge.eq_base = eq->abs_id - eq->cntxt_id;
4024 	else
4025 		KASSERT(eq->cntxt_id + sc->sge.eq_base == eq->abs_id,
4026 		    ("eq_base mismatch"));
4027 	KASSERT(sc->sge.eq_base == 0 || sc->flags & IS_VF,
4028 	    ("PF with non-zero eq_base"));
4029 
4030 	TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq);
4031 	txq->ifp = vi->ifp;
4032 	txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK);
4033 	if (sc->flags & IS_VF)
4034 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
4035 		    V_TXPKT_INTF(pi->tx_chan));
4036 	else
4037 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
4038 		    V_TXPKT_INTF(pi->tx_chan) |
4039 		    V_TXPKT_PF(G_FW_VIID_PFN(vi->viid)) |
4040 		    V_TXPKT_VF(G_FW_VIID_VIN(vi->viid)) |
4041 		    V_TXPKT_VF_VLD(G_FW_VIID_VIVLD(vi->viid)));
4042 	txq->tc_idx = -1;
4043 	txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE,
4044 	    M_ZERO | M_WAITOK);
4045 
4046 	snprintf(name, sizeof(name), "%d", idx);
4047 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
4048 	    NULL, "tx queue");
4049 	children = SYSCTL_CHILDREN(oid);
4050 
4051 	SYSCTL_ADD_UAUTO(&vi->ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
4052 	    &eq->ba, "bus address of descriptor ring");
4053 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
4054 	    eq->sidx * EQ_ESIZE + sc->params.sge.spg_len,
4055 	    "desc ring size in bytes");
4056 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
4057 	    &eq->abs_id, 0, "absolute id of the queue");
4058 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
4059 	    &eq->cntxt_id, 0, "SGE context id of the queue");
4060 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
4061 	    CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I",
4062 	    "consumer index");
4063 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
4064 	    CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I",
4065 	    "producer index");
4066 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
4067 	    eq->sidx, "status page index");
4068 
4069 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "tc",
4070 	    CTLTYPE_INT | CTLFLAG_RW, vi, idx, sysctl_tc, "I",
4071 	    "traffic class (-1 means none)");
4072 
4073 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
4074 	    &txq->txcsum, "# of times hardware assisted with checksum");
4075 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_insertion",
4076 	    CTLFLAG_RD, &txq->vlan_insertion,
4077 	    "# of times hardware inserted 802.1Q tag");
4078 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
4079 	    &txq->tso_wrs, "# of TSO work requests");
4080 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
4081 	    &txq->imm_wrs, "# of work requests with immediate data");
4082 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
4083 	    &txq->sgl_wrs, "# of work requests with direct SGL");
4084 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
4085 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
4086 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_wrs",
4087 	    CTLFLAG_RD, &txq->txpkts0_wrs,
4088 	    "# of txpkts (type 0) work requests");
4089 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_wrs",
4090 	    CTLFLAG_RD, &txq->txpkts1_wrs,
4091 	    "# of txpkts (type 1) work requests");
4092 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_pkts",
4093 	    CTLFLAG_RD, &txq->txpkts0_pkts,
4094 	    "# of frames tx'd using type0 txpkts work requests");
4095 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_pkts",
4096 	    CTLFLAG_RD, &txq->txpkts1_pkts,
4097 	    "# of frames tx'd using type1 txpkts work requests");
4098 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "raw_wrs", CTLFLAG_RD,
4099 	    &txq->raw_wrs, "# of raw work requests (non-packets)");
4100 
4101 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_enqueues",
4102 	    CTLFLAG_RD, &txq->r->enqueues,
4103 	    "# of enqueues to the mp_ring for this queue");
4104 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_drops",
4105 	    CTLFLAG_RD, &txq->r->drops,
4106 	    "# of drops in the mp_ring for this queue");
4107 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_starts",
4108 	    CTLFLAG_RD, &txq->r->starts,
4109 	    "# of normal consumer starts in the mp_ring for this queue");
4110 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_stalls",
4111 	    CTLFLAG_RD, &txq->r->stalls,
4112 	    "# of consumer stalls in the mp_ring for this queue");
4113 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_restarts",
4114 	    CTLFLAG_RD, &txq->r->restarts,
4115 	    "# of consumer restarts in the mp_ring for this queue");
4116 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_abdications",
4117 	    CTLFLAG_RD, &txq->r->abdications,
4118 	    "# of consumer abdications in the mp_ring for this queue");
4119 
4120 	return (0);
4121 }
4122 
4123 static int
4124 free_txq(struct vi_info *vi, struct sge_txq *txq)
4125 {
4126 	int rc;
4127 	struct adapter *sc = vi->pi->adapter;
4128 	struct sge_eq *eq = &txq->eq;
4129 
4130 	rc = free_eq(sc, eq);
4131 	if (rc)
4132 		return (rc);
4133 
4134 	sglist_free(txq->gl);
4135 	free(txq->sdesc, M_CXGBE);
4136 	mp_ring_free(txq->r);
4137 
4138 	bzero(txq, sizeof(*txq));
4139 	return (0);
4140 }
4141 
4142 static void
4143 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
4144 {
4145 	bus_addr_t *ba = arg;
4146 
4147 	KASSERT(nseg == 1,
4148 	    ("%s meant for single segment mappings only.", __func__));
4149 
4150 	*ba = error ? 0 : segs->ds_addr;
4151 }
4152 
4153 static inline void
4154 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
4155 {
4156 	uint32_t n, v;
4157 
4158 	n = IDXDIFF(fl->pidx / 8, fl->dbidx, fl->sidx);
4159 	MPASS(n > 0);
4160 
4161 	wmb();
4162 	v = fl->dbval | V_PIDX(n);
4163 	if (fl->udb)
4164 		*fl->udb = htole32(v);
4165 	else
4166 		t4_write_reg(sc, sc->sge_kdoorbell_reg, v);
4167 	IDXINCR(fl->dbidx, n, fl->sidx);
4168 }
4169 
4170 /*
4171  * Fills up the freelist by allocating up to 'n' buffers.  Buffers that are
4172  * recycled do not count towards this allocation budget.
4173  *
4174  * Returns non-zero to indicate that this freelist should be added to the list
4175  * of starving freelists.
4176  */
4177 static int
4178 refill_fl(struct adapter *sc, struct sge_fl *fl, int n)
4179 {
4180 	__be64 *d;
4181 	struct fl_sdesc *sd;
4182 	uintptr_t pa;
4183 	caddr_t cl;
4184 	struct cluster_layout *cll;
4185 	struct sw_zone_info *swz;
4186 	struct cluster_metadata *clm;
4187 	uint16_t max_pidx;
4188 	uint16_t hw_cidx = fl->hw_cidx;		/* stable snapshot */
4189 
4190 	FL_LOCK_ASSERT_OWNED(fl);
4191 
4192 	/*
4193 	 * We always stop at the beginning of the hardware descriptor that's just
4194 	 * before the one with the hw cidx.  This is to avoid hw pidx = hw cidx,
4195 	 * which would mean an empty freelist to the chip.
4196 	 */
4197 	max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1;
4198 	if (fl->pidx == max_pidx * 8)
4199 		return (0);
4200 
4201 	d = &fl->desc[fl->pidx];
4202 	sd = &fl->sdesc[fl->pidx];
4203 	cll = &fl->cll_def;	/* default layout */
4204 	swz = &sc->sge.sw_zone_info[cll->zidx];
4205 
4206 	while (n > 0) {
4207 
4208 		if (sd->cl != NULL) {
4209 
4210 			if (sd->nmbuf == 0) {
4211 				/*
4212 				 * Fast recycle without involving any atomics on
4213 				 * the cluster's metadata (if the cluster has
4214 				 * metadata).  This happens when all frames
4215 				 * received in the cluster were small enough to
4216 				 * fit within a single mbuf each.
4217 				 */
4218 				fl->cl_fast_recycled++;
4219 #ifdef INVARIANTS
4220 				clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
4221 				if (clm != NULL)
4222 					MPASS(clm->refcount == 1);
4223 #endif
4224 				goto recycled_fast;
4225 			}
4226 
4227 			/*
4228 			 * Cluster is guaranteed to have metadata.  Clusters
4229 			 * without metadata always take the fast recycle path
4230 			 * when they're recycled.
4231 			 */
4232 			clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
4233 			MPASS(clm != NULL);
4234 
4235 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4236 				fl->cl_recycled++;
4237 				counter_u64_add(extfree_rels, 1);
4238 				goto recycled;
4239 			}
4240 			sd->cl = NULL;	/* gave up my reference */
4241 		}
4242 		MPASS(sd->cl == NULL);
4243 alloc:
4244 		cl = uma_zalloc(swz->zone, M_NOWAIT);
4245 		if (__predict_false(cl == NULL)) {
4246 			if (cll == &fl->cll_alt || fl->cll_alt.zidx == -1 ||
4247 			    fl->cll_def.zidx == fl->cll_alt.zidx)
4248 				break;
4249 
4250 			/* fall back to the safe zone */
4251 			cll = &fl->cll_alt;
4252 			swz = &sc->sge.sw_zone_info[cll->zidx];
4253 			goto alloc;
4254 		}
4255 		fl->cl_allocated++;
4256 		n--;
4257 
4258 		pa = pmap_kextract((vm_offset_t)cl);
4259 		pa += cll->region1;
4260 		sd->cl = cl;
4261 		sd->cll = *cll;
4262 		*d = htobe64(pa | cll->hwidx);
4263 		clm = cl_metadata(sc, fl, cll, cl);
4264 		if (clm != NULL) {
4265 recycled:
4266 #ifdef INVARIANTS
4267 			clm->sd = sd;
4268 #endif
4269 			clm->refcount = 1;
4270 		}
4271 		sd->nmbuf = 0;
4272 recycled_fast:
4273 		d++;
4274 		sd++;
4275 		if (__predict_false(++fl->pidx % 8 == 0)) {
4276 			uint16_t pidx = fl->pidx / 8;
4277 
4278 			if (__predict_false(pidx == fl->sidx)) {
4279 				fl->pidx = 0;
4280 				pidx = 0;
4281 				sd = fl->sdesc;
4282 				d = fl->desc;
4283 			}
4284 			if (pidx == max_pidx)
4285 				break;
4286 
4287 			if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4)
4288 				ring_fl_db(sc, fl);
4289 		}
4290 	}
4291 
4292 	if (fl->pidx / 8 != fl->dbidx)
4293 		ring_fl_db(sc, fl);
4294 
4295 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
4296 }
4297 
4298 /*
4299  * Attempt to refill all starving freelists.
4300  */
4301 static void
4302 refill_sfl(void *arg)
4303 {
4304 	struct adapter *sc = arg;
4305 	struct sge_fl *fl, *fl_temp;
4306 
4307 	mtx_assert(&sc->sfl_lock, MA_OWNED);
4308 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
4309 		FL_LOCK(fl);
4310 		refill_fl(sc, fl, 64);
4311 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
4312 			TAILQ_REMOVE(&sc->sfl, fl, link);
4313 			fl->flags &= ~FL_STARVING;
4314 		}
4315 		FL_UNLOCK(fl);
4316 	}
4317 
4318 	if (!TAILQ_EMPTY(&sc->sfl))
4319 		callout_schedule(&sc->sfl_callout, hz / 5);
4320 }
4321 
4322 static int
4323 alloc_fl_sdesc(struct sge_fl *fl)
4324 {
4325 
4326 	fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc), M_CXGBE,
4327 	    M_ZERO | M_WAITOK);
4328 
4329 	return (0);
4330 }
4331 
4332 static void
4333 free_fl_sdesc(struct adapter *sc, struct sge_fl *fl)
4334 {
4335 	struct fl_sdesc *sd;
4336 	struct cluster_metadata *clm;
4337 	struct cluster_layout *cll;
4338 	int i;
4339 
4340 	sd = fl->sdesc;
4341 	for (i = 0; i < fl->sidx * 8; i++, sd++) {
4342 		if (sd->cl == NULL)
4343 			continue;
4344 
4345 		cll = &sd->cll;
4346 		clm = cl_metadata(sc, fl, cll, sd->cl);
4347 		if (sd->nmbuf == 0)
4348 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4349 		else if (clm && atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4350 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4351 			counter_u64_add(extfree_rels, 1);
4352 		}
4353 		sd->cl = NULL;
4354 	}
4355 
4356 	free(fl->sdesc, M_CXGBE);
4357 	fl->sdesc = NULL;
4358 }
4359 
4360 static inline void
4361 get_pkt_gl(struct mbuf *m, struct sglist *gl)
4362 {
4363 	int rc;
4364 
4365 	M_ASSERTPKTHDR(m);
4366 
4367 	sglist_reset(gl);
4368 	rc = sglist_append_mbuf(gl, m);
4369 	if (__predict_false(rc != 0)) {
4370 		panic("%s: mbuf %p (%d segs) was vetted earlier but now fails "
4371 		    "with %d.", __func__, m, mbuf_nsegs(m), rc);
4372 	}
4373 
4374 	KASSERT(gl->sg_nseg == mbuf_nsegs(m),
4375 	    ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m,
4376 	    mbuf_nsegs(m), gl->sg_nseg));
4377 	KASSERT(gl->sg_nseg > 0 &&
4378 	    gl->sg_nseg <= (needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS),
4379 	    ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__,
4380 		gl->sg_nseg, needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS));
4381 }
4382 
4383 /*
4384  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
4385  */
4386 static inline u_int
4387 txpkt_len16(u_int nsegs, u_int tso)
4388 {
4389 	u_int n;
4390 
4391 	MPASS(nsegs > 0);
4392 
4393 	nsegs--; /* first segment is part of ulptx_sgl */
4394 	n = sizeof(struct fw_eth_tx_pkt_wr) + sizeof(struct cpl_tx_pkt_core) +
4395 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4396 	if (tso)
4397 		n += sizeof(struct cpl_tx_pkt_lso_core);
4398 
4399 	return (howmany(n, 16));
4400 }
4401 
4402 /*
4403  * len16 for a txpkt_vm WR with a GL.  Includes the firmware work
4404  * request header.
4405  */
4406 static inline u_int
4407 txpkt_vm_len16(u_int nsegs, u_int tso)
4408 {
4409 	u_int n;
4410 
4411 	MPASS(nsegs > 0);
4412 
4413 	nsegs--; /* first segment is part of ulptx_sgl */
4414 	n = sizeof(struct fw_eth_tx_pkt_vm_wr) +
4415 	    sizeof(struct cpl_tx_pkt_core) +
4416 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4417 	if (tso)
4418 		n += sizeof(struct cpl_tx_pkt_lso_core);
4419 
4420 	return (howmany(n, 16));
4421 }
4422 
4423 /*
4424  * len16 for a txpkts type 0 WR with a GL.  Does not include the firmware work
4425  * request header.
4426  */
4427 static inline u_int
4428 txpkts0_len16(u_int nsegs)
4429 {
4430 	u_int n;
4431 
4432 	MPASS(nsegs > 0);
4433 
4434 	nsegs--; /* first segment is part of ulptx_sgl */
4435 	n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) +
4436 	    sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) +
4437 	    8 * ((3 * nsegs) / 2 + (nsegs & 1));
4438 
4439 	return (howmany(n, 16));
4440 }
4441 
4442 /*
4443  * len16 for a txpkts type 1 WR with a GL.  Does not include the firmware work
4444  * request header.
4445  */
4446 static inline u_int
4447 txpkts1_len16(void)
4448 {
4449 	u_int n;
4450 
4451 	n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl);
4452 
4453 	return (howmany(n, 16));
4454 }
4455 
4456 static inline u_int
4457 imm_payload(u_int ndesc)
4458 {
4459 	u_int n;
4460 
4461 	n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) -
4462 	    sizeof(struct cpl_tx_pkt_core);
4463 
4464 	return (n);
4465 }
4466 
4467 /*
4468  * Write a VM txpkt WR for this packet to the hardware descriptors, update the
4469  * software descriptor, and advance the pidx.  It is guaranteed that enough
4470  * descriptors are available.
4471  *
4472  * The return value is the # of hardware descriptors used.
4473  */
4474 static u_int
4475 write_txpkt_vm_wr(struct adapter *sc, struct sge_txq *txq,
4476     struct fw_eth_tx_pkt_vm_wr *wr, struct mbuf *m0, u_int available)
4477 {
4478 	struct sge_eq *eq = &txq->eq;
4479 	struct tx_sdesc *txsd;
4480 	struct cpl_tx_pkt_core *cpl;
4481 	uint32_t ctrl;	/* used in many unrelated places */
4482 	uint64_t ctrl1;
4483 	int csum_type, len16, ndesc, pktlen, nsegs;
4484 	caddr_t dst;
4485 
4486 	TXQ_LOCK_ASSERT_OWNED(txq);
4487 	M_ASSERTPKTHDR(m0);
4488 	MPASS(available > 0 && available < eq->sidx);
4489 
4490 	len16 = mbuf_len16(m0);
4491 	nsegs = mbuf_nsegs(m0);
4492 	pktlen = m0->m_pkthdr.len;
4493 	ctrl = sizeof(struct cpl_tx_pkt_core);
4494 	if (needs_tso(m0))
4495 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4496 	ndesc = howmany(len16, EQ_ESIZE / 16);
4497 	MPASS(ndesc <= available);
4498 
4499 	/* Firmware work request header */
4500 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4501 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_VM_WR) |
4502 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4503 
4504 	ctrl = V_FW_WR_LEN16(len16);
4505 	wr->equiq_to_len16 = htobe32(ctrl);
4506 	wr->r3[0] = 0;
4507 	wr->r3[1] = 0;
4508 
4509 	/*
4510 	 * Copy over ethmacdst, ethmacsrc, ethtype, and vlantci.
4511 	 * vlantci is ignored unless the ethtype is 0x8100, so it's
4512 	 * simpler to always copy it rather than making it
4513 	 * conditional.  Also, it seems that we do not have to set
4514 	 * vlantci or fake the ethtype when doing VLAN tag insertion.
4515 	 */
4516 	m_copydata(m0, 0, sizeof(struct ether_header) + 2, wr->ethmacdst);
4517 
4518 	csum_type = -1;
4519 	if (needs_tso(m0)) {
4520 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4521 
4522 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4523 		    m0->m_pkthdr.l4hlen > 0,
4524 		    ("%s: mbuf %p needs TSO but missing header lengths",
4525 			__func__, m0));
4526 
4527 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4528 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4529 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4530 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4531 			ctrl |= V_LSO_ETHHDR_LEN(1);
4532 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4533 			ctrl |= F_LSO_IPV6;
4534 
4535 		lso->lso_ctrl = htobe32(ctrl);
4536 		lso->ipid_ofst = htobe16(0);
4537 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4538 		lso->seqno_offset = htobe32(0);
4539 		lso->len = htobe32(pktlen);
4540 
4541 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4542 			csum_type = TX_CSUM_TCPIP6;
4543 		else
4544 			csum_type = TX_CSUM_TCPIP;
4545 
4546 		cpl = (void *)(lso + 1);
4547 
4548 		txq->tso_wrs++;
4549 	} else {
4550 		if (m0->m_pkthdr.csum_flags & CSUM_IP_TCP)
4551 			csum_type = TX_CSUM_TCPIP;
4552 		else if (m0->m_pkthdr.csum_flags & CSUM_IP_UDP)
4553 			csum_type = TX_CSUM_UDPIP;
4554 		else if (m0->m_pkthdr.csum_flags & CSUM_IP6_TCP)
4555 			csum_type = TX_CSUM_TCPIP6;
4556 		else if (m0->m_pkthdr.csum_flags & CSUM_IP6_UDP)
4557 			csum_type = TX_CSUM_UDPIP6;
4558 #if defined(INET)
4559 		else if (m0->m_pkthdr.csum_flags & CSUM_IP) {
4560 			/*
4561 			 * XXX: The firmware appears to stomp on the
4562 			 * fragment/flags field of the IP header when
4563 			 * using TX_CSUM_IP.  Fall back to doing
4564 			 * software checksums.
4565 			 */
4566 			u_short *sump;
4567 			struct mbuf *m;
4568 			int offset;
4569 
4570 			m = m0;
4571 			offset = 0;
4572 			sump = m_advance(&m, &offset, m0->m_pkthdr.l2hlen +
4573 			    offsetof(struct ip, ip_sum));
4574 			*sump = in_cksum_skip(m0, m0->m_pkthdr.l2hlen +
4575 			    m0->m_pkthdr.l3hlen, m0->m_pkthdr.l2hlen);
4576 			m0->m_pkthdr.csum_flags &= ~CSUM_IP;
4577 		}
4578 #endif
4579 
4580 		cpl = (void *)(wr + 1);
4581 	}
4582 
4583 	/* Checksum offload */
4584 	ctrl1 = 0;
4585 	if (needs_l3_csum(m0) == 0)
4586 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4587 	if (csum_type >= 0) {
4588 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0,
4589 	    ("%s: mbuf %p needs checksum offload but missing header lengths",
4590 			__func__, m0));
4591 
4592 		if (chip_id(sc) <= CHELSIO_T5) {
4593 			ctrl1 |= V_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4594 			    ETHER_HDR_LEN);
4595 		} else {
4596 			ctrl1 |= V_T6_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4597 			    ETHER_HDR_LEN);
4598 		}
4599 		ctrl1 |= V_TXPKT_IPHDR_LEN(m0->m_pkthdr.l3hlen);
4600 		ctrl1 |= V_TXPKT_CSUM_TYPE(csum_type);
4601 	} else
4602 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4603 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4604 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4605 		txq->txcsum++;	/* some hardware assistance provided */
4606 
4607 	/* VLAN tag insertion */
4608 	if (needs_vlan_insertion(m0)) {
4609 		ctrl1 |= F_TXPKT_VLAN_VLD |
4610 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4611 		txq->vlan_insertion++;
4612 	}
4613 
4614 	/* CPL header */
4615 	cpl->ctrl0 = txq->cpl_ctrl0;
4616 	cpl->pack = 0;
4617 	cpl->len = htobe16(pktlen);
4618 	cpl->ctrl1 = htobe64(ctrl1);
4619 
4620 	/* SGL */
4621 	dst = (void *)(cpl + 1);
4622 
4623 	/*
4624 	 * A packet using TSO will use up an entire descriptor for the
4625 	 * firmware work request header, LSO CPL, and TX_PKT_XT CPL.
4626 	 * If this descriptor is the last descriptor in the ring, wrap
4627 	 * around to the front of the ring explicitly for the start of
4628 	 * the sgl.
4629 	 */
4630 	if (dst == (void *)&eq->desc[eq->sidx]) {
4631 		dst = (void *)&eq->desc[0];
4632 		write_gl_to_txd(txq, m0, &dst, 0);
4633 	} else
4634 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4635 	txq->sgl_wrs++;
4636 
4637 	txq->txpkt_wrs++;
4638 
4639 	txsd = &txq->sdesc[eq->pidx];
4640 	txsd->m = m0;
4641 	txsd->desc_used = ndesc;
4642 
4643 	return (ndesc);
4644 }
4645 
4646 /*
4647  * Write a raw WR to the hardware descriptors, update the software
4648  * descriptor, and advance the pidx.  It is guaranteed that enough
4649  * descriptors are available.
4650  *
4651  * The return value is the # of hardware descriptors used.
4652  */
4653 static u_int
4654 write_raw_wr(struct sge_txq *txq, void *wr, struct mbuf *m0, u_int available)
4655 {
4656 	struct sge_eq *eq = &txq->eq;
4657 	struct tx_sdesc *txsd;
4658 	struct mbuf *m;
4659 	caddr_t dst;
4660 	int len16, ndesc;
4661 
4662 	len16 = mbuf_len16(m0);
4663 	ndesc = howmany(len16, EQ_ESIZE / 16);
4664 	MPASS(ndesc <= available);
4665 
4666 	dst = wr;
4667 	for (m = m0; m != NULL; m = m->m_next)
4668 		copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4669 
4670 	txq->raw_wrs++;
4671 
4672 	txsd = &txq->sdesc[eq->pidx];
4673 	txsd->m = m0;
4674 	txsd->desc_used = ndesc;
4675 
4676 	return (ndesc);
4677 }
4678 
4679 /*
4680  * Write a txpkt WR for this packet to the hardware descriptors, update the
4681  * software descriptor, and advance the pidx.  It is guaranteed that enough
4682  * descriptors are available.
4683  *
4684  * The return value is the # of hardware descriptors used.
4685  */
4686 static u_int
4687 write_txpkt_wr(struct sge_txq *txq, struct fw_eth_tx_pkt_wr *wr,
4688     struct mbuf *m0, u_int available)
4689 {
4690 	struct sge_eq *eq = &txq->eq;
4691 	struct tx_sdesc *txsd;
4692 	struct cpl_tx_pkt_core *cpl;
4693 	uint32_t ctrl;	/* used in many unrelated places */
4694 	uint64_t ctrl1;
4695 	int len16, ndesc, pktlen, nsegs;
4696 	caddr_t dst;
4697 
4698 	TXQ_LOCK_ASSERT_OWNED(txq);
4699 	M_ASSERTPKTHDR(m0);
4700 	MPASS(available > 0 && available < eq->sidx);
4701 
4702 	len16 = mbuf_len16(m0);
4703 	nsegs = mbuf_nsegs(m0);
4704 	pktlen = m0->m_pkthdr.len;
4705 	ctrl = sizeof(struct cpl_tx_pkt_core);
4706 	if (needs_tso(m0))
4707 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4708 	else if (pktlen <= imm_payload(2) && available >= 2) {
4709 		/* Immediate data.  Recalculate len16 and set nsegs to 0. */
4710 		ctrl += pktlen;
4711 		len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) +
4712 		    sizeof(struct cpl_tx_pkt_core) + pktlen, 16);
4713 		nsegs = 0;
4714 	}
4715 	ndesc = howmany(len16, EQ_ESIZE / 16);
4716 	MPASS(ndesc <= available);
4717 
4718 	/* Firmware work request header */
4719 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4720 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
4721 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4722 
4723 	ctrl = V_FW_WR_LEN16(len16);
4724 	wr->equiq_to_len16 = htobe32(ctrl);
4725 	wr->r3 = 0;
4726 
4727 	if (needs_tso(m0)) {
4728 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4729 
4730 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4731 		    m0->m_pkthdr.l4hlen > 0,
4732 		    ("%s: mbuf %p needs TSO but missing header lengths",
4733 			__func__, m0));
4734 
4735 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4736 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4737 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4738 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4739 			ctrl |= V_LSO_ETHHDR_LEN(1);
4740 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4741 			ctrl |= F_LSO_IPV6;
4742 
4743 		lso->lso_ctrl = htobe32(ctrl);
4744 		lso->ipid_ofst = htobe16(0);
4745 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4746 		lso->seqno_offset = htobe32(0);
4747 		lso->len = htobe32(pktlen);
4748 
4749 		cpl = (void *)(lso + 1);
4750 
4751 		txq->tso_wrs++;
4752 	} else
4753 		cpl = (void *)(wr + 1);
4754 
4755 	/* Checksum offload */
4756 	ctrl1 = 0;
4757 	if (needs_l3_csum(m0) == 0)
4758 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4759 	if (needs_l4_csum(m0) == 0)
4760 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4761 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4762 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4763 		txq->txcsum++;	/* some hardware assistance provided */
4764 
4765 	/* VLAN tag insertion */
4766 	if (needs_vlan_insertion(m0)) {
4767 		ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4768 		txq->vlan_insertion++;
4769 	}
4770 
4771 	/* CPL header */
4772 	cpl->ctrl0 = txq->cpl_ctrl0;
4773 	cpl->pack = 0;
4774 	cpl->len = htobe16(pktlen);
4775 	cpl->ctrl1 = htobe64(ctrl1);
4776 
4777 	/* SGL */
4778 	dst = (void *)(cpl + 1);
4779 	if (nsegs > 0) {
4780 
4781 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4782 		txq->sgl_wrs++;
4783 	} else {
4784 		struct mbuf *m;
4785 
4786 		for (m = m0; m != NULL; m = m->m_next) {
4787 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4788 #ifdef INVARIANTS
4789 			pktlen -= m->m_len;
4790 #endif
4791 		}
4792 #ifdef INVARIANTS
4793 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
4794 #endif
4795 		txq->imm_wrs++;
4796 	}
4797 
4798 	txq->txpkt_wrs++;
4799 
4800 	txsd = &txq->sdesc[eq->pidx];
4801 	txsd->m = m0;
4802 	txsd->desc_used = ndesc;
4803 
4804 	return (ndesc);
4805 }
4806 
4807 static int
4808 try_txpkts(struct mbuf *m, struct mbuf *n, struct txpkts *txp, u_int available)
4809 {
4810 	u_int needed, nsegs1, nsegs2, l1, l2;
4811 
4812 	if (cannot_use_txpkts(m) || cannot_use_txpkts(n))
4813 		return (1);
4814 
4815 	nsegs1 = mbuf_nsegs(m);
4816 	nsegs2 = mbuf_nsegs(n);
4817 	if (nsegs1 + nsegs2 == 2) {
4818 		txp->wr_type = 1;
4819 		l1 = l2 = txpkts1_len16();
4820 	} else {
4821 		txp->wr_type = 0;
4822 		l1 = txpkts0_len16(nsegs1);
4823 		l2 = txpkts0_len16(nsegs2);
4824 	}
4825 	txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) + l1 + l2;
4826 	needed = howmany(txp->len16, EQ_ESIZE / 16);
4827 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4828 		return (1);
4829 
4830 	txp->plen = m->m_pkthdr.len + n->m_pkthdr.len;
4831 	if (txp->plen > 65535)
4832 		return (1);
4833 
4834 	txp->npkt = 2;
4835 	set_mbuf_len16(m, l1);
4836 	set_mbuf_len16(n, l2);
4837 
4838 	return (0);
4839 }
4840 
4841 static int
4842 add_to_txpkts(struct mbuf *m, struct txpkts *txp, u_int available)
4843 {
4844 	u_int plen, len16, needed, nsegs;
4845 
4846 	MPASS(txp->wr_type == 0 || txp->wr_type == 1);
4847 
4848 	if (cannot_use_txpkts(m))
4849 		return (1);
4850 
4851 	nsegs = mbuf_nsegs(m);
4852 	if (txp->wr_type == 1 && nsegs != 1)
4853 		return (1);
4854 
4855 	plen = txp->plen + m->m_pkthdr.len;
4856 	if (plen > 65535)
4857 		return (1);
4858 
4859 	if (txp->wr_type == 0)
4860 		len16 = txpkts0_len16(nsegs);
4861 	else
4862 		len16 = txpkts1_len16();
4863 	needed = howmany(txp->len16 + len16, EQ_ESIZE / 16);
4864 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4865 		return (1);
4866 
4867 	txp->npkt++;
4868 	txp->plen = plen;
4869 	txp->len16 += len16;
4870 	set_mbuf_len16(m, len16);
4871 
4872 	return (0);
4873 }
4874 
4875 /*
4876  * Write a txpkts WR for the packets in txp to the hardware descriptors, update
4877  * the software descriptor, and advance the pidx.  It is guaranteed that enough
4878  * descriptors are available.
4879  *
4880  * The return value is the # of hardware descriptors used.
4881  */
4882 static u_int
4883 write_txpkts_wr(struct sge_txq *txq, struct fw_eth_tx_pkts_wr *wr,
4884     struct mbuf *m0, const struct txpkts *txp, u_int available)
4885 {
4886 	struct sge_eq *eq = &txq->eq;
4887 	struct tx_sdesc *txsd;
4888 	struct cpl_tx_pkt_core *cpl;
4889 	uint32_t ctrl;
4890 	uint64_t ctrl1;
4891 	int ndesc, checkwrap;
4892 	struct mbuf *m;
4893 	void *flitp;
4894 
4895 	TXQ_LOCK_ASSERT_OWNED(txq);
4896 	MPASS(txp->npkt > 0);
4897 	MPASS(txp->plen < 65536);
4898 	MPASS(m0 != NULL);
4899 	MPASS(m0->m_nextpkt != NULL);
4900 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
4901 	MPASS(available > 0 && available < eq->sidx);
4902 
4903 	ndesc = howmany(txp->len16, EQ_ESIZE / 16);
4904 	MPASS(ndesc <= available);
4905 
4906 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4907 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
4908 	ctrl = V_FW_WR_LEN16(txp->len16);
4909 	wr->equiq_to_len16 = htobe32(ctrl);
4910 	wr->plen = htobe16(txp->plen);
4911 	wr->npkt = txp->npkt;
4912 	wr->r3 = 0;
4913 	wr->type = txp->wr_type;
4914 	flitp = wr + 1;
4915 
4916 	/*
4917 	 * At this point we are 16B into a hardware descriptor.  If checkwrap is
4918 	 * set then we know the WR is going to wrap around somewhere.  We'll
4919 	 * check for that at appropriate points.
4920 	 */
4921 	checkwrap = eq->sidx - ndesc < eq->pidx;
4922 	for (m = m0; m != NULL; m = m->m_nextpkt) {
4923 		if (txp->wr_type == 0) {
4924 			struct ulp_txpkt *ulpmc;
4925 			struct ulptx_idata *ulpsc;
4926 
4927 			/* ULP master command */
4928 			ulpmc = flitp;
4929 			ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
4930 			    V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid));
4931 			ulpmc->len = htobe32(mbuf_len16(m));
4932 
4933 			/* ULP subcommand */
4934 			ulpsc = (void *)(ulpmc + 1);
4935 			ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
4936 			    F_ULP_TX_SC_MORE);
4937 			ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
4938 
4939 			cpl = (void *)(ulpsc + 1);
4940 			if (checkwrap &&
4941 			    (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx])
4942 				cpl = (void *)&eq->desc[0];
4943 		} else {
4944 			cpl = flitp;
4945 		}
4946 
4947 		/* Checksum offload */
4948 		ctrl1 = 0;
4949 		if (needs_l3_csum(m) == 0)
4950 			ctrl1 |= F_TXPKT_IPCSUM_DIS;
4951 		if (needs_l4_csum(m) == 0)
4952 			ctrl1 |= F_TXPKT_L4CSUM_DIS;
4953 		if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4954 		    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4955 			txq->txcsum++;	/* some hardware assistance provided */
4956 
4957 		/* VLAN tag insertion */
4958 		if (needs_vlan_insertion(m)) {
4959 			ctrl1 |= F_TXPKT_VLAN_VLD |
4960 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
4961 			txq->vlan_insertion++;
4962 		}
4963 
4964 		/* CPL header */
4965 		cpl->ctrl0 = txq->cpl_ctrl0;
4966 		cpl->pack = 0;
4967 		cpl->len = htobe16(m->m_pkthdr.len);
4968 		cpl->ctrl1 = htobe64(ctrl1);
4969 
4970 		flitp = cpl + 1;
4971 		if (checkwrap &&
4972 		    (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
4973 			flitp = (void *)&eq->desc[0];
4974 
4975 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap);
4976 
4977 	}
4978 
4979 	if (txp->wr_type == 0) {
4980 		txq->txpkts0_pkts += txp->npkt;
4981 		txq->txpkts0_wrs++;
4982 	} else {
4983 		txq->txpkts1_pkts += txp->npkt;
4984 		txq->txpkts1_wrs++;
4985 	}
4986 
4987 	txsd = &txq->sdesc[eq->pidx];
4988 	txsd->m = m0;
4989 	txsd->desc_used = ndesc;
4990 
4991 	return (ndesc);
4992 }
4993 
4994 /*
4995  * If the SGL ends on an address that is not 16 byte aligned, this function will
4996  * add a 0 filled flit at the end.
4997  */
4998 static void
4999 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap)
5000 {
5001 	struct sge_eq *eq = &txq->eq;
5002 	struct sglist *gl = txq->gl;
5003 	struct sglist_seg *seg;
5004 	__be64 *flitp, *wrap;
5005 	struct ulptx_sgl *usgl;
5006 	int i, nflits, nsegs;
5007 
5008 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
5009 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
5010 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
5011 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
5012 
5013 	get_pkt_gl(m, gl);
5014 	nsegs = gl->sg_nseg;
5015 	MPASS(nsegs > 0);
5016 
5017 	nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
5018 	flitp = (__be64 *)(*to);
5019 	wrap = (__be64 *)(&eq->desc[eq->sidx]);
5020 	seg = &gl->sg_segs[0];
5021 	usgl = (void *)flitp;
5022 
5023 	/*
5024 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
5025 	 * ring, so we're at least 16 bytes away from the status page.  There is
5026 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
5027 	 */
5028 
5029 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
5030 	    V_ULPTX_NSGE(nsegs));
5031 	usgl->len0 = htobe32(seg->ss_len);
5032 	usgl->addr0 = htobe64(seg->ss_paddr);
5033 	seg++;
5034 
5035 	if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) {
5036 
5037 		/* Won't wrap around at all */
5038 
5039 		for (i = 0; i < nsegs - 1; i++, seg++) {
5040 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
5041 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
5042 		}
5043 		if (i & 1)
5044 			usgl->sge[i / 2].len[1] = htobe32(0);
5045 		flitp += nflits;
5046 	} else {
5047 
5048 		/* Will wrap somewhere in the rest of the SGL */
5049 
5050 		/* 2 flits already written, write the rest flit by flit */
5051 		flitp = (void *)(usgl + 1);
5052 		for (i = 0; i < nflits - 2; i++) {
5053 			if (flitp == wrap)
5054 				flitp = (void *)eq->desc;
5055 			*flitp++ = get_flit(seg, nsegs - 1, i);
5056 		}
5057 	}
5058 
5059 	if (nflits & 1) {
5060 		MPASS(((uintptr_t)flitp) & 0xf);
5061 		*flitp++ = 0;
5062 	}
5063 
5064 	MPASS((((uintptr_t)flitp) & 0xf) == 0);
5065 	if (__predict_false(flitp == wrap))
5066 		*to = (void *)eq->desc;
5067 	else
5068 		*to = (void *)flitp;
5069 }
5070 
5071 static inline void
5072 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
5073 {
5074 
5075 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
5076 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
5077 
5078 	if (__predict_true((uintptr_t)(*to) + len <=
5079 	    (uintptr_t)&eq->desc[eq->sidx])) {
5080 		bcopy(from, *to, len);
5081 		(*to) += len;
5082 	} else {
5083 		int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
5084 
5085 		bcopy(from, *to, portion);
5086 		from += portion;
5087 		portion = len - portion;	/* remaining */
5088 		bcopy(from, (void *)eq->desc, portion);
5089 		(*to) = (caddr_t)eq->desc + portion;
5090 	}
5091 }
5092 
5093 static inline void
5094 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n)
5095 {
5096 	u_int db;
5097 
5098 	MPASS(n > 0);
5099 
5100 	db = eq->doorbells;
5101 	if (n > 1)
5102 		clrbit(&db, DOORBELL_WCWR);
5103 	wmb();
5104 
5105 	switch (ffs(db) - 1) {
5106 	case DOORBELL_UDB:
5107 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
5108 		break;
5109 
5110 	case DOORBELL_WCWR: {
5111 		volatile uint64_t *dst, *src;
5112 		int i;
5113 
5114 		/*
5115 		 * Queues whose 128B doorbell segment fits in the page do not
5116 		 * use relative qid (udb_qid is always 0).  Only queues with
5117 		 * doorbell segments can do WCWR.
5118 		 */
5119 		KASSERT(eq->udb_qid == 0 && n == 1,
5120 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
5121 		    __func__, eq->doorbells, n, eq->dbidx, eq));
5122 
5123 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
5124 		    UDBS_DB_OFFSET);
5125 		i = eq->dbidx;
5126 		src = (void *)&eq->desc[i];
5127 		while (src != (void *)&eq->desc[i + 1])
5128 			*dst++ = *src++;
5129 		wmb();
5130 		break;
5131 	}
5132 
5133 	case DOORBELL_UDBWC:
5134 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
5135 		wmb();
5136 		break;
5137 
5138 	case DOORBELL_KDB:
5139 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
5140 		    V_QID(eq->cntxt_id) | V_PIDX(n));
5141 		break;
5142 	}
5143 
5144 	IDXINCR(eq->dbidx, n, eq->sidx);
5145 }
5146 
5147 static inline u_int
5148 reclaimable_tx_desc(struct sge_eq *eq)
5149 {
5150 	uint16_t hw_cidx;
5151 
5152 	hw_cidx = read_hw_cidx(eq);
5153 	return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx));
5154 }
5155 
5156 static inline u_int
5157 total_available_tx_desc(struct sge_eq *eq)
5158 {
5159 	uint16_t hw_cidx, pidx;
5160 
5161 	hw_cidx = read_hw_cidx(eq);
5162 	pidx = eq->pidx;
5163 
5164 	if (pidx == hw_cidx)
5165 		return (eq->sidx - 1);
5166 	else
5167 		return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1);
5168 }
5169 
5170 static inline uint16_t
5171 read_hw_cidx(struct sge_eq *eq)
5172 {
5173 	struct sge_qstat *spg = (void *)&eq->desc[eq->sidx];
5174 	uint16_t cidx = spg->cidx;	/* stable snapshot */
5175 
5176 	return (be16toh(cidx));
5177 }
5178 
5179 /*
5180  * Reclaim 'n' descriptors approximately.
5181  */
5182 static u_int
5183 reclaim_tx_descs(struct sge_txq *txq, u_int n)
5184 {
5185 	struct tx_sdesc *txsd;
5186 	struct sge_eq *eq = &txq->eq;
5187 	u_int can_reclaim, reclaimed;
5188 
5189 	TXQ_LOCK_ASSERT_OWNED(txq);
5190 	MPASS(n > 0);
5191 
5192 	reclaimed = 0;
5193 	can_reclaim = reclaimable_tx_desc(eq);
5194 	while (can_reclaim && reclaimed < n) {
5195 		int ndesc;
5196 		struct mbuf *m, *nextpkt;
5197 
5198 		txsd = &txq->sdesc[eq->cidx];
5199 		ndesc = txsd->desc_used;
5200 
5201 		/* Firmware doesn't return "partial" credits. */
5202 		KASSERT(can_reclaim >= ndesc,
5203 		    ("%s: unexpected number of credits: %d, %d",
5204 		    __func__, can_reclaim, ndesc));
5205 		KASSERT(ndesc != 0,
5206 		    ("%s: descriptor with no credits: cidx %d",
5207 		    __func__, eq->cidx));
5208 
5209 		for (m = txsd->m; m != NULL; m = nextpkt) {
5210 			nextpkt = m->m_nextpkt;
5211 			m->m_nextpkt = NULL;
5212 			m_freem(m);
5213 		}
5214 		reclaimed += ndesc;
5215 		can_reclaim -= ndesc;
5216 		IDXINCR(eq->cidx, ndesc, eq->sidx);
5217 	}
5218 
5219 	return (reclaimed);
5220 }
5221 
5222 static void
5223 tx_reclaim(void *arg, int n)
5224 {
5225 	struct sge_txq *txq = arg;
5226 	struct sge_eq *eq = &txq->eq;
5227 
5228 	do {
5229 		if (TXQ_TRYLOCK(txq) == 0)
5230 			break;
5231 		n = reclaim_tx_descs(txq, 32);
5232 		if (eq->cidx == eq->pidx)
5233 			eq->equeqidx = eq->pidx;
5234 		TXQ_UNLOCK(txq);
5235 	} while (n > 0);
5236 }
5237 
5238 static __be64
5239 get_flit(struct sglist_seg *segs, int nsegs, int idx)
5240 {
5241 	int i = (idx / 3) * 2;
5242 
5243 	switch (idx % 3) {
5244 	case 0: {
5245 		uint64_t rc;
5246 
5247 		rc = (uint64_t)segs[i].ss_len << 32;
5248 		if (i + 1 < nsegs)
5249 			rc |= (uint64_t)(segs[i + 1].ss_len);
5250 
5251 		return (htobe64(rc));
5252 	}
5253 	case 1:
5254 		return (htobe64(segs[i].ss_paddr));
5255 	case 2:
5256 		return (htobe64(segs[i + 1].ss_paddr));
5257 	}
5258 
5259 	return (0);
5260 }
5261 
5262 static void
5263 find_best_refill_source(struct adapter *sc, struct sge_fl *fl, int maxp)
5264 {
5265 	int8_t zidx, hwidx, idx;
5266 	uint16_t region1, region3;
5267 	int spare, spare_needed, n;
5268 	struct sw_zone_info *swz;
5269 	struct hw_buf_info *hwb, *hwb_list = &sc->sge.hw_buf_info[0];
5270 
5271 	/*
5272 	 * Buffer Packing: Look for PAGE_SIZE or larger zone which has a bufsize
5273 	 * large enough for the max payload and cluster metadata.  Otherwise
5274 	 * settle for the largest bufsize that leaves enough room in the cluster
5275 	 * for metadata.
5276 	 *
5277 	 * Without buffer packing: Look for the smallest zone which has a
5278 	 * bufsize large enough for the max payload.  Settle for the largest
5279 	 * bufsize available if there's nothing big enough for max payload.
5280 	 */
5281 	spare_needed = fl->flags & FL_BUF_PACKING ? CL_METADATA_SIZE : 0;
5282 	swz = &sc->sge.sw_zone_info[0];
5283 	hwidx = -1;
5284 	for (zidx = 0; zidx < SW_ZONE_SIZES; zidx++, swz++) {
5285 		if (swz->size > largest_rx_cluster) {
5286 			if (__predict_true(hwidx != -1))
5287 				break;
5288 
5289 			/*
5290 			 * This is a misconfiguration.  largest_rx_cluster is
5291 			 * preventing us from finding a refill source.  See
5292 			 * dev.t5nex.<n>.buffer_sizes to figure out why.
5293 			 */
5294 			device_printf(sc->dev, "largest_rx_cluster=%u leaves no"
5295 			    " refill source for fl %p (dma %u).  Ignored.\n",
5296 			    largest_rx_cluster, fl, maxp);
5297 		}
5298 		for (idx = swz->head_hwidx; idx != -1; idx = hwb->next) {
5299 			hwb = &hwb_list[idx];
5300 			spare = swz->size - hwb->size;
5301 			if (spare < spare_needed)
5302 				continue;
5303 
5304 			hwidx = idx;		/* best option so far */
5305 			if (hwb->size >= maxp) {
5306 
5307 				if ((fl->flags & FL_BUF_PACKING) == 0)
5308 					goto done; /* stop looking (not packing) */
5309 
5310 				if (swz->size >= safest_rx_cluster)
5311 					goto done; /* stop looking (packing) */
5312 			}
5313 			break;		/* keep looking, next zone */
5314 		}
5315 	}
5316 done:
5317 	/* A usable hwidx has been located. */
5318 	MPASS(hwidx != -1);
5319 	hwb = &hwb_list[hwidx];
5320 	zidx = hwb->zidx;
5321 	swz = &sc->sge.sw_zone_info[zidx];
5322 	region1 = 0;
5323 	region3 = swz->size - hwb->size;
5324 
5325 	/*
5326 	 * Stay within this zone and see if there is a better match when mbuf
5327 	 * inlining is allowed.  Remember that the hwidx's are sorted in
5328 	 * decreasing order of size (so in increasing order of spare area).
5329 	 */
5330 	for (idx = hwidx; idx != -1; idx = hwb->next) {
5331 		hwb = &hwb_list[idx];
5332 		spare = swz->size - hwb->size;
5333 
5334 		if (allow_mbufs_in_cluster == 0 || hwb->size < maxp)
5335 			break;
5336 
5337 		/*
5338 		 * Do not inline mbufs if doing so would violate the pad/pack
5339 		 * boundary alignment requirement.
5340 		 */
5341 		if (fl_pad && (MSIZE % sc->params.sge.pad_boundary) != 0)
5342 			continue;
5343 		if (fl->flags & FL_BUF_PACKING &&
5344 		    (MSIZE % sc->params.sge.pack_boundary) != 0)
5345 			continue;
5346 
5347 		if (spare < CL_METADATA_SIZE + MSIZE)
5348 			continue;
5349 		n = (spare - CL_METADATA_SIZE) / MSIZE;
5350 		if (n > howmany(hwb->size, maxp))
5351 			break;
5352 
5353 		hwidx = idx;
5354 		if (fl->flags & FL_BUF_PACKING) {
5355 			region1 = n * MSIZE;
5356 			region3 = spare - region1;
5357 		} else {
5358 			region1 = MSIZE;
5359 			region3 = spare - region1;
5360 			break;
5361 		}
5362 	}
5363 
5364 	KASSERT(zidx >= 0 && zidx < SW_ZONE_SIZES,
5365 	    ("%s: bad zone %d for fl %p, maxp %d", __func__, zidx, fl, maxp));
5366 	KASSERT(hwidx >= 0 && hwidx <= SGE_FLBUF_SIZES,
5367 	    ("%s: bad hwidx %d for fl %p, maxp %d", __func__, hwidx, fl, maxp));
5368 	KASSERT(region1 + sc->sge.hw_buf_info[hwidx].size + region3 ==
5369 	    sc->sge.sw_zone_info[zidx].size,
5370 	    ("%s: bad buffer layout for fl %p, maxp %d. "
5371 		"cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5372 		sc->sge.sw_zone_info[zidx].size, region1,
5373 		sc->sge.hw_buf_info[hwidx].size, region3));
5374 	if (fl->flags & FL_BUF_PACKING || region1 > 0) {
5375 		KASSERT(region3 >= CL_METADATA_SIZE,
5376 		    ("%s: no room for metadata.  fl %p, maxp %d; "
5377 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5378 		    sc->sge.sw_zone_info[zidx].size, region1,
5379 		    sc->sge.hw_buf_info[hwidx].size, region3));
5380 		KASSERT(region1 % MSIZE == 0,
5381 		    ("%s: bad mbuf region for fl %p, maxp %d. "
5382 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5383 		    sc->sge.sw_zone_info[zidx].size, region1,
5384 		    sc->sge.hw_buf_info[hwidx].size, region3));
5385 	}
5386 
5387 	fl->cll_def.zidx = zidx;
5388 	fl->cll_def.hwidx = hwidx;
5389 	fl->cll_def.region1 = region1;
5390 	fl->cll_def.region3 = region3;
5391 }
5392 
5393 static void
5394 find_safe_refill_source(struct adapter *sc, struct sge_fl *fl)
5395 {
5396 	struct sge *s = &sc->sge;
5397 	struct hw_buf_info *hwb;
5398 	struct sw_zone_info *swz;
5399 	int spare;
5400 	int8_t hwidx;
5401 
5402 	if (fl->flags & FL_BUF_PACKING)
5403 		hwidx = s->safe_hwidx2;	/* with room for metadata */
5404 	else if (allow_mbufs_in_cluster && s->safe_hwidx2 != -1) {
5405 		hwidx = s->safe_hwidx2;
5406 		hwb = &s->hw_buf_info[hwidx];
5407 		swz = &s->sw_zone_info[hwb->zidx];
5408 		spare = swz->size - hwb->size;
5409 
5410 		/* no good if there isn't room for an mbuf as well */
5411 		if (spare < CL_METADATA_SIZE + MSIZE)
5412 			hwidx = s->safe_hwidx1;
5413 	} else
5414 		hwidx = s->safe_hwidx1;
5415 
5416 	if (hwidx == -1) {
5417 		/* No fallback source */
5418 		fl->cll_alt.hwidx = -1;
5419 		fl->cll_alt.zidx = -1;
5420 
5421 		return;
5422 	}
5423 
5424 	hwb = &s->hw_buf_info[hwidx];
5425 	swz = &s->sw_zone_info[hwb->zidx];
5426 	spare = swz->size - hwb->size;
5427 	fl->cll_alt.hwidx = hwidx;
5428 	fl->cll_alt.zidx = hwb->zidx;
5429 	if (allow_mbufs_in_cluster &&
5430 	    (fl_pad == 0 || (MSIZE % sc->params.sge.pad_boundary) == 0))
5431 		fl->cll_alt.region1 = ((spare - CL_METADATA_SIZE) / MSIZE) * MSIZE;
5432 	else
5433 		fl->cll_alt.region1 = 0;
5434 	fl->cll_alt.region3 = spare - fl->cll_alt.region1;
5435 }
5436 
5437 static void
5438 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
5439 {
5440 	mtx_lock(&sc->sfl_lock);
5441 	FL_LOCK(fl);
5442 	if ((fl->flags & FL_DOOMED) == 0) {
5443 		fl->flags |= FL_STARVING;
5444 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
5445 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
5446 	}
5447 	FL_UNLOCK(fl);
5448 	mtx_unlock(&sc->sfl_lock);
5449 }
5450 
5451 static void
5452 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq)
5453 {
5454 	struct sge_wrq *wrq = (void *)eq;
5455 
5456 	atomic_readandclear_int(&eq->equiq);
5457 	taskqueue_enqueue(sc->tq[eq->tx_chan], &wrq->wrq_tx_task);
5458 }
5459 
5460 static void
5461 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq)
5462 {
5463 	struct sge_txq *txq = (void *)eq;
5464 
5465 	MPASS((eq->flags & EQ_TYPEMASK) == EQ_ETH);
5466 
5467 	atomic_readandclear_int(&eq->equiq);
5468 	mp_ring_check_drainage(txq->r, 0);
5469 	taskqueue_enqueue(sc->tq[eq->tx_chan], &txq->tx_reclaim_task);
5470 }
5471 
5472 static int
5473 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
5474     struct mbuf *m)
5475 {
5476 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
5477 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
5478 	struct adapter *sc = iq->adapter;
5479 	struct sge *s = &sc->sge;
5480 	struct sge_eq *eq;
5481 	static void (*h[])(struct adapter *, struct sge_eq *) = {NULL,
5482 		&handle_wrq_egr_update, &handle_eth_egr_update,
5483 		&handle_wrq_egr_update};
5484 
5485 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5486 	    rss->opcode));
5487 
5488 	eq = s->eqmap[qid - s->eq_start - s->eq_base];
5489 	(*h[eq->flags & EQ_TYPEMASK])(sc, eq);
5490 
5491 	return (0);
5492 }
5493 
5494 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
5495 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
5496     offsetof(struct cpl_fw6_msg, data));
5497 
5498 static int
5499 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
5500 {
5501 	struct adapter *sc = iq->adapter;
5502 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
5503 
5504 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5505 	    rss->opcode));
5506 
5507 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
5508 		const struct rss_header *rss2;
5509 
5510 		rss2 = (const struct rss_header *)&cpl->data[0];
5511 		return (t4_cpl_handler[rss2->opcode](iq, rss2, m));
5512 	}
5513 
5514 	return (t4_fw_msg_handler[cpl->type](sc, &cpl->data[0]));
5515 }
5516 
5517 /**
5518  *	t4_handle_wrerr_rpl - process a FW work request error message
5519  *	@adap: the adapter
5520  *	@rpl: start of the FW message
5521  */
5522 static int
5523 t4_handle_wrerr_rpl(struct adapter *adap, const __be64 *rpl)
5524 {
5525 	u8 opcode = *(const u8 *)rpl;
5526 	const struct fw_error_cmd *e = (const void *)rpl;
5527 	unsigned int i;
5528 
5529 	if (opcode != FW_ERROR_CMD) {
5530 		log(LOG_ERR,
5531 		    "%s: Received WRERR_RPL message with opcode %#x\n",
5532 		    device_get_nameunit(adap->dev), opcode);
5533 		return (EINVAL);
5534 	}
5535 	log(LOG_ERR, "%s: FW_ERROR (%s) ", device_get_nameunit(adap->dev),
5536 	    G_FW_ERROR_CMD_FATAL(be32toh(e->op_to_type)) ? "fatal" :
5537 	    "non-fatal");
5538 	switch (G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type))) {
5539 	case FW_ERROR_TYPE_EXCEPTION:
5540 		log(LOG_ERR, "exception info:\n");
5541 		for (i = 0; i < nitems(e->u.exception.info); i++)
5542 			log(LOG_ERR, "%s%08x", i == 0 ? "\t" : " ",
5543 			    be32toh(e->u.exception.info[i]));
5544 		log(LOG_ERR, "\n");
5545 		break;
5546 	case FW_ERROR_TYPE_HWMODULE:
5547 		log(LOG_ERR, "HW module regaddr %08x regval %08x\n",
5548 		    be32toh(e->u.hwmodule.regaddr),
5549 		    be32toh(e->u.hwmodule.regval));
5550 		break;
5551 	case FW_ERROR_TYPE_WR:
5552 		log(LOG_ERR, "WR cidx %d PF %d VF %d eqid %d hdr:\n",
5553 		    be16toh(e->u.wr.cidx),
5554 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.wr.pfn_vfn)),
5555 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.wr.pfn_vfn)),
5556 		    be32toh(e->u.wr.eqid));
5557 		for (i = 0; i < nitems(e->u.wr.wrhdr); i++)
5558 			log(LOG_ERR, "%s%02x", i == 0 ? "\t" : " ",
5559 			    e->u.wr.wrhdr[i]);
5560 		log(LOG_ERR, "\n");
5561 		break;
5562 	case FW_ERROR_TYPE_ACL:
5563 		log(LOG_ERR, "ACL cidx %d PF %d VF %d eqid %d %s",
5564 		    be16toh(e->u.acl.cidx),
5565 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.acl.pfn_vfn)),
5566 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.acl.pfn_vfn)),
5567 		    be32toh(e->u.acl.eqid),
5568 		    G_FW_ERROR_CMD_MV(be16toh(e->u.acl.mv_pkd)) ? "vlanid" :
5569 		    "MAC");
5570 		for (i = 0; i < nitems(e->u.acl.val); i++)
5571 			log(LOG_ERR, " %02x", e->u.acl.val[i]);
5572 		log(LOG_ERR, "\n");
5573 		break;
5574 	default:
5575 		log(LOG_ERR, "type %#x\n",
5576 		    G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type)));
5577 		return (EINVAL);
5578 	}
5579 	return (0);
5580 }
5581 
5582 static int
5583 sysctl_uint16(SYSCTL_HANDLER_ARGS)
5584 {
5585 	uint16_t *id = arg1;
5586 	int i = *id;
5587 
5588 	return sysctl_handle_int(oidp, &i, 0, req);
5589 }
5590 
5591 static int
5592 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
5593 {
5594 	struct sge *s = arg1;
5595 	struct hw_buf_info *hwb = &s->hw_buf_info[0];
5596 	struct sw_zone_info *swz = &s->sw_zone_info[0];
5597 	int i, rc;
5598 	struct sbuf sb;
5599 	char c;
5600 
5601 	sbuf_new(&sb, NULL, 32, SBUF_AUTOEXTEND);
5602 	for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
5603 		if (hwb->zidx >= 0 && swz[hwb->zidx].size <= largest_rx_cluster)
5604 			c = '*';
5605 		else
5606 			c = '\0';
5607 
5608 		sbuf_printf(&sb, "%u%c ", hwb->size, c);
5609 	}
5610 	sbuf_trim(&sb);
5611 	sbuf_finish(&sb);
5612 	rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
5613 	sbuf_delete(&sb);
5614 	return (rc);
5615 }
5616 
5617 #ifdef RATELIMIT
5618 /*
5619  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
5620  */
5621 static inline u_int
5622 txpkt_eo_len16(u_int nsegs, u_int immhdrs, u_int tso)
5623 {
5624 	u_int n;
5625 
5626 	MPASS(immhdrs > 0);
5627 
5628 	n = roundup2(sizeof(struct fw_eth_tx_eo_wr) +
5629 	    sizeof(struct cpl_tx_pkt_core) + immhdrs, 16);
5630 	if (__predict_false(nsegs == 0))
5631 		goto done;
5632 
5633 	nsegs--; /* first segment is part of ulptx_sgl */
5634 	n += sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
5635 	if (tso)
5636 		n += sizeof(struct cpl_tx_pkt_lso_core);
5637 
5638 done:
5639 	return (howmany(n, 16));
5640 }
5641 
5642 #define ETID_FLOWC_NPARAMS 6
5643 #define ETID_FLOWC_LEN (roundup2((sizeof(struct fw_flowc_wr) + \
5644     ETID_FLOWC_NPARAMS * sizeof(struct fw_flowc_mnemval)), 16))
5645 #define ETID_FLOWC_LEN16 (howmany(ETID_FLOWC_LEN, 16))
5646 
5647 static int
5648 send_etid_flowc_wr(struct cxgbe_snd_tag *cst, struct port_info *pi,
5649     struct vi_info *vi)
5650 {
5651 	struct wrq_cookie cookie;
5652 	u_int pfvf = G_FW_VIID_PFN(vi->viid) << S_FW_VIID_PFN;
5653 	struct fw_flowc_wr *flowc;
5654 
5655 	mtx_assert(&cst->lock, MA_OWNED);
5656 	MPASS((cst->flags & (EO_FLOWC_PENDING | EO_FLOWC_RPL_PENDING)) ==
5657 	    EO_FLOWC_PENDING);
5658 
5659 	flowc = start_wrq_wr(cst->eo_txq, ETID_FLOWC_LEN16, &cookie);
5660 	if (__predict_false(flowc == NULL))
5661 		return (ENOMEM);
5662 
5663 	bzero(flowc, ETID_FLOWC_LEN);
5664 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
5665 	    V_FW_FLOWC_WR_NPARAMS(ETID_FLOWC_NPARAMS) | V_FW_WR_COMPL(0));
5666 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(ETID_FLOWC_LEN16) |
5667 	    V_FW_WR_FLOWID(cst->etid));
5668 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN;
5669 	flowc->mnemval[0].val = htobe32(pfvf);
5670 	flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH;
5671 	flowc->mnemval[1].val = htobe32(pi->tx_chan);
5672 	flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT;
5673 	flowc->mnemval[2].val = htobe32(pi->tx_chan);
5674 	flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID;
5675 	flowc->mnemval[3].val = htobe32(cst->iqid);
5676 	flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_EOSTATE;
5677 	flowc->mnemval[4].val = htobe32(FW_FLOWC_MNEM_EOSTATE_ESTABLISHED);
5678 	flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS;
5679 	flowc->mnemval[5].val = htobe32(cst->schedcl);
5680 
5681 	commit_wrq_wr(cst->eo_txq, flowc, &cookie);
5682 
5683 	cst->flags &= ~EO_FLOWC_PENDING;
5684 	cst->flags |= EO_FLOWC_RPL_PENDING;
5685 	MPASS(cst->tx_credits >= ETID_FLOWC_LEN16);	/* flowc is first WR. */
5686 	cst->tx_credits -= ETID_FLOWC_LEN16;
5687 
5688 	return (0);
5689 }
5690 
5691 #define ETID_FLUSH_LEN16 (howmany(sizeof (struct fw_flowc_wr), 16))
5692 
5693 void
5694 send_etid_flush_wr(struct cxgbe_snd_tag *cst)
5695 {
5696 	struct fw_flowc_wr *flowc;
5697 	struct wrq_cookie cookie;
5698 
5699 	mtx_assert(&cst->lock, MA_OWNED);
5700 
5701 	flowc = start_wrq_wr(cst->eo_txq, ETID_FLUSH_LEN16, &cookie);
5702 	if (__predict_false(flowc == NULL))
5703 		CXGBE_UNIMPLEMENTED(__func__);
5704 
5705 	bzero(flowc, ETID_FLUSH_LEN16 * 16);
5706 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
5707 	    V_FW_FLOWC_WR_NPARAMS(0) | F_FW_WR_COMPL);
5708 	flowc->flowid_len16 = htobe32(V_FW_WR_LEN16(ETID_FLUSH_LEN16) |
5709 	    V_FW_WR_FLOWID(cst->etid));
5710 
5711 	commit_wrq_wr(cst->eo_txq, flowc, &cookie);
5712 
5713 	cst->flags |= EO_FLUSH_RPL_PENDING;
5714 	MPASS(cst->tx_credits >= ETID_FLUSH_LEN16);
5715 	cst->tx_credits -= ETID_FLUSH_LEN16;
5716 	cst->ncompl++;
5717 }
5718 
5719 static void
5720 write_ethofld_wr(struct cxgbe_snd_tag *cst, struct fw_eth_tx_eo_wr *wr,
5721     struct mbuf *m0, int compl)
5722 {
5723 	struct cpl_tx_pkt_core *cpl;
5724 	uint64_t ctrl1;
5725 	uint32_t ctrl;	/* used in many unrelated places */
5726 	int len16, pktlen, nsegs, immhdrs;
5727 	caddr_t dst;
5728 	uintptr_t p;
5729 	struct ulptx_sgl *usgl;
5730 	struct sglist sg;
5731 	struct sglist_seg segs[38];	/* XXX: find real limit.  XXX: get off the stack */
5732 
5733 	mtx_assert(&cst->lock, MA_OWNED);
5734 	M_ASSERTPKTHDR(m0);
5735 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
5736 	    m0->m_pkthdr.l4hlen > 0,
5737 	    ("%s: ethofld mbuf %p is missing header lengths", __func__, m0));
5738 
5739 	len16 = mbuf_eo_len16(m0);
5740 	nsegs = mbuf_eo_nsegs(m0);
5741 	pktlen = m0->m_pkthdr.len;
5742 	ctrl = sizeof(struct cpl_tx_pkt_core);
5743 	if (needs_tso(m0))
5744 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
5745 	immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen + m0->m_pkthdr.l4hlen;
5746 	ctrl += immhdrs;
5747 
5748 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_EO_WR) |
5749 	    V_FW_ETH_TX_EO_WR_IMMDLEN(ctrl) | V_FW_WR_COMPL(!!compl));
5750 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(len16) |
5751 	    V_FW_WR_FLOWID(cst->etid));
5752 	wr->r3 = 0;
5753 	if (needs_udp_csum(m0)) {
5754 		wr->u.udpseg.type = FW_ETH_TX_EO_TYPE_UDPSEG;
5755 		wr->u.udpseg.ethlen = m0->m_pkthdr.l2hlen;
5756 		wr->u.udpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
5757 		wr->u.udpseg.udplen = m0->m_pkthdr.l4hlen;
5758 		wr->u.udpseg.rtplen = 0;
5759 		wr->u.udpseg.r4 = 0;
5760 		wr->u.udpseg.mss = htobe16(pktlen - immhdrs);
5761 		wr->u.udpseg.schedpktsize = wr->u.udpseg.mss;
5762 		wr->u.udpseg.plen = htobe32(pktlen - immhdrs);
5763 		cpl = (void *)(wr + 1);
5764 	} else {
5765 		MPASS(needs_tcp_csum(m0));
5766 		wr->u.tcpseg.type = FW_ETH_TX_EO_TYPE_TCPSEG;
5767 		wr->u.tcpseg.ethlen = m0->m_pkthdr.l2hlen;
5768 		wr->u.tcpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
5769 		wr->u.tcpseg.tcplen = m0->m_pkthdr.l4hlen;
5770 		wr->u.tcpseg.tsclk_tsoff = mbuf_eo_tsclk_tsoff(m0);
5771 		wr->u.tcpseg.r4 = 0;
5772 		wr->u.tcpseg.r5 = 0;
5773 		wr->u.tcpseg.plen = htobe32(pktlen - immhdrs);
5774 
5775 		if (needs_tso(m0)) {
5776 			struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
5777 
5778 			wr->u.tcpseg.mss = htobe16(m0->m_pkthdr.tso_segsz);
5779 
5780 			ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) |
5781 			    F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE |
5782 			    V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) |
5783 			    V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
5784 			if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
5785 				ctrl |= V_LSO_ETHHDR_LEN(1);
5786 			if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
5787 				ctrl |= F_LSO_IPV6;
5788 			lso->lso_ctrl = htobe32(ctrl);
5789 			lso->ipid_ofst = htobe16(0);
5790 			lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
5791 			lso->seqno_offset = htobe32(0);
5792 			lso->len = htobe32(pktlen);
5793 
5794 			cpl = (void *)(lso + 1);
5795 		} else {
5796 			wr->u.tcpseg.mss = htobe16(0xffff);
5797 			cpl = (void *)(wr + 1);
5798 		}
5799 	}
5800 
5801 	/* Checksum offload must be requested for ethofld. */
5802 	ctrl1 = 0;
5803 	MPASS(needs_l4_csum(m0));
5804 
5805 	/* VLAN tag insertion */
5806 	if (needs_vlan_insertion(m0)) {
5807 		ctrl1 |= F_TXPKT_VLAN_VLD |
5808 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
5809 	}
5810 
5811 	/* CPL header */
5812 	cpl->ctrl0 = cst->ctrl0;
5813 	cpl->pack = 0;
5814 	cpl->len = htobe16(pktlen);
5815 	cpl->ctrl1 = htobe64(ctrl1);
5816 
5817 	/* Copy Ethernet, IP & TCP/UDP hdrs as immediate data */
5818 	p = (uintptr_t)(cpl + 1);
5819 	m_copydata(m0, 0, immhdrs, (void *)p);
5820 
5821 	/* SGL */
5822 	dst = (void *)(cpl + 1);
5823 	if (nsegs > 0) {
5824 		int i, pad;
5825 
5826 		/* zero-pad upto next 16Byte boundary, if not 16Byte aligned */
5827 		p += immhdrs;
5828 		pad = 16 - (immhdrs & 0xf);
5829 		bzero((void *)p, pad);
5830 
5831 		usgl = (void *)(p + pad);
5832 		usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
5833 		    V_ULPTX_NSGE(nsegs));
5834 
5835 		sglist_init(&sg, nitems(segs), segs);
5836 		for (; m0 != NULL; m0 = m0->m_next) {
5837 			if (__predict_false(m0->m_len == 0))
5838 				continue;
5839 			if (immhdrs >= m0->m_len) {
5840 				immhdrs -= m0->m_len;
5841 				continue;
5842 			}
5843 
5844 			sglist_append(&sg, mtod(m0, char *) + immhdrs,
5845 			    m0->m_len - immhdrs);
5846 			immhdrs = 0;
5847 		}
5848 		MPASS(sg.sg_nseg == nsegs);
5849 
5850 		/*
5851 		 * Zero pad last 8B in case the WR doesn't end on a 16B
5852 		 * boundary.
5853 		 */
5854 		*(uint64_t *)((char *)wr + len16 * 16 - 8) = 0;
5855 
5856 		usgl->len0 = htobe32(segs[0].ss_len);
5857 		usgl->addr0 = htobe64(segs[0].ss_paddr);
5858 		for (i = 0; i < nsegs - 1; i++) {
5859 			usgl->sge[i / 2].len[i & 1] = htobe32(segs[i + 1].ss_len);
5860 			usgl->sge[i / 2].addr[i & 1] = htobe64(segs[i + 1].ss_paddr);
5861 		}
5862 		if (i & 1)
5863 			usgl->sge[i / 2].len[1] = htobe32(0);
5864 	}
5865 
5866 }
5867 
5868 static void
5869 ethofld_tx(struct cxgbe_snd_tag *cst)
5870 {
5871 	struct mbuf *m;
5872 	struct wrq_cookie cookie;
5873 	int next_credits, compl;
5874 	struct fw_eth_tx_eo_wr *wr;
5875 
5876 	mtx_assert(&cst->lock, MA_OWNED);
5877 
5878 	while ((m = mbufq_first(&cst->pending_tx)) != NULL) {
5879 		M_ASSERTPKTHDR(m);
5880 
5881 		/* How many len16 credits do we need to send this mbuf. */
5882 		next_credits = mbuf_eo_len16(m);
5883 		MPASS(next_credits > 0);
5884 		if (next_credits > cst->tx_credits) {
5885 			/*
5886 			 * Tx will make progress eventually because there is at
5887 			 * least one outstanding fw4_ack that will return
5888 			 * credits and kick the tx.
5889 			 */
5890 			MPASS(cst->ncompl > 0);
5891 			return;
5892 		}
5893 		wr = start_wrq_wr(cst->eo_txq, next_credits, &cookie);
5894 		if (__predict_false(wr == NULL)) {
5895 			/* XXX: wishful thinking, not a real assertion. */
5896 			MPASS(cst->ncompl > 0);
5897 			return;
5898 		}
5899 		cst->tx_credits -= next_credits;
5900 		cst->tx_nocompl += next_credits;
5901 		compl = cst->ncompl == 0 || cst->tx_nocompl >= cst->tx_total / 2;
5902 		ETHER_BPF_MTAP(cst->com.ifp, m);
5903 		write_ethofld_wr(cst, wr, m, compl);
5904 		commit_wrq_wr(cst->eo_txq, wr, &cookie);
5905 		if (compl) {
5906 			cst->ncompl++;
5907 			cst->tx_nocompl	= 0;
5908 		}
5909 		(void) mbufq_dequeue(&cst->pending_tx);
5910 		mbufq_enqueue(&cst->pending_fwack, m);
5911 	}
5912 }
5913 
5914 int
5915 ethofld_transmit(struct ifnet *ifp, struct mbuf *m0)
5916 {
5917 	struct cxgbe_snd_tag *cst;
5918 	int rc;
5919 
5920 	MPASS(m0->m_nextpkt == NULL);
5921 	MPASS(m0->m_pkthdr.snd_tag != NULL);
5922 	cst = mst_to_cst(m0->m_pkthdr.snd_tag);
5923 
5924 	mtx_lock(&cst->lock);
5925 	MPASS(cst->flags & EO_SND_TAG_REF);
5926 
5927 	if (__predict_false(cst->flags & EO_FLOWC_PENDING)) {
5928 		struct vi_info *vi = ifp->if_softc;
5929 		struct port_info *pi = vi->pi;
5930 		struct adapter *sc = pi->adapter;
5931 		const uint32_t rss_mask = vi->rss_size - 1;
5932 		uint32_t rss_hash;
5933 
5934 		cst->eo_txq = &sc->sge.ofld_txq[vi->first_ofld_txq];
5935 		if (M_HASHTYPE_ISHASH(m0))
5936 			rss_hash = m0->m_pkthdr.flowid;
5937 		else
5938 			rss_hash = arc4random();
5939 		/* We assume RSS hashing */
5940 		cst->iqid = vi->rss[rss_hash & rss_mask];
5941 		cst->eo_txq += rss_hash % vi->nofldtxq;
5942 		rc = send_etid_flowc_wr(cst, pi, vi);
5943 		if (rc != 0)
5944 			goto done;
5945 	}
5946 
5947 	if (__predict_false(cst->plen + m0->m_pkthdr.len > eo_max_backlog)) {
5948 		rc = ENOBUFS;
5949 		goto done;
5950 	}
5951 
5952 	mbufq_enqueue(&cst->pending_tx, m0);
5953 	cst->plen += m0->m_pkthdr.len;
5954 
5955 	ethofld_tx(cst);
5956 	rc = 0;
5957 done:
5958 	mtx_unlock(&cst->lock);
5959 	if (__predict_false(rc != 0))
5960 		m_freem(m0);
5961 	return (rc);
5962 }
5963 
5964 static int
5965 ethofld_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
5966 {
5967 	struct adapter *sc = iq->adapter;
5968 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
5969 	struct mbuf *m;
5970 	u_int etid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
5971 	struct cxgbe_snd_tag *cst;
5972 	uint8_t credits = cpl->credits;
5973 
5974 	cst = lookup_etid(sc, etid);
5975 	mtx_lock(&cst->lock);
5976 	if (__predict_false(cst->flags & EO_FLOWC_RPL_PENDING)) {
5977 		MPASS(credits >= ETID_FLOWC_LEN16);
5978 		credits -= ETID_FLOWC_LEN16;
5979 		cst->flags &= ~EO_FLOWC_RPL_PENDING;
5980 	}
5981 
5982 	KASSERT(cst->ncompl > 0,
5983 	    ("%s: etid %u (%p) wasn't expecting completion.",
5984 	    __func__, etid, cst));
5985 	cst->ncompl--;
5986 
5987 	while (credits > 0) {
5988 		m = mbufq_dequeue(&cst->pending_fwack);
5989 		if (__predict_false(m == NULL)) {
5990 			/*
5991 			 * The remaining credits are for the final flush that
5992 			 * was issued when the tag was freed by the kernel.
5993 			 */
5994 			MPASS((cst->flags &
5995 			    (EO_FLUSH_RPL_PENDING | EO_SND_TAG_REF)) ==
5996 			    EO_FLUSH_RPL_PENDING);
5997 			MPASS(credits == ETID_FLUSH_LEN16);
5998 			MPASS(cst->tx_credits + cpl->credits == cst->tx_total);
5999 			MPASS(cst->ncompl == 0);
6000 
6001 			cst->flags &= ~EO_FLUSH_RPL_PENDING;
6002 			cst->tx_credits += cpl->credits;
6003 freetag:
6004 			cxgbe_snd_tag_free_locked(cst);
6005 			return (0);	/* cst is gone. */
6006 		}
6007 		KASSERT(m != NULL,
6008 		    ("%s: too many credits (%u, %u)", __func__, cpl->credits,
6009 		    credits));
6010 		KASSERT(credits >= mbuf_eo_len16(m),
6011 		    ("%s: too few credits (%u, %u, %u)", __func__,
6012 		    cpl->credits, credits, mbuf_eo_len16(m)));
6013 		credits -= mbuf_eo_len16(m);
6014 		cst->plen -= m->m_pkthdr.len;
6015 		m_freem(m);
6016 	}
6017 
6018 	cst->tx_credits += cpl->credits;
6019 	MPASS(cst->tx_credits <= cst->tx_total);
6020 
6021 	m = mbufq_first(&cst->pending_tx);
6022 	if (m != NULL && cst->tx_credits >= mbuf_eo_len16(m))
6023 		ethofld_tx(cst);
6024 
6025 	if (__predict_false((cst->flags & EO_SND_TAG_REF) == 0) &&
6026 	    cst->ncompl == 0) {
6027 		if (cst->tx_credits == cst->tx_total)
6028 			goto freetag;
6029 		else {
6030 			MPASS((cst->flags & EO_FLUSH_RPL_PENDING) == 0);
6031 			send_etid_flush_wr(cst);
6032 		}
6033 	}
6034 
6035 	mtx_unlock(&cst->lock);
6036 
6037 	return (0);
6038 }
6039 #endif
6040