xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision e430d1ed78d021db4e9760d9800393b627156745)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2011 Chelsio Communications, Inc.
5  * All rights reserved.
6  * Written by: Navdeep Parhar <np@FreeBSD.org>
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include "opt_inet.h"
34 #include "opt_inet6.h"
35 #include "opt_ratelimit.h"
36 
37 #include <sys/types.h>
38 #include <sys/eventhandler.h>
39 #include <sys/mbuf.h>
40 #include <sys/socket.h>
41 #include <sys/kernel.h>
42 #include <sys/malloc.h>
43 #include <sys/queue.h>
44 #include <sys/sbuf.h>
45 #include <sys/taskqueue.h>
46 #include <sys/time.h>
47 #include <sys/sglist.h>
48 #include <sys/sysctl.h>
49 #include <sys/smp.h>
50 #include <sys/counter.h>
51 #include <net/bpf.h>
52 #include <net/ethernet.h>
53 #include <net/if.h>
54 #include <net/if_vlan_var.h>
55 #include <netinet/in.h>
56 #include <netinet/ip.h>
57 #include <netinet/ip6.h>
58 #include <netinet/tcp.h>
59 #include <netinet/udp.h>
60 #include <machine/in_cksum.h>
61 #include <machine/md_var.h>
62 #include <vm/vm.h>
63 #include <vm/pmap.h>
64 #ifdef DEV_NETMAP
65 #include <machine/bus.h>
66 #include <sys/selinfo.h>
67 #include <net/if_var.h>
68 #include <net/netmap.h>
69 #include <dev/netmap/netmap_kern.h>
70 #endif
71 
72 #include "common/common.h"
73 #include "common/t4_regs.h"
74 #include "common/t4_regs_values.h"
75 #include "common/t4_msg.h"
76 #include "t4_l2t.h"
77 #include "t4_mp_ring.h"
78 
79 #ifdef T4_PKT_TIMESTAMP
80 #define RX_COPY_THRESHOLD (MINCLSIZE - 8)
81 #else
82 #define RX_COPY_THRESHOLD MINCLSIZE
83 #endif
84 
85 /* Internal mbuf flags stored in PH_loc.eight[1]. */
86 #define	MC_RAW_WR		0x02
87 
88 /*
89  * Ethernet frames are DMA'd at this byte offset into the freelist buffer.
90  * 0-7 are valid values.
91  */
92 static int fl_pktshift = 0;
93 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pktshift, CTLFLAG_RDTUN, &fl_pktshift, 0,
94     "payload DMA offset in rx buffer (bytes)");
95 
96 /*
97  * Pad ethernet payload up to this boundary.
98  * -1: driver should figure out a good value.
99  *  0: disable padding.
100  *  Any power of 2 from 32 to 4096 (both inclusive) is also a valid value.
101  */
102 int fl_pad = -1;
103 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pad, CTLFLAG_RDTUN, &fl_pad, 0,
104     "payload pad boundary (bytes)");
105 
106 /*
107  * Status page length.
108  * -1: driver should figure out a good value.
109  *  64 or 128 are the only other valid values.
110  */
111 static int spg_len = -1;
112 SYSCTL_INT(_hw_cxgbe, OID_AUTO, spg_len, CTLFLAG_RDTUN, &spg_len, 0,
113     "status page size (bytes)");
114 
115 /*
116  * Congestion drops.
117  * -1: no congestion feedback (not recommended).
118  *  0: backpressure the channel instead of dropping packets right away.
119  *  1: no backpressure, drop packets for the congested queue immediately.
120  */
121 static int cong_drop = 0;
122 SYSCTL_INT(_hw_cxgbe, OID_AUTO, cong_drop, CTLFLAG_RDTUN, &cong_drop, 0,
123     "Congestion control for RX queues (0 = backpressure, 1 = drop");
124 
125 /*
126  * Deliver multiple frames in the same free list buffer if they fit.
127  * -1: let the driver decide whether to enable buffer packing or not.
128  *  0: disable buffer packing.
129  *  1: enable buffer packing.
130  */
131 static int buffer_packing = -1;
132 SYSCTL_INT(_hw_cxgbe, OID_AUTO, buffer_packing, CTLFLAG_RDTUN, &buffer_packing,
133     0, "Enable buffer packing");
134 
135 /*
136  * Start next frame in a packed buffer at this boundary.
137  * -1: driver should figure out a good value.
138  * T4: driver will ignore this and use the same value as fl_pad above.
139  * T5: 16, or a power of 2 from 64 to 4096 (both inclusive) is a valid value.
140  */
141 static int fl_pack = -1;
142 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pack, CTLFLAG_RDTUN, &fl_pack, 0,
143     "payload pack boundary (bytes)");
144 
145 /*
146  * Allow the driver to create mbuf(s) in a cluster allocated for rx.
147  * 0: never; always allocate mbufs from the zone_mbuf UMA zone.
148  * 1: ok to create mbuf(s) within a cluster if there is room.
149  */
150 static int allow_mbufs_in_cluster = 1;
151 SYSCTL_INT(_hw_cxgbe, OID_AUTO, allow_mbufs_in_cluster, CTLFLAG_RDTUN,
152     &allow_mbufs_in_cluster, 0,
153     "Allow driver to create mbufs within a rx cluster");
154 
155 /*
156  * Largest rx cluster size that the driver is allowed to allocate.
157  */
158 static int largest_rx_cluster = MJUM16BYTES;
159 SYSCTL_INT(_hw_cxgbe, OID_AUTO, largest_rx_cluster, CTLFLAG_RDTUN,
160     &largest_rx_cluster, 0, "Largest rx cluster (bytes)");
161 
162 /*
163  * Size of cluster allocation that's most likely to succeed.  The driver will
164  * fall back to this size if it fails to allocate clusters larger than this.
165  */
166 static int safest_rx_cluster = PAGE_SIZE;
167 SYSCTL_INT(_hw_cxgbe, OID_AUTO, safest_rx_cluster, CTLFLAG_RDTUN,
168     &safest_rx_cluster, 0, "Safe rx cluster (bytes)");
169 
170 #ifdef RATELIMIT
171 /*
172  * Knob to control TCP timestamp rewriting, and the granularity of the tick used
173  * for rewriting.  -1 and 0-3 are all valid values.
174  * -1: hardware should leave the TCP timestamps alone.
175  * 0: 1ms
176  * 1: 100us
177  * 2: 10us
178  * 3: 1us
179  */
180 static int tsclk = -1;
181 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tsclk, CTLFLAG_RDTUN, &tsclk, 0,
182     "Control TCP timestamp rewriting when using pacing");
183 
184 static int eo_max_backlog = 1024 * 1024;
185 SYSCTL_INT(_hw_cxgbe, OID_AUTO, eo_max_backlog, CTLFLAG_RDTUN, &eo_max_backlog,
186     0, "Maximum backlog of ratelimited data per flow");
187 #endif
188 
189 /*
190  * The interrupt holdoff timers are multiplied by this value on T6+.
191  * 1 and 3-17 (both inclusive) are legal values.
192  */
193 static int tscale = 1;
194 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tscale, CTLFLAG_RDTUN, &tscale, 0,
195     "Interrupt holdoff timer scale on T6+");
196 
197 /*
198  * Number of LRO entries in the lro_ctrl structure per rx queue.
199  */
200 static int lro_entries = TCP_LRO_ENTRIES;
201 SYSCTL_INT(_hw_cxgbe, OID_AUTO, lro_entries, CTLFLAG_RDTUN, &lro_entries, 0,
202     "Number of LRO entries per RX queue");
203 
204 /*
205  * This enables presorting of frames before they're fed into tcp_lro_rx.
206  */
207 static int lro_mbufs = 0;
208 SYSCTL_INT(_hw_cxgbe, OID_AUTO, lro_mbufs, CTLFLAG_RDTUN, &lro_mbufs, 0,
209     "Enable presorting of LRO frames");
210 
211 struct txpkts {
212 	u_int wr_type;		/* type 0 or type 1 */
213 	u_int npkt;		/* # of packets in this work request */
214 	u_int plen;		/* total payload (sum of all packets) */
215 	u_int len16;		/* # of 16B pieces used by this work request */
216 };
217 
218 /* A packet's SGL.  This + m_pkthdr has all info needed for tx */
219 struct sgl {
220 	struct sglist sg;
221 	struct sglist_seg seg[TX_SGL_SEGS];
222 };
223 
224 static int service_iq(struct sge_iq *, int);
225 static int service_iq_fl(struct sge_iq *, int);
226 static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t);
227 static int t4_eth_rx(struct sge_iq *, const struct rss_header *, struct mbuf *);
228 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int);
229 static inline void init_fl(struct adapter *, struct sge_fl *, int, int, char *);
230 static inline void init_eq(struct adapter *, struct sge_eq *, int, int, uint8_t,
231     uint16_t, char *);
232 static int alloc_ring(struct adapter *, size_t, bus_dma_tag_t *, bus_dmamap_t *,
233     bus_addr_t *, void **);
234 static int free_ring(struct adapter *, bus_dma_tag_t, bus_dmamap_t, bus_addr_t,
235     void *);
236 static int alloc_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *,
237     int, int);
238 static int free_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *);
239 static void add_iq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
240     struct sge_iq *);
241 static void add_fl_sysctls(struct adapter *, struct sysctl_ctx_list *,
242     struct sysctl_oid *, struct sge_fl *);
243 static int alloc_fwq(struct adapter *);
244 static int free_fwq(struct adapter *);
245 static int alloc_ctrlq(struct adapter *, struct sge_wrq *, int,
246     struct sysctl_oid *);
247 static int alloc_rxq(struct vi_info *, struct sge_rxq *, int, int,
248     struct sysctl_oid *);
249 static int free_rxq(struct vi_info *, struct sge_rxq *);
250 #ifdef TCP_OFFLOAD
251 static int alloc_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *, int, int,
252     struct sysctl_oid *);
253 static int free_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *);
254 #endif
255 #ifdef DEV_NETMAP
256 static int alloc_nm_rxq(struct vi_info *, struct sge_nm_rxq *, int, int,
257     struct sysctl_oid *);
258 static int free_nm_rxq(struct vi_info *, struct sge_nm_rxq *);
259 static int alloc_nm_txq(struct vi_info *, struct sge_nm_txq *, int, int,
260     struct sysctl_oid *);
261 static int free_nm_txq(struct vi_info *, struct sge_nm_txq *);
262 #endif
263 static int ctrl_eq_alloc(struct adapter *, struct sge_eq *);
264 static int eth_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
265 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
266 static int ofld_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
267 #endif
268 static int alloc_eq(struct adapter *, struct vi_info *, struct sge_eq *);
269 static int free_eq(struct adapter *, struct sge_eq *);
270 static int alloc_wrq(struct adapter *, struct vi_info *, struct sge_wrq *,
271     struct sysctl_oid *);
272 static int free_wrq(struct adapter *, struct sge_wrq *);
273 static int alloc_txq(struct vi_info *, struct sge_txq *, int,
274     struct sysctl_oid *);
275 static int free_txq(struct vi_info *, struct sge_txq *);
276 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int);
277 static inline void ring_fl_db(struct adapter *, struct sge_fl *);
278 static int refill_fl(struct adapter *, struct sge_fl *, int);
279 static void refill_sfl(void *);
280 static int alloc_fl_sdesc(struct sge_fl *);
281 static void free_fl_sdesc(struct adapter *, struct sge_fl *);
282 static void find_best_refill_source(struct adapter *, struct sge_fl *, int);
283 static void find_safe_refill_source(struct adapter *, struct sge_fl *);
284 static void add_fl_to_sfl(struct adapter *, struct sge_fl *);
285 
286 static inline void get_pkt_gl(struct mbuf *, struct sglist *);
287 static inline u_int txpkt_len16(u_int, u_int);
288 static inline u_int txpkt_vm_len16(u_int, u_int);
289 static inline u_int txpkts0_len16(u_int);
290 static inline u_int txpkts1_len16(void);
291 static u_int write_raw_wr(struct sge_txq *, void *, struct mbuf *, u_int);
292 static u_int write_txpkt_wr(struct sge_txq *, struct fw_eth_tx_pkt_wr *,
293     struct mbuf *, u_int);
294 static u_int write_txpkt_vm_wr(struct adapter *, struct sge_txq *,
295     struct fw_eth_tx_pkt_vm_wr *, struct mbuf *, u_int);
296 static int try_txpkts(struct mbuf *, struct mbuf *, struct txpkts *, u_int);
297 static int add_to_txpkts(struct mbuf *, struct txpkts *, u_int);
298 static u_int write_txpkts_wr(struct sge_txq *, struct fw_eth_tx_pkts_wr *,
299     struct mbuf *, const struct txpkts *, u_int);
300 static void write_gl_to_txd(struct sge_txq *, struct mbuf *, caddr_t *, int);
301 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int);
302 static inline void ring_eq_db(struct adapter *, struct sge_eq *, u_int);
303 static inline uint16_t read_hw_cidx(struct sge_eq *);
304 static inline u_int reclaimable_tx_desc(struct sge_eq *);
305 static inline u_int total_available_tx_desc(struct sge_eq *);
306 static u_int reclaim_tx_descs(struct sge_txq *, u_int);
307 static void tx_reclaim(void *, int);
308 static __be64 get_flit(struct sglist_seg *, int, int);
309 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
310     struct mbuf *);
311 static int handle_fw_msg(struct sge_iq *, const struct rss_header *,
312     struct mbuf *);
313 static int t4_handle_wrerr_rpl(struct adapter *, const __be64 *);
314 static void wrq_tx_drain(void *, int);
315 static void drain_wrq_wr_list(struct adapter *, struct sge_wrq *);
316 
317 static int sysctl_uint16(SYSCTL_HANDLER_ARGS);
318 static int sysctl_bufsizes(SYSCTL_HANDLER_ARGS);
319 #ifdef RATELIMIT
320 static inline u_int txpkt_eo_len16(u_int, u_int, u_int);
321 static int ethofld_fw4_ack(struct sge_iq *, const struct rss_header *,
322     struct mbuf *);
323 #endif
324 
325 static counter_u64_t extfree_refs;
326 static counter_u64_t extfree_rels;
327 
328 an_handler_t t4_an_handler;
329 fw_msg_handler_t t4_fw_msg_handler[NUM_FW6_TYPES];
330 cpl_handler_t t4_cpl_handler[NUM_CPL_CMDS];
331 cpl_handler_t set_tcb_rpl_handlers[NUM_CPL_COOKIES];
332 cpl_handler_t l2t_write_rpl_handlers[NUM_CPL_COOKIES];
333 cpl_handler_t act_open_rpl_handlers[NUM_CPL_COOKIES];
334 cpl_handler_t abort_rpl_rss_handlers[NUM_CPL_COOKIES];
335 cpl_handler_t fw4_ack_handlers[NUM_CPL_COOKIES];
336 
337 void
338 t4_register_an_handler(an_handler_t h)
339 {
340 	uintptr_t *loc;
341 
342 	MPASS(h == NULL || t4_an_handler == NULL);
343 
344 	loc = (uintptr_t *)&t4_an_handler;
345 	atomic_store_rel_ptr(loc, (uintptr_t)h);
346 }
347 
348 void
349 t4_register_fw_msg_handler(int type, fw_msg_handler_t h)
350 {
351 	uintptr_t *loc;
352 
353 	MPASS(type < nitems(t4_fw_msg_handler));
354 	MPASS(h == NULL || t4_fw_msg_handler[type] == NULL);
355 	/*
356 	 * These are dispatched by the handler for FW{4|6}_CPL_MSG using the CPL
357 	 * handler dispatch table.  Reject any attempt to install a handler for
358 	 * this subtype.
359 	 */
360 	MPASS(type != FW_TYPE_RSSCPL);
361 	MPASS(type != FW6_TYPE_RSSCPL);
362 
363 	loc = (uintptr_t *)&t4_fw_msg_handler[type];
364 	atomic_store_rel_ptr(loc, (uintptr_t)h);
365 }
366 
367 void
368 t4_register_cpl_handler(int opcode, cpl_handler_t h)
369 {
370 	uintptr_t *loc;
371 
372 	MPASS(opcode < nitems(t4_cpl_handler));
373 	MPASS(h == NULL || t4_cpl_handler[opcode] == NULL);
374 
375 	loc = (uintptr_t *)&t4_cpl_handler[opcode];
376 	atomic_store_rel_ptr(loc, (uintptr_t)h);
377 }
378 
379 static int
380 set_tcb_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
381     struct mbuf *m)
382 {
383 	const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
384 	u_int tid;
385 	int cookie;
386 
387 	MPASS(m == NULL);
388 
389 	tid = GET_TID(cpl);
390 	if (is_hpftid(iq->adapter, tid) || is_ftid(iq->adapter, tid)) {
391 		/*
392 		 * The return code for filter-write is put in the CPL cookie so
393 		 * we have to rely on the hardware tid (is_ftid) to determine
394 		 * that this is a response to a filter.
395 		 */
396 		cookie = CPL_COOKIE_FILTER;
397 	} else {
398 		cookie = G_COOKIE(cpl->cookie);
399 	}
400 	MPASS(cookie > CPL_COOKIE_RESERVED);
401 	MPASS(cookie < nitems(set_tcb_rpl_handlers));
402 
403 	return (set_tcb_rpl_handlers[cookie](iq, rss, m));
404 }
405 
406 static int
407 l2t_write_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
408     struct mbuf *m)
409 {
410 	const struct cpl_l2t_write_rpl *rpl = (const void *)(rss + 1);
411 	unsigned int cookie;
412 
413 	MPASS(m == NULL);
414 
415 	cookie = GET_TID(rpl) & F_SYNC_WR ? CPL_COOKIE_TOM : CPL_COOKIE_FILTER;
416 	return (l2t_write_rpl_handlers[cookie](iq, rss, m));
417 }
418 
419 static int
420 act_open_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
421     struct mbuf *m)
422 {
423 	const struct cpl_act_open_rpl *cpl = (const void *)(rss + 1);
424 	u_int cookie = G_TID_COOKIE(G_AOPEN_ATID(be32toh(cpl->atid_status)));
425 
426 	MPASS(m == NULL);
427 	MPASS(cookie != CPL_COOKIE_RESERVED);
428 
429 	return (act_open_rpl_handlers[cookie](iq, rss, m));
430 }
431 
432 static int
433 abort_rpl_rss_handler(struct sge_iq *iq, const struct rss_header *rss,
434     struct mbuf *m)
435 {
436 	struct adapter *sc = iq->adapter;
437 	u_int cookie;
438 
439 	MPASS(m == NULL);
440 	if (is_hashfilter(sc))
441 		cookie = CPL_COOKIE_HASHFILTER;
442 	else
443 		cookie = CPL_COOKIE_TOM;
444 
445 	return (abort_rpl_rss_handlers[cookie](iq, rss, m));
446 }
447 
448 static int
449 fw4_ack_handler(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
450 {
451 	struct adapter *sc = iq->adapter;
452 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
453 	unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
454 	u_int cookie;
455 
456 	MPASS(m == NULL);
457 	if (is_etid(sc, tid))
458 		cookie = CPL_COOKIE_ETHOFLD;
459 	else
460 		cookie = CPL_COOKIE_TOM;
461 
462 	return (fw4_ack_handlers[cookie](iq, rss, m));
463 }
464 
465 static void
466 t4_init_shared_cpl_handlers(void)
467 {
468 
469 	t4_register_cpl_handler(CPL_SET_TCB_RPL, set_tcb_rpl_handler);
470 	t4_register_cpl_handler(CPL_L2T_WRITE_RPL, l2t_write_rpl_handler);
471 	t4_register_cpl_handler(CPL_ACT_OPEN_RPL, act_open_rpl_handler);
472 	t4_register_cpl_handler(CPL_ABORT_RPL_RSS, abort_rpl_rss_handler);
473 	t4_register_cpl_handler(CPL_FW4_ACK, fw4_ack_handler);
474 }
475 
476 void
477 t4_register_shared_cpl_handler(int opcode, cpl_handler_t h, int cookie)
478 {
479 	uintptr_t *loc;
480 
481 	MPASS(opcode < nitems(t4_cpl_handler));
482 	MPASS(cookie > CPL_COOKIE_RESERVED);
483 	MPASS(cookie < NUM_CPL_COOKIES);
484 	MPASS(t4_cpl_handler[opcode] != NULL);
485 
486 	switch (opcode) {
487 	case CPL_SET_TCB_RPL:
488 		loc = (uintptr_t *)&set_tcb_rpl_handlers[cookie];
489 		break;
490 	case CPL_L2T_WRITE_RPL:
491 		loc = (uintptr_t *)&l2t_write_rpl_handlers[cookie];
492 		break;
493 	case CPL_ACT_OPEN_RPL:
494 		loc = (uintptr_t *)&act_open_rpl_handlers[cookie];
495 		break;
496 	case CPL_ABORT_RPL_RSS:
497 		loc = (uintptr_t *)&abort_rpl_rss_handlers[cookie];
498 		break;
499 	case CPL_FW4_ACK:
500 		loc = (uintptr_t *)&fw4_ack_handlers[cookie];
501 		break;
502 	default:
503 		MPASS(0);
504 		return;
505 	}
506 	MPASS(h == NULL || *loc == (uintptr_t)NULL);
507 	atomic_store_rel_ptr(loc, (uintptr_t)h);
508 }
509 
510 /*
511  * Called on MOD_LOAD.  Validates and calculates the SGE tunables.
512  */
513 void
514 t4_sge_modload(void)
515 {
516 
517 	if (fl_pktshift < 0 || fl_pktshift > 7) {
518 		printf("Invalid hw.cxgbe.fl_pktshift value (%d),"
519 		    " using 0 instead.\n", fl_pktshift);
520 		fl_pktshift = 0;
521 	}
522 
523 	if (spg_len != 64 && spg_len != 128) {
524 		int len;
525 
526 #if defined(__i386__) || defined(__amd64__)
527 		len = cpu_clflush_line_size > 64 ? 128 : 64;
528 #else
529 		len = 64;
530 #endif
531 		if (spg_len != -1) {
532 			printf("Invalid hw.cxgbe.spg_len value (%d),"
533 			    " using %d instead.\n", spg_len, len);
534 		}
535 		spg_len = len;
536 	}
537 
538 	if (cong_drop < -1 || cong_drop > 1) {
539 		printf("Invalid hw.cxgbe.cong_drop value (%d),"
540 		    " using 0 instead.\n", cong_drop);
541 		cong_drop = 0;
542 	}
543 
544 	if (tscale != 1 && (tscale < 3 || tscale > 17)) {
545 		printf("Invalid hw.cxgbe.tscale value (%d),"
546 		    " using 1 instead.\n", tscale);
547 		tscale = 1;
548 	}
549 
550 	extfree_refs = counter_u64_alloc(M_WAITOK);
551 	extfree_rels = counter_u64_alloc(M_WAITOK);
552 	counter_u64_zero(extfree_refs);
553 	counter_u64_zero(extfree_rels);
554 
555 	t4_init_shared_cpl_handlers();
556 	t4_register_cpl_handler(CPL_FW4_MSG, handle_fw_msg);
557 	t4_register_cpl_handler(CPL_FW6_MSG, handle_fw_msg);
558 	t4_register_cpl_handler(CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
559 	t4_register_cpl_handler(CPL_RX_PKT, t4_eth_rx);
560 #ifdef RATELIMIT
561 	t4_register_shared_cpl_handler(CPL_FW4_ACK, ethofld_fw4_ack,
562 	    CPL_COOKIE_ETHOFLD);
563 #endif
564 	t4_register_fw_msg_handler(FW6_TYPE_CMD_RPL, t4_handle_fw_rpl);
565 	t4_register_fw_msg_handler(FW6_TYPE_WRERR_RPL, t4_handle_wrerr_rpl);
566 }
567 
568 void
569 t4_sge_modunload(void)
570 {
571 
572 	counter_u64_free(extfree_refs);
573 	counter_u64_free(extfree_rels);
574 }
575 
576 uint64_t
577 t4_sge_extfree_refs(void)
578 {
579 	uint64_t refs, rels;
580 
581 	rels = counter_u64_fetch(extfree_rels);
582 	refs = counter_u64_fetch(extfree_refs);
583 
584 	return (refs - rels);
585 }
586 
587 static inline void
588 setup_pad_and_pack_boundaries(struct adapter *sc)
589 {
590 	uint32_t v, m;
591 	int pad, pack, pad_shift;
592 
593 	pad_shift = chip_id(sc) > CHELSIO_T5 ? X_T6_INGPADBOUNDARY_SHIFT :
594 	    X_INGPADBOUNDARY_SHIFT;
595 	pad = fl_pad;
596 	if (fl_pad < (1 << pad_shift) ||
597 	    fl_pad > (1 << (pad_shift + M_INGPADBOUNDARY)) ||
598 	    !powerof2(fl_pad)) {
599 		/*
600 		 * If there is any chance that we might use buffer packing and
601 		 * the chip is a T4, then pick 64 as the pad/pack boundary.  Set
602 		 * it to the minimum allowed in all other cases.
603 		 */
604 		pad = is_t4(sc) && buffer_packing ? 64 : 1 << pad_shift;
605 
606 		/*
607 		 * For fl_pad = 0 we'll still write a reasonable value to the
608 		 * register but all the freelists will opt out of padding.
609 		 * We'll complain here only if the user tried to set it to a
610 		 * value greater than 0 that was invalid.
611 		 */
612 		if (fl_pad > 0) {
613 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pad value"
614 			    " (%d), using %d instead.\n", fl_pad, pad);
615 		}
616 	}
617 	m = V_INGPADBOUNDARY(M_INGPADBOUNDARY);
618 	v = V_INGPADBOUNDARY(ilog2(pad) - pad_shift);
619 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
620 
621 	if (is_t4(sc)) {
622 		if (fl_pack != -1 && fl_pack != pad) {
623 			/* Complain but carry on. */
624 			device_printf(sc->dev, "hw.cxgbe.fl_pack (%d) ignored,"
625 			    " using %d instead.\n", fl_pack, pad);
626 		}
627 		return;
628 	}
629 
630 	pack = fl_pack;
631 	if (fl_pack < 16 || fl_pack == 32 || fl_pack > 4096 ||
632 	    !powerof2(fl_pack)) {
633 		pack = max(sc->params.pci.mps, CACHE_LINE_SIZE);
634 		MPASS(powerof2(pack));
635 		if (pack < 16)
636 			pack = 16;
637 		if (pack == 32)
638 			pack = 64;
639 		if (pack > 4096)
640 			pack = 4096;
641 		if (fl_pack != -1) {
642 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pack value"
643 			    " (%d), using %d instead.\n", fl_pack, pack);
644 		}
645 	}
646 	m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY);
647 	if (pack == 16)
648 		v = V_INGPACKBOUNDARY(0);
649 	else
650 		v = V_INGPACKBOUNDARY(ilog2(pack) - 5);
651 
652 	MPASS(!is_t4(sc));	/* T4 doesn't have SGE_CONTROL2 */
653 	t4_set_reg_field(sc, A_SGE_CONTROL2, m, v);
654 }
655 
656 /*
657  * adap->params.vpd.cclk must be set up before this is called.
658  */
659 void
660 t4_tweak_chip_settings(struct adapter *sc)
661 {
662 	int i;
663 	uint32_t v, m;
664 	int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200};
665 	int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk;
666 	int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */
667 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
668 	static int sge_flbuf_sizes[] = {
669 		MCLBYTES,
670 #if MJUMPAGESIZE != MCLBYTES
671 		MJUMPAGESIZE,
672 		MJUMPAGESIZE - CL_METADATA_SIZE,
673 		MJUMPAGESIZE - 2 * MSIZE - CL_METADATA_SIZE,
674 #endif
675 		MJUM9BYTES,
676 		MJUM16BYTES,
677 		MCLBYTES - MSIZE - CL_METADATA_SIZE,
678 		MJUM9BYTES - CL_METADATA_SIZE,
679 		MJUM16BYTES - CL_METADATA_SIZE,
680 	};
681 
682 	KASSERT(sc->flags & MASTER_PF,
683 	    ("%s: trying to change chip settings when not master.", __func__));
684 
685 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
686 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
687 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
688 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
689 
690 	setup_pad_and_pack_boundaries(sc);
691 
692 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
693 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
694 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
695 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
696 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
697 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
698 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
699 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
700 	t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v);
701 
702 	KASSERT(nitems(sge_flbuf_sizes) <= SGE_FLBUF_SIZES,
703 	    ("%s: hw buffer size table too big", __func__));
704 	t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0, 4096);
705 	t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE1, 65536);
706 	for (i = 0; i < min(nitems(sge_flbuf_sizes), SGE_FLBUF_SIZES); i++) {
707 		t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE15 - (4 * i),
708 		    sge_flbuf_sizes[i]);
709 	}
710 
711 	v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) |
712 	    V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]);
713 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v);
714 
715 	KASSERT(intr_timer[0] <= timer_max,
716 	    ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0],
717 	    timer_max));
718 	for (i = 1; i < nitems(intr_timer); i++) {
719 		KASSERT(intr_timer[i] >= intr_timer[i - 1],
720 		    ("%s: timers not listed in increasing order (%d)",
721 		    __func__, i));
722 
723 		while (intr_timer[i] > timer_max) {
724 			if (i == nitems(intr_timer) - 1) {
725 				intr_timer[i] = timer_max;
726 				break;
727 			}
728 			intr_timer[i] += intr_timer[i - 1];
729 			intr_timer[i] /= 2;
730 		}
731 	}
732 
733 	v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) |
734 	    V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1]));
735 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v);
736 	v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) |
737 	    V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3]));
738 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v);
739 	v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) |
740 	    V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5]));
741 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v);
742 
743 	if (chip_id(sc) >= CHELSIO_T6) {
744 		m = V_TSCALE(M_TSCALE);
745 		if (tscale == 1)
746 			v = 0;
747 		else
748 			v = V_TSCALE(tscale - 2);
749 		t4_set_reg_field(sc, A_SGE_ITP_CONTROL, m, v);
750 
751 		if (sc->debug_flags & DF_DISABLE_TCB_CACHE) {
752 			m = V_RDTHRESHOLD(M_RDTHRESHOLD) | F_WRTHRTHRESHEN |
753 			    V_WRTHRTHRESH(M_WRTHRTHRESH);
754 			t4_tp_pio_read(sc, &v, 1, A_TP_CMM_CONFIG, 1);
755 			v &= ~m;
756 			v |= V_RDTHRESHOLD(1) | F_WRTHRTHRESHEN |
757 			    V_WRTHRTHRESH(16);
758 			t4_tp_pio_write(sc, &v, 1, A_TP_CMM_CONFIG, 1);
759 		}
760 	}
761 
762 	/* 4K, 16K, 64K, 256K DDP "page sizes" for TDDP */
763 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
764 	t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v);
765 
766 	/*
767 	 * 4K, 8K, 16K, 64K DDP "page sizes" for iSCSI DDP.  These have been
768 	 * chosen with MAXPHYS = 128K in mind.  The largest DDP buffer that we
769 	 * may have to deal with is MAXPHYS + 1 page.
770 	 */
771 	v = V_HPZ0(0) | V_HPZ1(1) | V_HPZ2(2) | V_HPZ3(4);
772 	t4_write_reg(sc, A_ULP_RX_ISCSI_PSZ, v);
773 
774 	/* We use multiple DDP page sizes both in plain-TOE and ISCSI modes. */
775 	m = v = F_TDDPTAGTCB | F_ISCSITAGTCB;
776 	t4_set_reg_field(sc, A_ULP_RX_CTL, m, v);
777 
778 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
779 	    F_RESETDDPOFFSET;
780 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
781 	t4_set_reg_field(sc, A_TP_PARA_REG5, m, v);
782 }
783 
784 /*
785  * SGE wants the buffer to be at least 64B and then a multiple of 16.  If
786  * padding is in use, the buffer's start and end need to be aligned to the pad
787  * boundary as well.  We'll just make sure that the size is a multiple of the
788  * boundary here, it is up to the buffer allocation code to make sure the start
789  * of the buffer is aligned as well.
790  */
791 static inline int
792 hwsz_ok(struct adapter *sc, int hwsz)
793 {
794 	int mask = fl_pad ? sc->params.sge.pad_boundary - 1 : 16 - 1;
795 
796 	return (hwsz >= 64 && (hwsz & mask) == 0);
797 }
798 
799 /*
800  * XXX: driver really should be able to deal with unexpected settings.
801  */
802 int
803 t4_read_chip_settings(struct adapter *sc)
804 {
805 	struct sge *s = &sc->sge;
806 	struct sge_params *sp = &sc->params.sge;
807 	int i, j, n, rc = 0;
808 	uint32_t m, v, r;
809 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
810 	static int sw_buf_sizes[] = {	/* Sorted by size */
811 		MCLBYTES,
812 #if MJUMPAGESIZE != MCLBYTES
813 		MJUMPAGESIZE,
814 #endif
815 		MJUM9BYTES,
816 		MJUM16BYTES
817 	};
818 	struct sw_zone_info *swz, *safe_swz;
819 	struct hw_buf_info *hwb;
820 
821 	m = F_RXPKTCPLMODE;
822 	v = F_RXPKTCPLMODE;
823 	r = sc->params.sge.sge_control;
824 	if ((r & m) != v) {
825 		device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r);
826 		rc = EINVAL;
827 	}
828 
829 	/*
830 	 * If this changes then every single use of PAGE_SHIFT in the driver
831 	 * needs to be carefully reviewed for PAGE_SHIFT vs sp->page_shift.
832 	 */
833 	if (sp->page_shift != PAGE_SHIFT) {
834 		device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r);
835 		rc = EINVAL;
836 	}
837 
838 	/* Filter out unusable hw buffer sizes entirely (mark with -2). */
839 	hwb = &s->hw_buf_info[0];
840 	for (i = 0; i < nitems(s->hw_buf_info); i++, hwb++) {
841 		r = sc->params.sge.sge_fl_buffer_size[i];
842 		hwb->size = r;
843 		hwb->zidx = hwsz_ok(sc, r) ? -1 : -2;
844 		hwb->next = -1;
845 	}
846 
847 	/*
848 	 * Create a sorted list in decreasing order of hw buffer sizes (and so
849 	 * increasing order of spare area) for each software zone.
850 	 *
851 	 * If padding is enabled then the start and end of the buffer must align
852 	 * to the pad boundary; if packing is enabled then they must align with
853 	 * the pack boundary as well.  Allocations from the cluster zones are
854 	 * aligned to min(size, 4K), so the buffer starts at that alignment and
855 	 * ends at hwb->size alignment.  If mbuf inlining is allowed the
856 	 * starting alignment will be reduced to MSIZE and the driver will
857 	 * exercise appropriate caution when deciding on the best buffer layout
858 	 * to use.
859 	 */
860 	n = 0;	/* no usable buffer size to begin with */
861 	swz = &s->sw_zone_info[0];
862 	safe_swz = NULL;
863 	for (i = 0; i < SW_ZONE_SIZES; i++, swz++) {
864 		int8_t head = -1, tail = -1;
865 
866 		swz->size = sw_buf_sizes[i];
867 		swz->zone = m_getzone(swz->size);
868 		swz->type = m_gettype(swz->size);
869 
870 		if (swz->size < PAGE_SIZE) {
871 			MPASS(powerof2(swz->size));
872 			if (fl_pad && (swz->size % sp->pad_boundary != 0))
873 				continue;
874 		}
875 
876 		if (swz->size == safest_rx_cluster)
877 			safe_swz = swz;
878 
879 		hwb = &s->hw_buf_info[0];
880 		for (j = 0; j < SGE_FLBUF_SIZES; j++, hwb++) {
881 			if (hwb->zidx != -1 || hwb->size > swz->size)
882 				continue;
883 #ifdef INVARIANTS
884 			if (fl_pad)
885 				MPASS(hwb->size % sp->pad_boundary == 0);
886 #endif
887 			hwb->zidx = i;
888 			if (head == -1)
889 				head = tail = j;
890 			else if (hwb->size < s->hw_buf_info[tail].size) {
891 				s->hw_buf_info[tail].next = j;
892 				tail = j;
893 			} else {
894 				int8_t *cur;
895 				struct hw_buf_info *t;
896 
897 				for (cur = &head; *cur != -1; cur = &t->next) {
898 					t = &s->hw_buf_info[*cur];
899 					if (hwb->size == t->size) {
900 						hwb->zidx = -2;
901 						break;
902 					}
903 					if (hwb->size > t->size) {
904 						hwb->next = *cur;
905 						*cur = j;
906 						break;
907 					}
908 				}
909 			}
910 		}
911 		swz->head_hwidx = head;
912 		swz->tail_hwidx = tail;
913 
914 		if (tail != -1) {
915 			n++;
916 			if (swz->size - s->hw_buf_info[tail].size >=
917 			    CL_METADATA_SIZE)
918 				sc->flags |= BUF_PACKING_OK;
919 		}
920 	}
921 	if (n == 0) {
922 		device_printf(sc->dev, "no usable SGE FL buffer size.\n");
923 		rc = EINVAL;
924 	}
925 
926 	s->safe_hwidx1 = -1;
927 	s->safe_hwidx2 = -1;
928 	if (safe_swz != NULL) {
929 		s->safe_hwidx1 = safe_swz->head_hwidx;
930 		for (i = safe_swz->head_hwidx; i != -1; i = hwb->next) {
931 			int spare;
932 
933 			hwb = &s->hw_buf_info[i];
934 #ifdef INVARIANTS
935 			if (fl_pad)
936 				MPASS(hwb->size % sp->pad_boundary == 0);
937 #endif
938 			spare = safe_swz->size - hwb->size;
939 			if (spare >= CL_METADATA_SIZE) {
940 				s->safe_hwidx2 = i;
941 				break;
942 			}
943 		}
944 	}
945 
946 	if (sc->flags & IS_VF)
947 		return (0);
948 
949 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
950 	r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ);
951 	if (r != v) {
952 		device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r);
953 		rc = EINVAL;
954 	}
955 
956 	m = v = F_TDDPTAGTCB;
957 	r = t4_read_reg(sc, A_ULP_RX_CTL);
958 	if ((r & m) != v) {
959 		device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r);
960 		rc = EINVAL;
961 	}
962 
963 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
964 	    F_RESETDDPOFFSET;
965 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
966 	r = t4_read_reg(sc, A_TP_PARA_REG5);
967 	if ((r & m) != v) {
968 		device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r);
969 		rc = EINVAL;
970 	}
971 
972 	t4_init_tp_params(sc, 1);
973 
974 	t4_read_mtu_tbl(sc, sc->params.mtus, NULL);
975 	t4_load_mtus(sc, sc->params.mtus, sc->params.a_wnd, sc->params.b_wnd);
976 
977 	return (rc);
978 }
979 
980 int
981 t4_create_dma_tag(struct adapter *sc)
982 {
983 	int rc;
984 
985 	rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
986 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
987 	    BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
988 	    NULL, &sc->dmat);
989 	if (rc != 0) {
990 		device_printf(sc->dev,
991 		    "failed to create main DMA tag: %d\n", rc);
992 	}
993 
994 	return (rc);
995 }
996 
997 void
998 t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
999     struct sysctl_oid_list *children)
1000 {
1001 	struct sge_params *sp = &sc->params.sge;
1002 
1003 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "buffer_sizes",
1004 	    CTLTYPE_STRING | CTLFLAG_RD, &sc->sge, 0, sysctl_bufsizes, "A",
1005 	    "freelist buffer sizes");
1006 
1007 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD,
1008 	    NULL, sp->fl_pktshift, "payload DMA offset in rx buffer (bytes)");
1009 
1010 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD,
1011 	    NULL, sp->pad_boundary, "payload pad boundary (bytes)");
1012 
1013 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD,
1014 	    NULL, sp->spg_len, "status page size (bytes)");
1015 
1016 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD,
1017 	    NULL, cong_drop, "congestion drop setting");
1018 
1019 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD,
1020 	    NULL, sp->pack_boundary, "payload pack boundary (bytes)");
1021 }
1022 
1023 int
1024 t4_destroy_dma_tag(struct adapter *sc)
1025 {
1026 	if (sc->dmat)
1027 		bus_dma_tag_destroy(sc->dmat);
1028 
1029 	return (0);
1030 }
1031 
1032 /*
1033  * Allocate and initialize the firmware event queue, control queues, and special
1034  * purpose rx queues owned by the adapter.
1035  *
1036  * Returns errno on failure.  Resources allocated up to that point may still be
1037  * allocated.  Caller is responsible for cleanup in case this function fails.
1038  */
1039 int
1040 t4_setup_adapter_queues(struct adapter *sc)
1041 {
1042 	struct sysctl_oid *oid;
1043 	struct sysctl_oid_list *children;
1044 	int rc, i;
1045 
1046 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
1047 
1048 	sysctl_ctx_init(&sc->ctx);
1049 	sc->flags |= ADAP_SYSCTL_CTX;
1050 
1051 	/*
1052 	 * Firmware event queue
1053 	 */
1054 	rc = alloc_fwq(sc);
1055 	if (rc != 0)
1056 		return (rc);
1057 
1058 	/*
1059 	 * That's all for the VF driver.
1060 	 */
1061 	if (sc->flags & IS_VF)
1062 		return (rc);
1063 
1064 	oid = device_get_sysctl_tree(sc->dev);
1065 	children = SYSCTL_CHILDREN(oid);
1066 
1067 	/*
1068 	 * XXX: General purpose rx queues, one per port.
1069 	 */
1070 
1071 	/*
1072 	 * Control queues, one per port.
1073 	 */
1074 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "ctrlq",
1075 	    CTLFLAG_RD, NULL, "control queues");
1076 	for_each_port(sc, i) {
1077 		struct sge_wrq *ctrlq = &sc->sge.ctrlq[i];
1078 
1079 		rc = alloc_ctrlq(sc, ctrlq, i, oid);
1080 		if (rc != 0)
1081 			return (rc);
1082 	}
1083 
1084 	return (rc);
1085 }
1086 
1087 /*
1088  * Idempotent
1089  */
1090 int
1091 t4_teardown_adapter_queues(struct adapter *sc)
1092 {
1093 	int i;
1094 
1095 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
1096 
1097 	/* Do this before freeing the queue */
1098 	if (sc->flags & ADAP_SYSCTL_CTX) {
1099 		sysctl_ctx_free(&sc->ctx);
1100 		sc->flags &= ~ADAP_SYSCTL_CTX;
1101 	}
1102 
1103 	if (!(sc->flags & IS_VF)) {
1104 		for_each_port(sc, i)
1105 			free_wrq(sc, &sc->sge.ctrlq[i]);
1106 	}
1107 	free_fwq(sc);
1108 
1109 	return (0);
1110 }
1111 
1112 /* Maximum payload that can be delivered with a single iq descriptor */
1113 static inline int
1114 mtu_to_max_payload(struct adapter *sc, int mtu, const int toe)
1115 {
1116 	int payload;
1117 
1118 #ifdef TCP_OFFLOAD
1119 	if (toe) {
1120 		int rxcs = G_RXCOALESCESIZE(t4_read_reg(sc, A_TP_PARA_REG2));
1121 
1122 		/* Note that COP can set rx_coalesce on/off per connection. */
1123 		payload = max(mtu, rxcs);
1124 	} else {
1125 #endif
1126 		/* large enough even when hw VLAN extraction is disabled */
1127 		payload = sc->params.sge.fl_pktshift + ETHER_HDR_LEN +
1128 		    ETHER_VLAN_ENCAP_LEN + mtu;
1129 #ifdef TCP_OFFLOAD
1130 	}
1131 #endif
1132 
1133 	return (payload);
1134 }
1135 
1136 int
1137 t4_setup_vi_queues(struct vi_info *vi)
1138 {
1139 	int rc = 0, i, intr_idx, iqidx;
1140 	struct sge_rxq *rxq;
1141 	struct sge_txq *txq;
1142 #ifdef TCP_OFFLOAD
1143 	struct sge_ofld_rxq *ofld_rxq;
1144 #endif
1145 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1146 	struct sge_wrq *ofld_txq;
1147 #endif
1148 #ifdef DEV_NETMAP
1149 	int saved_idx;
1150 	struct sge_nm_rxq *nm_rxq;
1151 	struct sge_nm_txq *nm_txq;
1152 #endif
1153 	char name[16];
1154 	struct port_info *pi = vi->pi;
1155 	struct adapter *sc = pi->adapter;
1156 	struct ifnet *ifp = vi->ifp;
1157 	struct sysctl_oid *oid = device_get_sysctl_tree(vi->dev);
1158 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
1159 	int maxp, mtu = ifp->if_mtu;
1160 
1161 	/* Interrupt vector to start from (when using multiple vectors) */
1162 	intr_idx = vi->first_intr;
1163 
1164 #ifdef DEV_NETMAP
1165 	saved_idx = intr_idx;
1166 	if (ifp->if_capabilities & IFCAP_NETMAP) {
1167 
1168 		/* netmap is supported with direct interrupts only. */
1169 		MPASS(!forwarding_intr_to_fwq(sc));
1170 
1171 		/*
1172 		 * We don't have buffers to back the netmap rx queues
1173 		 * right now so we create the queues in a way that
1174 		 * doesn't set off any congestion signal in the chip.
1175 		 */
1176 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_rxq",
1177 		    CTLFLAG_RD, NULL, "rx queues");
1178 		for_each_nm_rxq(vi, i, nm_rxq) {
1179 			rc = alloc_nm_rxq(vi, nm_rxq, intr_idx, i, oid);
1180 			if (rc != 0)
1181 				goto done;
1182 			intr_idx++;
1183 		}
1184 
1185 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_txq",
1186 		    CTLFLAG_RD, NULL, "tx queues");
1187 		for_each_nm_txq(vi, i, nm_txq) {
1188 			iqidx = vi->first_nm_rxq + (i % vi->nnmrxq);
1189 			rc = alloc_nm_txq(vi, nm_txq, iqidx, i, oid);
1190 			if (rc != 0)
1191 				goto done;
1192 		}
1193 	}
1194 
1195 	/* Normal rx queues and netmap rx queues share the same interrupts. */
1196 	intr_idx = saved_idx;
1197 #endif
1198 
1199 	/*
1200 	 * Allocate rx queues first because a default iqid is required when
1201 	 * creating a tx queue.
1202 	 */
1203 	maxp = mtu_to_max_payload(sc, mtu, 0);
1204 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
1205 	    CTLFLAG_RD, NULL, "rx queues");
1206 	for_each_rxq(vi, i, rxq) {
1207 
1208 		init_iq(&rxq->iq, sc, vi->tmr_idx, vi->pktc_idx, vi->qsize_rxq);
1209 
1210 		snprintf(name, sizeof(name), "%s rxq%d-fl",
1211 		    device_get_nameunit(vi->dev), i);
1212 		init_fl(sc, &rxq->fl, vi->qsize_rxq / 8, maxp, name);
1213 
1214 		rc = alloc_rxq(vi, rxq,
1215 		    forwarding_intr_to_fwq(sc) ? -1 : intr_idx, i, oid);
1216 		if (rc != 0)
1217 			goto done;
1218 		intr_idx++;
1219 	}
1220 #ifdef DEV_NETMAP
1221 	if (ifp->if_capabilities & IFCAP_NETMAP)
1222 		intr_idx = saved_idx + max(vi->nrxq, vi->nnmrxq);
1223 #endif
1224 #ifdef TCP_OFFLOAD
1225 	maxp = mtu_to_max_payload(sc, mtu, 1);
1226 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq",
1227 	    CTLFLAG_RD, NULL, "rx queues for offloaded TCP connections");
1228 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1229 
1230 		init_iq(&ofld_rxq->iq, sc, vi->ofld_tmr_idx, vi->ofld_pktc_idx,
1231 		    vi->qsize_rxq);
1232 
1233 		snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
1234 		    device_get_nameunit(vi->dev), i);
1235 		init_fl(sc, &ofld_rxq->fl, vi->qsize_rxq / 8, maxp, name);
1236 
1237 		rc = alloc_ofld_rxq(vi, ofld_rxq,
1238 		    forwarding_intr_to_fwq(sc) ? -1 : intr_idx, i, oid);
1239 		if (rc != 0)
1240 			goto done;
1241 		intr_idx++;
1242 	}
1243 #endif
1244 
1245 	/*
1246 	 * Now the tx queues.
1247 	 */
1248 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "txq", CTLFLAG_RD,
1249 	    NULL, "tx queues");
1250 	for_each_txq(vi, i, txq) {
1251 		iqidx = vi->first_rxq + (i % vi->nrxq);
1252 		snprintf(name, sizeof(name), "%s txq%d",
1253 		    device_get_nameunit(vi->dev), i);
1254 		init_eq(sc, &txq->eq, EQ_ETH, vi->qsize_txq, pi->tx_chan,
1255 		    sc->sge.rxq[iqidx].iq.cntxt_id, name);
1256 
1257 		rc = alloc_txq(vi, txq, i, oid);
1258 		if (rc != 0)
1259 			goto done;
1260 	}
1261 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1262 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_txq",
1263 	    CTLFLAG_RD, NULL, "tx queues for TOE/ETHOFLD");
1264 	for_each_ofld_txq(vi, i, ofld_txq) {
1265 		struct sysctl_oid *oid2;
1266 
1267 		snprintf(name, sizeof(name), "%s ofld_txq%d",
1268 		    device_get_nameunit(vi->dev), i);
1269 		if (vi->nofldrxq > 0) {
1270 			iqidx = vi->first_ofld_rxq + (i % vi->nofldrxq);
1271 			init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq,
1272 			    pi->tx_chan, sc->sge.ofld_rxq[iqidx].iq.cntxt_id,
1273 			    name);
1274 		} else {
1275 			iqidx = vi->first_rxq + (i % vi->nrxq);
1276 			init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq,
1277 			    pi->tx_chan, sc->sge.rxq[iqidx].iq.cntxt_id, name);
1278 		}
1279 
1280 		snprintf(name, sizeof(name), "%d", i);
1281 		oid2 = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
1282 		    name, CTLFLAG_RD, NULL, "offload tx queue");
1283 
1284 		rc = alloc_wrq(sc, vi, ofld_txq, oid2);
1285 		if (rc != 0)
1286 			goto done;
1287 	}
1288 #endif
1289 done:
1290 	if (rc)
1291 		t4_teardown_vi_queues(vi);
1292 
1293 	return (rc);
1294 }
1295 
1296 /*
1297  * Idempotent
1298  */
1299 int
1300 t4_teardown_vi_queues(struct vi_info *vi)
1301 {
1302 	int i;
1303 	struct sge_rxq *rxq;
1304 	struct sge_txq *txq;
1305 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1306 	struct port_info *pi = vi->pi;
1307 	struct adapter *sc = pi->adapter;
1308 	struct sge_wrq *ofld_txq;
1309 #endif
1310 #ifdef TCP_OFFLOAD
1311 	struct sge_ofld_rxq *ofld_rxq;
1312 #endif
1313 #ifdef DEV_NETMAP
1314 	struct sge_nm_rxq *nm_rxq;
1315 	struct sge_nm_txq *nm_txq;
1316 #endif
1317 
1318 	/* Do this before freeing the queues */
1319 	if (vi->flags & VI_SYSCTL_CTX) {
1320 		sysctl_ctx_free(&vi->ctx);
1321 		vi->flags &= ~VI_SYSCTL_CTX;
1322 	}
1323 
1324 #ifdef DEV_NETMAP
1325 	if (vi->ifp->if_capabilities & IFCAP_NETMAP) {
1326 		for_each_nm_txq(vi, i, nm_txq) {
1327 			free_nm_txq(vi, nm_txq);
1328 		}
1329 
1330 		for_each_nm_rxq(vi, i, nm_rxq) {
1331 			free_nm_rxq(vi, nm_rxq);
1332 		}
1333 	}
1334 #endif
1335 
1336 	/*
1337 	 * Take down all the tx queues first, as they reference the rx queues
1338 	 * (for egress updates, etc.).
1339 	 */
1340 
1341 	for_each_txq(vi, i, txq) {
1342 		free_txq(vi, txq);
1343 	}
1344 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1345 	for_each_ofld_txq(vi, i, ofld_txq) {
1346 		free_wrq(sc, ofld_txq);
1347 	}
1348 #endif
1349 
1350 	/*
1351 	 * Then take down the rx queues.
1352 	 */
1353 
1354 	for_each_rxq(vi, i, rxq) {
1355 		free_rxq(vi, rxq);
1356 	}
1357 #ifdef TCP_OFFLOAD
1358 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1359 		free_ofld_rxq(vi, ofld_rxq);
1360 	}
1361 #endif
1362 
1363 	return (0);
1364 }
1365 
1366 /*
1367  * Interrupt handler when the driver is using only 1 interrupt.  This is a very
1368  * unusual scenario.
1369  *
1370  * a) Deals with errors, if any.
1371  * b) Services firmware event queue, which is taking interrupts for all other
1372  *    queues.
1373  */
1374 void
1375 t4_intr_all(void *arg)
1376 {
1377 	struct adapter *sc = arg;
1378 	struct sge_iq *fwq = &sc->sge.fwq;
1379 
1380 	MPASS(sc->intr_count == 1);
1381 
1382 	if (sc->intr_type == INTR_INTX)
1383 		t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
1384 
1385 	t4_intr_err(arg);
1386 	t4_intr_evt(fwq);
1387 }
1388 
1389 /*
1390  * Interrupt handler for errors (installed directly when multiple interrupts are
1391  * being used, or called by t4_intr_all).
1392  */
1393 void
1394 t4_intr_err(void *arg)
1395 {
1396 	struct adapter *sc = arg;
1397 	uint32_t v;
1398 	const bool verbose = (sc->debug_flags & DF_VERBOSE_SLOWINTR) != 0;
1399 
1400 	if (sc->flags & ADAP_ERR)
1401 		return;
1402 
1403 	v = t4_read_reg(sc, MYPF_REG(A_PL_PF_INT_CAUSE));
1404 	if (v & F_PFSW) {
1405 		sc->swintr++;
1406 		t4_write_reg(sc, MYPF_REG(A_PL_PF_INT_CAUSE), v);
1407 	}
1408 
1409 	t4_slow_intr_handler(sc, verbose);
1410 }
1411 
1412 /*
1413  * Interrupt handler for iq-only queues.  The firmware event queue is the only
1414  * such queue right now.
1415  */
1416 void
1417 t4_intr_evt(void *arg)
1418 {
1419 	struct sge_iq *iq = arg;
1420 
1421 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1422 		service_iq(iq, 0);
1423 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1424 	}
1425 }
1426 
1427 /*
1428  * Interrupt handler for iq+fl queues.
1429  */
1430 void
1431 t4_intr(void *arg)
1432 {
1433 	struct sge_iq *iq = arg;
1434 
1435 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1436 		service_iq_fl(iq, 0);
1437 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1438 	}
1439 }
1440 
1441 #ifdef DEV_NETMAP
1442 /*
1443  * Interrupt handler for netmap rx queues.
1444  */
1445 void
1446 t4_nm_intr(void *arg)
1447 {
1448 	struct sge_nm_rxq *nm_rxq = arg;
1449 
1450 	if (atomic_cmpset_int(&nm_rxq->nm_state, NM_ON, NM_BUSY)) {
1451 		service_nm_rxq(nm_rxq);
1452 		(void) atomic_cmpset_int(&nm_rxq->nm_state, NM_BUSY, NM_ON);
1453 	}
1454 }
1455 
1456 /*
1457  * Interrupt handler for vectors shared between NIC and netmap rx queues.
1458  */
1459 void
1460 t4_vi_intr(void *arg)
1461 {
1462 	struct irq *irq = arg;
1463 
1464 	MPASS(irq->nm_rxq != NULL);
1465 	t4_nm_intr(irq->nm_rxq);
1466 
1467 	MPASS(irq->rxq != NULL);
1468 	t4_intr(irq->rxq);
1469 }
1470 #endif
1471 
1472 /*
1473  * Deals with interrupts on an iq-only (no freelist) queue.
1474  */
1475 static int
1476 service_iq(struct sge_iq *iq, int budget)
1477 {
1478 	struct sge_iq *q;
1479 	struct adapter *sc = iq->adapter;
1480 	struct iq_desc *d = &iq->desc[iq->cidx];
1481 	int ndescs = 0, limit;
1482 	int rsp_type;
1483 	uint32_t lq;
1484 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1485 
1486 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1487 	KASSERT((iq->flags & IQ_HAS_FL) == 0,
1488 	    ("%s: called for iq %p with fl (iq->flags 0x%x)", __func__, iq,
1489 	    iq->flags));
1490 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1491 	MPASS((iq->flags & IQ_LRO_ENABLED) == 0);
1492 
1493 	limit = budget ? budget : iq->qsize / 16;
1494 
1495 	/*
1496 	 * We always come back and check the descriptor ring for new indirect
1497 	 * interrupts and other responses after running a single handler.
1498 	 */
1499 	for (;;) {
1500 		while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1501 
1502 			rmb();
1503 
1504 			rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1505 			lq = be32toh(d->rsp.pldbuflen_qid);
1506 
1507 			switch (rsp_type) {
1508 			case X_RSPD_TYPE_FLBUF:
1509 				panic("%s: data for an iq (%p) with no freelist",
1510 				    __func__, iq);
1511 
1512 				/* NOTREACHED */
1513 
1514 			case X_RSPD_TYPE_CPL:
1515 				KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1516 				    ("%s: bad opcode %02x.", __func__,
1517 				    d->rss.opcode));
1518 				t4_cpl_handler[d->rss.opcode](iq, &d->rss, NULL);
1519 				break;
1520 
1521 			case X_RSPD_TYPE_INTR:
1522 				/*
1523 				 * There are 1K interrupt-capable queues (qids 0
1524 				 * through 1023).  A response type indicating a
1525 				 * forwarded interrupt with a qid >= 1K is an
1526 				 * iWARP async notification.
1527 				 */
1528 				if (__predict_true(lq >= 1024)) {
1529 					t4_an_handler(iq, &d->rsp);
1530 					break;
1531 				}
1532 
1533 				q = sc->sge.iqmap[lq - sc->sge.iq_start -
1534 				    sc->sge.iq_base];
1535 				if (atomic_cmpset_int(&q->state, IQS_IDLE,
1536 				    IQS_BUSY)) {
1537 					if (service_iq_fl(q, q->qsize / 16) == 0) {
1538 						(void) atomic_cmpset_int(&q->state,
1539 						    IQS_BUSY, IQS_IDLE);
1540 					} else {
1541 						STAILQ_INSERT_TAIL(&iql, q,
1542 						    link);
1543 					}
1544 				}
1545 				break;
1546 
1547 			default:
1548 				KASSERT(0,
1549 				    ("%s: illegal response type %d on iq %p",
1550 				    __func__, rsp_type, iq));
1551 				log(LOG_ERR,
1552 				    "%s: illegal response type %d on iq %p",
1553 				    device_get_nameunit(sc->dev), rsp_type, iq);
1554 				break;
1555 			}
1556 
1557 			d++;
1558 			if (__predict_false(++iq->cidx == iq->sidx)) {
1559 				iq->cidx = 0;
1560 				iq->gen ^= F_RSPD_GEN;
1561 				d = &iq->desc[0];
1562 			}
1563 			if (__predict_false(++ndescs == limit)) {
1564 				t4_write_reg(sc, sc->sge_gts_reg,
1565 				    V_CIDXINC(ndescs) |
1566 				    V_INGRESSQID(iq->cntxt_id) |
1567 				    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1568 				ndescs = 0;
1569 
1570 				if (budget) {
1571 					return (EINPROGRESS);
1572 				}
1573 			}
1574 		}
1575 
1576 		if (STAILQ_EMPTY(&iql))
1577 			break;
1578 
1579 		/*
1580 		 * Process the head only, and send it to the back of the list if
1581 		 * it's still not done.
1582 		 */
1583 		q = STAILQ_FIRST(&iql);
1584 		STAILQ_REMOVE_HEAD(&iql, link);
1585 		if (service_iq_fl(q, q->qsize / 8) == 0)
1586 			(void) atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1587 		else
1588 			STAILQ_INSERT_TAIL(&iql, q, link);
1589 	}
1590 
1591 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1592 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1593 
1594 	return (0);
1595 }
1596 
1597 static inline int
1598 sort_before_lro(struct lro_ctrl *lro)
1599 {
1600 
1601 	return (lro->lro_mbuf_max != 0);
1602 }
1603 
1604 static inline uint64_t
1605 last_flit_to_ns(struct adapter *sc, uint64_t lf)
1606 {
1607 	uint64_t n = be64toh(lf) & 0xfffffffffffffff;	/* 60b, not 64b. */
1608 
1609 	if (n > UINT64_MAX / 1000000)
1610 		return (n / sc->params.vpd.cclk * 1000000);
1611 	else
1612 		return (n * 1000000 / sc->params.vpd.cclk);
1613 }
1614 
1615 /*
1616  * Deals with interrupts on an iq+fl queue.
1617  */
1618 static int
1619 service_iq_fl(struct sge_iq *iq, int budget)
1620 {
1621 	struct sge_rxq *rxq = iq_to_rxq(iq);
1622 	struct sge_fl *fl;
1623 	struct adapter *sc = iq->adapter;
1624 	struct iq_desc *d = &iq->desc[iq->cidx];
1625 	int ndescs = 0, limit;
1626 	int rsp_type, refill, starved;
1627 	uint32_t lq;
1628 	uint16_t fl_hw_cidx;
1629 	struct mbuf *m0;
1630 #if defined(INET) || defined(INET6)
1631 	const struct timeval lro_timeout = {0, sc->lro_timeout};
1632 	struct lro_ctrl *lro = &rxq->lro;
1633 #endif
1634 
1635 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1636 	MPASS(iq->flags & IQ_HAS_FL);
1637 
1638 	limit = budget ? budget : iq->qsize / 16;
1639 	fl = &rxq->fl;
1640 	fl_hw_cidx = fl->hw_cidx;	/* stable snapshot */
1641 
1642 #if defined(INET) || defined(INET6)
1643 	if (iq->flags & IQ_ADJ_CREDIT) {
1644 		MPASS(sort_before_lro(lro));
1645 		iq->flags &= ~IQ_ADJ_CREDIT;
1646 		if ((d->rsp.u.type_gen & F_RSPD_GEN) != iq->gen) {
1647 			tcp_lro_flush_all(lro);
1648 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(1) |
1649 			    V_INGRESSQID((u32)iq->cntxt_id) |
1650 			    V_SEINTARM(iq->intr_params));
1651 			return (0);
1652 		}
1653 		ndescs = 1;
1654 	}
1655 #else
1656 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1657 #endif
1658 
1659 	while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1660 
1661 		rmb();
1662 
1663 		refill = 0;
1664 		m0 = NULL;
1665 		rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1666 		lq = be32toh(d->rsp.pldbuflen_qid);
1667 
1668 		switch (rsp_type) {
1669 		case X_RSPD_TYPE_FLBUF:
1670 
1671 			m0 = get_fl_payload(sc, fl, lq);
1672 			if (__predict_false(m0 == NULL))
1673 				goto out;
1674 			refill = IDXDIFF(fl->hw_cidx, fl_hw_cidx, fl->sidx) > 2;
1675 
1676 			if (iq->flags & IQ_RX_TIMESTAMP) {
1677 				/*
1678 				 * Fill up rcv_tstmp but do not set M_TSTMP.
1679 				 * rcv_tstmp is not in the format that the
1680 				 * kernel expects and we don't want to mislead
1681 				 * it.  For now this is only for custom code
1682 				 * that knows how to interpret cxgbe's stamp.
1683 				 */
1684 				m0->m_pkthdr.rcv_tstmp =
1685 				    last_flit_to_ns(sc, d->rsp.u.last_flit);
1686 #ifdef notyet
1687 				m0->m_flags |= M_TSTMP;
1688 #endif
1689 			}
1690 
1691 			/* fall through */
1692 
1693 		case X_RSPD_TYPE_CPL:
1694 			KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1695 			    ("%s: bad opcode %02x.", __func__, d->rss.opcode));
1696 			t4_cpl_handler[d->rss.opcode](iq, &d->rss, m0);
1697 			break;
1698 
1699 		case X_RSPD_TYPE_INTR:
1700 
1701 			/*
1702 			 * There are 1K interrupt-capable queues (qids 0
1703 			 * through 1023).  A response type indicating a
1704 			 * forwarded interrupt with a qid >= 1K is an
1705 			 * iWARP async notification.  That is the only
1706 			 * acceptable indirect interrupt on this queue.
1707 			 */
1708 			if (__predict_false(lq < 1024)) {
1709 				panic("%s: indirect interrupt on iq_fl %p "
1710 				    "with qid %u", __func__, iq, lq);
1711 			}
1712 
1713 			t4_an_handler(iq, &d->rsp);
1714 			break;
1715 
1716 		default:
1717 			KASSERT(0, ("%s: illegal response type %d on iq %p",
1718 			    __func__, rsp_type, iq));
1719 			log(LOG_ERR, "%s: illegal response type %d on iq %p",
1720 			    device_get_nameunit(sc->dev), rsp_type, iq);
1721 			break;
1722 		}
1723 
1724 		d++;
1725 		if (__predict_false(++iq->cidx == iq->sidx)) {
1726 			iq->cidx = 0;
1727 			iq->gen ^= F_RSPD_GEN;
1728 			d = &iq->desc[0];
1729 		}
1730 		if (__predict_false(++ndescs == limit)) {
1731 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1732 			    V_INGRESSQID(iq->cntxt_id) |
1733 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1734 			ndescs = 0;
1735 
1736 #if defined(INET) || defined(INET6)
1737 			if (iq->flags & IQ_LRO_ENABLED &&
1738 			    !sort_before_lro(lro) &&
1739 			    sc->lro_timeout != 0) {
1740 				tcp_lro_flush_inactive(lro, &lro_timeout);
1741 			}
1742 #endif
1743 			if (budget) {
1744 				FL_LOCK(fl);
1745 				refill_fl(sc, fl, 32);
1746 				FL_UNLOCK(fl);
1747 
1748 				return (EINPROGRESS);
1749 			}
1750 		}
1751 		if (refill) {
1752 			FL_LOCK(fl);
1753 			refill_fl(sc, fl, 32);
1754 			FL_UNLOCK(fl);
1755 			fl_hw_cidx = fl->hw_cidx;
1756 		}
1757 	}
1758 out:
1759 #if defined(INET) || defined(INET6)
1760 	if (iq->flags & IQ_LRO_ENABLED) {
1761 		if (ndescs > 0 && lro->lro_mbuf_count > 8) {
1762 			MPASS(sort_before_lro(lro));
1763 			/* hold back one credit and don't flush LRO state */
1764 			iq->flags |= IQ_ADJ_CREDIT;
1765 			ndescs--;
1766 		} else {
1767 			tcp_lro_flush_all(lro);
1768 		}
1769 	}
1770 #endif
1771 
1772 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1773 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1774 
1775 	FL_LOCK(fl);
1776 	starved = refill_fl(sc, fl, 64);
1777 	FL_UNLOCK(fl);
1778 	if (__predict_false(starved != 0))
1779 		add_fl_to_sfl(sc, fl);
1780 
1781 	return (0);
1782 }
1783 
1784 static inline int
1785 cl_has_metadata(struct sge_fl *fl, struct cluster_layout *cll)
1786 {
1787 	int rc = fl->flags & FL_BUF_PACKING || cll->region1 > 0;
1788 
1789 	if (rc)
1790 		MPASS(cll->region3 >= CL_METADATA_SIZE);
1791 
1792 	return (rc);
1793 }
1794 
1795 static inline struct cluster_metadata *
1796 cl_metadata(struct adapter *sc, struct sge_fl *fl, struct cluster_layout *cll,
1797     caddr_t cl)
1798 {
1799 
1800 	if (cl_has_metadata(fl, cll)) {
1801 		struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1802 
1803 		return ((struct cluster_metadata *)(cl + swz->size) - 1);
1804 	}
1805 	return (NULL);
1806 }
1807 
1808 static void
1809 rxb_free(struct mbuf *m)
1810 {
1811 	uma_zone_t zone = m->m_ext.ext_arg1;
1812 	void *cl = m->m_ext.ext_arg2;
1813 
1814 	uma_zfree(zone, cl);
1815 	counter_u64_add(extfree_rels, 1);
1816 }
1817 
1818 /*
1819  * The mbuf returned by this function could be allocated from zone_mbuf or
1820  * constructed in spare room in the cluster.
1821  *
1822  * The mbuf carries the payload in one of these ways
1823  * a) frame inside the mbuf (mbuf from zone_mbuf)
1824  * b) m_cljset (for clusters without metadata) zone_mbuf
1825  * c) m_extaddref (cluster with metadata) inline mbuf
1826  * d) m_extaddref (cluster with metadata) zone_mbuf
1827  */
1828 static struct mbuf *
1829 get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1830     int remaining)
1831 {
1832 	struct mbuf *m;
1833 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1834 	struct cluster_layout *cll = &sd->cll;
1835 	struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1836 	struct hw_buf_info *hwb = &sc->sge.hw_buf_info[cll->hwidx];
1837 	struct cluster_metadata *clm = cl_metadata(sc, fl, cll, sd->cl);
1838 	int len, blen;
1839 	caddr_t payload;
1840 
1841 	blen = hwb->size - fl->rx_offset;	/* max possible in this buf */
1842 	len = min(remaining, blen);
1843 	payload = sd->cl + cll->region1 + fl->rx_offset;
1844 	if (fl->flags & FL_BUF_PACKING) {
1845 		const u_int l = fr_offset + len;
1846 		const u_int pad = roundup2(l, fl->buf_boundary) - l;
1847 
1848 		if (fl->rx_offset + len + pad < hwb->size)
1849 			blen = len + pad;
1850 		MPASS(fl->rx_offset + blen <= hwb->size);
1851 	} else {
1852 		MPASS(fl->rx_offset == 0);	/* not packing */
1853 	}
1854 
1855 
1856 	if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) {
1857 
1858 		/*
1859 		 * Copy payload into a freshly allocated mbuf.
1860 		 */
1861 
1862 		m = fr_offset == 0 ?
1863 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1864 		if (m == NULL)
1865 			return (NULL);
1866 		fl->mbuf_allocated++;
1867 
1868 		/* copy data to mbuf */
1869 		bcopy(payload, mtod(m, caddr_t), len);
1870 
1871 	} else if (sd->nmbuf * MSIZE < cll->region1) {
1872 
1873 		/*
1874 		 * There's spare room in the cluster for an mbuf.  Create one
1875 		 * and associate it with the payload that's in the cluster.
1876 		 */
1877 
1878 		MPASS(clm != NULL);
1879 		m = (struct mbuf *)(sd->cl + sd->nmbuf * MSIZE);
1880 		/* No bzero required */
1881 		if (m_init(m, M_NOWAIT, MT_DATA,
1882 		    fr_offset == 0 ? M_PKTHDR | M_NOFREE : M_NOFREE))
1883 			return (NULL);
1884 		fl->mbuf_inlined++;
1885 		m_extaddref(m, payload, blen, &clm->refcount, rxb_free,
1886 		    swz->zone, sd->cl);
1887 		if (sd->nmbuf++ == 0)
1888 			counter_u64_add(extfree_refs, 1);
1889 
1890 	} else {
1891 
1892 		/*
1893 		 * Grab an mbuf from zone_mbuf and associate it with the
1894 		 * payload in the cluster.
1895 		 */
1896 
1897 		m = fr_offset == 0 ?
1898 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1899 		if (m == NULL)
1900 			return (NULL);
1901 		fl->mbuf_allocated++;
1902 		if (clm != NULL) {
1903 			m_extaddref(m, payload, blen, &clm->refcount,
1904 			    rxb_free, swz->zone, sd->cl);
1905 			if (sd->nmbuf++ == 0)
1906 				counter_u64_add(extfree_refs, 1);
1907 		} else {
1908 			m_cljset(m, sd->cl, swz->type);
1909 			sd->cl = NULL;	/* consumed, not a recycle candidate */
1910 		}
1911 	}
1912 	if (fr_offset == 0)
1913 		m->m_pkthdr.len = remaining;
1914 	m->m_len = len;
1915 
1916 	if (fl->flags & FL_BUF_PACKING) {
1917 		fl->rx_offset += blen;
1918 		MPASS(fl->rx_offset <= hwb->size);
1919 		if (fl->rx_offset < hwb->size)
1920 			return (m);	/* without advancing the cidx */
1921 	}
1922 
1923 	if (__predict_false(++fl->cidx % 8 == 0)) {
1924 		uint16_t cidx = fl->cidx / 8;
1925 
1926 		if (__predict_false(cidx == fl->sidx))
1927 			fl->cidx = cidx = 0;
1928 		fl->hw_cidx = cidx;
1929 	}
1930 	fl->rx_offset = 0;
1931 
1932 	return (m);
1933 }
1934 
1935 static struct mbuf *
1936 get_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf)
1937 {
1938 	struct mbuf *m0, *m, **pnext;
1939 	u_int remaining;
1940 	const u_int total = G_RSPD_LEN(len_newbuf);
1941 
1942 	if (__predict_false(fl->flags & FL_BUF_RESUME)) {
1943 		M_ASSERTPKTHDR(fl->m0);
1944 		MPASS(fl->m0->m_pkthdr.len == total);
1945 		MPASS(fl->remaining < total);
1946 
1947 		m0 = fl->m0;
1948 		pnext = fl->pnext;
1949 		remaining = fl->remaining;
1950 		fl->flags &= ~FL_BUF_RESUME;
1951 		goto get_segment;
1952 	}
1953 
1954 	if (fl->rx_offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
1955 		fl->rx_offset = 0;
1956 		if (__predict_false(++fl->cidx % 8 == 0)) {
1957 			uint16_t cidx = fl->cidx / 8;
1958 
1959 			if (__predict_false(cidx == fl->sidx))
1960 				fl->cidx = cidx = 0;
1961 			fl->hw_cidx = cidx;
1962 		}
1963 	}
1964 
1965 	/*
1966 	 * Payload starts at rx_offset in the current hw buffer.  Its length is
1967 	 * 'len' and it may span multiple hw buffers.
1968 	 */
1969 
1970 	m0 = get_scatter_segment(sc, fl, 0, total);
1971 	if (m0 == NULL)
1972 		return (NULL);
1973 	remaining = total - m0->m_len;
1974 	pnext = &m0->m_next;
1975 	while (remaining > 0) {
1976 get_segment:
1977 		MPASS(fl->rx_offset == 0);
1978 		m = get_scatter_segment(sc, fl, total - remaining, remaining);
1979 		if (__predict_false(m == NULL)) {
1980 			fl->m0 = m0;
1981 			fl->pnext = pnext;
1982 			fl->remaining = remaining;
1983 			fl->flags |= FL_BUF_RESUME;
1984 			return (NULL);
1985 		}
1986 		*pnext = m;
1987 		pnext = &m->m_next;
1988 		remaining -= m->m_len;
1989 	}
1990 	*pnext = NULL;
1991 
1992 	M_ASSERTPKTHDR(m0);
1993 	return (m0);
1994 }
1995 
1996 static int
1997 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
1998 {
1999 	struct sge_rxq *rxq = iq_to_rxq(iq);
2000 	struct ifnet *ifp = rxq->ifp;
2001 	struct adapter *sc = iq->adapter;
2002 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
2003 #if defined(INET) || defined(INET6)
2004 	struct lro_ctrl *lro = &rxq->lro;
2005 #endif
2006 	static const int sw_hashtype[4][2] = {
2007 		{M_HASHTYPE_NONE, M_HASHTYPE_NONE},
2008 		{M_HASHTYPE_RSS_IPV4, M_HASHTYPE_RSS_IPV6},
2009 		{M_HASHTYPE_RSS_TCP_IPV4, M_HASHTYPE_RSS_TCP_IPV6},
2010 		{M_HASHTYPE_RSS_UDP_IPV4, M_HASHTYPE_RSS_UDP_IPV6},
2011 	};
2012 
2013 	KASSERT(m0 != NULL, ("%s: no payload with opcode %02x", __func__,
2014 	    rss->opcode));
2015 
2016 	m0->m_pkthdr.len -= sc->params.sge.fl_pktshift;
2017 	m0->m_len -= sc->params.sge.fl_pktshift;
2018 	m0->m_data += sc->params.sge.fl_pktshift;
2019 
2020 	m0->m_pkthdr.rcvif = ifp;
2021 	M_HASHTYPE_SET(m0, sw_hashtype[rss->hash_type][rss->ipv6]);
2022 	m0->m_pkthdr.flowid = be32toh(rss->hash_val);
2023 
2024 	if (cpl->csum_calc && !(cpl->err_vec & sc->params.tp.err_vec_mask)) {
2025 		if (ifp->if_capenable & IFCAP_RXCSUM &&
2026 		    cpl->l2info & htobe32(F_RXF_IP)) {
2027 			m0->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
2028 			    CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
2029 			rxq->rxcsum++;
2030 		} else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
2031 		    cpl->l2info & htobe32(F_RXF_IP6)) {
2032 			m0->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
2033 			    CSUM_PSEUDO_HDR);
2034 			rxq->rxcsum++;
2035 		}
2036 
2037 		if (__predict_false(cpl->ip_frag))
2038 			m0->m_pkthdr.csum_data = be16toh(cpl->csum);
2039 		else
2040 			m0->m_pkthdr.csum_data = 0xffff;
2041 	}
2042 
2043 	if (cpl->vlan_ex) {
2044 		m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
2045 		m0->m_flags |= M_VLANTAG;
2046 		rxq->vlan_extraction++;
2047 	}
2048 
2049 #ifdef NUMA
2050 	m0->m_pkthdr.numa_domain = ifp->if_numa_domain;
2051 #endif
2052 #if defined(INET) || defined(INET6)
2053 	if (iq->flags & IQ_LRO_ENABLED) {
2054 		if (sort_before_lro(lro)) {
2055 			tcp_lro_queue_mbuf(lro, m0);
2056 			return (0); /* queued for sort, then LRO */
2057 		}
2058 		if (tcp_lro_rx(lro, m0, 0) == 0)
2059 			return (0); /* queued for LRO */
2060 	}
2061 #endif
2062 	ifp->if_input(ifp, m0);
2063 
2064 	return (0);
2065 }
2066 
2067 /*
2068  * Must drain the wrq or make sure that someone else will.
2069  */
2070 static void
2071 wrq_tx_drain(void *arg, int n)
2072 {
2073 	struct sge_wrq *wrq = arg;
2074 	struct sge_eq *eq = &wrq->eq;
2075 
2076 	EQ_LOCK(eq);
2077 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2078 		drain_wrq_wr_list(wrq->adapter, wrq);
2079 	EQ_UNLOCK(eq);
2080 }
2081 
2082 static void
2083 drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq)
2084 {
2085 	struct sge_eq *eq = &wrq->eq;
2086 	u_int available, dbdiff;	/* # of hardware descriptors */
2087 	u_int n;
2088 	struct wrqe *wr;
2089 	struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
2090 
2091 	EQ_LOCK_ASSERT_OWNED(eq);
2092 	MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
2093 	wr = STAILQ_FIRST(&wrq->wr_list);
2094 	MPASS(wr != NULL);	/* Must be called with something useful to do */
2095 	MPASS(eq->pidx == eq->dbidx);
2096 	dbdiff = 0;
2097 
2098 	do {
2099 		eq->cidx = read_hw_cidx(eq);
2100 		if (eq->pidx == eq->cidx)
2101 			available = eq->sidx - 1;
2102 		else
2103 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2104 
2105 		MPASS(wr->wrq == wrq);
2106 		n = howmany(wr->wr_len, EQ_ESIZE);
2107 		if (available < n)
2108 			break;
2109 
2110 		dst = (void *)&eq->desc[eq->pidx];
2111 		if (__predict_true(eq->sidx - eq->pidx > n)) {
2112 			/* Won't wrap, won't end exactly at the status page. */
2113 			bcopy(&wr->wr[0], dst, wr->wr_len);
2114 			eq->pidx += n;
2115 		} else {
2116 			int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE;
2117 
2118 			bcopy(&wr->wr[0], dst, first_portion);
2119 			if (wr->wr_len > first_portion) {
2120 				bcopy(&wr->wr[first_portion], &eq->desc[0],
2121 				    wr->wr_len - first_portion);
2122 			}
2123 			eq->pidx = n - (eq->sidx - eq->pidx);
2124 		}
2125 		wrq->tx_wrs_copied++;
2126 
2127 		if (available < eq->sidx / 4 &&
2128 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2129 				/*
2130 				 * XXX: This is not 100% reliable with some
2131 				 * types of WRs.  But this is a very unusual
2132 				 * situation for an ofld/ctrl queue anyway.
2133 				 */
2134 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2135 			    F_FW_WR_EQUEQ);
2136 		}
2137 
2138 		dbdiff += n;
2139 		if (dbdiff >= 16) {
2140 			ring_eq_db(sc, eq, dbdiff);
2141 			dbdiff = 0;
2142 		}
2143 
2144 		STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
2145 		free_wrqe(wr);
2146 		MPASS(wrq->nwr_pending > 0);
2147 		wrq->nwr_pending--;
2148 		MPASS(wrq->ndesc_needed >= n);
2149 		wrq->ndesc_needed -= n;
2150 	} while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL);
2151 
2152 	if (dbdiff)
2153 		ring_eq_db(sc, eq, dbdiff);
2154 }
2155 
2156 /*
2157  * Doesn't fail.  Holds on to work requests it can't send right away.
2158  */
2159 void
2160 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
2161 {
2162 #ifdef INVARIANTS
2163 	struct sge_eq *eq = &wrq->eq;
2164 #endif
2165 
2166 	EQ_LOCK_ASSERT_OWNED(eq);
2167 	MPASS(wr != NULL);
2168 	MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN);
2169 	MPASS((wr->wr_len & 0x7) == 0);
2170 
2171 	STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
2172 	wrq->nwr_pending++;
2173 	wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE);
2174 
2175 	if (!TAILQ_EMPTY(&wrq->incomplete_wrs))
2176 		return;	/* commit_wrq_wr will drain wr_list as well. */
2177 
2178 	drain_wrq_wr_list(sc, wrq);
2179 
2180 	/* Doorbell must have caught up to the pidx. */
2181 	MPASS(eq->pidx == eq->dbidx);
2182 }
2183 
2184 void
2185 t4_update_fl_bufsize(struct ifnet *ifp)
2186 {
2187 	struct vi_info *vi = ifp->if_softc;
2188 	struct adapter *sc = vi->pi->adapter;
2189 	struct sge_rxq *rxq;
2190 #ifdef TCP_OFFLOAD
2191 	struct sge_ofld_rxq *ofld_rxq;
2192 #endif
2193 	struct sge_fl *fl;
2194 	int i, maxp, mtu = ifp->if_mtu;
2195 
2196 	maxp = mtu_to_max_payload(sc, mtu, 0);
2197 	for_each_rxq(vi, i, rxq) {
2198 		fl = &rxq->fl;
2199 
2200 		FL_LOCK(fl);
2201 		find_best_refill_source(sc, fl, maxp);
2202 		FL_UNLOCK(fl);
2203 	}
2204 #ifdef TCP_OFFLOAD
2205 	maxp = mtu_to_max_payload(sc, mtu, 1);
2206 	for_each_ofld_rxq(vi, i, ofld_rxq) {
2207 		fl = &ofld_rxq->fl;
2208 
2209 		FL_LOCK(fl);
2210 		find_best_refill_source(sc, fl, maxp);
2211 		FL_UNLOCK(fl);
2212 	}
2213 #endif
2214 }
2215 
2216 static inline int
2217 mbuf_nsegs(struct mbuf *m)
2218 {
2219 
2220 	M_ASSERTPKTHDR(m);
2221 	KASSERT(m->m_pkthdr.l5hlen > 0,
2222 	    ("%s: mbuf %p missing information on # of segments.", __func__, m));
2223 
2224 	return (m->m_pkthdr.l5hlen);
2225 }
2226 
2227 static inline void
2228 set_mbuf_nsegs(struct mbuf *m, uint8_t nsegs)
2229 {
2230 
2231 	M_ASSERTPKTHDR(m);
2232 	m->m_pkthdr.l5hlen = nsegs;
2233 }
2234 
2235 static inline int
2236 mbuf_cflags(struct mbuf *m)
2237 {
2238 
2239 	M_ASSERTPKTHDR(m);
2240 	return (m->m_pkthdr.PH_loc.eight[4]);
2241 }
2242 
2243 static inline void
2244 set_mbuf_cflags(struct mbuf *m, uint8_t flags)
2245 {
2246 
2247 	M_ASSERTPKTHDR(m);
2248 	m->m_pkthdr.PH_loc.eight[4] = flags;
2249 }
2250 
2251 static inline int
2252 mbuf_len16(struct mbuf *m)
2253 {
2254 	int n;
2255 
2256 	M_ASSERTPKTHDR(m);
2257 	n = m->m_pkthdr.PH_loc.eight[0];
2258 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2259 
2260 	return (n);
2261 }
2262 
2263 static inline void
2264 set_mbuf_len16(struct mbuf *m, uint8_t len16)
2265 {
2266 
2267 	M_ASSERTPKTHDR(m);
2268 	m->m_pkthdr.PH_loc.eight[0] = len16;
2269 }
2270 
2271 #ifdef RATELIMIT
2272 static inline int
2273 mbuf_eo_nsegs(struct mbuf *m)
2274 {
2275 
2276 	M_ASSERTPKTHDR(m);
2277 	return (m->m_pkthdr.PH_loc.eight[1]);
2278 }
2279 
2280 static inline void
2281 set_mbuf_eo_nsegs(struct mbuf *m, uint8_t nsegs)
2282 {
2283 
2284 	M_ASSERTPKTHDR(m);
2285 	m->m_pkthdr.PH_loc.eight[1] = nsegs;
2286 }
2287 
2288 static inline int
2289 mbuf_eo_len16(struct mbuf *m)
2290 {
2291 	int n;
2292 
2293 	M_ASSERTPKTHDR(m);
2294 	n = m->m_pkthdr.PH_loc.eight[2];
2295 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2296 
2297 	return (n);
2298 }
2299 
2300 static inline void
2301 set_mbuf_eo_len16(struct mbuf *m, uint8_t len16)
2302 {
2303 
2304 	M_ASSERTPKTHDR(m);
2305 	m->m_pkthdr.PH_loc.eight[2] = len16;
2306 }
2307 
2308 static inline int
2309 mbuf_eo_tsclk_tsoff(struct mbuf *m)
2310 {
2311 
2312 	M_ASSERTPKTHDR(m);
2313 	return (m->m_pkthdr.PH_loc.eight[3]);
2314 }
2315 
2316 static inline void
2317 set_mbuf_eo_tsclk_tsoff(struct mbuf *m, uint8_t tsclk_tsoff)
2318 {
2319 
2320 	M_ASSERTPKTHDR(m);
2321 	m->m_pkthdr.PH_loc.eight[3] = tsclk_tsoff;
2322 }
2323 
2324 static inline int
2325 needs_eo(struct mbuf *m)
2326 {
2327 
2328 	return (m->m_pkthdr.snd_tag != NULL);
2329 }
2330 #endif
2331 
2332 /*
2333  * Try to allocate an mbuf to contain a raw work request.  To make it
2334  * easy to construct the work request, don't allocate a chain but a
2335  * single mbuf.
2336  */
2337 struct mbuf *
2338 alloc_wr_mbuf(int len, int how)
2339 {
2340 	struct mbuf *m;
2341 
2342 	if (len <= MHLEN)
2343 		m = m_gethdr(how, MT_DATA);
2344 	else if (len <= MCLBYTES)
2345 		m = m_getcl(how, MT_DATA, M_PKTHDR);
2346 	else
2347 		m = NULL;
2348 	if (m == NULL)
2349 		return (NULL);
2350 	m->m_pkthdr.len = len;
2351 	m->m_len = len;
2352 	set_mbuf_cflags(m, MC_RAW_WR);
2353 	set_mbuf_len16(m, howmany(len, 16));
2354 	return (m);
2355 }
2356 
2357 static inline int
2358 needs_tso(struct mbuf *m)
2359 {
2360 
2361 	M_ASSERTPKTHDR(m);
2362 
2363 	return (m->m_pkthdr.csum_flags & CSUM_TSO);
2364 }
2365 
2366 static inline int
2367 needs_l3_csum(struct mbuf *m)
2368 {
2369 
2370 	M_ASSERTPKTHDR(m);
2371 
2372 	return (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO));
2373 }
2374 
2375 static inline int
2376 needs_l4_csum(struct mbuf *m)
2377 {
2378 
2379 	M_ASSERTPKTHDR(m);
2380 
2381 	return (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
2382 	    CSUM_TCP_IPV6 | CSUM_TSO));
2383 }
2384 
2385 static inline int
2386 needs_tcp_csum(struct mbuf *m)
2387 {
2388 
2389 	M_ASSERTPKTHDR(m);
2390 	return (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_TCP_IPV6 | CSUM_TSO));
2391 }
2392 
2393 #ifdef RATELIMIT
2394 static inline int
2395 needs_udp_csum(struct mbuf *m)
2396 {
2397 
2398 	M_ASSERTPKTHDR(m);
2399 	return (m->m_pkthdr.csum_flags & (CSUM_UDP | CSUM_UDP_IPV6));
2400 }
2401 #endif
2402 
2403 static inline int
2404 needs_vlan_insertion(struct mbuf *m)
2405 {
2406 
2407 	M_ASSERTPKTHDR(m);
2408 
2409 	return (m->m_flags & M_VLANTAG);
2410 }
2411 
2412 static void *
2413 m_advance(struct mbuf **pm, int *poffset, int len)
2414 {
2415 	struct mbuf *m = *pm;
2416 	int offset = *poffset;
2417 	uintptr_t p = 0;
2418 
2419 	MPASS(len > 0);
2420 
2421 	for (;;) {
2422 		if (offset + len < m->m_len) {
2423 			offset += len;
2424 			p = mtod(m, uintptr_t) + offset;
2425 			break;
2426 		}
2427 		len -= m->m_len - offset;
2428 		m = m->m_next;
2429 		offset = 0;
2430 		MPASS(m != NULL);
2431 	}
2432 	*poffset = offset;
2433 	*pm = m;
2434 	return ((void *)p);
2435 }
2436 
2437 /*
2438  * Can deal with empty mbufs in the chain that have m_len = 0, but the chain
2439  * must have at least one mbuf that's not empty.  It is possible for this
2440  * routine to return 0 if skip accounts for all the contents of the mbuf chain.
2441  */
2442 static inline int
2443 count_mbuf_nsegs(struct mbuf *m, int skip)
2444 {
2445 	vm_paddr_t lastb, next;
2446 	vm_offset_t va;
2447 	int len, nsegs;
2448 
2449 	M_ASSERTPKTHDR(m);
2450 	MPASS(m->m_pkthdr.len > 0);
2451 	MPASS(m->m_pkthdr.len >= skip);
2452 
2453 	nsegs = 0;
2454 	lastb = 0;
2455 	for (; m; m = m->m_next) {
2456 
2457 		len = m->m_len;
2458 		if (__predict_false(len == 0))
2459 			continue;
2460 		if (skip >= len) {
2461 			skip -= len;
2462 			continue;
2463 		}
2464 		va = mtod(m, vm_offset_t) + skip;
2465 		len -= skip;
2466 		skip = 0;
2467 		next = pmap_kextract(va);
2468 		nsegs += sglist_count((void *)(uintptr_t)va, len);
2469 		if (lastb + 1 == next)
2470 			nsegs--;
2471 		lastb = pmap_kextract(va + len - 1);
2472 	}
2473 
2474 	return (nsegs);
2475 }
2476 
2477 /*
2478  * Analyze the mbuf to determine its tx needs.  The mbuf passed in may change:
2479  * a) caller can assume it's been freed if this function returns with an error.
2480  * b) it may get defragged up if the gather list is too long for the hardware.
2481  */
2482 int
2483 parse_pkt(struct adapter *sc, struct mbuf **mp)
2484 {
2485 	struct mbuf *m0 = *mp, *m;
2486 	int rc, nsegs, defragged = 0, offset;
2487 	struct ether_header *eh;
2488 	void *l3hdr;
2489 #if defined(INET) || defined(INET6)
2490 	struct tcphdr *tcp;
2491 #endif
2492 	uint16_t eh_type;
2493 
2494 	M_ASSERTPKTHDR(m0);
2495 	if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) {
2496 		rc = EINVAL;
2497 fail:
2498 		m_freem(m0);
2499 		*mp = NULL;
2500 		return (rc);
2501 	}
2502 restart:
2503 	/*
2504 	 * First count the number of gather list segments in the payload.
2505 	 * Defrag the mbuf if nsegs exceeds the hardware limit.
2506 	 */
2507 	M_ASSERTPKTHDR(m0);
2508 	MPASS(m0->m_pkthdr.len > 0);
2509 	nsegs = count_mbuf_nsegs(m0, 0);
2510 	if (nsegs > (needs_tso(m0) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS)) {
2511 		if (defragged++ > 0 || (m = m_defrag(m0, M_NOWAIT)) == NULL) {
2512 			rc = EFBIG;
2513 			goto fail;
2514 		}
2515 		*mp = m0 = m;	/* update caller's copy after defrag */
2516 		goto restart;
2517 	}
2518 
2519 	if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN)) {
2520 		m0 = m_pullup(m0, m0->m_pkthdr.len);
2521 		if (m0 == NULL) {
2522 			/* Should have left well enough alone. */
2523 			rc = EFBIG;
2524 			goto fail;
2525 		}
2526 		*mp = m0;	/* update caller's copy after pullup */
2527 		goto restart;
2528 	}
2529 	set_mbuf_nsegs(m0, nsegs);
2530 	set_mbuf_cflags(m0, 0);
2531 	if (sc->flags & IS_VF)
2532 		set_mbuf_len16(m0, txpkt_vm_len16(nsegs, needs_tso(m0)));
2533 	else
2534 		set_mbuf_len16(m0, txpkt_len16(nsegs, needs_tso(m0)));
2535 
2536 #ifdef RATELIMIT
2537 	/*
2538 	 * Ethofld is limited to TCP and UDP for now, and only when L4 hw
2539 	 * checksumming is enabled.  needs_l4_csum happens to check for all the
2540 	 * right things.
2541 	 */
2542 	if (__predict_false(needs_eo(m0) && !needs_l4_csum(m0)))
2543 		m0->m_pkthdr.snd_tag = NULL;
2544 #endif
2545 
2546 	if (!needs_tso(m0) &&
2547 #ifdef RATELIMIT
2548 	    !needs_eo(m0) &&
2549 #endif
2550 	    !(sc->flags & IS_VF && (needs_l3_csum(m0) || needs_l4_csum(m0))))
2551 		return (0);
2552 
2553 	m = m0;
2554 	eh = mtod(m, struct ether_header *);
2555 	eh_type = ntohs(eh->ether_type);
2556 	if (eh_type == ETHERTYPE_VLAN) {
2557 		struct ether_vlan_header *evh = (void *)eh;
2558 
2559 		eh_type = ntohs(evh->evl_proto);
2560 		m0->m_pkthdr.l2hlen = sizeof(*evh);
2561 	} else
2562 		m0->m_pkthdr.l2hlen = sizeof(*eh);
2563 
2564 	offset = 0;
2565 	l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2566 
2567 	switch (eh_type) {
2568 #ifdef INET6
2569 	case ETHERTYPE_IPV6:
2570 	{
2571 		struct ip6_hdr *ip6 = l3hdr;
2572 
2573 		MPASS(!needs_tso(m0) || ip6->ip6_nxt == IPPROTO_TCP);
2574 
2575 		m0->m_pkthdr.l3hlen = sizeof(*ip6);
2576 		break;
2577 	}
2578 #endif
2579 #ifdef INET
2580 	case ETHERTYPE_IP:
2581 	{
2582 		struct ip *ip = l3hdr;
2583 
2584 		m0->m_pkthdr.l3hlen = ip->ip_hl * 4;
2585 		break;
2586 	}
2587 #endif
2588 	default:
2589 		panic("%s: ethertype 0x%04x unknown.  if_cxgbe must be compiled"
2590 		    " with the same INET/INET6 options as the kernel.",
2591 		    __func__, eh_type);
2592 	}
2593 
2594 #if defined(INET) || defined(INET6)
2595 	if (needs_tcp_csum(m0)) {
2596 		tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen);
2597 		m0->m_pkthdr.l4hlen = tcp->th_off * 4;
2598 #ifdef RATELIMIT
2599 		if (tsclk >= 0 && *(uint32_t *)(tcp + 1) == ntohl(0x0101080a)) {
2600 			set_mbuf_eo_tsclk_tsoff(m0,
2601 			    V_FW_ETH_TX_EO_WR_TSCLK(tsclk) |
2602 			    V_FW_ETH_TX_EO_WR_TSOFF(sizeof(*tcp) / 2 + 1));
2603 		} else
2604 			set_mbuf_eo_tsclk_tsoff(m0, 0);
2605 	} else if (needs_udp_csum(m)) {
2606 		m0->m_pkthdr.l4hlen = sizeof(struct udphdr);
2607 #endif
2608 	}
2609 #ifdef RATELIMIT
2610 	if (needs_eo(m0)) {
2611 		u_int immhdrs;
2612 
2613 		/* EO WRs have the headers in the WR and not the GL. */
2614 		immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen +
2615 		    m0->m_pkthdr.l4hlen;
2616 		nsegs = count_mbuf_nsegs(m0, immhdrs);
2617 		set_mbuf_eo_nsegs(m0, nsegs);
2618 		set_mbuf_eo_len16(m0,
2619 		    txpkt_eo_len16(nsegs, immhdrs, needs_tso(m0)));
2620 	}
2621 #endif
2622 #endif
2623 	MPASS(m0 == *mp);
2624 	return (0);
2625 }
2626 
2627 void *
2628 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie)
2629 {
2630 	struct sge_eq *eq = &wrq->eq;
2631 	struct adapter *sc = wrq->adapter;
2632 	int ndesc, available;
2633 	struct wrqe *wr;
2634 	void *w;
2635 
2636 	MPASS(len16 > 0);
2637 	ndesc = howmany(len16, EQ_ESIZE / 16);
2638 	MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC);
2639 
2640 	EQ_LOCK(eq);
2641 
2642 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2643 		drain_wrq_wr_list(sc, wrq);
2644 
2645 	if (!STAILQ_EMPTY(&wrq->wr_list)) {
2646 slowpath:
2647 		EQ_UNLOCK(eq);
2648 		wr = alloc_wrqe(len16 * 16, wrq);
2649 		if (__predict_false(wr == NULL))
2650 			return (NULL);
2651 		cookie->pidx = -1;
2652 		cookie->ndesc = ndesc;
2653 		return (&wr->wr);
2654 	}
2655 
2656 	eq->cidx = read_hw_cidx(eq);
2657 	if (eq->pidx == eq->cidx)
2658 		available = eq->sidx - 1;
2659 	else
2660 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2661 	if (available < ndesc)
2662 		goto slowpath;
2663 
2664 	cookie->pidx = eq->pidx;
2665 	cookie->ndesc = ndesc;
2666 	TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link);
2667 
2668 	w = &eq->desc[eq->pidx];
2669 	IDXINCR(eq->pidx, ndesc, eq->sidx);
2670 	if (__predict_false(cookie->pidx + ndesc > eq->sidx)) {
2671 		w = &wrq->ss[0];
2672 		wrq->ss_pidx = cookie->pidx;
2673 		wrq->ss_len = len16 * 16;
2674 	}
2675 
2676 	EQ_UNLOCK(eq);
2677 
2678 	return (w);
2679 }
2680 
2681 void
2682 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie)
2683 {
2684 	struct sge_eq *eq = &wrq->eq;
2685 	struct adapter *sc = wrq->adapter;
2686 	int ndesc, pidx;
2687 	struct wrq_cookie *prev, *next;
2688 
2689 	if (cookie->pidx == -1) {
2690 		struct wrqe *wr = __containerof(w, struct wrqe, wr);
2691 
2692 		t4_wrq_tx(sc, wr);
2693 		return;
2694 	}
2695 
2696 	if (__predict_false(w == &wrq->ss[0])) {
2697 		int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE;
2698 
2699 		MPASS(wrq->ss_len > n);	/* WR had better wrap around. */
2700 		bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n);
2701 		bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n);
2702 		wrq->tx_wrs_ss++;
2703 	} else
2704 		wrq->tx_wrs_direct++;
2705 
2706 	EQ_LOCK(eq);
2707 	ndesc = cookie->ndesc;	/* Can be more than SGE_MAX_WR_NDESC here. */
2708 	pidx = cookie->pidx;
2709 	MPASS(pidx >= 0 && pidx < eq->sidx);
2710 	prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link);
2711 	next = TAILQ_NEXT(cookie, link);
2712 	if (prev == NULL) {
2713 		MPASS(pidx == eq->dbidx);
2714 		if (next == NULL || ndesc >= 16) {
2715 			int available;
2716 			struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
2717 
2718 			/*
2719 			 * Note that the WR via which we'll request tx updates
2720 			 * is at pidx and not eq->pidx, which has moved on
2721 			 * already.
2722 			 */
2723 			dst = (void *)&eq->desc[pidx];
2724 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2725 			if (available < eq->sidx / 4 &&
2726 			    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2727 				/*
2728 				 * XXX: This is not 100% reliable with some
2729 				 * types of WRs.  But this is a very unusual
2730 				 * situation for an ofld/ctrl queue anyway.
2731 				 */
2732 				dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2733 				    F_FW_WR_EQUEQ);
2734 			}
2735 
2736 			ring_eq_db(wrq->adapter, eq, ndesc);
2737 		} else {
2738 			MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc);
2739 			next->pidx = pidx;
2740 			next->ndesc += ndesc;
2741 		}
2742 	} else {
2743 		MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc);
2744 		prev->ndesc += ndesc;
2745 	}
2746 	TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link);
2747 
2748 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2749 		drain_wrq_wr_list(sc, wrq);
2750 
2751 #ifdef INVARIANTS
2752 	if (TAILQ_EMPTY(&wrq->incomplete_wrs)) {
2753 		/* Doorbell must have caught up to the pidx. */
2754 		MPASS(wrq->eq.pidx == wrq->eq.dbidx);
2755 	}
2756 #endif
2757 	EQ_UNLOCK(eq);
2758 }
2759 
2760 static u_int
2761 can_resume_eth_tx(struct mp_ring *r)
2762 {
2763 	struct sge_eq *eq = r->cookie;
2764 
2765 	return (total_available_tx_desc(eq) > eq->sidx / 8);
2766 }
2767 
2768 static inline int
2769 cannot_use_txpkts(struct mbuf *m)
2770 {
2771 	/* maybe put a GL limit too, to avoid silliness? */
2772 
2773 	return (needs_tso(m) || (mbuf_cflags(m) & MC_RAW_WR) != 0);
2774 }
2775 
2776 static inline int
2777 discard_tx(struct sge_eq *eq)
2778 {
2779 
2780 	return ((eq->flags & (EQ_ENABLED | EQ_QFLUSH)) != EQ_ENABLED);
2781 }
2782 
2783 static inline int
2784 wr_can_update_eq(struct fw_eth_tx_pkts_wr *wr)
2785 {
2786 
2787 	switch (G_FW_WR_OP(be32toh(wr->op_pkd))) {
2788 	case FW_ULPTX_WR:
2789 	case FW_ETH_TX_PKT_WR:
2790 	case FW_ETH_TX_PKTS_WR:
2791 	case FW_ETH_TX_PKT_VM_WR:
2792 		return (1);
2793 	default:
2794 		return (0);
2795 	}
2796 }
2797 
2798 /*
2799  * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to
2800  * be consumed.  Return the actual number consumed.  0 indicates a stall.
2801  */
2802 static u_int
2803 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx)
2804 {
2805 	struct sge_txq *txq = r->cookie;
2806 	struct sge_eq *eq = &txq->eq;
2807 	struct ifnet *ifp = txq->ifp;
2808 	struct vi_info *vi = ifp->if_softc;
2809 	struct port_info *pi = vi->pi;
2810 	struct adapter *sc = pi->adapter;
2811 	u_int total, remaining;		/* # of packets */
2812 	u_int available, dbdiff;	/* # of hardware descriptors */
2813 	u_int n, next_cidx;
2814 	struct mbuf *m0, *tail;
2815 	struct txpkts txp;
2816 	struct fw_eth_tx_pkts_wr *wr;	/* any fw WR struct will do */
2817 
2818 	remaining = IDXDIFF(pidx, cidx, r->size);
2819 	MPASS(remaining > 0);	/* Must not be called without work to do. */
2820 	total = 0;
2821 
2822 	TXQ_LOCK(txq);
2823 	if (__predict_false(discard_tx(eq))) {
2824 		while (cidx != pidx) {
2825 			m0 = r->items[cidx];
2826 			m_freem(m0);
2827 			if (++cidx == r->size)
2828 				cidx = 0;
2829 		}
2830 		reclaim_tx_descs(txq, 2048);
2831 		total = remaining;
2832 		goto done;
2833 	}
2834 
2835 	/* How many hardware descriptors do we have readily available. */
2836 	if (eq->pidx == eq->cidx)
2837 		available = eq->sidx - 1;
2838 	else
2839 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2840 	dbdiff = IDXDIFF(eq->pidx, eq->dbidx, eq->sidx);
2841 
2842 	while (remaining > 0) {
2843 
2844 		m0 = r->items[cidx];
2845 		M_ASSERTPKTHDR(m0);
2846 		MPASS(m0->m_nextpkt == NULL);
2847 
2848 		if (available < SGE_MAX_WR_NDESC) {
2849 			available += reclaim_tx_descs(txq, 64);
2850 			if (available < howmany(mbuf_len16(m0), EQ_ESIZE / 16))
2851 				break;	/* out of descriptors */
2852 		}
2853 
2854 		next_cidx = cidx + 1;
2855 		if (__predict_false(next_cidx == r->size))
2856 			next_cidx = 0;
2857 
2858 		wr = (void *)&eq->desc[eq->pidx];
2859 		if (sc->flags & IS_VF) {
2860 			total++;
2861 			remaining--;
2862 			ETHER_BPF_MTAP(ifp, m0);
2863 			n = write_txpkt_vm_wr(sc, txq, (void *)wr, m0,
2864 			    available);
2865 		} else if (remaining > 1 &&
2866 		    try_txpkts(m0, r->items[next_cidx], &txp, available) == 0) {
2867 
2868 			/* pkts at cidx, next_cidx should both be in txp. */
2869 			MPASS(txp.npkt == 2);
2870 			tail = r->items[next_cidx];
2871 			MPASS(tail->m_nextpkt == NULL);
2872 			ETHER_BPF_MTAP(ifp, m0);
2873 			ETHER_BPF_MTAP(ifp, tail);
2874 			m0->m_nextpkt = tail;
2875 
2876 			if (__predict_false(++next_cidx == r->size))
2877 				next_cidx = 0;
2878 
2879 			while (next_cidx != pidx) {
2880 				if (add_to_txpkts(r->items[next_cidx], &txp,
2881 				    available) != 0)
2882 					break;
2883 				tail->m_nextpkt = r->items[next_cidx];
2884 				tail = tail->m_nextpkt;
2885 				ETHER_BPF_MTAP(ifp, tail);
2886 				if (__predict_false(++next_cidx == r->size))
2887 					next_cidx = 0;
2888 			}
2889 
2890 			n = write_txpkts_wr(txq, wr, m0, &txp, available);
2891 			total += txp.npkt;
2892 			remaining -= txp.npkt;
2893 		} else if (mbuf_cflags(m0) & MC_RAW_WR) {
2894 			total++;
2895 			remaining--;
2896 			n = write_raw_wr(txq, (void *)wr, m0, available);
2897 		} else {
2898 			total++;
2899 			remaining--;
2900 			ETHER_BPF_MTAP(ifp, m0);
2901 			n = write_txpkt_wr(txq, (void *)wr, m0, available);
2902 		}
2903 		MPASS(n >= 1 && n <= available && n <= SGE_MAX_WR_NDESC);
2904 
2905 		available -= n;
2906 		dbdiff += n;
2907 		IDXINCR(eq->pidx, n, eq->sidx);
2908 
2909 		if (wr_can_update_eq(wr)) {
2910 			if (total_available_tx_desc(eq) < eq->sidx / 4 &&
2911 			    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2912 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2913 				    F_FW_WR_EQUEQ);
2914 				eq->equeqidx = eq->pidx;
2915 			} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >=
2916 			    32) {
2917 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
2918 				eq->equeqidx = eq->pidx;
2919 			}
2920 		}
2921 
2922 		if (dbdiff >= 16 && remaining >= 4) {
2923 			ring_eq_db(sc, eq, dbdiff);
2924 			available += reclaim_tx_descs(txq, 4 * dbdiff);
2925 			dbdiff = 0;
2926 		}
2927 
2928 		cidx = next_cidx;
2929 	}
2930 	if (dbdiff != 0) {
2931 		ring_eq_db(sc, eq, dbdiff);
2932 		reclaim_tx_descs(txq, 32);
2933 	}
2934 done:
2935 	TXQ_UNLOCK(txq);
2936 
2937 	return (total);
2938 }
2939 
2940 static inline void
2941 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
2942     int qsize)
2943 {
2944 
2945 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
2946 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
2947 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
2948 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
2949 
2950 	iq->flags = 0;
2951 	iq->adapter = sc;
2952 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
2953 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
2954 	if (pktc_idx >= 0) {
2955 		iq->intr_params |= F_QINTR_CNT_EN;
2956 		iq->intr_pktc_idx = pktc_idx;
2957 	}
2958 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
2959 	iq->sidx = iq->qsize - sc->params.sge.spg_len / IQ_ESIZE;
2960 }
2961 
2962 static inline void
2963 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name)
2964 {
2965 
2966 	fl->qsize = qsize;
2967 	fl->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2968 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
2969 	if (sc->flags & BUF_PACKING_OK &&
2970 	    ((!is_t4(sc) && buffer_packing) ||	/* T5+: enabled unless 0 */
2971 	    (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */
2972 		fl->flags |= FL_BUF_PACKING;
2973 	find_best_refill_source(sc, fl, maxp);
2974 	find_safe_refill_source(sc, fl);
2975 }
2976 
2977 static inline void
2978 init_eq(struct adapter *sc, struct sge_eq *eq, int eqtype, int qsize,
2979     uint8_t tx_chan, uint16_t iqid, char *name)
2980 {
2981 	KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
2982 
2983 	eq->flags = eqtype & EQ_TYPEMASK;
2984 	eq->tx_chan = tx_chan;
2985 	eq->iqid = iqid;
2986 	eq->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2987 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
2988 }
2989 
2990 static int
2991 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
2992     bus_dmamap_t *map, bus_addr_t *pa, void **va)
2993 {
2994 	int rc;
2995 
2996 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
2997 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
2998 	if (rc != 0) {
2999 		device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
3000 		goto done;
3001 	}
3002 
3003 	rc = bus_dmamem_alloc(*tag, va,
3004 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
3005 	if (rc != 0) {
3006 		device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
3007 		goto done;
3008 	}
3009 
3010 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
3011 	if (rc != 0) {
3012 		device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
3013 		goto done;
3014 	}
3015 done:
3016 	if (rc)
3017 		free_ring(sc, *tag, *map, *pa, *va);
3018 
3019 	return (rc);
3020 }
3021 
3022 static int
3023 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
3024     bus_addr_t pa, void *va)
3025 {
3026 	if (pa)
3027 		bus_dmamap_unload(tag, map);
3028 	if (va)
3029 		bus_dmamem_free(tag, va, map);
3030 	if (tag)
3031 		bus_dma_tag_destroy(tag);
3032 
3033 	return (0);
3034 }
3035 
3036 /*
3037  * Allocates the ring for an ingress queue and an optional freelist.  If the
3038  * freelist is specified it will be allocated and then associated with the
3039  * ingress queue.
3040  *
3041  * Returns errno on failure.  Resources allocated up to that point may still be
3042  * allocated.  Caller is responsible for cleanup in case this function fails.
3043  *
3044  * If the ingress queue will take interrupts directly then the intr_idx
3045  * specifies the vector, starting from 0.  -1 means the interrupts for this
3046  * queue should be forwarded to the fwq.
3047  */
3048 static int
3049 alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl,
3050     int intr_idx, int cong)
3051 {
3052 	int rc, i, cntxt_id;
3053 	size_t len;
3054 	struct fw_iq_cmd c;
3055 	struct port_info *pi = vi->pi;
3056 	struct adapter *sc = iq->adapter;
3057 	struct sge_params *sp = &sc->params.sge;
3058 	__be32 v = 0;
3059 
3060 	len = iq->qsize * IQ_ESIZE;
3061 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
3062 	    (void **)&iq->desc);
3063 	if (rc != 0)
3064 		return (rc);
3065 
3066 	bzero(&c, sizeof(c));
3067 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
3068 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
3069 	    V_FW_IQ_CMD_VFN(0));
3070 
3071 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
3072 	    FW_LEN16(c));
3073 
3074 	/* Special handling for firmware event queue */
3075 	if (iq == &sc->sge.fwq)
3076 		v |= F_FW_IQ_CMD_IQASYNCH;
3077 
3078 	if (intr_idx < 0) {
3079 		/* Forwarded interrupts, all headed to fwq */
3080 		v |= F_FW_IQ_CMD_IQANDST;
3081 		v |= V_FW_IQ_CMD_IQANDSTINDEX(sc->sge.fwq.cntxt_id);
3082 	} else {
3083 		KASSERT(intr_idx < sc->intr_count,
3084 		    ("%s: invalid direct intr_idx %d", __func__, intr_idx));
3085 		v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
3086 	}
3087 
3088 	c.type_to_iqandstindex = htobe32(v |
3089 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
3090 	    V_FW_IQ_CMD_VIID(vi->viid) |
3091 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
3092 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
3093 	    F_FW_IQ_CMD_IQGTSMODE |
3094 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
3095 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
3096 	c.iqsize = htobe16(iq->qsize);
3097 	c.iqaddr = htobe64(iq->ba);
3098 	if (cong >= 0)
3099 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
3100 
3101 	if (fl) {
3102 		mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
3103 
3104 		len = fl->qsize * EQ_ESIZE;
3105 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
3106 		    &fl->ba, (void **)&fl->desc);
3107 		if (rc)
3108 			return (rc);
3109 
3110 		/* Allocate space for one software descriptor per buffer. */
3111 		rc = alloc_fl_sdesc(fl);
3112 		if (rc != 0) {
3113 			device_printf(sc->dev,
3114 			    "failed to setup fl software descriptors: %d\n",
3115 			    rc);
3116 			return (rc);
3117 		}
3118 
3119 		if (fl->flags & FL_BUF_PACKING) {
3120 			fl->lowat = roundup2(sp->fl_starve_threshold2, 8);
3121 			fl->buf_boundary = sp->pack_boundary;
3122 		} else {
3123 			fl->lowat = roundup2(sp->fl_starve_threshold, 8);
3124 			fl->buf_boundary = 16;
3125 		}
3126 		if (fl_pad && fl->buf_boundary < sp->pad_boundary)
3127 			fl->buf_boundary = sp->pad_boundary;
3128 
3129 		c.iqns_to_fl0congen |=
3130 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
3131 			F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
3132 			(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
3133 			(fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
3134 			    0));
3135 		if (cong >= 0) {
3136 			c.iqns_to_fl0congen |=
3137 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
3138 				    F_FW_IQ_CMD_FL0CONGCIF |
3139 				    F_FW_IQ_CMD_FL0CONGEN);
3140 		}
3141 		c.fl0dcaen_to_fl0cidxfthresh =
3142 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
3143 			X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B) |
3144 			V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
3145 			X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
3146 		c.fl0size = htobe16(fl->qsize);
3147 		c.fl0addr = htobe64(fl->ba);
3148 	}
3149 
3150 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3151 	if (rc != 0) {
3152 		device_printf(sc->dev,
3153 		    "failed to create ingress queue: %d\n", rc);
3154 		return (rc);
3155 	}
3156 
3157 	iq->cidx = 0;
3158 	iq->gen = F_RSPD_GEN;
3159 	iq->intr_next = iq->intr_params;
3160 	iq->cntxt_id = be16toh(c.iqid);
3161 	iq->abs_id = be16toh(c.physiqid);
3162 	iq->flags |= IQ_ALLOCATED;
3163 
3164 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
3165 	if (cntxt_id >= sc->sge.niq) {
3166 		panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
3167 		    cntxt_id, sc->sge.niq - 1);
3168 	}
3169 	sc->sge.iqmap[cntxt_id] = iq;
3170 
3171 	if (fl) {
3172 		u_int qid;
3173 
3174 		iq->flags |= IQ_HAS_FL;
3175 		fl->cntxt_id = be16toh(c.fl0id);
3176 		fl->pidx = fl->cidx = 0;
3177 
3178 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
3179 		if (cntxt_id >= sc->sge.neq) {
3180 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
3181 			    __func__, cntxt_id, sc->sge.neq - 1);
3182 		}
3183 		sc->sge.eqmap[cntxt_id] = (void *)fl;
3184 
3185 		qid = fl->cntxt_id;
3186 		if (isset(&sc->doorbells, DOORBELL_UDB)) {
3187 			uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3188 			uint32_t mask = (1 << s_qpp) - 1;
3189 			volatile uint8_t *udb;
3190 
3191 			udb = sc->udbs_base + UDBS_DB_OFFSET;
3192 			udb += (qid >> s_qpp) << PAGE_SHIFT;
3193 			qid &= mask;
3194 			if (qid < PAGE_SIZE / UDBS_SEG_SIZE) {
3195 				udb += qid << UDBS_SEG_SHIFT;
3196 				qid = 0;
3197 			}
3198 			fl->udb = (volatile void *)udb;
3199 		}
3200 		fl->dbval = V_QID(qid) | sc->chip_params->sge_fl_db;
3201 
3202 		FL_LOCK(fl);
3203 		/* Enough to make sure the SGE doesn't think it's starved */
3204 		refill_fl(sc, fl, fl->lowat);
3205 		FL_UNLOCK(fl);
3206 	}
3207 
3208 	if (chip_id(sc) >= CHELSIO_T5 && !(sc->flags & IS_VF) && cong >= 0) {
3209 		uint32_t param, val;
3210 
3211 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
3212 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
3213 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
3214 		if (cong == 0)
3215 			val = 1 << 19;
3216 		else {
3217 			val = 2 << 19;
3218 			for (i = 0; i < 4; i++) {
3219 				if (cong & (1 << i))
3220 					val |= 1 << (i << 2);
3221 			}
3222 		}
3223 
3224 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
3225 		if (rc != 0) {
3226 			/* report error but carry on */
3227 			device_printf(sc->dev,
3228 			    "failed to set congestion manager context for "
3229 			    "ingress queue %d: %d\n", iq->cntxt_id, rc);
3230 		}
3231 	}
3232 
3233 	/* Enable IQ interrupts */
3234 	atomic_store_rel_int(&iq->state, IQS_IDLE);
3235 	t4_write_reg(sc, sc->sge_gts_reg, V_SEINTARM(iq->intr_params) |
3236 	    V_INGRESSQID(iq->cntxt_id));
3237 
3238 	return (0);
3239 }
3240 
3241 static int
3242 free_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl)
3243 {
3244 	int rc;
3245 	struct adapter *sc = iq->adapter;
3246 	device_t dev;
3247 
3248 	if (sc == NULL)
3249 		return (0);	/* nothing to do */
3250 
3251 	dev = vi ? vi->dev : sc->dev;
3252 
3253 	if (iq->flags & IQ_ALLOCATED) {
3254 		rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
3255 		    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
3256 		    fl ? fl->cntxt_id : 0xffff, 0xffff);
3257 		if (rc != 0) {
3258 			device_printf(dev,
3259 			    "failed to free queue %p: %d\n", iq, rc);
3260 			return (rc);
3261 		}
3262 		iq->flags &= ~IQ_ALLOCATED;
3263 	}
3264 
3265 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
3266 
3267 	bzero(iq, sizeof(*iq));
3268 
3269 	if (fl) {
3270 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
3271 		    fl->desc);
3272 
3273 		if (fl->sdesc)
3274 			free_fl_sdesc(sc, fl);
3275 
3276 		if (mtx_initialized(&fl->fl_lock))
3277 			mtx_destroy(&fl->fl_lock);
3278 
3279 		bzero(fl, sizeof(*fl));
3280 	}
3281 
3282 	return (0);
3283 }
3284 
3285 static void
3286 add_iq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
3287     struct sge_iq *iq)
3288 {
3289 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3290 
3291 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD, &iq->ba,
3292 	    "bus address of descriptor ring");
3293 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3294 	    iq->qsize * IQ_ESIZE, "descriptor ring size in bytes");
3295 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3296 	    CTLTYPE_INT | CTLFLAG_RD, &iq->abs_id, 0, sysctl_uint16, "I",
3297 	    "absolute id of the queue");
3298 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3299 	    CTLTYPE_INT | CTLFLAG_RD, &iq->cntxt_id, 0, sysctl_uint16, "I",
3300 	    "SGE context id of the queue");
3301 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3302 	    CTLTYPE_INT | CTLFLAG_RD, &iq->cidx, 0, sysctl_uint16, "I",
3303 	    "consumer index");
3304 }
3305 
3306 static void
3307 add_fl_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
3308     struct sysctl_oid *oid, struct sge_fl *fl)
3309 {
3310 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3311 
3312 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3313 	    "freelist");
3314 	children = SYSCTL_CHILDREN(oid);
3315 
3316 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3317 	    &fl->ba, "bus address of descriptor ring");
3318 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3319 	    fl->sidx * EQ_ESIZE + sc->params.sge.spg_len,
3320 	    "desc ring size in bytes");
3321 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3322 	    CTLTYPE_INT | CTLFLAG_RD, &fl->cntxt_id, 0, sysctl_uint16, "I",
3323 	    "SGE context id of the freelist");
3324 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL,
3325 	    fl_pad ? 1 : 0, "padding enabled");
3326 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL,
3327 	    fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled");
3328 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
3329 	    0, "consumer index");
3330 	if (fl->flags & FL_BUF_PACKING) {
3331 		SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
3332 		    CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
3333 	}
3334 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
3335 	    0, "producer index");
3336 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_allocated",
3337 	    CTLFLAG_RD, &fl->mbuf_allocated, "# of mbuf allocated");
3338 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_inlined",
3339 	    CTLFLAG_RD, &fl->mbuf_inlined, "# of mbuf inlined in clusters");
3340 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
3341 	    CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
3342 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
3343 	    CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
3344 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
3345 	    CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
3346 }
3347 
3348 static int
3349 alloc_fwq(struct adapter *sc)
3350 {
3351 	int rc, intr_idx;
3352 	struct sge_iq *fwq = &sc->sge.fwq;
3353 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
3354 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3355 
3356 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE);
3357 	if (sc->flags & IS_VF)
3358 		intr_idx = 0;
3359 	else
3360 		intr_idx = sc->intr_count > 1 ? 1 : 0;
3361 	rc = alloc_iq_fl(&sc->port[0]->vi[0], fwq, NULL, intr_idx, -1);
3362 	if (rc != 0) {
3363 		device_printf(sc->dev,
3364 		    "failed to create firmware event queue: %d\n", rc);
3365 		return (rc);
3366 	}
3367 
3368 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD,
3369 	    NULL, "firmware event queue");
3370 	add_iq_sysctls(&sc->ctx, oid, fwq);
3371 
3372 	return (0);
3373 }
3374 
3375 static int
3376 free_fwq(struct adapter *sc)
3377 {
3378 	return free_iq_fl(NULL, &sc->sge.fwq, NULL);
3379 }
3380 
3381 static int
3382 alloc_ctrlq(struct adapter *sc, struct sge_wrq *ctrlq, int idx,
3383     struct sysctl_oid *oid)
3384 {
3385 	int rc;
3386 	char name[16];
3387 	struct sysctl_oid_list *children;
3388 
3389 	snprintf(name, sizeof(name), "%s ctrlq%d", device_get_nameunit(sc->dev),
3390 	    idx);
3391 	init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[idx]->tx_chan,
3392 	    sc->sge.fwq.cntxt_id, name);
3393 
3394 	children = SYSCTL_CHILDREN(oid);
3395 	snprintf(name, sizeof(name), "%d", idx);
3396 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3397 	    NULL, "ctrl queue");
3398 	rc = alloc_wrq(sc, NULL, ctrlq, oid);
3399 
3400 	return (rc);
3401 }
3402 
3403 int
3404 tnl_cong(struct port_info *pi, int drop)
3405 {
3406 
3407 	if (drop == -1)
3408 		return (-1);
3409 	else if (drop == 1)
3410 		return (0);
3411 	else
3412 		return (pi->rx_e_chan_map);
3413 }
3414 
3415 static int
3416 alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int intr_idx, int idx,
3417     struct sysctl_oid *oid)
3418 {
3419 	int rc;
3420 	struct adapter *sc = vi->pi->adapter;
3421 	struct sysctl_oid_list *children;
3422 	char name[16];
3423 
3424 	rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, intr_idx,
3425 	    tnl_cong(vi->pi, cong_drop));
3426 	if (rc != 0)
3427 		return (rc);
3428 
3429 	if (idx == 0)
3430 		sc->sge.iq_base = rxq->iq.abs_id - rxq->iq.cntxt_id;
3431 	else
3432 		KASSERT(rxq->iq.cntxt_id + sc->sge.iq_base == rxq->iq.abs_id,
3433 		    ("iq_base mismatch"));
3434 	KASSERT(sc->sge.iq_base == 0 || sc->flags & IS_VF,
3435 	    ("PF with non-zero iq_base"));
3436 
3437 	/*
3438 	 * The freelist is just barely above the starvation threshold right now,
3439 	 * fill it up a bit more.
3440 	 */
3441 	FL_LOCK(&rxq->fl);
3442 	refill_fl(sc, &rxq->fl, 128);
3443 	FL_UNLOCK(&rxq->fl);
3444 
3445 #if defined(INET) || defined(INET6)
3446 	rc = tcp_lro_init_args(&rxq->lro, vi->ifp, lro_entries, lro_mbufs);
3447 	if (rc != 0)
3448 		return (rc);
3449 	MPASS(rxq->lro.ifp == vi->ifp);	/* also indicates LRO init'ed */
3450 
3451 	if (vi->ifp->if_capenable & IFCAP_LRO)
3452 		rxq->iq.flags |= IQ_LRO_ENABLED;
3453 #endif
3454 	if (vi->ifp->if_capenable & IFCAP_HWRXTSTMP)
3455 		rxq->iq.flags |= IQ_RX_TIMESTAMP;
3456 	rxq->ifp = vi->ifp;
3457 
3458 	children = SYSCTL_CHILDREN(oid);
3459 
3460 	snprintf(name, sizeof(name), "%d", idx);
3461 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3462 	    NULL, "rx queue");
3463 	children = SYSCTL_CHILDREN(oid);
3464 
3465 	add_iq_sysctls(&vi->ctx, oid, &rxq->iq);
3466 #if defined(INET) || defined(INET6)
3467 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
3468 	    &rxq->lro.lro_queued, 0, NULL);
3469 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
3470 	    &rxq->lro.lro_flushed, 0, NULL);
3471 #endif
3472 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
3473 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
3474 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_extraction",
3475 	    CTLFLAG_RD, &rxq->vlan_extraction,
3476 	    "# of times hardware extracted 802.1Q tag");
3477 
3478 	add_fl_sysctls(sc, &vi->ctx, oid, &rxq->fl);
3479 
3480 	return (rc);
3481 }
3482 
3483 static int
3484 free_rxq(struct vi_info *vi, struct sge_rxq *rxq)
3485 {
3486 	int rc;
3487 
3488 #if defined(INET) || defined(INET6)
3489 	if (rxq->lro.ifp) {
3490 		tcp_lro_free(&rxq->lro);
3491 		rxq->lro.ifp = NULL;
3492 	}
3493 #endif
3494 
3495 	rc = free_iq_fl(vi, &rxq->iq, &rxq->fl);
3496 	if (rc == 0)
3497 		bzero(rxq, sizeof(*rxq));
3498 
3499 	return (rc);
3500 }
3501 
3502 #ifdef TCP_OFFLOAD
3503 static int
3504 alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq,
3505     int intr_idx, int idx, struct sysctl_oid *oid)
3506 {
3507 	struct port_info *pi = vi->pi;
3508 	int rc;
3509 	struct sysctl_oid_list *children;
3510 	char name[16];
3511 
3512 	rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx, 0);
3513 	if (rc != 0)
3514 		return (rc);
3515 
3516 	children = SYSCTL_CHILDREN(oid);
3517 
3518 	snprintf(name, sizeof(name), "%d", idx);
3519 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3520 	    NULL, "rx queue");
3521 	add_iq_sysctls(&vi->ctx, oid, &ofld_rxq->iq);
3522 	add_fl_sysctls(pi->adapter, &vi->ctx, oid, &ofld_rxq->fl);
3523 
3524 	return (rc);
3525 }
3526 
3527 static int
3528 free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq)
3529 {
3530 	int rc;
3531 
3532 	rc = free_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl);
3533 	if (rc == 0)
3534 		bzero(ofld_rxq, sizeof(*ofld_rxq));
3535 
3536 	return (rc);
3537 }
3538 #endif
3539 
3540 #ifdef DEV_NETMAP
3541 static int
3542 alloc_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int intr_idx,
3543     int idx, struct sysctl_oid *oid)
3544 {
3545 	int rc;
3546 	struct sysctl_oid_list *children;
3547 	struct sysctl_ctx_list *ctx;
3548 	char name[16];
3549 	size_t len;
3550 	struct adapter *sc = vi->pi->adapter;
3551 	struct netmap_adapter *na = NA(vi->ifp);
3552 
3553 	MPASS(na != NULL);
3554 
3555 	len = vi->qsize_rxq * IQ_ESIZE;
3556 	rc = alloc_ring(sc, len, &nm_rxq->iq_desc_tag, &nm_rxq->iq_desc_map,
3557 	    &nm_rxq->iq_ba, (void **)&nm_rxq->iq_desc);
3558 	if (rc != 0)
3559 		return (rc);
3560 
3561 	len = na->num_rx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3562 	rc = alloc_ring(sc, len, &nm_rxq->fl_desc_tag, &nm_rxq->fl_desc_map,
3563 	    &nm_rxq->fl_ba, (void **)&nm_rxq->fl_desc);
3564 	if (rc != 0)
3565 		return (rc);
3566 
3567 	nm_rxq->vi = vi;
3568 	nm_rxq->nid = idx;
3569 	nm_rxq->iq_cidx = 0;
3570 	nm_rxq->iq_sidx = vi->qsize_rxq - sc->params.sge.spg_len / IQ_ESIZE;
3571 	nm_rxq->iq_gen = F_RSPD_GEN;
3572 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
3573 	nm_rxq->fl_sidx = na->num_rx_desc;
3574 	nm_rxq->intr_idx = intr_idx;
3575 	nm_rxq->iq_cntxt_id = INVALID_NM_RXQ_CNTXT_ID;
3576 
3577 	ctx = &vi->ctx;
3578 	children = SYSCTL_CHILDREN(oid);
3579 
3580 	snprintf(name, sizeof(name), "%d", idx);
3581 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL,
3582 	    "rx queue");
3583 	children = SYSCTL_CHILDREN(oid);
3584 
3585 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3586 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_abs_id, 0, sysctl_uint16,
3587 	    "I", "absolute id of the queue");
3588 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3589 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cntxt_id, 0, sysctl_uint16,
3590 	    "I", "SGE context id of the queue");
3591 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3592 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cidx, 0, sysctl_uint16, "I",
3593 	    "consumer index");
3594 
3595 	children = SYSCTL_CHILDREN(oid);
3596 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3597 	    "freelist");
3598 	children = SYSCTL_CHILDREN(oid);
3599 
3600 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3601 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->fl_cntxt_id, 0, sysctl_uint16,
3602 	    "I", "SGE context id of the freelist");
3603 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
3604 	    &nm_rxq->fl_cidx, 0, "consumer index");
3605 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
3606 	    &nm_rxq->fl_pidx, 0, "producer index");
3607 
3608 	return (rc);
3609 }
3610 
3611 
3612 static int
3613 free_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
3614 {
3615 	struct adapter *sc = vi->pi->adapter;
3616 
3617 	if (vi->flags & VI_INIT_DONE)
3618 		MPASS(nm_rxq->iq_cntxt_id == INVALID_NM_RXQ_CNTXT_ID);
3619 	else
3620 		MPASS(nm_rxq->iq_cntxt_id == 0);
3621 
3622 	free_ring(sc, nm_rxq->iq_desc_tag, nm_rxq->iq_desc_map, nm_rxq->iq_ba,
3623 	    nm_rxq->iq_desc);
3624 	free_ring(sc, nm_rxq->fl_desc_tag, nm_rxq->fl_desc_map, nm_rxq->fl_ba,
3625 	    nm_rxq->fl_desc);
3626 
3627 	return (0);
3628 }
3629 
3630 static int
3631 alloc_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq, int iqidx, int idx,
3632     struct sysctl_oid *oid)
3633 {
3634 	int rc;
3635 	size_t len;
3636 	struct port_info *pi = vi->pi;
3637 	struct adapter *sc = pi->adapter;
3638 	struct netmap_adapter *na = NA(vi->ifp);
3639 	char name[16];
3640 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3641 
3642 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3643 	rc = alloc_ring(sc, len, &nm_txq->desc_tag, &nm_txq->desc_map,
3644 	    &nm_txq->ba, (void **)&nm_txq->desc);
3645 	if (rc)
3646 		return (rc);
3647 
3648 	nm_txq->pidx = nm_txq->cidx = 0;
3649 	nm_txq->sidx = na->num_tx_desc;
3650 	nm_txq->nid = idx;
3651 	nm_txq->iqidx = iqidx;
3652 	nm_txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3653 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(sc->pf) |
3654 	    V_TXPKT_VF(vi->vin) | V_TXPKT_VF_VLD(vi->vfvld));
3655 	nm_txq->cntxt_id = INVALID_NM_TXQ_CNTXT_ID;
3656 
3657 	snprintf(name, sizeof(name), "%d", idx);
3658 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3659 	    NULL, "netmap tx queue");
3660 	children = SYSCTL_CHILDREN(oid);
3661 
3662 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3663 	    &nm_txq->cntxt_id, 0, "SGE context id of the queue");
3664 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3665 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->cidx, 0, sysctl_uint16, "I",
3666 	    "consumer index");
3667 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3668 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->pidx, 0, sysctl_uint16, "I",
3669 	    "producer index");
3670 
3671 	return (rc);
3672 }
3673 
3674 static int
3675 free_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
3676 {
3677 	struct adapter *sc = vi->pi->adapter;
3678 
3679 	if (vi->flags & VI_INIT_DONE)
3680 		MPASS(nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID);
3681 	else
3682 		MPASS(nm_txq->cntxt_id == 0);
3683 
3684 	free_ring(sc, nm_txq->desc_tag, nm_txq->desc_map, nm_txq->ba,
3685 	    nm_txq->desc);
3686 
3687 	return (0);
3688 }
3689 #endif
3690 
3691 /*
3692  * Returns a reasonable automatic cidx flush threshold for a given queue size.
3693  */
3694 static u_int
3695 qsize_to_fthresh(int qsize)
3696 {
3697 	u_int fthresh;
3698 
3699 	while (!powerof2(qsize))
3700 		qsize++;
3701 	fthresh = ilog2(qsize);
3702 	if (fthresh > X_CIDXFLUSHTHRESH_128)
3703 		fthresh = X_CIDXFLUSHTHRESH_128;
3704 
3705 	return (fthresh);
3706 }
3707 
3708 static int
3709 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
3710 {
3711 	int rc, cntxt_id;
3712 	struct fw_eq_ctrl_cmd c;
3713 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3714 
3715 	bzero(&c, sizeof(c));
3716 
3717 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
3718 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
3719 	    V_FW_EQ_CTRL_CMD_VFN(0));
3720 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
3721 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
3722 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid));
3723 	c.physeqid_pkd = htobe32(0);
3724 	c.fetchszm_to_iqid =
3725 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
3726 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
3727 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
3728 	c.dcaen_to_eqsize =
3729 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3730 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3731 		V_FW_EQ_CTRL_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
3732 		V_FW_EQ_CTRL_CMD_EQSIZE(qsize));
3733 	c.eqaddr = htobe64(eq->ba);
3734 
3735 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3736 	if (rc != 0) {
3737 		device_printf(sc->dev,
3738 		    "failed to create control queue %d: %d\n", eq->tx_chan, rc);
3739 		return (rc);
3740 	}
3741 	eq->flags |= EQ_ALLOCATED;
3742 
3743 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
3744 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3745 	if (cntxt_id >= sc->sge.neq)
3746 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3747 		cntxt_id, sc->sge.neq - 1);
3748 	sc->sge.eqmap[cntxt_id] = eq;
3749 
3750 	return (rc);
3751 }
3752 
3753 static int
3754 eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3755 {
3756 	int rc, cntxt_id;
3757 	struct fw_eq_eth_cmd c;
3758 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3759 
3760 	bzero(&c, sizeof(c));
3761 
3762 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
3763 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
3764 	    V_FW_EQ_ETH_CMD_VFN(0));
3765 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
3766 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
3767 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
3768 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
3769 	c.fetchszm_to_iqid =
3770 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3771 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
3772 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
3773 	c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3774 	    V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3775 	    V_FW_EQ_ETH_CMD_EQSIZE(qsize));
3776 	c.eqaddr = htobe64(eq->ba);
3777 
3778 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3779 	if (rc != 0) {
3780 		device_printf(vi->dev,
3781 		    "failed to create Ethernet egress queue: %d\n", rc);
3782 		return (rc);
3783 	}
3784 	eq->flags |= EQ_ALLOCATED;
3785 
3786 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
3787 	eq->abs_id = G_FW_EQ_ETH_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
3788 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3789 	if (cntxt_id >= sc->sge.neq)
3790 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3791 		cntxt_id, sc->sge.neq - 1);
3792 	sc->sge.eqmap[cntxt_id] = eq;
3793 
3794 	return (rc);
3795 }
3796 
3797 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3798 static int
3799 ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3800 {
3801 	int rc, cntxt_id;
3802 	struct fw_eq_ofld_cmd c;
3803 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3804 
3805 	bzero(&c, sizeof(c));
3806 
3807 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
3808 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
3809 	    V_FW_EQ_OFLD_CMD_VFN(0));
3810 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
3811 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
3812 	c.fetchszm_to_iqid =
3813 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
3814 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
3815 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
3816 	c.dcaen_to_eqsize =
3817 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3818 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3819 		V_FW_EQ_OFLD_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
3820 		V_FW_EQ_OFLD_CMD_EQSIZE(qsize));
3821 	c.eqaddr = htobe64(eq->ba);
3822 
3823 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3824 	if (rc != 0) {
3825 		device_printf(vi->dev,
3826 		    "failed to create egress queue for TCP offload: %d\n", rc);
3827 		return (rc);
3828 	}
3829 	eq->flags |= EQ_ALLOCATED;
3830 
3831 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
3832 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3833 	if (cntxt_id >= sc->sge.neq)
3834 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3835 		cntxt_id, sc->sge.neq - 1);
3836 	sc->sge.eqmap[cntxt_id] = eq;
3837 
3838 	return (rc);
3839 }
3840 #endif
3841 
3842 static int
3843 alloc_eq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3844 {
3845 	int rc, qsize;
3846 	size_t len;
3847 
3848 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
3849 
3850 	qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3851 	len = qsize * EQ_ESIZE;
3852 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
3853 	    &eq->ba, (void **)&eq->desc);
3854 	if (rc)
3855 		return (rc);
3856 
3857 	eq->pidx = eq->cidx = eq->dbidx = 0;
3858 	/* Note that equeqidx is not used with sge_wrq (OFLD/CTRL) queues. */
3859 	eq->equeqidx = 0;
3860 	eq->doorbells = sc->doorbells;
3861 
3862 	switch (eq->flags & EQ_TYPEMASK) {
3863 	case EQ_CTRL:
3864 		rc = ctrl_eq_alloc(sc, eq);
3865 		break;
3866 
3867 	case EQ_ETH:
3868 		rc = eth_eq_alloc(sc, vi, eq);
3869 		break;
3870 
3871 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3872 	case EQ_OFLD:
3873 		rc = ofld_eq_alloc(sc, vi, eq);
3874 		break;
3875 #endif
3876 
3877 	default:
3878 		panic("%s: invalid eq type %d.", __func__,
3879 		    eq->flags & EQ_TYPEMASK);
3880 	}
3881 	if (rc != 0) {
3882 		device_printf(sc->dev,
3883 		    "failed to allocate egress queue(%d): %d\n",
3884 		    eq->flags & EQ_TYPEMASK, rc);
3885 	}
3886 
3887 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
3888 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
3889 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
3890 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3891 		uint32_t mask = (1 << s_qpp) - 1;
3892 		volatile uint8_t *udb;
3893 
3894 		udb = sc->udbs_base + UDBS_DB_OFFSET;
3895 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
3896 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
3897 		if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
3898 	    		clrbit(&eq->doorbells, DOORBELL_WCWR);
3899 		else {
3900 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
3901 			eq->udb_qid = 0;
3902 		}
3903 		eq->udb = (volatile void *)udb;
3904 	}
3905 
3906 	return (rc);
3907 }
3908 
3909 static int
3910 free_eq(struct adapter *sc, struct sge_eq *eq)
3911 {
3912 	int rc;
3913 
3914 	if (eq->flags & EQ_ALLOCATED) {
3915 		switch (eq->flags & EQ_TYPEMASK) {
3916 		case EQ_CTRL:
3917 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
3918 			    eq->cntxt_id);
3919 			break;
3920 
3921 		case EQ_ETH:
3922 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
3923 			    eq->cntxt_id);
3924 			break;
3925 
3926 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3927 		case EQ_OFLD:
3928 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
3929 			    eq->cntxt_id);
3930 			break;
3931 #endif
3932 
3933 		default:
3934 			panic("%s: invalid eq type %d.", __func__,
3935 			    eq->flags & EQ_TYPEMASK);
3936 		}
3937 		if (rc != 0) {
3938 			device_printf(sc->dev,
3939 			    "failed to free egress queue (%d): %d\n",
3940 			    eq->flags & EQ_TYPEMASK, rc);
3941 			return (rc);
3942 		}
3943 		eq->flags &= ~EQ_ALLOCATED;
3944 	}
3945 
3946 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
3947 
3948 	if (mtx_initialized(&eq->eq_lock))
3949 		mtx_destroy(&eq->eq_lock);
3950 
3951 	bzero(eq, sizeof(*eq));
3952 	return (0);
3953 }
3954 
3955 static int
3956 alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq,
3957     struct sysctl_oid *oid)
3958 {
3959 	int rc;
3960 	struct sysctl_ctx_list *ctx = vi ? &vi->ctx : &sc->ctx;
3961 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3962 
3963 	rc = alloc_eq(sc, vi, &wrq->eq);
3964 	if (rc)
3965 		return (rc);
3966 
3967 	wrq->adapter = sc;
3968 	TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq);
3969 	TAILQ_INIT(&wrq->incomplete_wrs);
3970 	STAILQ_INIT(&wrq->wr_list);
3971 	wrq->nwr_pending = 0;
3972 	wrq->ndesc_needed = 0;
3973 
3974 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3975 	    &wrq->eq.ba, "bus address of descriptor ring");
3976 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3977 	    wrq->eq.sidx * EQ_ESIZE + sc->params.sge.spg_len,
3978 	    "desc ring size in bytes");
3979 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3980 	    &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
3981 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3982 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I",
3983 	    "consumer index");
3984 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx",
3985 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I",
3986 	    "producer index");
3987 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
3988 	    wrq->eq.sidx, "status page index");
3989 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD,
3990 	    &wrq->tx_wrs_direct, "# of work requests (direct)");
3991 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD,
3992 	    &wrq->tx_wrs_copied, "# of work requests (copied)");
3993 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_sspace", CTLFLAG_RD,
3994 	    &wrq->tx_wrs_ss, "# of work requests (copied from scratch space)");
3995 
3996 	return (rc);
3997 }
3998 
3999 static int
4000 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
4001 {
4002 	int rc;
4003 
4004 	rc = free_eq(sc, &wrq->eq);
4005 	if (rc)
4006 		return (rc);
4007 
4008 	bzero(wrq, sizeof(*wrq));
4009 	return (0);
4010 }
4011 
4012 static int
4013 alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx,
4014     struct sysctl_oid *oid)
4015 {
4016 	int rc;
4017 	struct port_info *pi = vi->pi;
4018 	struct adapter *sc = pi->adapter;
4019 	struct sge_eq *eq = &txq->eq;
4020 	char name[16];
4021 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
4022 
4023 	rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx, can_resume_eth_tx,
4024 	    M_CXGBE, M_WAITOK);
4025 	if (rc != 0) {
4026 		device_printf(sc->dev, "failed to allocate mp_ring: %d\n", rc);
4027 		return (rc);
4028 	}
4029 
4030 	rc = alloc_eq(sc, vi, eq);
4031 	if (rc != 0) {
4032 		mp_ring_free(txq->r);
4033 		txq->r = NULL;
4034 		return (rc);
4035 	}
4036 
4037 	/* Can't fail after this point. */
4038 
4039 	if (idx == 0)
4040 		sc->sge.eq_base = eq->abs_id - eq->cntxt_id;
4041 	else
4042 		KASSERT(eq->cntxt_id + sc->sge.eq_base == eq->abs_id,
4043 		    ("eq_base mismatch"));
4044 	KASSERT(sc->sge.eq_base == 0 || sc->flags & IS_VF,
4045 	    ("PF with non-zero eq_base"));
4046 
4047 	TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq);
4048 	txq->ifp = vi->ifp;
4049 	txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK);
4050 	if (sc->flags & IS_VF)
4051 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
4052 		    V_TXPKT_INTF(pi->tx_chan));
4053 	else
4054 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
4055 		    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(sc->pf) |
4056 		    V_TXPKT_VF(vi->vin) | V_TXPKT_VF_VLD(vi->vfvld));
4057 	txq->tc_idx = -1;
4058 	txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE,
4059 	    M_ZERO | M_WAITOK);
4060 
4061 	snprintf(name, sizeof(name), "%d", idx);
4062 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
4063 	    NULL, "tx queue");
4064 	children = SYSCTL_CHILDREN(oid);
4065 
4066 	SYSCTL_ADD_UAUTO(&vi->ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
4067 	    &eq->ba, "bus address of descriptor ring");
4068 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
4069 	    eq->sidx * EQ_ESIZE + sc->params.sge.spg_len,
4070 	    "desc ring size in bytes");
4071 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
4072 	    &eq->abs_id, 0, "absolute id of the queue");
4073 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
4074 	    &eq->cntxt_id, 0, "SGE context id of the queue");
4075 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
4076 	    CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I",
4077 	    "consumer index");
4078 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
4079 	    CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I",
4080 	    "producer index");
4081 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
4082 	    eq->sidx, "status page index");
4083 
4084 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "tc",
4085 	    CTLTYPE_INT | CTLFLAG_RW, vi, idx, sysctl_tc, "I",
4086 	    "traffic class (-1 means none)");
4087 
4088 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
4089 	    &txq->txcsum, "# of times hardware assisted with checksum");
4090 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_insertion",
4091 	    CTLFLAG_RD, &txq->vlan_insertion,
4092 	    "# of times hardware inserted 802.1Q tag");
4093 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
4094 	    &txq->tso_wrs, "# of TSO work requests");
4095 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
4096 	    &txq->imm_wrs, "# of work requests with immediate data");
4097 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
4098 	    &txq->sgl_wrs, "# of work requests with direct SGL");
4099 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
4100 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
4101 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_wrs",
4102 	    CTLFLAG_RD, &txq->txpkts0_wrs,
4103 	    "# of txpkts (type 0) work requests");
4104 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_wrs",
4105 	    CTLFLAG_RD, &txq->txpkts1_wrs,
4106 	    "# of txpkts (type 1) work requests");
4107 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_pkts",
4108 	    CTLFLAG_RD, &txq->txpkts0_pkts,
4109 	    "# of frames tx'd using type0 txpkts work requests");
4110 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_pkts",
4111 	    CTLFLAG_RD, &txq->txpkts1_pkts,
4112 	    "# of frames tx'd using type1 txpkts work requests");
4113 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "raw_wrs", CTLFLAG_RD,
4114 	    &txq->raw_wrs, "# of raw work requests (non-packets)");
4115 
4116 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_enqueues",
4117 	    CTLFLAG_RD, &txq->r->enqueues,
4118 	    "# of enqueues to the mp_ring for this queue");
4119 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_drops",
4120 	    CTLFLAG_RD, &txq->r->drops,
4121 	    "# of drops in the mp_ring for this queue");
4122 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_starts",
4123 	    CTLFLAG_RD, &txq->r->starts,
4124 	    "# of normal consumer starts in the mp_ring for this queue");
4125 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_stalls",
4126 	    CTLFLAG_RD, &txq->r->stalls,
4127 	    "# of consumer stalls in the mp_ring for this queue");
4128 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_restarts",
4129 	    CTLFLAG_RD, &txq->r->restarts,
4130 	    "# of consumer restarts in the mp_ring for this queue");
4131 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_abdications",
4132 	    CTLFLAG_RD, &txq->r->abdications,
4133 	    "# of consumer abdications in the mp_ring for this queue");
4134 
4135 	return (0);
4136 }
4137 
4138 static int
4139 free_txq(struct vi_info *vi, struct sge_txq *txq)
4140 {
4141 	int rc;
4142 	struct adapter *sc = vi->pi->adapter;
4143 	struct sge_eq *eq = &txq->eq;
4144 
4145 	rc = free_eq(sc, eq);
4146 	if (rc)
4147 		return (rc);
4148 
4149 	sglist_free(txq->gl);
4150 	free(txq->sdesc, M_CXGBE);
4151 	mp_ring_free(txq->r);
4152 
4153 	bzero(txq, sizeof(*txq));
4154 	return (0);
4155 }
4156 
4157 static void
4158 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
4159 {
4160 	bus_addr_t *ba = arg;
4161 
4162 	KASSERT(nseg == 1,
4163 	    ("%s meant for single segment mappings only.", __func__));
4164 
4165 	*ba = error ? 0 : segs->ds_addr;
4166 }
4167 
4168 static inline void
4169 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
4170 {
4171 	uint32_t n, v;
4172 
4173 	n = IDXDIFF(fl->pidx / 8, fl->dbidx, fl->sidx);
4174 	MPASS(n > 0);
4175 
4176 	wmb();
4177 	v = fl->dbval | V_PIDX(n);
4178 	if (fl->udb)
4179 		*fl->udb = htole32(v);
4180 	else
4181 		t4_write_reg(sc, sc->sge_kdoorbell_reg, v);
4182 	IDXINCR(fl->dbidx, n, fl->sidx);
4183 }
4184 
4185 /*
4186  * Fills up the freelist by allocating up to 'n' buffers.  Buffers that are
4187  * recycled do not count towards this allocation budget.
4188  *
4189  * Returns non-zero to indicate that this freelist should be added to the list
4190  * of starving freelists.
4191  */
4192 static int
4193 refill_fl(struct adapter *sc, struct sge_fl *fl, int n)
4194 {
4195 	__be64 *d;
4196 	struct fl_sdesc *sd;
4197 	uintptr_t pa;
4198 	caddr_t cl;
4199 	struct cluster_layout *cll;
4200 	struct sw_zone_info *swz;
4201 	struct cluster_metadata *clm;
4202 	uint16_t max_pidx;
4203 	uint16_t hw_cidx = fl->hw_cidx;		/* stable snapshot */
4204 
4205 	FL_LOCK_ASSERT_OWNED(fl);
4206 
4207 	/*
4208 	 * We always stop at the beginning of the hardware descriptor that's just
4209 	 * before the one with the hw cidx.  This is to avoid hw pidx = hw cidx,
4210 	 * which would mean an empty freelist to the chip.
4211 	 */
4212 	max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1;
4213 	if (fl->pidx == max_pidx * 8)
4214 		return (0);
4215 
4216 	d = &fl->desc[fl->pidx];
4217 	sd = &fl->sdesc[fl->pidx];
4218 	cll = &fl->cll_def;	/* default layout */
4219 	swz = &sc->sge.sw_zone_info[cll->zidx];
4220 
4221 	while (n > 0) {
4222 
4223 		if (sd->cl != NULL) {
4224 
4225 			if (sd->nmbuf == 0) {
4226 				/*
4227 				 * Fast recycle without involving any atomics on
4228 				 * the cluster's metadata (if the cluster has
4229 				 * metadata).  This happens when all frames
4230 				 * received in the cluster were small enough to
4231 				 * fit within a single mbuf each.
4232 				 */
4233 				fl->cl_fast_recycled++;
4234 #ifdef INVARIANTS
4235 				clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
4236 				if (clm != NULL)
4237 					MPASS(clm->refcount == 1);
4238 #endif
4239 				goto recycled_fast;
4240 			}
4241 
4242 			/*
4243 			 * Cluster is guaranteed to have metadata.  Clusters
4244 			 * without metadata always take the fast recycle path
4245 			 * when they're recycled.
4246 			 */
4247 			clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
4248 			MPASS(clm != NULL);
4249 
4250 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4251 				fl->cl_recycled++;
4252 				counter_u64_add(extfree_rels, 1);
4253 				goto recycled;
4254 			}
4255 			sd->cl = NULL;	/* gave up my reference */
4256 		}
4257 		MPASS(sd->cl == NULL);
4258 alloc:
4259 		cl = uma_zalloc(swz->zone, M_NOWAIT);
4260 		if (__predict_false(cl == NULL)) {
4261 			if (cll == &fl->cll_alt || fl->cll_alt.zidx == -1 ||
4262 			    fl->cll_def.zidx == fl->cll_alt.zidx)
4263 				break;
4264 
4265 			/* fall back to the safe zone */
4266 			cll = &fl->cll_alt;
4267 			swz = &sc->sge.sw_zone_info[cll->zidx];
4268 			goto alloc;
4269 		}
4270 		fl->cl_allocated++;
4271 		n--;
4272 
4273 		pa = pmap_kextract((vm_offset_t)cl);
4274 		pa += cll->region1;
4275 		sd->cl = cl;
4276 		sd->cll = *cll;
4277 		*d = htobe64(pa | cll->hwidx);
4278 		clm = cl_metadata(sc, fl, cll, cl);
4279 		if (clm != NULL) {
4280 recycled:
4281 #ifdef INVARIANTS
4282 			clm->sd = sd;
4283 #endif
4284 			clm->refcount = 1;
4285 		}
4286 		sd->nmbuf = 0;
4287 recycled_fast:
4288 		d++;
4289 		sd++;
4290 		if (__predict_false(++fl->pidx % 8 == 0)) {
4291 			uint16_t pidx = fl->pidx / 8;
4292 
4293 			if (__predict_false(pidx == fl->sidx)) {
4294 				fl->pidx = 0;
4295 				pidx = 0;
4296 				sd = fl->sdesc;
4297 				d = fl->desc;
4298 			}
4299 			if (pidx == max_pidx)
4300 				break;
4301 
4302 			if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4)
4303 				ring_fl_db(sc, fl);
4304 		}
4305 	}
4306 
4307 	if (fl->pidx / 8 != fl->dbidx)
4308 		ring_fl_db(sc, fl);
4309 
4310 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
4311 }
4312 
4313 /*
4314  * Attempt to refill all starving freelists.
4315  */
4316 static void
4317 refill_sfl(void *arg)
4318 {
4319 	struct adapter *sc = arg;
4320 	struct sge_fl *fl, *fl_temp;
4321 
4322 	mtx_assert(&sc->sfl_lock, MA_OWNED);
4323 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
4324 		FL_LOCK(fl);
4325 		refill_fl(sc, fl, 64);
4326 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
4327 			TAILQ_REMOVE(&sc->sfl, fl, link);
4328 			fl->flags &= ~FL_STARVING;
4329 		}
4330 		FL_UNLOCK(fl);
4331 	}
4332 
4333 	if (!TAILQ_EMPTY(&sc->sfl))
4334 		callout_schedule(&sc->sfl_callout, hz / 5);
4335 }
4336 
4337 static int
4338 alloc_fl_sdesc(struct sge_fl *fl)
4339 {
4340 
4341 	fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc), M_CXGBE,
4342 	    M_ZERO | M_WAITOK);
4343 
4344 	return (0);
4345 }
4346 
4347 static void
4348 free_fl_sdesc(struct adapter *sc, struct sge_fl *fl)
4349 {
4350 	struct fl_sdesc *sd;
4351 	struct cluster_metadata *clm;
4352 	struct cluster_layout *cll;
4353 	int i;
4354 
4355 	sd = fl->sdesc;
4356 	for (i = 0; i < fl->sidx * 8; i++, sd++) {
4357 		if (sd->cl == NULL)
4358 			continue;
4359 
4360 		cll = &sd->cll;
4361 		clm = cl_metadata(sc, fl, cll, sd->cl);
4362 		if (sd->nmbuf == 0)
4363 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4364 		else if (clm && atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4365 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4366 			counter_u64_add(extfree_rels, 1);
4367 		}
4368 		sd->cl = NULL;
4369 	}
4370 
4371 	free(fl->sdesc, M_CXGBE);
4372 	fl->sdesc = NULL;
4373 }
4374 
4375 static inline void
4376 get_pkt_gl(struct mbuf *m, struct sglist *gl)
4377 {
4378 	int rc;
4379 
4380 	M_ASSERTPKTHDR(m);
4381 
4382 	sglist_reset(gl);
4383 	rc = sglist_append_mbuf(gl, m);
4384 	if (__predict_false(rc != 0)) {
4385 		panic("%s: mbuf %p (%d segs) was vetted earlier but now fails "
4386 		    "with %d.", __func__, m, mbuf_nsegs(m), rc);
4387 	}
4388 
4389 	KASSERT(gl->sg_nseg == mbuf_nsegs(m),
4390 	    ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m,
4391 	    mbuf_nsegs(m), gl->sg_nseg));
4392 	KASSERT(gl->sg_nseg > 0 &&
4393 	    gl->sg_nseg <= (needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS),
4394 	    ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__,
4395 		gl->sg_nseg, needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS));
4396 }
4397 
4398 /*
4399  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
4400  */
4401 static inline u_int
4402 txpkt_len16(u_int nsegs, u_int tso)
4403 {
4404 	u_int n;
4405 
4406 	MPASS(nsegs > 0);
4407 
4408 	nsegs--; /* first segment is part of ulptx_sgl */
4409 	n = sizeof(struct fw_eth_tx_pkt_wr) + sizeof(struct cpl_tx_pkt_core) +
4410 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4411 	if (tso)
4412 		n += sizeof(struct cpl_tx_pkt_lso_core);
4413 
4414 	return (howmany(n, 16));
4415 }
4416 
4417 /*
4418  * len16 for a txpkt_vm WR with a GL.  Includes the firmware work
4419  * request header.
4420  */
4421 static inline u_int
4422 txpkt_vm_len16(u_int nsegs, u_int tso)
4423 {
4424 	u_int n;
4425 
4426 	MPASS(nsegs > 0);
4427 
4428 	nsegs--; /* first segment is part of ulptx_sgl */
4429 	n = sizeof(struct fw_eth_tx_pkt_vm_wr) +
4430 	    sizeof(struct cpl_tx_pkt_core) +
4431 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4432 	if (tso)
4433 		n += sizeof(struct cpl_tx_pkt_lso_core);
4434 
4435 	return (howmany(n, 16));
4436 }
4437 
4438 /*
4439  * len16 for a txpkts type 0 WR with a GL.  Does not include the firmware work
4440  * request header.
4441  */
4442 static inline u_int
4443 txpkts0_len16(u_int nsegs)
4444 {
4445 	u_int n;
4446 
4447 	MPASS(nsegs > 0);
4448 
4449 	nsegs--; /* first segment is part of ulptx_sgl */
4450 	n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) +
4451 	    sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) +
4452 	    8 * ((3 * nsegs) / 2 + (nsegs & 1));
4453 
4454 	return (howmany(n, 16));
4455 }
4456 
4457 /*
4458  * len16 for a txpkts type 1 WR with a GL.  Does not include the firmware work
4459  * request header.
4460  */
4461 static inline u_int
4462 txpkts1_len16(void)
4463 {
4464 	u_int n;
4465 
4466 	n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl);
4467 
4468 	return (howmany(n, 16));
4469 }
4470 
4471 static inline u_int
4472 imm_payload(u_int ndesc)
4473 {
4474 	u_int n;
4475 
4476 	n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) -
4477 	    sizeof(struct cpl_tx_pkt_core);
4478 
4479 	return (n);
4480 }
4481 
4482 /*
4483  * Write a VM txpkt WR for this packet to the hardware descriptors, update the
4484  * software descriptor, and advance the pidx.  It is guaranteed that enough
4485  * descriptors are available.
4486  *
4487  * The return value is the # of hardware descriptors used.
4488  */
4489 static u_int
4490 write_txpkt_vm_wr(struct adapter *sc, struct sge_txq *txq,
4491     struct fw_eth_tx_pkt_vm_wr *wr, struct mbuf *m0, u_int available)
4492 {
4493 	struct sge_eq *eq = &txq->eq;
4494 	struct tx_sdesc *txsd;
4495 	struct cpl_tx_pkt_core *cpl;
4496 	uint32_t ctrl;	/* used in many unrelated places */
4497 	uint64_t ctrl1;
4498 	int csum_type, len16, ndesc, pktlen, nsegs;
4499 	caddr_t dst;
4500 
4501 	TXQ_LOCK_ASSERT_OWNED(txq);
4502 	M_ASSERTPKTHDR(m0);
4503 	MPASS(available > 0 && available < eq->sidx);
4504 
4505 	len16 = mbuf_len16(m0);
4506 	nsegs = mbuf_nsegs(m0);
4507 	pktlen = m0->m_pkthdr.len;
4508 	ctrl = sizeof(struct cpl_tx_pkt_core);
4509 	if (needs_tso(m0))
4510 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4511 	ndesc = howmany(len16, EQ_ESIZE / 16);
4512 	MPASS(ndesc <= available);
4513 
4514 	/* Firmware work request header */
4515 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4516 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_VM_WR) |
4517 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4518 
4519 	ctrl = V_FW_WR_LEN16(len16);
4520 	wr->equiq_to_len16 = htobe32(ctrl);
4521 	wr->r3[0] = 0;
4522 	wr->r3[1] = 0;
4523 
4524 	/*
4525 	 * Copy over ethmacdst, ethmacsrc, ethtype, and vlantci.
4526 	 * vlantci is ignored unless the ethtype is 0x8100, so it's
4527 	 * simpler to always copy it rather than making it
4528 	 * conditional.  Also, it seems that we do not have to set
4529 	 * vlantci or fake the ethtype when doing VLAN tag insertion.
4530 	 */
4531 	m_copydata(m0, 0, sizeof(struct ether_header) + 2, wr->ethmacdst);
4532 
4533 	csum_type = -1;
4534 	if (needs_tso(m0)) {
4535 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4536 
4537 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4538 		    m0->m_pkthdr.l4hlen > 0,
4539 		    ("%s: mbuf %p needs TSO but missing header lengths",
4540 			__func__, m0));
4541 
4542 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4543 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4544 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4545 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4546 			ctrl |= V_LSO_ETHHDR_LEN(1);
4547 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4548 			ctrl |= F_LSO_IPV6;
4549 
4550 		lso->lso_ctrl = htobe32(ctrl);
4551 		lso->ipid_ofst = htobe16(0);
4552 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4553 		lso->seqno_offset = htobe32(0);
4554 		lso->len = htobe32(pktlen);
4555 
4556 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4557 			csum_type = TX_CSUM_TCPIP6;
4558 		else
4559 			csum_type = TX_CSUM_TCPIP;
4560 
4561 		cpl = (void *)(lso + 1);
4562 
4563 		txq->tso_wrs++;
4564 	} else {
4565 		if (m0->m_pkthdr.csum_flags & CSUM_IP_TCP)
4566 			csum_type = TX_CSUM_TCPIP;
4567 		else if (m0->m_pkthdr.csum_flags & CSUM_IP_UDP)
4568 			csum_type = TX_CSUM_UDPIP;
4569 		else if (m0->m_pkthdr.csum_flags & CSUM_IP6_TCP)
4570 			csum_type = TX_CSUM_TCPIP6;
4571 		else if (m0->m_pkthdr.csum_flags & CSUM_IP6_UDP)
4572 			csum_type = TX_CSUM_UDPIP6;
4573 #if defined(INET)
4574 		else if (m0->m_pkthdr.csum_flags & CSUM_IP) {
4575 			/*
4576 			 * XXX: The firmware appears to stomp on the
4577 			 * fragment/flags field of the IP header when
4578 			 * using TX_CSUM_IP.  Fall back to doing
4579 			 * software checksums.
4580 			 */
4581 			u_short *sump;
4582 			struct mbuf *m;
4583 			int offset;
4584 
4585 			m = m0;
4586 			offset = 0;
4587 			sump = m_advance(&m, &offset, m0->m_pkthdr.l2hlen +
4588 			    offsetof(struct ip, ip_sum));
4589 			*sump = in_cksum_skip(m0, m0->m_pkthdr.l2hlen +
4590 			    m0->m_pkthdr.l3hlen, m0->m_pkthdr.l2hlen);
4591 			m0->m_pkthdr.csum_flags &= ~CSUM_IP;
4592 		}
4593 #endif
4594 
4595 		cpl = (void *)(wr + 1);
4596 	}
4597 
4598 	/* Checksum offload */
4599 	ctrl1 = 0;
4600 	if (needs_l3_csum(m0) == 0)
4601 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4602 	if (csum_type >= 0) {
4603 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0,
4604 	    ("%s: mbuf %p needs checksum offload but missing header lengths",
4605 			__func__, m0));
4606 
4607 		if (chip_id(sc) <= CHELSIO_T5) {
4608 			ctrl1 |= V_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4609 			    ETHER_HDR_LEN);
4610 		} else {
4611 			ctrl1 |= V_T6_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4612 			    ETHER_HDR_LEN);
4613 		}
4614 		ctrl1 |= V_TXPKT_IPHDR_LEN(m0->m_pkthdr.l3hlen);
4615 		ctrl1 |= V_TXPKT_CSUM_TYPE(csum_type);
4616 	} else
4617 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4618 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4619 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4620 		txq->txcsum++;	/* some hardware assistance provided */
4621 
4622 	/* VLAN tag insertion */
4623 	if (needs_vlan_insertion(m0)) {
4624 		ctrl1 |= F_TXPKT_VLAN_VLD |
4625 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4626 		txq->vlan_insertion++;
4627 	}
4628 
4629 	/* CPL header */
4630 	cpl->ctrl0 = txq->cpl_ctrl0;
4631 	cpl->pack = 0;
4632 	cpl->len = htobe16(pktlen);
4633 	cpl->ctrl1 = htobe64(ctrl1);
4634 
4635 	/* SGL */
4636 	dst = (void *)(cpl + 1);
4637 
4638 	/*
4639 	 * A packet using TSO will use up an entire descriptor for the
4640 	 * firmware work request header, LSO CPL, and TX_PKT_XT CPL.
4641 	 * If this descriptor is the last descriptor in the ring, wrap
4642 	 * around to the front of the ring explicitly for the start of
4643 	 * the sgl.
4644 	 */
4645 	if (dst == (void *)&eq->desc[eq->sidx]) {
4646 		dst = (void *)&eq->desc[0];
4647 		write_gl_to_txd(txq, m0, &dst, 0);
4648 	} else
4649 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4650 	txq->sgl_wrs++;
4651 
4652 	txq->txpkt_wrs++;
4653 
4654 	txsd = &txq->sdesc[eq->pidx];
4655 	txsd->m = m0;
4656 	txsd->desc_used = ndesc;
4657 
4658 	return (ndesc);
4659 }
4660 
4661 /*
4662  * Write a raw WR to the hardware descriptors, update the software
4663  * descriptor, and advance the pidx.  It is guaranteed that enough
4664  * descriptors are available.
4665  *
4666  * The return value is the # of hardware descriptors used.
4667  */
4668 static u_int
4669 write_raw_wr(struct sge_txq *txq, void *wr, struct mbuf *m0, u_int available)
4670 {
4671 	struct sge_eq *eq = &txq->eq;
4672 	struct tx_sdesc *txsd;
4673 	struct mbuf *m;
4674 	caddr_t dst;
4675 	int len16, ndesc;
4676 
4677 	len16 = mbuf_len16(m0);
4678 	ndesc = howmany(len16, EQ_ESIZE / 16);
4679 	MPASS(ndesc <= available);
4680 
4681 	dst = wr;
4682 	for (m = m0; m != NULL; m = m->m_next)
4683 		copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4684 
4685 	txq->raw_wrs++;
4686 
4687 	txsd = &txq->sdesc[eq->pidx];
4688 	txsd->m = m0;
4689 	txsd->desc_used = ndesc;
4690 
4691 	return (ndesc);
4692 }
4693 
4694 /*
4695  * Write a txpkt WR for this packet to the hardware descriptors, update the
4696  * software descriptor, and advance the pidx.  It is guaranteed that enough
4697  * descriptors are available.
4698  *
4699  * The return value is the # of hardware descriptors used.
4700  */
4701 static u_int
4702 write_txpkt_wr(struct sge_txq *txq, struct fw_eth_tx_pkt_wr *wr,
4703     struct mbuf *m0, u_int available)
4704 {
4705 	struct sge_eq *eq = &txq->eq;
4706 	struct tx_sdesc *txsd;
4707 	struct cpl_tx_pkt_core *cpl;
4708 	uint32_t ctrl;	/* used in many unrelated places */
4709 	uint64_t ctrl1;
4710 	int len16, ndesc, pktlen, nsegs;
4711 	caddr_t dst;
4712 
4713 	TXQ_LOCK_ASSERT_OWNED(txq);
4714 	M_ASSERTPKTHDR(m0);
4715 	MPASS(available > 0 && available < eq->sidx);
4716 
4717 	len16 = mbuf_len16(m0);
4718 	nsegs = mbuf_nsegs(m0);
4719 	pktlen = m0->m_pkthdr.len;
4720 	ctrl = sizeof(struct cpl_tx_pkt_core);
4721 	if (needs_tso(m0))
4722 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4723 	else if (pktlen <= imm_payload(2) && available >= 2) {
4724 		/* Immediate data.  Recalculate len16 and set nsegs to 0. */
4725 		ctrl += pktlen;
4726 		len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) +
4727 		    sizeof(struct cpl_tx_pkt_core) + pktlen, 16);
4728 		nsegs = 0;
4729 	}
4730 	ndesc = howmany(len16, EQ_ESIZE / 16);
4731 	MPASS(ndesc <= available);
4732 
4733 	/* Firmware work request header */
4734 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4735 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
4736 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4737 
4738 	ctrl = V_FW_WR_LEN16(len16);
4739 	wr->equiq_to_len16 = htobe32(ctrl);
4740 	wr->r3 = 0;
4741 
4742 	if (needs_tso(m0)) {
4743 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4744 
4745 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4746 		    m0->m_pkthdr.l4hlen > 0,
4747 		    ("%s: mbuf %p needs TSO but missing header lengths",
4748 			__func__, m0));
4749 
4750 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4751 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4752 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4753 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4754 			ctrl |= V_LSO_ETHHDR_LEN(1);
4755 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4756 			ctrl |= F_LSO_IPV6;
4757 
4758 		lso->lso_ctrl = htobe32(ctrl);
4759 		lso->ipid_ofst = htobe16(0);
4760 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4761 		lso->seqno_offset = htobe32(0);
4762 		lso->len = htobe32(pktlen);
4763 
4764 		cpl = (void *)(lso + 1);
4765 
4766 		txq->tso_wrs++;
4767 	} else
4768 		cpl = (void *)(wr + 1);
4769 
4770 	/* Checksum offload */
4771 	ctrl1 = 0;
4772 	if (needs_l3_csum(m0) == 0)
4773 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4774 	if (needs_l4_csum(m0) == 0)
4775 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4776 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4777 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4778 		txq->txcsum++;	/* some hardware assistance provided */
4779 
4780 	/* VLAN tag insertion */
4781 	if (needs_vlan_insertion(m0)) {
4782 		ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4783 		txq->vlan_insertion++;
4784 	}
4785 
4786 	/* CPL header */
4787 	cpl->ctrl0 = txq->cpl_ctrl0;
4788 	cpl->pack = 0;
4789 	cpl->len = htobe16(pktlen);
4790 	cpl->ctrl1 = htobe64(ctrl1);
4791 
4792 	/* SGL */
4793 	dst = (void *)(cpl + 1);
4794 	if (nsegs > 0) {
4795 
4796 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4797 		txq->sgl_wrs++;
4798 	} else {
4799 		struct mbuf *m;
4800 
4801 		for (m = m0; m != NULL; m = m->m_next) {
4802 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4803 #ifdef INVARIANTS
4804 			pktlen -= m->m_len;
4805 #endif
4806 		}
4807 #ifdef INVARIANTS
4808 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
4809 #endif
4810 		txq->imm_wrs++;
4811 	}
4812 
4813 	txq->txpkt_wrs++;
4814 
4815 	txsd = &txq->sdesc[eq->pidx];
4816 	txsd->m = m0;
4817 	txsd->desc_used = ndesc;
4818 
4819 	return (ndesc);
4820 }
4821 
4822 static int
4823 try_txpkts(struct mbuf *m, struct mbuf *n, struct txpkts *txp, u_int available)
4824 {
4825 	u_int needed, nsegs1, nsegs2, l1, l2;
4826 
4827 	if (cannot_use_txpkts(m) || cannot_use_txpkts(n))
4828 		return (1);
4829 
4830 	nsegs1 = mbuf_nsegs(m);
4831 	nsegs2 = mbuf_nsegs(n);
4832 	if (nsegs1 + nsegs2 == 2) {
4833 		txp->wr_type = 1;
4834 		l1 = l2 = txpkts1_len16();
4835 	} else {
4836 		txp->wr_type = 0;
4837 		l1 = txpkts0_len16(nsegs1);
4838 		l2 = txpkts0_len16(nsegs2);
4839 	}
4840 	txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) + l1 + l2;
4841 	needed = howmany(txp->len16, EQ_ESIZE / 16);
4842 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4843 		return (1);
4844 
4845 	txp->plen = m->m_pkthdr.len + n->m_pkthdr.len;
4846 	if (txp->plen > 65535)
4847 		return (1);
4848 
4849 	txp->npkt = 2;
4850 	set_mbuf_len16(m, l1);
4851 	set_mbuf_len16(n, l2);
4852 
4853 	return (0);
4854 }
4855 
4856 static int
4857 add_to_txpkts(struct mbuf *m, struct txpkts *txp, u_int available)
4858 {
4859 	u_int plen, len16, needed, nsegs;
4860 
4861 	MPASS(txp->wr_type == 0 || txp->wr_type == 1);
4862 
4863 	if (cannot_use_txpkts(m))
4864 		return (1);
4865 
4866 	nsegs = mbuf_nsegs(m);
4867 	if (txp->wr_type == 1 && nsegs != 1)
4868 		return (1);
4869 
4870 	plen = txp->plen + m->m_pkthdr.len;
4871 	if (plen > 65535)
4872 		return (1);
4873 
4874 	if (txp->wr_type == 0)
4875 		len16 = txpkts0_len16(nsegs);
4876 	else
4877 		len16 = txpkts1_len16();
4878 	needed = howmany(txp->len16 + len16, EQ_ESIZE / 16);
4879 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4880 		return (1);
4881 
4882 	txp->npkt++;
4883 	txp->plen = plen;
4884 	txp->len16 += len16;
4885 	set_mbuf_len16(m, len16);
4886 
4887 	return (0);
4888 }
4889 
4890 /*
4891  * Write a txpkts WR for the packets in txp to the hardware descriptors, update
4892  * the software descriptor, and advance the pidx.  It is guaranteed that enough
4893  * descriptors are available.
4894  *
4895  * The return value is the # of hardware descriptors used.
4896  */
4897 static u_int
4898 write_txpkts_wr(struct sge_txq *txq, struct fw_eth_tx_pkts_wr *wr,
4899     struct mbuf *m0, const struct txpkts *txp, u_int available)
4900 {
4901 	struct sge_eq *eq = &txq->eq;
4902 	struct tx_sdesc *txsd;
4903 	struct cpl_tx_pkt_core *cpl;
4904 	uint32_t ctrl;
4905 	uint64_t ctrl1;
4906 	int ndesc, checkwrap;
4907 	struct mbuf *m;
4908 	void *flitp;
4909 
4910 	TXQ_LOCK_ASSERT_OWNED(txq);
4911 	MPASS(txp->npkt > 0);
4912 	MPASS(txp->plen < 65536);
4913 	MPASS(m0 != NULL);
4914 	MPASS(m0->m_nextpkt != NULL);
4915 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
4916 	MPASS(available > 0 && available < eq->sidx);
4917 
4918 	ndesc = howmany(txp->len16, EQ_ESIZE / 16);
4919 	MPASS(ndesc <= available);
4920 
4921 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4922 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
4923 	ctrl = V_FW_WR_LEN16(txp->len16);
4924 	wr->equiq_to_len16 = htobe32(ctrl);
4925 	wr->plen = htobe16(txp->plen);
4926 	wr->npkt = txp->npkt;
4927 	wr->r3 = 0;
4928 	wr->type = txp->wr_type;
4929 	flitp = wr + 1;
4930 
4931 	/*
4932 	 * At this point we are 16B into a hardware descriptor.  If checkwrap is
4933 	 * set then we know the WR is going to wrap around somewhere.  We'll
4934 	 * check for that at appropriate points.
4935 	 */
4936 	checkwrap = eq->sidx - ndesc < eq->pidx;
4937 	for (m = m0; m != NULL; m = m->m_nextpkt) {
4938 		if (txp->wr_type == 0) {
4939 			struct ulp_txpkt *ulpmc;
4940 			struct ulptx_idata *ulpsc;
4941 
4942 			/* ULP master command */
4943 			ulpmc = flitp;
4944 			ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
4945 			    V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid));
4946 			ulpmc->len = htobe32(mbuf_len16(m));
4947 
4948 			/* ULP subcommand */
4949 			ulpsc = (void *)(ulpmc + 1);
4950 			ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
4951 			    F_ULP_TX_SC_MORE);
4952 			ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
4953 
4954 			cpl = (void *)(ulpsc + 1);
4955 			if (checkwrap &&
4956 			    (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx])
4957 				cpl = (void *)&eq->desc[0];
4958 		} else {
4959 			cpl = flitp;
4960 		}
4961 
4962 		/* Checksum offload */
4963 		ctrl1 = 0;
4964 		if (needs_l3_csum(m) == 0)
4965 			ctrl1 |= F_TXPKT_IPCSUM_DIS;
4966 		if (needs_l4_csum(m) == 0)
4967 			ctrl1 |= F_TXPKT_L4CSUM_DIS;
4968 		if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4969 		    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4970 			txq->txcsum++;	/* some hardware assistance provided */
4971 
4972 		/* VLAN tag insertion */
4973 		if (needs_vlan_insertion(m)) {
4974 			ctrl1 |= F_TXPKT_VLAN_VLD |
4975 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
4976 			txq->vlan_insertion++;
4977 		}
4978 
4979 		/* CPL header */
4980 		cpl->ctrl0 = txq->cpl_ctrl0;
4981 		cpl->pack = 0;
4982 		cpl->len = htobe16(m->m_pkthdr.len);
4983 		cpl->ctrl1 = htobe64(ctrl1);
4984 
4985 		flitp = cpl + 1;
4986 		if (checkwrap &&
4987 		    (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
4988 			flitp = (void *)&eq->desc[0];
4989 
4990 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap);
4991 
4992 	}
4993 
4994 	if (txp->wr_type == 0) {
4995 		txq->txpkts0_pkts += txp->npkt;
4996 		txq->txpkts0_wrs++;
4997 	} else {
4998 		txq->txpkts1_pkts += txp->npkt;
4999 		txq->txpkts1_wrs++;
5000 	}
5001 
5002 	txsd = &txq->sdesc[eq->pidx];
5003 	txsd->m = m0;
5004 	txsd->desc_used = ndesc;
5005 
5006 	return (ndesc);
5007 }
5008 
5009 /*
5010  * If the SGL ends on an address that is not 16 byte aligned, this function will
5011  * add a 0 filled flit at the end.
5012  */
5013 static void
5014 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap)
5015 {
5016 	struct sge_eq *eq = &txq->eq;
5017 	struct sglist *gl = txq->gl;
5018 	struct sglist_seg *seg;
5019 	__be64 *flitp, *wrap;
5020 	struct ulptx_sgl *usgl;
5021 	int i, nflits, nsegs;
5022 
5023 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
5024 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
5025 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
5026 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
5027 
5028 	get_pkt_gl(m, gl);
5029 	nsegs = gl->sg_nseg;
5030 	MPASS(nsegs > 0);
5031 
5032 	nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
5033 	flitp = (__be64 *)(*to);
5034 	wrap = (__be64 *)(&eq->desc[eq->sidx]);
5035 	seg = &gl->sg_segs[0];
5036 	usgl = (void *)flitp;
5037 
5038 	/*
5039 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
5040 	 * ring, so we're at least 16 bytes away from the status page.  There is
5041 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
5042 	 */
5043 
5044 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
5045 	    V_ULPTX_NSGE(nsegs));
5046 	usgl->len0 = htobe32(seg->ss_len);
5047 	usgl->addr0 = htobe64(seg->ss_paddr);
5048 	seg++;
5049 
5050 	if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) {
5051 
5052 		/* Won't wrap around at all */
5053 
5054 		for (i = 0; i < nsegs - 1; i++, seg++) {
5055 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
5056 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
5057 		}
5058 		if (i & 1)
5059 			usgl->sge[i / 2].len[1] = htobe32(0);
5060 		flitp += nflits;
5061 	} else {
5062 
5063 		/* Will wrap somewhere in the rest of the SGL */
5064 
5065 		/* 2 flits already written, write the rest flit by flit */
5066 		flitp = (void *)(usgl + 1);
5067 		for (i = 0; i < nflits - 2; i++) {
5068 			if (flitp == wrap)
5069 				flitp = (void *)eq->desc;
5070 			*flitp++ = get_flit(seg, nsegs - 1, i);
5071 		}
5072 	}
5073 
5074 	if (nflits & 1) {
5075 		MPASS(((uintptr_t)flitp) & 0xf);
5076 		*flitp++ = 0;
5077 	}
5078 
5079 	MPASS((((uintptr_t)flitp) & 0xf) == 0);
5080 	if (__predict_false(flitp == wrap))
5081 		*to = (void *)eq->desc;
5082 	else
5083 		*to = (void *)flitp;
5084 }
5085 
5086 static inline void
5087 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
5088 {
5089 
5090 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
5091 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
5092 
5093 	if (__predict_true((uintptr_t)(*to) + len <=
5094 	    (uintptr_t)&eq->desc[eq->sidx])) {
5095 		bcopy(from, *to, len);
5096 		(*to) += len;
5097 	} else {
5098 		int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
5099 
5100 		bcopy(from, *to, portion);
5101 		from += portion;
5102 		portion = len - portion;	/* remaining */
5103 		bcopy(from, (void *)eq->desc, portion);
5104 		(*to) = (caddr_t)eq->desc + portion;
5105 	}
5106 }
5107 
5108 static inline void
5109 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n)
5110 {
5111 	u_int db;
5112 
5113 	MPASS(n > 0);
5114 
5115 	db = eq->doorbells;
5116 	if (n > 1)
5117 		clrbit(&db, DOORBELL_WCWR);
5118 	wmb();
5119 
5120 	switch (ffs(db) - 1) {
5121 	case DOORBELL_UDB:
5122 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
5123 		break;
5124 
5125 	case DOORBELL_WCWR: {
5126 		volatile uint64_t *dst, *src;
5127 		int i;
5128 
5129 		/*
5130 		 * Queues whose 128B doorbell segment fits in the page do not
5131 		 * use relative qid (udb_qid is always 0).  Only queues with
5132 		 * doorbell segments can do WCWR.
5133 		 */
5134 		KASSERT(eq->udb_qid == 0 && n == 1,
5135 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
5136 		    __func__, eq->doorbells, n, eq->dbidx, eq));
5137 
5138 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
5139 		    UDBS_DB_OFFSET);
5140 		i = eq->dbidx;
5141 		src = (void *)&eq->desc[i];
5142 		while (src != (void *)&eq->desc[i + 1])
5143 			*dst++ = *src++;
5144 		wmb();
5145 		break;
5146 	}
5147 
5148 	case DOORBELL_UDBWC:
5149 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
5150 		wmb();
5151 		break;
5152 
5153 	case DOORBELL_KDB:
5154 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
5155 		    V_QID(eq->cntxt_id) | V_PIDX(n));
5156 		break;
5157 	}
5158 
5159 	IDXINCR(eq->dbidx, n, eq->sidx);
5160 }
5161 
5162 static inline u_int
5163 reclaimable_tx_desc(struct sge_eq *eq)
5164 {
5165 	uint16_t hw_cidx;
5166 
5167 	hw_cidx = read_hw_cidx(eq);
5168 	return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx));
5169 }
5170 
5171 static inline u_int
5172 total_available_tx_desc(struct sge_eq *eq)
5173 {
5174 	uint16_t hw_cidx, pidx;
5175 
5176 	hw_cidx = read_hw_cidx(eq);
5177 	pidx = eq->pidx;
5178 
5179 	if (pidx == hw_cidx)
5180 		return (eq->sidx - 1);
5181 	else
5182 		return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1);
5183 }
5184 
5185 static inline uint16_t
5186 read_hw_cidx(struct sge_eq *eq)
5187 {
5188 	struct sge_qstat *spg = (void *)&eq->desc[eq->sidx];
5189 	uint16_t cidx = spg->cidx;	/* stable snapshot */
5190 
5191 	return (be16toh(cidx));
5192 }
5193 
5194 /*
5195  * Reclaim 'n' descriptors approximately.
5196  */
5197 static u_int
5198 reclaim_tx_descs(struct sge_txq *txq, u_int n)
5199 {
5200 	struct tx_sdesc *txsd;
5201 	struct sge_eq *eq = &txq->eq;
5202 	u_int can_reclaim, reclaimed;
5203 
5204 	TXQ_LOCK_ASSERT_OWNED(txq);
5205 	MPASS(n > 0);
5206 
5207 	reclaimed = 0;
5208 	can_reclaim = reclaimable_tx_desc(eq);
5209 	while (can_reclaim && reclaimed < n) {
5210 		int ndesc;
5211 		struct mbuf *m, *nextpkt;
5212 
5213 		txsd = &txq->sdesc[eq->cidx];
5214 		ndesc = txsd->desc_used;
5215 
5216 		/* Firmware doesn't return "partial" credits. */
5217 		KASSERT(can_reclaim >= ndesc,
5218 		    ("%s: unexpected number of credits: %d, %d",
5219 		    __func__, can_reclaim, ndesc));
5220 		KASSERT(ndesc != 0,
5221 		    ("%s: descriptor with no credits: cidx %d",
5222 		    __func__, eq->cidx));
5223 
5224 		for (m = txsd->m; m != NULL; m = nextpkt) {
5225 			nextpkt = m->m_nextpkt;
5226 			m->m_nextpkt = NULL;
5227 			m_freem(m);
5228 		}
5229 		reclaimed += ndesc;
5230 		can_reclaim -= ndesc;
5231 		IDXINCR(eq->cidx, ndesc, eq->sidx);
5232 	}
5233 
5234 	return (reclaimed);
5235 }
5236 
5237 static void
5238 tx_reclaim(void *arg, int n)
5239 {
5240 	struct sge_txq *txq = arg;
5241 	struct sge_eq *eq = &txq->eq;
5242 
5243 	do {
5244 		if (TXQ_TRYLOCK(txq) == 0)
5245 			break;
5246 		n = reclaim_tx_descs(txq, 32);
5247 		if (eq->cidx == eq->pidx)
5248 			eq->equeqidx = eq->pidx;
5249 		TXQ_UNLOCK(txq);
5250 	} while (n > 0);
5251 }
5252 
5253 static __be64
5254 get_flit(struct sglist_seg *segs, int nsegs, int idx)
5255 {
5256 	int i = (idx / 3) * 2;
5257 
5258 	switch (idx % 3) {
5259 	case 0: {
5260 		uint64_t rc;
5261 
5262 		rc = (uint64_t)segs[i].ss_len << 32;
5263 		if (i + 1 < nsegs)
5264 			rc |= (uint64_t)(segs[i + 1].ss_len);
5265 
5266 		return (htobe64(rc));
5267 	}
5268 	case 1:
5269 		return (htobe64(segs[i].ss_paddr));
5270 	case 2:
5271 		return (htobe64(segs[i + 1].ss_paddr));
5272 	}
5273 
5274 	return (0);
5275 }
5276 
5277 static void
5278 find_best_refill_source(struct adapter *sc, struct sge_fl *fl, int maxp)
5279 {
5280 	int8_t zidx, hwidx, idx;
5281 	uint16_t region1, region3;
5282 	int spare, spare_needed, n;
5283 	struct sw_zone_info *swz;
5284 	struct hw_buf_info *hwb, *hwb_list = &sc->sge.hw_buf_info[0];
5285 
5286 	/*
5287 	 * Buffer Packing: Look for PAGE_SIZE or larger zone which has a bufsize
5288 	 * large enough for the max payload and cluster metadata.  Otherwise
5289 	 * settle for the largest bufsize that leaves enough room in the cluster
5290 	 * for metadata.
5291 	 *
5292 	 * Without buffer packing: Look for the smallest zone which has a
5293 	 * bufsize large enough for the max payload.  Settle for the largest
5294 	 * bufsize available if there's nothing big enough for max payload.
5295 	 */
5296 	spare_needed = fl->flags & FL_BUF_PACKING ? CL_METADATA_SIZE : 0;
5297 	swz = &sc->sge.sw_zone_info[0];
5298 	hwidx = -1;
5299 	for (zidx = 0; zidx < SW_ZONE_SIZES; zidx++, swz++) {
5300 		if (swz->size > largest_rx_cluster) {
5301 			if (__predict_true(hwidx != -1))
5302 				break;
5303 
5304 			/*
5305 			 * This is a misconfiguration.  largest_rx_cluster is
5306 			 * preventing us from finding a refill source.  See
5307 			 * dev.t5nex.<n>.buffer_sizes to figure out why.
5308 			 */
5309 			device_printf(sc->dev, "largest_rx_cluster=%u leaves no"
5310 			    " refill source for fl %p (dma %u).  Ignored.\n",
5311 			    largest_rx_cluster, fl, maxp);
5312 		}
5313 		for (idx = swz->head_hwidx; idx != -1; idx = hwb->next) {
5314 			hwb = &hwb_list[idx];
5315 			spare = swz->size - hwb->size;
5316 			if (spare < spare_needed)
5317 				continue;
5318 
5319 			hwidx = idx;		/* best option so far */
5320 			if (hwb->size >= maxp) {
5321 
5322 				if ((fl->flags & FL_BUF_PACKING) == 0)
5323 					goto done; /* stop looking (not packing) */
5324 
5325 				if (swz->size >= safest_rx_cluster)
5326 					goto done; /* stop looking (packing) */
5327 			}
5328 			break;		/* keep looking, next zone */
5329 		}
5330 	}
5331 done:
5332 	/* A usable hwidx has been located. */
5333 	MPASS(hwidx != -1);
5334 	hwb = &hwb_list[hwidx];
5335 	zidx = hwb->zidx;
5336 	swz = &sc->sge.sw_zone_info[zidx];
5337 	region1 = 0;
5338 	region3 = swz->size - hwb->size;
5339 
5340 	/*
5341 	 * Stay within this zone and see if there is a better match when mbuf
5342 	 * inlining is allowed.  Remember that the hwidx's are sorted in
5343 	 * decreasing order of size (so in increasing order of spare area).
5344 	 */
5345 	for (idx = hwidx; idx != -1; idx = hwb->next) {
5346 		hwb = &hwb_list[idx];
5347 		spare = swz->size - hwb->size;
5348 
5349 		if (allow_mbufs_in_cluster == 0 || hwb->size < maxp)
5350 			break;
5351 
5352 		/*
5353 		 * Do not inline mbufs if doing so would violate the pad/pack
5354 		 * boundary alignment requirement.
5355 		 */
5356 		if (fl_pad && (MSIZE % sc->params.sge.pad_boundary) != 0)
5357 			continue;
5358 		if (fl->flags & FL_BUF_PACKING &&
5359 		    (MSIZE % sc->params.sge.pack_boundary) != 0)
5360 			continue;
5361 
5362 		if (spare < CL_METADATA_SIZE + MSIZE)
5363 			continue;
5364 		n = (spare - CL_METADATA_SIZE) / MSIZE;
5365 		if (n > howmany(hwb->size, maxp))
5366 			break;
5367 
5368 		hwidx = idx;
5369 		if (fl->flags & FL_BUF_PACKING) {
5370 			region1 = n * MSIZE;
5371 			region3 = spare - region1;
5372 		} else {
5373 			region1 = MSIZE;
5374 			region3 = spare - region1;
5375 			break;
5376 		}
5377 	}
5378 
5379 	KASSERT(zidx >= 0 && zidx < SW_ZONE_SIZES,
5380 	    ("%s: bad zone %d for fl %p, maxp %d", __func__, zidx, fl, maxp));
5381 	KASSERT(hwidx >= 0 && hwidx <= SGE_FLBUF_SIZES,
5382 	    ("%s: bad hwidx %d for fl %p, maxp %d", __func__, hwidx, fl, maxp));
5383 	KASSERT(region1 + sc->sge.hw_buf_info[hwidx].size + region3 ==
5384 	    sc->sge.sw_zone_info[zidx].size,
5385 	    ("%s: bad buffer layout for fl %p, maxp %d. "
5386 		"cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5387 		sc->sge.sw_zone_info[zidx].size, region1,
5388 		sc->sge.hw_buf_info[hwidx].size, region3));
5389 	if (fl->flags & FL_BUF_PACKING || region1 > 0) {
5390 		KASSERT(region3 >= CL_METADATA_SIZE,
5391 		    ("%s: no room for metadata.  fl %p, maxp %d; "
5392 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5393 		    sc->sge.sw_zone_info[zidx].size, region1,
5394 		    sc->sge.hw_buf_info[hwidx].size, region3));
5395 		KASSERT(region1 % MSIZE == 0,
5396 		    ("%s: bad mbuf region for fl %p, maxp %d. "
5397 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5398 		    sc->sge.sw_zone_info[zidx].size, region1,
5399 		    sc->sge.hw_buf_info[hwidx].size, region3));
5400 	}
5401 
5402 	fl->cll_def.zidx = zidx;
5403 	fl->cll_def.hwidx = hwidx;
5404 	fl->cll_def.region1 = region1;
5405 	fl->cll_def.region3 = region3;
5406 }
5407 
5408 static void
5409 find_safe_refill_source(struct adapter *sc, struct sge_fl *fl)
5410 {
5411 	struct sge *s = &sc->sge;
5412 	struct hw_buf_info *hwb;
5413 	struct sw_zone_info *swz;
5414 	int spare;
5415 	int8_t hwidx;
5416 
5417 	if (fl->flags & FL_BUF_PACKING)
5418 		hwidx = s->safe_hwidx2;	/* with room for metadata */
5419 	else if (allow_mbufs_in_cluster && s->safe_hwidx2 != -1) {
5420 		hwidx = s->safe_hwidx2;
5421 		hwb = &s->hw_buf_info[hwidx];
5422 		swz = &s->sw_zone_info[hwb->zidx];
5423 		spare = swz->size - hwb->size;
5424 
5425 		/* no good if there isn't room for an mbuf as well */
5426 		if (spare < CL_METADATA_SIZE + MSIZE)
5427 			hwidx = s->safe_hwidx1;
5428 	} else
5429 		hwidx = s->safe_hwidx1;
5430 
5431 	if (hwidx == -1) {
5432 		/* No fallback source */
5433 		fl->cll_alt.hwidx = -1;
5434 		fl->cll_alt.zidx = -1;
5435 
5436 		return;
5437 	}
5438 
5439 	hwb = &s->hw_buf_info[hwidx];
5440 	swz = &s->sw_zone_info[hwb->zidx];
5441 	spare = swz->size - hwb->size;
5442 	fl->cll_alt.hwidx = hwidx;
5443 	fl->cll_alt.zidx = hwb->zidx;
5444 	if (allow_mbufs_in_cluster &&
5445 	    (fl_pad == 0 || (MSIZE % sc->params.sge.pad_boundary) == 0))
5446 		fl->cll_alt.region1 = ((spare - CL_METADATA_SIZE) / MSIZE) * MSIZE;
5447 	else
5448 		fl->cll_alt.region1 = 0;
5449 	fl->cll_alt.region3 = spare - fl->cll_alt.region1;
5450 }
5451 
5452 static void
5453 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
5454 {
5455 	mtx_lock(&sc->sfl_lock);
5456 	FL_LOCK(fl);
5457 	if ((fl->flags & FL_DOOMED) == 0) {
5458 		fl->flags |= FL_STARVING;
5459 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
5460 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
5461 	}
5462 	FL_UNLOCK(fl);
5463 	mtx_unlock(&sc->sfl_lock);
5464 }
5465 
5466 static void
5467 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq)
5468 {
5469 	struct sge_wrq *wrq = (void *)eq;
5470 
5471 	atomic_readandclear_int(&eq->equiq);
5472 	taskqueue_enqueue(sc->tq[eq->tx_chan], &wrq->wrq_tx_task);
5473 }
5474 
5475 static void
5476 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq)
5477 {
5478 	struct sge_txq *txq = (void *)eq;
5479 
5480 	MPASS((eq->flags & EQ_TYPEMASK) == EQ_ETH);
5481 
5482 	atomic_readandclear_int(&eq->equiq);
5483 	mp_ring_check_drainage(txq->r, 0);
5484 	taskqueue_enqueue(sc->tq[eq->tx_chan], &txq->tx_reclaim_task);
5485 }
5486 
5487 static int
5488 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
5489     struct mbuf *m)
5490 {
5491 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
5492 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
5493 	struct adapter *sc = iq->adapter;
5494 	struct sge *s = &sc->sge;
5495 	struct sge_eq *eq;
5496 	static void (*h[])(struct adapter *, struct sge_eq *) = {NULL,
5497 		&handle_wrq_egr_update, &handle_eth_egr_update,
5498 		&handle_wrq_egr_update};
5499 
5500 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5501 	    rss->opcode));
5502 
5503 	eq = s->eqmap[qid - s->eq_start - s->eq_base];
5504 	(*h[eq->flags & EQ_TYPEMASK])(sc, eq);
5505 
5506 	return (0);
5507 }
5508 
5509 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
5510 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
5511     offsetof(struct cpl_fw6_msg, data));
5512 
5513 static int
5514 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
5515 {
5516 	struct adapter *sc = iq->adapter;
5517 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
5518 
5519 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5520 	    rss->opcode));
5521 
5522 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
5523 		const struct rss_header *rss2;
5524 
5525 		rss2 = (const struct rss_header *)&cpl->data[0];
5526 		return (t4_cpl_handler[rss2->opcode](iq, rss2, m));
5527 	}
5528 
5529 	return (t4_fw_msg_handler[cpl->type](sc, &cpl->data[0]));
5530 }
5531 
5532 /**
5533  *	t4_handle_wrerr_rpl - process a FW work request error message
5534  *	@adap: the adapter
5535  *	@rpl: start of the FW message
5536  */
5537 static int
5538 t4_handle_wrerr_rpl(struct adapter *adap, const __be64 *rpl)
5539 {
5540 	u8 opcode = *(const u8 *)rpl;
5541 	const struct fw_error_cmd *e = (const void *)rpl;
5542 	unsigned int i;
5543 
5544 	if (opcode != FW_ERROR_CMD) {
5545 		log(LOG_ERR,
5546 		    "%s: Received WRERR_RPL message with opcode %#x\n",
5547 		    device_get_nameunit(adap->dev), opcode);
5548 		return (EINVAL);
5549 	}
5550 	log(LOG_ERR, "%s: FW_ERROR (%s) ", device_get_nameunit(adap->dev),
5551 	    G_FW_ERROR_CMD_FATAL(be32toh(e->op_to_type)) ? "fatal" :
5552 	    "non-fatal");
5553 	switch (G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type))) {
5554 	case FW_ERROR_TYPE_EXCEPTION:
5555 		log(LOG_ERR, "exception info:\n");
5556 		for (i = 0; i < nitems(e->u.exception.info); i++)
5557 			log(LOG_ERR, "%s%08x", i == 0 ? "\t" : " ",
5558 			    be32toh(e->u.exception.info[i]));
5559 		log(LOG_ERR, "\n");
5560 		break;
5561 	case FW_ERROR_TYPE_HWMODULE:
5562 		log(LOG_ERR, "HW module regaddr %08x regval %08x\n",
5563 		    be32toh(e->u.hwmodule.regaddr),
5564 		    be32toh(e->u.hwmodule.regval));
5565 		break;
5566 	case FW_ERROR_TYPE_WR:
5567 		log(LOG_ERR, "WR cidx %d PF %d VF %d eqid %d hdr:\n",
5568 		    be16toh(e->u.wr.cidx),
5569 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.wr.pfn_vfn)),
5570 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.wr.pfn_vfn)),
5571 		    be32toh(e->u.wr.eqid));
5572 		for (i = 0; i < nitems(e->u.wr.wrhdr); i++)
5573 			log(LOG_ERR, "%s%02x", i == 0 ? "\t" : " ",
5574 			    e->u.wr.wrhdr[i]);
5575 		log(LOG_ERR, "\n");
5576 		break;
5577 	case FW_ERROR_TYPE_ACL:
5578 		log(LOG_ERR, "ACL cidx %d PF %d VF %d eqid %d %s",
5579 		    be16toh(e->u.acl.cidx),
5580 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.acl.pfn_vfn)),
5581 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.acl.pfn_vfn)),
5582 		    be32toh(e->u.acl.eqid),
5583 		    G_FW_ERROR_CMD_MV(be16toh(e->u.acl.mv_pkd)) ? "vlanid" :
5584 		    "MAC");
5585 		for (i = 0; i < nitems(e->u.acl.val); i++)
5586 			log(LOG_ERR, " %02x", e->u.acl.val[i]);
5587 		log(LOG_ERR, "\n");
5588 		break;
5589 	default:
5590 		log(LOG_ERR, "type %#x\n",
5591 		    G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type)));
5592 		return (EINVAL);
5593 	}
5594 	return (0);
5595 }
5596 
5597 static int
5598 sysctl_uint16(SYSCTL_HANDLER_ARGS)
5599 {
5600 	uint16_t *id = arg1;
5601 	int i = *id;
5602 
5603 	return sysctl_handle_int(oidp, &i, 0, req);
5604 }
5605 
5606 static int
5607 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
5608 {
5609 	struct sge *s = arg1;
5610 	struct hw_buf_info *hwb = &s->hw_buf_info[0];
5611 	struct sw_zone_info *swz = &s->sw_zone_info[0];
5612 	int i, rc;
5613 	struct sbuf sb;
5614 	char c;
5615 
5616 	sbuf_new(&sb, NULL, 32, SBUF_AUTOEXTEND);
5617 	for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
5618 		if (hwb->zidx >= 0 && swz[hwb->zidx].size <= largest_rx_cluster)
5619 			c = '*';
5620 		else
5621 			c = '\0';
5622 
5623 		sbuf_printf(&sb, "%u%c ", hwb->size, c);
5624 	}
5625 	sbuf_trim(&sb);
5626 	sbuf_finish(&sb);
5627 	rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
5628 	sbuf_delete(&sb);
5629 	return (rc);
5630 }
5631 
5632 #ifdef RATELIMIT
5633 /*
5634  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
5635  */
5636 static inline u_int
5637 txpkt_eo_len16(u_int nsegs, u_int immhdrs, u_int tso)
5638 {
5639 	u_int n;
5640 
5641 	MPASS(immhdrs > 0);
5642 
5643 	n = roundup2(sizeof(struct fw_eth_tx_eo_wr) +
5644 	    sizeof(struct cpl_tx_pkt_core) + immhdrs, 16);
5645 	if (__predict_false(nsegs == 0))
5646 		goto done;
5647 
5648 	nsegs--; /* first segment is part of ulptx_sgl */
5649 	n += sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
5650 	if (tso)
5651 		n += sizeof(struct cpl_tx_pkt_lso_core);
5652 
5653 done:
5654 	return (howmany(n, 16));
5655 }
5656 
5657 #define ETID_FLOWC_NPARAMS 6
5658 #define ETID_FLOWC_LEN (roundup2((sizeof(struct fw_flowc_wr) + \
5659     ETID_FLOWC_NPARAMS * sizeof(struct fw_flowc_mnemval)), 16))
5660 #define ETID_FLOWC_LEN16 (howmany(ETID_FLOWC_LEN, 16))
5661 
5662 static int
5663 send_etid_flowc_wr(struct cxgbe_snd_tag *cst, struct port_info *pi,
5664     struct vi_info *vi)
5665 {
5666 	struct wrq_cookie cookie;
5667 	u_int pfvf = pi->adapter->pf << S_FW_VIID_PFN;
5668 	struct fw_flowc_wr *flowc;
5669 
5670 	mtx_assert(&cst->lock, MA_OWNED);
5671 	MPASS((cst->flags & (EO_FLOWC_PENDING | EO_FLOWC_RPL_PENDING)) ==
5672 	    EO_FLOWC_PENDING);
5673 
5674 	flowc = start_wrq_wr(cst->eo_txq, ETID_FLOWC_LEN16, &cookie);
5675 	if (__predict_false(flowc == NULL))
5676 		return (ENOMEM);
5677 
5678 	bzero(flowc, ETID_FLOWC_LEN);
5679 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
5680 	    V_FW_FLOWC_WR_NPARAMS(ETID_FLOWC_NPARAMS) | V_FW_WR_COMPL(0));
5681 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(ETID_FLOWC_LEN16) |
5682 	    V_FW_WR_FLOWID(cst->etid));
5683 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN;
5684 	flowc->mnemval[0].val = htobe32(pfvf);
5685 	flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH;
5686 	flowc->mnemval[1].val = htobe32(pi->tx_chan);
5687 	flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT;
5688 	flowc->mnemval[2].val = htobe32(pi->tx_chan);
5689 	flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID;
5690 	flowc->mnemval[3].val = htobe32(cst->iqid);
5691 	flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_EOSTATE;
5692 	flowc->mnemval[4].val = htobe32(FW_FLOWC_MNEM_EOSTATE_ESTABLISHED);
5693 	flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS;
5694 	flowc->mnemval[5].val = htobe32(cst->schedcl);
5695 
5696 	commit_wrq_wr(cst->eo_txq, flowc, &cookie);
5697 
5698 	cst->flags &= ~EO_FLOWC_PENDING;
5699 	cst->flags |= EO_FLOWC_RPL_PENDING;
5700 	MPASS(cst->tx_credits >= ETID_FLOWC_LEN16);	/* flowc is first WR. */
5701 	cst->tx_credits -= ETID_FLOWC_LEN16;
5702 
5703 	return (0);
5704 }
5705 
5706 #define ETID_FLUSH_LEN16 (howmany(sizeof (struct fw_flowc_wr), 16))
5707 
5708 void
5709 send_etid_flush_wr(struct cxgbe_snd_tag *cst)
5710 {
5711 	struct fw_flowc_wr *flowc;
5712 	struct wrq_cookie cookie;
5713 
5714 	mtx_assert(&cst->lock, MA_OWNED);
5715 
5716 	flowc = start_wrq_wr(cst->eo_txq, ETID_FLUSH_LEN16, &cookie);
5717 	if (__predict_false(flowc == NULL))
5718 		CXGBE_UNIMPLEMENTED(__func__);
5719 
5720 	bzero(flowc, ETID_FLUSH_LEN16 * 16);
5721 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
5722 	    V_FW_FLOWC_WR_NPARAMS(0) | F_FW_WR_COMPL);
5723 	flowc->flowid_len16 = htobe32(V_FW_WR_LEN16(ETID_FLUSH_LEN16) |
5724 	    V_FW_WR_FLOWID(cst->etid));
5725 
5726 	commit_wrq_wr(cst->eo_txq, flowc, &cookie);
5727 
5728 	cst->flags |= EO_FLUSH_RPL_PENDING;
5729 	MPASS(cst->tx_credits >= ETID_FLUSH_LEN16);
5730 	cst->tx_credits -= ETID_FLUSH_LEN16;
5731 	cst->ncompl++;
5732 }
5733 
5734 static void
5735 write_ethofld_wr(struct cxgbe_snd_tag *cst, struct fw_eth_tx_eo_wr *wr,
5736     struct mbuf *m0, int compl)
5737 {
5738 	struct cpl_tx_pkt_core *cpl;
5739 	uint64_t ctrl1;
5740 	uint32_t ctrl;	/* used in many unrelated places */
5741 	int len16, pktlen, nsegs, immhdrs;
5742 	caddr_t dst;
5743 	uintptr_t p;
5744 	struct ulptx_sgl *usgl;
5745 	struct sglist sg;
5746 	struct sglist_seg segs[38];	/* XXX: find real limit.  XXX: get off the stack */
5747 
5748 	mtx_assert(&cst->lock, MA_OWNED);
5749 	M_ASSERTPKTHDR(m0);
5750 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
5751 	    m0->m_pkthdr.l4hlen > 0,
5752 	    ("%s: ethofld mbuf %p is missing header lengths", __func__, m0));
5753 
5754 	len16 = mbuf_eo_len16(m0);
5755 	nsegs = mbuf_eo_nsegs(m0);
5756 	pktlen = m0->m_pkthdr.len;
5757 	ctrl = sizeof(struct cpl_tx_pkt_core);
5758 	if (needs_tso(m0))
5759 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
5760 	immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen + m0->m_pkthdr.l4hlen;
5761 	ctrl += immhdrs;
5762 
5763 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_EO_WR) |
5764 	    V_FW_ETH_TX_EO_WR_IMMDLEN(ctrl) | V_FW_WR_COMPL(!!compl));
5765 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(len16) |
5766 	    V_FW_WR_FLOWID(cst->etid));
5767 	wr->r3 = 0;
5768 	if (needs_udp_csum(m0)) {
5769 		wr->u.udpseg.type = FW_ETH_TX_EO_TYPE_UDPSEG;
5770 		wr->u.udpseg.ethlen = m0->m_pkthdr.l2hlen;
5771 		wr->u.udpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
5772 		wr->u.udpseg.udplen = m0->m_pkthdr.l4hlen;
5773 		wr->u.udpseg.rtplen = 0;
5774 		wr->u.udpseg.r4 = 0;
5775 		wr->u.udpseg.mss = htobe16(pktlen - immhdrs);
5776 		wr->u.udpseg.schedpktsize = wr->u.udpseg.mss;
5777 		wr->u.udpseg.plen = htobe32(pktlen - immhdrs);
5778 		cpl = (void *)(wr + 1);
5779 	} else {
5780 		MPASS(needs_tcp_csum(m0));
5781 		wr->u.tcpseg.type = FW_ETH_TX_EO_TYPE_TCPSEG;
5782 		wr->u.tcpseg.ethlen = m0->m_pkthdr.l2hlen;
5783 		wr->u.tcpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
5784 		wr->u.tcpseg.tcplen = m0->m_pkthdr.l4hlen;
5785 		wr->u.tcpseg.tsclk_tsoff = mbuf_eo_tsclk_tsoff(m0);
5786 		wr->u.tcpseg.r4 = 0;
5787 		wr->u.tcpseg.r5 = 0;
5788 		wr->u.tcpseg.plen = htobe32(pktlen - immhdrs);
5789 
5790 		if (needs_tso(m0)) {
5791 			struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
5792 
5793 			wr->u.tcpseg.mss = htobe16(m0->m_pkthdr.tso_segsz);
5794 
5795 			ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) |
5796 			    F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE |
5797 			    V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) |
5798 			    V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
5799 			if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
5800 				ctrl |= V_LSO_ETHHDR_LEN(1);
5801 			if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
5802 				ctrl |= F_LSO_IPV6;
5803 			lso->lso_ctrl = htobe32(ctrl);
5804 			lso->ipid_ofst = htobe16(0);
5805 			lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
5806 			lso->seqno_offset = htobe32(0);
5807 			lso->len = htobe32(pktlen);
5808 
5809 			cpl = (void *)(lso + 1);
5810 		} else {
5811 			wr->u.tcpseg.mss = htobe16(0xffff);
5812 			cpl = (void *)(wr + 1);
5813 		}
5814 	}
5815 
5816 	/* Checksum offload must be requested for ethofld. */
5817 	ctrl1 = 0;
5818 	MPASS(needs_l4_csum(m0));
5819 
5820 	/* VLAN tag insertion */
5821 	if (needs_vlan_insertion(m0)) {
5822 		ctrl1 |= F_TXPKT_VLAN_VLD |
5823 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
5824 	}
5825 
5826 	/* CPL header */
5827 	cpl->ctrl0 = cst->ctrl0;
5828 	cpl->pack = 0;
5829 	cpl->len = htobe16(pktlen);
5830 	cpl->ctrl1 = htobe64(ctrl1);
5831 
5832 	/* Copy Ethernet, IP & TCP/UDP hdrs as immediate data */
5833 	p = (uintptr_t)(cpl + 1);
5834 	m_copydata(m0, 0, immhdrs, (void *)p);
5835 
5836 	/* SGL */
5837 	dst = (void *)(cpl + 1);
5838 	if (nsegs > 0) {
5839 		int i, pad;
5840 
5841 		/* zero-pad upto next 16Byte boundary, if not 16Byte aligned */
5842 		p += immhdrs;
5843 		pad = 16 - (immhdrs & 0xf);
5844 		bzero((void *)p, pad);
5845 
5846 		usgl = (void *)(p + pad);
5847 		usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
5848 		    V_ULPTX_NSGE(nsegs));
5849 
5850 		sglist_init(&sg, nitems(segs), segs);
5851 		for (; m0 != NULL; m0 = m0->m_next) {
5852 			if (__predict_false(m0->m_len == 0))
5853 				continue;
5854 			if (immhdrs >= m0->m_len) {
5855 				immhdrs -= m0->m_len;
5856 				continue;
5857 			}
5858 
5859 			sglist_append(&sg, mtod(m0, char *) + immhdrs,
5860 			    m0->m_len - immhdrs);
5861 			immhdrs = 0;
5862 		}
5863 		MPASS(sg.sg_nseg == nsegs);
5864 
5865 		/*
5866 		 * Zero pad last 8B in case the WR doesn't end on a 16B
5867 		 * boundary.
5868 		 */
5869 		*(uint64_t *)((char *)wr + len16 * 16 - 8) = 0;
5870 
5871 		usgl->len0 = htobe32(segs[0].ss_len);
5872 		usgl->addr0 = htobe64(segs[0].ss_paddr);
5873 		for (i = 0; i < nsegs - 1; i++) {
5874 			usgl->sge[i / 2].len[i & 1] = htobe32(segs[i + 1].ss_len);
5875 			usgl->sge[i / 2].addr[i & 1] = htobe64(segs[i + 1].ss_paddr);
5876 		}
5877 		if (i & 1)
5878 			usgl->sge[i / 2].len[1] = htobe32(0);
5879 	}
5880 
5881 }
5882 
5883 static void
5884 ethofld_tx(struct cxgbe_snd_tag *cst)
5885 {
5886 	struct mbuf *m;
5887 	struct wrq_cookie cookie;
5888 	int next_credits, compl;
5889 	struct fw_eth_tx_eo_wr *wr;
5890 
5891 	mtx_assert(&cst->lock, MA_OWNED);
5892 
5893 	while ((m = mbufq_first(&cst->pending_tx)) != NULL) {
5894 		M_ASSERTPKTHDR(m);
5895 
5896 		/* How many len16 credits do we need to send this mbuf. */
5897 		next_credits = mbuf_eo_len16(m);
5898 		MPASS(next_credits > 0);
5899 		if (next_credits > cst->tx_credits) {
5900 			/*
5901 			 * Tx will make progress eventually because there is at
5902 			 * least one outstanding fw4_ack that will return
5903 			 * credits and kick the tx.
5904 			 */
5905 			MPASS(cst->ncompl > 0);
5906 			return;
5907 		}
5908 		wr = start_wrq_wr(cst->eo_txq, next_credits, &cookie);
5909 		if (__predict_false(wr == NULL)) {
5910 			/* XXX: wishful thinking, not a real assertion. */
5911 			MPASS(cst->ncompl > 0);
5912 			return;
5913 		}
5914 		cst->tx_credits -= next_credits;
5915 		cst->tx_nocompl += next_credits;
5916 		compl = cst->ncompl == 0 || cst->tx_nocompl >= cst->tx_total / 2;
5917 		ETHER_BPF_MTAP(cst->com.ifp, m);
5918 		write_ethofld_wr(cst, wr, m, compl);
5919 		commit_wrq_wr(cst->eo_txq, wr, &cookie);
5920 		if (compl) {
5921 			cst->ncompl++;
5922 			cst->tx_nocompl	= 0;
5923 		}
5924 		(void) mbufq_dequeue(&cst->pending_tx);
5925 		mbufq_enqueue(&cst->pending_fwack, m);
5926 	}
5927 }
5928 
5929 int
5930 ethofld_transmit(struct ifnet *ifp, struct mbuf *m0)
5931 {
5932 	struct cxgbe_snd_tag *cst;
5933 	int rc;
5934 
5935 	MPASS(m0->m_nextpkt == NULL);
5936 	MPASS(m0->m_pkthdr.snd_tag != NULL);
5937 	cst = mst_to_cst(m0->m_pkthdr.snd_tag);
5938 
5939 	mtx_lock(&cst->lock);
5940 	MPASS(cst->flags & EO_SND_TAG_REF);
5941 
5942 	if (__predict_false(cst->flags & EO_FLOWC_PENDING)) {
5943 		struct vi_info *vi = ifp->if_softc;
5944 		struct port_info *pi = vi->pi;
5945 		struct adapter *sc = pi->adapter;
5946 		const uint32_t rss_mask = vi->rss_size - 1;
5947 		uint32_t rss_hash;
5948 
5949 		cst->eo_txq = &sc->sge.ofld_txq[vi->first_ofld_txq];
5950 		if (M_HASHTYPE_ISHASH(m0))
5951 			rss_hash = m0->m_pkthdr.flowid;
5952 		else
5953 			rss_hash = arc4random();
5954 		/* We assume RSS hashing */
5955 		cst->iqid = vi->rss[rss_hash & rss_mask];
5956 		cst->eo_txq += rss_hash % vi->nofldtxq;
5957 		rc = send_etid_flowc_wr(cst, pi, vi);
5958 		if (rc != 0)
5959 			goto done;
5960 	}
5961 
5962 	if (__predict_false(cst->plen + m0->m_pkthdr.len > eo_max_backlog)) {
5963 		rc = ENOBUFS;
5964 		goto done;
5965 	}
5966 
5967 	mbufq_enqueue(&cst->pending_tx, m0);
5968 	cst->plen += m0->m_pkthdr.len;
5969 
5970 	ethofld_tx(cst);
5971 	rc = 0;
5972 done:
5973 	mtx_unlock(&cst->lock);
5974 	if (__predict_false(rc != 0))
5975 		m_freem(m0);
5976 	return (rc);
5977 }
5978 
5979 static int
5980 ethofld_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
5981 {
5982 	struct adapter *sc = iq->adapter;
5983 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
5984 	struct mbuf *m;
5985 	u_int etid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
5986 	struct cxgbe_snd_tag *cst;
5987 	uint8_t credits = cpl->credits;
5988 
5989 	cst = lookup_etid(sc, etid);
5990 	mtx_lock(&cst->lock);
5991 	if (__predict_false(cst->flags & EO_FLOWC_RPL_PENDING)) {
5992 		MPASS(credits >= ETID_FLOWC_LEN16);
5993 		credits -= ETID_FLOWC_LEN16;
5994 		cst->flags &= ~EO_FLOWC_RPL_PENDING;
5995 	}
5996 
5997 	KASSERT(cst->ncompl > 0,
5998 	    ("%s: etid %u (%p) wasn't expecting completion.",
5999 	    __func__, etid, cst));
6000 	cst->ncompl--;
6001 
6002 	while (credits > 0) {
6003 		m = mbufq_dequeue(&cst->pending_fwack);
6004 		if (__predict_false(m == NULL)) {
6005 			/*
6006 			 * The remaining credits are for the final flush that
6007 			 * was issued when the tag was freed by the kernel.
6008 			 */
6009 			MPASS((cst->flags &
6010 			    (EO_FLUSH_RPL_PENDING | EO_SND_TAG_REF)) ==
6011 			    EO_FLUSH_RPL_PENDING);
6012 			MPASS(credits == ETID_FLUSH_LEN16);
6013 			MPASS(cst->tx_credits + cpl->credits == cst->tx_total);
6014 			MPASS(cst->ncompl == 0);
6015 
6016 			cst->flags &= ~EO_FLUSH_RPL_PENDING;
6017 			cst->tx_credits += cpl->credits;
6018 freetag:
6019 			cxgbe_snd_tag_free_locked(cst);
6020 			return (0);	/* cst is gone. */
6021 		}
6022 		KASSERT(m != NULL,
6023 		    ("%s: too many credits (%u, %u)", __func__, cpl->credits,
6024 		    credits));
6025 		KASSERT(credits >= mbuf_eo_len16(m),
6026 		    ("%s: too few credits (%u, %u, %u)", __func__,
6027 		    cpl->credits, credits, mbuf_eo_len16(m)));
6028 		credits -= mbuf_eo_len16(m);
6029 		cst->plen -= m->m_pkthdr.len;
6030 		m_freem(m);
6031 	}
6032 
6033 	cst->tx_credits += cpl->credits;
6034 	MPASS(cst->tx_credits <= cst->tx_total);
6035 
6036 	m = mbufq_first(&cst->pending_tx);
6037 	if (m != NULL && cst->tx_credits >= mbuf_eo_len16(m))
6038 		ethofld_tx(cst);
6039 
6040 	if (__predict_false((cst->flags & EO_SND_TAG_REF) == 0) &&
6041 	    cst->ncompl == 0) {
6042 		if (cst->tx_credits == cst->tx_total)
6043 			goto freetag;
6044 		else {
6045 			MPASS((cst->flags & EO_FLUSH_RPL_PENDING) == 0);
6046 			send_etid_flush_wr(cst);
6047 		}
6048 	}
6049 
6050 	mtx_unlock(&cst->lock);
6051 
6052 	return (0);
6053 }
6054 #endif
6055