xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision 28f4385e45a2681c14bd04b83fe1796eaefe8265)
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 
1398 	t4_slow_intr_handler(sc);
1399 }
1400 
1401 /*
1402  * Interrupt handler for iq-only queues.  The firmware event queue is the only
1403  * such queue right now.
1404  */
1405 void
1406 t4_intr_evt(void *arg)
1407 {
1408 	struct sge_iq *iq = arg;
1409 
1410 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1411 		service_iq(iq, 0);
1412 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1413 	}
1414 }
1415 
1416 /*
1417  * Interrupt handler for iq+fl queues.
1418  */
1419 void
1420 t4_intr(void *arg)
1421 {
1422 	struct sge_iq *iq = arg;
1423 
1424 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1425 		service_iq_fl(iq, 0);
1426 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1427 	}
1428 }
1429 
1430 #ifdef DEV_NETMAP
1431 /*
1432  * Interrupt handler for netmap rx queues.
1433  */
1434 void
1435 t4_nm_intr(void *arg)
1436 {
1437 	struct sge_nm_rxq *nm_rxq = arg;
1438 
1439 	if (atomic_cmpset_int(&nm_rxq->nm_state, NM_ON, NM_BUSY)) {
1440 		service_nm_rxq(nm_rxq);
1441 		(void) atomic_cmpset_int(&nm_rxq->nm_state, NM_BUSY, NM_ON);
1442 	}
1443 }
1444 
1445 /*
1446  * Interrupt handler for vectors shared between NIC and netmap rx queues.
1447  */
1448 void
1449 t4_vi_intr(void *arg)
1450 {
1451 	struct irq *irq = arg;
1452 
1453 	MPASS(irq->nm_rxq != NULL);
1454 	t4_nm_intr(irq->nm_rxq);
1455 
1456 	MPASS(irq->rxq != NULL);
1457 	t4_intr(irq->rxq);
1458 }
1459 #endif
1460 
1461 /*
1462  * Deals with interrupts on an iq-only (no freelist) queue.
1463  */
1464 static int
1465 service_iq(struct sge_iq *iq, int budget)
1466 {
1467 	struct sge_iq *q;
1468 	struct adapter *sc = iq->adapter;
1469 	struct iq_desc *d = &iq->desc[iq->cidx];
1470 	int ndescs = 0, limit;
1471 	int rsp_type;
1472 	uint32_t lq;
1473 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1474 
1475 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1476 	KASSERT((iq->flags & IQ_HAS_FL) == 0,
1477 	    ("%s: called for iq %p with fl (iq->flags 0x%x)", __func__, iq,
1478 	    iq->flags));
1479 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1480 	MPASS((iq->flags & IQ_LRO_ENABLED) == 0);
1481 
1482 	limit = budget ? budget : iq->qsize / 16;
1483 
1484 	/*
1485 	 * We always come back and check the descriptor ring for new indirect
1486 	 * interrupts and other responses after running a single handler.
1487 	 */
1488 	for (;;) {
1489 		while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1490 
1491 			rmb();
1492 
1493 			rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1494 			lq = be32toh(d->rsp.pldbuflen_qid);
1495 
1496 			switch (rsp_type) {
1497 			case X_RSPD_TYPE_FLBUF:
1498 				panic("%s: data for an iq (%p) with no freelist",
1499 				    __func__, iq);
1500 
1501 				/* NOTREACHED */
1502 
1503 			case X_RSPD_TYPE_CPL:
1504 				KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1505 				    ("%s: bad opcode %02x.", __func__,
1506 				    d->rss.opcode));
1507 				t4_cpl_handler[d->rss.opcode](iq, &d->rss, NULL);
1508 				break;
1509 
1510 			case X_RSPD_TYPE_INTR:
1511 				/*
1512 				 * There are 1K interrupt-capable queues (qids 0
1513 				 * through 1023).  A response type indicating a
1514 				 * forwarded interrupt with a qid >= 1K is an
1515 				 * iWARP async notification.
1516 				 */
1517 				if (__predict_true(lq >= 1024)) {
1518 					t4_an_handler(iq, &d->rsp);
1519 					break;
1520 				}
1521 
1522 				q = sc->sge.iqmap[lq - sc->sge.iq_start -
1523 				    sc->sge.iq_base];
1524 				if (atomic_cmpset_int(&q->state, IQS_IDLE,
1525 				    IQS_BUSY)) {
1526 					if (service_iq_fl(q, q->qsize / 16) == 0) {
1527 						(void) atomic_cmpset_int(&q->state,
1528 						    IQS_BUSY, IQS_IDLE);
1529 					} else {
1530 						STAILQ_INSERT_TAIL(&iql, q,
1531 						    link);
1532 					}
1533 				}
1534 				break;
1535 
1536 			default:
1537 				KASSERT(0,
1538 				    ("%s: illegal response type %d on iq %p",
1539 				    __func__, rsp_type, iq));
1540 				log(LOG_ERR,
1541 				    "%s: illegal response type %d on iq %p",
1542 				    device_get_nameunit(sc->dev), rsp_type, iq);
1543 				break;
1544 			}
1545 
1546 			d++;
1547 			if (__predict_false(++iq->cidx == iq->sidx)) {
1548 				iq->cidx = 0;
1549 				iq->gen ^= F_RSPD_GEN;
1550 				d = &iq->desc[0];
1551 			}
1552 			if (__predict_false(++ndescs == limit)) {
1553 				t4_write_reg(sc, sc->sge_gts_reg,
1554 				    V_CIDXINC(ndescs) |
1555 				    V_INGRESSQID(iq->cntxt_id) |
1556 				    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1557 				ndescs = 0;
1558 
1559 				if (budget) {
1560 					return (EINPROGRESS);
1561 				}
1562 			}
1563 		}
1564 
1565 		if (STAILQ_EMPTY(&iql))
1566 			break;
1567 
1568 		/*
1569 		 * Process the head only, and send it to the back of the list if
1570 		 * it's still not done.
1571 		 */
1572 		q = STAILQ_FIRST(&iql);
1573 		STAILQ_REMOVE_HEAD(&iql, link);
1574 		if (service_iq_fl(q, q->qsize / 8) == 0)
1575 			(void) atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1576 		else
1577 			STAILQ_INSERT_TAIL(&iql, q, link);
1578 	}
1579 
1580 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1581 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1582 
1583 	return (0);
1584 }
1585 
1586 static inline int
1587 sort_before_lro(struct lro_ctrl *lro)
1588 {
1589 
1590 	return (lro->lro_mbuf_max != 0);
1591 }
1592 
1593 static inline uint64_t
1594 last_flit_to_ns(struct adapter *sc, uint64_t lf)
1595 {
1596 	uint64_t n = be64toh(lf) & 0xfffffffffffffff;	/* 60b, not 64b. */
1597 
1598 	if (n > UINT64_MAX / 1000000)
1599 		return (n / sc->params.vpd.cclk * 1000000);
1600 	else
1601 		return (n * 1000000 / sc->params.vpd.cclk);
1602 }
1603 
1604 /*
1605  * Deals with interrupts on an iq+fl queue.
1606  */
1607 static int
1608 service_iq_fl(struct sge_iq *iq, int budget)
1609 {
1610 	struct sge_rxq *rxq = iq_to_rxq(iq);
1611 	struct sge_fl *fl;
1612 	struct adapter *sc = iq->adapter;
1613 	struct iq_desc *d = &iq->desc[iq->cidx];
1614 	int ndescs = 0, limit;
1615 	int rsp_type, refill, starved;
1616 	uint32_t lq;
1617 	uint16_t fl_hw_cidx;
1618 	struct mbuf *m0;
1619 #if defined(INET) || defined(INET6)
1620 	const struct timeval lro_timeout = {0, sc->lro_timeout};
1621 	struct lro_ctrl *lro = &rxq->lro;
1622 #endif
1623 
1624 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1625 	MPASS(iq->flags & IQ_HAS_FL);
1626 
1627 	limit = budget ? budget : iq->qsize / 16;
1628 	fl = &rxq->fl;
1629 	fl_hw_cidx = fl->hw_cidx;	/* stable snapshot */
1630 
1631 #if defined(INET) || defined(INET6)
1632 	if (iq->flags & IQ_ADJ_CREDIT) {
1633 		MPASS(sort_before_lro(lro));
1634 		iq->flags &= ~IQ_ADJ_CREDIT;
1635 		if ((d->rsp.u.type_gen & F_RSPD_GEN) != iq->gen) {
1636 			tcp_lro_flush_all(lro);
1637 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(1) |
1638 			    V_INGRESSQID((u32)iq->cntxt_id) |
1639 			    V_SEINTARM(iq->intr_params));
1640 			return (0);
1641 		}
1642 		ndescs = 1;
1643 	}
1644 #else
1645 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1646 #endif
1647 
1648 	while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1649 
1650 		rmb();
1651 
1652 		refill = 0;
1653 		m0 = NULL;
1654 		rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1655 		lq = be32toh(d->rsp.pldbuflen_qid);
1656 
1657 		switch (rsp_type) {
1658 		case X_RSPD_TYPE_FLBUF:
1659 
1660 			m0 = get_fl_payload(sc, fl, lq);
1661 			if (__predict_false(m0 == NULL))
1662 				goto out;
1663 			refill = IDXDIFF(fl->hw_cidx, fl_hw_cidx, fl->sidx) > 2;
1664 
1665 			if (iq->flags & IQ_RX_TIMESTAMP) {
1666 				/*
1667 				 * Fill up rcv_tstmp but do not set M_TSTMP.
1668 				 * rcv_tstmp is not in the format that the
1669 				 * kernel expects and we don't want to mislead
1670 				 * it.  For now this is only for custom code
1671 				 * that knows how to interpret cxgbe's stamp.
1672 				 */
1673 				m0->m_pkthdr.rcv_tstmp =
1674 				    last_flit_to_ns(sc, d->rsp.u.last_flit);
1675 #ifdef notyet
1676 				m0->m_flags |= M_TSTMP;
1677 #endif
1678 			}
1679 
1680 			/* fall through */
1681 
1682 		case X_RSPD_TYPE_CPL:
1683 			KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1684 			    ("%s: bad opcode %02x.", __func__, d->rss.opcode));
1685 			t4_cpl_handler[d->rss.opcode](iq, &d->rss, m0);
1686 			break;
1687 
1688 		case X_RSPD_TYPE_INTR:
1689 
1690 			/*
1691 			 * There are 1K interrupt-capable queues (qids 0
1692 			 * through 1023).  A response type indicating a
1693 			 * forwarded interrupt with a qid >= 1K is an
1694 			 * iWARP async notification.  That is the only
1695 			 * acceptable indirect interrupt on this queue.
1696 			 */
1697 			if (__predict_false(lq < 1024)) {
1698 				panic("%s: indirect interrupt on iq_fl %p "
1699 				    "with qid %u", __func__, iq, lq);
1700 			}
1701 
1702 			t4_an_handler(iq, &d->rsp);
1703 			break;
1704 
1705 		default:
1706 			KASSERT(0, ("%s: illegal response type %d on iq %p",
1707 			    __func__, rsp_type, iq));
1708 			log(LOG_ERR, "%s: illegal response type %d on iq %p",
1709 			    device_get_nameunit(sc->dev), rsp_type, iq);
1710 			break;
1711 		}
1712 
1713 		d++;
1714 		if (__predict_false(++iq->cidx == iq->sidx)) {
1715 			iq->cidx = 0;
1716 			iq->gen ^= F_RSPD_GEN;
1717 			d = &iq->desc[0];
1718 		}
1719 		if (__predict_false(++ndescs == limit)) {
1720 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1721 			    V_INGRESSQID(iq->cntxt_id) |
1722 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1723 			ndescs = 0;
1724 
1725 #if defined(INET) || defined(INET6)
1726 			if (iq->flags & IQ_LRO_ENABLED &&
1727 			    !sort_before_lro(lro) &&
1728 			    sc->lro_timeout != 0) {
1729 				tcp_lro_flush_inactive(lro, &lro_timeout);
1730 			}
1731 #endif
1732 			if (budget) {
1733 				FL_LOCK(fl);
1734 				refill_fl(sc, fl, 32);
1735 				FL_UNLOCK(fl);
1736 
1737 				return (EINPROGRESS);
1738 			}
1739 		}
1740 		if (refill) {
1741 			FL_LOCK(fl);
1742 			refill_fl(sc, fl, 32);
1743 			FL_UNLOCK(fl);
1744 			fl_hw_cidx = fl->hw_cidx;
1745 		}
1746 	}
1747 out:
1748 #if defined(INET) || defined(INET6)
1749 	if (iq->flags & IQ_LRO_ENABLED) {
1750 		if (ndescs > 0 && lro->lro_mbuf_count > 8) {
1751 			MPASS(sort_before_lro(lro));
1752 			/* hold back one credit and don't flush LRO state */
1753 			iq->flags |= IQ_ADJ_CREDIT;
1754 			ndescs--;
1755 		} else {
1756 			tcp_lro_flush_all(lro);
1757 		}
1758 	}
1759 #endif
1760 
1761 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1762 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1763 
1764 	FL_LOCK(fl);
1765 	starved = refill_fl(sc, fl, 64);
1766 	FL_UNLOCK(fl);
1767 	if (__predict_false(starved != 0))
1768 		add_fl_to_sfl(sc, fl);
1769 
1770 	return (0);
1771 }
1772 
1773 static inline int
1774 cl_has_metadata(struct sge_fl *fl, struct cluster_layout *cll)
1775 {
1776 	int rc = fl->flags & FL_BUF_PACKING || cll->region1 > 0;
1777 
1778 	if (rc)
1779 		MPASS(cll->region3 >= CL_METADATA_SIZE);
1780 
1781 	return (rc);
1782 }
1783 
1784 static inline struct cluster_metadata *
1785 cl_metadata(struct adapter *sc, struct sge_fl *fl, struct cluster_layout *cll,
1786     caddr_t cl)
1787 {
1788 
1789 	if (cl_has_metadata(fl, cll)) {
1790 		struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1791 
1792 		return ((struct cluster_metadata *)(cl + swz->size) - 1);
1793 	}
1794 	return (NULL);
1795 }
1796 
1797 static void
1798 rxb_free(struct mbuf *m)
1799 {
1800 	uma_zone_t zone = m->m_ext.ext_arg1;
1801 	void *cl = m->m_ext.ext_arg2;
1802 
1803 	uma_zfree(zone, cl);
1804 	counter_u64_add(extfree_rels, 1);
1805 }
1806 
1807 /*
1808  * The mbuf returned by this function could be allocated from zone_mbuf or
1809  * constructed in spare room in the cluster.
1810  *
1811  * The mbuf carries the payload in one of these ways
1812  * a) frame inside the mbuf (mbuf from zone_mbuf)
1813  * b) m_cljset (for clusters without metadata) zone_mbuf
1814  * c) m_extaddref (cluster with metadata) inline mbuf
1815  * d) m_extaddref (cluster with metadata) zone_mbuf
1816  */
1817 static struct mbuf *
1818 get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1819     int remaining)
1820 {
1821 	struct mbuf *m;
1822 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1823 	struct cluster_layout *cll = &sd->cll;
1824 	struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1825 	struct hw_buf_info *hwb = &sc->sge.hw_buf_info[cll->hwidx];
1826 	struct cluster_metadata *clm = cl_metadata(sc, fl, cll, sd->cl);
1827 	int len, blen;
1828 	caddr_t payload;
1829 
1830 	blen = hwb->size - fl->rx_offset;	/* max possible in this buf */
1831 	len = min(remaining, blen);
1832 	payload = sd->cl + cll->region1 + fl->rx_offset;
1833 	if (fl->flags & FL_BUF_PACKING) {
1834 		const u_int l = fr_offset + len;
1835 		const u_int pad = roundup2(l, fl->buf_boundary) - l;
1836 
1837 		if (fl->rx_offset + len + pad < hwb->size)
1838 			blen = len + pad;
1839 		MPASS(fl->rx_offset + blen <= hwb->size);
1840 	} else {
1841 		MPASS(fl->rx_offset == 0);	/* not packing */
1842 	}
1843 
1844 
1845 	if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) {
1846 
1847 		/*
1848 		 * Copy payload into a freshly allocated mbuf.
1849 		 */
1850 
1851 		m = fr_offset == 0 ?
1852 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1853 		if (m == NULL)
1854 			return (NULL);
1855 		fl->mbuf_allocated++;
1856 
1857 		/* copy data to mbuf */
1858 		bcopy(payload, mtod(m, caddr_t), len);
1859 
1860 	} else if (sd->nmbuf * MSIZE < cll->region1) {
1861 
1862 		/*
1863 		 * There's spare room in the cluster for an mbuf.  Create one
1864 		 * and associate it with the payload that's in the cluster.
1865 		 */
1866 
1867 		MPASS(clm != NULL);
1868 		m = (struct mbuf *)(sd->cl + sd->nmbuf * MSIZE);
1869 		/* No bzero required */
1870 		if (m_init(m, M_NOWAIT, MT_DATA,
1871 		    fr_offset == 0 ? M_PKTHDR | M_NOFREE : M_NOFREE))
1872 			return (NULL);
1873 		fl->mbuf_inlined++;
1874 		m_extaddref(m, payload, blen, &clm->refcount, rxb_free,
1875 		    swz->zone, sd->cl);
1876 		if (sd->nmbuf++ == 0)
1877 			counter_u64_add(extfree_refs, 1);
1878 
1879 	} else {
1880 
1881 		/*
1882 		 * Grab an mbuf from zone_mbuf and associate it with the
1883 		 * payload in the cluster.
1884 		 */
1885 
1886 		m = fr_offset == 0 ?
1887 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1888 		if (m == NULL)
1889 			return (NULL);
1890 		fl->mbuf_allocated++;
1891 		if (clm != NULL) {
1892 			m_extaddref(m, payload, blen, &clm->refcount,
1893 			    rxb_free, swz->zone, sd->cl);
1894 			if (sd->nmbuf++ == 0)
1895 				counter_u64_add(extfree_refs, 1);
1896 		} else {
1897 			m_cljset(m, sd->cl, swz->type);
1898 			sd->cl = NULL;	/* consumed, not a recycle candidate */
1899 		}
1900 	}
1901 	if (fr_offset == 0)
1902 		m->m_pkthdr.len = remaining;
1903 	m->m_len = len;
1904 
1905 	if (fl->flags & FL_BUF_PACKING) {
1906 		fl->rx_offset += blen;
1907 		MPASS(fl->rx_offset <= hwb->size);
1908 		if (fl->rx_offset < hwb->size)
1909 			return (m);	/* without advancing the cidx */
1910 	}
1911 
1912 	if (__predict_false(++fl->cidx % 8 == 0)) {
1913 		uint16_t cidx = fl->cidx / 8;
1914 
1915 		if (__predict_false(cidx == fl->sidx))
1916 			fl->cidx = cidx = 0;
1917 		fl->hw_cidx = cidx;
1918 	}
1919 	fl->rx_offset = 0;
1920 
1921 	return (m);
1922 }
1923 
1924 static struct mbuf *
1925 get_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf)
1926 {
1927 	struct mbuf *m0, *m, **pnext;
1928 	u_int remaining;
1929 	const u_int total = G_RSPD_LEN(len_newbuf);
1930 
1931 	if (__predict_false(fl->flags & FL_BUF_RESUME)) {
1932 		M_ASSERTPKTHDR(fl->m0);
1933 		MPASS(fl->m0->m_pkthdr.len == total);
1934 		MPASS(fl->remaining < total);
1935 
1936 		m0 = fl->m0;
1937 		pnext = fl->pnext;
1938 		remaining = fl->remaining;
1939 		fl->flags &= ~FL_BUF_RESUME;
1940 		goto get_segment;
1941 	}
1942 
1943 	if (fl->rx_offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
1944 		fl->rx_offset = 0;
1945 		if (__predict_false(++fl->cidx % 8 == 0)) {
1946 			uint16_t cidx = fl->cidx / 8;
1947 
1948 			if (__predict_false(cidx == fl->sidx))
1949 				fl->cidx = cidx = 0;
1950 			fl->hw_cidx = cidx;
1951 		}
1952 	}
1953 
1954 	/*
1955 	 * Payload starts at rx_offset in the current hw buffer.  Its length is
1956 	 * 'len' and it may span multiple hw buffers.
1957 	 */
1958 
1959 	m0 = get_scatter_segment(sc, fl, 0, total);
1960 	if (m0 == NULL)
1961 		return (NULL);
1962 	remaining = total - m0->m_len;
1963 	pnext = &m0->m_next;
1964 	while (remaining > 0) {
1965 get_segment:
1966 		MPASS(fl->rx_offset == 0);
1967 		m = get_scatter_segment(sc, fl, total - remaining, remaining);
1968 		if (__predict_false(m == NULL)) {
1969 			fl->m0 = m0;
1970 			fl->pnext = pnext;
1971 			fl->remaining = remaining;
1972 			fl->flags |= FL_BUF_RESUME;
1973 			return (NULL);
1974 		}
1975 		*pnext = m;
1976 		pnext = &m->m_next;
1977 		remaining -= m->m_len;
1978 	}
1979 	*pnext = NULL;
1980 
1981 	M_ASSERTPKTHDR(m0);
1982 	return (m0);
1983 }
1984 
1985 static int
1986 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
1987 {
1988 	struct sge_rxq *rxq = iq_to_rxq(iq);
1989 	struct ifnet *ifp = rxq->ifp;
1990 	struct adapter *sc = iq->adapter;
1991 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
1992 #if defined(INET) || defined(INET6)
1993 	struct lro_ctrl *lro = &rxq->lro;
1994 #endif
1995 	static const int sw_hashtype[4][2] = {
1996 		{M_HASHTYPE_NONE, M_HASHTYPE_NONE},
1997 		{M_HASHTYPE_RSS_IPV4, M_HASHTYPE_RSS_IPV6},
1998 		{M_HASHTYPE_RSS_TCP_IPV4, M_HASHTYPE_RSS_TCP_IPV6},
1999 		{M_HASHTYPE_RSS_UDP_IPV4, M_HASHTYPE_RSS_UDP_IPV6},
2000 	};
2001 
2002 	KASSERT(m0 != NULL, ("%s: no payload with opcode %02x", __func__,
2003 	    rss->opcode));
2004 
2005 	m0->m_pkthdr.len -= sc->params.sge.fl_pktshift;
2006 	m0->m_len -= sc->params.sge.fl_pktshift;
2007 	m0->m_data += sc->params.sge.fl_pktshift;
2008 
2009 	m0->m_pkthdr.rcvif = ifp;
2010 	M_HASHTYPE_SET(m0, sw_hashtype[rss->hash_type][rss->ipv6]);
2011 	m0->m_pkthdr.flowid = be32toh(rss->hash_val);
2012 
2013 	if (cpl->csum_calc && !(cpl->err_vec & sc->params.tp.err_vec_mask)) {
2014 		if (ifp->if_capenable & IFCAP_RXCSUM &&
2015 		    cpl->l2info & htobe32(F_RXF_IP)) {
2016 			m0->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
2017 			    CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
2018 			rxq->rxcsum++;
2019 		} else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
2020 		    cpl->l2info & htobe32(F_RXF_IP6)) {
2021 			m0->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
2022 			    CSUM_PSEUDO_HDR);
2023 			rxq->rxcsum++;
2024 		}
2025 
2026 		if (__predict_false(cpl->ip_frag))
2027 			m0->m_pkthdr.csum_data = be16toh(cpl->csum);
2028 		else
2029 			m0->m_pkthdr.csum_data = 0xffff;
2030 	}
2031 
2032 	if (cpl->vlan_ex) {
2033 		m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
2034 		m0->m_flags |= M_VLANTAG;
2035 		rxq->vlan_extraction++;
2036 	}
2037 
2038 #if defined(INET) || defined(INET6)
2039 	if (iq->flags & IQ_LRO_ENABLED) {
2040 		if (sort_before_lro(lro)) {
2041 			tcp_lro_queue_mbuf(lro, m0);
2042 			return (0); /* queued for sort, then LRO */
2043 		}
2044 		if (tcp_lro_rx(lro, m0, 0) == 0)
2045 			return (0); /* queued for LRO */
2046 	}
2047 #endif
2048 	ifp->if_input(ifp, m0);
2049 
2050 	return (0);
2051 }
2052 
2053 /*
2054  * Must drain the wrq or make sure that someone else will.
2055  */
2056 static void
2057 wrq_tx_drain(void *arg, int n)
2058 {
2059 	struct sge_wrq *wrq = arg;
2060 	struct sge_eq *eq = &wrq->eq;
2061 
2062 	EQ_LOCK(eq);
2063 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2064 		drain_wrq_wr_list(wrq->adapter, wrq);
2065 	EQ_UNLOCK(eq);
2066 }
2067 
2068 static void
2069 drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq)
2070 {
2071 	struct sge_eq *eq = &wrq->eq;
2072 	u_int available, dbdiff;	/* # of hardware descriptors */
2073 	u_int n;
2074 	struct wrqe *wr;
2075 	struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
2076 
2077 	EQ_LOCK_ASSERT_OWNED(eq);
2078 	MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
2079 	wr = STAILQ_FIRST(&wrq->wr_list);
2080 	MPASS(wr != NULL);	/* Must be called with something useful to do */
2081 	MPASS(eq->pidx == eq->dbidx);
2082 	dbdiff = 0;
2083 
2084 	do {
2085 		eq->cidx = read_hw_cidx(eq);
2086 		if (eq->pidx == eq->cidx)
2087 			available = eq->sidx - 1;
2088 		else
2089 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2090 
2091 		MPASS(wr->wrq == wrq);
2092 		n = howmany(wr->wr_len, EQ_ESIZE);
2093 		if (available < n)
2094 			break;
2095 
2096 		dst = (void *)&eq->desc[eq->pidx];
2097 		if (__predict_true(eq->sidx - eq->pidx > n)) {
2098 			/* Won't wrap, won't end exactly at the status page. */
2099 			bcopy(&wr->wr[0], dst, wr->wr_len);
2100 			eq->pidx += n;
2101 		} else {
2102 			int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE;
2103 
2104 			bcopy(&wr->wr[0], dst, first_portion);
2105 			if (wr->wr_len > first_portion) {
2106 				bcopy(&wr->wr[first_portion], &eq->desc[0],
2107 				    wr->wr_len - first_portion);
2108 			}
2109 			eq->pidx = n - (eq->sidx - eq->pidx);
2110 		}
2111 		wrq->tx_wrs_copied++;
2112 
2113 		if (available < eq->sidx / 4 &&
2114 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2115 				/*
2116 				 * XXX: This is not 100% reliable with some
2117 				 * types of WRs.  But this is a very unusual
2118 				 * situation for an ofld/ctrl queue anyway.
2119 				 */
2120 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2121 			    F_FW_WR_EQUEQ);
2122 		}
2123 
2124 		dbdiff += n;
2125 		if (dbdiff >= 16) {
2126 			ring_eq_db(sc, eq, dbdiff);
2127 			dbdiff = 0;
2128 		}
2129 
2130 		STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
2131 		free_wrqe(wr);
2132 		MPASS(wrq->nwr_pending > 0);
2133 		wrq->nwr_pending--;
2134 		MPASS(wrq->ndesc_needed >= n);
2135 		wrq->ndesc_needed -= n;
2136 	} while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL);
2137 
2138 	if (dbdiff)
2139 		ring_eq_db(sc, eq, dbdiff);
2140 }
2141 
2142 /*
2143  * Doesn't fail.  Holds on to work requests it can't send right away.
2144  */
2145 void
2146 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
2147 {
2148 #ifdef INVARIANTS
2149 	struct sge_eq *eq = &wrq->eq;
2150 #endif
2151 
2152 	EQ_LOCK_ASSERT_OWNED(eq);
2153 	MPASS(wr != NULL);
2154 	MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN);
2155 	MPASS((wr->wr_len & 0x7) == 0);
2156 
2157 	STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
2158 	wrq->nwr_pending++;
2159 	wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE);
2160 
2161 	if (!TAILQ_EMPTY(&wrq->incomplete_wrs))
2162 		return;	/* commit_wrq_wr will drain wr_list as well. */
2163 
2164 	drain_wrq_wr_list(sc, wrq);
2165 
2166 	/* Doorbell must have caught up to the pidx. */
2167 	MPASS(eq->pidx == eq->dbidx);
2168 }
2169 
2170 void
2171 t4_update_fl_bufsize(struct ifnet *ifp)
2172 {
2173 	struct vi_info *vi = ifp->if_softc;
2174 	struct adapter *sc = vi->pi->adapter;
2175 	struct sge_rxq *rxq;
2176 #ifdef TCP_OFFLOAD
2177 	struct sge_ofld_rxq *ofld_rxq;
2178 #endif
2179 	struct sge_fl *fl;
2180 	int i, maxp, mtu = ifp->if_mtu;
2181 
2182 	maxp = mtu_to_max_payload(sc, mtu, 0);
2183 	for_each_rxq(vi, i, rxq) {
2184 		fl = &rxq->fl;
2185 
2186 		FL_LOCK(fl);
2187 		find_best_refill_source(sc, fl, maxp);
2188 		FL_UNLOCK(fl);
2189 	}
2190 #ifdef TCP_OFFLOAD
2191 	maxp = mtu_to_max_payload(sc, mtu, 1);
2192 	for_each_ofld_rxq(vi, i, ofld_rxq) {
2193 		fl = &ofld_rxq->fl;
2194 
2195 		FL_LOCK(fl);
2196 		find_best_refill_source(sc, fl, maxp);
2197 		FL_UNLOCK(fl);
2198 	}
2199 #endif
2200 }
2201 
2202 static inline int
2203 mbuf_nsegs(struct mbuf *m)
2204 {
2205 
2206 	M_ASSERTPKTHDR(m);
2207 	KASSERT(m->m_pkthdr.l5hlen > 0,
2208 	    ("%s: mbuf %p missing information on # of segments.", __func__, m));
2209 
2210 	return (m->m_pkthdr.l5hlen);
2211 }
2212 
2213 static inline void
2214 set_mbuf_nsegs(struct mbuf *m, uint8_t nsegs)
2215 {
2216 
2217 	M_ASSERTPKTHDR(m);
2218 	m->m_pkthdr.l5hlen = nsegs;
2219 }
2220 
2221 static inline int
2222 mbuf_cflags(struct mbuf *m)
2223 {
2224 
2225 	M_ASSERTPKTHDR(m);
2226 	return (m->m_pkthdr.PH_loc.eight[4]);
2227 }
2228 
2229 static inline void
2230 set_mbuf_cflags(struct mbuf *m, uint8_t flags)
2231 {
2232 
2233 	M_ASSERTPKTHDR(m);
2234 	m->m_pkthdr.PH_loc.eight[4] = flags;
2235 }
2236 
2237 static inline int
2238 mbuf_len16(struct mbuf *m)
2239 {
2240 	int n;
2241 
2242 	M_ASSERTPKTHDR(m);
2243 	n = m->m_pkthdr.PH_loc.eight[0];
2244 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2245 
2246 	return (n);
2247 }
2248 
2249 static inline void
2250 set_mbuf_len16(struct mbuf *m, uint8_t len16)
2251 {
2252 
2253 	M_ASSERTPKTHDR(m);
2254 	m->m_pkthdr.PH_loc.eight[0] = len16;
2255 }
2256 
2257 #ifdef RATELIMIT
2258 static inline int
2259 mbuf_eo_nsegs(struct mbuf *m)
2260 {
2261 
2262 	M_ASSERTPKTHDR(m);
2263 	return (m->m_pkthdr.PH_loc.eight[1]);
2264 }
2265 
2266 static inline void
2267 set_mbuf_eo_nsegs(struct mbuf *m, uint8_t nsegs)
2268 {
2269 
2270 	M_ASSERTPKTHDR(m);
2271 	m->m_pkthdr.PH_loc.eight[1] = nsegs;
2272 }
2273 
2274 static inline int
2275 mbuf_eo_len16(struct mbuf *m)
2276 {
2277 	int n;
2278 
2279 	M_ASSERTPKTHDR(m);
2280 	n = m->m_pkthdr.PH_loc.eight[2];
2281 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2282 
2283 	return (n);
2284 }
2285 
2286 static inline void
2287 set_mbuf_eo_len16(struct mbuf *m, uint8_t len16)
2288 {
2289 
2290 	M_ASSERTPKTHDR(m);
2291 	m->m_pkthdr.PH_loc.eight[2] = len16;
2292 }
2293 
2294 static inline int
2295 mbuf_eo_tsclk_tsoff(struct mbuf *m)
2296 {
2297 
2298 	M_ASSERTPKTHDR(m);
2299 	return (m->m_pkthdr.PH_loc.eight[3]);
2300 }
2301 
2302 static inline void
2303 set_mbuf_eo_tsclk_tsoff(struct mbuf *m, uint8_t tsclk_tsoff)
2304 {
2305 
2306 	M_ASSERTPKTHDR(m);
2307 	m->m_pkthdr.PH_loc.eight[3] = tsclk_tsoff;
2308 }
2309 
2310 static inline int
2311 needs_eo(struct mbuf *m)
2312 {
2313 
2314 	return (m->m_pkthdr.snd_tag != NULL);
2315 }
2316 #endif
2317 
2318 /*
2319  * Try to allocate an mbuf to contain a raw work request.  To make it
2320  * easy to construct the work request, don't allocate a chain but a
2321  * single mbuf.
2322  */
2323 struct mbuf *
2324 alloc_wr_mbuf(int len, int how)
2325 {
2326 	struct mbuf *m;
2327 
2328 	if (len <= MHLEN)
2329 		m = m_gethdr(how, MT_DATA);
2330 	else if (len <= MCLBYTES)
2331 		m = m_getcl(how, MT_DATA, M_PKTHDR);
2332 	else
2333 		m = NULL;
2334 	if (m == NULL)
2335 		return (NULL);
2336 	m->m_pkthdr.len = len;
2337 	m->m_len = len;
2338 	set_mbuf_cflags(m, MC_RAW_WR);
2339 	set_mbuf_len16(m, howmany(len, 16));
2340 	return (m);
2341 }
2342 
2343 static inline int
2344 needs_tso(struct mbuf *m)
2345 {
2346 
2347 	M_ASSERTPKTHDR(m);
2348 
2349 	return (m->m_pkthdr.csum_flags & CSUM_TSO);
2350 }
2351 
2352 static inline int
2353 needs_l3_csum(struct mbuf *m)
2354 {
2355 
2356 	M_ASSERTPKTHDR(m);
2357 
2358 	return (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO));
2359 }
2360 
2361 static inline int
2362 needs_l4_csum(struct mbuf *m)
2363 {
2364 
2365 	M_ASSERTPKTHDR(m);
2366 
2367 	return (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
2368 	    CSUM_TCP_IPV6 | CSUM_TSO));
2369 }
2370 
2371 static inline int
2372 needs_tcp_csum(struct mbuf *m)
2373 {
2374 
2375 	M_ASSERTPKTHDR(m);
2376 	return (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_TCP_IPV6 | CSUM_TSO));
2377 }
2378 
2379 #ifdef RATELIMIT
2380 static inline int
2381 needs_udp_csum(struct mbuf *m)
2382 {
2383 
2384 	M_ASSERTPKTHDR(m);
2385 	return (m->m_pkthdr.csum_flags & (CSUM_UDP | CSUM_UDP_IPV6));
2386 }
2387 #endif
2388 
2389 static inline int
2390 needs_vlan_insertion(struct mbuf *m)
2391 {
2392 
2393 	M_ASSERTPKTHDR(m);
2394 
2395 	return (m->m_flags & M_VLANTAG);
2396 }
2397 
2398 static void *
2399 m_advance(struct mbuf **pm, int *poffset, int len)
2400 {
2401 	struct mbuf *m = *pm;
2402 	int offset = *poffset;
2403 	uintptr_t p = 0;
2404 
2405 	MPASS(len > 0);
2406 
2407 	for (;;) {
2408 		if (offset + len < m->m_len) {
2409 			offset += len;
2410 			p = mtod(m, uintptr_t) + offset;
2411 			break;
2412 		}
2413 		len -= m->m_len - offset;
2414 		m = m->m_next;
2415 		offset = 0;
2416 		MPASS(m != NULL);
2417 	}
2418 	*poffset = offset;
2419 	*pm = m;
2420 	return ((void *)p);
2421 }
2422 
2423 /*
2424  * Can deal with empty mbufs in the chain that have m_len = 0, but the chain
2425  * must have at least one mbuf that's not empty.  It is possible for this
2426  * routine to return 0 if skip accounts for all the contents of the mbuf chain.
2427  */
2428 static inline int
2429 count_mbuf_nsegs(struct mbuf *m, int skip)
2430 {
2431 	vm_paddr_t lastb, next;
2432 	vm_offset_t va;
2433 	int len, nsegs;
2434 
2435 	M_ASSERTPKTHDR(m);
2436 	MPASS(m->m_pkthdr.len > 0);
2437 	MPASS(m->m_pkthdr.len >= skip);
2438 
2439 	nsegs = 0;
2440 	lastb = 0;
2441 	for (; m; m = m->m_next) {
2442 
2443 		len = m->m_len;
2444 		if (__predict_false(len == 0))
2445 			continue;
2446 		if (skip >= len) {
2447 			skip -= len;
2448 			continue;
2449 		}
2450 		va = mtod(m, vm_offset_t) + skip;
2451 		len -= skip;
2452 		skip = 0;
2453 		next = pmap_kextract(va);
2454 		nsegs += sglist_count((void *)(uintptr_t)va, len);
2455 		if (lastb + 1 == next)
2456 			nsegs--;
2457 		lastb = pmap_kextract(va + len - 1);
2458 	}
2459 
2460 	return (nsegs);
2461 }
2462 
2463 /*
2464  * Analyze the mbuf to determine its tx needs.  The mbuf passed in may change:
2465  * a) caller can assume it's been freed if this function returns with an error.
2466  * b) it may get defragged up if the gather list is too long for the hardware.
2467  */
2468 int
2469 parse_pkt(struct adapter *sc, struct mbuf **mp)
2470 {
2471 	struct mbuf *m0 = *mp, *m;
2472 	int rc, nsegs, defragged = 0, offset;
2473 	struct ether_header *eh;
2474 	void *l3hdr;
2475 #if defined(INET) || defined(INET6)
2476 	struct tcphdr *tcp;
2477 #endif
2478 	uint16_t eh_type;
2479 
2480 	M_ASSERTPKTHDR(m0);
2481 	if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) {
2482 		rc = EINVAL;
2483 fail:
2484 		m_freem(m0);
2485 		*mp = NULL;
2486 		return (rc);
2487 	}
2488 restart:
2489 	/*
2490 	 * First count the number of gather list segments in the payload.
2491 	 * Defrag the mbuf if nsegs exceeds the hardware limit.
2492 	 */
2493 	M_ASSERTPKTHDR(m0);
2494 	MPASS(m0->m_pkthdr.len > 0);
2495 	nsegs = count_mbuf_nsegs(m0, 0);
2496 	if (nsegs > (needs_tso(m0) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS)) {
2497 		if (defragged++ > 0 || (m = m_defrag(m0, M_NOWAIT)) == NULL) {
2498 			rc = EFBIG;
2499 			goto fail;
2500 		}
2501 		*mp = m0 = m;	/* update caller's copy after defrag */
2502 		goto restart;
2503 	}
2504 
2505 	if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN)) {
2506 		m0 = m_pullup(m0, m0->m_pkthdr.len);
2507 		if (m0 == NULL) {
2508 			/* Should have left well enough alone. */
2509 			rc = EFBIG;
2510 			goto fail;
2511 		}
2512 		*mp = m0;	/* update caller's copy after pullup */
2513 		goto restart;
2514 	}
2515 	set_mbuf_nsegs(m0, nsegs);
2516 	set_mbuf_cflags(m0, 0);
2517 	if (sc->flags & IS_VF)
2518 		set_mbuf_len16(m0, txpkt_vm_len16(nsegs, needs_tso(m0)));
2519 	else
2520 		set_mbuf_len16(m0, txpkt_len16(nsegs, needs_tso(m0)));
2521 
2522 #ifdef RATELIMIT
2523 	/*
2524 	 * Ethofld is limited to TCP and UDP for now, and only when L4 hw
2525 	 * checksumming is enabled.  needs_l4_csum happens to check for all the
2526 	 * right things.
2527 	 */
2528 	if (__predict_false(needs_eo(m0) && !needs_l4_csum(m0)))
2529 		m0->m_pkthdr.snd_tag = NULL;
2530 #endif
2531 
2532 	if (!needs_tso(m0) &&
2533 #ifdef RATELIMIT
2534 	    !needs_eo(m0) &&
2535 #endif
2536 	    !(sc->flags & IS_VF && (needs_l3_csum(m0) || needs_l4_csum(m0))))
2537 		return (0);
2538 
2539 	m = m0;
2540 	eh = mtod(m, struct ether_header *);
2541 	eh_type = ntohs(eh->ether_type);
2542 	if (eh_type == ETHERTYPE_VLAN) {
2543 		struct ether_vlan_header *evh = (void *)eh;
2544 
2545 		eh_type = ntohs(evh->evl_proto);
2546 		m0->m_pkthdr.l2hlen = sizeof(*evh);
2547 	} else
2548 		m0->m_pkthdr.l2hlen = sizeof(*eh);
2549 
2550 	offset = 0;
2551 	l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2552 
2553 	switch (eh_type) {
2554 #ifdef INET6
2555 	case ETHERTYPE_IPV6:
2556 	{
2557 		struct ip6_hdr *ip6 = l3hdr;
2558 
2559 		MPASS(!needs_tso(m0) || ip6->ip6_nxt == IPPROTO_TCP);
2560 
2561 		m0->m_pkthdr.l3hlen = sizeof(*ip6);
2562 		break;
2563 	}
2564 #endif
2565 #ifdef INET
2566 	case ETHERTYPE_IP:
2567 	{
2568 		struct ip *ip = l3hdr;
2569 
2570 		m0->m_pkthdr.l3hlen = ip->ip_hl * 4;
2571 		break;
2572 	}
2573 #endif
2574 	default:
2575 		panic("%s: ethertype 0x%04x unknown.  if_cxgbe must be compiled"
2576 		    " with the same INET/INET6 options as the kernel.",
2577 		    __func__, eh_type);
2578 	}
2579 
2580 #if defined(INET) || defined(INET6)
2581 	if (needs_tcp_csum(m0)) {
2582 		tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen);
2583 		m0->m_pkthdr.l4hlen = tcp->th_off * 4;
2584 #ifdef RATELIMIT
2585 		if (tsclk >= 0 && *(uint32_t *)(tcp + 1) == ntohl(0x0101080a)) {
2586 			set_mbuf_eo_tsclk_tsoff(m0,
2587 			    V_FW_ETH_TX_EO_WR_TSCLK(tsclk) |
2588 			    V_FW_ETH_TX_EO_WR_TSOFF(sizeof(*tcp) / 2 + 1));
2589 		} else
2590 			set_mbuf_eo_tsclk_tsoff(m0, 0);
2591 	} else if (needs_udp_csum(m)) {
2592 		m0->m_pkthdr.l4hlen = sizeof(struct udphdr);
2593 #endif
2594 	}
2595 #ifdef RATELIMIT
2596 	if (needs_eo(m0)) {
2597 		u_int immhdrs;
2598 
2599 		/* EO WRs have the headers in the WR and not the GL. */
2600 		immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen +
2601 		    m0->m_pkthdr.l4hlen;
2602 		nsegs = count_mbuf_nsegs(m0, immhdrs);
2603 		set_mbuf_eo_nsegs(m0, nsegs);
2604 		set_mbuf_eo_len16(m0,
2605 		    txpkt_eo_len16(nsegs, immhdrs, needs_tso(m0)));
2606 	}
2607 #endif
2608 #endif
2609 	MPASS(m0 == *mp);
2610 	return (0);
2611 }
2612 
2613 void *
2614 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie)
2615 {
2616 	struct sge_eq *eq = &wrq->eq;
2617 	struct adapter *sc = wrq->adapter;
2618 	int ndesc, available;
2619 	struct wrqe *wr;
2620 	void *w;
2621 
2622 	MPASS(len16 > 0);
2623 	ndesc = howmany(len16, EQ_ESIZE / 16);
2624 	MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC);
2625 
2626 	EQ_LOCK(eq);
2627 
2628 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2629 		drain_wrq_wr_list(sc, wrq);
2630 
2631 	if (!STAILQ_EMPTY(&wrq->wr_list)) {
2632 slowpath:
2633 		EQ_UNLOCK(eq);
2634 		wr = alloc_wrqe(len16 * 16, wrq);
2635 		if (__predict_false(wr == NULL))
2636 			return (NULL);
2637 		cookie->pidx = -1;
2638 		cookie->ndesc = ndesc;
2639 		return (&wr->wr);
2640 	}
2641 
2642 	eq->cidx = read_hw_cidx(eq);
2643 	if (eq->pidx == eq->cidx)
2644 		available = eq->sidx - 1;
2645 	else
2646 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2647 	if (available < ndesc)
2648 		goto slowpath;
2649 
2650 	cookie->pidx = eq->pidx;
2651 	cookie->ndesc = ndesc;
2652 	TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link);
2653 
2654 	w = &eq->desc[eq->pidx];
2655 	IDXINCR(eq->pidx, ndesc, eq->sidx);
2656 	if (__predict_false(cookie->pidx + ndesc > eq->sidx)) {
2657 		w = &wrq->ss[0];
2658 		wrq->ss_pidx = cookie->pidx;
2659 		wrq->ss_len = len16 * 16;
2660 	}
2661 
2662 	EQ_UNLOCK(eq);
2663 
2664 	return (w);
2665 }
2666 
2667 void
2668 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie)
2669 {
2670 	struct sge_eq *eq = &wrq->eq;
2671 	struct adapter *sc = wrq->adapter;
2672 	int ndesc, pidx;
2673 	struct wrq_cookie *prev, *next;
2674 
2675 	if (cookie->pidx == -1) {
2676 		struct wrqe *wr = __containerof(w, struct wrqe, wr);
2677 
2678 		t4_wrq_tx(sc, wr);
2679 		return;
2680 	}
2681 
2682 	if (__predict_false(w == &wrq->ss[0])) {
2683 		int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE;
2684 
2685 		MPASS(wrq->ss_len > n);	/* WR had better wrap around. */
2686 		bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n);
2687 		bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n);
2688 		wrq->tx_wrs_ss++;
2689 	} else
2690 		wrq->tx_wrs_direct++;
2691 
2692 	EQ_LOCK(eq);
2693 	ndesc = cookie->ndesc;	/* Can be more than SGE_MAX_WR_NDESC here. */
2694 	pidx = cookie->pidx;
2695 	MPASS(pidx >= 0 && pidx < eq->sidx);
2696 	prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link);
2697 	next = TAILQ_NEXT(cookie, link);
2698 	if (prev == NULL) {
2699 		MPASS(pidx == eq->dbidx);
2700 		if (next == NULL || ndesc >= 16) {
2701 			int available;
2702 			struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
2703 
2704 			/*
2705 			 * Note that the WR via which we'll request tx updates
2706 			 * is at pidx and not eq->pidx, which has moved on
2707 			 * already.
2708 			 */
2709 			dst = (void *)&eq->desc[pidx];
2710 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2711 			if (available < eq->sidx / 4 &&
2712 			    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2713 				/*
2714 				 * XXX: This is not 100% reliable with some
2715 				 * types of WRs.  But this is a very unusual
2716 				 * situation for an ofld/ctrl queue anyway.
2717 				 */
2718 				dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2719 				    F_FW_WR_EQUEQ);
2720 			}
2721 
2722 			ring_eq_db(wrq->adapter, eq, ndesc);
2723 		} else {
2724 			MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc);
2725 			next->pidx = pidx;
2726 			next->ndesc += ndesc;
2727 		}
2728 	} else {
2729 		MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc);
2730 		prev->ndesc += ndesc;
2731 	}
2732 	TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link);
2733 
2734 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2735 		drain_wrq_wr_list(sc, wrq);
2736 
2737 #ifdef INVARIANTS
2738 	if (TAILQ_EMPTY(&wrq->incomplete_wrs)) {
2739 		/* Doorbell must have caught up to the pidx. */
2740 		MPASS(wrq->eq.pidx == wrq->eq.dbidx);
2741 	}
2742 #endif
2743 	EQ_UNLOCK(eq);
2744 }
2745 
2746 static u_int
2747 can_resume_eth_tx(struct mp_ring *r)
2748 {
2749 	struct sge_eq *eq = r->cookie;
2750 
2751 	return (total_available_tx_desc(eq) > eq->sidx / 8);
2752 }
2753 
2754 static inline int
2755 cannot_use_txpkts(struct mbuf *m)
2756 {
2757 	/* maybe put a GL limit too, to avoid silliness? */
2758 
2759 	return (needs_tso(m) || (mbuf_cflags(m) & MC_RAW_WR) != 0);
2760 }
2761 
2762 static inline int
2763 discard_tx(struct sge_eq *eq)
2764 {
2765 
2766 	return ((eq->flags & (EQ_ENABLED | EQ_QFLUSH)) != EQ_ENABLED);
2767 }
2768 
2769 static inline int
2770 wr_can_update_eq(struct fw_eth_tx_pkts_wr *wr)
2771 {
2772 
2773 	switch (G_FW_WR_OP(be32toh(wr->op_pkd))) {
2774 	case FW_ULPTX_WR:
2775 	case FW_ETH_TX_PKT_WR:
2776 	case FW_ETH_TX_PKTS_WR:
2777 	case FW_ETH_TX_PKT_VM_WR:
2778 		return (1);
2779 	default:
2780 		return (0);
2781 	}
2782 }
2783 
2784 /*
2785  * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to
2786  * be consumed.  Return the actual number consumed.  0 indicates a stall.
2787  */
2788 static u_int
2789 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx)
2790 {
2791 	struct sge_txq *txq = r->cookie;
2792 	struct sge_eq *eq = &txq->eq;
2793 	struct ifnet *ifp = txq->ifp;
2794 	struct vi_info *vi = ifp->if_softc;
2795 	struct port_info *pi = vi->pi;
2796 	struct adapter *sc = pi->adapter;
2797 	u_int total, remaining;		/* # of packets */
2798 	u_int available, dbdiff;	/* # of hardware descriptors */
2799 	u_int n, next_cidx;
2800 	struct mbuf *m0, *tail;
2801 	struct txpkts txp;
2802 	struct fw_eth_tx_pkts_wr *wr;	/* any fw WR struct will do */
2803 
2804 	remaining = IDXDIFF(pidx, cidx, r->size);
2805 	MPASS(remaining > 0);	/* Must not be called without work to do. */
2806 	total = 0;
2807 
2808 	TXQ_LOCK(txq);
2809 	if (__predict_false(discard_tx(eq))) {
2810 		while (cidx != pidx) {
2811 			m0 = r->items[cidx];
2812 			m_freem(m0);
2813 			if (++cidx == r->size)
2814 				cidx = 0;
2815 		}
2816 		reclaim_tx_descs(txq, 2048);
2817 		total = remaining;
2818 		goto done;
2819 	}
2820 
2821 	/* How many hardware descriptors do we have readily available. */
2822 	if (eq->pidx == eq->cidx)
2823 		available = eq->sidx - 1;
2824 	else
2825 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2826 	dbdiff = IDXDIFF(eq->pidx, eq->dbidx, eq->sidx);
2827 
2828 	while (remaining > 0) {
2829 
2830 		m0 = r->items[cidx];
2831 		M_ASSERTPKTHDR(m0);
2832 		MPASS(m0->m_nextpkt == NULL);
2833 
2834 		if (available < SGE_MAX_WR_NDESC) {
2835 			available += reclaim_tx_descs(txq, 64);
2836 			if (available < howmany(mbuf_len16(m0), EQ_ESIZE / 16))
2837 				break;	/* out of descriptors */
2838 		}
2839 
2840 		next_cidx = cidx + 1;
2841 		if (__predict_false(next_cidx == r->size))
2842 			next_cidx = 0;
2843 
2844 		wr = (void *)&eq->desc[eq->pidx];
2845 		if (sc->flags & IS_VF) {
2846 			total++;
2847 			remaining--;
2848 			ETHER_BPF_MTAP(ifp, m0);
2849 			n = write_txpkt_vm_wr(sc, txq, (void *)wr, m0,
2850 			    available);
2851 		} else if (remaining > 1 &&
2852 		    try_txpkts(m0, r->items[next_cidx], &txp, available) == 0) {
2853 
2854 			/* pkts at cidx, next_cidx should both be in txp. */
2855 			MPASS(txp.npkt == 2);
2856 			tail = r->items[next_cidx];
2857 			MPASS(tail->m_nextpkt == NULL);
2858 			ETHER_BPF_MTAP(ifp, m0);
2859 			ETHER_BPF_MTAP(ifp, tail);
2860 			m0->m_nextpkt = tail;
2861 
2862 			if (__predict_false(++next_cidx == r->size))
2863 				next_cidx = 0;
2864 
2865 			while (next_cidx != pidx) {
2866 				if (add_to_txpkts(r->items[next_cidx], &txp,
2867 				    available) != 0)
2868 					break;
2869 				tail->m_nextpkt = r->items[next_cidx];
2870 				tail = tail->m_nextpkt;
2871 				ETHER_BPF_MTAP(ifp, tail);
2872 				if (__predict_false(++next_cidx == r->size))
2873 					next_cidx = 0;
2874 			}
2875 
2876 			n = write_txpkts_wr(txq, wr, m0, &txp, available);
2877 			total += txp.npkt;
2878 			remaining -= txp.npkt;
2879 		} else if (mbuf_cflags(m0) & MC_RAW_WR) {
2880 			total++;
2881 			remaining--;
2882 			n = write_raw_wr(txq, (void *)wr, m0, available);
2883 		} else {
2884 			total++;
2885 			remaining--;
2886 			ETHER_BPF_MTAP(ifp, m0);
2887 			n = write_txpkt_wr(txq, (void *)wr, m0, available);
2888 		}
2889 		MPASS(n >= 1 && n <= available && n <= SGE_MAX_WR_NDESC);
2890 
2891 		available -= n;
2892 		dbdiff += n;
2893 		IDXINCR(eq->pidx, n, eq->sidx);
2894 
2895 		if (wr_can_update_eq(wr)) {
2896 			if (total_available_tx_desc(eq) < eq->sidx / 4 &&
2897 			    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2898 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2899 				    F_FW_WR_EQUEQ);
2900 				eq->equeqidx = eq->pidx;
2901 			} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >=
2902 			    32) {
2903 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
2904 				eq->equeqidx = eq->pidx;
2905 			}
2906 		}
2907 
2908 		if (dbdiff >= 16 && remaining >= 4) {
2909 			ring_eq_db(sc, eq, dbdiff);
2910 			available += reclaim_tx_descs(txq, 4 * dbdiff);
2911 			dbdiff = 0;
2912 		}
2913 
2914 		cidx = next_cidx;
2915 	}
2916 	if (dbdiff != 0) {
2917 		ring_eq_db(sc, eq, dbdiff);
2918 		reclaim_tx_descs(txq, 32);
2919 	}
2920 done:
2921 	TXQ_UNLOCK(txq);
2922 
2923 	return (total);
2924 }
2925 
2926 static inline void
2927 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
2928     int qsize)
2929 {
2930 
2931 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
2932 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
2933 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
2934 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
2935 
2936 	iq->flags = 0;
2937 	iq->adapter = sc;
2938 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
2939 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
2940 	if (pktc_idx >= 0) {
2941 		iq->intr_params |= F_QINTR_CNT_EN;
2942 		iq->intr_pktc_idx = pktc_idx;
2943 	}
2944 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
2945 	iq->sidx = iq->qsize - sc->params.sge.spg_len / IQ_ESIZE;
2946 }
2947 
2948 static inline void
2949 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name)
2950 {
2951 
2952 	fl->qsize = qsize;
2953 	fl->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2954 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
2955 	if (sc->flags & BUF_PACKING_OK &&
2956 	    ((!is_t4(sc) && buffer_packing) ||	/* T5+: enabled unless 0 */
2957 	    (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */
2958 		fl->flags |= FL_BUF_PACKING;
2959 	find_best_refill_source(sc, fl, maxp);
2960 	find_safe_refill_source(sc, fl);
2961 }
2962 
2963 static inline void
2964 init_eq(struct adapter *sc, struct sge_eq *eq, int eqtype, int qsize,
2965     uint8_t tx_chan, uint16_t iqid, char *name)
2966 {
2967 	KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
2968 
2969 	eq->flags = eqtype & EQ_TYPEMASK;
2970 	eq->tx_chan = tx_chan;
2971 	eq->iqid = iqid;
2972 	eq->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2973 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
2974 }
2975 
2976 static int
2977 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
2978     bus_dmamap_t *map, bus_addr_t *pa, void **va)
2979 {
2980 	int rc;
2981 
2982 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
2983 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
2984 	if (rc != 0) {
2985 		device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
2986 		goto done;
2987 	}
2988 
2989 	rc = bus_dmamem_alloc(*tag, va,
2990 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
2991 	if (rc != 0) {
2992 		device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
2993 		goto done;
2994 	}
2995 
2996 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
2997 	if (rc != 0) {
2998 		device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
2999 		goto done;
3000 	}
3001 done:
3002 	if (rc)
3003 		free_ring(sc, *tag, *map, *pa, *va);
3004 
3005 	return (rc);
3006 }
3007 
3008 static int
3009 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
3010     bus_addr_t pa, void *va)
3011 {
3012 	if (pa)
3013 		bus_dmamap_unload(tag, map);
3014 	if (va)
3015 		bus_dmamem_free(tag, va, map);
3016 	if (tag)
3017 		bus_dma_tag_destroy(tag);
3018 
3019 	return (0);
3020 }
3021 
3022 /*
3023  * Allocates the ring for an ingress queue and an optional freelist.  If the
3024  * freelist is specified it will be allocated and then associated with the
3025  * ingress queue.
3026  *
3027  * Returns errno on failure.  Resources allocated up to that point may still be
3028  * allocated.  Caller is responsible for cleanup in case this function fails.
3029  *
3030  * If the ingress queue will take interrupts directly then the intr_idx
3031  * specifies the vector, starting from 0.  -1 means the interrupts for this
3032  * queue should be forwarded to the fwq.
3033  */
3034 static int
3035 alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl,
3036     int intr_idx, int cong)
3037 {
3038 	int rc, i, cntxt_id;
3039 	size_t len;
3040 	struct fw_iq_cmd c;
3041 	struct port_info *pi = vi->pi;
3042 	struct adapter *sc = iq->adapter;
3043 	struct sge_params *sp = &sc->params.sge;
3044 	__be32 v = 0;
3045 
3046 	len = iq->qsize * IQ_ESIZE;
3047 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
3048 	    (void **)&iq->desc);
3049 	if (rc != 0)
3050 		return (rc);
3051 
3052 	bzero(&c, sizeof(c));
3053 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
3054 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
3055 	    V_FW_IQ_CMD_VFN(0));
3056 
3057 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
3058 	    FW_LEN16(c));
3059 
3060 	/* Special handling for firmware event queue */
3061 	if (iq == &sc->sge.fwq)
3062 		v |= F_FW_IQ_CMD_IQASYNCH;
3063 
3064 	if (intr_idx < 0) {
3065 		/* Forwarded interrupts, all headed to fwq */
3066 		v |= F_FW_IQ_CMD_IQANDST;
3067 		v |= V_FW_IQ_CMD_IQANDSTINDEX(sc->sge.fwq.cntxt_id);
3068 	} else {
3069 		KASSERT(intr_idx < sc->intr_count,
3070 		    ("%s: invalid direct intr_idx %d", __func__, intr_idx));
3071 		v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
3072 	}
3073 
3074 	c.type_to_iqandstindex = htobe32(v |
3075 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
3076 	    V_FW_IQ_CMD_VIID(vi->viid) |
3077 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
3078 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
3079 	    F_FW_IQ_CMD_IQGTSMODE |
3080 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
3081 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
3082 	c.iqsize = htobe16(iq->qsize);
3083 	c.iqaddr = htobe64(iq->ba);
3084 	if (cong >= 0)
3085 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
3086 
3087 	if (fl) {
3088 		mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
3089 
3090 		len = fl->qsize * EQ_ESIZE;
3091 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
3092 		    &fl->ba, (void **)&fl->desc);
3093 		if (rc)
3094 			return (rc);
3095 
3096 		/* Allocate space for one software descriptor per buffer. */
3097 		rc = alloc_fl_sdesc(fl);
3098 		if (rc != 0) {
3099 			device_printf(sc->dev,
3100 			    "failed to setup fl software descriptors: %d\n",
3101 			    rc);
3102 			return (rc);
3103 		}
3104 
3105 		if (fl->flags & FL_BUF_PACKING) {
3106 			fl->lowat = roundup2(sp->fl_starve_threshold2, 8);
3107 			fl->buf_boundary = sp->pack_boundary;
3108 		} else {
3109 			fl->lowat = roundup2(sp->fl_starve_threshold, 8);
3110 			fl->buf_boundary = 16;
3111 		}
3112 		if (fl_pad && fl->buf_boundary < sp->pad_boundary)
3113 			fl->buf_boundary = sp->pad_boundary;
3114 
3115 		c.iqns_to_fl0congen |=
3116 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
3117 			F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
3118 			(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
3119 			(fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
3120 			    0));
3121 		if (cong >= 0) {
3122 			c.iqns_to_fl0congen |=
3123 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
3124 				    F_FW_IQ_CMD_FL0CONGCIF |
3125 				    F_FW_IQ_CMD_FL0CONGEN);
3126 		}
3127 		c.fl0dcaen_to_fl0cidxfthresh =
3128 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
3129 			X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B) |
3130 			V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
3131 			X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
3132 		c.fl0size = htobe16(fl->qsize);
3133 		c.fl0addr = htobe64(fl->ba);
3134 	}
3135 
3136 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3137 	if (rc != 0) {
3138 		device_printf(sc->dev,
3139 		    "failed to create ingress queue: %d\n", rc);
3140 		return (rc);
3141 	}
3142 
3143 	iq->cidx = 0;
3144 	iq->gen = F_RSPD_GEN;
3145 	iq->intr_next = iq->intr_params;
3146 	iq->cntxt_id = be16toh(c.iqid);
3147 	iq->abs_id = be16toh(c.physiqid);
3148 	iq->flags |= IQ_ALLOCATED;
3149 
3150 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
3151 	if (cntxt_id >= sc->sge.niq) {
3152 		panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
3153 		    cntxt_id, sc->sge.niq - 1);
3154 	}
3155 	sc->sge.iqmap[cntxt_id] = iq;
3156 
3157 	if (fl) {
3158 		u_int qid;
3159 
3160 		iq->flags |= IQ_HAS_FL;
3161 		fl->cntxt_id = be16toh(c.fl0id);
3162 		fl->pidx = fl->cidx = 0;
3163 
3164 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
3165 		if (cntxt_id >= sc->sge.neq) {
3166 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
3167 			    __func__, cntxt_id, sc->sge.neq - 1);
3168 		}
3169 		sc->sge.eqmap[cntxt_id] = (void *)fl;
3170 
3171 		qid = fl->cntxt_id;
3172 		if (isset(&sc->doorbells, DOORBELL_UDB)) {
3173 			uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3174 			uint32_t mask = (1 << s_qpp) - 1;
3175 			volatile uint8_t *udb;
3176 
3177 			udb = sc->udbs_base + UDBS_DB_OFFSET;
3178 			udb += (qid >> s_qpp) << PAGE_SHIFT;
3179 			qid &= mask;
3180 			if (qid < PAGE_SIZE / UDBS_SEG_SIZE) {
3181 				udb += qid << UDBS_SEG_SHIFT;
3182 				qid = 0;
3183 			}
3184 			fl->udb = (volatile void *)udb;
3185 		}
3186 		fl->dbval = V_QID(qid) | sc->chip_params->sge_fl_db;
3187 
3188 		FL_LOCK(fl);
3189 		/* Enough to make sure the SGE doesn't think it's starved */
3190 		refill_fl(sc, fl, fl->lowat);
3191 		FL_UNLOCK(fl);
3192 	}
3193 
3194 	if (chip_id(sc) >= CHELSIO_T5 && !(sc->flags & IS_VF) && cong >= 0) {
3195 		uint32_t param, val;
3196 
3197 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
3198 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
3199 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
3200 		if (cong == 0)
3201 			val = 1 << 19;
3202 		else {
3203 			val = 2 << 19;
3204 			for (i = 0; i < 4; i++) {
3205 				if (cong & (1 << i))
3206 					val |= 1 << (i << 2);
3207 			}
3208 		}
3209 
3210 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
3211 		if (rc != 0) {
3212 			/* report error but carry on */
3213 			device_printf(sc->dev,
3214 			    "failed to set congestion manager context for "
3215 			    "ingress queue %d: %d\n", iq->cntxt_id, rc);
3216 		}
3217 	}
3218 
3219 	/* Enable IQ interrupts */
3220 	atomic_store_rel_int(&iq->state, IQS_IDLE);
3221 	t4_write_reg(sc, sc->sge_gts_reg, V_SEINTARM(iq->intr_params) |
3222 	    V_INGRESSQID(iq->cntxt_id));
3223 
3224 	return (0);
3225 }
3226 
3227 static int
3228 free_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl)
3229 {
3230 	int rc;
3231 	struct adapter *sc = iq->adapter;
3232 	device_t dev;
3233 
3234 	if (sc == NULL)
3235 		return (0);	/* nothing to do */
3236 
3237 	dev = vi ? vi->dev : sc->dev;
3238 
3239 	if (iq->flags & IQ_ALLOCATED) {
3240 		rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
3241 		    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
3242 		    fl ? fl->cntxt_id : 0xffff, 0xffff);
3243 		if (rc != 0) {
3244 			device_printf(dev,
3245 			    "failed to free queue %p: %d\n", iq, rc);
3246 			return (rc);
3247 		}
3248 		iq->flags &= ~IQ_ALLOCATED;
3249 	}
3250 
3251 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
3252 
3253 	bzero(iq, sizeof(*iq));
3254 
3255 	if (fl) {
3256 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
3257 		    fl->desc);
3258 
3259 		if (fl->sdesc)
3260 			free_fl_sdesc(sc, fl);
3261 
3262 		if (mtx_initialized(&fl->fl_lock))
3263 			mtx_destroy(&fl->fl_lock);
3264 
3265 		bzero(fl, sizeof(*fl));
3266 	}
3267 
3268 	return (0);
3269 }
3270 
3271 static void
3272 add_iq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
3273     struct sge_iq *iq)
3274 {
3275 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3276 
3277 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD, &iq->ba,
3278 	    "bus address of descriptor ring");
3279 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3280 	    iq->qsize * IQ_ESIZE, "descriptor ring size in bytes");
3281 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3282 	    CTLTYPE_INT | CTLFLAG_RD, &iq->abs_id, 0, sysctl_uint16, "I",
3283 	    "absolute id of the queue");
3284 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3285 	    CTLTYPE_INT | CTLFLAG_RD, &iq->cntxt_id, 0, sysctl_uint16, "I",
3286 	    "SGE context id of the queue");
3287 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3288 	    CTLTYPE_INT | CTLFLAG_RD, &iq->cidx, 0, sysctl_uint16, "I",
3289 	    "consumer index");
3290 }
3291 
3292 static void
3293 add_fl_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
3294     struct sysctl_oid *oid, struct sge_fl *fl)
3295 {
3296 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3297 
3298 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3299 	    "freelist");
3300 	children = SYSCTL_CHILDREN(oid);
3301 
3302 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3303 	    &fl->ba, "bus address of descriptor ring");
3304 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3305 	    fl->sidx * EQ_ESIZE + sc->params.sge.spg_len,
3306 	    "desc ring size in bytes");
3307 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3308 	    CTLTYPE_INT | CTLFLAG_RD, &fl->cntxt_id, 0, sysctl_uint16, "I",
3309 	    "SGE context id of the freelist");
3310 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL,
3311 	    fl_pad ? 1 : 0, "padding enabled");
3312 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL,
3313 	    fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled");
3314 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
3315 	    0, "consumer index");
3316 	if (fl->flags & FL_BUF_PACKING) {
3317 		SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
3318 		    CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
3319 	}
3320 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
3321 	    0, "producer index");
3322 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_allocated",
3323 	    CTLFLAG_RD, &fl->mbuf_allocated, "# of mbuf allocated");
3324 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_inlined",
3325 	    CTLFLAG_RD, &fl->mbuf_inlined, "# of mbuf inlined in clusters");
3326 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
3327 	    CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
3328 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
3329 	    CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
3330 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
3331 	    CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
3332 }
3333 
3334 static int
3335 alloc_fwq(struct adapter *sc)
3336 {
3337 	int rc, intr_idx;
3338 	struct sge_iq *fwq = &sc->sge.fwq;
3339 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
3340 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3341 
3342 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE);
3343 	if (sc->flags & IS_VF)
3344 		intr_idx = 0;
3345 	else
3346 		intr_idx = sc->intr_count > 1 ? 1 : 0;
3347 	rc = alloc_iq_fl(&sc->port[0]->vi[0], fwq, NULL, intr_idx, -1);
3348 	if (rc != 0) {
3349 		device_printf(sc->dev,
3350 		    "failed to create firmware event queue: %d\n", rc);
3351 		return (rc);
3352 	}
3353 
3354 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD,
3355 	    NULL, "firmware event queue");
3356 	add_iq_sysctls(&sc->ctx, oid, fwq);
3357 
3358 	return (0);
3359 }
3360 
3361 static int
3362 free_fwq(struct adapter *sc)
3363 {
3364 	return free_iq_fl(NULL, &sc->sge.fwq, NULL);
3365 }
3366 
3367 static int
3368 alloc_ctrlq(struct adapter *sc, struct sge_wrq *ctrlq, int idx,
3369     struct sysctl_oid *oid)
3370 {
3371 	int rc;
3372 	char name[16];
3373 	struct sysctl_oid_list *children;
3374 
3375 	snprintf(name, sizeof(name), "%s ctrlq%d", device_get_nameunit(sc->dev),
3376 	    idx);
3377 	init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[idx]->tx_chan,
3378 	    sc->sge.fwq.cntxt_id, name);
3379 
3380 	children = SYSCTL_CHILDREN(oid);
3381 	snprintf(name, sizeof(name), "%d", idx);
3382 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3383 	    NULL, "ctrl queue");
3384 	rc = alloc_wrq(sc, NULL, ctrlq, oid);
3385 
3386 	return (rc);
3387 }
3388 
3389 int
3390 tnl_cong(struct port_info *pi, int drop)
3391 {
3392 
3393 	if (drop == -1)
3394 		return (-1);
3395 	else if (drop == 1)
3396 		return (0);
3397 	else
3398 		return (pi->rx_e_chan_map);
3399 }
3400 
3401 static int
3402 alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int intr_idx, int idx,
3403     struct sysctl_oid *oid)
3404 {
3405 	int rc;
3406 	struct adapter *sc = vi->pi->adapter;
3407 	struct sysctl_oid_list *children;
3408 	char name[16];
3409 
3410 	rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, intr_idx,
3411 	    tnl_cong(vi->pi, cong_drop));
3412 	if (rc != 0)
3413 		return (rc);
3414 
3415 	if (idx == 0)
3416 		sc->sge.iq_base = rxq->iq.abs_id - rxq->iq.cntxt_id;
3417 	else
3418 		KASSERT(rxq->iq.cntxt_id + sc->sge.iq_base == rxq->iq.abs_id,
3419 		    ("iq_base mismatch"));
3420 	KASSERT(sc->sge.iq_base == 0 || sc->flags & IS_VF,
3421 	    ("PF with non-zero iq_base"));
3422 
3423 	/*
3424 	 * The freelist is just barely above the starvation threshold right now,
3425 	 * fill it up a bit more.
3426 	 */
3427 	FL_LOCK(&rxq->fl);
3428 	refill_fl(sc, &rxq->fl, 128);
3429 	FL_UNLOCK(&rxq->fl);
3430 
3431 #if defined(INET) || defined(INET6)
3432 	rc = tcp_lro_init_args(&rxq->lro, vi->ifp, lro_entries, lro_mbufs);
3433 	if (rc != 0)
3434 		return (rc);
3435 	MPASS(rxq->lro.ifp == vi->ifp);	/* also indicates LRO init'ed */
3436 
3437 	if (vi->ifp->if_capenable & IFCAP_LRO)
3438 		rxq->iq.flags |= IQ_LRO_ENABLED;
3439 #endif
3440 	if (vi->ifp->if_capenable & IFCAP_HWRXTSTMP)
3441 		rxq->iq.flags |= IQ_RX_TIMESTAMP;
3442 	rxq->ifp = vi->ifp;
3443 
3444 	children = SYSCTL_CHILDREN(oid);
3445 
3446 	snprintf(name, sizeof(name), "%d", idx);
3447 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3448 	    NULL, "rx queue");
3449 	children = SYSCTL_CHILDREN(oid);
3450 
3451 	add_iq_sysctls(&vi->ctx, oid, &rxq->iq);
3452 #if defined(INET) || defined(INET6)
3453 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
3454 	    &rxq->lro.lro_queued, 0, NULL);
3455 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
3456 	    &rxq->lro.lro_flushed, 0, NULL);
3457 #endif
3458 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
3459 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
3460 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_extraction",
3461 	    CTLFLAG_RD, &rxq->vlan_extraction,
3462 	    "# of times hardware extracted 802.1Q tag");
3463 
3464 	add_fl_sysctls(sc, &vi->ctx, oid, &rxq->fl);
3465 
3466 	return (rc);
3467 }
3468 
3469 static int
3470 free_rxq(struct vi_info *vi, struct sge_rxq *rxq)
3471 {
3472 	int rc;
3473 
3474 #if defined(INET) || defined(INET6)
3475 	if (rxq->lro.ifp) {
3476 		tcp_lro_free(&rxq->lro);
3477 		rxq->lro.ifp = NULL;
3478 	}
3479 #endif
3480 
3481 	rc = free_iq_fl(vi, &rxq->iq, &rxq->fl);
3482 	if (rc == 0)
3483 		bzero(rxq, sizeof(*rxq));
3484 
3485 	return (rc);
3486 }
3487 
3488 #ifdef TCP_OFFLOAD
3489 static int
3490 alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq,
3491     int intr_idx, int idx, struct sysctl_oid *oid)
3492 {
3493 	struct port_info *pi = vi->pi;
3494 	int rc;
3495 	struct sysctl_oid_list *children;
3496 	char name[16];
3497 
3498 	rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx, 0);
3499 	if (rc != 0)
3500 		return (rc);
3501 
3502 	children = SYSCTL_CHILDREN(oid);
3503 
3504 	snprintf(name, sizeof(name), "%d", idx);
3505 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3506 	    NULL, "rx queue");
3507 	add_iq_sysctls(&vi->ctx, oid, &ofld_rxq->iq);
3508 	add_fl_sysctls(pi->adapter, &vi->ctx, oid, &ofld_rxq->fl);
3509 
3510 	return (rc);
3511 }
3512 
3513 static int
3514 free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq)
3515 {
3516 	int rc;
3517 
3518 	rc = free_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl);
3519 	if (rc == 0)
3520 		bzero(ofld_rxq, sizeof(*ofld_rxq));
3521 
3522 	return (rc);
3523 }
3524 #endif
3525 
3526 #ifdef DEV_NETMAP
3527 static int
3528 alloc_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int intr_idx,
3529     int idx, struct sysctl_oid *oid)
3530 {
3531 	int rc;
3532 	struct sysctl_oid_list *children;
3533 	struct sysctl_ctx_list *ctx;
3534 	char name[16];
3535 	size_t len;
3536 	struct adapter *sc = vi->pi->adapter;
3537 	struct netmap_adapter *na = NA(vi->ifp);
3538 
3539 	MPASS(na != NULL);
3540 
3541 	len = vi->qsize_rxq * IQ_ESIZE;
3542 	rc = alloc_ring(sc, len, &nm_rxq->iq_desc_tag, &nm_rxq->iq_desc_map,
3543 	    &nm_rxq->iq_ba, (void **)&nm_rxq->iq_desc);
3544 	if (rc != 0)
3545 		return (rc);
3546 
3547 	len = na->num_rx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3548 	rc = alloc_ring(sc, len, &nm_rxq->fl_desc_tag, &nm_rxq->fl_desc_map,
3549 	    &nm_rxq->fl_ba, (void **)&nm_rxq->fl_desc);
3550 	if (rc != 0)
3551 		return (rc);
3552 
3553 	nm_rxq->vi = vi;
3554 	nm_rxq->nid = idx;
3555 	nm_rxq->iq_cidx = 0;
3556 	nm_rxq->iq_sidx = vi->qsize_rxq - sc->params.sge.spg_len / IQ_ESIZE;
3557 	nm_rxq->iq_gen = F_RSPD_GEN;
3558 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
3559 	nm_rxq->fl_sidx = na->num_rx_desc;
3560 	nm_rxq->intr_idx = intr_idx;
3561 	nm_rxq->iq_cntxt_id = INVALID_NM_RXQ_CNTXT_ID;
3562 
3563 	ctx = &vi->ctx;
3564 	children = SYSCTL_CHILDREN(oid);
3565 
3566 	snprintf(name, sizeof(name), "%d", idx);
3567 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL,
3568 	    "rx queue");
3569 	children = SYSCTL_CHILDREN(oid);
3570 
3571 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3572 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_abs_id, 0, sysctl_uint16,
3573 	    "I", "absolute id of the queue");
3574 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3575 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cntxt_id, 0, sysctl_uint16,
3576 	    "I", "SGE context id of the queue");
3577 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3578 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cidx, 0, sysctl_uint16, "I",
3579 	    "consumer index");
3580 
3581 	children = SYSCTL_CHILDREN(oid);
3582 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3583 	    "freelist");
3584 	children = SYSCTL_CHILDREN(oid);
3585 
3586 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3587 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->fl_cntxt_id, 0, sysctl_uint16,
3588 	    "I", "SGE context id of the freelist");
3589 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
3590 	    &nm_rxq->fl_cidx, 0, "consumer index");
3591 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
3592 	    &nm_rxq->fl_pidx, 0, "producer index");
3593 
3594 	return (rc);
3595 }
3596 
3597 
3598 static int
3599 free_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
3600 {
3601 	struct adapter *sc = vi->pi->adapter;
3602 
3603 	if (vi->flags & VI_INIT_DONE)
3604 		MPASS(nm_rxq->iq_cntxt_id == INVALID_NM_RXQ_CNTXT_ID);
3605 	else
3606 		MPASS(nm_rxq->iq_cntxt_id == 0);
3607 
3608 	free_ring(sc, nm_rxq->iq_desc_tag, nm_rxq->iq_desc_map, nm_rxq->iq_ba,
3609 	    nm_rxq->iq_desc);
3610 	free_ring(sc, nm_rxq->fl_desc_tag, nm_rxq->fl_desc_map, nm_rxq->fl_ba,
3611 	    nm_rxq->fl_desc);
3612 
3613 	return (0);
3614 }
3615 
3616 static int
3617 alloc_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq, int iqidx, int idx,
3618     struct sysctl_oid *oid)
3619 {
3620 	int rc;
3621 	size_t len;
3622 	struct port_info *pi = vi->pi;
3623 	struct adapter *sc = pi->adapter;
3624 	struct netmap_adapter *na = NA(vi->ifp);
3625 	char name[16];
3626 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3627 
3628 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3629 	rc = alloc_ring(sc, len, &nm_txq->desc_tag, &nm_txq->desc_map,
3630 	    &nm_txq->ba, (void **)&nm_txq->desc);
3631 	if (rc)
3632 		return (rc);
3633 
3634 	nm_txq->pidx = nm_txq->cidx = 0;
3635 	nm_txq->sidx = na->num_tx_desc;
3636 	nm_txq->nid = idx;
3637 	nm_txq->iqidx = iqidx;
3638 	nm_txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3639 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(G_FW_VIID_PFN(vi->viid)) |
3640 	    V_TXPKT_VF(G_FW_VIID_VIN(vi->viid)) |
3641 	    V_TXPKT_VF_VLD(G_FW_VIID_VIVLD(vi->viid)));
3642 	nm_txq->cntxt_id = INVALID_NM_TXQ_CNTXT_ID;
3643 
3644 	snprintf(name, sizeof(name), "%d", idx);
3645 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3646 	    NULL, "netmap tx queue");
3647 	children = SYSCTL_CHILDREN(oid);
3648 
3649 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3650 	    &nm_txq->cntxt_id, 0, "SGE context id of the queue");
3651 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3652 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->cidx, 0, sysctl_uint16, "I",
3653 	    "consumer index");
3654 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3655 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->pidx, 0, sysctl_uint16, "I",
3656 	    "producer index");
3657 
3658 	return (rc);
3659 }
3660 
3661 static int
3662 free_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
3663 {
3664 	struct adapter *sc = vi->pi->adapter;
3665 
3666 	if (vi->flags & VI_INIT_DONE)
3667 		MPASS(nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID);
3668 	else
3669 		MPASS(nm_txq->cntxt_id == 0);
3670 
3671 	free_ring(sc, nm_txq->desc_tag, nm_txq->desc_map, nm_txq->ba,
3672 	    nm_txq->desc);
3673 
3674 	return (0);
3675 }
3676 #endif
3677 
3678 /*
3679  * Returns a reasonable automatic cidx flush threshold for a given queue size.
3680  */
3681 static u_int
3682 qsize_to_fthresh(int qsize)
3683 {
3684 	u_int fthresh;
3685 
3686 	while (!powerof2(qsize))
3687 		qsize++;
3688 	fthresh = ilog2(qsize);
3689 	if (fthresh > X_CIDXFLUSHTHRESH_128)
3690 		fthresh = X_CIDXFLUSHTHRESH_128;
3691 
3692 	return (fthresh);
3693 }
3694 
3695 static int
3696 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
3697 {
3698 	int rc, cntxt_id;
3699 	struct fw_eq_ctrl_cmd c;
3700 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3701 
3702 	bzero(&c, sizeof(c));
3703 
3704 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
3705 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
3706 	    V_FW_EQ_CTRL_CMD_VFN(0));
3707 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
3708 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
3709 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid));
3710 	c.physeqid_pkd = htobe32(0);
3711 	c.fetchszm_to_iqid =
3712 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
3713 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
3714 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
3715 	c.dcaen_to_eqsize =
3716 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3717 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3718 		V_FW_EQ_CTRL_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
3719 		V_FW_EQ_CTRL_CMD_EQSIZE(qsize));
3720 	c.eqaddr = htobe64(eq->ba);
3721 
3722 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3723 	if (rc != 0) {
3724 		device_printf(sc->dev,
3725 		    "failed to create control queue %d: %d\n", eq->tx_chan, rc);
3726 		return (rc);
3727 	}
3728 	eq->flags |= EQ_ALLOCATED;
3729 
3730 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
3731 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3732 	if (cntxt_id >= sc->sge.neq)
3733 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3734 		cntxt_id, sc->sge.neq - 1);
3735 	sc->sge.eqmap[cntxt_id] = eq;
3736 
3737 	return (rc);
3738 }
3739 
3740 static int
3741 eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3742 {
3743 	int rc, cntxt_id;
3744 	struct fw_eq_eth_cmd c;
3745 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3746 
3747 	bzero(&c, sizeof(c));
3748 
3749 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
3750 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
3751 	    V_FW_EQ_ETH_CMD_VFN(0));
3752 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
3753 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
3754 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
3755 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
3756 	c.fetchszm_to_iqid =
3757 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3758 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
3759 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
3760 	c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3761 	    V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3762 	    V_FW_EQ_ETH_CMD_EQSIZE(qsize));
3763 	c.eqaddr = htobe64(eq->ba);
3764 
3765 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3766 	if (rc != 0) {
3767 		device_printf(vi->dev,
3768 		    "failed to create Ethernet egress queue: %d\n", rc);
3769 		return (rc);
3770 	}
3771 	eq->flags |= EQ_ALLOCATED;
3772 
3773 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
3774 	eq->abs_id = G_FW_EQ_ETH_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
3775 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3776 	if (cntxt_id >= sc->sge.neq)
3777 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3778 		cntxt_id, sc->sge.neq - 1);
3779 	sc->sge.eqmap[cntxt_id] = eq;
3780 
3781 	return (rc);
3782 }
3783 
3784 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3785 static int
3786 ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3787 {
3788 	int rc, cntxt_id;
3789 	struct fw_eq_ofld_cmd c;
3790 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3791 
3792 	bzero(&c, sizeof(c));
3793 
3794 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
3795 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
3796 	    V_FW_EQ_OFLD_CMD_VFN(0));
3797 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
3798 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
3799 	c.fetchszm_to_iqid =
3800 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
3801 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
3802 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
3803 	c.dcaen_to_eqsize =
3804 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3805 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3806 		V_FW_EQ_OFLD_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
3807 		V_FW_EQ_OFLD_CMD_EQSIZE(qsize));
3808 	c.eqaddr = htobe64(eq->ba);
3809 
3810 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3811 	if (rc != 0) {
3812 		device_printf(vi->dev,
3813 		    "failed to create egress queue for TCP offload: %d\n", rc);
3814 		return (rc);
3815 	}
3816 	eq->flags |= EQ_ALLOCATED;
3817 
3818 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
3819 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3820 	if (cntxt_id >= sc->sge.neq)
3821 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3822 		cntxt_id, sc->sge.neq - 1);
3823 	sc->sge.eqmap[cntxt_id] = eq;
3824 
3825 	return (rc);
3826 }
3827 #endif
3828 
3829 static int
3830 alloc_eq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3831 {
3832 	int rc, qsize;
3833 	size_t len;
3834 
3835 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
3836 
3837 	qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3838 	len = qsize * EQ_ESIZE;
3839 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
3840 	    &eq->ba, (void **)&eq->desc);
3841 	if (rc)
3842 		return (rc);
3843 
3844 	eq->pidx = eq->cidx = eq->dbidx = 0;
3845 	/* Note that equeqidx is not used with sge_wrq (OFLD/CTRL) queues. */
3846 	eq->equeqidx = 0;
3847 	eq->doorbells = sc->doorbells;
3848 
3849 	switch (eq->flags & EQ_TYPEMASK) {
3850 	case EQ_CTRL:
3851 		rc = ctrl_eq_alloc(sc, eq);
3852 		break;
3853 
3854 	case EQ_ETH:
3855 		rc = eth_eq_alloc(sc, vi, eq);
3856 		break;
3857 
3858 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3859 	case EQ_OFLD:
3860 		rc = ofld_eq_alloc(sc, vi, eq);
3861 		break;
3862 #endif
3863 
3864 	default:
3865 		panic("%s: invalid eq type %d.", __func__,
3866 		    eq->flags & EQ_TYPEMASK);
3867 	}
3868 	if (rc != 0) {
3869 		device_printf(sc->dev,
3870 		    "failed to allocate egress queue(%d): %d\n",
3871 		    eq->flags & EQ_TYPEMASK, rc);
3872 	}
3873 
3874 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
3875 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
3876 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
3877 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3878 		uint32_t mask = (1 << s_qpp) - 1;
3879 		volatile uint8_t *udb;
3880 
3881 		udb = sc->udbs_base + UDBS_DB_OFFSET;
3882 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
3883 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
3884 		if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
3885 	    		clrbit(&eq->doorbells, DOORBELL_WCWR);
3886 		else {
3887 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
3888 			eq->udb_qid = 0;
3889 		}
3890 		eq->udb = (volatile void *)udb;
3891 	}
3892 
3893 	return (rc);
3894 }
3895 
3896 static int
3897 free_eq(struct adapter *sc, struct sge_eq *eq)
3898 {
3899 	int rc;
3900 
3901 	if (eq->flags & EQ_ALLOCATED) {
3902 		switch (eq->flags & EQ_TYPEMASK) {
3903 		case EQ_CTRL:
3904 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
3905 			    eq->cntxt_id);
3906 			break;
3907 
3908 		case EQ_ETH:
3909 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
3910 			    eq->cntxt_id);
3911 			break;
3912 
3913 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3914 		case EQ_OFLD:
3915 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
3916 			    eq->cntxt_id);
3917 			break;
3918 #endif
3919 
3920 		default:
3921 			panic("%s: invalid eq type %d.", __func__,
3922 			    eq->flags & EQ_TYPEMASK);
3923 		}
3924 		if (rc != 0) {
3925 			device_printf(sc->dev,
3926 			    "failed to free egress queue (%d): %d\n",
3927 			    eq->flags & EQ_TYPEMASK, rc);
3928 			return (rc);
3929 		}
3930 		eq->flags &= ~EQ_ALLOCATED;
3931 	}
3932 
3933 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
3934 
3935 	if (mtx_initialized(&eq->eq_lock))
3936 		mtx_destroy(&eq->eq_lock);
3937 
3938 	bzero(eq, sizeof(*eq));
3939 	return (0);
3940 }
3941 
3942 static int
3943 alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq,
3944     struct sysctl_oid *oid)
3945 {
3946 	int rc;
3947 	struct sysctl_ctx_list *ctx = vi ? &vi->ctx : &sc->ctx;
3948 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3949 
3950 	rc = alloc_eq(sc, vi, &wrq->eq);
3951 	if (rc)
3952 		return (rc);
3953 
3954 	wrq->adapter = sc;
3955 	TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq);
3956 	TAILQ_INIT(&wrq->incomplete_wrs);
3957 	STAILQ_INIT(&wrq->wr_list);
3958 	wrq->nwr_pending = 0;
3959 	wrq->ndesc_needed = 0;
3960 
3961 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3962 	    &wrq->eq.ba, "bus address of descriptor ring");
3963 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3964 	    wrq->eq.sidx * EQ_ESIZE + sc->params.sge.spg_len,
3965 	    "desc ring size in bytes");
3966 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3967 	    &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
3968 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3969 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I",
3970 	    "consumer index");
3971 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx",
3972 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I",
3973 	    "producer index");
3974 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
3975 	    wrq->eq.sidx, "status page index");
3976 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD,
3977 	    &wrq->tx_wrs_direct, "# of work requests (direct)");
3978 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD,
3979 	    &wrq->tx_wrs_copied, "# of work requests (copied)");
3980 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_sspace", CTLFLAG_RD,
3981 	    &wrq->tx_wrs_ss, "# of work requests (copied from scratch space)");
3982 
3983 	return (rc);
3984 }
3985 
3986 static int
3987 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
3988 {
3989 	int rc;
3990 
3991 	rc = free_eq(sc, &wrq->eq);
3992 	if (rc)
3993 		return (rc);
3994 
3995 	bzero(wrq, sizeof(*wrq));
3996 	return (0);
3997 }
3998 
3999 static int
4000 alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx,
4001     struct sysctl_oid *oid)
4002 {
4003 	int rc;
4004 	struct port_info *pi = vi->pi;
4005 	struct adapter *sc = pi->adapter;
4006 	struct sge_eq *eq = &txq->eq;
4007 	char name[16];
4008 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
4009 
4010 	rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx, can_resume_eth_tx,
4011 	    M_CXGBE, M_WAITOK);
4012 	if (rc != 0) {
4013 		device_printf(sc->dev, "failed to allocate mp_ring: %d\n", rc);
4014 		return (rc);
4015 	}
4016 
4017 	rc = alloc_eq(sc, vi, eq);
4018 	if (rc != 0) {
4019 		mp_ring_free(txq->r);
4020 		txq->r = NULL;
4021 		return (rc);
4022 	}
4023 
4024 	/* Can't fail after this point. */
4025 
4026 	if (idx == 0)
4027 		sc->sge.eq_base = eq->abs_id - eq->cntxt_id;
4028 	else
4029 		KASSERT(eq->cntxt_id + sc->sge.eq_base == eq->abs_id,
4030 		    ("eq_base mismatch"));
4031 	KASSERT(sc->sge.eq_base == 0 || sc->flags & IS_VF,
4032 	    ("PF with non-zero eq_base"));
4033 
4034 	TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq);
4035 	txq->ifp = vi->ifp;
4036 	txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK);
4037 	if (sc->flags & IS_VF)
4038 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
4039 		    V_TXPKT_INTF(pi->tx_chan));
4040 	else
4041 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
4042 		    V_TXPKT_INTF(pi->tx_chan) |
4043 		    V_TXPKT_PF(G_FW_VIID_PFN(vi->viid)) |
4044 		    V_TXPKT_VF(G_FW_VIID_VIN(vi->viid)) |
4045 		    V_TXPKT_VF_VLD(G_FW_VIID_VIVLD(vi->viid)));
4046 	txq->tc_idx = -1;
4047 	txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE,
4048 	    M_ZERO | M_WAITOK);
4049 
4050 	snprintf(name, sizeof(name), "%d", idx);
4051 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
4052 	    NULL, "tx queue");
4053 	children = SYSCTL_CHILDREN(oid);
4054 
4055 	SYSCTL_ADD_UAUTO(&vi->ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
4056 	    &eq->ba, "bus address of descriptor ring");
4057 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
4058 	    eq->sidx * EQ_ESIZE + sc->params.sge.spg_len,
4059 	    "desc ring size in bytes");
4060 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
4061 	    &eq->abs_id, 0, "absolute id of the queue");
4062 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
4063 	    &eq->cntxt_id, 0, "SGE context id of the queue");
4064 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
4065 	    CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I",
4066 	    "consumer index");
4067 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
4068 	    CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I",
4069 	    "producer index");
4070 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
4071 	    eq->sidx, "status page index");
4072 
4073 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "tc",
4074 	    CTLTYPE_INT | CTLFLAG_RW, vi, idx, sysctl_tc, "I",
4075 	    "traffic class (-1 means none)");
4076 
4077 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
4078 	    &txq->txcsum, "# of times hardware assisted with checksum");
4079 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_insertion",
4080 	    CTLFLAG_RD, &txq->vlan_insertion,
4081 	    "# of times hardware inserted 802.1Q tag");
4082 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
4083 	    &txq->tso_wrs, "# of TSO work requests");
4084 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
4085 	    &txq->imm_wrs, "# of work requests with immediate data");
4086 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
4087 	    &txq->sgl_wrs, "# of work requests with direct SGL");
4088 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
4089 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
4090 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_wrs",
4091 	    CTLFLAG_RD, &txq->txpkts0_wrs,
4092 	    "# of txpkts (type 0) work requests");
4093 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_wrs",
4094 	    CTLFLAG_RD, &txq->txpkts1_wrs,
4095 	    "# of txpkts (type 1) work requests");
4096 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_pkts",
4097 	    CTLFLAG_RD, &txq->txpkts0_pkts,
4098 	    "# of frames tx'd using type0 txpkts work requests");
4099 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_pkts",
4100 	    CTLFLAG_RD, &txq->txpkts1_pkts,
4101 	    "# of frames tx'd using type1 txpkts work requests");
4102 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "raw_wrs", CTLFLAG_RD,
4103 	    &txq->raw_wrs, "# of raw work requests (non-packets)");
4104 
4105 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_enqueues",
4106 	    CTLFLAG_RD, &txq->r->enqueues,
4107 	    "# of enqueues to the mp_ring for this queue");
4108 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_drops",
4109 	    CTLFLAG_RD, &txq->r->drops,
4110 	    "# of drops in the mp_ring for this queue");
4111 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_starts",
4112 	    CTLFLAG_RD, &txq->r->starts,
4113 	    "# of normal consumer starts in the mp_ring for this queue");
4114 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_stalls",
4115 	    CTLFLAG_RD, &txq->r->stalls,
4116 	    "# of consumer stalls in the mp_ring for this queue");
4117 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_restarts",
4118 	    CTLFLAG_RD, &txq->r->restarts,
4119 	    "# of consumer restarts in the mp_ring for this queue");
4120 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_abdications",
4121 	    CTLFLAG_RD, &txq->r->abdications,
4122 	    "# of consumer abdications in the mp_ring for this queue");
4123 
4124 	return (0);
4125 }
4126 
4127 static int
4128 free_txq(struct vi_info *vi, struct sge_txq *txq)
4129 {
4130 	int rc;
4131 	struct adapter *sc = vi->pi->adapter;
4132 	struct sge_eq *eq = &txq->eq;
4133 
4134 	rc = free_eq(sc, eq);
4135 	if (rc)
4136 		return (rc);
4137 
4138 	sglist_free(txq->gl);
4139 	free(txq->sdesc, M_CXGBE);
4140 	mp_ring_free(txq->r);
4141 
4142 	bzero(txq, sizeof(*txq));
4143 	return (0);
4144 }
4145 
4146 static void
4147 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
4148 {
4149 	bus_addr_t *ba = arg;
4150 
4151 	KASSERT(nseg == 1,
4152 	    ("%s meant for single segment mappings only.", __func__));
4153 
4154 	*ba = error ? 0 : segs->ds_addr;
4155 }
4156 
4157 static inline void
4158 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
4159 {
4160 	uint32_t n, v;
4161 
4162 	n = IDXDIFF(fl->pidx / 8, fl->dbidx, fl->sidx);
4163 	MPASS(n > 0);
4164 
4165 	wmb();
4166 	v = fl->dbval | V_PIDX(n);
4167 	if (fl->udb)
4168 		*fl->udb = htole32(v);
4169 	else
4170 		t4_write_reg(sc, sc->sge_kdoorbell_reg, v);
4171 	IDXINCR(fl->dbidx, n, fl->sidx);
4172 }
4173 
4174 /*
4175  * Fills up the freelist by allocating up to 'n' buffers.  Buffers that are
4176  * recycled do not count towards this allocation budget.
4177  *
4178  * Returns non-zero to indicate that this freelist should be added to the list
4179  * of starving freelists.
4180  */
4181 static int
4182 refill_fl(struct adapter *sc, struct sge_fl *fl, int n)
4183 {
4184 	__be64 *d;
4185 	struct fl_sdesc *sd;
4186 	uintptr_t pa;
4187 	caddr_t cl;
4188 	struct cluster_layout *cll;
4189 	struct sw_zone_info *swz;
4190 	struct cluster_metadata *clm;
4191 	uint16_t max_pidx;
4192 	uint16_t hw_cidx = fl->hw_cidx;		/* stable snapshot */
4193 
4194 	FL_LOCK_ASSERT_OWNED(fl);
4195 
4196 	/*
4197 	 * We always stop at the beginning of the hardware descriptor that's just
4198 	 * before the one with the hw cidx.  This is to avoid hw pidx = hw cidx,
4199 	 * which would mean an empty freelist to the chip.
4200 	 */
4201 	max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1;
4202 	if (fl->pidx == max_pidx * 8)
4203 		return (0);
4204 
4205 	d = &fl->desc[fl->pidx];
4206 	sd = &fl->sdesc[fl->pidx];
4207 	cll = &fl->cll_def;	/* default layout */
4208 	swz = &sc->sge.sw_zone_info[cll->zidx];
4209 
4210 	while (n > 0) {
4211 
4212 		if (sd->cl != NULL) {
4213 
4214 			if (sd->nmbuf == 0) {
4215 				/*
4216 				 * Fast recycle without involving any atomics on
4217 				 * the cluster's metadata (if the cluster has
4218 				 * metadata).  This happens when all frames
4219 				 * received in the cluster were small enough to
4220 				 * fit within a single mbuf each.
4221 				 */
4222 				fl->cl_fast_recycled++;
4223 #ifdef INVARIANTS
4224 				clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
4225 				if (clm != NULL)
4226 					MPASS(clm->refcount == 1);
4227 #endif
4228 				goto recycled_fast;
4229 			}
4230 
4231 			/*
4232 			 * Cluster is guaranteed to have metadata.  Clusters
4233 			 * without metadata always take the fast recycle path
4234 			 * when they're recycled.
4235 			 */
4236 			clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
4237 			MPASS(clm != NULL);
4238 
4239 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4240 				fl->cl_recycled++;
4241 				counter_u64_add(extfree_rels, 1);
4242 				goto recycled;
4243 			}
4244 			sd->cl = NULL;	/* gave up my reference */
4245 		}
4246 		MPASS(sd->cl == NULL);
4247 alloc:
4248 		cl = uma_zalloc(swz->zone, M_NOWAIT);
4249 		if (__predict_false(cl == NULL)) {
4250 			if (cll == &fl->cll_alt || fl->cll_alt.zidx == -1 ||
4251 			    fl->cll_def.zidx == fl->cll_alt.zidx)
4252 				break;
4253 
4254 			/* fall back to the safe zone */
4255 			cll = &fl->cll_alt;
4256 			swz = &sc->sge.sw_zone_info[cll->zidx];
4257 			goto alloc;
4258 		}
4259 		fl->cl_allocated++;
4260 		n--;
4261 
4262 		pa = pmap_kextract((vm_offset_t)cl);
4263 		pa += cll->region1;
4264 		sd->cl = cl;
4265 		sd->cll = *cll;
4266 		*d = htobe64(pa | cll->hwidx);
4267 		clm = cl_metadata(sc, fl, cll, cl);
4268 		if (clm != NULL) {
4269 recycled:
4270 #ifdef INVARIANTS
4271 			clm->sd = sd;
4272 #endif
4273 			clm->refcount = 1;
4274 		}
4275 		sd->nmbuf = 0;
4276 recycled_fast:
4277 		d++;
4278 		sd++;
4279 		if (__predict_false(++fl->pidx % 8 == 0)) {
4280 			uint16_t pidx = fl->pidx / 8;
4281 
4282 			if (__predict_false(pidx == fl->sidx)) {
4283 				fl->pidx = 0;
4284 				pidx = 0;
4285 				sd = fl->sdesc;
4286 				d = fl->desc;
4287 			}
4288 			if (pidx == max_pidx)
4289 				break;
4290 
4291 			if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4)
4292 				ring_fl_db(sc, fl);
4293 		}
4294 	}
4295 
4296 	if (fl->pidx / 8 != fl->dbidx)
4297 		ring_fl_db(sc, fl);
4298 
4299 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
4300 }
4301 
4302 /*
4303  * Attempt to refill all starving freelists.
4304  */
4305 static void
4306 refill_sfl(void *arg)
4307 {
4308 	struct adapter *sc = arg;
4309 	struct sge_fl *fl, *fl_temp;
4310 
4311 	mtx_assert(&sc->sfl_lock, MA_OWNED);
4312 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
4313 		FL_LOCK(fl);
4314 		refill_fl(sc, fl, 64);
4315 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
4316 			TAILQ_REMOVE(&sc->sfl, fl, link);
4317 			fl->flags &= ~FL_STARVING;
4318 		}
4319 		FL_UNLOCK(fl);
4320 	}
4321 
4322 	if (!TAILQ_EMPTY(&sc->sfl))
4323 		callout_schedule(&sc->sfl_callout, hz / 5);
4324 }
4325 
4326 static int
4327 alloc_fl_sdesc(struct sge_fl *fl)
4328 {
4329 
4330 	fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc), M_CXGBE,
4331 	    M_ZERO | M_WAITOK);
4332 
4333 	return (0);
4334 }
4335 
4336 static void
4337 free_fl_sdesc(struct adapter *sc, struct sge_fl *fl)
4338 {
4339 	struct fl_sdesc *sd;
4340 	struct cluster_metadata *clm;
4341 	struct cluster_layout *cll;
4342 	int i;
4343 
4344 	sd = fl->sdesc;
4345 	for (i = 0; i < fl->sidx * 8; i++, sd++) {
4346 		if (sd->cl == NULL)
4347 			continue;
4348 
4349 		cll = &sd->cll;
4350 		clm = cl_metadata(sc, fl, cll, sd->cl);
4351 		if (sd->nmbuf == 0)
4352 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4353 		else if (clm && atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4354 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4355 			counter_u64_add(extfree_rels, 1);
4356 		}
4357 		sd->cl = NULL;
4358 	}
4359 
4360 	free(fl->sdesc, M_CXGBE);
4361 	fl->sdesc = NULL;
4362 }
4363 
4364 static inline void
4365 get_pkt_gl(struct mbuf *m, struct sglist *gl)
4366 {
4367 	int rc;
4368 
4369 	M_ASSERTPKTHDR(m);
4370 
4371 	sglist_reset(gl);
4372 	rc = sglist_append_mbuf(gl, m);
4373 	if (__predict_false(rc != 0)) {
4374 		panic("%s: mbuf %p (%d segs) was vetted earlier but now fails "
4375 		    "with %d.", __func__, m, mbuf_nsegs(m), rc);
4376 	}
4377 
4378 	KASSERT(gl->sg_nseg == mbuf_nsegs(m),
4379 	    ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m,
4380 	    mbuf_nsegs(m), gl->sg_nseg));
4381 	KASSERT(gl->sg_nseg > 0 &&
4382 	    gl->sg_nseg <= (needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS),
4383 	    ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__,
4384 		gl->sg_nseg, needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS));
4385 }
4386 
4387 /*
4388  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
4389  */
4390 static inline u_int
4391 txpkt_len16(u_int nsegs, u_int tso)
4392 {
4393 	u_int n;
4394 
4395 	MPASS(nsegs > 0);
4396 
4397 	nsegs--; /* first segment is part of ulptx_sgl */
4398 	n = sizeof(struct fw_eth_tx_pkt_wr) + sizeof(struct cpl_tx_pkt_core) +
4399 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4400 	if (tso)
4401 		n += sizeof(struct cpl_tx_pkt_lso_core);
4402 
4403 	return (howmany(n, 16));
4404 }
4405 
4406 /*
4407  * len16 for a txpkt_vm WR with a GL.  Includes the firmware work
4408  * request header.
4409  */
4410 static inline u_int
4411 txpkt_vm_len16(u_int nsegs, u_int tso)
4412 {
4413 	u_int n;
4414 
4415 	MPASS(nsegs > 0);
4416 
4417 	nsegs--; /* first segment is part of ulptx_sgl */
4418 	n = sizeof(struct fw_eth_tx_pkt_vm_wr) +
4419 	    sizeof(struct cpl_tx_pkt_core) +
4420 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4421 	if (tso)
4422 		n += sizeof(struct cpl_tx_pkt_lso_core);
4423 
4424 	return (howmany(n, 16));
4425 }
4426 
4427 /*
4428  * len16 for a txpkts type 0 WR with a GL.  Does not include the firmware work
4429  * request header.
4430  */
4431 static inline u_int
4432 txpkts0_len16(u_int nsegs)
4433 {
4434 	u_int n;
4435 
4436 	MPASS(nsegs > 0);
4437 
4438 	nsegs--; /* first segment is part of ulptx_sgl */
4439 	n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) +
4440 	    sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) +
4441 	    8 * ((3 * nsegs) / 2 + (nsegs & 1));
4442 
4443 	return (howmany(n, 16));
4444 }
4445 
4446 /*
4447  * len16 for a txpkts type 1 WR with a GL.  Does not include the firmware work
4448  * request header.
4449  */
4450 static inline u_int
4451 txpkts1_len16(void)
4452 {
4453 	u_int n;
4454 
4455 	n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl);
4456 
4457 	return (howmany(n, 16));
4458 }
4459 
4460 static inline u_int
4461 imm_payload(u_int ndesc)
4462 {
4463 	u_int n;
4464 
4465 	n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) -
4466 	    sizeof(struct cpl_tx_pkt_core);
4467 
4468 	return (n);
4469 }
4470 
4471 /*
4472  * Write a VM txpkt WR for this packet to the hardware descriptors, update the
4473  * software descriptor, and advance the pidx.  It is guaranteed that enough
4474  * descriptors are available.
4475  *
4476  * The return value is the # of hardware descriptors used.
4477  */
4478 static u_int
4479 write_txpkt_vm_wr(struct adapter *sc, struct sge_txq *txq,
4480     struct fw_eth_tx_pkt_vm_wr *wr, struct mbuf *m0, u_int available)
4481 {
4482 	struct sge_eq *eq = &txq->eq;
4483 	struct tx_sdesc *txsd;
4484 	struct cpl_tx_pkt_core *cpl;
4485 	uint32_t ctrl;	/* used in many unrelated places */
4486 	uint64_t ctrl1;
4487 	int csum_type, len16, ndesc, pktlen, nsegs;
4488 	caddr_t dst;
4489 
4490 	TXQ_LOCK_ASSERT_OWNED(txq);
4491 	M_ASSERTPKTHDR(m0);
4492 	MPASS(available > 0 && available < eq->sidx);
4493 
4494 	len16 = mbuf_len16(m0);
4495 	nsegs = mbuf_nsegs(m0);
4496 	pktlen = m0->m_pkthdr.len;
4497 	ctrl = sizeof(struct cpl_tx_pkt_core);
4498 	if (needs_tso(m0))
4499 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4500 	ndesc = howmany(len16, EQ_ESIZE / 16);
4501 	MPASS(ndesc <= available);
4502 
4503 	/* Firmware work request header */
4504 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4505 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_VM_WR) |
4506 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4507 
4508 	ctrl = V_FW_WR_LEN16(len16);
4509 	wr->equiq_to_len16 = htobe32(ctrl);
4510 	wr->r3[0] = 0;
4511 	wr->r3[1] = 0;
4512 
4513 	/*
4514 	 * Copy over ethmacdst, ethmacsrc, ethtype, and vlantci.
4515 	 * vlantci is ignored unless the ethtype is 0x8100, so it's
4516 	 * simpler to always copy it rather than making it
4517 	 * conditional.  Also, it seems that we do not have to set
4518 	 * vlantci or fake the ethtype when doing VLAN tag insertion.
4519 	 */
4520 	m_copydata(m0, 0, sizeof(struct ether_header) + 2, wr->ethmacdst);
4521 
4522 	csum_type = -1;
4523 	if (needs_tso(m0)) {
4524 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4525 
4526 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4527 		    m0->m_pkthdr.l4hlen > 0,
4528 		    ("%s: mbuf %p needs TSO but missing header lengths",
4529 			__func__, m0));
4530 
4531 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4532 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4533 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4534 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4535 			ctrl |= V_LSO_ETHHDR_LEN(1);
4536 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4537 			ctrl |= F_LSO_IPV6;
4538 
4539 		lso->lso_ctrl = htobe32(ctrl);
4540 		lso->ipid_ofst = htobe16(0);
4541 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4542 		lso->seqno_offset = htobe32(0);
4543 		lso->len = htobe32(pktlen);
4544 
4545 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4546 			csum_type = TX_CSUM_TCPIP6;
4547 		else
4548 			csum_type = TX_CSUM_TCPIP;
4549 
4550 		cpl = (void *)(lso + 1);
4551 
4552 		txq->tso_wrs++;
4553 	} else {
4554 		if (m0->m_pkthdr.csum_flags & CSUM_IP_TCP)
4555 			csum_type = TX_CSUM_TCPIP;
4556 		else if (m0->m_pkthdr.csum_flags & CSUM_IP_UDP)
4557 			csum_type = TX_CSUM_UDPIP;
4558 		else if (m0->m_pkthdr.csum_flags & CSUM_IP6_TCP)
4559 			csum_type = TX_CSUM_TCPIP6;
4560 		else if (m0->m_pkthdr.csum_flags & CSUM_IP6_UDP)
4561 			csum_type = TX_CSUM_UDPIP6;
4562 #if defined(INET)
4563 		else if (m0->m_pkthdr.csum_flags & CSUM_IP) {
4564 			/*
4565 			 * XXX: The firmware appears to stomp on the
4566 			 * fragment/flags field of the IP header when
4567 			 * using TX_CSUM_IP.  Fall back to doing
4568 			 * software checksums.
4569 			 */
4570 			u_short *sump;
4571 			struct mbuf *m;
4572 			int offset;
4573 
4574 			m = m0;
4575 			offset = 0;
4576 			sump = m_advance(&m, &offset, m0->m_pkthdr.l2hlen +
4577 			    offsetof(struct ip, ip_sum));
4578 			*sump = in_cksum_skip(m0, m0->m_pkthdr.l2hlen +
4579 			    m0->m_pkthdr.l3hlen, m0->m_pkthdr.l2hlen);
4580 			m0->m_pkthdr.csum_flags &= ~CSUM_IP;
4581 		}
4582 #endif
4583 
4584 		cpl = (void *)(wr + 1);
4585 	}
4586 
4587 	/* Checksum offload */
4588 	ctrl1 = 0;
4589 	if (needs_l3_csum(m0) == 0)
4590 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4591 	if (csum_type >= 0) {
4592 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0,
4593 	    ("%s: mbuf %p needs checksum offload but missing header lengths",
4594 			__func__, m0));
4595 
4596 		if (chip_id(sc) <= CHELSIO_T5) {
4597 			ctrl1 |= V_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4598 			    ETHER_HDR_LEN);
4599 		} else {
4600 			ctrl1 |= V_T6_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4601 			    ETHER_HDR_LEN);
4602 		}
4603 		ctrl1 |= V_TXPKT_IPHDR_LEN(m0->m_pkthdr.l3hlen);
4604 		ctrl1 |= V_TXPKT_CSUM_TYPE(csum_type);
4605 	} else
4606 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4607 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4608 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4609 		txq->txcsum++;	/* some hardware assistance provided */
4610 
4611 	/* VLAN tag insertion */
4612 	if (needs_vlan_insertion(m0)) {
4613 		ctrl1 |= F_TXPKT_VLAN_VLD |
4614 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4615 		txq->vlan_insertion++;
4616 	}
4617 
4618 	/* CPL header */
4619 	cpl->ctrl0 = txq->cpl_ctrl0;
4620 	cpl->pack = 0;
4621 	cpl->len = htobe16(pktlen);
4622 	cpl->ctrl1 = htobe64(ctrl1);
4623 
4624 	/* SGL */
4625 	dst = (void *)(cpl + 1);
4626 
4627 	/*
4628 	 * A packet using TSO will use up an entire descriptor for the
4629 	 * firmware work request header, LSO CPL, and TX_PKT_XT CPL.
4630 	 * If this descriptor is the last descriptor in the ring, wrap
4631 	 * around to the front of the ring explicitly for the start of
4632 	 * the sgl.
4633 	 */
4634 	if (dst == (void *)&eq->desc[eq->sidx]) {
4635 		dst = (void *)&eq->desc[0];
4636 		write_gl_to_txd(txq, m0, &dst, 0);
4637 	} else
4638 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4639 	txq->sgl_wrs++;
4640 
4641 	txq->txpkt_wrs++;
4642 
4643 	txsd = &txq->sdesc[eq->pidx];
4644 	txsd->m = m0;
4645 	txsd->desc_used = ndesc;
4646 
4647 	return (ndesc);
4648 }
4649 
4650 /*
4651  * Write a raw WR to the hardware descriptors, update the software
4652  * descriptor, and advance the pidx.  It is guaranteed that enough
4653  * descriptors are available.
4654  *
4655  * The return value is the # of hardware descriptors used.
4656  */
4657 static u_int
4658 write_raw_wr(struct sge_txq *txq, void *wr, struct mbuf *m0, u_int available)
4659 {
4660 	struct sge_eq *eq = &txq->eq;
4661 	struct tx_sdesc *txsd;
4662 	struct mbuf *m;
4663 	caddr_t dst;
4664 	int len16, ndesc;
4665 
4666 	len16 = mbuf_len16(m0);
4667 	ndesc = howmany(len16, EQ_ESIZE / 16);
4668 	MPASS(ndesc <= available);
4669 
4670 	dst = wr;
4671 	for (m = m0; m != NULL; m = m->m_next)
4672 		copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4673 
4674 	txq->raw_wrs++;
4675 
4676 	txsd = &txq->sdesc[eq->pidx];
4677 	txsd->m = m0;
4678 	txsd->desc_used = ndesc;
4679 
4680 	return (ndesc);
4681 }
4682 
4683 /*
4684  * Write a txpkt WR for this packet to the hardware descriptors, update the
4685  * software descriptor, and advance the pidx.  It is guaranteed that enough
4686  * descriptors are available.
4687  *
4688  * The return value is the # of hardware descriptors used.
4689  */
4690 static u_int
4691 write_txpkt_wr(struct sge_txq *txq, struct fw_eth_tx_pkt_wr *wr,
4692     struct mbuf *m0, u_int available)
4693 {
4694 	struct sge_eq *eq = &txq->eq;
4695 	struct tx_sdesc *txsd;
4696 	struct cpl_tx_pkt_core *cpl;
4697 	uint32_t ctrl;	/* used in many unrelated places */
4698 	uint64_t ctrl1;
4699 	int len16, ndesc, pktlen, nsegs;
4700 	caddr_t dst;
4701 
4702 	TXQ_LOCK_ASSERT_OWNED(txq);
4703 	M_ASSERTPKTHDR(m0);
4704 	MPASS(available > 0 && available < eq->sidx);
4705 
4706 	len16 = mbuf_len16(m0);
4707 	nsegs = mbuf_nsegs(m0);
4708 	pktlen = m0->m_pkthdr.len;
4709 	ctrl = sizeof(struct cpl_tx_pkt_core);
4710 	if (needs_tso(m0))
4711 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4712 	else if (pktlen <= imm_payload(2) && available >= 2) {
4713 		/* Immediate data.  Recalculate len16 and set nsegs to 0. */
4714 		ctrl += pktlen;
4715 		len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) +
4716 		    sizeof(struct cpl_tx_pkt_core) + pktlen, 16);
4717 		nsegs = 0;
4718 	}
4719 	ndesc = howmany(len16, EQ_ESIZE / 16);
4720 	MPASS(ndesc <= available);
4721 
4722 	/* Firmware work request header */
4723 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4724 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
4725 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4726 
4727 	ctrl = V_FW_WR_LEN16(len16);
4728 	wr->equiq_to_len16 = htobe32(ctrl);
4729 	wr->r3 = 0;
4730 
4731 	if (needs_tso(m0)) {
4732 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4733 
4734 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4735 		    m0->m_pkthdr.l4hlen > 0,
4736 		    ("%s: mbuf %p needs TSO but missing header lengths",
4737 			__func__, m0));
4738 
4739 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4740 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4741 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4742 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4743 			ctrl |= V_LSO_ETHHDR_LEN(1);
4744 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4745 			ctrl |= F_LSO_IPV6;
4746 
4747 		lso->lso_ctrl = htobe32(ctrl);
4748 		lso->ipid_ofst = htobe16(0);
4749 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4750 		lso->seqno_offset = htobe32(0);
4751 		lso->len = htobe32(pktlen);
4752 
4753 		cpl = (void *)(lso + 1);
4754 
4755 		txq->tso_wrs++;
4756 	} else
4757 		cpl = (void *)(wr + 1);
4758 
4759 	/* Checksum offload */
4760 	ctrl1 = 0;
4761 	if (needs_l3_csum(m0) == 0)
4762 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4763 	if (needs_l4_csum(m0) == 0)
4764 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4765 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4766 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4767 		txq->txcsum++;	/* some hardware assistance provided */
4768 
4769 	/* VLAN tag insertion */
4770 	if (needs_vlan_insertion(m0)) {
4771 		ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4772 		txq->vlan_insertion++;
4773 	}
4774 
4775 	/* CPL header */
4776 	cpl->ctrl0 = txq->cpl_ctrl0;
4777 	cpl->pack = 0;
4778 	cpl->len = htobe16(pktlen);
4779 	cpl->ctrl1 = htobe64(ctrl1);
4780 
4781 	/* SGL */
4782 	dst = (void *)(cpl + 1);
4783 	if (nsegs > 0) {
4784 
4785 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4786 		txq->sgl_wrs++;
4787 	} else {
4788 		struct mbuf *m;
4789 
4790 		for (m = m0; m != NULL; m = m->m_next) {
4791 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4792 #ifdef INVARIANTS
4793 			pktlen -= m->m_len;
4794 #endif
4795 		}
4796 #ifdef INVARIANTS
4797 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
4798 #endif
4799 		txq->imm_wrs++;
4800 	}
4801 
4802 	txq->txpkt_wrs++;
4803 
4804 	txsd = &txq->sdesc[eq->pidx];
4805 	txsd->m = m0;
4806 	txsd->desc_used = ndesc;
4807 
4808 	return (ndesc);
4809 }
4810 
4811 static int
4812 try_txpkts(struct mbuf *m, struct mbuf *n, struct txpkts *txp, u_int available)
4813 {
4814 	u_int needed, nsegs1, nsegs2, l1, l2;
4815 
4816 	if (cannot_use_txpkts(m) || cannot_use_txpkts(n))
4817 		return (1);
4818 
4819 	nsegs1 = mbuf_nsegs(m);
4820 	nsegs2 = mbuf_nsegs(n);
4821 	if (nsegs1 + nsegs2 == 2) {
4822 		txp->wr_type = 1;
4823 		l1 = l2 = txpkts1_len16();
4824 	} else {
4825 		txp->wr_type = 0;
4826 		l1 = txpkts0_len16(nsegs1);
4827 		l2 = txpkts0_len16(nsegs2);
4828 	}
4829 	txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) + l1 + l2;
4830 	needed = howmany(txp->len16, EQ_ESIZE / 16);
4831 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4832 		return (1);
4833 
4834 	txp->plen = m->m_pkthdr.len + n->m_pkthdr.len;
4835 	if (txp->plen > 65535)
4836 		return (1);
4837 
4838 	txp->npkt = 2;
4839 	set_mbuf_len16(m, l1);
4840 	set_mbuf_len16(n, l2);
4841 
4842 	return (0);
4843 }
4844 
4845 static int
4846 add_to_txpkts(struct mbuf *m, struct txpkts *txp, u_int available)
4847 {
4848 	u_int plen, len16, needed, nsegs;
4849 
4850 	MPASS(txp->wr_type == 0 || txp->wr_type == 1);
4851 
4852 	if (cannot_use_txpkts(m))
4853 		return (1);
4854 
4855 	nsegs = mbuf_nsegs(m);
4856 	if (txp->wr_type == 1 && nsegs != 1)
4857 		return (1);
4858 
4859 	plen = txp->plen + m->m_pkthdr.len;
4860 	if (plen > 65535)
4861 		return (1);
4862 
4863 	if (txp->wr_type == 0)
4864 		len16 = txpkts0_len16(nsegs);
4865 	else
4866 		len16 = txpkts1_len16();
4867 	needed = howmany(txp->len16 + len16, EQ_ESIZE / 16);
4868 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4869 		return (1);
4870 
4871 	txp->npkt++;
4872 	txp->plen = plen;
4873 	txp->len16 += len16;
4874 	set_mbuf_len16(m, len16);
4875 
4876 	return (0);
4877 }
4878 
4879 /*
4880  * Write a txpkts WR for the packets in txp to the hardware descriptors, update
4881  * the software descriptor, and advance the pidx.  It is guaranteed that enough
4882  * descriptors are available.
4883  *
4884  * The return value is the # of hardware descriptors used.
4885  */
4886 static u_int
4887 write_txpkts_wr(struct sge_txq *txq, struct fw_eth_tx_pkts_wr *wr,
4888     struct mbuf *m0, const struct txpkts *txp, u_int available)
4889 {
4890 	struct sge_eq *eq = &txq->eq;
4891 	struct tx_sdesc *txsd;
4892 	struct cpl_tx_pkt_core *cpl;
4893 	uint32_t ctrl;
4894 	uint64_t ctrl1;
4895 	int ndesc, checkwrap;
4896 	struct mbuf *m;
4897 	void *flitp;
4898 
4899 	TXQ_LOCK_ASSERT_OWNED(txq);
4900 	MPASS(txp->npkt > 0);
4901 	MPASS(txp->plen < 65536);
4902 	MPASS(m0 != NULL);
4903 	MPASS(m0->m_nextpkt != NULL);
4904 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
4905 	MPASS(available > 0 && available < eq->sidx);
4906 
4907 	ndesc = howmany(txp->len16, EQ_ESIZE / 16);
4908 	MPASS(ndesc <= available);
4909 
4910 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4911 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
4912 	ctrl = V_FW_WR_LEN16(txp->len16);
4913 	wr->equiq_to_len16 = htobe32(ctrl);
4914 	wr->plen = htobe16(txp->plen);
4915 	wr->npkt = txp->npkt;
4916 	wr->r3 = 0;
4917 	wr->type = txp->wr_type;
4918 	flitp = wr + 1;
4919 
4920 	/*
4921 	 * At this point we are 16B into a hardware descriptor.  If checkwrap is
4922 	 * set then we know the WR is going to wrap around somewhere.  We'll
4923 	 * check for that at appropriate points.
4924 	 */
4925 	checkwrap = eq->sidx - ndesc < eq->pidx;
4926 	for (m = m0; m != NULL; m = m->m_nextpkt) {
4927 		if (txp->wr_type == 0) {
4928 			struct ulp_txpkt *ulpmc;
4929 			struct ulptx_idata *ulpsc;
4930 
4931 			/* ULP master command */
4932 			ulpmc = flitp;
4933 			ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
4934 			    V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid));
4935 			ulpmc->len = htobe32(mbuf_len16(m));
4936 
4937 			/* ULP subcommand */
4938 			ulpsc = (void *)(ulpmc + 1);
4939 			ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
4940 			    F_ULP_TX_SC_MORE);
4941 			ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
4942 
4943 			cpl = (void *)(ulpsc + 1);
4944 			if (checkwrap &&
4945 			    (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx])
4946 				cpl = (void *)&eq->desc[0];
4947 		} else {
4948 			cpl = flitp;
4949 		}
4950 
4951 		/* Checksum offload */
4952 		ctrl1 = 0;
4953 		if (needs_l3_csum(m) == 0)
4954 			ctrl1 |= F_TXPKT_IPCSUM_DIS;
4955 		if (needs_l4_csum(m) == 0)
4956 			ctrl1 |= F_TXPKT_L4CSUM_DIS;
4957 		if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4958 		    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4959 			txq->txcsum++;	/* some hardware assistance provided */
4960 
4961 		/* VLAN tag insertion */
4962 		if (needs_vlan_insertion(m)) {
4963 			ctrl1 |= F_TXPKT_VLAN_VLD |
4964 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
4965 			txq->vlan_insertion++;
4966 		}
4967 
4968 		/* CPL header */
4969 		cpl->ctrl0 = txq->cpl_ctrl0;
4970 		cpl->pack = 0;
4971 		cpl->len = htobe16(m->m_pkthdr.len);
4972 		cpl->ctrl1 = htobe64(ctrl1);
4973 
4974 		flitp = cpl + 1;
4975 		if (checkwrap &&
4976 		    (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
4977 			flitp = (void *)&eq->desc[0];
4978 
4979 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap);
4980 
4981 	}
4982 
4983 	if (txp->wr_type == 0) {
4984 		txq->txpkts0_pkts += txp->npkt;
4985 		txq->txpkts0_wrs++;
4986 	} else {
4987 		txq->txpkts1_pkts += txp->npkt;
4988 		txq->txpkts1_wrs++;
4989 	}
4990 
4991 	txsd = &txq->sdesc[eq->pidx];
4992 	txsd->m = m0;
4993 	txsd->desc_used = ndesc;
4994 
4995 	return (ndesc);
4996 }
4997 
4998 /*
4999  * If the SGL ends on an address that is not 16 byte aligned, this function will
5000  * add a 0 filled flit at the end.
5001  */
5002 static void
5003 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap)
5004 {
5005 	struct sge_eq *eq = &txq->eq;
5006 	struct sglist *gl = txq->gl;
5007 	struct sglist_seg *seg;
5008 	__be64 *flitp, *wrap;
5009 	struct ulptx_sgl *usgl;
5010 	int i, nflits, nsegs;
5011 
5012 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
5013 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
5014 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
5015 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
5016 
5017 	get_pkt_gl(m, gl);
5018 	nsegs = gl->sg_nseg;
5019 	MPASS(nsegs > 0);
5020 
5021 	nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
5022 	flitp = (__be64 *)(*to);
5023 	wrap = (__be64 *)(&eq->desc[eq->sidx]);
5024 	seg = &gl->sg_segs[0];
5025 	usgl = (void *)flitp;
5026 
5027 	/*
5028 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
5029 	 * ring, so we're at least 16 bytes away from the status page.  There is
5030 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
5031 	 */
5032 
5033 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
5034 	    V_ULPTX_NSGE(nsegs));
5035 	usgl->len0 = htobe32(seg->ss_len);
5036 	usgl->addr0 = htobe64(seg->ss_paddr);
5037 	seg++;
5038 
5039 	if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) {
5040 
5041 		/* Won't wrap around at all */
5042 
5043 		for (i = 0; i < nsegs - 1; i++, seg++) {
5044 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
5045 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
5046 		}
5047 		if (i & 1)
5048 			usgl->sge[i / 2].len[1] = htobe32(0);
5049 		flitp += nflits;
5050 	} else {
5051 
5052 		/* Will wrap somewhere in the rest of the SGL */
5053 
5054 		/* 2 flits already written, write the rest flit by flit */
5055 		flitp = (void *)(usgl + 1);
5056 		for (i = 0; i < nflits - 2; i++) {
5057 			if (flitp == wrap)
5058 				flitp = (void *)eq->desc;
5059 			*flitp++ = get_flit(seg, nsegs - 1, i);
5060 		}
5061 	}
5062 
5063 	if (nflits & 1) {
5064 		MPASS(((uintptr_t)flitp) & 0xf);
5065 		*flitp++ = 0;
5066 	}
5067 
5068 	MPASS((((uintptr_t)flitp) & 0xf) == 0);
5069 	if (__predict_false(flitp == wrap))
5070 		*to = (void *)eq->desc;
5071 	else
5072 		*to = (void *)flitp;
5073 }
5074 
5075 static inline void
5076 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
5077 {
5078 
5079 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
5080 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
5081 
5082 	if (__predict_true((uintptr_t)(*to) + len <=
5083 	    (uintptr_t)&eq->desc[eq->sidx])) {
5084 		bcopy(from, *to, len);
5085 		(*to) += len;
5086 	} else {
5087 		int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
5088 
5089 		bcopy(from, *to, portion);
5090 		from += portion;
5091 		portion = len - portion;	/* remaining */
5092 		bcopy(from, (void *)eq->desc, portion);
5093 		(*to) = (caddr_t)eq->desc + portion;
5094 	}
5095 }
5096 
5097 static inline void
5098 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n)
5099 {
5100 	u_int db;
5101 
5102 	MPASS(n > 0);
5103 
5104 	db = eq->doorbells;
5105 	if (n > 1)
5106 		clrbit(&db, DOORBELL_WCWR);
5107 	wmb();
5108 
5109 	switch (ffs(db) - 1) {
5110 	case DOORBELL_UDB:
5111 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
5112 		break;
5113 
5114 	case DOORBELL_WCWR: {
5115 		volatile uint64_t *dst, *src;
5116 		int i;
5117 
5118 		/*
5119 		 * Queues whose 128B doorbell segment fits in the page do not
5120 		 * use relative qid (udb_qid is always 0).  Only queues with
5121 		 * doorbell segments can do WCWR.
5122 		 */
5123 		KASSERT(eq->udb_qid == 0 && n == 1,
5124 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
5125 		    __func__, eq->doorbells, n, eq->dbidx, eq));
5126 
5127 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
5128 		    UDBS_DB_OFFSET);
5129 		i = eq->dbidx;
5130 		src = (void *)&eq->desc[i];
5131 		while (src != (void *)&eq->desc[i + 1])
5132 			*dst++ = *src++;
5133 		wmb();
5134 		break;
5135 	}
5136 
5137 	case DOORBELL_UDBWC:
5138 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
5139 		wmb();
5140 		break;
5141 
5142 	case DOORBELL_KDB:
5143 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
5144 		    V_QID(eq->cntxt_id) | V_PIDX(n));
5145 		break;
5146 	}
5147 
5148 	IDXINCR(eq->dbidx, n, eq->sidx);
5149 }
5150 
5151 static inline u_int
5152 reclaimable_tx_desc(struct sge_eq *eq)
5153 {
5154 	uint16_t hw_cidx;
5155 
5156 	hw_cidx = read_hw_cidx(eq);
5157 	return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx));
5158 }
5159 
5160 static inline u_int
5161 total_available_tx_desc(struct sge_eq *eq)
5162 {
5163 	uint16_t hw_cidx, pidx;
5164 
5165 	hw_cidx = read_hw_cidx(eq);
5166 	pidx = eq->pidx;
5167 
5168 	if (pidx == hw_cidx)
5169 		return (eq->sidx - 1);
5170 	else
5171 		return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1);
5172 }
5173 
5174 static inline uint16_t
5175 read_hw_cidx(struct sge_eq *eq)
5176 {
5177 	struct sge_qstat *spg = (void *)&eq->desc[eq->sidx];
5178 	uint16_t cidx = spg->cidx;	/* stable snapshot */
5179 
5180 	return (be16toh(cidx));
5181 }
5182 
5183 /*
5184  * Reclaim 'n' descriptors approximately.
5185  */
5186 static u_int
5187 reclaim_tx_descs(struct sge_txq *txq, u_int n)
5188 {
5189 	struct tx_sdesc *txsd;
5190 	struct sge_eq *eq = &txq->eq;
5191 	u_int can_reclaim, reclaimed;
5192 
5193 	TXQ_LOCK_ASSERT_OWNED(txq);
5194 	MPASS(n > 0);
5195 
5196 	reclaimed = 0;
5197 	can_reclaim = reclaimable_tx_desc(eq);
5198 	while (can_reclaim && reclaimed < n) {
5199 		int ndesc;
5200 		struct mbuf *m, *nextpkt;
5201 
5202 		txsd = &txq->sdesc[eq->cidx];
5203 		ndesc = txsd->desc_used;
5204 
5205 		/* Firmware doesn't return "partial" credits. */
5206 		KASSERT(can_reclaim >= ndesc,
5207 		    ("%s: unexpected number of credits: %d, %d",
5208 		    __func__, can_reclaim, ndesc));
5209 		KASSERT(ndesc != 0,
5210 		    ("%s: descriptor with no credits: cidx %d",
5211 		    __func__, eq->cidx));
5212 
5213 		for (m = txsd->m; m != NULL; m = nextpkt) {
5214 			nextpkt = m->m_nextpkt;
5215 			m->m_nextpkt = NULL;
5216 			m_freem(m);
5217 		}
5218 		reclaimed += ndesc;
5219 		can_reclaim -= ndesc;
5220 		IDXINCR(eq->cidx, ndesc, eq->sidx);
5221 	}
5222 
5223 	return (reclaimed);
5224 }
5225 
5226 static void
5227 tx_reclaim(void *arg, int n)
5228 {
5229 	struct sge_txq *txq = arg;
5230 	struct sge_eq *eq = &txq->eq;
5231 
5232 	do {
5233 		if (TXQ_TRYLOCK(txq) == 0)
5234 			break;
5235 		n = reclaim_tx_descs(txq, 32);
5236 		if (eq->cidx == eq->pidx)
5237 			eq->equeqidx = eq->pidx;
5238 		TXQ_UNLOCK(txq);
5239 	} while (n > 0);
5240 }
5241 
5242 static __be64
5243 get_flit(struct sglist_seg *segs, int nsegs, int idx)
5244 {
5245 	int i = (idx / 3) * 2;
5246 
5247 	switch (idx % 3) {
5248 	case 0: {
5249 		uint64_t rc;
5250 
5251 		rc = (uint64_t)segs[i].ss_len << 32;
5252 		if (i + 1 < nsegs)
5253 			rc |= (uint64_t)(segs[i + 1].ss_len);
5254 
5255 		return (htobe64(rc));
5256 	}
5257 	case 1:
5258 		return (htobe64(segs[i].ss_paddr));
5259 	case 2:
5260 		return (htobe64(segs[i + 1].ss_paddr));
5261 	}
5262 
5263 	return (0);
5264 }
5265 
5266 static void
5267 find_best_refill_source(struct adapter *sc, struct sge_fl *fl, int maxp)
5268 {
5269 	int8_t zidx, hwidx, idx;
5270 	uint16_t region1, region3;
5271 	int spare, spare_needed, n;
5272 	struct sw_zone_info *swz;
5273 	struct hw_buf_info *hwb, *hwb_list = &sc->sge.hw_buf_info[0];
5274 
5275 	/*
5276 	 * Buffer Packing: Look for PAGE_SIZE or larger zone which has a bufsize
5277 	 * large enough for the max payload and cluster metadata.  Otherwise
5278 	 * settle for the largest bufsize that leaves enough room in the cluster
5279 	 * for metadata.
5280 	 *
5281 	 * Without buffer packing: Look for the smallest zone which has a
5282 	 * bufsize large enough for the max payload.  Settle for the largest
5283 	 * bufsize available if there's nothing big enough for max payload.
5284 	 */
5285 	spare_needed = fl->flags & FL_BUF_PACKING ? CL_METADATA_SIZE : 0;
5286 	swz = &sc->sge.sw_zone_info[0];
5287 	hwidx = -1;
5288 	for (zidx = 0; zidx < SW_ZONE_SIZES; zidx++, swz++) {
5289 		if (swz->size > largest_rx_cluster) {
5290 			if (__predict_true(hwidx != -1))
5291 				break;
5292 
5293 			/*
5294 			 * This is a misconfiguration.  largest_rx_cluster is
5295 			 * preventing us from finding a refill source.  See
5296 			 * dev.t5nex.<n>.buffer_sizes to figure out why.
5297 			 */
5298 			device_printf(sc->dev, "largest_rx_cluster=%u leaves no"
5299 			    " refill source for fl %p (dma %u).  Ignored.\n",
5300 			    largest_rx_cluster, fl, maxp);
5301 		}
5302 		for (idx = swz->head_hwidx; idx != -1; idx = hwb->next) {
5303 			hwb = &hwb_list[idx];
5304 			spare = swz->size - hwb->size;
5305 			if (spare < spare_needed)
5306 				continue;
5307 
5308 			hwidx = idx;		/* best option so far */
5309 			if (hwb->size >= maxp) {
5310 
5311 				if ((fl->flags & FL_BUF_PACKING) == 0)
5312 					goto done; /* stop looking (not packing) */
5313 
5314 				if (swz->size >= safest_rx_cluster)
5315 					goto done; /* stop looking (packing) */
5316 			}
5317 			break;		/* keep looking, next zone */
5318 		}
5319 	}
5320 done:
5321 	/* A usable hwidx has been located. */
5322 	MPASS(hwidx != -1);
5323 	hwb = &hwb_list[hwidx];
5324 	zidx = hwb->zidx;
5325 	swz = &sc->sge.sw_zone_info[zidx];
5326 	region1 = 0;
5327 	region3 = swz->size - hwb->size;
5328 
5329 	/*
5330 	 * Stay within this zone and see if there is a better match when mbuf
5331 	 * inlining is allowed.  Remember that the hwidx's are sorted in
5332 	 * decreasing order of size (so in increasing order of spare area).
5333 	 */
5334 	for (idx = hwidx; idx != -1; idx = hwb->next) {
5335 		hwb = &hwb_list[idx];
5336 		spare = swz->size - hwb->size;
5337 
5338 		if (allow_mbufs_in_cluster == 0 || hwb->size < maxp)
5339 			break;
5340 
5341 		/*
5342 		 * Do not inline mbufs if doing so would violate the pad/pack
5343 		 * boundary alignment requirement.
5344 		 */
5345 		if (fl_pad && (MSIZE % sc->params.sge.pad_boundary) != 0)
5346 			continue;
5347 		if (fl->flags & FL_BUF_PACKING &&
5348 		    (MSIZE % sc->params.sge.pack_boundary) != 0)
5349 			continue;
5350 
5351 		if (spare < CL_METADATA_SIZE + MSIZE)
5352 			continue;
5353 		n = (spare - CL_METADATA_SIZE) / MSIZE;
5354 		if (n > howmany(hwb->size, maxp))
5355 			break;
5356 
5357 		hwidx = idx;
5358 		if (fl->flags & FL_BUF_PACKING) {
5359 			region1 = n * MSIZE;
5360 			region3 = spare - region1;
5361 		} else {
5362 			region1 = MSIZE;
5363 			region3 = spare - region1;
5364 			break;
5365 		}
5366 	}
5367 
5368 	KASSERT(zidx >= 0 && zidx < SW_ZONE_SIZES,
5369 	    ("%s: bad zone %d for fl %p, maxp %d", __func__, zidx, fl, maxp));
5370 	KASSERT(hwidx >= 0 && hwidx <= SGE_FLBUF_SIZES,
5371 	    ("%s: bad hwidx %d for fl %p, maxp %d", __func__, hwidx, fl, maxp));
5372 	KASSERT(region1 + sc->sge.hw_buf_info[hwidx].size + region3 ==
5373 	    sc->sge.sw_zone_info[zidx].size,
5374 	    ("%s: bad buffer layout for fl %p, maxp %d. "
5375 		"cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5376 		sc->sge.sw_zone_info[zidx].size, region1,
5377 		sc->sge.hw_buf_info[hwidx].size, region3));
5378 	if (fl->flags & FL_BUF_PACKING || region1 > 0) {
5379 		KASSERT(region3 >= CL_METADATA_SIZE,
5380 		    ("%s: no room for metadata.  fl %p, maxp %d; "
5381 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5382 		    sc->sge.sw_zone_info[zidx].size, region1,
5383 		    sc->sge.hw_buf_info[hwidx].size, region3));
5384 		KASSERT(region1 % MSIZE == 0,
5385 		    ("%s: bad mbuf region 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 	}
5390 
5391 	fl->cll_def.zidx = zidx;
5392 	fl->cll_def.hwidx = hwidx;
5393 	fl->cll_def.region1 = region1;
5394 	fl->cll_def.region3 = region3;
5395 }
5396 
5397 static void
5398 find_safe_refill_source(struct adapter *sc, struct sge_fl *fl)
5399 {
5400 	struct sge *s = &sc->sge;
5401 	struct hw_buf_info *hwb;
5402 	struct sw_zone_info *swz;
5403 	int spare;
5404 	int8_t hwidx;
5405 
5406 	if (fl->flags & FL_BUF_PACKING)
5407 		hwidx = s->safe_hwidx2;	/* with room for metadata */
5408 	else if (allow_mbufs_in_cluster && s->safe_hwidx2 != -1) {
5409 		hwidx = s->safe_hwidx2;
5410 		hwb = &s->hw_buf_info[hwidx];
5411 		swz = &s->sw_zone_info[hwb->zidx];
5412 		spare = swz->size - hwb->size;
5413 
5414 		/* no good if there isn't room for an mbuf as well */
5415 		if (spare < CL_METADATA_SIZE + MSIZE)
5416 			hwidx = s->safe_hwidx1;
5417 	} else
5418 		hwidx = s->safe_hwidx1;
5419 
5420 	if (hwidx == -1) {
5421 		/* No fallback source */
5422 		fl->cll_alt.hwidx = -1;
5423 		fl->cll_alt.zidx = -1;
5424 
5425 		return;
5426 	}
5427 
5428 	hwb = &s->hw_buf_info[hwidx];
5429 	swz = &s->sw_zone_info[hwb->zidx];
5430 	spare = swz->size - hwb->size;
5431 	fl->cll_alt.hwidx = hwidx;
5432 	fl->cll_alt.zidx = hwb->zidx;
5433 	if (allow_mbufs_in_cluster &&
5434 	    (fl_pad == 0 || (MSIZE % sc->params.sge.pad_boundary) == 0))
5435 		fl->cll_alt.region1 = ((spare - CL_METADATA_SIZE) / MSIZE) * MSIZE;
5436 	else
5437 		fl->cll_alt.region1 = 0;
5438 	fl->cll_alt.region3 = spare - fl->cll_alt.region1;
5439 }
5440 
5441 static void
5442 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
5443 {
5444 	mtx_lock(&sc->sfl_lock);
5445 	FL_LOCK(fl);
5446 	if ((fl->flags & FL_DOOMED) == 0) {
5447 		fl->flags |= FL_STARVING;
5448 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
5449 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
5450 	}
5451 	FL_UNLOCK(fl);
5452 	mtx_unlock(&sc->sfl_lock);
5453 }
5454 
5455 static void
5456 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq)
5457 {
5458 	struct sge_wrq *wrq = (void *)eq;
5459 
5460 	atomic_readandclear_int(&eq->equiq);
5461 	taskqueue_enqueue(sc->tq[eq->tx_chan], &wrq->wrq_tx_task);
5462 }
5463 
5464 static void
5465 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq)
5466 {
5467 	struct sge_txq *txq = (void *)eq;
5468 
5469 	MPASS((eq->flags & EQ_TYPEMASK) == EQ_ETH);
5470 
5471 	atomic_readandclear_int(&eq->equiq);
5472 	mp_ring_check_drainage(txq->r, 0);
5473 	taskqueue_enqueue(sc->tq[eq->tx_chan], &txq->tx_reclaim_task);
5474 }
5475 
5476 static int
5477 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
5478     struct mbuf *m)
5479 {
5480 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
5481 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
5482 	struct adapter *sc = iq->adapter;
5483 	struct sge *s = &sc->sge;
5484 	struct sge_eq *eq;
5485 	static void (*h[])(struct adapter *, struct sge_eq *) = {NULL,
5486 		&handle_wrq_egr_update, &handle_eth_egr_update,
5487 		&handle_wrq_egr_update};
5488 
5489 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5490 	    rss->opcode));
5491 
5492 	eq = s->eqmap[qid - s->eq_start - s->eq_base];
5493 	(*h[eq->flags & EQ_TYPEMASK])(sc, eq);
5494 
5495 	return (0);
5496 }
5497 
5498 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
5499 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
5500     offsetof(struct cpl_fw6_msg, data));
5501 
5502 static int
5503 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
5504 {
5505 	struct adapter *sc = iq->adapter;
5506 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
5507 
5508 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5509 	    rss->opcode));
5510 
5511 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
5512 		const struct rss_header *rss2;
5513 
5514 		rss2 = (const struct rss_header *)&cpl->data[0];
5515 		return (t4_cpl_handler[rss2->opcode](iq, rss2, m));
5516 	}
5517 
5518 	return (t4_fw_msg_handler[cpl->type](sc, &cpl->data[0]));
5519 }
5520 
5521 /**
5522  *	t4_handle_wrerr_rpl - process a FW work request error message
5523  *	@adap: the adapter
5524  *	@rpl: start of the FW message
5525  */
5526 static int
5527 t4_handle_wrerr_rpl(struct adapter *adap, const __be64 *rpl)
5528 {
5529 	u8 opcode = *(const u8 *)rpl;
5530 	const struct fw_error_cmd *e = (const void *)rpl;
5531 	unsigned int i;
5532 
5533 	if (opcode != FW_ERROR_CMD) {
5534 		log(LOG_ERR,
5535 		    "%s: Received WRERR_RPL message with opcode %#x\n",
5536 		    device_get_nameunit(adap->dev), opcode);
5537 		return (EINVAL);
5538 	}
5539 	log(LOG_ERR, "%s: FW_ERROR (%s) ", device_get_nameunit(adap->dev),
5540 	    G_FW_ERROR_CMD_FATAL(be32toh(e->op_to_type)) ? "fatal" :
5541 	    "non-fatal");
5542 	switch (G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type))) {
5543 	case FW_ERROR_TYPE_EXCEPTION:
5544 		log(LOG_ERR, "exception info:\n");
5545 		for (i = 0; i < nitems(e->u.exception.info); i++)
5546 			log(LOG_ERR, "%s%08x", i == 0 ? "\t" : " ",
5547 			    be32toh(e->u.exception.info[i]));
5548 		log(LOG_ERR, "\n");
5549 		break;
5550 	case FW_ERROR_TYPE_HWMODULE:
5551 		log(LOG_ERR, "HW module regaddr %08x regval %08x\n",
5552 		    be32toh(e->u.hwmodule.regaddr),
5553 		    be32toh(e->u.hwmodule.regval));
5554 		break;
5555 	case FW_ERROR_TYPE_WR:
5556 		log(LOG_ERR, "WR cidx %d PF %d VF %d eqid %d hdr:\n",
5557 		    be16toh(e->u.wr.cidx),
5558 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.wr.pfn_vfn)),
5559 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.wr.pfn_vfn)),
5560 		    be32toh(e->u.wr.eqid));
5561 		for (i = 0; i < nitems(e->u.wr.wrhdr); i++)
5562 			log(LOG_ERR, "%s%02x", i == 0 ? "\t" : " ",
5563 			    e->u.wr.wrhdr[i]);
5564 		log(LOG_ERR, "\n");
5565 		break;
5566 	case FW_ERROR_TYPE_ACL:
5567 		log(LOG_ERR, "ACL cidx %d PF %d VF %d eqid %d %s",
5568 		    be16toh(e->u.acl.cidx),
5569 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.acl.pfn_vfn)),
5570 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.acl.pfn_vfn)),
5571 		    be32toh(e->u.acl.eqid),
5572 		    G_FW_ERROR_CMD_MV(be16toh(e->u.acl.mv_pkd)) ? "vlanid" :
5573 		    "MAC");
5574 		for (i = 0; i < nitems(e->u.acl.val); i++)
5575 			log(LOG_ERR, " %02x", e->u.acl.val[i]);
5576 		log(LOG_ERR, "\n");
5577 		break;
5578 	default:
5579 		log(LOG_ERR, "type %#x\n",
5580 		    G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type)));
5581 		return (EINVAL);
5582 	}
5583 	return (0);
5584 }
5585 
5586 static int
5587 sysctl_uint16(SYSCTL_HANDLER_ARGS)
5588 {
5589 	uint16_t *id = arg1;
5590 	int i = *id;
5591 
5592 	return sysctl_handle_int(oidp, &i, 0, req);
5593 }
5594 
5595 static int
5596 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
5597 {
5598 	struct sge *s = arg1;
5599 	struct hw_buf_info *hwb = &s->hw_buf_info[0];
5600 	struct sw_zone_info *swz = &s->sw_zone_info[0];
5601 	int i, rc;
5602 	struct sbuf sb;
5603 	char c;
5604 
5605 	sbuf_new(&sb, NULL, 32, SBUF_AUTOEXTEND);
5606 	for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
5607 		if (hwb->zidx >= 0 && swz[hwb->zidx].size <= largest_rx_cluster)
5608 			c = '*';
5609 		else
5610 			c = '\0';
5611 
5612 		sbuf_printf(&sb, "%u%c ", hwb->size, c);
5613 	}
5614 	sbuf_trim(&sb);
5615 	sbuf_finish(&sb);
5616 	rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
5617 	sbuf_delete(&sb);
5618 	return (rc);
5619 }
5620 
5621 #ifdef RATELIMIT
5622 /*
5623  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
5624  */
5625 static inline u_int
5626 txpkt_eo_len16(u_int nsegs, u_int immhdrs, u_int tso)
5627 {
5628 	u_int n;
5629 
5630 	MPASS(immhdrs > 0);
5631 
5632 	n = roundup2(sizeof(struct fw_eth_tx_eo_wr) +
5633 	    sizeof(struct cpl_tx_pkt_core) + immhdrs, 16);
5634 	if (__predict_false(nsegs == 0))
5635 		goto done;
5636 
5637 	nsegs--; /* first segment is part of ulptx_sgl */
5638 	n += sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
5639 	if (tso)
5640 		n += sizeof(struct cpl_tx_pkt_lso_core);
5641 
5642 done:
5643 	return (howmany(n, 16));
5644 }
5645 
5646 #define ETID_FLOWC_NPARAMS 6
5647 #define ETID_FLOWC_LEN (roundup2((sizeof(struct fw_flowc_wr) + \
5648     ETID_FLOWC_NPARAMS * sizeof(struct fw_flowc_mnemval)), 16))
5649 #define ETID_FLOWC_LEN16 (howmany(ETID_FLOWC_LEN, 16))
5650 
5651 static int
5652 send_etid_flowc_wr(struct cxgbe_snd_tag *cst, struct port_info *pi,
5653     struct vi_info *vi)
5654 {
5655 	struct wrq_cookie cookie;
5656 	u_int pfvf = G_FW_VIID_PFN(vi->viid) << S_FW_VIID_PFN;
5657 	struct fw_flowc_wr *flowc;
5658 
5659 	mtx_assert(&cst->lock, MA_OWNED);
5660 	MPASS((cst->flags & (EO_FLOWC_PENDING | EO_FLOWC_RPL_PENDING)) ==
5661 	    EO_FLOWC_PENDING);
5662 
5663 	flowc = start_wrq_wr(cst->eo_txq, ETID_FLOWC_LEN16, &cookie);
5664 	if (__predict_false(flowc == NULL))
5665 		return (ENOMEM);
5666 
5667 	bzero(flowc, ETID_FLOWC_LEN);
5668 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
5669 	    V_FW_FLOWC_WR_NPARAMS(ETID_FLOWC_NPARAMS) | V_FW_WR_COMPL(0));
5670 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(ETID_FLOWC_LEN16) |
5671 	    V_FW_WR_FLOWID(cst->etid));
5672 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN;
5673 	flowc->mnemval[0].val = htobe32(pfvf);
5674 	flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH;
5675 	flowc->mnemval[1].val = htobe32(pi->tx_chan);
5676 	flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT;
5677 	flowc->mnemval[2].val = htobe32(pi->tx_chan);
5678 	flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID;
5679 	flowc->mnemval[3].val = htobe32(cst->iqid);
5680 	flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_EOSTATE;
5681 	flowc->mnemval[4].val = htobe32(FW_FLOWC_MNEM_EOSTATE_ESTABLISHED);
5682 	flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS;
5683 	flowc->mnemval[5].val = htobe32(cst->schedcl);
5684 
5685 	commit_wrq_wr(cst->eo_txq, flowc, &cookie);
5686 
5687 	cst->flags &= ~EO_FLOWC_PENDING;
5688 	cst->flags |= EO_FLOWC_RPL_PENDING;
5689 	MPASS(cst->tx_credits >= ETID_FLOWC_LEN16);	/* flowc is first WR. */
5690 	cst->tx_credits -= ETID_FLOWC_LEN16;
5691 
5692 	return (0);
5693 }
5694 
5695 #define ETID_FLUSH_LEN16 (howmany(sizeof (struct fw_flowc_wr), 16))
5696 
5697 void
5698 send_etid_flush_wr(struct cxgbe_snd_tag *cst)
5699 {
5700 	struct fw_flowc_wr *flowc;
5701 	struct wrq_cookie cookie;
5702 
5703 	mtx_assert(&cst->lock, MA_OWNED);
5704 
5705 	flowc = start_wrq_wr(cst->eo_txq, ETID_FLUSH_LEN16, &cookie);
5706 	if (__predict_false(flowc == NULL))
5707 		CXGBE_UNIMPLEMENTED(__func__);
5708 
5709 	bzero(flowc, ETID_FLUSH_LEN16 * 16);
5710 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
5711 	    V_FW_FLOWC_WR_NPARAMS(0) | F_FW_WR_COMPL);
5712 	flowc->flowid_len16 = htobe32(V_FW_WR_LEN16(ETID_FLUSH_LEN16) |
5713 	    V_FW_WR_FLOWID(cst->etid));
5714 
5715 	commit_wrq_wr(cst->eo_txq, flowc, &cookie);
5716 
5717 	cst->flags |= EO_FLUSH_RPL_PENDING;
5718 	MPASS(cst->tx_credits >= ETID_FLUSH_LEN16);
5719 	cst->tx_credits -= ETID_FLUSH_LEN16;
5720 	cst->ncompl++;
5721 }
5722 
5723 static void
5724 write_ethofld_wr(struct cxgbe_snd_tag *cst, struct fw_eth_tx_eo_wr *wr,
5725     struct mbuf *m0, int compl)
5726 {
5727 	struct cpl_tx_pkt_core *cpl;
5728 	uint64_t ctrl1;
5729 	uint32_t ctrl;	/* used in many unrelated places */
5730 	int len16, pktlen, nsegs, immhdrs;
5731 	caddr_t dst;
5732 	uintptr_t p;
5733 	struct ulptx_sgl *usgl;
5734 	struct sglist sg;
5735 	struct sglist_seg segs[38];	/* XXX: find real limit.  XXX: get off the stack */
5736 
5737 	mtx_assert(&cst->lock, MA_OWNED);
5738 	M_ASSERTPKTHDR(m0);
5739 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
5740 	    m0->m_pkthdr.l4hlen > 0,
5741 	    ("%s: ethofld mbuf %p is missing header lengths", __func__, m0));
5742 
5743 	len16 = mbuf_eo_len16(m0);
5744 	nsegs = mbuf_eo_nsegs(m0);
5745 	pktlen = m0->m_pkthdr.len;
5746 	ctrl = sizeof(struct cpl_tx_pkt_core);
5747 	if (needs_tso(m0))
5748 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
5749 	immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen + m0->m_pkthdr.l4hlen;
5750 	ctrl += immhdrs;
5751 
5752 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_EO_WR) |
5753 	    V_FW_ETH_TX_EO_WR_IMMDLEN(ctrl) | V_FW_WR_COMPL(!!compl));
5754 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(len16) |
5755 	    V_FW_WR_FLOWID(cst->etid));
5756 	wr->r3 = 0;
5757 	if (needs_udp_csum(m0)) {
5758 		wr->u.udpseg.type = FW_ETH_TX_EO_TYPE_UDPSEG;
5759 		wr->u.udpseg.ethlen = m0->m_pkthdr.l2hlen;
5760 		wr->u.udpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
5761 		wr->u.udpseg.udplen = m0->m_pkthdr.l4hlen;
5762 		wr->u.udpseg.rtplen = 0;
5763 		wr->u.udpseg.r4 = 0;
5764 		wr->u.udpseg.mss = htobe16(pktlen - immhdrs);
5765 		wr->u.udpseg.schedpktsize = wr->u.udpseg.mss;
5766 		wr->u.udpseg.plen = htobe32(pktlen - immhdrs);
5767 		cpl = (void *)(wr + 1);
5768 	} else {
5769 		MPASS(needs_tcp_csum(m0));
5770 		wr->u.tcpseg.type = FW_ETH_TX_EO_TYPE_TCPSEG;
5771 		wr->u.tcpseg.ethlen = m0->m_pkthdr.l2hlen;
5772 		wr->u.tcpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
5773 		wr->u.tcpseg.tcplen = m0->m_pkthdr.l4hlen;
5774 		wr->u.tcpseg.tsclk_tsoff = mbuf_eo_tsclk_tsoff(m0);
5775 		wr->u.tcpseg.r4 = 0;
5776 		wr->u.tcpseg.r5 = 0;
5777 		wr->u.tcpseg.plen = htobe32(pktlen - immhdrs);
5778 
5779 		if (needs_tso(m0)) {
5780 			struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
5781 
5782 			wr->u.tcpseg.mss = htobe16(m0->m_pkthdr.tso_segsz);
5783 
5784 			ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) |
5785 			    F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE |
5786 			    V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) |
5787 			    V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
5788 			if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
5789 				ctrl |= V_LSO_ETHHDR_LEN(1);
5790 			if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
5791 				ctrl |= F_LSO_IPV6;
5792 			lso->lso_ctrl = htobe32(ctrl);
5793 			lso->ipid_ofst = htobe16(0);
5794 			lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
5795 			lso->seqno_offset = htobe32(0);
5796 			lso->len = htobe32(pktlen);
5797 
5798 			cpl = (void *)(lso + 1);
5799 		} else {
5800 			wr->u.tcpseg.mss = htobe16(0xffff);
5801 			cpl = (void *)(wr + 1);
5802 		}
5803 	}
5804 
5805 	/* Checksum offload must be requested for ethofld. */
5806 	ctrl1 = 0;
5807 	MPASS(needs_l4_csum(m0));
5808 
5809 	/* VLAN tag insertion */
5810 	if (needs_vlan_insertion(m0)) {
5811 		ctrl1 |= F_TXPKT_VLAN_VLD |
5812 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
5813 	}
5814 
5815 	/* CPL header */
5816 	cpl->ctrl0 = cst->ctrl0;
5817 	cpl->pack = 0;
5818 	cpl->len = htobe16(pktlen);
5819 	cpl->ctrl1 = htobe64(ctrl1);
5820 
5821 	/* Copy Ethernet, IP & TCP/UDP hdrs as immediate data */
5822 	p = (uintptr_t)(cpl + 1);
5823 	m_copydata(m0, 0, immhdrs, (void *)p);
5824 
5825 	/* SGL */
5826 	dst = (void *)(cpl + 1);
5827 	if (nsegs > 0) {
5828 		int i, pad;
5829 
5830 		/* zero-pad upto next 16Byte boundary, if not 16Byte aligned */
5831 		p += immhdrs;
5832 		pad = 16 - (immhdrs & 0xf);
5833 		bzero((void *)p, pad);
5834 
5835 		usgl = (void *)(p + pad);
5836 		usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
5837 		    V_ULPTX_NSGE(nsegs));
5838 
5839 		sglist_init(&sg, nitems(segs), segs);
5840 		for (; m0 != NULL; m0 = m0->m_next) {
5841 			if (__predict_false(m0->m_len == 0))
5842 				continue;
5843 			if (immhdrs >= m0->m_len) {
5844 				immhdrs -= m0->m_len;
5845 				continue;
5846 			}
5847 
5848 			sglist_append(&sg, mtod(m0, char *) + immhdrs,
5849 			    m0->m_len - immhdrs);
5850 			immhdrs = 0;
5851 		}
5852 		MPASS(sg.sg_nseg == nsegs);
5853 
5854 		/*
5855 		 * Zero pad last 8B in case the WR doesn't end on a 16B
5856 		 * boundary.
5857 		 */
5858 		*(uint64_t *)((char *)wr + len16 * 16 - 8) = 0;
5859 
5860 		usgl->len0 = htobe32(segs[0].ss_len);
5861 		usgl->addr0 = htobe64(segs[0].ss_paddr);
5862 		for (i = 0; i < nsegs - 1; i++) {
5863 			usgl->sge[i / 2].len[i & 1] = htobe32(segs[i + 1].ss_len);
5864 			usgl->sge[i / 2].addr[i & 1] = htobe64(segs[i + 1].ss_paddr);
5865 		}
5866 		if (i & 1)
5867 			usgl->sge[i / 2].len[1] = htobe32(0);
5868 	}
5869 
5870 }
5871 
5872 static void
5873 ethofld_tx(struct cxgbe_snd_tag *cst)
5874 {
5875 	struct mbuf *m;
5876 	struct wrq_cookie cookie;
5877 	int next_credits, compl;
5878 	struct fw_eth_tx_eo_wr *wr;
5879 
5880 	mtx_assert(&cst->lock, MA_OWNED);
5881 
5882 	while ((m = mbufq_first(&cst->pending_tx)) != NULL) {
5883 		M_ASSERTPKTHDR(m);
5884 
5885 		/* How many len16 credits do we need to send this mbuf. */
5886 		next_credits = mbuf_eo_len16(m);
5887 		MPASS(next_credits > 0);
5888 		if (next_credits > cst->tx_credits) {
5889 			/*
5890 			 * Tx will make progress eventually because there is at
5891 			 * least one outstanding fw4_ack that will return
5892 			 * credits and kick the tx.
5893 			 */
5894 			MPASS(cst->ncompl > 0);
5895 			return;
5896 		}
5897 		wr = start_wrq_wr(cst->eo_txq, next_credits, &cookie);
5898 		if (__predict_false(wr == NULL)) {
5899 			/* XXX: wishful thinking, not a real assertion. */
5900 			MPASS(cst->ncompl > 0);
5901 			return;
5902 		}
5903 		cst->tx_credits -= next_credits;
5904 		cst->tx_nocompl += next_credits;
5905 		compl = cst->ncompl == 0 || cst->tx_nocompl >= cst->tx_total / 2;
5906 		ETHER_BPF_MTAP(cst->com.ifp, m);
5907 		write_ethofld_wr(cst, wr, m, compl);
5908 		commit_wrq_wr(cst->eo_txq, wr, &cookie);
5909 		if (compl) {
5910 			cst->ncompl++;
5911 			cst->tx_nocompl	= 0;
5912 		}
5913 		(void) mbufq_dequeue(&cst->pending_tx);
5914 		mbufq_enqueue(&cst->pending_fwack, m);
5915 	}
5916 }
5917 
5918 int
5919 ethofld_transmit(struct ifnet *ifp, struct mbuf *m0)
5920 {
5921 	struct cxgbe_snd_tag *cst;
5922 	int rc;
5923 
5924 	MPASS(m0->m_nextpkt == NULL);
5925 	MPASS(m0->m_pkthdr.snd_tag != NULL);
5926 	cst = mst_to_cst(m0->m_pkthdr.snd_tag);
5927 
5928 	mtx_lock(&cst->lock);
5929 	MPASS(cst->flags & EO_SND_TAG_REF);
5930 
5931 	if (__predict_false(cst->flags & EO_FLOWC_PENDING)) {
5932 		struct vi_info *vi = ifp->if_softc;
5933 		struct port_info *pi = vi->pi;
5934 		struct adapter *sc = pi->adapter;
5935 		const uint32_t rss_mask = vi->rss_size - 1;
5936 		uint32_t rss_hash;
5937 
5938 		cst->eo_txq = &sc->sge.ofld_txq[vi->first_ofld_txq];
5939 		if (M_HASHTYPE_ISHASH(m0))
5940 			rss_hash = m0->m_pkthdr.flowid;
5941 		else
5942 			rss_hash = arc4random();
5943 		/* We assume RSS hashing */
5944 		cst->iqid = vi->rss[rss_hash & rss_mask];
5945 		cst->eo_txq += rss_hash % vi->nofldtxq;
5946 		rc = send_etid_flowc_wr(cst, pi, vi);
5947 		if (rc != 0)
5948 			goto done;
5949 	}
5950 
5951 	if (__predict_false(cst->plen + m0->m_pkthdr.len > eo_max_backlog)) {
5952 		rc = ENOBUFS;
5953 		goto done;
5954 	}
5955 
5956 	mbufq_enqueue(&cst->pending_tx, m0);
5957 	cst->plen += m0->m_pkthdr.len;
5958 
5959 	ethofld_tx(cst);
5960 	rc = 0;
5961 done:
5962 	mtx_unlock(&cst->lock);
5963 	if (__predict_false(rc != 0))
5964 		m_freem(m0);
5965 	return (rc);
5966 }
5967 
5968 static int
5969 ethofld_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
5970 {
5971 	struct adapter *sc = iq->adapter;
5972 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
5973 	struct mbuf *m;
5974 	u_int etid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
5975 	struct cxgbe_snd_tag *cst;
5976 	uint8_t credits = cpl->credits;
5977 
5978 	cst = lookup_etid(sc, etid);
5979 	mtx_lock(&cst->lock);
5980 	if (__predict_false(cst->flags & EO_FLOWC_RPL_PENDING)) {
5981 		MPASS(credits >= ETID_FLOWC_LEN16);
5982 		credits -= ETID_FLOWC_LEN16;
5983 		cst->flags &= ~EO_FLOWC_RPL_PENDING;
5984 	}
5985 
5986 	KASSERT(cst->ncompl > 0,
5987 	    ("%s: etid %u (%p) wasn't expecting completion.",
5988 	    __func__, etid, cst));
5989 	cst->ncompl--;
5990 
5991 	while (credits > 0) {
5992 		m = mbufq_dequeue(&cst->pending_fwack);
5993 		if (__predict_false(m == NULL)) {
5994 			/*
5995 			 * The remaining credits are for the final flush that
5996 			 * was issued when the tag was freed by the kernel.
5997 			 */
5998 			MPASS((cst->flags &
5999 			    (EO_FLUSH_RPL_PENDING | EO_SND_TAG_REF)) ==
6000 			    EO_FLUSH_RPL_PENDING);
6001 			MPASS(credits == ETID_FLUSH_LEN16);
6002 			MPASS(cst->tx_credits + cpl->credits == cst->tx_total);
6003 			MPASS(cst->ncompl == 0);
6004 
6005 			cst->flags &= ~EO_FLUSH_RPL_PENDING;
6006 			cst->tx_credits += cpl->credits;
6007 freetag:
6008 			cxgbe_snd_tag_free_locked(cst);
6009 			return (0);	/* cst is gone. */
6010 		}
6011 		KASSERT(m != NULL,
6012 		    ("%s: too many credits (%u, %u)", __func__, cpl->credits,
6013 		    credits));
6014 		KASSERT(credits >= mbuf_eo_len16(m),
6015 		    ("%s: too few credits (%u, %u, %u)", __func__,
6016 		    cpl->credits, credits, mbuf_eo_len16(m)));
6017 		credits -= mbuf_eo_len16(m);
6018 		cst->plen -= m->m_pkthdr.len;
6019 		m_freem(m);
6020 	}
6021 
6022 	cst->tx_credits += cpl->credits;
6023 	MPASS(cst->tx_credits <= cst->tx_total);
6024 
6025 	m = mbufq_first(&cst->pending_tx);
6026 	if (m != NULL && cst->tx_credits >= mbuf_eo_len16(m))
6027 		ethofld_tx(cst);
6028 
6029 	if (__predict_false((cst->flags & EO_SND_TAG_REF) == 0) &&
6030 	    cst->ncompl == 0) {
6031 		if (cst->tx_credits == cst->tx_total)
6032 			goto freetag;
6033 		else {
6034 			MPASS((cst->flags & EO_FLUSH_RPL_PENDING) == 0);
6035 			send_etid_flush_wr(cst);
6036 		}
6037 	}
6038 
6039 	mtx_unlock(&cst->lock);
6040 
6041 	return (0);
6042 }
6043 #endif
6044