xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision ca987d4641cdcd7f27e153db17c5bf064934faf5)
1 /*-
2  * Copyright (c) 2011 Chelsio Communications, Inc.
3  * All rights reserved.
4  * Written by: Navdeep Parhar <np@FreeBSD.org>
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 
34 #include <sys/types.h>
35 #include <sys/eventhandler.h>
36 #include <sys/mbuf.h>
37 #include <sys/socket.h>
38 #include <sys/kernel.h>
39 #include <sys/malloc.h>
40 #include <sys/queue.h>
41 #include <sys/sbuf.h>
42 #include <sys/taskqueue.h>
43 #include <sys/time.h>
44 #include <sys/sglist.h>
45 #include <sys/sysctl.h>
46 #include <sys/smp.h>
47 #include <sys/counter.h>
48 #include <net/bpf.h>
49 #include <net/ethernet.h>
50 #include <net/if.h>
51 #include <net/if_vlan_var.h>
52 #include <netinet/in.h>
53 #include <netinet/ip.h>
54 #include <netinet/ip6.h>
55 #include <netinet/tcp.h>
56 #include <machine/in_cksum.h>
57 #include <machine/md_var.h>
58 #include <vm/vm.h>
59 #include <vm/pmap.h>
60 #ifdef DEV_NETMAP
61 #include <machine/bus.h>
62 #include <sys/selinfo.h>
63 #include <net/if_var.h>
64 #include <net/netmap.h>
65 #include <dev/netmap/netmap_kern.h>
66 #endif
67 
68 #include "common/common.h"
69 #include "common/t4_regs.h"
70 #include "common/t4_regs_values.h"
71 #include "common/t4_msg.h"
72 #include "t4_l2t.h"
73 #include "t4_mp_ring.h"
74 
75 #ifdef T4_PKT_TIMESTAMP
76 #define RX_COPY_THRESHOLD (MINCLSIZE - 8)
77 #else
78 #define RX_COPY_THRESHOLD MINCLSIZE
79 #endif
80 
81 /*
82  * Ethernet frames are DMA'd at this byte offset into the freelist buffer.
83  * 0-7 are valid values.
84  */
85 static int fl_pktshift = 2;
86 TUNABLE_INT("hw.cxgbe.fl_pktshift", &fl_pktshift);
87 
88 /*
89  * Pad ethernet payload up to this boundary.
90  * -1: driver should figure out a good value.
91  *  0: disable padding.
92  *  Any power of 2 from 32 to 4096 (both inclusive) is also a valid value.
93  */
94 int fl_pad = -1;
95 TUNABLE_INT("hw.cxgbe.fl_pad", &fl_pad);
96 
97 /*
98  * Status page length.
99  * -1: driver should figure out a good value.
100  *  64 or 128 are the only other valid values.
101  */
102 static int spg_len = -1;
103 TUNABLE_INT("hw.cxgbe.spg_len", &spg_len);
104 
105 /*
106  * Congestion drops.
107  * -1: no congestion feedback (not recommended).
108  *  0: backpressure the channel instead of dropping packets right away.
109  *  1: no backpressure, drop packets for the congested queue immediately.
110  */
111 static int cong_drop = 0;
112 TUNABLE_INT("hw.cxgbe.cong_drop", &cong_drop);
113 
114 /*
115  * Deliver multiple frames in the same free list buffer if they fit.
116  * -1: let the driver decide whether to enable buffer packing or not.
117  *  0: disable buffer packing.
118  *  1: enable buffer packing.
119  */
120 static int buffer_packing = -1;
121 TUNABLE_INT("hw.cxgbe.buffer_packing", &buffer_packing);
122 
123 /*
124  * Start next frame in a packed buffer at this boundary.
125  * -1: driver should figure out a good value.
126  * T4: driver will ignore this and use the same value as fl_pad above.
127  * T5: 16, or a power of 2 from 64 to 4096 (both inclusive) is a valid value.
128  */
129 static int fl_pack = -1;
130 TUNABLE_INT("hw.cxgbe.fl_pack", &fl_pack);
131 
132 /*
133  * Allow the driver to create mbuf(s) in a cluster allocated for rx.
134  * 0: never; always allocate mbufs from the zone_mbuf UMA zone.
135  * 1: ok to create mbuf(s) within a cluster if there is room.
136  */
137 static int allow_mbufs_in_cluster = 1;
138 TUNABLE_INT("hw.cxgbe.allow_mbufs_in_cluster", &allow_mbufs_in_cluster);
139 
140 /*
141  * Largest rx cluster size that the driver is allowed to allocate.
142  */
143 static int largest_rx_cluster = MJUM16BYTES;
144 TUNABLE_INT("hw.cxgbe.largest_rx_cluster", &largest_rx_cluster);
145 
146 /*
147  * Size of cluster allocation that's most likely to succeed.  The driver will
148  * fall back to this size if it fails to allocate clusters larger than this.
149  */
150 static int safest_rx_cluster = PAGE_SIZE;
151 TUNABLE_INT("hw.cxgbe.safest_rx_cluster", &safest_rx_cluster);
152 
153 /*
154  * The interrupt holdoff timers are multiplied by this value on T6+.
155  * 1 and 3-17 (both inclusive) are legal values.
156  */
157 static int tscale = 1;
158 TUNABLE_INT("hw.cxgbe.tscale", &tscale);
159 
160 /*
161  * Number of LRO entries in the lro_ctrl structure per rx queue.
162  */
163 static int lro_entries = TCP_LRO_ENTRIES;
164 TUNABLE_INT("hw.cxgbe.lro_entries", &lro_entries);
165 
166 /*
167  * This enables presorting of frames before they're fed into tcp_lro_rx.
168  */
169 static int lro_mbufs = 0;
170 TUNABLE_INT("hw.cxgbe.lro_mbufs", &lro_mbufs);
171 
172 struct txpkts {
173 	u_int wr_type;		/* type 0 or type 1 */
174 	u_int npkt;		/* # of packets in this work request */
175 	u_int plen;		/* total payload (sum of all packets) */
176 	u_int len16;		/* # of 16B pieces used by this work request */
177 };
178 
179 /* A packet's SGL.  This + m_pkthdr has all info needed for tx */
180 struct sgl {
181 	struct sglist sg;
182 	struct sglist_seg seg[TX_SGL_SEGS];
183 };
184 
185 static int service_iq(struct sge_iq *, int);
186 static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t);
187 static int t4_eth_rx(struct sge_iq *, const struct rss_header *, struct mbuf *);
188 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int);
189 static inline void init_fl(struct adapter *, struct sge_fl *, int, int, char *);
190 static inline void init_eq(struct adapter *, struct sge_eq *, int, int, uint8_t,
191     uint16_t, char *);
192 static int alloc_ring(struct adapter *, size_t, bus_dma_tag_t *, bus_dmamap_t *,
193     bus_addr_t *, void **);
194 static int free_ring(struct adapter *, bus_dma_tag_t, bus_dmamap_t, bus_addr_t,
195     void *);
196 static int alloc_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *,
197     int, int);
198 static int free_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *);
199 static void add_fl_sysctls(struct adapter *, struct sysctl_ctx_list *,
200     struct sysctl_oid *, struct sge_fl *);
201 static int alloc_fwq(struct adapter *);
202 static int free_fwq(struct adapter *);
203 static int alloc_mgmtq(struct adapter *);
204 static int free_mgmtq(struct adapter *);
205 static int alloc_rxq(struct vi_info *, struct sge_rxq *, int, int,
206     struct sysctl_oid *);
207 static int free_rxq(struct vi_info *, struct sge_rxq *);
208 #ifdef TCP_OFFLOAD
209 static int alloc_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *, int, int,
210     struct sysctl_oid *);
211 static int free_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *);
212 #endif
213 #ifdef DEV_NETMAP
214 static int alloc_nm_rxq(struct vi_info *, struct sge_nm_rxq *, int, int,
215     struct sysctl_oid *);
216 static int free_nm_rxq(struct vi_info *, struct sge_nm_rxq *);
217 static int alloc_nm_txq(struct vi_info *, struct sge_nm_txq *, int, int,
218     struct sysctl_oid *);
219 static int free_nm_txq(struct vi_info *, struct sge_nm_txq *);
220 #endif
221 static int ctrl_eq_alloc(struct adapter *, struct sge_eq *);
222 static int eth_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
223 #ifdef TCP_OFFLOAD
224 static int ofld_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
225 #endif
226 static int alloc_eq(struct adapter *, struct vi_info *, struct sge_eq *);
227 static int free_eq(struct adapter *, struct sge_eq *);
228 static int alloc_wrq(struct adapter *, struct vi_info *, struct sge_wrq *,
229     struct sysctl_oid *);
230 static int free_wrq(struct adapter *, struct sge_wrq *);
231 static int alloc_txq(struct vi_info *, struct sge_txq *, int,
232     struct sysctl_oid *);
233 static int free_txq(struct vi_info *, struct sge_txq *);
234 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int);
235 static inline void ring_fl_db(struct adapter *, struct sge_fl *);
236 static int refill_fl(struct adapter *, struct sge_fl *, int);
237 static void refill_sfl(void *);
238 static int alloc_fl_sdesc(struct sge_fl *);
239 static void free_fl_sdesc(struct adapter *, struct sge_fl *);
240 static void find_best_refill_source(struct adapter *, struct sge_fl *, int);
241 static void find_safe_refill_source(struct adapter *, struct sge_fl *);
242 static void add_fl_to_sfl(struct adapter *, struct sge_fl *);
243 
244 static inline void get_pkt_gl(struct mbuf *, struct sglist *);
245 static inline u_int txpkt_len16(u_int, u_int);
246 static inline u_int txpkt_vm_len16(u_int, u_int);
247 static inline u_int txpkts0_len16(u_int);
248 static inline u_int txpkts1_len16(void);
249 static u_int write_txpkt_wr(struct sge_txq *, struct fw_eth_tx_pkt_wr *,
250     struct mbuf *, u_int);
251 static u_int write_txpkt_vm_wr(struct adapter *, struct sge_txq *,
252     struct fw_eth_tx_pkt_vm_wr *, struct mbuf *, u_int);
253 static int try_txpkts(struct mbuf *, struct mbuf *, struct txpkts *, u_int);
254 static int add_to_txpkts(struct mbuf *, struct txpkts *, u_int);
255 static u_int write_txpkts_wr(struct sge_txq *, struct fw_eth_tx_pkts_wr *,
256     struct mbuf *, const struct txpkts *, u_int);
257 static void write_gl_to_txd(struct sge_txq *, struct mbuf *, caddr_t *, int);
258 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int);
259 static inline void ring_eq_db(struct adapter *, struct sge_eq *, u_int);
260 static inline uint16_t read_hw_cidx(struct sge_eq *);
261 static inline u_int reclaimable_tx_desc(struct sge_eq *);
262 static inline u_int total_available_tx_desc(struct sge_eq *);
263 static u_int reclaim_tx_descs(struct sge_txq *, u_int);
264 static void tx_reclaim(void *, int);
265 static __be64 get_flit(struct sglist_seg *, int, int);
266 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
267     struct mbuf *);
268 static int handle_fw_msg(struct sge_iq *, const struct rss_header *,
269     struct mbuf *);
270 static int t4_handle_wrerr_rpl(struct adapter *, const __be64 *);
271 static void wrq_tx_drain(void *, int);
272 static void drain_wrq_wr_list(struct adapter *, struct sge_wrq *);
273 
274 static int sysctl_uint16(SYSCTL_HANDLER_ARGS);
275 static int sysctl_bufsizes(SYSCTL_HANDLER_ARGS);
276 static int sysctl_tc(SYSCTL_HANDLER_ARGS);
277 
278 static counter_u64_t extfree_refs;
279 static counter_u64_t extfree_rels;
280 
281 an_handler_t t4_an_handler;
282 fw_msg_handler_t t4_fw_msg_handler[NUM_FW6_TYPES];
283 cpl_handler_t t4_cpl_handler[NUM_CPL_CMDS];
284 
285 
286 static int
287 an_not_handled(struct sge_iq *iq, const struct rsp_ctrl *ctrl)
288 {
289 
290 #ifdef INVARIANTS
291 	panic("%s: async notification on iq %p (ctrl %p)", __func__, iq, ctrl);
292 #else
293 	log(LOG_ERR, "%s: async notification on iq %p (ctrl %p)\n",
294 	    __func__, iq, ctrl);
295 #endif
296 	return (EDOOFUS);
297 }
298 
299 int
300 t4_register_an_handler(an_handler_t h)
301 {
302 	uintptr_t *loc, new;
303 
304 	new = h ? (uintptr_t)h : (uintptr_t)an_not_handled;
305 	loc = (uintptr_t *) &t4_an_handler;
306 	atomic_store_rel_ptr(loc, new);
307 
308 	return (0);
309 }
310 
311 static int
312 fw_msg_not_handled(struct adapter *sc, const __be64 *rpl)
313 {
314 	const struct cpl_fw6_msg *cpl =
315 	    __containerof(rpl, struct cpl_fw6_msg, data[0]);
316 
317 #ifdef INVARIANTS
318 	panic("%s: fw_msg type %d", __func__, cpl->type);
319 #else
320 	log(LOG_ERR, "%s: fw_msg type %d\n", __func__, cpl->type);
321 #endif
322 	return (EDOOFUS);
323 }
324 
325 int
326 t4_register_fw_msg_handler(int type, fw_msg_handler_t h)
327 {
328 	uintptr_t *loc, new;
329 
330 	if (type >= nitems(t4_fw_msg_handler))
331 		return (EINVAL);
332 
333 	/*
334 	 * These are dispatched by the handler for FW{4|6}_CPL_MSG using the CPL
335 	 * handler dispatch table.  Reject any attempt to install a handler for
336 	 * this subtype.
337 	 */
338 	if (type == FW_TYPE_RSSCPL || type == FW6_TYPE_RSSCPL)
339 		return (EINVAL);
340 
341 	new = h ? (uintptr_t)h : (uintptr_t)fw_msg_not_handled;
342 	loc = (uintptr_t *) &t4_fw_msg_handler[type];
343 	atomic_store_rel_ptr(loc, new);
344 
345 	return (0);
346 }
347 
348 static int
349 cpl_not_handled(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
350 {
351 
352 #ifdef INVARIANTS
353 	panic("%s: opcode 0x%02x on iq %p with payload %p",
354 	    __func__, rss->opcode, iq, m);
355 #else
356 	log(LOG_ERR, "%s: opcode 0x%02x on iq %p with payload %p\n",
357 	    __func__, rss->opcode, iq, m);
358 	m_freem(m);
359 #endif
360 	return (EDOOFUS);
361 }
362 
363 int
364 t4_register_cpl_handler(int opcode, cpl_handler_t h)
365 {
366 	uintptr_t *loc, new;
367 
368 	if (opcode >= nitems(t4_cpl_handler))
369 		return (EINVAL);
370 
371 	new = h ? (uintptr_t)h : (uintptr_t)cpl_not_handled;
372 	loc = (uintptr_t *) &t4_cpl_handler[opcode];
373 	atomic_store_rel_ptr(loc, new);
374 
375 	return (0);
376 }
377 
378 /*
379  * Called on MOD_LOAD.  Validates and calculates the SGE tunables.
380  */
381 void
382 t4_sge_modload(void)
383 {
384 	int i;
385 
386 	if (fl_pktshift < 0 || fl_pktshift > 7) {
387 		printf("Invalid hw.cxgbe.fl_pktshift value (%d),"
388 		    " using 2 instead.\n", fl_pktshift);
389 		fl_pktshift = 2;
390 	}
391 
392 	if (spg_len != 64 && spg_len != 128) {
393 		int len;
394 
395 #if defined(__i386__) || defined(__amd64__)
396 		len = cpu_clflush_line_size > 64 ? 128 : 64;
397 #else
398 		len = 64;
399 #endif
400 		if (spg_len != -1) {
401 			printf("Invalid hw.cxgbe.spg_len value (%d),"
402 			    " using %d instead.\n", spg_len, len);
403 		}
404 		spg_len = len;
405 	}
406 
407 	if (cong_drop < -1 || cong_drop > 1) {
408 		printf("Invalid hw.cxgbe.cong_drop value (%d),"
409 		    " using 0 instead.\n", cong_drop);
410 		cong_drop = 0;
411 	}
412 
413 	if (tscale != 1 && (tscale < 3 || tscale > 17)) {
414 		printf("Invalid hw.cxgbe.tscale value (%d),"
415 		    " using 1 instead.\n", tscale);
416 		tscale = 1;
417 	}
418 
419 	extfree_refs = counter_u64_alloc(M_WAITOK);
420 	extfree_rels = counter_u64_alloc(M_WAITOK);
421 	counter_u64_zero(extfree_refs);
422 	counter_u64_zero(extfree_rels);
423 
424 	t4_an_handler = an_not_handled;
425 	for (i = 0; i < nitems(t4_fw_msg_handler); i++)
426 		t4_fw_msg_handler[i] = fw_msg_not_handled;
427 	for (i = 0; i < nitems(t4_cpl_handler); i++)
428 		t4_cpl_handler[i] = cpl_not_handled;
429 
430 	t4_register_cpl_handler(CPL_FW4_MSG, handle_fw_msg);
431 	t4_register_cpl_handler(CPL_FW6_MSG, handle_fw_msg);
432 	t4_register_cpl_handler(CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
433 	t4_register_cpl_handler(CPL_RX_PKT, t4_eth_rx);
434 	t4_register_fw_msg_handler(FW6_TYPE_CMD_RPL, t4_handle_fw_rpl);
435 	t4_register_fw_msg_handler(FW6_TYPE_WRERR_RPL, t4_handle_wrerr_rpl);
436 }
437 
438 void
439 t4_sge_modunload(void)
440 {
441 
442 	counter_u64_free(extfree_refs);
443 	counter_u64_free(extfree_rels);
444 }
445 
446 uint64_t
447 t4_sge_extfree_refs(void)
448 {
449 	uint64_t refs, rels;
450 
451 	rels = counter_u64_fetch(extfree_rels);
452 	refs = counter_u64_fetch(extfree_refs);
453 
454 	return (refs - rels);
455 }
456 
457 static inline void
458 setup_pad_and_pack_boundaries(struct adapter *sc)
459 {
460 	uint32_t v, m;
461 	int pad, pack, pad_shift;
462 
463 	pad_shift = chip_id(sc) > CHELSIO_T5 ? X_T6_INGPADBOUNDARY_SHIFT :
464 	    X_INGPADBOUNDARY_SHIFT;
465 	pad = fl_pad;
466 	if (fl_pad < (1 << pad_shift) ||
467 	    fl_pad > (1 << (pad_shift + M_INGPADBOUNDARY)) ||
468 	    !powerof2(fl_pad)) {
469 		/*
470 		 * If there is any chance that we might use buffer packing and
471 		 * the chip is a T4, then pick 64 as the pad/pack boundary.  Set
472 		 * it to the minimum allowed in all other cases.
473 		 */
474 		pad = is_t4(sc) && buffer_packing ? 64 : 1 << pad_shift;
475 
476 		/*
477 		 * For fl_pad = 0 we'll still write a reasonable value to the
478 		 * register but all the freelists will opt out of padding.
479 		 * We'll complain here only if the user tried to set it to a
480 		 * value greater than 0 that was invalid.
481 		 */
482 		if (fl_pad > 0) {
483 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pad value"
484 			    " (%d), using %d instead.\n", fl_pad, pad);
485 		}
486 	}
487 	m = V_INGPADBOUNDARY(M_INGPADBOUNDARY);
488 	v = V_INGPADBOUNDARY(ilog2(pad) - pad_shift);
489 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
490 
491 	if (is_t4(sc)) {
492 		if (fl_pack != -1 && fl_pack != pad) {
493 			/* Complain but carry on. */
494 			device_printf(sc->dev, "hw.cxgbe.fl_pack (%d) ignored,"
495 			    " using %d instead.\n", fl_pack, pad);
496 		}
497 		return;
498 	}
499 
500 	pack = fl_pack;
501 	if (fl_pack < 16 || fl_pack == 32 || fl_pack > 4096 ||
502 	    !powerof2(fl_pack)) {
503 		pack = max(sc->params.pci.mps, CACHE_LINE_SIZE);
504 		MPASS(powerof2(pack));
505 		if (pack < 16)
506 			pack = 16;
507 		if (pack == 32)
508 			pack = 64;
509 		if (pack > 4096)
510 			pack = 4096;
511 		if (fl_pack != -1) {
512 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pack value"
513 			    " (%d), using %d instead.\n", fl_pack, pack);
514 		}
515 	}
516 	m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY);
517 	if (pack == 16)
518 		v = V_INGPACKBOUNDARY(0);
519 	else
520 		v = V_INGPACKBOUNDARY(ilog2(pack) - 5);
521 
522 	MPASS(!is_t4(sc));	/* T4 doesn't have SGE_CONTROL2 */
523 	t4_set_reg_field(sc, A_SGE_CONTROL2, m, v);
524 }
525 
526 /*
527  * adap->params.vpd.cclk must be set up before this is called.
528  */
529 void
530 t4_tweak_chip_settings(struct adapter *sc)
531 {
532 	int i;
533 	uint32_t v, m;
534 	int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200};
535 	int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk;
536 	int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */
537 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
538 	static int sge_flbuf_sizes[] = {
539 		MCLBYTES,
540 #if MJUMPAGESIZE != MCLBYTES
541 		MJUMPAGESIZE,
542 		MJUMPAGESIZE - CL_METADATA_SIZE,
543 		MJUMPAGESIZE - 2 * MSIZE - CL_METADATA_SIZE,
544 #endif
545 		MJUM9BYTES,
546 		MJUM16BYTES,
547 		MCLBYTES - MSIZE - CL_METADATA_SIZE,
548 		MJUM9BYTES - CL_METADATA_SIZE,
549 		MJUM16BYTES - CL_METADATA_SIZE,
550 	};
551 
552 	KASSERT(sc->flags & MASTER_PF,
553 	    ("%s: trying to change chip settings when not master.", __func__));
554 
555 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
556 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
557 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
558 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
559 
560 	setup_pad_and_pack_boundaries(sc);
561 
562 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
563 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
564 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
565 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
566 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
567 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
568 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
569 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
570 	t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v);
571 
572 	KASSERT(nitems(sge_flbuf_sizes) <= SGE_FLBUF_SIZES,
573 	    ("%s: hw buffer size table too big", __func__));
574 	for (i = 0; i < min(nitems(sge_flbuf_sizes), SGE_FLBUF_SIZES); i++) {
575 		t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i),
576 		    sge_flbuf_sizes[i]);
577 	}
578 
579 	v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) |
580 	    V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]);
581 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v);
582 
583 	KASSERT(intr_timer[0] <= timer_max,
584 	    ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0],
585 	    timer_max));
586 	for (i = 1; i < nitems(intr_timer); i++) {
587 		KASSERT(intr_timer[i] >= intr_timer[i - 1],
588 		    ("%s: timers not listed in increasing order (%d)",
589 		    __func__, i));
590 
591 		while (intr_timer[i] > timer_max) {
592 			if (i == nitems(intr_timer) - 1) {
593 				intr_timer[i] = timer_max;
594 				break;
595 			}
596 			intr_timer[i] += intr_timer[i - 1];
597 			intr_timer[i] /= 2;
598 		}
599 	}
600 
601 	v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) |
602 	    V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1]));
603 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v);
604 	v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) |
605 	    V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3]));
606 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v);
607 	v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) |
608 	    V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5]));
609 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v);
610 
611 	if (chip_id(sc) >= CHELSIO_T6) {
612 		m = V_TSCALE(M_TSCALE);
613 		if (tscale == 1)
614 			v = 0;
615 		else
616 			v = V_TSCALE(tscale - 2);
617 		t4_set_reg_field(sc, A_SGE_ITP_CONTROL, m, v);
618 
619 		if (sc->debug_flags & DF_DISABLE_TCB_CACHE) {
620 			m = V_RDTHRESHOLD(M_RDTHRESHOLD) | F_WRTHRTHRESHEN |
621 			    V_WRTHRTHRESH(M_WRTHRTHRESH);
622 			t4_tp_pio_read(sc, &v, 1, A_TP_CMM_CONFIG, 1);
623 			v &= ~m;
624 			v |= V_RDTHRESHOLD(1) | F_WRTHRTHRESHEN |
625 			    V_WRTHRTHRESH(16);
626 			t4_tp_pio_write(sc, &v, 1, A_TP_CMM_CONFIG, 1);
627 		}
628 	}
629 
630 	/* 4K, 16K, 64K, 256K DDP "page sizes" for TDDP */
631 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
632 	t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v);
633 
634 	/*
635 	 * 4K, 8K, 16K, 64K DDP "page sizes" for iSCSI DDP.  These have been
636 	 * chosen with MAXPHYS = 128K in mind.  The largest DDP buffer that we
637 	 * may have to deal with is MAXPHYS + 1 page.
638 	 */
639 	v = V_HPZ0(0) | V_HPZ1(1) | V_HPZ2(2) | V_HPZ3(4);
640 	t4_write_reg(sc, A_ULP_RX_ISCSI_PSZ, v);
641 
642 	/* We use multiple DDP page sizes both in plain-TOE and ISCSI modes. */
643 	m = v = F_TDDPTAGTCB | F_ISCSITAGTCB;
644 	t4_set_reg_field(sc, A_ULP_RX_CTL, m, v);
645 
646 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
647 	    F_RESETDDPOFFSET;
648 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
649 	t4_set_reg_field(sc, A_TP_PARA_REG5, m, v);
650 }
651 
652 /*
653  * SGE wants the buffer to be at least 64B and then a multiple of 16.  If
654  * padding is in use, the buffer's start and end need to be aligned to the pad
655  * boundary as well.  We'll just make sure that the size is a multiple of the
656  * boundary here, it is up to the buffer allocation code to make sure the start
657  * of the buffer is aligned as well.
658  */
659 static inline int
660 hwsz_ok(struct adapter *sc, int hwsz)
661 {
662 	int mask = fl_pad ? sc->params.sge.pad_boundary - 1 : 16 - 1;
663 
664 	return (hwsz >= 64 && (hwsz & mask) == 0);
665 }
666 
667 /*
668  * XXX: driver really should be able to deal with unexpected settings.
669  */
670 int
671 t4_read_chip_settings(struct adapter *sc)
672 {
673 	struct sge *s = &sc->sge;
674 	struct sge_params *sp = &sc->params.sge;
675 	int i, j, n, rc = 0;
676 	uint32_t m, v, r;
677 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
678 	static int sw_buf_sizes[] = {	/* Sorted by size */
679 		MCLBYTES,
680 #if MJUMPAGESIZE != MCLBYTES
681 		MJUMPAGESIZE,
682 #endif
683 		MJUM9BYTES,
684 		MJUM16BYTES
685 	};
686 	struct sw_zone_info *swz, *safe_swz;
687 	struct hw_buf_info *hwb;
688 
689 	m = F_RXPKTCPLMODE;
690 	v = F_RXPKTCPLMODE;
691 	r = sc->params.sge.sge_control;
692 	if ((r & m) != v) {
693 		device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r);
694 		rc = EINVAL;
695 	}
696 
697 	/*
698 	 * If this changes then every single use of PAGE_SHIFT in the driver
699 	 * needs to be carefully reviewed for PAGE_SHIFT vs sp->page_shift.
700 	 */
701 	if (sp->page_shift != PAGE_SHIFT) {
702 		device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r);
703 		rc = EINVAL;
704 	}
705 
706 	/* Filter out unusable hw buffer sizes entirely (mark with -2). */
707 	hwb = &s->hw_buf_info[0];
708 	for (i = 0; i < nitems(s->hw_buf_info); i++, hwb++) {
709 		r = sc->params.sge.sge_fl_buffer_size[i];
710 		hwb->size = r;
711 		hwb->zidx = hwsz_ok(sc, r) ? -1 : -2;
712 		hwb->next = -1;
713 	}
714 
715 	/*
716 	 * Create a sorted list in decreasing order of hw buffer sizes (and so
717 	 * increasing order of spare area) for each software zone.
718 	 *
719 	 * If padding is enabled then the start and end of the buffer must align
720 	 * to the pad boundary; if packing is enabled then they must align with
721 	 * the pack boundary as well.  Allocations from the cluster zones are
722 	 * aligned to min(size, 4K), so the buffer starts at that alignment and
723 	 * ends at hwb->size alignment.  If mbuf inlining is allowed the
724 	 * starting alignment will be reduced to MSIZE and the driver will
725 	 * exercise appropriate caution when deciding on the best buffer layout
726 	 * to use.
727 	 */
728 	n = 0;	/* no usable buffer size to begin with */
729 	swz = &s->sw_zone_info[0];
730 	safe_swz = NULL;
731 	for (i = 0; i < SW_ZONE_SIZES; i++, swz++) {
732 		int8_t head = -1, tail = -1;
733 
734 		swz->size = sw_buf_sizes[i];
735 		swz->zone = m_getzone(swz->size);
736 		swz->type = m_gettype(swz->size);
737 
738 		if (swz->size < PAGE_SIZE) {
739 			MPASS(powerof2(swz->size));
740 			if (fl_pad && (swz->size % sp->pad_boundary != 0))
741 				continue;
742 		}
743 
744 		if (swz->size == safest_rx_cluster)
745 			safe_swz = swz;
746 
747 		hwb = &s->hw_buf_info[0];
748 		for (j = 0; j < SGE_FLBUF_SIZES; j++, hwb++) {
749 			if (hwb->zidx != -1 || hwb->size > swz->size)
750 				continue;
751 #ifdef INVARIANTS
752 			if (fl_pad)
753 				MPASS(hwb->size % sp->pad_boundary == 0);
754 #endif
755 			hwb->zidx = i;
756 			if (head == -1)
757 				head = tail = j;
758 			else if (hwb->size < s->hw_buf_info[tail].size) {
759 				s->hw_buf_info[tail].next = j;
760 				tail = j;
761 			} else {
762 				int8_t *cur;
763 				struct hw_buf_info *t;
764 
765 				for (cur = &head; *cur != -1; cur = &t->next) {
766 					t = &s->hw_buf_info[*cur];
767 					if (hwb->size == t->size) {
768 						hwb->zidx = -2;
769 						break;
770 					}
771 					if (hwb->size > t->size) {
772 						hwb->next = *cur;
773 						*cur = j;
774 						break;
775 					}
776 				}
777 			}
778 		}
779 		swz->head_hwidx = head;
780 		swz->tail_hwidx = tail;
781 
782 		if (tail != -1) {
783 			n++;
784 			if (swz->size - s->hw_buf_info[tail].size >=
785 			    CL_METADATA_SIZE)
786 				sc->flags |= BUF_PACKING_OK;
787 		}
788 	}
789 	if (n == 0) {
790 		device_printf(sc->dev, "no usable SGE FL buffer size.\n");
791 		rc = EINVAL;
792 	}
793 
794 	s->safe_hwidx1 = -1;
795 	s->safe_hwidx2 = -1;
796 	if (safe_swz != NULL) {
797 		s->safe_hwidx1 = safe_swz->head_hwidx;
798 		for (i = safe_swz->head_hwidx; i != -1; i = hwb->next) {
799 			int spare;
800 
801 			hwb = &s->hw_buf_info[i];
802 #ifdef INVARIANTS
803 			if (fl_pad)
804 				MPASS(hwb->size % sp->pad_boundary == 0);
805 #endif
806 			spare = safe_swz->size - hwb->size;
807 			if (spare >= CL_METADATA_SIZE) {
808 				s->safe_hwidx2 = i;
809 				break;
810 			}
811 		}
812 	}
813 
814 	if (sc->flags & IS_VF)
815 		return (0);
816 
817 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
818 	r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ);
819 	if (r != v) {
820 		device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r);
821 		rc = EINVAL;
822 	}
823 
824 	m = v = F_TDDPTAGTCB;
825 	r = t4_read_reg(sc, A_ULP_RX_CTL);
826 	if ((r & m) != v) {
827 		device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r);
828 		rc = EINVAL;
829 	}
830 
831 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
832 	    F_RESETDDPOFFSET;
833 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
834 	r = t4_read_reg(sc, A_TP_PARA_REG5);
835 	if ((r & m) != v) {
836 		device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r);
837 		rc = EINVAL;
838 	}
839 
840 	t4_init_tp_params(sc, 1);
841 
842 	t4_read_mtu_tbl(sc, sc->params.mtus, NULL);
843 	t4_load_mtus(sc, sc->params.mtus, sc->params.a_wnd, sc->params.b_wnd);
844 
845 	return (rc);
846 }
847 
848 int
849 t4_create_dma_tag(struct adapter *sc)
850 {
851 	int rc;
852 
853 	rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
854 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
855 	    BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
856 	    NULL, &sc->dmat);
857 	if (rc != 0) {
858 		device_printf(sc->dev,
859 		    "failed to create main DMA tag: %d\n", rc);
860 	}
861 
862 	return (rc);
863 }
864 
865 void
866 t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
867     struct sysctl_oid_list *children)
868 {
869 	struct sge_params *sp = &sc->params.sge;
870 
871 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "buffer_sizes",
872 	    CTLTYPE_STRING | CTLFLAG_RD, &sc->sge, 0, sysctl_bufsizes, "A",
873 	    "freelist buffer sizes");
874 
875 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD,
876 	    NULL, sp->fl_pktshift, "payload DMA offset in rx buffer (bytes)");
877 
878 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD,
879 	    NULL, sp->pad_boundary, "payload pad boundary (bytes)");
880 
881 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD,
882 	    NULL, sp->spg_len, "status page size (bytes)");
883 
884 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD,
885 	    NULL, cong_drop, "congestion drop setting");
886 
887 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD,
888 	    NULL, sp->pack_boundary, "payload pack boundary (bytes)");
889 }
890 
891 int
892 t4_destroy_dma_tag(struct adapter *sc)
893 {
894 	if (sc->dmat)
895 		bus_dma_tag_destroy(sc->dmat);
896 
897 	return (0);
898 }
899 
900 /*
901  * Allocate and initialize the firmware event queue and the management queue.
902  *
903  * Returns errno on failure.  Resources allocated up to that point may still be
904  * allocated.  Caller is responsible for cleanup in case this function fails.
905  */
906 int
907 t4_setup_adapter_queues(struct adapter *sc)
908 {
909 	int rc;
910 
911 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
912 
913 	sysctl_ctx_init(&sc->ctx);
914 	sc->flags |= ADAP_SYSCTL_CTX;
915 
916 	/*
917 	 * Firmware event queue
918 	 */
919 	rc = alloc_fwq(sc);
920 	if (rc != 0)
921 		return (rc);
922 
923 	/*
924 	 * Management queue.  This is just a control queue that uses the fwq as
925 	 * its associated iq.
926 	 */
927 	if (!(sc->flags & IS_VF))
928 		rc = alloc_mgmtq(sc);
929 
930 	return (rc);
931 }
932 
933 /*
934  * Idempotent
935  */
936 int
937 t4_teardown_adapter_queues(struct adapter *sc)
938 {
939 
940 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
941 
942 	/* Do this before freeing the queue */
943 	if (sc->flags & ADAP_SYSCTL_CTX) {
944 		sysctl_ctx_free(&sc->ctx);
945 		sc->flags &= ~ADAP_SYSCTL_CTX;
946 	}
947 
948 	free_mgmtq(sc);
949 	free_fwq(sc);
950 
951 	return (0);
952 }
953 
954 static inline int
955 first_vector(struct vi_info *vi)
956 {
957 	struct adapter *sc = vi->pi->adapter;
958 
959 	if (sc->intr_count == 1)
960 		return (0);
961 
962 	return (vi->first_intr);
963 }
964 
965 /*
966  * Given an arbitrary "index," come up with an iq that can be used by other
967  * queues (of this VI) for interrupt forwarding, SGE egress updates, etc.
968  * The iq returned is guaranteed to be something that takes direct interrupts.
969  */
970 static struct sge_iq *
971 vi_intr_iq(struct vi_info *vi, int idx)
972 {
973 	struct adapter *sc = vi->pi->adapter;
974 	struct sge *s = &sc->sge;
975 	struct sge_iq *iq = NULL;
976 	int nintr, i;
977 
978 	if (sc->intr_count == 1)
979 		return (&sc->sge.fwq);
980 
981 	nintr = vi->nintr;
982 #ifdef DEV_NETMAP
983 	/* Do not consider any netmap-only interrupts */
984 	if (vi->flags & INTR_RXQ && vi->nnmrxq > vi->nrxq)
985 		nintr -= vi->nnmrxq - vi->nrxq;
986 #endif
987 	KASSERT(nintr != 0,
988 	    ("%s: vi %p has no exclusive interrupts, total interrupts = %d",
989 	    __func__, vi, sc->intr_count));
990 	i = idx % nintr;
991 
992 	if (vi->flags & INTR_RXQ) {
993 	       	if (i < vi->nrxq) {
994 			iq = &s->rxq[vi->first_rxq + i].iq;
995 			goto done;
996 		}
997 		i -= vi->nrxq;
998 	}
999 #ifdef TCP_OFFLOAD
1000 	if (vi->flags & INTR_OFLD_RXQ) {
1001 	       	if (i < vi->nofldrxq) {
1002 			iq = &s->ofld_rxq[vi->first_ofld_rxq + i].iq;
1003 			goto done;
1004 		}
1005 		i -= vi->nofldrxq;
1006 	}
1007 #endif
1008 	panic("%s: vi %p, intr_flags 0x%lx, idx %d, total intr %d\n", __func__,
1009 	    vi, vi->flags & INTR_ALL, idx, nintr);
1010 done:
1011 	MPASS(iq != NULL);
1012 	KASSERT(iq->flags & IQ_INTR,
1013 	    ("%s: iq %p (vi %p, intr_flags 0x%lx, idx %d)", __func__, iq, vi,
1014 	    vi->flags & INTR_ALL, idx));
1015 	return (iq);
1016 }
1017 
1018 /* Maximum payload that can be delivered with a single iq descriptor */
1019 static inline int
1020 mtu_to_max_payload(struct adapter *sc, int mtu, const int toe)
1021 {
1022 	int payload;
1023 
1024 #ifdef TCP_OFFLOAD
1025 	if (toe) {
1026 		payload = sc->tt.rx_coalesce ?
1027 		    G_RXCOALESCESIZE(t4_read_reg(sc, A_TP_PARA_REG2)) : mtu;
1028 	} else {
1029 #endif
1030 		/* large enough even when hw VLAN extraction is disabled */
1031 		payload = sc->params.sge.fl_pktshift + ETHER_HDR_LEN +
1032 		    ETHER_VLAN_ENCAP_LEN + mtu;
1033 #ifdef TCP_OFFLOAD
1034 	}
1035 #endif
1036 
1037 	return (payload);
1038 }
1039 
1040 int
1041 t4_setup_vi_queues(struct vi_info *vi)
1042 {
1043 	int rc = 0, i, j, intr_idx, iqid;
1044 	struct sge_rxq *rxq;
1045 	struct sge_txq *txq;
1046 	struct sge_wrq *ctrlq;
1047 #ifdef TCP_OFFLOAD
1048 	struct sge_ofld_rxq *ofld_rxq;
1049 	struct sge_wrq *ofld_txq;
1050 #endif
1051 #ifdef DEV_NETMAP
1052 	int saved_idx;
1053 	struct sge_nm_rxq *nm_rxq;
1054 	struct sge_nm_txq *nm_txq;
1055 #endif
1056 	char name[16];
1057 	struct port_info *pi = vi->pi;
1058 	struct adapter *sc = pi->adapter;
1059 	struct ifnet *ifp = vi->ifp;
1060 	struct sysctl_oid *oid = device_get_sysctl_tree(vi->dev);
1061 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
1062 	int maxp, mtu = ifp->if_mtu;
1063 
1064 	/* Interrupt vector to start from (when using multiple vectors) */
1065 	intr_idx = first_vector(vi);
1066 
1067 #ifdef DEV_NETMAP
1068 	saved_idx = intr_idx;
1069 	if (ifp->if_capabilities & IFCAP_NETMAP) {
1070 
1071 		/* netmap is supported with direct interrupts only. */
1072 		MPASS(vi->flags & INTR_RXQ);
1073 
1074 		/*
1075 		 * We don't have buffers to back the netmap rx queues
1076 		 * right now so we create the queues in a way that
1077 		 * doesn't set off any congestion signal in the chip.
1078 		 */
1079 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_rxq",
1080 		    CTLFLAG_RD, NULL, "rx queues");
1081 		for_each_nm_rxq(vi, i, nm_rxq) {
1082 			rc = alloc_nm_rxq(vi, nm_rxq, intr_idx, i, oid);
1083 			if (rc != 0)
1084 				goto done;
1085 			intr_idx++;
1086 		}
1087 
1088 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_txq",
1089 		    CTLFLAG_RD, NULL, "tx queues");
1090 		for_each_nm_txq(vi, i, nm_txq) {
1091 			iqid = vi->first_nm_rxq + (i % vi->nnmrxq);
1092 			rc = alloc_nm_txq(vi, nm_txq, iqid, i, oid);
1093 			if (rc != 0)
1094 				goto done;
1095 		}
1096 	}
1097 
1098 	/* Normal rx queues and netmap rx queues share the same interrupts. */
1099 	intr_idx = saved_idx;
1100 #endif
1101 
1102 	/*
1103 	 * First pass over all NIC and TOE rx queues:
1104 	 * a) initialize iq and fl
1105 	 * b) allocate queue iff it will take direct interrupts.
1106 	 */
1107 	maxp = mtu_to_max_payload(sc, mtu, 0);
1108 	if (vi->flags & INTR_RXQ) {
1109 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
1110 		    CTLFLAG_RD, NULL, "rx queues");
1111 	}
1112 	for_each_rxq(vi, i, rxq) {
1113 
1114 		init_iq(&rxq->iq, sc, vi->tmr_idx, vi->pktc_idx, vi->qsize_rxq);
1115 
1116 		snprintf(name, sizeof(name), "%s rxq%d-fl",
1117 		    device_get_nameunit(vi->dev), i);
1118 		init_fl(sc, &rxq->fl, vi->qsize_rxq / 8, maxp, name);
1119 
1120 		if (vi->flags & INTR_RXQ) {
1121 			rxq->iq.flags |= IQ_INTR;
1122 			rc = alloc_rxq(vi, rxq, intr_idx, i, oid);
1123 			if (rc != 0)
1124 				goto done;
1125 			intr_idx++;
1126 		}
1127 	}
1128 #ifdef DEV_NETMAP
1129 	if (ifp->if_capabilities & IFCAP_NETMAP)
1130 		intr_idx = saved_idx + max(vi->nrxq, vi->nnmrxq);
1131 #endif
1132 #ifdef TCP_OFFLOAD
1133 	maxp = mtu_to_max_payload(sc, mtu, 1);
1134 	if (vi->flags & INTR_OFLD_RXQ) {
1135 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq",
1136 		    CTLFLAG_RD, NULL,
1137 		    "rx queues for offloaded TCP connections");
1138 	}
1139 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1140 
1141 		init_iq(&ofld_rxq->iq, sc, vi->ofld_tmr_idx, vi->ofld_pktc_idx,
1142 		    vi->qsize_rxq);
1143 
1144 		snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
1145 		    device_get_nameunit(vi->dev), i);
1146 		init_fl(sc, &ofld_rxq->fl, vi->qsize_rxq / 8, maxp, name);
1147 
1148 		if (vi->flags & INTR_OFLD_RXQ) {
1149 			ofld_rxq->iq.flags |= IQ_INTR;
1150 			rc = alloc_ofld_rxq(vi, ofld_rxq, intr_idx, i, oid);
1151 			if (rc != 0)
1152 				goto done;
1153 			intr_idx++;
1154 		}
1155 	}
1156 #endif
1157 
1158 	/*
1159 	 * Second pass over all NIC and TOE rx queues.  The queues forwarding
1160 	 * their interrupts are allocated now.
1161 	 */
1162 	j = 0;
1163 	if (!(vi->flags & INTR_RXQ)) {
1164 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
1165 		    CTLFLAG_RD, NULL, "rx queues");
1166 		for_each_rxq(vi, i, rxq) {
1167 			MPASS(!(rxq->iq.flags & IQ_INTR));
1168 
1169 			intr_idx = vi_intr_iq(vi, j)->abs_id;
1170 
1171 			rc = alloc_rxq(vi, rxq, intr_idx, i, oid);
1172 			if (rc != 0)
1173 				goto done;
1174 			j++;
1175 		}
1176 	}
1177 #ifdef TCP_OFFLOAD
1178 	if (vi->nofldrxq != 0 && !(vi->flags & INTR_OFLD_RXQ)) {
1179 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq",
1180 		    CTLFLAG_RD, NULL,
1181 		    "rx queues for offloaded TCP connections");
1182 		for_each_ofld_rxq(vi, i, ofld_rxq) {
1183 			MPASS(!(ofld_rxq->iq.flags & IQ_INTR));
1184 
1185 			intr_idx = vi_intr_iq(vi, j)->abs_id;
1186 
1187 			rc = alloc_ofld_rxq(vi, ofld_rxq, intr_idx, i, oid);
1188 			if (rc != 0)
1189 				goto done;
1190 			j++;
1191 		}
1192 	}
1193 #endif
1194 
1195 	/*
1196 	 * Now the tx queues.  Only one pass needed.
1197 	 */
1198 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "txq", CTLFLAG_RD,
1199 	    NULL, "tx queues");
1200 	j = 0;
1201 	for_each_txq(vi, i, txq) {
1202 		iqid = vi_intr_iq(vi, j)->cntxt_id;
1203 		snprintf(name, sizeof(name), "%s txq%d",
1204 		    device_get_nameunit(vi->dev), i);
1205 		init_eq(sc, &txq->eq, EQ_ETH, vi->qsize_txq, pi->tx_chan, iqid,
1206 		    name);
1207 
1208 		rc = alloc_txq(vi, txq, i, oid);
1209 		if (rc != 0)
1210 			goto done;
1211 		j++;
1212 	}
1213 #ifdef TCP_OFFLOAD
1214 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_txq",
1215 	    CTLFLAG_RD, NULL, "tx queues for offloaded TCP connections");
1216 	for_each_ofld_txq(vi, i, ofld_txq) {
1217 		struct sysctl_oid *oid2;
1218 
1219 		iqid = vi_intr_iq(vi, j)->cntxt_id;
1220 		snprintf(name, sizeof(name), "%s ofld_txq%d",
1221 		    device_get_nameunit(vi->dev), i);
1222 		init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq, pi->tx_chan,
1223 		    iqid, name);
1224 
1225 		snprintf(name, sizeof(name), "%d", i);
1226 		oid2 = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
1227 		    name, CTLFLAG_RD, NULL, "offload tx queue");
1228 
1229 		rc = alloc_wrq(sc, vi, ofld_txq, oid2);
1230 		if (rc != 0)
1231 			goto done;
1232 		j++;
1233 	}
1234 #endif
1235 
1236 	/*
1237 	 * Finally, the control queue.
1238 	 */
1239 	if (!IS_MAIN_VI(vi) || sc->flags & IS_VF)
1240 		goto done;
1241 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ctrlq", CTLFLAG_RD,
1242 	    NULL, "ctrl queue");
1243 	ctrlq = &sc->sge.ctrlq[pi->port_id];
1244 	iqid = vi_intr_iq(vi, 0)->cntxt_id;
1245 	snprintf(name, sizeof(name), "%s ctrlq", device_get_nameunit(vi->dev));
1246 	init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, pi->tx_chan, iqid,
1247 	    name);
1248 	rc = alloc_wrq(sc, vi, ctrlq, oid);
1249 
1250 done:
1251 	if (rc)
1252 		t4_teardown_vi_queues(vi);
1253 
1254 	return (rc);
1255 }
1256 
1257 /*
1258  * Idempotent
1259  */
1260 int
1261 t4_teardown_vi_queues(struct vi_info *vi)
1262 {
1263 	int i;
1264 	struct port_info *pi = vi->pi;
1265 	struct adapter *sc = pi->adapter;
1266 	struct sge_rxq *rxq;
1267 	struct sge_txq *txq;
1268 #ifdef TCP_OFFLOAD
1269 	struct sge_ofld_rxq *ofld_rxq;
1270 	struct sge_wrq *ofld_txq;
1271 #endif
1272 #ifdef DEV_NETMAP
1273 	struct sge_nm_rxq *nm_rxq;
1274 	struct sge_nm_txq *nm_txq;
1275 #endif
1276 
1277 	/* Do this before freeing the queues */
1278 	if (vi->flags & VI_SYSCTL_CTX) {
1279 		sysctl_ctx_free(&vi->ctx);
1280 		vi->flags &= ~VI_SYSCTL_CTX;
1281 	}
1282 
1283 #ifdef DEV_NETMAP
1284 	if (vi->ifp->if_capabilities & IFCAP_NETMAP) {
1285 		for_each_nm_txq(vi, i, nm_txq) {
1286 			free_nm_txq(vi, nm_txq);
1287 		}
1288 
1289 		for_each_nm_rxq(vi, i, nm_rxq) {
1290 			free_nm_rxq(vi, nm_rxq);
1291 		}
1292 	}
1293 #endif
1294 
1295 	/*
1296 	 * Take down all the tx queues first, as they reference the rx queues
1297 	 * (for egress updates, etc.).
1298 	 */
1299 
1300 	if (IS_MAIN_VI(vi) && !(sc->flags & IS_VF))
1301 		free_wrq(sc, &sc->sge.ctrlq[pi->port_id]);
1302 
1303 	for_each_txq(vi, i, txq) {
1304 		free_txq(vi, txq);
1305 	}
1306 #ifdef TCP_OFFLOAD
1307 	for_each_ofld_txq(vi, i, ofld_txq) {
1308 		free_wrq(sc, ofld_txq);
1309 	}
1310 #endif
1311 
1312 	/*
1313 	 * Then take down the rx queues that forward their interrupts, as they
1314 	 * reference other rx queues.
1315 	 */
1316 
1317 	for_each_rxq(vi, i, rxq) {
1318 		if ((rxq->iq.flags & IQ_INTR) == 0)
1319 			free_rxq(vi, rxq);
1320 	}
1321 #ifdef TCP_OFFLOAD
1322 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1323 		if ((ofld_rxq->iq.flags & IQ_INTR) == 0)
1324 			free_ofld_rxq(vi, ofld_rxq);
1325 	}
1326 #endif
1327 
1328 	/*
1329 	 * Then take down the rx queues that take direct interrupts.
1330 	 */
1331 
1332 	for_each_rxq(vi, i, rxq) {
1333 		if (rxq->iq.flags & IQ_INTR)
1334 			free_rxq(vi, rxq);
1335 	}
1336 #ifdef TCP_OFFLOAD
1337 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1338 		if (ofld_rxq->iq.flags & IQ_INTR)
1339 			free_ofld_rxq(vi, ofld_rxq);
1340 	}
1341 #endif
1342 
1343 	return (0);
1344 }
1345 
1346 /*
1347  * Deals with errors and the firmware event queue.  All data rx queues forward
1348  * their interrupt to the firmware event queue.
1349  */
1350 void
1351 t4_intr_all(void *arg)
1352 {
1353 	struct adapter *sc = arg;
1354 	struct sge_iq *fwq = &sc->sge.fwq;
1355 
1356 	t4_intr_err(arg);
1357 	if (atomic_cmpset_int(&fwq->state, IQS_IDLE, IQS_BUSY)) {
1358 		service_iq(fwq, 0);
1359 		atomic_cmpset_int(&fwq->state, IQS_BUSY, IQS_IDLE);
1360 	}
1361 }
1362 
1363 /* Deals with error interrupts */
1364 void
1365 t4_intr_err(void *arg)
1366 {
1367 	struct adapter *sc = arg;
1368 
1369 	t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
1370 	t4_slow_intr_handler(sc);
1371 }
1372 
1373 void
1374 t4_intr_evt(void *arg)
1375 {
1376 	struct sge_iq *iq = arg;
1377 
1378 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1379 		service_iq(iq, 0);
1380 		atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1381 	}
1382 }
1383 
1384 void
1385 t4_intr(void *arg)
1386 {
1387 	struct sge_iq *iq = arg;
1388 
1389 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1390 		service_iq(iq, 0);
1391 		atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1392 	}
1393 }
1394 
1395 void
1396 t4_vi_intr(void *arg)
1397 {
1398 	struct irq *irq = arg;
1399 
1400 #ifdef DEV_NETMAP
1401 	if (atomic_cmpset_int(&irq->nm_state, NM_ON, NM_BUSY)) {
1402 		t4_nm_intr(irq->nm_rxq);
1403 		atomic_cmpset_int(&irq->nm_state, NM_BUSY, NM_ON);
1404 	}
1405 #endif
1406 	if (irq->rxq != NULL)
1407 		t4_intr(irq->rxq);
1408 }
1409 
1410 static inline int
1411 sort_before_lro(struct lro_ctrl *lro)
1412 {
1413 
1414 	return (lro->lro_mbuf_max != 0);
1415 }
1416 
1417 /*
1418  * Deals with anything and everything on the given ingress queue.
1419  */
1420 static int
1421 service_iq(struct sge_iq *iq, int budget)
1422 {
1423 	struct sge_iq *q;
1424 	struct sge_rxq *rxq = iq_to_rxq(iq);	/* Use iff iq is part of rxq */
1425 	struct sge_fl *fl;			/* Use iff IQ_HAS_FL */
1426 	struct adapter *sc = iq->adapter;
1427 	struct iq_desc *d = &iq->desc[iq->cidx];
1428 	int ndescs = 0, limit;
1429 	int rsp_type, refill;
1430 	uint32_t lq;
1431 	uint16_t fl_hw_cidx;
1432 	struct mbuf *m0;
1433 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1434 #if defined(INET) || defined(INET6)
1435 	const struct timeval lro_timeout = {0, sc->lro_timeout};
1436 	struct lro_ctrl *lro = &rxq->lro;
1437 #endif
1438 
1439 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1440 
1441 	limit = budget ? budget : iq->qsize / 16;
1442 
1443 	if (iq->flags & IQ_HAS_FL) {
1444 		fl = &rxq->fl;
1445 		fl_hw_cidx = fl->hw_cidx;	/* stable snapshot */
1446 	} else {
1447 		fl = NULL;
1448 		fl_hw_cidx = 0;			/* to silence gcc warning */
1449 	}
1450 
1451 #if defined(INET) || defined(INET6)
1452 	if (iq->flags & IQ_ADJ_CREDIT) {
1453 		MPASS(sort_before_lro(lro));
1454 		iq->flags &= ~IQ_ADJ_CREDIT;
1455 		if ((d->rsp.u.type_gen & F_RSPD_GEN) != iq->gen) {
1456 			tcp_lro_flush_all(lro);
1457 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(1) |
1458 			    V_INGRESSQID((u32)iq->cntxt_id) |
1459 			    V_SEINTARM(iq->intr_params));
1460 			return (0);
1461 		}
1462 		ndescs = 1;
1463 	}
1464 #else
1465 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1466 #endif
1467 
1468 	/*
1469 	 * We always come back and check the descriptor ring for new indirect
1470 	 * interrupts and other responses after running a single handler.
1471 	 */
1472 	for (;;) {
1473 		while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1474 
1475 			rmb();
1476 
1477 			refill = 0;
1478 			m0 = NULL;
1479 			rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1480 			lq = be32toh(d->rsp.pldbuflen_qid);
1481 
1482 			switch (rsp_type) {
1483 			case X_RSPD_TYPE_FLBUF:
1484 
1485 				KASSERT(iq->flags & IQ_HAS_FL,
1486 				    ("%s: data for an iq (%p) with no freelist",
1487 				    __func__, iq));
1488 
1489 				m0 = get_fl_payload(sc, fl, lq);
1490 				if (__predict_false(m0 == NULL))
1491 					goto process_iql;
1492 				refill = IDXDIFF(fl->hw_cidx, fl_hw_cidx, fl->sidx) > 2;
1493 #ifdef T4_PKT_TIMESTAMP
1494 				/*
1495 				 * 60 bit timestamp for the payload is
1496 				 * *(uint64_t *)m0->m_pktdat.  Note that it is
1497 				 * in the leading free-space in the mbuf.  The
1498 				 * kernel can clobber it during a pullup,
1499 				 * m_copymdata, etc.  You need to make sure that
1500 				 * the mbuf reaches you unmolested if you care
1501 				 * about the timestamp.
1502 				 */
1503 				*(uint64_t *)m0->m_pktdat =
1504 				    be64toh(ctrl->u.last_flit) &
1505 				    0xfffffffffffffff;
1506 #endif
1507 
1508 				/* fall through */
1509 
1510 			case X_RSPD_TYPE_CPL:
1511 				KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1512 				    ("%s: bad opcode %02x.", __func__,
1513 				    d->rss.opcode));
1514 				t4_cpl_handler[d->rss.opcode](iq, &d->rss, m0);
1515 				break;
1516 
1517 			case X_RSPD_TYPE_INTR:
1518 
1519 				/*
1520 				 * Interrupts should be forwarded only to queues
1521 				 * that are not forwarding their interrupts.
1522 				 * This means service_iq can recurse but only 1
1523 				 * level deep.
1524 				 */
1525 				KASSERT(budget == 0,
1526 				    ("%s: budget %u, rsp_type %u", __func__,
1527 				    budget, rsp_type));
1528 
1529 				/*
1530 				 * There are 1K interrupt-capable queues (qids 0
1531 				 * through 1023).  A response type indicating a
1532 				 * forwarded interrupt with a qid >= 1K is an
1533 				 * iWARP async notification.
1534 				 */
1535 				if (lq >= 1024) {
1536                                         t4_an_handler(iq, &d->rsp);
1537                                         break;
1538                                 }
1539 
1540 				q = sc->sge.iqmap[lq - sc->sge.iq_start -
1541 				    sc->sge.iq_base];
1542 				if (atomic_cmpset_int(&q->state, IQS_IDLE,
1543 				    IQS_BUSY)) {
1544 					if (service_iq(q, q->qsize / 16) == 0) {
1545 						atomic_cmpset_int(&q->state,
1546 						    IQS_BUSY, IQS_IDLE);
1547 					} else {
1548 						STAILQ_INSERT_TAIL(&iql, q,
1549 						    link);
1550 					}
1551 				}
1552 				break;
1553 
1554 			default:
1555 				KASSERT(0,
1556 				    ("%s: illegal response type %d on iq %p",
1557 				    __func__, rsp_type, iq));
1558 				log(LOG_ERR,
1559 				    "%s: illegal response type %d on iq %p",
1560 				    device_get_nameunit(sc->dev), rsp_type, iq);
1561 				break;
1562 			}
1563 
1564 			d++;
1565 			if (__predict_false(++iq->cidx == iq->sidx)) {
1566 				iq->cidx = 0;
1567 				iq->gen ^= F_RSPD_GEN;
1568 				d = &iq->desc[0];
1569 			}
1570 			if (__predict_false(++ndescs == limit)) {
1571 				t4_write_reg(sc, sc->sge_gts_reg,
1572 				    V_CIDXINC(ndescs) |
1573 				    V_INGRESSQID(iq->cntxt_id) |
1574 				    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1575 				ndescs = 0;
1576 
1577 #if defined(INET) || defined(INET6)
1578 				if (iq->flags & IQ_LRO_ENABLED &&
1579 				    !sort_before_lro(lro) &&
1580 				    sc->lro_timeout != 0) {
1581 					tcp_lro_flush_inactive(lro,
1582 					    &lro_timeout);
1583 				}
1584 #endif
1585 
1586 				if (budget) {
1587 					if (iq->flags & IQ_HAS_FL) {
1588 						FL_LOCK(fl);
1589 						refill_fl(sc, fl, 32);
1590 						FL_UNLOCK(fl);
1591 					}
1592 					return (EINPROGRESS);
1593 				}
1594 			}
1595 			if (refill) {
1596 				FL_LOCK(fl);
1597 				refill_fl(sc, fl, 32);
1598 				FL_UNLOCK(fl);
1599 				fl_hw_cidx = fl->hw_cidx;
1600 			}
1601 		}
1602 
1603 process_iql:
1604 		if (STAILQ_EMPTY(&iql))
1605 			break;
1606 
1607 		/*
1608 		 * Process the head only, and send it to the back of the list if
1609 		 * it's still not done.
1610 		 */
1611 		q = STAILQ_FIRST(&iql);
1612 		STAILQ_REMOVE_HEAD(&iql, link);
1613 		if (service_iq(q, q->qsize / 8) == 0)
1614 			atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1615 		else
1616 			STAILQ_INSERT_TAIL(&iql, q, link);
1617 	}
1618 
1619 #if defined(INET) || defined(INET6)
1620 	if (iq->flags & IQ_LRO_ENABLED) {
1621 		if (ndescs > 0 && lro->lro_mbuf_count > 8) {
1622 			MPASS(sort_before_lro(lro));
1623 			/* hold back one credit and don't flush LRO state */
1624 			iq->flags |= IQ_ADJ_CREDIT;
1625 			ndescs--;
1626 		} else {
1627 			tcp_lro_flush_all(lro);
1628 		}
1629 	}
1630 #endif
1631 
1632 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1633 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1634 
1635 	if (iq->flags & IQ_HAS_FL) {
1636 		int starved;
1637 
1638 		FL_LOCK(fl);
1639 		starved = refill_fl(sc, fl, 64);
1640 		FL_UNLOCK(fl);
1641 		if (__predict_false(starved != 0))
1642 			add_fl_to_sfl(sc, fl);
1643 	}
1644 
1645 	return (0);
1646 }
1647 
1648 static inline int
1649 cl_has_metadata(struct sge_fl *fl, struct cluster_layout *cll)
1650 {
1651 	int rc = fl->flags & FL_BUF_PACKING || cll->region1 > 0;
1652 
1653 	if (rc)
1654 		MPASS(cll->region3 >= CL_METADATA_SIZE);
1655 
1656 	return (rc);
1657 }
1658 
1659 static inline struct cluster_metadata *
1660 cl_metadata(struct adapter *sc, struct sge_fl *fl, struct cluster_layout *cll,
1661     caddr_t cl)
1662 {
1663 
1664 	if (cl_has_metadata(fl, cll)) {
1665 		struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1666 
1667 		return ((struct cluster_metadata *)(cl + swz->size) - 1);
1668 	}
1669 	return (NULL);
1670 }
1671 
1672 static void
1673 rxb_free(struct mbuf *m)
1674 {
1675 	uma_zone_t zone = m->m_ext.ext_arg1;
1676 	void *cl = m->m_ext.ext_arg2;
1677 
1678 	uma_zfree(zone, cl);
1679 	counter_u64_add(extfree_rels, 1);
1680 }
1681 
1682 /*
1683  * The mbuf returned by this function could be allocated from zone_mbuf or
1684  * constructed in spare room in the cluster.
1685  *
1686  * The mbuf carries the payload in one of these ways
1687  * a) frame inside the mbuf (mbuf from zone_mbuf)
1688  * b) m_cljset (for clusters without metadata) zone_mbuf
1689  * c) m_extaddref (cluster with metadata) inline mbuf
1690  * d) m_extaddref (cluster with metadata) zone_mbuf
1691  */
1692 static struct mbuf *
1693 get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1694     int remaining)
1695 {
1696 	struct mbuf *m;
1697 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1698 	struct cluster_layout *cll = &sd->cll;
1699 	struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1700 	struct hw_buf_info *hwb = &sc->sge.hw_buf_info[cll->hwidx];
1701 	struct cluster_metadata *clm = cl_metadata(sc, fl, cll, sd->cl);
1702 	int len, blen;
1703 	caddr_t payload;
1704 
1705 	blen = hwb->size - fl->rx_offset;	/* max possible in this buf */
1706 	len = min(remaining, blen);
1707 	payload = sd->cl + cll->region1 + fl->rx_offset;
1708 	if (fl->flags & FL_BUF_PACKING) {
1709 		const u_int l = fr_offset + len;
1710 		const u_int pad = roundup2(l, fl->buf_boundary) - l;
1711 
1712 		if (fl->rx_offset + len + pad < hwb->size)
1713 			blen = len + pad;
1714 		MPASS(fl->rx_offset + blen <= hwb->size);
1715 	} else {
1716 		MPASS(fl->rx_offset == 0);	/* not packing */
1717 	}
1718 
1719 
1720 	if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) {
1721 
1722 		/*
1723 		 * Copy payload into a freshly allocated mbuf.
1724 		 */
1725 
1726 		m = fr_offset == 0 ?
1727 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1728 		if (m == NULL)
1729 			return (NULL);
1730 		fl->mbuf_allocated++;
1731 #ifdef T4_PKT_TIMESTAMP
1732 		/* Leave room for a timestamp */
1733 		m->m_data += 8;
1734 #endif
1735 		/* copy data to mbuf */
1736 		bcopy(payload, mtod(m, caddr_t), len);
1737 
1738 	} else if (sd->nmbuf * MSIZE < cll->region1) {
1739 
1740 		/*
1741 		 * There's spare room in the cluster for an mbuf.  Create one
1742 		 * and associate it with the payload that's in the cluster.
1743 		 */
1744 
1745 		MPASS(clm != NULL);
1746 		m = (struct mbuf *)(sd->cl + sd->nmbuf * MSIZE);
1747 		/* No bzero required */
1748 		if (m_init(m, M_NOWAIT, MT_DATA,
1749 		    fr_offset == 0 ? M_PKTHDR | M_NOFREE : M_NOFREE))
1750 			return (NULL);
1751 		fl->mbuf_inlined++;
1752 		m_extaddref(m, payload, blen, &clm->refcount, rxb_free,
1753 		    swz->zone, sd->cl);
1754 		if (sd->nmbuf++ == 0)
1755 			counter_u64_add(extfree_refs, 1);
1756 
1757 	} else {
1758 
1759 		/*
1760 		 * Grab an mbuf from zone_mbuf and associate it with the
1761 		 * payload in the cluster.
1762 		 */
1763 
1764 		m = fr_offset == 0 ?
1765 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1766 		if (m == NULL)
1767 			return (NULL);
1768 		fl->mbuf_allocated++;
1769 		if (clm != NULL) {
1770 			m_extaddref(m, payload, blen, &clm->refcount,
1771 			    rxb_free, swz->zone, sd->cl);
1772 			if (sd->nmbuf++ == 0)
1773 				counter_u64_add(extfree_refs, 1);
1774 		} else {
1775 			m_cljset(m, sd->cl, swz->type);
1776 			sd->cl = NULL;	/* consumed, not a recycle candidate */
1777 		}
1778 	}
1779 	if (fr_offset == 0)
1780 		m->m_pkthdr.len = remaining;
1781 	m->m_len = len;
1782 
1783 	if (fl->flags & FL_BUF_PACKING) {
1784 		fl->rx_offset += blen;
1785 		MPASS(fl->rx_offset <= hwb->size);
1786 		if (fl->rx_offset < hwb->size)
1787 			return (m);	/* without advancing the cidx */
1788 	}
1789 
1790 	if (__predict_false(++fl->cidx % 8 == 0)) {
1791 		uint16_t cidx = fl->cidx / 8;
1792 
1793 		if (__predict_false(cidx == fl->sidx))
1794 			fl->cidx = cidx = 0;
1795 		fl->hw_cidx = cidx;
1796 	}
1797 	fl->rx_offset = 0;
1798 
1799 	return (m);
1800 }
1801 
1802 static struct mbuf *
1803 get_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf)
1804 {
1805 	struct mbuf *m0, *m, **pnext;
1806 	u_int remaining;
1807 	const u_int total = G_RSPD_LEN(len_newbuf);
1808 
1809 	if (__predict_false(fl->flags & FL_BUF_RESUME)) {
1810 		M_ASSERTPKTHDR(fl->m0);
1811 		MPASS(fl->m0->m_pkthdr.len == total);
1812 		MPASS(fl->remaining < total);
1813 
1814 		m0 = fl->m0;
1815 		pnext = fl->pnext;
1816 		remaining = fl->remaining;
1817 		fl->flags &= ~FL_BUF_RESUME;
1818 		goto get_segment;
1819 	}
1820 
1821 	if (fl->rx_offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
1822 		fl->rx_offset = 0;
1823 		if (__predict_false(++fl->cidx % 8 == 0)) {
1824 			uint16_t cidx = fl->cidx / 8;
1825 
1826 			if (__predict_false(cidx == fl->sidx))
1827 				fl->cidx = cidx = 0;
1828 			fl->hw_cidx = cidx;
1829 		}
1830 	}
1831 
1832 	/*
1833 	 * Payload starts at rx_offset in the current hw buffer.  Its length is
1834 	 * 'len' and it may span multiple hw buffers.
1835 	 */
1836 
1837 	m0 = get_scatter_segment(sc, fl, 0, total);
1838 	if (m0 == NULL)
1839 		return (NULL);
1840 	remaining = total - m0->m_len;
1841 	pnext = &m0->m_next;
1842 	while (remaining > 0) {
1843 get_segment:
1844 		MPASS(fl->rx_offset == 0);
1845 		m = get_scatter_segment(sc, fl, total - remaining, remaining);
1846 		if (__predict_false(m == NULL)) {
1847 			fl->m0 = m0;
1848 			fl->pnext = pnext;
1849 			fl->remaining = remaining;
1850 			fl->flags |= FL_BUF_RESUME;
1851 			return (NULL);
1852 		}
1853 		*pnext = m;
1854 		pnext = &m->m_next;
1855 		remaining -= m->m_len;
1856 	}
1857 	*pnext = NULL;
1858 
1859 	M_ASSERTPKTHDR(m0);
1860 	return (m0);
1861 }
1862 
1863 static int
1864 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
1865 {
1866 	struct sge_rxq *rxq = iq_to_rxq(iq);
1867 	struct ifnet *ifp = rxq->ifp;
1868 	struct adapter *sc = iq->adapter;
1869 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
1870 #if defined(INET) || defined(INET6)
1871 	struct lro_ctrl *lro = &rxq->lro;
1872 #endif
1873 	static const int sw_hashtype[4][2] = {
1874 		{M_HASHTYPE_NONE, M_HASHTYPE_NONE},
1875 		{M_HASHTYPE_RSS_IPV4, M_HASHTYPE_RSS_IPV6},
1876 		{M_HASHTYPE_RSS_TCP_IPV4, M_HASHTYPE_RSS_TCP_IPV6},
1877 		{M_HASHTYPE_RSS_UDP_IPV4, M_HASHTYPE_RSS_UDP_IPV6},
1878 	};
1879 
1880 	KASSERT(m0 != NULL, ("%s: no payload with opcode %02x", __func__,
1881 	    rss->opcode));
1882 
1883 	m0->m_pkthdr.len -= sc->params.sge.fl_pktshift;
1884 	m0->m_len -= sc->params.sge.fl_pktshift;
1885 	m0->m_data += sc->params.sge.fl_pktshift;
1886 
1887 	m0->m_pkthdr.rcvif = ifp;
1888 	M_HASHTYPE_SET(m0, sw_hashtype[rss->hash_type][rss->ipv6]);
1889 	m0->m_pkthdr.flowid = be32toh(rss->hash_val);
1890 
1891 	if (cpl->csum_calc && !(cpl->err_vec & sc->params.tp.err_vec_mask)) {
1892 		if (ifp->if_capenable & IFCAP_RXCSUM &&
1893 		    cpl->l2info & htobe32(F_RXF_IP)) {
1894 			m0->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
1895 			    CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
1896 			rxq->rxcsum++;
1897 		} else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
1898 		    cpl->l2info & htobe32(F_RXF_IP6)) {
1899 			m0->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
1900 			    CSUM_PSEUDO_HDR);
1901 			rxq->rxcsum++;
1902 		}
1903 
1904 		if (__predict_false(cpl->ip_frag))
1905 			m0->m_pkthdr.csum_data = be16toh(cpl->csum);
1906 		else
1907 			m0->m_pkthdr.csum_data = 0xffff;
1908 	}
1909 
1910 	if (cpl->vlan_ex) {
1911 		m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
1912 		m0->m_flags |= M_VLANTAG;
1913 		rxq->vlan_extraction++;
1914 	}
1915 
1916 #if defined(INET) || defined(INET6)
1917 	if (iq->flags & IQ_LRO_ENABLED) {
1918 		if (sort_before_lro(lro)) {
1919 			tcp_lro_queue_mbuf(lro, m0);
1920 			return (0); /* queued for sort, then LRO */
1921 		}
1922 		if (tcp_lro_rx(lro, m0, 0) == 0)
1923 			return (0); /* queued for LRO */
1924 	}
1925 #endif
1926 	ifp->if_input(ifp, m0);
1927 
1928 	return (0);
1929 }
1930 
1931 /*
1932  * Must drain the wrq or make sure that someone else will.
1933  */
1934 static void
1935 wrq_tx_drain(void *arg, int n)
1936 {
1937 	struct sge_wrq *wrq = arg;
1938 	struct sge_eq *eq = &wrq->eq;
1939 
1940 	EQ_LOCK(eq);
1941 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
1942 		drain_wrq_wr_list(wrq->adapter, wrq);
1943 	EQ_UNLOCK(eq);
1944 }
1945 
1946 static void
1947 drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq)
1948 {
1949 	struct sge_eq *eq = &wrq->eq;
1950 	u_int available, dbdiff;	/* # of hardware descriptors */
1951 	u_int n;
1952 	struct wrqe *wr;
1953 	struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
1954 
1955 	EQ_LOCK_ASSERT_OWNED(eq);
1956 	MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
1957 	wr = STAILQ_FIRST(&wrq->wr_list);
1958 	MPASS(wr != NULL);	/* Must be called with something useful to do */
1959 	MPASS(eq->pidx == eq->dbidx);
1960 	dbdiff = 0;
1961 
1962 	do {
1963 		eq->cidx = read_hw_cidx(eq);
1964 		if (eq->pidx == eq->cidx)
1965 			available = eq->sidx - 1;
1966 		else
1967 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
1968 
1969 		MPASS(wr->wrq == wrq);
1970 		n = howmany(wr->wr_len, EQ_ESIZE);
1971 		if (available < n)
1972 			break;
1973 
1974 		dst = (void *)&eq->desc[eq->pidx];
1975 		if (__predict_true(eq->sidx - eq->pidx > n)) {
1976 			/* Won't wrap, won't end exactly at the status page. */
1977 			bcopy(&wr->wr[0], dst, wr->wr_len);
1978 			eq->pidx += n;
1979 		} else {
1980 			int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE;
1981 
1982 			bcopy(&wr->wr[0], dst, first_portion);
1983 			if (wr->wr_len > first_portion) {
1984 				bcopy(&wr->wr[first_portion], &eq->desc[0],
1985 				    wr->wr_len - first_portion);
1986 			}
1987 			eq->pidx = n - (eq->sidx - eq->pidx);
1988 		}
1989 		wrq->tx_wrs_copied++;
1990 
1991 		if (available < eq->sidx / 4 &&
1992 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
1993 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
1994 			    F_FW_WR_EQUEQ);
1995 			eq->equeqidx = eq->pidx;
1996 		} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
1997 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
1998 			eq->equeqidx = eq->pidx;
1999 		}
2000 
2001 		dbdiff += n;
2002 		if (dbdiff >= 16) {
2003 			ring_eq_db(sc, eq, dbdiff);
2004 			dbdiff = 0;
2005 		}
2006 
2007 		STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
2008 		free_wrqe(wr);
2009 		MPASS(wrq->nwr_pending > 0);
2010 		wrq->nwr_pending--;
2011 		MPASS(wrq->ndesc_needed >= n);
2012 		wrq->ndesc_needed -= n;
2013 	} while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL);
2014 
2015 	if (dbdiff)
2016 		ring_eq_db(sc, eq, dbdiff);
2017 }
2018 
2019 /*
2020  * Doesn't fail.  Holds on to work requests it can't send right away.
2021  */
2022 void
2023 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
2024 {
2025 #ifdef INVARIANTS
2026 	struct sge_eq *eq = &wrq->eq;
2027 #endif
2028 
2029 	EQ_LOCK_ASSERT_OWNED(eq);
2030 	MPASS(wr != NULL);
2031 	MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN);
2032 	MPASS((wr->wr_len & 0x7) == 0);
2033 
2034 	STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
2035 	wrq->nwr_pending++;
2036 	wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE);
2037 
2038 	if (!TAILQ_EMPTY(&wrq->incomplete_wrs))
2039 		return;	/* commit_wrq_wr will drain wr_list as well. */
2040 
2041 	drain_wrq_wr_list(sc, wrq);
2042 
2043 	/* Doorbell must have caught up to the pidx. */
2044 	MPASS(eq->pidx == eq->dbidx);
2045 }
2046 
2047 void
2048 t4_update_fl_bufsize(struct ifnet *ifp)
2049 {
2050 	struct vi_info *vi = ifp->if_softc;
2051 	struct adapter *sc = vi->pi->adapter;
2052 	struct sge_rxq *rxq;
2053 #ifdef TCP_OFFLOAD
2054 	struct sge_ofld_rxq *ofld_rxq;
2055 #endif
2056 	struct sge_fl *fl;
2057 	int i, maxp, mtu = ifp->if_mtu;
2058 
2059 	maxp = mtu_to_max_payload(sc, mtu, 0);
2060 	for_each_rxq(vi, i, rxq) {
2061 		fl = &rxq->fl;
2062 
2063 		FL_LOCK(fl);
2064 		find_best_refill_source(sc, fl, maxp);
2065 		FL_UNLOCK(fl);
2066 	}
2067 #ifdef TCP_OFFLOAD
2068 	maxp = mtu_to_max_payload(sc, mtu, 1);
2069 	for_each_ofld_rxq(vi, i, ofld_rxq) {
2070 		fl = &ofld_rxq->fl;
2071 
2072 		FL_LOCK(fl);
2073 		find_best_refill_source(sc, fl, maxp);
2074 		FL_UNLOCK(fl);
2075 	}
2076 #endif
2077 }
2078 
2079 static inline int
2080 mbuf_nsegs(struct mbuf *m)
2081 {
2082 
2083 	M_ASSERTPKTHDR(m);
2084 	KASSERT(m->m_pkthdr.l5hlen > 0,
2085 	    ("%s: mbuf %p missing information on # of segments.", __func__, m));
2086 
2087 	return (m->m_pkthdr.l5hlen);
2088 }
2089 
2090 static inline void
2091 set_mbuf_nsegs(struct mbuf *m, uint8_t nsegs)
2092 {
2093 
2094 	M_ASSERTPKTHDR(m);
2095 	m->m_pkthdr.l5hlen = nsegs;
2096 }
2097 
2098 static inline int
2099 mbuf_len16(struct mbuf *m)
2100 {
2101 	int n;
2102 
2103 	M_ASSERTPKTHDR(m);
2104 	n = m->m_pkthdr.PH_loc.eight[0];
2105 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2106 
2107 	return (n);
2108 }
2109 
2110 static inline void
2111 set_mbuf_len16(struct mbuf *m, uint8_t len16)
2112 {
2113 
2114 	M_ASSERTPKTHDR(m);
2115 	m->m_pkthdr.PH_loc.eight[0] = len16;
2116 }
2117 
2118 static inline int
2119 needs_tso(struct mbuf *m)
2120 {
2121 
2122 	M_ASSERTPKTHDR(m);
2123 
2124 	if (m->m_pkthdr.csum_flags & CSUM_TSO) {
2125 		KASSERT(m->m_pkthdr.tso_segsz > 0,
2126 		    ("%s: TSO requested in mbuf %p but MSS not provided",
2127 		    __func__, m));
2128 		return (1);
2129 	}
2130 
2131 	return (0);
2132 }
2133 
2134 static inline int
2135 needs_l3_csum(struct mbuf *m)
2136 {
2137 
2138 	M_ASSERTPKTHDR(m);
2139 
2140 	if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO))
2141 		return (1);
2142 	return (0);
2143 }
2144 
2145 static inline int
2146 needs_l4_csum(struct mbuf *m)
2147 {
2148 
2149 	M_ASSERTPKTHDR(m);
2150 
2151 	if (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
2152 	    CSUM_TCP_IPV6 | CSUM_TSO))
2153 		return (1);
2154 	return (0);
2155 }
2156 
2157 static inline int
2158 needs_vlan_insertion(struct mbuf *m)
2159 {
2160 
2161 	M_ASSERTPKTHDR(m);
2162 
2163 	if (m->m_flags & M_VLANTAG) {
2164 		KASSERT(m->m_pkthdr.ether_vtag != 0,
2165 		    ("%s: HWVLAN requested in mbuf %p but tag not provided",
2166 		    __func__, m));
2167 		return (1);
2168 	}
2169 	return (0);
2170 }
2171 
2172 static void *
2173 m_advance(struct mbuf **pm, int *poffset, int len)
2174 {
2175 	struct mbuf *m = *pm;
2176 	int offset = *poffset;
2177 	uintptr_t p = 0;
2178 
2179 	MPASS(len > 0);
2180 
2181 	for (;;) {
2182 		if (offset + len < m->m_len) {
2183 			offset += len;
2184 			p = mtod(m, uintptr_t) + offset;
2185 			break;
2186 		}
2187 		len -= m->m_len - offset;
2188 		m = m->m_next;
2189 		offset = 0;
2190 		MPASS(m != NULL);
2191 	}
2192 	*poffset = offset;
2193 	*pm = m;
2194 	return ((void *)p);
2195 }
2196 
2197 /*
2198  * Can deal with empty mbufs in the chain that have m_len = 0, but the chain
2199  * must have at least one mbuf that's not empty.
2200  */
2201 static inline int
2202 count_mbuf_nsegs(struct mbuf *m)
2203 {
2204 	vm_paddr_t lastb, next;
2205 	vm_offset_t va;
2206 	int len, nsegs;
2207 
2208 	MPASS(m != NULL);
2209 
2210 	nsegs = 0;
2211 	lastb = 0;
2212 	for (; m; m = m->m_next) {
2213 
2214 		len = m->m_len;
2215 		if (__predict_false(len == 0))
2216 			continue;
2217 		va = mtod(m, vm_offset_t);
2218 		next = pmap_kextract(va);
2219 		nsegs += sglist_count(m->m_data, len);
2220 		if (lastb + 1 == next)
2221 			nsegs--;
2222 		lastb = pmap_kextract(va + len - 1);
2223 	}
2224 
2225 	MPASS(nsegs > 0);
2226 	return (nsegs);
2227 }
2228 
2229 /*
2230  * Analyze the mbuf to determine its tx needs.  The mbuf passed in may change:
2231  * a) caller can assume it's been freed if this function returns with an error.
2232  * b) it may get defragged up if the gather list is too long for the hardware.
2233  */
2234 int
2235 parse_pkt(struct adapter *sc, struct mbuf **mp)
2236 {
2237 	struct mbuf *m0 = *mp, *m;
2238 	int rc, nsegs, defragged = 0, offset;
2239 	struct ether_header *eh;
2240 	void *l3hdr;
2241 #if defined(INET) || defined(INET6)
2242 	struct tcphdr *tcp;
2243 #endif
2244 	uint16_t eh_type;
2245 
2246 	M_ASSERTPKTHDR(m0);
2247 	if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) {
2248 		rc = EINVAL;
2249 fail:
2250 		m_freem(m0);
2251 		*mp = NULL;
2252 		return (rc);
2253 	}
2254 restart:
2255 	/*
2256 	 * First count the number of gather list segments in the payload.
2257 	 * Defrag the mbuf if nsegs exceeds the hardware limit.
2258 	 */
2259 	M_ASSERTPKTHDR(m0);
2260 	MPASS(m0->m_pkthdr.len > 0);
2261 	nsegs = count_mbuf_nsegs(m0);
2262 	if (nsegs > (needs_tso(m0) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS)) {
2263 		if (defragged++ > 0 || (m = m_defrag(m0, M_NOWAIT)) == NULL) {
2264 			rc = EFBIG;
2265 			goto fail;
2266 		}
2267 		*mp = m0 = m;	/* update caller's copy after defrag */
2268 		goto restart;
2269 	}
2270 
2271 	if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN)) {
2272 		m0 = m_pullup(m0, m0->m_pkthdr.len);
2273 		if (m0 == NULL) {
2274 			/* Should have left well enough alone. */
2275 			rc = EFBIG;
2276 			goto fail;
2277 		}
2278 		*mp = m0;	/* update caller's copy after pullup */
2279 		goto restart;
2280 	}
2281 	set_mbuf_nsegs(m0, nsegs);
2282 	if (sc->flags & IS_VF)
2283 		set_mbuf_len16(m0, txpkt_vm_len16(nsegs, needs_tso(m0)));
2284 	else
2285 		set_mbuf_len16(m0, txpkt_len16(nsegs, needs_tso(m0)));
2286 
2287 	if (!needs_tso(m0) &&
2288 	    !(sc->flags & IS_VF && (needs_l3_csum(m0) || needs_l4_csum(m0))))
2289 		return (0);
2290 
2291 	m = m0;
2292 	eh = mtod(m, struct ether_header *);
2293 	eh_type = ntohs(eh->ether_type);
2294 	if (eh_type == ETHERTYPE_VLAN) {
2295 		struct ether_vlan_header *evh = (void *)eh;
2296 
2297 		eh_type = ntohs(evh->evl_proto);
2298 		m0->m_pkthdr.l2hlen = sizeof(*evh);
2299 	} else
2300 		m0->m_pkthdr.l2hlen = sizeof(*eh);
2301 
2302 	offset = 0;
2303 	l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2304 
2305 	switch (eh_type) {
2306 #ifdef INET6
2307 	case ETHERTYPE_IPV6:
2308 	{
2309 		struct ip6_hdr *ip6 = l3hdr;
2310 
2311 		MPASS(!needs_tso(m0) || ip6->ip6_nxt == IPPROTO_TCP);
2312 
2313 		m0->m_pkthdr.l3hlen = sizeof(*ip6);
2314 		break;
2315 	}
2316 #endif
2317 #ifdef INET
2318 	case ETHERTYPE_IP:
2319 	{
2320 		struct ip *ip = l3hdr;
2321 
2322 		m0->m_pkthdr.l3hlen = ip->ip_hl * 4;
2323 		break;
2324 	}
2325 #endif
2326 	default:
2327 		panic("%s: ethertype 0x%04x unknown.  if_cxgbe must be compiled"
2328 		    " with the same INET/INET6 options as the kernel.",
2329 		    __func__, eh_type);
2330 	}
2331 
2332 #if defined(INET) || defined(INET6)
2333 	if (needs_tso(m0)) {
2334 		tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen);
2335 		m0->m_pkthdr.l4hlen = tcp->th_off * 4;
2336 	}
2337 #endif
2338 	MPASS(m0 == *mp);
2339 	return (0);
2340 }
2341 
2342 void *
2343 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie)
2344 {
2345 	struct sge_eq *eq = &wrq->eq;
2346 	struct adapter *sc = wrq->adapter;
2347 	int ndesc, available;
2348 	struct wrqe *wr;
2349 	void *w;
2350 
2351 	MPASS(len16 > 0);
2352 	ndesc = howmany(len16, EQ_ESIZE / 16);
2353 	MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC);
2354 
2355 	EQ_LOCK(eq);
2356 
2357 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2358 		drain_wrq_wr_list(sc, wrq);
2359 
2360 	if (!STAILQ_EMPTY(&wrq->wr_list)) {
2361 slowpath:
2362 		EQ_UNLOCK(eq);
2363 		wr = alloc_wrqe(len16 * 16, wrq);
2364 		if (__predict_false(wr == NULL))
2365 			return (NULL);
2366 		cookie->pidx = -1;
2367 		cookie->ndesc = ndesc;
2368 		return (&wr->wr);
2369 	}
2370 
2371 	eq->cidx = read_hw_cidx(eq);
2372 	if (eq->pidx == eq->cidx)
2373 		available = eq->sidx - 1;
2374 	else
2375 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2376 	if (available < ndesc)
2377 		goto slowpath;
2378 
2379 	cookie->pidx = eq->pidx;
2380 	cookie->ndesc = ndesc;
2381 	TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link);
2382 
2383 	w = &eq->desc[eq->pidx];
2384 	IDXINCR(eq->pidx, ndesc, eq->sidx);
2385 	if (__predict_false(cookie->pidx + ndesc > eq->sidx)) {
2386 		w = &wrq->ss[0];
2387 		wrq->ss_pidx = cookie->pidx;
2388 		wrq->ss_len = len16 * 16;
2389 	}
2390 
2391 	EQ_UNLOCK(eq);
2392 
2393 	return (w);
2394 }
2395 
2396 void
2397 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie)
2398 {
2399 	struct sge_eq *eq = &wrq->eq;
2400 	struct adapter *sc = wrq->adapter;
2401 	int ndesc, pidx;
2402 	struct wrq_cookie *prev, *next;
2403 
2404 	if (cookie->pidx == -1) {
2405 		struct wrqe *wr = __containerof(w, struct wrqe, wr);
2406 
2407 		t4_wrq_tx(sc, wr);
2408 		return;
2409 	}
2410 
2411 	if (__predict_false(w == &wrq->ss[0])) {
2412 		int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE;
2413 
2414 		MPASS(wrq->ss_len > n);	/* WR had better wrap around. */
2415 		bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n);
2416 		bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n);
2417 		wrq->tx_wrs_ss++;
2418 	} else
2419 		wrq->tx_wrs_direct++;
2420 
2421 	EQ_LOCK(eq);
2422 	ndesc = cookie->ndesc;	/* Can be more than SGE_MAX_WR_NDESC here. */
2423 	pidx = cookie->pidx;
2424 	MPASS(pidx >= 0 && pidx < eq->sidx);
2425 	prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link);
2426 	next = TAILQ_NEXT(cookie, link);
2427 	if (prev == NULL) {
2428 		MPASS(pidx == eq->dbidx);
2429 		if (next == NULL || ndesc >= 16)
2430 			ring_eq_db(wrq->adapter, eq, ndesc);
2431 		else {
2432 			MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc);
2433 			next->pidx = pidx;
2434 			next->ndesc += ndesc;
2435 		}
2436 	} else {
2437 		MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc);
2438 		prev->ndesc += ndesc;
2439 	}
2440 	TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link);
2441 
2442 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2443 		drain_wrq_wr_list(sc, wrq);
2444 
2445 #ifdef INVARIANTS
2446 	if (TAILQ_EMPTY(&wrq->incomplete_wrs)) {
2447 		/* Doorbell must have caught up to the pidx. */
2448 		MPASS(wrq->eq.pidx == wrq->eq.dbidx);
2449 	}
2450 #endif
2451 	EQ_UNLOCK(eq);
2452 }
2453 
2454 static u_int
2455 can_resume_eth_tx(struct mp_ring *r)
2456 {
2457 	struct sge_eq *eq = r->cookie;
2458 
2459 	return (total_available_tx_desc(eq) > eq->sidx / 8);
2460 }
2461 
2462 static inline int
2463 cannot_use_txpkts(struct mbuf *m)
2464 {
2465 	/* maybe put a GL limit too, to avoid silliness? */
2466 
2467 	return (needs_tso(m));
2468 }
2469 
2470 static inline int
2471 discard_tx(struct sge_eq *eq)
2472 {
2473 
2474 	return ((eq->flags & (EQ_ENABLED | EQ_QFLUSH)) != EQ_ENABLED);
2475 }
2476 
2477 /*
2478  * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to
2479  * be consumed.  Return the actual number consumed.  0 indicates a stall.
2480  */
2481 static u_int
2482 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx)
2483 {
2484 	struct sge_txq *txq = r->cookie;
2485 	struct sge_eq *eq = &txq->eq;
2486 	struct ifnet *ifp = txq->ifp;
2487 	struct vi_info *vi = ifp->if_softc;
2488 	struct port_info *pi = vi->pi;
2489 	struct adapter *sc = pi->adapter;
2490 	u_int total, remaining;		/* # of packets */
2491 	u_int available, dbdiff;	/* # of hardware descriptors */
2492 	u_int n, next_cidx;
2493 	struct mbuf *m0, *tail;
2494 	struct txpkts txp;
2495 	struct fw_eth_tx_pkts_wr *wr;	/* any fw WR struct will do */
2496 
2497 	remaining = IDXDIFF(pidx, cidx, r->size);
2498 	MPASS(remaining > 0);	/* Must not be called without work to do. */
2499 	total = 0;
2500 
2501 	TXQ_LOCK(txq);
2502 	if (__predict_false(discard_tx(eq))) {
2503 		while (cidx != pidx) {
2504 			m0 = r->items[cidx];
2505 			m_freem(m0);
2506 			if (++cidx == r->size)
2507 				cidx = 0;
2508 		}
2509 		reclaim_tx_descs(txq, 2048);
2510 		total = remaining;
2511 		goto done;
2512 	}
2513 
2514 	/* How many hardware descriptors do we have readily available. */
2515 	if (eq->pidx == eq->cidx)
2516 		available = eq->sidx - 1;
2517 	else
2518 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2519 	dbdiff = IDXDIFF(eq->pidx, eq->dbidx, eq->sidx);
2520 
2521 	while (remaining > 0) {
2522 
2523 		m0 = r->items[cidx];
2524 		M_ASSERTPKTHDR(m0);
2525 		MPASS(m0->m_nextpkt == NULL);
2526 
2527 		if (available < SGE_MAX_WR_NDESC) {
2528 			available += reclaim_tx_descs(txq, 64);
2529 			if (available < howmany(mbuf_len16(m0), EQ_ESIZE / 16))
2530 				break;	/* out of descriptors */
2531 		}
2532 
2533 		next_cidx = cidx + 1;
2534 		if (__predict_false(next_cidx == r->size))
2535 			next_cidx = 0;
2536 
2537 		wr = (void *)&eq->desc[eq->pidx];
2538 		if (sc->flags & IS_VF) {
2539 			total++;
2540 			remaining--;
2541 			ETHER_BPF_MTAP(ifp, m0);
2542 			n = write_txpkt_vm_wr(sc, txq, (void *)wr, m0,
2543 			    available);
2544 		} else if (remaining > 1 &&
2545 		    try_txpkts(m0, r->items[next_cidx], &txp, available) == 0) {
2546 
2547 			/* pkts at cidx, next_cidx should both be in txp. */
2548 			MPASS(txp.npkt == 2);
2549 			tail = r->items[next_cidx];
2550 			MPASS(tail->m_nextpkt == NULL);
2551 			ETHER_BPF_MTAP(ifp, m0);
2552 			ETHER_BPF_MTAP(ifp, tail);
2553 			m0->m_nextpkt = tail;
2554 
2555 			if (__predict_false(++next_cidx == r->size))
2556 				next_cidx = 0;
2557 
2558 			while (next_cidx != pidx) {
2559 				if (add_to_txpkts(r->items[next_cidx], &txp,
2560 				    available) != 0)
2561 					break;
2562 				tail->m_nextpkt = r->items[next_cidx];
2563 				tail = tail->m_nextpkt;
2564 				ETHER_BPF_MTAP(ifp, tail);
2565 				if (__predict_false(++next_cidx == r->size))
2566 					next_cidx = 0;
2567 			}
2568 
2569 			n = write_txpkts_wr(txq, wr, m0, &txp, available);
2570 			total += txp.npkt;
2571 			remaining -= txp.npkt;
2572 		} else {
2573 			total++;
2574 			remaining--;
2575 			ETHER_BPF_MTAP(ifp, m0);
2576 			n = write_txpkt_wr(txq, (void *)wr, m0, available);
2577 		}
2578 		MPASS(n >= 1 && n <= available && n <= SGE_MAX_WR_NDESC);
2579 
2580 		available -= n;
2581 		dbdiff += n;
2582 		IDXINCR(eq->pidx, n, eq->sidx);
2583 
2584 		if (total_available_tx_desc(eq) < eq->sidx / 4 &&
2585 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2586 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2587 			    F_FW_WR_EQUEQ);
2588 			eq->equeqidx = eq->pidx;
2589 		} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
2590 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
2591 			eq->equeqidx = eq->pidx;
2592 		}
2593 
2594 		if (dbdiff >= 16 && remaining >= 4) {
2595 			ring_eq_db(sc, eq, dbdiff);
2596 			available += reclaim_tx_descs(txq, 4 * dbdiff);
2597 			dbdiff = 0;
2598 		}
2599 
2600 		cidx = next_cidx;
2601 	}
2602 	if (dbdiff != 0) {
2603 		ring_eq_db(sc, eq, dbdiff);
2604 		reclaim_tx_descs(txq, 32);
2605 	}
2606 done:
2607 	TXQ_UNLOCK(txq);
2608 
2609 	return (total);
2610 }
2611 
2612 static inline void
2613 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
2614     int qsize)
2615 {
2616 
2617 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
2618 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
2619 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
2620 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
2621 
2622 	iq->flags = 0;
2623 	iq->adapter = sc;
2624 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
2625 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
2626 	if (pktc_idx >= 0) {
2627 		iq->intr_params |= F_QINTR_CNT_EN;
2628 		iq->intr_pktc_idx = pktc_idx;
2629 	}
2630 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
2631 	iq->sidx = iq->qsize - sc->params.sge.spg_len / IQ_ESIZE;
2632 }
2633 
2634 static inline void
2635 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name)
2636 {
2637 
2638 	fl->qsize = qsize;
2639 	fl->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2640 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
2641 	if (sc->flags & BUF_PACKING_OK &&
2642 	    ((!is_t4(sc) && buffer_packing) ||	/* T5+: enabled unless 0 */
2643 	    (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */
2644 		fl->flags |= FL_BUF_PACKING;
2645 	find_best_refill_source(sc, fl, maxp);
2646 	find_safe_refill_source(sc, fl);
2647 }
2648 
2649 static inline void
2650 init_eq(struct adapter *sc, struct sge_eq *eq, int eqtype, int qsize,
2651     uint8_t tx_chan, uint16_t iqid, char *name)
2652 {
2653 	KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
2654 
2655 	eq->flags = eqtype & EQ_TYPEMASK;
2656 	eq->tx_chan = tx_chan;
2657 	eq->iqid = iqid;
2658 	eq->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2659 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
2660 }
2661 
2662 static int
2663 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
2664     bus_dmamap_t *map, bus_addr_t *pa, void **va)
2665 {
2666 	int rc;
2667 
2668 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
2669 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
2670 	if (rc != 0) {
2671 		device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
2672 		goto done;
2673 	}
2674 
2675 	rc = bus_dmamem_alloc(*tag, va,
2676 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
2677 	if (rc != 0) {
2678 		device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
2679 		goto done;
2680 	}
2681 
2682 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
2683 	if (rc != 0) {
2684 		device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
2685 		goto done;
2686 	}
2687 done:
2688 	if (rc)
2689 		free_ring(sc, *tag, *map, *pa, *va);
2690 
2691 	return (rc);
2692 }
2693 
2694 static int
2695 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
2696     bus_addr_t pa, void *va)
2697 {
2698 	if (pa)
2699 		bus_dmamap_unload(tag, map);
2700 	if (va)
2701 		bus_dmamem_free(tag, va, map);
2702 	if (tag)
2703 		bus_dma_tag_destroy(tag);
2704 
2705 	return (0);
2706 }
2707 
2708 /*
2709  * Allocates the ring for an ingress queue and an optional freelist.  If the
2710  * freelist is specified it will be allocated and then associated with the
2711  * ingress queue.
2712  *
2713  * Returns errno on failure.  Resources allocated up to that point may still be
2714  * allocated.  Caller is responsible for cleanup in case this function fails.
2715  *
2716  * If the ingress queue will take interrupts directly (iq->flags & IQ_INTR) then
2717  * the intr_idx specifies the vector, starting from 0.  Otherwise it specifies
2718  * the abs_id of the ingress queue to which its interrupts should be forwarded.
2719  */
2720 static int
2721 alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl,
2722     int intr_idx, int cong)
2723 {
2724 	int rc, i, cntxt_id;
2725 	size_t len;
2726 	struct fw_iq_cmd c;
2727 	struct port_info *pi = vi->pi;
2728 	struct adapter *sc = iq->adapter;
2729 	struct sge_params *sp = &sc->params.sge;
2730 	__be32 v = 0;
2731 
2732 	len = iq->qsize * IQ_ESIZE;
2733 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
2734 	    (void **)&iq->desc);
2735 	if (rc != 0)
2736 		return (rc);
2737 
2738 	bzero(&c, sizeof(c));
2739 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
2740 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
2741 	    V_FW_IQ_CMD_VFN(0));
2742 
2743 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
2744 	    FW_LEN16(c));
2745 
2746 	/* Special handling for firmware event queue */
2747 	if (iq == &sc->sge.fwq)
2748 		v |= F_FW_IQ_CMD_IQASYNCH;
2749 
2750 	if (iq->flags & IQ_INTR) {
2751 		KASSERT(intr_idx < sc->intr_count,
2752 		    ("%s: invalid direct intr_idx %d", __func__, intr_idx));
2753 	} else
2754 		v |= F_FW_IQ_CMD_IQANDST;
2755 	v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
2756 
2757 	c.type_to_iqandstindex = htobe32(v |
2758 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
2759 	    V_FW_IQ_CMD_VIID(vi->viid) |
2760 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
2761 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
2762 	    F_FW_IQ_CMD_IQGTSMODE |
2763 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
2764 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
2765 	c.iqsize = htobe16(iq->qsize);
2766 	c.iqaddr = htobe64(iq->ba);
2767 	if (cong >= 0)
2768 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
2769 
2770 	if (fl) {
2771 		mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
2772 
2773 		len = fl->qsize * EQ_ESIZE;
2774 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
2775 		    &fl->ba, (void **)&fl->desc);
2776 		if (rc)
2777 			return (rc);
2778 
2779 		/* Allocate space for one software descriptor per buffer. */
2780 		rc = alloc_fl_sdesc(fl);
2781 		if (rc != 0) {
2782 			device_printf(sc->dev,
2783 			    "failed to setup fl software descriptors: %d\n",
2784 			    rc);
2785 			return (rc);
2786 		}
2787 
2788 		if (fl->flags & FL_BUF_PACKING) {
2789 			fl->lowat = roundup2(sp->fl_starve_threshold2, 8);
2790 			fl->buf_boundary = sp->pack_boundary;
2791 		} else {
2792 			fl->lowat = roundup2(sp->fl_starve_threshold, 8);
2793 			fl->buf_boundary = 16;
2794 		}
2795 		if (fl_pad && fl->buf_boundary < sp->pad_boundary)
2796 			fl->buf_boundary = sp->pad_boundary;
2797 
2798 		c.iqns_to_fl0congen |=
2799 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
2800 			F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
2801 			(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
2802 			(fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
2803 			    0));
2804 		if (cong >= 0) {
2805 			c.iqns_to_fl0congen |=
2806 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
2807 				    F_FW_IQ_CMD_FL0CONGCIF |
2808 				    F_FW_IQ_CMD_FL0CONGEN);
2809 		}
2810 		c.fl0dcaen_to_fl0cidxfthresh =
2811 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
2812 			X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B) |
2813 			V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
2814 			X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
2815 		c.fl0size = htobe16(fl->qsize);
2816 		c.fl0addr = htobe64(fl->ba);
2817 	}
2818 
2819 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
2820 	if (rc != 0) {
2821 		device_printf(sc->dev,
2822 		    "failed to create ingress queue: %d\n", rc);
2823 		return (rc);
2824 	}
2825 
2826 	iq->cidx = 0;
2827 	iq->gen = F_RSPD_GEN;
2828 	iq->intr_next = iq->intr_params;
2829 	iq->cntxt_id = be16toh(c.iqid);
2830 	iq->abs_id = be16toh(c.physiqid);
2831 	iq->flags |= IQ_ALLOCATED;
2832 
2833 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
2834 	if (cntxt_id >= sc->sge.niq) {
2835 		panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
2836 		    cntxt_id, sc->sge.niq - 1);
2837 	}
2838 	sc->sge.iqmap[cntxt_id] = iq;
2839 
2840 	if (fl) {
2841 		u_int qid;
2842 
2843 		iq->flags |= IQ_HAS_FL;
2844 		fl->cntxt_id = be16toh(c.fl0id);
2845 		fl->pidx = fl->cidx = 0;
2846 
2847 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
2848 		if (cntxt_id >= sc->sge.neq) {
2849 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
2850 			    __func__, cntxt_id, sc->sge.neq - 1);
2851 		}
2852 		sc->sge.eqmap[cntxt_id] = (void *)fl;
2853 
2854 		qid = fl->cntxt_id;
2855 		if (isset(&sc->doorbells, DOORBELL_UDB)) {
2856 			uint32_t s_qpp = sc->params.sge.eq_s_qpp;
2857 			uint32_t mask = (1 << s_qpp) - 1;
2858 			volatile uint8_t *udb;
2859 
2860 			udb = sc->udbs_base + UDBS_DB_OFFSET;
2861 			udb += (qid >> s_qpp) << PAGE_SHIFT;
2862 			qid &= mask;
2863 			if (qid < PAGE_SIZE / UDBS_SEG_SIZE) {
2864 				udb += qid << UDBS_SEG_SHIFT;
2865 				qid = 0;
2866 			}
2867 			fl->udb = (volatile void *)udb;
2868 		}
2869 		fl->dbval = V_QID(qid) | sc->chip_params->sge_fl_db;
2870 
2871 		FL_LOCK(fl);
2872 		/* Enough to make sure the SGE doesn't think it's starved */
2873 		refill_fl(sc, fl, fl->lowat);
2874 		FL_UNLOCK(fl);
2875 	}
2876 
2877 	if (chip_id(sc) >= CHELSIO_T5 && !(sc->flags & IS_VF) && cong >= 0) {
2878 		uint32_t param, val;
2879 
2880 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
2881 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
2882 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
2883 		if (cong == 0)
2884 			val = 1 << 19;
2885 		else {
2886 			val = 2 << 19;
2887 			for (i = 0; i < 4; i++) {
2888 				if (cong & (1 << i))
2889 					val |= 1 << (i << 2);
2890 			}
2891 		}
2892 
2893 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
2894 		if (rc != 0) {
2895 			/* report error but carry on */
2896 			device_printf(sc->dev,
2897 			    "failed to set congestion manager context for "
2898 			    "ingress queue %d: %d\n", iq->cntxt_id, rc);
2899 		}
2900 	}
2901 
2902 	/* Enable IQ interrupts */
2903 	atomic_store_rel_int(&iq->state, IQS_IDLE);
2904 	t4_write_reg(sc, sc->sge_gts_reg, V_SEINTARM(iq->intr_params) |
2905 	    V_INGRESSQID(iq->cntxt_id));
2906 
2907 	return (0);
2908 }
2909 
2910 static int
2911 free_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl)
2912 {
2913 	int rc;
2914 	struct adapter *sc = iq->adapter;
2915 	device_t dev;
2916 
2917 	if (sc == NULL)
2918 		return (0);	/* nothing to do */
2919 
2920 	dev = vi ? vi->dev : sc->dev;
2921 
2922 	if (iq->flags & IQ_ALLOCATED) {
2923 		rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
2924 		    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
2925 		    fl ? fl->cntxt_id : 0xffff, 0xffff);
2926 		if (rc != 0) {
2927 			device_printf(dev,
2928 			    "failed to free queue %p: %d\n", iq, rc);
2929 			return (rc);
2930 		}
2931 		iq->flags &= ~IQ_ALLOCATED;
2932 	}
2933 
2934 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
2935 
2936 	bzero(iq, sizeof(*iq));
2937 
2938 	if (fl) {
2939 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
2940 		    fl->desc);
2941 
2942 		if (fl->sdesc)
2943 			free_fl_sdesc(sc, fl);
2944 
2945 		if (mtx_initialized(&fl->fl_lock))
2946 			mtx_destroy(&fl->fl_lock);
2947 
2948 		bzero(fl, sizeof(*fl));
2949 	}
2950 
2951 	return (0);
2952 }
2953 
2954 static void
2955 add_fl_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
2956     struct sysctl_oid *oid, struct sge_fl *fl)
2957 {
2958 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2959 
2960 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
2961 	    "freelist");
2962 	children = SYSCTL_CHILDREN(oid);
2963 
2964 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
2965 	    &fl->ba, "bus address of descriptor ring");
2966 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
2967 	    fl->sidx * EQ_ESIZE + sc->params.sge.spg_len,
2968 	    "desc ring size in bytes");
2969 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
2970 	    CTLTYPE_INT | CTLFLAG_RD, &fl->cntxt_id, 0, sysctl_uint16, "I",
2971 	    "SGE context id of the freelist");
2972 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL,
2973 	    fl_pad ? 1 : 0, "padding enabled");
2974 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL,
2975 	    fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled");
2976 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
2977 	    0, "consumer index");
2978 	if (fl->flags & FL_BUF_PACKING) {
2979 		SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
2980 		    CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
2981 	}
2982 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
2983 	    0, "producer index");
2984 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_allocated",
2985 	    CTLFLAG_RD, &fl->mbuf_allocated, "# of mbuf allocated");
2986 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_inlined",
2987 	    CTLFLAG_RD, &fl->mbuf_inlined, "# of mbuf inlined in clusters");
2988 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
2989 	    CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
2990 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
2991 	    CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
2992 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
2993 	    CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
2994 }
2995 
2996 static int
2997 alloc_fwq(struct adapter *sc)
2998 {
2999 	int rc, intr_idx;
3000 	struct sge_iq *fwq = &sc->sge.fwq;
3001 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
3002 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3003 
3004 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE);
3005 	fwq->flags |= IQ_INTR;	/* always */
3006 	if (sc->flags & IS_VF)
3007 		intr_idx = 0;
3008 	else {
3009 		intr_idx = sc->intr_count > 1 ? 1 : 0;
3010 		fwq->set_tcb_rpl = t4_filter_rpl;
3011 		fwq->l2t_write_rpl = do_l2t_write_rpl;
3012 	}
3013 	rc = alloc_iq_fl(&sc->port[0]->vi[0], fwq, NULL, intr_idx, -1);
3014 	if (rc != 0) {
3015 		device_printf(sc->dev,
3016 		    "failed to create firmware event queue: %d\n", rc);
3017 		return (rc);
3018 	}
3019 
3020 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD,
3021 	    NULL, "firmware event queue");
3022 	children = SYSCTL_CHILDREN(oid);
3023 
3024 	SYSCTL_ADD_UAUTO(&sc->ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3025 	    &fwq->ba, "bus address of descriptor ring");
3026 	SYSCTL_ADD_INT(&sc->ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3027 	    fwq->qsize * IQ_ESIZE, "descriptor ring size in bytes");
3028 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "abs_id",
3029 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->abs_id, 0, sysctl_uint16, "I",
3030 	    "absolute id of the queue");
3031 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cntxt_id",
3032 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->cntxt_id, 0, sysctl_uint16, "I",
3033 	    "SGE context id of the queue");
3034 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cidx",
3035 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->cidx, 0, sysctl_uint16, "I",
3036 	    "consumer index");
3037 
3038 	return (0);
3039 }
3040 
3041 static int
3042 free_fwq(struct adapter *sc)
3043 {
3044 	return free_iq_fl(NULL, &sc->sge.fwq, NULL);
3045 }
3046 
3047 static int
3048 alloc_mgmtq(struct adapter *sc)
3049 {
3050 	int rc;
3051 	struct sge_wrq *mgmtq = &sc->sge.mgmtq;
3052 	char name[16];
3053 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
3054 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3055 
3056 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "mgmtq", CTLFLAG_RD,
3057 	    NULL, "management queue");
3058 
3059 	snprintf(name, sizeof(name), "%s mgmtq", device_get_nameunit(sc->dev));
3060 	init_eq(sc, &mgmtq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[0]->tx_chan,
3061 	    sc->sge.fwq.cntxt_id, name);
3062 	rc = alloc_wrq(sc, NULL, mgmtq, oid);
3063 	if (rc != 0) {
3064 		device_printf(sc->dev,
3065 		    "failed to create management queue: %d\n", rc);
3066 		return (rc);
3067 	}
3068 
3069 	return (0);
3070 }
3071 
3072 static int
3073 free_mgmtq(struct adapter *sc)
3074 {
3075 
3076 	return free_wrq(sc, &sc->sge.mgmtq);
3077 }
3078 
3079 int
3080 tnl_cong(struct port_info *pi, int drop)
3081 {
3082 
3083 	if (drop == -1)
3084 		return (-1);
3085 	else if (drop == 1)
3086 		return (0);
3087 	else
3088 		return (pi->rx_e_chan_map);
3089 }
3090 
3091 static int
3092 alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int intr_idx, int idx,
3093     struct sysctl_oid *oid)
3094 {
3095 	int rc;
3096 	struct adapter *sc = vi->pi->adapter;
3097 	struct sysctl_oid_list *children;
3098 	char name[16];
3099 
3100 	rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, intr_idx,
3101 	    tnl_cong(vi->pi, cong_drop));
3102 	if (rc != 0)
3103 		return (rc);
3104 
3105 	if (idx == 0)
3106 		sc->sge.iq_base = rxq->iq.abs_id - rxq->iq.cntxt_id;
3107 	else
3108 		KASSERT(rxq->iq.cntxt_id + sc->sge.iq_base == rxq->iq.abs_id,
3109 		    ("iq_base mismatch"));
3110 	KASSERT(sc->sge.iq_base == 0 || sc->flags & IS_VF,
3111 	    ("PF with non-zero iq_base"));
3112 
3113 	/*
3114 	 * The freelist is just barely above the starvation threshold right now,
3115 	 * fill it up a bit more.
3116 	 */
3117 	FL_LOCK(&rxq->fl);
3118 	refill_fl(sc, &rxq->fl, 128);
3119 	FL_UNLOCK(&rxq->fl);
3120 
3121 #if defined(INET) || defined(INET6)
3122 	rc = tcp_lro_init_args(&rxq->lro, vi->ifp, lro_entries, lro_mbufs);
3123 	if (rc != 0)
3124 		return (rc);
3125 	MPASS(rxq->lro.ifp == vi->ifp);	/* also indicates LRO init'ed */
3126 
3127 	if (vi->ifp->if_capenable & IFCAP_LRO)
3128 		rxq->iq.flags |= IQ_LRO_ENABLED;
3129 #endif
3130 	rxq->ifp = vi->ifp;
3131 
3132 	children = SYSCTL_CHILDREN(oid);
3133 
3134 	snprintf(name, sizeof(name), "%d", idx);
3135 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3136 	    NULL, "rx queue");
3137 	children = SYSCTL_CHILDREN(oid);
3138 
3139 	SYSCTL_ADD_UAUTO(&vi->ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3140 	    &rxq->iq.ba, "bus address of descriptor ring");
3141 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3142 	    rxq->iq.qsize * IQ_ESIZE, "descriptor ring size in bytes");
3143 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "abs_id",
3144 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.abs_id, 0, sysctl_uint16, "I",
3145 	    "absolute id of the queue");
3146 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cntxt_id",
3147 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cntxt_id, 0, sysctl_uint16, "I",
3148 	    "SGE context id of the queue");
3149 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3150 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cidx, 0, sysctl_uint16, "I",
3151 	    "consumer index");
3152 #if defined(INET) || defined(INET6)
3153 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
3154 	    &rxq->lro.lro_queued, 0, NULL);
3155 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
3156 	    &rxq->lro.lro_flushed, 0, NULL);
3157 #endif
3158 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
3159 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
3160 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_extraction",
3161 	    CTLFLAG_RD, &rxq->vlan_extraction,
3162 	    "# of times hardware extracted 802.1Q tag");
3163 
3164 	add_fl_sysctls(sc, &vi->ctx, oid, &rxq->fl);
3165 
3166 	return (rc);
3167 }
3168 
3169 static int
3170 free_rxq(struct vi_info *vi, struct sge_rxq *rxq)
3171 {
3172 	int rc;
3173 
3174 #if defined(INET) || defined(INET6)
3175 	if (rxq->lro.ifp) {
3176 		tcp_lro_free(&rxq->lro);
3177 		rxq->lro.ifp = NULL;
3178 	}
3179 #endif
3180 
3181 	rc = free_iq_fl(vi, &rxq->iq, &rxq->fl);
3182 	if (rc == 0)
3183 		bzero(rxq, sizeof(*rxq));
3184 
3185 	return (rc);
3186 }
3187 
3188 #ifdef TCP_OFFLOAD
3189 static int
3190 alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq,
3191     int intr_idx, int idx, struct sysctl_oid *oid)
3192 {
3193 	struct port_info *pi = vi->pi;
3194 	int rc;
3195 	struct sysctl_oid_list *children;
3196 	char name[16];
3197 
3198 	rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx, 0);
3199 	if (rc != 0)
3200 		return (rc);
3201 
3202 	children = SYSCTL_CHILDREN(oid);
3203 
3204 	snprintf(name, sizeof(name), "%d", idx);
3205 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3206 	    NULL, "rx queue");
3207 	children = SYSCTL_CHILDREN(oid);
3208 
3209 	SYSCTL_ADD_UAUTO(&vi->ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3210 	    &ofld_rxq->iq.ba, "bus address of descriptor ring");
3211 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3212 	    ofld_rxq->iq.qsize * IQ_ESIZE, "descriptor ring size in bytes");
3213 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "abs_id",
3214 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.abs_id, 0, sysctl_uint16,
3215 	    "I", "absolute id of the queue");
3216 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cntxt_id",
3217 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cntxt_id, 0, sysctl_uint16,
3218 	    "I", "SGE context id of the queue");
3219 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3220 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cidx, 0, sysctl_uint16, "I",
3221 	    "consumer index");
3222 
3223 	add_fl_sysctls(pi->adapter, &vi->ctx, oid, &ofld_rxq->fl);
3224 
3225 	return (rc);
3226 }
3227 
3228 static int
3229 free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq)
3230 {
3231 	int rc;
3232 
3233 	rc = free_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl);
3234 	if (rc == 0)
3235 		bzero(ofld_rxq, sizeof(*ofld_rxq));
3236 
3237 	return (rc);
3238 }
3239 #endif
3240 
3241 #ifdef DEV_NETMAP
3242 static int
3243 alloc_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int intr_idx,
3244     int idx, struct sysctl_oid *oid)
3245 {
3246 	int rc;
3247 	struct sysctl_oid_list *children;
3248 	struct sysctl_ctx_list *ctx;
3249 	char name[16];
3250 	size_t len;
3251 	struct adapter *sc = vi->pi->adapter;
3252 	struct netmap_adapter *na = NA(vi->ifp);
3253 
3254 	MPASS(na != NULL);
3255 
3256 	len = vi->qsize_rxq * IQ_ESIZE;
3257 	rc = alloc_ring(sc, len, &nm_rxq->iq_desc_tag, &nm_rxq->iq_desc_map,
3258 	    &nm_rxq->iq_ba, (void **)&nm_rxq->iq_desc);
3259 	if (rc != 0)
3260 		return (rc);
3261 
3262 	len = na->num_rx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3263 	rc = alloc_ring(sc, len, &nm_rxq->fl_desc_tag, &nm_rxq->fl_desc_map,
3264 	    &nm_rxq->fl_ba, (void **)&nm_rxq->fl_desc);
3265 	if (rc != 0)
3266 		return (rc);
3267 
3268 	nm_rxq->vi = vi;
3269 	nm_rxq->nid = idx;
3270 	nm_rxq->iq_cidx = 0;
3271 	nm_rxq->iq_sidx = vi->qsize_rxq - sc->params.sge.spg_len / IQ_ESIZE;
3272 	nm_rxq->iq_gen = F_RSPD_GEN;
3273 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
3274 	nm_rxq->fl_sidx = na->num_rx_desc;
3275 	nm_rxq->intr_idx = intr_idx;
3276 	nm_rxq->iq_cntxt_id = INVALID_NM_RXQ_CNTXT_ID;
3277 
3278 	ctx = &vi->ctx;
3279 	children = SYSCTL_CHILDREN(oid);
3280 
3281 	snprintf(name, sizeof(name), "%d", idx);
3282 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL,
3283 	    "rx queue");
3284 	children = SYSCTL_CHILDREN(oid);
3285 
3286 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3287 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_abs_id, 0, sysctl_uint16,
3288 	    "I", "absolute id of the queue");
3289 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3290 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cntxt_id, 0, sysctl_uint16,
3291 	    "I", "SGE context id of the queue");
3292 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3293 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cidx, 0, sysctl_uint16, "I",
3294 	    "consumer index");
3295 
3296 	children = SYSCTL_CHILDREN(oid);
3297 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3298 	    "freelist");
3299 	children = SYSCTL_CHILDREN(oid);
3300 
3301 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3302 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->fl_cntxt_id, 0, sysctl_uint16,
3303 	    "I", "SGE context id of the freelist");
3304 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
3305 	    &nm_rxq->fl_cidx, 0, "consumer index");
3306 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
3307 	    &nm_rxq->fl_pidx, 0, "producer index");
3308 
3309 	return (rc);
3310 }
3311 
3312 
3313 static int
3314 free_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
3315 {
3316 	struct adapter *sc = vi->pi->adapter;
3317 
3318 	if (vi->flags & VI_INIT_DONE)
3319 		MPASS(nm_rxq->iq_cntxt_id == INVALID_NM_RXQ_CNTXT_ID);
3320 	else
3321 		MPASS(nm_rxq->iq_cntxt_id == 0);
3322 
3323 	free_ring(sc, nm_rxq->iq_desc_tag, nm_rxq->iq_desc_map, nm_rxq->iq_ba,
3324 	    nm_rxq->iq_desc);
3325 	free_ring(sc, nm_rxq->fl_desc_tag, nm_rxq->fl_desc_map, nm_rxq->fl_ba,
3326 	    nm_rxq->fl_desc);
3327 
3328 	return (0);
3329 }
3330 
3331 static int
3332 alloc_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq, int iqidx, int idx,
3333     struct sysctl_oid *oid)
3334 {
3335 	int rc;
3336 	size_t len;
3337 	struct port_info *pi = vi->pi;
3338 	struct adapter *sc = pi->adapter;
3339 	struct netmap_adapter *na = NA(vi->ifp);
3340 	char name[16];
3341 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3342 
3343 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3344 	rc = alloc_ring(sc, len, &nm_txq->desc_tag, &nm_txq->desc_map,
3345 	    &nm_txq->ba, (void **)&nm_txq->desc);
3346 	if (rc)
3347 		return (rc);
3348 
3349 	nm_txq->pidx = nm_txq->cidx = 0;
3350 	nm_txq->sidx = na->num_tx_desc;
3351 	nm_txq->nid = idx;
3352 	nm_txq->iqidx = iqidx;
3353 	nm_txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3354 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(G_FW_VIID_PFN(vi->viid)) |
3355 	    V_TXPKT_VF(G_FW_VIID_VIN(vi->viid)) |
3356 	    V_TXPKT_VF_VLD(G_FW_VIID_VIVLD(vi->viid)));
3357 	nm_txq->cntxt_id = INVALID_NM_TXQ_CNTXT_ID;
3358 
3359 	snprintf(name, sizeof(name), "%d", idx);
3360 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3361 	    NULL, "netmap tx queue");
3362 	children = SYSCTL_CHILDREN(oid);
3363 
3364 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3365 	    &nm_txq->cntxt_id, 0, "SGE context id of the queue");
3366 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3367 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->cidx, 0, sysctl_uint16, "I",
3368 	    "consumer index");
3369 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3370 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->pidx, 0, sysctl_uint16, "I",
3371 	    "producer index");
3372 
3373 	return (rc);
3374 }
3375 
3376 static int
3377 free_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
3378 {
3379 	struct adapter *sc = vi->pi->adapter;
3380 
3381 	if (vi->flags & VI_INIT_DONE)
3382 		MPASS(nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID);
3383 	else
3384 		MPASS(nm_txq->cntxt_id == 0);
3385 
3386 	free_ring(sc, nm_txq->desc_tag, nm_txq->desc_map, nm_txq->ba,
3387 	    nm_txq->desc);
3388 
3389 	return (0);
3390 }
3391 #endif
3392 
3393 static int
3394 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
3395 {
3396 	int rc, cntxt_id;
3397 	struct fw_eq_ctrl_cmd c;
3398 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3399 
3400 	bzero(&c, sizeof(c));
3401 
3402 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
3403 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
3404 	    V_FW_EQ_CTRL_CMD_VFN(0));
3405 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
3406 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
3407 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid));
3408 	c.physeqid_pkd = htobe32(0);
3409 	c.fetchszm_to_iqid =
3410 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
3411 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
3412 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
3413 	c.dcaen_to_eqsize =
3414 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3415 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3416 		V_FW_EQ_CTRL_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
3417 		V_FW_EQ_CTRL_CMD_EQSIZE(qsize));
3418 	c.eqaddr = htobe64(eq->ba);
3419 
3420 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3421 	if (rc != 0) {
3422 		device_printf(sc->dev,
3423 		    "failed to create control queue %d: %d\n", eq->tx_chan, rc);
3424 		return (rc);
3425 	}
3426 	eq->flags |= EQ_ALLOCATED;
3427 
3428 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
3429 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3430 	if (cntxt_id >= sc->sge.neq)
3431 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3432 		cntxt_id, sc->sge.neq - 1);
3433 	sc->sge.eqmap[cntxt_id] = eq;
3434 
3435 	return (rc);
3436 }
3437 
3438 static int
3439 eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3440 {
3441 	int rc, cntxt_id;
3442 	struct fw_eq_eth_cmd c;
3443 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3444 
3445 	bzero(&c, sizeof(c));
3446 
3447 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
3448 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
3449 	    V_FW_EQ_ETH_CMD_VFN(0));
3450 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
3451 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
3452 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
3453 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
3454 	c.fetchszm_to_iqid =
3455 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3456 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
3457 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
3458 	c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3459 	    V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3460 	    V_FW_EQ_ETH_CMD_EQSIZE(qsize));
3461 	c.eqaddr = htobe64(eq->ba);
3462 
3463 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3464 	if (rc != 0) {
3465 		device_printf(vi->dev,
3466 		    "failed to create Ethernet egress queue: %d\n", rc);
3467 		return (rc);
3468 	}
3469 	eq->flags |= EQ_ALLOCATED;
3470 
3471 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
3472 	eq->abs_id = G_FW_EQ_ETH_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
3473 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3474 	if (cntxt_id >= sc->sge.neq)
3475 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3476 		cntxt_id, sc->sge.neq - 1);
3477 	sc->sge.eqmap[cntxt_id] = eq;
3478 
3479 	return (rc);
3480 }
3481 
3482 #ifdef TCP_OFFLOAD
3483 static int
3484 ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3485 {
3486 	int rc, cntxt_id;
3487 	struct fw_eq_ofld_cmd c;
3488 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3489 
3490 	bzero(&c, sizeof(c));
3491 
3492 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
3493 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
3494 	    V_FW_EQ_OFLD_CMD_VFN(0));
3495 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
3496 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
3497 	c.fetchszm_to_iqid =
3498 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3499 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
3500 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
3501 	c.dcaen_to_eqsize =
3502 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3503 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3504 		V_FW_EQ_OFLD_CMD_EQSIZE(qsize));
3505 	c.eqaddr = htobe64(eq->ba);
3506 
3507 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3508 	if (rc != 0) {
3509 		device_printf(vi->dev,
3510 		    "failed to create egress queue for TCP offload: %d\n", rc);
3511 		return (rc);
3512 	}
3513 	eq->flags |= EQ_ALLOCATED;
3514 
3515 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
3516 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3517 	if (cntxt_id >= sc->sge.neq)
3518 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3519 		cntxt_id, sc->sge.neq - 1);
3520 	sc->sge.eqmap[cntxt_id] = eq;
3521 
3522 	return (rc);
3523 }
3524 #endif
3525 
3526 static int
3527 alloc_eq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3528 {
3529 	int rc, qsize;
3530 	size_t len;
3531 
3532 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
3533 
3534 	qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3535 	len = qsize * EQ_ESIZE;
3536 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
3537 	    &eq->ba, (void **)&eq->desc);
3538 	if (rc)
3539 		return (rc);
3540 
3541 	eq->pidx = eq->cidx = 0;
3542 	eq->equeqidx = eq->dbidx = 0;
3543 	eq->doorbells = sc->doorbells;
3544 
3545 	switch (eq->flags & EQ_TYPEMASK) {
3546 	case EQ_CTRL:
3547 		rc = ctrl_eq_alloc(sc, eq);
3548 		break;
3549 
3550 	case EQ_ETH:
3551 		rc = eth_eq_alloc(sc, vi, eq);
3552 		break;
3553 
3554 #ifdef TCP_OFFLOAD
3555 	case EQ_OFLD:
3556 		rc = ofld_eq_alloc(sc, vi, eq);
3557 		break;
3558 #endif
3559 
3560 	default:
3561 		panic("%s: invalid eq type %d.", __func__,
3562 		    eq->flags & EQ_TYPEMASK);
3563 	}
3564 	if (rc != 0) {
3565 		device_printf(sc->dev,
3566 		    "failed to allocate egress queue(%d): %d\n",
3567 		    eq->flags & EQ_TYPEMASK, rc);
3568 	}
3569 
3570 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
3571 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
3572 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
3573 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3574 		uint32_t mask = (1 << s_qpp) - 1;
3575 		volatile uint8_t *udb;
3576 
3577 		udb = sc->udbs_base + UDBS_DB_OFFSET;
3578 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
3579 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
3580 		if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
3581 	    		clrbit(&eq->doorbells, DOORBELL_WCWR);
3582 		else {
3583 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
3584 			eq->udb_qid = 0;
3585 		}
3586 		eq->udb = (volatile void *)udb;
3587 	}
3588 
3589 	return (rc);
3590 }
3591 
3592 static int
3593 free_eq(struct adapter *sc, struct sge_eq *eq)
3594 {
3595 	int rc;
3596 
3597 	if (eq->flags & EQ_ALLOCATED) {
3598 		switch (eq->flags & EQ_TYPEMASK) {
3599 		case EQ_CTRL:
3600 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
3601 			    eq->cntxt_id);
3602 			break;
3603 
3604 		case EQ_ETH:
3605 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
3606 			    eq->cntxt_id);
3607 			break;
3608 
3609 #ifdef TCP_OFFLOAD
3610 		case EQ_OFLD:
3611 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
3612 			    eq->cntxt_id);
3613 			break;
3614 #endif
3615 
3616 		default:
3617 			panic("%s: invalid eq type %d.", __func__,
3618 			    eq->flags & EQ_TYPEMASK);
3619 		}
3620 		if (rc != 0) {
3621 			device_printf(sc->dev,
3622 			    "failed to free egress queue (%d): %d\n",
3623 			    eq->flags & EQ_TYPEMASK, rc);
3624 			return (rc);
3625 		}
3626 		eq->flags &= ~EQ_ALLOCATED;
3627 	}
3628 
3629 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
3630 
3631 	if (mtx_initialized(&eq->eq_lock))
3632 		mtx_destroy(&eq->eq_lock);
3633 
3634 	bzero(eq, sizeof(*eq));
3635 	return (0);
3636 }
3637 
3638 static int
3639 alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq,
3640     struct sysctl_oid *oid)
3641 {
3642 	int rc;
3643 	struct sysctl_ctx_list *ctx = vi ? &vi->ctx : &sc->ctx;
3644 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3645 
3646 	rc = alloc_eq(sc, vi, &wrq->eq);
3647 	if (rc)
3648 		return (rc);
3649 
3650 	wrq->adapter = sc;
3651 	TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq);
3652 	TAILQ_INIT(&wrq->incomplete_wrs);
3653 	STAILQ_INIT(&wrq->wr_list);
3654 	wrq->nwr_pending = 0;
3655 	wrq->ndesc_needed = 0;
3656 
3657 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3658 	    &wrq->eq.ba, "bus address of descriptor ring");
3659 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3660 	    wrq->eq.sidx * EQ_ESIZE + sc->params.sge.spg_len,
3661 	    "desc ring size in bytes");
3662 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3663 	    &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
3664 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3665 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I",
3666 	    "consumer index");
3667 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx",
3668 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I",
3669 	    "producer index");
3670 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
3671 	    wrq->eq.sidx, "status page index");
3672 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD,
3673 	    &wrq->tx_wrs_direct, "# of work requests (direct)");
3674 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD,
3675 	    &wrq->tx_wrs_copied, "# of work requests (copied)");
3676 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_sspace", CTLFLAG_RD,
3677 	    &wrq->tx_wrs_ss, "# of work requests (copied from scratch space)");
3678 
3679 	return (rc);
3680 }
3681 
3682 static int
3683 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
3684 {
3685 	int rc;
3686 
3687 	rc = free_eq(sc, &wrq->eq);
3688 	if (rc)
3689 		return (rc);
3690 
3691 	bzero(wrq, sizeof(*wrq));
3692 	return (0);
3693 }
3694 
3695 static int
3696 alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx,
3697     struct sysctl_oid *oid)
3698 {
3699 	int rc;
3700 	struct port_info *pi = vi->pi;
3701 	struct adapter *sc = pi->adapter;
3702 	struct sge_eq *eq = &txq->eq;
3703 	char name[16];
3704 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3705 
3706 	rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx, can_resume_eth_tx,
3707 	    M_CXGBE, M_WAITOK);
3708 	if (rc != 0) {
3709 		device_printf(sc->dev, "failed to allocate mp_ring: %d\n", rc);
3710 		return (rc);
3711 	}
3712 
3713 	rc = alloc_eq(sc, vi, eq);
3714 	if (rc != 0) {
3715 		mp_ring_free(txq->r);
3716 		txq->r = NULL;
3717 		return (rc);
3718 	}
3719 
3720 	/* Can't fail after this point. */
3721 
3722 	if (idx == 0)
3723 		sc->sge.eq_base = eq->abs_id - eq->cntxt_id;
3724 	else
3725 		KASSERT(eq->cntxt_id + sc->sge.eq_base == eq->abs_id,
3726 		    ("eq_base mismatch"));
3727 	KASSERT(sc->sge.eq_base == 0 || sc->flags & IS_VF,
3728 	    ("PF with non-zero eq_base"));
3729 
3730 	TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq);
3731 	txq->ifp = vi->ifp;
3732 	txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK);
3733 	if (sc->flags & IS_VF)
3734 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
3735 		    V_TXPKT_INTF(pi->tx_chan));
3736 	else
3737 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3738 		    V_TXPKT_INTF(pi->tx_chan) |
3739 		    V_TXPKT_PF(G_FW_VIID_PFN(vi->viid)) |
3740 		    V_TXPKT_VF(G_FW_VIID_VIN(vi->viid)) |
3741 		    V_TXPKT_VF_VLD(G_FW_VIID_VIVLD(vi->viid)));
3742 	txq->tc_idx = -1;
3743 	txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE,
3744 	    M_ZERO | M_WAITOK);
3745 
3746 	snprintf(name, sizeof(name), "%d", idx);
3747 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3748 	    NULL, "tx queue");
3749 	children = SYSCTL_CHILDREN(oid);
3750 
3751 	SYSCTL_ADD_UAUTO(&vi->ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3752 	    &eq->ba, "bus address of descriptor ring");
3753 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3754 	    eq->sidx * EQ_ESIZE + sc->params.sge.spg_len,
3755 	    "desc ring size in bytes");
3756 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
3757 	    &eq->abs_id, 0, "absolute id of the queue");
3758 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3759 	    &eq->cntxt_id, 0, "SGE context id of the queue");
3760 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3761 	    CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I",
3762 	    "consumer index");
3763 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3764 	    CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I",
3765 	    "producer index");
3766 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
3767 	    eq->sidx, "status page index");
3768 
3769 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "tc",
3770 	    CTLTYPE_INT | CTLFLAG_RW, vi, idx, sysctl_tc, "I",
3771 	    "traffic class (-1 means none)");
3772 
3773 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
3774 	    &txq->txcsum, "# of times hardware assisted with checksum");
3775 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_insertion",
3776 	    CTLFLAG_RD, &txq->vlan_insertion,
3777 	    "# of times hardware inserted 802.1Q tag");
3778 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
3779 	    &txq->tso_wrs, "# of TSO work requests");
3780 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
3781 	    &txq->imm_wrs, "# of work requests with immediate data");
3782 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
3783 	    &txq->sgl_wrs, "# of work requests with direct SGL");
3784 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
3785 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
3786 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_wrs",
3787 	    CTLFLAG_RD, &txq->txpkts0_wrs,
3788 	    "# of txpkts (type 0) work requests");
3789 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_wrs",
3790 	    CTLFLAG_RD, &txq->txpkts1_wrs,
3791 	    "# of txpkts (type 1) work requests");
3792 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_pkts",
3793 	    CTLFLAG_RD, &txq->txpkts0_pkts,
3794 	    "# of frames tx'd using type0 txpkts work requests");
3795 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_pkts",
3796 	    CTLFLAG_RD, &txq->txpkts1_pkts,
3797 	    "# of frames tx'd using type1 txpkts work requests");
3798 
3799 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_enqueues",
3800 	    CTLFLAG_RD, &txq->r->enqueues,
3801 	    "# of enqueues to the mp_ring for this queue");
3802 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_drops",
3803 	    CTLFLAG_RD, &txq->r->drops,
3804 	    "# of drops in the mp_ring for this queue");
3805 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_starts",
3806 	    CTLFLAG_RD, &txq->r->starts,
3807 	    "# of normal consumer starts in the mp_ring for this queue");
3808 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_stalls",
3809 	    CTLFLAG_RD, &txq->r->stalls,
3810 	    "# of consumer stalls in the mp_ring for this queue");
3811 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_restarts",
3812 	    CTLFLAG_RD, &txq->r->restarts,
3813 	    "# of consumer restarts in the mp_ring for this queue");
3814 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_abdications",
3815 	    CTLFLAG_RD, &txq->r->abdications,
3816 	    "# of consumer abdications in the mp_ring for this queue");
3817 
3818 	return (0);
3819 }
3820 
3821 static int
3822 free_txq(struct vi_info *vi, struct sge_txq *txq)
3823 {
3824 	int rc;
3825 	struct adapter *sc = vi->pi->adapter;
3826 	struct sge_eq *eq = &txq->eq;
3827 
3828 	rc = free_eq(sc, eq);
3829 	if (rc)
3830 		return (rc);
3831 
3832 	sglist_free(txq->gl);
3833 	free(txq->sdesc, M_CXGBE);
3834 	mp_ring_free(txq->r);
3835 
3836 	bzero(txq, sizeof(*txq));
3837 	return (0);
3838 }
3839 
3840 static void
3841 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
3842 {
3843 	bus_addr_t *ba = arg;
3844 
3845 	KASSERT(nseg == 1,
3846 	    ("%s meant for single segment mappings only.", __func__));
3847 
3848 	*ba = error ? 0 : segs->ds_addr;
3849 }
3850 
3851 static inline void
3852 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
3853 {
3854 	uint32_t n, v;
3855 
3856 	n = IDXDIFF(fl->pidx / 8, fl->dbidx, fl->sidx);
3857 	MPASS(n > 0);
3858 
3859 	wmb();
3860 	v = fl->dbval | V_PIDX(n);
3861 	if (fl->udb)
3862 		*fl->udb = htole32(v);
3863 	else
3864 		t4_write_reg(sc, sc->sge_kdoorbell_reg, v);
3865 	IDXINCR(fl->dbidx, n, fl->sidx);
3866 }
3867 
3868 /*
3869  * Fills up the freelist by allocating up to 'n' buffers.  Buffers that are
3870  * recycled do not count towards this allocation budget.
3871  *
3872  * Returns non-zero to indicate that this freelist should be added to the list
3873  * of starving freelists.
3874  */
3875 static int
3876 refill_fl(struct adapter *sc, struct sge_fl *fl, int n)
3877 {
3878 	__be64 *d;
3879 	struct fl_sdesc *sd;
3880 	uintptr_t pa;
3881 	caddr_t cl;
3882 	struct cluster_layout *cll;
3883 	struct sw_zone_info *swz;
3884 	struct cluster_metadata *clm;
3885 	uint16_t max_pidx;
3886 	uint16_t hw_cidx = fl->hw_cidx;		/* stable snapshot */
3887 
3888 	FL_LOCK_ASSERT_OWNED(fl);
3889 
3890 	/*
3891 	 * We always stop at the beginning of the hardware descriptor that's just
3892 	 * before the one with the hw cidx.  This is to avoid hw pidx = hw cidx,
3893 	 * which would mean an empty freelist to the chip.
3894 	 */
3895 	max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1;
3896 	if (fl->pidx == max_pidx * 8)
3897 		return (0);
3898 
3899 	d = &fl->desc[fl->pidx];
3900 	sd = &fl->sdesc[fl->pidx];
3901 	cll = &fl->cll_def;	/* default layout */
3902 	swz = &sc->sge.sw_zone_info[cll->zidx];
3903 
3904 	while (n > 0) {
3905 
3906 		if (sd->cl != NULL) {
3907 
3908 			if (sd->nmbuf == 0) {
3909 				/*
3910 				 * Fast recycle without involving any atomics on
3911 				 * the cluster's metadata (if the cluster has
3912 				 * metadata).  This happens when all frames
3913 				 * received in the cluster were small enough to
3914 				 * fit within a single mbuf each.
3915 				 */
3916 				fl->cl_fast_recycled++;
3917 #ifdef INVARIANTS
3918 				clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
3919 				if (clm != NULL)
3920 					MPASS(clm->refcount == 1);
3921 #endif
3922 				goto recycled_fast;
3923 			}
3924 
3925 			/*
3926 			 * Cluster is guaranteed to have metadata.  Clusters
3927 			 * without metadata always take the fast recycle path
3928 			 * when they're recycled.
3929 			 */
3930 			clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
3931 			MPASS(clm != NULL);
3932 
3933 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
3934 				fl->cl_recycled++;
3935 				counter_u64_add(extfree_rels, 1);
3936 				goto recycled;
3937 			}
3938 			sd->cl = NULL;	/* gave up my reference */
3939 		}
3940 		MPASS(sd->cl == NULL);
3941 alloc:
3942 		cl = uma_zalloc(swz->zone, M_NOWAIT);
3943 		if (__predict_false(cl == NULL)) {
3944 			if (cll == &fl->cll_alt || fl->cll_alt.zidx == -1 ||
3945 			    fl->cll_def.zidx == fl->cll_alt.zidx)
3946 				break;
3947 
3948 			/* fall back to the safe zone */
3949 			cll = &fl->cll_alt;
3950 			swz = &sc->sge.sw_zone_info[cll->zidx];
3951 			goto alloc;
3952 		}
3953 		fl->cl_allocated++;
3954 		n--;
3955 
3956 		pa = pmap_kextract((vm_offset_t)cl);
3957 		pa += cll->region1;
3958 		sd->cl = cl;
3959 		sd->cll = *cll;
3960 		*d = htobe64(pa | cll->hwidx);
3961 		clm = cl_metadata(sc, fl, cll, cl);
3962 		if (clm != NULL) {
3963 recycled:
3964 #ifdef INVARIANTS
3965 			clm->sd = sd;
3966 #endif
3967 			clm->refcount = 1;
3968 		}
3969 		sd->nmbuf = 0;
3970 recycled_fast:
3971 		d++;
3972 		sd++;
3973 		if (__predict_false(++fl->pidx % 8 == 0)) {
3974 			uint16_t pidx = fl->pidx / 8;
3975 
3976 			if (__predict_false(pidx == fl->sidx)) {
3977 				fl->pidx = 0;
3978 				pidx = 0;
3979 				sd = fl->sdesc;
3980 				d = fl->desc;
3981 			}
3982 			if (pidx == max_pidx)
3983 				break;
3984 
3985 			if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4)
3986 				ring_fl_db(sc, fl);
3987 		}
3988 	}
3989 
3990 	if (fl->pidx / 8 != fl->dbidx)
3991 		ring_fl_db(sc, fl);
3992 
3993 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
3994 }
3995 
3996 /*
3997  * Attempt to refill all starving freelists.
3998  */
3999 static void
4000 refill_sfl(void *arg)
4001 {
4002 	struct adapter *sc = arg;
4003 	struct sge_fl *fl, *fl_temp;
4004 
4005 	mtx_assert(&sc->sfl_lock, MA_OWNED);
4006 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
4007 		FL_LOCK(fl);
4008 		refill_fl(sc, fl, 64);
4009 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
4010 			TAILQ_REMOVE(&sc->sfl, fl, link);
4011 			fl->flags &= ~FL_STARVING;
4012 		}
4013 		FL_UNLOCK(fl);
4014 	}
4015 
4016 	if (!TAILQ_EMPTY(&sc->sfl))
4017 		callout_schedule(&sc->sfl_callout, hz / 5);
4018 }
4019 
4020 static int
4021 alloc_fl_sdesc(struct sge_fl *fl)
4022 {
4023 
4024 	fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc), M_CXGBE,
4025 	    M_ZERO | M_WAITOK);
4026 
4027 	return (0);
4028 }
4029 
4030 static void
4031 free_fl_sdesc(struct adapter *sc, struct sge_fl *fl)
4032 {
4033 	struct fl_sdesc *sd;
4034 	struct cluster_metadata *clm;
4035 	struct cluster_layout *cll;
4036 	int i;
4037 
4038 	sd = fl->sdesc;
4039 	for (i = 0; i < fl->sidx * 8; i++, sd++) {
4040 		if (sd->cl == NULL)
4041 			continue;
4042 
4043 		cll = &sd->cll;
4044 		clm = cl_metadata(sc, fl, cll, sd->cl);
4045 		if (sd->nmbuf == 0)
4046 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4047 		else if (clm && atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4048 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4049 			counter_u64_add(extfree_rels, 1);
4050 		}
4051 		sd->cl = NULL;
4052 	}
4053 
4054 	free(fl->sdesc, M_CXGBE);
4055 	fl->sdesc = NULL;
4056 }
4057 
4058 static inline void
4059 get_pkt_gl(struct mbuf *m, struct sglist *gl)
4060 {
4061 	int rc;
4062 
4063 	M_ASSERTPKTHDR(m);
4064 
4065 	sglist_reset(gl);
4066 	rc = sglist_append_mbuf(gl, m);
4067 	if (__predict_false(rc != 0)) {
4068 		panic("%s: mbuf %p (%d segs) was vetted earlier but now fails "
4069 		    "with %d.", __func__, m, mbuf_nsegs(m), rc);
4070 	}
4071 
4072 	KASSERT(gl->sg_nseg == mbuf_nsegs(m),
4073 	    ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m,
4074 	    mbuf_nsegs(m), gl->sg_nseg));
4075 	KASSERT(gl->sg_nseg > 0 &&
4076 	    gl->sg_nseg <= (needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS),
4077 	    ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__,
4078 		gl->sg_nseg, needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS));
4079 }
4080 
4081 /*
4082  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
4083  */
4084 static inline u_int
4085 txpkt_len16(u_int nsegs, u_int tso)
4086 {
4087 	u_int n;
4088 
4089 	MPASS(nsegs > 0);
4090 
4091 	nsegs--; /* first segment is part of ulptx_sgl */
4092 	n = sizeof(struct fw_eth_tx_pkt_wr) + sizeof(struct cpl_tx_pkt_core) +
4093 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4094 	if (tso)
4095 		n += sizeof(struct cpl_tx_pkt_lso_core);
4096 
4097 	return (howmany(n, 16));
4098 }
4099 
4100 /*
4101  * len16 for a txpkt_vm WR with a GL.  Includes the firmware work
4102  * request header.
4103  */
4104 static inline u_int
4105 txpkt_vm_len16(u_int nsegs, u_int tso)
4106 {
4107 	u_int n;
4108 
4109 	MPASS(nsegs > 0);
4110 
4111 	nsegs--; /* first segment is part of ulptx_sgl */
4112 	n = sizeof(struct fw_eth_tx_pkt_vm_wr) +
4113 	    sizeof(struct cpl_tx_pkt_core) +
4114 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4115 	if (tso)
4116 		n += sizeof(struct cpl_tx_pkt_lso_core);
4117 
4118 	return (howmany(n, 16));
4119 }
4120 
4121 /*
4122  * len16 for a txpkts type 0 WR with a GL.  Does not include the firmware work
4123  * request header.
4124  */
4125 static inline u_int
4126 txpkts0_len16(u_int nsegs)
4127 {
4128 	u_int n;
4129 
4130 	MPASS(nsegs > 0);
4131 
4132 	nsegs--; /* first segment is part of ulptx_sgl */
4133 	n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) +
4134 	    sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) +
4135 	    8 * ((3 * nsegs) / 2 + (nsegs & 1));
4136 
4137 	return (howmany(n, 16));
4138 }
4139 
4140 /*
4141  * len16 for a txpkts type 1 WR with a GL.  Does not include the firmware work
4142  * request header.
4143  */
4144 static inline u_int
4145 txpkts1_len16(void)
4146 {
4147 	u_int n;
4148 
4149 	n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl);
4150 
4151 	return (howmany(n, 16));
4152 }
4153 
4154 static inline u_int
4155 imm_payload(u_int ndesc)
4156 {
4157 	u_int n;
4158 
4159 	n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) -
4160 	    sizeof(struct cpl_tx_pkt_core);
4161 
4162 	return (n);
4163 }
4164 
4165 /*
4166  * Write a VM txpkt WR for this packet to the hardware descriptors, update the
4167  * software descriptor, and advance the pidx.  It is guaranteed that enough
4168  * descriptors are available.
4169  *
4170  * The return value is the # of hardware descriptors used.
4171  */
4172 static u_int
4173 write_txpkt_vm_wr(struct adapter *sc, struct sge_txq *txq,
4174     struct fw_eth_tx_pkt_vm_wr *wr, struct mbuf *m0, u_int available)
4175 {
4176 	struct sge_eq *eq = &txq->eq;
4177 	struct tx_sdesc *txsd;
4178 	struct cpl_tx_pkt_core *cpl;
4179 	uint32_t ctrl;	/* used in many unrelated places */
4180 	uint64_t ctrl1;
4181 	int csum_type, len16, ndesc, pktlen, nsegs;
4182 	caddr_t dst;
4183 
4184 	TXQ_LOCK_ASSERT_OWNED(txq);
4185 	M_ASSERTPKTHDR(m0);
4186 	MPASS(available > 0 && available < eq->sidx);
4187 
4188 	len16 = mbuf_len16(m0);
4189 	nsegs = mbuf_nsegs(m0);
4190 	pktlen = m0->m_pkthdr.len;
4191 	ctrl = sizeof(struct cpl_tx_pkt_core);
4192 	if (needs_tso(m0))
4193 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4194 	ndesc = howmany(len16, EQ_ESIZE / 16);
4195 	MPASS(ndesc <= available);
4196 
4197 	/* Firmware work request header */
4198 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4199 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_VM_WR) |
4200 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4201 
4202 	ctrl = V_FW_WR_LEN16(len16);
4203 	wr->equiq_to_len16 = htobe32(ctrl);
4204 	wr->r3[0] = 0;
4205 	wr->r3[1] = 0;
4206 
4207 	/*
4208 	 * Copy over ethmacdst, ethmacsrc, ethtype, and vlantci.
4209 	 * vlantci is ignored unless the ethtype is 0x8100, so it's
4210 	 * simpler to always copy it rather than making it
4211 	 * conditional.  Also, it seems that we do not have to set
4212 	 * vlantci or fake the ethtype when doing VLAN tag insertion.
4213 	 */
4214 	m_copydata(m0, 0, sizeof(struct ether_header) + 2, wr->ethmacdst);
4215 
4216 	csum_type = -1;
4217 	if (needs_tso(m0)) {
4218 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4219 
4220 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4221 		    m0->m_pkthdr.l4hlen > 0,
4222 		    ("%s: mbuf %p needs TSO but missing header lengths",
4223 			__func__, m0));
4224 
4225 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4226 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4227 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4228 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4229 			ctrl |= V_LSO_ETHHDR_LEN(1);
4230 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4231 			ctrl |= F_LSO_IPV6;
4232 
4233 		lso->lso_ctrl = htobe32(ctrl);
4234 		lso->ipid_ofst = htobe16(0);
4235 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4236 		lso->seqno_offset = htobe32(0);
4237 		lso->len = htobe32(pktlen);
4238 
4239 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4240 			csum_type = TX_CSUM_TCPIP6;
4241 		else
4242 			csum_type = TX_CSUM_TCPIP;
4243 
4244 		cpl = (void *)(lso + 1);
4245 
4246 		txq->tso_wrs++;
4247 	} else {
4248 		if (m0->m_pkthdr.csum_flags & CSUM_IP_TCP)
4249 			csum_type = TX_CSUM_TCPIP;
4250 		else if (m0->m_pkthdr.csum_flags & CSUM_IP_UDP)
4251 			csum_type = TX_CSUM_UDPIP;
4252 		else if (m0->m_pkthdr.csum_flags & CSUM_IP6_TCP)
4253 			csum_type = TX_CSUM_TCPIP6;
4254 		else if (m0->m_pkthdr.csum_flags & CSUM_IP6_UDP)
4255 			csum_type = TX_CSUM_UDPIP6;
4256 #if defined(INET)
4257 		else if (m0->m_pkthdr.csum_flags & CSUM_IP) {
4258 			/*
4259 			 * XXX: The firmware appears to stomp on the
4260 			 * fragment/flags field of the IP header when
4261 			 * using TX_CSUM_IP.  Fall back to doing
4262 			 * software checksums.
4263 			 */
4264 			u_short *sump;
4265 			struct mbuf *m;
4266 			int offset;
4267 
4268 			m = m0;
4269 			offset = 0;
4270 			sump = m_advance(&m, &offset, m0->m_pkthdr.l2hlen +
4271 			    offsetof(struct ip, ip_sum));
4272 			*sump = in_cksum_skip(m0, m0->m_pkthdr.l2hlen +
4273 			    m0->m_pkthdr.l3hlen, m0->m_pkthdr.l2hlen);
4274 			m0->m_pkthdr.csum_flags &= ~CSUM_IP;
4275 		}
4276 #endif
4277 
4278 		cpl = (void *)(wr + 1);
4279 	}
4280 
4281 	/* Checksum offload */
4282 	ctrl1 = 0;
4283 	if (needs_l3_csum(m0) == 0)
4284 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4285 	if (csum_type >= 0) {
4286 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0,
4287 	    ("%s: mbuf %p needs checksum offload but missing header lengths",
4288 			__func__, m0));
4289 
4290 		if (chip_id(sc) <= CHELSIO_T5) {
4291 			ctrl1 |= V_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4292 			    ETHER_HDR_LEN);
4293 		} else {
4294 			ctrl1 |= V_T6_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4295 			    ETHER_HDR_LEN);
4296 		}
4297 		ctrl1 |= V_TXPKT_IPHDR_LEN(m0->m_pkthdr.l3hlen);
4298 		ctrl1 |= V_TXPKT_CSUM_TYPE(csum_type);
4299 	} else
4300 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4301 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4302 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4303 		txq->txcsum++;	/* some hardware assistance provided */
4304 
4305 	/* VLAN tag insertion */
4306 	if (needs_vlan_insertion(m0)) {
4307 		ctrl1 |= F_TXPKT_VLAN_VLD |
4308 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4309 		txq->vlan_insertion++;
4310 	}
4311 
4312 	/* CPL header */
4313 	cpl->ctrl0 = txq->cpl_ctrl0;
4314 	cpl->pack = 0;
4315 	cpl->len = htobe16(pktlen);
4316 	cpl->ctrl1 = htobe64(ctrl1);
4317 
4318 	/* SGL */
4319 	dst = (void *)(cpl + 1);
4320 
4321 	/*
4322 	 * A packet using TSO will use up an entire descriptor for the
4323 	 * firmware work request header, LSO CPL, and TX_PKT_XT CPL.
4324 	 * If this descriptor is the last descriptor in the ring, wrap
4325 	 * around to the front of the ring explicitly for the start of
4326 	 * the sgl.
4327 	 */
4328 	if (dst == (void *)&eq->desc[eq->sidx]) {
4329 		dst = (void *)&eq->desc[0];
4330 		write_gl_to_txd(txq, m0, &dst, 0);
4331 	} else
4332 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4333 	txq->sgl_wrs++;
4334 
4335 	txq->txpkt_wrs++;
4336 
4337 	txsd = &txq->sdesc[eq->pidx];
4338 	txsd->m = m0;
4339 	txsd->desc_used = ndesc;
4340 
4341 	return (ndesc);
4342 }
4343 
4344 /*
4345  * Write a txpkt WR for this packet to the hardware descriptors, update the
4346  * software descriptor, and advance the pidx.  It is guaranteed that enough
4347  * descriptors are available.
4348  *
4349  * The return value is the # of hardware descriptors used.
4350  */
4351 static u_int
4352 write_txpkt_wr(struct sge_txq *txq, struct fw_eth_tx_pkt_wr *wr,
4353     struct mbuf *m0, u_int available)
4354 {
4355 	struct sge_eq *eq = &txq->eq;
4356 	struct tx_sdesc *txsd;
4357 	struct cpl_tx_pkt_core *cpl;
4358 	uint32_t ctrl;	/* used in many unrelated places */
4359 	uint64_t ctrl1;
4360 	int len16, ndesc, pktlen, nsegs;
4361 	caddr_t dst;
4362 
4363 	TXQ_LOCK_ASSERT_OWNED(txq);
4364 	M_ASSERTPKTHDR(m0);
4365 	MPASS(available > 0 && available < eq->sidx);
4366 
4367 	len16 = mbuf_len16(m0);
4368 	nsegs = mbuf_nsegs(m0);
4369 	pktlen = m0->m_pkthdr.len;
4370 	ctrl = sizeof(struct cpl_tx_pkt_core);
4371 	if (needs_tso(m0))
4372 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4373 	else if (pktlen <= imm_payload(2) && available >= 2) {
4374 		/* Immediate data.  Recalculate len16 and set nsegs to 0. */
4375 		ctrl += pktlen;
4376 		len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) +
4377 		    sizeof(struct cpl_tx_pkt_core) + pktlen, 16);
4378 		nsegs = 0;
4379 	}
4380 	ndesc = howmany(len16, EQ_ESIZE / 16);
4381 	MPASS(ndesc <= available);
4382 
4383 	/* Firmware work request header */
4384 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4385 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
4386 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4387 
4388 	ctrl = V_FW_WR_LEN16(len16);
4389 	wr->equiq_to_len16 = htobe32(ctrl);
4390 	wr->r3 = 0;
4391 
4392 	if (needs_tso(m0)) {
4393 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4394 
4395 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4396 		    m0->m_pkthdr.l4hlen > 0,
4397 		    ("%s: mbuf %p needs TSO but missing header lengths",
4398 			__func__, m0));
4399 
4400 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4401 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4402 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4403 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4404 			ctrl |= V_LSO_ETHHDR_LEN(1);
4405 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4406 			ctrl |= F_LSO_IPV6;
4407 
4408 		lso->lso_ctrl = htobe32(ctrl);
4409 		lso->ipid_ofst = htobe16(0);
4410 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4411 		lso->seqno_offset = htobe32(0);
4412 		lso->len = htobe32(pktlen);
4413 
4414 		cpl = (void *)(lso + 1);
4415 
4416 		txq->tso_wrs++;
4417 	} else
4418 		cpl = (void *)(wr + 1);
4419 
4420 	/* Checksum offload */
4421 	ctrl1 = 0;
4422 	if (needs_l3_csum(m0) == 0)
4423 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4424 	if (needs_l4_csum(m0) == 0)
4425 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4426 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4427 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4428 		txq->txcsum++;	/* some hardware assistance provided */
4429 
4430 	/* VLAN tag insertion */
4431 	if (needs_vlan_insertion(m0)) {
4432 		ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4433 		txq->vlan_insertion++;
4434 	}
4435 
4436 	/* CPL header */
4437 	cpl->ctrl0 = txq->cpl_ctrl0;
4438 	cpl->pack = 0;
4439 	cpl->len = htobe16(pktlen);
4440 	cpl->ctrl1 = htobe64(ctrl1);
4441 
4442 	/* SGL */
4443 	dst = (void *)(cpl + 1);
4444 	if (nsegs > 0) {
4445 
4446 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4447 		txq->sgl_wrs++;
4448 	} else {
4449 		struct mbuf *m;
4450 
4451 		for (m = m0; m != NULL; m = m->m_next) {
4452 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4453 #ifdef INVARIANTS
4454 			pktlen -= m->m_len;
4455 #endif
4456 		}
4457 #ifdef INVARIANTS
4458 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
4459 #endif
4460 		txq->imm_wrs++;
4461 	}
4462 
4463 	txq->txpkt_wrs++;
4464 
4465 	txsd = &txq->sdesc[eq->pidx];
4466 	txsd->m = m0;
4467 	txsd->desc_used = ndesc;
4468 
4469 	return (ndesc);
4470 }
4471 
4472 static int
4473 try_txpkts(struct mbuf *m, struct mbuf *n, struct txpkts *txp, u_int available)
4474 {
4475 	u_int needed, nsegs1, nsegs2, l1, l2;
4476 
4477 	if (cannot_use_txpkts(m) || cannot_use_txpkts(n))
4478 		return (1);
4479 
4480 	nsegs1 = mbuf_nsegs(m);
4481 	nsegs2 = mbuf_nsegs(n);
4482 	if (nsegs1 + nsegs2 == 2) {
4483 		txp->wr_type = 1;
4484 		l1 = l2 = txpkts1_len16();
4485 	} else {
4486 		txp->wr_type = 0;
4487 		l1 = txpkts0_len16(nsegs1);
4488 		l2 = txpkts0_len16(nsegs2);
4489 	}
4490 	txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) + l1 + l2;
4491 	needed = howmany(txp->len16, EQ_ESIZE / 16);
4492 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4493 		return (1);
4494 
4495 	txp->plen = m->m_pkthdr.len + n->m_pkthdr.len;
4496 	if (txp->plen > 65535)
4497 		return (1);
4498 
4499 	txp->npkt = 2;
4500 	set_mbuf_len16(m, l1);
4501 	set_mbuf_len16(n, l2);
4502 
4503 	return (0);
4504 }
4505 
4506 static int
4507 add_to_txpkts(struct mbuf *m, struct txpkts *txp, u_int available)
4508 {
4509 	u_int plen, len16, needed, nsegs;
4510 
4511 	MPASS(txp->wr_type == 0 || txp->wr_type == 1);
4512 
4513 	nsegs = mbuf_nsegs(m);
4514 	if (needs_tso(m) || (txp->wr_type == 1 && nsegs != 1))
4515 		return (1);
4516 
4517 	plen = txp->plen + m->m_pkthdr.len;
4518 	if (plen > 65535)
4519 		return (1);
4520 
4521 	if (txp->wr_type == 0)
4522 		len16 = txpkts0_len16(nsegs);
4523 	else
4524 		len16 = txpkts1_len16();
4525 	needed = howmany(txp->len16 + len16, EQ_ESIZE / 16);
4526 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4527 		return (1);
4528 
4529 	txp->npkt++;
4530 	txp->plen = plen;
4531 	txp->len16 += len16;
4532 	set_mbuf_len16(m, len16);
4533 
4534 	return (0);
4535 }
4536 
4537 /*
4538  * Write a txpkts WR for the packets in txp to the hardware descriptors, update
4539  * the software descriptor, and advance the pidx.  It is guaranteed that enough
4540  * descriptors are available.
4541  *
4542  * The return value is the # of hardware descriptors used.
4543  */
4544 static u_int
4545 write_txpkts_wr(struct sge_txq *txq, struct fw_eth_tx_pkts_wr *wr,
4546     struct mbuf *m0, const struct txpkts *txp, u_int available)
4547 {
4548 	struct sge_eq *eq = &txq->eq;
4549 	struct tx_sdesc *txsd;
4550 	struct cpl_tx_pkt_core *cpl;
4551 	uint32_t ctrl;
4552 	uint64_t ctrl1;
4553 	int ndesc, checkwrap;
4554 	struct mbuf *m;
4555 	void *flitp;
4556 
4557 	TXQ_LOCK_ASSERT_OWNED(txq);
4558 	MPASS(txp->npkt > 0);
4559 	MPASS(txp->plen < 65536);
4560 	MPASS(m0 != NULL);
4561 	MPASS(m0->m_nextpkt != NULL);
4562 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
4563 	MPASS(available > 0 && available < eq->sidx);
4564 
4565 	ndesc = howmany(txp->len16, EQ_ESIZE / 16);
4566 	MPASS(ndesc <= available);
4567 
4568 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4569 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
4570 	ctrl = V_FW_WR_LEN16(txp->len16);
4571 	wr->equiq_to_len16 = htobe32(ctrl);
4572 	wr->plen = htobe16(txp->plen);
4573 	wr->npkt = txp->npkt;
4574 	wr->r3 = 0;
4575 	wr->type = txp->wr_type;
4576 	flitp = wr + 1;
4577 
4578 	/*
4579 	 * At this point we are 16B into a hardware descriptor.  If checkwrap is
4580 	 * set then we know the WR is going to wrap around somewhere.  We'll
4581 	 * check for that at appropriate points.
4582 	 */
4583 	checkwrap = eq->sidx - ndesc < eq->pidx;
4584 	for (m = m0; m != NULL; m = m->m_nextpkt) {
4585 		if (txp->wr_type == 0) {
4586 			struct ulp_txpkt *ulpmc;
4587 			struct ulptx_idata *ulpsc;
4588 
4589 			/* ULP master command */
4590 			ulpmc = flitp;
4591 			ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
4592 			    V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid));
4593 			ulpmc->len = htobe32(mbuf_len16(m));
4594 
4595 			/* ULP subcommand */
4596 			ulpsc = (void *)(ulpmc + 1);
4597 			ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
4598 			    F_ULP_TX_SC_MORE);
4599 			ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
4600 
4601 			cpl = (void *)(ulpsc + 1);
4602 			if (checkwrap &&
4603 			    (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx])
4604 				cpl = (void *)&eq->desc[0];
4605 		} else {
4606 			cpl = flitp;
4607 		}
4608 
4609 		/* Checksum offload */
4610 		ctrl1 = 0;
4611 		if (needs_l3_csum(m) == 0)
4612 			ctrl1 |= F_TXPKT_IPCSUM_DIS;
4613 		if (needs_l4_csum(m) == 0)
4614 			ctrl1 |= F_TXPKT_L4CSUM_DIS;
4615 		if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4616 		    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4617 			txq->txcsum++;	/* some hardware assistance provided */
4618 
4619 		/* VLAN tag insertion */
4620 		if (needs_vlan_insertion(m)) {
4621 			ctrl1 |= F_TXPKT_VLAN_VLD |
4622 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
4623 			txq->vlan_insertion++;
4624 		}
4625 
4626 		/* CPL header */
4627 		cpl->ctrl0 = txq->cpl_ctrl0;
4628 		cpl->pack = 0;
4629 		cpl->len = htobe16(m->m_pkthdr.len);
4630 		cpl->ctrl1 = htobe64(ctrl1);
4631 
4632 		flitp = cpl + 1;
4633 		if (checkwrap &&
4634 		    (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
4635 			flitp = (void *)&eq->desc[0];
4636 
4637 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap);
4638 
4639 	}
4640 
4641 	if (txp->wr_type == 0) {
4642 		txq->txpkts0_pkts += txp->npkt;
4643 		txq->txpkts0_wrs++;
4644 	} else {
4645 		txq->txpkts1_pkts += txp->npkt;
4646 		txq->txpkts1_wrs++;
4647 	}
4648 
4649 	txsd = &txq->sdesc[eq->pidx];
4650 	txsd->m = m0;
4651 	txsd->desc_used = ndesc;
4652 
4653 	return (ndesc);
4654 }
4655 
4656 /*
4657  * If the SGL ends on an address that is not 16 byte aligned, this function will
4658  * add a 0 filled flit at the end.
4659  */
4660 static void
4661 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap)
4662 {
4663 	struct sge_eq *eq = &txq->eq;
4664 	struct sglist *gl = txq->gl;
4665 	struct sglist_seg *seg;
4666 	__be64 *flitp, *wrap;
4667 	struct ulptx_sgl *usgl;
4668 	int i, nflits, nsegs;
4669 
4670 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
4671 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
4672 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
4673 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
4674 
4675 	get_pkt_gl(m, gl);
4676 	nsegs = gl->sg_nseg;
4677 	MPASS(nsegs > 0);
4678 
4679 	nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
4680 	flitp = (__be64 *)(*to);
4681 	wrap = (__be64 *)(&eq->desc[eq->sidx]);
4682 	seg = &gl->sg_segs[0];
4683 	usgl = (void *)flitp;
4684 
4685 	/*
4686 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
4687 	 * ring, so we're at least 16 bytes away from the status page.  There is
4688 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
4689 	 */
4690 
4691 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
4692 	    V_ULPTX_NSGE(nsegs));
4693 	usgl->len0 = htobe32(seg->ss_len);
4694 	usgl->addr0 = htobe64(seg->ss_paddr);
4695 	seg++;
4696 
4697 	if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) {
4698 
4699 		/* Won't wrap around at all */
4700 
4701 		for (i = 0; i < nsegs - 1; i++, seg++) {
4702 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
4703 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
4704 		}
4705 		if (i & 1)
4706 			usgl->sge[i / 2].len[1] = htobe32(0);
4707 		flitp += nflits;
4708 	} else {
4709 
4710 		/* Will wrap somewhere in the rest of the SGL */
4711 
4712 		/* 2 flits already written, write the rest flit by flit */
4713 		flitp = (void *)(usgl + 1);
4714 		for (i = 0; i < nflits - 2; i++) {
4715 			if (flitp == wrap)
4716 				flitp = (void *)eq->desc;
4717 			*flitp++ = get_flit(seg, nsegs - 1, i);
4718 		}
4719 	}
4720 
4721 	if (nflits & 1) {
4722 		MPASS(((uintptr_t)flitp) & 0xf);
4723 		*flitp++ = 0;
4724 	}
4725 
4726 	MPASS((((uintptr_t)flitp) & 0xf) == 0);
4727 	if (__predict_false(flitp == wrap))
4728 		*to = (void *)eq->desc;
4729 	else
4730 		*to = (void *)flitp;
4731 }
4732 
4733 static inline void
4734 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
4735 {
4736 
4737 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
4738 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
4739 
4740 	if (__predict_true((uintptr_t)(*to) + len <=
4741 	    (uintptr_t)&eq->desc[eq->sidx])) {
4742 		bcopy(from, *to, len);
4743 		(*to) += len;
4744 	} else {
4745 		int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
4746 
4747 		bcopy(from, *to, portion);
4748 		from += portion;
4749 		portion = len - portion;	/* remaining */
4750 		bcopy(from, (void *)eq->desc, portion);
4751 		(*to) = (caddr_t)eq->desc + portion;
4752 	}
4753 }
4754 
4755 static inline void
4756 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n)
4757 {
4758 	u_int db;
4759 
4760 	MPASS(n > 0);
4761 
4762 	db = eq->doorbells;
4763 	if (n > 1)
4764 		clrbit(&db, DOORBELL_WCWR);
4765 	wmb();
4766 
4767 	switch (ffs(db) - 1) {
4768 	case DOORBELL_UDB:
4769 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
4770 		break;
4771 
4772 	case DOORBELL_WCWR: {
4773 		volatile uint64_t *dst, *src;
4774 		int i;
4775 
4776 		/*
4777 		 * Queues whose 128B doorbell segment fits in the page do not
4778 		 * use relative qid (udb_qid is always 0).  Only queues with
4779 		 * doorbell segments can do WCWR.
4780 		 */
4781 		KASSERT(eq->udb_qid == 0 && n == 1,
4782 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
4783 		    __func__, eq->doorbells, n, eq->dbidx, eq));
4784 
4785 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
4786 		    UDBS_DB_OFFSET);
4787 		i = eq->dbidx;
4788 		src = (void *)&eq->desc[i];
4789 		while (src != (void *)&eq->desc[i + 1])
4790 			*dst++ = *src++;
4791 		wmb();
4792 		break;
4793 	}
4794 
4795 	case DOORBELL_UDBWC:
4796 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
4797 		wmb();
4798 		break;
4799 
4800 	case DOORBELL_KDB:
4801 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
4802 		    V_QID(eq->cntxt_id) | V_PIDX(n));
4803 		break;
4804 	}
4805 
4806 	IDXINCR(eq->dbidx, n, eq->sidx);
4807 }
4808 
4809 static inline u_int
4810 reclaimable_tx_desc(struct sge_eq *eq)
4811 {
4812 	uint16_t hw_cidx;
4813 
4814 	hw_cidx = read_hw_cidx(eq);
4815 	return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx));
4816 }
4817 
4818 static inline u_int
4819 total_available_tx_desc(struct sge_eq *eq)
4820 {
4821 	uint16_t hw_cidx, pidx;
4822 
4823 	hw_cidx = read_hw_cidx(eq);
4824 	pidx = eq->pidx;
4825 
4826 	if (pidx == hw_cidx)
4827 		return (eq->sidx - 1);
4828 	else
4829 		return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1);
4830 }
4831 
4832 static inline uint16_t
4833 read_hw_cidx(struct sge_eq *eq)
4834 {
4835 	struct sge_qstat *spg = (void *)&eq->desc[eq->sidx];
4836 	uint16_t cidx = spg->cidx;	/* stable snapshot */
4837 
4838 	return (be16toh(cidx));
4839 }
4840 
4841 /*
4842  * Reclaim 'n' descriptors approximately.
4843  */
4844 static u_int
4845 reclaim_tx_descs(struct sge_txq *txq, u_int n)
4846 {
4847 	struct tx_sdesc *txsd;
4848 	struct sge_eq *eq = &txq->eq;
4849 	u_int can_reclaim, reclaimed;
4850 
4851 	TXQ_LOCK_ASSERT_OWNED(txq);
4852 	MPASS(n > 0);
4853 
4854 	reclaimed = 0;
4855 	can_reclaim = reclaimable_tx_desc(eq);
4856 	while (can_reclaim && reclaimed < n) {
4857 		int ndesc;
4858 		struct mbuf *m, *nextpkt;
4859 
4860 		txsd = &txq->sdesc[eq->cidx];
4861 		ndesc = txsd->desc_used;
4862 
4863 		/* Firmware doesn't return "partial" credits. */
4864 		KASSERT(can_reclaim >= ndesc,
4865 		    ("%s: unexpected number of credits: %d, %d",
4866 		    __func__, can_reclaim, ndesc));
4867 
4868 		for (m = txsd->m; m != NULL; m = nextpkt) {
4869 			nextpkt = m->m_nextpkt;
4870 			m->m_nextpkt = NULL;
4871 			m_freem(m);
4872 		}
4873 		reclaimed += ndesc;
4874 		can_reclaim -= ndesc;
4875 		IDXINCR(eq->cidx, ndesc, eq->sidx);
4876 	}
4877 
4878 	return (reclaimed);
4879 }
4880 
4881 static void
4882 tx_reclaim(void *arg, int n)
4883 {
4884 	struct sge_txq *txq = arg;
4885 	struct sge_eq *eq = &txq->eq;
4886 
4887 	do {
4888 		if (TXQ_TRYLOCK(txq) == 0)
4889 			break;
4890 		n = reclaim_tx_descs(txq, 32);
4891 		if (eq->cidx == eq->pidx)
4892 			eq->equeqidx = eq->pidx;
4893 		TXQ_UNLOCK(txq);
4894 	} while (n > 0);
4895 }
4896 
4897 static __be64
4898 get_flit(struct sglist_seg *segs, int nsegs, int idx)
4899 {
4900 	int i = (idx / 3) * 2;
4901 
4902 	switch (idx % 3) {
4903 	case 0: {
4904 		__be64 rc;
4905 
4906 		rc = htobe32(segs[i].ss_len);
4907 		if (i + 1 < nsegs)
4908 			rc |= (uint64_t)htobe32(segs[i + 1].ss_len) << 32;
4909 
4910 		return (rc);
4911 	}
4912 	case 1:
4913 		return (htobe64(segs[i].ss_paddr));
4914 	case 2:
4915 		return (htobe64(segs[i + 1].ss_paddr));
4916 	}
4917 
4918 	return (0);
4919 }
4920 
4921 static void
4922 find_best_refill_source(struct adapter *sc, struct sge_fl *fl, int maxp)
4923 {
4924 	int8_t zidx, hwidx, idx;
4925 	uint16_t region1, region3;
4926 	int spare, spare_needed, n;
4927 	struct sw_zone_info *swz;
4928 	struct hw_buf_info *hwb, *hwb_list = &sc->sge.hw_buf_info[0];
4929 
4930 	/*
4931 	 * Buffer Packing: Look for PAGE_SIZE or larger zone which has a bufsize
4932 	 * large enough for the max payload and cluster metadata.  Otherwise
4933 	 * settle for the largest bufsize that leaves enough room in the cluster
4934 	 * for metadata.
4935 	 *
4936 	 * Without buffer packing: Look for the smallest zone which has a
4937 	 * bufsize large enough for the max payload.  Settle for the largest
4938 	 * bufsize available if there's nothing big enough for max payload.
4939 	 */
4940 	spare_needed = fl->flags & FL_BUF_PACKING ? CL_METADATA_SIZE : 0;
4941 	swz = &sc->sge.sw_zone_info[0];
4942 	hwidx = -1;
4943 	for (zidx = 0; zidx < SW_ZONE_SIZES; zidx++, swz++) {
4944 		if (swz->size > largest_rx_cluster) {
4945 			if (__predict_true(hwidx != -1))
4946 				break;
4947 
4948 			/*
4949 			 * This is a misconfiguration.  largest_rx_cluster is
4950 			 * preventing us from finding a refill source.  See
4951 			 * dev.t5nex.<n>.buffer_sizes to figure out why.
4952 			 */
4953 			device_printf(sc->dev, "largest_rx_cluster=%u leaves no"
4954 			    " refill source for fl %p (dma %u).  Ignored.\n",
4955 			    largest_rx_cluster, fl, maxp);
4956 		}
4957 		for (idx = swz->head_hwidx; idx != -1; idx = hwb->next) {
4958 			hwb = &hwb_list[idx];
4959 			spare = swz->size - hwb->size;
4960 			if (spare < spare_needed)
4961 				continue;
4962 
4963 			hwidx = idx;		/* best option so far */
4964 			if (hwb->size >= maxp) {
4965 
4966 				if ((fl->flags & FL_BUF_PACKING) == 0)
4967 					goto done; /* stop looking (not packing) */
4968 
4969 				if (swz->size >= safest_rx_cluster)
4970 					goto done; /* stop looking (packing) */
4971 			}
4972 			break;		/* keep looking, next zone */
4973 		}
4974 	}
4975 done:
4976 	/* A usable hwidx has been located. */
4977 	MPASS(hwidx != -1);
4978 	hwb = &hwb_list[hwidx];
4979 	zidx = hwb->zidx;
4980 	swz = &sc->sge.sw_zone_info[zidx];
4981 	region1 = 0;
4982 	region3 = swz->size - hwb->size;
4983 
4984 	/*
4985 	 * Stay within this zone and see if there is a better match when mbuf
4986 	 * inlining is allowed.  Remember that the hwidx's are sorted in
4987 	 * decreasing order of size (so in increasing order of spare area).
4988 	 */
4989 	for (idx = hwidx; idx != -1; idx = hwb->next) {
4990 		hwb = &hwb_list[idx];
4991 		spare = swz->size - hwb->size;
4992 
4993 		if (allow_mbufs_in_cluster == 0 || hwb->size < maxp)
4994 			break;
4995 
4996 		/*
4997 		 * Do not inline mbufs if doing so would violate the pad/pack
4998 		 * boundary alignment requirement.
4999 		 */
5000 		if (fl_pad && (MSIZE % sc->params.sge.pad_boundary) != 0)
5001 			continue;
5002 		if (fl->flags & FL_BUF_PACKING &&
5003 		    (MSIZE % sc->params.sge.pack_boundary) != 0)
5004 			continue;
5005 
5006 		if (spare < CL_METADATA_SIZE + MSIZE)
5007 			continue;
5008 		n = (spare - CL_METADATA_SIZE) / MSIZE;
5009 		if (n > howmany(hwb->size, maxp))
5010 			break;
5011 
5012 		hwidx = idx;
5013 		if (fl->flags & FL_BUF_PACKING) {
5014 			region1 = n * MSIZE;
5015 			region3 = spare - region1;
5016 		} else {
5017 			region1 = MSIZE;
5018 			region3 = spare - region1;
5019 			break;
5020 		}
5021 	}
5022 
5023 	KASSERT(zidx >= 0 && zidx < SW_ZONE_SIZES,
5024 	    ("%s: bad zone %d for fl %p, maxp %d", __func__, zidx, fl, maxp));
5025 	KASSERT(hwidx >= 0 && hwidx <= SGE_FLBUF_SIZES,
5026 	    ("%s: bad hwidx %d for fl %p, maxp %d", __func__, hwidx, fl, maxp));
5027 	KASSERT(region1 + sc->sge.hw_buf_info[hwidx].size + region3 ==
5028 	    sc->sge.sw_zone_info[zidx].size,
5029 	    ("%s: bad buffer layout for fl %p, maxp %d. "
5030 		"cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5031 		sc->sge.sw_zone_info[zidx].size, region1,
5032 		sc->sge.hw_buf_info[hwidx].size, region3));
5033 	if (fl->flags & FL_BUF_PACKING || region1 > 0) {
5034 		KASSERT(region3 >= CL_METADATA_SIZE,
5035 		    ("%s: no room for metadata.  fl %p, maxp %d; "
5036 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5037 		    sc->sge.sw_zone_info[zidx].size, region1,
5038 		    sc->sge.hw_buf_info[hwidx].size, region3));
5039 		KASSERT(region1 % MSIZE == 0,
5040 		    ("%s: bad mbuf region for fl %p, maxp %d. "
5041 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5042 		    sc->sge.sw_zone_info[zidx].size, region1,
5043 		    sc->sge.hw_buf_info[hwidx].size, region3));
5044 	}
5045 
5046 	fl->cll_def.zidx = zidx;
5047 	fl->cll_def.hwidx = hwidx;
5048 	fl->cll_def.region1 = region1;
5049 	fl->cll_def.region3 = region3;
5050 }
5051 
5052 static void
5053 find_safe_refill_source(struct adapter *sc, struct sge_fl *fl)
5054 {
5055 	struct sge *s = &sc->sge;
5056 	struct hw_buf_info *hwb;
5057 	struct sw_zone_info *swz;
5058 	int spare;
5059 	int8_t hwidx;
5060 
5061 	if (fl->flags & FL_BUF_PACKING)
5062 		hwidx = s->safe_hwidx2;	/* with room for metadata */
5063 	else if (allow_mbufs_in_cluster && s->safe_hwidx2 != -1) {
5064 		hwidx = s->safe_hwidx2;
5065 		hwb = &s->hw_buf_info[hwidx];
5066 		swz = &s->sw_zone_info[hwb->zidx];
5067 		spare = swz->size - hwb->size;
5068 
5069 		/* no good if there isn't room for an mbuf as well */
5070 		if (spare < CL_METADATA_SIZE + MSIZE)
5071 			hwidx = s->safe_hwidx1;
5072 	} else
5073 		hwidx = s->safe_hwidx1;
5074 
5075 	if (hwidx == -1) {
5076 		/* No fallback source */
5077 		fl->cll_alt.hwidx = -1;
5078 		fl->cll_alt.zidx = -1;
5079 
5080 		return;
5081 	}
5082 
5083 	hwb = &s->hw_buf_info[hwidx];
5084 	swz = &s->sw_zone_info[hwb->zidx];
5085 	spare = swz->size - hwb->size;
5086 	fl->cll_alt.hwidx = hwidx;
5087 	fl->cll_alt.zidx = hwb->zidx;
5088 	if (allow_mbufs_in_cluster &&
5089 	    (fl_pad == 0 || (MSIZE % sc->params.sge.pad_boundary) == 0))
5090 		fl->cll_alt.region1 = ((spare - CL_METADATA_SIZE) / MSIZE) * MSIZE;
5091 	else
5092 		fl->cll_alt.region1 = 0;
5093 	fl->cll_alt.region3 = spare - fl->cll_alt.region1;
5094 }
5095 
5096 static void
5097 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
5098 {
5099 	mtx_lock(&sc->sfl_lock);
5100 	FL_LOCK(fl);
5101 	if ((fl->flags & FL_DOOMED) == 0) {
5102 		fl->flags |= FL_STARVING;
5103 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
5104 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
5105 	}
5106 	FL_UNLOCK(fl);
5107 	mtx_unlock(&sc->sfl_lock);
5108 }
5109 
5110 static void
5111 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq)
5112 {
5113 	struct sge_wrq *wrq = (void *)eq;
5114 
5115 	atomic_readandclear_int(&eq->equiq);
5116 	taskqueue_enqueue(sc->tq[eq->tx_chan], &wrq->wrq_tx_task);
5117 }
5118 
5119 static void
5120 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq)
5121 {
5122 	struct sge_txq *txq = (void *)eq;
5123 
5124 	MPASS((eq->flags & EQ_TYPEMASK) == EQ_ETH);
5125 
5126 	atomic_readandclear_int(&eq->equiq);
5127 	mp_ring_check_drainage(txq->r, 0);
5128 	taskqueue_enqueue(sc->tq[eq->tx_chan], &txq->tx_reclaim_task);
5129 }
5130 
5131 static int
5132 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
5133     struct mbuf *m)
5134 {
5135 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
5136 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
5137 	struct adapter *sc = iq->adapter;
5138 	struct sge *s = &sc->sge;
5139 	struct sge_eq *eq;
5140 	static void (*h[])(struct adapter *, struct sge_eq *) = {NULL,
5141 		&handle_wrq_egr_update, &handle_eth_egr_update,
5142 		&handle_wrq_egr_update};
5143 
5144 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5145 	    rss->opcode));
5146 
5147 	eq = s->eqmap[qid - s->eq_start - s->eq_base];
5148 	(*h[eq->flags & EQ_TYPEMASK])(sc, eq);
5149 
5150 	return (0);
5151 }
5152 
5153 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
5154 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
5155     offsetof(struct cpl_fw6_msg, data));
5156 
5157 static int
5158 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
5159 {
5160 	struct adapter *sc = iq->adapter;
5161 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
5162 
5163 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5164 	    rss->opcode));
5165 
5166 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
5167 		const struct rss_header *rss2;
5168 
5169 		rss2 = (const struct rss_header *)&cpl->data[0];
5170 		return (t4_cpl_handler[rss2->opcode](iq, rss2, m));
5171 	}
5172 
5173 	return (t4_fw_msg_handler[cpl->type](sc, &cpl->data[0]));
5174 }
5175 
5176 /**
5177  *	t4_handle_wrerr_rpl - process a FW work request error message
5178  *	@adap: the adapter
5179  *	@rpl: start of the FW message
5180  */
5181 static int
5182 t4_handle_wrerr_rpl(struct adapter *adap, const __be64 *rpl)
5183 {
5184 	u8 opcode = *(const u8 *)rpl;
5185 	const struct fw_error_cmd *e = (const void *)rpl;
5186 	unsigned int i;
5187 
5188 	if (opcode != FW_ERROR_CMD) {
5189 		log(LOG_ERR,
5190 		    "%s: Received WRERR_RPL message with opcode %#x\n",
5191 		    device_get_nameunit(adap->dev), opcode);
5192 		return (EINVAL);
5193 	}
5194 	log(LOG_ERR, "%s: FW_ERROR (%s) ", device_get_nameunit(adap->dev),
5195 	    G_FW_ERROR_CMD_FATAL(be32toh(e->op_to_type)) ? "fatal" :
5196 	    "non-fatal");
5197 	switch (G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type))) {
5198 	case FW_ERROR_TYPE_EXCEPTION:
5199 		log(LOG_ERR, "exception info:\n");
5200 		for (i = 0; i < nitems(e->u.exception.info); i++)
5201 			log(LOG_ERR, "%s%08x", i == 0 ? "\t" : " ",
5202 			    be32toh(e->u.exception.info[i]));
5203 		log(LOG_ERR, "\n");
5204 		break;
5205 	case FW_ERROR_TYPE_HWMODULE:
5206 		log(LOG_ERR, "HW module regaddr %08x regval %08x\n",
5207 		    be32toh(e->u.hwmodule.regaddr),
5208 		    be32toh(e->u.hwmodule.regval));
5209 		break;
5210 	case FW_ERROR_TYPE_WR:
5211 		log(LOG_ERR, "WR cidx %d PF %d VF %d eqid %d hdr:\n",
5212 		    be16toh(e->u.wr.cidx),
5213 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.wr.pfn_vfn)),
5214 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.wr.pfn_vfn)),
5215 		    be32toh(e->u.wr.eqid));
5216 		for (i = 0; i < nitems(e->u.wr.wrhdr); i++)
5217 			log(LOG_ERR, "%s%02x", i == 0 ? "\t" : " ",
5218 			    e->u.wr.wrhdr[i]);
5219 		log(LOG_ERR, "\n");
5220 		break;
5221 	case FW_ERROR_TYPE_ACL:
5222 		log(LOG_ERR, "ACL cidx %d PF %d VF %d eqid %d %s",
5223 		    be16toh(e->u.acl.cidx),
5224 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.acl.pfn_vfn)),
5225 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.acl.pfn_vfn)),
5226 		    be32toh(e->u.acl.eqid),
5227 		    G_FW_ERROR_CMD_MV(be16toh(e->u.acl.mv_pkd)) ? "vlanid" :
5228 		    "MAC");
5229 		for (i = 0; i < nitems(e->u.acl.val); i++)
5230 			log(LOG_ERR, " %02x", e->u.acl.val[i]);
5231 		log(LOG_ERR, "\n");
5232 		break;
5233 	default:
5234 		log(LOG_ERR, "type %#x\n",
5235 		    G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type)));
5236 		return (EINVAL);
5237 	}
5238 	return (0);
5239 }
5240 
5241 static int
5242 sysctl_uint16(SYSCTL_HANDLER_ARGS)
5243 {
5244 	uint16_t *id = arg1;
5245 	int i = *id;
5246 
5247 	return sysctl_handle_int(oidp, &i, 0, req);
5248 }
5249 
5250 static int
5251 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
5252 {
5253 	struct sge *s = arg1;
5254 	struct hw_buf_info *hwb = &s->hw_buf_info[0];
5255 	struct sw_zone_info *swz = &s->sw_zone_info[0];
5256 	int i, rc;
5257 	struct sbuf sb;
5258 	char c;
5259 
5260 	sbuf_new(&sb, NULL, 32, SBUF_AUTOEXTEND);
5261 	for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
5262 		if (hwb->zidx >= 0 && swz[hwb->zidx].size <= largest_rx_cluster)
5263 			c = '*';
5264 		else
5265 			c = '\0';
5266 
5267 		sbuf_printf(&sb, "%u%c ", hwb->size, c);
5268 	}
5269 	sbuf_trim(&sb);
5270 	sbuf_finish(&sb);
5271 	rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
5272 	sbuf_delete(&sb);
5273 	return (rc);
5274 }
5275 
5276 static int
5277 sysctl_tc(SYSCTL_HANDLER_ARGS)
5278 {
5279 	struct vi_info *vi = arg1;
5280 	struct port_info *pi;
5281 	struct adapter *sc;
5282 	struct sge_txq *txq;
5283 	struct tx_cl_rl_params *tc;
5284 	int qidx = arg2, rc, tc_idx;
5285 	uint32_t fw_queue, fw_class;
5286 
5287 	MPASS(qidx >= 0 && qidx < vi->ntxq);
5288 	pi = vi->pi;
5289 	sc = pi->adapter;
5290 	txq = &sc->sge.txq[vi->first_txq + qidx];
5291 
5292 	tc_idx = txq->tc_idx;
5293 	rc = sysctl_handle_int(oidp, &tc_idx, 0, req);
5294 	if (rc != 0 || req->newptr == NULL)
5295 		return (rc);
5296 
5297 	if (sc->flags & IS_VF)
5298 		return (EPERM);
5299 
5300 	/* Note that -1 is legitimate input (it means unbind). */
5301 	if (tc_idx < -1 || tc_idx >= sc->chip_params->nsched_cls)
5302 		return (EINVAL);
5303 
5304 	mtx_lock(&sc->tc_lock);
5305 	if (tc_idx == txq->tc_idx) {
5306 		rc = 0;		/* No change, nothing to do. */
5307 		goto done;
5308 	}
5309 
5310 	fw_queue = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
5311 	    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_EQ_SCHEDCLASS_ETH) |
5312 	    V_FW_PARAMS_PARAM_YZ(txq->eq.cntxt_id);
5313 
5314 	if (tc_idx == -1)
5315 		fw_class = 0xffffffff;	/* Unbind. */
5316 	else {
5317 		/*
5318 		 * Bind to a different class.
5319 		 */
5320 		tc = &pi->sched_params->cl_rl[tc_idx];
5321 		if (tc->flags & TX_CLRL_ERROR) {
5322 			/* Previous attempt to set the cl-rl params failed. */
5323 			rc = EIO;
5324 			goto done;
5325 		} else {
5326 			/*
5327 			 * Ok to proceed.  Place a reference on the new class
5328 			 * while still holding on to the reference on the
5329 			 * previous class, if any.
5330 			 */
5331 			fw_class = tc_idx;
5332 			tc->refcount++;
5333 		}
5334 	}
5335 	mtx_unlock(&sc->tc_lock);
5336 
5337 	rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4stc");
5338 	if (rc)
5339 		return (rc);
5340 	rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &fw_queue, &fw_class);
5341 	end_synchronized_op(sc, 0);
5342 
5343 	mtx_lock(&sc->tc_lock);
5344 	if (rc == 0) {
5345 		if (txq->tc_idx != -1) {
5346 			tc = &pi->sched_params->cl_rl[txq->tc_idx];
5347 			MPASS(tc->refcount > 0);
5348 			tc->refcount--;
5349 		}
5350 		txq->tc_idx = tc_idx;
5351 	} else if (tc_idx != -1) {
5352 		tc = &pi->sched_params->cl_rl[tc_idx];
5353 		MPASS(tc->refcount > 0);
5354 		tc->refcount--;
5355 	}
5356 done:
5357 	mtx_unlock(&sc->tc_lock);
5358 	return (rc);
5359 }
5360