xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision 2c8d04d0228871c24017509cf039e7c5d97d97be)
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/md_var.h>
57 #include <vm/vm.h>
58 #include <vm/pmap.h>
59 #ifdef DEV_NETMAP
60 #include <machine/bus.h>
61 #include <sys/selinfo.h>
62 #include <net/if_var.h>
63 #include <net/netmap.h>
64 #include <dev/netmap/netmap_kern.h>
65 #endif
66 
67 #include "common/common.h"
68 #include "common/t4_regs.h"
69 #include "common/t4_regs_values.h"
70 #include "common/t4_msg.h"
71 #include "t4_l2t.h"
72 #include "t4_mp_ring.h"
73 
74 #ifdef T4_PKT_TIMESTAMP
75 #define RX_COPY_THRESHOLD (MINCLSIZE - 8)
76 #else
77 #define RX_COPY_THRESHOLD MINCLSIZE
78 #endif
79 
80 /*
81  * Ethernet frames are DMA'd at this byte offset into the freelist buffer.
82  * 0-7 are valid values.
83  */
84 int fl_pktshift = 2;
85 TUNABLE_INT("hw.cxgbe.fl_pktshift", &fl_pktshift);
86 
87 /*
88  * Pad ethernet payload up to this boundary.
89  * -1: driver should figure out a good value.
90  *  0: disable padding.
91  *  Any power of 2 from 32 to 4096 (both inclusive) is also a valid value.
92  */
93 int fl_pad = -1;
94 TUNABLE_INT("hw.cxgbe.fl_pad", &fl_pad);
95 
96 /*
97  * Status page length.
98  * -1: driver should figure out a good value.
99  *  64 or 128 are the only other valid values.
100  */
101 int spg_len = -1;
102 TUNABLE_INT("hw.cxgbe.spg_len", &spg_len);
103 
104 /*
105  * Congestion drops.
106  * -1: no congestion feedback (not recommended).
107  *  0: backpressure the channel instead of dropping packets right away.
108  *  1: no backpressure, drop packets for the congested queue immediately.
109  */
110 static int cong_drop = 0;
111 TUNABLE_INT("hw.cxgbe.cong_drop", &cong_drop);
112 
113 /*
114  * Deliver multiple frames in the same free list buffer if they fit.
115  * -1: let the driver decide whether to enable buffer packing or not.
116  *  0: disable buffer packing.
117  *  1: enable buffer packing.
118  */
119 static int buffer_packing = -1;
120 TUNABLE_INT("hw.cxgbe.buffer_packing", &buffer_packing);
121 
122 /*
123  * Start next frame in a packed buffer at this boundary.
124  * -1: driver should figure out a good value.
125  * T4: driver will ignore this and use the same value as fl_pad above.
126  * T5: 16, or a power of 2 from 64 to 4096 (both inclusive) is a valid value.
127  */
128 static int fl_pack = -1;
129 TUNABLE_INT("hw.cxgbe.fl_pack", &fl_pack);
130 
131 /*
132  * Allow the driver to create mbuf(s) in a cluster allocated for rx.
133  * 0: never; always allocate mbufs from the zone_mbuf UMA zone.
134  * 1: ok to create mbuf(s) within a cluster if there is room.
135  */
136 static int allow_mbufs_in_cluster = 1;
137 TUNABLE_INT("hw.cxgbe.allow_mbufs_in_cluster", &allow_mbufs_in_cluster);
138 
139 /*
140  * Largest rx cluster size that the driver is allowed to allocate.
141  */
142 static int largest_rx_cluster = MJUM16BYTES;
143 TUNABLE_INT("hw.cxgbe.largest_rx_cluster", &largest_rx_cluster);
144 
145 /*
146  * Size of cluster allocation that's most likely to succeed.  The driver will
147  * fall back to this size if it fails to allocate clusters larger than this.
148  */
149 static int safest_rx_cluster = PAGE_SIZE;
150 TUNABLE_INT("hw.cxgbe.safest_rx_cluster", &safest_rx_cluster);
151 
152 struct txpkts {
153 	u_int wr_type;		/* type 0 or type 1 */
154 	u_int npkt;		/* # of packets in this work request */
155 	u_int plen;		/* total payload (sum of all packets) */
156 	u_int len16;		/* # of 16B pieces used by this work request */
157 };
158 
159 /* A packet's SGL.  This + m_pkthdr has all info needed for tx */
160 struct sgl {
161 	struct sglist sg;
162 	struct sglist_seg seg[TX_SGL_SEGS];
163 };
164 
165 static int service_iq(struct sge_iq *, int);
166 static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t);
167 static int t4_eth_rx(struct sge_iq *, const struct rss_header *, struct mbuf *);
168 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int);
169 static inline void init_fl(struct adapter *, struct sge_fl *, int, int, char *);
170 static inline void init_eq(struct adapter *, struct sge_eq *, int, int, uint8_t,
171     uint16_t, char *);
172 static int alloc_ring(struct adapter *, size_t, bus_dma_tag_t *, bus_dmamap_t *,
173     bus_addr_t *, void **);
174 static int free_ring(struct adapter *, bus_dma_tag_t, bus_dmamap_t, bus_addr_t,
175     void *);
176 static int alloc_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *,
177     int, int);
178 static int free_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *);
179 static void add_fl_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
180     struct sge_fl *);
181 static int alloc_fwq(struct adapter *);
182 static int free_fwq(struct adapter *);
183 static int alloc_mgmtq(struct adapter *);
184 static int free_mgmtq(struct adapter *);
185 static int alloc_rxq(struct vi_info *, struct sge_rxq *, int, int,
186     struct sysctl_oid *);
187 static int free_rxq(struct vi_info *, struct sge_rxq *);
188 #ifdef TCP_OFFLOAD
189 static int alloc_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *, int, int,
190     struct sysctl_oid *);
191 static int free_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *);
192 #endif
193 #ifdef DEV_NETMAP
194 static int alloc_nm_rxq(struct vi_info *, struct sge_nm_rxq *, int, int,
195     struct sysctl_oid *);
196 static int free_nm_rxq(struct vi_info *, struct sge_nm_rxq *);
197 static int alloc_nm_txq(struct vi_info *, struct sge_nm_txq *, int, int,
198     struct sysctl_oid *);
199 static int free_nm_txq(struct vi_info *, struct sge_nm_txq *);
200 #endif
201 static int ctrl_eq_alloc(struct adapter *, struct sge_eq *);
202 static int eth_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
203 #ifdef TCP_OFFLOAD
204 static int ofld_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
205 #endif
206 static int alloc_eq(struct adapter *, struct vi_info *, struct sge_eq *);
207 static int free_eq(struct adapter *, struct sge_eq *);
208 static int alloc_wrq(struct adapter *, struct vi_info *, struct sge_wrq *,
209     struct sysctl_oid *);
210 static int free_wrq(struct adapter *, struct sge_wrq *);
211 static int alloc_txq(struct vi_info *, struct sge_txq *, int,
212     struct sysctl_oid *);
213 static int free_txq(struct vi_info *, struct sge_txq *);
214 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int);
215 static inline void ring_fl_db(struct adapter *, struct sge_fl *);
216 static int refill_fl(struct adapter *, struct sge_fl *, int);
217 static void refill_sfl(void *);
218 static int alloc_fl_sdesc(struct sge_fl *);
219 static void free_fl_sdesc(struct adapter *, struct sge_fl *);
220 static void find_best_refill_source(struct adapter *, struct sge_fl *, int);
221 static void find_safe_refill_source(struct adapter *, struct sge_fl *);
222 static void add_fl_to_sfl(struct adapter *, struct sge_fl *);
223 
224 static inline void get_pkt_gl(struct mbuf *, struct sglist *);
225 static inline u_int txpkt_len16(u_int, u_int);
226 static inline u_int txpkts0_len16(u_int);
227 static inline u_int txpkts1_len16(void);
228 static u_int write_txpkt_wr(struct sge_txq *, struct fw_eth_tx_pkt_wr *,
229     struct mbuf *, u_int);
230 static int try_txpkts(struct mbuf *, struct mbuf *, struct txpkts *, u_int);
231 static int add_to_txpkts(struct mbuf *, struct txpkts *, u_int);
232 static u_int write_txpkts_wr(struct sge_txq *, struct fw_eth_tx_pkts_wr *,
233     struct mbuf *, const struct txpkts *, u_int);
234 static void write_gl_to_txd(struct sge_txq *, struct mbuf *, caddr_t *, int);
235 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int);
236 static inline void ring_eq_db(struct adapter *, struct sge_eq *, u_int);
237 static inline uint16_t read_hw_cidx(struct sge_eq *);
238 static inline u_int reclaimable_tx_desc(struct sge_eq *);
239 static inline u_int total_available_tx_desc(struct sge_eq *);
240 static u_int reclaim_tx_descs(struct sge_txq *, u_int);
241 static void tx_reclaim(void *, int);
242 static __be64 get_flit(struct sglist_seg *, int, int);
243 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
244     struct mbuf *);
245 static int handle_fw_msg(struct sge_iq *, const struct rss_header *,
246     struct mbuf *);
247 static int t4_handle_wrerr_rpl(struct adapter *, const __be64 *);
248 static void wrq_tx_drain(void *, int);
249 static void drain_wrq_wr_list(struct adapter *, struct sge_wrq *);
250 
251 static int sysctl_uint16(SYSCTL_HANDLER_ARGS);
252 static int sysctl_bufsizes(SYSCTL_HANDLER_ARGS);
253 static int sysctl_tc(SYSCTL_HANDLER_ARGS);
254 
255 static counter_u64_t extfree_refs;
256 static counter_u64_t extfree_rels;
257 
258 an_handler_t t4_an_handler;
259 fw_msg_handler_t t4_fw_msg_handler[NUM_FW6_TYPES];
260 cpl_handler_t t4_cpl_handler[NUM_CPL_CMDS];
261 
262 
263 static int
264 an_not_handled(struct sge_iq *iq, const struct rsp_ctrl *ctrl)
265 {
266 
267 #ifdef INVARIANTS
268 	panic("%s: async notification on iq %p (ctrl %p)", __func__, iq, ctrl);
269 #else
270 	log(LOG_ERR, "%s: async notification on iq %p (ctrl %p)\n",
271 	    __func__, iq, ctrl);
272 #endif
273 	return (EDOOFUS);
274 }
275 
276 int
277 t4_register_an_handler(an_handler_t h)
278 {
279 	uintptr_t *loc, new;
280 
281 	new = h ? (uintptr_t)h : (uintptr_t)an_not_handled;
282 	loc = (uintptr_t *) &t4_an_handler;
283 	atomic_store_rel_ptr(loc, new);
284 
285 	return (0);
286 }
287 
288 static int
289 fw_msg_not_handled(struct adapter *sc, const __be64 *rpl)
290 {
291 	const struct cpl_fw6_msg *cpl =
292 	    __containerof(rpl, struct cpl_fw6_msg, data[0]);
293 
294 #ifdef INVARIANTS
295 	panic("%s: fw_msg type %d", __func__, cpl->type);
296 #else
297 	log(LOG_ERR, "%s: fw_msg type %d\n", __func__, cpl->type);
298 #endif
299 	return (EDOOFUS);
300 }
301 
302 int
303 t4_register_fw_msg_handler(int type, fw_msg_handler_t h)
304 {
305 	uintptr_t *loc, new;
306 
307 	if (type >= nitems(t4_fw_msg_handler))
308 		return (EINVAL);
309 
310 	/*
311 	 * These are dispatched by the handler for FW{4|6}_CPL_MSG using the CPL
312 	 * handler dispatch table.  Reject any attempt to install a handler for
313 	 * this subtype.
314 	 */
315 	if (type == FW_TYPE_RSSCPL || type == FW6_TYPE_RSSCPL)
316 		return (EINVAL);
317 
318 	new = h ? (uintptr_t)h : (uintptr_t)fw_msg_not_handled;
319 	loc = (uintptr_t *) &t4_fw_msg_handler[type];
320 	atomic_store_rel_ptr(loc, new);
321 
322 	return (0);
323 }
324 
325 static int
326 cpl_not_handled(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
327 {
328 
329 #ifdef INVARIANTS
330 	panic("%s: opcode 0x%02x on iq %p with payload %p",
331 	    __func__, rss->opcode, iq, m);
332 #else
333 	log(LOG_ERR, "%s: opcode 0x%02x on iq %p with payload %p\n",
334 	    __func__, rss->opcode, iq, m);
335 	m_freem(m);
336 #endif
337 	return (EDOOFUS);
338 }
339 
340 int
341 t4_register_cpl_handler(int opcode, cpl_handler_t h)
342 {
343 	uintptr_t *loc, new;
344 
345 	if (opcode >= nitems(t4_cpl_handler))
346 		return (EINVAL);
347 
348 	new = h ? (uintptr_t)h : (uintptr_t)cpl_not_handled;
349 	loc = (uintptr_t *) &t4_cpl_handler[opcode];
350 	atomic_store_rel_ptr(loc, new);
351 
352 	return (0);
353 }
354 
355 /*
356  * Called on MOD_LOAD.  Validates and calculates the SGE tunables.
357  */
358 void
359 t4_sge_modload(void)
360 {
361 	int i;
362 
363 	if (fl_pktshift < 0 || fl_pktshift > 7) {
364 		printf("Invalid hw.cxgbe.fl_pktshift value (%d),"
365 		    " using 2 instead.\n", fl_pktshift);
366 		fl_pktshift = 2;
367 	}
368 
369 	if (spg_len != 64 && spg_len != 128) {
370 		int len;
371 
372 #if defined(__i386__) || defined(__amd64__)
373 		len = cpu_clflush_line_size > 64 ? 128 : 64;
374 #else
375 		len = 64;
376 #endif
377 		if (spg_len != -1) {
378 			printf("Invalid hw.cxgbe.spg_len value (%d),"
379 			    " using %d instead.\n", spg_len, len);
380 		}
381 		spg_len = len;
382 	}
383 
384 	if (cong_drop < -1 || cong_drop > 1) {
385 		printf("Invalid hw.cxgbe.cong_drop value (%d),"
386 		    " using 0 instead.\n", cong_drop);
387 		cong_drop = 0;
388 	}
389 
390 	extfree_refs = counter_u64_alloc(M_WAITOK);
391 	extfree_rels = counter_u64_alloc(M_WAITOK);
392 	counter_u64_zero(extfree_refs);
393 	counter_u64_zero(extfree_rels);
394 
395 	t4_an_handler = an_not_handled;
396 	for (i = 0; i < nitems(t4_fw_msg_handler); i++)
397 		t4_fw_msg_handler[i] = fw_msg_not_handled;
398 	for (i = 0; i < nitems(t4_cpl_handler); i++)
399 		t4_cpl_handler[i] = cpl_not_handled;
400 
401 	t4_register_cpl_handler(CPL_FW4_MSG, handle_fw_msg);
402 	t4_register_cpl_handler(CPL_FW6_MSG, handle_fw_msg);
403 	t4_register_cpl_handler(CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
404 	t4_register_cpl_handler(CPL_RX_PKT, t4_eth_rx);
405 	t4_register_fw_msg_handler(FW6_TYPE_CMD_RPL, t4_handle_fw_rpl);
406 	t4_register_fw_msg_handler(FW6_TYPE_WRERR_RPL, t4_handle_wrerr_rpl);
407 }
408 
409 void
410 t4_sge_modunload(void)
411 {
412 
413 	counter_u64_free(extfree_refs);
414 	counter_u64_free(extfree_rels);
415 }
416 
417 uint64_t
418 t4_sge_extfree_refs(void)
419 {
420 	uint64_t refs, rels;
421 
422 	rels = counter_u64_fetch(extfree_rels);
423 	refs = counter_u64_fetch(extfree_refs);
424 
425 	return (refs - rels);
426 }
427 
428 static inline void
429 setup_pad_and_pack_boundaries(struct adapter *sc)
430 {
431 	uint32_t v, m;
432 	int pad, pack;
433 
434 	pad = fl_pad;
435 	if (fl_pad < 32 || fl_pad > 4096 || !powerof2(fl_pad)) {
436 		/*
437 		 * If there is any chance that we might use buffer packing and
438 		 * the chip is a T4, then pick 64 as the pad/pack boundary.  Set
439 		 * it to 32 in all other cases.
440 		 */
441 		pad = is_t4(sc) && buffer_packing ? 64 : 32;
442 
443 		/*
444 		 * For fl_pad = 0 we'll still write a reasonable value to the
445 		 * register but all the freelists will opt out of padding.
446 		 * We'll complain here only if the user tried to set it to a
447 		 * value greater than 0 that was invalid.
448 		 */
449 		if (fl_pad > 0) {
450 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pad value"
451 			    " (%d), using %d instead.\n", fl_pad, pad);
452 		}
453 	}
454 	m = V_INGPADBOUNDARY(M_INGPADBOUNDARY);
455 	v = V_INGPADBOUNDARY(ilog2(pad) - 5);
456 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
457 
458 	if (is_t4(sc)) {
459 		if (fl_pack != -1 && fl_pack != pad) {
460 			/* Complain but carry on. */
461 			device_printf(sc->dev, "hw.cxgbe.fl_pack (%d) ignored,"
462 			    " using %d instead.\n", fl_pack, pad);
463 		}
464 		return;
465 	}
466 
467 	pack = fl_pack;
468 	if (fl_pack < 16 || fl_pack == 32 || fl_pack > 4096 ||
469 	    !powerof2(fl_pack)) {
470 		pack = max(sc->params.pci.mps, CACHE_LINE_SIZE);
471 		MPASS(powerof2(pack));
472 		if (pack < 16)
473 			pack = 16;
474 		if (pack == 32)
475 			pack = 64;
476 		if (pack > 4096)
477 			pack = 4096;
478 		if (fl_pack != -1) {
479 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pack value"
480 			    " (%d), using %d instead.\n", fl_pack, pack);
481 		}
482 	}
483 	m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY);
484 	if (pack == 16)
485 		v = V_INGPACKBOUNDARY(0);
486 	else
487 		v = V_INGPACKBOUNDARY(ilog2(pack) - 5);
488 
489 	MPASS(!is_t4(sc));	/* T4 doesn't have SGE_CONTROL2 */
490 	t4_set_reg_field(sc, A_SGE_CONTROL2, m, v);
491 }
492 
493 /*
494  * adap->params.vpd.cclk must be set up before this is called.
495  */
496 void
497 t4_tweak_chip_settings(struct adapter *sc)
498 {
499 	int i;
500 	uint32_t v, m;
501 	int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200};
502 	int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk;
503 	int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */
504 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
505 	static int sge_flbuf_sizes[] = {
506 		MCLBYTES,
507 #if MJUMPAGESIZE != MCLBYTES
508 		MJUMPAGESIZE,
509 		MJUMPAGESIZE - CL_METADATA_SIZE,
510 		MJUMPAGESIZE - 2 * MSIZE - CL_METADATA_SIZE,
511 #endif
512 		MJUM9BYTES,
513 		MJUM16BYTES,
514 		MCLBYTES - MSIZE - CL_METADATA_SIZE,
515 		MJUM9BYTES - CL_METADATA_SIZE,
516 		MJUM16BYTES - CL_METADATA_SIZE,
517 	};
518 
519 	KASSERT(sc->flags & MASTER_PF,
520 	    ("%s: trying to change chip settings when not master.", __func__));
521 
522 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
523 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
524 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
525 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
526 
527 	setup_pad_and_pack_boundaries(sc);
528 
529 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
530 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
531 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
532 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
533 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
534 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
535 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
536 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
537 	t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v);
538 
539 	KASSERT(nitems(sge_flbuf_sizes) <= SGE_FLBUF_SIZES,
540 	    ("%s: hw buffer size table too big", __func__));
541 	for (i = 0; i < min(nitems(sge_flbuf_sizes), SGE_FLBUF_SIZES); i++) {
542 		t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i),
543 		    sge_flbuf_sizes[i]);
544 	}
545 
546 	v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) |
547 	    V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]);
548 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v);
549 
550 	KASSERT(intr_timer[0] <= timer_max,
551 	    ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0],
552 	    timer_max));
553 	for (i = 1; i < nitems(intr_timer); i++) {
554 		KASSERT(intr_timer[i] >= intr_timer[i - 1],
555 		    ("%s: timers not listed in increasing order (%d)",
556 		    __func__, i));
557 
558 		while (intr_timer[i] > timer_max) {
559 			if (i == nitems(intr_timer) - 1) {
560 				intr_timer[i] = timer_max;
561 				break;
562 			}
563 			intr_timer[i] += intr_timer[i - 1];
564 			intr_timer[i] /= 2;
565 		}
566 	}
567 
568 	v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) |
569 	    V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1]));
570 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v);
571 	v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) |
572 	    V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3]));
573 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v);
574 	v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) |
575 	    V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5]));
576 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v);
577 
578 	/* 4K, 16K, 64K, 256K DDP "page sizes" */
579 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
580 	t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v);
581 
582 	m = v = F_TDDPTAGTCB;
583 	t4_set_reg_field(sc, A_ULP_RX_CTL, m, v);
584 
585 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
586 	    F_RESETDDPOFFSET;
587 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
588 	t4_set_reg_field(sc, A_TP_PARA_REG5, m, v);
589 }
590 
591 /*
592  * SGE wants the buffer to be at least 64B and then a multiple of 16.  If
593  * padding is is use the buffer's start and end need to be aligned to the pad
594  * boundary as well.  We'll just make sure that the size is a multiple of the
595  * boundary here, it is up to the buffer allocation code to make sure the start
596  * of the buffer is aligned as well.
597  */
598 static inline int
599 hwsz_ok(struct adapter *sc, int hwsz)
600 {
601 	int mask = fl_pad ? sc->params.sge.pad_boundary - 1 : 16 - 1;
602 
603 	return (hwsz >= 64 && (hwsz & mask) == 0);
604 }
605 
606 /*
607  * XXX: driver really should be able to deal with unexpected settings.
608  */
609 int
610 t4_read_chip_settings(struct adapter *sc)
611 {
612 	struct sge *s = &sc->sge;
613 	struct sge_params *sp = &sc->params.sge;
614 	int i, j, n, rc = 0;
615 	uint32_t m, v, r;
616 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
617 	static int sw_buf_sizes[] = {	/* Sorted by size */
618 		MCLBYTES,
619 #if MJUMPAGESIZE != MCLBYTES
620 		MJUMPAGESIZE,
621 #endif
622 		MJUM9BYTES,
623 		MJUM16BYTES
624 	};
625 	struct sw_zone_info *swz, *safe_swz;
626 	struct hw_buf_info *hwb;
627 
628 	t4_init_sge_params(sc);
629 
630 	m = F_RXPKTCPLMODE;
631 	v = F_RXPKTCPLMODE;
632 	r = t4_read_reg(sc, A_SGE_CONTROL);
633 	if ((r & m) != v) {
634 		device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r);
635 		rc = EINVAL;
636 	}
637 
638 	/*
639 	 * If this changes then every single use of PAGE_SHIFT in the driver
640 	 * needs to be carefully reviewed for PAGE_SHIFT vs sp->page_shift.
641 	 */
642 	if (sp->page_shift != PAGE_SHIFT) {
643 		device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r);
644 		rc = EINVAL;
645 	}
646 
647 	/* Filter out unusable hw buffer sizes entirely (mark with -2). */
648 	hwb = &s->hw_buf_info[0];
649 	for (i = 0; i < nitems(s->hw_buf_info); i++, hwb++) {
650 		r = t4_read_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i));
651 		hwb->size = r;
652 		hwb->zidx = hwsz_ok(sc, r) ? -1 : -2;
653 		hwb->next = -1;
654 	}
655 
656 	/*
657 	 * Create a sorted list in decreasing order of hw buffer sizes (and so
658 	 * increasing order of spare area) for each software zone.
659 	 *
660 	 * If padding is enabled then the start and end of the buffer must align
661 	 * to the pad boundary; if packing is enabled then they must align with
662 	 * the pack boundary as well.  Allocations from the cluster zones are
663 	 * aligned to min(size, 4K), so the buffer starts at that alignment and
664 	 * ends at hwb->size alignment.  If mbuf inlining is allowed the
665 	 * starting alignment will be reduced to MSIZE and the driver will
666 	 * exercise appropriate caution when deciding on the best buffer layout
667 	 * to use.
668 	 */
669 	n = 0;	/* no usable buffer size to begin with */
670 	swz = &s->sw_zone_info[0];
671 	safe_swz = NULL;
672 	for (i = 0; i < SW_ZONE_SIZES; i++, swz++) {
673 		int8_t head = -1, tail = -1;
674 
675 		swz->size = sw_buf_sizes[i];
676 		swz->zone = m_getzone(swz->size);
677 		swz->type = m_gettype(swz->size);
678 
679 		if (swz->size < PAGE_SIZE) {
680 			MPASS(powerof2(swz->size));
681 			if (fl_pad && (swz->size % sp->pad_boundary != 0))
682 				continue;
683 		}
684 
685 		if (swz->size == safest_rx_cluster)
686 			safe_swz = swz;
687 
688 		hwb = &s->hw_buf_info[0];
689 		for (j = 0; j < SGE_FLBUF_SIZES; j++, hwb++) {
690 			if (hwb->zidx != -1 || hwb->size > swz->size)
691 				continue;
692 #ifdef INVARIANTS
693 			if (fl_pad)
694 				MPASS(hwb->size % sp->pad_boundary == 0);
695 #endif
696 			hwb->zidx = i;
697 			if (head == -1)
698 				head = tail = j;
699 			else if (hwb->size < s->hw_buf_info[tail].size) {
700 				s->hw_buf_info[tail].next = j;
701 				tail = j;
702 			} else {
703 				int8_t *cur;
704 				struct hw_buf_info *t;
705 
706 				for (cur = &head; *cur != -1; cur = &t->next) {
707 					t = &s->hw_buf_info[*cur];
708 					if (hwb->size == t->size) {
709 						hwb->zidx = -2;
710 						break;
711 					}
712 					if (hwb->size > t->size) {
713 						hwb->next = *cur;
714 						*cur = j;
715 						break;
716 					}
717 				}
718 			}
719 		}
720 		swz->head_hwidx = head;
721 		swz->tail_hwidx = tail;
722 
723 		if (tail != -1) {
724 			n++;
725 			if (swz->size - s->hw_buf_info[tail].size >=
726 			    CL_METADATA_SIZE)
727 				sc->flags |= BUF_PACKING_OK;
728 		}
729 	}
730 	if (n == 0) {
731 		device_printf(sc->dev, "no usable SGE FL buffer size.\n");
732 		rc = EINVAL;
733 	}
734 
735 	s->safe_hwidx1 = -1;
736 	s->safe_hwidx2 = -1;
737 	if (safe_swz != NULL) {
738 		s->safe_hwidx1 = safe_swz->head_hwidx;
739 		for (i = safe_swz->head_hwidx; i != -1; i = hwb->next) {
740 			int spare;
741 
742 			hwb = &s->hw_buf_info[i];
743 #ifdef INVARIANTS
744 			if (fl_pad)
745 				MPASS(hwb->size % sp->pad_boundary == 0);
746 #endif
747 			spare = safe_swz->size - hwb->size;
748 			if (spare >= CL_METADATA_SIZE) {
749 				s->safe_hwidx2 = i;
750 				break;
751 			}
752 		}
753 	}
754 
755 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
756 	r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ);
757 	if (r != v) {
758 		device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r);
759 		rc = EINVAL;
760 	}
761 
762 	m = v = F_TDDPTAGTCB;
763 	r = t4_read_reg(sc, A_ULP_RX_CTL);
764 	if ((r & m) != v) {
765 		device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r);
766 		rc = EINVAL;
767 	}
768 
769 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
770 	    F_RESETDDPOFFSET;
771 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
772 	r = t4_read_reg(sc, A_TP_PARA_REG5);
773 	if ((r & m) != v) {
774 		device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r);
775 		rc = EINVAL;
776 	}
777 
778 	t4_init_tp_params(sc);
779 
780 	t4_read_mtu_tbl(sc, sc->params.mtus, NULL);
781 	t4_load_mtus(sc, sc->params.mtus, sc->params.a_wnd, sc->params.b_wnd);
782 
783 	return (rc);
784 }
785 
786 int
787 t4_create_dma_tag(struct adapter *sc)
788 {
789 	int rc;
790 
791 	rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
792 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
793 	    BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
794 	    NULL, &sc->dmat);
795 	if (rc != 0) {
796 		device_printf(sc->dev,
797 		    "failed to create main DMA tag: %d\n", rc);
798 	}
799 
800 	return (rc);
801 }
802 
803 void
804 t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
805     struct sysctl_oid_list *children)
806 {
807 	struct sge_params *sp = &sc->params.sge;
808 
809 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "buffer_sizes",
810 	    CTLTYPE_STRING | CTLFLAG_RD, &sc->sge, 0, sysctl_bufsizes, "A",
811 	    "freelist buffer sizes");
812 
813 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD,
814 	    NULL, sp->fl_pktshift, "payload DMA offset in rx buffer (bytes)");
815 
816 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD,
817 	    NULL, sp->pad_boundary, "payload pad boundary (bytes)");
818 
819 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD,
820 	    NULL, sp->spg_len, "status page size (bytes)");
821 
822 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD,
823 	    NULL, cong_drop, "congestion drop setting");
824 
825 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD,
826 	    NULL, sp->pack_boundary, "payload pack boundary (bytes)");
827 }
828 
829 int
830 t4_destroy_dma_tag(struct adapter *sc)
831 {
832 	if (sc->dmat)
833 		bus_dma_tag_destroy(sc->dmat);
834 
835 	return (0);
836 }
837 
838 /*
839  * Allocate and initialize the firmware event queue and the management queue.
840  *
841  * Returns errno on failure.  Resources allocated up to that point may still be
842  * allocated.  Caller is responsible for cleanup in case this function fails.
843  */
844 int
845 t4_setup_adapter_queues(struct adapter *sc)
846 {
847 	int rc;
848 
849 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
850 
851 	sysctl_ctx_init(&sc->ctx);
852 	sc->flags |= ADAP_SYSCTL_CTX;
853 
854 	/*
855 	 * Firmware event queue
856 	 */
857 	rc = alloc_fwq(sc);
858 	if (rc != 0)
859 		return (rc);
860 
861 	/*
862 	 * Management queue.  This is just a control queue that uses the fwq as
863 	 * its associated iq.
864 	 */
865 	rc = alloc_mgmtq(sc);
866 
867 	return (rc);
868 }
869 
870 /*
871  * Idempotent
872  */
873 int
874 t4_teardown_adapter_queues(struct adapter *sc)
875 {
876 
877 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
878 
879 	/* Do this before freeing the queue */
880 	if (sc->flags & ADAP_SYSCTL_CTX) {
881 		sysctl_ctx_free(&sc->ctx);
882 		sc->flags &= ~ADAP_SYSCTL_CTX;
883 	}
884 
885 	free_mgmtq(sc);
886 	free_fwq(sc);
887 
888 	return (0);
889 }
890 
891 static inline int
892 first_vector(struct vi_info *vi)
893 {
894 	struct adapter *sc = vi->pi->adapter;
895 
896 	if (sc->intr_count == 1)
897 		return (0);
898 
899 	return (vi->first_intr);
900 }
901 
902 /*
903  * Given an arbitrary "index," come up with an iq that can be used by other
904  * queues (of this VI) for interrupt forwarding, SGE egress updates, etc.
905  * The iq returned is guaranteed to be something that takes direct interrupts.
906  */
907 static struct sge_iq *
908 vi_intr_iq(struct vi_info *vi, int idx)
909 {
910 	struct adapter *sc = vi->pi->adapter;
911 	struct sge *s = &sc->sge;
912 	struct sge_iq *iq = NULL;
913 	int nintr, i;
914 
915 	if (sc->intr_count == 1)
916 		return (&sc->sge.fwq);
917 
918 	nintr = vi->nintr;
919 	KASSERT(nintr != 0,
920 	    ("%s: vi %p has no exclusive interrupts, total interrupts = %d",
921 	    __func__, vi, sc->intr_count));
922 	i = idx % nintr;
923 
924 	if (vi->flags & INTR_RXQ) {
925 	       	if (i < vi->nrxq) {
926 			iq = &s->rxq[vi->first_rxq + i].iq;
927 			goto done;
928 		}
929 		i -= vi->nrxq;
930 	}
931 #ifdef TCP_OFFLOAD
932 	if (vi->flags & INTR_OFLD_RXQ) {
933 	       	if (i < vi->nofldrxq) {
934 			iq = &s->ofld_rxq[vi->first_ofld_rxq + i].iq;
935 			goto done;
936 		}
937 		i -= vi->nofldrxq;
938 	}
939 #endif
940 	panic("%s: vi %p, intr_flags 0x%lx, idx %d, total intr %d\n", __func__,
941 	    vi, vi->flags & INTR_ALL, idx, nintr);
942 done:
943 	MPASS(iq != NULL);
944 	KASSERT(iq->flags & IQ_INTR,
945 	    ("%s: iq %p (vi %p, intr_flags 0x%lx, idx %d)", __func__, iq, vi,
946 	    vi->flags & INTR_ALL, idx));
947 	return (iq);
948 }
949 
950 /* Maximum payload that can be delivered with a single iq descriptor */
951 static inline int
952 mtu_to_max_payload(struct adapter *sc, int mtu, const int toe)
953 {
954 	int payload;
955 
956 #ifdef TCP_OFFLOAD
957 	if (toe) {
958 		payload = sc->tt.rx_coalesce ?
959 		    G_RXCOALESCESIZE(t4_read_reg(sc, A_TP_PARA_REG2)) : mtu;
960 	} else {
961 #endif
962 		/* large enough even when hw VLAN extraction is disabled */
963 		payload = sc->params.sge.fl_pktshift + ETHER_HDR_LEN +
964 		    ETHER_VLAN_ENCAP_LEN + mtu;
965 #ifdef TCP_OFFLOAD
966 	}
967 #endif
968 
969 	return (payload);
970 }
971 
972 int
973 t4_setup_vi_queues(struct vi_info *vi)
974 {
975 	int rc = 0, i, j, intr_idx, iqid;
976 	struct sge_rxq *rxq;
977 	struct sge_txq *txq;
978 	struct sge_wrq *ctrlq;
979 #ifdef TCP_OFFLOAD
980 	struct sge_ofld_rxq *ofld_rxq;
981 	struct sge_wrq *ofld_txq;
982 #endif
983 #ifdef DEV_NETMAP
984 	int saved_idx;
985 	struct sge_nm_rxq *nm_rxq;
986 	struct sge_nm_txq *nm_txq;
987 #endif
988 	char name[16];
989 	struct port_info *pi = vi->pi;
990 	struct adapter *sc = pi->adapter;
991 	struct ifnet *ifp = vi->ifp;
992 	struct sysctl_oid *oid = device_get_sysctl_tree(vi->dev);
993 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
994 	int maxp, mtu = ifp->if_mtu;
995 
996 	/* Interrupt vector to start from (when using multiple vectors) */
997 	intr_idx = first_vector(vi);
998 
999 #ifdef DEV_NETMAP
1000 	saved_idx = intr_idx;
1001 	if (ifp->if_capabilities & IFCAP_NETMAP) {
1002 
1003 		/* netmap is supported with direct interrupts only. */
1004 		MPASS(vi->flags & INTR_RXQ);
1005 
1006 		/*
1007 		 * We don't have buffers to back the netmap rx queues
1008 		 * right now so we create the queues in a way that
1009 		 * doesn't set off any congestion signal in the chip.
1010 		 */
1011 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_rxq",
1012 		    CTLFLAG_RD, NULL, "rx queues");
1013 		for_each_nm_rxq(vi, i, nm_rxq) {
1014 			rc = alloc_nm_rxq(vi, nm_rxq, intr_idx, i, oid);
1015 			if (rc != 0)
1016 				goto done;
1017 			intr_idx++;
1018 		}
1019 
1020 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_txq",
1021 		    CTLFLAG_RD, NULL, "tx queues");
1022 		for_each_nm_txq(vi, i, nm_txq) {
1023 			iqid = vi->first_nm_rxq + (i % vi->nnmrxq);
1024 			rc = alloc_nm_txq(vi, nm_txq, iqid, i, oid);
1025 			if (rc != 0)
1026 				goto done;
1027 		}
1028 	}
1029 
1030 	/* Normal rx queues and netmap rx queues share the same interrupts. */
1031 	intr_idx = saved_idx;
1032 #endif
1033 
1034 	/*
1035 	 * First pass over all NIC and TOE rx queues:
1036 	 * a) initialize iq and fl
1037 	 * b) allocate queue iff it will take direct interrupts.
1038 	 */
1039 	maxp = mtu_to_max_payload(sc, mtu, 0);
1040 	if (vi->flags & INTR_RXQ) {
1041 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
1042 		    CTLFLAG_RD, NULL, "rx queues");
1043 	}
1044 	for_each_rxq(vi, i, rxq) {
1045 
1046 		init_iq(&rxq->iq, sc, vi->tmr_idx, vi->pktc_idx, vi->qsize_rxq);
1047 
1048 		snprintf(name, sizeof(name), "%s rxq%d-fl",
1049 		    device_get_nameunit(vi->dev), i);
1050 		init_fl(sc, &rxq->fl, vi->qsize_rxq / 8, maxp, name);
1051 
1052 		if (vi->flags & INTR_RXQ) {
1053 			rxq->iq.flags |= IQ_INTR;
1054 			rc = alloc_rxq(vi, rxq, intr_idx, i, oid);
1055 			if (rc != 0)
1056 				goto done;
1057 			intr_idx++;
1058 		}
1059 	}
1060 #ifdef DEV_NETMAP
1061 	if (ifp->if_capabilities & IFCAP_NETMAP)
1062 		intr_idx = saved_idx + max(vi->nrxq, vi->nnmrxq);
1063 #endif
1064 #ifdef TCP_OFFLOAD
1065 	maxp = mtu_to_max_payload(sc, mtu, 1);
1066 	if (vi->flags & INTR_OFLD_RXQ) {
1067 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq",
1068 		    CTLFLAG_RD, NULL,
1069 		    "rx queues for offloaded TCP connections");
1070 	}
1071 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1072 
1073 		init_iq(&ofld_rxq->iq, sc, vi->tmr_idx, vi->pktc_idx,
1074 		    vi->qsize_rxq);
1075 
1076 		snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
1077 		    device_get_nameunit(vi->dev), i);
1078 		init_fl(sc, &ofld_rxq->fl, vi->qsize_rxq / 8, maxp, name);
1079 
1080 		if (vi->flags & INTR_OFLD_RXQ) {
1081 			ofld_rxq->iq.flags |= IQ_INTR;
1082 			rc = alloc_ofld_rxq(vi, ofld_rxq, intr_idx, i, oid);
1083 			if (rc != 0)
1084 				goto done;
1085 			intr_idx++;
1086 		}
1087 	}
1088 #endif
1089 
1090 	/*
1091 	 * Second pass over all NIC and TOE rx queues.  The queues forwarding
1092 	 * their interrupts are allocated now.
1093 	 */
1094 	j = 0;
1095 	if (!(vi->flags & INTR_RXQ)) {
1096 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
1097 		    CTLFLAG_RD, NULL, "rx queues");
1098 		for_each_rxq(vi, i, rxq) {
1099 			MPASS(!(rxq->iq.flags & IQ_INTR));
1100 
1101 			intr_idx = vi_intr_iq(vi, j)->abs_id;
1102 
1103 			rc = alloc_rxq(vi, rxq, intr_idx, i, oid);
1104 			if (rc != 0)
1105 				goto done;
1106 			j++;
1107 		}
1108 	}
1109 #ifdef TCP_OFFLOAD
1110 	if (vi->nofldrxq != 0 && !(vi->flags & INTR_OFLD_RXQ)) {
1111 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq",
1112 		    CTLFLAG_RD, NULL,
1113 		    "rx queues for offloaded TCP connections");
1114 		for_each_ofld_rxq(vi, i, ofld_rxq) {
1115 			MPASS(!(ofld_rxq->iq.flags & IQ_INTR));
1116 
1117 			intr_idx = vi_intr_iq(vi, j)->abs_id;
1118 
1119 			rc = alloc_ofld_rxq(vi, ofld_rxq, intr_idx, i, oid);
1120 			if (rc != 0)
1121 				goto done;
1122 			j++;
1123 		}
1124 	}
1125 #endif
1126 
1127 	/*
1128 	 * Now the tx queues.  Only one pass needed.
1129 	 */
1130 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "txq", CTLFLAG_RD,
1131 	    NULL, "tx queues");
1132 	j = 0;
1133 	for_each_txq(vi, i, txq) {
1134 		iqid = vi_intr_iq(vi, j)->cntxt_id;
1135 		snprintf(name, sizeof(name), "%s txq%d",
1136 		    device_get_nameunit(vi->dev), i);
1137 		init_eq(sc, &txq->eq, EQ_ETH, vi->qsize_txq, pi->tx_chan, iqid,
1138 		    name);
1139 
1140 		rc = alloc_txq(vi, txq, i, oid);
1141 		if (rc != 0)
1142 			goto done;
1143 		j++;
1144 	}
1145 #ifdef TCP_OFFLOAD
1146 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_txq",
1147 	    CTLFLAG_RD, NULL, "tx queues for offloaded TCP connections");
1148 	for_each_ofld_txq(vi, i, ofld_txq) {
1149 		struct sysctl_oid *oid2;
1150 
1151 		iqid = vi_intr_iq(vi, j)->cntxt_id;
1152 		snprintf(name, sizeof(name), "%s ofld_txq%d",
1153 		    device_get_nameunit(vi->dev), i);
1154 		init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq, pi->tx_chan,
1155 		    iqid, name);
1156 
1157 		snprintf(name, sizeof(name), "%d", i);
1158 		oid2 = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
1159 		    name, CTLFLAG_RD, NULL, "offload tx queue");
1160 
1161 		rc = alloc_wrq(sc, vi, ofld_txq, oid2);
1162 		if (rc != 0)
1163 			goto done;
1164 		j++;
1165 	}
1166 #endif
1167 
1168 	/*
1169 	 * Finally, the control queue.
1170 	 */
1171 	if (!IS_MAIN_VI(vi))
1172 		goto done;
1173 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ctrlq", CTLFLAG_RD,
1174 	    NULL, "ctrl queue");
1175 	ctrlq = &sc->sge.ctrlq[pi->port_id];
1176 	iqid = vi_intr_iq(vi, 0)->cntxt_id;
1177 	snprintf(name, sizeof(name), "%s ctrlq", device_get_nameunit(vi->dev));
1178 	init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, pi->tx_chan, iqid,
1179 	    name);
1180 	rc = alloc_wrq(sc, vi, ctrlq, oid);
1181 
1182 done:
1183 	if (rc)
1184 		t4_teardown_vi_queues(vi);
1185 
1186 	return (rc);
1187 }
1188 
1189 /*
1190  * Idempotent
1191  */
1192 int
1193 t4_teardown_vi_queues(struct vi_info *vi)
1194 {
1195 	int i;
1196 	struct port_info *pi = vi->pi;
1197 	struct adapter *sc = pi->adapter;
1198 	struct sge_rxq *rxq;
1199 	struct sge_txq *txq;
1200 #ifdef TCP_OFFLOAD
1201 	struct sge_ofld_rxq *ofld_rxq;
1202 	struct sge_wrq *ofld_txq;
1203 #endif
1204 #ifdef DEV_NETMAP
1205 	struct sge_nm_rxq *nm_rxq;
1206 	struct sge_nm_txq *nm_txq;
1207 #endif
1208 
1209 	/* Do this before freeing the queues */
1210 	if (vi->flags & VI_SYSCTL_CTX) {
1211 		sysctl_ctx_free(&vi->ctx);
1212 		vi->flags &= ~VI_SYSCTL_CTX;
1213 	}
1214 
1215 #ifdef DEV_NETMAP
1216 	if (vi->ifp->if_capabilities & IFCAP_NETMAP) {
1217 		for_each_nm_txq(vi, i, nm_txq) {
1218 			free_nm_txq(vi, nm_txq);
1219 		}
1220 
1221 		for_each_nm_rxq(vi, i, nm_rxq) {
1222 			free_nm_rxq(vi, nm_rxq);
1223 		}
1224 	}
1225 #endif
1226 
1227 	/*
1228 	 * Take down all the tx queues first, as they reference the rx queues
1229 	 * (for egress updates, etc.).
1230 	 */
1231 
1232 	if (IS_MAIN_VI(vi))
1233 		free_wrq(sc, &sc->sge.ctrlq[pi->port_id]);
1234 
1235 	for_each_txq(vi, i, txq) {
1236 		free_txq(vi, txq);
1237 	}
1238 #ifdef TCP_OFFLOAD
1239 	for_each_ofld_txq(vi, i, ofld_txq) {
1240 		free_wrq(sc, ofld_txq);
1241 	}
1242 #endif
1243 
1244 	/*
1245 	 * Then take down the rx queues that forward their interrupts, as they
1246 	 * reference other rx queues.
1247 	 */
1248 
1249 	for_each_rxq(vi, i, rxq) {
1250 		if ((rxq->iq.flags & IQ_INTR) == 0)
1251 			free_rxq(vi, rxq);
1252 	}
1253 #ifdef TCP_OFFLOAD
1254 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1255 		if ((ofld_rxq->iq.flags & IQ_INTR) == 0)
1256 			free_ofld_rxq(vi, ofld_rxq);
1257 	}
1258 #endif
1259 
1260 	/*
1261 	 * Then take down the rx queues that take direct interrupts.
1262 	 */
1263 
1264 	for_each_rxq(vi, i, rxq) {
1265 		if (rxq->iq.flags & IQ_INTR)
1266 			free_rxq(vi, rxq);
1267 	}
1268 #ifdef TCP_OFFLOAD
1269 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1270 		if (ofld_rxq->iq.flags & IQ_INTR)
1271 			free_ofld_rxq(vi, ofld_rxq);
1272 	}
1273 #endif
1274 
1275 	return (0);
1276 }
1277 
1278 /*
1279  * Deals with errors and the firmware event queue.  All data rx queues forward
1280  * their interrupt to the firmware event queue.
1281  */
1282 void
1283 t4_intr_all(void *arg)
1284 {
1285 	struct adapter *sc = arg;
1286 	struct sge_iq *fwq = &sc->sge.fwq;
1287 
1288 	t4_intr_err(arg);
1289 	if (atomic_cmpset_int(&fwq->state, IQS_IDLE, IQS_BUSY)) {
1290 		service_iq(fwq, 0);
1291 		atomic_cmpset_int(&fwq->state, IQS_BUSY, IQS_IDLE);
1292 	}
1293 }
1294 
1295 /* Deals with error interrupts */
1296 void
1297 t4_intr_err(void *arg)
1298 {
1299 	struct adapter *sc = arg;
1300 
1301 	t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
1302 	t4_slow_intr_handler(sc);
1303 }
1304 
1305 void
1306 t4_intr_evt(void *arg)
1307 {
1308 	struct sge_iq *iq = arg;
1309 
1310 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1311 		service_iq(iq, 0);
1312 		atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1313 	}
1314 }
1315 
1316 void
1317 t4_intr(void *arg)
1318 {
1319 	struct sge_iq *iq = arg;
1320 
1321 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1322 		service_iq(iq, 0);
1323 		atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1324 	}
1325 }
1326 
1327 void
1328 t4_vi_intr(void *arg)
1329 {
1330 	struct irq *irq = arg;
1331 
1332 #ifdef DEV_NETMAP
1333 	if (atomic_cmpset_int(&irq->nm_state, NM_ON, NM_BUSY)) {
1334 		t4_nm_intr(irq->nm_rxq);
1335 		atomic_cmpset_int(&irq->nm_state, NM_BUSY, NM_ON);
1336 	}
1337 #endif
1338 	if (irq->rxq != NULL)
1339 		t4_intr(irq->rxq);
1340 }
1341 
1342 /*
1343  * Deals with anything and everything on the given ingress queue.
1344  */
1345 static int
1346 service_iq(struct sge_iq *iq, int budget)
1347 {
1348 	struct sge_iq *q;
1349 	struct sge_rxq *rxq = iq_to_rxq(iq);	/* Use iff iq is part of rxq */
1350 	struct sge_fl *fl;			/* Use iff IQ_HAS_FL */
1351 	struct adapter *sc = iq->adapter;
1352 	struct iq_desc *d = &iq->desc[iq->cidx];
1353 	int ndescs = 0, limit;
1354 	int rsp_type, refill;
1355 	uint32_t lq;
1356 	uint16_t fl_hw_cidx;
1357 	struct mbuf *m0;
1358 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1359 #if defined(INET) || defined(INET6)
1360 	const struct timeval lro_timeout = {0, sc->lro_timeout};
1361 #endif
1362 
1363 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1364 
1365 	limit = budget ? budget : iq->qsize / 16;
1366 
1367 	if (iq->flags & IQ_HAS_FL) {
1368 		fl = &rxq->fl;
1369 		fl_hw_cidx = fl->hw_cidx;	/* stable snapshot */
1370 	} else {
1371 		fl = NULL;
1372 		fl_hw_cidx = 0;			/* to silence gcc warning */
1373 	}
1374 
1375 	/*
1376 	 * We always come back and check the descriptor ring for new indirect
1377 	 * interrupts and other responses after running a single handler.
1378 	 */
1379 	for (;;) {
1380 		while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1381 
1382 			rmb();
1383 
1384 			refill = 0;
1385 			m0 = NULL;
1386 			rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1387 			lq = be32toh(d->rsp.pldbuflen_qid);
1388 
1389 			switch (rsp_type) {
1390 			case X_RSPD_TYPE_FLBUF:
1391 
1392 				KASSERT(iq->flags & IQ_HAS_FL,
1393 				    ("%s: data for an iq (%p) with no freelist",
1394 				    __func__, iq));
1395 
1396 				m0 = get_fl_payload(sc, fl, lq);
1397 				if (__predict_false(m0 == NULL))
1398 					goto process_iql;
1399 				refill = IDXDIFF(fl->hw_cidx, fl_hw_cidx, fl->sidx) > 2;
1400 #ifdef T4_PKT_TIMESTAMP
1401 				/*
1402 				 * 60 bit timestamp for the payload is
1403 				 * *(uint64_t *)m0->m_pktdat.  Note that it is
1404 				 * in the leading free-space in the mbuf.  The
1405 				 * kernel can clobber it during a pullup,
1406 				 * m_copymdata, etc.  You need to make sure that
1407 				 * the mbuf reaches you unmolested if you care
1408 				 * about the timestamp.
1409 				 */
1410 				*(uint64_t *)m0->m_pktdat =
1411 				    be64toh(ctrl->u.last_flit) &
1412 				    0xfffffffffffffff;
1413 #endif
1414 
1415 				/* fall through */
1416 
1417 			case X_RSPD_TYPE_CPL:
1418 				KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1419 				    ("%s: bad opcode %02x.", __func__,
1420 				    d->rss.opcode));
1421 				t4_cpl_handler[d->rss.opcode](iq, &d->rss, m0);
1422 				break;
1423 
1424 			case X_RSPD_TYPE_INTR:
1425 
1426 				/*
1427 				 * Interrupts should be forwarded only to queues
1428 				 * that are not forwarding their interrupts.
1429 				 * This means service_iq can recurse but only 1
1430 				 * level deep.
1431 				 */
1432 				KASSERT(budget == 0,
1433 				    ("%s: budget %u, rsp_type %u", __func__,
1434 				    budget, rsp_type));
1435 
1436 				/*
1437 				 * There are 1K interrupt-capable queues (qids 0
1438 				 * through 1023).  A response type indicating a
1439 				 * forwarded interrupt with a qid >= 1K is an
1440 				 * iWARP async notification.
1441 				 */
1442 				if (lq >= 1024) {
1443                                         t4_an_handler(iq, &d->rsp);
1444                                         break;
1445                                 }
1446 
1447 				q = sc->sge.iqmap[lq - sc->sge.iq_start];
1448 				if (atomic_cmpset_int(&q->state, IQS_IDLE,
1449 				    IQS_BUSY)) {
1450 					if (service_iq(q, q->qsize / 16) == 0) {
1451 						atomic_cmpset_int(&q->state,
1452 						    IQS_BUSY, IQS_IDLE);
1453 					} else {
1454 						STAILQ_INSERT_TAIL(&iql, q,
1455 						    link);
1456 					}
1457 				}
1458 				break;
1459 
1460 			default:
1461 				KASSERT(0,
1462 				    ("%s: illegal response type %d on iq %p",
1463 				    __func__, rsp_type, iq));
1464 				log(LOG_ERR,
1465 				    "%s: illegal response type %d on iq %p",
1466 				    device_get_nameunit(sc->dev), rsp_type, iq);
1467 				break;
1468 			}
1469 
1470 			d++;
1471 			if (__predict_false(++iq->cidx == iq->sidx)) {
1472 				iq->cidx = 0;
1473 				iq->gen ^= F_RSPD_GEN;
1474 				d = &iq->desc[0];
1475 			}
1476 			if (__predict_false(++ndescs == limit)) {
1477 				t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
1478 				    V_CIDXINC(ndescs) |
1479 				    V_INGRESSQID(iq->cntxt_id) |
1480 				    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1481 				ndescs = 0;
1482 
1483 #if defined(INET) || defined(INET6)
1484 				if (iq->flags & IQ_LRO_ENABLED &&
1485 				    sc->lro_timeout != 0) {
1486 					tcp_lro_flush_inactive(&rxq->lro,
1487 					    &lro_timeout);
1488 				}
1489 #endif
1490 
1491 				if (budget) {
1492 					if (iq->flags & IQ_HAS_FL) {
1493 						FL_LOCK(fl);
1494 						refill_fl(sc, fl, 32);
1495 						FL_UNLOCK(fl);
1496 					}
1497 					return (EINPROGRESS);
1498 				}
1499 			}
1500 			if (refill) {
1501 				FL_LOCK(fl);
1502 				refill_fl(sc, fl, 32);
1503 				FL_UNLOCK(fl);
1504 				fl_hw_cidx = fl->hw_cidx;
1505 			}
1506 		}
1507 
1508 process_iql:
1509 		if (STAILQ_EMPTY(&iql))
1510 			break;
1511 
1512 		/*
1513 		 * Process the head only, and send it to the back of the list if
1514 		 * it's still not done.
1515 		 */
1516 		q = STAILQ_FIRST(&iql);
1517 		STAILQ_REMOVE_HEAD(&iql, link);
1518 		if (service_iq(q, q->qsize / 8) == 0)
1519 			atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1520 		else
1521 			STAILQ_INSERT_TAIL(&iql, q, link);
1522 	}
1523 
1524 #if defined(INET) || defined(INET6)
1525 	if (iq->flags & IQ_LRO_ENABLED) {
1526 		struct lro_ctrl *lro = &rxq->lro;
1527 
1528 		tcp_lro_flush_all(lro);
1529 	}
1530 #endif
1531 
1532 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_CIDXINC(ndescs) |
1533 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1534 
1535 	if (iq->flags & IQ_HAS_FL) {
1536 		int starved;
1537 
1538 		FL_LOCK(fl);
1539 		starved = refill_fl(sc, fl, 64);
1540 		FL_UNLOCK(fl);
1541 		if (__predict_false(starved != 0))
1542 			add_fl_to_sfl(sc, fl);
1543 	}
1544 
1545 	return (0);
1546 }
1547 
1548 static inline int
1549 cl_has_metadata(struct sge_fl *fl, struct cluster_layout *cll)
1550 {
1551 	int rc = fl->flags & FL_BUF_PACKING || cll->region1 > 0;
1552 
1553 	if (rc)
1554 		MPASS(cll->region3 >= CL_METADATA_SIZE);
1555 
1556 	return (rc);
1557 }
1558 
1559 static inline struct cluster_metadata *
1560 cl_metadata(struct adapter *sc, struct sge_fl *fl, struct cluster_layout *cll,
1561     caddr_t cl)
1562 {
1563 
1564 	if (cl_has_metadata(fl, cll)) {
1565 		struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1566 
1567 		return ((struct cluster_metadata *)(cl + swz->size) - 1);
1568 	}
1569 	return (NULL);
1570 }
1571 
1572 static void
1573 rxb_free(struct mbuf *m, void *arg1, void *arg2)
1574 {
1575 	uma_zone_t zone = arg1;
1576 	caddr_t cl = arg2;
1577 
1578 	uma_zfree(zone, cl);
1579 	counter_u64_add(extfree_rels, 1);
1580 }
1581 
1582 /*
1583  * The mbuf returned by this function could be allocated from zone_mbuf or
1584  * constructed in spare room in the cluster.
1585  *
1586  * The mbuf carries the payload in one of these ways
1587  * a) frame inside the mbuf (mbuf from zone_mbuf)
1588  * b) m_cljset (for clusters without metadata) zone_mbuf
1589  * c) m_extaddref (cluster with metadata) inline mbuf
1590  * d) m_extaddref (cluster with metadata) zone_mbuf
1591  */
1592 static struct mbuf *
1593 get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1594     int remaining)
1595 {
1596 	struct mbuf *m;
1597 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1598 	struct cluster_layout *cll = &sd->cll;
1599 	struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1600 	struct hw_buf_info *hwb = &sc->sge.hw_buf_info[cll->hwidx];
1601 	struct cluster_metadata *clm = cl_metadata(sc, fl, cll, sd->cl);
1602 	int len, blen;
1603 	caddr_t payload;
1604 
1605 	blen = hwb->size - fl->rx_offset;	/* max possible in this buf */
1606 	len = min(remaining, blen);
1607 	payload = sd->cl + cll->region1 + fl->rx_offset;
1608 	if (fl->flags & FL_BUF_PACKING) {
1609 		const u_int l = fr_offset + len;
1610 		const u_int pad = roundup2(l, fl->buf_boundary) - l;
1611 
1612 		if (fl->rx_offset + len + pad < hwb->size)
1613 			blen = len + pad;
1614 		MPASS(fl->rx_offset + blen <= hwb->size);
1615 	} else {
1616 		MPASS(fl->rx_offset == 0);	/* not packing */
1617 	}
1618 
1619 
1620 	if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) {
1621 
1622 		/*
1623 		 * Copy payload into a freshly allocated mbuf.
1624 		 */
1625 
1626 		m = fr_offset == 0 ?
1627 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1628 		if (m == NULL)
1629 			return (NULL);
1630 		fl->mbuf_allocated++;
1631 #ifdef T4_PKT_TIMESTAMP
1632 		/* Leave room for a timestamp */
1633 		m->m_data += 8;
1634 #endif
1635 		/* copy data to mbuf */
1636 		bcopy(payload, mtod(m, caddr_t), len);
1637 
1638 	} else if (sd->nmbuf * MSIZE < cll->region1) {
1639 
1640 		/*
1641 		 * There's spare room in the cluster for an mbuf.  Create one
1642 		 * and associate it with the payload that's in the cluster.
1643 		 */
1644 
1645 		MPASS(clm != NULL);
1646 		m = (struct mbuf *)(sd->cl + sd->nmbuf * MSIZE);
1647 		/* No bzero required */
1648 		if (m_init(m, M_NOWAIT, MT_DATA,
1649 		    fr_offset == 0 ? M_PKTHDR | M_NOFREE : M_NOFREE))
1650 			return (NULL);
1651 		fl->mbuf_inlined++;
1652 		m_extaddref(m, payload, blen, &clm->refcount, rxb_free,
1653 		    swz->zone, sd->cl);
1654 		if (sd->nmbuf++ == 0)
1655 			counter_u64_add(extfree_refs, 1);
1656 
1657 	} else {
1658 
1659 		/*
1660 		 * Grab an mbuf from zone_mbuf and associate it with the
1661 		 * payload in the cluster.
1662 		 */
1663 
1664 		m = fr_offset == 0 ?
1665 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1666 		if (m == NULL)
1667 			return (NULL);
1668 		fl->mbuf_allocated++;
1669 		if (clm != NULL) {
1670 			m_extaddref(m, payload, blen, &clm->refcount,
1671 			    rxb_free, swz->zone, sd->cl);
1672 			if (sd->nmbuf++ == 0)
1673 				counter_u64_add(extfree_refs, 1);
1674 		} else {
1675 			m_cljset(m, sd->cl, swz->type);
1676 			sd->cl = NULL;	/* consumed, not a recycle candidate */
1677 		}
1678 	}
1679 	if (fr_offset == 0)
1680 		m->m_pkthdr.len = remaining;
1681 	m->m_len = len;
1682 
1683 	if (fl->flags & FL_BUF_PACKING) {
1684 		fl->rx_offset += blen;
1685 		MPASS(fl->rx_offset <= hwb->size);
1686 		if (fl->rx_offset < hwb->size)
1687 			return (m);	/* without advancing the cidx */
1688 	}
1689 
1690 	if (__predict_false(++fl->cidx % 8 == 0)) {
1691 		uint16_t cidx = fl->cidx / 8;
1692 
1693 		if (__predict_false(cidx == fl->sidx))
1694 			fl->cidx = cidx = 0;
1695 		fl->hw_cidx = cidx;
1696 	}
1697 	fl->rx_offset = 0;
1698 
1699 	return (m);
1700 }
1701 
1702 static struct mbuf *
1703 get_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf)
1704 {
1705 	struct mbuf *m0, *m, **pnext;
1706 	u_int remaining;
1707 	const u_int total = G_RSPD_LEN(len_newbuf);
1708 
1709 	if (__predict_false(fl->flags & FL_BUF_RESUME)) {
1710 		M_ASSERTPKTHDR(fl->m0);
1711 		MPASS(fl->m0->m_pkthdr.len == total);
1712 		MPASS(fl->remaining < total);
1713 
1714 		m0 = fl->m0;
1715 		pnext = fl->pnext;
1716 		remaining = fl->remaining;
1717 		fl->flags &= ~FL_BUF_RESUME;
1718 		goto get_segment;
1719 	}
1720 
1721 	if (fl->rx_offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
1722 		fl->rx_offset = 0;
1723 		if (__predict_false(++fl->cidx % 8 == 0)) {
1724 			uint16_t cidx = fl->cidx / 8;
1725 
1726 			if (__predict_false(cidx == fl->sidx))
1727 				fl->cidx = cidx = 0;
1728 			fl->hw_cidx = cidx;
1729 		}
1730 	}
1731 
1732 	/*
1733 	 * Payload starts at rx_offset in the current hw buffer.  Its length is
1734 	 * 'len' and it may span multiple hw buffers.
1735 	 */
1736 
1737 	m0 = get_scatter_segment(sc, fl, 0, total);
1738 	if (m0 == NULL)
1739 		return (NULL);
1740 	remaining = total - m0->m_len;
1741 	pnext = &m0->m_next;
1742 	while (remaining > 0) {
1743 get_segment:
1744 		MPASS(fl->rx_offset == 0);
1745 		m = get_scatter_segment(sc, fl, total - remaining, remaining);
1746 		if (__predict_false(m == NULL)) {
1747 			fl->m0 = m0;
1748 			fl->pnext = pnext;
1749 			fl->remaining = remaining;
1750 			fl->flags |= FL_BUF_RESUME;
1751 			return (NULL);
1752 		}
1753 		*pnext = m;
1754 		pnext = &m->m_next;
1755 		remaining -= m->m_len;
1756 	}
1757 	*pnext = NULL;
1758 
1759 	M_ASSERTPKTHDR(m0);
1760 	return (m0);
1761 }
1762 
1763 static int
1764 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
1765 {
1766 	struct sge_rxq *rxq = iq_to_rxq(iq);
1767 	struct ifnet *ifp = rxq->ifp;
1768 	struct adapter *sc = iq->adapter;
1769 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
1770 #if defined(INET) || defined(INET6)
1771 	struct lro_ctrl *lro = &rxq->lro;
1772 #endif
1773 	static const int sw_hashtype[4][2] = {
1774 		{M_HASHTYPE_NONE, M_HASHTYPE_NONE},
1775 		{M_HASHTYPE_RSS_IPV4, M_HASHTYPE_RSS_IPV6},
1776 		{M_HASHTYPE_RSS_TCP_IPV4, M_HASHTYPE_RSS_TCP_IPV6},
1777 		{M_HASHTYPE_RSS_UDP_IPV4, M_HASHTYPE_RSS_UDP_IPV6},
1778 	};
1779 
1780 	KASSERT(m0 != NULL, ("%s: no payload with opcode %02x", __func__,
1781 	    rss->opcode));
1782 
1783 	m0->m_pkthdr.len -= sc->params.sge.fl_pktshift;
1784 	m0->m_len -= sc->params.sge.fl_pktshift;
1785 	m0->m_data += sc->params.sge.fl_pktshift;
1786 
1787 	m0->m_pkthdr.rcvif = ifp;
1788 	M_HASHTYPE_SET(m0, sw_hashtype[rss->hash_type][rss->ipv6]);
1789 	m0->m_pkthdr.flowid = be32toh(rss->hash_val);
1790 
1791 	if (cpl->csum_calc && !cpl->err_vec) {
1792 		if (ifp->if_capenable & IFCAP_RXCSUM &&
1793 		    cpl->l2info & htobe32(F_RXF_IP)) {
1794 			m0->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
1795 			    CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
1796 			rxq->rxcsum++;
1797 		} else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
1798 		    cpl->l2info & htobe32(F_RXF_IP6)) {
1799 			m0->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
1800 			    CSUM_PSEUDO_HDR);
1801 			rxq->rxcsum++;
1802 		}
1803 
1804 		if (__predict_false(cpl->ip_frag))
1805 			m0->m_pkthdr.csum_data = be16toh(cpl->csum);
1806 		else
1807 			m0->m_pkthdr.csum_data = 0xffff;
1808 	}
1809 
1810 	if (cpl->vlan_ex) {
1811 		m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
1812 		m0->m_flags |= M_VLANTAG;
1813 		rxq->vlan_extraction++;
1814 	}
1815 
1816 #if defined(INET) || defined(INET6)
1817 	if (cpl->l2info & htobe32(F_RXF_LRO) &&
1818 	    iq->flags & IQ_LRO_ENABLED &&
1819 	    tcp_lro_rx(lro, m0, 0) == 0) {
1820 		/* queued for LRO */
1821 	} else
1822 #endif
1823 	ifp->if_input(ifp, m0);
1824 
1825 	return (0);
1826 }
1827 
1828 /*
1829  * Must drain the wrq or make sure that someone else will.
1830  */
1831 static void
1832 wrq_tx_drain(void *arg, int n)
1833 {
1834 	struct sge_wrq *wrq = arg;
1835 	struct sge_eq *eq = &wrq->eq;
1836 
1837 	EQ_LOCK(eq);
1838 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
1839 		drain_wrq_wr_list(wrq->adapter, wrq);
1840 	EQ_UNLOCK(eq);
1841 }
1842 
1843 static void
1844 drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq)
1845 {
1846 	struct sge_eq *eq = &wrq->eq;
1847 	u_int available, dbdiff;	/* # of hardware descriptors */
1848 	u_int n;
1849 	struct wrqe *wr;
1850 	struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
1851 
1852 	EQ_LOCK_ASSERT_OWNED(eq);
1853 	MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
1854 	wr = STAILQ_FIRST(&wrq->wr_list);
1855 	MPASS(wr != NULL);	/* Must be called with something useful to do */
1856 	MPASS(eq->pidx == eq->dbidx);
1857 	dbdiff = 0;
1858 
1859 	do {
1860 		eq->cidx = read_hw_cidx(eq);
1861 		if (eq->pidx == eq->cidx)
1862 			available = eq->sidx - 1;
1863 		else
1864 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
1865 
1866 		MPASS(wr->wrq == wrq);
1867 		n = howmany(wr->wr_len, EQ_ESIZE);
1868 		if (available < n)
1869 			break;
1870 
1871 		dst = (void *)&eq->desc[eq->pidx];
1872 		if (__predict_true(eq->sidx - eq->pidx > n)) {
1873 			/* Won't wrap, won't end exactly at the status page. */
1874 			bcopy(&wr->wr[0], dst, wr->wr_len);
1875 			eq->pidx += n;
1876 		} else {
1877 			int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE;
1878 
1879 			bcopy(&wr->wr[0], dst, first_portion);
1880 			if (wr->wr_len > first_portion) {
1881 				bcopy(&wr->wr[first_portion], &eq->desc[0],
1882 				    wr->wr_len - first_portion);
1883 			}
1884 			eq->pidx = n - (eq->sidx - eq->pidx);
1885 		}
1886 
1887 		if (available < eq->sidx / 4 &&
1888 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
1889 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
1890 			    F_FW_WR_EQUEQ);
1891 			eq->equeqidx = eq->pidx;
1892 		} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
1893 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
1894 			eq->equeqidx = eq->pidx;
1895 		}
1896 
1897 		dbdiff += n;
1898 		if (dbdiff >= 16) {
1899 			ring_eq_db(sc, eq, dbdiff);
1900 			dbdiff = 0;
1901 		}
1902 
1903 		STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
1904 		free_wrqe(wr);
1905 		MPASS(wrq->nwr_pending > 0);
1906 		wrq->nwr_pending--;
1907 		MPASS(wrq->ndesc_needed >= n);
1908 		wrq->ndesc_needed -= n;
1909 	} while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL);
1910 
1911 	if (dbdiff)
1912 		ring_eq_db(sc, eq, dbdiff);
1913 }
1914 
1915 /*
1916  * Doesn't fail.  Holds on to work requests it can't send right away.
1917  */
1918 void
1919 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
1920 {
1921 #ifdef INVARIANTS
1922 	struct sge_eq *eq = &wrq->eq;
1923 #endif
1924 
1925 	EQ_LOCK_ASSERT_OWNED(eq);
1926 	MPASS(wr != NULL);
1927 	MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN);
1928 	MPASS((wr->wr_len & 0x7) == 0);
1929 
1930 	STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
1931 	wrq->nwr_pending++;
1932 	wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE);
1933 
1934 	if (!TAILQ_EMPTY(&wrq->incomplete_wrs))
1935 		return;	/* commit_wrq_wr will drain wr_list as well. */
1936 
1937 	drain_wrq_wr_list(sc, wrq);
1938 
1939 	/* Doorbell must have caught up to the pidx. */
1940 	MPASS(eq->pidx == eq->dbidx);
1941 }
1942 
1943 void
1944 t4_update_fl_bufsize(struct ifnet *ifp)
1945 {
1946 	struct vi_info *vi = ifp->if_softc;
1947 	struct adapter *sc = vi->pi->adapter;
1948 	struct sge_rxq *rxq;
1949 #ifdef TCP_OFFLOAD
1950 	struct sge_ofld_rxq *ofld_rxq;
1951 #endif
1952 	struct sge_fl *fl;
1953 	int i, maxp, mtu = ifp->if_mtu;
1954 
1955 	maxp = mtu_to_max_payload(sc, mtu, 0);
1956 	for_each_rxq(vi, i, rxq) {
1957 		fl = &rxq->fl;
1958 
1959 		FL_LOCK(fl);
1960 		find_best_refill_source(sc, fl, maxp);
1961 		FL_UNLOCK(fl);
1962 	}
1963 #ifdef TCP_OFFLOAD
1964 	maxp = mtu_to_max_payload(sc, mtu, 1);
1965 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1966 		fl = &ofld_rxq->fl;
1967 
1968 		FL_LOCK(fl);
1969 		find_best_refill_source(sc, fl, maxp);
1970 		FL_UNLOCK(fl);
1971 	}
1972 #endif
1973 }
1974 
1975 static inline int
1976 mbuf_nsegs(struct mbuf *m)
1977 {
1978 
1979 	M_ASSERTPKTHDR(m);
1980 	KASSERT(m->m_pkthdr.l5hlen > 0,
1981 	    ("%s: mbuf %p missing information on # of segments.", __func__, m));
1982 
1983 	return (m->m_pkthdr.l5hlen);
1984 }
1985 
1986 static inline void
1987 set_mbuf_nsegs(struct mbuf *m, uint8_t nsegs)
1988 {
1989 
1990 	M_ASSERTPKTHDR(m);
1991 	m->m_pkthdr.l5hlen = nsegs;
1992 }
1993 
1994 static inline int
1995 mbuf_len16(struct mbuf *m)
1996 {
1997 	int n;
1998 
1999 	M_ASSERTPKTHDR(m);
2000 	n = m->m_pkthdr.PH_loc.eight[0];
2001 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2002 
2003 	return (n);
2004 }
2005 
2006 static inline void
2007 set_mbuf_len16(struct mbuf *m, uint8_t len16)
2008 {
2009 
2010 	M_ASSERTPKTHDR(m);
2011 	m->m_pkthdr.PH_loc.eight[0] = len16;
2012 }
2013 
2014 static inline int
2015 needs_tso(struct mbuf *m)
2016 {
2017 
2018 	M_ASSERTPKTHDR(m);
2019 
2020 	if (m->m_pkthdr.csum_flags & CSUM_TSO) {
2021 		KASSERT(m->m_pkthdr.tso_segsz > 0,
2022 		    ("%s: TSO requested in mbuf %p but MSS not provided",
2023 		    __func__, m));
2024 		return (1);
2025 	}
2026 
2027 	return (0);
2028 }
2029 
2030 static inline int
2031 needs_l3_csum(struct mbuf *m)
2032 {
2033 
2034 	M_ASSERTPKTHDR(m);
2035 
2036 	if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO))
2037 		return (1);
2038 	return (0);
2039 }
2040 
2041 static inline int
2042 needs_l4_csum(struct mbuf *m)
2043 {
2044 
2045 	M_ASSERTPKTHDR(m);
2046 
2047 	if (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
2048 	    CSUM_TCP_IPV6 | CSUM_TSO))
2049 		return (1);
2050 	return (0);
2051 }
2052 
2053 static inline int
2054 needs_vlan_insertion(struct mbuf *m)
2055 {
2056 
2057 	M_ASSERTPKTHDR(m);
2058 
2059 	if (m->m_flags & M_VLANTAG) {
2060 		KASSERT(m->m_pkthdr.ether_vtag != 0,
2061 		    ("%s: HWVLAN requested in mbuf %p but tag not provided",
2062 		    __func__, m));
2063 		return (1);
2064 	}
2065 	return (0);
2066 }
2067 
2068 static void *
2069 m_advance(struct mbuf **pm, int *poffset, int len)
2070 {
2071 	struct mbuf *m = *pm;
2072 	int offset = *poffset;
2073 	uintptr_t p = 0;
2074 
2075 	MPASS(len > 0);
2076 
2077 	while (len) {
2078 		if (offset + len < m->m_len) {
2079 			offset += len;
2080 			p = mtod(m, uintptr_t) + offset;
2081 			break;
2082 		}
2083 		len -= m->m_len - offset;
2084 		m = m->m_next;
2085 		offset = 0;
2086 		MPASS(m != NULL);
2087 	}
2088 	*poffset = offset;
2089 	*pm = m;
2090 	return ((void *)p);
2091 }
2092 
2093 static inline int
2094 same_paddr(char *a, char *b)
2095 {
2096 
2097 	if (a == b)
2098 		return (1);
2099 	else if (a != NULL && b != NULL) {
2100 		vm_offset_t x = (vm_offset_t)a;
2101 		vm_offset_t y = (vm_offset_t)b;
2102 
2103 		if ((x & PAGE_MASK) == (y & PAGE_MASK) &&
2104 		    pmap_kextract(x) == pmap_kextract(y))
2105 			return (1);
2106 	}
2107 
2108 	return (0);
2109 }
2110 
2111 /*
2112  * Can deal with empty mbufs in the chain that have m_len = 0, but the chain
2113  * must have at least one mbuf that's not empty.
2114  */
2115 static inline int
2116 count_mbuf_nsegs(struct mbuf *m)
2117 {
2118 	char *prev_end, *start;
2119 	int len, nsegs;
2120 
2121 	MPASS(m != NULL);
2122 
2123 	nsegs = 0;
2124 	prev_end = NULL;
2125 	for (; m; m = m->m_next) {
2126 
2127 		len = m->m_len;
2128 		if (__predict_false(len == 0))
2129 			continue;
2130 		start = mtod(m, char *);
2131 
2132 		nsegs += sglist_count(start, len);
2133 		if (same_paddr(prev_end, start))
2134 			nsegs--;
2135 		prev_end = start + len;
2136 	}
2137 
2138 	MPASS(nsegs > 0);
2139 	return (nsegs);
2140 }
2141 
2142 /*
2143  * Analyze the mbuf to determine its tx needs.  The mbuf passed in may change:
2144  * a) caller can assume it's been freed if this function returns with an error.
2145  * b) it may get defragged up if the gather list is too long for the hardware.
2146  */
2147 int
2148 parse_pkt(struct mbuf **mp)
2149 {
2150 	struct mbuf *m0 = *mp, *m;
2151 	int rc, nsegs, defragged = 0, offset;
2152 	struct ether_header *eh;
2153 	void *l3hdr;
2154 #if defined(INET) || defined(INET6)
2155 	struct tcphdr *tcp;
2156 #endif
2157 	uint16_t eh_type;
2158 
2159 	M_ASSERTPKTHDR(m0);
2160 	if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) {
2161 		rc = EINVAL;
2162 fail:
2163 		m_freem(m0);
2164 		*mp = NULL;
2165 		return (rc);
2166 	}
2167 restart:
2168 	/*
2169 	 * First count the number of gather list segments in the payload.
2170 	 * Defrag the mbuf if nsegs exceeds the hardware limit.
2171 	 */
2172 	M_ASSERTPKTHDR(m0);
2173 	MPASS(m0->m_pkthdr.len > 0);
2174 	nsegs = count_mbuf_nsegs(m0);
2175 	if (nsegs > (needs_tso(m0) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS)) {
2176 		if (defragged++ > 0 || (m = m_defrag(m0, M_NOWAIT)) == NULL) {
2177 			rc = EFBIG;
2178 			goto fail;
2179 		}
2180 		*mp = m0 = m;	/* update caller's copy after defrag */
2181 		goto restart;
2182 	}
2183 
2184 	if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN)) {
2185 		m0 = m_pullup(m0, m0->m_pkthdr.len);
2186 		if (m0 == NULL) {
2187 			/* Should have left well enough alone. */
2188 			rc = EFBIG;
2189 			goto fail;
2190 		}
2191 		*mp = m0;	/* update caller's copy after pullup */
2192 		goto restart;
2193 	}
2194 	set_mbuf_nsegs(m0, nsegs);
2195 	set_mbuf_len16(m0, txpkt_len16(nsegs, needs_tso(m0)));
2196 
2197 	if (!needs_tso(m0))
2198 		return (0);
2199 
2200 	m = m0;
2201 	eh = mtod(m, struct ether_header *);
2202 	eh_type = ntohs(eh->ether_type);
2203 	if (eh_type == ETHERTYPE_VLAN) {
2204 		struct ether_vlan_header *evh = (void *)eh;
2205 
2206 		eh_type = ntohs(evh->evl_proto);
2207 		m0->m_pkthdr.l2hlen = sizeof(*evh);
2208 	} else
2209 		m0->m_pkthdr.l2hlen = sizeof(*eh);
2210 
2211 	offset = 0;
2212 	l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2213 
2214 	switch (eh_type) {
2215 #ifdef INET6
2216 	case ETHERTYPE_IPV6:
2217 	{
2218 		struct ip6_hdr *ip6 = l3hdr;
2219 
2220 		MPASS(ip6->ip6_nxt == IPPROTO_TCP);
2221 
2222 		m0->m_pkthdr.l3hlen = sizeof(*ip6);
2223 		break;
2224 	}
2225 #endif
2226 #ifdef INET
2227 	case ETHERTYPE_IP:
2228 	{
2229 		struct ip *ip = l3hdr;
2230 
2231 		m0->m_pkthdr.l3hlen = ip->ip_hl * 4;
2232 		break;
2233 	}
2234 #endif
2235 	default:
2236 		panic("%s: ethertype 0x%04x unknown.  if_cxgbe must be compiled"
2237 		    " with the same INET/INET6 options as the kernel.",
2238 		    __func__, eh_type);
2239 	}
2240 
2241 #if defined(INET) || defined(INET6)
2242 	tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen);
2243 	m0->m_pkthdr.l4hlen = tcp->th_off * 4;
2244 #endif
2245 	MPASS(m0 == *mp);
2246 	return (0);
2247 }
2248 
2249 void *
2250 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie)
2251 {
2252 	struct sge_eq *eq = &wrq->eq;
2253 	struct adapter *sc = wrq->adapter;
2254 	int ndesc, available;
2255 	struct wrqe *wr;
2256 	void *w;
2257 
2258 	MPASS(len16 > 0);
2259 	ndesc = howmany(len16, EQ_ESIZE / 16);
2260 	MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC);
2261 
2262 	EQ_LOCK(eq);
2263 
2264 	if (!STAILQ_EMPTY(&wrq->wr_list))
2265 		drain_wrq_wr_list(sc, wrq);
2266 
2267 	if (!STAILQ_EMPTY(&wrq->wr_list)) {
2268 slowpath:
2269 		EQ_UNLOCK(eq);
2270 		wr = alloc_wrqe(len16 * 16, wrq);
2271 		if (__predict_false(wr == NULL))
2272 			return (NULL);
2273 		cookie->pidx = -1;
2274 		cookie->ndesc = ndesc;
2275 		return (&wr->wr);
2276 	}
2277 
2278 	eq->cidx = read_hw_cidx(eq);
2279 	if (eq->pidx == eq->cidx)
2280 		available = eq->sidx - 1;
2281 	else
2282 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2283 	if (available < ndesc)
2284 		goto slowpath;
2285 
2286 	cookie->pidx = eq->pidx;
2287 	cookie->ndesc = ndesc;
2288 	TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link);
2289 
2290 	w = &eq->desc[eq->pidx];
2291 	IDXINCR(eq->pidx, ndesc, eq->sidx);
2292 	if (__predict_false(eq->pidx < ndesc - 1)) {
2293 		w = &wrq->ss[0];
2294 		wrq->ss_pidx = cookie->pidx;
2295 		wrq->ss_len = len16 * 16;
2296 	}
2297 
2298 	EQ_UNLOCK(eq);
2299 
2300 	return (w);
2301 }
2302 
2303 void
2304 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie)
2305 {
2306 	struct sge_eq *eq = &wrq->eq;
2307 	struct adapter *sc = wrq->adapter;
2308 	int ndesc, pidx;
2309 	struct wrq_cookie *prev, *next;
2310 
2311 	if (cookie->pidx == -1) {
2312 		struct wrqe *wr = __containerof(w, struct wrqe, wr);
2313 
2314 		t4_wrq_tx(sc, wr);
2315 		return;
2316 	}
2317 
2318 	ndesc = cookie->ndesc;	/* Can be more than SGE_MAX_WR_NDESC here. */
2319 	pidx = cookie->pidx;
2320 	MPASS(pidx >= 0 && pidx < eq->sidx);
2321 	if (__predict_false(w == &wrq->ss[0])) {
2322 		int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE;
2323 
2324 		MPASS(wrq->ss_len > n);	/* WR had better wrap around. */
2325 		bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n);
2326 		bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n);
2327 		wrq->tx_wrs_ss++;
2328 	} else
2329 		wrq->tx_wrs_direct++;
2330 
2331 	EQ_LOCK(eq);
2332 	prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link);
2333 	next = TAILQ_NEXT(cookie, link);
2334 	if (prev == NULL) {
2335 		MPASS(pidx == eq->dbidx);
2336 		if (next == NULL || ndesc >= 16)
2337 			ring_eq_db(wrq->adapter, eq, ndesc);
2338 		else {
2339 			MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc);
2340 			next->pidx = pidx;
2341 			next->ndesc += ndesc;
2342 		}
2343 	} else {
2344 		MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc);
2345 		prev->ndesc += ndesc;
2346 	}
2347 	TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link);
2348 
2349 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2350 		drain_wrq_wr_list(sc, wrq);
2351 
2352 #ifdef INVARIANTS
2353 	if (TAILQ_EMPTY(&wrq->incomplete_wrs)) {
2354 		/* Doorbell must have caught up to the pidx. */
2355 		MPASS(wrq->eq.pidx == wrq->eq.dbidx);
2356 	}
2357 #endif
2358 	EQ_UNLOCK(eq);
2359 }
2360 
2361 static u_int
2362 can_resume_eth_tx(struct mp_ring *r)
2363 {
2364 	struct sge_eq *eq = r->cookie;
2365 
2366 	return (total_available_tx_desc(eq) > eq->sidx / 8);
2367 }
2368 
2369 static inline int
2370 cannot_use_txpkts(struct mbuf *m)
2371 {
2372 	/* maybe put a GL limit too, to avoid silliness? */
2373 
2374 	return (needs_tso(m));
2375 }
2376 
2377 /*
2378  * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to
2379  * be consumed.  Return the actual number consumed.  0 indicates a stall.
2380  */
2381 static u_int
2382 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx)
2383 {
2384 	struct sge_txq *txq = r->cookie;
2385 	struct sge_eq *eq = &txq->eq;
2386 	struct ifnet *ifp = txq->ifp;
2387 	struct vi_info *vi = ifp->if_softc;
2388 	struct port_info *pi = vi->pi;
2389 	struct adapter *sc = pi->adapter;
2390 	u_int total, remaining;		/* # of packets */
2391 	u_int available, dbdiff;	/* # of hardware descriptors */
2392 	u_int n, next_cidx;
2393 	struct mbuf *m0, *tail;
2394 	struct txpkts txp;
2395 	struct fw_eth_tx_pkts_wr *wr;	/* any fw WR struct will do */
2396 
2397 	remaining = IDXDIFF(pidx, cidx, r->size);
2398 	MPASS(remaining > 0);	/* Must not be called without work to do. */
2399 	total = 0;
2400 
2401 	TXQ_LOCK(txq);
2402 	if (__predict_false((eq->flags & EQ_ENABLED) == 0)) {
2403 		while (cidx != pidx) {
2404 			m0 = r->items[cidx];
2405 			m_freem(m0);
2406 			if (++cidx == r->size)
2407 				cidx = 0;
2408 		}
2409 		reclaim_tx_descs(txq, 2048);
2410 		total = remaining;
2411 		goto done;
2412 	}
2413 
2414 	/* How many hardware descriptors do we have readily available. */
2415 	if (eq->pidx == eq->cidx)
2416 		available = eq->sidx - 1;
2417 	else
2418 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2419 	dbdiff = IDXDIFF(eq->pidx, eq->dbidx, eq->sidx);
2420 
2421 	while (remaining > 0) {
2422 
2423 		m0 = r->items[cidx];
2424 		M_ASSERTPKTHDR(m0);
2425 		MPASS(m0->m_nextpkt == NULL);
2426 
2427 		if (available < SGE_MAX_WR_NDESC) {
2428 			available += reclaim_tx_descs(txq, 64);
2429 			if (available < howmany(mbuf_len16(m0), EQ_ESIZE / 16))
2430 				break;	/* out of descriptors */
2431 		}
2432 
2433 		next_cidx = cidx + 1;
2434 		if (__predict_false(next_cidx == r->size))
2435 			next_cidx = 0;
2436 
2437 		wr = (void *)&eq->desc[eq->pidx];
2438 		if (remaining > 1 &&
2439 		    try_txpkts(m0, r->items[next_cidx], &txp, available) == 0) {
2440 
2441 			/* pkts at cidx, next_cidx should both be in txp. */
2442 			MPASS(txp.npkt == 2);
2443 			tail = r->items[next_cidx];
2444 			MPASS(tail->m_nextpkt == NULL);
2445 			ETHER_BPF_MTAP(ifp, m0);
2446 			ETHER_BPF_MTAP(ifp, tail);
2447 			m0->m_nextpkt = tail;
2448 
2449 			if (__predict_false(++next_cidx == r->size))
2450 				next_cidx = 0;
2451 
2452 			while (next_cidx != pidx) {
2453 				if (add_to_txpkts(r->items[next_cidx], &txp,
2454 				    available) != 0)
2455 					break;
2456 				tail->m_nextpkt = r->items[next_cidx];
2457 				tail = tail->m_nextpkt;
2458 				ETHER_BPF_MTAP(ifp, tail);
2459 				if (__predict_false(++next_cidx == r->size))
2460 					next_cidx = 0;
2461 			}
2462 
2463 			n = write_txpkts_wr(txq, wr, m0, &txp, available);
2464 			total += txp.npkt;
2465 			remaining -= txp.npkt;
2466 		} else {
2467 			total++;
2468 			remaining--;
2469 			ETHER_BPF_MTAP(ifp, m0);
2470 			n = write_txpkt_wr(txq, (void *)wr, m0, available);
2471 		}
2472 		MPASS(n >= 1 && n <= available && n <= SGE_MAX_WR_NDESC);
2473 
2474 		available -= n;
2475 		dbdiff += n;
2476 		IDXINCR(eq->pidx, n, eq->sidx);
2477 
2478 		if (total_available_tx_desc(eq) < eq->sidx / 4 &&
2479 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2480 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2481 			    F_FW_WR_EQUEQ);
2482 			eq->equeqidx = eq->pidx;
2483 		} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
2484 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
2485 			eq->equeqidx = eq->pidx;
2486 		}
2487 
2488 		if (dbdiff >= 16 && remaining >= 4) {
2489 			ring_eq_db(sc, eq, dbdiff);
2490 			available += reclaim_tx_descs(txq, 4 * dbdiff);
2491 			dbdiff = 0;
2492 		}
2493 
2494 		cidx = next_cidx;
2495 	}
2496 	if (dbdiff != 0) {
2497 		ring_eq_db(sc, eq, dbdiff);
2498 		reclaim_tx_descs(txq, 32);
2499 	}
2500 done:
2501 	TXQ_UNLOCK(txq);
2502 
2503 	return (total);
2504 }
2505 
2506 static inline void
2507 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
2508     int qsize)
2509 {
2510 
2511 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
2512 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
2513 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
2514 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
2515 
2516 	iq->flags = 0;
2517 	iq->adapter = sc;
2518 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
2519 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
2520 	if (pktc_idx >= 0) {
2521 		iq->intr_params |= F_QINTR_CNT_EN;
2522 		iq->intr_pktc_idx = pktc_idx;
2523 	}
2524 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
2525 	iq->sidx = iq->qsize - sc->params.sge.spg_len / IQ_ESIZE;
2526 }
2527 
2528 static inline void
2529 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name)
2530 {
2531 
2532 	fl->qsize = qsize;
2533 	fl->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2534 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
2535 	if (sc->flags & BUF_PACKING_OK &&
2536 	    ((!is_t4(sc) && buffer_packing) ||	/* T5+: enabled unless 0 */
2537 	    (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */
2538 		fl->flags |= FL_BUF_PACKING;
2539 	find_best_refill_source(sc, fl, maxp);
2540 	find_safe_refill_source(sc, fl);
2541 }
2542 
2543 static inline void
2544 init_eq(struct adapter *sc, struct sge_eq *eq, int eqtype, int qsize,
2545     uint8_t tx_chan, uint16_t iqid, char *name)
2546 {
2547 	KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
2548 
2549 	eq->flags = eqtype & EQ_TYPEMASK;
2550 	eq->tx_chan = tx_chan;
2551 	eq->iqid = iqid;
2552 	eq->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2553 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
2554 }
2555 
2556 static int
2557 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
2558     bus_dmamap_t *map, bus_addr_t *pa, void **va)
2559 {
2560 	int rc;
2561 
2562 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
2563 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
2564 	if (rc != 0) {
2565 		device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
2566 		goto done;
2567 	}
2568 
2569 	rc = bus_dmamem_alloc(*tag, va,
2570 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
2571 	if (rc != 0) {
2572 		device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
2573 		goto done;
2574 	}
2575 
2576 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
2577 	if (rc != 0) {
2578 		device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
2579 		goto done;
2580 	}
2581 done:
2582 	if (rc)
2583 		free_ring(sc, *tag, *map, *pa, *va);
2584 
2585 	return (rc);
2586 }
2587 
2588 static int
2589 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
2590     bus_addr_t pa, void *va)
2591 {
2592 	if (pa)
2593 		bus_dmamap_unload(tag, map);
2594 	if (va)
2595 		bus_dmamem_free(tag, va, map);
2596 	if (tag)
2597 		bus_dma_tag_destroy(tag);
2598 
2599 	return (0);
2600 }
2601 
2602 /*
2603  * Allocates the ring for an ingress queue and an optional freelist.  If the
2604  * freelist is specified it will be allocated and then associated with the
2605  * ingress queue.
2606  *
2607  * Returns errno on failure.  Resources allocated up to that point may still be
2608  * allocated.  Caller is responsible for cleanup in case this function fails.
2609  *
2610  * If the ingress queue will take interrupts directly (iq->flags & IQ_INTR) then
2611  * the intr_idx specifies the vector, starting from 0.  Otherwise it specifies
2612  * the abs_id of the ingress queue to which its interrupts should be forwarded.
2613  */
2614 static int
2615 alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl,
2616     int intr_idx, int cong)
2617 {
2618 	int rc, i, cntxt_id;
2619 	size_t len;
2620 	struct fw_iq_cmd c;
2621 	struct port_info *pi = vi->pi;
2622 	struct adapter *sc = iq->adapter;
2623 	struct sge_params *sp = &sc->params.sge;
2624 	__be32 v = 0;
2625 
2626 	len = iq->qsize * IQ_ESIZE;
2627 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
2628 	    (void **)&iq->desc);
2629 	if (rc != 0)
2630 		return (rc);
2631 
2632 	bzero(&c, sizeof(c));
2633 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
2634 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
2635 	    V_FW_IQ_CMD_VFN(0));
2636 
2637 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
2638 	    FW_LEN16(c));
2639 
2640 	/* Special handling for firmware event queue */
2641 	if (iq == &sc->sge.fwq)
2642 		v |= F_FW_IQ_CMD_IQASYNCH;
2643 
2644 	if (iq->flags & IQ_INTR) {
2645 		KASSERT(intr_idx < sc->intr_count,
2646 		    ("%s: invalid direct intr_idx %d", __func__, intr_idx));
2647 	} else
2648 		v |= F_FW_IQ_CMD_IQANDST;
2649 	v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
2650 
2651 	c.type_to_iqandstindex = htobe32(v |
2652 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
2653 	    V_FW_IQ_CMD_VIID(vi->viid) |
2654 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
2655 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
2656 	    F_FW_IQ_CMD_IQGTSMODE |
2657 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
2658 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
2659 	c.iqsize = htobe16(iq->qsize);
2660 	c.iqaddr = htobe64(iq->ba);
2661 	if (cong >= 0)
2662 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
2663 
2664 	if (fl) {
2665 		mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
2666 
2667 		len = fl->qsize * EQ_ESIZE;
2668 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
2669 		    &fl->ba, (void **)&fl->desc);
2670 		if (rc)
2671 			return (rc);
2672 
2673 		/* Allocate space for one software descriptor per buffer. */
2674 		rc = alloc_fl_sdesc(fl);
2675 		if (rc != 0) {
2676 			device_printf(sc->dev,
2677 			    "failed to setup fl software descriptors: %d\n",
2678 			    rc);
2679 			return (rc);
2680 		}
2681 
2682 		if (fl->flags & FL_BUF_PACKING) {
2683 			fl->lowat = roundup2(sp->fl_starve_threshold2, 8);
2684 			fl->buf_boundary = sp->pack_boundary;
2685 		} else {
2686 			fl->lowat = roundup2(sp->fl_starve_threshold, 8);
2687 			fl->buf_boundary = 16;
2688 		}
2689 		if (fl_pad && fl->buf_boundary < sp->pad_boundary)
2690 			fl->buf_boundary = sp->pad_boundary;
2691 
2692 		c.iqns_to_fl0congen |=
2693 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
2694 			F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
2695 			(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
2696 			(fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
2697 			    0));
2698 		if (cong >= 0) {
2699 			c.iqns_to_fl0congen |=
2700 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
2701 				    F_FW_IQ_CMD_FL0CONGCIF |
2702 				    F_FW_IQ_CMD_FL0CONGEN);
2703 		}
2704 		c.fl0dcaen_to_fl0cidxfthresh =
2705 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(X_FETCHBURSTMIN_128B) |
2706 			V_FW_IQ_CMD_FL0FBMAX(X_FETCHBURSTMAX_512B));
2707 		c.fl0size = htobe16(fl->qsize);
2708 		c.fl0addr = htobe64(fl->ba);
2709 	}
2710 
2711 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
2712 	if (rc != 0) {
2713 		device_printf(sc->dev,
2714 		    "failed to create ingress queue: %d\n", rc);
2715 		return (rc);
2716 	}
2717 
2718 	iq->cidx = 0;
2719 	iq->gen = F_RSPD_GEN;
2720 	iq->intr_next = iq->intr_params;
2721 	iq->cntxt_id = be16toh(c.iqid);
2722 	iq->abs_id = be16toh(c.physiqid);
2723 	iq->flags |= IQ_ALLOCATED;
2724 
2725 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
2726 	if (cntxt_id >= sc->sge.niq) {
2727 		panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
2728 		    cntxt_id, sc->sge.niq - 1);
2729 	}
2730 	sc->sge.iqmap[cntxt_id] = iq;
2731 
2732 	if (fl) {
2733 		u_int qid;
2734 
2735 		iq->flags |= IQ_HAS_FL;
2736 		fl->cntxt_id = be16toh(c.fl0id);
2737 		fl->pidx = fl->cidx = 0;
2738 
2739 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
2740 		if (cntxt_id >= sc->sge.neq) {
2741 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
2742 			    __func__, cntxt_id, sc->sge.neq - 1);
2743 		}
2744 		sc->sge.eqmap[cntxt_id] = (void *)fl;
2745 
2746 		qid = fl->cntxt_id;
2747 		if (isset(&sc->doorbells, DOORBELL_UDB)) {
2748 			uint32_t s_qpp = sc->params.sge.eq_s_qpp;
2749 			uint32_t mask = (1 << s_qpp) - 1;
2750 			volatile uint8_t *udb;
2751 
2752 			udb = sc->udbs_base + UDBS_DB_OFFSET;
2753 			udb += (qid >> s_qpp) << PAGE_SHIFT;
2754 			qid &= mask;
2755 			if (qid < PAGE_SIZE / UDBS_SEG_SIZE) {
2756 				udb += qid << UDBS_SEG_SHIFT;
2757 				qid = 0;
2758 			}
2759 			fl->udb = (volatile void *)udb;
2760 		}
2761 		fl->dbval = V_QID(qid) | sc->chip_params->sge_fl_db;
2762 
2763 		FL_LOCK(fl);
2764 		/* Enough to make sure the SGE doesn't think it's starved */
2765 		refill_fl(sc, fl, fl->lowat);
2766 		FL_UNLOCK(fl);
2767 	}
2768 
2769 	if (is_t5(sc) && cong >= 0) {
2770 		uint32_t param, val;
2771 
2772 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
2773 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
2774 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
2775 		if (cong == 0)
2776 			val = 1 << 19;
2777 		else {
2778 			val = 2 << 19;
2779 			for (i = 0; i < 4; i++) {
2780 				if (cong & (1 << i))
2781 					val |= 1 << (i << 2);
2782 			}
2783 		}
2784 
2785 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
2786 		if (rc != 0) {
2787 			/* report error but carry on */
2788 			device_printf(sc->dev,
2789 			    "failed to set congestion manager context for "
2790 			    "ingress queue %d: %d\n", iq->cntxt_id, rc);
2791 		}
2792 	}
2793 
2794 	/* Enable IQ interrupts */
2795 	atomic_store_rel_int(&iq->state, IQS_IDLE);
2796 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_SEINTARM(iq->intr_params) |
2797 	    V_INGRESSQID(iq->cntxt_id));
2798 
2799 	return (0);
2800 }
2801 
2802 static int
2803 free_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl)
2804 {
2805 	int rc;
2806 	struct adapter *sc = iq->adapter;
2807 	device_t dev;
2808 
2809 	if (sc == NULL)
2810 		return (0);	/* nothing to do */
2811 
2812 	dev = vi ? vi->dev : sc->dev;
2813 
2814 	if (iq->flags & IQ_ALLOCATED) {
2815 		rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
2816 		    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
2817 		    fl ? fl->cntxt_id : 0xffff, 0xffff);
2818 		if (rc != 0) {
2819 			device_printf(dev,
2820 			    "failed to free queue %p: %d\n", iq, rc);
2821 			return (rc);
2822 		}
2823 		iq->flags &= ~IQ_ALLOCATED;
2824 	}
2825 
2826 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
2827 
2828 	bzero(iq, sizeof(*iq));
2829 
2830 	if (fl) {
2831 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
2832 		    fl->desc);
2833 
2834 		if (fl->sdesc)
2835 			free_fl_sdesc(sc, fl);
2836 
2837 		if (mtx_initialized(&fl->fl_lock))
2838 			mtx_destroy(&fl->fl_lock);
2839 
2840 		bzero(fl, sizeof(*fl));
2841 	}
2842 
2843 	return (0);
2844 }
2845 
2846 static void
2847 add_fl_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
2848     struct sge_fl *fl)
2849 {
2850 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2851 
2852 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
2853 	    "freelist");
2854 	children = SYSCTL_CHILDREN(oid);
2855 
2856 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
2857 	    CTLTYPE_INT | CTLFLAG_RD, &fl->cntxt_id, 0, sysctl_uint16, "I",
2858 	    "SGE context id of the freelist");
2859 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL,
2860 	    fl_pad ? 1 : 0, "padding enabled");
2861 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL,
2862 	    fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled");
2863 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
2864 	    0, "consumer index");
2865 	if (fl->flags & FL_BUF_PACKING) {
2866 		SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
2867 		    CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
2868 	}
2869 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
2870 	    0, "producer index");
2871 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_allocated",
2872 	    CTLFLAG_RD, &fl->mbuf_allocated, "# of mbuf allocated");
2873 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_inlined",
2874 	    CTLFLAG_RD, &fl->mbuf_inlined, "# of mbuf inlined in clusters");
2875 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
2876 	    CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
2877 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
2878 	    CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
2879 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
2880 	    CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
2881 }
2882 
2883 static int
2884 alloc_fwq(struct adapter *sc)
2885 {
2886 	int rc, intr_idx;
2887 	struct sge_iq *fwq = &sc->sge.fwq;
2888 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
2889 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2890 
2891 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE);
2892 	fwq->flags |= IQ_INTR;	/* always */
2893 	intr_idx = sc->intr_count > 1 ? 1 : 0;
2894 	fwq->set_tcb_rpl = t4_filter_rpl;
2895 	fwq->l2t_write_rpl = do_l2t_write_rpl;
2896 	rc = alloc_iq_fl(&sc->port[0]->vi[0], fwq, NULL, intr_idx, -1);
2897 	if (rc != 0) {
2898 		device_printf(sc->dev,
2899 		    "failed to create firmware event queue: %d\n", rc);
2900 		return (rc);
2901 	}
2902 
2903 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD,
2904 	    NULL, "firmware event queue");
2905 	children = SYSCTL_CHILDREN(oid);
2906 
2907 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "abs_id",
2908 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->abs_id, 0, sysctl_uint16, "I",
2909 	    "absolute id of the queue");
2910 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cntxt_id",
2911 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->cntxt_id, 0, sysctl_uint16, "I",
2912 	    "SGE context id of the queue");
2913 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cidx",
2914 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->cidx, 0, sysctl_uint16, "I",
2915 	    "consumer index");
2916 
2917 	return (0);
2918 }
2919 
2920 static int
2921 free_fwq(struct adapter *sc)
2922 {
2923 	return free_iq_fl(NULL, &sc->sge.fwq, NULL);
2924 }
2925 
2926 static int
2927 alloc_mgmtq(struct adapter *sc)
2928 {
2929 	int rc;
2930 	struct sge_wrq *mgmtq = &sc->sge.mgmtq;
2931 	char name[16];
2932 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
2933 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2934 
2935 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "mgmtq", CTLFLAG_RD,
2936 	    NULL, "management queue");
2937 
2938 	snprintf(name, sizeof(name), "%s mgmtq", device_get_nameunit(sc->dev));
2939 	init_eq(sc, &mgmtq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[0]->tx_chan,
2940 	    sc->sge.fwq.cntxt_id, name);
2941 	rc = alloc_wrq(sc, NULL, mgmtq, oid);
2942 	if (rc != 0) {
2943 		device_printf(sc->dev,
2944 		    "failed to create management queue: %d\n", rc);
2945 		return (rc);
2946 	}
2947 
2948 	return (0);
2949 }
2950 
2951 static int
2952 free_mgmtq(struct adapter *sc)
2953 {
2954 
2955 	return free_wrq(sc, &sc->sge.mgmtq);
2956 }
2957 
2958 int
2959 tnl_cong(struct port_info *pi, int drop)
2960 {
2961 
2962 	if (drop == -1)
2963 		return (-1);
2964 	else if (drop == 1)
2965 		return (0);
2966 	else
2967 		return (pi->rx_chan_map);
2968 }
2969 
2970 static int
2971 alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int intr_idx, int idx,
2972     struct sysctl_oid *oid)
2973 {
2974 	int rc;
2975 	struct sysctl_oid_list *children;
2976 	char name[16];
2977 
2978 	rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, intr_idx,
2979 	    tnl_cong(vi->pi, cong_drop));
2980 	if (rc != 0)
2981 		return (rc);
2982 
2983 	/*
2984 	 * The freelist is just barely above the starvation threshold right now,
2985 	 * fill it up a bit more.
2986 	 */
2987 	FL_LOCK(&rxq->fl);
2988 	refill_fl(vi->pi->adapter, &rxq->fl, 128);
2989 	FL_UNLOCK(&rxq->fl);
2990 
2991 #if defined(INET) || defined(INET6)
2992 	rc = tcp_lro_init(&rxq->lro);
2993 	if (rc != 0)
2994 		return (rc);
2995 	rxq->lro.ifp = vi->ifp; /* also indicates LRO init'ed */
2996 
2997 	if (vi->ifp->if_capenable & IFCAP_LRO)
2998 		rxq->iq.flags |= IQ_LRO_ENABLED;
2999 #endif
3000 	rxq->ifp = vi->ifp;
3001 
3002 	children = SYSCTL_CHILDREN(oid);
3003 
3004 	snprintf(name, sizeof(name), "%d", idx);
3005 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3006 	    NULL, "rx queue");
3007 	children = SYSCTL_CHILDREN(oid);
3008 
3009 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "abs_id",
3010 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.abs_id, 0, sysctl_uint16, "I",
3011 	    "absolute id of the queue");
3012 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cntxt_id",
3013 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cntxt_id, 0, sysctl_uint16, "I",
3014 	    "SGE context id of the queue");
3015 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3016 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cidx, 0, sysctl_uint16, "I",
3017 	    "consumer index");
3018 #if defined(INET) || defined(INET6)
3019 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
3020 	    &rxq->lro.lro_queued, 0, NULL);
3021 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
3022 	    &rxq->lro.lro_flushed, 0, NULL);
3023 #endif
3024 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
3025 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
3026 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_extraction",
3027 	    CTLFLAG_RD, &rxq->vlan_extraction,
3028 	    "# of times hardware extracted 802.1Q tag");
3029 
3030 	add_fl_sysctls(&vi->ctx, oid, &rxq->fl);
3031 
3032 	return (rc);
3033 }
3034 
3035 static int
3036 free_rxq(struct vi_info *vi, struct sge_rxq *rxq)
3037 {
3038 	int rc;
3039 
3040 #if defined(INET) || defined(INET6)
3041 	if (rxq->lro.ifp) {
3042 		tcp_lro_free(&rxq->lro);
3043 		rxq->lro.ifp = NULL;
3044 	}
3045 #endif
3046 
3047 	rc = free_iq_fl(vi, &rxq->iq, &rxq->fl);
3048 	if (rc == 0)
3049 		bzero(rxq, sizeof(*rxq));
3050 
3051 	return (rc);
3052 }
3053 
3054 #ifdef TCP_OFFLOAD
3055 static int
3056 alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq,
3057     int intr_idx, int idx, struct sysctl_oid *oid)
3058 {
3059 	int rc;
3060 	struct sysctl_oid_list *children;
3061 	char name[16];
3062 
3063 	rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx,
3064 	    vi->pi->rx_chan_map);
3065 	if (rc != 0)
3066 		return (rc);
3067 
3068 	children = SYSCTL_CHILDREN(oid);
3069 
3070 	snprintf(name, sizeof(name), "%d", idx);
3071 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3072 	    NULL, "rx queue");
3073 	children = SYSCTL_CHILDREN(oid);
3074 
3075 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "abs_id",
3076 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.abs_id, 0, sysctl_uint16,
3077 	    "I", "absolute id of the queue");
3078 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cntxt_id",
3079 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cntxt_id, 0, sysctl_uint16,
3080 	    "I", "SGE context id of the queue");
3081 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3082 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cidx, 0, sysctl_uint16, "I",
3083 	    "consumer index");
3084 
3085 	add_fl_sysctls(&vi->ctx, oid, &ofld_rxq->fl);
3086 
3087 	return (rc);
3088 }
3089 
3090 static int
3091 free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq)
3092 {
3093 	int rc;
3094 
3095 	rc = free_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl);
3096 	if (rc == 0)
3097 		bzero(ofld_rxq, sizeof(*ofld_rxq));
3098 
3099 	return (rc);
3100 }
3101 #endif
3102 
3103 #ifdef DEV_NETMAP
3104 static int
3105 alloc_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int intr_idx,
3106     int idx, struct sysctl_oid *oid)
3107 {
3108 	int rc;
3109 	struct sysctl_oid_list *children;
3110 	struct sysctl_ctx_list *ctx;
3111 	char name[16];
3112 	size_t len;
3113 	struct adapter *sc = vi->pi->adapter;
3114 	struct netmap_adapter *na = NA(vi->ifp);
3115 
3116 	MPASS(na != NULL);
3117 
3118 	len = vi->qsize_rxq * IQ_ESIZE;
3119 	rc = alloc_ring(sc, len, &nm_rxq->iq_desc_tag, &nm_rxq->iq_desc_map,
3120 	    &nm_rxq->iq_ba, (void **)&nm_rxq->iq_desc);
3121 	if (rc != 0)
3122 		return (rc);
3123 
3124 	len = na->num_rx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3125 	rc = alloc_ring(sc, len, &nm_rxq->fl_desc_tag, &nm_rxq->fl_desc_map,
3126 	    &nm_rxq->fl_ba, (void **)&nm_rxq->fl_desc);
3127 	if (rc != 0)
3128 		return (rc);
3129 
3130 	nm_rxq->vi = vi;
3131 	nm_rxq->nid = idx;
3132 	nm_rxq->iq_cidx = 0;
3133 	nm_rxq->iq_sidx = vi->qsize_rxq - sc->params.sge.spg_len / IQ_ESIZE;
3134 	nm_rxq->iq_gen = F_RSPD_GEN;
3135 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
3136 	nm_rxq->fl_sidx = na->num_rx_desc;
3137 	nm_rxq->intr_idx = intr_idx;
3138 
3139 	ctx = &vi->ctx;
3140 	children = SYSCTL_CHILDREN(oid);
3141 
3142 	snprintf(name, sizeof(name), "%d", idx);
3143 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL,
3144 	    "rx queue");
3145 	children = SYSCTL_CHILDREN(oid);
3146 
3147 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3148 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_abs_id, 0, sysctl_uint16,
3149 	    "I", "absolute id of the queue");
3150 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3151 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cntxt_id, 0, sysctl_uint16,
3152 	    "I", "SGE context id of the queue");
3153 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3154 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cidx, 0, sysctl_uint16, "I",
3155 	    "consumer index");
3156 
3157 	children = SYSCTL_CHILDREN(oid);
3158 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3159 	    "freelist");
3160 	children = SYSCTL_CHILDREN(oid);
3161 
3162 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3163 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->fl_cntxt_id, 0, sysctl_uint16,
3164 	    "I", "SGE context id of the freelist");
3165 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
3166 	    &nm_rxq->fl_cidx, 0, "consumer index");
3167 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
3168 	    &nm_rxq->fl_pidx, 0, "producer index");
3169 
3170 	return (rc);
3171 }
3172 
3173 
3174 static int
3175 free_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
3176 {
3177 	struct adapter *sc = vi->pi->adapter;
3178 
3179 	free_ring(sc, nm_rxq->iq_desc_tag, nm_rxq->iq_desc_map, nm_rxq->iq_ba,
3180 	    nm_rxq->iq_desc);
3181 	free_ring(sc, nm_rxq->fl_desc_tag, nm_rxq->fl_desc_map, nm_rxq->fl_ba,
3182 	    nm_rxq->fl_desc);
3183 
3184 	return (0);
3185 }
3186 
3187 static int
3188 alloc_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq, int iqidx, int idx,
3189     struct sysctl_oid *oid)
3190 {
3191 	int rc;
3192 	size_t len;
3193 	struct port_info *pi = vi->pi;
3194 	struct adapter *sc = pi->adapter;
3195 	struct netmap_adapter *na = NA(vi->ifp);
3196 	char name[16];
3197 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3198 
3199 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3200 	rc = alloc_ring(sc, len, &nm_txq->desc_tag, &nm_txq->desc_map,
3201 	    &nm_txq->ba, (void **)&nm_txq->desc);
3202 	if (rc)
3203 		return (rc);
3204 
3205 	nm_txq->pidx = nm_txq->cidx = 0;
3206 	nm_txq->sidx = na->num_tx_desc;
3207 	nm_txq->nid = idx;
3208 	nm_txq->iqidx = iqidx;
3209 	nm_txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3210 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_VF_VLD(1) |
3211 	    V_TXPKT_VF(vi->viid));
3212 
3213 	snprintf(name, sizeof(name), "%d", idx);
3214 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3215 	    NULL, "netmap tx queue");
3216 	children = SYSCTL_CHILDREN(oid);
3217 
3218 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3219 	    &nm_txq->cntxt_id, 0, "SGE context id of the queue");
3220 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3221 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->cidx, 0, sysctl_uint16, "I",
3222 	    "consumer index");
3223 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3224 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->pidx, 0, sysctl_uint16, "I",
3225 	    "producer index");
3226 
3227 	return (rc);
3228 }
3229 
3230 static int
3231 free_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
3232 {
3233 	struct adapter *sc = vi->pi->adapter;
3234 
3235 	free_ring(sc, nm_txq->desc_tag, nm_txq->desc_map, nm_txq->ba,
3236 	    nm_txq->desc);
3237 
3238 	return (0);
3239 }
3240 #endif
3241 
3242 static int
3243 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
3244 {
3245 	int rc, cntxt_id;
3246 	struct fw_eq_ctrl_cmd c;
3247 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3248 
3249 	bzero(&c, sizeof(c));
3250 
3251 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
3252 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
3253 	    V_FW_EQ_CTRL_CMD_VFN(0));
3254 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
3255 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
3256 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid));
3257 	c.physeqid_pkd = htobe32(0);
3258 	c.fetchszm_to_iqid =
3259 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3260 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
3261 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
3262 	c.dcaen_to_eqsize =
3263 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3264 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3265 		V_FW_EQ_CTRL_CMD_EQSIZE(qsize));
3266 	c.eqaddr = htobe64(eq->ba);
3267 
3268 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3269 	if (rc != 0) {
3270 		device_printf(sc->dev,
3271 		    "failed to create control queue %d: %d\n", eq->tx_chan, rc);
3272 		return (rc);
3273 	}
3274 	eq->flags |= EQ_ALLOCATED;
3275 
3276 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
3277 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3278 	if (cntxt_id >= sc->sge.neq)
3279 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3280 		cntxt_id, sc->sge.neq - 1);
3281 	sc->sge.eqmap[cntxt_id] = eq;
3282 
3283 	return (rc);
3284 }
3285 
3286 static int
3287 eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3288 {
3289 	int rc, cntxt_id;
3290 	struct fw_eq_eth_cmd c;
3291 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3292 
3293 	bzero(&c, sizeof(c));
3294 
3295 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
3296 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
3297 	    V_FW_EQ_ETH_CMD_VFN(0));
3298 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
3299 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
3300 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
3301 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
3302 	c.fetchszm_to_iqid =
3303 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3304 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
3305 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
3306 	c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3307 	    V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3308 	    V_FW_EQ_ETH_CMD_EQSIZE(qsize));
3309 	c.eqaddr = htobe64(eq->ba);
3310 
3311 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3312 	if (rc != 0) {
3313 		device_printf(vi->dev,
3314 		    "failed to create Ethernet egress queue: %d\n", rc);
3315 		return (rc);
3316 	}
3317 	eq->flags |= EQ_ALLOCATED;
3318 
3319 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
3320 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3321 	if (cntxt_id >= sc->sge.neq)
3322 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3323 		cntxt_id, sc->sge.neq - 1);
3324 	sc->sge.eqmap[cntxt_id] = eq;
3325 
3326 	return (rc);
3327 }
3328 
3329 #ifdef TCP_OFFLOAD
3330 static int
3331 ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3332 {
3333 	int rc, cntxt_id;
3334 	struct fw_eq_ofld_cmd c;
3335 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3336 
3337 	bzero(&c, sizeof(c));
3338 
3339 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
3340 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
3341 	    V_FW_EQ_OFLD_CMD_VFN(0));
3342 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
3343 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
3344 	c.fetchszm_to_iqid =
3345 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3346 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
3347 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
3348 	c.dcaen_to_eqsize =
3349 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3350 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3351 		V_FW_EQ_OFLD_CMD_EQSIZE(qsize));
3352 	c.eqaddr = htobe64(eq->ba);
3353 
3354 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3355 	if (rc != 0) {
3356 		device_printf(vi->dev,
3357 		    "failed to create egress queue for TCP offload: %d\n", rc);
3358 		return (rc);
3359 	}
3360 	eq->flags |= EQ_ALLOCATED;
3361 
3362 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
3363 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3364 	if (cntxt_id >= sc->sge.neq)
3365 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3366 		cntxt_id, sc->sge.neq - 1);
3367 	sc->sge.eqmap[cntxt_id] = eq;
3368 
3369 	return (rc);
3370 }
3371 #endif
3372 
3373 static int
3374 alloc_eq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3375 {
3376 	int rc, qsize;
3377 	size_t len;
3378 
3379 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
3380 
3381 	qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3382 	len = qsize * EQ_ESIZE;
3383 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
3384 	    &eq->ba, (void **)&eq->desc);
3385 	if (rc)
3386 		return (rc);
3387 
3388 	eq->pidx = eq->cidx = 0;
3389 	eq->equeqidx = eq->dbidx = 0;
3390 	eq->doorbells = sc->doorbells;
3391 
3392 	switch (eq->flags & EQ_TYPEMASK) {
3393 	case EQ_CTRL:
3394 		rc = ctrl_eq_alloc(sc, eq);
3395 		break;
3396 
3397 	case EQ_ETH:
3398 		rc = eth_eq_alloc(sc, vi, eq);
3399 		break;
3400 
3401 #ifdef TCP_OFFLOAD
3402 	case EQ_OFLD:
3403 		rc = ofld_eq_alloc(sc, vi, eq);
3404 		break;
3405 #endif
3406 
3407 	default:
3408 		panic("%s: invalid eq type %d.", __func__,
3409 		    eq->flags & EQ_TYPEMASK);
3410 	}
3411 	if (rc != 0) {
3412 		device_printf(sc->dev,
3413 		    "failed to allocate egress queue(%d): %d\n",
3414 		    eq->flags & EQ_TYPEMASK, rc);
3415 	}
3416 
3417 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
3418 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
3419 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
3420 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3421 		uint32_t mask = (1 << s_qpp) - 1;
3422 		volatile uint8_t *udb;
3423 
3424 		udb = sc->udbs_base + UDBS_DB_OFFSET;
3425 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
3426 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
3427 		if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
3428 	    		clrbit(&eq->doorbells, DOORBELL_WCWR);
3429 		else {
3430 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
3431 			eq->udb_qid = 0;
3432 		}
3433 		eq->udb = (volatile void *)udb;
3434 	}
3435 
3436 	return (rc);
3437 }
3438 
3439 static int
3440 free_eq(struct adapter *sc, struct sge_eq *eq)
3441 {
3442 	int rc;
3443 
3444 	if (eq->flags & EQ_ALLOCATED) {
3445 		switch (eq->flags & EQ_TYPEMASK) {
3446 		case EQ_CTRL:
3447 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
3448 			    eq->cntxt_id);
3449 			break;
3450 
3451 		case EQ_ETH:
3452 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
3453 			    eq->cntxt_id);
3454 			break;
3455 
3456 #ifdef TCP_OFFLOAD
3457 		case EQ_OFLD:
3458 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
3459 			    eq->cntxt_id);
3460 			break;
3461 #endif
3462 
3463 		default:
3464 			panic("%s: invalid eq type %d.", __func__,
3465 			    eq->flags & EQ_TYPEMASK);
3466 		}
3467 		if (rc != 0) {
3468 			device_printf(sc->dev,
3469 			    "failed to free egress queue (%d): %d\n",
3470 			    eq->flags & EQ_TYPEMASK, rc);
3471 			return (rc);
3472 		}
3473 		eq->flags &= ~EQ_ALLOCATED;
3474 	}
3475 
3476 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
3477 
3478 	if (mtx_initialized(&eq->eq_lock))
3479 		mtx_destroy(&eq->eq_lock);
3480 
3481 	bzero(eq, sizeof(*eq));
3482 	return (0);
3483 }
3484 
3485 static int
3486 alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq,
3487     struct sysctl_oid *oid)
3488 {
3489 	int rc;
3490 	struct sysctl_ctx_list *ctx = vi ? &vi->ctx : &sc->ctx;
3491 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3492 
3493 	rc = alloc_eq(sc, vi, &wrq->eq);
3494 	if (rc)
3495 		return (rc);
3496 
3497 	wrq->adapter = sc;
3498 	TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq);
3499 	TAILQ_INIT(&wrq->incomplete_wrs);
3500 	STAILQ_INIT(&wrq->wr_list);
3501 	wrq->nwr_pending = 0;
3502 	wrq->ndesc_needed = 0;
3503 
3504 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3505 	    &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
3506 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3507 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I",
3508 	    "consumer index");
3509 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx",
3510 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I",
3511 	    "producer index");
3512 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD,
3513 	    &wrq->tx_wrs_direct, "# of work requests (direct)");
3514 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD,
3515 	    &wrq->tx_wrs_copied, "# of work requests (copied)");
3516 
3517 	return (rc);
3518 }
3519 
3520 static int
3521 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
3522 {
3523 	int rc;
3524 
3525 	rc = free_eq(sc, &wrq->eq);
3526 	if (rc)
3527 		return (rc);
3528 
3529 	bzero(wrq, sizeof(*wrq));
3530 	return (0);
3531 }
3532 
3533 static int
3534 alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx,
3535     struct sysctl_oid *oid)
3536 {
3537 	int rc;
3538 	struct port_info *pi = vi->pi;
3539 	struct adapter *sc = pi->adapter;
3540 	struct sge_eq *eq = &txq->eq;
3541 	char name[16];
3542 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3543 
3544 	rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx, can_resume_eth_tx,
3545 	    M_CXGBE, M_WAITOK);
3546 	if (rc != 0) {
3547 		device_printf(sc->dev, "failed to allocate mp_ring: %d\n", rc);
3548 		return (rc);
3549 	}
3550 
3551 	rc = alloc_eq(sc, vi, eq);
3552 	if (rc != 0) {
3553 		mp_ring_free(txq->r);
3554 		txq->r = NULL;
3555 		return (rc);
3556 	}
3557 
3558 	/* Can't fail after this point. */
3559 
3560 	TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq);
3561 	txq->ifp = vi->ifp;
3562 	txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK);
3563 	txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3564 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_VF_VLD(1) |
3565 	    V_TXPKT_VF(vi->viid));
3566 	txq->tc_idx = -1;
3567 	txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE,
3568 	    M_ZERO | M_WAITOK);
3569 
3570 	snprintf(name, sizeof(name), "%d", idx);
3571 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3572 	    NULL, "tx queue");
3573 	children = SYSCTL_CHILDREN(oid);
3574 
3575 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3576 	    &eq->cntxt_id, 0, "SGE context id of the queue");
3577 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3578 	    CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I",
3579 	    "consumer index");
3580 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3581 	    CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I",
3582 	    "producer index");
3583 
3584 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "tc",
3585 	    CTLTYPE_INT | CTLFLAG_RW, vi, idx, sysctl_tc, "I",
3586 	    "traffic class (-1 means none)");
3587 
3588 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
3589 	    &txq->txcsum, "# of times hardware assisted with checksum");
3590 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_insertion",
3591 	    CTLFLAG_RD, &txq->vlan_insertion,
3592 	    "# of times hardware inserted 802.1Q tag");
3593 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
3594 	    &txq->tso_wrs, "# of TSO work requests");
3595 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
3596 	    &txq->imm_wrs, "# of work requests with immediate data");
3597 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
3598 	    &txq->sgl_wrs, "# of work requests with direct SGL");
3599 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
3600 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
3601 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_wrs",
3602 	    CTLFLAG_RD, &txq->txpkts0_wrs,
3603 	    "# of txpkts (type 0) work requests");
3604 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_wrs",
3605 	    CTLFLAG_RD, &txq->txpkts1_wrs,
3606 	    "# of txpkts (type 1) work requests");
3607 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_pkts",
3608 	    CTLFLAG_RD, &txq->txpkts0_pkts,
3609 	    "# of frames tx'd using type0 txpkts work requests");
3610 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_pkts",
3611 	    CTLFLAG_RD, &txq->txpkts1_pkts,
3612 	    "# of frames tx'd using type1 txpkts work requests");
3613 
3614 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_enqueues",
3615 	    CTLFLAG_RD, &txq->r->enqueues,
3616 	    "# of enqueues to the mp_ring for this queue");
3617 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_drops",
3618 	    CTLFLAG_RD, &txq->r->drops,
3619 	    "# of drops in the mp_ring for this queue");
3620 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_starts",
3621 	    CTLFLAG_RD, &txq->r->starts,
3622 	    "# of normal consumer starts in the mp_ring for this queue");
3623 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_stalls",
3624 	    CTLFLAG_RD, &txq->r->stalls,
3625 	    "# of consumer stalls in the mp_ring for this queue");
3626 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_restarts",
3627 	    CTLFLAG_RD, &txq->r->restarts,
3628 	    "# of consumer restarts in the mp_ring for this queue");
3629 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_abdications",
3630 	    CTLFLAG_RD, &txq->r->abdications,
3631 	    "# of consumer abdications in the mp_ring for this queue");
3632 
3633 	return (0);
3634 }
3635 
3636 static int
3637 free_txq(struct vi_info *vi, struct sge_txq *txq)
3638 {
3639 	int rc;
3640 	struct adapter *sc = vi->pi->adapter;
3641 	struct sge_eq *eq = &txq->eq;
3642 
3643 	rc = free_eq(sc, eq);
3644 	if (rc)
3645 		return (rc);
3646 
3647 	sglist_free(txq->gl);
3648 	free(txq->sdesc, M_CXGBE);
3649 	mp_ring_free(txq->r);
3650 
3651 	bzero(txq, sizeof(*txq));
3652 	return (0);
3653 }
3654 
3655 static void
3656 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
3657 {
3658 	bus_addr_t *ba = arg;
3659 
3660 	KASSERT(nseg == 1,
3661 	    ("%s meant for single segment mappings only.", __func__));
3662 
3663 	*ba = error ? 0 : segs->ds_addr;
3664 }
3665 
3666 static inline void
3667 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
3668 {
3669 	uint32_t n, v;
3670 
3671 	n = IDXDIFF(fl->pidx / 8, fl->dbidx, fl->sidx);
3672 	MPASS(n > 0);
3673 
3674 	wmb();
3675 	v = fl->dbval | V_PIDX(n);
3676 	if (fl->udb)
3677 		*fl->udb = htole32(v);
3678 	else
3679 		t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL), v);
3680 	IDXINCR(fl->dbidx, n, fl->sidx);
3681 }
3682 
3683 /*
3684  * Fills up the freelist by allocating up to 'n' buffers.  Buffers that are
3685  * recycled do not count towards this allocation budget.
3686  *
3687  * Returns non-zero to indicate that this freelist should be added to the list
3688  * of starving freelists.
3689  */
3690 static int
3691 refill_fl(struct adapter *sc, struct sge_fl *fl, int n)
3692 {
3693 	__be64 *d;
3694 	struct fl_sdesc *sd;
3695 	uintptr_t pa;
3696 	caddr_t cl;
3697 	struct cluster_layout *cll;
3698 	struct sw_zone_info *swz;
3699 	struct cluster_metadata *clm;
3700 	uint16_t max_pidx;
3701 	uint16_t hw_cidx = fl->hw_cidx;		/* stable snapshot */
3702 
3703 	FL_LOCK_ASSERT_OWNED(fl);
3704 
3705 	/*
3706 	 * We always stop at the beginning of the hardware descriptor that's just
3707 	 * before the one with the hw cidx.  This is to avoid hw pidx = hw cidx,
3708 	 * which would mean an empty freelist to the chip.
3709 	 */
3710 	max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1;
3711 	if (fl->pidx == max_pidx * 8)
3712 		return (0);
3713 
3714 	d = &fl->desc[fl->pidx];
3715 	sd = &fl->sdesc[fl->pidx];
3716 	cll = &fl->cll_def;	/* default layout */
3717 	swz = &sc->sge.sw_zone_info[cll->zidx];
3718 
3719 	while (n > 0) {
3720 
3721 		if (sd->cl != NULL) {
3722 
3723 			if (sd->nmbuf == 0) {
3724 				/*
3725 				 * Fast recycle without involving any atomics on
3726 				 * the cluster's metadata (if the cluster has
3727 				 * metadata).  This happens when all frames
3728 				 * received in the cluster were small enough to
3729 				 * fit within a single mbuf each.
3730 				 */
3731 				fl->cl_fast_recycled++;
3732 #ifdef INVARIANTS
3733 				clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
3734 				if (clm != NULL)
3735 					MPASS(clm->refcount == 1);
3736 #endif
3737 				goto recycled_fast;
3738 			}
3739 
3740 			/*
3741 			 * Cluster is guaranteed to have metadata.  Clusters
3742 			 * without metadata always take the fast recycle path
3743 			 * when they're recycled.
3744 			 */
3745 			clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
3746 			MPASS(clm != NULL);
3747 
3748 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
3749 				fl->cl_recycled++;
3750 				counter_u64_add(extfree_rels, 1);
3751 				goto recycled;
3752 			}
3753 			sd->cl = NULL;	/* gave up my reference */
3754 		}
3755 		MPASS(sd->cl == NULL);
3756 alloc:
3757 		cl = uma_zalloc(swz->zone, M_NOWAIT);
3758 		if (__predict_false(cl == NULL)) {
3759 			if (cll == &fl->cll_alt || fl->cll_alt.zidx == -1 ||
3760 			    fl->cll_def.zidx == fl->cll_alt.zidx)
3761 				break;
3762 
3763 			/* fall back to the safe zone */
3764 			cll = &fl->cll_alt;
3765 			swz = &sc->sge.sw_zone_info[cll->zidx];
3766 			goto alloc;
3767 		}
3768 		fl->cl_allocated++;
3769 		n--;
3770 
3771 		pa = pmap_kextract((vm_offset_t)cl);
3772 		pa += cll->region1;
3773 		sd->cl = cl;
3774 		sd->cll = *cll;
3775 		*d = htobe64(pa | cll->hwidx);
3776 		clm = cl_metadata(sc, fl, cll, cl);
3777 		if (clm != NULL) {
3778 recycled:
3779 #ifdef INVARIANTS
3780 			clm->sd = sd;
3781 #endif
3782 			clm->refcount = 1;
3783 		}
3784 		sd->nmbuf = 0;
3785 recycled_fast:
3786 		d++;
3787 		sd++;
3788 		if (__predict_false(++fl->pidx % 8 == 0)) {
3789 			uint16_t pidx = fl->pidx / 8;
3790 
3791 			if (__predict_false(pidx == fl->sidx)) {
3792 				fl->pidx = 0;
3793 				pidx = 0;
3794 				sd = fl->sdesc;
3795 				d = fl->desc;
3796 			}
3797 			if (pidx == max_pidx)
3798 				break;
3799 
3800 			if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4)
3801 				ring_fl_db(sc, fl);
3802 		}
3803 	}
3804 
3805 	if (fl->pidx / 8 != fl->dbidx)
3806 		ring_fl_db(sc, fl);
3807 
3808 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
3809 }
3810 
3811 /*
3812  * Attempt to refill all starving freelists.
3813  */
3814 static void
3815 refill_sfl(void *arg)
3816 {
3817 	struct adapter *sc = arg;
3818 	struct sge_fl *fl, *fl_temp;
3819 
3820 	mtx_assert(&sc->sfl_lock, MA_OWNED);
3821 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
3822 		FL_LOCK(fl);
3823 		refill_fl(sc, fl, 64);
3824 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
3825 			TAILQ_REMOVE(&sc->sfl, fl, link);
3826 			fl->flags &= ~FL_STARVING;
3827 		}
3828 		FL_UNLOCK(fl);
3829 	}
3830 
3831 	if (!TAILQ_EMPTY(&sc->sfl))
3832 		callout_schedule(&sc->sfl_callout, hz / 5);
3833 }
3834 
3835 static int
3836 alloc_fl_sdesc(struct sge_fl *fl)
3837 {
3838 
3839 	fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc), M_CXGBE,
3840 	    M_ZERO | M_WAITOK);
3841 
3842 	return (0);
3843 }
3844 
3845 static void
3846 free_fl_sdesc(struct adapter *sc, struct sge_fl *fl)
3847 {
3848 	struct fl_sdesc *sd;
3849 	struct cluster_metadata *clm;
3850 	struct cluster_layout *cll;
3851 	int i;
3852 
3853 	sd = fl->sdesc;
3854 	for (i = 0; i < fl->sidx * 8; i++, sd++) {
3855 		if (sd->cl == NULL)
3856 			continue;
3857 
3858 		cll = &sd->cll;
3859 		clm = cl_metadata(sc, fl, cll, sd->cl);
3860 		if (sd->nmbuf == 0)
3861 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
3862 		else if (clm && atomic_fetchadd_int(&clm->refcount, -1) == 1) {
3863 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
3864 			counter_u64_add(extfree_rels, 1);
3865 		}
3866 		sd->cl = NULL;
3867 	}
3868 
3869 	free(fl->sdesc, M_CXGBE);
3870 	fl->sdesc = NULL;
3871 }
3872 
3873 static inline void
3874 get_pkt_gl(struct mbuf *m, struct sglist *gl)
3875 {
3876 	int rc;
3877 
3878 	M_ASSERTPKTHDR(m);
3879 
3880 	sglist_reset(gl);
3881 	rc = sglist_append_mbuf(gl, m);
3882 	if (__predict_false(rc != 0)) {
3883 		panic("%s: mbuf %p (%d segs) was vetted earlier but now fails "
3884 		    "with %d.", __func__, m, mbuf_nsegs(m), rc);
3885 	}
3886 
3887 	KASSERT(gl->sg_nseg == mbuf_nsegs(m),
3888 	    ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m,
3889 	    mbuf_nsegs(m), gl->sg_nseg));
3890 	KASSERT(gl->sg_nseg > 0 &&
3891 	    gl->sg_nseg <= (needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS),
3892 	    ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__,
3893 		gl->sg_nseg, needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS));
3894 }
3895 
3896 /*
3897  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
3898  */
3899 static inline u_int
3900 txpkt_len16(u_int nsegs, u_int tso)
3901 {
3902 	u_int n;
3903 
3904 	MPASS(nsegs > 0);
3905 
3906 	nsegs--; /* first segment is part of ulptx_sgl */
3907 	n = sizeof(struct fw_eth_tx_pkt_wr) + sizeof(struct cpl_tx_pkt_core) +
3908 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
3909 	if (tso)
3910 		n += sizeof(struct cpl_tx_pkt_lso_core);
3911 
3912 	return (howmany(n, 16));
3913 }
3914 
3915 /*
3916  * len16 for a txpkts type 0 WR with a GL.  Does not include the firmware work
3917  * request header.
3918  */
3919 static inline u_int
3920 txpkts0_len16(u_int nsegs)
3921 {
3922 	u_int n;
3923 
3924 	MPASS(nsegs > 0);
3925 
3926 	nsegs--; /* first segment is part of ulptx_sgl */
3927 	n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) +
3928 	    sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) +
3929 	    8 * ((3 * nsegs) / 2 + (nsegs & 1));
3930 
3931 	return (howmany(n, 16));
3932 }
3933 
3934 /*
3935  * len16 for a txpkts type 1 WR with a GL.  Does not include the firmware work
3936  * request header.
3937  */
3938 static inline u_int
3939 txpkts1_len16(void)
3940 {
3941 	u_int n;
3942 
3943 	n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl);
3944 
3945 	return (howmany(n, 16));
3946 }
3947 
3948 static inline u_int
3949 imm_payload(u_int ndesc)
3950 {
3951 	u_int n;
3952 
3953 	n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) -
3954 	    sizeof(struct cpl_tx_pkt_core);
3955 
3956 	return (n);
3957 }
3958 
3959 /*
3960  * Write a txpkt WR for this packet to the hardware descriptors, update the
3961  * software descriptor, and advance the pidx.  It is guaranteed that enough
3962  * descriptors are available.
3963  *
3964  * The return value is the # of hardware descriptors used.
3965  */
3966 static u_int
3967 write_txpkt_wr(struct sge_txq *txq, struct fw_eth_tx_pkt_wr *wr,
3968     struct mbuf *m0, u_int available)
3969 {
3970 	struct sge_eq *eq = &txq->eq;
3971 	struct tx_sdesc *txsd;
3972 	struct cpl_tx_pkt_core *cpl;
3973 	uint32_t ctrl;	/* used in many unrelated places */
3974 	uint64_t ctrl1;
3975 	int len16, ndesc, pktlen, nsegs;
3976 	caddr_t dst;
3977 
3978 	TXQ_LOCK_ASSERT_OWNED(txq);
3979 	M_ASSERTPKTHDR(m0);
3980 	MPASS(available > 0 && available < eq->sidx);
3981 
3982 	len16 = mbuf_len16(m0);
3983 	nsegs = mbuf_nsegs(m0);
3984 	pktlen = m0->m_pkthdr.len;
3985 	ctrl = sizeof(struct cpl_tx_pkt_core);
3986 	if (needs_tso(m0))
3987 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
3988 	else if (pktlen <= imm_payload(2) && available >= 2) {
3989 		/* Immediate data.  Recalculate len16 and set nsegs to 0. */
3990 		ctrl += pktlen;
3991 		len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) +
3992 		    sizeof(struct cpl_tx_pkt_core) + pktlen, 16);
3993 		nsegs = 0;
3994 	}
3995 	ndesc = howmany(len16, EQ_ESIZE / 16);
3996 	MPASS(ndesc <= available);
3997 
3998 	/* Firmware work request header */
3999 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4000 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
4001 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4002 
4003 	ctrl = V_FW_WR_LEN16(len16);
4004 	wr->equiq_to_len16 = htobe32(ctrl);
4005 	wr->r3 = 0;
4006 
4007 	if (needs_tso(m0)) {
4008 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4009 
4010 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4011 		    m0->m_pkthdr.l4hlen > 0,
4012 		    ("%s: mbuf %p needs TSO but missing header lengths",
4013 			__func__, m0));
4014 
4015 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4016 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4017 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4018 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4019 			ctrl |= V_LSO_ETHHDR_LEN(1);
4020 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4021 			ctrl |= F_LSO_IPV6;
4022 
4023 		lso->lso_ctrl = htobe32(ctrl);
4024 		lso->ipid_ofst = htobe16(0);
4025 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4026 		lso->seqno_offset = htobe32(0);
4027 		lso->len = htobe32(pktlen);
4028 
4029 		cpl = (void *)(lso + 1);
4030 
4031 		txq->tso_wrs++;
4032 	} else
4033 		cpl = (void *)(wr + 1);
4034 
4035 	/* Checksum offload */
4036 	ctrl1 = 0;
4037 	if (needs_l3_csum(m0) == 0)
4038 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4039 	if (needs_l4_csum(m0) == 0)
4040 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4041 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4042 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4043 		txq->txcsum++;	/* some hardware assistance provided */
4044 
4045 	/* VLAN tag insertion */
4046 	if (needs_vlan_insertion(m0)) {
4047 		ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4048 		txq->vlan_insertion++;
4049 	}
4050 
4051 	/* CPL header */
4052 	cpl->ctrl0 = txq->cpl_ctrl0;
4053 	cpl->pack = 0;
4054 	cpl->len = htobe16(pktlen);
4055 	cpl->ctrl1 = htobe64(ctrl1);
4056 
4057 	/* SGL */
4058 	dst = (void *)(cpl + 1);
4059 	if (nsegs > 0) {
4060 
4061 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4062 		txq->sgl_wrs++;
4063 	} else {
4064 		struct mbuf *m;
4065 
4066 		for (m = m0; m != NULL; m = m->m_next) {
4067 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4068 #ifdef INVARIANTS
4069 			pktlen -= m->m_len;
4070 #endif
4071 		}
4072 #ifdef INVARIANTS
4073 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
4074 #endif
4075 		txq->imm_wrs++;
4076 	}
4077 
4078 	txq->txpkt_wrs++;
4079 
4080 	txsd = &txq->sdesc[eq->pidx];
4081 	txsd->m = m0;
4082 	txsd->desc_used = ndesc;
4083 
4084 	return (ndesc);
4085 }
4086 
4087 static int
4088 try_txpkts(struct mbuf *m, struct mbuf *n, struct txpkts *txp, u_int available)
4089 {
4090 	u_int needed, nsegs1, nsegs2, l1, l2;
4091 
4092 	if (cannot_use_txpkts(m) || cannot_use_txpkts(n))
4093 		return (1);
4094 
4095 	nsegs1 = mbuf_nsegs(m);
4096 	nsegs2 = mbuf_nsegs(n);
4097 	if (nsegs1 + nsegs2 == 2) {
4098 		txp->wr_type = 1;
4099 		l1 = l2 = txpkts1_len16();
4100 	} else {
4101 		txp->wr_type = 0;
4102 		l1 = txpkts0_len16(nsegs1);
4103 		l2 = txpkts0_len16(nsegs2);
4104 	}
4105 	txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) + l1 + l2;
4106 	needed = howmany(txp->len16, EQ_ESIZE / 16);
4107 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4108 		return (1);
4109 
4110 	txp->plen = m->m_pkthdr.len + n->m_pkthdr.len;
4111 	if (txp->plen > 65535)
4112 		return (1);
4113 
4114 	txp->npkt = 2;
4115 	set_mbuf_len16(m, l1);
4116 	set_mbuf_len16(n, l2);
4117 
4118 	return (0);
4119 }
4120 
4121 static int
4122 add_to_txpkts(struct mbuf *m, struct txpkts *txp, u_int available)
4123 {
4124 	u_int plen, len16, needed, nsegs;
4125 
4126 	MPASS(txp->wr_type == 0 || txp->wr_type == 1);
4127 
4128 	nsegs = mbuf_nsegs(m);
4129 	if (needs_tso(m) || (txp->wr_type == 1 && nsegs != 1))
4130 		return (1);
4131 
4132 	plen = txp->plen + m->m_pkthdr.len;
4133 	if (plen > 65535)
4134 		return (1);
4135 
4136 	if (txp->wr_type == 0)
4137 		len16 = txpkts0_len16(nsegs);
4138 	else
4139 		len16 = txpkts1_len16();
4140 	needed = howmany(txp->len16 + len16, EQ_ESIZE / 16);
4141 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4142 		return (1);
4143 
4144 	txp->npkt++;
4145 	txp->plen = plen;
4146 	txp->len16 += len16;
4147 	set_mbuf_len16(m, len16);
4148 
4149 	return (0);
4150 }
4151 
4152 /*
4153  * Write a txpkts WR for the packets in txp to the hardware descriptors, update
4154  * the software descriptor, and advance the pidx.  It is guaranteed that enough
4155  * descriptors are available.
4156  *
4157  * The return value is the # of hardware descriptors used.
4158  */
4159 static u_int
4160 write_txpkts_wr(struct sge_txq *txq, struct fw_eth_tx_pkts_wr *wr,
4161     struct mbuf *m0, const struct txpkts *txp, u_int available)
4162 {
4163 	struct sge_eq *eq = &txq->eq;
4164 	struct tx_sdesc *txsd;
4165 	struct cpl_tx_pkt_core *cpl;
4166 	uint32_t ctrl;
4167 	uint64_t ctrl1;
4168 	int ndesc, checkwrap;
4169 	struct mbuf *m;
4170 	void *flitp;
4171 
4172 	TXQ_LOCK_ASSERT_OWNED(txq);
4173 	MPASS(txp->npkt > 0);
4174 	MPASS(txp->plen < 65536);
4175 	MPASS(m0 != NULL);
4176 	MPASS(m0->m_nextpkt != NULL);
4177 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
4178 	MPASS(available > 0 && available < eq->sidx);
4179 
4180 	ndesc = howmany(txp->len16, EQ_ESIZE / 16);
4181 	MPASS(ndesc <= available);
4182 
4183 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4184 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
4185 	ctrl = V_FW_WR_LEN16(txp->len16);
4186 	wr->equiq_to_len16 = htobe32(ctrl);
4187 	wr->plen = htobe16(txp->plen);
4188 	wr->npkt = txp->npkt;
4189 	wr->r3 = 0;
4190 	wr->type = txp->wr_type;
4191 	flitp = wr + 1;
4192 
4193 	/*
4194 	 * At this point we are 16B into a hardware descriptor.  If checkwrap is
4195 	 * set then we know the WR is going to wrap around somewhere.  We'll
4196 	 * check for that at appropriate points.
4197 	 */
4198 	checkwrap = eq->sidx - ndesc < eq->pidx;
4199 	for (m = m0; m != NULL; m = m->m_nextpkt) {
4200 		if (txp->wr_type == 0) {
4201 			struct ulp_txpkt *ulpmc;
4202 			struct ulptx_idata *ulpsc;
4203 
4204 			/* ULP master command */
4205 			ulpmc = flitp;
4206 			ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
4207 			    V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid));
4208 			ulpmc->len = htobe32(mbuf_len16(m));
4209 
4210 			/* ULP subcommand */
4211 			ulpsc = (void *)(ulpmc + 1);
4212 			ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
4213 			    F_ULP_TX_SC_MORE);
4214 			ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
4215 
4216 			cpl = (void *)(ulpsc + 1);
4217 			if (checkwrap &&
4218 			    (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx])
4219 				cpl = (void *)&eq->desc[0];
4220 			txq->txpkts0_pkts += txp->npkt;
4221 			txq->txpkts0_wrs++;
4222 		} else {
4223 			cpl = flitp;
4224 			txq->txpkts1_pkts += txp->npkt;
4225 			txq->txpkts1_wrs++;
4226 		}
4227 
4228 		/* Checksum offload */
4229 		ctrl1 = 0;
4230 		if (needs_l3_csum(m) == 0)
4231 			ctrl1 |= F_TXPKT_IPCSUM_DIS;
4232 		if (needs_l4_csum(m) == 0)
4233 			ctrl1 |= F_TXPKT_L4CSUM_DIS;
4234 		if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4235 		    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4236 			txq->txcsum++;	/* some hardware assistance provided */
4237 
4238 		/* VLAN tag insertion */
4239 		if (needs_vlan_insertion(m)) {
4240 			ctrl1 |= F_TXPKT_VLAN_VLD |
4241 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
4242 			txq->vlan_insertion++;
4243 		}
4244 
4245 		/* CPL header */
4246 		cpl->ctrl0 = txq->cpl_ctrl0;
4247 		cpl->pack = 0;
4248 		cpl->len = htobe16(m->m_pkthdr.len);
4249 		cpl->ctrl1 = htobe64(ctrl1);
4250 
4251 		flitp = cpl + 1;
4252 		if (checkwrap &&
4253 		    (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
4254 			flitp = (void *)&eq->desc[0];
4255 
4256 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap);
4257 
4258 	}
4259 
4260 	txsd = &txq->sdesc[eq->pidx];
4261 	txsd->m = m0;
4262 	txsd->desc_used = ndesc;
4263 
4264 	return (ndesc);
4265 }
4266 
4267 /*
4268  * If the SGL ends on an address that is not 16 byte aligned, this function will
4269  * add a 0 filled flit at the end.
4270  */
4271 static void
4272 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap)
4273 {
4274 	struct sge_eq *eq = &txq->eq;
4275 	struct sglist *gl = txq->gl;
4276 	struct sglist_seg *seg;
4277 	__be64 *flitp, *wrap;
4278 	struct ulptx_sgl *usgl;
4279 	int i, nflits, nsegs;
4280 
4281 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
4282 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
4283 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
4284 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
4285 
4286 	get_pkt_gl(m, gl);
4287 	nsegs = gl->sg_nseg;
4288 	MPASS(nsegs > 0);
4289 
4290 	nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
4291 	flitp = (__be64 *)(*to);
4292 	wrap = (__be64 *)(&eq->desc[eq->sidx]);
4293 	seg = &gl->sg_segs[0];
4294 	usgl = (void *)flitp;
4295 
4296 	/*
4297 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
4298 	 * ring, so we're at least 16 bytes away from the status page.  There is
4299 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
4300 	 */
4301 
4302 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
4303 	    V_ULPTX_NSGE(nsegs));
4304 	usgl->len0 = htobe32(seg->ss_len);
4305 	usgl->addr0 = htobe64(seg->ss_paddr);
4306 	seg++;
4307 
4308 	if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) {
4309 
4310 		/* Won't wrap around at all */
4311 
4312 		for (i = 0; i < nsegs - 1; i++, seg++) {
4313 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
4314 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
4315 		}
4316 		if (i & 1)
4317 			usgl->sge[i / 2].len[1] = htobe32(0);
4318 		flitp += nflits;
4319 	} else {
4320 
4321 		/* Will wrap somewhere in the rest of the SGL */
4322 
4323 		/* 2 flits already written, write the rest flit by flit */
4324 		flitp = (void *)(usgl + 1);
4325 		for (i = 0; i < nflits - 2; i++) {
4326 			if (flitp == wrap)
4327 				flitp = (void *)eq->desc;
4328 			*flitp++ = get_flit(seg, nsegs - 1, i);
4329 		}
4330 	}
4331 
4332 	if (nflits & 1) {
4333 		MPASS(((uintptr_t)flitp) & 0xf);
4334 		*flitp++ = 0;
4335 	}
4336 
4337 	MPASS((((uintptr_t)flitp) & 0xf) == 0);
4338 	if (__predict_false(flitp == wrap))
4339 		*to = (void *)eq->desc;
4340 	else
4341 		*to = (void *)flitp;
4342 }
4343 
4344 static inline void
4345 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
4346 {
4347 
4348 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
4349 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
4350 
4351 	if (__predict_true((uintptr_t)(*to) + len <=
4352 	    (uintptr_t)&eq->desc[eq->sidx])) {
4353 		bcopy(from, *to, len);
4354 		(*to) += len;
4355 	} else {
4356 		int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
4357 
4358 		bcopy(from, *to, portion);
4359 		from += portion;
4360 		portion = len - portion;	/* remaining */
4361 		bcopy(from, (void *)eq->desc, portion);
4362 		(*to) = (caddr_t)eq->desc + portion;
4363 	}
4364 }
4365 
4366 static inline void
4367 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n)
4368 {
4369 	u_int db;
4370 
4371 	MPASS(n > 0);
4372 
4373 	db = eq->doorbells;
4374 	if (n > 1)
4375 		clrbit(&db, DOORBELL_WCWR);
4376 	wmb();
4377 
4378 	switch (ffs(db) - 1) {
4379 	case DOORBELL_UDB:
4380 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
4381 		break;
4382 
4383 	case DOORBELL_WCWR: {
4384 		volatile uint64_t *dst, *src;
4385 		int i;
4386 
4387 		/*
4388 		 * Queues whose 128B doorbell segment fits in the page do not
4389 		 * use relative qid (udb_qid is always 0).  Only queues with
4390 		 * doorbell segments can do WCWR.
4391 		 */
4392 		KASSERT(eq->udb_qid == 0 && n == 1,
4393 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
4394 		    __func__, eq->doorbells, n, eq->dbidx, eq));
4395 
4396 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
4397 		    UDBS_DB_OFFSET);
4398 		i = eq->dbidx;
4399 		src = (void *)&eq->desc[i];
4400 		while (src != (void *)&eq->desc[i + 1])
4401 			*dst++ = *src++;
4402 		wmb();
4403 		break;
4404 	}
4405 
4406 	case DOORBELL_UDBWC:
4407 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
4408 		wmb();
4409 		break;
4410 
4411 	case DOORBELL_KDB:
4412 		t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
4413 		    V_QID(eq->cntxt_id) | V_PIDX(n));
4414 		break;
4415 	}
4416 
4417 	IDXINCR(eq->dbidx, n, eq->sidx);
4418 }
4419 
4420 static inline u_int
4421 reclaimable_tx_desc(struct sge_eq *eq)
4422 {
4423 	uint16_t hw_cidx;
4424 
4425 	hw_cidx = read_hw_cidx(eq);
4426 	return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx));
4427 }
4428 
4429 static inline u_int
4430 total_available_tx_desc(struct sge_eq *eq)
4431 {
4432 	uint16_t hw_cidx, pidx;
4433 
4434 	hw_cidx = read_hw_cidx(eq);
4435 	pidx = eq->pidx;
4436 
4437 	if (pidx == hw_cidx)
4438 		return (eq->sidx - 1);
4439 	else
4440 		return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1);
4441 }
4442 
4443 static inline uint16_t
4444 read_hw_cidx(struct sge_eq *eq)
4445 {
4446 	struct sge_qstat *spg = (void *)&eq->desc[eq->sidx];
4447 	uint16_t cidx = spg->cidx;	/* stable snapshot */
4448 
4449 	return (be16toh(cidx));
4450 }
4451 
4452 /*
4453  * Reclaim 'n' descriptors approximately.
4454  */
4455 static u_int
4456 reclaim_tx_descs(struct sge_txq *txq, u_int n)
4457 {
4458 	struct tx_sdesc *txsd;
4459 	struct sge_eq *eq = &txq->eq;
4460 	u_int can_reclaim, reclaimed;
4461 
4462 	TXQ_LOCK_ASSERT_OWNED(txq);
4463 	MPASS(n > 0);
4464 
4465 	reclaimed = 0;
4466 	can_reclaim = reclaimable_tx_desc(eq);
4467 	while (can_reclaim && reclaimed < n) {
4468 		int ndesc;
4469 		struct mbuf *m, *nextpkt;
4470 
4471 		txsd = &txq->sdesc[eq->cidx];
4472 		ndesc = txsd->desc_used;
4473 
4474 		/* Firmware doesn't return "partial" credits. */
4475 		KASSERT(can_reclaim >= ndesc,
4476 		    ("%s: unexpected number of credits: %d, %d",
4477 		    __func__, can_reclaim, ndesc));
4478 
4479 		for (m = txsd->m; m != NULL; m = nextpkt) {
4480 			nextpkt = m->m_nextpkt;
4481 			m->m_nextpkt = NULL;
4482 			m_freem(m);
4483 		}
4484 		reclaimed += ndesc;
4485 		can_reclaim -= ndesc;
4486 		IDXINCR(eq->cidx, ndesc, eq->sidx);
4487 	}
4488 
4489 	return (reclaimed);
4490 }
4491 
4492 static void
4493 tx_reclaim(void *arg, int n)
4494 {
4495 	struct sge_txq *txq = arg;
4496 	struct sge_eq *eq = &txq->eq;
4497 
4498 	do {
4499 		if (TXQ_TRYLOCK(txq) == 0)
4500 			break;
4501 		n = reclaim_tx_descs(txq, 32);
4502 		if (eq->cidx == eq->pidx)
4503 			eq->equeqidx = eq->pidx;
4504 		TXQ_UNLOCK(txq);
4505 	} while (n > 0);
4506 }
4507 
4508 static __be64
4509 get_flit(struct sglist_seg *segs, int nsegs, int idx)
4510 {
4511 	int i = (idx / 3) * 2;
4512 
4513 	switch (idx % 3) {
4514 	case 0: {
4515 		__be64 rc;
4516 
4517 		rc = htobe32(segs[i].ss_len);
4518 		if (i + 1 < nsegs)
4519 			rc |= (uint64_t)htobe32(segs[i + 1].ss_len) << 32;
4520 
4521 		return (rc);
4522 	}
4523 	case 1:
4524 		return (htobe64(segs[i].ss_paddr));
4525 	case 2:
4526 		return (htobe64(segs[i + 1].ss_paddr));
4527 	}
4528 
4529 	return (0);
4530 }
4531 
4532 static void
4533 find_best_refill_source(struct adapter *sc, struct sge_fl *fl, int maxp)
4534 {
4535 	int8_t zidx, hwidx, idx;
4536 	uint16_t region1, region3;
4537 	int spare, spare_needed, n;
4538 	struct sw_zone_info *swz;
4539 	struct hw_buf_info *hwb, *hwb_list = &sc->sge.hw_buf_info[0];
4540 
4541 	/*
4542 	 * Buffer Packing: Look for PAGE_SIZE or larger zone which has a bufsize
4543 	 * large enough for the max payload and cluster metadata.  Otherwise
4544 	 * settle for the largest bufsize that leaves enough room in the cluster
4545 	 * for metadata.
4546 	 *
4547 	 * Without buffer packing: Look for the smallest zone which has a
4548 	 * bufsize large enough for the max payload.  Settle for the largest
4549 	 * bufsize available if there's nothing big enough for max payload.
4550 	 */
4551 	spare_needed = fl->flags & FL_BUF_PACKING ? CL_METADATA_SIZE : 0;
4552 	swz = &sc->sge.sw_zone_info[0];
4553 	hwidx = -1;
4554 	for (zidx = 0; zidx < SW_ZONE_SIZES; zidx++, swz++) {
4555 		if (swz->size > largest_rx_cluster) {
4556 			if (__predict_true(hwidx != -1))
4557 				break;
4558 
4559 			/*
4560 			 * This is a misconfiguration.  largest_rx_cluster is
4561 			 * preventing us from finding a refill source.  See
4562 			 * dev.t5nex.<n>.buffer_sizes to figure out why.
4563 			 */
4564 			device_printf(sc->dev, "largest_rx_cluster=%u leaves no"
4565 			    " refill source for fl %p (dma %u).  Ignored.\n",
4566 			    largest_rx_cluster, fl, maxp);
4567 		}
4568 		for (idx = swz->head_hwidx; idx != -1; idx = hwb->next) {
4569 			hwb = &hwb_list[idx];
4570 			spare = swz->size - hwb->size;
4571 			if (spare < spare_needed)
4572 				continue;
4573 
4574 			hwidx = idx;		/* best option so far */
4575 			if (hwb->size >= maxp) {
4576 
4577 				if ((fl->flags & FL_BUF_PACKING) == 0)
4578 					goto done; /* stop looking (not packing) */
4579 
4580 				if (swz->size >= safest_rx_cluster)
4581 					goto done; /* stop looking (packing) */
4582 			}
4583 			break;		/* keep looking, next zone */
4584 		}
4585 	}
4586 done:
4587 	/* A usable hwidx has been located. */
4588 	MPASS(hwidx != -1);
4589 	hwb = &hwb_list[hwidx];
4590 	zidx = hwb->zidx;
4591 	swz = &sc->sge.sw_zone_info[zidx];
4592 	region1 = 0;
4593 	region3 = swz->size - hwb->size;
4594 
4595 	/*
4596 	 * Stay within this zone and see if there is a better match when mbuf
4597 	 * inlining is allowed.  Remember that the hwidx's are sorted in
4598 	 * decreasing order of size (so in increasing order of spare area).
4599 	 */
4600 	for (idx = hwidx; idx != -1; idx = hwb->next) {
4601 		hwb = &hwb_list[idx];
4602 		spare = swz->size - hwb->size;
4603 
4604 		if (allow_mbufs_in_cluster == 0 || hwb->size < maxp)
4605 			break;
4606 
4607 		/*
4608 		 * Do not inline mbufs if doing so would violate the pad/pack
4609 		 * boundary alignment requirement.
4610 		 */
4611 		if (fl_pad && (MSIZE % sc->params.sge.pad_boundary) != 0)
4612 			continue;
4613 		if (fl->flags & FL_BUF_PACKING &&
4614 		    (MSIZE % sc->params.sge.pack_boundary) != 0)
4615 			continue;
4616 
4617 		if (spare < CL_METADATA_SIZE + MSIZE)
4618 			continue;
4619 		n = (spare - CL_METADATA_SIZE) / MSIZE;
4620 		if (n > howmany(hwb->size, maxp))
4621 			break;
4622 
4623 		hwidx = idx;
4624 		if (fl->flags & FL_BUF_PACKING) {
4625 			region1 = n * MSIZE;
4626 			region3 = spare - region1;
4627 		} else {
4628 			region1 = MSIZE;
4629 			region3 = spare - region1;
4630 			break;
4631 		}
4632 	}
4633 
4634 	KASSERT(zidx >= 0 && zidx < SW_ZONE_SIZES,
4635 	    ("%s: bad zone %d for fl %p, maxp %d", __func__, zidx, fl, maxp));
4636 	KASSERT(hwidx >= 0 && hwidx <= SGE_FLBUF_SIZES,
4637 	    ("%s: bad hwidx %d for fl %p, maxp %d", __func__, hwidx, fl, maxp));
4638 	KASSERT(region1 + sc->sge.hw_buf_info[hwidx].size + region3 ==
4639 	    sc->sge.sw_zone_info[zidx].size,
4640 	    ("%s: bad buffer layout for fl %p, maxp %d. "
4641 		"cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
4642 		sc->sge.sw_zone_info[zidx].size, region1,
4643 		sc->sge.hw_buf_info[hwidx].size, region3));
4644 	if (fl->flags & FL_BUF_PACKING || region1 > 0) {
4645 		KASSERT(region3 >= CL_METADATA_SIZE,
4646 		    ("%s: no room for metadata.  fl %p, maxp %d; "
4647 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
4648 		    sc->sge.sw_zone_info[zidx].size, region1,
4649 		    sc->sge.hw_buf_info[hwidx].size, region3));
4650 		KASSERT(region1 % MSIZE == 0,
4651 		    ("%s: bad mbuf region for fl %p, maxp %d. "
4652 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
4653 		    sc->sge.sw_zone_info[zidx].size, region1,
4654 		    sc->sge.hw_buf_info[hwidx].size, region3));
4655 	}
4656 
4657 	fl->cll_def.zidx = zidx;
4658 	fl->cll_def.hwidx = hwidx;
4659 	fl->cll_def.region1 = region1;
4660 	fl->cll_def.region3 = region3;
4661 }
4662 
4663 static void
4664 find_safe_refill_source(struct adapter *sc, struct sge_fl *fl)
4665 {
4666 	struct sge *s = &sc->sge;
4667 	struct hw_buf_info *hwb;
4668 	struct sw_zone_info *swz;
4669 	int spare;
4670 	int8_t hwidx;
4671 
4672 	if (fl->flags & FL_BUF_PACKING)
4673 		hwidx = s->safe_hwidx2;	/* with room for metadata */
4674 	else if (allow_mbufs_in_cluster && s->safe_hwidx2 != -1) {
4675 		hwidx = s->safe_hwidx2;
4676 		hwb = &s->hw_buf_info[hwidx];
4677 		swz = &s->sw_zone_info[hwb->zidx];
4678 		spare = swz->size - hwb->size;
4679 
4680 		/* no good if there isn't room for an mbuf as well */
4681 		if (spare < CL_METADATA_SIZE + MSIZE)
4682 			hwidx = s->safe_hwidx1;
4683 	} else
4684 		hwidx = s->safe_hwidx1;
4685 
4686 	if (hwidx == -1) {
4687 		/* No fallback source */
4688 		fl->cll_alt.hwidx = -1;
4689 		fl->cll_alt.zidx = -1;
4690 
4691 		return;
4692 	}
4693 
4694 	hwb = &s->hw_buf_info[hwidx];
4695 	swz = &s->sw_zone_info[hwb->zidx];
4696 	spare = swz->size - hwb->size;
4697 	fl->cll_alt.hwidx = hwidx;
4698 	fl->cll_alt.zidx = hwb->zidx;
4699 	if (allow_mbufs_in_cluster &&
4700 	    (fl_pad == 0 || (MSIZE % sc->params.sge.pad_boundary) == 0))
4701 		fl->cll_alt.region1 = ((spare - CL_METADATA_SIZE) / MSIZE) * MSIZE;
4702 	else
4703 		fl->cll_alt.region1 = 0;
4704 	fl->cll_alt.region3 = spare - fl->cll_alt.region1;
4705 }
4706 
4707 static void
4708 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
4709 {
4710 	mtx_lock(&sc->sfl_lock);
4711 	FL_LOCK(fl);
4712 	if ((fl->flags & FL_DOOMED) == 0) {
4713 		fl->flags |= FL_STARVING;
4714 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
4715 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
4716 	}
4717 	FL_UNLOCK(fl);
4718 	mtx_unlock(&sc->sfl_lock);
4719 }
4720 
4721 static void
4722 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq)
4723 {
4724 	struct sge_wrq *wrq = (void *)eq;
4725 
4726 	atomic_readandclear_int(&eq->equiq);
4727 	taskqueue_enqueue(sc->tq[eq->tx_chan], &wrq->wrq_tx_task);
4728 }
4729 
4730 static void
4731 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq)
4732 {
4733 	struct sge_txq *txq = (void *)eq;
4734 
4735 	MPASS((eq->flags & EQ_TYPEMASK) == EQ_ETH);
4736 
4737 	atomic_readandclear_int(&eq->equiq);
4738 	mp_ring_check_drainage(txq->r, 0);
4739 	taskqueue_enqueue(sc->tq[eq->tx_chan], &txq->tx_reclaim_task);
4740 }
4741 
4742 static int
4743 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
4744     struct mbuf *m)
4745 {
4746 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
4747 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
4748 	struct adapter *sc = iq->adapter;
4749 	struct sge *s = &sc->sge;
4750 	struct sge_eq *eq;
4751 	static void (*h[])(struct adapter *, struct sge_eq *) = {NULL,
4752 		&handle_wrq_egr_update, &handle_eth_egr_update,
4753 		&handle_wrq_egr_update};
4754 
4755 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
4756 	    rss->opcode));
4757 
4758 	eq = s->eqmap[qid - s->eq_start];
4759 	(*h[eq->flags & EQ_TYPEMASK])(sc, eq);
4760 
4761 	return (0);
4762 }
4763 
4764 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
4765 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
4766     offsetof(struct cpl_fw6_msg, data));
4767 
4768 static int
4769 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
4770 {
4771 	struct adapter *sc = iq->adapter;
4772 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
4773 
4774 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
4775 	    rss->opcode));
4776 
4777 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
4778 		const struct rss_header *rss2;
4779 
4780 		rss2 = (const struct rss_header *)&cpl->data[0];
4781 		return (t4_cpl_handler[rss2->opcode](iq, rss2, m));
4782 	}
4783 
4784 	return (t4_fw_msg_handler[cpl->type](sc, &cpl->data[0]));
4785 }
4786 
4787 /**
4788  *	t4_handle_wrerr_rpl - process a FW work request error message
4789  *	@adap: the adapter
4790  *	@rpl: start of the FW message
4791  */
4792 static int
4793 t4_handle_wrerr_rpl(struct adapter *adap, const __be64 *rpl)
4794 {
4795 	u8 opcode = *(const u8 *)rpl;
4796 	const struct fw_error_cmd *e = (const void *)rpl;
4797 	unsigned int i;
4798 
4799 	if (opcode != FW_ERROR_CMD) {
4800 		log(LOG_ERR,
4801 		    "%s: Received WRERR_RPL message with opcode %#x\n",
4802 		    device_get_nameunit(adap->dev), opcode);
4803 		return (EINVAL);
4804 	}
4805 	log(LOG_ERR, "%s: FW_ERROR (%s) ", device_get_nameunit(adap->dev),
4806 	    G_FW_ERROR_CMD_FATAL(be32toh(e->op_to_type)) ? "fatal" :
4807 	    "non-fatal");
4808 	switch (G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type))) {
4809 	case FW_ERROR_TYPE_EXCEPTION:
4810 		log(LOG_ERR, "exception info:\n");
4811 		for (i = 0; i < nitems(e->u.exception.info); i++)
4812 			log(LOG_ERR, "%s%08x", i == 0 ? "\t" : " ",
4813 			    be32toh(e->u.exception.info[i]));
4814 		log(LOG_ERR, "\n");
4815 		break;
4816 	case FW_ERROR_TYPE_HWMODULE:
4817 		log(LOG_ERR, "HW module regaddr %08x regval %08x\n",
4818 		    be32toh(e->u.hwmodule.regaddr),
4819 		    be32toh(e->u.hwmodule.regval));
4820 		break;
4821 	case FW_ERROR_TYPE_WR:
4822 		log(LOG_ERR, "WR cidx %d PF %d VF %d eqid %d hdr:\n",
4823 		    be16toh(e->u.wr.cidx),
4824 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.wr.pfn_vfn)),
4825 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.wr.pfn_vfn)),
4826 		    be32toh(e->u.wr.eqid));
4827 		for (i = 0; i < nitems(e->u.wr.wrhdr); i++)
4828 			log(LOG_ERR, "%s%02x", i == 0 ? "\t" : " ",
4829 			    e->u.wr.wrhdr[i]);
4830 		log(LOG_ERR, "\n");
4831 		break;
4832 	case FW_ERROR_TYPE_ACL:
4833 		log(LOG_ERR, "ACL cidx %d PF %d VF %d eqid %d %s",
4834 		    be16toh(e->u.acl.cidx),
4835 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.acl.pfn_vfn)),
4836 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.acl.pfn_vfn)),
4837 		    be32toh(e->u.acl.eqid),
4838 		    G_FW_ERROR_CMD_MV(be16toh(e->u.acl.mv_pkd)) ? "vlanid" :
4839 		    "MAC");
4840 		for (i = 0; i < nitems(e->u.acl.val); i++)
4841 			log(LOG_ERR, " %02x", e->u.acl.val[i]);
4842 		log(LOG_ERR, "\n");
4843 		break;
4844 	default:
4845 		log(LOG_ERR, "type %#x\n",
4846 		    G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type)));
4847 		return (EINVAL);
4848 	}
4849 	return (0);
4850 }
4851 
4852 static int
4853 sysctl_uint16(SYSCTL_HANDLER_ARGS)
4854 {
4855 	uint16_t *id = arg1;
4856 	int i = *id;
4857 
4858 	return sysctl_handle_int(oidp, &i, 0, req);
4859 }
4860 
4861 static int
4862 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
4863 {
4864 	struct sge *s = arg1;
4865 	struct hw_buf_info *hwb = &s->hw_buf_info[0];
4866 	struct sw_zone_info *swz = &s->sw_zone_info[0];
4867 	int i, rc;
4868 	struct sbuf sb;
4869 	char c;
4870 
4871 	sbuf_new(&sb, NULL, 32, SBUF_AUTOEXTEND);
4872 	for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
4873 		if (hwb->zidx >= 0 && swz[hwb->zidx].size <= largest_rx_cluster)
4874 			c = '*';
4875 		else
4876 			c = '\0';
4877 
4878 		sbuf_printf(&sb, "%u%c ", hwb->size, c);
4879 	}
4880 	sbuf_trim(&sb);
4881 	sbuf_finish(&sb);
4882 	rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
4883 	sbuf_delete(&sb);
4884 	return (rc);
4885 }
4886 
4887 static int
4888 sysctl_tc(SYSCTL_HANDLER_ARGS)
4889 {
4890 	struct vi_info *vi = arg1;
4891 	struct port_info *pi;
4892 	struct adapter *sc;
4893 	struct sge_txq *txq;
4894 	struct tx_sched_class *tc;
4895 	int qidx = arg2, rc, tc_idx;
4896 	uint32_t fw_queue, fw_class;
4897 
4898 	MPASS(qidx >= 0 && qidx < vi->ntxq);
4899 	pi = vi->pi;
4900 	sc = pi->adapter;
4901 	txq = &sc->sge.txq[vi->first_txq + qidx];
4902 
4903 	tc_idx = txq->tc_idx;
4904 	rc = sysctl_handle_int(oidp, &tc_idx, 0, req);
4905 	if (rc != 0 || req->newptr == NULL)
4906 		return (rc);
4907 
4908 	/* Note that -1 is legitimate input (it means unbind). */
4909 	if (tc_idx < -1 || tc_idx >= sc->chip_params->nsched_cls)
4910 		return (EINVAL);
4911 
4912 	rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4stc");
4913 	if (rc)
4914 		return (rc);
4915 
4916 	if (tc_idx == txq->tc_idx) {
4917 		rc = 0;		/* No change, nothing to do. */
4918 		goto done;
4919 	}
4920 
4921 	fw_queue = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
4922 	    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_EQ_SCHEDCLASS_ETH) |
4923 	    V_FW_PARAMS_PARAM_YZ(txq->eq.cntxt_id);
4924 
4925 	if (tc_idx == -1)
4926 		fw_class = 0xffffffff;	/* Unbind. */
4927 	else {
4928 		/*
4929 		 * Bind to a different class.  Ethernet txq's are only allowed
4930 		 * to bind to cl-rl mode-class for now.  XXX: too restrictive.
4931 		 */
4932 		tc = &pi->tc[tc_idx];
4933 		if (tc->flags & TX_SC_OK &&
4934 		    tc->params.level == SCHED_CLASS_LEVEL_CL_RL &&
4935 		    tc->params.mode == SCHED_CLASS_MODE_CLASS) {
4936 			/* Ok to proceed. */
4937 			fw_class = tc_idx;
4938 		} else {
4939 			rc = tc->flags & TX_SC_OK ? EBUSY : ENXIO;
4940 			goto done;
4941 		}
4942 	}
4943 
4944 	rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &fw_queue, &fw_class);
4945 	if (rc == 0) {
4946 		if (txq->tc_idx != -1) {
4947 			tc = &pi->tc[txq->tc_idx];
4948 			MPASS(tc->refcount > 0);
4949 			tc->refcount--;
4950 		}
4951 		if (tc_idx != -1) {
4952 			tc = &pi->tc[tc_idx];
4953 			tc->refcount++;
4954 		}
4955 		txq->tc_idx = tc_idx;
4956 	}
4957 done:
4958 	end_synchronized_op(sc, 0);
4959 	return (rc);
4960 }
4961