xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision 7be8de4271d5cb5d441e2757912c1824f6c3dc3b)
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 static 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 static 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" for TDDP */
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 	/*
583 	 * 4K, 8K, 16K, 64K DDP "page sizes" for iSCSI DDP.  These have been
584 	 * chosen with MAXPHYS = 128K in mind.  The largest DDP buffer that we
585 	 * may have to deal with is MAXPHYS + 1 page.
586 	 */
587 	v = V_HPZ0(0) | V_HPZ1(1) | V_HPZ2(2) | V_HPZ3(4);
588 	t4_write_reg(sc, A_ULP_RX_ISCSI_PSZ, v);
589 
590 	/* We use multiple DDP page sizes both in plain-TOE and ISCSI modes. */
591 	m = v = F_TDDPTAGTCB | F_ISCSITAGTCB;
592 	t4_set_reg_field(sc, A_ULP_RX_CTL, m, v);
593 
594 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
595 	    F_RESETDDPOFFSET;
596 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
597 	t4_set_reg_field(sc, A_TP_PARA_REG5, m, v);
598 }
599 
600 /*
601  * SGE wants the buffer to be at least 64B and then a multiple of 16.  If
602  * padding is in use, the buffer's start and end need to be aligned to the pad
603  * boundary as well.  We'll just make sure that the size is a multiple of the
604  * boundary here, it is up to the buffer allocation code to make sure the start
605  * of the buffer is aligned as well.
606  */
607 static inline int
608 hwsz_ok(struct adapter *sc, int hwsz)
609 {
610 	int mask = fl_pad ? sc->params.sge.pad_boundary - 1 : 16 - 1;
611 
612 	return (hwsz >= 64 && (hwsz & mask) == 0);
613 }
614 
615 /*
616  * XXX: driver really should be able to deal with unexpected settings.
617  */
618 int
619 t4_read_chip_settings(struct adapter *sc)
620 {
621 	struct sge *s = &sc->sge;
622 	struct sge_params *sp = &sc->params.sge;
623 	int i, j, n, rc = 0;
624 	uint32_t m, v, r;
625 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
626 	static int sw_buf_sizes[] = {	/* Sorted by size */
627 		MCLBYTES,
628 #if MJUMPAGESIZE != MCLBYTES
629 		MJUMPAGESIZE,
630 #endif
631 		MJUM9BYTES,
632 		MJUM16BYTES
633 	};
634 	struct sw_zone_info *swz, *safe_swz;
635 	struct hw_buf_info *hwb;
636 
637 	m = F_RXPKTCPLMODE;
638 	v = F_RXPKTCPLMODE;
639 	r = sc->params.sge.sge_control;
640 	if ((r & m) != v) {
641 		device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r);
642 		rc = EINVAL;
643 	}
644 
645 	/*
646 	 * If this changes then every single use of PAGE_SHIFT in the driver
647 	 * needs to be carefully reviewed for PAGE_SHIFT vs sp->page_shift.
648 	 */
649 	if (sp->page_shift != PAGE_SHIFT) {
650 		device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r);
651 		rc = EINVAL;
652 	}
653 
654 	/* Filter out unusable hw buffer sizes entirely (mark with -2). */
655 	hwb = &s->hw_buf_info[0];
656 	for (i = 0; i < nitems(s->hw_buf_info); i++, hwb++) {
657 		r = sc->params.sge.sge_fl_buffer_size[i];
658 		hwb->size = r;
659 		hwb->zidx = hwsz_ok(sc, r) ? -1 : -2;
660 		hwb->next = -1;
661 	}
662 
663 	/*
664 	 * Create a sorted list in decreasing order of hw buffer sizes (and so
665 	 * increasing order of spare area) for each software zone.
666 	 *
667 	 * If padding is enabled then the start and end of the buffer must align
668 	 * to the pad boundary; if packing is enabled then they must align with
669 	 * the pack boundary as well.  Allocations from the cluster zones are
670 	 * aligned to min(size, 4K), so the buffer starts at that alignment and
671 	 * ends at hwb->size alignment.  If mbuf inlining is allowed the
672 	 * starting alignment will be reduced to MSIZE and the driver will
673 	 * exercise appropriate caution when deciding on the best buffer layout
674 	 * to use.
675 	 */
676 	n = 0;	/* no usable buffer size to begin with */
677 	swz = &s->sw_zone_info[0];
678 	safe_swz = NULL;
679 	for (i = 0; i < SW_ZONE_SIZES; i++, swz++) {
680 		int8_t head = -1, tail = -1;
681 
682 		swz->size = sw_buf_sizes[i];
683 		swz->zone = m_getzone(swz->size);
684 		swz->type = m_gettype(swz->size);
685 
686 		if (swz->size < PAGE_SIZE) {
687 			MPASS(powerof2(swz->size));
688 			if (fl_pad && (swz->size % sp->pad_boundary != 0))
689 				continue;
690 		}
691 
692 		if (swz->size == safest_rx_cluster)
693 			safe_swz = swz;
694 
695 		hwb = &s->hw_buf_info[0];
696 		for (j = 0; j < SGE_FLBUF_SIZES; j++, hwb++) {
697 			if (hwb->zidx != -1 || hwb->size > swz->size)
698 				continue;
699 #ifdef INVARIANTS
700 			if (fl_pad)
701 				MPASS(hwb->size % sp->pad_boundary == 0);
702 #endif
703 			hwb->zidx = i;
704 			if (head == -1)
705 				head = tail = j;
706 			else if (hwb->size < s->hw_buf_info[tail].size) {
707 				s->hw_buf_info[tail].next = j;
708 				tail = j;
709 			} else {
710 				int8_t *cur;
711 				struct hw_buf_info *t;
712 
713 				for (cur = &head; *cur != -1; cur = &t->next) {
714 					t = &s->hw_buf_info[*cur];
715 					if (hwb->size == t->size) {
716 						hwb->zidx = -2;
717 						break;
718 					}
719 					if (hwb->size > t->size) {
720 						hwb->next = *cur;
721 						*cur = j;
722 						break;
723 					}
724 				}
725 			}
726 		}
727 		swz->head_hwidx = head;
728 		swz->tail_hwidx = tail;
729 
730 		if (tail != -1) {
731 			n++;
732 			if (swz->size - s->hw_buf_info[tail].size >=
733 			    CL_METADATA_SIZE)
734 				sc->flags |= BUF_PACKING_OK;
735 		}
736 	}
737 	if (n == 0) {
738 		device_printf(sc->dev, "no usable SGE FL buffer size.\n");
739 		rc = EINVAL;
740 	}
741 
742 	s->safe_hwidx1 = -1;
743 	s->safe_hwidx2 = -1;
744 	if (safe_swz != NULL) {
745 		s->safe_hwidx1 = safe_swz->head_hwidx;
746 		for (i = safe_swz->head_hwidx; i != -1; i = hwb->next) {
747 			int spare;
748 
749 			hwb = &s->hw_buf_info[i];
750 #ifdef INVARIANTS
751 			if (fl_pad)
752 				MPASS(hwb->size % sp->pad_boundary == 0);
753 #endif
754 			spare = safe_swz->size - hwb->size;
755 			if (spare >= CL_METADATA_SIZE) {
756 				s->safe_hwidx2 = i;
757 				break;
758 			}
759 		}
760 	}
761 
762 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
763 	r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ);
764 	if (r != v) {
765 		device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r);
766 		rc = EINVAL;
767 	}
768 
769 	m = v = F_TDDPTAGTCB;
770 	r = t4_read_reg(sc, A_ULP_RX_CTL);
771 	if ((r & m) != v) {
772 		device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r);
773 		rc = EINVAL;
774 	}
775 
776 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
777 	    F_RESETDDPOFFSET;
778 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
779 	r = t4_read_reg(sc, A_TP_PARA_REG5);
780 	if ((r & m) != v) {
781 		device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r);
782 		rc = EINVAL;
783 	}
784 
785 	t4_init_tp_params(sc);
786 
787 	t4_read_mtu_tbl(sc, sc->params.mtus, NULL);
788 	t4_load_mtus(sc, sc->params.mtus, sc->params.a_wnd, sc->params.b_wnd);
789 
790 	return (rc);
791 }
792 
793 int
794 t4_create_dma_tag(struct adapter *sc)
795 {
796 	int rc;
797 
798 	rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
799 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
800 	    BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
801 	    NULL, &sc->dmat);
802 	if (rc != 0) {
803 		device_printf(sc->dev,
804 		    "failed to create main DMA tag: %d\n", rc);
805 	}
806 
807 	return (rc);
808 }
809 
810 void
811 t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
812     struct sysctl_oid_list *children)
813 {
814 	struct sge_params *sp = &sc->params.sge;
815 
816 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "buffer_sizes",
817 	    CTLTYPE_STRING | CTLFLAG_RD, &sc->sge, 0, sysctl_bufsizes, "A",
818 	    "freelist buffer sizes");
819 
820 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD,
821 	    NULL, sp->fl_pktshift, "payload DMA offset in rx buffer (bytes)");
822 
823 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD,
824 	    NULL, sp->pad_boundary, "payload pad boundary (bytes)");
825 
826 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD,
827 	    NULL, sp->spg_len, "status page size (bytes)");
828 
829 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD,
830 	    NULL, cong_drop, "congestion drop setting");
831 
832 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD,
833 	    NULL, sp->pack_boundary, "payload pack boundary (bytes)");
834 }
835 
836 int
837 t4_destroy_dma_tag(struct adapter *sc)
838 {
839 	if (sc->dmat)
840 		bus_dma_tag_destroy(sc->dmat);
841 
842 	return (0);
843 }
844 
845 /*
846  * Allocate and initialize the firmware event queue and the management queue.
847  *
848  * Returns errno on failure.  Resources allocated up to that point may still be
849  * allocated.  Caller is responsible for cleanup in case this function fails.
850  */
851 int
852 t4_setup_adapter_queues(struct adapter *sc)
853 {
854 	int rc;
855 
856 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
857 
858 	sysctl_ctx_init(&sc->ctx);
859 	sc->flags |= ADAP_SYSCTL_CTX;
860 
861 	/*
862 	 * Firmware event queue
863 	 */
864 	rc = alloc_fwq(sc);
865 	if (rc != 0)
866 		return (rc);
867 
868 	/*
869 	 * Management queue.  This is just a control queue that uses the fwq as
870 	 * its associated iq.
871 	 */
872 	rc = alloc_mgmtq(sc);
873 
874 	return (rc);
875 }
876 
877 /*
878  * Idempotent
879  */
880 int
881 t4_teardown_adapter_queues(struct adapter *sc)
882 {
883 
884 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
885 
886 	/* Do this before freeing the queue */
887 	if (sc->flags & ADAP_SYSCTL_CTX) {
888 		sysctl_ctx_free(&sc->ctx);
889 		sc->flags &= ~ADAP_SYSCTL_CTX;
890 	}
891 
892 	free_mgmtq(sc);
893 	free_fwq(sc);
894 
895 	return (0);
896 }
897 
898 static inline int
899 first_vector(struct vi_info *vi)
900 {
901 	struct adapter *sc = vi->pi->adapter;
902 
903 	if (sc->intr_count == 1)
904 		return (0);
905 
906 	return (vi->first_intr);
907 }
908 
909 /*
910  * Given an arbitrary "index," come up with an iq that can be used by other
911  * queues (of this VI) for interrupt forwarding, SGE egress updates, etc.
912  * The iq returned is guaranteed to be something that takes direct interrupts.
913  */
914 static struct sge_iq *
915 vi_intr_iq(struct vi_info *vi, int idx)
916 {
917 	struct adapter *sc = vi->pi->adapter;
918 	struct sge *s = &sc->sge;
919 	struct sge_iq *iq = NULL;
920 	int nintr, i;
921 
922 	if (sc->intr_count == 1)
923 		return (&sc->sge.fwq);
924 
925 	nintr = vi->nintr;
926 	KASSERT(nintr != 0,
927 	    ("%s: vi %p has no exclusive interrupts, total interrupts = %d",
928 	    __func__, vi, sc->intr_count));
929 	i = idx % nintr;
930 
931 	if (vi->flags & INTR_RXQ) {
932 	       	if (i < vi->nrxq) {
933 			iq = &s->rxq[vi->first_rxq + i].iq;
934 			goto done;
935 		}
936 		i -= vi->nrxq;
937 	}
938 #ifdef TCP_OFFLOAD
939 	if (vi->flags & INTR_OFLD_RXQ) {
940 	       	if (i < vi->nofldrxq) {
941 			iq = &s->ofld_rxq[vi->first_ofld_rxq + i].iq;
942 			goto done;
943 		}
944 		i -= vi->nofldrxq;
945 	}
946 #endif
947 	panic("%s: vi %p, intr_flags 0x%lx, idx %d, total intr %d\n", __func__,
948 	    vi, vi->flags & INTR_ALL, idx, nintr);
949 done:
950 	MPASS(iq != NULL);
951 	KASSERT(iq->flags & IQ_INTR,
952 	    ("%s: iq %p (vi %p, intr_flags 0x%lx, idx %d)", __func__, iq, vi,
953 	    vi->flags & INTR_ALL, idx));
954 	return (iq);
955 }
956 
957 /* Maximum payload that can be delivered with a single iq descriptor */
958 static inline int
959 mtu_to_max_payload(struct adapter *sc, int mtu, const int toe)
960 {
961 	int payload;
962 
963 #ifdef TCP_OFFLOAD
964 	if (toe) {
965 		payload = sc->tt.rx_coalesce ?
966 		    G_RXCOALESCESIZE(t4_read_reg(sc, A_TP_PARA_REG2)) : mtu;
967 	} else {
968 #endif
969 		/* large enough even when hw VLAN extraction is disabled */
970 		payload = sc->params.sge.fl_pktshift + ETHER_HDR_LEN +
971 		    ETHER_VLAN_ENCAP_LEN + mtu;
972 #ifdef TCP_OFFLOAD
973 	}
974 #endif
975 
976 	return (payload);
977 }
978 
979 int
980 t4_setup_vi_queues(struct vi_info *vi)
981 {
982 	int rc = 0, i, j, intr_idx, iqid;
983 	struct sge_rxq *rxq;
984 	struct sge_txq *txq;
985 	struct sge_wrq *ctrlq;
986 #ifdef TCP_OFFLOAD
987 	struct sge_ofld_rxq *ofld_rxq;
988 	struct sge_wrq *ofld_txq;
989 #endif
990 #ifdef DEV_NETMAP
991 	int saved_idx;
992 	struct sge_nm_rxq *nm_rxq;
993 	struct sge_nm_txq *nm_txq;
994 #endif
995 	char name[16];
996 	struct port_info *pi = vi->pi;
997 	struct adapter *sc = pi->adapter;
998 	struct ifnet *ifp = vi->ifp;
999 	struct sysctl_oid *oid = device_get_sysctl_tree(vi->dev);
1000 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
1001 	int maxp, mtu = ifp->if_mtu;
1002 
1003 	/* Interrupt vector to start from (when using multiple vectors) */
1004 	intr_idx = first_vector(vi);
1005 
1006 #ifdef DEV_NETMAP
1007 	saved_idx = intr_idx;
1008 	if (ifp->if_capabilities & IFCAP_NETMAP) {
1009 
1010 		/* netmap is supported with direct interrupts only. */
1011 		MPASS(vi->flags & INTR_RXQ);
1012 
1013 		/*
1014 		 * We don't have buffers to back the netmap rx queues
1015 		 * right now so we create the queues in a way that
1016 		 * doesn't set off any congestion signal in the chip.
1017 		 */
1018 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_rxq",
1019 		    CTLFLAG_RD, NULL, "rx queues");
1020 		for_each_nm_rxq(vi, i, nm_rxq) {
1021 			rc = alloc_nm_rxq(vi, nm_rxq, intr_idx, i, oid);
1022 			if (rc != 0)
1023 				goto done;
1024 			intr_idx++;
1025 		}
1026 
1027 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_txq",
1028 		    CTLFLAG_RD, NULL, "tx queues");
1029 		for_each_nm_txq(vi, i, nm_txq) {
1030 			iqid = vi->first_nm_rxq + (i % vi->nnmrxq);
1031 			rc = alloc_nm_txq(vi, nm_txq, iqid, i, oid);
1032 			if (rc != 0)
1033 				goto done;
1034 		}
1035 	}
1036 
1037 	/* Normal rx queues and netmap rx queues share the same interrupts. */
1038 	intr_idx = saved_idx;
1039 #endif
1040 
1041 	/*
1042 	 * First pass over all NIC and TOE rx queues:
1043 	 * a) initialize iq and fl
1044 	 * b) allocate queue iff it will take direct interrupts.
1045 	 */
1046 	maxp = mtu_to_max_payload(sc, mtu, 0);
1047 	if (vi->flags & INTR_RXQ) {
1048 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
1049 		    CTLFLAG_RD, NULL, "rx queues");
1050 	}
1051 	for_each_rxq(vi, i, rxq) {
1052 
1053 		init_iq(&rxq->iq, sc, vi->tmr_idx, vi->pktc_idx, vi->qsize_rxq);
1054 
1055 		snprintf(name, sizeof(name), "%s rxq%d-fl",
1056 		    device_get_nameunit(vi->dev), i);
1057 		init_fl(sc, &rxq->fl, vi->qsize_rxq / 8, maxp, name);
1058 
1059 		if (vi->flags & INTR_RXQ) {
1060 			rxq->iq.flags |= IQ_INTR;
1061 			rc = alloc_rxq(vi, rxq, intr_idx, i, oid);
1062 			if (rc != 0)
1063 				goto done;
1064 			intr_idx++;
1065 		}
1066 	}
1067 #ifdef DEV_NETMAP
1068 	if (ifp->if_capabilities & IFCAP_NETMAP)
1069 		intr_idx = saved_idx + max(vi->nrxq, vi->nnmrxq);
1070 #endif
1071 #ifdef TCP_OFFLOAD
1072 	maxp = mtu_to_max_payload(sc, mtu, 1);
1073 	if (vi->flags & INTR_OFLD_RXQ) {
1074 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq",
1075 		    CTLFLAG_RD, NULL,
1076 		    "rx queues for offloaded TCP connections");
1077 	}
1078 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1079 
1080 		init_iq(&ofld_rxq->iq, sc, vi->tmr_idx, vi->pktc_idx,
1081 		    vi->qsize_rxq);
1082 
1083 		snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
1084 		    device_get_nameunit(vi->dev), i);
1085 		init_fl(sc, &ofld_rxq->fl, vi->qsize_rxq / 8, maxp, name);
1086 
1087 		if (vi->flags & INTR_OFLD_RXQ) {
1088 			ofld_rxq->iq.flags |= IQ_INTR;
1089 			rc = alloc_ofld_rxq(vi, ofld_rxq, intr_idx, i, oid);
1090 			if (rc != 0)
1091 				goto done;
1092 			intr_idx++;
1093 		}
1094 	}
1095 #endif
1096 
1097 	/*
1098 	 * Second pass over all NIC and TOE rx queues.  The queues forwarding
1099 	 * their interrupts are allocated now.
1100 	 */
1101 	j = 0;
1102 	if (!(vi->flags & INTR_RXQ)) {
1103 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
1104 		    CTLFLAG_RD, NULL, "rx queues");
1105 		for_each_rxq(vi, i, rxq) {
1106 			MPASS(!(rxq->iq.flags & IQ_INTR));
1107 
1108 			intr_idx = vi_intr_iq(vi, j)->abs_id;
1109 
1110 			rc = alloc_rxq(vi, rxq, intr_idx, i, oid);
1111 			if (rc != 0)
1112 				goto done;
1113 			j++;
1114 		}
1115 	}
1116 #ifdef TCP_OFFLOAD
1117 	if (vi->nofldrxq != 0 && !(vi->flags & INTR_OFLD_RXQ)) {
1118 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq",
1119 		    CTLFLAG_RD, NULL,
1120 		    "rx queues for offloaded TCP connections");
1121 		for_each_ofld_rxq(vi, i, ofld_rxq) {
1122 			MPASS(!(ofld_rxq->iq.flags & IQ_INTR));
1123 
1124 			intr_idx = vi_intr_iq(vi, j)->abs_id;
1125 
1126 			rc = alloc_ofld_rxq(vi, ofld_rxq, intr_idx, i, oid);
1127 			if (rc != 0)
1128 				goto done;
1129 			j++;
1130 		}
1131 	}
1132 #endif
1133 
1134 	/*
1135 	 * Now the tx queues.  Only one pass needed.
1136 	 */
1137 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "txq", CTLFLAG_RD,
1138 	    NULL, "tx queues");
1139 	j = 0;
1140 	for_each_txq(vi, i, txq) {
1141 		iqid = vi_intr_iq(vi, j)->cntxt_id;
1142 		snprintf(name, sizeof(name), "%s txq%d",
1143 		    device_get_nameunit(vi->dev), i);
1144 		init_eq(sc, &txq->eq, EQ_ETH, vi->qsize_txq, pi->tx_chan, iqid,
1145 		    name);
1146 
1147 		rc = alloc_txq(vi, txq, i, oid);
1148 		if (rc != 0)
1149 			goto done;
1150 		j++;
1151 	}
1152 #ifdef TCP_OFFLOAD
1153 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_txq",
1154 	    CTLFLAG_RD, NULL, "tx queues for offloaded TCP connections");
1155 	for_each_ofld_txq(vi, i, ofld_txq) {
1156 		struct sysctl_oid *oid2;
1157 
1158 		iqid = vi_intr_iq(vi, j)->cntxt_id;
1159 		snprintf(name, sizeof(name), "%s ofld_txq%d",
1160 		    device_get_nameunit(vi->dev), i);
1161 		init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq, pi->tx_chan,
1162 		    iqid, name);
1163 
1164 		snprintf(name, sizeof(name), "%d", i);
1165 		oid2 = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
1166 		    name, CTLFLAG_RD, NULL, "offload tx queue");
1167 
1168 		rc = alloc_wrq(sc, vi, ofld_txq, oid2);
1169 		if (rc != 0)
1170 			goto done;
1171 		j++;
1172 	}
1173 #endif
1174 
1175 	/*
1176 	 * Finally, the control queue.
1177 	 */
1178 	if (!IS_MAIN_VI(vi))
1179 		goto done;
1180 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ctrlq", CTLFLAG_RD,
1181 	    NULL, "ctrl queue");
1182 	ctrlq = &sc->sge.ctrlq[pi->port_id];
1183 	iqid = vi_intr_iq(vi, 0)->cntxt_id;
1184 	snprintf(name, sizeof(name), "%s ctrlq", device_get_nameunit(vi->dev));
1185 	init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, pi->tx_chan, iqid,
1186 	    name);
1187 	rc = alloc_wrq(sc, vi, ctrlq, oid);
1188 
1189 done:
1190 	if (rc)
1191 		t4_teardown_vi_queues(vi);
1192 
1193 	return (rc);
1194 }
1195 
1196 /*
1197  * Idempotent
1198  */
1199 int
1200 t4_teardown_vi_queues(struct vi_info *vi)
1201 {
1202 	int i;
1203 	struct port_info *pi = vi->pi;
1204 	struct adapter *sc = pi->adapter;
1205 	struct sge_rxq *rxq;
1206 	struct sge_txq *txq;
1207 #ifdef TCP_OFFLOAD
1208 	struct sge_ofld_rxq *ofld_rxq;
1209 	struct sge_wrq *ofld_txq;
1210 #endif
1211 #ifdef DEV_NETMAP
1212 	struct sge_nm_rxq *nm_rxq;
1213 	struct sge_nm_txq *nm_txq;
1214 #endif
1215 
1216 	/* Do this before freeing the queues */
1217 	if (vi->flags & VI_SYSCTL_CTX) {
1218 		sysctl_ctx_free(&vi->ctx);
1219 		vi->flags &= ~VI_SYSCTL_CTX;
1220 	}
1221 
1222 #ifdef DEV_NETMAP
1223 	if (vi->ifp->if_capabilities & IFCAP_NETMAP) {
1224 		for_each_nm_txq(vi, i, nm_txq) {
1225 			free_nm_txq(vi, nm_txq);
1226 		}
1227 
1228 		for_each_nm_rxq(vi, i, nm_rxq) {
1229 			free_nm_rxq(vi, nm_rxq);
1230 		}
1231 	}
1232 #endif
1233 
1234 	/*
1235 	 * Take down all the tx queues first, as they reference the rx queues
1236 	 * (for egress updates, etc.).
1237 	 */
1238 
1239 	if (IS_MAIN_VI(vi))
1240 		free_wrq(sc, &sc->sge.ctrlq[pi->port_id]);
1241 
1242 	for_each_txq(vi, i, txq) {
1243 		free_txq(vi, txq);
1244 	}
1245 #ifdef TCP_OFFLOAD
1246 	for_each_ofld_txq(vi, i, ofld_txq) {
1247 		free_wrq(sc, ofld_txq);
1248 	}
1249 #endif
1250 
1251 	/*
1252 	 * Then take down the rx queues that forward their interrupts, as they
1253 	 * reference other rx queues.
1254 	 */
1255 
1256 	for_each_rxq(vi, i, rxq) {
1257 		if ((rxq->iq.flags & IQ_INTR) == 0)
1258 			free_rxq(vi, rxq);
1259 	}
1260 #ifdef TCP_OFFLOAD
1261 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1262 		if ((ofld_rxq->iq.flags & IQ_INTR) == 0)
1263 			free_ofld_rxq(vi, ofld_rxq);
1264 	}
1265 #endif
1266 
1267 	/*
1268 	 * Then take down the rx queues that take direct interrupts.
1269 	 */
1270 
1271 	for_each_rxq(vi, i, rxq) {
1272 		if (rxq->iq.flags & IQ_INTR)
1273 			free_rxq(vi, rxq);
1274 	}
1275 #ifdef TCP_OFFLOAD
1276 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1277 		if (ofld_rxq->iq.flags & IQ_INTR)
1278 			free_ofld_rxq(vi, ofld_rxq);
1279 	}
1280 #endif
1281 
1282 	return (0);
1283 }
1284 
1285 /*
1286  * Deals with errors and the firmware event queue.  All data rx queues forward
1287  * their interrupt to the firmware event queue.
1288  */
1289 void
1290 t4_intr_all(void *arg)
1291 {
1292 	struct adapter *sc = arg;
1293 	struct sge_iq *fwq = &sc->sge.fwq;
1294 
1295 	t4_intr_err(arg);
1296 	if (atomic_cmpset_int(&fwq->state, IQS_IDLE, IQS_BUSY)) {
1297 		service_iq(fwq, 0);
1298 		atomic_cmpset_int(&fwq->state, IQS_BUSY, IQS_IDLE);
1299 	}
1300 }
1301 
1302 /* Deals with error interrupts */
1303 void
1304 t4_intr_err(void *arg)
1305 {
1306 	struct adapter *sc = arg;
1307 
1308 	t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
1309 	t4_slow_intr_handler(sc);
1310 }
1311 
1312 void
1313 t4_intr_evt(void *arg)
1314 {
1315 	struct sge_iq *iq = arg;
1316 
1317 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1318 		service_iq(iq, 0);
1319 		atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1320 	}
1321 }
1322 
1323 void
1324 t4_intr(void *arg)
1325 {
1326 	struct sge_iq *iq = arg;
1327 
1328 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1329 		service_iq(iq, 0);
1330 		atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1331 	}
1332 }
1333 
1334 void
1335 t4_vi_intr(void *arg)
1336 {
1337 	struct irq *irq = arg;
1338 
1339 #ifdef DEV_NETMAP
1340 	if (atomic_cmpset_int(&irq->nm_state, NM_ON, NM_BUSY)) {
1341 		t4_nm_intr(irq->nm_rxq);
1342 		atomic_cmpset_int(&irq->nm_state, NM_BUSY, NM_ON);
1343 	}
1344 #endif
1345 	if (irq->rxq != NULL)
1346 		t4_intr(irq->rxq);
1347 }
1348 
1349 /*
1350  * Deals with anything and everything on the given ingress queue.
1351  */
1352 static int
1353 service_iq(struct sge_iq *iq, int budget)
1354 {
1355 	struct sge_iq *q;
1356 	struct sge_rxq *rxq = iq_to_rxq(iq);	/* Use iff iq is part of rxq */
1357 	struct sge_fl *fl;			/* Use iff IQ_HAS_FL */
1358 	struct adapter *sc = iq->adapter;
1359 	struct iq_desc *d = &iq->desc[iq->cidx];
1360 	int ndescs = 0, limit;
1361 	int rsp_type, refill;
1362 	uint32_t lq;
1363 	uint16_t fl_hw_cidx;
1364 	struct mbuf *m0;
1365 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1366 #if defined(INET) || defined(INET6)
1367 	const struct timeval lro_timeout = {0, sc->lro_timeout};
1368 #endif
1369 
1370 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1371 
1372 	limit = budget ? budget : iq->qsize / 16;
1373 
1374 	if (iq->flags & IQ_HAS_FL) {
1375 		fl = &rxq->fl;
1376 		fl_hw_cidx = fl->hw_cidx;	/* stable snapshot */
1377 	} else {
1378 		fl = NULL;
1379 		fl_hw_cidx = 0;			/* to silence gcc warning */
1380 	}
1381 
1382 	/*
1383 	 * We always come back and check the descriptor ring for new indirect
1384 	 * interrupts and other responses after running a single handler.
1385 	 */
1386 	for (;;) {
1387 		while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1388 
1389 			rmb();
1390 
1391 			refill = 0;
1392 			m0 = NULL;
1393 			rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1394 			lq = be32toh(d->rsp.pldbuflen_qid);
1395 
1396 			switch (rsp_type) {
1397 			case X_RSPD_TYPE_FLBUF:
1398 
1399 				KASSERT(iq->flags & IQ_HAS_FL,
1400 				    ("%s: data for an iq (%p) with no freelist",
1401 				    __func__, iq));
1402 
1403 				m0 = get_fl_payload(sc, fl, lq);
1404 				if (__predict_false(m0 == NULL))
1405 					goto process_iql;
1406 				refill = IDXDIFF(fl->hw_cidx, fl_hw_cidx, fl->sidx) > 2;
1407 #ifdef T4_PKT_TIMESTAMP
1408 				/*
1409 				 * 60 bit timestamp for the payload is
1410 				 * *(uint64_t *)m0->m_pktdat.  Note that it is
1411 				 * in the leading free-space in the mbuf.  The
1412 				 * kernel can clobber it during a pullup,
1413 				 * m_copymdata, etc.  You need to make sure that
1414 				 * the mbuf reaches you unmolested if you care
1415 				 * about the timestamp.
1416 				 */
1417 				*(uint64_t *)m0->m_pktdat =
1418 				    be64toh(ctrl->u.last_flit) &
1419 				    0xfffffffffffffff;
1420 #endif
1421 
1422 				/* fall through */
1423 
1424 			case X_RSPD_TYPE_CPL:
1425 				KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1426 				    ("%s: bad opcode %02x.", __func__,
1427 				    d->rss.opcode));
1428 				t4_cpl_handler[d->rss.opcode](iq, &d->rss, m0);
1429 				break;
1430 
1431 			case X_RSPD_TYPE_INTR:
1432 
1433 				/*
1434 				 * Interrupts should be forwarded only to queues
1435 				 * that are not forwarding their interrupts.
1436 				 * This means service_iq can recurse but only 1
1437 				 * level deep.
1438 				 */
1439 				KASSERT(budget == 0,
1440 				    ("%s: budget %u, rsp_type %u", __func__,
1441 				    budget, rsp_type));
1442 
1443 				/*
1444 				 * There are 1K interrupt-capable queues (qids 0
1445 				 * through 1023).  A response type indicating a
1446 				 * forwarded interrupt with a qid >= 1K is an
1447 				 * iWARP async notification.
1448 				 */
1449 				if (lq >= 1024) {
1450                                         t4_an_handler(iq, &d->rsp);
1451                                         break;
1452                                 }
1453 
1454 				q = sc->sge.iqmap[lq - sc->sge.iq_start -
1455 				    sc->sge.iq_base];
1456 				if (atomic_cmpset_int(&q->state, IQS_IDLE,
1457 				    IQS_BUSY)) {
1458 					if (service_iq(q, q->qsize / 16) == 0) {
1459 						atomic_cmpset_int(&q->state,
1460 						    IQS_BUSY, IQS_IDLE);
1461 					} else {
1462 						STAILQ_INSERT_TAIL(&iql, q,
1463 						    link);
1464 					}
1465 				}
1466 				break;
1467 
1468 			default:
1469 				KASSERT(0,
1470 				    ("%s: illegal response type %d on iq %p",
1471 				    __func__, rsp_type, iq));
1472 				log(LOG_ERR,
1473 				    "%s: illegal response type %d on iq %p",
1474 				    device_get_nameunit(sc->dev), rsp_type, iq);
1475 				break;
1476 			}
1477 
1478 			d++;
1479 			if (__predict_false(++iq->cidx == iq->sidx)) {
1480 				iq->cidx = 0;
1481 				iq->gen ^= F_RSPD_GEN;
1482 				d = &iq->desc[0];
1483 			}
1484 			if (__predict_false(++ndescs == limit)) {
1485 				t4_write_reg(sc, sc->sge_gts_reg,
1486 				    V_CIDXINC(ndescs) |
1487 				    V_INGRESSQID(iq->cntxt_id) |
1488 				    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1489 				ndescs = 0;
1490 
1491 #if defined(INET) || defined(INET6)
1492 				if (iq->flags & IQ_LRO_ENABLED &&
1493 				    sc->lro_timeout != 0) {
1494 					tcp_lro_flush_inactive(&rxq->lro,
1495 					    &lro_timeout);
1496 				}
1497 #endif
1498 
1499 				if (budget) {
1500 					if (iq->flags & IQ_HAS_FL) {
1501 						FL_LOCK(fl);
1502 						refill_fl(sc, fl, 32);
1503 						FL_UNLOCK(fl);
1504 					}
1505 					return (EINPROGRESS);
1506 				}
1507 			}
1508 			if (refill) {
1509 				FL_LOCK(fl);
1510 				refill_fl(sc, fl, 32);
1511 				FL_UNLOCK(fl);
1512 				fl_hw_cidx = fl->hw_cidx;
1513 			}
1514 		}
1515 
1516 process_iql:
1517 		if (STAILQ_EMPTY(&iql))
1518 			break;
1519 
1520 		/*
1521 		 * Process the head only, and send it to the back of the list if
1522 		 * it's still not done.
1523 		 */
1524 		q = STAILQ_FIRST(&iql);
1525 		STAILQ_REMOVE_HEAD(&iql, link);
1526 		if (service_iq(q, q->qsize / 8) == 0)
1527 			atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1528 		else
1529 			STAILQ_INSERT_TAIL(&iql, q, link);
1530 	}
1531 
1532 #if defined(INET) || defined(INET6)
1533 	if (iq->flags & IQ_LRO_ENABLED) {
1534 		struct lro_ctrl *lro = &rxq->lro;
1535 
1536 		tcp_lro_flush_all(lro);
1537 	}
1538 #endif
1539 
1540 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1541 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1542 
1543 	if (iq->flags & IQ_HAS_FL) {
1544 		int starved;
1545 
1546 		FL_LOCK(fl);
1547 		starved = refill_fl(sc, fl, 64);
1548 		FL_UNLOCK(fl);
1549 		if (__predict_false(starved != 0))
1550 			add_fl_to_sfl(sc, fl);
1551 	}
1552 
1553 	return (0);
1554 }
1555 
1556 static inline int
1557 cl_has_metadata(struct sge_fl *fl, struct cluster_layout *cll)
1558 {
1559 	int rc = fl->flags & FL_BUF_PACKING || cll->region1 > 0;
1560 
1561 	if (rc)
1562 		MPASS(cll->region3 >= CL_METADATA_SIZE);
1563 
1564 	return (rc);
1565 }
1566 
1567 static inline struct cluster_metadata *
1568 cl_metadata(struct adapter *sc, struct sge_fl *fl, struct cluster_layout *cll,
1569     caddr_t cl)
1570 {
1571 
1572 	if (cl_has_metadata(fl, cll)) {
1573 		struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1574 
1575 		return ((struct cluster_metadata *)(cl + swz->size) - 1);
1576 	}
1577 	return (NULL);
1578 }
1579 
1580 static void
1581 rxb_free(struct mbuf *m, void *arg1, void *arg2)
1582 {
1583 	uma_zone_t zone = arg1;
1584 	caddr_t cl = arg2;
1585 
1586 	uma_zfree(zone, cl);
1587 	counter_u64_add(extfree_rels, 1);
1588 }
1589 
1590 /*
1591  * The mbuf returned by this function could be allocated from zone_mbuf or
1592  * constructed in spare room in the cluster.
1593  *
1594  * The mbuf carries the payload in one of these ways
1595  * a) frame inside the mbuf (mbuf from zone_mbuf)
1596  * b) m_cljset (for clusters without metadata) zone_mbuf
1597  * c) m_extaddref (cluster with metadata) inline mbuf
1598  * d) m_extaddref (cluster with metadata) zone_mbuf
1599  */
1600 static struct mbuf *
1601 get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1602     int remaining)
1603 {
1604 	struct mbuf *m;
1605 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1606 	struct cluster_layout *cll = &sd->cll;
1607 	struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1608 	struct hw_buf_info *hwb = &sc->sge.hw_buf_info[cll->hwidx];
1609 	struct cluster_metadata *clm = cl_metadata(sc, fl, cll, sd->cl);
1610 	int len, blen;
1611 	caddr_t payload;
1612 
1613 	blen = hwb->size - fl->rx_offset;	/* max possible in this buf */
1614 	len = min(remaining, blen);
1615 	payload = sd->cl + cll->region1 + fl->rx_offset;
1616 	if (fl->flags & FL_BUF_PACKING) {
1617 		const u_int l = fr_offset + len;
1618 		const u_int pad = roundup2(l, fl->buf_boundary) - l;
1619 
1620 		if (fl->rx_offset + len + pad < hwb->size)
1621 			blen = len + pad;
1622 		MPASS(fl->rx_offset + blen <= hwb->size);
1623 	} else {
1624 		MPASS(fl->rx_offset == 0);	/* not packing */
1625 	}
1626 
1627 
1628 	if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) {
1629 
1630 		/*
1631 		 * Copy payload into a freshly allocated mbuf.
1632 		 */
1633 
1634 		m = fr_offset == 0 ?
1635 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1636 		if (m == NULL)
1637 			return (NULL);
1638 		fl->mbuf_allocated++;
1639 #ifdef T4_PKT_TIMESTAMP
1640 		/* Leave room for a timestamp */
1641 		m->m_data += 8;
1642 #endif
1643 		/* copy data to mbuf */
1644 		bcopy(payload, mtod(m, caddr_t), len);
1645 
1646 	} else if (sd->nmbuf * MSIZE < cll->region1) {
1647 
1648 		/*
1649 		 * There's spare room in the cluster for an mbuf.  Create one
1650 		 * and associate it with the payload that's in the cluster.
1651 		 */
1652 
1653 		MPASS(clm != NULL);
1654 		m = (struct mbuf *)(sd->cl + sd->nmbuf * MSIZE);
1655 		/* No bzero required */
1656 		if (m_init(m, M_NOWAIT, MT_DATA,
1657 		    fr_offset == 0 ? M_PKTHDR | M_NOFREE : M_NOFREE))
1658 			return (NULL);
1659 		fl->mbuf_inlined++;
1660 		m_extaddref(m, payload, blen, &clm->refcount, rxb_free,
1661 		    swz->zone, sd->cl);
1662 		if (sd->nmbuf++ == 0)
1663 			counter_u64_add(extfree_refs, 1);
1664 
1665 	} else {
1666 
1667 		/*
1668 		 * Grab an mbuf from zone_mbuf and associate it with the
1669 		 * payload in the cluster.
1670 		 */
1671 
1672 		m = fr_offset == 0 ?
1673 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1674 		if (m == NULL)
1675 			return (NULL);
1676 		fl->mbuf_allocated++;
1677 		if (clm != NULL) {
1678 			m_extaddref(m, payload, blen, &clm->refcount,
1679 			    rxb_free, swz->zone, sd->cl);
1680 			if (sd->nmbuf++ == 0)
1681 				counter_u64_add(extfree_refs, 1);
1682 		} else {
1683 			m_cljset(m, sd->cl, swz->type);
1684 			sd->cl = NULL;	/* consumed, not a recycle candidate */
1685 		}
1686 	}
1687 	if (fr_offset == 0)
1688 		m->m_pkthdr.len = remaining;
1689 	m->m_len = len;
1690 
1691 	if (fl->flags & FL_BUF_PACKING) {
1692 		fl->rx_offset += blen;
1693 		MPASS(fl->rx_offset <= hwb->size);
1694 		if (fl->rx_offset < hwb->size)
1695 			return (m);	/* without advancing the cidx */
1696 	}
1697 
1698 	if (__predict_false(++fl->cidx % 8 == 0)) {
1699 		uint16_t cidx = fl->cidx / 8;
1700 
1701 		if (__predict_false(cidx == fl->sidx))
1702 			fl->cidx = cidx = 0;
1703 		fl->hw_cidx = cidx;
1704 	}
1705 	fl->rx_offset = 0;
1706 
1707 	return (m);
1708 }
1709 
1710 static struct mbuf *
1711 get_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf)
1712 {
1713 	struct mbuf *m0, *m, **pnext;
1714 	u_int remaining;
1715 	const u_int total = G_RSPD_LEN(len_newbuf);
1716 
1717 	if (__predict_false(fl->flags & FL_BUF_RESUME)) {
1718 		M_ASSERTPKTHDR(fl->m0);
1719 		MPASS(fl->m0->m_pkthdr.len == total);
1720 		MPASS(fl->remaining < total);
1721 
1722 		m0 = fl->m0;
1723 		pnext = fl->pnext;
1724 		remaining = fl->remaining;
1725 		fl->flags &= ~FL_BUF_RESUME;
1726 		goto get_segment;
1727 	}
1728 
1729 	if (fl->rx_offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
1730 		fl->rx_offset = 0;
1731 		if (__predict_false(++fl->cidx % 8 == 0)) {
1732 			uint16_t cidx = fl->cidx / 8;
1733 
1734 			if (__predict_false(cidx == fl->sidx))
1735 				fl->cidx = cidx = 0;
1736 			fl->hw_cidx = cidx;
1737 		}
1738 	}
1739 
1740 	/*
1741 	 * Payload starts at rx_offset in the current hw buffer.  Its length is
1742 	 * 'len' and it may span multiple hw buffers.
1743 	 */
1744 
1745 	m0 = get_scatter_segment(sc, fl, 0, total);
1746 	if (m0 == NULL)
1747 		return (NULL);
1748 	remaining = total - m0->m_len;
1749 	pnext = &m0->m_next;
1750 	while (remaining > 0) {
1751 get_segment:
1752 		MPASS(fl->rx_offset == 0);
1753 		m = get_scatter_segment(sc, fl, total - remaining, remaining);
1754 		if (__predict_false(m == NULL)) {
1755 			fl->m0 = m0;
1756 			fl->pnext = pnext;
1757 			fl->remaining = remaining;
1758 			fl->flags |= FL_BUF_RESUME;
1759 			return (NULL);
1760 		}
1761 		*pnext = m;
1762 		pnext = &m->m_next;
1763 		remaining -= m->m_len;
1764 	}
1765 	*pnext = NULL;
1766 
1767 	M_ASSERTPKTHDR(m0);
1768 	return (m0);
1769 }
1770 
1771 static int
1772 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
1773 {
1774 	struct sge_rxq *rxq = iq_to_rxq(iq);
1775 	struct ifnet *ifp = rxq->ifp;
1776 	struct adapter *sc = iq->adapter;
1777 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
1778 #if defined(INET) || defined(INET6)
1779 	struct lro_ctrl *lro = &rxq->lro;
1780 #endif
1781 	static const int sw_hashtype[4][2] = {
1782 		{M_HASHTYPE_NONE, M_HASHTYPE_NONE},
1783 		{M_HASHTYPE_RSS_IPV4, M_HASHTYPE_RSS_IPV6},
1784 		{M_HASHTYPE_RSS_TCP_IPV4, M_HASHTYPE_RSS_TCP_IPV6},
1785 		{M_HASHTYPE_RSS_UDP_IPV4, M_HASHTYPE_RSS_UDP_IPV6},
1786 	};
1787 
1788 	KASSERT(m0 != NULL, ("%s: no payload with opcode %02x", __func__,
1789 	    rss->opcode));
1790 
1791 	m0->m_pkthdr.len -= sc->params.sge.fl_pktshift;
1792 	m0->m_len -= sc->params.sge.fl_pktshift;
1793 	m0->m_data += sc->params.sge.fl_pktshift;
1794 
1795 	m0->m_pkthdr.rcvif = ifp;
1796 	M_HASHTYPE_SET(m0, sw_hashtype[rss->hash_type][rss->ipv6]);
1797 	m0->m_pkthdr.flowid = be32toh(rss->hash_val);
1798 
1799 	if (cpl->csum_calc && !cpl->err_vec) {
1800 		if (ifp->if_capenable & IFCAP_RXCSUM &&
1801 		    cpl->l2info & htobe32(F_RXF_IP)) {
1802 			m0->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
1803 			    CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
1804 			rxq->rxcsum++;
1805 		} else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
1806 		    cpl->l2info & htobe32(F_RXF_IP6)) {
1807 			m0->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
1808 			    CSUM_PSEUDO_HDR);
1809 			rxq->rxcsum++;
1810 		}
1811 
1812 		if (__predict_false(cpl->ip_frag))
1813 			m0->m_pkthdr.csum_data = be16toh(cpl->csum);
1814 		else
1815 			m0->m_pkthdr.csum_data = 0xffff;
1816 	}
1817 
1818 	if (cpl->vlan_ex) {
1819 		m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
1820 		m0->m_flags |= M_VLANTAG;
1821 		rxq->vlan_extraction++;
1822 	}
1823 
1824 #if defined(INET) || defined(INET6)
1825 	if (cpl->l2info & htobe32(F_RXF_LRO) &&
1826 	    iq->flags & IQ_LRO_ENABLED &&
1827 	    tcp_lro_rx(lro, m0, 0) == 0) {
1828 		/* queued for LRO */
1829 	} else
1830 #endif
1831 	ifp->if_input(ifp, m0);
1832 
1833 	return (0);
1834 }
1835 
1836 /*
1837  * Must drain the wrq or make sure that someone else will.
1838  */
1839 static void
1840 wrq_tx_drain(void *arg, int n)
1841 {
1842 	struct sge_wrq *wrq = arg;
1843 	struct sge_eq *eq = &wrq->eq;
1844 
1845 	EQ_LOCK(eq);
1846 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
1847 		drain_wrq_wr_list(wrq->adapter, wrq);
1848 	EQ_UNLOCK(eq);
1849 }
1850 
1851 static void
1852 drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq)
1853 {
1854 	struct sge_eq *eq = &wrq->eq;
1855 	u_int available, dbdiff;	/* # of hardware descriptors */
1856 	u_int n;
1857 	struct wrqe *wr;
1858 	struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
1859 
1860 	EQ_LOCK_ASSERT_OWNED(eq);
1861 	MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
1862 	wr = STAILQ_FIRST(&wrq->wr_list);
1863 	MPASS(wr != NULL);	/* Must be called with something useful to do */
1864 	MPASS(eq->pidx == eq->dbidx);
1865 	dbdiff = 0;
1866 
1867 	do {
1868 		eq->cidx = read_hw_cidx(eq);
1869 		if (eq->pidx == eq->cidx)
1870 			available = eq->sidx - 1;
1871 		else
1872 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
1873 
1874 		MPASS(wr->wrq == wrq);
1875 		n = howmany(wr->wr_len, EQ_ESIZE);
1876 		if (available < n)
1877 			break;
1878 
1879 		dst = (void *)&eq->desc[eq->pidx];
1880 		if (__predict_true(eq->sidx - eq->pidx > n)) {
1881 			/* Won't wrap, won't end exactly at the status page. */
1882 			bcopy(&wr->wr[0], dst, wr->wr_len);
1883 			eq->pidx += n;
1884 		} else {
1885 			int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE;
1886 
1887 			bcopy(&wr->wr[0], dst, first_portion);
1888 			if (wr->wr_len > first_portion) {
1889 				bcopy(&wr->wr[first_portion], &eq->desc[0],
1890 				    wr->wr_len - first_portion);
1891 			}
1892 			eq->pidx = n - (eq->sidx - eq->pidx);
1893 		}
1894 
1895 		if (available < eq->sidx / 4 &&
1896 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
1897 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
1898 			    F_FW_WR_EQUEQ);
1899 			eq->equeqidx = eq->pidx;
1900 		} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
1901 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
1902 			eq->equeqidx = eq->pidx;
1903 		}
1904 
1905 		dbdiff += n;
1906 		if (dbdiff >= 16) {
1907 			ring_eq_db(sc, eq, dbdiff);
1908 			dbdiff = 0;
1909 		}
1910 
1911 		STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
1912 		free_wrqe(wr);
1913 		MPASS(wrq->nwr_pending > 0);
1914 		wrq->nwr_pending--;
1915 		MPASS(wrq->ndesc_needed >= n);
1916 		wrq->ndesc_needed -= n;
1917 	} while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL);
1918 
1919 	if (dbdiff)
1920 		ring_eq_db(sc, eq, dbdiff);
1921 }
1922 
1923 /*
1924  * Doesn't fail.  Holds on to work requests it can't send right away.
1925  */
1926 void
1927 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
1928 {
1929 #ifdef INVARIANTS
1930 	struct sge_eq *eq = &wrq->eq;
1931 #endif
1932 
1933 	EQ_LOCK_ASSERT_OWNED(eq);
1934 	MPASS(wr != NULL);
1935 	MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN);
1936 	MPASS((wr->wr_len & 0x7) == 0);
1937 
1938 	STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
1939 	wrq->nwr_pending++;
1940 	wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE);
1941 
1942 	if (!TAILQ_EMPTY(&wrq->incomplete_wrs))
1943 		return;	/* commit_wrq_wr will drain wr_list as well. */
1944 
1945 	drain_wrq_wr_list(sc, wrq);
1946 
1947 	/* Doorbell must have caught up to the pidx. */
1948 	MPASS(eq->pidx == eq->dbidx);
1949 }
1950 
1951 void
1952 t4_update_fl_bufsize(struct ifnet *ifp)
1953 {
1954 	struct vi_info *vi = ifp->if_softc;
1955 	struct adapter *sc = vi->pi->adapter;
1956 	struct sge_rxq *rxq;
1957 #ifdef TCP_OFFLOAD
1958 	struct sge_ofld_rxq *ofld_rxq;
1959 #endif
1960 	struct sge_fl *fl;
1961 	int i, maxp, mtu = ifp->if_mtu;
1962 
1963 	maxp = mtu_to_max_payload(sc, mtu, 0);
1964 	for_each_rxq(vi, i, rxq) {
1965 		fl = &rxq->fl;
1966 
1967 		FL_LOCK(fl);
1968 		find_best_refill_source(sc, fl, maxp);
1969 		FL_UNLOCK(fl);
1970 	}
1971 #ifdef TCP_OFFLOAD
1972 	maxp = mtu_to_max_payload(sc, mtu, 1);
1973 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1974 		fl = &ofld_rxq->fl;
1975 
1976 		FL_LOCK(fl);
1977 		find_best_refill_source(sc, fl, maxp);
1978 		FL_UNLOCK(fl);
1979 	}
1980 #endif
1981 }
1982 
1983 static inline int
1984 mbuf_nsegs(struct mbuf *m)
1985 {
1986 
1987 	M_ASSERTPKTHDR(m);
1988 	KASSERT(m->m_pkthdr.l5hlen > 0,
1989 	    ("%s: mbuf %p missing information on # of segments.", __func__, m));
1990 
1991 	return (m->m_pkthdr.l5hlen);
1992 }
1993 
1994 static inline void
1995 set_mbuf_nsegs(struct mbuf *m, uint8_t nsegs)
1996 {
1997 
1998 	M_ASSERTPKTHDR(m);
1999 	m->m_pkthdr.l5hlen = nsegs;
2000 }
2001 
2002 static inline int
2003 mbuf_len16(struct mbuf *m)
2004 {
2005 	int n;
2006 
2007 	M_ASSERTPKTHDR(m);
2008 	n = m->m_pkthdr.PH_loc.eight[0];
2009 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2010 
2011 	return (n);
2012 }
2013 
2014 static inline void
2015 set_mbuf_len16(struct mbuf *m, uint8_t len16)
2016 {
2017 
2018 	M_ASSERTPKTHDR(m);
2019 	m->m_pkthdr.PH_loc.eight[0] = len16;
2020 }
2021 
2022 static inline int
2023 needs_tso(struct mbuf *m)
2024 {
2025 
2026 	M_ASSERTPKTHDR(m);
2027 
2028 	if (m->m_pkthdr.csum_flags & CSUM_TSO) {
2029 		KASSERT(m->m_pkthdr.tso_segsz > 0,
2030 		    ("%s: TSO requested in mbuf %p but MSS not provided",
2031 		    __func__, m));
2032 		return (1);
2033 	}
2034 
2035 	return (0);
2036 }
2037 
2038 static inline int
2039 needs_l3_csum(struct mbuf *m)
2040 {
2041 
2042 	M_ASSERTPKTHDR(m);
2043 
2044 	if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO))
2045 		return (1);
2046 	return (0);
2047 }
2048 
2049 static inline int
2050 needs_l4_csum(struct mbuf *m)
2051 {
2052 
2053 	M_ASSERTPKTHDR(m);
2054 
2055 	if (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
2056 	    CSUM_TCP_IPV6 | CSUM_TSO))
2057 		return (1);
2058 	return (0);
2059 }
2060 
2061 static inline int
2062 needs_vlan_insertion(struct mbuf *m)
2063 {
2064 
2065 	M_ASSERTPKTHDR(m);
2066 
2067 	if (m->m_flags & M_VLANTAG) {
2068 		KASSERT(m->m_pkthdr.ether_vtag != 0,
2069 		    ("%s: HWVLAN requested in mbuf %p but tag not provided",
2070 		    __func__, m));
2071 		return (1);
2072 	}
2073 	return (0);
2074 }
2075 
2076 static void *
2077 m_advance(struct mbuf **pm, int *poffset, int len)
2078 {
2079 	struct mbuf *m = *pm;
2080 	int offset = *poffset;
2081 	uintptr_t p = 0;
2082 
2083 	MPASS(len > 0);
2084 
2085 	while (len) {
2086 		if (offset + len < m->m_len) {
2087 			offset += len;
2088 			p = mtod(m, uintptr_t) + offset;
2089 			break;
2090 		}
2091 		len -= m->m_len - offset;
2092 		m = m->m_next;
2093 		offset = 0;
2094 		MPASS(m != NULL);
2095 	}
2096 	*poffset = offset;
2097 	*pm = m;
2098 	return ((void *)p);
2099 }
2100 
2101 static inline int
2102 same_paddr(char *a, char *b)
2103 {
2104 
2105 	if (a == b)
2106 		return (1);
2107 	else if (a != NULL && b != NULL) {
2108 		vm_offset_t x = (vm_offset_t)a;
2109 		vm_offset_t y = (vm_offset_t)b;
2110 
2111 		if ((x & PAGE_MASK) == (y & PAGE_MASK) &&
2112 		    pmap_kextract(x) == pmap_kextract(y))
2113 			return (1);
2114 	}
2115 
2116 	return (0);
2117 }
2118 
2119 /*
2120  * Can deal with empty mbufs in the chain that have m_len = 0, but the chain
2121  * must have at least one mbuf that's not empty.
2122  */
2123 static inline int
2124 count_mbuf_nsegs(struct mbuf *m)
2125 {
2126 	char *prev_end, *start;
2127 	int len, nsegs;
2128 
2129 	MPASS(m != NULL);
2130 
2131 	nsegs = 0;
2132 	prev_end = NULL;
2133 	for (; m; m = m->m_next) {
2134 
2135 		len = m->m_len;
2136 		if (__predict_false(len == 0))
2137 			continue;
2138 		start = mtod(m, char *);
2139 
2140 		nsegs += sglist_count(start, len);
2141 		if (same_paddr(prev_end, start))
2142 			nsegs--;
2143 		prev_end = start + len;
2144 	}
2145 
2146 	MPASS(nsegs > 0);
2147 	return (nsegs);
2148 }
2149 
2150 /*
2151  * Analyze the mbuf to determine its tx needs.  The mbuf passed in may change:
2152  * a) caller can assume it's been freed if this function returns with an error.
2153  * b) it may get defragged up if the gather list is too long for the hardware.
2154  */
2155 int
2156 parse_pkt(struct mbuf **mp)
2157 {
2158 	struct mbuf *m0 = *mp, *m;
2159 	int rc, nsegs, defragged = 0, offset;
2160 	struct ether_header *eh;
2161 	void *l3hdr;
2162 #if defined(INET) || defined(INET6)
2163 	struct tcphdr *tcp;
2164 #endif
2165 	uint16_t eh_type;
2166 
2167 	M_ASSERTPKTHDR(m0);
2168 	if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) {
2169 		rc = EINVAL;
2170 fail:
2171 		m_freem(m0);
2172 		*mp = NULL;
2173 		return (rc);
2174 	}
2175 restart:
2176 	/*
2177 	 * First count the number of gather list segments in the payload.
2178 	 * Defrag the mbuf if nsegs exceeds the hardware limit.
2179 	 */
2180 	M_ASSERTPKTHDR(m0);
2181 	MPASS(m0->m_pkthdr.len > 0);
2182 	nsegs = count_mbuf_nsegs(m0);
2183 	if (nsegs > (needs_tso(m0) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS)) {
2184 		if (defragged++ > 0 || (m = m_defrag(m0, M_NOWAIT)) == NULL) {
2185 			rc = EFBIG;
2186 			goto fail;
2187 		}
2188 		*mp = m0 = m;	/* update caller's copy after defrag */
2189 		goto restart;
2190 	}
2191 
2192 	if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN)) {
2193 		m0 = m_pullup(m0, m0->m_pkthdr.len);
2194 		if (m0 == NULL) {
2195 			/* Should have left well enough alone. */
2196 			rc = EFBIG;
2197 			goto fail;
2198 		}
2199 		*mp = m0;	/* update caller's copy after pullup */
2200 		goto restart;
2201 	}
2202 	set_mbuf_nsegs(m0, nsegs);
2203 	set_mbuf_len16(m0, txpkt_len16(nsegs, needs_tso(m0)));
2204 
2205 	if (!needs_tso(m0))
2206 		return (0);
2207 
2208 	m = m0;
2209 	eh = mtod(m, struct ether_header *);
2210 	eh_type = ntohs(eh->ether_type);
2211 	if (eh_type == ETHERTYPE_VLAN) {
2212 		struct ether_vlan_header *evh = (void *)eh;
2213 
2214 		eh_type = ntohs(evh->evl_proto);
2215 		m0->m_pkthdr.l2hlen = sizeof(*evh);
2216 	} else
2217 		m0->m_pkthdr.l2hlen = sizeof(*eh);
2218 
2219 	offset = 0;
2220 	l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2221 
2222 	switch (eh_type) {
2223 #ifdef INET6
2224 	case ETHERTYPE_IPV6:
2225 	{
2226 		struct ip6_hdr *ip6 = l3hdr;
2227 
2228 		MPASS(ip6->ip6_nxt == IPPROTO_TCP);
2229 
2230 		m0->m_pkthdr.l3hlen = sizeof(*ip6);
2231 		break;
2232 	}
2233 #endif
2234 #ifdef INET
2235 	case ETHERTYPE_IP:
2236 	{
2237 		struct ip *ip = l3hdr;
2238 
2239 		m0->m_pkthdr.l3hlen = ip->ip_hl * 4;
2240 		break;
2241 	}
2242 #endif
2243 	default:
2244 		panic("%s: ethertype 0x%04x unknown.  if_cxgbe must be compiled"
2245 		    " with the same INET/INET6 options as the kernel.",
2246 		    __func__, eh_type);
2247 	}
2248 
2249 #if defined(INET) || defined(INET6)
2250 	tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen);
2251 	m0->m_pkthdr.l4hlen = tcp->th_off * 4;
2252 #endif
2253 	MPASS(m0 == *mp);
2254 	return (0);
2255 }
2256 
2257 void *
2258 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie)
2259 {
2260 	struct sge_eq *eq = &wrq->eq;
2261 	struct adapter *sc = wrq->adapter;
2262 	int ndesc, available;
2263 	struct wrqe *wr;
2264 	void *w;
2265 
2266 	MPASS(len16 > 0);
2267 	ndesc = howmany(len16, EQ_ESIZE / 16);
2268 	MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC);
2269 
2270 	EQ_LOCK(eq);
2271 
2272 	if (!STAILQ_EMPTY(&wrq->wr_list))
2273 		drain_wrq_wr_list(sc, wrq);
2274 
2275 	if (!STAILQ_EMPTY(&wrq->wr_list)) {
2276 slowpath:
2277 		EQ_UNLOCK(eq);
2278 		wr = alloc_wrqe(len16 * 16, wrq);
2279 		if (__predict_false(wr == NULL))
2280 			return (NULL);
2281 		cookie->pidx = -1;
2282 		cookie->ndesc = ndesc;
2283 		return (&wr->wr);
2284 	}
2285 
2286 	eq->cidx = read_hw_cidx(eq);
2287 	if (eq->pidx == eq->cidx)
2288 		available = eq->sidx - 1;
2289 	else
2290 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2291 	if (available < ndesc)
2292 		goto slowpath;
2293 
2294 	cookie->pidx = eq->pidx;
2295 	cookie->ndesc = ndesc;
2296 	TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link);
2297 
2298 	w = &eq->desc[eq->pidx];
2299 	IDXINCR(eq->pidx, ndesc, eq->sidx);
2300 	if (__predict_false(eq->pidx < ndesc - 1)) {
2301 		w = &wrq->ss[0];
2302 		wrq->ss_pidx = cookie->pidx;
2303 		wrq->ss_len = len16 * 16;
2304 	}
2305 
2306 	EQ_UNLOCK(eq);
2307 
2308 	return (w);
2309 }
2310 
2311 void
2312 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie)
2313 {
2314 	struct sge_eq *eq = &wrq->eq;
2315 	struct adapter *sc = wrq->adapter;
2316 	int ndesc, pidx;
2317 	struct wrq_cookie *prev, *next;
2318 
2319 	if (cookie->pidx == -1) {
2320 		struct wrqe *wr = __containerof(w, struct wrqe, wr);
2321 
2322 		t4_wrq_tx(sc, wr);
2323 		return;
2324 	}
2325 
2326 	ndesc = cookie->ndesc;	/* Can be more than SGE_MAX_WR_NDESC here. */
2327 	pidx = cookie->pidx;
2328 	MPASS(pidx >= 0 && pidx < eq->sidx);
2329 	if (__predict_false(w == &wrq->ss[0])) {
2330 		int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE;
2331 
2332 		MPASS(wrq->ss_len > n);	/* WR had better wrap around. */
2333 		bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n);
2334 		bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n);
2335 		wrq->tx_wrs_ss++;
2336 	} else
2337 		wrq->tx_wrs_direct++;
2338 
2339 	EQ_LOCK(eq);
2340 	prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link);
2341 	next = TAILQ_NEXT(cookie, link);
2342 	if (prev == NULL) {
2343 		MPASS(pidx == eq->dbidx);
2344 		if (next == NULL || ndesc >= 16)
2345 			ring_eq_db(wrq->adapter, eq, ndesc);
2346 		else {
2347 			MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc);
2348 			next->pidx = pidx;
2349 			next->ndesc += ndesc;
2350 		}
2351 	} else {
2352 		MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc);
2353 		prev->ndesc += ndesc;
2354 	}
2355 	TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link);
2356 
2357 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2358 		drain_wrq_wr_list(sc, wrq);
2359 
2360 #ifdef INVARIANTS
2361 	if (TAILQ_EMPTY(&wrq->incomplete_wrs)) {
2362 		/* Doorbell must have caught up to the pidx. */
2363 		MPASS(wrq->eq.pidx == wrq->eq.dbidx);
2364 	}
2365 #endif
2366 	EQ_UNLOCK(eq);
2367 }
2368 
2369 static u_int
2370 can_resume_eth_tx(struct mp_ring *r)
2371 {
2372 	struct sge_eq *eq = r->cookie;
2373 
2374 	return (total_available_tx_desc(eq) > eq->sidx / 8);
2375 }
2376 
2377 static inline int
2378 cannot_use_txpkts(struct mbuf *m)
2379 {
2380 	/* maybe put a GL limit too, to avoid silliness? */
2381 
2382 	return (needs_tso(m));
2383 }
2384 
2385 /*
2386  * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to
2387  * be consumed.  Return the actual number consumed.  0 indicates a stall.
2388  */
2389 static u_int
2390 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx)
2391 {
2392 	struct sge_txq *txq = r->cookie;
2393 	struct sge_eq *eq = &txq->eq;
2394 	struct ifnet *ifp = txq->ifp;
2395 	struct vi_info *vi = ifp->if_softc;
2396 	struct port_info *pi = vi->pi;
2397 	struct adapter *sc = pi->adapter;
2398 	u_int total, remaining;		/* # of packets */
2399 	u_int available, dbdiff;	/* # of hardware descriptors */
2400 	u_int n, next_cidx;
2401 	struct mbuf *m0, *tail;
2402 	struct txpkts txp;
2403 	struct fw_eth_tx_pkts_wr *wr;	/* any fw WR struct will do */
2404 
2405 	remaining = IDXDIFF(pidx, cidx, r->size);
2406 	MPASS(remaining > 0);	/* Must not be called without work to do. */
2407 	total = 0;
2408 
2409 	TXQ_LOCK(txq);
2410 	if (__predict_false((eq->flags & EQ_ENABLED) == 0)) {
2411 		while (cidx != pidx) {
2412 			m0 = r->items[cidx];
2413 			m_freem(m0);
2414 			if (++cidx == r->size)
2415 				cidx = 0;
2416 		}
2417 		reclaim_tx_descs(txq, 2048);
2418 		total = remaining;
2419 		goto done;
2420 	}
2421 
2422 	/* How many hardware descriptors do we have readily available. */
2423 	if (eq->pidx == eq->cidx)
2424 		available = eq->sidx - 1;
2425 	else
2426 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2427 	dbdiff = IDXDIFF(eq->pidx, eq->dbidx, eq->sidx);
2428 
2429 	while (remaining > 0) {
2430 
2431 		m0 = r->items[cidx];
2432 		M_ASSERTPKTHDR(m0);
2433 		MPASS(m0->m_nextpkt == NULL);
2434 
2435 		if (available < SGE_MAX_WR_NDESC) {
2436 			available += reclaim_tx_descs(txq, 64);
2437 			if (available < howmany(mbuf_len16(m0), EQ_ESIZE / 16))
2438 				break;	/* out of descriptors */
2439 		}
2440 
2441 		next_cidx = cidx + 1;
2442 		if (__predict_false(next_cidx == r->size))
2443 			next_cidx = 0;
2444 
2445 		wr = (void *)&eq->desc[eq->pidx];
2446 		if (remaining > 1 &&
2447 		    try_txpkts(m0, r->items[next_cidx], &txp, available) == 0) {
2448 
2449 			/* pkts at cidx, next_cidx should both be in txp. */
2450 			MPASS(txp.npkt == 2);
2451 			tail = r->items[next_cidx];
2452 			MPASS(tail->m_nextpkt == NULL);
2453 			ETHER_BPF_MTAP(ifp, m0);
2454 			ETHER_BPF_MTAP(ifp, tail);
2455 			m0->m_nextpkt = tail;
2456 
2457 			if (__predict_false(++next_cidx == r->size))
2458 				next_cidx = 0;
2459 
2460 			while (next_cidx != pidx) {
2461 				if (add_to_txpkts(r->items[next_cidx], &txp,
2462 				    available) != 0)
2463 					break;
2464 				tail->m_nextpkt = r->items[next_cidx];
2465 				tail = tail->m_nextpkt;
2466 				ETHER_BPF_MTAP(ifp, tail);
2467 				if (__predict_false(++next_cidx == r->size))
2468 					next_cidx = 0;
2469 			}
2470 
2471 			n = write_txpkts_wr(txq, wr, m0, &txp, available);
2472 			total += txp.npkt;
2473 			remaining -= txp.npkt;
2474 		} else {
2475 			total++;
2476 			remaining--;
2477 			ETHER_BPF_MTAP(ifp, m0);
2478 			n = write_txpkt_wr(txq, (void *)wr, m0, available);
2479 		}
2480 		MPASS(n >= 1 && n <= available && n <= SGE_MAX_WR_NDESC);
2481 
2482 		available -= n;
2483 		dbdiff += n;
2484 		IDXINCR(eq->pidx, n, eq->sidx);
2485 
2486 		if (total_available_tx_desc(eq) < eq->sidx / 4 &&
2487 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2488 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2489 			    F_FW_WR_EQUEQ);
2490 			eq->equeqidx = eq->pidx;
2491 		} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
2492 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
2493 			eq->equeqidx = eq->pidx;
2494 		}
2495 
2496 		if (dbdiff >= 16 && remaining >= 4) {
2497 			ring_eq_db(sc, eq, dbdiff);
2498 			available += reclaim_tx_descs(txq, 4 * dbdiff);
2499 			dbdiff = 0;
2500 		}
2501 
2502 		cidx = next_cidx;
2503 	}
2504 	if (dbdiff != 0) {
2505 		ring_eq_db(sc, eq, dbdiff);
2506 		reclaim_tx_descs(txq, 32);
2507 	}
2508 done:
2509 	TXQ_UNLOCK(txq);
2510 
2511 	return (total);
2512 }
2513 
2514 static inline void
2515 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
2516     int qsize)
2517 {
2518 
2519 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
2520 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
2521 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
2522 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
2523 
2524 	iq->flags = 0;
2525 	iq->adapter = sc;
2526 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
2527 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
2528 	if (pktc_idx >= 0) {
2529 		iq->intr_params |= F_QINTR_CNT_EN;
2530 		iq->intr_pktc_idx = pktc_idx;
2531 	}
2532 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
2533 	iq->sidx = iq->qsize - sc->params.sge.spg_len / IQ_ESIZE;
2534 }
2535 
2536 static inline void
2537 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name)
2538 {
2539 
2540 	fl->qsize = qsize;
2541 	fl->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2542 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
2543 	if (sc->flags & BUF_PACKING_OK &&
2544 	    ((!is_t4(sc) && buffer_packing) ||	/* T5+: enabled unless 0 */
2545 	    (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */
2546 		fl->flags |= FL_BUF_PACKING;
2547 	find_best_refill_source(sc, fl, maxp);
2548 	find_safe_refill_source(sc, fl);
2549 }
2550 
2551 static inline void
2552 init_eq(struct adapter *sc, struct sge_eq *eq, int eqtype, int qsize,
2553     uint8_t tx_chan, uint16_t iqid, char *name)
2554 {
2555 	KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
2556 
2557 	eq->flags = eqtype & EQ_TYPEMASK;
2558 	eq->tx_chan = tx_chan;
2559 	eq->iqid = iqid;
2560 	eq->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2561 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
2562 }
2563 
2564 static int
2565 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
2566     bus_dmamap_t *map, bus_addr_t *pa, void **va)
2567 {
2568 	int rc;
2569 
2570 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
2571 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
2572 	if (rc != 0) {
2573 		device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
2574 		goto done;
2575 	}
2576 
2577 	rc = bus_dmamem_alloc(*tag, va,
2578 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
2579 	if (rc != 0) {
2580 		device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
2581 		goto done;
2582 	}
2583 
2584 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
2585 	if (rc != 0) {
2586 		device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
2587 		goto done;
2588 	}
2589 done:
2590 	if (rc)
2591 		free_ring(sc, *tag, *map, *pa, *va);
2592 
2593 	return (rc);
2594 }
2595 
2596 static int
2597 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
2598     bus_addr_t pa, void *va)
2599 {
2600 	if (pa)
2601 		bus_dmamap_unload(tag, map);
2602 	if (va)
2603 		bus_dmamem_free(tag, va, map);
2604 	if (tag)
2605 		bus_dma_tag_destroy(tag);
2606 
2607 	return (0);
2608 }
2609 
2610 /*
2611  * Allocates the ring for an ingress queue and an optional freelist.  If the
2612  * freelist is specified it will be allocated and then associated with the
2613  * ingress queue.
2614  *
2615  * Returns errno on failure.  Resources allocated up to that point may still be
2616  * allocated.  Caller is responsible for cleanup in case this function fails.
2617  *
2618  * If the ingress queue will take interrupts directly (iq->flags & IQ_INTR) then
2619  * the intr_idx specifies the vector, starting from 0.  Otherwise it specifies
2620  * the abs_id of the ingress queue to which its interrupts should be forwarded.
2621  */
2622 static int
2623 alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl,
2624     int intr_idx, int cong)
2625 {
2626 	int rc, i, cntxt_id;
2627 	size_t len;
2628 	struct fw_iq_cmd c;
2629 	struct port_info *pi = vi->pi;
2630 	struct adapter *sc = iq->adapter;
2631 	struct sge_params *sp = &sc->params.sge;
2632 	__be32 v = 0;
2633 
2634 	len = iq->qsize * IQ_ESIZE;
2635 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
2636 	    (void **)&iq->desc);
2637 	if (rc != 0)
2638 		return (rc);
2639 
2640 	bzero(&c, sizeof(c));
2641 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
2642 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
2643 	    V_FW_IQ_CMD_VFN(0));
2644 
2645 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
2646 	    FW_LEN16(c));
2647 
2648 	/* Special handling for firmware event queue */
2649 	if (iq == &sc->sge.fwq)
2650 		v |= F_FW_IQ_CMD_IQASYNCH;
2651 
2652 	if (iq->flags & IQ_INTR) {
2653 		KASSERT(intr_idx < sc->intr_count,
2654 		    ("%s: invalid direct intr_idx %d", __func__, intr_idx));
2655 	} else
2656 		v |= F_FW_IQ_CMD_IQANDST;
2657 	v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
2658 
2659 	c.type_to_iqandstindex = htobe32(v |
2660 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
2661 	    V_FW_IQ_CMD_VIID(vi->viid) |
2662 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
2663 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
2664 	    F_FW_IQ_CMD_IQGTSMODE |
2665 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
2666 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
2667 	c.iqsize = htobe16(iq->qsize);
2668 	c.iqaddr = htobe64(iq->ba);
2669 	if (cong >= 0)
2670 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
2671 
2672 	if (fl) {
2673 		mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
2674 
2675 		len = fl->qsize * EQ_ESIZE;
2676 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
2677 		    &fl->ba, (void **)&fl->desc);
2678 		if (rc)
2679 			return (rc);
2680 
2681 		/* Allocate space for one software descriptor per buffer. */
2682 		rc = alloc_fl_sdesc(fl);
2683 		if (rc != 0) {
2684 			device_printf(sc->dev,
2685 			    "failed to setup fl software descriptors: %d\n",
2686 			    rc);
2687 			return (rc);
2688 		}
2689 
2690 		if (fl->flags & FL_BUF_PACKING) {
2691 			fl->lowat = roundup2(sp->fl_starve_threshold2, 8);
2692 			fl->buf_boundary = sp->pack_boundary;
2693 		} else {
2694 			fl->lowat = roundup2(sp->fl_starve_threshold, 8);
2695 			fl->buf_boundary = 16;
2696 		}
2697 		if (fl_pad && fl->buf_boundary < sp->pad_boundary)
2698 			fl->buf_boundary = sp->pad_boundary;
2699 
2700 		c.iqns_to_fl0congen |=
2701 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
2702 			F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
2703 			(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
2704 			(fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
2705 			    0));
2706 		if (cong >= 0) {
2707 			c.iqns_to_fl0congen |=
2708 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
2709 				    F_FW_IQ_CMD_FL0CONGCIF |
2710 				    F_FW_IQ_CMD_FL0CONGEN);
2711 		}
2712 		c.fl0dcaen_to_fl0cidxfthresh =
2713 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(X_FETCHBURSTMIN_128B) |
2714 			V_FW_IQ_CMD_FL0FBMAX(X_FETCHBURSTMAX_512B));
2715 		c.fl0size = htobe16(fl->qsize);
2716 		c.fl0addr = htobe64(fl->ba);
2717 	}
2718 
2719 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
2720 	if (rc != 0) {
2721 		device_printf(sc->dev,
2722 		    "failed to create ingress queue: %d\n", rc);
2723 		return (rc);
2724 	}
2725 
2726 	iq->cidx = 0;
2727 	iq->gen = F_RSPD_GEN;
2728 	iq->intr_next = iq->intr_params;
2729 	iq->cntxt_id = be16toh(c.iqid);
2730 	iq->abs_id = be16toh(c.physiqid);
2731 	iq->flags |= IQ_ALLOCATED;
2732 
2733 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
2734 	if (cntxt_id >= sc->sge.niq) {
2735 		panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
2736 		    cntxt_id, sc->sge.niq - 1);
2737 	}
2738 	sc->sge.iqmap[cntxt_id] = iq;
2739 
2740 	if (fl) {
2741 		u_int qid;
2742 
2743 		iq->flags |= IQ_HAS_FL;
2744 		fl->cntxt_id = be16toh(c.fl0id);
2745 		fl->pidx = fl->cidx = 0;
2746 
2747 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
2748 		if (cntxt_id >= sc->sge.neq) {
2749 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
2750 			    __func__, cntxt_id, sc->sge.neq - 1);
2751 		}
2752 		sc->sge.eqmap[cntxt_id] = (void *)fl;
2753 
2754 		qid = fl->cntxt_id;
2755 		if (isset(&sc->doorbells, DOORBELL_UDB)) {
2756 			uint32_t s_qpp = sc->params.sge.eq_s_qpp;
2757 			uint32_t mask = (1 << s_qpp) - 1;
2758 			volatile uint8_t *udb;
2759 
2760 			udb = sc->udbs_base + UDBS_DB_OFFSET;
2761 			udb += (qid >> s_qpp) << PAGE_SHIFT;
2762 			qid &= mask;
2763 			if (qid < PAGE_SIZE / UDBS_SEG_SIZE) {
2764 				udb += qid << UDBS_SEG_SHIFT;
2765 				qid = 0;
2766 			}
2767 			fl->udb = (volatile void *)udb;
2768 		}
2769 		fl->dbval = V_QID(qid) | sc->chip_params->sge_fl_db;
2770 
2771 		FL_LOCK(fl);
2772 		/* Enough to make sure the SGE doesn't think it's starved */
2773 		refill_fl(sc, fl, fl->lowat);
2774 		FL_UNLOCK(fl);
2775 	}
2776 
2777 	if (is_t5(sc) && cong >= 0) {
2778 		uint32_t param, val;
2779 
2780 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
2781 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
2782 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
2783 		if (cong == 0)
2784 			val = 1 << 19;
2785 		else {
2786 			val = 2 << 19;
2787 			for (i = 0; i < 4; i++) {
2788 				if (cong & (1 << i))
2789 					val |= 1 << (i << 2);
2790 			}
2791 		}
2792 
2793 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
2794 		if (rc != 0) {
2795 			/* report error but carry on */
2796 			device_printf(sc->dev,
2797 			    "failed to set congestion manager context for "
2798 			    "ingress queue %d: %d\n", iq->cntxt_id, rc);
2799 		}
2800 	}
2801 
2802 	/* Enable IQ interrupts */
2803 	atomic_store_rel_int(&iq->state, IQS_IDLE);
2804 	t4_write_reg(sc, sc->sge_gts_reg, V_SEINTARM(iq->intr_params) |
2805 	    V_INGRESSQID(iq->cntxt_id));
2806 
2807 	return (0);
2808 }
2809 
2810 static int
2811 free_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl)
2812 {
2813 	int rc;
2814 	struct adapter *sc = iq->adapter;
2815 	device_t dev;
2816 
2817 	if (sc == NULL)
2818 		return (0);	/* nothing to do */
2819 
2820 	dev = vi ? vi->dev : sc->dev;
2821 
2822 	if (iq->flags & IQ_ALLOCATED) {
2823 		rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
2824 		    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
2825 		    fl ? fl->cntxt_id : 0xffff, 0xffff);
2826 		if (rc != 0) {
2827 			device_printf(dev,
2828 			    "failed to free queue %p: %d\n", iq, rc);
2829 			return (rc);
2830 		}
2831 		iq->flags &= ~IQ_ALLOCATED;
2832 	}
2833 
2834 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
2835 
2836 	bzero(iq, sizeof(*iq));
2837 
2838 	if (fl) {
2839 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
2840 		    fl->desc);
2841 
2842 		if (fl->sdesc)
2843 			free_fl_sdesc(sc, fl);
2844 
2845 		if (mtx_initialized(&fl->fl_lock))
2846 			mtx_destroy(&fl->fl_lock);
2847 
2848 		bzero(fl, sizeof(*fl));
2849 	}
2850 
2851 	return (0);
2852 }
2853 
2854 static void
2855 add_fl_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
2856     struct sge_fl *fl)
2857 {
2858 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2859 
2860 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
2861 	    "freelist");
2862 	children = SYSCTL_CHILDREN(oid);
2863 
2864 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
2865 	    CTLTYPE_INT | CTLFLAG_RD, &fl->cntxt_id, 0, sysctl_uint16, "I",
2866 	    "SGE context id of the freelist");
2867 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL,
2868 	    fl_pad ? 1 : 0, "padding enabled");
2869 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL,
2870 	    fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled");
2871 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
2872 	    0, "consumer index");
2873 	if (fl->flags & FL_BUF_PACKING) {
2874 		SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
2875 		    CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
2876 	}
2877 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
2878 	    0, "producer index");
2879 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_allocated",
2880 	    CTLFLAG_RD, &fl->mbuf_allocated, "# of mbuf allocated");
2881 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_inlined",
2882 	    CTLFLAG_RD, &fl->mbuf_inlined, "# of mbuf inlined in clusters");
2883 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
2884 	    CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
2885 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
2886 	    CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
2887 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
2888 	    CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
2889 }
2890 
2891 static int
2892 alloc_fwq(struct adapter *sc)
2893 {
2894 	int rc, intr_idx;
2895 	struct sge_iq *fwq = &sc->sge.fwq;
2896 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
2897 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2898 
2899 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE);
2900 	fwq->flags |= IQ_INTR;	/* always */
2901 	intr_idx = sc->intr_count > 1 ? 1 : 0;
2902 	fwq->set_tcb_rpl = t4_filter_rpl;
2903 	fwq->l2t_write_rpl = do_l2t_write_rpl;
2904 	rc = alloc_iq_fl(&sc->port[0]->vi[0], fwq, NULL, intr_idx, -1);
2905 	if (rc != 0) {
2906 		device_printf(sc->dev,
2907 		    "failed to create firmware event queue: %d\n", rc);
2908 		return (rc);
2909 	}
2910 
2911 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD,
2912 	    NULL, "firmware event queue");
2913 	children = SYSCTL_CHILDREN(oid);
2914 
2915 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "abs_id",
2916 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->abs_id, 0, sysctl_uint16, "I",
2917 	    "absolute id of the queue");
2918 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cntxt_id",
2919 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->cntxt_id, 0, sysctl_uint16, "I",
2920 	    "SGE context id of the queue");
2921 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cidx",
2922 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->cidx, 0, sysctl_uint16, "I",
2923 	    "consumer index");
2924 
2925 	return (0);
2926 }
2927 
2928 static int
2929 free_fwq(struct adapter *sc)
2930 {
2931 	return free_iq_fl(NULL, &sc->sge.fwq, NULL);
2932 }
2933 
2934 static int
2935 alloc_mgmtq(struct adapter *sc)
2936 {
2937 	int rc;
2938 	struct sge_wrq *mgmtq = &sc->sge.mgmtq;
2939 	char name[16];
2940 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
2941 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2942 
2943 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "mgmtq", CTLFLAG_RD,
2944 	    NULL, "management queue");
2945 
2946 	snprintf(name, sizeof(name), "%s mgmtq", device_get_nameunit(sc->dev));
2947 	init_eq(sc, &mgmtq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[0]->tx_chan,
2948 	    sc->sge.fwq.cntxt_id, name);
2949 	rc = alloc_wrq(sc, NULL, mgmtq, oid);
2950 	if (rc != 0) {
2951 		device_printf(sc->dev,
2952 		    "failed to create management queue: %d\n", rc);
2953 		return (rc);
2954 	}
2955 
2956 	return (0);
2957 }
2958 
2959 static int
2960 free_mgmtq(struct adapter *sc)
2961 {
2962 
2963 	return free_wrq(sc, &sc->sge.mgmtq);
2964 }
2965 
2966 int
2967 tnl_cong(struct port_info *pi, int drop)
2968 {
2969 
2970 	if (drop == -1)
2971 		return (-1);
2972 	else if (drop == 1)
2973 		return (0);
2974 	else
2975 		return (pi->rx_chan_map);
2976 }
2977 
2978 static int
2979 alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int intr_idx, int idx,
2980     struct sysctl_oid *oid)
2981 {
2982 	int rc;
2983 	struct adapter *sc = vi->pi->adapter;
2984 	struct sysctl_oid_list *children;
2985 	char name[16];
2986 
2987 	rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, intr_idx,
2988 	    tnl_cong(vi->pi, cong_drop));
2989 	if (rc != 0)
2990 		return (rc);
2991 
2992 	if (idx == 0)
2993 		sc->sge.iq_base = rxq->iq.abs_id - rxq->iq.cntxt_id;
2994 	else
2995 		KASSERT(rxq->iq.cntxt_id + sc->sge.iq_base == rxq->iq.abs_id,
2996 		    ("iq_base mismatch"));
2997 	KASSERT(sc->sge.iq_base == 0 || sc->flags & IS_VF,
2998 	    ("PF with non-zero iq_base"));
2999 
3000 	/*
3001 	 * The freelist is just barely above the starvation threshold right now,
3002 	 * fill it up a bit more.
3003 	 */
3004 	FL_LOCK(&rxq->fl);
3005 	refill_fl(sc, &rxq->fl, 128);
3006 	FL_UNLOCK(&rxq->fl);
3007 
3008 #if defined(INET) || defined(INET6)
3009 	rc = tcp_lro_init(&rxq->lro);
3010 	if (rc != 0)
3011 		return (rc);
3012 	rxq->lro.ifp = vi->ifp; /* also indicates LRO init'ed */
3013 
3014 	if (vi->ifp->if_capenable & IFCAP_LRO)
3015 		rxq->iq.flags |= IQ_LRO_ENABLED;
3016 #endif
3017 	rxq->ifp = vi->ifp;
3018 
3019 	children = SYSCTL_CHILDREN(oid);
3020 
3021 	snprintf(name, sizeof(name), "%d", idx);
3022 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3023 	    NULL, "rx queue");
3024 	children = SYSCTL_CHILDREN(oid);
3025 
3026 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "abs_id",
3027 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.abs_id, 0, sysctl_uint16, "I",
3028 	    "absolute id of the queue");
3029 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cntxt_id",
3030 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cntxt_id, 0, sysctl_uint16, "I",
3031 	    "SGE context id of the queue");
3032 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3033 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cidx, 0, sysctl_uint16, "I",
3034 	    "consumer index");
3035 #if defined(INET) || defined(INET6)
3036 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
3037 	    &rxq->lro.lro_queued, 0, NULL);
3038 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
3039 	    &rxq->lro.lro_flushed, 0, NULL);
3040 #endif
3041 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
3042 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
3043 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_extraction",
3044 	    CTLFLAG_RD, &rxq->vlan_extraction,
3045 	    "# of times hardware extracted 802.1Q tag");
3046 
3047 	add_fl_sysctls(&vi->ctx, oid, &rxq->fl);
3048 
3049 	return (rc);
3050 }
3051 
3052 static int
3053 free_rxq(struct vi_info *vi, struct sge_rxq *rxq)
3054 {
3055 	int rc;
3056 
3057 #if defined(INET) || defined(INET6)
3058 	if (rxq->lro.ifp) {
3059 		tcp_lro_free(&rxq->lro);
3060 		rxq->lro.ifp = NULL;
3061 	}
3062 #endif
3063 
3064 	rc = free_iq_fl(vi, &rxq->iq, &rxq->fl);
3065 	if (rc == 0)
3066 		bzero(rxq, sizeof(*rxq));
3067 
3068 	return (rc);
3069 }
3070 
3071 #ifdef TCP_OFFLOAD
3072 static int
3073 alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq,
3074     int intr_idx, int idx, struct sysctl_oid *oid)
3075 {
3076 	int rc;
3077 	struct sysctl_oid_list *children;
3078 	char name[16];
3079 
3080 	rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx,
3081 	    vi->pi->rx_chan_map);
3082 	if (rc != 0)
3083 		return (rc);
3084 
3085 	children = SYSCTL_CHILDREN(oid);
3086 
3087 	snprintf(name, sizeof(name), "%d", idx);
3088 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3089 	    NULL, "rx queue");
3090 	children = SYSCTL_CHILDREN(oid);
3091 
3092 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "abs_id",
3093 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.abs_id, 0, sysctl_uint16,
3094 	    "I", "absolute id of the queue");
3095 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cntxt_id",
3096 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cntxt_id, 0, sysctl_uint16,
3097 	    "I", "SGE context id of the queue");
3098 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3099 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cidx, 0, sysctl_uint16, "I",
3100 	    "consumer index");
3101 
3102 	add_fl_sysctls(&vi->ctx, oid, &ofld_rxq->fl);
3103 
3104 	return (rc);
3105 }
3106 
3107 static int
3108 free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq)
3109 {
3110 	int rc;
3111 
3112 	rc = free_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl);
3113 	if (rc == 0)
3114 		bzero(ofld_rxq, sizeof(*ofld_rxq));
3115 
3116 	return (rc);
3117 }
3118 #endif
3119 
3120 #ifdef DEV_NETMAP
3121 static int
3122 alloc_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int intr_idx,
3123     int idx, struct sysctl_oid *oid)
3124 {
3125 	int rc;
3126 	struct sysctl_oid_list *children;
3127 	struct sysctl_ctx_list *ctx;
3128 	char name[16];
3129 	size_t len;
3130 	struct adapter *sc = vi->pi->adapter;
3131 	struct netmap_adapter *na = NA(vi->ifp);
3132 
3133 	MPASS(na != NULL);
3134 
3135 	len = vi->qsize_rxq * IQ_ESIZE;
3136 	rc = alloc_ring(sc, len, &nm_rxq->iq_desc_tag, &nm_rxq->iq_desc_map,
3137 	    &nm_rxq->iq_ba, (void **)&nm_rxq->iq_desc);
3138 	if (rc != 0)
3139 		return (rc);
3140 
3141 	len = na->num_rx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3142 	rc = alloc_ring(sc, len, &nm_rxq->fl_desc_tag, &nm_rxq->fl_desc_map,
3143 	    &nm_rxq->fl_ba, (void **)&nm_rxq->fl_desc);
3144 	if (rc != 0)
3145 		return (rc);
3146 
3147 	nm_rxq->vi = vi;
3148 	nm_rxq->nid = idx;
3149 	nm_rxq->iq_cidx = 0;
3150 	nm_rxq->iq_sidx = vi->qsize_rxq - sc->params.sge.spg_len / IQ_ESIZE;
3151 	nm_rxq->iq_gen = F_RSPD_GEN;
3152 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
3153 	nm_rxq->fl_sidx = na->num_rx_desc;
3154 	nm_rxq->intr_idx = intr_idx;
3155 
3156 	ctx = &vi->ctx;
3157 	children = SYSCTL_CHILDREN(oid);
3158 
3159 	snprintf(name, sizeof(name), "%d", idx);
3160 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL,
3161 	    "rx queue");
3162 	children = SYSCTL_CHILDREN(oid);
3163 
3164 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3165 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_abs_id, 0, sysctl_uint16,
3166 	    "I", "absolute id of the queue");
3167 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3168 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cntxt_id, 0, sysctl_uint16,
3169 	    "I", "SGE context id of the queue");
3170 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3171 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cidx, 0, sysctl_uint16, "I",
3172 	    "consumer index");
3173 
3174 	children = SYSCTL_CHILDREN(oid);
3175 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3176 	    "freelist");
3177 	children = SYSCTL_CHILDREN(oid);
3178 
3179 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3180 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->fl_cntxt_id, 0, sysctl_uint16,
3181 	    "I", "SGE context id of the freelist");
3182 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
3183 	    &nm_rxq->fl_cidx, 0, "consumer index");
3184 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
3185 	    &nm_rxq->fl_pidx, 0, "producer index");
3186 
3187 	return (rc);
3188 }
3189 
3190 
3191 static int
3192 free_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
3193 {
3194 	struct adapter *sc = vi->pi->adapter;
3195 
3196 	free_ring(sc, nm_rxq->iq_desc_tag, nm_rxq->iq_desc_map, nm_rxq->iq_ba,
3197 	    nm_rxq->iq_desc);
3198 	free_ring(sc, nm_rxq->fl_desc_tag, nm_rxq->fl_desc_map, nm_rxq->fl_ba,
3199 	    nm_rxq->fl_desc);
3200 
3201 	return (0);
3202 }
3203 
3204 static int
3205 alloc_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq, int iqidx, int idx,
3206     struct sysctl_oid *oid)
3207 {
3208 	int rc;
3209 	size_t len;
3210 	struct port_info *pi = vi->pi;
3211 	struct adapter *sc = pi->adapter;
3212 	struct netmap_adapter *na = NA(vi->ifp);
3213 	char name[16];
3214 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3215 
3216 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3217 	rc = alloc_ring(sc, len, &nm_txq->desc_tag, &nm_txq->desc_map,
3218 	    &nm_txq->ba, (void **)&nm_txq->desc);
3219 	if (rc)
3220 		return (rc);
3221 
3222 	nm_txq->pidx = nm_txq->cidx = 0;
3223 	nm_txq->sidx = na->num_tx_desc;
3224 	nm_txq->nid = idx;
3225 	nm_txq->iqidx = iqidx;
3226 	nm_txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3227 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_VF_VLD(1) |
3228 	    V_TXPKT_VF(vi->viid));
3229 
3230 	snprintf(name, sizeof(name), "%d", idx);
3231 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3232 	    NULL, "netmap tx queue");
3233 	children = SYSCTL_CHILDREN(oid);
3234 
3235 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3236 	    &nm_txq->cntxt_id, 0, "SGE context id of the queue");
3237 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3238 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->cidx, 0, sysctl_uint16, "I",
3239 	    "consumer index");
3240 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3241 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->pidx, 0, sysctl_uint16, "I",
3242 	    "producer index");
3243 
3244 	return (rc);
3245 }
3246 
3247 static int
3248 free_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
3249 {
3250 	struct adapter *sc = vi->pi->adapter;
3251 
3252 	free_ring(sc, nm_txq->desc_tag, nm_txq->desc_map, nm_txq->ba,
3253 	    nm_txq->desc);
3254 
3255 	return (0);
3256 }
3257 #endif
3258 
3259 static int
3260 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
3261 {
3262 	int rc, cntxt_id;
3263 	struct fw_eq_ctrl_cmd c;
3264 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3265 
3266 	bzero(&c, sizeof(c));
3267 
3268 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
3269 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
3270 	    V_FW_EQ_CTRL_CMD_VFN(0));
3271 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
3272 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
3273 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid));
3274 	c.physeqid_pkd = htobe32(0);
3275 	c.fetchszm_to_iqid =
3276 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3277 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
3278 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
3279 	c.dcaen_to_eqsize =
3280 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3281 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3282 		V_FW_EQ_CTRL_CMD_EQSIZE(qsize));
3283 	c.eqaddr = htobe64(eq->ba);
3284 
3285 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3286 	if (rc != 0) {
3287 		device_printf(sc->dev,
3288 		    "failed to create control queue %d: %d\n", eq->tx_chan, rc);
3289 		return (rc);
3290 	}
3291 	eq->flags |= EQ_ALLOCATED;
3292 
3293 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
3294 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3295 	if (cntxt_id >= sc->sge.neq)
3296 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3297 		cntxt_id, sc->sge.neq - 1);
3298 	sc->sge.eqmap[cntxt_id] = eq;
3299 
3300 	return (rc);
3301 }
3302 
3303 static int
3304 eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3305 {
3306 	int rc, cntxt_id;
3307 	struct fw_eq_eth_cmd c;
3308 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3309 
3310 	bzero(&c, sizeof(c));
3311 
3312 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
3313 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
3314 	    V_FW_EQ_ETH_CMD_VFN(0));
3315 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
3316 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
3317 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
3318 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
3319 	c.fetchszm_to_iqid =
3320 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3321 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
3322 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
3323 	c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3324 	    V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3325 	    V_FW_EQ_ETH_CMD_EQSIZE(qsize));
3326 	c.eqaddr = htobe64(eq->ba);
3327 
3328 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3329 	if (rc != 0) {
3330 		device_printf(vi->dev,
3331 		    "failed to create Ethernet egress queue: %d\n", rc);
3332 		return (rc);
3333 	}
3334 	eq->flags |= EQ_ALLOCATED;
3335 
3336 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
3337 	eq->abs_id = G_FW_EQ_ETH_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
3338 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3339 	if (cntxt_id >= sc->sge.neq)
3340 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3341 		cntxt_id, sc->sge.neq - 1);
3342 	sc->sge.eqmap[cntxt_id] = eq;
3343 
3344 	return (rc);
3345 }
3346 
3347 #ifdef TCP_OFFLOAD
3348 static int
3349 ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3350 {
3351 	int rc, cntxt_id;
3352 	struct fw_eq_ofld_cmd c;
3353 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3354 
3355 	bzero(&c, sizeof(c));
3356 
3357 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
3358 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
3359 	    V_FW_EQ_OFLD_CMD_VFN(0));
3360 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
3361 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
3362 	c.fetchszm_to_iqid =
3363 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3364 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
3365 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
3366 	c.dcaen_to_eqsize =
3367 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3368 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3369 		V_FW_EQ_OFLD_CMD_EQSIZE(qsize));
3370 	c.eqaddr = htobe64(eq->ba);
3371 
3372 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3373 	if (rc != 0) {
3374 		device_printf(vi->dev,
3375 		    "failed to create egress queue for TCP offload: %d\n", rc);
3376 		return (rc);
3377 	}
3378 	eq->flags |= EQ_ALLOCATED;
3379 
3380 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
3381 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3382 	if (cntxt_id >= sc->sge.neq)
3383 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3384 		cntxt_id, sc->sge.neq - 1);
3385 	sc->sge.eqmap[cntxt_id] = eq;
3386 
3387 	return (rc);
3388 }
3389 #endif
3390 
3391 static int
3392 alloc_eq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3393 {
3394 	int rc, qsize;
3395 	size_t len;
3396 
3397 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
3398 
3399 	qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3400 	len = qsize * EQ_ESIZE;
3401 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
3402 	    &eq->ba, (void **)&eq->desc);
3403 	if (rc)
3404 		return (rc);
3405 
3406 	eq->pidx = eq->cidx = 0;
3407 	eq->equeqidx = eq->dbidx = 0;
3408 	eq->doorbells = sc->doorbells;
3409 
3410 	switch (eq->flags & EQ_TYPEMASK) {
3411 	case EQ_CTRL:
3412 		rc = ctrl_eq_alloc(sc, eq);
3413 		break;
3414 
3415 	case EQ_ETH:
3416 		rc = eth_eq_alloc(sc, vi, eq);
3417 		break;
3418 
3419 #ifdef TCP_OFFLOAD
3420 	case EQ_OFLD:
3421 		rc = ofld_eq_alloc(sc, vi, eq);
3422 		break;
3423 #endif
3424 
3425 	default:
3426 		panic("%s: invalid eq type %d.", __func__,
3427 		    eq->flags & EQ_TYPEMASK);
3428 	}
3429 	if (rc != 0) {
3430 		device_printf(sc->dev,
3431 		    "failed to allocate egress queue(%d): %d\n",
3432 		    eq->flags & EQ_TYPEMASK, rc);
3433 	}
3434 
3435 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
3436 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
3437 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
3438 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3439 		uint32_t mask = (1 << s_qpp) - 1;
3440 		volatile uint8_t *udb;
3441 
3442 		udb = sc->udbs_base + UDBS_DB_OFFSET;
3443 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
3444 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
3445 		if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
3446 	    		clrbit(&eq->doorbells, DOORBELL_WCWR);
3447 		else {
3448 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
3449 			eq->udb_qid = 0;
3450 		}
3451 		eq->udb = (volatile void *)udb;
3452 	}
3453 
3454 	return (rc);
3455 }
3456 
3457 static int
3458 free_eq(struct adapter *sc, struct sge_eq *eq)
3459 {
3460 	int rc;
3461 
3462 	if (eq->flags & EQ_ALLOCATED) {
3463 		switch (eq->flags & EQ_TYPEMASK) {
3464 		case EQ_CTRL:
3465 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
3466 			    eq->cntxt_id);
3467 			break;
3468 
3469 		case EQ_ETH:
3470 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
3471 			    eq->cntxt_id);
3472 			break;
3473 
3474 #ifdef TCP_OFFLOAD
3475 		case EQ_OFLD:
3476 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
3477 			    eq->cntxt_id);
3478 			break;
3479 #endif
3480 
3481 		default:
3482 			panic("%s: invalid eq type %d.", __func__,
3483 			    eq->flags & EQ_TYPEMASK);
3484 		}
3485 		if (rc != 0) {
3486 			device_printf(sc->dev,
3487 			    "failed to free egress queue (%d): %d\n",
3488 			    eq->flags & EQ_TYPEMASK, rc);
3489 			return (rc);
3490 		}
3491 		eq->flags &= ~EQ_ALLOCATED;
3492 	}
3493 
3494 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
3495 
3496 	if (mtx_initialized(&eq->eq_lock))
3497 		mtx_destroy(&eq->eq_lock);
3498 
3499 	bzero(eq, sizeof(*eq));
3500 	return (0);
3501 }
3502 
3503 static int
3504 alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq,
3505     struct sysctl_oid *oid)
3506 {
3507 	int rc;
3508 	struct sysctl_ctx_list *ctx = vi ? &vi->ctx : &sc->ctx;
3509 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3510 
3511 	rc = alloc_eq(sc, vi, &wrq->eq);
3512 	if (rc)
3513 		return (rc);
3514 
3515 	wrq->adapter = sc;
3516 	TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq);
3517 	TAILQ_INIT(&wrq->incomplete_wrs);
3518 	STAILQ_INIT(&wrq->wr_list);
3519 	wrq->nwr_pending = 0;
3520 	wrq->ndesc_needed = 0;
3521 
3522 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3523 	    &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
3524 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3525 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I",
3526 	    "consumer index");
3527 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx",
3528 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I",
3529 	    "producer index");
3530 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD,
3531 	    &wrq->tx_wrs_direct, "# of work requests (direct)");
3532 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD,
3533 	    &wrq->tx_wrs_copied, "# of work requests (copied)");
3534 
3535 	return (rc);
3536 }
3537 
3538 static int
3539 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
3540 {
3541 	int rc;
3542 
3543 	rc = free_eq(sc, &wrq->eq);
3544 	if (rc)
3545 		return (rc);
3546 
3547 	bzero(wrq, sizeof(*wrq));
3548 	return (0);
3549 }
3550 
3551 static int
3552 alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx,
3553     struct sysctl_oid *oid)
3554 {
3555 	int rc;
3556 	struct port_info *pi = vi->pi;
3557 	struct adapter *sc = pi->adapter;
3558 	struct sge_eq *eq = &txq->eq;
3559 	char name[16];
3560 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3561 
3562 	rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx, can_resume_eth_tx,
3563 	    M_CXGBE, M_WAITOK);
3564 	if (rc != 0) {
3565 		device_printf(sc->dev, "failed to allocate mp_ring: %d\n", rc);
3566 		return (rc);
3567 	}
3568 
3569 	rc = alloc_eq(sc, vi, eq);
3570 	if (rc != 0) {
3571 		mp_ring_free(txq->r);
3572 		txq->r = NULL;
3573 		return (rc);
3574 	}
3575 
3576 	/* Can't fail after this point. */
3577 
3578 	if (idx == 0)
3579 		sc->sge.eq_base = eq->abs_id - eq->cntxt_id;
3580 	else
3581 		KASSERT(eq->cntxt_id + sc->sge.eq_base == eq->abs_id,
3582 		    ("eq_base mismatch"));
3583 	KASSERT(sc->sge.eq_base == 0 || sc->flags & IS_VF,
3584 	    ("PF with non-zero eq_base"));
3585 
3586 	TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq);
3587 	txq->ifp = vi->ifp;
3588 	txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK);
3589 	txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3590 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_VF_VLD(1) |
3591 	    V_TXPKT_VF(vi->viid));
3592 	txq->tc_idx = -1;
3593 	txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE,
3594 	    M_ZERO | M_WAITOK);
3595 
3596 	snprintf(name, sizeof(name), "%d", idx);
3597 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3598 	    NULL, "tx queue");
3599 	children = SYSCTL_CHILDREN(oid);
3600 
3601 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
3602 	    &eq->abs_id, 0, "absolute id of the queue");
3603 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3604 	    &eq->cntxt_id, 0, "SGE context id of the queue");
3605 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3606 	    CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I",
3607 	    "consumer index");
3608 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3609 	    CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I",
3610 	    "producer index");
3611 
3612 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "tc",
3613 	    CTLTYPE_INT | CTLFLAG_RW, vi, idx, sysctl_tc, "I",
3614 	    "traffic class (-1 means none)");
3615 
3616 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
3617 	    &txq->txcsum, "# of times hardware assisted with checksum");
3618 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_insertion",
3619 	    CTLFLAG_RD, &txq->vlan_insertion,
3620 	    "# of times hardware inserted 802.1Q tag");
3621 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
3622 	    &txq->tso_wrs, "# of TSO work requests");
3623 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
3624 	    &txq->imm_wrs, "# of work requests with immediate data");
3625 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
3626 	    &txq->sgl_wrs, "# of work requests with direct SGL");
3627 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
3628 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
3629 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_wrs",
3630 	    CTLFLAG_RD, &txq->txpkts0_wrs,
3631 	    "# of txpkts (type 0) work requests");
3632 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_wrs",
3633 	    CTLFLAG_RD, &txq->txpkts1_wrs,
3634 	    "# of txpkts (type 1) work requests");
3635 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_pkts",
3636 	    CTLFLAG_RD, &txq->txpkts0_pkts,
3637 	    "# of frames tx'd using type0 txpkts work requests");
3638 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_pkts",
3639 	    CTLFLAG_RD, &txq->txpkts1_pkts,
3640 	    "# of frames tx'd using type1 txpkts work requests");
3641 
3642 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_enqueues",
3643 	    CTLFLAG_RD, &txq->r->enqueues,
3644 	    "# of enqueues to the mp_ring for this queue");
3645 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_drops",
3646 	    CTLFLAG_RD, &txq->r->drops,
3647 	    "# of drops in the mp_ring for this queue");
3648 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_starts",
3649 	    CTLFLAG_RD, &txq->r->starts,
3650 	    "# of normal consumer starts in the mp_ring for this queue");
3651 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_stalls",
3652 	    CTLFLAG_RD, &txq->r->stalls,
3653 	    "# of consumer stalls in the mp_ring for this queue");
3654 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_restarts",
3655 	    CTLFLAG_RD, &txq->r->restarts,
3656 	    "# of consumer restarts in the mp_ring for this queue");
3657 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_abdications",
3658 	    CTLFLAG_RD, &txq->r->abdications,
3659 	    "# of consumer abdications in the mp_ring for this queue");
3660 
3661 	return (0);
3662 }
3663 
3664 static int
3665 free_txq(struct vi_info *vi, struct sge_txq *txq)
3666 {
3667 	int rc;
3668 	struct adapter *sc = vi->pi->adapter;
3669 	struct sge_eq *eq = &txq->eq;
3670 
3671 	rc = free_eq(sc, eq);
3672 	if (rc)
3673 		return (rc);
3674 
3675 	sglist_free(txq->gl);
3676 	free(txq->sdesc, M_CXGBE);
3677 	mp_ring_free(txq->r);
3678 
3679 	bzero(txq, sizeof(*txq));
3680 	return (0);
3681 }
3682 
3683 static void
3684 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
3685 {
3686 	bus_addr_t *ba = arg;
3687 
3688 	KASSERT(nseg == 1,
3689 	    ("%s meant for single segment mappings only.", __func__));
3690 
3691 	*ba = error ? 0 : segs->ds_addr;
3692 }
3693 
3694 static inline void
3695 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
3696 {
3697 	uint32_t n, v;
3698 
3699 	n = IDXDIFF(fl->pidx / 8, fl->dbidx, fl->sidx);
3700 	MPASS(n > 0);
3701 
3702 	wmb();
3703 	v = fl->dbval | V_PIDX(n);
3704 	if (fl->udb)
3705 		*fl->udb = htole32(v);
3706 	else
3707 		t4_write_reg(sc, sc->sge_kdoorbell_reg, v);
3708 	IDXINCR(fl->dbidx, n, fl->sidx);
3709 }
3710 
3711 /*
3712  * Fills up the freelist by allocating up to 'n' buffers.  Buffers that are
3713  * recycled do not count towards this allocation budget.
3714  *
3715  * Returns non-zero to indicate that this freelist should be added to the list
3716  * of starving freelists.
3717  */
3718 static int
3719 refill_fl(struct adapter *sc, struct sge_fl *fl, int n)
3720 {
3721 	__be64 *d;
3722 	struct fl_sdesc *sd;
3723 	uintptr_t pa;
3724 	caddr_t cl;
3725 	struct cluster_layout *cll;
3726 	struct sw_zone_info *swz;
3727 	struct cluster_metadata *clm;
3728 	uint16_t max_pidx;
3729 	uint16_t hw_cidx = fl->hw_cidx;		/* stable snapshot */
3730 
3731 	FL_LOCK_ASSERT_OWNED(fl);
3732 
3733 	/*
3734 	 * We always stop at the beginning of the hardware descriptor that's just
3735 	 * before the one with the hw cidx.  This is to avoid hw pidx = hw cidx,
3736 	 * which would mean an empty freelist to the chip.
3737 	 */
3738 	max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1;
3739 	if (fl->pidx == max_pidx * 8)
3740 		return (0);
3741 
3742 	d = &fl->desc[fl->pidx];
3743 	sd = &fl->sdesc[fl->pidx];
3744 	cll = &fl->cll_def;	/* default layout */
3745 	swz = &sc->sge.sw_zone_info[cll->zidx];
3746 
3747 	while (n > 0) {
3748 
3749 		if (sd->cl != NULL) {
3750 
3751 			if (sd->nmbuf == 0) {
3752 				/*
3753 				 * Fast recycle without involving any atomics on
3754 				 * the cluster's metadata (if the cluster has
3755 				 * metadata).  This happens when all frames
3756 				 * received in the cluster were small enough to
3757 				 * fit within a single mbuf each.
3758 				 */
3759 				fl->cl_fast_recycled++;
3760 #ifdef INVARIANTS
3761 				clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
3762 				if (clm != NULL)
3763 					MPASS(clm->refcount == 1);
3764 #endif
3765 				goto recycled_fast;
3766 			}
3767 
3768 			/*
3769 			 * Cluster is guaranteed to have metadata.  Clusters
3770 			 * without metadata always take the fast recycle path
3771 			 * when they're recycled.
3772 			 */
3773 			clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
3774 			MPASS(clm != NULL);
3775 
3776 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
3777 				fl->cl_recycled++;
3778 				counter_u64_add(extfree_rels, 1);
3779 				goto recycled;
3780 			}
3781 			sd->cl = NULL;	/* gave up my reference */
3782 		}
3783 		MPASS(sd->cl == NULL);
3784 alloc:
3785 		cl = uma_zalloc(swz->zone, M_NOWAIT);
3786 		if (__predict_false(cl == NULL)) {
3787 			if (cll == &fl->cll_alt || fl->cll_alt.zidx == -1 ||
3788 			    fl->cll_def.zidx == fl->cll_alt.zidx)
3789 				break;
3790 
3791 			/* fall back to the safe zone */
3792 			cll = &fl->cll_alt;
3793 			swz = &sc->sge.sw_zone_info[cll->zidx];
3794 			goto alloc;
3795 		}
3796 		fl->cl_allocated++;
3797 		n--;
3798 
3799 		pa = pmap_kextract((vm_offset_t)cl);
3800 		pa += cll->region1;
3801 		sd->cl = cl;
3802 		sd->cll = *cll;
3803 		*d = htobe64(pa | cll->hwidx);
3804 		clm = cl_metadata(sc, fl, cll, cl);
3805 		if (clm != NULL) {
3806 recycled:
3807 #ifdef INVARIANTS
3808 			clm->sd = sd;
3809 #endif
3810 			clm->refcount = 1;
3811 		}
3812 		sd->nmbuf = 0;
3813 recycled_fast:
3814 		d++;
3815 		sd++;
3816 		if (__predict_false(++fl->pidx % 8 == 0)) {
3817 			uint16_t pidx = fl->pidx / 8;
3818 
3819 			if (__predict_false(pidx == fl->sidx)) {
3820 				fl->pidx = 0;
3821 				pidx = 0;
3822 				sd = fl->sdesc;
3823 				d = fl->desc;
3824 			}
3825 			if (pidx == max_pidx)
3826 				break;
3827 
3828 			if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4)
3829 				ring_fl_db(sc, fl);
3830 		}
3831 	}
3832 
3833 	if (fl->pidx / 8 != fl->dbidx)
3834 		ring_fl_db(sc, fl);
3835 
3836 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
3837 }
3838 
3839 /*
3840  * Attempt to refill all starving freelists.
3841  */
3842 static void
3843 refill_sfl(void *arg)
3844 {
3845 	struct adapter *sc = arg;
3846 	struct sge_fl *fl, *fl_temp;
3847 
3848 	mtx_assert(&sc->sfl_lock, MA_OWNED);
3849 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
3850 		FL_LOCK(fl);
3851 		refill_fl(sc, fl, 64);
3852 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
3853 			TAILQ_REMOVE(&sc->sfl, fl, link);
3854 			fl->flags &= ~FL_STARVING;
3855 		}
3856 		FL_UNLOCK(fl);
3857 	}
3858 
3859 	if (!TAILQ_EMPTY(&sc->sfl))
3860 		callout_schedule(&sc->sfl_callout, hz / 5);
3861 }
3862 
3863 static int
3864 alloc_fl_sdesc(struct sge_fl *fl)
3865 {
3866 
3867 	fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc), M_CXGBE,
3868 	    M_ZERO | M_WAITOK);
3869 
3870 	return (0);
3871 }
3872 
3873 static void
3874 free_fl_sdesc(struct adapter *sc, struct sge_fl *fl)
3875 {
3876 	struct fl_sdesc *sd;
3877 	struct cluster_metadata *clm;
3878 	struct cluster_layout *cll;
3879 	int i;
3880 
3881 	sd = fl->sdesc;
3882 	for (i = 0; i < fl->sidx * 8; i++, sd++) {
3883 		if (sd->cl == NULL)
3884 			continue;
3885 
3886 		cll = &sd->cll;
3887 		clm = cl_metadata(sc, fl, cll, sd->cl);
3888 		if (sd->nmbuf == 0)
3889 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
3890 		else if (clm && atomic_fetchadd_int(&clm->refcount, -1) == 1) {
3891 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
3892 			counter_u64_add(extfree_rels, 1);
3893 		}
3894 		sd->cl = NULL;
3895 	}
3896 
3897 	free(fl->sdesc, M_CXGBE);
3898 	fl->sdesc = NULL;
3899 }
3900 
3901 static inline void
3902 get_pkt_gl(struct mbuf *m, struct sglist *gl)
3903 {
3904 	int rc;
3905 
3906 	M_ASSERTPKTHDR(m);
3907 
3908 	sglist_reset(gl);
3909 	rc = sglist_append_mbuf(gl, m);
3910 	if (__predict_false(rc != 0)) {
3911 		panic("%s: mbuf %p (%d segs) was vetted earlier but now fails "
3912 		    "with %d.", __func__, m, mbuf_nsegs(m), rc);
3913 	}
3914 
3915 	KASSERT(gl->sg_nseg == mbuf_nsegs(m),
3916 	    ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m,
3917 	    mbuf_nsegs(m), gl->sg_nseg));
3918 	KASSERT(gl->sg_nseg > 0 &&
3919 	    gl->sg_nseg <= (needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS),
3920 	    ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__,
3921 		gl->sg_nseg, needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS));
3922 }
3923 
3924 /*
3925  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
3926  */
3927 static inline u_int
3928 txpkt_len16(u_int nsegs, u_int tso)
3929 {
3930 	u_int n;
3931 
3932 	MPASS(nsegs > 0);
3933 
3934 	nsegs--; /* first segment is part of ulptx_sgl */
3935 	n = sizeof(struct fw_eth_tx_pkt_wr) + sizeof(struct cpl_tx_pkt_core) +
3936 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
3937 	if (tso)
3938 		n += sizeof(struct cpl_tx_pkt_lso_core);
3939 
3940 	return (howmany(n, 16));
3941 }
3942 
3943 /*
3944  * len16 for a txpkts type 0 WR with a GL.  Does not include the firmware work
3945  * request header.
3946  */
3947 static inline u_int
3948 txpkts0_len16(u_int nsegs)
3949 {
3950 	u_int n;
3951 
3952 	MPASS(nsegs > 0);
3953 
3954 	nsegs--; /* first segment is part of ulptx_sgl */
3955 	n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) +
3956 	    sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) +
3957 	    8 * ((3 * nsegs) / 2 + (nsegs & 1));
3958 
3959 	return (howmany(n, 16));
3960 }
3961 
3962 /*
3963  * len16 for a txpkts type 1 WR with a GL.  Does not include the firmware work
3964  * request header.
3965  */
3966 static inline u_int
3967 txpkts1_len16(void)
3968 {
3969 	u_int n;
3970 
3971 	n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl);
3972 
3973 	return (howmany(n, 16));
3974 }
3975 
3976 static inline u_int
3977 imm_payload(u_int ndesc)
3978 {
3979 	u_int n;
3980 
3981 	n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) -
3982 	    sizeof(struct cpl_tx_pkt_core);
3983 
3984 	return (n);
3985 }
3986 
3987 /*
3988  * Write a txpkt WR for this packet to the hardware descriptors, update the
3989  * software descriptor, and advance the pidx.  It is guaranteed that enough
3990  * descriptors are available.
3991  *
3992  * The return value is the # of hardware descriptors used.
3993  */
3994 static u_int
3995 write_txpkt_wr(struct sge_txq *txq, struct fw_eth_tx_pkt_wr *wr,
3996     struct mbuf *m0, u_int available)
3997 {
3998 	struct sge_eq *eq = &txq->eq;
3999 	struct tx_sdesc *txsd;
4000 	struct cpl_tx_pkt_core *cpl;
4001 	uint32_t ctrl;	/* used in many unrelated places */
4002 	uint64_t ctrl1;
4003 	int len16, ndesc, pktlen, nsegs;
4004 	caddr_t dst;
4005 
4006 	TXQ_LOCK_ASSERT_OWNED(txq);
4007 	M_ASSERTPKTHDR(m0);
4008 	MPASS(available > 0 && available < eq->sidx);
4009 
4010 	len16 = mbuf_len16(m0);
4011 	nsegs = mbuf_nsegs(m0);
4012 	pktlen = m0->m_pkthdr.len;
4013 	ctrl = sizeof(struct cpl_tx_pkt_core);
4014 	if (needs_tso(m0))
4015 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4016 	else if (pktlen <= imm_payload(2) && available >= 2) {
4017 		/* Immediate data.  Recalculate len16 and set nsegs to 0. */
4018 		ctrl += pktlen;
4019 		len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) +
4020 		    sizeof(struct cpl_tx_pkt_core) + pktlen, 16);
4021 		nsegs = 0;
4022 	}
4023 	ndesc = howmany(len16, EQ_ESIZE / 16);
4024 	MPASS(ndesc <= available);
4025 
4026 	/* Firmware work request header */
4027 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4028 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
4029 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4030 
4031 	ctrl = V_FW_WR_LEN16(len16);
4032 	wr->equiq_to_len16 = htobe32(ctrl);
4033 	wr->r3 = 0;
4034 
4035 	if (needs_tso(m0)) {
4036 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4037 
4038 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4039 		    m0->m_pkthdr.l4hlen > 0,
4040 		    ("%s: mbuf %p needs TSO but missing header lengths",
4041 			__func__, m0));
4042 
4043 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4044 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4045 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4046 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4047 			ctrl |= V_LSO_ETHHDR_LEN(1);
4048 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4049 			ctrl |= F_LSO_IPV6;
4050 
4051 		lso->lso_ctrl = htobe32(ctrl);
4052 		lso->ipid_ofst = htobe16(0);
4053 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4054 		lso->seqno_offset = htobe32(0);
4055 		lso->len = htobe32(pktlen);
4056 
4057 		cpl = (void *)(lso + 1);
4058 
4059 		txq->tso_wrs++;
4060 	} else
4061 		cpl = (void *)(wr + 1);
4062 
4063 	/* Checksum offload */
4064 	ctrl1 = 0;
4065 	if (needs_l3_csum(m0) == 0)
4066 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4067 	if (needs_l4_csum(m0) == 0)
4068 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4069 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4070 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4071 		txq->txcsum++;	/* some hardware assistance provided */
4072 
4073 	/* VLAN tag insertion */
4074 	if (needs_vlan_insertion(m0)) {
4075 		ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4076 		txq->vlan_insertion++;
4077 	}
4078 
4079 	/* CPL header */
4080 	cpl->ctrl0 = txq->cpl_ctrl0;
4081 	cpl->pack = 0;
4082 	cpl->len = htobe16(pktlen);
4083 	cpl->ctrl1 = htobe64(ctrl1);
4084 
4085 	/* SGL */
4086 	dst = (void *)(cpl + 1);
4087 	if (nsegs > 0) {
4088 
4089 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4090 		txq->sgl_wrs++;
4091 	} else {
4092 		struct mbuf *m;
4093 
4094 		for (m = m0; m != NULL; m = m->m_next) {
4095 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4096 #ifdef INVARIANTS
4097 			pktlen -= m->m_len;
4098 #endif
4099 		}
4100 #ifdef INVARIANTS
4101 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
4102 #endif
4103 		txq->imm_wrs++;
4104 	}
4105 
4106 	txq->txpkt_wrs++;
4107 
4108 	txsd = &txq->sdesc[eq->pidx];
4109 	txsd->m = m0;
4110 	txsd->desc_used = ndesc;
4111 
4112 	return (ndesc);
4113 }
4114 
4115 static int
4116 try_txpkts(struct mbuf *m, struct mbuf *n, struct txpkts *txp, u_int available)
4117 {
4118 	u_int needed, nsegs1, nsegs2, l1, l2;
4119 
4120 	if (cannot_use_txpkts(m) || cannot_use_txpkts(n))
4121 		return (1);
4122 
4123 	nsegs1 = mbuf_nsegs(m);
4124 	nsegs2 = mbuf_nsegs(n);
4125 	if (nsegs1 + nsegs2 == 2) {
4126 		txp->wr_type = 1;
4127 		l1 = l2 = txpkts1_len16();
4128 	} else {
4129 		txp->wr_type = 0;
4130 		l1 = txpkts0_len16(nsegs1);
4131 		l2 = txpkts0_len16(nsegs2);
4132 	}
4133 	txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) + l1 + l2;
4134 	needed = howmany(txp->len16, EQ_ESIZE / 16);
4135 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4136 		return (1);
4137 
4138 	txp->plen = m->m_pkthdr.len + n->m_pkthdr.len;
4139 	if (txp->plen > 65535)
4140 		return (1);
4141 
4142 	txp->npkt = 2;
4143 	set_mbuf_len16(m, l1);
4144 	set_mbuf_len16(n, l2);
4145 
4146 	return (0);
4147 }
4148 
4149 static int
4150 add_to_txpkts(struct mbuf *m, struct txpkts *txp, u_int available)
4151 {
4152 	u_int plen, len16, needed, nsegs;
4153 
4154 	MPASS(txp->wr_type == 0 || txp->wr_type == 1);
4155 
4156 	nsegs = mbuf_nsegs(m);
4157 	if (needs_tso(m) || (txp->wr_type == 1 && nsegs != 1))
4158 		return (1);
4159 
4160 	plen = txp->plen + m->m_pkthdr.len;
4161 	if (plen > 65535)
4162 		return (1);
4163 
4164 	if (txp->wr_type == 0)
4165 		len16 = txpkts0_len16(nsegs);
4166 	else
4167 		len16 = txpkts1_len16();
4168 	needed = howmany(txp->len16 + len16, EQ_ESIZE / 16);
4169 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4170 		return (1);
4171 
4172 	txp->npkt++;
4173 	txp->plen = plen;
4174 	txp->len16 += len16;
4175 	set_mbuf_len16(m, len16);
4176 
4177 	return (0);
4178 }
4179 
4180 /*
4181  * Write a txpkts WR for the packets in txp to the hardware descriptors, update
4182  * the software descriptor, and advance the pidx.  It is guaranteed that enough
4183  * descriptors are available.
4184  *
4185  * The return value is the # of hardware descriptors used.
4186  */
4187 static u_int
4188 write_txpkts_wr(struct sge_txq *txq, struct fw_eth_tx_pkts_wr *wr,
4189     struct mbuf *m0, const struct txpkts *txp, u_int available)
4190 {
4191 	struct sge_eq *eq = &txq->eq;
4192 	struct tx_sdesc *txsd;
4193 	struct cpl_tx_pkt_core *cpl;
4194 	uint32_t ctrl;
4195 	uint64_t ctrl1;
4196 	int ndesc, checkwrap;
4197 	struct mbuf *m;
4198 	void *flitp;
4199 
4200 	TXQ_LOCK_ASSERT_OWNED(txq);
4201 	MPASS(txp->npkt > 0);
4202 	MPASS(txp->plen < 65536);
4203 	MPASS(m0 != NULL);
4204 	MPASS(m0->m_nextpkt != NULL);
4205 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
4206 	MPASS(available > 0 && available < eq->sidx);
4207 
4208 	ndesc = howmany(txp->len16, EQ_ESIZE / 16);
4209 	MPASS(ndesc <= available);
4210 
4211 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4212 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
4213 	ctrl = V_FW_WR_LEN16(txp->len16);
4214 	wr->equiq_to_len16 = htobe32(ctrl);
4215 	wr->plen = htobe16(txp->plen);
4216 	wr->npkt = txp->npkt;
4217 	wr->r3 = 0;
4218 	wr->type = txp->wr_type;
4219 	flitp = wr + 1;
4220 
4221 	/*
4222 	 * At this point we are 16B into a hardware descriptor.  If checkwrap is
4223 	 * set then we know the WR is going to wrap around somewhere.  We'll
4224 	 * check for that at appropriate points.
4225 	 */
4226 	checkwrap = eq->sidx - ndesc < eq->pidx;
4227 	for (m = m0; m != NULL; m = m->m_nextpkt) {
4228 		if (txp->wr_type == 0) {
4229 			struct ulp_txpkt *ulpmc;
4230 			struct ulptx_idata *ulpsc;
4231 
4232 			/* ULP master command */
4233 			ulpmc = flitp;
4234 			ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
4235 			    V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid));
4236 			ulpmc->len = htobe32(mbuf_len16(m));
4237 
4238 			/* ULP subcommand */
4239 			ulpsc = (void *)(ulpmc + 1);
4240 			ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
4241 			    F_ULP_TX_SC_MORE);
4242 			ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
4243 
4244 			cpl = (void *)(ulpsc + 1);
4245 			if (checkwrap &&
4246 			    (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx])
4247 				cpl = (void *)&eq->desc[0];
4248 			txq->txpkts0_pkts += txp->npkt;
4249 			txq->txpkts0_wrs++;
4250 		} else {
4251 			cpl = flitp;
4252 			txq->txpkts1_pkts += txp->npkt;
4253 			txq->txpkts1_wrs++;
4254 		}
4255 
4256 		/* Checksum offload */
4257 		ctrl1 = 0;
4258 		if (needs_l3_csum(m) == 0)
4259 			ctrl1 |= F_TXPKT_IPCSUM_DIS;
4260 		if (needs_l4_csum(m) == 0)
4261 			ctrl1 |= F_TXPKT_L4CSUM_DIS;
4262 		if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4263 		    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4264 			txq->txcsum++;	/* some hardware assistance provided */
4265 
4266 		/* VLAN tag insertion */
4267 		if (needs_vlan_insertion(m)) {
4268 			ctrl1 |= F_TXPKT_VLAN_VLD |
4269 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
4270 			txq->vlan_insertion++;
4271 		}
4272 
4273 		/* CPL header */
4274 		cpl->ctrl0 = txq->cpl_ctrl0;
4275 		cpl->pack = 0;
4276 		cpl->len = htobe16(m->m_pkthdr.len);
4277 		cpl->ctrl1 = htobe64(ctrl1);
4278 
4279 		flitp = cpl + 1;
4280 		if (checkwrap &&
4281 		    (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
4282 			flitp = (void *)&eq->desc[0];
4283 
4284 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap);
4285 
4286 	}
4287 
4288 	txsd = &txq->sdesc[eq->pidx];
4289 	txsd->m = m0;
4290 	txsd->desc_used = ndesc;
4291 
4292 	return (ndesc);
4293 }
4294 
4295 /*
4296  * If the SGL ends on an address that is not 16 byte aligned, this function will
4297  * add a 0 filled flit at the end.
4298  */
4299 static void
4300 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap)
4301 {
4302 	struct sge_eq *eq = &txq->eq;
4303 	struct sglist *gl = txq->gl;
4304 	struct sglist_seg *seg;
4305 	__be64 *flitp, *wrap;
4306 	struct ulptx_sgl *usgl;
4307 	int i, nflits, nsegs;
4308 
4309 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
4310 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
4311 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
4312 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
4313 
4314 	get_pkt_gl(m, gl);
4315 	nsegs = gl->sg_nseg;
4316 	MPASS(nsegs > 0);
4317 
4318 	nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
4319 	flitp = (__be64 *)(*to);
4320 	wrap = (__be64 *)(&eq->desc[eq->sidx]);
4321 	seg = &gl->sg_segs[0];
4322 	usgl = (void *)flitp;
4323 
4324 	/*
4325 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
4326 	 * ring, so we're at least 16 bytes away from the status page.  There is
4327 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
4328 	 */
4329 
4330 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
4331 	    V_ULPTX_NSGE(nsegs));
4332 	usgl->len0 = htobe32(seg->ss_len);
4333 	usgl->addr0 = htobe64(seg->ss_paddr);
4334 	seg++;
4335 
4336 	if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) {
4337 
4338 		/* Won't wrap around at all */
4339 
4340 		for (i = 0; i < nsegs - 1; i++, seg++) {
4341 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
4342 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
4343 		}
4344 		if (i & 1)
4345 			usgl->sge[i / 2].len[1] = htobe32(0);
4346 		flitp += nflits;
4347 	} else {
4348 
4349 		/* Will wrap somewhere in the rest of the SGL */
4350 
4351 		/* 2 flits already written, write the rest flit by flit */
4352 		flitp = (void *)(usgl + 1);
4353 		for (i = 0; i < nflits - 2; i++) {
4354 			if (flitp == wrap)
4355 				flitp = (void *)eq->desc;
4356 			*flitp++ = get_flit(seg, nsegs - 1, i);
4357 		}
4358 	}
4359 
4360 	if (nflits & 1) {
4361 		MPASS(((uintptr_t)flitp) & 0xf);
4362 		*flitp++ = 0;
4363 	}
4364 
4365 	MPASS((((uintptr_t)flitp) & 0xf) == 0);
4366 	if (__predict_false(flitp == wrap))
4367 		*to = (void *)eq->desc;
4368 	else
4369 		*to = (void *)flitp;
4370 }
4371 
4372 static inline void
4373 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
4374 {
4375 
4376 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
4377 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
4378 
4379 	if (__predict_true((uintptr_t)(*to) + len <=
4380 	    (uintptr_t)&eq->desc[eq->sidx])) {
4381 		bcopy(from, *to, len);
4382 		(*to) += len;
4383 	} else {
4384 		int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
4385 
4386 		bcopy(from, *to, portion);
4387 		from += portion;
4388 		portion = len - portion;	/* remaining */
4389 		bcopy(from, (void *)eq->desc, portion);
4390 		(*to) = (caddr_t)eq->desc + portion;
4391 	}
4392 }
4393 
4394 static inline void
4395 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n)
4396 {
4397 	u_int db;
4398 
4399 	MPASS(n > 0);
4400 
4401 	db = eq->doorbells;
4402 	if (n > 1)
4403 		clrbit(&db, DOORBELL_WCWR);
4404 	wmb();
4405 
4406 	switch (ffs(db) - 1) {
4407 	case DOORBELL_UDB:
4408 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
4409 		break;
4410 
4411 	case DOORBELL_WCWR: {
4412 		volatile uint64_t *dst, *src;
4413 		int i;
4414 
4415 		/*
4416 		 * Queues whose 128B doorbell segment fits in the page do not
4417 		 * use relative qid (udb_qid is always 0).  Only queues with
4418 		 * doorbell segments can do WCWR.
4419 		 */
4420 		KASSERT(eq->udb_qid == 0 && n == 1,
4421 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
4422 		    __func__, eq->doorbells, n, eq->dbidx, eq));
4423 
4424 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
4425 		    UDBS_DB_OFFSET);
4426 		i = eq->dbidx;
4427 		src = (void *)&eq->desc[i];
4428 		while (src != (void *)&eq->desc[i + 1])
4429 			*dst++ = *src++;
4430 		wmb();
4431 		break;
4432 	}
4433 
4434 	case DOORBELL_UDBWC:
4435 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
4436 		wmb();
4437 		break;
4438 
4439 	case DOORBELL_KDB:
4440 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
4441 		    V_QID(eq->cntxt_id) | V_PIDX(n));
4442 		break;
4443 	}
4444 
4445 	IDXINCR(eq->dbidx, n, eq->sidx);
4446 }
4447 
4448 static inline u_int
4449 reclaimable_tx_desc(struct sge_eq *eq)
4450 {
4451 	uint16_t hw_cidx;
4452 
4453 	hw_cidx = read_hw_cidx(eq);
4454 	return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx));
4455 }
4456 
4457 static inline u_int
4458 total_available_tx_desc(struct sge_eq *eq)
4459 {
4460 	uint16_t hw_cidx, pidx;
4461 
4462 	hw_cidx = read_hw_cidx(eq);
4463 	pidx = eq->pidx;
4464 
4465 	if (pidx == hw_cidx)
4466 		return (eq->sidx - 1);
4467 	else
4468 		return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1);
4469 }
4470 
4471 static inline uint16_t
4472 read_hw_cidx(struct sge_eq *eq)
4473 {
4474 	struct sge_qstat *spg = (void *)&eq->desc[eq->sidx];
4475 	uint16_t cidx = spg->cidx;	/* stable snapshot */
4476 
4477 	return (be16toh(cidx));
4478 }
4479 
4480 /*
4481  * Reclaim 'n' descriptors approximately.
4482  */
4483 static u_int
4484 reclaim_tx_descs(struct sge_txq *txq, u_int n)
4485 {
4486 	struct tx_sdesc *txsd;
4487 	struct sge_eq *eq = &txq->eq;
4488 	u_int can_reclaim, reclaimed;
4489 
4490 	TXQ_LOCK_ASSERT_OWNED(txq);
4491 	MPASS(n > 0);
4492 
4493 	reclaimed = 0;
4494 	can_reclaim = reclaimable_tx_desc(eq);
4495 	while (can_reclaim && reclaimed < n) {
4496 		int ndesc;
4497 		struct mbuf *m, *nextpkt;
4498 
4499 		txsd = &txq->sdesc[eq->cidx];
4500 		ndesc = txsd->desc_used;
4501 
4502 		/* Firmware doesn't return "partial" credits. */
4503 		KASSERT(can_reclaim >= ndesc,
4504 		    ("%s: unexpected number of credits: %d, %d",
4505 		    __func__, can_reclaim, ndesc));
4506 
4507 		for (m = txsd->m; m != NULL; m = nextpkt) {
4508 			nextpkt = m->m_nextpkt;
4509 			m->m_nextpkt = NULL;
4510 			m_freem(m);
4511 		}
4512 		reclaimed += ndesc;
4513 		can_reclaim -= ndesc;
4514 		IDXINCR(eq->cidx, ndesc, eq->sidx);
4515 	}
4516 
4517 	return (reclaimed);
4518 }
4519 
4520 static void
4521 tx_reclaim(void *arg, int n)
4522 {
4523 	struct sge_txq *txq = arg;
4524 	struct sge_eq *eq = &txq->eq;
4525 
4526 	do {
4527 		if (TXQ_TRYLOCK(txq) == 0)
4528 			break;
4529 		n = reclaim_tx_descs(txq, 32);
4530 		if (eq->cidx == eq->pidx)
4531 			eq->equeqidx = eq->pidx;
4532 		TXQ_UNLOCK(txq);
4533 	} while (n > 0);
4534 }
4535 
4536 static __be64
4537 get_flit(struct sglist_seg *segs, int nsegs, int idx)
4538 {
4539 	int i = (idx / 3) * 2;
4540 
4541 	switch (idx % 3) {
4542 	case 0: {
4543 		__be64 rc;
4544 
4545 		rc = htobe32(segs[i].ss_len);
4546 		if (i + 1 < nsegs)
4547 			rc |= (uint64_t)htobe32(segs[i + 1].ss_len) << 32;
4548 
4549 		return (rc);
4550 	}
4551 	case 1:
4552 		return (htobe64(segs[i].ss_paddr));
4553 	case 2:
4554 		return (htobe64(segs[i + 1].ss_paddr));
4555 	}
4556 
4557 	return (0);
4558 }
4559 
4560 static void
4561 find_best_refill_source(struct adapter *sc, struct sge_fl *fl, int maxp)
4562 {
4563 	int8_t zidx, hwidx, idx;
4564 	uint16_t region1, region3;
4565 	int spare, spare_needed, n;
4566 	struct sw_zone_info *swz;
4567 	struct hw_buf_info *hwb, *hwb_list = &sc->sge.hw_buf_info[0];
4568 
4569 	/*
4570 	 * Buffer Packing: Look for PAGE_SIZE or larger zone which has a bufsize
4571 	 * large enough for the max payload and cluster metadata.  Otherwise
4572 	 * settle for the largest bufsize that leaves enough room in the cluster
4573 	 * for metadata.
4574 	 *
4575 	 * Without buffer packing: Look for the smallest zone which has a
4576 	 * bufsize large enough for the max payload.  Settle for the largest
4577 	 * bufsize available if there's nothing big enough for max payload.
4578 	 */
4579 	spare_needed = fl->flags & FL_BUF_PACKING ? CL_METADATA_SIZE : 0;
4580 	swz = &sc->sge.sw_zone_info[0];
4581 	hwidx = -1;
4582 	for (zidx = 0; zidx < SW_ZONE_SIZES; zidx++, swz++) {
4583 		if (swz->size > largest_rx_cluster) {
4584 			if (__predict_true(hwidx != -1))
4585 				break;
4586 
4587 			/*
4588 			 * This is a misconfiguration.  largest_rx_cluster is
4589 			 * preventing us from finding a refill source.  See
4590 			 * dev.t5nex.<n>.buffer_sizes to figure out why.
4591 			 */
4592 			device_printf(sc->dev, "largest_rx_cluster=%u leaves no"
4593 			    " refill source for fl %p (dma %u).  Ignored.\n",
4594 			    largest_rx_cluster, fl, maxp);
4595 		}
4596 		for (idx = swz->head_hwidx; idx != -1; idx = hwb->next) {
4597 			hwb = &hwb_list[idx];
4598 			spare = swz->size - hwb->size;
4599 			if (spare < spare_needed)
4600 				continue;
4601 
4602 			hwidx = idx;		/* best option so far */
4603 			if (hwb->size >= maxp) {
4604 
4605 				if ((fl->flags & FL_BUF_PACKING) == 0)
4606 					goto done; /* stop looking (not packing) */
4607 
4608 				if (swz->size >= safest_rx_cluster)
4609 					goto done; /* stop looking (packing) */
4610 			}
4611 			break;		/* keep looking, next zone */
4612 		}
4613 	}
4614 done:
4615 	/* A usable hwidx has been located. */
4616 	MPASS(hwidx != -1);
4617 	hwb = &hwb_list[hwidx];
4618 	zidx = hwb->zidx;
4619 	swz = &sc->sge.sw_zone_info[zidx];
4620 	region1 = 0;
4621 	region3 = swz->size - hwb->size;
4622 
4623 	/*
4624 	 * Stay within this zone and see if there is a better match when mbuf
4625 	 * inlining is allowed.  Remember that the hwidx's are sorted in
4626 	 * decreasing order of size (so in increasing order of spare area).
4627 	 */
4628 	for (idx = hwidx; idx != -1; idx = hwb->next) {
4629 		hwb = &hwb_list[idx];
4630 		spare = swz->size - hwb->size;
4631 
4632 		if (allow_mbufs_in_cluster == 0 || hwb->size < maxp)
4633 			break;
4634 
4635 		/*
4636 		 * Do not inline mbufs if doing so would violate the pad/pack
4637 		 * boundary alignment requirement.
4638 		 */
4639 		if (fl_pad && (MSIZE % sc->params.sge.pad_boundary) != 0)
4640 			continue;
4641 		if (fl->flags & FL_BUF_PACKING &&
4642 		    (MSIZE % sc->params.sge.pack_boundary) != 0)
4643 			continue;
4644 
4645 		if (spare < CL_METADATA_SIZE + MSIZE)
4646 			continue;
4647 		n = (spare - CL_METADATA_SIZE) / MSIZE;
4648 		if (n > howmany(hwb->size, maxp))
4649 			break;
4650 
4651 		hwidx = idx;
4652 		if (fl->flags & FL_BUF_PACKING) {
4653 			region1 = n * MSIZE;
4654 			region3 = spare - region1;
4655 		} else {
4656 			region1 = MSIZE;
4657 			region3 = spare - region1;
4658 			break;
4659 		}
4660 	}
4661 
4662 	KASSERT(zidx >= 0 && zidx < SW_ZONE_SIZES,
4663 	    ("%s: bad zone %d for fl %p, maxp %d", __func__, zidx, fl, maxp));
4664 	KASSERT(hwidx >= 0 && hwidx <= SGE_FLBUF_SIZES,
4665 	    ("%s: bad hwidx %d for fl %p, maxp %d", __func__, hwidx, fl, maxp));
4666 	KASSERT(region1 + sc->sge.hw_buf_info[hwidx].size + region3 ==
4667 	    sc->sge.sw_zone_info[zidx].size,
4668 	    ("%s: bad buffer layout for fl %p, maxp %d. "
4669 		"cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
4670 		sc->sge.sw_zone_info[zidx].size, region1,
4671 		sc->sge.hw_buf_info[hwidx].size, region3));
4672 	if (fl->flags & FL_BUF_PACKING || region1 > 0) {
4673 		KASSERT(region3 >= CL_METADATA_SIZE,
4674 		    ("%s: no room for metadata.  fl %p, maxp %d; "
4675 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
4676 		    sc->sge.sw_zone_info[zidx].size, region1,
4677 		    sc->sge.hw_buf_info[hwidx].size, region3));
4678 		KASSERT(region1 % MSIZE == 0,
4679 		    ("%s: bad mbuf region for fl %p, maxp %d. "
4680 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
4681 		    sc->sge.sw_zone_info[zidx].size, region1,
4682 		    sc->sge.hw_buf_info[hwidx].size, region3));
4683 	}
4684 
4685 	fl->cll_def.zidx = zidx;
4686 	fl->cll_def.hwidx = hwidx;
4687 	fl->cll_def.region1 = region1;
4688 	fl->cll_def.region3 = region3;
4689 }
4690 
4691 static void
4692 find_safe_refill_source(struct adapter *sc, struct sge_fl *fl)
4693 {
4694 	struct sge *s = &sc->sge;
4695 	struct hw_buf_info *hwb;
4696 	struct sw_zone_info *swz;
4697 	int spare;
4698 	int8_t hwidx;
4699 
4700 	if (fl->flags & FL_BUF_PACKING)
4701 		hwidx = s->safe_hwidx2;	/* with room for metadata */
4702 	else if (allow_mbufs_in_cluster && s->safe_hwidx2 != -1) {
4703 		hwidx = s->safe_hwidx2;
4704 		hwb = &s->hw_buf_info[hwidx];
4705 		swz = &s->sw_zone_info[hwb->zidx];
4706 		spare = swz->size - hwb->size;
4707 
4708 		/* no good if there isn't room for an mbuf as well */
4709 		if (spare < CL_METADATA_SIZE + MSIZE)
4710 			hwidx = s->safe_hwidx1;
4711 	} else
4712 		hwidx = s->safe_hwidx1;
4713 
4714 	if (hwidx == -1) {
4715 		/* No fallback source */
4716 		fl->cll_alt.hwidx = -1;
4717 		fl->cll_alt.zidx = -1;
4718 
4719 		return;
4720 	}
4721 
4722 	hwb = &s->hw_buf_info[hwidx];
4723 	swz = &s->sw_zone_info[hwb->zidx];
4724 	spare = swz->size - hwb->size;
4725 	fl->cll_alt.hwidx = hwidx;
4726 	fl->cll_alt.zidx = hwb->zidx;
4727 	if (allow_mbufs_in_cluster &&
4728 	    (fl_pad == 0 || (MSIZE % sc->params.sge.pad_boundary) == 0))
4729 		fl->cll_alt.region1 = ((spare - CL_METADATA_SIZE) / MSIZE) * MSIZE;
4730 	else
4731 		fl->cll_alt.region1 = 0;
4732 	fl->cll_alt.region3 = spare - fl->cll_alt.region1;
4733 }
4734 
4735 static void
4736 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
4737 {
4738 	mtx_lock(&sc->sfl_lock);
4739 	FL_LOCK(fl);
4740 	if ((fl->flags & FL_DOOMED) == 0) {
4741 		fl->flags |= FL_STARVING;
4742 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
4743 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
4744 	}
4745 	FL_UNLOCK(fl);
4746 	mtx_unlock(&sc->sfl_lock);
4747 }
4748 
4749 static void
4750 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq)
4751 {
4752 	struct sge_wrq *wrq = (void *)eq;
4753 
4754 	atomic_readandclear_int(&eq->equiq);
4755 	taskqueue_enqueue(sc->tq[eq->tx_chan], &wrq->wrq_tx_task);
4756 }
4757 
4758 static void
4759 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq)
4760 {
4761 	struct sge_txq *txq = (void *)eq;
4762 
4763 	MPASS((eq->flags & EQ_TYPEMASK) == EQ_ETH);
4764 
4765 	atomic_readandclear_int(&eq->equiq);
4766 	mp_ring_check_drainage(txq->r, 0);
4767 	taskqueue_enqueue(sc->tq[eq->tx_chan], &txq->tx_reclaim_task);
4768 }
4769 
4770 static int
4771 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
4772     struct mbuf *m)
4773 {
4774 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
4775 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
4776 	struct adapter *sc = iq->adapter;
4777 	struct sge *s = &sc->sge;
4778 	struct sge_eq *eq;
4779 	static void (*h[])(struct adapter *, struct sge_eq *) = {NULL,
4780 		&handle_wrq_egr_update, &handle_eth_egr_update,
4781 		&handle_wrq_egr_update};
4782 
4783 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
4784 	    rss->opcode));
4785 
4786 	eq = s->eqmap[qid - s->eq_start - s->eq_base];
4787 	(*h[eq->flags & EQ_TYPEMASK])(sc, eq);
4788 
4789 	return (0);
4790 }
4791 
4792 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
4793 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
4794     offsetof(struct cpl_fw6_msg, data));
4795 
4796 static int
4797 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
4798 {
4799 	struct adapter *sc = iq->adapter;
4800 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
4801 
4802 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
4803 	    rss->opcode));
4804 
4805 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
4806 		const struct rss_header *rss2;
4807 
4808 		rss2 = (const struct rss_header *)&cpl->data[0];
4809 		return (t4_cpl_handler[rss2->opcode](iq, rss2, m));
4810 	}
4811 
4812 	return (t4_fw_msg_handler[cpl->type](sc, &cpl->data[0]));
4813 }
4814 
4815 /**
4816  *	t4_handle_wrerr_rpl - process a FW work request error message
4817  *	@adap: the adapter
4818  *	@rpl: start of the FW message
4819  */
4820 static int
4821 t4_handle_wrerr_rpl(struct adapter *adap, const __be64 *rpl)
4822 {
4823 	u8 opcode = *(const u8 *)rpl;
4824 	const struct fw_error_cmd *e = (const void *)rpl;
4825 	unsigned int i;
4826 
4827 	if (opcode != FW_ERROR_CMD) {
4828 		log(LOG_ERR,
4829 		    "%s: Received WRERR_RPL message with opcode %#x\n",
4830 		    device_get_nameunit(adap->dev), opcode);
4831 		return (EINVAL);
4832 	}
4833 	log(LOG_ERR, "%s: FW_ERROR (%s) ", device_get_nameunit(adap->dev),
4834 	    G_FW_ERROR_CMD_FATAL(be32toh(e->op_to_type)) ? "fatal" :
4835 	    "non-fatal");
4836 	switch (G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type))) {
4837 	case FW_ERROR_TYPE_EXCEPTION:
4838 		log(LOG_ERR, "exception info:\n");
4839 		for (i = 0; i < nitems(e->u.exception.info); i++)
4840 			log(LOG_ERR, "%s%08x", i == 0 ? "\t" : " ",
4841 			    be32toh(e->u.exception.info[i]));
4842 		log(LOG_ERR, "\n");
4843 		break;
4844 	case FW_ERROR_TYPE_HWMODULE:
4845 		log(LOG_ERR, "HW module regaddr %08x regval %08x\n",
4846 		    be32toh(e->u.hwmodule.regaddr),
4847 		    be32toh(e->u.hwmodule.regval));
4848 		break;
4849 	case FW_ERROR_TYPE_WR:
4850 		log(LOG_ERR, "WR cidx %d PF %d VF %d eqid %d hdr:\n",
4851 		    be16toh(e->u.wr.cidx),
4852 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.wr.pfn_vfn)),
4853 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.wr.pfn_vfn)),
4854 		    be32toh(e->u.wr.eqid));
4855 		for (i = 0; i < nitems(e->u.wr.wrhdr); i++)
4856 			log(LOG_ERR, "%s%02x", i == 0 ? "\t" : " ",
4857 			    e->u.wr.wrhdr[i]);
4858 		log(LOG_ERR, "\n");
4859 		break;
4860 	case FW_ERROR_TYPE_ACL:
4861 		log(LOG_ERR, "ACL cidx %d PF %d VF %d eqid %d %s",
4862 		    be16toh(e->u.acl.cidx),
4863 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.acl.pfn_vfn)),
4864 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.acl.pfn_vfn)),
4865 		    be32toh(e->u.acl.eqid),
4866 		    G_FW_ERROR_CMD_MV(be16toh(e->u.acl.mv_pkd)) ? "vlanid" :
4867 		    "MAC");
4868 		for (i = 0; i < nitems(e->u.acl.val); i++)
4869 			log(LOG_ERR, " %02x", e->u.acl.val[i]);
4870 		log(LOG_ERR, "\n");
4871 		break;
4872 	default:
4873 		log(LOG_ERR, "type %#x\n",
4874 		    G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type)));
4875 		return (EINVAL);
4876 	}
4877 	return (0);
4878 }
4879 
4880 static int
4881 sysctl_uint16(SYSCTL_HANDLER_ARGS)
4882 {
4883 	uint16_t *id = arg1;
4884 	int i = *id;
4885 
4886 	return sysctl_handle_int(oidp, &i, 0, req);
4887 }
4888 
4889 static int
4890 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
4891 {
4892 	struct sge *s = arg1;
4893 	struct hw_buf_info *hwb = &s->hw_buf_info[0];
4894 	struct sw_zone_info *swz = &s->sw_zone_info[0];
4895 	int i, rc;
4896 	struct sbuf sb;
4897 	char c;
4898 
4899 	sbuf_new(&sb, NULL, 32, SBUF_AUTOEXTEND);
4900 	for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
4901 		if (hwb->zidx >= 0 && swz[hwb->zidx].size <= largest_rx_cluster)
4902 			c = '*';
4903 		else
4904 			c = '\0';
4905 
4906 		sbuf_printf(&sb, "%u%c ", hwb->size, c);
4907 	}
4908 	sbuf_trim(&sb);
4909 	sbuf_finish(&sb);
4910 	rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
4911 	sbuf_delete(&sb);
4912 	return (rc);
4913 }
4914 
4915 static int
4916 sysctl_tc(SYSCTL_HANDLER_ARGS)
4917 {
4918 	struct vi_info *vi = arg1;
4919 	struct port_info *pi;
4920 	struct adapter *sc;
4921 	struct sge_txq *txq;
4922 	struct tx_sched_class *tc;
4923 	int qidx = arg2, rc, tc_idx;
4924 	uint32_t fw_queue, fw_class;
4925 
4926 	MPASS(qidx >= 0 && qidx < vi->ntxq);
4927 	pi = vi->pi;
4928 	sc = pi->adapter;
4929 	txq = &sc->sge.txq[vi->first_txq + qidx];
4930 
4931 	tc_idx = txq->tc_idx;
4932 	rc = sysctl_handle_int(oidp, &tc_idx, 0, req);
4933 	if (rc != 0 || req->newptr == NULL)
4934 		return (rc);
4935 
4936 	/* Note that -1 is legitimate input (it means unbind). */
4937 	if (tc_idx < -1 || tc_idx >= sc->chip_params->nsched_cls)
4938 		return (EINVAL);
4939 
4940 	rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4stc");
4941 	if (rc)
4942 		return (rc);
4943 
4944 	if (tc_idx == txq->tc_idx) {
4945 		rc = 0;		/* No change, nothing to do. */
4946 		goto done;
4947 	}
4948 
4949 	fw_queue = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
4950 	    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_EQ_SCHEDCLASS_ETH) |
4951 	    V_FW_PARAMS_PARAM_YZ(txq->eq.cntxt_id);
4952 
4953 	if (tc_idx == -1)
4954 		fw_class = 0xffffffff;	/* Unbind. */
4955 	else {
4956 		/*
4957 		 * Bind to a different class.  Ethernet txq's are only allowed
4958 		 * to bind to cl-rl mode-class for now.  XXX: too restrictive.
4959 		 */
4960 		tc = &pi->tc[tc_idx];
4961 		if (tc->flags & TX_SC_OK &&
4962 		    tc->params.level == SCHED_CLASS_LEVEL_CL_RL &&
4963 		    tc->params.mode == SCHED_CLASS_MODE_CLASS) {
4964 			/* Ok to proceed. */
4965 			fw_class = tc_idx;
4966 		} else {
4967 			rc = tc->flags & TX_SC_OK ? EBUSY : ENXIO;
4968 			goto done;
4969 		}
4970 	}
4971 
4972 	rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &fw_queue, &fw_class);
4973 	if (rc == 0) {
4974 		if (txq->tc_idx != -1) {
4975 			tc = &pi->tc[txq->tc_idx];
4976 			MPASS(tc->refcount > 0);
4977 			tc->refcount--;
4978 		}
4979 		if (tc_idx != -1) {
4980 			tc = &pi->tc[tc_idx];
4981 			tc->refcount++;
4982 		}
4983 		txq->tc_idx = tc_idx;
4984 	}
4985 done:
4986 	end_synchronized_op(sc, 0);
4987 	return (rc);
4988 }
4989