xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision 4b330699f819a81d8e34d471225143ffeb321855)
1 /*-
2  * Copyright (c) 2011 Chelsio Communications, Inc.
3  * All rights reserved.
4  * Written by: Navdeep Parhar <np@FreeBSD.org>
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 
34 #include <sys/types.h>
35 #include <sys/eventhandler.h>
36 #include <sys/mbuf.h>
37 #include <sys/socket.h>
38 #include <sys/kernel.h>
39 #include <sys/malloc.h>
40 #include <sys/queue.h>
41 #include <sys/sbuf.h>
42 #include <sys/taskqueue.h>
43 #include <sys/time.h>
44 #include <sys/sglist.h>
45 #include <sys/sysctl.h>
46 #include <sys/smp.h>
47 #include <sys/counter.h>
48 #include <net/bpf.h>
49 #include <net/ethernet.h>
50 #include <net/if.h>
51 #include <net/if_vlan_var.h>
52 #include <netinet/in.h>
53 #include <netinet/ip.h>
54 #include <netinet/ip6.h>
55 #include <netinet/tcp.h>
56 #include <machine/in_cksum.h>
57 #include <machine/md_var.h>
58 #include <vm/vm.h>
59 #include <vm/pmap.h>
60 #ifdef DEV_NETMAP
61 #include <machine/bus.h>
62 #include <sys/selinfo.h>
63 #include <net/if_var.h>
64 #include <net/netmap.h>
65 #include <dev/netmap/netmap_kern.h>
66 #endif
67 
68 #include "common/common.h"
69 #include "common/t4_regs.h"
70 #include "common/t4_regs_values.h"
71 #include "common/t4_msg.h"
72 #include "t4_l2t.h"
73 #include "t4_mp_ring.h"
74 
75 #ifdef T4_PKT_TIMESTAMP
76 #define RX_COPY_THRESHOLD (MINCLSIZE - 8)
77 #else
78 #define RX_COPY_THRESHOLD MINCLSIZE
79 #endif
80 
81 /*
82  * Ethernet frames are DMA'd at this byte offset into the freelist buffer.
83  * 0-7 are valid values.
84  */
85 static int fl_pktshift = 2;
86 TUNABLE_INT("hw.cxgbe.fl_pktshift", &fl_pktshift);
87 
88 /*
89  * Pad ethernet payload up to this boundary.
90  * -1: driver should figure out a good value.
91  *  0: disable padding.
92  *  Any power of 2 from 32 to 4096 (both inclusive) is also a valid value.
93  */
94 int fl_pad = -1;
95 TUNABLE_INT("hw.cxgbe.fl_pad", &fl_pad);
96 
97 /*
98  * Status page length.
99  * -1: driver should figure out a good value.
100  *  64 or 128 are the only other valid values.
101  */
102 static int spg_len = -1;
103 TUNABLE_INT("hw.cxgbe.spg_len", &spg_len);
104 
105 /*
106  * Congestion drops.
107  * -1: no congestion feedback (not recommended).
108  *  0: backpressure the channel instead of dropping packets right away.
109  *  1: no backpressure, drop packets for the congested queue immediately.
110  */
111 static int cong_drop = 0;
112 TUNABLE_INT("hw.cxgbe.cong_drop", &cong_drop);
113 
114 /*
115  * Deliver multiple frames in the same free list buffer if they fit.
116  * -1: let the driver decide whether to enable buffer packing or not.
117  *  0: disable buffer packing.
118  *  1: enable buffer packing.
119  */
120 static int buffer_packing = -1;
121 TUNABLE_INT("hw.cxgbe.buffer_packing", &buffer_packing);
122 
123 /*
124  * Start next frame in a packed buffer at this boundary.
125  * -1: driver should figure out a good value.
126  * T4: driver will ignore this and use the same value as fl_pad above.
127  * T5: 16, or a power of 2 from 64 to 4096 (both inclusive) is a valid value.
128  */
129 static int fl_pack = -1;
130 TUNABLE_INT("hw.cxgbe.fl_pack", &fl_pack);
131 
132 /*
133  * Allow the driver to create mbuf(s) in a cluster allocated for rx.
134  * 0: never; always allocate mbufs from the zone_mbuf UMA zone.
135  * 1: ok to create mbuf(s) within a cluster if there is room.
136  */
137 static int allow_mbufs_in_cluster = 1;
138 TUNABLE_INT("hw.cxgbe.allow_mbufs_in_cluster", &allow_mbufs_in_cluster);
139 
140 /*
141  * Largest rx cluster size that the driver is allowed to allocate.
142  */
143 static int largest_rx_cluster = MJUM16BYTES;
144 TUNABLE_INT("hw.cxgbe.largest_rx_cluster", &largest_rx_cluster);
145 
146 /*
147  * Size of cluster allocation that's most likely to succeed.  The driver will
148  * fall back to this size if it fails to allocate clusters larger than this.
149  */
150 static int safest_rx_cluster = PAGE_SIZE;
151 TUNABLE_INT("hw.cxgbe.safest_rx_cluster", &safest_rx_cluster);
152 
153 /*
154  * The interrupt holdoff timers are multiplied by this value on T6+.
155  * 1 and 3-17 (both inclusive) are legal values.
156  */
157 static int tscale = 1;
158 TUNABLE_INT("hw.cxgbe.tscale", &tscale);
159 
160 /*
161  * Number of LRO entries in the lro_ctrl structure per rx queue.
162  */
163 static int lro_entries = TCP_LRO_ENTRIES;
164 TUNABLE_INT("hw.cxgbe.lro_entries", &lro_entries);
165 
166 /*
167  * This enables presorting of frames before they're fed into tcp_lro_rx.
168  */
169 static int lro_mbufs = 0;
170 TUNABLE_INT("hw.cxgbe.lro_mbufs", &lro_mbufs);
171 
172 struct txpkts {
173 	u_int wr_type;		/* type 0 or type 1 */
174 	u_int npkt;		/* # of packets in this work request */
175 	u_int plen;		/* total payload (sum of all packets) */
176 	u_int len16;		/* # of 16B pieces used by this work request */
177 };
178 
179 /* A packet's SGL.  This + m_pkthdr has all info needed for tx */
180 struct sgl {
181 	struct sglist sg;
182 	struct sglist_seg seg[TX_SGL_SEGS];
183 };
184 
185 static int service_iq(struct sge_iq *, int);
186 static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t);
187 static int t4_eth_rx(struct sge_iq *, const struct rss_header *, struct mbuf *);
188 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int);
189 static inline void init_fl(struct adapter *, struct sge_fl *, int, int, char *);
190 static inline void init_eq(struct adapter *, struct sge_eq *, int, int, uint8_t,
191     uint16_t, char *);
192 static int alloc_ring(struct adapter *, size_t, bus_dma_tag_t *, bus_dmamap_t *,
193     bus_addr_t *, void **);
194 static int free_ring(struct adapter *, bus_dma_tag_t, bus_dmamap_t, bus_addr_t,
195     void *);
196 static int alloc_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *,
197     int, int);
198 static int free_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *);
199 static void add_fl_sysctls(struct adapter *, struct sysctl_ctx_list *,
200     struct sysctl_oid *, struct sge_fl *);
201 static int alloc_fwq(struct adapter *);
202 static int free_fwq(struct adapter *);
203 static int alloc_mgmtq(struct adapter *);
204 static int free_mgmtq(struct adapter *);
205 static int alloc_rxq(struct vi_info *, struct sge_rxq *, int, int,
206     struct sysctl_oid *);
207 static int free_rxq(struct vi_info *, struct sge_rxq *);
208 #ifdef TCP_OFFLOAD
209 static int alloc_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *, int, int,
210     struct sysctl_oid *);
211 static int free_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *);
212 #endif
213 #ifdef DEV_NETMAP
214 static int alloc_nm_rxq(struct vi_info *, struct sge_nm_rxq *, int, int,
215     struct sysctl_oid *);
216 static int free_nm_rxq(struct vi_info *, struct sge_nm_rxq *);
217 static int alloc_nm_txq(struct vi_info *, struct sge_nm_txq *, int, int,
218     struct sysctl_oid *);
219 static int free_nm_txq(struct vi_info *, struct sge_nm_txq *);
220 #endif
221 static int ctrl_eq_alloc(struct adapter *, struct sge_eq *);
222 static int eth_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
223 #ifdef TCP_OFFLOAD
224 static int ofld_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
225 #endif
226 static int alloc_eq(struct adapter *, struct vi_info *, struct sge_eq *);
227 static int free_eq(struct adapter *, struct sge_eq *);
228 static int alloc_wrq(struct adapter *, struct vi_info *, struct sge_wrq *,
229     struct sysctl_oid *);
230 static int free_wrq(struct adapter *, struct sge_wrq *);
231 static int alloc_txq(struct vi_info *, struct sge_txq *, int,
232     struct sysctl_oid *);
233 static int free_txq(struct vi_info *, struct sge_txq *);
234 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int);
235 static inline void ring_fl_db(struct adapter *, struct sge_fl *);
236 static int refill_fl(struct adapter *, struct sge_fl *, int);
237 static void refill_sfl(void *);
238 static int alloc_fl_sdesc(struct sge_fl *);
239 static void free_fl_sdesc(struct adapter *, struct sge_fl *);
240 static void find_best_refill_source(struct adapter *, struct sge_fl *, int);
241 static void find_safe_refill_source(struct adapter *, struct sge_fl *);
242 static void add_fl_to_sfl(struct adapter *, struct sge_fl *);
243 
244 static inline void get_pkt_gl(struct mbuf *, struct sglist *);
245 static inline u_int txpkt_len16(u_int, u_int);
246 static inline u_int txpkt_vm_len16(u_int, u_int);
247 static inline u_int txpkts0_len16(u_int);
248 static inline u_int txpkts1_len16(void);
249 static u_int write_txpkt_wr(struct sge_txq *, struct fw_eth_tx_pkt_wr *,
250     struct mbuf *, u_int);
251 static u_int write_txpkt_vm_wr(struct adapter *, struct sge_txq *,
252     struct fw_eth_tx_pkt_vm_wr *, struct mbuf *, u_int);
253 static int try_txpkts(struct mbuf *, struct mbuf *, struct txpkts *, u_int);
254 static int add_to_txpkts(struct mbuf *, struct txpkts *, u_int);
255 static u_int write_txpkts_wr(struct sge_txq *, struct fw_eth_tx_pkts_wr *,
256     struct mbuf *, const struct txpkts *, u_int);
257 static void write_gl_to_txd(struct sge_txq *, struct mbuf *, caddr_t *, int);
258 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int);
259 static inline void ring_eq_db(struct adapter *, struct sge_eq *, u_int);
260 static inline uint16_t read_hw_cidx(struct sge_eq *);
261 static inline u_int reclaimable_tx_desc(struct sge_eq *);
262 static inline u_int total_available_tx_desc(struct sge_eq *);
263 static u_int reclaim_tx_descs(struct sge_txq *, u_int);
264 static void tx_reclaim(void *, int);
265 static __be64 get_flit(struct sglist_seg *, int, int);
266 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
267     struct mbuf *);
268 static int handle_fw_msg(struct sge_iq *, const struct rss_header *,
269     struct mbuf *);
270 static int t4_handle_wrerr_rpl(struct adapter *, const __be64 *);
271 static void wrq_tx_drain(void *, int);
272 static void drain_wrq_wr_list(struct adapter *, struct sge_wrq *);
273 
274 static int sysctl_uint16(SYSCTL_HANDLER_ARGS);
275 static int sysctl_bufsizes(SYSCTL_HANDLER_ARGS);
276 static int sysctl_tc(SYSCTL_HANDLER_ARGS);
277 
278 static counter_u64_t extfree_refs;
279 static counter_u64_t extfree_rels;
280 
281 an_handler_t t4_an_handler;
282 fw_msg_handler_t t4_fw_msg_handler[NUM_FW6_TYPES];
283 cpl_handler_t t4_cpl_handler[NUM_CPL_CMDS];
284 
285 
286 static int
287 an_not_handled(struct sge_iq *iq, const struct rsp_ctrl *ctrl)
288 {
289 
290 #ifdef INVARIANTS
291 	panic("%s: async notification on iq %p (ctrl %p)", __func__, iq, ctrl);
292 #else
293 	log(LOG_ERR, "%s: async notification on iq %p (ctrl %p)\n",
294 	    __func__, iq, ctrl);
295 #endif
296 	return (EDOOFUS);
297 }
298 
299 int
300 t4_register_an_handler(an_handler_t h)
301 {
302 	uintptr_t *loc, new;
303 
304 	new = h ? (uintptr_t)h : (uintptr_t)an_not_handled;
305 	loc = (uintptr_t *) &t4_an_handler;
306 	atomic_store_rel_ptr(loc, new);
307 
308 	return (0);
309 }
310 
311 static int
312 fw_msg_not_handled(struct adapter *sc, const __be64 *rpl)
313 {
314 	const struct cpl_fw6_msg *cpl =
315 	    __containerof(rpl, struct cpl_fw6_msg, data[0]);
316 
317 #ifdef INVARIANTS
318 	panic("%s: fw_msg type %d", __func__, cpl->type);
319 #else
320 	log(LOG_ERR, "%s: fw_msg type %d\n", __func__, cpl->type);
321 #endif
322 	return (EDOOFUS);
323 }
324 
325 int
326 t4_register_fw_msg_handler(int type, fw_msg_handler_t h)
327 {
328 	uintptr_t *loc, new;
329 
330 	if (type >= nitems(t4_fw_msg_handler))
331 		return (EINVAL);
332 
333 	/*
334 	 * These are dispatched by the handler for FW{4|6}_CPL_MSG using the CPL
335 	 * handler dispatch table.  Reject any attempt to install a handler for
336 	 * this subtype.
337 	 */
338 	if (type == FW_TYPE_RSSCPL || type == FW6_TYPE_RSSCPL)
339 		return (EINVAL);
340 
341 	new = h ? (uintptr_t)h : (uintptr_t)fw_msg_not_handled;
342 	loc = (uintptr_t *) &t4_fw_msg_handler[type];
343 	atomic_store_rel_ptr(loc, new);
344 
345 	return (0);
346 }
347 
348 static int
349 cpl_not_handled(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
350 {
351 
352 #ifdef INVARIANTS
353 	panic("%s: opcode 0x%02x on iq %p with payload %p",
354 	    __func__, rss->opcode, iq, m);
355 #else
356 	log(LOG_ERR, "%s: opcode 0x%02x on iq %p with payload %p\n",
357 	    __func__, rss->opcode, iq, m);
358 	m_freem(m);
359 #endif
360 	return (EDOOFUS);
361 }
362 
363 int
364 t4_register_cpl_handler(int opcode, cpl_handler_t h)
365 {
366 	uintptr_t *loc, new;
367 
368 	if (opcode >= nitems(t4_cpl_handler))
369 		return (EINVAL);
370 
371 	new = h ? (uintptr_t)h : (uintptr_t)cpl_not_handled;
372 	loc = (uintptr_t *) &t4_cpl_handler[opcode];
373 	atomic_store_rel_ptr(loc, new);
374 
375 	return (0);
376 }
377 
378 /*
379  * Called on MOD_LOAD.  Validates and calculates the SGE tunables.
380  */
381 void
382 t4_sge_modload(void)
383 {
384 	int i;
385 
386 	if (fl_pktshift < 0 || fl_pktshift > 7) {
387 		printf("Invalid hw.cxgbe.fl_pktshift value (%d),"
388 		    " using 2 instead.\n", fl_pktshift);
389 		fl_pktshift = 2;
390 	}
391 
392 	if (spg_len != 64 && spg_len != 128) {
393 		int len;
394 
395 #if defined(__i386__) || defined(__amd64__)
396 		len = cpu_clflush_line_size > 64 ? 128 : 64;
397 #else
398 		len = 64;
399 #endif
400 		if (spg_len != -1) {
401 			printf("Invalid hw.cxgbe.spg_len value (%d),"
402 			    " using %d instead.\n", spg_len, len);
403 		}
404 		spg_len = len;
405 	}
406 
407 	if (cong_drop < -1 || cong_drop > 1) {
408 		printf("Invalid hw.cxgbe.cong_drop value (%d),"
409 		    " using 0 instead.\n", cong_drop);
410 		cong_drop = 0;
411 	}
412 
413 	if (tscale != 1 && (tscale < 3 || tscale > 17)) {
414 		printf("Invalid hw.cxgbe.tscale value (%d),"
415 		    " using 1 instead.\n", tscale);
416 		tscale = 1;
417 	}
418 
419 	extfree_refs = counter_u64_alloc(M_WAITOK);
420 	extfree_rels = counter_u64_alloc(M_WAITOK);
421 	counter_u64_zero(extfree_refs);
422 	counter_u64_zero(extfree_rels);
423 
424 	t4_an_handler = an_not_handled;
425 	for (i = 0; i < nitems(t4_fw_msg_handler); i++)
426 		t4_fw_msg_handler[i] = fw_msg_not_handled;
427 	for (i = 0; i < nitems(t4_cpl_handler); i++)
428 		t4_cpl_handler[i] = cpl_not_handled;
429 
430 	t4_register_cpl_handler(CPL_FW4_MSG, handle_fw_msg);
431 	t4_register_cpl_handler(CPL_FW6_MSG, handle_fw_msg);
432 	t4_register_cpl_handler(CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
433 	t4_register_cpl_handler(CPL_RX_PKT, t4_eth_rx);
434 	t4_register_fw_msg_handler(FW6_TYPE_CMD_RPL, t4_handle_fw_rpl);
435 	t4_register_fw_msg_handler(FW6_TYPE_WRERR_RPL, t4_handle_wrerr_rpl);
436 }
437 
438 void
439 t4_sge_modunload(void)
440 {
441 
442 	counter_u64_free(extfree_refs);
443 	counter_u64_free(extfree_rels);
444 }
445 
446 uint64_t
447 t4_sge_extfree_refs(void)
448 {
449 	uint64_t refs, rels;
450 
451 	rels = counter_u64_fetch(extfree_rels);
452 	refs = counter_u64_fetch(extfree_refs);
453 
454 	return (refs - rels);
455 }
456 
457 static inline void
458 setup_pad_and_pack_boundaries(struct adapter *sc)
459 {
460 	uint32_t v, m;
461 	int pad, pack, pad_shift;
462 
463 	pad_shift = chip_id(sc) > CHELSIO_T5 ? X_T6_INGPADBOUNDARY_SHIFT :
464 	    X_INGPADBOUNDARY_SHIFT;
465 	pad = fl_pad;
466 	if (fl_pad < (1 << pad_shift) ||
467 	    fl_pad > (1 << (pad_shift + M_INGPADBOUNDARY)) ||
468 	    !powerof2(fl_pad)) {
469 		/*
470 		 * If there is any chance that we might use buffer packing and
471 		 * the chip is a T4, then pick 64 as the pad/pack boundary.  Set
472 		 * it to the minimum allowed in all other cases.
473 		 */
474 		pad = is_t4(sc) && buffer_packing ? 64 : 1 << pad_shift;
475 
476 		/*
477 		 * For fl_pad = 0 we'll still write a reasonable value to the
478 		 * register but all the freelists will opt out of padding.
479 		 * We'll complain here only if the user tried to set it to a
480 		 * value greater than 0 that was invalid.
481 		 */
482 		if (fl_pad > 0) {
483 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pad value"
484 			    " (%d), using %d instead.\n", fl_pad, pad);
485 		}
486 	}
487 	m = V_INGPADBOUNDARY(M_INGPADBOUNDARY);
488 	v = V_INGPADBOUNDARY(ilog2(pad) - pad_shift);
489 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
490 
491 	if (is_t4(sc)) {
492 		if (fl_pack != -1 && fl_pack != pad) {
493 			/* Complain but carry on. */
494 			device_printf(sc->dev, "hw.cxgbe.fl_pack (%d) ignored,"
495 			    " using %d instead.\n", fl_pack, pad);
496 		}
497 		return;
498 	}
499 
500 	pack = fl_pack;
501 	if (fl_pack < 16 || fl_pack == 32 || fl_pack > 4096 ||
502 	    !powerof2(fl_pack)) {
503 		pack = max(sc->params.pci.mps, CACHE_LINE_SIZE);
504 		MPASS(powerof2(pack));
505 		if (pack < 16)
506 			pack = 16;
507 		if (pack == 32)
508 			pack = 64;
509 		if (pack > 4096)
510 			pack = 4096;
511 		if (fl_pack != -1) {
512 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pack value"
513 			    " (%d), using %d instead.\n", fl_pack, pack);
514 		}
515 	}
516 	m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY);
517 	if (pack == 16)
518 		v = V_INGPACKBOUNDARY(0);
519 	else
520 		v = V_INGPACKBOUNDARY(ilog2(pack) - 5);
521 
522 	MPASS(!is_t4(sc));	/* T4 doesn't have SGE_CONTROL2 */
523 	t4_set_reg_field(sc, A_SGE_CONTROL2, m, v);
524 }
525 
526 /*
527  * adap->params.vpd.cclk must be set up before this is called.
528  */
529 void
530 t4_tweak_chip_settings(struct adapter *sc)
531 {
532 	int i;
533 	uint32_t v, m;
534 	int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200};
535 	int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk;
536 	int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */
537 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
538 	static int sge_flbuf_sizes[] = {
539 		MCLBYTES,
540 #if MJUMPAGESIZE != MCLBYTES
541 		MJUMPAGESIZE,
542 		MJUMPAGESIZE - CL_METADATA_SIZE,
543 		MJUMPAGESIZE - 2 * MSIZE - CL_METADATA_SIZE,
544 #endif
545 		MJUM9BYTES,
546 		MJUM16BYTES,
547 		MCLBYTES - MSIZE - CL_METADATA_SIZE,
548 		MJUM9BYTES - CL_METADATA_SIZE,
549 		MJUM16BYTES - CL_METADATA_SIZE,
550 	};
551 
552 	KASSERT(sc->flags & MASTER_PF,
553 	    ("%s: trying to change chip settings when not master.", __func__));
554 
555 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
556 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
557 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
558 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
559 
560 	setup_pad_and_pack_boundaries(sc);
561 
562 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
563 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
564 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
565 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
566 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
567 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
568 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
569 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
570 	t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v);
571 
572 	KASSERT(nitems(sge_flbuf_sizes) <= SGE_FLBUF_SIZES,
573 	    ("%s: hw buffer size table too big", __func__));
574 	for (i = 0; i < min(nitems(sge_flbuf_sizes), SGE_FLBUF_SIZES); i++) {
575 		t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i),
576 		    sge_flbuf_sizes[i]);
577 	}
578 
579 	v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) |
580 	    V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]);
581 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v);
582 
583 	KASSERT(intr_timer[0] <= timer_max,
584 	    ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0],
585 	    timer_max));
586 	for (i = 1; i < nitems(intr_timer); i++) {
587 		KASSERT(intr_timer[i] >= intr_timer[i - 1],
588 		    ("%s: timers not listed in increasing order (%d)",
589 		    __func__, i));
590 
591 		while (intr_timer[i] > timer_max) {
592 			if (i == nitems(intr_timer) - 1) {
593 				intr_timer[i] = timer_max;
594 				break;
595 			}
596 			intr_timer[i] += intr_timer[i - 1];
597 			intr_timer[i] /= 2;
598 		}
599 	}
600 
601 	v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) |
602 	    V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1]));
603 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v);
604 	v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) |
605 	    V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3]));
606 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v);
607 	v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) |
608 	    V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5]));
609 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v);
610 
611 	if (chip_id(sc) >= CHELSIO_T6) {
612 		m = V_TSCALE(M_TSCALE);
613 		if (tscale == 1)
614 			v = 0;
615 		else
616 			v = V_TSCALE(tscale - 2);
617 		t4_set_reg_field(sc, A_SGE_ITP_CONTROL, m, v);
618 	}
619 
620 	/* 4K, 16K, 64K, 256K DDP "page sizes" for TDDP */
621 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
622 	t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v);
623 
624 	/*
625 	 * 4K, 8K, 16K, 64K DDP "page sizes" for iSCSI DDP.  These have been
626 	 * chosen with MAXPHYS = 128K in mind.  The largest DDP buffer that we
627 	 * may have to deal with is MAXPHYS + 1 page.
628 	 */
629 	v = V_HPZ0(0) | V_HPZ1(1) | V_HPZ2(2) | V_HPZ3(4);
630 	t4_write_reg(sc, A_ULP_RX_ISCSI_PSZ, v);
631 
632 	/* We use multiple DDP page sizes both in plain-TOE and ISCSI modes. */
633 	m = v = F_TDDPTAGTCB | F_ISCSITAGTCB;
634 	t4_set_reg_field(sc, A_ULP_RX_CTL, m, v);
635 
636 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
637 	    F_RESETDDPOFFSET;
638 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
639 	t4_set_reg_field(sc, A_TP_PARA_REG5, m, v);
640 }
641 
642 /*
643  * SGE wants the buffer to be at least 64B and then a multiple of 16.  If
644  * padding is in use, the buffer's start and end need to be aligned to the pad
645  * boundary as well.  We'll just make sure that the size is a multiple of the
646  * boundary here, it is up to the buffer allocation code to make sure the start
647  * of the buffer is aligned as well.
648  */
649 static inline int
650 hwsz_ok(struct adapter *sc, int hwsz)
651 {
652 	int mask = fl_pad ? sc->params.sge.pad_boundary - 1 : 16 - 1;
653 
654 	return (hwsz >= 64 && (hwsz & mask) == 0);
655 }
656 
657 /*
658  * XXX: driver really should be able to deal with unexpected settings.
659  */
660 int
661 t4_read_chip_settings(struct adapter *sc)
662 {
663 	struct sge *s = &sc->sge;
664 	struct sge_params *sp = &sc->params.sge;
665 	int i, j, n, rc = 0;
666 	uint32_t m, v, r;
667 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
668 	static int sw_buf_sizes[] = {	/* Sorted by size */
669 		MCLBYTES,
670 #if MJUMPAGESIZE != MCLBYTES
671 		MJUMPAGESIZE,
672 #endif
673 		MJUM9BYTES,
674 		MJUM16BYTES
675 	};
676 	struct sw_zone_info *swz, *safe_swz;
677 	struct hw_buf_info *hwb;
678 
679 	m = F_RXPKTCPLMODE;
680 	v = F_RXPKTCPLMODE;
681 	r = sc->params.sge.sge_control;
682 	if ((r & m) != v) {
683 		device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r);
684 		rc = EINVAL;
685 	}
686 
687 	/*
688 	 * If this changes then every single use of PAGE_SHIFT in the driver
689 	 * needs to be carefully reviewed for PAGE_SHIFT vs sp->page_shift.
690 	 */
691 	if (sp->page_shift != PAGE_SHIFT) {
692 		device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r);
693 		rc = EINVAL;
694 	}
695 
696 	/* Filter out unusable hw buffer sizes entirely (mark with -2). */
697 	hwb = &s->hw_buf_info[0];
698 	for (i = 0; i < nitems(s->hw_buf_info); i++, hwb++) {
699 		r = sc->params.sge.sge_fl_buffer_size[i];
700 		hwb->size = r;
701 		hwb->zidx = hwsz_ok(sc, r) ? -1 : -2;
702 		hwb->next = -1;
703 	}
704 
705 	/*
706 	 * Create a sorted list in decreasing order of hw buffer sizes (and so
707 	 * increasing order of spare area) for each software zone.
708 	 *
709 	 * If padding is enabled then the start and end of the buffer must align
710 	 * to the pad boundary; if packing is enabled then they must align with
711 	 * the pack boundary as well.  Allocations from the cluster zones are
712 	 * aligned to min(size, 4K), so the buffer starts at that alignment and
713 	 * ends at hwb->size alignment.  If mbuf inlining is allowed the
714 	 * starting alignment will be reduced to MSIZE and the driver will
715 	 * exercise appropriate caution when deciding on the best buffer layout
716 	 * to use.
717 	 */
718 	n = 0;	/* no usable buffer size to begin with */
719 	swz = &s->sw_zone_info[0];
720 	safe_swz = NULL;
721 	for (i = 0; i < SW_ZONE_SIZES; i++, swz++) {
722 		int8_t head = -1, tail = -1;
723 
724 		swz->size = sw_buf_sizes[i];
725 		swz->zone = m_getzone(swz->size);
726 		swz->type = m_gettype(swz->size);
727 
728 		if (swz->size < PAGE_SIZE) {
729 			MPASS(powerof2(swz->size));
730 			if (fl_pad && (swz->size % sp->pad_boundary != 0))
731 				continue;
732 		}
733 
734 		if (swz->size == safest_rx_cluster)
735 			safe_swz = swz;
736 
737 		hwb = &s->hw_buf_info[0];
738 		for (j = 0; j < SGE_FLBUF_SIZES; j++, hwb++) {
739 			if (hwb->zidx != -1 || hwb->size > swz->size)
740 				continue;
741 #ifdef INVARIANTS
742 			if (fl_pad)
743 				MPASS(hwb->size % sp->pad_boundary == 0);
744 #endif
745 			hwb->zidx = i;
746 			if (head == -1)
747 				head = tail = j;
748 			else if (hwb->size < s->hw_buf_info[tail].size) {
749 				s->hw_buf_info[tail].next = j;
750 				tail = j;
751 			} else {
752 				int8_t *cur;
753 				struct hw_buf_info *t;
754 
755 				for (cur = &head; *cur != -1; cur = &t->next) {
756 					t = &s->hw_buf_info[*cur];
757 					if (hwb->size == t->size) {
758 						hwb->zidx = -2;
759 						break;
760 					}
761 					if (hwb->size > t->size) {
762 						hwb->next = *cur;
763 						*cur = j;
764 						break;
765 					}
766 				}
767 			}
768 		}
769 		swz->head_hwidx = head;
770 		swz->tail_hwidx = tail;
771 
772 		if (tail != -1) {
773 			n++;
774 			if (swz->size - s->hw_buf_info[tail].size >=
775 			    CL_METADATA_SIZE)
776 				sc->flags |= BUF_PACKING_OK;
777 		}
778 	}
779 	if (n == 0) {
780 		device_printf(sc->dev, "no usable SGE FL buffer size.\n");
781 		rc = EINVAL;
782 	}
783 
784 	s->safe_hwidx1 = -1;
785 	s->safe_hwidx2 = -1;
786 	if (safe_swz != NULL) {
787 		s->safe_hwidx1 = safe_swz->head_hwidx;
788 		for (i = safe_swz->head_hwidx; i != -1; i = hwb->next) {
789 			int spare;
790 
791 			hwb = &s->hw_buf_info[i];
792 #ifdef INVARIANTS
793 			if (fl_pad)
794 				MPASS(hwb->size % sp->pad_boundary == 0);
795 #endif
796 			spare = safe_swz->size - hwb->size;
797 			if (spare >= CL_METADATA_SIZE) {
798 				s->safe_hwidx2 = i;
799 				break;
800 			}
801 		}
802 	}
803 
804 	if (sc->flags & IS_VF)
805 		return (0);
806 
807 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
808 	r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ);
809 	if (r != v) {
810 		device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r);
811 		rc = EINVAL;
812 	}
813 
814 	m = v = F_TDDPTAGTCB;
815 	r = t4_read_reg(sc, A_ULP_RX_CTL);
816 	if ((r & m) != v) {
817 		device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r);
818 		rc = EINVAL;
819 	}
820 
821 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
822 	    F_RESETDDPOFFSET;
823 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
824 	r = t4_read_reg(sc, A_TP_PARA_REG5);
825 	if ((r & m) != v) {
826 		device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r);
827 		rc = EINVAL;
828 	}
829 
830 	t4_init_tp_params(sc, 1);
831 
832 	t4_read_mtu_tbl(sc, sc->params.mtus, NULL);
833 	t4_load_mtus(sc, sc->params.mtus, sc->params.a_wnd, sc->params.b_wnd);
834 
835 	return (rc);
836 }
837 
838 int
839 t4_create_dma_tag(struct adapter *sc)
840 {
841 	int rc;
842 
843 	rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
844 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
845 	    BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
846 	    NULL, &sc->dmat);
847 	if (rc != 0) {
848 		device_printf(sc->dev,
849 		    "failed to create main DMA tag: %d\n", rc);
850 	}
851 
852 	return (rc);
853 }
854 
855 void
856 t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
857     struct sysctl_oid_list *children)
858 {
859 	struct sge_params *sp = &sc->params.sge;
860 
861 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "buffer_sizes",
862 	    CTLTYPE_STRING | CTLFLAG_RD, &sc->sge, 0, sysctl_bufsizes, "A",
863 	    "freelist buffer sizes");
864 
865 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD,
866 	    NULL, sp->fl_pktshift, "payload DMA offset in rx buffer (bytes)");
867 
868 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD,
869 	    NULL, sp->pad_boundary, "payload pad boundary (bytes)");
870 
871 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD,
872 	    NULL, sp->spg_len, "status page size (bytes)");
873 
874 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD,
875 	    NULL, cong_drop, "congestion drop setting");
876 
877 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD,
878 	    NULL, sp->pack_boundary, "payload pack boundary (bytes)");
879 }
880 
881 int
882 t4_destroy_dma_tag(struct adapter *sc)
883 {
884 	if (sc->dmat)
885 		bus_dma_tag_destroy(sc->dmat);
886 
887 	return (0);
888 }
889 
890 /*
891  * Allocate and initialize the firmware event queue and the management queue.
892  *
893  * Returns errno on failure.  Resources allocated up to that point may still be
894  * allocated.  Caller is responsible for cleanup in case this function fails.
895  */
896 int
897 t4_setup_adapter_queues(struct adapter *sc)
898 {
899 	int rc;
900 
901 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
902 
903 	sysctl_ctx_init(&sc->ctx);
904 	sc->flags |= ADAP_SYSCTL_CTX;
905 
906 	/*
907 	 * Firmware event queue
908 	 */
909 	rc = alloc_fwq(sc);
910 	if (rc != 0)
911 		return (rc);
912 
913 	/*
914 	 * Management queue.  This is just a control queue that uses the fwq as
915 	 * its associated iq.
916 	 */
917 	if (!(sc->flags & IS_VF))
918 		rc = alloc_mgmtq(sc);
919 
920 	return (rc);
921 }
922 
923 /*
924  * Idempotent
925  */
926 int
927 t4_teardown_adapter_queues(struct adapter *sc)
928 {
929 
930 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
931 
932 	/* Do this before freeing the queue */
933 	if (sc->flags & ADAP_SYSCTL_CTX) {
934 		sysctl_ctx_free(&sc->ctx);
935 		sc->flags &= ~ADAP_SYSCTL_CTX;
936 	}
937 
938 	free_mgmtq(sc);
939 	free_fwq(sc);
940 
941 	return (0);
942 }
943 
944 static inline int
945 first_vector(struct vi_info *vi)
946 {
947 	struct adapter *sc = vi->pi->adapter;
948 
949 	if (sc->intr_count == 1)
950 		return (0);
951 
952 	return (vi->first_intr);
953 }
954 
955 /*
956  * Given an arbitrary "index," come up with an iq that can be used by other
957  * queues (of this VI) for interrupt forwarding, SGE egress updates, etc.
958  * The iq returned is guaranteed to be something that takes direct interrupts.
959  */
960 static struct sge_iq *
961 vi_intr_iq(struct vi_info *vi, int idx)
962 {
963 	struct adapter *sc = vi->pi->adapter;
964 	struct sge *s = &sc->sge;
965 	struct sge_iq *iq = NULL;
966 	int nintr, i;
967 
968 	if (sc->intr_count == 1)
969 		return (&sc->sge.fwq);
970 
971 	nintr = vi->nintr;
972 #ifdef DEV_NETMAP
973 	/* Do not consider any netmap-only interrupts */
974 	if (vi->flags & INTR_RXQ && vi->nnmrxq > vi->nrxq)
975 		nintr -= vi->nnmrxq - vi->nrxq;
976 #endif
977 	KASSERT(nintr != 0,
978 	    ("%s: vi %p has no exclusive interrupts, total interrupts = %d",
979 	    __func__, vi, sc->intr_count));
980 	i = idx % nintr;
981 
982 	if (vi->flags & INTR_RXQ) {
983 	       	if (i < vi->nrxq) {
984 			iq = &s->rxq[vi->first_rxq + i].iq;
985 			goto done;
986 		}
987 		i -= vi->nrxq;
988 	}
989 #ifdef TCP_OFFLOAD
990 	if (vi->flags & INTR_OFLD_RXQ) {
991 	       	if (i < vi->nofldrxq) {
992 			iq = &s->ofld_rxq[vi->first_ofld_rxq + i].iq;
993 			goto done;
994 		}
995 		i -= vi->nofldrxq;
996 	}
997 #endif
998 	panic("%s: vi %p, intr_flags 0x%lx, idx %d, total intr %d\n", __func__,
999 	    vi, vi->flags & INTR_ALL, idx, nintr);
1000 done:
1001 	MPASS(iq != NULL);
1002 	KASSERT(iq->flags & IQ_INTR,
1003 	    ("%s: iq %p (vi %p, intr_flags 0x%lx, idx %d)", __func__, iq, vi,
1004 	    vi->flags & INTR_ALL, idx));
1005 	return (iq);
1006 }
1007 
1008 /* Maximum payload that can be delivered with a single iq descriptor */
1009 static inline int
1010 mtu_to_max_payload(struct adapter *sc, int mtu, const int toe)
1011 {
1012 	int payload;
1013 
1014 #ifdef TCP_OFFLOAD
1015 	if (toe) {
1016 		payload = sc->tt.rx_coalesce ?
1017 		    G_RXCOALESCESIZE(t4_read_reg(sc, A_TP_PARA_REG2)) : mtu;
1018 	} else {
1019 #endif
1020 		/* large enough even when hw VLAN extraction is disabled */
1021 		payload = sc->params.sge.fl_pktshift + ETHER_HDR_LEN +
1022 		    ETHER_VLAN_ENCAP_LEN + mtu;
1023 #ifdef TCP_OFFLOAD
1024 	}
1025 #endif
1026 
1027 	return (payload);
1028 }
1029 
1030 int
1031 t4_setup_vi_queues(struct vi_info *vi)
1032 {
1033 	int rc = 0, i, j, intr_idx, iqid;
1034 	struct sge_rxq *rxq;
1035 	struct sge_txq *txq;
1036 	struct sge_wrq *ctrlq;
1037 #ifdef TCP_OFFLOAD
1038 	struct sge_ofld_rxq *ofld_rxq;
1039 	struct sge_wrq *ofld_txq;
1040 #endif
1041 #ifdef DEV_NETMAP
1042 	int saved_idx;
1043 	struct sge_nm_rxq *nm_rxq;
1044 	struct sge_nm_txq *nm_txq;
1045 #endif
1046 	char name[16];
1047 	struct port_info *pi = vi->pi;
1048 	struct adapter *sc = pi->adapter;
1049 	struct ifnet *ifp = vi->ifp;
1050 	struct sysctl_oid *oid = device_get_sysctl_tree(vi->dev);
1051 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
1052 	int maxp, mtu = ifp->if_mtu;
1053 
1054 	/* Interrupt vector to start from (when using multiple vectors) */
1055 	intr_idx = first_vector(vi);
1056 
1057 #ifdef DEV_NETMAP
1058 	saved_idx = intr_idx;
1059 	if (ifp->if_capabilities & IFCAP_NETMAP) {
1060 
1061 		/* netmap is supported with direct interrupts only. */
1062 		MPASS(vi->flags & INTR_RXQ);
1063 
1064 		/*
1065 		 * We don't have buffers to back the netmap rx queues
1066 		 * right now so we create the queues in a way that
1067 		 * doesn't set off any congestion signal in the chip.
1068 		 */
1069 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_rxq",
1070 		    CTLFLAG_RD, NULL, "rx queues");
1071 		for_each_nm_rxq(vi, i, nm_rxq) {
1072 			rc = alloc_nm_rxq(vi, nm_rxq, intr_idx, i, oid);
1073 			if (rc != 0)
1074 				goto done;
1075 			intr_idx++;
1076 		}
1077 
1078 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_txq",
1079 		    CTLFLAG_RD, NULL, "tx queues");
1080 		for_each_nm_txq(vi, i, nm_txq) {
1081 			iqid = vi->first_nm_rxq + (i % vi->nnmrxq);
1082 			rc = alloc_nm_txq(vi, nm_txq, iqid, i, oid);
1083 			if (rc != 0)
1084 				goto done;
1085 		}
1086 	}
1087 
1088 	/* Normal rx queues and netmap rx queues share the same interrupts. */
1089 	intr_idx = saved_idx;
1090 #endif
1091 
1092 	/*
1093 	 * First pass over all NIC and TOE rx queues:
1094 	 * a) initialize iq and fl
1095 	 * b) allocate queue iff it will take direct interrupts.
1096 	 */
1097 	maxp = mtu_to_max_payload(sc, mtu, 0);
1098 	if (vi->flags & INTR_RXQ) {
1099 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
1100 		    CTLFLAG_RD, NULL, "rx queues");
1101 	}
1102 	for_each_rxq(vi, i, rxq) {
1103 
1104 		init_iq(&rxq->iq, sc, vi->tmr_idx, vi->pktc_idx, vi->qsize_rxq);
1105 
1106 		snprintf(name, sizeof(name), "%s rxq%d-fl",
1107 		    device_get_nameunit(vi->dev), i);
1108 		init_fl(sc, &rxq->fl, vi->qsize_rxq / 8, maxp, name);
1109 
1110 		if (vi->flags & INTR_RXQ) {
1111 			rxq->iq.flags |= IQ_INTR;
1112 			rc = alloc_rxq(vi, rxq, intr_idx, i, oid);
1113 			if (rc != 0)
1114 				goto done;
1115 			intr_idx++;
1116 		}
1117 	}
1118 #ifdef DEV_NETMAP
1119 	if (ifp->if_capabilities & IFCAP_NETMAP)
1120 		intr_idx = saved_idx + max(vi->nrxq, vi->nnmrxq);
1121 #endif
1122 #ifdef TCP_OFFLOAD
1123 	maxp = mtu_to_max_payload(sc, mtu, 1);
1124 	if (vi->flags & INTR_OFLD_RXQ) {
1125 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq",
1126 		    CTLFLAG_RD, NULL,
1127 		    "rx queues for offloaded TCP connections");
1128 	}
1129 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1130 
1131 		init_iq(&ofld_rxq->iq, sc, vi->tmr_idx, vi->pktc_idx,
1132 		    vi->qsize_rxq);
1133 
1134 		snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
1135 		    device_get_nameunit(vi->dev), i);
1136 		init_fl(sc, &ofld_rxq->fl, vi->qsize_rxq / 8, maxp, name);
1137 
1138 		if (vi->flags & INTR_OFLD_RXQ) {
1139 			ofld_rxq->iq.flags |= IQ_INTR;
1140 			rc = alloc_ofld_rxq(vi, ofld_rxq, intr_idx, i, oid);
1141 			if (rc != 0)
1142 				goto done;
1143 			intr_idx++;
1144 		}
1145 	}
1146 #endif
1147 
1148 	/*
1149 	 * Second pass over all NIC and TOE rx queues.  The queues forwarding
1150 	 * their interrupts are allocated now.
1151 	 */
1152 	j = 0;
1153 	if (!(vi->flags & INTR_RXQ)) {
1154 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
1155 		    CTLFLAG_RD, NULL, "rx queues");
1156 		for_each_rxq(vi, i, rxq) {
1157 			MPASS(!(rxq->iq.flags & IQ_INTR));
1158 
1159 			intr_idx = vi_intr_iq(vi, j)->abs_id;
1160 
1161 			rc = alloc_rxq(vi, rxq, intr_idx, i, oid);
1162 			if (rc != 0)
1163 				goto done;
1164 			j++;
1165 		}
1166 	}
1167 #ifdef TCP_OFFLOAD
1168 	if (vi->nofldrxq != 0 && !(vi->flags & INTR_OFLD_RXQ)) {
1169 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq",
1170 		    CTLFLAG_RD, NULL,
1171 		    "rx queues for offloaded TCP connections");
1172 		for_each_ofld_rxq(vi, i, ofld_rxq) {
1173 			MPASS(!(ofld_rxq->iq.flags & IQ_INTR));
1174 
1175 			intr_idx = vi_intr_iq(vi, j)->abs_id;
1176 
1177 			rc = alloc_ofld_rxq(vi, ofld_rxq, intr_idx, i, oid);
1178 			if (rc != 0)
1179 				goto done;
1180 			j++;
1181 		}
1182 	}
1183 #endif
1184 
1185 	/*
1186 	 * Now the tx queues.  Only one pass needed.
1187 	 */
1188 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "txq", CTLFLAG_RD,
1189 	    NULL, "tx queues");
1190 	j = 0;
1191 	for_each_txq(vi, i, txq) {
1192 		iqid = vi_intr_iq(vi, j)->cntxt_id;
1193 		snprintf(name, sizeof(name), "%s txq%d",
1194 		    device_get_nameunit(vi->dev), i);
1195 		init_eq(sc, &txq->eq, EQ_ETH, vi->qsize_txq, pi->tx_chan, iqid,
1196 		    name);
1197 
1198 		rc = alloc_txq(vi, txq, i, oid);
1199 		if (rc != 0)
1200 			goto done;
1201 		j++;
1202 	}
1203 #ifdef TCP_OFFLOAD
1204 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_txq",
1205 	    CTLFLAG_RD, NULL, "tx queues for offloaded TCP connections");
1206 	for_each_ofld_txq(vi, i, ofld_txq) {
1207 		struct sysctl_oid *oid2;
1208 
1209 		iqid = vi_intr_iq(vi, j)->cntxt_id;
1210 		snprintf(name, sizeof(name), "%s ofld_txq%d",
1211 		    device_get_nameunit(vi->dev), i);
1212 		init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq, pi->tx_chan,
1213 		    iqid, name);
1214 
1215 		snprintf(name, sizeof(name), "%d", i);
1216 		oid2 = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
1217 		    name, CTLFLAG_RD, NULL, "offload tx queue");
1218 
1219 		rc = alloc_wrq(sc, vi, ofld_txq, oid2);
1220 		if (rc != 0)
1221 			goto done;
1222 		j++;
1223 	}
1224 #endif
1225 
1226 	/*
1227 	 * Finally, the control queue.
1228 	 */
1229 	if (!IS_MAIN_VI(vi) || sc->flags & IS_VF)
1230 		goto done;
1231 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ctrlq", CTLFLAG_RD,
1232 	    NULL, "ctrl queue");
1233 	ctrlq = &sc->sge.ctrlq[pi->port_id];
1234 	iqid = vi_intr_iq(vi, 0)->cntxt_id;
1235 	snprintf(name, sizeof(name), "%s ctrlq", device_get_nameunit(vi->dev));
1236 	init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, pi->tx_chan, iqid,
1237 	    name);
1238 	rc = alloc_wrq(sc, vi, ctrlq, oid);
1239 
1240 done:
1241 	if (rc)
1242 		t4_teardown_vi_queues(vi);
1243 
1244 	return (rc);
1245 }
1246 
1247 /*
1248  * Idempotent
1249  */
1250 int
1251 t4_teardown_vi_queues(struct vi_info *vi)
1252 {
1253 	int i;
1254 	struct port_info *pi = vi->pi;
1255 	struct adapter *sc = pi->adapter;
1256 	struct sge_rxq *rxq;
1257 	struct sge_txq *txq;
1258 #ifdef TCP_OFFLOAD
1259 	struct sge_ofld_rxq *ofld_rxq;
1260 	struct sge_wrq *ofld_txq;
1261 #endif
1262 #ifdef DEV_NETMAP
1263 	struct sge_nm_rxq *nm_rxq;
1264 	struct sge_nm_txq *nm_txq;
1265 #endif
1266 
1267 	/* Do this before freeing the queues */
1268 	if (vi->flags & VI_SYSCTL_CTX) {
1269 		sysctl_ctx_free(&vi->ctx);
1270 		vi->flags &= ~VI_SYSCTL_CTX;
1271 	}
1272 
1273 #ifdef DEV_NETMAP
1274 	if (vi->ifp->if_capabilities & IFCAP_NETMAP) {
1275 		for_each_nm_txq(vi, i, nm_txq) {
1276 			free_nm_txq(vi, nm_txq);
1277 		}
1278 
1279 		for_each_nm_rxq(vi, i, nm_rxq) {
1280 			free_nm_rxq(vi, nm_rxq);
1281 		}
1282 	}
1283 #endif
1284 
1285 	/*
1286 	 * Take down all the tx queues first, as they reference the rx queues
1287 	 * (for egress updates, etc.).
1288 	 */
1289 
1290 	if (IS_MAIN_VI(vi) && !(sc->flags & IS_VF))
1291 		free_wrq(sc, &sc->sge.ctrlq[pi->port_id]);
1292 
1293 	for_each_txq(vi, i, txq) {
1294 		free_txq(vi, txq);
1295 	}
1296 #ifdef TCP_OFFLOAD
1297 	for_each_ofld_txq(vi, i, ofld_txq) {
1298 		free_wrq(sc, ofld_txq);
1299 	}
1300 #endif
1301 
1302 	/*
1303 	 * Then take down the rx queues that forward their interrupts, as they
1304 	 * reference other rx queues.
1305 	 */
1306 
1307 	for_each_rxq(vi, i, rxq) {
1308 		if ((rxq->iq.flags & IQ_INTR) == 0)
1309 			free_rxq(vi, rxq);
1310 	}
1311 #ifdef TCP_OFFLOAD
1312 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1313 		if ((ofld_rxq->iq.flags & IQ_INTR) == 0)
1314 			free_ofld_rxq(vi, ofld_rxq);
1315 	}
1316 #endif
1317 
1318 	/*
1319 	 * Then take down the rx queues that take direct interrupts.
1320 	 */
1321 
1322 	for_each_rxq(vi, i, rxq) {
1323 		if (rxq->iq.flags & IQ_INTR)
1324 			free_rxq(vi, rxq);
1325 	}
1326 #ifdef TCP_OFFLOAD
1327 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1328 		if (ofld_rxq->iq.flags & IQ_INTR)
1329 			free_ofld_rxq(vi, ofld_rxq);
1330 	}
1331 #endif
1332 
1333 	return (0);
1334 }
1335 
1336 /*
1337  * Deals with errors and the firmware event queue.  All data rx queues forward
1338  * their interrupt to the firmware event queue.
1339  */
1340 void
1341 t4_intr_all(void *arg)
1342 {
1343 	struct adapter *sc = arg;
1344 	struct sge_iq *fwq = &sc->sge.fwq;
1345 
1346 	t4_intr_err(arg);
1347 	if (atomic_cmpset_int(&fwq->state, IQS_IDLE, IQS_BUSY)) {
1348 		service_iq(fwq, 0);
1349 		atomic_cmpset_int(&fwq->state, IQS_BUSY, IQS_IDLE);
1350 	}
1351 }
1352 
1353 /* Deals with error interrupts */
1354 void
1355 t4_intr_err(void *arg)
1356 {
1357 	struct adapter *sc = arg;
1358 
1359 	t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
1360 	t4_slow_intr_handler(sc);
1361 }
1362 
1363 void
1364 t4_intr_evt(void *arg)
1365 {
1366 	struct sge_iq *iq = arg;
1367 
1368 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1369 		service_iq(iq, 0);
1370 		atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1371 	}
1372 }
1373 
1374 void
1375 t4_intr(void *arg)
1376 {
1377 	struct sge_iq *iq = arg;
1378 
1379 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1380 		service_iq(iq, 0);
1381 		atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1382 	}
1383 }
1384 
1385 void
1386 t4_vi_intr(void *arg)
1387 {
1388 	struct irq *irq = arg;
1389 
1390 #ifdef DEV_NETMAP
1391 	if (atomic_cmpset_int(&irq->nm_state, NM_ON, NM_BUSY)) {
1392 		t4_nm_intr(irq->nm_rxq);
1393 		atomic_cmpset_int(&irq->nm_state, NM_BUSY, NM_ON);
1394 	}
1395 #endif
1396 	if (irq->rxq != NULL)
1397 		t4_intr(irq->rxq);
1398 }
1399 
1400 static inline int
1401 sort_before_lro(struct lro_ctrl *lro)
1402 {
1403 
1404 	return (lro->lro_mbuf_max != 0);
1405 }
1406 
1407 /*
1408  * Deals with anything and everything on the given ingress queue.
1409  */
1410 static int
1411 service_iq(struct sge_iq *iq, int budget)
1412 {
1413 	struct sge_iq *q;
1414 	struct sge_rxq *rxq = iq_to_rxq(iq);	/* Use iff iq is part of rxq */
1415 	struct sge_fl *fl;			/* Use iff IQ_HAS_FL */
1416 	struct adapter *sc = iq->adapter;
1417 	struct iq_desc *d = &iq->desc[iq->cidx];
1418 	int ndescs = 0, limit;
1419 	int rsp_type, refill;
1420 	uint32_t lq;
1421 	uint16_t fl_hw_cidx;
1422 	struct mbuf *m0;
1423 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1424 #if defined(INET) || defined(INET6)
1425 	const struct timeval lro_timeout = {0, sc->lro_timeout};
1426 	struct lro_ctrl *lro = &rxq->lro;
1427 #endif
1428 
1429 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1430 
1431 	limit = budget ? budget : iq->qsize / 16;
1432 
1433 	if (iq->flags & IQ_HAS_FL) {
1434 		fl = &rxq->fl;
1435 		fl_hw_cidx = fl->hw_cidx;	/* stable snapshot */
1436 	} else {
1437 		fl = NULL;
1438 		fl_hw_cidx = 0;			/* to silence gcc warning */
1439 	}
1440 
1441 #if defined(INET) || defined(INET6)
1442 	if (iq->flags & IQ_ADJ_CREDIT) {
1443 		MPASS(sort_before_lro(lro));
1444 		iq->flags &= ~IQ_ADJ_CREDIT;
1445 		if ((d->rsp.u.type_gen & F_RSPD_GEN) != iq->gen) {
1446 			tcp_lro_flush_all(lro);
1447 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(1) |
1448 			    V_INGRESSQID((u32)iq->cntxt_id) |
1449 			    V_SEINTARM(iq->intr_params));
1450 			return (0);
1451 		}
1452 		ndescs = 1;
1453 	}
1454 #else
1455 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1456 #endif
1457 
1458 	/*
1459 	 * We always come back and check the descriptor ring for new indirect
1460 	 * interrupts and other responses after running a single handler.
1461 	 */
1462 	for (;;) {
1463 		while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1464 
1465 			rmb();
1466 
1467 			refill = 0;
1468 			m0 = NULL;
1469 			rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1470 			lq = be32toh(d->rsp.pldbuflen_qid);
1471 
1472 			switch (rsp_type) {
1473 			case X_RSPD_TYPE_FLBUF:
1474 
1475 				KASSERT(iq->flags & IQ_HAS_FL,
1476 				    ("%s: data for an iq (%p) with no freelist",
1477 				    __func__, iq));
1478 
1479 				m0 = get_fl_payload(sc, fl, lq);
1480 				if (__predict_false(m0 == NULL))
1481 					goto process_iql;
1482 				refill = IDXDIFF(fl->hw_cidx, fl_hw_cidx, fl->sidx) > 2;
1483 #ifdef T4_PKT_TIMESTAMP
1484 				/*
1485 				 * 60 bit timestamp for the payload is
1486 				 * *(uint64_t *)m0->m_pktdat.  Note that it is
1487 				 * in the leading free-space in the mbuf.  The
1488 				 * kernel can clobber it during a pullup,
1489 				 * m_copymdata, etc.  You need to make sure that
1490 				 * the mbuf reaches you unmolested if you care
1491 				 * about the timestamp.
1492 				 */
1493 				*(uint64_t *)m0->m_pktdat =
1494 				    be64toh(ctrl->u.last_flit) &
1495 				    0xfffffffffffffff;
1496 #endif
1497 
1498 				/* fall through */
1499 
1500 			case X_RSPD_TYPE_CPL:
1501 				KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1502 				    ("%s: bad opcode %02x.", __func__,
1503 				    d->rss.opcode));
1504 				t4_cpl_handler[d->rss.opcode](iq, &d->rss, m0);
1505 				break;
1506 
1507 			case X_RSPD_TYPE_INTR:
1508 
1509 				/*
1510 				 * Interrupts should be forwarded only to queues
1511 				 * that are not forwarding their interrupts.
1512 				 * This means service_iq can recurse but only 1
1513 				 * level deep.
1514 				 */
1515 				KASSERT(budget == 0,
1516 				    ("%s: budget %u, rsp_type %u", __func__,
1517 				    budget, rsp_type));
1518 
1519 				/*
1520 				 * There are 1K interrupt-capable queues (qids 0
1521 				 * through 1023).  A response type indicating a
1522 				 * forwarded interrupt with a qid >= 1K is an
1523 				 * iWARP async notification.
1524 				 */
1525 				if (lq >= 1024) {
1526                                         t4_an_handler(iq, &d->rsp);
1527                                         break;
1528                                 }
1529 
1530 				q = sc->sge.iqmap[lq - sc->sge.iq_start -
1531 				    sc->sge.iq_base];
1532 				if (atomic_cmpset_int(&q->state, IQS_IDLE,
1533 				    IQS_BUSY)) {
1534 					if (service_iq(q, q->qsize / 16) == 0) {
1535 						atomic_cmpset_int(&q->state,
1536 						    IQS_BUSY, IQS_IDLE);
1537 					} else {
1538 						STAILQ_INSERT_TAIL(&iql, q,
1539 						    link);
1540 					}
1541 				}
1542 				break;
1543 
1544 			default:
1545 				KASSERT(0,
1546 				    ("%s: illegal response type %d on iq %p",
1547 				    __func__, rsp_type, iq));
1548 				log(LOG_ERR,
1549 				    "%s: illegal response type %d on iq %p",
1550 				    device_get_nameunit(sc->dev), rsp_type, iq);
1551 				break;
1552 			}
1553 
1554 			d++;
1555 			if (__predict_false(++iq->cidx == iq->sidx)) {
1556 				iq->cidx = 0;
1557 				iq->gen ^= F_RSPD_GEN;
1558 				d = &iq->desc[0];
1559 			}
1560 			if (__predict_false(++ndescs == limit)) {
1561 				t4_write_reg(sc, sc->sge_gts_reg,
1562 				    V_CIDXINC(ndescs) |
1563 				    V_INGRESSQID(iq->cntxt_id) |
1564 				    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1565 				ndescs = 0;
1566 
1567 #if defined(INET) || defined(INET6)
1568 				if (iq->flags & IQ_LRO_ENABLED &&
1569 				    !sort_before_lro(lro) &&
1570 				    sc->lro_timeout != 0) {
1571 					tcp_lro_flush_inactive(lro,
1572 					    &lro_timeout);
1573 				}
1574 #endif
1575 
1576 				if (budget) {
1577 					if (iq->flags & IQ_HAS_FL) {
1578 						FL_LOCK(fl);
1579 						refill_fl(sc, fl, 32);
1580 						FL_UNLOCK(fl);
1581 					}
1582 					return (EINPROGRESS);
1583 				}
1584 			}
1585 			if (refill) {
1586 				FL_LOCK(fl);
1587 				refill_fl(sc, fl, 32);
1588 				FL_UNLOCK(fl);
1589 				fl_hw_cidx = fl->hw_cidx;
1590 			}
1591 		}
1592 
1593 process_iql:
1594 		if (STAILQ_EMPTY(&iql))
1595 			break;
1596 
1597 		/*
1598 		 * Process the head only, and send it to the back of the list if
1599 		 * it's still not done.
1600 		 */
1601 		q = STAILQ_FIRST(&iql);
1602 		STAILQ_REMOVE_HEAD(&iql, link);
1603 		if (service_iq(q, q->qsize / 8) == 0)
1604 			atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1605 		else
1606 			STAILQ_INSERT_TAIL(&iql, q, link);
1607 	}
1608 
1609 #if defined(INET) || defined(INET6)
1610 	if (iq->flags & IQ_LRO_ENABLED) {
1611 		if (ndescs > 0 && lro->lro_mbuf_count > 8) {
1612 			MPASS(sort_before_lro(lro));
1613 			/* hold back one credit and don't flush LRO state */
1614 			iq->flags |= IQ_ADJ_CREDIT;
1615 			ndescs--;
1616 		} else {
1617 			tcp_lro_flush_all(lro);
1618 		}
1619 	}
1620 #endif
1621 
1622 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1623 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1624 
1625 	if (iq->flags & IQ_HAS_FL) {
1626 		int starved;
1627 
1628 		FL_LOCK(fl);
1629 		starved = refill_fl(sc, fl, 64);
1630 		FL_UNLOCK(fl);
1631 		if (__predict_false(starved != 0))
1632 			add_fl_to_sfl(sc, fl);
1633 	}
1634 
1635 	return (0);
1636 }
1637 
1638 static inline int
1639 cl_has_metadata(struct sge_fl *fl, struct cluster_layout *cll)
1640 {
1641 	int rc = fl->flags & FL_BUF_PACKING || cll->region1 > 0;
1642 
1643 	if (rc)
1644 		MPASS(cll->region3 >= CL_METADATA_SIZE);
1645 
1646 	return (rc);
1647 }
1648 
1649 static inline struct cluster_metadata *
1650 cl_metadata(struct adapter *sc, struct sge_fl *fl, struct cluster_layout *cll,
1651     caddr_t cl)
1652 {
1653 
1654 	if (cl_has_metadata(fl, cll)) {
1655 		struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1656 
1657 		return ((struct cluster_metadata *)(cl + swz->size) - 1);
1658 	}
1659 	return (NULL);
1660 }
1661 
1662 static void
1663 rxb_free(struct mbuf *m, void *arg1, void *arg2)
1664 {
1665 	uma_zone_t zone = arg1;
1666 	caddr_t cl = arg2;
1667 
1668 	uma_zfree(zone, cl);
1669 	counter_u64_add(extfree_rels, 1);
1670 }
1671 
1672 /*
1673  * The mbuf returned by this function could be allocated from zone_mbuf or
1674  * constructed in spare room in the cluster.
1675  *
1676  * The mbuf carries the payload in one of these ways
1677  * a) frame inside the mbuf (mbuf from zone_mbuf)
1678  * b) m_cljset (for clusters without metadata) zone_mbuf
1679  * c) m_extaddref (cluster with metadata) inline mbuf
1680  * d) m_extaddref (cluster with metadata) zone_mbuf
1681  */
1682 static struct mbuf *
1683 get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1684     int remaining)
1685 {
1686 	struct mbuf *m;
1687 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1688 	struct cluster_layout *cll = &sd->cll;
1689 	struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1690 	struct hw_buf_info *hwb = &sc->sge.hw_buf_info[cll->hwidx];
1691 	struct cluster_metadata *clm = cl_metadata(sc, fl, cll, sd->cl);
1692 	int len, blen;
1693 	caddr_t payload;
1694 
1695 	blen = hwb->size - fl->rx_offset;	/* max possible in this buf */
1696 	len = min(remaining, blen);
1697 	payload = sd->cl + cll->region1 + fl->rx_offset;
1698 	if (fl->flags & FL_BUF_PACKING) {
1699 		const u_int l = fr_offset + len;
1700 		const u_int pad = roundup2(l, fl->buf_boundary) - l;
1701 
1702 		if (fl->rx_offset + len + pad < hwb->size)
1703 			blen = len + pad;
1704 		MPASS(fl->rx_offset + blen <= hwb->size);
1705 	} else {
1706 		MPASS(fl->rx_offset == 0);	/* not packing */
1707 	}
1708 
1709 
1710 	if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) {
1711 
1712 		/*
1713 		 * Copy payload into a freshly allocated mbuf.
1714 		 */
1715 
1716 		m = fr_offset == 0 ?
1717 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1718 		if (m == NULL)
1719 			return (NULL);
1720 		fl->mbuf_allocated++;
1721 #ifdef T4_PKT_TIMESTAMP
1722 		/* Leave room for a timestamp */
1723 		m->m_data += 8;
1724 #endif
1725 		/* copy data to mbuf */
1726 		bcopy(payload, mtod(m, caddr_t), len);
1727 
1728 	} else if (sd->nmbuf * MSIZE < cll->region1) {
1729 
1730 		/*
1731 		 * There's spare room in the cluster for an mbuf.  Create one
1732 		 * and associate it with the payload that's in the cluster.
1733 		 */
1734 
1735 		MPASS(clm != NULL);
1736 		m = (struct mbuf *)(sd->cl + sd->nmbuf * MSIZE);
1737 		/* No bzero required */
1738 		if (m_init(m, M_NOWAIT, MT_DATA,
1739 		    fr_offset == 0 ? M_PKTHDR | M_NOFREE : M_NOFREE))
1740 			return (NULL);
1741 		fl->mbuf_inlined++;
1742 		m_extaddref(m, payload, blen, &clm->refcount, rxb_free,
1743 		    swz->zone, sd->cl);
1744 		if (sd->nmbuf++ == 0)
1745 			counter_u64_add(extfree_refs, 1);
1746 
1747 	} else {
1748 
1749 		/*
1750 		 * Grab an mbuf from zone_mbuf and associate it with the
1751 		 * payload in the cluster.
1752 		 */
1753 
1754 		m = fr_offset == 0 ?
1755 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1756 		if (m == NULL)
1757 			return (NULL);
1758 		fl->mbuf_allocated++;
1759 		if (clm != NULL) {
1760 			m_extaddref(m, payload, blen, &clm->refcount,
1761 			    rxb_free, swz->zone, sd->cl);
1762 			if (sd->nmbuf++ == 0)
1763 				counter_u64_add(extfree_refs, 1);
1764 		} else {
1765 			m_cljset(m, sd->cl, swz->type);
1766 			sd->cl = NULL;	/* consumed, not a recycle candidate */
1767 		}
1768 	}
1769 	if (fr_offset == 0)
1770 		m->m_pkthdr.len = remaining;
1771 	m->m_len = len;
1772 
1773 	if (fl->flags & FL_BUF_PACKING) {
1774 		fl->rx_offset += blen;
1775 		MPASS(fl->rx_offset <= hwb->size);
1776 		if (fl->rx_offset < hwb->size)
1777 			return (m);	/* without advancing the cidx */
1778 	}
1779 
1780 	if (__predict_false(++fl->cidx % 8 == 0)) {
1781 		uint16_t cidx = fl->cidx / 8;
1782 
1783 		if (__predict_false(cidx == fl->sidx))
1784 			fl->cidx = cidx = 0;
1785 		fl->hw_cidx = cidx;
1786 	}
1787 	fl->rx_offset = 0;
1788 
1789 	return (m);
1790 }
1791 
1792 static struct mbuf *
1793 get_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf)
1794 {
1795 	struct mbuf *m0, *m, **pnext;
1796 	u_int remaining;
1797 	const u_int total = G_RSPD_LEN(len_newbuf);
1798 
1799 	if (__predict_false(fl->flags & FL_BUF_RESUME)) {
1800 		M_ASSERTPKTHDR(fl->m0);
1801 		MPASS(fl->m0->m_pkthdr.len == total);
1802 		MPASS(fl->remaining < total);
1803 
1804 		m0 = fl->m0;
1805 		pnext = fl->pnext;
1806 		remaining = fl->remaining;
1807 		fl->flags &= ~FL_BUF_RESUME;
1808 		goto get_segment;
1809 	}
1810 
1811 	if (fl->rx_offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
1812 		fl->rx_offset = 0;
1813 		if (__predict_false(++fl->cidx % 8 == 0)) {
1814 			uint16_t cidx = fl->cidx / 8;
1815 
1816 			if (__predict_false(cidx == fl->sidx))
1817 				fl->cidx = cidx = 0;
1818 			fl->hw_cidx = cidx;
1819 		}
1820 	}
1821 
1822 	/*
1823 	 * Payload starts at rx_offset in the current hw buffer.  Its length is
1824 	 * 'len' and it may span multiple hw buffers.
1825 	 */
1826 
1827 	m0 = get_scatter_segment(sc, fl, 0, total);
1828 	if (m0 == NULL)
1829 		return (NULL);
1830 	remaining = total - m0->m_len;
1831 	pnext = &m0->m_next;
1832 	while (remaining > 0) {
1833 get_segment:
1834 		MPASS(fl->rx_offset == 0);
1835 		m = get_scatter_segment(sc, fl, total - remaining, remaining);
1836 		if (__predict_false(m == NULL)) {
1837 			fl->m0 = m0;
1838 			fl->pnext = pnext;
1839 			fl->remaining = remaining;
1840 			fl->flags |= FL_BUF_RESUME;
1841 			return (NULL);
1842 		}
1843 		*pnext = m;
1844 		pnext = &m->m_next;
1845 		remaining -= m->m_len;
1846 	}
1847 	*pnext = NULL;
1848 
1849 	M_ASSERTPKTHDR(m0);
1850 	return (m0);
1851 }
1852 
1853 static int
1854 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
1855 {
1856 	struct sge_rxq *rxq = iq_to_rxq(iq);
1857 	struct ifnet *ifp = rxq->ifp;
1858 	struct adapter *sc = iq->adapter;
1859 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
1860 #if defined(INET) || defined(INET6)
1861 	struct lro_ctrl *lro = &rxq->lro;
1862 #endif
1863 	static const int sw_hashtype[4][2] = {
1864 		{M_HASHTYPE_NONE, M_HASHTYPE_NONE},
1865 		{M_HASHTYPE_RSS_IPV4, M_HASHTYPE_RSS_IPV6},
1866 		{M_HASHTYPE_RSS_TCP_IPV4, M_HASHTYPE_RSS_TCP_IPV6},
1867 		{M_HASHTYPE_RSS_UDP_IPV4, M_HASHTYPE_RSS_UDP_IPV6},
1868 	};
1869 
1870 	KASSERT(m0 != NULL, ("%s: no payload with opcode %02x", __func__,
1871 	    rss->opcode));
1872 
1873 	m0->m_pkthdr.len -= sc->params.sge.fl_pktshift;
1874 	m0->m_len -= sc->params.sge.fl_pktshift;
1875 	m0->m_data += sc->params.sge.fl_pktshift;
1876 
1877 	m0->m_pkthdr.rcvif = ifp;
1878 	M_HASHTYPE_SET(m0, sw_hashtype[rss->hash_type][rss->ipv6]);
1879 	m0->m_pkthdr.flowid = be32toh(rss->hash_val);
1880 
1881 	if (cpl->csum_calc && !(cpl->err_vec & sc->params.tp.err_vec_mask)) {
1882 		if (ifp->if_capenable & IFCAP_RXCSUM &&
1883 		    cpl->l2info & htobe32(F_RXF_IP)) {
1884 			m0->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
1885 			    CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
1886 			rxq->rxcsum++;
1887 		} else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
1888 		    cpl->l2info & htobe32(F_RXF_IP6)) {
1889 			m0->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
1890 			    CSUM_PSEUDO_HDR);
1891 			rxq->rxcsum++;
1892 		}
1893 
1894 		if (__predict_false(cpl->ip_frag))
1895 			m0->m_pkthdr.csum_data = be16toh(cpl->csum);
1896 		else
1897 			m0->m_pkthdr.csum_data = 0xffff;
1898 	}
1899 
1900 	if (cpl->vlan_ex) {
1901 		m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
1902 		m0->m_flags |= M_VLANTAG;
1903 		rxq->vlan_extraction++;
1904 	}
1905 
1906 #if defined(INET) || defined(INET6)
1907 	if (iq->flags & IQ_LRO_ENABLED) {
1908 		if (sort_before_lro(lro)) {
1909 			tcp_lro_queue_mbuf(lro, m0);
1910 			return (0); /* queued for sort, then LRO */
1911 		}
1912 		if (tcp_lro_rx(lro, m0, 0) == 0)
1913 			return (0); /* queued for LRO */
1914 	}
1915 #endif
1916 	ifp->if_input(ifp, m0);
1917 
1918 	return (0);
1919 }
1920 
1921 /*
1922  * Must drain the wrq or make sure that someone else will.
1923  */
1924 static void
1925 wrq_tx_drain(void *arg, int n)
1926 {
1927 	struct sge_wrq *wrq = arg;
1928 	struct sge_eq *eq = &wrq->eq;
1929 
1930 	EQ_LOCK(eq);
1931 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
1932 		drain_wrq_wr_list(wrq->adapter, wrq);
1933 	EQ_UNLOCK(eq);
1934 }
1935 
1936 static void
1937 drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq)
1938 {
1939 	struct sge_eq *eq = &wrq->eq;
1940 	u_int available, dbdiff;	/* # of hardware descriptors */
1941 	u_int n;
1942 	struct wrqe *wr;
1943 	struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
1944 
1945 	EQ_LOCK_ASSERT_OWNED(eq);
1946 	MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
1947 	wr = STAILQ_FIRST(&wrq->wr_list);
1948 	MPASS(wr != NULL);	/* Must be called with something useful to do */
1949 	MPASS(eq->pidx == eq->dbidx);
1950 	dbdiff = 0;
1951 
1952 	do {
1953 		eq->cidx = read_hw_cidx(eq);
1954 		if (eq->pidx == eq->cidx)
1955 			available = eq->sidx - 1;
1956 		else
1957 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
1958 
1959 		MPASS(wr->wrq == wrq);
1960 		n = howmany(wr->wr_len, EQ_ESIZE);
1961 		if (available < n)
1962 			break;
1963 
1964 		dst = (void *)&eq->desc[eq->pidx];
1965 		if (__predict_true(eq->sidx - eq->pidx > n)) {
1966 			/* Won't wrap, won't end exactly at the status page. */
1967 			bcopy(&wr->wr[0], dst, wr->wr_len);
1968 			eq->pidx += n;
1969 		} else {
1970 			int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE;
1971 
1972 			bcopy(&wr->wr[0], dst, first_portion);
1973 			if (wr->wr_len > first_portion) {
1974 				bcopy(&wr->wr[first_portion], &eq->desc[0],
1975 				    wr->wr_len - first_portion);
1976 			}
1977 			eq->pidx = n - (eq->sidx - eq->pidx);
1978 		}
1979 		wrq->tx_wrs_copied++;
1980 
1981 		if (available < eq->sidx / 4 &&
1982 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
1983 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
1984 			    F_FW_WR_EQUEQ);
1985 			eq->equeqidx = eq->pidx;
1986 		} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
1987 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
1988 			eq->equeqidx = eq->pidx;
1989 		}
1990 
1991 		dbdiff += n;
1992 		if (dbdiff >= 16) {
1993 			ring_eq_db(sc, eq, dbdiff);
1994 			dbdiff = 0;
1995 		}
1996 
1997 		STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
1998 		free_wrqe(wr);
1999 		MPASS(wrq->nwr_pending > 0);
2000 		wrq->nwr_pending--;
2001 		MPASS(wrq->ndesc_needed >= n);
2002 		wrq->ndesc_needed -= n;
2003 	} while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL);
2004 
2005 	if (dbdiff)
2006 		ring_eq_db(sc, eq, dbdiff);
2007 }
2008 
2009 /*
2010  * Doesn't fail.  Holds on to work requests it can't send right away.
2011  */
2012 void
2013 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
2014 {
2015 #ifdef INVARIANTS
2016 	struct sge_eq *eq = &wrq->eq;
2017 #endif
2018 
2019 	EQ_LOCK_ASSERT_OWNED(eq);
2020 	MPASS(wr != NULL);
2021 	MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN);
2022 	MPASS((wr->wr_len & 0x7) == 0);
2023 
2024 	STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
2025 	wrq->nwr_pending++;
2026 	wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE);
2027 
2028 	if (!TAILQ_EMPTY(&wrq->incomplete_wrs))
2029 		return;	/* commit_wrq_wr will drain wr_list as well. */
2030 
2031 	drain_wrq_wr_list(sc, wrq);
2032 
2033 	/* Doorbell must have caught up to the pidx. */
2034 	MPASS(eq->pidx == eq->dbidx);
2035 }
2036 
2037 void
2038 t4_update_fl_bufsize(struct ifnet *ifp)
2039 {
2040 	struct vi_info *vi = ifp->if_softc;
2041 	struct adapter *sc = vi->pi->adapter;
2042 	struct sge_rxq *rxq;
2043 #ifdef TCP_OFFLOAD
2044 	struct sge_ofld_rxq *ofld_rxq;
2045 #endif
2046 	struct sge_fl *fl;
2047 	int i, maxp, mtu = ifp->if_mtu;
2048 
2049 	maxp = mtu_to_max_payload(sc, mtu, 0);
2050 	for_each_rxq(vi, i, rxq) {
2051 		fl = &rxq->fl;
2052 
2053 		FL_LOCK(fl);
2054 		find_best_refill_source(sc, fl, maxp);
2055 		FL_UNLOCK(fl);
2056 	}
2057 #ifdef TCP_OFFLOAD
2058 	maxp = mtu_to_max_payload(sc, mtu, 1);
2059 	for_each_ofld_rxq(vi, i, ofld_rxq) {
2060 		fl = &ofld_rxq->fl;
2061 
2062 		FL_LOCK(fl);
2063 		find_best_refill_source(sc, fl, maxp);
2064 		FL_UNLOCK(fl);
2065 	}
2066 #endif
2067 }
2068 
2069 static inline int
2070 mbuf_nsegs(struct mbuf *m)
2071 {
2072 
2073 	M_ASSERTPKTHDR(m);
2074 	KASSERT(m->m_pkthdr.l5hlen > 0,
2075 	    ("%s: mbuf %p missing information on # of segments.", __func__, m));
2076 
2077 	return (m->m_pkthdr.l5hlen);
2078 }
2079 
2080 static inline void
2081 set_mbuf_nsegs(struct mbuf *m, uint8_t nsegs)
2082 {
2083 
2084 	M_ASSERTPKTHDR(m);
2085 	m->m_pkthdr.l5hlen = nsegs;
2086 }
2087 
2088 static inline int
2089 mbuf_len16(struct mbuf *m)
2090 {
2091 	int n;
2092 
2093 	M_ASSERTPKTHDR(m);
2094 	n = m->m_pkthdr.PH_loc.eight[0];
2095 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2096 
2097 	return (n);
2098 }
2099 
2100 static inline void
2101 set_mbuf_len16(struct mbuf *m, uint8_t len16)
2102 {
2103 
2104 	M_ASSERTPKTHDR(m);
2105 	m->m_pkthdr.PH_loc.eight[0] = len16;
2106 }
2107 
2108 static inline int
2109 needs_tso(struct mbuf *m)
2110 {
2111 
2112 	M_ASSERTPKTHDR(m);
2113 
2114 	if (m->m_pkthdr.csum_flags & CSUM_TSO) {
2115 		KASSERT(m->m_pkthdr.tso_segsz > 0,
2116 		    ("%s: TSO requested in mbuf %p but MSS not provided",
2117 		    __func__, m));
2118 		return (1);
2119 	}
2120 
2121 	return (0);
2122 }
2123 
2124 static inline int
2125 needs_l3_csum(struct mbuf *m)
2126 {
2127 
2128 	M_ASSERTPKTHDR(m);
2129 
2130 	if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO))
2131 		return (1);
2132 	return (0);
2133 }
2134 
2135 static inline int
2136 needs_l4_csum(struct mbuf *m)
2137 {
2138 
2139 	M_ASSERTPKTHDR(m);
2140 
2141 	if (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
2142 	    CSUM_TCP_IPV6 | CSUM_TSO))
2143 		return (1);
2144 	return (0);
2145 }
2146 
2147 static inline int
2148 needs_vlan_insertion(struct mbuf *m)
2149 {
2150 
2151 	M_ASSERTPKTHDR(m);
2152 
2153 	if (m->m_flags & M_VLANTAG) {
2154 		KASSERT(m->m_pkthdr.ether_vtag != 0,
2155 		    ("%s: HWVLAN requested in mbuf %p but tag not provided",
2156 		    __func__, m));
2157 		return (1);
2158 	}
2159 	return (0);
2160 }
2161 
2162 static void *
2163 m_advance(struct mbuf **pm, int *poffset, int len)
2164 {
2165 	struct mbuf *m = *pm;
2166 	int offset = *poffset;
2167 	uintptr_t p = 0;
2168 
2169 	MPASS(len > 0);
2170 
2171 	for (;;) {
2172 		if (offset + len < m->m_len) {
2173 			offset += len;
2174 			p = mtod(m, uintptr_t) + offset;
2175 			break;
2176 		}
2177 		len -= m->m_len - offset;
2178 		m = m->m_next;
2179 		offset = 0;
2180 		MPASS(m != NULL);
2181 	}
2182 	*poffset = offset;
2183 	*pm = m;
2184 	return ((void *)p);
2185 }
2186 
2187 /*
2188  * Can deal with empty mbufs in the chain that have m_len = 0, but the chain
2189  * must have at least one mbuf that's not empty.
2190  */
2191 static inline int
2192 count_mbuf_nsegs(struct mbuf *m)
2193 {
2194 	vm_paddr_t lastb, next;
2195 	vm_offset_t va;
2196 	int len, nsegs;
2197 
2198 	MPASS(m != NULL);
2199 
2200 	nsegs = 0;
2201 	lastb = 0;
2202 	for (; m; m = m->m_next) {
2203 
2204 		len = m->m_len;
2205 		if (__predict_false(len == 0))
2206 			continue;
2207 		va = mtod(m, vm_offset_t);
2208 		next = pmap_kextract(va);
2209 		nsegs += sglist_count(m->m_data, len);
2210 		if (lastb + 1 == next)
2211 			nsegs--;
2212 		lastb = pmap_kextract(va + len - 1);
2213 	}
2214 
2215 	MPASS(nsegs > 0);
2216 	return (nsegs);
2217 }
2218 
2219 /*
2220  * Analyze the mbuf to determine its tx needs.  The mbuf passed in may change:
2221  * a) caller can assume it's been freed if this function returns with an error.
2222  * b) it may get defragged up if the gather list is too long for the hardware.
2223  */
2224 int
2225 parse_pkt(struct adapter *sc, struct mbuf **mp)
2226 {
2227 	struct mbuf *m0 = *mp, *m;
2228 	int rc, nsegs, defragged = 0, offset;
2229 	struct ether_header *eh;
2230 	void *l3hdr;
2231 #if defined(INET) || defined(INET6)
2232 	struct tcphdr *tcp;
2233 #endif
2234 	uint16_t eh_type;
2235 
2236 	M_ASSERTPKTHDR(m0);
2237 	if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) {
2238 		rc = EINVAL;
2239 fail:
2240 		m_freem(m0);
2241 		*mp = NULL;
2242 		return (rc);
2243 	}
2244 restart:
2245 	/*
2246 	 * First count the number of gather list segments in the payload.
2247 	 * Defrag the mbuf if nsegs exceeds the hardware limit.
2248 	 */
2249 	M_ASSERTPKTHDR(m0);
2250 	MPASS(m0->m_pkthdr.len > 0);
2251 	nsegs = count_mbuf_nsegs(m0);
2252 	if (nsegs > (needs_tso(m0) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS)) {
2253 		if (defragged++ > 0 || (m = m_defrag(m0, M_NOWAIT)) == NULL) {
2254 			rc = EFBIG;
2255 			goto fail;
2256 		}
2257 		*mp = m0 = m;	/* update caller's copy after defrag */
2258 		goto restart;
2259 	}
2260 
2261 	if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN)) {
2262 		m0 = m_pullup(m0, m0->m_pkthdr.len);
2263 		if (m0 == NULL) {
2264 			/* Should have left well enough alone. */
2265 			rc = EFBIG;
2266 			goto fail;
2267 		}
2268 		*mp = m0;	/* update caller's copy after pullup */
2269 		goto restart;
2270 	}
2271 	set_mbuf_nsegs(m0, nsegs);
2272 	if (sc->flags & IS_VF)
2273 		set_mbuf_len16(m0, txpkt_vm_len16(nsegs, needs_tso(m0)));
2274 	else
2275 		set_mbuf_len16(m0, txpkt_len16(nsegs, needs_tso(m0)));
2276 
2277 	if (!needs_tso(m0) &&
2278 	    !(sc->flags & IS_VF && (needs_l3_csum(m0) || needs_l4_csum(m0))))
2279 		return (0);
2280 
2281 	m = m0;
2282 	eh = mtod(m, struct ether_header *);
2283 	eh_type = ntohs(eh->ether_type);
2284 	if (eh_type == ETHERTYPE_VLAN) {
2285 		struct ether_vlan_header *evh = (void *)eh;
2286 
2287 		eh_type = ntohs(evh->evl_proto);
2288 		m0->m_pkthdr.l2hlen = sizeof(*evh);
2289 	} else
2290 		m0->m_pkthdr.l2hlen = sizeof(*eh);
2291 
2292 	offset = 0;
2293 	l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2294 
2295 	switch (eh_type) {
2296 #ifdef INET6
2297 	case ETHERTYPE_IPV6:
2298 	{
2299 		struct ip6_hdr *ip6 = l3hdr;
2300 
2301 		MPASS(!needs_tso(m0) || ip6->ip6_nxt == IPPROTO_TCP);
2302 
2303 		m0->m_pkthdr.l3hlen = sizeof(*ip6);
2304 		break;
2305 	}
2306 #endif
2307 #ifdef INET
2308 	case ETHERTYPE_IP:
2309 	{
2310 		struct ip *ip = l3hdr;
2311 
2312 		m0->m_pkthdr.l3hlen = ip->ip_hl * 4;
2313 		break;
2314 	}
2315 #endif
2316 	default:
2317 		panic("%s: ethertype 0x%04x unknown.  if_cxgbe must be compiled"
2318 		    " with the same INET/INET6 options as the kernel.",
2319 		    __func__, eh_type);
2320 	}
2321 
2322 #if defined(INET) || defined(INET6)
2323 	if (needs_tso(m0)) {
2324 		tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen);
2325 		m0->m_pkthdr.l4hlen = tcp->th_off * 4;
2326 	}
2327 #endif
2328 	MPASS(m0 == *mp);
2329 	return (0);
2330 }
2331 
2332 void *
2333 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie)
2334 {
2335 	struct sge_eq *eq = &wrq->eq;
2336 	struct adapter *sc = wrq->adapter;
2337 	int ndesc, available;
2338 	struct wrqe *wr;
2339 	void *w;
2340 
2341 	MPASS(len16 > 0);
2342 	ndesc = howmany(len16, EQ_ESIZE / 16);
2343 	MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC);
2344 
2345 	EQ_LOCK(eq);
2346 
2347 	if (!STAILQ_EMPTY(&wrq->wr_list))
2348 		drain_wrq_wr_list(sc, wrq);
2349 
2350 	if (!STAILQ_EMPTY(&wrq->wr_list)) {
2351 slowpath:
2352 		EQ_UNLOCK(eq);
2353 		wr = alloc_wrqe(len16 * 16, wrq);
2354 		if (__predict_false(wr == NULL))
2355 			return (NULL);
2356 		cookie->pidx = -1;
2357 		cookie->ndesc = ndesc;
2358 		return (&wr->wr);
2359 	}
2360 
2361 	eq->cidx = read_hw_cidx(eq);
2362 	if (eq->pidx == eq->cidx)
2363 		available = eq->sidx - 1;
2364 	else
2365 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2366 	if (available < ndesc)
2367 		goto slowpath;
2368 
2369 	cookie->pidx = eq->pidx;
2370 	cookie->ndesc = ndesc;
2371 	TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link);
2372 
2373 	w = &eq->desc[eq->pidx];
2374 	IDXINCR(eq->pidx, ndesc, eq->sidx);
2375 	if (__predict_false(cookie->pidx + ndesc > eq->sidx)) {
2376 		w = &wrq->ss[0];
2377 		wrq->ss_pidx = cookie->pidx;
2378 		wrq->ss_len = len16 * 16;
2379 	}
2380 
2381 	EQ_UNLOCK(eq);
2382 
2383 	return (w);
2384 }
2385 
2386 void
2387 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie)
2388 {
2389 	struct sge_eq *eq = &wrq->eq;
2390 	struct adapter *sc = wrq->adapter;
2391 	int ndesc, pidx;
2392 	struct wrq_cookie *prev, *next;
2393 
2394 	if (cookie->pidx == -1) {
2395 		struct wrqe *wr = __containerof(w, struct wrqe, wr);
2396 
2397 		t4_wrq_tx(sc, wr);
2398 		return;
2399 	}
2400 
2401 	ndesc = cookie->ndesc;	/* Can be more than SGE_MAX_WR_NDESC here. */
2402 	pidx = cookie->pidx;
2403 	MPASS(pidx >= 0 && pidx < eq->sidx);
2404 	if (__predict_false(w == &wrq->ss[0])) {
2405 		int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE;
2406 
2407 		MPASS(wrq->ss_len > n);	/* WR had better wrap around. */
2408 		bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n);
2409 		bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n);
2410 		wrq->tx_wrs_ss++;
2411 	} else
2412 		wrq->tx_wrs_direct++;
2413 
2414 	EQ_LOCK(eq);
2415 	prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link);
2416 	next = TAILQ_NEXT(cookie, link);
2417 	if (prev == NULL) {
2418 		MPASS(pidx == eq->dbidx);
2419 		if (next == NULL || ndesc >= 16)
2420 			ring_eq_db(wrq->adapter, eq, ndesc);
2421 		else {
2422 			MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc);
2423 			next->pidx = pidx;
2424 			next->ndesc += ndesc;
2425 		}
2426 	} else {
2427 		MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc);
2428 		prev->ndesc += ndesc;
2429 	}
2430 	TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link);
2431 
2432 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2433 		drain_wrq_wr_list(sc, wrq);
2434 
2435 #ifdef INVARIANTS
2436 	if (TAILQ_EMPTY(&wrq->incomplete_wrs)) {
2437 		/* Doorbell must have caught up to the pidx. */
2438 		MPASS(wrq->eq.pidx == wrq->eq.dbidx);
2439 	}
2440 #endif
2441 	EQ_UNLOCK(eq);
2442 }
2443 
2444 static u_int
2445 can_resume_eth_tx(struct mp_ring *r)
2446 {
2447 	struct sge_eq *eq = r->cookie;
2448 
2449 	return (total_available_tx_desc(eq) > eq->sidx / 8);
2450 }
2451 
2452 static inline int
2453 cannot_use_txpkts(struct mbuf *m)
2454 {
2455 	/* maybe put a GL limit too, to avoid silliness? */
2456 
2457 	return (needs_tso(m));
2458 }
2459 
2460 static inline int
2461 discard_tx(struct sge_eq *eq)
2462 {
2463 
2464 	return ((eq->flags & (EQ_ENABLED | EQ_QFLUSH)) != EQ_ENABLED);
2465 }
2466 
2467 /*
2468  * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to
2469  * be consumed.  Return the actual number consumed.  0 indicates a stall.
2470  */
2471 static u_int
2472 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx)
2473 {
2474 	struct sge_txq *txq = r->cookie;
2475 	struct sge_eq *eq = &txq->eq;
2476 	struct ifnet *ifp = txq->ifp;
2477 	struct vi_info *vi = ifp->if_softc;
2478 	struct port_info *pi = vi->pi;
2479 	struct adapter *sc = pi->adapter;
2480 	u_int total, remaining;		/* # of packets */
2481 	u_int available, dbdiff;	/* # of hardware descriptors */
2482 	u_int n, next_cidx;
2483 	struct mbuf *m0, *tail;
2484 	struct txpkts txp;
2485 	struct fw_eth_tx_pkts_wr *wr;	/* any fw WR struct will do */
2486 
2487 	remaining = IDXDIFF(pidx, cidx, r->size);
2488 	MPASS(remaining > 0);	/* Must not be called without work to do. */
2489 	total = 0;
2490 
2491 	TXQ_LOCK(txq);
2492 	if (__predict_false(discard_tx(eq))) {
2493 		while (cidx != pidx) {
2494 			m0 = r->items[cidx];
2495 			m_freem(m0);
2496 			if (++cidx == r->size)
2497 				cidx = 0;
2498 		}
2499 		reclaim_tx_descs(txq, 2048);
2500 		total = remaining;
2501 		goto done;
2502 	}
2503 
2504 	/* How many hardware descriptors do we have readily available. */
2505 	if (eq->pidx == eq->cidx)
2506 		available = eq->sidx - 1;
2507 	else
2508 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2509 	dbdiff = IDXDIFF(eq->pidx, eq->dbidx, eq->sidx);
2510 
2511 	while (remaining > 0) {
2512 
2513 		m0 = r->items[cidx];
2514 		M_ASSERTPKTHDR(m0);
2515 		MPASS(m0->m_nextpkt == NULL);
2516 
2517 		if (available < SGE_MAX_WR_NDESC) {
2518 			available += reclaim_tx_descs(txq, 64);
2519 			if (available < howmany(mbuf_len16(m0), EQ_ESIZE / 16))
2520 				break;	/* out of descriptors */
2521 		}
2522 
2523 		next_cidx = cidx + 1;
2524 		if (__predict_false(next_cidx == r->size))
2525 			next_cidx = 0;
2526 
2527 		wr = (void *)&eq->desc[eq->pidx];
2528 		if (sc->flags & IS_VF) {
2529 			total++;
2530 			remaining--;
2531 			ETHER_BPF_MTAP(ifp, m0);
2532 			n = write_txpkt_vm_wr(sc, txq, (void *)wr, m0,
2533 			    available);
2534 		} else if (remaining > 1 &&
2535 		    try_txpkts(m0, r->items[next_cidx], &txp, available) == 0) {
2536 
2537 			/* pkts at cidx, next_cidx should both be in txp. */
2538 			MPASS(txp.npkt == 2);
2539 			tail = r->items[next_cidx];
2540 			MPASS(tail->m_nextpkt == NULL);
2541 			ETHER_BPF_MTAP(ifp, m0);
2542 			ETHER_BPF_MTAP(ifp, tail);
2543 			m0->m_nextpkt = tail;
2544 
2545 			if (__predict_false(++next_cidx == r->size))
2546 				next_cidx = 0;
2547 
2548 			while (next_cidx != pidx) {
2549 				if (add_to_txpkts(r->items[next_cidx], &txp,
2550 				    available) != 0)
2551 					break;
2552 				tail->m_nextpkt = r->items[next_cidx];
2553 				tail = tail->m_nextpkt;
2554 				ETHER_BPF_MTAP(ifp, tail);
2555 				if (__predict_false(++next_cidx == r->size))
2556 					next_cidx = 0;
2557 			}
2558 
2559 			n = write_txpkts_wr(txq, wr, m0, &txp, available);
2560 			total += txp.npkt;
2561 			remaining -= txp.npkt;
2562 		} else {
2563 			total++;
2564 			remaining--;
2565 			ETHER_BPF_MTAP(ifp, m0);
2566 			n = write_txpkt_wr(txq, (void *)wr, m0, available);
2567 		}
2568 		MPASS(n >= 1 && n <= available && n <= SGE_MAX_WR_NDESC);
2569 
2570 		available -= n;
2571 		dbdiff += n;
2572 		IDXINCR(eq->pidx, n, eq->sidx);
2573 
2574 		if (total_available_tx_desc(eq) < eq->sidx / 4 &&
2575 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2576 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2577 			    F_FW_WR_EQUEQ);
2578 			eq->equeqidx = eq->pidx;
2579 		} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
2580 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
2581 			eq->equeqidx = eq->pidx;
2582 		}
2583 
2584 		if (dbdiff >= 16 && remaining >= 4) {
2585 			ring_eq_db(sc, eq, dbdiff);
2586 			available += reclaim_tx_descs(txq, 4 * dbdiff);
2587 			dbdiff = 0;
2588 		}
2589 
2590 		cidx = next_cidx;
2591 	}
2592 	if (dbdiff != 0) {
2593 		ring_eq_db(sc, eq, dbdiff);
2594 		reclaim_tx_descs(txq, 32);
2595 	}
2596 done:
2597 	TXQ_UNLOCK(txq);
2598 
2599 	return (total);
2600 }
2601 
2602 static inline void
2603 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
2604     int qsize)
2605 {
2606 
2607 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
2608 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
2609 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
2610 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
2611 
2612 	iq->flags = 0;
2613 	iq->adapter = sc;
2614 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
2615 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
2616 	if (pktc_idx >= 0) {
2617 		iq->intr_params |= F_QINTR_CNT_EN;
2618 		iq->intr_pktc_idx = pktc_idx;
2619 	}
2620 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
2621 	iq->sidx = iq->qsize - sc->params.sge.spg_len / IQ_ESIZE;
2622 }
2623 
2624 static inline void
2625 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name)
2626 {
2627 
2628 	fl->qsize = qsize;
2629 	fl->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2630 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
2631 	if (sc->flags & BUF_PACKING_OK &&
2632 	    ((!is_t4(sc) && buffer_packing) ||	/* T5+: enabled unless 0 */
2633 	    (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */
2634 		fl->flags |= FL_BUF_PACKING;
2635 	find_best_refill_source(sc, fl, maxp);
2636 	find_safe_refill_source(sc, fl);
2637 }
2638 
2639 static inline void
2640 init_eq(struct adapter *sc, struct sge_eq *eq, int eqtype, int qsize,
2641     uint8_t tx_chan, uint16_t iqid, char *name)
2642 {
2643 	KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
2644 
2645 	eq->flags = eqtype & EQ_TYPEMASK;
2646 	eq->tx_chan = tx_chan;
2647 	eq->iqid = iqid;
2648 	eq->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
2649 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
2650 }
2651 
2652 static int
2653 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
2654     bus_dmamap_t *map, bus_addr_t *pa, void **va)
2655 {
2656 	int rc;
2657 
2658 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
2659 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
2660 	if (rc != 0) {
2661 		device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
2662 		goto done;
2663 	}
2664 
2665 	rc = bus_dmamem_alloc(*tag, va,
2666 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
2667 	if (rc != 0) {
2668 		device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
2669 		goto done;
2670 	}
2671 
2672 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
2673 	if (rc != 0) {
2674 		device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
2675 		goto done;
2676 	}
2677 done:
2678 	if (rc)
2679 		free_ring(sc, *tag, *map, *pa, *va);
2680 
2681 	return (rc);
2682 }
2683 
2684 static int
2685 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
2686     bus_addr_t pa, void *va)
2687 {
2688 	if (pa)
2689 		bus_dmamap_unload(tag, map);
2690 	if (va)
2691 		bus_dmamem_free(tag, va, map);
2692 	if (tag)
2693 		bus_dma_tag_destroy(tag);
2694 
2695 	return (0);
2696 }
2697 
2698 /*
2699  * Allocates the ring for an ingress queue and an optional freelist.  If the
2700  * freelist is specified it will be allocated and then associated with the
2701  * ingress queue.
2702  *
2703  * Returns errno on failure.  Resources allocated up to that point may still be
2704  * allocated.  Caller is responsible for cleanup in case this function fails.
2705  *
2706  * If the ingress queue will take interrupts directly (iq->flags & IQ_INTR) then
2707  * the intr_idx specifies the vector, starting from 0.  Otherwise it specifies
2708  * the abs_id of the ingress queue to which its interrupts should be forwarded.
2709  */
2710 static int
2711 alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl,
2712     int intr_idx, int cong)
2713 {
2714 	int rc, i, cntxt_id;
2715 	size_t len;
2716 	struct fw_iq_cmd c;
2717 	struct port_info *pi = vi->pi;
2718 	struct adapter *sc = iq->adapter;
2719 	struct sge_params *sp = &sc->params.sge;
2720 	__be32 v = 0;
2721 
2722 	len = iq->qsize * IQ_ESIZE;
2723 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
2724 	    (void **)&iq->desc);
2725 	if (rc != 0)
2726 		return (rc);
2727 
2728 	bzero(&c, sizeof(c));
2729 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
2730 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
2731 	    V_FW_IQ_CMD_VFN(0));
2732 
2733 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
2734 	    FW_LEN16(c));
2735 
2736 	/* Special handling for firmware event queue */
2737 	if (iq == &sc->sge.fwq)
2738 		v |= F_FW_IQ_CMD_IQASYNCH;
2739 
2740 	if (iq->flags & IQ_INTR) {
2741 		KASSERT(intr_idx < sc->intr_count,
2742 		    ("%s: invalid direct intr_idx %d", __func__, intr_idx));
2743 	} else
2744 		v |= F_FW_IQ_CMD_IQANDST;
2745 	v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
2746 
2747 	c.type_to_iqandstindex = htobe32(v |
2748 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
2749 	    V_FW_IQ_CMD_VIID(vi->viid) |
2750 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
2751 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
2752 	    F_FW_IQ_CMD_IQGTSMODE |
2753 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
2754 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
2755 	c.iqsize = htobe16(iq->qsize);
2756 	c.iqaddr = htobe64(iq->ba);
2757 	if (cong >= 0)
2758 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
2759 
2760 	if (fl) {
2761 		mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
2762 
2763 		len = fl->qsize * EQ_ESIZE;
2764 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
2765 		    &fl->ba, (void **)&fl->desc);
2766 		if (rc)
2767 			return (rc);
2768 
2769 		/* Allocate space for one software descriptor per buffer. */
2770 		rc = alloc_fl_sdesc(fl);
2771 		if (rc != 0) {
2772 			device_printf(sc->dev,
2773 			    "failed to setup fl software descriptors: %d\n",
2774 			    rc);
2775 			return (rc);
2776 		}
2777 
2778 		if (fl->flags & FL_BUF_PACKING) {
2779 			fl->lowat = roundup2(sp->fl_starve_threshold2, 8);
2780 			fl->buf_boundary = sp->pack_boundary;
2781 		} else {
2782 			fl->lowat = roundup2(sp->fl_starve_threshold, 8);
2783 			fl->buf_boundary = 16;
2784 		}
2785 		if (fl_pad && fl->buf_boundary < sp->pad_boundary)
2786 			fl->buf_boundary = sp->pad_boundary;
2787 
2788 		c.iqns_to_fl0congen |=
2789 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
2790 			F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
2791 			(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
2792 			(fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
2793 			    0));
2794 		if (cong >= 0) {
2795 			c.iqns_to_fl0congen |=
2796 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
2797 				    F_FW_IQ_CMD_FL0CONGCIF |
2798 				    F_FW_IQ_CMD_FL0CONGEN);
2799 		}
2800 		c.fl0dcaen_to_fl0cidxfthresh =
2801 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
2802 			X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B) |
2803 			V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
2804 			X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
2805 		c.fl0size = htobe16(fl->qsize);
2806 		c.fl0addr = htobe64(fl->ba);
2807 	}
2808 
2809 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
2810 	if (rc != 0) {
2811 		device_printf(sc->dev,
2812 		    "failed to create ingress queue: %d\n", rc);
2813 		return (rc);
2814 	}
2815 
2816 	iq->cidx = 0;
2817 	iq->gen = F_RSPD_GEN;
2818 	iq->intr_next = iq->intr_params;
2819 	iq->cntxt_id = be16toh(c.iqid);
2820 	iq->abs_id = be16toh(c.physiqid);
2821 	iq->flags |= IQ_ALLOCATED;
2822 
2823 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
2824 	if (cntxt_id >= sc->sge.niq) {
2825 		panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
2826 		    cntxt_id, sc->sge.niq - 1);
2827 	}
2828 	sc->sge.iqmap[cntxt_id] = iq;
2829 
2830 	if (fl) {
2831 		u_int qid;
2832 
2833 		iq->flags |= IQ_HAS_FL;
2834 		fl->cntxt_id = be16toh(c.fl0id);
2835 		fl->pidx = fl->cidx = 0;
2836 
2837 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
2838 		if (cntxt_id >= sc->sge.neq) {
2839 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
2840 			    __func__, cntxt_id, sc->sge.neq - 1);
2841 		}
2842 		sc->sge.eqmap[cntxt_id] = (void *)fl;
2843 
2844 		qid = fl->cntxt_id;
2845 		if (isset(&sc->doorbells, DOORBELL_UDB)) {
2846 			uint32_t s_qpp = sc->params.sge.eq_s_qpp;
2847 			uint32_t mask = (1 << s_qpp) - 1;
2848 			volatile uint8_t *udb;
2849 
2850 			udb = sc->udbs_base + UDBS_DB_OFFSET;
2851 			udb += (qid >> s_qpp) << PAGE_SHIFT;
2852 			qid &= mask;
2853 			if (qid < PAGE_SIZE / UDBS_SEG_SIZE) {
2854 				udb += qid << UDBS_SEG_SHIFT;
2855 				qid = 0;
2856 			}
2857 			fl->udb = (volatile void *)udb;
2858 		}
2859 		fl->dbval = V_QID(qid) | sc->chip_params->sge_fl_db;
2860 
2861 		FL_LOCK(fl);
2862 		/* Enough to make sure the SGE doesn't think it's starved */
2863 		refill_fl(sc, fl, fl->lowat);
2864 		FL_UNLOCK(fl);
2865 	}
2866 
2867 	if (chip_id(sc) >= CHELSIO_T5 && !(sc->flags & IS_VF) && cong >= 0) {
2868 		uint32_t param, val;
2869 
2870 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
2871 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
2872 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
2873 		if (cong == 0)
2874 			val = 1 << 19;
2875 		else {
2876 			val = 2 << 19;
2877 			for (i = 0; i < 4; i++) {
2878 				if (cong & (1 << i))
2879 					val |= 1 << (i << 2);
2880 			}
2881 		}
2882 
2883 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
2884 		if (rc != 0) {
2885 			/* report error but carry on */
2886 			device_printf(sc->dev,
2887 			    "failed to set congestion manager context for "
2888 			    "ingress queue %d: %d\n", iq->cntxt_id, rc);
2889 		}
2890 	}
2891 
2892 	/* Enable IQ interrupts */
2893 	atomic_store_rel_int(&iq->state, IQS_IDLE);
2894 	t4_write_reg(sc, sc->sge_gts_reg, V_SEINTARM(iq->intr_params) |
2895 	    V_INGRESSQID(iq->cntxt_id));
2896 
2897 	return (0);
2898 }
2899 
2900 static int
2901 free_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl)
2902 {
2903 	int rc;
2904 	struct adapter *sc = iq->adapter;
2905 	device_t dev;
2906 
2907 	if (sc == NULL)
2908 		return (0);	/* nothing to do */
2909 
2910 	dev = vi ? vi->dev : sc->dev;
2911 
2912 	if (iq->flags & IQ_ALLOCATED) {
2913 		rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
2914 		    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
2915 		    fl ? fl->cntxt_id : 0xffff, 0xffff);
2916 		if (rc != 0) {
2917 			device_printf(dev,
2918 			    "failed to free queue %p: %d\n", iq, rc);
2919 			return (rc);
2920 		}
2921 		iq->flags &= ~IQ_ALLOCATED;
2922 	}
2923 
2924 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
2925 
2926 	bzero(iq, sizeof(*iq));
2927 
2928 	if (fl) {
2929 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
2930 		    fl->desc);
2931 
2932 		if (fl->sdesc)
2933 			free_fl_sdesc(sc, fl);
2934 
2935 		if (mtx_initialized(&fl->fl_lock))
2936 			mtx_destroy(&fl->fl_lock);
2937 
2938 		bzero(fl, sizeof(*fl));
2939 	}
2940 
2941 	return (0);
2942 }
2943 
2944 static void
2945 add_fl_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
2946     struct sysctl_oid *oid, struct sge_fl *fl)
2947 {
2948 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2949 
2950 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
2951 	    "freelist");
2952 	children = SYSCTL_CHILDREN(oid);
2953 
2954 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
2955 	    &fl->ba, "bus address of descriptor ring");
2956 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
2957 	    fl->sidx * EQ_ESIZE + sc->params.sge.spg_len,
2958 	    "desc ring size in bytes");
2959 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
2960 	    CTLTYPE_INT | CTLFLAG_RD, &fl->cntxt_id, 0, sysctl_uint16, "I",
2961 	    "SGE context id of the freelist");
2962 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL,
2963 	    fl_pad ? 1 : 0, "padding enabled");
2964 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL,
2965 	    fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled");
2966 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
2967 	    0, "consumer index");
2968 	if (fl->flags & FL_BUF_PACKING) {
2969 		SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
2970 		    CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
2971 	}
2972 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
2973 	    0, "producer index");
2974 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_allocated",
2975 	    CTLFLAG_RD, &fl->mbuf_allocated, "# of mbuf allocated");
2976 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_inlined",
2977 	    CTLFLAG_RD, &fl->mbuf_inlined, "# of mbuf inlined in clusters");
2978 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
2979 	    CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
2980 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
2981 	    CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
2982 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
2983 	    CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
2984 }
2985 
2986 static int
2987 alloc_fwq(struct adapter *sc)
2988 {
2989 	int rc, intr_idx;
2990 	struct sge_iq *fwq = &sc->sge.fwq;
2991 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
2992 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2993 
2994 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE);
2995 	fwq->flags |= IQ_INTR;	/* always */
2996 	if (sc->flags & IS_VF)
2997 		intr_idx = 0;
2998 	else {
2999 		intr_idx = sc->intr_count > 1 ? 1 : 0;
3000 		fwq->set_tcb_rpl = t4_filter_rpl;
3001 		fwq->l2t_write_rpl = do_l2t_write_rpl;
3002 	}
3003 	rc = alloc_iq_fl(&sc->port[0]->vi[0], fwq, NULL, intr_idx, -1);
3004 	if (rc != 0) {
3005 		device_printf(sc->dev,
3006 		    "failed to create firmware event queue: %d\n", rc);
3007 		return (rc);
3008 	}
3009 
3010 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD,
3011 	    NULL, "firmware event queue");
3012 	children = SYSCTL_CHILDREN(oid);
3013 
3014 	SYSCTL_ADD_UAUTO(&sc->ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3015 	    &fwq->ba, "bus address of descriptor ring");
3016 	SYSCTL_ADD_INT(&sc->ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3017 	    fwq->qsize * IQ_ESIZE, "descriptor ring size in bytes");
3018 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "abs_id",
3019 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->abs_id, 0, sysctl_uint16, "I",
3020 	    "absolute id of the queue");
3021 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cntxt_id",
3022 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->cntxt_id, 0, sysctl_uint16, "I",
3023 	    "SGE context id of the queue");
3024 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cidx",
3025 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->cidx, 0, sysctl_uint16, "I",
3026 	    "consumer index");
3027 
3028 	return (0);
3029 }
3030 
3031 static int
3032 free_fwq(struct adapter *sc)
3033 {
3034 	return free_iq_fl(NULL, &sc->sge.fwq, NULL);
3035 }
3036 
3037 static int
3038 alloc_mgmtq(struct adapter *sc)
3039 {
3040 	int rc;
3041 	struct sge_wrq *mgmtq = &sc->sge.mgmtq;
3042 	char name[16];
3043 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
3044 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3045 
3046 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "mgmtq", CTLFLAG_RD,
3047 	    NULL, "management queue");
3048 
3049 	snprintf(name, sizeof(name), "%s mgmtq", device_get_nameunit(sc->dev));
3050 	init_eq(sc, &mgmtq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[0]->tx_chan,
3051 	    sc->sge.fwq.cntxt_id, name);
3052 	rc = alloc_wrq(sc, NULL, mgmtq, oid);
3053 	if (rc != 0) {
3054 		device_printf(sc->dev,
3055 		    "failed to create management queue: %d\n", rc);
3056 		return (rc);
3057 	}
3058 
3059 	return (0);
3060 }
3061 
3062 static int
3063 free_mgmtq(struct adapter *sc)
3064 {
3065 
3066 	return free_wrq(sc, &sc->sge.mgmtq);
3067 }
3068 
3069 int
3070 tnl_cong(struct port_info *pi, int drop)
3071 {
3072 
3073 	if (drop == -1)
3074 		return (-1);
3075 	else if (drop == 1)
3076 		return (0);
3077 	else
3078 		return (pi->rx_chan_map);
3079 }
3080 
3081 static int
3082 alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int intr_idx, int idx,
3083     struct sysctl_oid *oid)
3084 {
3085 	int rc;
3086 	struct adapter *sc = vi->pi->adapter;
3087 	struct sysctl_oid_list *children;
3088 	char name[16];
3089 
3090 	rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, intr_idx,
3091 	    tnl_cong(vi->pi, cong_drop));
3092 	if (rc != 0)
3093 		return (rc);
3094 
3095 	if (idx == 0)
3096 		sc->sge.iq_base = rxq->iq.abs_id - rxq->iq.cntxt_id;
3097 	else
3098 		KASSERT(rxq->iq.cntxt_id + sc->sge.iq_base == rxq->iq.abs_id,
3099 		    ("iq_base mismatch"));
3100 	KASSERT(sc->sge.iq_base == 0 || sc->flags & IS_VF,
3101 	    ("PF with non-zero iq_base"));
3102 
3103 	/*
3104 	 * The freelist is just barely above the starvation threshold right now,
3105 	 * fill it up a bit more.
3106 	 */
3107 	FL_LOCK(&rxq->fl);
3108 	refill_fl(sc, &rxq->fl, 128);
3109 	FL_UNLOCK(&rxq->fl);
3110 
3111 #if defined(INET) || defined(INET6)
3112 	rc = tcp_lro_init_args(&rxq->lro, vi->ifp, lro_entries, lro_mbufs);
3113 	if (rc != 0)
3114 		return (rc);
3115 	MPASS(rxq->lro.ifp == vi->ifp);	/* also indicates LRO init'ed */
3116 
3117 	if (vi->ifp->if_capenable & IFCAP_LRO)
3118 		rxq->iq.flags |= IQ_LRO_ENABLED;
3119 #endif
3120 	rxq->ifp = vi->ifp;
3121 
3122 	children = SYSCTL_CHILDREN(oid);
3123 
3124 	snprintf(name, sizeof(name), "%d", idx);
3125 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3126 	    NULL, "rx queue");
3127 	children = SYSCTL_CHILDREN(oid);
3128 
3129 	SYSCTL_ADD_UAUTO(&vi->ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3130 	    &rxq->iq.ba, "bus address of descriptor ring");
3131 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3132 	    rxq->iq.qsize * IQ_ESIZE, "descriptor ring size in bytes");
3133 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "abs_id",
3134 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.abs_id, 0, sysctl_uint16, "I",
3135 	    "absolute id of the queue");
3136 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cntxt_id",
3137 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cntxt_id, 0, sysctl_uint16, "I",
3138 	    "SGE context id of the queue");
3139 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3140 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cidx, 0, sysctl_uint16, "I",
3141 	    "consumer index");
3142 #if defined(INET) || defined(INET6)
3143 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
3144 	    &rxq->lro.lro_queued, 0, NULL);
3145 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
3146 	    &rxq->lro.lro_flushed, 0, NULL);
3147 #endif
3148 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
3149 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
3150 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_extraction",
3151 	    CTLFLAG_RD, &rxq->vlan_extraction,
3152 	    "# of times hardware extracted 802.1Q tag");
3153 
3154 	add_fl_sysctls(sc, &vi->ctx, oid, &rxq->fl);
3155 
3156 	return (rc);
3157 }
3158 
3159 static int
3160 free_rxq(struct vi_info *vi, struct sge_rxq *rxq)
3161 {
3162 	int rc;
3163 
3164 #if defined(INET) || defined(INET6)
3165 	if (rxq->lro.ifp) {
3166 		tcp_lro_free(&rxq->lro);
3167 		rxq->lro.ifp = NULL;
3168 	}
3169 #endif
3170 
3171 	rc = free_iq_fl(vi, &rxq->iq, &rxq->fl);
3172 	if (rc == 0)
3173 		bzero(rxq, sizeof(*rxq));
3174 
3175 	return (rc);
3176 }
3177 
3178 #ifdef TCP_OFFLOAD
3179 static int
3180 alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq,
3181     int intr_idx, int idx, struct sysctl_oid *oid)
3182 {
3183 	struct port_info *pi = vi->pi;
3184 	int rc;
3185 	struct sysctl_oid_list *children;
3186 	char name[16];
3187 
3188 	rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx,
3189 	    pi->rx_chan_map);
3190 	if (rc != 0)
3191 		return (rc);
3192 
3193 	children = SYSCTL_CHILDREN(oid);
3194 
3195 	snprintf(name, sizeof(name), "%d", idx);
3196 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3197 	    NULL, "rx queue");
3198 	children = SYSCTL_CHILDREN(oid);
3199 
3200 	SYSCTL_ADD_UAUTO(&vi->ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3201 	    &ofld_rxq->iq.ba, "bus address of descriptor ring");
3202 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3203 	    ofld_rxq->iq.qsize * IQ_ESIZE, "descriptor ring size in bytes");
3204 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "abs_id",
3205 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.abs_id, 0, sysctl_uint16,
3206 	    "I", "absolute id of the queue");
3207 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cntxt_id",
3208 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cntxt_id, 0, sysctl_uint16,
3209 	    "I", "SGE context id of the queue");
3210 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3211 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cidx, 0, sysctl_uint16, "I",
3212 	    "consumer index");
3213 
3214 	add_fl_sysctls(pi->adapter, &vi->ctx, oid, &ofld_rxq->fl);
3215 
3216 	return (rc);
3217 }
3218 
3219 static int
3220 free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq)
3221 {
3222 	int rc;
3223 
3224 	rc = free_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl);
3225 	if (rc == 0)
3226 		bzero(ofld_rxq, sizeof(*ofld_rxq));
3227 
3228 	return (rc);
3229 }
3230 #endif
3231 
3232 #ifdef DEV_NETMAP
3233 static int
3234 alloc_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int intr_idx,
3235     int idx, struct sysctl_oid *oid)
3236 {
3237 	int rc;
3238 	struct sysctl_oid_list *children;
3239 	struct sysctl_ctx_list *ctx;
3240 	char name[16];
3241 	size_t len;
3242 	struct adapter *sc = vi->pi->adapter;
3243 	struct netmap_adapter *na = NA(vi->ifp);
3244 
3245 	MPASS(na != NULL);
3246 
3247 	len = vi->qsize_rxq * IQ_ESIZE;
3248 	rc = alloc_ring(sc, len, &nm_rxq->iq_desc_tag, &nm_rxq->iq_desc_map,
3249 	    &nm_rxq->iq_ba, (void **)&nm_rxq->iq_desc);
3250 	if (rc != 0)
3251 		return (rc);
3252 
3253 	len = na->num_rx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3254 	rc = alloc_ring(sc, len, &nm_rxq->fl_desc_tag, &nm_rxq->fl_desc_map,
3255 	    &nm_rxq->fl_ba, (void **)&nm_rxq->fl_desc);
3256 	if (rc != 0)
3257 		return (rc);
3258 
3259 	nm_rxq->vi = vi;
3260 	nm_rxq->nid = idx;
3261 	nm_rxq->iq_cidx = 0;
3262 	nm_rxq->iq_sidx = vi->qsize_rxq - sc->params.sge.spg_len / IQ_ESIZE;
3263 	nm_rxq->iq_gen = F_RSPD_GEN;
3264 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
3265 	nm_rxq->fl_sidx = na->num_rx_desc;
3266 	nm_rxq->intr_idx = intr_idx;
3267 	nm_rxq->iq_cntxt_id = INVALID_NM_RXQ_CNTXT_ID;
3268 
3269 	ctx = &vi->ctx;
3270 	children = SYSCTL_CHILDREN(oid);
3271 
3272 	snprintf(name, sizeof(name), "%d", idx);
3273 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL,
3274 	    "rx queue");
3275 	children = SYSCTL_CHILDREN(oid);
3276 
3277 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3278 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_abs_id, 0, sysctl_uint16,
3279 	    "I", "absolute id of the queue");
3280 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3281 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cntxt_id, 0, sysctl_uint16,
3282 	    "I", "SGE context id of the queue");
3283 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3284 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cidx, 0, sysctl_uint16, "I",
3285 	    "consumer index");
3286 
3287 	children = SYSCTL_CHILDREN(oid);
3288 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3289 	    "freelist");
3290 	children = SYSCTL_CHILDREN(oid);
3291 
3292 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3293 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->fl_cntxt_id, 0, sysctl_uint16,
3294 	    "I", "SGE context id of the freelist");
3295 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
3296 	    &nm_rxq->fl_cidx, 0, "consumer index");
3297 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
3298 	    &nm_rxq->fl_pidx, 0, "producer index");
3299 
3300 	return (rc);
3301 }
3302 
3303 
3304 static int
3305 free_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
3306 {
3307 	struct adapter *sc = vi->pi->adapter;
3308 
3309 	MPASS(nm_rxq->iq_cntxt_id == INVALID_NM_RXQ_CNTXT_ID);
3310 
3311 	free_ring(sc, nm_rxq->iq_desc_tag, nm_rxq->iq_desc_map, nm_rxq->iq_ba,
3312 	    nm_rxq->iq_desc);
3313 	free_ring(sc, nm_rxq->fl_desc_tag, nm_rxq->fl_desc_map, nm_rxq->fl_ba,
3314 	    nm_rxq->fl_desc);
3315 
3316 	return (0);
3317 }
3318 
3319 static int
3320 alloc_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq, int iqidx, int idx,
3321     struct sysctl_oid *oid)
3322 {
3323 	int rc;
3324 	size_t len;
3325 	struct port_info *pi = vi->pi;
3326 	struct adapter *sc = pi->adapter;
3327 	struct netmap_adapter *na = NA(vi->ifp);
3328 	char name[16];
3329 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3330 
3331 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3332 	rc = alloc_ring(sc, len, &nm_txq->desc_tag, &nm_txq->desc_map,
3333 	    &nm_txq->ba, (void **)&nm_txq->desc);
3334 	if (rc)
3335 		return (rc);
3336 
3337 	nm_txq->pidx = nm_txq->cidx = 0;
3338 	nm_txq->sidx = na->num_tx_desc;
3339 	nm_txq->nid = idx;
3340 	nm_txq->iqidx = iqidx;
3341 	nm_txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3342 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(G_FW_VIID_PFN(vi->viid)) |
3343 	    V_TXPKT_VF(G_FW_VIID_VIN(vi->viid)) |
3344 	    V_TXPKT_VF_VLD(G_FW_VIID_VIVLD(vi->viid)));
3345 	nm_txq->cntxt_id = INVALID_NM_TXQ_CNTXT_ID;
3346 
3347 	snprintf(name, sizeof(name), "%d", idx);
3348 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3349 	    NULL, "netmap tx queue");
3350 	children = SYSCTL_CHILDREN(oid);
3351 
3352 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3353 	    &nm_txq->cntxt_id, 0, "SGE context id of the queue");
3354 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3355 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->cidx, 0, sysctl_uint16, "I",
3356 	    "consumer index");
3357 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3358 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->pidx, 0, sysctl_uint16, "I",
3359 	    "producer index");
3360 
3361 	return (rc);
3362 }
3363 
3364 static int
3365 free_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
3366 {
3367 	struct adapter *sc = vi->pi->adapter;
3368 
3369 	MPASS(nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID);
3370 
3371 	free_ring(sc, nm_txq->desc_tag, nm_txq->desc_map, nm_txq->ba,
3372 	    nm_txq->desc);
3373 
3374 	return (0);
3375 }
3376 #endif
3377 
3378 static int
3379 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
3380 {
3381 	int rc, cntxt_id;
3382 	struct fw_eq_ctrl_cmd c;
3383 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3384 
3385 	bzero(&c, sizeof(c));
3386 
3387 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
3388 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
3389 	    V_FW_EQ_CTRL_CMD_VFN(0));
3390 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
3391 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
3392 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid));
3393 	c.physeqid_pkd = htobe32(0);
3394 	c.fetchszm_to_iqid =
3395 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
3396 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
3397 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
3398 	c.dcaen_to_eqsize =
3399 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3400 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3401 		V_FW_EQ_CTRL_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
3402 		V_FW_EQ_CTRL_CMD_EQSIZE(qsize));
3403 	c.eqaddr = htobe64(eq->ba);
3404 
3405 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3406 	if (rc != 0) {
3407 		device_printf(sc->dev,
3408 		    "failed to create control queue %d: %d\n", eq->tx_chan, rc);
3409 		return (rc);
3410 	}
3411 	eq->flags |= EQ_ALLOCATED;
3412 
3413 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
3414 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3415 	if (cntxt_id >= sc->sge.neq)
3416 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3417 		cntxt_id, sc->sge.neq - 1);
3418 	sc->sge.eqmap[cntxt_id] = eq;
3419 
3420 	return (rc);
3421 }
3422 
3423 static int
3424 eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3425 {
3426 	int rc, cntxt_id;
3427 	struct fw_eq_eth_cmd c;
3428 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3429 
3430 	bzero(&c, sizeof(c));
3431 
3432 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
3433 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
3434 	    V_FW_EQ_ETH_CMD_VFN(0));
3435 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
3436 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
3437 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
3438 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
3439 	c.fetchszm_to_iqid =
3440 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3441 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
3442 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
3443 	c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3444 	    V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3445 	    V_FW_EQ_ETH_CMD_EQSIZE(qsize));
3446 	c.eqaddr = htobe64(eq->ba);
3447 
3448 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3449 	if (rc != 0) {
3450 		device_printf(vi->dev,
3451 		    "failed to create Ethernet egress queue: %d\n", rc);
3452 		return (rc);
3453 	}
3454 	eq->flags |= EQ_ALLOCATED;
3455 
3456 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
3457 	eq->abs_id = G_FW_EQ_ETH_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
3458 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3459 	if (cntxt_id >= sc->sge.neq)
3460 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3461 		cntxt_id, sc->sge.neq - 1);
3462 	sc->sge.eqmap[cntxt_id] = eq;
3463 
3464 	return (rc);
3465 }
3466 
3467 #ifdef TCP_OFFLOAD
3468 static int
3469 ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3470 {
3471 	int rc, cntxt_id;
3472 	struct fw_eq_ofld_cmd c;
3473 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3474 
3475 	bzero(&c, sizeof(c));
3476 
3477 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
3478 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
3479 	    V_FW_EQ_OFLD_CMD_VFN(0));
3480 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
3481 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
3482 	c.fetchszm_to_iqid =
3483 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3484 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
3485 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
3486 	c.dcaen_to_eqsize =
3487 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3488 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3489 		V_FW_EQ_OFLD_CMD_EQSIZE(qsize));
3490 	c.eqaddr = htobe64(eq->ba);
3491 
3492 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3493 	if (rc != 0) {
3494 		device_printf(vi->dev,
3495 		    "failed to create egress queue for TCP offload: %d\n", rc);
3496 		return (rc);
3497 	}
3498 	eq->flags |= EQ_ALLOCATED;
3499 
3500 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
3501 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3502 	if (cntxt_id >= sc->sge.neq)
3503 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3504 		cntxt_id, sc->sge.neq - 1);
3505 	sc->sge.eqmap[cntxt_id] = eq;
3506 
3507 	return (rc);
3508 }
3509 #endif
3510 
3511 static int
3512 alloc_eq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3513 {
3514 	int rc, qsize;
3515 	size_t len;
3516 
3517 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
3518 
3519 	qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3520 	len = qsize * EQ_ESIZE;
3521 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
3522 	    &eq->ba, (void **)&eq->desc);
3523 	if (rc)
3524 		return (rc);
3525 
3526 	eq->pidx = eq->cidx = 0;
3527 	eq->equeqidx = eq->dbidx = 0;
3528 	eq->doorbells = sc->doorbells;
3529 
3530 	switch (eq->flags & EQ_TYPEMASK) {
3531 	case EQ_CTRL:
3532 		rc = ctrl_eq_alloc(sc, eq);
3533 		break;
3534 
3535 	case EQ_ETH:
3536 		rc = eth_eq_alloc(sc, vi, eq);
3537 		break;
3538 
3539 #ifdef TCP_OFFLOAD
3540 	case EQ_OFLD:
3541 		rc = ofld_eq_alloc(sc, vi, eq);
3542 		break;
3543 #endif
3544 
3545 	default:
3546 		panic("%s: invalid eq type %d.", __func__,
3547 		    eq->flags & EQ_TYPEMASK);
3548 	}
3549 	if (rc != 0) {
3550 		device_printf(sc->dev,
3551 		    "failed to allocate egress queue(%d): %d\n",
3552 		    eq->flags & EQ_TYPEMASK, rc);
3553 	}
3554 
3555 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
3556 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
3557 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
3558 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3559 		uint32_t mask = (1 << s_qpp) - 1;
3560 		volatile uint8_t *udb;
3561 
3562 		udb = sc->udbs_base + UDBS_DB_OFFSET;
3563 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
3564 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
3565 		if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
3566 	    		clrbit(&eq->doorbells, DOORBELL_WCWR);
3567 		else {
3568 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
3569 			eq->udb_qid = 0;
3570 		}
3571 		eq->udb = (volatile void *)udb;
3572 	}
3573 
3574 	return (rc);
3575 }
3576 
3577 static int
3578 free_eq(struct adapter *sc, struct sge_eq *eq)
3579 {
3580 	int rc;
3581 
3582 	if (eq->flags & EQ_ALLOCATED) {
3583 		switch (eq->flags & EQ_TYPEMASK) {
3584 		case EQ_CTRL:
3585 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
3586 			    eq->cntxt_id);
3587 			break;
3588 
3589 		case EQ_ETH:
3590 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
3591 			    eq->cntxt_id);
3592 			break;
3593 
3594 #ifdef TCP_OFFLOAD
3595 		case EQ_OFLD:
3596 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
3597 			    eq->cntxt_id);
3598 			break;
3599 #endif
3600 
3601 		default:
3602 			panic("%s: invalid eq type %d.", __func__,
3603 			    eq->flags & EQ_TYPEMASK);
3604 		}
3605 		if (rc != 0) {
3606 			device_printf(sc->dev,
3607 			    "failed to free egress queue (%d): %d\n",
3608 			    eq->flags & EQ_TYPEMASK, rc);
3609 			return (rc);
3610 		}
3611 		eq->flags &= ~EQ_ALLOCATED;
3612 	}
3613 
3614 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
3615 
3616 	if (mtx_initialized(&eq->eq_lock))
3617 		mtx_destroy(&eq->eq_lock);
3618 
3619 	bzero(eq, sizeof(*eq));
3620 	return (0);
3621 }
3622 
3623 static int
3624 alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq,
3625     struct sysctl_oid *oid)
3626 {
3627 	int rc;
3628 	struct sysctl_ctx_list *ctx = vi ? &vi->ctx : &sc->ctx;
3629 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3630 
3631 	rc = alloc_eq(sc, vi, &wrq->eq);
3632 	if (rc)
3633 		return (rc);
3634 
3635 	wrq->adapter = sc;
3636 	TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq);
3637 	TAILQ_INIT(&wrq->incomplete_wrs);
3638 	STAILQ_INIT(&wrq->wr_list);
3639 	wrq->nwr_pending = 0;
3640 	wrq->ndesc_needed = 0;
3641 
3642 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3643 	    &wrq->eq.ba, "bus address of descriptor ring");
3644 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3645 	    wrq->eq.sidx * EQ_ESIZE + sc->params.sge.spg_len,
3646 	    "desc ring size in bytes");
3647 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3648 	    &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
3649 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3650 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I",
3651 	    "consumer index");
3652 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx",
3653 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I",
3654 	    "producer index");
3655 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
3656 	    wrq->eq.sidx, "status page index");
3657 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD,
3658 	    &wrq->tx_wrs_direct, "# of work requests (direct)");
3659 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD,
3660 	    &wrq->tx_wrs_copied, "# of work requests (copied)");
3661 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_sspace", CTLFLAG_RD,
3662 	    &wrq->tx_wrs_ss, "# of work requests (copied from scratch space)");
3663 
3664 	return (rc);
3665 }
3666 
3667 static int
3668 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
3669 {
3670 	int rc;
3671 
3672 	rc = free_eq(sc, &wrq->eq);
3673 	if (rc)
3674 		return (rc);
3675 
3676 	bzero(wrq, sizeof(*wrq));
3677 	return (0);
3678 }
3679 
3680 static int
3681 alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx,
3682     struct sysctl_oid *oid)
3683 {
3684 	int rc;
3685 	struct port_info *pi = vi->pi;
3686 	struct adapter *sc = pi->adapter;
3687 	struct sge_eq *eq = &txq->eq;
3688 	char name[16];
3689 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3690 
3691 	rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx, can_resume_eth_tx,
3692 	    M_CXGBE, M_WAITOK);
3693 	if (rc != 0) {
3694 		device_printf(sc->dev, "failed to allocate mp_ring: %d\n", rc);
3695 		return (rc);
3696 	}
3697 
3698 	rc = alloc_eq(sc, vi, eq);
3699 	if (rc != 0) {
3700 		mp_ring_free(txq->r);
3701 		txq->r = NULL;
3702 		return (rc);
3703 	}
3704 
3705 	/* Can't fail after this point. */
3706 
3707 	if (idx == 0)
3708 		sc->sge.eq_base = eq->abs_id - eq->cntxt_id;
3709 	else
3710 		KASSERT(eq->cntxt_id + sc->sge.eq_base == eq->abs_id,
3711 		    ("eq_base mismatch"));
3712 	KASSERT(sc->sge.eq_base == 0 || sc->flags & IS_VF,
3713 	    ("PF with non-zero eq_base"));
3714 
3715 	TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq);
3716 	txq->ifp = vi->ifp;
3717 	txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK);
3718 	if (sc->flags & IS_VF)
3719 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
3720 		    V_TXPKT_INTF(pi->tx_chan));
3721 	else
3722 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3723 		    V_TXPKT_INTF(pi->tx_chan) |
3724 		    V_TXPKT_PF(G_FW_VIID_PFN(vi->viid)) |
3725 		    V_TXPKT_VF(G_FW_VIID_VIN(vi->viid)) |
3726 		    V_TXPKT_VF_VLD(G_FW_VIID_VIVLD(vi->viid)));
3727 	txq->tc_idx = -1;
3728 	txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE,
3729 	    M_ZERO | M_WAITOK);
3730 
3731 	snprintf(name, sizeof(name), "%d", idx);
3732 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3733 	    NULL, "tx queue");
3734 	children = SYSCTL_CHILDREN(oid);
3735 
3736 	SYSCTL_ADD_UAUTO(&vi->ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3737 	    &eq->ba, "bus address of descriptor ring");
3738 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3739 	    eq->sidx * EQ_ESIZE + sc->params.sge.spg_len,
3740 	    "desc ring size in bytes");
3741 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
3742 	    &eq->abs_id, 0, "absolute id of the queue");
3743 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3744 	    &eq->cntxt_id, 0, "SGE context id of the queue");
3745 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3746 	    CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I",
3747 	    "consumer index");
3748 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3749 	    CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I",
3750 	    "producer index");
3751 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
3752 	    eq->sidx, "status page index");
3753 
3754 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "tc",
3755 	    CTLTYPE_INT | CTLFLAG_RW, vi, idx, sysctl_tc, "I",
3756 	    "traffic class (-1 means none)");
3757 
3758 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
3759 	    &txq->txcsum, "# of times hardware assisted with checksum");
3760 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_insertion",
3761 	    CTLFLAG_RD, &txq->vlan_insertion,
3762 	    "# of times hardware inserted 802.1Q tag");
3763 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
3764 	    &txq->tso_wrs, "# of TSO work requests");
3765 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
3766 	    &txq->imm_wrs, "# of work requests with immediate data");
3767 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
3768 	    &txq->sgl_wrs, "# of work requests with direct SGL");
3769 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
3770 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
3771 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_wrs",
3772 	    CTLFLAG_RD, &txq->txpkts0_wrs,
3773 	    "# of txpkts (type 0) work requests");
3774 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_wrs",
3775 	    CTLFLAG_RD, &txq->txpkts1_wrs,
3776 	    "# of txpkts (type 1) work requests");
3777 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_pkts",
3778 	    CTLFLAG_RD, &txq->txpkts0_pkts,
3779 	    "# of frames tx'd using type0 txpkts work requests");
3780 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_pkts",
3781 	    CTLFLAG_RD, &txq->txpkts1_pkts,
3782 	    "# of frames tx'd using type1 txpkts work requests");
3783 
3784 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_enqueues",
3785 	    CTLFLAG_RD, &txq->r->enqueues,
3786 	    "# of enqueues to the mp_ring for this queue");
3787 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_drops",
3788 	    CTLFLAG_RD, &txq->r->drops,
3789 	    "# of drops in the mp_ring for this queue");
3790 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_starts",
3791 	    CTLFLAG_RD, &txq->r->starts,
3792 	    "# of normal consumer starts in the mp_ring for this queue");
3793 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_stalls",
3794 	    CTLFLAG_RD, &txq->r->stalls,
3795 	    "# of consumer stalls in the mp_ring for this queue");
3796 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_restarts",
3797 	    CTLFLAG_RD, &txq->r->restarts,
3798 	    "# of consumer restarts in the mp_ring for this queue");
3799 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_abdications",
3800 	    CTLFLAG_RD, &txq->r->abdications,
3801 	    "# of consumer abdications in the mp_ring for this queue");
3802 
3803 	return (0);
3804 }
3805 
3806 static int
3807 free_txq(struct vi_info *vi, struct sge_txq *txq)
3808 {
3809 	int rc;
3810 	struct adapter *sc = vi->pi->adapter;
3811 	struct sge_eq *eq = &txq->eq;
3812 
3813 	rc = free_eq(sc, eq);
3814 	if (rc)
3815 		return (rc);
3816 
3817 	sglist_free(txq->gl);
3818 	free(txq->sdesc, M_CXGBE);
3819 	mp_ring_free(txq->r);
3820 
3821 	bzero(txq, sizeof(*txq));
3822 	return (0);
3823 }
3824 
3825 static void
3826 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
3827 {
3828 	bus_addr_t *ba = arg;
3829 
3830 	KASSERT(nseg == 1,
3831 	    ("%s meant for single segment mappings only.", __func__));
3832 
3833 	*ba = error ? 0 : segs->ds_addr;
3834 }
3835 
3836 static inline void
3837 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
3838 {
3839 	uint32_t n, v;
3840 
3841 	n = IDXDIFF(fl->pidx / 8, fl->dbidx, fl->sidx);
3842 	MPASS(n > 0);
3843 
3844 	wmb();
3845 	v = fl->dbval | V_PIDX(n);
3846 	if (fl->udb)
3847 		*fl->udb = htole32(v);
3848 	else
3849 		t4_write_reg(sc, sc->sge_kdoorbell_reg, v);
3850 	IDXINCR(fl->dbidx, n, fl->sidx);
3851 }
3852 
3853 /*
3854  * Fills up the freelist by allocating up to 'n' buffers.  Buffers that are
3855  * recycled do not count towards this allocation budget.
3856  *
3857  * Returns non-zero to indicate that this freelist should be added to the list
3858  * of starving freelists.
3859  */
3860 static int
3861 refill_fl(struct adapter *sc, struct sge_fl *fl, int n)
3862 {
3863 	__be64 *d;
3864 	struct fl_sdesc *sd;
3865 	uintptr_t pa;
3866 	caddr_t cl;
3867 	struct cluster_layout *cll;
3868 	struct sw_zone_info *swz;
3869 	struct cluster_metadata *clm;
3870 	uint16_t max_pidx;
3871 	uint16_t hw_cidx = fl->hw_cidx;		/* stable snapshot */
3872 
3873 	FL_LOCK_ASSERT_OWNED(fl);
3874 
3875 	/*
3876 	 * We always stop at the beginning of the hardware descriptor that's just
3877 	 * before the one with the hw cidx.  This is to avoid hw pidx = hw cidx,
3878 	 * which would mean an empty freelist to the chip.
3879 	 */
3880 	max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1;
3881 	if (fl->pidx == max_pidx * 8)
3882 		return (0);
3883 
3884 	d = &fl->desc[fl->pidx];
3885 	sd = &fl->sdesc[fl->pidx];
3886 	cll = &fl->cll_def;	/* default layout */
3887 	swz = &sc->sge.sw_zone_info[cll->zidx];
3888 
3889 	while (n > 0) {
3890 
3891 		if (sd->cl != NULL) {
3892 
3893 			if (sd->nmbuf == 0) {
3894 				/*
3895 				 * Fast recycle without involving any atomics on
3896 				 * the cluster's metadata (if the cluster has
3897 				 * metadata).  This happens when all frames
3898 				 * received in the cluster were small enough to
3899 				 * fit within a single mbuf each.
3900 				 */
3901 				fl->cl_fast_recycled++;
3902 #ifdef INVARIANTS
3903 				clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
3904 				if (clm != NULL)
3905 					MPASS(clm->refcount == 1);
3906 #endif
3907 				goto recycled_fast;
3908 			}
3909 
3910 			/*
3911 			 * Cluster is guaranteed to have metadata.  Clusters
3912 			 * without metadata always take the fast recycle path
3913 			 * when they're recycled.
3914 			 */
3915 			clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
3916 			MPASS(clm != NULL);
3917 
3918 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
3919 				fl->cl_recycled++;
3920 				counter_u64_add(extfree_rels, 1);
3921 				goto recycled;
3922 			}
3923 			sd->cl = NULL;	/* gave up my reference */
3924 		}
3925 		MPASS(sd->cl == NULL);
3926 alloc:
3927 		cl = uma_zalloc(swz->zone, M_NOWAIT);
3928 		if (__predict_false(cl == NULL)) {
3929 			if (cll == &fl->cll_alt || fl->cll_alt.zidx == -1 ||
3930 			    fl->cll_def.zidx == fl->cll_alt.zidx)
3931 				break;
3932 
3933 			/* fall back to the safe zone */
3934 			cll = &fl->cll_alt;
3935 			swz = &sc->sge.sw_zone_info[cll->zidx];
3936 			goto alloc;
3937 		}
3938 		fl->cl_allocated++;
3939 		n--;
3940 
3941 		pa = pmap_kextract((vm_offset_t)cl);
3942 		pa += cll->region1;
3943 		sd->cl = cl;
3944 		sd->cll = *cll;
3945 		*d = htobe64(pa | cll->hwidx);
3946 		clm = cl_metadata(sc, fl, cll, cl);
3947 		if (clm != NULL) {
3948 recycled:
3949 #ifdef INVARIANTS
3950 			clm->sd = sd;
3951 #endif
3952 			clm->refcount = 1;
3953 		}
3954 		sd->nmbuf = 0;
3955 recycled_fast:
3956 		d++;
3957 		sd++;
3958 		if (__predict_false(++fl->pidx % 8 == 0)) {
3959 			uint16_t pidx = fl->pidx / 8;
3960 
3961 			if (__predict_false(pidx == fl->sidx)) {
3962 				fl->pidx = 0;
3963 				pidx = 0;
3964 				sd = fl->sdesc;
3965 				d = fl->desc;
3966 			}
3967 			if (pidx == max_pidx)
3968 				break;
3969 
3970 			if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4)
3971 				ring_fl_db(sc, fl);
3972 		}
3973 	}
3974 
3975 	if (fl->pidx / 8 != fl->dbidx)
3976 		ring_fl_db(sc, fl);
3977 
3978 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
3979 }
3980 
3981 /*
3982  * Attempt to refill all starving freelists.
3983  */
3984 static void
3985 refill_sfl(void *arg)
3986 {
3987 	struct adapter *sc = arg;
3988 	struct sge_fl *fl, *fl_temp;
3989 
3990 	mtx_assert(&sc->sfl_lock, MA_OWNED);
3991 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
3992 		FL_LOCK(fl);
3993 		refill_fl(sc, fl, 64);
3994 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
3995 			TAILQ_REMOVE(&sc->sfl, fl, link);
3996 			fl->flags &= ~FL_STARVING;
3997 		}
3998 		FL_UNLOCK(fl);
3999 	}
4000 
4001 	if (!TAILQ_EMPTY(&sc->sfl))
4002 		callout_schedule(&sc->sfl_callout, hz / 5);
4003 }
4004 
4005 static int
4006 alloc_fl_sdesc(struct sge_fl *fl)
4007 {
4008 
4009 	fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc), M_CXGBE,
4010 	    M_ZERO | M_WAITOK);
4011 
4012 	return (0);
4013 }
4014 
4015 static void
4016 free_fl_sdesc(struct adapter *sc, struct sge_fl *fl)
4017 {
4018 	struct fl_sdesc *sd;
4019 	struct cluster_metadata *clm;
4020 	struct cluster_layout *cll;
4021 	int i;
4022 
4023 	sd = fl->sdesc;
4024 	for (i = 0; i < fl->sidx * 8; i++, sd++) {
4025 		if (sd->cl == NULL)
4026 			continue;
4027 
4028 		cll = &sd->cll;
4029 		clm = cl_metadata(sc, fl, cll, sd->cl);
4030 		if (sd->nmbuf == 0)
4031 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4032 		else if (clm && atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4033 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4034 			counter_u64_add(extfree_rels, 1);
4035 		}
4036 		sd->cl = NULL;
4037 	}
4038 
4039 	free(fl->sdesc, M_CXGBE);
4040 	fl->sdesc = NULL;
4041 }
4042 
4043 static inline void
4044 get_pkt_gl(struct mbuf *m, struct sglist *gl)
4045 {
4046 	int rc;
4047 
4048 	M_ASSERTPKTHDR(m);
4049 
4050 	sglist_reset(gl);
4051 	rc = sglist_append_mbuf(gl, m);
4052 	if (__predict_false(rc != 0)) {
4053 		panic("%s: mbuf %p (%d segs) was vetted earlier but now fails "
4054 		    "with %d.", __func__, m, mbuf_nsegs(m), rc);
4055 	}
4056 
4057 	KASSERT(gl->sg_nseg == mbuf_nsegs(m),
4058 	    ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m,
4059 	    mbuf_nsegs(m), gl->sg_nseg));
4060 	KASSERT(gl->sg_nseg > 0 &&
4061 	    gl->sg_nseg <= (needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS),
4062 	    ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__,
4063 		gl->sg_nseg, needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS));
4064 }
4065 
4066 /*
4067  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
4068  */
4069 static inline u_int
4070 txpkt_len16(u_int nsegs, u_int tso)
4071 {
4072 	u_int n;
4073 
4074 	MPASS(nsegs > 0);
4075 
4076 	nsegs--; /* first segment is part of ulptx_sgl */
4077 	n = sizeof(struct fw_eth_tx_pkt_wr) + sizeof(struct cpl_tx_pkt_core) +
4078 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4079 	if (tso)
4080 		n += sizeof(struct cpl_tx_pkt_lso_core);
4081 
4082 	return (howmany(n, 16));
4083 }
4084 
4085 /*
4086  * len16 for a txpkt_vm WR with a GL.  Includes the firmware work
4087  * request header.
4088  */
4089 static inline u_int
4090 txpkt_vm_len16(u_int nsegs, u_int tso)
4091 {
4092 	u_int n;
4093 
4094 	MPASS(nsegs > 0);
4095 
4096 	nsegs--; /* first segment is part of ulptx_sgl */
4097 	n = sizeof(struct fw_eth_tx_pkt_vm_wr) +
4098 	    sizeof(struct cpl_tx_pkt_core) +
4099 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4100 	if (tso)
4101 		n += sizeof(struct cpl_tx_pkt_lso_core);
4102 
4103 	return (howmany(n, 16));
4104 }
4105 
4106 /*
4107  * len16 for a txpkts type 0 WR with a GL.  Does not include the firmware work
4108  * request header.
4109  */
4110 static inline u_int
4111 txpkts0_len16(u_int nsegs)
4112 {
4113 	u_int n;
4114 
4115 	MPASS(nsegs > 0);
4116 
4117 	nsegs--; /* first segment is part of ulptx_sgl */
4118 	n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) +
4119 	    sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) +
4120 	    8 * ((3 * nsegs) / 2 + (nsegs & 1));
4121 
4122 	return (howmany(n, 16));
4123 }
4124 
4125 /*
4126  * len16 for a txpkts type 1 WR with a GL.  Does not include the firmware work
4127  * request header.
4128  */
4129 static inline u_int
4130 txpkts1_len16(void)
4131 {
4132 	u_int n;
4133 
4134 	n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl);
4135 
4136 	return (howmany(n, 16));
4137 }
4138 
4139 static inline u_int
4140 imm_payload(u_int ndesc)
4141 {
4142 	u_int n;
4143 
4144 	n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) -
4145 	    sizeof(struct cpl_tx_pkt_core);
4146 
4147 	return (n);
4148 }
4149 
4150 /*
4151  * Write a VM txpkt WR for this packet to the hardware descriptors, update the
4152  * software descriptor, and advance the pidx.  It is guaranteed that enough
4153  * descriptors are available.
4154  *
4155  * The return value is the # of hardware descriptors used.
4156  */
4157 static u_int
4158 write_txpkt_vm_wr(struct adapter *sc, struct sge_txq *txq,
4159     struct fw_eth_tx_pkt_vm_wr *wr, struct mbuf *m0, u_int available)
4160 {
4161 	struct sge_eq *eq = &txq->eq;
4162 	struct tx_sdesc *txsd;
4163 	struct cpl_tx_pkt_core *cpl;
4164 	uint32_t ctrl;	/* used in many unrelated places */
4165 	uint64_t ctrl1;
4166 	int csum_type, len16, ndesc, pktlen, nsegs;
4167 	caddr_t dst;
4168 
4169 	TXQ_LOCK_ASSERT_OWNED(txq);
4170 	M_ASSERTPKTHDR(m0);
4171 	MPASS(available > 0 && available < eq->sidx);
4172 
4173 	len16 = mbuf_len16(m0);
4174 	nsegs = mbuf_nsegs(m0);
4175 	pktlen = m0->m_pkthdr.len;
4176 	ctrl = sizeof(struct cpl_tx_pkt_core);
4177 	if (needs_tso(m0))
4178 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4179 	ndesc = howmany(len16, EQ_ESIZE / 16);
4180 	MPASS(ndesc <= available);
4181 
4182 	/* Firmware work request header */
4183 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4184 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_VM_WR) |
4185 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4186 
4187 	ctrl = V_FW_WR_LEN16(len16);
4188 	wr->equiq_to_len16 = htobe32(ctrl);
4189 	wr->r3[0] = 0;
4190 	wr->r3[1] = 0;
4191 
4192 	/*
4193 	 * Copy over ethmacdst, ethmacsrc, ethtype, and vlantci.
4194 	 * vlantci is ignored unless the ethtype is 0x8100, so it's
4195 	 * simpler to always copy it rather than making it
4196 	 * conditional.  Also, it seems that we do not have to set
4197 	 * vlantci or fake the ethtype when doing VLAN tag insertion.
4198 	 */
4199 	m_copydata(m0, 0, sizeof(struct ether_header) + 2, wr->ethmacdst);
4200 
4201 	csum_type = -1;
4202 	if (needs_tso(m0)) {
4203 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4204 
4205 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4206 		    m0->m_pkthdr.l4hlen > 0,
4207 		    ("%s: mbuf %p needs TSO but missing header lengths",
4208 			__func__, m0));
4209 
4210 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4211 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4212 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4213 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4214 			ctrl |= V_LSO_ETHHDR_LEN(1);
4215 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4216 			ctrl |= F_LSO_IPV6;
4217 
4218 		lso->lso_ctrl = htobe32(ctrl);
4219 		lso->ipid_ofst = htobe16(0);
4220 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4221 		lso->seqno_offset = htobe32(0);
4222 		lso->len = htobe32(pktlen);
4223 
4224 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4225 			csum_type = TX_CSUM_TCPIP6;
4226 		else
4227 			csum_type = TX_CSUM_TCPIP;
4228 
4229 		cpl = (void *)(lso + 1);
4230 
4231 		txq->tso_wrs++;
4232 	} else {
4233 		if (m0->m_pkthdr.csum_flags & CSUM_IP_TCP)
4234 			csum_type = TX_CSUM_TCPIP;
4235 		else if (m0->m_pkthdr.csum_flags & CSUM_IP_UDP)
4236 			csum_type = TX_CSUM_UDPIP;
4237 		else if (m0->m_pkthdr.csum_flags & CSUM_IP6_TCP)
4238 			csum_type = TX_CSUM_TCPIP6;
4239 		else if (m0->m_pkthdr.csum_flags & CSUM_IP6_UDP)
4240 			csum_type = TX_CSUM_UDPIP6;
4241 #if defined(INET)
4242 		else if (m0->m_pkthdr.csum_flags & CSUM_IP) {
4243 			/*
4244 			 * XXX: The firmware appears to stomp on the
4245 			 * fragment/flags field of the IP header when
4246 			 * using TX_CSUM_IP.  Fall back to doing
4247 			 * software checksums.
4248 			 */
4249 			u_short *sump;
4250 			struct mbuf *m;
4251 			int offset;
4252 
4253 			m = m0;
4254 			offset = 0;
4255 			sump = m_advance(&m, &offset, m0->m_pkthdr.l2hlen +
4256 			    offsetof(struct ip, ip_sum));
4257 			*sump = in_cksum_skip(m0, m0->m_pkthdr.l2hlen +
4258 			    m0->m_pkthdr.l3hlen, m0->m_pkthdr.l2hlen);
4259 			m0->m_pkthdr.csum_flags &= ~CSUM_IP;
4260 		}
4261 #endif
4262 
4263 		cpl = (void *)(wr + 1);
4264 	}
4265 
4266 	/* Checksum offload */
4267 	ctrl1 = 0;
4268 	if (needs_l3_csum(m0) == 0)
4269 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4270 	if (csum_type >= 0) {
4271 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0,
4272 	    ("%s: mbuf %p needs checksum offload but missing header lengths",
4273 			__func__, m0));
4274 
4275 		if (chip_id(sc) <= CHELSIO_T5) {
4276 			ctrl1 |= V_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4277 			    ETHER_HDR_LEN);
4278 		} else {
4279 			ctrl1 |= V_T6_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4280 			    ETHER_HDR_LEN);
4281 		}
4282 		ctrl1 |= V_TXPKT_IPHDR_LEN(m0->m_pkthdr.l3hlen);
4283 		ctrl1 |= V_TXPKT_CSUM_TYPE(csum_type);
4284 	} else
4285 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4286 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4287 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4288 		txq->txcsum++;	/* some hardware assistance provided */
4289 
4290 	/* VLAN tag insertion */
4291 	if (needs_vlan_insertion(m0)) {
4292 		ctrl1 |= F_TXPKT_VLAN_VLD |
4293 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4294 		txq->vlan_insertion++;
4295 	}
4296 
4297 	/* CPL header */
4298 	cpl->ctrl0 = txq->cpl_ctrl0;
4299 	cpl->pack = 0;
4300 	cpl->len = htobe16(pktlen);
4301 	cpl->ctrl1 = htobe64(ctrl1);
4302 
4303 	/* SGL */
4304 	dst = (void *)(cpl + 1);
4305 
4306 	/*
4307 	 * A packet using TSO will use up an entire descriptor for the
4308 	 * firmware work request header, LSO CPL, and TX_PKT_XT CPL.
4309 	 * If this descriptor is the last descriptor in the ring, wrap
4310 	 * around to the front of the ring explicitly for the start of
4311 	 * the sgl.
4312 	 */
4313 	if (dst == (void *)&eq->desc[eq->sidx]) {
4314 		dst = (void *)&eq->desc[0];
4315 		write_gl_to_txd(txq, m0, &dst, 0);
4316 	} else
4317 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4318 	txq->sgl_wrs++;
4319 
4320 	txq->txpkt_wrs++;
4321 
4322 	txsd = &txq->sdesc[eq->pidx];
4323 	txsd->m = m0;
4324 	txsd->desc_used = ndesc;
4325 
4326 	return (ndesc);
4327 }
4328 
4329 /*
4330  * Write a txpkt WR for this packet to the hardware descriptors, update the
4331  * software descriptor, and advance the pidx.  It is guaranteed that enough
4332  * descriptors are available.
4333  *
4334  * The return value is the # of hardware descriptors used.
4335  */
4336 static u_int
4337 write_txpkt_wr(struct sge_txq *txq, struct fw_eth_tx_pkt_wr *wr,
4338     struct mbuf *m0, u_int available)
4339 {
4340 	struct sge_eq *eq = &txq->eq;
4341 	struct tx_sdesc *txsd;
4342 	struct cpl_tx_pkt_core *cpl;
4343 	uint32_t ctrl;	/* used in many unrelated places */
4344 	uint64_t ctrl1;
4345 	int len16, ndesc, pktlen, nsegs;
4346 	caddr_t dst;
4347 
4348 	TXQ_LOCK_ASSERT_OWNED(txq);
4349 	M_ASSERTPKTHDR(m0);
4350 	MPASS(available > 0 && available < eq->sidx);
4351 
4352 	len16 = mbuf_len16(m0);
4353 	nsegs = mbuf_nsegs(m0);
4354 	pktlen = m0->m_pkthdr.len;
4355 	ctrl = sizeof(struct cpl_tx_pkt_core);
4356 	if (needs_tso(m0))
4357 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4358 	else if (pktlen <= imm_payload(2) && available >= 2) {
4359 		/* Immediate data.  Recalculate len16 and set nsegs to 0. */
4360 		ctrl += pktlen;
4361 		len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) +
4362 		    sizeof(struct cpl_tx_pkt_core) + pktlen, 16);
4363 		nsegs = 0;
4364 	}
4365 	ndesc = howmany(len16, EQ_ESIZE / 16);
4366 	MPASS(ndesc <= available);
4367 
4368 	/* Firmware work request header */
4369 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4370 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
4371 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4372 
4373 	ctrl = V_FW_WR_LEN16(len16);
4374 	wr->equiq_to_len16 = htobe32(ctrl);
4375 	wr->r3 = 0;
4376 
4377 	if (needs_tso(m0)) {
4378 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4379 
4380 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4381 		    m0->m_pkthdr.l4hlen > 0,
4382 		    ("%s: mbuf %p needs TSO but missing header lengths",
4383 			__func__, m0));
4384 
4385 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4386 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4387 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4388 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4389 			ctrl |= V_LSO_ETHHDR_LEN(1);
4390 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4391 			ctrl |= F_LSO_IPV6;
4392 
4393 		lso->lso_ctrl = htobe32(ctrl);
4394 		lso->ipid_ofst = htobe16(0);
4395 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4396 		lso->seqno_offset = htobe32(0);
4397 		lso->len = htobe32(pktlen);
4398 
4399 		cpl = (void *)(lso + 1);
4400 
4401 		txq->tso_wrs++;
4402 	} else
4403 		cpl = (void *)(wr + 1);
4404 
4405 	/* Checksum offload */
4406 	ctrl1 = 0;
4407 	if (needs_l3_csum(m0) == 0)
4408 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4409 	if (needs_l4_csum(m0) == 0)
4410 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4411 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4412 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4413 		txq->txcsum++;	/* some hardware assistance provided */
4414 
4415 	/* VLAN tag insertion */
4416 	if (needs_vlan_insertion(m0)) {
4417 		ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4418 		txq->vlan_insertion++;
4419 	}
4420 
4421 	/* CPL header */
4422 	cpl->ctrl0 = txq->cpl_ctrl0;
4423 	cpl->pack = 0;
4424 	cpl->len = htobe16(pktlen);
4425 	cpl->ctrl1 = htobe64(ctrl1);
4426 
4427 	/* SGL */
4428 	dst = (void *)(cpl + 1);
4429 	if (nsegs > 0) {
4430 
4431 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4432 		txq->sgl_wrs++;
4433 	} else {
4434 		struct mbuf *m;
4435 
4436 		for (m = m0; m != NULL; m = m->m_next) {
4437 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4438 #ifdef INVARIANTS
4439 			pktlen -= m->m_len;
4440 #endif
4441 		}
4442 #ifdef INVARIANTS
4443 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
4444 #endif
4445 		txq->imm_wrs++;
4446 	}
4447 
4448 	txq->txpkt_wrs++;
4449 
4450 	txsd = &txq->sdesc[eq->pidx];
4451 	txsd->m = m0;
4452 	txsd->desc_used = ndesc;
4453 
4454 	return (ndesc);
4455 }
4456 
4457 static int
4458 try_txpkts(struct mbuf *m, struct mbuf *n, struct txpkts *txp, u_int available)
4459 {
4460 	u_int needed, nsegs1, nsegs2, l1, l2;
4461 
4462 	if (cannot_use_txpkts(m) || cannot_use_txpkts(n))
4463 		return (1);
4464 
4465 	nsegs1 = mbuf_nsegs(m);
4466 	nsegs2 = mbuf_nsegs(n);
4467 	if (nsegs1 + nsegs2 == 2) {
4468 		txp->wr_type = 1;
4469 		l1 = l2 = txpkts1_len16();
4470 	} else {
4471 		txp->wr_type = 0;
4472 		l1 = txpkts0_len16(nsegs1);
4473 		l2 = txpkts0_len16(nsegs2);
4474 	}
4475 	txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) + l1 + l2;
4476 	needed = howmany(txp->len16, EQ_ESIZE / 16);
4477 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4478 		return (1);
4479 
4480 	txp->plen = m->m_pkthdr.len + n->m_pkthdr.len;
4481 	if (txp->plen > 65535)
4482 		return (1);
4483 
4484 	txp->npkt = 2;
4485 	set_mbuf_len16(m, l1);
4486 	set_mbuf_len16(n, l2);
4487 
4488 	return (0);
4489 }
4490 
4491 static int
4492 add_to_txpkts(struct mbuf *m, struct txpkts *txp, u_int available)
4493 {
4494 	u_int plen, len16, needed, nsegs;
4495 
4496 	MPASS(txp->wr_type == 0 || txp->wr_type == 1);
4497 
4498 	nsegs = mbuf_nsegs(m);
4499 	if (needs_tso(m) || (txp->wr_type == 1 && nsegs != 1))
4500 		return (1);
4501 
4502 	plen = txp->plen + m->m_pkthdr.len;
4503 	if (plen > 65535)
4504 		return (1);
4505 
4506 	if (txp->wr_type == 0)
4507 		len16 = txpkts0_len16(nsegs);
4508 	else
4509 		len16 = txpkts1_len16();
4510 	needed = howmany(txp->len16 + len16, EQ_ESIZE / 16);
4511 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4512 		return (1);
4513 
4514 	txp->npkt++;
4515 	txp->plen = plen;
4516 	txp->len16 += len16;
4517 	set_mbuf_len16(m, len16);
4518 
4519 	return (0);
4520 }
4521 
4522 /*
4523  * Write a txpkts WR for the packets in txp to the hardware descriptors, update
4524  * the software descriptor, and advance the pidx.  It is guaranteed that enough
4525  * descriptors are available.
4526  *
4527  * The return value is the # of hardware descriptors used.
4528  */
4529 static u_int
4530 write_txpkts_wr(struct sge_txq *txq, struct fw_eth_tx_pkts_wr *wr,
4531     struct mbuf *m0, const struct txpkts *txp, u_int available)
4532 {
4533 	struct sge_eq *eq = &txq->eq;
4534 	struct tx_sdesc *txsd;
4535 	struct cpl_tx_pkt_core *cpl;
4536 	uint32_t ctrl;
4537 	uint64_t ctrl1;
4538 	int ndesc, checkwrap;
4539 	struct mbuf *m;
4540 	void *flitp;
4541 
4542 	TXQ_LOCK_ASSERT_OWNED(txq);
4543 	MPASS(txp->npkt > 0);
4544 	MPASS(txp->plen < 65536);
4545 	MPASS(m0 != NULL);
4546 	MPASS(m0->m_nextpkt != NULL);
4547 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
4548 	MPASS(available > 0 && available < eq->sidx);
4549 
4550 	ndesc = howmany(txp->len16, EQ_ESIZE / 16);
4551 	MPASS(ndesc <= available);
4552 
4553 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4554 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
4555 	ctrl = V_FW_WR_LEN16(txp->len16);
4556 	wr->equiq_to_len16 = htobe32(ctrl);
4557 	wr->plen = htobe16(txp->plen);
4558 	wr->npkt = txp->npkt;
4559 	wr->r3 = 0;
4560 	wr->type = txp->wr_type;
4561 	flitp = wr + 1;
4562 
4563 	/*
4564 	 * At this point we are 16B into a hardware descriptor.  If checkwrap is
4565 	 * set then we know the WR is going to wrap around somewhere.  We'll
4566 	 * check for that at appropriate points.
4567 	 */
4568 	checkwrap = eq->sidx - ndesc < eq->pidx;
4569 	for (m = m0; m != NULL; m = m->m_nextpkt) {
4570 		if (txp->wr_type == 0) {
4571 			struct ulp_txpkt *ulpmc;
4572 			struct ulptx_idata *ulpsc;
4573 
4574 			/* ULP master command */
4575 			ulpmc = flitp;
4576 			ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
4577 			    V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid));
4578 			ulpmc->len = htobe32(mbuf_len16(m));
4579 
4580 			/* ULP subcommand */
4581 			ulpsc = (void *)(ulpmc + 1);
4582 			ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
4583 			    F_ULP_TX_SC_MORE);
4584 			ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
4585 
4586 			cpl = (void *)(ulpsc + 1);
4587 			if (checkwrap &&
4588 			    (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx])
4589 				cpl = (void *)&eq->desc[0];
4590 		} else {
4591 			cpl = flitp;
4592 		}
4593 
4594 		/* Checksum offload */
4595 		ctrl1 = 0;
4596 		if (needs_l3_csum(m) == 0)
4597 			ctrl1 |= F_TXPKT_IPCSUM_DIS;
4598 		if (needs_l4_csum(m) == 0)
4599 			ctrl1 |= F_TXPKT_L4CSUM_DIS;
4600 		if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4601 		    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4602 			txq->txcsum++;	/* some hardware assistance provided */
4603 
4604 		/* VLAN tag insertion */
4605 		if (needs_vlan_insertion(m)) {
4606 			ctrl1 |= F_TXPKT_VLAN_VLD |
4607 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
4608 			txq->vlan_insertion++;
4609 		}
4610 
4611 		/* CPL header */
4612 		cpl->ctrl0 = txq->cpl_ctrl0;
4613 		cpl->pack = 0;
4614 		cpl->len = htobe16(m->m_pkthdr.len);
4615 		cpl->ctrl1 = htobe64(ctrl1);
4616 
4617 		flitp = cpl + 1;
4618 		if (checkwrap &&
4619 		    (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
4620 			flitp = (void *)&eq->desc[0];
4621 
4622 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap);
4623 
4624 	}
4625 
4626 	if (txp->wr_type == 0) {
4627 		txq->txpkts0_pkts += txp->npkt;
4628 		txq->txpkts0_wrs++;
4629 	} else {
4630 		txq->txpkts1_pkts += txp->npkt;
4631 		txq->txpkts1_wrs++;
4632 	}
4633 
4634 	txsd = &txq->sdesc[eq->pidx];
4635 	txsd->m = m0;
4636 	txsd->desc_used = ndesc;
4637 
4638 	return (ndesc);
4639 }
4640 
4641 /*
4642  * If the SGL ends on an address that is not 16 byte aligned, this function will
4643  * add a 0 filled flit at the end.
4644  */
4645 static void
4646 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap)
4647 {
4648 	struct sge_eq *eq = &txq->eq;
4649 	struct sglist *gl = txq->gl;
4650 	struct sglist_seg *seg;
4651 	__be64 *flitp, *wrap;
4652 	struct ulptx_sgl *usgl;
4653 	int i, nflits, nsegs;
4654 
4655 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
4656 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
4657 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
4658 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
4659 
4660 	get_pkt_gl(m, gl);
4661 	nsegs = gl->sg_nseg;
4662 	MPASS(nsegs > 0);
4663 
4664 	nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
4665 	flitp = (__be64 *)(*to);
4666 	wrap = (__be64 *)(&eq->desc[eq->sidx]);
4667 	seg = &gl->sg_segs[0];
4668 	usgl = (void *)flitp;
4669 
4670 	/*
4671 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
4672 	 * ring, so we're at least 16 bytes away from the status page.  There is
4673 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
4674 	 */
4675 
4676 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
4677 	    V_ULPTX_NSGE(nsegs));
4678 	usgl->len0 = htobe32(seg->ss_len);
4679 	usgl->addr0 = htobe64(seg->ss_paddr);
4680 	seg++;
4681 
4682 	if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) {
4683 
4684 		/* Won't wrap around at all */
4685 
4686 		for (i = 0; i < nsegs - 1; i++, seg++) {
4687 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
4688 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
4689 		}
4690 		if (i & 1)
4691 			usgl->sge[i / 2].len[1] = htobe32(0);
4692 		flitp += nflits;
4693 	} else {
4694 
4695 		/* Will wrap somewhere in the rest of the SGL */
4696 
4697 		/* 2 flits already written, write the rest flit by flit */
4698 		flitp = (void *)(usgl + 1);
4699 		for (i = 0; i < nflits - 2; i++) {
4700 			if (flitp == wrap)
4701 				flitp = (void *)eq->desc;
4702 			*flitp++ = get_flit(seg, nsegs - 1, i);
4703 		}
4704 	}
4705 
4706 	if (nflits & 1) {
4707 		MPASS(((uintptr_t)flitp) & 0xf);
4708 		*flitp++ = 0;
4709 	}
4710 
4711 	MPASS((((uintptr_t)flitp) & 0xf) == 0);
4712 	if (__predict_false(flitp == wrap))
4713 		*to = (void *)eq->desc;
4714 	else
4715 		*to = (void *)flitp;
4716 }
4717 
4718 static inline void
4719 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
4720 {
4721 
4722 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
4723 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
4724 
4725 	if (__predict_true((uintptr_t)(*to) + len <=
4726 	    (uintptr_t)&eq->desc[eq->sidx])) {
4727 		bcopy(from, *to, len);
4728 		(*to) += len;
4729 	} else {
4730 		int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
4731 
4732 		bcopy(from, *to, portion);
4733 		from += portion;
4734 		portion = len - portion;	/* remaining */
4735 		bcopy(from, (void *)eq->desc, portion);
4736 		(*to) = (caddr_t)eq->desc + portion;
4737 	}
4738 }
4739 
4740 static inline void
4741 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n)
4742 {
4743 	u_int db;
4744 
4745 	MPASS(n > 0);
4746 
4747 	db = eq->doorbells;
4748 	if (n > 1)
4749 		clrbit(&db, DOORBELL_WCWR);
4750 	wmb();
4751 
4752 	switch (ffs(db) - 1) {
4753 	case DOORBELL_UDB:
4754 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
4755 		break;
4756 
4757 	case DOORBELL_WCWR: {
4758 		volatile uint64_t *dst, *src;
4759 		int i;
4760 
4761 		/*
4762 		 * Queues whose 128B doorbell segment fits in the page do not
4763 		 * use relative qid (udb_qid is always 0).  Only queues with
4764 		 * doorbell segments can do WCWR.
4765 		 */
4766 		KASSERT(eq->udb_qid == 0 && n == 1,
4767 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
4768 		    __func__, eq->doorbells, n, eq->dbidx, eq));
4769 
4770 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
4771 		    UDBS_DB_OFFSET);
4772 		i = eq->dbidx;
4773 		src = (void *)&eq->desc[i];
4774 		while (src != (void *)&eq->desc[i + 1])
4775 			*dst++ = *src++;
4776 		wmb();
4777 		break;
4778 	}
4779 
4780 	case DOORBELL_UDBWC:
4781 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
4782 		wmb();
4783 		break;
4784 
4785 	case DOORBELL_KDB:
4786 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
4787 		    V_QID(eq->cntxt_id) | V_PIDX(n));
4788 		break;
4789 	}
4790 
4791 	IDXINCR(eq->dbidx, n, eq->sidx);
4792 }
4793 
4794 static inline u_int
4795 reclaimable_tx_desc(struct sge_eq *eq)
4796 {
4797 	uint16_t hw_cidx;
4798 
4799 	hw_cidx = read_hw_cidx(eq);
4800 	return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx));
4801 }
4802 
4803 static inline u_int
4804 total_available_tx_desc(struct sge_eq *eq)
4805 {
4806 	uint16_t hw_cidx, pidx;
4807 
4808 	hw_cidx = read_hw_cidx(eq);
4809 	pidx = eq->pidx;
4810 
4811 	if (pidx == hw_cidx)
4812 		return (eq->sidx - 1);
4813 	else
4814 		return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1);
4815 }
4816 
4817 static inline uint16_t
4818 read_hw_cidx(struct sge_eq *eq)
4819 {
4820 	struct sge_qstat *spg = (void *)&eq->desc[eq->sidx];
4821 	uint16_t cidx = spg->cidx;	/* stable snapshot */
4822 
4823 	return (be16toh(cidx));
4824 }
4825 
4826 /*
4827  * Reclaim 'n' descriptors approximately.
4828  */
4829 static u_int
4830 reclaim_tx_descs(struct sge_txq *txq, u_int n)
4831 {
4832 	struct tx_sdesc *txsd;
4833 	struct sge_eq *eq = &txq->eq;
4834 	u_int can_reclaim, reclaimed;
4835 
4836 	TXQ_LOCK_ASSERT_OWNED(txq);
4837 	MPASS(n > 0);
4838 
4839 	reclaimed = 0;
4840 	can_reclaim = reclaimable_tx_desc(eq);
4841 	while (can_reclaim && reclaimed < n) {
4842 		int ndesc;
4843 		struct mbuf *m, *nextpkt;
4844 
4845 		txsd = &txq->sdesc[eq->cidx];
4846 		ndesc = txsd->desc_used;
4847 
4848 		/* Firmware doesn't return "partial" credits. */
4849 		KASSERT(can_reclaim >= ndesc,
4850 		    ("%s: unexpected number of credits: %d, %d",
4851 		    __func__, can_reclaim, ndesc));
4852 
4853 		for (m = txsd->m; m != NULL; m = nextpkt) {
4854 			nextpkt = m->m_nextpkt;
4855 			m->m_nextpkt = NULL;
4856 			m_freem(m);
4857 		}
4858 		reclaimed += ndesc;
4859 		can_reclaim -= ndesc;
4860 		IDXINCR(eq->cidx, ndesc, eq->sidx);
4861 	}
4862 
4863 	return (reclaimed);
4864 }
4865 
4866 static void
4867 tx_reclaim(void *arg, int n)
4868 {
4869 	struct sge_txq *txq = arg;
4870 	struct sge_eq *eq = &txq->eq;
4871 
4872 	do {
4873 		if (TXQ_TRYLOCK(txq) == 0)
4874 			break;
4875 		n = reclaim_tx_descs(txq, 32);
4876 		if (eq->cidx == eq->pidx)
4877 			eq->equeqidx = eq->pidx;
4878 		TXQ_UNLOCK(txq);
4879 	} while (n > 0);
4880 }
4881 
4882 static __be64
4883 get_flit(struct sglist_seg *segs, int nsegs, int idx)
4884 {
4885 	int i = (idx / 3) * 2;
4886 
4887 	switch (idx % 3) {
4888 	case 0: {
4889 		__be64 rc;
4890 
4891 		rc = htobe32(segs[i].ss_len);
4892 		if (i + 1 < nsegs)
4893 			rc |= (uint64_t)htobe32(segs[i + 1].ss_len) << 32;
4894 
4895 		return (rc);
4896 	}
4897 	case 1:
4898 		return (htobe64(segs[i].ss_paddr));
4899 	case 2:
4900 		return (htobe64(segs[i + 1].ss_paddr));
4901 	}
4902 
4903 	return (0);
4904 }
4905 
4906 static void
4907 find_best_refill_source(struct adapter *sc, struct sge_fl *fl, int maxp)
4908 {
4909 	int8_t zidx, hwidx, idx;
4910 	uint16_t region1, region3;
4911 	int spare, spare_needed, n;
4912 	struct sw_zone_info *swz;
4913 	struct hw_buf_info *hwb, *hwb_list = &sc->sge.hw_buf_info[0];
4914 
4915 	/*
4916 	 * Buffer Packing: Look for PAGE_SIZE or larger zone which has a bufsize
4917 	 * large enough for the max payload and cluster metadata.  Otherwise
4918 	 * settle for the largest bufsize that leaves enough room in the cluster
4919 	 * for metadata.
4920 	 *
4921 	 * Without buffer packing: Look for the smallest zone which has a
4922 	 * bufsize large enough for the max payload.  Settle for the largest
4923 	 * bufsize available if there's nothing big enough for max payload.
4924 	 */
4925 	spare_needed = fl->flags & FL_BUF_PACKING ? CL_METADATA_SIZE : 0;
4926 	swz = &sc->sge.sw_zone_info[0];
4927 	hwidx = -1;
4928 	for (zidx = 0; zidx < SW_ZONE_SIZES; zidx++, swz++) {
4929 		if (swz->size > largest_rx_cluster) {
4930 			if (__predict_true(hwidx != -1))
4931 				break;
4932 
4933 			/*
4934 			 * This is a misconfiguration.  largest_rx_cluster is
4935 			 * preventing us from finding a refill source.  See
4936 			 * dev.t5nex.<n>.buffer_sizes to figure out why.
4937 			 */
4938 			device_printf(sc->dev, "largest_rx_cluster=%u leaves no"
4939 			    " refill source for fl %p (dma %u).  Ignored.\n",
4940 			    largest_rx_cluster, fl, maxp);
4941 		}
4942 		for (idx = swz->head_hwidx; idx != -1; idx = hwb->next) {
4943 			hwb = &hwb_list[idx];
4944 			spare = swz->size - hwb->size;
4945 			if (spare < spare_needed)
4946 				continue;
4947 
4948 			hwidx = idx;		/* best option so far */
4949 			if (hwb->size >= maxp) {
4950 
4951 				if ((fl->flags & FL_BUF_PACKING) == 0)
4952 					goto done; /* stop looking (not packing) */
4953 
4954 				if (swz->size >= safest_rx_cluster)
4955 					goto done; /* stop looking (packing) */
4956 			}
4957 			break;		/* keep looking, next zone */
4958 		}
4959 	}
4960 done:
4961 	/* A usable hwidx has been located. */
4962 	MPASS(hwidx != -1);
4963 	hwb = &hwb_list[hwidx];
4964 	zidx = hwb->zidx;
4965 	swz = &sc->sge.sw_zone_info[zidx];
4966 	region1 = 0;
4967 	region3 = swz->size - hwb->size;
4968 
4969 	/*
4970 	 * Stay within this zone and see if there is a better match when mbuf
4971 	 * inlining is allowed.  Remember that the hwidx's are sorted in
4972 	 * decreasing order of size (so in increasing order of spare area).
4973 	 */
4974 	for (idx = hwidx; idx != -1; idx = hwb->next) {
4975 		hwb = &hwb_list[idx];
4976 		spare = swz->size - hwb->size;
4977 
4978 		if (allow_mbufs_in_cluster == 0 || hwb->size < maxp)
4979 			break;
4980 
4981 		/*
4982 		 * Do not inline mbufs if doing so would violate the pad/pack
4983 		 * boundary alignment requirement.
4984 		 */
4985 		if (fl_pad && (MSIZE % sc->params.sge.pad_boundary) != 0)
4986 			continue;
4987 		if (fl->flags & FL_BUF_PACKING &&
4988 		    (MSIZE % sc->params.sge.pack_boundary) != 0)
4989 			continue;
4990 
4991 		if (spare < CL_METADATA_SIZE + MSIZE)
4992 			continue;
4993 		n = (spare - CL_METADATA_SIZE) / MSIZE;
4994 		if (n > howmany(hwb->size, maxp))
4995 			break;
4996 
4997 		hwidx = idx;
4998 		if (fl->flags & FL_BUF_PACKING) {
4999 			region1 = n * MSIZE;
5000 			region3 = spare - region1;
5001 		} else {
5002 			region1 = MSIZE;
5003 			region3 = spare - region1;
5004 			break;
5005 		}
5006 	}
5007 
5008 	KASSERT(zidx >= 0 && zidx < SW_ZONE_SIZES,
5009 	    ("%s: bad zone %d for fl %p, maxp %d", __func__, zidx, fl, maxp));
5010 	KASSERT(hwidx >= 0 && hwidx <= SGE_FLBUF_SIZES,
5011 	    ("%s: bad hwidx %d for fl %p, maxp %d", __func__, hwidx, fl, maxp));
5012 	KASSERT(region1 + sc->sge.hw_buf_info[hwidx].size + region3 ==
5013 	    sc->sge.sw_zone_info[zidx].size,
5014 	    ("%s: bad buffer layout for fl %p, maxp %d. "
5015 		"cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5016 		sc->sge.sw_zone_info[zidx].size, region1,
5017 		sc->sge.hw_buf_info[hwidx].size, region3));
5018 	if (fl->flags & FL_BUF_PACKING || region1 > 0) {
5019 		KASSERT(region3 >= CL_METADATA_SIZE,
5020 		    ("%s: no room for metadata.  fl %p, maxp %d; "
5021 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5022 		    sc->sge.sw_zone_info[zidx].size, region1,
5023 		    sc->sge.hw_buf_info[hwidx].size, region3));
5024 		KASSERT(region1 % MSIZE == 0,
5025 		    ("%s: bad mbuf region for fl %p, maxp %d. "
5026 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5027 		    sc->sge.sw_zone_info[zidx].size, region1,
5028 		    sc->sge.hw_buf_info[hwidx].size, region3));
5029 	}
5030 
5031 	fl->cll_def.zidx = zidx;
5032 	fl->cll_def.hwidx = hwidx;
5033 	fl->cll_def.region1 = region1;
5034 	fl->cll_def.region3 = region3;
5035 }
5036 
5037 static void
5038 find_safe_refill_source(struct adapter *sc, struct sge_fl *fl)
5039 {
5040 	struct sge *s = &sc->sge;
5041 	struct hw_buf_info *hwb;
5042 	struct sw_zone_info *swz;
5043 	int spare;
5044 	int8_t hwidx;
5045 
5046 	if (fl->flags & FL_BUF_PACKING)
5047 		hwidx = s->safe_hwidx2;	/* with room for metadata */
5048 	else if (allow_mbufs_in_cluster && s->safe_hwidx2 != -1) {
5049 		hwidx = s->safe_hwidx2;
5050 		hwb = &s->hw_buf_info[hwidx];
5051 		swz = &s->sw_zone_info[hwb->zidx];
5052 		spare = swz->size - hwb->size;
5053 
5054 		/* no good if there isn't room for an mbuf as well */
5055 		if (spare < CL_METADATA_SIZE + MSIZE)
5056 			hwidx = s->safe_hwidx1;
5057 	} else
5058 		hwidx = s->safe_hwidx1;
5059 
5060 	if (hwidx == -1) {
5061 		/* No fallback source */
5062 		fl->cll_alt.hwidx = -1;
5063 		fl->cll_alt.zidx = -1;
5064 
5065 		return;
5066 	}
5067 
5068 	hwb = &s->hw_buf_info[hwidx];
5069 	swz = &s->sw_zone_info[hwb->zidx];
5070 	spare = swz->size - hwb->size;
5071 	fl->cll_alt.hwidx = hwidx;
5072 	fl->cll_alt.zidx = hwb->zidx;
5073 	if (allow_mbufs_in_cluster &&
5074 	    (fl_pad == 0 || (MSIZE % sc->params.sge.pad_boundary) == 0))
5075 		fl->cll_alt.region1 = ((spare - CL_METADATA_SIZE) / MSIZE) * MSIZE;
5076 	else
5077 		fl->cll_alt.region1 = 0;
5078 	fl->cll_alt.region3 = spare - fl->cll_alt.region1;
5079 }
5080 
5081 static void
5082 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
5083 {
5084 	mtx_lock(&sc->sfl_lock);
5085 	FL_LOCK(fl);
5086 	if ((fl->flags & FL_DOOMED) == 0) {
5087 		fl->flags |= FL_STARVING;
5088 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
5089 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
5090 	}
5091 	FL_UNLOCK(fl);
5092 	mtx_unlock(&sc->sfl_lock);
5093 }
5094 
5095 static void
5096 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq)
5097 {
5098 	struct sge_wrq *wrq = (void *)eq;
5099 
5100 	atomic_readandclear_int(&eq->equiq);
5101 	taskqueue_enqueue(sc->tq[eq->tx_chan], &wrq->wrq_tx_task);
5102 }
5103 
5104 static void
5105 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq)
5106 {
5107 	struct sge_txq *txq = (void *)eq;
5108 
5109 	MPASS((eq->flags & EQ_TYPEMASK) == EQ_ETH);
5110 
5111 	atomic_readandclear_int(&eq->equiq);
5112 	mp_ring_check_drainage(txq->r, 0);
5113 	taskqueue_enqueue(sc->tq[eq->tx_chan], &txq->tx_reclaim_task);
5114 }
5115 
5116 static int
5117 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
5118     struct mbuf *m)
5119 {
5120 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
5121 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
5122 	struct adapter *sc = iq->adapter;
5123 	struct sge *s = &sc->sge;
5124 	struct sge_eq *eq;
5125 	static void (*h[])(struct adapter *, struct sge_eq *) = {NULL,
5126 		&handle_wrq_egr_update, &handle_eth_egr_update,
5127 		&handle_wrq_egr_update};
5128 
5129 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5130 	    rss->opcode));
5131 
5132 	eq = s->eqmap[qid - s->eq_start - s->eq_base];
5133 	(*h[eq->flags & EQ_TYPEMASK])(sc, eq);
5134 
5135 	return (0);
5136 }
5137 
5138 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
5139 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
5140     offsetof(struct cpl_fw6_msg, data));
5141 
5142 static int
5143 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
5144 {
5145 	struct adapter *sc = iq->adapter;
5146 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
5147 
5148 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5149 	    rss->opcode));
5150 
5151 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
5152 		const struct rss_header *rss2;
5153 
5154 		rss2 = (const struct rss_header *)&cpl->data[0];
5155 		return (t4_cpl_handler[rss2->opcode](iq, rss2, m));
5156 	}
5157 
5158 	return (t4_fw_msg_handler[cpl->type](sc, &cpl->data[0]));
5159 }
5160 
5161 /**
5162  *	t4_handle_wrerr_rpl - process a FW work request error message
5163  *	@adap: the adapter
5164  *	@rpl: start of the FW message
5165  */
5166 static int
5167 t4_handle_wrerr_rpl(struct adapter *adap, const __be64 *rpl)
5168 {
5169 	u8 opcode = *(const u8 *)rpl;
5170 	const struct fw_error_cmd *e = (const void *)rpl;
5171 	unsigned int i;
5172 
5173 	if (opcode != FW_ERROR_CMD) {
5174 		log(LOG_ERR,
5175 		    "%s: Received WRERR_RPL message with opcode %#x\n",
5176 		    device_get_nameunit(adap->dev), opcode);
5177 		return (EINVAL);
5178 	}
5179 	log(LOG_ERR, "%s: FW_ERROR (%s) ", device_get_nameunit(adap->dev),
5180 	    G_FW_ERROR_CMD_FATAL(be32toh(e->op_to_type)) ? "fatal" :
5181 	    "non-fatal");
5182 	switch (G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type))) {
5183 	case FW_ERROR_TYPE_EXCEPTION:
5184 		log(LOG_ERR, "exception info:\n");
5185 		for (i = 0; i < nitems(e->u.exception.info); i++)
5186 			log(LOG_ERR, "%s%08x", i == 0 ? "\t" : " ",
5187 			    be32toh(e->u.exception.info[i]));
5188 		log(LOG_ERR, "\n");
5189 		break;
5190 	case FW_ERROR_TYPE_HWMODULE:
5191 		log(LOG_ERR, "HW module regaddr %08x regval %08x\n",
5192 		    be32toh(e->u.hwmodule.regaddr),
5193 		    be32toh(e->u.hwmodule.regval));
5194 		break;
5195 	case FW_ERROR_TYPE_WR:
5196 		log(LOG_ERR, "WR cidx %d PF %d VF %d eqid %d hdr:\n",
5197 		    be16toh(e->u.wr.cidx),
5198 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.wr.pfn_vfn)),
5199 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.wr.pfn_vfn)),
5200 		    be32toh(e->u.wr.eqid));
5201 		for (i = 0; i < nitems(e->u.wr.wrhdr); i++)
5202 			log(LOG_ERR, "%s%02x", i == 0 ? "\t" : " ",
5203 			    e->u.wr.wrhdr[i]);
5204 		log(LOG_ERR, "\n");
5205 		break;
5206 	case FW_ERROR_TYPE_ACL:
5207 		log(LOG_ERR, "ACL cidx %d PF %d VF %d eqid %d %s",
5208 		    be16toh(e->u.acl.cidx),
5209 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.acl.pfn_vfn)),
5210 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.acl.pfn_vfn)),
5211 		    be32toh(e->u.acl.eqid),
5212 		    G_FW_ERROR_CMD_MV(be16toh(e->u.acl.mv_pkd)) ? "vlanid" :
5213 		    "MAC");
5214 		for (i = 0; i < nitems(e->u.acl.val); i++)
5215 			log(LOG_ERR, " %02x", e->u.acl.val[i]);
5216 		log(LOG_ERR, "\n");
5217 		break;
5218 	default:
5219 		log(LOG_ERR, "type %#x\n",
5220 		    G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type)));
5221 		return (EINVAL);
5222 	}
5223 	return (0);
5224 }
5225 
5226 static int
5227 sysctl_uint16(SYSCTL_HANDLER_ARGS)
5228 {
5229 	uint16_t *id = arg1;
5230 	int i = *id;
5231 
5232 	return sysctl_handle_int(oidp, &i, 0, req);
5233 }
5234 
5235 static int
5236 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
5237 {
5238 	struct sge *s = arg1;
5239 	struct hw_buf_info *hwb = &s->hw_buf_info[0];
5240 	struct sw_zone_info *swz = &s->sw_zone_info[0];
5241 	int i, rc;
5242 	struct sbuf sb;
5243 	char c;
5244 
5245 	sbuf_new(&sb, NULL, 32, SBUF_AUTOEXTEND);
5246 	for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
5247 		if (hwb->zidx >= 0 && swz[hwb->zidx].size <= largest_rx_cluster)
5248 			c = '*';
5249 		else
5250 			c = '\0';
5251 
5252 		sbuf_printf(&sb, "%u%c ", hwb->size, c);
5253 	}
5254 	sbuf_trim(&sb);
5255 	sbuf_finish(&sb);
5256 	rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
5257 	sbuf_delete(&sb);
5258 	return (rc);
5259 }
5260 
5261 static int
5262 sysctl_tc(SYSCTL_HANDLER_ARGS)
5263 {
5264 	struct vi_info *vi = arg1;
5265 	struct port_info *pi;
5266 	struct adapter *sc;
5267 	struct sge_txq *txq;
5268 	struct tx_cl_rl_params *tc;
5269 	int qidx = arg2, rc, tc_idx;
5270 	uint32_t fw_queue, fw_class;
5271 
5272 	MPASS(qidx >= 0 && qidx < vi->ntxq);
5273 	pi = vi->pi;
5274 	sc = pi->adapter;
5275 	txq = &sc->sge.txq[vi->first_txq + qidx];
5276 
5277 	tc_idx = txq->tc_idx;
5278 	rc = sysctl_handle_int(oidp, &tc_idx, 0, req);
5279 	if (rc != 0 || req->newptr == NULL)
5280 		return (rc);
5281 
5282 	if (sc->flags & IS_VF)
5283 		return (EPERM);
5284 
5285 	/* Note that -1 is legitimate input (it means unbind). */
5286 	if (tc_idx < -1 || tc_idx >= sc->chip_params->nsched_cls)
5287 		return (EINVAL);
5288 
5289 	mtx_lock(&sc->tc_lock);
5290 	if (tc_idx == txq->tc_idx) {
5291 		rc = 0;		/* No change, nothing to do. */
5292 		goto done;
5293 	}
5294 
5295 	fw_queue = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
5296 	    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_EQ_SCHEDCLASS_ETH) |
5297 	    V_FW_PARAMS_PARAM_YZ(txq->eq.cntxt_id);
5298 
5299 	if (tc_idx == -1)
5300 		fw_class = 0xffffffff;	/* Unbind. */
5301 	else {
5302 		/*
5303 		 * Bind to a different class.
5304 		 */
5305 		tc = &pi->sched_params->cl_rl[tc_idx];
5306 		if (tc->flags & TX_CLRL_ERROR) {
5307 			/* Previous attempt to set the cl-rl params failed. */
5308 			rc = EIO;
5309 			goto done;
5310 		} else {
5311 			/*
5312 			 * Ok to proceed.  Place a reference on the new class
5313 			 * while still holding on to the reference on the
5314 			 * previous class, if any.
5315 			 */
5316 			fw_class = tc_idx;
5317 			tc->refcount++;
5318 		}
5319 	}
5320 	mtx_unlock(&sc->tc_lock);
5321 
5322 	rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4stc");
5323 	if (rc)
5324 		return (rc);
5325 	rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &fw_queue, &fw_class);
5326 	end_synchronized_op(sc, 0);
5327 
5328 	mtx_lock(&sc->tc_lock);
5329 	if (rc == 0) {
5330 		if (txq->tc_idx != -1) {
5331 			tc = &pi->sched_params->cl_rl[txq->tc_idx];
5332 			MPASS(tc->refcount > 0);
5333 			tc->refcount--;
5334 		}
5335 		txq->tc_idx = tc_idx;
5336 	} else if (tc_idx != -1) {
5337 		tc = &pi->sched_params->cl_rl[tc_idx];
5338 		MPASS(tc->refcount > 0);
5339 		tc->refcount--;
5340 	}
5341 done:
5342 	mtx_unlock(&sc->tc_lock);
5343 	return (rc);
5344 }
5345