xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision ec0e626bafb335b30c499d06066997f54b10c092)
1 /*-
2  * Copyright (c) 2011 Chelsio Communications, Inc.
3  * All rights reserved.
4  * Written by: Navdeep Parhar <np@FreeBSD.org>
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 
34 #include <sys/types.h>
35 #include <sys/eventhandler.h>
36 #include <sys/mbuf.h>
37 #include <sys/socket.h>
38 #include <sys/kernel.h>
39 #include <sys/malloc.h>
40 #include <sys/queue.h>
41 #include <sys/sbuf.h>
42 #include <sys/taskqueue.h>
43 #include <sys/time.h>
44 #include <sys/sglist.h>
45 #include <sys/sysctl.h>
46 #include <sys/smp.h>
47 #include <sys/counter.h>
48 #include <net/bpf.h>
49 #include <net/ethernet.h>
50 #include <net/if.h>
51 #include <net/if_vlan_var.h>
52 #include <netinet/in.h>
53 #include <netinet/ip.h>
54 #include <netinet/ip6.h>
55 #include <netinet/tcp.h>
56 #include <machine/md_var.h>
57 #include <vm/vm.h>
58 #include <vm/pmap.h>
59 #ifdef DEV_NETMAP
60 #include <machine/bus.h>
61 #include <sys/selinfo.h>
62 #include <net/if_var.h>
63 #include <net/netmap.h>
64 #include <dev/netmap/netmap_kern.h>
65 #endif
66 
67 #include "common/common.h"
68 #include "common/t4_regs.h"
69 #include "common/t4_regs_values.h"
70 #include "common/t4_msg.h"
71 #include "t4_mp_ring.h"
72 
73 #ifdef T4_PKT_TIMESTAMP
74 #define RX_COPY_THRESHOLD (MINCLSIZE - 8)
75 #else
76 #define RX_COPY_THRESHOLD MINCLSIZE
77 #endif
78 
79 /*
80  * Ethernet frames are DMA'd at this byte offset into the freelist buffer.
81  * 0-7 are valid values.
82  */
83 int fl_pktshift = 2;
84 TUNABLE_INT("hw.cxgbe.fl_pktshift", &fl_pktshift);
85 
86 /*
87  * Pad ethernet payload up to this boundary.
88  * -1: driver should figure out a good value.
89  *  0: disable padding.
90  *  Any power of 2 from 32 to 4096 (both inclusive) is also a valid value.
91  */
92 int fl_pad = -1;
93 TUNABLE_INT("hw.cxgbe.fl_pad", &fl_pad);
94 
95 /*
96  * Status page length.
97  * -1: driver should figure out a good value.
98  *  64 or 128 are the only other valid values.
99  */
100 int spg_len = -1;
101 TUNABLE_INT("hw.cxgbe.spg_len", &spg_len);
102 
103 /*
104  * Congestion drops.
105  * -1: no congestion feedback (not recommended).
106  *  0: backpressure the channel instead of dropping packets right away.
107  *  1: no backpressure, drop packets for the congested queue immediately.
108  */
109 static int cong_drop = 0;
110 TUNABLE_INT("hw.cxgbe.cong_drop", &cong_drop);
111 
112 /*
113  * Deliver multiple frames in the same free list buffer if they fit.
114  * -1: let the driver decide whether to enable buffer packing or not.
115  *  0: disable buffer packing.
116  *  1: enable buffer packing.
117  */
118 static int buffer_packing = -1;
119 TUNABLE_INT("hw.cxgbe.buffer_packing", &buffer_packing);
120 
121 /*
122  * Start next frame in a packed buffer at this boundary.
123  * -1: driver should figure out a good value.
124  * T4: driver will ignore this and use the same value as fl_pad above.
125  * T5: 16, or a power of 2 from 64 to 4096 (both inclusive) is a valid value.
126  */
127 static int fl_pack = -1;
128 TUNABLE_INT("hw.cxgbe.fl_pack", &fl_pack);
129 
130 /*
131  * Allow the driver to create mbuf(s) in a cluster allocated for rx.
132  * 0: never; always allocate mbufs from the zone_mbuf UMA zone.
133  * 1: ok to create mbuf(s) within a cluster if there is room.
134  */
135 static int allow_mbufs_in_cluster = 1;
136 TUNABLE_INT("hw.cxgbe.allow_mbufs_in_cluster", &allow_mbufs_in_cluster);
137 
138 /*
139  * Largest rx cluster size that the driver is allowed to allocate.
140  */
141 static int largest_rx_cluster = MJUM16BYTES;
142 TUNABLE_INT("hw.cxgbe.largest_rx_cluster", &largest_rx_cluster);
143 
144 /*
145  * Size of cluster allocation that's most likely to succeed.  The driver will
146  * fall back to this size if it fails to allocate clusters larger than this.
147  */
148 static int safest_rx_cluster = PAGE_SIZE;
149 TUNABLE_INT("hw.cxgbe.safest_rx_cluster", &safest_rx_cluster);
150 
151 struct txpkts {
152 	u_int wr_type;		/* type 0 or type 1 */
153 	u_int npkt;		/* # of packets in this work request */
154 	u_int plen;		/* total payload (sum of all packets) */
155 	u_int len16;		/* # of 16B pieces used by this work request */
156 };
157 
158 /* A packet's SGL.  This + m_pkthdr has all info needed for tx */
159 struct sgl {
160 	struct sglist sg;
161 	struct sglist_seg seg[TX_SGL_SEGS];
162 };
163 
164 static int service_iq(struct sge_iq *, int);
165 static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t);
166 static int t4_eth_rx(struct sge_iq *, const struct rss_header *, struct mbuf *);
167 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int);
168 static inline void init_fl(struct adapter *, struct sge_fl *, int, int, char *);
169 static inline void init_eq(struct sge_eq *, int, int, uint8_t, uint16_t,
170     char *);
171 static int alloc_ring(struct adapter *, size_t, bus_dma_tag_t *, bus_dmamap_t *,
172     bus_addr_t *, void **);
173 static int free_ring(struct adapter *, bus_dma_tag_t, bus_dmamap_t, bus_addr_t,
174     void *);
175 static int alloc_iq_fl(struct port_info *, struct sge_iq *, struct sge_fl *,
176     int, int);
177 static int free_iq_fl(struct port_info *, struct sge_iq *, struct sge_fl *);
178 static void add_fl_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
179     struct sge_fl *);
180 static int alloc_fwq(struct adapter *);
181 static int free_fwq(struct adapter *);
182 static int alloc_mgmtq(struct adapter *);
183 static int free_mgmtq(struct adapter *);
184 static int alloc_rxq(struct port_info *, struct sge_rxq *, int, int,
185     struct sysctl_oid *);
186 static int free_rxq(struct port_info *, struct sge_rxq *);
187 #ifdef TCP_OFFLOAD
188 static int alloc_ofld_rxq(struct port_info *, struct sge_ofld_rxq *, int, int,
189     struct sysctl_oid *);
190 static int free_ofld_rxq(struct port_info *, struct sge_ofld_rxq *);
191 #endif
192 #ifdef DEV_NETMAP
193 static int alloc_nm_rxq(struct port_info *, struct sge_nm_rxq *, int, int,
194     struct sysctl_oid *);
195 static int free_nm_rxq(struct port_info *, struct sge_nm_rxq *);
196 static int alloc_nm_txq(struct port_info *, struct sge_nm_txq *, int, int,
197     struct sysctl_oid *);
198 static int free_nm_txq(struct port_info *, struct sge_nm_txq *);
199 #endif
200 static int ctrl_eq_alloc(struct adapter *, struct sge_eq *);
201 static int eth_eq_alloc(struct adapter *, struct port_info *, struct sge_eq *);
202 #ifdef TCP_OFFLOAD
203 static int ofld_eq_alloc(struct adapter *, struct port_info *, struct sge_eq *);
204 #endif
205 static int alloc_eq(struct adapter *, struct port_info *, struct sge_eq *);
206 static int free_eq(struct adapter *, struct sge_eq *);
207 static int alloc_wrq(struct adapter *, struct port_info *, struct sge_wrq *,
208     struct sysctl_oid *);
209 static int free_wrq(struct adapter *, struct sge_wrq *);
210 static int alloc_txq(struct port_info *, struct sge_txq *, int,
211     struct sysctl_oid *);
212 static int free_txq(struct port_info *, struct sge_txq *);
213 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int);
214 static inline void ring_fl_db(struct adapter *, struct sge_fl *);
215 static int refill_fl(struct adapter *, struct sge_fl *, int);
216 static void refill_sfl(void *);
217 static int alloc_fl_sdesc(struct sge_fl *);
218 static void free_fl_sdesc(struct adapter *, struct sge_fl *);
219 static void find_best_refill_source(struct adapter *, struct sge_fl *, int);
220 static void find_safe_refill_source(struct adapter *, struct sge_fl *);
221 static void add_fl_to_sfl(struct adapter *, struct sge_fl *);
222 
223 static inline void get_pkt_gl(struct mbuf *, struct sglist *);
224 static inline u_int txpkt_len16(u_int, u_int);
225 static inline u_int txpkts0_len16(u_int);
226 static inline u_int txpkts1_len16(void);
227 static u_int write_txpkt_wr(struct sge_txq *, struct fw_eth_tx_pkt_wr *,
228     struct mbuf *, u_int);
229 static int try_txpkts(struct mbuf *, struct mbuf *, struct txpkts *, u_int);
230 static int add_to_txpkts(struct mbuf *, struct txpkts *, u_int);
231 static u_int write_txpkts_wr(struct sge_txq *, struct fw_eth_tx_pkts_wr *,
232     struct mbuf *, const struct txpkts *, u_int);
233 static void write_gl_to_txd(struct sge_txq *, struct mbuf *, caddr_t *, int);
234 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int);
235 static inline void ring_eq_db(struct adapter *, struct sge_eq *, u_int);
236 static inline uint16_t read_hw_cidx(struct sge_eq *);
237 static inline u_int reclaimable_tx_desc(struct sge_eq *);
238 static inline u_int total_available_tx_desc(struct sge_eq *);
239 static u_int reclaim_tx_descs(struct sge_txq *, u_int);
240 static void tx_reclaim(void *, int);
241 static __be64 get_flit(struct sglist_seg *, int, int);
242 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
243     struct mbuf *);
244 static int handle_fw_msg(struct sge_iq *, const struct rss_header *,
245     struct mbuf *);
246 static void wrq_tx_drain(void *, int);
247 static void drain_wrq_wr_list(struct adapter *, struct sge_wrq *);
248 
249 static int sysctl_uint16(SYSCTL_HANDLER_ARGS);
250 static int sysctl_bufsizes(SYSCTL_HANDLER_ARGS);
251 
252 static counter_u64_t extfree_refs;
253 static counter_u64_t extfree_rels;
254 
255 /*
256  * Called on MOD_LOAD.  Validates and calculates the SGE tunables.
257  */
258 void
259 t4_sge_modload(void)
260 {
261 
262 	if (fl_pktshift < 0 || fl_pktshift > 7) {
263 		printf("Invalid hw.cxgbe.fl_pktshift value (%d),"
264 		    " using 2 instead.\n", fl_pktshift);
265 		fl_pktshift = 2;
266 	}
267 
268 	if (spg_len != 64 && spg_len != 128) {
269 		int len;
270 
271 #if defined(__i386__) || defined(__amd64__)
272 		len = cpu_clflush_line_size > 64 ? 128 : 64;
273 #else
274 		len = 64;
275 #endif
276 		if (spg_len != -1) {
277 			printf("Invalid hw.cxgbe.spg_len value (%d),"
278 			    " using %d instead.\n", spg_len, len);
279 		}
280 		spg_len = len;
281 	}
282 
283 	if (cong_drop < -1 || cong_drop > 1) {
284 		printf("Invalid hw.cxgbe.cong_drop value (%d),"
285 		    " using 0 instead.\n", cong_drop);
286 		cong_drop = 0;
287 	}
288 
289 	extfree_refs = counter_u64_alloc(M_WAITOK);
290 	extfree_rels = counter_u64_alloc(M_WAITOK);
291 	counter_u64_zero(extfree_refs);
292 	counter_u64_zero(extfree_rels);
293 }
294 
295 void
296 t4_sge_modunload(void)
297 {
298 
299 	counter_u64_free(extfree_refs);
300 	counter_u64_free(extfree_rels);
301 }
302 
303 uint64_t
304 t4_sge_extfree_refs(void)
305 {
306 	uint64_t refs, rels;
307 
308 	rels = counter_u64_fetch(extfree_rels);
309 	refs = counter_u64_fetch(extfree_refs);
310 
311 	return (refs - rels);
312 }
313 
314 void
315 t4_init_sge_cpl_handlers(struct adapter *sc)
316 {
317 
318 	t4_register_cpl_handler(sc, CPL_FW4_MSG, handle_fw_msg);
319 	t4_register_cpl_handler(sc, CPL_FW6_MSG, handle_fw_msg);
320 	t4_register_cpl_handler(sc, CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
321 	t4_register_cpl_handler(sc, CPL_RX_PKT, t4_eth_rx);
322 	t4_register_fw_msg_handler(sc, FW6_TYPE_CMD_RPL, t4_handle_fw_rpl);
323 }
324 
325 static inline void
326 setup_pad_and_pack_boundaries(struct adapter *sc)
327 {
328 	uint32_t v, m;
329 	int pad, pack;
330 
331 	pad = fl_pad;
332 	if (fl_pad < 32 || fl_pad > 4096 || !powerof2(fl_pad)) {
333 		/*
334 		 * If there is any chance that we might use buffer packing and
335 		 * the chip is a T4, then pick 64 as the pad/pack boundary.  Set
336 		 * it to 32 in all other cases.
337 		 */
338 		pad = is_t4(sc) && buffer_packing ? 64 : 32;
339 
340 		/*
341 		 * For fl_pad = 0 we'll still write a reasonable value to the
342 		 * register but all the freelists will opt out of padding.
343 		 * We'll complain here only if the user tried to set it to a
344 		 * value greater than 0 that was invalid.
345 		 */
346 		if (fl_pad > 0) {
347 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pad value"
348 			    " (%d), using %d instead.\n", fl_pad, pad);
349 		}
350 	}
351 	m = V_INGPADBOUNDARY(M_INGPADBOUNDARY);
352 	v = V_INGPADBOUNDARY(ilog2(pad) - 5);
353 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
354 
355 	if (is_t4(sc)) {
356 		if (fl_pack != -1 && fl_pack != pad) {
357 			/* Complain but carry on. */
358 			device_printf(sc->dev, "hw.cxgbe.fl_pack (%d) ignored,"
359 			    " using %d instead.\n", fl_pack, pad);
360 		}
361 		return;
362 	}
363 
364 	pack = fl_pack;
365 	if (fl_pack < 16 || fl_pack == 32 || fl_pack > 4096 ||
366 	    !powerof2(fl_pack)) {
367 		pack = max(sc->params.pci.mps, CACHE_LINE_SIZE);
368 		MPASS(powerof2(pack));
369 		if (pack < 16)
370 			pack = 16;
371 		if (pack == 32)
372 			pack = 64;
373 		if (pack > 4096)
374 			pack = 4096;
375 		if (fl_pack != -1) {
376 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pack value"
377 			    " (%d), using %d instead.\n", fl_pack, pack);
378 		}
379 	}
380 	m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY);
381 	if (pack == 16)
382 		v = V_INGPACKBOUNDARY(0);
383 	else
384 		v = V_INGPACKBOUNDARY(ilog2(pack) - 5);
385 
386 	MPASS(!is_t4(sc));	/* T4 doesn't have SGE_CONTROL2 */
387 	t4_set_reg_field(sc, A_SGE_CONTROL2, m, v);
388 }
389 
390 /*
391  * adap->params.vpd.cclk must be set up before this is called.
392  */
393 void
394 t4_tweak_chip_settings(struct adapter *sc)
395 {
396 	int i;
397 	uint32_t v, m;
398 	int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200};
399 	int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk;
400 	int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */
401 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
402 	static int sge_flbuf_sizes[] = {
403 		MCLBYTES,
404 #if MJUMPAGESIZE != MCLBYTES
405 		MJUMPAGESIZE,
406 		MJUMPAGESIZE - CL_METADATA_SIZE,
407 		MJUMPAGESIZE - 2 * MSIZE - CL_METADATA_SIZE,
408 #endif
409 		MJUM9BYTES,
410 		MJUM16BYTES,
411 		MCLBYTES - MSIZE - CL_METADATA_SIZE,
412 		MJUM9BYTES - CL_METADATA_SIZE,
413 		MJUM16BYTES - CL_METADATA_SIZE,
414 	};
415 
416 	KASSERT(sc->flags & MASTER_PF,
417 	    ("%s: trying to change chip settings when not master.", __func__));
418 
419 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
420 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
421 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
422 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
423 
424 	setup_pad_and_pack_boundaries(sc);
425 
426 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
427 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
428 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
429 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
430 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
431 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
432 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
433 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
434 	t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v);
435 
436 	KASSERT(nitems(sge_flbuf_sizes) <= SGE_FLBUF_SIZES,
437 	    ("%s: hw buffer size table too big", __func__));
438 	for (i = 0; i < min(nitems(sge_flbuf_sizes), SGE_FLBUF_SIZES); i++) {
439 		t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i),
440 		    sge_flbuf_sizes[i]);
441 	}
442 
443 	v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) |
444 	    V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]);
445 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v);
446 
447 	KASSERT(intr_timer[0] <= timer_max,
448 	    ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0],
449 	    timer_max));
450 	for (i = 1; i < nitems(intr_timer); i++) {
451 		KASSERT(intr_timer[i] >= intr_timer[i - 1],
452 		    ("%s: timers not listed in increasing order (%d)",
453 		    __func__, i));
454 
455 		while (intr_timer[i] > timer_max) {
456 			if (i == nitems(intr_timer) - 1) {
457 				intr_timer[i] = timer_max;
458 				break;
459 			}
460 			intr_timer[i] += intr_timer[i - 1];
461 			intr_timer[i] /= 2;
462 		}
463 	}
464 
465 	v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) |
466 	    V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1]));
467 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v);
468 	v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) |
469 	    V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3]));
470 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v);
471 	v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) |
472 	    V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5]));
473 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v);
474 
475 	if (cong_drop == 0) {
476 		m = F_TUNNELCNGDROP0 | F_TUNNELCNGDROP1 | F_TUNNELCNGDROP2 |
477 		    F_TUNNELCNGDROP3;
478 		t4_set_reg_field(sc, A_TP_PARA_REG3, m, 0);
479 	}
480 
481 	/* 4K, 16K, 64K, 256K DDP "page sizes" */
482 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
483 	t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v);
484 
485 	m = v = F_TDDPTAGTCB;
486 	t4_set_reg_field(sc, A_ULP_RX_CTL, m, v);
487 
488 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
489 	    F_RESETDDPOFFSET;
490 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
491 	t4_set_reg_field(sc, A_TP_PARA_REG5, m, v);
492 }
493 
494 /*
495  * SGE wants the buffer to be at least 64B and then a multiple of 16.  If
496  * padding is is use the buffer's start and end need to be aligned to the pad
497  * boundary as well.  We'll just make sure that the size is a multiple of the
498  * boundary here, it is up to the buffer allocation code to make sure the start
499  * of the buffer is aligned as well.
500  */
501 static inline int
502 hwsz_ok(struct adapter *sc, int hwsz)
503 {
504 	int mask = fl_pad ? sc->sge.pad_boundary - 1 : 16 - 1;
505 
506 	return (hwsz >= 64 && (hwsz & mask) == 0);
507 }
508 
509 /*
510  * XXX: driver really should be able to deal with unexpected settings.
511  */
512 int
513 t4_read_chip_settings(struct adapter *sc)
514 {
515 	struct sge *s = &sc->sge;
516 	int i, j, n, rc = 0;
517 	uint32_t m, v, r;
518 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
519 	static int sw_buf_sizes[] = {	/* Sorted by size */
520 		MCLBYTES,
521 #if MJUMPAGESIZE != MCLBYTES
522 		MJUMPAGESIZE,
523 #endif
524 		MJUM9BYTES,
525 		MJUM16BYTES
526 	};
527 	struct sw_zone_info *swz, *safe_swz;
528 	struct hw_buf_info *hwb;
529 
530 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
531 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
532 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
533 	r = t4_read_reg(sc, A_SGE_CONTROL);
534 	if ((r & m) != v) {
535 		device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r);
536 		rc = EINVAL;
537 	}
538 	s->pad_boundary = 1 << (G_INGPADBOUNDARY(r) + 5);
539 
540 	if (is_t4(sc))
541 		s->pack_boundary = s->pad_boundary;
542 	else {
543 		r = t4_read_reg(sc, A_SGE_CONTROL2);
544 		if (G_INGPACKBOUNDARY(r) == 0)
545 			s->pack_boundary = 16;
546 		else
547 			s->pack_boundary = 1 << (G_INGPACKBOUNDARY(r) + 5);
548 	}
549 
550 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
551 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
552 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
553 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
554 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
555 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
556 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
557 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
558 	r = t4_read_reg(sc, A_SGE_HOST_PAGE_SIZE);
559 	if (r != v) {
560 		device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r);
561 		rc = EINVAL;
562 	}
563 
564 	/* Filter out unusable hw buffer sizes entirely (mark with -2). */
565 	hwb = &s->hw_buf_info[0];
566 	for (i = 0; i < nitems(s->hw_buf_info); i++, hwb++) {
567 		r = t4_read_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i));
568 		hwb->size = r;
569 		hwb->zidx = hwsz_ok(sc, r) ? -1 : -2;
570 		hwb->next = -1;
571 	}
572 
573 	/*
574 	 * Create a sorted list in decreasing order of hw buffer sizes (and so
575 	 * increasing order of spare area) for each software zone.
576 	 *
577 	 * If padding is enabled then the start and end of the buffer must align
578 	 * to the pad boundary; if packing is enabled then they must align with
579 	 * the pack boundary as well.  Allocations from the cluster zones are
580 	 * aligned to min(size, 4K), so the buffer starts at that alignment and
581 	 * ends at hwb->size alignment.  If mbuf inlining is allowed the
582 	 * starting alignment will be reduced to MSIZE and the driver will
583 	 * exercise appropriate caution when deciding on the best buffer layout
584 	 * to use.
585 	 */
586 	n = 0;	/* no usable buffer size to begin with */
587 	swz = &s->sw_zone_info[0];
588 	safe_swz = NULL;
589 	for (i = 0; i < SW_ZONE_SIZES; i++, swz++) {
590 		int8_t head = -1, tail = -1;
591 
592 		swz->size = sw_buf_sizes[i];
593 		swz->zone = m_getzone(swz->size);
594 		swz->type = m_gettype(swz->size);
595 
596 		if (swz->size < PAGE_SIZE) {
597 			MPASS(powerof2(swz->size));
598 			if (fl_pad && (swz->size % sc->sge.pad_boundary != 0))
599 				continue;
600 		}
601 
602 		if (swz->size == safest_rx_cluster)
603 			safe_swz = swz;
604 
605 		hwb = &s->hw_buf_info[0];
606 		for (j = 0; j < SGE_FLBUF_SIZES; j++, hwb++) {
607 			if (hwb->zidx != -1 || hwb->size > swz->size)
608 				continue;
609 #ifdef INVARIANTS
610 			if (fl_pad)
611 				MPASS(hwb->size % sc->sge.pad_boundary == 0);
612 #endif
613 			hwb->zidx = i;
614 			if (head == -1)
615 				head = tail = j;
616 			else if (hwb->size < s->hw_buf_info[tail].size) {
617 				s->hw_buf_info[tail].next = j;
618 				tail = j;
619 			} else {
620 				int8_t *cur;
621 				struct hw_buf_info *t;
622 
623 				for (cur = &head; *cur != -1; cur = &t->next) {
624 					t = &s->hw_buf_info[*cur];
625 					if (hwb->size == t->size) {
626 						hwb->zidx = -2;
627 						break;
628 					}
629 					if (hwb->size > t->size) {
630 						hwb->next = *cur;
631 						*cur = j;
632 						break;
633 					}
634 				}
635 			}
636 		}
637 		swz->head_hwidx = head;
638 		swz->tail_hwidx = tail;
639 
640 		if (tail != -1) {
641 			n++;
642 			if (swz->size - s->hw_buf_info[tail].size >=
643 			    CL_METADATA_SIZE)
644 				sc->flags |= BUF_PACKING_OK;
645 		}
646 	}
647 	if (n == 0) {
648 		device_printf(sc->dev, "no usable SGE FL buffer size.\n");
649 		rc = EINVAL;
650 	}
651 
652 	s->safe_hwidx1 = -1;
653 	s->safe_hwidx2 = -1;
654 	if (safe_swz != NULL) {
655 		s->safe_hwidx1 = safe_swz->head_hwidx;
656 		for (i = safe_swz->head_hwidx; i != -1; i = hwb->next) {
657 			int spare;
658 
659 			hwb = &s->hw_buf_info[i];
660 #ifdef INVARIANTS
661 			if (fl_pad)
662 				MPASS(hwb->size % sc->sge.pad_boundary == 0);
663 #endif
664 			spare = safe_swz->size - hwb->size;
665 			if (spare >= CL_METADATA_SIZE) {
666 				s->safe_hwidx2 = i;
667 				break;
668 			}
669 		}
670 	}
671 
672 	r = t4_read_reg(sc, A_SGE_INGRESS_RX_THRESHOLD);
673 	s->counter_val[0] = G_THRESHOLD_0(r);
674 	s->counter_val[1] = G_THRESHOLD_1(r);
675 	s->counter_val[2] = G_THRESHOLD_2(r);
676 	s->counter_val[3] = G_THRESHOLD_3(r);
677 
678 	r = t4_read_reg(sc, A_SGE_TIMER_VALUE_0_AND_1);
679 	s->timer_val[0] = G_TIMERVALUE0(r) / core_ticks_per_usec(sc);
680 	s->timer_val[1] = G_TIMERVALUE1(r) / core_ticks_per_usec(sc);
681 	r = t4_read_reg(sc, A_SGE_TIMER_VALUE_2_AND_3);
682 	s->timer_val[2] = G_TIMERVALUE2(r) / core_ticks_per_usec(sc);
683 	s->timer_val[3] = G_TIMERVALUE3(r) / core_ticks_per_usec(sc);
684 	r = t4_read_reg(sc, A_SGE_TIMER_VALUE_4_AND_5);
685 	s->timer_val[4] = G_TIMERVALUE4(r) / core_ticks_per_usec(sc);
686 	s->timer_val[5] = G_TIMERVALUE5(r) / core_ticks_per_usec(sc);
687 
688 	if (cong_drop == 0) {
689 		m = F_TUNNELCNGDROP0 | F_TUNNELCNGDROP1 | F_TUNNELCNGDROP2 |
690 		    F_TUNNELCNGDROP3;
691 		r = t4_read_reg(sc, A_TP_PARA_REG3);
692 		if (r & m) {
693 			device_printf(sc->dev,
694 			    "invalid TP_PARA_REG3(0x%x)\n", r);
695 			rc = EINVAL;
696 		}
697 	}
698 
699 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
700 	r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ);
701 	if (r != v) {
702 		device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r);
703 		rc = EINVAL;
704 	}
705 
706 	m = v = F_TDDPTAGTCB;
707 	r = t4_read_reg(sc, A_ULP_RX_CTL);
708 	if ((r & m) != v) {
709 		device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r);
710 		rc = EINVAL;
711 	}
712 
713 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
714 	    F_RESETDDPOFFSET;
715 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
716 	r = t4_read_reg(sc, A_TP_PARA_REG5);
717 	if ((r & m) != v) {
718 		device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r);
719 		rc = EINVAL;
720 	}
721 
722 	r = t4_read_reg(sc, A_SGE_CONM_CTRL);
723 	s->fl_starve_threshold = G_EGRTHRESHOLD(r) * 2 + 1;
724 	if (is_t4(sc))
725 		s->fl_starve_threshold2 = s->fl_starve_threshold;
726 	else
727 		s->fl_starve_threshold2 = G_EGRTHRESHOLDPACKING(r) * 2 + 1;
728 
729 	/* egress queues: log2 of # of doorbells per BAR2 page */
730 	r = t4_read_reg(sc, A_SGE_EGRESS_QUEUES_PER_PAGE_PF);
731 	r >>= S_QUEUESPERPAGEPF0 +
732 	    (S_QUEUESPERPAGEPF1 - S_QUEUESPERPAGEPF0) * sc->pf;
733 	s->eq_s_qpp = r & M_QUEUESPERPAGEPF0;
734 
735 	/* ingress queues: log2 of # of doorbells per BAR2 page */
736 	r = t4_read_reg(sc, A_SGE_INGRESS_QUEUES_PER_PAGE_PF);
737 	r >>= S_QUEUESPERPAGEPF0 +
738 	    (S_QUEUESPERPAGEPF1 - S_QUEUESPERPAGEPF0) * sc->pf;
739 	s->iq_s_qpp = r & M_QUEUESPERPAGEPF0;
740 
741 	t4_init_tp_params(sc);
742 
743 	t4_read_mtu_tbl(sc, sc->params.mtus, NULL);
744 	t4_load_mtus(sc, sc->params.mtus, sc->params.a_wnd, sc->params.b_wnd);
745 
746 	return (rc);
747 }
748 
749 int
750 t4_create_dma_tag(struct adapter *sc)
751 {
752 	int rc;
753 
754 	rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
755 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
756 	    BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
757 	    NULL, &sc->dmat);
758 	if (rc != 0) {
759 		device_printf(sc->dev,
760 		    "failed to create main DMA tag: %d\n", rc);
761 	}
762 
763 	return (rc);
764 }
765 
766 void
767 t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
768     struct sysctl_oid_list *children)
769 {
770 
771 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "buffer_sizes",
772 	    CTLTYPE_STRING | CTLFLAG_RD, &sc->sge, 0, sysctl_bufsizes, "A",
773 	    "freelist buffer sizes");
774 
775 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD,
776 	    NULL, fl_pktshift, "payload DMA offset in rx buffer (bytes)");
777 
778 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD,
779 	    NULL, sc->sge.pad_boundary, "payload pad boundary (bytes)");
780 
781 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD,
782 	    NULL, spg_len, "status page size (bytes)");
783 
784 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD,
785 	    NULL, cong_drop, "congestion drop setting");
786 
787 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD,
788 	    NULL, sc->sge.pack_boundary, "payload pack boundary (bytes)");
789 }
790 
791 int
792 t4_destroy_dma_tag(struct adapter *sc)
793 {
794 	if (sc->dmat)
795 		bus_dma_tag_destroy(sc->dmat);
796 
797 	return (0);
798 }
799 
800 /*
801  * Allocate and initialize the firmware event queue and the management queue.
802  *
803  * Returns errno on failure.  Resources allocated up to that point may still be
804  * allocated.  Caller is responsible for cleanup in case this function fails.
805  */
806 int
807 t4_setup_adapter_queues(struct adapter *sc)
808 {
809 	int rc;
810 
811 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
812 
813 	sysctl_ctx_init(&sc->ctx);
814 	sc->flags |= ADAP_SYSCTL_CTX;
815 
816 	/*
817 	 * Firmware event queue
818 	 */
819 	rc = alloc_fwq(sc);
820 	if (rc != 0)
821 		return (rc);
822 
823 	/*
824 	 * Management queue.  This is just a control queue that uses the fwq as
825 	 * its associated iq.
826 	 */
827 	rc = alloc_mgmtq(sc);
828 
829 	return (rc);
830 }
831 
832 /*
833  * Idempotent
834  */
835 int
836 t4_teardown_adapter_queues(struct adapter *sc)
837 {
838 
839 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
840 
841 	/* Do this before freeing the queue */
842 	if (sc->flags & ADAP_SYSCTL_CTX) {
843 		sysctl_ctx_free(&sc->ctx);
844 		sc->flags &= ~ADAP_SYSCTL_CTX;
845 	}
846 
847 	free_mgmtq(sc);
848 	free_fwq(sc);
849 
850 	return (0);
851 }
852 
853 static inline int
854 port_intr_count(struct port_info *pi)
855 {
856 	int rc = 0;
857 
858 	if (pi->flags & INTR_RXQ)
859 		rc += pi->nrxq;
860 #ifdef TCP_OFFLOAD
861 	if (pi->flags & INTR_OFLD_RXQ)
862 		rc += pi->nofldrxq;
863 #endif
864 #ifdef DEV_NETMAP
865 	if (pi->flags & INTR_NM_RXQ)
866 		rc += pi->nnmrxq;
867 #endif
868 	return (rc);
869 }
870 
871 static inline int
872 first_vector(struct port_info *pi)
873 {
874 	struct adapter *sc = pi->adapter;
875 	int rc = T4_EXTRA_INTR, i;
876 
877 	if (sc->intr_count == 1)
878 		return (0);
879 
880 	for_each_port(sc, i) {
881 		if (i == pi->port_id)
882 			break;
883 
884 		rc += port_intr_count(sc->port[i]);
885 	}
886 
887 	return (rc);
888 }
889 
890 /*
891  * Given an arbitrary "index," come up with an iq that can be used by other
892  * queues (of this port) for interrupt forwarding, SGE egress updates, etc.
893  * The iq returned is guaranteed to be something that takes direct interrupts.
894  */
895 static struct sge_iq *
896 port_intr_iq(struct port_info *pi, int idx)
897 {
898 	struct adapter *sc = pi->adapter;
899 	struct sge *s = &sc->sge;
900 	struct sge_iq *iq = NULL;
901 	int nintr, i;
902 
903 	if (sc->intr_count == 1)
904 		return (&sc->sge.fwq);
905 
906 	nintr = port_intr_count(pi);
907 	KASSERT(nintr != 0,
908 	    ("%s: pi %p has no exclusive interrupts, total interrupts = %d",
909 	    __func__, pi, sc->intr_count));
910 #ifdef DEV_NETMAP
911 	/* Exclude netmap queues as they can't take anyone else's interrupts */
912 	if (pi->flags & INTR_NM_RXQ)
913 		nintr -= pi->nnmrxq;
914 	KASSERT(nintr > 0,
915 	    ("%s: pi %p has nintr %d after netmap adjustment of %d", __func__,
916 	    pi, nintr, pi->nnmrxq));
917 #endif
918 	i = idx % nintr;
919 
920 	if (pi->flags & INTR_RXQ) {
921 	       	if (i < pi->nrxq) {
922 			iq = &s->rxq[pi->first_rxq + i].iq;
923 			goto done;
924 		}
925 		i -= pi->nrxq;
926 	}
927 #ifdef TCP_OFFLOAD
928 	if (pi->flags & INTR_OFLD_RXQ) {
929 	       	if (i < pi->nofldrxq) {
930 			iq = &s->ofld_rxq[pi->first_ofld_rxq + i].iq;
931 			goto done;
932 		}
933 		i -= pi->nofldrxq;
934 	}
935 #endif
936 	panic("%s: pi %p, intr_flags 0x%lx, idx %d, total intr %d\n", __func__,
937 	    pi, pi->flags & INTR_ALL, idx, nintr);
938 done:
939 	MPASS(iq != NULL);
940 	KASSERT(iq->flags & IQ_INTR,
941 	    ("%s: iq %p (port %p, intr_flags 0x%lx, idx %d)", __func__, iq, pi,
942 	    pi->flags & INTR_ALL, idx));
943 	return (iq);
944 }
945 
946 /* Maximum payload that can be delivered with a single iq descriptor */
947 static inline int
948 mtu_to_max_payload(struct adapter *sc, int mtu, const int toe)
949 {
950 	int payload;
951 
952 #ifdef TCP_OFFLOAD
953 	if (toe) {
954 		payload = sc->tt.rx_coalesce ?
955 		    G_RXCOALESCESIZE(t4_read_reg(sc, A_TP_PARA_REG2)) : mtu;
956 	} else {
957 #endif
958 		/* large enough even when hw VLAN extraction is disabled */
959 		payload = fl_pktshift + ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN +
960 		    mtu;
961 #ifdef TCP_OFFLOAD
962 	}
963 #endif
964 
965 	return (payload);
966 }
967 
968 int
969 t4_setup_port_queues(struct port_info *pi)
970 {
971 	int rc = 0, i, j, intr_idx, iqid;
972 	struct sge_rxq *rxq;
973 	struct sge_txq *txq;
974 	struct sge_wrq *ctrlq;
975 #ifdef TCP_OFFLOAD
976 	struct sge_ofld_rxq *ofld_rxq;
977 	struct sge_wrq *ofld_txq;
978 #endif
979 #ifdef DEV_NETMAP
980 	struct sge_nm_rxq *nm_rxq;
981 	struct sge_nm_txq *nm_txq;
982 #endif
983 	char name[16];
984 	struct adapter *sc = pi->adapter;
985 	struct ifnet *ifp = pi->ifp;
986 	struct sysctl_oid *oid = device_get_sysctl_tree(pi->dev);
987 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
988 	int maxp, mtu = ifp->if_mtu;
989 
990 	/* Interrupt vector to start from (when using multiple vectors) */
991 	intr_idx = first_vector(pi);
992 
993 	/*
994 	 * First pass over all NIC and TOE rx queues:
995 	 * a) initialize iq and fl
996 	 * b) allocate queue iff it will take direct interrupts.
997 	 */
998 	maxp = mtu_to_max_payload(sc, mtu, 0);
999 	if (pi->flags & INTR_RXQ) {
1000 		oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "rxq",
1001 		    CTLFLAG_RD, NULL, "rx queues");
1002 	}
1003 	for_each_rxq(pi, i, rxq) {
1004 
1005 		init_iq(&rxq->iq, sc, pi->tmr_idx, pi->pktc_idx, pi->qsize_rxq);
1006 
1007 		snprintf(name, sizeof(name), "%s rxq%d-fl",
1008 		    device_get_nameunit(pi->dev), i);
1009 		init_fl(sc, &rxq->fl, pi->qsize_rxq / 8, maxp, name);
1010 
1011 		if (pi->flags & INTR_RXQ) {
1012 			rxq->iq.flags |= IQ_INTR;
1013 			rc = alloc_rxq(pi, rxq, intr_idx, i, oid);
1014 			if (rc != 0)
1015 				goto done;
1016 			intr_idx++;
1017 		}
1018 	}
1019 #ifdef TCP_OFFLOAD
1020 	maxp = mtu_to_max_payload(sc, mtu, 1);
1021 	if (is_offload(sc) && pi->flags & INTR_OFLD_RXQ) {
1022 		oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "ofld_rxq",
1023 		    CTLFLAG_RD, NULL,
1024 		    "rx queues for offloaded TCP connections");
1025 	}
1026 	for_each_ofld_rxq(pi, i, ofld_rxq) {
1027 
1028 		init_iq(&ofld_rxq->iq, sc, pi->tmr_idx, pi->pktc_idx,
1029 		    pi->qsize_rxq);
1030 
1031 		snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
1032 		    device_get_nameunit(pi->dev), i);
1033 		init_fl(sc, &ofld_rxq->fl, pi->qsize_rxq / 8, maxp, name);
1034 
1035 		if (pi->flags & INTR_OFLD_RXQ) {
1036 			ofld_rxq->iq.flags |= IQ_INTR;
1037 			rc = alloc_ofld_rxq(pi, ofld_rxq, intr_idx, i, oid);
1038 			if (rc != 0)
1039 				goto done;
1040 			intr_idx++;
1041 		}
1042 	}
1043 #endif
1044 #ifdef DEV_NETMAP
1045 	/*
1046 	 * We don't have buffers to back the netmap rx queues right now so we
1047 	 * create the queues in a way that doesn't set off any congestion signal
1048 	 * in the chip.
1049 	 */
1050 	if (pi->flags & INTR_NM_RXQ) {
1051 		oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "nm_rxq",
1052 		    CTLFLAG_RD, NULL, "rx queues for netmap");
1053 		for_each_nm_rxq(pi, i, nm_rxq) {
1054 			rc = alloc_nm_rxq(pi, nm_rxq, intr_idx, i, oid);
1055 			if (rc != 0)
1056 				goto done;
1057 			intr_idx++;
1058 		}
1059 	}
1060 #endif
1061 
1062 	/*
1063 	 * Second pass over all NIC and TOE rx queues.  The queues forwarding
1064 	 * their interrupts are allocated now.
1065 	 */
1066 	j = 0;
1067 	if (!(pi->flags & INTR_RXQ)) {
1068 		oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "rxq",
1069 		    CTLFLAG_RD, NULL, "rx queues");
1070 		for_each_rxq(pi, i, rxq) {
1071 			MPASS(!(rxq->iq.flags & IQ_INTR));
1072 
1073 			intr_idx = port_intr_iq(pi, j)->abs_id;
1074 
1075 			rc = alloc_rxq(pi, rxq, intr_idx, i, oid);
1076 			if (rc != 0)
1077 				goto done;
1078 			j++;
1079 		}
1080 	}
1081 #ifdef TCP_OFFLOAD
1082 	if (is_offload(sc) && !(pi->flags & INTR_OFLD_RXQ)) {
1083 		oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "ofld_rxq",
1084 		    CTLFLAG_RD, NULL,
1085 		    "rx queues for offloaded TCP connections");
1086 		for_each_ofld_rxq(pi, i, ofld_rxq) {
1087 			MPASS(!(ofld_rxq->iq.flags & IQ_INTR));
1088 
1089 			intr_idx = port_intr_iq(pi, j)->abs_id;
1090 
1091 			rc = alloc_ofld_rxq(pi, ofld_rxq, intr_idx, i, oid);
1092 			if (rc != 0)
1093 				goto done;
1094 			j++;
1095 		}
1096 	}
1097 #endif
1098 #ifdef DEV_NETMAP
1099 	if (!(pi->flags & INTR_NM_RXQ))
1100 		CXGBE_UNIMPLEMENTED(__func__);
1101 #endif
1102 
1103 	/*
1104 	 * Now the tx queues.  Only one pass needed.
1105 	 */
1106 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "txq", CTLFLAG_RD,
1107 	    NULL, "tx queues");
1108 	j = 0;
1109 	for_each_txq(pi, i, txq) {
1110 		iqid = port_intr_iq(pi, j)->cntxt_id;
1111 		snprintf(name, sizeof(name), "%s txq%d",
1112 		    device_get_nameunit(pi->dev), i);
1113 		init_eq(&txq->eq, EQ_ETH, pi->qsize_txq, pi->tx_chan, iqid,
1114 		    name);
1115 
1116 		rc = alloc_txq(pi, txq, i, oid);
1117 		if (rc != 0)
1118 			goto done;
1119 		j++;
1120 	}
1121 #ifdef TCP_OFFLOAD
1122 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "ofld_txq",
1123 	    CTLFLAG_RD, NULL, "tx queues for offloaded TCP connections");
1124 	for_each_ofld_txq(pi, i, ofld_txq) {
1125 		struct sysctl_oid *oid2;
1126 
1127 		iqid = port_intr_iq(pi, j)->cntxt_id;
1128 		snprintf(name, sizeof(name), "%s ofld_txq%d",
1129 		    device_get_nameunit(pi->dev), i);
1130 		init_eq(&ofld_txq->eq, EQ_OFLD, pi->qsize_txq, pi->tx_chan,
1131 		    iqid, name);
1132 
1133 		snprintf(name, sizeof(name), "%d", i);
1134 		oid2 = SYSCTL_ADD_NODE(&pi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
1135 		    name, CTLFLAG_RD, NULL, "offload tx queue");
1136 
1137 		rc = alloc_wrq(sc, pi, ofld_txq, oid2);
1138 		if (rc != 0)
1139 			goto done;
1140 		j++;
1141 	}
1142 #endif
1143 #ifdef DEV_NETMAP
1144 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "nm_txq",
1145 	    CTLFLAG_RD, NULL, "tx queues for netmap use");
1146 	for_each_nm_txq(pi, i, nm_txq) {
1147 		iqid = pi->first_nm_rxq + (j % pi->nnmrxq);
1148 		rc = alloc_nm_txq(pi, nm_txq, iqid, i, oid);
1149 		if (rc != 0)
1150 			goto done;
1151 		j++;
1152 	}
1153 #endif
1154 
1155 	/*
1156 	 * Finally, the control queue.
1157 	 */
1158 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "ctrlq", CTLFLAG_RD,
1159 	    NULL, "ctrl queue");
1160 	ctrlq = &sc->sge.ctrlq[pi->port_id];
1161 	iqid = port_intr_iq(pi, 0)->cntxt_id;
1162 	snprintf(name, sizeof(name), "%s ctrlq", device_get_nameunit(pi->dev));
1163 	init_eq(&ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, pi->tx_chan, iqid, name);
1164 	rc = alloc_wrq(sc, pi, ctrlq, oid);
1165 
1166 done:
1167 	if (rc)
1168 		t4_teardown_port_queues(pi);
1169 
1170 	return (rc);
1171 }
1172 
1173 /*
1174  * Idempotent
1175  */
1176 int
1177 t4_teardown_port_queues(struct port_info *pi)
1178 {
1179 	int i;
1180 	struct adapter *sc = pi->adapter;
1181 	struct sge_rxq *rxq;
1182 	struct sge_txq *txq;
1183 #ifdef TCP_OFFLOAD
1184 	struct sge_ofld_rxq *ofld_rxq;
1185 	struct sge_wrq *ofld_txq;
1186 #endif
1187 #ifdef DEV_NETMAP
1188 	struct sge_nm_rxq *nm_rxq;
1189 	struct sge_nm_txq *nm_txq;
1190 #endif
1191 
1192 	/* Do this before freeing the queues */
1193 	if (pi->flags & PORT_SYSCTL_CTX) {
1194 		sysctl_ctx_free(&pi->ctx);
1195 		pi->flags &= ~PORT_SYSCTL_CTX;
1196 	}
1197 
1198 	/*
1199 	 * Take down all the tx queues first, as they reference the rx queues
1200 	 * (for egress updates, etc.).
1201 	 */
1202 
1203 	free_wrq(sc, &sc->sge.ctrlq[pi->port_id]);
1204 
1205 	for_each_txq(pi, i, txq) {
1206 		free_txq(pi, txq);
1207 	}
1208 #ifdef TCP_OFFLOAD
1209 	for_each_ofld_txq(pi, i, ofld_txq) {
1210 		free_wrq(sc, ofld_txq);
1211 	}
1212 #endif
1213 #ifdef DEV_NETMAP
1214 	for_each_nm_txq(pi, i, nm_txq)
1215 	    free_nm_txq(pi, nm_txq);
1216 #endif
1217 
1218 	/*
1219 	 * Then take down the rx queues that forward their interrupts, as they
1220 	 * reference other rx queues.
1221 	 */
1222 
1223 	for_each_rxq(pi, i, rxq) {
1224 		if ((rxq->iq.flags & IQ_INTR) == 0)
1225 			free_rxq(pi, rxq);
1226 	}
1227 #ifdef TCP_OFFLOAD
1228 	for_each_ofld_rxq(pi, i, ofld_rxq) {
1229 		if ((ofld_rxq->iq.flags & IQ_INTR) == 0)
1230 			free_ofld_rxq(pi, ofld_rxq);
1231 	}
1232 #endif
1233 #ifdef DEV_NETMAP
1234 	for_each_nm_rxq(pi, i, nm_rxq)
1235 	    free_nm_rxq(pi, nm_rxq);
1236 #endif
1237 
1238 	/*
1239 	 * Then take down the rx queues that take direct interrupts.
1240 	 */
1241 
1242 	for_each_rxq(pi, i, rxq) {
1243 		if (rxq->iq.flags & IQ_INTR)
1244 			free_rxq(pi, rxq);
1245 	}
1246 #ifdef TCP_OFFLOAD
1247 	for_each_ofld_rxq(pi, i, ofld_rxq) {
1248 		if (ofld_rxq->iq.flags & IQ_INTR)
1249 			free_ofld_rxq(pi, ofld_rxq);
1250 	}
1251 #endif
1252 
1253 	return (0);
1254 }
1255 
1256 /*
1257  * Deals with errors and the firmware event queue.  All data rx queues forward
1258  * their interrupt to the firmware event queue.
1259  */
1260 void
1261 t4_intr_all(void *arg)
1262 {
1263 	struct adapter *sc = arg;
1264 	struct sge_iq *fwq = &sc->sge.fwq;
1265 
1266 	t4_intr_err(arg);
1267 	if (atomic_cmpset_int(&fwq->state, IQS_IDLE, IQS_BUSY)) {
1268 		service_iq(fwq, 0);
1269 		atomic_cmpset_int(&fwq->state, IQS_BUSY, IQS_IDLE);
1270 	}
1271 }
1272 
1273 /* Deals with error interrupts */
1274 void
1275 t4_intr_err(void *arg)
1276 {
1277 	struct adapter *sc = arg;
1278 
1279 	t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
1280 	t4_slow_intr_handler(sc);
1281 }
1282 
1283 void
1284 t4_intr_evt(void *arg)
1285 {
1286 	struct sge_iq *iq = arg;
1287 
1288 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1289 		service_iq(iq, 0);
1290 		atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1291 	}
1292 }
1293 
1294 void
1295 t4_intr(void *arg)
1296 {
1297 	struct sge_iq *iq = arg;
1298 
1299 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1300 		service_iq(iq, 0);
1301 		atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1302 	}
1303 }
1304 
1305 /*
1306  * Deals with anything and everything on the given ingress queue.
1307  */
1308 static int
1309 service_iq(struct sge_iq *iq, int budget)
1310 {
1311 	struct sge_iq *q;
1312 	struct sge_rxq *rxq = iq_to_rxq(iq);	/* Use iff iq is part of rxq */
1313 	struct sge_fl *fl;			/* Use iff IQ_HAS_FL */
1314 	struct adapter *sc = iq->adapter;
1315 	struct iq_desc *d = &iq->desc[iq->cidx];
1316 	int ndescs = 0, limit;
1317 	int rsp_type, refill;
1318 	uint32_t lq;
1319 	uint16_t fl_hw_cidx;
1320 	struct mbuf *m0;
1321 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1322 #if defined(INET) || defined(INET6)
1323 	const struct timeval lro_timeout = {0, sc->lro_timeout};
1324 #endif
1325 
1326 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1327 
1328 	limit = budget ? budget : iq->qsize / 16;
1329 
1330 	if (iq->flags & IQ_HAS_FL) {
1331 		fl = &rxq->fl;
1332 		fl_hw_cidx = fl->hw_cidx;	/* stable snapshot */
1333 	} else {
1334 		fl = NULL;
1335 		fl_hw_cidx = 0;			/* to silence gcc warning */
1336 	}
1337 
1338 	/*
1339 	 * We always come back and check the descriptor ring for new indirect
1340 	 * interrupts and other responses after running a single handler.
1341 	 */
1342 	for (;;) {
1343 		while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1344 
1345 			rmb();
1346 
1347 			refill = 0;
1348 			m0 = NULL;
1349 			rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1350 			lq = be32toh(d->rsp.pldbuflen_qid);
1351 
1352 			switch (rsp_type) {
1353 			case X_RSPD_TYPE_FLBUF:
1354 
1355 				KASSERT(iq->flags & IQ_HAS_FL,
1356 				    ("%s: data for an iq (%p) with no freelist",
1357 				    __func__, iq));
1358 
1359 				m0 = get_fl_payload(sc, fl, lq);
1360 				if (__predict_false(m0 == NULL))
1361 					goto process_iql;
1362 				refill = IDXDIFF(fl->hw_cidx, fl_hw_cidx, fl->sidx) > 2;
1363 #ifdef T4_PKT_TIMESTAMP
1364 				/*
1365 				 * 60 bit timestamp for the payload is
1366 				 * *(uint64_t *)m0->m_pktdat.  Note that it is
1367 				 * in the leading free-space in the mbuf.  The
1368 				 * kernel can clobber it during a pullup,
1369 				 * m_copymdata, etc.  You need to make sure that
1370 				 * the mbuf reaches you unmolested if you care
1371 				 * about the timestamp.
1372 				 */
1373 				*(uint64_t *)m0->m_pktdat =
1374 				    be64toh(ctrl->u.last_flit) &
1375 				    0xfffffffffffffff;
1376 #endif
1377 
1378 				/* fall through */
1379 
1380 			case X_RSPD_TYPE_CPL:
1381 				KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1382 				    ("%s: bad opcode %02x.", __func__,
1383 				    d->rss.opcode));
1384 				sc->cpl_handler[d->rss.opcode](iq, &d->rss, m0);
1385 				break;
1386 
1387 			case X_RSPD_TYPE_INTR:
1388 
1389 				/*
1390 				 * Interrupts should be forwarded only to queues
1391 				 * that are not forwarding their interrupts.
1392 				 * This means service_iq can recurse but only 1
1393 				 * level deep.
1394 				 */
1395 				KASSERT(budget == 0,
1396 				    ("%s: budget %u, rsp_type %u", __func__,
1397 				    budget, rsp_type));
1398 
1399 				/*
1400 				 * There are 1K interrupt-capable queues (qids 0
1401 				 * through 1023).  A response type indicating a
1402 				 * forwarded interrupt with a qid >= 1K is an
1403 				 * iWARP async notification.
1404 				 */
1405 				if (lq >= 1024) {
1406                                         sc->an_handler(iq, &d->rsp);
1407                                         break;
1408                                 }
1409 
1410 				q = sc->sge.iqmap[lq - sc->sge.iq_start];
1411 				if (atomic_cmpset_int(&q->state, IQS_IDLE,
1412 				    IQS_BUSY)) {
1413 					if (service_iq(q, q->qsize / 16) == 0) {
1414 						atomic_cmpset_int(&q->state,
1415 						    IQS_BUSY, IQS_IDLE);
1416 					} else {
1417 						STAILQ_INSERT_TAIL(&iql, q,
1418 						    link);
1419 					}
1420 				}
1421 				break;
1422 
1423 			default:
1424 				KASSERT(0,
1425 				    ("%s: illegal response type %d on iq %p",
1426 				    __func__, rsp_type, iq));
1427 				log(LOG_ERR,
1428 				    "%s: illegal response type %d on iq %p",
1429 				    device_get_nameunit(sc->dev), rsp_type, iq);
1430 				break;
1431 			}
1432 
1433 			d++;
1434 			if (__predict_false(++iq->cidx == iq->sidx)) {
1435 				iq->cidx = 0;
1436 				iq->gen ^= F_RSPD_GEN;
1437 				d = &iq->desc[0];
1438 			}
1439 			if (__predict_false(++ndescs == limit)) {
1440 				t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
1441 				    V_CIDXINC(ndescs) |
1442 				    V_INGRESSQID(iq->cntxt_id) |
1443 				    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1444 				ndescs = 0;
1445 
1446 #if defined(INET) || defined(INET6)
1447 				if (iq->flags & IQ_LRO_ENABLED &&
1448 				    sc->lro_timeout != 0) {
1449 					tcp_lro_flush_inactive(&rxq->lro,
1450 					    &lro_timeout);
1451 				}
1452 #endif
1453 
1454 				if (budget) {
1455 					if (iq->flags & IQ_HAS_FL) {
1456 						FL_LOCK(fl);
1457 						refill_fl(sc, fl, 32);
1458 						FL_UNLOCK(fl);
1459 					}
1460 					return (EINPROGRESS);
1461 				}
1462 			}
1463 			if (refill) {
1464 				FL_LOCK(fl);
1465 				refill_fl(sc, fl, 32);
1466 				FL_UNLOCK(fl);
1467 				fl_hw_cidx = fl->hw_cidx;
1468 			}
1469 		}
1470 
1471 process_iql:
1472 		if (STAILQ_EMPTY(&iql))
1473 			break;
1474 
1475 		/*
1476 		 * Process the head only, and send it to the back of the list if
1477 		 * it's still not done.
1478 		 */
1479 		q = STAILQ_FIRST(&iql);
1480 		STAILQ_REMOVE_HEAD(&iql, link);
1481 		if (service_iq(q, q->qsize / 8) == 0)
1482 			atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1483 		else
1484 			STAILQ_INSERT_TAIL(&iql, q, link);
1485 	}
1486 
1487 #if defined(INET) || defined(INET6)
1488 	if (iq->flags & IQ_LRO_ENABLED) {
1489 		struct lro_ctrl *lro = &rxq->lro;
1490 		struct lro_entry *l;
1491 
1492 		while (!SLIST_EMPTY(&lro->lro_active)) {
1493 			l = SLIST_FIRST(&lro->lro_active);
1494 			SLIST_REMOVE_HEAD(&lro->lro_active, next);
1495 			tcp_lro_flush(lro, l);
1496 		}
1497 	}
1498 #endif
1499 
1500 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_CIDXINC(ndescs) |
1501 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1502 
1503 	if (iq->flags & IQ_HAS_FL) {
1504 		int starved;
1505 
1506 		FL_LOCK(fl);
1507 		starved = refill_fl(sc, fl, 64);
1508 		FL_UNLOCK(fl);
1509 		if (__predict_false(starved != 0))
1510 			add_fl_to_sfl(sc, fl);
1511 	}
1512 
1513 	return (0);
1514 }
1515 
1516 static inline int
1517 cl_has_metadata(struct sge_fl *fl, struct cluster_layout *cll)
1518 {
1519 	int rc = fl->flags & FL_BUF_PACKING || cll->region1 > 0;
1520 
1521 	if (rc)
1522 		MPASS(cll->region3 >= CL_METADATA_SIZE);
1523 
1524 	return (rc);
1525 }
1526 
1527 static inline struct cluster_metadata *
1528 cl_metadata(struct adapter *sc, struct sge_fl *fl, struct cluster_layout *cll,
1529     caddr_t cl)
1530 {
1531 
1532 	if (cl_has_metadata(fl, cll)) {
1533 		struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1534 
1535 		return ((struct cluster_metadata *)(cl + swz->size) - 1);
1536 	}
1537 	return (NULL);
1538 }
1539 
1540 static void
1541 rxb_free(struct mbuf *m, void *arg1, void *arg2)
1542 {
1543 	uma_zone_t zone = arg1;
1544 	caddr_t cl = arg2;
1545 
1546 	uma_zfree(zone, cl);
1547 	counter_u64_add(extfree_rels, 1);
1548 }
1549 
1550 /*
1551  * The mbuf returned by this function could be allocated from zone_mbuf or
1552  * constructed in spare room in the cluster.
1553  *
1554  * The mbuf carries the payload in one of these ways
1555  * a) frame inside the mbuf (mbuf from zone_mbuf)
1556  * b) m_cljset (for clusters without metadata) zone_mbuf
1557  * c) m_extaddref (cluster with metadata) inline mbuf
1558  * d) m_extaddref (cluster with metadata) zone_mbuf
1559  */
1560 static struct mbuf *
1561 get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1562     int remaining)
1563 {
1564 	struct mbuf *m;
1565 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1566 	struct cluster_layout *cll = &sd->cll;
1567 	struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1568 	struct hw_buf_info *hwb = &sc->sge.hw_buf_info[cll->hwidx];
1569 	struct cluster_metadata *clm = cl_metadata(sc, fl, cll, sd->cl);
1570 	int len, blen;
1571 	caddr_t payload;
1572 
1573 	blen = hwb->size - fl->rx_offset;	/* max possible in this buf */
1574 	len = min(remaining, blen);
1575 	payload = sd->cl + cll->region1 + fl->rx_offset;
1576 	if (fl->flags & FL_BUF_PACKING) {
1577 		const u_int l = fr_offset + len;
1578 		const u_int pad = roundup2(l, fl->buf_boundary) - l;
1579 
1580 		if (fl->rx_offset + len + pad < hwb->size)
1581 			blen = len + pad;
1582 		MPASS(fl->rx_offset + blen <= hwb->size);
1583 	} else {
1584 		MPASS(fl->rx_offset == 0);	/* not packing */
1585 	}
1586 
1587 
1588 	if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) {
1589 
1590 		/*
1591 		 * Copy payload into a freshly allocated mbuf.
1592 		 */
1593 
1594 		m = fr_offset == 0 ?
1595 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1596 		if (m == NULL)
1597 			return (NULL);
1598 		fl->mbuf_allocated++;
1599 #ifdef T4_PKT_TIMESTAMP
1600 		/* Leave room for a timestamp */
1601 		m->m_data += 8;
1602 #endif
1603 		/* copy data to mbuf */
1604 		bcopy(payload, mtod(m, caddr_t), len);
1605 
1606 	} else if (sd->nmbuf * MSIZE < cll->region1) {
1607 
1608 		/*
1609 		 * There's spare room in the cluster for an mbuf.  Create one
1610 		 * and associate it with the payload that's in the cluster.
1611 		 */
1612 
1613 		MPASS(clm != NULL);
1614 		m = (struct mbuf *)(sd->cl + sd->nmbuf * MSIZE);
1615 		/* No bzero required */
1616 		if (m_init(m, NULL, 0, M_NOWAIT, MT_DATA,
1617 		    fr_offset == 0 ? M_PKTHDR | M_NOFREE : M_NOFREE))
1618 			return (NULL);
1619 		fl->mbuf_inlined++;
1620 		m_extaddref(m, payload, blen, &clm->refcount, rxb_free,
1621 		    swz->zone, sd->cl);
1622 		if (sd->nmbuf++ == 0)
1623 			counter_u64_add(extfree_refs, 1);
1624 
1625 	} else {
1626 
1627 		/*
1628 		 * Grab an mbuf from zone_mbuf and associate it with the
1629 		 * payload in the cluster.
1630 		 */
1631 
1632 		m = fr_offset == 0 ?
1633 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1634 		if (m == NULL)
1635 			return (NULL);
1636 		fl->mbuf_allocated++;
1637 		if (clm != NULL) {
1638 			m_extaddref(m, payload, blen, &clm->refcount,
1639 			    rxb_free, swz->zone, sd->cl);
1640 			if (sd->nmbuf++ == 0)
1641 				counter_u64_add(extfree_refs, 1);
1642 		} else {
1643 			m_cljset(m, sd->cl, swz->type);
1644 			sd->cl = NULL;	/* consumed, not a recycle candidate */
1645 		}
1646 	}
1647 	if (fr_offset == 0)
1648 		m->m_pkthdr.len = remaining;
1649 	m->m_len = len;
1650 
1651 	if (fl->flags & FL_BUF_PACKING) {
1652 		fl->rx_offset += blen;
1653 		MPASS(fl->rx_offset <= hwb->size);
1654 		if (fl->rx_offset < hwb->size)
1655 			return (m);	/* without advancing the cidx */
1656 	}
1657 
1658 	if (__predict_false(++fl->cidx % 8 == 0)) {
1659 		uint16_t cidx = fl->cidx / 8;
1660 
1661 		if (__predict_false(cidx == fl->sidx))
1662 			fl->cidx = cidx = 0;
1663 		fl->hw_cidx = cidx;
1664 	}
1665 	fl->rx_offset = 0;
1666 
1667 	return (m);
1668 }
1669 
1670 static struct mbuf *
1671 get_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf)
1672 {
1673 	struct mbuf *m0, *m, **pnext;
1674 	u_int remaining;
1675 	const u_int total = G_RSPD_LEN(len_newbuf);
1676 
1677 	if (__predict_false(fl->flags & FL_BUF_RESUME)) {
1678 		M_ASSERTPKTHDR(fl->m0);
1679 		MPASS(fl->m0->m_pkthdr.len == total);
1680 		MPASS(fl->remaining < total);
1681 
1682 		m0 = fl->m0;
1683 		pnext = fl->pnext;
1684 		remaining = fl->remaining;
1685 		fl->flags &= ~FL_BUF_RESUME;
1686 		goto get_segment;
1687 	}
1688 
1689 	if (fl->rx_offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
1690 		fl->rx_offset = 0;
1691 		if (__predict_false(++fl->cidx % 8 == 0)) {
1692 			uint16_t cidx = fl->cidx / 8;
1693 
1694 			if (__predict_false(cidx == fl->sidx))
1695 				fl->cidx = cidx = 0;
1696 			fl->hw_cidx = cidx;
1697 		}
1698 	}
1699 
1700 	/*
1701 	 * Payload starts at rx_offset in the current hw buffer.  Its length is
1702 	 * 'len' and it may span multiple hw buffers.
1703 	 */
1704 
1705 	m0 = get_scatter_segment(sc, fl, 0, total);
1706 	if (m0 == NULL)
1707 		return (NULL);
1708 	remaining = total - m0->m_len;
1709 	pnext = &m0->m_next;
1710 	while (remaining > 0) {
1711 get_segment:
1712 		MPASS(fl->rx_offset == 0);
1713 		m = get_scatter_segment(sc, fl, total - remaining, remaining);
1714 		if (__predict_false(m == NULL)) {
1715 			fl->m0 = m0;
1716 			fl->pnext = pnext;
1717 			fl->remaining = remaining;
1718 			fl->flags |= FL_BUF_RESUME;
1719 			return (NULL);
1720 		}
1721 		*pnext = m;
1722 		pnext = &m->m_next;
1723 		remaining -= m->m_len;
1724 	}
1725 	*pnext = NULL;
1726 
1727 	return (m0);
1728 }
1729 
1730 static int
1731 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
1732 {
1733 	struct sge_rxq *rxq = iq_to_rxq(iq);
1734 	struct ifnet *ifp = rxq->ifp;
1735 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
1736 #if defined(INET) || defined(INET6)
1737 	struct lro_ctrl *lro = &rxq->lro;
1738 #endif
1739 
1740 	KASSERT(m0 != NULL, ("%s: no payload with opcode %02x", __func__,
1741 	    rss->opcode));
1742 
1743 	m0->m_pkthdr.len -= fl_pktshift;
1744 	m0->m_len -= fl_pktshift;
1745 	m0->m_data += fl_pktshift;
1746 
1747 	m0->m_pkthdr.rcvif = ifp;
1748 	M_HASHTYPE_SET(m0, M_HASHTYPE_OPAQUE);
1749 	m0->m_pkthdr.flowid = be32toh(rss->hash_val);
1750 
1751 	if (cpl->csum_calc && !cpl->err_vec) {
1752 		if (ifp->if_capenable & IFCAP_RXCSUM &&
1753 		    cpl->l2info & htobe32(F_RXF_IP)) {
1754 			m0->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
1755 			    CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
1756 			rxq->rxcsum++;
1757 		} else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
1758 		    cpl->l2info & htobe32(F_RXF_IP6)) {
1759 			m0->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
1760 			    CSUM_PSEUDO_HDR);
1761 			rxq->rxcsum++;
1762 		}
1763 
1764 		if (__predict_false(cpl->ip_frag))
1765 			m0->m_pkthdr.csum_data = be16toh(cpl->csum);
1766 		else
1767 			m0->m_pkthdr.csum_data = 0xffff;
1768 	}
1769 
1770 	if (cpl->vlan_ex) {
1771 		m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
1772 		m0->m_flags |= M_VLANTAG;
1773 		rxq->vlan_extraction++;
1774 	}
1775 
1776 #if defined(INET) || defined(INET6)
1777 	if (cpl->l2info & htobe32(F_RXF_LRO) &&
1778 	    iq->flags & IQ_LRO_ENABLED &&
1779 	    tcp_lro_rx(lro, m0, 0) == 0) {
1780 		/* queued for LRO */
1781 	} else
1782 #endif
1783 	ifp->if_input(ifp, m0);
1784 
1785 	return (0);
1786 }
1787 
1788 /*
1789  * Must drain the wrq or make sure that someone else will.
1790  */
1791 static void
1792 wrq_tx_drain(void *arg, int n)
1793 {
1794 	struct sge_wrq *wrq = arg;
1795 	struct sge_eq *eq = &wrq->eq;
1796 
1797 	EQ_LOCK(eq);
1798 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
1799 		drain_wrq_wr_list(wrq->adapter, wrq);
1800 	EQ_UNLOCK(eq);
1801 }
1802 
1803 static void
1804 drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq)
1805 {
1806 	struct sge_eq *eq = &wrq->eq;
1807 	u_int available, dbdiff;	/* # of hardware descriptors */
1808 	u_int n;
1809 	struct wrqe *wr;
1810 	struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
1811 
1812 	EQ_LOCK_ASSERT_OWNED(eq);
1813 	MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
1814 	wr = STAILQ_FIRST(&wrq->wr_list);
1815 	MPASS(wr != NULL);	/* Must be called with something useful to do */
1816 	dbdiff = IDXDIFF(eq->pidx, eq->dbidx, eq->sidx);
1817 
1818 	do {
1819 		eq->cidx = read_hw_cidx(eq);
1820 		if (eq->pidx == eq->cidx)
1821 			available = eq->sidx - 1;
1822 		else
1823 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
1824 
1825 		MPASS(wr->wrq == wrq);
1826 		n = howmany(wr->wr_len, EQ_ESIZE);
1827 		if (available < n)
1828 			return;
1829 
1830 		dst = (void *)&eq->desc[eq->pidx];
1831 		if (__predict_true(eq->sidx - eq->pidx > n)) {
1832 			/* Won't wrap, won't end exactly at the status page. */
1833 			bcopy(&wr->wr[0], dst, wr->wr_len);
1834 			eq->pidx += n;
1835 		} else {
1836 			int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE;
1837 
1838 			bcopy(&wr->wr[0], dst, first_portion);
1839 			if (wr->wr_len > first_portion) {
1840 				bcopy(&wr->wr[first_portion], &eq->desc[0],
1841 				    wr->wr_len - first_portion);
1842 			}
1843 			eq->pidx = n - (eq->sidx - eq->pidx);
1844 		}
1845 
1846 		if (available < eq->sidx / 4 &&
1847 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
1848 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
1849 			    F_FW_WR_EQUEQ);
1850 			eq->equeqidx = eq->pidx;
1851 		} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
1852 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
1853 			eq->equeqidx = eq->pidx;
1854 		}
1855 
1856 		dbdiff += n;
1857 		if (dbdiff >= 16) {
1858 			ring_eq_db(sc, eq, dbdiff);
1859 			dbdiff = 0;
1860 		}
1861 
1862 		STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
1863 		free_wrqe(wr);
1864 		MPASS(wrq->nwr_pending > 0);
1865 		wrq->nwr_pending--;
1866 		MPASS(wrq->ndesc_needed >= n);
1867 		wrq->ndesc_needed -= n;
1868 	} while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL);
1869 
1870 	if (dbdiff)
1871 		ring_eq_db(sc, eq, dbdiff);
1872 }
1873 
1874 /*
1875  * Doesn't fail.  Holds on to work requests it can't send right away.
1876  */
1877 void
1878 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
1879 {
1880 #ifdef INVARIANTS
1881 	struct sge_eq *eq = &wrq->eq;
1882 #endif
1883 
1884 	EQ_LOCK_ASSERT_OWNED(eq);
1885 	MPASS(wr != NULL);
1886 	MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN);
1887 	MPASS((wr->wr_len & 0x7) == 0);
1888 
1889 	STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
1890 	wrq->nwr_pending++;
1891 	wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE);
1892 
1893 	if (!TAILQ_EMPTY(&wrq->incomplete_wrs))
1894 		return;	/* commit_wrq_wr will drain wr_list as well. */
1895 
1896 	drain_wrq_wr_list(sc, wrq);
1897 
1898 	/* Doorbell must have caught up to the pidx. */
1899 	MPASS(eq->pidx == eq->dbidx);
1900 }
1901 
1902 void
1903 t4_update_fl_bufsize(struct ifnet *ifp)
1904 {
1905 	struct port_info *pi = ifp->if_softc;
1906 	struct adapter *sc = pi->adapter;
1907 	struct sge_rxq *rxq;
1908 #ifdef TCP_OFFLOAD
1909 	struct sge_ofld_rxq *ofld_rxq;
1910 #endif
1911 	struct sge_fl *fl;
1912 	int i, maxp, mtu = ifp->if_mtu;
1913 
1914 	maxp = mtu_to_max_payload(sc, mtu, 0);
1915 	for_each_rxq(pi, i, rxq) {
1916 		fl = &rxq->fl;
1917 
1918 		FL_LOCK(fl);
1919 		find_best_refill_source(sc, fl, maxp);
1920 		FL_UNLOCK(fl);
1921 	}
1922 #ifdef TCP_OFFLOAD
1923 	maxp = mtu_to_max_payload(sc, mtu, 1);
1924 	for_each_ofld_rxq(pi, i, ofld_rxq) {
1925 		fl = &ofld_rxq->fl;
1926 
1927 		FL_LOCK(fl);
1928 		find_best_refill_source(sc, fl, maxp);
1929 		FL_UNLOCK(fl);
1930 	}
1931 #endif
1932 }
1933 
1934 static inline int
1935 mbuf_nsegs(struct mbuf *m)
1936 {
1937 
1938 	M_ASSERTPKTHDR(m);
1939 	KASSERT(m->m_pkthdr.l5hlen > 0,
1940 	    ("%s: mbuf %p missing information on # of segments.", __func__, m));
1941 
1942 	return (m->m_pkthdr.l5hlen);
1943 }
1944 
1945 static inline void
1946 set_mbuf_nsegs(struct mbuf *m, uint8_t nsegs)
1947 {
1948 
1949 	M_ASSERTPKTHDR(m);
1950 	m->m_pkthdr.l5hlen = nsegs;
1951 }
1952 
1953 static inline int
1954 mbuf_len16(struct mbuf *m)
1955 {
1956 	int n;
1957 
1958 	M_ASSERTPKTHDR(m);
1959 	n = m->m_pkthdr.PH_loc.eight[0];
1960 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
1961 
1962 	return (n);
1963 }
1964 
1965 static inline void
1966 set_mbuf_len16(struct mbuf *m, uint8_t len16)
1967 {
1968 
1969 	M_ASSERTPKTHDR(m);
1970 	m->m_pkthdr.PH_loc.eight[0] = len16;
1971 }
1972 
1973 static inline int
1974 needs_tso(struct mbuf *m)
1975 {
1976 
1977 	M_ASSERTPKTHDR(m);
1978 
1979 	if (m->m_pkthdr.csum_flags & CSUM_TSO) {
1980 		KASSERT(m->m_pkthdr.tso_segsz > 0,
1981 		    ("%s: TSO requested in mbuf %p but MSS not provided",
1982 		    __func__, m));
1983 		return (1);
1984 	}
1985 
1986 	return (0);
1987 }
1988 
1989 static inline int
1990 needs_l3_csum(struct mbuf *m)
1991 {
1992 
1993 	M_ASSERTPKTHDR(m);
1994 
1995 	if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO))
1996 		return (1);
1997 	return (0);
1998 }
1999 
2000 static inline int
2001 needs_l4_csum(struct mbuf *m)
2002 {
2003 
2004 	M_ASSERTPKTHDR(m);
2005 
2006 	if (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
2007 	    CSUM_TCP_IPV6 | CSUM_TSO))
2008 		return (1);
2009 	return (0);
2010 }
2011 
2012 static inline int
2013 needs_vlan_insertion(struct mbuf *m)
2014 {
2015 
2016 	M_ASSERTPKTHDR(m);
2017 
2018 	if (m->m_flags & M_VLANTAG) {
2019 		KASSERT(m->m_pkthdr.ether_vtag != 0,
2020 		    ("%s: HWVLAN requested in mbuf %p but tag not provided",
2021 		    __func__, m));
2022 		return (1);
2023 	}
2024 	return (0);
2025 }
2026 
2027 static void *
2028 m_advance(struct mbuf **pm, int *poffset, int len)
2029 {
2030 	struct mbuf *m = *pm;
2031 	int offset = *poffset;
2032 	uintptr_t p = 0;
2033 
2034 	MPASS(len > 0);
2035 
2036 	while (len) {
2037 		if (offset + len < m->m_len) {
2038 			offset += len;
2039 			p = mtod(m, uintptr_t) + offset;
2040 			break;
2041 		}
2042 		len -= m->m_len - offset;
2043 		m = m->m_next;
2044 		offset = 0;
2045 		MPASS(m != NULL);
2046 	}
2047 	*poffset = offset;
2048 	*pm = m;
2049 	return ((void *)p);
2050 }
2051 
2052 static inline int
2053 same_paddr(char *a, char *b)
2054 {
2055 
2056 	if (a == b)
2057 		return (1);
2058 	else if (a != NULL && b != NULL) {
2059 		vm_offset_t x = (vm_offset_t)a;
2060 		vm_offset_t y = (vm_offset_t)b;
2061 
2062 		if ((x & PAGE_MASK) == (y & PAGE_MASK) &&
2063 		    pmap_kextract(x) == pmap_kextract(y))
2064 			return (1);
2065 	}
2066 
2067 	return (0);
2068 }
2069 
2070 /*
2071  * Can deal with empty mbufs in the chain that have m_len = 0, but the chain
2072  * must have at least one mbuf that's not empty.
2073  */
2074 static inline int
2075 count_mbuf_nsegs(struct mbuf *m)
2076 {
2077 	char *prev_end, *start;
2078 	int len, nsegs;
2079 
2080 	MPASS(m != NULL);
2081 
2082 	nsegs = 0;
2083 	prev_end = NULL;
2084 	for (; m; m = m->m_next) {
2085 
2086 		len = m->m_len;
2087 		if (__predict_false(len == 0))
2088 			continue;
2089 		start = mtod(m, char *);
2090 
2091 		nsegs += sglist_count(start, len);
2092 		if (same_paddr(prev_end, start))
2093 			nsegs--;
2094 		prev_end = start + len;
2095 	}
2096 
2097 	MPASS(nsegs > 0);
2098 	return (nsegs);
2099 }
2100 
2101 /*
2102  * Analyze the mbuf to determine its tx needs.  The mbuf passed in may change:
2103  * a) caller can assume it's been freed if this function returns with an error.
2104  * b) it may get defragged up if the gather list is too long for the hardware.
2105  */
2106 int
2107 parse_pkt(struct mbuf **mp)
2108 {
2109 	struct mbuf *m0 = *mp, *m;
2110 	int rc, nsegs, defragged = 0, offset;
2111 	struct ether_header *eh;
2112 	void *l3hdr;
2113 #if defined(INET) || defined(INET6)
2114 	struct tcphdr *tcp;
2115 #endif
2116 	uint16_t eh_type;
2117 
2118 	M_ASSERTPKTHDR(m0);
2119 	if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) {
2120 		rc = EINVAL;
2121 fail:
2122 		m_freem(m0);
2123 		*mp = NULL;
2124 		return (rc);
2125 	}
2126 restart:
2127 	/*
2128 	 * First count the number of gather list segments in the payload.
2129 	 * Defrag the mbuf if nsegs exceeds the hardware limit.
2130 	 */
2131 	M_ASSERTPKTHDR(m0);
2132 	MPASS(m0->m_pkthdr.len > 0);
2133 	nsegs = count_mbuf_nsegs(m0);
2134 	if (nsegs > (needs_tso(m0) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS)) {
2135 		if (defragged++ > 0 || (m = m_defrag(m0, M_NOWAIT)) == NULL) {
2136 			rc = EFBIG;
2137 			goto fail;
2138 		}
2139 		*mp = m0 = m;	/* update caller's copy after defrag */
2140 		goto restart;
2141 	}
2142 
2143 	if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN)) {
2144 		m0 = m_pullup(m0, m0->m_pkthdr.len);
2145 		if (m0 == NULL) {
2146 			/* Should have left well enough alone. */
2147 			rc = EFBIG;
2148 			goto fail;
2149 		}
2150 		*mp = m0;	/* update caller's copy after pullup */
2151 		goto restart;
2152 	}
2153 	set_mbuf_nsegs(m0, nsegs);
2154 	set_mbuf_len16(m0, txpkt_len16(nsegs, needs_tso(m0)));
2155 
2156 	if (!needs_tso(m0))
2157 		return (0);
2158 
2159 	m = m0;
2160 	eh = mtod(m, struct ether_header *);
2161 	eh_type = ntohs(eh->ether_type);
2162 	if (eh_type == ETHERTYPE_VLAN) {
2163 		struct ether_vlan_header *evh = (void *)eh;
2164 
2165 		eh_type = ntohs(evh->evl_proto);
2166 		m0->m_pkthdr.l2hlen = sizeof(*evh);
2167 	} else
2168 		m0->m_pkthdr.l2hlen = sizeof(*eh);
2169 
2170 	offset = 0;
2171 	l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2172 
2173 	switch (eh_type) {
2174 #ifdef INET6
2175 	case ETHERTYPE_IPV6:
2176 	{
2177 		struct ip6_hdr *ip6 = l3hdr;
2178 
2179 		MPASS(ip6->ip6_nxt == IPPROTO_TCP);
2180 
2181 		m0->m_pkthdr.l3hlen = sizeof(*ip6);
2182 		break;
2183 	}
2184 #endif
2185 #ifdef INET
2186 	case ETHERTYPE_IP:
2187 	{
2188 		struct ip *ip = l3hdr;
2189 
2190 		m0->m_pkthdr.l3hlen = ip->ip_hl * 4;
2191 		break;
2192 	}
2193 #endif
2194 	default:
2195 		panic("%s: ethertype 0x%04x unknown.  if_cxgbe must be compiled"
2196 		    " with the same INET/INET6 options as the kernel.",
2197 		    __func__, eh_type);
2198 	}
2199 
2200 #if defined(INET) || defined(INET6)
2201 	tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen);
2202 	m0->m_pkthdr.l4hlen = tcp->th_off * 4;
2203 #endif
2204 	MPASS(m0 == *mp);
2205 	return (0);
2206 }
2207 
2208 void *
2209 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie)
2210 {
2211 	struct sge_eq *eq = &wrq->eq;
2212 	struct adapter *sc = wrq->adapter;
2213 	int ndesc, available;
2214 	struct wrqe *wr;
2215 	void *w;
2216 
2217 	MPASS(len16 > 0);
2218 	ndesc = howmany(len16, EQ_ESIZE / 16);
2219 	MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC);
2220 
2221 	EQ_LOCK(eq);
2222 
2223 	if (!STAILQ_EMPTY(&wrq->wr_list))
2224 		drain_wrq_wr_list(sc, wrq);
2225 
2226 	if (!STAILQ_EMPTY(&wrq->wr_list)) {
2227 slowpath:
2228 		EQ_UNLOCK(eq);
2229 		wr = alloc_wrqe(len16 * 16, wrq);
2230 		if (__predict_false(wr == NULL))
2231 			return (NULL);
2232 		cookie->pidx = -1;
2233 		cookie->ndesc = ndesc;
2234 		return (&wr->wr);
2235 	}
2236 
2237 	eq->cidx = read_hw_cidx(eq);
2238 	if (eq->pidx == eq->cidx)
2239 		available = eq->sidx - 1;
2240 	else
2241 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2242 	if (available < ndesc)
2243 		goto slowpath;
2244 
2245 	cookie->pidx = eq->pidx;
2246 	cookie->ndesc = ndesc;
2247 	TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link);
2248 
2249 	w = &eq->desc[eq->pidx];
2250 	IDXINCR(eq->pidx, ndesc, eq->sidx);
2251 	if (__predict_false(eq->pidx < ndesc - 1)) {
2252 		w = &wrq->ss[0];
2253 		wrq->ss_pidx = cookie->pidx;
2254 		wrq->ss_len = len16 * 16;
2255 	}
2256 
2257 	EQ_UNLOCK(eq);
2258 
2259 	return (w);
2260 }
2261 
2262 void
2263 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie)
2264 {
2265 	struct sge_eq *eq = &wrq->eq;
2266 	struct adapter *sc = wrq->adapter;
2267 	int ndesc, pidx;
2268 	struct wrq_cookie *prev, *next;
2269 
2270 	if (cookie->pidx == -1) {
2271 		struct wrqe *wr = __containerof(w, struct wrqe, wr);
2272 
2273 		t4_wrq_tx(sc, wr);
2274 		return;
2275 	}
2276 
2277 	ndesc = cookie->ndesc;	/* Can be more than SGE_MAX_WR_NDESC here. */
2278 	pidx = cookie->pidx;
2279 	MPASS(pidx >= 0 && pidx < eq->sidx);
2280 	if (__predict_false(w == &wrq->ss[0])) {
2281 		int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE;
2282 
2283 		MPASS(wrq->ss_len > n);	/* WR had better wrap around. */
2284 		bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n);
2285 		bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n);
2286 		wrq->tx_wrs_ss++;
2287 	} else
2288 		wrq->tx_wrs_direct++;
2289 
2290 	EQ_LOCK(eq);
2291 	prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link);
2292 	next = TAILQ_NEXT(cookie, link);
2293 	if (prev == NULL) {
2294 		MPASS(pidx == eq->dbidx);
2295 		if (next == NULL || ndesc >= 16)
2296 			ring_eq_db(wrq->adapter, eq, ndesc);
2297 		else {
2298 			MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc);
2299 			next->pidx = pidx;
2300 			next->ndesc += ndesc;
2301 		}
2302 	} else {
2303 		MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc);
2304 		prev->ndesc += ndesc;
2305 	}
2306 	TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link);
2307 
2308 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2309 		drain_wrq_wr_list(sc, wrq);
2310 
2311 #ifdef INVARIANTS
2312 	if (TAILQ_EMPTY(&wrq->incomplete_wrs)) {
2313 		/* Doorbell must have caught up to the pidx. */
2314 		MPASS(wrq->eq.pidx == wrq->eq.dbidx);
2315 	}
2316 #endif
2317 	EQ_UNLOCK(eq);
2318 }
2319 
2320 static u_int
2321 can_resume_eth_tx(struct mp_ring *r)
2322 {
2323 	struct sge_eq *eq = r->cookie;
2324 
2325 	return (total_available_tx_desc(eq) > eq->sidx / 8);
2326 }
2327 
2328 static inline int
2329 cannot_use_txpkts(struct mbuf *m)
2330 {
2331 	/* maybe put a GL limit too, to avoid silliness? */
2332 
2333 	return (needs_tso(m));
2334 }
2335 
2336 /*
2337  * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to
2338  * be consumed.  Return the actual number consumed.  0 indicates a stall.
2339  */
2340 static u_int
2341 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx)
2342 {
2343 	struct sge_txq *txq = r->cookie;
2344 	struct sge_eq *eq = &txq->eq;
2345 	struct ifnet *ifp = txq->ifp;
2346 	struct port_info *pi = (void *)ifp->if_softc;
2347 	struct adapter *sc = pi->adapter;
2348 	u_int total, remaining;		/* # of packets */
2349 	u_int available, dbdiff;	/* # of hardware descriptors */
2350 	u_int n, next_cidx;
2351 	struct mbuf *m0, *tail;
2352 	struct txpkts txp;
2353 	struct fw_eth_tx_pkts_wr *wr;	/* any fw WR struct will do */
2354 
2355 	remaining = IDXDIFF(pidx, cidx, r->size);
2356 	MPASS(remaining > 0);	/* Must not be called without work to do. */
2357 	total = 0;
2358 
2359 	TXQ_LOCK(txq);
2360 	if (__predict_false((eq->flags & EQ_ENABLED) == 0)) {
2361 		while (cidx != pidx) {
2362 			m0 = r->items[cidx];
2363 			m_freem(m0);
2364 			if (++cidx == r->size)
2365 				cidx = 0;
2366 		}
2367 		reclaim_tx_descs(txq, 2048);
2368 		total = remaining;
2369 		goto done;
2370 	}
2371 
2372 	/* How many hardware descriptors do we have readily available. */
2373 	if (eq->pidx == eq->cidx)
2374 		available = eq->sidx - 1;
2375 	else
2376 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2377 	dbdiff = IDXDIFF(eq->pidx, eq->dbidx, eq->sidx);
2378 
2379 	while (remaining > 0) {
2380 
2381 		m0 = r->items[cidx];
2382 		M_ASSERTPKTHDR(m0);
2383 		MPASS(m0->m_nextpkt == NULL);
2384 
2385 		if (available < SGE_MAX_WR_NDESC) {
2386 			available += reclaim_tx_descs(txq, 64);
2387 			if (available < howmany(mbuf_len16(m0), EQ_ESIZE / 16))
2388 				break;	/* out of descriptors */
2389 		}
2390 
2391 		next_cidx = cidx + 1;
2392 		if (__predict_false(next_cidx == r->size))
2393 			next_cidx = 0;
2394 
2395 		wr = (void *)&eq->desc[eq->pidx];
2396 		if (remaining > 1 &&
2397 		    try_txpkts(m0, r->items[next_cidx], &txp, available) == 0) {
2398 
2399 			/* pkts at cidx, next_cidx should both be in txp. */
2400 			MPASS(txp.npkt == 2);
2401 			tail = r->items[next_cidx];
2402 			MPASS(tail->m_nextpkt == NULL);
2403 			ETHER_BPF_MTAP(ifp, m0);
2404 			ETHER_BPF_MTAP(ifp, tail);
2405 			m0->m_nextpkt = tail;
2406 
2407 			if (__predict_false(++next_cidx == r->size))
2408 				next_cidx = 0;
2409 
2410 			while (next_cidx != pidx) {
2411 				if (add_to_txpkts(r->items[next_cidx], &txp,
2412 				    available) != 0)
2413 					break;
2414 				tail->m_nextpkt = r->items[next_cidx];
2415 				tail = tail->m_nextpkt;
2416 				ETHER_BPF_MTAP(ifp, tail);
2417 				if (__predict_false(++next_cidx == r->size))
2418 					next_cidx = 0;
2419 			}
2420 
2421 			n = write_txpkts_wr(txq, wr, m0, &txp, available);
2422 			total += txp.npkt;
2423 			remaining -= txp.npkt;
2424 		} else {
2425 			total++;
2426 			remaining--;
2427 			n = write_txpkt_wr(txq, (void *)wr, m0, available);
2428 			ETHER_BPF_MTAP(ifp, m0);
2429 		}
2430 		MPASS(n >= 1 && n <= available && n <= SGE_MAX_WR_NDESC);
2431 
2432 		available -= n;
2433 		dbdiff += n;
2434 		IDXINCR(eq->pidx, n, eq->sidx);
2435 
2436 		if (total_available_tx_desc(eq) < eq->sidx / 4 &&
2437 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2438 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2439 			    F_FW_WR_EQUEQ);
2440 			eq->equeqidx = eq->pidx;
2441 		} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
2442 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
2443 			eq->equeqidx = eq->pidx;
2444 		}
2445 
2446 		if (dbdiff >= 16 && remaining >= 4) {
2447 			ring_eq_db(sc, eq, dbdiff);
2448 			available += reclaim_tx_descs(txq, 4 * dbdiff);
2449 			dbdiff = 0;
2450 		}
2451 
2452 		cidx = next_cidx;
2453 	}
2454 	if (dbdiff != 0) {
2455 		ring_eq_db(sc, eq, dbdiff);
2456 		reclaim_tx_descs(txq, 32);
2457 	}
2458 done:
2459 	TXQ_UNLOCK(txq);
2460 
2461 	return (total);
2462 }
2463 
2464 static inline void
2465 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
2466     int qsize)
2467 {
2468 
2469 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
2470 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
2471 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
2472 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
2473 
2474 	iq->flags = 0;
2475 	iq->adapter = sc;
2476 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
2477 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
2478 	if (pktc_idx >= 0) {
2479 		iq->intr_params |= F_QINTR_CNT_EN;
2480 		iq->intr_pktc_idx = pktc_idx;
2481 	}
2482 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
2483 	iq->sidx = iq->qsize - spg_len / IQ_ESIZE;
2484 }
2485 
2486 static inline void
2487 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name)
2488 {
2489 
2490 	fl->qsize = qsize;
2491 	fl->sidx = qsize - spg_len / EQ_ESIZE;
2492 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
2493 	if (sc->flags & BUF_PACKING_OK &&
2494 	    ((!is_t4(sc) && buffer_packing) ||	/* T5+: enabled unless 0 */
2495 	    (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */
2496 		fl->flags |= FL_BUF_PACKING;
2497 	find_best_refill_source(sc, fl, maxp);
2498 	find_safe_refill_source(sc, fl);
2499 }
2500 
2501 static inline void
2502 init_eq(struct sge_eq *eq, int eqtype, int qsize, uint8_t tx_chan,
2503     uint16_t iqid, char *name)
2504 {
2505 	KASSERT(tx_chan < NCHAN, ("%s: bad tx channel %d", __func__, tx_chan));
2506 	KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
2507 
2508 	eq->flags = eqtype & EQ_TYPEMASK;
2509 	eq->tx_chan = tx_chan;
2510 	eq->iqid = iqid;
2511 	eq->sidx = qsize - spg_len / EQ_ESIZE;
2512 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
2513 }
2514 
2515 static int
2516 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
2517     bus_dmamap_t *map, bus_addr_t *pa, void **va)
2518 {
2519 	int rc;
2520 
2521 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
2522 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
2523 	if (rc != 0) {
2524 		device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
2525 		goto done;
2526 	}
2527 
2528 	rc = bus_dmamem_alloc(*tag, va,
2529 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
2530 	if (rc != 0) {
2531 		device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
2532 		goto done;
2533 	}
2534 
2535 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
2536 	if (rc != 0) {
2537 		device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
2538 		goto done;
2539 	}
2540 done:
2541 	if (rc)
2542 		free_ring(sc, *tag, *map, *pa, *va);
2543 
2544 	return (rc);
2545 }
2546 
2547 static int
2548 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
2549     bus_addr_t pa, void *va)
2550 {
2551 	if (pa)
2552 		bus_dmamap_unload(tag, map);
2553 	if (va)
2554 		bus_dmamem_free(tag, va, map);
2555 	if (tag)
2556 		bus_dma_tag_destroy(tag);
2557 
2558 	return (0);
2559 }
2560 
2561 /*
2562  * Allocates the ring for an ingress queue and an optional freelist.  If the
2563  * freelist is specified it will be allocated and then associated with the
2564  * ingress queue.
2565  *
2566  * Returns errno on failure.  Resources allocated up to that point may still be
2567  * allocated.  Caller is responsible for cleanup in case this function fails.
2568  *
2569  * If the ingress queue will take interrupts directly (iq->flags & IQ_INTR) then
2570  * the intr_idx specifies the vector, starting from 0.  Otherwise it specifies
2571  * the abs_id of the ingress queue to which its interrupts should be forwarded.
2572  */
2573 static int
2574 alloc_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl,
2575     int intr_idx, int cong)
2576 {
2577 	int rc, i, cntxt_id;
2578 	size_t len;
2579 	struct fw_iq_cmd c;
2580 	struct adapter *sc = iq->adapter;
2581 	__be32 v = 0;
2582 
2583 	len = iq->qsize * IQ_ESIZE;
2584 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
2585 	    (void **)&iq->desc);
2586 	if (rc != 0)
2587 		return (rc);
2588 
2589 	bzero(&c, sizeof(c));
2590 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
2591 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
2592 	    V_FW_IQ_CMD_VFN(0));
2593 
2594 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
2595 	    FW_LEN16(c));
2596 
2597 	/* Special handling for firmware event queue */
2598 	if (iq == &sc->sge.fwq)
2599 		v |= F_FW_IQ_CMD_IQASYNCH;
2600 
2601 	if (iq->flags & IQ_INTR) {
2602 		KASSERT(intr_idx < sc->intr_count,
2603 		    ("%s: invalid direct intr_idx %d", __func__, intr_idx));
2604 	} else
2605 		v |= F_FW_IQ_CMD_IQANDST;
2606 	v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
2607 
2608 	c.type_to_iqandstindex = htobe32(v |
2609 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
2610 	    V_FW_IQ_CMD_VIID(pi->viid) |
2611 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
2612 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
2613 	    F_FW_IQ_CMD_IQGTSMODE |
2614 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
2615 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
2616 	c.iqsize = htobe16(iq->qsize);
2617 	c.iqaddr = htobe64(iq->ba);
2618 	if (cong >= 0)
2619 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
2620 
2621 	if (fl) {
2622 		mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
2623 
2624 		len = fl->qsize * EQ_ESIZE;
2625 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
2626 		    &fl->ba, (void **)&fl->desc);
2627 		if (rc)
2628 			return (rc);
2629 
2630 		/* Allocate space for one software descriptor per buffer. */
2631 		rc = alloc_fl_sdesc(fl);
2632 		if (rc != 0) {
2633 			device_printf(sc->dev,
2634 			    "failed to setup fl software descriptors: %d\n",
2635 			    rc);
2636 			return (rc);
2637 		}
2638 
2639 		if (fl->flags & FL_BUF_PACKING) {
2640 			fl->lowat = roundup2(sc->sge.fl_starve_threshold2, 8);
2641 			fl->buf_boundary = sc->sge.pack_boundary;
2642 		} else {
2643 			fl->lowat = roundup2(sc->sge.fl_starve_threshold, 8);
2644 			fl->buf_boundary = 16;
2645 		}
2646 		if (fl_pad && fl->buf_boundary < sc->sge.pad_boundary)
2647 			fl->buf_boundary = sc->sge.pad_boundary;
2648 
2649 		c.iqns_to_fl0congen |=
2650 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
2651 			F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
2652 			(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
2653 			(fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
2654 			    0));
2655 		if (cong >= 0) {
2656 			c.iqns_to_fl0congen |=
2657 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
2658 				    F_FW_IQ_CMD_FL0CONGCIF |
2659 				    F_FW_IQ_CMD_FL0CONGEN);
2660 		}
2661 		c.fl0dcaen_to_fl0cidxfthresh =
2662 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(X_FETCHBURSTMIN_64B) |
2663 			V_FW_IQ_CMD_FL0FBMAX(X_FETCHBURSTMAX_512B));
2664 		c.fl0size = htobe16(fl->qsize);
2665 		c.fl0addr = htobe64(fl->ba);
2666 	}
2667 
2668 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
2669 	if (rc != 0) {
2670 		device_printf(sc->dev,
2671 		    "failed to create ingress queue: %d\n", rc);
2672 		return (rc);
2673 	}
2674 
2675 	iq->cidx = 0;
2676 	iq->gen = F_RSPD_GEN;
2677 	iq->intr_next = iq->intr_params;
2678 	iq->cntxt_id = be16toh(c.iqid);
2679 	iq->abs_id = be16toh(c.physiqid);
2680 	iq->flags |= IQ_ALLOCATED;
2681 
2682 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
2683 	if (cntxt_id >= sc->sge.niq) {
2684 		panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
2685 		    cntxt_id, sc->sge.niq - 1);
2686 	}
2687 	sc->sge.iqmap[cntxt_id] = iq;
2688 
2689 	if (fl) {
2690 		u_int qid;
2691 
2692 		iq->flags |= IQ_HAS_FL;
2693 		fl->cntxt_id = be16toh(c.fl0id);
2694 		fl->pidx = fl->cidx = 0;
2695 
2696 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
2697 		if (cntxt_id >= sc->sge.neq) {
2698 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
2699 			    __func__, cntxt_id, sc->sge.neq - 1);
2700 		}
2701 		sc->sge.eqmap[cntxt_id] = (void *)fl;
2702 
2703 		qid = fl->cntxt_id;
2704 		if (isset(&sc->doorbells, DOORBELL_UDB)) {
2705 			uint32_t s_qpp = sc->sge.eq_s_qpp;
2706 			uint32_t mask = (1 << s_qpp) - 1;
2707 			volatile uint8_t *udb;
2708 
2709 			udb = sc->udbs_base + UDBS_DB_OFFSET;
2710 			udb += (qid >> s_qpp) << PAGE_SHIFT;
2711 			qid &= mask;
2712 			if (qid < PAGE_SIZE / UDBS_SEG_SIZE) {
2713 				udb += qid << UDBS_SEG_SHIFT;
2714 				qid = 0;
2715 			}
2716 			fl->udb = (volatile void *)udb;
2717 		}
2718 		fl->dbval = F_DBPRIO | V_QID(qid);
2719 		if (is_t5(sc))
2720 			fl->dbval |= F_DBTYPE;
2721 
2722 		FL_LOCK(fl);
2723 		/* Enough to make sure the SGE doesn't think it's starved */
2724 		refill_fl(sc, fl, fl->lowat);
2725 		FL_UNLOCK(fl);
2726 	}
2727 
2728 	if (is_t5(sc) && cong >= 0) {
2729 		uint32_t param, val;
2730 
2731 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
2732 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
2733 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
2734 		if (cong == 0)
2735 			val = 1 << 19;
2736 		else {
2737 			val = 2 << 19;
2738 			for (i = 0; i < 4; i++) {
2739 				if (cong & (1 << i))
2740 					val |= 1 << (i << 2);
2741 			}
2742 		}
2743 
2744 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
2745 		if (rc != 0) {
2746 			/* report error but carry on */
2747 			device_printf(sc->dev,
2748 			    "failed to set congestion manager context for "
2749 			    "ingress queue %d: %d\n", iq->cntxt_id, rc);
2750 		}
2751 	}
2752 
2753 	/* Enable IQ interrupts */
2754 	atomic_store_rel_int(&iq->state, IQS_IDLE);
2755 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_SEINTARM(iq->intr_params) |
2756 	    V_INGRESSQID(iq->cntxt_id));
2757 
2758 	return (0);
2759 }
2760 
2761 static int
2762 free_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl)
2763 {
2764 	int rc;
2765 	struct adapter *sc = iq->adapter;
2766 	device_t dev;
2767 
2768 	if (sc == NULL)
2769 		return (0);	/* nothing to do */
2770 
2771 	dev = pi ? pi->dev : sc->dev;
2772 
2773 	if (iq->flags & IQ_ALLOCATED) {
2774 		rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
2775 		    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
2776 		    fl ? fl->cntxt_id : 0xffff, 0xffff);
2777 		if (rc != 0) {
2778 			device_printf(dev,
2779 			    "failed to free queue %p: %d\n", iq, rc);
2780 			return (rc);
2781 		}
2782 		iq->flags &= ~IQ_ALLOCATED;
2783 	}
2784 
2785 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
2786 
2787 	bzero(iq, sizeof(*iq));
2788 
2789 	if (fl) {
2790 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
2791 		    fl->desc);
2792 
2793 		if (fl->sdesc)
2794 			free_fl_sdesc(sc, fl);
2795 
2796 		if (mtx_initialized(&fl->fl_lock))
2797 			mtx_destroy(&fl->fl_lock);
2798 
2799 		bzero(fl, sizeof(*fl));
2800 	}
2801 
2802 	return (0);
2803 }
2804 
2805 static void
2806 add_fl_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
2807     struct sge_fl *fl)
2808 {
2809 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2810 
2811 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
2812 	    "freelist");
2813 	children = SYSCTL_CHILDREN(oid);
2814 
2815 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
2816 	    CTLTYPE_INT | CTLFLAG_RD, &fl->cntxt_id, 0, sysctl_uint16, "I",
2817 	    "SGE context id of the freelist");
2818 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL,
2819 	    fl_pad ? 1 : 0, "padding enabled");
2820 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL,
2821 	    fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled");
2822 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
2823 	    0, "consumer index");
2824 	if (fl->flags & FL_BUF_PACKING) {
2825 		SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
2826 		    CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
2827 	}
2828 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
2829 	    0, "producer index");
2830 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_allocated",
2831 	    CTLFLAG_RD, &fl->mbuf_allocated, "# of mbuf allocated");
2832 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_inlined",
2833 	    CTLFLAG_RD, &fl->mbuf_inlined, "# of mbuf inlined in clusters");
2834 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
2835 	    CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
2836 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
2837 	    CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
2838 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
2839 	    CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
2840 }
2841 
2842 static int
2843 alloc_fwq(struct adapter *sc)
2844 {
2845 	int rc, intr_idx;
2846 	struct sge_iq *fwq = &sc->sge.fwq;
2847 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
2848 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2849 
2850 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE);
2851 	fwq->flags |= IQ_INTR;	/* always */
2852 	intr_idx = sc->intr_count > 1 ? 1 : 0;
2853 	rc = alloc_iq_fl(sc->port[0], fwq, NULL, intr_idx, -1);
2854 	if (rc != 0) {
2855 		device_printf(sc->dev,
2856 		    "failed to create firmware event queue: %d\n", rc);
2857 		return (rc);
2858 	}
2859 
2860 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD,
2861 	    NULL, "firmware event queue");
2862 	children = SYSCTL_CHILDREN(oid);
2863 
2864 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "abs_id",
2865 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->abs_id, 0, sysctl_uint16, "I",
2866 	    "absolute id of the queue");
2867 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cntxt_id",
2868 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->cntxt_id, 0, sysctl_uint16, "I",
2869 	    "SGE context id of the queue");
2870 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cidx",
2871 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->cidx, 0, sysctl_uint16, "I",
2872 	    "consumer index");
2873 
2874 	return (0);
2875 }
2876 
2877 static int
2878 free_fwq(struct adapter *sc)
2879 {
2880 	return free_iq_fl(NULL, &sc->sge.fwq, NULL);
2881 }
2882 
2883 static int
2884 alloc_mgmtq(struct adapter *sc)
2885 {
2886 	int rc;
2887 	struct sge_wrq *mgmtq = &sc->sge.mgmtq;
2888 	char name[16];
2889 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
2890 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2891 
2892 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "mgmtq", CTLFLAG_RD,
2893 	    NULL, "management queue");
2894 
2895 	snprintf(name, sizeof(name), "%s mgmtq", device_get_nameunit(sc->dev));
2896 	init_eq(&mgmtq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[0]->tx_chan,
2897 	    sc->sge.fwq.cntxt_id, name);
2898 	rc = alloc_wrq(sc, NULL, mgmtq, oid);
2899 	if (rc != 0) {
2900 		device_printf(sc->dev,
2901 		    "failed to create management queue: %d\n", rc);
2902 		return (rc);
2903 	}
2904 
2905 	return (0);
2906 }
2907 
2908 static int
2909 free_mgmtq(struct adapter *sc)
2910 {
2911 
2912 	return free_wrq(sc, &sc->sge.mgmtq);
2913 }
2914 
2915 int
2916 tnl_cong(struct port_info *pi)
2917 {
2918 
2919 	if (cong_drop == -1)
2920 		return (-1);
2921 	else if (cong_drop == 1)
2922 		return (0);
2923 	else
2924 		return (pi->rx_chan_map);
2925 }
2926 
2927 static int
2928 alloc_rxq(struct port_info *pi, struct sge_rxq *rxq, int intr_idx, int idx,
2929     struct sysctl_oid *oid)
2930 {
2931 	int rc;
2932 	struct sysctl_oid_list *children;
2933 	char name[16];
2934 
2935 	rc = alloc_iq_fl(pi, &rxq->iq, &rxq->fl, intr_idx, tnl_cong(pi));
2936 	if (rc != 0)
2937 		return (rc);
2938 
2939 	/*
2940 	 * The freelist is just barely above the starvation threshold right now,
2941 	 * fill it up a bit more.
2942 	 */
2943 	FL_LOCK(&rxq->fl);
2944 	refill_fl(pi->adapter, &rxq->fl, 128);
2945 	FL_UNLOCK(&rxq->fl);
2946 
2947 #if defined(INET) || defined(INET6)
2948 	rc = tcp_lro_init(&rxq->lro);
2949 	if (rc != 0)
2950 		return (rc);
2951 	rxq->lro.ifp = pi->ifp; /* also indicates LRO init'ed */
2952 
2953 	if (pi->ifp->if_capenable & IFCAP_LRO)
2954 		rxq->iq.flags |= IQ_LRO_ENABLED;
2955 #endif
2956 	rxq->ifp = pi->ifp;
2957 
2958 	children = SYSCTL_CHILDREN(oid);
2959 
2960 	snprintf(name, sizeof(name), "%d", idx);
2961 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
2962 	    NULL, "rx queue");
2963 	children = SYSCTL_CHILDREN(oid);
2964 
2965 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "abs_id",
2966 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.abs_id, 0, sysctl_uint16, "I",
2967 	    "absolute id of the queue");
2968 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cntxt_id",
2969 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cntxt_id, 0, sysctl_uint16, "I",
2970 	    "SGE context id of the queue");
2971 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cidx",
2972 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cidx, 0, sysctl_uint16, "I",
2973 	    "consumer index");
2974 #if defined(INET) || defined(INET6)
2975 	SYSCTL_ADD_INT(&pi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
2976 	    &rxq->lro.lro_queued, 0, NULL);
2977 	SYSCTL_ADD_INT(&pi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
2978 	    &rxq->lro.lro_flushed, 0, NULL);
2979 #endif
2980 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
2981 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
2982 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "vlan_extraction",
2983 	    CTLFLAG_RD, &rxq->vlan_extraction,
2984 	    "# of times hardware extracted 802.1Q tag");
2985 
2986 	add_fl_sysctls(&pi->ctx, oid, &rxq->fl);
2987 
2988 	return (rc);
2989 }
2990 
2991 static int
2992 free_rxq(struct port_info *pi, struct sge_rxq *rxq)
2993 {
2994 	int rc;
2995 
2996 #if defined(INET) || defined(INET6)
2997 	if (rxq->lro.ifp) {
2998 		tcp_lro_free(&rxq->lro);
2999 		rxq->lro.ifp = NULL;
3000 	}
3001 #endif
3002 
3003 	rc = free_iq_fl(pi, &rxq->iq, &rxq->fl);
3004 	if (rc == 0)
3005 		bzero(rxq, sizeof(*rxq));
3006 
3007 	return (rc);
3008 }
3009 
3010 #ifdef TCP_OFFLOAD
3011 static int
3012 alloc_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq,
3013     int intr_idx, int idx, struct sysctl_oid *oid)
3014 {
3015 	int rc;
3016 	struct sysctl_oid_list *children;
3017 	char name[16];
3018 
3019 	rc = alloc_iq_fl(pi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx,
3020 	    pi->rx_chan_map);
3021 	if (rc != 0)
3022 		return (rc);
3023 
3024 	children = SYSCTL_CHILDREN(oid);
3025 
3026 	snprintf(name, sizeof(name), "%d", idx);
3027 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3028 	    NULL, "rx queue");
3029 	children = SYSCTL_CHILDREN(oid);
3030 
3031 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "abs_id",
3032 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.abs_id, 0, sysctl_uint16,
3033 	    "I", "absolute id of the queue");
3034 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cntxt_id",
3035 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cntxt_id, 0, sysctl_uint16,
3036 	    "I", "SGE context id of the queue");
3037 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cidx",
3038 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cidx, 0, sysctl_uint16, "I",
3039 	    "consumer index");
3040 
3041 	add_fl_sysctls(&pi->ctx, oid, &ofld_rxq->fl);
3042 
3043 	return (rc);
3044 }
3045 
3046 static int
3047 free_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq)
3048 {
3049 	int rc;
3050 
3051 	rc = free_iq_fl(pi, &ofld_rxq->iq, &ofld_rxq->fl);
3052 	if (rc == 0)
3053 		bzero(ofld_rxq, sizeof(*ofld_rxq));
3054 
3055 	return (rc);
3056 }
3057 #endif
3058 
3059 #ifdef DEV_NETMAP
3060 static int
3061 alloc_nm_rxq(struct port_info *pi, struct sge_nm_rxq *nm_rxq, int intr_idx,
3062     int idx, struct sysctl_oid *oid)
3063 {
3064 	int rc;
3065 	struct sysctl_oid_list *children;
3066 	struct sysctl_ctx_list *ctx;
3067 	char name[16];
3068 	size_t len;
3069 	struct adapter *sc = pi->adapter;
3070 	struct netmap_adapter *na = NA(pi->nm_ifp);
3071 
3072 	MPASS(na != NULL);
3073 
3074 	len = pi->qsize_rxq * IQ_ESIZE;
3075 	rc = alloc_ring(sc, len, &nm_rxq->iq_desc_tag, &nm_rxq->iq_desc_map,
3076 	    &nm_rxq->iq_ba, (void **)&nm_rxq->iq_desc);
3077 	if (rc != 0)
3078 		return (rc);
3079 
3080 	len = na->num_rx_desc * EQ_ESIZE + spg_len;
3081 	rc = alloc_ring(sc, len, &nm_rxq->fl_desc_tag, &nm_rxq->fl_desc_map,
3082 	    &nm_rxq->fl_ba, (void **)&nm_rxq->fl_desc);
3083 	if (rc != 0)
3084 		return (rc);
3085 
3086 	nm_rxq->pi = pi;
3087 	nm_rxq->nid = idx;
3088 	nm_rxq->iq_cidx = 0;
3089 	nm_rxq->iq_sidx = pi->qsize_rxq - spg_len / IQ_ESIZE;
3090 	nm_rxq->iq_gen = F_RSPD_GEN;
3091 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
3092 	nm_rxq->fl_sidx = na->num_rx_desc;
3093 	nm_rxq->intr_idx = intr_idx;
3094 
3095 	ctx = &pi->ctx;
3096 	children = SYSCTL_CHILDREN(oid);
3097 
3098 	snprintf(name, sizeof(name), "%d", idx);
3099 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL,
3100 	    "rx queue");
3101 	children = SYSCTL_CHILDREN(oid);
3102 
3103 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3104 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_abs_id, 0, sysctl_uint16,
3105 	    "I", "absolute id of the queue");
3106 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3107 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cntxt_id, 0, sysctl_uint16,
3108 	    "I", "SGE context id of the queue");
3109 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3110 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cidx, 0, sysctl_uint16, "I",
3111 	    "consumer index");
3112 
3113 	children = SYSCTL_CHILDREN(oid);
3114 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3115 	    "freelist");
3116 	children = SYSCTL_CHILDREN(oid);
3117 
3118 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3119 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->fl_cntxt_id, 0, sysctl_uint16,
3120 	    "I", "SGE context id of the freelist");
3121 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
3122 	    &nm_rxq->fl_cidx, 0, "consumer index");
3123 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
3124 	    &nm_rxq->fl_pidx, 0, "producer index");
3125 
3126 	return (rc);
3127 }
3128 
3129 
3130 static int
3131 free_nm_rxq(struct port_info *pi, struct sge_nm_rxq *nm_rxq)
3132 {
3133 	struct adapter *sc = pi->adapter;
3134 
3135 	free_ring(sc, nm_rxq->iq_desc_tag, nm_rxq->iq_desc_map, nm_rxq->iq_ba,
3136 	    nm_rxq->iq_desc);
3137 	free_ring(sc, nm_rxq->fl_desc_tag, nm_rxq->fl_desc_map, nm_rxq->fl_ba,
3138 	    nm_rxq->fl_desc);
3139 
3140 	return (0);
3141 }
3142 
3143 static int
3144 alloc_nm_txq(struct port_info *pi, struct sge_nm_txq *nm_txq, int iqidx, int idx,
3145     struct sysctl_oid *oid)
3146 {
3147 	int rc;
3148 	size_t len;
3149 	struct adapter *sc = pi->adapter;
3150 	struct netmap_adapter *na = NA(pi->nm_ifp);
3151 	char name[16];
3152 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3153 
3154 	len = na->num_tx_desc * EQ_ESIZE + spg_len;
3155 	rc = alloc_ring(sc, len, &nm_txq->desc_tag, &nm_txq->desc_map,
3156 	    &nm_txq->ba, (void **)&nm_txq->desc);
3157 	if (rc)
3158 		return (rc);
3159 
3160 	nm_txq->pidx = nm_txq->cidx = 0;
3161 	nm_txq->sidx = na->num_tx_desc;
3162 	nm_txq->nid = idx;
3163 	nm_txq->iqidx = iqidx;
3164 	nm_txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3165 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(sc->pf));
3166 
3167 	snprintf(name, sizeof(name), "%d", idx);
3168 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3169 	    NULL, "netmap tx queue");
3170 	children = SYSCTL_CHILDREN(oid);
3171 
3172 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3173 	    &nm_txq->cntxt_id, 0, "SGE context id of the queue");
3174 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cidx",
3175 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->cidx, 0, sysctl_uint16, "I",
3176 	    "consumer index");
3177 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "pidx",
3178 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->pidx, 0, sysctl_uint16, "I",
3179 	    "producer index");
3180 
3181 	return (rc);
3182 }
3183 
3184 static int
3185 free_nm_txq(struct port_info *pi, struct sge_nm_txq *nm_txq)
3186 {
3187 	struct adapter *sc = pi->adapter;
3188 
3189 	free_ring(sc, nm_txq->desc_tag, nm_txq->desc_map, nm_txq->ba,
3190 	    nm_txq->desc);
3191 
3192 	return (0);
3193 }
3194 #endif
3195 
3196 static int
3197 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
3198 {
3199 	int rc, cntxt_id;
3200 	struct fw_eq_ctrl_cmd c;
3201 	int qsize = eq->sidx + spg_len / EQ_ESIZE;
3202 
3203 	bzero(&c, sizeof(c));
3204 
3205 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
3206 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
3207 	    V_FW_EQ_CTRL_CMD_VFN(0));
3208 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
3209 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
3210 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid));
3211 	c.physeqid_pkd = htobe32(0);
3212 	c.fetchszm_to_iqid =
3213 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3214 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
3215 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
3216 	c.dcaen_to_eqsize =
3217 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3218 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3219 		V_FW_EQ_CTRL_CMD_EQSIZE(qsize));
3220 	c.eqaddr = htobe64(eq->ba);
3221 
3222 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3223 	if (rc != 0) {
3224 		device_printf(sc->dev,
3225 		    "failed to create control queue %d: %d\n", eq->tx_chan, rc);
3226 		return (rc);
3227 	}
3228 	eq->flags |= EQ_ALLOCATED;
3229 
3230 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
3231 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3232 	if (cntxt_id >= sc->sge.neq)
3233 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3234 		cntxt_id, sc->sge.neq - 1);
3235 	sc->sge.eqmap[cntxt_id] = eq;
3236 
3237 	return (rc);
3238 }
3239 
3240 static int
3241 eth_eq_alloc(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
3242 {
3243 	int rc, cntxt_id;
3244 	struct fw_eq_eth_cmd c;
3245 	int qsize = eq->sidx + spg_len / EQ_ESIZE;
3246 
3247 	bzero(&c, sizeof(c));
3248 
3249 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
3250 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
3251 	    V_FW_EQ_ETH_CMD_VFN(0));
3252 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
3253 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
3254 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
3255 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(pi->viid));
3256 	c.fetchszm_to_iqid =
3257 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3258 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
3259 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
3260 	c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3261 	    V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3262 	    V_FW_EQ_ETH_CMD_EQSIZE(qsize));
3263 	c.eqaddr = htobe64(eq->ba);
3264 
3265 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3266 	if (rc != 0) {
3267 		device_printf(pi->dev,
3268 		    "failed to create Ethernet egress queue: %d\n", rc);
3269 		return (rc);
3270 	}
3271 	eq->flags |= EQ_ALLOCATED;
3272 
3273 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
3274 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3275 	if (cntxt_id >= sc->sge.neq)
3276 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3277 		cntxt_id, sc->sge.neq - 1);
3278 	sc->sge.eqmap[cntxt_id] = eq;
3279 
3280 	return (rc);
3281 }
3282 
3283 #ifdef TCP_OFFLOAD
3284 static int
3285 ofld_eq_alloc(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
3286 {
3287 	int rc, cntxt_id;
3288 	struct fw_eq_ofld_cmd c;
3289 	int qsize = eq->sidx + spg_len / EQ_ESIZE;
3290 
3291 	bzero(&c, sizeof(c));
3292 
3293 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
3294 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
3295 	    V_FW_EQ_OFLD_CMD_VFN(0));
3296 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
3297 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
3298 	c.fetchszm_to_iqid =
3299 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3300 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
3301 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
3302 	c.dcaen_to_eqsize =
3303 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3304 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3305 		V_FW_EQ_OFLD_CMD_EQSIZE(qsize));
3306 	c.eqaddr = htobe64(eq->ba);
3307 
3308 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3309 	if (rc != 0) {
3310 		device_printf(pi->dev,
3311 		    "failed to create egress queue for TCP offload: %d\n", rc);
3312 		return (rc);
3313 	}
3314 	eq->flags |= EQ_ALLOCATED;
3315 
3316 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
3317 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3318 	if (cntxt_id >= sc->sge.neq)
3319 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3320 		cntxt_id, sc->sge.neq - 1);
3321 	sc->sge.eqmap[cntxt_id] = eq;
3322 
3323 	return (rc);
3324 }
3325 #endif
3326 
3327 static int
3328 alloc_eq(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
3329 {
3330 	int rc, qsize;
3331 	size_t len;
3332 
3333 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
3334 
3335 	qsize = eq->sidx + spg_len / EQ_ESIZE;
3336 	len = qsize * EQ_ESIZE;
3337 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
3338 	    &eq->ba, (void **)&eq->desc);
3339 	if (rc)
3340 		return (rc);
3341 
3342 	eq->pidx = eq->cidx = 0;
3343 	eq->equeqidx = eq->dbidx = 0;
3344 	eq->doorbells = sc->doorbells;
3345 
3346 	switch (eq->flags & EQ_TYPEMASK) {
3347 	case EQ_CTRL:
3348 		rc = ctrl_eq_alloc(sc, eq);
3349 		break;
3350 
3351 	case EQ_ETH:
3352 		rc = eth_eq_alloc(sc, pi, eq);
3353 		break;
3354 
3355 #ifdef TCP_OFFLOAD
3356 	case EQ_OFLD:
3357 		rc = ofld_eq_alloc(sc, pi, eq);
3358 		break;
3359 #endif
3360 
3361 	default:
3362 		panic("%s: invalid eq type %d.", __func__,
3363 		    eq->flags & EQ_TYPEMASK);
3364 	}
3365 	if (rc != 0) {
3366 		device_printf(sc->dev,
3367 		    "failed to allocate egress queue(%d): %d\n",
3368 		    eq->flags & EQ_TYPEMASK, rc);
3369 	}
3370 
3371 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
3372 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
3373 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
3374 		uint32_t s_qpp = sc->sge.eq_s_qpp;
3375 		uint32_t mask = (1 << s_qpp) - 1;
3376 		volatile uint8_t *udb;
3377 
3378 		udb = sc->udbs_base + UDBS_DB_OFFSET;
3379 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
3380 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
3381 		if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
3382 	    		clrbit(&eq->doorbells, DOORBELL_WCWR);
3383 		else {
3384 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
3385 			eq->udb_qid = 0;
3386 		}
3387 		eq->udb = (volatile void *)udb;
3388 	}
3389 
3390 	return (rc);
3391 }
3392 
3393 static int
3394 free_eq(struct adapter *sc, struct sge_eq *eq)
3395 {
3396 	int rc;
3397 
3398 	if (eq->flags & EQ_ALLOCATED) {
3399 		switch (eq->flags & EQ_TYPEMASK) {
3400 		case EQ_CTRL:
3401 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
3402 			    eq->cntxt_id);
3403 			break;
3404 
3405 		case EQ_ETH:
3406 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
3407 			    eq->cntxt_id);
3408 			break;
3409 
3410 #ifdef TCP_OFFLOAD
3411 		case EQ_OFLD:
3412 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
3413 			    eq->cntxt_id);
3414 			break;
3415 #endif
3416 
3417 		default:
3418 			panic("%s: invalid eq type %d.", __func__,
3419 			    eq->flags & EQ_TYPEMASK);
3420 		}
3421 		if (rc != 0) {
3422 			device_printf(sc->dev,
3423 			    "failed to free egress queue (%d): %d\n",
3424 			    eq->flags & EQ_TYPEMASK, rc);
3425 			return (rc);
3426 		}
3427 		eq->flags &= ~EQ_ALLOCATED;
3428 	}
3429 
3430 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
3431 
3432 	if (mtx_initialized(&eq->eq_lock))
3433 		mtx_destroy(&eq->eq_lock);
3434 
3435 	bzero(eq, sizeof(*eq));
3436 	return (0);
3437 }
3438 
3439 static int
3440 alloc_wrq(struct adapter *sc, struct port_info *pi, struct sge_wrq *wrq,
3441     struct sysctl_oid *oid)
3442 {
3443 	int rc;
3444 	struct sysctl_ctx_list *ctx = pi ? &pi->ctx : &sc->ctx;
3445 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3446 
3447 	rc = alloc_eq(sc, pi, &wrq->eq);
3448 	if (rc)
3449 		return (rc);
3450 
3451 	wrq->adapter = sc;
3452 	TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq);
3453 	TAILQ_INIT(&wrq->incomplete_wrs);
3454 	STAILQ_INIT(&wrq->wr_list);
3455 	wrq->nwr_pending = 0;
3456 	wrq->ndesc_needed = 0;
3457 
3458 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3459 	    &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
3460 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3461 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I",
3462 	    "consumer index");
3463 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx",
3464 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I",
3465 	    "producer index");
3466 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD,
3467 	    &wrq->tx_wrs_direct, "# of work requests (direct)");
3468 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD,
3469 	    &wrq->tx_wrs_copied, "# of work requests (copied)");
3470 
3471 	return (rc);
3472 }
3473 
3474 static int
3475 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
3476 {
3477 	int rc;
3478 
3479 	rc = free_eq(sc, &wrq->eq);
3480 	if (rc)
3481 		return (rc);
3482 
3483 	bzero(wrq, sizeof(*wrq));
3484 	return (0);
3485 }
3486 
3487 static int
3488 alloc_txq(struct port_info *pi, struct sge_txq *txq, int idx,
3489     struct sysctl_oid *oid)
3490 {
3491 	int rc;
3492 	struct adapter *sc = pi->adapter;
3493 	struct sge_eq *eq = &txq->eq;
3494 	char name[16];
3495 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3496 
3497 	rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx, can_resume_eth_tx,
3498 	    M_CXGBE, M_WAITOK);
3499 	if (rc != 0) {
3500 		device_printf(sc->dev, "failed to allocate mp_ring: %d\n", rc);
3501 		return (rc);
3502 	}
3503 
3504 	rc = alloc_eq(sc, pi, eq);
3505 	if (rc != 0) {
3506 		mp_ring_free(txq->r);
3507 		txq->r = NULL;
3508 		return (rc);
3509 	}
3510 
3511 	/* Can't fail after this point. */
3512 
3513 	TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq);
3514 	txq->ifp = pi->ifp;
3515 	txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK);
3516 	txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3517 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(sc->pf));
3518 	txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE,
3519 	    M_ZERO | M_WAITOK);
3520 
3521 	snprintf(name, sizeof(name), "%d", idx);
3522 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3523 	    NULL, "tx queue");
3524 	children = SYSCTL_CHILDREN(oid);
3525 
3526 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3527 	    &eq->cntxt_id, 0, "SGE context id of the queue");
3528 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cidx",
3529 	    CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I",
3530 	    "consumer index");
3531 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "pidx",
3532 	    CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I",
3533 	    "producer index");
3534 
3535 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
3536 	    &txq->txcsum, "# of times hardware assisted with checksum");
3537 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "vlan_insertion",
3538 	    CTLFLAG_RD, &txq->vlan_insertion,
3539 	    "# of times hardware inserted 802.1Q tag");
3540 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
3541 	    &txq->tso_wrs, "# of TSO work requests");
3542 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
3543 	    &txq->imm_wrs, "# of work requests with immediate data");
3544 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
3545 	    &txq->sgl_wrs, "# of work requests with direct SGL");
3546 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
3547 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
3548 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txpkts0_wrs",
3549 	    CTLFLAG_RD, &txq->txpkts0_wrs,
3550 	    "# of txpkts (type 0) work requests");
3551 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txpkts1_wrs",
3552 	    CTLFLAG_RD, &txq->txpkts1_wrs,
3553 	    "# of txpkts (type 1) work requests");
3554 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txpkts0_pkts",
3555 	    CTLFLAG_RD, &txq->txpkts0_pkts,
3556 	    "# of frames tx'd using type0 txpkts work requests");
3557 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txpkts1_pkts",
3558 	    CTLFLAG_RD, &txq->txpkts1_pkts,
3559 	    "# of frames tx'd using type1 txpkts work requests");
3560 
3561 	SYSCTL_ADD_COUNTER_U64(&pi->ctx, children, OID_AUTO, "r_enqueues",
3562 	    CTLFLAG_RD, &txq->r->enqueues,
3563 	    "# of enqueues to the mp_ring for this queue");
3564 	SYSCTL_ADD_COUNTER_U64(&pi->ctx, children, OID_AUTO, "r_drops",
3565 	    CTLFLAG_RD, &txq->r->drops,
3566 	    "# of drops in the mp_ring for this queue");
3567 	SYSCTL_ADD_COUNTER_U64(&pi->ctx, children, OID_AUTO, "r_starts",
3568 	    CTLFLAG_RD, &txq->r->starts,
3569 	    "# of normal consumer starts in the mp_ring for this queue");
3570 	SYSCTL_ADD_COUNTER_U64(&pi->ctx, children, OID_AUTO, "r_stalls",
3571 	    CTLFLAG_RD, &txq->r->stalls,
3572 	    "# of consumer stalls in the mp_ring for this queue");
3573 	SYSCTL_ADD_COUNTER_U64(&pi->ctx, children, OID_AUTO, "r_restarts",
3574 	    CTLFLAG_RD, &txq->r->restarts,
3575 	    "# of consumer restarts in the mp_ring for this queue");
3576 	SYSCTL_ADD_COUNTER_U64(&pi->ctx, children, OID_AUTO, "r_abdications",
3577 	    CTLFLAG_RD, &txq->r->abdications,
3578 	    "# of consumer abdications in the mp_ring for this queue");
3579 
3580 	return (0);
3581 }
3582 
3583 static int
3584 free_txq(struct port_info *pi, struct sge_txq *txq)
3585 {
3586 	int rc;
3587 	struct adapter *sc = pi->adapter;
3588 	struct sge_eq *eq = &txq->eq;
3589 
3590 	rc = free_eq(sc, eq);
3591 	if (rc)
3592 		return (rc);
3593 
3594 	sglist_free(txq->gl);
3595 	free(txq->sdesc, M_CXGBE);
3596 	mp_ring_free(txq->r);
3597 
3598 	bzero(txq, sizeof(*txq));
3599 	return (0);
3600 }
3601 
3602 static void
3603 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
3604 {
3605 	bus_addr_t *ba = arg;
3606 
3607 	KASSERT(nseg == 1,
3608 	    ("%s meant for single segment mappings only.", __func__));
3609 
3610 	*ba = error ? 0 : segs->ds_addr;
3611 }
3612 
3613 static inline void
3614 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
3615 {
3616 	uint32_t n, v;
3617 
3618 	n = IDXDIFF(fl->pidx / 8, fl->dbidx, fl->sidx);
3619 	MPASS(n > 0);
3620 
3621 	wmb();
3622 	v = fl->dbval | V_PIDX(n);
3623 	if (fl->udb)
3624 		*fl->udb = htole32(v);
3625 	else
3626 		t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL), v);
3627 	IDXINCR(fl->dbidx, n, fl->sidx);
3628 }
3629 
3630 /*
3631  * Fills up the freelist by allocating upto 'n' buffers.  Buffers that are
3632  * recycled do not count towards this allocation budget.
3633  *
3634  * Returns non-zero to indicate that this freelist should be added to the list
3635  * of starving freelists.
3636  */
3637 static int
3638 refill_fl(struct adapter *sc, struct sge_fl *fl, int n)
3639 {
3640 	__be64 *d;
3641 	struct fl_sdesc *sd;
3642 	uintptr_t pa;
3643 	caddr_t cl;
3644 	struct cluster_layout *cll;
3645 	struct sw_zone_info *swz;
3646 	struct cluster_metadata *clm;
3647 	uint16_t max_pidx;
3648 	uint16_t hw_cidx = fl->hw_cidx;		/* stable snapshot */
3649 
3650 	FL_LOCK_ASSERT_OWNED(fl);
3651 
3652 	/*
3653 	 * We always stop at the begining of the hardware descriptor that's just
3654 	 * before the one with the hw cidx.  This is to avoid hw pidx = hw cidx,
3655 	 * which would mean an empty freelist to the chip.
3656 	 */
3657 	max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1;
3658 	if (fl->pidx == max_pidx * 8)
3659 		return (0);
3660 
3661 	d = &fl->desc[fl->pidx];
3662 	sd = &fl->sdesc[fl->pidx];
3663 	cll = &fl->cll_def;	/* default layout */
3664 	swz = &sc->sge.sw_zone_info[cll->zidx];
3665 
3666 	while (n > 0) {
3667 
3668 		if (sd->cl != NULL) {
3669 
3670 			if (sd->nmbuf == 0) {
3671 				/*
3672 				 * Fast recycle without involving any atomics on
3673 				 * the cluster's metadata (if the cluster has
3674 				 * metadata).  This happens when all frames
3675 				 * received in the cluster were small enough to
3676 				 * fit within a single mbuf each.
3677 				 */
3678 				fl->cl_fast_recycled++;
3679 #ifdef INVARIANTS
3680 				clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
3681 				if (clm != NULL)
3682 					MPASS(clm->refcount == 1);
3683 #endif
3684 				goto recycled_fast;
3685 			}
3686 
3687 			/*
3688 			 * Cluster is guaranteed to have metadata.  Clusters
3689 			 * without metadata always take the fast recycle path
3690 			 * when they're recycled.
3691 			 */
3692 			clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
3693 			MPASS(clm != NULL);
3694 
3695 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
3696 				fl->cl_recycled++;
3697 				counter_u64_add(extfree_rels, 1);
3698 				goto recycled;
3699 			}
3700 			sd->cl = NULL;	/* gave up my reference */
3701 		}
3702 		MPASS(sd->cl == NULL);
3703 alloc:
3704 		cl = uma_zalloc(swz->zone, M_NOWAIT);
3705 		if (__predict_false(cl == NULL)) {
3706 			if (cll == &fl->cll_alt || fl->cll_alt.zidx == -1 ||
3707 			    fl->cll_def.zidx == fl->cll_alt.zidx)
3708 				break;
3709 
3710 			/* fall back to the safe zone */
3711 			cll = &fl->cll_alt;
3712 			swz = &sc->sge.sw_zone_info[cll->zidx];
3713 			goto alloc;
3714 		}
3715 		fl->cl_allocated++;
3716 		n--;
3717 
3718 		pa = pmap_kextract((vm_offset_t)cl);
3719 		pa += cll->region1;
3720 		sd->cl = cl;
3721 		sd->cll = *cll;
3722 		*d = htobe64(pa | cll->hwidx);
3723 		clm = cl_metadata(sc, fl, cll, cl);
3724 		if (clm != NULL) {
3725 recycled:
3726 #ifdef INVARIANTS
3727 			clm->sd = sd;
3728 #endif
3729 			clm->refcount = 1;
3730 		}
3731 		sd->nmbuf = 0;
3732 recycled_fast:
3733 		d++;
3734 		sd++;
3735 		if (__predict_false(++fl->pidx % 8 == 0)) {
3736 			uint16_t pidx = fl->pidx / 8;
3737 
3738 			if (__predict_false(pidx == fl->sidx)) {
3739 				fl->pidx = 0;
3740 				pidx = 0;
3741 				sd = fl->sdesc;
3742 				d = fl->desc;
3743 			}
3744 			if (pidx == max_pidx)
3745 				break;
3746 
3747 			if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4)
3748 				ring_fl_db(sc, fl);
3749 		}
3750 	}
3751 
3752 	if (fl->pidx / 8 != fl->dbidx)
3753 		ring_fl_db(sc, fl);
3754 
3755 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
3756 }
3757 
3758 /*
3759  * Attempt to refill all starving freelists.
3760  */
3761 static void
3762 refill_sfl(void *arg)
3763 {
3764 	struct adapter *sc = arg;
3765 	struct sge_fl *fl, *fl_temp;
3766 
3767 	mtx_lock(&sc->sfl_lock);
3768 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
3769 		FL_LOCK(fl);
3770 		refill_fl(sc, fl, 64);
3771 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
3772 			TAILQ_REMOVE(&sc->sfl, fl, link);
3773 			fl->flags &= ~FL_STARVING;
3774 		}
3775 		FL_UNLOCK(fl);
3776 	}
3777 
3778 	if (!TAILQ_EMPTY(&sc->sfl))
3779 		callout_schedule(&sc->sfl_callout, hz / 5);
3780 	mtx_unlock(&sc->sfl_lock);
3781 }
3782 
3783 static int
3784 alloc_fl_sdesc(struct sge_fl *fl)
3785 {
3786 
3787 	fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc), M_CXGBE,
3788 	    M_ZERO | M_WAITOK);
3789 
3790 	return (0);
3791 }
3792 
3793 static void
3794 free_fl_sdesc(struct adapter *sc, struct sge_fl *fl)
3795 {
3796 	struct fl_sdesc *sd;
3797 	struct cluster_metadata *clm;
3798 	struct cluster_layout *cll;
3799 	int i;
3800 
3801 	sd = fl->sdesc;
3802 	for (i = 0; i < fl->sidx * 8; i++, sd++) {
3803 		if (sd->cl == NULL)
3804 			continue;
3805 
3806 		cll = &sd->cll;
3807 		clm = cl_metadata(sc, fl, cll, sd->cl);
3808 		if (sd->nmbuf == 0)
3809 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
3810 		else if (clm && atomic_fetchadd_int(&clm->refcount, -1) == 1) {
3811 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
3812 			counter_u64_add(extfree_rels, 1);
3813 		}
3814 		sd->cl = NULL;
3815 	}
3816 
3817 	free(fl->sdesc, M_CXGBE);
3818 	fl->sdesc = NULL;
3819 }
3820 
3821 static inline void
3822 get_pkt_gl(struct mbuf *m, struct sglist *gl)
3823 {
3824 	int rc;
3825 
3826 	M_ASSERTPKTHDR(m);
3827 
3828 	sglist_reset(gl);
3829 	rc = sglist_append_mbuf(gl, m);
3830 	if (__predict_false(rc != 0)) {
3831 		panic("%s: mbuf %p (%d segs) was vetted earlier but now fails "
3832 		    "with %d.", __func__, m, mbuf_nsegs(m), rc);
3833 	}
3834 
3835 	KASSERT(gl->sg_nseg == mbuf_nsegs(m),
3836 	    ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m,
3837 	    mbuf_nsegs(m), gl->sg_nseg));
3838 	KASSERT(gl->sg_nseg > 0 &&
3839 	    gl->sg_nseg <= (needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS),
3840 	    ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__,
3841 		gl->sg_nseg, needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS));
3842 }
3843 
3844 /*
3845  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
3846  */
3847 static inline u_int
3848 txpkt_len16(u_int nsegs, u_int tso)
3849 {
3850 	u_int n;
3851 
3852 	MPASS(nsegs > 0);
3853 
3854 	nsegs--; /* first segment is part of ulptx_sgl */
3855 	n = sizeof(struct fw_eth_tx_pkt_wr) + sizeof(struct cpl_tx_pkt_core) +
3856 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
3857 	if (tso)
3858 		n += sizeof(struct cpl_tx_pkt_lso_core);
3859 
3860 	return (howmany(n, 16));
3861 }
3862 
3863 /*
3864  * len16 for a txpkts type 0 WR with a GL.  Does not include the firmware work
3865  * request header.
3866  */
3867 static inline u_int
3868 txpkts0_len16(u_int nsegs)
3869 {
3870 	u_int n;
3871 
3872 	MPASS(nsegs > 0);
3873 
3874 	nsegs--; /* first segment is part of ulptx_sgl */
3875 	n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) +
3876 	    sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) +
3877 	    8 * ((3 * nsegs) / 2 + (nsegs & 1));
3878 
3879 	return (howmany(n, 16));
3880 }
3881 
3882 /*
3883  * len16 for a txpkts type 1 WR with a GL.  Does not include the firmware work
3884  * request header.
3885  */
3886 static inline u_int
3887 txpkts1_len16(void)
3888 {
3889 	u_int n;
3890 
3891 	n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl);
3892 
3893 	return (howmany(n, 16));
3894 }
3895 
3896 static inline u_int
3897 imm_payload(u_int ndesc)
3898 {
3899 	u_int n;
3900 
3901 	n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) -
3902 	    sizeof(struct cpl_tx_pkt_core);
3903 
3904 	return (n);
3905 }
3906 
3907 /*
3908  * Write a txpkt WR for this packet to the hardware descriptors, update the
3909  * software descriptor, and advance the pidx.  It is guaranteed that enough
3910  * descriptors are available.
3911  *
3912  * The return value is the # of hardware descriptors used.
3913  */
3914 static u_int
3915 write_txpkt_wr(struct sge_txq *txq, struct fw_eth_tx_pkt_wr *wr,
3916     struct mbuf *m0, u_int available)
3917 {
3918 	struct sge_eq *eq = &txq->eq;
3919 	struct tx_sdesc *txsd;
3920 	struct cpl_tx_pkt_core *cpl;
3921 	uint32_t ctrl;	/* used in many unrelated places */
3922 	uint64_t ctrl1;
3923 	int len16, ndesc, pktlen, nsegs;
3924 	caddr_t dst;
3925 
3926 	TXQ_LOCK_ASSERT_OWNED(txq);
3927 	M_ASSERTPKTHDR(m0);
3928 	MPASS(available > 0 && available < eq->sidx);
3929 
3930 	len16 = mbuf_len16(m0);
3931 	nsegs = mbuf_nsegs(m0);
3932 	pktlen = m0->m_pkthdr.len;
3933 	ctrl = sizeof(struct cpl_tx_pkt_core);
3934 	if (needs_tso(m0))
3935 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
3936 	else if (pktlen <= imm_payload(2) && available >= 2) {
3937 		/* Immediate data.  Recalculate len16 and set nsegs to 0. */
3938 		ctrl += pktlen;
3939 		len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) +
3940 		    sizeof(struct cpl_tx_pkt_core) + pktlen, 16);
3941 		nsegs = 0;
3942 	}
3943 	ndesc = howmany(len16, EQ_ESIZE / 16);
3944 	MPASS(ndesc <= available);
3945 
3946 	/* Firmware work request header */
3947 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
3948 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
3949 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
3950 
3951 	ctrl = V_FW_WR_LEN16(len16);
3952 	wr->equiq_to_len16 = htobe32(ctrl);
3953 	wr->r3 = 0;
3954 
3955 	if (needs_tso(m0)) {
3956 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
3957 
3958 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
3959 		    m0->m_pkthdr.l4hlen > 0,
3960 		    ("%s: mbuf %p needs TSO but missing header lengths",
3961 			__func__, m0));
3962 
3963 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
3964 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
3965 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
3966 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
3967 			ctrl |= V_LSO_ETHHDR_LEN(1);
3968 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
3969 			ctrl |= F_LSO_IPV6;
3970 
3971 		lso->lso_ctrl = htobe32(ctrl);
3972 		lso->ipid_ofst = htobe16(0);
3973 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
3974 		lso->seqno_offset = htobe32(0);
3975 		lso->len = htobe32(pktlen);
3976 
3977 		cpl = (void *)(lso + 1);
3978 
3979 		txq->tso_wrs++;
3980 	} else
3981 		cpl = (void *)(wr + 1);
3982 
3983 	/* Checksum offload */
3984 	ctrl1 = 0;
3985 	if (needs_l3_csum(m0) == 0)
3986 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
3987 	if (needs_l4_csum(m0) == 0)
3988 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
3989 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
3990 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
3991 		txq->txcsum++;	/* some hardware assistance provided */
3992 
3993 	/* VLAN tag insertion */
3994 	if (needs_vlan_insertion(m0)) {
3995 		ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
3996 		txq->vlan_insertion++;
3997 	}
3998 
3999 	/* CPL header */
4000 	cpl->ctrl0 = txq->cpl_ctrl0;
4001 	cpl->pack = 0;
4002 	cpl->len = htobe16(pktlen);
4003 	cpl->ctrl1 = htobe64(ctrl1);
4004 
4005 	/* SGL */
4006 	dst = (void *)(cpl + 1);
4007 	if (nsegs > 0) {
4008 
4009 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4010 		txq->sgl_wrs++;
4011 	} else {
4012 		struct mbuf *m;
4013 
4014 		for (m = m0; m != NULL; m = m->m_next) {
4015 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4016 #ifdef INVARIANTS
4017 			pktlen -= m->m_len;
4018 #endif
4019 		}
4020 #ifdef INVARIANTS
4021 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
4022 #endif
4023 		txq->imm_wrs++;
4024 	}
4025 
4026 	txq->txpkt_wrs++;
4027 
4028 	txsd = &txq->sdesc[eq->pidx];
4029 	txsd->m = m0;
4030 	txsd->desc_used = ndesc;
4031 
4032 	return (ndesc);
4033 }
4034 
4035 static int
4036 try_txpkts(struct mbuf *m, struct mbuf *n, struct txpkts *txp, u_int available)
4037 {
4038 	u_int needed, nsegs1, nsegs2, l1, l2;
4039 
4040 	if (cannot_use_txpkts(m) || cannot_use_txpkts(n))
4041 		return (1);
4042 
4043 	nsegs1 = mbuf_nsegs(m);
4044 	nsegs2 = mbuf_nsegs(n);
4045 	if (nsegs1 + nsegs2 == 2) {
4046 		txp->wr_type = 1;
4047 		l1 = l2 = txpkts1_len16();
4048 	} else {
4049 		txp->wr_type = 0;
4050 		l1 = txpkts0_len16(nsegs1);
4051 		l2 = txpkts0_len16(nsegs2);
4052 	}
4053 	txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) + l1 + l2;
4054 	needed = howmany(txp->len16, EQ_ESIZE / 16);
4055 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4056 		return (1);
4057 
4058 	txp->plen = m->m_pkthdr.len + n->m_pkthdr.len;
4059 	if (txp->plen > 65535)
4060 		return (1);
4061 
4062 	txp->npkt = 2;
4063 	set_mbuf_len16(m, l1);
4064 	set_mbuf_len16(n, l2);
4065 
4066 	return (0);
4067 }
4068 
4069 static int
4070 add_to_txpkts(struct mbuf *m, struct txpkts *txp, u_int available)
4071 {
4072 	u_int plen, len16, needed, nsegs;
4073 
4074 	MPASS(txp->wr_type == 0 || txp->wr_type == 1);
4075 
4076 	nsegs = mbuf_nsegs(m);
4077 	if (needs_tso(m) || (txp->wr_type == 1 && nsegs != 1))
4078 		return (1);
4079 
4080 	plen = txp->plen + m->m_pkthdr.len;
4081 	if (plen > 65535)
4082 		return (1);
4083 
4084 	if (txp->wr_type == 0)
4085 		len16 = txpkts0_len16(nsegs);
4086 	else
4087 		len16 = txpkts1_len16();
4088 	needed = howmany(txp->len16 + len16, EQ_ESIZE / 16);
4089 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4090 		return (1);
4091 
4092 	txp->npkt++;
4093 	txp->plen = plen;
4094 	txp->len16 += len16;
4095 	set_mbuf_len16(m, len16);
4096 
4097 	return (0);
4098 }
4099 
4100 /*
4101  * Write a txpkts WR for the packets in txp to the hardware descriptors, update
4102  * the software descriptor, and advance the pidx.  It is guaranteed that enough
4103  * descriptors are available.
4104  *
4105  * The return value is the # of hardware descriptors used.
4106  */
4107 static u_int
4108 write_txpkts_wr(struct sge_txq *txq, struct fw_eth_tx_pkts_wr *wr,
4109     struct mbuf *m0, const struct txpkts *txp, u_int available)
4110 {
4111 	struct sge_eq *eq = &txq->eq;
4112 	struct tx_sdesc *txsd;
4113 	struct cpl_tx_pkt_core *cpl;
4114 	uint32_t ctrl;
4115 	uint64_t ctrl1;
4116 	int ndesc, checkwrap;
4117 	struct mbuf *m;
4118 	void *flitp;
4119 
4120 	TXQ_LOCK_ASSERT_OWNED(txq);
4121 	MPASS(txp->npkt > 0);
4122 	MPASS(txp->plen < 65536);
4123 	MPASS(m0 != NULL);
4124 	MPASS(m0->m_nextpkt != NULL);
4125 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
4126 	MPASS(available > 0 && available < eq->sidx);
4127 
4128 	ndesc = howmany(txp->len16, EQ_ESIZE / 16);
4129 	MPASS(ndesc <= available);
4130 
4131 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4132 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
4133 	ctrl = V_FW_WR_LEN16(txp->len16);
4134 	wr->equiq_to_len16 = htobe32(ctrl);
4135 	wr->plen = htobe16(txp->plen);
4136 	wr->npkt = txp->npkt;
4137 	wr->r3 = 0;
4138 	wr->type = txp->wr_type;
4139 	flitp = wr + 1;
4140 
4141 	/*
4142 	 * At this point we are 16B into a hardware descriptor.  If checkwrap is
4143 	 * set then we know the WR is going to wrap around somewhere.  We'll
4144 	 * check for that at appropriate points.
4145 	 */
4146 	checkwrap = eq->sidx - ndesc < eq->pidx;
4147 	for (m = m0; m != NULL; m = m->m_nextpkt) {
4148 		if (txp->wr_type == 0) {
4149 			struct ulp_txpkt *ulpmc;
4150 			struct ulptx_idata *ulpsc;
4151 
4152 			/* ULP master command */
4153 			ulpmc = flitp;
4154 			ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
4155 			    V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid));
4156 			ulpmc->len = htobe32(mbuf_len16(m));
4157 
4158 			/* ULP subcommand */
4159 			ulpsc = (void *)(ulpmc + 1);
4160 			ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
4161 			    F_ULP_TX_SC_MORE);
4162 			ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
4163 
4164 			cpl = (void *)(ulpsc + 1);
4165 			if (checkwrap &&
4166 			    (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx])
4167 				cpl = (void *)&eq->desc[0];
4168 			txq->txpkts0_pkts += txp->npkt;
4169 			txq->txpkts0_wrs++;
4170 		} else {
4171 			cpl = flitp;
4172 			txq->txpkts1_pkts += txp->npkt;
4173 			txq->txpkts1_wrs++;
4174 		}
4175 
4176 		/* Checksum offload */
4177 		ctrl1 = 0;
4178 		if (needs_l3_csum(m) == 0)
4179 			ctrl1 |= F_TXPKT_IPCSUM_DIS;
4180 		if (needs_l4_csum(m) == 0)
4181 			ctrl1 |= F_TXPKT_L4CSUM_DIS;
4182 		if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4183 		    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4184 			txq->txcsum++;	/* some hardware assistance provided */
4185 
4186 		/* VLAN tag insertion */
4187 		if (needs_vlan_insertion(m)) {
4188 			ctrl1 |= F_TXPKT_VLAN_VLD |
4189 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
4190 			txq->vlan_insertion++;
4191 		}
4192 
4193 		/* CPL header */
4194 		cpl->ctrl0 = txq->cpl_ctrl0;
4195 		cpl->pack = 0;
4196 		cpl->len = htobe16(m->m_pkthdr.len);
4197 		cpl->ctrl1 = htobe64(ctrl1);
4198 
4199 		flitp = cpl + 1;
4200 		if (checkwrap &&
4201 		    (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
4202 			flitp = (void *)&eq->desc[0];
4203 
4204 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap);
4205 
4206 	}
4207 
4208 	txsd = &txq->sdesc[eq->pidx];
4209 	txsd->m = m0;
4210 	txsd->desc_used = ndesc;
4211 
4212 	return (ndesc);
4213 }
4214 
4215 /*
4216  * If the SGL ends on an address that is not 16 byte aligned, this function will
4217  * add a 0 filled flit at the end.
4218  */
4219 static void
4220 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap)
4221 {
4222 	struct sge_eq *eq = &txq->eq;
4223 	struct sglist *gl = txq->gl;
4224 	struct sglist_seg *seg;
4225 	__be64 *flitp, *wrap;
4226 	struct ulptx_sgl *usgl;
4227 	int i, nflits, nsegs;
4228 
4229 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
4230 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
4231 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
4232 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
4233 
4234 	get_pkt_gl(m, gl);
4235 	nsegs = gl->sg_nseg;
4236 	MPASS(nsegs > 0);
4237 
4238 	nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
4239 	flitp = (__be64 *)(*to);
4240 	wrap = (__be64 *)(&eq->desc[eq->sidx]);
4241 	seg = &gl->sg_segs[0];
4242 	usgl = (void *)flitp;
4243 
4244 	/*
4245 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
4246 	 * ring, so we're at least 16 bytes away from the status page.  There is
4247 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
4248 	 */
4249 
4250 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
4251 	    V_ULPTX_NSGE(nsegs));
4252 	usgl->len0 = htobe32(seg->ss_len);
4253 	usgl->addr0 = htobe64(seg->ss_paddr);
4254 	seg++;
4255 
4256 	if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) {
4257 
4258 		/* Won't wrap around at all */
4259 
4260 		for (i = 0; i < nsegs - 1; i++, seg++) {
4261 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
4262 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
4263 		}
4264 		if (i & 1)
4265 			usgl->sge[i / 2].len[1] = htobe32(0);
4266 		flitp += nflits;
4267 	} else {
4268 
4269 		/* Will wrap somewhere in the rest of the SGL */
4270 
4271 		/* 2 flits already written, write the rest flit by flit */
4272 		flitp = (void *)(usgl + 1);
4273 		for (i = 0; i < nflits - 2; i++) {
4274 			if (flitp == wrap)
4275 				flitp = (void *)eq->desc;
4276 			*flitp++ = get_flit(seg, nsegs - 1, i);
4277 		}
4278 	}
4279 
4280 	if (nflits & 1) {
4281 		MPASS(((uintptr_t)flitp) & 0xf);
4282 		*flitp++ = 0;
4283 	}
4284 
4285 	MPASS((((uintptr_t)flitp) & 0xf) == 0);
4286 	if (__predict_false(flitp == wrap))
4287 		*to = (void *)eq->desc;
4288 	else
4289 		*to = (void *)flitp;
4290 }
4291 
4292 static inline void
4293 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
4294 {
4295 
4296 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
4297 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
4298 
4299 	if (__predict_true((uintptr_t)(*to) + len <=
4300 	    (uintptr_t)&eq->desc[eq->sidx])) {
4301 		bcopy(from, *to, len);
4302 		(*to) += len;
4303 	} else {
4304 		int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
4305 
4306 		bcopy(from, *to, portion);
4307 		from += portion;
4308 		portion = len - portion;	/* remaining */
4309 		bcopy(from, (void *)eq->desc, portion);
4310 		(*to) = (caddr_t)eq->desc + portion;
4311 	}
4312 }
4313 
4314 static inline void
4315 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n)
4316 {
4317 	u_int db;
4318 
4319 	MPASS(n > 0);
4320 
4321 	db = eq->doorbells;
4322 	if (n > 1)
4323 		clrbit(&db, DOORBELL_WCWR);
4324 	wmb();
4325 
4326 	switch (ffs(db) - 1) {
4327 	case DOORBELL_UDB:
4328 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
4329 		break;
4330 
4331 	case DOORBELL_WCWR: {
4332 		volatile uint64_t *dst, *src;
4333 		int i;
4334 
4335 		/*
4336 		 * Queues whose 128B doorbell segment fits in the page do not
4337 		 * use relative qid (udb_qid is always 0).  Only queues with
4338 		 * doorbell segments can do WCWR.
4339 		 */
4340 		KASSERT(eq->udb_qid == 0 && n == 1,
4341 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
4342 		    __func__, eq->doorbells, n, eq->dbidx, eq));
4343 
4344 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
4345 		    UDBS_DB_OFFSET);
4346 		i = eq->dbidx;
4347 		src = (void *)&eq->desc[i];
4348 		while (src != (void *)&eq->desc[i + 1])
4349 			*dst++ = *src++;
4350 		wmb();
4351 		break;
4352 	}
4353 
4354 	case DOORBELL_UDBWC:
4355 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
4356 		wmb();
4357 		break;
4358 
4359 	case DOORBELL_KDB:
4360 		t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
4361 		    V_QID(eq->cntxt_id) | V_PIDX(n));
4362 		break;
4363 	}
4364 
4365 	IDXINCR(eq->dbidx, n, eq->sidx);
4366 }
4367 
4368 static inline u_int
4369 reclaimable_tx_desc(struct sge_eq *eq)
4370 {
4371 	uint16_t hw_cidx;
4372 
4373 	hw_cidx = read_hw_cidx(eq);
4374 	return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx));
4375 }
4376 
4377 static inline u_int
4378 total_available_tx_desc(struct sge_eq *eq)
4379 {
4380 	uint16_t hw_cidx, pidx;
4381 
4382 	hw_cidx = read_hw_cidx(eq);
4383 	pidx = eq->pidx;
4384 
4385 	if (pidx == hw_cidx)
4386 		return (eq->sidx - 1);
4387 	else
4388 		return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1);
4389 }
4390 
4391 static inline uint16_t
4392 read_hw_cidx(struct sge_eq *eq)
4393 {
4394 	struct sge_qstat *spg = (void *)&eq->desc[eq->sidx];
4395 	uint16_t cidx = spg->cidx;	/* stable snapshot */
4396 
4397 	return (be16toh(cidx));
4398 }
4399 
4400 /*
4401  * Reclaim 'n' descriptors approximately.
4402  */
4403 static u_int
4404 reclaim_tx_descs(struct sge_txq *txq, u_int n)
4405 {
4406 	struct tx_sdesc *txsd;
4407 	struct sge_eq *eq = &txq->eq;
4408 	u_int can_reclaim, reclaimed;
4409 
4410 	TXQ_LOCK_ASSERT_OWNED(txq);
4411 	MPASS(n > 0);
4412 
4413 	reclaimed = 0;
4414 	can_reclaim = reclaimable_tx_desc(eq);
4415 	while (can_reclaim && reclaimed < n) {
4416 		int ndesc;
4417 		struct mbuf *m, *nextpkt;
4418 
4419 		txsd = &txq->sdesc[eq->cidx];
4420 		ndesc = txsd->desc_used;
4421 
4422 		/* Firmware doesn't return "partial" credits. */
4423 		KASSERT(can_reclaim >= ndesc,
4424 		    ("%s: unexpected number of credits: %d, %d",
4425 		    __func__, can_reclaim, ndesc));
4426 
4427 		for (m = txsd->m; m != NULL; m = nextpkt) {
4428 			nextpkt = m->m_nextpkt;
4429 			m->m_nextpkt = NULL;
4430 			m_freem(m);
4431 		}
4432 		reclaimed += ndesc;
4433 		can_reclaim -= ndesc;
4434 		IDXINCR(eq->cidx, ndesc, eq->sidx);
4435 	}
4436 
4437 	return (reclaimed);
4438 }
4439 
4440 static void
4441 tx_reclaim(void *arg, int n)
4442 {
4443 	struct sge_txq *txq = arg;
4444 	struct sge_eq *eq = &txq->eq;
4445 
4446 	do {
4447 		if (TXQ_TRYLOCK(txq) == 0)
4448 			break;
4449 		n = reclaim_tx_descs(txq, 32);
4450 		if (eq->cidx == eq->pidx)
4451 			eq->equeqidx = eq->pidx;
4452 		TXQ_UNLOCK(txq);
4453 	} while (n > 0);
4454 }
4455 
4456 static __be64
4457 get_flit(struct sglist_seg *segs, int nsegs, int idx)
4458 {
4459 	int i = (idx / 3) * 2;
4460 
4461 	switch (idx % 3) {
4462 	case 0: {
4463 		__be64 rc;
4464 
4465 		rc = htobe32(segs[i].ss_len);
4466 		if (i + 1 < nsegs)
4467 			rc |= (uint64_t)htobe32(segs[i + 1].ss_len) << 32;
4468 
4469 		return (rc);
4470 	}
4471 	case 1:
4472 		return (htobe64(segs[i].ss_paddr));
4473 	case 2:
4474 		return (htobe64(segs[i + 1].ss_paddr));
4475 	}
4476 
4477 	return (0);
4478 }
4479 
4480 static void
4481 find_best_refill_source(struct adapter *sc, struct sge_fl *fl, int maxp)
4482 {
4483 	int8_t zidx, hwidx, idx;
4484 	uint16_t region1, region3;
4485 	int spare, spare_needed, n;
4486 	struct sw_zone_info *swz;
4487 	struct hw_buf_info *hwb, *hwb_list = &sc->sge.hw_buf_info[0];
4488 
4489 	/*
4490 	 * Buffer Packing: Look for PAGE_SIZE or larger zone which has a bufsize
4491 	 * large enough for the max payload and cluster metadata.  Otherwise
4492 	 * settle for the largest bufsize that leaves enough room in the cluster
4493 	 * for metadata.
4494 	 *
4495 	 * Without buffer packing: Look for the smallest zone which has a
4496 	 * bufsize large enough for the max payload.  Settle for the largest
4497 	 * bufsize available if there's nothing big enough for max payload.
4498 	 */
4499 	spare_needed = fl->flags & FL_BUF_PACKING ? CL_METADATA_SIZE : 0;
4500 	swz = &sc->sge.sw_zone_info[0];
4501 	hwidx = -1;
4502 	for (zidx = 0; zidx < SW_ZONE_SIZES; zidx++, swz++) {
4503 		if (swz->size > largest_rx_cluster) {
4504 			if (__predict_true(hwidx != -1))
4505 				break;
4506 
4507 			/*
4508 			 * This is a misconfiguration.  largest_rx_cluster is
4509 			 * preventing us from finding a refill source.  See
4510 			 * dev.t5nex.<n>.buffer_sizes to figure out why.
4511 			 */
4512 			device_printf(sc->dev, "largest_rx_cluster=%u leaves no"
4513 			    " refill source for fl %p (dma %u).  Ignored.\n",
4514 			    largest_rx_cluster, fl, maxp);
4515 		}
4516 		for (idx = swz->head_hwidx; idx != -1; idx = hwb->next) {
4517 			hwb = &hwb_list[idx];
4518 			spare = swz->size - hwb->size;
4519 			if (spare < spare_needed)
4520 				continue;
4521 
4522 			hwidx = idx;		/* best option so far */
4523 			if (hwb->size >= maxp) {
4524 
4525 				if ((fl->flags & FL_BUF_PACKING) == 0)
4526 					goto done; /* stop looking (not packing) */
4527 
4528 				if (swz->size >= safest_rx_cluster)
4529 					goto done; /* stop looking (packing) */
4530 			}
4531 			break;		/* keep looking, next zone */
4532 		}
4533 	}
4534 done:
4535 	/* A usable hwidx has been located. */
4536 	MPASS(hwidx != -1);
4537 	hwb = &hwb_list[hwidx];
4538 	zidx = hwb->zidx;
4539 	swz = &sc->sge.sw_zone_info[zidx];
4540 	region1 = 0;
4541 	region3 = swz->size - hwb->size;
4542 
4543 	/*
4544 	 * Stay within this zone and see if there is a better match when mbuf
4545 	 * inlining is allowed.  Remember that the hwidx's are sorted in
4546 	 * decreasing order of size (so in increasing order of spare area).
4547 	 */
4548 	for (idx = hwidx; idx != -1; idx = hwb->next) {
4549 		hwb = &hwb_list[idx];
4550 		spare = swz->size - hwb->size;
4551 
4552 		if (allow_mbufs_in_cluster == 0 || hwb->size < maxp)
4553 			break;
4554 
4555 		/*
4556 		 * Do not inline mbufs if doing so would violate the pad/pack
4557 		 * boundary alignment requirement.
4558 		 */
4559 		if (fl_pad && (MSIZE % sc->sge.pad_boundary) != 0)
4560 			continue;
4561 		if (fl->flags & FL_BUF_PACKING &&
4562 		    (MSIZE % sc->sge.pack_boundary) != 0)
4563 			continue;
4564 
4565 		if (spare < CL_METADATA_SIZE + MSIZE)
4566 			continue;
4567 		n = (spare - CL_METADATA_SIZE) / MSIZE;
4568 		if (n > howmany(hwb->size, maxp))
4569 			break;
4570 
4571 		hwidx = idx;
4572 		if (fl->flags & FL_BUF_PACKING) {
4573 			region1 = n * MSIZE;
4574 			region3 = spare - region1;
4575 		} else {
4576 			region1 = MSIZE;
4577 			region3 = spare - region1;
4578 			break;
4579 		}
4580 	}
4581 
4582 	KASSERT(zidx >= 0 && zidx < SW_ZONE_SIZES,
4583 	    ("%s: bad zone %d for fl %p, maxp %d", __func__, zidx, fl, maxp));
4584 	KASSERT(hwidx >= 0 && hwidx <= SGE_FLBUF_SIZES,
4585 	    ("%s: bad hwidx %d for fl %p, maxp %d", __func__, hwidx, fl, maxp));
4586 	KASSERT(region1 + sc->sge.hw_buf_info[hwidx].size + region3 ==
4587 	    sc->sge.sw_zone_info[zidx].size,
4588 	    ("%s: bad buffer layout for fl %p, maxp %d. "
4589 		"cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
4590 		sc->sge.sw_zone_info[zidx].size, region1,
4591 		sc->sge.hw_buf_info[hwidx].size, region3));
4592 	if (fl->flags & FL_BUF_PACKING || region1 > 0) {
4593 		KASSERT(region3 >= CL_METADATA_SIZE,
4594 		    ("%s: no room for metadata.  fl %p, maxp %d; "
4595 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
4596 		    sc->sge.sw_zone_info[zidx].size, region1,
4597 		    sc->sge.hw_buf_info[hwidx].size, region3));
4598 		KASSERT(region1 % MSIZE == 0,
4599 		    ("%s: bad mbuf region for fl %p, maxp %d. "
4600 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
4601 		    sc->sge.sw_zone_info[zidx].size, region1,
4602 		    sc->sge.hw_buf_info[hwidx].size, region3));
4603 	}
4604 
4605 	fl->cll_def.zidx = zidx;
4606 	fl->cll_def.hwidx = hwidx;
4607 	fl->cll_def.region1 = region1;
4608 	fl->cll_def.region3 = region3;
4609 }
4610 
4611 static void
4612 find_safe_refill_source(struct adapter *sc, struct sge_fl *fl)
4613 {
4614 	struct sge *s = &sc->sge;
4615 	struct hw_buf_info *hwb;
4616 	struct sw_zone_info *swz;
4617 	int spare;
4618 	int8_t hwidx;
4619 
4620 	if (fl->flags & FL_BUF_PACKING)
4621 		hwidx = s->safe_hwidx2;	/* with room for metadata */
4622 	else if (allow_mbufs_in_cluster && s->safe_hwidx2 != -1) {
4623 		hwidx = s->safe_hwidx2;
4624 		hwb = &s->hw_buf_info[hwidx];
4625 		swz = &s->sw_zone_info[hwb->zidx];
4626 		spare = swz->size - hwb->size;
4627 
4628 		/* no good if there isn't room for an mbuf as well */
4629 		if (spare < CL_METADATA_SIZE + MSIZE)
4630 			hwidx = s->safe_hwidx1;
4631 	} else
4632 		hwidx = s->safe_hwidx1;
4633 
4634 	if (hwidx == -1) {
4635 		/* No fallback source */
4636 		fl->cll_alt.hwidx = -1;
4637 		fl->cll_alt.zidx = -1;
4638 
4639 		return;
4640 	}
4641 
4642 	hwb = &s->hw_buf_info[hwidx];
4643 	swz = &s->sw_zone_info[hwb->zidx];
4644 	spare = swz->size - hwb->size;
4645 	fl->cll_alt.hwidx = hwidx;
4646 	fl->cll_alt.zidx = hwb->zidx;
4647 	if (allow_mbufs_in_cluster &&
4648 	    (fl_pad == 0 || (MSIZE % sc->sge.pad_boundary) == 0))
4649 		fl->cll_alt.region1 = ((spare - CL_METADATA_SIZE) / MSIZE) * MSIZE;
4650 	else
4651 		fl->cll_alt.region1 = 0;
4652 	fl->cll_alt.region3 = spare - fl->cll_alt.region1;
4653 }
4654 
4655 static void
4656 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
4657 {
4658 	mtx_lock(&sc->sfl_lock);
4659 	FL_LOCK(fl);
4660 	if ((fl->flags & FL_DOOMED) == 0) {
4661 		fl->flags |= FL_STARVING;
4662 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
4663 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
4664 	}
4665 	FL_UNLOCK(fl);
4666 	mtx_unlock(&sc->sfl_lock);
4667 }
4668 
4669 static void
4670 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq)
4671 {
4672 	struct sge_wrq *wrq = (void *)eq;
4673 
4674 	atomic_readandclear_int(&eq->equiq);
4675 	taskqueue_enqueue(sc->tq[eq->tx_chan], &wrq->wrq_tx_task);
4676 }
4677 
4678 static void
4679 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq)
4680 {
4681 	struct sge_txq *txq = (void *)eq;
4682 
4683 	MPASS((eq->flags & EQ_TYPEMASK) == EQ_ETH);
4684 
4685 	atomic_readandclear_int(&eq->equiq);
4686 	mp_ring_check_drainage(txq->r, 0);
4687 	taskqueue_enqueue(sc->tq[eq->tx_chan], &txq->tx_reclaim_task);
4688 }
4689 
4690 static int
4691 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
4692     struct mbuf *m)
4693 {
4694 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
4695 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
4696 	struct adapter *sc = iq->adapter;
4697 	struct sge *s = &sc->sge;
4698 	struct sge_eq *eq;
4699 	static void (*h[])(struct adapter *, struct sge_eq *) = {NULL,
4700 		&handle_wrq_egr_update, &handle_eth_egr_update,
4701 		&handle_wrq_egr_update};
4702 
4703 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
4704 	    rss->opcode));
4705 
4706 	eq = s->eqmap[qid - s->eq_start];
4707 	(*h[eq->flags & EQ_TYPEMASK])(sc, eq);
4708 
4709 	return (0);
4710 }
4711 
4712 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
4713 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
4714     offsetof(struct cpl_fw6_msg, data));
4715 
4716 static int
4717 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
4718 {
4719 	struct adapter *sc = iq->adapter;
4720 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
4721 
4722 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
4723 	    rss->opcode));
4724 
4725 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
4726 		const struct rss_header *rss2;
4727 
4728 		rss2 = (const struct rss_header *)&cpl->data[0];
4729 		return (sc->cpl_handler[rss2->opcode](iq, rss2, m));
4730 	}
4731 
4732 	return (sc->fw_msg_handler[cpl->type](sc, &cpl->data[0]));
4733 }
4734 
4735 static int
4736 sysctl_uint16(SYSCTL_HANDLER_ARGS)
4737 {
4738 	uint16_t *id = arg1;
4739 	int i = *id;
4740 
4741 	return sysctl_handle_int(oidp, &i, 0, req);
4742 }
4743 
4744 static int
4745 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
4746 {
4747 	struct sge *s = arg1;
4748 	struct hw_buf_info *hwb = &s->hw_buf_info[0];
4749 	struct sw_zone_info *swz = &s->sw_zone_info[0];
4750 	int i, rc;
4751 	struct sbuf sb;
4752 	char c;
4753 
4754 	sbuf_new(&sb, NULL, 32, SBUF_AUTOEXTEND);
4755 	for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
4756 		if (hwb->zidx >= 0 && swz[hwb->zidx].size <= largest_rx_cluster)
4757 			c = '*';
4758 		else
4759 			c = '\0';
4760 
4761 		sbuf_printf(&sb, "%u%c ", hwb->size, c);
4762 	}
4763 	sbuf_trim(&sb);
4764 	sbuf_finish(&sb);
4765 	rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
4766 	sbuf_delete(&sb);
4767 	return (rc);
4768 }
4769