xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision 95d45410b5100e07f6f98450bcd841a8945d4726)
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/kdb.h>
40 #include <sys/malloc.h>
41 #include <sys/queue.h>
42 #include <sys/sbuf.h>
43 #include <sys/taskqueue.h>
44 #include <sys/time.h>
45 #include <sys/sysctl.h>
46 #include <sys/smp.h>
47 #include <net/bpf.h>
48 #include <net/ethernet.h>
49 #include <net/if.h>
50 #include <net/if_vlan_var.h>
51 #include <netinet/in.h>
52 #include <netinet/ip.h>
53 #include <netinet/ip6.h>
54 #include <netinet/tcp.h>
55 #include <machine/md_var.h>
56 #include <vm/vm.h>
57 #include <vm/pmap.h>
58 #ifdef DEV_NETMAP
59 #include <machine/bus.h>
60 #include <sys/selinfo.h>
61 #include <net/if_var.h>
62 #include <net/netmap.h>
63 #include <dev/netmap/netmap_kern.h>
64 #endif
65 
66 #include "common/common.h"
67 #include "common/t4_regs.h"
68 #include "common/t4_regs_values.h"
69 #include "common/t4_msg.h"
70 
71 #ifdef T4_PKT_TIMESTAMP
72 #define RX_COPY_THRESHOLD (MINCLSIZE - 8)
73 #else
74 #define RX_COPY_THRESHOLD MINCLSIZE
75 #endif
76 
77 /*
78  * Ethernet frames are DMA'd at this byte offset into the freelist buffer.
79  * 0-7 are valid values.
80  */
81 int fl_pktshift = 2;
82 TUNABLE_INT("hw.cxgbe.fl_pktshift", &fl_pktshift);
83 
84 /*
85  * Pad ethernet payload up to this boundary.
86  * -1: driver should figure out a good value.
87  *  0: disable padding.
88  *  Any power of 2 from 32 to 4096 (both inclusive) is also a valid value.
89  */
90 int fl_pad = -1;
91 TUNABLE_INT("hw.cxgbe.fl_pad", &fl_pad);
92 
93 /*
94  * Status page length.
95  * -1: driver should figure out a good value.
96  *  64 or 128 are the only other valid values.
97  */
98 int spg_len = -1;
99 TUNABLE_INT("hw.cxgbe.spg_len", &spg_len);
100 
101 /*
102  * Congestion drops.
103  * -1: no congestion feedback (not recommended).
104  *  0: backpressure the channel instead of dropping packets right away.
105  *  1: no backpressure, drop packets for the congested queue immediately.
106  */
107 static int cong_drop = 0;
108 TUNABLE_INT("hw.cxgbe.cong_drop", &cong_drop);
109 
110 /*
111  * Deliver multiple frames in the same free list buffer if they fit.
112  * -1: let the driver decide whether to enable buffer packing or not.
113  *  0: disable buffer packing.
114  *  1: enable buffer packing.
115  */
116 static int buffer_packing = -1;
117 TUNABLE_INT("hw.cxgbe.buffer_packing", &buffer_packing);
118 
119 /*
120  * Start next frame in a packed buffer at this boundary.
121  * -1: driver should figure out a good value.
122  * T4:
123  * ---
124  * if fl_pad != 0
125  * 	value specified here will be overridden by fl_pad.
126  * else
127  * 	power of 2 from 32 to 4096 (both inclusive) is a valid value here.
128  * T5:
129  * ---
130  * 16, or a power of 2 from 64 to 4096 (both inclusive) is a valid value.
131  */
132 static int fl_pack = -1;
133 static int t4_fl_pack;
134 static int t5_fl_pack;
135 TUNABLE_INT("hw.cxgbe.fl_pack", &fl_pack);
136 
137 /*
138  * Allow the driver to create mbuf(s) in a cluster allocated for rx.
139  * 0: never; always allocate mbufs from the zone_mbuf UMA zone.
140  * 1: ok to create mbuf(s) within a cluster if there is room.
141  */
142 static int allow_mbufs_in_cluster = 1;
143 TUNABLE_INT("hw.cxgbe.allow_mbufs_in_cluster", &allow_mbufs_in_cluster);
144 
145 /*
146  * Largest rx cluster size that the driver is allowed to allocate.
147  */
148 static int largest_rx_cluster = MJUM16BYTES;
149 TUNABLE_INT("hw.cxgbe.largest_rx_cluster", &largest_rx_cluster);
150 
151 /*
152  * Size of cluster allocation that's most likely to succeed.  The driver will
153  * fall back to this size if it fails to allocate clusters larger than this.
154  */
155 static int safest_rx_cluster = PAGE_SIZE;
156 TUNABLE_INT("hw.cxgbe.safest_rx_cluster", &safest_rx_cluster);
157 
158 /* Used to track coalesced tx work request */
159 struct txpkts {
160 	uint64_t *flitp;	/* ptr to flit where next pkt should start */
161 	uint8_t npkt;		/* # of packets in this work request */
162 	uint8_t nflits;		/* # of flits used by this work request */
163 	uint16_t plen;		/* total payload (sum of all packets) */
164 };
165 
166 /* A packet's SGL.  This + m_pkthdr has all info needed for tx */
167 struct sgl {
168 	int nsegs;		/* # of segments in the SGL, 0 means imm. tx */
169 	int nflits;		/* # of flits needed for the SGL */
170 	bus_dma_segment_t seg[TX_SGL_SEGS];
171 };
172 
173 static int service_iq(struct sge_iq *, int);
174 static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t,
175     int *);
176 static int t4_eth_rx(struct sge_iq *, const struct rss_header *, struct mbuf *);
177 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int,
178     int);
179 static inline void init_fl(struct adapter *, struct sge_fl *, int, int, int,
180     char *);
181 static inline void init_eq(struct sge_eq *, int, int, uint8_t, uint16_t,
182     char *);
183 static int alloc_ring(struct adapter *, size_t, bus_dma_tag_t *, bus_dmamap_t *,
184     bus_addr_t *, void **);
185 static int free_ring(struct adapter *, bus_dma_tag_t, bus_dmamap_t, bus_addr_t,
186     void *);
187 static int alloc_iq_fl(struct port_info *, struct sge_iq *, struct sge_fl *,
188     int, int);
189 static int free_iq_fl(struct port_info *, struct sge_iq *, struct sge_fl *);
190 static void add_fl_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
191     struct sge_fl *);
192 static int alloc_fwq(struct adapter *);
193 static int free_fwq(struct adapter *);
194 static int alloc_mgmtq(struct adapter *);
195 static int free_mgmtq(struct adapter *);
196 static int alloc_rxq(struct port_info *, struct sge_rxq *, int, int,
197     struct sysctl_oid *);
198 static int free_rxq(struct port_info *, struct sge_rxq *);
199 #ifdef TCP_OFFLOAD
200 static int alloc_ofld_rxq(struct port_info *, struct sge_ofld_rxq *, int, int,
201     struct sysctl_oid *);
202 static int free_ofld_rxq(struct port_info *, struct sge_ofld_rxq *);
203 #endif
204 #ifdef DEV_NETMAP
205 static int alloc_nm_rxq(struct port_info *, struct sge_nm_rxq *, int, int,
206     struct sysctl_oid *);
207 static int free_nm_rxq(struct port_info *, struct sge_nm_rxq *);
208 static int alloc_nm_txq(struct port_info *, struct sge_nm_txq *, int, int,
209     struct sysctl_oid *);
210 static int free_nm_txq(struct port_info *, struct sge_nm_txq *);
211 #endif
212 static int ctrl_eq_alloc(struct adapter *, struct sge_eq *);
213 static int eth_eq_alloc(struct adapter *, struct port_info *, struct sge_eq *);
214 #ifdef TCP_OFFLOAD
215 static int ofld_eq_alloc(struct adapter *, struct port_info *, struct sge_eq *);
216 #endif
217 static int alloc_eq(struct adapter *, struct port_info *, struct sge_eq *);
218 static int free_eq(struct adapter *, struct sge_eq *);
219 static int alloc_wrq(struct adapter *, struct port_info *, struct sge_wrq *,
220     struct sysctl_oid *);
221 static int free_wrq(struct adapter *, struct sge_wrq *);
222 static int alloc_txq(struct port_info *, struct sge_txq *, int,
223     struct sysctl_oid *);
224 static int free_txq(struct port_info *, struct sge_txq *);
225 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int);
226 static inline bool is_new_response(const struct sge_iq *, struct rsp_ctrl **);
227 static inline void iq_next(struct sge_iq *);
228 static inline void ring_fl_db(struct adapter *, struct sge_fl *);
229 static int refill_fl(struct adapter *, struct sge_fl *, int);
230 static void refill_sfl(void *);
231 static int alloc_fl_sdesc(struct sge_fl *);
232 static void free_fl_sdesc(struct adapter *, struct sge_fl *);
233 static void find_best_refill_source(struct adapter *, struct sge_fl *, int);
234 static void find_safe_refill_source(struct adapter *, struct sge_fl *);
235 static void add_fl_to_sfl(struct adapter *, struct sge_fl *);
236 
237 static int get_pkt_sgl(struct sge_txq *, struct mbuf **, struct sgl *, int);
238 static int free_pkt_sgl(struct sge_txq *, struct sgl *);
239 static int write_txpkt_wr(struct port_info *, struct sge_txq *, struct mbuf *,
240     struct sgl *);
241 static int add_to_txpkts(struct port_info *, struct sge_txq *, struct txpkts *,
242     struct mbuf *, struct sgl *);
243 static void write_txpkts_wr(struct sge_txq *, struct txpkts *);
244 static inline void write_ulp_cpl_sgl(struct port_info *, struct sge_txq *,
245     struct txpkts *, struct mbuf *, struct sgl *);
246 static int write_sgl_to_txd(struct sge_eq *, struct sgl *, caddr_t *);
247 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int);
248 static inline void ring_eq_db(struct adapter *, struct sge_eq *);
249 static inline int reclaimable(struct sge_eq *);
250 static int reclaim_tx_descs(struct sge_txq *, int, int);
251 static void write_eqflush_wr(struct sge_eq *);
252 static __be64 get_flit(bus_dma_segment_t *, int, int);
253 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
254     struct mbuf *);
255 static int handle_fw_msg(struct sge_iq *, const struct rss_header *,
256     struct mbuf *);
257 
258 static int sysctl_uint16(SYSCTL_HANDLER_ARGS);
259 static int sysctl_bufsizes(SYSCTL_HANDLER_ARGS);
260 
261 /*
262  * Called on MOD_LOAD.  Validates and calculates the SGE tunables.
263  */
264 void
265 t4_sge_modload(void)
266 {
267 	int pad;
268 
269 	/* set pad to a reasonable powerof2 between 16 and 4096 (inclusive) */
270 #if defined(__i386__) || defined(__amd64__)
271 	pad = max(cpu_clflush_line_size, 16);
272 #else
273 	pad = max(CACHE_LINE_SIZE, 16);
274 #endif
275 	pad = min(pad, 4096);
276 
277 	if (fl_pktshift < 0 || fl_pktshift > 7) {
278 		printf("Invalid hw.cxgbe.fl_pktshift value (%d),"
279 		    " using 2 instead.\n", fl_pktshift);
280 		fl_pktshift = 2;
281 	}
282 
283 	if (fl_pad != 0 &&
284 	    (fl_pad < 32 || fl_pad > 4096 || !powerof2(fl_pad))) {
285 
286 		if (fl_pad != -1) {
287 			printf("Invalid hw.cxgbe.fl_pad value (%d),"
288 			    " using %d instead.\n", fl_pad, max(pad, 32));
289 		}
290 		fl_pad = max(pad, 32);
291 	}
292 
293 	/*
294 	 * T4 has the same pad and pack boundary.  If a pad boundary is set,
295 	 * pack boundary must be set to the same value.  Otherwise take the
296 	 * specified value or auto-calculate something reasonable.
297 	 */
298 	if (fl_pad)
299 		t4_fl_pack = fl_pad;
300 	else if (fl_pack < 32 || fl_pack > 4096 || !powerof2(fl_pack))
301 		t4_fl_pack = max(pad, 32);
302 	else
303 		t4_fl_pack = fl_pack;
304 
305 	/* T5's pack boundary is independent of the pad boundary. */
306 	if (fl_pack < 16 || fl_pack == 32 || fl_pack > 4096 ||
307 	    !powerof2(fl_pack))
308 	       t5_fl_pack = max(pad, CACHE_LINE_SIZE);
309 	else
310 	       t5_fl_pack = fl_pack;
311 
312 	if (spg_len != 64 && spg_len != 128) {
313 		int len;
314 
315 #if defined(__i386__) || defined(__amd64__)
316 		len = cpu_clflush_line_size > 64 ? 128 : 64;
317 #else
318 		len = 64;
319 #endif
320 		if (spg_len != -1) {
321 			printf("Invalid hw.cxgbe.spg_len value (%d),"
322 			    " using %d instead.\n", spg_len, len);
323 		}
324 		spg_len = len;
325 	}
326 
327 	if (cong_drop < -1 || cong_drop > 1) {
328 		printf("Invalid hw.cxgbe.cong_drop value (%d),"
329 		    " using 0 instead.\n", cong_drop);
330 		cong_drop = 0;
331 	}
332 }
333 
334 void
335 t4_init_sge_cpl_handlers(struct adapter *sc)
336 {
337 
338 	t4_register_cpl_handler(sc, CPL_FW4_MSG, handle_fw_msg);
339 	t4_register_cpl_handler(sc, CPL_FW6_MSG, handle_fw_msg);
340 	t4_register_cpl_handler(sc, CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
341 	t4_register_cpl_handler(sc, CPL_RX_PKT, t4_eth_rx);
342 	t4_register_fw_msg_handler(sc, FW6_TYPE_CMD_RPL, t4_handle_fw_rpl);
343 }
344 
345 /*
346  * adap->params.vpd.cclk must be set up before this is called.
347  */
348 void
349 t4_tweak_chip_settings(struct adapter *sc)
350 {
351 	int i;
352 	uint32_t v, m;
353 	int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200};
354 	int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk;
355 	int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */
356 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
357 	static int sge_flbuf_sizes[] = {
358 		MCLBYTES,
359 #if MJUMPAGESIZE != MCLBYTES
360 		MJUMPAGESIZE,
361 		MJUMPAGESIZE - CL_METADATA_SIZE,
362 		MJUMPAGESIZE - 2 * MSIZE - CL_METADATA_SIZE,
363 #endif
364 		MJUM9BYTES,
365 		MJUM16BYTES,
366 		MCLBYTES - MSIZE - CL_METADATA_SIZE,
367 		MJUM9BYTES - CL_METADATA_SIZE,
368 		MJUM16BYTES - CL_METADATA_SIZE,
369 	};
370 
371 	KASSERT(sc->flags & MASTER_PF,
372 	    ("%s: trying to change chip settings when not master.", __func__));
373 
374 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
375 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
376 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
377 	if (is_t4(sc) && (fl_pad || buffer_packing)) {
378 		/* t4_fl_pack has the correct value even when fl_pad = 0 */
379 		m |= V_INGPADBOUNDARY(M_INGPADBOUNDARY);
380 		v |= V_INGPADBOUNDARY(ilog2(t4_fl_pack) - 5);
381 	} else if (is_t5(sc) && fl_pad) {
382 		m |= V_INGPADBOUNDARY(M_INGPADBOUNDARY);
383 		v |= V_INGPADBOUNDARY(ilog2(fl_pad) - 5);
384 	}
385 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
386 
387 	if (is_t5(sc) && buffer_packing) {
388 		m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY);
389 		if (t5_fl_pack == 16)
390 			v = V_INGPACKBOUNDARY(0);
391 		else
392 			v = V_INGPACKBOUNDARY(ilog2(t5_fl_pack) - 5);
393 		t4_set_reg_field(sc, A_SGE_CONTROL2, m, v);
394 	}
395 
396 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
397 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
398 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
399 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
400 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
401 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
402 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
403 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
404 	t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v);
405 
406 	KASSERT(nitems(sge_flbuf_sizes) <= SGE_FLBUF_SIZES,
407 	    ("%s: hw buffer size table too big", __func__));
408 	for (i = 0; i < min(nitems(sge_flbuf_sizes), SGE_FLBUF_SIZES); i++) {
409 		t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i),
410 		    sge_flbuf_sizes[i]);
411 	}
412 
413 	v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) |
414 	    V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]);
415 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v);
416 
417 	KASSERT(intr_timer[0] <= timer_max,
418 	    ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0],
419 	    timer_max));
420 	for (i = 1; i < nitems(intr_timer); i++) {
421 		KASSERT(intr_timer[i] >= intr_timer[i - 1],
422 		    ("%s: timers not listed in increasing order (%d)",
423 		    __func__, i));
424 
425 		while (intr_timer[i] > timer_max) {
426 			if (i == nitems(intr_timer) - 1) {
427 				intr_timer[i] = timer_max;
428 				break;
429 			}
430 			intr_timer[i] += intr_timer[i - 1];
431 			intr_timer[i] /= 2;
432 		}
433 	}
434 
435 	v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) |
436 	    V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1]));
437 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v);
438 	v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) |
439 	    V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3]));
440 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v);
441 	v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) |
442 	    V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5]));
443 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v);
444 
445 	if (cong_drop == 0) {
446 		m = F_TUNNELCNGDROP0 | F_TUNNELCNGDROP1 | F_TUNNELCNGDROP2 |
447 		    F_TUNNELCNGDROP3;
448 		t4_set_reg_field(sc, A_TP_PARA_REG3, m, 0);
449 	}
450 
451 	/* 4K, 16K, 64K, 256K DDP "page sizes" */
452 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
453 	t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v);
454 
455 	m = v = F_TDDPTAGTCB;
456 	t4_set_reg_field(sc, A_ULP_RX_CTL, m, v);
457 
458 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
459 	    F_RESETDDPOFFSET;
460 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
461 	t4_set_reg_field(sc, A_TP_PARA_REG5, m, v);
462 }
463 
464 /*
465  * SGE wants the buffer to be at least 64B and then a multiple of the pad
466  * boundary or 16, whichever is greater.
467  */
468 static inline int
469 hwsz_ok(int hwsz)
470 {
471 	int mask = max(fl_pad, 16) - 1;
472 
473 	return (hwsz >= 64 && (hwsz & mask) == 0);
474 }
475 
476 /*
477  * XXX: driver really should be able to deal with unexpected settings.
478  */
479 int
480 t4_read_chip_settings(struct adapter *sc)
481 {
482 	struct sge *s = &sc->sge;
483 	int i, j, n, rc = 0;
484 	uint32_t m, v, r;
485 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
486 	static int sw_buf_sizes[] = {	/* Sorted by size */
487 		MCLBYTES,
488 #if MJUMPAGESIZE != MCLBYTES
489 		MJUMPAGESIZE,
490 #endif
491 		MJUM9BYTES,
492 		MJUM16BYTES
493 	};
494 	struct sw_zone_info *swz, *safe_swz;
495 	struct hw_buf_info *hwb;
496 
497 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
498 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
499 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
500 	if (is_t4(sc) && (fl_pad || buffer_packing)) {
501 		m |= V_INGPADBOUNDARY(M_INGPADBOUNDARY);
502 		v |= V_INGPADBOUNDARY(ilog2(t4_fl_pack) - 5);
503 	} else if (is_t5(sc) && fl_pad) {
504 		m |= V_INGPADBOUNDARY(M_INGPADBOUNDARY);
505 		v |= V_INGPADBOUNDARY(ilog2(fl_pad) - 5);
506 	}
507 	r = t4_read_reg(sc, A_SGE_CONTROL);
508 	if ((r & m) != v) {
509 		device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r);
510 		rc = EINVAL;
511 	}
512 
513 	if (is_t5(sc) && buffer_packing) {
514 		m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY);
515 		if (t5_fl_pack == 16)
516 			v = V_INGPACKBOUNDARY(0);
517 		else
518 			v = V_INGPACKBOUNDARY(ilog2(t5_fl_pack) - 5);
519 		r = t4_read_reg(sc, A_SGE_CONTROL2);
520 		if ((r & m) != v) {
521 			device_printf(sc->dev,
522 			    "invalid SGE_CONTROL2(0x%x)\n", r);
523 			rc = EINVAL;
524 		}
525 	}
526 	s->pack_boundary = is_t4(sc) ? t4_fl_pack : t5_fl_pack;
527 
528 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
529 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
530 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
531 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
532 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
533 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
534 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
535 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
536 	r = t4_read_reg(sc, A_SGE_HOST_PAGE_SIZE);
537 	if (r != v) {
538 		device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r);
539 		rc = EINVAL;
540 	}
541 
542 	/* Filter out unusable hw buffer sizes entirely (mark with -2). */
543 	hwb = &s->hw_buf_info[0];
544 	for (i = 0; i < nitems(s->hw_buf_info); i++, hwb++) {
545 		r = t4_read_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i));
546 		hwb->size = r;
547 		hwb->zidx = hwsz_ok(r) ? -1 : -2;
548 		hwb->next = -1;
549 	}
550 
551 	/*
552 	 * Create a sorted list in decreasing order of hw buffer sizes (and so
553 	 * increasing order of spare area) for each software zone.
554 	 */
555 	n = 0;	/* no usable buffer size to begin with */
556 	swz = &s->sw_zone_info[0];
557 	safe_swz = NULL;
558 	for (i = 0; i < SW_ZONE_SIZES; i++, swz++) {
559 		int8_t head = -1, tail = -1;
560 
561 		swz->size = sw_buf_sizes[i];
562 		swz->zone = m_getzone(swz->size);
563 		swz->type = m_gettype(swz->size);
564 
565 		if (swz->size == safest_rx_cluster)
566 			safe_swz = swz;
567 
568 		hwb = &s->hw_buf_info[0];
569 		for (j = 0; j < SGE_FLBUF_SIZES; j++, hwb++) {
570 			if (hwb->zidx != -1 || hwb->size > swz->size)
571 				continue;
572 			hwb->zidx = i;
573 			if (head == -1)
574 				head = tail = j;
575 			else if (hwb->size < s->hw_buf_info[tail].size) {
576 				s->hw_buf_info[tail].next = j;
577 				tail = j;
578 			} else {
579 				int8_t *cur;
580 				struct hw_buf_info *t;
581 
582 				for (cur = &head; *cur != -1; cur = &t->next) {
583 					t = &s->hw_buf_info[*cur];
584 					if (hwb->size == t->size) {
585 						hwb->zidx = -2;
586 						break;
587 					}
588 					if (hwb->size > t->size) {
589 						hwb->next = *cur;
590 						*cur = j;
591 						break;
592 					}
593 				}
594 			}
595 		}
596 		swz->head_hwidx = head;
597 		swz->tail_hwidx = tail;
598 
599 		if (tail != -1) {
600 			n++;
601 			if (swz->size - s->hw_buf_info[tail].size >=
602 			    CL_METADATA_SIZE)
603 				sc->flags |= BUF_PACKING_OK;
604 		}
605 	}
606 	if (n == 0) {
607 		device_printf(sc->dev, "no usable SGE FL buffer size.\n");
608 		rc = EINVAL;
609 	}
610 
611 	s->safe_hwidx1 = -1;
612 	s->safe_hwidx2 = -1;
613 	if (safe_swz != NULL) {
614 		s->safe_hwidx1 = safe_swz->head_hwidx;
615 		for (i = safe_swz->head_hwidx; i != -1; i = hwb->next) {
616 			int spare;
617 
618 			hwb = &s->hw_buf_info[i];
619 			spare = safe_swz->size - hwb->size;
620 			if (spare < CL_METADATA_SIZE)
621 				continue;
622 			if (s->safe_hwidx2 == -1 ||
623 			    spare == CL_METADATA_SIZE + MSIZE)
624 				s->safe_hwidx2 = i;
625 			if (spare >= CL_METADATA_SIZE + MSIZE)
626 				break;
627 		}
628 	}
629 
630 	r = t4_read_reg(sc, A_SGE_INGRESS_RX_THRESHOLD);
631 	s->counter_val[0] = G_THRESHOLD_0(r);
632 	s->counter_val[1] = G_THRESHOLD_1(r);
633 	s->counter_val[2] = G_THRESHOLD_2(r);
634 	s->counter_val[3] = G_THRESHOLD_3(r);
635 
636 	r = t4_read_reg(sc, A_SGE_TIMER_VALUE_0_AND_1);
637 	s->timer_val[0] = G_TIMERVALUE0(r) / core_ticks_per_usec(sc);
638 	s->timer_val[1] = G_TIMERVALUE1(r) / core_ticks_per_usec(sc);
639 	r = t4_read_reg(sc, A_SGE_TIMER_VALUE_2_AND_3);
640 	s->timer_val[2] = G_TIMERVALUE2(r) / core_ticks_per_usec(sc);
641 	s->timer_val[3] = G_TIMERVALUE3(r) / core_ticks_per_usec(sc);
642 	r = t4_read_reg(sc, A_SGE_TIMER_VALUE_4_AND_5);
643 	s->timer_val[4] = G_TIMERVALUE4(r) / core_ticks_per_usec(sc);
644 	s->timer_val[5] = G_TIMERVALUE5(r) / core_ticks_per_usec(sc);
645 
646 	if (cong_drop == 0) {
647 		m = F_TUNNELCNGDROP0 | F_TUNNELCNGDROP1 | F_TUNNELCNGDROP2 |
648 		    F_TUNNELCNGDROP3;
649 		r = t4_read_reg(sc, A_TP_PARA_REG3);
650 		if (r & m) {
651 			device_printf(sc->dev,
652 			    "invalid TP_PARA_REG3(0x%x)\n", r);
653 			rc = EINVAL;
654 		}
655 	}
656 
657 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
658 	r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ);
659 	if (r != v) {
660 		device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r);
661 		rc = EINVAL;
662 	}
663 
664 	m = v = F_TDDPTAGTCB;
665 	r = t4_read_reg(sc, A_ULP_RX_CTL);
666 	if ((r & m) != v) {
667 		device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r);
668 		rc = EINVAL;
669 	}
670 
671 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
672 	    F_RESETDDPOFFSET;
673 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
674 	r = t4_read_reg(sc, A_TP_PARA_REG5);
675 	if ((r & m) != v) {
676 		device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r);
677 		rc = EINVAL;
678 	}
679 
680 	r = t4_read_reg(sc, A_SGE_CONM_CTRL);
681 	s->fl_starve_threshold = G_EGRTHRESHOLD(r) * 2 + 1;
682 	if (is_t4(sc))
683 		s->fl_starve_threshold2 = s->fl_starve_threshold;
684 	else
685 		s->fl_starve_threshold2 = G_EGRTHRESHOLDPACKING(r) * 2 + 1;
686 
687 	/* egress queues: log2 of # of doorbells per BAR2 page */
688 	r = t4_read_reg(sc, A_SGE_EGRESS_QUEUES_PER_PAGE_PF);
689 	r >>= S_QUEUESPERPAGEPF0 +
690 	    (S_QUEUESPERPAGEPF1 - S_QUEUESPERPAGEPF0) * sc->pf;
691 	s->eq_s_qpp = r & M_QUEUESPERPAGEPF0;
692 
693 	/* ingress queues: log2 of # of doorbells per BAR2 page */
694 	r = t4_read_reg(sc, A_SGE_INGRESS_QUEUES_PER_PAGE_PF);
695 	r >>= S_QUEUESPERPAGEPF0 +
696 	    (S_QUEUESPERPAGEPF1 - S_QUEUESPERPAGEPF0) * sc->pf;
697 	s->iq_s_qpp = r & M_QUEUESPERPAGEPF0;
698 
699 	t4_init_tp_params(sc);
700 
701 	t4_read_mtu_tbl(sc, sc->params.mtus, NULL);
702 	t4_load_mtus(sc, sc->params.mtus, sc->params.a_wnd, sc->params.b_wnd);
703 
704 	return (rc);
705 }
706 
707 int
708 t4_create_dma_tag(struct adapter *sc)
709 {
710 	int rc;
711 
712 	rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
713 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
714 	    BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
715 	    NULL, &sc->dmat);
716 	if (rc != 0) {
717 		device_printf(sc->dev,
718 		    "failed to create main DMA tag: %d\n", rc);
719 	}
720 
721 	return (rc);
722 }
723 
724 static inline int
725 enable_buffer_packing(struct adapter *sc)
726 {
727 
728 	if (sc->flags & BUF_PACKING_OK &&
729 	    ((is_t5(sc) && buffer_packing) ||	/* 1 or -1 both ok for T5 */
730 	    (is_t4(sc) && buffer_packing == 1)))
731 		return (1);
732 	return (0);
733 }
734 
735 void
736 t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
737     struct sysctl_oid_list *children)
738 {
739 
740 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "buffer_sizes",
741 	    CTLTYPE_STRING | CTLFLAG_RD, &sc->sge, 0, sysctl_bufsizes, "A",
742 	    "freelist buffer sizes");
743 
744 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD,
745 	    NULL, fl_pktshift, "payload DMA offset in rx buffer (bytes)");
746 
747 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD,
748 	    NULL, fl_pad, "payload pad boundary (bytes)");
749 
750 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD,
751 	    NULL, spg_len, "status page size (bytes)");
752 
753 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD,
754 	    NULL, cong_drop, "congestion drop setting");
755 
756 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "buffer_packing", CTLFLAG_RD,
757 	    NULL, enable_buffer_packing(sc),
758 	    "pack multiple frames in one fl buffer");
759 
760 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD,
761 	    NULL, sc->sge.pack_boundary, "payload pack boundary (bytes)");
762 }
763 
764 int
765 t4_destroy_dma_tag(struct adapter *sc)
766 {
767 	if (sc->dmat)
768 		bus_dma_tag_destroy(sc->dmat);
769 
770 	return (0);
771 }
772 
773 /*
774  * Allocate and initialize the firmware event queue and the management queue.
775  *
776  * Returns errno on failure.  Resources allocated up to that point may still be
777  * allocated.  Caller is responsible for cleanup in case this function fails.
778  */
779 int
780 t4_setup_adapter_queues(struct adapter *sc)
781 {
782 	int rc;
783 
784 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
785 
786 	sysctl_ctx_init(&sc->ctx);
787 	sc->flags |= ADAP_SYSCTL_CTX;
788 
789 	/*
790 	 * Firmware event queue
791 	 */
792 	rc = alloc_fwq(sc);
793 	if (rc != 0)
794 		return (rc);
795 
796 	/*
797 	 * Management queue.  This is just a control queue that uses the fwq as
798 	 * its associated iq.
799 	 */
800 	rc = alloc_mgmtq(sc);
801 
802 	return (rc);
803 }
804 
805 /*
806  * Idempotent
807  */
808 int
809 t4_teardown_adapter_queues(struct adapter *sc)
810 {
811 
812 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
813 
814 	/* Do this before freeing the queue */
815 	if (sc->flags & ADAP_SYSCTL_CTX) {
816 		sysctl_ctx_free(&sc->ctx);
817 		sc->flags &= ~ADAP_SYSCTL_CTX;
818 	}
819 
820 	free_mgmtq(sc);
821 	free_fwq(sc);
822 
823 	return (0);
824 }
825 
826 static inline int
827 port_intr_count(struct port_info *pi)
828 {
829 	int rc = 0;
830 
831 	if (pi->flags & INTR_RXQ)
832 		rc += pi->nrxq;
833 #ifdef TCP_OFFLOAD
834 	if (pi->flags & INTR_OFLD_RXQ)
835 		rc += pi->nofldrxq;
836 #endif
837 #ifdef DEV_NETMAP
838 	if (pi->flags & INTR_NM_RXQ)
839 		rc += pi->nnmrxq;
840 #endif
841 	return (rc);
842 }
843 
844 static inline int
845 first_vector(struct port_info *pi)
846 {
847 	struct adapter *sc = pi->adapter;
848 	int rc = T4_EXTRA_INTR, i;
849 
850 	if (sc->intr_count == 1)
851 		return (0);
852 
853 	for_each_port(sc, i) {
854 		if (i == pi->port_id)
855 			break;
856 
857 		rc += port_intr_count(sc->port[i]);
858 	}
859 
860 	return (rc);
861 }
862 
863 /*
864  * Given an arbitrary "index," come up with an iq that can be used by other
865  * queues (of this port) for interrupt forwarding, SGE egress updates, etc.
866  * The iq returned is guaranteed to be something that takes direct interrupts.
867  */
868 static struct sge_iq *
869 port_intr_iq(struct port_info *pi, int idx)
870 {
871 	struct adapter *sc = pi->adapter;
872 	struct sge *s = &sc->sge;
873 	struct sge_iq *iq = NULL;
874 	int nintr, i;
875 
876 	if (sc->intr_count == 1)
877 		return (&sc->sge.fwq);
878 
879 	nintr = port_intr_count(pi);
880 	KASSERT(nintr != 0,
881 	    ("%s: pi %p has no exclusive interrupts, total interrupts = %d",
882 	    __func__, pi, sc->intr_count));
883 #ifdef DEV_NETMAP
884 	/* Exclude netmap queues as they can't take anyone else's interrupts */
885 	if (pi->flags & INTR_NM_RXQ)
886 		nintr -= pi->nnmrxq;
887 	KASSERT(nintr > 0,
888 	    ("%s: pi %p has nintr %d after netmap adjustment of %d", __func__,
889 	    pi, nintr, pi->nnmrxq));
890 #endif
891 	i = idx % nintr;
892 
893 	if (pi->flags & INTR_RXQ) {
894 	       	if (i < pi->nrxq) {
895 			iq = &s->rxq[pi->first_rxq + i].iq;
896 			goto done;
897 		}
898 		i -= pi->nrxq;
899 	}
900 #ifdef TCP_OFFLOAD
901 	if (pi->flags & INTR_OFLD_RXQ) {
902 	       	if (i < pi->nofldrxq) {
903 			iq = &s->ofld_rxq[pi->first_ofld_rxq + i].iq;
904 			goto done;
905 		}
906 		i -= pi->nofldrxq;
907 	}
908 #endif
909 	panic("%s: pi %p, intr_flags 0x%lx, idx %d, total intr %d\n", __func__,
910 	    pi, pi->flags & INTR_ALL, idx, nintr);
911 done:
912 	MPASS(iq != NULL);
913 	KASSERT(iq->flags & IQ_INTR,
914 	    ("%s: iq %p (port %p, intr_flags 0x%lx, idx %d)", __func__, iq, pi,
915 	    pi->flags & INTR_ALL, idx));
916 	return (iq);
917 }
918 
919 /* Maximum payload that can be delivered with a single iq descriptor */
920 static inline int
921 mtu_to_max_payload(struct adapter *sc, int mtu, const int toe)
922 {
923 	int payload;
924 
925 #ifdef TCP_OFFLOAD
926 	if (toe) {
927 		payload = sc->tt.rx_coalesce ?
928 		    G_RXCOALESCESIZE(t4_read_reg(sc, A_TP_PARA_REG2)) : mtu;
929 	} else {
930 #endif
931 		/* large enough even when hw VLAN extraction is disabled */
932 		payload = fl_pktshift + ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN +
933 		    mtu;
934 #ifdef TCP_OFFLOAD
935 	}
936 #endif
937 	payload = roundup2(payload, fl_pad);
938 
939 	return (payload);
940 }
941 
942 int
943 t4_setup_port_queues(struct port_info *pi)
944 {
945 	int rc = 0, i, j, intr_idx, iqid;
946 	struct sge_rxq *rxq;
947 	struct sge_txq *txq;
948 	struct sge_wrq *ctrlq;
949 #ifdef TCP_OFFLOAD
950 	struct sge_ofld_rxq *ofld_rxq;
951 	struct sge_wrq *ofld_txq;
952 #endif
953 #ifdef DEV_NETMAP
954 	struct sge_nm_rxq *nm_rxq;
955 	struct sge_nm_txq *nm_txq;
956 #endif
957 	char name[16];
958 	struct adapter *sc = pi->adapter;
959 	struct ifnet *ifp = pi->ifp;
960 	struct sysctl_oid *oid = device_get_sysctl_tree(pi->dev);
961 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
962 	int maxp, pack, mtu = ifp->if_mtu;
963 
964 	/* Interrupt vector to start from (when using multiple vectors) */
965 	intr_idx = first_vector(pi);
966 
967 	/*
968 	 * First pass over all NIC and TOE rx queues:
969 	 * a) initialize iq and fl
970 	 * b) allocate queue iff it will take direct interrupts.
971 	 */
972 	maxp = mtu_to_max_payload(sc, mtu, 0);
973 	pack = enable_buffer_packing(sc);
974 	if (pi->flags & INTR_RXQ) {
975 		oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "rxq",
976 		    CTLFLAG_RD, NULL, "rx queues");
977 	}
978 	for_each_rxq(pi, i, rxq) {
979 
980 		init_iq(&rxq->iq, sc, pi->tmr_idx, pi->pktc_idx, pi->qsize_rxq,
981 		    RX_IQ_ESIZE);
982 
983 		snprintf(name, sizeof(name), "%s rxq%d-fl",
984 		    device_get_nameunit(pi->dev), i);
985 		init_fl(sc, &rxq->fl, pi->qsize_rxq / 8, maxp, pack, name);
986 
987 		if (pi->flags & INTR_RXQ) {
988 			rxq->iq.flags |= IQ_INTR;
989 			rc = alloc_rxq(pi, rxq, intr_idx, i, oid);
990 			if (rc != 0)
991 				goto done;
992 			intr_idx++;
993 		}
994 	}
995 #ifdef TCP_OFFLOAD
996 	maxp = mtu_to_max_payload(sc, mtu, 1);
997 	if (is_offload(sc) && pi->flags & INTR_OFLD_RXQ) {
998 		oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "ofld_rxq",
999 		    CTLFLAG_RD, NULL,
1000 		    "rx queues for offloaded TCP connections");
1001 	}
1002 	for_each_ofld_rxq(pi, i, ofld_rxq) {
1003 
1004 		init_iq(&ofld_rxq->iq, sc, pi->tmr_idx, pi->pktc_idx,
1005 		    pi->qsize_rxq, RX_IQ_ESIZE);
1006 
1007 		snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
1008 		    device_get_nameunit(pi->dev), i);
1009 		init_fl(sc, &ofld_rxq->fl, pi->qsize_rxq / 8, maxp, pack, name);
1010 
1011 		if (pi->flags & INTR_OFLD_RXQ) {
1012 			ofld_rxq->iq.flags |= IQ_INTR;
1013 			rc = alloc_ofld_rxq(pi, ofld_rxq, intr_idx, i, oid);
1014 			if (rc != 0)
1015 				goto done;
1016 			intr_idx++;
1017 		}
1018 	}
1019 #endif
1020 #ifdef DEV_NETMAP
1021 	/*
1022 	 * We don't have buffers to back the netmap rx queues right now so we
1023 	 * create the queues in a way that doesn't set off any congestion signal
1024 	 * in the chip.
1025 	 */
1026 	if (pi->flags & INTR_NM_RXQ) {
1027 		oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "nm_rxq",
1028 		    CTLFLAG_RD, NULL, "rx queues for netmap");
1029 		for_each_nm_rxq(pi, i, nm_rxq) {
1030 			rc = alloc_nm_rxq(pi, nm_rxq, intr_idx, i, oid);
1031 			if (rc != 0)
1032 				goto done;
1033 			intr_idx++;
1034 		}
1035 	}
1036 #endif
1037 
1038 	/*
1039 	 * Second pass over all NIC and TOE rx queues.  The queues forwarding
1040 	 * their interrupts are allocated now.
1041 	 */
1042 	j = 0;
1043 	if (!(pi->flags & INTR_RXQ)) {
1044 		oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "rxq",
1045 		    CTLFLAG_RD, NULL, "rx queues");
1046 		for_each_rxq(pi, i, rxq) {
1047 			MPASS(!(rxq->iq.flags & IQ_INTR));
1048 
1049 			intr_idx = port_intr_iq(pi, j)->abs_id;
1050 
1051 			rc = alloc_rxq(pi, rxq, intr_idx, i, oid);
1052 			if (rc != 0)
1053 				goto done;
1054 			j++;
1055 		}
1056 	}
1057 #ifdef TCP_OFFLOAD
1058 	if (is_offload(sc) && !(pi->flags & INTR_OFLD_RXQ)) {
1059 		oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "ofld_rxq",
1060 		    CTLFLAG_RD, NULL,
1061 		    "rx queues for offloaded TCP connections");
1062 		for_each_ofld_rxq(pi, i, ofld_rxq) {
1063 			MPASS(!(ofld_rxq->iq.flags & IQ_INTR));
1064 
1065 			intr_idx = port_intr_iq(pi, j)->abs_id;
1066 
1067 			rc = alloc_ofld_rxq(pi, ofld_rxq, intr_idx, i, oid);
1068 			if (rc != 0)
1069 				goto done;
1070 			j++;
1071 		}
1072 	}
1073 #endif
1074 #ifdef DEV_NETMAP
1075 	if (!(pi->flags & INTR_NM_RXQ))
1076 		CXGBE_UNIMPLEMENTED(__func__);
1077 #endif
1078 
1079 	/*
1080 	 * Now the tx queues.  Only one pass needed.
1081 	 */
1082 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "txq", CTLFLAG_RD,
1083 	    NULL, "tx queues");
1084 	j = 0;
1085 	for_each_txq(pi, i, txq) {
1086 		iqid = port_intr_iq(pi, j)->cntxt_id;
1087 		snprintf(name, sizeof(name), "%s txq%d",
1088 		    device_get_nameunit(pi->dev), i);
1089 		init_eq(&txq->eq, EQ_ETH, pi->qsize_txq, pi->tx_chan, iqid,
1090 		    name);
1091 
1092 		rc = alloc_txq(pi, txq, i, oid);
1093 		if (rc != 0)
1094 			goto done;
1095 		j++;
1096 	}
1097 #ifdef TCP_OFFLOAD
1098 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "ofld_txq",
1099 	    CTLFLAG_RD, NULL, "tx queues for offloaded TCP connections");
1100 	for_each_ofld_txq(pi, i, ofld_txq) {
1101 		struct sysctl_oid *oid2;
1102 
1103 		iqid = port_intr_iq(pi, j)->cntxt_id;
1104 		snprintf(name, sizeof(name), "%s ofld_txq%d",
1105 		    device_get_nameunit(pi->dev), i);
1106 		init_eq(&ofld_txq->eq, EQ_OFLD, pi->qsize_txq, pi->tx_chan,
1107 		    iqid, name);
1108 
1109 		snprintf(name, sizeof(name), "%d", i);
1110 		oid2 = SYSCTL_ADD_NODE(&pi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
1111 		    name, CTLFLAG_RD, NULL, "offload tx queue");
1112 
1113 		rc = alloc_wrq(sc, pi, ofld_txq, oid2);
1114 		if (rc != 0)
1115 			goto done;
1116 		j++;
1117 	}
1118 #endif
1119 #ifdef DEV_NETMAP
1120 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "nm_txq",
1121 	    CTLFLAG_RD, NULL, "tx queues for netmap use");
1122 	for_each_nm_txq(pi, i, nm_txq) {
1123 		iqid = pi->first_nm_rxq + (j % pi->nnmrxq);
1124 		rc = alloc_nm_txq(pi, nm_txq, iqid, i, oid);
1125 		if (rc != 0)
1126 			goto done;
1127 		j++;
1128 	}
1129 #endif
1130 
1131 	/*
1132 	 * Finally, the control queue.
1133 	 */
1134 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, "ctrlq", CTLFLAG_RD,
1135 	    NULL, "ctrl queue");
1136 	ctrlq = &sc->sge.ctrlq[pi->port_id];
1137 	iqid = port_intr_iq(pi, 0)->cntxt_id;
1138 	snprintf(name, sizeof(name), "%s ctrlq", device_get_nameunit(pi->dev));
1139 	init_eq(&ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, pi->tx_chan, iqid, name);
1140 	rc = alloc_wrq(sc, pi, ctrlq, oid);
1141 
1142 done:
1143 	if (rc)
1144 		t4_teardown_port_queues(pi);
1145 
1146 	return (rc);
1147 }
1148 
1149 /*
1150  * Idempotent
1151  */
1152 int
1153 t4_teardown_port_queues(struct port_info *pi)
1154 {
1155 	int i;
1156 	struct adapter *sc = pi->adapter;
1157 	struct sge_rxq *rxq;
1158 	struct sge_txq *txq;
1159 #ifdef TCP_OFFLOAD
1160 	struct sge_ofld_rxq *ofld_rxq;
1161 	struct sge_wrq *ofld_txq;
1162 #endif
1163 #ifdef DEV_NETMAP
1164 	struct sge_nm_rxq *nm_rxq;
1165 	struct sge_nm_txq *nm_txq;
1166 #endif
1167 
1168 	/* Do this before freeing the queues */
1169 	if (pi->flags & PORT_SYSCTL_CTX) {
1170 		sysctl_ctx_free(&pi->ctx);
1171 		pi->flags &= ~PORT_SYSCTL_CTX;
1172 	}
1173 
1174 	/*
1175 	 * Take down all the tx queues first, as they reference the rx queues
1176 	 * (for egress updates, etc.).
1177 	 */
1178 
1179 	free_wrq(sc, &sc->sge.ctrlq[pi->port_id]);
1180 
1181 	for_each_txq(pi, i, txq) {
1182 		free_txq(pi, txq);
1183 	}
1184 #ifdef TCP_OFFLOAD
1185 	for_each_ofld_txq(pi, i, ofld_txq) {
1186 		free_wrq(sc, ofld_txq);
1187 	}
1188 #endif
1189 #ifdef DEV_NETMAP
1190 	for_each_nm_txq(pi, i, nm_txq)
1191 	    free_nm_txq(pi, nm_txq);
1192 #endif
1193 
1194 	/*
1195 	 * Then take down the rx queues that forward their interrupts, as they
1196 	 * reference other rx queues.
1197 	 */
1198 
1199 	for_each_rxq(pi, i, rxq) {
1200 		if ((rxq->iq.flags & IQ_INTR) == 0)
1201 			free_rxq(pi, rxq);
1202 	}
1203 #ifdef TCP_OFFLOAD
1204 	for_each_ofld_rxq(pi, i, ofld_rxq) {
1205 		if ((ofld_rxq->iq.flags & IQ_INTR) == 0)
1206 			free_ofld_rxq(pi, ofld_rxq);
1207 	}
1208 #endif
1209 #ifdef DEV_NETMAP
1210 	for_each_nm_rxq(pi, i, nm_rxq)
1211 	    free_nm_rxq(pi, nm_rxq);
1212 #endif
1213 
1214 	/*
1215 	 * Then take down the rx queues that take direct interrupts.
1216 	 */
1217 
1218 	for_each_rxq(pi, i, rxq) {
1219 		if (rxq->iq.flags & IQ_INTR)
1220 			free_rxq(pi, rxq);
1221 	}
1222 #ifdef TCP_OFFLOAD
1223 	for_each_ofld_rxq(pi, i, ofld_rxq) {
1224 		if (ofld_rxq->iq.flags & IQ_INTR)
1225 			free_ofld_rxq(pi, ofld_rxq);
1226 	}
1227 #endif
1228 #ifdef DEV_NETMAP
1229 	CXGBE_UNIMPLEMENTED(__func__);
1230 #endif
1231 
1232 	return (0);
1233 }
1234 
1235 /*
1236  * Deals with errors and the firmware event queue.  All data rx queues forward
1237  * their interrupt to the firmware event queue.
1238  */
1239 void
1240 t4_intr_all(void *arg)
1241 {
1242 	struct adapter *sc = arg;
1243 	struct sge_iq *fwq = &sc->sge.fwq;
1244 
1245 	t4_intr_err(arg);
1246 	if (atomic_cmpset_int(&fwq->state, IQS_IDLE, IQS_BUSY)) {
1247 		service_iq(fwq, 0);
1248 		atomic_cmpset_int(&fwq->state, IQS_BUSY, IQS_IDLE);
1249 	}
1250 }
1251 
1252 /* Deals with error interrupts */
1253 void
1254 t4_intr_err(void *arg)
1255 {
1256 	struct adapter *sc = arg;
1257 
1258 	t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
1259 	t4_slow_intr_handler(sc);
1260 }
1261 
1262 void
1263 t4_intr_evt(void *arg)
1264 {
1265 	struct sge_iq *iq = arg;
1266 
1267 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1268 		service_iq(iq, 0);
1269 		atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1270 	}
1271 }
1272 
1273 void
1274 t4_intr(void *arg)
1275 {
1276 	struct sge_iq *iq = arg;
1277 
1278 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1279 		service_iq(iq, 0);
1280 		atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1281 	}
1282 }
1283 
1284 /*
1285  * Deals with anything and everything on the given ingress queue.
1286  */
1287 static int
1288 service_iq(struct sge_iq *iq, int budget)
1289 {
1290 	struct sge_iq *q;
1291 	struct sge_rxq *rxq = iq_to_rxq(iq);	/* Use iff iq is part of rxq */
1292 	struct sge_fl *fl = &rxq->fl;		/* Use iff IQ_HAS_FL */
1293 	struct adapter *sc = iq->adapter;
1294 	struct rsp_ctrl *ctrl;
1295 	const struct rss_header *rss;
1296 	int ndescs = 0, limit, fl_bufs_used = 0;
1297 	int rsp_type;
1298 	uint32_t lq;
1299 	struct mbuf *m0;
1300 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1301 #if defined(INET) || defined(INET6)
1302 	const struct timeval lro_timeout = {0, sc->lro_timeout};
1303 #endif
1304 
1305 	limit = budget ? budget : iq->qsize / 8;
1306 
1307 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1308 
1309 	/*
1310 	 * We always come back and check the descriptor ring for new indirect
1311 	 * interrupts and other responses after running a single handler.
1312 	 */
1313 	for (;;) {
1314 		while (is_new_response(iq, &ctrl)) {
1315 
1316 			rmb();
1317 
1318 			m0 = NULL;
1319 			rsp_type = G_RSPD_TYPE(ctrl->u.type_gen);
1320 			lq = be32toh(ctrl->pldbuflen_qid);
1321 			rss = (const void *)iq->cdesc;
1322 
1323 			switch (rsp_type) {
1324 			case X_RSPD_TYPE_FLBUF:
1325 
1326 				KASSERT(iq->flags & IQ_HAS_FL,
1327 				    ("%s: data for an iq (%p) with no freelist",
1328 				    __func__, iq));
1329 
1330 				m0 = get_fl_payload(sc, fl, lq, &fl_bufs_used);
1331 				if (__predict_false(m0 == NULL))
1332 					goto process_iql;
1333 #ifdef T4_PKT_TIMESTAMP
1334 				/*
1335 				 * 60 bit timestamp for the payload is
1336 				 * *(uint64_t *)m0->m_pktdat.  Note that it is
1337 				 * in the leading free-space in the mbuf.  The
1338 				 * kernel can clobber it during a pullup,
1339 				 * m_copymdata, etc.  You need to make sure that
1340 				 * the mbuf reaches you unmolested if you care
1341 				 * about the timestamp.
1342 				 */
1343 				*(uint64_t *)m0->m_pktdat =
1344 				    be64toh(ctrl->u.last_flit) &
1345 				    0xfffffffffffffff;
1346 #endif
1347 
1348 				/* fall through */
1349 
1350 			case X_RSPD_TYPE_CPL:
1351 				KASSERT(rss->opcode < NUM_CPL_CMDS,
1352 				    ("%s: bad opcode %02x.", __func__,
1353 				    rss->opcode));
1354 				sc->cpl_handler[rss->opcode](iq, rss, m0);
1355 				break;
1356 
1357 			case X_RSPD_TYPE_INTR:
1358 
1359 				/*
1360 				 * Interrupts should be forwarded only to queues
1361 				 * that are not forwarding their interrupts.
1362 				 * This means service_iq can recurse but only 1
1363 				 * level deep.
1364 				 */
1365 				KASSERT(budget == 0,
1366 				    ("%s: budget %u, rsp_type %u", __func__,
1367 				    budget, rsp_type));
1368 
1369 				/*
1370 				 * There are 1K interrupt-capable queues (qids 0
1371 				 * through 1023).  A response type indicating a
1372 				 * forwarded interrupt with a qid >= 1K is an
1373 				 * iWARP async notification.
1374 				 */
1375 				if (lq >= 1024) {
1376                                         sc->an_handler(iq, ctrl);
1377                                         break;
1378                                 }
1379 
1380 				q = sc->sge.iqmap[lq - sc->sge.iq_start];
1381 				if (atomic_cmpset_int(&q->state, IQS_IDLE,
1382 				    IQS_BUSY)) {
1383 					if (service_iq(q, q->qsize / 8) == 0) {
1384 						atomic_cmpset_int(&q->state,
1385 						    IQS_BUSY, IQS_IDLE);
1386 					} else {
1387 						STAILQ_INSERT_TAIL(&iql, q,
1388 						    link);
1389 					}
1390 				}
1391 				break;
1392 
1393 			default:
1394 				KASSERT(0,
1395 				    ("%s: illegal response type %d on iq %p",
1396 				    __func__, rsp_type, iq));
1397 				log(LOG_ERR,
1398 				    "%s: illegal response type %d on iq %p",
1399 				    device_get_nameunit(sc->dev), rsp_type, iq);
1400 				break;
1401 			}
1402 
1403 			if (fl_bufs_used >= 16) {
1404 				FL_LOCK(fl);
1405 				fl->needed += fl_bufs_used;
1406 				refill_fl(sc, fl, 32);
1407 				FL_UNLOCK(fl);
1408 				fl_bufs_used = 0;
1409 			}
1410 
1411 			iq_next(iq);
1412 			if (++ndescs == limit) {
1413 				t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
1414 				    V_CIDXINC(ndescs) |
1415 				    V_INGRESSQID(iq->cntxt_id) |
1416 				    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1417 				ndescs = 0;
1418 
1419 #if defined(INET) || defined(INET6)
1420 				if (iq->flags & IQ_LRO_ENABLED &&
1421 				    sc->lro_timeout != 0) {
1422 					tcp_lro_flush_inactive(&rxq->lro,
1423 					    &lro_timeout);
1424 				}
1425 #endif
1426 
1427 				if (budget) {
1428 					if (fl_bufs_used) {
1429 						FL_LOCK(fl);
1430 						fl->needed += fl_bufs_used;
1431 						refill_fl(sc, fl, 32);
1432 						FL_UNLOCK(fl);
1433 					}
1434 					return (EINPROGRESS);
1435 				}
1436 			}
1437 		}
1438 
1439 process_iql:
1440 		if (STAILQ_EMPTY(&iql))
1441 			break;
1442 
1443 		/*
1444 		 * Process the head only, and send it to the back of the list if
1445 		 * it's still not done.
1446 		 */
1447 		q = STAILQ_FIRST(&iql);
1448 		STAILQ_REMOVE_HEAD(&iql, link);
1449 		if (service_iq(q, q->qsize / 8) == 0)
1450 			atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1451 		else
1452 			STAILQ_INSERT_TAIL(&iql, q, link);
1453 	}
1454 
1455 #if defined(INET) || defined(INET6)
1456 	if (iq->flags & IQ_LRO_ENABLED) {
1457 		struct lro_ctrl *lro = &rxq->lro;
1458 		struct lro_entry *l;
1459 
1460 		while (!SLIST_EMPTY(&lro->lro_active)) {
1461 			l = SLIST_FIRST(&lro->lro_active);
1462 			SLIST_REMOVE_HEAD(&lro->lro_active, next);
1463 			tcp_lro_flush(lro, l);
1464 		}
1465 	}
1466 #endif
1467 
1468 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_CIDXINC(ndescs) |
1469 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1470 
1471 	if (iq->flags & IQ_HAS_FL) {
1472 		int starved;
1473 
1474 		FL_LOCK(fl);
1475 		fl->needed += fl_bufs_used;
1476 		starved = refill_fl(sc, fl, 64);
1477 		FL_UNLOCK(fl);
1478 		if (__predict_false(starved != 0))
1479 			add_fl_to_sfl(sc, fl);
1480 	}
1481 
1482 	return (0);
1483 }
1484 
1485 static inline int
1486 cl_has_metadata(struct sge_fl *fl, struct cluster_layout *cll)
1487 {
1488 	int rc = fl->flags & FL_BUF_PACKING || cll->region1 > 0;
1489 
1490 	if (rc)
1491 		MPASS(cll->region3 >= CL_METADATA_SIZE);
1492 
1493 	return (rc);
1494 }
1495 
1496 static inline struct cluster_metadata *
1497 cl_metadata(struct adapter *sc, struct sge_fl *fl, struct cluster_layout *cll,
1498     caddr_t cl)
1499 {
1500 
1501 	if (cl_has_metadata(fl, cll)) {
1502 		struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1503 
1504 		return ((struct cluster_metadata *)(cl + swz->size) - 1);
1505 	}
1506 	return (NULL);
1507 }
1508 
1509 static void
1510 rxb_free(struct mbuf *m, void *arg1, void *arg2)
1511 {
1512 	uma_zone_t zone = arg1;
1513 	caddr_t cl = arg2;
1514 
1515 	uma_zfree(zone, cl);
1516 }
1517 
1518 /*
1519  * The mbuf returned by this function could be allocated from zone_mbuf or
1520  * constructed in spare room in the cluster.
1521  *
1522  * The mbuf carries the payload in one of these ways
1523  * a) frame inside the mbuf (mbuf from zone_mbuf)
1524  * b) m_cljset (for clusters without metadata) zone_mbuf
1525  * c) m_extaddref (cluster with metadata) inline mbuf
1526  * d) m_extaddref (cluster with metadata) zone_mbuf
1527  */
1528 static struct mbuf *
1529 get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int total, int flags)
1530 {
1531 	struct mbuf *m;
1532 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1533 	struct cluster_layout *cll = &sd->cll;
1534 	struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1535 	struct hw_buf_info *hwb = &sc->sge.hw_buf_info[cll->hwidx];
1536 	struct cluster_metadata *clm = cl_metadata(sc, fl, cll, sd->cl);
1537 	int len, padded_len;
1538 	caddr_t payload;
1539 
1540 	len = min(total, hwb->size - fl->rx_offset);
1541 	padded_len = roundup2(len, fl_pad);
1542 	payload = sd->cl + cll->region1 + fl->rx_offset;
1543 
1544 	if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) {
1545 
1546 		/*
1547 		 * Copy payload into a freshly allocated mbuf.
1548 		 */
1549 
1550 		m = flags & M_PKTHDR ?
1551 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1552 		if (m == NULL)
1553 			return (NULL);
1554 		fl->mbuf_allocated++;
1555 #ifdef T4_PKT_TIMESTAMP
1556 		/* Leave room for a timestamp */
1557 		m->m_data += 8;
1558 #endif
1559 		/* copy data to mbuf */
1560 		bcopy(payload, mtod(m, caddr_t), len);
1561 
1562 	} else if (sd->nmbuf * MSIZE < cll->region1) {
1563 
1564 		/*
1565 		 * There's spare room in the cluster for an mbuf.  Create one
1566 		 * and associate it with the payload that's in the cluster.
1567 		 */
1568 
1569 		MPASS(clm != NULL);
1570 		m = (struct mbuf *)(sd->cl + sd->nmbuf * MSIZE);
1571 		/* No bzero required */
1572 		if (m_init(m, NULL, 0, M_NOWAIT, MT_DATA, flags | M_NOFREE))
1573 			return (NULL);
1574 		fl->mbuf_inlined++;
1575 		m_extaddref(m, payload, padded_len, &clm->refcount, rxb_free,
1576 		    swz->zone, sd->cl);
1577 		sd->nmbuf++;
1578 
1579 	} else {
1580 
1581 		/*
1582 		 * Grab an mbuf from zone_mbuf and associate it with the
1583 		 * payload in the cluster.
1584 		 */
1585 
1586 		m = flags & M_PKTHDR ?
1587 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1588 		if (m == NULL)
1589 			return (NULL);
1590 		fl->mbuf_allocated++;
1591 		if (clm != NULL) {
1592 			m_extaddref(m, payload, padded_len, &clm->refcount,
1593 			    rxb_free, swz->zone, sd->cl);
1594 			sd->nmbuf++;
1595 		} else {
1596 			m_cljset(m, sd->cl, swz->type);
1597 			sd->cl = NULL;	/* consumed, not a recycle candidate */
1598 		}
1599 	}
1600 	if (flags & M_PKTHDR)
1601 		m->m_pkthdr.len = total;
1602 	m->m_len = len;
1603 
1604 	if (fl->flags & FL_BUF_PACKING) {
1605 		fl->rx_offset += roundup2(padded_len, sc->sge.pack_boundary);
1606 		MPASS(fl->rx_offset <= hwb->size);
1607 		if (fl->rx_offset < hwb->size)
1608 			return (m);	/* without advancing the cidx */
1609 	}
1610 
1611 	if (__predict_false(++fl->cidx == fl->cap))
1612 		fl->cidx = 0;
1613 	fl->rx_offset = 0;
1614 
1615 	return (m);
1616 }
1617 
1618 static struct mbuf *
1619 get_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf,
1620     int *fl_bufs_used)
1621 {
1622 	struct mbuf *m0, *m, **pnext;
1623 	u_int nbuf, len;
1624 
1625 	/*
1626 	 * No assertion for the fl lock because we don't need it.  This routine
1627 	 * is called only from the rx interrupt handler and it only updates
1628 	 * fl->cidx.  (Contrast that with fl->pidx/fl->needed which could be
1629 	 * updated in the rx interrupt handler or the starvation helper routine.
1630 	 * That's why code that manipulates fl->pidx/fl->needed needs the fl
1631 	 * lock but this routine does not).
1632 	 */
1633 
1634 	nbuf = 0;
1635 	len = G_RSPD_LEN(len_newbuf);
1636 	if (__predict_false(fl->m0 != NULL)) {
1637 		M_ASSERTPKTHDR(fl->m0);
1638 		MPASS(len == fl->m0->m_pkthdr.len);
1639 		MPASS(fl->remaining < len);
1640 
1641 		m0 = fl->m0;
1642 		pnext = fl->pnext;
1643 		len = fl->remaining;
1644 		fl->m0 = NULL;
1645 		goto get_segment;
1646 	}
1647 
1648 	if (fl->rx_offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
1649 		nbuf++;
1650 		fl->rx_offset = 0;
1651 		if (__predict_false(++fl->cidx == fl->cap))
1652 			fl->cidx = 0;
1653 	}
1654 
1655 	/*
1656 	 * Payload starts at rx_offset in the current hw buffer.  Its length is
1657 	 * 'len' and it may span multiple hw buffers.
1658 	 */
1659 
1660 	m0 = get_scatter_segment(sc, fl, len, M_PKTHDR);
1661 	if (m0 == NULL)
1662 		goto done;
1663 	len -= m0->m_len;
1664 	pnext = &m0->m_next;
1665 	while (len > 0) {
1666 		nbuf++;
1667 get_segment:
1668 		MPASS(fl->rx_offset == 0);
1669 		m = get_scatter_segment(sc, fl, len, 0);
1670 		if (m == NULL) {
1671 			fl->m0 = m0;
1672 			fl->pnext = pnext;
1673 			fl->remaining = len;
1674 			m0 = NULL;
1675 			goto done;
1676 		}
1677 		*pnext = m;
1678 		pnext = &m->m_next;
1679 		len -= m->m_len;
1680 	}
1681 	*pnext = NULL;
1682 	if (fl->rx_offset == 0)
1683 		nbuf++;
1684 done:
1685 	(*fl_bufs_used) += nbuf;
1686 	return (m0);
1687 }
1688 
1689 static int
1690 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
1691 {
1692 	struct sge_rxq *rxq = iq_to_rxq(iq);
1693 	struct ifnet *ifp = rxq->ifp;
1694 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
1695 #if defined(INET) || defined(INET6)
1696 	struct lro_ctrl *lro = &rxq->lro;
1697 #endif
1698 
1699 	KASSERT(m0 != NULL, ("%s: no payload with opcode %02x", __func__,
1700 	    rss->opcode));
1701 
1702 	m0->m_pkthdr.len -= fl_pktshift;
1703 	m0->m_len -= fl_pktshift;
1704 	m0->m_data += fl_pktshift;
1705 
1706 	m0->m_pkthdr.rcvif = ifp;
1707 	m0->m_flags |= M_FLOWID;
1708 	m0->m_pkthdr.flowid = be32toh(rss->hash_val);
1709 
1710 	if (cpl->csum_calc && !cpl->err_vec) {
1711 		if (ifp->if_capenable & IFCAP_RXCSUM &&
1712 		    cpl->l2info & htobe32(F_RXF_IP)) {
1713 			m0->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
1714 			    CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
1715 			rxq->rxcsum++;
1716 		} else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
1717 		    cpl->l2info & htobe32(F_RXF_IP6)) {
1718 			m0->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
1719 			    CSUM_PSEUDO_HDR);
1720 			rxq->rxcsum++;
1721 		}
1722 
1723 		if (__predict_false(cpl->ip_frag))
1724 			m0->m_pkthdr.csum_data = be16toh(cpl->csum);
1725 		else
1726 			m0->m_pkthdr.csum_data = 0xffff;
1727 	}
1728 
1729 	if (cpl->vlan_ex) {
1730 		m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
1731 		m0->m_flags |= M_VLANTAG;
1732 		rxq->vlan_extraction++;
1733 	}
1734 
1735 #if defined(INET) || defined(INET6)
1736 	if (cpl->l2info & htobe32(F_RXF_LRO) &&
1737 	    iq->flags & IQ_LRO_ENABLED &&
1738 	    tcp_lro_rx(lro, m0, 0) == 0) {
1739 		/* queued for LRO */
1740 	} else
1741 #endif
1742 	ifp->if_input(ifp, m0);
1743 
1744 	return (0);
1745 }
1746 
1747 /*
1748  * Doesn't fail.  Holds on to work requests it can't send right away.
1749  */
1750 void
1751 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
1752 {
1753 	struct sge_eq *eq = &wrq->eq;
1754 	int can_reclaim;
1755 	caddr_t dst;
1756 
1757 	TXQ_LOCK_ASSERT_OWNED(wrq);
1758 #ifdef TCP_OFFLOAD
1759 	KASSERT((eq->flags & EQ_TYPEMASK) == EQ_OFLD ||
1760 	    (eq->flags & EQ_TYPEMASK) == EQ_CTRL,
1761 	    ("%s: eq type %d", __func__, eq->flags & EQ_TYPEMASK));
1762 #else
1763 	KASSERT((eq->flags & EQ_TYPEMASK) == EQ_CTRL,
1764 	    ("%s: eq type %d", __func__, eq->flags & EQ_TYPEMASK));
1765 #endif
1766 
1767 	if (__predict_true(wr != NULL))
1768 		STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
1769 
1770 	can_reclaim = reclaimable(eq);
1771 	if (__predict_false(eq->flags & EQ_STALLED)) {
1772 		if (eq->avail + can_reclaim < tx_resume_threshold(eq))
1773 			return;
1774 		eq->flags &= ~EQ_STALLED;
1775 		eq->unstalled++;
1776 	}
1777 	eq->cidx += can_reclaim;
1778 	eq->avail += can_reclaim;
1779 	if (__predict_false(eq->cidx >= eq->cap))
1780 		eq->cidx -= eq->cap;
1781 
1782 	while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL) {
1783 		int ndesc;
1784 
1785 		if (__predict_false(wr->wr_len < 0 ||
1786 		    wr->wr_len > SGE_MAX_WR_LEN || (wr->wr_len & 0x7))) {
1787 
1788 #ifdef INVARIANTS
1789 			panic("%s: work request with length %d", __func__,
1790 			    wr->wr_len);
1791 #endif
1792 #ifdef KDB
1793 			kdb_backtrace();
1794 #endif
1795 			log(LOG_ERR, "%s: %s work request with length %d",
1796 			    device_get_nameunit(sc->dev), __func__, wr->wr_len);
1797 			STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
1798 			free_wrqe(wr);
1799 			continue;
1800 		}
1801 
1802 		ndesc = howmany(wr->wr_len, EQ_ESIZE);
1803 		if (eq->avail < ndesc) {
1804 			wrq->no_desc++;
1805 			break;
1806 		}
1807 
1808 		dst = (void *)&eq->desc[eq->pidx];
1809 		copy_to_txd(eq, wrtod(wr), &dst, wr->wr_len);
1810 
1811 		eq->pidx += ndesc;
1812 		eq->avail -= ndesc;
1813 		if (__predict_false(eq->pidx >= eq->cap))
1814 			eq->pidx -= eq->cap;
1815 
1816 		eq->pending += ndesc;
1817 		if (eq->pending >= 8)
1818 			ring_eq_db(sc, eq);
1819 
1820 		wrq->tx_wrs++;
1821 		STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
1822 		free_wrqe(wr);
1823 
1824 		if (eq->avail < 8) {
1825 			can_reclaim = reclaimable(eq);
1826 			eq->cidx += can_reclaim;
1827 			eq->avail += can_reclaim;
1828 			if (__predict_false(eq->cidx >= eq->cap))
1829 				eq->cidx -= eq->cap;
1830 		}
1831 	}
1832 
1833 	if (eq->pending)
1834 		ring_eq_db(sc, eq);
1835 
1836 	if (wr != NULL) {
1837 		eq->flags |= EQ_STALLED;
1838 		if (callout_pending(&eq->tx_callout) == 0)
1839 			callout_reset(&eq->tx_callout, 1, t4_tx_callout, eq);
1840 	}
1841 }
1842 
1843 /* Per-packet header in a coalesced tx WR, before the SGL starts (in flits) */
1844 #define TXPKTS_PKT_HDR ((\
1845     sizeof(struct ulp_txpkt) + \
1846     sizeof(struct ulptx_idata) + \
1847     sizeof(struct cpl_tx_pkt_core) \
1848     ) / 8)
1849 
1850 /* Header of a coalesced tx WR, before SGL of first packet (in flits) */
1851 #define TXPKTS_WR_HDR (\
1852     sizeof(struct fw_eth_tx_pkts_wr) / 8 + \
1853     TXPKTS_PKT_HDR)
1854 
1855 /* Header of a tx WR, before SGL of first packet (in flits) */
1856 #define TXPKT_WR_HDR ((\
1857     sizeof(struct fw_eth_tx_pkt_wr) + \
1858     sizeof(struct cpl_tx_pkt_core) \
1859     ) / 8 )
1860 
1861 /* Header of a tx LSO WR, before SGL of first packet (in flits) */
1862 #define TXPKT_LSO_WR_HDR ((\
1863     sizeof(struct fw_eth_tx_pkt_wr) + \
1864     sizeof(struct cpl_tx_pkt_lso_core) + \
1865     sizeof(struct cpl_tx_pkt_core) \
1866     ) / 8 )
1867 
1868 int
1869 t4_eth_tx(struct ifnet *ifp, struct sge_txq *txq, struct mbuf *m)
1870 {
1871 	struct port_info *pi = (void *)ifp->if_softc;
1872 	struct adapter *sc = pi->adapter;
1873 	struct sge_eq *eq = &txq->eq;
1874 	struct buf_ring *br = txq->br;
1875 	struct mbuf *next;
1876 	int rc, coalescing, can_reclaim;
1877 	struct txpkts txpkts;
1878 	struct sgl sgl;
1879 
1880 	TXQ_LOCK_ASSERT_OWNED(txq);
1881 	KASSERT(m, ("%s: called with nothing to do.", __func__));
1882 	KASSERT((eq->flags & EQ_TYPEMASK) == EQ_ETH,
1883 	    ("%s: eq type %d", __func__, eq->flags & EQ_TYPEMASK));
1884 
1885 	prefetch(&eq->desc[eq->pidx]);
1886 	prefetch(&txq->sdesc[eq->pidx]);
1887 
1888 	txpkts.npkt = 0;/* indicates there's nothing in txpkts */
1889 	coalescing = 0;
1890 
1891 	can_reclaim = reclaimable(eq);
1892 	if (__predict_false(eq->flags & EQ_STALLED)) {
1893 		if (eq->avail + can_reclaim < tx_resume_threshold(eq)) {
1894 			txq->m = m;
1895 			return (0);
1896 		}
1897 		eq->flags &= ~EQ_STALLED;
1898 		eq->unstalled++;
1899 	}
1900 
1901 	if (__predict_false(eq->flags & EQ_DOOMED)) {
1902 		m_freem(m);
1903 		while ((m = buf_ring_dequeue_sc(txq->br)) != NULL)
1904 			m_freem(m);
1905 		return (ENETDOWN);
1906 	}
1907 
1908 	if (eq->avail < 8 && can_reclaim)
1909 		reclaim_tx_descs(txq, can_reclaim, 32);
1910 
1911 	for (; m; m = next ? next : drbr_dequeue(ifp, br)) {
1912 
1913 		if (eq->avail < 8)
1914 			break;
1915 
1916 		next = m->m_nextpkt;
1917 		m->m_nextpkt = NULL;
1918 
1919 		if (next || buf_ring_peek(br))
1920 			coalescing = 1;
1921 
1922 		rc = get_pkt_sgl(txq, &m, &sgl, coalescing);
1923 		if (rc != 0) {
1924 			if (rc == ENOMEM) {
1925 
1926 				/* Short of resources, suspend tx */
1927 
1928 				m->m_nextpkt = next;
1929 				break;
1930 			}
1931 
1932 			/*
1933 			 * Unrecoverable error for this packet, throw it away
1934 			 * and move on to the next.  get_pkt_sgl may already
1935 			 * have freed m (it will be NULL in that case and the
1936 			 * m_freem here is still safe).
1937 			 */
1938 
1939 			m_freem(m);
1940 			continue;
1941 		}
1942 
1943 		if (coalescing &&
1944 		    add_to_txpkts(pi, txq, &txpkts, m, &sgl) == 0) {
1945 
1946 			/* Successfully absorbed into txpkts */
1947 
1948 			write_ulp_cpl_sgl(pi, txq, &txpkts, m, &sgl);
1949 			goto doorbell;
1950 		}
1951 
1952 		/*
1953 		 * We weren't coalescing to begin with, or current frame could
1954 		 * not be coalesced (add_to_txpkts flushes txpkts if a frame
1955 		 * given to it can't be coalesced).  Either way there should be
1956 		 * nothing in txpkts.
1957 		 */
1958 		KASSERT(txpkts.npkt == 0,
1959 		    ("%s: txpkts not empty: %d", __func__, txpkts.npkt));
1960 
1961 		/* We're sending out individual packets now */
1962 		coalescing = 0;
1963 
1964 		if (eq->avail < 8)
1965 			reclaim_tx_descs(txq, 0, 8);
1966 		rc = write_txpkt_wr(pi, txq, m, &sgl);
1967 		if (rc != 0) {
1968 
1969 			/* Short of hardware descriptors, suspend tx */
1970 
1971 			/*
1972 			 * This is an unlikely but expensive failure.  We've
1973 			 * done all the hard work (DMA mappings etc.) and now we
1974 			 * can't send out the packet.  What's worse, we have to
1975 			 * spend even more time freeing up everything in sgl.
1976 			 */
1977 			txq->no_desc++;
1978 			free_pkt_sgl(txq, &sgl);
1979 
1980 			m->m_nextpkt = next;
1981 			break;
1982 		}
1983 
1984 		ETHER_BPF_MTAP(ifp, m);
1985 		if (sgl.nsegs == 0)
1986 			m_freem(m);
1987 doorbell:
1988 		if (eq->pending >= 8)
1989 			ring_eq_db(sc, eq);
1990 
1991 		can_reclaim = reclaimable(eq);
1992 		if (can_reclaim >= 32)
1993 			reclaim_tx_descs(txq, can_reclaim, 64);
1994 	}
1995 
1996 	if (txpkts.npkt > 0)
1997 		write_txpkts_wr(txq, &txpkts);
1998 
1999 	/*
2000 	 * m not NULL means there was an error but we haven't thrown it away.
2001 	 * This can happen when we're short of tx descriptors (no_desc) or maybe
2002 	 * even DMA maps (no_dmamap).  Either way, a credit flush and reclaim
2003 	 * will get things going again.
2004 	 */
2005 	if (m && !(eq->flags & EQ_CRFLUSHED)) {
2006 		struct tx_sdesc *txsd = &txq->sdesc[eq->pidx];
2007 
2008 		/*
2009 		 * If EQ_CRFLUSHED is not set then we know we have at least one
2010 		 * available descriptor because any WR that reduces eq->avail to
2011 		 * 0 also sets EQ_CRFLUSHED.
2012 		 */
2013 		KASSERT(eq->avail > 0, ("%s: no space for eqflush.", __func__));
2014 
2015 		txsd->desc_used = 1;
2016 		txsd->credits = 0;
2017 		write_eqflush_wr(eq);
2018 	}
2019 	txq->m = m;
2020 
2021 	if (eq->pending)
2022 		ring_eq_db(sc, eq);
2023 
2024 	reclaim_tx_descs(txq, 0, 128);
2025 
2026 	if (eq->flags & EQ_STALLED && callout_pending(&eq->tx_callout) == 0)
2027 		callout_reset(&eq->tx_callout, 1, t4_tx_callout, eq);
2028 
2029 	return (0);
2030 }
2031 
2032 void
2033 t4_update_fl_bufsize(struct ifnet *ifp)
2034 {
2035 	struct port_info *pi = ifp->if_softc;
2036 	struct adapter *sc = pi->adapter;
2037 	struct sge_rxq *rxq;
2038 #ifdef TCP_OFFLOAD
2039 	struct sge_ofld_rxq *ofld_rxq;
2040 #endif
2041 	struct sge_fl *fl;
2042 	int i, maxp, mtu = ifp->if_mtu;
2043 
2044 	maxp = mtu_to_max_payload(sc, mtu, 0);
2045 	for_each_rxq(pi, i, rxq) {
2046 		fl = &rxq->fl;
2047 
2048 		FL_LOCK(fl);
2049 		find_best_refill_source(sc, fl, maxp);
2050 		FL_UNLOCK(fl);
2051 	}
2052 #ifdef TCP_OFFLOAD
2053 	maxp = mtu_to_max_payload(sc, mtu, 1);
2054 	for_each_ofld_rxq(pi, i, ofld_rxq) {
2055 		fl = &ofld_rxq->fl;
2056 
2057 		FL_LOCK(fl);
2058 		find_best_refill_source(sc, fl, maxp);
2059 		FL_UNLOCK(fl);
2060 	}
2061 #endif
2062 }
2063 
2064 int
2065 can_resume_tx(struct sge_eq *eq)
2066 {
2067 
2068 	return (eq->avail + reclaimable(eq) >= tx_resume_threshold(eq));
2069 }
2070 
2071 static inline void
2072 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
2073     int qsize, int esize)
2074 {
2075 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
2076 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
2077 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
2078 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
2079 
2080 	iq->flags = 0;
2081 	iq->adapter = sc;
2082 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
2083 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
2084 	if (pktc_idx >= 0) {
2085 		iq->intr_params |= F_QINTR_CNT_EN;
2086 		iq->intr_pktc_idx = pktc_idx;
2087 	}
2088 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
2089 	iq->esize = max(esize, 16);		/* See FW_IQ_CMD/iqesize */
2090 }
2091 
2092 static inline void
2093 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, int pack,
2094     char *name)
2095 {
2096 
2097 	fl->qsize = qsize;
2098 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
2099 	if (pack)
2100 		fl->flags |= FL_BUF_PACKING;
2101 	find_best_refill_source(sc, fl, maxp);
2102 	find_safe_refill_source(sc, fl);
2103 }
2104 
2105 static inline void
2106 init_eq(struct sge_eq *eq, int eqtype, int qsize, uint8_t tx_chan,
2107     uint16_t iqid, char *name)
2108 {
2109 	KASSERT(tx_chan < NCHAN, ("%s: bad tx channel %d", __func__, tx_chan));
2110 	KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
2111 
2112 	eq->flags = eqtype & EQ_TYPEMASK;
2113 	eq->tx_chan = tx_chan;
2114 	eq->iqid = iqid;
2115 	eq->qsize = qsize;
2116 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
2117 
2118 	TASK_INIT(&eq->tx_task, 0, t4_tx_task, eq);
2119 	callout_init(&eq->tx_callout, CALLOUT_MPSAFE);
2120 }
2121 
2122 static int
2123 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
2124     bus_dmamap_t *map, bus_addr_t *pa, void **va)
2125 {
2126 	int rc;
2127 
2128 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
2129 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
2130 	if (rc != 0) {
2131 		device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
2132 		goto done;
2133 	}
2134 
2135 	rc = bus_dmamem_alloc(*tag, va,
2136 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
2137 	if (rc != 0) {
2138 		device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
2139 		goto done;
2140 	}
2141 
2142 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
2143 	if (rc != 0) {
2144 		device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
2145 		goto done;
2146 	}
2147 done:
2148 	if (rc)
2149 		free_ring(sc, *tag, *map, *pa, *va);
2150 
2151 	return (rc);
2152 }
2153 
2154 static int
2155 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
2156     bus_addr_t pa, void *va)
2157 {
2158 	if (pa)
2159 		bus_dmamap_unload(tag, map);
2160 	if (va)
2161 		bus_dmamem_free(tag, va, map);
2162 	if (tag)
2163 		bus_dma_tag_destroy(tag);
2164 
2165 	return (0);
2166 }
2167 
2168 /*
2169  * Allocates the ring for an ingress queue and an optional freelist.  If the
2170  * freelist is specified it will be allocated and then associated with the
2171  * ingress queue.
2172  *
2173  * Returns errno on failure.  Resources allocated up to that point may still be
2174  * allocated.  Caller is responsible for cleanup in case this function fails.
2175  *
2176  * If the ingress queue will take interrupts directly (iq->flags & IQ_INTR) then
2177  * the intr_idx specifies the vector, starting from 0.  Otherwise it specifies
2178  * the abs_id of the ingress queue to which its interrupts should be forwarded.
2179  */
2180 static int
2181 alloc_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl,
2182     int intr_idx, int cong)
2183 {
2184 	int rc, i, cntxt_id;
2185 	size_t len;
2186 	struct fw_iq_cmd c;
2187 	struct adapter *sc = iq->adapter;
2188 	__be32 v = 0;
2189 
2190 	len = iq->qsize * iq->esize;
2191 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
2192 	    (void **)&iq->desc);
2193 	if (rc != 0)
2194 		return (rc);
2195 
2196 	bzero(&c, sizeof(c));
2197 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
2198 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
2199 	    V_FW_IQ_CMD_VFN(0));
2200 
2201 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
2202 	    FW_LEN16(c));
2203 
2204 	/* Special handling for firmware event queue */
2205 	if (iq == &sc->sge.fwq)
2206 		v |= F_FW_IQ_CMD_IQASYNCH;
2207 
2208 	if (iq->flags & IQ_INTR) {
2209 		KASSERT(intr_idx < sc->intr_count,
2210 		    ("%s: invalid direct intr_idx %d", __func__, intr_idx));
2211 	} else
2212 		v |= F_FW_IQ_CMD_IQANDST;
2213 	v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
2214 
2215 	c.type_to_iqandstindex = htobe32(v |
2216 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
2217 	    V_FW_IQ_CMD_VIID(pi->viid) |
2218 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
2219 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
2220 	    F_FW_IQ_CMD_IQGTSMODE |
2221 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
2222 	    V_FW_IQ_CMD_IQESIZE(ilog2(iq->esize) - 4));
2223 	c.iqsize = htobe16(iq->qsize);
2224 	c.iqaddr = htobe64(iq->ba);
2225 	if (cong >= 0)
2226 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
2227 
2228 	if (fl) {
2229 		mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
2230 
2231 		len = fl->qsize * RX_FL_ESIZE;
2232 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
2233 		    &fl->ba, (void **)&fl->desc);
2234 		if (rc)
2235 			return (rc);
2236 
2237 		/* Allocate space for one software descriptor per buffer. */
2238 		fl->cap = (fl->qsize - spg_len / RX_FL_ESIZE) * 8;
2239 		rc = alloc_fl_sdesc(fl);
2240 		if (rc != 0) {
2241 			device_printf(sc->dev,
2242 			    "failed to setup fl software descriptors: %d\n",
2243 			    rc);
2244 			return (rc);
2245 		}
2246 		fl->needed = fl->cap;
2247 		fl->lowat = fl->flags & FL_BUF_PACKING ?
2248 		    roundup2(sc->sge.fl_starve_threshold2, 8) :
2249 		    roundup2(sc->sge.fl_starve_threshold, 8);
2250 
2251 		c.iqns_to_fl0congen |=
2252 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
2253 			F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
2254 			(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
2255 			(fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
2256 			    0));
2257 		if (cong >= 0) {
2258 			c.iqns_to_fl0congen |=
2259 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
2260 				    F_FW_IQ_CMD_FL0CONGCIF |
2261 				    F_FW_IQ_CMD_FL0CONGEN);
2262 		}
2263 		c.fl0dcaen_to_fl0cidxfthresh =
2264 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(X_FETCHBURSTMIN_64B) |
2265 			V_FW_IQ_CMD_FL0FBMAX(X_FETCHBURSTMAX_512B));
2266 		c.fl0size = htobe16(fl->qsize);
2267 		c.fl0addr = htobe64(fl->ba);
2268 	}
2269 
2270 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
2271 	if (rc != 0) {
2272 		device_printf(sc->dev,
2273 		    "failed to create ingress queue: %d\n", rc);
2274 		return (rc);
2275 	}
2276 
2277 	iq->cdesc = iq->desc;
2278 	iq->cidx = 0;
2279 	iq->gen = 1;
2280 	iq->intr_next = iq->intr_params;
2281 	iq->cntxt_id = be16toh(c.iqid);
2282 	iq->abs_id = be16toh(c.physiqid);
2283 	iq->flags |= IQ_ALLOCATED;
2284 
2285 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
2286 	if (cntxt_id >= sc->sge.niq) {
2287 		panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
2288 		    cntxt_id, sc->sge.niq - 1);
2289 	}
2290 	sc->sge.iqmap[cntxt_id] = iq;
2291 
2292 	if (fl) {
2293 		fl->cntxt_id = be16toh(c.fl0id);
2294 		fl->pidx = fl->cidx = 0;
2295 
2296 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
2297 		if (cntxt_id >= sc->sge.neq) {
2298 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
2299 			    __func__, cntxt_id, sc->sge.neq - 1);
2300 		}
2301 		sc->sge.eqmap[cntxt_id] = (void *)fl;
2302 
2303 		FL_LOCK(fl);
2304 		/* Enough to make sure the SGE doesn't think it's starved */
2305 		refill_fl(sc, fl, fl->lowat);
2306 		FL_UNLOCK(fl);
2307 
2308 		iq->flags |= IQ_HAS_FL;
2309 	}
2310 
2311 	if (is_t5(sc) && cong >= 0) {
2312 		uint32_t param, val;
2313 
2314 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
2315 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
2316 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
2317 		if (cong == 0)
2318 			val = 1 << 19;
2319 		else {
2320 			val = 2 << 19;
2321 			for (i = 0; i < 4; i++) {
2322 				if (cong & (1 << i))
2323 					val |= 1 << (i << 2);
2324 			}
2325 		}
2326 
2327 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
2328 		if (rc != 0) {
2329 			/* report error but carry on */
2330 			device_printf(sc->dev,
2331 			    "failed to set congestion manager context for "
2332 			    "ingress queue %d: %d\n", iq->cntxt_id, rc);
2333 		}
2334 	}
2335 
2336 	/* Enable IQ interrupts */
2337 	atomic_store_rel_int(&iq->state, IQS_IDLE);
2338 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_SEINTARM(iq->intr_params) |
2339 	    V_INGRESSQID(iq->cntxt_id));
2340 
2341 	return (0);
2342 }
2343 
2344 static int
2345 free_iq_fl(struct port_info *pi, struct sge_iq *iq, struct sge_fl *fl)
2346 {
2347 	int rc;
2348 	struct adapter *sc = iq->adapter;
2349 	device_t dev;
2350 
2351 	if (sc == NULL)
2352 		return (0);	/* nothing to do */
2353 
2354 	dev = pi ? pi->dev : sc->dev;
2355 
2356 	if (iq->flags & IQ_ALLOCATED) {
2357 		rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
2358 		    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
2359 		    fl ? fl->cntxt_id : 0xffff, 0xffff);
2360 		if (rc != 0) {
2361 			device_printf(dev,
2362 			    "failed to free queue %p: %d\n", iq, rc);
2363 			return (rc);
2364 		}
2365 		iq->flags &= ~IQ_ALLOCATED;
2366 	}
2367 
2368 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
2369 
2370 	bzero(iq, sizeof(*iq));
2371 
2372 	if (fl) {
2373 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
2374 		    fl->desc);
2375 
2376 		if (fl->sdesc)
2377 			free_fl_sdesc(sc, fl);
2378 
2379 		if (mtx_initialized(&fl->fl_lock))
2380 			mtx_destroy(&fl->fl_lock);
2381 
2382 		bzero(fl, sizeof(*fl));
2383 	}
2384 
2385 	return (0);
2386 }
2387 
2388 static void
2389 add_fl_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
2390     struct sge_fl *fl)
2391 {
2392 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2393 
2394 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
2395 	    "freelist");
2396 	children = SYSCTL_CHILDREN(oid);
2397 
2398 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
2399 	    CTLTYPE_INT | CTLFLAG_RD, &fl->cntxt_id, 0, sysctl_uint16, "I",
2400 	    "SGE context id of the freelist");
2401 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
2402 	    0, "consumer index");
2403 	if (fl->flags & FL_BUF_PACKING) {
2404 		SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
2405 		    CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
2406 	}
2407 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
2408 	    0, "producer index");
2409 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_allocated",
2410 	    CTLFLAG_RD, &fl->mbuf_allocated, "# of mbuf allocated");
2411 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_inlined",
2412 	    CTLFLAG_RD, &fl->mbuf_inlined, "# of mbuf inlined in clusters");
2413 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
2414 	    CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
2415 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
2416 	    CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
2417 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
2418 	    CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
2419 }
2420 
2421 static int
2422 alloc_fwq(struct adapter *sc)
2423 {
2424 	int rc, intr_idx;
2425 	struct sge_iq *fwq = &sc->sge.fwq;
2426 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
2427 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2428 
2429 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE, FW_IQ_ESIZE);
2430 	fwq->flags |= IQ_INTR;	/* always */
2431 	intr_idx = sc->intr_count > 1 ? 1 : 0;
2432 	rc = alloc_iq_fl(sc->port[0], fwq, NULL, intr_idx, -1);
2433 	if (rc != 0) {
2434 		device_printf(sc->dev,
2435 		    "failed to create firmware event queue: %d\n", rc);
2436 		return (rc);
2437 	}
2438 
2439 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD,
2440 	    NULL, "firmware event queue");
2441 	children = SYSCTL_CHILDREN(oid);
2442 
2443 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "abs_id",
2444 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->abs_id, 0, sysctl_uint16, "I",
2445 	    "absolute id of the queue");
2446 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cntxt_id",
2447 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->cntxt_id, 0, sysctl_uint16, "I",
2448 	    "SGE context id of the queue");
2449 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cidx",
2450 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->cidx, 0, sysctl_uint16, "I",
2451 	    "consumer index");
2452 
2453 	return (0);
2454 }
2455 
2456 static int
2457 free_fwq(struct adapter *sc)
2458 {
2459 	return free_iq_fl(NULL, &sc->sge.fwq, NULL);
2460 }
2461 
2462 static int
2463 alloc_mgmtq(struct adapter *sc)
2464 {
2465 	int rc;
2466 	struct sge_wrq *mgmtq = &sc->sge.mgmtq;
2467 	char name[16];
2468 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
2469 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2470 
2471 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "mgmtq", CTLFLAG_RD,
2472 	    NULL, "management queue");
2473 
2474 	snprintf(name, sizeof(name), "%s mgmtq", device_get_nameunit(sc->dev));
2475 	init_eq(&mgmtq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[0]->tx_chan,
2476 	    sc->sge.fwq.cntxt_id, name);
2477 	rc = alloc_wrq(sc, NULL, mgmtq, oid);
2478 	if (rc != 0) {
2479 		device_printf(sc->dev,
2480 		    "failed to create management queue: %d\n", rc);
2481 		return (rc);
2482 	}
2483 
2484 	return (0);
2485 }
2486 
2487 static int
2488 free_mgmtq(struct adapter *sc)
2489 {
2490 
2491 	return free_wrq(sc, &sc->sge.mgmtq);
2492 }
2493 
2494 static inline int
2495 tnl_cong(struct port_info *pi)
2496 {
2497 
2498 	if (cong_drop == -1)
2499 		return (-1);
2500 	else if (cong_drop == 1)
2501 		return (0);
2502 	else
2503 		return (pi->rx_chan_map);
2504 }
2505 
2506 static int
2507 alloc_rxq(struct port_info *pi, struct sge_rxq *rxq, int intr_idx, int idx,
2508     struct sysctl_oid *oid)
2509 {
2510 	int rc;
2511 	struct sysctl_oid_list *children;
2512 	char name[16];
2513 
2514 	rc = alloc_iq_fl(pi, &rxq->iq, &rxq->fl, intr_idx, tnl_cong(pi));
2515 	if (rc != 0)
2516 		return (rc);
2517 
2518 	FL_LOCK(&rxq->fl);
2519 	refill_fl(pi->adapter, &rxq->fl, rxq->fl.needed / 8);
2520 	FL_UNLOCK(&rxq->fl);
2521 
2522 #if defined(INET) || defined(INET6)
2523 	rc = tcp_lro_init(&rxq->lro);
2524 	if (rc != 0)
2525 		return (rc);
2526 	rxq->lro.ifp = pi->ifp; /* also indicates LRO init'ed */
2527 
2528 	if (pi->ifp->if_capenable & IFCAP_LRO)
2529 		rxq->iq.flags |= IQ_LRO_ENABLED;
2530 #endif
2531 	rxq->ifp = pi->ifp;
2532 
2533 	children = SYSCTL_CHILDREN(oid);
2534 
2535 	snprintf(name, sizeof(name), "%d", idx);
2536 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
2537 	    NULL, "rx queue");
2538 	children = SYSCTL_CHILDREN(oid);
2539 
2540 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "abs_id",
2541 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.abs_id, 0, sysctl_uint16, "I",
2542 	    "absolute id of the queue");
2543 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cntxt_id",
2544 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cntxt_id, 0, sysctl_uint16, "I",
2545 	    "SGE context id of the queue");
2546 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cidx",
2547 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cidx, 0, sysctl_uint16, "I",
2548 	    "consumer index");
2549 #if defined(INET) || defined(INET6)
2550 	SYSCTL_ADD_INT(&pi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
2551 	    &rxq->lro.lro_queued, 0, NULL);
2552 	SYSCTL_ADD_INT(&pi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
2553 	    &rxq->lro.lro_flushed, 0, NULL);
2554 #endif
2555 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
2556 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
2557 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "vlan_extraction",
2558 	    CTLFLAG_RD, &rxq->vlan_extraction,
2559 	    "# of times hardware extracted 802.1Q tag");
2560 
2561 	add_fl_sysctls(&pi->ctx, oid, &rxq->fl);
2562 
2563 	return (rc);
2564 }
2565 
2566 static int
2567 free_rxq(struct port_info *pi, struct sge_rxq *rxq)
2568 {
2569 	int rc;
2570 
2571 #if defined(INET) || defined(INET6)
2572 	if (rxq->lro.ifp) {
2573 		tcp_lro_free(&rxq->lro);
2574 		rxq->lro.ifp = NULL;
2575 	}
2576 #endif
2577 
2578 	rc = free_iq_fl(pi, &rxq->iq, &rxq->fl);
2579 	if (rc == 0)
2580 		bzero(rxq, sizeof(*rxq));
2581 
2582 	return (rc);
2583 }
2584 
2585 #ifdef TCP_OFFLOAD
2586 static int
2587 alloc_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq,
2588     int intr_idx, int idx, struct sysctl_oid *oid)
2589 {
2590 	int rc;
2591 	struct sysctl_oid_list *children;
2592 	char name[16];
2593 
2594 	rc = alloc_iq_fl(pi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx,
2595 	    pi->rx_chan_map);
2596 	if (rc != 0)
2597 		return (rc);
2598 
2599 	children = SYSCTL_CHILDREN(oid);
2600 
2601 	snprintf(name, sizeof(name), "%d", idx);
2602 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
2603 	    NULL, "rx queue");
2604 	children = SYSCTL_CHILDREN(oid);
2605 
2606 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "abs_id",
2607 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.abs_id, 0, sysctl_uint16,
2608 	    "I", "absolute id of the queue");
2609 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cntxt_id",
2610 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cntxt_id, 0, sysctl_uint16,
2611 	    "I", "SGE context id of the queue");
2612 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cidx",
2613 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cidx, 0, sysctl_uint16, "I",
2614 	    "consumer index");
2615 
2616 	add_fl_sysctls(&pi->ctx, oid, &ofld_rxq->fl);
2617 
2618 	return (rc);
2619 }
2620 
2621 static int
2622 free_ofld_rxq(struct port_info *pi, struct sge_ofld_rxq *ofld_rxq)
2623 {
2624 	int rc;
2625 
2626 	rc = free_iq_fl(pi, &ofld_rxq->iq, &ofld_rxq->fl);
2627 	if (rc == 0)
2628 		bzero(ofld_rxq, sizeof(*ofld_rxq));
2629 
2630 	return (rc);
2631 }
2632 #endif
2633 
2634 #ifdef DEV_NETMAP
2635 static int
2636 alloc_nm_rxq(struct port_info *pi, struct sge_nm_rxq *nm_rxq, int intr_idx,
2637     int idx, struct sysctl_oid *oid)
2638 {
2639 	int rc;
2640 	struct sysctl_oid_list *children;
2641 	struct sysctl_ctx_list *ctx;
2642 	char name[16];
2643 	size_t len;
2644 	struct adapter *sc = pi->adapter;
2645 	struct netmap_adapter *na = NA(pi->nm_ifp);
2646 
2647 	MPASS(na != NULL);
2648 
2649 	len = pi->qsize_rxq * RX_IQ_ESIZE;
2650 	rc = alloc_ring(sc, len, &nm_rxq->iq_desc_tag, &nm_rxq->iq_desc_map,
2651 	    &nm_rxq->iq_ba, (void **)&nm_rxq->iq_desc);
2652 	if (rc != 0)
2653 		return (rc);
2654 
2655 	len = na->num_rx_desc * RX_FL_ESIZE + spg_len;
2656 	rc = alloc_ring(sc, len, &nm_rxq->fl_desc_tag, &nm_rxq->fl_desc_map,
2657 	    &nm_rxq->fl_ba, (void **)&nm_rxq->fl_desc);
2658 	if (rc != 0)
2659 		return (rc);
2660 
2661 	nm_rxq->pi = pi;
2662 	nm_rxq->nid = idx;
2663 	nm_rxq->iq_cidx = 0;
2664 	nm_rxq->iq_sidx = pi->qsize_rxq - spg_len / RX_IQ_ESIZE;
2665 	nm_rxq->iq_gen = F_RSPD_GEN;
2666 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
2667 	nm_rxq->fl_sidx = na->num_rx_desc;
2668 	nm_rxq->intr_idx = intr_idx;
2669 
2670 	ctx = &pi->ctx;
2671 	children = SYSCTL_CHILDREN(oid);
2672 
2673 	snprintf(name, sizeof(name), "%d", idx);
2674 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL,
2675 	    "rx queue");
2676 	children = SYSCTL_CHILDREN(oid);
2677 
2678 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
2679 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_abs_id, 0, sysctl_uint16,
2680 	    "I", "absolute id of the queue");
2681 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
2682 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cntxt_id, 0, sysctl_uint16,
2683 	    "I", "SGE context id of the queue");
2684 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
2685 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cidx, 0, sysctl_uint16, "I",
2686 	    "consumer index");
2687 
2688 	children = SYSCTL_CHILDREN(oid);
2689 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
2690 	    "freelist");
2691 	children = SYSCTL_CHILDREN(oid);
2692 
2693 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
2694 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->fl_cntxt_id, 0, sysctl_uint16,
2695 	    "I", "SGE context id of the freelist");
2696 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
2697 	    &nm_rxq->fl_cidx, 0, "consumer index");
2698 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
2699 	    &nm_rxq->fl_pidx, 0, "producer index");
2700 
2701 	return (rc);
2702 }
2703 
2704 
2705 static int
2706 free_nm_rxq(struct port_info *pi, struct sge_nm_rxq *nm_rxq)
2707 {
2708 	struct adapter *sc = pi->adapter;
2709 
2710 	free_ring(sc, nm_rxq->iq_desc_tag, nm_rxq->iq_desc_map, nm_rxq->iq_ba,
2711 	    nm_rxq->iq_desc);
2712 	free_ring(sc, nm_rxq->fl_desc_tag, nm_rxq->fl_desc_map, nm_rxq->fl_ba,
2713 	    nm_rxq->fl_desc);
2714 
2715 	return (0);
2716 }
2717 
2718 static int
2719 alloc_nm_txq(struct port_info *pi, struct sge_nm_txq *nm_txq, int iqidx, int idx,
2720     struct sysctl_oid *oid)
2721 {
2722 	int rc;
2723 	size_t len;
2724 	struct adapter *sc = pi->adapter;
2725 	struct netmap_adapter *na = NA(pi->nm_ifp);
2726 	char name[16];
2727 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2728 
2729 	len = na->num_tx_desc * EQ_ESIZE + spg_len;
2730 	rc = alloc_ring(sc, len, &nm_txq->desc_tag, &nm_txq->desc_map,
2731 	    &nm_txq->ba, (void **)&nm_txq->desc);
2732 	if (rc)
2733 		return (rc);
2734 
2735 	nm_txq->pidx = nm_txq->cidx = 0;
2736 	nm_txq->sidx = na->num_tx_desc;
2737 	nm_txq->nid = idx;
2738 	nm_txq->iqidx = iqidx;
2739 	nm_txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
2740 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(sc->pf));
2741 
2742 	snprintf(name, sizeof(name), "%d", idx);
2743 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
2744 	    NULL, "netmap tx queue");
2745 	children = SYSCTL_CHILDREN(oid);
2746 
2747 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
2748 	    &nm_txq->cntxt_id, 0, "SGE context id of the queue");
2749 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cidx",
2750 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->cidx, 0, sysctl_uint16, "I",
2751 	    "consumer index");
2752 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "pidx",
2753 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->pidx, 0, sysctl_uint16, "I",
2754 	    "producer index");
2755 
2756 	return (rc);
2757 }
2758 
2759 static int
2760 free_nm_txq(struct port_info *pi, struct sge_nm_txq *nm_txq)
2761 {
2762 	struct adapter *sc = pi->adapter;
2763 
2764 	free_ring(sc, nm_txq->desc_tag, nm_txq->desc_map, nm_txq->ba,
2765 	    nm_txq->desc);
2766 
2767 	return (0);
2768 }
2769 #endif
2770 
2771 static int
2772 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
2773 {
2774 	int rc, cntxt_id;
2775 	struct fw_eq_ctrl_cmd c;
2776 
2777 	bzero(&c, sizeof(c));
2778 
2779 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
2780 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
2781 	    V_FW_EQ_CTRL_CMD_VFN(0));
2782 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
2783 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
2784 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid)); /* XXX */
2785 	c.physeqid_pkd = htobe32(0);
2786 	c.fetchszm_to_iqid =
2787 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
2788 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
2789 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
2790 	c.dcaen_to_eqsize =
2791 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
2792 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
2793 		V_FW_EQ_CTRL_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
2794 		V_FW_EQ_CTRL_CMD_EQSIZE(eq->qsize));
2795 	c.eqaddr = htobe64(eq->ba);
2796 
2797 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
2798 	if (rc != 0) {
2799 		device_printf(sc->dev,
2800 		    "failed to create control queue %d: %d\n", eq->tx_chan, rc);
2801 		return (rc);
2802 	}
2803 	eq->flags |= EQ_ALLOCATED;
2804 
2805 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
2806 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
2807 	if (cntxt_id >= sc->sge.neq)
2808 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
2809 		cntxt_id, sc->sge.neq - 1);
2810 	sc->sge.eqmap[cntxt_id] = eq;
2811 
2812 	return (rc);
2813 }
2814 
2815 static int
2816 eth_eq_alloc(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
2817 {
2818 	int rc, cntxt_id;
2819 	struct fw_eq_eth_cmd c;
2820 
2821 	bzero(&c, sizeof(c));
2822 
2823 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
2824 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
2825 	    V_FW_EQ_ETH_CMD_VFN(0));
2826 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
2827 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
2828 	c.autoequiqe_to_viid = htobe32(V_FW_EQ_ETH_CMD_VIID(pi->viid));
2829 	c.fetchszm_to_iqid =
2830 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
2831 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
2832 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
2833 	c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
2834 		      V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
2835 		      V_FW_EQ_ETH_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
2836 		      V_FW_EQ_ETH_CMD_EQSIZE(eq->qsize));
2837 	c.eqaddr = htobe64(eq->ba);
2838 
2839 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
2840 	if (rc != 0) {
2841 		device_printf(pi->dev,
2842 		    "failed to create Ethernet egress queue: %d\n", rc);
2843 		return (rc);
2844 	}
2845 	eq->flags |= EQ_ALLOCATED;
2846 
2847 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
2848 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
2849 	if (cntxt_id >= sc->sge.neq)
2850 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
2851 		cntxt_id, sc->sge.neq - 1);
2852 	sc->sge.eqmap[cntxt_id] = eq;
2853 
2854 	return (rc);
2855 }
2856 
2857 #ifdef TCP_OFFLOAD
2858 static int
2859 ofld_eq_alloc(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
2860 {
2861 	int rc, cntxt_id;
2862 	struct fw_eq_ofld_cmd c;
2863 
2864 	bzero(&c, sizeof(c));
2865 
2866 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
2867 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
2868 	    V_FW_EQ_OFLD_CMD_VFN(0));
2869 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
2870 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
2871 	c.fetchszm_to_iqid =
2872 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
2873 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
2874 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
2875 	c.dcaen_to_eqsize =
2876 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
2877 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
2878 		V_FW_EQ_OFLD_CMD_CIDXFTHRESH(X_CIDXFLUSHTHRESH_32) |
2879 		V_FW_EQ_OFLD_CMD_EQSIZE(eq->qsize));
2880 	c.eqaddr = htobe64(eq->ba);
2881 
2882 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
2883 	if (rc != 0) {
2884 		device_printf(pi->dev,
2885 		    "failed to create egress queue for TCP offload: %d\n", rc);
2886 		return (rc);
2887 	}
2888 	eq->flags |= EQ_ALLOCATED;
2889 
2890 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
2891 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
2892 	if (cntxt_id >= sc->sge.neq)
2893 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
2894 		cntxt_id, sc->sge.neq - 1);
2895 	sc->sge.eqmap[cntxt_id] = eq;
2896 
2897 	return (rc);
2898 }
2899 #endif
2900 
2901 static int
2902 alloc_eq(struct adapter *sc, struct port_info *pi, struct sge_eq *eq)
2903 {
2904 	int rc;
2905 	size_t len;
2906 
2907 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
2908 
2909 	len = eq->qsize * EQ_ESIZE;
2910 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
2911 	    &eq->ba, (void **)&eq->desc);
2912 	if (rc)
2913 		return (rc);
2914 
2915 	eq->cap = eq->qsize - spg_len / EQ_ESIZE;
2916 	eq->spg = (void *)&eq->desc[eq->cap];
2917 	eq->avail = eq->cap - 1;	/* one less to avoid cidx = pidx */
2918 	eq->pidx = eq->cidx = 0;
2919 	eq->doorbells = sc->doorbells;
2920 
2921 	switch (eq->flags & EQ_TYPEMASK) {
2922 	case EQ_CTRL:
2923 		rc = ctrl_eq_alloc(sc, eq);
2924 		break;
2925 
2926 	case EQ_ETH:
2927 		rc = eth_eq_alloc(sc, pi, eq);
2928 		break;
2929 
2930 #ifdef TCP_OFFLOAD
2931 	case EQ_OFLD:
2932 		rc = ofld_eq_alloc(sc, pi, eq);
2933 		break;
2934 #endif
2935 
2936 	default:
2937 		panic("%s: invalid eq type %d.", __func__,
2938 		    eq->flags & EQ_TYPEMASK);
2939 	}
2940 	if (rc != 0) {
2941 		device_printf(sc->dev,
2942 		    "failed to allocate egress queue(%d): %d\n",
2943 		    eq->flags & EQ_TYPEMASK, rc);
2944 	}
2945 
2946 	eq->tx_callout.c_cpu = eq->cntxt_id % mp_ncpus;
2947 
2948 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
2949 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
2950 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
2951 		uint32_t s_qpp = sc->sge.eq_s_qpp;
2952 		uint32_t mask = (1 << s_qpp) - 1;
2953 		volatile uint8_t *udb;
2954 
2955 		udb = sc->udbs_base + UDBS_DB_OFFSET;
2956 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
2957 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
2958 		if (eq->udb_qid > PAGE_SIZE / UDBS_SEG_SIZE)
2959 	    		clrbit(&eq->doorbells, DOORBELL_WCWR);
2960 		else {
2961 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
2962 			eq->udb_qid = 0;
2963 		}
2964 		eq->udb = (volatile void *)udb;
2965 	}
2966 
2967 	return (rc);
2968 }
2969 
2970 static int
2971 free_eq(struct adapter *sc, struct sge_eq *eq)
2972 {
2973 	int rc;
2974 
2975 	if (eq->flags & EQ_ALLOCATED) {
2976 		switch (eq->flags & EQ_TYPEMASK) {
2977 		case EQ_CTRL:
2978 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
2979 			    eq->cntxt_id);
2980 			break;
2981 
2982 		case EQ_ETH:
2983 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
2984 			    eq->cntxt_id);
2985 			break;
2986 
2987 #ifdef TCP_OFFLOAD
2988 		case EQ_OFLD:
2989 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
2990 			    eq->cntxt_id);
2991 			break;
2992 #endif
2993 
2994 		default:
2995 			panic("%s: invalid eq type %d.", __func__,
2996 			    eq->flags & EQ_TYPEMASK);
2997 		}
2998 		if (rc != 0) {
2999 			device_printf(sc->dev,
3000 			    "failed to free egress queue (%d): %d\n",
3001 			    eq->flags & EQ_TYPEMASK, rc);
3002 			return (rc);
3003 		}
3004 		eq->flags &= ~EQ_ALLOCATED;
3005 	}
3006 
3007 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
3008 
3009 	if (mtx_initialized(&eq->eq_lock))
3010 		mtx_destroy(&eq->eq_lock);
3011 
3012 	bzero(eq, sizeof(*eq));
3013 	return (0);
3014 }
3015 
3016 static int
3017 alloc_wrq(struct adapter *sc, struct port_info *pi, struct sge_wrq *wrq,
3018     struct sysctl_oid *oid)
3019 {
3020 	int rc;
3021 	struct sysctl_ctx_list *ctx = pi ? &pi->ctx : &sc->ctx;
3022 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3023 
3024 	rc = alloc_eq(sc, pi, &wrq->eq);
3025 	if (rc)
3026 		return (rc);
3027 
3028 	wrq->adapter = sc;
3029 	STAILQ_INIT(&wrq->wr_list);
3030 
3031 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3032 	    &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
3033 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3034 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I",
3035 	    "consumer index");
3036 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx",
3037 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I",
3038 	    "producer index");
3039 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs", CTLFLAG_RD,
3040 	    &wrq->tx_wrs, "# of work requests");
3041 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "no_desc", CTLFLAG_RD,
3042 	    &wrq->no_desc, 0,
3043 	    "# of times queue ran out of hardware descriptors");
3044 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "unstalled", CTLFLAG_RD,
3045 	    &wrq->eq.unstalled, 0, "# of times queue recovered after stall");
3046 
3047 	return (rc);
3048 }
3049 
3050 static int
3051 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
3052 {
3053 	int rc;
3054 
3055 	rc = free_eq(sc, &wrq->eq);
3056 	if (rc)
3057 		return (rc);
3058 
3059 	bzero(wrq, sizeof(*wrq));
3060 	return (0);
3061 }
3062 
3063 static int
3064 alloc_txq(struct port_info *pi, struct sge_txq *txq, int idx,
3065     struct sysctl_oid *oid)
3066 {
3067 	int rc;
3068 	struct adapter *sc = pi->adapter;
3069 	struct sge_eq *eq = &txq->eq;
3070 	char name[16];
3071 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3072 
3073 	rc = alloc_eq(sc, pi, eq);
3074 	if (rc)
3075 		return (rc);
3076 
3077 	txq->ifp = pi->ifp;
3078 
3079 	txq->sdesc = malloc(eq->cap * sizeof(struct tx_sdesc), M_CXGBE,
3080 	    M_ZERO | M_WAITOK);
3081 	txq->br = buf_ring_alloc(eq->qsize, M_CXGBE, M_WAITOK, &eq->eq_lock);
3082 
3083 	rc = bus_dma_tag_create(sc->dmat, 1, 0, BUS_SPACE_MAXADDR,
3084 	    BUS_SPACE_MAXADDR, NULL, NULL, 64 * 1024, TX_SGL_SEGS,
3085 	    BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL, NULL, &txq->tx_tag);
3086 	if (rc != 0) {
3087 		device_printf(sc->dev,
3088 		    "failed to create tx DMA tag: %d\n", rc);
3089 		return (rc);
3090 	}
3091 
3092 	/*
3093 	 * We can stuff ~10 frames in an 8-descriptor txpkts WR (8 is the SGE
3094 	 * limit for any WR).  txq->no_dmamap events shouldn't occur if maps is
3095 	 * sized for the worst case.
3096 	 */
3097 	rc = t4_alloc_tx_maps(&txq->txmaps, txq->tx_tag, eq->qsize * 10 / 8,
3098 	    M_WAITOK);
3099 	if (rc != 0) {
3100 		device_printf(sc->dev, "failed to setup tx DMA maps: %d\n", rc);
3101 		return (rc);
3102 	}
3103 
3104 	snprintf(name, sizeof(name), "%d", idx);
3105 	oid = SYSCTL_ADD_NODE(&pi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3106 	    NULL, "tx queue");
3107 	children = SYSCTL_CHILDREN(oid);
3108 
3109 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3110 	    &eq->cntxt_id, 0, "SGE context id of the queue");
3111 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "cidx",
3112 	    CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I",
3113 	    "consumer index");
3114 	SYSCTL_ADD_PROC(&pi->ctx, children, OID_AUTO, "pidx",
3115 	    CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I",
3116 	    "producer index");
3117 
3118 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
3119 	    &txq->txcsum, "# of times hardware assisted with checksum");
3120 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "vlan_insertion",
3121 	    CTLFLAG_RD, &txq->vlan_insertion,
3122 	    "# of times hardware inserted 802.1Q tag");
3123 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
3124 	    &txq->tso_wrs, "# of TSO work requests");
3125 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
3126 	    &txq->imm_wrs, "# of work requests with immediate data");
3127 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
3128 	    &txq->sgl_wrs, "# of work requests with direct SGL");
3129 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
3130 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
3131 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txpkts_wrs", CTLFLAG_RD,
3132 	    &txq->txpkts_wrs, "# of txpkts work requests (multiple pkts/WR)");
3133 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "txpkts_pkts", CTLFLAG_RD,
3134 	    &txq->txpkts_pkts, "# of frames tx'd using txpkts work requests");
3135 
3136 	SYSCTL_ADD_UQUAD(&pi->ctx, children, OID_AUTO, "br_drops", CTLFLAG_RD,
3137 	    &txq->br->br_drops, "# of drops in the buf_ring for this queue");
3138 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "no_dmamap", CTLFLAG_RD,
3139 	    &txq->no_dmamap, 0, "# of times txq ran out of DMA maps");
3140 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "no_desc", CTLFLAG_RD,
3141 	    &txq->no_desc, 0, "# of times txq ran out of hardware descriptors");
3142 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "egr_update", CTLFLAG_RD,
3143 	    &eq->egr_update, 0, "egress update notifications from the SGE");
3144 	SYSCTL_ADD_UINT(&pi->ctx, children, OID_AUTO, "unstalled", CTLFLAG_RD,
3145 	    &eq->unstalled, 0, "# of times txq recovered after stall");
3146 
3147 	return (rc);
3148 }
3149 
3150 static int
3151 free_txq(struct port_info *pi, struct sge_txq *txq)
3152 {
3153 	int rc;
3154 	struct adapter *sc = pi->adapter;
3155 	struct sge_eq *eq = &txq->eq;
3156 
3157 	rc = free_eq(sc, eq);
3158 	if (rc)
3159 		return (rc);
3160 
3161 	free(txq->sdesc, M_CXGBE);
3162 
3163 	if (txq->txmaps.maps)
3164 		t4_free_tx_maps(&txq->txmaps, txq->tx_tag);
3165 
3166 	buf_ring_free(txq->br, M_CXGBE);
3167 
3168 	if (txq->tx_tag)
3169 		bus_dma_tag_destroy(txq->tx_tag);
3170 
3171 	bzero(txq, sizeof(*txq));
3172 	return (0);
3173 }
3174 
3175 static void
3176 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
3177 {
3178 	bus_addr_t *ba = arg;
3179 
3180 	KASSERT(nseg == 1,
3181 	    ("%s meant for single segment mappings only.", __func__));
3182 
3183 	*ba = error ? 0 : segs->ds_addr;
3184 }
3185 
3186 static inline bool
3187 is_new_response(const struct sge_iq *iq, struct rsp_ctrl **ctrl)
3188 {
3189 	*ctrl = (void *)((uintptr_t)iq->cdesc +
3190 	    (iq->esize - sizeof(struct rsp_ctrl)));
3191 
3192 	return (((*ctrl)->u.type_gen >> S_RSPD_GEN) == iq->gen);
3193 }
3194 
3195 static inline void
3196 iq_next(struct sge_iq *iq)
3197 {
3198 	iq->cdesc = (void *) ((uintptr_t)iq->cdesc + iq->esize);
3199 	if (__predict_false(++iq->cidx == iq->qsize - spg_len / iq->esize)) {
3200 		iq->cidx = 0;
3201 		iq->gen ^= 1;
3202 		iq->cdesc = iq->desc;
3203 	}
3204 }
3205 
3206 #define FL_HW_IDX(x) ((x) >> 3)
3207 static inline void
3208 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
3209 {
3210 	int ndesc = fl->pending / 8;
3211 	uint32_t v;
3212 
3213 	if (FL_HW_IDX(fl->pidx) == FL_HW_IDX(fl->cidx))
3214 		ndesc--;	/* hold back one credit */
3215 
3216 	if (ndesc <= 0)
3217 		return;		/* nothing to do */
3218 
3219 	v = F_DBPRIO | V_QID(fl->cntxt_id) | V_PIDX(ndesc);
3220 	if (is_t5(sc))
3221 		v |= F_DBTYPE;
3222 
3223 	wmb();
3224 
3225 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL), v);
3226 	fl->pending -= ndesc * 8;
3227 }
3228 
3229 /*
3230  * Fill up the freelist by upto nbufs and maybe ring its doorbell.
3231  *
3232  * Returns non-zero to indicate that it should be added to the list of starving
3233  * freelists.
3234  */
3235 static int
3236 refill_fl(struct adapter *sc, struct sge_fl *fl, int nbufs)
3237 {
3238 	__be64 *d = &fl->desc[fl->pidx];
3239 	struct fl_sdesc *sd = &fl->sdesc[fl->pidx];
3240 	uintptr_t pa;
3241 	caddr_t cl;
3242 	struct cluster_layout *cll = &fl->cll_def;	/* default layout */
3243 	struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
3244 	struct cluster_metadata *clm;
3245 
3246 	FL_LOCK_ASSERT_OWNED(fl);
3247 
3248 	if (nbufs > fl->needed)
3249 		nbufs = fl->needed;
3250 	nbufs -= (fl->pidx + nbufs) % 8;
3251 
3252 	while (nbufs--) {
3253 
3254 		if (sd->cl != NULL) {
3255 
3256 			if (sd->nmbuf == 0) {
3257 				/*
3258 				 * Fast recycle without involving any atomics on
3259 				 * the cluster's metadata (if the cluster has
3260 				 * metadata).  This happens when all frames
3261 				 * received in the cluster were small enough to
3262 				 * fit within a single mbuf each.
3263 				 */
3264 				fl->cl_fast_recycled++;
3265 #ifdef INVARIANTS
3266 				clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
3267 				if (clm != NULL)
3268 					MPASS(clm->refcount == 1);
3269 #endif
3270 				goto recycled_fast;
3271 			}
3272 
3273 			/*
3274 			 * Cluster is guaranteed to have metadata.  Clusters
3275 			 * without metadata always take the fast recycle path
3276 			 * when they're recycled.
3277 			 */
3278 			clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
3279 			MPASS(clm != NULL);
3280 
3281 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
3282 				fl->cl_recycled++;
3283 				goto recycled;
3284 			}
3285 			sd->cl = NULL;	/* gave up my reference */
3286 		}
3287 		MPASS(sd->cl == NULL);
3288 alloc:
3289 		cl = uma_zalloc(swz->zone, M_NOWAIT);
3290 		if (__predict_false(cl == NULL)) {
3291 			if (cll == &fl->cll_alt || fl->cll_alt.zidx == -1 ||
3292 			    fl->cll_def.zidx == fl->cll_alt.zidx)
3293 				break;
3294 
3295 			/* fall back to the safe zone */
3296 			cll = &fl->cll_alt;
3297 			swz = &sc->sge.sw_zone_info[cll->zidx];
3298 			goto alloc;
3299 		}
3300 		fl->cl_allocated++;
3301 
3302 		pa = pmap_kextract((vm_offset_t)cl);
3303 		pa += cll->region1;
3304 		sd->cl = cl;
3305 		sd->cll = *cll;
3306 		*d = htobe64(pa | cll->hwidx);
3307 		clm = cl_metadata(sc, fl, cll, cl);
3308 		if (clm != NULL) {
3309 recycled:
3310 #ifdef INVARIANTS
3311 			clm->sd = sd;
3312 #endif
3313 			clm->refcount = 1;
3314 		}
3315 		sd->nmbuf = 0;
3316 recycled_fast:
3317 		fl->pending++;
3318 		fl->needed--;
3319 		d++;
3320 		sd++;
3321 		if (__predict_false(++fl->pidx == fl->cap)) {
3322 			fl->pidx = 0;
3323 			sd = fl->sdesc;
3324 			d = fl->desc;
3325 		}
3326 	}
3327 
3328 	if (fl->pending >= 8)
3329 		ring_fl_db(sc, fl);
3330 
3331 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
3332 }
3333 
3334 /*
3335  * Attempt to refill all starving freelists.
3336  */
3337 static void
3338 refill_sfl(void *arg)
3339 {
3340 	struct adapter *sc = arg;
3341 	struct sge_fl *fl, *fl_temp;
3342 
3343 	mtx_lock(&sc->sfl_lock);
3344 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
3345 		FL_LOCK(fl);
3346 		refill_fl(sc, fl, 64);
3347 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
3348 			TAILQ_REMOVE(&sc->sfl, fl, link);
3349 			fl->flags &= ~FL_STARVING;
3350 		}
3351 		FL_UNLOCK(fl);
3352 	}
3353 
3354 	if (!TAILQ_EMPTY(&sc->sfl))
3355 		callout_schedule(&sc->sfl_callout, hz / 5);
3356 	mtx_unlock(&sc->sfl_lock);
3357 }
3358 
3359 static int
3360 alloc_fl_sdesc(struct sge_fl *fl)
3361 {
3362 
3363 	fl->sdesc = malloc(fl->cap * sizeof(struct fl_sdesc), M_CXGBE,
3364 	    M_ZERO | M_WAITOK);
3365 
3366 	return (0);
3367 }
3368 
3369 static void
3370 free_fl_sdesc(struct adapter *sc, struct sge_fl *fl)
3371 {
3372 	struct fl_sdesc *sd;
3373 	struct cluster_metadata *clm;
3374 	struct cluster_layout *cll;
3375 	int i;
3376 
3377 	sd = fl->sdesc;
3378 	for (i = 0; i < fl->cap; i++, sd++) {
3379 		if (sd->cl == NULL)
3380 			continue;
3381 
3382 		cll = &sd->cll;
3383 		clm = cl_metadata(sc, fl, cll, sd->cl);
3384 		if (sd->nmbuf == 0 ||
3385 		    (clm && atomic_fetchadd_int(&clm->refcount, -1) == 1)) {
3386 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
3387 		}
3388 		sd->cl = NULL;
3389 	}
3390 
3391 	free(fl->sdesc, M_CXGBE);
3392 	fl->sdesc = NULL;
3393 }
3394 
3395 int
3396 t4_alloc_tx_maps(struct tx_maps *txmaps, bus_dma_tag_t tx_tag, int count,
3397     int flags)
3398 {
3399 	struct tx_map *txm;
3400 	int i, rc;
3401 
3402 	txmaps->map_total = txmaps->map_avail = count;
3403 	txmaps->map_cidx = txmaps->map_pidx = 0;
3404 
3405 	txmaps->maps = malloc(count * sizeof(struct tx_map), M_CXGBE,
3406 	    M_ZERO | flags);
3407 
3408 	txm = txmaps->maps;
3409 	for (i = 0; i < count; i++, txm++) {
3410 		rc = bus_dmamap_create(tx_tag, 0, &txm->map);
3411 		if (rc != 0)
3412 			goto failed;
3413 	}
3414 
3415 	return (0);
3416 failed:
3417 	while (--i >= 0) {
3418 		txm--;
3419 		bus_dmamap_destroy(tx_tag, txm->map);
3420 	}
3421 	KASSERT(txm == txmaps->maps, ("%s: EDOOFUS", __func__));
3422 
3423 	free(txmaps->maps, M_CXGBE);
3424 	txmaps->maps = NULL;
3425 
3426 	return (rc);
3427 }
3428 
3429 void
3430 t4_free_tx_maps(struct tx_maps *txmaps, bus_dma_tag_t tx_tag)
3431 {
3432 	struct tx_map *txm;
3433 	int i;
3434 
3435 	txm = txmaps->maps;
3436 	for (i = 0; i < txmaps->map_total; i++, txm++) {
3437 
3438 		if (txm->m) {
3439 			bus_dmamap_unload(tx_tag, txm->map);
3440 			m_freem(txm->m);
3441 			txm->m = NULL;
3442 		}
3443 
3444 		bus_dmamap_destroy(tx_tag, txm->map);
3445 	}
3446 
3447 	free(txmaps->maps, M_CXGBE);
3448 	txmaps->maps = NULL;
3449 }
3450 
3451 /*
3452  * We'll do immediate data tx for non-TSO, but only when not coalescing.  We're
3453  * willing to use upto 2 hardware descriptors which means a maximum of 96 bytes
3454  * of immediate data.
3455  */
3456 #define IMM_LEN ( \
3457       2 * EQ_ESIZE \
3458     - sizeof(struct fw_eth_tx_pkt_wr) \
3459     - sizeof(struct cpl_tx_pkt_core))
3460 
3461 /*
3462  * Returns non-zero on failure, no need to cleanup anything in that case.
3463  *
3464  * Note 1: We always try to defrag the mbuf if required and return EFBIG only
3465  * if the resulting chain still won't fit in a tx descriptor.
3466  *
3467  * Note 2: We'll pullup the mbuf chain if TSO is requested and the first mbuf
3468  * does not have the TCP header in it.
3469  */
3470 static int
3471 get_pkt_sgl(struct sge_txq *txq, struct mbuf **fp, struct sgl *sgl,
3472     int sgl_only)
3473 {
3474 	struct mbuf *m = *fp;
3475 	struct tx_maps *txmaps;
3476 	struct tx_map *txm;
3477 	int rc, defragged = 0, n;
3478 
3479 	TXQ_LOCK_ASSERT_OWNED(txq);
3480 
3481 	if (m->m_pkthdr.tso_segsz)
3482 		sgl_only = 1;	/* Do not allow immediate data with LSO */
3483 
3484 start:	sgl->nsegs = 0;
3485 
3486 	if (m->m_pkthdr.len <= IMM_LEN && !sgl_only)
3487 		return (0);	/* nsegs = 0 tells caller to use imm. tx */
3488 
3489 	txmaps = &txq->txmaps;
3490 	if (txmaps->map_avail == 0) {
3491 		txq->no_dmamap++;
3492 		return (ENOMEM);
3493 	}
3494 	txm = &txmaps->maps[txmaps->map_pidx];
3495 
3496 	if (m->m_pkthdr.tso_segsz && m->m_len < 50) {
3497 		*fp = m_pullup(m, 50);
3498 		m = *fp;
3499 		if (m == NULL)
3500 			return (ENOBUFS);
3501 	}
3502 
3503 	rc = bus_dmamap_load_mbuf_sg(txq->tx_tag, txm->map, m, sgl->seg,
3504 	    &sgl->nsegs, BUS_DMA_NOWAIT);
3505 	if (rc == EFBIG && defragged == 0) {
3506 		m = m_defrag(m, M_NOWAIT);
3507 		if (m == NULL)
3508 			return (EFBIG);
3509 
3510 		defragged = 1;
3511 		*fp = m;
3512 		goto start;
3513 	}
3514 	if (rc != 0)
3515 		return (rc);
3516 
3517 	txm->m = m;
3518 	txmaps->map_avail--;
3519 	if (++txmaps->map_pidx == txmaps->map_total)
3520 		txmaps->map_pidx = 0;
3521 
3522 	KASSERT(sgl->nsegs > 0 && sgl->nsegs <= TX_SGL_SEGS,
3523 	    ("%s: bad DMA mapping (%d segments)", __func__, sgl->nsegs));
3524 
3525 	/*
3526 	 * Store the # of flits required to hold this frame's SGL in nflits.  An
3527 	 * SGL has a (ULPTX header + len0, addr0) tuple optionally followed by
3528 	 * multiple (len0 + len1, addr0, addr1) tuples.  If addr1 is not used
3529 	 * then len1 must be set to 0.
3530 	 */
3531 	n = sgl->nsegs - 1;
3532 	sgl->nflits = (3 * n) / 2 + (n & 1) + 2;
3533 
3534 	return (0);
3535 }
3536 
3537 
3538 /*
3539  * Releases all the txq resources used up in the specified sgl.
3540  */
3541 static int
3542 free_pkt_sgl(struct sge_txq *txq, struct sgl *sgl)
3543 {
3544 	struct tx_maps *txmaps;
3545 	struct tx_map *txm;
3546 
3547 	TXQ_LOCK_ASSERT_OWNED(txq);
3548 
3549 	if (sgl->nsegs == 0)
3550 		return (0);	/* didn't use any map */
3551 
3552 	txmaps = &txq->txmaps;
3553 
3554 	/* 1 pkt uses exactly 1 map, back it out */
3555 
3556 	txmaps->map_avail++;
3557 	if (txmaps->map_pidx > 0)
3558 		txmaps->map_pidx--;
3559 	else
3560 		txmaps->map_pidx = txmaps->map_total - 1;
3561 
3562 	txm = &txmaps->maps[txmaps->map_pidx];
3563 	bus_dmamap_unload(txq->tx_tag, txm->map);
3564 	txm->m = NULL;
3565 
3566 	return (0);
3567 }
3568 
3569 static int
3570 write_txpkt_wr(struct port_info *pi, struct sge_txq *txq, struct mbuf *m,
3571     struct sgl *sgl)
3572 {
3573 	struct sge_eq *eq = &txq->eq;
3574 	struct fw_eth_tx_pkt_wr *wr;
3575 	struct cpl_tx_pkt_core *cpl;
3576 	uint32_t ctrl;	/* used in many unrelated places */
3577 	uint64_t ctrl1;
3578 	int nflits, ndesc, pktlen;
3579 	struct tx_sdesc *txsd;
3580 	caddr_t dst;
3581 
3582 	TXQ_LOCK_ASSERT_OWNED(txq);
3583 
3584 	pktlen = m->m_pkthdr.len;
3585 
3586 	/*
3587 	 * Do we have enough flits to send this frame out?
3588 	 */
3589 	ctrl = sizeof(struct cpl_tx_pkt_core);
3590 	if (m->m_pkthdr.tso_segsz) {
3591 		nflits = TXPKT_LSO_WR_HDR;
3592 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
3593 	} else
3594 		nflits = TXPKT_WR_HDR;
3595 	if (sgl->nsegs > 0)
3596 		nflits += sgl->nflits;
3597 	else {
3598 		nflits += howmany(pktlen, 8);
3599 		ctrl += pktlen;
3600 	}
3601 	ndesc = howmany(nflits, 8);
3602 	if (ndesc > eq->avail)
3603 		return (ENOMEM);
3604 
3605 	/* Firmware work request header */
3606 	wr = (void *)&eq->desc[eq->pidx];
3607 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
3608 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
3609 	ctrl = V_FW_WR_LEN16(howmany(nflits, 2));
3610 	if (eq->avail == ndesc) {
3611 		if (!(eq->flags & EQ_CRFLUSHED)) {
3612 			ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ;
3613 			eq->flags |= EQ_CRFLUSHED;
3614 		}
3615 		eq->flags |= EQ_STALLED;
3616 	}
3617 
3618 	wr->equiq_to_len16 = htobe32(ctrl);
3619 	wr->r3 = 0;
3620 
3621 	if (m->m_pkthdr.tso_segsz) {
3622 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
3623 		struct ether_header *eh;
3624 		void *l3hdr;
3625 #if defined(INET) || defined(INET6)
3626 		struct tcphdr *tcp;
3627 #endif
3628 		uint16_t eh_type;
3629 
3630 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
3631 		    F_LSO_LAST_SLICE;
3632 
3633 		eh = mtod(m, struct ether_header *);
3634 		eh_type = ntohs(eh->ether_type);
3635 		if (eh_type == ETHERTYPE_VLAN) {
3636 			struct ether_vlan_header *evh = (void *)eh;
3637 
3638 			ctrl |= V_LSO_ETHHDR_LEN(1);
3639 			l3hdr = evh + 1;
3640 			eh_type = ntohs(evh->evl_proto);
3641 		} else
3642 			l3hdr = eh + 1;
3643 
3644 		switch (eh_type) {
3645 #ifdef INET6
3646 		case ETHERTYPE_IPV6:
3647 		{
3648 			struct ip6_hdr *ip6 = l3hdr;
3649 
3650 			/*
3651 			 * XXX-BZ For now we do not pretend to support
3652 			 * IPv6 extension headers.
3653 			 */
3654 			KASSERT(ip6->ip6_nxt == IPPROTO_TCP, ("%s: CSUM_TSO "
3655 			    "with ip6_nxt != TCP: %u", __func__, ip6->ip6_nxt));
3656 			tcp = (struct tcphdr *)(ip6 + 1);
3657 			ctrl |= F_LSO_IPV6;
3658 			ctrl |= V_LSO_IPHDR_LEN(sizeof(*ip6) >> 2) |
3659 			    V_LSO_TCPHDR_LEN(tcp->th_off);
3660 			break;
3661 		}
3662 #endif
3663 #ifdef INET
3664 		case ETHERTYPE_IP:
3665 		{
3666 			struct ip *ip = l3hdr;
3667 
3668 			tcp = (void *)((uintptr_t)ip + ip->ip_hl * 4);
3669 			ctrl |= V_LSO_IPHDR_LEN(ip->ip_hl) |
3670 			    V_LSO_TCPHDR_LEN(tcp->th_off);
3671 			break;
3672 		}
3673 #endif
3674 		default:
3675 			panic("%s: CSUM_TSO but no supported IP version "
3676 			    "(0x%04x)", __func__, eh_type);
3677 		}
3678 
3679 		lso->lso_ctrl = htobe32(ctrl);
3680 		lso->ipid_ofst = htobe16(0);
3681 		lso->mss = htobe16(m->m_pkthdr.tso_segsz);
3682 		lso->seqno_offset = htobe32(0);
3683 		lso->len = htobe32(pktlen);
3684 
3685 		cpl = (void *)(lso + 1);
3686 
3687 		txq->tso_wrs++;
3688 	} else
3689 		cpl = (void *)(wr + 1);
3690 
3691 	/* Checksum offload */
3692 	ctrl1 = 0;
3693 	if (!(m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO)))
3694 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
3695 	if (!(m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
3696 	    CSUM_TCP_IPV6 | CSUM_TSO)))
3697 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
3698 	if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
3699 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
3700 		txq->txcsum++;	/* some hardware assistance provided */
3701 
3702 	/* VLAN tag insertion */
3703 	if (m->m_flags & M_VLANTAG) {
3704 		ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
3705 		txq->vlan_insertion++;
3706 	}
3707 
3708 	/* CPL header */
3709 	cpl->ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3710 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf));
3711 	cpl->pack = 0;
3712 	cpl->len = htobe16(pktlen);
3713 	cpl->ctrl1 = htobe64(ctrl1);
3714 
3715 	/* Software descriptor */
3716 	txsd = &txq->sdesc[eq->pidx];
3717 	txsd->desc_used = ndesc;
3718 
3719 	eq->pending += ndesc;
3720 	eq->avail -= ndesc;
3721 	eq->pidx += ndesc;
3722 	if (eq->pidx >= eq->cap)
3723 		eq->pidx -= eq->cap;
3724 
3725 	/* SGL */
3726 	dst = (void *)(cpl + 1);
3727 	if (sgl->nsegs > 0) {
3728 		txsd->credits = 1;
3729 		txq->sgl_wrs++;
3730 		write_sgl_to_txd(eq, sgl, &dst);
3731 	} else {
3732 		txsd->credits = 0;
3733 		txq->imm_wrs++;
3734 		for (; m; m = m->m_next) {
3735 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
3736 #ifdef INVARIANTS
3737 			pktlen -= m->m_len;
3738 #endif
3739 		}
3740 #ifdef INVARIANTS
3741 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
3742 #endif
3743 
3744 	}
3745 
3746 	txq->txpkt_wrs++;
3747 	return (0);
3748 }
3749 
3750 /*
3751  * Returns 0 to indicate that m has been accepted into a coalesced tx work
3752  * request.  It has either been folded into txpkts or txpkts was flushed and m
3753  * has started a new coalesced work request (as the first frame in a fresh
3754  * txpkts).
3755  *
3756  * Returns non-zero to indicate a failure - caller is responsible for
3757  * transmitting m, if there was anything in txpkts it has been flushed.
3758  */
3759 static int
3760 add_to_txpkts(struct port_info *pi, struct sge_txq *txq, struct txpkts *txpkts,
3761     struct mbuf *m, struct sgl *sgl)
3762 {
3763 	struct sge_eq *eq = &txq->eq;
3764 	int can_coalesce;
3765 	struct tx_sdesc *txsd;
3766 	int flits;
3767 
3768 	TXQ_LOCK_ASSERT_OWNED(txq);
3769 
3770 	KASSERT(sgl->nsegs, ("%s: can't coalesce imm data", __func__));
3771 
3772 	if (txpkts->npkt > 0) {
3773 		flits = TXPKTS_PKT_HDR + sgl->nflits;
3774 		can_coalesce = m->m_pkthdr.tso_segsz == 0 &&
3775 		    txpkts->nflits + flits <= TX_WR_FLITS &&
3776 		    txpkts->nflits + flits <= eq->avail * 8 &&
3777 		    txpkts->plen + m->m_pkthdr.len < 65536;
3778 
3779 		if (can_coalesce) {
3780 			txpkts->npkt++;
3781 			txpkts->nflits += flits;
3782 			txpkts->plen += m->m_pkthdr.len;
3783 
3784 			txsd = &txq->sdesc[eq->pidx];
3785 			txsd->credits++;
3786 
3787 			return (0);
3788 		}
3789 
3790 		/*
3791 		 * Couldn't coalesce m into txpkts.  The first order of business
3792 		 * is to send txpkts on its way.  Then we'll revisit m.
3793 		 */
3794 		write_txpkts_wr(txq, txpkts);
3795 	}
3796 
3797 	/*
3798 	 * Check if we can start a new coalesced tx work request with m as
3799 	 * the first packet in it.
3800 	 */
3801 
3802 	KASSERT(txpkts->npkt == 0, ("%s: txpkts not empty", __func__));
3803 
3804 	flits = TXPKTS_WR_HDR + sgl->nflits;
3805 	can_coalesce = m->m_pkthdr.tso_segsz == 0 &&
3806 	    flits <= eq->avail * 8 && flits <= TX_WR_FLITS;
3807 
3808 	if (can_coalesce == 0)
3809 		return (EINVAL);
3810 
3811 	/*
3812 	 * Start a fresh coalesced tx WR with m as the first frame in it.
3813 	 */
3814 	txpkts->npkt = 1;
3815 	txpkts->nflits = flits;
3816 	txpkts->flitp = &eq->desc[eq->pidx].flit[2];
3817 	txpkts->plen = m->m_pkthdr.len;
3818 
3819 	txsd = &txq->sdesc[eq->pidx];
3820 	txsd->credits = 1;
3821 
3822 	return (0);
3823 }
3824 
3825 /*
3826  * Note that write_txpkts_wr can never run out of hardware descriptors (but
3827  * write_txpkt_wr can).  add_to_txpkts ensures that a frame is accepted for
3828  * coalescing only if sufficient hardware descriptors are available.
3829  */
3830 static void
3831 write_txpkts_wr(struct sge_txq *txq, struct txpkts *txpkts)
3832 {
3833 	struct sge_eq *eq = &txq->eq;
3834 	struct fw_eth_tx_pkts_wr *wr;
3835 	struct tx_sdesc *txsd;
3836 	uint32_t ctrl;
3837 	int ndesc;
3838 
3839 	TXQ_LOCK_ASSERT_OWNED(txq);
3840 
3841 	ndesc = howmany(txpkts->nflits, 8);
3842 
3843 	wr = (void *)&eq->desc[eq->pidx];
3844 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
3845 	ctrl = V_FW_WR_LEN16(howmany(txpkts->nflits, 2));
3846 	if (eq->avail == ndesc) {
3847 		if (!(eq->flags & EQ_CRFLUSHED)) {
3848 			ctrl |= F_FW_WR_EQUEQ | F_FW_WR_EQUIQ;
3849 			eq->flags |= EQ_CRFLUSHED;
3850 		}
3851 		eq->flags |= EQ_STALLED;
3852 	}
3853 	wr->equiq_to_len16 = htobe32(ctrl);
3854 	wr->plen = htobe16(txpkts->plen);
3855 	wr->npkt = txpkts->npkt;
3856 	wr->r3 = wr->type = 0;
3857 
3858 	/* Everything else already written */
3859 
3860 	txsd = &txq->sdesc[eq->pidx];
3861 	txsd->desc_used = ndesc;
3862 
3863 	KASSERT(eq->avail >= ndesc, ("%s: out of descriptors", __func__));
3864 
3865 	eq->pending += ndesc;
3866 	eq->avail -= ndesc;
3867 	eq->pidx += ndesc;
3868 	if (eq->pidx >= eq->cap)
3869 		eq->pidx -= eq->cap;
3870 
3871 	txq->txpkts_pkts += txpkts->npkt;
3872 	txq->txpkts_wrs++;
3873 	txpkts->npkt = 0;	/* emptied */
3874 }
3875 
3876 static inline void
3877 write_ulp_cpl_sgl(struct port_info *pi, struct sge_txq *txq,
3878     struct txpkts *txpkts, struct mbuf *m, struct sgl *sgl)
3879 {
3880 	struct ulp_txpkt *ulpmc;
3881 	struct ulptx_idata *ulpsc;
3882 	struct cpl_tx_pkt_core *cpl;
3883 	struct sge_eq *eq = &txq->eq;
3884 	uintptr_t flitp, start, end;
3885 	uint64_t ctrl;
3886 	caddr_t dst;
3887 
3888 	KASSERT(txpkts->npkt > 0, ("%s: txpkts is empty", __func__));
3889 
3890 	start = (uintptr_t)eq->desc;
3891 	end = (uintptr_t)eq->spg;
3892 
3893 	/* Checksum offload */
3894 	ctrl = 0;
3895 	if (!(m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO)))
3896 		ctrl |= F_TXPKT_IPCSUM_DIS;
3897 	if (!(m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
3898 	    CSUM_TCP_IPV6 | CSUM_TSO)))
3899 		ctrl |= F_TXPKT_L4CSUM_DIS;
3900 	if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
3901 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
3902 		txq->txcsum++;	/* some hardware assistance provided */
3903 
3904 	/* VLAN tag insertion */
3905 	if (m->m_flags & M_VLANTAG) {
3906 		ctrl |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
3907 		txq->vlan_insertion++;
3908 	}
3909 
3910 	/*
3911 	 * The previous packet's SGL must have ended at a 16 byte boundary (this
3912 	 * is required by the firmware/hardware).  It follows that flitp cannot
3913 	 * wrap around between the ULPTX master command and ULPTX subcommand (8
3914 	 * bytes each), and that it can not wrap around in the middle of the
3915 	 * cpl_tx_pkt_core either.
3916 	 */
3917 	flitp = (uintptr_t)txpkts->flitp;
3918 	KASSERT((flitp & 0xf) == 0,
3919 	    ("%s: last SGL did not end at 16 byte boundary: %p",
3920 	    __func__, txpkts->flitp));
3921 
3922 	/* ULP master command */
3923 	ulpmc = (void *)flitp;
3924 	ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0) |
3925 	    V_ULP_TXPKT_FID(eq->iqid));
3926 	ulpmc->len = htonl(howmany(sizeof(*ulpmc) + sizeof(*ulpsc) +
3927 	    sizeof(*cpl) + 8 * sgl->nflits, 16));
3928 
3929 	/* ULP subcommand */
3930 	ulpsc = (void *)(ulpmc + 1);
3931 	ulpsc->cmd_more = htobe32(V_ULPTX_CMD((u32)ULP_TX_SC_IMM) |
3932 	    F_ULP_TX_SC_MORE);
3933 	ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
3934 
3935 	flitp += sizeof(*ulpmc) + sizeof(*ulpsc);
3936 	if (flitp == end)
3937 		flitp = start;
3938 
3939 	/* CPL_TX_PKT */
3940 	cpl = (void *)flitp;
3941 	cpl->ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3942 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(pi->adapter->pf));
3943 	cpl->pack = 0;
3944 	cpl->len = htobe16(m->m_pkthdr.len);
3945 	cpl->ctrl1 = htobe64(ctrl);
3946 
3947 	flitp += sizeof(*cpl);
3948 	if (flitp == end)
3949 		flitp = start;
3950 
3951 	/* SGL for this frame */
3952 	dst = (caddr_t)flitp;
3953 	txpkts->nflits += write_sgl_to_txd(eq, sgl, &dst);
3954 	txpkts->flitp = (void *)dst;
3955 
3956 	KASSERT(((uintptr_t)dst & 0xf) == 0,
3957 	    ("%s: SGL ends at %p (not a 16 byte boundary)", __func__, dst));
3958 }
3959 
3960 /*
3961  * If the SGL ends on an address that is not 16 byte aligned, this function will
3962  * add a 0 filled flit at the end.  It returns 1 in that case.
3963  */
3964 static int
3965 write_sgl_to_txd(struct sge_eq *eq, struct sgl *sgl, caddr_t *to)
3966 {
3967 	__be64 *flitp, *end;
3968 	struct ulptx_sgl *usgl;
3969 	bus_dma_segment_t *seg;
3970 	int i, padded;
3971 
3972 	KASSERT(sgl->nsegs > 0 && sgl->nflits > 0,
3973 	    ("%s: bad SGL - nsegs=%d, nflits=%d",
3974 	    __func__, sgl->nsegs, sgl->nflits));
3975 
3976 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
3977 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
3978 
3979 	flitp = (__be64 *)(*to);
3980 	end = flitp + sgl->nflits;
3981 	seg = &sgl->seg[0];
3982 	usgl = (void *)flitp;
3983 
3984 	/*
3985 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
3986 	 * ring, so we're at least 16 bytes away from the status page.  There is
3987 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
3988 	 */
3989 
3990 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
3991 	    V_ULPTX_NSGE(sgl->nsegs));
3992 	usgl->len0 = htobe32(seg->ds_len);
3993 	usgl->addr0 = htobe64(seg->ds_addr);
3994 	seg++;
3995 
3996 	if ((uintptr_t)end <= (uintptr_t)eq->spg) {
3997 
3998 		/* Won't wrap around at all */
3999 
4000 		for (i = 0; i < sgl->nsegs - 1; i++, seg++) {
4001 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ds_len);
4002 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ds_addr);
4003 		}
4004 		if (i & 1)
4005 			usgl->sge[i / 2].len[1] = htobe32(0);
4006 	} else {
4007 
4008 		/* Will wrap somewhere in the rest of the SGL */
4009 
4010 		/* 2 flits already written, write the rest flit by flit */
4011 		flitp = (void *)(usgl + 1);
4012 		for (i = 0; i < sgl->nflits - 2; i++) {
4013 			if ((uintptr_t)flitp == (uintptr_t)eq->spg)
4014 				flitp = (void *)eq->desc;
4015 			*flitp++ = get_flit(seg, sgl->nsegs - 1, i);
4016 		}
4017 		end = flitp;
4018 	}
4019 
4020 	if ((uintptr_t)end & 0xf) {
4021 		*(uint64_t *)end = 0;
4022 		end++;
4023 		padded = 1;
4024 	} else
4025 		padded = 0;
4026 
4027 	if ((uintptr_t)end == (uintptr_t)eq->spg)
4028 		*to = (void *)eq->desc;
4029 	else
4030 		*to = (void *)end;
4031 
4032 	return (padded);
4033 }
4034 
4035 static inline void
4036 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
4037 {
4038 	if (__predict_true((uintptr_t)(*to) + len <= (uintptr_t)eq->spg)) {
4039 		bcopy(from, *to, len);
4040 		(*to) += len;
4041 	} else {
4042 		int portion = (uintptr_t)eq->spg - (uintptr_t)(*to);
4043 
4044 		bcopy(from, *to, portion);
4045 		from += portion;
4046 		portion = len - portion;	/* remaining */
4047 		bcopy(from, (void *)eq->desc, portion);
4048 		(*to) = (caddr_t)eq->desc + portion;
4049 	}
4050 }
4051 
4052 static inline void
4053 ring_eq_db(struct adapter *sc, struct sge_eq *eq)
4054 {
4055 	u_int db, pending;
4056 
4057 	db = eq->doorbells;
4058 	pending = eq->pending;
4059 	if (pending > 1)
4060 		clrbit(&db, DOORBELL_WCWR);
4061 	eq->pending = 0;
4062 	wmb();
4063 
4064 	switch (ffs(db) - 1) {
4065 	case DOORBELL_UDB:
4066 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(pending));
4067 		return;
4068 
4069 	case DOORBELL_WCWR: {
4070 		volatile uint64_t *dst, *src;
4071 		int i;
4072 
4073 		/*
4074 		 * Queues whose 128B doorbell segment fits in the page do not
4075 		 * use relative qid (udb_qid is always 0).  Only queues with
4076 		 * doorbell segments can do WCWR.
4077 		 */
4078 		KASSERT(eq->udb_qid == 0 && pending == 1,
4079 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
4080 		    __func__, eq->doorbells, pending, eq->pidx, eq));
4081 
4082 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
4083 		    UDBS_DB_OFFSET);
4084 		i = eq->pidx ? eq->pidx - 1 : eq->cap - 1;
4085 		src = (void *)&eq->desc[i];
4086 		while (src != (void *)&eq->desc[i + 1])
4087 			*dst++ = *src++;
4088 		wmb();
4089 		return;
4090 	}
4091 
4092 	case DOORBELL_UDBWC:
4093 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(pending));
4094 		wmb();
4095 		return;
4096 
4097 	case DOORBELL_KDB:
4098 		t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
4099 		    V_QID(eq->cntxt_id) | V_PIDX(pending));
4100 		return;
4101 	}
4102 }
4103 
4104 static inline int
4105 reclaimable(struct sge_eq *eq)
4106 {
4107 	unsigned int cidx;
4108 
4109 	cidx = eq->spg->cidx;	/* stable snapshot */
4110 	cidx = be16toh(cidx);
4111 
4112 	if (cidx >= eq->cidx)
4113 		return (cidx - eq->cidx);
4114 	else
4115 		return (cidx + eq->cap - eq->cidx);
4116 }
4117 
4118 /*
4119  * There are "can_reclaim" tx descriptors ready to be reclaimed.  Reclaim as
4120  * many as possible but stop when there are around "n" mbufs to free.
4121  *
4122  * The actual number reclaimed is provided as the return value.
4123  */
4124 static int
4125 reclaim_tx_descs(struct sge_txq *txq, int can_reclaim, int n)
4126 {
4127 	struct tx_sdesc *txsd;
4128 	struct tx_maps *txmaps;
4129 	struct tx_map *txm;
4130 	unsigned int reclaimed, maps;
4131 	struct sge_eq *eq = &txq->eq;
4132 
4133 	TXQ_LOCK_ASSERT_OWNED(txq);
4134 
4135 	if (can_reclaim == 0)
4136 		can_reclaim = reclaimable(eq);
4137 
4138 	maps = reclaimed = 0;
4139 	while (can_reclaim && maps < n) {
4140 		int ndesc;
4141 
4142 		txsd = &txq->sdesc[eq->cidx];
4143 		ndesc = txsd->desc_used;
4144 
4145 		/* Firmware doesn't return "partial" credits. */
4146 		KASSERT(can_reclaim >= ndesc,
4147 		    ("%s: unexpected number of credits: %d, %d",
4148 		    __func__, can_reclaim, ndesc));
4149 
4150 		maps += txsd->credits;
4151 
4152 		reclaimed += ndesc;
4153 		can_reclaim -= ndesc;
4154 
4155 		eq->cidx += ndesc;
4156 		if (__predict_false(eq->cidx >= eq->cap))
4157 			eq->cidx -= eq->cap;
4158 	}
4159 
4160 	txmaps = &txq->txmaps;
4161 	txm = &txmaps->maps[txmaps->map_cidx];
4162 	if (maps)
4163 		prefetch(txm->m);
4164 
4165 	eq->avail += reclaimed;
4166 	KASSERT(eq->avail < eq->cap,	/* avail tops out at (cap - 1) */
4167 	    ("%s: too many descriptors available", __func__));
4168 
4169 	txmaps->map_avail += maps;
4170 	KASSERT(txmaps->map_avail <= txmaps->map_total,
4171 	    ("%s: too many maps available", __func__));
4172 
4173 	while (maps--) {
4174 		struct tx_map *next;
4175 
4176 		next = txm + 1;
4177 		if (__predict_false(txmaps->map_cidx + 1 == txmaps->map_total))
4178 			next = txmaps->maps;
4179 		prefetch(next->m);
4180 
4181 		bus_dmamap_unload(txq->tx_tag, txm->map);
4182 		m_freem(txm->m);
4183 		txm->m = NULL;
4184 
4185 		txm = next;
4186 		if (__predict_false(++txmaps->map_cidx == txmaps->map_total))
4187 			txmaps->map_cidx = 0;
4188 	}
4189 
4190 	return (reclaimed);
4191 }
4192 
4193 static void
4194 write_eqflush_wr(struct sge_eq *eq)
4195 {
4196 	struct fw_eq_flush_wr *wr;
4197 
4198 	EQ_LOCK_ASSERT_OWNED(eq);
4199 	KASSERT(eq->avail > 0, ("%s: no descriptors left.", __func__));
4200 	KASSERT(!(eq->flags & EQ_CRFLUSHED), ("%s: flushed already", __func__));
4201 
4202 	wr = (void *)&eq->desc[eq->pidx];
4203 	bzero(wr, sizeof(*wr));
4204 	wr->opcode = FW_EQ_FLUSH_WR;
4205 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(sizeof(*wr) / 16) |
4206 	    F_FW_WR_EQUEQ | F_FW_WR_EQUIQ);
4207 
4208 	eq->flags |= (EQ_CRFLUSHED | EQ_STALLED);
4209 	eq->pending++;
4210 	eq->avail--;
4211 	if (++eq->pidx == eq->cap)
4212 		eq->pidx = 0;
4213 }
4214 
4215 static __be64
4216 get_flit(bus_dma_segment_t *sgl, int nsegs, int idx)
4217 {
4218 	int i = (idx / 3) * 2;
4219 
4220 	switch (idx % 3) {
4221 	case 0: {
4222 		__be64 rc;
4223 
4224 		rc = htobe32(sgl[i].ds_len);
4225 		if (i + 1 < nsegs)
4226 			rc |= (uint64_t)htobe32(sgl[i + 1].ds_len) << 32;
4227 
4228 		return (rc);
4229 	}
4230 	case 1:
4231 		return htobe64(sgl[i].ds_addr);
4232 	case 2:
4233 		return htobe64(sgl[i + 1].ds_addr);
4234 	}
4235 
4236 	return (0);
4237 }
4238 
4239 static void
4240 find_best_refill_source(struct adapter *sc, struct sge_fl *fl, int maxp)
4241 {
4242 	int8_t zidx, hwidx, idx;
4243 	uint16_t region1, region3;
4244 	int spare, spare_needed, n;
4245 	struct sw_zone_info *swz;
4246 	struct hw_buf_info *hwb, *hwb_list = &sc->sge.hw_buf_info[0];
4247 
4248 	/*
4249 	 * Buffer Packing: Look for PAGE_SIZE or larger zone which has a bufsize
4250 	 * large enough for the max payload and cluster metadata.  Otherwise
4251 	 * settle for the largest bufsize that leaves enough room in the cluster
4252 	 * for metadata.
4253 	 *
4254 	 * Without buffer packing: Look for the smallest zone which has a
4255 	 * bufsize large enough for the max payload.  Settle for the largest
4256 	 * bufsize available if there's nothing big enough for max payload.
4257 	 */
4258 	spare_needed = fl->flags & FL_BUF_PACKING ? CL_METADATA_SIZE : 0;
4259 	swz = &sc->sge.sw_zone_info[0];
4260 	hwidx = -1;
4261 	for (zidx = 0; zidx < SW_ZONE_SIZES; zidx++, swz++) {
4262 		if (swz->size > largest_rx_cluster) {
4263 			if (__predict_true(hwidx != -1))
4264 				break;
4265 
4266 			/*
4267 			 * This is a misconfiguration.  largest_rx_cluster is
4268 			 * preventing us from finding a refill source.  See
4269 			 * dev.t5nex.<n>.buffer_sizes to figure out why.
4270 			 */
4271 			device_printf(sc->dev, "largest_rx_cluster=%u leaves no"
4272 			    " refill source for fl %p (dma %u).  Ignored.\n",
4273 			    largest_rx_cluster, fl, maxp);
4274 		}
4275 		for (idx = swz->head_hwidx; idx != -1; idx = hwb->next) {
4276 			hwb = &hwb_list[idx];
4277 			spare = swz->size - hwb->size;
4278 			if (spare < spare_needed)
4279 				continue;
4280 
4281 			hwidx = idx;		/* best option so far */
4282 			if (hwb->size >= maxp) {
4283 
4284 				if ((fl->flags & FL_BUF_PACKING) == 0)
4285 					goto done; /* stop looking (not packing) */
4286 
4287 				if (swz->size >= safest_rx_cluster)
4288 					goto done; /* stop looking (packing) */
4289 			}
4290 			break;		/* keep looking, next zone */
4291 		}
4292 	}
4293 done:
4294 	/* A usable hwidx has been located. */
4295 	MPASS(hwidx != -1);
4296 	hwb = &hwb_list[hwidx];
4297 	zidx = hwb->zidx;
4298 	swz = &sc->sge.sw_zone_info[zidx];
4299 	region1 = 0;
4300 	region3 = swz->size - hwb->size;
4301 
4302 	/*
4303 	 * Stay within this zone and see if there is a better match when mbuf
4304 	 * inlining is allowed.  Remember that the hwidx's are sorted in
4305 	 * decreasing order of size (so in increasing order of spare area).
4306 	 */
4307 	for (idx = hwidx; idx != -1; idx = hwb->next) {
4308 		hwb = &hwb_list[idx];
4309 		spare = swz->size - hwb->size;
4310 
4311 		if (allow_mbufs_in_cluster == 0 || hwb->size < maxp)
4312 			break;
4313 		if (spare < CL_METADATA_SIZE + MSIZE)
4314 			continue;
4315 		n = (spare - CL_METADATA_SIZE) / MSIZE;
4316 		if (n > howmany(hwb->size, maxp))
4317 			break;
4318 
4319 		hwidx = idx;
4320 		if (fl->flags & FL_BUF_PACKING) {
4321 			region1 = n * MSIZE;
4322 			region3 = spare - region1;
4323 		} else {
4324 			region1 = MSIZE;
4325 			region3 = spare - region1;
4326 			break;
4327 		}
4328 	}
4329 
4330 	KASSERT(zidx >= 0 && zidx < SW_ZONE_SIZES,
4331 	    ("%s: bad zone %d for fl %p, maxp %d", __func__, zidx, fl, maxp));
4332 	KASSERT(hwidx >= 0 && hwidx <= SGE_FLBUF_SIZES,
4333 	    ("%s: bad hwidx %d for fl %p, maxp %d", __func__, hwidx, fl, maxp));
4334 	KASSERT(region1 + sc->sge.hw_buf_info[hwidx].size + region3 ==
4335 	    sc->sge.sw_zone_info[zidx].size,
4336 	    ("%s: bad buffer layout for fl %p, maxp %d. "
4337 		"cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
4338 		sc->sge.sw_zone_info[zidx].size, region1,
4339 		sc->sge.hw_buf_info[hwidx].size, region3));
4340 	if (fl->flags & FL_BUF_PACKING || region1 > 0) {
4341 		KASSERT(region3 >= CL_METADATA_SIZE,
4342 		    ("%s: no room for metadata.  fl %p, maxp %d; "
4343 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
4344 		    sc->sge.sw_zone_info[zidx].size, region1,
4345 		    sc->sge.hw_buf_info[hwidx].size, region3));
4346 		KASSERT(region1 % MSIZE == 0,
4347 		    ("%s: bad mbuf region for fl %p, maxp %d. "
4348 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
4349 		    sc->sge.sw_zone_info[zidx].size, region1,
4350 		    sc->sge.hw_buf_info[hwidx].size, region3));
4351 	}
4352 
4353 	fl->cll_def.zidx = zidx;
4354 	fl->cll_def.hwidx = hwidx;
4355 	fl->cll_def.region1 = region1;
4356 	fl->cll_def.region3 = region3;
4357 }
4358 
4359 static void
4360 find_safe_refill_source(struct adapter *sc, struct sge_fl *fl)
4361 {
4362 	struct sge *s = &sc->sge;
4363 	struct hw_buf_info *hwb;
4364 	struct sw_zone_info *swz;
4365 	int spare;
4366 	int8_t hwidx;
4367 
4368 	if (fl->flags & FL_BUF_PACKING)
4369 		hwidx = s->safe_hwidx2;	/* with room for metadata */
4370 	else if (allow_mbufs_in_cluster && s->safe_hwidx2 != -1) {
4371 		hwidx = s->safe_hwidx2;
4372 		hwb = &s->hw_buf_info[hwidx];
4373 		swz = &s->sw_zone_info[hwb->zidx];
4374 		spare = swz->size - hwb->size;
4375 
4376 		/* no good if there isn't room for an mbuf as well */
4377 		if (spare < CL_METADATA_SIZE + MSIZE)
4378 			hwidx = s->safe_hwidx1;
4379 	} else
4380 		hwidx = s->safe_hwidx1;
4381 
4382 	if (hwidx == -1) {
4383 		/* No fallback source */
4384 		fl->cll_alt.hwidx = -1;
4385 		fl->cll_alt.zidx = -1;
4386 
4387 		return;
4388 	}
4389 
4390 	hwb = &s->hw_buf_info[hwidx];
4391 	swz = &s->sw_zone_info[hwb->zidx];
4392 	spare = swz->size - hwb->size;
4393 	fl->cll_alt.hwidx = hwidx;
4394 	fl->cll_alt.zidx = hwb->zidx;
4395 	if (allow_mbufs_in_cluster)
4396 		fl->cll_alt.region1 = ((spare - CL_METADATA_SIZE) / MSIZE) * MSIZE;
4397 	else
4398 		fl->cll_alt.region1 = 0;
4399 	fl->cll_alt.region3 = spare - fl->cll_alt.region1;
4400 }
4401 
4402 static void
4403 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
4404 {
4405 	mtx_lock(&sc->sfl_lock);
4406 	FL_LOCK(fl);
4407 	if ((fl->flags & FL_DOOMED) == 0) {
4408 		fl->flags |= FL_STARVING;
4409 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
4410 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
4411 	}
4412 	FL_UNLOCK(fl);
4413 	mtx_unlock(&sc->sfl_lock);
4414 }
4415 
4416 static int
4417 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
4418     struct mbuf *m)
4419 {
4420 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
4421 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
4422 	struct adapter *sc = iq->adapter;
4423 	struct sge *s = &sc->sge;
4424 	struct sge_eq *eq;
4425 
4426 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
4427 	    rss->opcode));
4428 
4429 	eq = s->eqmap[qid - s->eq_start];
4430 	EQ_LOCK(eq);
4431 	KASSERT(eq->flags & EQ_CRFLUSHED,
4432 	    ("%s: unsolicited egress update", __func__));
4433 	eq->flags &= ~EQ_CRFLUSHED;
4434 	eq->egr_update++;
4435 
4436 	if (__predict_false(eq->flags & EQ_DOOMED))
4437 		wakeup_one(eq);
4438 	else if (eq->flags & EQ_STALLED && can_resume_tx(eq))
4439 		taskqueue_enqueue(sc->tq[eq->tx_chan], &eq->tx_task);
4440 	EQ_UNLOCK(eq);
4441 
4442 	return (0);
4443 }
4444 
4445 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
4446 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
4447     offsetof(struct cpl_fw6_msg, data));
4448 
4449 static int
4450 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
4451 {
4452 	struct adapter *sc = iq->adapter;
4453 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
4454 
4455 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
4456 	    rss->opcode));
4457 
4458 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
4459 		const struct rss_header *rss2;
4460 
4461 		rss2 = (const struct rss_header *)&cpl->data[0];
4462 		return (sc->cpl_handler[rss2->opcode](iq, rss2, m));
4463 	}
4464 
4465 	return (sc->fw_msg_handler[cpl->type](sc, &cpl->data[0]));
4466 }
4467 
4468 static int
4469 sysctl_uint16(SYSCTL_HANDLER_ARGS)
4470 {
4471 	uint16_t *id = arg1;
4472 	int i = *id;
4473 
4474 	return sysctl_handle_int(oidp, &i, 0, req);
4475 }
4476 
4477 static int
4478 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
4479 {
4480 	struct sge *s = arg1;
4481 	struct hw_buf_info *hwb = &s->hw_buf_info[0];
4482 	struct sw_zone_info *swz = &s->sw_zone_info[0];
4483 	int i, rc;
4484 	struct sbuf sb;
4485 	char c;
4486 
4487 	sbuf_new(&sb, NULL, 32, SBUF_AUTOEXTEND);
4488 	for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
4489 		if (hwb->zidx >= 0 && swz[hwb->zidx].size <= largest_rx_cluster)
4490 			c = '*';
4491 		else
4492 			c = '\0';
4493 
4494 		sbuf_printf(&sb, "%u%c ", hwb->size, c);
4495 	}
4496 	sbuf_trim(&sb);
4497 	sbuf_finish(&sb);
4498 	rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
4499 	sbuf_delete(&sb);
4500 	return (rc);
4501 }
4502