xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision 43faedc1339a9624c7acedb7f3e5624e64da5b99)
1 /*-
2  * Copyright (c) 2011 Chelsio Communications, Inc.
3  * All rights reserved.
4  * Written by: Navdeep Parhar <np@FreeBSD.org>
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 
34 #include <sys/types.h>
35 #include <sys/eventhandler.h>
36 #include <sys/mbuf.h>
37 #include <sys/socket.h>
38 #include <sys/kernel.h>
39 #include <sys/malloc.h>
40 #include <sys/queue.h>
41 #include <sys/sbuf.h>
42 #include <sys/taskqueue.h>
43 #include <sys/time.h>
44 #include <sys/sglist.h>
45 #include <sys/sysctl.h>
46 #include <sys/smp.h>
47 #include <sys/counter.h>
48 #include <net/bpf.h>
49 #include <net/ethernet.h>
50 #include <net/if.h>
51 #include <net/if_vlan_var.h>
52 #include <netinet/in.h>
53 #include <netinet/ip.h>
54 #include <netinet/ip6.h>
55 #include <netinet/tcp.h>
56 #include <machine/md_var.h>
57 #include <vm/vm.h>
58 #include <vm/pmap.h>
59 #ifdef DEV_NETMAP
60 #include <machine/bus.h>
61 #include <sys/selinfo.h>
62 #include <net/if_var.h>
63 #include <net/netmap.h>
64 #include <dev/netmap/netmap_kern.h>
65 #endif
66 
67 #include "common/common.h"
68 #include "common/t4_regs.h"
69 #include "common/t4_regs_values.h"
70 #include "common/t4_msg.h"
71 #include "t4_mp_ring.h"
72 
73 #ifdef T4_PKT_TIMESTAMP
74 #define RX_COPY_THRESHOLD (MINCLSIZE - 8)
75 #else
76 #define RX_COPY_THRESHOLD MINCLSIZE
77 #endif
78 
79 /*
80  * Ethernet frames are DMA'd at this byte offset into the freelist buffer.
81  * 0-7 are valid values.
82  */
83 int fl_pktshift = 2;
84 TUNABLE_INT("hw.cxgbe.fl_pktshift", &fl_pktshift);
85 
86 /*
87  * Pad ethernet payload up to this boundary.
88  * -1: driver should figure out a good value.
89  *  0: disable padding.
90  *  Any power of 2 from 32 to 4096 (both inclusive) is also a valid value.
91  */
92 int fl_pad = -1;
93 TUNABLE_INT("hw.cxgbe.fl_pad", &fl_pad);
94 
95 /*
96  * Status page length.
97  * -1: driver should figure out a good value.
98  *  64 or 128 are the only other valid values.
99  */
100 int spg_len = -1;
101 TUNABLE_INT("hw.cxgbe.spg_len", &spg_len);
102 
103 /*
104  * Congestion drops.
105  * -1: no congestion feedback (not recommended).
106  *  0: backpressure the channel instead of dropping packets right away.
107  *  1: no backpressure, drop packets for the congested queue immediately.
108  */
109 static int cong_drop = 0;
110 TUNABLE_INT("hw.cxgbe.cong_drop", &cong_drop);
111 
112 /*
113  * Deliver multiple frames in the same free list buffer if they fit.
114  * -1: let the driver decide whether to enable buffer packing or not.
115  *  0: disable buffer packing.
116  *  1: enable buffer packing.
117  */
118 static int buffer_packing = -1;
119 TUNABLE_INT("hw.cxgbe.buffer_packing", &buffer_packing);
120 
121 /*
122  * Start next frame in a packed buffer at this boundary.
123  * -1: driver should figure out a good value.
124  * T4: driver will ignore this and use the same value as fl_pad above.
125  * T5: 16, or a power of 2 from 64 to 4096 (both inclusive) is a valid value.
126  */
127 static int fl_pack = -1;
128 TUNABLE_INT("hw.cxgbe.fl_pack", &fl_pack);
129 
130 /*
131  * Allow the driver to create mbuf(s) in a cluster allocated for rx.
132  * 0: never; always allocate mbufs from the zone_mbuf UMA zone.
133  * 1: ok to create mbuf(s) within a cluster if there is room.
134  */
135 static int allow_mbufs_in_cluster = 1;
136 TUNABLE_INT("hw.cxgbe.allow_mbufs_in_cluster", &allow_mbufs_in_cluster);
137 
138 /*
139  * Largest rx cluster size that the driver is allowed to allocate.
140  */
141 static int largest_rx_cluster = MJUM16BYTES;
142 TUNABLE_INT("hw.cxgbe.largest_rx_cluster", &largest_rx_cluster);
143 
144 /*
145  * Size of cluster allocation that's most likely to succeed.  The driver will
146  * fall back to this size if it fails to allocate clusters larger than this.
147  */
148 static int safest_rx_cluster = PAGE_SIZE;
149 TUNABLE_INT("hw.cxgbe.safest_rx_cluster", &safest_rx_cluster);
150 
151 struct txpkts {
152 	u_int wr_type;		/* type 0 or type 1 */
153 	u_int npkt;		/* # of packets in this work request */
154 	u_int plen;		/* total payload (sum of all packets) */
155 	u_int len16;		/* # of 16B pieces used by this work request */
156 };
157 
158 /* A packet's SGL.  This + m_pkthdr has all info needed for tx */
159 struct sgl {
160 	struct sglist sg;
161 	struct sglist_seg seg[TX_SGL_SEGS];
162 };
163 
164 static int service_iq(struct sge_iq *, int);
165 static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t);
166 static int t4_eth_rx(struct sge_iq *, const struct rss_header *, struct mbuf *);
167 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int);
168 static inline void init_fl(struct adapter *, struct sge_fl *, int, int, char *);
169 static inline void init_eq(struct sge_eq *, int, int, uint8_t, uint16_t,
170     char *);
171 static int alloc_ring(struct adapter *, size_t, bus_dma_tag_t *, bus_dmamap_t *,
172     bus_addr_t *, void **);
173 static int free_ring(struct adapter *, bus_dma_tag_t, bus_dmamap_t, bus_addr_t,
174     void *);
175 static int alloc_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *,
176     int, int);
177 static int free_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *);
178 static void add_fl_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
179     struct sge_fl *);
180 static int alloc_fwq(struct adapter *);
181 static int free_fwq(struct adapter *);
182 static int alloc_mgmtq(struct adapter *);
183 static int free_mgmtq(struct adapter *);
184 static int alloc_rxq(struct vi_info *, struct sge_rxq *, int, int,
185     struct sysctl_oid *);
186 static int free_rxq(struct vi_info *, struct sge_rxq *);
187 #ifdef TCP_OFFLOAD
188 static int alloc_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *, int, int,
189     struct sysctl_oid *);
190 static int free_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *);
191 #endif
192 #ifdef DEV_NETMAP
193 static int alloc_nm_rxq(struct vi_info *, struct sge_nm_rxq *, int, int,
194     struct sysctl_oid *);
195 static int free_nm_rxq(struct vi_info *, struct sge_nm_rxq *);
196 static int alloc_nm_txq(struct vi_info *, struct sge_nm_txq *, int, int,
197     struct sysctl_oid *);
198 static int free_nm_txq(struct vi_info *, struct sge_nm_txq *);
199 #endif
200 static int ctrl_eq_alloc(struct adapter *, struct sge_eq *);
201 static int eth_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
202 #ifdef TCP_OFFLOAD
203 static int ofld_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
204 #endif
205 static int alloc_eq(struct adapter *, struct vi_info *, struct sge_eq *);
206 static int free_eq(struct adapter *, struct sge_eq *);
207 static int alloc_wrq(struct adapter *, struct vi_info *, struct sge_wrq *,
208     struct sysctl_oid *);
209 static int free_wrq(struct adapter *, struct sge_wrq *);
210 static int alloc_txq(struct vi_info *, struct sge_txq *, int,
211     struct sysctl_oid *);
212 static int free_txq(struct vi_info *, struct sge_txq *);
213 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int);
214 static inline void ring_fl_db(struct adapter *, struct sge_fl *);
215 static int refill_fl(struct adapter *, struct sge_fl *, int);
216 static void refill_sfl(void *);
217 static int alloc_fl_sdesc(struct sge_fl *);
218 static void free_fl_sdesc(struct adapter *, struct sge_fl *);
219 static void find_best_refill_source(struct adapter *, struct sge_fl *, int);
220 static void find_safe_refill_source(struct adapter *, struct sge_fl *);
221 static void add_fl_to_sfl(struct adapter *, struct sge_fl *);
222 
223 static inline void get_pkt_gl(struct mbuf *, struct sglist *);
224 static inline u_int txpkt_len16(u_int, u_int);
225 static inline u_int txpkts0_len16(u_int);
226 static inline u_int txpkts1_len16(void);
227 static u_int write_txpkt_wr(struct sge_txq *, struct fw_eth_tx_pkt_wr *,
228     struct mbuf *, u_int);
229 static int try_txpkts(struct mbuf *, struct mbuf *, struct txpkts *, u_int);
230 static int add_to_txpkts(struct mbuf *, struct txpkts *, u_int);
231 static u_int write_txpkts_wr(struct sge_txq *, struct fw_eth_tx_pkts_wr *,
232     struct mbuf *, const struct txpkts *, u_int);
233 static void write_gl_to_txd(struct sge_txq *, struct mbuf *, caddr_t *, int);
234 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int);
235 static inline void ring_eq_db(struct adapter *, struct sge_eq *, u_int);
236 static inline uint16_t read_hw_cidx(struct sge_eq *);
237 static inline u_int reclaimable_tx_desc(struct sge_eq *);
238 static inline u_int total_available_tx_desc(struct sge_eq *);
239 static u_int reclaim_tx_descs(struct sge_txq *, u_int);
240 static void tx_reclaim(void *, int);
241 static __be64 get_flit(struct sglist_seg *, int, int);
242 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
243     struct mbuf *);
244 static int handle_fw_msg(struct sge_iq *, const struct rss_header *,
245     struct mbuf *);
246 static void wrq_tx_drain(void *, int);
247 static void drain_wrq_wr_list(struct adapter *, struct sge_wrq *);
248 
249 static int sysctl_uint16(SYSCTL_HANDLER_ARGS);
250 static int sysctl_bufsizes(SYSCTL_HANDLER_ARGS);
251 
252 static counter_u64_t extfree_refs;
253 static counter_u64_t extfree_rels;
254 
255 /*
256  * Called on MOD_LOAD.  Validates and calculates the SGE tunables.
257  */
258 void
259 t4_sge_modload(void)
260 {
261 
262 	if (fl_pktshift < 0 || fl_pktshift > 7) {
263 		printf("Invalid hw.cxgbe.fl_pktshift value (%d),"
264 		    " using 2 instead.\n", fl_pktshift);
265 		fl_pktshift = 2;
266 	}
267 
268 	if (spg_len != 64 && spg_len != 128) {
269 		int len;
270 
271 #if defined(__i386__) || defined(__amd64__)
272 		len = cpu_clflush_line_size > 64 ? 128 : 64;
273 #else
274 		len = 64;
275 #endif
276 		if (spg_len != -1) {
277 			printf("Invalid hw.cxgbe.spg_len value (%d),"
278 			    " using %d instead.\n", spg_len, len);
279 		}
280 		spg_len = len;
281 	}
282 
283 	if (cong_drop < -1 || cong_drop > 1) {
284 		printf("Invalid hw.cxgbe.cong_drop value (%d),"
285 		    " using 0 instead.\n", cong_drop);
286 		cong_drop = 0;
287 	}
288 
289 	extfree_refs = counter_u64_alloc(M_WAITOK);
290 	extfree_rels = counter_u64_alloc(M_WAITOK);
291 	counter_u64_zero(extfree_refs);
292 	counter_u64_zero(extfree_rels);
293 }
294 
295 void
296 t4_sge_modunload(void)
297 {
298 
299 	counter_u64_free(extfree_refs);
300 	counter_u64_free(extfree_rels);
301 }
302 
303 uint64_t
304 t4_sge_extfree_refs(void)
305 {
306 	uint64_t refs, rels;
307 
308 	rels = counter_u64_fetch(extfree_rels);
309 	refs = counter_u64_fetch(extfree_refs);
310 
311 	return (refs - rels);
312 }
313 
314 void
315 t4_init_sge_cpl_handlers(struct adapter *sc)
316 {
317 
318 	t4_register_cpl_handler(sc, CPL_FW4_MSG, handle_fw_msg);
319 	t4_register_cpl_handler(sc, CPL_FW6_MSG, handle_fw_msg);
320 	t4_register_cpl_handler(sc, CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
321 	t4_register_cpl_handler(sc, CPL_RX_PKT, t4_eth_rx);
322 	t4_register_fw_msg_handler(sc, FW6_TYPE_CMD_RPL, t4_handle_fw_rpl);
323 }
324 
325 static inline void
326 setup_pad_and_pack_boundaries(struct adapter *sc)
327 {
328 	uint32_t v, m;
329 	int pad, pack;
330 
331 	pad = fl_pad;
332 	if (fl_pad < 32 || fl_pad > 4096 || !powerof2(fl_pad)) {
333 		/*
334 		 * If there is any chance that we might use buffer packing and
335 		 * the chip is a T4, then pick 64 as the pad/pack boundary.  Set
336 		 * it to 32 in all other cases.
337 		 */
338 		pad = is_t4(sc) && buffer_packing ? 64 : 32;
339 
340 		/*
341 		 * For fl_pad = 0 we'll still write a reasonable value to the
342 		 * register but all the freelists will opt out of padding.
343 		 * We'll complain here only if the user tried to set it to a
344 		 * value greater than 0 that was invalid.
345 		 */
346 		if (fl_pad > 0) {
347 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pad value"
348 			    " (%d), using %d instead.\n", fl_pad, pad);
349 		}
350 	}
351 	m = V_INGPADBOUNDARY(M_INGPADBOUNDARY);
352 	v = V_INGPADBOUNDARY(ilog2(pad) - 5);
353 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
354 
355 	if (is_t4(sc)) {
356 		if (fl_pack != -1 && fl_pack != pad) {
357 			/* Complain but carry on. */
358 			device_printf(sc->dev, "hw.cxgbe.fl_pack (%d) ignored,"
359 			    " using %d instead.\n", fl_pack, pad);
360 		}
361 		return;
362 	}
363 
364 	pack = fl_pack;
365 	if (fl_pack < 16 || fl_pack == 32 || fl_pack > 4096 ||
366 	    !powerof2(fl_pack)) {
367 		pack = max(sc->params.pci.mps, CACHE_LINE_SIZE);
368 		MPASS(powerof2(pack));
369 		if (pack < 16)
370 			pack = 16;
371 		if (pack == 32)
372 			pack = 64;
373 		if (pack > 4096)
374 			pack = 4096;
375 		if (fl_pack != -1) {
376 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pack value"
377 			    " (%d), using %d instead.\n", fl_pack, pack);
378 		}
379 	}
380 	m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY);
381 	if (pack == 16)
382 		v = V_INGPACKBOUNDARY(0);
383 	else
384 		v = V_INGPACKBOUNDARY(ilog2(pack) - 5);
385 
386 	MPASS(!is_t4(sc));	/* T4 doesn't have SGE_CONTROL2 */
387 	t4_set_reg_field(sc, A_SGE_CONTROL2, m, v);
388 }
389 
390 /*
391  * adap->params.vpd.cclk must be set up before this is called.
392  */
393 void
394 t4_tweak_chip_settings(struct adapter *sc)
395 {
396 	int i;
397 	uint32_t v, m;
398 	int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200};
399 	int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk;
400 	int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */
401 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
402 	static int sge_flbuf_sizes[] = {
403 		MCLBYTES,
404 #if MJUMPAGESIZE != MCLBYTES
405 		MJUMPAGESIZE,
406 		MJUMPAGESIZE - CL_METADATA_SIZE,
407 		MJUMPAGESIZE - 2 * MSIZE - CL_METADATA_SIZE,
408 #endif
409 		MJUM9BYTES,
410 		MJUM16BYTES,
411 		MCLBYTES - MSIZE - CL_METADATA_SIZE,
412 		MJUM9BYTES - CL_METADATA_SIZE,
413 		MJUM16BYTES - CL_METADATA_SIZE,
414 	};
415 
416 	KASSERT(sc->flags & MASTER_PF,
417 	    ("%s: trying to change chip settings when not master.", __func__));
418 
419 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
420 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
421 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
422 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
423 
424 	setup_pad_and_pack_boundaries(sc);
425 
426 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
427 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
428 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
429 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
430 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
431 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
432 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
433 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
434 	t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v);
435 
436 	KASSERT(nitems(sge_flbuf_sizes) <= SGE_FLBUF_SIZES,
437 	    ("%s: hw buffer size table too big", __func__));
438 	for (i = 0; i < min(nitems(sge_flbuf_sizes), SGE_FLBUF_SIZES); i++) {
439 		t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i),
440 		    sge_flbuf_sizes[i]);
441 	}
442 
443 	v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) |
444 	    V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]);
445 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v);
446 
447 	KASSERT(intr_timer[0] <= timer_max,
448 	    ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0],
449 	    timer_max));
450 	for (i = 1; i < nitems(intr_timer); i++) {
451 		KASSERT(intr_timer[i] >= intr_timer[i - 1],
452 		    ("%s: timers not listed in increasing order (%d)",
453 		    __func__, i));
454 
455 		while (intr_timer[i] > timer_max) {
456 			if (i == nitems(intr_timer) - 1) {
457 				intr_timer[i] = timer_max;
458 				break;
459 			}
460 			intr_timer[i] += intr_timer[i - 1];
461 			intr_timer[i] /= 2;
462 		}
463 	}
464 
465 	v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) |
466 	    V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1]));
467 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v);
468 	v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) |
469 	    V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3]));
470 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v);
471 	v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) |
472 	    V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5]));
473 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v);
474 
475 	/* 4K, 16K, 64K, 256K DDP "page sizes" */
476 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
477 	t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v);
478 
479 	m = v = F_TDDPTAGTCB;
480 	t4_set_reg_field(sc, A_ULP_RX_CTL, m, v);
481 
482 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
483 	    F_RESETDDPOFFSET;
484 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
485 	t4_set_reg_field(sc, A_TP_PARA_REG5, m, v);
486 }
487 
488 /*
489  * SGE wants the buffer to be at least 64B and then a multiple of 16.  If
490  * padding is is use the buffer's start and end need to be aligned to the pad
491  * boundary as well.  We'll just make sure that the size is a multiple of the
492  * boundary here, it is up to the buffer allocation code to make sure the start
493  * of the buffer is aligned as well.
494  */
495 static inline int
496 hwsz_ok(struct adapter *sc, int hwsz)
497 {
498 	int mask = fl_pad ? sc->sge.pad_boundary - 1 : 16 - 1;
499 
500 	return (hwsz >= 64 && (hwsz & mask) == 0);
501 }
502 
503 /*
504  * XXX: driver really should be able to deal with unexpected settings.
505  */
506 int
507 t4_read_chip_settings(struct adapter *sc)
508 {
509 	struct sge *s = &sc->sge;
510 	int i, j, n, rc = 0;
511 	uint32_t m, v, r;
512 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
513 	static int sw_buf_sizes[] = {	/* Sorted by size */
514 		MCLBYTES,
515 #if MJUMPAGESIZE != MCLBYTES
516 		MJUMPAGESIZE,
517 #endif
518 		MJUM9BYTES,
519 		MJUM16BYTES
520 	};
521 	struct sw_zone_info *swz, *safe_swz;
522 	struct hw_buf_info *hwb;
523 
524 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
525 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
526 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
527 	r = t4_read_reg(sc, A_SGE_CONTROL);
528 	if ((r & m) != v) {
529 		device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r);
530 		rc = EINVAL;
531 	}
532 	s->pad_boundary = 1 << (G_INGPADBOUNDARY(r) + 5);
533 
534 	if (is_t4(sc))
535 		s->pack_boundary = s->pad_boundary;
536 	else {
537 		r = t4_read_reg(sc, A_SGE_CONTROL2);
538 		if (G_INGPACKBOUNDARY(r) == 0)
539 			s->pack_boundary = 16;
540 		else
541 			s->pack_boundary = 1 << (G_INGPACKBOUNDARY(r) + 5);
542 	}
543 
544 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
545 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
546 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
547 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
548 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
549 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
550 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
551 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
552 	r = t4_read_reg(sc, A_SGE_HOST_PAGE_SIZE);
553 	if (r != v) {
554 		device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r);
555 		rc = EINVAL;
556 	}
557 
558 	/* Filter out unusable hw buffer sizes entirely (mark with -2). */
559 	hwb = &s->hw_buf_info[0];
560 	for (i = 0; i < nitems(s->hw_buf_info); i++, hwb++) {
561 		r = t4_read_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i));
562 		hwb->size = r;
563 		hwb->zidx = hwsz_ok(sc, r) ? -1 : -2;
564 		hwb->next = -1;
565 	}
566 
567 	/*
568 	 * Create a sorted list in decreasing order of hw buffer sizes (and so
569 	 * increasing order of spare area) for each software zone.
570 	 *
571 	 * If padding is enabled then the start and end of the buffer must align
572 	 * to the pad boundary; if packing is enabled then they must align with
573 	 * the pack boundary as well.  Allocations from the cluster zones are
574 	 * aligned to min(size, 4K), so the buffer starts at that alignment and
575 	 * ends at hwb->size alignment.  If mbuf inlining is allowed the
576 	 * starting alignment will be reduced to MSIZE and the driver will
577 	 * exercise appropriate caution when deciding on the best buffer layout
578 	 * to use.
579 	 */
580 	n = 0;	/* no usable buffer size to begin with */
581 	swz = &s->sw_zone_info[0];
582 	safe_swz = NULL;
583 	for (i = 0; i < SW_ZONE_SIZES; i++, swz++) {
584 		int8_t head = -1, tail = -1;
585 
586 		swz->size = sw_buf_sizes[i];
587 		swz->zone = m_getzone(swz->size);
588 		swz->type = m_gettype(swz->size);
589 
590 		if (swz->size < PAGE_SIZE) {
591 			MPASS(powerof2(swz->size));
592 			if (fl_pad && (swz->size % sc->sge.pad_boundary != 0))
593 				continue;
594 		}
595 
596 		if (swz->size == safest_rx_cluster)
597 			safe_swz = swz;
598 
599 		hwb = &s->hw_buf_info[0];
600 		for (j = 0; j < SGE_FLBUF_SIZES; j++, hwb++) {
601 			if (hwb->zidx != -1 || hwb->size > swz->size)
602 				continue;
603 #ifdef INVARIANTS
604 			if (fl_pad)
605 				MPASS(hwb->size % sc->sge.pad_boundary == 0);
606 #endif
607 			hwb->zidx = i;
608 			if (head == -1)
609 				head = tail = j;
610 			else if (hwb->size < s->hw_buf_info[tail].size) {
611 				s->hw_buf_info[tail].next = j;
612 				tail = j;
613 			} else {
614 				int8_t *cur;
615 				struct hw_buf_info *t;
616 
617 				for (cur = &head; *cur != -1; cur = &t->next) {
618 					t = &s->hw_buf_info[*cur];
619 					if (hwb->size == t->size) {
620 						hwb->zidx = -2;
621 						break;
622 					}
623 					if (hwb->size > t->size) {
624 						hwb->next = *cur;
625 						*cur = j;
626 						break;
627 					}
628 				}
629 			}
630 		}
631 		swz->head_hwidx = head;
632 		swz->tail_hwidx = tail;
633 
634 		if (tail != -1) {
635 			n++;
636 			if (swz->size - s->hw_buf_info[tail].size >=
637 			    CL_METADATA_SIZE)
638 				sc->flags |= BUF_PACKING_OK;
639 		}
640 	}
641 	if (n == 0) {
642 		device_printf(sc->dev, "no usable SGE FL buffer size.\n");
643 		rc = EINVAL;
644 	}
645 
646 	s->safe_hwidx1 = -1;
647 	s->safe_hwidx2 = -1;
648 	if (safe_swz != NULL) {
649 		s->safe_hwidx1 = safe_swz->head_hwidx;
650 		for (i = safe_swz->head_hwidx; i != -1; i = hwb->next) {
651 			int spare;
652 
653 			hwb = &s->hw_buf_info[i];
654 #ifdef INVARIANTS
655 			if (fl_pad)
656 				MPASS(hwb->size % sc->sge.pad_boundary == 0);
657 #endif
658 			spare = safe_swz->size - hwb->size;
659 			if (spare >= CL_METADATA_SIZE) {
660 				s->safe_hwidx2 = i;
661 				break;
662 			}
663 		}
664 	}
665 
666 	r = t4_read_reg(sc, A_SGE_INGRESS_RX_THRESHOLD);
667 	s->counter_val[0] = G_THRESHOLD_0(r);
668 	s->counter_val[1] = G_THRESHOLD_1(r);
669 	s->counter_val[2] = G_THRESHOLD_2(r);
670 	s->counter_val[3] = G_THRESHOLD_3(r);
671 
672 	r = t4_read_reg(sc, A_SGE_TIMER_VALUE_0_AND_1);
673 	s->timer_val[0] = G_TIMERVALUE0(r) / core_ticks_per_usec(sc);
674 	s->timer_val[1] = G_TIMERVALUE1(r) / core_ticks_per_usec(sc);
675 	r = t4_read_reg(sc, A_SGE_TIMER_VALUE_2_AND_3);
676 	s->timer_val[2] = G_TIMERVALUE2(r) / core_ticks_per_usec(sc);
677 	s->timer_val[3] = G_TIMERVALUE3(r) / core_ticks_per_usec(sc);
678 	r = t4_read_reg(sc, A_SGE_TIMER_VALUE_4_AND_5);
679 	s->timer_val[4] = G_TIMERVALUE4(r) / core_ticks_per_usec(sc);
680 	s->timer_val[5] = G_TIMERVALUE5(r) / core_ticks_per_usec(sc);
681 
682 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
683 	r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ);
684 	if (r != v) {
685 		device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r);
686 		rc = EINVAL;
687 	}
688 
689 	m = v = F_TDDPTAGTCB;
690 	r = t4_read_reg(sc, A_ULP_RX_CTL);
691 	if ((r & m) != v) {
692 		device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r);
693 		rc = EINVAL;
694 	}
695 
696 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
697 	    F_RESETDDPOFFSET;
698 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
699 	r = t4_read_reg(sc, A_TP_PARA_REG5);
700 	if ((r & m) != v) {
701 		device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r);
702 		rc = EINVAL;
703 	}
704 
705 	r = t4_read_reg(sc, A_SGE_CONM_CTRL);
706 	s->fl_starve_threshold = G_EGRTHRESHOLD(r) * 2 + 1;
707 	if (is_t4(sc))
708 		s->fl_starve_threshold2 = s->fl_starve_threshold;
709 	else
710 		s->fl_starve_threshold2 = G_EGRTHRESHOLDPACKING(r) * 2 + 1;
711 
712 	/* egress queues: log2 of # of doorbells per BAR2 page */
713 	r = t4_read_reg(sc, A_SGE_EGRESS_QUEUES_PER_PAGE_PF);
714 	r >>= S_QUEUESPERPAGEPF0 +
715 	    (S_QUEUESPERPAGEPF1 - S_QUEUESPERPAGEPF0) * sc->pf;
716 	s->eq_s_qpp = r & M_QUEUESPERPAGEPF0;
717 
718 	/* ingress queues: log2 of # of doorbells per BAR2 page */
719 	r = t4_read_reg(sc, A_SGE_INGRESS_QUEUES_PER_PAGE_PF);
720 	r >>= S_QUEUESPERPAGEPF0 +
721 	    (S_QUEUESPERPAGEPF1 - S_QUEUESPERPAGEPF0) * sc->pf;
722 	s->iq_s_qpp = r & M_QUEUESPERPAGEPF0;
723 
724 	t4_init_tp_params(sc);
725 
726 	t4_read_mtu_tbl(sc, sc->params.mtus, NULL);
727 	t4_load_mtus(sc, sc->params.mtus, sc->params.a_wnd, sc->params.b_wnd);
728 
729 	return (rc);
730 }
731 
732 int
733 t4_create_dma_tag(struct adapter *sc)
734 {
735 	int rc;
736 
737 	rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
738 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
739 	    BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
740 	    NULL, &sc->dmat);
741 	if (rc != 0) {
742 		device_printf(sc->dev,
743 		    "failed to create main DMA tag: %d\n", rc);
744 	}
745 
746 	return (rc);
747 }
748 
749 void
750 t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
751     struct sysctl_oid_list *children)
752 {
753 
754 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "buffer_sizes",
755 	    CTLTYPE_STRING | CTLFLAG_RD, &sc->sge, 0, sysctl_bufsizes, "A",
756 	    "freelist buffer sizes");
757 
758 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD,
759 	    NULL, fl_pktshift, "payload DMA offset in rx buffer (bytes)");
760 
761 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD,
762 	    NULL, sc->sge.pad_boundary, "payload pad boundary (bytes)");
763 
764 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD,
765 	    NULL, spg_len, "status page size (bytes)");
766 
767 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD,
768 	    NULL, cong_drop, "congestion drop setting");
769 
770 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD,
771 	    NULL, sc->sge.pack_boundary, "payload pack boundary (bytes)");
772 }
773 
774 int
775 t4_destroy_dma_tag(struct adapter *sc)
776 {
777 	if (sc->dmat)
778 		bus_dma_tag_destroy(sc->dmat);
779 
780 	return (0);
781 }
782 
783 /*
784  * Allocate and initialize the firmware event queue and the management queue.
785  *
786  * Returns errno on failure.  Resources allocated up to that point may still be
787  * allocated.  Caller is responsible for cleanup in case this function fails.
788  */
789 int
790 t4_setup_adapter_queues(struct adapter *sc)
791 {
792 	int rc;
793 
794 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
795 
796 	sysctl_ctx_init(&sc->ctx);
797 	sc->flags |= ADAP_SYSCTL_CTX;
798 
799 	/*
800 	 * Firmware event queue
801 	 */
802 	rc = alloc_fwq(sc);
803 	if (rc != 0)
804 		return (rc);
805 
806 	/*
807 	 * Management queue.  This is just a control queue that uses the fwq as
808 	 * its associated iq.
809 	 */
810 	rc = alloc_mgmtq(sc);
811 
812 	return (rc);
813 }
814 
815 /*
816  * Idempotent
817  */
818 int
819 t4_teardown_adapter_queues(struct adapter *sc)
820 {
821 
822 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
823 
824 	/* Do this before freeing the queue */
825 	if (sc->flags & ADAP_SYSCTL_CTX) {
826 		sysctl_ctx_free(&sc->ctx);
827 		sc->flags &= ~ADAP_SYSCTL_CTX;
828 	}
829 
830 	free_mgmtq(sc);
831 	free_fwq(sc);
832 
833 	return (0);
834 }
835 
836 static inline int
837 first_vector(struct vi_info *vi)
838 {
839 	struct adapter *sc = vi->pi->adapter;
840 
841 	if (sc->intr_count == 1)
842 		return (0);
843 
844 	return (vi->first_intr);
845 }
846 
847 /*
848  * Given an arbitrary "index," come up with an iq that can be used by other
849  * queues (of this VI) for interrupt forwarding, SGE egress updates, etc.
850  * The iq returned is guaranteed to be something that takes direct interrupts.
851  */
852 static struct sge_iq *
853 vi_intr_iq(struct vi_info *vi, int idx)
854 {
855 	struct adapter *sc = vi->pi->adapter;
856 	struct sge *s = &sc->sge;
857 	struct sge_iq *iq = NULL;
858 	int nintr, i;
859 
860 	if (sc->intr_count == 1)
861 		return (&sc->sge.fwq);
862 
863 	KASSERT(!(vi->flags & VI_NETMAP),
864 	    ("%s: called on netmap VI", __func__));
865 	nintr = vi->nintr;
866 	KASSERT(nintr != 0,
867 	    ("%s: vi %p has no exclusive interrupts, total interrupts = %d",
868 	    __func__, vi, sc->intr_count));
869 	i = idx % nintr;
870 
871 	if (vi->flags & INTR_RXQ) {
872 	       	if (i < vi->nrxq) {
873 			iq = &s->rxq[vi->first_rxq + i].iq;
874 			goto done;
875 		}
876 		i -= vi->nrxq;
877 	}
878 #ifdef TCP_OFFLOAD
879 	if (vi->flags & INTR_OFLD_RXQ) {
880 	       	if (i < vi->nofldrxq) {
881 			iq = &s->ofld_rxq[vi->first_ofld_rxq + i].iq;
882 			goto done;
883 		}
884 		i -= vi->nofldrxq;
885 	}
886 #endif
887 	panic("%s: vi %p, intr_flags 0x%lx, idx %d, total intr %d\n", __func__,
888 	    vi, vi->flags & INTR_ALL, idx, nintr);
889 done:
890 	MPASS(iq != NULL);
891 	KASSERT(iq->flags & IQ_INTR,
892 	    ("%s: iq %p (vi %p, intr_flags 0x%lx, idx %d)", __func__, iq, vi,
893 	    vi->flags & INTR_ALL, idx));
894 	return (iq);
895 }
896 
897 /* Maximum payload that can be delivered with a single iq descriptor */
898 static inline int
899 mtu_to_max_payload(struct adapter *sc, int mtu, const int toe)
900 {
901 	int payload;
902 
903 #ifdef TCP_OFFLOAD
904 	if (toe) {
905 		payload = sc->tt.rx_coalesce ?
906 		    G_RXCOALESCESIZE(t4_read_reg(sc, A_TP_PARA_REG2)) : mtu;
907 	} else {
908 #endif
909 		/* large enough even when hw VLAN extraction is disabled */
910 		payload = fl_pktshift + ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN +
911 		    mtu;
912 #ifdef TCP_OFFLOAD
913 	}
914 #endif
915 
916 	return (payload);
917 }
918 
919 int
920 t4_setup_vi_queues(struct vi_info *vi)
921 {
922 	int rc = 0, i, j, intr_idx, iqid;
923 	struct sge_rxq *rxq;
924 	struct sge_txq *txq;
925 	struct sge_wrq *ctrlq;
926 #ifdef TCP_OFFLOAD
927 	struct sge_ofld_rxq *ofld_rxq;
928 	struct sge_wrq *ofld_txq;
929 #endif
930 #ifdef DEV_NETMAP
931 	struct sge_nm_rxq *nm_rxq;
932 	struct sge_nm_txq *nm_txq;
933 #endif
934 	char name[16];
935 	struct port_info *pi = vi->pi;
936 	struct adapter *sc = pi->adapter;
937 	struct ifnet *ifp = vi->ifp;
938 	struct sysctl_oid *oid = device_get_sysctl_tree(vi->dev);
939 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
940 	int maxp, mtu = ifp->if_mtu;
941 
942 	/* Interrupt vector to start from (when using multiple vectors) */
943 	intr_idx = first_vector(vi);
944 
945 #ifdef DEV_NETMAP
946 	if (vi->flags & VI_NETMAP) {
947 		/*
948 		 * We don't have buffers to back the netmap rx queues
949 		 * right now so we create the queues in a way that
950 		 * doesn't set off any congestion signal in the chip.
951 		 */
952 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
953 		    CTLFLAG_RD, NULL, "rx queues");
954 		for_each_nm_rxq(vi, i, nm_rxq) {
955 			rc = alloc_nm_rxq(vi, nm_rxq, intr_idx, i, oid);
956 			if (rc != 0)
957 				goto done;
958 			intr_idx++;
959 		}
960 
961 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "txq",
962 		    CTLFLAG_RD, NULL, "tx queues");
963 		for_each_nm_txq(vi, i, nm_txq) {
964 			iqid = vi->first_rxq + (i % vi->nrxq);
965 			rc = alloc_nm_txq(vi, nm_txq, iqid, i, oid);
966 			if (rc != 0)
967 				goto done;
968 		}
969 		goto done;
970 	}
971 #endif
972 
973 	/*
974 	 * First pass over all NIC and TOE rx queues:
975 	 * a) initialize iq and fl
976 	 * b) allocate queue iff it will take direct interrupts.
977 	 */
978 	maxp = mtu_to_max_payload(sc, mtu, 0);
979 	if (vi->flags & INTR_RXQ) {
980 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
981 		    CTLFLAG_RD, NULL, "rx queues");
982 	}
983 	for_each_rxq(vi, i, rxq) {
984 
985 		init_iq(&rxq->iq, sc, vi->tmr_idx, vi->pktc_idx, vi->qsize_rxq);
986 
987 		snprintf(name, sizeof(name), "%s rxq%d-fl",
988 		    device_get_nameunit(vi->dev), i);
989 		init_fl(sc, &rxq->fl, vi->qsize_rxq / 8, maxp, name);
990 
991 		if (vi->flags & INTR_RXQ) {
992 			rxq->iq.flags |= IQ_INTR;
993 			rc = alloc_rxq(vi, rxq, intr_idx, i, oid);
994 			if (rc != 0)
995 				goto done;
996 			intr_idx++;
997 		}
998 	}
999 #ifdef TCP_OFFLOAD
1000 	maxp = mtu_to_max_payload(sc, mtu, 1);
1001 	if (vi->flags & INTR_OFLD_RXQ) {
1002 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq",
1003 		    CTLFLAG_RD, NULL,
1004 		    "rx queues for offloaded TCP connections");
1005 	}
1006 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1007 
1008 		init_iq(&ofld_rxq->iq, sc, vi->tmr_idx, vi->pktc_idx,
1009 		    vi->qsize_rxq);
1010 
1011 		snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
1012 		    device_get_nameunit(vi->dev), i);
1013 		init_fl(sc, &ofld_rxq->fl, vi->qsize_rxq / 8, maxp, name);
1014 
1015 		if (vi->flags & INTR_OFLD_RXQ) {
1016 			ofld_rxq->iq.flags |= IQ_INTR;
1017 			rc = alloc_ofld_rxq(vi, ofld_rxq, intr_idx, i, oid);
1018 			if (rc != 0)
1019 				goto done;
1020 			intr_idx++;
1021 		}
1022 	}
1023 #endif
1024 
1025 	/*
1026 	 * Second pass over all NIC and TOE rx queues.  The queues forwarding
1027 	 * their interrupts are allocated now.
1028 	 */
1029 	j = 0;
1030 	if (!(vi->flags & INTR_RXQ)) {
1031 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
1032 		    CTLFLAG_RD, NULL, "rx queues");
1033 		for_each_rxq(vi, i, rxq) {
1034 			MPASS(!(rxq->iq.flags & IQ_INTR));
1035 
1036 			intr_idx = vi_intr_iq(vi, j)->abs_id;
1037 
1038 			rc = alloc_rxq(vi, rxq, intr_idx, i, oid);
1039 			if (rc != 0)
1040 				goto done;
1041 			j++;
1042 		}
1043 	}
1044 #ifdef TCP_OFFLOAD
1045 	if (vi->nofldrxq != 0 && !(vi->flags & INTR_OFLD_RXQ)) {
1046 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq",
1047 		    CTLFLAG_RD, NULL,
1048 		    "rx queues for offloaded TCP connections");
1049 		for_each_ofld_rxq(vi, i, ofld_rxq) {
1050 			MPASS(!(ofld_rxq->iq.flags & IQ_INTR));
1051 
1052 			intr_idx = vi_intr_iq(vi, j)->abs_id;
1053 
1054 			rc = alloc_ofld_rxq(vi, ofld_rxq, intr_idx, i, oid);
1055 			if (rc != 0)
1056 				goto done;
1057 			j++;
1058 		}
1059 	}
1060 #endif
1061 
1062 	/*
1063 	 * Now the tx queues.  Only one pass needed.
1064 	 */
1065 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "txq", CTLFLAG_RD,
1066 	    NULL, "tx queues");
1067 	j = 0;
1068 	for_each_txq(vi, i, txq) {
1069 		iqid = vi_intr_iq(vi, j)->cntxt_id;
1070 		snprintf(name, sizeof(name), "%s txq%d",
1071 		    device_get_nameunit(vi->dev), i);
1072 		init_eq(&txq->eq, EQ_ETH, vi->qsize_txq, pi->tx_chan, iqid,
1073 		    name);
1074 
1075 		rc = alloc_txq(vi, txq, i, oid);
1076 		if (rc != 0)
1077 			goto done;
1078 		j++;
1079 	}
1080 #ifdef TCP_OFFLOAD
1081 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_txq",
1082 	    CTLFLAG_RD, NULL, "tx queues for offloaded TCP connections");
1083 	for_each_ofld_txq(vi, i, ofld_txq) {
1084 		struct sysctl_oid *oid2;
1085 
1086 		iqid = vi_intr_iq(vi, j)->cntxt_id;
1087 		snprintf(name, sizeof(name), "%s ofld_txq%d",
1088 		    device_get_nameunit(vi->dev), i);
1089 		init_eq(&ofld_txq->eq, EQ_OFLD, vi->qsize_txq, pi->tx_chan,
1090 		    iqid, name);
1091 
1092 		snprintf(name, sizeof(name), "%d", i);
1093 		oid2 = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
1094 		    name, CTLFLAG_RD, NULL, "offload tx queue");
1095 
1096 		rc = alloc_wrq(sc, vi, ofld_txq, oid2);
1097 		if (rc != 0)
1098 			goto done;
1099 		j++;
1100 	}
1101 #endif
1102 
1103 	/*
1104 	 * Finally, the control queue.
1105 	 */
1106 	if (!IS_MAIN_VI(vi))
1107 		goto done;
1108 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ctrlq", CTLFLAG_RD,
1109 	    NULL, "ctrl queue");
1110 	ctrlq = &sc->sge.ctrlq[pi->port_id];
1111 	iqid = vi_intr_iq(vi, 0)->cntxt_id;
1112 	snprintf(name, sizeof(name), "%s ctrlq", device_get_nameunit(vi->dev));
1113 	init_eq(&ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, pi->tx_chan, iqid, name);
1114 	rc = alloc_wrq(sc, vi, ctrlq, oid);
1115 
1116 done:
1117 	if (rc)
1118 		t4_teardown_vi_queues(vi);
1119 
1120 	return (rc);
1121 }
1122 
1123 /*
1124  * Idempotent
1125  */
1126 int
1127 t4_teardown_vi_queues(struct vi_info *vi)
1128 {
1129 	int i;
1130 	struct port_info *pi = vi->pi;
1131 	struct adapter *sc = pi->adapter;
1132 	struct sge_rxq *rxq;
1133 	struct sge_txq *txq;
1134 #ifdef TCP_OFFLOAD
1135 	struct sge_ofld_rxq *ofld_rxq;
1136 	struct sge_wrq *ofld_txq;
1137 #endif
1138 #ifdef DEV_NETMAP
1139 	struct sge_nm_rxq *nm_rxq;
1140 	struct sge_nm_txq *nm_txq;
1141 #endif
1142 
1143 	/* Do this before freeing the queues */
1144 	if (vi->flags & VI_SYSCTL_CTX) {
1145 		sysctl_ctx_free(&vi->ctx);
1146 		vi->flags &= ~VI_SYSCTL_CTX;
1147 	}
1148 
1149 #ifdef DEV_NETMAP
1150 	if (vi->flags & VI_NETMAP) {
1151 		for_each_nm_txq(vi, i, nm_txq) {
1152 			free_nm_txq(vi, nm_txq);
1153 		}
1154 
1155 		for_each_nm_rxq(vi, i, nm_rxq) {
1156 			free_nm_rxq(vi, nm_rxq);
1157 		}
1158 		return (0);
1159 	}
1160 #endif
1161 
1162 	/*
1163 	 * Take down all the tx queues first, as they reference the rx queues
1164 	 * (for egress updates, etc.).
1165 	 */
1166 
1167 	if (IS_MAIN_VI(vi))
1168 		free_wrq(sc, &sc->sge.ctrlq[pi->port_id]);
1169 
1170 	for_each_txq(vi, i, txq) {
1171 		free_txq(vi, txq);
1172 	}
1173 #ifdef TCP_OFFLOAD
1174 	for_each_ofld_txq(vi, i, ofld_txq) {
1175 		free_wrq(sc, ofld_txq);
1176 	}
1177 #endif
1178 
1179 	/*
1180 	 * Then take down the rx queues that forward their interrupts, as they
1181 	 * reference other rx queues.
1182 	 */
1183 
1184 	for_each_rxq(vi, i, rxq) {
1185 		if ((rxq->iq.flags & IQ_INTR) == 0)
1186 			free_rxq(vi, rxq);
1187 	}
1188 #ifdef TCP_OFFLOAD
1189 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1190 		if ((ofld_rxq->iq.flags & IQ_INTR) == 0)
1191 			free_ofld_rxq(vi, ofld_rxq);
1192 	}
1193 #endif
1194 
1195 	/*
1196 	 * Then take down the rx queues that take direct interrupts.
1197 	 */
1198 
1199 	for_each_rxq(vi, i, rxq) {
1200 		if (rxq->iq.flags & IQ_INTR)
1201 			free_rxq(vi, rxq);
1202 	}
1203 #ifdef TCP_OFFLOAD
1204 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1205 		if (ofld_rxq->iq.flags & IQ_INTR)
1206 			free_ofld_rxq(vi, ofld_rxq);
1207 	}
1208 #endif
1209 
1210 	return (0);
1211 }
1212 
1213 /*
1214  * Deals with errors and the firmware event queue.  All data rx queues forward
1215  * their interrupt to the firmware event queue.
1216  */
1217 void
1218 t4_intr_all(void *arg)
1219 {
1220 	struct adapter *sc = arg;
1221 	struct sge_iq *fwq = &sc->sge.fwq;
1222 
1223 	t4_intr_err(arg);
1224 	if (atomic_cmpset_int(&fwq->state, IQS_IDLE, IQS_BUSY)) {
1225 		service_iq(fwq, 0);
1226 		atomic_cmpset_int(&fwq->state, IQS_BUSY, IQS_IDLE);
1227 	}
1228 }
1229 
1230 /* Deals with error interrupts */
1231 void
1232 t4_intr_err(void *arg)
1233 {
1234 	struct adapter *sc = arg;
1235 
1236 	t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
1237 	t4_slow_intr_handler(sc);
1238 }
1239 
1240 void
1241 t4_intr_evt(void *arg)
1242 {
1243 	struct sge_iq *iq = arg;
1244 
1245 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1246 		service_iq(iq, 0);
1247 		atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1248 	}
1249 }
1250 
1251 void
1252 t4_intr(void *arg)
1253 {
1254 	struct sge_iq *iq = arg;
1255 
1256 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1257 		service_iq(iq, 0);
1258 		atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1259 	}
1260 }
1261 
1262 /*
1263  * Deals with anything and everything on the given ingress queue.
1264  */
1265 static int
1266 service_iq(struct sge_iq *iq, int budget)
1267 {
1268 	struct sge_iq *q;
1269 	struct sge_rxq *rxq = iq_to_rxq(iq);	/* Use iff iq is part of rxq */
1270 	struct sge_fl *fl;			/* Use iff IQ_HAS_FL */
1271 	struct adapter *sc = iq->adapter;
1272 	struct iq_desc *d = &iq->desc[iq->cidx];
1273 	int ndescs = 0, limit;
1274 	int rsp_type, refill;
1275 	uint32_t lq;
1276 	uint16_t fl_hw_cidx;
1277 	struct mbuf *m0;
1278 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1279 #if defined(INET) || defined(INET6)
1280 	const struct timeval lro_timeout = {0, sc->lro_timeout};
1281 #endif
1282 
1283 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1284 
1285 	limit = budget ? budget : iq->qsize / 16;
1286 
1287 	if (iq->flags & IQ_HAS_FL) {
1288 		fl = &rxq->fl;
1289 		fl_hw_cidx = fl->hw_cidx;	/* stable snapshot */
1290 	} else {
1291 		fl = NULL;
1292 		fl_hw_cidx = 0;			/* to silence gcc warning */
1293 	}
1294 
1295 	/*
1296 	 * We always come back and check the descriptor ring for new indirect
1297 	 * interrupts and other responses after running a single handler.
1298 	 */
1299 	for (;;) {
1300 		while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1301 
1302 			rmb();
1303 
1304 			refill = 0;
1305 			m0 = NULL;
1306 			rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1307 			lq = be32toh(d->rsp.pldbuflen_qid);
1308 
1309 			switch (rsp_type) {
1310 			case X_RSPD_TYPE_FLBUF:
1311 
1312 				KASSERT(iq->flags & IQ_HAS_FL,
1313 				    ("%s: data for an iq (%p) with no freelist",
1314 				    __func__, iq));
1315 
1316 				m0 = get_fl_payload(sc, fl, lq);
1317 				if (__predict_false(m0 == NULL))
1318 					goto process_iql;
1319 				refill = IDXDIFF(fl->hw_cidx, fl_hw_cidx, fl->sidx) > 2;
1320 #ifdef T4_PKT_TIMESTAMP
1321 				/*
1322 				 * 60 bit timestamp for the payload is
1323 				 * *(uint64_t *)m0->m_pktdat.  Note that it is
1324 				 * in the leading free-space in the mbuf.  The
1325 				 * kernel can clobber it during a pullup,
1326 				 * m_copymdata, etc.  You need to make sure that
1327 				 * the mbuf reaches you unmolested if you care
1328 				 * about the timestamp.
1329 				 */
1330 				*(uint64_t *)m0->m_pktdat =
1331 				    be64toh(ctrl->u.last_flit) &
1332 				    0xfffffffffffffff;
1333 #endif
1334 
1335 				/* fall through */
1336 
1337 			case X_RSPD_TYPE_CPL:
1338 				KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1339 				    ("%s: bad opcode %02x.", __func__,
1340 				    d->rss.opcode));
1341 				sc->cpl_handler[d->rss.opcode](iq, &d->rss, m0);
1342 				break;
1343 
1344 			case X_RSPD_TYPE_INTR:
1345 
1346 				/*
1347 				 * Interrupts should be forwarded only to queues
1348 				 * that are not forwarding their interrupts.
1349 				 * This means service_iq can recurse but only 1
1350 				 * level deep.
1351 				 */
1352 				KASSERT(budget == 0,
1353 				    ("%s: budget %u, rsp_type %u", __func__,
1354 				    budget, rsp_type));
1355 
1356 				/*
1357 				 * There are 1K interrupt-capable queues (qids 0
1358 				 * through 1023).  A response type indicating a
1359 				 * forwarded interrupt with a qid >= 1K is an
1360 				 * iWARP async notification.
1361 				 */
1362 				if (lq >= 1024) {
1363                                         sc->an_handler(iq, &d->rsp);
1364                                         break;
1365                                 }
1366 
1367 				q = sc->sge.iqmap[lq - sc->sge.iq_start];
1368 				if (atomic_cmpset_int(&q->state, IQS_IDLE,
1369 				    IQS_BUSY)) {
1370 					if (service_iq(q, q->qsize / 16) == 0) {
1371 						atomic_cmpset_int(&q->state,
1372 						    IQS_BUSY, IQS_IDLE);
1373 					} else {
1374 						STAILQ_INSERT_TAIL(&iql, q,
1375 						    link);
1376 					}
1377 				}
1378 				break;
1379 
1380 			default:
1381 				KASSERT(0,
1382 				    ("%s: illegal response type %d on iq %p",
1383 				    __func__, rsp_type, iq));
1384 				log(LOG_ERR,
1385 				    "%s: illegal response type %d on iq %p",
1386 				    device_get_nameunit(sc->dev), rsp_type, iq);
1387 				break;
1388 			}
1389 
1390 			d++;
1391 			if (__predict_false(++iq->cidx == iq->sidx)) {
1392 				iq->cidx = 0;
1393 				iq->gen ^= F_RSPD_GEN;
1394 				d = &iq->desc[0];
1395 			}
1396 			if (__predict_false(++ndescs == limit)) {
1397 				t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS),
1398 				    V_CIDXINC(ndescs) |
1399 				    V_INGRESSQID(iq->cntxt_id) |
1400 				    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1401 				ndescs = 0;
1402 
1403 #if defined(INET) || defined(INET6)
1404 				if (iq->flags & IQ_LRO_ENABLED &&
1405 				    sc->lro_timeout != 0) {
1406 					tcp_lro_flush_inactive(&rxq->lro,
1407 					    &lro_timeout);
1408 				}
1409 #endif
1410 
1411 				if (budget) {
1412 					if (iq->flags & IQ_HAS_FL) {
1413 						FL_LOCK(fl);
1414 						refill_fl(sc, fl, 32);
1415 						FL_UNLOCK(fl);
1416 					}
1417 					return (EINPROGRESS);
1418 				}
1419 			}
1420 			if (refill) {
1421 				FL_LOCK(fl);
1422 				refill_fl(sc, fl, 32);
1423 				FL_UNLOCK(fl);
1424 				fl_hw_cidx = fl->hw_cidx;
1425 			}
1426 		}
1427 
1428 process_iql:
1429 		if (STAILQ_EMPTY(&iql))
1430 			break;
1431 
1432 		/*
1433 		 * Process the head only, and send it to the back of the list if
1434 		 * it's still not done.
1435 		 */
1436 		q = STAILQ_FIRST(&iql);
1437 		STAILQ_REMOVE_HEAD(&iql, link);
1438 		if (service_iq(q, q->qsize / 8) == 0)
1439 			atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1440 		else
1441 			STAILQ_INSERT_TAIL(&iql, q, link);
1442 	}
1443 
1444 #if defined(INET) || defined(INET6)
1445 	if (iq->flags & IQ_LRO_ENABLED) {
1446 		struct lro_ctrl *lro = &rxq->lro;
1447 		struct lro_entry *l;
1448 
1449 		while (!SLIST_EMPTY(&lro->lro_active)) {
1450 			l = SLIST_FIRST(&lro->lro_active);
1451 			SLIST_REMOVE_HEAD(&lro->lro_active, next);
1452 			tcp_lro_flush(lro, l);
1453 		}
1454 	}
1455 #endif
1456 
1457 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_CIDXINC(ndescs) |
1458 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1459 
1460 	if (iq->flags & IQ_HAS_FL) {
1461 		int starved;
1462 
1463 		FL_LOCK(fl);
1464 		starved = refill_fl(sc, fl, 64);
1465 		FL_UNLOCK(fl);
1466 		if (__predict_false(starved != 0))
1467 			add_fl_to_sfl(sc, fl);
1468 	}
1469 
1470 	return (0);
1471 }
1472 
1473 static inline int
1474 cl_has_metadata(struct sge_fl *fl, struct cluster_layout *cll)
1475 {
1476 	int rc = fl->flags & FL_BUF_PACKING || cll->region1 > 0;
1477 
1478 	if (rc)
1479 		MPASS(cll->region3 >= CL_METADATA_SIZE);
1480 
1481 	return (rc);
1482 }
1483 
1484 static inline struct cluster_metadata *
1485 cl_metadata(struct adapter *sc, struct sge_fl *fl, struct cluster_layout *cll,
1486     caddr_t cl)
1487 {
1488 
1489 	if (cl_has_metadata(fl, cll)) {
1490 		struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1491 
1492 		return ((struct cluster_metadata *)(cl + swz->size) - 1);
1493 	}
1494 	return (NULL);
1495 }
1496 
1497 static void
1498 rxb_free(struct mbuf *m, void *arg1, void *arg2)
1499 {
1500 	uma_zone_t zone = arg1;
1501 	caddr_t cl = arg2;
1502 
1503 	uma_zfree(zone, cl);
1504 	counter_u64_add(extfree_rels, 1);
1505 }
1506 
1507 /*
1508  * The mbuf returned by this function could be allocated from zone_mbuf or
1509  * constructed in spare room in the cluster.
1510  *
1511  * The mbuf carries the payload in one of these ways
1512  * a) frame inside the mbuf (mbuf from zone_mbuf)
1513  * b) m_cljset (for clusters without metadata) zone_mbuf
1514  * c) m_extaddref (cluster with metadata) inline mbuf
1515  * d) m_extaddref (cluster with metadata) zone_mbuf
1516  */
1517 static struct mbuf *
1518 get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1519     int remaining)
1520 {
1521 	struct mbuf *m;
1522 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1523 	struct cluster_layout *cll = &sd->cll;
1524 	struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1525 	struct hw_buf_info *hwb = &sc->sge.hw_buf_info[cll->hwidx];
1526 	struct cluster_metadata *clm = cl_metadata(sc, fl, cll, sd->cl);
1527 	int len, blen;
1528 	caddr_t payload;
1529 
1530 	blen = hwb->size - fl->rx_offset;	/* max possible in this buf */
1531 	len = min(remaining, blen);
1532 	payload = sd->cl + cll->region1 + fl->rx_offset;
1533 	if (fl->flags & FL_BUF_PACKING) {
1534 		const u_int l = fr_offset + len;
1535 		const u_int pad = roundup2(l, fl->buf_boundary) - l;
1536 
1537 		if (fl->rx_offset + len + pad < hwb->size)
1538 			blen = len + pad;
1539 		MPASS(fl->rx_offset + blen <= hwb->size);
1540 	} else {
1541 		MPASS(fl->rx_offset == 0);	/* not packing */
1542 	}
1543 
1544 
1545 	if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) {
1546 
1547 		/*
1548 		 * Copy payload into a freshly allocated mbuf.
1549 		 */
1550 
1551 		m = fr_offset == 0 ?
1552 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1553 		if (m == NULL)
1554 			return (NULL);
1555 		fl->mbuf_allocated++;
1556 #ifdef T4_PKT_TIMESTAMP
1557 		/* Leave room for a timestamp */
1558 		m->m_data += 8;
1559 #endif
1560 		/* copy data to mbuf */
1561 		bcopy(payload, mtod(m, caddr_t), len);
1562 
1563 	} else if (sd->nmbuf * MSIZE < cll->region1) {
1564 
1565 		/*
1566 		 * There's spare room in the cluster for an mbuf.  Create one
1567 		 * and associate it with the payload that's in the cluster.
1568 		 */
1569 
1570 		MPASS(clm != NULL);
1571 		m = (struct mbuf *)(sd->cl + sd->nmbuf * MSIZE);
1572 		/* No bzero required */
1573 		if (m_init(m, NULL, 0, M_NOWAIT, MT_DATA,
1574 		    fr_offset == 0 ? M_PKTHDR | M_NOFREE : M_NOFREE))
1575 			return (NULL);
1576 		fl->mbuf_inlined++;
1577 		m_extaddref(m, payload, blen, &clm->refcount, rxb_free,
1578 		    swz->zone, sd->cl);
1579 		if (sd->nmbuf++ == 0)
1580 			counter_u64_add(extfree_refs, 1);
1581 
1582 	} else {
1583 
1584 		/*
1585 		 * Grab an mbuf from zone_mbuf and associate it with the
1586 		 * payload in the cluster.
1587 		 */
1588 
1589 		m = fr_offset == 0 ?
1590 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1591 		if (m == NULL)
1592 			return (NULL);
1593 		fl->mbuf_allocated++;
1594 		if (clm != NULL) {
1595 			m_extaddref(m, payload, blen, &clm->refcount,
1596 			    rxb_free, swz->zone, sd->cl);
1597 			if (sd->nmbuf++ == 0)
1598 				counter_u64_add(extfree_refs, 1);
1599 		} else {
1600 			m_cljset(m, sd->cl, swz->type);
1601 			sd->cl = NULL;	/* consumed, not a recycle candidate */
1602 		}
1603 	}
1604 	if (fr_offset == 0)
1605 		m->m_pkthdr.len = remaining;
1606 	m->m_len = len;
1607 
1608 	if (fl->flags & FL_BUF_PACKING) {
1609 		fl->rx_offset += blen;
1610 		MPASS(fl->rx_offset <= hwb->size);
1611 		if (fl->rx_offset < hwb->size)
1612 			return (m);	/* without advancing the cidx */
1613 	}
1614 
1615 	if (__predict_false(++fl->cidx % 8 == 0)) {
1616 		uint16_t cidx = fl->cidx / 8;
1617 
1618 		if (__predict_false(cidx == fl->sidx))
1619 			fl->cidx = cidx = 0;
1620 		fl->hw_cidx = cidx;
1621 	}
1622 	fl->rx_offset = 0;
1623 
1624 	return (m);
1625 }
1626 
1627 static struct mbuf *
1628 get_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf)
1629 {
1630 	struct mbuf *m0, *m, **pnext;
1631 	u_int remaining;
1632 	const u_int total = G_RSPD_LEN(len_newbuf);
1633 
1634 	if (__predict_false(fl->flags & FL_BUF_RESUME)) {
1635 		M_ASSERTPKTHDR(fl->m0);
1636 		MPASS(fl->m0->m_pkthdr.len == total);
1637 		MPASS(fl->remaining < total);
1638 
1639 		m0 = fl->m0;
1640 		pnext = fl->pnext;
1641 		remaining = fl->remaining;
1642 		fl->flags &= ~FL_BUF_RESUME;
1643 		goto get_segment;
1644 	}
1645 
1646 	if (fl->rx_offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
1647 		fl->rx_offset = 0;
1648 		if (__predict_false(++fl->cidx % 8 == 0)) {
1649 			uint16_t cidx = fl->cidx / 8;
1650 
1651 			if (__predict_false(cidx == fl->sidx))
1652 				fl->cidx = cidx = 0;
1653 			fl->hw_cidx = cidx;
1654 		}
1655 	}
1656 
1657 	/*
1658 	 * Payload starts at rx_offset in the current hw buffer.  Its length is
1659 	 * 'len' and it may span multiple hw buffers.
1660 	 */
1661 
1662 	m0 = get_scatter_segment(sc, fl, 0, total);
1663 	if (m0 == NULL)
1664 		return (NULL);
1665 	remaining = total - m0->m_len;
1666 	pnext = &m0->m_next;
1667 	while (remaining > 0) {
1668 get_segment:
1669 		MPASS(fl->rx_offset == 0);
1670 		m = get_scatter_segment(sc, fl, total - remaining, remaining);
1671 		if (__predict_false(m == NULL)) {
1672 			fl->m0 = m0;
1673 			fl->pnext = pnext;
1674 			fl->remaining = remaining;
1675 			fl->flags |= FL_BUF_RESUME;
1676 			return (NULL);
1677 		}
1678 		*pnext = m;
1679 		pnext = &m->m_next;
1680 		remaining -= m->m_len;
1681 	}
1682 	*pnext = NULL;
1683 
1684 	M_ASSERTPKTHDR(m0);
1685 	return (m0);
1686 }
1687 
1688 static int
1689 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
1690 {
1691 	struct sge_rxq *rxq = iq_to_rxq(iq);
1692 	struct ifnet *ifp = rxq->ifp;
1693 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
1694 #if defined(INET) || defined(INET6)
1695 	struct lro_ctrl *lro = &rxq->lro;
1696 #endif
1697 	static const int sw_hashtype[4][2] = {
1698 		{M_HASHTYPE_NONE, M_HASHTYPE_NONE},
1699 		{M_HASHTYPE_RSS_IPV4, M_HASHTYPE_RSS_IPV6},
1700 		{M_HASHTYPE_RSS_TCP_IPV4, M_HASHTYPE_RSS_TCP_IPV6},
1701 		{M_HASHTYPE_RSS_UDP_IPV4, M_HASHTYPE_RSS_UDP_IPV6},
1702 	};
1703 
1704 	KASSERT(m0 != NULL, ("%s: no payload with opcode %02x", __func__,
1705 	    rss->opcode));
1706 
1707 	m0->m_pkthdr.len -= fl_pktshift;
1708 	m0->m_len -= fl_pktshift;
1709 	m0->m_data += fl_pktshift;
1710 
1711 	m0->m_pkthdr.rcvif = ifp;
1712 	M_HASHTYPE_SET(m0, sw_hashtype[rss->hash_type][rss->ipv6]);
1713 	m0->m_pkthdr.flowid = be32toh(rss->hash_val);
1714 
1715 	if (cpl->csum_calc && !cpl->err_vec) {
1716 		if (ifp->if_capenable & IFCAP_RXCSUM &&
1717 		    cpl->l2info & htobe32(F_RXF_IP)) {
1718 			m0->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
1719 			    CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
1720 			rxq->rxcsum++;
1721 		} else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
1722 		    cpl->l2info & htobe32(F_RXF_IP6)) {
1723 			m0->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
1724 			    CSUM_PSEUDO_HDR);
1725 			rxq->rxcsum++;
1726 		}
1727 
1728 		if (__predict_false(cpl->ip_frag))
1729 			m0->m_pkthdr.csum_data = be16toh(cpl->csum);
1730 		else
1731 			m0->m_pkthdr.csum_data = 0xffff;
1732 	}
1733 
1734 	if (cpl->vlan_ex) {
1735 		m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
1736 		m0->m_flags |= M_VLANTAG;
1737 		rxq->vlan_extraction++;
1738 	}
1739 
1740 #if defined(INET) || defined(INET6)
1741 	if (cpl->l2info & htobe32(F_RXF_LRO) &&
1742 	    iq->flags & IQ_LRO_ENABLED &&
1743 	    tcp_lro_rx(lro, m0, 0) == 0) {
1744 		/* queued for LRO */
1745 	} else
1746 #endif
1747 	ifp->if_input(ifp, m0);
1748 
1749 	return (0);
1750 }
1751 
1752 /*
1753  * Must drain the wrq or make sure that someone else will.
1754  */
1755 static void
1756 wrq_tx_drain(void *arg, int n)
1757 {
1758 	struct sge_wrq *wrq = arg;
1759 	struct sge_eq *eq = &wrq->eq;
1760 
1761 	EQ_LOCK(eq);
1762 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
1763 		drain_wrq_wr_list(wrq->adapter, wrq);
1764 	EQ_UNLOCK(eq);
1765 }
1766 
1767 static void
1768 drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq)
1769 {
1770 	struct sge_eq *eq = &wrq->eq;
1771 	u_int available, dbdiff;	/* # of hardware descriptors */
1772 	u_int n;
1773 	struct wrqe *wr;
1774 	struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
1775 
1776 	EQ_LOCK_ASSERT_OWNED(eq);
1777 	MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
1778 	wr = STAILQ_FIRST(&wrq->wr_list);
1779 	MPASS(wr != NULL);	/* Must be called with something useful to do */
1780 	dbdiff = IDXDIFF(eq->pidx, eq->dbidx, eq->sidx);
1781 
1782 	do {
1783 		eq->cidx = read_hw_cidx(eq);
1784 		if (eq->pidx == eq->cidx)
1785 			available = eq->sidx - 1;
1786 		else
1787 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
1788 
1789 		MPASS(wr->wrq == wrq);
1790 		n = howmany(wr->wr_len, EQ_ESIZE);
1791 		if (available < n)
1792 			return;
1793 
1794 		dst = (void *)&eq->desc[eq->pidx];
1795 		if (__predict_true(eq->sidx - eq->pidx > n)) {
1796 			/* Won't wrap, won't end exactly at the status page. */
1797 			bcopy(&wr->wr[0], dst, wr->wr_len);
1798 			eq->pidx += n;
1799 		} else {
1800 			int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE;
1801 
1802 			bcopy(&wr->wr[0], dst, first_portion);
1803 			if (wr->wr_len > first_portion) {
1804 				bcopy(&wr->wr[first_portion], &eq->desc[0],
1805 				    wr->wr_len - first_portion);
1806 			}
1807 			eq->pidx = n - (eq->sidx - eq->pidx);
1808 		}
1809 
1810 		if (available < eq->sidx / 4 &&
1811 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
1812 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
1813 			    F_FW_WR_EQUEQ);
1814 			eq->equeqidx = eq->pidx;
1815 		} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
1816 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
1817 			eq->equeqidx = eq->pidx;
1818 		}
1819 
1820 		dbdiff += n;
1821 		if (dbdiff >= 16) {
1822 			ring_eq_db(sc, eq, dbdiff);
1823 			dbdiff = 0;
1824 		}
1825 
1826 		STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
1827 		free_wrqe(wr);
1828 		MPASS(wrq->nwr_pending > 0);
1829 		wrq->nwr_pending--;
1830 		MPASS(wrq->ndesc_needed >= n);
1831 		wrq->ndesc_needed -= n;
1832 	} while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL);
1833 
1834 	if (dbdiff)
1835 		ring_eq_db(sc, eq, dbdiff);
1836 }
1837 
1838 /*
1839  * Doesn't fail.  Holds on to work requests it can't send right away.
1840  */
1841 void
1842 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
1843 {
1844 #ifdef INVARIANTS
1845 	struct sge_eq *eq = &wrq->eq;
1846 #endif
1847 
1848 	EQ_LOCK_ASSERT_OWNED(eq);
1849 	MPASS(wr != NULL);
1850 	MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN);
1851 	MPASS((wr->wr_len & 0x7) == 0);
1852 
1853 	STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
1854 	wrq->nwr_pending++;
1855 	wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE);
1856 
1857 	if (!TAILQ_EMPTY(&wrq->incomplete_wrs))
1858 		return;	/* commit_wrq_wr will drain wr_list as well. */
1859 
1860 	drain_wrq_wr_list(sc, wrq);
1861 
1862 	/* Doorbell must have caught up to the pidx. */
1863 	MPASS(eq->pidx == eq->dbidx);
1864 }
1865 
1866 void
1867 t4_update_fl_bufsize(struct ifnet *ifp)
1868 {
1869 	struct vi_info *vi = ifp->if_softc;
1870 	struct adapter *sc = vi->pi->adapter;
1871 	struct sge_rxq *rxq;
1872 #ifdef TCP_OFFLOAD
1873 	struct sge_ofld_rxq *ofld_rxq;
1874 #endif
1875 	struct sge_fl *fl;
1876 	int i, maxp, mtu = ifp->if_mtu;
1877 
1878 	maxp = mtu_to_max_payload(sc, mtu, 0);
1879 	for_each_rxq(vi, i, rxq) {
1880 		fl = &rxq->fl;
1881 
1882 		FL_LOCK(fl);
1883 		find_best_refill_source(sc, fl, maxp);
1884 		FL_UNLOCK(fl);
1885 	}
1886 #ifdef TCP_OFFLOAD
1887 	maxp = mtu_to_max_payload(sc, mtu, 1);
1888 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1889 		fl = &ofld_rxq->fl;
1890 
1891 		FL_LOCK(fl);
1892 		find_best_refill_source(sc, fl, maxp);
1893 		FL_UNLOCK(fl);
1894 	}
1895 #endif
1896 }
1897 
1898 static inline int
1899 mbuf_nsegs(struct mbuf *m)
1900 {
1901 
1902 	M_ASSERTPKTHDR(m);
1903 	KASSERT(m->m_pkthdr.l5hlen > 0,
1904 	    ("%s: mbuf %p missing information on # of segments.", __func__, m));
1905 
1906 	return (m->m_pkthdr.l5hlen);
1907 }
1908 
1909 static inline void
1910 set_mbuf_nsegs(struct mbuf *m, uint8_t nsegs)
1911 {
1912 
1913 	M_ASSERTPKTHDR(m);
1914 	m->m_pkthdr.l5hlen = nsegs;
1915 }
1916 
1917 static inline int
1918 mbuf_len16(struct mbuf *m)
1919 {
1920 	int n;
1921 
1922 	M_ASSERTPKTHDR(m);
1923 	n = m->m_pkthdr.PH_loc.eight[0];
1924 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
1925 
1926 	return (n);
1927 }
1928 
1929 static inline void
1930 set_mbuf_len16(struct mbuf *m, uint8_t len16)
1931 {
1932 
1933 	M_ASSERTPKTHDR(m);
1934 	m->m_pkthdr.PH_loc.eight[0] = len16;
1935 }
1936 
1937 static inline int
1938 needs_tso(struct mbuf *m)
1939 {
1940 
1941 	M_ASSERTPKTHDR(m);
1942 
1943 	if (m->m_pkthdr.csum_flags & CSUM_TSO) {
1944 		KASSERT(m->m_pkthdr.tso_segsz > 0,
1945 		    ("%s: TSO requested in mbuf %p but MSS not provided",
1946 		    __func__, m));
1947 		return (1);
1948 	}
1949 
1950 	return (0);
1951 }
1952 
1953 static inline int
1954 needs_l3_csum(struct mbuf *m)
1955 {
1956 
1957 	M_ASSERTPKTHDR(m);
1958 
1959 	if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO))
1960 		return (1);
1961 	return (0);
1962 }
1963 
1964 static inline int
1965 needs_l4_csum(struct mbuf *m)
1966 {
1967 
1968 	M_ASSERTPKTHDR(m);
1969 
1970 	if (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
1971 	    CSUM_TCP_IPV6 | CSUM_TSO))
1972 		return (1);
1973 	return (0);
1974 }
1975 
1976 static inline int
1977 needs_vlan_insertion(struct mbuf *m)
1978 {
1979 
1980 	M_ASSERTPKTHDR(m);
1981 
1982 	if (m->m_flags & M_VLANTAG) {
1983 		KASSERT(m->m_pkthdr.ether_vtag != 0,
1984 		    ("%s: HWVLAN requested in mbuf %p but tag not provided",
1985 		    __func__, m));
1986 		return (1);
1987 	}
1988 	return (0);
1989 }
1990 
1991 static void *
1992 m_advance(struct mbuf **pm, int *poffset, int len)
1993 {
1994 	struct mbuf *m = *pm;
1995 	int offset = *poffset;
1996 	uintptr_t p = 0;
1997 
1998 	MPASS(len > 0);
1999 
2000 	while (len) {
2001 		if (offset + len < m->m_len) {
2002 			offset += len;
2003 			p = mtod(m, uintptr_t) + offset;
2004 			break;
2005 		}
2006 		len -= m->m_len - offset;
2007 		m = m->m_next;
2008 		offset = 0;
2009 		MPASS(m != NULL);
2010 	}
2011 	*poffset = offset;
2012 	*pm = m;
2013 	return ((void *)p);
2014 }
2015 
2016 static inline int
2017 same_paddr(char *a, char *b)
2018 {
2019 
2020 	if (a == b)
2021 		return (1);
2022 	else if (a != NULL && b != NULL) {
2023 		vm_offset_t x = (vm_offset_t)a;
2024 		vm_offset_t y = (vm_offset_t)b;
2025 
2026 		if ((x & PAGE_MASK) == (y & PAGE_MASK) &&
2027 		    pmap_kextract(x) == pmap_kextract(y))
2028 			return (1);
2029 	}
2030 
2031 	return (0);
2032 }
2033 
2034 /*
2035  * Can deal with empty mbufs in the chain that have m_len = 0, but the chain
2036  * must have at least one mbuf that's not empty.
2037  */
2038 static inline int
2039 count_mbuf_nsegs(struct mbuf *m)
2040 {
2041 	char *prev_end, *start;
2042 	int len, nsegs;
2043 
2044 	MPASS(m != NULL);
2045 
2046 	nsegs = 0;
2047 	prev_end = NULL;
2048 	for (; m; m = m->m_next) {
2049 
2050 		len = m->m_len;
2051 		if (__predict_false(len == 0))
2052 			continue;
2053 		start = mtod(m, char *);
2054 
2055 		nsegs += sglist_count(start, len);
2056 		if (same_paddr(prev_end, start))
2057 			nsegs--;
2058 		prev_end = start + len;
2059 	}
2060 
2061 	MPASS(nsegs > 0);
2062 	return (nsegs);
2063 }
2064 
2065 /*
2066  * Analyze the mbuf to determine its tx needs.  The mbuf passed in may change:
2067  * a) caller can assume it's been freed if this function returns with an error.
2068  * b) it may get defragged up if the gather list is too long for the hardware.
2069  */
2070 int
2071 parse_pkt(struct mbuf **mp)
2072 {
2073 	struct mbuf *m0 = *mp, *m;
2074 	int rc, nsegs, defragged = 0, offset;
2075 	struct ether_header *eh;
2076 	void *l3hdr;
2077 #if defined(INET) || defined(INET6)
2078 	struct tcphdr *tcp;
2079 #endif
2080 	uint16_t eh_type;
2081 
2082 	M_ASSERTPKTHDR(m0);
2083 	if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) {
2084 		rc = EINVAL;
2085 fail:
2086 		m_freem(m0);
2087 		*mp = NULL;
2088 		return (rc);
2089 	}
2090 restart:
2091 	/*
2092 	 * First count the number of gather list segments in the payload.
2093 	 * Defrag the mbuf if nsegs exceeds the hardware limit.
2094 	 */
2095 	M_ASSERTPKTHDR(m0);
2096 	MPASS(m0->m_pkthdr.len > 0);
2097 	nsegs = count_mbuf_nsegs(m0);
2098 	if (nsegs > (needs_tso(m0) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS)) {
2099 		if (defragged++ > 0 || (m = m_defrag(m0, M_NOWAIT)) == NULL) {
2100 			rc = EFBIG;
2101 			goto fail;
2102 		}
2103 		*mp = m0 = m;	/* update caller's copy after defrag */
2104 		goto restart;
2105 	}
2106 
2107 	if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN)) {
2108 		m0 = m_pullup(m0, m0->m_pkthdr.len);
2109 		if (m0 == NULL) {
2110 			/* Should have left well enough alone. */
2111 			rc = EFBIG;
2112 			goto fail;
2113 		}
2114 		*mp = m0;	/* update caller's copy after pullup */
2115 		goto restart;
2116 	}
2117 	set_mbuf_nsegs(m0, nsegs);
2118 	set_mbuf_len16(m0, txpkt_len16(nsegs, needs_tso(m0)));
2119 
2120 	if (!needs_tso(m0))
2121 		return (0);
2122 
2123 	m = m0;
2124 	eh = mtod(m, struct ether_header *);
2125 	eh_type = ntohs(eh->ether_type);
2126 	if (eh_type == ETHERTYPE_VLAN) {
2127 		struct ether_vlan_header *evh = (void *)eh;
2128 
2129 		eh_type = ntohs(evh->evl_proto);
2130 		m0->m_pkthdr.l2hlen = sizeof(*evh);
2131 	} else
2132 		m0->m_pkthdr.l2hlen = sizeof(*eh);
2133 
2134 	offset = 0;
2135 	l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2136 
2137 	switch (eh_type) {
2138 #ifdef INET6
2139 	case ETHERTYPE_IPV6:
2140 	{
2141 		struct ip6_hdr *ip6 = l3hdr;
2142 
2143 		MPASS(ip6->ip6_nxt == IPPROTO_TCP);
2144 
2145 		m0->m_pkthdr.l3hlen = sizeof(*ip6);
2146 		break;
2147 	}
2148 #endif
2149 #ifdef INET
2150 	case ETHERTYPE_IP:
2151 	{
2152 		struct ip *ip = l3hdr;
2153 
2154 		m0->m_pkthdr.l3hlen = ip->ip_hl * 4;
2155 		break;
2156 	}
2157 #endif
2158 	default:
2159 		panic("%s: ethertype 0x%04x unknown.  if_cxgbe must be compiled"
2160 		    " with the same INET/INET6 options as the kernel.",
2161 		    __func__, eh_type);
2162 	}
2163 
2164 #if defined(INET) || defined(INET6)
2165 	tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen);
2166 	m0->m_pkthdr.l4hlen = tcp->th_off * 4;
2167 #endif
2168 	MPASS(m0 == *mp);
2169 	return (0);
2170 }
2171 
2172 void *
2173 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie)
2174 {
2175 	struct sge_eq *eq = &wrq->eq;
2176 	struct adapter *sc = wrq->adapter;
2177 	int ndesc, available;
2178 	struct wrqe *wr;
2179 	void *w;
2180 
2181 	MPASS(len16 > 0);
2182 	ndesc = howmany(len16, EQ_ESIZE / 16);
2183 	MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC);
2184 
2185 	EQ_LOCK(eq);
2186 
2187 	if (!STAILQ_EMPTY(&wrq->wr_list))
2188 		drain_wrq_wr_list(sc, wrq);
2189 
2190 	if (!STAILQ_EMPTY(&wrq->wr_list)) {
2191 slowpath:
2192 		EQ_UNLOCK(eq);
2193 		wr = alloc_wrqe(len16 * 16, wrq);
2194 		if (__predict_false(wr == NULL))
2195 			return (NULL);
2196 		cookie->pidx = -1;
2197 		cookie->ndesc = ndesc;
2198 		return (&wr->wr);
2199 	}
2200 
2201 	eq->cidx = read_hw_cidx(eq);
2202 	if (eq->pidx == eq->cidx)
2203 		available = eq->sidx - 1;
2204 	else
2205 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2206 	if (available < ndesc)
2207 		goto slowpath;
2208 
2209 	cookie->pidx = eq->pidx;
2210 	cookie->ndesc = ndesc;
2211 	TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link);
2212 
2213 	w = &eq->desc[eq->pidx];
2214 	IDXINCR(eq->pidx, ndesc, eq->sidx);
2215 	if (__predict_false(eq->pidx < ndesc - 1)) {
2216 		w = &wrq->ss[0];
2217 		wrq->ss_pidx = cookie->pidx;
2218 		wrq->ss_len = len16 * 16;
2219 	}
2220 
2221 	EQ_UNLOCK(eq);
2222 
2223 	return (w);
2224 }
2225 
2226 void
2227 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie)
2228 {
2229 	struct sge_eq *eq = &wrq->eq;
2230 	struct adapter *sc = wrq->adapter;
2231 	int ndesc, pidx;
2232 	struct wrq_cookie *prev, *next;
2233 
2234 	if (cookie->pidx == -1) {
2235 		struct wrqe *wr = __containerof(w, struct wrqe, wr);
2236 
2237 		t4_wrq_tx(sc, wr);
2238 		return;
2239 	}
2240 
2241 	ndesc = cookie->ndesc;	/* Can be more than SGE_MAX_WR_NDESC here. */
2242 	pidx = cookie->pidx;
2243 	MPASS(pidx >= 0 && pidx < eq->sidx);
2244 	if (__predict_false(w == &wrq->ss[0])) {
2245 		int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE;
2246 
2247 		MPASS(wrq->ss_len > n);	/* WR had better wrap around. */
2248 		bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n);
2249 		bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n);
2250 		wrq->tx_wrs_ss++;
2251 	} else
2252 		wrq->tx_wrs_direct++;
2253 
2254 	EQ_LOCK(eq);
2255 	prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link);
2256 	next = TAILQ_NEXT(cookie, link);
2257 	if (prev == NULL) {
2258 		MPASS(pidx == eq->dbidx);
2259 		if (next == NULL || ndesc >= 16)
2260 			ring_eq_db(wrq->adapter, eq, ndesc);
2261 		else {
2262 			MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc);
2263 			next->pidx = pidx;
2264 			next->ndesc += ndesc;
2265 		}
2266 	} else {
2267 		MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc);
2268 		prev->ndesc += ndesc;
2269 	}
2270 	TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link);
2271 
2272 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2273 		drain_wrq_wr_list(sc, wrq);
2274 
2275 #ifdef INVARIANTS
2276 	if (TAILQ_EMPTY(&wrq->incomplete_wrs)) {
2277 		/* Doorbell must have caught up to the pidx. */
2278 		MPASS(wrq->eq.pidx == wrq->eq.dbidx);
2279 	}
2280 #endif
2281 	EQ_UNLOCK(eq);
2282 }
2283 
2284 static u_int
2285 can_resume_eth_tx(struct mp_ring *r)
2286 {
2287 	struct sge_eq *eq = r->cookie;
2288 
2289 	return (total_available_tx_desc(eq) > eq->sidx / 8);
2290 }
2291 
2292 static inline int
2293 cannot_use_txpkts(struct mbuf *m)
2294 {
2295 	/* maybe put a GL limit too, to avoid silliness? */
2296 
2297 	return (needs_tso(m));
2298 }
2299 
2300 /*
2301  * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to
2302  * be consumed.  Return the actual number consumed.  0 indicates a stall.
2303  */
2304 static u_int
2305 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx)
2306 {
2307 	struct sge_txq *txq = r->cookie;
2308 	struct sge_eq *eq = &txq->eq;
2309 	struct ifnet *ifp = txq->ifp;
2310 	struct vi_info *vi = ifp->if_softc;
2311 	struct port_info *pi = vi->pi;
2312 	struct adapter *sc = pi->adapter;
2313 	u_int total, remaining;		/* # of packets */
2314 	u_int available, dbdiff;	/* # of hardware descriptors */
2315 	u_int n, next_cidx;
2316 	struct mbuf *m0, *tail;
2317 	struct txpkts txp;
2318 	struct fw_eth_tx_pkts_wr *wr;	/* any fw WR struct will do */
2319 
2320 	remaining = IDXDIFF(pidx, cidx, r->size);
2321 	MPASS(remaining > 0);	/* Must not be called without work to do. */
2322 	total = 0;
2323 
2324 	TXQ_LOCK(txq);
2325 	if (__predict_false((eq->flags & EQ_ENABLED) == 0)) {
2326 		while (cidx != pidx) {
2327 			m0 = r->items[cidx];
2328 			m_freem(m0);
2329 			if (++cidx == r->size)
2330 				cidx = 0;
2331 		}
2332 		reclaim_tx_descs(txq, 2048);
2333 		total = remaining;
2334 		goto done;
2335 	}
2336 
2337 	/* How many hardware descriptors do we have readily available. */
2338 	if (eq->pidx == eq->cidx)
2339 		available = eq->sidx - 1;
2340 	else
2341 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2342 	dbdiff = IDXDIFF(eq->pidx, eq->dbidx, eq->sidx);
2343 
2344 	while (remaining > 0) {
2345 
2346 		m0 = r->items[cidx];
2347 		M_ASSERTPKTHDR(m0);
2348 		MPASS(m0->m_nextpkt == NULL);
2349 
2350 		if (available < SGE_MAX_WR_NDESC) {
2351 			available += reclaim_tx_descs(txq, 64);
2352 			if (available < howmany(mbuf_len16(m0), EQ_ESIZE / 16))
2353 				break;	/* out of descriptors */
2354 		}
2355 
2356 		next_cidx = cidx + 1;
2357 		if (__predict_false(next_cidx == r->size))
2358 			next_cidx = 0;
2359 
2360 		wr = (void *)&eq->desc[eq->pidx];
2361 		if (remaining > 1 &&
2362 		    try_txpkts(m0, r->items[next_cidx], &txp, available) == 0) {
2363 
2364 			/* pkts at cidx, next_cidx should both be in txp. */
2365 			MPASS(txp.npkt == 2);
2366 			tail = r->items[next_cidx];
2367 			MPASS(tail->m_nextpkt == NULL);
2368 			ETHER_BPF_MTAP(ifp, m0);
2369 			ETHER_BPF_MTAP(ifp, tail);
2370 			m0->m_nextpkt = tail;
2371 
2372 			if (__predict_false(++next_cidx == r->size))
2373 				next_cidx = 0;
2374 
2375 			while (next_cidx != pidx) {
2376 				if (add_to_txpkts(r->items[next_cidx], &txp,
2377 				    available) != 0)
2378 					break;
2379 				tail->m_nextpkt = r->items[next_cidx];
2380 				tail = tail->m_nextpkt;
2381 				ETHER_BPF_MTAP(ifp, tail);
2382 				if (__predict_false(++next_cidx == r->size))
2383 					next_cidx = 0;
2384 			}
2385 
2386 			n = write_txpkts_wr(txq, wr, m0, &txp, available);
2387 			total += txp.npkt;
2388 			remaining -= txp.npkt;
2389 		} else {
2390 			total++;
2391 			remaining--;
2392 			n = write_txpkt_wr(txq, (void *)wr, m0, available);
2393 			ETHER_BPF_MTAP(ifp, m0);
2394 		}
2395 		MPASS(n >= 1 && n <= available && n <= SGE_MAX_WR_NDESC);
2396 
2397 		available -= n;
2398 		dbdiff += n;
2399 		IDXINCR(eq->pidx, n, eq->sidx);
2400 
2401 		if (total_available_tx_desc(eq) < eq->sidx / 4 &&
2402 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2403 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2404 			    F_FW_WR_EQUEQ);
2405 			eq->equeqidx = eq->pidx;
2406 		} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) {
2407 			wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
2408 			eq->equeqidx = eq->pidx;
2409 		}
2410 
2411 		if (dbdiff >= 16 && remaining >= 4) {
2412 			ring_eq_db(sc, eq, dbdiff);
2413 			available += reclaim_tx_descs(txq, 4 * dbdiff);
2414 			dbdiff = 0;
2415 		}
2416 
2417 		cidx = next_cidx;
2418 	}
2419 	if (dbdiff != 0) {
2420 		ring_eq_db(sc, eq, dbdiff);
2421 		reclaim_tx_descs(txq, 32);
2422 	}
2423 done:
2424 	TXQ_UNLOCK(txq);
2425 
2426 	return (total);
2427 }
2428 
2429 static inline void
2430 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
2431     int qsize)
2432 {
2433 
2434 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
2435 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
2436 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
2437 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
2438 
2439 	iq->flags = 0;
2440 	iq->adapter = sc;
2441 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
2442 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
2443 	if (pktc_idx >= 0) {
2444 		iq->intr_params |= F_QINTR_CNT_EN;
2445 		iq->intr_pktc_idx = pktc_idx;
2446 	}
2447 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
2448 	iq->sidx = iq->qsize - spg_len / IQ_ESIZE;
2449 }
2450 
2451 static inline void
2452 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name)
2453 {
2454 
2455 	fl->qsize = qsize;
2456 	fl->sidx = qsize - spg_len / EQ_ESIZE;
2457 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
2458 	if (sc->flags & BUF_PACKING_OK &&
2459 	    ((!is_t4(sc) && buffer_packing) ||	/* T5+: enabled unless 0 */
2460 	    (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */
2461 		fl->flags |= FL_BUF_PACKING;
2462 	find_best_refill_source(sc, fl, maxp);
2463 	find_safe_refill_source(sc, fl);
2464 }
2465 
2466 static inline void
2467 init_eq(struct sge_eq *eq, int eqtype, int qsize, uint8_t tx_chan,
2468     uint16_t iqid, char *name)
2469 {
2470 	KASSERT(tx_chan < NCHAN, ("%s: bad tx channel %d", __func__, tx_chan));
2471 	KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
2472 
2473 	eq->flags = eqtype & EQ_TYPEMASK;
2474 	eq->tx_chan = tx_chan;
2475 	eq->iqid = iqid;
2476 	eq->sidx = qsize - spg_len / EQ_ESIZE;
2477 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
2478 }
2479 
2480 static int
2481 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
2482     bus_dmamap_t *map, bus_addr_t *pa, void **va)
2483 {
2484 	int rc;
2485 
2486 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
2487 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
2488 	if (rc != 0) {
2489 		device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
2490 		goto done;
2491 	}
2492 
2493 	rc = bus_dmamem_alloc(*tag, va,
2494 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
2495 	if (rc != 0) {
2496 		device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
2497 		goto done;
2498 	}
2499 
2500 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
2501 	if (rc != 0) {
2502 		device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
2503 		goto done;
2504 	}
2505 done:
2506 	if (rc)
2507 		free_ring(sc, *tag, *map, *pa, *va);
2508 
2509 	return (rc);
2510 }
2511 
2512 static int
2513 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
2514     bus_addr_t pa, void *va)
2515 {
2516 	if (pa)
2517 		bus_dmamap_unload(tag, map);
2518 	if (va)
2519 		bus_dmamem_free(tag, va, map);
2520 	if (tag)
2521 		bus_dma_tag_destroy(tag);
2522 
2523 	return (0);
2524 }
2525 
2526 /*
2527  * Allocates the ring for an ingress queue and an optional freelist.  If the
2528  * freelist is specified it will be allocated and then associated with the
2529  * ingress queue.
2530  *
2531  * Returns errno on failure.  Resources allocated up to that point may still be
2532  * allocated.  Caller is responsible for cleanup in case this function fails.
2533  *
2534  * If the ingress queue will take interrupts directly (iq->flags & IQ_INTR) then
2535  * the intr_idx specifies the vector, starting from 0.  Otherwise it specifies
2536  * the abs_id of the ingress queue to which its interrupts should be forwarded.
2537  */
2538 static int
2539 alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl,
2540     int intr_idx, int cong)
2541 {
2542 	int rc, i, cntxt_id;
2543 	size_t len;
2544 	struct fw_iq_cmd c;
2545 	struct port_info *pi = vi->pi;
2546 	struct adapter *sc = iq->adapter;
2547 	__be32 v = 0;
2548 
2549 	len = iq->qsize * IQ_ESIZE;
2550 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
2551 	    (void **)&iq->desc);
2552 	if (rc != 0)
2553 		return (rc);
2554 
2555 	bzero(&c, sizeof(c));
2556 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
2557 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
2558 	    V_FW_IQ_CMD_VFN(0));
2559 
2560 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
2561 	    FW_LEN16(c));
2562 
2563 	/* Special handling for firmware event queue */
2564 	if (iq == &sc->sge.fwq)
2565 		v |= F_FW_IQ_CMD_IQASYNCH;
2566 
2567 	if (iq->flags & IQ_INTR) {
2568 		KASSERT(intr_idx < sc->intr_count,
2569 		    ("%s: invalid direct intr_idx %d", __func__, intr_idx));
2570 	} else
2571 		v |= F_FW_IQ_CMD_IQANDST;
2572 	v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
2573 
2574 	c.type_to_iqandstindex = htobe32(v |
2575 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
2576 	    V_FW_IQ_CMD_VIID(vi->viid) |
2577 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
2578 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
2579 	    F_FW_IQ_CMD_IQGTSMODE |
2580 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
2581 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
2582 	c.iqsize = htobe16(iq->qsize);
2583 	c.iqaddr = htobe64(iq->ba);
2584 	if (cong >= 0)
2585 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
2586 
2587 	if (fl) {
2588 		mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
2589 
2590 		len = fl->qsize * EQ_ESIZE;
2591 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
2592 		    &fl->ba, (void **)&fl->desc);
2593 		if (rc)
2594 			return (rc);
2595 
2596 		/* Allocate space for one software descriptor per buffer. */
2597 		rc = alloc_fl_sdesc(fl);
2598 		if (rc != 0) {
2599 			device_printf(sc->dev,
2600 			    "failed to setup fl software descriptors: %d\n",
2601 			    rc);
2602 			return (rc);
2603 		}
2604 
2605 		if (fl->flags & FL_BUF_PACKING) {
2606 			fl->lowat = roundup2(sc->sge.fl_starve_threshold2, 8);
2607 			fl->buf_boundary = sc->sge.pack_boundary;
2608 		} else {
2609 			fl->lowat = roundup2(sc->sge.fl_starve_threshold, 8);
2610 			fl->buf_boundary = 16;
2611 		}
2612 		if (fl_pad && fl->buf_boundary < sc->sge.pad_boundary)
2613 			fl->buf_boundary = sc->sge.pad_boundary;
2614 
2615 		c.iqns_to_fl0congen |=
2616 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
2617 			F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
2618 			(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
2619 			(fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
2620 			    0));
2621 		if (cong >= 0) {
2622 			c.iqns_to_fl0congen |=
2623 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
2624 				    F_FW_IQ_CMD_FL0CONGCIF |
2625 				    F_FW_IQ_CMD_FL0CONGEN);
2626 		}
2627 		c.fl0dcaen_to_fl0cidxfthresh =
2628 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(X_FETCHBURSTMIN_128B) |
2629 			V_FW_IQ_CMD_FL0FBMAX(X_FETCHBURSTMAX_512B));
2630 		c.fl0size = htobe16(fl->qsize);
2631 		c.fl0addr = htobe64(fl->ba);
2632 	}
2633 
2634 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
2635 	if (rc != 0) {
2636 		device_printf(sc->dev,
2637 		    "failed to create ingress queue: %d\n", rc);
2638 		return (rc);
2639 	}
2640 
2641 	iq->cidx = 0;
2642 	iq->gen = F_RSPD_GEN;
2643 	iq->intr_next = iq->intr_params;
2644 	iq->cntxt_id = be16toh(c.iqid);
2645 	iq->abs_id = be16toh(c.physiqid);
2646 	iq->flags |= IQ_ALLOCATED;
2647 
2648 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
2649 	if (cntxt_id >= sc->sge.niq) {
2650 		panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
2651 		    cntxt_id, sc->sge.niq - 1);
2652 	}
2653 	sc->sge.iqmap[cntxt_id] = iq;
2654 
2655 	if (fl) {
2656 		u_int qid;
2657 
2658 		iq->flags |= IQ_HAS_FL;
2659 		fl->cntxt_id = be16toh(c.fl0id);
2660 		fl->pidx = fl->cidx = 0;
2661 
2662 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
2663 		if (cntxt_id >= sc->sge.neq) {
2664 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
2665 			    __func__, cntxt_id, sc->sge.neq - 1);
2666 		}
2667 		sc->sge.eqmap[cntxt_id] = (void *)fl;
2668 
2669 		qid = fl->cntxt_id;
2670 		if (isset(&sc->doorbells, DOORBELL_UDB)) {
2671 			uint32_t s_qpp = sc->sge.eq_s_qpp;
2672 			uint32_t mask = (1 << s_qpp) - 1;
2673 			volatile uint8_t *udb;
2674 
2675 			udb = sc->udbs_base + UDBS_DB_OFFSET;
2676 			udb += (qid >> s_qpp) << PAGE_SHIFT;
2677 			qid &= mask;
2678 			if (qid < PAGE_SIZE / UDBS_SEG_SIZE) {
2679 				udb += qid << UDBS_SEG_SHIFT;
2680 				qid = 0;
2681 			}
2682 			fl->udb = (volatile void *)udb;
2683 		}
2684 		fl->dbval = F_DBPRIO | V_QID(qid);
2685 		if (is_t5(sc))
2686 			fl->dbval |= F_DBTYPE;
2687 
2688 		FL_LOCK(fl);
2689 		/* Enough to make sure the SGE doesn't think it's starved */
2690 		refill_fl(sc, fl, fl->lowat);
2691 		FL_UNLOCK(fl);
2692 	}
2693 
2694 	if (is_t5(sc) && cong >= 0) {
2695 		uint32_t param, val;
2696 
2697 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
2698 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
2699 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
2700 		if (cong == 0)
2701 			val = 1 << 19;
2702 		else {
2703 			val = 2 << 19;
2704 			for (i = 0; i < 4; i++) {
2705 				if (cong & (1 << i))
2706 					val |= 1 << (i << 2);
2707 			}
2708 		}
2709 
2710 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
2711 		if (rc != 0) {
2712 			/* report error but carry on */
2713 			device_printf(sc->dev,
2714 			    "failed to set congestion manager context for "
2715 			    "ingress queue %d: %d\n", iq->cntxt_id, rc);
2716 		}
2717 	}
2718 
2719 	/* Enable IQ interrupts */
2720 	atomic_store_rel_int(&iq->state, IQS_IDLE);
2721 	t4_write_reg(sc, MYPF_REG(A_SGE_PF_GTS), V_SEINTARM(iq->intr_params) |
2722 	    V_INGRESSQID(iq->cntxt_id));
2723 
2724 	return (0);
2725 }
2726 
2727 static int
2728 free_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl)
2729 {
2730 	int rc;
2731 	struct adapter *sc = iq->adapter;
2732 	device_t dev;
2733 
2734 	if (sc == NULL)
2735 		return (0);	/* nothing to do */
2736 
2737 	dev = vi ? vi->dev : sc->dev;
2738 
2739 	if (iq->flags & IQ_ALLOCATED) {
2740 		rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
2741 		    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
2742 		    fl ? fl->cntxt_id : 0xffff, 0xffff);
2743 		if (rc != 0) {
2744 			device_printf(dev,
2745 			    "failed to free queue %p: %d\n", iq, rc);
2746 			return (rc);
2747 		}
2748 		iq->flags &= ~IQ_ALLOCATED;
2749 	}
2750 
2751 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
2752 
2753 	bzero(iq, sizeof(*iq));
2754 
2755 	if (fl) {
2756 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
2757 		    fl->desc);
2758 
2759 		if (fl->sdesc)
2760 			free_fl_sdesc(sc, fl);
2761 
2762 		if (mtx_initialized(&fl->fl_lock))
2763 			mtx_destroy(&fl->fl_lock);
2764 
2765 		bzero(fl, sizeof(*fl));
2766 	}
2767 
2768 	return (0);
2769 }
2770 
2771 static void
2772 add_fl_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
2773     struct sge_fl *fl)
2774 {
2775 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2776 
2777 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
2778 	    "freelist");
2779 	children = SYSCTL_CHILDREN(oid);
2780 
2781 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
2782 	    CTLTYPE_INT | CTLFLAG_RD, &fl->cntxt_id, 0, sysctl_uint16, "I",
2783 	    "SGE context id of the freelist");
2784 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL,
2785 	    fl_pad ? 1 : 0, "padding enabled");
2786 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL,
2787 	    fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled");
2788 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
2789 	    0, "consumer index");
2790 	if (fl->flags & FL_BUF_PACKING) {
2791 		SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
2792 		    CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
2793 	}
2794 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
2795 	    0, "producer index");
2796 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_allocated",
2797 	    CTLFLAG_RD, &fl->mbuf_allocated, "# of mbuf allocated");
2798 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_inlined",
2799 	    CTLFLAG_RD, &fl->mbuf_inlined, "# of mbuf inlined in clusters");
2800 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
2801 	    CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
2802 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
2803 	    CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
2804 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
2805 	    CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
2806 }
2807 
2808 static int
2809 alloc_fwq(struct adapter *sc)
2810 {
2811 	int rc, intr_idx;
2812 	struct sge_iq *fwq = &sc->sge.fwq;
2813 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
2814 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2815 
2816 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE);
2817 	fwq->flags |= IQ_INTR;	/* always */
2818 	intr_idx = sc->intr_count > 1 ? 1 : 0;
2819 	rc = alloc_iq_fl(&sc->port[0]->vi[0], fwq, NULL, intr_idx, -1);
2820 	if (rc != 0) {
2821 		device_printf(sc->dev,
2822 		    "failed to create firmware event queue: %d\n", rc);
2823 		return (rc);
2824 	}
2825 
2826 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD,
2827 	    NULL, "firmware event queue");
2828 	children = SYSCTL_CHILDREN(oid);
2829 
2830 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "abs_id",
2831 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->abs_id, 0, sysctl_uint16, "I",
2832 	    "absolute id of the queue");
2833 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cntxt_id",
2834 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->cntxt_id, 0, sysctl_uint16, "I",
2835 	    "SGE context id of the queue");
2836 	SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cidx",
2837 	    CTLTYPE_INT | CTLFLAG_RD, &fwq->cidx, 0, sysctl_uint16, "I",
2838 	    "consumer index");
2839 
2840 	return (0);
2841 }
2842 
2843 static int
2844 free_fwq(struct adapter *sc)
2845 {
2846 	return free_iq_fl(NULL, &sc->sge.fwq, NULL);
2847 }
2848 
2849 static int
2850 alloc_mgmtq(struct adapter *sc)
2851 {
2852 	int rc;
2853 	struct sge_wrq *mgmtq = &sc->sge.mgmtq;
2854 	char name[16];
2855 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
2856 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
2857 
2858 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "mgmtq", CTLFLAG_RD,
2859 	    NULL, "management queue");
2860 
2861 	snprintf(name, sizeof(name), "%s mgmtq", device_get_nameunit(sc->dev));
2862 	init_eq(&mgmtq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[0]->tx_chan,
2863 	    sc->sge.fwq.cntxt_id, name);
2864 	rc = alloc_wrq(sc, NULL, mgmtq, oid);
2865 	if (rc != 0) {
2866 		device_printf(sc->dev,
2867 		    "failed to create management queue: %d\n", rc);
2868 		return (rc);
2869 	}
2870 
2871 	return (0);
2872 }
2873 
2874 static int
2875 free_mgmtq(struct adapter *sc)
2876 {
2877 
2878 	return free_wrq(sc, &sc->sge.mgmtq);
2879 }
2880 
2881 int
2882 tnl_cong(struct port_info *pi, int drop)
2883 {
2884 
2885 	if (drop == -1)
2886 		return (-1);
2887 	else if (drop == 1)
2888 		return (0);
2889 	else
2890 		return (pi->rx_chan_map);
2891 }
2892 
2893 static int
2894 alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int intr_idx, int idx,
2895     struct sysctl_oid *oid)
2896 {
2897 	int rc;
2898 	struct sysctl_oid_list *children;
2899 	char name[16];
2900 
2901 	rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, intr_idx,
2902 	    tnl_cong(vi->pi, cong_drop));
2903 	if (rc != 0)
2904 		return (rc);
2905 
2906 	/*
2907 	 * The freelist is just barely above the starvation threshold right now,
2908 	 * fill it up a bit more.
2909 	 */
2910 	FL_LOCK(&rxq->fl);
2911 	refill_fl(vi->pi->adapter, &rxq->fl, 128);
2912 	FL_UNLOCK(&rxq->fl);
2913 
2914 #if defined(INET) || defined(INET6)
2915 	rc = tcp_lro_init(&rxq->lro);
2916 	if (rc != 0)
2917 		return (rc);
2918 	rxq->lro.ifp = vi->ifp; /* also indicates LRO init'ed */
2919 
2920 	if (vi->ifp->if_capenable & IFCAP_LRO)
2921 		rxq->iq.flags |= IQ_LRO_ENABLED;
2922 #endif
2923 	rxq->ifp = vi->ifp;
2924 
2925 	children = SYSCTL_CHILDREN(oid);
2926 
2927 	snprintf(name, sizeof(name), "%d", idx);
2928 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
2929 	    NULL, "rx queue");
2930 	children = SYSCTL_CHILDREN(oid);
2931 
2932 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "abs_id",
2933 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.abs_id, 0, sysctl_uint16, "I",
2934 	    "absolute id of the queue");
2935 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cntxt_id",
2936 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cntxt_id, 0, sysctl_uint16, "I",
2937 	    "SGE context id of the queue");
2938 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
2939 	    CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cidx, 0, sysctl_uint16, "I",
2940 	    "consumer index");
2941 #if defined(INET) || defined(INET6)
2942 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
2943 	    &rxq->lro.lro_queued, 0, NULL);
2944 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
2945 	    &rxq->lro.lro_flushed, 0, NULL);
2946 #endif
2947 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
2948 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
2949 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_extraction",
2950 	    CTLFLAG_RD, &rxq->vlan_extraction,
2951 	    "# of times hardware extracted 802.1Q tag");
2952 
2953 	add_fl_sysctls(&vi->ctx, oid, &rxq->fl);
2954 
2955 	return (rc);
2956 }
2957 
2958 static int
2959 free_rxq(struct vi_info *vi, struct sge_rxq *rxq)
2960 {
2961 	int rc;
2962 
2963 #if defined(INET) || defined(INET6)
2964 	if (rxq->lro.ifp) {
2965 		tcp_lro_free(&rxq->lro);
2966 		rxq->lro.ifp = NULL;
2967 	}
2968 #endif
2969 
2970 	rc = free_iq_fl(vi, &rxq->iq, &rxq->fl);
2971 	if (rc == 0)
2972 		bzero(rxq, sizeof(*rxq));
2973 
2974 	return (rc);
2975 }
2976 
2977 #ifdef TCP_OFFLOAD
2978 static int
2979 alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq,
2980     int intr_idx, int idx, struct sysctl_oid *oid)
2981 {
2982 	int rc;
2983 	struct sysctl_oid_list *children;
2984 	char name[16];
2985 
2986 	rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx,
2987 	    vi->pi->rx_chan_map);
2988 	if (rc != 0)
2989 		return (rc);
2990 
2991 	children = SYSCTL_CHILDREN(oid);
2992 
2993 	snprintf(name, sizeof(name), "%d", idx);
2994 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
2995 	    NULL, "rx queue");
2996 	children = SYSCTL_CHILDREN(oid);
2997 
2998 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "abs_id",
2999 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.abs_id, 0, sysctl_uint16,
3000 	    "I", "absolute id of the queue");
3001 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cntxt_id",
3002 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cntxt_id, 0, sysctl_uint16,
3003 	    "I", "SGE context id of the queue");
3004 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3005 	    CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cidx, 0, sysctl_uint16, "I",
3006 	    "consumer index");
3007 
3008 	add_fl_sysctls(&vi->ctx, oid, &ofld_rxq->fl);
3009 
3010 	return (rc);
3011 }
3012 
3013 static int
3014 free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq)
3015 {
3016 	int rc;
3017 
3018 	rc = free_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl);
3019 	if (rc == 0)
3020 		bzero(ofld_rxq, sizeof(*ofld_rxq));
3021 
3022 	return (rc);
3023 }
3024 #endif
3025 
3026 #ifdef DEV_NETMAP
3027 static int
3028 alloc_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int intr_idx,
3029     int idx, struct sysctl_oid *oid)
3030 {
3031 	int rc;
3032 	struct sysctl_oid_list *children;
3033 	struct sysctl_ctx_list *ctx;
3034 	char name[16];
3035 	size_t len;
3036 	struct adapter *sc = vi->pi->adapter;
3037 	struct netmap_adapter *na = NA(vi->ifp);
3038 
3039 	MPASS(na != NULL);
3040 
3041 	len = vi->qsize_rxq * IQ_ESIZE;
3042 	rc = alloc_ring(sc, len, &nm_rxq->iq_desc_tag, &nm_rxq->iq_desc_map,
3043 	    &nm_rxq->iq_ba, (void **)&nm_rxq->iq_desc);
3044 	if (rc != 0)
3045 		return (rc);
3046 
3047 	len = na->num_rx_desc * EQ_ESIZE + spg_len;
3048 	rc = alloc_ring(sc, len, &nm_rxq->fl_desc_tag, &nm_rxq->fl_desc_map,
3049 	    &nm_rxq->fl_ba, (void **)&nm_rxq->fl_desc);
3050 	if (rc != 0)
3051 		return (rc);
3052 
3053 	nm_rxq->vi = vi;
3054 	nm_rxq->nid = idx;
3055 	nm_rxq->iq_cidx = 0;
3056 	nm_rxq->iq_sidx = vi->qsize_rxq - spg_len / IQ_ESIZE;
3057 	nm_rxq->iq_gen = F_RSPD_GEN;
3058 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
3059 	nm_rxq->fl_sidx = na->num_rx_desc;
3060 	nm_rxq->intr_idx = intr_idx;
3061 
3062 	ctx = &vi->ctx;
3063 	children = SYSCTL_CHILDREN(oid);
3064 
3065 	snprintf(name, sizeof(name), "%d", idx);
3066 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL,
3067 	    "rx queue");
3068 	children = SYSCTL_CHILDREN(oid);
3069 
3070 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3071 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_abs_id, 0, sysctl_uint16,
3072 	    "I", "absolute id of the queue");
3073 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3074 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cntxt_id, 0, sysctl_uint16,
3075 	    "I", "SGE context id of the queue");
3076 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3077 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cidx, 0, sysctl_uint16, "I",
3078 	    "consumer index");
3079 
3080 	children = SYSCTL_CHILDREN(oid);
3081 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3082 	    "freelist");
3083 	children = SYSCTL_CHILDREN(oid);
3084 
3085 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3086 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->fl_cntxt_id, 0, sysctl_uint16,
3087 	    "I", "SGE context id of the freelist");
3088 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
3089 	    &nm_rxq->fl_cidx, 0, "consumer index");
3090 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
3091 	    &nm_rxq->fl_pidx, 0, "producer index");
3092 
3093 	return (rc);
3094 }
3095 
3096 
3097 static int
3098 free_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
3099 {
3100 	struct adapter *sc = vi->pi->adapter;
3101 
3102 	free_ring(sc, nm_rxq->iq_desc_tag, nm_rxq->iq_desc_map, nm_rxq->iq_ba,
3103 	    nm_rxq->iq_desc);
3104 	free_ring(sc, nm_rxq->fl_desc_tag, nm_rxq->fl_desc_map, nm_rxq->fl_ba,
3105 	    nm_rxq->fl_desc);
3106 
3107 	return (0);
3108 }
3109 
3110 static int
3111 alloc_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq, int iqidx, int idx,
3112     struct sysctl_oid *oid)
3113 {
3114 	int rc;
3115 	size_t len;
3116 	struct port_info *pi = vi->pi;
3117 	struct adapter *sc = pi->adapter;
3118 	struct netmap_adapter *na = NA(vi->ifp);
3119 	char name[16];
3120 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3121 
3122 	len = na->num_tx_desc * EQ_ESIZE + spg_len;
3123 	rc = alloc_ring(sc, len, &nm_txq->desc_tag, &nm_txq->desc_map,
3124 	    &nm_txq->ba, (void **)&nm_txq->desc);
3125 	if (rc)
3126 		return (rc);
3127 
3128 	nm_txq->pidx = nm_txq->cidx = 0;
3129 	nm_txq->sidx = na->num_tx_desc;
3130 	nm_txq->nid = idx;
3131 	nm_txq->iqidx = iqidx;
3132 	nm_txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3133 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_VF_VLD(1) |
3134 	    V_TXPKT_VF(vi->viid));
3135 
3136 	snprintf(name, sizeof(name), "%d", idx);
3137 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3138 	    NULL, "netmap tx queue");
3139 	children = SYSCTL_CHILDREN(oid);
3140 
3141 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3142 	    &nm_txq->cntxt_id, 0, "SGE context id of the queue");
3143 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3144 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->cidx, 0, sysctl_uint16, "I",
3145 	    "consumer index");
3146 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3147 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->pidx, 0, sysctl_uint16, "I",
3148 	    "producer index");
3149 
3150 	return (rc);
3151 }
3152 
3153 static int
3154 free_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
3155 {
3156 	struct adapter *sc = vi->pi->adapter;
3157 
3158 	free_ring(sc, nm_txq->desc_tag, nm_txq->desc_map, nm_txq->ba,
3159 	    nm_txq->desc);
3160 
3161 	return (0);
3162 }
3163 #endif
3164 
3165 static int
3166 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
3167 {
3168 	int rc, cntxt_id;
3169 	struct fw_eq_ctrl_cmd c;
3170 	int qsize = eq->sidx + spg_len / EQ_ESIZE;
3171 
3172 	bzero(&c, sizeof(c));
3173 
3174 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
3175 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
3176 	    V_FW_EQ_CTRL_CMD_VFN(0));
3177 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
3178 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
3179 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid));
3180 	c.physeqid_pkd = htobe32(0);
3181 	c.fetchszm_to_iqid =
3182 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3183 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
3184 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
3185 	c.dcaen_to_eqsize =
3186 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3187 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3188 		V_FW_EQ_CTRL_CMD_EQSIZE(qsize));
3189 	c.eqaddr = htobe64(eq->ba);
3190 
3191 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3192 	if (rc != 0) {
3193 		device_printf(sc->dev,
3194 		    "failed to create control queue %d: %d\n", eq->tx_chan, rc);
3195 		return (rc);
3196 	}
3197 	eq->flags |= EQ_ALLOCATED;
3198 
3199 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
3200 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3201 	if (cntxt_id >= sc->sge.neq)
3202 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3203 		cntxt_id, sc->sge.neq - 1);
3204 	sc->sge.eqmap[cntxt_id] = eq;
3205 
3206 	return (rc);
3207 }
3208 
3209 static int
3210 eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3211 {
3212 	int rc, cntxt_id;
3213 	struct fw_eq_eth_cmd c;
3214 	int qsize = eq->sidx + spg_len / EQ_ESIZE;
3215 
3216 	bzero(&c, sizeof(c));
3217 
3218 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
3219 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
3220 	    V_FW_EQ_ETH_CMD_VFN(0));
3221 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
3222 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
3223 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
3224 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
3225 	c.fetchszm_to_iqid =
3226 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3227 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
3228 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
3229 	c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3230 	    V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3231 	    V_FW_EQ_ETH_CMD_EQSIZE(qsize));
3232 	c.eqaddr = htobe64(eq->ba);
3233 
3234 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3235 	if (rc != 0) {
3236 		device_printf(vi->dev,
3237 		    "failed to create Ethernet egress queue: %d\n", rc);
3238 		return (rc);
3239 	}
3240 	eq->flags |= EQ_ALLOCATED;
3241 
3242 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
3243 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3244 	if (cntxt_id >= sc->sge.neq)
3245 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3246 		cntxt_id, sc->sge.neq - 1);
3247 	sc->sge.eqmap[cntxt_id] = eq;
3248 
3249 	return (rc);
3250 }
3251 
3252 #ifdef TCP_OFFLOAD
3253 static int
3254 ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3255 {
3256 	int rc, cntxt_id;
3257 	struct fw_eq_ofld_cmd c;
3258 	int qsize = eq->sidx + spg_len / EQ_ESIZE;
3259 
3260 	bzero(&c, sizeof(c));
3261 
3262 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
3263 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
3264 	    V_FW_EQ_OFLD_CMD_VFN(0));
3265 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
3266 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
3267 	c.fetchszm_to_iqid =
3268 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3269 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
3270 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
3271 	c.dcaen_to_eqsize =
3272 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) |
3273 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3274 		V_FW_EQ_OFLD_CMD_EQSIZE(qsize));
3275 	c.eqaddr = htobe64(eq->ba);
3276 
3277 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3278 	if (rc != 0) {
3279 		device_printf(vi->dev,
3280 		    "failed to create egress queue for TCP offload: %d\n", rc);
3281 		return (rc);
3282 	}
3283 	eq->flags |= EQ_ALLOCATED;
3284 
3285 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
3286 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3287 	if (cntxt_id >= sc->sge.neq)
3288 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3289 		cntxt_id, sc->sge.neq - 1);
3290 	sc->sge.eqmap[cntxt_id] = eq;
3291 
3292 	return (rc);
3293 }
3294 #endif
3295 
3296 static int
3297 alloc_eq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3298 {
3299 	int rc, qsize;
3300 	size_t len;
3301 
3302 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
3303 
3304 	qsize = eq->sidx + spg_len / EQ_ESIZE;
3305 	len = qsize * EQ_ESIZE;
3306 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
3307 	    &eq->ba, (void **)&eq->desc);
3308 	if (rc)
3309 		return (rc);
3310 
3311 	eq->pidx = eq->cidx = 0;
3312 	eq->equeqidx = eq->dbidx = 0;
3313 	eq->doorbells = sc->doorbells;
3314 
3315 	switch (eq->flags & EQ_TYPEMASK) {
3316 	case EQ_CTRL:
3317 		rc = ctrl_eq_alloc(sc, eq);
3318 		break;
3319 
3320 	case EQ_ETH:
3321 		rc = eth_eq_alloc(sc, vi, eq);
3322 		break;
3323 
3324 #ifdef TCP_OFFLOAD
3325 	case EQ_OFLD:
3326 		rc = ofld_eq_alloc(sc, vi, eq);
3327 		break;
3328 #endif
3329 
3330 	default:
3331 		panic("%s: invalid eq type %d.", __func__,
3332 		    eq->flags & EQ_TYPEMASK);
3333 	}
3334 	if (rc != 0) {
3335 		device_printf(sc->dev,
3336 		    "failed to allocate egress queue(%d): %d\n",
3337 		    eq->flags & EQ_TYPEMASK, rc);
3338 	}
3339 
3340 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
3341 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
3342 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
3343 		uint32_t s_qpp = sc->sge.eq_s_qpp;
3344 		uint32_t mask = (1 << s_qpp) - 1;
3345 		volatile uint8_t *udb;
3346 
3347 		udb = sc->udbs_base + UDBS_DB_OFFSET;
3348 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
3349 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
3350 		if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
3351 	    		clrbit(&eq->doorbells, DOORBELL_WCWR);
3352 		else {
3353 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
3354 			eq->udb_qid = 0;
3355 		}
3356 		eq->udb = (volatile void *)udb;
3357 	}
3358 
3359 	return (rc);
3360 }
3361 
3362 static int
3363 free_eq(struct adapter *sc, struct sge_eq *eq)
3364 {
3365 	int rc;
3366 
3367 	if (eq->flags & EQ_ALLOCATED) {
3368 		switch (eq->flags & EQ_TYPEMASK) {
3369 		case EQ_CTRL:
3370 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
3371 			    eq->cntxt_id);
3372 			break;
3373 
3374 		case EQ_ETH:
3375 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
3376 			    eq->cntxt_id);
3377 			break;
3378 
3379 #ifdef TCP_OFFLOAD
3380 		case EQ_OFLD:
3381 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
3382 			    eq->cntxt_id);
3383 			break;
3384 #endif
3385 
3386 		default:
3387 			panic("%s: invalid eq type %d.", __func__,
3388 			    eq->flags & EQ_TYPEMASK);
3389 		}
3390 		if (rc != 0) {
3391 			device_printf(sc->dev,
3392 			    "failed to free egress queue (%d): %d\n",
3393 			    eq->flags & EQ_TYPEMASK, rc);
3394 			return (rc);
3395 		}
3396 		eq->flags &= ~EQ_ALLOCATED;
3397 	}
3398 
3399 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
3400 
3401 	if (mtx_initialized(&eq->eq_lock))
3402 		mtx_destroy(&eq->eq_lock);
3403 
3404 	bzero(eq, sizeof(*eq));
3405 	return (0);
3406 }
3407 
3408 static int
3409 alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq,
3410     struct sysctl_oid *oid)
3411 {
3412 	int rc;
3413 	struct sysctl_ctx_list *ctx = vi ? &vi->ctx : &sc->ctx;
3414 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3415 
3416 	rc = alloc_eq(sc, vi, &wrq->eq);
3417 	if (rc)
3418 		return (rc);
3419 
3420 	wrq->adapter = sc;
3421 	TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq);
3422 	TAILQ_INIT(&wrq->incomplete_wrs);
3423 	STAILQ_INIT(&wrq->wr_list);
3424 	wrq->nwr_pending = 0;
3425 	wrq->ndesc_needed = 0;
3426 
3427 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3428 	    &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
3429 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3430 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I",
3431 	    "consumer index");
3432 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx",
3433 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I",
3434 	    "producer index");
3435 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD,
3436 	    &wrq->tx_wrs_direct, "# of work requests (direct)");
3437 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD,
3438 	    &wrq->tx_wrs_copied, "# of work requests (copied)");
3439 
3440 	return (rc);
3441 }
3442 
3443 static int
3444 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
3445 {
3446 	int rc;
3447 
3448 	rc = free_eq(sc, &wrq->eq);
3449 	if (rc)
3450 		return (rc);
3451 
3452 	bzero(wrq, sizeof(*wrq));
3453 	return (0);
3454 }
3455 
3456 static int
3457 alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx,
3458     struct sysctl_oid *oid)
3459 {
3460 	int rc;
3461 	struct port_info *pi = vi->pi;
3462 	struct adapter *sc = pi->adapter;
3463 	struct sge_eq *eq = &txq->eq;
3464 	char name[16];
3465 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3466 
3467 	rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx, can_resume_eth_tx,
3468 	    M_CXGBE, M_WAITOK);
3469 	if (rc != 0) {
3470 		device_printf(sc->dev, "failed to allocate mp_ring: %d\n", rc);
3471 		return (rc);
3472 	}
3473 
3474 	rc = alloc_eq(sc, vi, eq);
3475 	if (rc != 0) {
3476 		mp_ring_free(txq->r);
3477 		txq->r = NULL;
3478 		return (rc);
3479 	}
3480 
3481 	/* Can't fail after this point. */
3482 
3483 	TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq);
3484 	txq->ifp = vi->ifp;
3485 	txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK);
3486 	txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3487 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_VF_VLD(1) |
3488 	    V_TXPKT_VF(vi->viid));
3489 	txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE,
3490 	    M_ZERO | M_WAITOK);
3491 
3492 	snprintf(name, sizeof(name), "%d", idx);
3493 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3494 	    NULL, "tx queue");
3495 	children = SYSCTL_CHILDREN(oid);
3496 
3497 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3498 	    &eq->cntxt_id, 0, "SGE context id of the queue");
3499 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3500 	    CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I",
3501 	    "consumer index");
3502 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3503 	    CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I",
3504 	    "producer index");
3505 
3506 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
3507 	    &txq->txcsum, "# of times hardware assisted with checksum");
3508 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_insertion",
3509 	    CTLFLAG_RD, &txq->vlan_insertion,
3510 	    "# of times hardware inserted 802.1Q tag");
3511 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
3512 	    &txq->tso_wrs, "# of TSO work requests");
3513 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
3514 	    &txq->imm_wrs, "# of work requests with immediate data");
3515 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
3516 	    &txq->sgl_wrs, "# of work requests with direct SGL");
3517 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
3518 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
3519 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_wrs",
3520 	    CTLFLAG_RD, &txq->txpkts0_wrs,
3521 	    "# of txpkts (type 0) work requests");
3522 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_wrs",
3523 	    CTLFLAG_RD, &txq->txpkts1_wrs,
3524 	    "# of txpkts (type 1) work requests");
3525 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_pkts",
3526 	    CTLFLAG_RD, &txq->txpkts0_pkts,
3527 	    "# of frames tx'd using type0 txpkts work requests");
3528 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_pkts",
3529 	    CTLFLAG_RD, &txq->txpkts1_pkts,
3530 	    "# of frames tx'd using type1 txpkts work requests");
3531 
3532 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_enqueues",
3533 	    CTLFLAG_RD, &txq->r->enqueues,
3534 	    "# of enqueues to the mp_ring for this queue");
3535 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_drops",
3536 	    CTLFLAG_RD, &txq->r->drops,
3537 	    "# of drops in the mp_ring for this queue");
3538 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_starts",
3539 	    CTLFLAG_RD, &txq->r->starts,
3540 	    "# of normal consumer starts in the mp_ring for this queue");
3541 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_stalls",
3542 	    CTLFLAG_RD, &txq->r->stalls,
3543 	    "# of consumer stalls in the mp_ring for this queue");
3544 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_restarts",
3545 	    CTLFLAG_RD, &txq->r->restarts,
3546 	    "# of consumer restarts in the mp_ring for this queue");
3547 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_abdications",
3548 	    CTLFLAG_RD, &txq->r->abdications,
3549 	    "# of consumer abdications in the mp_ring for this queue");
3550 
3551 	return (0);
3552 }
3553 
3554 static int
3555 free_txq(struct vi_info *vi, struct sge_txq *txq)
3556 {
3557 	int rc;
3558 	struct adapter *sc = vi->pi->adapter;
3559 	struct sge_eq *eq = &txq->eq;
3560 
3561 	rc = free_eq(sc, eq);
3562 	if (rc)
3563 		return (rc);
3564 
3565 	sglist_free(txq->gl);
3566 	free(txq->sdesc, M_CXGBE);
3567 	mp_ring_free(txq->r);
3568 
3569 	bzero(txq, sizeof(*txq));
3570 	return (0);
3571 }
3572 
3573 static void
3574 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
3575 {
3576 	bus_addr_t *ba = arg;
3577 
3578 	KASSERT(nseg == 1,
3579 	    ("%s meant for single segment mappings only.", __func__));
3580 
3581 	*ba = error ? 0 : segs->ds_addr;
3582 }
3583 
3584 static inline void
3585 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
3586 {
3587 	uint32_t n, v;
3588 
3589 	n = IDXDIFF(fl->pidx / 8, fl->dbidx, fl->sidx);
3590 	MPASS(n > 0);
3591 
3592 	wmb();
3593 	v = fl->dbval | V_PIDX(n);
3594 	if (fl->udb)
3595 		*fl->udb = htole32(v);
3596 	else
3597 		t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL), v);
3598 	IDXINCR(fl->dbidx, n, fl->sidx);
3599 }
3600 
3601 /*
3602  * Fills up the freelist by allocating upto 'n' buffers.  Buffers that are
3603  * recycled do not count towards this allocation budget.
3604  *
3605  * Returns non-zero to indicate that this freelist should be added to the list
3606  * of starving freelists.
3607  */
3608 static int
3609 refill_fl(struct adapter *sc, struct sge_fl *fl, int n)
3610 {
3611 	__be64 *d;
3612 	struct fl_sdesc *sd;
3613 	uintptr_t pa;
3614 	caddr_t cl;
3615 	struct cluster_layout *cll;
3616 	struct sw_zone_info *swz;
3617 	struct cluster_metadata *clm;
3618 	uint16_t max_pidx;
3619 	uint16_t hw_cidx = fl->hw_cidx;		/* stable snapshot */
3620 
3621 	FL_LOCK_ASSERT_OWNED(fl);
3622 
3623 	/*
3624 	 * We always stop at the begining of the hardware descriptor that's just
3625 	 * before the one with the hw cidx.  This is to avoid hw pidx = hw cidx,
3626 	 * which would mean an empty freelist to the chip.
3627 	 */
3628 	max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1;
3629 	if (fl->pidx == max_pidx * 8)
3630 		return (0);
3631 
3632 	d = &fl->desc[fl->pidx];
3633 	sd = &fl->sdesc[fl->pidx];
3634 	cll = &fl->cll_def;	/* default layout */
3635 	swz = &sc->sge.sw_zone_info[cll->zidx];
3636 
3637 	while (n > 0) {
3638 
3639 		if (sd->cl != NULL) {
3640 
3641 			if (sd->nmbuf == 0) {
3642 				/*
3643 				 * Fast recycle without involving any atomics on
3644 				 * the cluster's metadata (if the cluster has
3645 				 * metadata).  This happens when all frames
3646 				 * received in the cluster were small enough to
3647 				 * fit within a single mbuf each.
3648 				 */
3649 				fl->cl_fast_recycled++;
3650 #ifdef INVARIANTS
3651 				clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
3652 				if (clm != NULL)
3653 					MPASS(clm->refcount == 1);
3654 #endif
3655 				goto recycled_fast;
3656 			}
3657 
3658 			/*
3659 			 * Cluster is guaranteed to have metadata.  Clusters
3660 			 * without metadata always take the fast recycle path
3661 			 * when they're recycled.
3662 			 */
3663 			clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
3664 			MPASS(clm != NULL);
3665 
3666 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
3667 				fl->cl_recycled++;
3668 				counter_u64_add(extfree_rels, 1);
3669 				goto recycled;
3670 			}
3671 			sd->cl = NULL;	/* gave up my reference */
3672 		}
3673 		MPASS(sd->cl == NULL);
3674 alloc:
3675 		cl = uma_zalloc(swz->zone, M_NOWAIT);
3676 		if (__predict_false(cl == NULL)) {
3677 			if (cll == &fl->cll_alt || fl->cll_alt.zidx == -1 ||
3678 			    fl->cll_def.zidx == fl->cll_alt.zidx)
3679 				break;
3680 
3681 			/* fall back to the safe zone */
3682 			cll = &fl->cll_alt;
3683 			swz = &sc->sge.sw_zone_info[cll->zidx];
3684 			goto alloc;
3685 		}
3686 		fl->cl_allocated++;
3687 		n--;
3688 
3689 		pa = pmap_kextract((vm_offset_t)cl);
3690 		pa += cll->region1;
3691 		sd->cl = cl;
3692 		sd->cll = *cll;
3693 		*d = htobe64(pa | cll->hwidx);
3694 		clm = cl_metadata(sc, fl, cll, cl);
3695 		if (clm != NULL) {
3696 recycled:
3697 #ifdef INVARIANTS
3698 			clm->sd = sd;
3699 #endif
3700 			clm->refcount = 1;
3701 		}
3702 		sd->nmbuf = 0;
3703 recycled_fast:
3704 		d++;
3705 		sd++;
3706 		if (__predict_false(++fl->pidx % 8 == 0)) {
3707 			uint16_t pidx = fl->pidx / 8;
3708 
3709 			if (__predict_false(pidx == fl->sidx)) {
3710 				fl->pidx = 0;
3711 				pidx = 0;
3712 				sd = fl->sdesc;
3713 				d = fl->desc;
3714 			}
3715 			if (pidx == max_pidx)
3716 				break;
3717 
3718 			if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4)
3719 				ring_fl_db(sc, fl);
3720 		}
3721 	}
3722 
3723 	if (fl->pidx / 8 != fl->dbidx)
3724 		ring_fl_db(sc, fl);
3725 
3726 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
3727 }
3728 
3729 /*
3730  * Attempt to refill all starving freelists.
3731  */
3732 static void
3733 refill_sfl(void *arg)
3734 {
3735 	struct adapter *sc = arg;
3736 	struct sge_fl *fl, *fl_temp;
3737 
3738 	mtx_assert(&sc->sfl_lock, MA_OWNED);
3739 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
3740 		FL_LOCK(fl);
3741 		refill_fl(sc, fl, 64);
3742 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
3743 			TAILQ_REMOVE(&sc->sfl, fl, link);
3744 			fl->flags &= ~FL_STARVING;
3745 		}
3746 		FL_UNLOCK(fl);
3747 	}
3748 
3749 	if (!TAILQ_EMPTY(&sc->sfl))
3750 		callout_schedule(&sc->sfl_callout, hz / 5);
3751 }
3752 
3753 static int
3754 alloc_fl_sdesc(struct sge_fl *fl)
3755 {
3756 
3757 	fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc), M_CXGBE,
3758 	    M_ZERO | M_WAITOK);
3759 
3760 	return (0);
3761 }
3762 
3763 static void
3764 free_fl_sdesc(struct adapter *sc, struct sge_fl *fl)
3765 {
3766 	struct fl_sdesc *sd;
3767 	struct cluster_metadata *clm;
3768 	struct cluster_layout *cll;
3769 	int i;
3770 
3771 	sd = fl->sdesc;
3772 	for (i = 0; i < fl->sidx * 8; i++, sd++) {
3773 		if (sd->cl == NULL)
3774 			continue;
3775 
3776 		cll = &sd->cll;
3777 		clm = cl_metadata(sc, fl, cll, sd->cl);
3778 		if (sd->nmbuf == 0)
3779 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
3780 		else if (clm && atomic_fetchadd_int(&clm->refcount, -1) == 1) {
3781 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
3782 			counter_u64_add(extfree_rels, 1);
3783 		}
3784 		sd->cl = NULL;
3785 	}
3786 
3787 	free(fl->sdesc, M_CXGBE);
3788 	fl->sdesc = NULL;
3789 }
3790 
3791 static inline void
3792 get_pkt_gl(struct mbuf *m, struct sglist *gl)
3793 {
3794 	int rc;
3795 
3796 	M_ASSERTPKTHDR(m);
3797 
3798 	sglist_reset(gl);
3799 	rc = sglist_append_mbuf(gl, m);
3800 	if (__predict_false(rc != 0)) {
3801 		panic("%s: mbuf %p (%d segs) was vetted earlier but now fails "
3802 		    "with %d.", __func__, m, mbuf_nsegs(m), rc);
3803 	}
3804 
3805 	KASSERT(gl->sg_nseg == mbuf_nsegs(m),
3806 	    ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m,
3807 	    mbuf_nsegs(m), gl->sg_nseg));
3808 	KASSERT(gl->sg_nseg > 0 &&
3809 	    gl->sg_nseg <= (needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS),
3810 	    ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__,
3811 		gl->sg_nseg, needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS));
3812 }
3813 
3814 /*
3815  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
3816  */
3817 static inline u_int
3818 txpkt_len16(u_int nsegs, u_int tso)
3819 {
3820 	u_int n;
3821 
3822 	MPASS(nsegs > 0);
3823 
3824 	nsegs--; /* first segment is part of ulptx_sgl */
3825 	n = sizeof(struct fw_eth_tx_pkt_wr) + sizeof(struct cpl_tx_pkt_core) +
3826 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
3827 	if (tso)
3828 		n += sizeof(struct cpl_tx_pkt_lso_core);
3829 
3830 	return (howmany(n, 16));
3831 }
3832 
3833 /*
3834  * len16 for a txpkts type 0 WR with a GL.  Does not include the firmware work
3835  * request header.
3836  */
3837 static inline u_int
3838 txpkts0_len16(u_int nsegs)
3839 {
3840 	u_int n;
3841 
3842 	MPASS(nsegs > 0);
3843 
3844 	nsegs--; /* first segment is part of ulptx_sgl */
3845 	n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) +
3846 	    sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) +
3847 	    8 * ((3 * nsegs) / 2 + (nsegs & 1));
3848 
3849 	return (howmany(n, 16));
3850 }
3851 
3852 /*
3853  * len16 for a txpkts type 1 WR with a GL.  Does not include the firmware work
3854  * request header.
3855  */
3856 static inline u_int
3857 txpkts1_len16(void)
3858 {
3859 	u_int n;
3860 
3861 	n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl);
3862 
3863 	return (howmany(n, 16));
3864 }
3865 
3866 static inline u_int
3867 imm_payload(u_int ndesc)
3868 {
3869 	u_int n;
3870 
3871 	n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) -
3872 	    sizeof(struct cpl_tx_pkt_core);
3873 
3874 	return (n);
3875 }
3876 
3877 /*
3878  * Write a txpkt WR for this packet to the hardware descriptors, update the
3879  * software descriptor, and advance the pidx.  It is guaranteed that enough
3880  * descriptors are available.
3881  *
3882  * The return value is the # of hardware descriptors used.
3883  */
3884 static u_int
3885 write_txpkt_wr(struct sge_txq *txq, struct fw_eth_tx_pkt_wr *wr,
3886     struct mbuf *m0, u_int available)
3887 {
3888 	struct sge_eq *eq = &txq->eq;
3889 	struct tx_sdesc *txsd;
3890 	struct cpl_tx_pkt_core *cpl;
3891 	uint32_t ctrl;	/* used in many unrelated places */
3892 	uint64_t ctrl1;
3893 	int len16, ndesc, pktlen, nsegs;
3894 	caddr_t dst;
3895 
3896 	TXQ_LOCK_ASSERT_OWNED(txq);
3897 	M_ASSERTPKTHDR(m0);
3898 	MPASS(available > 0 && available < eq->sidx);
3899 
3900 	len16 = mbuf_len16(m0);
3901 	nsegs = mbuf_nsegs(m0);
3902 	pktlen = m0->m_pkthdr.len;
3903 	ctrl = sizeof(struct cpl_tx_pkt_core);
3904 	if (needs_tso(m0))
3905 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
3906 	else if (pktlen <= imm_payload(2) && available >= 2) {
3907 		/* Immediate data.  Recalculate len16 and set nsegs to 0. */
3908 		ctrl += pktlen;
3909 		len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) +
3910 		    sizeof(struct cpl_tx_pkt_core) + pktlen, 16);
3911 		nsegs = 0;
3912 	}
3913 	ndesc = howmany(len16, EQ_ESIZE / 16);
3914 	MPASS(ndesc <= available);
3915 
3916 	/* Firmware work request header */
3917 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
3918 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
3919 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
3920 
3921 	ctrl = V_FW_WR_LEN16(len16);
3922 	wr->equiq_to_len16 = htobe32(ctrl);
3923 	wr->r3 = 0;
3924 
3925 	if (needs_tso(m0)) {
3926 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
3927 
3928 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
3929 		    m0->m_pkthdr.l4hlen > 0,
3930 		    ("%s: mbuf %p needs TSO but missing header lengths",
3931 			__func__, m0));
3932 
3933 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
3934 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
3935 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
3936 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
3937 			ctrl |= V_LSO_ETHHDR_LEN(1);
3938 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
3939 			ctrl |= F_LSO_IPV6;
3940 
3941 		lso->lso_ctrl = htobe32(ctrl);
3942 		lso->ipid_ofst = htobe16(0);
3943 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
3944 		lso->seqno_offset = htobe32(0);
3945 		lso->len = htobe32(pktlen);
3946 
3947 		cpl = (void *)(lso + 1);
3948 
3949 		txq->tso_wrs++;
3950 	} else
3951 		cpl = (void *)(wr + 1);
3952 
3953 	/* Checksum offload */
3954 	ctrl1 = 0;
3955 	if (needs_l3_csum(m0) == 0)
3956 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
3957 	if (needs_l4_csum(m0) == 0)
3958 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
3959 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
3960 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
3961 		txq->txcsum++;	/* some hardware assistance provided */
3962 
3963 	/* VLAN tag insertion */
3964 	if (needs_vlan_insertion(m0)) {
3965 		ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
3966 		txq->vlan_insertion++;
3967 	}
3968 
3969 	/* CPL header */
3970 	cpl->ctrl0 = txq->cpl_ctrl0;
3971 	cpl->pack = 0;
3972 	cpl->len = htobe16(pktlen);
3973 	cpl->ctrl1 = htobe64(ctrl1);
3974 
3975 	/* SGL */
3976 	dst = (void *)(cpl + 1);
3977 	if (nsegs > 0) {
3978 
3979 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
3980 		txq->sgl_wrs++;
3981 	} else {
3982 		struct mbuf *m;
3983 
3984 		for (m = m0; m != NULL; m = m->m_next) {
3985 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
3986 #ifdef INVARIANTS
3987 			pktlen -= m->m_len;
3988 #endif
3989 		}
3990 #ifdef INVARIANTS
3991 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
3992 #endif
3993 		txq->imm_wrs++;
3994 	}
3995 
3996 	txq->txpkt_wrs++;
3997 
3998 	txsd = &txq->sdesc[eq->pidx];
3999 	txsd->m = m0;
4000 	txsd->desc_used = ndesc;
4001 
4002 	return (ndesc);
4003 }
4004 
4005 static int
4006 try_txpkts(struct mbuf *m, struct mbuf *n, struct txpkts *txp, u_int available)
4007 {
4008 	u_int needed, nsegs1, nsegs2, l1, l2;
4009 
4010 	if (cannot_use_txpkts(m) || cannot_use_txpkts(n))
4011 		return (1);
4012 
4013 	nsegs1 = mbuf_nsegs(m);
4014 	nsegs2 = mbuf_nsegs(n);
4015 	if (nsegs1 + nsegs2 == 2) {
4016 		txp->wr_type = 1;
4017 		l1 = l2 = txpkts1_len16();
4018 	} else {
4019 		txp->wr_type = 0;
4020 		l1 = txpkts0_len16(nsegs1);
4021 		l2 = txpkts0_len16(nsegs2);
4022 	}
4023 	txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) + l1 + l2;
4024 	needed = howmany(txp->len16, EQ_ESIZE / 16);
4025 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4026 		return (1);
4027 
4028 	txp->plen = m->m_pkthdr.len + n->m_pkthdr.len;
4029 	if (txp->plen > 65535)
4030 		return (1);
4031 
4032 	txp->npkt = 2;
4033 	set_mbuf_len16(m, l1);
4034 	set_mbuf_len16(n, l2);
4035 
4036 	return (0);
4037 }
4038 
4039 static int
4040 add_to_txpkts(struct mbuf *m, struct txpkts *txp, u_int available)
4041 {
4042 	u_int plen, len16, needed, nsegs;
4043 
4044 	MPASS(txp->wr_type == 0 || txp->wr_type == 1);
4045 
4046 	nsegs = mbuf_nsegs(m);
4047 	if (needs_tso(m) || (txp->wr_type == 1 && nsegs != 1))
4048 		return (1);
4049 
4050 	plen = txp->plen + m->m_pkthdr.len;
4051 	if (plen > 65535)
4052 		return (1);
4053 
4054 	if (txp->wr_type == 0)
4055 		len16 = txpkts0_len16(nsegs);
4056 	else
4057 		len16 = txpkts1_len16();
4058 	needed = howmany(txp->len16 + len16, EQ_ESIZE / 16);
4059 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4060 		return (1);
4061 
4062 	txp->npkt++;
4063 	txp->plen = plen;
4064 	txp->len16 += len16;
4065 	set_mbuf_len16(m, len16);
4066 
4067 	return (0);
4068 }
4069 
4070 /*
4071  * Write a txpkts WR for the packets in txp to the hardware descriptors, update
4072  * the software descriptor, and advance the pidx.  It is guaranteed that enough
4073  * descriptors are available.
4074  *
4075  * The return value is the # of hardware descriptors used.
4076  */
4077 static u_int
4078 write_txpkts_wr(struct sge_txq *txq, struct fw_eth_tx_pkts_wr *wr,
4079     struct mbuf *m0, const struct txpkts *txp, u_int available)
4080 {
4081 	struct sge_eq *eq = &txq->eq;
4082 	struct tx_sdesc *txsd;
4083 	struct cpl_tx_pkt_core *cpl;
4084 	uint32_t ctrl;
4085 	uint64_t ctrl1;
4086 	int ndesc, checkwrap;
4087 	struct mbuf *m;
4088 	void *flitp;
4089 
4090 	TXQ_LOCK_ASSERT_OWNED(txq);
4091 	MPASS(txp->npkt > 0);
4092 	MPASS(txp->plen < 65536);
4093 	MPASS(m0 != NULL);
4094 	MPASS(m0->m_nextpkt != NULL);
4095 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
4096 	MPASS(available > 0 && available < eq->sidx);
4097 
4098 	ndesc = howmany(txp->len16, EQ_ESIZE / 16);
4099 	MPASS(ndesc <= available);
4100 
4101 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4102 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
4103 	ctrl = V_FW_WR_LEN16(txp->len16);
4104 	wr->equiq_to_len16 = htobe32(ctrl);
4105 	wr->plen = htobe16(txp->plen);
4106 	wr->npkt = txp->npkt;
4107 	wr->r3 = 0;
4108 	wr->type = txp->wr_type;
4109 	flitp = wr + 1;
4110 
4111 	/*
4112 	 * At this point we are 16B into a hardware descriptor.  If checkwrap is
4113 	 * set then we know the WR is going to wrap around somewhere.  We'll
4114 	 * check for that at appropriate points.
4115 	 */
4116 	checkwrap = eq->sidx - ndesc < eq->pidx;
4117 	for (m = m0; m != NULL; m = m->m_nextpkt) {
4118 		if (txp->wr_type == 0) {
4119 			struct ulp_txpkt *ulpmc;
4120 			struct ulptx_idata *ulpsc;
4121 
4122 			/* ULP master command */
4123 			ulpmc = flitp;
4124 			ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
4125 			    V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid));
4126 			ulpmc->len = htobe32(mbuf_len16(m));
4127 
4128 			/* ULP subcommand */
4129 			ulpsc = (void *)(ulpmc + 1);
4130 			ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
4131 			    F_ULP_TX_SC_MORE);
4132 			ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
4133 
4134 			cpl = (void *)(ulpsc + 1);
4135 			if (checkwrap &&
4136 			    (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx])
4137 				cpl = (void *)&eq->desc[0];
4138 			txq->txpkts0_pkts += txp->npkt;
4139 			txq->txpkts0_wrs++;
4140 		} else {
4141 			cpl = flitp;
4142 			txq->txpkts1_pkts += txp->npkt;
4143 			txq->txpkts1_wrs++;
4144 		}
4145 
4146 		/* Checksum offload */
4147 		ctrl1 = 0;
4148 		if (needs_l3_csum(m) == 0)
4149 			ctrl1 |= F_TXPKT_IPCSUM_DIS;
4150 		if (needs_l4_csum(m) == 0)
4151 			ctrl1 |= F_TXPKT_L4CSUM_DIS;
4152 		if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4153 		    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4154 			txq->txcsum++;	/* some hardware assistance provided */
4155 
4156 		/* VLAN tag insertion */
4157 		if (needs_vlan_insertion(m)) {
4158 			ctrl1 |= F_TXPKT_VLAN_VLD |
4159 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
4160 			txq->vlan_insertion++;
4161 		}
4162 
4163 		/* CPL header */
4164 		cpl->ctrl0 = txq->cpl_ctrl0;
4165 		cpl->pack = 0;
4166 		cpl->len = htobe16(m->m_pkthdr.len);
4167 		cpl->ctrl1 = htobe64(ctrl1);
4168 
4169 		flitp = cpl + 1;
4170 		if (checkwrap &&
4171 		    (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
4172 			flitp = (void *)&eq->desc[0];
4173 
4174 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap);
4175 
4176 	}
4177 
4178 	txsd = &txq->sdesc[eq->pidx];
4179 	txsd->m = m0;
4180 	txsd->desc_used = ndesc;
4181 
4182 	return (ndesc);
4183 }
4184 
4185 /*
4186  * If the SGL ends on an address that is not 16 byte aligned, this function will
4187  * add a 0 filled flit at the end.
4188  */
4189 static void
4190 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap)
4191 {
4192 	struct sge_eq *eq = &txq->eq;
4193 	struct sglist *gl = txq->gl;
4194 	struct sglist_seg *seg;
4195 	__be64 *flitp, *wrap;
4196 	struct ulptx_sgl *usgl;
4197 	int i, nflits, nsegs;
4198 
4199 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
4200 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
4201 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
4202 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
4203 
4204 	get_pkt_gl(m, gl);
4205 	nsegs = gl->sg_nseg;
4206 	MPASS(nsegs > 0);
4207 
4208 	nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
4209 	flitp = (__be64 *)(*to);
4210 	wrap = (__be64 *)(&eq->desc[eq->sidx]);
4211 	seg = &gl->sg_segs[0];
4212 	usgl = (void *)flitp;
4213 
4214 	/*
4215 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
4216 	 * ring, so we're at least 16 bytes away from the status page.  There is
4217 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
4218 	 */
4219 
4220 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
4221 	    V_ULPTX_NSGE(nsegs));
4222 	usgl->len0 = htobe32(seg->ss_len);
4223 	usgl->addr0 = htobe64(seg->ss_paddr);
4224 	seg++;
4225 
4226 	if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) {
4227 
4228 		/* Won't wrap around at all */
4229 
4230 		for (i = 0; i < nsegs - 1; i++, seg++) {
4231 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
4232 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
4233 		}
4234 		if (i & 1)
4235 			usgl->sge[i / 2].len[1] = htobe32(0);
4236 		flitp += nflits;
4237 	} else {
4238 
4239 		/* Will wrap somewhere in the rest of the SGL */
4240 
4241 		/* 2 flits already written, write the rest flit by flit */
4242 		flitp = (void *)(usgl + 1);
4243 		for (i = 0; i < nflits - 2; i++) {
4244 			if (flitp == wrap)
4245 				flitp = (void *)eq->desc;
4246 			*flitp++ = get_flit(seg, nsegs - 1, i);
4247 		}
4248 	}
4249 
4250 	if (nflits & 1) {
4251 		MPASS(((uintptr_t)flitp) & 0xf);
4252 		*flitp++ = 0;
4253 	}
4254 
4255 	MPASS((((uintptr_t)flitp) & 0xf) == 0);
4256 	if (__predict_false(flitp == wrap))
4257 		*to = (void *)eq->desc;
4258 	else
4259 		*to = (void *)flitp;
4260 }
4261 
4262 static inline void
4263 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
4264 {
4265 
4266 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
4267 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
4268 
4269 	if (__predict_true((uintptr_t)(*to) + len <=
4270 	    (uintptr_t)&eq->desc[eq->sidx])) {
4271 		bcopy(from, *to, len);
4272 		(*to) += len;
4273 	} else {
4274 		int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
4275 
4276 		bcopy(from, *to, portion);
4277 		from += portion;
4278 		portion = len - portion;	/* remaining */
4279 		bcopy(from, (void *)eq->desc, portion);
4280 		(*to) = (caddr_t)eq->desc + portion;
4281 	}
4282 }
4283 
4284 static inline void
4285 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n)
4286 {
4287 	u_int db;
4288 
4289 	MPASS(n > 0);
4290 
4291 	db = eq->doorbells;
4292 	if (n > 1)
4293 		clrbit(&db, DOORBELL_WCWR);
4294 	wmb();
4295 
4296 	switch (ffs(db) - 1) {
4297 	case DOORBELL_UDB:
4298 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
4299 		break;
4300 
4301 	case DOORBELL_WCWR: {
4302 		volatile uint64_t *dst, *src;
4303 		int i;
4304 
4305 		/*
4306 		 * Queues whose 128B doorbell segment fits in the page do not
4307 		 * use relative qid (udb_qid is always 0).  Only queues with
4308 		 * doorbell segments can do WCWR.
4309 		 */
4310 		KASSERT(eq->udb_qid == 0 && n == 1,
4311 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
4312 		    __func__, eq->doorbells, n, eq->dbidx, eq));
4313 
4314 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
4315 		    UDBS_DB_OFFSET);
4316 		i = eq->dbidx;
4317 		src = (void *)&eq->desc[i];
4318 		while (src != (void *)&eq->desc[i + 1])
4319 			*dst++ = *src++;
4320 		wmb();
4321 		break;
4322 	}
4323 
4324 	case DOORBELL_UDBWC:
4325 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
4326 		wmb();
4327 		break;
4328 
4329 	case DOORBELL_KDB:
4330 		t4_write_reg(sc, MYPF_REG(A_SGE_PF_KDOORBELL),
4331 		    V_QID(eq->cntxt_id) | V_PIDX(n));
4332 		break;
4333 	}
4334 
4335 	IDXINCR(eq->dbidx, n, eq->sidx);
4336 }
4337 
4338 static inline u_int
4339 reclaimable_tx_desc(struct sge_eq *eq)
4340 {
4341 	uint16_t hw_cidx;
4342 
4343 	hw_cidx = read_hw_cidx(eq);
4344 	return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx));
4345 }
4346 
4347 static inline u_int
4348 total_available_tx_desc(struct sge_eq *eq)
4349 {
4350 	uint16_t hw_cidx, pidx;
4351 
4352 	hw_cidx = read_hw_cidx(eq);
4353 	pidx = eq->pidx;
4354 
4355 	if (pidx == hw_cidx)
4356 		return (eq->sidx - 1);
4357 	else
4358 		return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1);
4359 }
4360 
4361 static inline uint16_t
4362 read_hw_cidx(struct sge_eq *eq)
4363 {
4364 	struct sge_qstat *spg = (void *)&eq->desc[eq->sidx];
4365 	uint16_t cidx = spg->cidx;	/* stable snapshot */
4366 
4367 	return (be16toh(cidx));
4368 }
4369 
4370 /*
4371  * Reclaim 'n' descriptors approximately.
4372  */
4373 static u_int
4374 reclaim_tx_descs(struct sge_txq *txq, u_int n)
4375 {
4376 	struct tx_sdesc *txsd;
4377 	struct sge_eq *eq = &txq->eq;
4378 	u_int can_reclaim, reclaimed;
4379 
4380 	TXQ_LOCK_ASSERT_OWNED(txq);
4381 	MPASS(n > 0);
4382 
4383 	reclaimed = 0;
4384 	can_reclaim = reclaimable_tx_desc(eq);
4385 	while (can_reclaim && reclaimed < n) {
4386 		int ndesc;
4387 		struct mbuf *m, *nextpkt;
4388 
4389 		txsd = &txq->sdesc[eq->cidx];
4390 		ndesc = txsd->desc_used;
4391 
4392 		/* Firmware doesn't return "partial" credits. */
4393 		KASSERT(can_reclaim >= ndesc,
4394 		    ("%s: unexpected number of credits: %d, %d",
4395 		    __func__, can_reclaim, ndesc));
4396 
4397 		for (m = txsd->m; m != NULL; m = nextpkt) {
4398 			nextpkt = m->m_nextpkt;
4399 			m->m_nextpkt = NULL;
4400 			m_freem(m);
4401 		}
4402 		reclaimed += ndesc;
4403 		can_reclaim -= ndesc;
4404 		IDXINCR(eq->cidx, ndesc, eq->sidx);
4405 	}
4406 
4407 	return (reclaimed);
4408 }
4409 
4410 static void
4411 tx_reclaim(void *arg, int n)
4412 {
4413 	struct sge_txq *txq = arg;
4414 	struct sge_eq *eq = &txq->eq;
4415 
4416 	do {
4417 		if (TXQ_TRYLOCK(txq) == 0)
4418 			break;
4419 		n = reclaim_tx_descs(txq, 32);
4420 		if (eq->cidx == eq->pidx)
4421 			eq->equeqidx = eq->pidx;
4422 		TXQ_UNLOCK(txq);
4423 	} while (n > 0);
4424 }
4425 
4426 static __be64
4427 get_flit(struct sglist_seg *segs, int nsegs, int idx)
4428 {
4429 	int i = (idx / 3) * 2;
4430 
4431 	switch (idx % 3) {
4432 	case 0: {
4433 		__be64 rc;
4434 
4435 		rc = htobe32(segs[i].ss_len);
4436 		if (i + 1 < nsegs)
4437 			rc |= (uint64_t)htobe32(segs[i + 1].ss_len) << 32;
4438 
4439 		return (rc);
4440 	}
4441 	case 1:
4442 		return (htobe64(segs[i].ss_paddr));
4443 	case 2:
4444 		return (htobe64(segs[i + 1].ss_paddr));
4445 	}
4446 
4447 	return (0);
4448 }
4449 
4450 static void
4451 find_best_refill_source(struct adapter *sc, struct sge_fl *fl, int maxp)
4452 {
4453 	int8_t zidx, hwidx, idx;
4454 	uint16_t region1, region3;
4455 	int spare, spare_needed, n;
4456 	struct sw_zone_info *swz;
4457 	struct hw_buf_info *hwb, *hwb_list = &sc->sge.hw_buf_info[0];
4458 
4459 	/*
4460 	 * Buffer Packing: Look for PAGE_SIZE or larger zone which has a bufsize
4461 	 * large enough for the max payload and cluster metadata.  Otherwise
4462 	 * settle for the largest bufsize that leaves enough room in the cluster
4463 	 * for metadata.
4464 	 *
4465 	 * Without buffer packing: Look for the smallest zone which has a
4466 	 * bufsize large enough for the max payload.  Settle for the largest
4467 	 * bufsize available if there's nothing big enough for max payload.
4468 	 */
4469 	spare_needed = fl->flags & FL_BUF_PACKING ? CL_METADATA_SIZE : 0;
4470 	swz = &sc->sge.sw_zone_info[0];
4471 	hwidx = -1;
4472 	for (zidx = 0; zidx < SW_ZONE_SIZES; zidx++, swz++) {
4473 		if (swz->size > largest_rx_cluster) {
4474 			if (__predict_true(hwidx != -1))
4475 				break;
4476 
4477 			/*
4478 			 * This is a misconfiguration.  largest_rx_cluster is
4479 			 * preventing us from finding a refill source.  See
4480 			 * dev.t5nex.<n>.buffer_sizes to figure out why.
4481 			 */
4482 			device_printf(sc->dev, "largest_rx_cluster=%u leaves no"
4483 			    " refill source for fl %p (dma %u).  Ignored.\n",
4484 			    largest_rx_cluster, fl, maxp);
4485 		}
4486 		for (idx = swz->head_hwidx; idx != -1; idx = hwb->next) {
4487 			hwb = &hwb_list[idx];
4488 			spare = swz->size - hwb->size;
4489 			if (spare < spare_needed)
4490 				continue;
4491 
4492 			hwidx = idx;		/* best option so far */
4493 			if (hwb->size >= maxp) {
4494 
4495 				if ((fl->flags & FL_BUF_PACKING) == 0)
4496 					goto done; /* stop looking (not packing) */
4497 
4498 				if (swz->size >= safest_rx_cluster)
4499 					goto done; /* stop looking (packing) */
4500 			}
4501 			break;		/* keep looking, next zone */
4502 		}
4503 	}
4504 done:
4505 	/* A usable hwidx has been located. */
4506 	MPASS(hwidx != -1);
4507 	hwb = &hwb_list[hwidx];
4508 	zidx = hwb->zidx;
4509 	swz = &sc->sge.sw_zone_info[zidx];
4510 	region1 = 0;
4511 	region3 = swz->size - hwb->size;
4512 
4513 	/*
4514 	 * Stay within this zone and see if there is a better match when mbuf
4515 	 * inlining is allowed.  Remember that the hwidx's are sorted in
4516 	 * decreasing order of size (so in increasing order of spare area).
4517 	 */
4518 	for (idx = hwidx; idx != -1; idx = hwb->next) {
4519 		hwb = &hwb_list[idx];
4520 		spare = swz->size - hwb->size;
4521 
4522 		if (allow_mbufs_in_cluster == 0 || hwb->size < maxp)
4523 			break;
4524 
4525 		/*
4526 		 * Do not inline mbufs if doing so would violate the pad/pack
4527 		 * boundary alignment requirement.
4528 		 */
4529 		if (fl_pad && (MSIZE % sc->sge.pad_boundary) != 0)
4530 			continue;
4531 		if (fl->flags & FL_BUF_PACKING &&
4532 		    (MSIZE % sc->sge.pack_boundary) != 0)
4533 			continue;
4534 
4535 		if (spare < CL_METADATA_SIZE + MSIZE)
4536 			continue;
4537 		n = (spare - CL_METADATA_SIZE) / MSIZE;
4538 		if (n > howmany(hwb->size, maxp))
4539 			break;
4540 
4541 		hwidx = idx;
4542 		if (fl->flags & FL_BUF_PACKING) {
4543 			region1 = n * MSIZE;
4544 			region3 = spare - region1;
4545 		} else {
4546 			region1 = MSIZE;
4547 			region3 = spare - region1;
4548 			break;
4549 		}
4550 	}
4551 
4552 	KASSERT(zidx >= 0 && zidx < SW_ZONE_SIZES,
4553 	    ("%s: bad zone %d for fl %p, maxp %d", __func__, zidx, fl, maxp));
4554 	KASSERT(hwidx >= 0 && hwidx <= SGE_FLBUF_SIZES,
4555 	    ("%s: bad hwidx %d for fl %p, maxp %d", __func__, hwidx, fl, maxp));
4556 	KASSERT(region1 + sc->sge.hw_buf_info[hwidx].size + region3 ==
4557 	    sc->sge.sw_zone_info[zidx].size,
4558 	    ("%s: bad buffer layout for fl %p, maxp %d. "
4559 		"cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
4560 		sc->sge.sw_zone_info[zidx].size, region1,
4561 		sc->sge.hw_buf_info[hwidx].size, region3));
4562 	if (fl->flags & FL_BUF_PACKING || region1 > 0) {
4563 		KASSERT(region3 >= CL_METADATA_SIZE,
4564 		    ("%s: no room for metadata.  fl %p, maxp %d; "
4565 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
4566 		    sc->sge.sw_zone_info[zidx].size, region1,
4567 		    sc->sge.hw_buf_info[hwidx].size, region3));
4568 		KASSERT(region1 % MSIZE == 0,
4569 		    ("%s: bad mbuf region for fl %p, maxp %d. "
4570 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
4571 		    sc->sge.sw_zone_info[zidx].size, region1,
4572 		    sc->sge.hw_buf_info[hwidx].size, region3));
4573 	}
4574 
4575 	fl->cll_def.zidx = zidx;
4576 	fl->cll_def.hwidx = hwidx;
4577 	fl->cll_def.region1 = region1;
4578 	fl->cll_def.region3 = region3;
4579 }
4580 
4581 static void
4582 find_safe_refill_source(struct adapter *sc, struct sge_fl *fl)
4583 {
4584 	struct sge *s = &sc->sge;
4585 	struct hw_buf_info *hwb;
4586 	struct sw_zone_info *swz;
4587 	int spare;
4588 	int8_t hwidx;
4589 
4590 	if (fl->flags & FL_BUF_PACKING)
4591 		hwidx = s->safe_hwidx2;	/* with room for metadata */
4592 	else if (allow_mbufs_in_cluster && s->safe_hwidx2 != -1) {
4593 		hwidx = s->safe_hwidx2;
4594 		hwb = &s->hw_buf_info[hwidx];
4595 		swz = &s->sw_zone_info[hwb->zidx];
4596 		spare = swz->size - hwb->size;
4597 
4598 		/* no good if there isn't room for an mbuf as well */
4599 		if (spare < CL_METADATA_SIZE + MSIZE)
4600 			hwidx = s->safe_hwidx1;
4601 	} else
4602 		hwidx = s->safe_hwidx1;
4603 
4604 	if (hwidx == -1) {
4605 		/* No fallback source */
4606 		fl->cll_alt.hwidx = -1;
4607 		fl->cll_alt.zidx = -1;
4608 
4609 		return;
4610 	}
4611 
4612 	hwb = &s->hw_buf_info[hwidx];
4613 	swz = &s->sw_zone_info[hwb->zidx];
4614 	spare = swz->size - hwb->size;
4615 	fl->cll_alt.hwidx = hwidx;
4616 	fl->cll_alt.zidx = hwb->zidx;
4617 	if (allow_mbufs_in_cluster &&
4618 	    (fl_pad == 0 || (MSIZE % sc->sge.pad_boundary) == 0))
4619 		fl->cll_alt.region1 = ((spare - CL_METADATA_SIZE) / MSIZE) * MSIZE;
4620 	else
4621 		fl->cll_alt.region1 = 0;
4622 	fl->cll_alt.region3 = spare - fl->cll_alt.region1;
4623 }
4624 
4625 static void
4626 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
4627 {
4628 	mtx_lock(&sc->sfl_lock);
4629 	FL_LOCK(fl);
4630 	if ((fl->flags & FL_DOOMED) == 0) {
4631 		fl->flags |= FL_STARVING;
4632 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
4633 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
4634 	}
4635 	FL_UNLOCK(fl);
4636 	mtx_unlock(&sc->sfl_lock);
4637 }
4638 
4639 static void
4640 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq)
4641 {
4642 	struct sge_wrq *wrq = (void *)eq;
4643 
4644 	atomic_readandclear_int(&eq->equiq);
4645 	taskqueue_enqueue(sc->tq[eq->tx_chan], &wrq->wrq_tx_task);
4646 }
4647 
4648 static void
4649 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq)
4650 {
4651 	struct sge_txq *txq = (void *)eq;
4652 
4653 	MPASS((eq->flags & EQ_TYPEMASK) == EQ_ETH);
4654 
4655 	atomic_readandclear_int(&eq->equiq);
4656 	mp_ring_check_drainage(txq->r, 0);
4657 	taskqueue_enqueue(sc->tq[eq->tx_chan], &txq->tx_reclaim_task);
4658 }
4659 
4660 static int
4661 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
4662     struct mbuf *m)
4663 {
4664 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
4665 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
4666 	struct adapter *sc = iq->adapter;
4667 	struct sge *s = &sc->sge;
4668 	struct sge_eq *eq;
4669 	static void (*h[])(struct adapter *, struct sge_eq *) = {NULL,
4670 		&handle_wrq_egr_update, &handle_eth_egr_update,
4671 		&handle_wrq_egr_update};
4672 
4673 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
4674 	    rss->opcode));
4675 
4676 	eq = s->eqmap[qid - s->eq_start];
4677 	(*h[eq->flags & EQ_TYPEMASK])(sc, eq);
4678 
4679 	return (0);
4680 }
4681 
4682 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
4683 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
4684     offsetof(struct cpl_fw6_msg, data));
4685 
4686 static int
4687 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
4688 {
4689 	struct adapter *sc = iq->adapter;
4690 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
4691 
4692 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
4693 	    rss->opcode));
4694 
4695 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
4696 		const struct rss_header *rss2;
4697 
4698 		rss2 = (const struct rss_header *)&cpl->data[0];
4699 		return (sc->cpl_handler[rss2->opcode](iq, rss2, m));
4700 	}
4701 
4702 	return (sc->fw_msg_handler[cpl->type](sc, &cpl->data[0]));
4703 }
4704 
4705 static int
4706 sysctl_uint16(SYSCTL_HANDLER_ARGS)
4707 {
4708 	uint16_t *id = arg1;
4709 	int i = *id;
4710 
4711 	return sysctl_handle_int(oidp, &i, 0, req);
4712 }
4713 
4714 static int
4715 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
4716 {
4717 	struct sge *s = arg1;
4718 	struct hw_buf_info *hwb = &s->hw_buf_info[0];
4719 	struct sw_zone_info *swz = &s->sw_zone_info[0];
4720 	int i, rc;
4721 	struct sbuf sb;
4722 	char c;
4723 
4724 	sbuf_new(&sb, NULL, 32, SBUF_AUTOEXTEND);
4725 	for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
4726 		if (hwb->zidx >= 0 && swz[hwb->zidx].size <= largest_rx_cluster)
4727 			c = '*';
4728 		else
4729 			c = '\0';
4730 
4731 		sbuf_printf(&sb, "%u%c ", hwb->size, c);
4732 	}
4733 	sbuf_trim(&sb);
4734 	sbuf_finish(&sb);
4735 	rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
4736 	sbuf_delete(&sb);
4737 	return (rc);
4738 }
4739