xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision a7dea1671b87c07d2d266f836bfa8b58efc7c134)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2011 Chelsio Communications, Inc.
5  * All rights reserved.
6  * Written by: Navdeep Parhar <np@FreeBSD.org>
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include "opt_inet.h"
34 #include "opt_inet6.h"
35 #include "opt_kern_tls.h"
36 #include "opt_ratelimit.h"
37 
38 #include <sys/types.h>
39 #include <sys/eventhandler.h>
40 #include <sys/mbuf.h>
41 #include <sys/socket.h>
42 #include <sys/kernel.h>
43 #include <sys/ktls.h>
44 #include <sys/malloc.h>
45 #include <sys/queue.h>
46 #include <sys/sbuf.h>
47 #include <sys/taskqueue.h>
48 #include <sys/time.h>
49 #include <sys/sglist.h>
50 #include <sys/sysctl.h>
51 #include <sys/smp.h>
52 #include <sys/socketvar.h>
53 #include <sys/counter.h>
54 #include <net/bpf.h>
55 #include <net/ethernet.h>
56 #include <net/if.h>
57 #include <net/if_vlan_var.h>
58 #include <netinet/in.h>
59 #include <netinet/ip.h>
60 #include <netinet/ip6.h>
61 #include <netinet/tcp.h>
62 #include <netinet/udp.h>
63 #include <machine/in_cksum.h>
64 #include <machine/md_var.h>
65 #include <vm/vm.h>
66 #include <vm/pmap.h>
67 #ifdef DEV_NETMAP
68 #include <machine/bus.h>
69 #include <sys/selinfo.h>
70 #include <net/if_var.h>
71 #include <net/netmap.h>
72 #include <dev/netmap/netmap_kern.h>
73 #endif
74 
75 #include "common/common.h"
76 #include "common/t4_regs.h"
77 #include "common/t4_regs_values.h"
78 #include "common/t4_msg.h"
79 #include "t4_l2t.h"
80 #include "t4_mp_ring.h"
81 
82 #ifdef T4_PKT_TIMESTAMP
83 #define RX_COPY_THRESHOLD (MINCLSIZE - 8)
84 #else
85 #define RX_COPY_THRESHOLD MINCLSIZE
86 #endif
87 
88 /* Internal mbuf flags stored in PH_loc.eight[1]. */
89 #define	MC_NOMAP		0x01
90 #define	MC_RAW_WR		0x02
91 #define	MC_TLS			0x04
92 
93 /*
94  * Ethernet frames are DMA'd at this byte offset into the freelist buffer.
95  * 0-7 are valid values.
96  */
97 static int fl_pktshift = 0;
98 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pktshift, CTLFLAG_RDTUN, &fl_pktshift, 0,
99     "payload DMA offset in rx buffer (bytes)");
100 
101 /*
102  * Pad ethernet payload up to this boundary.
103  * -1: driver should figure out a good value.
104  *  0: disable padding.
105  *  Any power of 2 from 32 to 4096 (both inclusive) is also a valid value.
106  */
107 int fl_pad = -1;
108 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pad, CTLFLAG_RDTUN, &fl_pad, 0,
109     "payload pad boundary (bytes)");
110 
111 /*
112  * Status page length.
113  * -1: driver should figure out a good value.
114  *  64 or 128 are the only other valid values.
115  */
116 static int spg_len = -1;
117 SYSCTL_INT(_hw_cxgbe, OID_AUTO, spg_len, CTLFLAG_RDTUN, &spg_len, 0,
118     "status page size (bytes)");
119 
120 /*
121  * Congestion drops.
122  * -1: no congestion feedback (not recommended).
123  *  0: backpressure the channel instead of dropping packets right away.
124  *  1: no backpressure, drop packets for the congested queue immediately.
125  */
126 static int cong_drop = 0;
127 SYSCTL_INT(_hw_cxgbe, OID_AUTO, cong_drop, CTLFLAG_RDTUN, &cong_drop, 0,
128     "Congestion control for RX queues (0 = backpressure, 1 = drop");
129 
130 /*
131  * Deliver multiple frames in the same free list buffer if they fit.
132  * -1: let the driver decide whether to enable buffer packing or not.
133  *  0: disable buffer packing.
134  *  1: enable buffer packing.
135  */
136 static int buffer_packing = -1;
137 SYSCTL_INT(_hw_cxgbe, OID_AUTO, buffer_packing, CTLFLAG_RDTUN, &buffer_packing,
138     0, "Enable buffer packing");
139 
140 /*
141  * Start next frame in a packed buffer at this boundary.
142  * -1: driver should figure out a good value.
143  * T4: driver will ignore this and use the same value as fl_pad above.
144  * T5: 16, or a power of 2 from 64 to 4096 (both inclusive) is a valid value.
145  */
146 static int fl_pack = -1;
147 SYSCTL_INT(_hw_cxgbe, OID_AUTO, fl_pack, CTLFLAG_RDTUN, &fl_pack, 0,
148     "payload pack boundary (bytes)");
149 
150 /*
151  * Allow the driver to create mbuf(s) in a cluster allocated for rx.
152  * 0: never; always allocate mbufs from the zone_mbuf UMA zone.
153  * 1: ok to create mbuf(s) within a cluster if there is room.
154  */
155 static int allow_mbufs_in_cluster = 1;
156 SYSCTL_INT(_hw_cxgbe, OID_AUTO, allow_mbufs_in_cluster, CTLFLAG_RDTUN,
157     &allow_mbufs_in_cluster, 0,
158     "Allow driver to create mbufs within a rx cluster");
159 
160 /*
161  * Largest rx cluster size that the driver is allowed to allocate.
162  */
163 static int largest_rx_cluster = MJUM16BYTES;
164 SYSCTL_INT(_hw_cxgbe, OID_AUTO, largest_rx_cluster, CTLFLAG_RDTUN,
165     &largest_rx_cluster, 0, "Largest rx cluster (bytes)");
166 
167 /*
168  * Size of cluster allocation that's most likely to succeed.  The driver will
169  * fall back to this size if it fails to allocate clusters larger than this.
170  */
171 static int safest_rx_cluster = PAGE_SIZE;
172 SYSCTL_INT(_hw_cxgbe, OID_AUTO, safest_rx_cluster, CTLFLAG_RDTUN,
173     &safest_rx_cluster, 0, "Safe rx cluster (bytes)");
174 
175 #ifdef RATELIMIT
176 /*
177  * Knob to control TCP timestamp rewriting, and the granularity of the tick used
178  * for rewriting.  -1 and 0-3 are all valid values.
179  * -1: hardware should leave the TCP timestamps alone.
180  * 0: 1ms
181  * 1: 100us
182  * 2: 10us
183  * 3: 1us
184  */
185 static int tsclk = -1;
186 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tsclk, CTLFLAG_RDTUN, &tsclk, 0,
187     "Control TCP timestamp rewriting when using pacing");
188 
189 static int eo_max_backlog = 1024 * 1024;
190 SYSCTL_INT(_hw_cxgbe, OID_AUTO, eo_max_backlog, CTLFLAG_RDTUN, &eo_max_backlog,
191     0, "Maximum backlog of ratelimited data per flow");
192 #endif
193 
194 /*
195  * The interrupt holdoff timers are multiplied by this value on T6+.
196  * 1 and 3-17 (both inclusive) are legal values.
197  */
198 static int tscale = 1;
199 SYSCTL_INT(_hw_cxgbe, OID_AUTO, tscale, CTLFLAG_RDTUN, &tscale, 0,
200     "Interrupt holdoff timer scale on T6+");
201 
202 /*
203  * Number of LRO entries in the lro_ctrl structure per rx queue.
204  */
205 static int lro_entries = TCP_LRO_ENTRIES;
206 SYSCTL_INT(_hw_cxgbe, OID_AUTO, lro_entries, CTLFLAG_RDTUN, &lro_entries, 0,
207     "Number of LRO entries per RX queue");
208 
209 /*
210  * This enables presorting of frames before they're fed into tcp_lro_rx.
211  */
212 static int lro_mbufs = 0;
213 SYSCTL_INT(_hw_cxgbe, OID_AUTO, lro_mbufs, CTLFLAG_RDTUN, &lro_mbufs, 0,
214     "Enable presorting of LRO frames");
215 
216 struct txpkts {
217 	u_int wr_type;		/* type 0 or type 1 */
218 	u_int npkt;		/* # of packets in this work request */
219 	u_int plen;		/* total payload (sum of all packets) */
220 	u_int len16;		/* # of 16B pieces used by this work request */
221 };
222 
223 /* A packet's SGL.  This + m_pkthdr has all info needed for tx */
224 struct sgl {
225 	struct sglist sg;
226 	struct sglist_seg seg[TX_SGL_SEGS];
227 };
228 
229 static int service_iq(struct sge_iq *, int);
230 static int service_iq_fl(struct sge_iq *, int);
231 static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t);
232 static int t4_eth_rx(struct sge_iq *, const struct rss_header *, struct mbuf *);
233 static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int);
234 static inline void init_fl(struct adapter *, struct sge_fl *, int, int, char *);
235 static inline void init_eq(struct adapter *, struct sge_eq *, int, int, uint8_t,
236     uint16_t, char *);
237 static int alloc_ring(struct adapter *, size_t, bus_dma_tag_t *, bus_dmamap_t *,
238     bus_addr_t *, void **);
239 static int free_ring(struct adapter *, bus_dma_tag_t, bus_dmamap_t, bus_addr_t,
240     void *);
241 static int alloc_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *,
242     int, int);
243 static int free_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *);
244 static void add_iq_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *,
245     struct sge_iq *);
246 static void add_fl_sysctls(struct adapter *, struct sysctl_ctx_list *,
247     struct sysctl_oid *, struct sge_fl *);
248 static int alloc_fwq(struct adapter *);
249 static int free_fwq(struct adapter *);
250 static int alloc_ctrlq(struct adapter *, struct sge_wrq *, int,
251     struct sysctl_oid *);
252 static int alloc_rxq(struct vi_info *, struct sge_rxq *, int, int,
253     struct sysctl_oid *);
254 static int free_rxq(struct vi_info *, struct sge_rxq *);
255 #ifdef TCP_OFFLOAD
256 static int alloc_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *, int, int,
257     struct sysctl_oid *);
258 static int free_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *);
259 #endif
260 #ifdef DEV_NETMAP
261 static int alloc_nm_rxq(struct vi_info *, struct sge_nm_rxq *, int, int,
262     struct sysctl_oid *);
263 static int free_nm_rxq(struct vi_info *, struct sge_nm_rxq *);
264 static int alloc_nm_txq(struct vi_info *, struct sge_nm_txq *, int, int,
265     struct sysctl_oid *);
266 static int free_nm_txq(struct vi_info *, struct sge_nm_txq *);
267 #endif
268 static int ctrl_eq_alloc(struct adapter *, struct sge_eq *);
269 static int eth_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
270 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
271 static int ofld_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *);
272 #endif
273 static int alloc_eq(struct adapter *, struct vi_info *, struct sge_eq *);
274 static int free_eq(struct adapter *, struct sge_eq *);
275 static int alloc_wrq(struct adapter *, struct vi_info *, struct sge_wrq *,
276     struct sysctl_oid *);
277 static int free_wrq(struct adapter *, struct sge_wrq *);
278 static int alloc_txq(struct vi_info *, struct sge_txq *, int,
279     struct sysctl_oid *);
280 static int free_txq(struct vi_info *, struct sge_txq *);
281 static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int);
282 static inline void ring_fl_db(struct adapter *, struct sge_fl *);
283 static int refill_fl(struct adapter *, struct sge_fl *, int);
284 static void refill_sfl(void *);
285 static int alloc_fl_sdesc(struct sge_fl *);
286 static void free_fl_sdesc(struct adapter *, struct sge_fl *);
287 static void find_best_refill_source(struct adapter *, struct sge_fl *, int);
288 static void find_safe_refill_source(struct adapter *, struct sge_fl *);
289 static void add_fl_to_sfl(struct adapter *, struct sge_fl *);
290 
291 static inline void get_pkt_gl(struct mbuf *, struct sglist *);
292 static inline u_int txpkt_len16(u_int, u_int);
293 static inline u_int txpkt_vm_len16(u_int, u_int);
294 static inline u_int txpkts0_len16(u_int);
295 static inline u_int txpkts1_len16(void);
296 static u_int write_raw_wr(struct sge_txq *, void *, struct mbuf *, u_int);
297 static u_int write_txpkt_wr(struct adapter *, struct sge_txq *,
298     struct fw_eth_tx_pkt_wr *, struct mbuf *, u_int);
299 static u_int write_txpkt_vm_wr(struct adapter *, struct sge_txq *,
300     struct fw_eth_tx_pkt_vm_wr *, struct mbuf *, u_int);
301 static int try_txpkts(struct mbuf *, struct mbuf *, struct txpkts *, u_int);
302 static int add_to_txpkts(struct mbuf *, struct txpkts *, u_int);
303 static u_int write_txpkts_wr(struct adapter *, struct sge_txq *,
304     struct fw_eth_tx_pkts_wr *, struct mbuf *, const struct txpkts *, u_int);
305 static void write_gl_to_txd(struct sge_txq *, struct mbuf *, caddr_t *, int);
306 static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int);
307 static inline void ring_eq_db(struct adapter *, struct sge_eq *, u_int);
308 static inline uint16_t read_hw_cidx(struct sge_eq *);
309 static inline u_int reclaimable_tx_desc(struct sge_eq *);
310 static inline u_int total_available_tx_desc(struct sge_eq *);
311 static u_int reclaim_tx_descs(struct sge_txq *, u_int);
312 static void tx_reclaim(void *, int);
313 static __be64 get_flit(struct sglist_seg *, int, int);
314 static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *,
315     struct mbuf *);
316 static int handle_fw_msg(struct sge_iq *, const struct rss_header *,
317     struct mbuf *);
318 static int t4_handle_wrerr_rpl(struct adapter *, const __be64 *);
319 static void wrq_tx_drain(void *, int);
320 static void drain_wrq_wr_list(struct adapter *, struct sge_wrq *);
321 
322 static int sysctl_uint16(SYSCTL_HANDLER_ARGS);
323 static int sysctl_bufsizes(SYSCTL_HANDLER_ARGS);
324 #ifdef RATELIMIT
325 static inline u_int txpkt_eo_len16(u_int, u_int, u_int);
326 static int ethofld_fw4_ack(struct sge_iq *, const struct rss_header *,
327     struct mbuf *);
328 #endif
329 
330 static counter_u64_t extfree_refs;
331 static counter_u64_t extfree_rels;
332 
333 an_handler_t t4_an_handler;
334 fw_msg_handler_t t4_fw_msg_handler[NUM_FW6_TYPES];
335 cpl_handler_t t4_cpl_handler[NUM_CPL_CMDS];
336 cpl_handler_t set_tcb_rpl_handlers[NUM_CPL_COOKIES];
337 cpl_handler_t l2t_write_rpl_handlers[NUM_CPL_COOKIES];
338 cpl_handler_t act_open_rpl_handlers[NUM_CPL_COOKIES];
339 cpl_handler_t abort_rpl_rss_handlers[NUM_CPL_COOKIES];
340 cpl_handler_t fw4_ack_handlers[NUM_CPL_COOKIES];
341 
342 void
343 t4_register_an_handler(an_handler_t h)
344 {
345 	uintptr_t *loc;
346 
347 	MPASS(h == NULL || t4_an_handler == NULL);
348 
349 	loc = (uintptr_t *)&t4_an_handler;
350 	atomic_store_rel_ptr(loc, (uintptr_t)h);
351 }
352 
353 void
354 t4_register_fw_msg_handler(int type, fw_msg_handler_t h)
355 {
356 	uintptr_t *loc;
357 
358 	MPASS(type < nitems(t4_fw_msg_handler));
359 	MPASS(h == NULL || t4_fw_msg_handler[type] == NULL);
360 	/*
361 	 * These are dispatched by the handler for FW{4|6}_CPL_MSG using the CPL
362 	 * handler dispatch table.  Reject any attempt to install a handler for
363 	 * this subtype.
364 	 */
365 	MPASS(type != FW_TYPE_RSSCPL);
366 	MPASS(type != FW6_TYPE_RSSCPL);
367 
368 	loc = (uintptr_t *)&t4_fw_msg_handler[type];
369 	atomic_store_rel_ptr(loc, (uintptr_t)h);
370 }
371 
372 void
373 t4_register_cpl_handler(int opcode, cpl_handler_t h)
374 {
375 	uintptr_t *loc;
376 
377 	MPASS(opcode < nitems(t4_cpl_handler));
378 	MPASS(h == NULL || t4_cpl_handler[opcode] == NULL);
379 
380 	loc = (uintptr_t *)&t4_cpl_handler[opcode];
381 	atomic_store_rel_ptr(loc, (uintptr_t)h);
382 }
383 
384 static int
385 set_tcb_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
386     struct mbuf *m)
387 {
388 	const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
389 	u_int tid;
390 	int cookie;
391 
392 	MPASS(m == NULL);
393 
394 	tid = GET_TID(cpl);
395 	if (is_hpftid(iq->adapter, tid) || is_ftid(iq->adapter, tid)) {
396 		/*
397 		 * The return code for filter-write is put in the CPL cookie so
398 		 * we have to rely on the hardware tid (is_ftid) to determine
399 		 * that this is a response to a filter.
400 		 */
401 		cookie = CPL_COOKIE_FILTER;
402 	} else {
403 		cookie = G_COOKIE(cpl->cookie);
404 	}
405 	MPASS(cookie > CPL_COOKIE_RESERVED);
406 	MPASS(cookie < nitems(set_tcb_rpl_handlers));
407 
408 	return (set_tcb_rpl_handlers[cookie](iq, rss, m));
409 }
410 
411 static int
412 l2t_write_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
413     struct mbuf *m)
414 {
415 	const struct cpl_l2t_write_rpl *rpl = (const void *)(rss + 1);
416 	unsigned int cookie;
417 
418 	MPASS(m == NULL);
419 
420 	cookie = GET_TID(rpl) & F_SYNC_WR ? CPL_COOKIE_TOM : CPL_COOKIE_FILTER;
421 	return (l2t_write_rpl_handlers[cookie](iq, rss, m));
422 }
423 
424 static int
425 act_open_rpl_handler(struct sge_iq *iq, const struct rss_header *rss,
426     struct mbuf *m)
427 {
428 	const struct cpl_act_open_rpl *cpl = (const void *)(rss + 1);
429 	u_int cookie = G_TID_COOKIE(G_AOPEN_ATID(be32toh(cpl->atid_status)));
430 
431 	MPASS(m == NULL);
432 	MPASS(cookie != CPL_COOKIE_RESERVED);
433 
434 	return (act_open_rpl_handlers[cookie](iq, rss, m));
435 }
436 
437 static int
438 abort_rpl_rss_handler(struct sge_iq *iq, const struct rss_header *rss,
439     struct mbuf *m)
440 {
441 	struct adapter *sc = iq->adapter;
442 	u_int cookie;
443 
444 	MPASS(m == NULL);
445 	if (is_hashfilter(sc))
446 		cookie = CPL_COOKIE_HASHFILTER;
447 	else
448 		cookie = CPL_COOKIE_TOM;
449 
450 	return (abort_rpl_rss_handlers[cookie](iq, rss, m));
451 }
452 
453 static int
454 fw4_ack_handler(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
455 {
456 	struct adapter *sc = iq->adapter;
457 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
458 	unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
459 	u_int cookie;
460 
461 	MPASS(m == NULL);
462 	if (is_etid(sc, tid))
463 		cookie = CPL_COOKIE_ETHOFLD;
464 	else
465 		cookie = CPL_COOKIE_TOM;
466 
467 	return (fw4_ack_handlers[cookie](iq, rss, m));
468 }
469 
470 static void
471 t4_init_shared_cpl_handlers(void)
472 {
473 
474 	t4_register_cpl_handler(CPL_SET_TCB_RPL, set_tcb_rpl_handler);
475 	t4_register_cpl_handler(CPL_L2T_WRITE_RPL, l2t_write_rpl_handler);
476 	t4_register_cpl_handler(CPL_ACT_OPEN_RPL, act_open_rpl_handler);
477 	t4_register_cpl_handler(CPL_ABORT_RPL_RSS, abort_rpl_rss_handler);
478 	t4_register_cpl_handler(CPL_FW4_ACK, fw4_ack_handler);
479 }
480 
481 void
482 t4_register_shared_cpl_handler(int opcode, cpl_handler_t h, int cookie)
483 {
484 	uintptr_t *loc;
485 
486 	MPASS(opcode < nitems(t4_cpl_handler));
487 	MPASS(cookie > CPL_COOKIE_RESERVED);
488 	MPASS(cookie < NUM_CPL_COOKIES);
489 	MPASS(t4_cpl_handler[opcode] != NULL);
490 
491 	switch (opcode) {
492 	case CPL_SET_TCB_RPL:
493 		loc = (uintptr_t *)&set_tcb_rpl_handlers[cookie];
494 		break;
495 	case CPL_L2T_WRITE_RPL:
496 		loc = (uintptr_t *)&l2t_write_rpl_handlers[cookie];
497 		break;
498 	case CPL_ACT_OPEN_RPL:
499 		loc = (uintptr_t *)&act_open_rpl_handlers[cookie];
500 		break;
501 	case CPL_ABORT_RPL_RSS:
502 		loc = (uintptr_t *)&abort_rpl_rss_handlers[cookie];
503 		break;
504 	case CPL_FW4_ACK:
505 		loc = (uintptr_t *)&fw4_ack_handlers[cookie];
506 		break;
507 	default:
508 		MPASS(0);
509 		return;
510 	}
511 	MPASS(h == NULL || *loc == (uintptr_t)NULL);
512 	atomic_store_rel_ptr(loc, (uintptr_t)h);
513 }
514 
515 /*
516  * Called on MOD_LOAD.  Validates and calculates the SGE tunables.
517  */
518 void
519 t4_sge_modload(void)
520 {
521 
522 	if (fl_pktshift < 0 || fl_pktshift > 7) {
523 		printf("Invalid hw.cxgbe.fl_pktshift value (%d),"
524 		    " using 0 instead.\n", fl_pktshift);
525 		fl_pktshift = 0;
526 	}
527 
528 	if (spg_len != 64 && spg_len != 128) {
529 		int len;
530 
531 #if defined(__i386__) || defined(__amd64__)
532 		len = cpu_clflush_line_size > 64 ? 128 : 64;
533 #else
534 		len = 64;
535 #endif
536 		if (spg_len != -1) {
537 			printf("Invalid hw.cxgbe.spg_len value (%d),"
538 			    " using %d instead.\n", spg_len, len);
539 		}
540 		spg_len = len;
541 	}
542 
543 	if (cong_drop < -1 || cong_drop > 1) {
544 		printf("Invalid hw.cxgbe.cong_drop value (%d),"
545 		    " using 0 instead.\n", cong_drop);
546 		cong_drop = 0;
547 	}
548 
549 	if (tscale != 1 && (tscale < 3 || tscale > 17)) {
550 		printf("Invalid hw.cxgbe.tscale value (%d),"
551 		    " using 1 instead.\n", tscale);
552 		tscale = 1;
553 	}
554 
555 	extfree_refs = counter_u64_alloc(M_WAITOK);
556 	extfree_rels = counter_u64_alloc(M_WAITOK);
557 	counter_u64_zero(extfree_refs);
558 	counter_u64_zero(extfree_rels);
559 
560 	t4_init_shared_cpl_handlers();
561 	t4_register_cpl_handler(CPL_FW4_MSG, handle_fw_msg);
562 	t4_register_cpl_handler(CPL_FW6_MSG, handle_fw_msg);
563 	t4_register_cpl_handler(CPL_SGE_EGR_UPDATE, handle_sge_egr_update);
564 	t4_register_cpl_handler(CPL_RX_PKT, t4_eth_rx);
565 #ifdef RATELIMIT
566 	t4_register_shared_cpl_handler(CPL_FW4_ACK, ethofld_fw4_ack,
567 	    CPL_COOKIE_ETHOFLD);
568 #endif
569 	t4_register_fw_msg_handler(FW6_TYPE_CMD_RPL, t4_handle_fw_rpl);
570 	t4_register_fw_msg_handler(FW6_TYPE_WRERR_RPL, t4_handle_wrerr_rpl);
571 }
572 
573 void
574 t4_sge_modunload(void)
575 {
576 
577 	counter_u64_free(extfree_refs);
578 	counter_u64_free(extfree_rels);
579 }
580 
581 uint64_t
582 t4_sge_extfree_refs(void)
583 {
584 	uint64_t refs, rels;
585 
586 	rels = counter_u64_fetch(extfree_rels);
587 	refs = counter_u64_fetch(extfree_refs);
588 
589 	return (refs - rels);
590 }
591 
592 static inline void
593 setup_pad_and_pack_boundaries(struct adapter *sc)
594 {
595 	uint32_t v, m;
596 	int pad, pack, pad_shift;
597 
598 	pad_shift = chip_id(sc) > CHELSIO_T5 ? X_T6_INGPADBOUNDARY_SHIFT :
599 	    X_INGPADBOUNDARY_SHIFT;
600 	pad = fl_pad;
601 	if (fl_pad < (1 << pad_shift) ||
602 	    fl_pad > (1 << (pad_shift + M_INGPADBOUNDARY)) ||
603 	    !powerof2(fl_pad)) {
604 		/*
605 		 * If there is any chance that we might use buffer packing and
606 		 * the chip is a T4, then pick 64 as the pad/pack boundary.  Set
607 		 * it to the minimum allowed in all other cases.
608 		 */
609 		pad = is_t4(sc) && buffer_packing ? 64 : 1 << pad_shift;
610 
611 		/*
612 		 * For fl_pad = 0 we'll still write a reasonable value to the
613 		 * register but all the freelists will opt out of padding.
614 		 * We'll complain here only if the user tried to set it to a
615 		 * value greater than 0 that was invalid.
616 		 */
617 		if (fl_pad > 0) {
618 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pad value"
619 			    " (%d), using %d instead.\n", fl_pad, pad);
620 		}
621 	}
622 	m = V_INGPADBOUNDARY(M_INGPADBOUNDARY);
623 	v = V_INGPADBOUNDARY(ilog2(pad) - pad_shift);
624 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
625 
626 	if (is_t4(sc)) {
627 		if (fl_pack != -1 && fl_pack != pad) {
628 			/* Complain but carry on. */
629 			device_printf(sc->dev, "hw.cxgbe.fl_pack (%d) ignored,"
630 			    " using %d instead.\n", fl_pack, pad);
631 		}
632 		return;
633 	}
634 
635 	pack = fl_pack;
636 	if (fl_pack < 16 || fl_pack == 32 || fl_pack > 4096 ||
637 	    !powerof2(fl_pack)) {
638 		pack = max(sc->params.pci.mps, CACHE_LINE_SIZE);
639 		MPASS(powerof2(pack));
640 		if (pack < 16)
641 			pack = 16;
642 		if (pack == 32)
643 			pack = 64;
644 		if (pack > 4096)
645 			pack = 4096;
646 		if (fl_pack != -1) {
647 			device_printf(sc->dev, "Invalid hw.cxgbe.fl_pack value"
648 			    " (%d), using %d instead.\n", fl_pack, pack);
649 		}
650 	}
651 	m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY);
652 	if (pack == 16)
653 		v = V_INGPACKBOUNDARY(0);
654 	else
655 		v = V_INGPACKBOUNDARY(ilog2(pack) - 5);
656 
657 	MPASS(!is_t4(sc));	/* T4 doesn't have SGE_CONTROL2 */
658 	t4_set_reg_field(sc, A_SGE_CONTROL2, m, v);
659 }
660 
661 /*
662  * adap->params.vpd.cclk must be set up before this is called.
663  */
664 void
665 t4_tweak_chip_settings(struct adapter *sc)
666 {
667 	int i;
668 	uint32_t v, m;
669 	int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200};
670 	int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk;
671 	int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */
672 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
673 	static int sge_flbuf_sizes[] = {
674 		MCLBYTES,
675 #if MJUMPAGESIZE != MCLBYTES
676 		MJUMPAGESIZE,
677 		MJUMPAGESIZE - CL_METADATA_SIZE,
678 		MJUMPAGESIZE - 2 * MSIZE - CL_METADATA_SIZE,
679 #endif
680 		MJUM9BYTES,
681 		MJUM16BYTES,
682 		MCLBYTES - MSIZE - CL_METADATA_SIZE,
683 		MJUM9BYTES - CL_METADATA_SIZE,
684 		MJUM16BYTES - CL_METADATA_SIZE,
685 	};
686 
687 	KASSERT(sc->flags & MASTER_PF,
688 	    ("%s: trying to change chip settings when not master.", __func__));
689 
690 	m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE;
691 	v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE |
692 	    V_EGRSTATUSPAGESIZE(spg_len == 128);
693 	t4_set_reg_field(sc, A_SGE_CONTROL, m, v);
694 
695 	setup_pad_and_pack_boundaries(sc);
696 
697 	v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) |
698 	    V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) |
699 	    V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) |
700 	    V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) |
701 	    V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) |
702 	    V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) |
703 	    V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) |
704 	    V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10);
705 	t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v);
706 
707 	KASSERT(nitems(sge_flbuf_sizes) <= SGE_FLBUF_SIZES,
708 	    ("%s: hw buffer size table too big", __func__));
709 	t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0, 4096);
710 	t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE1, 65536);
711 	for (i = 0; i < min(nitems(sge_flbuf_sizes), SGE_FLBUF_SIZES); i++) {
712 		t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE15 - (4 * i),
713 		    sge_flbuf_sizes[i]);
714 	}
715 
716 	v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) |
717 	    V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]);
718 	t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v);
719 
720 	KASSERT(intr_timer[0] <= timer_max,
721 	    ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0],
722 	    timer_max));
723 	for (i = 1; i < nitems(intr_timer); i++) {
724 		KASSERT(intr_timer[i] >= intr_timer[i - 1],
725 		    ("%s: timers not listed in increasing order (%d)",
726 		    __func__, i));
727 
728 		while (intr_timer[i] > timer_max) {
729 			if (i == nitems(intr_timer) - 1) {
730 				intr_timer[i] = timer_max;
731 				break;
732 			}
733 			intr_timer[i] += intr_timer[i - 1];
734 			intr_timer[i] /= 2;
735 		}
736 	}
737 
738 	v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) |
739 	    V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1]));
740 	t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v);
741 	v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) |
742 	    V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3]));
743 	t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v);
744 	v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) |
745 	    V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5]));
746 	t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v);
747 
748 	if (chip_id(sc) >= CHELSIO_T6) {
749 		m = V_TSCALE(M_TSCALE);
750 		if (tscale == 1)
751 			v = 0;
752 		else
753 			v = V_TSCALE(tscale - 2);
754 		t4_set_reg_field(sc, A_SGE_ITP_CONTROL, m, v);
755 
756 		if (sc->debug_flags & DF_DISABLE_TCB_CACHE) {
757 			m = V_RDTHRESHOLD(M_RDTHRESHOLD) | F_WRTHRTHRESHEN |
758 			    V_WRTHRTHRESH(M_WRTHRTHRESH);
759 			t4_tp_pio_read(sc, &v, 1, A_TP_CMM_CONFIG, 1);
760 			v &= ~m;
761 			v |= V_RDTHRESHOLD(1) | F_WRTHRTHRESHEN |
762 			    V_WRTHRTHRESH(16);
763 			t4_tp_pio_write(sc, &v, 1, A_TP_CMM_CONFIG, 1);
764 		}
765 	}
766 
767 	/* 4K, 16K, 64K, 256K DDP "page sizes" for TDDP */
768 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
769 	t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v);
770 
771 	/*
772 	 * 4K, 8K, 16K, 64K DDP "page sizes" for iSCSI DDP.  These have been
773 	 * chosen with MAXPHYS = 128K in mind.  The largest DDP buffer that we
774 	 * may have to deal with is MAXPHYS + 1 page.
775 	 */
776 	v = V_HPZ0(0) | V_HPZ1(1) | V_HPZ2(2) | V_HPZ3(4);
777 	t4_write_reg(sc, A_ULP_RX_ISCSI_PSZ, v);
778 
779 	/* We use multiple DDP page sizes both in plain-TOE and ISCSI modes. */
780 	m = v = F_TDDPTAGTCB | F_ISCSITAGTCB;
781 	t4_set_reg_field(sc, A_ULP_RX_CTL, m, v);
782 
783 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
784 	    F_RESETDDPOFFSET;
785 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
786 	t4_set_reg_field(sc, A_TP_PARA_REG5, m, v);
787 }
788 
789 /*
790  * SGE wants the buffer to be at least 64B and then a multiple of 16.  If
791  * padding is in use, the buffer's start and end need to be aligned to the pad
792  * boundary as well.  We'll just make sure that the size is a multiple of the
793  * boundary here, it is up to the buffer allocation code to make sure the start
794  * of the buffer is aligned as well.
795  */
796 static inline int
797 hwsz_ok(struct adapter *sc, int hwsz)
798 {
799 	int mask = fl_pad ? sc->params.sge.pad_boundary - 1 : 16 - 1;
800 
801 	return (hwsz >= 64 && (hwsz & mask) == 0);
802 }
803 
804 /*
805  * XXX: driver really should be able to deal with unexpected settings.
806  */
807 int
808 t4_read_chip_settings(struct adapter *sc)
809 {
810 	struct sge *s = &sc->sge;
811 	struct sge_params *sp = &sc->params.sge;
812 	int i, j, n, rc = 0;
813 	uint32_t m, v, r;
814 	uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE);
815 	static int sw_buf_sizes[] = {	/* Sorted by size */
816 		MCLBYTES,
817 #if MJUMPAGESIZE != MCLBYTES
818 		MJUMPAGESIZE,
819 #endif
820 		MJUM9BYTES,
821 		MJUM16BYTES
822 	};
823 	struct sw_zone_info *swz, *safe_swz;
824 	struct hw_buf_info *hwb;
825 
826 	m = F_RXPKTCPLMODE;
827 	v = F_RXPKTCPLMODE;
828 	r = sc->params.sge.sge_control;
829 	if ((r & m) != v) {
830 		device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r);
831 		rc = EINVAL;
832 	}
833 
834 	/*
835 	 * If this changes then every single use of PAGE_SHIFT in the driver
836 	 * needs to be carefully reviewed for PAGE_SHIFT vs sp->page_shift.
837 	 */
838 	if (sp->page_shift != PAGE_SHIFT) {
839 		device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r);
840 		rc = EINVAL;
841 	}
842 
843 	/* Filter out unusable hw buffer sizes entirely (mark with -2). */
844 	hwb = &s->hw_buf_info[0];
845 	for (i = 0; i < nitems(s->hw_buf_info); i++, hwb++) {
846 		r = sc->params.sge.sge_fl_buffer_size[i];
847 		hwb->size = r;
848 		hwb->zidx = hwsz_ok(sc, r) ? -1 : -2;
849 		hwb->next = -1;
850 	}
851 
852 	/*
853 	 * Create a sorted list in decreasing order of hw buffer sizes (and so
854 	 * increasing order of spare area) for each software zone.
855 	 *
856 	 * If padding is enabled then the start and end of the buffer must align
857 	 * to the pad boundary; if packing is enabled then they must align with
858 	 * the pack boundary as well.  Allocations from the cluster zones are
859 	 * aligned to min(size, 4K), so the buffer starts at that alignment and
860 	 * ends at hwb->size alignment.  If mbuf inlining is allowed the
861 	 * starting alignment will be reduced to MSIZE and the driver will
862 	 * exercise appropriate caution when deciding on the best buffer layout
863 	 * to use.
864 	 */
865 	n = 0;	/* no usable buffer size to begin with */
866 	swz = &s->sw_zone_info[0];
867 	safe_swz = NULL;
868 	for (i = 0; i < SW_ZONE_SIZES; i++, swz++) {
869 		int8_t head = -1, tail = -1;
870 
871 		swz->size = sw_buf_sizes[i];
872 		swz->zone = m_getzone(swz->size);
873 		swz->type = m_gettype(swz->size);
874 
875 		if (swz->size < PAGE_SIZE) {
876 			MPASS(powerof2(swz->size));
877 			if (fl_pad && (swz->size % sp->pad_boundary != 0))
878 				continue;
879 		}
880 
881 		if (swz->size == safest_rx_cluster)
882 			safe_swz = swz;
883 
884 		hwb = &s->hw_buf_info[0];
885 		for (j = 0; j < SGE_FLBUF_SIZES; j++, hwb++) {
886 			if (hwb->zidx != -1 || hwb->size > swz->size)
887 				continue;
888 #ifdef INVARIANTS
889 			if (fl_pad)
890 				MPASS(hwb->size % sp->pad_boundary == 0);
891 #endif
892 			hwb->zidx = i;
893 			if (head == -1)
894 				head = tail = j;
895 			else if (hwb->size < s->hw_buf_info[tail].size) {
896 				s->hw_buf_info[tail].next = j;
897 				tail = j;
898 			} else {
899 				int8_t *cur;
900 				struct hw_buf_info *t;
901 
902 				for (cur = &head; *cur != -1; cur = &t->next) {
903 					t = &s->hw_buf_info[*cur];
904 					if (hwb->size == t->size) {
905 						hwb->zidx = -2;
906 						break;
907 					}
908 					if (hwb->size > t->size) {
909 						hwb->next = *cur;
910 						*cur = j;
911 						break;
912 					}
913 				}
914 			}
915 		}
916 		swz->head_hwidx = head;
917 		swz->tail_hwidx = tail;
918 
919 		if (tail != -1) {
920 			n++;
921 			if (swz->size - s->hw_buf_info[tail].size >=
922 			    CL_METADATA_SIZE)
923 				sc->flags |= BUF_PACKING_OK;
924 		}
925 	}
926 	if (n == 0) {
927 		device_printf(sc->dev, "no usable SGE FL buffer size.\n");
928 		rc = EINVAL;
929 	}
930 
931 	s->safe_hwidx1 = -1;
932 	s->safe_hwidx2 = -1;
933 	if (safe_swz != NULL) {
934 		s->safe_hwidx1 = safe_swz->head_hwidx;
935 		for (i = safe_swz->head_hwidx; i != -1; i = hwb->next) {
936 			int spare;
937 
938 			hwb = &s->hw_buf_info[i];
939 #ifdef INVARIANTS
940 			if (fl_pad)
941 				MPASS(hwb->size % sp->pad_boundary == 0);
942 #endif
943 			spare = safe_swz->size - hwb->size;
944 			if (spare >= CL_METADATA_SIZE) {
945 				s->safe_hwidx2 = i;
946 				break;
947 			}
948 		}
949 	}
950 
951 	if (sc->flags & IS_VF)
952 		return (0);
953 
954 	v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6);
955 	r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ);
956 	if (r != v) {
957 		device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r);
958 		rc = EINVAL;
959 	}
960 
961 	m = v = F_TDDPTAGTCB;
962 	r = t4_read_reg(sc, A_ULP_RX_CTL);
963 	if ((r & m) != v) {
964 		device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r);
965 		rc = EINVAL;
966 	}
967 
968 	m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET |
969 	    F_RESETDDPOFFSET;
970 	v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET;
971 	r = t4_read_reg(sc, A_TP_PARA_REG5);
972 	if ((r & m) != v) {
973 		device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r);
974 		rc = EINVAL;
975 	}
976 
977 	t4_init_tp_params(sc, 1);
978 
979 	t4_read_mtu_tbl(sc, sc->params.mtus, NULL);
980 	t4_load_mtus(sc, sc->params.mtus, sc->params.a_wnd, sc->params.b_wnd);
981 
982 	return (rc);
983 }
984 
985 int
986 t4_create_dma_tag(struct adapter *sc)
987 {
988 	int rc;
989 
990 	rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0,
991 	    BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE,
992 	    BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL,
993 	    NULL, &sc->dmat);
994 	if (rc != 0) {
995 		device_printf(sc->dev,
996 		    "failed to create main DMA tag: %d\n", rc);
997 	}
998 
999 	return (rc);
1000 }
1001 
1002 void
1003 t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
1004     struct sysctl_oid_list *children)
1005 {
1006 	struct sge_params *sp = &sc->params.sge;
1007 
1008 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "buffer_sizes",
1009 	    CTLTYPE_STRING | CTLFLAG_RD, &sc->sge, 0, sysctl_bufsizes, "A",
1010 	    "freelist buffer sizes");
1011 
1012 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD,
1013 	    NULL, sp->fl_pktshift, "payload DMA offset in rx buffer (bytes)");
1014 
1015 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD,
1016 	    NULL, sp->pad_boundary, "payload pad boundary (bytes)");
1017 
1018 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD,
1019 	    NULL, sp->spg_len, "status page size (bytes)");
1020 
1021 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD,
1022 	    NULL, cong_drop, "congestion drop setting");
1023 
1024 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD,
1025 	    NULL, sp->pack_boundary, "payload pack boundary (bytes)");
1026 }
1027 
1028 int
1029 t4_destroy_dma_tag(struct adapter *sc)
1030 {
1031 	if (sc->dmat)
1032 		bus_dma_tag_destroy(sc->dmat);
1033 
1034 	return (0);
1035 }
1036 
1037 /*
1038  * Allocate and initialize the firmware event queue, control queues, and special
1039  * purpose rx queues owned by the adapter.
1040  *
1041  * Returns errno on failure.  Resources allocated up to that point may still be
1042  * allocated.  Caller is responsible for cleanup in case this function fails.
1043  */
1044 int
1045 t4_setup_adapter_queues(struct adapter *sc)
1046 {
1047 	struct sysctl_oid *oid;
1048 	struct sysctl_oid_list *children;
1049 	int rc, i;
1050 
1051 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
1052 
1053 	sysctl_ctx_init(&sc->ctx);
1054 	sc->flags |= ADAP_SYSCTL_CTX;
1055 
1056 	/*
1057 	 * Firmware event queue
1058 	 */
1059 	rc = alloc_fwq(sc);
1060 	if (rc != 0)
1061 		return (rc);
1062 
1063 	/*
1064 	 * That's all for the VF driver.
1065 	 */
1066 	if (sc->flags & IS_VF)
1067 		return (rc);
1068 
1069 	oid = device_get_sysctl_tree(sc->dev);
1070 	children = SYSCTL_CHILDREN(oid);
1071 
1072 	/*
1073 	 * XXX: General purpose rx queues, one per port.
1074 	 */
1075 
1076 	/*
1077 	 * Control queues, one per port.
1078 	 */
1079 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "ctrlq",
1080 	    CTLFLAG_RD, NULL, "control queues");
1081 	for_each_port(sc, i) {
1082 		struct sge_wrq *ctrlq = &sc->sge.ctrlq[i];
1083 
1084 		rc = alloc_ctrlq(sc, ctrlq, i, oid);
1085 		if (rc != 0)
1086 			return (rc);
1087 	}
1088 
1089 	return (rc);
1090 }
1091 
1092 /*
1093  * Idempotent
1094  */
1095 int
1096 t4_teardown_adapter_queues(struct adapter *sc)
1097 {
1098 	int i;
1099 
1100 	ADAPTER_LOCK_ASSERT_NOTOWNED(sc);
1101 
1102 	/* Do this before freeing the queue */
1103 	if (sc->flags & ADAP_SYSCTL_CTX) {
1104 		sysctl_ctx_free(&sc->ctx);
1105 		sc->flags &= ~ADAP_SYSCTL_CTX;
1106 	}
1107 
1108 	if (!(sc->flags & IS_VF)) {
1109 		for_each_port(sc, i)
1110 			free_wrq(sc, &sc->sge.ctrlq[i]);
1111 	}
1112 	free_fwq(sc);
1113 
1114 	return (0);
1115 }
1116 
1117 /* Maximum payload that can be delivered with a single iq descriptor */
1118 static inline int
1119 mtu_to_max_payload(struct adapter *sc, int mtu)
1120 {
1121 
1122 	/* large enough even when hw VLAN extraction is disabled */
1123 	return (sc->params.sge.fl_pktshift + ETHER_HDR_LEN +
1124 	    ETHER_VLAN_ENCAP_LEN + mtu);
1125 }
1126 
1127 int
1128 t4_setup_vi_queues(struct vi_info *vi)
1129 {
1130 	int rc = 0, i, intr_idx, iqidx;
1131 	struct sge_rxq *rxq;
1132 	struct sge_txq *txq;
1133 #ifdef TCP_OFFLOAD
1134 	struct sge_ofld_rxq *ofld_rxq;
1135 #endif
1136 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1137 	struct sge_wrq *ofld_txq;
1138 #endif
1139 #ifdef DEV_NETMAP
1140 	int saved_idx;
1141 	struct sge_nm_rxq *nm_rxq;
1142 	struct sge_nm_txq *nm_txq;
1143 #endif
1144 	char name[16];
1145 	struct port_info *pi = vi->pi;
1146 	struct adapter *sc = pi->adapter;
1147 	struct ifnet *ifp = vi->ifp;
1148 	struct sysctl_oid *oid = device_get_sysctl_tree(vi->dev);
1149 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
1150 	int maxp, mtu = ifp->if_mtu;
1151 
1152 	/* Interrupt vector to start from (when using multiple vectors) */
1153 	intr_idx = vi->first_intr;
1154 
1155 #ifdef DEV_NETMAP
1156 	saved_idx = intr_idx;
1157 	if (ifp->if_capabilities & IFCAP_NETMAP) {
1158 
1159 		/* netmap is supported with direct interrupts only. */
1160 		MPASS(!forwarding_intr_to_fwq(sc));
1161 
1162 		/*
1163 		 * We don't have buffers to back the netmap rx queues
1164 		 * right now so we create the queues in a way that
1165 		 * doesn't set off any congestion signal in the chip.
1166 		 */
1167 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_rxq",
1168 		    CTLFLAG_RD, NULL, "rx queues");
1169 		for_each_nm_rxq(vi, i, nm_rxq) {
1170 			rc = alloc_nm_rxq(vi, nm_rxq, intr_idx, i, oid);
1171 			if (rc != 0)
1172 				goto done;
1173 			intr_idx++;
1174 		}
1175 
1176 		oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_txq",
1177 		    CTLFLAG_RD, NULL, "tx queues");
1178 		for_each_nm_txq(vi, i, nm_txq) {
1179 			iqidx = vi->first_nm_rxq + (i % vi->nnmrxq);
1180 			rc = alloc_nm_txq(vi, nm_txq, iqidx, i, oid);
1181 			if (rc != 0)
1182 				goto done;
1183 		}
1184 	}
1185 
1186 	/* Normal rx queues and netmap rx queues share the same interrupts. */
1187 	intr_idx = saved_idx;
1188 #endif
1189 
1190 	/*
1191 	 * Allocate rx queues first because a default iqid is required when
1192 	 * creating a tx queue.
1193 	 */
1194 	maxp = mtu_to_max_payload(sc, mtu);
1195 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq",
1196 	    CTLFLAG_RD, NULL, "rx queues");
1197 	for_each_rxq(vi, i, rxq) {
1198 
1199 		init_iq(&rxq->iq, sc, vi->tmr_idx, vi->pktc_idx, vi->qsize_rxq);
1200 
1201 		snprintf(name, sizeof(name), "%s rxq%d-fl",
1202 		    device_get_nameunit(vi->dev), i);
1203 		init_fl(sc, &rxq->fl, vi->qsize_rxq / 8, maxp, name);
1204 
1205 		rc = alloc_rxq(vi, rxq,
1206 		    forwarding_intr_to_fwq(sc) ? -1 : intr_idx, i, oid);
1207 		if (rc != 0)
1208 			goto done;
1209 		intr_idx++;
1210 	}
1211 #ifdef DEV_NETMAP
1212 	if (ifp->if_capabilities & IFCAP_NETMAP)
1213 		intr_idx = saved_idx + max(vi->nrxq, vi->nnmrxq);
1214 #endif
1215 #ifdef TCP_OFFLOAD
1216 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq",
1217 	    CTLFLAG_RD, NULL, "rx queues for offloaded TCP connections");
1218 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1219 
1220 		init_iq(&ofld_rxq->iq, sc, vi->ofld_tmr_idx, vi->ofld_pktc_idx,
1221 		    vi->qsize_rxq);
1222 
1223 		snprintf(name, sizeof(name), "%s ofld_rxq%d-fl",
1224 		    device_get_nameunit(vi->dev), i);
1225 		init_fl(sc, &ofld_rxq->fl, vi->qsize_rxq / 8, maxp, name);
1226 
1227 		rc = alloc_ofld_rxq(vi, ofld_rxq,
1228 		    forwarding_intr_to_fwq(sc) ? -1 : intr_idx, i, oid);
1229 		if (rc != 0)
1230 			goto done;
1231 		intr_idx++;
1232 	}
1233 #endif
1234 
1235 	/*
1236 	 * Now the tx queues.
1237 	 */
1238 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "txq", CTLFLAG_RD,
1239 	    NULL, "tx queues");
1240 	for_each_txq(vi, i, txq) {
1241 		iqidx = vi->first_rxq + (i % vi->nrxq);
1242 		snprintf(name, sizeof(name), "%s txq%d",
1243 		    device_get_nameunit(vi->dev), i);
1244 		init_eq(sc, &txq->eq, EQ_ETH, vi->qsize_txq, pi->tx_chan,
1245 		    sc->sge.rxq[iqidx].iq.cntxt_id, name);
1246 
1247 		rc = alloc_txq(vi, txq, i, oid);
1248 		if (rc != 0)
1249 			goto done;
1250 	}
1251 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1252 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_txq",
1253 	    CTLFLAG_RD, NULL, "tx queues for TOE/ETHOFLD");
1254 	for_each_ofld_txq(vi, i, ofld_txq) {
1255 		struct sysctl_oid *oid2;
1256 
1257 		snprintf(name, sizeof(name), "%s ofld_txq%d",
1258 		    device_get_nameunit(vi->dev), i);
1259 		if (vi->nofldrxq > 0) {
1260 			iqidx = vi->first_ofld_rxq + (i % vi->nofldrxq);
1261 			init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq,
1262 			    pi->tx_chan, sc->sge.ofld_rxq[iqidx].iq.cntxt_id,
1263 			    name);
1264 		} else {
1265 			iqidx = vi->first_rxq + (i % vi->nrxq);
1266 			init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq,
1267 			    pi->tx_chan, sc->sge.rxq[iqidx].iq.cntxt_id, name);
1268 		}
1269 
1270 		snprintf(name, sizeof(name), "%d", i);
1271 		oid2 = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO,
1272 		    name, CTLFLAG_RD, NULL, "offload tx queue");
1273 
1274 		rc = alloc_wrq(sc, vi, ofld_txq, oid2);
1275 		if (rc != 0)
1276 			goto done;
1277 	}
1278 #endif
1279 done:
1280 	if (rc)
1281 		t4_teardown_vi_queues(vi);
1282 
1283 	return (rc);
1284 }
1285 
1286 /*
1287  * Idempotent
1288  */
1289 int
1290 t4_teardown_vi_queues(struct vi_info *vi)
1291 {
1292 	int i;
1293 	struct sge_rxq *rxq;
1294 	struct sge_txq *txq;
1295 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1296 	struct port_info *pi = vi->pi;
1297 	struct adapter *sc = pi->adapter;
1298 	struct sge_wrq *ofld_txq;
1299 #endif
1300 #ifdef TCP_OFFLOAD
1301 	struct sge_ofld_rxq *ofld_rxq;
1302 #endif
1303 #ifdef DEV_NETMAP
1304 	struct sge_nm_rxq *nm_rxq;
1305 	struct sge_nm_txq *nm_txq;
1306 #endif
1307 
1308 	/* Do this before freeing the queues */
1309 	if (vi->flags & VI_SYSCTL_CTX) {
1310 		sysctl_ctx_free(&vi->ctx);
1311 		vi->flags &= ~VI_SYSCTL_CTX;
1312 	}
1313 
1314 #ifdef DEV_NETMAP
1315 	if (vi->ifp->if_capabilities & IFCAP_NETMAP) {
1316 		for_each_nm_txq(vi, i, nm_txq) {
1317 			free_nm_txq(vi, nm_txq);
1318 		}
1319 
1320 		for_each_nm_rxq(vi, i, nm_rxq) {
1321 			free_nm_rxq(vi, nm_rxq);
1322 		}
1323 	}
1324 #endif
1325 
1326 	/*
1327 	 * Take down all the tx queues first, as they reference the rx queues
1328 	 * (for egress updates, etc.).
1329 	 */
1330 
1331 	for_each_txq(vi, i, txq) {
1332 		free_txq(vi, txq);
1333 	}
1334 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
1335 	for_each_ofld_txq(vi, i, ofld_txq) {
1336 		free_wrq(sc, ofld_txq);
1337 	}
1338 #endif
1339 
1340 	/*
1341 	 * Then take down the rx queues.
1342 	 */
1343 
1344 	for_each_rxq(vi, i, rxq) {
1345 		free_rxq(vi, rxq);
1346 	}
1347 #ifdef TCP_OFFLOAD
1348 	for_each_ofld_rxq(vi, i, ofld_rxq) {
1349 		free_ofld_rxq(vi, ofld_rxq);
1350 	}
1351 #endif
1352 
1353 	return (0);
1354 }
1355 
1356 /*
1357  * Interrupt handler when the driver is using only 1 interrupt.  This is a very
1358  * unusual scenario.
1359  *
1360  * a) Deals with errors, if any.
1361  * b) Services firmware event queue, which is taking interrupts for all other
1362  *    queues.
1363  */
1364 void
1365 t4_intr_all(void *arg)
1366 {
1367 	struct adapter *sc = arg;
1368 	struct sge_iq *fwq = &sc->sge.fwq;
1369 
1370 	MPASS(sc->intr_count == 1);
1371 
1372 	if (sc->intr_type == INTR_INTX)
1373 		t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0);
1374 
1375 	t4_intr_err(arg);
1376 	t4_intr_evt(fwq);
1377 }
1378 
1379 /*
1380  * Interrupt handler for errors (installed directly when multiple interrupts are
1381  * being used, or called by t4_intr_all).
1382  */
1383 void
1384 t4_intr_err(void *arg)
1385 {
1386 	struct adapter *sc = arg;
1387 	uint32_t v;
1388 	const bool verbose = (sc->debug_flags & DF_VERBOSE_SLOWINTR) != 0;
1389 
1390 	if (sc->flags & ADAP_ERR)
1391 		return;
1392 
1393 	v = t4_read_reg(sc, MYPF_REG(A_PL_PF_INT_CAUSE));
1394 	if (v & F_PFSW) {
1395 		sc->swintr++;
1396 		t4_write_reg(sc, MYPF_REG(A_PL_PF_INT_CAUSE), v);
1397 	}
1398 
1399 	t4_slow_intr_handler(sc, verbose);
1400 }
1401 
1402 /*
1403  * Interrupt handler for iq-only queues.  The firmware event queue is the only
1404  * such queue right now.
1405  */
1406 void
1407 t4_intr_evt(void *arg)
1408 {
1409 	struct sge_iq *iq = arg;
1410 
1411 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1412 		service_iq(iq, 0);
1413 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1414 	}
1415 }
1416 
1417 /*
1418  * Interrupt handler for iq+fl queues.
1419  */
1420 void
1421 t4_intr(void *arg)
1422 {
1423 	struct sge_iq *iq = arg;
1424 
1425 	if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) {
1426 		service_iq_fl(iq, 0);
1427 		(void) atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE);
1428 	}
1429 }
1430 
1431 #ifdef DEV_NETMAP
1432 /*
1433  * Interrupt handler for netmap rx queues.
1434  */
1435 void
1436 t4_nm_intr(void *arg)
1437 {
1438 	struct sge_nm_rxq *nm_rxq = arg;
1439 
1440 	if (atomic_cmpset_int(&nm_rxq->nm_state, NM_ON, NM_BUSY)) {
1441 		service_nm_rxq(nm_rxq);
1442 		(void) atomic_cmpset_int(&nm_rxq->nm_state, NM_BUSY, NM_ON);
1443 	}
1444 }
1445 
1446 /*
1447  * Interrupt handler for vectors shared between NIC and netmap rx queues.
1448  */
1449 void
1450 t4_vi_intr(void *arg)
1451 {
1452 	struct irq *irq = arg;
1453 
1454 	MPASS(irq->nm_rxq != NULL);
1455 	t4_nm_intr(irq->nm_rxq);
1456 
1457 	MPASS(irq->rxq != NULL);
1458 	t4_intr(irq->rxq);
1459 }
1460 #endif
1461 
1462 /*
1463  * Deals with interrupts on an iq-only (no freelist) queue.
1464  */
1465 static int
1466 service_iq(struct sge_iq *iq, int budget)
1467 {
1468 	struct sge_iq *q;
1469 	struct adapter *sc = iq->adapter;
1470 	struct iq_desc *d = &iq->desc[iq->cidx];
1471 	int ndescs = 0, limit;
1472 	int rsp_type;
1473 	uint32_t lq;
1474 	STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql);
1475 
1476 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1477 	KASSERT((iq->flags & IQ_HAS_FL) == 0,
1478 	    ("%s: called for iq %p with fl (iq->flags 0x%x)", __func__, iq,
1479 	    iq->flags));
1480 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1481 	MPASS((iq->flags & IQ_LRO_ENABLED) == 0);
1482 
1483 	limit = budget ? budget : iq->qsize / 16;
1484 
1485 	/*
1486 	 * We always come back and check the descriptor ring for new indirect
1487 	 * interrupts and other responses after running a single handler.
1488 	 */
1489 	for (;;) {
1490 		while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1491 
1492 			rmb();
1493 
1494 			rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1495 			lq = be32toh(d->rsp.pldbuflen_qid);
1496 
1497 			switch (rsp_type) {
1498 			case X_RSPD_TYPE_FLBUF:
1499 				panic("%s: data for an iq (%p) with no freelist",
1500 				    __func__, iq);
1501 
1502 				/* NOTREACHED */
1503 
1504 			case X_RSPD_TYPE_CPL:
1505 				KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1506 				    ("%s: bad opcode %02x.", __func__,
1507 				    d->rss.opcode));
1508 				t4_cpl_handler[d->rss.opcode](iq, &d->rss, NULL);
1509 				break;
1510 
1511 			case X_RSPD_TYPE_INTR:
1512 				/*
1513 				 * There are 1K interrupt-capable queues (qids 0
1514 				 * through 1023).  A response type indicating a
1515 				 * forwarded interrupt with a qid >= 1K is an
1516 				 * iWARP async notification.
1517 				 */
1518 				if (__predict_true(lq >= 1024)) {
1519 					t4_an_handler(iq, &d->rsp);
1520 					break;
1521 				}
1522 
1523 				q = sc->sge.iqmap[lq - sc->sge.iq_start -
1524 				    sc->sge.iq_base];
1525 				if (atomic_cmpset_int(&q->state, IQS_IDLE,
1526 				    IQS_BUSY)) {
1527 					if (service_iq_fl(q, q->qsize / 16) == 0) {
1528 						(void) atomic_cmpset_int(&q->state,
1529 						    IQS_BUSY, IQS_IDLE);
1530 					} else {
1531 						STAILQ_INSERT_TAIL(&iql, q,
1532 						    link);
1533 					}
1534 				}
1535 				break;
1536 
1537 			default:
1538 				KASSERT(0,
1539 				    ("%s: illegal response type %d on iq %p",
1540 				    __func__, rsp_type, iq));
1541 				log(LOG_ERR,
1542 				    "%s: illegal response type %d on iq %p",
1543 				    device_get_nameunit(sc->dev), rsp_type, iq);
1544 				break;
1545 			}
1546 
1547 			d++;
1548 			if (__predict_false(++iq->cidx == iq->sidx)) {
1549 				iq->cidx = 0;
1550 				iq->gen ^= F_RSPD_GEN;
1551 				d = &iq->desc[0];
1552 			}
1553 			if (__predict_false(++ndescs == limit)) {
1554 				t4_write_reg(sc, sc->sge_gts_reg,
1555 				    V_CIDXINC(ndescs) |
1556 				    V_INGRESSQID(iq->cntxt_id) |
1557 				    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1558 				ndescs = 0;
1559 
1560 				if (budget) {
1561 					return (EINPROGRESS);
1562 				}
1563 			}
1564 		}
1565 
1566 		if (STAILQ_EMPTY(&iql))
1567 			break;
1568 
1569 		/*
1570 		 * Process the head only, and send it to the back of the list if
1571 		 * it's still not done.
1572 		 */
1573 		q = STAILQ_FIRST(&iql);
1574 		STAILQ_REMOVE_HEAD(&iql, link);
1575 		if (service_iq_fl(q, q->qsize / 8) == 0)
1576 			(void) atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE);
1577 		else
1578 			STAILQ_INSERT_TAIL(&iql, q, link);
1579 	}
1580 
1581 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1582 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1583 
1584 	return (0);
1585 }
1586 
1587 static inline int
1588 sort_before_lro(struct lro_ctrl *lro)
1589 {
1590 
1591 	return (lro->lro_mbuf_max != 0);
1592 }
1593 
1594 static inline uint64_t
1595 last_flit_to_ns(struct adapter *sc, uint64_t lf)
1596 {
1597 	uint64_t n = be64toh(lf) & 0xfffffffffffffff;	/* 60b, not 64b. */
1598 
1599 	if (n > UINT64_MAX / 1000000)
1600 		return (n / sc->params.vpd.cclk * 1000000);
1601 	else
1602 		return (n * 1000000 / sc->params.vpd.cclk);
1603 }
1604 
1605 /*
1606  * Deals with interrupts on an iq+fl queue.
1607  */
1608 static int
1609 service_iq_fl(struct sge_iq *iq, int budget)
1610 {
1611 	struct sge_rxq *rxq = iq_to_rxq(iq);
1612 	struct sge_fl *fl;
1613 	struct adapter *sc = iq->adapter;
1614 	struct iq_desc *d = &iq->desc[iq->cidx];
1615 	int ndescs = 0, limit;
1616 	int rsp_type, refill, starved;
1617 	uint32_t lq;
1618 	uint16_t fl_hw_cidx;
1619 	struct mbuf *m0;
1620 #if defined(INET) || defined(INET6)
1621 	const struct timeval lro_timeout = {0, sc->lro_timeout};
1622 	struct lro_ctrl *lro = &rxq->lro;
1623 #endif
1624 
1625 	KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq));
1626 	MPASS(iq->flags & IQ_HAS_FL);
1627 
1628 	limit = budget ? budget : iq->qsize / 16;
1629 	fl = &rxq->fl;
1630 	fl_hw_cidx = fl->hw_cidx;	/* stable snapshot */
1631 
1632 #if defined(INET) || defined(INET6)
1633 	if (iq->flags & IQ_ADJ_CREDIT) {
1634 		MPASS(sort_before_lro(lro));
1635 		iq->flags &= ~IQ_ADJ_CREDIT;
1636 		if ((d->rsp.u.type_gen & F_RSPD_GEN) != iq->gen) {
1637 			tcp_lro_flush_all(lro);
1638 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(1) |
1639 			    V_INGRESSQID((u32)iq->cntxt_id) |
1640 			    V_SEINTARM(iq->intr_params));
1641 			return (0);
1642 		}
1643 		ndescs = 1;
1644 	}
1645 #else
1646 	MPASS((iq->flags & IQ_ADJ_CREDIT) == 0);
1647 #endif
1648 
1649 	while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) {
1650 
1651 		rmb();
1652 
1653 		refill = 0;
1654 		m0 = NULL;
1655 		rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen);
1656 		lq = be32toh(d->rsp.pldbuflen_qid);
1657 
1658 		switch (rsp_type) {
1659 		case X_RSPD_TYPE_FLBUF:
1660 
1661 			m0 = get_fl_payload(sc, fl, lq);
1662 			if (__predict_false(m0 == NULL))
1663 				goto out;
1664 			refill = IDXDIFF(fl->hw_cidx, fl_hw_cidx, fl->sidx) > 2;
1665 
1666 			if (iq->flags & IQ_RX_TIMESTAMP) {
1667 				/*
1668 				 * Fill up rcv_tstmp but do not set M_TSTMP.
1669 				 * rcv_tstmp is not in the format that the
1670 				 * kernel expects and we don't want to mislead
1671 				 * it.  For now this is only for custom code
1672 				 * that knows how to interpret cxgbe's stamp.
1673 				 */
1674 				m0->m_pkthdr.rcv_tstmp =
1675 				    last_flit_to_ns(sc, d->rsp.u.last_flit);
1676 #ifdef notyet
1677 				m0->m_flags |= M_TSTMP;
1678 #endif
1679 			}
1680 
1681 			/* fall through */
1682 
1683 		case X_RSPD_TYPE_CPL:
1684 			KASSERT(d->rss.opcode < NUM_CPL_CMDS,
1685 			    ("%s: bad opcode %02x.", __func__, d->rss.opcode));
1686 			t4_cpl_handler[d->rss.opcode](iq, &d->rss, m0);
1687 			break;
1688 
1689 		case X_RSPD_TYPE_INTR:
1690 
1691 			/*
1692 			 * There are 1K interrupt-capable queues (qids 0
1693 			 * through 1023).  A response type indicating a
1694 			 * forwarded interrupt with a qid >= 1K is an
1695 			 * iWARP async notification.  That is the only
1696 			 * acceptable indirect interrupt on this queue.
1697 			 */
1698 			if (__predict_false(lq < 1024)) {
1699 				panic("%s: indirect interrupt on iq_fl %p "
1700 				    "with qid %u", __func__, iq, lq);
1701 			}
1702 
1703 			t4_an_handler(iq, &d->rsp);
1704 			break;
1705 
1706 		default:
1707 			KASSERT(0, ("%s: illegal response type %d on iq %p",
1708 			    __func__, rsp_type, iq));
1709 			log(LOG_ERR, "%s: illegal response type %d on iq %p",
1710 			    device_get_nameunit(sc->dev), rsp_type, iq);
1711 			break;
1712 		}
1713 
1714 		d++;
1715 		if (__predict_false(++iq->cidx == iq->sidx)) {
1716 			iq->cidx = 0;
1717 			iq->gen ^= F_RSPD_GEN;
1718 			d = &iq->desc[0];
1719 		}
1720 		if (__predict_false(++ndescs == limit)) {
1721 			t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1722 			    V_INGRESSQID(iq->cntxt_id) |
1723 			    V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX)));
1724 			ndescs = 0;
1725 
1726 #if defined(INET) || defined(INET6)
1727 			if (iq->flags & IQ_LRO_ENABLED &&
1728 			    !sort_before_lro(lro) &&
1729 			    sc->lro_timeout != 0) {
1730 				tcp_lro_flush_inactive(lro, &lro_timeout);
1731 			}
1732 #endif
1733 			if (budget) {
1734 				FL_LOCK(fl);
1735 				refill_fl(sc, fl, 32);
1736 				FL_UNLOCK(fl);
1737 
1738 				return (EINPROGRESS);
1739 			}
1740 		}
1741 		if (refill) {
1742 			FL_LOCK(fl);
1743 			refill_fl(sc, fl, 32);
1744 			FL_UNLOCK(fl);
1745 			fl_hw_cidx = fl->hw_cidx;
1746 		}
1747 	}
1748 out:
1749 #if defined(INET) || defined(INET6)
1750 	if (iq->flags & IQ_LRO_ENABLED) {
1751 		if (ndescs > 0 && lro->lro_mbuf_count > 8) {
1752 			MPASS(sort_before_lro(lro));
1753 			/* hold back one credit and don't flush LRO state */
1754 			iq->flags |= IQ_ADJ_CREDIT;
1755 			ndescs--;
1756 		} else {
1757 			tcp_lro_flush_all(lro);
1758 		}
1759 	}
1760 #endif
1761 
1762 	t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) |
1763 	    V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params));
1764 
1765 	FL_LOCK(fl);
1766 	starved = refill_fl(sc, fl, 64);
1767 	FL_UNLOCK(fl);
1768 	if (__predict_false(starved != 0))
1769 		add_fl_to_sfl(sc, fl);
1770 
1771 	return (0);
1772 }
1773 
1774 static inline int
1775 cl_has_metadata(struct sge_fl *fl, struct cluster_layout *cll)
1776 {
1777 	int rc = fl->flags & FL_BUF_PACKING || cll->region1 > 0;
1778 
1779 	if (rc)
1780 		MPASS(cll->region3 >= CL_METADATA_SIZE);
1781 
1782 	return (rc);
1783 }
1784 
1785 static inline struct cluster_metadata *
1786 cl_metadata(struct adapter *sc, struct sge_fl *fl, struct cluster_layout *cll,
1787     caddr_t cl)
1788 {
1789 
1790 	if (cl_has_metadata(fl, cll)) {
1791 		struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1792 
1793 		return ((struct cluster_metadata *)(cl + swz->size) - 1);
1794 	}
1795 	return (NULL);
1796 }
1797 
1798 static void
1799 rxb_free(struct mbuf *m)
1800 {
1801 	uma_zone_t zone = m->m_ext.ext_arg1;
1802 	void *cl = m->m_ext.ext_arg2;
1803 
1804 	uma_zfree(zone, cl);
1805 	counter_u64_add(extfree_rels, 1);
1806 }
1807 
1808 /*
1809  * The mbuf returned by this function could be allocated from zone_mbuf or
1810  * constructed in spare room in the cluster.
1811  *
1812  * The mbuf carries the payload in one of these ways
1813  * a) frame inside the mbuf (mbuf from zone_mbuf)
1814  * b) m_cljset (for clusters without metadata) zone_mbuf
1815  * c) m_extaddref (cluster with metadata) inline mbuf
1816  * d) m_extaddref (cluster with metadata) zone_mbuf
1817  */
1818 static struct mbuf *
1819 get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset,
1820     int remaining)
1821 {
1822 	struct mbuf *m;
1823 	struct fl_sdesc *sd = &fl->sdesc[fl->cidx];
1824 	struct cluster_layout *cll = &sd->cll;
1825 	struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx];
1826 	struct hw_buf_info *hwb = &sc->sge.hw_buf_info[cll->hwidx];
1827 	struct cluster_metadata *clm = cl_metadata(sc, fl, cll, sd->cl);
1828 	int len, blen;
1829 	caddr_t payload;
1830 
1831 	blen = hwb->size - fl->rx_offset;	/* max possible in this buf */
1832 	len = min(remaining, blen);
1833 	payload = sd->cl + cll->region1 + fl->rx_offset;
1834 	if (fl->flags & FL_BUF_PACKING) {
1835 		const u_int l = fr_offset + len;
1836 		const u_int pad = roundup2(l, fl->buf_boundary) - l;
1837 
1838 		if (fl->rx_offset + len + pad < hwb->size)
1839 			blen = len + pad;
1840 		MPASS(fl->rx_offset + blen <= hwb->size);
1841 	} else {
1842 		MPASS(fl->rx_offset == 0);	/* not packing */
1843 	}
1844 
1845 
1846 	if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) {
1847 
1848 		/*
1849 		 * Copy payload into a freshly allocated mbuf.
1850 		 */
1851 
1852 		m = fr_offset == 0 ?
1853 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1854 		if (m == NULL)
1855 			return (NULL);
1856 		fl->mbuf_allocated++;
1857 
1858 		/* copy data to mbuf */
1859 		bcopy(payload, mtod(m, caddr_t), len);
1860 
1861 	} else if (sd->nmbuf * MSIZE < cll->region1) {
1862 
1863 		/*
1864 		 * There's spare room in the cluster for an mbuf.  Create one
1865 		 * and associate it with the payload that's in the cluster.
1866 		 */
1867 
1868 		MPASS(clm != NULL);
1869 		m = (struct mbuf *)(sd->cl + sd->nmbuf * MSIZE);
1870 		/* No bzero required */
1871 		if (m_init(m, M_NOWAIT, MT_DATA,
1872 		    fr_offset == 0 ? M_PKTHDR | M_NOFREE : M_NOFREE))
1873 			return (NULL);
1874 		fl->mbuf_inlined++;
1875 		m_extaddref(m, payload, blen, &clm->refcount, rxb_free,
1876 		    swz->zone, sd->cl);
1877 		if (sd->nmbuf++ == 0)
1878 			counter_u64_add(extfree_refs, 1);
1879 
1880 	} else {
1881 
1882 		/*
1883 		 * Grab an mbuf from zone_mbuf and associate it with the
1884 		 * payload in the cluster.
1885 		 */
1886 
1887 		m = fr_offset == 0 ?
1888 		    m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA);
1889 		if (m == NULL)
1890 			return (NULL);
1891 		fl->mbuf_allocated++;
1892 		if (clm != NULL) {
1893 			m_extaddref(m, payload, blen, &clm->refcount,
1894 			    rxb_free, swz->zone, sd->cl);
1895 			if (sd->nmbuf++ == 0)
1896 				counter_u64_add(extfree_refs, 1);
1897 		} else {
1898 			m_cljset(m, sd->cl, swz->type);
1899 			sd->cl = NULL;	/* consumed, not a recycle candidate */
1900 		}
1901 	}
1902 	if (fr_offset == 0)
1903 		m->m_pkthdr.len = remaining;
1904 	m->m_len = len;
1905 
1906 	if (fl->flags & FL_BUF_PACKING) {
1907 		fl->rx_offset += blen;
1908 		MPASS(fl->rx_offset <= hwb->size);
1909 		if (fl->rx_offset < hwb->size)
1910 			return (m);	/* without advancing the cidx */
1911 	}
1912 
1913 	if (__predict_false(++fl->cidx % 8 == 0)) {
1914 		uint16_t cidx = fl->cidx / 8;
1915 
1916 		if (__predict_false(cidx == fl->sidx))
1917 			fl->cidx = cidx = 0;
1918 		fl->hw_cidx = cidx;
1919 	}
1920 	fl->rx_offset = 0;
1921 
1922 	return (m);
1923 }
1924 
1925 static struct mbuf *
1926 get_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf)
1927 {
1928 	struct mbuf *m0, *m, **pnext;
1929 	u_int remaining;
1930 	const u_int total = G_RSPD_LEN(len_newbuf);
1931 
1932 	if (__predict_false(fl->flags & FL_BUF_RESUME)) {
1933 		M_ASSERTPKTHDR(fl->m0);
1934 		MPASS(fl->m0->m_pkthdr.len == total);
1935 		MPASS(fl->remaining < total);
1936 
1937 		m0 = fl->m0;
1938 		pnext = fl->pnext;
1939 		remaining = fl->remaining;
1940 		fl->flags &= ~FL_BUF_RESUME;
1941 		goto get_segment;
1942 	}
1943 
1944 	if (fl->rx_offset > 0 && len_newbuf & F_RSPD_NEWBUF) {
1945 		fl->rx_offset = 0;
1946 		if (__predict_false(++fl->cidx % 8 == 0)) {
1947 			uint16_t cidx = fl->cidx / 8;
1948 
1949 			if (__predict_false(cidx == fl->sidx))
1950 				fl->cidx = cidx = 0;
1951 			fl->hw_cidx = cidx;
1952 		}
1953 	}
1954 
1955 	/*
1956 	 * Payload starts at rx_offset in the current hw buffer.  Its length is
1957 	 * 'len' and it may span multiple hw buffers.
1958 	 */
1959 
1960 	m0 = get_scatter_segment(sc, fl, 0, total);
1961 	if (m0 == NULL)
1962 		return (NULL);
1963 	remaining = total - m0->m_len;
1964 	pnext = &m0->m_next;
1965 	while (remaining > 0) {
1966 get_segment:
1967 		MPASS(fl->rx_offset == 0);
1968 		m = get_scatter_segment(sc, fl, total - remaining, remaining);
1969 		if (__predict_false(m == NULL)) {
1970 			fl->m0 = m0;
1971 			fl->pnext = pnext;
1972 			fl->remaining = remaining;
1973 			fl->flags |= FL_BUF_RESUME;
1974 			return (NULL);
1975 		}
1976 		*pnext = m;
1977 		pnext = &m->m_next;
1978 		remaining -= m->m_len;
1979 	}
1980 	*pnext = NULL;
1981 
1982 	M_ASSERTPKTHDR(m0);
1983 	return (m0);
1984 }
1985 
1986 static int
1987 t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
1988 {
1989 	struct sge_rxq *rxq = iq_to_rxq(iq);
1990 	struct ifnet *ifp = rxq->ifp;
1991 	struct adapter *sc = iq->adapter;
1992 	const struct cpl_rx_pkt *cpl = (const void *)(rss + 1);
1993 #if defined(INET) || defined(INET6)
1994 	struct lro_ctrl *lro = &rxq->lro;
1995 #endif
1996 	static const int sw_hashtype[4][2] = {
1997 		{M_HASHTYPE_NONE, M_HASHTYPE_NONE},
1998 		{M_HASHTYPE_RSS_IPV4, M_HASHTYPE_RSS_IPV6},
1999 		{M_HASHTYPE_RSS_TCP_IPV4, M_HASHTYPE_RSS_TCP_IPV6},
2000 		{M_HASHTYPE_RSS_UDP_IPV4, M_HASHTYPE_RSS_UDP_IPV6},
2001 	};
2002 
2003 	KASSERT(m0 != NULL, ("%s: no payload with opcode %02x", __func__,
2004 	    rss->opcode));
2005 
2006 	m0->m_pkthdr.len -= sc->params.sge.fl_pktshift;
2007 	m0->m_len -= sc->params.sge.fl_pktshift;
2008 	m0->m_data += sc->params.sge.fl_pktshift;
2009 
2010 	m0->m_pkthdr.rcvif = ifp;
2011 	M_HASHTYPE_SET(m0, sw_hashtype[rss->hash_type][rss->ipv6]);
2012 	m0->m_pkthdr.flowid = be32toh(rss->hash_val);
2013 
2014 	if (cpl->csum_calc && !(cpl->err_vec & sc->params.tp.err_vec_mask)) {
2015 		if (ifp->if_capenable & IFCAP_RXCSUM &&
2016 		    cpl->l2info & htobe32(F_RXF_IP)) {
2017 			m0->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
2018 			    CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
2019 			rxq->rxcsum++;
2020 		} else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
2021 		    cpl->l2info & htobe32(F_RXF_IP6)) {
2022 			m0->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
2023 			    CSUM_PSEUDO_HDR);
2024 			rxq->rxcsum++;
2025 		}
2026 
2027 		if (__predict_false(cpl->ip_frag))
2028 			m0->m_pkthdr.csum_data = be16toh(cpl->csum);
2029 		else
2030 			m0->m_pkthdr.csum_data = 0xffff;
2031 	}
2032 
2033 	if (cpl->vlan_ex) {
2034 		m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan);
2035 		m0->m_flags |= M_VLANTAG;
2036 		rxq->vlan_extraction++;
2037 	}
2038 
2039 #ifdef NUMA
2040 	m0->m_pkthdr.numa_domain = ifp->if_numa_domain;
2041 #endif
2042 #if defined(INET) || defined(INET6)
2043 	if (iq->flags & IQ_LRO_ENABLED) {
2044 		if (sort_before_lro(lro)) {
2045 			tcp_lro_queue_mbuf(lro, m0);
2046 			return (0); /* queued for sort, then LRO */
2047 		}
2048 		if (tcp_lro_rx(lro, m0, 0) == 0)
2049 			return (0); /* queued for LRO */
2050 	}
2051 #endif
2052 	ifp->if_input(ifp, m0);
2053 
2054 	return (0);
2055 }
2056 
2057 /*
2058  * Must drain the wrq or make sure that someone else will.
2059  */
2060 static void
2061 wrq_tx_drain(void *arg, int n)
2062 {
2063 	struct sge_wrq *wrq = arg;
2064 	struct sge_eq *eq = &wrq->eq;
2065 
2066 	EQ_LOCK(eq);
2067 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2068 		drain_wrq_wr_list(wrq->adapter, wrq);
2069 	EQ_UNLOCK(eq);
2070 }
2071 
2072 static void
2073 drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq)
2074 {
2075 	struct sge_eq *eq = &wrq->eq;
2076 	u_int available, dbdiff;	/* # of hardware descriptors */
2077 	u_int n;
2078 	struct wrqe *wr;
2079 	struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
2080 
2081 	EQ_LOCK_ASSERT_OWNED(eq);
2082 	MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs));
2083 	wr = STAILQ_FIRST(&wrq->wr_list);
2084 	MPASS(wr != NULL);	/* Must be called with something useful to do */
2085 	MPASS(eq->pidx == eq->dbidx);
2086 	dbdiff = 0;
2087 
2088 	do {
2089 		eq->cidx = read_hw_cidx(eq);
2090 		if (eq->pidx == eq->cidx)
2091 			available = eq->sidx - 1;
2092 		else
2093 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2094 
2095 		MPASS(wr->wrq == wrq);
2096 		n = howmany(wr->wr_len, EQ_ESIZE);
2097 		if (available < n)
2098 			break;
2099 
2100 		dst = (void *)&eq->desc[eq->pidx];
2101 		if (__predict_true(eq->sidx - eq->pidx > n)) {
2102 			/* Won't wrap, won't end exactly at the status page. */
2103 			bcopy(&wr->wr[0], dst, wr->wr_len);
2104 			eq->pidx += n;
2105 		} else {
2106 			int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE;
2107 
2108 			bcopy(&wr->wr[0], dst, first_portion);
2109 			if (wr->wr_len > first_portion) {
2110 				bcopy(&wr->wr[first_portion], &eq->desc[0],
2111 				    wr->wr_len - first_portion);
2112 			}
2113 			eq->pidx = n - (eq->sidx - eq->pidx);
2114 		}
2115 		wrq->tx_wrs_copied++;
2116 
2117 		if (available < eq->sidx / 4 &&
2118 		    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2119 				/*
2120 				 * XXX: This is not 100% reliable with some
2121 				 * types of WRs.  But this is a very unusual
2122 				 * situation for an ofld/ctrl queue anyway.
2123 				 */
2124 			dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2125 			    F_FW_WR_EQUEQ);
2126 		}
2127 
2128 		dbdiff += n;
2129 		if (dbdiff >= 16) {
2130 			ring_eq_db(sc, eq, dbdiff);
2131 			dbdiff = 0;
2132 		}
2133 
2134 		STAILQ_REMOVE_HEAD(&wrq->wr_list, link);
2135 		free_wrqe(wr);
2136 		MPASS(wrq->nwr_pending > 0);
2137 		wrq->nwr_pending--;
2138 		MPASS(wrq->ndesc_needed >= n);
2139 		wrq->ndesc_needed -= n;
2140 	} while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL);
2141 
2142 	if (dbdiff)
2143 		ring_eq_db(sc, eq, dbdiff);
2144 }
2145 
2146 /*
2147  * Doesn't fail.  Holds on to work requests it can't send right away.
2148  */
2149 void
2150 t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr)
2151 {
2152 #ifdef INVARIANTS
2153 	struct sge_eq *eq = &wrq->eq;
2154 #endif
2155 
2156 	EQ_LOCK_ASSERT_OWNED(eq);
2157 	MPASS(wr != NULL);
2158 	MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN);
2159 	MPASS((wr->wr_len & 0x7) == 0);
2160 
2161 	STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link);
2162 	wrq->nwr_pending++;
2163 	wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE);
2164 
2165 	if (!TAILQ_EMPTY(&wrq->incomplete_wrs))
2166 		return;	/* commit_wrq_wr will drain wr_list as well. */
2167 
2168 	drain_wrq_wr_list(sc, wrq);
2169 
2170 	/* Doorbell must have caught up to the pidx. */
2171 	MPASS(eq->pidx == eq->dbidx);
2172 }
2173 
2174 void
2175 t4_update_fl_bufsize(struct ifnet *ifp)
2176 {
2177 	struct vi_info *vi = ifp->if_softc;
2178 	struct adapter *sc = vi->pi->adapter;
2179 	struct sge_rxq *rxq;
2180 #ifdef TCP_OFFLOAD
2181 	struct sge_ofld_rxq *ofld_rxq;
2182 #endif
2183 	struct sge_fl *fl;
2184 	int i, maxp, mtu = ifp->if_mtu;
2185 
2186 	maxp = mtu_to_max_payload(sc, mtu);
2187 	for_each_rxq(vi, i, rxq) {
2188 		fl = &rxq->fl;
2189 
2190 		FL_LOCK(fl);
2191 		find_best_refill_source(sc, fl, maxp);
2192 		FL_UNLOCK(fl);
2193 	}
2194 #ifdef TCP_OFFLOAD
2195 	for_each_ofld_rxq(vi, i, ofld_rxq) {
2196 		fl = &ofld_rxq->fl;
2197 
2198 		FL_LOCK(fl);
2199 		find_best_refill_source(sc, fl, maxp);
2200 		FL_UNLOCK(fl);
2201 	}
2202 #endif
2203 }
2204 
2205 static inline int
2206 mbuf_nsegs(struct mbuf *m)
2207 {
2208 
2209 	M_ASSERTPKTHDR(m);
2210 	KASSERT(m->m_pkthdr.l5hlen > 0,
2211 	    ("%s: mbuf %p missing information on # of segments.", __func__, m));
2212 
2213 	return (m->m_pkthdr.l5hlen);
2214 }
2215 
2216 static inline void
2217 set_mbuf_nsegs(struct mbuf *m, uint8_t nsegs)
2218 {
2219 
2220 	M_ASSERTPKTHDR(m);
2221 	m->m_pkthdr.l5hlen = nsegs;
2222 }
2223 
2224 static inline int
2225 mbuf_cflags(struct mbuf *m)
2226 {
2227 
2228 	M_ASSERTPKTHDR(m);
2229 	return (m->m_pkthdr.PH_loc.eight[4]);
2230 }
2231 
2232 static inline void
2233 set_mbuf_cflags(struct mbuf *m, uint8_t flags)
2234 {
2235 
2236 	M_ASSERTPKTHDR(m);
2237 	m->m_pkthdr.PH_loc.eight[4] = flags;
2238 }
2239 
2240 static inline int
2241 mbuf_len16(struct mbuf *m)
2242 {
2243 	int n;
2244 
2245 	M_ASSERTPKTHDR(m);
2246 	n = m->m_pkthdr.PH_loc.eight[0];
2247 	if (!(mbuf_cflags(m) & MC_TLS))
2248 		MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2249 
2250 	return (n);
2251 }
2252 
2253 static inline void
2254 set_mbuf_len16(struct mbuf *m, uint8_t len16)
2255 {
2256 
2257 	M_ASSERTPKTHDR(m);
2258 	m->m_pkthdr.PH_loc.eight[0] = len16;
2259 }
2260 
2261 #ifdef RATELIMIT
2262 static inline int
2263 mbuf_eo_nsegs(struct mbuf *m)
2264 {
2265 
2266 	M_ASSERTPKTHDR(m);
2267 	return (m->m_pkthdr.PH_loc.eight[1]);
2268 }
2269 
2270 static inline void
2271 set_mbuf_eo_nsegs(struct mbuf *m, uint8_t nsegs)
2272 {
2273 
2274 	M_ASSERTPKTHDR(m);
2275 	m->m_pkthdr.PH_loc.eight[1] = nsegs;
2276 }
2277 
2278 static inline int
2279 mbuf_eo_len16(struct mbuf *m)
2280 {
2281 	int n;
2282 
2283 	M_ASSERTPKTHDR(m);
2284 	n = m->m_pkthdr.PH_loc.eight[2];
2285 	MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16);
2286 
2287 	return (n);
2288 }
2289 
2290 static inline void
2291 set_mbuf_eo_len16(struct mbuf *m, uint8_t len16)
2292 {
2293 
2294 	M_ASSERTPKTHDR(m);
2295 	m->m_pkthdr.PH_loc.eight[2] = len16;
2296 }
2297 
2298 static inline int
2299 mbuf_eo_tsclk_tsoff(struct mbuf *m)
2300 {
2301 
2302 	M_ASSERTPKTHDR(m);
2303 	return (m->m_pkthdr.PH_loc.eight[3]);
2304 }
2305 
2306 static inline void
2307 set_mbuf_eo_tsclk_tsoff(struct mbuf *m, uint8_t tsclk_tsoff)
2308 {
2309 
2310 	M_ASSERTPKTHDR(m);
2311 	m->m_pkthdr.PH_loc.eight[3] = tsclk_tsoff;
2312 }
2313 
2314 static inline int
2315 needs_eo(struct cxgbe_snd_tag *cst)
2316 {
2317 
2318 	return (cst != NULL && cst->type == IF_SND_TAG_TYPE_RATE_LIMIT);
2319 }
2320 #endif
2321 
2322 /*
2323  * Try to allocate an mbuf to contain a raw work request.  To make it
2324  * easy to construct the work request, don't allocate a chain but a
2325  * single mbuf.
2326  */
2327 struct mbuf *
2328 alloc_wr_mbuf(int len, int how)
2329 {
2330 	struct mbuf *m;
2331 
2332 	if (len <= MHLEN)
2333 		m = m_gethdr(how, MT_DATA);
2334 	else if (len <= MCLBYTES)
2335 		m = m_getcl(how, MT_DATA, M_PKTHDR);
2336 	else
2337 		m = NULL;
2338 	if (m == NULL)
2339 		return (NULL);
2340 	m->m_pkthdr.len = len;
2341 	m->m_len = len;
2342 	set_mbuf_cflags(m, MC_RAW_WR);
2343 	set_mbuf_len16(m, howmany(len, 16));
2344 	return (m);
2345 }
2346 
2347 static inline int
2348 needs_hwcsum(struct mbuf *m)
2349 {
2350 
2351 	M_ASSERTPKTHDR(m);
2352 
2353 	return (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_IP |
2354 	    CSUM_TSO | CSUM_UDP_IPV6 | CSUM_TCP_IPV6));
2355 }
2356 
2357 static inline int
2358 needs_tso(struct mbuf *m)
2359 {
2360 
2361 	M_ASSERTPKTHDR(m);
2362 
2363 	return (m->m_pkthdr.csum_flags & CSUM_TSO);
2364 }
2365 
2366 static inline int
2367 needs_l3_csum(struct mbuf *m)
2368 {
2369 
2370 	M_ASSERTPKTHDR(m);
2371 
2372 	return (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO));
2373 }
2374 
2375 static inline int
2376 needs_tcp_csum(struct mbuf *m)
2377 {
2378 
2379 	M_ASSERTPKTHDR(m);
2380 	return (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_TCP_IPV6 | CSUM_TSO));
2381 }
2382 
2383 #ifdef RATELIMIT
2384 static inline int
2385 needs_l4_csum(struct mbuf *m)
2386 {
2387 
2388 	M_ASSERTPKTHDR(m);
2389 
2390 	return (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
2391 	    CSUM_TCP_IPV6 | CSUM_TSO));
2392 }
2393 
2394 static inline int
2395 needs_udp_csum(struct mbuf *m)
2396 {
2397 
2398 	M_ASSERTPKTHDR(m);
2399 	return (m->m_pkthdr.csum_flags & (CSUM_UDP | CSUM_UDP_IPV6));
2400 }
2401 #endif
2402 
2403 static inline int
2404 needs_vlan_insertion(struct mbuf *m)
2405 {
2406 
2407 	M_ASSERTPKTHDR(m);
2408 
2409 	return (m->m_flags & M_VLANTAG);
2410 }
2411 
2412 static void *
2413 m_advance(struct mbuf **pm, int *poffset, int len)
2414 {
2415 	struct mbuf *m = *pm;
2416 	int offset = *poffset;
2417 	uintptr_t p = 0;
2418 
2419 	MPASS(len > 0);
2420 
2421 	for (;;) {
2422 		if (offset + len < m->m_len) {
2423 			offset += len;
2424 			p = mtod(m, uintptr_t) + offset;
2425 			break;
2426 		}
2427 		len -= m->m_len - offset;
2428 		m = m->m_next;
2429 		offset = 0;
2430 		MPASS(m != NULL);
2431 	}
2432 	*poffset = offset;
2433 	*pm = m;
2434 	return ((void *)p);
2435 }
2436 
2437 static inline int
2438 count_mbuf_ext_pgs(struct mbuf *m, int skip, vm_paddr_t *nextaddr)
2439 {
2440 	struct mbuf_ext_pgs *ext_pgs;
2441 	vm_paddr_t paddr;
2442 	int i, len, off, pglen, pgoff, seglen, segoff;
2443 	int nsegs = 0;
2444 
2445 	MBUF_EXT_PGS_ASSERT(m);
2446 	ext_pgs = m->m_ext.ext_pgs;
2447 	off = mtod(m, vm_offset_t);
2448 	len = m->m_len;
2449 	off += skip;
2450 	len -= skip;
2451 
2452 	if (ext_pgs->hdr_len != 0) {
2453 		if (off >= ext_pgs->hdr_len) {
2454 			off -= ext_pgs->hdr_len;
2455 		} else {
2456 			seglen = ext_pgs->hdr_len - off;
2457 			segoff = off;
2458 			seglen = min(seglen, len);
2459 			off = 0;
2460 			len -= seglen;
2461 			paddr = pmap_kextract(
2462 			    (vm_offset_t)&ext_pgs->hdr[segoff]);
2463 			if (*nextaddr != paddr)
2464 				nsegs++;
2465 			*nextaddr = paddr + seglen;
2466 		}
2467 	}
2468 	pgoff = ext_pgs->first_pg_off;
2469 	for (i = 0; i < ext_pgs->npgs && len > 0; i++) {
2470 		pglen = mbuf_ext_pg_len(ext_pgs, i, pgoff);
2471 		if (off >= pglen) {
2472 			off -= pglen;
2473 			pgoff = 0;
2474 			continue;
2475 		}
2476 		seglen = pglen - off;
2477 		segoff = pgoff + off;
2478 		off = 0;
2479 		seglen = min(seglen, len);
2480 		len -= seglen;
2481 		paddr = ext_pgs->pa[i] + segoff;
2482 		if (*nextaddr != paddr)
2483 			nsegs++;
2484 		*nextaddr = paddr + seglen;
2485 		pgoff = 0;
2486 	};
2487 	if (len != 0) {
2488 		seglen = min(len, ext_pgs->trail_len - off);
2489 		len -= seglen;
2490 		paddr = pmap_kextract((vm_offset_t)&ext_pgs->trail[off]);
2491 		if (*nextaddr != paddr)
2492 			nsegs++;
2493 		*nextaddr = paddr + seglen;
2494 	}
2495 
2496 	return (nsegs);
2497 }
2498 
2499 
2500 /*
2501  * Can deal with empty mbufs in the chain that have m_len = 0, but the chain
2502  * must have at least one mbuf that's not empty.  It is possible for this
2503  * routine to return 0 if skip accounts for all the contents of the mbuf chain.
2504  */
2505 static inline int
2506 count_mbuf_nsegs(struct mbuf *m, int skip, uint8_t *cflags)
2507 {
2508 	vm_paddr_t nextaddr, paddr;
2509 	vm_offset_t va;
2510 	int len, nsegs;
2511 
2512 	M_ASSERTPKTHDR(m);
2513 	MPASS(m->m_pkthdr.len > 0);
2514 	MPASS(m->m_pkthdr.len >= skip);
2515 
2516 	nsegs = 0;
2517 	nextaddr = 0;
2518 	for (; m; m = m->m_next) {
2519 		len = m->m_len;
2520 		if (__predict_false(len == 0))
2521 			continue;
2522 		if (skip >= len) {
2523 			skip -= len;
2524 			continue;
2525 		}
2526 		if ((m->m_flags & M_NOMAP) != 0) {
2527 			*cflags |= MC_NOMAP;
2528 			nsegs += count_mbuf_ext_pgs(m, skip, &nextaddr);
2529 			skip = 0;
2530 			continue;
2531 		}
2532 		va = mtod(m, vm_offset_t) + skip;
2533 		len -= skip;
2534 		skip = 0;
2535 		paddr = pmap_kextract(va);
2536 		nsegs += sglist_count((void *)(uintptr_t)va, len);
2537 		if (paddr == nextaddr)
2538 			nsegs--;
2539 		nextaddr = pmap_kextract(va + len - 1) + 1;
2540 	}
2541 
2542 	return (nsegs);
2543 }
2544 
2545 /*
2546  * Analyze the mbuf to determine its tx needs.  The mbuf passed in may change:
2547  * a) caller can assume it's been freed if this function returns with an error.
2548  * b) it may get defragged up if the gather list is too long for the hardware.
2549  */
2550 int
2551 parse_pkt(struct adapter *sc, struct mbuf **mp)
2552 {
2553 	struct mbuf *m0 = *mp, *m;
2554 	int rc, nsegs, defragged = 0, offset;
2555 	struct ether_header *eh;
2556 	void *l3hdr;
2557 #if defined(INET) || defined(INET6)
2558 	struct tcphdr *tcp;
2559 #endif
2560 #if defined(KERN_TLS) || defined(RATELIMIT)
2561 	struct cxgbe_snd_tag *cst;
2562 #endif
2563 	uint16_t eh_type;
2564 	uint8_t cflags;
2565 
2566 	cflags = 0;
2567 	M_ASSERTPKTHDR(m0);
2568 	if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) {
2569 		rc = EINVAL;
2570 fail:
2571 		m_freem(m0);
2572 		*mp = NULL;
2573 		return (rc);
2574 	}
2575 restart:
2576 	/*
2577 	 * First count the number of gather list segments in the payload.
2578 	 * Defrag the mbuf if nsegs exceeds the hardware limit.
2579 	 */
2580 	M_ASSERTPKTHDR(m0);
2581 	MPASS(m0->m_pkthdr.len > 0);
2582 	nsegs = count_mbuf_nsegs(m0, 0, &cflags);
2583 #if defined(KERN_TLS) || defined(RATELIMIT)
2584 	if (m0->m_pkthdr.csum_flags & CSUM_SND_TAG)
2585 		cst = mst_to_cst(m0->m_pkthdr.snd_tag);
2586 	else
2587 		cst = NULL;
2588 #endif
2589 #ifdef KERN_TLS
2590 	if (cst != NULL && cst->type == IF_SND_TAG_TYPE_TLS) {
2591 		int len16;
2592 
2593 		cflags |= MC_TLS;
2594 		set_mbuf_cflags(m0, cflags);
2595 		rc = t6_ktls_parse_pkt(m0, &nsegs, &len16);
2596 		if (rc != 0)
2597 			goto fail;
2598 		set_mbuf_nsegs(m0, nsegs);
2599 		set_mbuf_len16(m0, len16);
2600 		return (0);
2601 	}
2602 #endif
2603 	if (nsegs > (needs_tso(m0) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS)) {
2604 		if (defragged++ > 0 || (m = m_defrag(m0, M_NOWAIT)) == NULL) {
2605 			rc = EFBIG;
2606 			goto fail;
2607 		}
2608 		*mp = m0 = m;	/* update caller's copy after defrag */
2609 		goto restart;
2610 	}
2611 
2612 	if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN &&
2613 	    !(cflags & MC_NOMAP))) {
2614 		m0 = m_pullup(m0, m0->m_pkthdr.len);
2615 		if (m0 == NULL) {
2616 			/* Should have left well enough alone. */
2617 			rc = EFBIG;
2618 			goto fail;
2619 		}
2620 		*mp = m0;	/* update caller's copy after pullup */
2621 		goto restart;
2622 	}
2623 	set_mbuf_nsegs(m0, nsegs);
2624 	set_mbuf_cflags(m0, cflags);
2625 	if (sc->flags & IS_VF)
2626 		set_mbuf_len16(m0, txpkt_vm_len16(nsegs, needs_tso(m0)));
2627 	else
2628 		set_mbuf_len16(m0, txpkt_len16(nsegs, needs_tso(m0)));
2629 
2630 #ifdef RATELIMIT
2631 	/*
2632 	 * Ethofld is limited to TCP and UDP for now, and only when L4 hw
2633 	 * checksumming is enabled.  needs_l4_csum happens to check for all the
2634 	 * right things.
2635 	 */
2636 	if (__predict_false(needs_eo(cst) && !needs_l4_csum(m0))) {
2637 		m_snd_tag_rele(m0->m_pkthdr.snd_tag);
2638 		m0->m_pkthdr.snd_tag = NULL;
2639 		m0->m_pkthdr.csum_flags &= ~CSUM_SND_TAG;
2640 		cst = NULL;
2641 	}
2642 #endif
2643 
2644 	if (!needs_hwcsum(m0)
2645 #ifdef RATELIMIT
2646    		 && !needs_eo(cst)
2647 #endif
2648 	)
2649 		return (0);
2650 
2651 	m = m0;
2652 	eh = mtod(m, struct ether_header *);
2653 	eh_type = ntohs(eh->ether_type);
2654 	if (eh_type == ETHERTYPE_VLAN) {
2655 		struct ether_vlan_header *evh = (void *)eh;
2656 
2657 		eh_type = ntohs(evh->evl_proto);
2658 		m0->m_pkthdr.l2hlen = sizeof(*evh);
2659 	} else
2660 		m0->m_pkthdr.l2hlen = sizeof(*eh);
2661 
2662 	offset = 0;
2663 	l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2664 
2665 	switch (eh_type) {
2666 #ifdef INET6
2667 	case ETHERTYPE_IPV6:
2668 	{
2669 		struct ip6_hdr *ip6 = l3hdr;
2670 
2671 		MPASS(!needs_tso(m0) || ip6->ip6_nxt == IPPROTO_TCP);
2672 
2673 		m0->m_pkthdr.l3hlen = sizeof(*ip6);
2674 		break;
2675 	}
2676 #endif
2677 #ifdef INET
2678 	case ETHERTYPE_IP:
2679 	{
2680 		struct ip *ip = l3hdr;
2681 
2682 		m0->m_pkthdr.l3hlen = ip->ip_hl * 4;
2683 		break;
2684 	}
2685 #endif
2686 	default:
2687 		panic("%s: ethertype 0x%04x unknown.  if_cxgbe must be compiled"
2688 		    " with the same INET/INET6 options as the kernel.",
2689 		    __func__, eh_type);
2690 	}
2691 
2692 #if defined(INET) || defined(INET6)
2693 	if (needs_tcp_csum(m0)) {
2694 		tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen);
2695 		m0->m_pkthdr.l4hlen = tcp->th_off * 4;
2696 #ifdef RATELIMIT
2697 		if (tsclk >= 0 && *(uint32_t *)(tcp + 1) == ntohl(0x0101080a)) {
2698 			set_mbuf_eo_tsclk_tsoff(m0,
2699 			    V_FW_ETH_TX_EO_WR_TSCLK(tsclk) |
2700 			    V_FW_ETH_TX_EO_WR_TSOFF(sizeof(*tcp) / 2 + 1));
2701 		} else
2702 			set_mbuf_eo_tsclk_tsoff(m0, 0);
2703 	} else if (needs_udp_csum(m0)) {
2704 		m0->m_pkthdr.l4hlen = sizeof(struct udphdr);
2705 #endif
2706 	}
2707 #ifdef RATELIMIT
2708 	if (needs_eo(cst)) {
2709 		u_int immhdrs;
2710 
2711 		/* EO WRs have the headers in the WR and not the GL. */
2712 		immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen +
2713 		    m0->m_pkthdr.l4hlen;
2714 		cflags = 0;
2715 		nsegs = count_mbuf_nsegs(m0, immhdrs, &cflags);
2716 		MPASS(cflags == mbuf_cflags(m0));
2717 		set_mbuf_eo_nsegs(m0, nsegs);
2718 		set_mbuf_eo_len16(m0,
2719 		    txpkt_eo_len16(nsegs, immhdrs, needs_tso(m0)));
2720 	}
2721 #endif
2722 #endif
2723 	MPASS(m0 == *mp);
2724 	return (0);
2725 }
2726 
2727 void *
2728 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie)
2729 {
2730 	struct sge_eq *eq = &wrq->eq;
2731 	struct adapter *sc = wrq->adapter;
2732 	int ndesc, available;
2733 	struct wrqe *wr;
2734 	void *w;
2735 
2736 	MPASS(len16 > 0);
2737 	ndesc = howmany(len16, EQ_ESIZE / 16);
2738 	MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC);
2739 
2740 	EQ_LOCK(eq);
2741 
2742 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2743 		drain_wrq_wr_list(sc, wrq);
2744 
2745 	if (!STAILQ_EMPTY(&wrq->wr_list)) {
2746 slowpath:
2747 		EQ_UNLOCK(eq);
2748 		wr = alloc_wrqe(len16 * 16, wrq);
2749 		if (__predict_false(wr == NULL))
2750 			return (NULL);
2751 		cookie->pidx = -1;
2752 		cookie->ndesc = ndesc;
2753 		return (&wr->wr);
2754 	}
2755 
2756 	eq->cidx = read_hw_cidx(eq);
2757 	if (eq->pidx == eq->cidx)
2758 		available = eq->sidx - 1;
2759 	else
2760 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2761 	if (available < ndesc)
2762 		goto slowpath;
2763 
2764 	cookie->pidx = eq->pidx;
2765 	cookie->ndesc = ndesc;
2766 	TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link);
2767 
2768 	w = &eq->desc[eq->pidx];
2769 	IDXINCR(eq->pidx, ndesc, eq->sidx);
2770 	if (__predict_false(cookie->pidx + ndesc > eq->sidx)) {
2771 		w = &wrq->ss[0];
2772 		wrq->ss_pidx = cookie->pidx;
2773 		wrq->ss_len = len16 * 16;
2774 	}
2775 
2776 	EQ_UNLOCK(eq);
2777 
2778 	return (w);
2779 }
2780 
2781 void
2782 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie)
2783 {
2784 	struct sge_eq *eq = &wrq->eq;
2785 	struct adapter *sc = wrq->adapter;
2786 	int ndesc, pidx;
2787 	struct wrq_cookie *prev, *next;
2788 
2789 	if (cookie->pidx == -1) {
2790 		struct wrqe *wr = __containerof(w, struct wrqe, wr);
2791 
2792 		t4_wrq_tx(sc, wr);
2793 		return;
2794 	}
2795 
2796 	if (__predict_false(w == &wrq->ss[0])) {
2797 		int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE;
2798 
2799 		MPASS(wrq->ss_len > n);	/* WR had better wrap around. */
2800 		bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n);
2801 		bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n);
2802 		wrq->tx_wrs_ss++;
2803 	} else
2804 		wrq->tx_wrs_direct++;
2805 
2806 	EQ_LOCK(eq);
2807 	ndesc = cookie->ndesc;	/* Can be more than SGE_MAX_WR_NDESC here. */
2808 	pidx = cookie->pidx;
2809 	MPASS(pidx >= 0 && pidx < eq->sidx);
2810 	prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link);
2811 	next = TAILQ_NEXT(cookie, link);
2812 	if (prev == NULL) {
2813 		MPASS(pidx == eq->dbidx);
2814 		if (next == NULL || ndesc >= 16) {
2815 			int available;
2816 			struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
2817 
2818 			/*
2819 			 * Note that the WR via which we'll request tx updates
2820 			 * is at pidx and not eq->pidx, which has moved on
2821 			 * already.
2822 			 */
2823 			dst = (void *)&eq->desc[pidx];
2824 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2825 			if (available < eq->sidx / 4 &&
2826 			    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2827 				/*
2828 				 * XXX: This is not 100% reliable with some
2829 				 * types of WRs.  But this is a very unusual
2830 				 * situation for an ofld/ctrl queue anyway.
2831 				 */
2832 				dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2833 				    F_FW_WR_EQUEQ);
2834 			}
2835 
2836 			ring_eq_db(wrq->adapter, eq, ndesc);
2837 		} else {
2838 			MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc);
2839 			next->pidx = pidx;
2840 			next->ndesc += ndesc;
2841 		}
2842 	} else {
2843 		MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc);
2844 		prev->ndesc += ndesc;
2845 	}
2846 	TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link);
2847 
2848 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2849 		drain_wrq_wr_list(sc, wrq);
2850 
2851 #ifdef INVARIANTS
2852 	if (TAILQ_EMPTY(&wrq->incomplete_wrs)) {
2853 		/* Doorbell must have caught up to the pidx. */
2854 		MPASS(wrq->eq.pidx == wrq->eq.dbidx);
2855 	}
2856 #endif
2857 	EQ_UNLOCK(eq);
2858 }
2859 
2860 static u_int
2861 can_resume_eth_tx(struct mp_ring *r)
2862 {
2863 	struct sge_eq *eq = r->cookie;
2864 
2865 	return (total_available_tx_desc(eq) > eq->sidx / 8);
2866 }
2867 
2868 static inline int
2869 cannot_use_txpkts(struct mbuf *m)
2870 {
2871 	/* maybe put a GL limit too, to avoid silliness? */
2872 
2873 	return (needs_tso(m) || (mbuf_cflags(m) & (MC_RAW_WR | MC_TLS)) != 0);
2874 }
2875 
2876 static inline int
2877 discard_tx(struct sge_eq *eq)
2878 {
2879 
2880 	return ((eq->flags & (EQ_ENABLED | EQ_QFLUSH)) != EQ_ENABLED);
2881 }
2882 
2883 static inline int
2884 wr_can_update_eq(struct fw_eth_tx_pkts_wr *wr)
2885 {
2886 
2887 	switch (G_FW_WR_OP(be32toh(wr->op_pkd))) {
2888 	case FW_ULPTX_WR:
2889 	case FW_ETH_TX_PKT_WR:
2890 	case FW_ETH_TX_PKTS_WR:
2891 	case FW_ETH_TX_PKTS2_WR:
2892 	case FW_ETH_TX_PKT_VM_WR:
2893 		return (1);
2894 	default:
2895 		return (0);
2896 	}
2897 }
2898 
2899 /*
2900  * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to
2901  * be consumed.  Return the actual number consumed.  0 indicates a stall.
2902  */
2903 static u_int
2904 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx)
2905 {
2906 	struct sge_txq *txq = r->cookie;
2907 	struct sge_eq *eq = &txq->eq;
2908 	struct ifnet *ifp = txq->ifp;
2909 	struct vi_info *vi = ifp->if_softc;
2910 	struct port_info *pi = vi->pi;
2911 	struct adapter *sc = pi->adapter;
2912 	u_int total, remaining;		/* # of packets */
2913 	u_int available, dbdiff;	/* # of hardware descriptors */
2914 	u_int n, next_cidx;
2915 	struct mbuf *m0, *tail;
2916 	struct txpkts txp;
2917 	struct fw_eth_tx_pkts_wr *wr;	/* any fw WR struct will do */
2918 
2919 	remaining = IDXDIFF(pidx, cidx, r->size);
2920 	MPASS(remaining > 0);	/* Must not be called without work to do. */
2921 	total = 0;
2922 
2923 	TXQ_LOCK(txq);
2924 	if (__predict_false(discard_tx(eq))) {
2925 		while (cidx != pidx) {
2926 			m0 = r->items[cidx];
2927 			m_freem(m0);
2928 			if (++cidx == r->size)
2929 				cidx = 0;
2930 		}
2931 		reclaim_tx_descs(txq, 2048);
2932 		total = remaining;
2933 		goto done;
2934 	}
2935 
2936 	/* How many hardware descriptors do we have readily available. */
2937 	if (eq->pidx == eq->cidx)
2938 		available = eq->sidx - 1;
2939 	else
2940 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2941 	dbdiff = IDXDIFF(eq->pidx, eq->dbidx, eq->sidx);
2942 
2943 	while (remaining > 0) {
2944 
2945 		m0 = r->items[cidx];
2946 		M_ASSERTPKTHDR(m0);
2947 		MPASS(m0->m_nextpkt == NULL);
2948 
2949 		if (available < howmany(mbuf_len16(m0), EQ_ESIZE / 16)) {
2950 			MPASS(howmany(mbuf_len16(m0), EQ_ESIZE / 16) <= 64);
2951 			available += reclaim_tx_descs(txq, 64);
2952 			if (available < howmany(mbuf_len16(m0), EQ_ESIZE / 16))
2953 				break;	/* out of descriptors */
2954 		}
2955 
2956 		next_cidx = cidx + 1;
2957 		if (__predict_false(next_cidx == r->size))
2958 			next_cidx = 0;
2959 
2960 		wr = (void *)&eq->desc[eq->pidx];
2961 		if (mbuf_cflags(m0) & MC_RAW_WR) {
2962 			total++;
2963 			remaining--;
2964 			n = write_raw_wr(txq, (void *)wr, m0, available);
2965 #ifdef KERN_TLS
2966 		} else if (mbuf_cflags(m0) & MC_TLS) {
2967 			total++;
2968 			remaining--;
2969 			ETHER_BPF_MTAP(ifp, m0);
2970 			n = t6_ktls_write_wr(txq,(void *)wr, m0,
2971 			    mbuf_nsegs(m0), available);
2972 #endif
2973 		} else if (sc->flags & IS_VF) {
2974 			total++;
2975 			remaining--;
2976 			ETHER_BPF_MTAP(ifp, m0);
2977 			n = write_txpkt_vm_wr(sc, txq, (void *)wr, m0,
2978 			    available);
2979 		} else if (remaining > 1 &&
2980 		    try_txpkts(m0, r->items[next_cidx], &txp, available) == 0) {
2981 
2982 			/* pkts at cidx, next_cidx should both be in txp. */
2983 			MPASS(txp.npkt == 2);
2984 			tail = r->items[next_cidx];
2985 			MPASS(tail->m_nextpkt == NULL);
2986 			ETHER_BPF_MTAP(ifp, m0);
2987 			ETHER_BPF_MTAP(ifp, tail);
2988 			m0->m_nextpkt = tail;
2989 
2990 			if (__predict_false(++next_cidx == r->size))
2991 				next_cidx = 0;
2992 
2993 			while (next_cidx != pidx) {
2994 				if (add_to_txpkts(r->items[next_cidx], &txp,
2995 				    available) != 0)
2996 					break;
2997 				tail->m_nextpkt = r->items[next_cidx];
2998 				tail = tail->m_nextpkt;
2999 				ETHER_BPF_MTAP(ifp, tail);
3000 				if (__predict_false(++next_cidx == r->size))
3001 					next_cidx = 0;
3002 			}
3003 
3004 			n = write_txpkts_wr(sc, txq, wr, m0, &txp, available);
3005 			total += txp.npkt;
3006 			remaining -= txp.npkt;
3007 		} else {
3008 			total++;
3009 			remaining--;
3010 			ETHER_BPF_MTAP(ifp, m0);
3011 			n = write_txpkt_wr(sc, txq, (void *)wr, m0, available);
3012 		}
3013 		MPASS(n >= 1 && n <= available);
3014 		if (!(mbuf_cflags(m0) & MC_TLS))
3015 			MPASS(n <= SGE_MAX_WR_NDESC);
3016 
3017 		available -= n;
3018 		dbdiff += n;
3019 		IDXINCR(eq->pidx, n, eq->sidx);
3020 
3021 		if (wr_can_update_eq(wr)) {
3022 			if (total_available_tx_desc(eq) < eq->sidx / 4 &&
3023 			    atomic_cmpset_int(&eq->equiq, 0, 1)) {
3024 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
3025 				    F_FW_WR_EQUEQ);
3026 				eq->equeqidx = eq->pidx;
3027 			} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >=
3028 			    32) {
3029 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
3030 				eq->equeqidx = eq->pidx;
3031 			}
3032 		}
3033 
3034 		if (dbdiff >= 16 && remaining >= 4) {
3035 			ring_eq_db(sc, eq, dbdiff);
3036 			available += reclaim_tx_descs(txq, 4 * dbdiff);
3037 			dbdiff = 0;
3038 		}
3039 
3040 		cidx = next_cidx;
3041 	}
3042 	if (dbdiff != 0) {
3043 		ring_eq_db(sc, eq, dbdiff);
3044 		reclaim_tx_descs(txq, 32);
3045 	}
3046 done:
3047 	TXQ_UNLOCK(txq);
3048 
3049 	return (total);
3050 }
3051 
3052 static inline void
3053 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
3054     int qsize)
3055 {
3056 
3057 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
3058 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
3059 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
3060 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
3061 
3062 	iq->flags = 0;
3063 	iq->adapter = sc;
3064 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
3065 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
3066 	if (pktc_idx >= 0) {
3067 		iq->intr_params |= F_QINTR_CNT_EN;
3068 		iq->intr_pktc_idx = pktc_idx;
3069 	}
3070 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
3071 	iq->sidx = iq->qsize - sc->params.sge.spg_len / IQ_ESIZE;
3072 }
3073 
3074 static inline void
3075 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name)
3076 {
3077 
3078 	fl->qsize = qsize;
3079 	fl->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
3080 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
3081 	if (sc->flags & BUF_PACKING_OK &&
3082 	    ((!is_t4(sc) && buffer_packing) ||	/* T5+: enabled unless 0 */
3083 	    (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */
3084 		fl->flags |= FL_BUF_PACKING;
3085 	find_best_refill_source(sc, fl, maxp);
3086 	find_safe_refill_source(sc, fl);
3087 }
3088 
3089 static inline void
3090 init_eq(struct adapter *sc, struct sge_eq *eq, int eqtype, int qsize,
3091     uint8_t tx_chan, uint16_t iqid, char *name)
3092 {
3093 	KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
3094 
3095 	eq->flags = eqtype & EQ_TYPEMASK;
3096 	eq->tx_chan = tx_chan;
3097 	eq->iqid = iqid;
3098 	eq->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
3099 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
3100 }
3101 
3102 static int
3103 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
3104     bus_dmamap_t *map, bus_addr_t *pa, void **va)
3105 {
3106 	int rc;
3107 
3108 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
3109 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
3110 	if (rc != 0) {
3111 		device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
3112 		goto done;
3113 	}
3114 
3115 	rc = bus_dmamem_alloc(*tag, va,
3116 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
3117 	if (rc != 0) {
3118 		device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
3119 		goto done;
3120 	}
3121 
3122 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
3123 	if (rc != 0) {
3124 		device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
3125 		goto done;
3126 	}
3127 done:
3128 	if (rc)
3129 		free_ring(sc, *tag, *map, *pa, *va);
3130 
3131 	return (rc);
3132 }
3133 
3134 static int
3135 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
3136     bus_addr_t pa, void *va)
3137 {
3138 	if (pa)
3139 		bus_dmamap_unload(tag, map);
3140 	if (va)
3141 		bus_dmamem_free(tag, va, map);
3142 	if (tag)
3143 		bus_dma_tag_destroy(tag);
3144 
3145 	return (0);
3146 }
3147 
3148 /*
3149  * Allocates the ring for an ingress queue and an optional freelist.  If the
3150  * freelist is specified it will be allocated and then associated with the
3151  * ingress queue.
3152  *
3153  * Returns errno on failure.  Resources allocated up to that point may still be
3154  * allocated.  Caller is responsible for cleanup in case this function fails.
3155  *
3156  * If the ingress queue will take interrupts directly then the intr_idx
3157  * specifies the vector, starting from 0.  -1 means the interrupts for this
3158  * queue should be forwarded to the fwq.
3159  */
3160 static int
3161 alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl,
3162     int intr_idx, int cong)
3163 {
3164 	int rc, i, cntxt_id;
3165 	size_t len;
3166 	struct fw_iq_cmd c;
3167 	struct port_info *pi = vi->pi;
3168 	struct adapter *sc = iq->adapter;
3169 	struct sge_params *sp = &sc->params.sge;
3170 	__be32 v = 0;
3171 
3172 	len = iq->qsize * IQ_ESIZE;
3173 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
3174 	    (void **)&iq->desc);
3175 	if (rc != 0)
3176 		return (rc);
3177 
3178 	bzero(&c, sizeof(c));
3179 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
3180 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
3181 	    V_FW_IQ_CMD_VFN(0));
3182 
3183 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
3184 	    FW_LEN16(c));
3185 
3186 	/* Special handling for firmware event queue */
3187 	if (iq == &sc->sge.fwq)
3188 		v |= F_FW_IQ_CMD_IQASYNCH;
3189 
3190 	if (intr_idx < 0) {
3191 		/* Forwarded interrupts, all headed to fwq */
3192 		v |= F_FW_IQ_CMD_IQANDST;
3193 		v |= V_FW_IQ_CMD_IQANDSTINDEX(sc->sge.fwq.cntxt_id);
3194 	} else {
3195 		KASSERT(intr_idx < sc->intr_count,
3196 		    ("%s: invalid direct intr_idx %d", __func__, intr_idx));
3197 		v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
3198 	}
3199 
3200 	c.type_to_iqandstindex = htobe32(v |
3201 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
3202 	    V_FW_IQ_CMD_VIID(vi->viid) |
3203 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
3204 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
3205 	    F_FW_IQ_CMD_IQGTSMODE |
3206 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
3207 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
3208 	c.iqsize = htobe16(iq->qsize);
3209 	c.iqaddr = htobe64(iq->ba);
3210 	if (cong >= 0)
3211 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
3212 
3213 	if (fl) {
3214 		mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
3215 
3216 		len = fl->qsize * EQ_ESIZE;
3217 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
3218 		    &fl->ba, (void **)&fl->desc);
3219 		if (rc)
3220 			return (rc);
3221 
3222 		/* Allocate space for one software descriptor per buffer. */
3223 		rc = alloc_fl_sdesc(fl);
3224 		if (rc != 0) {
3225 			device_printf(sc->dev,
3226 			    "failed to setup fl software descriptors: %d\n",
3227 			    rc);
3228 			return (rc);
3229 		}
3230 
3231 		if (fl->flags & FL_BUF_PACKING) {
3232 			fl->lowat = roundup2(sp->fl_starve_threshold2, 8);
3233 			fl->buf_boundary = sp->pack_boundary;
3234 		} else {
3235 			fl->lowat = roundup2(sp->fl_starve_threshold, 8);
3236 			fl->buf_boundary = 16;
3237 		}
3238 		if (fl_pad && fl->buf_boundary < sp->pad_boundary)
3239 			fl->buf_boundary = sp->pad_boundary;
3240 
3241 		c.iqns_to_fl0congen |=
3242 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
3243 			F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
3244 			(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
3245 			(fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
3246 			    0));
3247 		if (cong >= 0) {
3248 			c.iqns_to_fl0congen |=
3249 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
3250 				    F_FW_IQ_CMD_FL0CONGCIF |
3251 				    F_FW_IQ_CMD_FL0CONGEN);
3252 		}
3253 		c.fl0dcaen_to_fl0cidxfthresh =
3254 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
3255 			X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B_T6) |
3256 			V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
3257 			X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
3258 		c.fl0size = htobe16(fl->qsize);
3259 		c.fl0addr = htobe64(fl->ba);
3260 	}
3261 
3262 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3263 	if (rc != 0) {
3264 		device_printf(sc->dev,
3265 		    "failed to create ingress queue: %d\n", rc);
3266 		return (rc);
3267 	}
3268 
3269 	iq->cidx = 0;
3270 	iq->gen = F_RSPD_GEN;
3271 	iq->intr_next = iq->intr_params;
3272 	iq->cntxt_id = be16toh(c.iqid);
3273 	iq->abs_id = be16toh(c.physiqid);
3274 	iq->flags |= IQ_ALLOCATED;
3275 
3276 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
3277 	if (cntxt_id >= sc->sge.niq) {
3278 		panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
3279 		    cntxt_id, sc->sge.niq - 1);
3280 	}
3281 	sc->sge.iqmap[cntxt_id] = iq;
3282 
3283 	if (fl) {
3284 		u_int qid;
3285 
3286 		iq->flags |= IQ_HAS_FL;
3287 		fl->cntxt_id = be16toh(c.fl0id);
3288 		fl->pidx = fl->cidx = 0;
3289 
3290 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
3291 		if (cntxt_id >= sc->sge.neq) {
3292 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
3293 			    __func__, cntxt_id, sc->sge.neq - 1);
3294 		}
3295 		sc->sge.eqmap[cntxt_id] = (void *)fl;
3296 
3297 		qid = fl->cntxt_id;
3298 		if (isset(&sc->doorbells, DOORBELL_UDB)) {
3299 			uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3300 			uint32_t mask = (1 << s_qpp) - 1;
3301 			volatile uint8_t *udb;
3302 
3303 			udb = sc->udbs_base + UDBS_DB_OFFSET;
3304 			udb += (qid >> s_qpp) << PAGE_SHIFT;
3305 			qid &= mask;
3306 			if (qid < PAGE_SIZE / UDBS_SEG_SIZE) {
3307 				udb += qid << UDBS_SEG_SHIFT;
3308 				qid = 0;
3309 			}
3310 			fl->udb = (volatile void *)udb;
3311 		}
3312 		fl->dbval = V_QID(qid) | sc->chip_params->sge_fl_db;
3313 
3314 		FL_LOCK(fl);
3315 		/* Enough to make sure the SGE doesn't think it's starved */
3316 		refill_fl(sc, fl, fl->lowat);
3317 		FL_UNLOCK(fl);
3318 	}
3319 
3320 	if (chip_id(sc) >= CHELSIO_T5 && !(sc->flags & IS_VF) && cong >= 0) {
3321 		uint32_t param, val;
3322 
3323 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
3324 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
3325 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
3326 		if (cong == 0)
3327 			val = 1 << 19;
3328 		else {
3329 			val = 2 << 19;
3330 			for (i = 0; i < 4; i++) {
3331 				if (cong & (1 << i))
3332 					val |= 1 << (i << 2);
3333 			}
3334 		}
3335 
3336 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
3337 		if (rc != 0) {
3338 			/* report error but carry on */
3339 			device_printf(sc->dev,
3340 			    "failed to set congestion manager context for "
3341 			    "ingress queue %d: %d\n", iq->cntxt_id, rc);
3342 		}
3343 	}
3344 
3345 	/* Enable IQ interrupts */
3346 	atomic_store_rel_int(&iq->state, IQS_IDLE);
3347 	t4_write_reg(sc, sc->sge_gts_reg, V_SEINTARM(iq->intr_params) |
3348 	    V_INGRESSQID(iq->cntxt_id));
3349 
3350 	return (0);
3351 }
3352 
3353 static int
3354 free_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl)
3355 {
3356 	int rc;
3357 	struct adapter *sc = iq->adapter;
3358 	device_t dev;
3359 
3360 	if (sc == NULL)
3361 		return (0);	/* nothing to do */
3362 
3363 	dev = vi ? vi->dev : sc->dev;
3364 
3365 	if (iq->flags & IQ_ALLOCATED) {
3366 		rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
3367 		    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
3368 		    fl ? fl->cntxt_id : 0xffff, 0xffff);
3369 		if (rc != 0) {
3370 			device_printf(dev,
3371 			    "failed to free queue %p: %d\n", iq, rc);
3372 			return (rc);
3373 		}
3374 		iq->flags &= ~IQ_ALLOCATED;
3375 	}
3376 
3377 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
3378 
3379 	bzero(iq, sizeof(*iq));
3380 
3381 	if (fl) {
3382 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
3383 		    fl->desc);
3384 
3385 		if (fl->sdesc)
3386 			free_fl_sdesc(sc, fl);
3387 
3388 		if (mtx_initialized(&fl->fl_lock))
3389 			mtx_destroy(&fl->fl_lock);
3390 
3391 		bzero(fl, sizeof(*fl));
3392 	}
3393 
3394 	return (0);
3395 }
3396 
3397 static void
3398 add_iq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
3399     struct sge_iq *iq)
3400 {
3401 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3402 
3403 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD, &iq->ba,
3404 	    "bus address of descriptor ring");
3405 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3406 	    iq->qsize * IQ_ESIZE, "descriptor ring size in bytes");
3407 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3408 	    CTLTYPE_INT | CTLFLAG_RD, &iq->abs_id, 0, sysctl_uint16, "I",
3409 	    "absolute id of the queue");
3410 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3411 	    CTLTYPE_INT | CTLFLAG_RD, &iq->cntxt_id, 0, sysctl_uint16, "I",
3412 	    "SGE context id of the queue");
3413 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3414 	    CTLTYPE_INT | CTLFLAG_RD, &iq->cidx, 0, sysctl_uint16, "I",
3415 	    "consumer index");
3416 }
3417 
3418 static void
3419 add_fl_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
3420     struct sysctl_oid *oid, struct sge_fl *fl)
3421 {
3422 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3423 
3424 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3425 	    "freelist");
3426 	children = SYSCTL_CHILDREN(oid);
3427 
3428 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3429 	    &fl->ba, "bus address of descriptor ring");
3430 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3431 	    fl->sidx * EQ_ESIZE + sc->params.sge.spg_len,
3432 	    "desc ring size in bytes");
3433 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3434 	    CTLTYPE_INT | CTLFLAG_RD, &fl->cntxt_id, 0, sysctl_uint16, "I",
3435 	    "SGE context id of the freelist");
3436 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL,
3437 	    fl_pad ? 1 : 0, "padding enabled");
3438 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL,
3439 	    fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled");
3440 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
3441 	    0, "consumer index");
3442 	if (fl->flags & FL_BUF_PACKING) {
3443 		SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
3444 		    CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
3445 	}
3446 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
3447 	    0, "producer index");
3448 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_allocated",
3449 	    CTLFLAG_RD, &fl->mbuf_allocated, "# of mbuf allocated");
3450 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_inlined",
3451 	    CTLFLAG_RD, &fl->mbuf_inlined, "# of mbuf inlined in clusters");
3452 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
3453 	    CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
3454 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
3455 	    CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
3456 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
3457 	    CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
3458 }
3459 
3460 static int
3461 alloc_fwq(struct adapter *sc)
3462 {
3463 	int rc, intr_idx;
3464 	struct sge_iq *fwq = &sc->sge.fwq;
3465 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
3466 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3467 
3468 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE);
3469 	if (sc->flags & IS_VF)
3470 		intr_idx = 0;
3471 	else
3472 		intr_idx = sc->intr_count > 1 ? 1 : 0;
3473 	rc = alloc_iq_fl(&sc->port[0]->vi[0], fwq, NULL, intr_idx, -1);
3474 	if (rc != 0) {
3475 		device_printf(sc->dev,
3476 		    "failed to create firmware event queue: %d\n", rc);
3477 		return (rc);
3478 	}
3479 
3480 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD,
3481 	    NULL, "firmware event queue");
3482 	add_iq_sysctls(&sc->ctx, oid, fwq);
3483 
3484 	return (0);
3485 }
3486 
3487 static int
3488 free_fwq(struct adapter *sc)
3489 {
3490 	return free_iq_fl(NULL, &sc->sge.fwq, NULL);
3491 }
3492 
3493 static int
3494 alloc_ctrlq(struct adapter *sc, struct sge_wrq *ctrlq, int idx,
3495     struct sysctl_oid *oid)
3496 {
3497 	int rc;
3498 	char name[16];
3499 	struct sysctl_oid_list *children;
3500 
3501 	snprintf(name, sizeof(name), "%s ctrlq%d", device_get_nameunit(sc->dev),
3502 	    idx);
3503 	init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[idx]->tx_chan,
3504 	    sc->sge.fwq.cntxt_id, name);
3505 
3506 	children = SYSCTL_CHILDREN(oid);
3507 	snprintf(name, sizeof(name), "%d", idx);
3508 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3509 	    NULL, "ctrl queue");
3510 	rc = alloc_wrq(sc, NULL, ctrlq, oid);
3511 
3512 	return (rc);
3513 }
3514 
3515 int
3516 tnl_cong(struct port_info *pi, int drop)
3517 {
3518 
3519 	if (drop == -1)
3520 		return (-1);
3521 	else if (drop == 1)
3522 		return (0);
3523 	else
3524 		return (pi->rx_e_chan_map);
3525 }
3526 
3527 static int
3528 alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int intr_idx, int idx,
3529     struct sysctl_oid *oid)
3530 {
3531 	int rc;
3532 	struct adapter *sc = vi->pi->adapter;
3533 	struct sysctl_oid_list *children;
3534 	char name[16];
3535 
3536 	rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, intr_idx,
3537 	    tnl_cong(vi->pi, cong_drop));
3538 	if (rc != 0)
3539 		return (rc);
3540 
3541 	if (idx == 0)
3542 		sc->sge.iq_base = rxq->iq.abs_id - rxq->iq.cntxt_id;
3543 	else
3544 		KASSERT(rxq->iq.cntxt_id + sc->sge.iq_base == rxq->iq.abs_id,
3545 		    ("iq_base mismatch"));
3546 	KASSERT(sc->sge.iq_base == 0 || sc->flags & IS_VF,
3547 	    ("PF with non-zero iq_base"));
3548 
3549 	/*
3550 	 * The freelist is just barely above the starvation threshold right now,
3551 	 * fill it up a bit more.
3552 	 */
3553 	FL_LOCK(&rxq->fl);
3554 	refill_fl(sc, &rxq->fl, 128);
3555 	FL_UNLOCK(&rxq->fl);
3556 
3557 #if defined(INET) || defined(INET6)
3558 	rc = tcp_lro_init_args(&rxq->lro, vi->ifp, lro_entries, lro_mbufs);
3559 	if (rc != 0)
3560 		return (rc);
3561 	MPASS(rxq->lro.ifp == vi->ifp);	/* also indicates LRO init'ed */
3562 
3563 	if (vi->ifp->if_capenable & IFCAP_LRO)
3564 		rxq->iq.flags |= IQ_LRO_ENABLED;
3565 #endif
3566 	if (vi->ifp->if_capenable & IFCAP_HWRXTSTMP)
3567 		rxq->iq.flags |= IQ_RX_TIMESTAMP;
3568 	rxq->ifp = vi->ifp;
3569 
3570 	children = SYSCTL_CHILDREN(oid);
3571 
3572 	snprintf(name, sizeof(name), "%d", idx);
3573 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3574 	    NULL, "rx queue");
3575 	children = SYSCTL_CHILDREN(oid);
3576 
3577 	add_iq_sysctls(&vi->ctx, oid, &rxq->iq);
3578 #if defined(INET) || defined(INET6)
3579 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
3580 	    &rxq->lro.lro_queued, 0, NULL);
3581 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
3582 	    &rxq->lro.lro_flushed, 0, NULL);
3583 #endif
3584 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
3585 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
3586 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_extraction",
3587 	    CTLFLAG_RD, &rxq->vlan_extraction,
3588 	    "# of times hardware extracted 802.1Q tag");
3589 
3590 	add_fl_sysctls(sc, &vi->ctx, oid, &rxq->fl);
3591 
3592 	return (rc);
3593 }
3594 
3595 static int
3596 free_rxq(struct vi_info *vi, struct sge_rxq *rxq)
3597 {
3598 	int rc;
3599 
3600 #if defined(INET) || defined(INET6)
3601 	if (rxq->lro.ifp) {
3602 		tcp_lro_free(&rxq->lro);
3603 		rxq->lro.ifp = NULL;
3604 	}
3605 #endif
3606 
3607 	rc = free_iq_fl(vi, &rxq->iq, &rxq->fl);
3608 	if (rc == 0)
3609 		bzero(rxq, sizeof(*rxq));
3610 
3611 	return (rc);
3612 }
3613 
3614 #ifdef TCP_OFFLOAD
3615 static int
3616 alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq,
3617     int intr_idx, int idx, struct sysctl_oid *oid)
3618 {
3619 	struct port_info *pi = vi->pi;
3620 	int rc;
3621 	struct sysctl_oid_list *children;
3622 	char name[16];
3623 
3624 	rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx, 0);
3625 	if (rc != 0)
3626 		return (rc);
3627 
3628 	children = SYSCTL_CHILDREN(oid);
3629 
3630 	snprintf(name, sizeof(name), "%d", idx);
3631 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3632 	    NULL, "rx queue");
3633 	add_iq_sysctls(&vi->ctx, oid, &ofld_rxq->iq);
3634 	add_fl_sysctls(pi->adapter, &vi->ctx, oid, &ofld_rxq->fl);
3635 
3636 	return (rc);
3637 }
3638 
3639 static int
3640 free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq)
3641 {
3642 	int rc;
3643 
3644 	rc = free_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl);
3645 	if (rc == 0)
3646 		bzero(ofld_rxq, sizeof(*ofld_rxq));
3647 
3648 	return (rc);
3649 }
3650 #endif
3651 
3652 #ifdef DEV_NETMAP
3653 static int
3654 alloc_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int intr_idx,
3655     int idx, struct sysctl_oid *oid)
3656 {
3657 	int rc;
3658 	struct sysctl_oid_list *children;
3659 	struct sysctl_ctx_list *ctx;
3660 	char name[16];
3661 	size_t len;
3662 	struct adapter *sc = vi->pi->adapter;
3663 	struct netmap_adapter *na = NA(vi->ifp);
3664 
3665 	MPASS(na != NULL);
3666 
3667 	len = vi->qsize_rxq * IQ_ESIZE;
3668 	rc = alloc_ring(sc, len, &nm_rxq->iq_desc_tag, &nm_rxq->iq_desc_map,
3669 	    &nm_rxq->iq_ba, (void **)&nm_rxq->iq_desc);
3670 	if (rc != 0)
3671 		return (rc);
3672 
3673 	len = na->num_rx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3674 	rc = alloc_ring(sc, len, &nm_rxq->fl_desc_tag, &nm_rxq->fl_desc_map,
3675 	    &nm_rxq->fl_ba, (void **)&nm_rxq->fl_desc);
3676 	if (rc != 0)
3677 		return (rc);
3678 
3679 	nm_rxq->vi = vi;
3680 	nm_rxq->nid = idx;
3681 	nm_rxq->iq_cidx = 0;
3682 	nm_rxq->iq_sidx = vi->qsize_rxq - sc->params.sge.spg_len / IQ_ESIZE;
3683 	nm_rxq->iq_gen = F_RSPD_GEN;
3684 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
3685 	nm_rxq->fl_sidx = na->num_rx_desc;
3686 	nm_rxq->intr_idx = intr_idx;
3687 	nm_rxq->iq_cntxt_id = INVALID_NM_RXQ_CNTXT_ID;
3688 
3689 	ctx = &vi->ctx;
3690 	children = SYSCTL_CHILDREN(oid);
3691 
3692 	snprintf(name, sizeof(name), "%d", idx);
3693 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL,
3694 	    "rx queue");
3695 	children = SYSCTL_CHILDREN(oid);
3696 
3697 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3698 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_abs_id, 0, sysctl_uint16,
3699 	    "I", "absolute id of the queue");
3700 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3701 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cntxt_id, 0, sysctl_uint16,
3702 	    "I", "SGE context id of the queue");
3703 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3704 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cidx, 0, sysctl_uint16, "I",
3705 	    "consumer index");
3706 
3707 	children = SYSCTL_CHILDREN(oid);
3708 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3709 	    "freelist");
3710 	children = SYSCTL_CHILDREN(oid);
3711 
3712 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3713 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->fl_cntxt_id, 0, sysctl_uint16,
3714 	    "I", "SGE context id of the freelist");
3715 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
3716 	    &nm_rxq->fl_cidx, 0, "consumer index");
3717 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
3718 	    &nm_rxq->fl_pidx, 0, "producer index");
3719 
3720 	return (rc);
3721 }
3722 
3723 
3724 static int
3725 free_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
3726 {
3727 	struct adapter *sc = vi->pi->adapter;
3728 
3729 	if (vi->flags & VI_INIT_DONE)
3730 		MPASS(nm_rxq->iq_cntxt_id == INVALID_NM_RXQ_CNTXT_ID);
3731 	else
3732 		MPASS(nm_rxq->iq_cntxt_id == 0);
3733 
3734 	free_ring(sc, nm_rxq->iq_desc_tag, nm_rxq->iq_desc_map, nm_rxq->iq_ba,
3735 	    nm_rxq->iq_desc);
3736 	free_ring(sc, nm_rxq->fl_desc_tag, nm_rxq->fl_desc_map, nm_rxq->fl_ba,
3737 	    nm_rxq->fl_desc);
3738 
3739 	return (0);
3740 }
3741 
3742 static int
3743 alloc_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq, int iqidx, int idx,
3744     struct sysctl_oid *oid)
3745 {
3746 	int rc;
3747 	size_t len;
3748 	struct port_info *pi = vi->pi;
3749 	struct adapter *sc = pi->adapter;
3750 	struct netmap_adapter *na = NA(vi->ifp);
3751 	char name[16];
3752 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3753 
3754 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3755 	rc = alloc_ring(sc, len, &nm_txq->desc_tag, &nm_txq->desc_map,
3756 	    &nm_txq->ba, (void **)&nm_txq->desc);
3757 	if (rc)
3758 		return (rc);
3759 
3760 	nm_txq->pidx = nm_txq->cidx = 0;
3761 	nm_txq->sidx = na->num_tx_desc;
3762 	nm_txq->nid = idx;
3763 	nm_txq->iqidx = iqidx;
3764 	nm_txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3765 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(sc->pf) |
3766 	    V_TXPKT_VF(vi->vin) | V_TXPKT_VF_VLD(vi->vfvld));
3767 	if (sc->params.fw_vers >= FW_VERSION32(1, 24, 11, 0))
3768 		nm_txq->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS2_WR));
3769 	else
3770 		nm_txq->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
3771 	nm_txq->cntxt_id = INVALID_NM_TXQ_CNTXT_ID;
3772 
3773 	snprintf(name, sizeof(name), "%d", idx);
3774 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3775 	    NULL, "netmap tx queue");
3776 	children = SYSCTL_CHILDREN(oid);
3777 
3778 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3779 	    &nm_txq->cntxt_id, 0, "SGE context id of the queue");
3780 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3781 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->cidx, 0, sysctl_uint16, "I",
3782 	    "consumer index");
3783 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3784 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->pidx, 0, sysctl_uint16, "I",
3785 	    "producer index");
3786 
3787 	return (rc);
3788 }
3789 
3790 static int
3791 free_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
3792 {
3793 	struct adapter *sc = vi->pi->adapter;
3794 
3795 	if (vi->flags & VI_INIT_DONE)
3796 		MPASS(nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID);
3797 	else
3798 		MPASS(nm_txq->cntxt_id == 0);
3799 
3800 	free_ring(sc, nm_txq->desc_tag, nm_txq->desc_map, nm_txq->ba,
3801 	    nm_txq->desc);
3802 
3803 	return (0);
3804 }
3805 #endif
3806 
3807 /*
3808  * Returns a reasonable automatic cidx flush threshold for a given queue size.
3809  */
3810 static u_int
3811 qsize_to_fthresh(int qsize)
3812 {
3813 	u_int fthresh;
3814 
3815 	while (!powerof2(qsize))
3816 		qsize++;
3817 	fthresh = ilog2(qsize);
3818 	if (fthresh > X_CIDXFLUSHTHRESH_128)
3819 		fthresh = X_CIDXFLUSHTHRESH_128;
3820 
3821 	return (fthresh);
3822 }
3823 
3824 static int
3825 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
3826 {
3827 	int rc, cntxt_id;
3828 	struct fw_eq_ctrl_cmd c;
3829 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3830 
3831 	bzero(&c, sizeof(c));
3832 
3833 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
3834 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
3835 	    V_FW_EQ_CTRL_CMD_VFN(0));
3836 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
3837 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
3838 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid));
3839 	c.physeqid_pkd = htobe32(0);
3840 	c.fetchszm_to_iqid =
3841 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
3842 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
3843 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
3844 	c.dcaen_to_eqsize =
3845 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
3846 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
3847 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3848 		V_FW_EQ_CTRL_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
3849 		V_FW_EQ_CTRL_CMD_EQSIZE(qsize));
3850 	c.eqaddr = htobe64(eq->ba);
3851 
3852 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3853 	if (rc != 0) {
3854 		device_printf(sc->dev,
3855 		    "failed to create control queue %d: %d\n", eq->tx_chan, rc);
3856 		return (rc);
3857 	}
3858 	eq->flags |= EQ_ALLOCATED;
3859 
3860 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
3861 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3862 	if (cntxt_id >= sc->sge.neq)
3863 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3864 		cntxt_id, sc->sge.neq - 1);
3865 	sc->sge.eqmap[cntxt_id] = eq;
3866 
3867 	return (rc);
3868 }
3869 
3870 static int
3871 eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3872 {
3873 	int rc, cntxt_id;
3874 	struct fw_eq_eth_cmd c;
3875 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3876 
3877 	bzero(&c, sizeof(c));
3878 
3879 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
3880 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
3881 	    V_FW_EQ_ETH_CMD_VFN(0));
3882 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
3883 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
3884 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
3885 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
3886 	c.fetchszm_to_iqid =
3887 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3888 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
3889 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
3890 	c.dcaen_to_eqsize =
3891 	    htobe32(V_FW_EQ_ETH_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
3892 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
3893 		V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3894 		V_FW_EQ_ETH_CMD_EQSIZE(qsize));
3895 	c.eqaddr = htobe64(eq->ba);
3896 
3897 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3898 	if (rc != 0) {
3899 		device_printf(vi->dev,
3900 		    "failed to create Ethernet egress queue: %d\n", rc);
3901 		return (rc);
3902 	}
3903 	eq->flags |= EQ_ALLOCATED;
3904 
3905 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
3906 	eq->abs_id = G_FW_EQ_ETH_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
3907 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3908 	if (cntxt_id >= sc->sge.neq)
3909 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3910 		cntxt_id, sc->sge.neq - 1);
3911 	sc->sge.eqmap[cntxt_id] = eq;
3912 
3913 	return (rc);
3914 }
3915 
3916 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3917 static int
3918 ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3919 {
3920 	int rc, cntxt_id;
3921 	struct fw_eq_ofld_cmd c;
3922 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3923 
3924 	bzero(&c, sizeof(c));
3925 
3926 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
3927 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
3928 	    V_FW_EQ_OFLD_CMD_VFN(0));
3929 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
3930 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
3931 	c.fetchszm_to_iqid =
3932 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
3933 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
3934 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
3935 	c.dcaen_to_eqsize =
3936 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
3937 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
3938 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3939 		V_FW_EQ_OFLD_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
3940 		V_FW_EQ_OFLD_CMD_EQSIZE(qsize));
3941 	c.eqaddr = htobe64(eq->ba);
3942 
3943 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3944 	if (rc != 0) {
3945 		device_printf(vi->dev,
3946 		    "failed to create egress queue for TCP offload: %d\n", rc);
3947 		return (rc);
3948 	}
3949 	eq->flags |= EQ_ALLOCATED;
3950 
3951 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
3952 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3953 	if (cntxt_id >= sc->sge.neq)
3954 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3955 		cntxt_id, sc->sge.neq - 1);
3956 	sc->sge.eqmap[cntxt_id] = eq;
3957 
3958 	return (rc);
3959 }
3960 #endif
3961 
3962 static int
3963 alloc_eq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3964 {
3965 	int rc, qsize;
3966 	size_t len;
3967 
3968 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
3969 
3970 	qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3971 	len = qsize * EQ_ESIZE;
3972 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
3973 	    &eq->ba, (void **)&eq->desc);
3974 	if (rc)
3975 		return (rc);
3976 
3977 	eq->pidx = eq->cidx = eq->dbidx = 0;
3978 	/* Note that equeqidx is not used with sge_wrq (OFLD/CTRL) queues. */
3979 	eq->equeqidx = 0;
3980 	eq->doorbells = sc->doorbells;
3981 
3982 	switch (eq->flags & EQ_TYPEMASK) {
3983 	case EQ_CTRL:
3984 		rc = ctrl_eq_alloc(sc, eq);
3985 		break;
3986 
3987 	case EQ_ETH:
3988 		rc = eth_eq_alloc(sc, vi, eq);
3989 		break;
3990 
3991 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3992 	case EQ_OFLD:
3993 		rc = ofld_eq_alloc(sc, vi, eq);
3994 		break;
3995 #endif
3996 
3997 	default:
3998 		panic("%s: invalid eq type %d.", __func__,
3999 		    eq->flags & EQ_TYPEMASK);
4000 	}
4001 	if (rc != 0) {
4002 		device_printf(sc->dev,
4003 		    "failed to allocate egress queue(%d): %d\n",
4004 		    eq->flags & EQ_TYPEMASK, rc);
4005 	}
4006 
4007 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
4008 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
4009 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
4010 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
4011 		uint32_t mask = (1 << s_qpp) - 1;
4012 		volatile uint8_t *udb;
4013 
4014 		udb = sc->udbs_base + UDBS_DB_OFFSET;
4015 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
4016 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
4017 		if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
4018 	    		clrbit(&eq->doorbells, DOORBELL_WCWR);
4019 		else {
4020 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
4021 			eq->udb_qid = 0;
4022 		}
4023 		eq->udb = (volatile void *)udb;
4024 	}
4025 
4026 	return (rc);
4027 }
4028 
4029 static int
4030 free_eq(struct adapter *sc, struct sge_eq *eq)
4031 {
4032 	int rc;
4033 
4034 	if (eq->flags & EQ_ALLOCATED) {
4035 		switch (eq->flags & EQ_TYPEMASK) {
4036 		case EQ_CTRL:
4037 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
4038 			    eq->cntxt_id);
4039 			break;
4040 
4041 		case EQ_ETH:
4042 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
4043 			    eq->cntxt_id);
4044 			break;
4045 
4046 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
4047 		case EQ_OFLD:
4048 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
4049 			    eq->cntxt_id);
4050 			break;
4051 #endif
4052 
4053 		default:
4054 			panic("%s: invalid eq type %d.", __func__,
4055 			    eq->flags & EQ_TYPEMASK);
4056 		}
4057 		if (rc != 0) {
4058 			device_printf(sc->dev,
4059 			    "failed to free egress queue (%d): %d\n",
4060 			    eq->flags & EQ_TYPEMASK, rc);
4061 			return (rc);
4062 		}
4063 		eq->flags &= ~EQ_ALLOCATED;
4064 	}
4065 
4066 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
4067 
4068 	if (mtx_initialized(&eq->eq_lock))
4069 		mtx_destroy(&eq->eq_lock);
4070 
4071 	bzero(eq, sizeof(*eq));
4072 	return (0);
4073 }
4074 
4075 static int
4076 alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq,
4077     struct sysctl_oid *oid)
4078 {
4079 	int rc;
4080 	struct sysctl_ctx_list *ctx = vi ? &vi->ctx : &sc->ctx;
4081 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
4082 
4083 	rc = alloc_eq(sc, vi, &wrq->eq);
4084 	if (rc)
4085 		return (rc);
4086 
4087 	wrq->adapter = sc;
4088 	TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq);
4089 	TAILQ_INIT(&wrq->incomplete_wrs);
4090 	STAILQ_INIT(&wrq->wr_list);
4091 	wrq->nwr_pending = 0;
4092 	wrq->ndesc_needed = 0;
4093 
4094 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
4095 	    &wrq->eq.ba, "bus address of descriptor ring");
4096 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
4097 	    wrq->eq.sidx * EQ_ESIZE + sc->params.sge.spg_len,
4098 	    "desc ring size in bytes");
4099 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
4100 	    &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
4101 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
4102 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I",
4103 	    "consumer index");
4104 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx",
4105 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I",
4106 	    "producer index");
4107 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
4108 	    wrq->eq.sidx, "status page index");
4109 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD,
4110 	    &wrq->tx_wrs_direct, "# of work requests (direct)");
4111 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD,
4112 	    &wrq->tx_wrs_copied, "# of work requests (copied)");
4113 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_sspace", CTLFLAG_RD,
4114 	    &wrq->tx_wrs_ss, "# of work requests (copied from scratch space)");
4115 
4116 	return (rc);
4117 }
4118 
4119 static int
4120 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
4121 {
4122 	int rc;
4123 
4124 	rc = free_eq(sc, &wrq->eq);
4125 	if (rc)
4126 		return (rc);
4127 
4128 	bzero(wrq, sizeof(*wrq));
4129 	return (0);
4130 }
4131 
4132 static int
4133 alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx,
4134     struct sysctl_oid *oid)
4135 {
4136 	int rc;
4137 	struct port_info *pi = vi->pi;
4138 	struct adapter *sc = pi->adapter;
4139 	struct sge_eq *eq = &txq->eq;
4140 	char name[16];
4141 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
4142 
4143 	rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx, can_resume_eth_tx,
4144 	    M_CXGBE, M_WAITOK);
4145 	if (rc != 0) {
4146 		device_printf(sc->dev, "failed to allocate mp_ring: %d\n", rc);
4147 		return (rc);
4148 	}
4149 
4150 	rc = alloc_eq(sc, vi, eq);
4151 	if (rc != 0) {
4152 		mp_ring_free(txq->r);
4153 		txq->r = NULL;
4154 		return (rc);
4155 	}
4156 
4157 	/* Can't fail after this point. */
4158 
4159 	if (idx == 0)
4160 		sc->sge.eq_base = eq->abs_id - eq->cntxt_id;
4161 	else
4162 		KASSERT(eq->cntxt_id + sc->sge.eq_base == eq->abs_id,
4163 		    ("eq_base mismatch"));
4164 	KASSERT(sc->sge.eq_base == 0 || sc->flags & IS_VF,
4165 	    ("PF with non-zero eq_base"));
4166 
4167 	TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq);
4168 	txq->ifp = vi->ifp;
4169 	txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK);
4170 	if (sc->flags & IS_VF)
4171 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
4172 		    V_TXPKT_INTF(pi->tx_chan));
4173 	else
4174 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
4175 		    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(sc->pf) |
4176 		    V_TXPKT_VF(vi->vin) | V_TXPKT_VF_VLD(vi->vfvld));
4177 	txq->tc_idx = -1;
4178 	txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE,
4179 	    M_ZERO | M_WAITOK);
4180 
4181 	snprintf(name, sizeof(name), "%d", idx);
4182 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
4183 	    NULL, "tx queue");
4184 	children = SYSCTL_CHILDREN(oid);
4185 
4186 	SYSCTL_ADD_UAUTO(&vi->ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
4187 	    &eq->ba, "bus address of descriptor ring");
4188 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
4189 	    eq->sidx * EQ_ESIZE + sc->params.sge.spg_len,
4190 	    "desc ring size in bytes");
4191 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
4192 	    &eq->abs_id, 0, "absolute id of the queue");
4193 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
4194 	    &eq->cntxt_id, 0, "SGE context id of the queue");
4195 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
4196 	    CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I",
4197 	    "consumer index");
4198 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
4199 	    CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I",
4200 	    "producer index");
4201 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
4202 	    eq->sidx, "status page index");
4203 
4204 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "tc",
4205 	    CTLTYPE_INT | CTLFLAG_RW, vi, idx, sysctl_tc, "I",
4206 	    "traffic class (-1 means none)");
4207 
4208 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
4209 	    &txq->txcsum, "# of times hardware assisted with checksum");
4210 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_insertion",
4211 	    CTLFLAG_RD, &txq->vlan_insertion,
4212 	    "# of times hardware inserted 802.1Q tag");
4213 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
4214 	    &txq->tso_wrs, "# of TSO work requests");
4215 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
4216 	    &txq->imm_wrs, "# of work requests with immediate data");
4217 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
4218 	    &txq->sgl_wrs, "# of work requests with direct SGL");
4219 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
4220 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
4221 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_wrs",
4222 	    CTLFLAG_RD, &txq->txpkts0_wrs,
4223 	    "# of txpkts (type 0) work requests");
4224 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_wrs",
4225 	    CTLFLAG_RD, &txq->txpkts1_wrs,
4226 	    "# of txpkts (type 1) work requests");
4227 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_pkts",
4228 	    CTLFLAG_RD, &txq->txpkts0_pkts,
4229 	    "# of frames tx'd using type0 txpkts work requests");
4230 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_pkts",
4231 	    CTLFLAG_RD, &txq->txpkts1_pkts,
4232 	    "# of frames tx'd using type1 txpkts work requests");
4233 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "raw_wrs", CTLFLAG_RD,
4234 	    &txq->raw_wrs, "# of raw work requests (non-packets)");
4235 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "tls_wrs", CTLFLAG_RD,
4236 	    &txq->tls_wrs, "# of TLS work requests (TLS records)");
4237 
4238 #ifdef KERN_TLS
4239 	if (sc->flags & KERN_TLS_OK) {
4240 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4241 		    "kern_tls_records", CTLFLAG_RD, &txq->kern_tls_records,
4242 		    "# of NIC TLS records transmitted");
4243 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4244 		    "kern_tls_short", CTLFLAG_RD, &txq->kern_tls_short,
4245 		    "# of short NIC TLS records transmitted");
4246 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4247 		    "kern_tls_partial", CTLFLAG_RD, &txq->kern_tls_partial,
4248 		    "# of partial NIC TLS records transmitted");
4249 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4250 		    "kern_tls_full", CTLFLAG_RD, &txq->kern_tls_full,
4251 		    "# of full NIC TLS records transmitted");
4252 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4253 		    "kern_tls_octets", CTLFLAG_RD, &txq->kern_tls_octets,
4254 		    "# of payload octets in transmitted NIC TLS records");
4255 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4256 		    "kern_tls_waste", CTLFLAG_RD, &txq->kern_tls_waste,
4257 		    "# of octets DMAd but not transmitted in NIC TLS records");
4258 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4259 		    "kern_tls_options", CTLFLAG_RD, &txq->kern_tls_options,
4260 		    "# of NIC TLS options-only packets transmitted");
4261 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4262 		    "kern_tls_header", CTLFLAG_RD, &txq->kern_tls_header,
4263 		    "# of NIC TLS header-only packets transmitted");
4264 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4265 		    "kern_tls_fin", CTLFLAG_RD, &txq->kern_tls_fin,
4266 		    "# of NIC TLS FIN-only packets transmitted");
4267 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4268 		    "kern_tls_fin_short", CTLFLAG_RD, &txq->kern_tls_fin_short,
4269 		    "# of NIC TLS padded FIN packets on short TLS records");
4270 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4271 		    "kern_tls_cbc", CTLFLAG_RD, &txq->kern_tls_cbc,
4272 		    "# of NIC TLS sessions using AES-CBC");
4273 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4274 		    "kern_tls_gcm", CTLFLAG_RD, &txq->kern_tls_gcm,
4275 		    "# of NIC TLS sessions using AES-GCM");
4276 	}
4277 #endif
4278 
4279 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_enqueues",
4280 	    CTLFLAG_RD, &txq->r->enqueues,
4281 	    "# of enqueues to the mp_ring for this queue");
4282 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_drops",
4283 	    CTLFLAG_RD, &txq->r->drops,
4284 	    "# of drops in the mp_ring for this queue");
4285 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_starts",
4286 	    CTLFLAG_RD, &txq->r->starts,
4287 	    "# of normal consumer starts in the mp_ring for this queue");
4288 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_stalls",
4289 	    CTLFLAG_RD, &txq->r->stalls,
4290 	    "# of consumer stalls in the mp_ring for this queue");
4291 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_restarts",
4292 	    CTLFLAG_RD, &txq->r->restarts,
4293 	    "# of consumer restarts in the mp_ring for this queue");
4294 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_abdications",
4295 	    CTLFLAG_RD, &txq->r->abdications,
4296 	    "# of consumer abdications in the mp_ring for this queue");
4297 
4298 	return (0);
4299 }
4300 
4301 static int
4302 free_txq(struct vi_info *vi, struct sge_txq *txq)
4303 {
4304 	int rc;
4305 	struct adapter *sc = vi->pi->adapter;
4306 	struct sge_eq *eq = &txq->eq;
4307 
4308 	rc = free_eq(sc, eq);
4309 	if (rc)
4310 		return (rc);
4311 
4312 	sglist_free(txq->gl);
4313 	free(txq->sdesc, M_CXGBE);
4314 	mp_ring_free(txq->r);
4315 
4316 	bzero(txq, sizeof(*txq));
4317 	return (0);
4318 }
4319 
4320 static void
4321 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
4322 {
4323 	bus_addr_t *ba = arg;
4324 
4325 	KASSERT(nseg == 1,
4326 	    ("%s meant for single segment mappings only.", __func__));
4327 
4328 	*ba = error ? 0 : segs->ds_addr;
4329 }
4330 
4331 static inline void
4332 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
4333 {
4334 	uint32_t n, v;
4335 
4336 	n = IDXDIFF(fl->pidx / 8, fl->dbidx, fl->sidx);
4337 	MPASS(n > 0);
4338 
4339 	wmb();
4340 	v = fl->dbval | V_PIDX(n);
4341 	if (fl->udb)
4342 		*fl->udb = htole32(v);
4343 	else
4344 		t4_write_reg(sc, sc->sge_kdoorbell_reg, v);
4345 	IDXINCR(fl->dbidx, n, fl->sidx);
4346 }
4347 
4348 /*
4349  * Fills up the freelist by allocating up to 'n' buffers.  Buffers that are
4350  * recycled do not count towards this allocation budget.
4351  *
4352  * Returns non-zero to indicate that this freelist should be added to the list
4353  * of starving freelists.
4354  */
4355 static int
4356 refill_fl(struct adapter *sc, struct sge_fl *fl, int n)
4357 {
4358 	__be64 *d;
4359 	struct fl_sdesc *sd;
4360 	uintptr_t pa;
4361 	caddr_t cl;
4362 	struct cluster_layout *cll;
4363 	struct sw_zone_info *swz;
4364 	struct cluster_metadata *clm;
4365 	uint16_t max_pidx;
4366 	uint16_t hw_cidx = fl->hw_cidx;		/* stable snapshot */
4367 
4368 	FL_LOCK_ASSERT_OWNED(fl);
4369 
4370 	/*
4371 	 * We always stop at the beginning of the hardware descriptor that's just
4372 	 * before the one with the hw cidx.  This is to avoid hw pidx = hw cidx,
4373 	 * which would mean an empty freelist to the chip.
4374 	 */
4375 	max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1;
4376 	if (fl->pidx == max_pidx * 8)
4377 		return (0);
4378 
4379 	d = &fl->desc[fl->pidx];
4380 	sd = &fl->sdesc[fl->pidx];
4381 	cll = &fl->cll_def;	/* default layout */
4382 	swz = &sc->sge.sw_zone_info[cll->zidx];
4383 
4384 	while (n > 0) {
4385 
4386 		if (sd->cl != NULL) {
4387 
4388 			if (sd->nmbuf == 0) {
4389 				/*
4390 				 * Fast recycle without involving any atomics on
4391 				 * the cluster's metadata (if the cluster has
4392 				 * metadata).  This happens when all frames
4393 				 * received in the cluster were small enough to
4394 				 * fit within a single mbuf each.
4395 				 */
4396 				fl->cl_fast_recycled++;
4397 #ifdef INVARIANTS
4398 				clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
4399 				if (clm != NULL)
4400 					MPASS(clm->refcount == 1);
4401 #endif
4402 				goto recycled_fast;
4403 			}
4404 
4405 			/*
4406 			 * Cluster is guaranteed to have metadata.  Clusters
4407 			 * without metadata always take the fast recycle path
4408 			 * when they're recycled.
4409 			 */
4410 			clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
4411 			MPASS(clm != NULL);
4412 
4413 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4414 				fl->cl_recycled++;
4415 				counter_u64_add(extfree_rels, 1);
4416 				goto recycled;
4417 			}
4418 			sd->cl = NULL;	/* gave up my reference */
4419 		}
4420 		MPASS(sd->cl == NULL);
4421 alloc:
4422 		cl = uma_zalloc(swz->zone, M_NOWAIT);
4423 		if (__predict_false(cl == NULL)) {
4424 			if (cll == &fl->cll_alt || fl->cll_alt.zidx == -1 ||
4425 			    fl->cll_def.zidx == fl->cll_alt.zidx)
4426 				break;
4427 
4428 			/* fall back to the safe zone */
4429 			cll = &fl->cll_alt;
4430 			swz = &sc->sge.sw_zone_info[cll->zidx];
4431 			goto alloc;
4432 		}
4433 		fl->cl_allocated++;
4434 		n--;
4435 
4436 		pa = pmap_kextract((vm_offset_t)cl);
4437 		pa += cll->region1;
4438 		sd->cl = cl;
4439 		sd->cll = *cll;
4440 		*d = htobe64(pa | cll->hwidx);
4441 		clm = cl_metadata(sc, fl, cll, cl);
4442 		if (clm != NULL) {
4443 recycled:
4444 #ifdef INVARIANTS
4445 			clm->sd = sd;
4446 #endif
4447 			clm->refcount = 1;
4448 		}
4449 		sd->nmbuf = 0;
4450 recycled_fast:
4451 		d++;
4452 		sd++;
4453 		if (__predict_false(++fl->pidx % 8 == 0)) {
4454 			uint16_t pidx = fl->pidx / 8;
4455 
4456 			if (__predict_false(pidx == fl->sidx)) {
4457 				fl->pidx = 0;
4458 				pidx = 0;
4459 				sd = fl->sdesc;
4460 				d = fl->desc;
4461 			}
4462 			if (pidx == max_pidx)
4463 				break;
4464 
4465 			if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4)
4466 				ring_fl_db(sc, fl);
4467 		}
4468 	}
4469 
4470 	if (fl->pidx / 8 != fl->dbidx)
4471 		ring_fl_db(sc, fl);
4472 
4473 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
4474 }
4475 
4476 /*
4477  * Attempt to refill all starving freelists.
4478  */
4479 static void
4480 refill_sfl(void *arg)
4481 {
4482 	struct adapter *sc = arg;
4483 	struct sge_fl *fl, *fl_temp;
4484 
4485 	mtx_assert(&sc->sfl_lock, MA_OWNED);
4486 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
4487 		FL_LOCK(fl);
4488 		refill_fl(sc, fl, 64);
4489 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
4490 			TAILQ_REMOVE(&sc->sfl, fl, link);
4491 			fl->flags &= ~FL_STARVING;
4492 		}
4493 		FL_UNLOCK(fl);
4494 	}
4495 
4496 	if (!TAILQ_EMPTY(&sc->sfl))
4497 		callout_schedule(&sc->sfl_callout, hz / 5);
4498 }
4499 
4500 static int
4501 alloc_fl_sdesc(struct sge_fl *fl)
4502 {
4503 
4504 	fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc), M_CXGBE,
4505 	    M_ZERO | M_WAITOK);
4506 
4507 	return (0);
4508 }
4509 
4510 static void
4511 free_fl_sdesc(struct adapter *sc, struct sge_fl *fl)
4512 {
4513 	struct fl_sdesc *sd;
4514 	struct cluster_metadata *clm;
4515 	struct cluster_layout *cll;
4516 	int i;
4517 
4518 	sd = fl->sdesc;
4519 	for (i = 0; i < fl->sidx * 8; i++, sd++) {
4520 		if (sd->cl == NULL)
4521 			continue;
4522 
4523 		cll = &sd->cll;
4524 		clm = cl_metadata(sc, fl, cll, sd->cl);
4525 		if (sd->nmbuf == 0)
4526 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4527 		else if (clm && atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4528 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4529 			counter_u64_add(extfree_rels, 1);
4530 		}
4531 		sd->cl = NULL;
4532 	}
4533 
4534 	free(fl->sdesc, M_CXGBE);
4535 	fl->sdesc = NULL;
4536 }
4537 
4538 static inline void
4539 get_pkt_gl(struct mbuf *m, struct sglist *gl)
4540 {
4541 	int rc;
4542 
4543 	M_ASSERTPKTHDR(m);
4544 
4545 	sglist_reset(gl);
4546 	rc = sglist_append_mbuf(gl, m);
4547 	if (__predict_false(rc != 0)) {
4548 		panic("%s: mbuf %p (%d segs) was vetted earlier but now fails "
4549 		    "with %d.", __func__, m, mbuf_nsegs(m), rc);
4550 	}
4551 
4552 	KASSERT(gl->sg_nseg == mbuf_nsegs(m),
4553 	    ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m,
4554 	    mbuf_nsegs(m), gl->sg_nseg));
4555 	KASSERT(gl->sg_nseg > 0 &&
4556 	    gl->sg_nseg <= (needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS),
4557 	    ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__,
4558 		gl->sg_nseg, needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS));
4559 }
4560 
4561 /*
4562  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
4563  */
4564 static inline u_int
4565 txpkt_len16(u_int nsegs, u_int tso)
4566 {
4567 	u_int n;
4568 
4569 	MPASS(nsegs > 0);
4570 
4571 	nsegs--; /* first segment is part of ulptx_sgl */
4572 	n = sizeof(struct fw_eth_tx_pkt_wr) + sizeof(struct cpl_tx_pkt_core) +
4573 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4574 	if (tso)
4575 		n += sizeof(struct cpl_tx_pkt_lso_core);
4576 
4577 	return (howmany(n, 16));
4578 }
4579 
4580 /*
4581  * len16 for a txpkt_vm WR with a GL.  Includes the firmware work
4582  * request header.
4583  */
4584 static inline u_int
4585 txpkt_vm_len16(u_int nsegs, u_int tso)
4586 {
4587 	u_int n;
4588 
4589 	MPASS(nsegs > 0);
4590 
4591 	nsegs--; /* first segment is part of ulptx_sgl */
4592 	n = sizeof(struct fw_eth_tx_pkt_vm_wr) +
4593 	    sizeof(struct cpl_tx_pkt_core) +
4594 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4595 	if (tso)
4596 		n += sizeof(struct cpl_tx_pkt_lso_core);
4597 
4598 	return (howmany(n, 16));
4599 }
4600 
4601 /*
4602  * len16 for a txpkts type 0 WR with a GL.  Does not include the firmware work
4603  * request header.
4604  */
4605 static inline u_int
4606 txpkts0_len16(u_int nsegs)
4607 {
4608 	u_int n;
4609 
4610 	MPASS(nsegs > 0);
4611 
4612 	nsegs--; /* first segment is part of ulptx_sgl */
4613 	n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) +
4614 	    sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) +
4615 	    8 * ((3 * nsegs) / 2 + (nsegs & 1));
4616 
4617 	return (howmany(n, 16));
4618 }
4619 
4620 /*
4621  * len16 for a txpkts type 1 WR with a GL.  Does not include the firmware work
4622  * request header.
4623  */
4624 static inline u_int
4625 txpkts1_len16(void)
4626 {
4627 	u_int n;
4628 
4629 	n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl);
4630 
4631 	return (howmany(n, 16));
4632 }
4633 
4634 static inline u_int
4635 imm_payload(u_int ndesc)
4636 {
4637 	u_int n;
4638 
4639 	n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) -
4640 	    sizeof(struct cpl_tx_pkt_core);
4641 
4642 	return (n);
4643 }
4644 
4645 static inline uint64_t
4646 csum_to_ctrl(struct adapter *sc, struct mbuf *m)
4647 {
4648 	uint64_t ctrl;
4649 	int csum_type;
4650 
4651 	M_ASSERTPKTHDR(m);
4652 
4653 	if (needs_hwcsum(m) == 0)
4654 		return (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS);
4655 
4656 	ctrl = 0;
4657 	if (needs_l3_csum(m) == 0)
4658 		ctrl |= F_TXPKT_IPCSUM_DIS;
4659 	switch (m->m_pkthdr.csum_flags &
4660 	    (CSUM_IP_TCP | CSUM_IP_UDP | CSUM_IP6_TCP | CSUM_IP6_UDP)) {
4661 	case CSUM_IP_TCP:
4662 		csum_type = TX_CSUM_TCPIP;
4663 		break;
4664 	case CSUM_IP_UDP:
4665 		csum_type = TX_CSUM_UDPIP;
4666 		break;
4667 	case CSUM_IP6_TCP:
4668 		csum_type = TX_CSUM_TCPIP6;
4669 		break;
4670 	case CSUM_IP6_UDP:
4671 		csum_type = TX_CSUM_UDPIP6;
4672 		break;
4673 	default:
4674 		/* needs_hwcsum told us that at least some hwcsum is needed. */
4675 		MPASS(ctrl == 0);
4676 		MPASS(m->m_pkthdr.csum_flags & CSUM_IP);
4677 		ctrl |= F_TXPKT_L4CSUM_DIS;
4678 		csum_type = TX_CSUM_IP;
4679 		break;
4680 	}
4681 
4682 	MPASS(m->m_pkthdr.l2hlen > 0);
4683 	MPASS(m->m_pkthdr.l3hlen > 0);
4684 	ctrl |= V_TXPKT_CSUM_TYPE(csum_type) |
4685 	    V_TXPKT_IPHDR_LEN(m->m_pkthdr.l3hlen);
4686 	if (chip_id(sc) <= CHELSIO_T5)
4687 		ctrl |= V_TXPKT_ETHHDR_LEN(m->m_pkthdr.l2hlen - ETHER_HDR_LEN);
4688 	else
4689 		ctrl |= V_T6_TXPKT_ETHHDR_LEN(m->m_pkthdr.l2hlen - ETHER_HDR_LEN);
4690 
4691 	return (ctrl);
4692 }
4693 
4694 /*
4695  * Write a VM txpkt WR for this packet to the hardware descriptors, update the
4696  * software descriptor, and advance the pidx.  It is guaranteed that enough
4697  * descriptors are available.
4698  *
4699  * The return value is the # of hardware descriptors used.
4700  */
4701 static u_int
4702 write_txpkt_vm_wr(struct adapter *sc, struct sge_txq *txq,
4703     struct fw_eth_tx_pkt_vm_wr *wr, struct mbuf *m0, u_int available)
4704 {
4705 	struct sge_eq *eq = &txq->eq;
4706 	struct tx_sdesc *txsd;
4707 	struct cpl_tx_pkt_core *cpl;
4708 	uint32_t ctrl;	/* used in many unrelated places */
4709 	uint64_t ctrl1;
4710 	int len16, ndesc, pktlen, nsegs;
4711 	caddr_t dst;
4712 
4713 	TXQ_LOCK_ASSERT_OWNED(txq);
4714 	M_ASSERTPKTHDR(m0);
4715 	MPASS(available > 0 && available < eq->sidx);
4716 
4717 	len16 = mbuf_len16(m0);
4718 	nsegs = mbuf_nsegs(m0);
4719 	pktlen = m0->m_pkthdr.len;
4720 	ctrl = sizeof(struct cpl_tx_pkt_core);
4721 	if (needs_tso(m0))
4722 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4723 	ndesc = howmany(len16, EQ_ESIZE / 16);
4724 	MPASS(ndesc <= available);
4725 
4726 	/* Firmware work request header */
4727 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4728 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_VM_WR) |
4729 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4730 
4731 	ctrl = V_FW_WR_LEN16(len16);
4732 	wr->equiq_to_len16 = htobe32(ctrl);
4733 	wr->r3[0] = 0;
4734 	wr->r3[1] = 0;
4735 
4736 	/*
4737 	 * Copy over ethmacdst, ethmacsrc, ethtype, and vlantci.
4738 	 * vlantci is ignored unless the ethtype is 0x8100, so it's
4739 	 * simpler to always copy it rather than making it
4740 	 * conditional.  Also, it seems that we do not have to set
4741 	 * vlantci or fake the ethtype when doing VLAN tag insertion.
4742 	 */
4743 	m_copydata(m0, 0, sizeof(struct ether_header) + 2, wr->ethmacdst);
4744 
4745 	if (needs_tso(m0)) {
4746 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4747 
4748 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4749 		    m0->m_pkthdr.l4hlen > 0,
4750 		    ("%s: mbuf %p needs TSO but missing header lengths",
4751 			__func__, m0));
4752 
4753 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4754 		    F_LSO_LAST_SLICE | V_LSO_ETHHDR_LEN((m0->m_pkthdr.l2hlen -
4755 			ETHER_HDR_LEN) >> 2) |
4756 		    V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) |
4757 		    V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4758 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4759 			ctrl |= F_LSO_IPV6;
4760 
4761 		lso->lso_ctrl = htobe32(ctrl);
4762 		lso->ipid_ofst = htobe16(0);
4763 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4764 		lso->seqno_offset = htobe32(0);
4765 		lso->len = htobe32(pktlen);
4766 
4767 		cpl = (void *)(lso + 1);
4768 
4769 		txq->tso_wrs++;
4770 	} else
4771 		cpl = (void *)(wr + 1);
4772 
4773 	/* Checksum offload */
4774 	ctrl1 = csum_to_ctrl(sc, m0);
4775 	if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS))
4776 		txq->txcsum++;	/* some hardware assistance provided */
4777 
4778 	/* VLAN tag insertion */
4779 	if (needs_vlan_insertion(m0)) {
4780 		ctrl1 |= F_TXPKT_VLAN_VLD |
4781 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4782 		txq->vlan_insertion++;
4783 	}
4784 
4785 	/* CPL header */
4786 	cpl->ctrl0 = txq->cpl_ctrl0;
4787 	cpl->pack = 0;
4788 	cpl->len = htobe16(pktlen);
4789 	cpl->ctrl1 = htobe64(ctrl1);
4790 
4791 	/* SGL */
4792 	dst = (void *)(cpl + 1);
4793 
4794 	/*
4795 	 * A packet using TSO will use up an entire descriptor for the
4796 	 * firmware work request header, LSO CPL, and TX_PKT_XT CPL.
4797 	 * If this descriptor is the last descriptor in the ring, wrap
4798 	 * around to the front of the ring explicitly for the start of
4799 	 * the sgl.
4800 	 */
4801 	if (dst == (void *)&eq->desc[eq->sidx]) {
4802 		dst = (void *)&eq->desc[0];
4803 		write_gl_to_txd(txq, m0, &dst, 0);
4804 	} else
4805 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4806 	txq->sgl_wrs++;
4807 
4808 	txq->txpkt_wrs++;
4809 
4810 	txsd = &txq->sdesc[eq->pidx];
4811 	txsd->m = m0;
4812 	txsd->desc_used = ndesc;
4813 
4814 	return (ndesc);
4815 }
4816 
4817 /*
4818  * Write a raw WR to the hardware descriptors, update the software
4819  * descriptor, and advance the pidx.  It is guaranteed that enough
4820  * descriptors are available.
4821  *
4822  * The return value is the # of hardware descriptors used.
4823  */
4824 static u_int
4825 write_raw_wr(struct sge_txq *txq, void *wr, struct mbuf *m0, u_int available)
4826 {
4827 	struct sge_eq *eq = &txq->eq;
4828 	struct tx_sdesc *txsd;
4829 	struct mbuf *m;
4830 	caddr_t dst;
4831 	int len16, ndesc;
4832 
4833 	len16 = mbuf_len16(m0);
4834 	ndesc = howmany(len16, EQ_ESIZE / 16);
4835 	MPASS(ndesc <= available);
4836 
4837 	dst = wr;
4838 	for (m = m0; m != NULL; m = m->m_next)
4839 		copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4840 
4841 	txq->raw_wrs++;
4842 
4843 	txsd = &txq->sdesc[eq->pidx];
4844 	txsd->m = m0;
4845 	txsd->desc_used = ndesc;
4846 
4847 	return (ndesc);
4848 }
4849 
4850 /*
4851  * Write a txpkt WR for this packet to the hardware descriptors, update the
4852  * software descriptor, and advance the pidx.  It is guaranteed that enough
4853  * descriptors are available.
4854  *
4855  * The return value is the # of hardware descriptors used.
4856  */
4857 static u_int
4858 write_txpkt_wr(struct adapter *sc, struct sge_txq *txq,
4859     struct fw_eth_tx_pkt_wr *wr, struct mbuf *m0, u_int available)
4860 {
4861 	struct sge_eq *eq = &txq->eq;
4862 	struct tx_sdesc *txsd;
4863 	struct cpl_tx_pkt_core *cpl;
4864 	uint32_t ctrl;	/* used in many unrelated places */
4865 	uint64_t ctrl1;
4866 	int len16, ndesc, pktlen, nsegs;
4867 	caddr_t dst;
4868 
4869 	TXQ_LOCK_ASSERT_OWNED(txq);
4870 	M_ASSERTPKTHDR(m0);
4871 	MPASS(available > 0 && available < eq->sidx);
4872 
4873 	len16 = mbuf_len16(m0);
4874 	nsegs = mbuf_nsegs(m0);
4875 	pktlen = m0->m_pkthdr.len;
4876 	ctrl = sizeof(struct cpl_tx_pkt_core);
4877 	if (needs_tso(m0))
4878 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4879 	else if (!(mbuf_cflags(m0) & MC_NOMAP) && pktlen <= imm_payload(2) &&
4880 	    available >= 2) {
4881 		/* Immediate data.  Recalculate len16 and set nsegs to 0. */
4882 		ctrl += pktlen;
4883 		len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) +
4884 		    sizeof(struct cpl_tx_pkt_core) + pktlen, 16);
4885 		nsegs = 0;
4886 	}
4887 	ndesc = howmany(len16, EQ_ESIZE / 16);
4888 	MPASS(ndesc <= available);
4889 
4890 	/* Firmware work request header */
4891 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4892 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
4893 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4894 
4895 	ctrl = V_FW_WR_LEN16(len16);
4896 	wr->equiq_to_len16 = htobe32(ctrl);
4897 	wr->r3 = 0;
4898 
4899 	if (needs_tso(m0)) {
4900 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4901 
4902 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4903 		    m0->m_pkthdr.l4hlen > 0,
4904 		    ("%s: mbuf %p needs TSO but missing header lengths",
4905 			__func__, m0));
4906 
4907 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4908 		    F_LSO_LAST_SLICE | V_LSO_ETHHDR_LEN((m0->m_pkthdr.l2hlen -
4909 			ETHER_HDR_LEN) >> 2) |
4910 		    V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) |
4911 		    V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4912 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4913 			ctrl |= F_LSO_IPV6;
4914 
4915 		lso->lso_ctrl = htobe32(ctrl);
4916 		lso->ipid_ofst = htobe16(0);
4917 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4918 		lso->seqno_offset = htobe32(0);
4919 		lso->len = htobe32(pktlen);
4920 
4921 		cpl = (void *)(lso + 1);
4922 
4923 		txq->tso_wrs++;
4924 	} else
4925 		cpl = (void *)(wr + 1);
4926 
4927 	/* Checksum offload */
4928 	ctrl1 = csum_to_ctrl(sc, m0);
4929 	if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS))
4930 		txq->txcsum++;	/* some hardware assistance provided */
4931 
4932 	/* VLAN tag insertion */
4933 	if (needs_vlan_insertion(m0)) {
4934 		ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4935 		txq->vlan_insertion++;
4936 	}
4937 
4938 	/* CPL header */
4939 	cpl->ctrl0 = txq->cpl_ctrl0;
4940 	cpl->pack = 0;
4941 	cpl->len = htobe16(pktlen);
4942 	cpl->ctrl1 = htobe64(ctrl1);
4943 
4944 	/* SGL */
4945 	dst = (void *)(cpl + 1);
4946 	if (nsegs > 0) {
4947 
4948 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4949 		txq->sgl_wrs++;
4950 	} else {
4951 		struct mbuf *m;
4952 
4953 		for (m = m0; m != NULL; m = m->m_next) {
4954 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4955 #ifdef INVARIANTS
4956 			pktlen -= m->m_len;
4957 #endif
4958 		}
4959 #ifdef INVARIANTS
4960 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
4961 #endif
4962 		txq->imm_wrs++;
4963 	}
4964 
4965 	txq->txpkt_wrs++;
4966 
4967 	txsd = &txq->sdesc[eq->pidx];
4968 	txsd->m = m0;
4969 	txsd->desc_used = ndesc;
4970 
4971 	return (ndesc);
4972 }
4973 
4974 static int
4975 try_txpkts(struct mbuf *m, struct mbuf *n, struct txpkts *txp, u_int available)
4976 {
4977 	u_int needed, nsegs1, nsegs2, l1, l2;
4978 
4979 	if (cannot_use_txpkts(m) || cannot_use_txpkts(n))
4980 		return (1);
4981 
4982 	nsegs1 = mbuf_nsegs(m);
4983 	nsegs2 = mbuf_nsegs(n);
4984 	if (nsegs1 + nsegs2 == 2) {
4985 		txp->wr_type = 1;
4986 		l1 = l2 = txpkts1_len16();
4987 	} else {
4988 		txp->wr_type = 0;
4989 		l1 = txpkts0_len16(nsegs1);
4990 		l2 = txpkts0_len16(nsegs2);
4991 	}
4992 	txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) + l1 + l2;
4993 	needed = howmany(txp->len16, EQ_ESIZE / 16);
4994 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4995 		return (1);
4996 
4997 	txp->plen = m->m_pkthdr.len + n->m_pkthdr.len;
4998 	if (txp->plen > 65535)
4999 		return (1);
5000 
5001 	txp->npkt = 2;
5002 	set_mbuf_len16(m, l1);
5003 	set_mbuf_len16(n, l2);
5004 
5005 	return (0);
5006 }
5007 
5008 static int
5009 add_to_txpkts(struct mbuf *m, struct txpkts *txp, u_int available)
5010 {
5011 	u_int plen, len16, needed, nsegs;
5012 
5013 	MPASS(txp->wr_type == 0 || txp->wr_type == 1);
5014 
5015 	if (cannot_use_txpkts(m))
5016 		return (1);
5017 
5018 	nsegs = mbuf_nsegs(m);
5019 	if (txp->wr_type == 1 && nsegs != 1)
5020 		return (1);
5021 
5022 	plen = txp->plen + m->m_pkthdr.len;
5023 	if (plen > 65535)
5024 		return (1);
5025 
5026 	if (txp->wr_type == 0)
5027 		len16 = txpkts0_len16(nsegs);
5028 	else
5029 		len16 = txpkts1_len16();
5030 	needed = howmany(txp->len16 + len16, EQ_ESIZE / 16);
5031 	if (needed > SGE_MAX_WR_NDESC || needed > available)
5032 		return (1);
5033 
5034 	txp->npkt++;
5035 	txp->plen = plen;
5036 	txp->len16 += len16;
5037 	set_mbuf_len16(m, len16);
5038 
5039 	return (0);
5040 }
5041 
5042 /*
5043  * Write a txpkts WR for the packets in txp to the hardware descriptors, update
5044  * the software descriptor, and advance the pidx.  It is guaranteed that enough
5045  * descriptors are available.
5046  *
5047  * The return value is the # of hardware descriptors used.
5048  */
5049 static u_int
5050 write_txpkts_wr(struct adapter *sc, struct sge_txq *txq,
5051     struct fw_eth_tx_pkts_wr *wr, struct mbuf *m0, const struct txpkts *txp,
5052     u_int available)
5053 {
5054 	struct sge_eq *eq = &txq->eq;
5055 	struct tx_sdesc *txsd;
5056 	struct cpl_tx_pkt_core *cpl;
5057 	uint32_t ctrl;
5058 	uint64_t ctrl1;
5059 	int ndesc, checkwrap;
5060 	struct mbuf *m;
5061 	void *flitp;
5062 
5063 	TXQ_LOCK_ASSERT_OWNED(txq);
5064 	MPASS(txp->npkt > 0);
5065 	MPASS(txp->plen < 65536);
5066 	MPASS(m0 != NULL);
5067 	MPASS(m0->m_nextpkt != NULL);
5068 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
5069 	MPASS(available > 0 && available < eq->sidx);
5070 
5071 	ndesc = howmany(txp->len16, EQ_ESIZE / 16);
5072 	MPASS(ndesc <= available);
5073 
5074 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
5075 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
5076 	ctrl = V_FW_WR_LEN16(txp->len16);
5077 	wr->equiq_to_len16 = htobe32(ctrl);
5078 	wr->plen = htobe16(txp->plen);
5079 	wr->npkt = txp->npkt;
5080 	wr->r3 = 0;
5081 	wr->type = txp->wr_type;
5082 	flitp = wr + 1;
5083 
5084 	/*
5085 	 * At this point we are 16B into a hardware descriptor.  If checkwrap is
5086 	 * set then we know the WR is going to wrap around somewhere.  We'll
5087 	 * check for that at appropriate points.
5088 	 */
5089 	checkwrap = eq->sidx - ndesc < eq->pidx;
5090 	for (m = m0; m != NULL; m = m->m_nextpkt) {
5091 		if (txp->wr_type == 0) {
5092 			struct ulp_txpkt *ulpmc;
5093 			struct ulptx_idata *ulpsc;
5094 
5095 			/* ULP master command */
5096 			ulpmc = flitp;
5097 			ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
5098 			    V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid));
5099 			ulpmc->len = htobe32(mbuf_len16(m));
5100 
5101 			/* ULP subcommand */
5102 			ulpsc = (void *)(ulpmc + 1);
5103 			ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
5104 			    F_ULP_TX_SC_MORE);
5105 			ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
5106 
5107 			cpl = (void *)(ulpsc + 1);
5108 			if (checkwrap &&
5109 			    (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx])
5110 				cpl = (void *)&eq->desc[0];
5111 		} else {
5112 			cpl = flitp;
5113 		}
5114 
5115 		/* Checksum offload */
5116 		ctrl1 = csum_to_ctrl(sc, m);
5117 		if (ctrl1 != (F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS))
5118 			txq->txcsum++;	/* some hardware assistance provided */
5119 
5120 		/* VLAN tag insertion */
5121 		if (needs_vlan_insertion(m)) {
5122 			ctrl1 |= F_TXPKT_VLAN_VLD |
5123 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
5124 			txq->vlan_insertion++;
5125 		}
5126 
5127 		/* CPL header */
5128 		cpl->ctrl0 = txq->cpl_ctrl0;
5129 		cpl->pack = 0;
5130 		cpl->len = htobe16(m->m_pkthdr.len);
5131 		cpl->ctrl1 = htobe64(ctrl1);
5132 
5133 		flitp = cpl + 1;
5134 		if (checkwrap &&
5135 		    (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
5136 			flitp = (void *)&eq->desc[0];
5137 
5138 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap);
5139 
5140 	}
5141 
5142 	if (txp->wr_type == 0) {
5143 		txq->txpkts0_pkts += txp->npkt;
5144 		txq->txpkts0_wrs++;
5145 	} else {
5146 		txq->txpkts1_pkts += txp->npkt;
5147 		txq->txpkts1_wrs++;
5148 	}
5149 
5150 	txsd = &txq->sdesc[eq->pidx];
5151 	txsd->m = m0;
5152 	txsd->desc_used = ndesc;
5153 
5154 	return (ndesc);
5155 }
5156 
5157 /*
5158  * If the SGL ends on an address that is not 16 byte aligned, this function will
5159  * add a 0 filled flit at the end.
5160  */
5161 static void
5162 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap)
5163 {
5164 	struct sge_eq *eq = &txq->eq;
5165 	struct sglist *gl = txq->gl;
5166 	struct sglist_seg *seg;
5167 	__be64 *flitp, *wrap;
5168 	struct ulptx_sgl *usgl;
5169 	int i, nflits, nsegs;
5170 
5171 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
5172 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
5173 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
5174 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
5175 
5176 	get_pkt_gl(m, gl);
5177 	nsegs = gl->sg_nseg;
5178 	MPASS(nsegs > 0);
5179 
5180 	nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
5181 	flitp = (__be64 *)(*to);
5182 	wrap = (__be64 *)(&eq->desc[eq->sidx]);
5183 	seg = &gl->sg_segs[0];
5184 	usgl = (void *)flitp;
5185 
5186 	/*
5187 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
5188 	 * ring, so we're at least 16 bytes away from the status page.  There is
5189 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
5190 	 */
5191 
5192 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
5193 	    V_ULPTX_NSGE(nsegs));
5194 	usgl->len0 = htobe32(seg->ss_len);
5195 	usgl->addr0 = htobe64(seg->ss_paddr);
5196 	seg++;
5197 
5198 	if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) {
5199 
5200 		/* Won't wrap around at all */
5201 
5202 		for (i = 0; i < nsegs - 1; i++, seg++) {
5203 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
5204 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
5205 		}
5206 		if (i & 1)
5207 			usgl->sge[i / 2].len[1] = htobe32(0);
5208 		flitp += nflits;
5209 	} else {
5210 
5211 		/* Will wrap somewhere in the rest of the SGL */
5212 
5213 		/* 2 flits already written, write the rest flit by flit */
5214 		flitp = (void *)(usgl + 1);
5215 		for (i = 0; i < nflits - 2; i++) {
5216 			if (flitp == wrap)
5217 				flitp = (void *)eq->desc;
5218 			*flitp++ = get_flit(seg, nsegs - 1, i);
5219 		}
5220 	}
5221 
5222 	if (nflits & 1) {
5223 		MPASS(((uintptr_t)flitp) & 0xf);
5224 		*flitp++ = 0;
5225 	}
5226 
5227 	MPASS((((uintptr_t)flitp) & 0xf) == 0);
5228 	if (__predict_false(flitp == wrap))
5229 		*to = (void *)eq->desc;
5230 	else
5231 		*to = (void *)flitp;
5232 }
5233 
5234 static inline void
5235 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
5236 {
5237 
5238 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
5239 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
5240 
5241 	if (__predict_true((uintptr_t)(*to) + len <=
5242 	    (uintptr_t)&eq->desc[eq->sidx])) {
5243 		bcopy(from, *to, len);
5244 		(*to) += len;
5245 	} else {
5246 		int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
5247 
5248 		bcopy(from, *to, portion);
5249 		from += portion;
5250 		portion = len - portion;	/* remaining */
5251 		bcopy(from, (void *)eq->desc, portion);
5252 		(*to) = (caddr_t)eq->desc + portion;
5253 	}
5254 }
5255 
5256 static inline void
5257 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n)
5258 {
5259 	u_int db;
5260 
5261 	MPASS(n > 0);
5262 
5263 	db = eq->doorbells;
5264 	if (n > 1)
5265 		clrbit(&db, DOORBELL_WCWR);
5266 	wmb();
5267 
5268 	switch (ffs(db) - 1) {
5269 	case DOORBELL_UDB:
5270 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
5271 		break;
5272 
5273 	case DOORBELL_WCWR: {
5274 		volatile uint64_t *dst, *src;
5275 		int i;
5276 
5277 		/*
5278 		 * Queues whose 128B doorbell segment fits in the page do not
5279 		 * use relative qid (udb_qid is always 0).  Only queues with
5280 		 * doorbell segments can do WCWR.
5281 		 */
5282 		KASSERT(eq->udb_qid == 0 && n == 1,
5283 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
5284 		    __func__, eq->doorbells, n, eq->dbidx, eq));
5285 
5286 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
5287 		    UDBS_DB_OFFSET);
5288 		i = eq->dbidx;
5289 		src = (void *)&eq->desc[i];
5290 		while (src != (void *)&eq->desc[i + 1])
5291 			*dst++ = *src++;
5292 		wmb();
5293 		break;
5294 	}
5295 
5296 	case DOORBELL_UDBWC:
5297 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
5298 		wmb();
5299 		break;
5300 
5301 	case DOORBELL_KDB:
5302 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
5303 		    V_QID(eq->cntxt_id) | V_PIDX(n));
5304 		break;
5305 	}
5306 
5307 	IDXINCR(eq->dbidx, n, eq->sidx);
5308 }
5309 
5310 static inline u_int
5311 reclaimable_tx_desc(struct sge_eq *eq)
5312 {
5313 	uint16_t hw_cidx;
5314 
5315 	hw_cidx = read_hw_cidx(eq);
5316 	return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx));
5317 }
5318 
5319 static inline u_int
5320 total_available_tx_desc(struct sge_eq *eq)
5321 {
5322 	uint16_t hw_cidx, pidx;
5323 
5324 	hw_cidx = read_hw_cidx(eq);
5325 	pidx = eq->pidx;
5326 
5327 	if (pidx == hw_cidx)
5328 		return (eq->sidx - 1);
5329 	else
5330 		return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1);
5331 }
5332 
5333 static inline uint16_t
5334 read_hw_cidx(struct sge_eq *eq)
5335 {
5336 	struct sge_qstat *spg = (void *)&eq->desc[eq->sidx];
5337 	uint16_t cidx = spg->cidx;	/* stable snapshot */
5338 
5339 	return (be16toh(cidx));
5340 }
5341 
5342 /*
5343  * Reclaim 'n' descriptors approximately.
5344  */
5345 static u_int
5346 reclaim_tx_descs(struct sge_txq *txq, u_int n)
5347 {
5348 	struct tx_sdesc *txsd;
5349 	struct sge_eq *eq = &txq->eq;
5350 	u_int can_reclaim, reclaimed;
5351 
5352 	TXQ_LOCK_ASSERT_OWNED(txq);
5353 	MPASS(n > 0);
5354 
5355 	reclaimed = 0;
5356 	can_reclaim = reclaimable_tx_desc(eq);
5357 	while (can_reclaim && reclaimed < n) {
5358 		int ndesc;
5359 		struct mbuf *m, *nextpkt;
5360 
5361 		txsd = &txq->sdesc[eq->cidx];
5362 		ndesc = txsd->desc_used;
5363 
5364 		/* Firmware doesn't return "partial" credits. */
5365 		KASSERT(can_reclaim >= ndesc,
5366 		    ("%s: unexpected number of credits: %d, %d",
5367 		    __func__, can_reclaim, ndesc));
5368 		KASSERT(ndesc != 0,
5369 		    ("%s: descriptor with no credits: cidx %d",
5370 		    __func__, eq->cidx));
5371 
5372 		for (m = txsd->m; m != NULL; m = nextpkt) {
5373 			nextpkt = m->m_nextpkt;
5374 			m->m_nextpkt = NULL;
5375 			m_freem(m);
5376 		}
5377 		reclaimed += ndesc;
5378 		can_reclaim -= ndesc;
5379 		IDXINCR(eq->cidx, ndesc, eq->sidx);
5380 	}
5381 
5382 	return (reclaimed);
5383 }
5384 
5385 static void
5386 tx_reclaim(void *arg, int n)
5387 {
5388 	struct sge_txq *txq = arg;
5389 	struct sge_eq *eq = &txq->eq;
5390 
5391 	do {
5392 		if (TXQ_TRYLOCK(txq) == 0)
5393 			break;
5394 		n = reclaim_tx_descs(txq, 32);
5395 		if (eq->cidx == eq->pidx)
5396 			eq->equeqidx = eq->pidx;
5397 		TXQ_UNLOCK(txq);
5398 	} while (n > 0);
5399 }
5400 
5401 static __be64
5402 get_flit(struct sglist_seg *segs, int nsegs, int idx)
5403 {
5404 	int i = (idx / 3) * 2;
5405 
5406 	switch (idx % 3) {
5407 	case 0: {
5408 		uint64_t rc;
5409 
5410 		rc = (uint64_t)segs[i].ss_len << 32;
5411 		if (i + 1 < nsegs)
5412 			rc |= (uint64_t)(segs[i + 1].ss_len);
5413 
5414 		return (htobe64(rc));
5415 	}
5416 	case 1:
5417 		return (htobe64(segs[i].ss_paddr));
5418 	case 2:
5419 		return (htobe64(segs[i + 1].ss_paddr));
5420 	}
5421 
5422 	return (0);
5423 }
5424 
5425 static void
5426 find_best_refill_source(struct adapter *sc, struct sge_fl *fl, int maxp)
5427 {
5428 	int8_t zidx, hwidx, idx;
5429 	uint16_t region1, region3;
5430 	int spare, spare_needed, n;
5431 	struct sw_zone_info *swz;
5432 	struct hw_buf_info *hwb, *hwb_list = &sc->sge.hw_buf_info[0];
5433 
5434 	/*
5435 	 * Buffer Packing: Look for PAGE_SIZE or larger zone which has a bufsize
5436 	 * large enough for the max payload and cluster metadata.  Otherwise
5437 	 * settle for the largest bufsize that leaves enough room in the cluster
5438 	 * for metadata.
5439 	 *
5440 	 * Without buffer packing: Look for the smallest zone which has a
5441 	 * bufsize large enough for the max payload.  Settle for the largest
5442 	 * bufsize available if there's nothing big enough for max payload.
5443 	 */
5444 	spare_needed = fl->flags & FL_BUF_PACKING ? CL_METADATA_SIZE : 0;
5445 	swz = &sc->sge.sw_zone_info[0];
5446 	hwidx = -1;
5447 	for (zidx = 0; zidx < SW_ZONE_SIZES; zidx++, swz++) {
5448 		if (swz->size > largest_rx_cluster) {
5449 			if (__predict_true(hwidx != -1))
5450 				break;
5451 
5452 			/*
5453 			 * This is a misconfiguration.  largest_rx_cluster is
5454 			 * preventing us from finding a refill source.  See
5455 			 * dev.t5nex.<n>.buffer_sizes to figure out why.
5456 			 */
5457 			device_printf(sc->dev, "largest_rx_cluster=%u leaves no"
5458 			    " refill source for fl %p (dma %u).  Ignored.\n",
5459 			    largest_rx_cluster, fl, maxp);
5460 		}
5461 		for (idx = swz->head_hwidx; idx != -1; idx = hwb->next) {
5462 			hwb = &hwb_list[idx];
5463 			spare = swz->size - hwb->size;
5464 			if (spare < spare_needed)
5465 				continue;
5466 
5467 			hwidx = idx;		/* best option so far */
5468 			if (hwb->size >= maxp) {
5469 
5470 				if ((fl->flags & FL_BUF_PACKING) == 0)
5471 					goto done; /* stop looking (not packing) */
5472 
5473 				if (swz->size >= safest_rx_cluster)
5474 					goto done; /* stop looking (packing) */
5475 			}
5476 			break;		/* keep looking, next zone */
5477 		}
5478 	}
5479 done:
5480 	/* A usable hwidx has been located. */
5481 	MPASS(hwidx != -1);
5482 	hwb = &hwb_list[hwidx];
5483 	zidx = hwb->zidx;
5484 	swz = &sc->sge.sw_zone_info[zidx];
5485 	region1 = 0;
5486 	region3 = swz->size - hwb->size;
5487 
5488 	/*
5489 	 * Stay within this zone and see if there is a better match when mbuf
5490 	 * inlining is allowed.  Remember that the hwidx's are sorted in
5491 	 * decreasing order of size (so in increasing order of spare area).
5492 	 */
5493 	for (idx = hwidx; idx != -1; idx = hwb->next) {
5494 		hwb = &hwb_list[idx];
5495 		spare = swz->size - hwb->size;
5496 
5497 		if (allow_mbufs_in_cluster == 0 || hwb->size < maxp)
5498 			break;
5499 
5500 		/*
5501 		 * Do not inline mbufs if doing so would violate the pad/pack
5502 		 * boundary alignment requirement.
5503 		 */
5504 		if (fl_pad && (MSIZE % sc->params.sge.pad_boundary) != 0)
5505 			continue;
5506 		if (fl->flags & FL_BUF_PACKING &&
5507 		    (MSIZE % sc->params.sge.pack_boundary) != 0)
5508 			continue;
5509 
5510 		if (spare < CL_METADATA_SIZE + MSIZE)
5511 			continue;
5512 		n = (spare - CL_METADATA_SIZE) / MSIZE;
5513 		if (n > howmany(hwb->size, maxp))
5514 			break;
5515 
5516 		hwidx = idx;
5517 		if (fl->flags & FL_BUF_PACKING) {
5518 			region1 = n * MSIZE;
5519 			region3 = spare - region1;
5520 		} else {
5521 			region1 = MSIZE;
5522 			region3 = spare - region1;
5523 			break;
5524 		}
5525 	}
5526 
5527 	KASSERT(zidx >= 0 && zidx < SW_ZONE_SIZES,
5528 	    ("%s: bad zone %d for fl %p, maxp %d", __func__, zidx, fl, maxp));
5529 	KASSERT(hwidx >= 0 && hwidx <= SGE_FLBUF_SIZES,
5530 	    ("%s: bad hwidx %d for fl %p, maxp %d", __func__, hwidx, fl, maxp));
5531 	KASSERT(region1 + sc->sge.hw_buf_info[hwidx].size + region3 ==
5532 	    sc->sge.sw_zone_info[zidx].size,
5533 	    ("%s: bad buffer layout for fl %p, maxp %d. "
5534 		"cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5535 		sc->sge.sw_zone_info[zidx].size, region1,
5536 		sc->sge.hw_buf_info[hwidx].size, region3));
5537 	if (fl->flags & FL_BUF_PACKING || region1 > 0) {
5538 		KASSERT(region3 >= CL_METADATA_SIZE,
5539 		    ("%s: no room for metadata.  fl %p, maxp %d; "
5540 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5541 		    sc->sge.sw_zone_info[zidx].size, region1,
5542 		    sc->sge.hw_buf_info[hwidx].size, region3));
5543 		KASSERT(region1 % MSIZE == 0,
5544 		    ("%s: bad mbuf region for fl %p, maxp %d. "
5545 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5546 		    sc->sge.sw_zone_info[zidx].size, region1,
5547 		    sc->sge.hw_buf_info[hwidx].size, region3));
5548 	}
5549 
5550 	fl->cll_def.zidx = zidx;
5551 	fl->cll_def.hwidx = hwidx;
5552 	fl->cll_def.region1 = region1;
5553 	fl->cll_def.region3 = region3;
5554 }
5555 
5556 static void
5557 find_safe_refill_source(struct adapter *sc, struct sge_fl *fl)
5558 {
5559 	struct sge *s = &sc->sge;
5560 	struct hw_buf_info *hwb;
5561 	struct sw_zone_info *swz;
5562 	int spare;
5563 	int8_t hwidx;
5564 
5565 	if (fl->flags & FL_BUF_PACKING)
5566 		hwidx = s->safe_hwidx2;	/* with room for metadata */
5567 	else if (allow_mbufs_in_cluster && s->safe_hwidx2 != -1) {
5568 		hwidx = s->safe_hwidx2;
5569 		hwb = &s->hw_buf_info[hwidx];
5570 		swz = &s->sw_zone_info[hwb->zidx];
5571 		spare = swz->size - hwb->size;
5572 
5573 		/* no good if there isn't room for an mbuf as well */
5574 		if (spare < CL_METADATA_SIZE + MSIZE)
5575 			hwidx = s->safe_hwidx1;
5576 	} else
5577 		hwidx = s->safe_hwidx1;
5578 
5579 	if (hwidx == -1) {
5580 		/* No fallback source */
5581 		fl->cll_alt.hwidx = -1;
5582 		fl->cll_alt.zidx = -1;
5583 
5584 		return;
5585 	}
5586 
5587 	hwb = &s->hw_buf_info[hwidx];
5588 	swz = &s->sw_zone_info[hwb->zidx];
5589 	spare = swz->size - hwb->size;
5590 	fl->cll_alt.hwidx = hwidx;
5591 	fl->cll_alt.zidx = hwb->zidx;
5592 	if (allow_mbufs_in_cluster &&
5593 	    (fl_pad == 0 || (MSIZE % sc->params.sge.pad_boundary) == 0))
5594 		fl->cll_alt.region1 = ((spare - CL_METADATA_SIZE) / MSIZE) * MSIZE;
5595 	else
5596 		fl->cll_alt.region1 = 0;
5597 	fl->cll_alt.region3 = spare - fl->cll_alt.region1;
5598 }
5599 
5600 static void
5601 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
5602 {
5603 	mtx_lock(&sc->sfl_lock);
5604 	FL_LOCK(fl);
5605 	if ((fl->flags & FL_DOOMED) == 0) {
5606 		fl->flags |= FL_STARVING;
5607 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
5608 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
5609 	}
5610 	FL_UNLOCK(fl);
5611 	mtx_unlock(&sc->sfl_lock);
5612 }
5613 
5614 static void
5615 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq)
5616 {
5617 	struct sge_wrq *wrq = (void *)eq;
5618 
5619 	atomic_readandclear_int(&eq->equiq);
5620 	taskqueue_enqueue(sc->tq[eq->tx_chan], &wrq->wrq_tx_task);
5621 }
5622 
5623 static void
5624 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq)
5625 {
5626 	struct sge_txq *txq = (void *)eq;
5627 
5628 	MPASS((eq->flags & EQ_TYPEMASK) == EQ_ETH);
5629 
5630 	atomic_readandclear_int(&eq->equiq);
5631 	mp_ring_check_drainage(txq->r, 0);
5632 	taskqueue_enqueue(sc->tq[eq->tx_chan], &txq->tx_reclaim_task);
5633 }
5634 
5635 static int
5636 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
5637     struct mbuf *m)
5638 {
5639 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
5640 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
5641 	struct adapter *sc = iq->adapter;
5642 	struct sge *s = &sc->sge;
5643 	struct sge_eq *eq;
5644 	static void (*h[])(struct adapter *, struct sge_eq *) = {NULL,
5645 		&handle_wrq_egr_update, &handle_eth_egr_update,
5646 		&handle_wrq_egr_update};
5647 
5648 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5649 	    rss->opcode));
5650 
5651 	eq = s->eqmap[qid - s->eq_start - s->eq_base];
5652 	(*h[eq->flags & EQ_TYPEMASK])(sc, eq);
5653 
5654 	return (0);
5655 }
5656 
5657 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
5658 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
5659     offsetof(struct cpl_fw6_msg, data));
5660 
5661 static int
5662 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
5663 {
5664 	struct adapter *sc = iq->adapter;
5665 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
5666 
5667 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5668 	    rss->opcode));
5669 
5670 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
5671 		const struct rss_header *rss2;
5672 
5673 		rss2 = (const struct rss_header *)&cpl->data[0];
5674 		return (t4_cpl_handler[rss2->opcode](iq, rss2, m));
5675 	}
5676 
5677 	return (t4_fw_msg_handler[cpl->type](sc, &cpl->data[0]));
5678 }
5679 
5680 /**
5681  *	t4_handle_wrerr_rpl - process a FW work request error message
5682  *	@adap: the adapter
5683  *	@rpl: start of the FW message
5684  */
5685 static int
5686 t4_handle_wrerr_rpl(struct adapter *adap, const __be64 *rpl)
5687 {
5688 	u8 opcode = *(const u8 *)rpl;
5689 	const struct fw_error_cmd *e = (const void *)rpl;
5690 	unsigned int i;
5691 
5692 	if (opcode != FW_ERROR_CMD) {
5693 		log(LOG_ERR,
5694 		    "%s: Received WRERR_RPL message with opcode %#x\n",
5695 		    device_get_nameunit(adap->dev), opcode);
5696 		return (EINVAL);
5697 	}
5698 	log(LOG_ERR, "%s: FW_ERROR (%s) ", device_get_nameunit(adap->dev),
5699 	    G_FW_ERROR_CMD_FATAL(be32toh(e->op_to_type)) ? "fatal" :
5700 	    "non-fatal");
5701 	switch (G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type))) {
5702 	case FW_ERROR_TYPE_EXCEPTION:
5703 		log(LOG_ERR, "exception info:\n");
5704 		for (i = 0; i < nitems(e->u.exception.info); i++)
5705 			log(LOG_ERR, "%s%08x", i == 0 ? "\t" : " ",
5706 			    be32toh(e->u.exception.info[i]));
5707 		log(LOG_ERR, "\n");
5708 		break;
5709 	case FW_ERROR_TYPE_HWMODULE:
5710 		log(LOG_ERR, "HW module regaddr %08x regval %08x\n",
5711 		    be32toh(e->u.hwmodule.regaddr),
5712 		    be32toh(e->u.hwmodule.regval));
5713 		break;
5714 	case FW_ERROR_TYPE_WR:
5715 		log(LOG_ERR, "WR cidx %d PF %d VF %d eqid %d hdr:\n",
5716 		    be16toh(e->u.wr.cidx),
5717 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.wr.pfn_vfn)),
5718 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.wr.pfn_vfn)),
5719 		    be32toh(e->u.wr.eqid));
5720 		for (i = 0; i < nitems(e->u.wr.wrhdr); i++)
5721 			log(LOG_ERR, "%s%02x", i == 0 ? "\t" : " ",
5722 			    e->u.wr.wrhdr[i]);
5723 		log(LOG_ERR, "\n");
5724 		break;
5725 	case FW_ERROR_TYPE_ACL:
5726 		log(LOG_ERR, "ACL cidx %d PF %d VF %d eqid %d %s",
5727 		    be16toh(e->u.acl.cidx),
5728 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.acl.pfn_vfn)),
5729 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.acl.pfn_vfn)),
5730 		    be32toh(e->u.acl.eqid),
5731 		    G_FW_ERROR_CMD_MV(be16toh(e->u.acl.mv_pkd)) ? "vlanid" :
5732 		    "MAC");
5733 		for (i = 0; i < nitems(e->u.acl.val); i++)
5734 			log(LOG_ERR, " %02x", e->u.acl.val[i]);
5735 		log(LOG_ERR, "\n");
5736 		break;
5737 	default:
5738 		log(LOG_ERR, "type %#x\n",
5739 		    G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type)));
5740 		return (EINVAL);
5741 	}
5742 	return (0);
5743 }
5744 
5745 static int
5746 sysctl_uint16(SYSCTL_HANDLER_ARGS)
5747 {
5748 	uint16_t *id = arg1;
5749 	int i = *id;
5750 
5751 	return sysctl_handle_int(oidp, &i, 0, req);
5752 }
5753 
5754 static int
5755 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
5756 {
5757 	struct sge *s = arg1;
5758 	struct hw_buf_info *hwb = &s->hw_buf_info[0];
5759 	struct sw_zone_info *swz = &s->sw_zone_info[0];
5760 	int i, rc;
5761 	struct sbuf sb;
5762 	char c;
5763 
5764 	sbuf_new(&sb, NULL, 32, SBUF_AUTOEXTEND);
5765 	for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
5766 		if (hwb->zidx >= 0 && swz[hwb->zidx].size <= largest_rx_cluster)
5767 			c = '*';
5768 		else
5769 			c = '\0';
5770 
5771 		sbuf_printf(&sb, "%u%c ", hwb->size, c);
5772 	}
5773 	sbuf_trim(&sb);
5774 	sbuf_finish(&sb);
5775 	rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
5776 	sbuf_delete(&sb);
5777 	return (rc);
5778 }
5779 
5780 #ifdef RATELIMIT
5781 /*
5782  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
5783  */
5784 static inline u_int
5785 txpkt_eo_len16(u_int nsegs, u_int immhdrs, u_int tso)
5786 {
5787 	u_int n;
5788 
5789 	MPASS(immhdrs > 0);
5790 
5791 	n = roundup2(sizeof(struct fw_eth_tx_eo_wr) +
5792 	    sizeof(struct cpl_tx_pkt_core) + immhdrs, 16);
5793 	if (__predict_false(nsegs == 0))
5794 		goto done;
5795 
5796 	nsegs--; /* first segment is part of ulptx_sgl */
5797 	n += sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
5798 	if (tso)
5799 		n += sizeof(struct cpl_tx_pkt_lso_core);
5800 
5801 done:
5802 	return (howmany(n, 16));
5803 }
5804 
5805 #define ETID_FLOWC_NPARAMS 6
5806 #define ETID_FLOWC_LEN (roundup2((sizeof(struct fw_flowc_wr) + \
5807     ETID_FLOWC_NPARAMS * sizeof(struct fw_flowc_mnemval)), 16))
5808 #define ETID_FLOWC_LEN16 (howmany(ETID_FLOWC_LEN, 16))
5809 
5810 static int
5811 send_etid_flowc_wr(struct cxgbe_rate_tag *cst, struct port_info *pi,
5812     struct vi_info *vi)
5813 {
5814 	struct wrq_cookie cookie;
5815 	u_int pfvf = pi->adapter->pf << S_FW_VIID_PFN;
5816 	struct fw_flowc_wr *flowc;
5817 
5818 	mtx_assert(&cst->lock, MA_OWNED);
5819 	MPASS((cst->flags & (EO_FLOWC_PENDING | EO_FLOWC_RPL_PENDING)) ==
5820 	    EO_FLOWC_PENDING);
5821 
5822 	flowc = start_wrq_wr(cst->eo_txq, ETID_FLOWC_LEN16, &cookie);
5823 	if (__predict_false(flowc == NULL))
5824 		return (ENOMEM);
5825 
5826 	bzero(flowc, ETID_FLOWC_LEN);
5827 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
5828 	    V_FW_FLOWC_WR_NPARAMS(ETID_FLOWC_NPARAMS) | V_FW_WR_COMPL(0));
5829 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(ETID_FLOWC_LEN16) |
5830 	    V_FW_WR_FLOWID(cst->etid));
5831 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN;
5832 	flowc->mnemval[0].val = htobe32(pfvf);
5833 	flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH;
5834 	flowc->mnemval[1].val = htobe32(pi->tx_chan);
5835 	flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT;
5836 	flowc->mnemval[2].val = htobe32(pi->tx_chan);
5837 	flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID;
5838 	flowc->mnemval[3].val = htobe32(cst->iqid);
5839 	flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_EOSTATE;
5840 	flowc->mnemval[4].val = htobe32(FW_FLOWC_MNEM_EOSTATE_ESTABLISHED);
5841 	flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS;
5842 	flowc->mnemval[5].val = htobe32(cst->schedcl);
5843 
5844 	commit_wrq_wr(cst->eo_txq, flowc, &cookie);
5845 
5846 	cst->flags &= ~EO_FLOWC_PENDING;
5847 	cst->flags |= EO_FLOWC_RPL_PENDING;
5848 	MPASS(cst->tx_credits >= ETID_FLOWC_LEN16);	/* flowc is first WR. */
5849 	cst->tx_credits -= ETID_FLOWC_LEN16;
5850 
5851 	return (0);
5852 }
5853 
5854 #define ETID_FLUSH_LEN16 (howmany(sizeof (struct fw_flowc_wr), 16))
5855 
5856 void
5857 send_etid_flush_wr(struct cxgbe_rate_tag *cst)
5858 {
5859 	struct fw_flowc_wr *flowc;
5860 	struct wrq_cookie cookie;
5861 
5862 	mtx_assert(&cst->lock, MA_OWNED);
5863 
5864 	flowc = start_wrq_wr(cst->eo_txq, ETID_FLUSH_LEN16, &cookie);
5865 	if (__predict_false(flowc == NULL))
5866 		CXGBE_UNIMPLEMENTED(__func__);
5867 
5868 	bzero(flowc, ETID_FLUSH_LEN16 * 16);
5869 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
5870 	    V_FW_FLOWC_WR_NPARAMS(0) | F_FW_WR_COMPL);
5871 	flowc->flowid_len16 = htobe32(V_FW_WR_LEN16(ETID_FLUSH_LEN16) |
5872 	    V_FW_WR_FLOWID(cst->etid));
5873 
5874 	commit_wrq_wr(cst->eo_txq, flowc, &cookie);
5875 
5876 	cst->flags |= EO_FLUSH_RPL_PENDING;
5877 	MPASS(cst->tx_credits >= ETID_FLUSH_LEN16);
5878 	cst->tx_credits -= ETID_FLUSH_LEN16;
5879 	cst->ncompl++;
5880 }
5881 
5882 static void
5883 write_ethofld_wr(struct cxgbe_rate_tag *cst, struct fw_eth_tx_eo_wr *wr,
5884     struct mbuf *m0, int compl)
5885 {
5886 	struct cpl_tx_pkt_core *cpl;
5887 	uint64_t ctrl1;
5888 	uint32_t ctrl;	/* used in many unrelated places */
5889 	int len16, pktlen, nsegs, immhdrs;
5890 	caddr_t dst;
5891 	uintptr_t p;
5892 	struct ulptx_sgl *usgl;
5893 	struct sglist sg;
5894 	struct sglist_seg segs[38];	/* XXX: find real limit.  XXX: get off the stack */
5895 
5896 	mtx_assert(&cst->lock, MA_OWNED);
5897 	M_ASSERTPKTHDR(m0);
5898 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
5899 	    m0->m_pkthdr.l4hlen > 0,
5900 	    ("%s: ethofld mbuf %p is missing header lengths", __func__, m0));
5901 
5902 	len16 = mbuf_eo_len16(m0);
5903 	nsegs = mbuf_eo_nsegs(m0);
5904 	pktlen = m0->m_pkthdr.len;
5905 	ctrl = sizeof(struct cpl_tx_pkt_core);
5906 	if (needs_tso(m0))
5907 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
5908 	immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen + m0->m_pkthdr.l4hlen;
5909 	ctrl += immhdrs;
5910 
5911 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_EO_WR) |
5912 	    V_FW_ETH_TX_EO_WR_IMMDLEN(ctrl) | V_FW_WR_COMPL(!!compl));
5913 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(len16) |
5914 	    V_FW_WR_FLOWID(cst->etid));
5915 	wr->r3 = 0;
5916 	if (needs_udp_csum(m0)) {
5917 		wr->u.udpseg.type = FW_ETH_TX_EO_TYPE_UDPSEG;
5918 		wr->u.udpseg.ethlen = m0->m_pkthdr.l2hlen;
5919 		wr->u.udpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
5920 		wr->u.udpseg.udplen = m0->m_pkthdr.l4hlen;
5921 		wr->u.udpseg.rtplen = 0;
5922 		wr->u.udpseg.r4 = 0;
5923 		wr->u.udpseg.mss = htobe16(pktlen - immhdrs);
5924 		wr->u.udpseg.schedpktsize = wr->u.udpseg.mss;
5925 		wr->u.udpseg.plen = htobe32(pktlen - immhdrs);
5926 		cpl = (void *)(wr + 1);
5927 	} else {
5928 		MPASS(needs_tcp_csum(m0));
5929 		wr->u.tcpseg.type = FW_ETH_TX_EO_TYPE_TCPSEG;
5930 		wr->u.tcpseg.ethlen = m0->m_pkthdr.l2hlen;
5931 		wr->u.tcpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
5932 		wr->u.tcpseg.tcplen = m0->m_pkthdr.l4hlen;
5933 		wr->u.tcpseg.tsclk_tsoff = mbuf_eo_tsclk_tsoff(m0);
5934 		wr->u.tcpseg.r4 = 0;
5935 		wr->u.tcpseg.r5 = 0;
5936 		wr->u.tcpseg.plen = htobe32(pktlen - immhdrs);
5937 
5938 		if (needs_tso(m0)) {
5939 			struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
5940 
5941 			wr->u.tcpseg.mss = htobe16(m0->m_pkthdr.tso_segsz);
5942 
5943 			ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) |
5944 			    F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE |
5945 			    V_LSO_ETHHDR_LEN((m0->m_pkthdr.l2hlen -
5946 				ETHER_HDR_LEN) >> 2) |
5947 			    V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) |
5948 			    V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
5949 			if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
5950 				ctrl |= F_LSO_IPV6;
5951 			lso->lso_ctrl = htobe32(ctrl);
5952 			lso->ipid_ofst = htobe16(0);
5953 			lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
5954 			lso->seqno_offset = htobe32(0);
5955 			lso->len = htobe32(pktlen);
5956 
5957 			cpl = (void *)(lso + 1);
5958 		} else {
5959 			wr->u.tcpseg.mss = htobe16(0xffff);
5960 			cpl = (void *)(wr + 1);
5961 		}
5962 	}
5963 
5964 	/* Checksum offload must be requested for ethofld. */
5965 	MPASS(needs_l4_csum(m0));
5966 	ctrl1 = csum_to_ctrl(cst->adapter, m0);
5967 
5968 	/* VLAN tag insertion */
5969 	if (needs_vlan_insertion(m0)) {
5970 		ctrl1 |= F_TXPKT_VLAN_VLD |
5971 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
5972 	}
5973 
5974 	/* CPL header */
5975 	cpl->ctrl0 = cst->ctrl0;
5976 	cpl->pack = 0;
5977 	cpl->len = htobe16(pktlen);
5978 	cpl->ctrl1 = htobe64(ctrl1);
5979 
5980 	/* Copy Ethernet, IP & TCP/UDP hdrs as immediate data */
5981 	p = (uintptr_t)(cpl + 1);
5982 	m_copydata(m0, 0, immhdrs, (void *)p);
5983 
5984 	/* SGL */
5985 	dst = (void *)(cpl + 1);
5986 	if (nsegs > 0) {
5987 		int i, pad;
5988 
5989 		/* zero-pad upto next 16Byte boundary, if not 16Byte aligned */
5990 		p += immhdrs;
5991 		pad = 16 - (immhdrs & 0xf);
5992 		bzero((void *)p, pad);
5993 
5994 		usgl = (void *)(p + pad);
5995 		usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
5996 		    V_ULPTX_NSGE(nsegs));
5997 
5998 		sglist_init(&sg, nitems(segs), segs);
5999 		for (; m0 != NULL; m0 = m0->m_next) {
6000 			if (__predict_false(m0->m_len == 0))
6001 				continue;
6002 			if (immhdrs >= m0->m_len) {
6003 				immhdrs -= m0->m_len;
6004 				continue;
6005 			}
6006 
6007 			sglist_append(&sg, mtod(m0, char *) + immhdrs,
6008 			    m0->m_len - immhdrs);
6009 			immhdrs = 0;
6010 		}
6011 		MPASS(sg.sg_nseg == nsegs);
6012 
6013 		/*
6014 		 * Zero pad last 8B in case the WR doesn't end on a 16B
6015 		 * boundary.
6016 		 */
6017 		*(uint64_t *)((char *)wr + len16 * 16 - 8) = 0;
6018 
6019 		usgl->len0 = htobe32(segs[0].ss_len);
6020 		usgl->addr0 = htobe64(segs[0].ss_paddr);
6021 		for (i = 0; i < nsegs - 1; i++) {
6022 			usgl->sge[i / 2].len[i & 1] = htobe32(segs[i + 1].ss_len);
6023 			usgl->sge[i / 2].addr[i & 1] = htobe64(segs[i + 1].ss_paddr);
6024 		}
6025 		if (i & 1)
6026 			usgl->sge[i / 2].len[1] = htobe32(0);
6027 	}
6028 
6029 }
6030 
6031 static void
6032 ethofld_tx(struct cxgbe_rate_tag *cst)
6033 {
6034 	struct mbuf *m;
6035 	struct wrq_cookie cookie;
6036 	int next_credits, compl;
6037 	struct fw_eth_tx_eo_wr *wr;
6038 
6039 	mtx_assert(&cst->lock, MA_OWNED);
6040 
6041 	while ((m = mbufq_first(&cst->pending_tx)) != NULL) {
6042 		M_ASSERTPKTHDR(m);
6043 
6044 		/* How many len16 credits do we need to send this mbuf. */
6045 		next_credits = mbuf_eo_len16(m);
6046 		MPASS(next_credits > 0);
6047 		if (next_credits > cst->tx_credits) {
6048 			/*
6049 			 * Tx will make progress eventually because there is at
6050 			 * least one outstanding fw4_ack that will return
6051 			 * credits and kick the tx.
6052 			 */
6053 			MPASS(cst->ncompl > 0);
6054 			return;
6055 		}
6056 		wr = start_wrq_wr(cst->eo_txq, next_credits, &cookie);
6057 		if (__predict_false(wr == NULL)) {
6058 			/* XXX: wishful thinking, not a real assertion. */
6059 			MPASS(cst->ncompl > 0);
6060 			return;
6061 		}
6062 		cst->tx_credits -= next_credits;
6063 		cst->tx_nocompl += next_credits;
6064 		compl = cst->ncompl == 0 || cst->tx_nocompl >= cst->tx_total / 2;
6065 		ETHER_BPF_MTAP(cst->com.com.ifp, m);
6066 		write_ethofld_wr(cst, wr, m, compl);
6067 		commit_wrq_wr(cst->eo_txq, wr, &cookie);
6068 		if (compl) {
6069 			cst->ncompl++;
6070 			cst->tx_nocompl	= 0;
6071 		}
6072 		(void) mbufq_dequeue(&cst->pending_tx);
6073 
6074 		/*
6075 		 * Drop the mbuf's reference on the tag now rather
6076 		 * than waiting until m_freem().  This ensures that
6077 		 * cxgbe_rate_tag_free gets called when the inp drops
6078 		 * its reference on the tag and there are no more
6079 		 * mbufs in the pending_tx queue and can flush any
6080 		 * pending requests.  Otherwise if the last mbuf
6081 		 * doesn't request a completion the etid will never be
6082 		 * released.
6083 		 */
6084 		m->m_pkthdr.snd_tag = NULL;
6085 		m->m_pkthdr.csum_flags &= ~CSUM_SND_TAG;
6086 		m_snd_tag_rele(&cst->com.com);
6087 
6088 		mbufq_enqueue(&cst->pending_fwack, m);
6089 	}
6090 }
6091 
6092 int
6093 ethofld_transmit(struct ifnet *ifp, struct mbuf *m0)
6094 {
6095 	struct cxgbe_rate_tag *cst;
6096 	int rc;
6097 
6098 	MPASS(m0->m_nextpkt == NULL);
6099 	MPASS(m0->m_pkthdr.csum_flags & CSUM_SND_TAG);
6100 	MPASS(m0->m_pkthdr.snd_tag != NULL);
6101 	cst = mst_to_crt(m0->m_pkthdr.snd_tag);
6102 
6103 	mtx_lock(&cst->lock);
6104 	MPASS(cst->flags & EO_SND_TAG_REF);
6105 
6106 	if (__predict_false(cst->flags & EO_FLOWC_PENDING)) {
6107 		struct vi_info *vi = ifp->if_softc;
6108 		struct port_info *pi = vi->pi;
6109 		struct adapter *sc = pi->adapter;
6110 		const uint32_t rss_mask = vi->rss_size - 1;
6111 		uint32_t rss_hash;
6112 
6113 		cst->eo_txq = &sc->sge.ofld_txq[vi->first_ofld_txq];
6114 		if (M_HASHTYPE_ISHASH(m0))
6115 			rss_hash = m0->m_pkthdr.flowid;
6116 		else
6117 			rss_hash = arc4random();
6118 		/* We assume RSS hashing */
6119 		cst->iqid = vi->rss[rss_hash & rss_mask];
6120 		cst->eo_txq += rss_hash % vi->nofldtxq;
6121 		rc = send_etid_flowc_wr(cst, pi, vi);
6122 		if (rc != 0)
6123 			goto done;
6124 	}
6125 
6126 	if (__predict_false(cst->plen + m0->m_pkthdr.len > eo_max_backlog)) {
6127 		rc = ENOBUFS;
6128 		goto done;
6129 	}
6130 
6131 	mbufq_enqueue(&cst->pending_tx, m0);
6132 	cst->plen += m0->m_pkthdr.len;
6133 
6134 	/*
6135 	 * Hold an extra reference on the tag while generating work
6136 	 * requests to ensure that we don't try to free the tag during
6137 	 * ethofld_tx() in case we are sending the final mbuf after
6138 	 * the inp was freed.
6139 	 */
6140 	m_snd_tag_ref(&cst->com.com);
6141 	ethofld_tx(cst);
6142 	mtx_unlock(&cst->lock);
6143 	m_snd_tag_rele(&cst->com.com);
6144 	return (0);
6145 
6146 done:
6147 	mtx_unlock(&cst->lock);
6148 	if (__predict_false(rc != 0))
6149 		m_freem(m0);
6150 	return (rc);
6151 }
6152 
6153 static int
6154 ethofld_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
6155 {
6156 	struct adapter *sc = iq->adapter;
6157 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
6158 	struct mbuf *m;
6159 	u_int etid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
6160 	struct cxgbe_rate_tag *cst;
6161 	uint8_t credits = cpl->credits;
6162 
6163 	cst = lookup_etid(sc, etid);
6164 	mtx_lock(&cst->lock);
6165 	if (__predict_false(cst->flags & EO_FLOWC_RPL_PENDING)) {
6166 		MPASS(credits >= ETID_FLOWC_LEN16);
6167 		credits -= ETID_FLOWC_LEN16;
6168 		cst->flags &= ~EO_FLOWC_RPL_PENDING;
6169 	}
6170 
6171 	KASSERT(cst->ncompl > 0,
6172 	    ("%s: etid %u (%p) wasn't expecting completion.",
6173 	    __func__, etid, cst));
6174 	cst->ncompl--;
6175 
6176 	while (credits > 0) {
6177 		m = mbufq_dequeue(&cst->pending_fwack);
6178 		if (__predict_false(m == NULL)) {
6179 			/*
6180 			 * The remaining credits are for the final flush that
6181 			 * was issued when the tag was freed by the kernel.
6182 			 */
6183 			MPASS((cst->flags &
6184 			    (EO_FLUSH_RPL_PENDING | EO_SND_TAG_REF)) ==
6185 			    EO_FLUSH_RPL_PENDING);
6186 			MPASS(credits == ETID_FLUSH_LEN16);
6187 			MPASS(cst->tx_credits + cpl->credits == cst->tx_total);
6188 			MPASS(cst->ncompl == 0);
6189 
6190 			cst->flags &= ~EO_FLUSH_RPL_PENDING;
6191 			cst->tx_credits += cpl->credits;
6192 			cxgbe_rate_tag_free_locked(cst);
6193 			return (0);	/* cst is gone. */
6194 		}
6195 		KASSERT(m != NULL,
6196 		    ("%s: too many credits (%u, %u)", __func__, cpl->credits,
6197 		    credits));
6198 		KASSERT(credits >= mbuf_eo_len16(m),
6199 		    ("%s: too few credits (%u, %u, %u)", __func__,
6200 		    cpl->credits, credits, mbuf_eo_len16(m)));
6201 		credits -= mbuf_eo_len16(m);
6202 		cst->plen -= m->m_pkthdr.len;
6203 		m_freem(m);
6204 	}
6205 
6206 	cst->tx_credits += cpl->credits;
6207 	MPASS(cst->tx_credits <= cst->tx_total);
6208 
6209 	if (cst->flags & EO_SND_TAG_REF) {
6210 		/*
6211 		 * As with ethofld_transmit(), hold an extra reference
6212 		 * so that the tag is stable across ethold_tx().
6213 		 */
6214 		m_snd_tag_ref(&cst->com.com);
6215 		m = mbufq_first(&cst->pending_tx);
6216 		if (m != NULL && cst->tx_credits >= mbuf_eo_len16(m))
6217 			ethofld_tx(cst);
6218 		mtx_unlock(&cst->lock);
6219 		m_snd_tag_rele(&cst->com.com);
6220 	} else {
6221 		/*
6222 		 * There shouldn't be any pending packets if the tag
6223 		 * was freed by the kernel since any pending packet
6224 		 * should hold a reference to the tag.
6225 		 */
6226 		MPASS(mbufq_first(&cst->pending_tx) == NULL);
6227 		mtx_unlock(&cst->lock);
6228 	}
6229 
6230 	return (0);
6231 }
6232 #endif
6233