xref: /freebsd/sys/dev/cxgbe/t4_sge.c (revision 2c9a9dfc187d171de6b92654d71b977f067ed641)
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 sge_txq *, struct fw_eth_tx_pkt_wr *,
298     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 sge_txq *, struct fw_eth_tx_pkts_wr *,
304     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_tso(struct mbuf *m)
2349 {
2350 
2351 	M_ASSERTPKTHDR(m);
2352 
2353 	return (m->m_pkthdr.csum_flags & CSUM_TSO);
2354 }
2355 
2356 static inline int
2357 needs_l3_csum(struct mbuf *m)
2358 {
2359 
2360 	M_ASSERTPKTHDR(m);
2361 
2362 	return (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO));
2363 }
2364 
2365 static inline int
2366 needs_l4_csum(struct mbuf *m)
2367 {
2368 
2369 	M_ASSERTPKTHDR(m);
2370 
2371 	return (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 |
2372 	    CSUM_TCP_IPV6 | 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_udp_csum(struct mbuf *m)
2386 {
2387 
2388 	M_ASSERTPKTHDR(m);
2389 	return (m->m_pkthdr.csum_flags & (CSUM_UDP | CSUM_UDP_IPV6));
2390 }
2391 #endif
2392 
2393 static inline int
2394 needs_vlan_insertion(struct mbuf *m)
2395 {
2396 
2397 	M_ASSERTPKTHDR(m);
2398 
2399 	return (m->m_flags & M_VLANTAG);
2400 }
2401 
2402 static void *
2403 m_advance(struct mbuf **pm, int *poffset, int len)
2404 {
2405 	struct mbuf *m = *pm;
2406 	int offset = *poffset;
2407 	uintptr_t p = 0;
2408 
2409 	MPASS(len > 0);
2410 
2411 	for (;;) {
2412 		if (offset + len < m->m_len) {
2413 			offset += len;
2414 			p = mtod(m, uintptr_t) + offset;
2415 			break;
2416 		}
2417 		len -= m->m_len - offset;
2418 		m = m->m_next;
2419 		offset = 0;
2420 		MPASS(m != NULL);
2421 	}
2422 	*poffset = offset;
2423 	*pm = m;
2424 	return ((void *)p);
2425 }
2426 
2427 static inline int
2428 count_mbuf_ext_pgs(struct mbuf *m, int skip, vm_paddr_t *nextaddr)
2429 {
2430 	struct mbuf_ext_pgs *ext_pgs;
2431 	vm_paddr_t paddr;
2432 	int i, len, off, pglen, pgoff, seglen, segoff;
2433 	int nsegs = 0;
2434 
2435 	MBUF_EXT_PGS_ASSERT(m);
2436 	ext_pgs = m->m_ext.ext_pgs;
2437 	off = mtod(m, vm_offset_t);
2438 	len = m->m_len;
2439 	off += skip;
2440 	len -= skip;
2441 
2442 	if (ext_pgs->hdr_len != 0) {
2443 		if (off >= ext_pgs->hdr_len) {
2444 			off -= ext_pgs->hdr_len;
2445 		} else {
2446 			seglen = ext_pgs->hdr_len - off;
2447 			segoff = off;
2448 			seglen = min(seglen, len);
2449 			off = 0;
2450 			len -= seglen;
2451 			paddr = pmap_kextract(
2452 			    (vm_offset_t)&ext_pgs->hdr[segoff]);
2453 			if (*nextaddr != paddr)
2454 				nsegs++;
2455 			*nextaddr = paddr + seglen;
2456 		}
2457 	}
2458 	pgoff = ext_pgs->first_pg_off;
2459 	for (i = 0; i < ext_pgs->npgs && len > 0; i++) {
2460 		pglen = mbuf_ext_pg_len(ext_pgs, i, pgoff);
2461 		if (off >= pglen) {
2462 			off -= pglen;
2463 			pgoff = 0;
2464 			continue;
2465 		}
2466 		seglen = pglen - off;
2467 		segoff = pgoff + off;
2468 		off = 0;
2469 		seglen = min(seglen, len);
2470 		len -= seglen;
2471 		paddr = ext_pgs->pa[i] + segoff;
2472 		if (*nextaddr != paddr)
2473 			nsegs++;
2474 		*nextaddr = paddr + seglen;
2475 		pgoff = 0;
2476 	};
2477 	if (len != 0) {
2478 		seglen = min(len, ext_pgs->trail_len - off);
2479 		len -= seglen;
2480 		paddr = pmap_kextract((vm_offset_t)&ext_pgs->trail[off]);
2481 		if (*nextaddr != paddr)
2482 			nsegs++;
2483 		*nextaddr = paddr + seglen;
2484 	}
2485 
2486 	return (nsegs);
2487 }
2488 
2489 
2490 /*
2491  * Can deal with empty mbufs in the chain that have m_len = 0, but the chain
2492  * must have at least one mbuf that's not empty.  It is possible for this
2493  * routine to return 0 if skip accounts for all the contents of the mbuf chain.
2494  */
2495 static inline int
2496 count_mbuf_nsegs(struct mbuf *m, int skip, uint8_t *cflags)
2497 {
2498 	vm_paddr_t nextaddr, paddr;
2499 	vm_offset_t va;
2500 	int len, nsegs;
2501 
2502 	M_ASSERTPKTHDR(m);
2503 	MPASS(m->m_pkthdr.len > 0);
2504 	MPASS(m->m_pkthdr.len >= skip);
2505 
2506 	nsegs = 0;
2507 	nextaddr = 0;
2508 	for (; m; m = m->m_next) {
2509 		len = m->m_len;
2510 		if (__predict_false(len == 0))
2511 			continue;
2512 		if (skip >= len) {
2513 			skip -= len;
2514 			continue;
2515 		}
2516 		if ((m->m_flags & M_NOMAP) != 0) {
2517 			*cflags |= MC_NOMAP;
2518 			nsegs += count_mbuf_ext_pgs(m, skip, &nextaddr);
2519 			skip = 0;
2520 			continue;
2521 		}
2522 		va = mtod(m, vm_offset_t) + skip;
2523 		len -= skip;
2524 		skip = 0;
2525 		paddr = pmap_kextract(va);
2526 		nsegs += sglist_count((void *)(uintptr_t)va, len);
2527 		if (paddr == nextaddr)
2528 			nsegs--;
2529 		nextaddr = pmap_kextract(va + len - 1) + 1;
2530 	}
2531 
2532 	return (nsegs);
2533 }
2534 
2535 /*
2536  * Analyze the mbuf to determine its tx needs.  The mbuf passed in may change:
2537  * a) caller can assume it's been freed if this function returns with an error.
2538  * b) it may get defragged up if the gather list is too long for the hardware.
2539  */
2540 int
2541 parse_pkt(struct adapter *sc, struct mbuf **mp)
2542 {
2543 	struct mbuf *m0 = *mp, *m;
2544 	int rc, nsegs, defragged = 0, offset;
2545 	struct ether_header *eh;
2546 	void *l3hdr;
2547 #if defined(INET) || defined(INET6)
2548 	struct tcphdr *tcp;
2549 #endif
2550 #if defined(KERN_TLS) || defined(RATELIMIT)
2551 	struct cxgbe_snd_tag *cst;
2552 #endif
2553 	uint16_t eh_type;
2554 	uint8_t cflags;
2555 
2556 	cflags = 0;
2557 	M_ASSERTPKTHDR(m0);
2558 	if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) {
2559 		rc = EINVAL;
2560 fail:
2561 		m_freem(m0);
2562 		*mp = NULL;
2563 		return (rc);
2564 	}
2565 restart:
2566 	/*
2567 	 * First count the number of gather list segments in the payload.
2568 	 * Defrag the mbuf if nsegs exceeds the hardware limit.
2569 	 */
2570 	M_ASSERTPKTHDR(m0);
2571 	MPASS(m0->m_pkthdr.len > 0);
2572 	nsegs = count_mbuf_nsegs(m0, 0, &cflags);
2573 #if defined(KERN_TLS) || defined(RATELIMIT)
2574 	if (m0->m_pkthdr.csum_flags & CSUM_SND_TAG)
2575 		cst = mst_to_cst(m0->m_pkthdr.snd_tag);
2576 	else
2577 		cst = NULL;
2578 #endif
2579 #ifdef KERN_TLS
2580 	if (cst != NULL && cst->type == IF_SND_TAG_TYPE_TLS) {
2581 		int len16;
2582 
2583 		cflags |= MC_TLS;
2584 		set_mbuf_cflags(m0, cflags);
2585 		rc = t6_ktls_parse_pkt(m0, &nsegs, &len16);
2586 		if (rc != 0)
2587 			goto fail;
2588 		set_mbuf_nsegs(m0, nsegs);
2589 		set_mbuf_len16(m0, len16);
2590 		return (0);
2591 	}
2592 #endif
2593 	if (nsegs > (needs_tso(m0) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS)) {
2594 		if (defragged++ > 0 || (m = m_defrag(m0, M_NOWAIT)) == NULL) {
2595 			rc = EFBIG;
2596 			goto fail;
2597 		}
2598 		*mp = m0 = m;	/* update caller's copy after defrag */
2599 		goto restart;
2600 	}
2601 
2602 	if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN &&
2603 	    !(cflags & MC_NOMAP))) {
2604 		m0 = m_pullup(m0, m0->m_pkthdr.len);
2605 		if (m0 == NULL) {
2606 			/* Should have left well enough alone. */
2607 			rc = EFBIG;
2608 			goto fail;
2609 		}
2610 		*mp = m0;	/* update caller's copy after pullup */
2611 		goto restart;
2612 	}
2613 	set_mbuf_nsegs(m0, nsegs);
2614 	set_mbuf_cflags(m0, cflags);
2615 	if (sc->flags & IS_VF)
2616 		set_mbuf_len16(m0, txpkt_vm_len16(nsegs, needs_tso(m0)));
2617 	else
2618 		set_mbuf_len16(m0, txpkt_len16(nsegs, needs_tso(m0)));
2619 
2620 #ifdef RATELIMIT
2621 	/*
2622 	 * Ethofld is limited to TCP and UDP for now, and only when L4 hw
2623 	 * checksumming is enabled.  needs_l4_csum happens to check for all the
2624 	 * right things.
2625 	 */
2626 	if (__predict_false(needs_eo(cst) && !needs_l4_csum(m0))) {
2627 		m_snd_tag_rele(m0->m_pkthdr.snd_tag);
2628 		m0->m_pkthdr.snd_tag = NULL;
2629 		m0->m_pkthdr.csum_flags &= ~CSUM_SND_TAG;
2630 		cst = NULL;
2631 	}
2632 #endif
2633 
2634 	if (!needs_tso(m0) &&
2635 #ifdef RATELIMIT
2636 	    !needs_eo(cst) &&
2637 #endif
2638 	    !(sc->flags & IS_VF && (needs_l3_csum(m0) || needs_l4_csum(m0))))
2639 		return (0);
2640 
2641 	m = m0;
2642 	eh = mtod(m, struct ether_header *);
2643 	eh_type = ntohs(eh->ether_type);
2644 	if (eh_type == ETHERTYPE_VLAN) {
2645 		struct ether_vlan_header *evh = (void *)eh;
2646 
2647 		eh_type = ntohs(evh->evl_proto);
2648 		m0->m_pkthdr.l2hlen = sizeof(*evh);
2649 	} else
2650 		m0->m_pkthdr.l2hlen = sizeof(*eh);
2651 
2652 	offset = 0;
2653 	l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen);
2654 
2655 	switch (eh_type) {
2656 #ifdef INET6
2657 	case ETHERTYPE_IPV6:
2658 	{
2659 		struct ip6_hdr *ip6 = l3hdr;
2660 
2661 		MPASS(!needs_tso(m0) || ip6->ip6_nxt == IPPROTO_TCP);
2662 
2663 		m0->m_pkthdr.l3hlen = sizeof(*ip6);
2664 		break;
2665 	}
2666 #endif
2667 #ifdef INET
2668 	case ETHERTYPE_IP:
2669 	{
2670 		struct ip *ip = l3hdr;
2671 
2672 		m0->m_pkthdr.l3hlen = ip->ip_hl * 4;
2673 		break;
2674 	}
2675 #endif
2676 	default:
2677 		panic("%s: ethertype 0x%04x unknown.  if_cxgbe must be compiled"
2678 		    " with the same INET/INET6 options as the kernel.",
2679 		    __func__, eh_type);
2680 	}
2681 
2682 #if defined(INET) || defined(INET6)
2683 	if (needs_tcp_csum(m0)) {
2684 		tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen);
2685 		m0->m_pkthdr.l4hlen = tcp->th_off * 4;
2686 #ifdef RATELIMIT
2687 		if (tsclk >= 0 && *(uint32_t *)(tcp + 1) == ntohl(0x0101080a)) {
2688 			set_mbuf_eo_tsclk_tsoff(m0,
2689 			    V_FW_ETH_TX_EO_WR_TSCLK(tsclk) |
2690 			    V_FW_ETH_TX_EO_WR_TSOFF(sizeof(*tcp) / 2 + 1));
2691 		} else
2692 			set_mbuf_eo_tsclk_tsoff(m0, 0);
2693 	} else if (needs_udp_csum(m)) {
2694 		m0->m_pkthdr.l4hlen = sizeof(struct udphdr);
2695 #endif
2696 	}
2697 #ifdef RATELIMIT
2698 	if (needs_eo(cst)) {
2699 		u_int immhdrs;
2700 
2701 		/* EO WRs have the headers in the WR and not the GL. */
2702 		immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen +
2703 		    m0->m_pkthdr.l4hlen;
2704 		cflags = 0;
2705 		nsegs = count_mbuf_nsegs(m0, immhdrs, &cflags);
2706 		MPASS(cflags == mbuf_cflags(m0));
2707 		set_mbuf_eo_nsegs(m0, nsegs);
2708 		set_mbuf_eo_len16(m0,
2709 		    txpkt_eo_len16(nsegs, immhdrs, needs_tso(m0)));
2710 	}
2711 #endif
2712 #endif
2713 	MPASS(m0 == *mp);
2714 	return (0);
2715 }
2716 
2717 void *
2718 start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie)
2719 {
2720 	struct sge_eq *eq = &wrq->eq;
2721 	struct adapter *sc = wrq->adapter;
2722 	int ndesc, available;
2723 	struct wrqe *wr;
2724 	void *w;
2725 
2726 	MPASS(len16 > 0);
2727 	ndesc = howmany(len16, EQ_ESIZE / 16);
2728 	MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC);
2729 
2730 	EQ_LOCK(eq);
2731 
2732 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2733 		drain_wrq_wr_list(sc, wrq);
2734 
2735 	if (!STAILQ_EMPTY(&wrq->wr_list)) {
2736 slowpath:
2737 		EQ_UNLOCK(eq);
2738 		wr = alloc_wrqe(len16 * 16, wrq);
2739 		if (__predict_false(wr == NULL))
2740 			return (NULL);
2741 		cookie->pidx = -1;
2742 		cookie->ndesc = ndesc;
2743 		return (&wr->wr);
2744 	}
2745 
2746 	eq->cidx = read_hw_cidx(eq);
2747 	if (eq->pidx == eq->cidx)
2748 		available = eq->sidx - 1;
2749 	else
2750 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2751 	if (available < ndesc)
2752 		goto slowpath;
2753 
2754 	cookie->pidx = eq->pidx;
2755 	cookie->ndesc = ndesc;
2756 	TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link);
2757 
2758 	w = &eq->desc[eq->pidx];
2759 	IDXINCR(eq->pidx, ndesc, eq->sidx);
2760 	if (__predict_false(cookie->pidx + ndesc > eq->sidx)) {
2761 		w = &wrq->ss[0];
2762 		wrq->ss_pidx = cookie->pidx;
2763 		wrq->ss_len = len16 * 16;
2764 	}
2765 
2766 	EQ_UNLOCK(eq);
2767 
2768 	return (w);
2769 }
2770 
2771 void
2772 commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie)
2773 {
2774 	struct sge_eq *eq = &wrq->eq;
2775 	struct adapter *sc = wrq->adapter;
2776 	int ndesc, pidx;
2777 	struct wrq_cookie *prev, *next;
2778 
2779 	if (cookie->pidx == -1) {
2780 		struct wrqe *wr = __containerof(w, struct wrqe, wr);
2781 
2782 		t4_wrq_tx(sc, wr);
2783 		return;
2784 	}
2785 
2786 	if (__predict_false(w == &wrq->ss[0])) {
2787 		int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE;
2788 
2789 		MPASS(wrq->ss_len > n);	/* WR had better wrap around. */
2790 		bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n);
2791 		bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n);
2792 		wrq->tx_wrs_ss++;
2793 	} else
2794 		wrq->tx_wrs_direct++;
2795 
2796 	EQ_LOCK(eq);
2797 	ndesc = cookie->ndesc;	/* Can be more than SGE_MAX_WR_NDESC here. */
2798 	pidx = cookie->pidx;
2799 	MPASS(pidx >= 0 && pidx < eq->sidx);
2800 	prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link);
2801 	next = TAILQ_NEXT(cookie, link);
2802 	if (prev == NULL) {
2803 		MPASS(pidx == eq->dbidx);
2804 		if (next == NULL || ndesc >= 16) {
2805 			int available;
2806 			struct fw_eth_tx_pkt_wr *dst;	/* any fw WR struct will do */
2807 
2808 			/*
2809 			 * Note that the WR via which we'll request tx updates
2810 			 * is at pidx and not eq->pidx, which has moved on
2811 			 * already.
2812 			 */
2813 			dst = (void *)&eq->desc[pidx];
2814 			available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2815 			if (available < eq->sidx / 4 &&
2816 			    atomic_cmpset_int(&eq->equiq, 0, 1)) {
2817 				/*
2818 				 * XXX: This is not 100% reliable with some
2819 				 * types of WRs.  But this is a very unusual
2820 				 * situation for an ofld/ctrl queue anyway.
2821 				 */
2822 				dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
2823 				    F_FW_WR_EQUEQ);
2824 			}
2825 
2826 			ring_eq_db(wrq->adapter, eq, ndesc);
2827 		} else {
2828 			MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc);
2829 			next->pidx = pidx;
2830 			next->ndesc += ndesc;
2831 		}
2832 	} else {
2833 		MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc);
2834 		prev->ndesc += ndesc;
2835 	}
2836 	TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link);
2837 
2838 	if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list))
2839 		drain_wrq_wr_list(sc, wrq);
2840 
2841 #ifdef INVARIANTS
2842 	if (TAILQ_EMPTY(&wrq->incomplete_wrs)) {
2843 		/* Doorbell must have caught up to the pidx. */
2844 		MPASS(wrq->eq.pidx == wrq->eq.dbidx);
2845 	}
2846 #endif
2847 	EQ_UNLOCK(eq);
2848 }
2849 
2850 static u_int
2851 can_resume_eth_tx(struct mp_ring *r)
2852 {
2853 	struct sge_eq *eq = r->cookie;
2854 
2855 	return (total_available_tx_desc(eq) > eq->sidx / 8);
2856 }
2857 
2858 static inline int
2859 cannot_use_txpkts(struct mbuf *m)
2860 {
2861 	/* maybe put a GL limit too, to avoid silliness? */
2862 
2863 	return (needs_tso(m) || (mbuf_cflags(m) & (MC_RAW_WR | MC_TLS)) != 0);
2864 }
2865 
2866 static inline int
2867 discard_tx(struct sge_eq *eq)
2868 {
2869 
2870 	return ((eq->flags & (EQ_ENABLED | EQ_QFLUSH)) != EQ_ENABLED);
2871 }
2872 
2873 static inline int
2874 wr_can_update_eq(struct fw_eth_tx_pkts_wr *wr)
2875 {
2876 
2877 	switch (G_FW_WR_OP(be32toh(wr->op_pkd))) {
2878 	case FW_ULPTX_WR:
2879 	case FW_ETH_TX_PKT_WR:
2880 	case FW_ETH_TX_PKTS_WR:
2881 	case FW_ETH_TX_PKTS2_WR:
2882 	case FW_ETH_TX_PKT_VM_WR:
2883 		return (1);
2884 	default:
2885 		return (0);
2886 	}
2887 }
2888 
2889 /*
2890  * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to
2891  * be consumed.  Return the actual number consumed.  0 indicates a stall.
2892  */
2893 static u_int
2894 eth_tx(struct mp_ring *r, u_int cidx, u_int pidx)
2895 {
2896 	struct sge_txq *txq = r->cookie;
2897 	struct sge_eq *eq = &txq->eq;
2898 	struct ifnet *ifp = txq->ifp;
2899 	struct vi_info *vi = ifp->if_softc;
2900 	struct port_info *pi = vi->pi;
2901 	struct adapter *sc = pi->adapter;
2902 	u_int total, remaining;		/* # of packets */
2903 	u_int available, dbdiff;	/* # of hardware descriptors */
2904 	u_int n, next_cidx;
2905 	struct mbuf *m0, *tail;
2906 	struct txpkts txp;
2907 	struct fw_eth_tx_pkts_wr *wr;	/* any fw WR struct will do */
2908 
2909 	remaining = IDXDIFF(pidx, cidx, r->size);
2910 	MPASS(remaining > 0);	/* Must not be called without work to do. */
2911 	total = 0;
2912 
2913 	TXQ_LOCK(txq);
2914 	if (__predict_false(discard_tx(eq))) {
2915 		while (cidx != pidx) {
2916 			m0 = r->items[cidx];
2917 			m_freem(m0);
2918 			if (++cidx == r->size)
2919 				cidx = 0;
2920 		}
2921 		reclaim_tx_descs(txq, 2048);
2922 		total = remaining;
2923 		goto done;
2924 	}
2925 
2926 	/* How many hardware descriptors do we have readily available. */
2927 	if (eq->pidx == eq->cidx)
2928 		available = eq->sidx - 1;
2929 	else
2930 		available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1;
2931 	dbdiff = IDXDIFF(eq->pidx, eq->dbidx, eq->sidx);
2932 
2933 	while (remaining > 0) {
2934 
2935 		m0 = r->items[cidx];
2936 		M_ASSERTPKTHDR(m0);
2937 		MPASS(m0->m_nextpkt == NULL);
2938 
2939 		if (available < howmany(mbuf_len16(m0), EQ_ESIZE / 16)) {
2940 			MPASS(howmany(mbuf_len16(m0), EQ_ESIZE / 16) <= 64);
2941 			available += reclaim_tx_descs(txq, 64);
2942 			if (available < howmany(mbuf_len16(m0), EQ_ESIZE / 16))
2943 				break;	/* out of descriptors */
2944 		}
2945 
2946 		next_cidx = cidx + 1;
2947 		if (__predict_false(next_cidx == r->size))
2948 			next_cidx = 0;
2949 
2950 		wr = (void *)&eq->desc[eq->pidx];
2951 		if (mbuf_cflags(m0) & MC_RAW_WR) {
2952 			total++;
2953 			remaining--;
2954 			n = write_raw_wr(txq, (void *)wr, m0, available);
2955 #ifdef KERN_TLS
2956 		} else if (mbuf_cflags(m0) & MC_TLS) {
2957 			total++;
2958 			remaining--;
2959 			ETHER_BPF_MTAP(ifp, m0);
2960 			n = t6_ktls_write_wr(txq,(void *)wr, m0,
2961 			    mbuf_nsegs(m0), available);
2962 #endif
2963 		} else if (sc->flags & IS_VF) {
2964 			total++;
2965 			remaining--;
2966 			ETHER_BPF_MTAP(ifp, m0);
2967 			n = write_txpkt_vm_wr(sc, txq, (void *)wr, m0,
2968 			    available);
2969 		} else if (remaining > 1 &&
2970 		    try_txpkts(m0, r->items[next_cidx], &txp, available) == 0) {
2971 
2972 			/* pkts at cidx, next_cidx should both be in txp. */
2973 			MPASS(txp.npkt == 2);
2974 			tail = r->items[next_cidx];
2975 			MPASS(tail->m_nextpkt == NULL);
2976 			ETHER_BPF_MTAP(ifp, m0);
2977 			ETHER_BPF_MTAP(ifp, tail);
2978 			m0->m_nextpkt = tail;
2979 
2980 			if (__predict_false(++next_cidx == r->size))
2981 				next_cidx = 0;
2982 
2983 			while (next_cidx != pidx) {
2984 				if (add_to_txpkts(r->items[next_cidx], &txp,
2985 				    available) != 0)
2986 					break;
2987 				tail->m_nextpkt = r->items[next_cidx];
2988 				tail = tail->m_nextpkt;
2989 				ETHER_BPF_MTAP(ifp, tail);
2990 				if (__predict_false(++next_cidx == r->size))
2991 					next_cidx = 0;
2992 			}
2993 
2994 			n = write_txpkts_wr(txq, wr, m0, &txp, available);
2995 			total += txp.npkt;
2996 			remaining -= txp.npkt;
2997 		} else {
2998 			total++;
2999 			remaining--;
3000 			ETHER_BPF_MTAP(ifp, m0);
3001 			n = write_txpkt_wr(txq, (void *)wr, m0, available);
3002 		}
3003 		MPASS(n >= 1 && n <= available);
3004 		if (!(mbuf_cflags(m0) & MC_TLS))
3005 			MPASS(n <= SGE_MAX_WR_NDESC);
3006 
3007 		available -= n;
3008 		dbdiff += n;
3009 		IDXINCR(eq->pidx, n, eq->sidx);
3010 
3011 		if (wr_can_update_eq(wr)) {
3012 			if (total_available_tx_desc(eq) < eq->sidx / 4 &&
3013 			    atomic_cmpset_int(&eq->equiq, 0, 1)) {
3014 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ |
3015 				    F_FW_WR_EQUEQ);
3016 				eq->equeqidx = eq->pidx;
3017 			} else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >=
3018 			    32) {
3019 				wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ);
3020 				eq->equeqidx = eq->pidx;
3021 			}
3022 		}
3023 
3024 		if (dbdiff >= 16 && remaining >= 4) {
3025 			ring_eq_db(sc, eq, dbdiff);
3026 			available += reclaim_tx_descs(txq, 4 * dbdiff);
3027 			dbdiff = 0;
3028 		}
3029 
3030 		cidx = next_cidx;
3031 	}
3032 	if (dbdiff != 0) {
3033 		ring_eq_db(sc, eq, dbdiff);
3034 		reclaim_tx_descs(txq, 32);
3035 	}
3036 done:
3037 	TXQ_UNLOCK(txq);
3038 
3039 	return (total);
3040 }
3041 
3042 static inline void
3043 init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx,
3044     int qsize)
3045 {
3046 
3047 	KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS,
3048 	    ("%s: bad tmr_idx %d", __func__, tmr_idx));
3049 	KASSERT(pktc_idx < SGE_NCOUNTERS,	/* -ve is ok, means don't use */
3050 	    ("%s: bad pktc_idx %d", __func__, pktc_idx));
3051 
3052 	iq->flags = 0;
3053 	iq->adapter = sc;
3054 	iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx);
3055 	iq->intr_pktc_idx = SGE_NCOUNTERS - 1;
3056 	if (pktc_idx >= 0) {
3057 		iq->intr_params |= F_QINTR_CNT_EN;
3058 		iq->intr_pktc_idx = pktc_idx;
3059 	}
3060 	iq->qsize = roundup2(qsize, 16);	/* See FW_IQ_CMD/iqsize */
3061 	iq->sidx = iq->qsize - sc->params.sge.spg_len / IQ_ESIZE;
3062 }
3063 
3064 static inline void
3065 init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name)
3066 {
3067 
3068 	fl->qsize = qsize;
3069 	fl->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
3070 	strlcpy(fl->lockname, name, sizeof(fl->lockname));
3071 	if (sc->flags & BUF_PACKING_OK &&
3072 	    ((!is_t4(sc) && buffer_packing) ||	/* T5+: enabled unless 0 */
3073 	    (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */
3074 		fl->flags |= FL_BUF_PACKING;
3075 	find_best_refill_source(sc, fl, maxp);
3076 	find_safe_refill_source(sc, fl);
3077 }
3078 
3079 static inline void
3080 init_eq(struct adapter *sc, struct sge_eq *eq, int eqtype, int qsize,
3081     uint8_t tx_chan, uint16_t iqid, char *name)
3082 {
3083 	KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype));
3084 
3085 	eq->flags = eqtype & EQ_TYPEMASK;
3086 	eq->tx_chan = tx_chan;
3087 	eq->iqid = iqid;
3088 	eq->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE;
3089 	strlcpy(eq->lockname, name, sizeof(eq->lockname));
3090 }
3091 
3092 static int
3093 alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag,
3094     bus_dmamap_t *map, bus_addr_t *pa, void **va)
3095 {
3096 	int rc;
3097 
3098 	rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR,
3099 	    BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag);
3100 	if (rc != 0) {
3101 		device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc);
3102 		goto done;
3103 	}
3104 
3105 	rc = bus_dmamem_alloc(*tag, va,
3106 	    BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map);
3107 	if (rc != 0) {
3108 		device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc);
3109 		goto done;
3110 	}
3111 
3112 	rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0);
3113 	if (rc != 0) {
3114 		device_printf(sc->dev, "cannot load DMA map: %d\n", rc);
3115 		goto done;
3116 	}
3117 done:
3118 	if (rc)
3119 		free_ring(sc, *tag, *map, *pa, *va);
3120 
3121 	return (rc);
3122 }
3123 
3124 static int
3125 free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map,
3126     bus_addr_t pa, void *va)
3127 {
3128 	if (pa)
3129 		bus_dmamap_unload(tag, map);
3130 	if (va)
3131 		bus_dmamem_free(tag, va, map);
3132 	if (tag)
3133 		bus_dma_tag_destroy(tag);
3134 
3135 	return (0);
3136 }
3137 
3138 /*
3139  * Allocates the ring for an ingress queue and an optional freelist.  If the
3140  * freelist is specified it will be allocated and then associated with the
3141  * ingress queue.
3142  *
3143  * Returns errno on failure.  Resources allocated up to that point may still be
3144  * allocated.  Caller is responsible for cleanup in case this function fails.
3145  *
3146  * If the ingress queue will take interrupts directly then the intr_idx
3147  * specifies the vector, starting from 0.  -1 means the interrupts for this
3148  * queue should be forwarded to the fwq.
3149  */
3150 static int
3151 alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl,
3152     int intr_idx, int cong)
3153 {
3154 	int rc, i, cntxt_id;
3155 	size_t len;
3156 	struct fw_iq_cmd c;
3157 	struct port_info *pi = vi->pi;
3158 	struct adapter *sc = iq->adapter;
3159 	struct sge_params *sp = &sc->params.sge;
3160 	__be32 v = 0;
3161 
3162 	len = iq->qsize * IQ_ESIZE;
3163 	rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba,
3164 	    (void **)&iq->desc);
3165 	if (rc != 0)
3166 		return (rc);
3167 
3168 	bzero(&c, sizeof(c));
3169 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST |
3170 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) |
3171 	    V_FW_IQ_CMD_VFN(0));
3172 
3173 	c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART |
3174 	    FW_LEN16(c));
3175 
3176 	/* Special handling for firmware event queue */
3177 	if (iq == &sc->sge.fwq)
3178 		v |= F_FW_IQ_CMD_IQASYNCH;
3179 
3180 	if (intr_idx < 0) {
3181 		/* Forwarded interrupts, all headed to fwq */
3182 		v |= F_FW_IQ_CMD_IQANDST;
3183 		v |= V_FW_IQ_CMD_IQANDSTINDEX(sc->sge.fwq.cntxt_id);
3184 	} else {
3185 		KASSERT(intr_idx < sc->intr_count,
3186 		    ("%s: invalid direct intr_idx %d", __func__, intr_idx));
3187 		v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx);
3188 	}
3189 
3190 	c.type_to_iqandstindex = htobe32(v |
3191 	    V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) |
3192 	    V_FW_IQ_CMD_VIID(vi->viid) |
3193 	    V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT));
3194 	c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) |
3195 	    F_FW_IQ_CMD_IQGTSMODE |
3196 	    V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) |
3197 	    V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4));
3198 	c.iqsize = htobe16(iq->qsize);
3199 	c.iqaddr = htobe64(iq->ba);
3200 	if (cong >= 0)
3201 		c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN);
3202 
3203 	if (fl) {
3204 		mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF);
3205 
3206 		len = fl->qsize * EQ_ESIZE;
3207 		rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map,
3208 		    &fl->ba, (void **)&fl->desc);
3209 		if (rc)
3210 			return (rc);
3211 
3212 		/* Allocate space for one software descriptor per buffer. */
3213 		rc = alloc_fl_sdesc(fl);
3214 		if (rc != 0) {
3215 			device_printf(sc->dev,
3216 			    "failed to setup fl software descriptors: %d\n",
3217 			    rc);
3218 			return (rc);
3219 		}
3220 
3221 		if (fl->flags & FL_BUF_PACKING) {
3222 			fl->lowat = roundup2(sp->fl_starve_threshold2, 8);
3223 			fl->buf_boundary = sp->pack_boundary;
3224 		} else {
3225 			fl->lowat = roundup2(sp->fl_starve_threshold, 8);
3226 			fl->buf_boundary = 16;
3227 		}
3228 		if (fl_pad && fl->buf_boundary < sp->pad_boundary)
3229 			fl->buf_boundary = sp->pad_boundary;
3230 
3231 		c.iqns_to_fl0congen |=
3232 		    htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) |
3233 			F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO |
3234 			(fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) |
3235 			(fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN :
3236 			    0));
3237 		if (cong >= 0) {
3238 			c.iqns_to_fl0congen |=
3239 				htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) |
3240 				    F_FW_IQ_CMD_FL0CONGCIF |
3241 				    F_FW_IQ_CMD_FL0CONGEN);
3242 		}
3243 		c.fl0dcaen_to_fl0cidxfthresh =
3244 		    htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ?
3245 			X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B_T6) |
3246 			V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ?
3247 			X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B));
3248 		c.fl0size = htobe16(fl->qsize);
3249 		c.fl0addr = htobe64(fl->ba);
3250 	}
3251 
3252 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3253 	if (rc != 0) {
3254 		device_printf(sc->dev,
3255 		    "failed to create ingress queue: %d\n", rc);
3256 		return (rc);
3257 	}
3258 
3259 	iq->cidx = 0;
3260 	iq->gen = F_RSPD_GEN;
3261 	iq->intr_next = iq->intr_params;
3262 	iq->cntxt_id = be16toh(c.iqid);
3263 	iq->abs_id = be16toh(c.physiqid);
3264 	iq->flags |= IQ_ALLOCATED;
3265 
3266 	cntxt_id = iq->cntxt_id - sc->sge.iq_start;
3267 	if (cntxt_id >= sc->sge.niq) {
3268 		panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__,
3269 		    cntxt_id, sc->sge.niq - 1);
3270 	}
3271 	sc->sge.iqmap[cntxt_id] = iq;
3272 
3273 	if (fl) {
3274 		u_int qid;
3275 
3276 		iq->flags |= IQ_HAS_FL;
3277 		fl->cntxt_id = be16toh(c.fl0id);
3278 		fl->pidx = fl->cidx = 0;
3279 
3280 		cntxt_id = fl->cntxt_id - sc->sge.eq_start;
3281 		if (cntxt_id >= sc->sge.neq) {
3282 			panic("%s: fl->cntxt_id (%d) more than the max (%d)",
3283 			    __func__, cntxt_id, sc->sge.neq - 1);
3284 		}
3285 		sc->sge.eqmap[cntxt_id] = (void *)fl;
3286 
3287 		qid = fl->cntxt_id;
3288 		if (isset(&sc->doorbells, DOORBELL_UDB)) {
3289 			uint32_t s_qpp = sc->params.sge.eq_s_qpp;
3290 			uint32_t mask = (1 << s_qpp) - 1;
3291 			volatile uint8_t *udb;
3292 
3293 			udb = sc->udbs_base + UDBS_DB_OFFSET;
3294 			udb += (qid >> s_qpp) << PAGE_SHIFT;
3295 			qid &= mask;
3296 			if (qid < PAGE_SIZE / UDBS_SEG_SIZE) {
3297 				udb += qid << UDBS_SEG_SHIFT;
3298 				qid = 0;
3299 			}
3300 			fl->udb = (volatile void *)udb;
3301 		}
3302 		fl->dbval = V_QID(qid) | sc->chip_params->sge_fl_db;
3303 
3304 		FL_LOCK(fl);
3305 		/* Enough to make sure the SGE doesn't think it's starved */
3306 		refill_fl(sc, fl, fl->lowat);
3307 		FL_UNLOCK(fl);
3308 	}
3309 
3310 	if (chip_id(sc) >= CHELSIO_T5 && !(sc->flags & IS_VF) && cong >= 0) {
3311 		uint32_t param, val;
3312 
3313 		param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
3314 		    V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) |
3315 		    V_FW_PARAMS_PARAM_YZ(iq->cntxt_id);
3316 		if (cong == 0)
3317 			val = 1 << 19;
3318 		else {
3319 			val = 2 << 19;
3320 			for (i = 0; i < 4; i++) {
3321 				if (cong & (1 << i))
3322 					val |= 1 << (i << 2);
3323 			}
3324 		}
3325 
3326 		rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &param, &val);
3327 		if (rc != 0) {
3328 			/* report error but carry on */
3329 			device_printf(sc->dev,
3330 			    "failed to set congestion manager context for "
3331 			    "ingress queue %d: %d\n", iq->cntxt_id, rc);
3332 		}
3333 	}
3334 
3335 	/* Enable IQ interrupts */
3336 	atomic_store_rel_int(&iq->state, IQS_IDLE);
3337 	t4_write_reg(sc, sc->sge_gts_reg, V_SEINTARM(iq->intr_params) |
3338 	    V_INGRESSQID(iq->cntxt_id));
3339 
3340 	return (0);
3341 }
3342 
3343 static int
3344 free_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl)
3345 {
3346 	int rc;
3347 	struct adapter *sc = iq->adapter;
3348 	device_t dev;
3349 
3350 	if (sc == NULL)
3351 		return (0);	/* nothing to do */
3352 
3353 	dev = vi ? vi->dev : sc->dev;
3354 
3355 	if (iq->flags & IQ_ALLOCATED) {
3356 		rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0,
3357 		    FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id,
3358 		    fl ? fl->cntxt_id : 0xffff, 0xffff);
3359 		if (rc != 0) {
3360 			device_printf(dev,
3361 			    "failed to free queue %p: %d\n", iq, rc);
3362 			return (rc);
3363 		}
3364 		iq->flags &= ~IQ_ALLOCATED;
3365 	}
3366 
3367 	free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc);
3368 
3369 	bzero(iq, sizeof(*iq));
3370 
3371 	if (fl) {
3372 		free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba,
3373 		    fl->desc);
3374 
3375 		if (fl->sdesc)
3376 			free_fl_sdesc(sc, fl);
3377 
3378 		if (mtx_initialized(&fl->fl_lock))
3379 			mtx_destroy(&fl->fl_lock);
3380 
3381 		bzero(fl, sizeof(*fl));
3382 	}
3383 
3384 	return (0);
3385 }
3386 
3387 static void
3388 add_iq_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid,
3389     struct sge_iq *iq)
3390 {
3391 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3392 
3393 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD, &iq->ba,
3394 	    "bus address of descriptor ring");
3395 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3396 	    iq->qsize * IQ_ESIZE, "descriptor ring size in bytes");
3397 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3398 	    CTLTYPE_INT | CTLFLAG_RD, &iq->abs_id, 0, sysctl_uint16, "I",
3399 	    "absolute id of the queue");
3400 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3401 	    CTLTYPE_INT | CTLFLAG_RD, &iq->cntxt_id, 0, sysctl_uint16, "I",
3402 	    "SGE context id of the queue");
3403 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3404 	    CTLTYPE_INT | CTLFLAG_RD, &iq->cidx, 0, sysctl_uint16, "I",
3405 	    "consumer index");
3406 }
3407 
3408 static void
3409 add_fl_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx,
3410     struct sysctl_oid *oid, struct sge_fl *fl)
3411 {
3412 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3413 
3414 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3415 	    "freelist");
3416 	children = SYSCTL_CHILDREN(oid);
3417 
3418 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
3419 	    &fl->ba, "bus address of descriptor ring");
3420 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
3421 	    fl->sidx * EQ_ESIZE + sc->params.sge.spg_len,
3422 	    "desc ring size in bytes");
3423 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3424 	    CTLTYPE_INT | CTLFLAG_RD, &fl->cntxt_id, 0, sysctl_uint16, "I",
3425 	    "SGE context id of the freelist");
3426 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL,
3427 	    fl_pad ? 1 : 0, "padding enabled");
3428 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL,
3429 	    fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled");
3430 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx,
3431 	    0, "consumer index");
3432 	if (fl->flags & FL_BUF_PACKING) {
3433 		SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset",
3434 		    CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset");
3435 	}
3436 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx,
3437 	    0, "producer index");
3438 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_allocated",
3439 	    CTLFLAG_RD, &fl->mbuf_allocated, "# of mbuf allocated");
3440 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_inlined",
3441 	    CTLFLAG_RD, &fl->mbuf_inlined, "# of mbuf inlined in clusters");
3442 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated",
3443 	    CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated");
3444 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled",
3445 	    CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled");
3446 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled",
3447 	    CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)");
3448 }
3449 
3450 static int
3451 alloc_fwq(struct adapter *sc)
3452 {
3453 	int rc, intr_idx;
3454 	struct sge_iq *fwq = &sc->sge.fwq;
3455 	struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev);
3456 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3457 
3458 	init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE);
3459 	if (sc->flags & IS_VF)
3460 		intr_idx = 0;
3461 	else
3462 		intr_idx = sc->intr_count > 1 ? 1 : 0;
3463 	rc = alloc_iq_fl(&sc->port[0]->vi[0], fwq, NULL, intr_idx, -1);
3464 	if (rc != 0) {
3465 		device_printf(sc->dev,
3466 		    "failed to create firmware event queue: %d\n", rc);
3467 		return (rc);
3468 	}
3469 
3470 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD,
3471 	    NULL, "firmware event queue");
3472 	add_iq_sysctls(&sc->ctx, oid, fwq);
3473 
3474 	return (0);
3475 }
3476 
3477 static int
3478 free_fwq(struct adapter *sc)
3479 {
3480 	return free_iq_fl(NULL, &sc->sge.fwq, NULL);
3481 }
3482 
3483 static int
3484 alloc_ctrlq(struct adapter *sc, struct sge_wrq *ctrlq, int idx,
3485     struct sysctl_oid *oid)
3486 {
3487 	int rc;
3488 	char name[16];
3489 	struct sysctl_oid_list *children;
3490 
3491 	snprintf(name, sizeof(name), "%s ctrlq%d", device_get_nameunit(sc->dev),
3492 	    idx);
3493 	init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[idx]->tx_chan,
3494 	    sc->sge.fwq.cntxt_id, name);
3495 
3496 	children = SYSCTL_CHILDREN(oid);
3497 	snprintf(name, sizeof(name), "%d", idx);
3498 	oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3499 	    NULL, "ctrl queue");
3500 	rc = alloc_wrq(sc, NULL, ctrlq, oid);
3501 
3502 	return (rc);
3503 }
3504 
3505 int
3506 tnl_cong(struct port_info *pi, int drop)
3507 {
3508 
3509 	if (drop == -1)
3510 		return (-1);
3511 	else if (drop == 1)
3512 		return (0);
3513 	else
3514 		return (pi->rx_e_chan_map);
3515 }
3516 
3517 static int
3518 alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int intr_idx, int idx,
3519     struct sysctl_oid *oid)
3520 {
3521 	int rc;
3522 	struct adapter *sc = vi->pi->adapter;
3523 	struct sysctl_oid_list *children;
3524 	char name[16];
3525 
3526 	rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, intr_idx,
3527 	    tnl_cong(vi->pi, cong_drop));
3528 	if (rc != 0)
3529 		return (rc);
3530 
3531 	if (idx == 0)
3532 		sc->sge.iq_base = rxq->iq.abs_id - rxq->iq.cntxt_id;
3533 	else
3534 		KASSERT(rxq->iq.cntxt_id + sc->sge.iq_base == rxq->iq.abs_id,
3535 		    ("iq_base mismatch"));
3536 	KASSERT(sc->sge.iq_base == 0 || sc->flags & IS_VF,
3537 	    ("PF with non-zero iq_base"));
3538 
3539 	/*
3540 	 * The freelist is just barely above the starvation threshold right now,
3541 	 * fill it up a bit more.
3542 	 */
3543 	FL_LOCK(&rxq->fl);
3544 	refill_fl(sc, &rxq->fl, 128);
3545 	FL_UNLOCK(&rxq->fl);
3546 
3547 #if defined(INET) || defined(INET6)
3548 	rc = tcp_lro_init_args(&rxq->lro, vi->ifp, lro_entries, lro_mbufs);
3549 	if (rc != 0)
3550 		return (rc);
3551 	MPASS(rxq->lro.ifp == vi->ifp);	/* also indicates LRO init'ed */
3552 
3553 	if (vi->ifp->if_capenable & IFCAP_LRO)
3554 		rxq->iq.flags |= IQ_LRO_ENABLED;
3555 #endif
3556 	if (vi->ifp->if_capenable & IFCAP_HWRXTSTMP)
3557 		rxq->iq.flags |= IQ_RX_TIMESTAMP;
3558 	rxq->ifp = vi->ifp;
3559 
3560 	children = SYSCTL_CHILDREN(oid);
3561 
3562 	snprintf(name, sizeof(name), "%d", idx);
3563 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3564 	    NULL, "rx queue");
3565 	children = SYSCTL_CHILDREN(oid);
3566 
3567 	add_iq_sysctls(&vi->ctx, oid, &rxq->iq);
3568 #if defined(INET) || defined(INET6)
3569 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD,
3570 	    &rxq->lro.lro_queued, 0, NULL);
3571 	SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD,
3572 	    &rxq->lro.lro_flushed, 0, NULL);
3573 #endif
3574 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD,
3575 	    &rxq->rxcsum, "# of times hardware assisted with checksum");
3576 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_extraction",
3577 	    CTLFLAG_RD, &rxq->vlan_extraction,
3578 	    "# of times hardware extracted 802.1Q tag");
3579 
3580 	add_fl_sysctls(sc, &vi->ctx, oid, &rxq->fl);
3581 
3582 	return (rc);
3583 }
3584 
3585 static int
3586 free_rxq(struct vi_info *vi, struct sge_rxq *rxq)
3587 {
3588 	int rc;
3589 
3590 #if defined(INET) || defined(INET6)
3591 	if (rxq->lro.ifp) {
3592 		tcp_lro_free(&rxq->lro);
3593 		rxq->lro.ifp = NULL;
3594 	}
3595 #endif
3596 
3597 	rc = free_iq_fl(vi, &rxq->iq, &rxq->fl);
3598 	if (rc == 0)
3599 		bzero(rxq, sizeof(*rxq));
3600 
3601 	return (rc);
3602 }
3603 
3604 #ifdef TCP_OFFLOAD
3605 static int
3606 alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq,
3607     int intr_idx, int idx, struct sysctl_oid *oid)
3608 {
3609 	struct port_info *pi = vi->pi;
3610 	int rc;
3611 	struct sysctl_oid_list *children;
3612 	char name[16];
3613 
3614 	rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx, 0);
3615 	if (rc != 0)
3616 		return (rc);
3617 
3618 	children = SYSCTL_CHILDREN(oid);
3619 
3620 	snprintf(name, sizeof(name), "%d", idx);
3621 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3622 	    NULL, "rx queue");
3623 	add_iq_sysctls(&vi->ctx, oid, &ofld_rxq->iq);
3624 	add_fl_sysctls(pi->adapter, &vi->ctx, oid, &ofld_rxq->fl);
3625 
3626 	return (rc);
3627 }
3628 
3629 static int
3630 free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq)
3631 {
3632 	int rc;
3633 
3634 	rc = free_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl);
3635 	if (rc == 0)
3636 		bzero(ofld_rxq, sizeof(*ofld_rxq));
3637 
3638 	return (rc);
3639 }
3640 #endif
3641 
3642 #ifdef DEV_NETMAP
3643 static int
3644 alloc_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int intr_idx,
3645     int idx, struct sysctl_oid *oid)
3646 {
3647 	int rc;
3648 	struct sysctl_oid_list *children;
3649 	struct sysctl_ctx_list *ctx;
3650 	char name[16];
3651 	size_t len;
3652 	struct adapter *sc = vi->pi->adapter;
3653 	struct netmap_adapter *na = NA(vi->ifp);
3654 
3655 	MPASS(na != NULL);
3656 
3657 	len = vi->qsize_rxq * IQ_ESIZE;
3658 	rc = alloc_ring(sc, len, &nm_rxq->iq_desc_tag, &nm_rxq->iq_desc_map,
3659 	    &nm_rxq->iq_ba, (void **)&nm_rxq->iq_desc);
3660 	if (rc != 0)
3661 		return (rc);
3662 
3663 	len = na->num_rx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3664 	rc = alloc_ring(sc, len, &nm_rxq->fl_desc_tag, &nm_rxq->fl_desc_map,
3665 	    &nm_rxq->fl_ba, (void **)&nm_rxq->fl_desc);
3666 	if (rc != 0)
3667 		return (rc);
3668 
3669 	nm_rxq->vi = vi;
3670 	nm_rxq->nid = idx;
3671 	nm_rxq->iq_cidx = 0;
3672 	nm_rxq->iq_sidx = vi->qsize_rxq - sc->params.sge.spg_len / IQ_ESIZE;
3673 	nm_rxq->iq_gen = F_RSPD_GEN;
3674 	nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0;
3675 	nm_rxq->fl_sidx = na->num_rx_desc;
3676 	nm_rxq->intr_idx = intr_idx;
3677 	nm_rxq->iq_cntxt_id = INVALID_NM_RXQ_CNTXT_ID;
3678 
3679 	ctx = &vi->ctx;
3680 	children = SYSCTL_CHILDREN(oid);
3681 
3682 	snprintf(name, sizeof(name), "%d", idx);
3683 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL,
3684 	    "rx queue");
3685 	children = SYSCTL_CHILDREN(oid);
3686 
3687 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id",
3688 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_abs_id, 0, sysctl_uint16,
3689 	    "I", "absolute id of the queue");
3690 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3691 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cntxt_id, 0, sysctl_uint16,
3692 	    "I", "SGE context id of the queue");
3693 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
3694 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cidx, 0, sysctl_uint16, "I",
3695 	    "consumer index");
3696 
3697 	children = SYSCTL_CHILDREN(oid);
3698 	oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL,
3699 	    "freelist");
3700 	children = SYSCTL_CHILDREN(oid);
3701 
3702 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id",
3703 	    CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->fl_cntxt_id, 0, sysctl_uint16,
3704 	    "I", "SGE context id of the freelist");
3705 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD,
3706 	    &nm_rxq->fl_cidx, 0, "consumer index");
3707 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD,
3708 	    &nm_rxq->fl_pidx, 0, "producer index");
3709 
3710 	return (rc);
3711 }
3712 
3713 
3714 static int
3715 free_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq)
3716 {
3717 	struct adapter *sc = vi->pi->adapter;
3718 
3719 	if (vi->flags & VI_INIT_DONE)
3720 		MPASS(nm_rxq->iq_cntxt_id == INVALID_NM_RXQ_CNTXT_ID);
3721 	else
3722 		MPASS(nm_rxq->iq_cntxt_id == 0);
3723 
3724 	free_ring(sc, nm_rxq->iq_desc_tag, nm_rxq->iq_desc_map, nm_rxq->iq_ba,
3725 	    nm_rxq->iq_desc);
3726 	free_ring(sc, nm_rxq->fl_desc_tag, nm_rxq->fl_desc_map, nm_rxq->fl_ba,
3727 	    nm_rxq->fl_desc);
3728 
3729 	return (0);
3730 }
3731 
3732 static int
3733 alloc_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq, int iqidx, int idx,
3734     struct sysctl_oid *oid)
3735 {
3736 	int rc;
3737 	size_t len;
3738 	struct port_info *pi = vi->pi;
3739 	struct adapter *sc = pi->adapter;
3740 	struct netmap_adapter *na = NA(vi->ifp);
3741 	char name[16];
3742 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
3743 
3744 	len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len;
3745 	rc = alloc_ring(sc, len, &nm_txq->desc_tag, &nm_txq->desc_map,
3746 	    &nm_txq->ba, (void **)&nm_txq->desc);
3747 	if (rc)
3748 		return (rc);
3749 
3750 	nm_txq->pidx = nm_txq->cidx = 0;
3751 	nm_txq->sidx = na->num_tx_desc;
3752 	nm_txq->nid = idx;
3753 	nm_txq->iqidx = iqidx;
3754 	nm_txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
3755 	    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(sc->pf) |
3756 	    V_TXPKT_VF(vi->vin) | V_TXPKT_VF_VLD(vi->vfvld));
3757 	if (sc->params.fw_vers >= FW_VERSION32(1, 24, 11, 0))
3758 		nm_txq->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS2_WR));
3759 	else
3760 		nm_txq->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
3761 	nm_txq->cntxt_id = INVALID_NM_TXQ_CNTXT_ID;
3762 
3763 	snprintf(name, sizeof(name), "%d", idx);
3764 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
3765 	    NULL, "netmap tx queue");
3766 	children = SYSCTL_CHILDREN(oid);
3767 
3768 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
3769 	    &nm_txq->cntxt_id, 0, "SGE context id of the queue");
3770 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
3771 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->cidx, 0, sysctl_uint16, "I",
3772 	    "consumer index");
3773 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
3774 	    CTLTYPE_INT | CTLFLAG_RD, &nm_txq->pidx, 0, sysctl_uint16, "I",
3775 	    "producer index");
3776 
3777 	return (rc);
3778 }
3779 
3780 static int
3781 free_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq)
3782 {
3783 	struct adapter *sc = vi->pi->adapter;
3784 
3785 	if (vi->flags & VI_INIT_DONE)
3786 		MPASS(nm_txq->cntxt_id == INVALID_NM_TXQ_CNTXT_ID);
3787 	else
3788 		MPASS(nm_txq->cntxt_id == 0);
3789 
3790 	free_ring(sc, nm_txq->desc_tag, nm_txq->desc_map, nm_txq->ba,
3791 	    nm_txq->desc);
3792 
3793 	return (0);
3794 }
3795 #endif
3796 
3797 /*
3798  * Returns a reasonable automatic cidx flush threshold for a given queue size.
3799  */
3800 static u_int
3801 qsize_to_fthresh(int qsize)
3802 {
3803 	u_int fthresh;
3804 
3805 	while (!powerof2(qsize))
3806 		qsize++;
3807 	fthresh = ilog2(qsize);
3808 	if (fthresh > X_CIDXFLUSHTHRESH_128)
3809 		fthresh = X_CIDXFLUSHTHRESH_128;
3810 
3811 	return (fthresh);
3812 }
3813 
3814 static int
3815 ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq)
3816 {
3817 	int rc, cntxt_id;
3818 	struct fw_eq_ctrl_cmd c;
3819 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3820 
3821 	bzero(&c, sizeof(c));
3822 
3823 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST |
3824 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) |
3825 	    V_FW_EQ_CTRL_CMD_VFN(0));
3826 	c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC |
3827 	    F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c));
3828 	c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid));
3829 	c.physeqid_pkd = htobe32(0);
3830 	c.fetchszm_to_iqid =
3831 	    htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
3832 		V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) |
3833 		F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid));
3834 	c.dcaen_to_eqsize =
3835 	    htobe32(V_FW_EQ_CTRL_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
3836 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
3837 		V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3838 		V_FW_EQ_CTRL_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
3839 		V_FW_EQ_CTRL_CMD_EQSIZE(qsize));
3840 	c.eqaddr = htobe64(eq->ba);
3841 
3842 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3843 	if (rc != 0) {
3844 		device_printf(sc->dev,
3845 		    "failed to create control queue %d: %d\n", eq->tx_chan, rc);
3846 		return (rc);
3847 	}
3848 	eq->flags |= EQ_ALLOCATED;
3849 
3850 	eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid));
3851 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3852 	if (cntxt_id >= sc->sge.neq)
3853 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3854 		cntxt_id, sc->sge.neq - 1);
3855 	sc->sge.eqmap[cntxt_id] = eq;
3856 
3857 	return (rc);
3858 }
3859 
3860 static int
3861 eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3862 {
3863 	int rc, cntxt_id;
3864 	struct fw_eq_eth_cmd c;
3865 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3866 
3867 	bzero(&c, sizeof(c));
3868 
3869 	c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST |
3870 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) |
3871 	    V_FW_EQ_ETH_CMD_VFN(0));
3872 	c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC |
3873 	    F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c));
3874 	c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE |
3875 	    F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid));
3876 	c.fetchszm_to_iqid =
3877 	    htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) |
3878 		V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO |
3879 		V_FW_EQ_ETH_CMD_IQID(eq->iqid));
3880 	c.dcaen_to_eqsize =
3881 	    htobe32(V_FW_EQ_ETH_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
3882 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
3883 		V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3884 		V_FW_EQ_ETH_CMD_EQSIZE(qsize));
3885 	c.eqaddr = htobe64(eq->ba);
3886 
3887 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3888 	if (rc != 0) {
3889 		device_printf(vi->dev,
3890 		    "failed to create Ethernet egress queue: %d\n", rc);
3891 		return (rc);
3892 	}
3893 	eq->flags |= EQ_ALLOCATED;
3894 
3895 	eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd));
3896 	eq->abs_id = G_FW_EQ_ETH_CMD_PHYSEQID(be32toh(c.physeqid_pkd));
3897 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3898 	if (cntxt_id >= sc->sge.neq)
3899 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3900 		cntxt_id, sc->sge.neq - 1);
3901 	sc->sge.eqmap[cntxt_id] = eq;
3902 
3903 	return (rc);
3904 }
3905 
3906 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3907 static int
3908 ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3909 {
3910 	int rc, cntxt_id;
3911 	struct fw_eq_ofld_cmd c;
3912 	int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3913 
3914 	bzero(&c, sizeof(c));
3915 
3916 	c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST |
3917 	    F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) |
3918 	    V_FW_EQ_OFLD_CMD_VFN(0));
3919 	c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC |
3920 	    F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c));
3921 	c.fetchszm_to_iqid =
3922 		htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_STATUS_PAGE) |
3923 		    V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) |
3924 		    F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid));
3925 	c.dcaen_to_eqsize =
3926 	    htobe32(V_FW_EQ_OFLD_CMD_FBMIN(chip_id(sc) <= CHELSIO_T5 ?
3927 		X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) |
3928 		V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) |
3929 		V_FW_EQ_OFLD_CMD_CIDXFTHRESH(qsize_to_fthresh(qsize)) |
3930 		V_FW_EQ_OFLD_CMD_EQSIZE(qsize));
3931 	c.eqaddr = htobe64(eq->ba);
3932 
3933 	rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c);
3934 	if (rc != 0) {
3935 		device_printf(vi->dev,
3936 		    "failed to create egress queue for TCP offload: %d\n", rc);
3937 		return (rc);
3938 	}
3939 	eq->flags |= EQ_ALLOCATED;
3940 
3941 	eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd));
3942 	cntxt_id = eq->cntxt_id - sc->sge.eq_start;
3943 	if (cntxt_id >= sc->sge.neq)
3944 	    panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__,
3945 		cntxt_id, sc->sge.neq - 1);
3946 	sc->sge.eqmap[cntxt_id] = eq;
3947 
3948 	return (rc);
3949 }
3950 #endif
3951 
3952 static int
3953 alloc_eq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq)
3954 {
3955 	int rc, qsize;
3956 	size_t len;
3957 
3958 	mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF);
3959 
3960 	qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE;
3961 	len = qsize * EQ_ESIZE;
3962 	rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map,
3963 	    &eq->ba, (void **)&eq->desc);
3964 	if (rc)
3965 		return (rc);
3966 
3967 	eq->pidx = eq->cidx = eq->dbidx = 0;
3968 	/* Note that equeqidx is not used with sge_wrq (OFLD/CTRL) queues. */
3969 	eq->equeqidx = 0;
3970 	eq->doorbells = sc->doorbells;
3971 
3972 	switch (eq->flags & EQ_TYPEMASK) {
3973 	case EQ_CTRL:
3974 		rc = ctrl_eq_alloc(sc, eq);
3975 		break;
3976 
3977 	case EQ_ETH:
3978 		rc = eth_eq_alloc(sc, vi, eq);
3979 		break;
3980 
3981 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
3982 	case EQ_OFLD:
3983 		rc = ofld_eq_alloc(sc, vi, eq);
3984 		break;
3985 #endif
3986 
3987 	default:
3988 		panic("%s: invalid eq type %d.", __func__,
3989 		    eq->flags & EQ_TYPEMASK);
3990 	}
3991 	if (rc != 0) {
3992 		device_printf(sc->dev,
3993 		    "failed to allocate egress queue(%d): %d\n",
3994 		    eq->flags & EQ_TYPEMASK, rc);
3995 	}
3996 
3997 	if (isset(&eq->doorbells, DOORBELL_UDB) ||
3998 	    isset(&eq->doorbells, DOORBELL_UDBWC) ||
3999 	    isset(&eq->doorbells, DOORBELL_WCWR)) {
4000 		uint32_t s_qpp = sc->params.sge.eq_s_qpp;
4001 		uint32_t mask = (1 << s_qpp) - 1;
4002 		volatile uint8_t *udb;
4003 
4004 		udb = sc->udbs_base + UDBS_DB_OFFSET;
4005 		udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT;	/* pg offset */
4006 		eq->udb_qid = eq->cntxt_id & mask;		/* id in page */
4007 		if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE)
4008 	    		clrbit(&eq->doorbells, DOORBELL_WCWR);
4009 		else {
4010 			udb += eq->udb_qid << UDBS_SEG_SHIFT;	/* seg offset */
4011 			eq->udb_qid = 0;
4012 		}
4013 		eq->udb = (volatile void *)udb;
4014 	}
4015 
4016 	return (rc);
4017 }
4018 
4019 static int
4020 free_eq(struct adapter *sc, struct sge_eq *eq)
4021 {
4022 	int rc;
4023 
4024 	if (eq->flags & EQ_ALLOCATED) {
4025 		switch (eq->flags & EQ_TYPEMASK) {
4026 		case EQ_CTRL:
4027 			rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0,
4028 			    eq->cntxt_id);
4029 			break;
4030 
4031 		case EQ_ETH:
4032 			rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0,
4033 			    eq->cntxt_id);
4034 			break;
4035 
4036 #if defined(TCP_OFFLOAD) || defined(RATELIMIT)
4037 		case EQ_OFLD:
4038 			rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0,
4039 			    eq->cntxt_id);
4040 			break;
4041 #endif
4042 
4043 		default:
4044 			panic("%s: invalid eq type %d.", __func__,
4045 			    eq->flags & EQ_TYPEMASK);
4046 		}
4047 		if (rc != 0) {
4048 			device_printf(sc->dev,
4049 			    "failed to free egress queue (%d): %d\n",
4050 			    eq->flags & EQ_TYPEMASK, rc);
4051 			return (rc);
4052 		}
4053 		eq->flags &= ~EQ_ALLOCATED;
4054 	}
4055 
4056 	free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc);
4057 
4058 	if (mtx_initialized(&eq->eq_lock))
4059 		mtx_destroy(&eq->eq_lock);
4060 
4061 	bzero(eq, sizeof(*eq));
4062 	return (0);
4063 }
4064 
4065 static int
4066 alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq,
4067     struct sysctl_oid *oid)
4068 {
4069 	int rc;
4070 	struct sysctl_ctx_list *ctx = vi ? &vi->ctx : &sc->ctx;
4071 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
4072 
4073 	rc = alloc_eq(sc, vi, &wrq->eq);
4074 	if (rc)
4075 		return (rc);
4076 
4077 	wrq->adapter = sc;
4078 	TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq);
4079 	TAILQ_INIT(&wrq->incomplete_wrs);
4080 	STAILQ_INIT(&wrq->wr_list);
4081 	wrq->nwr_pending = 0;
4082 	wrq->ndesc_needed = 0;
4083 
4084 	SYSCTL_ADD_UAUTO(ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
4085 	    &wrq->eq.ba, "bus address of descriptor ring");
4086 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
4087 	    wrq->eq.sidx * EQ_ESIZE + sc->params.sge.spg_len,
4088 	    "desc ring size in bytes");
4089 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
4090 	    &wrq->eq.cntxt_id, 0, "SGE context id of the queue");
4091 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx",
4092 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I",
4093 	    "consumer index");
4094 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx",
4095 	    CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I",
4096 	    "producer index");
4097 	SYSCTL_ADD_INT(ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
4098 	    wrq->eq.sidx, "status page index");
4099 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD,
4100 	    &wrq->tx_wrs_direct, "# of work requests (direct)");
4101 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD,
4102 	    &wrq->tx_wrs_copied, "# of work requests (copied)");
4103 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_sspace", CTLFLAG_RD,
4104 	    &wrq->tx_wrs_ss, "# of work requests (copied from scratch space)");
4105 
4106 	return (rc);
4107 }
4108 
4109 static int
4110 free_wrq(struct adapter *sc, struct sge_wrq *wrq)
4111 {
4112 	int rc;
4113 
4114 	rc = free_eq(sc, &wrq->eq);
4115 	if (rc)
4116 		return (rc);
4117 
4118 	bzero(wrq, sizeof(*wrq));
4119 	return (0);
4120 }
4121 
4122 static int
4123 alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx,
4124     struct sysctl_oid *oid)
4125 {
4126 	int rc;
4127 	struct port_info *pi = vi->pi;
4128 	struct adapter *sc = pi->adapter;
4129 	struct sge_eq *eq = &txq->eq;
4130 	char name[16];
4131 	struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid);
4132 
4133 	rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx, can_resume_eth_tx,
4134 	    M_CXGBE, M_WAITOK);
4135 	if (rc != 0) {
4136 		device_printf(sc->dev, "failed to allocate mp_ring: %d\n", rc);
4137 		return (rc);
4138 	}
4139 
4140 	rc = alloc_eq(sc, vi, eq);
4141 	if (rc != 0) {
4142 		mp_ring_free(txq->r);
4143 		txq->r = NULL;
4144 		return (rc);
4145 	}
4146 
4147 	/* Can't fail after this point. */
4148 
4149 	if (idx == 0)
4150 		sc->sge.eq_base = eq->abs_id - eq->cntxt_id;
4151 	else
4152 		KASSERT(eq->cntxt_id + sc->sge.eq_base == eq->abs_id,
4153 		    ("eq_base mismatch"));
4154 	KASSERT(sc->sge.eq_base == 0 || sc->flags & IS_VF,
4155 	    ("PF with non-zero eq_base"));
4156 
4157 	TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq);
4158 	txq->ifp = vi->ifp;
4159 	txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK);
4160 	if (sc->flags & IS_VF)
4161 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) |
4162 		    V_TXPKT_INTF(pi->tx_chan));
4163 	else
4164 		txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) |
4165 		    V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_PF(sc->pf) |
4166 		    V_TXPKT_VF(vi->vin) | V_TXPKT_VF_VLD(vi->vfvld));
4167 	txq->tc_idx = -1;
4168 	txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE,
4169 	    M_ZERO | M_WAITOK);
4170 
4171 	snprintf(name, sizeof(name), "%d", idx);
4172 	oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD,
4173 	    NULL, "tx queue");
4174 	children = SYSCTL_CHILDREN(oid);
4175 
4176 	SYSCTL_ADD_UAUTO(&vi->ctx, children, OID_AUTO, "ba", CTLFLAG_RD,
4177 	    &eq->ba, "bus address of descriptor ring");
4178 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "dmalen", CTLFLAG_RD, NULL,
4179 	    eq->sidx * EQ_ESIZE + sc->params.sge.spg_len,
4180 	    "desc ring size in bytes");
4181 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD,
4182 	    &eq->abs_id, 0, "absolute id of the queue");
4183 	SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD,
4184 	    &eq->cntxt_id, 0, "SGE context id of the queue");
4185 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx",
4186 	    CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I",
4187 	    "consumer index");
4188 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx",
4189 	    CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I",
4190 	    "producer index");
4191 	SYSCTL_ADD_INT(&vi->ctx, children, OID_AUTO, "sidx", CTLFLAG_RD, NULL,
4192 	    eq->sidx, "status page index");
4193 
4194 	SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "tc",
4195 	    CTLTYPE_INT | CTLFLAG_RW, vi, idx, sysctl_tc, "I",
4196 	    "traffic class (-1 means none)");
4197 
4198 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD,
4199 	    &txq->txcsum, "# of times hardware assisted with checksum");
4200 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_insertion",
4201 	    CTLFLAG_RD, &txq->vlan_insertion,
4202 	    "# of times hardware inserted 802.1Q tag");
4203 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD,
4204 	    &txq->tso_wrs, "# of TSO work requests");
4205 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD,
4206 	    &txq->imm_wrs, "# of work requests with immediate data");
4207 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD,
4208 	    &txq->sgl_wrs, "# of work requests with direct SGL");
4209 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD,
4210 	    &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)");
4211 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_wrs",
4212 	    CTLFLAG_RD, &txq->txpkts0_wrs,
4213 	    "# of txpkts (type 0) work requests");
4214 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_wrs",
4215 	    CTLFLAG_RD, &txq->txpkts1_wrs,
4216 	    "# of txpkts (type 1) work requests");
4217 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_pkts",
4218 	    CTLFLAG_RD, &txq->txpkts0_pkts,
4219 	    "# of frames tx'd using type0 txpkts work requests");
4220 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_pkts",
4221 	    CTLFLAG_RD, &txq->txpkts1_pkts,
4222 	    "# of frames tx'd using type1 txpkts work requests");
4223 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "raw_wrs", CTLFLAG_RD,
4224 	    &txq->raw_wrs, "# of raw work requests (non-packets)");
4225 	SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "tls_wrs", CTLFLAG_RD,
4226 	    &txq->tls_wrs, "# of TLS work requests (TLS records)");
4227 
4228 #ifdef KERN_TLS
4229 	if (sc->flags & KERN_TLS_OK) {
4230 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4231 		    "kern_tls_records", CTLFLAG_RD, &txq->kern_tls_records,
4232 		    "# of NIC TLS records transmitted");
4233 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4234 		    "kern_tls_short", CTLFLAG_RD, &txq->kern_tls_short,
4235 		    "# of short NIC TLS records transmitted");
4236 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4237 		    "kern_tls_partial", CTLFLAG_RD, &txq->kern_tls_partial,
4238 		    "# of partial NIC TLS records transmitted");
4239 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4240 		    "kern_tls_full", CTLFLAG_RD, &txq->kern_tls_full,
4241 		    "# of full NIC TLS records transmitted");
4242 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4243 		    "kern_tls_octets", CTLFLAG_RD, &txq->kern_tls_octets,
4244 		    "# of payload octets in transmitted NIC TLS records");
4245 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4246 		    "kern_tls_waste", CTLFLAG_RD, &txq->kern_tls_waste,
4247 		    "# of octets DMAd but not transmitted in NIC TLS records");
4248 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4249 		    "kern_tls_options", CTLFLAG_RD, &txq->kern_tls_options,
4250 		    "# of NIC TLS options-only packets transmitted");
4251 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4252 		    "kern_tls_header", CTLFLAG_RD, &txq->kern_tls_header,
4253 		    "# of NIC TLS header-only packets transmitted");
4254 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4255 		    "kern_tls_fin", CTLFLAG_RD, &txq->kern_tls_fin,
4256 		    "# of NIC TLS FIN-only packets transmitted");
4257 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4258 		    "kern_tls_fin_short", CTLFLAG_RD, &txq->kern_tls_fin_short,
4259 		    "# of NIC TLS padded FIN packets on short TLS records");
4260 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4261 		    "kern_tls_cbc", CTLFLAG_RD, &txq->kern_tls_cbc,
4262 		    "# of NIC TLS sessions using AES-CBC");
4263 		SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO,
4264 		    "kern_tls_gcm", CTLFLAG_RD, &txq->kern_tls_gcm,
4265 		    "# of NIC TLS sessions using AES-GCM");
4266 	}
4267 #endif
4268 
4269 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_enqueues",
4270 	    CTLFLAG_RD, &txq->r->enqueues,
4271 	    "# of enqueues to the mp_ring for this queue");
4272 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_drops",
4273 	    CTLFLAG_RD, &txq->r->drops,
4274 	    "# of drops in the mp_ring for this queue");
4275 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_starts",
4276 	    CTLFLAG_RD, &txq->r->starts,
4277 	    "# of normal consumer starts in the mp_ring for this queue");
4278 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_stalls",
4279 	    CTLFLAG_RD, &txq->r->stalls,
4280 	    "# of consumer stalls in the mp_ring for this queue");
4281 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_restarts",
4282 	    CTLFLAG_RD, &txq->r->restarts,
4283 	    "# of consumer restarts in the mp_ring for this queue");
4284 	SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_abdications",
4285 	    CTLFLAG_RD, &txq->r->abdications,
4286 	    "# of consumer abdications in the mp_ring for this queue");
4287 
4288 	return (0);
4289 }
4290 
4291 static int
4292 free_txq(struct vi_info *vi, struct sge_txq *txq)
4293 {
4294 	int rc;
4295 	struct adapter *sc = vi->pi->adapter;
4296 	struct sge_eq *eq = &txq->eq;
4297 
4298 	rc = free_eq(sc, eq);
4299 	if (rc)
4300 		return (rc);
4301 
4302 	sglist_free(txq->gl);
4303 	free(txq->sdesc, M_CXGBE);
4304 	mp_ring_free(txq->r);
4305 
4306 	bzero(txq, sizeof(*txq));
4307 	return (0);
4308 }
4309 
4310 static void
4311 oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error)
4312 {
4313 	bus_addr_t *ba = arg;
4314 
4315 	KASSERT(nseg == 1,
4316 	    ("%s meant for single segment mappings only.", __func__));
4317 
4318 	*ba = error ? 0 : segs->ds_addr;
4319 }
4320 
4321 static inline void
4322 ring_fl_db(struct adapter *sc, struct sge_fl *fl)
4323 {
4324 	uint32_t n, v;
4325 
4326 	n = IDXDIFF(fl->pidx / 8, fl->dbidx, fl->sidx);
4327 	MPASS(n > 0);
4328 
4329 	wmb();
4330 	v = fl->dbval | V_PIDX(n);
4331 	if (fl->udb)
4332 		*fl->udb = htole32(v);
4333 	else
4334 		t4_write_reg(sc, sc->sge_kdoorbell_reg, v);
4335 	IDXINCR(fl->dbidx, n, fl->sidx);
4336 }
4337 
4338 /*
4339  * Fills up the freelist by allocating up to 'n' buffers.  Buffers that are
4340  * recycled do not count towards this allocation budget.
4341  *
4342  * Returns non-zero to indicate that this freelist should be added to the list
4343  * of starving freelists.
4344  */
4345 static int
4346 refill_fl(struct adapter *sc, struct sge_fl *fl, int n)
4347 {
4348 	__be64 *d;
4349 	struct fl_sdesc *sd;
4350 	uintptr_t pa;
4351 	caddr_t cl;
4352 	struct cluster_layout *cll;
4353 	struct sw_zone_info *swz;
4354 	struct cluster_metadata *clm;
4355 	uint16_t max_pidx;
4356 	uint16_t hw_cidx = fl->hw_cidx;		/* stable snapshot */
4357 
4358 	FL_LOCK_ASSERT_OWNED(fl);
4359 
4360 	/*
4361 	 * We always stop at the beginning of the hardware descriptor that's just
4362 	 * before the one with the hw cidx.  This is to avoid hw pidx = hw cidx,
4363 	 * which would mean an empty freelist to the chip.
4364 	 */
4365 	max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1;
4366 	if (fl->pidx == max_pidx * 8)
4367 		return (0);
4368 
4369 	d = &fl->desc[fl->pidx];
4370 	sd = &fl->sdesc[fl->pidx];
4371 	cll = &fl->cll_def;	/* default layout */
4372 	swz = &sc->sge.sw_zone_info[cll->zidx];
4373 
4374 	while (n > 0) {
4375 
4376 		if (sd->cl != NULL) {
4377 
4378 			if (sd->nmbuf == 0) {
4379 				/*
4380 				 * Fast recycle without involving any atomics on
4381 				 * the cluster's metadata (if the cluster has
4382 				 * metadata).  This happens when all frames
4383 				 * received in the cluster were small enough to
4384 				 * fit within a single mbuf each.
4385 				 */
4386 				fl->cl_fast_recycled++;
4387 #ifdef INVARIANTS
4388 				clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
4389 				if (clm != NULL)
4390 					MPASS(clm->refcount == 1);
4391 #endif
4392 				goto recycled_fast;
4393 			}
4394 
4395 			/*
4396 			 * Cluster is guaranteed to have metadata.  Clusters
4397 			 * without metadata always take the fast recycle path
4398 			 * when they're recycled.
4399 			 */
4400 			clm = cl_metadata(sc, fl, &sd->cll, sd->cl);
4401 			MPASS(clm != NULL);
4402 
4403 			if (atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4404 				fl->cl_recycled++;
4405 				counter_u64_add(extfree_rels, 1);
4406 				goto recycled;
4407 			}
4408 			sd->cl = NULL;	/* gave up my reference */
4409 		}
4410 		MPASS(sd->cl == NULL);
4411 alloc:
4412 		cl = uma_zalloc(swz->zone, M_NOWAIT);
4413 		if (__predict_false(cl == NULL)) {
4414 			if (cll == &fl->cll_alt || fl->cll_alt.zidx == -1 ||
4415 			    fl->cll_def.zidx == fl->cll_alt.zidx)
4416 				break;
4417 
4418 			/* fall back to the safe zone */
4419 			cll = &fl->cll_alt;
4420 			swz = &sc->sge.sw_zone_info[cll->zidx];
4421 			goto alloc;
4422 		}
4423 		fl->cl_allocated++;
4424 		n--;
4425 
4426 		pa = pmap_kextract((vm_offset_t)cl);
4427 		pa += cll->region1;
4428 		sd->cl = cl;
4429 		sd->cll = *cll;
4430 		*d = htobe64(pa | cll->hwidx);
4431 		clm = cl_metadata(sc, fl, cll, cl);
4432 		if (clm != NULL) {
4433 recycled:
4434 #ifdef INVARIANTS
4435 			clm->sd = sd;
4436 #endif
4437 			clm->refcount = 1;
4438 		}
4439 		sd->nmbuf = 0;
4440 recycled_fast:
4441 		d++;
4442 		sd++;
4443 		if (__predict_false(++fl->pidx % 8 == 0)) {
4444 			uint16_t pidx = fl->pidx / 8;
4445 
4446 			if (__predict_false(pidx == fl->sidx)) {
4447 				fl->pidx = 0;
4448 				pidx = 0;
4449 				sd = fl->sdesc;
4450 				d = fl->desc;
4451 			}
4452 			if (pidx == max_pidx)
4453 				break;
4454 
4455 			if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4)
4456 				ring_fl_db(sc, fl);
4457 		}
4458 	}
4459 
4460 	if (fl->pidx / 8 != fl->dbidx)
4461 		ring_fl_db(sc, fl);
4462 
4463 	return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING));
4464 }
4465 
4466 /*
4467  * Attempt to refill all starving freelists.
4468  */
4469 static void
4470 refill_sfl(void *arg)
4471 {
4472 	struct adapter *sc = arg;
4473 	struct sge_fl *fl, *fl_temp;
4474 
4475 	mtx_assert(&sc->sfl_lock, MA_OWNED);
4476 	TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) {
4477 		FL_LOCK(fl);
4478 		refill_fl(sc, fl, 64);
4479 		if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) {
4480 			TAILQ_REMOVE(&sc->sfl, fl, link);
4481 			fl->flags &= ~FL_STARVING;
4482 		}
4483 		FL_UNLOCK(fl);
4484 	}
4485 
4486 	if (!TAILQ_EMPTY(&sc->sfl))
4487 		callout_schedule(&sc->sfl_callout, hz / 5);
4488 }
4489 
4490 static int
4491 alloc_fl_sdesc(struct sge_fl *fl)
4492 {
4493 
4494 	fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc), M_CXGBE,
4495 	    M_ZERO | M_WAITOK);
4496 
4497 	return (0);
4498 }
4499 
4500 static void
4501 free_fl_sdesc(struct adapter *sc, struct sge_fl *fl)
4502 {
4503 	struct fl_sdesc *sd;
4504 	struct cluster_metadata *clm;
4505 	struct cluster_layout *cll;
4506 	int i;
4507 
4508 	sd = fl->sdesc;
4509 	for (i = 0; i < fl->sidx * 8; i++, sd++) {
4510 		if (sd->cl == NULL)
4511 			continue;
4512 
4513 		cll = &sd->cll;
4514 		clm = cl_metadata(sc, fl, cll, sd->cl);
4515 		if (sd->nmbuf == 0)
4516 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4517 		else if (clm && atomic_fetchadd_int(&clm->refcount, -1) == 1) {
4518 			uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl);
4519 			counter_u64_add(extfree_rels, 1);
4520 		}
4521 		sd->cl = NULL;
4522 	}
4523 
4524 	free(fl->sdesc, M_CXGBE);
4525 	fl->sdesc = NULL;
4526 }
4527 
4528 static inline void
4529 get_pkt_gl(struct mbuf *m, struct sglist *gl)
4530 {
4531 	int rc;
4532 
4533 	M_ASSERTPKTHDR(m);
4534 
4535 	sglist_reset(gl);
4536 	rc = sglist_append_mbuf(gl, m);
4537 	if (__predict_false(rc != 0)) {
4538 		panic("%s: mbuf %p (%d segs) was vetted earlier but now fails "
4539 		    "with %d.", __func__, m, mbuf_nsegs(m), rc);
4540 	}
4541 
4542 	KASSERT(gl->sg_nseg == mbuf_nsegs(m),
4543 	    ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m,
4544 	    mbuf_nsegs(m), gl->sg_nseg));
4545 	KASSERT(gl->sg_nseg > 0 &&
4546 	    gl->sg_nseg <= (needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS),
4547 	    ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__,
4548 		gl->sg_nseg, needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS));
4549 }
4550 
4551 /*
4552  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
4553  */
4554 static inline u_int
4555 txpkt_len16(u_int nsegs, u_int tso)
4556 {
4557 	u_int n;
4558 
4559 	MPASS(nsegs > 0);
4560 
4561 	nsegs--; /* first segment is part of ulptx_sgl */
4562 	n = sizeof(struct fw_eth_tx_pkt_wr) + sizeof(struct cpl_tx_pkt_core) +
4563 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4564 	if (tso)
4565 		n += sizeof(struct cpl_tx_pkt_lso_core);
4566 
4567 	return (howmany(n, 16));
4568 }
4569 
4570 /*
4571  * len16 for a txpkt_vm WR with a GL.  Includes the firmware work
4572  * request header.
4573  */
4574 static inline u_int
4575 txpkt_vm_len16(u_int nsegs, u_int tso)
4576 {
4577 	u_int n;
4578 
4579 	MPASS(nsegs > 0);
4580 
4581 	nsegs--; /* first segment is part of ulptx_sgl */
4582 	n = sizeof(struct fw_eth_tx_pkt_vm_wr) +
4583 	    sizeof(struct cpl_tx_pkt_core) +
4584 	    sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
4585 	if (tso)
4586 		n += sizeof(struct cpl_tx_pkt_lso_core);
4587 
4588 	return (howmany(n, 16));
4589 }
4590 
4591 /*
4592  * len16 for a txpkts type 0 WR with a GL.  Does not include the firmware work
4593  * request header.
4594  */
4595 static inline u_int
4596 txpkts0_len16(u_int nsegs)
4597 {
4598 	u_int n;
4599 
4600 	MPASS(nsegs > 0);
4601 
4602 	nsegs--; /* first segment is part of ulptx_sgl */
4603 	n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) +
4604 	    sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) +
4605 	    8 * ((3 * nsegs) / 2 + (nsegs & 1));
4606 
4607 	return (howmany(n, 16));
4608 }
4609 
4610 /*
4611  * len16 for a txpkts type 1 WR with a GL.  Does not include the firmware work
4612  * request header.
4613  */
4614 static inline u_int
4615 txpkts1_len16(void)
4616 {
4617 	u_int n;
4618 
4619 	n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl);
4620 
4621 	return (howmany(n, 16));
4622 }
4623 
4624 static inline u_int
4625 imm_payload(u_int ndesc)
4626 {
4627 	u_int n;
4628 
4629 	n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) -
4630 	    sizeof(struct cpl_tx_pkt_core);
4631 
4632 	return (n);
4633 }
4634 
4635 /*
4636  * Write a VM txpkt WR for this packet to the hardware descriptors, update the
4637  * software descriptor, and advance the pidx.  It is guaranteed that enough
4638  * descriptors are available.
4639  *
4640  * The return value is the # of hardware descriptors used.
4641  */
4642 static u_int
4643 write_txpkt_vm_wr(struct adapter *sc, struct sge_txq *txq,
4644     struct fw_eth_tx_pkt_vm_wr *wr, struct mbuf *m0, u_int available)
4645 {
4646 	struct sge_eq *eq = &txq->eq;
4647 	struct tx_sdesc *txsd;
4648 	struct cpl_tx_pkt_core *cpl;
4649 	uint32_t ctrl;	/* used in many unrelated places */
4650 	uint64_t ctrl1;
4651 	int csum_type, len16, ndesc, pktlen, nsegs;
4652 	caddr_t dst;
4653 
4654 	TXQ_LOCK_ASSERT_OWNED(txq);
4655 	M_ASSERTPKTHDR(m0);
4656 	MPASS(available > 0 && available < eq->sidx);
4657 
4658 	len16 = mbuf_len16(m0);
4659 	nsegs = mbuf_nsegs(m0);
4660 	pktlen = m0->m_pkthdr.len;
4661 	ctrl = sizeof(struct cpl_tx_pkt_core);
4662 	if (needs_tso(m0))
4663 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4664 	ndesc = howmany(len16, EQ_ESIZE / 16);
4665 	MPASS(ndesc <= available);
4666 
4667 	/* Firmware work request header */
4668 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4669 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_VM_WR) |
4670 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4671 
4672 	ctrl = V_FW_WR_LEN16(len16);
4673 	wr->equiq_to_len16 = htobe32(ctrl);
4674 	wr->r3[0] = 0;
4675 	wr->r3[1] = 0;
4676 
4677 	/*
4678 	 * Copy over ethmacdst, ethmacsrc, ethtype, and vlantci.
4679 	 * vlantci is ignored unless the ethtype is 0x8100, so it's
4680 	 * simpler to always copy it rather than making it
4681 	 * conditional.  Also, it seems that we do not have to set
4682 	 * vlantci or fake the ethtype when doing VLAN tag insertion.
4683 	 */
4684 	m_copydata(m0, 0, sizeof(struct ether_header) + 2, wr->ethmacdst);
4685 
4686 	csum_type = -1;
4687 	if (needs_tso(m0)) {
4688 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4689 
4690 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4691 		    m0->m_pkthdr.l4hlen > 0,
4692 		    ("%s: mbuf %p needs TSO but missing header lengths",
4693 			__func__, m0));
4694 
4695 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4696 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4697 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4698 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4699 			ctrl |= V_LSO_ETHHDR_LEN(1);
4700 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4701 			ctrl |= F_LSO_IPV6;
4702 
4703 		lso->lso_ctrl = htobe32(ctrl);
4704 		lso->ipid_ofst = htobe16(0);
4705 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4706 		lso->seqno_offset = htobe32(0);
4707 		lso->len = htobe32(pktlen);
4708 
4709 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4710 			csum_type = TX_CSUM_TCPIP6;
4711 		else
4712 			csum_type = TX_CSUM_TCPIP;
4713 
4714 		cpl = (void *)(lso + 1);
4715 
4716 		txq->tso_wrs++;
4717 	} else {
4718 		if (m0->m_pkthdr.csum_flags & CSUM_IP_TCP)
4719 			csum_type = TX_CSUM_TCPIP;
4720 		else if (m0->m_pkthdr.csum_flags & CSUM_IP_UDP)
4721 			csum_type = TX_CSUM_UDPIP;
4722 		else if (m0->m_pkthdr.csum_flags & CSUM_IP6_TCP)
4723 			csum_type = TX_CSUM_TCPIP6;
4724 		else if (m0->m_pkthdr.csum_flags & CSUM_IP6_UDP)
4725 			csum_type = TX_CSUM_UDPIP6;
4726 #if defined(INET)
4727 		else if (m0->m_pkthdr.csum_flags & CSUM_IP) {
4728 			/*
4729 			 * XXX: The firmware appears to stomp on the
4730 			 * fragment/flags field of the IP header when
4731 			 * using TX_CSUM_IP.  Fall back to doing
4732 			 * software checksums.
4733 			 */
4734 			u_short *sump;
4735 			struct mbuf *m;
4736 			int offset;
4737 
4738 			m = m0;
4739 			offset = 0;
4740 			sump = m_advance(&m, &offset, m0->m_pkthdr.l2hlen +
4741 			    offsetof(struct ip, ip_sum));
4742 			*sump = in_cksum_skip(m0, m0->m_pkthdr.l2hlen +
4743 			    m0->m_pkthdr.l3hlen, m0->m_pkthdr.l2hlen);
4744 			m0->m_pkthdr.csum_flags &= ~CSUM_IP;
4745 		}
4746 #endif
4747 
4748 		cpl = (void *)(wr + 1);
4749 	}
4750 
4751 	/* Checksum offload */
4752 	ctrl1 = 0;
4753 	if (needs_l3_csum(m0) == 0)
4754 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4755 	if (csum_type >= 0) {
4756 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0,
4757 	    ("%s: mbuf %p needs checksum offload but missing header lengths",
4758 			__func__, m0));
4759 
4760 		if (chip_id(sc) <= CHELSIO_T5) {
4761 			ctrl1 |= V_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4762 			    ETHER_HDR_LEN);
4763 		} else {
4764 			ctrl1 |= V_T6_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen -
4765 			    ETHER_HDR_LEN);
4766 		}
4767 		ctrl1 |= V_TXPKT_IPHDR_LEN(m0->m_pkthdr.l3hlen);
4768 		ctrl1 |= V_TXPKT_CSUM_TYPE(csum_type);
4769 	} else
4770 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4771 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4772 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4773 		txq->txcsum++;	/* some hardware assistance provided */
4774 
4775 	/* VLAN tag insertion */
4776 	if (needs_vlan_insertion(m0)) {
4777 		ctrl1 |= F_TXPKT_VLAN_VLD |
4778 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4779 		txq->vlan_insertion++;
4780 	}
4781 
4782 	/* CPL header */
4783 	cpl->ctrl0 = txq->cpl_ctrl0;
4784 	cpl->pack = 0;
4785 	cpl->len = htobe16(pktlen);
4786 	cpl->ctrl1 = htobe64(ctrl1);
4787 
4788 	/* SGL */
4789 	dst = (void *)(cpl + 1);
4790 
4791 	/*
4792 	 * A packet using TSO will use up an entire descriptor for the
4793 	 * firmware work request header, LSO CPL, and TX_PKT_XT CPL.
4794 	 * If this descriptor is the last descriptor in the ring, wrap
4795 	 * around to the front of the ring explicitly for the start of
4796 	 * the sgl.
4797 	 */
4798 	if (dst == (void *)&eq->desc[eq->sidx]) {
4799 		dst = (void *)&eq->desc[0];
4800 		write_gl_to_txd(txq, m0, &dst, 0);
4801 	} else
4802 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4803 	txq->sgl_wrs++;
4804 
4805 	txq->txpkt_wrs++;
4806 
4807 	txsd = &txq->sdesc[eq->pidx];
4808 	txsd->m = m0;
4809 	txsd->desc_used = ndesc;
4810 
4811 	return (ndesc);
4812 }
4813 
4814 /*
4815  * Write a raw WR to the hardware descriptors, update the software
4816  * descriptor, and advance the pidx.  It is guaranteed that enough
4817  * descriptors are available.
4818  *
4819  * The return value is the # of hardware descriptors used.
4820  */
4821 static u_int
4822 write_raw_wr(struct sge_txq *txq, void *wr, struct mbuf *m0, u_int available)
4823 {
4824 	struct sge_eq *eq = &txq->eq;
4825 	struct tx_sdesc *txsd;
4826 	struct mbuf *m;
4827 	caddr_t dst;
4828 	int len16, ndesc;
4829 
4830 	len16 = mbuf_len16(m0);
4831 	ndesc = howmany(len16, EQ_ESIZE / 16);
4832 	MPASS(ndesc <= available);
4833 
4834 	dst = wr;
4835 	for (m = m0; m != NULL; m = m->m_next)
4836 		copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4837 
4838 	txq->raw_wrs++;
4839 
4840 	txsd = &txq->sdesc[eq->pidx];
4841 	txsd->m = m0;
4842 	txsd->desc_used = ndesc;
4843 
4844 	return (ndesc);
4845 }
4846 
4847 /*
4848  * Write a txpkt WR for this packet to the hardware descriptors, update the
4849  * software descriptor, and advance the pidx.  It is guaranteed that enough
4850  * descriptors are available.
4851  *
4852  * The return value is the # of hardware descriptors used.
4853  */
4854 static u_int
4855 write_txpkt_wr(struct sge_txq *txq, struct fw_eth_tx_pkt_wr *wr,
4856     struct mbuf *m0, u_int available)
4857 {
4858 	struct sge_eq *eq = &txq->eq;
4859 	struct tx_sdesc *txsd;
4860 	struct cpl_tx_pkt_core *cpl;
4861 	uint32_t ctrl;	/* used in many unrelated places */
4862 	uint64_t ctrl1;
4863 	int len16, ndesc, pktlen, nsegs;
4864 	caddr_t dst;
4865 
4866 	TXQ_LOCK_ASSERT_OWNED(txq);
4867 	M_ASSERTPKTHDR(m0);
4868 	MPASS(available > 0 && available < eq->sidx);
4869 
4870 	len16 = mbuf_len16(m0);
4871 	nsegs = mbuf_nsegs(m0);
4872 	pktlen = m0->m_pkthdr.len;
4873 	ctrl = sizeof(struct cpl_tx_pkt_core);
4874 	if (needs_tso(m0))
4875 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
4876 	else if (!(mbuf_cflags(m0) & MC_NOMAP) && pktlen <= imm_payload(2) &&
4877 	    available >= 2) {
4878 		/* Immediate data.  Recalculate len16 and set nsegs to 0. */
4879 		ctrl += pktlen;
4880 		len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) +
4881 		    sizeof(struct cpl_tx_pkt_core) + pktlen, 16);
4882 		nsegs = 0;
4883 	}
4884 	ndesc = howmany(len16, EQ_ESIZE / 16);
4885 	MPASS(ndesc <= available);
4886 
4887 	/* Firmware work request header */
4888 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
4889 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) |
4890 	    V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl));
4891 
4892 	ctrl = V_FW_WR_LEN16(len16);
4893 	wr->equiq_to_len16 = htobe32(ctrl);
4894 	wr->r3 = 0;
4895 
4896 	if (needs_tso(m0)) {
4897 		struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
4898 
4899 		KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
4900 		    m0->m_pkthdr.l4hlen > 0,
4901 		    ("%s: mbuf %p needs TSO but missing header lengths",
4902 			__func__, m0));
4903 
4904 		ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE |
4905 		    F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2)
4906 		    | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
4907 		if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
4908 			ctrl |= V_LSO_ETHHDR_LEN(1);
4909 		if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
4910 			ctrl |= F_LSO_IPV6;
4911 
4912 		lso->lso_ctrl = htobe32(ctrl);
4913 		lso->ipid_ofst = htobe16(0);
4914 		lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
4915 		lso->seqno_offset = htobe32(0);
4916 		lso->len = htobe32(pktlen);
4917 
4918 		cpl = (void *)(lso + 1);
4919 
4920 		txq->tso_wrs++;
4921 	} else
4922 		cpl = (void *)(wr + 1);
4923 
4924 	/* Checksum offload */
4925 	ctrl1 = 0;
4926 	if (needs_l3_csum(m0) == 0)
4927 		ctrl1 |= F_TXPKT_IPCSUM_DIS;
4928 	if (needs_l4_csum(m0) == 0)
4929 		ctrl1 |= F_TXPKT_L4CSUM_DIS;
4930 	if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
4931 	    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
4932 		txq->txcsum++;	/* some hardware assistance provided */
4933 
4934 	/* VLAN tag insertion */
4935 	if (needs_vlan_insertion(m0)) {
4936 		ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
4937 		txq->vlan_insertion++;
4938 	}
4939 
4940 	/* CPL header */
4941 	cpl->ctrl0 = txq->cpl_ctrl0;
4942 	cpl->pack = 0;
4943 	cpl->len = htobe16(pktlen);
4944 	cpl->ctrl1 = htobe64(ctrl1);
4945 
4946 	/* SGL */
4947 	dst = (void *)(cpl + 1);
4948 	if (nsegs > 0) {
4949 
4950 		write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx);
4951 		txq->sgl_wrs++;
4952 	} else {
4953 		struct mbuf *m;
4954 
4955 		for (m = m0; m != NULL; m = m->m_next) {
4956 			copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len);
4957 #ifdef INVARIANTS
4958 			pktlen -= m->m_len;
4959 #endif
4960 		}
4961 #ifdef INVARIANTS
4962 		KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen));
4963 #endif
4964 		txq->imm_wrs++;
4965 	}
4966 
4967 	txq->txpkt_wrs++;
4968 
4969 	txsd = &txq->sdesc[eq->pidx];
4970 	txsd->m = m0;
4971 	txsd->desc_used = ndesc;
4972 
4973 	return (ndesc);
4974 }
4975 
4976 static int
4977 try_txpkts(struct mbuf *m, struct mbuf *n, struct txpkts *txp, u_int available)
4978 {
4979 	u_int needed, nsegs1, nsegs2, l1, l2;
4980 
4981 	if (cannot_use_txpkts(m) || cannot_use_txpkts(n))
4982 		return (1);
4983 
4984 	nsegs1 = mbuf_nsegs(m);
4985 	nsegs2 = mbuf_nsegs(n);
4986 	if (nsegs1 + nsegs2 == 2) {
4987 		txp->wr_type = 1;
4988 		l1 = l2 = txpkts1_len16();
4989 	} else {
4990 		txp->wr_type = 0;
4991 		l1 = txpkts0_len16(nsegs1);
4992 		l2 = txpkts0_len16(nsegs2);
4993 	}
4994 	txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) + l1 + l2;
4995 	needed = howmany(txp->len16, EQ_ESIZE / 16);
4996 	if (needed > SGE_MAX_WR_NDESC || needed > available)
4997 		return (1);
4998 
4999 	txp->plen = m->m_pkthdr.len + n->m_pkthdr.len;
5000 	if (txp->plen > 65535)
5001 		return (1);
5002 
5003 	txp->npkt = 2;
5004 	set_mbuf_len16(m, l1);
5005 	set_mbuf_len16(n, l2);
5006 
5007 	return (0);
5008 }
5009 
5010 static int
5011 add_to_txpkts(struct mbuf *m, struct txpkts *txp, u_int available)
5012 {
5013 	u_int plen, len16, needed, nsegs;
5014 
5015 	MPASS(txp->wr_type == 0 || txp->wr_type == 1);
5016 
5017 	if (cannot_use_txpkts(m))
5018 		return (1);
5019 
5020 	nsegs = mbuf_nsegs(m);
5021 	if (txp->wr_type == 1 && nsegs != 1)
5022 		return (1);
5023 
5024 	plen = txp->plen + m->m_pkthdr.len;
5025 	if (plen > 65535)
5026 		return (1);
5027 
5028 	if (txp->wr_type == 0)
5029 		len16 = txpkts0_len16(nsegs);
5030 	else
5031 		len16 = txpkts1_len16();
5032 	needed = howmany(txp->len16 + len16, EQ_ESIZE / 16);
5033 	if (needed > SGE_MAX_WR_NDESC || needed > available)
5034 		return (1);
5035 
5036 	txp->npkt++;
5037 	txp->plen = plen;
5038 	txp->len16 += len16;
5039 	set_mbuf_len16(m, len16);
5040 
5041 	return (0);
5042 }
5043 
5044 /*
5045  * Write a txpkts WR for the packets in txp to the hardware descriptors, update
5046  * the software descriptor, and advance the pidx.  It is guaranteed that enough
5047  * descriptors are available.
5048  *
5049  * The return value is the # of hardware descriptors used.
5050  */
5051 static u_int
5052 write_txpkts_wr(struct sge_txq *txq, struct fw_eth_tx_pkts_wr *wr,
5053     struct mbuf *m0, const struct txpkts *txp, u_int available)
5054 {
5055 	struct sge_eq *eq = &txq->eq;
5056 	struct tx_sdesc *txsd;
5057 	struct cpl_tx_pkt_core *cpl;
5058 	uint32_t ctrl;
5059 	uint64_t ctrl1;
5060 	int ndesc, checkwrap;
5061 	struct mbuf *m;
5062 	void *flitp;
5063 
5064 	TXQ_LOCK_ASSERT_OWNED(txq);
5065 	MPASS(txp->npkt > 0);
5066 	MPASS(txp->plen < 65536);
5067 	MPASS(m0 != NULL);
5068 	MPASS(m0->m_nextpkt != NULL);
5069 	MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16));
5070 	MPASS(available > 0 && available < eq->sidx);
5071 
5072 	ndesc = howmany(txp->len16, EQ_ESIZE / 16);
5073 	MPASS(ndesc <= available);
5074 
5075 	MPASS(wr == (void *)&eq->desc[eq->pidx]);
5076 	wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR));
5077 	ctrl = V_FW_WR_LEN16(txp->len16);
5078 	wr->equiq_to_len16 = htobe32(ctrl);
5079 	wr->plen = htobe16(txp->plen);
5080 	wr->npkt = txp->npkt;
5081 	wr->r3 = 0;
5082 	wr->type = txp->wr_type;
5083 	flitp = wr + 1;
5084 
5085 	/*
5086 	 * At this point we are 16B into a hardware descriptor.  If checkwrap is
5087 	 * set then we know the WR is going to wrap around somewhere.  We'll
5088 	 * check for that at appropriate points.
5089 	 */
5090 	checkwrap = eq->sidx - ndesc < eq->pidx;
5091 	for (m = m0; m != NULL; m = m->m_nextpkt) {
5092 		if (txp->wr_type == 0) {
5093 			struct ulp_txpkt *ulpmc;
5094 			struct ulptx_idata *ulpsc;
5095 
5096 			/* ULP master command */
5097 			ulpmc = flitp;
5098 			ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) |
5099 			    V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid));
5100 			ulpmc->len = htobe32(mbuf_len16(m));
5101 
5102 			/* ULP subcommand */
5103 			ulpsc = (void *)(ulpmc + 1);
5104 			ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) |
5105 			    F_ULP_TX_SC_MORE);
5106 			ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core));
5107 
5108 			cpl = (void *)(ulpsc + 1);
5109 			if (checkwrap &&
5110 			    (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx])
5111 				cpl = (void *)&eq->desc[0];
5112 		} else {
5113 			cpl = flitp;
5114 		}
5115 
5116 		/* Checksum offload */
5117 		ctrl1 = 0;
5118 		if (needs_l3_csum(m) == 0)
5119 			ctrl1 |= F_TXPKT_IPCSUM_DIS;
5120 		if (needs_l4_csum(m) == 0)
5121 			ctrl1 |= F_TXPKT_L4CSUM_DIS;
5122 		if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP |
5123 		    CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO))
5124 			txq->txcsum++;	/* some hardware assistance provided */
5125 
5126 		/* VLAN tag insertion */
5127 		if (needs_vlan_insertion(m)) {
5128 			ctrl1 |= F_TXPKT_VLAN_VLD |
5129 			    V_TXPKT_VLAN(m->m_pkthdr.ether_vtag);
5130 			txq->vlan_insertion++;
5131 		}
5132 
5133 		/* CPL header */
5134 		cpl->ctrl0 = txq->cpl_ctrl0;
5135 		cpl->pack = 0;
5136 		cpl->len = htobe16(m->m_pkthdr.len);
5137 		cpl->ctrl1 = htobe64(ctrl1);
5138 
5139 		flitp = cpl + 1;
5140 		if (checkwrap &&
5141 		    (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx])
5142 			flitp = (void *)&eq->desc[0];
5143 
5144 		write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap);
5145 
5146 	}
5147 
5148 	if (txp->wr_type == 0) {
5149 		txq->txpkts0_pkts += txp->npkt;
5150 		txq->txpkts0_wrs++;
5151 	} else {
5152 		txq->txpkts1_pkts += txp->npkt;
5153 		txq->txpkts1_wrs++;
5154 	}
5155 
5156 	txsd = &txq->sdesc[eq->pidx];
5157 	txsd->m = m0;
5158 	txsd->desc_used = ndesc;
5159 
5160 	return (ndesc);
5161 }
5162 
5163 /*
5164  * If the SGL ends on an address that is not 16 byte aligned, this function will
5165  * add a 0 filled flit at the end.
5166  */
5167 static void
5168 write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap)
5169 {
5170 	struct sge_eq *eq = &txq->eq;
5171 	struct sglist *gl = txq->gl;
5172 	struct sglist_seg *seg;
5173 	__be64 *flitp, *wrap;
5174 	struct ulptx_sgl *usgl;
5175 	int i, nflits, nsegs;
5176 
5177 	KASSERT(((uintptr_t)(*to) & 0xf) == 0,
5178 	    ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to));
5179 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
5180 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
5181 
5182 	get_pkt_gl(m, gl);
5183 	nsegs = gl->sg_nseg;
5184 	MPASS(nsegs > 0);
5185 
5186 	nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2;
5187 	flitp = (__be64 *)(*to);
5188 	wrap = (__be64 *)(&eq->desc[eq->sidx]);
5189 	seg = &gl->sg_segs[0];
5190 	usgl = (void *)flitp;
5191 
5192 	/*
5193 	 * We start at a 16 byte boundary somewhere inside the tx descriptor
5194 	 * ring, so we're at least 16 bytes away from the status page.  There is
5195 	 * no chance of a wrap around in the middle of usgl (which is 16 bytes).
5196 	 */
5197 
5198 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
5199 	    V_ULPTX_NSGE(nsegs));
5200 	usgl->len0 = htobe32(seg->ss_len);
5201 	usgl->addr0 = htobe64(seg->ss_paddr);
5202 	seg++;
5203 
5204 	if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) {
5205 
5206 		/* Won't wrap around at all */
5207 
5208 		for (i = 0; i < nsegs - 1; i++, seg++) {
5209 			usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len);
5210 			usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr);
5211 		}
5212 		if (i & 1)
5213 			usgl->sge[i / 2].len[1] = htobe32(0);
5214 		flitp += nflits;
5215 	} else {
5216 
5217 		/* Will wrap somewhere in the rest of the SGL */
5218 
5219 		/* 2 flits already written, write the rest flit by flit */
5220 		flitp = (void *)(usgl + 1);
5221 		for (i = 0; i < nflits - 2; i++) {
5222 			if (flitp == wrap)
5223 				flitp = (void *)eq->desc;
5224 			*flitp++ = get_flit(seg, nsegs - 1, i);
5225 		}
5226 	}
5227 
5228 	if (nflits & 1) {
5229 		MPASS(((uintptr_t)flitp) & 0xf);
5230 		*flitp++ = 0;
5231 	}
5232 
5233 	MPASS((((uintptr_t)flitp) & 0xf) == 0);
5234 	if (__predict_false(flitp == wrap))
5235 		*to = (void *)eq->desc;
5236 	else
5237 		*to = (void *)flitp;
5238 }
5239 
5240 static inline void
5241 copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len)
5242 {
5243 
5244 	MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]);
5245 	MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]);
5246 
5247 	if (__predict_true((uintptr_t)(*to) + len <=
5248 	    (uintptr_t)&eq->desc[eq->sidx])) {
5249 		bcopy(from, *to, len);
5250 		(*to) += len;
5251 	} else {
5252 		int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to);
5253 
5254 		bcopy(from, *to, portion);
5255 		from += portion;
5256 		portion = len - portion;	/* remaining */
5257 		bcopy(from, (void *)eq->desc, portion);
5258 		(*to) = (caddr_t)eq->desc + portion;
5259 	}
5260 }
5261 
5262 static inline void
5263 ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n)
5264 {
5265 	u_int db;
5266 
5267 	MPASS(n > 0);
5268 
5269 	db = eq->doorbells;
5270 	if (n > 1)
5271 		clrbit(&db, DOORBELL_WCWR);
5272 	wmb();
5273 
5274 	switch (ffs(db) - 1) {
5275 	case DOORBELL_UDB:
5276 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
5277 		break;
5278 
5279 	case DOORBELL_WCWR: {
5280 		volatile uint64_t *dst, *src;
5281 		int i;
5282 
5283 		/*
5284 		 * Queues whose 128B doorbell segment fits in the page do not
5285 		 * use relative qid (udb_qid is always 0).  Only queues with
5286 		 * doorbell segments can do WCWR.
5287 		 */
5288 		KASSERT(eq->udb_qid == 0 && n == 1,
5289 		    ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p",
5290 		    __func__, eq->doorbells, n, eq->dbidx, eq));
5291 
5292 		dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET -
5293 		    UDBS_DB_OFFSET);
5294 		i = eq->dbidx;
5295 		src = (void *)&eq->desc[i];
5296 		while (src != (void *)&eq->desc[i + 1])
5297 			*dst++ = *src++;
5298 		wmb();
5299 		break;
5300 	}
5301 
5302 	case DOORBELL_UDBWC:
5303 		*eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n));
5304 		wmb();
5305 		break;
5306 
5307 	case DOORBELL_KDB:
5308 		t4_write_reg(sc, sc->sge_kdoorbell_reg,
5309 		    V_QID(eq->cntxt_id) | V_PIDX(n));
5310 		break;
5311 	}
5312 
5313 	IDXINCR(eq->dbidx, n, eq->sidx);
5314 }
5315 
5316 static inline u_int
5317 reclaimable_tx_desc(struct sge_eq *eq)
5318 {
5319 	uint16_t hw_cidx;
5320 
5321 	hw_cidx = read_hw_cidx(eq);
5322 	return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx));
5323 }
5324 
5325 static inline u_int
5326 total_available_tx_desc(struct sge_eq *eq)
5327 {
5328 	uint16_t hw_cidx, pidx;
5329 
5330 	hw_cidx = read_hw_cidx(eq);
5331 	pidx = eq->pidx;
5332 
5333 	if (pidx == hw_cidx)
5334 		return (eq->sidx - 1);
5335 	else
5336 		return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1);
5337 }
5338 
5339 static inline uint16_t
5340 read_hw_cidx(struct sge_eq *eq)
5341 {
5342 	struct sge_qstat *spg = (void *)&eq->desc[eq->sidx];
5343 	uint16_t cidx = spg->cidx;	/* stable snapshot */
5344 
5345 	return (be16toh(cidx));
5346 }
5347 
5348 /*
5349  * Reclaim 'n' descriptors approximately.
5350  */
5351 static u_int
5352 reclaim_tx_descs(struct sge_txq *txq, u_int n)
5353 {
5354 	struct tx_sdesc *txsd;
5355 	struct sge_eq *eq = &txq->eq;
5356 	u_int can_reclaim, reclaimed;
5357 
5358 	TXQ_LOCK_ASSERT_OWNED(txq);
5359 	MPASS(n > 0);
5360 
5361 	reclaimed = 0;
5362 	can_reclaim = reclaimable_tx_desc(eq);
5363 	while (can_reclaim && reclaimed < n) {
5364 		int ndesc;
5365 		struct mbuf *m, *nextpkt;
5366 
5367 		txsd = &txq->sdesc[eq->cidx];
5368 		ndesc = txsd->desc_used;
5369 
5370 		/* Firmware doesn't return "partial" credits. */
5371 		KASSERT(can_reclaim >= ndesc,
5372 		    ("%s: unexpected number of credits: %d, %d",
5373 		    __func__, can_reclaim, ndesc));
5374 		KASSERT(ndesc != 0,
5375 		    ("%s: descriptor with no credits: cidx %d",
5376 		    __func__, eq->cidx));
5377 
5378 		for (m = txsd->m; m != NULL; m = nextpkt) {
5379 			nextpkt = m->m_nextpkt;
5380 			m->m_nextpkt = NULL;
5381 			m_freem(m);
5382 		}
5383 		reclaimed += ndesc;
5384 		can_reclaim -= ndesc;
5385 		IDXINCR(eq->cidx, ndesc, eq->sidx);
5386 	}
5387 
5388 	return (reclaimed);
5389 }
5390 
5391 static void
5392 tx_reclaim(void *arg, int n)
5393 {
5394 	struct sge_txq *txq = arg;
5395 	struct sge_eq *eq = &txq->eq;
5396 
5397 	do {
5398 		if (TXQ_TRYLOCK(txq) == 0)
5399 			break;
5400 		n = reclaim_tx_descs(txq, 32);
5401 		if (eq->cidx == eq->pidx)
5402 			eq->equeqidx = eq->pidx;
5403 		TXQ_UNLOCK(txq);
5404 	} while (n > 0);
5405 }
5406 
5407 static __be64
5408 get_flit(struct sglist_seg *segs, int nsegs, int idx)
5409 {
5410 	int i = (idx / 3) * 2;
5411 
5412 	switch (idx % 3) {
5413 	case 0: {
5414 		uint64_t rc;
5415 
5416 		rc = (uint64_t)segs[i].ss_len << 32;
5417 		if (i + 1 < nsegs)
5418 			rc |= (uint64_t)(segs[i + 1].ss_len);
5419 
5420 		return (htobe64(rc));
5421 	}
5422 	case 1:
5423 		return (htobe64(segs[i].ss_paddr));
5424 	case 2:
5425 		return (htobe64(segs[i + 1].ss_paddr));
5426 	}
5427 
5428 	return (0);
5429 }
5430 
5431 static void
5432 find_best_refill_source(struct adapter *sc, struct sge_fl *fl, int maxp)
5433 {
5434 	int8_t zidx, hwidx, idx;
5435 	uint16_t region1, region3;
5436 	int spare, spare_needed, n;
5437 	struct sw_zone_info *swz;
5438 	struct hw_buf_info *hwb, *hwb_list = &sc->sge.hw_buf_info[0];
5439 
5440 	/*
5441 	 * Buffer Packing: Look for PAGE_SIZE or larger zone which has a bufsize
5442 	 * large enough for the max payload and cluster metadata.  Otherwise
5443 	 * settle for the largest bufsize that leaves enough room in the cluster
5444 	 * for metadata.
5445 	 *
5446 	 * Without buffer packing: Look for the smallest zone which has a
5447 	 * bufsize large enough for the max payload.  Settle for the largest
5448 	 * bufsize available if there's nothing big enough for max payload.
5449 	 */
5450 	spare_needed = fl->flags & FL_BUF_PACKING ? CL_METADATA_SIZE : 0;
5451 	swz = &sc->sge.sw_zone_info[0];
5452 	hwidx = -1;
5453 	for (zidx = 0; zidx < SW_ZONE_SIZES; zidx++, swz++) {
5454 		if (swz->size > largest_rx_cluster) {
5455 			if (__predict_true(hwidx != -1))
5456 				break;
5457 
5458 			/*
5459 			 * This is a misconfiguration.  largest_rx_cluster is
5460 			 * preventing us from finding a refill source.  See
5461 			 * dev.t5nex.<n>.buffer_sizes to figure out why.
5462 			 */
5463 			device_printf(sc->dev, "largest_rx_cluster=%u leaves no"
5464 			    " refill source for fl %p (dma %u).  Ignored.\n",
5465 			    largest_rx_cluster, fl, maxp);
5466 		}
5467 		for (idx = swz->head_hwidx; idx != -1; idx = hwb->next) {
5468 			hwb = &hwb_list[idx];
5469 			spare = swz->size - hwb->size;
5470 			if (spare < spare_needed)
5471 				continue;
5472 
5473 			hwidx = idx;		/* best option so far */
5474 			if (hwb->size >= maxp) {
5475 
5476 				if ((fl->flags & FL_BUF_PACKING) == 0)
5477 					goto done; /* stop looking (not packing) */
5478 
5479 				if (swz->size >= safest_rx_cluster)
5480 					goto done; /* stop looking (packing) */
5481 			}
5482 			break;		/* keep looking, next zone */
5483 		}
5484 	}
5485 done:
5486 	/* A usable hwidx has been located. */
5487 	MPASS(hwidx != -1);
5488 	hwb = &hwb_list[hwidx];
5489 	zidx = hwb->zidx;
5490 	swz = &sc->sge.sw_zone_info[zidx];
5491 	region1 = 0;
5492 	region3 = swz->size - hwb->size;
5493 
5494 	/*
5495 	 * Stay within this zone and see if there is a better match when mbuf
5496 	 * inlining is allowed.  Remember that the hwidx's are sorted in
5497 	 * decreasing order of size (so in increasing order of spare area).
5498 	 */
5499 	for (idx = hwidx; idx != -1; idx = hwb->next) {
5500 		hwb = &hwb_list[idx];
5501 		spare = swz->size - hwb->size;
5502 
5503 		if (allow_mbufs_in_cluster == 0 || hwb->size < maxp)
5504 			break;
5505 
5506 		/*
5507 		 * Do not inline mbufs if doing so would violate the pad/pack
5508 		 * boundary alignment requirement.
5509 		 */
5510 		if (fl_pad && (MSIZE % sc->params.sge.pad_boundary) != 0)
5511 			continue;
5512 		if (fl->flags & FL_BUF_PACKING &&
5513 		    (MSIZE % sc->params.sge.pack_boundary) != 0)
5514 			continue;
5515 
5516 		if (spare < CL_METADATA_SIZE + MSIZE)
5517 			continue;
5518 		n = (spare - CL_METADATA_SIZE) / MSIZE;
5519 		if (n > howmany(hwb->size, maxp))
5520 			break;
5521 
5522 		hwidx = idx;
5523 		if (fl->flags & FL_BUF_PACKING) {
5524 			region1 = n * MSIZE;
5525 			region3 = spare - region1;
5526 		} else {
5527 			region1 = MSIZE;
5528 			region3 = spare - region1;
5529 			break;
5530 		}
5531 	}
5532 
5533 	KASSERT(zidx >= 0 && zidx < SW_ZONE_SIZES,
5534 	    ("%s: bad zone %d for fl %p, maxp %d", __func__, zidx, fl, maxp));
5535 	KASSERT(hwidx >= 0 && hwidx <= SGE_FLBUF_SIZES,
5536 	    ("%s: bad hwidx %d for fl %p, maxp %d", __func__, hwidx, fl, maxp));
5537 	KASSERT(region1 + sc->sge.hw_buf_info[hwidx].size + region3 ==
5538 	    sc->sge.sw_zone_info[zidx].size,
5539 	    ("%s: bad buffer layout for 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 	if (fl->flags & FL_BUF_PACKING || region1 > 0) {
5544 		KASSERT(region3 >= CL_METADATA_SIZE,
5545 		    ("%s: no room for metadata.  fl %p, maxp %d; "
5546 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5547 		    sc->sge.sw_zone_info[zidx].size, region1,
5548 		    sc->sge.hw_buf_info[hwidx].size, region3));
5549 		KASSERT(region1 % MSIZE == 0,
5550 		    ("%s: bad mbuf region for fl %p, maxp %d. "
5551 		    "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp,
5552 		    sc->sge.sw_zone_info[zidx].size, region1,
5553 		    sc->sge.hw_buf_info[hwidx].size, region3));
5554 	}
5555 
5556 	fl->cll_def.zidx = zidx;
5557 	fl->cll_def.hwidx = hwidx;
5558 	fl->cll_def.region1 = region1;
5559 	fl->cll_def.region3 = region3;
5560 }
5561 
5562 static void
5563 find_safe_refill_source(struct adapter *sc, struct sge_fl *fl)
5564 {
5565 	struct sge *s = &sc->sge;
5566 	struct hw_buf_info *hwb;
5567 	struct sw_zone_info *swz;
5568 	int spare;
5569 	int8_t hwidx;
5570 
5571 	if (fl->flags & FL_BUF_PACKING)
5572 		hwidx = s->safe_hwidx2;	/* with room for metadata */
5573 	else if (allow_mbufs_in_cluster && s->safe_hwidx2 != -1) {
5574 		hwidx = s->safe_hwidx2;
5575 		hwb = &s->hw_buf_info[hwidx];
5576 		swz = &s->sw_zone_info[hwb->zidx];
5577 		spare = swz->size - hwb->size;
5578 
5579 		/* no good if there isn't room for an mbuf as well */
5580 		if (spare < CL_METADATA_SIZE + MSIZE)
5581 			hwidx = s->safe_hwidx1;
5582 	} else
5583 		hwidx = s->safe_hwidx1;
5584 
5585 	if (hwidx == -1) {
5586 		/* No fallback source */
5587 		fl->cll_alt.hwidx = -1;
5588 		fl->cll_alt.zidx = -1;
5589 
5590 		return;
5591 	}
5592 
5593 	hwb = &s->hw_buf_info[hwidx];
5594 	swz = &s->sw_zone_info[hwb->zidx];
5595 	spare = swz->size - hwb->size;
5596 	fl->cll_alt.hwidx = hwidx;
5597 	fl->cll_alt.zidx = hwb->zidx;
5598 	if (allow_mbufs_in_cluster &&
5599 	    (fl_pad == 0 || (MSIZE % sc->params.sge.pad_boundary) == 0))
5600 		fl->cll_alt.region1 = ((spare - CL_METADATA_SIZE) / MSIZE) * MSIZE;
5601 	else
5602 		fl->cll_alt.region1 = 0;
5603 	fl->cll_alt.region3 = spare - fl->cll_alt.region1;
5604 }
5605 
5606 static void
5607 add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl)
5608 {
5609 	mtx_lock(&sc->sfl_lock);
5610 	FL_LOCK(fl);
5611 	if ((fl->flags & FL_DOOMED) == 0) {
5612 		fl->flags |= FL_STARVING;
5613 		TAILQ_INSERT_TAIL(&sc->sfl, fl, link);
5614 		callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc);
5615 	}
5616 	FL_UNLOCK(fl);
5617 	mtx_unlock(&sc->sfl_lock);
5618 }
5619 
5620 static void
5621 handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq)
5622 {
5623 	struct sge_wrq *wrq = (void *)eq;
5624 
5625 	atomic_readandclear_int(&eq->equiq);
5626 	taskqueue_enqueue(sc->tq[eq->tx_chan], &wrq->wrq_tx_task);
5627 }
5628 
5629 static void
5630 handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq)
5631 {
5632 	struct sge_txq *txq = (void *)eq;
5633 
5634 	MPASS((eq->flags & EQ_TYPEMASK) == EQ_ETH);
5635 
5636 	atomic_readandclear_int(&eq->equiq);
5637 	mp_ring_check_drainage(txq->r, 0);
5638 	taskqueue_enqueue(sc->tq[eq->tx_chan], &txq->tx_reclaim_task);
5639 }
5640 
5641 static int
5642 handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss,
5643     struct mbuf *m)
5644 {
5645 	const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1);
5646 	unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid));
5647 	struct adapter *sc = iq->adapter;
5648 	struct sge *s = &sc->sge;
5649 	struct sge_eq *eq;
5650 	static void (*h[])(struct adapter *, struct sge_eq *) = {NULL,
5651 		&handle_wrq_egr_update, &handle_eth_egr_update,
5652 		&handle_wrq_egr_update};
5653 
5654 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5655 	    rss->opcode));
5656 
5657 	eq = s->eqmap[qid - s->eq_start - s->eq_base];
5658 	(*h[eq->flags & EQ_TYPEMASK])(sc, eq);
5659 
5660 	return (0);
5661 }
5662 
5663 /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */
5664 CTASSERT(offsetof(struct cpl_fw4_msg, data) == \
5665     offsetof(struct cpl_fw6_msg, data));
5666 
5667 static int
5668 handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
5669 {
5670 	struct adapter *sc = iq->adapter;
5671 	const struct cpl_fw6_msg *cpl = (const void *)(rss + 1);
5672 
5673 	KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__,
5674 	    rss->opcode));
5675 
5676 	if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) {
5677 		const struct rss_header *rss2;
5678 
5679 		rss2 = (const struct rss_header *)&cpl->data[0];
5680 		return (t4_cpl_handler[rss2->opcode](iq, rss2, m));
5681 	}
5682 
5683 	return (t4_fw_msg_handler[cpl->type](sc, &cpl->data[0]));
5684 }
5685 
5686 /**
5687  *	t4_handle_wrerr_rpl - process a FW work request error message
5688  *	@adap: the adapter
5689  *	@rpl: start of the FW message
5690  */
5691 static int
5692 t4_handle_wrerr_rpl(struct adapter *adap, const __be64 *rpl)
5693 {
5694 	u8 opcode = *(const u8 *)rpl;
5695 	const struct fw_error_cmd *e = (const void *)rpl;
5696 	unsigned int i;
5697 
5698 	if (opcode != FW_ERROR_CMD) {
5699 		log(LOG_ERR,
5700 		    "%s: Received WRERR_RPL message with opcode %#x\n",
5701 		    device_get_nameunit(adap->dev), opcode);
5702 		return (EINVAL);
5703 	}
5704 	log(LOG_ERR, "%s: FW_ERROR (%s) ", device_get_nameunit(adap->dev),
5705 	    G_FW_ERROR_CMD_FATAL(be32toh(e->op_to_type)) ? "fatal" :
5706 	    "non-fatal");
5707 	switch (G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type))) {
5708 	case FW_ERROR_TYPE_EXCEPTION:
5709 		log(LOG_ERR, "exception info:\n");
5710 		for (i = 0; i < nitems(e->u.exception.info); i++)
5711 			log(LOG_ERR, "%s%08x", i == 0 ? "\t" : " ",
5712 			    be32toh(e->u.exception.info[i]));
5713 		log(LOG_ERR, "\n");
5714 		break;
5715 	case FW_ERROR_TYPE_HWMODULE:
5716 		log(LOG_ERR, "HW module regaddr %08x regval %08x\n",
5717 		    be32toh(e->u.hwmodule.regaddr),
5718 		    be32toh(e->u.hwmodule.regval));
5719 		break;
5720 	case FW_ERROR_TYPE_WR:
5721 		log(LOG_ERR, "WR cidx %d PF %d VF %d eqid %d hdr:\n",
5722 		    be16toh(e->u.wr.cidx),
5723 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.wr.pfn_vfn)),
5724 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.wr.pfn_vfn)),
5725 		    be32toh(e->u.wr.eqid));
5726 		for (i = 0; i < nitems(e->u.wr.wrhdr); i++)
5727 			log(LOG_ERR, "%s%02x", i == 0 ? "\t" : " ",
5728 			    e->u.wr.wrhdr[i]);
5729 		log(LOG_ERR, "\n");
5730 		break;
5731 	case FW_ERROR_TYPE_ACL:
5732 		log(LOG_ERR, "ACL cidx %d PF %d VF %d eqid %d %s",
5733 		    be16toh(e->u.acl.cidx),
5734 		    G_FW_ERROR_CMD_PFN(be16toh(e->u.acl.pfn_vfn)),
5735 		    G_FW_ERROR_CMD_VFN(be16toh(e->u.acl.pfn_vfn)),
5736 		    be32toh(e->u.acl.eqid),
5737 		    G_FW_ERROR_CMD_MV(be16toh(e->u.acl.mv_pkd)) ? "vlanid" :
5738 		    "MAC");
5739 		for (i = 0; i < nitems(e->u.acl.val); i++)
5740 			log(LOG_ERR, " %02x", e->u.acl.val[i]);
5741 		log(LOG_ERR, "\n");
5742 		break;
5743 	default:
5744 		log(LOG_ERR, "type %#x\n",
5745 		    G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type)));
5746 		return (EINVAL);
5747 	}
5748 	return (0);
5749 }
5750 
5751 static int
5752 sysctl_uint16(SYSCTL_HANDLER_ARGS)
5753 {
5754 	uint16_t *id = arg1;
5755 	int i = *id;
5756 
5757 	return sysctl_handle_int(oidp, &i, 0, req);
5758 }
5759 
5760 static int
5761 sysctl_bufsizes(SYSCTL_HANDLER_ARGS)
5762 {
5763 	struct sge *s = arg1;
5764 	struct hw_buf_info *hwb = &s->hw_buf_info[0];
5765 	struct sw_zone_info *swz = &s->sw_zone_info[0];
5766 	int i, rc;
5767 	struct sbuf sb;
5768 	char c;
5769 
5770 	sbuf_new(&sb, NULL, 32, SBUF_AUTOEXTEND);
5771 	for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) {
5772 		if (hwb->zidx >= 0 && swz[hwb->zidx].size <= largest_rx_cluster)
5773 			c = '*';
5774 		else
5775 			c = '\0';
5776 
5777 		sbuf_printf(&sb, "%u%c ", hwb->size, c);
5778 	}
5779 	sbuf_trim(&sb);
5780 	sbuf_finish(&sb);
5781 	rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req);
5782 	sbuf_delete(&sb);
5783 	return (rc);
5784 }
5785 
5786 #ifdef RATELIMIT
5787 /*
5788  * len16 for a txpkt WR with a GL.  Includes the firmware work request header.
5789  */
5790 static inline u_int
5791 txpkt_eo_len16(u_int nsegs, u_int immhdrs, u_int tso)
5792 {
5793 	u_int n;
5794 
5795 	MPASS(immhdrs > 0);
5796 
5797 	n = roundup2(sizeof(struct fw_eth_tx_eo_wr) +
5798 	    sizeof(struct cpl_tx_pkt_core) + immhdrs, 16);
5799 	if (__predict_false(nsegs == 0))
5800 		goto done;
5801 
5802 	nsegs--; /* first segment is part of ulptx_sgl */
5803 	n += sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1));
5804 	if (tso)
5805 		n += sizeof(struct cpl_tx_pkt_lso_core);
5806 
5807 done:
5808 	return (howmany(n, 16));
5809 }
5810 
5811 #define ETID_FLOWC_NPARAMS 6
5812 #define ETID_FLOWC_LEN (roundup2((sizeof(struct fw_flowc_wr) + \
5813     ETID_FLOWC_NPARAMS * sizeof(struct fw_flowc_mnemval)), 16))
5814 #define ETID_FLOWC_LEN16 (howmany(ETID_FLOWC_LEN, 16))
5815 
5816 static int
5817 send_etid_flowc_wr(struct cxgbe_rate_tag *cst, struct port_info *pi,
5818     struct vi_info *vi)
5819 {
5820 	struct wrq_cookie cookie;
5821 	u_int pfvf = pi->adapter->pf << S_FW_VIID_PFN;
5822 	struct fw_flowc_wr *flowc;
5823 
5824 	mtx_assert(&cst->lock, MA_OWNED);
5825 	MPASS((cst->flags & (EO_FLOWC_PENDING | EO_FLOWC_RPL_PENDING)) ==
5826 	    EO_FLOWC_PENDING);
5827 
5828 	flowc = start_wrq_wr(cst->eo_txq, ETID_FLOWC_LEN16, &cookie);
5829 	if (__predict_false(flowc == NULL))
5830 		return (ENOMEM);
5831 
5832 	bzero(flowc, ETID_FLOWC_LEN);
5833 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
5834 	    V_FW_FLOWC_WR_NPARAMS(ETID_FLOWC_NPARAMS) | V_FW_WR_COMPL(0));
5835 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(ETID_FLOWC_LEN16) |
5836 	    V_FW_WR_FLOWID(cst->etid));
5837 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN;
5838 	flowc->mnemval[0].val = htobe32(pfvf);
5839 	flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH;
5840 	flowc->mnemval[1].val = htobe32(pi->tx_chan);
5841 	flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT;
5842 	flowc->mnemval[2].val = htobe32(pi->tx_chan);
5843 	flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID;
5844 	flowc->mnemval[3].val = htobe32(cst->iqid);
5845 	flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_EOSTATE;
5846 	flowc->mnemval[4].val = htobe32(FW_FLOWC_MNEM_EOSTATE_ESTABLISHED);
5847 	flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS;
5848 	flowc->mnemval[5].val = htobe32(cst->schedcl);
5849 
5850 	commit_wrq_wr(cst->eo_txq, flowc, &cookie);
5851 
5852 	cst->flags &= ~EO_FLOWC_PENDING;
5853 	cst->flags |= EO_FLOWC_RPL_PENDING;
5854 	MPASS(cst->tx_credits >= ETID_FLOWC_LEN16);	/* flowc is first WR. */
5855 	cst->tx_credits -= ETID_FLOWC_LEN16;
5856 
5857 	return (0);
5858 }
5859 
5860 #define ETID_FLUSH_LEN16 (howmany(sizeof (struct fw_flowc_wr), 16))
5861 
5862 void
5863 send_etid_flush_wr(struct cxgbe_rate_tag *cst)
5864 {
5865 	struct fw_flowc_wr *flowc;
5866 	struct wrq_cookie cookie;
5867 
5868 	mtx_assert(&cst->lock, MA_OWNED);
5869 
5870 	flowc = start_wrq_wr(cst->eo_txq, ETID_FLUSH_LEN16, &cookie);
5871 	if (__predict_false(flowc == NULL))
5872 		CXGBE_UNIMPLEMENTED(__func__);
5873 
5874 	bzero(flowc, ETID_FLUSH_LEN16 * 16);
5875 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
5876 	    V_FW_FLOWC_WR_NPARAMS(0) | F_FW_WR_COMPL);
5877 	flowc->flowid_len16 = htobe32(V_FW_WR_LEN16(ETID_FLUSH_LEN16) |
5878 	    V_FW_WR_FLOWID(cst->etid));
5879 
5880 	commit_wrq_wr(cst->eo_txq, flowc, &cookie);
5881 
5882 	cst->flags |= EO_FLUSH_RPL_PENDING;
5883 	MPASS(cst->tx_credits >= ETID_FLUSH_LEN16);
5884 	cst->tx_credits -= ETID_FLUSH_LEN16;
5885 	cst->ncompl++;
5886 }
5887 
5888 static void
5889 write_ethofld_wr(struct cxgbe_rate_tag *cst, struct fw_eth_tx_eo_wr *wr,
5890     struct mbuf *m0, int compl)
5891 {
5892 	struct cpl_tx_pkt_core *cpl;
5893 	uint64_t ctrl1;
5894 	uint32_t ctrl;	/* used in many unrelated places */
5895 	int len16, pktlen, nsegs, immhdrs;
5896 	caddr_t dst;
5897 	uintptr_t p;
5898 	struct ulptx_sgl *usgl;
5899 	struct sglist sg;
5900 	struct sglist_seg segs[38];	/* XXX: find real limit.  XXX: get off the stack */
5901 
5902 	mtx_assert(&cst->lock, MA_OWNED);
5903 	M_ASSERTPKTHDR(m0);
5904 	KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 &&
5905 	    m0->m_pkthdr.l4hlen > 0,
5906 	    ("%s: ethofld mbuf %p is missing header lengths", __func__, m0));
5907 
5908 	len16 = mbuf_eo_len16(m0);
5909 	nsegs = mbuf_eo_nsegs(m0);
5910 	pktlen = m0->m_pkthdr.len;
5911 	ctrl = sizeof(struct cpl_tx_pkt_core);
5912 	if (needs_tso(m0))
5913 		ctrl += sizeof(struct cpl_tx_pkt_lso_core);
5914 	immhdrs = m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen + m0->m_pkthdr.l4hlen;
5915 	ctrl += immhdrs;
5916 
5917 	wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_EO_WR) |
5918 	    V_FW_ETH_TX_EO_WR_IMMDLEN(ctrl) | V_FW_WR_COMPL(!!compl));
5919 	wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(len16) |
5920 	    V_FW_WR_FLOWID(cst->etid));
5921 	wr->r3 = 0;
5922 	if (needs_udp_csum(m0)) {
5923 		wr->u.udpseg.type = FW_ETH_TX_EO_TYPE_UDPSEG;
5924 		wr->u.udpseg.ethlen = m0->m_pkthdr.l2hlen;
5925 		wr->u.udpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
5926 		wr->u.udpseg.udplen = m0->m_pkthdr.l4hlen;
5927 		wr->u.udpseg.rtplen = 0;
5928 		wr->u.udpseg.r4 = 0;
5929 		wr->u.udpseg.mss = htobe16(pktlen - immhdrs);
5930 		wr->u.udpseg.schedpktsize = wr->u.udpseg.mss;
5931 		wr->u.udpseg.plen = htobe32(pktlen - immhdrs);
5932 		cpl = (void *)(wr + 1);
5933 	} else {
5934 		MPASS(needs_tcp_csum(m0));
5935 		wr->u.tcpseg.type = FW_ETH_TX_EO_TYPE_TCPSEG;
5936 		wr->u.tcpseg.ethlen = m0->m_pkthdr.l2hlen;
5937 		wr->u.tcpseg.iplen = htobe16(m0->m_pkthdr.l3hlen);
5938 		wr->u.tcpseg.tcplen = m0->m_pkthdr.l4hlen;
5939 		wr->u.tcpseg.tsclk_tsoff = mbuf_eo_tsclk_tsoff(m0);
5940 		wr->u.tcpseg.r4 = 0;
5941 		wr->u.tcpseg.r5 = 0;
5942 		wr->u.tcpseg.plen = htobe32(pktlen - immhdrs);
5943 
5944 		if (needs_tso(m0)) {
5945 			struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1);
5946 
5947 			wr->u.tcpseg.mss = htobe16(m0->m_pkthdr.tso_segsz);
5948 
5949 			ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) |
5950 			    F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE |
5951 			    V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) |
5952 			    V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2);
5953 			if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header))
5954 				ctrl |= V_LSO_ETHHDR_LEN(1);
5955 			if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr))
5956 				ctrl |= F_LSO_IPV6;
5957 			lso->lso_ctrl = htobe32(ctrl);
5958 			lso->ipid_ofst = htobe16(0);
5959 			lso->mss = htobe16(m0->m_pkthdr.tso_segsz);
5960 			lso->seqno_offset = htobe32(0);
5961 			lso->len = htobe32(pktlen);
5962 
5963 			cpl = (void *)(lso + 1);
5964 		} else {
5965 			wr->u.tcpseg.mss = htobe16(0xffff);
5966 			cpl = (void *)(wr + 1);
5967 		}
5968 	}
5969 
5970 	/* Checksum offload must be requested for ethofld. */
5971 	ctrl1 = 0;
5972 	MPASS(needs_l4_csum(m0));
5973 
5974 	/* VLAN tag insertion */
5975 	if (needs_vlan_insertion(m0)) {
5976 		ctrl1 |= F_TXPKT_VLAN_VLD |
5977 		    V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag);
5978 	}
5979 
5980 	/* CPL header */
5981 	cpl->ctrl0 = cst->ctrl0;
5982 	cpl->pack = 0;
5983 	cpl->len = htobe16(pktlen);
5984 	cpl->ctrl1 = htobe64(ctrl1);
5985 
5986 	/* Copy Ethernet, IP & TCP/UDP hdrs as immediate data */
5987 	p = (uintptr_t)(cpl + 1);
5988 	m_copydata(m0, 0, immhdrs, (void *)p);
5989 
5990 	/* SGL */
5991 	dst = (void *)(cpl + 1);
5992 	if (nsegs > 0) {
5993 		int i, pad;
5994 
5995 		/* zero-pad upto next 16Byte boundary, if not 16Byte aligned */
5996 		p += immhdrs;
5997 		pad = 16 - (immhdrs & 0xf);
5998 		bzero((void *)p, pad);
5999 
6000 		usgl = (void *)(p + pad);
6001 		usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
6002 		    V_ULPTX_NSGE(nsegs));
6003 
6004 		sglist_init(&sg, nitems(segs), segs);
6005 		for (; m0 != NULL; m0 = m0->m_next) {
6006 			if (__predict_false(m0->m_len == 0))
6007 				continue;
6008 			if (immhdrs >= m0->m_len) {
6009 				immhdrs -= m0->m_len;
6010 				continue;
6011 			}
6012 
6013 			sglist_append(&sg, mtod(m0, char *) + immhdrs,
6014 			    m0->m_len - immhdrs);
6015 			immhdrs = 0;
6016 		}
6017 		MPASS(sg.sg_nseg == nsegs);
6018 
6019 		/*
6020 		 * Zero pad last 8B in case the WR doesn't end on a 16B
6021 		 * boundary.
6022 		 */
6023 		*(uint64_t *)((char *)wr + len16 * 16 - 8) = 0;
6024 
6025 		usgl->len0 = htobe32(segs[0].ss_len);
6026 		usgl->addr0 = htobe64(segs[0].ss_paddr);
6027 		for (i = 0; i < nsegs - 1; i++) {
6028 			usgl->sge[i / 2].len[i & 1] = htobe32(segs[i + 1].ss_len);
6029 			usgl->sge[i / 2].addr[i & 1] = htobe64(segs[i + 1].ss_paddr);
6030 		}
6031 		if (i & 1)
6032 			usgl->sge[i / 2].len[1] = htobe32(0);
6033 	}
6034 
6035 }
6036 
6037 static void
6038 ethofld_tx(struct cxgbe_rate_tag *cst)
6039 {
6040 	struct mbuf *m;
6041 	struct wrq_cookie cookie;
6042 	int next_credits, compl;
6043 	struct fw_eth_tx_eo_wr *wr;
6044 
6045 	mtx_assert(&cst->lock, MA_OWNED);
6046 
6047 	while ((m = mbufq_first(&cst->pending_tx)) != NULL) {
6048 		M_ASSERTPKTHDR(m);
6049 
6050 		/* How many len16 credits do we need to send this mbuf. */
6051 		next_credits = mbuf_eo_len16(m);
6052 		MPASS(next_credits > 0);
6053 		if (next_credits > cst->tx_credits) {
6054 			/*
6055 			 * Tx will make progress eventually because there is at
6056 			 * least one outstanding fw4_ack that will return
6057 			 * credits and kick the tx.
6058 			 */
6059 			MPASS(cst->ncompl > 0);
6060 			return;
6061 		}
6062 		wr = start_wrq_wr(cst->eo_txq, next_credits, &cookie);
6063 		if (__predict_false(wr == NULL)) {
6064 			/* XXX: wishful thinking, not a real assertion. */
6065 			MPASS(cst->ncompl > 0);
6066 			return;
6067 		}
6068 		cst->tx_credits -= next_credits;
6069 		cst->tx_nocompl += next_credits;
6070 		compl = cst->ncompl == 0 || cst->tx_nocompl >= cst->tx_total / 2;
6071 		ETHER_BPF_MTAP(cst->com.com.ifp, m);
6072 		write_ethofld_wr(cst, wr, m, compl);
6073 		commit_wrq_wr(cst->eo_txq, wr, &cookie);
6074 		if (compl) {
6075 			cst->ncompl++;
6076 			cst->tx_nocompl	= 0;
6077 		}
6078 		(void) mbufq_dequeue(&cst->pending_tx);
6079 
6080 		/*
6081 		 * Drop the mbuf's reference on the tag now rather
6082 		 * than waiting until m_freem().  This ensures that
6083 		 * cxgbe_rate_tag_free gets called when the inp drops
6084 		 * its reference on the tag and there are no more
6085 		 * mbufs in the pending_tx queue and can flush any
6086 		 * pending requests.  Otherwise if the last mbuf
6087 		 * doesn't request a completion the etid will never be
6088 		 * released.
6089 		 */
6090 		m->m_pkthdr.snd_tag = NULL;
6091 		m->m_pkthdr.csum_flags &= ~CSUM_SND_TAG;
6092 		m_snd_tag_rele(&cst->com.com);
6093 
6094 		mbufq_enqueue(&cst->pending_fwack, m);
6095 	}
6096 }
6097 
6098 int
6099 ethofld_transmit(struct ifnet *ifp, struct mbuf *m0)
6100 {
6101 	struct cxgbe_rate_tag *cst;
6102 	int rc;
6103 
6104 	MPASS(m0->m_nextpkt == NULL);
6105 	MPASS(m0->m_pkthdr.csum_flags & CSUM_SND_TAG);
6106 	MPASS(m0->m_pkthdr.snd_tag != NULL);
6107 	cst = mst_to_crt(m0->m_pkthdr.snd_tag);
6108 
6109 	mtx_lock(&cst->lock);
6110 	MPASS(cst->flags & EO_SND_TAG_REF);
6111 
6112 	if (__predict_false(cst->flags & EO_FLOWC_PENDING)) {
6113 		struct vi_info *vi = ifp->if_softc;
6114 		struct port_info *pi = vi->pi;
6115 		struct adapter *sc = pi->adapter;
6116 		const uint32_t rss_mask = vi->rss_size - 1;
6117 		uint32_t rss_hash;
6118 
6119 		cst->eo_txq = &sc->sge.ofld_txq[vi->first_ofld_txq];
6120 		if (M_HASHTYPE_ISHASH(m0))
6121 			rss_hash = m0->m_pkthdr.flowid;
6122 		else
6123 			rss_hash = arc4random();
6124 		/* We assume RSS hashing */
6125 		cst->iqid = vi->rss[rss_hash & rss_mask];
6126 		cst->eo_txq += rss_hash % vi->nofldtxq;
6127 		rc = send_etid_flowc_wr(cst, pi, vi);
6128 		if (rc != 0)
6129 			goto done;
6130 	}
6131 
6132 	if (__predict_false(cst->plen + m0->m_pkthdr.len > eo_max_backlog)) {
6133 		rc = ENOBUFS;
6134 		goto done;
6135 	}
6136 
6137 	mbufq_enqueue(&cst->pending_tx, m0);
6138 	cst->plen += m0->m_pkthdr.len;
6139 
6140 	/*
6141 	 * Hold an extra reference on the tag while generating work
6142 	 * requests to ensure that we don't try to free the tag during
6143 	 * ethofld_tx() in case we are sending the final mbuf after
6144 	 * the inp was freed.
6145 	 */
6146 	m_snd_tag_ref(&cst->com.com);
6147 	ethofld_tx(cst);
6148 	mtx_unlock(&cst->lock);
6149 	m_snd_tag_rele(&cst->com.com);
6150 	return (0);
6151 
6152 done:
6153 	mtx_unlock(&cst->lock);
6154 	if (__predict_false(rc != 0))
6155 		m_freem(m0);
6156 	return (rc);
6157 }
6158 
6159 static int
6160 ethofld_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0)
6161 {
6162 	struct adapter *sc = iq->adapter;
6163 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
6164 	struct mbuf *m;
6165 	u_int etid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
6166 	struct cxgbe_rate_tag *cst;
6167 	uint8_t credits = cpl->credits;
6168 
6169 	cst = lookup_etid(sc, etid);
6170 	mtx_lock(&cst->lock);
6171 	if (__predict_false(cst->flags & EO_FLOWC_RPL_PENDING)) {
6172 		MPASS(credits >= ETID_FLOWC_LEN16);
6173 		credits -= ETID_FLOWC_LEN16;
6174 		cst->flags &= ~EO_FLOWC_RPL_PENDING;
6175 	}
6176 
6177 	KASSERT(cst->ncompl > 0,
6178 	    ("%s: etid %u (%p) wasn't expecting completion.",
6179 	    __func__, etid, cst));
6180 	cst->ncompl--;
6181 
6182 	while (credits > 0) {
6183 		m = mbufq_dequeue(&cst->pending_fwack);
6184 		if (__predict_false(m == NULL)) {
6185 			/*
6186 			 * The remaining credits are for the final flush that
6187 			 * was issued when the tag was freed by the kernel.
6188 			 */
6189 			MPASS((cst->flags &
6190 			    (EO_FLUSH_RPL_PENDING | EO_SND_TAG_REF)) ==
6191 			    EO_FLUSH_RPL_PENDING);
6192 			MPASS(credits == ETID_FLUSH_LEN16);
6193 			MPASS(cst->tx_credits + cpl->credits == cst->tx_total);
6194 			MPASS(cst->ncompl == 0);
6195 
6196 			cst->flags &= ~EO_FLUSH_RPL_PENDING;
6197 			cst->tx_credits += cpl->credits;
6198 			cxgbe_rate_tag_free_locked(cst);
6199 			return (0);	/* cst is gone. */
6200 		}
6201 		KASSERT(m != NULL,
6202 		    ("%s: too many credits (%u, %u)", __func__, cpl->credits,
6203 		    credits));
6204 		KASSERT(credits >= mbuf_eo_len16(m),
6205 		    ("%s: too few credits (%u, %u, %u)", __func__,
6206 		    cpl->credits, credits, mbuf_eo_len16(m)));
6207 		credits -= mbuf_eo_len16(m);
6208 		cst->plen -= m->m_pkthdr.len;
6209 		m_freem(m);
6210 	}
6211 
6212 	cst->tx_credits += cpl->credits;
6213 	MPASS(cst->tx_credits <= cst->tx_total);
6214 
6215 	if (cst->flags & EO_SND_TAG_REF) {
6216 		/*
6217 		 * As with ethofld_transmit(), hold an extra reference
6218 		 * so that the tag is stable across ethold_tx().
6219 		 */
6220 		m_snd_tag_ref(&cst->com.com);
6221 		m = mbufq_first(&cst->pending_tx);
6222 		if (m != NULL && cst->tx_credits >= mbuf_eo_len16(m))
6223 			ethofld_tx(cst);
6224 		mtx_unlock(&cst->lock);
6225 		m_snd_tag_rele(&cst->com.com);
6226 	} else {
6227 		/*
6228 		 * There shouldn't be any pending packets if the tag
6229 		 * was freed by the kernel since any pending packet
6230 		 * should hold a reference to the tag.
6231 		 */
6232 		MPASS(mbufq_first(&cst->pending_tx) == NULL);
6233 		mtx_unlock(&cst->lock);
6234 	}
6235 
6236 	return (0);
6237 }
6238 #endif
6239