xref: /freebsd/sys/dev/cxgb/cxgb_sge.c (revision 7ef62cebc2f965b0f640263e179276928885e33d)
1 /**************************************************************************
2 SPDX-License-Identifier: BSD-2-Clause
3 
4 Copyright (c) 2007-2009, Chelsio Inc.
5 All rights reserved.
6 
7 Redistribution and use in source and binary forms, with or without
8 modification, are permitted provided that the following conditions are met:
9 
10  1. Redistributions of source code must retain the above copyright notice,
11     this list of conditions and the following disclaimer.
12 
13  2. Neither the name of the Chelsio Corporation nor the names of its
14     contributors may be used to endorse or promote products derived from
15     this software without specific prior written permission.
16 
17 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
18 AND 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 COPYRIGHT OWNER OR CONTRIBUTORS BE
21 LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
22 CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
23 SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
24 INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
25 CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
26 ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
27 POSSIBILITY OF SUCH DAMAGE.
28 
29 ***************************************************************************/
30 
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
33 
34 #include "opt_inet6.h"
35 #include "opt_inet.h"
36 
37 #include <sys/param.h>
38 #include <sys/systm.h>
39 #include <sys/kernel.h>
40 #include <sys/module.h>
41 #include <sys/bus.h>
42 #include <sys/conf.h>
43 #include <machine/bus.h>
44 #include <machine/resource.h>
45 #include <sys/rman.h>
46 #include <sys/queue.h>
47 #include <sys/sysctl.h>
48 #include <sys/taskqueue.h>
49 
50 #include <sys/proc.h>
51 #include <sys/sbuf.h>
52 #include <sys/sched.h>
53 #include <sys/smp.h>
54 #include <sys/systm.h>
55 #include <sys/syslog.h>
56 #include <sys/socket.h>
57 #include <sys/sglist.h>
58 
59 #include <net/if.h>
60 #include <net/if_var.h>
61 #include <net/bpf.h>
62 #include <net/ethernet.h>
63 #include <net/if_vlan_var.h>
64 
65 #include <netinet/in_systm.h>
66 #include <netinet/in.h>
67 #include <netinet/ip.h>
68 #include <netinet/ip6.h>
69 #include <netinet/tcp.h>
70 
71 #include <dev/pci/pcireg.h>
72 #include <dev/pci/pcivar.h>
73 
74 #include <vm/vm.h>
75 #include <vm/pmap.h>
76 
77 #include <cxgb_include.h>
78 #include <sys/mvec.h>
79 
80 int	txq_fills = 0;
81 int	multiq_tx_enable = 1;
82 
83 #ifdef TCP_OFFLOAD
84 CTASSERT(NUM_CPL_HANDLERS >= NUM_CPL_CMDS);
85 #endif
86 
87 extern struct sysctl_oid_list sysctl__hw_cxgb_children;
88 int cxgb_txq_buf_ring_size = TX_ETH_Q_SIZE;
89 SYSCTL_INT(_hw_cxgb, OID_AUTO, txq_mr_size, CTLFLAG_RDTUN, &cxgb_txq_buf_ring_size, 0,
90     "size of per-queue mbuf ring");
91 
92 static int cxgb_tx_coalesce_force = 0;
93 SYSCTL_INT(_hw_cxgb, OID_AUTO, tx_coalesce_force, CTLFLAG_RWTUN,
94     &cxgb_tx_coalesce_force, 0,
95     "coalesce small packets into a single work request regardless of ring state");
96 
97 #define	COALESCE_START_DEFAULT		TX_ETH_Q_SIZE>>1
98 #define	COALESCE_START_MAX		(TX_ETH_Q_SIZE-(TX_ETH_Q_SIZE>>3))
99 #define	COALESCE_STOP_DEFAULT		TX_ETH_Q_SIZE>>2
100 #define	COALESCE_STOP_MIN		TX_ETH_Q_SIZE>>5
101 #define	TX_RECLAIM_DEFAULT		TX_ETH_Q_SIZE>>5
102 #define	TX_RECLAIM_MAX			TX_ETH_Q_SIZE>>2
103 #define	TX_RECLAIM_MIN			TX_ETH_Q_SIZE>>6
104 
105 
106 static int cxgb_tx_coalesce_enable_start = COALESCE_START_DEFAULT;
107 SYSCTL_INT(_hw_cxgb, OID_AUTO, tx_coalesce_enable_start, CTLFLAG_RWTUN,
108     &cxgb_tx_coalesce_enable_start, 0,
109     "coalesce enable threshold");
110 static int cxgb_tx_coalesce_enable_stop = COALESCE_STOP_DEFAULT;
111 SYSCTL_INT(_hw_cxgb, OID_AUTO, tx_coalesce_enable_stop, CTLFLAG_RWTUN,
112     &cxgb_tx_coalesce_enable_stop, 0,
113     "coalesce disable threshold");
114 static int cxgb_tx_reclaim_threshold = TX_RECLAIM_DEFAULT;
115 SYSCTL_INT(_hw_cxgb, OID_AUTO, tx_reclaim_threshold, CTLFLAG_RWTUN,
116     &cxgb_tx_reclaim_threshold, 0,
117     "tx cleaning minimum threshold");
118 
119 /*
120  * XXX don't re-enable this until TOE stops assuming
121  * we have an m_ext
122  */
123 static int recycle_enable = 0;
124 
125 extern int cxgb_use_16k_clusters;
126 extern int nmbjumbop;
127 extern int nmbjumbo9;
128 extern int nmbjumbo16;
129 
130 #define USE_GTS 0
131 
132 #define SGE_RX_SM_BUF_SIZE	1536
133 #define SGE_RX_DROP_THRES	16
134 #define SGE_RX_COPY_THRES	128
135 
136 /*
137  * Period of the Tx buffer reclaim timer.  This timer does not need to run
138  * frequently as Tx buffers are usually reclaimed by new Tx packets.
139  */
140 #define TX_RECLAIM_PERIOD       (hz >> 1)
141 
142 /*
143  * Values for sge_txq.flags
144  */
145 enum {
146 	TXQ_RUNNING	= 1 << 0,  /* fetch engine is running */
147 	TXQ_LAST_PKT_DB = 1 << 1,  /* last packet rang the doorbell */
148 };
149 
150 struct tx_desc {
151 	uint64_t	flit[TX_DESC_FLITS];
152 } __packed;
153 
154 struct rx_desc {
155 	uint32_t	addr_lo;
156 	uint32_t	len_gen;
157 	uint32_t	gen2;
158 	uint32_t	addr_hi;
159 } __packed;
160 
161 struct rsp_desc {               /* response queue descriptor */
162 	struct rss_header	rss_hdr;
163 	uint32_t		flags;
164 	uint32_t		len_cq;
165 	uint8_t			imm_data[47];
166 	uint8_t			intr_gen;
167 } __packed;
168 
169 #define RX_SW_DESC_MAP_CREATED	(1 << 0)
170 #define TX_SW_DESC_MAP_CREATED	(1 << 1)
171 #define RX_SW_DESC_INUSE        (1 << 3)
172 #define TX_SW_DESC_MAPPED       (1 << 4)
173 
174 #define RSPQ_NSOP_NEOP           G_RSPD_SOP_EOP(0)
175 #define RSPQ_EOP                 G_RSPD_SOP_EOP(F_RSPD_EOP)
176 #define RSPQ_SOP                 G_RSPD_SOP_EOP(F_RSPD_SOP)
177 #define RSPQ_SOP_EOP             G_RSPD_SOP_EOP(F_RSPD_SOP|F_RSPD_EOP)
178 
179 struct tx_sw_desc {                /* SW state per Tx descriptor */
180 	struct mbuf	*m;
181 	bus_dmamap_t	map;
182 	int		flags;
183 };
184 
185 struct rx_sw_desc {                /* SW state per Rx descriptor */
186 	caddr_t		rxsd_cl;
187 	struct mbuf	*m;
188 	bus_dmamap_t	map;
189 	int		flags;
190 };
191 
192 struct txq_state {
193 	unsigned int	compl;
194 	unsigned int	gen;
195 	unsigned int	pidx;
196 };
197 
198 struct refill_fl_cb_arg {
199 	int               error;
200 	bus_dma_segment_t seg;
201 	int               nseg;
202 };
203 
204 
205 /*
206  * Maps a number of flits to the number of Tx descriptors that can hold them.
207  * The formula is
208  *
209  * desc = 1 + (flits - 2) / (WR_FLITS - 1).
210  *
211  * HW allows up to 4 descriptors to be combined into a WR.
212  */
213 static uint8_t flit_desc_map[] = {
214 	0,
215 #if SGE_NUM_GENBITS == 1
216 	1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
217 	2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
218 	3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
219 	4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4
220 #elif SGE_NUM_GENBITS == 2
221 	1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
222 	2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
223 	3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
224 	4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4,
225 #else
226 # error "SGE_NUM_GENBITS must be 1 or 2"
227 #endif
228 };
229 
230 #define	TXQ_LOCK_ASSERT(qs)	mtx_assert(&(qs)->lock, MA_OWNED)
231 #define	TXQ_TRYLOCK(qs)		mtx_trylock(&(qs)->lock)
232 #define	TXQ_LOCK(qs)		mtx_lock(&(qs)->lock)
233 #define	TXQ_UNLOCK(qs)		mtx_unlock(&(qs)->lock)
234 #define	TXQ_RING_EMPTY(qs)	drbr_empty((qs)->port->ifp, (qs)->txq[TXQ_ETH].txq_mr)
235 #define	TXQ_RING_NEEDS_ENQUEUE(qs)					\
236 	drbr_needs_enqueue((qs)->port->ifp, (qs)->txq[TXQ_ETH].txq_mr)
237 #define	TXQ_RING_FLUSH(qs)	drbr_flush((qs)->port->ifp, (qs)->txq[TXQ_ETH].txq_mr)
238 #define	TXQ_RING_DEQUEUE_COND(qs, func, arg)				\
239 	drbr_dequeue_cond((qs)->port->ifp, (qs)->txq[TXQ_ETH].txq_mr, func, arg)
240 #define	TXQ_RING_DEQUEUE(qs) \
241 	drbr_dequeue((qs)->port->ifp, (qs)->txq[TXQ_ETH].txq_mr)
242 
243 int cxgb_debug = 0;
244 
245 static void sge_timer_cb(void *arg);
246 static void sge_timer_reclaim(void *arg, int ncount);
247 static void sge_txq_reclaim_handler(void *arg, int ncount);
248 static void cxgb_start_locked(struct sge_qset *qs);
249 
250 /*
251  * XXX need to cope with bursty scheduling by looking at a wider
252  * window than we are now for determining the need for coalescing
253  *
254  */
255 static __inline uint64_t
256 check_pkt_coalesce(struct sge_qset *qs)
257 {
258         struct adapter *sc;
259         struct sge_txq *txq;
260 	uint8_t *fill;
261 
262 	if (__predict_false(cxgb_tx_coalesce_force))
263 		return (1);
264 	txq = &qs->txq[TXQ_ETH];
265         sc = qs->port->adapter;
266 	fill = &sc->tunq_fill[qs->idx];
267 
268 	if (cxgb_tx_coalesce_enable_start > COALESCE_START_MAX)
269 		cxgb_tx_coalesce_enable_start = COALESCE_START_MAX;
270 	if (cxgb_tx_coalesce_enable_stop < COALESCE_STOP_MIN)
271 		cxgb_tx_coalesce_enable_start = COALESCE_STOP_MIN;
272 	/*
273 	 * if the hardware transmit queue is more than 1/8 full
274 	 * we mark it as coalescing - we drop back from coalescing
275 	 * when we go below 1/32 full and there are no packets enqueued,
276 	 * this provides us with some degree of hysteresis
277 	 */
278         if (*fill != 0 && (txq->in_use <= cxgb_tx_coalesce_enable_stop) &&
279 	    TXQ_RING_EMPTY(qs) && (qs->coalescing == 0))
280                 *fill = 0;
281         else if (*fill == 0 && (txq->in_use >= cxgb_tx_coalesce_enable_start))
282                 *fill = 1;
283 
284 	return (sc->tunq_coalesce);
285 }
286 
287 #ifdef __LP64__
288 static void
289 set_wr_hdr(struct work_request_hdr *wrp, uint32_t wr_hi, uint32_t wr_lo)
290 {
291 	uint64_t wr_hilo;
292 #if _BYTE_ORDER == _LITTLE_ENDIAN
293 	wr_hilo = wr_hi;
294 	wr_hilo |= (((uint64_t)wr_lo)<<32);
295 #else
296 	wr_hilo = wr_lo;
297 	wr_hilo |= (((uint64_t)wr_hi)<<32);
298 #endif
299 	wrp->wrh_hilo = wr_hilo;
300 }
301 #else
302 static void
303 set_wr_hdr(struct work_request_hdr *wrp, uint32_t wr_hi, uint32_t wr_lo)
304 {
305 
306 	wrp->wrh_hi = wr_hi;
307 	wmb();
308 	wrp->wrh_lo = wr_lo;
309 }
310 #endif
311 
312 struct coalesce_info {
313 	int count;
314 	int nbytes;
315 	int noncoal;
316 };
317 
318 static int
319 coalesce_check(struct mbuf *m, void *arg)
320 {
321 	struct coalesce_info *ci = arg;
322 
323 	if ((m->m_next != NULL) ||
324 	    ((mtod(m, vm_offset_t) & PAGE_MASK) + m->m_len > PAGE_SIZE))
325 		ci->noncoal = 1;
326 
327 	if ((ci->count == 0) || (ci->noncoal == 0 && (ci->count < 7) &&
328 	    (ci->nbytes + m->m_len <= 10500))) {
329 		ci->count++;
330 		ci->nbytes += m->m_len;
331 		return (1);
332 	}
333 	return (0);
334 }
335 
336 static struct mbuf *
337 cxgb_dequeue(struct sge_qset *qs)
338 {
339 	struct mbuf *m, *m_head, *m_tail;
340 	struct coalesce_info ci;
341 
342 
343 	if (check_pkt_coalesce(qs) == 0)
344 		return TXQ_RING_DEQUEUE(qs);
345 
346 	m_head = m_tail = NULL;
347 	ci.count = ci.nbytes = ci.noncoal = 0;
348 	do {
349 		m = TXQ_RING_DEQUEUE_COND(qs, coalesce_check, &ci);
350 		if (m_head == NULL) {
351 			m_tail = m_head = m;
352 		} else if (m != NULL) {
353 			m_tail->m_nextpkt = m;
354 			m_tail = m;
355 		}
356 	} while (m != NULL);
357 	if (ci.count > 7)
358 		panic("trying to coalesce %d packets in to one WR", ci.count);
359 	return (m_head);
360 }
361 
362 /**
363  *	reclaim_completed_tx - reclaims completed Tx descriptors
364  *	@adapter: the adapter
365  *	@q: the Tx queue to reclaim completed descriptors from
366  *
367  *	Reclaims Tx descriptors that the SGE has indicated it has processed,
368  *	and frees the associated buffers if possible.  Called with the Tx
369  *	queue's lock held.
370  */
371 static __inline int
372 reclaim_completed_tx(struct sge_qset *qs, int reclaim_min, int queue)
373 {
374 	struct sge_txq *q = &qs->txq[queue];
375 	int reclaim = desc_reclaimable(q);
376 
377 	if ((cxgb_tx_reclaim_threshold > TX_RECLAIM_MAX) ||
378 	    (cxgb_tx_reclaim_threshold < TX_RECLAIM_MIN))
379 		cxgb_tx_reclaim_threshold = TX_RECLAIM_DEFAULT;
380 
381 	if (reclaim < reclaim_min)
382 		return (0);
383 
384 	mtx_assert(&qs->lock, MA_OWNED);
385 	if (reclaim > 0) {
386 		t3_free_tx_desc(qs, reclaim, queue);
387 		q->cleaned += reclaim;
388 		q->in_use -= reclaim;
389 	}
390 	if (isset(&qs->txq_stopped, TXQ_ETH))
391                 clrbit(&qs->txq_stopped, TXQ_ETH);
392 
393 	return (reclaim);
394 }
395 
396 #ifdef DEBUGNET
397 int
398 cxgb_debugnet_poll_tx(struct sge_qset *qs)
399 {
400 
401 	return (reclaim_completed_tx(qs, TX_RECLAIM_MAX, TXQ_ETH));
402 }
403 #endif
404 
405 /**
406  *	should_restart_tx - are there enough resources to restart a Tx queue?
407  *	@q: the Tx queue
408  *
409  *	Checks if there are enough descriptors to restart a suspended Tx queue.
410  */
411 static __inline int
412 should_restart_tx(const struct sge_txq *q)
413 {
414 	unsigned int r = q->processed - q->cleaned;
415 
416 	return q->in_use - r < (q->size >> 1);
417 }
418 
419 /**
420  *	t3_sge_init - initialize SGE
421  *	@adap: the adapter
422  *	@p: the SGE parameters
423  *
424  *	Performs SGE initialization needed every time after a chip reset.
425  *	We do not initialize any of the queue sets here, instead the driver
426  *	top-level must request those individually.  We also do not enable DMA
427  *	here, that should be done after the queues have been set up.
428  */
429 void
430 t3_sge_init(adapter_t *adap, struct sge_params *p)
431 {
432 	u_int ctrl, ups;
433 
434 	ups = 0; /* = ffs(pci_resource_len(adap->pdev, 2) >> 12); */
435 
436 	ctrl = F_DROPPKT | V_PKTSHIFT(2) | F_FLMODE | F_AVOIDCQOVFL |
437 	       F_CQCRDTCTRL | F_CONGMODE | F_TNLFLMODE | F_FATLPERREN |
438 	       V_HOSTPAGESIZE(PAGE_SHIFT - 11) | F_BIGENDIANINGRESS |
439 	       V_USERSPACESIZE(ups ? ups - 1 : 0) | F_ISCSICOALESCING;
440 #if SGE_NUM_GENBITS == 1
441 	ctrl |= F_EGRGENCTRL;
442 #endif
443 	if (adap->params.rev > 0) {
444 		if (!(adap->flags & (USING_MSIX | USING_MSI)))
445 			ctrl |= F_ONEINTMULTQ | F_OPTONEINTMULTQ;
446 	}
447 	t3_write_reg(adap, A_SG_CONTROL, ctrl);
448 	t3_write_reg(adap, A_SG_EGR_RCQ_DRB_THRSH, V_HIRCQDRBTHRSH(512) |
449 		     V_LORCQDRBTHRSH(512));
450 	t3_write_reg(adap, A_SG_TIMER_TICK, core_ticks_per_usec(adap) / 10);
451 	t3_write_reg(adap, A_SG_CMDQ_CREDIT_TH, V_THRESHOLD(32) |
452 		     V_TIMEOUT(200 * core_ticks_per_usec(adap)));
453 	t3_write_reg(adap, A_SG_HI_DRB_HI_THRSH,
454 		     adap->params.rev < T3_REV_C ? 1000 : 500);
455 	t3_write_reg(adap, A_SG_HI_DRB_LO_THRSH, 256);
456 	t3_write_reg(adap, A_SG_LO_DRB_HI_THRSH, 1000);
457 	t3_write_reg(adap, A_SG_LO_DRB_LO_THRSH, 256);
458 	t3_write_reg(adap, A_SG_OCO_BASE, V_BASE1(0xfff));
459 	t3_write_reg(adap, A_SG_DRB_PRI_THRESH, 63 * 1024);
460 }
461 
462 
463 /**
464  *	sgl_len - calculates the size of an SGL of the given capacity
465  *	@n: the number of SGL entries
466  *
467  *	Calculates the number of flits needed for a scatter/gather list that
468  *	can hold the given number of entries.
469  */
470 static __inline unsigned int
471 sgl_len(unsigned int n)
472 {
473 	return ((3 * n) / 2 + (n & 1));
474 }
475 
476 /**
477  *	get_imm_packet - return the next ingress packet buffer from a response
478  *	@resp: the response descriptor containing the packet data
479  *
480  *	Return a packet containing the immediate data of the given response.
481  */
482 static int
483 get_imm_packet(adapter_t *sc, const struct rsp_desc *resp, struct mbuf *m)
484 {
485 
486 	if (resp->rss_hdr.opcode == CPL_RX_DATA) {
487 		const struct cpl_rx_data *cpl = (const void *)&resp->imm_data[0];
488 		m->m_len = sizeof(*cpl) + ntohs(cpl->len);
489 	} else if (resp->rss_hdr.opcode == CPL_RX_PKT) {
490 		const struct cpl_rx_pkt *cpl = (const void *)&resp->imm_data[0];
491 		m->m_len = sizeof(*cpl) + ntohs(cpl->len);
492 	} else
493 		m->m_len = IMMED_PKT_SIZE;
494 	m->m_ext.ext_buf = NULL;
495 	m->m_ext.ext_type = 0;
496 	memcpy(mtod(m, uint8_t *), resp->imm_data, m->m_len);
497 	return (0);
498 }
499 
500 static __inline u_int
501 flits_to_desc(u_int n)
502 {
503 	return (flit_desc_map[n]);
504 }
505 
506 #define SGE_PARERR (F_CPPARITYERROR | F_OCPARITYERROR | F_RCPARITYERROR | \
507 		    F_IRPARITYERROR | V_ITPARITYERROR(M_ITPARITYERROR) | \
508 		    V_FLPARITYERROR(M_FLPARITYERROR) | F_LODRBPARITYERROR | \
509 		    F_HIDRBPARITYERROR | F_LORCQPARITYERROR | \
510 		    F_HIRCQPARITYERROR)
511 #define SGE_FRAMINGERR (F_UC_REQ_FRAMINGERROR | F_R_REQ_FRAMINGERROR)
512 #define SGE_FATALERR (SGE_PARERR | SGE_FRAMINGERR | F_RSPQCREDITOVERFOW | \
513 		      F_RSPQDISABLED)
514 
515 /**
516  *	t3_sge_err_intr_handler - SGE async event interrupt handler
517  *	@adapter: the adapter
518  *
519  *	Interrupt handler for SGE asynchronous (non-data) events.
520  */
521 void
522 t3_sge_err_intr_handler(adapter_t *adapter)
523 {
524 	unsigned int v, status;
525 
526 	status = t3_read_reg(adapter, A_SG_INT_CAUSE);
527 	if (status & SGE_PARERR)
528 		CH_ALERT(adapter, "SGE parity error (0x%x)\n",
529 			 status & SGE_PARERR);
530 	if (status & SGE_FRAMINGERR)
531 		CH_ALERT(adapter, "SGE framing error (0x%x)\n",
532 			 status & SGE_FRAMINGERR);
533 	if (status & F_RSPQCREDITOVERFOW)
534 		CH_ALERT(adapter, "SGE response queue credit overflow\n");
535 
536 	if (status & F_RSPQDISABLED) {
537 		v = t3_read_reg(adapter, A_SG_RSPQ_FL_STATUS);
538 
539 		CH_ALERT(adapter,
540 			 "packet delivered to disabled response queue (0x%x)\n",
541 			 (v >> S_RSPQ0DISABLED) & 0xff);
542 	}
543 
544 	t3_write_reg(adapter, A_SG_INT_CAUSE, status);
545 	if (status & SGE_FATALERR)
546 		t3_fatal_err(adapter);
547 }
548 
549 void
550 t3_sge_prep(adapter_t *adap, struct sge_params *p)
551 {
552 	int i, nqsets, fl_q_size, jumbo_q_size, use_16k, jumbo_buf_size;
553 
554 	nqsets = min(SGE_QSETS / adap->params.nports, mp_ncpus);
555 	nqsets *= adap->params.nports;
556 
557 	fl_q_size = min(nmbclusters/(3*nqsets), FL_Q_SIZE);
558 
559 	while (!powerof2(fl_q_size))
560 		fl_q_size--;
561 
562 	use_16k = cxgb_use_16k_clusters != -1 ? cxgb_use_16k_clusters :
563 	    is_offload(adap);
564 
565 	if (use_16k) {
566 		jumbo_q_size = min(nmbjumbo16/(3*nqsets), JUMBO_Q_SIZE);
567 		jumbo_buf_size = MJUM16BYTES;
568 	} else {
569 		jumbo_q_size = min(nmbjumbo9/(3*nqsets), JUMBO_Q_SIZE);
570 		jumbo_buf_size = MJUM9BYTES;
571 	}
572 	while (!powerof2(jumbo_q_size))
573 		jumbo_q_size--;
574 
575 	if (fl_q_size < (FL_Q_SIZE / 4) || jumbo_q_size < (JUMBO_Q_SIZE / 2))
576 		device_printf(adap->dev,
577 		    "Insufficient clusters and/or jumbo buffers.\n");
578 
579 	p->max_pkt_size = jumbo_buf_size - sizeof(struct cpl_rx_data);
580 
581 	for (i = 0; i < SGE_QSETS; ++i) {
582 		struct qset_params *q = p->qset + i;
583 
584 		if (adap->params.nports > 2) {
585 			q->coalesce_usecs = 50;
586 		} else {
587 #ifdef INVARIANTS
588 			q->coalesce_usecs = 10;
589 #else
590 			q->coalesce_usecs = 5;
591 #endif
592 		}
593 		q->polling = 0;
594 		q->rspq_size = RSPQ_Q_SIZE;
595 		q->fl_size = fl_q_size;
596 		q->jumbo_size = jumbo_q_size;
597 		q->jumbo_buf_size = jumbo_buf_size;
598 		q->txq_size[TXQ_ETH] = TX_ETH_Q_SIZE;
599 		q->txq_size[TXQ_OFLD] = is_offload(adap) ? TX_OFLD_Q_SIZE : 16;
600 		q->txq_size[TXQ_CTRL] = TX_CTRL_Q_SIZE;
601 		q->cong_thres = 0;
602 	}
603 }
604 
605 int
606 t3_sge_alloc(adapter_t *sc)
607 {
608 
609 	/* The parent tag. */
610 	if (bus_dma_tag_create( bus_get_dma_tag(sc->dev),/* PCI parent */
611 				1, 0,			/* algnmnt, boundary */
612 				BUS_SPACE_MAXADDR,	/* lowaddr */
613 				BUS_SPACE_MAXADDR,	/* highaddr */
614 				NULL, NULL,		/* filter, filterarg */
615 				BUS_SPACE_MAXSIZE_32BIT,/* maxsize */
616 				BUS_SPACE_UNRESTRICTED, /* nsegments */
617 				BUS_SPACE_MAXSIZE_32BIT,/* maxsegsize */
618 				0,			/* flags */
619 				NULL, NULL,		/* lock, lockarg */
620 				&sc->parent_dmat)) {
621 		device_printf(sc->dev, "Cannot allocate parent DMA tag\n");
622 		return (ENOMEM);
623 	}
624 
625 	/*
626 	 * DMA tag for normal sized RX frames
627 	 */
628 	if (bus_dma_tag_create(sc->parent_dmat, MCLBYTES, 0, BUS_SPACE_MAXADDR,
629 		BUS_SPACE_MAXADDR, NULL, NULL, MCLBYTES, 1,
630 		MCLBYTES, BUS_DMA_ALLOCNOW, NULL, NULL, &sc->rx_dmat)) {
631 		device_printf(sc->dev, "Cannot allocate RX DMA tag\n");
632 		return (ENOMEM);
633 	}
634 
635 	/*
636 	 * DMA tag for jumbo sized RX frames.
637 	 */
638 	if (bus_dma_tag_create(sc->parent_dmat, MJUM16BYTES, 0, BUS_SPACE_MAXADDR,
639 		BUS_SPACE_MAXADDR, NULL, NULL, MJUM16BYTES, 1, MJUM16BYTES,
640 		BUS_DMA_ALLOCNOW, NULL, NULL, &sc->rx_jumbo_dmat)) {
641 		device_printf(sc->dev, "Cannot allocate RX jumbo DMA tag\n");
642 		return (ENOMEM);
643 	}
644 
645 	/*
646 	 * DMA tag for TX frames.
647 	 */
648 	if (bus_dma_tag_create(sc->parent_dmat, 1, 0, BUS_SPACE_MAXADDR,
649 		BUS_SPACE_MAXADDR, NULL, NULL, TX_MAX_SIZE, TX_MAX_SEGS,
650 		TX_MAX_SIZE, BUS_DMA_ALLOCNOW,
651 		NULL, NULL, &sc->tx_dmat)) {
652 		device_printf(sc->dev, "Cannot allocate TX DMA tag\n");
653 		return (ENOMEM);
654 	}
655 
656 	return (0);
657 }
658 
659 int
660 t3_sge_free(struct adapter * sc)
661 {
662 
663 	if (sc->tx_dmat != NULL)
664 		bus_dma_tag_destroy(sc->tx_dmat);
665 
666 	if (sc->rx_jumbo_dmat != NULL)
667 		bus_dma_tag_destroy(sc->rx_jumbo_dmat);
668 
669 	if (sc->rx_dmat != NULL)
670 		bus_dma_tag_destroy(sc->rx_dmat);
671 
672 	if (sc->parent_dmat != NULL)
673 		bus_dma_tag_destroy(sc->parent_dmat);
674 
675 	return (0);
676 }
677 
678 void
679 t3_update_qset_coalesce(struct sge_qset *qs, const struct qset_params *p)
680 {
681 
682 	qs->rspq.holdoff_tmr = max(p->coalesce_usecs * 10, 1U);
683 	qs->rspq.polling = 0 /* p->polling */;
684 }
685 
686 #if !defined(__i386__) && !defined(__amd64__)
687 static void
688 refill_fl_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
689 {
690 	struct refill_fl_cb_arg *cb_arg = arg;
691 
692 	cb_arg->error = error;
693 	cb_arg->seg = segs[0];
694 	cb_arg->nseg = nseg;
695 
696 }
697 #endif
698 /**
699  *	refill_fl - refill an SGE free-buffer list
700  *	@sc: the controller softc
701  *	@q: the free-list to refill
702  *	@n: the number of new buffers to allocate
703  *
704  *	(Re)populate an SGE free-buffer list with up to @n new packet buffers.
705  *	The caller must assure that @n does not exceed the queue's capacity.
706  */
707 static void
708 refill_fl(adapter_t *sc, struct sge_fl *q, int n)
709 {
710 	struct rx_sw_desc *sd = &q->sdesc[q->pidx];
711 	struct rx_desc *d = &q->desc[q->pidx];
712 	struct refill_fl_cb_arg cb_arg;
713 	struct mbuf *m;
714 	caddr_t cl;
715 	int err;
716 
717 	cb_arg.error = 0;
718 	while (n--) {
719 		/*
720 		 * We allocate an uninitialized mbuf + cluster, mbuf is
721 		 * initialized after rx.
722 		 */
723 		if (q->zone == zone_pack) {
724 			if ((m = m_getcl(M_NOWAIT, MT_NOINIT, M_PKTHDR)) == NULL)
725 				break;
726 			cl = m->m_ext.ext_buf;
727 		} else {
728 			if ((cl = m_cljget(NULL, M_NOWAIT, q->buf_size)) == NULL)
729 				break;
730 			if ((m = m_gethdr_raw(M_NOWAIT, 0)) == NULL) {
731 				uma_zfree(q->zone, cl);
732 				break;
733 			}
734 		}
735 		if ((sd->flags & RX_SW_DESC_MAP_CREATED) == 0) {
736 			if ((err = bus_dmamap_create(q->entry_tag, 0, &sd->map))) {
737 				log(LOG_WARNING, "bus_dmamap_create failed %d\n", err);
738 				uma_zfree(q->zone, cl);
739 				goto done;
740 			}
741 			sd->flags |= RX_SW_DESC_MAP_CREATED;
742 		}
743 #if !defined(__i386__) && !defined(__amd64__)
744 		err = bus_dmamap_load(q->entry_tag, sd->map,
745 		    cl, q->buf_size, refill_fl_cb, &cb_arg, 0);
746 
747 		if (err != 0 || cb_arg.error) {
748 			if (q->zone != zone_pack)
749 				uma_zfree(q->zone, cl);
750 			m_free(m);
751 			goto done;
752 		}
753 #else
754 		cb_arg.seg.ds_addr = pmap_kextract((vm_offset_t)cl);
755 #endif
756 		sd->flags |= RX_SW_DESC_INUSE;
757 		sd->rxsd_cl = cl;
758 		sd->m = m;
759 		d->addr_lo = htobe32(cb_arg.seg.ds_addr & 0xffffffff);
760 		d->addr_hi = htobe32(((uint64_t)cb_arg.seg.ds_addr >>32) & 0xffffffff);
761 		d->len_gen = htobe32(V_FLD_GEN1(q->gen));
762 		d->gen2 = htobe32(V_FLD_GEN2(q->gen));
763 
764 		d++;
765 		sd++;
766 
767 		if (++q->pidx == q->size) {
768 			q->pidx = 0;
769 			q->gen ^= 1;
770 			sd = q->sdesc;
771 			d = q->desc;
772 		}
773 		q->credits++;
774 		q->db_pending++;
775 	}
776 
777 done:
778 	if (q->db_pending >= 32) {
779 		q->db_pending = 0;
780 		t3_write_reg(sc, A_SG_KDOORBELL, V_EGRCNTX(q->cntxt_id));
781 	}
782 }
783 
784 
785 /**
786  *	free_rx_bufs - free the Rx buffers on an SGE free list
787  *	@sc: the controle softc
788  *	@q: the SGE free list to clean up
789  *
790  *	Release the buffers on an SGE free-buffer Rx queue.  HW fetching from
791  *	this queue should be stopped before calling this function.
792  */
793 static void
794 free_rx_bufs(adapter_t *sc, struct sge_fl *q)
795 {
796 	u_int cidx = q->cidx;
797 
798 	while (q->credits--) {
799 		struct rx_sw_desc *d = &q->sdesc[cidx];
800 
801 		if (d->flags & RX_SW_DESC_INUSE) {
802 			bus_dmamap_unload(q->entry_tag, d->map);
803 			bus_dmamap_destroy(q->entry_tag, d->map);
804 			if (q->zone == zone_pack) {
805 				m_init(d->m, M_NOWAIT, MT_DATA, M_EXT);
806 				uma_zfree(zone_pack, d->m);
807 			} else {
808 				m_init(d->m, M_NOWAIT, MT_DATA, 0);
809 				m_free_raw(d->m);
810 				uma_zfree(q->zone, d->rxsd_cl);
811 			}
812 		}
813 
814 		d->rxsd_cl = NULL;
815 		d->m = NULL;
816 		if (++cidx == q->size)
817 			cidx = 0;
818 	}
819 }
820 
821 static __inline void
822 __refill_fl(adapter_t *adap, struct sge_fl *fl)
823 {
824 	refill_fl(adap, fl, min(16U, fl->size - fl->credits));
825 }
826 
827 static __inline void
828 __refill_fl_lt(adapter_t *adap, struct sge_fl *fl, int max)
829 {
830 	uint32_t reclaimable = fl->size - fl->credits;
831 
832 	if (reclaimable > 0)
833 		refill_fl(adap, fl, min(max, reclaimable));
834 }
835 
836 /**
837  *	recycle_rx_buf - recycle a receive buffer
838  *	@adapter: the adapter
839  *	@q: the SGE free list
840  *	@idx: index of buffer to recycle
841  *
842  *	Recycles the specified buffer on the given free list by adding it at
843  *	the next available slot on the list.
844  */
845 static void
846 recycle_rx_buf(adapter_t *adap, struct sge_fl *q, unsigned int idx)
847 {
848 	struct rx_desc *from = &q->desc[idx];
849 	struct rx_desc *to   = &q->desc[q->pidx];
850 
851 	q->sdesc[q->pidx] = q->sdesc[idx];
852 	to->addr_lo = from->addr_lo;        // already big endian
853 	to->addr_hi = from->addr_hi;        // likewise
854 	wmb();	/* necessary ? */
855 	to->len_gen = htobe32(V_FLD_GEN1(q->gen));
856 	to->gen2 = htobe32(V_FLD_GEN2(q->gen));
857 	q->credits++;
858 
859 	if (++q->pidx == q->size) {
860 		q->pidx = 0;
861 		q->gen ^= 1;
862 	}
863 	t3_write_reg(adap, A_SG_KDOORBELL, V_EGRCNTX(q->cntxt_id));
864 }
865 
866 static void
867 alloc_ring_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
868 {
869 	uint32_t *addr;
870 
871 	addr = arg;
872 	*addr = segs[0].ds_addr;
873 }
874 
875 static int
876 alloc_ring(adapter_t *sc, size_t nelem, size_t elem_size, size_t sw_size,
877     bus_addr_t *phys, void *desc, void *sdesc, bus_dma_tag_t *tag,
878     bus_dmamap_t *map, bus_dma_tag_t parent_entry_tag, bus_dma_tag_t *entry_tag)
879 {
880 	size_t len = nelem * elem_size;
881 	void *s = NULL;
882 	void *p = NULL;
883 	int err;
884 
885 	if ((err = bus_dma_tag_create(sc->parent_dmat, PAGE_SIZE, 0,
886 				      BUS_SPACE_MAXADDR_32BIT,
887 				      BUS_SPACE_MAXADDR, NULL, NULL, len, 1,
888 				      len, 0, NULL, NULL, tag)) != 0) {
889 		device_printf(sc->dev, "Cannot allocate descriptor tag\n");
890 		return (ENOMEM);
891 	}
892 
893 	if ((err = bus_dmamem_alloc(*tag, (void **)&p, BUS_DMA_NOWAIT,
894 				    map)) != 0) {
895 		device_printf(sc->dev, "Cannot allocate descriptor memory\n");
896 		return (ENOMEM);
897 	}
898 
899 	bus_dmamap_load(*tag, *map, p, len, alloc_ring_cb, phys, 0);
900 	bzero(p, len);
901 	*(void **)desc = p;
902 
903 	if (sw_size) {
904 		len = nelem * sw_size;
905 		s = malloc(len, M_DEVBUF, M_WAITOK|M_ZERO);
906 		*(void **)sdesc = s;
907 	}
908 	if (parent_entry_tag == NULL)
909 		return (0);
910 
911 	if ((err = bus_dma_tag_create(parent_entry_tag, 1, 0,
912 				      BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR,
913 		                      NULL, NULL, TX_MAX_SIZE, TX_MAX_SEGS,
914 				      TX_MAX_SIZE, BUS_DMA_ALLOCNOW,
915 		                      NULL, NULL, entry_tag)) != 0) {
916 		device_printf(sc->dev, "Cannot allocate descriptor entry tag\n");
917 		return (ENOMEM);
918 	}
919 	return (0);
920 }
921 
922 static void
923 sge_slow_intr_handler(void *arg, int ncount)
924 {
925 	adapter_t *sc = arg;
926 
927 	t3_slow_intr_handler(sc);
928 	t3_write_reg(sc, A_PL_INT_ENABLE0, sc->slow_intr_mask);
929 	(void) t3_read_reg(sc, A_PL_INT_ENABLE0);
930 }
931 
932 /**
933  *	sge_timer_cb - perform periodic maintenance of an SGE qset
934  *	@data: the SGE queue set to maintain
935  *
936  *	Runs periodically from a timer to perform maintenance of an SGE queue
937  *	set.  It performs two tasks:
938  *
939  *	a) Cleans up any completed Tx descriptors that may still be pending.
940  *	Normal descriptor cleanup happens when new packets are added to a Tx
941  *	queue so this timer is relatively infrequent and does any cleanup only
942  *	if the Tx queue has not seen any new packets in a while.  We make a
943  *	best effort attempt to reclaim descriptors, in that we don't wait
944  *	around if we cannot get a queue's lock (which most likely is because
945  *	someone else is queueing new packets and so will also handle the clean
946  *	up).  Since control queues use immediate data exclusively we don't
947  *	bother cleaning them up here.
948  *
949  *	b) Replenishes Rx queues that have run out due to memory shortage.
950  *	Normally new Rx buffers are added when existing ones are consumed but
951  *	when out of memory a queue can become empty.  We try to add only a few
952  *	buffers here, the queue will be replenished fully as these new buffers
953  *	are used up if memory shortage has subsided.
954  *
955  *	c) Return coalesced response queue credits in case a response queue is
956  *	starved.
957  *
958  *	d) Ring doorbells for T304 tunnel queues since we have seen doorbell
959  *	fifo overflows and the FW doesn't implement any recovery scheme yet.
960  */
961 static void
962 sge_timer_cb(void *arg)
963 {
964 	adapter_t *sc = arg;
965 	if ((sc->flags & USING_MSIX) == 0) {
966 
967 		struct port_info *pi;
968 		struct sge_qset *qs;
969 		struct sge_txq  *txq;
970 		int i, j;
971 		int reclaim_ofl, refill_rx;
972 
973 		if (sc->open_device_map == 0)
974 			return;
975 
976 		for (i = 0; i < sc->params.nports; i++) {
977 			pi = &sc->port[i];
978 			for (j = 0; j < pi->nqsets; j++) {
979 				qs = &sc->sge.qs[pi->first_qset + j];
980 				txq = &qs->txq[0];
981 				reclaim_ofl = txq[TXQ_OFLD].processed - txq[TXQ_OFLD].cleaned;
982 				refill_rx = ((qs->fl[0].credits < qs->fl[0].size) ||
983 				    (qs->fl[1].credits < qs->fl[1].size));
984 				if (reclaim_ofl || refill_rx) {
985 					taskqueue_enqueue(sc->tq, &pi->timer_reclaim_task);
986 					break;
987 				}
988 			}
989 		}
990 	}
991 
992 	if (sc->params.nports > 2) {
993 		int i;
994 
995 		for_each_port(sc, i) {
996 			struct port_info *pi = &sc->port[i];
997 
998 			t3_write_reg(sc, A_SG_KDOORBELL,
999 				     F_SELEGRCNTX |
1000 				     (FW_TUNNEL_SGEEC_START + pi->first_qset));
1001 		}
1002 	}
1003 	if (((sc->flags & USING_MSIX) == 0 || sc->params.nports > 2) &&
1004 	    sc->open_device_map != 0)
1005 		callout_reset(&sc->sge_timer_ch, TX_RECLAIM_PERIOD, sge_timer_cb, sc);
1006 }
1007 
1008 /*
1009  * This is meant to be a catch-all function to keep sge state private
1010  * to sge.c
1011  *
1012  */
1013 int
1014 t3_sge_init_adapter(adapter_t *sc)
1015 {
1016 	callout_init(&sc->sge_timer_ch, 1);
1017 	callout_reset(&sc->sge_timer_ch, TX_RECLAIM_PERIOD, sge_timer_cb, sc);
1018 	TASK_INIT(&sc->slow_intr_task, 0, sge_slow_intr_handler, sc);
1019 	return (0);
1020 }
1021 
1022 int
1023 t3_sge_reset_adapter(adapter_t *sc)
1024 {
1025 	callout_reset(&sc->sge_timer_ch, TX_RECLAIM_PERIOD, sge_timer_cb, sc);
1026 	return (0);
1027 }
1028 
1029 int
1030 t3_sge_init_port(struct port_info *pi)
1031 {
1032 	TASK_INIT(&pi->timer_reclaim_task, 0, sge_timer_reclaim, pi);
1033 	return (0);
1034 }
1035 
1036 /**
1037  *	refill_rspq - replenish an SGE response queue
1038  *	@adapter: the adapter
1039  *	@q: the response queue to replenish
1040  *	@credits: how many new responses to make available
1041  *
1042  *	Replenishes a response queue by making the supplied number of responses
1043  *	available to HW.
1044  */
1045 static __inline void
1046 refill_rspq(adapter_t *sc, const struct sge_rspq *q, u_int credits)
1047 {
1048 
1049 	/* mbufs are allocated on demand when a rspq entry is processed. */
1050 	t3_write_reg(sc, A_SG_RSPQ_CREDIT_RETURN,
1051 		     V_RSPQ(q->cntxt_id) | V_CREDITS(credits));
1052 }
1053 
1054 static void
1055 sge_txq_reclaim_handler(void *arg, int ncount)
1056 {
1057 	struct sge_qset *qs = arg;
1058 	int i;
1059 
1060 	for (i = 0; i < 3; i++)
1061 		reclaim_completed_tx(qs, 16, i);
1062 }
1063 
1064 static void
1065 sge_timer_reclaim(void *arg, int ncount)
1066 {
1067 	struct port_info *pi = arg;
1068 	int i, nqsets = pi->nqsets;
1069 	adapter_t *sc = pi->adapter;
1070 	struct sge_qset *qs;
1071 	struct mtx *lock;
1072 
1073 	KASSERT((sc->flags & USING_MSIX) == 0,
1074 	    ("can't call timer reclaim for msi-x"));
1075 
1076 	for (i = 0; i < nqsets; i++) {
1077 		qs = &sc->sge.qs[pi->first_qset + i];
1078 
1079 		reclaim_completed_tx(qs, 16, TXQ_OFLD);
1080 		lock = (sc->flags & USING_MSIX) ? &qs->rspq.lock :
1081 			    &sc->sge.qs[0].rspq.lock;
1082 
1083 		if (mtx_trylock(lock)) {
1084 			/* XXX currently assume that we are *NOT* polling */
1085 			uint32_t status = t3_read_reg(sc, A_SG_RSPQ_FL_STATUS);
1086 
1087 			if (qs->fl[0].credits < qs->fl[0].size - 16)
1088 				__refill_fl(sc, &qs->fl[0]);
1089 			if (qs->fl[1].credits < qs->fl[1].size - 16)
1090 				__refill_fl(sc, &qs->fl[1]);
1091 
1092 			if (status & (1 << qs->rspq.cntxt_id)) {
1093 				if (qs->rspq.credits) {
1094 					refill_rspq(sc, &qs->rspq, 1);
1095 					qs->rspq.credits--;
1096 					t3_write_reg(sc, A_SG_RSPQ_FL_STATUS,
1097 					    1 << qs->rspq.cntxt_id);
1098 				}
1099 			}
1100 			mtx_unlock(lock);
1101 		}
1102 	}
1103 }
1104 
1105 /**
1106  *	init_qset_cntxt - initialize an SGE queue set context info
1107  *	@qs: the queue set
1108  *	@id: the queue set id
1109  *
1110  *	Initializes the TIDs and context ids for the queues of a queue set.
1111  */
1112 static void
1113 init_qset_cntxt(struct sge_qset *qs, u_int id)
1114 {
1115 
1116 	qs->rspq.cntxt_id = id;
1117 	qs->fl[0].cntxt_id = 2 * id;
1118 	qs->fl[1].cntxt_id = 2 * id + 1;
1119 	qs->txq[TXQ_ETH].cntxt_id = FW_TUNNEL_SGEEC_START + id;
1120 	qs->txq[TXQ_ETH].token = FW_TUNNEL_TID_START + id;
1121 	qs->txq[TXQ_OFLD].cntxt_id = FW_OFLD_SGEEC_START + id;
1122 	qs->txq[TXQ_CTRL].cntxt_id = FW_CTRL_SGEEC_START + id;
1123 	qs->txq[TXQ_CTRL].token = FW_CTRL_TID_START + id;
1124 
1125 	/* XXX: a sane limit is needed instead of INT_MAX */
1126 	mbufq_init(&qs->txq[TXQ_ETH].sendq, INT_MAX);
1127 	mbufq_init(&qs->txq[TXQ_OFLD].sendq, INT_MAX);
1128 	mbufq_init(&qs->txq[TXQ_CTRL].sendq, INT_MAX);
1129 }
1130 
1131 
1132 static void
1133 txq_prod(struct sge_txq *txq, unsigned int ndesc, struct txq_state *txqs)
1134 {
1135 	txq->in_use += ndesc;
1136 	/*
1137 	 * XXX we don't handle stopping of queue
1138 	 * presumably start handles this when we bump against the end
1139 	 */
1140 	txqs->gen = txq->gen;
1141 	txq->unacked += ndesc;
1142 	txqs->compl = (txq->unacked & 32) << (S_WR_COMPL - 5);
1143 	txq->unacked &= 31;
1144 	txqs->pidx = txq->pidx;
1145 	txq->pidx += ndesc;
1146 #ifdef INVARIANTS
1147 	if (((txqs->pidx > txq->cidx) &&
1148 		(txq->pidx < txqs->pidx) &&
1149 		(txq->pidx >= txq->cidx)) ||
1150 	    ((txqs->pidx < txq->cidx) &&
1151 		(txq->pidx >= txq-> cidx)) ||
1152 	    ((txqs->pidx < txq->cidx) &&
1153 		(txq->cidx < txqs->pidx)))
1154 		panic("txqs->pidx=%d txq->pidx=%d txq->cidx=%d",
1155 		    txqs->pidx, txq->pidx, txq->cidx);
1156 #endif
1157 	if (txq->pidx >= txq->size) {
1158 		txq->pidx -= txq->size;
1159 		txq->gen ^= 1;
1160 	}
1161 
1162 }
1163 
1164 /**
1165  *	calc_tx_descs - calculate the number of Tx descriptors for a packet
1166  *	@m: the packet mbufs
1167  *      @nsegs: the number of segments
1168  *
1169  * 	Returns the number of Tx descriptors needed for the given Ethernet
1170  * 	packet.  Ethernet packets require addition of WR and CPL headers.
1171  */
1172 static __inline unsigned int
1173 calc_tx_descs(const struct mbuf *m, int nsegs)
1174 {
1175 	unsigned int flits;
1176 
1177 	if (m->m_pkthdr.len <= PIO_LEN)
1178 		return 1;
1179 
1180 	flits = sgl_len(nsegs) + 2;
1181 	if (m->m_pkthdr.csum_flags & CSUM_TSO)
1182 		flits++;
1183 
1184 	return flits_to_desc(flits);
1185 }
1186 
1187 /**
1188  *	make_sgl - populate a scatter/gather list for a packet
1189  *	@sgp: the SGL to populate
1190  *	@segs: the packet dma segments
1191  *	@nsegs: the number of segments
1192  *
1193  *	Generates a scatter/gather list for the buffers that make up a packet
1194  *	and returns the SGL size in 8-byte words.  The caller must size the SGL
1195  *	appropriately.
1196  */
1197 static __inline void
1198 make_sgl(struct sg_ent *sgp, bus_dma_segment_t *segs, int nsegs)
1199 {
1200 	int i, idx;
1201 
1202 	for (idx = 0, i = 0; i < nsegs; i++) {
1203 		/*
1204 		 * firmware doesn't like empty segments
1205 		 */
1206 		if (segs[i].ds_len == 0)
1207 			continue;
1208 		if (i && idx == 0)
1209 			++sgp;
1210 
1211 		sgp->len[idx] = htobe32(segs[i].ds_len);
1212 		sgp->addr[idx] = htobe64(segs[i].ds_addr);
1213 		idx ^= 1;
1214 	}
1215 
1216 	if (idx) {
1217 		sgp->len[idx] = 0;
1218 		sgp->addr[idx] = 0;
1219 	}
1220 }
1221 
1222 /**
1223  *	check_ring_tx_db - check and potentially ring a Tx queue's doorbell
1224  *	@adap: the adapter
1225  *	@q: the Tx queue
1226  *
1227  *	Ring the doorbell if a Tx queue is asleep.  There is a natural race,
1228  *	where the HW is going to sleep just after we checked, however,
1229  *	then the interrupt handler will detect the outstanding TX packet
1230  *	and ring the doorbell for us.
1231  *
1232  *	When GTS is disabled we unconditionally ring the doorbell.
1233  */
1234 static __inline void
1235 check_ring_tx_db(adapter_t *adap, struct sge_txq *q, int mustring)
1236 {
1237 #if USE_GTS
1238 	clear_bit(TXQ_LAST_PKT_DB, &q->flags);
1239 	if (test_and_set_bit(TXQ_RUNNING, &q->flags) == 0) {
1240 		set_bit(TXQ_LAST_PKT_DB, &q->flags);
1241 #ifdef T3_TRACE
1242 		T3_TRACE1(adap->tb[q->cntxt_id & 7], "doorbell Tx, cntxt %d",
1243 			  q->cntxt_id);
1244 #endif
1245 		t3_write_reg(adap, A_SG_KDOORBELL,
1246 			     F_SELEGRCNTX | V_EGRCNTX(q->cntxt_id));
1247 	}
1248 #else
1249 	if (mustring || ++q->db_pending >= 32) {
1250 		wmb();            /* write descriptors before telling HW */
1251 		t3_write_reg(adap, A_SG_KDOORBELL,
1252 		    F_SELEGRCNTX | V_EGRCNTX(q->cntxt_id));
1253 		q->db_pending = 0;
1254 	}
1255 #endif
1256 }
1257 
1258 static __inline void
1259 wr_gen2(struct tx_desc *d, unsigned int gen)
1260 {
1261 #if SGE_NUM_GENBITS == 2
1262 	d->flit[TX_DESC_FLITS - 1] = htobe64(gen);
1263 #endif
1264 }
1265 
1266 /**
1267  *	write_wr_hdr_sgl - write a WR header and, optionally, SGL
1268  *	@ndesc: number of Tx descriptors spanned by the SGL
1269  *	@txd: first Tx descriptor to be written
1270  *	@txqs: txq state (generation and producer index)
1271  *	@txq: the SGE Tx queue
1272  *	@sgl: the SGL
1273  *	@flits: number of flits to the start of the SGL in the first descriptor
1274  *	@sgl_flits: the SGL size in flits
1275  *	@wr_hi: top 32 bits of WR header based on WR type (big endian)
1276  *	@wr_lo: low 32 bits of WR header based on WR type (big endian)
1277  *
1278  *	Write a work request header and an associated SGL.  If the SGL is
1279  *	small enough to fit into one Tx descriptor it has already been written
1280  *	and we just need to write the WR header.  Otherwise we distribute the
1281  *	SGL across the number of descriptors it spans.
1282  */
1283 static void
1284 write_wr_hdr_sgl(unsigned int ndesc, struct tx_desc *txd, struct txq_state *txqs,
1285     const struct sge_txq *txq, const struct sg_ent *sgl, unsigned int flits,
1286     unsigned int sgl_flits, unsigned int wr_hi, unsigned int wr_lo)
1287 {
1288 
1289 	struct work_request_hdr *wrp = (struct work_request_hdr *)txd;
1290 
1291 	if (__predict_true(ndesc == 1)) {
1292 		set_wr_hdr(wrp, htonl(F_WR_SOP | F_WR_EOP | V_WR_DATATYPE(1) |
1293 		    V_WR_SGLSFLT(flits)) | wr_hi,
1294 		    htonl(V_WR_LEN(flits + sgl_flits) | V_WR_GEN(txqs->gen)) |
1295 		    wr_lo);
1296 
1297 		wr_gen2(txd, txqs->gen);
1298 
1299 	} else {
1300 		unsigned int ogen = txqs->gen;
1301 		const uint64_t *fp = (const uint64_t *)sgl;
1302 		struct work_request_hdr *wp = wrp;
1303 
1304 		wrp->wrh_hi = htonl(F_WR_SOP | V_WR_DATATYPE(1) |
1305 		    V_WR_SGLSFLT(flits)) | wr_hi;
1306 
1307 		while (sgl_flits) {
1308 			unsigned int avail = WR_FLITS - flits;
1309 
1310 			if (avail > sgl_flits)
1311 				avail = sgl_flits;
1312 			memcpy(&txd->flit[flits], fp, avail * sizeof(*fp));
1313 			sgl_flits -= avail;
1314 			ndesc--;
1315 			if (!sgl_flits)
1316 				break;
1317 
1318 			fp += avail;
1319 			txd++;
1320 			if (++txqs->pidx == txq->size) {
1321 				txqs->pidx = 0;
1322 				txqs->gen ^= 1;
1323 				txd = txq->desc;
1324 			}
1325 
1326 			/*
1327 			 * when the head of the mbuf chain
1328 			 * is freed all clusters will be freed
1329 			 * with it
1330 			 */
1331 			wrp = (struct work_request_hdr *)txd;
1332 			wrp->wrh_hi = htonl(V_WR_DATATYPE(1) |
1333 			    V_WR_SGLSFLT(1)) | wr_hi;
1334 			wrp->wrh_lo = htonl(V_WR_LEN(min(WR_FLITS,
1335 				    sgl_flits + 1)) |
1336 			    V_WR_GEN(txqs->gen)) | wr_lo;
1337 			wr_gen2(txd, txqs->gen);
1338 			flits = 1;
1339 		}
1340 		wrp->wrh_hi |= htonl(F_WR_EOP);
1341 		wmb();
1342 		wp->wrh_lo = htonl(V_WR_LEN(WR_FLITS) | V_WR_GEN(ogen)) | wr_lo;
1343 		wr_gen2((struct tx_desc *)wp, ogen);
1344 	}
1345 }
1346 
1347 /* sizeof(*eh) + sizeof(*ip) + sizeof(*tcp) */
1348 #define TCPPKTHDRSIZE (ETHER_HDR_LEN + 20 + 20)
1349 
1350 #define GET_VTAG(cntrl, m) \
1351 do { \
1352 	if ((m)->m_flags & M_VLANTAG)					            \
1353 		cntrl |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN((m)->m_pkthdr.ether_vtag); \
1354 } while (0)
1355 
1356 static int
1357 t3_encap(struct sge_qset *qs, struct mbuf **m)
1358 {
1359 	adapter_t *sc;
1360 	struct mbuf *m0;
1361 	struct sge_txq *txq;
1362 	struct txq_state txqs;
1363 	struct port_info *pi;
1364 	unsigned int ndesc, flits, cntrl, mlen;
1365 	int err, nsegs, tso_info = 0;
1366 
1367 	struct work_request_hdr *wrp;
1368 	struct tx_sw_desc *txsd;
1369 	struct sg_ent *sgp, *sgl;
1370 	uint32_t wr_hi, wr_lo, sgl_flits;
1371 	bus_dma_segment_t segs[TX_MAX_SEGS];
1372 
1373 	struct tx_desc *txd;
1374 
1375 	pi = qs->port;
1376 	sc = pi->adapter;
1377 	txq = &qs->txq[TXQ_ETH];
1378 	txd = &txq->desc[txq->pidx];
1379 	txsd = &txq->sdesc[txq->pidx];
1380 	sgl = txq->txq_sgl;
1381 
1382 	prefetch(txd);
1383 	m0 = *m;
1384 
1385 	mtx_assert(&qs->lock, MA_OWNED);
1386 	cntrl = V_TXPKT_INTF(pi->txpkt_intf);
1387 	KASSERT(m0->m_flags & M_PKTHDR, ("not packet header\n"));
1388 
1389 	if  (m0->m_nextpkt == NULL && m0->m_next != NULL &&
1390 	    m0->m_pkthdr.csum_flags & (CSUM_TSO))
1391 		tso_info = V_LSO_MSS(m0->m_pkthdr.tso_segsz);
1392 
1393 	if (m0->m_nextpkt != NULL) {
1394 		busdma_map_sg_vec(txq->entry_tag, txsd->map, m0, segs, &nsegs);
1395 		ndesc = 1;
1396 		mlen = 0;
1397 	} else {
1398 		if ((err = busdma_map_sg_collapse(txq->entry_tag, txsd->map,
1399 		    &m0, segs, &nsegs))) {
1400 			if (cxgb_debug)
1401 				printf("failed ... err=%d\n", err);
1402 			return (err);
1403 		}
1404 		mlen = m0->m_pkthdr.len;
1405 		ndesc = calc_tx_descs(m0, nsegs);
1406 	}
1407 	txq_prod(txq, ndesc, &txqs);
1408 
1409 	KASSERT(m0->m_pkthdr.len, ("empty packet nsegs=%d", nsegs));
1410 	txsd->m = m0;
1411 
1412 	if (m0->m_nextpkt != NULL) {
1413 		struct cpl_tx_pkt_batch *cpl_batch = (struct cpl_tx_pkt_batch *)txd;
1414 		int i, fidx;
1415 
1416 		if (nsegs > 7)
1417 			panic("trying to coalesce %d packets in to one WR", nsegs);
1418 		txq->txq_coalesced += nsegs;
1419 		wrp = (struct work_request_hdr *)txd;
1420 		flits = nsegs*2 + 1;
1421 
1422 		for (fidx = 1, i = 0; i < nsegs; i++, fidx += 2) {
1423 			struct cpl_tx_pkt_batch_entry *cbe;
1424 			uint64_t flit;
1425 			uint32_t *hflit = (uint32_t *)&flit;
1426 			int cflags = m0->m_pkthdr.csum_flags;
1427 
1428 			cntrl = V_TXPKT_INTF(pi->txpkt_intf);
1429 			GET_VTAG(cntrl, m0);
1430 			cntrl |= V_TXPKT_OPCODE(CPL_TX_PKT);
1431 			if (__predict_false(!(cflags & CSUM_IP)))
1432 				cntrl |= F_TXPKT_IPCSUM_DIS;
1433 			if (__predict_false(!(cflags & (CSUM_TCP | CSUM_UDP |
1434 			    CSUM_UDP_IPV6 | CSUM_TCP_IPV6))))
1435 				cntrl |= F_TXPKT_L4CSUM_DIS;
1436 
1437 			hflit[0] = htonl(cntrl);
1438 			hflit[1] = htonl(segs[i].ds_len | 0x80000000);
1439 			flit |= htobe64(1 << 24);
1440 			cbe = &cpl_batch->pkt_entry[i];
1441 			cbe->cntrl = hflit[0];
1442 			cbe->len = hflit[1];
1443 			cbe->addr = htobe64(segs[i].ds_addr);
1444 		}
1445 
1446 		wr_hi = htonl(F_WR_SOP | F_WR_EOP | V_WR_DATATYPE(1) |
1447 		    V_WR_SGLSFLT(flits)) |
1448 		    htonl(V_WR_OP(FW_WROPCODE_TUNNEL_TX_PKT) | txqs.compl);
1449 		wr_lo = htonl(V_WR_LEN(flits) |
1450 		    V_WR_GEN(txqs.gen)) | htonl(V_WR_TID(txq->token));
1451 		set_wr_hdr(wrp, wr_hi, wr_lo);
1452 		wmb();
1453 		ETHER_BPF_MTAP(pi->ifp, m0);
1454 		wr_gen2(txd, txqs.gen);
1455 		check_ring_tx_db(sc, txq, 0);
1456 		return (0);
1457 	} else if (tso_info) {
1458 		uint16_t eth_type;
1459 		struct cpl_tx_pkt_lso *hdr = (struct cpl_tx_pkt_lso *)txd;
1460 		struct ether_header *eh;
1461 		void *l3hdr;
1462 		struct tcphdr *tcp;
1463 
1464 		txd->flit[2] = 0;
1465 		GET_VTAG(cntrl, m0);
1466 		cntrl |= V_TXPKT_OPCODE(CPL_TX_PKT_LSO);
1467 		hdr->cntrl = htonl(cntrl);
1468 		hdr->len = htonl(mlen | 0x80000000);
1469 
1470 		if (__predict_false(mlen < TCPPKTHDRSIZE)) {
1471 			printf("mbuf=%p,len=%d,tso_segsz=%d,csum_flags=%b,flags=%#x",
1472 			    m0, mlen, m0->m_pkthdr.tso_segsz,
1473 			    (int)m0->m_pkthdr.csum_flags, CSUM_BITS, m0->m_flags);
1474 			panic("tx tso packet too small");
1475 		}
1476 
1477 		/* Make sure that ether, ip, tcp headers are all in m0 */
1478 		if (__predict_false(m0->m_len < TCPPKTHDRSIZE)) {
1479 			m0 = m_pullup(m0, TCPPKTHDRSIZE);
1480 			if (__predict_false(m0 == NULL)) {
1481 				/* XXX panic probably an overreaction */
1482 				panic("couldn't fit header into mbuf");
1483 			}
1484 		}
1485 
1486 		eh = mtod(m0, struct ether_header *);
1487 		eth_type = eh->ether_type;
1488 		if (eth_type == htons(ETHERTYPE_VLAN)) {
1489 			struct ether_vlan_header *evh = (void *)eh;
1490 
1491 			tso_info |= V_LSO_ETH_TYPE(CPL_ETH_II_VLAN);
1492 			l3hdr = evh + 1;
1493 			eth_type = evh->evl_proto;
1494 		} else {
1495 			tso_info |= V_LSO_ETH_TYPE(CPL_ETH_II);
1496 			l3hdr = eh + 1;
1497 		}
1498 
1499 		if (eth_type == htons(ETHERTYPE_IP)) {
1500 			struct ip *ip = l3hdr;
1501 
1502 			tso_info |= V_LSO_IPHDR_WORDS(ip->ip_hl);
1503 			tcp = (struct tcphdr *)(ip + 1);
1504 		} else if (eth_type == htons(ETHERTYPE_IPV6)) {
1505 			struct ip6_hdr *ip6 = l3hdr;
1506 
1507 			KASSERT(ip6->ip6_nxt == IPPROTO_TCP,
1508 			    ("%s: CSUM_TSO with ip6_nxt %d",
1509 			    __func__, ip6->ip6_nxt));
1510 
1511 			tso_info |= F_LSO_IPV6;
1512 			tso_info |= V_LSO_IPHDR_WORDS(sizeof(*ip6) >> 2);
1513 			tcp = (struct tcphdr *)(ip6 + 1);
1514 		} else
1515 			panic("%s: CSUM_TSO but neither ip nor ip6", __func__);
1516 
1517 		tso_info |= V_LSO_TCPHDR_WORDS(tcp->th_off);
1518 		hdr->lso_info = htonl(tso_info);
1519 
1520 		if (__predict_false(mlen <= PIO_LEN)) {
1521 			/*
1522 			 * pkt not undersized but fits in PIO_LEN
1523 			 * Indicates a TSO bug at the higher levels.
1524 			 */
1525 			txsd->m = NULL;
1526 			m_copydata(m0, 0, mlen, (caddr_t)&txd->flit[3]);
1527 			flits = (mlen + 7) / 8 + 3;
1528 			wr_hi = htonl(V_WR_BCNTLFLT(mlen & 7) |
1529 					  V_WR_OP(FW_WROPCODE_TUNNEL_TX_PKT) |
1530 					  F_WR_SOP | F_WR_EOP | txqs.compl);
1531 			wr_lo = htonl(V_WR_LEN(flits) |
1532 			    V_WR_GEN(txqs.gen) | V_WR_TID(txq->token));
1533 			set_wr_hdr(&hdr->wr, wr_hi, wr_lo);
1534 			wmb();
1535 			ETHER_BPF_MTAP(pi->ifp, m0);
1536 			wr_gen2(txd, txqs.gen);
1537 			check_ring_tx_db(sc, txq, 0);
1538 			m_freem(m0);
1539 			return (0);
1540 		}
1541 		flits = 3;
1542 	} else {
1543 		struct cpl_tx_pkt *cpl = (struct cpl_tx_pkt *)txd;
1544 
1545 		GET_VTAG(cntrl, m0);
1546 		cntrl |= V_TXPKT_OPCODE(CPL_TX_PKT);
1547 		if (__predict_false(!(m0->m_pkthdr.csum_flags & CSUM_IP)))
1548 			cntrl |= F_TXPKT_IPCSUM_DIS;
1549 		if (__predict_false(!(m0->m_pkthdr.csum_flags & (CSUM_TCP |
1550 		    CSUM_UDP | CSUM_UDP_IPV6 | CSUM_TCP_IPV6))))
1551 			cntrl |= F_TXPKT_L4CSUM_DIS;
1552 		cpl->cntrl = htonl(cntrl);
1553 		cpl->len = htonl(mlen | 0x80000000);
1554 
1555 		if (mlen <= PIO_LEN) {
1556 			txsd->m = NULL;
1557 			m_copydata(m0, 0, mlen, (caddr_t)&txd->flit[2]);
1558 			flits = (mlen + 7) / 8 + 2;
1559 
1560 			wr_hi = htonl(V_WR_BCNTLFLT(mlen & 7) |
1561 			    V_WR_OP(FW_WROPCODE_TUNNEL_TX_PKT) |
1562 					  F_WR_SOP | F_WR_EOP | txqs.compl);
1563 			wr_lo = htonl(V_WR_LEN(flits) |
1564 			    V_WR_GEN(txqs.gen) | V_WR_TID(txq->token));
1565 			set_wr_hdr(&cpl->wr, wr_hi, wr_lo);
1566 			wmb();
1567 			ETHER_BPF_MTAP(pi->ifp, m0);
1568 			wr_gen2(txd, txqs.gen);
1569 			check_ring_tx_db(sc, txq, 0);
1570 			m_freem(m0);
1571 			return (0);
1572 		}
1573 		flits = 2;
1574 	}
1575 	wrp = (struct work_request_hdr *)txd;
1576 	sgp = (ndesc == 1) ? (struct sg_ent *)&txd->flit[flits] : sgl;
1577 	make_sgl(sgp, segs, nsegs);
1578 
1579 	sgl_flits = sgl_len(nsegs);
1580 
1581 	ETHER_BPF_MTAP(pi->ifp, m0);
1582 
1583 	KASSERT(ndesc <= 4, ("ndesc too large %d", ndesc));
1584 	wr_hi = htonl(V_WR_OP(FW_WROPCODE_TUNNEL_TX_PKT) | txqs.compl);
1585 	wr_lo = htonl(V_WR_TID(txq->token));
1586 	write_wr_hdr_sgl(ndesc, txd, &txqs, txq, sgl, flits,
1587 	    sgl_flits, wr_hi, wr_lo);
1588 	check_ring_tx_db(sc, txq, 0);
1589 
1590 	return (0);
1591 }
1592 
1593 #ifdef DEBUGNET
1594 int
1595 cxgb_debugnet_encap(struct sge_qset *qs, struct mbuf **m)
1596 {
1597 	int error;
1598 
1599 	error = t3_encap(qs, m);
1600 	if (error == 0)
1601 		check_ring_tx_db(qs->port->adapter, &qs->txq[TXQ_ETH], 1);
1602 	else if (*m != NULL) {
1603 		m_freem(*m);
1604 		*m = NULL;
1605 	}
1606 	return (error);
1607 }
1608 #endif
1609 
1610 void
1611 cxgb_tx_watchdog(void *arg)
1612 {
1613 	struct sge_qset *qs = arg;
1614 	struct sge_txq *txq = &qs->txq[TXQ_ETH];
1615 
1616         if (qs->coalescing != 0 &&
1617 	    (txq->in_use <= cxgb_tx_coalesce_enable_stop) &&
1618 	    TXQ_RING_EMPTY(qs))
1619                 qs->coalescing = 0;
1620         else if (qs->coalescing == 0 &&
1621 	    (txq->in_use >= cxgb_tx_coalesce_enable_start))
1622                 qs->coalescing = 1;
1623 	if (TXQ_TRYLOCK(qs)) {
1624 		qs->qs_flags |= QS_FLUSHING;
1625 		cxgb_start_locked(qs);
1626 		qs->qs_flags &= ~QS_FLUSHING;
1627 		TXQ_UNLOCK(qs);
1628 	}
1629 	if (if_getdrvflags(qs->port->ifp) & IFF_DRV_RUNNING)
1630 		callout_reset_on(&txq->txq_watchdog, hz/4, cxgb_tx_watchdog,
1631 		    qs, txq->txq_watchdog.c_cpu);
1632 }
1633 
1634 static void
1635 cxgb_tx_timeout(void *arg)
1636 {
1637 	struct sge_qset *qs = arg;
1638 	struct sge_txq *txq = &qs->txq[TXQ_ETH];
1639 
1640 	if (qs->coalescing == 0 && (txq->in_use >= (txq->size>>3)))
1641                 qs->coalescing = 1;
1642 	if (TXQ_TRYLOCK(qs)) {
1643 		qs->qs_flags |= QS_TIMEOUT;
1644 		cxgb_start_locked(qs);
1645 		qs->qs_flags &= ~QS_TIMEOUT;
1646 		TXQ_UNLOCK(qs);
1647 	}
1648 }
1649 
1650 static void
1651 cxgb_start_locked(struct sge_qset *qs)
1652 {
1653 	struct mbuf *m_head = NULL;
1654 	struct sge_txq *txq = &qs->txq[TXQ_ETH];
1655 	struct port_info *pi = qs->port;
1656 	if_t ifp = pi->ifp;
1657 
1658 	if (qs->qs_flags & (QS_FLUSHING|QS_TIMEOUT))
1659 		reclaim_completed_tx(qs, 0, TXQ_ETH);
1660 
1661 	if (!pi->link_config.link_ok) {
1662 		TXQ_RING_FLUSH(qs);
1663 		return;
1664 	}
1665 	TXQ_LOCK_ASSERT(qs);
1666 	while (!TXQ_RING_EMPTY(qs) && (if_getdrvflags(ifp) & IFF_DRV_RUNNING) &&
1667 	    pi->link_config.link_ok) {
1668 		reclaim_completed_tx(qs, cxgb_tx_reclaim_threshold, TXQ_ETH);
1669 
1670 		if (txq->size - txq->in_use <= TX_MAX_DESC)
1671 			break;
1672 
1673 		if ((m_head = cxgb_dequeue(qs)) == NULL)
1674 			break;
1675 		/*
1676 		 *  Encapsulation can modify our pointer, and or make it
1677 		 *  NULL on failure.  In that event, we can't requeue.
1678 		 */
1679 		if (t3_encap(qs, &m_head) || m_head == NULL)
1680 			break;
1681 
1682 		m_head = NULL;
1683 	}
1684 
1685 	if (txq->db_pending)
1686 		check_ring_tx_db(pi->adapter, txq, 1);
1687 
1688 	if (!TXQ_RING_EMPTY(qs) && callout_pending(&txq->txq_timer) == 0 &&
1689 	    pi->link_config.link_ok)
1690 		callout_reset_on(&txq->txq_timer, 1, cxgb_tx_timeout,
1691 		    qs, txq->txq_timer.c_cpu);
1692 	if (m_head != NULL)
1693 		m_freem(m_head);
1694 }
1695 
1696 static int
1697 cxgb_transmit_locked(if_t ifp, struct sge_qset *qs, struct mbuf *m)
1698 {
1699 	struct port_info *pi = qs->port;
1700 	struct sge_txq *txq = &qs->txq[TXQ_ETH];
1701 	struct buf_ring *br = txq->txq_mr;
1702 	int error, avail;
1703 
1704 	avail = txq->size - txq->in_use;
1705 	TXQ_LOCK_ASSERT(qs);
1706 
1707 	/*
1708 	 * We can only do a direct transmit if the following are true:
1709 	 * - we aren't coalescing (ring < 3/4 full)
1710 	 * - the link is up -- checked in caller
1711 	 * - there are no packets enqueued already
1712 	 * - there is space in hardware transmit queue
1713 	 */
1714 	if (check_pkt_coalesce(qs) == 0 &&
1715 	    !TXQ_RING_NEEDS_ENQUEUE(qs) && avail > TX_MAX_DESC) {
1716 		if (t3_encap(qs, &m)) {
1717 			if (m != NULL &&
1718 			    (error = drbr_enqueue(ifp, br, m)) != 0)
1719 				return (error);
1720 		} else {
1721 			if (txq->db_pending)
1722 				check_ring_tx_db(pi->adapter, txq, 1);
1723 
1724 			/*
1725 			 * We've bypassed the buf ring so we need to update
1726 			 * the stats directly
1727 			 */
1728 			txq->txq_direct_packets++;
1729 			txq->txq_direct_bytes += m->m_pkthdr.len;
1730 		}
1731 	} else if ((error = drbr_enqueue(ifp, br, m)) != 0)
1732 		return (error);
1733 
1734 	reclaim_completed_tx(qs, cxgb_tx_reclaim_threshold, TXQ_ETH);
1735 	if (!TXQ_RING_EMPTY(qs) && pi->link_config.link_ok &&
1736 	    (!check_pkt_coalesce(qs) || (drbr_inuse(ifp, br) >= 7)))
1737 		cxgb_start_locked(qs);
1738 	else if (!TXQ_RING_EMPTY(qs) && !callout_pending(&txq->txq_timer))
1739 		callout_reset_on(&txq->txq_timer, 1, cxgb_tx_timeout,
1740 		    qs, txq->txq_timer.c_cpu);
1741 	return (0);
1742 }
1743 
1744 int
1745 cxgb_transmit(if_t ifp, struct mbuf *m)
1746 {
1747 	struct sge_qset *qs;
1748 	struct port_info *pi = if_getsoftc(ifp);
1749 	int error, qidx = pi->first_qset;
1750 
1751 	if ((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0
1752 	    ||(!pi->link_config.link_ok)) {
1753 		m_freem(m);
1754 		return (0);
1755 	}
1756 
1757 	/* check if flowid is set */
1758 	if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
1759 		qidx = (m->m_pkthdr.flowid % pi->nqsets) + pi->first_qset;
1760 
1761 	qs = &pi->adapter->sge.qs[qidx];
1762 
1763 	if (TXQ_TRYLOCK(qs)) {
1764 		/* XXX running */
1765 		error = cxgb_transmit_locked(ifp, qs, m);
1766 		TXQ_UNLOCK(qs);
1767 	} else
1768 		error = drbr_enqueue(ifp, qs->txq[TXQ_ETH].txq_mr, m);
1769 	return (error);
1770 }
1771 
1772 void
1773 cxgb_qflush(if_t ifp)
1774 {
1775 	/*
1776 	 * flush any enqueued mbufs in the buf_rings
1777 	 * and in the transmit queues
1778 	 * no-op for now
1779 	 */
1780 	return;
1781 }
1782 
1783 /**
1784  *	write_imm - write a packet into a Tx descriptor as immediate data
1785  *	@d: the Tx descriptor to write
1786  *	@m: the packet
1787  *	@len: the length of packet data to write as immediate data
1788  *	@gen: the generation bit value to write
1789  *
1790  *	Writes a packet as immediate data into a Tx descriptor.  The packet
1791  *	contains a work request at its beginning.  We must write the packet
1792  *	carefully so the SGE doesn't read accidentally before it's written in
1793  *	its entirety.
1794  */
1795 static __inline void
1796 write_imm(struct tx_desc *d, caddr_t src,
1797 	  unsigned int len, unsigned int gen)
1798 {
1799 	struct work_request_hdr *from = (struct work_request_hdr *)src;
1800 	struct work_request_hdr *to = (struct work_request_hdr *)d;
1801 	uint32_t wr_hi, wr_lo;
1802 
1803 	KASSERT(len <= WR_LEN && len >= sizeof(*from),
1804 	    ("%s: invalid len %d", __func__, len));
1805 
1806 	memcpy(&to[1], &from[1], len - sizeof(*from));
1807 	wr_hi = from->wrh_hi | htonl(F_WR_SOP | F_WR_EOP |
1808 	    V_WR_BCNTLFLT(len & 7));
1809 	wr_lo = from->wrh_lo | htonl(V_WR_GEN(gen) | V_WR_LEN((len + 7) / 8));
1810 	set_wr_hdr(to, wr_hi, wr_lo);
1811 	wmb();
1812 	wr_gen2(d, gen);
1813 }
1814 
1815 /**
1816  *	check_desc_avail - check descriptor availability on a send queue
1817  *	@adap: the adapter
1818  *	@q: the TX queue
1819  *	@m: the packet needing the descriptors
1820  *	@ndesc: the number of Tx descriptors needed
1821  *	@qid: the Tx queue number in its queue set (TXQ_OFLD or TXQ_CTRL)
1822  *
1823  *	Checks if the requested number of Tx descriptors is available on an
1824  *	SGE send queue.  If the queue is already suspended or not enough
1825  *	descriptors are available the packet is queued for later transmission.
1826  *	Must be called with the Tx queue locked.
1827  *
1828  *	Returns 0 if enough descriptors are available, 1 if there aren't
1829  *	enough descriptors and the packet has been queued, and 2 if the caller
1830  *	needs to retry because there weren't enough descriptors at the
1831  *	beginning of the call but some freed up in the mean time.
1832  */
1833 static __inline int
1834 check_desc_avail(adapter_t *adap, struct sge_txq *q,
1835 		 struct mbuf *m, unsigned int ndesc,
1836 		 unsigned int qid)
1837 {
1838 	/*
1839 	 * XXX We currently only use this for checking the control queue
1840 	 * the control queue is only used for binding qsets which happens
1841 	 * at init time so we are guaranteed enough descriptors
1842 	 */
1843 	if (__predict_false(mbufq_len(&q->sendq))) {
1844 addq_exit:	(void )mbufq_enqueue(&q->sendq, m);
1845 		return 1;
1846 	}
1847 	if (__predict_false(q->size - q->in_use < ndesc)) {
1848 
1849 		struct sge_qset *qs = txq_to_qset(q, qid);
1850 
1851 		setbit(&qs->txq_stopped, qid);
1852 		if (should_restart_tx(q) &&
1853 		    test_and_clear_bit(qid, &qs->txq_stopped))
1854 			return 2;
1855 
1856 		q->stops++;
1857 		goto addq_exit;
1858 	}
1859 	return 0;
1860 }
1861 
1862 
1863 /**
1864  *	reclaim_completed_tx_imm - reclaim completed control-queue Tx descs
1865  *	@q: the SGE control Tx queue
1866  *
1867  *	This is a variant of reclaim_completed_tx() that is used for Tx queues
1868  *	that send only immediate data (presently just the control queues) and
1869  *	thus do not have any mbufs
1870  */
1871 static __inline void
1872 reclaim_completed_tx_imm(struct sge_txq *q)
1873 {
1874 	unsigned int reclaim = q->processed - q->cleaned;
1875 
1876 	q->in_use -= reclaim;
1877 	q->cleaned += reclaim;
1878 }
1879 
1880 /**
1881  *	ctrl_xmit - send a packet through an SGE control Tx queue
1882  *	@adap: the adapter
1883  *	@q: the control queue
1884  *	@m: the packet
1885  *
1886  *	Send a packet through an SGE control Tx queue.  Packets sent through
1887  *	a control queue must fit entirely as immediate data in a single Tx
1888  *	descriptor and have no page fragments.
1889  */
1890 static int
1891 ctrl_xmit(adapter_t *adap, struct sge_qset *qs, struct mbuf *m)
1892 {
1893 	int ret;
1894 	struct work_request_hdr *wrp = mtod(m, struct work_request_hdr *);
1895 	struct sge_txq *q = &qs->txq[TXQ_CTRL];
1896 
1897 	KASSERT(m->m_len <= WR_LEN, ("%s: bad tx data", __func__));
1898 
1899 	wrp->wrh_hi |= htonl(F_WR_SOP | F_WR_EOP);
1900 	wrp->wrh_lo = htonl(V_WR_TID(q->token));
1901 
1902 	TXQ_LOCK(qs);
1903 again:	reclaim_completed_tx_imm(q);
1904 
1905 	ret = check_desc_avail(adap, q, m, 1, TXQ_CTRL);
1906 	if (__predict_false(ret)) {
1907 		if (ret == 1) {
1908 			TXQ_UNLOCK(qs);
1909 			return (ENOSPC);
1910 		}
1911 		goto again;
1912 	}
1913 	write_imm(&q->desc[q->pidx], m->m_data, m->m_len, q->gen);
1914 
1915 	q->in_use++;
1916 	if (++q->pidx >= q->size) {
1917 		q->pidx = 0;
1918 		q->gen ^= 1;
1919 	}
1920 	TXQ_UNLOCK(qs);
1921 	wmb();
1922 	t3_write_reg(adap, A_SG_KDOORBELL,
1923 	    F_SELEGRCNTX | V_EGRCNTX(q->cntxt_id));
1924 
1925 	m_free(m);
1926 	return (0);
1927 }
1928 
1929 
1930 /**
1931  *	restart_ctrlq - restart a suspended control queue
1932  *	@qs: the queue set cotaining the control queue
1933  *
1934  *	Resumes transmission on a suspended Tx control queue.
1935  */
1936 static void
1937 restart_ctrlq(void *data, int npending)
1938 {
1939 	struct mbuf *m;
1940 	struct sge_qset *qs = (struct sge_qset *)data;
1941 	struct sge_txq *q = &qs->txq[TXQ_CTRL];
1942 	adapter_t *adap = qs->port->adapter;
1943 
1944 	TXQ_LOCK(qs);
1945 again:	reclaim_completed_tx_imm(q);
1946 
1947 	while (q->in_use < q->size &&
1948 	       (m = mbufq_dequeue(&q->sendq)) != NULL) {
1949 
1950 		write_imm(&q->desc[q->pidx], m->m_data, m->m_len, q->gen);
1951 		m_free(m);
1952 
1953 		if (++q->pidx >= q->size) {
1954 			q->pidx = 0;
1955 			q->gen ^= 1;
1956 		}
1957 		q->in_use++;
1958 	}
1959 	if (mbufq_len(&q->sendq)) {
1960 		setbit(&qs->txq_stopped, TXQ_CTRL);
1961 
1962 		if (should_restart_tx(q) &&
1963 		    test_and_clear_bit(TXQ_CTRL, &qs->txq_stopped))
1964 			goto again;
1965 		q->stops++;
1966 	}
1967 	TXQ_UNLOCK(qs);
1968 	t3_write_reg(adap, A_SG_KDOORBELL,
1969 		     F_SELEGRCNTX | V_EGRCNTX(q->cntxt_id));
1970 }
1971 
1972 
1973 /*
1974  * Send a management message through control queue 0
1975  */
1976 int
1977 t3_mgmt_tx(struct adapter *adap, struct mbuf *m)
1978 {
1979 	return ctrl_xmit(adap, &adap->sge.qs[0], m);
1980 }
1981 
1982 /**
1983  *	free_qset - free the resources of an SGE queue set
1984  *	@sc: the controller owning the queue set
1985  *	@q: the queue set
1986  *
1987  *	Release the HW and SW resources associated with an SGE queue set, such
1988  *	as HW contexts, packet buffers, and descriptor rings.  Traffic to the
1989  *	queue set must be quiesced prior to calling this.
1990  */
1991 static void
1992 t3_free_qset(adapter_t *sc, struct sge_qset *q)
1993 {
1994 	int i;
1995 
1996 	reclaim_completed_tx(q, 0, TXQ_ETH);
1997 	if (q->txq[TXQ_ETH].txq_mr != NULL)
1998 		buf_ring_free(q->txq[TXQ_ETH].txq_mr, M_DEVBUF);
1999 	if (q->txq[TXQ_ETH].txq_ifq != NULL) {
2000 		ifq_delete(q->txq[TXQ_ETH].txq_ifq);
2001 		free(q->txq[TXQ_ETH].txq_ifq, M_DEVBUF);
2002 	}
2003 
2004 	for (i = 0; i < SGE_RXQ_PER_SET; ++i) {
2005 		if (q->fl[i].desc) {
2006 			mtx_lock_spin(&sc->sge.reg_lock);
2007 			t3_sge_disable_fl(sc, q->fl[i].cntxt_id);
2008 			mtx_unlock_spin(&sc->sge.reg_lock);
2009 			bus_dmamap_unload(q->fl[i].desc_tag, q->fl[i].desc_map);
2010 			bus_dmamem_free(q->fl[i].desc_tag, q->fl[i].desc,
2011 					q->fl[i].desc_map);
2012 			bus_dma_tag_destroy(q->fl[i].desc_tag);
2013 			bus_dma_tag_destroy(q->fl[i].entry_tag);
2014 		}
2015 		if (q->fl[i].sdesc) {
2016 			free_rx_bufs(sc, &q->fl[i]);
2017 			free(q->fl[i].sdesc, M_DEVBUF);
2018 		}
2019 	}
2020 
2021 	mtx_unlock(&q->lock);
2022 	MTX_DESTROY(&q->lock);
2023 	for (i = 0; i < SGE_TXQ_PER_SET; i++) {
2024 		if (q->txq[i].desc) {
2025 			mtx_lock_spin(&sc->sge.reg_lock);
2026 			t3_sge_enable_ecntxt(sc, q->txq[i].cntxt_id, 0);
2027 			mtx_unlock_spin(&sc->sge.reg_lock);
2028 			bus_dmamap_unload(q->txq[i].desc_tag,
2029 					q->txq[i].desc_map);
2030 			bus_dmamem_free(q->txq[i].desc_tag, q->txq[i].desc,
2031 					q->txq[i].desc_map);
2032 			bus_dma_tag_destroy(q->txq[i].desc_tag);
2033 			bus_dma_tag_destroy(q->txq[i].entry_tag);
2034 		}
2035 		if (q->txq[i].sdesc) {
2036 			free(q->txq[i].sdesc, M_DEVBUF);
2037 		}
2038 	}
2039 
2040 	if (q->rspq.desc) {
2041 		mtx_lock_spin(&sc->sge.reg_lock);
2042 		t3_sge_disable_rspcntxt(sc, q->rspq.cntxt_id);
2043 		mtx_unlock_spin(&sc->sge.reg_lock);
2044 
2045 		bus_dmamap_unload(q->rspq.desc_tag, q->rspq.desc_map);
2046 		bus_dmamem_free(q->rspq.desc_tag, q->rspq.desc,
2047 			        q->rspq.desc_map);
2048 		bus_dma_tag_destroy(q->rspq.desc_tag);
2049 		MTX_DESTROY(&q->rspq.lock);
2050 	}
2051 
2052 #if defined(INET6) || defined(INET)
2053 	tcp_lro_free(&q->lro.ctrl);
2054 #endif
2055 
2056 	bzero(q, sizeof(*q));
2057 }
2058 
2059 /**
2060  *	t3_free_sge_resources - free SGE resources
2061  *	@sc: the adapter softc
2062  *
2063  *	Frees resources used by the SGE queue sets.
2064  */
2065 void
2066 t3_free_sge_resources(adapter_t *sc, int nqsets)
2067 {
2068 	int i;
2069 
2070 	for (i = 0; i < nqsets; ++i) {
2071 		TXQ_LOCK(&sc->sge.qs[i]);
2072 		t3_free_qset(sc, &sc->sge.qs[i]);
2073 	}
2074 }
2075 
2076 /**
2077  *	t3_sge_start - enable SGE
2078  *	@sc: the controller softc
2079  *
2080  *	Enables the SGE for DMAs.  This is the last step in starting packet
2081  *	transfers.
2082  */
2083 void
2084 t3_sge_start(adapter_t *sc)
2085 {
2086 	t3_set_reg_field(sc, A_SG_CONTROL, F_GLOBALENABLE, F_GLOBALENABLE);
2087 }
2088 
2089 /**
2090  *	t3_sge_stop - disable SGE operation
2091  *	@sc: the adapter
2092  *
2093  *	Disables the DMA engine.  This can be called in emeregencies (e.g.,
2094  *	from error interrupts) or from normal process context.  In the latter
2095  *	case it also disables any pending queue restart tasklets.  Note that
2096  *	if it is called in interrupt context it cannot disable the restart
2097  *	tasklets as it cannot wait, however the tasklets will have no effect
2098  *	since the doorbells are disabled and the driver will call this again
2099  *	later from process context, at which time the tasklets will be stopped
2100  *	if they are still running.
2101  */
2102 void
2103 t3_sge_stop(adapter_t *sc)
2104 {
2105 
2106 	t3_set_reg_field(sc, A_SG_CONTROL, F_GLOBALENABLE, 0);
2107 }
2108 
2109 /**
2110  *	t3_free_tx_desc - reclaims Tx descriptors and their buffers
2111  *	@adapter: the adapter
2112  *	@q: the Tx queue to reclaim descriptors from
2113  *	@reclaimable: the number of descriptors to reclaim
2114  *      @m_vec_size: maximum number of buffers to reclaim
2115  *      @desc_reclaimed: returns the number of descriptors reclaimed
2116  *
2117  *	Reclaims Tx descriptors from an SGE Tx queue and frees the associated
2118  *	Tx buffers.  Called with the Tx queue lock held.
2119  *
2120  *      Returns number of buffers of reclaimed
2121  */
2122 void
2123 t3_free_tx_desc(struct sge_qset *qs, int reclaimable, int queue)
2124 {
2125 	struct tx_sw_desc *txsd;
2126 	unsigned int cidx, mask;
2127 	struct sge_txq *q = &qs->txq[queue];
2128 
2129 #ifdef T3_TRACE
2130 	T3_TRACE2(sc->tb[q->cntxt_id & 7],
2131 		  "reclaiming %u Tx descriptors at cidx %u", reclaimable, cidx);
2132 #endif
2133 	cidx = q->cidx;
2134 	mask = q->size - 1;
2135 	txsd = &q->sdesc[cidx];
2136 
2137 	mtx_assert(&qs->lock, MA_OWNED);
2138 	while (reclaimable--) {
2139 		prefetch(q->sdesc[(cidx + 1) & mask].m);
2140 		prefetch(q->sdesc[(cidx + 2) & mask].m);
2141 
2142 		if (txsd->m != NULL) {
2143 			if (txsd->flags & TX_SW_DESC_MAPPED) {
2144 				bus_dmamap_unload(q->entry_tag, txsd->map);
2145 				txsd->flags &= ~TX_SW_DESC_MAPPED;
2146 			}
2147 			m_freem_list(txsd->m);
2148 			txsd->m = NULL;
2149 		} else
2150 			q->txq_skipped++;
2151 
2152 		++txsd;
2153 		if (++cidx == q->size) {
2154 			cidx = 0;
2155 			txsd = q->sdesc;
2156 		}
2157 	}
2158 	q->cidx = cidx;
2159 
2160 }
2161 
2162 /**
2163  *	is_new_response - check if a response is newly written
2164  *	@r: the response descriptor
2165  *	@q: the response queue
2166  *
2167  *	Returns true if a response descriptor contains a yet unprocessed
2168  *	response.
2169  */
2170 static __inline int
2171 is_new_response(const struct rsp_desc *r,
2172     const struct sge_rspq *q)
2173 {
2174 	return (r->intr_gen & F_RSPD_GEN2) == q->gen;
2175 }
2176 
2177 #define RSPD_GTS_MASK  (F_RSPD_TXQ0_GTS | F_RSPD_TXQ1_GTS)
2178 #define RSPD_CTRL_MASK (RSPD_GTS_MASK | \
2179 			V_RSPD_TXQ0_CR(M_RSPD_TXQ0_CR) | \
2180 			V_RSPD_TXQ1_CR(M_RSPD_TXQ1_CR) | \
2181 			V_RSPD_TXQ2_CR(M_RSPD_TXQ2_CR))
2182 
2183 /* How long to delay the next interrupt in case of memory shortage, in 0.1us. */
2184 #define NOMEM_INTR_DELAY 2500
2185 
2186 #ifdef TCP_OFFLOAD
2187 /**
2188  *	write_ofld_wr - write an offload work request
2189  *	@adap: the adapter
2190  *	@m: the packet to send
2191  *	@q: the Tx queue
2192  *	@pidx: index of the first Tx descriptor to write
2193  *	@gen: the generation value to use
2194  *	@ndesc: number of descriptors the packet will occupy
2195  *
2196  *	Write an offload work request to send the supplied packet.  The packet
2197  *	data already carry the work request with most fields populated.
2198  */
2199 static void
2200 write_ofld_wr(adapter_t *adap, struct mbuf *m, struct sge_txq *q,
2201     unsigned int pidx, unsigned int gen, unsigned int ndesc)
2202 {
2203 	unsigned int sgl_flits, flits;
2204 	int i, idx, nsegs, wrlen;
2205 	struct work_request_hdr *from;
2206 	struct sg_ent *sgp, t3sgl[TX_MAX_SEGS / 2 + 1];
2207 	struct tx_desc *d = &q->desc[pidx];
2208 	struct txq_state txqs;
2209 	struct sglist_seg *segs;
2210 	struct ofld_hdr *oh = mtod(m, struct ofld_hdr *);
2211 	struct sglist *sgl;
2212 
2213 	from = (void *)(oh + 1);	/* Start of WR within mbuf */
2214 	wrlen = m->m_len - sizeof(*oh);
2215 
2216 	if (!(oh->flags & F_HDR_SGL)) {
2217 		write_imm(d, (caddr_t)from, wrlen, gen);
2218 
2219 		/*
2220 		 * mbuf with "real" immediate tx data will be enqueue_wr'd by
2221 		 * t3_push_frames and freed in wr_ack.  Others, like those sent
2222 		 * down by close_conn, t3_send_reset, etc. should be freed here.
2223 		 */
2224 		if (!(oh->flags & F_HDR_DF))
2225 			m_free(m);
2226 		return;
2227 	}
2228 
2229 	memcpy(&d->flit[1], &from[1], wrlen - sizeof(*from));
2230 
2231 	sgl = oh->sgl;
2232 	flits = wrlen / 8;
2233 	sgp = (ndesc == 1) ? (struct sg_ent *)&d->flit[flits] : t3sgl;
2234 
2235 	nsegs = sgl->sg_nseg;
2236 	segs = sgl->sg_segs;
2237 	for (idx = 0, i = 0; i < nsegs; i++) {
2238 		KASSERT(segs[i].ss_len, ("%s: 0 len in sgl", __func__));
2239 		if (i && idx == 0)
2240 			++sgp;
2241 		sgp->len[idx] = htobe32(segs[i].ss_len);
2242 		sgp->addr[idx] = htobe64(segs[i].ss_paddr);
2243 		idx ^= 1;
2244 	}
2245 	if (idx) {
2246 		sgp->len[idx] = 0;
2247 		sgp->addr[idx] = 0;
2248 	}
2249 
2250 	sgl_flits = sgl_len(nsegs);
2251 	txqs.gen = gen;
2252 	txqs.pidx = pidx;
2253 	txqs.compl = 0;
2254 
2255 	write_wr_hdr_sgl(ndesc, d, &txqs, q, t3sgl, flits, sgl_flits,
2256 	    from->wrh_hi, from->wrh_lo);
2257 }
2258 
2259 /**
2260  *	ofld_xmit - send a packet through an offload queue
2261  *	@adap: the adapter
2262  *	@q: the Tx offload queue
2263  *	@m: the packet
2264  *
2265  *	Send an offload packet through an SGE offload queue.
2266  */
2267 static int
2268 ofld_xmit(adapter_t *adap, struct sge_qset *qs, struct mbuf *m)
2269 {
2270 	int ret;
2271 	unsigned int ndesc;
2272 	unsigned int pidx, gen;
2273 	struct sge_txq *q = &qs->txq[TXQ_OFLD];
2274 	struct ofld_hdr *oh = mtod(m, struct ofld_hdr *);
2275 
2276 	ndesc = G_HDR_NDESC(oh->flags);
2277 
2278 	TXQ_LOCK(qs);
2279 again:	reclaim_completed_tx(qs, 16, TXQ_OFLD);
2280 	ret = check_desc_avail(adap, q, m, ndesc, TXQ_OFLD);
2281 	if (__predict_false(ret)) {
2282 		if (ret == 1) {
2283 			TXQ_UNLOCK(qs);
2284 			return (EINTR);
2285 		}
2286 		goto again;
2287 	}
2288 
2289 	gen = q->gen;
2290 	q->in_use += ndesc;
2291 	pidx = q->pidx;
2292 	q->pidx += ndesc;
2293 	if (q->pidx >= q->size) {
2294 		q->pidx -= q->size;
2295 		q->gen ^= 1;
2296 	}
2297 
2298 	write_ofld_wr(adap, m, q, pidx, gen, ndesc);
2299 	check_ring_tx_db(adap, q, 1);
2300 	TXQ_UNLOCK(qs);
2301 
2302 	return (0);
2303 }
2304 
2305 /**
2306  *	restart_offloadq - restart a suspended offload queue
2307  *	@qs: the queue set cotaining the offload queue
2308  *
2309  *	Resumes transmission on a suspended Tx offload queue.
2310  */
2311 static void
2312 restart_offloadq(void *data, int npending)
2313 {
2314 	struct mbuf *m;
2315 	struct sge_qset *qs = data;
2316 	struct sge_txq *q = &qs->txq[TXQ_OFLD];
2317 	adapter_t *adap = qs->port->adapter;
2318 
2319 	TXQ_LOCK(qs);
2320 again:
2321 	while ((m = mbufq_first(&q->sendq)) != NULL) {
2322 		unsigned int gen, pidx;
2323 		struct ofld_hdr *oh = mtod(m, struct ofld_hdr *);
2324 		unsigned int ndesc = G_HDR_NDESC(oh->flags);
2325 
2326 		if (__predict_false(q->size - q->in_use < ndesc)) {
2327 			setbit(&qs->txq_stopped, TXQ_OFLD);
2328 			if (should_restart_tx(q) &&
2329 			    test_and_clear_bit(TXQ_OFLD, &qs->txq_stopped))
2330 				goto again;
2331 			q->stops++;
2332 			break;
2333 		}
2334 
2335 		gen = q->gen;
2336 		q->in_use += ndesc;
2337 		pidx = q->pidx;
2338 		q->pidx += ndesc;
2339 		if (q->pidx >= q->size) {
2340 			q->pidx -= q->size;
2341 			q->gen ^= 1;
2342 		}
2343 
2344 		(void)mbufq_dequeue(&q->sendq);
2345 		TXQ_UNLOCK(qs);
2346 		write_ofld_wr(adap, m, q, pidx, gen, ndesc);
2347 		TXQ_LOCK(qs);
2348 	}
2349 #if USE_GTS
2350 	set_bit(TXQ_RUNNING, &q->flags);
2351 	set_bit(TXQ_LAST_PKT_DB, &q->flags);
2352 #endif
2353 	TXQ_UNLOCK(qs);
2354 	wmb();
2355 	t3_write_reg(adap, A_SG_KDOORBELL,
2356 		     F_SELEGRCNTX | V_EGRCNTX(q->cntxt_id));
2357 }
2358 
2359 /**
2360  *	t3_offload_tx - send an offload packet
2361  *	@m: the packet
2362  *
2363  *	Sends an offload packet.  We use the packet priority to select the
2364  *	appropriate Tx queue as follows: bit 0 indicates whether the packet
2365  *	should be sent as regular or control, bits 1-3 select the queue set.
2366  */
2367 int
2368 t3_offload_tx(struct adapter *sc, struct mbuf *m)
2369 {
2370 	struct ofld_hdr *oh = mtod(m, struct ofld_hdr *);
2371 	struct sge_qset *qs = &sc->sge.qs[G_HDR_QSET(oh->flags)];
2372 
2373 	if (oh->flags & F_HDR_CTRL) {
2374 		m_adj(m, sizeof (*oh));	/* trim ofld_hdr off */
2375 		return (ctrl_xmit(sc, qs, m));
2376 	} else
2377 		return (ofld_xmit(sc, qs, m));
2378 }
2379 #endif
2380 
2381 static void
2382 restart_tx(struct sge_qset *qs)
2383 {
2384 	struct adapter *sc = qs->port->adapter;
2385 
2386 	if (isset(&qs->txq_stopped, TXQ_OFLD) &&
2387 	    should_restart_tx(&qs->txq[TXQ_OFLD]) &&
2388 	    test_and_clear_bit(TXQ_OFLD, &qs->txq_stopped)) {
2389 		qs->txq[TXQ_OFLD].restarts++;
2390 		taskqueue_enqueue(sc->tq, &qs->txq[TXQ_OFLD].qresume_task);
2391 	}
2392 
2393 	if (isset(&qs->txq_stopped, TXQ_CTRL) &&
2394 	    should_restart_tx(&qs->txq[TXQ_CTRL]) &&
2395 	    test_and_clear_bit(TXQ_CTRL, &qs->txq_stopped)) {
2396 		qs->txq[TXQ_CTRL].restarts++;
2397 		taskqueue_enqueue(sc->tq, &qs->txq[TXQ_CTRL].qresume_task);
2398 	}
2399 }
2400 
2401 /**
2402  *	t3_sge_alloc_qset - initialize an SGE queue set
2403  *	@sc: the controller softc
2404  *	@id: the queue set id
2405  *	@nports: how many Ethernet ports will be using this queue set
2406  *	@irq_vec_idx: the IRQ vector index for response queue interrupts
2407  *	@p: configuration parameters for this queue set
2408  *	@ntxq: number of Tx queues for the queue set
2409  *	@pi: port info for queue set
2410  *
2411  *	Allocate resources and initialize an SGE queue set.  A queue set
2412  *	comprises a response queue, two Rx free-buffer queues, and up to 3
2413  *	Tx queues.  The Tx queues are assigned roles in the order Ethernet
2414  *	queue, offload queue, and control queue.
2415  */
2416 int
2417 t3_sge_alloc_qset(adapter_t *sc, u_int id, int nports, int irq_vec_idx,
2418 		  const struct qset_params *p, int ntxq, struct port_info *pi)
2419 {
2420 	struct sge_qset *q = &sc->sge.qs[id];
2421 	int i, ret = 0;
2422 
2423 	MTX_INIT(&q->lock, q->namebuf, NULL, MTX_DEF);
2424 	q->port = pi;
2425 	q->adap = sc;
2426 
2427 	if ((q->txq[TXQ_ETH].txq_mr = buf_ring_alloc(cxgb_txq_buf_ring_size,
2428 	    M_DEVBUF, M_WAITOK, &q->lock)) == NULL) {
2429 		device_printf(sc->dev, "failed to allocate mbuf ring\n");
2430 		goto err;
2431 	}
2432 	if ((q->txq[TXQ_ETH].txq_ifq = malloc(sizeof(struct ifaltq), M_DEVBUF,
2433 	    M_NOWAIT | M_ZERO)) == NULL) {
2434 		device_printf(sc->dev, "failed to allocate ifq\n");
2435 		goto err;
2436 	}
2437 	ifq_init(q->txq[TXQ_ETH].txq_ifq, pi->ifp);
2438 	callout_init(&q->txq[TXQ_ETH].txq_timer, 1);
2439 	callout_init(&q->txq[TXQ_ETH].txq_watchdog, 1);
2440 	q->txq[TXQ_ETH].txq_timer.c_cpu = id % mp_ncpus;
2441 	q->txq[TXQ_ETH].txq_watchdog.c_cpu = id % mp_ncpus;
2442 
2443 	init_qset_cntxt(q, id);
2444 	q->idx = id;
2445 	if ((ret = alloc_ring(sc, p->fl_size, sizeof(struct rx_desc),
2446 		    sizeof(struct rx_sw_desc), &q->fl[0].phys_addr,
2447 		    &q->fl[0].desc, &q->fl[0].sdesc,
2448 		    &q->fl[0].desc_tag, &q->fl[0].desc_map,
2449 		    sc->rx_dmat, &q->fl[0].entry_tag)) != 0) {
2450 		printf("error %d from alloc ring fl0\n", ret);
2451 		goto err;
2452 	}
2453 
2454 	if ((ret = alloc_ring(sc, p->jumbo_size, sizeof(struct rx_desc),
2455 		    sizeof(struct rx_sw_desc), &q->fl[1].phys_addr,
2456 		    &q->fl[1].desc, &q->fl[1].sdesc,
2457 		    &q->fl[1].desc_tag, &q->fl[1].desc_map,
2458 		    sc->rx_jumbo_dmat, &q->fl[1].entry_tag)) != 0) {
2459 		printf("error %d from alloc ring fl1\n", ret);
2460 		goto err;
2461 	}
2462 
2463 	if ((ret = alloc_ring(sc, p->rspq_size, sizeof(struct rsp_desc), 0,
2464 		    &q->rspq.phys_addr, &q->rspq.desc, NULL,
2465 		    &q->rspq.desc_tag, &q->rspq.desc_map,
2466 		    NULL, NULL)) != 0) {
2467 		printf("error %d from alloc ring rspq\n", ret);
2468 		goto err;
2469 	}
2470 
2471 	snprintf(q->rspq.lockbuf, RSPQ_NAME_LEN, "t3 rspq lock %d:%d",
2472 	    device_get_unit(sc->dev), irq_vec_idx);
2473 	MTX_INIT(&q->rspq.lock, q->rspq.lockbuf, NULL, MTX_DEF);
2474 
2475 	for (i = 0; i < ntxq; ++i) {
2476 		size_t sz = i == TXQ_CTRL ? 0 : sizeof(struct tx_sw_desc);
2477 
2478 		if ((ret = alloc_ring(sc, p->txq_size[i],
2479 			    sizeof(struct tx_desc), sz,
2480 			    &q->txq[i].phys_addr, &q->txq[i].desc,
2481 			    &q->txq[i].sdesc, &q->txq[i].desc_tag,
2482 			    &q->txq[i].desc_map,
2483 			    sc->tx_dmat, &q->txq[i].entry_tag)) != 0) {
2484 			printf("error %d from alloc ring tx %i\n", ret, i);
2485 			goto err;
2486 		}
2487 		mbufq_init(&q->txq[i].sendq, INT_MAX);
2488 		q->txq[i].gen = 1;
2489 		q->txq[i].size = p->txq_size[i];
2490 	}
2491 
2492 #ifdef TCP_OFFLOAD
2493 	TASK_INIT(&q->txq[TXQ_OFLD].qresume_task, 0, restart_offloadq, q);
2494 #endif
2495 	TASK_INIT(&q->txq[TXQ_CTRL].qresume_task, 0, restart_ctrlq, q);
2496 	TASK_INIT(&q->txq[TXQ_ETH].qreclaim_task, 0, sge_txq_reclaim_handler, q);
2497 	TASK_INIT(&q->txq[TXQ_OFLD].qreclaim_task, 0, sge_txq_reclaim_handler, q);
2498 
2499 	q->fl[0].gen = q->fl[1].gen = 1;
2500 	q->fl[0].size = p->fl_size;
2501 	q->fl[1].size = p->jumbo_size;
2502 
2503 	q->rspq.gen = 1;
2504 	q->rspq.cidx = 0;
2505 	q->rspq.size = p->rspq_size;
2506 
2507 	q->txq[TXQ_ETH].stop_thres = nports *
2508 	    flits_to_desc(sgl_len(TX_MAX_SEGS + 1) + 3);
2509 
2510 	q->fl[0].buf_size = MCLBYTES;
2511 	q->fl[0].zone = zone_pack;
2512 	q->fl[0].type = EXT_PACKET;
2513 
2514 	if (p->jumbo_buf_size ==  MJUM16BYTES) {
2515 		q->fl[1].zone = zone_jumbo16;
2516 		q->fl[1].type = EXT_JUMBO16;
2517 	} else if (p->jumbo_buf_size ==  MJUM9BYTES) {
2518 		q->fl[1].zone = zone_jumbo9;
2519 		q->fl[1].type = EXT_JUMBO9;
2520 	} else if (p->jumbo_buf_size ==  MJUMPAGESIZE) {
2521 		q->fl[1].zone = zone_jumbop;
2522 		q->fl[1].type = EXT_JUMBOP;
2523 	} else {
2524 		KASSERT(0, ("can't deal with jumbo_buf_size %d.", p->jumbo_buf_size));
2525 		ret = EDOOFUS;
2526 		goto err;
2527 	}
2528 	q->fl[1].buf_size = p->jumbo_buf_size;
2529 
2530 	/* Allocate and setup the lro_ctrl structure */
2531 	q->lro.enabled = !!(if_getcapenable(pi->ifp) & IFCAP_LRO);
2532 #if defined(INET6) || defined(INET)
2533 	ret = tcp_lro_init(&q->lro.ctrl);
2534 	if (ret) {
2535 		printf("error %d from tcp_lro_init\n", ret);
2536 		goto err;
2537 	}
2538 #endif
2539 	q->lro.ctrl.ifp = pi->ifp;
2540 
2541 	mtx_lock_spin(&sc->sge.reg_lock);
2542 	ret = -t3_sge_init_rspcntxt(sc, q->rspq.cntxt_id, irq_vec_idx,
2543 				   q->rspq.phys_addr, q->rspq.size,
2544 				   q->fl[0].buf_size, 1, 0);
2545 	if (ret) {
2546 		printf("error %d from t3_sge_init_rspcntxt\n", ret);
2547 		goto err_unlock;
2548 	}
2549 
2550 	for (i = 0; i < SGE_RXQ_PER_SET; ++i) {
2551 		ret = -t3_sge_init_flcntxt(sc, q->fl[i].cntxt_id, 0,
2552 					  q->fl[i].phys_addr, q->fl[i].size,
2553 					  q->fl[i].buf_size, p->cong_thres, 1,
2554 					  0);
2555 		if (ret) {
2556 			printf("error %d from t3_sge_init_flcntxt for index i=%d\n", ret, i);
2557 			goto err_unlock;
2558 		}
2559 	}
2560 
2561 	ret = -t3_sge_init_ecntxt(sc, q->txq[TXQ_ETH].cntxt_id, USE_GTS,
2562 				 SGE_CNTXT_ETH, id, q->txq[TXQ_ETH].phys_addr,
2563 				 q->txq[TXQ_ETH].size, q->txq[TXQ_ETH].token,
2564 				 1, 0);
2565 	if (ret) {
2566 		printf("error %d from t3_sge_init_ecntxt\n", ret);
2567 		goto err_unlock;
2568 	}
2569 
2570 	if (ntxq > 1) {
2571 		ret = -t3_sge_init_ecntxt(sc, q->txq[TXQ_OFLD].cntxt_id,
2572 					 USE_GTS, SGE_CNTXT_OFLD, id,
2573 					 q->txq[TXQ_OFLD].phys_addr,
2574 					 q->txq[TXQ_OFLD].size, 0, 1, 0);
2575 		if (ret) {
2576 			printf("error %d from t3_sge_init_ecntxt\n", ret);
2577 			goto err_unlock;
2578 		}
2579 	}
2580 
2581 	if (ntxq > 2) {
2582 		ret = -t3_sge_init_ecntxt(sc, q->txq[TXQ_CTRL].cntxt_id, 0,
2583 					 SGE_CNTXT_CTRL, id,
2584 					 q->txq[TXQ_CTRL].phys_addr,
2585 					 q->txq[TXQ_CTRL].size,
2586 					 q->txq[TXQ_CTRL].token, 1, 0);
2587 		if (ret) {
2588 			printf("error %d from t3_sge_init_ecntxt\n", ret);
2589 			goto err_unlock;
2590 		}
2591 	}
2592 
2593 	mtx_unlock_spin(&sc->sge.reg_lock);
2594 	t3_update_qset_coalesce(q, p);
2595 
2596 	refill_fl(sc, &q->fl[0], q->fl[0].size);
2597 	refill_fl(sc, &q->fl[1], q->fl[1].size);
2598 	refill_rspq(sc, &q->rspq, q->rspq.size - 1);
2599 
2600 	t3_write_reg(sc, A_SG_GTS, V_RSPQ(q->rspq.cntxt_id) |
2601 		     V_NEWTIMER(q->rspq.holdoff_tmr));
2602 
2603 	return (0);
2604 
2605 err_unlock:
2606 	mtx_unlock_spin(&sc->sge.reg_lock);
2607 err:
2608 	TXQ_LOCK(q);
2609 	t3_free_qset(sc, q);
2610 
2611 	return (ret);
2612 }
2613 
2614 /*
2615  * Remove CPL_RX_PKT headers from the mbuf and reduce it to a regular mbuf with
2616  * ethernet data.  Hardware assistance with various checksums and any vlan tag
2617  * will also be taken into account here.
2618  */
2619 void
2620 t3_rx_eth(struct adapter *adap, struct mbuf *m, int ethpad)
2621 {
2622 	struct cpl_rx_pkt *cpl = (struct cpl_rx_pkt *)(mtod(m, uint8_t *) + ethpad);
2623 	struct port_info *pi = &adap->port[adap->rxpkt_map[cpl->iff]];
2624 	if_t ifp = pi->ifp;
2625 
2626 	if (cpl->vlan_valid) {
2627 		m->m_pkthdr.ether_vtag = ntohs(cpl->vlan);
2628 		m->m_flags |= M_VLANTAG;
2629 	}
2630 
2631 	m->m_pkthdr.rcvif = ifp;
2632 	/*
2633 	 * adjust after conversion to mbuf chain
2634 	 */
2635 	m->m_pkthdr.len -= (sizeof(*cpl) + ethpad);
2636 	m->m_len -= (sizeof(*cpl) + ethpad);
2637 	m->m_data += (sizeof(*cpl) + ethpad);
2638 
2639 	if (!cpl->fragment && cpl->csum_valid && cpl->csum == 0xffff) {
2640 		struct ether_header *eh = mtod(m, void *);
2641 		uint16_t eh_type;
2642 
2643 		if (eh->ether_type == htons(ETHERTYPE_VLAN)) {
2644 			struct ether_vlan_header *evh = mtod(m, void *);
2645 
2646 			eh_type = evh->evl_proto;
2647 		} else
2648 			eh_type = eh->ether_type;
2649 
2650 		if (if_getcapenable(ifp) & IFCAP_RXCSUM &&
2651 		    eh_type == htons(ETHERTYPE_IP)) {
2652 			m->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
2653 			    CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
2654 			m->m_pkthdr.csum_data = 0xffff;
2655 		} else if (if_getcapenable(ifp) & IFCAP_RXCSUM_IPV6 &&
2656 		    eh_type == htons(ETHERTYPE_IPV6)) {
2657 			m->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
2658 			    CSUM_PSEUDO_HDR);
2659 			m->m_pkthdr.csum_data = 0xffff;
2660 		}
2661 	}
2662 }
2663 
2664 /**
2665  *	get_packet - return the next ingress packet buffer from a free list
2666  *	@adap: the adapter that received the packet
2667  *	@drop_thres: # of remaining buffers before we start dropping packets
2668  *	@qs: the qset that the SGE free list holding the packet belongs to
2669  *      @mh: the mbuf header, contains a pointer to the head and tail of the mbuf chain
2670  *      @r: response descriptor
2671  *
2672  *	Get the next packet from a free list and complete setup of the
2673  *	sk_buff.  If the packet is small we make a copy and recycle the
2674  *	original buffer, otherwise we use the original buffer itself.  If a
2675  *	positive drop threshold is supplied packets are dropped and their
2676  *	buffers recycled if (a) the number of remaining buffers is under the
2677  *	threshold and the packet is too big to copy, or (b) the packet should
2678  *	be copied but there is no memory for the copy.
2679  */
2680 static int
2681 get_packet(adapter_t *adap, unsigned int drop_thres, struct sge_qset *qs,
2682     struct t3_mbuf_hdr *mh, struct rsp_desc *r)
2683 {
2684 
2685 	unsigned int len_cq =  ntohl(r->len_cq);
2686 	struct sge_fl *fl = (len_cq & F_RSPD_FLQ) ? &qs->fl[1] : &qs->fl[0];
2687 	int mask, cidx = fl->cidx;
2688 	struct rx_sw_desc *sd = &fl->sdesc[cidx];
2689 	uint32_t len = G_RSPD_LEN(len_cq);
2690 	uint32_t flags = M_EXT;
2691 	uint8_t sopeop = G_RSPD_SOP_EOP(ntohl(r->flags));
2692 	caddr_t cl;
2693 	struct mbuf *m;
2694 	int ret = 0;
2695 
2696 	mask = fl->size - 1;
2697 	prefetch(fl->sdesc[(cidx + 1) & mask].m);
2698 	prefetch(fl->sdesc[(cidx + 2) & mask].m);
2699 	prefetch(fl->sdesc[(cidx + 1) & mask].rxsd_cl);
2700 	prefetch(fl->sdesc[(cidx + 2) & mask].rxsd_cl);
2701 
2702 	fl->credits--;
2703 	bus_dmamap_sync(fl->entry_tag, sd->map, BUS_DMASYNC_POSTREAD);
2704 
2705 	if (recycle_enable && len <= SGE_RX_COPY_THRES &&
2706 	    sopeop == RSPQ_SOP_EOP) {
2707 		if ((m = m_gethdr(M_NOWAIT, MT_DATA)) == NULL)
2708 			goto skip_recycle;
2709 		cl = mtod(m, void *);
2710 		memcpy(cl, sd->rxsd_cl, len);
2711 		recycle_rx_buf(adap, fl, fl->cidx);
2712 		m->m_pkthdr.len = m->m_len = len;
2713 		m->m_flags = 0;
2714 		mh->mh_head = mh->mh_tail = m;
2715 		ret = 1;
2716 		goto done;
2717 	} else {
2718 	skip_recycle:
2719 		bus_dmamap_unload(fl->entry_tag, sd->map);
2720 		cl = sd->rxsd_cl;
2721 		m = sd->m;
2722 
2723 		if ((sopeop == RSPQ_SOP_EOP) ||
2724 		    (sopeop == RSPQ_SOP))
2725 			flags |= M_PKTHDR;
2726 		m_init(m, M_NOWAIT, MT_DATA, flags);
2727 		if (fl->zone == zone_pack) {
2728 			/*
2729 			 * restore clobbered data pointer
2730 			 */
2731 			m->m_data = m->m_ext.ext_buf;
2732 		} else {
2733 			m_cljset(m, cl, fl->type);
2734 		}
2735 		m->m_len = len;
2736 	}
2737 	switch(sopeop) {
2738 	case RSPQ_SOP_EOP:
2739 		ret = 1;
2740 		/* FALLTHROUGH */
2741 	case RSPQ_SOP:
2742 		mh->mh_head = mh->mh_tail = m;
2743 		m->m_pkthdr.len = len;
2744 		break;
2745 	case RSPQ_EOP:
2746 		ret = 1;
2747 		/* FALLTHROUGH */
2748 	case RSPQ_NSOP_NEOP:
2749 		if (mh->mh_tail == NULL) {
2750 			log(LOG_ERR, "discarding intermediate descriptor entry\n");
2751 			m_freem(m);
2752 			m = NULL;
2753 			break;
2754 		}
2755 		mh->mh_tail->m_next = m;
2756 		mh->mh_tail = m;
2757 		mh->mh_head->m_pkthdr.len += len;
2758 		break;
2759 	}
2760 	if (cxgb_debug && m != NULL)
2761 		printf("len=%d pktlen=%d\n", m->m_len, m->m_pkthdr.len);
2762 done:
2763 	if (++fl->cidx == fl->size)
2764 		fl->cidx = 0;
2765 
2766 	return (ret);
2767 }
2768 
2769 /**
2770  *	handle_rsp_cntrl_info - handles control information in a response
2771  *	@qs: the queue set corresponding to the response
2772  *	@flags: the response control flags
2773  *
2774  *	Handles the control information of an SGE response, such as GTS
2775  *	indications and completion credits for the queue set's Tx queues.
2776  *	HW coalesces credits, we don't do any extra SW coalescing.
2777  */
2778 static __inline void
2779 handle_rsp_cntrl_info(struct sge_qset *qs, uint32_t flags)
2780 {
2781 	unsigned int credits;
2782 
2783 #if USE_GTS
2784 	if (flags & F_RSPD_TXQ0_GTS)
2785 		clear_bit(TXQ_RUNNING, &qs->txq[TXQ_ETH].flags);
2786 #endif
2787 	credits = G_RSPD_TXQ0_CR(flags);
2788 	if (credits)
2789 		qs->txq[TXQ_ETH].processed += credits;
2790 
2791 	credits = G_RSPD_TXQ2_CR(flags);
2792 	if (credits)
2793 		qs->txq[TXQ_CTRL].processed += credits;
2794 
2795 # if USE_GTS
2796 	if (flags & F_RSPD_TXQ1_GTS)
2797 		clear_bit(TXQ_RUNNING, &qs->txq[TXQ_OFLD].flags);
2798 # endif
2799 	credits = G_RSPD_TXQ1_CR(flags);
2800 	if (credits)
2801 		qs->txq[TXQ_OFLD].processed += credits;
2802 
2803 }
2804 
2805 static void
2806 check_ring_db(adapter_t *adap, struct sge_qset *qs,
2807     unsigned int sleeping)
2808 {
2809 	;
2810 }
2811 
2812 /**
2813  *	process_responses - process responses from an SGE response queue
2814  *	@adap: the adapter
2815  *	@qs: the queue set to which the response queue belongs
2816  *	@budget: how many responses can be processed in this round
2817  *
2818  *	Process responses from an SGE response queue up to the supplied budget.
2819  *	Responses include received packets as well as credits and other events
2820  *	for the queues that belong to the response queue's queue set.
2821  *	A negative budget is effectively unlimited.
2822  *
2823  *	Additionally choose the interrupt holdoff time for the next interrupt
2824  *	on this queue.  If the system is under memory shortage use a fairly
2825  *	long delay to help recovery.
2826  */
2827 static int
2828 process_responses(adapter_t *adap, struct sge_qset *qs, int budget)
2829 {
2830 	struct sge_rspq *rspq = &qs->rspq;
2831 	struct rsp_desc *r = &rspq->desc[rspq->cidx];
2832 	int budget_left = budget;
2833 	unsigned int sleeping = 0;
2834 #if defined(INET6) || defined(INET)
2835 	int lro_enabled = qs->lro.enabled;
2836 	int skip_lro;
2837 	struct lro_ctrl *lro_ctrl = &qs->lro.ctrl;
2838 #endif
2839 	struct t3_mbuf_hdr *mh = &rspq->rspq_mh;
2840 #ifdef DEBUG
2841 	static int last_holdoff = 0;
2842 	if (cxgb_debug && rspq->holdoff_tmr != last_holdoff) {
2843 		printf("next_holdoff=%d\n", rspq->holdoff_tmr);
2844 		last_holdoff = rspq->holdoff_tmr;
2845 	}
2846 #endif
2847 	rspq->next_holdoff = rspq->holdoff_tmr;
2848 
2849 	while (__predict_true(budget_left && is_new_response(r, rspq))) {
2850 		int eth, eop = 0, ethpad = 0;
2851 		uint32_t flags = ntohl(r->flags);
2852 		uint32_t rss_hash = be32toh(r->rss_hdr.rss_hash_val);
2853 		uint8_t opcode = r->rss_hdr.opcode;
2854 
2855 		eth = (opcode == CPL_RX_PKT);
2856 
2857 		if (__predict_false(flags & F_RSPD_ASYNC_NOTIF)) {
2858 			struct mbuf *m;
2859 
2860 			if (cxgb_debug)
2861 				printf("async notification\n");
2862 
2863 			if (mh->mh_head == NULL) {
2864 				mh->mh_head = m_gethdr(M_NOWAIT, MT_DATA);
2865 				m = mh->mh_head;
2866 			} else {
2867 				m = m_gethdr(M_NOWAIT, MT_DATA);
2868 			}
2869 			if (m == NULL)
2870 				goto no_mem;
2871 
2872                         memcpy(mtod(m, char *), r, AN_PKT_SIZE);
2873 			m->m_len = m->m_pkthdr.len = AN_PKT_SIZE;
2874                         *mtod(m, uint8_t *) = CPL_ASYNC_NOTIF;
2875 			opcode = CPL_ASYNC_NOTIF;
2876 			eop = 1;
2877                         rspq->async_notif++;
2878 			goto skip;
2879 		} else if  (flags & F_RSPD_IMM_DATA_VALID) {
2880 			struct mbuf *m = m_gethdr(M_NOWAIT, MT_DATA);
2881 
2882 			if (m == NULL) {
2883 		no_mem:
2884 				rspq->next_holdoff = NOMEM_INTR_DELAY;
2885 				budget_left--;
2886 				break;
2887 			}
2888 			if (mh->mh_head == NULL)
2889 				mh->mh_head = m;
2890                         else
2891 				mh->mh_tail->m_next = m;
2892 			mh->mh_tail = m;
2893 
2894 			get_imm_packet(adap, r, m);
2895 			mh->mh_head->m_pkthdr.len += m->m_len;
2896 			eop = 1;
2897 			rspq->imm_data++;
2898 		} else if (r->len_cq) {
2899 			int drop_thresh = eth ? SGE_RX_DROP_THRES : 0;
2900 
2901 			eop = get_packet(adap, drop_thresh, qs, mh, r);
2902 			if (eop) {
2903 				if (r->rss_hdr.hash_type && !adap->timestamp) {
2904 					M_HASHTYPE_SET(mh->mh_head,
2905 					    M_HASHTYPE_OPAQUE_HASH);
2906 					mh->mh_head->m_pkthdr.flowid = rss_hash;
2907 				}
2908 			}
2909 
2910 			ethpad = 2;
2911 		} else {
2912 			rspq->pure_rsps++;
2913 		}
2914 	skip:
2915 		if (flags & RSPD_CTRL_MASK) {
2916 			sleeping |= flags & RSPD_GTS_MASK;
2917 			handle_rsp_cntrl_info(qs, flags);
2918 		}
2919 
2920 		if (!eth && eop) {
2921 			rspq->offload_pkts++;
2922 #ifdef TCP_OFFLOAD
2923 			adap->cpl_handler[opcode](qs, r, mh->mh_head);
2924 #else
2925 			m_freem(mh->mh_head);
2926 #endif
2927 			mh->mh_head = NULL;
2928 		} else if (eth && eop) {
2929 			struct mbuf *m = mh->mh_head;
2930 
2931 			t3_rx_eth(adap, m, ethpad);
2932 
2933 			/*
2934 			 * The T304 sends incoming packets on any qset.  If LRO
2935 			 * is also enabled, we could end up sending packet up
2936 			 * lro_ctrl->ifp's input.  That is incorrect.
2937 			 *
2938 			 * The mbuf's rcvif was derived from the cpl header and
2939 			 * is accurate.  Skip LRO and just use that.
2940 			 */
2941 #if defined(INET6) || defined(INET)
2942 			skip_lro = __predict_false(qs->port->ifp != m->m_pkthdr.rcvif);
2943 
2944 			if (lro_enabled && lro_ctrl->lro_cnt && !skip_lro
2945 			    && (tcp_lro_rx(lro_ctrl, m, 0) == 0)
2946 			    ) {
2947 				/* successfully queue'd for LRO */
2948 			} else
2949 #endif
2950 			{
2951 				/*
2952 				 * LRO not enabled, packet unsuitable for LRO,
2953 				 * or unable to queue.  Pass it up right now in
2954 				 * either case.
2955 				 */
2956 				if_t ifp = m->m_pkthdr.rcvif;
2957 				if_input(ifp, m);
2958 			}
2959 			mh->mh_head = NULL;
2960 
2961 		}
2962 
2963 		r++;
2964 		if (__predict_false(++rspq->cidx == rspq->size)) {
2965 			rspq->cidx = 0;
2966 			rspq->gen ^= 1;
2967 			r = rspq->desc;
2968 		}
2969 
2970 		if (++rspq->credits >= 64) {
2971 			refill_rspq(adap, rspq, rspq->credits);
2972 			rspq->credits = 0;
2973 		}
2974 		__refill_fl_lt(adap, &qs->fl[0], 32);
2975 		__refill_fl_lt(adap, &qs->fl[1], 32);
2976 		--budget_left;
2977 	}
2978 
2979 #if defined(INET6) || defined(INET)
2980 	/* Flush LRO */
2981 	tcp_lro_flush_all(lro_ctrl);
2982 #endif
2983 
2984 	if (sleeping)
2985 		check_ring_db(adap, qs, sleeping);
2986 
2987 	mb();  /* commit Tx queue processed updates */
2988 	if (__predict_false(qs->txq_stopped > 1))
2989 		restart_tx(qs);
2990 
2991 	__refill_fl_lt(adap, &qs->fl[0], 512);
2992 	__refill_fl_lt(adap, &qs->fl[1], 512);
2993 	budget -= budget_left;
2994 	return (budget);
2995 }
2996 
2997 /*
2998  * A helper function that processes responses and issues GTS.
2999  */
3000 static __inline int
3001 process_responses_gts(adapter_t *adap, struct sge_rspq *rq)
3002 {
3003 	int work;
3004 	static int last_holdoff = 0;
3005 
3006 	work = process_responses(adap, rspq_to_qset(rq), -1);
3007 
3008 	if (cxgb_debug && (rq->next_holdoff != last_holdoff)) {
3009 		printf("next_holdoff=%d\n", rq->next_holdoff);
3010 		last_holdoff = rq->next_holdoff;
3011 	}
3012 	t3_write_reg(adap, A_SG_GTS, V_RSPQ(rq->cntxt_id) |
3013 	    V_NEWTIMER(rq->next_holdoff) | V_NEWINDEX(rq->cidx));
3014 
3015 	return (work);
3016 }
3017 
3018 #ifdef DEBUGNET
3019 int
3020 cxgb_debugnet_poll_rx(adapter_t *adap, struct sge_qset *qs)
3021 {
3022 
3023 	return (process_responses_gts(adap, &qs->rspq));
3024 }
3025 #endif
3026 
3027 /*
3028  * Interrupt handler for legacy INTx interrupts for T3B-based cards.
3029  * Handles data events from SGE response queues as well as error and other
3030  * async events as they all use the same interrupt pin.  We use one SGE
3031  * response queue per port in this mode and protect all response queues with
3032  * queue 0's lock.
3033  */
3034 void
3035 t3b_intr(void *data)
3036 {
3037 	uint32_t i, map;
3038 	adapter_t *adap = data;
3039 	struct sge_rspq *q0 = &adap->sge.qs[0].rspq;
3040 
3041 	t3_write_reg(adap, A_PL_CLI, 0);
3042 	map = t3_read_reg(adap, A_SG_DATA_INTR);
3043 
3044 	if (!map)
3045 		return;
3046 
3047 	if (__predict_false(map & F_ERRINTR)) {
3048 		t3_write_reg(adap, A_PL_INT_ENABLE0, 0);
3049 		(void) t3_read_reg(adap, A_PL_INT_ENABLE0);
3050 		taskqueue_enqueue(adap->tq, &adap->slow_intr_task);
3051 	}
3052 
3053 	mtx_lock(&q0->lock);
3054 	for_each_port(adap, i)
3055 	    if (map & (1 << i))
3056 			process_responses_gts(adap, &adap->sge.qs[i].rspq);
3057 	mtx_unlock(&q0->lock);
3058 }
3059 
3060 /*
3061  * The MSI interrupt handler.  This needs to handle data events from SGE
3062  * response queues as well as error and other async events as they all use
3063  * the same MSI vector.  We use one SGE response queue per port in this mode
3064  * and protect all response queues with queue 0's lock.
3065  */
3066 void
3067 t3_intr_msi(void *data)
3068 {
3069 	adapter_t *adap = data;
3070 	struct sge_rspq *q0 = &adap->sge.qs[0].rspq;
3071 	int i, new_packets = 0;
3072 
3073 	mtx_lock(&q0->lock);
3074 
3075 	for_each_port(adap, i)
3076 	    if (process_responses_gts(adap, &adap->sge.qs[i].rspq))
3077 		    new_packets = 1;
3078 	mtx_unlock(&q0->lock);
3079 	if (new_packets == 0) {
3080 		t3_write_reg(adap, A_PL_INT_ENABLE0, 0);
3081 		(void) t3_read_reg(adap, A_PL_INT_ENABLE0);
3082 		taskqueue_enqueue(adap->tq, &adap->slow_intr_task);
3083 	}
3084 }
3085 
3086 void
3087 t3_intr_msix(void *data)
3088 {
3089 	struct sge_qset *qs = data;
3090 	adapter_t *adap = qs->port->adapter;
3091 	struct sge_rspq *rspq = &qs->rspq;
3092 
3093 	if (process_responses_gts(adap, rspq) == 0)
3094 		rspq->unhandled_irqs++;
3095 }
3096 
3097 #define QDUMP_SBUF_SIZE		32 * 400
3098 static int
3099 t3_dump_rspq(SYSCTL_HANDLER_ARGS)
3100 {
3101 	struct sge_rspq *rspq;
3102 	struct sge_qset *qs;
3103 	int i, err, dump_end, idx;
3104 	struct sbuf *sb;
3105 	struct rsp_desc *rspd;
3106 	uint32_t data[4];
3107 
3108 	rspq = arg1;
3109 	qs = rspq_to_qset(rspq);
3110 	if (rspq->rspq_dump_count == 0)
3111 		return (0);
3112 	if (rspq->rspq_dump_count > RSPQ_Q_SIZE) {
3113 		log(LOG_WARNING,
3114 		    "dump count is too large %d\n", rspq->rspq_dump_count);
3115 		rspq->rspq_dump_count = 0;
3116 		return (EINVAL);
3117 	}
3118 	if (rspq->rspq_dump_start > (RSPQ_Q_SIZE-1)) {
3119 		log(LOG_WARNING,
3120 		    "dump start of %d is greater than queue size\n",
3121 		    rspq->rspq_dump_start);
3122 		rspq->rspq_dump_start = 0;
3123 		return (EINVAL);
3124 	}
3125 	err = t3_sge_read_rspq(qs->port->adapter, rspq->cntxt_id, data);
3126 	if (err)
3127 		return (err);
3128 	err = sysctl_wire_old_buffer(req, 0);
3129 	if (err)
3130 		return (err);
3131 	sb = sbuf_new_for_sysctl(NULL, NULL, QDUMP_SBUF_SIZE, req);
3132 
3133 	sbuf_printf(sb, " \n index=%u size=%u MSI-X/RspQ=%u intr enable=%u intr armed=%u\n",
3134 	    (data[0] & 0xffff), data[0] >> 16, ((data[2] >> 20) & 0x3f),
3135 	    ((data[2] >> 26) & 1), ((data[2] >> 27) & 1));
3136 	sbuf_printf(sb, " generation=%u CQ mode=%u FL threshold=%u\n",
3137 	    ((data[2] >> 28) & 1), ((data[2] >> 31) & 1), data[3]);
3138 
3139 	sbuf_printf(sb, " start=%d -> end=%d\n", rspq->rspq_dump_start,
3140 	    (rspq->rspq_dump_start + rspq->rspq_dump_count) & (RSPQ_Q_SIZE-1));
3141 
3142 	dump_end = rspq->rspq_dump_start + rspq->rspq_dump_count;
3143 	for (i = rspq->rspq_dump_start; i < dump_end; i++) {
3144 		idx = i & (RSPQ_Q_SIZE-1);
3145 
3146 		rspd = &rspq->desc[idx];
3147 		sbuf_printf(sb, "\tidx=%04d opcode=%02x cpu_idx=%x hash_type=%x cq_idx=%x\n",
3148 		    idx, rspd->rss_hdr.opcode, rspd->rss_hdr.cpu_idx,
3149 		    rspd->rss_hdr.hash_type, be16toh(rspd->rss_hdr.cq_idx));
3150 		sbuf_printf(sb, "\trss_hash_val=%x flags=%08x len_cq=%x intr_gen=%x\n",
3151 		    rspd->rss_hdr.rss_hash_val, be32toh(rspd->flags),
3152 		    be32toh(rspd->len_cq), rspd->intr_gen);
3153 	}
3154 
3155 	err = sbuf_finish(sb);
3156 	sbuf_delete(sb);
3157 	return (err);
3158 }
3159 
3160 static int
3161 t3_dump_txq_eth(SYSCTL_HANDLER_ARGS)
3162 {
3163 	struct sge_txq *txq;
3164 	struct sge_qset *qs;
3165 	int i, j, err, dump_end;
3166 	struct sbuf *sb;
3167 	struct tx_desc *txd;
3168 	uint32_t *WR, wr_hi, wr_lo, gen;
3169 	uint32_t data[4];
3170 
3171 	txq = arg1;
3172 	qs = txq_to_qset(txq, TXQ_ETH);
3173 	if (txq->txq_dump_count == 0) {
3174 		return (0);
3175 	}
3176 	if (txq->txq_dump_count > TX_ETH_Q_SIZE) {
3177 		log(LOG_WARNING,
3178 		    "dump count is too large %d\n", txq->txq_dump_count);
3179 		txq->txq_dump_count = 1;
3180 		return (EINVAL);
3181 	}
3182 	if (txq->txq_dump_start > (TX_ETH_Q_SIZE-1)) {
3183 		log(LOG_WARNING,
3184 		    "dump start of %d is greater than queue size\n",
3185 		    txq->txq_dump_start);
3186 		txq->txq_dump_start = 0;
3187 		return (EINVAL);
3188 	}
3189 	err = t3_sge_read_ecntxt(qs->port->adapter, qs->rspq.cntxt_id, data);
3190 	if (err)
3191 		return (err);
3192 	err = sysctl_wire_old_buffer(req, 0);
3193 	if (err)
3194 		return (err);
3195 	sb = sbuf_new_for_sysctl(NULL, NULL, QDUMP_SBUF_SIZE, req);
3196 
3197 	sbuf_printf(sb, " \n credits=%u GTS=%u index=%u size=%u rspq#=%u cmdq#=%u\n",
3198 	    (data[0] & 0x7fff), ((data[0] >> 15) & 1), (data[0] >> 16),
3199 	    (data[1] & 0xffff), ((data[3] >> 4) & 7), ((data[3] >> 7) & 1));
3200 	sbuf_printf(sb, " TUN=%u TOE=%u generation%u uP token=%u valid=%u\n",
3201 	    ((data[3] >> 8) & 1), ((data[3] >> 9) & 1), ((data[3] >> 10) & 1),
3202 	    ((data[3] >> 11) & 0xfffff), ((data[3] >> 31) & 1));
3203 	sbuf_printf(sb, " qid=%d start=%d -> end=%d\n", qs->idx,
3204 	    txq->txq_dump_start,
3205 	    (txq->txq_dump_start + txq->txq_dump_count) & (TX_ETH_Q_SIZE-1));
3206 
3207 	dump_end = txq->txq_dump_start + txq->txq_dump_count;
3208 	for (i = txq->txq_dump_start; i < dump_end; i++) {
3209 		txd = &txq->desc[i & (TX_ETH_Q_SIZE-1)];
3210 		WR = (uint32_t *)txd->flit;
3211 		wr_hi = ntohl(WR[0]);
3212 		wr_lo = ntohl(WR[1]);
3213 		gen = G_WR_GEN(wr_lo);
3214 
3215 		sbuf_printf(sb," wr_hi %08x wr_lo %08x gen %d\n",
3216 		    wr_hi, wr_lo, gen);
3217 		for (j = 2; j < 30; j += 4)
3218 			sbuf_printf(sb, "\t%08x %08x %08x %08x \n",
3219 			    WR[j], WR[j + 1], WR[j + 2], WR[j + 3]);
3220 
3221 	}
3222 	err = sbuf_finish(sb);
3223 	sbuf_delete(sb);
3224 	return (err);
3225 }
3226 
3227 static int
3228 t3_dump_txq_ctrl(SYSCTL_HANDLER_ARGS)
3229 {
3230 	struct sge_txq *txq;
3231 	struct sge_qset *qs;
3232 	int i, j, err, dump_end;
3233 	struct sbuf *sb;
3234 	struct tx_desc *txd;
3235 	uint32_t *WR, wr_hi, wr_lo, gen;
3236 
3237 	txq = arg1;
3238 	qs = txq_to_qset(txq, TXQ_CTRL);
3239 	if (txq->txq_dump_count == 0) {
3240 		return (0);
3241 	}
3242 	if (txq->txq_dump_count > 256) {
3243 		log(LOG_WARNING,
3244 		    "dump count is too large %d\n", txq->txq_dump_count);
3245 		txq->txq_dump_count = 1;
3246 		return (EINVAL);
3247 	}
3248 	if (txq->txq_dump_start > 255) {
3249 		log(LOG_WARNING,
3250 		    "dump start of %d is greater than queue size\n",
3251 		    txq->txq_dump_start);
3252 		txq->txq_dump_start = 0;
3253 		return (EINVAL);
3254 	}
3255 
3256 	err = sysctl_wire_old_buffer(req, 0);
3257 	if (err != 0)
3258 		return (err);
3259 	sb = sbuf_new_for_sysctl(NULL, NULL, QDUMP_SBUF_SIZE, req);
3260 	sbuf_printf(sb, " qid=%d start=%d -> end=%d\n", qs->idx,
3261 	    txq->txq_dump_start,
3262 	    (txq->txq_dump_start + txq->txq_dump_count) & 255);
3263 
3264 	dump_end = txq->txq_dump_start + txq->txq_dump_count;
3265 	for (i = txq->txq_dump_start; i < dump_end; i++) {
3266 		txd = &txq->desc[i & (255)];
3267 		WR = (uint32_t *)txd->flit;
3268 		wr_hi = ntohl(WR[0]);
3269 		wr_lo = ntohl(WR[1]);
3270 		gen = G_WR_GEN(wr_lo);
3271 
3272 		sbuf_printf(sb," wr_hi %08x wr_lo %08x gen %d\n",
3273 		    wr_hi, wr_lo, gen);
3274 		for (j = 2; j < 30; j += 4)
3275 			sbuf_printf(sb, "\t%08x %08x %08x %08x \n",
3276 			    WR[j], WR[j + 1], WR[j + 2], WR[j + 3]);
3277 
3278 	}
3279 	err = sbuf_finish(sb);
3280 	sbuf_delete(sb);
3281 	return (err);
3282 }
3283 
3284 static int
3285 t3_set_coalesce_usecs(SYSCTL_HANDLER_ARGS)
3286 {
3287 	adapter_t *sc = arg1;
3288 	struct qset_params *qsp = &sc->params.sge.qset[0];
3289 	int coalesce_usecs;
3290 	struct sge_qset *qs;
3291 	int i, j, err, nqsets = 0;
3292 	struct mtx *lock;
3293 
3294 	if ((sc->flags & FULL_INIT_DONE) == 0)
3295 		return (ENXIO);
3296 
3297 	coalesce_usecs = qsp->coalesce_usecs;
3298         err = sysctl_handle_int(oidp, &coalesce_usecs, arg2, req);
3299 
3300 	if (err != 0) {
3301 		return (err);
3302 	}
3303 	if (coalesce_usecs == qsp->coalesce_usecs)
3304 		return (0);
3305 
3306 	for (i = 0; i < sc->params.nports; i++)
3307 		for (j = 0; j < sc->port[i].nqsets; j++)
3308 			nqsets++;
3309 
3310 	coalesce_usecs = max(1, coalesce_usecs);
3311 
3312 	for (i = 0; i < nqsets; i++) {
3313 		qs = &sc->sge.qs[i];
3314 		qsp = &sc->params.sge.qset[i];
3315 		qsp->coalesce_usecs = coalesce_usecs;
3316 
3317 		lock = (sc->flags & USING_MSIX) ? &qs->rspq.lock :
3318 			    &sc->sge.qs[0].rspq.lock;
3319 
3320 		mtx_lock(lock);
3321 		t3_update_qset_coalesce(qs, qsp);
3322 		t3_write_reg(sc, A_SG_GTS, V_RSPQ(qs->rspq.cntxt_id) |
3323 		    V_NEWTIMER(qs->rspq.holdoff_tmr));
3324 		mtx_unlock(lock);
3325 	}
3326 
3327 	return (0);
3328 }
3329 
3330 static int
3331 t3_pkt_timestamp(SYSCTL_HANDLER_ARGS)
3332 {
3333 	adapter_t *sc = arg1;
3334 	int rc, timestamp;
3335 
3336 	if ((sc->flags & FULL_INIT_DONE) == 0)
3337 		return (ENXIO);
3338 
3339 	timestamp = sc->timestamp;
3340 	rc = sysctl_handle_int(oidp, &timestamp, arg2, req);
3341 
3342 	if (rc != 0)
3343 		return (rc);
3344 
3345 	if (timestamp != sc->timestamp) {
3346 		t3_set_reg_field(sc, A_TP_PC_CONFIG2, F_ENABLERXPKTTMSTPRSS,
3347 		    timestamp ? F_ENABLERXPKTTMSTPRSS : 0);
3348 		sc->timestamp = timestamp;
3349 	}
3350 
3351 	return (0);
3352 }
3353 
3354 void
3355 t3_add_attach_sysctls(adapter_t *sc)
3356 {
3357 	struct sysctl_ctx_list *ctx;
3358 	struct sysctl_oid_list *children;
3359 
3360 	ctx = device_get_sysctl_ctx(sc->dev);
3361 	children = SYSCTL_CHILDREN(device_get_sysctl_tree(sc->dev));
3362 
3363 	/* random information */
3364 	SYSCTL_ADD_STRING(ctx, children, OID_AUTO,
3365 	    "firmware_version",
3366 	    CTLFLAG_RD, sc->fw_version,
3367 	    0, "firmware version");
3368 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO,
3369 	    "hw_revision",
3370 	    CTLFLAG_RD, &sc->params.rev,
3371 	    0, "chip model");
3372 	SYSCTL_ADD_STRING(ctx, children, OID_AUTO,
3373 	    "port_types",
3374 	    CTLFLAG_RD, sc->port_types,
3375 	    0, "type of ports");
3376 	SYSCTL_ADD_INT(ctx, children, OID_AUTO,
3377 	    "enable_debug",
3378 	    CTLFLAG_RW, &cxgb_debug,
3379 	    0, "enable verbose debugging output");
3380 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tunq_coalesce",
3381 	    CTLFLAG_RD, &sc->tunq_coalesce,
3382 	    "#tunneled packets freed");
3383 	SYSCTL_ADD_INT(ctx, children, OID_AUTO,
3384 	    "txq_overrun",
3385 	    CTLFLAG_RD, &txq_fills,
3386 	    0, "#times txq overrun");
3387 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO,
3388 	    "core_clock",
3389 	    CTLFLAG_RD, &sc->params.vpd.cclk,
3390 	    0, "core clock frequency (in KHz)");
3391 }
3392 
3393 
3394 static const char *rspq_name = "rspq";
3395 static const char *txq_names[] =
3396 {
3397 	"txq_eth",
3398 	"txq_ofld",
3399 	"txq_ctrl"
3400 };
3401 
3402 static int
3403 sysctl_handle_macstat(SYSCTL_HANDLER_ARGS)
3404 {
3405 	struct port_info *p = arg1;
3406 	uint64_t *parg;
3407 
3408 	if (!p)
3409 		return (EINVAL);
3410 
3411 	cxgb_refresh_stats(p);
3412 	parg = (uint64_t *) ((uint8_t *)&p->mac.stats + arg2);
3413 
3414 	return (sysctl_handle_64(oidp, parg, 0, req));
3415 }
3416 
3417 void
3418 t3_add_configured_sysctls(adapter_t *sc)
3419 {
3420 	struct sysctl_ctx_list *ctx;
3421 	struct sysctl_oid_list *children;
3422 	int i, j;
3423 
3424 	ctx = device_get_sysctl_ctx(sc->dev);
3425 	children = SYSCTL_CHILDREN(device_get_sysctl_tree(sc->dev));
3426 
3427 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO,
3428 	    "intr_coal",
3429 	    CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, sc,
3430 	    0, t3_set_coalesce_usecs,
3431 	    "I", "interrupt coalescing timer (us)");
3432 
3433 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO,
3434 	    "pkt_timestamp",
3435 	    CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, sc,
3436 	    0, t3_pkt_timestamp,
3437 	    "I", "provide packet timestamp instead of connection hash");
3438 
3439 	for (i = 0; i < sc->params.nports; i++) {
3440 		struct port_info *pi = &sc->port[i];
3441 		struct sysctl_oid *poid;
3442 		struct sysctl_oid_list *poidlist;
3443 		struct mac_stats *mstats = &pi->mac.stats;
3444 
3445 		snprintf(pi->namebuf, PORT_NAME_LEN, "port%d", i);
3446 		poid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO,
3447 		    pi->namebuf, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
3448 		    "port statistics");
3449 		poidlist = SYSCTL_CHILDREN(poid);
3450 		SYSCTL_ADD_UINT(ctx, poidlist, OID_AUTO,
3451 		    "nqsets", CTLFLAG_RD, &pi->nqsets,
3452 		    0, "#queue sets");
3453 
3454 		for (j = 0; j < pi->nqsets; j++) {
3455 			struct sge_qset *qs = &sc->sge.qs[pi->first_qset + j];
3456 			struct sysctl_oid *qspoid, *rspqpoid, *txqpoid,
3457 					  *ctrlqpoid, *lropoid;
3458 			struct sysctl_oid_list *qspoidlist, *rspqpoidlist,
3459 					       *txqpoidlist, *ctrlqpoidlist,
3460 					       *lropoidlist;
3461 			struct sge_txq *txq = &qs->txq[TXQ_ETH];
3462 
3463 			snprintf(qs->namebuf, QS_NAME_LEN, "qs%d", j);
3464 
3465 			qspoid = SYSCTL_ADD_NODE(ctx, poidlist, OID_AUTO,
3466 			    qs->namebuf, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
3467 			    "qset statistics");
3468 			qspoidlist = SYSCTL_CHILDREN(qspoid);
3469 
3470 			SYSCTL_ADD_UINT(ctx, qspoidlist, OID_AUTO, "fl0_empty",
3471 					CTLFLAG_RD, &qs->fl[0].empty, 0,
3472 					"freelist #0 empty");
3473 			SYSCTL_ADD_UINT(ctx, qspoidlist, OID_AUTO, "fl1_empty",
3474 					CTLFLAG_RD, &qs->fl[1].empty, 0,
3475 					"freelist #1 empty");
3476 
3477 			rspqpoid = SYSCTL_ADD_NODE(ctx, qspoidlist, OID_AUTO,
3478 			    rspq_name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
3479 			    "rspq statistics");
3480 			rspqpoidlist = SYSCTL_CHILDREN(rspqpoid);
3481 
3482 			txqpoid = SYSCTL_ADD_NODE(ctx, qspoidlist, OID_AUTO,
3483 			    txq_names[0], CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
3484 			    "txq statistics");
3485 			txqpoidlist = SYSCTL_CHILDREN(txqpoid);
3486 
3487 			ctrlqpoid = SYSCTL_ADD_NODE(ctx, qspoidlist, OID_AUTO,
3488 			    txq_names[2], CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
3489 			    "ctrlq statistics");
3490 			ctrlqpoidlist = SYSCTL_CHILDREN(ctrlqpoid);
3491 
3492 			lropoid = SYSCTL_ADD_NODE(ctx, qspoidlist, OID_AUTO,
3493 			    "lro_stats", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
3494 			    "LRO statistics");
3495 			lropoidlist = SYSCTL_CHILDREN(lropoid);
3496 
3497 			SYSCTL_ADD_UINT(ctx, rspqpoidlist, OID_AUTO, "size",
3498 			    CTLFLAG_RD, &qs->rspq.size,
3499 			    0, "#entries in response queue");
3500 			SYSCTL_ADD_UINT(ctx, rspqpoidlist, OID_AUTO, "cidx",
3501 			    CTLFLAG_RD, &qs->rspq.cidx,
3502 			    0, "consumer index");
3503 			SYSCTL_ADD_UINT(ctx, rspqpoidlist, OID_AUTO, "credits",
3504 			    CTLFLAG_RD, &qs->rspq.credits,
3505 			    0, "#credits");
3506 			SYSCTL_ADD_UINT(ctx, rspqpoidlist, OID_AUTO, "starved",
3507 			    CTLFLAG_RD, &qs->rspq.starved,
3508 			    0, "#times starved");
3509 			SYSCTL_ADD_UAUTO(ctx, rspqpoidlist, OID_AUTO, "phys_addr",
3510 			    CTLFLAG_RD, &qs->rspq.phys_addr,
3511 			    "physical_address_of the queue");
3512 			SYSCTL_ADD_UINT(ctx, rspqpoidlist, OID_AUTO, "dump_start",
3513 			    CTLFLAG_RW, &qs->rspq.rspq_dump_start,
3514 			    0, "start rspq dump entry");
3515 			SYSCTL_ADD_UINT(ctx, rspqpoidlist, OID_AUTO, "dump_count",
3516 			    CTLFLAG_RW, &qs->rspq.rspq_dump_count,
3517 			    0, "#rspq entries to dump");
3518 			SYSCTL_ADD_PROC(ctx, rspqpoidlist, OID_AUTO, "qdump",
3519 			    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
3520 			    &qs->rspq, 0, t3_dump_rspq, "A",
3521 			    "dump of the response queue");
3522 
3523 			SYSCTL_ADD_UQUAD(ctx, txqpoidlist, OID_AUTO, "dropped",
3524 			    CTLFLAG_RD, &qs->txq[TXQ_ETH].txq_mr->br_drops,
3525 			    "#tunneled packets dropped");
3526 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "sendqlen",
3527 			    CTLFLAG_RD, &qs->txq[TXQ_ETH].sendq.mq_len,
3528 			    0, "#tunneled packets waiting to be sent");
3529 #if 0
3530 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "queue_pidx",
3531 			    CTLFLAG_RD, (uint32_t *)(uintptr_t)&qs->txq[TXQ_ETH].txq_mr.br_prod,
3532 			    0, "#tunneled packets queue producer index");
3533 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "queue_cidx",
3534 			    CTLFLAG_RD, (uint32_t *)(uintptr_t)&qs->txq[TXQ_ETH].txq_mr.br_cons,
3535 			    0, "#tunneled packets queue consumer index");
3536 #endif
3537 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "processed",
3538 			    CTLFLAG_RD, &qs->txq[TXQ_ETH].processed,
3539 			    0, "#tunneled packets processed by the card");
3540 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "cleaned",
3541 			    CTLFLAG_RD, &txq->cleaned,
3542 			    0, "#tunneled packets cleaned");
3543 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "in_use",
3544 			    CTLFLAG_RD, &txq->in_use,
3545 			    0, "#tunneled packet slots in use");
3546 			SYSCTL_ADD_UQUAD(ctx, txqpoidlist, OID_AUTO, "frees",
3547 			    CTLFLAG_RD, &txq->txq_frees,
3548 			    "#tunneled packets freed");
3549 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "skipped",
3550 			    CTLFLAG_RD, &txq->txq_skipped,
3551 			    0, "#tunneled packet descriptors skipped");
3552 			SYSCTL_ADD_UQUAD(ctx, txqpoidlist, OID_AUTO, "coalesced",
3553 			    CTLFLAG_RD, &txq->txq_coalesced,
3554 			    "#tunneled packets coalesced");
3555 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "enqueued",
3556 			    CTLFLAG_RD, &txq->txq_enqueued,
3557 			    0, "#tunneled packets enqueued to hardware");
3558 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "stopped_flags",
3559 			    CTLFLAG_RD, &qs->txq_stopped,
3560 			    0, "tx queues stopped");
3561 			SYSCTL_ADD_UAUTO(ctx, txqpoidlist, OID_AUTO, "phys_addr",
3562 			    CTLFLAG_RD, &txq->phys_addr,
3563 			    "physical_address_of the queue");
3564 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "qgen",
3565 			    CTLFLAG_RW, &qs->txq[TXQ_ETH].gen,
3566 			    0, "txq generation");
3567 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "hw_cidx",
3568 			    CTLFLAG_RD, &txq->cidx,
3569 			    0, "hardware queue cidx");
3570 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "hw_pidx",
3571 			    CTLFLAG_RD, &txq->pidx,
3572 			    0, "hardware queue pidx");
3573 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "dump_start",
3574 			    CTLFLAG_RW, &qs->txq[TXQ_ETH].txq_dump_start,
3575 			    0, "txq start idx for dump");
3576 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "dump_count",
3577 			    CTLFLAG_RW, &qs->txq[TXQ_ETH].txq_dump_count,
3578 			    0, "txq #entries to dump");
3579 			SYSCTL_ADD_PROC(ctx, txqpoidlist, OID_AUTO, "qdump",
3580 			    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
3581 			    &qs->txq[TXQ_ETH], 0, t3_dump_txq_eth, "A",
3582 			    "dump of the transmit queue");
3583 
3584 			SYSCTL_ADD_UINT(ctx, ctrlqpoidlist, OID_AUTO, "dump_start",
3585 			    CTLFLAG_RW, &qs->txq[TXQ_CTRL].txq_dump_start,
3586 			    0, "ctrlq start idx for dump");
3587 			SYSCTL_ADD_UINT(ctx, ctrlqpoidlist, OID_AUTO, "dump_count",
3588 			    CTLFLAG_RW, &qs->txq[TXQ_CTRL].txq_dump_count,
3589 			    0, "ctrl #entries to dump");
3590 			SYSCTL_ADD_PROC(ctx, ctrlqpoidlist, OID_AUTO, "qdump",
3591 			    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
3592 			    &qs->txq[TXQ_CTRL], 0, t3_dump_txq_ctrl, "A",
3593 			    "dump of the transmit queue");
3594 
3595 			SYSCTL_ADD_U64(ctx, lropoidlist, OID_AUTO, "lro_queued",
3596 			    CTLFLAG_RD, &qs->lro.ctrl.lro_queued, 0, NULL);
3597 			SYSCTL_ADD_U64(ctx, lropoidlist, OID_AUTO, "lro_flushed",
3598 			    CTLFLAG_RD, &qs->lro.ctrl.lro_flushed, 0, NULL);
3599 			SYSCTL_ADD_U64(ctx, lropoidlist, OID_AUTO, "lro_bad_csum",
3600 			    CTLFLAG_RD, &qs->lro.ctrl.lro_bad_csum, 0, NULL);
3601 			SYSCTL_ADD_INT(ctx, lropoidlist, OID_AUTO, "lro_cnt",
3602 			    CTLFLAG_RD, &qs->lro.ctrl.lro_cnt, 0, NULL);
3603 		}
3604 
3605 		/* Now add a node for mac stats. */
3606 		poid = SYSCTL_ADD_NODE(ctx, poidlist, OID_AUTO, "mac_stats",
3607 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "MAC statistics");
3608 		poidlist = SYSCTL_CHILDREN(poid);
3609 
3610 		/*
3611 		 * We (ab)use the length argument (arg2) to pass on the offset
3612 		 * of the data that we are interested in.  This is only required
3613 		 * for the quad counters that are updated from the hardware (we
3614 		 * make sure that we return the latest value).
3615 		 * sysctl_handle_macstat first updates *all* the counters from
3616 		 * the hardware, and then returns the latest value of the
3617 		 * requested counter.  Best would be to update only the
3618 		 * requested counter from hardware, but t3_mac_update_stats()
3619 		 * hides all the register details and we don't want to dive into
3620 		 * all that here.
3621 		 */
3622 #define CXGB_SYSCTL_ADD_QUAD(a)	SYSCTL_ADD_OID(ctx, poidlist, OID_AUTO, #a, \
3623     CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_NEEDGIANT, pi, \
3624     offsetof(struct mac_stats, a), sysctl_handle_macstat, "QU", 0)
3625 		CXGB_SYSCTL_ADD_QUAD(tx_octets);
3626 		CXGB_SYSCTL_ADD_QUAD(tx_octets_bad);
3627 		CXGB_SYSCTL_ADD_QUAD(tx_frames);
3628 		CXGB_SYSCTL_ADD_QUAD(tx_mcast_frames);
3629 		CXGB_SYSCTL_ADD_QUAD(tx_bcast_frames);
3630 		CXGB_SYSCTL_ADD_QUAD(tx_pause);
3631 		CXGB_SYSCTL_ADD_QUAD(tx_deferred);
3632 		CXGB_SYSCTL_ADD_QUAD(tx_late_collisions);
3633 		CXGB_SYSCTL_ADD_QUAD(tx_total_collisions);
3634 		CXGB_SYSCTL_ADD_QUAD(tx_excess_collisions);
3635 		CXGB_SYSCTL_ADD_QUAD(tx_underrun);
3636 		CXGB_SYSCTL_ADD_QUAD(tx_len_errs);
3637 		CXGB_SYSCTL_ADD_QUAD(tx_mac_internal_errs);
3638 		CXGB_SYSCTL_ADD_QUAD(tx_excess_deferral);
3639 		CXGB_SYSCTL_ADD_QUAD(tx_fcs_errs);
3640 		CXGB_SYSCTL_ADD_QUAD(tx_frames_64);
3641 		CXGB_SYSCTL_ADD_QUAD(tx_frames_65_127);
3642 		CXGB_SYSCTL_ADD_QUAD(tx_frames_128_255);
3643 		CXGB_SYSCTL_ADD_QUAD(tx_frames_256_511);
3644 		CXGB_SYSCTL_ADD_QUAD(tx_frames_512_1023);
3645 		CXGB_SYSCTL_ADD_QUAD(tx_frames_1024_1518);
3646 		CXGB_SYSCTL_ADD_QUAD(tx_frames_1519_max);
3647 		CXGB_SYSCTL_ADD_QUAD(rx_octets);
3648 		CXGB_SYSCTL_ADD_QUAD(rx_octets_bad);
3649 		CXGB_SYSCTL_ADD_QUAD(rx_frames);
3650 		CXGB_SYSCTL_ADD_QUAD(rx_mcast_frames);
3651 		CXGB_SYSCTL_ADD_QUAD(rx_bcast_frames);
3652 		CXGB_SYSCTL_ADD_QUAD(rx_pause);
3653 		CXGB_SYSCTL_ADD_QUAD(rx_fcs_errs);
3654 		CXGB_SYSCTL_ADD_QUAD(rx_align_errs);
3655 		CXGB_SYSCTL_ADD_QUAD(rx_symbol_errs);
3656 		CXGB_SYSCTL_ADD_QUAD(rx_data_errs);
3657 		CXGB_SYSCTL_ADD_QUAD(rx_sequence_errs);
3658 		CXGB_SYSCTL_ADD_QUAD(rx_runt);
3659 		CXGB_SYSCTL_ADD_QUAD(rx_jabber);
3660 		CXGB_SYSCTL_ADD_QUAD(rx_short);
3661 		CXGB_SYSCTL_ADD_QUAD(rx_too_long);
3662 		CXGB_SYSCTL_ADD_QUAD(rx_mac_internal_errs);
3663 		CXGB_SYSCTL_ADD_QUAD(rx_cong_drops);
3664 		CXGB_SYSCTL_ADD_QUAD(rx_frames_64);
3665 		CXGB_SYSCTL_ADD_QUAD(rx_frames_65_127);
3666 		CXGB_SYSCTL_ADD_QUAD(rx_frames_128_255);
3667 		CXGB_SYSCTL_ADD_QUAD(rx_frames_256_511);
3668 		CXGB_SYSCTL_ADD_QUAD(rx_frames_512_1023);
3669 		CXGB_SYSCTL_ADD_QUAD(rx_frames_1024_1518);
3670 		CXGB_SYSCTL_ADD_QUAD(rx_frames_1519_max);
3671 #undef CXGB_SYSCTL_ADD_QUAD
3672 
3673 #define CXGB_SYSCTL_ADD_ULONG(a) SYSCTL_ADD_ULONG(ctx, poidlist, OID_AUTO, #a, \
3674     CTLFLAG_RD, &mstats->a, 0)
3675 		CXGB_SYSCTL_ADD_ULONG(tx_fifo_parity_err);
3676 		CXGB_SYSCTL_ADD_ULONG(rx_fifo_parity_err);
3677 		CXGB_SYSCTL_ADD_ULONG(tx_fifo_urun);
3678 		CXGB_SYSCTL_ADD_ULONG(rx_fifo_ovfl);
3679 		CXGB_SYSCTL_ADD_ULONG(serdes_signal_loss);
3680 		CXGB_SYSCTL_ADD_ULONG(xaui_pcs_ctc_err);
3681 		CXGB_SYSCTL_ADD_ULONG(xaui_pcs_align_change);
3682 		CXGB_SYSCTL_ADD_ULONG(num_toggled);
3683 		CXGB_SYSCTL_ADD_ULONG(num_resets);
3684 		CXGB_SYSCTL_ADD_ULONG(link_faults);
3685 #undef CXGB_SYSCTL_ADD_ULONG
3686 	}
3687 }
3688 
3689 /**
3690  *	t3_get_desc - dump an SGE descriptor for debugging purposes
3691  *	@qs: the queue set
3692  *	@qnum: identifies the specific queue (0..2: Tx, 3:response, 4..5: Rx)
3693  *	@idx: the descriptor index in the queue
3694  *	@data: where to dump the descriptor contents
3695  *
3696  *	Dumps the contents of a HW descriptor of an SGE queue.  Returns the
3697  *	size of the descriptor.
3698  */
3699 int
3700 t3_get_desc(const struct sge_qset *qs, unsigned int qnum, unsigned int idx,
3701 		unsigned char *data)
3702 {
3703 	if (qnum >= 6)
3704 		return (EINVAL);
3705 
3706 	if (qnum < 3) {
3707 		if (!qs->txq[qnum].desc || idx >= qs->txq[qnum].size)
3708 			return -EINVAL;
3709 		memcpy(data, &qs->txq[qnum].desc[idx], sizeof(struct tx_desc));
3710 		return sizeof(struct tx_desc);
3711 	}
3712 
3713 	if (qnum == 3) {
3714 		if (!qs->rspq.desc || idx >= qs->rspq.size)
3715 			return (EINVAL);
3716 		memcpy(data, &qs->rspq.desc[idx], sizeof(struct rsp_desc));
3717 		return sizeof(struct rsp_desc);
3718 	}
3719 
3720 	qnum -= 4;
3721 	if (!qs->fl[qnum].desc || idx >= qs->fl[qnum].size)
3722 		return (EINVAL);
3723 	memcpy(data, &qs->fl[qnum].desc[idx], sizeof(struct rx_desc));
3724 	return sizeof(struct rx_desc);
3725 }
3726