xref: /freebsd/sys/dev/cxgb/cxgb_sge.c (revision 48c5129f93c5eb5419c87b08e4677d51513f1dc0)
1 /**************************************************************************
2 
3 Copyright (c) 2007-2009, Chelsio Inc.
4 All rights reserved.
5 
6 Redistribution and use in source and binary forms, with or without
7 modification, are permitted provided that the following conditions are met:
8 
9  1. Redistributions of source code must retain the above copyright notice,
10     this list of conditions and the following disclaimer.
11 
12  2. Neither the name of the Chelsio Corporation nor the names of its
13     contributors may be used to endorse or promote products derived from
14     this software without specific prior written permission.
15 
16 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
17 AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19 ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
20 LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
21 CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
22 SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
23 INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
24 CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
25 ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
26 POSSIBILITY OF SUCH DAMAGE.
27 
28 ***************************************************************************/
29 
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include "opt_inet6.h"
34 #include "opt_inet.h"
35 
36 #include <sys/param.h>
37 #include <sys/systm.h>
38 #include <sys/kernel.h>
39 #include <sys/module.h>
40 #include <sys/bus.h>
41 #include <sys/conf.h>
42 #include <machine/bus.h>
43 #include <machine/resource.h>
44 #include <sys/bus_dma.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/bpf.h>
60 #include <net/ethernet.h>
61 #include <net/if.h>
62 #include <net/if_vlan_var.h>
63 
64 #include <netinet/in_systm.h>
65 #include <netinet/in.h>
66 #include <netinet/ip.h>
67 #include <netinet/ip6.h>
68 #include <netinet/tcp.h>
69 
70 #include <dev/pci/pcireg.h>
71 #include <dev/pci/pcivar.h>
72 
73 #include <vm/vm.h>
74 #include <vm/pmap.h>
75 
76 #include <cxgb_include.h>
77 #include <sys/mvec.h>
78 
79 int	txq_fills = 0;
80 int	multiq_tx_enable = 1;
81 
82 #ifdef TCP_OFFLOAD
83 CTASSERT(NUM_CPL_HANDLERS >= NUM_CPL_CMDS);
84 #endif
85 
86 extern struct sysctl_oid_list sysctl__hw_cxgb_children;
87 int cxgb_txq_buf_ring_size = TX_ETH_Q_SIZE;
88 TUNABLE_INT("hw.cxgb.txq_mr_size", &cxgb_txq_buf_ring_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 TUNABLE_INT("hw.cxgb.tx_coalesce_force", &cxgb_tx_coalesce_force);
94 SYSCTL_INT(_hw_cxgb, OID_AUTO, tx_coalesce_force, CTLFLAG_RW,
95     &cxgb_tx_coalesce_force, 0,
96     "coalesce small packets into a single work request regardless of ring state");
97 
98 #define	COALESCE_START_DEFAULT		TX_ETH_Q_SIZE>>1
99 #define	COALESCE_START_MAX		(TX_ETH_Q_SIZE-(TX_ETH_Q_SIZE>>3))
100 #define	COALESCE_STOP_DEFAULT		TX_ETH_Q_SIZE>>2
101 #define	COALESCE_STOP_MIN		TX_ETH_Q_SIZE>>5
102 #define	TX_RECLAIM_DEFAULT		TX_ETH_Q_SIZE>>5
103 #define	TX_RECLAIM_MAX			TX_ETH_Q_SIZE>>2
104 #define	TX_RECLAIM_MIN			TX_ETH_Q_SIZE>>6
105 
106 
107 static int cxgb_tx_coalesce_enable_start = COALESCE_START_DEFAULT;
108 TUNABLE_INT("hw.cxgb.tx_coalesce_enable_start",
109     &cxgb_tx_coalesce_enable_start);
110 SYSCTL_INT(_hw_cxgb, OID_AUTO, tx_coalesce_enable_start, CTLFLAG_RW,
111     &cxgb_tx_coalesce_enable_start, 0,
112     "coalesce enable threshold");
113 static int cxgb_tx_coalesce_enable_stop = COALESCE_STOP_DEFAULT;
114 TUNABLE_INT("hw.cxgb.tx_coalesce_enable_stop", &cxgb_tx_coalesce_enable_stop);
115 SYSCTL_INT(_hw_cxgb, OID_AUTO, tx_coalesce_enable_stop, CTLFLAG_RW,
116     &cxgb_tx_coalesce_enable_stop, 0,
117     "coalesce disable threshold");
118 static int cxgb_tx_reclaim_threshold = TX_RECLAIM_DEFAULT;
119 TUNABLE_INT("hw.cxgb.tx_reclaim_threshold", &cxgb_tx_reclaim_threshold);
120 SYSCTL_INT(_hw_cxgb, OID_AUTO, tx_reclaim_threshold, CTLFLAG_RW,
121     &cxgb_tx_reclaim_threshold, 0,
122     "tx cleaning minimum threshold");
123 
124 /*
125  * XXX don't re-enable this until TOE stops assuming
126  * we have an m_ext
127  */
128 static int recycle_enable = 0;
129 
130 extern int cxgb_use_16k_clusters;
131 extern int nmbjumbop;
132 extern int nmbjumbo9;
133 extern int nmbjumbo16;
134 
135 #define USE_GTS 0
136 
137 #define SGE_RX_SM_BUF_SIZE	1536
138 #define SGE_RX_DROP_THRES	16
139 #define SGE_RX_COPY_THRES	128
140 
141 /*
142  * Period of the Tx buffer reclaim timer.  This timer does not need to run
143  * frequently as Tx buffers are usually reclaimed by new Tx packets.
144  */
145 #define TX_RECLAIM_PERIOD       (hz >> 1)
146 
147 /*
148  * Values for sge_txq.flags
149  */
150 enum {
151 	TXQ_RUNNING	= 1 << 0,  /* fetch engine is running */
152 	TXQ_LAST_PKT_DB = 1 << 1,  /* last packet rang the doorbell */
153 };
154 
155 struct tx_desc {
156 	uint64_t	flit[TX_DESC_FLITS];
157 } __packed;
158 
159 struct rx_desc {
160 	uint32_t	addr_lo;
161 	uint32_t	len_gen;
162 	uint32_t	gen2;
163 	uint32_t	addr_hi;
164 } __packed;
165 
166 struct rsp_desc {               /* response queue descriptor */
167 	struct rss_header	rss_hdr;
168 	uint32_t		flags;
169 	uint32_t		len_cq;
170 	uint8_t			imm_data[47];
171 	uint8_t			intr_gen;
172 } __packed;
173 
174 #define RX_SW_DESC_MAP_CREATED	(1 << 0)
175 #define TX_SW_DESC_MAP_CREATED	(1 << 1)
176 #define RX_SW_DESC_INUSE        (1 << 3)
177 #define TX_SW_DESC_MAPPED       (1 << 4)
178 
179 #define RSPQ_NSOP_NEOP           G_RSPD_SOP_EOP(0)
180 #define RSPQ_EOP                 G_RSPD_SOP_EOP(F_RSPD_EOP)
181 #define RSPQ_SOP                 G_RSPD_SOP_EOP(F_RSPD_SOP)
182 #define RSPQ_SOP_EOP             G_RSPD_SOP_EOP(F_RSPD_SOP|F_RSPD_EOP)
183 
184 struct tx_sw_desc {                /* SW state per Tx descriptor */
185 	struct mbuf	*m;
186 	bus_dmamap_t	map;
187 	int		flags;
188 };
189 
190 struct rx_sw_desc {                /* SW state per Rx descriptor */
191 	caddr_t		rxsd_cl;
192 	struct mbuf	*m;
193 	bus_dmamap_t	map;
194 	int		flags;
195 };
196 
197 struct txq_state {
198 	unsigned int	compl;
199 	unsigned int	gen;
200 	unsigned int	pidx;
201 };
202 
203 struct refill_fl_cb_arg {
204 	int               error;
205 	bus_dma_segment_t seg;
206 	int               nseg;
207 };
208 
209 
210 /*
211  * Maps a number of flits to the number of Tx descriptors that can hold them.
212  * The formula is
213  *
214  * desc = 1 + (flits - 2) / (WR_FLITS - 1).
215  *
216  * HW allows up to 4 descriptors to be combined into a WR.
217  */
218 static uint8_t flit_desc_map[] = {
219 	0,
220 #if SGE_NUM_GENBITS == 1
221 	1, 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, 2,
223 	3, 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, 4
225 #elif SGE_NUM_GENBITS == 2
226 	1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
227 	2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
228 	3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
229 	4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4,
230 #else
231 # error "SGE_NUM_GENBITS must be 1 or 2"
232 #endif
233 };
234 
235 #define	TXQ_LOCK_ASSERT(qs)	mtx_assert(&(qs)->lock, MA_OWNED)
236 #define	TXQ_TRYLOCK(qs)		mtx_trylock(&(qs)->lock)
237 #define	TXQ_LOCK(qs)		mtx_lock(&(qs)->lock)
238 #define	TXQ_UNLOCK(qs)		mtx_unlock(&(qs)->lock)
239 #define	TXQ_RING_EMPTY(qs)	drbr_empty((qs)->port->ifp, (qs)->txq[TXQ_ETH].txq_mr)
240 #define	TXQ_RING_NEEDS_ENQUEUE(qs)					\
241 	drbr_needs_enqueue((qs)->port->ifp, (qs)->txq[TXQ_ETH].txq_mr)
242 #define	TXQ_RING_FLUSH(qs)	drbr_flush((qs)->port->ifp, (qs)->txq[TXQ_ETH].txq_mr)
243 #define	TXQ_RING_DEQUEUE_COND(qs, func, arg)				\
244 	drbr_dequeue_cond((qs)->port->ifp, (qs)->txq[TXQ_ETH].txq_mr, func, arg)
245 #define	TXQ_RING_DEQUEUE(qs) \
246 	drbr_dequeue((qs)->port->ifp, (qs)->txq[TXQ_ETH].txq_mr)
247 
248 int cxgb_debug = 0;
249 
250 static void sge_timer_cb(void *arg);
251 static void sge_timer_reclaim(void *arg, int ncount);
252 static void sge_txq_reclaim_handler(void *arg, int ncount);
253 static void cxgb_start_locked(struct sge_qset *qs);
254 
255 /*
256  * XXX need to cope with bursty scheduling by looking at a wider
257  * window than we are now for determining the need for coalescing
258  *
259  */
260 static __inline uint64_t
261 check_pkt_coalesce(struct sge_qset *qs)
262 {
263         struct adapter *sc;
264         struct sge_txq *txq;
265 	uint8_t *fill;
266 
267 	if (__predict_false(cxgb_tx_coalesce_force))
268 		return (1);
269 	txq = &qs->txq[TXQ_ETH];
270         sc = qs->port->adapter;
271 	fill = &sc->tunq_fill[qs->idx];
272 
273 	if (cxgb_tx_coalesce_enable_start > COALESCE_START_MAX)
274 		cxgb_tx_coalesce_enable_start = COALESCE_START_MAX;
275 	if (cxgb_tx_coalesce_enable_stop < COALESCE_STOP_MIN)
276 		cxgb_tx_coalesce_enable_start = COALESCE_STOP_MIN;
277 	/*
278 	 * if the hardware transmit queue is more than 1/8 full
279 	 * we mark it as coalescing - we drop back from coalescing
280 	 * when we go below 1/32 full and there are no packets enqueued,
281 	 * this provides us with some degree of hysteresis
282 	 */
283         if (*fill != 0 && (txq->in_use <= cxgb_tx_coalesce_enable_stop) &&
284 	    TXQ_RING_EMPTY(qs) && (qs->coalescing == 0))
285                 *fill = 0;
286         else if (*fill == 0 && (txq->in_use >= cxgb_tx_coalesce_enable_start))
287                 *fill = 1;
288 
289 	return (sc->tunq_coalesce);
290 }
291 
292 #ifdef __LP64__
293 static void
294 set_wr_hdr(struct work_request_hdr *wrp, uint32_t wr_hi, uint32_t wr_lo)
295 {
296 	uint64_t wr_hilo;
297 #if _BYTE_ORDER == _LITTLE_ENDIAN
298 	wr_hilo = wr_hi;
299 	wr_hilo |= (((uint64_t)wr_lo)<<32);
300 #else
301 	wr_hilo = wr_lo;
302 	wr_hilo |= (((uint64_t)wr_hi)<<32);
303 #endif
304 	wrp->wrh_hilo = wr_hilo;
305 }
306 #else
307 static void
308 set_wr_hdr(struct work_request_hdr *wrp, uint32_t wr_hi, uint32_t wr_lo)
309 {
310 
311 	wrp->wrh_hi = wr_hi;
312 	wmb();
313 	wrp->wrh_lo = wr_lo;
314 }
315 #endif
316 
317 struct coalesce_info {
318 	int count;
319 	int nbytes;
320 };
321 
322 static int
323 coalesce_check(struct mbuf *m, void *arg)
324 {
325 	struct coalesce_info *ci = arg;
326 	int *count = &ci->count;
327 	int *nbytes = &ci->nbytes;
328 
329 	if ((*nbytes == 0) || ((*nbytes + m->m_len <= 10500) &&
330 		(*count < 7) && (m->m_next == NULL))) {
331 		*count += 1;
332 		*nbytes += m->m_len;
333 		return (1);
334 	}
335 	return (0);
336 }
337 
338 static struct mbuf *
339 cxgb_dequeue(struct sge_qset *qs)
340 {
341 	struct mbuf *m, *m_head, *m_tail;
342 	struct coalesce_info ci;
343 
344 
345 	if (check_pkt_coalesce(qs) == 0)
346 		return TXQ_RING_DEQUEUE(qs);
347 
348 	m_head = m_tail = NULL;
349 	ci.count = ci.nbytes = 0;
350 	do {
351 		m = TXQ_RING_DEQUEUE_COND(qs, coalesce_check, &ci);
352 		if (m_head == NULL) {
353 			m_tail = m_head = m;
354 		} else if (m != NULL) {
355 			m_tail->m_nextpkt = m;
356 			m_tail = m;
357 		}
358 	} while (m != NULL);
359 	if (ci.count > 7)
360 		panic("trying to coalesce %d packets in to one WR", ci.count);
361 	return (m_head);
362 }
363 
364 /**
365  *	reclaim_completed_tx - reclaims completed Tx descriptors
366  *	@adapter: the adapter
367  *	@q: the Tx queue to reclaim completed descriptors from
368  *
369  *	Reclaims Tx descriptors that the SGE has indicated it has processed,
370  *	and frees the associated buffers if possible.  Called with the Tx
371  *	queue's lock held.
372  */
373 static __inline int
374 reclaim_completed_tx(struct sge_qset *qs, int reclaim_min, int queue)
375 {
376 	struct sge_txq *q = &qs->txq[queue];
377 	int reclaim = desc_reclaimable(q);
378 
379 	if ((cxgb_tx_reclaim_threshold > TX_RECLAIM_MAX) ||
380 	    (cxgb_tx_reclaim_threshold < TX_RECLAIM_MIN))
381 		cxgb_tx_reclaim_threshold = TX_RECLAIM_DEFAULT;
382 
383 	if (reclaim < reclaim_min)
384 		return (0);
385 
386 	mtx_assert(&qs->lock, MA_OWNED);
387 	if (reclaim > 0) {
388 		t3_free_tx_desc(qs, reclaim, queue);
389 		q->cleaned += reclaim;
390 		q->in_use -= reclaim;
391 	}
392 	if (isset(&qs->txq_stopped, TXQ_ETH))
393                 clrbit(&qs->txq_stopped, TXQ_ETH);
394 
395 	return (reclaim);
396 }
397 
398 /**
399  *	should_restart_tx - are there enough resources to restart a Tx queue?
400  *	@q: the Tx queue
401  *
402  *	Checks if there are enough descriptors to restart a suspended Tx queue.
403  */
404 static __inline int
405 should_restart_tx(const struct sge_txq *q)
406 {
407 	unsigned int r = q->processed - q->cleaned;
408 
409 	return q->in_use - r < (q->size >> 1);
410 }
411 
412 /**
413  *	t3_sge_init - initialize SGE
414  *	@adap: the adapter
415  *	@p: the SGE parameters
416  *
417  *	Performs SGE initialization needed every time after a chip reset.
418  *	We do not initialize any of the queue sets here, instead the driver
419  *	top-level must request those individually.  We also do not enable DMA
420  *	here, that should be done after the queues have been set up.
421  */
422 void
423 t3_sge_init(adapter_t *adap, struct sge_params *p)
424 {
425 	u_int ctrl, ups;
426 
427 	ups = 0; /* = ffs(pci_resource_len(adap->pdev, 2) >> 12); */
428 
429 	ctrl = F_DROPPKT | V_PKTSHIFT(2) | F_FLMODE | F_AVOIDCQOVFL |
430 	       F_CQCRDTCTRL | F_CONGMODE | F_TNLFLMODE | F_FATLPERREN |
431 	       V_HOSTPAGESIZE(PAGE_SHIFT - 11) | F_BIGENDIANINGRESS |
432 	       V_USERSPACESIZE(ups ? ups - 1 : 0) | F_ISCSICOALESCING;
433 #if SGE_NUM_GENBITS == 1
434 	ctrl |= F_EGRGENCTRL;
435 #endif
436 	if (adap->params.rev > 0) {
437 		if (!(adap->flags & (USING_MSIX | USING_MSI)))
438 			ctrl |= F_ONEINTMULTQ | F_OPTONEINTMULTQ;
439 	}
440 	t3_write_reg(adap, A_SG_CONTROL, ctrl);
441 	t3_write_reg(adap, A_SG_EGR_RCQ_DRB_THRSH, V_HIRCQDRBTHRSH(512) |
442 		     V_LORCQDRBTHRSH(512));
443 	t3_write_reg(adap, A_SG_TIMER_TICK, core_ticks_per_usec(adap) / 10);
444 	t3_write_reg(adap, A_SG_CMDQ_CREDIT_TH, V_THRESHOLD(32) |
445 		     V_TIMEOUT(200 * core_ticks_per_usec(adap)));
446 	t3_write_reg(adap, A_SG_HI_DRB_HI_THRSH,
447 		     adap->params.rev < T3_REV_C ? 1000 : 500);
448 	t3_write_reg(adap, A_SG_HI_DRB_LO_THRSH, 256);
449 	t3_write_reg(adap, A_SG_LO_DRB_HI_THRSH, 1000);
450 	t3_write_reg(adap, A_SG_LO_DRB_LO_THRSH, 256);
451 	t3_write_reg(adap, A_SG_OCO_BASE, V_BASE1(0xfff));
452 	t3_write_reg(adap, A_SG_DRB_PRI_THRESH, 63 * 1024);
453 }
454 
455 
456 /**
457  *	sgl_len - calculates the size of an SGL of the given capacity
458  *	@n: the number of SGL entries
459  *
460  *	Calculates the number of flits needed for a scatter/gather list that
461  *	can hold the given number of entries.
462  */
463 static __inline unsigned int
464 sgl_len(unsigned int n)
465 {
466 	return ((3 * n) / 2 + (n & 1));
467 }
468 
469 /**
470  *	get_imm_packet - return the next ingress packet buffer from a response
471  *	@resp: the response descriptor containing the packet data
472  *
473  *	Return a packet containing the immediate data of the given response.
474  */
475 static int
476 get_imm_packet(adapter_t *sc, const struct rsp_desc *resp, struct mbuf *m)
477 {
478 
479 	if (resp->rss_hdr.opcode == CPL_RX_DATA) {
480 		const struct cpl_rx_data *cpl = (const void *)&resp->imm_data[0];
481 		m->m_len = sizeof(*cpl) + ntohs(cpl->len);
482 	} else if (resp->rss_hdr.opcode == CPL_RX_PKT) {
483 		const struct cpl_rx_pkt *cpl = (const void *)&resp->imm_data[0];
484 		m->m_len = sizeof(*cpl) + ntohs(cpl->len);
485 	} else
486 		m->m_len = IMMED_PKT_SIZE;
487 	m->m_ext.ext_buf = NULL;
488 	m->m_ext.ext_type = 0;
489 	memcpy(mtod(m, uint8_t *), resp->imm_data, m->m_len);
490 	return (0);
491 }
492 
493 static __inline u_int
494 flits_to_desc(u_int n)
495 {
496 	return (flit_desc_map[n]);
497 }
498 
499 #define SGE_PARERR (F_CPPARITYERROR | F_OCPARITYERROR | F_RCPARITYERROR | \
500 		    F_IRPARITYERROR | V_ITPARITYERROR(M_ITPARITYERROR) | \
501 		    V_FLPARITYERROR(M_FLPARITYERROR) | F_LODRBPARITYERROR | \
502 		    F_HIDRBPARITYERROR | F_LORCQPARITYERROR | \
503 		    F_HIRCQPARITYERROR)
504 #define SGE_FRAMINGERR (F_UC_REQ_FRAMINGERROR | F_R_REQ_FRAMINGERROR)
505 #define SGE_FATALERR (SGE_PARERR | SGE_FRAMINGERR | F_RSPQCREDITOVERFOW | \
506 		      F_RSPQDISABLED)
507 
508 /**
509  *	t3_sge_err_intr_handler - SGE async event interrupt handler
510  *	@adapter: the adapter
511  *
512  *	Interrupt handler for SGE asynchronous (non-data) events.
513  */
514 void
515 t3_sge_err_intr_handler(adapter_t *adapter)
516 {
517 	unsigned int v, status;
518 
519 	status = t3_read_reg(adapter, A_SG_INT_CAUSE);
520 	if (status & SGE_PARERR)
521 		CH_ALERT(adapter, "SGE parity error (0x%x)\n",
522 			 status & SGE_PARERR);
523 	if (status & SGE_FRAMINGERR)
524 		CH_ALERT(adapter, "SGE framing error (0x%x)\n",
525 			 status & SGE_FRAMINGERR);
526 	if (status & F_RSPQCREDITOVERFOW)
527 		CH_ALERT(adapter, "SGE response queue credit overflow\n");
528 
529 	if (status & F_RSPQDISABLED) {
530 		v = t3_read_reg(adapter, A_SG_RSPQ_FL_STATUS);
531 
532 		CH_ALERT(adapter,
533 			 "packet delivered to disabled response queue (0x%x)\n",
534 			 (v >> S_RSPQ0DISABLED) & 0xff);
535 	}
536 
537 	t3_write_reg(adapter, A_SG_INT_CAUSE, status);
538 	if (status & SGE_FATALERR)
539 		t3_fatal_err(adapter);
540 }
541 
542 void
543 t3_sge_prep(adapter_t *adap, struct sge_params *p)
544 {
545 	int i, nqsets, fl_q_size, jumbo_q_size, use_16k, jumbo_buf_size;
546 
547 	nqsets = min(SGE_QSETS / adap->params.nports, mp_ncpus);
548 	nqsets *= adap->params.nports;
549 
550 	fl_q_size = min(nmbclusters/(3*nqsets), FL_Q_SIZE);
551 
552 	while (!powerof2(fl_q_size))
553 		fl_q_size--;
554 
555 	use_16k = cxgb_use_16k_clusters != -1 ? cxgb_use_16k_clusters :
556 	    is_offload(adap);
557 
558 #if __FreeBSD_version >= 700111
559 	if (use_16k) {
560 		jumbo_q_size = min(nmbjumbo16/(3*nqsets), JUMBO_Q_SIZE);
561 		jumbo_buf_size = MJUM16BYTES;
562 	} else {
563 		jumbo_q_size = min(nmbjumbo9/(3*nqsets), JUMBO_Q_SIZE);
564 		jumbo_buf_size = MJUM9BYTES;
565 	}
566 #else
567 	jumbo_q_size = min(nmbjumbop/(3*nqsets), JUMBO_Q_SIZE);
568 	jumbo_buf_size = MJUMPAGESIZE;
569 #endif
570 	while (!powerof2(jumbo_q_size))
571 		jumbo_q_size--;
572 
573 	if (fl_q_size < (FL_Q_SIZE / 4) || jumbo_q_size < (JUMBO_Q_SIZE / 2))
574 		device_printf(adap->dev,
575 		    "Insufficient clusters and/or jumbo buffers.\n");
576 
577 	p->max_pkt_size = jumbo_buf_size - sizeof(struct cpl_rx_data);
578 
579 	for (i = 0; i < SGE_QSETS; ++i) {
580 		struct qset_params *q = p->qset + i;
581 
582 		if (adap->params.nports > 2) {
583 			q->coalesce_usecs = 50;
584 		} else {
585 #ifdef INVARIANTS
586 			q->coalesce_usecs = 10;
587 #else
588 			q->coalesce_usecs = 5;
589 #endif
590 		}
591 		q->polling = 0;
592 		q->rspq_size = RSPQ_Q_SIZE;
593 		q->fl_size = fl_q_size;
594 		q->jumbo_size = jumbo_q_size;
595 		q->jumbo_buf_size = jumbo_buf_size;
596 		q->txq_size[TXQ_ETH] = TX_ETH_Q_SIZE;
597 		q->txq_size[TXQ_OFLD] = is_offload(adap) ? TX_OFLD_Q_SIZE : 16;
598 		q->txq_size[TXQ_CTRL] = TX_CTRL_Q_SIZE;
599 		q->cong_thres = 0;
600 	}
601 }
602 
603 int
604 t3_sge_alloc(adapter_t *sc)
605 {
606 
607 	/* The parent tag. */
608 	if (bus_dma_tag_create( bus_get_dma_tag(sc->dev),/* PCI parent */
609 				1, 0,			/* algnmnt, boundary */
610 				BUS_SPACE_MAXADDR,	/* lowaddr */
611 				BUS_SPACE_MAXADDR,	/* highaddr */
612 				NULL, NULL,		/* filter, filterarg */
613 				BUS_SPACE_MAXSIZE_32BIT,/* maxsize */
614 				BUS_SPACE_UNRESTRICTED, /* nsegments */
615 				BUS_SPACE_MAXSIZE_32BIT,/* maxsegsize */
616 				0,			/* flags */
617 				NULL, NULL,		/* lock, lockarg */
618 				&sc->parent_dmat)) {
619 		device_printf(sc->dev, "Cannot allocate parent DMA tag\n");
620 		return (ENOMEM);
621 	}
622 
623 	/*
624 	 * DMA tag for normal sized RX frames
625 	 */
626 	if (bus_dma_tag_create(sc->parent_dmat, MCLBYTES, 0, BUS_SPACE_MAXADDR,
627 		BUS_SPACE_MAXADDR, NULL, NULL, MCLBYTES, 1,
628 		MCLBYTES, BUS_DMA_ALLOCNOW, NULL, NULL, &sc->rx_dmat)) {
629 		device_printf(sc->dev, "Cannot allocate RX DMA tag\n");
630 		return (ENOMEM);
631 	}
632 
633 	/*
634 	 * DMA tag for jumbo sized RX frames.
635 	 */
636 	if (bus_dma_tag_create(sc->parent_dmat, MJUM16BYTES, 0, BUS_SPACE_MAXADDR,
637 		BUS_SPACE_MAXADDR, NULL, NULL, MJUM16BYTES, 1, MJUM16BYTES,
638 		BUS_DMA_ALLOCNOW, NULL, NULL, &sc->rx_jumbo_dmat)) {
639 		device_printf(sc->dev, "Cannot allocate RX jumbo DMA tag\n");
640 		return (ENOMEM);
641 	}
642 
643 	/*
644 	 * DMA tag for TX frames.
645 	 */
646 	if (bus_dma_tag_create(sc->parent_dmat, 1, 0, BUS_SPACE_MAXADDR,
647 		BUS_SPACE_MAXADDR, NULL, NULL, TX_MAX_SIZE, TX_MAX_SEGS,
648 		TX_MAX_SIZE, BUS_DMA_ALLOCNOW,
649 		NULL, NULL, &sc->tx_dmat)) {
650 		device_printf(sc->dev, "Cannot allocate TX DMA tag\n");
651 		return (ENOMEM);
652 	}
653 
654 	return (0);
655 }
656 
657 int
658 t3_sge_free(struct adapter * sc)
659 {
660 
661 	if (sc->tx_dmat != NULL)
662 		bus_dma_tag_destroy(sc->tx_dmat);
663 
664 	if (sc->rx_jumbo_dmat != NULL)
665 		bus_dma_tag_destroy(sc->rx_jumbo_dmat);
666 
667 	if (sc->rx_dmat != NULL)
668 		bus_dma_tag_destroy(sc->rx_dmat);
669 
670 	if (sc->parent_dmat != NULL)
671 		bus_dma_tag_destroy(sc->parent_dmat);
672 
673 	return (0);
674 }
675 
676 void
677 t3_update_qset_coalesce(struct sge_qset *qs, const struct qset_params *p)
678 {
679 
680 	qs->rspq.holdoff_tmr = max(p->coalesce_usecs * 10, 1U);
681 	qs->rspq.polling = 0 /* p->polling */;
682 }
683 
684 #if !defined(__i386__) && !defined(__amd64__)
685 static void
686 refill_fl_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
687 {
688 	struct refill_fl_cb_arg *cb_arg = arg;
689 
690 	cb_arg->error = error;
691 	cb_arg->seg = segs[0];
692 	cb_arg->nseg = nseg;
693 
694 }
695 #endif
696 /**
697  *	refill_fl - refill an SGE free-buffer list
698  *	@sc: the controller softc
699  *	@q: the free-list to refill
700  *	@n: the number of new buffers to allocate
701  *
702  *	(Re)populate an SGE free-buffer list with up to @n new packet buffers.
703  *	The caller must assure that @n does not exceed the queue's capacity.
704  */
705 static void
706 refill_fl(adapter_t *sc, struct sge_fl *q, int n)
707 {
708 	struct rx_sw_desc *sd = &q->sdesc[q->pidx];
709 	struct rx_desc *d = &q->desc[q->pidx];
710 	struct refill_fl_cb_arg cb_arg;
711 	struct mbuf *m;
712 	caddr_t cl;
713 	int err;
714 
715 	cb_arg.error = 0;
716 	while (n--) {
717 		/*
718 		 * We allocate an uninitialized mbuf + cluster, mbuf is
719 		 * initialized after rx.
720 		 */
721 		if (q->zone == zone_pack) {
722 			if ((m = m_getcl(M_NOWAIT, MT_NOINIT, M_PKTHDR)) == NULL)
723 				break;
724 			cl = m->m_ext.ext_buf;
725 		} else {
726 			if ((cl = m_cljget(NULL, M_NOWAIT, q->buf_size)) == NULL)
727 				break;
728 			if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
729 				uma_zfree(q->zone, cl);
730 				break;
731 			}
732 		}
733 		if ((sd->flags & RX_SW_DESC_MAP_CREATED) == 0) {
734 			if ((err = bus_dmamap_create(q->entry_tag, 0, &sd->map))) {
735 				log(LOG_WARNING, "bus_dmamap_create failed %d\n", err);
736 				uma_zfree(q->zone, cl);
737 				goto done;
738 			}
739 			sd->flags |= RX_SW_DESC_MAP_CREATED;
740 		}
741 #if !defined(__i386__) && !defined(__amd64__)
742 		err = bus_dmamap_load(q->entry_tag, sd->map,
743 		    cl, q->buf_size, refill_fl_cb, &cb_arg, 0);
744 
745 		if (err != 0 || cb_arg.error) {
746 			if (q->zone == zone_pack)
747 				uma_zfree(q->zone, cl);
748 			m_free(m);
749 			goto done;
750 		}
751 #else
752 		cb_arg.seg.ds_addr = pmap_kextract((vm_offset_t)cl);
753 #endif
754 		sd->flags |= RX_SW_DESC_INUSE;
755 		sd->rxsd_cl = cl;
756 		sd->m = m;
757 		d->addr_lo = htobe32(cb_arg.seg.ds_addr & 0xffffffff);
758 		d->addr_hi = htobe32(((uint64_t)cb_arg.seg.ds_addr >>32) & 0xffffffff);
759 		d->len_gen = htobe32(V_FLD_GEN1(q->gen));
760 		d->gen2 = htobe32(V_FLD_GEN2(q->gen));
761 
762 		d++;
763 		sd++;
764 
765 		if (++q->pidx == q->size) {
766 			q->pidx = 0;
767 			q->gen ^= 1;
768 			sd = q->sdesc;
769 			d = q->desc;
770 		}
771 		q->credits++;
772 		q->db_pending++;
773 	}
774 
775 done:
776 	if (q->db_pending >= 32) {
777 		q->db_pending = 0;
778 		t3_write_reg(sc, A_SG_KDOORBELL, V_EGRCNTX(q->cntxt_id));
779 	}
780 }
781 
782 
783 /**
784  *	free_rx_bufs - free the Rx buffers on an SGE free list
785  *	@sc: the controle softc
786  *	@q: the SGE free list to clean up
787  *
788  *	Release the buffers on an SGE free-buffer Rx queue.  HW fetching from
789  *	this queue should be stopped before calling this function.
790  */
791 static void
792 free_rx_bufs(adapter_t *sc, struct sge_fl *q)
793 {
794 	u_int cidx = q->cidx;
795 
796 	while (q->credits--) {
797 		struct rx_sw_desc *d = &q->sdesc[cidx];
798 
799 		if (d->flags & RX_SW_DESC_INUSE) {
800 			bus_dmamap_unload(q->entry_tag, d->map);
801 			bus_dmamap_destroy(q->entry_tag, d->map);
802 			if (q->zone == zone_pack) {
803 				m_init(d->m, zone_pack, MCLBYTES,
804 				    M_NOWAIT, MT_DATA, M_EXT);
805 				uma_zfree(zone_pack, d->m);
806 			} else {
807 				m_init(d->m, zone_mbuf, MLEN,
808 				    M_NOWAIT, MT_DATA, 0);
809 				uma_zfree(zone_mbuf, 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, CALLOUT_MPSAFE);
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 	mbufq_init(&qs->txq[TXQ_ETH].sendq);
1126 	mbufq_init(&qs->txq[TXQ_OFLD].sendq);
1127 	mbufq_init(&qs->txq[TXQ_CTRL].sendq);
1128 }
1129 
1130 
1131 static void
1132 txq_prod(struct sge_txq *txq, unsigned int ndesc, struct txq_state *txqs)
1133 {
1134 	txq->in_use += ndesc;
1135 	/*
1136 	 * XXX we don't handle stopping of queue
1137 	 * presumably start handles this when we bump against the end
1138 	 */
1139 	txqs->gen = txq->gen;
1140 	txq->unacked += ndesc;
1141 	txqs->compl = (txq->unacked & 32) << (S_WR_COMPL - 5);
1142 	txq->unacked &= 31;
1143 	txqs->pidx = txq->pidx;
1144 	txq->pidx += ndesc;
1145 #ifdef INVARIANTS
1146 	if (((txqs->pidx > txq->cidx) &&
1147 		(txq->pidx < txqs->pidx) &&
1148 		(txq->pidx >= txq->cidx)) ||
1149 	    ((txqs->pidx < txq->cidx) &&
1150 		(txq->pidx >= txq-> cidx)) ||
1151 	    ((txqs->pidx < txq->cidx) &&
1152 		(txq->cidx < txqs->pidx)))
1153 		panic("txqs->pidx=%d txq->pidx=%d txq->cidx=%d",
1154 		    txqs->pidx, txq->pidx, txq->cidx);
1155 #endif
1156 	if (txq->pidx >= txq->size) {
1157 		txq->pidx -= txq->size;
1158 		txq->gen ^= 1;
1159 	}
1160 
1161 }
1162 
1163 /**
1164  *	calc_tx_descs - calculate the number of Tx descriptors for a packet
1165  *	@m: the packet mbufs
1166  *      @nsegs: the number of segments
1167  *
1168  * 	Returns the number of Tx descriptors needed for the given Ethernet
1169  * 	packet.  Ethernet packets require addition of WR and CPL headers.
1170  */
1171 static __inline unsigned int
1172 calc_tx_descs(const struct mbuf *m, int nsegs)
1173 {
1174 	unsigned int flits;
1175 
1176 	if (m->m_pkthdr.len <= PIO_LEN)
1177 		return 1;
1178 
1179 	flits = sgl_len(nsegs) + 2;
1180 	if (m->m_pkthdr.csum_flags & CSUM_TSO)
1181 		flits++;
1182 
1183 	return flits_to_desc(flits);
1184 }
1185 
1186 /**
1187  *	make_sgl - populate a scatter/gather list for a packet
1188  *	@sgp: the SGL to populate
1189  *	@segs: the packet dma segments
1190  *	@nsegs: the number of segments
1191  *
1192  *	Generates a scatter/gather list for the buffers that make up a packet
1193  *	and returns the SGL size in 8-byte words.  The caller must size the SGL
1194  *	appropriately.
1195  */
1196 static __inline void
1197 make_sgl(struct sg_ent *sgp, bus_dma_segment_t *segs, int nsegs)
1198 {
1199 	int i, idx;
1200 
1201 	for (idx = 0, i = 0; i < nsegs; i++) {
1202 		/*
1203 		 * firmware doesn't like empty segments
1204 		 */
1205 		if (segs[i].ds_len == 0)
1206 			continue;
1207 		if (i && idx == 0)
1208 			++sgp;
1209 
1210 		sgp->len[idx] = htobe32(segs[i].ds_len);
1211 		sgp->addr[idx] = htobe64(segs[i].ds_addr);
1212 		idx ^= 1;
1213 	}
1214 
1215 	if (idx) {
1216 		sgp->len[idx] = 0;
1217 		sgp->addr[idx] = 0;
1218 	}
1219 }
1220 
1221 /**
1222  *	check_ring_tx_db - check and potentially ring a Tx queue's doorbell
1223  *	@adap: the adapter
1224  *	@q: the Tx queue
1225  *
1226  *	Ring the doorbell if a Tx queue is asleep.  There is a natural race,
1227  *	where the HW is going to sleep just after we checked, however,
1228  *	then the interrupt handler will detect the outstanding TX packet
1229  *	and ring the doorbell for us.
1230  *
1231  *	When GTS is disabled we unconditionally ring the doorbell.
1232  */
1233 static __inline void
1234 check_ring_tx_db(adapter_t *adap, struct sge_txq *q, int mustring)
1235 {
1236 #if USE_GTS
1237 	clear_bit(TXQ_LAST_PKT_DB, &q->flags);
1238 	if (test_and_set_bit(TXQ_RUNNING, &q->flags) == 0) {
1239 		set_bit(TXQ_LAST_PKT_DB, &q->flags);
1240 #ifdef T3_TRACE
1241 		T3_TRACE1(adap->tb[q->cntxt_id & 7], "doorbell Tx, cntxt %d",
1242 			  q->cntxt_id);
1243 #endif
1244 		t3_write_reg(adap, A_SG_KDOORBELL,
1245 			     F_SELEGRCNTX | V_EGRCNTX(q->cntxt_id));
1246 	}
1247 #else
1248 	if (mustring || ++q->db_pending >= 32) {
1249 		wmb();            /* write descriptors before telling HW */
1250 		t3_write_reg(adap, A_SG_KDOORBELL,
1251 		    F_SELEGRCNTX | V_EGRCNTX(q->cntxt_id));
1252 		q->db_pending = 0;
1253 	}
1254 #endif
1255 }
1256 
1257 static __inline void
1258 wr_gen2(struct tx_desc *d, unsigned int gen)
1259 {
1260 #if SGE_NUM_GENBITS == 2
1261 	d->flit[TX_DESC_FLITS - 1] = htobe64(gen);
1262 #endif
1263 }
1264 
1265 /**
1266  *	write_wr_hdr_sgl - write a WR header and, optionally, SGL
1267  *	@ndesc: number of Tx descriptors spanned by the SGL
1268  *	@txd: first Tx descriptor to be written
1269  *	@txqs: txq state (generation and producer index)
1270  *	@txq: the SGE Tx queue
1271  *	@sgl: the SGL
1272  *	@flits: number of flits to the start of the SGL in the first descriptor
1273  *	@sgl_flits: the SGL size in flits
1274  *	@wr_hi: top 32 bits of WR header based on WR type (big endian)
1275  *	@wr_lo: low 32 bits of WR header based on WR type (big endian)
1276  *
1277  *	Write a work request header and an associated SGL.  If the SGL is
1278  *	small enough to fit into one Tx descriptor it has already been written
1279  *	and we just need to write the WR header.  Otherwise we distribute the
1280  *	SGL across the number of descriptors it spans.
1281  */
1282 static void
1283 write_wr_hdr_sgl(unsigned int ndesc, struct tx_desc *txd, struct txq_state *txqs,
1284     const struct sge_txq *txq, const struct sg_ent *sgl, unsigned int flits,
1285     unsigned int sgl_flits, unsigned int wr_hi, unsigned int wr_lo)
1286 {
1287 
1288 	struct work_request_hdr *wrp = (struct work_request_hdr *)txd;
1289 	struct tx_sw_desc *txsd = &txq->sdesc[txqs->pidx];
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 			txsd++;
1321 			if (++txqs->pidx == txq->size) {
1322 				txqs->pidx = 0;
1323 				txqs->gen ^= 1;
1324 				txd = txq->desc;
1325 				txsd = txq->sdesc;
1326 			}
1327 
1328 			/*
1329 			 * when the head of the mbuf chain
1330 			 * is freed all clusters will be freed
1331 			 * with it
1332 			 */
1333 			wrp = (struct work_request_hdr *)txd;
1334 			wrp->wrh_hi = htonl(V_WR_DATATYPE(1) |
1335 			    V_WR_SGLSFLT(1)) | wr_hi;
1336 			wrp->wrh_lo = htonl(V_WR_LEN(min(WR_FLITS,
1337 				    sgl_flits + 1)) |
1338 			    V_WR_GEN(txqs->gen)) | wr_lo;
1339 			wr_gen2(txd, txqs->gen);
1340 			flits = 1;
1341 		}
1342 		wrp->wrh_hi |= htonl(F_WR_EOP);
1343 		wmb();
1344 		wp->wrh_lo = htonl(V_WR_LEN(WR_FLITS) | V_WR_GEN(ogen)) | wr_lo;
1345 		wr_gen2((struct tx_desc *)wp, ogen);
1346 	}
1347 }
1348 
1349 /* sizeof(*eh) + sizeof(*ip) + sizeof(*tcp) */
1350 #define TCPPKTHDRSIZE (ETHER_HDR_LEN + 20 + 20)
1351 
1352 #define GET_VTAG(cntrl, m) \
1353 do { \
1354 	if ((m)->m_flags & M_VLANTAG)					            \
1355 		cntrl |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN((m)->m_pkthdr.ether_vtag); \
1356 } while (0)
1357 
1358 static int
1359 t3_encap(struct sge_qset *qs, struct mbuf **m)
1360 {
1361 	adapter_t *sc;
1362 	struct mbuf *m0;
1363 	struct sge_txq *txq;
1364 	struct txq_state txqs;
1365 	struct port_info *pi;
1366 	unsigned int ndesc, flits, cntrl, mlen;
1367 	int err, nsegs, tso_info = 0;
1368 
1369 	struct work_request_hdr *wrp;
1370 	struct tx_sw_desc *txsd;
1371 	struct sg_ent *sgp, *sgl;
1372 	uint32_t wr_hi, wr_lo, sgl_flits;
1373 	bus_dma_segment_t segs[TX_MAX_SEGS];
1374 
1375 	struct tx_desc *txd;
1376 
1377 	pi = qs->port;
1378 	sc = pi->adapter;
1379 	txq = &qs->txq[TXQ_ETH];
1380 	txd = &txq->desc[txq->pidx];
1381 	txsd = &txq->sdesc[txq->pidx];
1382 	sgl = txq->txq_sgl;
1383 
1384 	prefetch(txd);
1385 	m0 = *m;
1386 
1387 	mtx_assert(&qs->lock, MA_OWNED);
1388 	cntrl = V_TXPKT_INTF(pi->txpkt_intf);
1389 	KASSERT(m0->m_flags & M_PKTHDR, ("not packet header\n"));
1390 
1391 	if  (m0->m_nextpkt == NULL && m0->m_next != NULL &&
1392 	    m0->m_pkthdr.csum_flags & (CSUM_TSO))
1393 		tso_info = V_LSO_MSS(m0->m_pkthdr.tso_segsz);
1394 
1395 	if (m0->m_nextpkt != NULL) {
1396 		busdma_map_sg_vec(txq->entry_tag, txsd->map, m0, segs, &nsegs);
1397 		ndesc = 1;
1398 		mlen = 0;
1399 	} else {
1400 		if ((err = busdma_map_sg_collapse(txq->entry_tag, txsd->map,
1401 		    &m0, segs, &nsegs))) {
1402 			if (cxgb_debug)
1403 				printf("failed ... err=%d\n", err);
1404 			return (err);
1405 		}
1406 		mlen = m0->m_pkthdr.len;
1407 		ndesc = calc_tx_descs(m0, nsegs);
1408 	}
1409 	txq_prod(txq, ndesc, &txqs);
1410 
1411 	KASSERT(m0->m_pkthdr.len, ("empty packet nsegs=%d", nsegs));
1412 	txsd->m = m0;
1413 
1414 	if (m0->m_nextpkt != NULL) {
1415 		struct cpl_tx_pkt_batch *cpl_batch = (struct cpl_tx_pkt_batch *)txd;
1416 		int i, fidx;
1417 
1418 		if (nsegs > 7)
1419 			panic("trying to coalesce %d packets in to one WR", nsegs);
1420 		txq->txq_coalesced += nsegs;
1421 		wrp = (struct work_request_hdr *)txd;
1422 		flits = nsegs*2 + 1;
1423 
1424 		for (fidx = 1, i = 0; i < nsegs; i++, fidx += 2) {
1425 			struct cpl_tx_pkt_batch_entry *cbe;
1426 			uint64_t flit;
1427 			uint32_t *hflit = (uint32_t *)&flit;
1428 			int cflags = m0->m_pkthdr.csum_flags;
1429 
1430 			cntrl = V_TXPKT_INTF(pi->txpkt_intf);
1431 			GET_VTAG(cntrl, m0);
1432 			cntrl |= V_TXPKT_OPCODE(CPL_TX_PKT);
1433 			if (__predict_false(!(cflags & CSUM_IP)))
1434 				cntrl |= F_TXPKT_IPCSUM_DIS;
1435 			if (__predict_false(!(cflags & (CSUM_TCP | CSUM_UDP |
1436 			    CSUM_UDP_IPV6 | CSUM_TCP_IPV6))))
1437 				cntrl |= F_TXPKT_L4CSUM_DIS;
1438 
1439 			hflit[0] = htonl(cntrl);
1440 			hflit[1] = htonl(segs[i].ds_len | 0x80000000);
1441 			flit |= htobe64(1 << 24);
1442 			cbe = &cpl_batch->pkt_entry[i];
1443 			cbe->cntrl = hflit[0];
1444 			cbe->len = hflit[1];
1445 			cbe->addr = htobe64(segs[i].ds_addr);
1446 		}
1447 
1448 		wr_hi = htonl(F_WR_SOP | F_WR_EOP | V_WR_DATATYPE(1) |
1449 		    V_WR_SGLSFLT(flits)) |
1450 		    htonl(V_WR_OP(FW_WROPCODE_TUNNEL_TX_PKT) | txqs.compl);
1451 		wr_lo = htonl(V_WR_LEN(flits) |
1452 		    V_WR_GEN(txqs.gen)) | htonl(V_WR_TID(txq->token));
1453 		set_wr_hdr(wrp, wr_hi, wr_lo);
1454 		wmb();
1455 		ETHER_BPF_MTAP(pi->ifp, m0);
1456 		wr_gen2(txd, txqs.gen);
1457 		check_ring_tx_db(sc, txq, 0);
1458 		return (0);
1459 	} else if (tso_info) {
1460 		uint16_t eth_type;
1461 		struct cpl_tx_pkt_lso *hdr = (struct cpl_tx_pkt_lso *)txd;
1462 		struct ether_header *eh;
1463 		void *l3hdr;
1464 		struct tcphdr *tcp;
1465 
1466 		txd->flit[2] = 0;
1467 		GET_VTAG(cntrl, m0);
1468 		cntrl |= V_TXPKT_OPCODE(CPL_TX_PKT_LSO);
1469 		hdr->cntrl = htonl(cntrl);
1470 		hdr->len = htonl(mlen | 0x80000000);
1471 
1472 		if (__predict_false(mlen < TCPPKTHDRSIZE)) {
1473 			printf("mbuf=%p,len=%d,tso_segsz=%d,csum_flags=%#x,flags=%#x",
1474 			    m0, mlen, m0->m_pkthdr.tso_segsz,
1475 			    m0->m_pkthdr.csum_flags, m0->m_flags);
1476 			panic("tx tso packet too small");
1477 		}
1478 
1479 		/* Make sure that ether, ip, tcp headers are all in m0 */
1480 		if (__predict_false(m0->m_len < TCPPKTHDRSIZE)) {
1481 			m0 = m_pullup(m0, TCPPKTHDRSIZE);
1482 			if (__predict_false(m0 == NULL)) {
1483 				/* XXX panic probably an overreaction */
1484 				panic("couldn't fit header into mbuf");
1485 			}
1486 		}
1487 
1488 		eh = mtod(m0, struct ether_header *);
1489 		eth_type = eh->ether_type;
1490 		if (eth_type == htons(ETHERTYPE_VLAN)) {
1491 			struct ether_vlan_header *evh = (void *)eh;
1492 
1493 			tso_info |= V_LSO_ETH_TYPE(CPL_ETH_II_VLAN);
1494 			l3hdr = evh + 1;
1495 			eth_type = evh->evl_proto;
1496 		} else {
1497 			tso_info |= V_LSO_ETH_TYPE(CPL_ETH_II);
1498 			l3hdr = eh + 1;
1499 		}
1500 
1501 		if (eth_type == htons(ETHERTYPE_IP)) {
1502 			struct ip *ip = l3hdr;
1503 
1504 			tso_info |= V_LSO_IPHDR_WORDS(ip->ip_hl);
1505 			tcp = (struct tcphdr *)(ip + 1);
1506 		} else if (eth_type == htons(ETHERTYPE_IPV6)) {
1507 			struct ip6_hdr *ip6 = l3hdr;
1508 
1509 			KASSERT(ip6->ip6_nxt == IPPROTO_TCP,
1510 			    ("%s: CSUM_TSO with ip6_nxt %d",
1511 			    __func__, ip6->ip6_nxt));
1512 
1513 			tso_info |= F_LSO_IPV6;
1514 			tso_info |= V_LSO_IPHDR_WORDS(sizeof(*ip6) >> 2);
1515 			tcp = (struct tcphdr *)(ip6 + 1);
1516 		} else
1517 			panic("%s: CSUM_TSO but neither ip nor ip6", __func__);
1518 
1519 		tso_info |= V_LSO_TCPHDR_WORDS(tcp->th_off);
1520 		hdr->lso_info = htonl(tso_info);
1521 
1522 		if (__predict_false(mlen <= PIO_LEN)) {
1523 			/*
1524 			 * pkt not undersized but fits in PIO_LEN
1525 			 * Indicates a TSO bug at the higher levels.
1526 			 */
1527 			txsd->m = NULL;
1528 			m_copydata(m0, 0, mlen, (caddr_t)&txd->flit[3]);
1529 			flits = (mlen + 7) / 8 + 3;
1530 			wr_hi = htonl(V_WR_BCNTLFLT(mlen & 7) |
1531 					  V_WR_OP(FW_WROPCODE_TUNNEL_TX_PKT) |
1532 					  F_WR_SOP | F_WR_EOP | txqs.compl);
1533 			wr_lo = htonl(V_WR_LEN(flits) |
1534 			    V_WR_GEN(txqs.gen) | V_WR_TID(txq->token));
1535 			set_wr_hdr(&hdr->wr, wr_hi, wr_lo);
1536 			wmb();
1537 			ETHER_BPF_MTAP(pi->ifp, m0);
1538 			wr_gen2(txd, txqs.gen);
1539 			check_ring_tx_db(sc, txq, 0);
1540 			m_freem(m0);
1541 			return (0);
1542 		}
1543 		flits = 3;
1544 	} else {
1545 		struct cpl_tx_pkt *cpl = (struct cpl_tx_pkt *)txd;
1546 
1547 		GET_VTAG(cntrl, m0);
1548 		cntrl |= V_TXPKT_OPCODE(CPL_TX_PKT);
1549 		if (__predict_false(!(m0->m_pkthdr.csum_flags & CSUM_IP)))
1550 			cntrl |= F_TXPKT_IPCSUM_DIS;
1551 		if (__predict_false(!(m0->m_pkthdr.csum_flags & (CSUM_TCP |
1552 		    CSUM_UDP | CSUM_UDP_IPV6 | CSUM_TCP_IPV6))))
1553 			cntrl |= F_TXPKT_L4CSUM_DIS;
1554 		cpl->cntrl = htonl(cntrl);
1555 		cpl->len = htonl(mlen | 0x80000000);
1556 
1557 		if (mlen <= PIO_LEN) {
1558 			txsd->m = NULL;
1559 			m_copydata(m0, 0, mlen, (caddr_t)&txd->flit[2]);
1560 			flits = (mlen + 7) / 8 + 2;
1561 
1562 			wr_hi = htonl(V_WR_BCNTLFLT(mlen & 7) |
1563 			    V_WR_OP(FW_WROPCODE_TUNNEL_TX_PKT) |
1564 					  F_WR_SOP | F_WR_EOP | txqs.compl);
1565 			wr_lo = htonl(V_WR_LEN(flits) |
1566 			    V_WR_GEN(txqs.gen) | V_WR_TID(txq->token));
1567 			set_wr_hdr(&cpl->wr, wr_hi, wr_lo);
1568 			wmb();
1569 			ETHER_BPF_MTAP(pi->ifp, m0);
1570 			wr_gen2(txd, txqs.gen);
1571 			check_ring_tx_db(sc, txq, 0);
1572 			m_freem(m0);
1573 			return (0);
1574 		}
1575 		flits = 2;
1576 	}
1577 	wrp = (struct work_request_hdr *)txd;
1578 	sgp = (ndesc == 1) ? (struct sg_ent *)&txd->flit[flits] : sgl;
1579 	make_sgl(sgp, segs, nsegs);
1580 
1581 	sgl_flits = sgl_len(nsegs);
1582 
1583 	ETHER_BPF_MTAP(pi->ifp, m0);
1584 
1585 	KASSERT(ndesc <= 4, ("ndesc too large %d", ndesc));
1586 	wr_hi = htonl(V_WR_OP(FW_WROPCODE_TUNNEL_TX_PKT) | txqs.compl);
1587 	wr_lo = htonl(V_WR_TID(txq->token));
1588 	write_wr_hdr_sgl(ndesc, txd, &txqs, txq, sgl, flits,
1589 	    sgl_flits, wr_hi, wr_lo);
1590 	check_ring_tx_db(sc, txq, 0);
1591 
1592 	return (0);
1593 }
1594 
1595 void
1596 cxgb_tx_watchdog(void *arg)
1597 {
1598 	struct sge_qset *qs = arg;
1599 	struct sge_txq *txq = &qs->txq[TXQ_ETH];
1600 
1601         if (qs->coalescing != 0 &&
1602 	    (txq->in_use <= cxgb_tx_coalesce_enable_stop) &&
1603 	    TXQ_RING_EMPTY(qs))
1604                 qs->coalescing = 0;
1605         else if (qs->coalescing == 0 &&
1606 	    (txq->in_use >= cxgb_tx_coalesce_enable_start))
1607                 qs->coalescing = 1;
1608 	if (TXQ_TRYLOCK(qs)) {
1609 		qs->qs_flags |= QS_FLUSHING;
1610 		cxgb_start_locked(qs);
1611 		qs->qs_flags &= ~QS_FLUSHING;
1612 		TXQ_UNLOCK(qs);
1613 	}
1614 	if (qs->port->ifp->if_drv_flags & IFF_DRV_RUNNING)
1615 		callout_reset_on(&txq->txq_watchdog, hz/4, cxgb_tx_watchdog,
1616 		    qs, txq->txq_watchdog.c_cpu);
1617 }
1618 
1619 static void
1620 cxgb_tx_timeout(void *arg)
1621 {
1622 	struct sge_qset *qs = arg;
1623 	struct sge_txq *txq = &qs->txq[TXQ_ETH];
1624 
1625 	if (qs->coalescing == 0 && (txq->in_use >= (txq->size>>3)))
1626                 qs->coalescing = 1;
1627 	if (TXQ_TRYLOCK(qs)) {
1628 		qs->qs_flags |= QS_TIMEOUT;
1629 		cxgb_start_locked(qs);
1630 		qs->qs_flags &= ~QS_TIMEOUT;
1631 		TXQ_UNLOCK(qs);
1632 	}
1633 }
1634 
1635 static void
1636 cxgb_start_locked(struct sge_qset *qs)
1637 {
1638 	struct mbuf *m_head = NULL;
1639 	struct sge_txq *txq = &qs->txq[TXQ_ETH];
1640 	struct port_info *pi = qs->port;
1641 	struct ifnet *ifp = pi->ifp;
1642 
1643 	if (qs->qs_flags & (QS_FLUSHING|QS_TIMEOUT))
1644 		reclaim_completed_tx(qs, 0, TXQ_ETH);
1645 
1646 	if (!pi->link_config.link_ok) {
1647 		TXQ_RING_FLUSH(qs);
1648 		return;
1649 	}
1650 	TXQ_LOCK_ASSERT(qs);
1651 	while (!TXQ_RING_EMPTY(qs) && (ifp->if_drv_flags & IFF_DRV_RUNNING) &&
1652 	    pi->link_config.link_ok) {
1653 		reclaim_completed_tx(qs, cxgb_tx_reclaim_threshold, TXQ_ETH);
1654 
1655 		if (txq->size - txq->in_use <= TX_MAX_DESC)
1656 			break;
1657 
1658 		if ((m_head = cxgb_dequeue(qs)) == NULL)
1659 			break;
1660 		/*
1661 		 *  Encapsulation can modify our pointer, and or make it
1662 		 *  NULL on failure.  In that event, we can't requeue.
1663 		 */
1664 		if (t3_encap(qs, &m_head) || m_head == NULL)
1665 			break;
1666 
1667 		m_head = NULL;
1668 	}
1669 
1670 	if (txq->db_pending)
1671 		check_ring_tx_db(pi->adapter, txq, 1);
1672 
1673 	if (!TXQ_RING_EMPTY(qs) && callout_pending(&txq->txq_timer) == 0 &&
1674 	    pi->link_config.link_ok)
1675 		callout_reset_on(&txq->txq_timer, 1, cxgb_tx_timeout,
1676 		    qs, txq->txq_timer.c_cpu);
1677 	if (m_head != NULL)
1678 		m_freem(m_head);
1679 }
1680 
1681 static int
1682 cxgb_transmit_locked(struct ifnet *ifp, struct sge_qset *qs, struct mbuf *m)
1683 {
1684 	struct port_info *pi = qs->port;
1685 	struct sge_txq *txq = &qs->txq[TXQ_ETH];
1686 	struct buf_ring *br = txq->txq_mr;
1687 	int error, avail;
1688 
1689 	avail = txq->size - txq->in_use;
1690 	TXQ_LOCK_ASSERT(qs);
1691 
1692 	/*
1693 	 * We can only do a direct transmit if the following are true:
1694 	 * - we aren't coalescing (ring < 3/4 full)
1695 	 * - the link is up -- checked in caller
1696 	 * - there are no packets enqueued already
1697 	 * - there is space in hardware transmit queue
1698 	 */
1699 	if (check_pkt_coalesce(qs) == 0 &&
1700 	    !TXQ_RING_NEEDS_ENQUEUE(qs) && avail > TX_MAX_DESC) {
1701 		if (t3_encap(qs, &m)) {
1702 			if (m != NULL &&
1703 			    (error = drbr_enqueue(ifp, br, m)) != 0)
1704 				return (error);
1705 		} else {
1706 			if (txq->db_pending)
1707 				check_ring_tx_db(pi->adapter, txq, 1);
1708 
1709 			/*
1710 			 * We've bypassed the buf ring so we need to update
1711 			 * the stats directly
1712 			 */
1713 			txq->txq_direct_packets++;
1714 			txq->txq_direct_bytes += m->m_pkthdr.len;
1715 		}
1716 	} else if ((error = drbr_enqueue(ifp, br, m)) != 0)
1717 		return (error);
1718 
1719 	reclaim_completed_tx(qs, cxgb_tx_reclaim_threshold, TXQ_ETH);
1720 	if (!TXQ_RING_EMPTY(qs) && pi->link_config.link_ok &&
1721 	    (!check_pkt_coalesce(qs) || (drbr_inuse(ifp, br) >= 7)))
1722 		cxgb_start_locked(qs);
1723 	else if (!TXQ_RING_EMPTY(qs) && !callout_pending(&txq->txq_timer))
1724 		callout_reset_on(&txq->txq_timer, 1, cxgb_tx_timeout,
1725 		    qs, txq->txq_timer.c_cpu);
1726 	return (0);
1727 }
1728 
1729 int
1730 cxgb_transmit(struct ifnet *ifp, struct mbuf *m)
1731 {
1732 	struct sge_qset *qs;
1733 	struct port_info *pi = ifp->if_softc;
1734 	int error, qidx = pi->first_qset;
1735 
1736 	if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0
1737 	    ||(!pi->link_config.link_ok)) {
1738 		m_freem(m);
1739 		return (0);
1740 	}
1741 
1742 	if (m->m_flags & M_FLOWID)
1743 		qidx = (m->m_pkthdr.flowid % pi->nqsets) + pi->first_qset;
1744 
1745 	qs = &pi->adapter->sge.qs[qidx];
1746 
1747 	if (TXQ_TRYLOCK(qs)) {
1748 		/* XXX running */
1749 		error = cxgb_transmit_locked(ifp, qs, m);
1750 		TXQ_UNLOCK(qs);
1751 	} else
1752 		error = drbr_enqueue(ifp, qs->txq[TXQ_ETH].txq_mr, m);
1753 	return (error);
1754 }
1755 
1756 void
1757 cxgb_qflush(struct ifnet *ifp)
1758 {
1759 	/*
1760 	 * flush any enqueued mbufs in the buf_rings
1761 	 * and in the transmit queues
1762 	 * no-op for now
1763 	 */
1764 	return;
1765 }
1766 
1767 /**
1768  *	write_imm - write a packet into a Tx descriptor as immediate data
1769  *	@d: the Tx descriptor to write
1770  *	@m: the packet
1771  *	@len: the length of packet data to write as immediate data
1772  *	@gen: the generation bit value to write
1773  *
1774  *	Writes a packet as immediate data into a Tx descriptor.  The packet
1775  *	contains a work request at its beginning.  We must write the packet
1776  *	carefully so the SGE doesn't read accidentally before it's written in
1777  *	its entirety.
1778  */
1779 static __inline void
1780 write_imm(struct tx_desc *d, caddr_t src,
1781 	  unsigned int len, unsigned int gen)
1782 {
1783 	struct work_request_hdr *from = (struct work_request_hdr *)src;
1784 	struct work_request_hdr *to = (struct work_request_hdr *)d;
1785 	uint32_t wr_hi, wr_lo;
1786 
1787 	KASSERT(len <= WR_LEN && len >= sizeof(*from),
1788 	    ("%s: invalid len %d", __func__, len));
1789 
1790 	memcpy(&to[1], &from[1], len - sizeof(*from));
1791 	wr_hi = from->wrh_hi | htonl(F_WR_SOP | F_WR_EOP |
1792 	    V_WR_BCNTLFLT(len & 7));
1793 	wr_lo = from->wrh_lo | htonl(V_WR_GEN(gen) | V_WR_LEN((len + 7) / 8));
1794 	set_wr_hdr(to, wr_hi, wr_lo);
1795 	wmb();
1796 	wr_gen2(d, gen);
1797 }
1798 
1799 /**
1800  *	check_desc_avail - check descriptor availability on a send queue
1801  *	@adap: the adapter
1802  *	@q: the TX queue
1803  *	@m: the packet needing the descriptors
1804  *	@ndesc: the number of Tx descriptors needed
1805  *	@qid: the Tx queue number in its queue set (TXQ_OFLD or TXQ_CTRL)
1806  *
1807  *	Checks if the requested number of Tx descriptors is available on an
1808  *	SGE send queue.  If the queue is already suspended or not enough
1809  *	descriptors are available the packet is queued for later transmission.
1810  *	Must be called with the Tx queue locked.
1811  *
1812  *	Returns 0 if enough descriptors are available, 1 if there aren't
1813  *	enough descriptors and the packet has been queued, and 2 if the caller
1814  *	needs to retry because there weren't enough descriptors at the
1815  *	beginning of the call but some freed up in the mean time.
1816  */
1817 static __inline int
1818 check_desc_avail(adapter_t *adap, struct sge_txq *q,
1819 		 struct mbuf *m, unsigned int ndesc,
1820 		 unsigned int qid)
1821 {
1822 	/*
1823 	 * XXX We currently only use this for checking the control queue
1824 	 * the control queue is only used for binding qsets which happens
1825 	 * at init time so we are guaranteed enough descriptors
1826 	 */
1827 	if (__predict_false(!mbufq_empty(&q->sendq))) {
1828 addq_exit:	mbufq_tail(&q->sendq, m);
1829 		return 1;
1830 	}
1831 	if (__predict_false(q->size - q->in_use < ndesc)) {
1832 
1833 		struct sge_qset *qs = txq_to_qset(q, qid);
1834 
1835 		setbit(&qs->txq_stopped, qid);
1836 		if (should_restart_tx(q) &&
1837 		    test_and_clear_bit(qid, &qs->txq_stopped))
1838 			return 2;
1839 
1840 		q->stops++;
1841 		goto addq_exit;
1842 	}
1843 	return 0;
1844 }
1845 
1846 
1847 /**
1848  *	reclaim_completed_tx_imm - reclaim completed control-queue Tx descs
1849  *	@q: the SGE control Tx queue
1850  *
1851  *	This is a variant of reclaim_completed_tx() that is used for Tx queues
1852  *	that send only immediate data (presently just the control queues) and
1853  *	thus do not have any mbufs
1854  */
1855 static __inline void
1856 reclaim_completed_tx_imm(struct sge_txq *q)
1857 {
1858 	unsigned int reclaim = q->processed - q->cleaned;
1859 
1860 	q->in_use -= reclaim;
1861 	q->cleaned += reclaim;
1862 }
1863 
1864 /**
1865  *	ctrl_xmit - send a packet through an SGE control Tx queue
1866  *	@adap: the adapter
1867  *	@q: the control queue
1868  *	@m: the packet
1869  *
1870  *	Send a packet through an SGE control Tx queue.  Packets sent through
1871  *	a control queue must fit entirely as immediate data in a single Tx
1872  *	descriptor and have no page fragments.
1873  */
1874 static int
1875 ctrl_xmit(adapter_t *adap, struct sge_qset *qs, struct mbuf *m)
1876 {
1877 	int ret;
1878 	struct work_request_hdr *wrp = mtod(m, struct work_request_hdr *);
1879 	struct sge_txq *q = &qs->txq[TXQ_CTRL];
1880 
1881 	KASSERT(m->m_len <= WR_LEN, ("%s: bad tx data", __func__));
1882 
1883 	wrp->wrh_hi |= htonl(F_WR_SOP | F_WR_EOP);
1884 	wrp->wrh_lo = htonl(V_WR_TID(q->token));
1885 
1886 	TXQ_LOCK(qs);
1887 again:	reclaim_completed_tx_imm(q);
1888 
1889 	ret = check_desc_avail(adap, q, m, 1, TXQ_CTRL);
1890 	if (__predict_false(ret)) {
1891 		if (ret == 1) {
1892 			TXQ_UNLOCK(qs);
1893 			return (ENOSPC);
1894 		}
1895 		goto again;
1896 	}
1897 	write_imm(&q->desc[q->pidx], m->m_data, m->m_len, q->gen);
1898 
1899 	q->in_use++;
1900 	if (++q->pidx >= q->size) {
1901 		q->pidx = 0;
1902 		q->gen ^= 1;
1903 	}
1904 	TXQ_UNLOCK(qs);
1905 	wmb();
1906 	t3_write_reg(adap, A_SG_KDOORBELL,
1907 	    F_SELEGRCNTX | V_EGRCNTX(q->cntxt_id));
1908 
1909 	m_free(m);
1910 	return (0);
1911 }
1912 
1913 
1914 /**
1915  *	restart_ctrlq - restart a suspended control queue
1916  *	@qs: the queue set cotaining the control queue
1917  *
1918  *	Resumes transmission on a suspended Tx control queue.
1919  */
1920 static void
1921 restart_ctrlq(void *data, int npending)
1922 {
1923 	struct mbuf *m;
1924 	struct sge_qset *qs = (struct sge_qset *)data;
1925 	struct sge_txq *q = &qs->txq[TXQ_CTRL];
1926 	adapter_t *adap = qs->port->adapter;
1927 
1928 	TXQ_LOCK(qs);
1929 again:	reclaim_completed_tx_imm(q);
1930 
1931 	while (q->in_use < q->size &&
1932 	       (m = mbufq_dequeue(&q->sendq)) != NULL) {
1933 
1934 		write_imm(&q->desc[q->pidx], m->m_data, m->m_len, q->gen);
1935 		m_free(m);
1936 
1937 		if (++q->pidx >= q->size) {
1938 			q->pidx = 0;
1939 			q->gen ^= 1;
1940 		}
1941 		q->in_use++;
1942 	}
1943 	if (!mbufq_empty(&q->sendq)) {
1944 		setbit(&qs->txq_stopped, TXQ_CTRL);
1945 
1946 		if (should_restart_tx(q) &&
1947 		    test_and_clear_bit(TXQ_CTRL, &qs->txq_stopped))
1948 			goto again;
1949 		q->stops++;
1950 	}
1951 	TXQ_UNLOCK(qs);
1952 	t3_write_reg(adap, A_SG_KDOORBELL,
1953 		     F_SELEGRCNTX | V_EGRCNTX(q->cntxt_id));
1954 }
1955 
1956 
1957 /*
1958  * Send a management message through control queue 0
1959  */
1960 int
1961 t3_mgmt_tx(struct adapter *adap, struct mbuf *m)
1962 {
1963 	return ctrl_xmit(adap, &adap->sge.qs[0], m);
1964 }
1965 
1966 /**
1967  *	free_qset - free the resources of an SGE queue set
1968  *	@sc: the controller owning the queue set
1969  *	@q: the queue set
1970  *
1971  *	Release the HW and SW resources associated with an SGE queue set, such
1972  *	as HW contexts, packet buffers, and descriptor rings.  Traffic to the
1973  *	queue set must be quiesced prior to calling this.
1974  */
1975 static void
1976 t3_free_qset(adapter_t *sc, struct sge_qset *q)
1977 {
1978 	int i;
1979 
1980 	reclaim_completed_tx(q, 0, TXQ_ETH);
1981 	if (q->txq[TXQ_ETH].txq_mr != NULL)
1982 		buf_ring_free(q->txq[TXQ_ETH].txq_mr, M_DEVBUF);
1983 	if (q->txq[TXQ_ETH].txq_ifq != NULL) {
1984 		ifq_delete(q->txq[TXQ_ETH].txq_ifq);
1985 		free(q->txq[TXQ_ETH].txq_ifq, M_DEVBUF);
1986 	}
1987 
1988 	for (i = 0; i < SGE_RXQ_PER_SET; ++i) {
1989 		if (q->fl[i].desc) {
1990 			mtx_lock_spin(&sc->sge.reg_lock);
1991 			t3_sge_disable_fl(sc, q->fl[i].cntxt_id);
1992 			mtx_unlock_spin(&sc->sge.reg_lock);
1993 			bus_dmamap_unload(q->fl[i].desc_tag, q->fl[i].desc_map);
1994 			bus_dmamem_free(q->fl[i].desc_tag, q->fl[i].desc,
1995 					q->fl[i].desc_map);
1996 			bus_dma_tag_destroy(q->fl[i].desc_tag);
1997 			bus_dma_tag_destroy(q->fl[i].entry_tag);
1998 		}
1999 		if (q->fl[i].sdesc) {
2000 			free_rx_bufs(sc, &q->fl[i]);
2001 			free(q->fl[i].sdesc, M_DEVBUF);
2002 		}
2003 	}
2004 
2005 	mtx_unlock(&q->lock);
2006 	MTX_DESTROY(&q->lock);
2007 	for (i = 0; i < SGE_TXQ_PER_SET; i++) {
2008 		if (q->txq[i].desc) {
2009 			mtx_lock_spin(&sc->sge.reg_lock);
2010 			t3_sge_enable_ecntxt(sc, q->txq[i].cntxt_id, 0);
2011 			mtx_unlock_spin(&sc->sge.reg_lock);
2012 			bus_dmamap_unload(q->txq[i].desc_tag,
2013 					q->txq[i].desc_map);
2014 			bus_dmamem_free(q->txq[i].desc_tag, q->txq[i].desc,
2015 					q->txq[i].desc_map);
2016 			bus_dma_tag_destroy(q->txq[i].desc_tag);
2017 			bus_dma_tag_destroy(q->txq[i].entry_tag);
2018 		}
2019 		if (q->txq[i].sdesc) {
2020 			free(q->txq[i].sdesc, M_DEVBUF);
2021 		}
2022 	}
2023 
2024 	if (q->rspq.desc) {
2025 		mtx_lock_spin(&sc->sge.reg_lock);
2026 		t3_sge_disable_rspcntxt(sc, q->rspq.cntxt_id);
2027 		mtx_unlock_spin(&sc->sge.reg_lock);
2028 
2029 		bus_dmamap_unload(q->rspq.desc_tag, q->rspq.desc_map);
2030 		bus_dmamem_free(q->rspq.desc_tag, q->rspq.desc,
2031 			        q->rspq.desc_map);
2032 		bus_dma_tag_destroy(q->rspq.desc_tag);
2033 		MTX_DESTROY(&q->rspq.lock);
2034 	}
2035 
2036 #if defined(INET6) || defined(INET)
2037 	tcp_lro_free(&q->lro.ctrl);
2038 #endif
2039 
2040 	bzero(q, sizeof(*q));
2041 }
2042 
2043 /**
2044  *	t3_free_sge_resources - free SGE resources
2045  *	@sc: the adapter softc
2046  *
2047  *	Frees resources used by the SGE queue sets.
2048  */
2049 void
2050 t3_free_sge_resources(adapter_t *sc, int nqsets)
2051 {
2052 	int i;
2053 
2054 	for (i = 0; i < nqsets; ++i) {
2055 		TXQ_LOCK(&sc->sge.qs[i]);
2056 		t3_free_qset(sc, &sc->sge.qs[i]);
2057 	}
2058 }
2059 
2060 /**
2061  *	t3_sge_start - enable SGE
2062  *	@sc: the controller softc
2063  *
2064  *	Enables the SGE for DMAs.  This is the last step in starting packet
2065  *	transfers.
2066  */
2067 void
2068 t3_sge_start(adapter_t *sc)
2069 {
2070 	t3_set_reg_field(sc, A_SG_CONTROL, F_GLOBALENABLE, F_GLOBALENABLE);
2071 }
2072 
2073 /**
2074  *	t3_sge_stop - disable SGE operation
2075  *	@sc: the adapter
2076  *
2077  *	Disables the DMA engine.  This can be called in emeregencies (e.g.,
2078  *	from error interrupts) or from normal process context.  In the latter
2079  *	case it also disables any pending queue restart tasklets.  Note that
2080  *	if it is called in interrupt context it cannot disable the restart
2081  *	tasklets as it cannot wait, however the tasklets will have no effect
2082  *	since the doorbells are disabled and the driver will call this again
2083  *	later from process context, at which time the tasklets will be stopped
2084  *	if they are still running.
2085  */
2086 void
2087 t3_sge_stop(adapter_t *sc)
2088 {
2089 	int i, nqsets;
2090 
2091 	t3_set_reg_field(sc, A_SG_CONTROL, F_GLOBALENABLE, 0);
2092 
2093 	if (sc->tq == NULL)
2094 		return;
2095 
2096 	for (nqsets = i = 0; i < (sc)->params.nports; i++)
2097 		nqsets += sc->port[i].nqsets;
2098 #ifdef notyet
2099 	/*
2100 	 *
2101 	 * XXX
2102 	 */
2103 	for (i = 0; i < nqsets; ++i) {
2104 		struct sge_qset *qs = &sc->sge.qs[i];
2105 
2106 		taskqueue_drain(sc->tq, &qs->txq[TXQ_OFLD].qresume_task);
2107 		taskqueue_drain(sc->tq, &qs->txq[TXQ_CTRL].qresume_task);
2108 	}
2109 #endif
2110 }
2111 
2112 /**
2113  *	t3_free_tx_desc - reclaims Tx descriptors and their buffers
2114  *	@adapter: the adapter
2115  *	@q: the Tx queue to reclaim descriptors from
2116  *	@reclaimable: the number of descriptors to reclaim
2117  *      @m_vec_size: maximum number of buffers to reclaim
2118  *      @desc_reclaimed: returns the number of descriptors reclaimed
2119  *
2120  *	Reclaims Tx descriptors from an SGE Tx queue and frees the associated
2121  *	Tx buffers.  Called with the Tx queue lock held.
2122  *
2123  *      Returns number of buffers of reclaimed
2124  */
2125 void
2126 t3_free_tx_desc(struct sge_qset *qs, int reclaimable, int queue)
2127 {
2128 	struct tx_sw_desc *txsd;
2129 	unsigned int cidx, mask;
2130 	struct sge_txq *q = &qs->txq[queue];
2131 
2132 #ifdef T3_TRACE
2133 	T3_TRACE2(sc->tb[q->cntxt_id & 7],
2134 		  "reclaiming %u Tx descriptors at cidx %u", reclaimable, cidx);
2135 #endif
2136 	cidx = q->cidx;
2137 	mask = q->size - 1;
2138 	txsd = &q->sdesc[cidx];
2139 
2140 	mtx_assert(&qs->lock, MA_OWNED);
2141 	while (reclaimable--) {
2142 		prefetch(q->sdesc[(cidx + 1) & mask].m);
2143 		prefetch(q->sdesc[(cidx + 2) & mask].m);
2144 
2145 		if (txsd->m != NULL) {
2146 			if (txsd->flags & TX_SW_DESC_MAPPED) {
2147 				bus_dmamap_unload(q->entry_tag, txsd->map);
2148 				txsd->flags &= ~TX_SW_DESC_MAPPED;
2149 			}
2150 			m_freem_list(txsd->m);
2151 			txsd->m = NULL;
2152 		} else
2153 			q->txq_skipped++;
2154 
2155 		++txsd;
2156 		if (++cidx == q->size) {
2157 			cidx = 0;
2158 			txsd = q->sdesc;
2159 		}
2160 	}
2161 	q->cidx = cidx;
2162 
2163 }
2164 
2165 /**
2166  *	is_new_response - check if a response is newly written
2167  *	@r: the response descriptor
2168  *	@q: the response queue
2169  *
2170  *	Returns true if a response descriptor contains a yet unprocessed
2171  *	response.
2172  */
2173 static __inline int
2174 is_new_response(const struct rsp_desc *r,
2175     const struct sge_rspq *q)
2176 {
2177 	return (r->intr_gen & F_RSPD_GEN2) == q->gen;
2178 }
2179 
2180 #define RSPD_GTS_MASK  (F_RSPD_TXQ0_GTS | F_RSPD_TXQ1_GTS)
2181 #define RSPD_CTRL_MASK (RSPD_GTS_MASK | \
2182 			V_RSPD_TXQ0_CR(M_RSPD_TXQ0_CR) | \
2183 			V_RSPD_TXQ1_CR(M_RSPD_TXQ1_CR) | \
2184 			V_RSPD_TXQ2_CR(M_RSPD_TXQ2_CR))
2185 
2186 /* How long to delay the next interrupt in case of memory shortage, in 0.1us. */
2187 #define NOMEM_INTR_DELAY 2500
2188 
2189 #ifdef TCP_OFFLOAD
2190 /**
2191  *	write_ofld_wr - write an offload work request
2192  *	@adap: the adapter
2193  *	@m: the packet to send
2194  *	@q: the Tx queue
2195  *	@pidx: index of the first Tx descriptor to write
2196  *	@gen: the generation value to use
2197  *	@ndesc: number of descriptors the packet will occupy
2198  *
2199  *	Write an offload work request to send the supplied packet.  The packet
2200  *	data already carry the work request with most fields populated.
2201  */
2202 static void
2203 write_ofld_wr(adapter_t *adap, struct mbuf *m, struct sge_txq *q,
2204     unsigned int pidx, unsigned int gen, unsigned int ndesc)
2205 {
2206 	unsigned int sgl_flits, flits;
2207 	int i, idx, nsegs, wrlen;
2208 	struct work_request_hdr *from;
2209 	struct sg_ent *sgp, t3sgl[TX_MAX_SEGS / 2 + 1];
2210 	struct tx_desc *d = &q->desc[pidx];
2211 	struct txq_state txqs;
2212 	struct sglist_seg *segs;
2213 	struct ofld_hdr *oh = mtod(m, struct ofld_hdr *);
2214 	struct sglist *sgl;
2215 
2216 	from = (void *)(oh + 1);	/* Start of WR within mbuf */
2217 	wrlen = m->m_len - sizeof(*oh);
2218 
2219 	if (!(oh->flags & F_HDR_SGL)) {
2220 		write_imm(d, (caddr_t)from, wrlen, gen);
2221 
2222 		/*
2223 		 * mbuf with "real" immediate tx data will be enqueue_wr'd by
2224 		 * t3_push_frames and freed in wr_ack.  Others, like those sent
2225 		 * down by close_conn, t3_send_reset, etc. should be freed here.
2226 		 */
2227 		if (!(oh->flags & F_HDR_DF))
2228 			m_free(m);
2229 		return;
2230 	}
2231 
2232 	memcpy(&d->flit[1], &from[1], wrlen - sizeof(*from));
2233 
2234 	sgl = oh->sgl;
2235 	flits = wrlen / 8;
2236 	sgp = (ndesc == 1) ? (struct sg_ent *)&d->flit[flits] : t3sgl;
2237 
2238 	nsegs = sgl->sg_nseg;
2239 	segs = sgl->sg_segs;
2240 	for (idx = 0, i = 0; i < nsegs; i++) {
2241 		KASSERT(segs[i].ss_len, ("%s: 0 len in sgl", __func__));
2242 		if (i && idx == 0)
2243 			++sgp;
2244 		sgp->len[idx] = htobe32(segs[i].ss_len);
2245 		sgp->addr[idx] = htobe64(segs[i].ss_paddr);
2246 		idx ^= 1;
2247 	}
2248 	if (idx) {
2249 		sgp->len[idx] = 0;
2250 		sgp->addr[idx] = 0;
2251 	}
2252 
2253 	sgl_flits = sgl_len(nsegs);
2254 	txqs.gen = gen;
2255 	txqs.pidx = pidx;
2256 	txqs.compl = 0;
2257 
2258 	write_wr_hdr_sgl(ndesc, d, &txqs, q, t3sgl, flits, sgl_flits,
2259 	    from->wrh_hi, from->wrh_lo);
2260 }
2261 
2262 /**
2263  *	ofld_xmit - send a packet through an offload queue
2264  *	@adap: the adapter
2265  *	@q: the Tx offload queue
2266  *	@m: the packet
2267  *
2268  *	Send an offload packet through an SGE offload queue.
2269  */
2270 static int
2271 ofld_xmit(adapter_t *adap, struct sge_qset *qs, struct mbuf *m)
2272 {
2273 	int ret;
2274 	unsigned int ndesc;
2275 	unsigned int pidx, gen;
2276 	struct sge_txq *q = &qs->txq[TXQ_OFLD];
2277 	struct ofld_hdr *oh = mtod(m, struct ofld_hdr *);
2278 
2279 	ndesc = G_HDR_NDESC(oh->flags);
2280 
2281 	TXQ_LOCK(qs);
2282 again:	reclaim_completed_tx(qs, 16, TXQ_OFLD);
2283 	ret = check_desc_avail(adap, q, m, ndesc, TXQ_OFLD);
2284 	if (__predict_false(ret)) {
2285 		if (ret == 1) {
2286 			TXQ_UNLOCK(qs);
2287 			return (EINTR);
2288 		}
2289 		goto again;
2290 	}
2291 
2292 	gen = q->gen;
2293 	q->in_use += ndesc;
2294 	pidx = q->pidx;
2295 	q->pidx += ndesc;
2296 	if (q->pidx >= q->size) {
2297 		q->pidx -= q->size;
2298 		q->gen ^= 1;
2299 	}
2300 
2301 	write_ofld_wr(adap, m, q, pidx, gen, ndesc);
2302 	check_ring_tx_db(adap, q, 1);
2303 	TXQ_UNLOCK(qs);
2304 
2305 	return (0);
2306 }
2307 
2308 /**
2309  *	restart_offloadq - restart a suspended offload queue
2310  *	@qs: the queue set cotaining the offload queue
2311  *
2312  *	Resumes transmission on a suspended Tx offload queue.
2313  */
2314 static void
2315 restart_offloadq(void *data, int npending)
2316 {
2317 	struct mbuf *m;
2318 	struct sge_qset *qs = data;
2319 	struct sge_txq *q = &qs->txq[TXQ_OFLD];
2320 	adapter_t *adap = qs->port->adapter;
2321 	int cleaned;
2322 
2323 	TXQ_LOCK(qs);
2324 again:	cleaned = reclaim_completed_tx(qs, 16, TXQ_OFLD);
2325 
2326 	while ((m = mbufq_peek(&q->sendq)) != NULL) {
2327 		unsigned int gen, pidx;
2328 		struct ofld_hdr *oh = mtod(m, struct ofld_hdr *);
2329 		unsigned int ndesc = G_HDR_NDESC(oh->flags);
2330 
2331 		if (__predict_false(q->size - q->in_use < ndesc)) {
2332 			setbit(&qs->txq_stopped, TXQ_OFLD);
2333 			if (should_restart_tx(q) &&
2334 			    test_and_clear_bit(TXQ_OFLD, &qs->txq_stopped))
2335 				goto again;
2336 			q->stops++;
2337 			break;
2338 		}
2339 
2340 		gen = q->gen;
2341 		q->in_use += ndesc;
2342 		pidx = q->pidx;
2343 		q->pidx += ndesc;
2344 		if (q->pidx >= q->size) {
2345 			q->pidx -= q->size;
2346 			q->gen ^= 1;
2347 		}
2348 
2349 		(void)mbufq_dequeue(&q->sendq);
2350 		TXQ_UNLOCK(qs);
2351 		write_ofld_wr(adap, m, q, pidx, gen, ndesc);
2352 		TXQ_LOCK(qs);
2353 	}
2354 #if USE_GTS
2355 	set_bit(TXQ_RUNNING, &q->flags);
2356 	set_bit(TXQ_LAST_PKT_DB, &q->flags);
2357 #endif
2358 	TXQ_UNLOCK(qs);
2359 	wmb();
2360 	t3_write_reg(adap, A_SG_KDOORBELL,
2361 		     F_SELEGRCNTX | V_EGRCNTX(q->cntxt_id));
2362 }
2363 
2364 /**
2365  *	t3_offload_tx - send an offload packet
2366  *	@m: the packet
2367  *
2368  *	Sends an offload packet.  We use the packet priority to select the
2369  *	appropriate Tx queue as follows: bit 0 indicates whether the packet
2370  *	should be sent as regular or control, bits 1-3 select the queue set.
2371  */
2372 int
2373 t3_offload_tx(struct adapter *sc, struct mbuf *m)
2374 {
2375 	struct ofld_hdr *oh = mtod(m, struct ofld_hdr *);
2376 	struct sge_qset *qs = &sc->sge.qs[G_HDR_QSET(oh->flags)];
2377 
2378 	if (oh->flags & F_HDR_CTRL) {
2379 		m_adj(m, sizeof (*oh));	/* trim ofld_hdr off */
2380 		return (ctrl_xmit(sc, qs, m));
2381 	} else
2382 		return (ofld_xmit(sc, qs, m));
2383 }
2384 #endif
2385 
2386 static void
2387 restart_tx(struct sge_qset *qs)
2388 {
2389 	struct adapter *sc = qs->port->adapter;
2390 
2391 	if (isset(&qs->txq_stopped, TXQ_OFLD) &&
2392 	    should_restart_tx(&qs->txq[TXQ_OFLD]) &&
2393 	    test_and_clear_bit(TXQ_OFLD, &qs->txq_stopped)) {
2394 		qs->txq[TXQ_OFLD].restarts++;
2395 		taskqueue_enqueue(sc->tq, &qs->txq[TXQ_OFLD].qresume_task);
2396 	}
2397 
2398 	if (isset(&qs->txq_stopped, TXQ_CTRL) &&
2399 	    should_restart_tx(&qs->txq[TXQ_CTRL]) &&
2400 	    test_and_clear_bit(TXQ_CTRL, &qs->txq_stopped)) {
2401 		qs->txq[TXQ_CTRL].restarts++;
2402 		taskqueue_enqueue(sc->tq, &qs->txq[TXQ_CTRL].qresume_task);
2403 	}
2404 }
2405 
2406 /**
2407  *	t3_sge_alloc_qset - initialize an SGE queue set
2408  *	@sc: the controller softc
2409  *	@id: the queue set id
2410  *	@nports: how many Ethernet ports will be using this queue set
2411  *	@irq_vec_idx: the IRQ vector index for response queue interrupts
2412  *	@p: configuration parameters for this queue set
2413  *	@ntxq: number of Tx queues for the queue set
2414  *	@pi: port info for queue set
2415  *
2416  *	Allocate resources and initialize an SGE queue set.  A queue set
2417  *	comprises a response queue, two Rx free-buffer queues, and up to 3
2418  *	Tx queues.  The Tx queues are assigned roles in the order Ethernet
2419  *	queue, offload queue, and control queue.
2420  */
2421 int
2422 t3_sge_alloc_qset(adapter_t *sc, u_int id, int nports, int irq_vec_idx,
2423 		  const struct qset_params *p, int ntxq, struct port_info *pi)
2424 {
2425 	struct sge_qset *q = &sc->sge.qs[id];
2426 	int i, ret = 0;
2427 
2428 	MTX_INIT(&q->lock, q->namebuf, NULL, MTX_DEF);
2429 	q->port = pi;
2430 	q->adap = sc;
2431 
2432 	if ((q->txq[TXQ_ETH].txq_mr = buf_ring_alloc(cxgb_txq_buf_ring_size,
2433 	    M_DEVBUF, M_WAITOK, &q->lock)) == NULL) {
2434 		device_printf(sc->dev, "failed to allocate mbuf ring\n");
2435 		goto err;
2436 	}
2437 	if ((q->txq[TXQ_ETH].txq_ifq = malloc(sizeof(struct ifaltq), M_DEVBUF,
2438 	    M_NOWAIT | M_ZERO)) == NULL) {
2439 		device_printf(sc->dev, "failed to allocate ifq\n");
2440 		goto err;
2441 	}
2442 	ifq_init(q->txq[TXQ_ETH].txq_ifq, pi->ifp);
2443 	callout_init(&q->txq[TXQ_ETH].txq_timer, 1);
2444 	callout_init(&q->txq[TXQ_ETH].txq_watchdog, 1);
2445 	q->txq[TXQ_ETH].txq_timer.c_cpu = id % mp_ncpus;
2446 	q->txq[TXQ_ETH].txq_watchdog.c_cpu = id % mp_ncpus;
2447 
2448 	init_qset_cntxt(q, id);
2449 	q->idx = id;
2450 	if ((ret = alloc_ring(sc, p->fl_size, sizeof(struct rx_desc),
2451 		    sizeof(struct rx_sw_desc), &q->fl[0].phys_addr,
2452 		    &q->fl[0].desc, &q->fl[0].sdesc,
2453 		    &q->fl[0].desc_tag, &q->fl[0].desc_map,
2454 		    sc->rx_dmat, &q->fl[0].entry_tag)) != 0) {
2455 		printf("error %d from alloc ring fl0\n", ret);
2456 		goto err;
2457 	}
2458 
2459 	if ((ret = alloc_ring(sc, p->jumbo_size, sizeof(struct rx_desc),
2460 		    sizeof(struct rx_sw_desc), &q->fl[1].phys_addr,
2461 		    &q->fl[1].desc, &q->fl[1].sdesc,
2462 		    &q->fl[1].desc_tag, &q->fl[1].desc_map,
2463 		    sc->rx_jumbo_dmat, &q->fl[1].entry_tag)) != 0) {
2464 		printf("error %d from alloc ring fl1\n", ret);
2465 		goto err;
2466 	}
2467 
2468 	if ((ret = alloc_ring(sc, p->rspq_size, sizeof(struct rsp_desc), 0,
2469 		    &q->rspq.phys_addr, &q->rspq.desc, NULL,
2470 		    &q->rspq.desc_tag, &q->rspq.desc_map,
2471 		    NULL, NULL)) != 0) {
2472 		printf("error %d from alloc ring rspq\n", ret);
2473 		goto err;
2474 	}
2475 
2476 	snprintf(q->rspq.lockbuf, RSPQ_NAME_LEN, "t3 rspq lock %d:%d",
2477 	    device_get_unit(sc->dev), irq_vec_idx);
2478 	MTX_INIT(&q->rspq.lock, q->rspq.lockbuf, NULL, MTX_DEF);
2479 
2480 	for (i = 0; i < ntxq; ++i) {
2481 		size_t sz = i == TXQ_CTRL ? 0 : sizeof(struct tx_sw_desc);
2482 
2483 		if ((ret = alloc_ring(sc, p->txq_size[i],
2484 			    sizeof(struct tx_desc), sz,
2485 			    &q->txq[i].phys_addr, &q->txq[i].desc,
2486 			    &q->txq[i].sdesc, &q->txq[i].desc_tag,
2487 			    &q->txq[i].desc_map,
2488 			    sc->tx_dmat, &q->txq[i].entry_tag)) != 0) {
2489 			printf("error %d from alloc ring tx %i\n", ret, i);
2490 			goto err;
2491 		}
2492 		mbufq_init(&q->txq[i].sendq);
2493 		q->txq[i].gen = 1;
2494 		q->txq[i].size = p->txq_size[i];
2495 	}
2496 
2497 #ifdef TCP_OFFLOAD
2498 	TASK_INIT(&q->txq[TXQ_OFLD].qresume_task, 0, restart_offloadq, q);
2499 #endif
2500 	TASK_INIT(&q->txq[TXQ_CTRL].qresume_task, 0, restart_ctrlq, q);
2501 	TASK_INIT(&q->txq[TXQ_ETH].qreclaim_task, 0, sge_txq_reclaim_handler, q);
2502 	TASK_INIT(&q->txq[TXQ_OFLD].qreclaim_task, 0, sge_txq_reclaim_handler, q);
2503 
2504 	q->fl[0].gen = q->fl[1].gen = 1;
2505 	q->fl[0].size = p->fl_size;
2506 	q->fl[1].size = p->jumbo_size;
2507 
2508 	q->rspq.gen = 1;
2509 	q->rspq.cidx = 0;
2510 	q->rspq.size = p->rspq_size;
2511 
2512 	q->txq[TXQ_ETH].stop_thres = nports *
2513 	    flits_to_desc(sgl_len(TX_MAX_SEGS + 1) + 3);
2514 
2515 	q->fl[0].buf_size = MCLBYTES;
2516 	q->fl[0].zone = zone_pack;
2517 	q->fl[0].type = EXT_PACKET;
2518 
2519 	if (p->jumbo_buf_size ==  MJUM16BYTES) {
2520 		q->fl[1].zone = zone_jumbo16;
2521 		q->fl[1].type = EXT_JUMBO16;
2522 	} else if (p->jumbo_buf_size ==  MJUM9BYTES) {
2523 		q->fl[1].zone = zone_jumbo9;
2524 		q->fl[1].type = EXT_JUMBO9;
2525 	} else if (p->jumbo_buf_size ==  MJUMPAGESIZE) {
2526 		q->fl[1].zone = zone_jumbop;
2527 		q->fl[1].type = EXT_JUMBOP;
2528 	} else {
2529 		KASSERT(0, ("can't deal with jumbo_buf_size %d.", p->jumbo_buf_size));
2530 		ret = EDOOFUS;
2531 		goto err;
2532 	}
2533 	q->fl[1].buf_size = p->jumbo_buf_size;
2534 
2535 	/* Allocate and setup the lro_ctrl structure */
2536 	q->lro.enabled = !!(pi->ifp->if_capenable & IFCAP_LRO);
2537 #if defined(INET6) || defined(INET)
2538 	ret = tcp_lro_init(&q->lro.ctrl);
2539 	if (ret) {
2540 		printf("error %d from tcp_lro_init\n", ret);
2541 		goto err;
2542 	}
2543 #endif
2544 	q->lro.ctrl.ifp = pi->ifp;
2545 
2546 	mtx_lock_spin(&sc->sge.reg_lock);
2547 	ret = -t3_sge_init_rspcntxt(sc, q->rspq.cntxt_id, irq_vec_idx,
2548 				   q->rspq.phys_addr, q->rspq.size,
2549 				   q->fl[0].buf_size, 1, 0);
2550 	if (ret) {
2551 		printf("error %d from t3_sge_init_rspcntxt\n", ret);
2552 		goto err_unlock;
2553 	}
2554 
2555 	for (i = 0; i < SGE_RXQ_PER_SET; ++i) {
2556 		ret = -t3_sge_init_flcntxt(sc, q->fl[i].cntxt_id, 0,
2557 					  q->fl[i].phys_addr, q->fl[i].size,
2558 					  q->fl[i].buf_size, p->cong_thres, 1,
2559 					  0);
2560 		if (ret) {
2561 			printf("error %d from t3_sge_init_flcntxt for index i=%d\n", ret, i);
2562 			goto err_unlock;
2563 		}
2564 	}
2565 
2566 	ret = -t3_sge_init_ecntxt(sc, q->txq[TXQ_ETH].cntxt_id, USE_GTS,
2567 				 SGE_CNTXT_ETH, id, q->txq[TXQ_ETH].phys_addr,
2568 				 q->txq[TXQ_ETH].size, q->txq[TXQ_ETH].token,
2569 				 1, 0);
2570 	if (ret) {
2571 		printf("error %d from t3_sge_init_ecntxt\n", ret);
2572 		goto err_unlock;
2573 	}
2574 
2575 	if (ntxq > 1) {
2576 		ret = -t3_sge_init_ecntxt(sc, q->txq[TXQ_OFLD].cntxt_id,
2577 					 USE_GTS, SGE_CNTXT_OFLD, id,
2578 					 q->txq[TXQ_OFLD].phys_addr,
2579 					 q->txq[TXQ_OFLD].size, 0, 1, 0);
2580 		if (ret) {
2581 			printf("error %d from t3_sge_init_ecntxt\n", ret);
2582 			goto err_unlock;
2583 		}
2584 	}
2585 
2586 	if (ntxq > 2) {
2587 		ret = -t3_sge_init_ecntxt(sc, q->txq[TXQ_CTRL].cntxt_id, 0,
2588 					 SGE_CNTXT_CTRL, id,
2589 					 q->txq[TXQ_CTRL].phys_addr,
2590 					 q->txq[TXQ_CTRL].size,
2591 					 q->txq[TXQ_CTRL].token, 1, 0);
2592 		if (ret) {
2593 			printf("error %d from t3_sge_init_ecntxt\n", ret);
2594 			goto err_unlock;
2595 		}
2596 	}
2597 
2598 	mtx_unlock_spin(&sc->sge.reg_lock);
2599 	t3_update_qset_coalesce(q, p);
2600 
2601 	refill_fl(sc, &q->fl[0], q->fl[0].size);
2602 	refill_fl(sc, &q->fl[1], q->fl[1].size);
2603 	refill_rspq(sc, &q->rspq, q->rspq.size - 1);
2604 
2605 	t3_write_reg(sc, A_SG_GTS, V_RSPQ(q->rspq.cntxt_id) |
2606 		     V_NEWTIMER(q->rspq.holdoff_tmr));
2607 
2608 	return (0);
2609 
2610 err_unlock:
2611 	mtx_unlock_spin(&sc->sge.reg_lock);
2612 err:
2613 	TXQ_LOCK(q);
2614 	t3_free_qset(sc, q);
2615 
2616 	return (ret);
2617 }
2618 
2619 /*
2620  * Remove CPL_RX_PKT headers from the mbuf and reduce it to a regular mbuf with
2621  * ethernet data.  Hardware assistance with various checksums and any vlan tag
2622  * will also be taken into account here.
2623  */
2624 void
2625 t3_rx_eth(struct adapter *adap, struct mbuf *m, int ethpad)
2626 {
2627 	struct cpl_rx_pkt *cpl = (struct cpl_rx_pkt *)(mtod(m, uint8_t *) + ethpad);
2628 	struct port_info *pi = &adap->port[adap->rxpkt_map[cpl->iff]];
2629 	struct ifnet *ifp = pi->ifp;
2630 
2631 	if (cpl->vlan_valid) {
2632 		m->m_pkthdr.ether_vtag = ntohs(cpl->vlan);
2633 		m->m_flags |= M_VLANTAG;
2634 	}
2635 
2636 	m->m_pkthdr.rcvif = ifp;
2637 	m->m_pkthdr.header = mtod(m, uint8_t *) + sizeof(*cpl) + ethpad;
2638 	/*
2639 	 * adjust after conversion to mbuf chain
2640 	 */
2641 	m->m_pkthdr.len -= (sizeof(*cpl) + ethpad);
2642 	m->m_len -= (sizeof(*cpl) + ethpad);
2643 	m->m_data += (sizeof(*cpl) + ethpad);
2644 
2645 	if (!cpl->fragment && cpl->csum_valid && cpl->csum == 0xffff) {
2646 		struct ether_header *eh = mtod(m, void *);
2647 		uint16_t eh_type;
2648 
2649 		if (eh->ether_type == htons(ETHERTYPE_VLAN)) {
2650 			struct ether_vlan_header *evh = mtod(m, void *);
2651 
2652 			eh_type = evh->evl_proto;
2653 		} else
2654 			eh_type = eh->ether_type;
2655 
2656 		if (ifp->if_capenable & IFCAP_RXCSUM &&
2657 		    eh_type == htons(ETHERTYPE_IP)) {
2658 			m->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
2659 			    CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
2660 			m->m_pkthdr.csum_data = 0xffff;
2661 		} else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
2662 		    eh_type == htons(ETHERTYPE_IPV6)) {
2663 			m->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
2664 			    CSUM_PSEUDO_HDR);
2665 			m->m_pkthdr.csum_data = 0xffff;
2666 		}
2667 	}
2668 }
2669 
2670 /**
2671  *	get_packet - return the next ingress packet buffer from a free list
2672  *	@adap: the adapter that received the packet
2673  *	@drop_thres: # of remaining buffers before we start dropping packets
2674  *	@qs: the qset that the SGE free list holding the packet belongs to
2675  *      @mh: the mbuf header, contains a pointer to the head and tail of the mbuf chain
2676  *      @r: response descriptor
2677  *
2678  *	Get the next packet from a free list and complete setup of the
2679  *	sk_buff.  If the packet is small we make a copy and recycle the
2680  *	original buffer, otherwise we use the original buffer itself.  If a
2681  *	positive drop threshold is supplied packets are dropped and their
2682  *	buffers recycled if (a) the number of remaining buffers is under the
2683  *	threshold and the packet is too big to copy, or (b) the packet should
2684  *	be copied but there is no memory for the copy.
2685  */
2686 static int
2687 get_packet(adapter_t *adap, unsigned int drop_thres, struct sge_qset *qs,
2688     struct t3_mbuf_hdr *mh, struct rsp_desc *r)
2689 {
2690 
2691 	unsigned int len_cq =  ntohl(r->len_cq);
2692 	struct sge_fl *fl = (len_cq & F_RSPD_FLQ) ? &qs->fl[1] : &qs->fl[0];
2693 	int mask, cidx = fl->cidx;
2694 	struct rx_sw_desc *sd = &fl->sdesc[cidx];
2695 	uint32_t len = G_RSPD_LEN(len_cq);
2696 	uint32_t flags = M_EXT;
2697 	uint8_t sopeop = G_RSPD_SOP_EOP(ntohl(r->flags));
2698 	caddr_t cl;
2699 	struct mbuf *m;
2700 	int ret = 0;
2701 
2702 	mask = fl->size - 1;
2703 	prefetch(fl->sdesc[(cidx + 1) & mask].m);
2704 	prefetch(fl->sdesc[(cidx + 2) & mask].m);
2705 	prefetch(fl->sdesc[(cidx + 1) & mask].rxsd_cl);
2706 	prefetch(fl->sdesc[(cidx + 2) & mask].rxsd_cl);
2707 
2708 	fl->credits--;
2709 	bus_dmamap_sync(fl->entry_tag, sd->map, BUS_DMASYNC_POSTREAD);
2710 
2711 	if (recycle_enable && len <= SGE_RX_COPY_THRES &&
2712 	    sopeop == RSPQ_SOP_EOP) {
2713 		if ((m = m_gethdr(M_DONTWAIT, MT_DATA)) == NULL)
2714 			goto skip_recycle;
2715 		cl = mtod(m, void *);
2716 		memcpy(cl, sd->rxsd_cl, len);
2717 		recycle_rx_buf(adap, fl, fl->cidx);
2718 		m->m_pkthdr.len = m->m_len = len;
2719 		m->m_flags = 0;
2720 		mh->mh_head = mh->mh_tail = m;
2721 		ret = 1;
2722 		goto done;
2723 	} else {
2724 	skip_recycle:
2725 		bus_dmamap_unload(fl->entry_tag, sd->map);
2726 		cl = sd->rxsd_cl;
2727 		m = sd->m;
2728 
2729 		if ((sopeop == RSPQ_SOP_EOP) ||
2730 		    (sopeop == RSPQ_SOP))
2731 			flags |= M_PKTHDR;
2732 		m_init(m, fl->zone, fl->buf_size, M_NOWAIT, MT_DATA, flags);
2733 		if (fl->zone == zone_pack) {
2734 			/*
2735 			 * restore clobbered data pointer
2736 			 */
2737 			m->m_data = m->m_ext.ext_buf;
2738 		} else {
2739 			m_cljset(m, cl, fl->type);
2740 		}
2741 		m->m_len = len;
2742 	}
2743 	switch(sopeop) {
2744 	case RSPQ_SOP_EOP:
2745 		ret = 1;
2746 		/* FALLTHROUGH */
2747 	case RSPQ_SOP:
2748 		mh->mh_head = mh->mh_tail = m;
2749 		m->m_pkthdr.len = len;
2750 		break;
2751 	case RSPQ_EOP:
2752 		ret = 1;
2753 		/* FALLTHROUGH */
2754 	case RSPQ_NSOP_NEOP:
2755 		if (mh->mh_tail == NULL) {
2756 			log(LOG_ERR, "discarding intermediate descriptor entry\n");
2757 			m_freem(m);
2758 			break;
2759 		}
2760 		mh->mh_tail->m_next = m;
2761 		mh->mh_tail = m;
2762 		mh->mh_head->m_pkthdr.len += len;
2763 		break;
2764 	}
2765 	if (cxgb_debug)
2766 		printf("len=%d pktlen=%d\n", m->m_len, m->m_pkthdr.len);
2767 done:
2768 	if (++fl->cidx == fl->size)
2769 		fl->cidx = 0;
2770 
2771 	return (ret);
2772 }
2773 
2774 /**
2775  *	handle_rsp_cntrl_info - handles control information in a response
2776  *	@qs: the queue set corresponding to the response
2777  *	@flags: the response control flags
2778  *
2779  *	Handles the control information of an SGE response, such as GTS
2780  *	indications and completion credits for the queue set's Tx queues.
2781  *	HW coalesces credits, we don't do any extra SW coalescing.
2782  */
2783 static __inline void
2784 handle_rsp_cntrl_info(struct sge_qset *qs, uint32_t flags)
2785 {
2786 	unsigned int credits;
2787 
2788 #if USE_GTS
2789 	if (flags & F_RSPD_TXQ0_GTS)
2790 		clear_bit(TXQ_RUNNING, &qs->txq[TXQ_ETH].flags);
2791 #endif
2792 	credits = G_RSPD_TXQ0_CR(flags);
2793 	if (credits)
2794 		qs->txq[TXQ_ETH].processed += credits;
2795 
2796 	credits = G_RSPD_TXQ2_CR(flags);
2797 	if (credits)
2798 		qs->txq[TXQ_CTRL].processed += credits;
2799 
2800 # if USE_GTS
2801 	if (flags & F_RSPD_TXQ1_GTS)
2802 		clear_bit(TXQ_RUNNING, &qs->txq[TXQ_OFLD].flags);
2803 # endif
2804 	credits = G_RSPD_TXQ1_CR(flags);
2805 	if (credits)
2806 		qs->txq[TXQ_OFLD].processed += credits;
2807 
2808 }
2809 
2810 static void
2811 check_ring_db(adapter_t *adap, struct sge_qset *qs,
2812     unsigned int sleeping)
2813 {
2814 	;
2815 }
2816 
2817 /**
2818  *	process_responses - process responses from an SGE response queue
2819  *	@adap: the adapter
2820  *	@qs: the queue set to which the response queue belongs
2821  *	@budget: how many responses can be processed in this round
2822  *
2823  *	Process responses from an SGE response queue up to the supplied budget.
2824  *	Responses include received packets as well as credits and other events
2825  *	for the queues that belong to the response queue's queue set.
2826  *	A negative budget is effectively unlimited.
2827  *
2828  *	Additionally choose the interrupt holdoff time for the next interrupt
2829  *	on this queue.  If the system is under memory shortage use a fairly
2830  *	long delay to help recovery.
2831  */
2832 static int
2833 process_responses(adapter_t *adap, struct sge_qset *qs, int budget)
2834 {
2835 	struct sge_rspq *rspq = &qs->rspq;
2836 	struct rsp_desc *r = &rspq->desc[rspq->cidx];
2837 	int budget_left = budget;
2838 	unsigned int sleeping = 0;
2839 #if defined(INET6) || defined(INET)
2840 	int lro_enabled = qs->lro.enabled;
2841 	int skip_lro;
2842 	struct lro_ctrl *lro_ctrl = &qs->lro.ctrl;
2843 #endif
2844 	struct t3_mbuf_hdr *mh = &rspq->rspq_mh;
2845 #ifdef DEBUG
2846 	static int last_holdoff = 0;
2847 	if (cxgb_debug && rspq->holdoff_tmr != last_holdoff) {
2848 		printf("next_holdoff=%d\n", rspq->holdoff_tmr);
2849 		last_holdoff = rspq->holdoff_tmr;
2850 	}
2851 #endif
2852 	rspq->next_holdoff = rspq->holdoff_tmr;
2853 
2854 	while (__predict_true(budget_left && is_new_response(r, rspq))) {
2855 		int eth, eop = 0, ethpad = 0;
2856 		uint32_t flags = ntohl(r->flags);
2857 		uint32_t rss_hash = be32toh(r->rss_hdr.rss_hash_val);
2858 		uint8_t opcode = r->rss_hdr.opcode;
2859 
2860 		eth = (opcode == CPL_RX_PKT);
2861 
2862 		if (__predict_false(flags & F_RSPD_ASYNC_NOTIF)) {
2863 			struct mbuf *m;
2864 
2865 			if (cxgb_debug)
2866 				printf("async notification\n");
2867 
2868 			if (mh->mh_head == NULL) {
2869 				mh->mh_head = m_gethdr(M_DONTWAIT, MT_DATA);
2870 				m = mh->mh_head;
2871 			} else {
2872 				m = m_gethdr(M_DONTWAIT, MT_DATA);
2873 			}
2874 			if (m == NULL)
2875 				goto no_mem;
2876 
2877                         memcpy(mtod(m, char *), r, AN_PKT_SIZE);
2878 			m->m_len = m->m_pkthdr.len = AN_PKT_SIZE;
2879                         *mtod(m, char *) = CPL_ASYNC_NOTIF;
2880 			opcode = CPL_ASYNC_NOTIF;
2881 			eop = 1;
2882                         rspq->async_notif++;
2883 			goto skip;
2884 		} else if  (flags & F_RSPD_IMM_DATA_VALID) {
2885 			struct mbuf *m = m_gethdr(M_DONTWAIT, MT_DATA);
2886 
2887 			if (m == NULL) {
2888 		no_mem:
2889 				rspq->next_holdoff = NOMEM_INTR_DELAY;
2890 				budget_left--;
2891 				break;
2892 			}
2893 			if (mh->mh_head == NULL)
2894 				mh->mh_head = m;
2895                         else
2896 				mh->mh_tail->m_next = m;
2897 			mh->mh_tail = m;
2898 
2899 			get_imm_packet(adap, r, m);
2900 			mh->mh_head->m_pkthdr.len += m->m_len;
2901 			eop = 1;
2902 			rspq->imm_data++;
2903 		} else if (r->len_cq) {
2904 			int drop_thresh = eth ? SGE_RX_DROP_THRES : 0;
2905 
2906 			eop = get_packet(adap, drop_thresh, qs, mh, r);
2907 			if (eop) {
2908 				if (r->rss_hdr.hash_type && !adap->timestamp)
2909 					mh->mh_head->m_flags |= M_FLOWID;
2910 				mh->mh_head->m_pkthdr.flowid = rss_hash;
2911 			}
2912 
2913 			ethpad = 2;
2914 		} else {
2915 			rspq->pure_rsps++;
2916 		}
2917 	skip:
2918 		if (flags & RSPD_CTRL_MASK) {
2919 			sleeping |= flags & RSPD_GTS_MASK;
2920 			handle_rsp_cntrl_info(qs, flags);
2921 		}
2922 
2923 		if (!eth && eop) {
2924 			rspq->offload_pkts++;
2925 #ifdef TCP_OFFLOAD
2926 			adap->cpl_handler[opcode](qs, r, mh->mh_head);
2927 #else
2928 			m_freem(mh->mh_head);
2929 #endif
2930 			mh->mh_head = NULL;
2931 		} else if (eth && eop) {
2932 			struct mbuf *m = mh->mh_head;
2933 
2934 			t3_rx_eth(adap, m, ethpad);
2935 
2936 			/*
2937 			 * The T304 sends incoming packets on any qset.  If LRO
2938 			 * is also enabled, we could end up sending packet up
2939 			 * lro_ctrl->ifp's input.  That is incorrect.
2940 			 *
2941 			 * The mbuf's rcvif was derived from the cpl header and
2942 			 * is accurate.  Skip LRO and just use that.
2943 			 */
2944 #if defined(INET6) || defined(INET)
2945 			skip_lro = __predict_false(qs->port->ifp != m->m_pkthdr.rcvif);
2946 
2947 			if (lro_enabled && lro_ctrl->lro_cnt && !skip_lro
2948 			    && (tcp_lro_rx(lro_ctrl, m, 0) == 0)
2949 			    ) {
2950 				/* successfully queue'd for LRO */
2951 			} else
2952 #endif
2953 			{
2954 				/*
2955 				 * LRO not enabled, packet unsuitable for LRO,
2956 				 * or unable to queue.  Pass it up right now in
2957 				 * either case.
2958 				 */
2959 				struct ifnet *ifp = m->m_pkthdr.rcvif;
2960 				(*ifp->if_input)(ifp, m);
2961 			}
2962 			mh->mh_head = NULL;
2963 
2964 		}
2965 
2966 		r++;
2967 		if (__predict_false(++rspq->cidx == rspq->size)) {
2968 			rspq->cidx = 0;
2969 			rspq->gen ^= 1;
2970 			r = rspq->desc;
2971 		}
2972 
2973 		if (++rspq->credits >= 64) {
2974 			refill_rspq(adap, rspq, rspq->credits);
2975 			rspq->credits = 0;
2976 		}
2977 		__refill_fl_lt(adap, &qs->fl[0], 32);
2978 		__refill_fl_lt(adap, &qs->fl[1], 32);
2979 		--budget_left;
2980 	}
2981 
2982 #if defined(INET6) || defined(INET)
2983 	/* Flush LRO */
2984 	while (!SLIST_EMPTY(&lro_ctrl->lro_active)) {
2985 		struct lro_entry *queued = SLIST_FIRST(&lro_ctrl->lro_active);
2986 		SLIST_REMOVE_HEAD(&lro_ctrl->lro_active, next);
2987 		tcp_lro_flush(lro_ctrl, queued);
2988 	}
2989 #endif
2990 
2991 	if (sleeping)
2992 		check_ring_db(adap, qs, sleeping);
2993 
2994 	mb();  /* commit Tx queue processed updates */
2995 	if (__predict_false(qs->txq_stopped > 1))
2996 		restart_tx(qs);
2997 
2998 	__refill_fl_lt(adap, &qs->fl[0], 512);
2999 	__refill_fl_lt(adap, &qs->fl[1], 512);
3000 	budget -= budget_left;
3001 	return (budget);
3002 }
3003 
3004 /*
3005  * A helper function that processes responses and issues GTS.
3006  */
3007 static __inline int
3008 process_responses_gts(adapter_t *adap, struct sge_rspq *rq)
3009 {
3010 	int work;
3011 	static int last_holdoff = 0;
3012 
3013 	work = process_responses(adap, rspq_to_qset(rq), -1);
3014 
3015 	if (cxgb_debug && (rq->next_holdoff != last_holdoff)) {
3016 		printf("next_holdoff=%d\n", rq->next_holdoff);
3017 		last_holdoff = rq->next_holdoff;
3018 	}
3019 	t3_write_reg(adap, A_SG_GTS, V_RSPQ(rq->cntxt_id) |
3020 	    V_NEWTIMER(rq->next_holdoff) | V_NEWINDEX(rq->cidx));
3021 
3022 	return (work);
3023 }
3024 
3025 
3026 /*
3027  * Interrupt handler for legacy INTx interrupts for T3B-based cards.
3028  * Handles data events from SGE response queues as well as error and other
3029  * async events as they all use the same interrupt pin.  We use one SGE
3030  * response queue per port in this mode and protect all response queues with
3031  * queue 0's lock.
3032  */
3033 void
3034 t3b_intr(void *data)
3035 {
3036 	uint32_t i, map;
3037 	adapter_t *adap = data;
3038 	struct sge_rspq *q0 = &adap->sge.qs[0].rspq;
3039 
3040 	t3_write_reg(adap, A_PL_CLI, 0);
3041 	map = t3_read_reg(adap, A_SG_DATA_INTR);
3042 
3043 	if (!map)
3044 		return;
3045 
3046 	if (__predict_false(map & F_ERRINTR)) {
3047 		t3_write_reg(adap, A_PL_INT_ENABLE0, 0);
3048 		(void) t3_read_reg(adap, A_PL_INT_ENABLE0);
3049 		taskqueue_enqueue(adap->tq, &adap->slow_intr_task);
3050 	}
3051 
3052 	mtx_lock(&q0->lock);
3053 	for_each_port(adap, i)
3054 	    if (map & (1 << i))
3055 			process_responses_gts(adap, &adap->sge.qs[i].rspq);
3056 	mtx_unlock(&q0->lock);
3057 }
3058 
3059 /*
3060  * The MSI interrupt handler.  This needs to handle data events from SGE
3061  * response queues as well as error and other async events as they all use
3062  * the same MSI vector.  We use one SGE response queue per port in this mode
3063  * and protect all response queues with queue 0's lock.
3064  */
3065 void
3066 t3_intr_msi(void *data)
3067 {
3068 	adapter_t *adap = data;
3069 	struct sge_rspq *q0 = &adap->sge.qs[0].rspq;
3070 	int i, new_packets = 0;
3071 
3072 	mtx_lock(&q0->lock);
3073 
3074 	for_each_port(adap, i)
3075 	    if (process_responses_gts(adap, &adap->sge.qs[i].rspq))
3076 		    new_packets = 1;
3077 	mtx_unlock(&q0->lock);
3078 	if (new_packets == 0) {
3079 		t3_write_reg(adap, A_PL_INT_ENABLE0, 0);
3080 		(void) t3_read_reg(adap, A_PL_INT_ENABLE0);
3081 		taskqueue_enqueue(adap->tq, &adap->slow_intr_task);
3082 	}
3083 }
3084 
3085 void
3086 t3_intr_msix(void *data)
3087 {
3088 	struct sge_qset *qs = data;
3089 	adapter_t *adap = qs->port->adapter;
3090 	struct sge_rspq *rspq = &qs->rspq;
3091 
3092 	if (process_responses_gts(adap, rspq) == 0)
3093 		rspq->unhandled_irqs++;
3094 }
3095 
3096 #define QDUMP_SBUF_SIZE		32 * 400
3097 static int
3098 t3_dump_rspq(SYSCTL_HANDLER_ARGS)
3099 {
3100 	struct sge_rspq *rspq;
3101 	struct sge_qset *qs;
3102 	int i, err, dump_end, idx;
3103 	struct sbuf *sb;
3104 	struct rsp_desc *rspd;
3105 	uint32_t data[4];
3106 
3107 	rspq = arg1;
3108 	qs = rspq_to_qset(rspq);
3109 	if (rspq->rspq_dump_count == 0)
3110 		return (0);
3111 	if (rspq->rspq_dump_count > RSPQ_Q_SIZE) {
3112 		log(LOG_WARNING,
3113 		    "dump count is too large %d\n", rspq->rspq_dump_count);
3114 		rspq->rspq_dump_count = 0;
3115 		return (EINVAL);
3116 	}
3117 	if (rspq->rspq_dump_start > (RSPQ_Q_SIZE-1)) {
3118 		log(LOG_WARNING,
3119 		    "dump start of %d is greater than queue size\n",
3120 		    rspq->rspq_dump_start);
3121 		rspq->rspq_dump_start = 0;
3122 		return (EINVAL);
3123 	}
3124 	err = t3_sge_read_rspq(qs->port->adapter, rspq->cntxt_id, data);
3125 	if (err)
3126 		return (err);
3127 	err = sysctl_wire_old_buffer(req, 0);
3128 	if (err)
3129 		return (err);
3130 	sb = sbuf_new_for_sysctl(NULL, NULL, QDUMP_SBUF_SIZE, req);
3131 
3132 	sbuf_printf(sb, " \n index=%u size=%u MSI-X/RspQ=%u intr enable=%u intr armed=%u\n",
3133 	    (data[0] & 0xffff), data[0] >> 16, ((data[2] >> 20) & 0x3f),
3134 	    ((data[2] >> 26) & 1), ((data[2] >> 27) & 1));
3135 	sbuf_printf(sb, " generation=%u CQ mode=%u FL threshold=%u\n",
3136 	    ((data[2] >> 28) & 1), ((data[2] >> 31) & 1), data[3]);
3137 
3138 	sbuf_printf(sb, " start=%d -> end=%d\n", rspq->rspq_dump_start,
3139 	    (rspq->rspq_dump_start + rspq->rspq_dump_count) & (RSPQ_Q_SIZE-1));
3140 
3141 	dump_end = rspq->rspq_dump_start + rspq->rspq_dump_count;
3142 	for (i = rspq->rspq_dump_start; i < dump_end; i++) {
3143 		idx = i & (RSPQ_Q_SIZE-1);
3144 
3145 		rspd = &rspq->desc[idx];
3146 		sbuf_printf(sb, "\tidx=%04d opcode=%02x cpu_idx=%x hash_type=%x cq_idx=%x\n",
3147 		    idx, rspd->rss_hdr.opcode, rspd->rss_hdr.cpu_idx,
3148 		    rspd->rss_hdr.hash_type, be16toh(rspd->rss_hdr.cq_idx));
3149 		sbuf_printf(sb, "\trss_hash_val=%x flags=%08x len_cq=%x intr_gen=%x\n",
3150 		    rspd->rss_hdr.rss_hash_val, be32toh(rspd->flags),
3151 		    be32toh(rspd->len_cq), rspd->intr_gen);
3152 	}
3153 
3154 	err = sbuf_finish(sb);
3155 	/* Output a trailing NUL. */
3156 	if (err == 0)
3157 		err = SYSCTL_OUT(req, "", 1);
3158 	sbuf_delete(sb);
3159 	return (err);
3160 }
3161 
3162 static int
3163 t3_dump_txq_eth(SYSCTL_HANDLER_ARGS)
3164 {
3165 	struct sge_txq *txq;
3166 	struct sge_qset *qs;
3167 	int i, j, err, dump_end;
3168 	struct sbuf *sb;
3169 	struct tx_desc *txd;
3170 	uint32_t *WR, wr_hi, wr_lo, gen;
3171 	uint32_t data[4];
3172 
3173 	txq = arg1;
3174 	qs = txq_to_qset(txq, TXQ_ETH);
3175 	if (txq->txq_dump_count == 0) {
3176 		return (0);
3177 	}
3178 	if (txq->txq_dump_count > TX_ETH_Q_SIZE) {
3179 		log(LOG_WARNING,
3180 		    "dump count is too large %d\n", txq->txq_dump_count);
3181 		txq->txq_dump_count = 1;
3182 		return (EINVAL);
3183 	}
3184 	if (txq->txq_dump_start > (TX_ETH_Q_SIZE-1)) {
3185 		log(LOG_WARNING,
3186 		    "dump start of %d is greater than queue size\n",
3187 		    txq->txq_dump_start);
3188 		txq->txq_dump_start = 0;
3189 		return (EINVAL);
3190 	}
3191 	err = t3_sge_read_ecntxt(qs->port->adapter, qs->rspq.cntxt_id, data);
3192 	if (err)
3193 		return (err);
3194 	err = sysctl_wire_old_buffer(req, 0);
3195 	if (err)
3196 		return (err);
3197 	sb = sbuf_new_for_sysctl(NULL, NULL, QDUMP_SBUF_SIZE, req);
3198 
3199 	sbuf_printf(sb, " \n credits=%u GTS=%u index=%u size=%u rspq#=%u cmdq#=%u\n",
3200 	    (data[0] & 0x7fff), ((data[0] >> 15) & 1), (data[0] >> 16),
3201 	    (data[1] & 0xffff), ((data[3] >> 4) & 7), ((data[3] >> 7) & 1));
3202 	sbuf_printf(sb, " TUN=%u TOE=%u generation%u uP token=%u valid=%u\n",
3203 	    ((data[3] >> 8) & 1), ((data[3] >> 9) & 1), ((data[3] >> 10) & 1),
3204 	    ((data[3] >> 11) & 0xfffff), ((data[3] >> 31) & 1));
3205 	sbuf_printf(sb, " qid=%d start=%d -> end=%d\n", qs->idx,
3206 	    txq->txq_dump_start,
3207 	    (txq->txq_dump_start + txq->txq_dump_count) & (TX_ETH_Q_SIZE-1));
3208 
3209 	dump_end = txq->txq_dump_start + txq->txq_dump_count;
3210 	for (i = txq->txq_dump_start; i < dump_end; i++) {
3211 		txd = &txq->desc[i & (TX_ETH_Q_SIZE-1)];
3212 		WR = (uint32_t *)txd->flit;
3213 		wr_hi = ntohl(WR[0]);
3214 		wr_lo = ntohl(WR[1]);
3215 		gen = G_WR_GEN(wr_lo);
3216 
3217 		sbuf_printf(sb," wr_hi %08x wr_lo %08x gen %d\n",
3218 		    wr_hi, wr_lo, gen);
3219 		for (j = 2; j < 30; j += 4)
3220 			sbuf_printf(sb, "\t%08x %08x %08x %08x \n",
3221 			    WR[j], WR[j + 1], WR[j + 2], WR[j + 3]);
3222 
3223 	}
3224 	err = sbuf_finish(sb);
3225 	/* Output a trailing NUL. */
3226 	if (err == 0)
3227 		err = SYSCTL_OUT(req, "", 1);
3228 	sbuf_delete(sb);
3229 	return (err);
3230 }
3231 
3232 static int
3233 t3_dump_txq_ctrl(SYSCTL_HANDLER_ARGS)
3234 {
3235 	struct sge_txq *txq;
3236 	struct sge_qset *qs;
3237 	int i, j, err, dump_end;
3238 	struct sbuf *sb;
3239 	struct tx_desc *txd;
3240 	uint32_t *WR, wr_hi, wr_lo, gen;
3241 
3242 	txq = arg1;
3243 	qs = txq_to_qset(txq, TXQ_CTRL);
3244 	if (txq->txq_dump_count == 0) {
3245 		return (0);
3246 	}
3247 	if (txq->txq_dump_count > 256) {
3248 		log(LOG_WARNING,
3249 		    "dump count is too large %d\n", txq->txq_dump_count);
3250 		txq->txq_dump_count = 1;
3251 		return (EINVAL);
3252 	}
3253 	if (txq->txq_dump_start > 255) {
3254 		log(LOG_WARNING,
3255 		    "dump start of %d is greater than queue size\n",
3256 		    txq->txq_dump_start);
3257 		txq->txq_dump_start = 0;
3258 		return (EINVAL);
3259 	}
3260 
3261 	err = sysctl_wire_old_buffer(req, 0);
3262 	if (err != 0)
3263 		return (err);
3264 	sb = sbuf_new_for_sysctl(NULL, NULL, QDUMP_SBUF_SIZE, req);
3265 	sbuf_printf(sb, " qid=%d start=%d -> end=%d\n", qs->idx,
3266 	    txq->txq_dump_start,
3267 	    (txq->txq_dump_start + txq->txq_dump_count) & 255);
3268 
3269 	dump_end = txq->txq_dump_start + txq->txq_dump_count;
3270 	for (i = txq->txq_dump_start; i < dump_end; i++) {
3271 		txd = &txq->desc[i & (255)];
3272 		WR = (uint32_t *)txd->flit;
3273 		wr_hi = ntohl(WR[0]);
3274 		wr_lo = ntohl(WR[1]);
3275 		gen = G_WR_GEN(wr_lo);
3276 
3277 		sbuf_printf(sb," wr_hi %08x wr_lo %08x gen %d\n",
3278 		    wr_hi, wr_lo, gen);
3279 		for (j = 2; j < 30; j += 4)
3280 			sbuf_printf(sb, "\t%08x %08x %08x %08x \n",
3281 			    WR[j], WR[j + 1], WR[j + 2], WR[j + 3]);
3282 
3283 	}
3284 	err = sbuf_finish(sb);
3285 	/* Output a trailing NUL. */
3286 	if (err == 0)
3287 		err = SYSCTL_OUT(req, "", 1);
3288 	sbuf_delete(sb);
3289 	return (err);
3290 }
3291 
3292 static int
3293 t3_set_coalesce_usecs(SYSCTL_HANDLER_ARGS)
3294 {
3295 	adapter_t *sc = arg1;
3296 	struct qset_params *qsp = &sc->params.sge.qset[0];
3297 	int coalesce_usecs;
3298 	struct sge_qset *qs;
3299 	int i, j, err, nqsets = 0;
3300 	struct mtx *lock;
3301 
3302 	if ((sc->flags & FULL_INIT_DONE) == 0)
3303 		return (ENXIO);
3304 
3305 	coalesce_usecs = qsp->coalesce_usecs;
3306         err = sysctl_handle_int(oidp, &coalesce_usecs, arg2, req);
3307 
3308 	if (err != 0) {
3309 		return (err);
3310 	}
3311 	if (coalesce_usecs == qsp->coalesce_usecs)
3312 		return (0);
3313 
3314 	for (i = 0; i < sc->params.nports; i++)
3315 		for (j = 0; j < sc->port[i].nqsets; j++)
3316 			nqsets++;
3317 
3318 	coalesce_usecs = max(1, coalesce_usecs);
3319 
3320 	for (i = 0; i < nqsets; i++) {
3321 		qs = &sc->sge.qs[i];
3322 		qsp = &sc->params.sge.qset[i];
3323 		qsp->coalesce_usecs = coalesce_usecs;
3324 
3325 		lock = (sc->flags & USING_MSIX) ? &qs->rspq.lock :
3326 			    &sc->sge.qs[0].rspq.lock;
3327 
3328 		mtx_lock(lock);
3329 		t3_update_qset_coalesce(qs, qsp);
3330 		t3_write_reg(sc, A_SG_GTS, V_RSPQ(qs->rspq.cntxt_id) |
3331 		    V_NEWTIMER(qs->rspq.holdoff_tmr));
3332 		mtx_unlock(lock);
3333 	}
3334 
3335 	return (0);
3336 }
3337 
3338 static int
3339 t3_pkt_timestamp(SYSCTL_HANDLER_ARGS)
3340 {
3341 	adapter_t *sc = arg1;
3342 	int rc, timestamp;
3343 
3344 	if ((sc->flags & FULL_INIT_DONE) == 0)
3345 		return (ENXIO);
3346 
3347 	timestamp = sc->timestamp;
3348 	rc = sysctl_handle_int(oidp, &timestamp, arg2, req);
3349 
3350 	if (rc != 0)
3351 		return (rc);
3352 
3353 	if (timestamp != sc->timestamp) {
3354 		t3_set_reg_field(sc, A_TP_PC_CONFIG2, F_ENABLERXPKTTMSTPRSS,
3355 		    timestamp ? F_ENABLERXPKTTMSTPRSS : 0);
3356 		sc->timestamp = timestamp;
3357 	}
3358 
3359 	return (0);
3360 }
3361 
3362 void
3363 t3_add_attach_sysctls(adapter_t *sc)
3364 {
3365 	struct sysctl_ctx_list *ctx;
3366 	struct sysctl_oid_list *children;
3367 
3368 	ctx = device_get_sysctl_ctx(sc->dev);
3369 	children = SYSCTL_CHILDREN(device_get_sysctl_tree(sc->dev));
3370 
3371 	/* random information */
3372 	SYSCTL_ADD_STRING(ctx, children, OID_AUTO,
3373 	    "firmware_version",
3374 	    CTLFLAG_RD, &sc->fw_version,
3375 	    0, "firmware version");
3376 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO,
3377 	    "hw_revision",
3378 	    CTLFLAG_RD, &sc->params.rev,
3379 	    0, "chip model");
3380 	SYSCTL_ADD_STRING(ctx, children, OID_AUTO,
3381 	    "port_types",
3382 	    CTLFLAG_RD, &sc->port_types,
3383 	    0, "type of ports");
3384 	SYSCTL_ADD_INT(ctx, children, OID_AUTO,
3385 	    "enable_debug",
3386 	    CTLFLAG_RW, &cxgb_debug,
3387 	    0, "enable verbose debugging output");
3388 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tunq_coalesce",
3389 	    CTLFLAG_RD, &sc->tunq_coalesce,
3390 	    "#tunneled packets freed");
3391 	SYSCTL_ADD_INT(ctx, children, OID_AUTO,
3392 	    "txq_overrun",
3393 	    CTLFLAG_RD, &txq_fills,
3394 	    0, "#times txq overrun");
3395 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO,
3396 	    "core_clock",
3397 	    CTLFLAG_RD, &sc->params.vpd.cclk,
3398 	    0, "core clock frequency (in KHz)");
3399 }
3400 
3401 
3402 static const char *rspq_name = "rspq";
3403 static const char *txq_names[] =
3404 {
3405 	"txq_eth",
3406 	"txq_ofld",
3407 	"txq_ctrl"
3408 };
3409 
3410 static int
3411 sysctl_handle_macstat(SYSCTL_HANDLER_ARGS)
3412 {
3413 	struct port_info *p = arg1;
3414 	uint64_t *parg;
3415 
3416 	if (!p)
3417 		return (EINVAL);
3418 
3419 	parg = (uint64_t *) ((uint8_t *)&p->mac.stats + arg2);
3420 	PORT_LOCK(p);
3421 	t3_mac_update_stats(&p->mac);
3422 	PORT_UNLOCK(p);
3423 
3424 	return (sysctl_handle_64(oidp, parg, 0, req));
3425 }
3426 
3427 void
3428 t3_add_configured_sysctls(adapter_t *sc)
3429 {
3430 	struct sysctl_ctx_list *ctx;
3431 	struct sysctl_oid_list *children;
3432 	int i, j;
3433 
3434 	ctx = device_get_sysctl_ctx(sc->dev);
3435 	children = SYSCTL_CHILDREN(device_get_sysctl_tree(sc->dev));
3436 
3437 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO,
3438 	    "intr_coal",
3439 	    CTLTYPE_INT|CTLFLAG_RW, sc,
3440 	    0, t3_set_coalesce_usecs,
3441 	    "I", "interrupt coalescing timer (us)");
3442 
3443 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO,
3444 	    "pkt_timestamp",
3445 	    CTLTYPE_INT | CTLFLAG_RW, sc,
3446 	    0, t3_pkt_timestamp,
3447 	    "I", "provide packet timestamp instead of connection hash");
3448 
3449 	for (i = 0; i < sc->params.nports; i++) {
3450 		struct port_info *pi = &sc->port[i];
3451 		struct sysctl_oid *poid;
3452 		struct sysctl_oid_list *poidlist;
3453 		struct mac_stats *mstats = &pi->mac.stats;
3454 
3455 		snprintf(pi->namebuf, PORT_NAME_LEN, "port%d", i);
3456 		poid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO,
3457 		    pi->namebuf, CTLFLAG_RD, NULL, "port statistics");
3458 		poidlist = SYSCTL_CHILDREN(poid);
3459 		SYSCTL_ADD_UINT(ctx, poidlist, OID_AUTO,
3460 		    "nqsets", CTLFLAG_RD, &pi->nqsets,
3461 		    0, "#queue sets");
3462 
3463 		for (j = 0; j < pi->nqsets; j++) {
3464 			struct sge_qset *qs = &sc->sge.qs[pi->first_qset + j];
3465 			struct sysctl_oid *qspoid, *rspqpoid, *txqpoid,
3466 					  *ctrlqpoid, *lropoid;
3467 			struct sysctl_oid_list *qspoidlist, *rspqpoidlist,
3468 					       *txqpoidlist, *ctrlqpoidlist,
3469 					       *lropoidlist;
3470 			struct sge_txq *txq = &qs->txq[TXQ_ETH];
3471 
3472 			snprintf(qs->namebuf, QS_NAME_LEN, "qs%d", j);
3473 
3474 			qspoid = SYSCTL_ADD_NODE(ctx, poidlist, OID_AUTO,
3475 			    qs->namebuf, CTLFLAG_RD, NULL, "qset statistics");
3476 			qspoidlist = SYSCTL_CHILDREN(qspoid);
3477 
3478 			SYSCTL_ADD_UINT(ctx, qspoidlist, OID_AUTO, "fl0_empty",
3479 					CTLFLAG_RD, &qs->fl[0].empty, 0,
3480 					"freelist #0 empty");
3481 			SYSCTL_ADD_UINT(ctx, qspoidlist, OID_AUTO, "fl1_empty",
3482 					CTLFLAG_RD, &qs->fl[1].empty, 0,
3483 					"freelist #1 empty");
3484 
3485 			rspqpoid = SYSCTL_ADD_NODE(ctx, qspoidlist, OID_AUTO,
3486 			    rspq_name, CTLFLAG_RD, NULL, "rspq statistics");
3487 			rspqpoidlist = SYSCTL_CHILDREN(rspqpoid);
3488 
3489 			txqpoid = SYSCTL_ADD_NODE(ctx, qspoidlist, OID_AUTO,
3490 			    txq_names[0], CTLFLAG_RD, NULL, "txq statistics");
3491 			txqpoidlist = SYSCTL_CHILDREN(txqpoid);
3492 
3493 			ctrlqpoid = SYSCTL_ADD_NODE(ctx, qspoidlist, OID_AUTO,
3494 			    txq_names[2], CTLFLAG_RD, NULL, "ctrlq statistics");
3495 			ctrlqpoidlist = SYSCTL_CHILDREN(ctrlqpoid);
3496 
3497 			lropoid = SYSCTL_ADD_NODE(ctx, qspoidlist, OID_AUTO,
3498 			    "lro_stats", CTLFLAG_RD, NULL, "LRO statistics");
3499 			lropoidlist = SYSCTL_CHILDREN(lropoid);
3500 
3501 			SYSCTL_ADD_UINT(ctx, rspqpoidlist, OID_AUTO, "size",
3502 			    CTLFLAG_RD, &qs->rspq.size,
3503 			    0, "#entries in response queue");
3504 			SYSCTL_ADD_UINT(ctx, rspqpoidlist, OID_AUTO, "cidx",
3505 			    CTLFLAG_RD, &qs->rspq.cidx,
3506 			    0, "consumer index");
3507 			SYSCTL_ADD_UINT(ctx, rspqpoidlist, OID_AUTO, "credits",
3508 			    CTLFLAG_RD, &qs->rspq.credits,
3509 			    0, "#credits");
3510 			SYSCTL_ADD_UINT(ctx, rspqpoidlist, OID_AUTO, "starved",
3511 			    CTLFLAG_RD, &qs->rspq.starved,
3512 			    0, "#times starved");
3513 			SYSCTL_ADD_ULONG(ctx, rspqpoidlist, OID_AUTO, "phys_addr",
3514 			    CTLFLAG_RD, &qs->rspq.phys_addr,
3515 			    "physical_address_of the queue");
3516 			SYSCTL_ADD_UINT(ctx, rspqpoidlist, OID_AUTO, "dump_start",
3517 			    CTLFLAG_RW, &qs->rspq.rspq_dump_start,
3518 			    0, "start rspq dump entry");
3519 			SYSCTL_ADD_UINT(ctx, rspqpoidlist, OID_AUTO, "dump_count",
3520 			    CTLFLAG_RW, &qs->rspq.rspq_dump_count,
3521 			    0, "#rspq entries to dump");
3522 			SYSCTL_ADD_PROC(ctx, rspqpoidlist, OID_AUTO, "qdump",
3523 			    CTLTYPE_STRING | CTLFLAG_RD, &qs->rspq,
3524 			    0, t3_dump_rspq, "A", "dump of the response queue");
3525 
3526 			SYSCTL_ADD_UQUAD(ctx, txqpoidlist, OID_AUTO, "dropped",
3527 			    CTLFLAG_RD, &qs->txq[TXQ_ETH].txq_mr->br_drops,
3528 			    "#tunneled packets dropped");
3529 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "sendqlen",
3530 			    CTLFLAG_RD, &qs->txq[TXQ_ETH].sendq.qlen,
3531 			    0, "#tunneled packets waiting to be sent");
3532 #if 0
3533 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "queue_pidx",
3534 			    CTLFLAG_RD, (uint32_t *)(uintptr_t)&qs->txq[TXQ_ETH].txq_mr.br_prod,
3535 			    0, "#tunneled packets queue producer index");
3536 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "queue_cidx",
3537 			    CTLFLAG_RD, (uint32_t *)(uintptr_t)&qs->txq[TXQ_ETH].txq_mr.br_cons,
3538 			    0, "#tunneled packets queue consumer index");
3539 #endif
3540 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "processed",
3541 			    CTLFLAG_RD, &qs->txq[TXQ_ETH].processed,
3542 			    0, "#tunneled packets processed by the card");
3543 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "cleaned",
3544 			    CTLFLAG_RD, &txq->cleaned,
3545 			    0, "#tunneled packets cleaned");
3546 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "in_use",
3547 			    CTLFLAG_RD, &txq->in_use,
3548 			    0, "#tunneled packet slots in use");
3549 			SYSCTL_ADD_ULONG(ctx, txqpoidlist, OID_AUTO, "frees",
3550 			    CTLFLAG_RD, &txq->txq_frees,
3551 			    "#tunneled packets freed");
3552 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "skipped",
3553 			    CTLFLAG_RD, &txq->txq_skipped,
3554 			    0, "#tunneled packet descriptors skipped");
3555 			SYSCTL_ADD_UQUAD(ctx, txqpoidlist, OID_AUTO, "coalesced",
3556 			    CTLFLAG_RD, &txq->txq_coalesced,
3557 			    "#tunneled packets coalesced");
3558 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "enqueued",
3559 			    CTLFLAG_RD, &txq->txq_enqueued,
3560 			    0, "#tunneled packets enqueued to hardware");
3561 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "stopped_flags",
3562 			    CTLFLAG_RD, &qs->txq_stopped,
3563 			    0, "tx queues stopped");
3564 			SYSCTL_ADD_ULONG(ctx, txqpoidlist, OID_AUTO, "phys_addr",
3565 			    CTLFLAG_RD, &txq->phys_addr,
3566 			    "physical_address_of the queue");
3567 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "qgen",
3568 			    CTLFLAG_RW, &qs->txq[TXQ_ETH].gen,
3569 			    0, "txq generation");
3570 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "hw_cidx",
3571 			    CTLFLAG_RD, &txq->cidx,
3572 			    0, "hardware queue cidx");
3573 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "hw_pidx",
3574 			    CTLFLAG_RD, &txq->pidx,
3575 			    0, "hardware queue pidx");
3576 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "dump_start",
3577 			    CTLFLAG_RW, &qs->txq[TXQ_ETH].txq_dump_start,
3578 			    0, "txq start idx for dump");
3579 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "dump_count",
3580 			    CTLFLAG_RW, &qs->txq[TXQ_ETH].txq_dump_count,
3581 			    0, "txq #entries to dump");
3582 			SYSCTL_ADD_PROC(ctx, txqpoidlist, OID_AUTO, "qdump",
3583 			    CTLTYPE_STRING | CTLFLAG_RD, &qs->txq[TXQ_ETH],
3584 			    0, t3_dump_txq_eth, "A", "dump of the transmit queue");
3585 
3586 			SYSCTL_ADD_UINT(ctx, ctrlqpoidlist, OID_AUTO, "dump_start",
3587 			    CTLFLAG_RW, &qs->txq[TXQ_CTRL].txq_dump_start,
3588 			    0, "ctrlq start idx for dump");
3589 			SYSCTL_ADD_UINT(ctx, ctrlqpoidlist, OID_AUTO, "dump_count",
3590 			    CTLFLAG_RW, &qs->txq[TXQ_CTRL].txq_dump_count,
3591 			    0, "ctrl #entries to dump");
3592 			SYSCTL_ADD_PROC(ctx, ctrlqpoidlist, OID_AUTO, "qdump",
3593 			    CTLTYPE_STRING | CTLFLAG_RD, &qs->txq[TXQ_CTRL],
3594 			    0, t3_dump_txq_ctrl, "A", "dump of the transmit queue");
3595 
3596 			SYSCTL_ADD_INT(ctx, lropoidlist, OID_AUTO, "lro_queued",
3597 			    CTLFLAG_RD, &qs->lro.ctrl.lro_queued, 0, NULL);
3598 			SYSCTL_ADD_INT(ctx, lropoidlist, OID_AUTO, "lro_flushed",
3599 			    CTLFLAG_RD, &qs->lro.ctrl.lro_flushed, 0, NULL);
3600 			SYSCTL_ADD_INT(ctx, lropoidlist, OID_AUTO, "lro_bad_csum",
3601 			    CTLFLAG_RD, &qs->lro.ctrl.lro_bad_csum, 0, NULL);
3602 			SYSCTL_ADD_INT(ctx, lropoidlist, OID_AUTO, "lro_cnt",
3603 			    CTLFLAG_RD, &qs->lro.ctrl.lro_cnt, 0, NULL);
3604 		}
3605 
3606 		/* Now add a node for mac stats. */
3607 		poid = SYSCTL_ADD_NODE(ctx, poidlist, OID_AUTO, "mac_stats",
3608 		    CTLFLAG_RD, NULL, "MAC statistics");
3609 		poidlist = SYSCTL_CHILDREN(poid);
3610 
3611 		/*
3612 		 * We (ab)use the length argument (arg2) to pass on the offset
3613 		 * of the data that we are interested in.  This is only required
3614 		 * for the quad counters that are updated from the hardware (we
3615 		 * make sure that we return the latest value).
3616 		 * sysctl_handle_macstat first updates *all* the counters from
3617 		 * the hardware, and then returns the latest value of the
3618 		 * requested counter.  Best would be to update only the
3619 		 * requested counter from hardware, but t3_mac_update_stats()
3620 		 * hides all the register details and we don't want to dive into
3621 		 * all that here.
3622 		 */
3623 #define CXGB_SYSCTL_ADD_QUAD(a)	SYSCTL_ADD_OID(ctx, poidlist, OID_AUTO, #a, \
3624     (CTLTYPE_U64 | CTLFLAG_RD), pi, offsetof(struct mac_stats, a), \
3625     sysctl_handle_macstat, "QU", 0)
3626 		CXGB_SYSCTL_ADD_QUAD(tx_octets);
3627 		CXGB_SYSCTL_ADD_QUAD(tx_octets_bad);
3628 		CXGB_SYSCTL_ADD_QUAD(tx_frames);
3629 		CXGB_SYSCTL_ADD_QUAD(tx_mcast_frames);
3630 		CXGB_SYSCTL_ADD_QUAD(tx_bcast_frames);
3631 		CXGB_SYSCTL_ADD_QUAD(tx_pause);
3632 		CXGB_SYSCTL_ADD_QUAD(tx_deferred);
3633 		CXGB_SYSCTL_ADD_QUAD(tx_late_collisions);
3634 		CXGB_SYSCTL_ADD_QUAD(tx_total_collisions);
3635 		CXGB_SYSCTL_ADD_QUAD(tx_excess_collisions);
3636 		CXGB_SYSCTL_ADD_QUAD(tx_underrun);
3637 		CXGB_SYSCTL_ADD_QUAD(tx_len_errs);
3638 		CXGB_SYSCTL_ADD_QUAD(tx_mac_internal_errs);
3639 		CXGB_SYSCTL_ADD_QUAD(tx_excess_deferral);
3640 		CXGB_SYSCTL_ADD_QUAD(tx_fcs_errs);
3641 		CXGB_SYSCTL_ADD_QUAD(tx_frames_64);
3642 		CXGB_SYSCTL_ADD_QUAD(tx_frames_65_127);
3643 		CXGB_SYSCTL_ADD_QUAD(tx_frames_128_255);
3644 		CXGB_SYSCTL_ADD_QUAD(tx_frames_256_511);
3645 		CXGB_SYSCTL_ADD_QUAD(tx_frames_512_1023);
3646 		CXGB_SYSCTL_ADD_QUAD(tx_frames_1024_1518);
3647 		CXGB_SYSCTL_ADD_QUAD(tx_frames_1519_max);
3648 		CXGB_SYSCTL_ADD_QUAD(rx_octets);
3649 		CXGB_SYSCTL_ADD_QUAD(rx_octets_bad);
3650 		CXGB_SYSCTL_ADD_QUAD(rx_frames);
3651 		CXGB_SYSCTL_ADD_QUAD(rx_mcast_frames);
3652 		CXGB_SYSCTL_ADD_QUAD(rx_bcast_frames);
3653 		CXGB_SYSCTL_ADD_QUAD(rx_pause);
3654 		CXGB_SYSCTL_ADD_QUAD(rx_fcs_errs);
3655 		CXGB_SYSCTL_ADD_QUAD(rx_align_errs);
3656 		CXGB_SYSCTL_ADD_QUAD(rx_symbol_errs);
3657 		CXGB_SYSCTL_ADD_QUAD(rx_data_errs);
3658 		CXGB_SYSCTL_ADD_QUAD(rx_sequence_errs);
3659 		CXGB_SYSCTL_ADD_QUAD(rx_runt);
3660 		CXGB_SYSCTL_ADD_QUAD(rx_jabber);
3661 		CXGB_SYSCTL_ADD_QUAD(rx_short);
3662 		CXGB_SYSCTL_ADD_QUAD(rx_too_long);
3663 		CXGB_SYSCTL_ADD_QUAD(rx_mac_internal_errs);
3664 		CXGB_SYSCTL_ADD_QUAD(rx_cong_drops);
3665 		CXGB_SYSCTL_ADD_QUAD(rx_frames_64);
3666 		CXGB_SYSCTL_ADD_QUAD(rx_frames_65_127);
3667 		CXGB_SYSCTL_ADD_QUAD(rx_frames_128_255);
3668 		CXGB_SYSCTL_ADD_QUAD(rx_frames_256_511);
3669 		CXGB_SYSCTL_ADD_QUAD(rx_frames_512_1023);
3670 		CXGB_SYSCTL_ADD_QUAD(rx_frames_1024_1518);
3671 		CXGB_SYSCTL_ADD_QUAD(rx_frames_1519_max);
3672 #undef CXGB_SYSCTL_ADD_QUAD
3673 
3674 #define CXGB_SYSCTL_ADD_ULONG(a) SYSCTL_ADD_ULONG(ctx, poidlist, OID_AUTO, #a, \
3675     CTLFLAG_RD, &mstats->a, 0)
3676 		CXGB_SYSCTL_ADD_ULONG(tx_fifo_parity_err);
3677 		CXGB_SYSCTL_ADD_ULONG(rx_fifo_parity_err);
3678 		CXGB_SYSCTL_ADD_ULONG(tx_fifo_urun);
3679 		CXGB_SYSCTL_ADD_ULONG(rx_fifo_ovfl);
3680 		CXGB_SYSCTL_ADD_ULONG(serdes_signal_loss);
3681 		CXGB_SYSCTL_ADD_ULONG(xaui_pcs_ctc_err);
3682 		CXGB_SYSCTL_ADD_ULONG(xaui_pcs_align_change);
3683 		CXGB_SYSCTL_ADD_ULONG(num_toggled);
3684 		CXGB_SYSCTL_ADD_ULONG(num_resets);
3685 		CXGB_SYSCTL_ADD_ULONG(link_faults);
3686 #undef CXGB_SYSCTL_ADD_ULONG
3687 	}
3688 }
3689 
3690 /**
3691  *	t3_get_desc - dump an SGE descriptor for debugging purposes
3692  *	@qs: the queue set
3693  *	@qnum: identifies the specific queue (0..2: Tx, 3:response, 4..5: Rx)
3694  *	@idx: the descriptor index in the queue
3695  *	@data: where to dump the descriptor contents
3696  *
3697  *	Dumps the contents of a HW descriptor of an SGE queue.  Returns the
3698  *	size of the descriptor.
3699  */
3700 int
3701 t3_get_desc(const struct sge_qset *qs, unsigned int qnum, unsigned int idx,
3702 		unsigned char *data)
3703 {
3704 	if (qnum >= 6)
3705 		return (EINVAL);
3706 
3707 	if (qnum < 3) {
3708 		if (!qs->txq[qnum].desc || idx >= qs->txq[qnum].size)
3709 			return -EINVAL;
3710 		memcpy(data, &qs->txq[qnum].desc[idx], sizeof(struct tx_desc));
3711 		return sizeof(struct tx_desc);
3712 	}
3713 
3714 	if (qnum == 3) {
3715 		if (!qs->rspq.desc || idx >= qs->rspq.size)
3716 			return (EINVAL);
3717 		memcpy(data, &qs->rspq.desc[idx], sizeof(struct rsp_desc));
3718 		return sizeof(struct rsp_desc);
3719 	}
3720 
3721 	qnum -= 4;
3722 	if (!qs->fl[qnum].desc || idx >= qs->fl[qnum].size)
3723 		return (EINVAL);
3724 	memcpy(data, &qs->fl[qnum].desc[idx], sizeof(struct rx_desc));
3725 	return sizeof(struct rx_desc);
3726 }
3727