xref: /freebsd/sys/dev/cxgb/cxgb_sge.c (revision 3fc36ee018bb836bd1796067cf4ef8683f166ebc)
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/if.h>
60 #include <net/if_var.h>
61 #include <net/bpf.h>
62 #include <net/ethernet.h>
63 #include <net/if_vlan_var.h>
64 
65 #include <netinet/in_systm.h>
66 #include <netinet/in.h>
67 #include <netinet/ip.h>
68 #include <netinet/ip6.h>
69 #include <netinet/tcp.h>
70 
71 #include <dev/pci/pcireg.h>
72 #include <dev/pci/pcivar.h>
73 
74 #include <vm/vm.h>
75 #include <vm/pmap.h>
76 
77 #include <cxgb_include.h>
78 #include <sys/mvec.h>
79 
80 int	txq_fills = 0;
81 int	multiq_tx_enable = 1;
82 
83 #ifdef TCP_OFFLOAD
84 CTASSERT(NUM_CPL_HANDLERS >= NUM_CPL_CMDS);
85 #endif
86 
87 extern struct sysctl_oid_list sysctl__hw_cxgb_children;
88 int cxgb_txq_buf_ring_size = TX_ETH_Q_SIZE;
89 SYSCTL_INT(_hw_cxgb, OID_AUTO, txq_mr_size, CTLFLAG_RDTUN, &cxgb_txq_buf_ring_size, 0,
90     "size of per-queue mbuf ring");
91 
92 static int cxgb_tx_coalesce_force = 0;
93 SYSCTL_INT(_hw_cxgb, OID_AUTO, tx_coalesce_force, CTLFLAG_RWTUN,
94     &cxgb_tx_coalesce_force, 0,
95     "coalesce small packets into a single work request regardless of ring state");
96 
97 #define	COALESCE_START_DEFAULT		TX_ETH_Q_SIZE>>1
98 #define	COALESCE_START_MAX		(TX_ETH_Q_SIZE-(TX_ETH_Q_SIZE>>3))
99 #define	COALESCE_STOP_DEFAULT		TX_ETH_Q_SIZE>>2
100 #define	COALESCE_STOP_MIN		TX_ETH_Q_SIZE>>5
101 #define	TX_RECLAIM_DEFAULT		TX_ETH_Q_SIZE>>5
102 #define	TX_RECLAIM_MAX			TX_ETH_Q_SIZE>>2
103 #define	TX_RECLAIM_MIN			TX_ETH_Q_SIZE>>6
104 
105 
106 static int cxgb_tx_coalesce_enable_start = COALESCE_START_DEFAULT;
107 SYSCTL_INT(_hw_cxgb, OID_AUTO, tx_coalesce_enable_start, CTLFLAG_RWTUN,
108     &cxgb_tx_coalesce_enable_start, 0,
109     "coalesce enable threshold");
110 static int cxgb_tx_coalesce_enable_stop = COALESCE_STOP_DEFAULT;
111 SYSCTL_INT(_hw_cxgb, OID_AUTO, tx_coalesce_enable_stop, CTLFLAG_RWTUN,
112     &cxgb_tx_coalesce_enable_stop, 0,
113     "coalesce disable threshold");
114 static int cxgb_tx_reclaim_threshold = TX_RECLAIM_DEFAULT;
115 SYSCTL_INT(_hw_cxgb, OID_AUTO, tx_reclaim_threshold, CTLFLAG_RWTUN,
116     &cxgb_tx_reclaim_threshold, 0,
117     "tx cleaning minimum threshold");
118 
119 /*
120  * XXX don't re-enable this until TOE stops assuming
121  * we have an m_ext
122  */
123 static int recycle_enable = 0;
124 
125 extern int cxgb_use_16k_clusters;
126 extern int nmbjumbop;
127 extern int nmbjumbo9;
128 extern int nmbjumbo16;
129 
130 #define USE_GTS 0
131 
132 #define SGE_RX_SM_BUF_SIZE	1536
133 #define SGE_RX_DROP_THRES	16
134 #define SGE_RX_COPY_THRES	128
135 
136 /*
137  * Period of the Tx buffer reclaim timer.  This timer does not need to run
138  * frequently as Tx buffers are usually reclaimed by new Tx packets.
139  */
140 #define TX_RECLAIM_PERIOD       (hz >> 1)
141 
142 /*
143  * Values for sge_txq.flags
144  */
145 enum {
146 	TXQ_RUNNING	= 1 << 0,  /* fetch engine is running */
147 	TXQ_LAST_PKT_DB = 1 << 1,  /* last packet rang the doorbell */
148 };
149 
150 struct tx_desc {
151 	uint64_t	flit[TX_DESC_FLITS];
152 } __packed;
153 
154 struct rx_desc {
155 	uint32_t	addr_lo;
156 	uint32_t	len_gen;
157 	uint32_t	gen2;
158 	uint32_t	addr_hi;
159 } __packed;
160 
161 struct rsp_desc {               /* response queue descriptor */
162 	struct rss_header	rss_hdr;
163 	uint32_t		flags;
164 	uint32_t		len_cq;
165 	uint8_t			imm_data[47];
166 	uint8_t			intr_gen;
167 } __packed;
168 
169 #define RX_SW_DESC_MAP_CREATED	(1 << 0)
170 #define TX_SW_DESC_MAP_CREATED	(1 << 1)
171 #define RX_SW_DESC_INUSE        (1 << 3)
172 #define TX_SW_DESC_MAPPED       (1 << 4)
173 
174 #define RSPQ_NSOP_NEOP           G_RSPD_SOP_EOP(0)
175 #define RSPQ_EOP                 G_RSPD_SOP_EOP(F_RSPD_EOP)
176 #define RSPQ_SOP                 G_RSPD_SOP_EOP(F_RSPD_SOP)
177 #define RSPQ_SOP_EOP             G_RSPD_SOP_EOP(F_RSPD_SOP|F_RSPD_EOP)
178 
179 struct tx_sw_desc {                /* SW state per Tx descriptor */
180 	struct mbuf	*m;
181 	bus_dmamap_t	map;
182 	int		flags;
183 };
184 
185 struct rx_sw_desc {                /* SW state per Rx descriptor */
186 	caddr_t		rxsd_cl;
187 	struct mbuf	*m;
188 	bus_dmamap_t	map;
189 	int		flags;
190 };
191 
192 struct txq_state {
193 	unsigned int	compl;
194 	unsigned int	gen;
195 	unsigned int	pidx;
196 };
197 
198 struct refill_fl_cb_arg {
199 	int               error;
200 	bus_dma_segment_t seg;
201 	int               nseg;
202 };
203 
204 
205 /*
206  * Maps a number of flits to the number of Tx descriptors that can hold them.
207  * The formula is
208  *
209  * desc = 1 + (flits - 2) / (WR_FLITS - 1).
210  *
211  * HW allows up to 4 descriptors to be combined into a WR.
212  */
213 static uint8_t flit_desc_map[] = {
214 	0,
215 #if SGE_NUM_GENBITS == 1
216 	1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
217 	2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
218 	3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
219 	4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4
220 #elif SGE_NUM_GENBITS == 2
221 	1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
222 	2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
223 	3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
224 	4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4,
225 #else
226 # error "SGE_NUM_GENBITS must be 1 or 2"
227 #endif
228 };
229 
230 #define	TXQ_LOCK_ASSERT(qs)	mtx_assert(&(qs)->lock, MA_OWNED)
231 #define	TXQ_TRYLOCK(qs)		mtx_trylock(&(qs)->lock)
232 #define	TXQ_LOCK(qs)		mtx_lock(&(qs)->lock)
233 #define	TXQ_UNLOCK(qs)		mtx_unlock(&(qs)->lock)
234 #define	TXQ_RING_EMPTY(qs)	drbr_empty((qs)->port->ifp, (qs)->txq[TXQ_ETH].txq_mr)
235 #define	TXQ_RING_NEEDS_ENQUEUE(qs)					\
236 	drbr_needs_enqueue((qs)->port->ifp, (qs)->txq[TXQ_ETH].txq_mr)
237 #define	TXQ_RING_FLUSH(qs)	drbr_flush((qs)->port->ifp, (qs)->txq[TXQ_ETH].txq_mr)
238 #define	TXQ_RING_DEQUEUE_COND(qs, func, arg)				\
239 	drbr_dequeue_cond((qs)->port->ifp, (qs)->txq[TXQ_ETH].txq_mr, func, arg)
240 #define	TXQ_RING_DEQUEUE(qs) \
241 	drbr_dequeue((qs)->port->ifp, (qs)->txq[TXQ_ETH].txq_mr)
242 
243 int cxgb_debug = 0;
244 
245 static void sge_timer_cb(void *arg);
246 static void sge_timer_reclaim(void *arg, int ncount);
247 static void sge_txq_reclaim_handler(void *arg, int ncount);
248 static void cxgb_start_locked(struct sge_qset *qs);
249 
250 /*
251  * XXX need to cope with bursty scheduling by looking at a wider
252  * window than we are now for determining the need for coalescing
253  *
254  */
255 static __inline uint64_t
256 check_pkt_coalesce(struct sge_qset *qs)
257 {
258         struct adapter *sc;
259         struct sge_txq *txq;
260 	uint8_t *fill;
261 
262 	if (__predict_false(cxgb_tx_coalesce_force))
263 		return (1);
264 	txq = &qs->txq[TXQ_ETH];
265         sc = qs->port->adapter;
266 	fill = &sc->tunq_fill[qs->idx];
267 
268 	if (cxgb_tx_coalesce_enable_start > COALESCE_START_MAX)
269 		cxgb_tx_coalesce_enable_start = COALESCE_START_MAX;
270 	if (cxgb_tx_coalesce_enable_stop < COALESCE_STOP_MIN)
271 		cxgb_tx_coalesce_enable_start = COALESCE_STOP_MIN;
272 	/*
273 	 * if the hardware transmit queue is more than 1/8 full
274 	 * we mark it as coalescing - we drop back from coalescing
275 	 * when we go below 1/32 full and there are no packets enqueued,
276 	 * this provides us with some degree of hysteresis
277 	 */
278         if (*fill != 0 && (txq->in_use <= cxgb_tx_coalesce_enable_stop) &&
279 	    TXQ_RING_EMPTY(qs) && (qs->coalescing == 0))
280                 *fill = 0;
281         else if (*fill == 0 && (txq->in_use >= cxgb_tx_coalesce_enable_start))
282                 *fill = 1;
283 
284 	return (sc->tunq_coalesce);
285 }
286 
287 #ifdef __LP64__
288 static void
289 set_wr_hdr(struct work_request_hdr *wrp, uint32_t wr_hi, uint32_t wr_lo)
290 {
291 	uint64_t wr_hilo;
292 #if _BYTE_ORDER == _LITTLE_ENDIAN
293 	wr_hilo = wr_hi;
294 	wr_hilo |= (((uint64_t)wr_lo)<<32);
295 #else
296 	wr_hilo = wr_lo;
297 	wr_hilo |= (((uint64_t)wr_hi)<<32);
298 #endif
299 	wrp->wrh_hilo = wr_hilo;
300 }
301 #else
302 static void
303 set_wr_hdr(struct work_request_hdr *wrp, uint32_t wr_hi, uint32_t wr_lo)
304 {
305 
306 	wrp->wrh_hi = wr_hi;
307 	wmb();
308 	wrp->wrh_lo = wr_lo;
309 }
310 #endif
311 
312 struct coalesce_info {
313 	int count;
314 	int nbytes;
315 };
316 
317 static int
318 coalesce_check(struct mbuf *m, void *arg)
319 {
320 	struct coalesce_info *ci = arg;
321 	int *count = &ci->count;
322 	int *nbytes = &ci->nbytes;
323 
324 	if ((*nbytes == 0) || ((*nbytes + m->m_len <= 10500) &&
325 		(*count < 7) && (m->m_next == NULL))) {
326 		*count += 1;
327 		*nbytes += m->m_len;
328 		return (1);
329 	}
330 	return (0);
331 }
332 
333 static struct mbuf *
334 cxgb_dequeue(struct sge_qset *qs)
335 {
336 	struct mbuf *m, *m_head, *m_tail;
337 	struct coalesce_info ci;
338 
339 
340 	if (check_pkt_coalesce(qs) == 0)
341 		return TXQ_RING_DEQUEUE(qs);
342 
343 	m_head = m_tail = NULL;
344 	ci.count = ci.nbytes = 0;
345 	do {
346 		m = TXQ_RING_DEQUEUE_COND(qs, coalesce_check, &ci);
347 		if (m_head == NULL) {
348 			m_tail = m_head = m;
349 		} else if (m != NULL) {
350 			m_tail->m_nextpkt = m;
351 			m_tail = m;
352 		}
353 	} while (m != NULL);
354 	if (ci.count > 7)
355 		panic("trying to coalesce %d packets in to one WR", ci.count);
356 	return (m_head);
357 }
358 
359 /**
360  *	reclaim_completed_tx - reclaims completed Tx descriptors
361  *	@adapter: the adapter
362  *	@q: the Tx queue to reclaim completed descriptors from
363  *
364  *	Reclaims Tx descriptors that the SGE has indicated it has processed,
365  *	and frees the associated buffers if possible.  Called with the Tx
366  *	queue's lock held.
367  */
368 static __inline int
369 reclaim_completed_tx(struct sge_qset *qs, int reclaim_min, int queue)
370 {
371 	struct sge_txq *q = &qs->txq[queue];
372 	int reclaim = desc_reclaimable(q);
373 
374 	if ((cxgb_tx_reclaim_threshold > TX_RECLAIM_MAX) ||
375 	    (cxgb_tx_reclaim_threshold < TX_RECLAIM_MIN))
376 		cxgb_tx_reclaim_threshold = TX_RECLAIM_DEFAULT;
377 
378 	if (reclaim < reclaim_min)
379 		return (0);
380 
381 	mtx_assert(&qs->lock, MA_OWNED);
382 	if (reclaim > 0) {
383 		t3_free_tx_desc(qs, reclaim, queue);
384 		q->cleaned += reclaim;
385 		q->in_use -= reclaim;
386 	}
387 	if (isset(&qs->txq_stopped, TXQ_ETH))
388                 clrbit(&qs->txq_stopped, TXQ_ETH);
389 
390 	return (reclaim);
391 }
392 
393 /**
394  *	should_restart_tx - are there enough resources to restart a Tx queue?
395  *	@q: the Tx queue
396  *
397  *	Checks if there are enough descriptors to restart a suspended Tx queue.
398  */
399 static __inline int
400 should_restart_tx(const struct sge_txq *q)
401 {
402 	unsigned int r = q->processed - q->cleaned;
403 
404 	return q->in_use - r < (q->size >> 1);
405 }
406 
407 /**
408  *	t3_sge_init - initialize SGE
409  *	@adap: the adapter
410  *	@p: the SGE parameters
411  *
412  *	Performs SGE initialization needed every time after a chip reset.
413  *	We do not initialize any of the queue sets here, instead the driver
414  *	top-level must request those individually.  We also do not enable DMA
415  *	here, that should be done after the queues have been set up.
416  */
417 void
418 t3_sge_init(adapter_t *adap, struct sge_params *p)
419 {
420 	u_int ctrl, ups;
421 
422 	ups = 0; /* = ffs(pci_resource_len(adap->pdev, 2) >> 12); */
423 
424 	ctrl = F_DROPPKT | V_PKTSHIFT(2) | F_FLMODE | F_AVOIDCQOVFL |
425 	       F_CQCRDTCTRL | F_CONGMODE | F_TNLFLMODE | F_FATLPERREN |
426 	       V_HOSTPAGESIZE(PAGE_SHIFT - 11) | F_BIGENDIANINGRESS |
427 	       V_USERSPACESIZE(ups ? ups - 1 : 0) | F_ISCSICOALESCING;
428 #if SGE_NUM_GENBITS == 1
429 	ctrl |= F_EGRGENCTRL;
430 #endif
431 	if (adap->params.rev > 0) {
432 		if (!(adap->flags & (USING_MSIX | USING_MSI)))
433 			ctrl |= F_ONEINTMULTQ | F_OPTONEINTMULTQ;
434 	}
435 	t3_write_reg(adap, A_SG_CONTROL, ctrl);
436 	t3_write_reg(adap, A_SG_EGR_RCQ_DRB_THRSH, V_HIRCQDRBTHRSH(512) |
437 		     V_LORCQDRBTHRSH(512));
438 	t3_write_reg(adap, A_SG_TIMER_TICK, core_ticks_per_usec(adap) / 10);
439 	t3_write_reg(adap, A_SG_CMDQ_CREDIT_TH, V_THRESHOLD(32) |
440 		     V_TIMEOUT(200 * core_ticks_per_usec(adap)));
441 	t3_write_reg(adap, A_SG_HI_DRB_HI_THRSH,
442 		     adap->params.rev < T3_REV_C ? 1000 : 500);
443 	t3_write_reg(adap, A_SG_HI_DRB_LO_THRSH, 256);
444 	t3_write_reg(adap, A_SG_LO_DRB_HI_THRSH, 1000);
445 	t3_write_reg(adap, A_SG_LO_DRB_LO_THRSH, 256);
446 	t3_write_reg(adap, A_SG_OCO_BASE, V_BASE1(0xfff));
447 	t3_write_reg(adap, A_SG_DRB_PRI_THRESH, 63 * 1024);
448 }
449 
450 
451 /**
452  *	sgl_len - calculates the size of an SGL of the given capacity
453  *	@n: the number of SGL entries
454  *
455  *	Calculates the number of flits needed for a scatter/gather list that
456  *	can hold the given number of entries.
457  */
458 static __inline unsigned int
459 sgl_len(unsigned int n)
460 {
461 	return ((3 * n) / 2 + (n & 1));
462 }
463 
464 /**
465  *	get_imm_packet - return the next ingress packet buffer from a response
466  *	@resp: the response descriptor containing the packet data
467  *
468  *	Return a packet containing the immediate data of the given response.
469  */
470 static int
471 get_imm_packet(adapter_t *sc, const struct rsp_desc *resp, struct mbuf *m)
472 {
473 
474 	if (resp->rss_hdr.opcode == CPL_RX_DATA) {
475 		const struct cpl_rx_data *cpl = (const void *)&resp->imm_data[0];
476 		m->m_len = sizeof(*cpl) + ntohs(cpl->len);
477 	} else if (resp->rss_hdr.opcode == CPL_RX_PKT) {
478 		const struct cpl_rx_pkt *cpl = (const void *)&resp->imm_data[0];
479 		m->m_len = sizeof(*cpl) + ntohs(cpl->len);
480 	} else
481 		m->m_len = IMMED_PKT_SIZE;
482 	m->m_ext.ext_buf = NULL;
483 	m->m_ext.ext_type = 0;
484 	memcpy(mtod(m, uint8_t *), resp->imm_data, m->m_len);
485 	return (0);
486 }
487 
488 static __inline u_int
489 flits_to_desc(u_int n)
490 {
491 	return (flit_desc_map[n]);
492 }
493 
494 #define SGE_PARERR (F_CPPARITYERROR | F_OCPARITYERROR | F_RCPARITYERROR | \
495 		    F_IRPARITYERROR | V_ITPARITYERROR(M_ITPARITYERROR) | \
496 		    V_FLPARITYERROR(M_FLPARITYERROR) | F_LODRBPARITYERROR | \
497 		    F_HIDRBPARITYERROR | F_LORCQPARITYERROR | \
498 		    F_HIRCQPARITYERROR)
499 #define SGE_FRAMINGERR (F_UC_REQ_FRAMINGERROR | F_R_REQ_FRAMINGERROR)
500 #define SGE_FATALERR (SGE_PARERR | SGE_FRAMINGERR | F_RSPQCREDITOVERFOW | \
501 		      F_RSPQDISABLED)
502 
503 /**
504  *	t3_sge_err_intr_handler - SGE async event interrupt handler
505  *	@adapter: the adapter
506  *
507  *	Interrupt handler for SGE asynchronous (non-data) events.
508  */
509 void
510 t3_sge_err_intr_handler(adapter_t *adapter)
511 {
512 	unsigned int v, status;
513 
514 	status = t3_read_reg(adapter, A_SG_INT_CAUSE);
515 	if (status & SGE_PARERR)
516 		CH_ALERT(adapter, "SGE parity error (0x%x)\n",
517 			 status & SGE_PARERR);
518 	if (status & SGE_FRAMINGERR)
519 		CH_ALERT(adapter, "SGE framing error (0x%x)\n",
520 			 status & SGE_FRAMINGERR);
521 	if (status & F_RSPQCREDITOVERFOW)
522 		CH_ALERT(adapter, "SGE response queue credit overflow\n");
523 
524 	if (status & F_RSPQDISABLED) {
525 		v = t3_read_reg(adapter, A_SG_RSPQ_FL_STATUS);
526 
527 		CH_ALERT(adapter,
528 			 "packet delivered to disabled response queue (0x%x)\n",
529 			 (v >> S_RSPQ0DISABLED) & 0xff);
530 	}
531 
532 	t3_write_reg(adapter, A_SG_INT_CAUSE, status);
533 	if (status & SGE_FATALERR)
534 		t3_fatal_err(adapter);
535 }
536 
537 void
538 t3_sge_prep(adapter_t *adap, struct sge_params *p)
539 {
540 	int i, nqsets, fl_q_size, jumbo_q_size, use_16k, jumbo_buf_size;
541 
542 	nqsets = min(SGE_QSETS / adap->params.nports, mp_ncpus);
543 	nqsets *= adap->params.nports;
544 
545 	fl_q_size = min(nmbclusters/(3*nqsets), FL_Q_SIZE);
546 
547 	while (!powerof2(fl_q_size))
548 		fl_q_size--;
549 
550 	use_16k = cxgb_use_16k_clusters != -1 ? cxgb_use_16k_clusters :
551 	    is_offload(adap);
552 
553 #if __FreeBSD_version >= 700111
554 	if (use_16k) {
555 		jumbo_q_size = min(nmbjumbo16/(3*nqsets), JUMBO_Q_SIZE);
556 		jumbo_buf_size = MJUM16BYTES;
557 	} else {
558 		jumbo_q_size = min(nmbjumbo9/(3*nqsets), JUMBO_Q_SIZE);
559 		jumbo_buf_size = MJUM9BYTES;
560 	}
561 #else
562 	jumbo_q_size = min(nmbjumbop/(3*nqsets), JUMBO_Q_SIZE);
563 	jumbo_buf_size = MJUMPAGESIZE;
564 #endif
565 	while (!powerof2(jumbo_q_size))
566 		jumbo_q_size--;
567 
568 	if (fl_q_size < (FL_Q_SIZE / 4) || jumbo_q_size < (JUMBO_Q_SIZE / 2))
569 		device_printf(adap->dev,
570 		    "Insufficient clusters and/or jumbo buffers.\n");
571 
572 	p->max_pkt_size = jumbo_buf_size - sizeof(struct cpl_rx_data);
573 
574 	for (i = 0; i < SGE_QSETS; ++i) {
575 		struct qset_params *q = p->qset + i;
576 
577 		if (adap->params.nports > 2) {
578 			q->coalesce_usecs = 50;
579 		} else {
580 #ifdef INVARIANTS
581 			q->coalesce_usecs = 10;
582 #else
583 			q->coalesce_usecs = 5;
584 #endif
585 		}
586 		q->polling = 0;
587 		q->rspq_size = RSPQ_Q_SIZE;
588 		q->fl_size = fl_q_size;
589 		q->jumbo_size = jumbo_q_size;
590 		q->jumbo_buf_size = jumbo_buf_size;
591 		q->txq_size[TXQ_ETH] = TX_ETH_Q_SIZE;
592 		q->txq_size[TXQ_OFLD] = is_offload(adap) ? TX_OFLD_Q_SIZE : 16;
593 		q->txq_size[TXQ_CTRL] = TX_CTRL_Q_SIZE;
594 		q->cong_thres = 0;
595 	}
596 }
597 
598 int
599 t3_sge_alloc(adapter_t *sc)
600 {
601 
602 	/* The parent tag. */
603 	if (bus_dma_tag_create( bus_get_dma_tag(sc->dev),/* PCI parent */
604 				1, 0,			/* algnmnt, boundary */
605 				BUS_SPACE_MAXADDR,	/* lowaddr */
606 				BUS_SPACE_MAXADDR,	/* highaddr */
607 				NULL, NULL,		/* filter, filterarg */
608 				BUS_SPACE_MAXSIZE_32BIT,/* maxsize */
609 				BUS_SPACE_UNRESTRICTED, /* nsegments */
610 				BUS_SPACE_MAXSIZE_32BIT,/* maxsegsize */
611 				0,			/* flags */
612 				NULL, NULL,		/* lock, lockarg */
613 				&sc->parent_dmat)) {
614 		device_printf(sc->dev, "Cannot allocate parent DMA tag\n");
615 		return (ENOMEM);
616 	}
617 
618 	/*
619 	 * DMA tag for normal sized RX frames
620 	 */
621 	if (bus_dma_tag_create(sc->parent_dmat, MCLBYTES, 0, BUS_SPACE_MAXADDR,
622 		BUS_SPACE_MAXADDR, NULL, NULL, MCLBYTES, 1,
623 		MCLBYTES, BUS_DMA_ALLOCNOW, NULL, NULL, &sc->rx_dmat)) {
624 		device_printf(sc->dev, "Cannot allocate RX DMA tag\n");
625 		return (ENOMEM);
626 	}
627 
628 	/*
629 	 * DMA tag for jumbo sized RX frames.
630 	 */
631 	if (bus_dma_tag_create(sc->parent_dmat, MJUM16BYTES, 0, BUS_SPACE_MAXADDR,
632 		BUS_SPACE_MAXADDR, NULL, NULL, MJUM16BYTES, 1, MJUM16BYTES,
633 		BUS_DMA_ALLOCNOW, NULL, NULL, &sc->rx_jumbo_dmat)) {
634 		device_printf(sc->dev, "Cannot allocate RX jumbo DMA tag\n");
635 		return (ENOMEM);
636 	}
637 
638 	/*
639 	 * DMA tag for TX frames.
640 	 */
641 	if (bus_dma_tag_create(sc->parent_dmat, 1, 0, BUS_SPACE_MAXADDR,
642 		BUS_SPACE_MAXADDR, NULL, NULL, TX_MAX_SIZE, TX_MAX_SEGS,
643 		TX_MAX_SIZE, BUS_DMA_ALLOCNOW,
644 		NULL, NULL, &sc->tx_dmat)) {
645 		device_printf(sc->dev, "Cannot allocate TX DMA tag\n");
646 		return (ENOMEM);
647 	}
648 
649 	return (0);
650 }
651 
652 int
653 t3_sge_free(struct adapter * sc)
654 {
655 
656 	if (sc->tx_dmat != NULL)
657 		bus_dma_tag_destroy(sc->tx_dmat);
658 
659 	if (sc->rx_jumbo_dmat != NULL)
660 		bus_dma_tag_destroy(sc->rx_jumbo_dmat);
661 
662 	if (sc->rx_dmat != NULL)
663 		bus_dma_tag_destroy(sc->rx_dmat);
664 
665 	if (sc->parent_dmat != NULL)
666 		bus_dma_tag_destroy(sc->parent_dmat);
667 
668 	return (0);
669 }
670 
671 void
672 t3_update_qset_coalesce(struct sge_qset *qs, const struct qset_params *p)
673 {
674 
675 	qs->rspq.holdoff_tmr = max(p->coalesce_usecs * 10, 1U);
676 	qs->rspq.polling = 0 /* p->polling */;
677 }
678 
679 #if !defined(__i386__) && !defined(__amd64__)
680 static void
681 refill_fl_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
682 {
683 	struct refill_fl_cb_arg *cb_arg = arg;
684 
685 	cb_arg->error = error;
686 	cb_arg->seg = segs[0];
687 	cb_arg->nseg = nseg;
688 
689 }
690 #endif
691 /**
692  *	refill_fl - refill an SGE free-buffer list
693  *	@sc: the controller softc
694  *	@q: the free-list to refill
695  *	@n: the number of new buffers to allocate
696  *
697  *	(Re)populate an SGE free-buffer list with up to @n new packet buffers.
698  *	The caller must assure that @n does not exceed the queue's capacity.
699  */
700 static void
701 refill_fl(adapter_t *sc, struct sge_fl *q, int n)
702 {
703 	struct rx_sw_desc *sd = &q->sdesc[q->pidx];
704 	struct rx_desc *d = &q->desc[q->pidx];
705 	struct refill_fl_cb_arg cb_arg;
706 	struct mbuf *m;
707 	caddr_t cl;
708 	int err;
709 
710 	cb_arg.error = 0;
711 	while (n--) {
712 		/*
713 		 * We allocate an uninitialized mbuf + cluster, mbuf is
714 		 * initialized after rx.
715 		 */
716 		if (q->zone == zone_pack) {
717 			if ((m = m_getcl(M_NOWAIT, MT_NOINIT, M_PKTHDR)) == NULL)
718 				break;
719 			cl = m->m_ext.ext_buf;
720 		} else {
721 			if ((cl = m_cljget(NULL, M_NOWAIT, q->buf_size)) == NULL)
722 				break;
723 			if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
724 				uma_zfree(q->zone, cl);
725 				break;
726 			}
727 		}
728 		if ((sd->flags & RX_SW_DESC_MAP_CREATED) == 0) {
729 			if ((err = bus_dmamap_create(q->entry_tag, 0, &sd->map))) {
730 				log(LOG_WARNING, "bus_dmamap_create failed %d\n", err);
731 				uma_zfree(q->zone, cl);
732 				goto done;
733 			}
734 			sd->flags |= RX_SW_DESC_MAP_CREATED;
735 		}
736 #if !defined(__i386__) && !defined(__amd64__)
737 		err = bus_dmamap_load(q->entry_tag, sd->map,
738 		    cl, q->buf_size, refill_fl_cb, &cb_arg, 0);
739 
740 		if (err != 0 || cb_arg.error) {
741 			if (q->zone != zone_pack)
742 				uma_zfree(q->zone, cl);
743 			m_free(m);
744 			goto done;
745 		}
746 #else
747 		cb_arg.seg.ds_addr = pmap_kextract((vm_offset_t)cl);
748 #endif
749 		sd->flags |= RX_SW_DESC_INUSE;
750 		sd->rxsd_cl = cl;
751 		sd->m = m;
752 		d->addr_lo = htobe32(cb_arg.seg.ds_addr & 0xffffffff);
753 		d->addr_hi = htobe32(((uint64_t)cb_arg.seg.ds_addr >>32) & 0xffffffff);
754 		d->len_gen = htobe32(V_FLD_GEN1(q->gen));
755 		d->gen2 = htobe32(V_FLD_GEN2(q->gen));
756 
757 		d++;
758 		sd++;
759 
760 		if (++q->pidx == q->size) {
761 			q->pidx = 0;
762 			q->gen ^= 1;
763 			sd = q->sdesc;
764 			d = q->desc;
765 		}
766 		q->credits++;
767 		q->db_pending++;
768 	}
769 
770 done:
771 	if (q->db_pending >= 32) {
772 		q->db_pending = 0;
773 		t3_write_reg(sc, A_SG_KDOORBELL, V_EGRCNTX(q->cntxt_id));
774 	}
775 }
776 
777 
778 /**
779  *	free_rx_bufs - free the Rx buffers on an SGE free list
780  *	@sc: the controle softc
781  *	@q: the SGE free list to clean up
782  *
783  *	Release the buffers on an SGE free-buffer Rx queue.  HW fetching from
784  *	this queue should be stopped before calling this function.
785  */
786 static void
787 free_rx_bufs(adapter_t *sc, struct sge_fl *q)
788 {
789 	u_int cidx = q->cidx;
790 
791 	while (q->credits--) {
792 		struct rx_sw_desc *d = &q->sdesc[cidx];
793 
794 		if (d->flags & RX_SW_DESC_INUSE) {
795 			bus_dmamap_unload(q->entry_tag, d->map);
796 			bus_dmamap_destroy(q->entry_tag, d->map);
797 			if (q->zone == zone_pack) {
798 				m_init(d->m, M_NOWAIT, MT_DATA, M_EXT);
799 				uma_zfree(zone_pack, d->m);
800 			} else {
801 				m_init(d->m, M_NOWAIT, MT_DATA, 0);
802 				uma_zfree(zone_mbuf, d->m);
803 				uma_zfree(q->zone, d->rxsd_cl);
804 			}
805 		}
806 
807 		d->rxsd_cl = NULL;
808 		d->m = NULL;
809 		if (++cidx == q->size)
810 			cidx = 0;
811 	}
812 }
813 
814 static __inline void
815 __refill_fl(adapter_t *adap, struct sge_fl *fl)
816 {
817 	refill_fl(adap, fl, min(16U, fl->size - fl->credits));
818 }
819 
820 static __inline void
821 __refill_fl_lt(adapter_t *adap, struct sge_fl *fl, int max)
822 {
823 	uint32_t reclaimable = fl->size - fl->credits;
824 
825 	if (reclaimable > 0)
826 		refill_fl(adap, fl, min(max, reclaimable));
827 }
828 
829 /**
830  *	recycle_rx_buf - recycle a receive buffer
831  *	@adapter: the adapter
832  *	@q: the SGE free list
833  *	@idx: index of buffer to recycle
834  *
835  *	Recycles the specified buffer on the given free list by adding it at
836  *	the next available slot on the list.
837  */
838 static void
839 recycle_rx_buf(adapter_t *adap, struct sge_fl *q, unsigned int idx)
840 {
841 	struct rx_desc *from = &q->desc[idx];
842 	struct rx_desc *to   = &q->desc[q->pidx];
843 
844 	q->sdesc[q->pidx] = q->sdesc[idx];
845 	to->addr_lo = from->addr_lo;        // already big endian
846 	to->addr_hi = from->addr_hi;        // likewise
847 	wmb();	/* necessary ? */
848 	to->len_gen = htobe32(V_FLD_GEN1(q->gen));
849 	to->gen2 = htobe32(V_FLD_GEN2(q->gen));
850 	q->credits++;
851 
852 	if (++q->pidx == q->size) {
853 		q->pidx = 0;
854 		q->gen ^= 1;
855 	}
856 	t3_write_reg(adap, A_SG_KDOORBELL, V_EGRCNTX(q->cntxt_id));
857 }
858 
859 static void
860 alloc_ring_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int error)
861 {
862 	uint32_t *addr;
863 
864 	addr = arg;
865 	*addr = segs[0].ds_addr;
866 }
867 
868 static int
869 alloc_ring(adapter_t *sc, size_t nelem, size_t elem_size, size_t sw_size,
870     bus_addr_t *phys, void *desc, void *sdesc, bus_dma_tag_t *tag,
871     bus_dmamap_t *map, bus_dma_tag_t parent_entry_tag, bus_dma_tag_t *entry_tag)
872 {
873 	size_t len = nelem * elem_size;
874 	void *s = NULL;
875 	void *p = NULL;
876 	int err;
877 
878 	if ((err = bus_dma_tag_create(sc->parent_dmat, PAGE_SIZE, 0,
879 				      BUS_SPACE_MAXADDR_32BIT,
880 				      BUS_SPACE_MAXADDR, NULL, NULL, len, 1,
881 				      len, 0, NULL, NULL, tag)) != 0) {
882 		device_printf(sc->dev, "Cannot allocate descriptor tag\n");
883 		return (ENOMEM);
884 	}
885 
886 	if ((err = bus_dmamem_alloc(*tag, (void **)&p, BUS_DMA_NOWAIT,
887 				    map)) != 0) {
888 		device_printf(sc->dev, "Cannot allocate descriptor memory\n");
889 		return (ENOMEM);
890 	}
891 
892 	bus_dmamap_load(*tag, *map, p, len, alloc_ring_cb, phys, 0);
893 	bzero(p, len);
894 	*(void **)desc = p;
895 
896 	if (sw_size) {
897 		len = nelem * sw_size;
898 		s = malloc(len, M_DEVBUF, M_WAITOK|M_ZERO);
899 		*(void **)sdesc = s;
900 	}
901 	if (parent_entry_tag == NULL)
902 		return (0);
903 
904 	if ((err = bus_dma_tag_create(parent_entry_tag, 1, 0,
905 				      BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR,
906 		                      NULL, NULL, TX_MAX_SIZE, TX_MAX_SEGS,
907 				      TX_MAX_SIZE, BUS_DMA_ALLOCNOW,
908 		                      NULL, NULL, entry_tag)) != 0) {
909 		device_printf(sc->dev, "Cannot allocate descriptor entry tag\n");
910 		return (ENOMEM);
911 	}
912 	return (0);
913 }
914 
915 static void
916 sge_slow_intr_handler(void *arg, int ncount)
917 {
918 	adapter_t *sc = arg;
919 
920 	t3_slow_intr_handler(sc);
921 	t3_write_reg(sc, A_PL_INT_ENABLE0, sc->slow_intr_mask);
922 	(void) t3_read_reg(sc, A_PL_INT_ENABLE0);
923 }
924 
925 /**
926  *	sge_timer_cb - perform periodic maintenance of an SGE qset
927  *	@data: the SGE queue set to maintain
928  *
929  *	Runs periodically from a timer to perform maintenance of an SGE queue
930  *	set.  It performs two tasks:
931  *
932  *	a) Cleans up any completed Tx descriptors that may still be pending.
933  *	Normal descriptor cleanup happens when new packets are added to a Tx
934  *	queue so this timer is relatively infrequent and does any cleanup only
935  *	if the Tx queue has not seen any new packets in a while.  We make a
936  *	best effort attempt to reclaim descriptors, in that we don't wait
937  *	around if we cannot get a queue's lock (which most likely is because
938  *	someone else is queueing new packets and so will also handle the clean
939  *	up).  Since control queues use immediate data exclusively we don't
940  *	bother cleaning them up here.
941  *
942  *	b) Replenishes Rx queues that have run out due to memory shortage.
943  *	Normally new Rx buffers are added when existing ones are consumed but
944  *	when out of memory a queue can become empty.  We try to add only a few
945  *	buffers here, the queue will be replenished fully as these new buffers
946  *	are used up if memory shortage has subsided.
947  *
948  *	c) Return coalesced response queue credits in case a response queue is
949  *	starved.
950  *
951  *	d) Ring doorbells for T304 tunnel queues since we have seen doorbell
952  *	fifo overflows and the FW doesn't implement any recovery scheme yet.
953  */
954 static void
955 sge_timer_cb(void *arg)
956 {
957 	adapter_t *sc = arg;
958 	if ((sc->flags & USING_MSIX) == 0) {
959 
960 		struct port_info *pi;
961 		struct sge_qset *qs;
962 		struct sge_txq  *txq;
963 		int i, j;
964 		int reclaim_ofl, refill_rx;
965 
966 		if (sc->open_device_map == 0)
967 			return;
968 
969 		for (i = 0; i < sc->params.nports; i++) {
970 			pi = &sc->port[i];
971 			for (j = 0; j < pi->nqsets; j++) {
972 				qs = &sc->sge.qs[pi->first_qset + j];
973 				txq = &qs->txq[0];
974 				reclaim_ofl = txq[TXQ_OFLD].processed - txq[TXQ_OFLD].cleaned;
975 				refill_rx = ((qs->fl[0].credits < qs->fl[0].size) ||
976 				    (qs->fl[1].credits < qs->fl[1].size));
977 				if (reclaim_ofl || refill_rx) {
978 					taskqueue_enqueue(sc->tq, &pi->timer_reclaim_task);
979 					break;
980 				}
981 			}
982 		}
983 	}
984 
985 	if (sc->params.nports > 2) {
986 		int i;
987 
988 		for_each_port(sc, i) {
989 			struct port_info *pi = &sc->port[i];
990 
991 			t3_write_reg(sc, A_SG_KDOORBELL,
992 				     F_SELEGRCNTX |
993 				     (FW_TUNNEL_SGEEC_START + pi->first_qset));
994 		}
995 	}
996 	if (((sc->flags & USING_MSIX) == 0 || sc->params.nports > 2) &&
997 	    sc->open_device_map != 0)
998 		callout_reset(&sc->sge_timer_ch, TX_RECLAIM_PERIOD, sge_timer_cb, sc);
999 }
1000 
1001 /*
1002  * This is meant to be a catch-all function to keep sge state private
1003  * to sge.c
1004  *
1005  */
1006 int
1007 t3_sge_init_adapter(adapter_t *sc)
1008 {
1009 	callout_init(&sc->sge_timer_ch, 1);
1010 	callout_reset(&sc->sge_timer_ch, TX_RECLAIM_PERIOD, sge_timer_cb, sc);
1011 	TASK_INIT(&sc->slow_intr_task, 0, sge_slow_intr_handler, sc);
1012 	return (0);
1013 }
1014 
1015 int
1016 t3_sge_reset_adapter(adapter_t *sc)
1017 {
1018 	callout_reset(&sc->sge_timer_ch, TX_RECLAIM_PERIOD, sge_timer_cb, sc);
1019 	return (0);
1020 }
1021 
1022 int
1023 t3_sge_init_port(struct port_info *pi)
1024 {
1025 	TASK_INIT(&pi->timer_reclaim_task, 0, sge_timer_reclaim, pi);
1026 	return (0);
1027 }
1028 
1029 /**
1030  *	refill_rspq - replenish an SGE response queue
1031  *	@adapter: the adapter
1032  *	@q: the response queue to replenish
1033  *	@credits: how many new responses to make available
1034  *
1035  *	Replenishes a response queue by making the supplied number of responses
1036  *	available to HW.
1037  */
1038 static __inline void
1039 refill_rspq(adapter_t *sc, const struct sge_rspq *q, u_int credits)
1040 {
1041 
1042 	/* mbufs are allocated on demand when a rspq entry is processed. */
1043 	t3_write_reg(sc, A_SG_RSPQ_CREDIT_RETURN,
1044 		     V_RSPQ(q->cntxt_id) | V_CREDITS(credits));
1045 }
1046 
1047 static void
1048 sge_txq_reclaim_handler(void *arg, int ncount)
1049 {
1050 	struct sge_qset *qs = arg;
1051 	int i;
1052 
1053 	for (i = 0; i < 3; i++)
1054 		reclaim_completed_tx(qs, 16, i);
1055 }
1056 
1057 static void
1058 sge_timer_reclaim(void *arg, int ncount)
1059 {
1060 	struct port_info *pi = arg;
1061 	int i, nqsets = pi->nqsets;
1062 	adapter_t *sc = pi->adapter;
1063 	struct sge_qset *qs;
1064 	struct mtx *lock;
1065 
1066 	KASSERT((sc->flags & USING_MSIX) == 0,
1067 	    ("can't call timer reclaim for msi-x"));
1068 
1069 	for (i = 0; i < nqsets; i++) {
1070 		qs = &sc->sge.qs[pi->first_qset + i];
1071 
1072 		reclaim_completed_tx(qs, 16, TXQ_OFLD);
1073 		lock = (sc->flags & USING_MSIX) ? &qs->rspq.lock :
1074 			    &sc->sge.qs[0].rspq.lock;
1075 
1076 		if (mtx_trylock(lock)) {
1077 			/* XXX currently assume that we are *NOT* polling */
1078 			uint32_t status = t3_read_reg(sc, A_SG_RSPQ_FL_STATUS);
1079 
1080 			if (qs->fl[0].credits < qs->fl[0].size - 16)
1081 				__refill_fl(sc, &qs->fl[0]);
1082 			if (qs->fl[1].credits < qs->fl[1].size - 16)
1083 				__refill_fl(sc, &qs->fl[1]);
1084 
1085 			if (status & (1 << qs->rspq.cntxt_id)) {
1086 				if (qs->rspq.credits) {
1087 					refill_rspq(sc, &qs->rspq, 1);
1088 					qs->rspq.credits--;
1089 					t3_write_reg(sc, A_SG_RSPQ_FL_STATUS,
1090 					    1 << qs->rspq.cntxt_id);
1091 				}
1092 			}
1093 			mtx_unlock(lock);
1094 		}
1095 	}
1096 }
1097 
1098 /**
1099  *	init_qset_cntxt - initialize an SGE queue set context info
1100  *	@qs: the queue set
1101  *	@id: the queue set id
1102  *
1103  *	Initializes the TIDs and context ids for the queues of a queue set.
1104  */
1105 static void
1106 init_qset_cntxt(struct sge_qset *qs, u_int id)
1107 {
1108 
1109 	qs->rspq.cntxt_id = id;
1110 	qs->fl[0].cntxt_id = 2 * id;
1111 	qs->fl[1].cntxt_id = 2 * id + 1;
1112 	qs->txq[TXQ_ETH].cntxt_id = FW_TUNNEL_SGEEC_START + id;
1113 	qs->txq[TXQ_ETH].token = FW_TUNNEL_TID_START + id;
1114 	qs->txq[TXQ_OFLD].cntxt_id = FW_OFLD_SGEEC_START + id;
1115 	qs->txq[TXQ_CTRL].cntxt_id = FW_CTRL_SGEEC_START + id;
1116 	qs->txq[TXQ_CTRL].token = FW_CTRL_TID_START + id;
1117 
1118 	/* XXX: a sane limit is needed instead of INT_MAX */
1119 	mbufq_init(&qs->txq[TXQ_ETH].sendq, INT_MAX);
1120 	mbufq_init(&qs->txq[TXQ_OFLD].sendq, INT_MAX);
1121 	mbufq_init(&qs->txq[TXQ_CTRL].sendq, INT_MAX);
1122 }
1123 
1124 
1125 static void
1126 txq_prod(struct sge_txq *txq, unsigned int ndesc, struct txq_state *txqs)
1127 {
1128 	txq->in_use += ndesc;
1129 	/*
1130 	 * XXX we don't handle stopping of queue
1131 	 * presumably start handles this when we bump against the end
1132 	 */
1133 	txqs->gen = txq->gen;
1134 	txq->unacked += ndesc;
1135 	txqs->compl = (txq->unacked & 32) << (S_WR_COMPL - 5);
1136 	txq->unacked &= 31;
1137 	txqs->pidx = txq->pidx;
1138 	txq->pidx += ndesc;
1139 #ifdef INVARIANTS
1140 	if (((txqs->pidx > txq->cidx) &&
1141 		(txq->pidx < txqs->pidx) &&
1142 		(txq->pidx >= txq->cidx)) ||
1143 	    ((txqs->pidx < txq->cidx) &&
1144 		(txq->pidx >= txq-> cidx)) ||
1145 	    ((txqs->pidx < txq->cidx) &&
1146 		(txq->cidx < txqs->pidx)))
1147 		panic("txqs->pidx=%d txq->pidx=%d txq->cidx=%d",
1148 		    txqs->pidx, txq->pidx, txq->cidx);
1149 #endif
1150 	if (txq->pidx >= txq->size) {
1151 		txq->pidx -= txq->size;
1152 		txq->gen ^= 1;
1153 	}
1154 
1155 }
1156 
1157 /**
1158  *	calc_tx_descs - calculate the number of Tx descriptors for a packet
1159  *	@m: the packet mbufs
1160  *      @nsegs: the number of segments
1161  *
1162  * 	Returns the number of Tx descriptors needed for the given Ethernet
1163  * 	packet.  Ethernet packets require addition of WR and CPL headers.
1164  */
1165 static __inline unsigned int
1166 calc_tx_descs(const struct mbuf *m, int nsegs)
1167 {
1168 	unsigned int flits;
1169 
1170 	if (m->m_pkthdr.len <= PIO_LEN)
1171 		return 1;
1172 
1173 	flits = sgl_len(nsegs) + 2;
1174 	if (m->m_pkthdr.csum_flags & CSUM_TSO)
1175 		flits++;
1176 
1177 	return flits_to_desc(flits);
1178 }
1179 
1180 /**
1181  *	make_sgl - populate a scatter/gather list for a packet
1182  *	@sgp: the SGL to populate
1183  *	@segs: the packet dma segments
1184  *	@nsegs: the number of segments
1185  *
1186  *	Generates a scatter/gather list for the buffers that make up a packet
1187  *	and returns the SGL size in 8-byte words.  The caller must size the SGL
1188  *	appropriately.
1189  */
1190 static __inline void
1191 make_sgl(struct sg_ent *sgp, bus_dma_segment_t *segs, int nsegs)
1192 {
1193 	int i, idx;
1194 
1195 	for (idx = 0, i = 0; i < nsegs; i++) {
1196 		/*
1197 		 * firmware doesn't like empty segments
1198 		 */
1199 		if (segs[i].ds_len == 0)
1200 			continue;
1201 		if (i && idx == 0)
1202 			++sgp;
1203 
1204 		sgp->len[idx] = htobe32(segs[i].ds_len);
1205 		sgp->addr[idx] = htobe64(segs[i].ds_addr);
1206 		idx ^= 1;
1207 	}
1208 
1209 	if (idx) {
1210 		sgp->len[idx] = 0;
1211 		sgp->addr[idx] = 0;
1212 	}
1213 }
1214 
1215 /**
1216  *	check_ring_tx_db - check and potentially ring a Tx queue's doorbell
1217  *	@adap: the adapter
1218  *	@q: the Tx queue
1219  *
1220  *	Ring the doorbell if a Tx queue is asleep.  There is a natural race,
1221  *	where the HW is going to sleep just after we checked, however,
1222  *	then the interrupt handler will detect the outstanding TX packet
1223  *	and ring the doorbell for us.
1224  *
1225  *	When GTS is disabled we unconditionally ring the doorbell.
1226  */
1227 static __inline void
1228 check_ring_tx_db(adapter_t *adap, struct sge_txq *q, int mustring)
1229 {
1230 #if USE_GTS
1231 	clear_bit(TXQ_LAST_PKT_DB, &q->flags);
1232 	if (test_and_set_bit(TXQ_RUNNING, &q->flags) == 0) {
1233 		set_bit(TXQ_LAST_PKT_DB, &q->flags);
1234 #ifdef T3_TRACE
1235 		T3_TRACE1(adap->tb[q->cntxt_id & 7], "doorbell Tx, cntxt %d",
1236 			  q->cntxt_id);
1237 #endif
1238 		t3_write_reg(adap, A_SG_KDOORBELL,
1239 			     F_SELEGRCNTX | V_EGRCNTX(q->cntxt_id));
1240 	}
1241 #else
1242 	if (mustring || ++q->db_pending >= 32) {
1243 		wmb();            /* write descriptors before telling HW */
1244 		t3_write_reg(adap, A_SG_KDOORBELL,
1245 		    F_SELEGRCNTX | V_EGRCNTX(q->cntxt_id));
1246 		q->db_pending = 0;
1247 	}
1248 #endif
1249 }
1250 
1251 static __inline void
1252 wr_gen2(struct tx_desc *d, unsigned int gen)
1253 {
1254 #if SGE_NUM_GENBITS == 2
1255 	d->flit[TX_DESC_FLITS - 1] = htobe64(gen);
1256 #endif
1257 }
1258 
1259 /**
1260  *	write_wr_hdr_sgl - write a WR header and, optionally, SGL
1261  *	@ndesc: number of Tx descriptors spanned by the SGL
1262  *	@txd: first Tx descriptor to be written
1263  *	@txqs: txq state (generation and producer index)
1264  *	@txq: the SGE Tx queue
1265  *	@sgl: the SGL
1266  *	@flits: number of flits to the start of the SGL in the first descriptor
1267  *	@sgl_flits: the SGL size in flits
1268  *	@wr_hi: top 32 bits of WR header based on WR type (big endian)
1269  *	@wr_lo: low 32 bits of WR header based on WR type (big endian)
1270  *
1271  *	Write a work request header and an associated SGL.  If the SGL is
1272  *	small enough to fit into one Tx descriptor it has already been written
1273  *	and we just need to write the WR header.  Otherwise we distribute the
1274  *	SGL across the number of descriptors it spans.
1275  */
1276 static void
1277 write_wr_hdr_sgl(unsigned int ndesc, struct tx_desc *txd, struct txq_state *txqs,
1278     const struct sge_txq *txq, const struct sg_ent *sgl, unsigned int flits,
1279     unsigned int sgl_flits, unsigned int wr_hi, unsigned int wr_lo)
1280 {
1281 
1282 	struct work_request_hdr *wrp = (struct work_request_hdr *)txd;
1283 	struct tx_sw_desc *txsd = &txq->sdesc[txqs->pidx];
1284 
1285 	if (__predict_true(ndesc == 1)) {
1286 		set_wr_hdr(wrp, htonl(F_WR_SOP | F_WR_EOP | V_WR_DATATYPE(1) |
1287 		    V_WR_SGLSFLT(flits)) | wr_hi,
1288 		    htonl(V_WR_LEN(flits + sgl_flits) | V_WR_GEN(txqs->gen)) |
1289 		    wr_lo);
1290 
1291 		wr_gen2(txd, txqs->gen);
1292 
1293 	} else {
1294 		unsigned int ogen = txqs->gen;
1295 		const uint64_t *fp = (const uint64_t *)sgl;
1296 		struct work_request_hdr *wp = wrp;
1297 
1298 		wrp->wrh_hi = htonl(F_WR_SOP | V_WR_DATATYPE(1) |
1299 		    V_WR_SGLSFLT(flits)) | wr_hi;
1300 
1301 		while (sgl_flits) {
1302 			unsigned int avail = WR_FLITS - flits;
1303 
1304 			if (avail > sgl_flits)
1305 				avail = sgl_flits;
1306 			memcpy(&txd->flit[flits], fp, avail * sizeof(*fp));
1307 			sgl_flits -= avail;
1308 			ndesc--;
1309 			if (!sgl_flits)
1310 				break;
1311 
1312 			fp += avail;
1313 			txd++;
1314 			txsd++;
1315 			if (++txqs->pidx == txq->size) {
1316 				txqs->pidx = 0;
1317 				txqs->gen ^= 1;
1318 				txd = txq->desc;
1319 				txsd = txq->sdesc;
1320 			}
1321 
1322 			/*
1323 			 * when the head of the mbuf chain
1324 			 * is freed all clusters will be freed
1325 			 * with it
1326 			 */
1327 			wrp = (struct work_request_hdr *)txd;
1328 			wrp->wrh_hi = htonl(V_WR_DATATYPE(1) |
1329 			    V_WR_SGLSFLT(1)) | wr_hi;
1330 			wrp->wrh_lo = htonl(V_WR_LEN(min(WR_FLITS,
1331 				    sgl_flits + 1)) |
1332 			    V_WR_GEN(txqs->gen)) | wr_lo;
1333 			wr_gen2(txd, txqs->gen);
1334 			flits = 1;
1335 		}
1336 		wrp->wrh_hi |= htonl(F_WR_EOP);
1337 		wmb();
1338 		wp->wrh_lo = htonl(V_WR_LEN(WR_FLITS) | V_WR_GEN(ogen)) | wr_lo;
1339 		wr_gen2((struct tx_desc *)wp, ogen);
1340 	}
1341 }
1342 
1343 /* sizeof(*eh) + sizeof(*ip) + sizeof(*tcp) */
1344 #define TCPPKTHDRSIZE (ETHER_HDR_LEN + 20 + 20)
1345 
1346 #define GET_VTAG(cntrl, m) \
1347 do { \
1348 	if ((m)->m_flags & M_VLANTAG)					            \
1349 		cntrl |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN((m)->m_pkthdr.ether_vtag); \
1350 } while (0)
1351 
1352 static int
1353 t3_encap(struct sge_qset *qs, struct mbuf **m)
1354 {
1355 	adapter_t *sc;
1356 	struct mbuf *m0;
1357 	struct sge_txq *txq;
1358 	struct txq_state txqs;
1359 	struct port_info *pi;
1360 	unsigned int ndesc, flits, cntrl, mlen;
1361 	int err, nsegs, tso_info = 0;
1362 
1363 	struct work_request_hdr *wrp;
1364 	struct tx_sw_desc *txsd;
1365 	struct sg_ent *sgp, *sgl;
1366 	uint32_t wr_hi, wr_lo, sgl_flits;
1367 	bus_dma_segment_t segs[TX_MAX_SEGS];
1368 
1369 	struct tx_desc *txd;
1370 
1371 	pi = qs->port;
1372 	sc = pi->adapter;
1373 	txq = &qs->txq[TXQ_ETH];
1374 	txd = &txq->desc[txq->pidx];
1375 	txsd = &txq->sdesc[txq->pidx];
1376 	sgl = txq->txq_sgl;
1377 
1378 	prefetch(txd);
1379 	m0 = *m;
1380 
1381 	mtx_assert(&qs->lock, MA_OWNED);
1382 	cntrl = V_TXPKT_INTF(pi->txpkt_intf);
1383 	KASSERT(m0->m_flags & M_PKTHDR, ("not packet header\n"));
1384 
1385 	if  (m0->m_nextpkt == NULL && m0->m_next != NULL &&
1386 	    m0->m_pkthdr.csum_flags & (CSUM_TSO))
1387 		tso_info = V_LSO_MSS(m0->m_pkthdr.tso_segsz);
1388 
1389 	if (m0->m_nextpkt != NULL) {
1390 		busdma_map_sg_vec(txq->entry_tag, txsd->map, m0, segs, &nsegs);
1391 		ndesc = 1;
1392 		mlen = 0;
1393 	} else {
1394 		if ((err = busdma_map_sg_collapse(txq->entry_tag, txsd->map,
1395 		    &m0, segs, &nsegs))) {
1396 			if (cxgb_debug)
1397 				printf("failed ... err=%d\n", err);
1398 			return (err);
1399 		}
1400 		mlen = m0->m_pkthdr.len;
1401 		ndesc = calc_tx_descs(m0, nsegs);
1402 	}
1403 	txq_prod(txq, ndesc, &txqs);
1404 
1405 	KASSERT(m0->m_pkthdr.len, ("empty packet nsegs=%d", nsegs));
1406 	txsd->m = m0;
1407 
1408 	if (m0->m_nextpkt != NULL) {
1409 		struct cpl_tx_pkt_batch *cpl_batch = (struct cpl_tx_pkt_batch *)txd;
1410 		int i, fidx;
1411 
1412 		if (nsegs > 7)
1413 			panic("trying to coalesce %d packets in to one WR", nsegs);
1414 		txq->txq_coalesced += nsegs;
1415 		wrp = (struct work_request_hdr *)txd;
1416 		flits = nsegs*2 + 1;
1417 
1418 		for (fidx = 1, i = 0; i < nsegs; i++, fidx += 2) {
1419 			struct cpl_tx_pkt_batch_entry *cbe;
1420 			uint64_t flit;
1421 			uint32_t *hflit = (uint32_t *)&flit;
1422 			int cflags = m0->m_pkthdr.csum_flags;
1423 
1424 			cntrl = V_TXPKT_INTF(pi->txpkt_intf);
1425 			GET_VTAG(cntrl, m0);
1426 			cntrl |= V_TXPKT_OPCODE(CPL_TX_PKT);
1427 			if (__predict_false(!(cflags & CSUM_IP)))
1428 				cntrl |= F_TXPKT_IPCSUM_DIS;
1429 			if (__predict_false(!(cflags & (CSUM_TCP | CSUM_UDP |
1430 			    CSUM_UDP_IPV6 | CSUM_TCP_IPV6))))
1431 				cntrl |= F_TXPKT_L4CSUM_DIS;
1432 
1433 			hflit[0] = htonl(cntrl);
1434 			hflit[1] = htonl(segs[i].ds_len | 0x80000000);
1435 			flit |= htobe64(1 << 24);
1436 			cbe = &cpl_batch->pkt_entry[i];
1437 			cbe->cntrl = hflit[0];
1438 			cbe->len = hflit[1];
1439 			cbe->addr = htobe64(segs[i].ds_addr);
1440 		}
1441 
1442 		wr_hi = htonl(F_WR_SOP | F_WR_EOP | V_WR_DATATYPE(1) |
1443 		    V_WR_SGLSFLT(flits)) |
1444 		    htonl(V_WR_OP(FW_WROPCODE_TUNNEL_TX_PKT) | txqs.compl);
1445 		wr_lo = htonl(V_WR_LEN(flits) |
1446 		    V_WR_GEN(txqs.gen)) | htonl(V_WR_TID(txq->token));
1447 		set_wr_hdr(wrp, wr_hi, wr_lo);
1448 		wmb();
1449 		ETHER_BPF_MTAP(pi->ifp, m0);
1450 		wr_gen2(txd, txqs.gen);
1451 		check_ring_tx_db(sc, txq, 0);
1452 		return (0);
1453 	} else if (tso_info) {
1454 		uint16_t eth_type;
1455 		struct cpl_tx_pkt_lso *hdr = (struct cpl_tx_pkt_lso *)txd;
1456 		struct ether_header *eh;
1457 		void *l3hdr;
1458 		struct tcphdr *tcp;
1459 
1460 		txd->flit[2] = 0;
1461 		GET_VTAG(cntrl, m0);
1462 		cntrl |= V_TXPKT_OPCODE(CPL_TX_PKT_LSO);
1463 		hdr->cntrl = htonl(cntrl);
1464 		hdr->len = htonl(mlen | 0x80000000);
1465 
1466 		if (__predict_false(mlen < TCPPKTHDRSIZE)) {
1467 			printf("mbuf=%p,len=%d,tso_segsz=%d,csum_flags=%b,flags=%#x",
1468 			    m0, mlen, m0->m_pkthdr.tso_segsz,
1469 			    (int)m0->m_pkthdr.csum_flags, CSUM_BITS, m0->m_flags);
1470 			panic("tx tso packet too small");
1471 		}
1472 
1473 		/* Make sure that ether, ip, tcp headers are all in m0 */
1474 		if (__predict_false(m0->m_len < TCPPKTHDRSIZE)) {
1475 			m0 = m_pullup(m0, TCPPKTHDRSIZE);
1476 			if (__predict_false(m0 == NULL)) {
1477 				/* XXX panic probably an overreaction */
1478 				panic("couldn't fit header into mbuf");
1479 			}
1480 		}
1481 
1482 		eh = mtod(m0, struct ether_header *);
1483 		eth_type = eh->ether_type;
1484 		if (eth_type == htons(ETHERTYPE_VLAN)) {
1485 			struct ether_vlan_header *evh = (void *)eh;
1486 
1487 			tso_info |= V_LSO_ETH_TYPE(CPL_ETH_II_VLAN);
1488 			l3hdr = evh + 1;
1489 			eth_type = evh->evl_proto;
1490 		} else {
1491 			tso_info |= V_LSO_ETH_TYPE(CPL_ETH_II);
1492 			l3hdr = eh + 1;
1493 		}
1494 
1495 		if (eth_type == htons(ETHERTYPE_IP)) {
1496 			struct ip *ip = l3hdr;
1497 
1498 			tso_info |= V_LSO_IPHDR_WORDS(ip->ip_hl);
1499 			tcp = (struct tcphdr *)(ip + 1);
1500 		} else if (eth_type == htons(ETHERTYPE_IPV6)) {
1501 			struct ip6_hdr *ip6 = l3hdr;
1502 
1503 			KASSERT(ip6->ip6_nxt == IPPROTO_TCP,
1504 			    ("%s: CSUM_TSO with ip6_nxt %d",
1505 			    __func__, ip6->ip6_nxt));
1506 
1507 			tso_info |= F_LSO_IPV6;
1508 			tso_info |= V_LSO_IPHDR_WORDS(sizeof(*ip6) >> 2);
1509 			tcp = (struct tcphdr *)(ip6 + 1);
1510 		} else
1511 			panic("%s: CSUM_TSO but neither ip nor ip6", __func__);
1512 
1513 		tso_info |= V_LSO_TCPHDR_WORDS(tcp->th_off);
1514 		hdr->lso_info = htonl(tso_info);
1515 
1516 		if (__predict_false(mlen <= PIO_LEN)) {
1517 			/*
1518 			 * pkt not undersized but fits in PIO_LEN
1519 			 * Indicates a TSO bug at the higher levels.
1520 			 */
1521 			txsd->m = NULL;
1522 			m_copydata(m0, 0, mlen, (caddr_t)&txd->flit[3]);
1523 			flits = (mlen + 7) / 8 + 3;
1524 			wr_hi = htonl(V_WR_BCNTLFLT(mlen & 7) |
1525 					  V_WR_OP(FW_WROPCODE_TUNNEL_TX_PKT) |
1526 					  F_WR_SOP | F_WR_EOP | txqs.compl);
1527 			wr_lo = htonl(V_WR_LEN(flits) |
1528 			    V_WR_GEN(txqs.gen) | V_WR_TID(txq->token));
1529 			set_wr_hdr(&hdr->wr, wr_hi, wr_lo);
1530 			wmb();
1531 			ETHER_BPF_MTAP(pi->ifp, m0);
1532 			wr_gen2(txd, txqs.gen);
1533 			check_ring_tx_db(sc, txq, 0);
1534 			m_freem(m0);
1535 			return (0);
1536 		}
1537 		flits = 3;
1538 	} else {
1539 		struct cpl_tx_pkt *cpl = (struct cpl_tx_pkt *)txd;
1540 
1541 		GET_VTAG(cntrl, m0);
1542 		cntrl |= V_TXPKT_OPCODE(CPL_TX_PKT);
1543 		if (__predict_false(!(m0->m_pkthdr.csum_flags & CSUM_IP)))
1544 			cntrl |= F_TXPKT_IPCSUM_DIS;
1545 		if (__predict_false(!(m0->m_pkthdr.csum_flags & (CSUM_TCP |
1546 		    CSUM_UDP | CSUM_UDP_IPV6 | CSUM_TCP_IPV6))))
1547 			cntrl |= F_TXPKT_L4CSUM_DIS;
1548 		cpl->cntrl = htonl(cntrl);
1549 		cpl->len = htonl(mlen | 0x80000000);
1550 
1551 		if (mlen <= PIO_LEN) {
1552 			txsd->m = NULL;
1553 			m_copydata(m0, 0, mlen, (caddr_t)&txd->flit[2]);
1554 			flits = (mlen + 7) / 8 + 2;
1555 
1556 			wr_hi = htonl(V_WR_BCNTLFLT(mlen & 7) |
1557 			    V_WR_OP(FW_WROPCODE_TUNNEL_TX_PKT) |
1558 					  F_WR_SOP | F_WR_EOP | txqs.compl);
1559 			wr_lo = htonl(V_WR_LEN(flits) |
1560 			    V_WR_GEN(txqs.gen) | V_WR_TID(txq->token));
1561 			set_wr_hdr(&cpl->wr, wr_hi, wr_lo);
1562 			wmb();
1563 			ETHER_BPF_MTAP(pi->ifp, m0);
1564 			wr_gen2(txd, txqs.gen);
1565 			check_ring_tx_db(sc, txq, 0);
1566 			m_freem(m0);
1567 			return (0);
1568 		}
1569 		flits = 2;
1570 	}
1571 	wrp = (struct work_request_hdr *)txd;
1572 	sgp = (ndesc == 1) ? (struct sg_ent *)&txd->flit[flits] : sgl;
1573 	make_sgl(sgp, segs, nsegs);
1574 
1575 	sgl_flits = sgl_len(nsegs);
1576 
1577 	ETHER_BPF_MTAP(pi->ifp, m0);
1578 
1579 	KASSERT(ndesc <= 4, ("ndesc too large %d", ndesc));
1580 	wr_hi = htonl(V_WR_OP(FW_WROPCODE_TUNNEL_TX_PKT) | txqs.compl);
1581 	wr_lo = htonl(V_WR_TID(txq->token));
1582 	write_wr_hdr_sgl(ndesc, txd, &txqs, txq, sgl, flits,
1583 	    sgl_flits, wr_hi, wr_lo);
1584 	check_ring_tx_db(sc, txq, 0);
1585 
1586 	return (0);
1587 }
1588 
1589 void
1590 cxgb_tx_watchdog(void *arg)
1591 {
1592 	struct sge_qset *qs = arg;
1593 	struct sge_txq *txq = &qs->txq[TXQ_ETH];
1594 
1595         if (qs->coalescing != 0 &&
1596 	    (txq->in_use <= cxgb_tx_coalesce_enable_stop) &&
1597 	    TXQ_RING_EMPTY(qs))
1598                 qs->coalescing = 0;
1599         else if (qs->coalescing == 0 &&
1600 	    (txq->in_use >= cxgb_tx_coalesce_enable_start))
1601                 qs->coalescing = 1;
1602 	if (TXQ_TRYLOCK(qs)) {
1603 		qs->qs_flags |= QS_FLUSHING;
1604 		cxgb_start_locked(qs);
1605 		qs->qs_flags &= ~QS_FLUSHING;
1606 		TXQ_UNLOCK(qs);
1607 	}
1608 	if (qs->port->ifp->if_drv_flags & IFF_DRV_RUNNING)
1609 		callout_reset_on(&txq->txq_watchdog, hz/4, cxgb_tx_watchdog,
1610 		    qs, txq->txq_watchdog.c_cpu);
1611 }
1612 
1613 static void
1614 cxgb_tx_timeout(void *arg)
1615 {
1616 	struct sge_qset *qs = arg;
1617 	struct sge_txq *txq = &qs->txq[TXQ_ETH];
1618 
1619 	if (qs->coalescing == 0 && (txq->in_use >= (txq->size>>3)))
1620                 qs->coalescing = 1;
1621 	if (TXQ_TRYLOCK(qs)) {
1622 		qs->qs_flags |= QS_TIMEOUT;
1623 		cxgb_start_locked(qs);
1624 		qs->qs_flags &= ~QS_TIMEOUT;
1625 		TXQ_UNLOCK(qs);
1626 	}
1627 }
1628 
1629 static void
1630 cxgb_start_locked(struct sge_qset *qs)
1631 {
1632 	struct mbuf *m_head = NULL;
1633 	struct sge_txq *txq = &qs->txq[TXQ_ETH];
1634 	struct port_info *pi = qs->port;
1635 	struct ifnet *ifp = pi->ifp;
1636 
1637 	if (qs->qs_flags & (QS_FLUSHING|QS_TIMEOUT))
1638 		reclaim_completed_tx(qs, 0, TXQ_ETH);
1639 
1640 	if (!pi->link_config.link_ok) {
1641 		TXQ_RING_FLUSH(qs);
1642 		return;
1643 	}
1644 	TXQ_LOCK_ASSERT(qs);
1645 	while (!TXQ_RING_EMPTY(qs) && (ifp->if_drv_flags & IFF_DRV_RUNNING) &&
1646 	    pi->link_config.link_ok) {
1647 		reclaim_completed_tx(qs, cxgb_tx_reclaim_threshold, TXQ_ETH);
1648 
1649 		if (txq->size - txq->in_use <= TX_MAX_DESC)
1650 			break;
1651 
1652 		if ((m_head = cxgb_dequeue(qs)) == NULL)
1653 			break;
1654 		/*
1655 		 *  Encapsulation can modify our pointer, and or make it
1656 		 *  NULL on failure.  In that event, we can't requeue.
1657 		 */
1658 		if (t3_encap(qs, &m_head) || m_head == NULL)
1659 			break;
1660 
1661 		m_head = NULL;
1662 	}
1663 
1664 	if (txq->db_pending)
1665 		check_ring_tx_db(pi->adapter, txq, 1);
1666 
1667 	if (!TXQ_RING_EMPTY(qs) && callout_pending(&txq->txq_timer) == 0 &&
1668 	    pi->link_config.link_ok)
1669 		callout_reset_on(&txq->txq_timer, 1, cxgb_tx_timeout,
1670 		    qs, txq->txq_timer.c_cpu);
1671 	if (m_head != NULL)
1672 		m_freem(m_head);
1673 }
1674 
1675 static int
1676 cxgb_transmit_locked(struct ifnet *ifp, struct sge_qset *qs, struct mbuf *m)
1677 {
1678 	struct port_info *pi = qs->port;
1679 	struct sge_txq *txq = &qs->txq[TXQ_ETH];
1680 	struct buf_ring *br = txq->txq_mr;
1681 	int error, avail;
1682 
1683 	avail = txq->size - txq->in_use;
1684 	TXQ_LOCK_ASSERT(qs);
1685 
1686 	/*
1687 	 * We can only do a direct transmit if the following are true:
1688 	 * - we aren't coalescing (ring < 3/4 full)
1689 	 * - the link is up -- checked in caller
1690 	 * - there are no packets enqueued already
1691 	 * - there is space in hardware transmit queue
1692 	 */
1693 	if (check_pkt_coalesce(qs) == 0 &&
1694 	    !TXQ_RING_NEEDS_ENQUEUE(qs) && avail > TX_MAX_DESC) {
1695 		if (t3_encap(qs, &m)) {
1696 			if (m != NULL &&
1697 			    (error = drbr_enqueue(ifp, br, m)) != 0)
1698 				return (error);
1699 		} else {
1700 			if (txq->db_pending)
1701 				check_ring_tx_db(pi->adapter, txq, 1);
1702 
1703 			/*
1704 			 * We've bypassed the buf ring so we need to update
1705 			 * the stats directly
1706 			 */
1707 			txq->txq_direct_packets++;
1708 			txq->txq_direct_bytes += m->m_pkthdr.len;
1709 		}
1710 	} else if ((error = drbr_enqueue(ifp, br, m)) != 0)
1711 		return (error);
1712 
1713 	reclaim_completed_tx(qs, cxgb_tx_reclaim_threshold, TXQ_ETH);
1714 	if (!TXQ_RING_EMPTY(qs) && pi->link_config.link_ok &&
1715 	    (!check_pkt_coalesce(qs) || (drbr_inuse(ifp, br) >= 7)))
1716 		cxgb_start_locked(qs);
1717 	else if (!TXQ_RING_EMPTY(qs) && !callout_pending(&txq->txq_timer))
1718 		callout_reset_on(&txq->txq_timer, 1, cxgb_tx_timeout,
1719 		    qs, txq->txq_timer.c_cpu);
1720 	return (0);
1721 }
1722 
1723 int
1724 cxgb_transmit(struct ifnet *ifp, struct mbuf *m)
1725 {
1726 	struct sge_qset *qs;
1727 	struct port_info *pi = ifp->if_softc;
1728 	int error, qidx = pi->first_qset;
1729 
1730 	if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0
1731 	    ||(!pi->link_config.link_ok)) {
1732 		m_freem(m);
1733 		return (0);
1734 	}
1735 
1736 	/* check if flowid is set */
1737 	if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
1738 		qidx = (m->m_pkthdr.flowid % pi->nqsets) + pi->first_qset;
1739 
1740 	qs = &pi->adapter->sge.qs[qidx];
1741 
1742 	if (TXQ_TRYLOCK(qs)) {
1743 		/* XXX running */
1744 		error = cxgb_transmit_locked(ifp, qs, m);
1745 		TXQ_UNLOCK(qs);
1746 	} else
1747 		error = drbr_enqueue(ifp, qs->txq[TXQ_ETH].txq_mr, m);
1748 	return (error);
1749 }
1750 
1751 void
1752 cxgb_qflush(struct ifnet *ifp)
1753 {
1754 	/*
1755 	 * flush any enqueued mbufs in the buf_rings
1756 	 * and in the transmit queues
1757 	 * no-op for now
1758 	 */
1759 	return;
1760 }
1761 
1762 /**
1763  *	write_imm - write a packet into a Tx descriptor as immediate data
1764  *	@d: the Tx descriptor to write
1765  *	@m: the packet
1766  *	@len: the length of packet data to write as immediate data
1767  *	@gen: the generation bit value to write
1768  *
1769  *	Writes a packet as immediate data into a Tx descriptor.  The packet
1770  *	contains a work request at its beginning.  We must write the packet
1771  *	carefully so the SGE doesn't read accidentally before it's written in
1772  *	its entirety.
1773  */
1774 static __inline void
1775 write_imm(struct tx_desc *d, caddr_t src,
1776 	  unsigned int len, unsigned int gen)
1777 {
1778 	struct work_request_hdr *from = (struct work_request_hdr *)src;
1779 	struct work_request_hdr *to = (struct work_request_hdr *)d;
1780 	uint32_t wr_hi, wr_lo;
1781 
1782 	KASSERT(len <= WR_LEN && len >= sizeof(*from),
1783 	    ("%s: invalid len %d", __func__, len));
1784 
1785 	memcpy(&to[1], &from[1], len - sizeof(*from));
1786 	wr_hi = from->wrh_hi | htonl(F_WR_SOP | F_WR_EOP |
1787 	    V_WR_BCNTLFLT(len & 7));
1788 	wr_lo = from->wrh_lo | htonl(V_WR_GEN(gen) | V_WR_LEN((len + 7) / 8));
1789 	set_wr_hdr(to, wr_hi, wr_lo);
1790 	wmb();
1791 	wr_gen2(d, gen);
1792 }
1793 
1794 /**
1795  *	check_desc_avail - check descriptor availability on a send queue
1796  *	@adap: the adapter
1797  *	@q: the TX queue
1798  *	@m: the packet needing the descriptors
1799  *	@ndesc: the number of Tx descriptors needed
1800  *	@qid: the Tx queue number in its queue set (TXQ_OFLD or TXQ_CTRL)
1801  *
1802  *	Checks if the requested number of Tx descriptors is available on an
1803  *	SGE send queue.  If the queue is already suspended or not enough
1804  *	descriptors are available the packet is queued for later transmission.
1805  *	Must be called with the Tx queue locked.
1806  *
1807  *	Returns 0 if enough descriptors are available, 1 if there aren't
1808  *	enough descriptors and the packet has been queued, and 2 if the caller
1809  *	needs to retry because there weren't enough descriptors at the
1810  *	beginning of the call but some freed up in the mean time.
1811  */
1812 static __inline int
1813 check_desc_avail(adapter_t *adap, struct sge_txq *q,
1814 		 struct mbuf *m, unsigned int ndesc,
1815 		 unsigned int qid)
1816 {
1817 	/*
1818 	 * XXX We currently only use this for checking the control queue
1819 	 * the control queue is only used for binding qsets which happens
1820 	 * at init time so we are guaranteed enough descriptors
1821 	 */
1822 	if (__predict_false(mbufq_len(&q->sendq))) {
1823 addq_exit:	(void )mbufq_enqueue(&q->sendq, m);
1824 		return 1;
1825 	}
1826 	if (__predict_false(q->size - q->in_use < ndesc)) {
1827 
1828 		struct sge_qset *qs = txq_to_qset(q, qid);
1829 
1830 		setbit(&qs->txq_stopped, qid);
1831 		if (should_restart_tx(q) &&
1832 		    test_and_clear_bit(qid, &qs->txq_stopped))
1833 			return 2;
1834 
1835 		q->stops++;
1836 		goto addq_exit;
1837 	}
1838 	return 0;
1839 }
1840 
1841 
1842 /**
1843  *	reclaim_completed_tx_imm - reclaim completed control-queue Tx descs
1844  *	@q: the SGE control Tx queue
1845  *
1846  *	This is a variant of reclaim_completed_tx() that is used for Tx queues
1847  *	that send only immediate data (presently just the control queues) and
1848  *	thus do not have any mbufs
1849  */
1850 static __inline void
1851 reclaim_completed_tx_imm(struct sge_txq *q)
1852 {
1853 	unsigned int reclaim = q->processed - q->cleaned;
1854 
1855 	q->in_use -= reclaim;
1856 	q->cleaned += reclaim;
1857 }
1858 
1859 /**
1860  *	ctrl_xmit - send a packet through an SGE control Tx queue
1861  *	@adap: the adapter
1862  *	@q: the control queue
1863  *	@m: the packet
1864  *
1865  *	Send a packet through an SGE control Tx queue.  Packets sent through
1866  *	a control queue must fit entirely as immediate data in a single Tx
1867  *	descriptor and have no page fragments.
1868  */
1869 static int
1870 ctrl_xmit(adapter_t *adap, struct sge_qset *qs, struct mbuf *m)
1871 {
1872 	int ret;
1873 	struct work_request_hdr *wrp = mtod(m, struct work_request_hdr *);
1874 	struct sge_txq *q = &qs->txq[TXQ_CTRL];
1875 
1876 	KASSERT(m->m_len <= WR_LEN, ("%s: bad tx data", __func__));
1877 
1878 	wrp->wrh_hi |= htonl(F_WR_SOP | F_WR_EOP);
1879 	wrp->wrh_lo = htonl(V_WR_TID(q->token));
1880 
1881 	TXQ_LOCK(qs);
1882 again:	reclaim_completed_tx_imm(q);
1883 
1884 	ret = check_desc_avail(adap, q, m, 1, TXQ_CTRL);
1885 	if (__predict_false(ret)) {
1886 		if (ret == 1) {
1887 			TXQ_UNLOCK(qs);
1888 			return (ENOSPC);
1889 		}
1890 		goto again;
1891 	}
1892 	write_imm(&q->desc[q->pidx], m->m_data, m->m_len, q->gen);
1893 
1894 	q->in_use++;
1895 	if (++q->pidx >= q->size) {
1896 		q->pidx = 0;
1897 		q->gen ^= 1;
1898 	}
1899 	TXQ_UNLOCK(qs);
1900 	wmb();
1901 	t3_write_reg(adap, A_SG_KDOORBELL,
1902 	    F_SELEGRCNTX | V_EGRCNTX(q->cntxt_id));
1903 
1904 	m_free(m);
1905 	return (0);
1906 }
1907 
1908 
1909 /**
1910  *	restart_ctrlq - restart a suspended control queue
1911  *	@qs: the queue set cotaining the control queue
1912  *
1913  *	Resumes transmission on a suspended Tx control queue.
1914  */
1915 static void
1916 restart_ctrlq(void *data, int npending)
1917 {
1918 	struct mbuf *m;
1919 	struct sge_qset *qs = (struct sge_qset *)data;
1920 	struct sge_txq *q = &qs->txq[TXQ_CTRL];
1921 	adapter_t *adap = qs->port->adapter;
1922 
1923 	TXQ_LOCK(qs);
1924 again:	reclaim_completed_tx_imm(q);
1925 
1926 	while (q->in_use < q->size &&
1927 	       (m = mbufq_dequeue(&q->sendq)) != NULL) {
1928 
1929 		write_imm(&q->desc[q->pidx], m->m_data, m->m_len, q->gen);
1930 		m_free(m);
1931 
1932 		if (++q->pidx >= q->size) {
1933 			q->pidx = 0;
1934 			q->gen ^= 1;
1935 		}
1936 		q->in_use++;
1937 	}
1938 	if (mbufq_len(&q->sendq)) {
1939 		setbit(&qs->txq_stopped, TXQ_CTRL);
1940 
1941 		if (should_restart_tx(q) &&
1942 		    test_and_clear_bit(TXQ_CTRL, &qs->txq_stopped))
1943 			goto again;
1944 		q->stops++;
1945 	}
1946 	TXQ_UNLOCK(qs);
1947 	t3_write_reg(adap, A_SG_KDOORBELL,
1948 		     F_SELEGRCNTX | V_EGRCNTX(q->cntxt_id));
1949 }
1950 
1951 
1952 /*
1953  * Send a management message through control queue 0
1954  */
1955 int
1956 t3_mgmt_tx(struct adapter *adap, struct mbuf *m)
1957 {
1958 	return ctrl_xmit(adap, &adap->sge.qs[0], m);
1959 }
1960 
1961 /**
1962  *	free_qset - free the resources of an SGE queue set
1963  *	@sc: the controller owning the queue set
1964  *	@q: the queue set
1965  *
1966  *	Release the HW and SW resources associated with an SGE queue set, such
1967  *	as HW contexts, packet buffers, and descriptor rings.  Traffic to the
1968  *	queue set must be quiesced prior to calling this.
1969  */
1970 static void
1971 t3_free_qset(adapter_t *sc, struct sge_qset *q)
1972 {
1973 	int i;
1974 
1975 	reclaim_completed_tx(q, 0, TXQ_ETH);
1976 	if (q->txq[TXQ_ETH].txq_mr != NULL)
1977 		buf_ring_free(q->txq[TXQ_ETH].txq_mr, M_DEVBUF);
1978 	if (q->txq[TXQ_ETH].txq_ifq != NULL) {
1979 		ifq_delete(q->txq[TXQ_ETH].txq_ifq);
1980 		free(q->txq[TXQ_ETH].txq_ifq, M_DEVBUF);
1981 	}
1982 
1983 	for (i = 0; i < SGE_RXQ_PER_SET; ++i) {
1984 		if (q->fl[i].desc) {
1985 			mtx_lock_spin(&sc->sge.reg_lock);
1986 			t3_sge_disable_fl(sc, q->fl[i].cntxt_id);
1987 			mtx_unlock_spin(&sc->sge.reg_lock);
1988 			bus_dmamap_unload(q->fl[i].desc_tag, q->fl[i].desc_map);
1989 			bus_dmamem_free(q->fl[i].desc_tag, q->fl[i].desc,
1990 					q->fl[i].desc_map);
1991 			bus_dma_tag_destroy(q->fl[i].desc_tag);
1992 			bus_dma_tag_destroy(q->fl[i].entry_tag);
1993 		}
1994 		if (q->fl[i].sdesc) {
1995 			free_rx_bufs(sc, &q->fl[i]);
1996 			free(q->fl[i].sdesc, M_DEVBUF);
1997 		}
1998 	}
1999 
2000 	mtx_unlock(&q->lock);
2001 	MTX_DESTROY(&q->lock);
2002 	for (i = 0; i < SGE_TXQ_PER_SET; i++) {
2003 		if (q->txq[i].desc) {
2004 			mtx_lock_spin(&sc->sge.reg_lock);
2005 			t3_sge_enable_ecntxt(sc, q->txq[i].cntxt_id, 0);
2006 			mtx_unlock_spin(&sc->sge.reg_lock);
2007 			bus_dmamap_unload(q->txq[i].desc_tag,
2008 					q->txq[i].desc_map);
2009 			bus_dmamem_free(q->txq[i].desc_tag, q->txq[i].desc,
2010 					q->txq[i].desc_map);
2011 			bus_dma_tag_destroy(q->txq[i].desc_tag);
2012 			bus_dma_tag_destroy(q->txq[i].entry_tag);
2013 		}
2014 		if (q->txq[i].sdesc) {
2015 			free(q->txq[i].sdesc, M_DEVBUF);
2016 		}
2017 	}
2018 
2019 	if (q->rspq.desc) {
2020 		mtx_lock_spin(&sc->sge.reg_lock);
2021 		t3_sge_disable_rspcntxt(sc, q->rspq.cntxt_id);
2022 		mtx_unlock_spin(&sc->sge.reg_lock);
2023 
2024 		bus_dmamap_unload(q->rspq.desc_tag, q->rspq.desc_map);
2025 		bus_dmamem_free(q->rspq.desc_tag, q->rspq.desc,
2026 			        q->rspq.desc_map);
2027 		bus_dma_tag_destroy(q->rspq.desc_tag);
2028 		MTX_DESTROY(&q->rspq.lock);
2029 	}
2030 
2031 #if defined(INET6) || defined(INET)
2032 	tcp_lro_free(&q->lro.ctrl);
2033 #endif
2034 
2035 	bzero(q, sizeof(*q));
2036 }
2037 
2038 /**
2039  *	t3_free_sge_resources - free SGE resources
2040  *	@sc: the adapter softc
2041  *
2042  *	Frees resources used by the SGE queue sets.
2043  */
2044 void
2045 t3_free_sge_resources(adapter_t *sc, int nqsets)
2046 {
2047 	int i;
2048 
2049 	for (i = 0; i < nqsets; ++i) {
2050 		TXQ_LOCK(&sc->sge.qs[i]);
2051 		t3_free_qset(sc, &sc->sge.qs[i]);
2052 	}
2053 }
2054 
2055 /**
2056  *	t3_sge_start - enable SGE
2057  *	@sc: the controller softc
2058  *
2059  *	Enables the SGE for DMAs.  This is the last step in starting packet
2060  *	transfers.
2061  */
2062 void
2063 t3_sge_start(adapter_t *sc)
2064 {
2065 	t3_set_reg_field(sc, A_SG_CONTROL, F_GLOBALENABLE, F_GLOBALENABLE);
2066 }
2067 
2068 /**
2069  *	t3_sge_stop - disable SGE operation
2070  *	@sc: the adapter
2071  *
2072  *	Disables the DMA engine.  This can be called in emeregencies (e.g.,
2073  *	from error interrupts) or from normal process context.  In the latter
2074  *	case it also disables any pending queue restart tasklets.  Note that
2075  *	if it is called in interrupt context it cannot disable the restart
2076  *	tasklets as it cannot wait, however the tasklets will have no effect
2077  *	since the doorbells are disabled and the driver will call this again
2078  *	later from process context, at which time the tasklets will be stopped
2079  *	if they are still running.
2080  */
2081 void
2082 t3_sge_stop(adapter_t *sc)
2083 {
2084 	int i, nqsets;
2085 
2086 	t3_set_reg_field(sc, A_SG_CONTROL, F_GLOBALENABLE, 0);
2087 
2088 	if (sc->tq == NULL)
2089 		return;
2090 
2091 	for (nqsets = i = 0; i < (sc)->params.nports; i++)
2092 		nqsets += sc->port[i].nqsets;
2093 #ifdef notyet
2094 	/*
2095 	 *
2096 	 * XXX
2097 	 */
2098 	for (i = 0; i < nqsets; ++i) {
2099 		struct sge_qset *qs = &sc->sge.qs[i];
2100 
2101 		taskqueue_drain(sc->tq, &qs->txq[TXQ_OFLD].qresume_task);
2102 		taskqueue_drain(sc->tq, &qs->txq[TXQ_CTRL].qresume_task);
2103 	}
2104 #endif
2105 }
2106 
2107 /**
2108  *	t3_free_tx_desc - reclaims Tx descriptors and their buffers
2109  *	@adapter: the adapter
2110  *	@q: the Tx queue to reclaim descriptors from
2111  *	@reclaimable: the number of descriptors to reclaim
2112  *      @m_vec_size: maximum number of buffers to reclaim
2113  *      @desc_reclaimed: returns the number of descriptors reclaimed
2114  *
2115  *	Reclaims Tx descriptors from an SGE Tx queue and frees the associated
2116  *	Tx buffers.  Called with the Tx queue lock held.
2117  *
2118  *      Returns number of buffers of reclaimed
2119  */
2120 void
2121 t3_free_tx_desc(struct sge_qset *qs, int reclaimable, int queue)
2122 {
2123 	struct tx_sw_desc *txsd;
2124 	unsigned int cidx, mask;
2125 	struct sge_txq *q = &qs->txq[queue];
2126 
2127 #ifdef T3_TRACE
2128 	T3_TRACE2(sc->tb[q->cntxt_id & 7],
2129 		  "reclaiming %u Tx descriptors at cidx %u", reclaimable, cidx);
2130 #endif
2131 	cidx = q->cidx;
2132 	mask = q->size - 1;
2133 	txsd = &q->sdesc[cidx];
2134 
2135 	mtx_assert(&qs->lock, MA_OWNED);
2136 	while (reclaimable--) {
2137 		prefetch(q->sdesc[(cidx + 1) & mask].m);
2138 		prefetch(q->sdesc[(cidx + 2) & mask].m);
2139 
2140 		if (txsd->m != NULL) {
2141 			if (txsd->flags & TX_SW_DESC_MAPPED) {
2142 				bus_dmamap_unload(q->entry_tag, txsd->map);
2143 				txsd->flags &= ~TX_SW_DESC_MAPPED;
2144 			}
2145 			m_freem_list(txsd->m);
2146 			txsd->m = NULL;
2147 		} else
2148 			q->txq_skipped++;
2149 
2150 		++txsd;
2151 		if (++cidx == q->size) {
2152 			cidx = 0;
2153 			txsd = q->sdesc;
2154 		}
2155 	}
2156 	q->cidx = cidx;
2157 
2158 }
2159 
2160 /**
2161  *	is_new_response - check if a response is newly written
2162  *	@r: the response descriptor
2163  *	@q: the response queue
2164  *
2165  *	Returns true if a response descriptor contains a yet unprocessed
2166  *	response.
2167  */
2168 static __inline int
2169 is_new_response(const struct rsp_desc *r,
2170     const struct sge_rspq *q)
2171 {
2172 	return (r->intr_gen & F_RSPD_GEN2) == q->gen;
2173 }
2174 
2175 #define RSPD_GTS_MASK  (F_RSPD_TXQ0_GTS | F_RSPD_TXQ1_GTS)
2176 #define RSPD_CTRL_MASK (RSPD_GTS_MASK | \
2177 			V_RSPD_TXQ0_CR(M_RSPD_TXQ0_CR) | \
2178 			V_RSPD_TXQ1_CR(M_RSPD_TXQ1_CR) | \
2179 			V_RSPD_TXQ2_CR(M_RSPD_TXQ2_CR))
2180 
2181 /* How long to delay the next interrupt in case of memory shortage, in 0.1us. */
2182 #define NOMEM_INTR_DELAY 2500
2183 
2184 #ifdef TCP_OFFLOAD
2185 /**
2186  *	write_ofld_wr - write an offload work request
2187  *	@adap: the adapter
2188  *	@m: the packet to send
2189  *	@q: the Tx queue
2190  *	@pidx: index of the first Tx descriptor to write
2191  *	@gen: the generation value to use
2192  *	@ndesc: number of descriptors the packet will occupy
2193  *
2194  *	Write an offload work request to send the supplied packet.  The packet
2195  *	data already carry the work request with most fields populated.
2196  */
2197 static void
2198 write_ofld_wr(adapter_t *adap, struct mbuf *m, struct sge_txq *q,
2199     unsigned int pidx, unsigned int gen, unsigned int ndesc)
2200 {
2201 	unsigned int sgl_flits, flits;
2202 	int i, idx, nsegs, wrlen;
2203 	struct work_request_hdr *from;
2204 	struct sg_ent *sgp, t3sgl[TX_MAX_SEGS / 2 + 1];
2205 	struct tx_desc *d = &q->desc[pidx];
2206 	struct txq_state txqs;
2207 	struct sglist_seg *segs;
2208 	struct ofld_hdr *oh = mtod(m, struct ofld_hdr *);
2209 	struct sglist *sgl;
2210 
2211 	from = (void *)(oh + 1);	/* Start of WR within mbuf */
2212 	wrlen = m->m_len - sizeof(*oh);
2213 
2214 	if (!(oh->flags & F_HDR_SGL)) {
2215 		write_imm(d, (caddr_t)from, wrlen, gen);
2216 
2217 		/*
2218 		 * mbuf with "real" immediate tx data will be enqueue_wr'd by
2219 		 * t3_push_frames and freed in wr_ack.  Others, like those sent
2220 		 * down by close_conn, t3_send_reset, etc. should be freed here.
2221 		 */
2222 		if (!(oh->flags & F_HDR_DF))
2223 			m_free(m);
2224 		return;
2225 	}
2226 
2227 	memcpy(&d->flit[1], &from[1], wrlen - sizeof(*from));
2228 
2229 	sgl = oh->sgl;
2230 	flits = wrlen / 8;
2231 	sgp = (ndesc == 1) ? (struct sg_ent *)&d->flit[flits] : t3sgl;
2232 
2233 	nsegs = sgl->sg_nseg;
2234 	segs = sgl->sg_segs;
2235 	for (idx = 0, i = 0; i < nsegs; i++) {
2236 		KASSERT(segs[i].ss_len, ("%s: 0 len in sgl", __func__));
2237 		if (i && idx == 0)
2238 			++sgp;
2239 		sgp->len[idx] = htobe32(segs[i].ss_len);
2240 		sgp->addr[idx] = htobe64(segs[i].ss_paddr);
2241 		idx ^= 1;
2242 	}
2243 	if (idx) {
2244 		sgp->len[idx] = 0;
2245 		sgp->addr[idx] = 0;
2246 	}
2247 
2248 	sgl_flits = sgl_len(nsegs);
2249 	txqs.gen = gen;
2250 	txqs.pidx = pidx;
2251 	txqs.compl = 0;
2252 
2253 	write_wr_hdr_sgl(ndesc, d, &txqs, q, t3sgl, flits, sgl_flits,
2254 	    from->wrh_hi, from->wrh_lo);
2255 }
2256 
2257 /**
2258  *	ofld_xmit - send a packet through an offload queue
2259  *	@adap: the adapter
2260  *	@q: the Tx offload queue
2261  *	@m: the packet
2262  *
2263  *	Send an offload packet through an SGE offload queue.
2264  */
2265 static int
2266 ofld_xmit(adapter_t *adap, struct sge_qset *qs, struct mbuf *m)
2267 {
2268 	int ret;
2269 	unsigned int ndesc;
2270 	unsigned int pidx, gen;
2271 	struct sge_txq *q = &qs->txq[TXQ_OFLD];
2272 	struct ofld_hdr *oh = mtod(m, struct ofld_hdr *);
2273 
2274 	ndesc = G_HDR_NDESC(oh->flags);
2275 
2276 	TXQ_LOCK(qs);
2277 again:	reclaim_completed_tx(qs, 16, TXQ_OFLD);
2278 	ret = check_desc_avail(adap, q, m, ndesc, TXQ_OFLD);
2279 	if (__predict_false(ret)) {
2280 		if (ret == 1) {
2281 			TXQ_UNLOCK(qs);
2282 			return (EINTR);
2283 		}
2284 		goto again;
2285 	}
2286 
2287 	gen = q->gen;
2288 	q->in_use += ndesc;
2289 	pidx = q->pidx;
2290 	q->pidx += ndesc;
2291 	if (q->pidx >= q->size) {
2292 		q->pidx -= q->size;
2293 		q->gen ^= 1;
2294 	}
2295 
2296 	write_ofld_wr(adap, m, q, pidx, gen, ndesc);
2297 	check_ring_tx_db(adap, q, 1);
2298 	TXQ_UNLOCK(qs);
2299 
2300 	return (0);
2301 }
2302 
2303 /**
2304  *	restart_offloadq - restart a suspended offload queue
2305  *	@qs: the queue set cotaining the offload queue
2306  *
2307  *	Resumes transmission on a suspended Tx offload queue.
2308  */
2309 static void
2310 restart_offloadq(void *data, int npending)
2311 {
2312 	struct mbuf *m;
2313 	struct sge_qset *qs = data;
2314 	struct sge_txq *q = &qs->txq[TXQ_OFLD];
2315 	adapter_t *adap = qs->port->adapter;
2316 	int cleaned;
2317 
2318 	TXQ_LOCK(qs);
2319 again:	cleaned = reclaim_completed_tx(qs, 16, TXQ_OFLD);
2320 
2321 	while ((m = mbufq_first(&q->sendq)) != NULL) {
2322 		unsigned int gen, pidx;
2323 		struct ofld_hdr *oh = mtod(m, struct ofld_hdr *);
2324 		unsigned int ndesc = G_HDR_NDESC(oh->flags);
2325 
2326 		if (__predict_false(q->size - q->in_use < ndesc)) {
2327 			setbit(&qs->txq_stopped, TXQ_OFLD);
2328 			if (should_restart_tx(q) &&
2329 			    test_and_clear_bit(TXQ_OFLD, &qs->txq_stopped))
2330 				goto again;
2331 			q->stops++;
2332 			break;
2333 		}
2334 
2335 		gen = q->gen;
2336 		q->in_use += ndesc;
2337 		pidx = q->pidx;
2338 		q->pidx += ndesc;
2339 		if (q->pidx >= q->size) {
2340 			q->pidx -= q->size;
2341 			q->gen ^= 1;
2342 		}
2343 
2344 		(void)mbufq_dequeue(&q->sendq);
2345 		TXQ_UNLOCK(qs);
2346 		write_ofld_wr(adap, m, q, pidx, gen, ndesc);
2347 		TXQ_LOCK(qs);
2348 	}
2349 #if USE_GTS
2350 	set_bit(TXQ_RUNNING, &q->flags);
2351 	set_bit(TXQ_LAST_PKT_DB, &q->flags);
2352 #endif
2353 	TXQ_UNLOCK(qs);
2354 	wmb();
2355 	t3_write_reg(adap, A_SG_KDOORBELL,
2356 		     F_SELEGRCNTX | V_EGRCNTX(q->cntxt_id));
2357 }
2358 
2359 /**
2360  *	t3_offload_tx - send an offload packet
2361  *	@m: the packet
2362  *
2363  *	Sends an offload packet.  We use the packet priority to select the
2364  *	appropriate Tx queue as follows: bit 0 indicates whether the packet
2365  *	should be sent as regular or control, bits 1-3 select the queue set.
2366  */
2367 int
2368 t3_offload_tx(struct adapter *sc, struct mbuf *m)
2369 {
2370 	struct ofld_hdr *oh = mtod(m, struct ofld_hdr *);
2371 	struct sge_qset *qs = &sc->sge.qs[G_HDR_QSET(oh->flags)];
2372 
2373 	if (oh->flags & F_HDR_CTRL) {
2374 		m_adj(m, sizeof (*oh));	/* trim ofld_hdr off */
2375 		return (ctrl_xmit(sc, qs, m));
2376 	} else
2377 		return (ofld_xmit(sc, qs, m));
2378 }
2379 #endif
2380 
2381 static void
2382 restart_tx(struct sge_qset *qs)
2383 {
2384 	struct adapter *sc = qs->port->adapter;
2385 
2386 	if (isset(&qs->txq_stopped, TXQ_OFLD) &&
2387 	    should_restart_tx(&qs->txq[TXQ_OFLD]) &&
2388 	    test_and_clear_bit(TXQ_OFLD, &qs->txq_stopped)) {
2389 		qs->txq[TXQ_OFLD].restarts++;
2390 		taskqueue_enqueue(sc->tq, &qs->txq[TXQ_OFLD].qresume_task);
2391 	}
2392 
2393 	if (isset(&qs->txq_stopped, TXQ_CTRL) &&
2394 	    should_restart_tx(&qs->txq[TXQ_CTRL]) &&
2395 	    test_and_clear_bit(TXQ_CTRL, &qs->txq_stopped)) {
2396 		qs->txq[TXQ_CTRL].restarts++;
2397 		taskqueue_enqueue(sc->tq, &qs->txq[TXQ_CTRL].qresume_task);
2398 	}
2399 }
2400 
2401 /**
2402  *	t3_sge_alloc_qset - initialize an SGE queue set
2403  *	@sc: the controller softc
2404  *	@id: the queue set id
2405  *	@nports: how many Ethernet ports will be using this queue set
2406  *	@irq_vec_idx: the IRQ vector index for response queue interrupts
2407  *	@p: configuration parameters for this queue set
2408  *	@ntxq: number of Tx queues for the queue set
2409  *	@pi: port info for queue set
2410  *
2411  *	Allocate resources and initialize an SGE queue set.  A queue set
2412  *	comprises a response queue, two Rx free-buffer queues, and up to 3
2413  *	Tx queues.  The Tx queues are assigned roles in the order Ethernet
2414  *	queue, offload queue, and control queue.
2415  */
2416 int
2417 t3_sge_alloc_qset(adapter_t *sc, u_int id, int nports, int irq_vec_idx,
2418 		  const struct qset_params *p, int ntxq, struct port_info *pi)
2419 {
2420 	struct sge_qset *q = &sc->sge.qs[id];
2421 	int i, ret = 0;
2422 
2423 	MTX_INIT(&q->lock, q->namebuf, NULL, MTX_DEF);
2424 	q->port = pi;
2425 	q->adap = sc;
2426 
2427 	if ((q->txq[TXQ_ETH].txq_mr = buf_ring_alloc(cxgb_txq_buf_ring_size,
2428 	    M_DEVBUF, M_WAITOK, &q->lock)) == NULL) {
2429 		device_printf(sc->dev, "failed to allocate mbuf ring\n");
2430 		goto err;
2431 	}
2432 	if ((q->txq[TXQ_ETH].txq_ifq = malloc(sizeof(struct ifaltq), M_DEVBUF,
2433 	    M_NOWAIT | M_ZERO)) == NULL) {
2434 		device_printf(sc->dev, "failed to allocate ifq\n");
2435 		goto err;
2436 	}
2437 	ifq_init(q->txq[TXQ_ETH].txq_ifq, pi->ifp);
2438 	callout_init(&q->txq[TXQ_ETH].txq_timer, 1);
2439 	callout_init(&q->txq[TXQ_ETH].txq_watchdog, 1);
2440 	q->txq[TXQ_ETH].txq_timer.c_cpu = id % mp_ncpus;
2441 	q->txq[TXQ_ETH].txq_watchdog.c_cpu = id % mp_ncpus;
2442 
2443 	init_qset_cntxt(q, id);
2444 	q->idx = id;
2445 	if ((ret = alloc_ring(sc, p->fl_size, sizeof(struct rx_desc),
2446 		    sizeof(struct rx_sw_desc), &q->fl[0].phys_addr,
2447 		    &q->fl[0].desc, &q->fl[0].sdesc,
2448 		    &q->fl[0].desc_tag, &q->fl[0].desc_map,
2449 		    sc->rx_dmat, &q->fl[0].entry_tag)) != 0) {
2450 		printf("error %d from alloc ring fl0\n", ret);
2451 		goto err;
2452 	}
2453 
2454 	if ((ret = alloc_ring(sc, p->jumbo_size, sizeof(struct rx_desc),
2455 		    sizeof(struct rx_sw_desc), &q->fl[1].phys_addr,
2456 		    &q->fl[1].desc, &q->fl[1].sdesc,
2457 		    &q->fl[1].desc_tag, &q->fl[1].desc_map,
2458 		    sc->rx_jumbo_dmat, &q->fl[1].entry_tag)) != 0) {
2459 		printf("error %d from alloc ring fl1\n", ret);
2460 		goto err;
2461 	}
2462 
2463 	if ((ret = alloc_ring(sc, p->rspq_size, sizeof(struct rsp_desc), 0,
2464 		    &q->rspq.phys_addr, &q->rspq.desc, NULL,
2465 		    &q->rspq.desc_tag, &q->rspq.desc_map,
2466 		    NULL, NULL)) != 0) {
2467 		printf("error %d from alloc ring rspq\n", ret);
2468 		goto err;
2469 	}
2470 
2471 	snprintf(q->rspq.lockbuf, RSPQ_NAME_LEN, "t3 rspq lock %d:%d",
2472 	    device_get_unit(sc->dev), irq_vec_idx);
2473 	MTX_INIT(&q->rspq.lock, q->rspq.lockbuf, NULL, MTX_DEF);
2474 
2475 	for (i = 0; i < ntxq; ++i) {
2476 		size_t sz = i == TXQ_CTRL ? 0 : sizeof(struct tx_sw_desc);
2477 
2478 		if ((ret = alloc_ring(sc, p->txq_size[i],
2479 			    sizeof(struct tx_desc), sz,
2480 			    &q->txq[i].phys_addr, &q->txq[i].desc,
2481 			    &q->txq[i].sdesc, &q->txq[i].desc_tag,
2482 			    &q->txq[i].desc_map,
2483 			    sc->tx_dmat, &q->txq[i].entry_tag)) != 0) {
2484 			printf("error %d from alloc ring tx %i\n", ret, i);
2485 			goto err;
2486 		}
2487 		mbufq_init(&q->txq[i].sendq, INT_MAX);
2488 		q->txq[i].gen = 1;
2489 		q->txq[i].size = p->txq_size[i];
2490 	}
2491 
2492 #ifdef TCP_OFFLOAD
2493 	TASK_INIT(&q->txq[TXQ_OFLD].qresume_task, 0, restart_offloadq, q);
2494 #endif
2495 	TASK_INIT(&q->txq[TXQ_CTRL].qresume_task, 0, restart_ctrlq, q);
2496 	TASK_INIT(&q->txq[TXQ_ETH].qreclaim_task, 0, sge_txq_reclaim_handler, q);
2497 	TASK_INIT(&q->txq[TXQ_OFLD].qreclaim_task, 0, sge_txq_reclaim_handler, q);
2498 
2499 	q->fl[0].gen = q->fl[1].gen = 1;
2500 	q->fl[0].size = p->fl_size;
2501 	q->fl[1].size = p->jumbo_size;
2502 
2503 	q->rspq.gen = 1;
2504 	q->rspq.cidx = 0;
2505 	q->rspq.size = p->rspq_size;
2506 
2507 	q->txq[TXQ_ETH].stop_thres = nports *
2508 	    flits_to_desc(sgl_len(TX_MAX_SEGS + 1) + 3);
2509 
2510 	q->fl[0].buf_size = MCLBYTES;
2511 	q->fl[0].zone = zone_pack;
2512 	q->fl[0].type = EXT_PACKET;
2513 
2514 	if (p->jumbo_buf_size ==  MJUM16BYTES) {
2515 		q->fl[1].zone = zone_jumbo16;
2516 		q->fl[1].type = EXT_JUMBO16;
2517 	} else if (p->jumbo_buf_size ==  MJUM9BYTES) {
2518 		q->fl[1].zone = zone_jumbo9;
2519 		q->fl[1].type = EXT_JUMBO9;
2520 	} else if (p->jumbo_buf_size ==  MJUMPAGESIZE) {
2521 		q->fl[1].zone = zone_jumbop;
2522 		q->fl[1].type = EXT_JUMBOP;
2523 	} else {
2524 		KASSERT(0, ("can't deal with jumbo_buf_size %d.", p->jumbo_buf_size));
2525 		ret = EDOOFUS;
2526 		goto err;
2527 	}
2528 	q->fl[1].buf_size = p->jumbo_buf_size;
2529 
2530 	/* Allocate and setup the lro_ctrl structure */
2531 	q->lro.enabled = !!(pi->ifp->if_capenable & IFCAP_LRO);
2532 #if defined(INET6) || defined(INET)
2533 	ret = tcp_lro_init(&q->lro.ctrl);
2534 	if (ret) {
2535 		printf("error %d from tcp_lro_init\n", ret);
2536 		goto err;
2537 	}
2538 #endif
2539 	q->lro.ctrl.ifp = pi->ifp;
2540 
2541 	mtx_lock_spin(&sc->sge.reg_lock);
2542 	ret = -t3_sge_init_rspcntxt(sc, q->rspq.cntxt_id, irq_vec_idx,
2543 				   q->rspq.phys_addr, q->rspq.size,
2544 				   q->fl[0].buf_size, 1, 0);
2545 	if (ret) {
2546 		printf("error %d from t3_sge_init_rspcntxt\n", ret);
2547 		goto err_unlock;
2548 	}
2549 
2550 	for (i = 0; i < SGE_RXQ_PER_SET; ++i) {
2551 		ret = -t3_sge_init_flcntxt(sc, q->fl[i].cntxt_id, 0,
2552 					  q->fl[i].phys_addr, q->fl[i].size,
2553 					  q->fl[i].buf_size, p->cong_thres, 1,
2554 					  0);
2555 		if (ret) {
2556 			printf("error %d from t3_sge_init_flcntxt for index i=%d\n", ret, i);
2557 			goto err_unlock;
2558 		}
2559 	}
2560 
2561 	ret = -t3_sge_init_ecntxt(sc, q->txq[TXQ_ETH].cntxt_id, USE_GTS,
2562 				 SGE_CNTXT_ETH, id, q->txq[TXQ_ETH].phys_addr,
2563 				 q->txq[TXQ_ETH].size, q->txq[TXQ_ETH].token,
2564 				 1, 0);
2565 	if (ret) {
2566 		printf("error %d from t3_sge_init_ecntxt\n", ret);
2567 		goto err_unlock;
2568 	}
2569 
2570 	if (ntxq > 1) {
2571 		ret = -t3_sge_init_ecntxt(sc, q->txq[TXQ_OFLD].cntxt_id,
2572 					 USE_GTS, SGE_CNTXT_OFLD, id,
2573 					 q->txq[TXQ_OFLD].phys_addr,
2574 					 q->txq[TXQ_OFLD].size, 0, 1, 0);
2575 		if (ret) {
2576 			printf("error %d from t3_sge_init_ecntxt\n", ret);
2577 			goto err_unlock;
2578 		}
2579 	}
2580 
2581 	if (ntxq > 2) {
2582 		ret = -t3_sge_init_ecntxt(sc, q->txq[TXQ_CTRL].cntxt_id, 0,
2583 					 SGE_CNTXT_CTRL, id,
2584 					 q->txq[TXQ_CTRL].phys_addr,
2585 					 q->txq[TXQ_CTRL].size,
2586 					 q->txq[TXQ_CTRL].token, 1, 0);
2587 		if (ret) {
2588 			printf("error %d from t3_sge_init_ecntxt\n", ret);
2589 			goto err_unlock;
2590 		}
2591 	}
2592 
2593 	mtx_unlock_spin(&sc->sge.reg_lock);
2594 	t3_update_qset_coalesce(q, p);
2595 
2596 	refill_fl(sc, &q->fl[0], q->fl[0].size);
2597 	refill_fl(sc, &q->fl[1], q->fl[1].size);
2598 	refill_rspq(sc, &q->rspq, q->rspq.size - 1);
2599 
2600 	t3_write_reg(sc, A_SG_GTS, V_RSPQ(q->rspq.cntxt_id) |
2601 		     V_NEWTIMER(q->rspq.holdoff_tmr));
2602 
2603 	return (0);
2604 
2605 err_unlock:
2606 	mtx_unlock_spin(&sc->sge.reg_lock);
2607 err:
2608 	TXQ_LOCK(q);
2609 	t3_free_qset(sc, q);
2610 
2611 	return (ret);
2612 }
2613 
2614 /*
2615  * Remove CPL_RX_PKT headers from the mbuf and reduce it to a regular mbuf with
2616  * ethernet data.  Hardware assistance with various checksums and any vlan tag
2617  * will also be taken into account here.
2618  */
2619 void
2620 t3_rx_eth(struct adapter *adap, struct mbuf *m, int ethpad)
2621 {
2622 	struct cpl_rx_pkt *cpl = (struct cpl_rx_pkt *)(mtod(m, uint8_t *) + ethpad);
2623 	struct port_info *pi = &adap->port[adap->rxpkt_map[cpl->iff]];
2624 	struct ifnet *ifp = pi->ifp;
2625 
2626 	if (cpl->vlan_valid) {
2627 		m->m_pkthdr.ether_vtag = ntohs(cpl->vlan);
2628 		m->m_flags |= M_VLANTAG;
2629 	}
2630 
2631 	m->m_pkthdr.rcvif = ifp;
2632 	/*
2633 	 * adjust after conversion to mbuf chain
2634 	 */
2635 	m->m_pkthdr.len -= (sizeof(*cpl) + ethpad);
2636 	m->m_len -= (sizeof(*cpl) + ethpad);
2637 	m->m_data += (sizeof(*cpl) + ethpad);
2638 
2639 	if (!cpl->fragment && cpl->csum_valid && cpl->csum == 0xffff) {
2640 		struct ether_header *eh = mtod(m, void *);
2641 		uint16_t eh_type;
2642 
2643 		if (eh->ether_type == htons(ETHERTYPE_VLAN)) {
2644 			struct ether_vlan_header *evh = mtod(m, void *);
2645 
2646 			eh_type = evh->evl_proto;
2647 		} else
2648 			eh_type = eh->ether_type;
2649 
2650 		if (ifp->if_capenable & IFCAP_RXCSUM &&
2651 		    eh_type == htons(ETHERTYPE_IP)) {
2652 			m->m_pkthdr.csum_flags = (CSUM_IP_CHECKED |
2653 			    CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
2654 			m->m_pkthdr.csum_data = 0xffff;
2655 		} else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 &&
2656 		    eh_type == htons(ETHERTYPE_IPV6)) {
2657 			m->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 |
2658 			    CSUM_PSEUDO_HDR);
2659 			m->m_pkthdr.csum_data = 0xffff;
2660 		}
2661 	}
2662 }
2663 
2664 /**
2665  *	get_packet - return the next ingress packet buffer from a free list
2666  *	@adap: the adapter that received the packet
2667  *	@drop_thres: # of remaining buffers before we start dropping packets
2668  *	@qs: the qset that the SGE free list holding the packet belongs to
2669  *      @mh: the mbuf header, contains a pointer to the head and tail of the mbuf chain
2670  *      @r: response descriptor
2671  *
2672  *	Get the next packet from a free list and complete setup of the
2673  *	sk_buff.  If the packet is small we make a copy and recycle the
2674  *	original buffer, otherwise we use the original buffer itself.  If a
2675  *	positive drop threshold is supplied packets are dropped and their
2676  *	buffers recycled if (a) the number of remaining buffers is under the
2677  *	threshold and the packet is too big to copy, or (b) the packet should
2678  *	be copied but there is no memory for the copy.
2679  */
2680 static int
2681 get_packet(adapter_t *adap, unsigned int drop_thres, struct sge_qset *qs,
2682     struct t3_mbuf_hdr *mh, struct rsp_desc *r)
2683 {
2684 
2685 	unsigned int len_cq =  ntohl(r->len_cq);
2686 	struct sge_fl *fl = (len_cq & F_RSPD_FLQ) ? &qs->fl[1] : &qs->fl[0];
2687 	int mask, cidx = fl->cidx;
2688 	struct rx_sw_desc *sd = &fl->sdesc[cidx];
2689 	uint32_t len = G_RSPD_LEN(len_cq);
2690 	uint32_t flags = M_EXT;
2691 	uint8_t sopeop = G_RSPD_SOP_EOP(ntohl(r->flags));
2692 	caddr_t cl;
2693 	struct mbuf *m;
2694 	int ret = 0;
2695 
2696 	mask = fl->size - 1;
2697 	prefetch(fl->sdesc[(cidx + 1) & mask].m);
2698 	prefetch(fl->sdesc[(cidx + 2) & mask].m);
2699 	prefetch(fl->sdesc[(cidx + 1) & mask].rxsd_cl);
2700 	prefetch(fl->sdesc[(cidx + 2) & mask].rxsd_cl);
2701 
2702 	fl->credits--;
2703 	bus_dmamap_sync(fl->entry_tag, sd->map, BUS_DMASYNC_POSTREAD);
2704 
2705 	if (recycle_enable && len <= SGE_RX_COPY_THRES &&
2706 	    sopeop == RSPQ_SOP_EOP) {
2707 		if ((m = m_gethdr(M_NOWAIT, MT_DATA)) == NULL)
2708 			goto skip_recycle;
2709 		cl = mtod(m, void *);
2710 		memcpy(cl, sd->rxsd_cl, len);
2711 		recycle_rx_buf(adap, fl, fl->cidx);
2712 		m->m_pkthdr.len = m->m_len = len;
2713 		m->m_flags = 0;
2714 		mh->mh_head = mh->mh_tail = m;
2715 		ret = 1;
2716 		goto done;
2717 	} else {
2718 	skip_recycle:
2719 		bus_dmamap_unload(fl->entry_tag, sd->map);
2720 		cl = sd->rxsd_cl;
2721 		m = sd->m;
2722 
2723 		if ((sopeop == RSPQ_SOP_EOP) ||
2724 		    (sopeop == RSPQ_SOP))
2725 			flags |= M_PKTHDR;
2726 		m_init(m, M_NOWAIT, MT_DATA, flags);
2727 		if (fl->zone == zone_pack) {
2728 			/*
2729 			 * restore clobbered data pointer
2730 			 */
2731 			m->m_data = m->m_ext.ext_buf;
2732 		} else {
2733 			m_cljset(m, cl, fl->type);
2734 		}
2735 		m->m_len = len;
2736 	}
2737 	switch(sopeop) {
2738 	case RSPQ_SOP_EOP:
2739 		ret = 1;
2740 		/* FALLTHROUGH */
2741 	case RSPQ_SOP:
2742 		mh->mh_head = mh->mh_tail = m;
2743 		m->m_pkthdr.len = len;
2744 		break;
2745 	case RSPQ_EOP:
2746 		ret = 1;
2747 		/* FALLTHROUGH */
2748 	case RSPQ_NSOP_NEOP:
2749 		if (mh->mh_tail == NULL) {
2750 			log(LOG_ERR, "discarding intermediate descriptor entry\n");
2751 			m_freem(m);
2752 			break;
2753 		}
2754 		mh->mh_tail->m_next = m;
2755 		mh->mh_tail = m;
2756 		mh->mh_head->m_pkthdr.len += len;
2757 		break;
2758 	}
2759 	if (cxgb_debug)
2760 		printf("len=%d pktlen=%d\n", m->m_len, m->m_pkthdr.len);
2761 done:
2762 	if (++fl->cidx == fl->size)
2763 		fl->cidx = 0;
2764 
2765 	return (ret);
2766 }
2767 
2768 /**
2769  *	handle_rsp_cntrl_info - handles control information in a response
2770  *	@qs: the queue set corresponding to the response
2771  *	@flags: the response control flags
2772  *
2773  *	Handles the control information of an SGE response, such as GTS
2774  *	indications and completion credits for the queue set's Tx queues.
2775  *	HW coalesces credits, we don't do any extra SW coalescing.
2776  */
2777 static __inline void
2778 handle_rsp_cntrl_info(struct sge_qset *qs, uint32_t flags)
2779 {
2780 	unsigned int credits;
2781 
2782 #if USE_GTS
2783 	if (flags & F_RSPD_TXQ0_GTS)
2784 		clear_bit(TXQ_RUNNING, &qs->txq[TXQ_ETH].flags);
2785 #endif
2786 	credits = G_RSPD_TXQ0_CR(flags);
2787 	if (credits)
2788 		qs->txq[TXQ_ETH].processed += credits;
2789 
2790 	credits = G_RSPD_TXQ2_CR(flags);
2791 	if (credits)
2792 		qs->txq[TXQ_CTRL].processed += credits;
2793 
2794 # if USE_GTS
2795 	if (flags & F_RSPD_TXQ1_GTS)
2796 		clear_bit(TXQ_RUNNING, &qs->txq[TXQ_OFLD].flags);
2797 # endif
2798 	credits = G_RSPD_TXQ1_CR(flags);
2799 	if (credits)
2800 		qs->txq[TXQ_OFLD].processed += credits;
2801 
2802 }
2803 
2804 static void
2805 check_ring_db(adapter_t *adap, struct sge_qset *qs,
2806     unsigned int sleeping)
2807 {
2808 	;
2809 }
2810 
2811 /**
2812  *	process_responses - process responses from an SGE response queue
2813  *	@adap: the adapter
2814  *	@qs: the queue set to which the response queue belongs
2815  *	@budget: how many responses can be processed in this round
2816  *
2817  *	Process responses from an SGE response queue up to the supplied budget.
2818  *	Responses include received packets as well as credits and other events
2819  *	for the queues that belong to the response queue's queue set.
2820  *	A negative budget is effectively unlimited.
2821  *
2822  *	Additionally choose the interrupt holdoff time for the next interrupt
2823  *	on this queue.  If the system is under memory shortage use a fairly
2824  *	long delay to help recovery.
2825  */
2826 static int
2827 process_responses(adapter_t *adap, struct sge_qset *qs, int budget)
2828 {
2829 	struct sge_rspq *rspq = &qs->rspq;
2830 	struct rsp_desc *r = &rspq->desc[rspq->cidx];
2831 	int budget_left = budget;
2832 	unsigned int sleeping = 0;
2833 #if defined(INET6) || defined(INET)
2834 	int lro_enabled = qs->lro.enabled;
2835 	int skip_lro;
2836 	struct lro_ctrl *lro_ctrl = &qs->lro.ctrl;
2837 #endif
2838 	struct t3_mbuf_hdr *mh = &rspq->rspq_mh;
2839 #ifdef DEBUG
2840 	static int last_holdoff = 0;
2841 	if (cxgb_debug && rspq->holdoff_tmr != last_holdoff) {
2842 		printf("next_holdoff=%d\n", rspq->holdoff_tmr);
2843 		last_holdoff = rspq->holdoff_tmr;
2844 	}
2845 #endif
2846 	rspq->next_holdoff = rspq->holdoff_tmr;
2847 
2848 	while (__predict_true(budget_left && is_new_response(r, rspq))) {
2849 		int eth, eop = 0, ethpad = 0;
2850 		uint32_t flags = ntohl(r->flags);
2851 		uint32_t rss_hash = be32toh(r->rss_hdr.rss_hash_val);
2852 		uint8_t opcode = r->rss_hdr.opcode;
2853 
2854 		eth = (opcode == CPL_RX_PKT);
2855 
2856 		if (__predict_false(flags & F_RSPD_ASYNC_NOTIF)) {
2857 			struct mbuf *m;
2858 
2859 			if (cxgb_debug)
2860 				printf("async notification\n");
2861 
2862 			if (mh->mh_head == NULL) {
2863 				mh->mh_head = m_gethdr(M_NOWAIT, MT_DATA);
2864 				m = mh->mh_head;
2865 			} else {
2866 				m = m_gethdr(M_NOWAIT, MT_DATA);
2867 			}
2868 			if (m == NULL)
2869 				goto no_mem;
2870 
2871                         memcpy(mtod(m, char *), r, AN_PKT_SIZE);
2872 			m->m_len = m->m_pkthdr.len = AN_PKT_SIZE;
2873                         *mtod(m, uint8_t *) = CPL_ASYNC_NOTIF;
2874 			opcode = CPL_ASYNC_NOTIF;
2875 			eop = 1;
2876                         rspq->async_notif++;
2877 			goto skip;
2878 		} else if  (flags & F_RSPD_IMM_DATA_VALID) {
2879 			struct mbuf *m = m_gethdr(M_NOWAIT, MT_DATA);
2880 
2881 			if (m == NULL) {
2882 		no_mem:
2883 				rspq->next_holdoff = NOMEM_INTR_DELAY;
2884 				budget_left--;
2885 				break;
2886 			}
2887 			if (mh->mh_head == NULL)
2888 				mh->mh_head = m;
2889                         else
2890 				mh->mh_tail->m_next = m;
2891 			mh->mh_tail = m;
2892 
2893 			get_imm_packet(adap, r, m);
2894 			mh->mh_head->m_pkthdr.len += m->m_len;
2895 			eop = 1;
2896 			rspq->imm_data++;
2897 		} else if (r->len_cq) {
2898 			int drop_thresh = eth ? SGE_RX_DROP_THRES : 0;
2899 
2900 			eop = get_packet(adap, drop_thresh, qs, mh, r);
2901 			if (eop) {
2902 				if (r->rss_hdr.hash_type && !adap->timestamp) {
2903 					M_HASHTYPE_SET(mh->mh_head,
2904 					    M_HASHTYPE_OPAQUE_HASH);
2905 					mh->mh_head->m_pkthdr.flowid = rss_hash;
2906 				}
2907 			}
2908 
2909 			ethpad = 2;
2910 		} else {
2911 			rspq->pure_rsps++;
2912 		}
2913 	skip:
2914 		if (flags & RSPD_CTRL_MASK) {
2915 			sleeping |= flags & RSPD_GTS_MASK;
2916 			handle_rsp_cntrl_info(qs, flags);
2917 		}
2918 
2919 		if (!eth && eop) {
2920 			rspq->offload_pkts++;
2921 #ifdef TCP_OFFLOAD
2922 			adap->cpl_handler[opcode](qs, r, mh->mh_head);
2923 #else
2924 			m_freem(mh->mh_head);
2925 #endif
2926 			mh->mh_head = NULL;
2927 		} else if (eth && eop) {
2928 			struct mbuf *m = mh->mh_head;
2929 
2930 			t3_rx_eth(adap, m, ethpad);
2931 
2932 			/*
2933 			 * The T304 sends incoming packets on any qset.  If LRO
2934 			 * is also enabled, we could end up sending packet up
2935 			 * lro_ctrl->ifp's input.  That is incorrect.
2936 			 *
2937 			 * The mbuf's rcvif was derived from the cpl header and
2938 			 * is accurate.  Skip LRO and just use that.
2939 			 */
2940 #if defined(INET6) || defined(INET)
2941 			skip_lro = __predict_false(qs->port->ifp != m->m_pkthdr.rcvif);
2942 
2943 			if (lro_enabled && lro_ctrl->lro_cnt && !skip_lro
2944 			    && (tcp_lro_rx(lro_ctrl, m, 0) == 0)
2945 			    ) {
2946 				/* successfully queue'd for LRO */
2947 			} else
2948 #endif
2949 			{
2950 				/*
2951 				 * LRO not enabled, packet unsuitable for LRO,
2952 				 * or unable to queue.  Pass it up right now in
2953 				 * either case.
2954 				 */
2955 				struct ifnet *ifp = m->m_pkthdr.rcvif;
2956 				(*ifp->if_input)(ifp, m);
2957 			}
2958 			mh->mh_head = NULL;
2959 
2960 		}
2961 
2962 		r++;
2963 		if (__predict_false(++rspq->cidx == rspq->size)) {
2964 			rspq->cidx = 0;
2965 			rspq->gen ^= 1;
2966 			r = rspq->desc;
2967 		}
2968 
2969 		if (++rspq->credits >= 64) {
2970 			refill_rspq(adap, rspq, rspq->credits);
2971 			rspq->credits = 0;
2972 		}
2973 		__refill_fl_lt(adap, &qs->fl[0], 32);
2974 		__refill_fl_lt(adap, &qs->fl[1], 32);
2975 		--budget_left;
2976 	}
2977 
2978 #if defined(INET6) || defined(INET)
2979 	/* Flush LRO */
2980 	tcp_lro_flush_all(lro_ctrl);
2981 #endif
2982 
2983 	if (sleeping)
2984 		check_ring_db(adap, qs, sleeping);
2985 
2986 	mb();  /* commit Tx queue processed updates */
2987 	if (__predict_false(qs->txq_stopped > 1))
2988 		restart_tx(qs);
2989 
2990 	__refill_fl_lt(adap, &qs->fl[0], 512);
2991 	__refill_fl_lt(adap, &qs->fl[1], 512);
2992 	budget -= budget_left;
2993 	return (budget);
2994 }
2995 
2996 /*
2997  * A helper function that processes responses and issues GTS.
2998  */
2999 static __inline int
3000 process_responses_gts(adapter_t *adap, struct sge_rspq *rq)
3001 {
3002 	int work;
3003 	static int last_holdoff = 0;
3004 
3005 	work = process_responses(adap, rspq_to_qset(rq), -1);
3006 
3007 	if (cxgb_debug && (rq->next_holdoff != last_holdoff)) {
3008 		printf("next_holdoff=%d\n", rq->next_holdoff);
3009 		last_holdoff = rq->next_holdoff;
3010 	}
3011 	t3_write_reg(adap, A_SG_GTS, V_RSPQ(rq->cntxt_id) |
3012 	    V_NEWTIMER(rq->next_holdoff) | V_NEWINDEX(rq->cidx));
3013 
3014 	return (work);
3015 }
3016 
3017 
3018 /*
3019  * Interrupt handler for legacy INTx interrupts for T3B-based cards.
3020  * Handles data events from SGE response queues as well as error and other
3021  * async events as they all use the same interrupt pin.  We use one SGE
3022  * response queue per port in this mode and protect all response queues with
3023  * queue 0's lock.
3024  */
3025 void
3026 t3b_intr(void *data)
3027 {
3028 	uint32_t i, map;
3029 	adapter_t *adap = data;
3030 	struct sge_rspq *q0 = &adap->sge.qs[0].rspq;
3031 
3032 	t3_write_reg(adap, A_PL_CLI, 0);
3033 	map = t3_read_reg(adap, A_SG_DATA_INTR);
3034 
3035 	if (!map)
3036 		return;
3037 
3038 	if (__predict_false(map & F_ERRINTR)) {
3039 		t3_write_reg(adap, A_PL_INT_ENABLE0, 0);
3040 		(void) t3_read_reg(adap, A_PL_INT_ENABLE0);
3041 		taskqueue_enqueue(adap->tq, &adap->slow_intr_task);
3042 	}
3043 
3044 	mtx_lock(&q0->lock);
3045 	for_each_port(adap, i)
3046 	    if (map & (1 << i))
3047 			process_responses_gts(adap, &adap->sge.qs[i].rspq);
3048 	mtx_unlock(&q0->lock);
3049 }
3050 
3051 /*
3052  * The MSI interrupt handler.  This needs to handle data events from SGE
3053  * response queues as well as error and other async events as they all use
3054  * the same MSI vector.  We use one SGE response queue per port in this mode
3055  * and protect all response queues with queue 0's lock.
3056  */
3057 void
3058 t3_intr_msi(void *data)
3059 {
3060 	adapter_t *adap = data;
3061 	struct sge_rspq *q0 = &adap->sge.qs[0].rspq;
3062 	int i, new_packets = 0;
3063 
3064 	mtx_lock(&q0->lock);
3065 
3066 	for_each_port(adap, i)
3067 	    if (process_responses_gts(adap, &adap->sge.qs[i].rspq))
3068 		    new_packets = 1;
3069 	mtx_unlock(&q0->lock);
3070 	if (new_packets == 0) {
3071 		t3_write_reg(adap, A_PL_INT_ENABLE0, 0);
3072 		(void) t3_read_reg(adap, A_PL_INT_ENABLE0);
3073 		taskqueue_enqueue(adap->tq, &adap->slow_intr_task);
3074 	}
3075 }
3076 
3077 void
3078 t3_intr_msix(void *data)
3079 {
3080 	struct sge_qset *qs = data;
3081 	adapter_t *adap = qs->port->adapter;
3082 	struct sge_rspq *rspq = &qs->rspq;
3083 
3084 	if (process_responses_gts(adap, rspq) == 0)
3085 		rspq->unhandled_irqs++;
3086 }
3087 
3088 #define QDUMP_SBUF_SIZE		32 * 400
3089 static int
3090 t3_dump_rspq(SYSCTL_HANDLER_ARGS)
3091 {
3092 	struct sge_rspq *rspq;
3093 	struct sge_qset *qs;
3094 	int i, err, dump_end, idx;
3095 	struct sbuf *sb;
3096 	struct rsp_desc *rspd;
3097 	uint32_t data[4];
3098 
3099 	rspq = arg1;
3100 	qs = rspq_to_qset(rspq);
3101 	if (rspq->rspq_dump_count == 0)
3102 		return (0);
3103 	if (rspq->rspq_dump_count > RSPQ_Q_SIZE) {
3104 		log(LOG_WARNING,
3105 		    "dump count is too large %d\n", rspq->rspq_dump_count);
3106 		rspq->rspq_dump_count = 0;
3107 		return (EINVAL);
3108 	}
3109 	if (rspq->rspq_dump_start > (RSPQ_Q_SIZE-1)) {
3110 		log(LOG_WARNING,
3111 		    "dump start of %d is greater than queue size\n",
3112 		    rspq->rspq_dump_start);
3113 		rspq->rspq_dump_start = 0;
3114 		return (EINVAL);
3115 	}
3116 	err = t3_sge_read_rspq(qs->port->adapter, rspq->cntxt_id, data);
3117 	if (err)
3118 		return (err);
3119 	err = sysctl_wire_old_buffer(req, 0);
3120 	if (err)
3121 		return (err);
3122 	sb = sbuf_new_for_sysctl(NULL, NULL, QDUMP_SBUF_SIZE, req);
3123 
3124 	sbuf_printf(sb, " \n index=%u size=%u MSI-X/RspQ=%u intr enable=%u intr armed=%u\n",
3125 	    (data[0] & 0xffff), data[0] >> 16, ((data[2] >> 20) & 0x3f),
3126 	    ((data[2] >> 26) & 1), ((data[2] >> 27) & 1));
3127 	sbuf_printf(sb, " generation=%u CQ mode=%u FL threshold=%u\n",
3128 	    ((data[2] >> 28) & 1), ((data[2] >> 31) & 1), data[3]);
3129 
3130 	sbuf_printf(sb, " start=%d -> end=%d\n", rspq->rspq_dump_start,
3131 	    (rspq->rspq_dump_start + rspq->rspq_dump_count) & (RSPQ_Q_SIZE-1));
3132 
3133 	dump_end = rspq->rspq_dump_start + rspq->rspq_dump_count;
3134 	for (i = rspq->rspq_dump_start; i < dump_end; i++) {
3135 		idx = i & (RSPQ_Q_SIZE-1);
3136 
3137 		rspd = &rspq->desc[idx];
3138 		sbuf_printf(sb, "\tidx=%04d opcode=%02x cpu_idx=%x hash_type=%x cq_idx=%x\n",
3139 		    idx, rspd->rss_hdr.opcode, rspd->rss_hdr.cpu_idx,
3140 		    rspd->rss_hdr.hash_type, be16toh(rspd->rss_hdr.cq_idx));
3141 		sbuf_printf(sb, "\trss_hash_val=%x flags=%08x len_cq=%x intr_gen=%x\n",
3142 		    rspd->rss_hdr.rss_hash_val, be32toh(rspd->flags),
3143 		    be32toh(rspd->len_cq), rspd->intr_gen);
3144 	}
3145 
3146 	err = sbuf_finish(sb);
3147 	sbuf_delete(sb);
3148 	return (err);
3149 }
3150 
3151 static int
3152 t3_dump_txq_eth(SYSCTL_HANDLER_ARGS)
3153 {
3154 	struct sge_txq *txq;
3155 	struct sge_qset *qs;
3156 	int i, j, err, dump_end;
3157 	struct sbuf *sb;
3158 	struct tx_desc *txd;
3159 	uint32_t *WR, wr_hi, wr_lo, gen;
3160 	uint32_t data[4];
3161 
3162 	txq = arg1;
3163 	qs = txq_to_qset(txq, TXQ_ETH);
3164 	if (txq->txq_dump_count == 0) {
3165 		return (0);
3166 	}
3167 	if (txq->txq_dump_count > TX_ETH_Q_SIZE) {
3168 		log(LOG_WARNING,
3169 		    "dump count is too large %d\n", txq->txq_dump_count);
3170 		txq->txq_dump_count = 1;
3171 		return (EINVAL);
3172 	}
3173 	if (txq->txq_dump_start > (TX_ETH_Q_SIZE-1)) {
3174 		log(LOG_WARNING,
3175 		    "dump start of %d is greater than queue size\n",
3176 		    txq->txq_dump_start);
3177 		txq->txq_dump_start = 0;
3178 		return (EINVAL);
3179 	}
3180 	err = t3_sge_read_ecntxt(qs->port->adapter, qs->rspq.cntxt_id, data);
3181 	if (err)
3182 		return (err);
3183 	err = sysctl_wire_old_buffer(req, 0);
3184 	if (err)
3185 		return (err);
3186 	sb = sbuf_new_for_sysctl(NULL, NULL, QDUMP_SBUF_SIZE, req);
3187 
3188 	sbuf_printf(sb, " \n credits=%u GTS=%u index=%u size=%u rspq#=%u cmdq#=%u\n",
3189 	    (data[0] & 0x7fff), ((data[0] >> 15) & 1), (data[0] >> 16),
3190 	    (data[1] & 0xffff), ((data[3] >> 4) & 7), ((data[3] >> 7) & 1));
3191 	sbuf_printf(sb, " TUN=%u TOE=%u generation%u uP token=%u valid=%u\n",
3192 	    ((data[3] >> 8) & 1), ((data[3] >> 9) & 1), ((data[3] >> 10) & 1),
3193 	    ((data[3] >> 11) & 0xfffff), ((data[3] >> 31) & 1));
3194 	sbuf_printf(sb, " qid=%d start=%d -> end=%d\n", qs->idx,
3195 	    txq->txq_dump_start,
3196 	    (txq->txq_dump_start + txq->txq_dump_count) & (TX_ETH_Q_SIZE-1));
3197 
3198 	dump_end = txq->txq_dump_start + txq->txq_dump_count;
3199 	for (i = txq->txq_dump_start; i < dump_end; i++) {
3200 		txd = &txq->desc[i & (TX_ETH_Q_SIZE-1)];
3201 		WR = (uint32_t *)txd->flit;
3202 		wr_hi = ntohl(WR[0]);
3203 		wr_lo = ntohl(WR[1]);
3204 		gen = G_WR_GEN(wr_lo);
3205 
3206 		sbuf_printf(sb," wr_hi %08x wr_lo %08x gen %d\n",
3207 		    wr_hi, wr_lo, gen);
3208 		for (j = 2; j < 30; j += 4)
3209 			sbuf_printf(sb, "\t%08x %08x %08x %08x \n",
3210 			    WR[j], WR[j + 1], WR[j + 2], WR[j + 3]);
3211 
3212 	}
3213 	err = sbuf_finish(sb);
3214 	sbuf_delete(sb);
3215 	return (err);
3216 }
3217 
3218 static int
3219 t3_dump_txq_ctrl(SYSCTL_HANDLER_ARGS)
3220 {
3221 	struct sge_txq *txq;
3222 	struct sge_qset *qs;
3223 	int i, j, err, dump_end;
3224 	struct sbuf *sb;
3225 	struct tx_desc *txd;
3226 	uint32_t *WR, wr_hi, wr_lo, gen;
3227 
3228 	txq = arg1;
3229 	qs = txq_to_qset(txq, TXQ_CTRL);
3230 	if (txq->txq_dump_count == 0) {
3231 		return (0);
3232 	}
3233 	if (txq->txq_dump_count > 256) {
3234 		log(LOG_WARNING,
3235 		    "dump count is too large %d\n", txq->txq_dump_count);
3236 		txq->txq_dump_count = 1;
3237 		return (EINVAL);
3238 	}
3239 	if (txq->txq_dump_start > 255) {
3240 		log(LOG_WARNING,
3241 		    "dump start of %d is greater than queue size\n",
3242 		    txq->txq_dump_start);
3243 		txq->txq_dump_start = 0;
3244 		return (EINVAL);
3245 	}
3246 
3247 	err = sysctl_wire_old_buffer(req, 0);
3248 	if (err != 0)
3249 		return (err);
3250 	sb = sbuf_new_for_sysctl(NULL, NULL, QDUMP_SBUF_SIZE, req);
3251 	sbuf_printf(sb, " qid=%d start=%d -> end=%d\n", qs->idx,
3252 	    txq->txq_dump_start,
3253 	    (txq->txq_dump_start + txq->txq_dump_count) & 255);
3254 
3255 	dump_end = txq->txq_dump_start + txq->txq_dump_count;
3256 	for (i = txq->txq_dump_start; i < dump_end; i++) {
3257 		txd = &txq->desc[i & (255)];
3258 		WR = (uint32_t *)txd->flit;
3259 		wr_hi = ntohl(WR[0]);
3260 		wr_lo = ntohl(WR[1]);
3261 		gen = G_WR_GEN(wr_lo);
3262 
3263 		sbuf_printf(sb," wr_hi %08x wr_lo %08x gen %d\n",
3264 		    wr_hi, wr_lo, gen);
3265 		for (j = 2; j < 30; j += 4)
3266 			sbuf_printf(sb, "\t%08x %08x %08x %08x \n",
3267 			    WR[j], WR[j + 1], WR[j + 2], WR[j + 3]);
3268 
3269 	}
3270 	err = sbuf_finish(sb);
3271 	sbuf_delete(sb);
3272 	return (err);
3273 }
3274 
3275 static int
3276 t3_set_coalesce_usecs(SYSCTL_HANDLER_ARGS)
3277 {
3278 	adapter_t *sc = arg1;
3279 	struct qset_params *qsp = &sc->params.sge.qset[0];
3280 	int coalesce_usecs;
3281 	struct sge_qset *qs;
3282 	int i, j, err, nqsets = 0;
3283 	struct mtx *lock;
3284 
3285 	if ((sc->flags & FULL_INIT_DONE) == 0)
3286 		return (ENXIO);
3287 
3288 	coalesce_usecs = qsp->coalesce_usecs;
3289         err = sysctl_handle_int(oidp, &coalesce_usecs, arg2, req);
3290 
3291 	if (err != 0) {
3292 		return (err);
3293 	}
3294 	if (coalesce_usecs == qsp->coalesce_usecs)
3295 		return (0);
3296 
3297 	for (i = 0; i < sc->params.nports; i++)
3298 		for (j = 0; j < sc->port[i].nqsets; j++)
3299 			nqsets++;
3300 
3301 	coalesce_usecs = max(1, coalesce_usecs);
3302 
3303 	for (i = 0; i < nqsets; i++) {
3304 		qs = &sc->sge.qs[i];
3305 		qsp = &sc->params.sge.qset[i];
3306 		qsp->coalesce_usecs = coalesce_usecs;
3307 
3308 		lock = (sc->flags & USING_MSIX) ? &qs->rspq.lock :
3309 			    &sc->sge.qs[0].rspq.lock;
3310 
3311 		mtx_lock(lock);
3312 		t3_update_qset_coalesce(qs, qsp);
3313 		t3_write_reg(sc, A_SG_GTS, V_RSPQ(qs->rspq.cntxt_id) |
3314 		    V_NEWTIMER(qs->rspq.holdoff_tmr));
3315 		mtx_unlock(lock);
3316 	}
3317 
3318 	return (0);
3319 }
3320 
3321 static int
3322 t3_pkt_timestamp(SYSCTL_HANDLER_ARGS)
3323 {
3324 	adapter_t *sc = arg1;
3325 	int rc, timestamp;
3326 
3327 	if ((sc->flags & FULL_INIT_DONE) == 0)
3328 		return (ENXIO);
3329 
3330 	timestamp = sc->timestamp;
3331 	rc = sysctl_handle_int(oidp, &timestamp, arg2, req);
3332 
3333 	if (rc != 0)
3334 		return (rc);
3335 
3336 	if (timestamp != sc->timestamp) {
3337 		t3_set_reg_field(sc, A_TP_PC_CONFIG2, F_ENABLERXPKTTMSTPRSS,
3338 		    timestamp ? F_ENABLERXPKTTMSTPRSS : 0);
3339 		sc->timestamp = timestamp;
3340 	}
3341 
3342 	return (0);
3343 }
3344 
3345 void
3346 t3_add_attach_sysctls(adapter_t *sc)
3347 {
3348 	struct sysctl_ctx_list *ctx;
3349 	struct sysctl_oid_list *children;
3350 
3351 	ctx = device_get_sysctl_ctx(sc->dev);
3352 	children = SYSCTL_CHILDREN(device_get_sysctl_tree(sc->dev));
3353 
3354 	/* random information */
3355 	SYSCTL_ADD_STRING(ctx, children, OID_AUTO,
3356 	    "firmware_version",
3357 	    CTLFLAG_RD, sc->fw_version,
3358 	    0, "firmware version");
3359 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO,
3360 	    "hw_revision",
3361 	    CTLFLAG_RD, &sc->params.rev,
3362 	    0, "chip model");
3363 	SYSCTL_ADD_STRING(ctx, children, OID_AUTO,
3364 	    "port_types",
3365 	    CTLFLAG_RD, sc->port_types,
3366 	    0, "type of ports");
3367 	SYSCTL_ADD_INT(ctx, children, OID_AUTO,
3368 	    "enable_debug",
3369 	    CTLFLAG_RW, &cxgb_debug,
3370 	    0, "enable verbose debugging output");
3371 	SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tunq_coalesce",
3372 	    CTLFLAG_RD, &sc->tunq_coalesce,
3373 	    "#tunneled packets freed");
3374 	SYSCTL_ADD_INT(ctx, children, OID_AUTO,
3375 	    "txq_overrun",
3376 	    CTLFLAG_RD, &txq_fills,
3377 	    0, "#times txq overrun");
3378 	SYSCTL_ADD_UINT(ctx, children, OID_AUTO,
3379 	    "core_clock",
3380 	    CTLFLAG_RD, &sc->params.vpd.cclk,
3381 	    0, "core clock frequency (in KHz)");
3382 }
3383 
3384 
3385 static const char *rspq_name = "rspq";
3386 static const char *txq_names[] =
3387 {
3388 	"txq_eth",
3389 	"txq_ofld",
3390 	"txq_ctrl"
3391 };
3392 
3393 static int
3394 sysctl_handle_macstat(SYSCTL_HANDLER_ARGS)
3395 {
3396 	struct port_info *p = arg1;
3397 	uint64_t *parg;
3398 
3399 	if (!p)
3400 		return (EINVAL);
3401 
3402 	cxgb_refresh_stats(p);
3403 	parg = (uint64_t *) ((uint8_t *)&p->mac.stats + arg2);
3404 
3405 	return (sysctl_handle_64(oidp, parg, 0, req));
3406 }
3407 
3408 void
3409 t3_add_configured_sysctls(adapter_t *sc)
3410 {
3411 	struct sysctl_ctx_list *ctx;
3412 	struct sysctl_oid_list *children;
3413 	int i, j;
3414 
3415 	ctx = device_get_sysctl_ctx(sc->dev);
3416 	children = SYSCTL_CHILDREN(device_get_sysctl_tree(sc->dev));
3417 
3418 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO,
3419 	    "intr_coal",
3420 	    CTLTYPE_INT|CTLFLAG_RW, sc,
3421 	    0, t3_set_coalesce_usecs,
3422 	    "I", "interrupt coalescing timer (us)");
3423 
3424 	SYSCTL_ADD_PROC(ctx, children, OID_AUTO,
3425 	    "pkt_timestamp",
3426 	    CTLTYPE_INT | CTLFLAG_RW, sc,
3427 	    0, t3_pkt_timestamp,
3428 	    "I", "provide packet timestamp instead of connection hash");
3429 
3430 	for (i = 0; i < sc->params.nports; i++) {
3431 		struct port_info *pi = &sc->port[i];
3432 		struct sysctl_oid *poid;
3433 		struct sysctl_oid_list *poidlist;
3434 		struct mac_stats *mstats = &pi->mac.stats;
3435 
3436 		snprintf(pi->namebuf, PORT_NAME_LEN, "port%d", i);
3437 		poid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO,
3438 		    pi->namebuf, CTLFLAG_RD, NULL, "port statistics");
3439 		poidlist = SYSCTL_CHILDREN(poid);
3440 		SYSCTL_ADD_UINT(ctx, poidlist, OID_AUTO,
3441 		    "nqsets", CTLFLAG_RD, &pi->nqsets,
3442 		    0, "#queue sets");
3443 
3444 		for (j = 0; j < pi->nqsets; j++) {
3445 			struct sge_qset *qs = &sc->sge.qs[pi->first_qset + j];
3446 			struct sysctl_oid *qspoid, *rspqpoid, *txqpoid,
3447 					  *ctrlqpoid, *lropoid;
3448 			struct sysctl_oid_list *qspoidlist, *rspqpoidlist,
3449 					       *txqpoidlist, *ctrlqpoidlist,
3450 					       *lropoidlist;
3451 			struct sge_txq *txq = &qs->txq[TXQ_ETH];
3452 
3453 			snprintf(qs->namebuf, QS_NAME_LEN, "qs%d", j);
3454 
3455 			qspoid = SYSCTL_ADD_NODE(ctx, poidlist, OID_AUTO,
3456 			    qs->namebuf, CTLFLAG_RD, NULL, "qset statistics");
3457 			qspoidlist = SYSCTL_CHILDREN(qspoid);
3458 
3459 			SYSCTL_ADD_UINT(ctx, qspoidlist, OID_AUTO, "fl0_empty",
3460 					CTLFLAG_RD, &qs->fl[0].empty, 0,
3461 					"freelist #0 empty");
3462 			SYSCTL_ADD_UINT(ctx, qspoidlist, OID_AUTO, "fl1_empty",
3463 					CTLFLAG_RD, &qs->fl[1].empty, 0,
3464 					"freelist #1 empty");
3465 
3466 			rspqpoid = SYSCTL_ADD_NODE(ctx, qspoidlist, OID_AUTO,
3467 			    rspq_name, CTLFLAG_RD, NULL, "rspq statistics");
3468 			rspqpoidlist = SYSCTL_CHILDREN(rspqpoid);
3469 
3470 			txqpoid = SYSCTL_ADD_NODE(ctx, qspoidlist, OID_AUTO,
3471 			    txq_names[0], CTLFLAG_RD, NULL, "txq statistics");
3472 			txqpoidlist = SYSCTL_CHILDREN(txqpoid);
3473 
3474 			ctrlqpoid = SYSCTL_ADD_NODE(ctx, qspoidlist, OID_AUTO,
3475 			    txq_names[2], CTLFLAG_RD, NULL, "ctrlq statistics");
3476 			ctrlqpoidlist = SYSCTL_CHILDREN(ctrlqpoid);
3477 
3478 			lropoid = SYSCTL_ADD_NODE(ctx, qspoidlist, OID_AUTO,
3479 			    "lro_stats", CTLFLAG_RD, NULL, "LRO statistics");
3480 			lropoidlist = SYSCTL_CHILDREN(lropoid);
3481 
3482 			SYSCTL_ADD_UINT(ctx, rspqpoidlist, OID_AUTO, "size",
3483 			    CTLFLAG_RD, &qs->rspq.size,
3484 			    0, "#entries in response queue");
3485 			SYSCTL_ADD_UINT(ctx, rspqpoidlist, OID_AUTO, "cidx",
3486 			    CTLFLAG_RD, &qs->rspq.cidx,
3487 			    0, "consumer index");
3488 			SYSCTL_ADD_UINT(ctx, rspqpoidlist, OID_AUTO, "credits",
3489 			    CTLFLAG_RD, &qs->rspq.credits,
3490 			    0, "#credits");
3491 			SYSCTL_ADD_UINT(ctx, rspqpoidlist, OID_AUTO, "starved",
3492 			    CTLFLAG_RD, &qs->rspq.starved,
3493 			    0, "#times starved");
3494 			SYSCTL_ADD_UAUTO(ctx, rspqpoidlist, OID_AUTO, "phys_addr",
3495 			    CTLFLAG_RD, &qs->rspq.phys_addr,
3496 			    "physical_address_of the queue");
3497 			SYSCTL_ADD_UINT(ctx, rspqpoidlist, OID_AUTO, "dump_start",
3498 			    CTLFLAG_RW, &qs->rspq.rspq_dump_start,
3499 			    0, "start rspq dump entry");
3500 			SYSCTL_ADD_UINT(ctx, rspqpoidlist, OID_AUTO, "dump_count",
3501 			    CTLFLAG_RW, &qs->rspq.rspq_dump_count,
3502 			    0, "#rspq entries to dump");
3503 			SYSCTL_ADD_PROC(ctx, rspqpoidlist, OID_AUTO, "qdump",
3504 			    CTLTYPE_STRING | CTLFLAG_RD, &qs->rspq,
3505 			    0, t3_dump_rspq, "A", "dump of the response queue");
3506 
3507 			SYSCTL_ADD_UQUAD(ctx, txqpoidlist, OID_AUTO, "dropped",
3508 			    CTLFLAG_RD, &qs->txq[TXQ_ETH].txq_mr->br_drops,
3509 			    "#tunneled packets dropped");
3510 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "sendqlen",
3511 			    CTLFLAG_RD, &qs->txq[TXQ_ETH].sendq.mq_len,
3512 			    0, "#tunneled packets waiting to be sent");
3513 #if 0
3514 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "queue_pidx",
3515 			    CTLFLAG_RD, (uint32_t *)(uintptr_t)&qs->txq[TXQ_ETH].txq_mr.br_prod,
3516 			    0, "#tunneled packets queue producer index");
3517 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "queue_cidx",
3518 			    CTLFLAG_RD, (uint32_t *)(uintptr_t)&qs->txq[TXQ_ETH].txq_mr.br_cons,
3519 			    0, "#tunneled packets queue consumer index");
3520 #endif
3521 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "processed",
3522 			    CTLFLAG_RD, &qs->txq[TXQ_ETH].processed,
3523 			    0, "#tunneled packets processed by the card");
3524 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "cleaned",
3525 			    CTLFLAG_RD, &txq->cleaned,
3526 			    0, "#tunneled packets cleaned");
3527 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "in_use",
3528 			    CTLFLAG_RD, &txq->in_use,
3529 			    0, "#tunneled packet slots in use");
3530 			SYSCTL_ADD_UQUAD(ctx, txqpoidlist, OID_AUTO, "frees",
3531 			    CTLFLAG_RD, &txq->txq_frees,
3532 			    "#tunneled packets freed");
3533 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "skipped",
3534 			    CTLFLAG_RD, &txq->txq_skipped,
3535 			    0, "#tunneled packet descriptors skipped");
3536 			SYSCTL_ADD_UQUAD(ctx, txqpoidlist, OID_AUTO, "coalesced",
3537 			    CTLFLAG_RD, &txq->txq_coalesced,
3538 			    "#tunneled packets coalesced");
3539 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "enqueued",
3540 			    CTLFLAG_RD, &txq->txq_enqueued,
3541 			    0, "#tunneled packets enqueued to hardware");
3542 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "stopped_flags",
3543 			    CTLFLAG_RD, &qs->txq_stopped,
3544 			    0, "tx queues stopped");
3545 			SYSCTL_ADD_UAUTO(ctx, txqpoidlist, OID_AUTO, "phys_addr",
3546 			    CTLFLAG_RD, &txq->phys_addr,
3547 			    "physical_address_of the queue");
3548 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "qgen",
3549 			    CTLFLAG_RW, &qs->txq[TXQ_ETH].gen,
3550 			    0, "txq generation");
3551 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "hw_cidx",
3552 			    CTLFLAG_RD, &txq->cidx,
3553 			    0, "hardware queue cidx");
3554 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "hw_pidx",
3555 			    CTLFLAG_RD, &txq->pidx,
3556 			    0, "hardware queue pidx");
3557 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "dump_start",
3558 			    CTLFLAG_RW, &qs->txq[TXQ_ETH].txq_dump_start,
3559 			    0, "txq start idx for dump");
3560 			SYSCTL_ADD_UINT(ctx, txqpoidlist, OID_AUTO, "dump_count",
3561 			    CTLFLAG_RW, &qs->txq[TXQ_ETH].txq_dump_count,
3562 			    0, "txq #entries to dump");
3563 			SYSCTL_ADD_PROC(ctx, txqpoidlist, OID_AUTO, "qdump",
3564 			    CTLTYPE_STRING | CTLFLAG_RD, &qs->txq[TXQ_ETH],
3565 			    0, t3_dump_txq_eth, "A", "dump of the transmit queue");
3566 
3567 			SYSCTL_ADD_UINT(ctx, ctrlqpoidlist, OID_AUTO, "dump_start",
3568 			    CTLFLAG_RW, &qs->txq[TXQ_CTRL].txq_dump_start,
3569 			    0, "ctrlq start idx for dump");
3570 			SYSCTL_ADD_UINT(ctx, ctrlqpoidlist, OID_AUTO, "dump_count",
3571 			    CTLFLAG_RW, &qs->txq[TXQ_CTRL].txq_dump_count,
3572 			    0, "ctrl #entries to dump");
3573 			SYSCTL_ADD_PROC(ctx, ctrlqpoidlist, OID_AUTO, "qdump",
3574 			    CTLTYPE_STRING | CTLFLAG_RD, &qs->txq[TXQ_CTRL],
3575 			    0, t3_dump_txq_ctrl, "A", "dump of the transmit queue");
3576 
3577 			SYSCTL_ADD_U64(ctx, lropoidlist, OID_AUTO, "lro_queued",
3578 			    CTLFLAG_RD, &qs->lro.ctrl.lro_queued, 0, NULL);
3579 			SYSCTL_ADD_U64(ctx, lropoidlist, OID_AUTO, "lro_flushed",
3580 			    CTLFLAG_RD, &qs->lro.ctrl.lro_flushed, 0, NULL);
3581 			SYSCTL_ADD_U64(ctx, lropoidlist, OID_AUTO, "lro_bad_csum",
3582 			    CTLFLAG_RD, &qs->lro.ctrl.lro_bad_csum, 0, NULL);
3583 			SYSCTL_ADD_INT(ctx, lropoidlist, OID_AUTO, "lro_cnt",
3584 			    CTLFLAG_RD, &qs->lro.ctrl.lro_cnt, 0, NULL);
3585 		}
3586 
3587 		/* Now add a node for mac stats. */
3588 		poid = SYSCTL_ADD_NODE(ctx, poidlist, OID_AUTO, "mac_stats",
3589 		    CTLFLAG_RD, NULL, "MAC statistics");
3590 		poidlist = SYSCTL_CHILDREN(poid);
3591 
3592 		/*
3593 		 * We (ab)use the length argument (arg2) to pass on the offset
3594 		 * of the data that we are interested in.  This is only required
3595 		 * for the quad counters that are updated from the hardware (we
3596 		 * make sure that we return the latest value).
3597 		 * sysctl_handle_macstat first updates *all* the counters from
3598 		 * the hardware, and then returns the latest value of the
3599 		 * requested counter.  Best would be to update only the
3600 		 * requested counter from hardware, but t3_mac_update_stats()
3601 		 * hides all the register details and we don't want to dive into
3602 		 * all that here.
3603 		 */
3604 #define CXGB_SYSCTL_ADD_QUAD(a)	SYSCTL_ADD_OID(ctx, poidlist, OID_AUTO, #a, \
3605     (CTLTYPE_U64 | CTLFLAG_RD), pi, offsetof(struct mac_stats, a), \
3606     sysctl_handle_macstat, "QU", 0)
3607 		CXGB_SYSCTL_ADD_QUAD(tx_octets);
3608 		CXGB_SYSCTL_ADD_QUAD(tx_octets_bad);
3609 		CXGB_SYSCTL_ADD_QUAD(tx_frames);
3610 		CXGB_SYSCTL_ADD_QUAD(tx_mcast_frames);
3611 		CXGB_SYSCTL_ADD_QUAD(tx_bcast_frames);
3612 		CXGB_SYSCTL_ADD_QUAD(tx_pause);
3613 		CXGB_SYSCTL_ADD_QUAD(tx_deferred);
3614 		CXGB_SYSCTL_ADD_QUAD(tx_late_collisions);
3615 		CXGB_SYSCTL_ADD_QUAD(tx_total_collisions);
3616 		CXGB_SYSCTL_ADD_QUAD(tx_excess_collisions);
3617 		CXGB_SYSCTL_ADD_QUAD(tx_underrun);
3618 		CXGB_SYSCTL_ADD_QUAD(tx_len_errs);
3619 		CXGB_SYSCTL_ADD_QUAD(tx_mac_internal_errs);
3620 		CXGB_SYSCTL_ADD_QUAD(tx_excess_deferral);
3621 		CXGB_SYSCTL_ADD_QUAD(tx_fcs_errs);
3622 		CXGB_SYSCTL_ADD_QUAD(tx_frames_64);
3623 		CXGB_SYSCTL_ADD_QUAD(tx_frames_65_127);
3624 		CXGB_SYSCTL_ADD_QUAD(tx_frames_128_255);
3625 		CXGB_SYSCTL_ADD_QUAD(tx_frames_256_511);
3626 		CXGB_SYSCTL_ADD_QUAD(tx_frames_512_1023);
3627 		CXGB_SYSCTL_ADD_QUAD(tx_frames_1024_1518);
3628 		CXGB_SYSCTL_ADD_QUAD(tx_frames_1519_max);
3629 		CXGB_SYSCTL_ADD_QUAD(rx_octets);
3630 		CXGB_SYSCTL_ADD_QUAD(rx_octets_bad);
3631 		CXGB_SYSCTL_ADD_QUAD(rx_frames);
3632 		CXGB_SYSCTL_ADD_QUAD(rx_mcast_frames);
3633 		CXGB_SYSCTL_ADD_QUAD(rx_bcast_frames);
3634 		CXGB_SYSCTL_ADD_QUAD(rx_pause);
3635 		CXGB_SYSCTL_ADD_QUAD(rx_fcs_errs);
3636 		CXGB_SYSCTL_ADD_QUAD(rx_align_errs);
3637 		CXGB_SYSCTL_ADD_QUAD(rx_symbol_errs);
3638 		CXGB_SYSCTL_ADD_QUAD(rx_data_errs);
3639 		CXGB_SYSCTL_ADD_QUAD(rx_sequence_errs);
3640 		CXGB_SYSCTL_ADD_QUAD(rx_runt);
3641 		CXGB_SYSCTL_ADD_QUAD(rx_jabber);
3642 		CXGB_SYSCTL_ADD_QUAD(rx_short);
3643 		CXGB_SYSCTL_ADD_QUAD(rx_too_long);
3644 		CXGB_SYSCTL_ADD_QUAD(rx_mac_internal_errs);
3645 		CXGB_SYSCTL_ADD_QUAD(rx_cong_drops);
3646 		CXGB_SYSCTL_ADD_QUAD(rx_frames_64);
3647 		CXGB_SYSCTL_ADD_QUAD(rx_frames_65_127);
3648 		CXGB_SYSCTL_ADD_QUAD(rx_frames_128_255);
3649 		CXGB_SYSCTL_ADD_QUAD(rx_frames_256_511);
3650 		CXGB_SYSCTL_ADD_QUAD(rx_frames_512_1023);
3651 		CXGB_SYSCTL_ADD_QUAD(rx_frames_1024_1518);
3652 		CXGB_SYSCTL_ADD_QUAD(rx_frames_1519_max);
3653 #undef CXGB_SYSCTL_ADD_QUAD
3654 
3655 #define CXGB_SYSCTL_ADD_ULONG(a) SYSCTL_ADD_ULONG(ctx, poidlist, OID_AUTO, #a, \
3656     CTLFLAG_RD, &mstats->a, 0)
3657 		CXGB_SYSCTL_ADD_ULONG(tx_fifo_parity_err);
3658 		CXGB_SYSCTL_ADD_ULONG(rx_fifo_parity_err);
3659 		CXGB_SYSCTL_ADD_ULONG(tx_fifo_urun);
3660 		CXGB_SYSCTL_ADD_ULONG(rx_fifo_ovfl);
3661 		CXGB_SYSCTL_ADD_ULONG(serdes_signal_loss);
3662 		CXGB_SYSCTL_ADD_ULONG(xaui_pcs_ctc_err);
3663 		CXGB_SYSCTL_ADD_ULONG(xaui_pcs_align_change);
3664 		CXGB_SYSCTL_ADD_ULONG(num_toggled);
3665 		CXGB_SYSCTL_ADD_ULONG(num_resets);
3666 		CXGB_SYSCTL_ADD_ULONG(link_faults);
3667 #undef CXGB_SYSCTL_ADD_ULONG
3668 	}
3669 }
3670 
3671 /**
3672  *	t3_get_desc - dump an SGE descriptor for debugging purposes
3673  *	@qs: the queue set
3674  *	@qnum: identifies the specific queue (0..2: Tx, 3:response, 4..5: Rx)
3675  *	@idx: the descriptor index in the queue
3676  *	@data: where to dump the descriptor contents
3677  *
3678  *	Dumps the contents of a HW descriptor of an SGE queue.  Returns the
3679  *	size of the descriptor.
3680  */
3681 int
3682 t3_get_desc(const struct sge_qset *qs, unsigned int qnum, unsigned int idx,
3683 		unsigned char *data)
3684 {
3685 	if (qnum >= 6)
3686 		return (EINVAL);
3687 
3688 	if (qnum < 3) {
3689 		if (!qs->txq[qnum].desc || idx >= qs->txq[qnum].size)
3690 			return -EINVAL;
3691 		memcpy(data, &qs->txq[qnum].desc[idx], sizeof(struct tx_desc));
3692 		return sizeof(struct tx_desc);
3693 	}
3694 
3695 	if (qnum == 3) {
3696 		if (!qs->rspq.desc || idx >= qs->rspq.size)
3697 			return (EINVAL);
3698 		memcpy(data, &qs->rspq.desc[idx], sizeof(struct rsp_desc));
3699 		return sizeof(struct rsp_desc);
3700 	}
3701 
3702 	qnum -= 4;
3703 	if (!qs->fl[qnum].desc || idx >= qs->fl[qnum].size)
3704 		return (EINVAL);
3705 	memcpy(data, &qs->fl[qnum].desc[idx], sizeof(struct rx_desc));
3706 	return sizeof(struct rx_desc);
3707 }
3708