xref: /illumos-gate/usr/src/uts/common/io/chxge/sge.c (revision 8119dad84d6416f13557b0ba8e2aaf9064cbcfd3)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 /*
28  * This file is part of the Chelsio T1 Ethernet driver.
29  *
30  * Copyright (C) 2003-2005 Chelsio Communications.  All rights reserved.
31  */
32 
33 #include <sys/types.h>
34 #include <sys/param.h>
35 #include <sys/cmn_err.h>
36 #include <sys/sunddi.h>
37 #include <sys/kmem.h>
38 #include <sys/cmn_err.h>
39 #include <sys/byteorder.h>
40 #include <sys/atomic.h>
41 #include <sys/stropts.h>
42 #include <sys/stream.h>
43 #include <sys/strsubr.h>
44 #include <sys/dlpi.h>
45 #include <sys/kstat.h>
46 #include <sys/ethernet.h>
47 #include <netinet/in.h>
48 #include <netinet/udp.h>
49 #include <inet/common.h>
50 #include <inet/nd.h>
51 #include <inet/ip.h>
52 #include <inet/tcp.h>
53 #include <netinet/udp.h>
54 #include <sys/gld.h>
55 #include "ostypes.h"
56 #include "common.h"
57 #ifdef CONFIG_CHELSIO_T1_1G
58 #include "fpga_defs.h"
59 #endif
60 #include "regs.h"
61 #include "suni1x10gexp_regs.h"
62 #include "sge.h"
63 #include "espi.h"
64 
65 #include "ch.h"
66 
67 extern uint32_t buffers_in_use[];
68 
69 uint32_t sge_cmdq0_cnt = SGE_CMDQ0_E_N;
70 uint32_t sge_cmdq1_cnt = SGE_CMDQ1_E_N;
71 uint32_t sge_flq0_cnt = SGE_FREELQ0_E_N;
72 uint32_t sge_flq1_cnt = SGE_FREELQ1_E_N;
73 uint32_t sge_respq_cnt = SGE_RESPQ_E_N;
74 
75 uint32_t sge_cmdq0_cnt_orig = SGE_CMDQ0_E_N;
76 uint32_t sge_cmdq1_cnt_orig = SGE_CMDQ1_E_N;
77 uint32_t sge_flq0_cnt_orig = SGE_FREELQ0_E_N;
78 uint32_t sge_flq1_cnt_orig = SGE_FREELQ1_E_N;
79 uint32_t sge_respq_cnt_orig = SGE_RESPQ_E_N;
80 
81 #ifdef HOST_PAUSE
82 uint32_t do_host_pause = 1;
83 uint32_t flq_pause_window = 64;
84 #endif
85 
86 static uint64_t os_freelist_buffer_alloc(ch_t *sa, int sz, mblk_t **mb,
87     ulong_t *dh);
88 void pe_os_free_contig(ch_t *, size_t, void *, uint64_t, ulong_t, ulong_t);
89 
90 static inline uint32_t t1_sge_rx(pesge *sge, freelQ_t *Q,
91     unsigned int len, unsigned int offload);
92 #ifdef HOST_PAUSE
93 static void t1_sge_check_pause(pesge *sge, struct freelQ *Q);
94 #endif
95 static void alloc_freelQ_buffers(pesge *sge, struct freelQ *Q);
96 static void freelQs_empty(pesge *sge);
97 static void free_cmdQ_buffers(pesge *sge, cmdQ_t *Q, uint32_t credits_pend);
98 static int alloc_rx_resources(pesge *sge, struct sge_params *p);
99 static int alloc_tx_resources(pesge *sge, struct sge_params *p);
100 static inline void setup_ring_params(ch_t *adapter, u64 addr, u32 size,
101     int base_reg_lo, int base_reg_hi, int size_reg);
102 static void configure_sge(pesge *sge, struct sge_params *p);
103 static void free_freelQ_buffers(pesge *sge, struct freelQ *Q);
104 static void free_rx_resources(pesge *sge);
105 static void free_tx_resources(pesge *sge);
106 static inline unsigned int jumbo_payload_capacity(pesge *sge);
107 #ifdef SUN_KSTATS
108 static int sge_kstat_setup(pesge *);
109 static void sge_kstat_remove(pesge *);
110 static int sge_kstat_update(p_kstat_t, int);
111 #endif
112 static uint16_t calc_ocsum(mblk_t *, int);
113 
114 /*
115  * Local routines.
116  */
117 static inline void sge_ring_doorbell(pesge *sge, u32 control_reg);
118 
119 static inline void
120 sge_ring_doorbell(pesge *sge, u32 control_reg)
121 {
122 	membar_producer();
123 	t1_write_reg_4(sge->obj, A_SG_DOORBELL, control_reg);
124 }
125 
126 /*
127  * DESC:
128  *
129  * NOTES:   Must have at least 1 command queue and 1 freelist queue.
130  *
131  */
132 pesge *
133 t1_sge_create(ch_t *sa, struct sge_params *p)
134 {
135 	pesge *sge;
136 
137 	sge = t1_os_malloc_wait_zero(sizeof (pesge));
138 
139 	if (sge == NULL)
140 		goto error_no_mem;
141 
142 	memset(sge, 0, sizeof (*sge));
143 
144 	/*
145 	 * PR2928 & PR3309
146 	 * set default timeout value - 20 msec
147 	 * we set the initial value to 2 which gurantees at least one tick.
148 	 */
149 	if (is_T2(sa))
150 		sge->ptimeout = 1;
151 
152 	sge->obj = sa;
153 #ifdef SUN_KSTATS
154 	if (sge_kstat_setup(sge) != 0)
155 		goto t1_sge_create_fail1;
156 #endif
157 	p->cmdQ_size[0] = sge_cmdq0_cnt;
158 	p->cmdQ_size[1] = sge_cmdq1_cnt;
159 
160 	/* note that jumbo frame index is inverted for T2 */
161 	if (is_T2(sa)) {
162 		p->freelQ_size[1] = sge_flq0_cnt;
163 		p->freelQ_size[0] = sge_flq1_cnt;
164 	} else {
165 		p->freelQ_size[0] = sge_flq0_cnt;
166 		p->freelQ_size[1] = sge_flq1_cnt;
167 	}
168 
169 #if CH_DEBUG
170 	/* DEBUG only */
171 	cmn_err(CE_NOTE, "sge: %p\n", sge);
172 	cmn_err(CE_NOTE, "&sge->cmdQ[0]: %p\n", &sge->cmdQ[0]);
173 	cmn_err(CE_NOTE, "&sge->freelQ[0]: %p\n", &sge->freelQ[0]);
174 	cmn_err(CE_NOTE, "&sge->freelQ[1]: %p\n", &sge->freelQ[1]);
175 	cmn_err(CE_NOTE, "&sge->respQ: %p\n", &sge->respQ);
176 	cmn_err(CE_NOTE, "&sge->intr_cnt: %p\n", &sge->intr_cnt);
177 #endif
178 #ifdef SUN_KSTATS
179 	goto error_no_mem;
180 
181 t1_sge_create_fail1:
182 	t1_os_free(sge, sizeof (pesge));
183 	sge = NULL;
184 #endif
185 error_no_mem:
186 	return (sge);
187 }
188 
189 int
190 t1_sge_destroy(pesge* sge)
191 {
192 	if (sge != NULL) {
193 		free_tx_resources(sge);
194 		free_rx_resources(sge);
195 
196 		/* PR2928 & PR3309 */
197 		if ((is_T2(sge->obj)) && (sge->pskb))
198 			pe_free_fake_arp(sge->pskb);
199 #ifdef SUN_KSTATS
200 		sge_kstat_remove(sge);
201 #endif
202 		t1_os_free(sge, sizeof (pesge));
203 	}
204 	return (0);
205 }
206 
207 /*
208  * PR2928 & PR3309
209  * call out event from timeout
210  *
211  * there is a potential race between the timeout and the close.
212  * unless we protect the timeout, the close could occur at the
213  * same time. Then if the timeout service routine was slow or
214  * interrupted, the sge_stop() could complete with a timeoutID
215  * that has expired, thus letting another timeout occur. If the
216  * service routine was delayed still further, a detach could occur.
217  * the second time could then end up accessing memory that has been
218  * released back to the system. Bad things could then occur. We
219  * set a flag in sge_stop() to tell the service routine not to
220  * issue further timeouts. sge_stop() will block until a timeout
221  * has occured. If the command Q is full then we shouldn't put out
222  * an arp.
223  */
224 
225 void
226 t1_espi_workaround(ch_t *adapter)
227 {
228 	pesge *sge = adapter->sge;
229 	ch_t *chp = (ch_t *)sge->obj;
230 	int rv = 1;
231 
232 	if ((chp->ch_state == PERUNNING) &&
233 	    atomic_read(sge->cmdQ[0].cq_asleep)) {
234 		u32 seop;
235 		seop = t1_espi_get_mon(adapter, 0x930, 0);
236 		if ((seop & 0xfff0fff) == 0xfff) {
237 			/* after first arp */
238 			if (sge->pskb) {
239 				rv = pe_start(adapter, (mblk_t *)sge->pskb,
240 				    CH_ARP);
241 				if (!rv)
242 					sge->intr_cnt.arp_sent++;
243 			}
244 		}
245 	}
246 #ifdef HOST_PAUSE
247 	/*
248 	 * If we are already in sge_data_in, then we can skip calling
249 	 * t1_sge_check_pause() this clock cycle. lockstat showed that
250 	 * we were blocking on the mutex ~ 2% of the time.
251 	 */
252 	if (mutex_tryenter(&adapter->ch_intr)) {
253 		t1_sge_check_pause(sge, &sge->freelQ[0]);
254 		t1_sge_check_pause(sge, &sge->freelQ[1]);
255 		mutex_exit(&adapter->ch_intr);
256 	}
257 #endif
258 }
259 
260 int
261 sge_start(pesge *sge)
262 {
263 	t1_write_reg_4(sge->obj, A_SG_CONTROL, sge->sge_control);
264 	/* PR2928 & PR3309, also need to avoid Pause deadlock */
265 	ch_init_cyclic(sge->obj, &sge->espi_wa_cyclic,
266 	    (void (*)(void *))t1_espi_workaround, sge->obj);
267 	ch_start_cyclic(&sge->espi_wa_cyclic, sge->ptimeout);
268 	return (0);
269 }
270 
271 /*
272  * Disables SGE queues.
273  */
274 int
275 sge_stop(pesge *sge)
276 {
277 	uint32_t status;
278 	int loops;
279 
280 	DBGASSERT(sge);
281 
282 	/* PR2928 & PR3309, also need to avoid Pause deadlock */
283 	t1_write_reg_4(sge->obj, A_SG_CONTROL, 0x0);
284 
285 	/* wait until there's no more outstanding interrupts pending */
286 	loops = 0;
287 	do {
288 		status = t1_read_reg_4(sge->obj, A_SG_INT_CAUSE);
289 		t1_write_reg_4(sge->obj, A_SG_INT_CAUSE, status);
290 		drv_usecwait(125);
291 		loops++;
292 	} while (status && (loops < 1000));
293 
294 	ch_stop_cyclic(&sge->espi_wa_cyclic);
295 
296 	return (0);
297 }
298 
299 uint32_t sge_cmdq_send_fail;
300 
301 int
302 sge_data_out(pesge* sge, int qid, mblk_t *m0,
303     cmdQ_ce_t *cmp, int count, uint32_t flg)
304 {
305 	struct cmdQ *Q = &sge->cmdQ[qid];
306 	ddi_dma_handle_t dh = (ddi_dma_handle_t)sge->cmdQ[qid].cq_dh;
307 	spinlock_t *qlock = &Q->cq_qlock;
308 	cmdQ_e *e;
309 	cmdQ_e *q = Q->cq_entries;
310 	uint32_t credits;
311 	uint32_t pidx;
312 	uint32_t genbit;
313 	uint32_t entries_n = Q->cq_entries_n;
314 	cmdQ_ce_t *ce;
315 	cmdQ_ce_t *cq = Q->cq_centries;
316 	dma_addr_t mapping;
317 	uint32_t j = 0;
318 	uint32_t offset;
319 #if defined(TX_CKSUM_FIX)
320 	uint16_t csum;
321 	uint16_t *csum_loc;
322 #endif
323 #ifdef TX_THREAD_RECLAIM
324 	uint32_t reclaim_cnt;
325 #endif
326 
327 	/*
328 	 * We must exit if we don't have enough free command queue entries
329 	 * available.
330 	 */
331 
332 	spin_lock(qlock);
333 
334 #if defined(TX_CKSUM_FIX)
335 	/*
336 	 * This checksum fix will address a fragmented datagram
337 	 * checksum error. Which will lead to the next packet after
338 	 * the last packet with the More fragment bit set having its
339 	 * checksum corrupted. When the packet reaches this point
340 	 * the 'flg' variable indicates whether a checksum is needed
341 	 * or not. The algorithm is as follows, if the current packet
342 	 * is a More fragment set the count of packets to be checksummed
343 	 * after it to 3. If it't not and the count of is more than 0
344 	 * then calculate the checksum in software, if a hardware checksum
345 	 * was requested. Then decrment the count. Same algorithm applies
346 	 * to TCP.
347 	 */
348 	if (flg & CH_UDP_MF) {
349 		sge->do_udp_csum = 3;
350 	} else if ((flg & CH_UDP) && (sge->do_udp_csum != 0)) {
351 		if ((flg & CH_NO_HWCKSUM) == 0) {
352 			/*
353 			 *  Calc Checksum here.
354 			 */
355 			csum = calc_ocsum(m0,
356 			    sizeof (struct ether_header) + CPL_FORMAT_0_SIZE);
357 			csum_loc = (uint16_t *)(m0->b_rptr +
358 			    sizeof (struct ether_header) + CPL_FORMAT_0_SIZE);
359 			csum_loc += (((*(char *)csum_loc) & 0x0f) << 1);
360 
361 			sge->intr_cnt.tx_soft_cksums++;
362 			((struct udphdr *)(csum_loc))->uh_sum = csum;
363 			((struct cpl_tx_pkt *)m0->b_rptr)->l4_csum_dis = 1;
364 		}
365 		sge->do_udp_csum--;
366 	} else if (flg & CH_TCP_MF) {
367 		sge->do_tcp_csum = 3;
368 	} else if (sge->do_tcp_csum != 0) {
369 		if ((flg & CH_NO_HWCKSUM) == 0) {
370 			sge->intr_cnt.tx_soft_cksums++;
371 			/*
372 			 *  Calc Checksum here.
373 			 */
374 		}
375 		sge->do_tcp_csum--;
376 	}
377 #endif	/* TX_CKSUM_FIX */
378 #ifdef TX_THREAD_RECLAIM
379 	reclaim_cnt = Q->cq_complete;
380 	if (reclaim_cnt > SGE_BATCH_THRESH) {
381 		sge->intr_cnt.tx_reclaims[qid]++;
382 		free_cmdQ_buffers(sge, Q, reclaim_cnt);
383 		Q->cq_complete = 0;
384 	}
385 #endif
386 	genbit = Q->cq_genbit;
387 	pidx = Q->cq_pidx;
388 	credits = Q->cq_credits;
389 
390 	if ((credits - 1) < count) {
391 		spin_unlock(qlock);
392 		sge->intr_cnt.cmdQ_full[qid]++;
393 		return (1);
394 	}
395 
396 	atomic_sub(count, &Q->cq_credits);
397 	Q->cq_pidx += count;
398 	if (Q->cq_pidx >= entries_n) {
399 		Q->cq_pidx -= entries_n;
400 		Q->cq_genbit ^= 1;
401 	}
402 
403 	spin_unlock(qlock);
404 
405 #ifdef SUN_KSTATS
406 	if (count > MBLK_MAX)
407 		sge->intr_cnt.tx_descs[MBLK_MAX - 1]++;
408 	else
409 		sge->intr_cnt.tx_descs[count]++;
410 #endif
411 
412 	ce = &cq[pidx];
413 	*ce = *cmp;
414 	mapping = cmp->ce_pa;
415 	j++;
416 
417 	e = &q[pidx];
418 
419 	offset = (caddr_t)e - (caddr_t)q;
420 
421 	e->Sop =  1;
422 	e->DataValid = 1;
423 	e->BufferLength = cmp->ce_len;
424 	e->AddrHigh = ((u64)mapping >> 32);
425 	e->AddrLow = ((u64)mapping & 0xffffffff);
426 
427 	--count;
428 	if (count > 0) {
429 		unsigned int i;
430 
431 		e->Eop = 0;
432 		wmb();
433 		e->GenerationBit = e->GenerationBit2 = genbit;
434 
435 		for (i = 0; i < count; i++) {
436 
437 			ce++;
438 			e++;
439 			cmp++;
440 			if (++pidx == entries_n) {
441 				pidx = 0;
442 				genbit ^= 1;
443 				/* sync from offset to end of cmdQ */
444 				(void) ddi_dma_sync(dh, (off_t)(offset),
445 				    j*sizeof (*e), DDI_DMA_SYNC_FORDEV);
446 				offset = j = 0;
447 				ce = cq;
448 				e = q;
449 			}
450 
451 			*ce = *cmp;
452 			mapping = cmp->ce_pa;
453 			j++;
454 			e->Sop = 0;
455 			e->DataValid = 1;
456 			e->BufferLength = cmp->ce_len;
457 			e->AddrHigh = ((u64)mapping >> 32);
458 			e->AddrLow = ((u64)mapping & 0xffffffff);
459 
460 			if (i < (count - 1)) {
461 				e->Eop = 0;
462 				wmb();
463 				e->GenerationBit = e->GenerationBit2 = genbit;
464 			}
465 		}
466 	}
467 
468 	ce->ce_mp = m0;
469 
470 	e->Eop = 1;
471 	wmb();
472 	e->GenerationBit = e->GenerationBit2 = genbit;
473 
474 	(void) ddi_dma_sync(dh, (off_t)(offset), j*sizeof (*e),
475 	    DDI_DMA_SYNC_FORDEV);
476 
477 	/*
478 	 * We always ring the doorbell for cmdQ1.  For cmdQ0, we only ring
479 	 * the doorbell if the Q is asleep. There is a natural race, where
480 	 * the hardware is going to sleep just after we checked, however,
481 	 * then the interrupt handler will detect the outstanding TX packet
482 	 * and ring the doorbell for us.
483 	 */
484 	if (qid) {
485 		doorbell_pio(sge, F_CMDQ1_ENABLE);
486 	} else {
487 		if (atomic_read(Q->cq_asleep)) {
488 			atomic_set(&Q->cq_asleep, 0);
489 /* NOT YET		doorbell_pio(sge, F_CMDQ0_ENABLE); */
490 			atomic_set(&Q->cq_pio_pidx, Q->cq_pidx);
491 		}
492 	}
493 	doorbell_pio(sge, F_CMDQ0_ENABLE);
494 
495 	return (0);
496 }
497 
498 #define	SGE_PL_INTR_MASK (F_PL_INTR_SGE_ERR | F_PL_INTR_SGE_DATA)
499 
500 /*
501  * Disable SGE error interrupts.
502  */
503 int
504 t1_sge_intr_disable(pesge* sge)
505 {
506 	u32 val = t1_read_reg_4(sge->obj, A_PL_ENABLE);
507 
508 	t1_write_reg_4(sge->obj, A_PL_ENABLE, val & ~SGE_PL_INTR_MASK);
509 	t1_write_reg_4(sge->obj, A_SG_INT_ENABLE, 0);
510 	return (0);
511 }
512 
513 #define	SGE_INT_ENABLE (F_RESPQ_EXHAUSTED | F_RESPQ_OVERFLOW | \
514 	F_FL_EXHAUSTED | F_PACKET_TOO_BIG | F_PACKET_MISMATCH)
515 
516 /*
517  * Enable SGE error interrupts.
518  */
519 int
520 t1_sge_intr_enable(pesge* sge)
521 {
522 	u32 en = SGE_INT_ENABLE;
523 	u32 val = t1_read_reg_4(sge->obj, A_PL_ENABLE);
524 
525 	t1_write_reg_4(sge->obj, A_PL_ENABLE, val | SGE_PL_INTR_MASK);
526 
527 	if (sge->obj->ch_flags & TSO_CAPABLE)
528 		en &= ~F_PACKET_TOO_BIG;
529 	t1_write_reg_4(sge->obj, A_SG_INT_ENABLE, en);
530 	return (0);
531 }
532 
533 /*
534  * Clear SGE error interrupts.
535  */
536 int
537 t1_sge_intr_clear(pesge* sge)
538 {
539 	t1_write_reg_4(sge->obj, A_PL_CAUSE, SGE_PL_INTR_MASK);
540 	t1_write_reg_4(sge->obj, A_SG_INT_CAUSE, 0xffffffff);
541 	return (0);
542 }
543 
544 #define	SGE_INT_FATAL (F_RESPQ_OVERFLOW | F_PACKET_TOO_BIG | F_PACKET_MISMATCH)
545 
546 int
547 t1_sge_intr_error_handler(pesge *sge)
548 {
549 	peobj *obj = sge->obj;
550 	u32 cause = t1_read_reg_4(obj, A_SG_INT_CAUSE);
551 
552 	if (cause & F_RESPQ_EXHAUSTED)
553 		sge->intr_cnt.respQ_empty++;
554 	if (cause & F_RESPQ_OVERFLOW) {
555 		sge->intr_cnt.respQ_overflow++;
556 		cmn_err(CE_WARN, "%s: SGE response queue overflow\n",
557 		    obj->ch_name);
558 	}
559 	if (cause & F_FL_EXHAUSTED) {
560 		sge->intr_cnt.freelistQ_empty++;
561 		freelQs_empty(sge);
562 	}
563 	if (cause & F_PACKET_TOO_BIG) {
564 		sge->intr_cnt.pkt_too_big++;
565 		cmn_err(CE_WARN, "%s: SGE max packet size exceeded\n",
566 		    obj->ch_name);
567 	}
568 	if (cause & F_PACKET_MISMATCH) {
569 		sge->intr_cnt.pkt_mismatch++;
570 		cmn_err(CE_WARN, "%s: SGE packet mismatch\n",
571 		    obj->ch_name);
572 	}
573 	if (cause & SGE_INT_FATAL)
574 		t1_fatal_err(obj);
575 
576 	t1_write_reg_4(obj, A_SG_INT_CAUSE, cause);
577 	return (0);
578 }
579 
580 /*
581  *
582  * PARAM:   sge     - SGE instance pointer.
583  */
584 int
585 sge_data_in(pesge *sge)
586 {
587 	peobj *adapter = sge->obj;
588 	struct respQ *Q = &sge->respQ;
589 	respQ_e *e;				/* response queue entry */
590 	respQ_e *q = Q->rq_entries;		/* base response queue */
591 	uint32_t cidx = Q->rq_cidx;
592 	uint32_t genbit = Q->rq_genbit;
593 	uint32_t entries_n = Q->rq_entries_n;
594 	uint32_t credits = Q->rq_credits;
595 	uint32_t credits_thresh = Q->rq_credits_thresh;
596 	uint32_t ret = 0;
597 #ifndef TX_THREAD_RECLAIM
598 	uint32_t credits_pend[2] = {0, 0};
599 #endif
600 	uint32_t flags = 0;
601 	uint32_t flagt;
602 	ddi_dma_handle_t dh = (ddi_dma_handle_t)Q->rq_dh;
603 
604 	t1_write_reg_4(adapter, A_PL_CAUSE, F_PL_INTR_SGE_DATA);
605 
606 	/*
607 	 * Catch the case where an interrupt arrives
608 	 * early.
609 	 */
610 	if ((q == NULL) || (dh == NULL)) {
611 		goto check_slow_ints;
612 	}
613 
614 	/* initial response queue entry */
615 	e = &q[cidx];
616 
617 	/* pull physical memory of response queue entry into cache */
618 	(void) ddi_dma_sync(dh, (off_t)((caddr_t)e - (caddr_t)q),
619 	    sizeof (*e), DDI_DMA_SYNC_FORKERNEL);
620 
621 	while (e->GenerationBit == genbit) {
622 		if (--credits < credits_thresh) {
623 			uint32_t n = entries_n - credits - 1;
624 			t1_write_reg_4(adapter, A_SG_RSPQUEUECREDIT, n);
625 			credits += n;
626 		}
627 		if (likely(e->DataValid)) {
628 			(void) t1_sge_rx(sge, &sge->freelQ[e->FreelistQid],
629 			    e->BufferLength, e->Offload);
630 			if ((e->Sop != 1) || (e->Eop != 1)) {
631 				sge->intr_cnt.rx_badEopSop++;
632 				cmn_err(CE_WARN, "bad Sop %d or Eop %d: %d",
633 				    e->Sop, e->Eop, e->BufferLength);
634 			}
635 		}
636 		flagt = e->Qsleeping;
637 		flags |= flagt;
638 		if (flagt & F_CMDQ0_ENABLE)
639 			sge->intr_cnt.rx_cmdq0++;
640 		if (flagt & F_CMDQ1_ENABLE)
641 			sge->intr_cnt.rx_cmdq1++;
642 		if (flagt & F_FL0_ENABLE)
643 			sge->intr_cnt.rx_flq0++;
644 		if (flagt & F_FL1_ENABLE)
645 			sge->intr_cnt.rx_flq1++;
646 #ifdef TX_THREAD_RECLAIM
647 		spin_lock(&sge->cmdQ[0].cq_qlock);
648 		sge->cmdQ[0].cq_complete += e->Cmdq0CreditReturn;
649 		spin_unlock(&sge->cmdQ[0].cq_qlock);
650 		spin_lock(&sge->cmdQ[1].cq_qlock);
651 		sge->cmdQ[1].cq_complete += e->Cmdq1CreditReturn;
652 		if ((adapter->ch_blked) &&
653 		    (sge->cmdQ[0].cq_complete +
654 		    sge->cmdQ[1].cq_complete) > 16) {
655 			adapter->ch_blked = 0;
656 			ch_gld_ok(adapter);
657 		}
658 		spin_unlock(&sge->cmdQ[1].cq_qlock);
659 #else
660 		credits_pend[0] += e->Cmdq0CreditReturn;
661 		credits_pend[1] += e->Cmdq1CreditReturn;
662 #ifdef CONFIG_SMP
663 		if (unlikely(credits_pend[0] > SGE_BATCH_THRESH)) {
664 			free_cmdQ_buffers(sge, &sge->cmdQ[0], credits_pend[0]);
665 			credits_pend[0] = 0;
666 		}
667 		if (unlikely(credits_pend[1] > SGE_BATCH_THRESH)) {
668 			free_cmdQ_buffers(sge, &sge->cmdQ[1], credits_pend[1]);
669 			credits_pend[1] = 0;
670 		}
671 #endif
672 #endif
673 #ifdef HOST_PAUSE
674 		t1_sge_check_pause(sge, &sge->freelQ[e->FreelistQid]);
675 #endif
676 		e++;
677 		if (unlikely(++cidx == entries_n)) {
678 			cidx = 0;
679 			genbit ^= 1;
680 			e = q;
681 		}
682 
683 		/* pull physical memory of response queue entry into cache */
684 		(void) ddi_dma_sync(dh, (off_t)((caddr_t)e - (caddr_t)q),
685 		    sizeof (*e), DDI_DMA_SYNC_FORKERNEL);
686 
687 		ret = 1;
688 	}
689 
690 #ifndef TX_THREAD_RECLAIM
691 	if (credits_pend[0])
692 		free_cmdQ_buffers(sge, &sge->cmdQ[0], credits_pend[0]);
693 	if (credits_pend[1])
694 		free_cmdQ_buffers(sge, &sge->cmdQ[1], credits_pend[1]);
695 #endif
696 	if (flags & F_CMDQ0_ENABLE) {
697 		struct cmdQ *cmdQ = &sge->cmdQ[0];
698 		atomic_set(&cmdQ->cq_asleep, 1);
699 		if (atomic_read(cmdQ->cq_pio_pidx) != cmdQ->cq_pidx) {
700 			doorbell_pio(sge, F_CMDQ0_ENABLE);
701 			atomic_set(&cmdQ->cq_pio_pidx, cmdQ->cq_pidx);
702 		}
703 	}
704 
705 	/* the SGE told us one of the free lists is empty */
706 	if (unlikely(flags & (F_FL0_ENABLE | F_FL1_ENABLE)))
707 		freelQs_empty(sge);
708 
709 #ifdef CONFIG_CHELSIO_T1_OFFLOAD
710 	if (adapter->ch_tx_overflow_mutex)
711 		mutex_enter(adapter->ch_tx_overflow_mutex);
712 	if (adapter->ch_blked &&
713 	    (sge->cmdQ[0].cq_credits > (sge->cmdQ[0].cq_entries_n>>2)) &&
714 	    (sge->cmdQ[1].cq_credits > (sge->cmdQ[1].cq_entries_n>>2))) {
715 		adapter->ch_blked = 0;
716 		if (adapter->ch_tx_overflow_cv)
717 			cv_broadcast(adapter->ch_tx_overflow_cv);
718 		ch_gld_ok(adapter);
719 	}
720 	if (adapter->ch_tx_overflow_mutex)
721 		mutex_exit(adapter->ch_tx_overflow_mutex);
722 #else
723 #ifndef TX_THREAD_RECLAIM
724 	if (adapter->ch_blked &&
725 	    (sge->cmdQ[0].cq_credits > (sge->cmdQ[0].cq_entries_n>>1)) &&
726 	    (sge->cmdQ[1].cq_credits > (sge->cmdQ[1].cq_entries_n>>1))) {
727 		adapter->ch_blked = 0;
728 		ch_gld_ok(adapter);
729 	}
730 #endif
731 #endif	/* CONFIG_CHELSIO_T1_OFFLOAD */
732 
733 	Q->rq_genbit = genbit;
734 	Q->rq_cidx = cidx;
735 	Q->rq_credits = credits;
736 
737 	t1_write_reg_4(adapter, A_SG_SLEEPING, cidx);
738 
739 check_slow_ints:
740 	/* handle non-data interrupts */
741 	if (unlikely(!ret))
742 		ret = t1_slow_intr_handler(adapter);
743 
744 	return (ret);
745 }
746 
747 /*
748  * allocate a mblk with DMA mapped mblk.
749  * When checksum offload is enabled, we start the DMA at a 2 byte offset so
750  * the IP header will be aligned. We do this for sparc only.
751  */
752 static uint64_t
753 os_freelist_buffer_alloc(ch_t *sa, int sz, mblk_t **mb, ulong_t *dh)
754 {
755 	ch_esb_t *ch_get_small_rbuf(ch_t *sa);
756 	ch_esb_t *ch_get_big_rbuf(ch_t *sa);
757 	ch_esb_t *rbp;
758 	uint32_t rxoff = sa->sge->rx_offset;
759 
760 	if (sz == SGE_SM_BUF_SZ(sa)) {
761 		/* get pre-mapped buffer */
762 		if ((rbp = ch_get_small_rbuf(sa)) == NULL) {
763 			sa->norcvbuf++;
764 			return ((uint64_t)0);
765 		}
766 
767 		*mb = desballoc((unsigned char *)rbp->cs_buf + rxoff,
768 		    SGE_SM_BUF_SZ(sa)-rxoff, BPRI_MED, &rbp->cs_frtn);
769 		if (*mb == NULL) {
770 			mutex_enter(&sa->ch_small_esbl);
771 			rbp->cs_next = sa->ch_small_esb_free;
772 			sa->ch_small_esb_free = rbp;
773 			mutex_exit(&sa->ch_small_esbl);
774 			return ((uint64_t)0);
775 		}
776 		*dh = rbp->cs_dh;
777 
778 		return (rbp->cs_pa + rxoff);
779 	} else {
780 		/* get pre-mapped buffer */
781 		if ((rbp = ch_get_big_rbuf(sa)) == NULL) {
782 			sa->norcvbuf++;
783 			return ((uint64_t)0);
784 		}
785 
786 		*mb = desballoc((unsigned char *)rbp->cs_buf + rxoff,
787 		    SGE_BG_BUF_SZ(sa)-rxoff, BPRI_MED, &rbp->cs_frtn);
788 		if (*mb == NULL) {
789 			mutex_enter(&sa->ch_big_esbl);
790 			rbp->cs_next = sa->ch_big_esb_free;
791 			sa->ch_big_esb_free = rbp;
792 			mutex_exit(&sa->ch_big_esbl);
793 			return ((uint64_t)0);
794 		}
795 		*dh = rbp->cs_dh;
796 
797 		return (rbp->cs_pa + rxoff);
798 	}
799 }
800 
801 static inline unsigned int
802 t1_sge_rx(pesge *sge, struct freelQ *Q, unsigned int len, unsigned int offload)
803 {
804 	mblk_t *skb;
805 	peobj *adapter = sge->obj;
806 	struct freelQ_ce *cq = Q->fq_centries;
807 	struct freelQ_ce *ce = &cq[Q->fq_cidx];
808 	ddi_dma_handle_t dh = (ddi_dma_handle_t)ce->fe_dh;
809 	uint32_t cidx = Q->fq_cidx;
810 	uint32_t entries_n = Q->fq_entries_n;
811 	uint32_t sz = Q->fq_rx_buffer_size;
812 	uint32_t useit = 1;
813 	uint32_t rxoff = sge->rx_offset;
814 #ifdef CONFIG_CHELSIO_T1_OFFLOAD
815 	uint32_t rv;
816 #endif
817 
818 	if (Q->fq_id)
819 		sge->intr_cnt.rx_flq1_cnt++;
820 	else
821 		sge->intr_cnt.rx_flq0_cnt++;
822 	/*
823 	 * If pkt size falls below threshold, then we'll copy data to
824 	 * an blk and reuse mblk.
825 	 *
826 	 * NOTE that rxoff is 2 for T1 adapters. We align the the start
827 	 * of the DMA buffer begin at rxoff offset for T1 cards instead of
828 	 * at the beginning of the buffer, thus the length of the received
829 	 * data does not include this offset. We therefore always add
830 	 * SGE_RX_OFFSET to the allocb size so we have space to provide the
831 	 * offset for the copied data.
832 	 */
833 #ifdef HOST_PAUSE
834 	/*
835 	 * If we have Host pause compiled in, then we look at the
836 	 * free list, if the pause is on and we're not in offload
837 	 * mode then we drop packets, this is designed to avoid
838 	 * overwhelming the machine. If the machine is powerfull enough
839 	 * this will not happen. The 'rx_pkt_drops' will show when
840 	 * packets are being dropped and how much.
841 	 */
842 	if ((offload == 0) && adapter->pause_on) {
843 		freelQ_e *e;
844 		/* Ditch the packet and reuse original buffer */
845 		e = &Q->fq_entries[cidx];
846 		e->GenerationBit  ^= 1;
847 		e->GenerationBit2 ^= 1;
848 		sge->intr_cnt.rx_pkt_drops++;
849 		goto rx_entry_consumed;
850 	} else if (((adapter->pause_on ||
851 	    (len <= SGE_RX_COPY_THRESHOLD)) &&
852 	    (skb = allocb(len + SGE_RX_OFFSET, BPRI_HI))))
853 #else
854 	if ((len <= SGE_RX_COPY_THRESHOLD) &&
855 	    (skb = allocb(len + SGE_RX_OFFSET, BPRI_HI)))
856 #endif
857 	{
858 		freelQ_e *e;
859 		char *src = (char *)((mblk_t *)ce->fe_mp)->b_rptr;
860 
861 		/*
862 		 * pull physical memory of pkt data into cache
863 		 * Note that len does not include offset for T1.
864 		 */
865 		(void) ddi_dma_sync(dh, (off_t)(rxoff), len,
866 		    DDI_DMA_SYNC_FORKERNEL);
867 
868 		if (offload == 0) {
869 			/*
870 			 * create 2 byte offset so IP header aligned on
871 			 * 4 byte boundry
872 			 */
873 			skb_reserve(skb, SGE_RX_OFFSET);
874 			/*
875 			 * if hardware inserted 2 byte offset then need to
876 			 * start copying with extra offset
877 			 */
878 			src += sge->rx_pkt_pad;
879 		}
880 		memcpy(skb->b_rptr, src, len);
881 		useit = 0;	/* mblk copy, don't inc esballoc in use cnt */
882 
883 		/* so we can reuse original buffer */
884 		e = &Q->fq_entries[cidx];
885 		e->GenerationBit  ^= 1;
886 		e->GenerationBit2 ^= 1;
887 		sge->intr_cnt.rx_pkt_copied++;
888 	} else {
889 		/* consume buffer off the ring */
890 		skb = ce->fe_mp;
891 		ce->fe_mp = NULL;
892 
893 		/*
894 		 * if not offload (tunneled pkt), & hardward padded, then
895 		 * adjust start of pkt to point to start of data i.e.
896 		 * skip pad (2 bytes).
897 		 */
898 		if (!offload && sge->rx_pkt_pad)
899 			__skb_pull(skb, SGE_RX_OFFSET);
900 
901 		/*
902 		 * pull physical memory of pkt data into cache
903 		 * Note that len does not include offset for T1.
904 		 */
905 		(void) ddi_dma_sync(dh, (off_t)(rxoff), len,
906 		    DDI_DMA_SYNC_FORKERNEL);
907 	}
908 
909 	/* set length of data in skb */
910 	skb_put(skb, len);
911 
912 #ifdef CONFIG_CHELSIO_T1_OFFLOAD
913 	if (likely(offload)) {
914 		if (likely(toe_running(adapter))) {
915 			/* sends pkt upstream to toe layer */
916 			if (useit) {
917 				uint_t index;
918 				if (sz == SGE_SM_BUF_SZ(adapter))
919 					index = adapter->ch_sm_index;
920 				else
921 					index = adapter->ch_big_index;
922 				atomic_add(1, &buffers_in_use[index]);
923 			}
924 			if (adapter->toe_rcv)
925 				adapter->toe_rcv(adapter->ch_toeinst, skb);
926 			else
927 				freemsg(skb);
928 		} else {
929 			cmn_err(CE_WARN,
930 			    "%s: unexpected offloaded packet, cmd %u\n",
931 			    adapter->ch_name, *skb->b_rptr);
932 
933 			/* discard packet */
934 			freemsg(skb);
935 		}
936 	}
937 #else
938 	if (unlikely(offload)) {
939 		cmn_err(CE_WARN,
940 		    "%s: unexpected offloaded packet, cmd %u\n",
941 		    adapter->ch_name, *skb->b_rptr);
942 
943 		/* discard paket */
944 		freemsg(skb);
945 	}
946 #endif
947 	else {
948 		struct cpl_rx_pkt *p = (struct cpl_rx_pkt *)skb->b_rptr;
949 		int flg = 0;
950 		uint32_t cksum;
951 
952 		/* adjust beginning of data to skip CPL header */
953 		skb_pull(skb, SZ_CPL_RX_PKT);
954 
955 		/* extract checksum from CPL header here */
956 
957 		/*
958 		 * bump count of mlbks in used by protocol stack(s)
959 		 */
960 		if (useit) {
961 			if (sz == SGE_SM_BUF_SZ(adapter)) {
962 				atomic_add(1,
963 				    &buffers_in_use[adapter->ch_sm_index]);
964 			} else {
965 				atomic_add(1,
966 				    &buffers_in_use[adapter->ch_big_index]);
967 			}
968 		}
969 
970 #ifdef CONFIG_CHELSIO_T1_OFFLOAD
971 		/*
972 		 * let the TOE layer have a crack at the packet first.
973 		 */
974 		if (adapter->toe_tunnel) {
975 			rv = adapter->toe_tunnel(adapter->ch_toeinst, skb);
976 			/*
977 			 * The TOE may have consumed the packet.
978 			 */
979 			if (rv)
980 				goto rx_entry_consumed;
981 		}
982 #endif	/* CONFIG_CHELSIO_T1_OFFLOAD */
983 
984 		cksum = p->csum;
985 
986 		/*
987 		 * NOTE: 14+9 = size of MAC + offset to IP protocol field
988 		 */
989 		if (adapter->ch_config.cksum_enabled &&
990 		    (ntohs(((struct ether_header *)skb->b_rptr)->ether_type) ==
991 		    ETHERTYPE_IP) &&
992 		    ((skb->b_rptr[14+9] == IPPROTO_TCP) ||
993 		    (skb->b_rptr[14+9] == IPPROTO_UDP))) {
994 			flg = 1;
995 		}
996 
997 		ch_send_up(adapter, skb, cksum, flg);
998 	}
999 
1000 rx_entry_consumed:
1001 
1002 	if (++cidx == entries_n)
1003 		cidx = 0;
1004 
1005 	Q->fq_cidx = cidx;
1006 
1007 	if (unlikely(--Q->fq_credits < (entries_n>>2)))
1008 		/* allocate new buffers on the free list */
1009 		alloc_freelQ_buffers(sge, Q);
1010 	return (1);
1011 }
1012 
1013 #ifdef HOST_PAUSE
1014 static void
1015 t1_sge_check_pause(pesge *sge, struct freelQ *Q)
1016 {
1017 	peobj *adapter = sge->obj;
1018 
1019 	/*
1020 	 * If the number of available credits shrinks below
1021 	 * the Pause on threshold then enable the pause and
1022 	 * try and allocate more buffers.
1023 	 * On the next pass, if there's more credits returned
1024 	 * then check that you've went above the pause
1025 	 * threshold and then disable the pause.
1026 	 */
1027 	if (Q->fq_credits < Q->fq_pause_on_thresh) {
1028 		if (do_host_pause) {
1029 			sge->intr_cnt.rx_pause_on++;
1030 			adapter->txxg_cfg1 |=
1031 			    SUNI1x10GEXP_BITMSK_TXXG_HOSTPAUSE;
1032 			(void) t1_tpi_write(adapter,
1033 			    SUNI1x10GEXP_REG_TXXG_CONFIG_1 << 2,
1034 			    adapter->txxg_cfg1);
1035 			adapter->pause_on = 1;
1036 			adapter->pause_time = gethrtime();
1037 		}
1038 		alloc_freelQ_buffers(sge, Q);
1039 	} else if ((adapter->pause_on) &&
1040 	    (Q->fq_credits > Q->fq_pause_off_thresh)) {
1041 		hrtime_t time;
1042 		sge->intr_cnt.rx_pause_off++;
1043 		adapter->txxg_cfg1 &= ~SUNI1x10GEXP_BITMSK_TXXG_HOSTPAUSE;
1044 		(void) t1_tpi_write(adapter,
1045 		    SUNI1x10GEXP_REG_TXXG_CONFIG_1 << 2,
1046 		    adapter->txxg_cfg1);
1047 		adapter->pause_on = 0;
1048 		time = (gethrtime() - adapter->pause_time)/1000;
1049 		sge->intr_cnt.rx_pause_ms += time;
1050 		if (time > sge->intr_cnt.rx_pause_spike)
1051 			sge->intr_cnt.rx_pause_spike = (uint32_t)time;
1052 	}
1053 	sge->intr_cnt.rx_fl_credits = Q->fq_credits;
1054 }
1055 #endif	/* HOST_PAUSE */
1056 
1057 static void
1058 alloc_freelQ_buffers(pesge *sge, struct freelQ *Q)
1059 {
1060 	uint32_t pidx = Q->fq_pidx;
1061 	struct freelQ_ce *ce = &Q->fq_centries[pidx];
1062 	freelQ_e *fq = Q->fq_entries;		/* base of freelist Q */
1063 	freelQ_e *e = &Q->fq_entries[pidx];
1064 	uint32_t sz = Q->fq_rx_buffer_size;
1065 	uint32_t rxoff = sge->rx_offset;
1066 	uint32_t credits = Q->fq_credits;
1067 	uint32_t entries_n = Q->fq_entries_n;
1068 	uint32_t genbit = Q->fq_genbit;
1069 	ddi_dma_handle_t th = (ddi_dma_handle_t)Q->fq_dh;
1070 	ulong_t dh;
1071 	uint64_t mapping;
1072 	off_t offset = (off_t)((caddr_t)e - (caddr_t)fq);
1073 	size_t len = 0;
1074 
1075 	while (credits < entries_n) {
1076 		if (e->GenerationBit != genbit) {
1077 			mblk_t *skb;
1078 
1079 			mapping = os_freelist_buffer_alloc(sge->obj, sz,
1080 			    &skb, &dh);
1081 			if (mapping == 0) {
1082 				sge->intr_cnt.rx_flbuf_fails++;
1083 				break;
1084 			}
1085 			sge->intr_cnt.rx_flbuf_allocs++;
1086 
1087 			ce->fe_mp = skb;
1088 			ce->fe_dh = dh;
1089 
1090 			/*
1091 			 * Note that for T1, we've started the beginning of
1092 			 * of the buffer by an offset of 2 bytes. We thus
1093 			 * decrement the length to account for this.
1094 			 */
1095 			e->AddrLow = (u32)mapping;
1096 			e->AddrHigh = (u64)mapping >> 32;
1097 			e->BufferLength = sz - rxoff;
1098 			wmb();
1099 			e->GenerationBit = e->GenerationBit2 = genbit;
1100 		}
1101 
1102 		len += sizeof (*e);
1103 
1104 		ce++;
1105 		e++;
1106 		credits++;
1107 		if (++pidx == entries_n) {
1108 			/*
1109 			 * sync freelist entries to physical memory up to
1110 			 * end of the table.
1111 			 */
1112 			(void) ddi_dma_sync(th, offset, len,
1113 			    DDI_DMA_SYNC_FORDEV);
1114 			offset = 0;
1115 			len = 0;
1116 
1117 			pidx = 0;
1118 			genbit ^= 1;
1119 			ce = Q->fq_centries;
1120 			e = Q->fq_entries;
1121 		}
1122 	}
1123 
1124 	/* sync freelist entries that have been modified. */
1125 	if (len)
1126 		(void) ddi_dma_sync(th, offset, len, DDI_DMA_SYNC_FORDEV);
1127 
1128 	Q->fq_genbit = genbit;
1129 	Q->fq_pidx = pidx;
1130 	Q->fq_credits = credits;
1131 }
1132 
1133 static void
1134 freelQs_empty(pesge *sge)
1135 {
1136 	u32 irq_reg = t1_read_reg_4(sge->obj, A_SG_INT_ENABLE);
1137 	u32 irqholdoff_reg;
1138 
1139 	alloc_freelQ_buffers(sge, &sge->freelQ[0]);
1140 	alloc_freelQ_buffers(sge, &sge->freelQ[1]);
1141 
1142 	if ((sge->freelQ[0].fq_credits > sge->freelQ[0].fq_entries_n >> 2) &&
1143 	    (sge->freelQ[1].fq_credits > sge->freelQ[1].fq_entries_n >> 2)) {
1144 		irq_reg |= F_FL_EXHAUSTED;
1145 		irqholdoff_reg = sge->intrtimer[sge->currIndex];
1146 	} else {
1147 		/* Clear the F_FL_EXHAUSTED interrupts for now */
1148 		irq_reg &= ~F_FL_EXHAUSTED;
1149 		irqholdoff_reg = sge->intrtimer_nres;
1150 	}
1151 	t1_write_reg_4(sge->obj, A_SG_INTRTIMER, irqholdoff_reg);
1152 	t1_write_reg_4(sge->obj, A_SG_INT_ENABLE, irq_reg);
1153 
1154 	/* We reenable the Qs to force an Freelist GTS interrupt later */
1155 	doorbell_pio(sge, F_FL0_ENABLE | F_FL1_ENABLE);
1156 }
1157 
1158 /*
1159  * Frees 'credits_pend' TX buffers and returns the credits to Q->credits.
1160  * Free xmit buffers
1161  */
1162 static void
1163 free_cmdQ_buffers(pesge *sge, struct cmdQ *Q, unsigned int credits_pend)
1164 {
1165 	mblk_t *skb;
1166 	struct cmdQ_ce *ce;
1167 	struct cmdQ_ce *cq = Q->cq_centries;
1168 	uint32_t entries_n = Q->cq_entries_n;
1169 	uint32_t cidx = Q->cq_cidx;
1170 	uint32_t i = credits_pend;
1171 #ifdef CONFIG_CHELSIO_T1_OFFLOAD
1172 	ch_t *chp = sge->obj;
1173 #endif
1174 	ce = &cq[cidx];
1175 
1176 	while (i--) {
1177 #ifdef CONFIG_CHELSIO_T1_OFFLOAD
1178 		/* if flag set, then toe buffer */
1179 		switch (ce->ce_flg & 0x7) {
1180 		case DH_DMA:
1181 			if (ce->ce_dh) {
1182 				ch_unbind_dma_handle(sge->obj, ce->ce_dh);
1183 				ce->ce_dh = NULL;	/* may not be needed */
1184 			}
1185 			skb = ce->ce_mp;
1186 			if (skb && ((ce->ce_flg & CH_ARP) == 0)) {
1187 				freemsg(skb);
1188 			}
1189 			ce->ce_mp = NULL;
1190 			break;
1191 
1192 #if defined(__sparc)
1193 		case DH_DVMA:
1194 			if (ce->ce_dh) {
1195 				ch_unbind_dvma_handle(sge->obj, ce->ce_dh);
1196 				ce->ce_dh = NULL;	/* may not be needed */
1197 			}
1198 			skb = ce->ce_mp;
1199 			if (skb && ((ce->ce_flg & CH_ARP) == 0)) {
1200 				freemsg(skb);
1201 			}
1202 			ce->ce_mp = NULL;
1203 			break;
1204 #endif	/* __sparc */
1205 
1206 		case DH_TOE:
1207 			chp->toe_free(chp->ch_toeinst, (tbuf_t *)(ce->ce_mp));
1208 			ce->ce_mp = NULL;
1209 			break;
1210 		}
1211 #else	/* CONFIG_CHELSIO_T1_OFFLOAD */
1212 		if (ce->ce_dh) {
1213 			if ((ce->ce_flg & 7) == DH_DMA) {
1214 				ch_unbind_dma_handle(sge->obj, ce->ce_dh);
1215 			}
1216 #if defined(__sparc)
1217 			else {
1218 				ch_unbind_dvma_handle(sge->obj, ce->ce_dh);
1219 			}
1220 #endif	/* __sparc */
1221 			ce->ce_dh = NULL; /* may not be needed */
1222 		}
1223 
1224 		skb = ce->ce_mp;
1225 		if (skb && ((ce->ce_flg & CH_ARP) == 0)) {
1226 			freemsg(skb);
1227 		}
1228 		ce->ce_mp = NULL;
1229 #endif	/* !CONFIG_CHELSIO_T1_OFFLOAD */
1230 
1231 		ce++;
1232 		if (++cidx == entries_n) {
1233 			cidx = 0;
1234 			ce = cq;
1235 		}
1236 	}
1237 
1238 	Q->cq_cidx = cidx;
1239 	atomic_add(credits_pend, &Q->cq_credits);
1240 }
1241 
1242 struct sge_intr_counts *
1243 sge_get_stat(pesge *sge)
1244 {
1245 	return (&sge->intr_cnt);
1246 }
1247 
1248 /*
1249  * Allocates both RX and TX resources and configures the SGE. However,
1250  * the hardware is not enabled yet.
1251  *
1252  * rx_pkt_pad is set, if the hardware supports aligning non-offload traffic.
1253  * jumbo_fl is set to the index of the freelist containing the jumbo buffers.
1254  */
1255 int
1256 t1_sge_configure(pesge *sge, struct sge_params *p)
1257 {
1258 	sge->rx_pkt_pad = t1_is_T1B(sge->obj) ? 0 : SGE_RX_OFFSET;
1259 	sge->jumbo_fl = t1_is_T1B(sge->obj) ? 1 : 0;
1260 	/* if we're a T2 card, then we have hardware offset support */
1261 	sge->rx_offset = t1_is_T1B(sge->obj) ? SGE_RX_OFFSET: 0;
1262 
1263 	if (alloc_rx_resources(sge, p))
1264 		return (-ENOMEM);
1265 	if (alloc_tx_resources(sge, p)) {
1266 		free_rx_resources(sge);
1267 		return (-ENOMEM);
1268 	}
1269 	configure_sge(sge, p);
1270 
1271 	/*
1272 	 * Now that we have sized the free lists calculate the payload
1273 	 * capacity of the large buffers.  Other parts of the driver use
1274 	 * this to set the max offload coalescing size so that RX packets
1275 	 * do not overflow our large buffers.
1276 	 */
1277 	p->large_buf_capacity = jumbo_payload_capacity(sge);
1278 	return (0);
1279 }
1280 
1281 /*
1282  * Allocates basic RX resources, consisting of memory mapped freelist Qs and a
1283  * response Q.
1284  */
1285 static int
1286 alloc_rx_resources(pesge *sge, struct sge_params *p)
1287 {
1288 	unsigned int size, i;
1289 
1290 	for (i = 0; i < SGE_FREELQ_N; i++) {
1291 		struct freelQ *Q = &sge->freelQ[i];
1292 
1293 		Q->fq_id = i;
1294 		Q->fq_genbit = 1;
1295 		Q->fq_entries_n = p->freelQ_size[i];
1296 #ifdef HOST_PAUSE
1297 		Q->fq_pause_on_thresh = flq_pause_window;
1298 		Q->fq_pause_off_thresh = Q->fq_entries_n >> 1;
1299 #endif
1300 		size = sizeof (freelQ_e) * Q->fq_entries_n;
1301 
1302 		Q->fq_entries = pe_os_malloc_contig_wait_zero(sge->obj,
1303 		    size, &Q->fq_pa, &Q->fq_dh, &Q->fq_ah, DMA_OUT);
1304 
1305 
1306 		if (!Q->fq_entries)
1307 			goto err_no_mem;
1308 		memset(Q->fq_entries, 0, size);
1309 		size = sizeof (struct freelQ_ce) * Q->fq_entries_n;
1310 		Q->fq_centries = t1_os_malloc_wait_zero(size);
1311 		if (!Q->fq_centries)
1312 			goto err_no_mem;
1313 		memset(Q->fq_centries, 0, size);
1314 	}
1315 
1316 	/*
1317 	 * Calculate the buffer sizes for the two free lists.  FL0 accommodates
1318 	 * regular sized Ethernet frames, FL1 is sized not to exceed 16K,
1319 	 * including all the sk_buff overhead.
1320 	 * For T1C FL0 and FL1 are reversed.
1321 	 */
1322 #ifdef NOTYET
1323 	sge->freelQ[1 ^ sge->jumbo_fl].fq_rx_buffer_size = SGE_RX_SM_BUF_SIZE +
1324 	    sizeof (struct cpl_rx_data) +
1325 	    SGE_RX_OFFSET - sge->rx_pkt_pad;
1326 #else
1327 	sge->freelQ[1 ^ sge->jumbo_fl].fq_rx_buffer_size =
1328 	    sge->obj->ch_sm_buf_sz;
1329 	if (is_T2(sge->obj))
1330 		sge->intr_cnt.rx_flq1_sz = sge->obj->ch_sm_buf_sz;
1331 	else
1332 		sge->intr_cnt.rx_flq0_sz = sge->obj->ch_sm_buf_sz;
1333 #endif
1334 #ifdef NOTYET
1335 	sge->freelQ[sge->jumbo_fl].fq_rx_buffer_size = (16 * 1024) -
1336 	    SKB_DATA_ALIGN(sizeof (struct skb_shared_info));
1337 #else
1338 	sge->freelQ[sge->jumbo_fl].fq_rx_buffer_size = sge->obj->ch_bg_buf_sz;
1339 	if (is_T2(sge->obj))
1340 		sge->intr_cnt.rx_flq0_sz = sge->obj->ch_bg_buf_sz;
1341 	else
1342 		sge->intr_cnt.rx_flq1_sz = sge->obj->ch_bg_buf_sz;
1343 #endif
1344 
1345 	sge->respQ.rq_genbit = 1;
1346 	sge->respQ.rq_entries_n = sge_respq_cnt;
1347 	sge->respQ.rq_credits = sge_respq_cnt;
1348 	sge->respQ.rq_credits_thresh = sge_respq_cnt - (sge_respq_cnt >> 2);
1349 	size = sizeof (respQ_e) * sge->respQ.rq_entries_n;
1350 
1351 	sge->respQ.rq_entries = pe_os_malloc_contig_wait_zero(sge->obj,
1352 	    size, &(sge->respQ.rq_pa), &(sge->respQ.rq_dh),
1353 	    &(sge->respQ.rq_ah), 0);
1354 
1355 	if (!sge->respQ.rq_entries)
1356 		goto err_no_mem;
1357 	memset(sge->respQ.rq_entries, 0, size);
1358 	return (0);
1359 
1360 err_no_mem:
1361 	free_rx_resources(sge);
1362 	return (1);
1363 }
1364 
1365 /*
1366  * Allocates basic TX resources, consisting of memory mapped command Qs.
1367  */
1368 static int
1369 alloc_tx_resources(pesge *sge, struct sge_params *p)
1370 {
1371 	unsigned int size, i;
1372 
1373 	for (i = 0; i < SGE_CMDQ_N; i++) {
1374 		struct cmdQ *Q = &sge->cmdQ[i];
1375 
1376 		Q->cq_genbit = 1;
1377 		Q->cq_entries_n = p->cmdQ_size[i];
1378 		atomic_set(&Q->cq_credits, Q->cq_entries_n);
1379 		atomic_set(&Q->cq_asleep, 1);
1380 
1381 		mutex_init(&Q->cq_qlock, NULL, MUTEX_DRIVER,
1382 		    sge->obj->ch_icookp);
1383 
1384 		size = sizeof (cmdQ_e) * Q->cq_entries_n;
1385 		Q->cq_entries = pe_os_malloc_contig_wait_zero(sge->obj,
1386 		    size, &Q->cq_pa, &Q->cq_dh, &Q->cq_ah, DMA_OUT);
1387 
1388 		if (!Q->cq_entries)
1389 			goto err_no_mem;
1390 		memset(Q->cq_entries, 0, size);
1391 		size = sizeof (struct cmdQ_ce) * Q->cq_entries_n;
1392 		Q->cq_centries = t1_os_malloc_wait_zero(size);
1393 		if (!Q->cq_centries)
1394 			goto err_no_mem;
1395 		memset(Q->cq_centries, 0, size);
1396 
1397 		/* allocate pre-mapped dma headers */
1398 		pe_dma_handle_init(sge->obj, Q->cq_entries_n);
1399 	}
1400 
1401 	return (0);
1402 
1403 err_no_mem:
1404 	free_tx_resources(sge);
1405 	return (1);
1406 }
1407 
1408 /*
1409  * Sets the interrupt latency timer when the adaptive Rx coalescing
1410  * is turned off. Do nothing when it is turned on again.
1411  *
1412  * This routine relies on the fact that the caller has already set
1413  * the adaptive policy in adapter->sge_params before calling it.
1414  */
1415 int
1416 t1_sge_set_coalesce_params(pesge *sge, struct sge_params *p)
1417 {
1418 	if (!p->coalesce_enable) {
1419 		u32 newTimer = p->rx_coalesce_usecs *
1420 		    (board_info(sge->obj)->clock_core / 1000000);
1421 
1422 		t1_write_reg_4(sge->obj, A_SG_INTRTIMER, newTimer);
1423 	}
1424 	return (0);
1425 }
1426 
1427 /*
1428  * Programs the various SGE registers. However, the engine is not yet enabled,
1429  * but sge->sge_control is setup and ready to go.
1430  */
1431 static void
1432 configure_sge(pesge *sge, struct sge_params *p)
1433 {
1434 	ch_t *ap = sge->obj;
1435 	int i;
1436 
1437 	t1_write_reg_4(ap, A_SG_CONTROL, 0);
1438 
1439 	setup_ring_params(ap, sge->cmdQ[0].cq_pa, sge->cmdQ[0].cq_entries_n,
1440 	    A_SG_CMD0BASELWR, A_SG_CMD0BASEUPR, A_SG_CMD0SIZE);
1441 	setup_ring_params(ap, sge->cmdQ[1].cq_pa, sge->cmdQ[1].cq_entries_n,
1442 	    A_SG_CMD1BASELWR, A_SG_CMD1BASEUPR, A_SG_CMD1SIZE);
1443 	setup_ring_params(ap, sge->freelQ[0].fq_pa,
1444 	    sge->freelQ[0].fq_entries_n, A_SG_FL0BASELWR,
1445 	    A_SG_FL0BASEUPR, A_SG_FL0SIZE);
1446 	setup_ring_params(ap, sge->freelQ[1].fq_pa,
1447 	    sge->freelQ[1].fq_entries_n, A_SG_FL1BASELWR,
1448 	    A_SG_FL1BASEUPR, A_SG_FL1SIZE);
1449 
1450 	/* The threshold comparison uses <. */
1451 	t1_write_reg_4(ap, A_SG_FLTHRESHOLD, SGE_RX_SM_BUF_SIZE(ap) -
1452 	    SZ_CPL_RX_PKT - sge->rx_pkt_pad - sge->rx_offset + 1);
1453 	setup_ring_params(ap, sge->respQ.rq_pa, sge->respQ.rq_entries_n,
1454 	    A_SG_RSPBASELWR, A_SG_RSPBASEUPR, A_SG_RSPSIZE);
1455 	t1_write_reg_4(ap, A_SG_RSPQUEUECREDIT, (u32)sge->respQ.rq_entries_n);
1456 	sge->sge_control = F_CMDQ0_ENABLE | F_CMDQ1_ENABLE | F_FL0_ENABLE |
1457 	    F_FL1_ENABLE | F_CPL_ENABLE | F_RESPONSE_QUEUE_ENABLE |
1458 	    V_CMDQ_PRIORITY(2) | F_DISABLE_CMDQ1_GTS | F_ISCSI_COALESCE |
1459 #if 1
1460 		/*
1461 		 * if the the following bit is not set, then we'll get an
1462 		 * interrupt everytime command Q 0 goes empty. Since we're
1463 		 * always ringing the doorbell, we can turn it on.
1464 		 */
1465 	    F_DISABLE_CMDQ0_GTS |
1466 #endif
1467 	    V_RX_PKT_OFFSET(sge->rx_pkt_pad);
1468 
1469 #if BYTE_ORDER == BIG_ENDIAN
1470 	sge->sge_control |= F_ENABLE_BIG_ENDIAN;
1471 #endif
1472 
1473 	/*
1474 	 * Initialize the SGE Interrupt Timer arrray:
1475 	 * intrtimer[0] = (SGE_INTRTIMER0) usec
1476 	 * intrtimer[0<i<10] = (SGE_INTRTIMER0 + 2*i) usec
1477 	 * intrtimer[10] = (SGE_INTRTIMER1) usec
1478 	 *
1479 	 */
1480 	sge->intrtimer[0] = board_info(sge->obj)->clock_core / 1000000;
1481 	for (i = 1; i < SGE_INTR_MAXBUCKETS - 1; ++i) {
1482 		sge->intrtimer[i] = SGE_INTRTIMER0 + (2 * i);
1483 		sge->intrtimer[i] *= sge->intrtimer[0];
1484 	}
1485 	sge->intrtimer[SGE_INTR_MAXBUCKETS - 1] =
1486 	    sge->intrtimer[0] * SGE_INTRTIMER1;
1487 	/* Initialize resource timer */
1488 	sge->intrtimer_nres = (uint32_t)(sge->intrtimer[0] *
1489 	    SGE_INTRTIMER_NRES);
1490 	/* Finally finish initialization of intrtimer[0] */
1491 	sge->intrtimer[0] = (uint32_t)(sge->intrtimer[0] * SGE_INTRTIMER0);
1492 	/* Initialize for a throughput oriented workload */
1493 	sge->currIndex = SGE_INTR_MAXBUCKETS - 1;
1494 
1495 	if (p->coalesce_enable)
1496 		t1_write_reg_4(ap, A_SG_INTRTIMER,
1497 		    sge->intrtimer[sge->currIndex]);
1498 	else
1499 		(void) t1_sge_set_coalesce_params(sge, p);
1500 }
1501 
1502 static inline void
1503 setup_ring_params(ch_t *adapter, u64 addr, u32 size, int base_reg_lo,
1504     int base_reg_hi, int size_reg)
1505 {
1506 	t1_write_reg_4(adapter, base_reg_lo, (u32)addr);
1507 	t1_write_reg_4(adapter, base_reg_hi, addr >> 32);
1508 	t1_write_reg_4(adapter, size_reg, size);
1509 }
1510 
1511 /*
1512  * Frees RX resources.
1513  */
1514 static void
1515 free_rx_resources(pesge *sge)
1516 {
1517 	unsigned int size, i;
1518 
1519 	if (sge->respQ.rq_entries) {
1520 		size = sizeof (respQ_e) * sge->respQ.rq_entries_n;
1521 
1522 		pe_os_free_contig(sge->obj, size, sge->respQ.rq_entries,
1523 		    sge->respQ.rq_pa, sge->respQ.rq_dh, sge->respQ.rq_ah);
1524 	}
1525 
1526 	for (i = 0; i < SGE_FREELQ_N; i++) {
1527 		struct freelQ *Q = &sge->freelQ[i];
1528 
1529 		if (Q->fq_centries) {
1530 			free_freelQ_buffers(sge, Q);
1531 
1532 			t1_os_free(Q->fq_centries,
1533 			    Q->fq_entries_n * sizeof (freelQ_ce_t));
1534 		}
1535 		if (Q->fq_entries) {
1536 			size = sizeof (freelQ_e) * Q->fq_entries_n;
1537 
1538 			/* free the freelist queue */
1539 			pe_os_free_contig(sge->obj, size, Q->fq_entries,
1540 			    Q->fq_pa, Q->fq_dh, Q->fq_ah);
1541 
1542 		}
1543 	}
1544 }
1545 
1546 /*
1547  * Frees all RX buffers on the freelist Q. The caller must make sure that
1548  * the SGE is turned off before calling this function.
1549  */
1550 static void
1551 free_freelQ_buffers(pesge *sge, struct freelQ *Q)
1552 {
1553 	struct freelQ_ce *ce;
1554 	struct freelQ_ce *cq = Q->fq_centries;
1555 	uint32_t credits = Q->fq_credits;
1556 	uint32_t entries_n = Q->fq_entries_n;
1557 	uint32_t cidx = Q->fq_cidx;
1558 	uint32_t i = Q->fq_id;
1559 
1560 	ce = &cq[cidx];
1561 
1562 	credits = entries_n;
1563 	while (credits--) {
1564 		mblk_t *mp;
1565 		if ((mp = ce->fe_mp) != NULL) {
1566 			/* bump in-use count of receive buffers */
1567 			if (i != sge->jumbo_fl) {
1568 				atomic_add(1,
1569 				    &buffers_in_use[sge->obj->ch_sm_index]);
1570 			} else {
1571 				atomic_add(1,
1572 				    &buffers_in_use[sge->obj->ch_big_index]);
1573 			}
1574 
1575 			/*
1576 			 * note. freeb() callback of esb-alloced mblk will
1577 			 * cause receive buffer to be put back on sa free list.
1578 			 */
1579 			freeb(mp);
1580 			ce->fe_mp = NULL;
1581 		}
1582 
1583 		ce++;
1584 		if (++cidx == entries_n) {
1585 			cidx = 0;
1586 			ce = cq;
1587 		}
1588 	}
1589 
1590 	Q->fq_cidx = cidx;
1591 	Q->fq_credits = credits;
1592 }
1593 
1594 /*
1595  * Free TX resources.
1596  *
1597  * Assumes that SGE is stopped and all interrupts are disabled.
1598  */
1599 static void
1600 free_tx_resources(pesge *sge)
1601 {
1602 	unsigned int size;
1603 	uint32_t i;
1604 
1605 	for (i = 0; i < SGE_CMDQ_N; i++) {
1606 		struct cmdQ *Q = &sge->cmdQ[i];
1607 
1608 		if (Q->cq_centries) {
1609 			unsigned int pending = Q->cq_entries_n -
1610 			    atomic_read(Q->cq_credits);
1611 
1612 			mutex_destroy(&Q->cq_qlock);
1613 
1614 			if (pending)
1615 				free_cmdQ_buffers(sge, Q, pending);
1616 
1617 			size = sizeof (struct cmdQ_ce) * Q->cq_entries_n;
1618 			t1_os_free(Q->cq_centries, size);
1619 		}
1620 
1621 		if (Q->cq_entries) {
1622 			size = sizeof (cmdQ_e) * Q->cq_entries_n;
1623 			pe_os_free_contig(sge->obj, size, Q->cq_entries,
1624 			    Q->cq_pa, Q->cq_dh, Q->cq_ah);
1625 		}
1626 	}
1627 }
1628 
1629 /*
1630  * Return the payload capacity of the jumbo free-list buffers.
1631  */
1632 static inline unsigned int jumbo_payload_capacity(pesge *sge)
1633 {
1634 	return (sge->freelQ[sge->jumbo_fl].fq_rx_buffer_size -
1635 	    sizeof (struct cpl_rx_data) - sge->rx_pkt_pad - sge->rx_offset);
1636 }
1637 
1638 /* PR2928 & PR3309 */
1639 void
1640 t1_sge_set_ptimeout(adapter_t *adapter, u32 val)
1641 {
1642 	pesge *sge = adapter->sge;
1643 
1644 	if (is_T2(adapter))
1645 		sge->ptimeout = max(val, 1);
1646 }
1647 
1648 /* PR2928 & PR3309 */
1649 u32
1650 t1_sge_get_ptimeout(adapter_t *adapter)
1651 {
1652 	pesge *sge = adapter->sge;
1653 
1654 	return (is_T2(adapter) ? sge->ptimeout : 0);
1655 }
1656 
1657 void
1658 sge_add_fake_arp(pesge *sge, void *bp)
1659 {
1660 	sge->pskb = bp;
1661 }
1662 
1663 #ifdef SUN_KSTATS
1664 static int
1665 sge_kstat_setup(pesge *sge)
1666 {
1667 	int status;
1668 	p_kstat_t ksp;
1669 	size_t ch_kstat_sz;
1670 	p_ch_kstat_t chkp;
1671 	char kstat_name[32];
1672 	int instance;
1673 	int i;
1674 
1675 	status = -1;
1676 	ch_kstat_sz = sizeof (ch_kstat_t);
1677 	instance = ddi_get_instance(sge->obj->ch_dip);
1678 	if ((ksp = kstat_create(CHNAME "_debug", instance,
1679 	    NULL, "net_debug", KSTAT_TYPE_NAMED,
1680 	    ch_kstat_sz / sizeof (kstat_named_t), 0)) == NULL)
1681 		goto sge_kstat_setup_exit;
1682 	chkp = (p_ch_kstat_t)ksp->ks_data;
1683 	kstat_named_init(&chkp->respQ_empty,		"respQ_empty",
1684 	    KSTAT_DATA_UINT32);
1685 	kstat_named_init(&chkp->respQ_overflow,		"respQ_overflow",
1686 	    KSTAT_DATA_UINT32);
1687 	kstat_named_init(&chkp->freelistQ_empty,	"freelistQ_empty",
1688 	    KSTAT_DATA_UINT32);
1689 	kstat_named_init(&chkp->pkt_too_big,		"pkt_too_big",
1690 	    KSTAT_DATA_UINT32);
1691 	kstat_named_init(&chkp->pkt_mismatch,		"pkt_mismatch",
1692 	    KSTAT_DATA_UINT32);
1693 	kstat_named_init(&chkp->cmdQ_full[0],		"cmdQ_full[0]",
1694 	    KSTAT_DATA_UINT32);
1695 	kstat_named_init(&chkp->cmdQ_full[1],		"cmdQ_full[1]",
1696 	    KSTAT_DATA_UINT32);
1697 	kstat_named_init(&chkp->tx_reclaims[0],		"tx_reclaims[0]",
1698 	    KSTAT_DATA_UINT32);
1699 	kstat_named_init(&chkp->tx_reclaims[1],		"tx_reclaims[1]",
1700 	    KSTAT_DATA_UINT32);
1701 	kstat_named_init(&chkp->tx_msg_pullups,		"tx_msg_pullups",
1702 	    KSTAT_DATA_UINT32);
1703 	kstat_named_init(&chkp->tx_hdr_pullups,		"tx_hdr_pullups",
1704 	    KSTAT_DATA_UINT32);
1705 	kstat_named_init(&chkp->tx_tcp_ip_frag,		"tx_tcp_ip_frag",
1706 	    KSTAT_DATA_UINT32);
1707 	kstat_named_init(&chkp->tx_udp_ip_frag,		"tx_udp_ip_frag",
1708 	    KSTAT_DATA_UINT32);
1709 	kstat_named_init(&chkp->tx_soft_cksums,		"tx_soft_cksums",
1710 	    KSTAT_DATA_UINT32);
1711 	kstat_named_init(&chkp->tx_need_cpl_space,	"tx_need_cpl_space",
1712 	    KSTAT_DATA_UINT32);
1713 	kstat_named_init(&chkp->tx_multi_mblks,		"tx_multi_mblks",
1714 	    KSTAT_DATA_UINT32);
1715 	kstat_named_init(&chkp->tx_no_dvma1,	"tx_num_multi_dvma_fails",
1716 	    KSTAT_DATA_UINT32);
1717 	kstat_named_init(&chkp->tx_no_dvma2,	"tx_num_single_dvma_fails",
1718 	    KSTAT_DATA_UINT32);
1719 	kstat_named_init(&chkp->tx_no_dma1,	"tx_num_multi_dma_fails",
1720 	    KSTAT_DATA_UINT32);
1721 	kstat_named_init(&chkp->tx_no_dma2,	"tx_num_single_dma_fails",
1722 	    KSTAT_DATA_UINT32);
1723 	kstat_named_init(&chkp->rx_cmdq0,		"rx_cmdq0",
1724 	    KSTAT_DATA_UINT32);
1725 	kstat_named_init(&chkp->rx_cmdq1,		"rx_cmdq1",
1726 	    KSTAT_DATA_UINT32);
1727 	kstat_named_init(&chkp->rx_flq0,		"rx_flq0",
1728 	    KSTAT_DATA_UINT32);
1729 	kstat_named_init(&chkp->rx_flq1,		"rx_flq1",
1730 	    KSTAT_DATA_UINT32);
1731 	kstat_named_init(&chkp->rx_flq0_sz,		"rx_flq0_buffer_sz",
1732 	    KSTAT_DATA_UINT32);
1733 	kstat_named_init(&chkp->rx_flq1_sz,		"rx_flq1_buffer_sz",
1734 	    KSTAT_DATA_UINT32);
1735 	kstat_named_init(&chkp->rx_pkt_drops,		"rx_pkt_drops",
1736 	    KSTAT_DATA_UINT32);
1737 	kstat_named_init(&chkp->rx_pkt_copied,		"rx_pkt_copied",
1738 	    KSTAT_DATA_UINT32);
1739 	kstat_named_init(&chkp->rx_pause_on,		"rx_pause_on",
1740 	    KSTAT_DATA_UINT32);
1741 	kstat_named_init(&chkp->rx_pause_off,		"rx_pause_off",
1742 	    KSTAT_DATA_UINT32);
1743 	kstat_named_init(&chkp->rx_pause_ms,		"rx_pause_ms",
1744 	    KSTAT_DATA_UINT32);
1745 	kstat_named_init(&chkp->rx_pause_spike,		"rx_pause_spike",
1746 	    KSTAT_DATA_UINT32);
1747 	kstat_named_init(&chkp->rx_fl_credits,		"rx_fl_credits",
1748 	    KSTAT_DATA_UINT32);
1749 	kstat_named_init(&chkp->rx_flbuf_fails,		"rx_flbuf_fails",
1750 	    KSTAT_DATA_UINT32);
1751 	kstat_named_init(&chkp->rx_flbuf_allocs,	"rx_flbuf_allocs",
1752 	    KSTAT_DATA_UINT32);
1753 	kstat_named_init(&chkp->rx_badEopSop,		"rx_badEopSop",
1754 	    KSTAT_DATA_UINT32);
1755 	kstat_named_init(&chkp->rx_flq0_cnt,		"rx_flq0_cnt",
1756 	    KSTAT_DATA_UINT32);
1757 	kstat_named_init(&chkp->rx_flq1_cnt,		"rx_flq1_cnt",
1758 	    KSTAT_DATA_UINT32);
1759 	kstat_named_init(&chkp->arp_sent,		"arp_sent",
1760 	    KSTAT_DATA_UINT32);
1761 	kstat_named_init(&chkp->tx_doorbells,		"tx_doorbells",
1762 	    KSTAT_DATA_UINT32);
1763 	kstat_named_init(&chkp->intr_doorbells,		"intr_doorbells",
1764 	    KSTAT_DATA_UINT32);
1765 	kstat_named_init(&chkp->intr1_doorbells,	"intr1_doorbells",
1766 	    KSTAT_DATA_UINT32);
1767 	kstat_named_init(&chkp->sleep_cnt,		"sleep_cnt",
1768 	    KSTAT_DATA_UINT32);
1769 	kstat_named_init(&chkp->pe_allocb_cnt,		"pe_allocb_cnt",
1770 	    KSTAT_DATA_UINT32);
1771 	for (i = 0; i < MBLK_MAX; i++) {
1772 		(void) sprintf(kstat_name, "tx_descs[%02d]", i);
1773 		kstat_named_init(&chkp->tx_descs[i],
1774 		    kstat_name, KSTAT_DATA_UINT32);
1775 	}
1776 	ksp->ks_update = sge_kstat_update;
1777 	ksp->ks_private = (void *)sge;
1778 	sge->ksp = ksp;
1779 	kstat_install(ksp);
1780 	status = 0;
1781 
1782 sge_kstat_setup_exit:
1783 	return (status);
1784 }
1785 
1786 static void
1787 sge_kstat_remove(pesge *sge)
1788 {
1789 	if (sge->ksp)
1790 		kstat_delete(sge->ksp);
1791 }
1792 
1793 static int
1794 sge_kstat_update(p_kstat_t ksp, int rw)
1795 {
1796 	pesge *sge;
1797 	p_ch_stats_t statsp;
1798 	p_ch_kstat_t chkp;
1799 	int i;
1800 
1801 	sge = (pesge *)ksp->ks_private;
1802 	statsp = (p_ch_stats_t)&sge->intr_cnt;
1803 	chkp = (p_ch_kstat_t)ksp->ks_data;
1804 	if (rw == KSTAT_WRITE) {
1805 		statsp->respQ_empty	= chkp->respQ_empty.value.ui32;
1806 		statsp->respQ_overflow	= chkp->respQ_overflow.value.ui32;
1807 		statsp->freelistQ_empty	= chkp->freelistQ_empty.value.ui32;
1808 		statsp->pkt_too_big	= chkp->pkt_too_big.value.ui32;
1809 		statsp->pkt_mismatch	= chkp->pkt_mismatch.value.ui32;
1810 		statsp->cmdQ_full[0]	= chkp->cmdQ_full[0].value.ui32;
1811 		statsp->cmdQ_full[1]	= chkp->cmdQ_full[1].value.ui32;
1812 		statsp->tx_reclaims[0]	= chkp->tx_reclaims[0].value.ui32;
1813 		statsp->tx_reclaims[1]	= chkp->tx_reclaims[1].value.ui32;
1814 		statsp->tx_msg_pullups	= chkp->tx_msg_pullups.value.ui32;
1815 		statsp->tx_hdr_pullups	= chkp->tx_hdr_pullups.value.ui32;
1816 		statsp->tx_tcp_ip_frag	= chkp->tx_tcp_ip_frag.value.ui32;
1817 		statsp->tx_udp_ip_frag	= chkp->tx_udp_ip_frag.value.ui32;
1818 		statsp->tx_soft_cksums	= chkp->tx_soft_cksums.value.ui32;
1819 		statsp->tx_need_cpl_space
1820 		    = chkp->tx_need_cpl_space.value.ui32;
1821 		statsp->tx_multi_mblks	= chkp->tx_multi_mblks.value.ui32;
1822 		statsp->tx_no_dvma1	= chkp->tx_no_dvma1.value.ui32;
1823 		statsp->tx_no_dvma2	= chkp->tx_no_dvma2.value.ui32;
1824 		statsp->tx_no_dma1	= chkp->tx_no_dma1.value.ui32;
1825 		statsp->tx_no_dma2	= chkp->tx_no_dma2.value.ui32;
1826 		statsp->rx_cmdq0	= chkp->rx_cmdq0.value.ui32;
1827 		statsp->rx_cmdq1	= chkp->rx_cmdq1.value.ui32;
1828 		statsp->rx_flq0		= chkp->rx_flq0.value.ui32;
1829 		statsp->rx_flq1		= chkp->rx_flq1.value.ui32;
1830 		statsp->rx_flq0_sz	= chkp->rx_flq0_sz.value.ui32;
1831 		statsp->rx_flq1_sz	= chkp->rx_flq1_sz.value.ui32;
1832 		statsp->rx_pkt_drops	= chkp->rx_pkt_drops.value.ui32;
1833 		statsp->rx_pkt_copied	= chkp->rx_pkt_copied.value.ui32;
1834 		statsp->rx_pause_on	= chkp->rx_pause_on.value.ui32;
1835 		statsp->rx_pause_off	= chkp->rx_pause_off.value.ui32;
1836 		statsp->rx_pause_ms	= chkp->rx_pause_ms.value.ui32;
1837 		statsp->rx_pause_spike	= chkp->rx_pause_spike.value.ui32;
1838 		statsp->rx_fl_credits	= chkp->rx_fl_credits.value.ui32;
1839 		statsp->rx_flbuf_fails	= chkp->rx_flbuf_fails.value.ui32;
1840 		statsp->rx_flbuf_allocs	= chkp->rx_flbuf_allocs.value.ui32;
1841 		statsp->rx_badEopSop	= chkp->rx_badEopSop.value.ui32;
1842 		statsp->rx_flq0_cnt	= chkp->rx_flq0_cnt.value.ui32;
1843 		statsp->rx_flq1_cnt	= chkp->rx_flq1_cnt.value.ui32;
1844 		statsp->arp_sent	= chkp->arp_sent.value.ui32;
1845 		statsp->tx_doorbells	= chkp->tx_doorbells.value.ui32;
1846 		statsp->intr_doorbells	= chkp->intr_doorbells.value.ui32;
1847 		statsp->intr1_doorbells = chkp->intr1_doorbells.value.ui32;
1848 		statsp->sleep_cnt	= chkp->sleep_cnt.value.ui32;
1849 		statsp->pe_allocb_cnt	= chkp->pe_allocb_cnt.value.ui32;
1850 		for (i = 0; i < MBLK_MAX; i++) {
1851 			statsp->tx_descs[i] = chkp->tx_descs[i].value.ui32;
1852 		}
1853 	} else {
1854 		chkp->respQ_empty.value.ui32	= statsp->respQ_empty;
1855 		chkp->respQ_overflow.value.ui32	= statsp->respQ_overflow;
1856 		chkp->freelistQ_empty.value.ui32
1857 		    = statsp->freelistQ_empty;
1858 		chkp->pkt_too_big.value.ui32	= statsp->pkt_too_big;
1859 		chkp->pkt_mismatch.value.ui32	= statsp->pkt_mismatch;
1860 		chkp->cmdQ_full[0].value.ui32	= statsp->cmdQ_full[0];
1861 		chkp->cmdQ_full[1].value.ui32	= statsp->cmdQ_full[1];
1862 		chkp->tx_reclaims[0].value.ui32	= statsp->tx_reclaims[0];
1863 		chkp->tx_reclaims[1].value.ui32	= statsp->tx_reclaims[1];
1864 		chkp->tx_msg_pullups.value.ui32	= statsp->tx_msg_pullups;
1865 		chkp->tx_hdr_pullups.value.ui32	= statsp->tx_hdr_pullups;
1866 		chkp->tx_tcp_ip_frag.value.ui32	= statsp->tx_tcp_ip_frag;
1867 		chkp->tx_udp_ip_frag.value.ui32	= statsp->tx_udp_ip_frag;
1868 		chkp->tx_soft_cksums.value.ui32	= statsp->tx_soft_cksums;
1869 		chkp->tx_need_cpl_space.value.ui32
1870 		    = statsp->tx_need_cpl_space;
1871 		chkp->tx_multi_mblks.value.ui32	= statsp->tx_multi_mblks;
1872 		chkp->tx_no_dvma1.value.ui32	= statsp->tx_no_dvma1;
1873 		chkp->tx_no_dvma2.value.ui32	= statsp->tx_no_dvma2;
1874 		chkp->tx_no_dma1.value.ui32	= statsp->tx_no_dma1;
1875 		chkp->tx_no_dma2.value.ui32	= statsp->tx_no_dma2;
1876 		chkp->rx_cmdq0.value.ui32	= statsp->rx_cmdq0;
1877 		chkp->rx_cmdq1.value.ui32	= statsp->rx_cmdq1;
1878 		chkp->rx_flq0.value.ui32	= statsp->rx_flq0;
1879 		chkp->rx_flq1.value.ui32	= statsp->rx_flq1;
1880 		chkp->rx_flq0_sz.value.ui32	= statsp->rx_flq0_sz;
1881 		chkp->rx_flq1_sz.value.ui32	= statsp->rx_flq1_sz;
1882 		chkp->rx_pkt_drops.value.ui32	= statsp->rx_pkt_drops;
1883 		chkp->rx_pkt_copied.value.ui32	= statsp->rx_pkt_copied;
1884 		chkp->rx_pause_on.value.ui32	= statsp->rx_pause_on;
1885 		chkp->rx_pause_off.value.ui32	= statsp->rx_pause_off;
1886 		chkp->rx_pause_ms.value.ui32	= statsp->rx_pause_ms;
1887 		chkp->rx_pause_spike.value.ui32	= statsp->rx_pause_spike;
1888 		chkp->rx_fl_credits.value.ui32	= statsp->rx_fl_credits;
1889 		chkp->rx_flbuf_fails.value.ui32
1890 		    = statsp->rx_flbuf_fails;
1891 		chkp->rx_flbuf_allocs.value.ui32
1892 		    = statsp->rx_flbuf_allocs;
1893 		chkp->rx_badEopSop.value.ui32	= statsp->rx_badEopSop;
1894 		chkp->rx_flq0_cnt.value.ui32	= statsp->rx_flq0_cnt;
1895 		chkp->rx_flq1_cnt.value.ui32	= statsp->rx_flq1_cnt;
1896 		chkp->arp_sent.value.ui32	= statsp->arp_sent;
1897 		chkp->tx_doorbells.value.ui32	= statsp->tx_doorbells;
1898 		chkp->intr_doorbells.value.ui32	= statsp->intr_doorbells;
1899 		chkp->intr1_doorbells.value.ui32
1900 		    = statsp->intr1_doorbells;
1901 		chkp->sleep_cnt.value.ui32	= statsp->sleep_cnt;
1902 		chkp->pe_allocb_cnt.value.ui32	= statsp->pe_allocb_cnt;
1903 		for (i = 0; i < MBLK_MAX; i++) {
1904 			chkp->tx_descs[i].value.ui32 = statsp->tx_descs[i];
1905 		}
1906 	}
1907 	return (0);
1908 }
1909 #endif
1910 
1911 static uint16_t
1912 calc_ocsum(mblk_t *mp, int offset)
1913 {
1914 	uint8_t *addrp;
1915 	uint32_t src;
1916 	uint32_t dst;
1917 
1918 	ipha_t *ihdr = (ipha_t *)(mp->b_rptr + offset);
1919 	uint32_t sum;
1920 	int iplen = IPH_HDR_LENGTH(ihdr);
1921 	struct udphdr *udpp = (struct udphdr *)(mp->b_rptr + offset + iplen);
1922 	uchar_t *byte;
1923 	int len;
1924 
1925 	addrp = (uint8_t *)&ihdr->ipha_src;
1926 	src =  ((uint32_t)(addrp[0]) << 24) | ((uint32_t)(addrp[1]) << 16) |
1927 	    ((uint32_t)(addrp[2]) << 8) | (uint32_t)(addrp[3]);
1928 
1929 	addrp = (uint8_t *)&ihdr->ipha_dst;
1930 	dst =  ((uint32_t)(addrp[0]) << 24) | ((uint32_t)(addrp[1]) << 16) |
1931 	    ((uint32_t)(addrp[2]) << 8) | (uint32_t)(addrp[3]);
1932 
1933 	sum = (uint16_t)(src >> 16) +
1934 	    (uint16_t)(src) +
1935 	    (uint16_t)(dst >> 16) +
1936 	    (uint16_t)(dst) + (udpp->uh_ulen + htons(IPPROTO_UDP));
1937 
1938 	sum = (uint16_t)(sum >> 16) + (uint16_t)(sum);
1939 
1940 	if (sum > 0xffff)
1941 		sum -= 0xffff;
1942 
1943 	udpp->uh_sum = 0;
1944 	byte = mp->b_rptr + offset + iplen;
1945 	do {
1946 		len = (mp->b_wptr - byte);
1947 		sum = bcksum(byte, len, sum);
1948 		if (sum > 0xffff)
1949 			sum -= 0xffff;
1950 		mp = mp->b_cont;
1951 		if (mp)
1952 			byte = mp->b_rptr;
1953 	} while (mp);
1954 
1955 	sum = ~sum & 0xffff;
1956 
1957 	return (sum);
1958 }
1959