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 (c) 2010-2013, by Broadcom, Inc.
24 * All Rights Reserved.
25 */
26
27 /*
28 * Copyright (c) 2002, 2010, Oracle and/or its affiliates.
29 * All rights reserved.
30 */
31
32 #include "bge_impl.h"
33
34
35 /*
36 * The transmit-side code uses an allocation process which is similar
37 * to some theme park roller-coaster rides, where riders sit in cars
38 * that can go individually, but work better in a train.
39 *
40 * 1) RESERVE a place - this doesn't refer to any specific car or
41 * seat, just that you will get a ride. The attempt to RESERVE a
42 * place can fail if all spaces in all cars are already committed.
43 *
44 * 2) Prepare yourself; this may take an arbitrary (but not unbounded)
45 * time, and you can back out at this stage, in which case you must
46 * give up (RENOUNCE) your place.
47 *
48 * 3) CLAIM your space - a specific car (the next sequentially
49 * numbered one) is allocated at this stage, and is guaranteed
50 * to be part of the next train to depart. Once you've done
51 * this, you can't back out, nor wait for any external event
52 * or resource.
53 *
54 * 4) Occupy your car - when all CLAIMED cars are OCCUPIED, they
55 * all depart together as a single train!
56 *
57 * 5) At the end of the ride, you climb out of the car and RENOUNCE
58 * your right to it, so that it can be recycled for another rider.
59 *
60 * For each rider, these have to occur in this order, but the riders
61 * don't have to stay in the same order at each stage. In particular,
62 * they may overtake each other between RESERVING a place and CLAIMING
63 * it, or between CLAIMING and OCCUPYING a space.
64 *
65 * Once a car is CLAIMED, the train currently being assembled can't go
66 * without that car (this guarantees that the cars in a single train
67 * make up a consecutively-numbered set). Therefore, when any train
68 * leaves, we know there can't be any riders in transit between CLAIMING
69 * and OCCUPYING their cars. There can be some who have RESERVED but
70 * not yet CLAIMED their places. That's OK, though, because they'll go
71 * into the next train.
72 */
73
74 #define BGE_DBG BGE_DBG_SEND /* debug flag for this code */
75
76 /*
77 * ========== Send-side recycle routines ==========
78 */
79
80 /*
81 * Recycle all the completed buffers in the specified send ring up to
82 * (but not including) the consumer index in the status block.
83 *
84 * This function must advance (srp->tc_next) AND adjust (srp->tx_free)
85 * to account for the packets it has recycled.
86 *
87 * This is a trivial version that just does that and nothing more, but
88 * it suffices while there's only one method for sending messages (by
89 * copying) and that method doesn't need any special per-buffer action
90 * for recycling.
91 */
92 static boolean_t
bge_recycle_ring(bge_t * bgep,send_ring_t * srp)93 bge_recycle_ring(bge_t *bgep, send_ring_t *srp)
94 {
95 sw_sbd_t *ssbdp;
96 bge_queue_item_t *buf_item;
97 bge_queue_item_t *buf_item_head;
98 bge_queue_item_t *buf_item_tail;
99 bge_queue_t *txbuf_queue;
100 uint64_t slot;
101 uint64_t n;
102
103 ASSERT(mutex_owned(srp->tc_lock));
104
105 /*
106 * We're about to release one or more places :-)
107 * These ASSERTions check that our invariants still hold:
108 * there must always be at least one free place
109 * at this point, there must be at least one place NOT free
110 * we're not about to free more places than were claimed!
111 */
112 ASSERT(srp->tx_free <= srp->desc.nslots);
113
114 buf_item_head = buf_item_tail = NULL;
115 for (n = 0, slot = srp->tc_next; slot != *srp->cons_index_p;
116 slot = NEXT(slot, srp->desc.nslots)) {
117 ssbdp = &srp->sw_sbds[slot];
118 ASSERT(ssbdp->pbuf != NULL);
119 buf_item = ssbdp->pbuf;
120 if (buf_item_head == NULL)
121 buf_item_head = buf_item_tail = buf_item;
122 else {
123 buf_item_tail->next = buf_item;
124 buf_item_tail = buf_item;
125 }
126 ssbdp->pbuf = NULL;
127 n++;
128 }
129 if (n == 0)
130 return (B_FALSE);
131
132 /*
133 * Reset the watchdog count: to 0 if all buffers are
134 * now free, or to 1 if some are still outstanding.
135 * Note: non-synchonised access here means we may get
136 * the "wrong" answer, but only in a harmless fashion
137 * (i.e. we deactivate the watchdog because all buffers
138 * are apparently free, even though another thread may
139 * have claimed one before we leave here; in this case
140 * the watchdog will restart on the next send() call).
141 */
142 bgep->watchdog = (slot == srp->tx_next) ? 0 : 1;
143
144 /*
145 * Update recycle index and free tx BD number
146 */
147 srp->tc_next = slot;
148 ASSERT(srp->tx_free + n <= srp->desc.nslots);
149 bge_atomic_renounce(&srp->tx_free, n);
150
151 /*
152 * Return tx buffers to buffer push queue
153 */
154 txbuf_queue = srp->txbuf_push_queue;
155 mutex_enter(txbuf_queue->lock);
156 buf_item_tail->next = txbuf_queue->head;
157 txbuf_queue->head = buf_item_head;
158 txbuf_queue->count += n;
159 mutex_exit(txbuf_queue->lock);
160
161 /*
162 * Check if we need exchange the tx buffer push and pop queue
163 */
164 if ((srp->txbuf_pop_queue->count < srp->tx_buffers_low) &&
165 (srp->txbuf_pop_queue->count < txbuf_queue->count)) {
166 srp->txbuf_push_queue = srp->txbuf_pop_queue;
167 srp->txbuf_pop_queue = txbuf_queue;
168 }
169
170 if (srp->tx_flow != 0 || bgep->tx_resched_needed)
171 ddi_trigger_softintr(bgep->drain_id);
172
173 return (B_TRUE);
174 }
175
176 /*
177 * Recycle all returned slots in all rings.
178 *
179 * To give priority to low-numbered rings, whenever we have recycled any
180 * slots in any ring except 0, we restart scanning again from ring 0.
181 * Thus, for example, if rings 0, 3, and 10 are carrying traffic, the
182 * pattern of recycles might go 0, 3, 10, 3, 0, 10, 0:
183 *
184 * 0 found some - recycle them
185 * 1..2 none found
186 * 3 found some - recycle them and restart scan
187 * 0..9 none found
188 * 10 found some - recycle them and restart scan
189 * 0..2 none found
190 * 3 found some more - recycle them and restart scan
191 * 0 found some more - recycle them
192 * 0..9 none found
193 * 10 found some more - recycle them and restart scan
194 * 0 found some more - recycle them
195 * 1..15 none found
196 *
197 * The routine returns only when a complete scan has been performed
198 * without finding any slots to recycle.
199 *
200 * Note: the expression (BGE_SEND_RINGS_USED > 1) yields a compile-time
201 * constant and allows the compiler to optimise away the outer do-loop
202 * if only one send ring is being used.
203 */
204 boolean_t bge_recycle(bge_t *bgep, bge_status_t *bsp);
205
206 boolean_t
bge_recycle(bge_t * bgep,bge_status_t * bsp)207 bge_recycle(bge_t *bgep, bge_status_t *bsp)
208 {
209 send_ring_t *srp;
210 uint64_t ring;
211 uint64_t tx_rings = bgep->chipid.tx_rings;
212 boolean_t tx_done = B_FALSE;
213
214 restart:
215 ring = 0;
216 srp = &bgep->send[ring];
217 do {
218 /*
219 * For each ring, (srp->cons_index_p) points to the
220 * proper index within the status block (which has
221 * already been sync'd by the caller).
222 */
223 ASSERT(srp->cons_index_p == SEND_INDEX_P(bsp, ring));
224
225 if (*srp->cons_index_p == srp->tc_next)
226 continue; /* no slots to recycle */
227 if (mutex_tryenter(srp->tc_lock) == 0)
228 continue; /* already in process */
229 tx_done |= bge_recycle_ring(bgep, srp);
230 mutex_exit(srp->tc_lock);
231
232 /*
233 * Restart from ring 0, if we're not on ring 0 already.
234 * As H/W selects send BDs totally based on priority and
235 * available BDs on the higher priority ring are always
236 * selected first, driver should keep consistence with H/W
237 * and gives lower-numbered ring with higher priority.
238 */
239 if (tx_rings > 1 && ring > 0)
240 goto restart;
241
242 /*
243 * Loop over all rings (if there *are* multiple rings)
244 */
245 } while (++srp, ++ring < tx_rings);
246
247 return (tx_done);
248 }
249
250
251 /*
252 * ========== Send-side transmit routines ==========
253 */
254 #define TCP_CKSUM_OFFSET 16
255 #define UDP_CKSUM_OFFSET 6
256
257 static void
bge_pseudo_cksum(uint8_t * buf)258 bge_pseudo_cksum(uint8_t *buf)
259 {
260 uint32_t cksum;
261 uint16_t iphl;
262 uint16_t proto;
263
264 /*
265 * Point it to the ip header.
266 */
267 buf += sizeof (struct ether_header);
268
269 /*
270 * Calculate the pseudo-header checksum.
271 */
272 iphl = 4 * (buf[0] & 0xF);
273 cksum = (((uint16_t)buf[2])<<8) + buf[3] - iphl;
274 cksum += proto = buf[9];
275 cksum += (((uint16_t)buf[12])<<8) + buf[13];
276 cksum += (((uint16_t)buf[14])<<8) + buf[15];
277 cksum += (((uint16_t)buf[16])<<8) + buf[17];
278 cksum += (((uint16_t)buf[18])<<8) + buf[19];
279 cksum = (cksum>>16) + (cksum & 0xFFFF);
280 cksum = (cksum>>16) + (cksum & 0xFFFF);
281
282 /*
283 * Point it to the TCP/UDP header, and
284 * update the checksum field.
285 */
286 buf += iphl + ((proto == IPPROTO_TCP) ?
287 TCP_CKSUM_OFFSET : UDP_CKSUM_OFFSET);
288
289 /*
290 * A real possibility that pointer cast is a problem.
291 * Should be fixed when we know the code better.
292 * E_BAD_PTR_CAST_ALIGN is added to make it temporarily clean.
293 */
294 *(uint16_t *)buf = htons((uint16_t)cksum);
295 }
296
297 static bge_queue_item_t *
bge_get_txbuf(bge_t * bgep,send_ring_t * srp)298 bge_get_txbuf(bge_t *bgep, send_ring_t *srp)
299 {
300 bge_queue_item_t *txbuf_item;
301 bge_queue_t *txbuf_queue;
302
303 txbuf_queue = srp->txbuf_pop_queue;
304 mutex_enter(txbuf_queue->lock);
305 if (txbuf_queue->count == 0) {
306 mutex_exit(txbuf_queue->lock);
307 txbuf_queue = srp->txbuf_push_queue;
308 mutex_enter(txbuf_queue->lock);
309 if (txbuf_queue->count == 0) {
310 mutex_exit(txbuf_queue->lock);
311 /* Try to allocate more tx buffers */
312 if (srp->tx_array < srp->tx_array_max) {
313 mutex_enter(srp->tx_lock);
314 txbuf_item = bge_alloc_txbuf_array(bgep, srp);
315 mutex_exit(srp->tx_lock);
316 } else
317 txbuf_item = NULL;
318 return (txbuf_item);
319 }
320 }
321 txbuf_item = txbuf_queue->head;
322 txbuf_queue->head = (bge_queue_item_t *)txbuf_item->next;
323 txbuf_queue->count--;
324 mutex_exit(txbuf_queue->lock);
325 txbuf_item->next = NULL;
326
327 return (txbuf_item);
328 }
329
330 /*
331 * Send a message by copying it into a preallocated (and premapped) buffer
332 */
333 static void bge_send_copy(bge_t *bgep, sw_txbuf_t *txbuf, mblk_t *mp);
334
335 static void
bge_send_copy(bge_t * bgep,sw_txbuf_t * txbuf,mblk_t * mp)336 bge_send_copy(bge_t *bgep, sw_txbuf_t *txbuf, mblk_t *mp)
337 {
338 mblk_t *bp;
339 uint32_t mblen;
340 char *pbuf;
341
342 txbuf->copy_len = 0;
343 pbuf = DMA_VPTR(txbuf->buf);
344 for (bp = mp; bp != NULL; bp = bp->b_cont) {
345 if ((mblen = MBLKL(bp)) == 0)
346 continue;
347 ASSERT(txbuf->copy_len + mblen <=
348 bgep->chipid.snd_buff_size);
349 bcopy(bp->b_rptr, pbuf, mblen);
350 pbuf += mblen;
351 txbuf->copy_len += mblen;
352 }
353 }
354
355 /*
356 * Fill the Tx buffer descriptors and trigger the h/w transmission
357 */
358 static void
bge_send_serial(bge_t * bgep,send_ring_t * srp)359 bge_send_serial(bge_t *bgep, send_ring_t *srp)
360 {
361 send_pkt_t *pktp;
362 uint64_t txfill_next;
363 uint32_t count;
364 uint32_t tx_next;
365 sw_sbd_t *ssbdp;
366 bge_status_t *bsp;
367 bge_sbd_t *hw_sbd_p;
368 bge_queue_item_t *txbuf_item;
369 sw_txbuf_t *txbuf;
370
371 /*
372 * Try to hold the tx lock:
373 * If we are in an interrupt context, use mutex_enter() to
374 * ensure quick response for tx in interrupt context;
375 * Otherwise, use mutex_tryenter() to serialize this h/w tx
376 * BD filling and transmission triggering task.
377 */
378 if (servicing_interrupt() != 0)
379 mutex_enter(srp->tx_lock);
380 else if (mutex_tryenter(srp->tx_lock) == 0)
381 return; /* already in process */
382
383 bsp = DMA_VPTR(bgep->status_block);
384 txfill_next = srp->txfill_next;
385 tx_next = srp->tx_next;
386 start_tx:
387 for (count = 0; count < bgep->param_drain_max; ++count) {
388 pktp = &srp->pktp[txfill_next];
389 if (!pktp->tx_ready) {
390 if (count == 0)
391 srp->tx_block++;
392 break;
393 }
394
395 /*
396 * If there are no enough BDs: try to recycle more
397 */
398 if (srp->tx_free <= 1)
399 (void) bge_recycle(bgep, bsp);
400
401 /*
402 * Reserved required BDs: 1 is enough
403 */
404 if (!bge_atomic_reserve(&srp->tx_free, 1)) {
405 srp->tx_nobd++;
406 break;
407 }
408
409 /*
410 * Filling the tx BD
411 */
412
413 /*
414 * Go straight to claiming our already-reserved places
415 * on the train!
416 */
417 ASSERT(pktp->txbuf_item != NULL);
418 txbuf_item = pktp->txbuf_item;
419 pktp->txbuf_item = NULL;
420 pktp->tx_ready = B_FALSE;
421
422 txbuf = txbuf_item->item;
423 ASSERT(txbuf->copy_len != 0);
424 (void) ddi_dma_sync(txbuf->buf.dma_hdl, 0,
425 txbuf->copy_len, DDI_DMA_SYNC_FORDEV);
426
427 ssbdp = &srp->sw_sbds[tx_next];
428 ASSERT(ssbdp->pbuf == NULL);
429 ssbdp->pbuf = txbuf_item;
430
431 /*
432 * Setting hardware send buffer descriptor
433 */
434 hw_sbd_p = DMA_VPTR(ssbdp->desc);
435 hw_sbd_p->flags = 0;
436 hw_sbd_p->host_buf_addr = txbuf->buf.cookie.dmac_laddress;
437 hw_sbd_p->len = txbuf->copy_len;
438 if (pktp->vlan_tci != 0) {
439 hw_sbd_p->vlan_tci = pktp->vlan_tci;
440 hw_sbd_p->host_buf_addr += VLAN_TAGSZ;
441 hw_sbd_p->flags |= SBD_FLAG_VLAN_TAG;
442 }
443 if (pktp->pflags & HCK_IPV4_HDRCKSUM)
444 hw_sbd_p->flags |= SBD_FLAG_IP_CKSUM;
445 if (pktp->pflags & HCK_FULLCKSUM)
446 hw_sbd_p->flags |= SBD_FLAG_TCP_UDP_CKSUM;
447 if (!(bgep->chipid.flags & CHIP_FLAG_NO_JUMBO) &&
448 (DEVICE_5717_SERIES_CHIPSETS(bgep) ||
449 DEVICE_5725_SERIES_CHIPSETS(bgep) ||
450 DEVICE_57765_SERIES_CHIPSETS(bgep)) &&
451 (txbuf->copy_len > ETHERMAX))
452 hw_sbd_p->flags |= SBD_FLAG_JMB_PKT;
453 hw_sbd_p->flags |= SBD_FLAG_PACKET_END;
454
455 txfill_next = NEXT(txfill_next, BGE_SEND_BUF_MAX);
456 tx_next = NEXT(tx_next, srp->desc.nslots);
457 }
458
459 /*
460 * Trigger h/w to start transmission.
461 */
462 if (count != 0) {
463 bge_atomic_sub64(&srp->tx_flow, count);
464 srp->txfill_next = txfill_next;
465
466 if (srp->tx_next > tx_next) {
467 (void) ddi_dma_sync(ssbdp->desc.dma_hdl, 0,
468 (srp->desc.nslots - srp->tx_next) *
469 sizeof (bge_sbd_t),
470 DDI_DMA_SYNC_FORDEV);
471 count -= srp->desc.nslots - srp->tx_next;
472 ssbdp = &srp->sw_sbds[0];
473 }
474 (void) ddi_dma_sync(ssbdp->desc.dma_hdl, 0,
475 count*sizeof (bge_sbd_t), DDI_DMA_SYNC_FORDEV);
476 bge_mbx_put(bgep, srp->chip_mbx_reg, tx_next);
477 srp->tx_next = tx_next;
478 atomic_or_32(&bgep->watchdog, 1);
479
480 if (srp->tx_flow != 0 && srp->tx_free > 1)
481 goto start_tx;
482 }
483
484 mutex_exit(srp->tx_lock);
485 }
486
487 mblk_t *
bge_ring_tx(void * arg,mblk_t * mp)488 bge_ring_tx(void *arg, mblk_t *mp)
489 {
490 send_ring_t *srp = arg;
491 bge_t *bgep = srp->bgep;
492 struct ether_vlan_header *ehp;
493 bge_queue_item_t *txbuf_item;
494 sw_txbuf_t *txbuf;
495 send_pkt_t *pktp;
496 uint64_t pkt_slot;
497 uint16_t vlan_tci;
498 uint32_t pflags;
499 char *pbuf;
500
501 ASSERT(mp->b_next == NULL);
502
503 /*
504 * Get a s/w tx buffer first
505 */
506 txbuf_item = bge_get_txbuf(bgep, srp);
507 if (txbuf_item == NULL) {
508 /* no tx buffer available */
509 srp->tx_nobuf++;
510 bgep->tx_resched_needed = B_TRUE;
511 bge_send_serial(bgep, srp);
512 return (mp);
513 }
514
515 /*
516 * Copy all mp fragments to the pkt buffer
517 */
518 txbuf = txbuf_item->item;
519 bge_send_copy(bgep, txbuf, mp);
520
521 /*
522 * Determine if the packet is VLAN tagged.
523 */
524 ASSERT(txbuf->copy_len >= sizeof (struct ether_header));
525 pbuf = DMA_VPTR(txbuf->buf);
526
527 ehp = (void *)pbuf;
528 if (ehp->ether_tpid == htons(ETHERTYPE_VLAN)) {
529 /* Strip the vlan tag */
530 vlan_tci = ntohs(ehp->ether_tci);
531 pbuf = memmove(pbuf + VLAN_TAGSZ, pbuf, 2 * ETHERADDRL);
532 txbuf->copy_len -= VLAN_TAGSZ;
533 } else
534 vlan_tci = 0;
535
536 /*
537 * Retrieve checksum offloading info.
538 */
539 mac_hcksum_get(mp, NULL, NULL, NULL, NULL, &pflags);
540
541 /*
542 * Calculate pseudo checksum if needed.
543 */
544 if ((pflags & HCK_FULLCKSUM) &&
545 (bgep->chipid.flags & CHIP_FLAG_PARTIAL_CSUM))
546 bge_pseudo_cksum((uint8_t *)pbuf);
547
548 /*
549 * Packet buffer is ready to send: get and fill pkt info
550 */
551 pkt_slot = bge_atomic_next(&srp->txpkt_next, BGE_SEND_BUF_MAX);
552 pktp = &srp->pktp[pkt_slot];
553 ASSERT(pktp->txbuf_item == NULL);
554 pktp->txbuf_item = txbuf_item;
555 pktp->vlan_tci = vlan_tci;
556 pktp->pflags = pflags;
557 atomic_inc_64(&srp->tx_flow);
558 ASSERT(pktp->tx_ready == B_FALSE);
559 pktp->tx_ready = B_TRUE;
560
561 /*
562 * Filling the h/w bd and trigger the h/w to start transmission
563 */
564 bge_send_serial(bgep, srp);
565
566 srp->pushed_bytes += MBLKL(mp);
567
568 /*
569 * We've copied the contents, the message can be freed right away
570 */
571 freemsg(mp);
572 return (NULL);
573 }
574
575 static mblk_t *
bge_send(bge_t * bgep,mblk_t * mp)576 bge_send(bge_t *bgep, mblk_t *mp)
577 {
578 send_ring_t *ring;
579
580 ring = &bgep->send[0]; /* ring 0 */
581
582 return (bge_ring_tx(ring, mp));
583 }
584
585 uint_t
bge_send_drain(caddr_t arg)586 bge_send_drain(caddr_t arg)
587 {
588 uint_t ring = 0; /* use ring 0 */
589 bge_t *bgep;
590 send_ring_t *srp;
591
592 bgep = (void *)arg;
593 BGE_TRACE(("bge_send_drain($%p)", (void *)bgep));
594
595 srp = &bgep->send[ring];
596 bge_send_serial(bgep, srp);
597
598 if (bgep->tx_resched_needed &&
599 (srp->tx_flow < srp->tx_buffers_low) &&
600 (bgep->bge_mac_state == BGE_MAC_STARTED)) {
601 mac_tx_update(bgep->mh);
602 bgep->tx_resched_needed = B_FALSE;
603 bgep->tx_resched++;
604 }
605
606 return (DDI_INTR_CLAIMED);
607 }
608
609 /*
610 * bge_m_tx() - send a chain of packets
611 */
612 mblk_t *
bge_m_tx(void * arg,mblk_t * mp)613 bge_m_tx(void *arg, mblk_t *mp)
614 {
615 bge_t *bgep = arg; /* private device info */
616 mblk_t *next;
617
618 BGE_TRACE(("bge_m_tx($%p, $%p)", arg, (void *)mp));
619
620 ASSERT(mp != NULL);
621 ASSERT(bgep->bge_mac_state == BGE_MAC_STARTED);
622
623 rw_enter(bgep->errlock, RW_READER);
624 if ((bgep->bge_chip_state != BGE_CHIP_RUNNING) ||
625 !(bgep->param_link_up)) {
626 BGE_DEBUG(("bge_m_tx: chip not running or link down"));
627 freemsgchain(mp);
628 mp = NULL;
629 }
630
631 while (mp != NULL) {
632 next = mp->b_next;
633 mp->b_next = NULL;
634
635 if ((mp = bge_send(bgep, mp)) != NULL) {
636 mp->b_next = next;
637 break;
638 }
639
640 mp = next;
641 }
642 rw_exit(bgep->errlock);
643
644 return (mp);
645 }
646