1 /*
2  * Copyright (c) 2012 Intel Corporation. All rights reserved.
3  * Copyright (c) 2007 - 2012 QLogic Corporation. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33 
34 #include <linux/spinlock.h>
35 #include <linux/netdevice.h>
36 #include <linux/moduleparam.h>
37 
38 #include "qib.h"
39 #include "qib_common.h"
40 
41 /* default pio off, sdma on */
42 static ushort sdma_descq_cnt = 256;
43 module_param_named(sdma_descq_cnt, sdma_descq_cnt, ushort, S_IRUGO);
44 MODULE_PARM_DESC(sdma_descq_cnt, "Number of SDMA descq entries");
45 
46 /*
47  * Bits defined in the send DMA descriptor.
48  */
49 #define SDMA_DESC_LAST          (1ULL << 11)
50 #define SDMA_DESC_FIRST         (1ULL << 12)
51 #define SDMA_DESC_DMA_HEAD      (1ULL << 13)
52 #define SDMA_DESC_USE_LARGE_BUF (1ULL << 14)
53 #define SDMA_DESC_INTR          (1ULL << 15)
54 #define SDMA_DESC_COUNT_LSB     16
55 #define SDMA_DESC_GEN_LSB       30
56 
57 /* declare all statics here rather than keep sorting */
58 static int alloc_sdma(struct qib_pportdata *);
59 static void sdma_complete(struct kref *);
60 static void sdma_finalput(struct qib_sdma_state *);
61 static void sdma_get(struct qib_sdma_state *);
62 static void sdma_put(struct qib_sdma_state *);
63 static void sdma_set_state(struct qib_pportdata *, enum qib_sdma_states);
64 static void sdma_start_sw_clean_up(struct qib_pportdata *);
65 static void sdma_sw_clean_up_task(unsigned long);
66 static void unmap_desc(struct qib_pportdata *, unsigned);
67 
68 static void sdma_get(struct qib_sdma_state *ss)
69 {
70 	kref_get(&ss->kref);
71 }
72 
73 static void sdma_complete(struct kref *kref)
74 {
75 	struct qib_sdma_state *ss =
76 		container_of(kref, struct qib_sdma_state, kref);
77 
78 	complete(&ss->comp);
79 }
80 
81 static void sdma_put(struct qib_sdma_state *ss)
82 {
83 	kref_put(&ss->kref, sdma_complete);
84 }
85 
86 static void sdma_finalput(struct qib_sdma_state *ss)
87 {
88 	sdma_put(ss);
89 	wait_for_completion(&ss->comp);
90 }
91 
92 /*
93  * Complete all the sdma requests on the active list, in the correct
94  * order, and with appropriate processing.   Called when cleaning up
95  * after sdma shutdown, and when new sdma requests are submitted for
96  * a link that is down.   This matches what is done for requests
97  * that complete normally, it's just the full list.
98  *
99  * Must be called with sdma_lock held
100  */
101 static void clear_sdma_activelist(struct qib_pportdata *ppd)
102 {
103 	struct qib_sdma_txreq *txp, *txp_next;
104 
105 	list_for_each_entry_safe(txp, txp_next, &ppd->sdma_activelist, list) {
106 		list_del_init(&txp->list);
107 		if (txp->flags & QIB_SDMA_TXREQ_F_FREEDESC) {
108 			unsigned idx;
109 
110 			idx = txp->start_idx;
111 			while (idx != txp->next_descq_idx) {
112 				unmap_desc(ppd, idx);
113 				if (++idx == ppd->sdma_descq_cnt)
114 					idx = 0;
115 			}
116 		}
117 		if (txp->callback)
118 			(*txp->callback)(txp, QIB_SDMA_TXREQ_S_ABORTED);
119 	}
120 }
121 
122 static void sdma_sw_clean_up_task(unsigned long opaque)
123 {
124 	struct qib_pportdata *ppd = (struct qib_pportdata *) opaque;
125 	unsigned long flags;
126 
127 	spin_lock_irqsave(&ppd->sdma_lock, flags);
128 
129 	/*
130 	 * At this point, the following should always be true:
131 	 * - We are halted, so no more descriptors are getting retired.
132 	 * - We are not running, so no one is submitting new work.
133 	 * - Only we can send the e40_sw_cleaned, so we can't start
134 	 *   running again until we say so.  So, the active list and
135 	 *   descq are ours to play with.
136 	 */
137 
138 	/* Process all retired requests. */
139 	qib_sdma_make_progress(ppd);
140 
141 	clear_sdma_activelist(ppd);
142 
143 	/*
144 	 * Resync count of added and removed.  It is VERY important that
145 	 * sdma_descq_removed NEVER decrement - user_sdma depends on it.
146 	 */
147 	ppd->sdma_descq_removed = ppd->sdma_descq_added;
148 
149 	/*
150 	 * Reset our notion of head and tail.
151 	 * Note that the HW registers will be reset when switching states
152 	 * due to calling __qib_sdma_process_event() below.
153 	 */
154 	ppd->sdma_descq_tail = 0;
155 	ppd->sdma_descq_head = 0;
156 	ppd->sdma_head_dma[0] = 0;
157 	ppd->sdma_generation = 0;
158 
159 	__qib_sdma_process_event(ppd, qib_sdma_event_e40_sw_cleaned);
160 
161 	spin_unlock_irqrestore(&ppd->sdma_lock, flags);
162 }
163 
164 /*
165  * This is called when changing to state qib_sdma_state_s10_hw_start_up_wait
166  * as a result of send buffer errors or send DMA descriptor errors.
167  * We want to disarm the buffers in these cases.
168  */
169 static void sdma_hw_start_up(struct qib_pportdata *ppd)
170 {
171 	struct qib_sdma_state *ss = &ppd->sdma_state;
172 	unsigned bufno;
173 
174 	for (bufno = ss->first_sendbuf; bufno < ss->last_sendbuf; ++bufno)
175 		ppd->dd->f_sendctrl(ppd, QIB_SENDCTRL_DISARM_BUF(bufno));
176 
177 	ppd->dd->f_sdma_hw_start_up(ppd);
178 }
179 
180 static void sdma_sw_tear_down(struct qib_pportdata *ppd)
181 {
182 	struct qib_sdma_state *ss = &ppd->sdma_state;
183 
184 	/* Releasing this reference means the state machine has stopped. */
185 	sdma_put(ss);
186 }
187 
188 static void sdma_start_sw_clean_up(struct qib_pportdata *ppd)
189 {
190 	tasklet_hi_schedule(&ppd->sdma_sw_clean_up_task);
191 }
192 
193 static void sdma_set_state(struct qib_pportdata *ppd,
194 	enum qib_sdma_states next_state)
195 {
196 	struct qib_sdma_state *ss = &ppd->sdma_state;
197 	struct sdma_set_state_action *action = ss->set_state_action;
198 	unsigned op = 0;
199 
200 	/* debugging bookkeeping */
201 	ss->previous_state = ss->current_state;
202 	ss->previous_op = ss->current_op;
203 
204 	ss->current_state = next_state;
205 
206 	if (action[next_state].op_enable)
207 		op |= QIB_SDMA_SENDCTRL_OP_ENABLE;
208 
209 	if (action[next_state].op_intenable)
210 		op |= QIB_SDMA_SENDCTRL_OP_INTENABLE;
211 
212 	if (action[next_state].op_halt)
213 		op |= QIB_SDMA_SENDCTRL_OP_HALT;
214 
215 	if (action[next_state].op_drain)
216 		op |= QIB_SDMA_SENDCTRL_OP_DRAIN;
217 
218 	if (action[next_state].go_s99_running_tofalse)
219 		ss->go_s99_running = 0;
220 
221 	if (action[next_state].go_s99_running_totrue)
222 		ss->go_s99_running = 1;
223 
224 	ss->current_op = op;
225 
226 	ppd->dd->f_sdma_sendctrl(ppd, ss->current_op);
227 }
228 
229 static void unmap_desc(struct qib_pportdata *ppd, unsigned head)
230 {
231 	__le64 *descqp = &ppd->sdma_descq[head].qw[0];
232 	u64 desc[2];
233 	dma_addr_t addr;
234 	size_t len;
235 
236 	desc[0] = le64_to_cpu(descqp[0]);
237 	desc[1] = le64_to_cpu(descqp[1]);
238 
239 	addr = (desc[1] << 32) | (desc[0] >> 32);
240 	len = (desc[0] >> 14) & (0x7ffULL << 2);
241 	dma_unmap_single(&ppd->dd->pcidev->dev, addr, len, DMA_TO_DEVICE);
242 }
243 
244 static int alloc_sdma(struct qib_pportdata *ppd)
245 {
246 	ppd->sdma_descq_cnt = sdma_descq_cnt;
247 	if (!ppd->sdma_descq_cnt)
248 		ppd->sdma_descq_cnt = 256;
249 
250 	/* Allocate memory for SendDMA descriptor FIFO */
251 	ppd->sdma_descq = dma_alloc_coherent(&ppd->dd->pcidev->dev,
252 		ppd->sdma_descq_cnt * sizeof(u64[2]), &ppd->sdma_descq_phys,
253 		GFP_KERNEL);
254 
255 	if (!ppd->sdma_descq) {
256 		qib_dev_err(ppd->dd,
257 			"failed to allocate SendDMA descriptor FIFO memory\n");
258 		goto bail;
259 	}
260 
261 	/* Allocate memory for DMA of head register to memory */
262 	ppd->sdma_head_dma = dma_alloc_coherent(&ppd->dd->pcidev->dev,
263 		PAGE_SIZE, &ppd->sdma_head_phys, GFP_KERNEL);
264 	if (!ppd->sdma_head_dma) {
265 		qib_dev_err(ppd->dd,
266 			"failed to allocate SendDMA head memory\n");
267 		goto cleanup_descq;
268 	}
269 	ppd->sdma_head_dma[0] = 0;
270 	return 0;
271 
272 cleanup_descq:
273 	dma_free_coherent(&ppd->dd->pcidev->dev,
274 		ppd->sdma_descq_cnt * sizeof(u64[2]), (void *)ppd->sdma_descq,
275 		ppd->sdma_descq_phys);
276 	ppd->sdma_descq = NULL;
277 	ppd->sdma_descq_phys = 0;
278 bail:
279 	ppd->sdma_descq_cnt = 0;
280 	return -ENOMEM;
281 }
282 
283 static void free_sdma(struct qib_pportdata *ppd)
284 {
285 	struct qib_devdata *dd = ppd->dd;
286 
287 	if (ppd->sdma_head_dma) {
288 		dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
289 				  (void *)ppd->sdma_head_dma,
290 				  ppd->sdma_head_phys);
291 		ppd->sdma_head_dma = NULL;
292 		ppd->sdma_head_phys = 0;
293 	}
294 
295 	if (ppd->sdma_descq) {
296 		dma_free_coherent(&dd->pcidev->dev,
297 				  ppd->sdma_descq_cnt * sizeof(u64[2]),
298 				  ppd->sdma_descq, ppd->sdma_descq_phys);
299 		ppd->sdma_descq = NULL;
300 		ppd->sdma_descq_phys = 0;
301 	}
302 }
303 
304 static inline void make_sdma_desc(struct qib_pportdata *ppd,
305 				  u64 *sdmadesc, u64 addr, u64 dwlen,
306 				  u64 dwoffset)
307 {
308 
309 	WARN_ON(addr & 3);
310 	/* SDmaPhyAddr[47:32] */
311 	sdmadesc[1] = addr >> 32;
312 	/* SDmaPhyAddr[31:0] */
313 	sdmadesc[0] = (addr & 0xfffffffcULL) << 32;
314 	/* SDmaGeneration[1:0] */
315 	sdmadesc[0] |= (ppd->sdma_generation & 3ULL) <<
316 		SDMA_DESC_GEN_LSB;
317 	/* SDmaDwordCount[10:0] */
318 	sdmadesc[0] |= (dwlen & 0x7ffULL) << SDMA_DESC_COUNT_LSB;
319 	/* SDmaBufOffset[12:2] */
320 	sdmadesc[0] |= dwoffset & 0x7ffULL;
321 }
322 
323 /* sdma_lock must be held */
324 int qib_sdma_make_progress(struct qib_pportdata *ppd)
325 {
326 	struct list_head *lp = NULL;
327 	struct qib_sdma_txreq *txp = NULL;
328 	struct qib_devdata *dd = ppd->dd;
329 	int progress = 0;
330 	u16 hwhead;
331 	u16 idx = 0;
332 
333 	hwhead = dd->f_sdma_gethead(ppd);
334 
335 	/* The reason for some of the complexity of this code is that
336 	 * not all descriptors have corresponding txps.  So, we have to
337 	 * be able to skip over descs until we wander into the range of
338 	 * the next txp on the list.
339 	 */
340 
341 	if (!list_empty(&ppd->sdma_activelist)) {
342 		lp = ppd->sdma_activelist.next;
343 		txp = list_entry(lp, struct qib_sdma_txreq, list);
344 		idx = txp->start_idx;
345 	}
346 
347 	while (ppd->sdma_descq_head != hwhead) {
348 		/* if desc is part of this txp, unmap if needed */
349 		if (txp && (txp->flags & QIB_SDMA_TXREQ_F_FREEDESC) &&
350 		    (idx == ppd->sdma_descq_head)) {
351 			unmap_desc(ppd, ppd->sdma_descq_head);
352 			if (++idx == ppd->sdma_descq_cnt)
353 				idx = 0;
354 		}
355 
356 		/* increment dequed desc count */
357 		ppd->sdma_descq_removed++;
358 
359 		/* advance head, wrap if needed */
360 		if (++ppd->sdma_descq_head == ppd->sdma_descq_cnt)
361 			ppd->sdma_descq_head = 0;
362 
363 		/* if now past this txp's descs, do the callback */
364 		if (txp && txp->next_descq_idx == ppd->sdma_descq_head) {
365 			/* remove from active list */
366 			list_del_init(&txp->list);
367 			if (txp->callback)
368 				(*txp->callback)(txp, QIB_SDMA_TXREQ_S_OK);
369 			/* see if there is another txp */
370 			if (list_empty(&ppd->sdma_activelist))
371 				txp = NULL;
372 			else {
373 				lp = ppd->sdma_activelist.next;
374 				txp = list_entry(lp, struct qib_sdma_txreq,
375 					list);
376 				idx = txp->start_idx;
377 			}
378 		}
379 		progress = 1;
380 	}
381 	if (progress)
382 		qib_verbs_sdma_desc_avail(ppd, qib_sdma_descq_freecnt(ppd));
383 	return progress;
384 }
385 
386 /*
387  * This is called from interrupt context.
388  */
389 void qib_sdma_intr(struct qib_pportdata *ppd)
390 {
391 	unsigned long flags;
392 
393 	spin_lock_irqsave(&ppd->sdma_lock, flags);
394 
395 	__qib_sdma_intr(ppd);
396 
397 	spin_unlock_irqrestore(&ppd->sdma_lock, flags);
398 }
399 
400 void __qib_sdma_intr(struct qib_pportdata *ppd)
401 {
402 	if (__qib_sdma_running(ppd)) {
403 		qib_sdma_make_progress(ppd);
404 		if (!list_empty(&ppd->sdma_userpending))
405 			qib_user_sdma_send_desc(ppd, &ppd->sdma_userpending);
406 	}
407 }
408 
409 int qib_setup_sdma(struct qib_pportdata *ppd)
410 {
411 	struct qib_devdata *dd = ppd->dd;
412 	unsigned long flags;
413 	int ret = 0;
414 
415 	ret = alloc_sdma(ppd);
416 	if (ret)
417 		goto bail;
418 
419 	/* set consistent sdma state */
420 	ppd->dd->f_sdma_init_early(ppd);
421 	spin_lock_irqsave(&ppd->sdma_lock, flags);
422 	sdma_set_state(ppd, qib_sdma_state_s00_hw_down);
423 	spin_unlock_irqrestore(&ppd->sdma_lock, flags);
424 
425 	/* set up reference counting */
426 	kref_init(&ppd->sdma_state.kref);
427 	init_completion(&ppd->sdma_state.comp);
428 
429 	ppd->sdma_generation = 0;
430 	ppd->sdma_descq_head = 0;
431 	ppd->sdma_descq_removed = 0;
432 	ppd->sdma_descq_added = 0;
433 
434 	ppd->sdma_intrequest = 0;
435 	INIT_LIST_HEAD(&ppd->sdma_userpending);
436 
437 	INIT_LIST_HEAD(&ppd->sdma_activelist);
438 
439 	tasklet_init(&ppd->sdma_sw_clean_up_task, sdma_sw_clean_up_task,
440 		(unsigned long)ppd);
441 
442 	ret = dd->f_init_sdma_regs(ppd);
443 	if (ret)
444 		goto bail_alloc;
445 
446 	qib_sdma_process_event(ppd, qib_sdma_event_e10_go_hw_start);
447 
448 	return 0;
449 
450 bail_alloc:
451 	qib_teardown_sdma(ppd);
452 bail:
453 	return ret;
454 }
455 
456 void qib_teardown_sdma(struct qib_pportdata *ppd)
457 {
458 	qib_sdma_process_event(ppd, qib_sdma_event_e00_go_hw_down);
459 
460 	/*
461 	 * This waits for the state machine to exit so it is not
462 	 * necessary to kill the sdma_sw_clean_up_task to make sure
463 	 * it is not running.
464 	 */
465 	sdma_finalput(&ppd->sdma_state);
466 
467 	free_sdma(ppd);
468 }
469 
470 int qib_sdma_running(struct qib_pportdata *ppd)
471 {
472 	unsigned long flags;
473 	int ret;
474 
475 	spin_lock_irqsave(&ppd->sdma_lock, flags);
476 	ret = __qib_sdma_running(ppd);
477 	spin_unlock_irqrestore(&ppd->sdma_lock, flags);
478 
479 	return ret;
480 }
481 
482 /*
483  * Complete a request when sdma not running; likely only request
484  * but to simplify the code, always queue it, then process the full
485  * activelist.  We process the entire list to ensure that this particular
486  * request does get it's callback, but in the correct order.
487  * Must be called with sdma_lock held
488  */
489 static void complete_sdma_err_req(struct qib_pportdata *ppd,
490 				  struct qib_verbs_txreq *tx)
491 {
492 	struct qib_qp_priv *priv = tx->qp->priv;
493 
494 	atomic_inc(&priv->s_dma_busy);
495 	/* no sdma descriptors, so no unmap_desc */
496 	tx->txreq.start_idx = 0;
497 	tx->txreq.next_descq_idx = 0;
498 	list_add_tail(&tx->txreq.list, &ppd->sdma_activelist);
499 	clear_sdma_activelist(ppd);
500 }
501 
502 /*
503  * This function queues one IB packet onto the send DMA queue per call.
504  * The caller is responsible for checking:
505  * 1) The number of send DMA descriptor entries is less than the size of
506  *    the descriptor queue.
507  * 2) The IB SGE addresses and lengths are 32-bit aligned
508  *    (except possibly the last SGE's length)
509  * 3) The SGE addresses are suitable for passing to dma_map_single().
510  */
511 int qib_sdma_verbs_send(struct qib_pportdata *ppd,
512 			struct rvt_sge_state *ss, u32 dwords,
513 			struct qib_verbs_txreq *tx)
514 {
515 	unsigned long flags;
516 	struct rvt_sge *sge;
517 	struct rvt_qp *qp;
518 	int ret = 0;
519 	u16 tail;
520 	__le64 *descqp;
521 	u64 sdmadesc[2];
522 	u32 dwoffset;
523 	dma_addr_t addr;
524 	struct qib_qp_priv *priv;
525 
526 	spin_lock_irqsave(&ppd->sdma_lock, flags);
527 
528 retry:
529 	if (unlikely(!__qib_sdma_running(ppd))) {
530 		complete_sdma_err_req(ppd, tx);
531 		goto unlock;
532 	}
533 
534 	if (tx->txreq.sg_count > qib_sdma_descq_freecnt(ppd)) {
535 		if (qib_sdma_make_progress(ppd))
536 			goto retry;
537 		if (ppd->dd->flags & QIB_HAS_SDMA_TIMEOUT)
538 			ppd->dd->f_sdma_set_desc_cnt(ppd,
539 					ppd->sdma_descq_cnt / 2);
540 		goto busy;
541 	}
542 
543 	dwoffset = tx->hdr_dwords;
544 	make_sdma_desc(ppd, sdmadesc, (u64) tx->txreq.addr, dwoffset, 0);
545 
546 	sdmadesc[0] |= SDMA_DESC_FIRST;
547 	if (tx->txreq.flags & QIB_SDMA_TXREQ_F_USELARGEBUF)
548 		sdmadesc[0] |= SDMA_DESC_USE_LARGE_BUF;
549 
550 	/* write to the descq */
551 	tail = ppd->sdma_descq_tail;
552 	descqp = &ppd->sdma_descq[tail].qw[0];
553 	*descqp++ = cpu_to_le64(sdmadesc[0]);
554 	*descqp++ = cpu_to_le64(sdmadesc[1]);
555 
556 	/* increment the tail */
557 	if (++tail == ppd->sdma_descq_cnt) {
558 		tail = 0;
559 		descqp = &ppd->sdma_descq[0].qw[0];
560 		++ppd->sdma_generation;
561 	}
562 
563 	tx->txreq.start_idx = tail;
564 
565 	sge = &ss->sge;
566 	while (dwords) {
567 		u32 dw;
568 		u32 len;
569 
570 		len = dwords << 2;
571 		if (len > sge->length)
572 			len = sge->length;
573 		if (len > sge->sge_length)
574 			len = sge->sge_length;
575 		dw = (len + 3) >> 2;
576 		addr = dma_map_single(&ppd->dd->pcidev->dev, sge->vaddr,
577 				      dw << 2, DMA_TO_DEVICE);
578 		if (dma_mapping_error(&ppd->dd->pcidev->dev, addr)) {
579 			ret = -ENOMEM;
580 			goto unmap;
581 		}
582 		sdmadesc[0] = 0;
583 		make_sdma_desc(ppd, sdmadesc, (u64) addr, dw, dwoffset);
584 		/* SDmaUseLargeBuf has to be set in every descriptor */
585 		if (tx->txreq.flags & QIB_SDMA_TXREQ_F_USELARGEBUF)
586 			sdmadesc[0] |= SDMA_DESC_USE_LARGE_BUF;
587 		/* write to the descq */
588 		*descqp++ = cpu_to_le64(sdmadesc[0]);
589 		*descqp++ = cpu_to_le64(sdmadesc[1]);
590 
591 		/* increment the tail */
592 		if (++tail == ppd->sdma_descq_cnt) {
593 			tail = 0;
594 			descqp = &ppd->sdma_descq[0].qw[0];
595 			++ppd->sdma_generation;
596 		}
597 		sge->vaddr += len;
598 		sge->length -= len;
599 		sge->sge_length -= len;
600 		if (sge->sge_length == 0) {
601 			if (--ss->num_sge)
602 				*sge = *ss->sg_list++;
603 		} else if (sge->length == 0 && sge->mr->lkey) {
604 			if (++sge->n >= RVT_SEGSZ) {
605 				if (++sge->m >= sge->mr->mapsz)
606 					break;
607 				sge->n = 0;
608 			}
609 			sge->vaddr =
610 				sge->mr->map[sge->m]->segs[sge->n].vaddr;
611 			sge->length =
612 				sge->mr->map[sge->m]->segs[sge->n].length;
613 		}
614 
615 		dwoffset += dw;
616 		dwords -= dw;
617 	}
618 
619 	if (!tail)
620 		descqp = &ppd->sdma_descq[ppd->sdma_descq_cnt].qw[0];
621 	descqp -= 2;
622 	descqp[0] |= cpu_to_le64(SDMA_DESC_LAST);
623 	if (tx->txreq.flags & QIB_SDMA_TXREQ_F_HEADTOHOST)
624 		descqp[0] |= cpu_to_le64(SDMA_DESC_DMA_HEAD);
625 	if (tx->txreq.flags & QIB_SDMA_TXREQ_F_INTREQ)
626 		descqp[0] |= cpu_to_le64(SDMA_DESC_INTR);
627 	priv = tx->qp->priv;
628 	atomic_inc(&priv->s_dma_busy);
629 	tx->txreq.next_descq_idx = tail;
630 	ppd->dd->f_sdma_update_tail(ppd, tail);
631 	ppd->sdma_descq_added += tx->txreq.sg_count;
632 	list_add_tail(&tx->txreq.list, &ppd->sdma_activelist);
633 	goto unlock;
634 
635 unmap:
636 	for (;;) {
637 		if (!tail)
638 			tail = ppd->sdma_descq_cnt - 1;
639 		else
640 			tail--;
641 		if (tail == ppd->sdma_descq_tail)
642 			break;
643 		unmap_desc(ppd, tail);
644 	}
645 	qp = tx->qp;
646 	priv = qp->priv;
647 	qib_put_txreq(tx);
648 	spin_lock(&qp->r_lock);
649 	spin_lock(&qp->s_lock);
650 	if (qp->ibqp.qp_type == IB_QPT_RC) {
651 		/* XXX what about error sending RDMA read responses? */
652 		if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)
653 			rvt_error_qp(qp, IB_WC_GENERAL_ERR);
654 	} else if (qp->s_wqe)
655 		rvt_send_complete(qp, qp->s_wqe, IB_WC_GENERAL_ERR);
656 	spin_unlock(&qp->s_lock);
657 	spin_unlock(&qp->r_lock);
658 	/* return zero to process the next send work request */
659 	goto unlock;
660 
661 busy:
662 	qp = tx->qp;
663 	priv = qp->priv;
664 	spin_lock(&qp->s_lock);
665 	if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
666 		struct qib_ibdev *dev;
667 
668 		/*
669 		 * If we couldn't queue the DMA request, save the info
670 		 * and try again later rather than destroying the
671 		 * buffer and undoing the side effects of the copy.
672 		 */
673 		tx->ss = ss;
674 		tx->dwords = dwords;
675 		priv->s_tx = tx;
676 		dev = &ppd->dd->verbs_dev;
677 		spin_lock(&dev->rdi.pending_lock);
678 		if (list_empty(&priv->iowait)) {
679 			struct qib_ibport *ibp;
680 
681 			ibp = &ppd->ibport_data;
682 			ibp->rvp.n_dmawait++;
683 			qp->s_flags |= RVT_S_WAIT_DMA_DESC;
684 			list_add_tail(&priv->iowait, &dev->dmawait);
685 		}
686 		spin_unlock(&dev->rdi.pending_lock);
687 		qp->s_flags &= ~RVT_S_BUSY;
688 		spin_unlock(&qp->s_lock);
689 		ret = -EBUSY;
690 	} else {
691 		spin_unlock(&qp->s_lock);
692 		qib_put_txreq(tx);
693 	}
694 unlock:
695 	spin_unlock_irqrestore(&ppd->sdma_lock, flags);
696 	return ret;
697 }
698 
699 /*
700  * sdma_lock should be acquired before calling this routine
701  */
702 void dump_sdma_state(struct qib_pportdata *ppd)
703 {
704 	struct qib_sdma_desc *descq;
705 	struct qib_sdma_txreq *txp, *txpnext;
706 	__le64 *descqp;
707 	u64 desc[2];
708 	u64 addr;
709 	u16 gen, dwlen, dwoffset;
710 	u16 head, tail, cnt;
711 
712 	head = ppd->sdma_descq_head;
713 	tail = ppd->sdma_descq_tail;
714 	cnt = qib_sdma_descq_freecnt(ppd);
715 	descq = ppd->sdma_descq;
716 
717 	qib_dev_porterr(ppd->dd, ppd->port,
718 		"SDMA ppd->sdma_descq_head: %u\n", head);
719 	qib_dev_porterr(ppd->dd, ppd->port,
720 		"SDMA ppd->sdma_descq_tail: %u\n", tail);
721 	qib_dev_porterr(ppd->dd, ppd->port,
722 		"SDMA sdma_descq_freecnt: %u\n", cnt);
723 
724 	/* print info for each entry in the descriptor queue */
725 	while (head != tail) {
726 		char flags[6] = { 'x', 'x', 'x', 'x', 'x', 0 };
727 
728 		descqp = &descq[head].qw[0];
729 		desc[0] = le64_to_cpu(descqp[0]);
730 		desc[1] = le64_to_cpu(descqp[1]);
731 		flags[0] = (desc[0] & 1<<15) ? 'I' : '-';
732 		flags[1] = (desc[0] & 1<<14) ? 'L' : 'S';
733 		flags[2] = (desc[0] & 1<<13) ? 'H' : '-';
734 		flags[3] = (desc[0] & 1<<12) ? 'F' : '-';
735 		flags[4] = (desc[0] & 1<<11) ? 'L' : '-';
736 		addr = (desc[1] << 32) | ((desc[0] >> 32) & 0xfffffffcULL);
737 		gen = (desc[0] >> 30) & 3ULL;
738 		dwlen = (desc[0] >> 14) & (0x7ffULL << 2);
739 		dwoffset = (desc[0] & 0x7ffULL) << 2;
740 		qib_dev_porterr(ppd->dd, ppd->port,
741 			"SDMA sdmadesc[%u]: flags:%s addr:0x%016llx gen:%u len:%u bytes offset:%u bytes\n",
742 			 head, flags, addr, gen, dwlen, dwoffset);
743 		if (++head == ppd->sdma_descq_cnt)
744 			head = 0;
745 	}
746 
747 	/* print dma descriptor indices from the TX requests */
748 	list_for_each_entry_safe(txp, txpnext, &ppd->sdma_activelist,
749 				 list)
750 		qib_dev_porterr(ppd->dd, ppd->port,
751 			"SDMA txp->start_idx: %u txp->next_descq_idx: %u\n",
752 			txp->start_idx, txp->next_descq_idx);
753 }
754 
755 void qib_sdma_process_event(struct qib_pportdata *ppd,
756 	enum qib_sdma_events event)
757 {
758 	unsigned long flags;
759 
760 	spin_lock_irqsave(&ppd->sdma_lock, flags);
761 
762 	__qib_sdma_process_event(ppd, event);
763 
764 	if (ppd->sdma_state.current_state == qib_sdma_state_s99_running)
765 		qib_verbs_sdma_desc_avail(ppd, qib_sdma_descq_freecnt(ppd));
766 
767 	spin_unlock_irqrestore(&ppd->sdma_lock, flags);
768 }
769 
770 void __qib_sdma_process_event(struct qib_pportdata *ppd,
771 	enum qib_sdma_events event)
772 {
773 	struct qib_sdma_state *ss = &ppd->sdma_state;
774 
775 	switch (ss->current_state) {
776 	case qib_sdma_state_s00_hw_down:
777 		switch (event) {
778 		case qib_sdma_event_e00_go_hw_down:
779 			break;
780 		case qib_sdma_event_e30_go_running:
781 			/*
782 			 * If down, but running requested (usually result
783 			 * of link up, then we need to start up.
784 			 * This can happen when hw down is requested while
785 			 * bringing the link up with traffic active on
786 			 * 7220, e.g. */
787 			ss->go_s99_running = 1;
788 			/* fall through -- and start dma engine */
789 		case qib_sdma_event_e10_go_hw_start:
790 			/* This reference means the state machine is started */
791 			sdma_get(&ppd->sdma_state);
792 			sdma_set_state(ppd,
793 				       qib_sdma_state_s10_hw_start_up_wait);
794 			break;
795 		case qib_sdma_event_e20_hw_started:
796 			break;
797 		case qib_sdma_event_e40_sw_cleaned:
798 			sdma_sw_tear_down(ppd);
799 			break;
800 		case qib_sdma_event_e50_hw_cleaned:
801 			break;
802 		case qib_sdma_event_e60_hw_halted:
803 			break;
804 		case qib_sdma_event_e70_go_idle:
805 			break;
806 		case qib_sdma_event_e7220_err_halted:
807 			break;
808 		case qib_sdma_event_e7322_err_halted:
809 			break;
810 		case qib_sdma_event_e90_timer_tick:
811 			break;
812 		}
813 		break;
814 
815 	case qib_sdma_state_s10_hw_start_up_wait:
816 		switch (event) {
817 		case qib_sdma_event_e00_go_hw_down:
818 			sdma_set_state(ppd, qib_sdma_state_s00_hw_down);
819 			sdma_sw_tear_down(ppd);
820 			break;
821 		case qib_sdma_event_e10_go_hw_start:
822 			break;
823 		case qib_sdma_event_e20_hw_started:
824 			sdma_set_state(ppd, ss->go_s99_running ?
825 				       qib_sdma_state_s99_running :
826 				       qib_sdma_state_s20_idle);
827 			break;
828 		case qib_sdma_event_e30_go_running:
829 			ss->go_s99_running = 1;
830 			break;
831 		case qib_sdma_event_e40_sw_cleaned:
832 			break;
833 		case qib_sdma_event_e50_hw_cleaned:
834 			break;
835 		case qib_sdma_event_e60_hw_halted:
836 			break;
837 		case qib_sdma_event_e70_go_idle:
838 			ss->go_s99_running = 0;
839 			break;
840 		case qib_sdma_event_e7220_err_halted:
841 			break;
842 		case qib_sdma_event_e7322_err_halted:
843 			break;
844 		case qib_sdma_event_e90_timer_tick:
845 			break;
846 		}
847 		break;
848 
849 	case qib_sdma_state_s20_idle:
850 		switch (event) {
851 		case qib_sdma_event_e00_go_hw_down:
852 			sdma_set_state(ppd, qib_sdma_state_s00_hw_down);
853 			sdma_sw_tear_down(ppd);
854 			break;
855 		case qib_sdma_event_e10_go_hw_start:
856 			break;
857 		case qib_sdma_event_e20_hw_started:
858 			break;
859 		case qib_sdma_event_e30_go_running:
860 			sdma_set_state(ppd, qib_sdma_state_s99_running);
861 			ss->go_s99_running = 1;
862 			break;
863 		case qib_sdma_event_e40_sw_cleaned:
864 			break;
865 		case qib_sdma_event_e50_hw_cleaned:
866 			break;
867 		case qib_sdma_event_e60_hw_halted:
868 			break;
869 		case qib_sdma_event_e70_go_idle:
870 			break;
871 		case qib_sdma_event_e7220_err_halted:
872 			break;
873 		case qib_sdma_event_e7322_err_halted:
874 			break;
875 		case qib_sdma_event_e90_timer_tick:
876 			break;
877 		}
878 		break;
879 
880 	case qib_sdma_state_s30_sw_clean_up_wait:
881 		switch (event) {
882 		case qib_sdma_event_e00_go_hw_down:
883 			sdma_set_state(ppd, qib_sdma_state_s00_hw_down);
884 			break;
885 		case qib_sdma_event_e10_go_hw_start:
886 			break;
887 		case qib_sdma_event_e20_hw_started:
888 			break;
889 		case qib_sdma_event_e30_go_running:
890 			ss->go_s99_running = 1;
891 			break;
892 		case qib_sdma_event_e40_sw_cleaned:
893 			sdma_set_state(ppd,
894 				       qib_sdma_state_s10_hw_start_up_wait);
895 			sdma_hw_start_up(ppd);
896 			break;
897 		case qib_sdma_event_e50_hw_cleaned:
898 			break;
899 		case qib_sdma_event_e60_hw_halted:
900 			break;
901 		case qib_sdma_event_e70_go_idle:
902 			ss->go_s99_running = 0;
903 			break;
904 		case qib_sdma_event_e7220_err_halted:
905 			break;
906 		case qib_sdma_event_e7322_err_halted:
907 			break;
908 		case qib_sdma_event_e90_timer_tick:
909 			break;
910 		}
911 		break;
912 
913 	case qib_sdma_state_s40_hw_clean_up_wait:
914 		switch (event) {
915 		case qib_sdma_event_e00_go_hw_down:
916 			sdma_set_state(ppd, qib_sdma_state_s00_hw_down);
917 			sdma_start_sw_clean_up(ppd);
918 			break;
919 		case qib_sdma_event_e10_go_hw_start:
920 			break;
921 		case qib_sdma_event_e20_hw_started:
922 			break;
923 		case qib_sdma_event_e30_go_running:
924 			ss->go_s99_running = 1;
925 			break;
926 		case qib_sdma_event_e40_sw_cleaned:
927 			break;
928 		case qib_sdma_event_e50_hw_cleaned:
929 			sdma_set_state(ppd,
930 				       qib_sdma_state_s30_sw_clean_up_wait);
931 			sdma_start_sw_clean_up(ppd);
932 			break;
933 		case qib_sdma_event_e60_hw_halted:
934 			break;
935 		case qib_sdma_event_e70_go_idle:
936 			ss->go_s99_running = 0;
937 			break;
938 		case qib_sdma_event_e7220_err_halted:
939 			break;
940 		case qib_sdma_event_e7322_err_halted:
941 			break;
942 		case qib_sdma_event_e90_timer_tick:
943 			break;
944 		}
945 		break;
946 
947 	case qib_sdma_state_s50_hw_halt_wait:
948 		switch (event) {
949 		case qib_sdma_event_e00_go_hw_down:
950 			sdma_set_state(ppd, qib_sdma_state_s00_hw_down);
951 			sdma_start_sw_clean_up(ppd);
952 			break;
953 		case qib_sdma_event_e10_go_hw_start:
954 			break;
955 		case qib_sdma_event_e20_hw_started:
956 			break;
957 		case qib_sdma_event_e30_go_running:
958 			ss->go_s99_running = 1;
959 			break;
960 		case qib_sdma_event_e40_sw_cleaned:
961 			break;
962 		case qib_sdma_event_e50_hw_cleaned:
963 			break;
964 		case qib_sdma_event_e60_hw_halted:
965 			sdma_set_state(ppd,
966 				       qib_sdma_state_s40_hw_clean_up_wait);
967 			ppd->dd->f_sdma_hw_clean_up(ppd);
968 			break;
969 		case qib_sdma_event_e70_go_idle:
970 			ss->go_s99_running = 0;
971 			break;
972 		case qib_sdma_event_e7220_err_halted:
973 			break;
974 		case qib_sdma_event_e7322_err_halted:
975 			break;
976 		case qib_sdma_event_e90_timer_tick:
977 			break;
978 		}
979 		break;
980 
981 	case qib_sdma_state_s99_running:
982 		switch (event) {
983 		case qib_sdma_event_e00_go_hw_down:
984 			sdma_set_state(ppd, qib_sdma_state_s00_hw_down);
985 			sdma_start_sw_clean_up(ppd);
986 			break;
987 		case qib_sdma_event_e10_go_hw_start:
988 			break;
989 		case qib_sdma_event_e20_hw_started:
990 			break;
991 		case qib_sdma_event_e30_go_running:
992 			break;
993 		case qib_sdma_event_e40_sw_cleaned:
994 			break;
995 		case qib_sdma_event_e50_hw_cleaned:
996 			break;
997 		case qib_sdma_event_e60_hw_halted:
998 			sdma_set_state(ppd,
999 				       qib_sdma_state_s30_sw_clean_up_wait);
1000 			sdma_start_sw_clean_up(ppd);
1001 			break;
1002 		case qib_sdma_event_e70_go_idle:
1003 			sdma_set_state(ppd, qib_sdma_state_s50_hw_halt_wait);
1004 			ss->go_s99_running = 0;
1005 			break;
1006 		case qib_sdma_event_e7220_err_halted:
1007 			sdma_set_state(ppd,
1008 				       qib_sdma_state_s30_sw_clean_up_wait);
1009 			sdma_start_sw_clean_up(ppd);
1010 			break;
1011 		case qib_sdma_event_e7322_err_halted:
1012 			sdma_set_state(ppd, qib_sdma_state_s50_hw_halt_wait);
1013 			break;
1014 		case qib_sdma_event_e90_timer_tick:
1015 			break;
1016 		}
1017 		break;
1018 	}
1019 
1020 	ss->last_event = event;
1021 }
1022