?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
mm, page_alloc: set alloc_flags only once in slowpath
[linux/fpc-iii.git] / drivers / block / rsxx / dma.c
blob5a20385f87d045af1704205dea18b0aa9e7a1260
1 /*
2 * Filename: dma.c
3 *
4 *
5 * Authors: Joshua Morris <josh.h.morris@us.ibm.com>
6 * Philip Kelleher <pjk1939@linux.vnet.ibm.com>
7 *
8 * (C) Copyright 2013 IBM Corporation
9 *
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public License as
12 * published by the Free Software Foundation; either version 2 of the
13 * License, or (at your option) any later version.
14 *
15 * This program is distributed in the hope that it will be useful, but
16 * WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
19 *
20 * You should have received a copy of the GNU General Public License
21 * along with this program; if not, write to the Free Software Foundation,
22 * Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
23 */
24
25 #include <linux/slab.h>
26 #include "rsxx_priv.h"
27
28 struct rsxx_dma {
29 struct list_head list;
30 u8 cmd;
31 unsigned int laddr; /* Logical address */
32 struct {
33 u32 off;
34 u32 cnt;
35 } sub_page;
36 dma_addr_t dma_addr;
37 struct page *page;
38 unsigned int pg_off; /* Page Offset */
39 rsxx_dma_cb cb;
40 void *cb_data;
41 };
42
43 /* This timeout is used to detect a stalled DMA channel */
44 #define DMA_ACTIVITY_TIMEOUT msecs_to_jiffies(10000)
45
46 struct hw_status {
47 u8 status;
48 u8 tag;
49 __le16 count;
50 __le32 _rsvd2;
51 __le64 _rsvd3;
52 } __packed;
53
54 enum rsxx_dma_status {
55 DMA_SW_ERR = 0x1,
56 DMA_HW_FAULT = 0x2,
57 DMA_CANCELLED = 0x4,
58 };
59
60 struct hw_cmd {
61 u8 command;
62 u8 tag;
63 u8 _rsvd;
64 u8 sub_page; /* Bit[0:2]: 512byte offset */
65 /* Bit[4:6]: 512byte count */
66 __le32 device_addr;
67 __le64 host_addr;
68 } __packed;
69
70 enum rsxx_hw_cmd {
71 HW_CMD_BLK_DISCARD = 0x70,
72 HW_CMD_BLK_WRITE = 0x80,
73 HW_CMD_BLK_READ = 0xC0,
74 HW_CMD_BLK_RECON_READ = 0xE0,
75 };
76
77 enum rsxx_hw_status {
78 HW_STATUS_CRC = 0x01,
79 HW_STATUS_HARD_ERR = 0x02,
80 HW_STATUS_SOFT_ERR = 0x04,
81 HW_STATUS_FAULT = 0x08,
82 };
83
84 static struct kmem_cache *rsxx_dma_pool;
85
86 struct dma_tracker {
87 int next_tag;
88 struct rsxx_dma *dma;
89 };
90
91 #define DMA_TRACKER_LIST_SIZE8 (sizeof(struct dma_tracker_list) + \
92 (sizeof(struct dma_tracker) * RSXX_MAX_OUTSTANDING_CMDS))
93
94 struct dma_tracker_list {
95 spinlock_t lock;
96 int head;
97 struct dma_tracker list[0];
98 };
99
100
101 /*----------------- Misc Utility Functions -------------------*/
102 static unsigned int rsxx_addr8_to_laddr(u64 addr8, struct rsxx_cardinfo *card)
103 {
104 unsigned long long tgt_addr8;
105
106 tgt_addr8 = ((addr8 >> card->_stripe.upper_shift) &
107 card->_stripe.upper_mask) |
108 ((addr8) & card->_stripe.lower_mask);
109 do_div(tgt_addr8, RSXX_HW_BLK_SIZE);
110 return tgt_addr8;
111 }
112
113 static unsigned int rsxx_get_dma_tgt(struct rsxx_cardinfo *card, u64 addr8)
114 {
115 unsigned int tgt;
116
117 tgt = (addr8 >> card->_stripe.target_shift) & card->_stripe.target_mask;
118
119 return tgt;
120 }
121
122 void rsxx_dma_queue_reset(struct rsxx_cardinfo *card)
123 {
124 /* Reset all DMA Command/Status Queues */
125 iowrite32(DMA_QUEUE_RESET, card->regmap + RESET);
126 }
127
128 static unsigned int get_dma_size(struct rsxx_dma *dma)
129 {
130 if (dma->sub_page.cnt)
131 return dma->sub_page.cnt << 9;
132 else
133 return RSXX_HW_BLK_SIZE;
134 }
135
136
137 /*----------------- DMA Tracker -------------------*/
138 static void set_tracker_dma(struct dma_tracker_list *trackers,
139 int tag,
140 struct rsxx_dma *dma)
141 {
142 trackers->list[tag].dma = dma;
143 }
144
145 static struct rsxx_dma *get_tracker_dma(struct dma_tracker_list *trackers,
146 int tag)
147 {
148 return trackers->list[tag].dma;
149 }
150
151 static int pop_tracker(struct dma_tracker_list *trackers)
152 {
153 int tag;
154
155 spin_lock(&trackers->lock);
156 tag = trackers->head;
157 if (tag != -1) {
158 trackers->head = trackers->list[tag].next_tag;
159 trackers->list[tag].next_tag = -1;
160 }
161 spin_unlock(&trackers->lock);
162
163 return tag;
164 }
165
166 static void push_tracker(struct dma_tracker_list *trackers, int tag)
167 {
168 spin_lock(&trackers->lock);
169 trackers->list[tag].next_tag = trackers->head;
170 trackers->head = tag;
171 trackers->list[tag].dma = NULL;
172 spin_unlock(&trackers->lock);
173 }
174
175
176 /*----------------- Interrupt Coalescing -------------*/
177 /*
178 * Interrupt Coalescing Register Format:
179 * Interrupt Timer (64ns units) [15:0]
180 * Interrupt Count [24:16]
181 * Reserved [31:25]
182 */
183 #define INTR_COAL_LATENCY_MASK (0x0000ffff)
184
185 #define INTR_COAL_COUNT_SHIFT 16
186 #define INTR_COAL_COUNT_BITS 9
187 #define INTR_COAL_COUNT_MASK (((1 << INTR_COAL_COUNT_BITS) - 1) << \
188 INTR_COAL_COUNT_SHIFT)
189 #define INTR_COAL_LATENCY_UNITS_NS 64
190
191
192 static u32 dma_intr_coal_val(u32 mode, u32 count, u32 latency)
193 {
194 u32 latency_units = latency / INTR_COAL_LATENCY_UNITS_NS;
195
196 if (mode == RSXX_INTR_COAL_DISABLED)
197 return 0;
198
199 return ((count << INTR_COAL_COUNT_SHIFT) & INTR_COAL_COUNT_MASK) |
200 (latency_units & INTR_COAL_LATENCY_MASK);
201
202 }
203
204 static void dma_intr_coal_auto_tune(struct rsxx_cardinfo *card)
205 {
206 int i;
207 u32 q_depth = 0;
208 u32 intr_coal;
209
210 if (card->config.data.intr_coal.mode != RSXX_INTR_COAL_AUTO_TUNE ||
211 unlikely(card->eeh_state))
212 return;
213
214 for (i = 0; i < card->n_targets; i++)
215 q_depth += atomic_read(&card->ctrl[i].stats.hw_q_depth);
216
217 intr_coal = dma_intr_coal_val(card->config.data.intr_coal.mode,
218 q_depth / 2,
219 card->config.data.intr_coal.latency);
220 iowrite32(intr_coal, card->regmap + INTR_COAL);
221 }
222
223 /*----------------- RSXX DMA Handling -------------------*/
224 static void rsxx_free_dma(struct rsxx_dma_ctrl *ctrl, struct rsxx_dma *dma)
225 {
226 if (dma->cmd != HW_CMD_BLK_DISCARD) {
227 if (!pci_dma_mapping_error(ctrl->card->dev, dma->dma_addr)) {
228 pci_unmap_page(ctrl->card->dev, dma->dma_addr,
229 get_dma_size(dma),
230 dma->cmd == HW_CMD_BLK_WRITE ?
231 PCI_DMA_TODEVICE :
232 PCI_DMA_FROMDEVICE);
233 }
234 }
235
236 kmem_cache_free(rsxx_dma_pool, dma);
237 }
238
239 static void rsxx_complete_dma(struct rsxx_dma_ctrl *ctrl,
240 struct rsxx_dma *dma,
241 unsigned int status)
242 {
243 if (status & DMA_SW_ERR)
244 ctrl->stats.dma_sw_err++;
245 if (status & DMA_HW_FAULT)
246 ctrl->stats.dma_hw_fault++;
247 if (status & DMA_CANCELLED)
248 ctrl->stats.dma_cancelled++;
249
250 if (dma->cb)
251 dma->cb(ctrl->card, dma->cb_data, status ? 1 : 0);
252
253 rsxx_free_dma(ctrl, dma);
254 }
255
256 int rsxx_cleanup_dma_queue(struct rsxx_dma_ctrl *ctrl,
257 struct list_head *q, unsigned int done)
258 {
259 struct rsxx_dma *dma;
260 struct rsxx_dma *tmp;
261 int cnt = 0;
262
263 list_for_each_entry_safe(dma, tmp, q, list) {
264 list_del(&dma->list);
265 if (done & COMPLETE_DMA)
266 rsxx_complete_dma(ctrl, dma, DMA_CANCELLED);
267 else
268 rsxx_free_dma(ctrl, dma);
269 cnt++;
270 }
271
272 return cnt;
273 }
274
275 static void rsxx_requeue_dma(struct rsxx_dma_ctrl *ctrl,
276 struct rsxx_dma *dma)
277 {
278 /*
279 * Requeued DMAs go to the front of the queue so they are issued
280 * first.
281 */
282 spin_lock_bh(&ctrl->queue_lock);
283 ctrl->stats.sw_q_depth++;
284 list_add(&dma->list, &ctrl->queue);
285 spin_unlock_bh(&ctrl->queue_lock);
286 }
287
288 static void rsxx_handle_dma_error(struct rsxx_dma_ctrl *ctrl,
289 struct rsxx_dma *dma,
290 u8 hw_st)
291 {
292 unsigned int status = 0;
293 int requeue_cmd = 0;
294
295 dev_dbg(CARD_TO_DEV(ctrl->card),
296 "Handling DMA error(cmd x%02x, laddr x%08x st:x%02x)\n",
297 dma->cmd, dma->laddr, hw_st);
298
299 if (hw_st & HW_STATUS_CRC)
300 ctrl->stats.crc_errors++;
301 if (hw_st & HW_STATUS_HARD_ERR)
302 ctrl->stats.hard_errors++;
303 if (hw_st & HW_STATUS_SOFT_ERR)
304 ctrl->stats.soft_errors++;
305
306 switch (dma->cmd) {
307 case HW_CMD_BLK_READ:
308 if (hw_st & (HW_STATUS_CRC | HW_STATUS_HARD_ERR)) {
309 if (ctrl->card->scrub_hard) {
310 dma->cmd = HW_CMD_BLK_RECON_READ;
311 requeue_cmd = 1;
312 ctrl->stats.reads_retried++;
313 } else {
314 status |= DMA_HW_FAULT;
315 ctrl->stats.reads_failed++;
316 }
317 } else if (hw_st & HW_STATUS_FAULT) {
318 status |= DMA_HW_FAULT;
319 ctrl->stats.reads_failed++;
320 }
321
322 break;
323 case HW_CMD_BLK_RECON_READ:
324 if (hw_st & (HW_STATUS_CRC | HW_STATUS_HARD_ERR)) {
325 /* Data could not be reconstructed. */
326 status |= DMA_HW_FAULT;
327 ctrl->stats.reads_failed++;
328 }
329
330 break;
331 case HW_CMD_BLK_WRITE:
332 status |= DMA_HW_FAULT;
333 ctrl->stats.writes_failed++;
334
335 break;
336 case HW_CMD_BLK_DISCARD:
337 status |= DMA_HW_FAULT;
338 ctrl->stats.discards_failed++;
339
340 break;
341 default:
342 dev_err(CARD_TO_DEV(ctrl->card),
343 "Unknown command in DMA!(cmd: x%02x "
344 "laddr x%08x st: x%02x\n",
345 dma->cmd, dma->laddr, hw_st);
346 status |= DMA_SW_ERR;
347
348 break;
349 }
350
351 if (requeue_cmd)
352 rsxx_requeue_dma(ctrl, dma);
353 else
354 rsxx_complete_dma(ctrl, dma, status);
355 }
356
357 static void dma_engine_stalled(unsigned long data)
358 {
359 struct rsxx_dma_ctrl *ctrl = (struct rsxx_dma_ctrl *)data;
360 int cnt;
361
362 if (atomic_read(&ctrl->stats.hw_q_depth) == 0 ||
363 unlikely(ctrl->card->eeh_state))
364 return;
365
366 if (ctrl->cmd.idx != ioread32(ctrl->regmap + SW_CMD_IDX)) {
367 /*
368 * The dma engine was stalled because the SW_CMD_IDX write
369 * was lost. Issue it again to recover.
370 */
371 dev_warn(CARD_TO_DEV(ctrl->card),
372 "SW_CMD_IDX write was lost, re-writing...\n");
373 iowrite32(ctrl->cmd.idx, ctrl->regmap + SW_CMD_IDX);
374 mod_timer(&ctrl->activity_timer,
375 jiffies + DMA_ACTIVITY_TIMEOUT);
376 } else {
377 dev_warn(CARD_TO_DEV(ctrl->card),
378 "DMA channel %d has stalled, faulting interface.\n",
379 ctrl->id);
380 ctrl->card->dma_fault = 1;
381
382 /* Clean up the DMA queue */
383 spin_lock(&ctrl->queue_lock);
384 cnt = rsxx_cleanup_dma_queue(ctrl, &ctrl->queue, COMPLETE_DMA);
385 spin_unlock(&ctrl->queue_lock);
386
387 cnt += rsxx_dma_cancel(ctrl);
388
389 if (cnt)
390 dev_info(CARD_TO_DEV(ctrl->card),
391 "Freed %d queued DMAs on channel %d\n",
392 cnt, ctrl->id);
393 }
394 }
395
396 static void rsxx_issue_dmas(struct rsxx_dma_ctrl *ctrl)
397 {
398 struct rsxx_dma *dma;
399 int tag;
400 int cmds_pending = 0;
401 struct hw_cmd *hw_cmd_buf;
402 int dir;
403
404 hw_cmd_buf = ctrl->cmd.buf;
405
406 if (unlikely(ctrl->card->halt) ||
407 unlikely(ctrl->card->eeh_state))
408 return;
409
410 while (1) {
411 spin_lock_bh(&ctrl->queue_lock);
412 if (list_empty(&ctrl->queue)) {
413 spin_unlock_bh(&ctrl->queue_lock);
414 break;
415 }
416 spin_unlock_bh(&ctrl->queue_lock);
417
418 tag = pop_tracker(ctrl->trackers);
419 if (tag == -1)
420 break;
421
422 spin_lock_bh(&ctrl->queue_lock);
423 dma = list_entry(ctrl->queue.next, struct rsxx_dma, list);
424 list_del(&dma->list);
425 ctrl->stats.sw_q_depth--;
426 spin_unlock_bh(&ctrl->queue_lock);
427
428 /*
429 * This will catch any DMAs that slipped in right before the
430 * fault, but was queued after all the other DMAs were
431 * cancelled.
432 */
433 if (unlikely(ctrl->card->dma_fault)) {
434 push_tracker(ctrl->trackers, tag);
435 rsxx_complete_dma(ctrl, dma, DMA_CANCELLED);
436 continue;
437 }
438
439 if (dma->cmd != HW_CMD_BLK_DISCARD) {
440 if (dma->cmd == HW_CMD_BLK_WRITE)
441 dir = PCI_DMA_TODEVICE;
442 else
443 dir = PCI_DMA_FROMDEVICE;
444
445 /*
446 * The function pci_map_page is placed here because we
447 * can only, by design, issue up to 255 commands to the
448 * hardware at one time per DMA channel. So the maximum
449 * amount of mapped memory would be 255 * 4 channels *
450 * 4096 Bytes which is less than 2GB, the limit of a x8
451 * Non-HWWD PCIe slot. This way the pci_map_page
452 * function should never fail because of a lack of
453 * mappable memory.
454 */
455 dma->dma_addr = pci_map_page(ctrl->card->dev, dma->page,
456 dma->pg_off, dma->sub_page.cnt << 9, dir);
457 if (pci_dma_mapping_error(ctrl->card->dev, dma->dma_addr)) {
458 push_tracker(ctrl->trackers, tag);
459 rsxx_complete_dma(ctrl, dma, DMA_CANCELLED);
460 continue;
461 }
462 }
463
464 set_tracker_dma(ctrl->trackers, tag, dma);
465 hw_cmd_buf[ctrl->cmd.idx].command = dma->cmd;
466 hw_cmd_buf[ctrl->cmd.idx].tag = tag;
467 hw_cmd_buf[ctrl->cmd.idx]._rsvd = 0;
468 hw_cmd_buf[ctrl->cmd.idx].sub_page =
469 ((dma->sub_page.cnt & 0x7) << 4) |
470 (dma->sub_page.off & 0x7);
471
472 hw_cmd_buf[ctrl->cmd.idx].device_addr =
473 cpu_to_le32(dma->laddr);
474
475 hw_cmd_buf[ctrl->cmd.idx].host_addr =
476 cpu_to_le64(dma->dma_addr);
477
478 dev_dbg(CARD_TO_DEV(ctrl->card),
479 "Issue DMA%d(laddr %d tag %d) to idx %d\n",
480 ctrl->id, dma->laddr, tag, ctrl->cmd.idx);
481
482 ctrl->cmd.idx = (ctrl->cmd.idx + 1) & RSXX_CS_IDX_MASK;
483 cmds_pending++;
484
485 if (dma->cmd == HW_CMD_BLK_WRITE)
486 ctrl->stats.writes_issued++;
487 else if (dma->cmd == HW_CMD_BLK_DISCARD)
488 ctrl->stats.discards_issued++;
489 else
490 ctrl->stats.reads_issued++;
491 }
492
493 /* Let HW know we've queued commands. */
494 if (cmds_pending) {
495 atomic_add(cmds_pending, &ctrl->stats.hw_q_depth);
496 mod_timer(&ctrl->activity_timer,
497 jiffies + DMA_ACTIVITY_TIMEOUT);
498
499 if (unlikely(ctrl->card->eeh_state)) {
500 del_timer_sync(&ctrl->activity_timer);
501 return;
502 }
503
504 iowrite32(ctrl->cmd.idx, ctrl->regmap + SW_CMD_IDX);
505 }
506 }
507
508 static void rsxx_dma_done(struct rsxx_dma_ctrl *ctrl)
509 {
510 struct rsxx_dma *dma;
511 unsigned long flags;
512 u16 count;
513 u8 status;
514 u8 tag;
515 struct hw_status *hw_st_buf;
516
517 hw_st_buf = ctrl->status.buf;
518
519 if (unlikely(ctrl->card->halt) ||
520 unlikely(ctrl->card->dma_fault) ||
521 unlikely(ctrl->card->eeh_state))
522 return;
523
524 count = le16_to_cpu(hw_st_buf[ctrl->status.idx].count);
525
526 while (count == ctrl->e_cnt) {
527 /*
528 * The read memory-barrier is necessary to keep aggressive
529 * processors/optimizers (such as the PPC Apple G5) from
530 * reordering the following status-buffer tag & status read
531 * *before* the count read on subsequent iterations of the
532 * loop!
533 */
534 rmb();
535
536 status = hw_st_buf[ctrl->status.idx].status;
537 tag = hw_st_buf[ctrl->status.idx].tag;
538
539 dma = get_tracker_dma(ctrl->trackers, tag);
540 if (dma == NULL) {
541 spin_lock_irqsave(&ctrl->card->irq_lock, flags);
542 rsxx_disable_ier(ctrl->card, CR_INTR_DMA_ALL);
543 spin_unlock_irqrestore(&ctrl->card->irq_lock, flags);
544
545 dev_err(CARD_TO_DEV(ctrl->card),
546 "No tracker for tag %d "
547 "(idx %d id %d)\n",
548 tag, ctrl->status.idx, ctrl->id);
549 return;
550 }
551
552 dev_dbg(CARD_TO_DEV(ctrl->card),
553 "Completing DMA%d"
554 "(laddr x%x tag %d st: x%x cnt: x%04x) from idx %d.\n",
555 ctrl->id, dma->laddr, tag, status, count,
556 ctrl->status.idx);
557
558 atomic_dec(&ctrl->stats.hw_q_depth);
559
560 mod_timer(&ctrl->activity_timer,
561 jiffies + DMA_ACTIVITY_TIMEOUT);
562
563 if (status)
564 rsxx_handle_dma_error(ctrl, dma, status);
565 else
566 rsxx_complete_dma(ctrl, dma, 0);
567
568 push_tracker(ctrl->trackers, tag);
569
570 ctrl->status.idx = (ctrl->status.idx + 1) &
571 RSXX_CS_IDX_MASK;
572 ctrl->e_cnt++;
573
574 count = le16_to_cpu(hw_st_buf[ctrl->status.idx].count);
575 }
576
577 dma_intr_coal_auto_tune(ctrl->card);
578
579 if (atomic_read(&ctrl->stats.hw_q_depth) == 0)
580 del_timer_sync(&ctrl->activity_timer);
581
582 spin_lock_irqsave(&ctrl->card->irq_lock, flags);
583 rsxx_enable_ier(ctrl->card, CR_INTR_DMA(ctrl->id));
584 spin_unlock_irqrestore(&ctrl->card->irq_lock, flags);
585
586 spin_lock_bh(&ctrl->queue_lock);
587 if (ctrl->stats.sw_q_depth)
588 queue_work(ctrl->issue_wq, &ctrl->issue_dma_work);
589 spin_unlock_bh(&ctrl->queue_lock);
590 }
591
592 static void rsxx_schedule_issue(struct work_struct *work)
593 {
594 struct rsxx_dma_ctrl *ctrl;
595
596 ctrl = container_of(work, struct rsxx_dma_ctrl, issue_dma_work);
597
598 mutex_lock(&ctrl->work_lock);
599 rsxx_issue_dmas(ctrl);
600 mutex_unlock(&ctrl->work_lock);
601 }
602
603 static void rsxx_schedule_done(struct work_struct *work)
604 {
605 struct rsxx_dma_ctrl *ctrl;
606
607 ctrl = container_of(work, struct rsxx_dma_ctrl, dma_done_work);
608
609 mutex_lock(&ctrl->work_lock);
610 rsxx_dma_done(ctrl);
611 mutex_unlock(&ctrl->work_lock);
612 }
613
614 static int rsxx_queue_discard(struct rsxx_cardinfo *card,
615 struct list_head *q,
616 unsigned int laddr,
617 rsxx_dma_cb cb,
618 void *cb_data)
619 {
620 struct rsxx_dma *dma;
621
622 dma = kmem_cache_alloc(rsxx_dma_pool, GFP_KERNEL);
623 if (!dma)
624 return -ENOMEM;
625
626 dma->cmd = HW_CMD_BLK_DISCARD;
627 dma->laddr = laddr;
628 dma->dma_addr = 0;
629 dma->sub_page.off = 0;
630 dma->sub_page.cnt = 0;
631 dma->page = NULL;
632 dma->pg_off = 0;
633 dma->cb = cb;
634 dma->cb_data = cb_data;
635
636 dev_dbg(CARD_TO_DEV(card), "Queuing[D] laddr %x\n", dma->laddr);
637
638 list_add_tail(&dma->list, q);
639
640 return 0;
641 }
642
643 static int rsxx_queue_dma(struct rsxx_cardinfo *card,
644 struct list_head *q,
645 int dir,
646 unsigned int dma_off,
647 unsigned int dma_len,
648 unsigned int laddr,
649 struct page *page,
650 unsigned int pg_off,
651 rsxx_dma_cb cb,
652 void *cb_data)
653 {
654 struct rsxx_dma *dma;
655
656 dma = kmem_cache_alloc(rsxx_dma_pool, GFP_KERNEL);
657 if (!dma)
658 return -ENOMEM;
659
660 dma->cmd = dir ? HW_CMD_BLK_WRITE : HW_CMD_BLK_READ;
661 dma->laddr = laddr;
662 dma->sub_page.off = (dma_off >> 9);
663 dma->sub_page.cnt = (dma_len >> 9);
664 dma->page = page;
665 dma->pg_off = pg_off;
666 dma->cb = cb;
667 dma->cb_data = cb_data;
668
669 dev_dbg(CARD_TO_DEV(card),
670 "Queuing[%c] laddr %x off %d cnt %d page %p pg_off %d\n",
671 dir ? 'W' : 'R', dma->laddr, dma->sub_page.off,
672 dma->sub_page.cnt, dma->page, dma->pg_off);
673
674 /* Queue the DMA */
675 list_add_tail(&dma->list, q);
676
677 return 0;
678 }
679
680 int rsxx_dma_queue_bio(struct rsxx_cardinfo *card,
681 struct bio *bio,
682 atomic_t *n_dmas,
683 rsxx_dma_cb cb,
684 void *cb_data)
685 {
686 struct list_head dma_list[RSXX_MAX_TARGETS];
687 struct bio_vec bvec;
688 struct bvec_iter iter;
689 unsigned long long addr8;
690 unsigned int laddr;
691 unsigned int bv_len;
692 unsigned int bv_off;
693 unsigned int dma_off;
694 unsigned int dma_len;
695 int dma_cnt[RSXX_MAX_TARGETS];
696 int tgt;
697 int st;
698 int i;
699
700 addr8 = bio->bi_iter.bi_sector << 9; /* sectors are 512 bytes */
701 atomic_set(n_dmas, 0);
702
703 for (i = 0; i < card->n_targets; i++) {
704 INIT_LIST_HEAD(&dma_list[i]);
705 dma_cnt[i] = 0;
706 }
707
708 if (bio_op(bio) == REQ_OP_DISCARD) {
709 bv_len = bio->bi_iter.bi_size;
710
711 while (bv_len > 0) {
712 tgt = rsxx_get_dma_tgt(card, addr8);
713 laddr = rsxx_addr8_to_laddr(addr8, card);
714
715 st = rsxx_queue_discard(card, &dma_list[tgt], laddr,
716 cb, cb_data);
717 if (st)
718 goto bvec_err;
719
720 dma_cnt[tgt]++;
721 atomic_inc(n_dmas);
722 addr8 += RSXX_HW_BLK_SIZE;
723 bv_len -= RSXX_HW_BLK_SIZE;
724 }
725 } else {
726 bio_for_each_segment(bvec, bio, iter) {
727 bv_len = bvec.bv_len;
728 bv_off = bvec.bv_offset;
729
730 while (bv_len > 0) {
731 tgt = rsxx_get_dma_tgt(card, addr8);
732 laddr = rsxx_addr8_to_laddr(addr8, card);
733 dma_off = addr8 & RSXX_HW_BLK_MASK;
734 dma_len = min(bv_len,
735 RSXX_HW_BLK_SIZE - dma_off);
736
737 st = rsxx_queue_dma(card, &dma_list[tgt],
738 bio_data_dir(bio),
739 dma_off, dma_len,
740 laddr, bvec.bv_page,
741 bv_off, cb, cb_data);
742 if (st)
743 goto bvec_err;
744
745 dma_cnt[tgt]++;
746 atomic_inc(n_dmas);
747 addr8 += dma_len;
748 bv_off += dma_len;
749 bv_len -= dma_len;
750 }
751 }
752 }
753
754 for (i = 0; i < card->n_targets; i++) {
755 if (!list_empty(&dma_list[i])) {
756 spin_lock_bh(&card->ctrl[i].queue_lock);
757 card->ctrl[i].stats.sw_q_depth += dma_cnt[i];
758 list_splice_tail(&dma_list[i], &card->ctrl[i].queue);
759 spin_unlock_bh(&card->ctrl[i].queue_lock);
760
761 queue_work(card->ctrl[i].issue_wq,
762 &card->ctrl[i].issue_dma_work);
763 }
764 }
765
766 return 0;
767
768 bvec_err:
769 for (i = 0; i < card->n_targets; i++)
770 rsxx_cleanup_dma_queue(&card->ctrl[i], &dma_list[i],
771 FREE_DMA);
772
773 return st;
774 }
775
776
777 /*----------------- DMA Engine Initialization & Setup -------------------*/
778 int rsxx_hw_buffers_init(struct pci_dev *dev, struct rsxx_dma_ctrl *ctrl)
779 {
780 ctrl->status.buf = pci_alloc_consistent(dev, STATUS_BUFFER_SIZE8,
781 &ctrl->status.dma_addr);
782 ctrl->cmd.buf = pci_alloc_consistent(dev, COMMAND_BUFFER_SIZE8,
783 &ctrl->cmd.dma_addr);
784 if (ctrl->status.buf == NULL || ctrl->cmd.buf == NULL)
785 return -ENOMEM;
786
787 memset(ctrl->status.buf, 0xac, STATUS_BUFFER_SIZE8);
788 iowrite32(lower_32_bits(ctrl->status.dma_addr),
789 ctrl->regmap + SB_ADD_LO);
790 iowrite32(upper_32_bits(ctrl->status.dma_addr),
791 ctrl->regmap + SB_ADD_HI);
792
793 memset(ctrl->cmd.buf, 0x83, COMMAND_BUFFER_SIZE8);
794 iowrite32(lower_32_bits(ctrl->cmd.dma_addr), ctrl->regmap + CB_ADD_LO);
795 iowrite32(upper_32_bits(ctrl->cmd.dma_addr), ctrl->regmap + CB_ADD_HI);
796
797 ctrl->status.idx = ioread32(ctrl->regmap + HW_STATUS_CNT);
798 if (ctrl->status.idx > RSXX_MAX_OUTSTANDING_CMDS) {
799 dev_crit(&dev->dev, "Failed reading status cnt x%x\n",
800 ctrl->status.idx);
801 return -EINVAL;
802 }
803 iowrite32(ctrl->status.idx, ctrl->regmap + HW_STATUS_CNT);
804 iowrite32(ctrl->status.idx, ctrl->regmap + SW_STATUS_CNT);
805
806 ctrl->cmd.idx = ioread32(ctrl->regmap + HW_CMD_IDX);
807 if (ctrl->cmd.idx > RSXX_MAX_OUTSTANDING_CMDS) {
808 dev_crit(&dev->dev, "Failed reading cmd cnt x%x\n",
809 ctrl->status.idx);
810 return -EINVAL;
811 }
812 iowrite32(ctrl->cmd.idx, ctrl->regmap + HW_CMD_IDX);
813 iowrite32(ctrl->cmd.idx, ctrl->regmap + SW_CMD_IDX);
814
815 return 0;
816 }
817
818 static int rsxx_dma_ctrl_init(struct pci_dev *dev,
819 struct rsxx_dma_ctrl *ctrl)
820 {
821 int i;
822 int st;
823
824 memset(&ctrl->stats, 0, sizeof(ctrl->stats));
825
826 ctrl->trackers = vmalloc(DMA_TRACKER_LIST_SIZE8);
827 if (!ctrl->trackers)
828 return -ENOMEM;
829
830 ctrl->trackers->head = 0;
831 for (i = 0; i < RSXX_MAX_OUTSTANDING_CMDS; i++) {
832 ctrl->trackers->list[i].next_tag = i + 1;
833 ctrl->trackers->list[i].dma = NULL;
834 }
835 ctrl->trackers->list[RSXX_MAX_OUTSTANDING_CMDS-1].next_tag = -1;
836 spin_lock_init(&ctrl->trackers->lock);
837
838 spin_lock_init(&ctrl->queue_lock);
839 mutex_init(&ctrl->work_lock);
840 INIT_LIST_HEAD(&ctrl->queue);
841
842 setup_timer(&ctrl->activity_timer, dma_engine_stalled,
843 (unsigned long)ctrl);
844
845 ctrl->issue_wq = alloc_ordered_workqueue(DRIVER_NAME"_issue", 0);
846 if (!ctrl->issue_wq)
847 return -ENOMEM;
848
849 ctrl->done_wq = alloc_ordered_workqueue(DRIVER_NAME"_done", 0);
850 if (!ctrl->done_wq)
851 return -ENOMEM;
852
853 INIT_WORK(&ctrl->issue_dma_work, rsxx_schedule_issue);
854 INIT_WORK(&ctrl->dma_done_work, rsxx_schedule_done);
855
856 st = rsxx_hw_buffers_init(dev, ctrl);
857 if (st)
858 return st;
859
860 return 0;
861 }
862
863 static int rsxx_dma_stripe_setup(struct rsxx_cardinfo *card,
864 unsigned int stripe_size8)
865 {
866 if (!is_power_of_2(stripe_size8)) {
867 dev_err(CARD_TO_DEV(card),
868 "stripe_size is NOT a power of 2!\n");
869 return -EINVAL;
870 }
871
872 card->_stripe.lower_mask = stripe_size8 - 1;
873
874 card->_stripe.upper_mask = ~(card->_stripe.lower_mask);
875 card->_stripe.upper_shift = ffs(card->n_targets) - 1;
876
877 card->_stripe.target_mask = card->n_targets - 1;
878 card->_stripe.target_shift = ffs(stripe_size8) - 1;
879
880 dev_dbg(CARD_TO_DEV(card), "_stripe.lower_mask = x%016llx\n",
881 card->_stripe.lower_mask);
882 dev_dbg(CARD_TO_DEV(card), "_stripe.upper_shift = x%016llx\n",
883 card->_stripe.upper_shift);
884 dev_dbg(CARD_TO_DEV(card), "_stripe.upper_mask = x%016llx\n",
885 card->_stripe.upper_mask);
886 dev_dbg(CARD_TO_DEV(card), "_stripe.target_mask = x%016llx\n",
887 card->_stripe.target_mask);
888 dev_dbg(CARD_TO_DEV(card), "_stripe.target_shift = x%016llx\n",
889 card->_stripe.target_shift);
890
891 return 0;
892 }
893
894 int rsxx_dma_configure(struct rsxx_cardinfo *card)
895 {
896 u32 intr_coal;
897
898 intr_coal = dma_intr_coal_val(card->config.data.intr_coal.mode,
899 card->config.data.intr_coal.count,
900 card->config.data.intr_coal.latency);
901 iowrite32(intr_coal, card->regmap + INTR_COAL);
902
903 return rsxx_dma_stripe_setup(card, card->config.data.stripe_size);
904 }
905
906 int rsxx_dma_setup(struct rsxx_cardinfo *card)
907 {
908 unsigned long flags;
909 int st;
910 int i;
911
912 dev_info(CARD_TO_DEV(card),
913 "Initializing %d DMA targets\n",
914 card->n_targets);
915
916 /* Regmap is divided up into 4K chunks. One for each DMA channel */
917 for (i = 0; i < card->n_targets; i++)
918 card->ctrl[i].regmap = card->regmap + (i * 4096);
919
920 card->dma_fault = 0;
921
922 /* Reset the DMA queues */
923 rsxx_dma_queue_reset(card);
924
925 /************* Setup DMA Control *************/
926 for (i = 0; i < card->n_targets; i++) {
927 st = rsxx_dma_ctrl_init(card->dev, &card->ctrl[i]);
928 if (st)
929 goto failed_dma_setup;
930
931 card->ctrl[i].card = card;
932 card->ctrl[i].id = i;
933 }
934
935 card->scrub_hard = 1;
936
937 if (card->config_valid)
938 rsxx_dma_configure(card);
939
940 /* Enable the interrupts after all setup has completed. */
941 for (i = 0; i < card->n_targets; i++) {
942 spin_lock_irqsave(&card->irq_lock, flags);
943 rsxx_enable_ier_and_isr(card, CR_INTR_DMA(i));
944 spin_unlock_irqrestore(&card->irq_lock, flags);
945 }
946
947 return 0;
948
949 failed_dma_setup:
950 for (i = 0; i < card->n_targets; i++) {
951 struct rsxx_dma_ctrl *ctrl = &card->ctrl[i];
952
953 if (ctrl->issue_wq) {
954 destroy_workqueue(ctrl->issue_wq);
955 ctrl->issue_wq = NULL;
956 }
957
958 if (ctrl->done_wq) {
959 destroy_workqueue(ctrl->done_wq);
960 ctrl->done_wq = NULL;
961 }
962
963 if (ctrl->trackers)
964 vfree(ctrl->trackers);
965
966 if (ctrl->status.buf)
967 pci_free_consistent(card->dev, STATUS_BUFFER_SIZE8,
968 ctrl->status.buf,
969 ctrl->status.dma_addr);
970 if (ctrl->cmd.buf)
971 pci_free_consistent(card->dev, COMMAND_BUFFER_SIZE8,
972 ctrl->cmd.buf, ctrl->cmd.dma_addr);
973 }
974
975 return st;
976 }
977
978 int rsxx_dma_cancel(struct rsxx_dma_ctrl *ctrl)
979 {
980 struct rsxx_dma *dma;
981 int i;
982 int cnt = 0;
983
984 /* Clean up issued DMAs */
985 for (i = 0; i < RSXX_MAX_OUTSTANDING_CMDS; i++) {
986 dma = get_tracker_dma(ctrl->trackers, i);
987 if (dma) {
988 atomic_dec(&ctrl->stats.hw_q_depth);
989 rsxx_complete_dma(ctrl, dma, DMA_CANCELLED);
990 push_tracker(ctrl->trackers, i);
991 cnt++;
992 }
993 }
994
995 return cnt;
996 }
997
998 void rsxx_dma_destroy(struct rsxx_cardinfo *card)
999 {
1000 struct rsxx_dma_ctrl *ctrl;
1001 int i;
1002
1003 for (i = 0; i < card->n_targets; i++) {
1004 ctrl = &card->ctrl[i];
1005
1006 if (ctrl->issue_wq) {
1007 destroy_workqueue(ctrl->issue_wq);
1008 ctrl->issue_wq = NULL;
1009 }
1010
1011 if (ctrl->done_wq) {
1012 destroy_workqueue(ctrl->done_wq);
1013 ctrl->done_wq = NULL;
1014 }
1015
1016 if (timer_pending(&ctrl->activity_timer))
1017 del_timer_sync(&ctrl->activity_timer);
1018
1019 /* Clean up the DMA queue */
1020 spin_lock_bh(&ctrl->queue_lock);
1021 rsxx_cleanup_dma_queue(ctrl, &ctrl->queue, COMPLETE_DMA);
1022 spin_unlock_bh(&ctrl->queue_lock);
1023
1024 rsxx_dma_cancel(ctrl);
1025
1026 vfree(ctrl->trackers);
1027
1028 pci_free_consistent(card->dev, STATUS_BUFFER_SIZE8,
1029 ctrl->status.buf, ctrl->status.dma_addr);
1030 pci_free_consistent(card->dev, COMMAND_BUFFER_SIZE8,
1031 ctrl->cmd.buf, ctrl->cmd.dma_addr);
1032 }
1033 }
1034
1035 int rsxx_eeh_save_issued_dmas(struct rsxx_cardinfo *card)
1036 {
1037 int i;
1038 int j;
1039 int cnt;
1040 struct rsxx_dma *dma;
1041 struct list_head *issued_dmas;
1042
1043 issued_dmas = kzalloc(sizeof(*issued_dmas) * card->n_targets,
1044 GFP_KERNEL);
1045 if (!issued_dmas)
1046 return -ENOMEM;
1047
1048 for (i = 0; i < card->n_targets; i++) {
1049 INIT_LIST_HEAD(&issued_dmas[i]);
1050 cnt = 0;
1051 for (j = 0; j < RSXX_MAX_OUTSTANDING_CMDS; j++) {
1052 dma = get_tracker_dma(card->ctrl[i].trackers, j);
1053 if (dma == NULL)
1054 continue;
1055
1056 if (dma->cmd == HW_CMD_BLK_WRITE)
1057 card->ctrl[i].stats.writes_issued--;
1058 else if (dma->cmd == HW_CMD_BLK_DISCARD)
1059 card->ctrl[i].stats.discards_issued--;
1060 else
1061 card->ctrl[i].stats.reads_issued--;
1062
1063 if (dma->cmd != HW_CMD_BLK_DISCARD) {
1064 pci_unmap_page(card->dev, dma->dma_addr,
1065 get_dma_size(dma),
1066 dma->cmd == HW_CMD_BLK_WRITE ?
1067 PCI_DMA_TODEVICE :
1068 PCI_DMA_FROMDEVICE);
1069 }
1070
1071 list_add_tail(&dma->list, &issued_dmas[i]);
1072 push_tracker(card->ctrl[i].trackers, j);
1073 cnt++;
1074 }
1075
1076 spin_lock_bh(&card->ctrl[i].queue_lock);
1077 list_splice(&issued_dmas[i], &card->ctrl[i].queue);
1078
1079 atomic_sub(cnt, &card->ctrl[i].stats.hw_q_depth);
1080 card->ctrl[i].stats.sw_q_depth += cnt;
1081 card->ctrl[i].e_cnt = 0;
1082 spin_unlock_bh(&card->ctrl[i].queue_lock);
1083 }
1084
1085 kfree(issued_dmas);
1086
1087 return 0;
1088 }
1089
1090 int rsxx_dma_init(void)
1091 {
1092 rsxx_dma_pool = KMEM_CACHE(rsxx_dma, SLAB_HWCACHE_ALIGN);
1093 if (!rsxx_dma_pool)
1094 return -ENOMEM;
1095
1096 return 0;
1097 }
1098
1099
1100 void rsxx_dma_cleanup(void)
1101 {
1102 kmem_cache_destroy(rsxx_dma_pool);
1103 }
1104
Linux with added FPC-III support