?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge branch 'i7300_idle' into release
[linux-2.6/mini2440.git] / drivers / dma / ioat_dma.c
blobb0438c4f0c305128b9ab5111de2311d193c259b5
1 /*
2 * Intel I/OAT DMA Linux driver
3 * Copyright(c) 2004 - 2007 Intel Corporation.
4 *
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms and conditions of the GNU General Public License,
7 * version 2, as published by the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
12 * more details.
13 *
14 * You should have received a copy of the GNU General Public License along with
15 * this program; if not, write to the Free Software Foundation, Inc.,
16 * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
17 *
18 * The full GNU General Public License is included in this distribution in
19 * the file called "COPYING".
20 *
21 */
22
23 /*
24 * This driver supports an Intel I/OAT DMA engine, which does asynchronous
25 * copy operations.
26 */
27
28 #include <linux/init.h>
29 #include <linux/module.h>
30 #include <linux/pci.h>
31 #include <linux/interrupt.h>
32 #include <linux/dmaengine.h>
33 #include <linux/delay.h>
34 #include <linux/dma-mapping.h>
35 #include <linux/workqueue.h>
36 #include <linux/i7300_idle.h>
37 #include "ioatdma.h"
38 #include "ioatdma_registers.h"
39 #include "ioatdma_hw.h"
40
41 #define to_ioat_chan(chan) container_of(chan, struct ioat_dma_chan, common)
42 #define to_ioatdma_device(dev) container_of(dev, struct ioatdma_device, common)
43 #define to_ioat_desc(lh) container_of(lh, struct ioat_desc_sw, node)
44 #define tx_to_ioat_desc(tx) container_of(tx, struct ioat_desc_sw, async_tx)
45
46 #define chan_num(ch) ((int)((ch)->reg_base - (ch)->device->reg_base) / 0x80)
47 static int ioat_pending_level = 4;
48 module_param(ioat_pending_level, int, 0644);
49 MODULE_PARM_DESC(ioat_pending_level,
50 "high-water mark for pushing ioat descriptors (default: 4)");
51
52 #define RESET_DELAY msecs_to_jiffies(100)
53 #define WATCHDOG_DELAY round_jiffies(msecs_to_jiffies(2000))
54 static void ioat_dma_chan_reset_part2(struct work_struct *work);
55 static void ioat_dma_chan_watchdog(struct work_struct *work);
56
57 /*
58 * workaround for IOAT ver.3.0 null descriptor issue
59 * (channel returns error when size is 0)
60 */
61 #define NULL_DESC_BUFFER_SIZE 1
62
63 /* internal functions */
64 static void ioat_dma_start_null_desc(struct ioat_dma_chan *ioat_chan);
65 static void ioat_dma_memcpy_cleanup(struct ioat_dma_chan *ioat_chan);
66
67 static struct ioat_desc_sw *
68 ioat1_dma_get_next_descriptor(struct ioat_dma_chan *ioat_chan);
69 static struct ioat_desc_sw *
70 ioat2_dma_get_next_descriptor(struct ioat_dma_chan *ioat_chan);
71
72 static inline struct ioat_dma_chan *ioat_lookup_chan_by_index(
73 struct ioatdma_device *device,
74 int index)
75 {
76 return device->idx[index];
77 }
78
79 /**
80 * ioat_dma_do_interrupt - handler used for single vector interrupt mode
81 * @irq: interrupt id
82 * @data: interrupt data
83 */
84 static irqreturn_t ioat_dma_do_interrupt(int irq, void *data)
85 {
86 struct ioatdma_device *instance = data;
87 struct ioat_dma_chan *ioat_chan;
88 unsigned long attnstatus;
89 int bit;
90 u8 intrctrl;
91
92 intrctrl = readb(instance->reg_base + IOAT_INTRCTRL_OFFSET);
93
94 if (!(intrctrl & IOAT_INTRCTRL_MASTER_INT_EN))
95 return IRQ_NONE;
96
97 if (!(intrctrl & IOAT_INTRCTRL_INT_STATUS)) {
98 writeb(intrctrl, instance->reg_base + IOAT_INTRCTRL_OFFSET);
99 return IRQ_NONE;
100 }
101
102 attnstatus = readl(instance->reg_base + IOAT_ATTNSTATUS_OFFSET);
103 for_each_bit(bit, &attnstatus, BITS_PER_LONG) {
104 ioat_chan = ioat_lookup_chan_by_index(instance, bit);
105 tasklet_schedule(&ioat_chan->cleanup_task);
106 }
107
108 writeb(intrctrl, instance->reg_base + IOAT_INTRCTRL_OFFSET);
109 return IRQ_HANDLED;
110 }
111
112 /**
113 * ioat_dma_do_interrupt_msix - handler used for vector-per-channel interrupt mode
114 * @irq: interrupt id
115 * @data: interrupt data
116 */
117 static irqreturn_t ioat_dma_do_interrupt_msix(int irq, void *data)
118 {
119 struct ioat_dma_chan *ioat_chan = data;
120
121 tasklet_schedule(&ioat_chan->cleanup_task);
122
123 return IRQ_HANDLED;
124 }
125
126 static void ioat_dma_cleanup_tasklet(unsigned long data);
127
128 /**
129 * ioat_dma_enumerate_channels - find and initialize the device's channels
130 * @device: the device to be enumerated
131 */
132 static int ioat_dma_enumerate_channels(struct ioatdma_device *device)
133 {
134 u8 xfercap_scale;
135 u32 xfercap;
136 int i;
137 struct ioat_dma_chan *ioat_chan;
138
139 /*
140 * IOAT ver.3 workarounds
141 */
142 if (device->version == IOAT_VER_3_0) {
143 u32 chan_err_mask;
144 u16 dev_id;
145 u32 dmauncerrsts;
146
147 /*
148 * Write CHANERRMSK_INT with 3E07h to mask out the errors
149 * that can cause stability issues for IOAT ver.3
150 */
151 chan_err_mask = 0x3E07;
152 pci_write_config_dword(device->pdev,
153 IOAT_PCI_CHANERRMASK_INT_OFFSET,
154 chan_err_mask);
155
156 /*
157 * Clear DMAUNCERRSTS Cfg-Reg Parity Error status bit
158 * (workaround for spurious config parity error after restart)
159 */
160 pci_read_config_word(device->pdev,
161 IOAT_PCI_DEVICE_ID_OFFSET,
162 &dev_id);
163 if (dev_id == PCI_DEVICE_ID_INTEL_IOAT_TBG0) {
164 dmauncerrsts = 0x10;
165 pci_write_config_dword(device->pdev,
166 IOAT_PCI_DMAUNCERRSTS_OFFSET,
167 dmauncerrsts);
168 }
169 }
170
171 device->common.chancnt = readb(device->reg_base + IOAT_CHANCNT_OFFSET);
172 xfercap_scale = readb(device->reg_base + IOAT_XFERCAP_OFFSET);
173 xfercap = (xfercap_scale == 0 ? -1 : (1UL << xfercap_scale));
174
175 #ifdef CONFIG_I7300_IDLE_IOAT_CHANNEL
176 if (i7300_idle_platform_probe(NULL, NULL) == 0) {
177 device->common.chancnt--;
178 }
179 #endif
180 for (i = 0; i < device->common.chancnt; i++) {
181 ioat_chan = kzalloc(sizeof(*ioat_chan), GFP_KERNEL);
182 if (!ioat_chan) {
183 device->common.chancnt = i;
184 break;
185 }
186
187 ioat_chan->device = device;
188 ioat_chan->reg_base = device->reg_base + (0x80 * (i + 1));
189 ioat_chan->xfercap = xfercap;
190 ioat_chan->desccount = 0;
191 INIT_DELAYED_WORK(&ioat_chan->work, ioat_dma_chan_reset_part2);
192 if (ioat_chan->device->version != IOAT_VER_1_2) {
193 writel(IOAT_DCACTRL_CMPL_WRITE_ENABLE
194 | IOAT_DMA_DCA_ANY_CPU,
195 ioat_chan->reg_base + IOAT_DCACTRL_OFFSET);
196 }
197 spin_lock_init(&ioat_chan->cleanup_lock);
198 spin_lock_init(&ioat_chan->desc_lock);
199 INIT_LIST_HEAD(&ioat_chan->free_desc);
200 INIT_LIST_HEAD(&ioat_chan->used_desc);
201 /* This should be made common somewhere in dmaengine.c */
202 ioat_chan->common.device = &device->common;
203 list_add_tail(&ioat_chan->common.device_node,
204 &device->common.channels);
205 device->idx[i] = ioat_chan;
206 tasklet_init(&ioat_chan->cleanup_task,
207 ioat_dma_cleanup_tasklet,
208 (unsigned long) ioat_chan);
209 tasklet_disable(&ioat_chan->cleanup_task);
210 }
211 return device->common.chancnt;
212 }
213
214 /**
215 * ioat_dma_memcpy_issue_pending - push potentially unrecognized appended
216 * descriptors to hw
217 * @chan: DMA channel handle
218 */
219 static inline void __ioat1_dma_memcpy_issue_pending(
220 struct ioat_dma_chan *ioat_chan)
221 {
222 ioat_chan->pending = 0;
223 writeb(IOAT_CHANCMD_APPEND, ioat_chan->reg_base + IOAT1_CHANCMD_OFFSET);
224 }
225
226 static void ioat1_dma_memcpy_issue_pending(struct dma_chan *chan)
227 {
228 struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
229
230 if (ioat_chan->pending > 0) {
231 spin_lock_bh(&ioat_chan->desc_lock);
232 __ioat1_dma_memcpy_issue_pending(ioat_chan);
233 spin_unlock_bh(&ioat_chan->desc_lock);
234 }
235 }
236
237 static inline void __ioat2_dma_memcpy_issue_pending(
238 struct ioat_dma_chan *ioat_chan)
239 {
240 ioat_chan->pending = 0;
241 writew(ioat_chan->dmacount,
242 ioat_chan->reg_base + IOAT_CHAN_DMACOUNT_OFFSET);
243 }
244
245 static void ioat2_dma_memcpy_issue_pending(struct dma_chan *chan)
246 {
247 struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
248
249 if (ioat_chan->pending > 0) {
250 spin_lock_bh(&ioat_chan->desc_lock);
251 __ioat2_dma_memcpy_issue_pending(ioat_chan);
252 spin_unlock_bh(&ioat_chan->desc_lock);
253 }
254 }
255
256
257 /**
258 * ioat_dma_chan_reset_part2 - reinit the channel after a reset
259 */
260 static void ioat_dma_chan_reset_part2(struct work_struct *work)
261 {
262 struct ioat_dma_chan *ioat_chan =
263 container_of(work, struct ioat_dma_chan, work.work);
264 struct ioat_desc_sw *desc;
265
266 spin_lock_bh(&ioat_chan->cleanup_lock);
267 spin_lock_bh(&ioat_chan->desc_lock);
268
269 ioat_chan->completion_virt->low = 0;
270 ioat_chan->completion_virt->high = 0;
271 ioat_chan->pending = 0;
272
273 /*
274 * count the descriptors waiting, and be sure to do it
275 * right for both the CB1 line and the CB2 ring
276 */
277 ioat_chan->dmacount = 0;
278 if (ioat_chan->used_desc.prev) {
279 desc = to_ioat_desc(ioat_chan->used_desc.prev);
280 do {
281 ioat_chan->dmacount++;
282 desc = to_ioat_desc(desc->node.next);
283 } while (&desc->node != ioat_chan->used_desc.next);
284 }
285
286 /*
287 * write the new starting descriptor address
288 * this puts channel engine into ARMED state
289 */
290 desc = to_ioat_desc(ioat_chan->used_desc.prev);
291 switch (ioat_chan->device->version) {
292 case IOAT_VER_1_2:
293 writel(((u64) desc->async_tx.phys) & 0x00000000FFFFFFFF,
294 ioat_chan->reg_base + IOAT1_CHAINADDR_OFFSET_LOW);
295 writel(((u64) desc->async_tx.phys) >> 32,
296 ioat_chan->reg_base + IOAT1_CHAINADDR_OFFSET_HIGH);
297
298 writeb(IOAT_CHANCMD_START, ioat_chan->reg_base
299 + IOAT_CHANCMD_OFFSET(ioat_chan->device->version));
300 break;
301 case IOAT_VER_2_0:
302 writel(((u64) desc->async_tx.phys) & 0x00000000FFFFFFFF,
303 ioat_chan->reg_base + IOAT2_CHAINADDR_OFFSET_LOW);
304 writel(((u64) desc->async_tx.phys) >> 32,
305 ioat_chan->reg_base + IOAT2_CHAINADDR_OFFSET_HIGH);
306
307 /* tell the engine to go with what's left to be done */
308 writew(ioat_chan->dmacount,
309 ioat_chan->reg_base + IOAT_CHAN_DMACOUNT_OFFSET);
310
311 break;
312 }
313 dev_err(&ioat_chan->device->pdev->dev,
314 "chan%d reset - %d descs waiting, %d total desc\n",
315 chan_num(ioat_chan), ioat_chan->dmacount, ioat_chan->desccount);
316
317 spin_unlock_bh(&ioat_chan->desc_lock);
318 spin_unlock_bh(&ioat_chan->cleanup_lock);
319 }
320
321 /**
322 * ioat_dma_reset_channel - restart a channel
323 * @ioat_chan: IOAT DMA channel handle
324 */
325 static void ioat_dma_reset_channel(struct ioat_dma_chan *ioat_chan)
326 {
327 u32 chansts, chanerr;
328
329 if (!ioat_chan->used_desc.prev)
330 return;
331
332 chanerr = readl(ioat_chan->reg_base + IOAT_CHANERR_OFFSET);
333 chansts = (ioat_chan->completion_virt->low
334 & IOAT_CHANSTS_DMA_TRANSFER_STATUS);
335 if (chanerr) {
336 dev_err(&ioat_chan->device->pdev->dev,
337 "chan%d, CHANSTS = 0x%08x CHANERR = 0x%04x, clearing\n",
338 chan_num(ioat_chan), chansts, chanerr);
339 writel(chanerr, ioat_chan->reg_base + IOAT_CHANERR_OFFSET);
340 }
341
342 /*
343 * whack it upside the head with a reset
344 * and wait for things to settle out.
345 * force the pending count to a really big negative
346 * to make sure no one forces an issue_pending
347 * while we're waiting.
348 */
349
350 spin_lock_bh(&ioat_chan->desc_lock);
351 ioat_chan->pending = INT_MIN;
352 writeb(IOAT_CHANCMD_RESET,
353 ioat_chan->reg_base
354 + IOAT_CHANCMD_OFFSET(ioat_chan->device->version));
355 spin_unlock_bh(&ioat_chan->desc_lock);
356
357 /* schedule the 2nd half instead of sleeping a long time */
358 schedule_delayed_work(&ioat_chan->work, RESET_DELAY);
359 }
360
361 /**
362 * ioat_dma_chan_watchdog - watch for stuck channels
363 */
364 static void ioat_dma_chan_watchdog(struct work_struct *work)
365 {
366 struct ioatdma_device *device =
367 container_of(work, struct ioatdma_device, work.work);
368 struct ioat_dma_chan *ioat_chan;
369 int i;
370
371 union {
372 u64 full;
373 struct {
374 u32 low;
375 u32 high;
376 };
377 } completion_hw;
378 unsigned long compl_desc_addr_hw;
379
380 for (i = 0; i < device->common.chancnt; i++) {
381 ioat_chan = ioat_lookup_chan_by_index(device, i);
382
383 if (ioat_chan->device->version == IOAT_VER_1_2
384 /* have we started processing anything yet */
385 && ioat_chan->last_completion
386 /* have we completed any since last watchdog cycle? */
387 && (ioat_chan->last_completion ==
388 ioat_chan->watchdog_completion)
389 /* has TCP stuck on one cookie since last watchdog? */
390 && (ioat_chan->watchdog_tcp_cookie ==
391 ioat_chan->watchdog_last_tcp_cookie)
392 && (ioat_chan->watchdog_tcp_cookie !=
393 ioat_chan->completed_cookie)
394 /* is there something in the chain to be processed? */
395 /* CB1 chain always has at least the last one processed */
396 && (ioat_chan->used_desc.prev != ioat_chan->used_desc.next)
397 && ioat_chan->pending == 0) {
398
399 /*
400 * check CHANSTS register for completed
401 * descriptor address.
402 * if it is different than completion writeback,
403 * it is not zero
404 * and it has changed since the last watchdog
405 * we can assume that channel
406 * is still working correctly
407 * and the problem is in completion writeback.
408 * update completion writeback
409 * with actual CHANSTS value
410 * else
411 * try resetting the channel
412 */
413
414 completion_hw.low = readl(ioat_chan->reg_base +
415 IOAT_CHANSTS_OFFSET_LOW(ioat_chan->device->version));
416 completion_hw.high = readl(ioat_chan->reg_base +
417 IOAT_CHANSTS_OFFSET_HIGH(ioat_chan->device->version));
418 #if (BITS_PER_LONG == 64)
419 compl_desc_addr_hw =
420 completion_hw.full
421 & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_ADDR;
422 #else
423 compl_desc_addr_hw =
424 completion_hw.low & IOAT_LOW_COMPLETION_MASK;
425 #endif
426
427 if ((compl_desc_addr_hw != 0)
428 && (compl_desc_addr_hw != ioat_chan->watchdog_completion)
429 && (compl_desc_addr_hw != ioat_chan->last_compl_desc_addr_hw)) {
430 ioat_chan->last_compl_desc_addr_hw = compl_desc_addr_hw;
431 ioat_chan->completion_virt->low = completion_hw.low;
432 ioat_chan->completion_virt->high = completion_hw.high;
433 } else {
434 ioat_dma_reset_channel(ioat_chan);
435 ioat_chan->watchdog_completion = 0;
436 ioat_chan->last_compl_desc_addr_hw = 0;
437 }
438
439 /*
440 * for version 2.0 if there are descriptors yet to be processed
441 * and the last completed hasn't changed since the last watchdog
442 * if they haven't hit the pending level
443 * issue the pending to push them through
444 * else
445 * try resetting the channel
446 */
447 } else if (ioat_chan->device->version == IOAT_VER_2_0
448 && ioat_chan->used_desc.prev
449 && ioat_chan->last_completion
450 && ioat_chan->last_completion == ioat_chan->watchdog_completion) {
451
452 if (ioat_chan->pending < ioat_pending_level)
453 ioat2_dma_memcpy_issue_pending(&ioat_chan->common);
454 else {
455 ioat_dma_reset_channel(ioat_chan);
456 ioat_chan->watchdog_completion = 0;
457 }
458 } else {
459 ioat_chan->last_compl_desc_addr_hw = 0;
460 ioat_chan->watchdog_completion
461 = ioat_chan->last_completion;
462 }
463
464 ioat_chan->watchdog_last_tcp_cookie =
465 ioat_chan->watchdog_tcp_cookie;
466 }
467
468 schedule_delayed_work(&device->work, WATCHDOG_DELAY);
469 }
470
471 static dma_cookie_t ioat1_tx_submit(struct dma_async_tx_descriptor *tx)
472 {
473 struct ioat_dma_chan *ioat_chan = to_ioat_chan(tx->chan);
474 struct ioat_desc_sw *first = tx_to_ioat_desc(tx);
475 struct ioat_desc_sw *prev, *new;
476 struct ioat_dma_descriptor *hw;
477 dma_cookie_t cookie;
478 LIST_HEAD(new_chain);
479 u32 copy;
480 size_t len;
481 dma_addr_t src, dst;
482 unsigned long orig_flags;
483 unsigned int desc_count = 0;
484
485 /* src and dest and len are stored in the initial descriptor */
486 len = first->len;
487 src = first->src;
488 dst = first->dst;
489 orig_flags = first->async_tx.flags;
490 new = first;
491
492 spin_lock_bh(&ioat_chan->desc_lock);
493 prev = to_ioat_desc(ioat_chan->used_desc.prev);
494 prefetch(prev->hw);
495 do {
496 copy = min_t(size_t, len, ioat_chan->xfercap);
497
498 async_tx_ack(&new->async_tx);
499
500 hw = new->hw;
501 hw->size = copy;
502 hw->ctl = 0;
503 hw->src_addr = src;
504 hw->dst_addr = dst;
505 hw->next = 0;
506
507 /* chain together the physical address list for the HW */
508 wmb();
509 prev->hw->next = (u64) new->async_tx.phys;
510
511 len -= copy;
512 dst += copy;
513 src += copy;
514
515 list_add_tail(&new->node, &new_chain);
516 desc_count++;
517 prev = new;
518 } while (len && (new = ioat1_dma_get_next_descriptor(ioat_chan)));
519
520 if (!new) {
521 dev_err(&ioat_chan->device->pdev->dev,
522 "tx submit failed\n");
523 spin_unlock_bh(&ioat_chan->desc_lock);
524 return -ENOMEM;
525 }
526
527 hw->ctl = IOAT_DMA_DESCRIPTOR_CTL_CP_STS;
528 if (new->async_tx.callback) {
529 hw->ctl |= IOAT_DMA_DESCRIPTOR_CTL_INT_GN;
530 if (first != new) {
531 /* move callback into to last desc */
532 new->async_tx.callback = first->async_tx.callback;
533 new->async_tx.callback_param
534 = first->async_tx.callback_param;
535 first->async_tx.callback = NULL;
536 first->async_tx.callback_param = NULL;
537 }
538 }
539
540 new->tx_cnt = desc_count;
541 new->async_tx.flags = orig_flags; /* client is in control of this ack */
542
543 /* store the original values for use in later cleanup */
544 if (new != first) {
545 new->src = first->src;
546 new->dst = first->dst;
547 new->len = first->len;
548 }
549
550 /* cookie incr and addition to used_list must be atomic */
551 cookie = ioat_chan->common.cookie;
552 cookie++;
553 if (cookie < 0)
554 cookie = 1;
555 ioat_chan->common.cookie = new->async_tx.cookie = cookie;
556
557 /* write address into NextDescriptor field of last desc in chain */
558 to_ioat_desc(ioat_chan->used_desc.prev)->hw->next =
559 first->async_tx.phys;
560 list_splice_tail(&new_chain, &ioat_chan->used_desc);
561
562 ioat_chan->dmacount += desc_count;
563 ioat_chan->pending += desc_count;
564 if (ioat_chan->pending >= ioat_pending_level)
565 __ioat1_dma_memcpy_issue_pending(ioat_chan);
566 spin_unlock_bh(&ioat_chan->desc_lock);
567
568 return cookie;
569 }
570
571 static dma_cookie_t ioat2_tx_submit(struct dma_async_tx_descriptor *tx)
572 {
573 struct ioat_dma_chan *ioat_chan = to_ioat_chan(tx->chan);
574 struct ioat_desc_sw *first = tx_to_ioat_desc(tx);
575 struct ioat_desc_sw *new;
576 struct ioat_dma_descriptor *hw;
577 dma_cookie_t cookie;
578 u32 copy;
579 size_t len;
580 dma_addr_t src, dst;
581 unsigned long orig_flags;
582 unsigned int desc_count = 0;
583
584 /* src and dest and len are stored in the initial descriptor */
585 len = first->len;
586 src = first->src;
587 dst = first->dst;
588 orig_flags = first->async_tx.flags;
589 new = first;
590
591 /*
592 * ioat_chan->desc_lock is still in force in version 2 path
593 * it gets unlocked at end of this function
594 */
595 do {
596 copy = min_t(size_t, len, ioat_chan->xfercap);
597
598 async_tx_ack(&new->async_tx);
599
600 hw = new->hw;
601 hw->size = copy;
602 hw->ctl = 0;
603 hw->src_addr = src;
604 hw->dst_addr = dst;
605
606 len -= copy;
607 dst += copy;
608 src += copy;
609 desc_count++;
610 } while (len && (new = ioat2_dma_get_next_descriptor(ioat_chan)));
611
612 if (!new) {
613 dev_err(&ioat_chan->device->pdev->dev,
614 "tx submit failed\n");
615 spin_unlock_bh(&ioat_chan->desc_lock);
616 return -ENOMEM;
617 }
618
619 hw->ctl |= IOAT_DMA_DESCRIPTOR_CTL_CP_STS;
620 if (new->async_tx.callback) {
621 hw->ctl |= IOAT_DMA_DESCRIPTOR_CTL_INT_GN;
622 if (first != new) {
623 /* move callback into to last desc */
624 new->async_tx.callback = first->async_tx.callback;
625 new->async_tx.callback_param
626 = first->async_tx.callback_param;
627 first->async_tx.callback = NULL;
628 first->async_tx.callback_param = NULL;
629 }
630 }
631
632 new->tx_cnt = desc_count;
633 new->async_tx.flags = orig_flags; /* client is in control of this ack */
634
635 /* store the original values for use in later cleanup */
636 if (new != first) {
637 new->src = first->src;
638 new->dst = first->dst;
639 new->len = first->len;
640 }
641
642 /* cookie incr and addition to used_list must be atomic */
643 cookie = ioat_chan->common.cookie;
644 cookie++;
645 if (cookie < 0)
646 cookie = 1;
647 ioat_chan->common.cookie = new->async_tx.cookie = cookie;
648
649 ioat_chan->dmacount += desc_count;
650 ioat_chan->pending += desc_count;
651 if (ioat_chan->pending >= ioat_pending_level)
652 __ioat2_dma_memcpy_issue_pending(ioat_chan);
653 spin_unlock_bh(&ioat_chan->desc_lock);
654
655 return cookie;
656 }
657
658 /**
659 * ioat_dma_alloc_descriptor - allocate and return a sw and hw descriptor pair
660 * @ioat_chan: the channel supplying the memory pool for the descriptors
661 * @flags: allocation flags
662 */
663 static struct ioat_desc_sw *ioat_dma_alloc_descriptor(
664 struct ioat_dma_chan *ioat_chan,
665 gfp_t flags)
666 {
667 struct ioat_dma_descriptor *desc;
668 struct ioat_desc_sw *desc_sw;
669 struct ioatdma_device *ioatdma_device;
670 dma_addr_t phys;
671
672 ioatdma_device = to_ioatdma_device(ioat_chan->common.device);
673 desc = pci_pool_alloc(ioatdma_device->dma_pool, flags, &phys);
674 if (unlikely(!desc))
675 return NULL;
676
677 desc_sw = kzalloc(sizeof(*desc_sw), flags);
678 if (unlikely(!desc_sw)) {
679 pci_pool_free(ioatdma_device->dma_pool, desc, phys);
680 return NULL;
681 }
682
683 memset(desc, 0, sizeof(*desc));
684 dma_async_tx_descriptor_init(&desc_sw->async_tx, &ioat_chan->common);
685 switch (ioat_chan->device->version) {
686 case IOAT_VER_1_2:
687 desc_sw->async_tx.tx_submit = ioat1_tx_submit;
688 break;
689 case IOAT_VER_2_0:
690 case IOAT_VER_3_0:
691 desc_sw->async_tx.tx_submit = ioat2_tx_submit;
692 break;
693 }
694 INIT_LIST_HEAD(&desc_sw->async_tx.tx_list);
695
696 desc_sw->hw = desc;
697 desc_sw->async_tx.phys = phys;
698
699 return desc_sw;
700 }
701
702 static int ioat_initial_desc_count = 256;
703 module_param(ioat_initial_desc_count, int, 0644);
704 MODULE_PARM_DESC(ioat_initial_desc_count,
705 "initial descriptors per channel (default: 256)");
706
707 /**
708 * ioat2_dma_massage_chan_desc - link the descriptors into a circle
709 * @ioat_chan: the channel to be massaged
710 */
711 static void ioat2_dma_massage_chan_desc(struct ioat_dma_chan *ioat_chan)
712 {
713 struct ioat_desc_sw *desc, *_desc;
714
715 /* setup used_desc */
716 ioat_chan->used_desc.next = ioat_chan->free_desc.next;
717 ioat_chan->used_desc.prev = NULL;
718
719 /* pull free_desc out of the circle so that every node is a hw
720 * descriptor, but leave it pointing to the list
721 */
722 ioat_chan->free_desc.prev->next = ioat_chan->free_desc.next;
723 ioat_chan->free_desc.next->prev = ioat_chan->free_desc.prev;
724
725 /* circle link the hw descriptors */
726 desc = to_ioat_desc(ioat_chan->free_desc.next);
727 desc->hw->next = to_ioat_desc(desc->node.next)->async_tx.phys;
728 list_for_each_entry_safe(desc, _desc, ioat_chan->free_desc.next, node) {
729 desc->hw->next = to_ioat_desc(desc->node.next)->async_tx.phys;
730 }
731 }
732
733 /**
734 * ioat_dma_alloc_chan_resources - returns the number of allocated descriptors
735 * @chan: the channel to be filled out
736 */
737 static int ioat_dma_alloc_chan_resources(struct dma_chan *chan,
738 struct dma_client *client)
739 {
740 struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
741 struct ioat_desc_sw *desc;
742 u16 chanctrl;
743 u32 chanerr;
744 int i;
745 LIST_HEAD(tmp_list);
746
747 /* have we already been set up? */
748 if (!list_empty(&ioat_chan->free_desc))
749 return ioat_chan->desccount;
750
751 /* Setup register to interrupt and write completion status on error */
752 chanctrl = IOAT_CHANCTRL_ERR_INT_EN |
753 IOAT_CHANCTRL_ANY_ERR_ABORT_EN |
754 IOAT_CHANCTRL_ERR_COMPLETION_EN;
755 writew(chanctrl, ioat_chan->reg_base + IOAT_CHANCTRL_OFFSET);
756
757 chanerr = readl(ioat_chan->reg_base + IOAT_CHANERR_OFFSET);
758 if (chanerr) {
759 dev_err(&ioat_chan->device->pdev->dev,
760 "CHANERR = %x, clearing\n", chanerr);
761 writel(chanerr, ioat_chan->reg_base + IOAT_CHANERR_OFFSET);
762 }
763
764 /* Allocate descriptors */
765 for (i = 0; i < ioat_initial_desc_count; i++) {
766 desc = ioat_dma_alloc_descriptor(ioat_chan, GFP_KERNEL);
767 if (!desc) {
768 dev_err(&ioat_chan->device->pdev->dev,
769 "Only %d initial descriptors\n", i);
770 break;
771 }
772 list_add_tail(&desc->node, &tmp_list);
773 }
774 spin_lock_bh(&ioat_chan->desc_lock);
775 ioat_chan->desccount = i;
776 list_splice(&tmp_list, &ioat_chan->free_desc);
777 if (ioat_chan->device->version != IOAT_VER_1_2)
778 ioat2_dma_massage_chan_desc(ioat_chan);
779 spin_unlock_bh(&ioat_chan->desc_lock);
780
781 /* allocate a completion writeback area */
782 /* doing 2 32bit writes to mmio since 1 64b write doesn't work */
783 ioat_chan->completion_virt =
784 pci_pool_alloc(ioat_chan->device->completion_pool,
785 GFP_KERNEL,
786 &ioat_chan->completion_addr);
787 memset(ioat_chan->completion_virt, 0,
788 sizeof(*ioat_chan->completion_virt));
789 writel(((u64) ioat_chan->completion_addr) & 0x00000000FFFFFFFF,
790 ioat_chan->reg_base + IOAT_CHANCMP_OFFSET_LOW);
791 writel(((u64) ioat_chan->completion_addr) >> 32,
792 ioat_chan->reg_base + IOAT_CHANCMP_OFFSET_HIGH);
793
794 tasklet_enable(&ioat_chan->cleanup_task);
795 ioat_dma_start_null_desc(ioat_chan); /* give chain to dma device */
796 return ioat_chan->desccount;
797 }
798
799 /**
800 * ioat_dma_free_chan_resources - release all the descriptors
801 * @chan: the channel to be cleaned
802 */
803 static void ioat_dma_free_chan_resources(struct dma_chan *chan)
804 {
805 struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
806 struct ioatdma_device *ioatdma_device = to_ioatdma_device(chan->device);
807 struct ioat_desc_sw *desc, *_desc;
808 int in_use_descs = 0;
809
810 tasklet_disable(&ioat_chan->cleanup_task);
811 ioat_dma_memcpy_cleanup(ioat_chan);
812
813 /* Delay 100ms after reset to allow internal DMA logic to quiesce
814 * before removing DMA descriptor resources.
815 */
816 writeb(IOAT_CHANCMD_RESET,
817 ioat_chan->reg_base
818 + IOAT_CHANCMD_OFFSET(ioat_chan->device->version));
819 mdelay(100);
820
821 spin_lock_bh(&ioat_chan->desc_lock);
822 switch (ioat_chan->device->version) {
823 case IOAT_VER_1_2:
824 list_for_each_entry_safe(desc, _desc,
825 &ioat_chan->used_desc, node) {
826 in_use_descs++;
827 list_del(&desc->node);
828 pci_pool_free(ioatdma_device->dma_pool, desc->hw,
829 desc->async_tx.phys);
830 kfree(desc);
831 }
832 list_for_each_entry_safe(desc, _desc,
833 &ioat_chan->free_desc, node) {
834 list_del(&desc->node);
835 pci_pool_free(ioatdma_device->dma_pool, desc->hw,
836 desc->async_tx.phys);
837 kfree(desc);
838 }
839 break;
840 case IOAT_VER_2_0:
841 case IOAT_VER_3_0:
842 list_for_each_entry_safe(desc, _desc,
843 ioat_chan->free_desc.next, node) {
844 list_del(&desc->node);
845 pci_pool_free(ioatdma_device->dma_pool, desc->hw,
846 desc->async_tx.phys);
847 kfree(desc);
848 }
849 desc = to_ioat_desc(ioat_chan->free_desc.next);
850 pci_pool_free(ioatdma_device->dma_pool, desc->hw,
851 desc->async_tx.phys);
852 kfree(desc);
853 INIT_LIST_HEAD(&ioat_chan->free_desc);
854 INIT_LIST_HEAD(&ioat_chan->used_desc);
855 break;
856 }
857 spin_unlock_bh(&ioat_chan->desc_lock);
858
859 pci_pool_free(ioatdma_device->completion_pool,
860 ioat_chan->completion_virt,
861 ioat_chan->completion_addr);
862
863 /* one is ok since we left it on there on purpose */
864 if (in_use_descs > 1)
865 dev_err(&ioat_chan->device->pdev->dev,
866 "Freeing %d in use descriptors!\n",
867 in_use_descs - 1);
868
869 ioat_chan->last_completion = ioat_chan->completion_addr = 0;
870 ioat_chan->pending = 0;
871 ioat_chan->dmacount = 0;
872 ioat_chan->watchdog_completion = 0;
873 ioat_chan->last_compl_desc_addr_hw = 0;
874 ioat_chan->watchdog_tcp_cookie =
875 ioat_chan->watchdog_last_tcp_cookie = 0;
876 }
877
878 /**
879 * ioat_dma_get_next_descriptor - return the next available descriptor
880 * @ioat_chan: IOAT DMA channel handle
881 *
882 * Gets the next descriptor from the chain, and must be called with the
883 * channel's desc_lock held. Allocates more descriptors if the channel
884 * has run out.
885 */
886 static struct ioat_desc_sw *
887 ioat1_dma_get_next_descriptor(struct ioat_dma_chan *ioat_chan)
888 {
889 struct ioat_desc_sw *new;
890
891 if (!list_empty(&ioat_chan->free_desc)) {
892 new = to_ioat_desc(ioat_chan->free_desc.next);
893 list_del(&new->node);
894 } else {
895 /* try to get another desc */
896 new = ioat_dma_alloc_descriptor(ioat_chan, GFP_ATOMIC);
897 if (!new) {
898 dev_err(&ioat_chan->device->pdev->dev,
899 "alloc failed\n");
900 return NULL;
901 }
902 }
903
904 prefetch(new->hw);
905 return new;
906 }
907
908 static struct ioat_desc_sw *
909 ioat2_dma_get_next_descriptor(struct ioat_dma_chan *ioat_chan)
910 {
911 struct ioat_desc_sw *new;
912
913 /*
914 * used.prev points to where to start processing
915 * used.next points to next free descriptor
916 * if used.prev == NULL, there are none waiting to be processed
917 * if used.next == used.prev.prev, there is only one free descriptor,
918 * and we need to use it to as a noop descriptor before
919 * linking in a new set of descriptors, since the device
920 * has probably already read the pointer to it
921 */
922 if (ioat_chan->used_desc.prev &&
923 ioat_chan->used_desc.next == ioat_chan->used_desc.prev->prev) {
924
925 struct ioat_desc_sw *desc;
926 struct ioat_desc_sw *noop_desc;
927 int i;
928
929 /* set up the noop descriptor */
930 noop_desc = to_ioat_desc(ioat_chan->used_desc.next);
931 /* set size to non-zero value (channel returns error when size is 0) */
932 noop_desc->hw->size = NULL_DESC_BUFFER_SIZE;
933 noop_desc->hw->ctl = IOAT_DMA_DESCRIPTOR_NUL;
934 noop_desc->hw->src_addr = 0;
935 noop_desc->hw->dst_addr = 0;
936
937 ioat_chan->used_desc.next = ioat_chan->used_desc.next->next;
938 ioat_chan->pending++;
939 ioat_chan->dmacount++;
940
941 /* try to get a few more descriptors */
942 for (i = 16; i; i--) {
943 desc = ioat_dma_alloc_descriptor(ioat_chan, GFP_ATOMIC);
944 if (!desc) {
945 dev_err(&ioat_chan->device->pdev->dev,
946 "alloc failed\n");
947 break;
948 }
949 list_add_tail(&desc->node, ioat_chan->used_desc.next);
950
951 desc->hw->next
952 = to_ioat_desc(desc->node.next)->async_tx.phys;
953 to_ioat_desc(desc->node.prev)->hw->next
954 = desc->async_tx.phys;
955 ioat_chan->desccount++;
956 }
957
958 ioat_chan->used_desc.next = noop_desc->node.next;
959 }
960 new = to_ioat_desc(ioat_chan->used_desc.next);
961 prefetch(new);
962 ioat_chan->used_desc.next = new->node.next;
963
964 if (ioat_chan->used_desc.prev == NULL)
965 ioat_chan->used_desc.prev = &new->node;
966
967 prefetch(new->hw);
968 return new;
969 }
970
971 static struct ioat_desc_sw *ioat_dma_get_next_descriptor(
972 struct ioat_dma_chan *ioat_chan)
973 {
974 if (!ioat_chan)
975 return NULL;
976
977 switch (ioat_chan->device->version) {
978 case IOAT_VER_1_2:
979 return ioat1_dma_get_next_descriptor(ioat_chan);
980 case IOAT_VER_2_0:
981 case IOAT_VER_3_0:
982 return ioat2_dma_get_next_descriptor(ioat_chan);
983 }
984 return NULL;
985 }
986
987 static struct dma_async_tx_descriptor *ioat1_dma_prep_memcpy(
988 struct dma_chan *chan,
989 dma_addr_t dma_dest,
990 dma_addr_t dma_src,
991 size_t len,
992 unsigned long flags)
993 {
994 struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
995 struct ioat_desc_sw *new;
996
997 spin_lock_bh(&ioat_chan->desc_lock);
998 new = ioat_dma_get_next_descriptor(ioat_chan);
999 spin_unlock_bh(&ioat_chan->desc_lock);
1000
1001 if (new) {
1002 new->len = len;
1003 new->dst = dma_dest;
1004 new->src = dma_src;
1005 new->async_tx.flags = flags;
1006 return &new->async_tx;
1007 } else {
1008 dev_err(&ioat_chan->device->pdev->dev,
1009 "chan%d - get_next_desc failed: %d descs waiting, %d total desc\n",
1010 chan_num(ioat_chan), ioat_chan->dmacount, ioat_chan->desccount);
1011 return NULL;
1012 }
1013 }
1014
1015 static struct dma_async_tx_descriptor *ioat2_dma_prep_memcpy(
1016 struct dma_chan *chan,
1017 dma_addr_t dma_dest,
1018 dma_addr_t dma_src,
1019 size_t len,
1020 unsigned long flags)
1021 {
1022 struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
1023 struct ioat_desc_sw *new;
1024
1025 spin_lock_bh(&ioat_chan->desc_lock);
1026 new = ioat2_dma_get_next_descriptor(ioat_chan);
1027
1028 /*
1029 * leave ioat_chan->desc_lock set in ioat 2 path
1030 * it will get unlocked at end of tx_submit
1031 */
1032
1033 if (new) {
1034 new->len = len;
1035 new->dst = dma_dest;
1036 new->src = dma_src;
1037 new->async_tx.flags = flags;
1038 return &new->async_tx;
1039 } else {
1040 spin_unlock_bh(&ioat_chan->desc_lock);
1041 dev_err(&ioat_chan->device->pdev->dev,
1042 "chan%d - get_next_desc failed: %d descs waiting, %d total desc\n",
1043 chan_num(ioat_chan), ioat_chan->dmacount, ioat_chan->desccount);
1044 return NULL;
1045 }
1046 }
1047
1048 static void ioat_dma_cleanup_tasklet(unsigned long data)
1049 {
1050 struct ioat_dma_chan *chan = (void *)data;
1051 ioat_dma_memcpy_cleanup(chan);
1052 writew(IOAT_CHANCTRL_INT_DISABLE,
1053 chan->reg_base + IOAT_CHANCTRL_OFFSET);
1054 }
1055
1056 static void
1057 ioat_dma_unmap(struct ioat_dma_chan *ioat_chan, struct ioat_desc_sw *desc)
1058 {
1059 /*
1060 * yes we are unmapping both _page and _single
1061 * alloc'd regions with unmap_page. Is this
1062 * *really* that bad?
1063 */
1064 if (!(desc->async_tx.flags & DMA_COMPL_SKIP_DEST_UNMAP))
1065 pci_unmap_page(ioat_chan->device->pdev,
1066 pci_unmap_addr(desc, dst),
1067 pci_unmap_len(desc, len),
1068 PCI_DMA_FROMDEVICE);
1069
1070 if (!(desc->async_tx.flags & DMA_COMPL_SKIP_SRC_UNMAP))
1071 pci_unmap_page(ioat_chan->device->pdev,
1072 pci_unmap_addr(desc, src),
1073 pci_unmap_len(desc, len),
1074 PCI_DMA_TODEVICE);
1075 }
1076
1077 /**
1078 * ioat_dma_memcpy_cleanup - cleanup up finished descriptors
1079 * @chan: ioat channel to be cleaned up
1080 */
1081 static void ioat_dma_memcpy_cleanup(struct ioat_dma_chan *ioat_chan)
1082 {
1083 unsigned long phys_complete;
1084 struct ioat_desc_sw *desc, *_desc;
1085 dma_cookie_t cookie = 0;
1086 unsigned long desc_phys;
1087 struct ioat_desc_sw *latest_desc;
1088
1089 prefetch(ioat_chan->completion_virt);
1090
1091 if (!spin_trylock_bh(&ioat_chan->cleanup_lock))
1092 return;
1093
1094 /* The completion writeback can happen at any time,
1095 so reads by the driver need to be atomic operations
1096 The descriptor physical addresses are limited to 32-bits
1097 when the CPU can only do a 32-bit mov */
1098
1099 #if (BITS_PER_LONG == 64)
1100 phys_complete =
1101 ioat_chan->completion_virt->full
1102 & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_ADDR;
1103 #else
1104 phys_complete =
1105 ioat_chan->completion_virt->low & IOAT_LOW_COMPLETION_MASK;
1106 #endif
1107
1108 if ((ioat_chan->completion_virt->full
1109 & IOAT_CHANSTS_DMA_TRANSFER_STATUS) ==
1110 IOAT_CHANSTS_DMA_TRANSFER_STATUS_HALTED) {
1111 dev_err(&ioat_chan->device->pdev->dev,
1112 "Channel halted, chanerr = %x\n",
1113 readl(ioat_chan->reg_base + IOAT_CHANERR_OFFSET));
1114
1115 /* TODO do something to salvage the situation */
1116 }
1117
1118 if (phys_complete == ioat_chan->last_completion) {
1119 spin_unlock_bh(&ioat_chan->cleanup_lock);
1120 /*
1121 * perhaps we're stuck so hard that the watchdog can't go off?
1122 * try to catch it after 2 seconds
1123 */
1124 if (ioat_chan->device->version != IOAT_VER_3_0) {
1125 if (time_after(jiffies,
1126 ioat_chan->last_completion_time + HZ*WATCHDOG_DELAY)) {
1127 ioat_dma_chan_watchdog(&(ioat_chan->device->work.work));
1128 ioat_chan->last_completion_time = jiffies;
1129 }
1130 }
1131 return;
1132 }
1133 ioat_chan->last_completion_time = jiffies;
1134
1135 cookie = 0;
1136 if (!spin_trylock_bh(&ioat_chan->desc_lock)) {
1137 spin_unlock_bh(&ioat_chan->cleanup_lock);
1138 return;
1139 }
1140
1141 switch (ioat_chan->device->version) {
1142 case IOAT_VER_1_2:
1143 list_for_each_entry_safe(desc, _desc,
1144 &ioat_chan->used_desc, node) {
1145
1146 /*
1147 * Incoming DMA requests may use multiple descriptors,
1148 * due to exceeding xfercap, perhaps. If so, only the
1149 * last one will have a cookie, and require unmapping.
1150 */
1151 if (desc->async_tx.cookie) {
1152 cookie = desc->async_tx.cookie;
1153 ioat_dma_unmap(ioat_chan, desc);
1154 if (desc->async_tx.callback) {
1155 desc->async_tx.callback(desc->async_tx.callback_param);
1156 desc->async_tx.callback = NULL;
1157 }
1158 }
1159
1160 if (desc->async_tx.phys != phys_complete) {
1161 /*
1162 * a completed entry, but not the last, so clean
1163 * up if the client is done with the descriptor
1164 */
1165 if (async_tx_test_ack(&desc->async_tx)) {
1166 list_del(&desc->node);
1167 list_add_tail(&desc->node,
1168 &ioat_chan->free_desc);
1169 } else
1170 desc->async_tx.cookie = 0;
1171 } else {
1172 /*
1173 * last used desc. Do not remove, so we can
1174 * append from it, but don't look at it next
1175 * time, either
1176 */
1177 desc->async_tx.cookie = 0;
1178
1179 /* TODO check status bits? */
1180 break;
1181 }
1182 }
1183 break;
1184 case IOAT_VER_2_0:
1185 case IOAT_VER_3_0:
1186 /* has some other thread has already cleaned up? */
1187 if (ioat_chan->used_desc.prev == NULL)
1188 break;
1189
1190 /* work backwards to find latest finished desc */
1191 desc = to_ioat_desc(ioat_chan->used_desc.next);
1192 latest_desc = NULL;
1193 do {
1194 desc = to_ioat_desc(desc->node.prev);
1195 desc_phys = (unsigned long)desc->async_tx.phys
1196 & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_ADDR;
1197 if (desc_phys == phys_complete) {
1198 latest_desc = desc;
1199 break;
1200 }
1201 } while (&desc->node != ioat_chan->used_desc.prev);
1202
1203 if (latest_desc != NULL) {
1204
1205 /* work forwards to clear finished descriptors */
1206 for (desc = to_ioat_desc(ioat_chan->used_desc.prev);
1207 &desc->node != latest_desc->node.next &&
1208 &desc->node != ioat_chan->used_desc.next;
1209 desc = to_ioat_desc(desc->node.next)) {
1210 if (desc->async_tx.cookie) {
1211 cookie = desc->async_tx.cookie;
1212 desc->async_tx.cookie = 0;
1213 ioat_dma_unmap(ioat_chan, desc);
1214 if (desc->async_tx.callback) {
1215 desc->async_tx.callback(desc->async_tx.callback_param);
1216 desc->async_tx.callback = NULL;
1217 }
1218 }
1219 }
1220
1221 /* move used.prev up beyond those that are finished */
1222 if (&desc->node == ioat_chan->used_desc.next)
1223 ioat_chan->used_desc.prev = NULL;
1224 else
1225 ioat_chan->used_desc.prev = &desc->node;
1226 }
1227 break;
1228 }
1229
1230 spin_unlock_bh(&ioat_chan->desc_lock);
1231
1232 ioat_chan->last_completion = phys_complete;
1233 if (cookie != 0)
1234 ioat_chan->completed_cookie = cookie;
1235
1236 spin_unlock_bh(&ioat_chan->cleanup_lock);
1237 }
1238
1239 /**
1240 * ioat_dma_is_complete - poll the status of a IOAT DMA transaction
1241 * @chan: IOAT DMA channel handle
1242 * @cookie: DMA transaction identifier
1243 * @done: if not %NULL, updated with last completed transaction
1244 * @used: if not %NULL, updated with last used transaction
1245 */
1246 static enum dma_status ioat_dma_is_complete(struct dma_chan *chan,
1247 dma_cookie_t cookie,
1248 dma_cookie_t *done,
1249 dma_cookie_t *used)
1250 {
1251 struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
1252 dma_cookie_t last_used;
1253 dma_cookie_t last_complete;
1254 enum dma_status ret;
1255
1256 last_used = chan->cookie;
1257 last_complete = ioat_chan->completed_cookie;
1258 ioat_chan->watchdog_tcp_cookie = cookie;
1259
1260 if (done)
1261 *done = last_complete;
1262 if (used)
1263 *used = last_used;
1264
1265 ret = dma_async_is_complete(cookie, last_complete, last_used);
1266 if (ret == DMA_SUCCESS)
1267 return ret;
1268
1269 ioat_dma_memcpy_cleanup(ioat_chan);
1270
1271 last_used = chan->cookie;
1272 last_complete = ioat_chan->completed_cookie;
1273
1274 if (done)
1275 *done = last_complete;
1276 if (used)
1277 *used = last_used;
1278
1279 return dma_async_is_complete(cookie, last_complete, last_used);
1280 }
1281
1282 static void ioat_dma_start_null_desc(struct ioat_dma_chan *ioat_chan)
1283 {
1284 struct ioat_desc_sw *desc;
1285
1286 spin_lock_bh(&ioat_chan->desc_lock);
1287
1288 desc = ioat_dma_get_next_descriptor(ioat_chan);
1289
1290 if (!desc) {
1291 dev_err(&ioat_chan->device->pdev->dev,
1292 "Unable to start null desc - get next desc failed\n");
1293 spin_unlock_bh(&ioat_chan->desc_lock);
1294 return;
1295 }
1296
1297 desc->hw->ctl = IOAT_DMA_DESCRIPTOR_NUL
1298 | IOAT_DMA_DESCRIPTOR_CTL_INT_GN
1299 | IOAT_DMA_DESCRIPTOR_CTL_CP_STS;
1300 /* set size to non-zero value (channel returns error when size is 0) */
1301 desc->hw->size = NULL_DESC_BUFFER_SIZE;
1302 desc->hw->src_addr = 0;
1303 desc->hw->dst_addr = 0;
1304 async_tx_ack(&desc->async_tx);
1305 switch (ioat_chan->device->version) {
1306 case IOAT_VER_1_2:
1307 desc->hw->next = 0;
1308 list_add_tail(&desc->node, &ioat_chan->used_desc);
1309
1310 writel(((u64) desc->async_tx.phys) & 0x00000000FFFFFFFF,
1311 ioat_chan->reg_base + IOAT1_CHAINADDR_OFFSET_LOW);
1312 writel(((u64) desc->async_tx.phys) >> 32,
1313 ioat_chan->reg_base + IOAT1_CHAINADDR_OFFSET_HIGH);
1314
1315 writeb(IOAT_CHANCMD_START, ioat_chan->reg_base
1316 + IOAT_CHANCMD_OFFSET(ioat_chan->device->version));
1317 break;
1318 case IOAT_VER_2_0:
1319 case IOAT_VER_3_0:
1320 writel(((u64) desc->async_tx.phys) & 0x00000000FFFFFFFF,
1321 ioat_chan->reg_base + IOAT2_CHAINADDR_OFFSET_LOW);
1322 writel(((u64) desc->async_tx.phys) >> 32,
1323 ioat_chan->reg_base + IOAT2_CHAINADDR_OFFSET_HIGH);
1324
1325 ioat_chan->dmacount++;
1326 __ioat2_dma_memcpy_issue_pending(ioat_chan);
1327 break;
1328 }
1329 spin_unlock_bh(&ioat_chan->desc_lock);
1330 }
1331
1332 /*
1333 * Perform a IOAT transaction to verify the HW works.
1334 */
1335 #define IOAT_TEST_SIZE 2000
1336
1337 static void ioat_dma_test_callback(void *dma_async_param)
1338 {
1339 printk(KERN_ERR "ioatdma: ioat_dma_test_callback(%p)\n",
1340 dma_async_param);
1341 }
1342
1343 /**
1344 * ioat_dma_self_test - Perform a IOAT transaction to verify the HW works.
1345 * @device: device to be tested
1346 */
1347 static int ioat_dma_self_test(struct ioatdma_device *device)
1348 {
1349 int i;
1350 u8 *src;
1351 u8 *dest;
1352 struct dma_chan *dma_chan;
1353 struct dma_async_tx_descriptor *tx;
1354 dma_addr_t dma_dest, dma_src;
1355 dma_cookie_t cookie;
1356 int err = 0;
1357
1358 src = kzalloc(sizeof(u8) * IOAT_TEST_SIZE, GFP_KERNEL);
1359 if (!src)
1360 return -ENOMEM;
1361 dest = kzalloc(sizeof(u8) * IOAT_TEST_SIZE, GFP_KERNEL);
1362 if (!dest) {
1363 kfree(src);
1364 return -ENOMEM;
1365 }
1366
1367 /* Fill in src buffer */
1368 for (i = 0; i < IOAT_TEST_SIZE; i++)
1369 src[i] = (u8)i;
1370
1371 /* Start copy, using first DMA channel */
1372 dma_chan = container_of(device->common.channels.next,
1373 struct dma_chan,
1374 device_node);
1375 if (device->common.device_alloc_chan_resources(dma_chan, NULL) < 1) {
1376 dev_err(&device->pdev->dev,
1377 "selftest cannot allocate chan resource\n");
1378 err = -ENODEV;
1379 goto out;
1380 }
1381
1382 dma_src = dma_map_single(dma_chan->device->dev, src, IOAT_TEST_SIZE,
1383 DMA_TO_DEVICE);
1384 dma_dest = dma_map_single(dma_chan->device->dev, dest, IOAT_TEST_SIZE,
1385 DMA_FROM_DEVICE);
1386 tx = device->common.device_prep_dma_memcpy(dma_chan, dma_dest, dma_src,
1387 IOAT_TEST_SIZE, 0);
1388 if (!tx) {
1389 dev_err(&device->pdev->dev,
1390 "Self-test prep failed, disabling\n");
1391 err = -ENODEV;
1392 goto free_resources;
1393 }
1394
1395 async_tx_ack(tx);
1396 tx->callback = ioat_dma_test_callback;
1397 tx->callback_param = (void *)0x8086;
1398 cookie = tx->tx_submit(tx);
1399 if (cookie < 0) {
1400 dev_err(&device->pdev->dev,
1401 "Self-test setup failed, disabling\n");
1402 err = -ENODEV;
1403 goto free_resources;
1404 }
1405 device->common.device_issue_pending(dma_chan);
1406 msleep(1);
1407
1408 if (device->common.device_is_tx_complete(dma_chan, cookie, NULL, NULL)
1409 != DMA_SUCCESS) {
1410 dev_err(&device->pdev->dev,
1411 "Self-test copy timed out, disabling\n");
1412 err = -ENODEV;
1413 goto free_resources;
1414 }
1415 if (memcmp(src, dest, IOAT_TEST_SIZE)) {
1416 dev_err(&device->pdev->dev,
1417 "Self-test copy failed compare, disabling\n");
1418 err = -ENODEV;
1419 goto free_resources;
1420 }
1421
1422 free_resources:
1423 device->common.device_free_chan_resources(dma_chan);
1424 out:
1425 kfree(src);
1426 kfree(dest);
1427 return err;
1428 }
1429
1430 static char ioat_interrupt_style[32] = "msix";
1431 module_param_string(ioat_interrupt_style, ioat_interrupt_style,
1432 sizeof(ioat_interrupt_style), 0644);
1433 MODULE_PARM_DESC(ioat_interrupt_style,
1434 "set ioat interrupt style: msix (default), "
1435 "msix-single-vector, msi, intx)");
1436
1437 /**
1438 * ioat_dma_setup_interrupts - setup interrupt handler
1439 * @device: ioat device
1440 */
1441 static int ioat_dma_setup_interrupts(struct ioatdma_device *device)
1442 {
1443 struct ioat_dma_chan *ioat_chan;
1444 int err, i, j, msixcnt;
1445 u8 intrctrl = 0;
1446
1447 if (!strcmp(ioat_interrupt_style, "msix"))
1448 goto msix;
1449 if (!strcmp(ioat_interrupt_style, "msix-single-vector"))
1450 goto msix_single_vector;
1451 if (!strcmp(ioat_interrupt_style, "msi"))
1452 goto msi;
1453 if (!strcmp(ioat_interrupt_style, "intx"))
1454 goto intx;
1455 dev_err(&device->pdev->dev, "invalid ioat_interrupt_style %s\n",
1456 ioat_interrupt_style);
1457 goto err_no_irq;
1458
1459 msix:
1460 /* The number of MSI-X vectors should equal the number of channels */
1461 msixcnt = device->common.chancnt;
1462 for (i = 0; i < msixcnt; i++)
1463 device->msix_entries[i].entry = i;
1464
1465 err = pci_enable_msix(device->pdev, device->msix_entries, msixcnt);
1466 if (err < 0)
1467 goto msi;
1468 if (err > 0)
1469 goto msix_single_vector;
1470
1471 for (i = 0; i < msixcnt; i++) {
1472 ioat_chan = ioat_lookup_chan_by_index(device, i);
1473 err = request_irq(device->msix_entries[i].vector,
1474 ioat_dma_do_interrupt_msix,
1475 0, "ioat-msix", ioat_chan);
1476 if (err) {
1477 for (j = 0; j < i; j++) {
1478 ioat_chan =
1479 ioat_lookup_chan_by_index(device, j);
1480 free_irq(device->msix_entries[j].vector,
1481 ioat_chan);
1482 }
1483 goto msix_single_vector;
1484 }
1485 }
1486 intrctrl |= IOAT_INTRCTRL_MSIX_VECTOR_CONTROL;
1487 device->irq_mode = msix_multi_vector;
1488 goto done;
1489
1490 msix_single_vector:
1491 device->msix_entries[0].entry = 0;
1492 err = pci_enable_msix(device->pdev, device->msix_entries, 1);
1493 if (err)
1494 goto msi;
1495
1496 err = request_irq(device->msix_entries[0].vector, ioat_dma_do_interrupt,
1497 0, "ioat-msix", device);
1498 if (err) {
1499 pci_disable_msix(device->pdev);
1500 goto msi;
1501 }
1502 device->irq_mode = msix_single_vector;
1503 goto done;
1504
1505 msi:
1506 err = pci_enable_msi(device->pdev);
1507 if (err)
1508 goto intx;
1509
1510 err = request_irq(device->pdev->irq, ioat_dma_do_interrupt,
1511 0, "ioat-msi", device);
1512 if (err) {
1513 pci_disable_msi(device->pdev);
1514 goto intx;
1515 }
1516 /*
1517 * CB 1.2 devices need a bit set in configuration space to enable MSI
1518 */
1519 if (device->version == IOAT_VER_1_2) {
1520 u32 dmactrl;
1521 pci_read_config_dword(device->pdev,
1522 IOAT_PCI_DMACTRL_OFFSET, &dmactrl);
1523 dmactrl |= IOAT_PCI_DMACTRL_MSI_EN;
1524 pci_write_config_dword(device->pdev,
1525 IOAT_PCI_DMACTRL_OFFSET, dmactrl);
1526 }
1527 device->irq_mode = msi;
1528 goto done;
1529
1530 intx:
1531 err = request_irq(device->pdev->irq, ioat_dma_do_interrupt,
1532 IRQF_SHARED, "ioat-intx", device);
1533 if (err)
1534 goto err_no_irq;
1535 device->irq_mode = intx;
1536
1537 done:
1538 intrctrl |= IOAT_INTRCTRL_MASTER_INT_EN;
1539 writeb(intrctrl, device->reg_base + IOAT_INTRCTRL_OFFSET);
1540 return 0;
1541
1542 err_no_irq:
1543 /* Disable all interrupt generation */
1544 writeb(0, device->reg_base + IOAT_INTRCTRL_OFFSET);
1545 dev_err(&device->pdev->dev, "no usable interrupts\n");
1546 device->irq_mode = none;
1547 return -1;
1548 }
1549
1550 /**
1551 * ioat_dma_remove_interrupts - remove whatever interrupts were set
1552 * @device: ioat device
1553 */
1554 static void ioat_dma_remove_interrupts(struct ioatdma_device *device)
1555 {
1556 struct ioat_dma_chan *ioat_chan;
1557 int i;
1558
1559 /* Disable all interrupt generation */
1560 writeb(0, device->reg_base + IOAT_INTRCTRL_OFFSET);
1561
1562 switch (device->irq_mode) {
1563 case msix_multi_vector:
1564 for (i = 0; i < device->common.chancnt; i++) {
1565 ioat_chan = ioat_lookup_chan_by_index(device, i);
1566 free_irq(device->msix_entries[i].vector, ioat_chan);
1567 }
1568 pci_disable_msix(device->pdev);
1569 break;
1570 case msix_single_vector:
1571 free_irq(device->msix_entries[0].vector, device);
1572 pci_disable_msix(device->pdev);
1573 break;
1574 case msi:
1575 free_irq(device->pdev->irq, device);
1576 pci_disable_msi(device->pdev);
1577 break;
1578 case intx:
1579 free_irq(device->pdev->irq, device);
1580 break;
1581 case none:
1582 dev_warn(&device->pdev->dev,
1583 "call to %s without interrupts setup\n", __func__);
1584 }
1585 device->irq_mode = none;
1586 }
1587
1588 struct ioatdma_device *ioat_dma_probe(struct pci_dev *pdev,
1589 void __iomem *iobase)
1590 {
1591 int err;
1592 struct ioatdma_device *device;
1593
1594 device = kzalloc(sizeof(*device), GFP_KERNEL);
1595 if (!device) {
1596 err = -ENOMEM;
1597 goto err_kzalloc;
1598 }
1599 device->pdev = pdev;
1600 device->reg_base = iobase;
1601 device->version = readb(device->reg_base + IOAT_VER_OFFSET);
1602
1603 /* DMA coherent memory pool for DMA descriptor allocations */
1604 device->dma_pool = pci_pool_create("dma_desc_pool", pdev,
1605 sizeof(struct ioat_dma_descriptor),
1606 64, 0);
1607 if (!device->dma_pool) {
1608 err = -ENOMEM;
1609 goto err_dma_pool;
1610 }
1611
1612 device->completion_pool = pci_pool_create("completion_pool", pdev,
1613 sizeof(u64), SMP_CACHE_BYTES,
1614 SMP_CACHE_BYTES);
1615 if (!device->completion_pool) {
1616 err = -ENOMEM;
1617 goto err_completion_pool;
1618 }
1619
1620 INIT_LIST_HEAD(&device->common.channels);
1621 ioat_dma_enumerate_channels(device);
1622
1623 device->common.device_alloc_chan_resources =
1624 ioat_dma_alloc_chan_resources;
1625 device->common.device_free_chan_resources =
1626 ioat_dma_free_chan_resources;
1627 device->common.dev = &pdev->dev;
1628
1629 dma_cap_set(DMA_MEMCPY, device->common.cap_mask);
1630 device->common.device_is_tx_complete = ioat_dma_is_complete;
1631 switch (device->version) {
1632 case IOAT_VER_1_2:
1633 device->common.device_prep_dma_memcpy = ioat1_dma_prep_memcpy;
1634 device->common.device_issue_pending =
1635 ioat1_dma_memcpy_issue_pending;
1636 break;
1637 case IOAT_VER_2_0:
1638 case IOAT_VER_3_0:
1639 device->common.device_prep_dma_memcpy = ioat2_dma_prep_memcpy;
1640 device->common.device_issue_pending =
1641 ioat2_dma_memcpy_issue_pending;
1642 break;
1643 }
1644
1645 dev_err(&device->pdev->dev,
1646 "Intel(R) I/OAT DMA Engine found,"
1647 " %d channels, device version 0x%02x, driver version %s\n",
1648 device->common.chancnt, device->version, IOAT_DMA_VERSION);
1649
1650 err = ioat_dma_setup_interrupts(device);
1651 if (err)
1652 goto err_setup_interrupts;
1653
1654 err = ioat_dma_self_test(device);
1655 if (err)
1656 goto err_self_test;
1657
1658 ioat_set_tcp_copy_break(device);
1659
1660 dma_async_device_register(&device->common);
1661
1662 if (device->version != IOAT_VER_3_0) {
1663 INIT_DELAYED_WORK(&device->work, ioat_dma_chan_watchdog);
1664 schedule_delayed_work(&device->work,
1665 WATCHDOG_DELAY);
1666 }
1667
1668 return device;
1669
1670 err_self_test:
1671 ioat_dma_remove_interrupts(device);
1672 err_setup_interrupts:
1673 pci_pool_destroy(device->completion_pool);
1674 err_completion_pool:
1675 pci_pool_destroy(device->dma_pool);
1676 err_dma_pool:
1677 kfree(device);
1678 err_kzalloc:
1679 dev_err(&pdev->dev,
1680 "Intel(R) I/OAT DMA Engine initialization failed\n");
1681 return NULL;
1682 }
1683
1684 void ioat_dma_remove(struct ioatdma_device *device)
1685 {
1686 struct dma_chan *chan, *_chan;
1687 struct ioat_dma_chan *ioat_chan;
1688
1689 ioat_dma_remove_interrupts(device);
1690
1691 dma_async_device_unregister(&device->common);
1692
1693 pci_pool_destroy(device->dma_pool);
1694 pci_pool_destroy(device->completion_pool);
1695
1696 iounmap(device->reg_base);
1697 pci_release_regions(device->pdev);
1698 pci_disable_device(device->pdev);
1699
1700 if (device->version != IOAT_VER_3_0) {
1701 cancel_delayed_work(&device->work);
1702 }
1703
1704 list_for_each_entry_safe(chan, _chan,
1705 &device->common.channels, device_node) {
1706 ioat_chan = to_ioat_chan(chan);
1707 list_del(&chan->device_node);
1708 kfree(ioat_chan);
1709 }
1710 kfree(device);
1711 }
1712
Support for the Chinese Samsung S3C2440 based development boards