?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
I/OAT: I/OAT version 3.0 support
[linux-2.6.git] / drivers / dma / ioat_dma.c
bloba52156e568867e827c30f0008de9477df493b730
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 "ioatdma.h"
37 #include "ioatdma_registers.h"
38 #include "ioatdma_hw.h"
39
40 #define to_ioat_chan(chan) container_of(chan, struct ioat_dma_chan, common)
41 #define to_ioatdma_device(dev) container_of(dev, struct ioatdma_device, common)
42 #define to_ioat_desc(lh) container_of(lh, struct ioat_desc_sw, node)
43 #define tx_to_ioat_desc(tx) container_of(tx, struct ioat_desc_sw, async_tx)
44
45 #define chan_num(ch) ((int)((ch)->reg_base - (ch)->device->reg_base) / 0x80)
46 static int ioat_pending_level = 4;
47 module_param(ioat_pending_level, int, 0644);
48 MODULE_PARM_DESC(ioat_pending_level,
49 "high-water mark for pushing ioat descriptors (default: 4)");
50
51 #define RESET_DELAY msecs_to_jiffies(100)
52 #define WATCHDOG_DELAY round_jiffies(msecs_to_jiffies(2000))
53 static void ioat_dma_chan_reset_part2(struct work_struct *work);
54 static void ioat_dma_chan_watchdog(struct work_struct *work);
55
56 /*
57 * workaround for IOAT ver.3.0 null descriptor issue
58 * (channel returns error when size is 0)
59 */
60 #define NULL_DESC_BUFFER_SIZE 1
61
62 /* internal functions */
63 static void ioat_dma_start_null_desc(struct ioat_dma_chan *ioat_chan);
64 static void ioat_dma_memcpy_cleanup(struct ioat_dma_chan *ioat_chan);
65
66 static struct ioat_desc_sw *
67 ioat1_dma_get_next_descriptor(struct ioat_dma_chan *ioat_chan);
68 static struct ioat_desc_sw *
69 ioat2_dma_get_next_descriptor(struct ioat_dma_chan *ioat_chan);
70
71 static inline struct ioat_dma_chan *ioat_lookup_chan_by_index(
72 struct ioatdma_device *device,
73 int index)
74 {
75 return device->idx[index];
76 }
77
78 /**
79 * ioat_dma_do_interrupt - handler used for single vector interrupt mode
80 * @irq: interrupt id
81 * @data: interrupt data
82 */
83 static irqreturn_t ioat_dma_do_interrupt(int irq, void *data)
84 {
85 struct ioatdma_device *instance = data;
86 struct ioat_dma_chan *ioat_chan;
87 unsigned long attnstatus;
88 int bit;
89 u8 intrctrl;
90
91 intrctrl = readb(instance->reg_base + IOAT_INTRCTRL_OFFSET);
92
93 if (!(intrctrl & IOAT_INTRCTRL_MASTER_INT_EN))
94 return IRQ_NONE;
95
96 if (!(intrctrl & IOAT_INTRCTRL_INT_STATUS)) {
97 writeb(intrctrl, instance->reg_base + IOAT_INTRCTRL_OFFSET);
98 return IRQ_NONE;
99 }
100
101 attnstatus = readl(instance->reg_base + IOAT_ATTNSTATUS_OFFSET);
102 for_each_bit(bit, &attnstatus, BITS_PER_LONG) {
103 ioat_chan = ioat_lookup_chan_by_index(instance, bit);
104 tasklet_schedule(&ioat_chan->cleanup_task);
105 }
106
107 writeb(intrctrl, instance->reg_base + IOAT_INTRCTRL_OFFSET);
108 return IRQ_HANDLED;
109 }
110
111 /**
112 * ioat_dma_do_interrupt_msix - handler used for vector-per-channel interrupt mode
113 * @irq: interrupt id
114 * @data: interrupt data
115 */
116 static irqreturn_t ioat_dma_do_interrupt_msix(int irq, void *data)
117 {
118 struct ioat_dma_chan *ioat_chan = data;
119
120 tasklet_schedule(&ioat_chan->cleanup_task);
121
122 return IRQ_HANDLED;
123 }
124
125 static void ioat_dma_cleanup_tasklet(unsigned long data);
126
127 /**
128 * ioat_dma_enumerate_channels - find and initialize the device's channels
129 * @device: the device to be enumerated
130 */
131 static int ioat_dma_enumerate_channels(struct ioatdma_device *device)
132 {
133 u8 xfercap_scale;
134 u32 xfercap;
135 int i;
136 struct ioat_dma_chan *ioat_chan;
137
138 /*
139 * IOAT ver.3 workarounds
140 */
141 if (device->version == IOAT_VER_3_0) {
142 u32 chan_err_mask;
143 u16 dev_id;
144 u32 dmauncerrsts;
145
146 /*
147 * Write CHANERRMSK_INT with 3E07h to mask out the errors
148 * that can cause stability issues for IOAT ver.3
149 */
150 chan_err_mask = 0x3E07;
151 pci_write_config_dword(device->pdev,
152 IOAT_PCI_CHANERRMASK_INT_OFFSET,
153 chan_err_mask);
154
155 /*
156 * Clear DMAUNCERRSTS Cfg-Reg Parity Error status bit
157 * (workaround for spurious config parity error after restart)
158 */
159 pci_read_config_word(device->pdev,
160 IOAT_PCI_DEVICE_ID_OFFSET,
161 &dev_id);
162 if (dev_id == PCI_DEVICE_ID_INTEL_IOAT_TBG0) {
163 dmauncerrsts = 0x10;
164 pci_write_config_dword(device->pdev,
165 IOAT_PCI_DMAUNCERRSTS_OFFSET,
166 dmauncerrsts);
167 }
168 }
169
170 device->common.chancnt = readb(device->reg_base + IOAT_CHANCNT_OFFSET);
171 xfercap_scale = readb(device->reg_base + IOAT_XFERCAP_OFFSET);
172 xfercap = (xfercap_scale == 0 ? -1 : (1UL << xfercap_scale));
173
174 for (i = 0; i < device->common.chancnt; i++) {
175 ioat_chan = kzalloc(sizeof(*ioat_chan), GFP_KERNEL);
176 if (!ioat_chan) {
177 device->common.chancnt = i;
178 break;
179 }
180
181 ioat_chan->device = device;
182 ioat_chan->reg_base = device->reg_base + (0x80 * (i + 1));
183 ioat_chan->xfercap = xfercap;
184 ioat_chan->desccount = 0;
185 INIT_DELAYED_WORK(&ioat_chan->work, ioat_dma_chan_reset_part2);
186 if (ioat_chan->device->version != IOAT_VER_1_2) {
187 writel(IOAT_DCACTRL_CMPL_WRITE_ENABLE
188 | IOAT_DMA_DCA_ANY_CPU,
189 ioat_chan->reg_base + IOAT_DCACTRL_OFFSET);
190 }
191 spin_lock_init(&ioat_chan->cleanup_lock);
192 spin_lock_init(&ioat_chan->desc_lock);
193 INIT_LIST_HEAD(&ioat_chan->free_desc);
194 INIT_LIST_HEAD(&ioat_chan->used_desc);
195 /* This should be made common somewhere in dmaengine.c */
196 ioat_chan->common.device = &device->common;
197 list_add_tail(&ioat_chan->common.device_node,
198 &device->common.channels);
199 device->idx[i] = ioat_chan;
200 tasklet_init(&ioat_chan->cleanup_task,
201 ioat_dma_cleanup_tasklet,
202 (unsigned long) ioat_chan);
203 tasklet_disable(&ioat_chan->cleanup_task);
204 }
205 return device->common.chancnt;
206 }
207
208 /**
209 * ioat_dma_memcpy_issue_pending - push potentially unrecognized appended
210 * descriptors to hw
211 * @chan: DMA channel handle
212 */
213 static inline void __ioat1_dma_memcpy_issue_pending(
214 struct ioat_dma_chan *ioat_chan)
215 {
216 ioat_chan->pending = 0;
217 writeb(IOAT_CHANCMD_APPEND, ioat_chan->reg_base + IOAT1_CHANCMD_OFFSET);
218 }
219
220 static void ioat1_dma_memcpy_issue_pending(struct dma_chan *chan)
221 {
222 struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
223
224 if (ioat_chan->pending > 0) {
225 spin_lock_bh(&ioat_chan->desc_lock);
226 __ioat1_dma_memcpy_issue_pending(ioat_chan);
227 spin_unlock_bh(&ioat_chan->desc_lock);
228 }
229 }
230
231 static inline void __ioat2_dma_memcpy_issue_pending(
232 struct ioat_dma_chan *ioat_chan)
233 {
234 ioat_chan->pending = 0;
235 writew(ioat_chan->dmacount,
236 ioat_chan->reg_base + IOAT_CHAN_DMACOUNT_OFFSET);
237 }
238
239 static void ioat2_dma_memcpy_issue_pending(struct dma_chan *chan)
240 {
241 struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
242
243 if (ioat_chan->pending > 0) {
244 spin_lock_bh(&ioat_chan->desc_lock);
245 __ioat2_dma_memcpy_issue_pending(ioat_chan);
246 spin_unlock_bh(&ioat_chan->desc_lock);
247 }
248 }
249
250
251 /**
252 * ioat_dma_chan_reset_part2 - reinit the channel after a reset
253 */
254 static void ioat_dma_chan_reset_part2(struct work_struct *work)
255 {
256 struct ioat_dma_chan *ioat_chan =
257 container_of(work, struct ioat_dma_chan, work.work);
258 struct ioat_desc_sw *desc;
259
260 spin_lock_bh(&ioat_chan->cleanup_lock);
261 spin_lock_bh(&ioat_chan->desc_lock);
262
263 ioat_chan->completion_virt->low = 0;
264 ioat_chan->completion_virt->high = 0;
265 ioat_chan->pending = 0;
266
267 /*
268 * count the descriptors waiting, and be sure to do it
269 * right for both the CB1 line and the CB2 ring
270 */
271 ioat_chan->dmacount = 0;
272 if (ioat_chan->used_desc.prev) {
273 desc = to_ioat_desc(ioat_chan->used_desc.prev);
274 do {
275 ioat_chan->dmacount++;
276 desc = to_ioat_desc(desc->node.next);
277 } while (&desc->node != ioat_chan->used_desc.next);
278 }
279
280 /*
281 * write the new starting descriptor address
282 * this puts channel engine into ARMED state
283 */
284 desc = to_ioat_desc(ioat_chan->used_desc.prev);
285 switch (ioat_chan->device->version) {
286 case IOAT_VER_1_2:
287 writel(((u64) desc->async_tx.phys) & 0x00000000FFFFFFFF,
288 ioat_chan->reg_base + IOAT1_CHAINADDR_OFFSET_LOW);
289 writel(((u64) desc->async_tx.phys) >> 32,
290 ioat_chan->reg_base + IOAT1_CHAINADDR_OFFSET_HIGH);
291
292 writeb(IOAT_CHANCMD_START, ioat_chan->reg_base
293 + IOAT_CHANCMD_OFFSET(ioat_chan->device->version));
294 break;
295 case IOAT_VER_2_0:
296 writel(((u64) desc->async_tx.phys) & 0x00000000FFFFFFFF,
297 ioat_chan->reg_base + IOAT2_CHAINADDR_OFFSET_LOW);
298 writel(((u64) desc->async_tx.phys) >> 32,
299 ioat_chan->reg_base + IOAT2_CHAINADDR_OFFSET_HIGH);
300
301 /* tell the engine to go with what's left to be done */
302 writew(ioat_chan->dmacount,
303 ioat_chan->reg_base + IOAT_CHAN_DMACOUNT_OFFSET);
304
305 break;
306 }
307 dev_err(&ioat_chan->device->pdev->dev,
308 "chan%d reset - %d descs waiting, %d total desc\n",
309 chan_num(ioat_chan), ioat_chan->dmacount, ioat_chan->desccount);
310
311 spin_unlock_bh(&ioat_chan->desc_lock);
312 spin_unlock_bh(&ioat_chan->cleanup_lock);
313 }
314
315 /**
316 * ioat_dma_reset_channel - restart a channel
317 * @ioat_chan: IOAT DMA channel handle
318 */
319 static void ioat_dma_reset_channel(struct ioat_dma_chan *ioat_chan)
320 {
321 u32 chansts, chanerr;
322
323 if (!ioat_chan->used_desc.prev)
324 return;
325
326 chanerr = readl(ioat_chan->reg_base + IOAT_CHANERR_OFFSET);
327 chansts = (ioat_chan->completion_virt->low
328 & IOAT_CHANSTS_DMA_TRANSFER_STATUS);
329 if (chanerr) {
330 dev_err(&ioat_chan->device->pdev->dev,
331 "chan%d, CHANSTS = 0x%08x CHANERR = 0x%04x, clearing\n",
332 chan_num(ioat_chan), chansts, chanerr);
333 writel(chanerr, ioat_chan->reg_base + IOAT_CHANERR_OFFSET);
334 }
335
336 /*
337 * whack it upside the head with a reset
338 * and wait for things to settle out.
339 * force the pending count to a really big negative
340 * to make sure no one forces an issue_pending
341 * while we're waiting.
342 */
343
344 spin_lock_bh(&ioat_chan->desc_lock);
345 ioat_chan->pending = INT_MIN;
346 writeb(IOAT_CHANCMD_RESET,
347 ioat_chan->reg_base
348 + IOAT_CHANCMD_OFFSET(ioat_chan->device->version));
349 spin_unlock_bh(&ioat_chan->desc_lock);
350
351 /* schedule the 2nd half instead of sleeping a long time */
352 schedule_delayed_work(&ioat_chan->work, RESET_DELAY);
353 }
354
355 /**
356 * ioat_dma_chan_watchdog - watch for stuck channels
357 */
358 static void ioat_dma_chan_watchdog(struct work_struct *work)
359 {
360 struct ioatdma_device *device =
361 container_of(work, struct ioatdma_device, work.work);
362 struct ioat_dma_chan *ioat_chan;
363 int i;
364
365 union {
366 u64 full;
367 struct {
368 u32 low;
369 u32 high;
370 };
371 } completion_hw;
372 unsigned long compl_desc_addr_hw;
373
374 for (i = 0; i < device->common.chancnt; i++) {
375 ioat_chan = ioat_lookup_chan_by_index(device, i);
376
377 if (ioat_chan->device->version == IOAT_VER_1_2
378 /* have we started processing anything yet */
379 && ioat_chan->last_completion
380 /* have we completed any since last watchdog cycle? */
381 && (ioat_chan->last_completion ==
382 ioat_chan->watchdog_completion)
383 /* has TCP stuck on one cookie since last watchdog? */
384 && (ioat_chan->watchdog_tcp_cookie ==
385 ioat_chan->watchdog_last_tcp_cookie)
386 && (ioat_chan->watchdog_tcp_cookie !=
387 ioat_chan->completed_cookie)
388 /* is there something in the chain to be processed? */
389 /* CB1 chain always has at least the last one processed */
390 && (ioat_chan->used_desc.prev != ioat_chan->used_desc.next)
391 && ioat_chan->pending == 0) {
392
393 /*
394 * check CHANSTS register for completed
395 * descriptor address.
396 * if it is different than completion writeback,
397 * it is not zero
398 * and it has changed since the last watchdog
399 * we can assume that channel
400 * is still working correctly
401 * and the problem is in completion writeback.
402 * update completion writeback
403 * with actual CHANSTS value
404 * else
405 * try resetting the channel
406 */
407
408 completion_hw.low = readl(ioat_chan->reg_base +
409 IOAT_CHANSTS_OFFSET_LOW(ioat_chan->device->version));
410 completion_hw.high = readl(ioat_chan->reg_base +
411 IOAT_CHANSTS_OFFSET_HIGH(ioat_chan->device->version));
412 #if (BITS_PER_LONG == 64)
413 compl_desc_addr_hw =
414 completion_hw.full
415 & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_ADDR;
416 #else
417 compl_desc_addr_hw =
418 completion_hw.low & IOAT_LOW_COMPLETION_MASK;
419 #endif
420
421 if ((compl_desc_addr_hw != 0)
422 && (compl_desc_addr_hw != ioat_chan->watchdog_completion)
423 && (compl_desc_addr_hw != ioat_chan->last_compl_desc_addr_hw)) {
424 ioat_chan->last_compl_desc_addr_hw = compl_desc_addr_hw;
425 ioat_chan->completion_virt->low = completion_hw.low;
426 ioat_chan->completion_virt->high = completion_hw.high;
427 } else {
428 ioat_dma_reset_channel(ioat_chan);
429 ioat_chan->watchdog_completion = 0;
430 ioat_chan->last_compl_desc_addr_hw = 0;
431 }
432
433 /*
434 * for version 2.0 if there are descriptors yet to be processed
435 * and the last completed hasn't changed since the last watchdog
436 * if they haven't hit the pending level
437 * issue the pending to push them through
438 * else
439 * try resetting the channel
440 */
441 } else if (ioat_chan->device->version == IOAT_VER_2_0
442 && ioat_chan->used_desc.prev
443 && ioat_chan->last_completion
444 && ioat_chan->last_completion == ioat_chan->watchdog_completion) {
445
446 if (ioat_chan->pending < ioat_pending_level)
447 ioat2_dma_memcpy_issue_pending(&ioat_chan->common);
448 else {
449 ioat_dma_reset_channel(ioat_chan);
450 ioat_chan->watchdog_completion = 0;
451 }
452 } else {
453 ioat_chan->last_compl_desc_addr_hw = 0;
454 ioat_chan->watchdog_completion
455 = ioat_chan->last_completion;
456 }
457
458 ioat_chan->watchdog_last_tcp_cookie =
459 ioat_chan->watchdog_tcp_cookie;
460 }
461
462 schedule_delayed_work(&device->work, WATCHDOG_DELAY);
463 }
464
465 static dma_cookie_t ioat1_tx_submit(struct dma_async_tx_descriptor *tx)
466 {
467 struct ioat_dma_chan *ioat_chan = to_ioat_chan(tx->chan);
468 struct ioat_desc_sw *first = tx_to_ioat_desc(tx);
469 struct ioat_desc_sw *prev, *new;
470 struct ioat_dma_descriptor *hw;
471 dma_cookie_t cookie;
472 LIST_HEAD(new_chain);
473 u32 copy;
474 size_t len;
475 dma_addr_t src, dst;
476 unsigned long orig_flags;
477 unsigned int desc_count = 0;
478
479 /* src and dest and len are stored in the initial descriptor */
480 len = first->len;
481 src = first->src;
482 dst = first->dst;
483 orig_flags = first->async_tx.flags;
484 new = first;
485
486 spin_lock_bh(&ioat_chan->desc_lock);
487 prev = to_ioat_desc(ioat_chan->used_desc.prev);
488 prefetch(prev->hw);
489 do {
490 copy = min_t(size_t, len, ioat_chan->xfercap);
491
492 async_tx_ack(&new->async_tx);
493
494 hw = new->hw;
495 hw->size = copy;
496 hw->ctl = 0;
497 hw->src_addr = src;
498 hw->dst_addr = dst;
499 hw->next = 0;
500
501 /* chain together the physical address list for the HW */
502 wmb();
503 prev->hw->next = (u64) new->async_tx.phys;
504
505 len -= copy;
506 dst += copy;
507 src += copy;
508
509 list_add_tail(&new->node, &new_chain);
510 desc_count++;
511 prev = new;
512 } while (len && (new = ioat1_dma_get_next_descriptor(ioat_chan)));
513
514 if (!new) {
515 dev_err(&ioat_chan->device->pdev->dev,
516 "tx submit failed\n");
517 spin_unlock_bh(&ioat_chan->desc_lock);
518 return -ENOMEM;
519 }
520
521 hw->ctl = IOAT_DMA_DESCRIPTOR_CTL_CP_STS;
522 if (new->async_tx.callback) {
523 hw->ctl |= IOAT_DMA_DESCRIPTOR_CTL_INT_GN;
524 if (first != new) {
525 /* move callback into to last desc */
526 new->async_tx.callback = first->async_tx.callback;
527 new->async_tx.callback_param
528 = first->async_tx.callback_param;
529 first->async_tx.callback = NULL;
530 first->async_tx.callback_param = NULL;
531 }
532 }
533
534 new->tx_cnt = desc_count;
535 new->async_tx.flags = orig_flags; /* client is in control of this ack */
536
537 /* store the original values for use in later cleanup */
538 if (new != first) {
539 new->src = first->src;
540 new->dst = first->dst;
541 new->len = first->len;
542 }
543
544 /* cookie incr and addition to used_list must be atomic */
545 cookie = ioat_chan->common.cookie;
546 cookie++;
547 if (cookie < 0)
548 cookie = 1;
549 ioat_chan->common.cookie = new->async_tx.cookie = cookie;
550
551 /* write address into NextDescriptor field of last desc in chain */
552 to_ioat_desc(ioat_chan->used_desc.prev)->hw->next =
553 first->async_tx.phys;
554 __list_splice(&new_chain, ioat_chan->used_desc.prev);
555
556 ioat_chan->dmacount += desc_count;
557 ioat_chan->pending += desc_count;
558 if (ioat_chan->pending >= ioat_pending_level)
559 __ioat1_dma_memcpy_issue_pending(ioat_chan);
560 spin_unlock_bh(&ioat_chan->desc_lock);
561
562 return cookie;
563 }
564
565 static dma_cookie_t ioat2_tx_submit(struct dma_async_tx_descriptor *tx)
566 {
567 struct ioat_dma_chan *ioat_chan = to_ioat_chan(tx->chan);
568 struct ioat_desc_sw *first = tx_to_ioat_desc(tx);
569 struct ioat_desc_sw *new;
570 struct ioat_dma_descriptor *hw;
571 dma_cookie_t cookie;
572 u32 copy;
573 size_t len;
574 dma_addr_t src, dst;
575 unsigned long orig_flags;
576 unsigned int desc_count = 0;
577
578 /* src and dest and len are stored in the initial descriptor */
579 len = first->len;
580 src = first->src;
581 dst = first->dst;
582 orig_flags = first->async_tx.flags;
583 new = first;
584
585 /*
586 * ioat_chan->desc_lock is still in force in version 2 path
587 * it gets unlocked at end of this function
588 */
589 do {
590 copy = min_t(size_t, len, ioat_chan->xfercap);
591
592 async_tx_ack(&new->async_tx);
593
594 hw = new->hw;
595 hw->size = copy;
596 hw->ctl = 0;
597 hw->src_addr = src;
598 hw->dst_addr = dst;
599
600 len -= copy;
601 dst += copy;
602 src += copy;
603 desc_count++;
604 } while (len && (new = ioat2_dma_get_next_descriptor(ioat_chan)));
605
606 if (!new) {
607 dev_err(&ioat_chan->device->pdev->dev,
608 "tx submit failed\n");
609 spin_unlock_bh(&ioat_chan->desc_lock);
610 return -ENOMEM;
611 }
612
613 hw->ctl |= IOAT_DMA_DESCRIPTOR_CTL_CP_STS;
614 if (new->async_tx.callback) {
615 hw->ctl |= IOAT_DMA_DESCRIPTOR_CTL_INT_GN;
616 if (first != new) {
617 /* move callback into to last desc */
618 new->async_tx.callback = first->async_tx.callback;
619 new->async_tx.callback_param
620 = first->async_tx.callback_param;
621 first->async_tx.callback = NULL;
622 first->async_tx.callback_param = NULL;
623 }
624 }
625
626 new->tx_cnt = desc_count;
627 new->async_tx.flags = orig_flags; /* client is in control of this ack */
628
629 /* store the original values for use in later cleanup */
630 if (new != first) {
631 new->src = first->src;
632 new->dst = first->dst;
633 new->len = first->len;
634 }
635
636 /* cookie incr and addition to used_list must be atomic */
637 cookie = ioat_chan->common.cookie;
638 cookie++;
639 if (cookie < 0)
640 cookie = 1;
641 ioat_chan->common.cookie = new->async_tx.cookie = cookie;
642
643 ioat_chan->dmacount += desc_count;
644 ioat_chan->pending += desc_count;
645 if (ioat_chan->pending >= ioat_pending_level)
646 __ioat2_dma_memcpy_issue_pending(ioat_chan);
647 spin_unlock_bh(&ioat_chan->desc_lock);
648
649 return cookie;
650 }
651
652 /**
653 * ioat_dma_alloc_descriptor - allocate and return a sw and hw descriptor pair
654 * @ioat_chan: the channel supplying the memory pool for the descriptors
655 * @flags: allocation flags
656 */
657 static struct ioat_desc_sw *ioat_dma_alloc_descriptor(
658 struct ioat_dma_chan *ioat_chan,
659 gfp_t flags)
660 {
661 struct ioat_dma_descriptor *desc;
662 struct ioat_desc_sw *desc_sw;
663 struct ioatdma_device *ioatdma_device;
664 dma_addr_t phys;
665
666 ioatdma_device = to_ioatdma_device(ioat_chan->common.device);
667 desc = pci_pool_alloc(ioatdma_device->dma_pool, flags, &phys);
668 if (unlikely(!desc))
669 return NULL;
670
671 desc_sw = kzalloc(sizeof(*desc_sw), flags);
672 if (unlikely(!desc_sw)) {
673 pci_pool_free(ioatdma_device->dma_pool, desc, phys);
674 return NULL;
675 }
676
677 memset(desc, 0, sizeof(*desc));
678 dma_async_tx_descriptor_init(&desc_sw->async_tx, &ioat_chan->common);
679 switch (ioat_chan->device->version) {
680 case IOAT_VER_1_2:
681 desc_sw->async_tx.tx_submit = ioat1_tx_submit;
682 break;
683 case IOAT_VER_2_0:
684 case IOAT_VER_3_0:
685 desc_sw->async_tx.tx_submit = ioat2_tx_submit;
686 break;
687 }
688 INIT_LIST_HEAD(&desc_sw->async_tx.tx_list);
689
690 desc_sw->hw = desc;
691 desc_sw->async_tx.phys = phys;
692
693 return desc_sw;
694 }
695
696 static int ioat_initial_desc_count = 256;
697 module_param(ioat_initial_desc_count, int, 0644);
698 MODULE_PARM_DESC(ioat_initial_desc_count,
699 "initial descriptors per channel (default: 256)");
700
701 /**
702 * ioat2_dma_massage_chan_desc - link the descriptors into a circle
703 * @ioat_chan: the channel to be massaged
704 */
705 static void ioat2_dma_massage_chan_desc(struct ioat_dma_chan *ioat_chan)
706 {
707 struct ioat_desc_sw *desc, *_desc;
708
709 /* setup used_desc */
710 ioat_chan->used_desc.next = ioat_chan->free_desc.next;
711 ioat_chan->used_desc.prev = NULL;
712
713 /* pull free_desc out of the circle so that every node is a hw
714 * descriptor, but leave it pointing to the list
715 */
716 ioat_chan->free_desc.prev->next = ioat_chan->free_desc.next;
717 ioat_chan->free_desc.next->prev = ioat_chan->free_desc.prev;
718
719 /* circle link the hw descriptors */
720 desc = to_ioat_desc(ioat_chan->free_desc.next);
721 desc->hw->next = to_ioat_desc(desc->node.next)->async_tx.phys;
722 list_for_each_entry_safe(desc, _desc, ioat_chan->free_desc.next, node) {
723 desc->hw->next = to_ioat_desc(desc->node.next)->async_tx.phys;
724 }
725 }
726
727 /**
728 * ioat_dma_alloc_chan_resources - returns the number of allocated descriptors
729 * @chan: the channel to be filled out
730 */
731 static int ioat_dma_alloc_chan_resources(struct dma_chan *chan,
732 struct dma_client *client)
733 {
734 struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
735 struct ioat_desc_sw *desc;
736 u16 chanctrl;
737 u32 chanerr;
738 int i;
739 LIST_HEAD(tmp_list);
740
741 /* have we already been set up? */
742 if (!list_empty(&ioat_chan->free_desc))
743 return ioat_chan->desccount;
744
745 /* Setup register to interrupt and write completion status on error */
746 chanctrl = IOAT_CHANCTRL_ERR_INT_EN |
747 IOAT_CHANCTRL_ANY_ERR_ABORT_EN |
748 IOAT_CHANCTRL_ERR_COMPLETION_EN;
749 writew(chanctrl, ioat_chan->reg_base + IOAT_CHANCTRL_OFFSET);
750
751 chanerr = readl(ioat_chan->reg_base + IOAT_CHANERR_OFFSET);
752 if (chanerr) {
753 dev_err(&ioat_chan->device->pdev->dev,
754 "CHANERR = %x, clearing\n", chanerr);
755 writel(chanerr, ioat_chan->reg_base + IOAT_CHANERR_OFFSET);
756 }
757
758 /* Allocate descriptors */
759 for (i = 0; i < ioat_initial_desc_count; i++) {
760 desc = ioat_dma_alloc_descriptor(ioat_chan, GFP_KERNEL);
761 if (!desc) {
762 dev_err(&ioat_chan->device->pdev->dev,
763 "Only %d initial descriptors\n", i);
764 break;
765 }
766 list_add_tail(&desc->node, &tmp_list);
767 }
768 spin_lock_bh(&ioat_chan->desc_lock);
769 ioat_chan->desccount = i;
770 list_splice(&tmp_list, &ioat_chan->free_desc);
771 if (ioat_chan->device->version != IOAT_VER_1_2)
772 ioat2_dma_massage_chan_desc(ioat_chan);
773 spin_unlock_bh(&ioat_chan->desc_lock);
774
775 /* allocate a completion writeback area */
776 /* doing 2 32bit writes to mmio since 1 64b write doesn't work */
777 ioat_chan->completion_virt =
778 pci_pool_alloc(ioat_chan->device->completion_pool,
779 GFP_KERNEL,
780 &ioat_chan->completion_addr);
781 memset(ioat_chan->completion_virt, 0,
782 sizeof(*ioat_chan->completion_virt));
783 writel(((u64) ioat_chan->completion_addr) & 0x00000000FFFFFFFF,
784 ioat_chan->reg_base + IOAT_CHANCMP_OFFSET_LOW);
785 writel(((u64) ioat_chan->completion_addr) >> 32,
786 ioat_chan->reg_base + IOAT_CHANCMP_OFFSET_HIGH);
787
788 tasklet_enable(&ioat_chan->cleanup_task);
789 ioat_dma_start_null_desc(ioat_chan); /* give chain to dma device */
790 return ioat_chan->desccount;
791 }
792
793 /**
794 * ioat_dma_free_chan_resources - release all the descriptors
795 * @chan: the channel to be cleaned
796 */
797 static void ioat_dma_free_chan_resources(struct dma_chan *chan)
798 {
799 struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
800 struct ioatdma_device *ioatdma_device = to_ioatdma_device(chan->device);
801 struct ioat_desc_sw *desc, *_desc;
802 int in_use_descs = 0;
803
804 tasklet_disable(&ioat_chan->cleanup_task);
805 ioat_dma_memcpy_cleanup(ioat_chan);
806
807 /* Delay 100ms after reset to allow internal DMA logic to quiesce
808 * before removing DMA descriptor resources.
809 */
810 writeb(IOAT_CHANCMD_RESET,
811 ioat_chan->reg_base
812 + IOAT_CHANCMD_OFFSET(ioat_chan->device->version));
813 mdelay(100);
814
815 spin_lock_bh(&ioat_chan->desc_lock);
816 switch (ioat_chan->device->version) {
817 case IOAT_VER_1_2:
818 list_for_each_entry_safe(desc, _desc,
819 &ioat_chan->used_desc, node) {
820 in_use_descs++;
821 list_del(&desc->node);
822 pci_pool_free(ioatdma_device->dma_pool, desc->hw,
823 desc->async_tx.phys);
824 kfree(desc);
825 }
826 list_for_each_entry_safe(desc, _desc,
827 &ioat_chan->free_desc, node) {
828 list_del(&desc->node);
829 pci_pool_free(ioatdma_device->dma_pool, desc->hw,
830 desc->async_tx.phys);
831 kfree(desc);
832 }
833 break;
834 case IOAT_VER_2_0:
835 case IOAT_VER_3_0:
836 list_for_each_entry_safe(desc, _desc,
837 ioat_chan->free_desc.next, node) {
838 list_del(&desc->node);
839 pci_pool_free(ioatdma_device->dma_pool, desc->hw,
840 desc->async_tx.phys);
841 kfree(desc);
842 }
843 desc = to_ioat_desc(ioat_chan->free_desc.next);
844 pci_pool_free(ioatdma_device->dma_pool, desc->hw,
845 desc->async_tx.phys);
846 kfree(desc);
847 INIT_LIST_HEAD(&ioat_chan->free_desc);
848 INIT_LIST_HEAD(&ioat_chan->used_desc);
849 break;
850 }
851 spin_unlock_bh(&ioat_chan->desc_lock);
852
853 pci_pool_free(ioatdma_device->completion_pool,
854 ioat_chan->completion_virt,
855 ioat_chan->completion_addr);
856
857 /* one is ok since we left it on there on purpose */
858 if (in_use_descs > 1)
859 dev_err(&ioat_chan->device->pdev->dev,
860 "Freeing %d in use descriptors!\n",
861 in_use_descs - 1);
862
863 ioat_chan->last_completion = ioat_chan->completion_addr = 0;
864 ioat_chan->pending = 0;
865 ioat_chan->dmacount = 0;
866 ioat_chan->watchdog_completion = 0;
867 ioat_chan->last_compl_desc_addr_hw = 0;
868 ioat_chan->watchdog_tcp_cookie =
869 ioat_chan->watchdog_last_tcp_cookie = 0;
870 }
871
872 /**
873 * ioat_dma_get_next_descriptor - return the next available descriptor
874 * @ioat_chan: IOAT DMA channel handle
875 *
876 * Gets the next descriptor from the chain, and must be called with the
877 * channel's desc_lock held. Allocates more descriptors if the channel
878 * has run out.
879 */
880 static struct ioat_desc_sw *
881 ioat1_dma_get_next_descriptor(struct ioat_dma_chan *ioat_chan)
882 {
883 struct ioat_desc_sw *new;
884
885 if (!list_empty(&ioat_chan->free_desc)) {
886 new = to_ioat_desc(ioat_chan->free_desc.next);
887 list_del(&new->node);
888 } else {
889 /* try to get another desc */
890 new = ioat_dma_alloc_descriptor(ioat_chan, GFP_ATOMIC);
891 if (!new) {
892 dev_err(&ioat_chan->device->pdev->dev,
893 "alloc failed\n");
894 return NULL;
895 }
896 }
897
898 prefetch(new->hw);
899 return new;
900 }
901
902 static struct ioat_desc_sw *
903 ioat2_dma_get_next_descriptor(struct ioat_dma_chan *ioat_chan)
904 {
905 struct ioat_desc_sw *new;
906
907 /*
908 * used.prev points to where to start processing
909 * used.next points to next free descriptor
910 * if used.prev == NULL, there are none waiting to be processed
911 * if used.next == used.prev.prev, there is only one free descriptor,
912 * and we need to use it to as a noop descriptor before
913 * linking in a new set of descriptors, since the device
914 * has probably already read the pointer to it
915 */
916 if (ioat_chan->used_desc.prev &&
917 ioat_chan->used_desc.next == ioat_chan->used_desc.prev->prev) {
918
919 struct ioat_desc_sw *desc;
920 struct ioat_desc_sw *noop_desc;
921 int i;
922
923 /* set up the noop descriptor */
924 noop_desc = to_ioat_desc(ioat_chan->used_desc.next);
925 /* set size to non-zero value (channel returns error when size is 0) */
926 noop_desc->hw->size = NULL_DESC_BUFFER_SIZE;
927 noop_desc->hw->ctl = IOAT_DMA_DESCRIPTOR_NUL;
928 noop_desc->hw->src_addr = 0;
929 noop_desc->hw->dst_addr = 0;
930
931 ioat_chan->used_desc.next = ioat_chan->used_desc.next->next;
932 ioat_chan->pending++;
933 ioat_chan->dmacount++;
934
935 /* try to get a few more descriptors */
936 for (i = 16; i; i--) {
937 desc = ioat_dma_alloc_descriptor(ioat_chan, GFP_ATOMIC);
938 if (!desc) {
939 dev_err(&ioat_chan->device->pdev->dev,
940 "alloc failed\n");
941 break;
942 }
943 list_add_tail(&desc->node, ioat_chan->used_desc.next);
944
945 desc->hw->next
946 = to_ioat_desc(desc->node.next)->async_tx.phys;
947 to_ioat_desc(desc->node.prev)->hw->next
948 = desc->async_tx.phys;
949 ioat_chan->desccount++;
950 }
951
952 ioat_chan->used_desc.next = noop_desc->node.next;
953 }
954 new = to_ioat_desc(ioat_chan->used_desc.next);
955 prefetch(new);
956 ioat_chan->used_desc.next = new->node.next;
957
958 if (ioat_chan->used_desc.prev == NULL)
959 ioat_chan->used_desc.prev = &new->node;
960
961 prefetch(new->hw);
962 return new;
963 }
964
965 static struct ioat_desc_sw *ioat_dma_get_next_descriptor(
966 struct ioat_dma_chan *ioat_chan)
967 {
968 if (!ioat_chan)
969 return NULL;
970
971 switch (ioat_chan->device->version) {
972 case IOAT_VER_1_2:
973 return ioat1_dma_get_next_descriptor(ioat_chan);
974 break;
975 case IOAT_VER_2_0:
976 case IOAT_VER_3_0:
977 return ioat2_dma_get_next_descriptor(ioat_chan);
978 break;
979 }
980 return NULL;
981 }
982
983 static struct dma_async_tx_descriptor *ioat1_dma_prep_memcpy(
984 struct dma_chan *chan,
985 dma_addr_t dma_dest,
986 dma_addr_t dma_src,
987 size_t len,
988 unsigned long flags)
989 {
990 struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
991 struct ioat_desc_sw *new;
992
993 spin_lock_bh(&ioat_chan->desc_lock);
994 new = ioat_dma_get_next_descriptor(ioat_chan);
995 spin_unlock_bh(&ioat_chan->desc_lock);
996
997 if (new) {
998 new->len = len;
999 new->dst = dma_dest;
1000 new->src = dma_src;
1001 new->async_tx.flags = flags;
1002 return &new->async_tx;
1003 } else {
1004 dev_err(&ioat_chan->device->pdev->dev,
1005 "chan%d - get_next_desc failed: %d descs waiting, %d total desc\n",
1006 chan_num(ioat_chan), ioat_chan->dmacount, ioat_chan->desccount);
1007 return NULL;
1008 }
1009 }
1010
1011 static struct dma_async_tx_descriptor *ioat2_dma_prep_memcpy(
1012 struct dma_chan *chan,
1013 dma_addr_t dma_dest,
1014 dma_addr_t dma_src,
1015 size_t len,
1016 unsigned long flags)
1017 {
1018 struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
1019 struct ioat_desc_sw *new;
1020
1021 spin_lock_bh(&ioat_chan->desc_lock);
1022 new = ioat2_dma_get_next_descriptor(ioat_chan);
1023
1024 /*
1025 * leave ioat_chan->desc_lock set in ioat 2 path
1026 * it will get unlocked at end of tx_submit
1027 */
1028
1029 if (new) {
1030 new->len = len;
1031 new->dst = dma_dest;
1032 new->src = dma_src;
1033 new->async_tx.flags = flags;
1034 return &new->async_tx;
1035 } else {
1036 spin_unlock_bh(&ioat_chan->desc_lock);
1037 dev_err(&ioat_chan->device->pdev->dev,
1038 "chan%d - get_next_desc failed: %d descs waiting, %d total desc\n",
1039 chan_num(ioat_chan), ioat_chan->dmacount, ioat_chan->desccount);
1040 return NULL;
1041 }
1042 }
1043
1044 static void ioat_dma_cleanup_tasklet(unsigned long data)
1045 {
1046 struct ioat_dma_chan *chan = (void *)data;
1047 ioat_dma_memcpy_cleanup(chan);
1048 writew(IOAT_CHANCTRL_INT_DISABLE,
1049 chan->reg_base + IOAT_CHANCTRL_OFFSET);
1050 }
1051
1052 static void
1053 ioat_dma_unmap(struct ioat_dma_chan *ioat_chan, struct ioat_desc_sw *desc)
1054 {
1055 /*
1056 * yes we are unmapping both _page and _single
1057 * alloc'd regions with unmap_page. Is this
1058 * *really* that bad?
1059 */
1060 if (!(desc->async_tx.flags & DMA_COMPL_SKIP_DEST_UNMAP))
1061 pci_unmap_page(ioat_chan->device->pdev,
1062 pci_unmap_addr(desc, dst),
1063 pci_unmap_len(desc, len),
1064 PCI_DMA_FROMDEVICE);
1065
1066 if (!(desc->async_tx.flags & DMA_COMPL_SKIP_SRC_UNMAP))
1067 pci_unmap_page(ioat_chan->device->pdev,
1068 pci_unmap_addr(desc, src),
1069 pci_unmap_len(desc, len),
1070 PCI_DMA_TODEVICE);
1071 }
1072
1073 /**
1074 * ioat_dma_memcpy_cleanup - cleanup up finished descriptors
1075 * @chan: ioat channel to be cleaned up
1076 */
1077 static void ioat_dma_memcpy_cleanup(struct ioat_dma_chan *ioat_chan)
1078 {
1079 unsigned long phys_complete;
1080 struct ioat_desc_sw *desc, *_desc;
1081 dma_cookie_t cookie = 0;
1082 unsigned long desc_phys;
1083 struct ioat_desc_sw *latest_desc;
1084
1085 prefetch(ioat_chan->completion_virt);
1086
1087 if (!spin_trylock_bh(&ioat_chan->cleanup_lock))
1088 return;
1089
1090 /* The completion writeback can happen at any time,
1091 so reads by the driver need to be atomic operations
1092 The descriptor physical addresses are limited to 32-bits
1093 when the CPU can only do a 32-bit mov */
1094
1095 #if (BITS_PER_LONG == 64)
1096 phys_complete =
1097 ioat_chan->completion_virt->full
1098 & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_ADDR;
1099 #else
1100 phys_complete =
1101 ioat_chan->completion_virt->low & IOAT_LOW_COMPLETION_MASK;
1102 #endif
1103
1104 if ((ioat_chan->completion_virt->full
1105 & IOAT_CHANSTS_DMA_TRANSFER_STATUS) ==
1106 IOAT_CHANSTS_DMA_TRANSFER_STATUS_HALTED) {
1107 dev_err(&ioat_chan->device->pdev->dev,
1108 "Channel halted, chanerr = %x\n",
1109 readl(ioat_chan->reg_base + IOAT_CHANERR_OFFSET));
1110
1111 /* TODO do something to salvage the situation */
1112 }
1113
1114 if (phys_complete == ioat_chan->last_completion) {
1115 spin_unlock_bh(&ioat_chan->cleanup_lock);
1116 /*
1117 * perhaps we're stuck so hard that the watchdog can't go off?
1118 * try to catch it after 2 seconds
1119 */
1120 if (ioat_chan->device->version != IOAT_VER_3_0) {
1121 if (time_after(jiffies,
1122 ioat_chan->last_completion_time + HZ*WATCHDOG_DELAY)) {
1123 ioat_dma_chan_watchdog(&(ioat_chan->device->work.work));
1124 ioat_chan->last_completion_time = jiffies;
1125 }
1126 }
1127 return;
1128 }
1129 ioat_chan->last_completion_time = jiffies;
1130
1131 cookie = 0;
1132 if (!spin_trylock_bh(&ioat_chan->desc_lock)) {
1133 spin_unlock_bh(&ioat_chan->cleanup_lock);
1134 return;
1135 }
1136
1137 switch (ioat_chan->device->version) {
1138 case IOAT_VER_1_2:
1139 list_for_each_entry_safe(desc, _desc,
1140 &ioat_chan->used_desc, node) {
1141
1142 /*
1143 * Incoming DMA requests may use multiple descriptors,
1144 * due to exceeding xfercap, perhaps. If so, only the
1145 * last one will have a cookie, and require unmapping.
1146 */
1147 if (desc->async_tx.cookie) {
1148 cookie = desc->async_tx.cookie;
1149 ioat_dma_unmap(ioat_chan, desc);
1150 if (desc->async_tx.callback) {
1151 desc->async_tx.callback(desc->async_tx.callback_param);
1152 desc->async_tx.callback = NULL;
1153 }
1154 }
1155
1156 if (desc->async_tx.phys != phys_complete) {
1157 /*
1158 * a completed entry, but not the last, so clean
1159 * up if the client is done with the descriptor
1160 */
1161 if (async_tx_test_ack(&desc->async_tx)) {
1162 list_del(&desc->node);
1163 list_add_tail(&desc->node,
1164 &ioat_chan->free_desc);
1165 } else
1166 desc->async_tx.cookie = 0;
1167 } else {
1168 /*
1169 * last used desc. Do not remove, so we can
1170 * append from it, but don't look at it next
1171 * time, either
1172 */
1173 desc->async_tx.cookie = 0;
1174
1175 /* TODO check status bits? */
1176 break;
1177 }
1178 }
1179 break;
1180 case IOAT_VER_2_0:
1181 case IOAT_VER_3_0:
1182 /* has some other thread has already cleaned up? */
1183 if (ioat_chan->used_desc.prev == NULL)
1184 break;
1185
1186 /* work backwards to find latest finished desc */
1187 desc = to_ioat_desc(ioat_chan->used_desc.next);
1188 latest_desc = NULL;
1189 do {
1190 desc = to_ioat_desc(desc->node.prev);
1191 desc_phys = (unsigned long)desc->async_tx.phys
1192 & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_ADDR;
1193 if (desc_phys == phys_complete) {
1194 latest_desc = desc;
1195 break;
1196 }
1197 } while (&desc->node != ioat_chan->used_desc.prev);
1198
1199 if (latest_desc != NULL) {
1200
1201 /* work forwards to clear finished descriptors */
1202 for (desc = to_ioat_desc(ioat_chan->used_desc.prev);
1203 &desc->node != latest_desc->node.next &&
1204 &desc->node != ioat_chan->used_desc.next;
1205 desc = to_ioat_desc(desc->node.next)) {
1206 if (desc->async_tx.cookie) {
1207 cookie = desc->async_tx.cookie;
1208 desc->async_tx.cookie = 0;
1209 ioat_dma_unmap(ioat_chan, desc);
1210 if (desc->async_tx.callback) {
1211 desc->async_tx.callback(desc->async_tx.callback_param);
1212 desc->async_tx.callback = NULL;
1213 }
1214 }
1215 }
1216
1217 /* move used.prev up beyond those that are finished */
1218 if (&desc->node == ioat_chan->used_desc.next)
1219 ioat_chan->used_desc.prev = NULL;
1220 else
1221 ioat_chan->used_desc.prev = &desc->node;
1222 }
1223 break;
1224 }
1225
1226 spin_unlock_bh(&ioat_chan->desc_lock);
1227
1228 ioat_chan->last_completion = phys_complete;
1229 if (cookie != 0)
1230 ioat_chan->completed_cookie = cookie;
1231
1232 spin_unlock_bh(&ioat_chan->cleanup_lock);
1233 }
1234
1235 /**
1236 * ioat_dma_is_complete - poll the status of a IOAT DMA transaction
1237 * @chan: IOAT DMA channel handle
1238 * @cookie: DMA transaction identifier
1239 * @done: if not %NULL, updated with last completed transaction
1240 * @used: if not %NULL, updated with last used transaction
1241 */
1242 static enum dma_status ioat_dma_is_complete(struct dma_chan *chan,
1243 dma_cookie_t cookie,
1244 dma_cookie_t *done,
1245 dma_cookie_t *used)
1246 {
1247 struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
1248 dma_cookie_t last_used;
1249 dma_cookie_t last_complete;
1250 enum dma_status ret;
1251
1252 last_used = chan->cookie;
1253 last_complete = ioat_chan->completed_cookie;
1254 ioat_chan->watchdog_tcp_cookie = cookie;
1255
1256 if (done)
1257 *done = last_complete;
1258 if (used)
1259 *used = last_used;
1260
1261 ret = dma_async_is_complete(cookie, last_complete, last_used);
1262 if (ret == DMA_SUCCESS)
1263 return ret;
1264
1265 ioat_dma_memcpy_cleanup(ioat_chan);
1266
1267 last_used = chan->cookie;
1268 last_complete = ioat_chan->completed_cookie;
1269
1270 if (done)
1271 *done = last_complete;
1272 if (used)
1273 *used = last_used;
1274
1275 return dma_async_is_complete(cookie, last_complete, last_used);
1276 }
1277
1278 static void ioat_dma_start_null_desc(struct ioat_dma_chan *ioat_chan)
1279 {
1280 struct ioat_desc_sw *desc;
1281
1282 spin_lock_bh(&ioat_chan->desc_lock);
1283
1284 desc = ioat_dma_get_next_descriptor(ioat_chan);
1285
1286 if (!desc) {
1287 dev_err(&ioat_chan->device->pdev->dev,
1288 "Unable to start null desc - get next desc failed\n");
1289 spin_unlock_bh(&ioat_chan->desc_lock);
1290 return;
1291 }
1292
1293 desc->hw->ctl = IOAT_DMA_DESCRIPTOR_NUL
1294 | IOAT_DMA_DESCRIPTOR_CTL_INT_GN
1295 | IOAT_DMA_DESCRIPTOR_CTL_CP_STS;
1296 /* set size to non-zero value (channel returns error when size is 0) */
1297 desc->hw->size = NULL_DESC_BUFFER_SIZE;
1298 desc->hw->src_addr = 0;
1299 desc->hw->dst_addr = 0;
1300 async_tx_ack(&desc->async_tx);
1301 switch (ioat_chan->device->version) {
1302 case IOAT_VER_1_2:
1303 desc->hw->next = 0;
1304 list_add_tail(&desc->node, &ioat_chan->used_desc);
1305
1306 writel(((u64) desc->async_tx.phys) & 0x00000000FFFFFFFF,
1307 ioat_chan->reg_base + IOAT1_CHAINADDR_OFFSET_LOW);
1308 writel(((u64) desc->async_tx.phys) >> 32,
1309 ioat_chan->reg_base + IOAT1_CHAINADDR_OFFSET_HIGH);
1310
1311 writeb(IOAT_CHANCMD_START, ioat_chan->reg_base
1312 + IOAT_CHANCMD_OFFSET(ioat_chan->device->version));
1313 break;
1314 case IOAT_VER_2_0:
1315 case IOAT_VER_3_0:
1316 writel(((u64) desc->async_tx.phys) & 0x00000000FFFFFFFF,
1317 ioat_chan->reg_base + IOAT2_CHAINADDR_OFFSET_LOW);
1318 writel(((u64) desc->async_tx.phys) >> 32,
1319 ioat_chan->reg_base + IOAT2_CHAINADDR_OFFSET_HIGH);
1320
1321 ioat_chan->dmacount++;
1322 __ioat2_dma_memcpy_issue_pending(ioat_chan);
1323 break;
1324 }
1325 spin_unlock_bh(&ioat_chan->desc_lock);
1326 }
1327
1328 /*
1329 * Perform a IOAT transaction to verify the HW works.
1330 */
1331 #define IOAT_TEST_SIZE 2000
1332
1333 static void ioat_dma_test_callback(void *dma_async_param)
1334 {
1335 printk(KERN_ERR "ioatdma: ioat_dma_test_callback(%p)\n",
1336 dma_async_param);
1337 }
1338
1339 /**
1340 * ioat_dma_self_test - Perform a IOAT transaction to verify the HW works.
1341 * @device: device to be tested
1342 */
1343 static int ioat_dma_self_test(struct ioatdma_device *device)
1344 {
1345 int i;
1346 u8 *src;
1347 u8 *dest;
1348 struct dma_chan *dma_chan;
1349 struct dma_async_tx_descriptor *tx;
1350 dma_addr_t dma_dest, dma_src;
1351 dma_cookie_t cookie;
1352 int err = 0;
1353
1354 src = kzalloc(sizeof(u8) * IOAT_TEST_SIZE, GFP_KERNEL);
1355 if (!src)
1356 return -ENOMEM;
1357 dest = kzalloc(sizeof(u8) * IOAT_TEST_SIZE, GFP_KERNEL);
1358 if (!dest) {
1359 kfree(src);
1360 return -ENOMEM;
1361 }
1362
1363 /* Fill in src buffer */
1364 for (i = 0; i < IOAT_TEST_SIZE; i++)
1365 src[i] = (u8)i;
1366
1367 /* Start copy, using first DMA channel */
1368 dma_chan = container_of(device->common.channels.next,
1369 struct dma_chan,
1370 device_node);
1371 if (device->common.device_alloc_chan_resources(dma_chan, NULL) < 1) {
1372 dev_err(&device->pdev->dev,
1373 "selftest cannot allocate chan resource\n");
1374 err = -ENODEV;
1375 goto out;
1376 }
1377
1378 dma_src = dma_map_single(dma_chan->device->dev, src, IOAT_TEST_SIZE,
1379 DMA_TO_DEVICE);
1380 dma_dest = dma_map_single(dma_chan->device->dev, dest, IOAT_TEST_SIZE,
1381 DMA_FROM_DEVICE);
1382 tx = device->common.device_prep_dma_memcpy(dma_chan, dma_dest, dma_src,
1383 IOAT_TEST_SIZE, 0);
1384 if (!tx) {
1385 dev_err(&device->pdev->dev,
1386 "Self-test prep failed, disabling\n");
1387 err = -ENODEV;
1388 goto free_resources;
1389 }
1390
1391 async_tx_ack(tx);
1392 tx->callback = ioat_dma_test_callback;
1393 tx->callback_param = (void *)0x8086;
1394 cookie = tx->tx_submit(tx);
1395 if (cookie < 0) {
1396 dev_err(&device->pdev->dev,
1397 "Self-test setup failed, disabling\n");
1398 err = -ENODEV;
1399 goto free_resources;
1400 }
1401 device->common.device_issue_pending(dma_chan);
1402 msleep(1);
1403
1404 if (device->common.device_is_tx_complete(dma_chan, cookie, NULL, NULL)
1405 != DMA_SUCCESS) {
1406 dev_err(&device->pdev->dev,
1407 "Self-test copy timed out, disabling\n");
1408 err = -ENODEV;
1409 goto free_resources;
1410 }
1411 if (memcmp(src, dest, IOAT_TEST_SIZE)) {
1412 dev_err(&device->pdev->dev,
1413 "Self-test copy failed compare, disabling\n");
1414 err = -ENODEV;
1415 goto free_resources;
1416 }
1417
1418 free_resources:
1419 device->common.device_free_chan_resources(dma_chan);
1420 out:
1421 kfree(src);
1422 kfree(dest);
1423 return err;
1424 }
1425
1426 static char ioat_interrupt_style[32] = "msix";
1427 module_param_string(ioat_interrupt_style, ioat_interrupt_style,
1428 sizeof(ioat_interrupt_style), 0644);
1429 MODULE_PARM_DESC(ioat_interrupt_style,
1430 "set ioat interrupt style: msix (default), "
1431 "msix-single-vector, msi, intx)");
1432
1433 /**
1434 * ioat_dma_setup_interrupts - setup interrupt handler
1435 * @device: ioat device
1436 */
1437 static int ioat_dma_setup_interrupts(struct ioatdma_device *device)
1438 {
1439 struct ioat_dma_chan *ioat_chan;
1440 int err, i, j, msixcnt;
1441 u8 intrctrl = 0;
1442
1443 if (!strcmp(ioat_interrupt_style, "msix"))
1444 goto msix;
1445 if (!strcmp(ioat_interrupt_style, "msix-single-vector"))
1446 goto msix_single_vector;
1447 if (!strcmp(ioat_interrupt_style, "msi"))
1448 goto msi;
1449 if (!strcmp(ioat_interrupt_style, "intx"))
1450 goto intx;
1451 dev_err(&device->pdev->dev, "invalid ioat_interrupt_style %s\n",
1452 ioat_interrupt_style);
1453 goto err_no_irq;
1454
1455 msix:
1456 /* The number of MSI-X vectors should equal the number of channels */
1457 msixcnt = device->common.chancnt;
1458 for (i = 0; i < msixcnt; i++)
1459 device->msix_entries[i].entry = i;
1460
1461 err = pci_enable_msix(device->pdev, device->msix_entries, msixcnt);
1462 if (err < 0)
1463 goto msi;
1464 if (err > 0)
1465 goto msix_single_vector;
1466
1467 for (i = 0; i < msixcnt; i++) {
1468 ioat_chan = ioat_lookup_chan_by_index(device, i);
1469 err = request_irq(device->msix_entries[i].vector,
1470 ioat_dma_do_interrupt_msix,
1471 0, "ioat-msix", ioat_chan);
1472 if (err) {
1473 for (j = 0; j < i; j++) {
1474 ioat_chan =
1475 ioat_lookup_chan_by_index(device, j);
1476 free_irq(device->msix_entries[j].vector,
1477 ioat_chan);
1478 }
1479 goto msix_single_vector;
1480 }
1481 }
1482 intrctrl |= IOAT_INTRCTRL_MSIX_VECTOR_CONTROL;
1483 device->irq_mode = msix_multi_vector;
1484 goto done;
1485
1486 msix_single_vector:
1487 device->msix_entries[0].entry = 0;
1488 err = pci_enable_msix(device->pdev, device->msix_entries, 1);
1489 if (err)
1490 goto msi;
1491
1492 err = request_irq(device->msix_entries[0].vector, ioat_dma_do_interrupt,
1493 0, "ioat-msix", device);
1494 if (err) {
1495 pci_disable_msix(device->pdev);
1496 goto msi;
1497 }
1498 device->irq_mode = msix_single_vector;
1499 goto done;
1500
1501 msi:
1502 err = pci_enable_msi(device->pdev);
1503 if (err)
1504 goto intx;
1505
1506 err = request_irq(device->pdev->irq, ioat_dma_do_interrupt,
1507 0, "ioat-msi", device);
1508 if (err) {
1509 pci_disable_msi(device->pdev);
1510 goto intx;
1511 }
1512 /*
1513 * CB 1.2 devices need a bit set in configuration space to enable MSI
1514 */
1515 if (device->version == IOAT_VER_1_2) {
1516 u32 dmactrl;
1517 pci_read_config_dword(device->pdev,
1518 IOAT_PCI_DMACTRL_OFFSET, &dmactrl);
1519 dmactrl |= IOAT_PCI_DMACTRL_MSI_EN;
1520 pci_write_config_dword(device->pdev,
1521 IOAT_PCI_DMACTRL_OFFSET, dmactrl);
1522 }
1523 device->irq_mode = msi;
1524 goto done;
1525
1526 intx:
1527 err = request_irq(device->pdev->irq, ioat_dma_do_interrupt,
1528 IRQF_SHARED, "ioat-intx", device);
1529 if (err)
1530 goto err_no_irq;
1531 device->irq_mode = intx;
1532
1533 done:
1534 intrctrl |= IOAT_INTRCTRL_MASTER_INT_EN;
1535 writeb(intrctrl, device->reg_base + IOAT_INTRCTRL_OFFSET);
1536 return 0;
1537
1538 err_no_irq:
1539 /* Disable all interrupt generation */
1540 writeb(0, device->reg_base + IOAT_INTRCTRL_OFFSET);
1541 dev_err(&device->pdev->dev, "no usable interrupts\n");
1542 device->irq_mode = none;
1543 return -1;
1544 }
1545
1546 /**
1547 * ioat_dma_remove_interrupts - remove whatever interrupts were set
1548 * @device: ioat device
1549 */
1550 static void ioat_dma_remove_interrupts(struct ioatdma_device *device)
1551 {
1552 struct ioat_dma_chan *ioat_chan;
1553 int i;
1554
1555 /* Disable all interrupt generation */
1556 writeb(0, device->reg_base + IOAT_INTRCTRL_OFFSET);
1557
1558 switch (device->irq_mode) {
1559 case msix_multi_vector:
1560 for (i = 0; i < device->common.chancnt; i++) {
1561 ioat_chan = ioat_lookup_chan_by_index(device, i);
1562 free_irq(device->msix_entries[i].vector, ioat_chan);
1563 }
1564 pci_disable_msix(device->pdev);
1565 break;
1566 case msix_single_vector:
1567 free_irq(device->msix_entries[0].vector, device);
1568 pci_disable_msix(device->pdev);
1569 break;
1570 case msi:
1571 free_irq(device->pdev->irq, device);
1572 pci_disable_msi(device->pdev);
1573 break;
1574 case intx:
1575 free_irq(device->pdev->irq, device);
1576 break;
1577 case none:
1578 dev_warn(&device->pdev->dev,
1579 "call to %s without interrupts setup\n", __func__);
1580 }
1581 device->irq_mode = none;
1582 }
1583
1584 struct ioatdma_device *ioat_dma_probe(struct pci_dev *pdev,
1585 void __iomem *iobase)
1586 {
1587 int err;
1588 struct ioatdma_device *device;
1589
1590 device = kzalloc(sizeof(*device), GFP_KERNEL);
1591 if (!device) {
1592 err = -ENOMEM;
1593 goto err_kzalloc;
1594 }
1595 device->pdev = pdev;
1596 device->reg_base = iobase;
1597 device->version = readb(device->reg_base + IOAT_VER_OFFSET);
1598
1599 /* DMA coherent memory pool for DMA descriptor allocations */
1600 device->dma_pool = pci_pool_create("dma_desc_pool", pdev,
1601 sizeof(struct ioat_dma_descriptor),
1602 64, 0);
1603 if (!device->dma_pool) {
1604 err = -ENOMEM;
1605 goto err_dma_pool;
1606 }
1607
1608 device->completion_pool = pci_pool_create("completion_pool", pdev,
1609 sizeof(u64), SMP_CACHE_BYTES,
1610 SMP_CACHE_BYTES);
1611 if (!device->completion_pool) {
1612 err = -ENOMEM;
1613 goto err_completion_pool;
1614 }
1615
1616 INIT_LIST_HEAD(&device->common.channels);
1617 ioat_dma_enumerate_channels(device);
1618
1619 device->common.device_alloc_chan_resources =
1620 ioat_dma_alloc_chan_resources;
1621 device->common.device_free_chan_resources =
1622 ioat_dma_free_chan_resources;
1623 device->common.dev = &pdev->dev;
1624
1625 dma_cap_set(DMA_MEMCPY, device->common.cap_mask);
1626 device->common.device_is_tx_complete = ioat_dma_is_complete;
1627 switch (device->version) {
1628 case IOAT_VER_1_2:
1629 device->common.device_prep_dma_memcpy = ioat1_dma_prep_memcpy;
1630 device->common.device_issue_pending =
1631 ioat1_dma_memcpy_issue_pending;
1632 break;
1633 case IOAT_VER_2_0:
1634 case IOAT_VER_3_0:
1635 device->common.device_prep_dma_memcpy = ioat2_dma_prep_memcpy;
1636 device->common.device_issue_pending =
1637 ioat2_dma_memcpy_issue_pending;
1638 break;
1639 }
1640
1641 dev_err(&device->pdev->dev,
1642 "Intel(R) I/OAT DMA Engine found,"
1643 " %d channels, device version 0x%02x, driver version %s\n",
1644 device->common.chancnt, device->version, IOAT_DMA_VERSION);
1645
1646 err = ioat_dma_setup_interrupts(device);
1647 if (err)
1648 goto err_setup_interrupts;
1649
1650 err = ioat_dma_self_test(device);
1651 if (err)
1652 goto err_self_test;
1653
1654 ioat_set_tcp_copy_break(device);
1655
1656 dma_async_device_register(&device->common);
1657
1658 if (device->version != IOAT_VER_3_0) {
1659 INIT_DELAYED_WORK(&device->work, ioat_dma_chan_watchdog);
1660 schedule_delayed_work(&device->work,
1661 WATCHDOG_DELAY);
1662 }
1663
1664 return device;
1665
1666 err_self_test:
1667 ioat_dma_remove_interrupts(device);
1668 err_setup_interrupts:
1669 pci_pool_destroy(device->completion_pool);
1670 err_completion_pool:
1671 pci_pool_destroy(device->dma_pool);
1672 err_dma_pool:
1673 kfree(device);
1674 err_kzalloc:
1675 dev_err(&pdev->dev,
1676 "Intel(R) I/OAT DMA Engine initialization failed\n");
1677 return NULL;
1678 }
1679
1680 void ioat_dma_remove(struct ioatdma_device *device)
1681 {
1682 struct dma_chan *chan, *_chan;
1683 struct ioat_dma_chan *ioat_chan;
1684
1685 ioat_dma_remove_interrupts(device);
1686
1687 dma_async_device_unregister(&device->common);
1688
1689 pci_pool_destroy(device->dma_pool);
1690 pci_pool_destroy(device->completion_pool);
1691
1692 iounmap(device->reg_base);
1693 pci_release_regions(device->pdev);
1694 pci_disable_device(device->pdev);
1695
1696 if (device->version != IOAT_VER_3_0) {
1697 cancel_delayed_work(&device->work);
1698 }
1699
1700 list_for_each_entry_safe(chan, _chan,
1701 &device->common.channels, device_node) {
1702 ioat_chan = to_ioat_chan(chan);
1703 list_del(&chan->device_node);
1704 kfree(ioat_chan);
1705 }
1706 kfree(device);
1707 }
1708
Linus' kernel tree