Merge branch 'oprofile-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git...
[deliverable/linux.git] / drivers / dma / ioat_dma.c
1 /*
2 * Intel I/OAT DMA Linux driver
3 * Copyright(c) 2004 - 2009 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, 1) == 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_2_0)
193 writel(IOAT_DCACTRL_CMPL_WRITE_ENABLE |
194 IOAT_DMA_DCA_ANY_CPU,
195 ioat_chan->reg_base + IOAT_DCACTRL_OFFSET);
196 else if (ioat_chan->device->version == IOAT_VER_3_0)
197 writel(IOAT_DMA_DCA_ANY_CPU,
198 ioat_chan->reg_base + IOAT_DCACTRL_OFFSET);
199 spin_lock_init(&ioat_chan->cleanup_lock);
200 spin_lock_init(&ioat_chan->desc_lock);
201 INIT_LIST_HEAD(&ioat_chan->free_desc);
202 INIT_LIST_HEAD(&ioat_chan->used_desc);
203 /* This should be made common somewhere in dmaengine.c */
204 ioat_chan->common.device = &device->common;
205 list_add_tail(&ioat_chan->common.device_node,
206 &device->common.channels);
207 device->idx[i] = ioat_chan;
208 tasklet_init(&ioat_chan->cleanup_task,
209 ioat_dma_cleanup_tasklet,
210 (unsigned long) ioat_chan);
211 tasklet_disable(&ioat_chan->cleanup_task);
212 }
213 return device->common.chancnt;
214 }
215
216 /**
217 * ioat_dma_memcpy_issue_pending - push potentially unrecognized appended
218 * descriptors to hw
219 * @chan: DMA channel handle
220 */
221 static inline void __ioat1_dma_memcpy_issue_pending(
222 struct ioat_dma_chan *ioat_chan)
223 {
224 ioat_chan->pending = 0;
225 writeb(IOAT_CHANCMD_APPEND, ioat_chan->reg_base + IOAT1_CHANCMD_OFFSET);
226 }
227
228 static void ioat1_dma_memcpy_issue_pending(struct dma_chan *chan)
229 {
230 struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
231
232 if (ioat_chan->pending > 0) {
233 spin_lock_bh(&ioat_chan->desc_lock);
234 __ioat1_dma_memcpy_issue_pending(ioat_chan);
235 spin_unlock_bh(&ioat_chan->desc_lock);
236 }
237 }
238
239 static inline void __ioat2_dma_memcpy_issue_pending(
240 struct ioat_dma_chan *ioat_chan)
241 {
242 ioat_chan->pending = 0;
243 writew(ioat_chan->dmacount,
244 ioat_chan->reg_base + IOAT_CHAN_DMACOUNT_OFFSET);
245 }
246
247 static void ioat2_dma_memcpy_issue_pending(struct dma_chan *chan)
248 {
249 struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
250
251 if (ioat_chan->pending > 0) {
252 spin_lock_bh(&ioat_chan->desc_lock);
253 __ioat2_dma_memcpy_issue_pending(ioat_chan);
254 spin_unlock_bh(&ioat_chan->desc_lock);
255 }
256 }
257
258
259 /**
260 * ioat_dma_chan_reset_part2 - reinit the channel after a reset
261 */
262 static void ioat_dma_chan_reset_part2(struct work_struct *work)
263 {
264 struct ioat_dma_chan *ioat_chan =
265 container_of(work, struct ioat_dma_chan, work.work);
266 struct ioat_desc_sw *desc;
267
268 spin_lock_bh(&ioat_chan->cleanup_lock);
269 spin_lock_bh(&ioat_chan->desc_lock);
270
271 ioat_chan->completion_virt->low = 0;
272 ioat_chan->completion_virt->high = 0;
273 ioat_chan->pending = 0;
274
275 /*
276 * count the descriptors waiting, and be sure to do it
277 * right for both the CB1 line and the CB2 ring
278 */
279 ioat_chan->dmacount = 0;
280 if (ioat_chan->used_desc.prev) {
281 desc = to_ioat_desc(ioat_chan->used_desc.prev);
282 do {
283 ioat_chan->dmacount++;
284 desc = to_ioat_desc(desc->node.next);
285 } while (&desc->node != ioat_chan->used_desc.next);
286 }
287
288 /*
289 * write the new starting descriptor address
290 * this puts channel engine into ARMED state
291 */
292 desc = to_ioat_desc(ioat_chan->used_desc.prev);
293 switch (ioat_chan->device->version) {
294 case IOAT_VER_1_2:
295 writel(((u64) desc->async_tx.phys) & 0x00000000FFFFFFFF,
296 ioat_chan->reg_base + IOAT1_CHAINADDR_OFFSET_LOW);
297 writel(((u64) desc->async_tx.phys) >> 32,
298 ioat_chan->reg_base + IOAT1_CHAINADDR_OFFSET_HIGH);
299
300 writeb(IOAT_CHANCMD_START, ioat_chan->reg_base
301 + IOAT_CHANCMD_OFFSET(ioat_chan->device->version));
302 break;
303 case IOAT_VER_2_0:
304 writel(((u64) desc->async_tx.phys) & 0x00000000FFFFFFFF,
305 ioat_chan->reg_base + IOAT2_CHAINADDR_OFFSET_LOW);
306 writel(((u64) desc->async_tx.phys) >> 32,
307 ioat_chan->reg_base + IOAT2_CHAINADDR_OFFSET_HIGH);
308
309 /* tell the engine to go with what's left to be done */
310 writew(ioat_chan->dmacount,
311 ioat_chan->reg_base + IOAT_CHAN_DMACOUNT_OFFSET);
312
313 break;
314 }
315 dev_err(&ioat_chan->device->pdev->dev,
316 "chan%d reset - %d descs waiting, %d total desc\n",
317 chan_num(ioat_chan), ioat_chan->dmacount, ioat_chan->desccount);
318
319 spin_unlock_bh(&ioat_chan->desc_lock);
320 spin_unlock_bh(&ioat_chan->cleanup_lock);
321 }
322
323 /**
324 * ioat_dma_reset_channel - restart a channel
325 * @ioat_chan: IOAT DMA channel handle
326 */
327 static void ioat_dma_reset_channel(struct ioat_dma_chan *ioat_chan)
328 {
329 u32 chansts, chanerr;
330
331 if (!ioat_chan->used_desc.prev)
332 return;
333
334 chanerr = readl(ioat_chan->reg_base + IOAT_CHANERR_OFFSET);
335 chansts = (ioat_chan->completion_virt->low
336 & IOAT_CHANSTS_DMA_TRANSFER_STATUS);
337 if (chanerr) {
338 dev_err(&ioat_chan->device->pdev->dev,
339 "chan%d, CHANSTS = 0x%08x CHANERR = 0x%04x, clearing\n",
340 chan_num(ioat_chan), chansts, chanerr);
341 writel(chanerr, ioat_chan->reg_base + IOAT_CHANERR_OFFSET);
342 }
343
344 /*
345 * whack it upside the head with a reset
346 * and wait for things to settle out.
347 * force the pending count to a really big negative
348 * to make sure no one forces an issue_pending
349 * while we're waiting.
350 */
351
352 spin_lock_bh(&ioat_chan->desc_lock);
353 ioat_chan->pending = INT_MIN;
354 writeb(IOAT_CHANCMD_RESET,
355 ioat_chan->reg_base
356 + IOAT_CHANCMD_OFFSET(ioat_chan->device->version));
357 spin_unlock_bh(&ioat_chan->desc_lock);
358
359 /* schedule the 2nd half instead of sleeping a long time */
360 schedule_delayed_work(&ioat_chan->work, RESET_DELAY);
361 }
362
363 /**
364 * ioat_dma_chan_watchdog - watch for stuck channels
365 */
366 static void ioat_dma_chan_watchdog(struct work_struct *work)
367 {
368 struct ioatdma_device *device =
369 container_of(work, struct ioatdma_device, work.work);
370 struct ioat_dma_chan *ioat_chan;
371 int i;
372
373 union {
374 u64 full;
375 struct {
376 u32 low;
377 u32 high;
378 };
379 } completion_hw;
380 unsigned long compl_desc_addr_hw;
381
382 for (i = 0; i < device->common.chancnt; i++) {
383 ioat_chan = ioat_lookup_chan_by_index(device, i);
384
385 if (ioat_chan->device->version == IOAT_VER_1_2
386 /* have we started processing anything yet */
387 && ioat_chan->last_completion
388 /* have we completed any since last watchdog cycle? */
389 && (ioat_chan->last_completion ==
390 ioat_chan->watchdog_completion)
391 /* has TCP stuck on one cookie since last watchdog? */
392 && (ioat_chan->watchdog_tcp_cookie ==
393 ioat_chan->watchdog_last_tcp_cookie)
394 && (ioat_chan->watchdog_tcp_cookie !=
395 ioat_chan->completed_cookie)
396 /* is there something in the chain to be processed? */
397 /* CB1 chain always has at least the last one processed */
398 && (ioat_chan->used_desc.prev != ioat_chan->used_desc.next)
399 && ioat_chan->pending == 0) {
400
401 /*
402 * check CHANSTS register for completed
403 * descriptor address.
404 * if it is different than completion writeback,
405 * it is not zero
406 * and it has changed since the last watchdog
407 * we can assume that channel
408 * is still working correctly
409 * and the problem is in completion writeback.
410 * update completion writeback
411 * with actual CHANSTS value
412 * else
413 * try resetting the channel
414 */
415
416 completion_hw.low = readl(ioat_chan->reg_base +
417 IOAT_CHANSTS_OFFSET_LOW(ioat_chan->device->version));
418 completion_hw.high = readl(ioat_chan->reg_base +
419 IOAT_CHANSTS_OFFSET_HIGH(ioat_chan->device->version));
420 #if (BITS_PER_LONG == 64)
421 compl_desc_addr_hw =
422 completion_hw.full
423 & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_ADDR;
424 #else
425 compl_desc_addr_hw =
426 completion_hw.low & IOAT_LOW_COMPLETION_MASK;
427 #endif
428
429 if ((compl_desc_addr_hw != 0)
430 && (compl_desc_addr_hw != ioat_chan->watchdog_completion)
431 && (compl_desc_addr_hw != ioat_chan->last_compl_desc_addr_hw)) {
432 ioat_chan->last_compl_desc_addr_hw = compl_desc_addr_hw;
433 ioat_chan->completion_virt->low = completion_hw.low;
434 ioat_chan->completion_virt->high = completion_hw.high;
435 } else {
436 ioat_dma_reset_channel(ioat_chan);
437 ioat_chan->watchdog_completion = 0;
438 ioat_chan->last_compl_desc_addr_hw = 0;
439 }
440
441 /*
442 * for version 2.0 if there are descriptors yet to be processed
443 * and the last completed hasn't changed since the last watchdog
444 * if they haven't hit the pending level
445 * issue the pending to push them through
446 * else
447 * try resetting the channel
448 */
449 } else if (ioat_chan->device->version == IOAT_VER_2_0
450 && ioat_chan->used_desc.prev
451 && ioat_chan->last_completion
452 && ioat_chan->last_completion == ioat_chan->watchdog_completion) {
453
454 if (ioat_chan->pending < ioat_pending_level)
455 ioat2_dma_memcpy_issue_pending(&ioat_chan->common);
456 else {
457 ioat_dma_reset_channel(ioat_chan);
458 ioat_chan->watchdog_completion = 0;
459 }
460 } else {
461 ioat_chan->last_compl_desc_addr_hw = 0;
462 ioat_chan->watchdog_completion
463 = ioat_chan->last_completion;
464 }
465
466 ioat_chan->watchdog_last_tcp_cookie =
467 ioat_chan->watchdog_tcp_cookie;
468 }
469
470 schedule_delayed_work(&device->work, WATCHDOG_DELAY);
471 }
472
473 static dma_cookie_t ioat1_tx_submit(struct dma_async_tx_descriptor *tx)
474 {
475 struct ioat_dma_chan *ioat_chan = to_ioat_chan(tx->chan);
476 struct ioat_desc_sw *first = tx_to_ioat_desc(tx);
477 struct ioat_desc_sw *prev, *new;
478 struct ioat_dma_descriptor *hw;
479 dma_cookie_t cookie;
480 LIST_HEAD(new_chain);
481 u32 copy;
482 size_t len;
483 dma_addr_t src, dst;
484 unsigned long orig_flags;
485 unsigned int desc_count = 0;
486
487 /* src and dest and len are stored in the initial descriptor */
488 len = first->len;
489 src = first->src;
490 dst = first->dst;
491 orig_flags = first->async_tx.flags;
492 new = first;
493
494 spin_lock_bh(&ioat_chan->desc_lock);
495 prev = to_ioat_desc(ioat_chan->used_desc.prev);
496 prefetch(prev->hw);
497 do {
498 copy = min_t(size_t, len, ioat_chan->xfercap);
499
500 async_tx_ack(&new->async_tx);
501
502 hw = new->hw;
503 hw->size = copy;
504 hw->ctl = 0;
505 hw->src_addr = src;
506 hw->dst_addr = dst;
507 hw->next = 0;
508
509 /* chain together the physical address list for the HW */
510 wmb();
511 prev->hw->next = (u64) new->async_tx.phys;
512
513 len -= copy;
514 dst += copy;
515 src += copy;
516
517 list_add_tail(&new->node, &new_chain);
518 desc_count++;
519 prev = new;
520 } while (len && (new = ioat1_dma_get_next_descriptor(ioat_chan)));
521
522 if (!new) {
523 dev_err(&ioat_chan->device->pdev->dev,
524 "tx submit failed\n");
525 spin_unlock_bh(&ioat_chan->desc_lock);
526 return -ENOMEM;
527 }
528
529 hw->ctl = IOAT_DMA_DESCRIPTOR_CTL_CP_STS;
530 if (first->async_tx.callback) {
531 hw->ctl |= IOAT_DMA_DESCRIPTOR_CTL_INT_GN;
532 if (first != new) {
533 /* move callback into to last desc */
534 new->async_tx.callback = first->async_tx.callback;
535 new->async_tx.callback_param
536 = first->async_tx.callback_param;
537 first->async_tx.callback = NULL;
538 first->async_tx.callback_param = NULL;
539 }
540 }
541
542 new->tx_cnt = desc_count;
543 new->async_tx.flags = orig_flags; /* client is in control of this ack */
544
545 /* store the original values for use in later cleanup */
546 if (new != first) {
547 new->src = first->src;
548 new->dst = first->dst;
549 new->len = first->len;
550 }
551
552 /* cookie incr and addition to used_list must be atomic */
553 cookie = ioat_chan->common.cookie;
554 cookie++;
555 if (cookie < 0)
556 cookie = 1;
557 ioat_chan->common.cookie = new->async_tx.cookie = cookie;
558
559 /* write address into NextDescriptor field of last desc in chain */
560 to_ioat_desc(ioat_chan->used_desc.prev)->hw->next =
561 first->async_tx.phys;
562 list_splice_tail(&new_chain, &ioat_chan->used_desc);
563
564 ioat_chan->dmacount += desc_count;
565 ioat_chan->pending += desc_count;
566 if (ioat_chan->pending >= ioat_pending_level)
567 __ioat1_dma_memcpy_issue_pending(ioat_chan);
568 spin_unlock_bh(&ioat_chan->desc_lock);
569
570 return cookie;
571 }
572
573 static dma_cookie_t ioat2_tx_submit(struct dma_async_tx_descriptor *tx)
574 {
575 struct ioat_dma_chan *ioat_chan = to_ioat_chan(tx->chan);
576 struct ioat_desc_sw *first = tx_to_ioat_desc(tx);
577 struct ioat_desc_sw *new;
578 struct ioat_dma_descriptor *hw;
579 dma_cookie_t cookie;
580 u32 copy;
581 size_t len;
582 dma_addr_t src, dst;
583 unsigned long orig_flags;
584 unsigned int desc_count = 0;
585
586 /* src and dest and len are stored in the initial descriptor */
587 len = first->len;
588 src = first->src;
589 dst = first->dst;
590 orig_flags = first->async_tx.flags;
591 new = first;
592
593 /*
594 * ioat_chan->desc_lock is still in force in version 2 path
595 * it gets unlocked at end of this function
596 */
597 do {
598 copy = min_t(size_t, len, ioat_chan->xfercap);
599
600 async_tx_ack(&new->async_tx);
601
602 hw = new->hw;
603 hw->size = copy;
604 hw->ctl = 0;
605 hw->src_addr = src;
606 hw->dst_addr = dst;
607
608 len -= copy;
609 dst += copy;
610 src += copy;
611 desc_count++;
612 } while (len && (new = ioat2_dma_get_next_descriptor(ioat_chan)));
613
614 if (!new) {
615 dev_err(&ioat_chan->device->pdev->dev,
616 "tx submit failed\n");
617 spin_unlock_bh(&ioat_chan->desc_lock);
618 return -ENOMEM;
619 }
620
621 hw->ctl |= IOAT_DMA_DESCRIPTOR_CTL_CP_STS;
622 if (first->async_tx.callback) {
623 hw->ctl |= IOAT_DMA_DESCRIPTOR_CTL_INT_GN;
624 if (first != new) {
625 /* move callback into to last desc */
626 new->async_tx.callback = first->async_tx.callback;
627 new->async_tx.callback_param
628 = first->async_tx.callback_param;
629 first->async_tx.callback = NULL;
630 first->async_tx.callback_param = NULL;
631 }
632 }
633
634 new->tx_cnt = desc_count;
635 new->async_tx.flags = orig_flags; /* client is in control of this ack */
636
637 /* store the original values for use in later cleanup */
638 if (new != first) {
639 new->src = first->src;
640 new->dst = first->dst;
641 new->len = first->len;
642 }
643
644 /* cookie incr and addition to used_list must be atomic */
645 cookie = ioat_chan->common.cookie;
646 cookie++;
647 if (cookie < 0)
648 cookie = 1;
649 ioat_chan->common.cookie = new->async_tx.cookie = cookie;
650
651 ioat_chan->dmacount += desc_count;
652 ioat_chan->pending += desc_count;
653 if (ioat_chan->pending >= ioat_pending_level)
654 __ioat2_dma_memcpy_issue_pending(ioat_chan);
655 spin_unlock_bh(&ioat_chan->desc_lock);
656
657 return cookie;
658 }
659
660 /**
661 * ioat_dma_alloc_descriptor - allocate and return a sw and hw descriptor pair
662 * @ioat_chan: the channel supplying the memory pool for the descriptors
663 * @flags: allocation flags
664 */
665 static struct ioat_desc_sw *ioat_dma_alloc_descriptor(
666 struct ioat_dma_chan *ioat_chan,
667 gfp_t flags)
668 {
669 struct ioat_dma_descriptor *desc;
670 struct ioat_desc_sw *desc_sw;
671 struct ioatdma_device *ioatdma_device;
672 dma_addr_t phys;
673
674 ioatdma_device = to_ioatdma_device(ioat_chan->common.device);
675 desc = pci_pool_alloc(ioatdma_device->dma_pool, flags, &phys);
676 if (unlikely(!desc))
677 return NULL;
678
679 desc_sw = kzalloc(sizeof(*desc_sw), flags);
680 if (unlikely(!desc_sw)) {
681 pci_pool_free(ioatdma_device->dma_pool, desc, phys);
682 return NULL;
683 }
684
685 memset(desc, 0, sizeof(*desc));
686 dma_async_tx_descriptor_init(&desc_sw->async_tx, &ioat_chan->common);
687 switch (ioat_chan->device->version) {
688 case IOAT_VER_1_2:
689 desc_sw->async_tx.tx_submit = ioat1_tx_submit;
690 break;
691 case IOAT_VER_2_0:
692 case IOAT_VER_3_0:
693 desc_sw->async_tx.tx_submit = ioat2_tx_submit;
694 break;
695 }
696
697 desc_sw->hw = desc;
698 desc_sw->async_tx.phys = phys;
699
700 return desc_sw;
701 }
702
703 static int ioat_initial_desc_count = 256;
704 module_param(ioat_initial_desc_count, int, 0644);
705 MODULE_PARM_DESC(ioat_initial_desc_count,
706 "initial descriptors per channel (default: 256)");
707
708 /**
709 * ioat2_dma_massage_chan_desc - link the descriptors into a circle
710 * @ioat_chan: the channel to be massaged
711 */
712 static void ioat2_dma_massage_chan_desc(struct ioat_dma_chan *ioat_chan)
713 {
714 struct ioat_desc_sw *desc, *_desc;
715
716 /* setup used_desc */
717 ioat_chan->used_desc.next = ioat_chan->free_desc.next;
718 ioat_chan->used_desc.prev = NULL;
719
720 /* pull free_desc out of the circle so that every node is a hw
721 * descriptor, but leave it pointing to the list
722 */
723 ioat_chan->free_desc.prev->next = ioat_chan->free_desc.next;
724 ioat_chan->free_desc.next->prev = ioat_chan->free_desc.prev;
725
726 /* circle link the hw descriptors */
727 desc = to_ioat_desc(ioat_chan->free_desc.next);
728 desc->hw->next = to_ioat_desc(desc->node.next)->async_tx.phys;
729 list_for_each_entry_safe(desc, _desc, ioat_chan->free_desc.next, node) {
730 desc->hw->next = to_ioat_desc(desc->node.next)->async_tx.phys;
731 }
732 }
733
734 /**
735 * ioat_dma_alloc_chan_resources - returns the number of allocated descriptors
736 * @chan: the channel to be filled out
737 */
738 static int ioat_dma_alloc_chan_resources(struct dma_chan *chan)
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 /* Before freeing channel resources first check
811 * if they have been previously allocated for this channel.
812 */
813 if (ioat_chan->desccount == 0)
814 return;
815
816 tasklet_disable(&ioat_chan->cleanup_task);
817 ioat_dma_memcpy_cleanup(ioat_chan);
818
819 /* Delay 100ms after reset to allow internal DMA logic to quiesce
820 * before removing DMA descriptor resources.
821 */
822 writeb(IOAT_CHANCMD_RESET,
823 ioat_chan->reg_base
824 + IOAT_CHANCMD_OFFSET(ioat_chan->device->version));
825 mdelay(100);
826
827 spin_lock_bh(&ioat_chan->desc_lock);
828 switch (ioat_chan->device->version) {
829 case IOAT_VER_1_2:
830 list_for_each_entry_safe(desc, _desc,
831 &ioat_chan->used_desc, node) {
832 in_use_descs++;
833 list_del(&desc->node);
834 pci_pool_free(ioatdma_device->dma_pool, desc->hw,
835 desc->async_tx.phys);
836 kfree(desc);
837 }
838 list_for_each_entry_safe(desc, _desc,
839 &ioat_chan->free_desc, node) {
840 list_del(&desc->node);
841 pci_pool_free(ioatdma_device->dma_pool, desc->hw,
842 desc->async_tx.phys);
843 kfree(desc);
844 }
845 break;
846 case IOAT_VER_2_0:
847 case IOAT_VER_3_0:
848 list_for_each_entry_safe(desc, _desc,
849 ioat_chan->free_desc.next, node) {
850 list_del(&desc->node);
851 pci_pool_free(ioatdma_device->dma_pool, desc->hw,
852 desc->async_tx.phys);
853 kfree(desc);
854 }
855 desc = to_ioat_desc(ioat_chan->free_desc.next);
856 pci_pool_free(ioatdma_device->dma_pool, desc->hw,
857 desc->async_tx.phys);
858 kfree(desc);
859 INIT_LIST_HEAD(&ioat_chan->free_desc);
860 INIT_LIST_HEAD(&ioat_chan->used_desc);
861 break;
862 }
863 spin_unlock_bh(&ioat_chan->desc_lock);
864
865 pci_pool_free(ioatdma_device->completion_pool,
866 ioat_chan->completion_virt,
867 ioat_chan->completion_addr);
868
869 /* one is ok since we left it on there on purpose */
870 if (in_use_descs > 1)
871 dev_err(&ioat_chan->device->pdev->dev,
872 "Freeing %d in use descriptors!\n",
873 in_use_descs - 1);
874
875 ioat_chan->last_completion = ioat_chan->completion_addr = 0;
876 ioat_chan->pending = 0;
877 ioat_chan->dmacount = 0;
878 ioat_chan->desccount = 0;
879 ioat_chan->watchdog_completion = 0;
880 ioat_chan->last_compl_desc_addr_hw = 0;
881 ioat_chan->watchdog_tcp_cookie =
882 ioat_chan->watchdog_last_tcp_cookie = 0;
883 }
884
885 /**
886 * ioat_dma_get_next_descriptor - return the next available descriptor
887 * @ioat_chan: IOAT DMA channel handle
888 *
889 * Gets the next descriptor from the chain, and must be called with the
890 * channel's desc_lock held. Allocates more descriptors if the channel
891 * has run out.
892 */
893 static struct ioat_desc_sw *
894 ioat1_dma_get_next_descriptor(struct ioat_dma_chan *ioat_chan)
895 {
896 struct ioat_desc_sw *new;
897
898 if (!list_empty(&ioat_chan->free_desc)) {
899 new = to_ioat_desc(ioat_chan->free_desc.next);
900 list_del(&new->node);
901 } else {
902 /* try to get another desc */
903 new = ioat_dma_alloc_descriptor(ioat_chan, GFP_ATOMIC);
904 if (!new) {
905 dev_err(&ioat_chan->device->pdev->dev,
906 "alloc failed\n");
907 return NULL;
908 }
909 }
910
911 prefetch(new->hw);
912 return new;
913 }
914
915 static struct ioat_desc_sw *
916 ioat2_dma_get_next_descriptor(struct ioat_dma_chan *ioat_chan)
917 {
918 struct ioat_desc_sw *new;
919
920 /*
921 * used.prev points to where to start processing
922 * used.next points to next free descriptor
923 * if used.prev == NULL, there are none waiting to be processed
924 * if used.next == used.prev.prev, there is only one free descriptor,
925 * and we need to use it to as a noop descriptor before
926 * linking in a new set of descriptors, since the device
927 * has probably already read the pointer to it
928 */
929 if (ioat_chan->used_desc.prev &&
930 ioat_chan->used_desc.next == ioat_chan->used_desc.prev->prev) {
931
932 struct ioat_desc_sw *desc;
933 struct ioat_desc_sw *noop_desc;
934 int i;
935
936 /* set up the noop descriptor */
937 noop_desc = to_ioat_desc(ioat_chan->used_desc.next);
938 /* set size to non-zero value (channel returns error when size is 0) */
939 noop_desc->hw->size = NULL_DESC_BUFFER_SIZE;
940 noop_desc->hw->ctl = IOAT_DMA_DESCRIPTOR_NUL;
941 noop_desc->hw->src_addr = 0;
942 noop_desc->hw->dst_addr = 0;
943
944 ioat_chan->used_desc.next = ioat_chan->used_desc.next->next;
945 ioat_chan->pending++;
946 ioat_chan->dmacount++;
947
948 /* try to get a few more descriptors */
949 for (i = 16; i; i--) {
950 desc = ioat_dma_alloc_descriptor(ioat_chan, GFP_ATOMIC);
951 if (!desc) {
952 dev_err(&ioat_chan->device->pdev->dev,
953 "alloc failed\n");
954 break;
955 }
956 list_add_tail(&desc->node, ioat_chan->used_desc.next);
957
958 desc->hw->next
959 = to_ioat_desc(desc->node.next)->async_tx.phys;
960 to_ioat_desc(desc->node.prev)->hw->next
961 = desc->async_tx.phys;
962 ioat_chan->desccount++;
963 }
964
965 ioat_chan->used_desc.next = noop_desc->node.next;
966 }
967 new = to_ioat_desc(ioat_chan->used_desc.next);
968 prefetch(new);
969 ioat_chan->used_desc.next = new->node.next;
970
971 if (ioat_chan->used_desc.prev == NULL)
972 ioat_chan->used_desc.prev = &new->node;
973
974 prefetch(new->hw);
975 return new;
976 }
977
978 static struct ioat_desc_sw *ioat_dma_get_next_descriptor(
979 struct ioat_dma_chan *ioat_chan)
980 {
981 if (!ioat_chan)
982 return NULL;
983
984 switch (ioat_chan->device->version) {
985 case IOAT_VER_1_2:
986 return ioat1_dma_get_next_descriptor(ioat_chan);
987 case IOAT_VER_2_0:
988 case IOAT_VER_3_0:
989 return ioat2_dma_get_next_descriptor(ioat_chan);
990 }
991 return NULL;
992 }
993
994 static struct dma_async_tx_descriptor *ioat1_dma_prep_memcpy(
995 struct dma_chan *chan,
996 dma_addr_t dma_dest,
997 dma_addr_t dma_src,
998 size_t len,
999 unsigned long flags)
1000 {
1001 struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
1002 struct ioat_desc_sw *new;
1003
1004 spin_lock_bh(&ioat_chan->desc_lock);
1005 new = ioat_dma_get_next_descriptor(ioat_chan);
1006 spin_unlock_bh(&ioat_chan->desc_lock);
1007
1008 if (new) {
1009 new->len = len;
1010 new->dst = dma_dest;
1011 new->src = dma_src;
1012 new->async_tx.flags = flags;
1013 return &new->async_tx;
1014 } else {
1015 dev_err(&ioat_chan->device->pdev->dev,
1016 "chan%d - get_next_desc failed: %d descs waiting, %d total desc\n",
1017 chan_num(ioat_chan), ioat_chan->dmacount, ioat_chan->desccount);
1018 return NULL;
1019 }
1020 }
1021
1022 static struct dma_async_tx_descriptor *ioat2_dma_prep_memcpy(
1023 struct dma_chan *chan,
1024 dma_addr_t dma_dest,
1025 dma_addr_t dma_src,
1026 size_t len,
1027 unsigned long flags)
1028 {
1029 struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
1030 struct ioat_desc_sw *new;
1031
1032 spin_lock_bh(&ioat_chan->desc_lock);
1033 new = ioat2_dma_get_next_descriptor(ioat_chan);
1034
1035 /*
1036 * leave ioat_chan->desc_lock set in ioat 2 path
1037 * it will get unlocked at end of tx_submit
1038 */
1039
1040 if (new) {
1041 new->len = len;
1042 new->dst = dma_dest;
1043 new->src = dma_src;
1044 new->async_tx.flags = flags;
1045 return &new->async_tx;
1046 } else {
1047 spin_unlock_bh(&ioat_chan->desc_lock);
1048 dev_err(&ioat_chan->device->pdev->dev,
1049 "chan%d - get_next_desc failed: %d descs waiting, %d total desc\n",
1050 chan_num(ioat_chan), ioat_chan->dmacount, ioat_chan->desccount);
1051 return NULL;
1052 }
1053 }
1054
1055 static void ioat_dma_cleanup_tasklet(unsigned long data)
1056 {
1057 struct ioat_dma_chan *chan = (void *)data;
1058 ioat_dma_memcpy_cleanup(chan);
1059 writew(IOAT_CHANCTRL_INT_DISABLE,
1060 chan->reg_base + IOAT_CHANCTRL_OFFSET);
1061 }
1062
1063 static void
1064 ioat_dma_unmap(struct ioat_dma_chan *ioat_chan, struct ioat_desc_sw *desc)
1065 {
1066 if (!(desc->async_tx.flags & DMA_COMPL_SKIP_DEST_UNMAP)) {
1067 if (desc->async_tx.flags & DMA_COMPL_DEST_UNMAP_SINGLE)
1068 pci_unmap_single(ioat_chan->device->pdev,
1069 pci_unmap_addr(desc, dst),
1070 pci_unmap_len(desc, len),
1071 PCI_DMA_FROMDEVICE);
1072 else
1073 pci_unmap_page(ioat_chan->device->pdev,
1074 pci_unmap_addr(desc, dst),
1075 pci_unmap_len(desc, len),
1076 PCI_DMA_FROMDEVICE);
1077 }
1078
1079 if (!(desc->async_tx.flags & DMA_COMPL_SKIP_SRC_UNMAP)) {
1080 if (desc->async_tx.flags & DMA_COMPL_SRC_UNMAP_SINGLE)
1081 pci_unmap_single(ioat_chan->device->pdev,
1082 pci_unmap_addr(desc, src),
1083 pci_unmap_len(desc, len),
1084 PCI_DMA_TODEVICE);
1085 else
1086 pci_unmap_page(ioat_chan->device->pdev,
1087 pci_unmap_addr(desc, src),
1088 pci_unmap_len(desc, len),
1089 PCI_DMA_TODEVICE);
1090 }
1091 }
1092
1093 /**
1094 * ioat_dma_memcpy_cleanup - cleanup up finished descriptors
1095 * @chan: ioat channel to be cleaned up
1096 */
1097 static void ioat_dma_memcpy_cleanup(struct ioat_dma_chan *ioat_chan)
1098 {
1099 unsigned long phys_complete;
1100 struct ioat_desc_sw *desc, *_desc;
1101 dma_cookie_t cookie = 0;
1102 unsigned long desc_phys;
1103 struct ioat_desc_sw *latest_desc;
1104
1105 prefetch(ioat_chan->completion_virt);
1106
1107 if (!spin_trylock_bh(&ioat_chan->cleanup_lock))
1108 return;
1109
1110 /* The completion writeback can happen at any time,
1111 so reads by the driver need to be atomic operations
1112 The descriptor physical addresses are limited to 32-bits
1113 when the CPU can only do a 32-bit mov */
1114
1115 #if (BITS_PER_LONG == 64)
1116 phys_complete =
1117 ioat_chan->completion_virt->full
1118 & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_ADDR;
1119 #else
1120 phys_complete =
1121 ioat_chan->completion_virt->low & IOAT_LOW_COMPLETION_MASK;
1122 #endif
1123
1124 if ((ioat_chan->completion_virt->full
1125 & IOAT_CHANSTS_DMA_TRANSFER_STATUS) ==
1126 IOAT_CHANSTS_DMA_TRANSFER_STATUS_HALTED) {
1127 dev_err(&ioat_chan->device->pdev->dev,
1128 "Channel halted, chanerr = %x\n",
1129 readl(ioat_chan->reg_base + IOAT_CHANERR_OFFSET));
1130
1131 /* TODO do something to salvage the situation */
1132 }
1133
1134 if (phys_complete == ioat_chan->last_completion) {
1135 spin_unlock_bh(&ioat_chan->cleanup_lock);
1136 /*
1137 * perhaps we're stuck so hard that the watchdog can't go off?
1138 * try to catch it after 2 seconds
1139 */
1140 if (ioat_chan->device->version != IOAT_VER_3_0) {
1141 if (time_after(jiffies,
1142 ioat_chan->last_completion_time + HZ*WATCHDOG_DELAY)) {
1143 ioat_dma_chan_watchdog(&(ioat_chan->device->work.work));
1144 ioat_chan->last_completion_time = jiffies;
1145 }
1146 }
1147 return;
1148 }
1149 ioat_chan->last_completion_time = jiffies;
1150
1151 cookie = 0;
1152 if (!spin_trylock_bh(&ioat_chan->desc_lock)) {
1153 spin_unlock_bh(&ioat_chan->cleanup_lock);
1154 return;
1155 }
1156
1157 switch (ioat_chan->device->version) {
1158 case IOAT_VER_1_2:
1159 list_for_each_entry_safe(desc, _desc,
1160 &ioat_chan->used_desc, node) {
1161
1162 /*
1163 * Incoming DMA requests may use multiple descriptors,
1164 * due to exceeding xfercap, perhaps. If so, only the
1165 * last one will have a cookie, and require unmapping.
1166 */
1167 if (desc->async_tx.cookie) {
1168 cookie = desc->async_tx.cookie;
1169 ioat_dma_unmap(ioat_chan, desc);
1170 if (desc->async_tx.callback) {
1171 desc->async_tx.callback(desc->async_tx.callback_param);
1172 desc->async_tx.callback = NULL;
1173 }
1174 }
1175
1176 if (desc->async_tx.phys != phys_complete) {
1177 /*
1178 * a completed entry, but not the last, so clean
1179 * up if the client is done with the descriptor
1180 */
1181 if (async_tx_test_ack(&desc->async_tx)) {
1182 list_move_tail(&desc->node,
1183 &ioat_chan->free_desc);
1184 } else
1185 desc->async_tx.cookie = 0;
1186 } else {
1187 /*
1188 * last used desc. Do not remove, so we can
1189 * append from it, but don't look at it next
1190 * time, either
1191 */
1192 desc->async_tx.cookie = 0;
1193
1194 /* TODO check status bits? */
1195 break;
1196 }
1197 }
1198 break;
1199 case IOAT_VER_2_0:
1200 case IOAT_VER_3_0:
1201 /* has some other thread has already cleaned up? */
1202 if (ioat_chan->used_desc.prev == NULL)
1203 break;
1204
1205 /* work backwards to find latest finished desc */
1206 desc = to_ioat_desc(ioat_chan->used_desc.next);
1207 latest_desc = NULL;
1208 do {
1209 desc = to_ioat_desc(desc->node.prev);
1210 desc_phys = (unsigned long)desc->async_tx.phys
1211 & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_ADDR;
1212 if (desc_phys == phys_complete) {
1213 latest_desc = desc;
1214 break;
1215 }
1216 } while (&desc->node != ioat_chan->used_desc.prev);
1217
1218 if (latest_desc != NULL) {
1219
1220 /* work forwards to clear finished descriptors */
1221 for (desc = to_ioat_desc(ioat_chan->used_desc.prev);
1222 &desc->node != latest_desc->node.next &&
1223 &desc->node != ioat_chan->used_desc.next;
1224 desc = to_ioat_desc(desc->node.next)) {
1225 if (desc->async_tx.cookie) {
1226 cookie = desc->async_tx.cookie;
1227 desc->async_tx.cookie = 0;
1228 ioat_dma_unmap(ioat_chan, desc);
1229 if (desc->async_tx.callback) {
1230 desc->async_tx.callback(desc->async_tx.callback_param);
1231 desc->async_tx.callback = NULL;
1232 }
1233 }
1234 }
1235
1236 /* move used.prev up beyond those that are finished */
1237 if (&desc->node == ioat_chan->used_desc.next)
1238 ioat_chan->used_desc.prev = NULL;
1239 else
1240 ioat_chan->used_desc.prev = &desc->node;
1241 }
1242 break;
1243 }
1244
1245 spin_unlock_bh(&ioat_chan->desc_lock);
1246
1247 ioat_chan->last_completion = phys_complete;
1248 if (cookie != 0)
1249 ioat_chan->completed_cookie = cookie;
1250
1251 spin_unlock_bh(&ioat_chan->cleanup_lock);
1252 }
1253
1254 /**
1255 * ioat_dma_is_complete - poll the status of a IOAT DMA transaction
1256 * @chan: IOAT DMA channel handle
1257 * @cookie: DMA transaction identifier
1258 * @done: if not %NULL, updated with last completed transaction
1259 * @used: if not %NULL, updated with last used transaction
1260 */
1261 static enum dma_status ioat_dma_is_complete(struct dma_chan *chan,
1262 dma_cookie_t cookie,
1263 dma_cookie_t *done,
1264 dma_cookie_t *used)
1265 {
1266 struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
1267 dma_cookie_t last_used;
1268 dma_cookie_t last_complete;
1269 enum dma_status ret;
1270
1271 last_used = chan->cookie;
1272 last_complete = ioat_chan->completed_cookie;
1273 ioat_chan->watchdog_tcp_cookie = cookie;
1274
1275 if (done)
1276 *done = last_complete;
1277 if (used)
1278 *used = last_used;
1279
1280 ret = dma_async_is_complete(cookie, last_complete, last_used);
1281 if (ret == DMA_SUCCESS)
1282 return ret;
1283
1284 ioat_dma_memcpy_cleanup(ioat_chan);
1285
1286 last_used = chan->cookie;
1287 last_complete = ioat_chan->completed_cookie;
1288
1289 if (done)
1290 *done = last_complete;
1291 if (used)
1292 *used = last_used;
1293
1294 return dma_async_is_complete(cookie, last_complete, last_used);
1295 }
1296
1297 static void ioat_dma_start_null_desc(struct ioat_dma_chan *ioat_chan)
1298 {
1299 struct ioat_desc_sw *desc;
1300
1301 spin_lock_bh(&ioat_chan->desc_lock);
1302
1303 desc = ioat_dma_get_next_descriptor(ioat_chan);
1304
1305 if (!desc) {
1306 dev_err(&ioat_chan->device->pdev->dev,
1307 "Unable to start null desc - get next desc failed\n");
1308 spin_unlock_bh(&ioat_chan->desc_lock);
1309 return;
1310 }
1311
1312 desc->hw->ctl = IOAT_DMA_DESCRIPTOR_NUL
1313 | IOAT_DMA_DESCRIPTOR_CTL_INT_GN
1314 | IOAT_DMA_DESCRIPTOR_CTL_CP_STS;
1315 /* set size to non-zero value (channel returns error when size is 0) */
1316 desc->hw->size = NULL_DESC_BUFFER_SIZE;
1317 desc->hw->src_addr = 0;
1318 desc->hw->dst_addr = 0;
1319 async_tx_ack(&desc->async_tx);
1320 switch (ioat_chan->device->version) {
1321 case IOAT_VER_1_2:
1322 desc->hw->next = 0;
1323 list_add_tail(&desc->node, &ioat_chan->used_desc);
1324
1325 writel(((u64) desc->async_tx.phys) & 0x00000000FFFFFFFF,
1326 ioat_chan->reg_base + IOAT1_CHAINADDR_OFFSET_LOW);
1327 writel(((u64) desc->async_tx.phys) >> 32,
1328 ioat_chan->reg_base + IOAT1_CHAINADDR_OFFSET_HIGH);
1329
1330 writeb(IOAT_CHANCMD_START, ioat_chan->reg_base
1331 + IOAT_CHANCMD_OFFSET(ioat_chan->device->version));
1332 break;
1333 case IOAT_VER_2_0:
1334 case IOAT_VER_3_0:
1335 writel(((u64) desc->async_tx.phys) & 0x00000000FFFFFFFF,
1336 ioat_chan->reg_base + IOAT2_CHAINADDR_OFFSET_LOW);
1337 writel(((u64) desc->async_tx.phys) >> 32,
1338 ioat_chan->reg_base + IOAT2_CHAINADDR_OFFSET_HIGH);
1339
1340 ioat_chan->dmacount++;
1341 __ioat2_dma_memcpy_issue_pending(ioat_chan);
1342 break;
1343 }
1344 spin_unlock_bh(&ioat_chan->desc_lock);
1345 }
1346
1347 /*
1348 * Perform a IOAT transaction to verify the HW works.
1349 */
1350 #define IOAT_TEST_SIZE 2000
1351
1352 static void ioat_dma_test_callback(void *dma_async_param)
1353 {
1354 struct completion *cmp = dma_async_param;
1355
1356 complete(cmp);
1357 }
1358
1359 /**
1360 * ioat_dma_self_test - Perform a IOAT transaction to verify the HW works.
1361 * @device: device to be tested
1362 */
1363 static int ioat_dma_self_test(struct ioatdma_device *device)
1364 {
1365 int i;
1366 u8 *src;
1367 u8 *dest;
1368 struct dma_chan *dma_chan;
1369 struct dma_async_tx_descriptor *tx;
1370 dma_addr_t dma_dest, dma_src;
1371 dma_cookie_t cookie;
1372 int err = 0;
1373 struct completion cmp;
1374 unsigned long tmo;
1375 unsigned long flags;
1376
1377 src = kzalloc(sizeof(u8) * IOAT_TEST_SIZE, GFP_KERNEL);
1378 if (!src)
1379 return -ENOMEM;
1380 dest = kzalloc(sizeof(u8) * IOAT_TEST_SIZE, GFP_KERNEL);
1381 if (!dest) {
1382 kfree(src);
1383 return -ENOMEM;
1384 }
1385
1386 /* Fill in src buffer */
1387 for (i = 0; i < IOAT_TEST_SIZE; i++)
1388 src[i] = (u8)i;
1389
1390 /* Start copy, using first DMA channel */
1391 dma_chan = container_of(device->common.channels.next,
1392 struct dma_chan,
1393 device_node);
1394 if (device->common.device_alloc_chan_resources(dma_chan) < 1) {
1395 dev_err(&device->pdev->dev,
1396 "selftest cannot allocate chan resource\n");
1397 err = -ENODEV;
1398 goto out;
1399 }
1400
1401 dma_src = dma_map_single(dma_chan->device->dev, src, IOAT_TEST_SIZE,
1402 DMA_TO_DEVICE);
1403 dma_dest = dma_map_single(dma_chan->device->dev, dest, IOAT_TEST_SIZE,
1404 DMA_FROM_DEVICE);
1405 flags = DMA_COMPL_SRC_UNMAP_SINGLE | DMA_COMPL_DEST_UNMAP_SINGLE;
1406 tx = device->common.device_prep_dma_memcpy(dma_chan, dma_dest, dma_src,
1407 IOAT_TEST_SIZE, flags);
1408 if (!tx) {
1409 dev_err(&device->pdev->dev,
1410 "Self-test prep failed, disabling\n");
1411 err = -ENODEV;
1412 goto free_resources;
1413 }
1414
1415 async_tx_ack(tx);
1416 init_completion(&cmp);
1417 tx->callback = ioat_dma_test_callback;
1418 tx->callback_param = &cmp;
1419 cookie = tx->tx_submit(tx);
1420 if (cookie < 0) {
1421 dev_err(&device->pdev->dev,
1422 "Self-test setup failed, disabling\n");
1423 err = -ENODEV;
1424 goto free_resources;
1425 }
1426 device->common.device_issue_pending(dma_chan);
1427
1428 tmo = wait_for_completion_timeout(&cmp, msecs_to_jiffies(3000));
1429
1430 if (tmo == 0 ||
1431 device->common.device_is_tx_complete(dma_chan, cookie, NULL, NULL)
1432 != DMA_SUCCESS) {
1433 dev_err(&device->pdev->dev,
1434 "Self-test copy timed out, disabling\n");
1435 err = -ENODEV;
1436 goto free_resources;
1437 }
1438 if (memcmp(src, dest, IOAT_TEST_SIZE)) {
1439 dev_err(&device->pdev->dev,
1440 "Self-test copy failed compare, disabling\n");
1441 err = -ENODEV;
1442 goto free_resources;
1443 }
1444
1445 free_resources:
1446 device->common.device_free_chan_resources(dma_chan);
1447 out:
1448 kfree(src);
1449 kfree(dest);
1450 return err;
1451 }
1452
1453 static char ioat_interrupt_style[32] = "msix";
1454 module_param_string(ioat_interrupt_style, ioat_interrupt_style,
1455 sizeof(ioat_interrupt_style), 0644);
1456 MODULE_PARM_DESC(ioat_interrupt_style,
1457 "set ioat interrupt style: msix (default), "
1458 "msix-single-vector, msi, intx)");
1459
1460 /**
1461 * ioat_dma_setup_interrupts - setup interrupt handler
1462 * @device: ioat device
1463 */
1464 static int ioat_dma_setup_interrupts(struct ioatdma_device *device)
1465 {
1466 struct ioat_dma_chan *ioat_chan;
1467 int err, i, j, msixcnt;
1468 u8 intrctrl = 0;
1469
1470 if (!strcmp(ioat_interrupt_style, "msix"))
1471 goto msix;
1472 if (!strcmp(ioat_interrupt_style, "msix-single-vector"))
1473 goto msix_single_vector;
1474 if (!strcmp(ioat_interrupt_style, "msi"))
1475 goto msi;
1476 if (!strcmp(ioat_interrupt_style, "intx"))
1477 goto intx;
1478 dev_err(&device->pdev->dev, "invalid ioat_interrupt_style %s\n",
1479 ioat_interrupt_style);
1480 goto err_no_irq;
1481
1482 msix:
1483 /* The number of MSI-X vectors should equal the number of channels */
1484 msixcnt = device->common.chancnt;
1485 for (i = 0; i < msixcnt; i++)
1486 device->msix_entries[i].entry = i;
1487
1488 err = pci_enable_msix(device->pdev, device->msix_entries, msixcnt);
1489 if (err < 0)
1490 goto msi;
1491 if (err > 0)
1492 goto msix_single_vector;
1493
1494 for (i = 0; i < msixcnt; i++) {
1495 ioat_chan = ioat_lookup_chan_by_index(device, i);
1496 err = request_irq(device->msix_entries[i].vector,
1497 ioat_dma_do_interrupt_msix,
1498 0, "ioat-msix", ioat_chan);
1499 if (err) {
1500 for (j = 0; j < i; j++) {
1501 ioat_chan =
1502 ioat_lookup_chan_by_index(device, j);
1503 free_irq(device->msix_entries[j].vector,
1504 ioat_chan);
1505 }
1506 goto msix_single_vector;
1507 }
1508 }
1509 intrctrl |= IOAT_INTRCTRL_MSIX_VECTOR_CONTROL;
1510 device->irq_mode = msix_multi_vector;
1511 goto done;
1512
1513 msix_single_vector:
1514 device->msix_entries[0].entry = 0;
1515 err = pci_enable_msix(device->pdev, device->msix_entries, 1);
1516 if (err)
1517 goto msi;
1518
1519 err = request_irq(device->msix_entries[0].vector, ioat_dma_do_interrupt,
1520 0, "ioat-msix", device);
1521 if (err) {
1522 pci_disable_msix(device->pdev);
1523 goto msi;
1524 }
1525 device->irq_mode = msix_single_vector;
1526 goto done;
1527
1528 msi:
1529 err = pci_enable_msi(device->pdev);
1530 if (err)
1531 goto intx;
1532
1533 err = request_irq(device->pdev->irq, ioat_dma_do_interrupt,
1534 0, "ioat-msi", device);
1535 if (err) {
1536 pci_disable_msi(device->pdev);
1537 goto intx;
1538 }
1539 /*
1540 * CB 1.2 devices need a bit set in configuration space to enable MSI
1541 */
1542 if (device->version == IOAT_VER_1_2) {
1543 u32 dmactrl;
1544 pci_read_config_dword(device->pdev,
1545 IOAT_PCI_DMACTRL_OFFSET, &dmactrl);
1546 dmactrl |= IOAT_PCI_DMACTRL_MSI_EN;
1547 pci_write_config_dword(device->pdev,
1548 IOAT_PCI_DMACTRL_OFFSET, dmactrl);
1549 }
1550 device->irq_mode = msi;
1551 goto done;
1552
1553 intx:
1554 err = request_irq(device->pdev->irq, ioat_dma_do_interrupt,
1555 IRQF_SHARED, "ioat-intx", device);
1556 if (err)
1557 goto err_no_irq;
1558 device->irq_mode = intx;
1559
1560 done:
1561 intrctrl |= IOAT_INTRCTRL_MASTER_INT_EN;
1562 writeb(intrctrl, device->reg_base + IOAT_INTRCTRL_OFFSET);
1563 return 0;
1564
1565 err_no_irq:
1566 /* Disable all interrupt generation */
1567 writeb(0, device->reg_base + IOAT_INTRCTRL_OFFSET);
1568 dev_err(&device->pdev->dev, "no usable interrupts\n");
1569 device->irq_mode = none;
1570 return -1;
1571 }
1572
1573 /**
1574 * ioat_dma_remove_interrupts - remove whatever interrupts were set
1575 * @device: ioat device
1576 */
1577 static void ioat_dma_remove_interrupts(struct ioatdma_device *device)
1578 {
1579 struct ioat_dma_chan *ioat_chan;
1580 int i;
1581
1582 /* Disable all interrupt generation */
1583 writeb(0, device->reg_base + IOAT_INTRCTRL_OFFSET);
1584
1585 switch (device->irq_mode) {
1586 case msix_multi_vector:
1587 for (i = 0; i < device->common.chancnt; i++) {
1588 ioat_chan = ioat_lookup_chan_by_index(device, i);
1589 free_irq(device->msix_entries[i].vector, ioat_chan);
1590 }
1591 pci_disable_msix(device->pdev);
1592 break;
1593 case msix_single_vector:
1594 free_irq(device->msix_entries[0].vector, device);
1595 pci_disable_msix(device->pdev);
1596 break;
1597 case msi:
1598 free_irq(device->pdev->irq, device);
1599 pci_disable_msi(device->pdev);
1600 break;
1601 case intx:
1602 free_irq(device->pdev->irq, device);
1603 break;
1604 case none:
1605 dev_warn(&device->pdev->dev,
1606 "call to %s without interrupts setup\n", __func__);
1607 }
1608 device->irq_mode = none;
1609 }
1610
1611 struct ioatdma_device *ioat_dma_probe(struct pci_dev *pdev,
1612 void __iomem *iobase)
1613 {
1614 int err;
1615 struct ioatdma_device *device;
1616
1617 device = kzalloc(sizeof(*device), GFP_KERNEL);
1618 if (!device) {
1619 err = -ENOMEM;
1620 goto err_kzalloc;
1621 }
1622 device->pdev = pdev;
1623 device->reg_base = iobase;
1624 device->version = readb(device->reg_base + IOAT_VER_OFFSET);
1625
1626 /* DMA coherent memory pool for DMA descriptor allocations */
1627 device->dma_pool = pci_pool_create("dma_desc_pool", pdev,
1628 sizeof(struct ioat_dma_descriptor),
1629 64, 0);
1630 if (!device->dma_pool) {
1631 err = -ENOMEM;
1632 goto err_dma_pool;
1633 }
1634
1635 device->completion_pool = pci_pool_create("completion_pool", pdev,
1636 sizeof(u64), SMP_CACHE_BYTES,
1637 SMP_CACHE_BYTES);
1638 if (!device->completion_pool) {
1639 err = -ENOMEM;
1640 goto err_completion_pool;
1641 }
1642
1643 INIT_LIST_HEAD(&device->common.channels);
1644 ioat_dma_enumerate_channels(device);
1645
1646 device->common.device_alloc_chan_resources =
1647 ioat_dma_alloc_chan_resources;
1648 device->common.device_free_chan_resources =
1649 ioat_dma_free_chan_resources;
1650 device->common.dev = &pdev->dev;
1651
1652 dma_cap_set(DMA_MEMCPY, device->common.cap_mask);
1653 device->common.device_is_tx_complete = ioat_dma_is_complete;
1654 switch (device->version) {
1655 case IOAT_VER_1_2:
1656 device->common.device_prep_dma_memcpy = ioat1_dma_prep_memcpy;
1657 device->common.device_issue_pending =
1658 ioat1_dma_memcpy_issue_pending;
1659 break;
1660 case IOAT_VER_2_0:
1661 case IOAT_VER_3_0:
1662 device->common.device_prep_dma_memcpy = ioat2_dma_prep_memcpy;
1663 device->common.device_issue_pending =
1664 ioat2_dma_memcpy_issue_pending;
1665 break;
1666 }
1667
1668 dev_err(&device->pdev->dev,
1669 "Intel(R) I/OAT DMA Engine found,"
1670 " %d channels, device version 0x%02x, driver version %s\n",
1671 device->common.chancnt, device->version, IOAT_DMA_VERSION);
1672
1673 if (!device->common.chancnt) {
1674 dev_err(&device->pdev->dev,
1675 "Intel(R) I/OAT DMA Engine problem found: "
1676 "zero channels detected\n");
1677 goto err_setup_interrupts;
1678 }
1679
1680 err = ioat_dma_setup_interrupts(device);
1681 if (err)
1682 goto err_setup_interrupts;
1683
1684 err = ioat_dma_self_test(device);
1685 if (err)
1686 goto err_self_test;
1687
1688 ioat_set_tcp_copy_break(device);
1689
1690 dma_async_device_register(&device->common);
1691
1692 if (device->version != IOAT_VER_3_0) {
1693 INIT_DELAYED_WORK(&device->work, ioat_dma_chan_watchdog);
1694 schedule_delayed_work(&device->work,
1695 WATCHDOG_DELAY);
1696 }
1697
1698 return device;
1699
1700 err_self_test:
1701 ioat_dma_remove_interrupts(device);
1702 err_setup_interrupts:
1703 pci_pool_destroy(device->completion_pool);
1704 err_completion_pool:
1705 pci_pool_destroy(device->dma_pool);
1706 err_dma_pool:
1707 kfree(device);
1708 err_kzalloc:
1709 dev_err(&pdev->dev,
1710 "Intel(R) I/OAT DMA Engine initialization failed\n");
1711 return NULL;
1712 }
1713
1714 void ioat_dma_remove(struct ioatdma_device *device)
1715 {
1716 struct dma_chan *chan, *_chan;
1717 struct ioat_dma_chan *ioat_chan;
1718
1719 if (device->version != IOAT_VER_3_0)
1720 cancel_delayed_work(&device->work);
1721
1722 ioat_dma_remove_interrupts(device);
1723
1724 dma_async_device_unregister(&device->common);
1725
1726 pci_pool_destroy(device->dma_pool);
1727 pci_pool_destroy(device->completion_pool);
1728
1729 iounmap(device->reg_base);
1730 pci_release_regions(device->pdev);
1731 pci_disable_device(device->pdev);
1732
1733 list_for_each_entry_safe(chan, _chan,
1734 &device->common.channels, device_node) {
1735 ioat_chan = to_ioat_chan(chan);
1736 list_del(&chan->device_node);
1737 kfree(ioat_chan);
1738 }
1739 kfree(device);
1740 }
1741
This page took 0.069226 seconds and 5 git commands to generate.