projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Merge branches 'x86/apic', 'x86/cleanups', 'x86/cpufeature', 'x86/crashdump', 'x86...
[deliverable/linux.git] / drivers / dma / ioat_dma.c
1 /*
2 * Intel I/OAT DMA Linux driver
3 * Copyright(c) 2004 - 2007 Intel Corporation.
4 *
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms and conditions of the GNU General Public License,
7 * version 2, as published by the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
12 * more details.
13 *
14 * You should have received a copy of the GNU General Public License along with
15 * this program; if not, write to the Free Software Foundation, Inc.,
16 * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
17 *
18 * The full GNU General Public License is included in this distribution in
19 * the file called "COPYING".
20 *
21 */
22
23 /*
24 * This driver supports an Intel I/OAT DMA engine, which does asynchronous
25 * copy operations.
26 */
27
28 #include <linux/init.h>
29 #include <linux/module.h>
30 #include <linux/pci.h>
31 #include <linux/interrupt.h>
32 #include <linux/dmaengine.h>
33 #include <linux/delay.h>
34 #include <linux/dma-mapping.h>
35 #include <linux/workqueue.h>
36 #include <linux/i7300_idle.h>
37 #include "ioatdma.h"
38 #include "ioatdma_registers.h"
39 #include "ioatdma_hw.h"
40
41 #define to_ioat_chan(chan) container_of(chan, struct ioat_dma_chan, common)
42 #define to_ioatdma_device(dev) container_of(dev, struct ioatdma_device, common)
43 #define to_ioat_desc(lh) container_of(lh, struct ioat_desc_sw, node)
44 #define tx_to_ioat_desc(tx) container_of(tx, struct ioat_desc_sw, async_tx)
45
46 #define chan_num(ch) ((int)((ch)->reg_base - (ch)->device->reg_base) / 0x80)
47 static int ioat_pending_level = 4;
48 module_param(ioat_pending_level, int, 0644);
49 MODULE_PARM_DESC(ioat_pending_level,
50 "high-water mark for pushing ioat descriptors (default: 4)");
51
52 #define RESET_DELAY msecs_to_jiffies(100)
53 #define WATCHDOG_DELAY round_jiffies(msecs_to_jiffies(2000))
54 static void ioat_dma_chan_reset_part2(struct work_struct *work);
55 static void ioat_dma_chan_watchdog(struct work_struct *work);
56
57 /*
58 * workaround for IOAT ver.3.0 null descriptor issue
59 * (channel returns error when size is 0)
60 */
61 #define NULL_DESC_BUFFER_SIZE 1
62
63 /* internal functions */
64 static void ioat_dma_start_null_desc(struct ioat_dma_chan *ioat_chan);
65 static void ioat_dma_memcpy_cleanup(struct ioat_dma_chan *ioat_chan);
66
67 static struct ioat_desc_sw *
68 ioat1_dma_get_next_descriptor(struct ioat_dma_chan *ioat_chan);
69 static struct ioat_desc_sw *
70 ioat2_dma_get_next_descriptor(struct ioat_dma_chan *ioat_chan);
71
72 static inline struct ioat_dma_chan *ioat_lookup_chan_by_index(
73 struct ioatdma_device *device,
74 int index)
75 {
76 return device->idx[index];
77 }
78
79 /**
80 * ioat_dma_do_interrupt - handler used for single vector interrupt mode
81 * @irq: interrupt id
82 * @data: interrupt data
83 */
84 static irqreturn_t ioat_dma_do_interrupt(int irq, void *data)
85 {
86 struct ioatdma_device *instance = data;
87 struct ioat_dma_chan *ioat_chan;
88 unsigned long attnstatus;
89 int bit;
90 u8 intrctrl;
91
92 intrctrl = readb(instance->reg_base + IOAT_INTRCTRL_OFFSET);
93
94 if (!(intrctrl & IOAT_INTRCTRL_MASTER_INT_EN))
95 return IRQ_NONE;
96
97 if (!(intrctrl & IOAT_INTRCTRL_INT_STATUS)) {
98 writeb(intrctrl, instance->reg_base + IOAT_INTRCTRL_OFFSET);
99 return IRQ_NONE;
100 }
101
102 attnstatus = readl(instance->reg_base + IOAT_ATTNSTATUS_OFFSET);
103 for_each_bit(bit, &attnstatus, BITS_PER_LONG) {
104 ioat_chan = ioat_lookup_chan_by_index(instance, bit);
105 tasklet_schedule(&ioat_chan->cleanup_task);
106 }
107
108 writeb(intrctrl, instance->reg_base + IOAT_INTRCTRL_OFFSET);
109 return IRQ_HANDLED;
110 }
111
112 /**
113 * ioat_dma_do_interrupt_msix - handler used for vector-per-channel interrupt mode
114 * @irq: interrupt id
115 * @data: interrupt data
116 */
117 static irqreturn_t ioat_dma_do_interrupt_msix(int irq, void *data)
118 {
119 struct ioat_dma_chan *ioat_chan = data;
120
121 tasklet_schedule(&ioat_chan->cleanup_task);
122
123 return IRQ_HANDLED;
124 }
125
126 static void ioat_dma_cleanup_tasklet(unsigned long data);
127
128 /**
129 * ioat_dma_enumerate_channels - find and initialize the device's channels
130 * @device: the device to be enumerated
131 */
132 static int ioat_dma_enumerate_channels(struct ioatdma_device *device)
133 {
134 u8 xfercap_scale;
135 u32 xfercap;
136 int i;
137 struct ioat_dma_chan *ioat_chan;
138
139 /*
140 * IOAT ver.3 workarounds
141 */
142 if (device->version == IOAT_VER_3_0) {
143 u32 chan_err_mask;
144 u16 dev_id;
145 u32 dmauncerrsts;
146
147 /*
148 * Write CHANERRMSK_INT with 3E07h to mask out the errors
149 * that can cause stability issues for IOAT ver.3
150 */
151 chan_err_mask = 0x3E07;
152 pci_write_config_dword(device->pdev,
153 IOAT_PCI_CHANERRMASK_INT_OFFSET,
154 chan_err_mask);
155
156 /*
157 * Clear DMAUNCERRSTS Cfg-Reg Parity Error status bit
158 * (workaround for spurious config parity error after restart)
159 */
160 pci_read_config_word(device->pdev,
161 IOAT_PCI_DEVICE_ID_OFFSET,
162 &dev_id);
163 if (dev_id == PCI_DEVICE_ID_INTEL_IOAT_TBG0) {
164 dmauncerrsts = 0x10;
165 pci_write_config_dword(device->pdev,
166 IOAT_PCI_DMAUNCERRSTS_OFFSET,
167 dmauncerrsts);
168 }
169 }
170
171 device->common.chancnt = readb(device->reg_base + IOAT_CHANCNT_OFFSET);
172 xfercap_scale = readb(device->reg_base + IOAT_XFERCAP_OFFSET);
173 xfercap = (xfercap_scale == 0 ? -1 : (1UL << xfercap_scale));
174
175 #ifdef CONFIG_I7300_IDLE_IOAT_CHANNEL
176 if (i7300_idle_platform_probe(NULL, NULL) == 0) {
177 device->common.chancnt--;
178 }
179 #endif
180 for (i = 0; i < device->common.chancnt; i++) {
181 ioat_chan = kzalloc(sizeof(*ioat_chan), GFP_KERNEL);
182 if (!ioat_chan) {
183 device->common.chancnt = i;
184 break;
185 }
186
187 ioat_chan->device = device;
188 ioat_chan->reg_base = device->reg_base + (0x80 * (i + 1));
189 ioat_chan->xfercap = xfercap;
190 ioat_chan->desccount = 0;
191 INIT_DELAYED_WORK(&ioat_chan->work, ioat_dma_chan_reset_part2);
192 if (ioat_chan->device->version != IOAT_VER_1_2) {
193 writel(IOAT_DCACTRL_CMPL_WRITE_ENABLE
194 | IOAT_DMA_DCA_ANY_CPU,
195 ioat_chan->reg_base + IOAT_DCACTRL_OFFSET);
196 }
197 spin_lock_init(&ioat_chan->cleanup_lock);
198 spin_lock_init(&ioat_chan->desc_lock);
199 INIT_LIST_HEAD(&ioat_chan->free_desc);
200 INIT_LIST_HEAD(&ioat_chan->used_desc);
201 /* This should be made common somewhere in dmaengine.c */
202 ioat_chan->common.device = &device->common;
203 list_add_tail(&ioat_chan->common.device_node,
204 &device->common.channels);
205 device->idx[i] = ioat_chan;
206 tasklet_init(&ioat_chan->cleanup_task,
207 ioat_dma_cleanup_tasklet,
208 (unsigned long) ioat_chan);
209 tasklet_disable(&ioat_chan->cleanup_task);
210 }
211 return device->common.chancnt;
212 }
213
214 /**
215 * ioat_dma_memcpy_issue_pending - push potentially unrecognized appended
216 * descriptors to hw
217 * @chan: DMA channel handle
218 */
219 static inline void __ioat1_dma_memcpy_issue_pending(
220 struct ioat_dma_chan *ioat_chan)
221 {
222 ioat_chan->pending = 0;
223 writeb(IOAT_CHANCMD_APPEND, ioat_chan->reg_base + IOAT1_CHANCMD_OFFSET);
224 }
225
226 static void ioat1_dma_memcpy_issue_pending(struct dma_chan *chan)
227 {
228 struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
229
230 if (ioat_chan->pending > 0) {
231 spin_lock_bh(&ioat_chan->desc_lock);
232 __ioat1_dma_memcpy_issue_pending(ioat_chan);
233 spin_unlock_bh(&ioat_chan->desc_lock);
234 }
235 }
236
237 static inline void __ioat2_dma_memcpy_issue_pending(
238 struct ioat_dma_chan *ioat_chan)
239 {
240 ioat_chan->pending = 0;
241 writew(ioat_chan->dmacount,
242 ioat_chan->reg_base + IOAT_CHAN_DMACOUNT_OFFSET);
243 }
244
245 static void ioat2_dma_memcpy_issue_pending(struct dma_chan *chan)
246 {
247 struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
248
249 if (ioat_chan->pending > 0) {
250 spin_lock_bh(&ioat_chan->desc_lock);
251 __ioat2_dma_memcpy_issue_pending(ioat_chan);
252 spin_unlock_bh(&ioat_chan->desc_lock);
253 }
254 }
255
256
257 /**
258 * ioat_dma_chan_reset_part2 - reinit the channel after a reset
259 */
260 static void ioat_dma_chan_reset_part2(struct work_struct *work)
261 {
262 struct ioat_dma_chan *ioat_chan =
263 container_of(work, struct ioat_dma_chan, work.work);
264 struct ioat_desc_sw *desc;
265
266 spin_lock_bh(&ioat_chan->cleanup_lock);
267 spin_lock_bh(&ioat_chan->desc_lock);
268
269 ioat_chan->completion_virt->low = 0;
270 ioat_chan->completion_virt->high = 0;
271 ioat_chan->pending = 0;
272
273 /*
274 * count the descriptors waiting, and be sure to do it
275 * right for both the CB1 line and the CB2 ring
276 */
277 ioat_chan->dmacount = 0;
278 if (ioat_chan->used_desc.prev) {
279 desc = to_ioat_desc(ioat_chan->used_desc.prev);
280 do {
281 ioat_chan->dmacount++;
282 desc = to_ioat_desc(desc->node.next);
283 } while (&desc->node != ioat_chan->used_desc.next);
284 }
285
286 /*
287 * write the new starting descriptor address
288 * this puts channel engine into ARMED state
289 */
290 desc = to_ioat_desc(ioat_chan->used_desc.prev);
291 switch (ioat_chan->device->version) {
292 case IOAT_VER_1_2:
293 writel(((u64) desc->async_tx.phys) & 0x00000000FFFFFFFF,
294 ioat_chan->reg_base + IOAT1_CHAINADDR_OFFSET_LOW);
295 writel(((u64) desc->async_tx.phys) >> 32,
296 ioat_chan->reg_base + IOAT1_CHAINADDR_OFFSET_HIGH);
297
298 writeb(IOAT_CHANCMD_START, ioat_chan->reg_base
299 + IOAT_CHANCMD_OFFSET(ioat_chan->device->version));
300 break;
301 case IOAT_VER_2_0:
302 writel(((u64) desc->async_tx.phys) & 0x00000000FFFFFFFF,
303 ioat_chan->reg_base + IOAT2_CHAINADDR_OFFSET_LOW);
304 writel(((u64) desc->async_tx.phys) >> 32,
305 ioat_chan->reg_base + IOAT2_CHAINADDR_OFFSET_HIGH);
306
307 /* tell the engine to go with what's left to be done */
308 writew(ioat_chan->dmacount,
309 ioat_chan->reg_base + IOAT_CHAN_DMACOUNT_OFFSET);
310
311 break;
312 }
313 dev_err(&ioat_chan->device->pdev->dev,
314 "chan%d reset - %d descs waiting, %d total desc\n",
315 chan_num(ioat_chan), ioat_chan->dmacount, ioat_chan->desccount);
316
317 spin_unlock_bh(&ioat_chan->desc_lock);
318 spin_unlock_bh(&ioat_chan->cleanup_lock);
319 }
320
321 /**
322 * ioat_dma_reset_channel - restart a channel
323 * @ioat_chan: IOAT DMA channel handle
324 */
325 static void ioat_dma_reset_channel(struct ioat_dma_chan *ioat_chan)
326 {
327 u32 chansts, chanerr;
328
329 if (!ioat_chan->used_desc.prev)
330 return;
331
332 chanerr = readl(ioat_chan->reg_base + IOAT_CHANERR_OFFSET);
333 chansts = (ioat_chan->completion_virt->low
334 & IOAT_CHANSTS_DMA_TRANSFER_STATUS);
335 if (chanerr) {
336 dev_err(&ioat_chan->device->pdev->dev,
337 "chan%d, CHANSTS = 0x%08x CHANERR = 0x%04x, clearing\n",
338 chan_num(ioat_chan), chansts, chanerr);
339 writel(chanerr, ioat_chan->reg_base + IOAT_CHANERR_OFFSET);
340 }
341
342 /*
343 * whack it upside the head with a reset
344 * and wait for things to settle out.
345 * force the pending count to a really big negative
346 * to make sure no one forces an issue_pending
347 * while we're waiting.
348 */
349
350 spin_lock_bh(&ioat_chan->desc_lock);
351 ioat_chan->pending = INT_MIN;
352 writeb(IOAT_CHANCMD_RESET,
353 ioat_chan->reg_base
354 + IOAT_CHANCMD_OFFSET(ioat_chan->device->version));
355 spin_unlock_bh(&ioat_chan->desc_lock);
356
357 /* schedule the 2nd half instead of sleeping a long time */
358 schedule_delayed_work(&ioat_chan->work, RESET_DELAY);
359 }
360
361 /**
362 * ioat_dma_chan_watchdog - watch for stuck channels
363 */
364 static void ioat_dma_chan_watchdog(struct work_struct *work)
365 {
366 struct ioatdma_device *device =
367 container_of(work, struct ioatdma_device, work.work);
368 struct ioat_dma_chan *ioat_chan;
369 int i;
370
371 union {
372 u64 full;
373 struct {
374 u32 low;
375 u32 high;
376 };
377 } completion_hw;
378 unsigned long compl_desc_addr_hw;
379
380 for (i = 0; i < device->common.chancnt; i++) {
381 ioat_chan = ioat_lookup_chan_by_index(device, i);
382
383 if (ioat_chan->device->version == IOAT_VER_1_2
384 /* have we started processing anything yet */
385 && ioat_chan->last_completion
386 /* have we completed any since last watchdog cycle? */
387 && (ioat_chan->last_completion ==
388 ioat_chan->watchdog_completion)
389 /* has TCP stuck on one cookie since last watchdog? */
390 && (ioat_chan->watchdog_tcp_cookie ==
391 ioat_chan->watchdog_last_tcp_cookie)
392 && (ioat_chan->watchdog_tcp_cookie !=
393 ioat_chan->completed_cookie)
394 /* is there something in the chain to be processed? */
395 /* CB1 chain always has at least the last one processed */
396 && (ioat_chan->used_desc.prev != ioat_chan->used_desc.next)
397 && ioat_chan->pending == 0) {
398
399 /*
400 * check CHANSTS register for completed
401 * descriptor address.
402 * if it is different than completion writeback,
403 * it is not zero
404 * and it has changed since the last watchdog
405 * we can assume that channel
406 * is still working correctly
407 * and the problem is in completion writeback.
408 * update completion writeback
409 * with actual CHANSTS value
410 * else
411 * try resetting the channel
412 */
413
414 completion_hw.low = readl(ioat_chan->reg_base +
415 IOAT_CHANSTS_OFFSET_LOW(ioat_chan->device->version));
416 completion_hw.high = readl(ioat_chan->reg_base +
417 IOAT_CHANSTS_OFFSET_HIGH(ioat_chan->device->version));
418 #if (BITS_PER_LONG == 64)
419 compl_desc_addr_hw =
420 completion_hw.full
421 & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_ADDR;
422 #else
423 compl_desc_addr_hw =
424 completion_hw.low & IOAT_LOW_COMPLETION_MASK;
425 #endif
426
427 if ((compl_desc_addr_hw != 0)
428 && (compl_desc_addr_hw != ioat_chan->watchdog_completion)
429 && (compl_desc_addr_hw != ioat_chan->last_compl_desc_addr_hw)) {
430 ioat_chan->last_compl_desc_addr_hw = compl_desc_addr_hw;
431 ioat_chan->completion_virt->low = completion_hw.low;
432 ioat_chan->completion_virt->high = completion_hw.high;
433 } else {
434 ioat_dma_reset_channel(ioat_chan);
435 ioat_chan->watchdog_completion = 0;
436 ioat_chan->last_compl_desc_addr_hw = 0;
437 }
438
439 /*
440 * for version 2.0 if there are descriptors yet to be processed
441 * and the last completed hasn't changed since the last watchdog
442 * if they haven't hit the pending level
443 * issue the pending to push them through
444 * else
445 * try resetting the channel
446 */
447 } else if (ioat_chan->device->version == IOAT_VER_2_0
448 && ioat_chan->used_desc.prev
449 && ioat_chan->last_completion
450 && ioat_chan->last_completion == ioat_chan->watchdog_completion) {
451
452 if (ioat_chan->pending < ioat_pending_level)
453 ioat2_dma_memcpy_issue_pending(&ioat_chan->common);
454 else {
455 ioat_dma_reset_channel(ioat_chan);
456 ioat_chan->watchdog_completion = 0;
457 }
458 } else {
459 ioat_chan->last_compl_desc_addr_hw = 0;
460 ioat_chan->watchdog_completion
461 = ioat_chan->last_completion;
462 }
463
464 ioat_chan->watchdog_last_tcp_cookie =
465 ioat_chan->watchdog_tcp_cookie;
466 }
467
468 schedule_delayed_work(&device->work, WATCHDOG_DELAY);
469 }
470
471 static dma_cookie_t ioat1_tx_submit(struct dma_async_tx_descriptor *tx)
472 {
473 struct ioat_dma_chan *ioat_chan = to_ioat_chan(tx->chan);
474 struct ioat_desc_sw *first = tx_to_ioat_desc(tx);
475 struct ioat_desc_sw *prev, *new;
476 struct ioat_dma_descriptor *hw;
477 dma_cookie_t cookie;
478 LIST_HEAD(new_chain);
479 u32 copy;
480 size_t len;
481 dma_addr_t src, dst;
482 unsigned long orig_flags;
483 unsigned int desc_count = 0;
484
485 /* src and dest and len are stored in the initial descriptor */
486 len = first->len;
487 src = first->src;
488 dst = first->dst;
489 orig_flags = first->async_tx.flags;
490 new = first;
491
492 spin_lock_bh(&ioat_chan->desc_lock);
493 prev = to_ioat_desc(ioat_chan->used_desc.prev);
494 prefetch(prev->hw);
495 do {
496 copy = min_t(size_t, len, ioat_chan->xfercap);
497
498 async_tx_ack(&new->async_tx);
499
500 hw = new->hw;
501 hw->size = copy;
502 hw->ctl = 0;
503 hw->src_addr = src;
504 hw->dst_addr = dst;
505 hw->next = 0;
506
507 /* chain together the physical address list for the HW */
508 wmb();
509 prev->hw->next = (u64) new->async_tx.phys;
510
511 len -= copy;
512 dst += copy;
513 src += copy;
514
515 list_add_tail(&new->node, &new_chain);
516 desc_count++;
517 prev = new;
518 } while (len && (new = ioat1_dma_get_next_descriptor(ioat_chan)));
519
520 if (!new) {
521 dev_err(&ioat_chan->device->pdev->dev,
522 "tx submit failed\n");
523 spin_unlock_bh(&ioat_chan->desc_lock);
524 return -ENOMEM;
525 }
526
527 hw->ctl = IOAT_DMA_DESCRIPTOR_CTL_CP_STS;
528 if (first->async_tx.callback) {
529 hw->ctl |= IOAT_DMA_DESCRIPTOR_CTL_INT_GN;
530 if (first != new) {
531 /* move callback into to last desc */
532 new->async_tx.callback = first->async_tx.callback;
533 new->async_tx.callback_param
534 = first->async_tx.callback_param;
535 first->async_tx.callback = NULL;
536 first->async_tx.callback_param = NULL;
537 }
538 }
539
540 new->tx_cnt = desc_count;
541 new->async_tx.flags = orig_flags; /* client is in control of this ack */
542
543 /* store the original values for use in later cleanup */
544 if (new != first) {
545 new->src = first->src;
546 new->dst = first->dst;
547 new->len = first->len;
548 }
549
550 /* cookie incr and addition to used_list must be atomic */
551 cookie = ioat_chan->common.cookie;
552 cookie++;
553 if (cookie < 0)
554 cookie = 1;
555 ioat_chan->common.cookie = new->async_tx.cookie = cookie;
556
557 /* write address into NextDescriptor field of last desc in chain */
558 to_ioat_desc(ioat_chan->used_desc.prev)->hw->next =
559 first->async_tx.phys;
560 list_splice_tail(&new_chain, &ioat_chan->used_desc);
561
562 ioat_chan->dmacount += desc_count;
563 ioat_chan->pending += desc_count;
564 if (ioat_chan->pending >= ioat_pending_level)
565 __ioat1_dma_memcpy_issue_pending(ioat_chan);
566 spin_unlock_bh(&ioat_chan->desc_lock);
567
568 return cookie;
569 }
570
571 static dma_cookie_t ioat2_tx_submit(struct dma_async_tx_descriptor *tx)
572 {
573 struct ioat_dma_chan *ioat_chan = to_ioat_chan(tx->chan);
574 struct ioat_desc_sw *first = tx_to_ioat_desc(tx);
575 struct ioat_desc_sw *new;
576 struct ioat_dma_descriptor *hw;
577 dma_cookie_t cookie;
578 u32 copy;
579 size_t len;
580 dma_addr_t src, dst;
581 unsigned long orig_flags;
582 unsigned int desc_count = 0;
583
584 /* src and dest and len are stored in the initial descriptor */
585 len = first->len;
586 src = first->src;
587 dst = first->dst;
588 orig_flags = first->async_tx.flags;
589 new = first;
590
591 /*
592 * ioat_chan->desc_lock is still in force in version 2 path
593 * it gets unlocked at end of this function
594 */
595 do {
596 copy = min_t(size_t, len, ioat_chan->xfercap);
597
598 async_tx_ack(&new->async_tx);
599
600 hw = new->hw;
601 hw->size = copy;
602 hw->ctl = 0;
603 hw->src_addr = src;
604 hw->dst_addr = dst;
605
606 len -= copy;
607 dst += copy;
608 src += copy;
609 desc_count++;
610 } while (len && (new = ioat2_dma_get_next_descriptor(ioat_chan)));
611
612 if (!new) {
613 dev_err(&ioat_chan->device->pdev->dev,
614 "tx submit failed\n");
615 spin_unlock_bh(&ioat_chan->desc_lock);
616 return -ENOMEM;
617 }
618
619 hw->ctl |= IOAT_DMA_DESCRIPTOR_CTL_CP_STS;
620 if (first->async_tx.callback) {
621 hw->ctl |= IOAT_DMA_DESCRIPTOR_CTL_INT_GN;
622 if (first != new) {
623 /* move callback into to last desc */
624 new->async_tx.callback = first->async_tx.callback;
625 new->async_tx.callback_param
626 = first->async_tx.callback_param;
627 first->async_tx.callback = NULL;
628 first->async_tx.callback_param = NULL;
629 }
630 }
631
632 new->tx_cnt = desc_count;
633 new->async_tx.flags = orig_flags; /* client is in control of this ack */
634
635 /* store the original values for use in later cleanup */
636 if (new != first) {
637 new->src = first->src;
638 new->dst = first->dst;
639 new->len = first->len;
640 }
641
642 /* cookie incr and addition to used_list must be atomic */
643 cookie = ioat_chan->common.cookie;
644 cookie++;
645 if (cookie < 0)
646 cookie = 1;
647 ioat_chan->common.cookie = new->async_tx.cookie = cookie;
648
649 ioat_chan->dmacount += desc_count;
650 ioat_chan->pending += desc_count;
651 if (ioat_chan->pending >= ioat_pending_level)
652 __ioat2_dma_memcpy_issue_pending(ioat_chan);
653 spin_unlock_bh(&ioat_chan->desc_lock);
654
655 return cookie;
656 }
657
658 /**
659 * ioat_dma_alloc_descriptor - allocate and return a sw and hw descriptor pair
660 * @ioat_chan: the channel supplying the memory pool for the descriptors
661 * @flags: allocation flags
662 */
663 static struct ioat_desc_sw *ioat_dma_alloc_descriptor(
664 struct ioat_dma_chan *ioat_chan,
665 gfp_t flags)
666 {
667 struct ioat_dma_descriptor *desc;
668 struct ioat_desc_sw *desc_sw;
669 struct ioatdma_device *ioatdma_device;
670 dma_addr_t phys;
671
672 ioatdma_device = to_ioatdma_device(ioat_chan->common.device);
673 desc = pci_pool_alloc(ioatdma_device->dma_pool, flags, &phys);
674 if (unlikely(!desc))
675 return NULL;
676
677 desc_sw = kzalloc(sizeof(*desc_sw), flags);
678 if (unlikely(!desc_sw)) {
679 pci_pool_free(ioatdma_device->dma_pool, desc, phys);
680 return NULL;
681 }
682
683 memset(desc, 0, sizeof(*desc));
684 dma_async_tx_descriptor_init(&desc_sw->async_tx, &ioat_chan->common);
685 switch (ioat_chan->device->version) {
686 case IOAT_VER_1_2:
687 desc_sw->async_tx.tx_submit = ioat1_tx_submit;
688 break;
689 case IOAT_VER_2_0:
690 case IOAT_VER_3_0:
691 desc_sw->async_tx.tx_submit = ioat2_tx_submit;
692 break;
693 }
694 INIT_LIST_HEAD(&desc_sw->async_tx.tx_list);
695
696 desc_sw->hw = desc;
697 desc_sw->async_tx.phys = phys;
698
699 return desc_sw;
700 }
701
702 static int ioat_initial_desc_count = 256;
703 module_param(ioat_initial_desc_count, int, 0644);
704 MODULE_PARM_DESC(ioat_initial_desc_count,
705 "initial descriptors per channel (default: 256)");
706
707 /**
708 * ioat2_dma_massage_chan_desc - link the descriptors into a circle
709 * @ioat_chan: the channel to be massaged
710 */
711 static void ioat2_dma_massage_chan_desc(struct ioat_dma_chan *ioat_chan)
712 {
713 struct ioat_desc_sw *desc, *_desc;
714
715 /* setup used_desc */
716 ioat_chan->used_desc.next = ioat_chan->free_desc.next;
717 ioat_chan->used_desc.prev = NULL;
718
719 /* pull free_desc out of the circle so that every node is a hw
720 * descriptor, but leave it pointing to the list
721 */
722 ioat_chan->free_desc.prev->next = ioat_chan->free_desc.next;
723 ioat_chan->free_desc.next->prev = ioat_chan->free_desc.prev;
724
725 /* circle link the hw descriptors */
726 desc = to_ioat_desc(ioat_chan->free_desc.next);
727 desc->hw->next = to_ioat_desc(desc->node.next)->async_tx.phys;
728 list_for_each_entry_safe(desc, _desc, ioat_chan->free_desc.next, node) {
729 desc->hw->next = to_ioat_desc(desc->node.next)->async_tx.phys;
730 }
731 }
732
733 /**
734 * ioat_dma_alloc_chan_resources - returns the number of allocated descriptors
735 * @chan: the channel to be filled out
736 */
737 static int ioat_dma_alloc_chan_resources(struct dma_chan *chan,
738 struct dma_client *client)
739 {
740 struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
741 struct ioat_desc_sw *desc;
742 u16 chanctrl;
743 u32 chanerr;
744 int i;
745 LIST_HEAD(tmp_list);
746
747 /* have we already been set up? */
748 if (!list_empty(&ioat_chan->free_desc))
749 return ioat_chan->desccount;
750
751 /* Setup register to interrupt and write completion status on error */
752 chanctrl = IOAT_CHANCTRL_ERR_INT_EN |
753 IOAT_CHANCTRL_ANY_ERR_ABORT_EN |
754 IOAT_CHANCTRL_ERR_COMPLETION_EN;
755 writew(chanctrl, ioat_chan->reg_base + IOAT_CHANCTRL_OFFSET);
756
757 chanerr = readl(ioat_chan->reg_base + IOAT_CHANERR_OFFSET);
758 if (chanerr) {
759 dev_err(&ioat_chan->device->pdev->dev,
760 "CHANERR = %x, clearing\n", chanerr);
761 writel(chanerr, ioat_chan->reg_base + IOAT_CHANERR_OFFSET);
762 }
763
764 /* Allocate descriptors */
765 for (i = 0; i < ioat_initial_desc_count; i++) {
766 desc = ioat_dma_alloc_descriptor(ioat_chan, GFP_KERNEL);
767 if (!desc) {
768 dev_err(&ioat_chan->device->pdev->dev,
769 "Only %d initial descriptors\n", i);
770 break;
771 }
772 list_add_tail(&desc->node, &tmp_list);
773 }
774 spin_lock_bh(&ioat_chan->desc_lock);
775 ioat_chan->desccount = i;
776 list_splice(&tmp_list, &ioat_chan->free_desc);
777 if (ioat_chan->device->version != IOAT_VER_1_2)
778 ioat2_dma_massage_chan_desc(ioat_chan);
779 spin_unlock_bh(&ioat_chan->desc_lock);
780
781 /* allocate a completion writeback area */
782 /* doing 2 32bit writes to mmio since 1 64b write doesn't work */
783 ioat_chan->completion_virt =
784 pci_pool_alloc(ioat_chan->device->completion_pool,
785 GFP_KERNEL,
786 &ioat_chan->completion_addr);
787 memset(ioat_chan->completion_virt, 0,
788 sizeof(*ioat_chan->completion_virt));
789 writel(((u64) ioat_chan->completion_addr) & 0x00000000FFFFFFFF,
790 ioat_chan->reg_base + IOAT_CHANCMP_OFFSET_LOW);
791 writel(((u64) ioat_chan->completion_addr) >> 32,
792 ioat_chan->reg_base + IOAT_CHANCMP_OFFSET_HIGH);
793
794 tasklet_enable(&ioat_chan->cleanup_task);
795 ioat_dma_start_null_desc(ioat_chan); /* give chain to dma device */
796 return ioat_chan->desccount;
797 }
798
799 /**
800 * ioat_dma_free_chan_resources - release all the descriptors
801 * @chan: the channel to be cleaned
802 */
803 static void ioat_dma_free_chan_resources(struct dma_chan *chan)
804 {
805 struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
806 struct ioatdma_device *ioatdma_device = to_ioatdma_device(chan->device);
807 struct ioat_desc_sw *desc, *_desc;
808 int in_use_descs = 0;
809
810 /* 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 /*
1067 * yes we are unmapping both _page and _single
1068 * alloc'd regions with unmap_page. Is this
1069 * *really* that bad?
1070 */
1071 if (!(desc->async_tx.flags & DMA_COMPL_SKIP_DEST_UNMAP))
1072 pci_unmap_page(ioat_chan->device->pdev,
1073 pci_unmap_addr(desc, dst),
1074 pci_unmap_len(desc, len),
1075 PCI_DMA_FROMDEVICE);
1076
1077 if (!(desc->async_tx.flags & DMA_COMPL_SKIP_SRC_UNMAP))
1078 pci_unmap_page(ioat_chan->device->pdev,
1079 pci_unmap_addr(desc, src),
1080 pci_unmap_len(desc, len),
1081 PCI_DMA_TODEVICE);
1082 }
1083
1084 /**
1085 * ioat_dma_memcpy_cleanup - cleanup up finished descriptors
1086 * @chan: ioat channel to be cleaned up
1087 */
1088 static void ioat_dma_memcpy_cleanup(struct ioat_dma_chan *ioat_chan)
1089 {
1090 unsigned long phys_complete;
1091 struct ioat_desc_sw *desc, *_desc;
1092 dma_cookie_t cookie = 0;
1093 unsigned long desc_phys;
1094 struct ioat_desc_sw *latest_desc;
1095
1096 prefetch(ioat_chan->completion_virt);
1097
1098 if (!spin_trylock_bh(&ioat_chan->cleanup_lock))
1099 return;
1100
1101 /* The completion writeback can happen at any time,
1102 so reads by the driver need to be atomic operations
1103 The descriptor physical addresses are limited to 32-bits
1104 when the CPU can only do a 32-bit mov */
1105
1106 #if (BITS_PER_LONG == 64)
1107 phys_complete =
1108 ioat_chan->completion_virt->full
1109 & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_ADDR;
1110 #else
1111 phys_complete =
1112 ioat_chan->completion_virt->low & IOAT_LOW_COMPLETION_MASK;
1113 #endif
1114
1115 if ((ioat_chan->completion_virt->full
1116 & IOAT_CHANSTS_DMA_TRANSFER_STATUS) ==
1117 IOAT_CHANSTS_DMA_TRANSFER_STATUS_HALTED) {
1118 dev_err(&ioat_chan->device->pdev->dev,
1119 "Channel halted, chanerr = %x\n",
1120 readl(ioat_chan->reg_base + IOAT_CHANERR_OFFSET));
1121
1122 /* TODO do something to salvage the situation */
1123 }
1124
1125 if (phys_complete == ioat_chan->last_completion) {
1126 spin_unlock_bh(&ioat_chan->cleanup_lock);
1127 /*
1128 * perhaps we're stuck so hard that the watchdog can't go off?
1129 * try to catch it after 2 seconds
1130 */
1131 if (ioat_chan->device->version != IOAT_VER_3_0) {
1132 if (time_after(jiffies,
1133 ioat_chan->last_completion_time + HZ*WATCHDOG_DELAY)) {
1134 ioat_dma_chan_watchdog(&(ioat_chan->device->work.work));
1135 ioat_chan->last_completion_time = jiffies;
1136 }
1137 }
1138 return;
1139 }
1140 ioat_chan->last_completion_time = jiffies;
1141
1142 cookie = 0;
1143 if (!spin_trylock_bh(&ioat_chan->desc_lock)) {
1144 spin_unlock_bh(&ioat_chan->cleanup_lock);
1145 return;
1146 }
1147
1148 switch (ioat_chan->device->version) {
1149 case IOAT_VER_1_2:
1150 list_for_each_entry_safe(desc, _desc,
1151 &ioat_chan->used_desc, node) {
1152
1153 /*
1154 * Incoming DMA requests may use multiple descriptors,
1155 * due to exceeding xfercap, perhaps. If so, only the
1156 * last one will have a cookie, and require unmapping.
1157 */
1158 if (desc->async_tx.cookie) {
1159 cookie = desc->async_tx.cookie;
1160 ioat_dma_unmap(ioat_chan, desc);
1161 if (desc->async_tx.callback) {
1162 desc->async_tx.callback(desc->async_tx.callback_param);
1163 desc->async_tx.callback = NULL;
1164 }
1165 }
1166
1167 if (desc->async_tx.phys != phys_complete) {
1168 /*
1169 * a completed entry, but not the last, so clean
1170 * up if the client is done with the descriptor
1171 */
1172 if (async_tx_test_ack(&desc->async_tx)) {
1173 list_del(&desc->node);
1174 list_add_tail(&desc->node,
1175 &ioat_chan->free_desc);
1176 } else
1177 desc->async_tx.cookie = 0;
1178 } else {
1179 /*
1180 * last used desc. Do not remove, so we can
1181 * append from it, but don't look at it next
1182 * time, either
1183 */
1184 desc->async_tx.cookie = 0;
1185
1186 /* TODO check status bits? */
1187 break;
1188 }
1189 }
1190 break;
1191 case IOAT_VER_2_0:
1192 case IOAT_VER_3_0:
1193 /* has some other thread has already cleaned up? */
1194 if (ioat_chan->used_desc.prev == NULL)
1195 break;
1196
1197 /* work backwards to find latest finished desc */
1198 desc = to_ioat_desc(ioat_chan->used_desc.next);
1199 latest_desc = NULL;
1200 do {
1201 desc = to_ioat_desc(desc->node.prev);
1202 desc_phys = (unsigned long)desc->async_tx.phys
1203 & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_ADDR;
1204 if (desc_phys == phys_complete) {
1205 latest_desc = desc;
1206 break;
1207 }
1208 } while (&desc->node != ioat_chan->used_desc.prev);
1209
1210 if (latest_desc != NULL) {
1211
1212 /* work forwards to clear finished descriptors */
1213 for (desc = to_ioat_desc(ioat_chan->used_desc.prev);
1214 &desc->node != latest_desc->node.next &&
1215 &desc->node != ioat_chan->used_desc.next;
1216 desc = to_ioat_desc(desc->node.next)) {
1217 if (desc->async_tx.cookie) {
1218 cookie = desc->async_tx.cookie;
1219 desc->async_tx.cookie = 0;
1220 ioat_dma_unmap(ioat_chan, desc);
1221 if (desc->async_tx.callback) {
1222 desc->async_tx.callback(desc->async_tx.callback_param);
1223 desc->async_tx.callback = NULL;
1224 }
1225 }
1226 }
1227
1228 /* move used.prev up beyond those that are finished */
1229 if (&desc->node == ioat_chan->used_desc.next)
1230 ioat_chan->used_desc.prev = NULL;
1231 else
1232 ioat_chan->used_desc.prev = &desc->node;
1233 }
1234 break;
1235 }
1236
1237 spin_unlock_bh(&ioat_chan->desc_lock);
1238
1239 ioat_chan->last_completion = phys_complete;
1240 if (cookie != 0)
1241 ioat_chan->completed_cookie = cookie;
1242
1243 spin_unlock_bh(&ioat_chan->cleanup_lock);
1244 }
1245
1246 /**
1247 * ioat_dma_is_complete - poll the status of a IOAT DMA transaction
1248 * @chan: IOAT DMA channel handle
1249 * @cookie: DMA transaction identifier
1250 * @done: if not %NULL, updated with last completed transaction
1251 * @used: if not %NULL, updated with last used transaction
1252 */
1253 static enum dma_status ioat_dma_is_complete(struct dma_chan *chan,
1254 dma_cookie_t cookie,
1255 dma_cookie_t *done,
1256 dma_cookie_t *used)
1257 {
1258 struct ioat_dma_chan *ioat_chan = to_ioat_chan(chan);
1259 dma_cookie_t last_used;
1260 dma_cookie_t last_complete;
1261 enum dma_status ret;
1262
1263 last_used = chan->cookie;
1264 last_complete = ioat_chan->completed_cookie;
1265 ioat_chan->watchdog_tcp_cookie = cookie;
1266
1267 if (done)
1268 *done = last_complete;
1269 if (used)
1270 *used = last_used;
1271
1272 ret = dma_async_is_complete(cookie, last_complete, last_used);
1273 if (ret == DMA_SUCCESS)
1274 return ret;
1275
1276 ioat_dma_memcpy_cleanup(ioat_chan);
1277
1278 last_used = chan->cookie;
1279 last_complete = ioat_chan->completed_cookie;
1280
1281 if (done)
1282 *done = last_complete;
1283 if (used)
1284 *used = last_used;
1285
1286 return dma_async_is_complete(cookie, last_complete, last_used);
1287 }
1288
1289 static void ioat_dma_start_null_desc(struct ioat_dma_chan *ioat_chan)
1290 {
1291 struct ioat_desc_sw *desc;
1292
1293 spin_lock_bh(&ioat_chan->desc_lock);
1294
1295 desc = ioat_dma_get_next_descriptor(ioat_chan);
1296
1297 if (!desc) {
1298 dev_err(&ioat_chan->device->pdev->dev,
1299 "Unable to start null desc - get next desc failed\n");
1300 spin_unlock_bh(&ioat_chan->desc_lock);
1301 return;
1302 }
1303
1304 desc->hw->ctl = IOAT_DMA_DESCRIPTOR_NUL
1305 | IOAT_DMA_DESCRIPTOR_CTL_INT_GN
1306 | IOAT_DMA_DESCRIPTOR_CTL_CP_STS;
1307 /* set size to non-zero value (channel returns error when size is 0) */
1308 desc->hw->size = NULL_DESC_BUFFER_SIZE;
1309 desc->hw->src_addr = 0;
1310 desc->hw->dst_addr = 0;
1311 async_tx_ack(&desc->async_tx);
1312 switch (ioat_chan->device->version) {
1313 case IOAT_VER_1_2:
1314 desc->hw->next = 0;
1315 list_add_tail(&desc->node, &ioat_chan->used_desc);
1316
1317 writel(((u64) desc->async_tx.phys) & 0x00000000FFFFFFFF,
1318 ioat_chan->reg_base + IOAT1_CHAINADDR_OFFSET_LOW);
1319 writel(((u64) desc->async_tx.phys) >> 32,
1320 ioat_chan->reg_base + IOAT1_CHAINADDR_OFFSET_HIGH);
1321
1322 writeb(IOAT_CHANCMD_START, ioat_chan->reg_base
1323 + IOAT_CHANCMD_OFFSET(ioat_chan->device->version));
1324 break;
1325 case IOAT_VER_2_0:
1326 case IOAT_VER_3_0:
1327 writel(((u64) desc->async_tx.phys) & 0x00000000FFFFFFFF,
1328 ioat_chan->reg_base + IOAT2_CHAINADDR_OFFSET_LOW);
1329 writel(((u64) desc->async_tx.phys) >> 32,
1330 ioat_chan->reg_base + IOAT2_CHAINADDR_OFFSET_HIGH);
1331
1332 ioat_chan->dmacount++;
1333 __ioat2_dma_memcpy_issue_pending(ioat_chan);
1334 break;
1335 }
1336 spin_unlock_bh(&ioat_chan->desc_lock);
1337 }
1338
1339 /*
1340 * Perform a IOAT transaction to verify the HW works.
1341 */
1342 #define IOAT_TEST_SIZE 2000
1343
1344 DECLARE_COMPLETION(test_completion);
1345 static void ioat_dma_test_callback(void *dma_async_param)
1346 {
1347 printk(KERN_ERR "ioatdma: ioat_dma_test_callback(%p)\n",
1348 dma_async_param);
1349 complete(&test_completion);
1350 }
1351
1352 /**
1353 * ioat_dma_self_test - Perform a IOAT transaction to verify the HW works.
1354 * @device: device to be tested
1355 */
1356 static int ioat_dma_self_test(struct ioatdma_device *device)
1357 {
1358 int i;
1359 u8 *src;
1360 u8 *dest;
1361 struct dma_chan *dma_chan;
1362 struct dma_async_tx_descriptor *tx;
1363 dma_addr_t dma_dest, dma_src;
1364 dma_cookie_t cookie;
1365 int err = 0;
1366
1367 src = kzalloc(sizeof(u8) * IOAT_TEST_SIZE, GFP_KERNEL);
1368 if (!src)
1369 return -ENOMEM;
1370 dest = kzalloc(sizeof(u8) * IOAT_TEST_SIZE, GFP_KERNEL);
1371 if (!dest) {
1372 kfree(src);
1373 return -ENOMEM;
1374 }
1375
1376 /* Fill in src buffer */
1377 for (i = 0; i < IOAT_TEST_SIZE; i++)
1378 src[i] = (u8)i;
1379
1380 /* Start copy, using first DMA channel */
1381 dma_chan = container_of(device->common.channels.next,
1382 struct dma_chan,
1383 device_node);
1384 if (device->common.device_alloc_chan_resources(dma_chan, NULL) < 1) {
1385 dev_err(&device->pdev->dev,
1386 "selftest cannot allocate chan resource\n");
1387 err = -ENODEV;
1388 goto out;
1389 }
1390
1391 dma_src = dma_map_single(dma_chan->device->dev, src, IOAT_TEST_SIZE,
1392 DMA_TO_DEVICE);
1393 dma_dest = dma_map_single(dma_chan->device->dev, dest, IOAT_TEST_SIZE,
1394 DMA_FROM_DEVICE);
1395 tx = device->common.device_prep_dma_memcpy(dma_chan, dma_dest, dma_src,
1396 IOAT_TEST_SIZE, 0);
1397 if (!tx) {
1398 dev_err(&device->pdev->dev,
1399 "Self-test prep failed, disabling\n");
1400 err = -ENODEV;
1401 goto free_resources;
1402 }
1403
1404 async_tx_ack(tx);
1405 tx->callback = ioat_dma_test_callback;
1406 tx->callback_param = (void *)0x8086;
1407 cookie = tx->tx_submit(tx);
1408 if (cookie < 0) {
1409 dev_err(&device->pdev->dev,
1410 "Self-test setup failed, disabling\n");
1411 err = -ENODEV;
1412 goto free_resources;
1413 }
1414 device->common.device_issue_pending(dma_chan);
1415
1416 wait_for_completion_timeout(&test_completion, msecs_to_jiffies(3000));
1417
1418 if (device->common.device_is_tx_complete(dma_chan, cookie, NULL, NULL)
1419 != DMA_SUCCESS) {
1420 dev_err(&device->pdev->dev,
1421 "Self-test copy timed out, disabling\n");
1422 err = -ENODEV;
1423 goto free_resources;
1424 }
1425 if (memcmp(src, dest, IOAT_TEST_SIZE)) {
1426 dev_err(&device->pdev->dev,
1427 "Self-test copy failed compare, disabling\n");
1428 err = -ENODEV;
1429 goto free_resources;
1430 }
1431
1432 free_resources:
1433 device->common.device_free_chan_resources(dma_chan);
1434 out:
1435 kfree(src);
1436 kfree(dest);
1437 return err;
1438 }
1439
1440 static char ioat_interrupt_style[32] = "msix";
1441 module_param_string(ioat_interrupt_style, ioat_interrupt_style,
1442 sizeof(ioat_interrupt_style), 0644);
1443 MODULE_PARM_DESC(ioat_interrupt_style,
1444 "set ioat interrupt style: msix (default), "
1445 "msix-single-vector, msi, intx)");
1446
1447 /**
1448 * ioat_dma_setup_interrupts - setup interrupt handler
1449 * @device: ioat device
1450 */
1451 static int ioat_dma_setup_interrupts(struct ioatdma_device *device)
1452 {
1453 struct ioat_dma_chan *ioat_chan;
1454 int err, i, j, msixcnt;
1455 u8 intrctrl = 0;
1456
1457 if (!strcmp(ioat_interrupt_style, "msix"))
1458 goto msix;
1459 if (!strcmp(ioat_interrupt_style, "msix-single-vector"))
1460 goto msix_single_vector;
1461 if (!strcmp(ioat_interrupt_style, "msi"))
1462 goto msi;
1463 if (!strcmp(ioat_interrupt_style, "intx"))
1464 goto intx;
1465 dev_err(&device->pdev->dev, "invalid ioat_interrupt_style %s\n",
1466 ioat_interrupt_style);
1467 goto err_no_irq;
1468
1469 msix:
1470 /* The number of MSI-X vectors should equal the number of channels */
1471 msixcnt = device->common.chancnt;
1472 for (i = 0; i < msixcnt; i++)
1473 device->msix_entries[i].entry = i;
1474
1475 err = pci_enable_msix(device->pdev, device->msix_entries, msixcnt);
1476 if (err < 0)
1477 goto msi;
1478 if (err > 0)
1479 goto msix_single_vector;
1480
1481 for (i = 0; i < msixcnt; i++) {
1482 ioat_chan = ioat_lookup_chan_by_index(device, i);
1483 err = request_irq(device->msix_entries[i].vector,
1484 ioat_dma_do_interrupt_msix,
1485 0, "ioat-msix", ioat_chan);
1486 if (err) {
1487 for (j = 0; j < i; j++) {
1488 ioat_chan =
1489 ioat_lookup_chan_by_index(device, j);
1490 free_irq(device->msix_entries[j].vector,
1491 ioat_chan);
1492 }
1493 goto msix_single_vector;
1494 }
1495 }
1496 intrctrl |= IOAT_INTRCTRL_MSIX_VECTOR_CONTROL;
1497 device->irq_mode = msix_multi_vector;
1498 goto done;
1499
1500 msix_single_vector:
1501 device->msix_entries[0].entry = 0;
1502 err = pci_enable_msix(device->pdev, device->msix_entries, 1);
1503 if (err)
1504 goto msi;
1505
1506 err = request_irq(device->msix_entries[0].vector, ioat_dma_do_interrupt,
1507 0, "ioat-msix", device);
1508 if (err) {
1509 pci_disable_msix(device->pdev);
1510 goto msi;
1511 }
1512 device->irq_mode = msix_single_vector;
1513 goto done;
1514
1515 msi:
1516 err = pci_enable_msi(device->pdev);
1517 if (err)
1518 goto intx;
1519
1520 err = request_irq(device->pdev->irq, ioat_dma_do_interrupt,
1521 0, "ioat-msi", device);
1522 if (err) {
1523 pci_disable_msi(device->pdev);
1524 goto intx;
1525 }
1526 /*
1527 * CB 1.2 devices need a bit set in configuration space to enable MSI
1528 */
1529 if (device->version == IOAT_VER_1_2) {
1530 u32 dmactrl;
1531 pci_read_config_dword(device->pdev,
1532 IOAT_PCI_DMACTRL_OFFSET, &dmactrl);
1533 dmactrl |= IOAT_PCI_DMACTRL_MSI_EN;
1534 pci_write_config_dword(device->pdev,
1535 IOAT_PCI_DMACTRL_OFFSET, dmactrl);
1536 }
1537 device->irq_mode = msi;
1538 goto done;
1539
1540 intx:
1541 err = request_irq(device->pdev->irq, ioat_dma_do_interrupt,
1542 IRQF_SHARED, "ioat-intx", device);
1543 if (err)
1544 goto err_no_irq;
1545 device->irq_mode = intx;
1546
1547 done:
1548 intrctrl |= IOAT_INTRCTRL_MASTER_INT_EN;
1549 writeb(intrctrl, device->reg_base + IOAT_INTRCTRL_OFFSET);
1550 return 0;
1551
1552 err_no_irq:
1553 /* Disable all interrupt generation */
1554 writeb(0, device->reg_base + IOAT_INTRCTRL_OFFSET);
1555 dev_err(&device->pdev->dev, "no usable interrupts\n");
1556 device->irq_mode = none;
1557 return -1;
1558 }
1559
1560 /**
1561 * ioat_dma_remove_interrupts - remove whatever interrupts were set
1562 * @device: ioat device
1563 */
1564 static void ioat_dma_remove_interrupts(struct ioatdma_device *device)
1565 {
1566 struct ioat_dma_chan *ioat_chan;
1567 int i;
1568
1569 /* Disable all interrupt generation */
1570 writeb(0, device->reg_base + IOAT_INTRCTRL_OFFSET);
1571
1572 switch (device->irq_mode) {
1573 case msix_multi_vector:
1574 for (i = 0; i < device->common.chancnt; i++) {
1575 ioat_chan = ioat_lookup_chan_by_index(device, i);
1576 free_irq(device->msix_entries[i].vector, ioat_chan);
1577 }
1578 pci_disable_msix(device->pdev);
1579 break;
1580 case msix_single_vector:
1581 free_irq(device->msix_entries[0].vector, device);
1582 pci_disable_msix(device->pdev);
1583 break;
1584 case msi:
1585 free_irq(device->pdev->irq, device);
1586 pci_disable_msi(device->pdev);
1587 break;
1588 case intx:
1589 free_irq(device->pdev->irq, device);
1590 break;
1591 case none:
1592 dev_warn(&device->pdev->dev,
1593 "call to %s without interrupts setup\n", __func__);
1594 }
1595 device->irq_mode = none;
1596 }
1597
1598 struct ioatdma_device *ioat_dma_probe(struct pci_dev *pdev,
1599 void __iomem *iobase)
1600 {
1601 int err;
1602 struct ioatdma_device *device;
1603
1604 device = kzalloc(sizeof(*device), GFP_KERNEL);
1605 if (!device) {
1606 err = -ENOMEM;
1607 goto err_kzalloc;
1608 }
1609 device->pdev = pdev;
1610 device->reg_base = iobase;
1611 device->version = readb(device->reg_base + IOAT_VER_OFFSET);
1612
1613 /* DMA coherent memory pool for DMA descriptor allocations */
1614 device->dma_pool = pci_pool_create("dma_desc_pool", pdev,
1615 sizeof(struct ioat_dma_descriptor),
1616 64, 0);
1617 if (!device->dma_pool) {
1618 err = -ENOMEM;
1619 goto err_dma_pool;
1620 }
1621
1622 device->completion_pool = pci_pool_create("completion_pool", pdev,
1623 sizeof(u64), SMP_CACHE_BYTES,
1624 SMP_CACHE_BYTES);
1625 if (!device->completion_pool) {
1626 err = -ENOMEM;
1627 goto err_completion_pool;
1628 }
1629
1630 INIT_LIST_HEAD(&device->common.channels);
1631 ioat_dma_enumerate_channels(device);
1632
1633 device->common.device_alloc_chan_resources =
1634 ioat_dma_alloc_chan_resources;
1635 device->common.device_free_chan_resources =
1636 ioat_dma_free_chan_resources;
1637 device->common.dev = &pdev->dev;
1638
1639 dma_cap_set(DMA_MEMCPY, device->common.cap_mask);
1640 device->common.device_is_tx_complete = ioat_dma_is_complete;
1641 switch (device->version) {
1642 case IOAT_VER_1_2:
1643 device->common.device_prep_dma_memcpy = ioat1_dma_prep_memcpy;
1644 device->common.device_issue_pending =
1645 ioat1_dma_memcpy_issue_pending;
1646 break;
1647 case IOAT_VER_2_0:
1648 case IOAT_VER_3_0:
1649 device->common.device_prep_dma_memcpy = ioat2_dma_prep_memcpy;
1650 device->common.device_issue_pending =
1651 ioat2_dma_memcpy_issue_pending;
1652 break;
1653 }
1654
1655 dev_err(&device->pdev->dev,
1656 "Intel(R) I/OAT DMA Engine found,"
1657 " %d channels, device version 0x%02x, driver version %s\n",
1658 device->common.chancnt, device->version, IOAT_DMA_VERSION);
1659
1660 err = ioat_dma_setup_interrupts(device);
1661 if (err)
1662 goto err_setup_interrupts;
1663
1664 err = ioat_dma_self_test(device);
1665 if (err)
1666 goto err_self_test;
1667
1668 ioat_set_tcp_copy_break(device);
1669
1670 dma_async_device_register(&device->common);
1671
1672 if (device->version != IOAT_VER_3_0) {
1673 INIT_DELAYED_WORK(&device->work, ioat_dma_chan_watchdog);
1674 schedule_delayed_work(&device->work,
1675 WATCHDOG_DELAY);
1676 }
1677
1678 return device;
1679
1680 err_self_test:
1681 ioat_dma_remove_interrupts(device);
1682 err_setup_interrupts:
1683 pci_pool_destroy(device->completion_pool);
1684 err_completion_pool:
1685 pci_pool_destroy(device->dma_pool);
1686 err_dma_pool:
1687 kfree(device);
1688 err_kzalloc:
1689 dev_err(&pdev->dev,
1690 "Intel(R) I/OAT DMA Engine initialization failed\n");
1691 return NULL;
1692 }
1693
1694 void ioat_dma_remove(struct ioatdma_device *device)
1695 {
1696 struct dma_chan *chan, *_chan;
1697 struct ioat_dma_chan *ioat_chan;
1698
1699 ioat_dma_remove_interrupts(device);
1700
1701 dma_async_device_unregister(&device->common);
1702
1703 pci_pool_destroy(device->dma_pool);
1704 pci_pool_destroy(device->completion_pool);
1705
1706 iounmap(device->reg_base);
1707 pci_release_regions(device->pdev);
1708 pci_disable_device(device->pdev);
1709
1710 if (device->version != IOAT_VER_3_0) {
1711 cancel_delayed_work(&device->work);
1712 }
1713
1714 list_for_each_entry_safe(chan, _chan,
1715 &device->common.channels, device_node) {
1716 ioat_chan = to_ioat_chan(chan);
1717 list_del(&chan->device_node);
1718 kfree(ioat_chan);
1719 }
1720 kfree(device);
1721 }
1722
Linux kernel - Deliverables
This page took 0.126574 seconds and 5 git commands to generate.