ioat2,3: convert to a true ring buffer
[linux-2.6.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 "dma.h"
38 #include "registers.h"
39 #include "hw.h"
40
41 int ioat_pending_level = 4;
42 module_param(ioat_pending_level, int, 0644);
43 MODULE_PARM_DESC(ioat_pending_level,
44                  "high-water mark for pushing ioat descriptors (default: 4)");
45
46 /* internal functions */
47 static void ioat1_cleanup(struct ioat_dma_chan *ioat);
48 static void ioat1_dma_start_null_desc(struct ioat_dma_chan *ioat);
49
50 /**
51  * ioat_dma_do_interrupt - handler used for single vector interrupt mode
52  * @irq: interrupt id
53  * @data: interrupt data
54  */
55 static irqreturn_t ioat_dma_do_interrupt(int irq, void *data)
56 {
57         struct ioatdma_device *instance = data;
58         struct ioat_chan_common *chan;
59         unsigned long attnstatus;
60         int bit;
61         u8 intrctrl;
62
63         intrctrl = readb(instance->reg_base + IOAT_INTRCTRL_OFFSET);
64
65         if (!(intrctrl & IOAT_INTRCTRL_MASTER_INT_EN))
66                 return IRQ_NONE;
67
68         if (!(intrctrl & IOAT_INTRCTRL_INT_STATUS)) {
69                 writeb(intrctrl, instance->reg_base + IOAT_INTRCTRL_OFFSET);
70                 return IRQ_NONE;
71         }
72
73         attnstatus = readl(instance->reg_base + IOAT_ATTNSTATUS_OFFSET);
74         for_each_bit(bit, &attnstatus, BITS_PER_LONG) {
75                 chan = ioat_chan_by_index(instance, bit);
76                 tasklet_schedule(&chan->cleanup_task);
77         }
78
79         writeb(intrctrl, instance->reg_base + IOAT_INTRCTRL_OFFSET);
80         return IRQ_HANDLED;
81 }
82
83 /**
84  * ioat_dma_do_interrupt_msix - handler used for vector-per-channel interrupt mode
85  * @irq: interrupt id
86  * @data: interrupt data
87  */
88 static irqreturn_t ioat_dma_do_interrupt_msix(int irq, void *data)
89 {
90         struct ioat_chan_common *chan = data;
91
92         tasklet_schedule(&chan->cleanup_task);
93
94         return IRQ_HANDLED;
95 }
96
97 static void ioat1_cleanup_tasklet(unsigned long data);
98
99 /* common channel initialization */
100 void ioat_init_channel(struct ioatdma_device *device,
101                        struct ioat_chan_common *chan, int idx,
102                        work_func_t work_fn, void (*tasklet)(unsigned long),
103                        unsigned long tasklet_data)
104 {
105         struct dma_device *dma = &device->common;
106
107         chan->device = device;
108         chan->reg_base = device->reg_base + (0x80 * (idx + 1));
109         INIT_DELAYED_WORK(&chan->work, work_fn);
110         spin_lock_init(&chan->cleanup_lock);
111         chan->common.device = dma;
112         list_add_tail(&chan->common.device_node, &dma->channels);
113         device->idx[idx] = chan;
114         tasklet_init(&chan->cleanup_task, tasklet, tasklet_data);
115         tasklet_disable(&chan->cleanup_task);
116 }
117
118 static void ioat1_reset_part2(struct work_struct *work);
119
120 /**
121  * ioat1_dma_enumerate_channels - find and initialize the device's channels
122  * @device: the device to be enumerated
123  */
124 static int ioat1_enumerate_channels(struct ioatdma_device *device)
125 {
126         u8 xfercap_scale;
127         u32 xfercap;
128         int i;
129         struct ioat_dma_chan *ioat;
130         struct device *dev = &device->pdev->dev;
131         struct dma_device *dma = &device->common;
132
133         INIT_LIST_HEAD(&dma->channels);
134         dma->chancnt = readb(device->reg_base + IOAT_CHANCNT_OFFSET);
135         xfercap_scale = readb(device->reg_base + IOAT_XFERCAP_OFFSET);
136         xfercap = (xfercap_scale == 0 ? -1 : (1UL << xfercap_scale));
137
138 #ifdef  CONFIG_I7300_IDLE_IOAT_CHANNEL
139         if (i7300_idle_platform_probe(NULL, NULL, 1) == 0)
140                 dma->chancnt--;
141 #endif
142         for (i = 0; i < dma->chancnt; i++) {
143                 ioat = devm_kzalloc(dev, sizeof(*ioat), GFP_KERNEL);
144                 if (!ioat)
145                         break;
146
147                 ioat_init_channel(device, &ioat->base, i,
148                                   ioat1_reset_part2,
149                                   ioat1_cleanup_tasklet,
150                                   (unsigned long) ioat);
151                 ioat->xfercap = xfercap;
152                 spin_lock_init(&ioat->desc_lock);
153                 INIT_LIST_HEAD(&ioat->free_desc);
154                 INIT_LIST_HEAD(&ioat->used_desc);
155         }
156         dma->chancnt = i;
157         return i;
158 }
159
160 /**
161  * ioat_dma_memcpy_issue_pending - push potentially unrecognized appended
162  *                                 descriptors to hw
163  * @chan: DMA channel handle
164  */
165 static inline void
166 __ioat1_dma_memcpy_issue_pending(struct ioat_dma_chan *ioat)
167 {
168         void __iomem *reg_base = ioat->base.reg_base;
169
170         ioat->pending = 0;
171         writeb(IOAT_CHANCMD_APPEND, reg_base + IOAT1_CHANCMD_OFFSET);
172 }
173
174 static void ioat1_dma_memcpy_issue_pending(struct dma_chan *chan)
175 {
176         struct ioat_dma_chan *ioat = to_ioat_chan(chan);
177
178         if (ioat->pending > 0) {
179                 spin_lock_bh(&ioat->desc_lock);
180                 __ioat1_dma_memcpy_issue_pending(ioat);
181                 spin_unlock_bh(&ioat->desc_lock);
182         }
183 }
184
185 /**
186  * ioat1_reset_part2 - reinit the channel after a reset
187  */
188 static void ioat1_reset_part2(struct work_struct *work)
189 {
190         struct ioat_chan_common *chan;
191         struct ioat_dma_chan *ioat;
192         struct ioat_desc_sw *desc;
193         int dmacount;
194         bool start_null = false;
195
196         chan = container_of(work, struct ioat_chan_common, work.work);
197         ioat = container_of(chan, struct ioat_dma_chan, base);
198         spin_lock_bh(&chan->cleanup_lock);
199         spin_lock_bh(&ioat->desc_lock);
200
201         chan->completion_virt->low = 0;
202         chan->completion_virt->high = 0;
203         ioat->pending = 0;
204
205         /* count the descriptors waiting */
206         dmacount = 0;
207         if (ioat->used_desc.prev) {
208                 desc = to_ioat_desc(ioat->used_desc.prev);
209                 do {
210                         dmacount++;
211                         desc = to_ioat_desc(desc->node.next);
212                 } while (&desc->node != ioat->used_desc.next);
213         }
214
215         if (dmacount) {
216                 /*
217                  * write the new starting descriptor address
218                  * this puts channel engine into ARMED state
219                  */
220                 desc = to_ioat_desc(ioat->used_desc.prev);
221                 writel(((u64) desc->txd.phys) & 0x00000000FFFFFFFF,
222                        chan->reg_base + IOAT1_CHAINADDR_OFFSET_LOW);
223                 writel(((u64) desc->txd.phys) >> 32,
224                        chan->reg_base + IOAT1_CHAINADDR_OFFSET_HIGH);
225
226                 writeb(IOAT_CHANCMD_START, chan->reg_base
227                         + IOAT_CHANCMD_OFFSET(chan->device->version));
228         } else
229                 start_null = true;
230         spin_unlock_bh(&ioat->desc_lock);
231         spin_unlock_bh(&chan->cleanup_lock);
232
233         dev_err(to_dev(chan),
234                 "chan%d reset - %d descs waiting, %d total desc\n",
235                 chan_num(chan), dmacount, ioat->desccount);
236
237         if (start_null)
238                 ioat1_dma_start_null_desc(ioat);
239 }
240
241 /**
242  * ioat1_reset_channel - restart a channel
243  * @ioat: IOAT DMA channel handle
244  */
245 static void ioat1_reset_channel(struct ioat_dma_chan *ioat)
246 {
247         struct ioat_chan_common *chan = &ioat->base;
248         void __iomem *reg_base = chan->reg_base;
249         u32 chansts, chanerr;
250
251         if (!ioat->used_desc.prev)
252                 return;
253
254         chanerr = readl(reg_base + IOAT_CHANERR_OFFSET);
255         chansts = (chan->completion_virt->low
256                                         & IOAT_CHANSTS_DMA_TRANSFER_STATUS);
257         if (chanerr) {
258                 dev_err(to_dev(chan),
259                         "chan%d, CHANSTS = 0x%08x CHANERR = 0x%04x, clearing\n",
260                         chan_num(chan), chansts, chanerr);
261                 writel(chanerr, reg_base + IOAT_CHANERR_OFFSET);
262         }
263
264         /*
265          * whack it upside the head with a reset
266          * and wait for things to settle out.
267          * force the pending count to a really big negative
268          * to make sure no one forces an issue_pending
269          * while we're waiting.
270          */
271
272         spin_lock_bh(&ioat->desc_lock);
273         ioat->pending = INT_MIN;
274         writeb(IOAT_CHANCMD_RESET,
275                reg_base + IOAT_CHANCMD_OFFSET(chan->device->version));
276         spin_unlock_bh(&ioat->desc_lock);
277
278         /* schedule the 2nd half instead of sleeping a long time */
279         schedule_delayed_work(&chan->work, RESET_DELAY);
280 }
281
282 /**
283  * ioat1_chan_watchdog - watch for stuck channels
284  */
285 static void ioat1_chan_watchdog(struct work_struct *work)
286 {
287         struct ioatdma_device *device =
288                 container_of(work, struct ioatdma_device, work.work);
289         struct ioat_dma_chan *ioat;
290         struct ioat_chan_common *chan;
291         int i;
292
293         union {
294                 u64 full;
295                 struct {
296                         u32 low;
297                         u32 high;
298                 };
299         } completion_hw;
300         unsigned long compl_desc_addr_hw;
301
302         for (i = 0; i < device->common.chancnt; i++) {
303                 chan = ioat_chan_by_index(device, i);
304                 ioat = container_of(chan, struct ioat_dma_chan, base);
305
306                 if (/* have we started processing anything yet */
307                     chan->last_completion
308                     /* have we completed any since last watchdog cycle? */
309                     && (chan->last_completion == chan->watchdog_completion)
310                     /* has TCP stuck on one cookie since last watchdog? */
311                     && (chan->watchdog_tcp_cookie == chan->watchdog_last_tcp_cookie)
312                     && (chan->watchdog_tcp_cookie != chan->completed_cookie)
313                     /* is there something in the chain to be processed? */
314                     /* CB1 chain always has at least the last one processed */
315                     && (ioat->used_desc.prev != ioat->used_desc.next)
316                     && ioat->pending == 0) {
317
318                         /*
319                          * check CHANSTS register for completed
320                          * descriptor address.
321                          * if it is different than completion writeback,
322                          * it is not zero
323                          * and it has changed since the last watchdog
324                          *     we can assume that channel
325                          *     is still working correctly
326                          *     and the problem is in completion writeback.
327                          *     update completion writeback
328                          *     with actual CHANSTS value
329                          * else
330                          *     try resetting the channel
331                          */
332
333                         completion_hw.low = readl(chan->reg_base +
334                                 IOAT_CHANSTS_OFFSET_LOW(chan->device->version));
335                         completion_hw.high = readl(chan->reg_base +
336                                 IOAT_CHANSTS_OFFSET_HIGH(chan->device->version));
337 #if (BITS_PER_LONG == 64)
338                         compl_desc_addr_hw =
339                                 completion_hw.full
340                                 & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_ADDR;
341 #else
342                         compl_desc_addr_hw =
343                                 completion_hw.low & IOAT_LOW_COMPLETION_MASK;
344 #endif
345
346                         if ((compl_desc_addr_hw != 0)
347                            && (compl_desc_addr_hw != chan->watchdog_completion)
348                            && (compl_desc_addr_hw != chan->last_compl_desc_addr_hw)) {
349                                 chan->last_compl_desc_addr_hw = compl_desc_addr_hw;
350                                 chan->completion_virt->low = completion_hw.low;
351                                 chan->completion_virt->high = completion_hw.high;
352                         } else {
353                                 ioat1_reset_channel(ioat);
354                                 chan->watchdog_completion = 0;
355                                 chan->last_compl_desc_addr_hw = 0;
356                         }
357                 } else {
358                         chan->last_compl_desc_addr_hw = 0;
359                         chan->watchdog_completion = chan->last_completion;
360                 }
361
362                 chan->watchdog_last_tcp_cookie = chan->watchdog_tcp_cookie;
363         }
364
365         schedule_delayed_work(&device->work, WATCHDOG_DELAY);
366 }
367
368 static dma_cookie_t ioat1_tx_submit(struct dma_async_tx_descriptor *tx)
369 {
370         struct dma_chan *c = tx->chan;
371         struct ioat_dma_chan *ioat = to_ioat_chan(c);
372         struct ioat_desc_sw *desc = tx_to_ioat_desc(tx);
373         struct ioat_desc_sw *first;
374         struct ioat_desc_sw *chain_tail;
375         dma_cookie_t cookie;
376
377         spin_lock_bh(&ioat->desc_lock);
378         /* cookie incr and addition to used_list must be atomic */
379         cookie = c->cookie;
380         cookie++;
381         if (cookie < 0)
382                 cookie = 1;
383         c->cookie = cookie;
384         tx->cookie = cookie;
385
386         /* write address into NextDescriptor field of last desc in chain */
387         first = to_ioat_desc(tx->tx_list.next);
388         chain_tail = to_ioat_desc(ioat->used_desc.prev);
389         /* make descriptor updates globally visible before chaining */
390         wmb();
391         chain_tail->hw->next = first->txd.phys;
392         list_splice_tail_init(&tx->tx_list, &ioat->used_desc);
393
394         ioat->pending += desc->tx_cnt;
395         if (ioat->pending >= ioat_pending_level)
396                 __ioat1_dma_memcpy_issue_pending(ioat);
397         spin_unlock_bh(&ioat->desc_lock);
398
399         return cookie;
400 }
401
402 /**
403  * ioat_dma_alloc_descriptor - allocate and return a sw and hw descriptor pair
404  * @ioat: the channel supplying the memory pool for the descriptors
405  * @flags: allocation flags
406  */
407 static struct ioat_desc_sw *
408 ioat_dma_alloc_descriptor(struct ioat_dma_chan *ioat, gfp_t flags)
409 {
410         struct ioat_dma_descriptor *desc;
411         struct ioat_desc_sw *desc_sw;
412         struct ioatdma_device *ioatdma_device;
413         dma_addr_t phys;
414
415         ioatdma_device = ioat->base.device;
416         desc = pci_pool_alloc(ioatdma_device->dma_pool, flags, &phys);
417         if (unlikely(!desc))
418                 return NULL;
419
420         desc_sw = kzalloc(sizeof(*desc_sw), flags);
421         if (unlikely(!desc_sw)) {
422                 pci_pool_free(ioatdma_device->dma_pool, desc, phys);
423                 return NULL;
424         }
425
426         memset(desc, 0, sizeof(*desc));
427
428         dma_async_tx_descriptor_init(&desc_sw->txd, &ioat->base.common);
429         desc_sw->txd.tx_submit = ioat1_tx_submit;
430         desc_sw->hw = desc;
431         desc_sw->txd.phys = phys;
432
433         return desc_sw;
434 }
435
436 static int ioat_initial_desc_count = 256;
437 module_param(ioat_initial_desc_count, int, 0644);
438 MODULE_PARM_DESC(ioat_initial_desc_count,
439                  "ioat1: initial descriptors per channel (default: 256)");
440 /**
441  * ioat1_dma_alloc_chan_resources - returns the number of allocated descriptors
442  * @chan: the channel to be filled out
443  */
444 static int ioat1_dma_alloc_chan_resources(struct dma_chan *c)
445 {
446         struct ioat_dma_chan *ioat = to_ioat_chan(c);
447         struct ioat_chan_common *chan = &ioat->base;
448         struct ioat_desc_sw *desc;
449         u16 chanctrl;
450         u32 chanerr;
451         int i;
452         LIST_HEAD(tmp_list);
453
454         /* have we already been set up? */
455         if (!list_empty(&ioat->free_desc))
456                 return ioat->desccount;
457
458         /* Setup register to interrupt and write completion status on error */
459         chanctrl = IOAT_CHANCTRL_ERR_INT_EN |
460                 IOAT_CHANCTRL_ANY_ERR_ABORT_EN |
461                 IOAT_CHANCTRL_ERR_COMPLETION_EN;
462         writew(chanctrl, chan->reg_base + IOAT_CHANCTRL_OFFSET);
463
464         chanerr = readl(chan->reg_base + IOAT_CHANERR_OFFSET);
465         if (chanerr) {
466                 dev_err(to_dev(chan), "CHANERR = %x, clearing\n", chanerr);
467                 writel(chanerr, chan->reg_base + IOAT_CHANERR_OFFSET);
468         }
469
470         /* Allocate descriptors */
471         for (i = 0; i < ioat_initial_desc_count; i++) {
472                 desc = ioat_dma_alloc_descriptor(ioat, GFP_KERNEL);
473                 if (!desc) {
474                         dev_err(to_dev(chan), "Only %d initial descriptors\n", i);
475                         break;
476                 }
477                 list_add_tail(&desc->node, &tmp_list);
478         }
479         spin_lock_bh(&ioat->desc_lock);
480         ioat->desccount = i;
481         list_splice(&tmp_list, &ioat->free_desc);
482         spin_unlock_bh(&ioat->desc_lock);
483
484         /* allocate a completion writeback area */
485         /* doing 2 32bit writes to mmio since 1 64b write doesn't work */
486         chan->completion_virt = pci_pool_alloc(chan->device->completion_pool,
487                                                GFP_KERNEL,
488                                                &chan->completion_addr);
489         memset(chan->completion_virt, 0,
490                sizeof(*chan->completion_virt));
491         writel(((u64) chan->completion_addr) & 0x00000000FFFFFFFF,
492                chan->reg_base + IOAT_CHANCMP_OFFSET_LOW);
493         writel(((u64) chan->completion_addr) >> 32,
494                chan->reg_base + IOAT_CHANCMP_OFFSET_HIGH);
495
496         tasklet_enable(&chan->cleanup_task);
497         ioat1_dma_start_null_desc(ioat);  /* give chain to dma device */
498         return ioat->desccount;
499 }
500
501 /**
502  * ioat1_dma_free_chan_resources - release all the descriptors
503  * @chan: the channel to be cleaned
504  */
505 static void ioat1_dma_free_chan_resources(struct dma_chan *c)
506 {
507         struct ioat_dma_chan *ioat = to_ioat_chan(c);
508         struct ioat_chan_common *chan = &ioat->base;
509         struct ioatdma_device *ioatdma_device = chan->device;
510         struct ioat_desc_sw *desc, *_desc;
511         int in_use_descs = 0;
512
513         /* Before freeing channel resources first check
514          * if they have been previously allocated for this channel.
515          */
516         if (ioat->desccount == 0)
517                 return;
518
519         tasklet_disable(&chan->cleanup_task);
520         ioat1_cleanup(ioat);
521
522         /* Delay 100ms after reset to allow internal DMA logic to quiesce
523          * before removing DMA descriptor resources.
524          */
525         writeb(IOAT_CHANCMD_RESET,
526                chan->reg_base + IOAT_CHANCMD_OFFSET(chan->device->version));
527         mdelay(100);
528
529         spin_lock_bh(&ioat->desc_lock);
530         list_for_each_entry_safe(desc, _desc,
531                                  &ioat->used_desc, node) {
532                 in_use_descs++;
533                 list_del(&desc->node);
534                 pci_pool_free(ioatdma_device->dma_pool, desc->hw,
535                               desc->txd.phys);
536                 kfree(desc);
537         }
538         list_for_each_entry_safe(desc, _desc,
539                                  &ioat->free_desc, node) {
540                 list_del(&desc->node);
541                 pci_pool_free(ioatdma_device->dma_pool, desc->hw,
542                               desc->txd.phys);
543                 kfree(desc);
544         }
545         spin_unlock_bh(&ioat->desc_lock);
546
547         pci_pool_free(ioatdma_device->completion_pool,
548                       chan->completion_virt,
549                       chan->completion_addr);
550
551         /* one is ok since we left it on there on purpose */
552         if (in_use_descs > 1)
553                 dev_err(to_dev(chan), "Freeing %d in use descriptors!\n",
554                         in_use_descs - 1);
555
556         chan->last_completion = chan->completion_addr = 0;
557         chan->watchdog_completion = 0;
558         chan->last_compl_desc_addr_hw = 0;
559         chan->watchdog_tcp_cookie = chan->watchdog_last_tcp_cookie = 0;
560         ioat->pending = 0;
561         ioat->desccount = 0;
562 }
563
564 /**
565  * ioat1_dma_get_next_descriptor - return the next available descriptor
566  * @ioat: IOAT DMA channel handle
567  *
568  * Gets the next descriptor from the chain, and must be called with the
569  * channel's desc_lock held.  Allocates more descriptors if the channel
570  * has run out.
571  */
572 static struct ioat_desc_sw *
573 ioat1_dma_get_next_descriptor(struct ioat_dma_chan *ioat)
574 {
575         struct ioat_desc_sw *new;
576
577         if (!list_empty(&ioat->free_desc)) {
578                 new = to_ioat_desc(ioat->free_desc.next);
579                 list_del(&new->node);
580         } else {
581                 /* try to get another desc */
582                 new = ioat_dma_alloc_descriptor(ioat, GFP_ATOMIC);
583                 if (!new) {
584                         dev_err(to_dev(&ioat->base), "alloc failed\n");
585                         return NULL;
586                 }
587         }
588
589         prefetch(new->hw);
590         return new;
591 }
592
593 static struct dma_async_tx_descriptor *
594 ioat1_dma_prep_memcpy(struct dma_chan *c, dma_addr_t dma_dest,
595                       dma_addr_t dma_src, size_t len, unsigned long flags)
596 {
597         struct ioat_dma_chan *ioat = to_ioat_chan(c);
598         struct ioat_desc_sw *desc;
599         size_t copy;
600         LIST_HEAD(chain);
601         dma_addr_t src = dma_src;
602         dma_addr_t dest = dma_dest;
603         size_t total_len = len;
604         struct ioat_dma_descriptor *hw = NULL;
605         int tx_cnt = 0;
606
607         spin_lock_bh(&ioat->desc_lock);
608         desc = ioat1_dma_get_next_descriptor(ioat);
609         do {
610                 if (!desc)
611                         break;
612
613                 tx_cnt++;
614                 copy = min_t(size_t, len, ioat->xfercap);
615
616                 hw = desc->hw;
617                 hw->size = copy;
618                 hw->ctl = 0;
619                 hw->src_addr = src;
620                 hw->dst_addr = dest;
621
622                 list_add_tail(&desc->node, &chain);
623
624                 len -= copy;
625                 dest += copy;
626                 src += copy;
627                 if (len) {
628                         struct ioat_desc_sw *next;
629
630                         async_tx_ack(&desc->txd);
631                         next = ioat1_dma_get_next_descriptor(ioat);
632                         hw->next = next ? next->txd.phys : 0;
633                         desc = next;
634                 } else
635                         hw->next = 0;
636         } while (len);
637
638         if (!desc) {
639                 struct ioat_chan_common *chan = &ioat->base;
640
641                 dev_err(to_dev(chan),
642                         "chan%d - get_next_desc failed\n", chan_num(chan));
643                 list_splice(&chain, &ioat->free_desc);
644                 spin_unlock_bh(&ioat->desc_lock);
645                 return NULL;
646         }
647         spin_unlock_bh(&ioat->desc_lock);
648
649         desc->txd.flags = flags;
650         desc->tx_cnt = tx_cnt;
651         desc->src = dma_src;
652         desc->dst = dma_dest;
653         desc->len = total_len;
654         list_splice(&chain, &desc->txd.tx_list);
655         hw->ctl_f.int_en = !!(flags & DMA_PREP_INTERRUPT);
656         hw->ctl_f.compl_write = 1;
657
658         return &desc->txd;
659 }
660
661 static void ioat1_cleanup_tasklet(unsigned long data)
662 {
663         struct ioat_dma_chan *chan = (void *)data;
664         ioat1_cleanup(chan);
665         writew(IOAT_CHANCTRL_INT_DISABLE,
666                chan->base.reg_base + IOAT_CHANCTRL_OFFSET);
667 }
668
669 static void ioat_unmap(struct pci_dev *pdev, dma_addr_t addr, size_t len,
670                        int direction, enum dma_ctrl_flags flags, bool dst)
671 {
672         if ((dst && (flags & DMA_COMPL_DEST_UNMAP_SINGLE)) ||
673             (!dst && (flags & DMA_COMPL_SRC_UNMAP_SINGLE)))
674                 pci_unmap_single(pdev, addr, len, direction);
675         else
676                 pci_unmap_page(pdev, addr, len, direction);
677 }
678
679
680 void ioat_dma_unmap(struct ioat_chan_common *chan, enum dma_ctrl_flags flags,
681                     size_t len, struct ioat_dma_descriptor *hw)
682 {
683         struct pci_dev *pdev = chan->device->pdev;
684         size_t offset = len - hw->size;
685
686         if (!(flags & DMA_COMPL_SKIP_DEST_UNMAP))
687                 ioat_unmap(pdev, hw->dst_addr - offset, len,
688                            PCI_DMA_FROMDEVICE, flags, 1);
689
690         if (!(flags & DMA_COMPL_SKIP_SRC_UNMAP))
691                 ioat_unmap(pdev, hw->src_addr - offset, len,
692                            PCI_DMA_TODEVICE, flags, 0);
693 }
694
695 unsigned long ioat_get_current_completion(struct ioat_chan_common *chan)
696 {
697         unsigned long phys_complete;
698
699         /* The completion writeback can happen at any time,
700            so reads by the driver need to be atomic operations
701            The descriptor physical addresses are limited to 32-bits
702            when the CPU can only do a 32-bit mov */
703
704 #if (BITS_PER_LONG == 64)
705         phys_complete =
706                 chan->completion_virt->full
707                 & IOAT_CHANSTS_COMPLETED_DESCRIPTOR_ADDR;
708 #else
709         phys_complete = chan->completion_virt->low & IOAT_LOW_COMPLETION_MASK;
710 #endif
711
712         if ((chan->completion_virt->full
713                 & IOAT_CHANSTS_DMA_TRANSFER_STATUS) ==
714                                 IOAT_CHANSTS_DMA_TRANSFER_STATUS_HALTED) {
715                 dev_err(to_dev(chan), "Channel halted, chanerr = %x\n",
716                         readl(chan->reg_base + IOAT_CHANERR_OFFSET));
717
718                 /* TODO do something to salvage the situation */
719         }
720
721         return phys_complete;
722 }
723
724 /**
725  * ioat1_cleanup - cleanup up finished descriptors
726  * @chan: ioat channel to be cleaned up
727  */
728 static void ioat1_cleanup(struct ioat_dma_chan *ioat)
729 {
730         struct ioat_chan_common *chan = &ioat->base;
731         unsigned long phys_complete;
732         struct ioat_desc_sw *desc, *_desc;
733         dma_cookie_t cookie = 0;
734         struct dma_async_tx_descriptor *tx;
735
736         prefetch(chan->completion_virt);
737
738         if (!spin_trylock_bh(&chan->cleanup_lock))
739                 return;
740
741         phys_complete = ioat_get_current_completion(chan);
742         if (phys_complete == chan->last_completion) {
743                 spin_unlock_bh(&chan->cleanup_lock);
744                 /*
745                  * perhaps we're stuck so hard that the watchdog can't go off?
746                  * try to catch it after 2 seconds
747                  */
748                 if (time_after(jiffies,
749                                chan->last_completion_time + HZ*WATCHDOG_DELAY)) {
750                         ioat1_chan_watchdog(&(chan->device->work.work));
751                         chan->last_completion_time = jiffies;
752                 }
753                 return;
754         }
755         chan->last_completion_time = jiffies;
756
757         cookie = 0;
758         if (!spin_trylock_bh(&ioat->desc_lock)) {
759                 spin_unlock_bh(&chan->cleanup_lock);
760                 return;
761         }
762
763         list_for_each_entry_safe(desc, _desc, &ioat->used_desc, node) {
764                 tx = &desc->txd;
765                 /*
766                  * Incoming DMA requests may use multiple descriptors,
767                  * due to exceeding xfercap, perhaps. If so, only the
768                  * last one will have a cookie, and require unmapping.
769                  */
770                 if (tx->cookie) {
771                         cookie = tx->cookie;
772                         ioat_dma_unmap(chan, tx->flags, desc->len, desc->hw);
773                         if (tx->callback) {
774                                 tx->callback(tx->callback_param);
775                                 tx->callback = NULL;
776                         }
777                 }
778
779                 if (tx->phys != phys_complete) {
780                         /*
781                          * a completed entry, but not the last, so clean
782                          * up if the client is done with the descriptor
783                          */
784                         if (async_tx_test_ack(tx))
785                                 list_move_tail(&desc->node, &ioat->free_desc);
786                         else
787                                 tx->cookie = 0;
788                 } else {
789                         /*
790                          * last used desc. Do not remove, so we can
791                          * append from it, but don't look at it next
792                          * time, either
793                          */
794                         tx->cookie = 0;
795
796                         /* TODO check status bits? */
797                         break;
798                 }
799         }
800
801         spin_unlock_bh(&ioat->desc_lock);
802
803         chan->last_completion = phys_complete;
804         if (cookie != 0)
805                 chan->completed_cookie = cookie;
806
807         spin_unlock_bh(&chan->cleanup_lock);
808 }
809
810 static enum dma_status
811 ioat1_dma_is_complete(struct dma_chan *c, dma_cookie_t cookie,
812                       dma_cookie_t *done, dma_cookie_t *used)
813 {
814         struct ioat_dma_chan *ioat = to_ioat_chan(c);
815
816         if (ioat_is_complete(c, cookie, done, used) == DMA_SUCCESS)
817                 return DMA_SUCCESS;
818
819         ioat1_cleanup(ioat);
820
821         return ioat_is_complete(c, cookie, done, used);
822 }
823
824 static void ioat1_dma_start_null_desc(struct ioat_dma_chan *ioat)
825 {
826         struct ioat_chan_common *chan = &ioat->base;
827         struct ioat_desc_sw *desc;
828         struct ioat_dma_descriptor *hw;
829
830         spin_lock_bh(&ioat->desc_lock);
831
832         desc = ioat1_dma_get_next_descriptor(ioat);
833
834         if (!desc) {
835                 dev_err(to_dev(chan),
836                         "Unable to start null desc - get next desc failed\n");
837                 spin_unlock_bh(&ioat->desc_lock);
838                 return;
839         }
840
841         hw = desc->hw;
842         hw->ctl = 0;
843         hw->ctl_f.null = 1;
844         hw->ctl_f.int_en = 1;
845         hw->ctl_f.compl_write = 1;
846         /* set size to non-zero value (channel returns error when size is 0) */
847         hw->size = NULL_DESC_BUFFER_SIZE;
848         hw->src_addr = 0;
849         hw->dst_addr = 0;
850         async_tx_ack(&desc->txd);
851         hw->next = 0;
852         list_add_tail(&desc->node, &ioat->used_desc);
853
854         writel(((u64) desc->txd.phys) & 0x00000000FFFFFFFF,
855                chan->reg_base + IOAT1_CHAINADDR_OFFSET_LOW);
856         writel(((u64) desc->txd.phys) >> 32,
857                chan->reg_base + IOAT1_CHAINADDR_OFFSET_HIGH);
858
859         writeb(IOAT_CHANCMD_START, chan->reg_base
860                 + IOAT_CHANCMD_OFFSET(chan->device->version));
861         spin_unlock_bh(&ioat->desc_lock);
862 }
863
864 /*
865  * Perform a IOAT transaction to verify the HW works.
866  */
867 #define IOAT_TEST_SIZE 2000
868
869 static void ioat_dma_test_callback(void *dma_async_param)
870 {
871         struct completion *cmp = dma_async_param;
872
873         complete(cmp);
874 }
875
876 /**
877  * ioat_dma_self_test - Perform a IOAT transaction to verify the HW works.
878  * @device: device to be tested
879  */
880 static int ioat_dma_self_test(struct ioatdma_device *device)
881 {
882         int i;
883         u8 *src;
884         u8 *dest;
885         struct dma_device *dma = &device->common;
886         struct device *dev = &device->pdev->dev;
887         struct dma_chan *dma_chan;
888         struct dma_async_tx_descriptor *tx;
889         dma_addr_t dma_dest, dma_src;
890         dma_cookie_t cookie;
891         int err = 0;
892         struct completion cmp;
893         unsigned long tmo;
894         unsigned long flags;
895
896         src = kzalloc(sizeof(u8) * IOAT_TEST_SIZE, GFP_KERNEL);
897         if (!src)
898                 return -ENOMEM;
899         dest = kzalloc(sizeof(u8) * IOAT_TEST_SIZE, GFP_KERNEL);
900         if (!dest) {
901                 kfree(src);
902                 return -ENOMEM;
903         }
904
905         /* Fill in src buffer */
906         for (i = 0; i < IOAT_TEST_SIZE; i++)
907                 src[i] = (u8)i;
908
909         /* Start copy, using first DMA channel */
910         dma_chan = container_of(dma->channels.next, struct dma_chan,
911                                 device_node);
912         if (dma->device_alloc_chan_resources(dma_chan) < 1) {
913                 dev_err(dev, "selftest cannot allocate chan resource\n");
914                 err = -ENODEV;
915                 goto out;
916         }
917
918         dma_src = dma_map_single(dev, src, IOAT_TEST_SIZE, DMA_TO_DEVICE);
919         dma_dest = dma_map_single(dev, dest, IOAT_TEST_SIZE, DMA_FROM_DEVICE);
920         flags = DMA_COMPL_SRC_UNMAP_SINGLE | DMA_COMPL_DEST_UNMAP_SINGLE |
921                 DMA_PREP_INTERRUPT;
922         tx = device->common.device_prep_dma_memcpy(dma_chan, dma_dest, dma_src,
923                                                    IOAT_TEST_SIZE, flags);
924         if (!tx) {
925                 dev_err(dev, "Self-test prep failed, disabling\n");
926                 err = -ENODEV;
927                 goto free_resources;
928         }
929
930         async_tx_ack(tx);
931         init_completion(&cmp);
932         tx->callback = ioat_dma_test_callback;
933         tx->callback_param = &cmp;
934         cookie = tx->tx_submit(tx);
935         if (cookie < 0) {
936                 dev_err(dev, "Self-test setup failed, disabling\n");
937                 err = -ENODEV;
938                 goto free_resources;
939         }
940         dma->device_issue_pending(dma_chan);
941
942         tmo = wait_for_completion_timeout(&cmp, msecs_to_jiffies(3000));
943
944         if (tmo == 0 ||
945             dma->device_is_tx_complete(dma_chan, cookie, NULL, NULL)
946                                         != DMA_SUCCESS) {
947                 dev_err(dev, "Self-test copy timed out, disabling\n");
948                 err = -ENODEV;
949                 goto free_resources;
950         }
951         if (memcmp(src, dest, IOAT_TEST_SIZE)) {
952                 dev_err(dev, "Self-test copy failed compare, disabling\n");
953                 err = -ENODEV;
954                 goto free_resources;
955         }
956
957 free_resources:
958         dma->device_free_chan_resources(dma_chan);
959 out:
960         kfree(src);
961         kfree(dest);
962         return err;
963 }
964
965 static char ioat_interrupt_style[32] = "msix";
966 module_param_string(ioat_interrupt_style, ioat_interrupt_style,
967                     sizeof(ioat_interrupt_style), 0644);
968 MODULE_PARM_DESC(ioat_interrupt_style,
969                  "set ioat interrupt style: msix (default), "
970                  "msix-single-vector, msi, intx)");
971
972 /**
973  * ioat_dma_setup_interrupts - setup interrupt handler
974  * @device: ioat device
975  */
976 static int ioat_dma_setup_interrupts(struct ioatdma_device *device)
977 {
978         struct ioat_chan_common *chan;
979         struct pci_dev *pdev = device->pdev;
980         struct device *dev = &pdev->dev;
981         struct msix_entry *msix;
982         int i, j, msixcnt;
983         int err = -EINVAL;
984         u8 intrctrl = 0;
985
986         if (!strcmp(ioat_interrupt_style, "msix"))
987                 goto msix;
988         if (!strcmp(ioat_interrupt_style, "msix-single-vector"))
989                 goto msix_single_vector;
990         if (!strcmp(ioat_interrupt_style, "msi"))
991                 goto msi;
992         if (!strcmp(ioat_interrupt_style, "intx"))
993                 goto intx;
994         dev_err(dev, "invalid ioat_interrupt_style %s\n", ioat_interrupt_style);
995         goto err_no_irq;
996
997 msix:
998         /* The number of MSI-X vectors should equal the number of channels */
999         msixcnt = device->common.chancnt;
1000         for (i = 0; i < msixcnt; i++)
1001                 device->msix_entries[i].entry = i;
1002
1003         err = pci_enable_msix(pdev, device->msix_entries, msixcnt);
1004         if (err < 0)
1005                 goto msi;
1006         if (err > 0)
1007                 goto msix_single_vector;
1008
1009         for (i = 0; i < msixcnt; i++) {
1010                 msix = &device->msix_entries[i];
1011                 chan = ioat_chan_by_index(device, i);
1012                 err = devm_request_irq(dev, msix->vector,
1013                                        ioat_dma_do_interrupt_msix, 0,
1014                                        "ioat-msix", chan);
1015                 if (err) {
1016                         for (j = 0; j < i; j++) {
1017                                 msix = &device->msix_entries[j];
1018                                 chan = ioat_chan_by_index(device, j);
1019                                 devm_free_irq(dev, msix->vector, chan);
1020                         }
1021                         goto msix_single_vector;
1022                 }
1023         }
1024         intrctrl |= IOAT_INTRCTRL_MSIX_VECTOR_CONTROL;
1025         goto done;
1026
1027 msix_single_vector:
1028         msix = &device->msix_entries[0];
1029         msix->entry = 0;
1030         err = pci_enable_msix(pdev, device->msix_entries, 1);
1031         if (err)
1032                 goto msi;
1033
1034         err = devm_request_irq(dev, msix->vector, ioat_dma_do_interrupt, 0,
1035                                "ioat-msix", device);
1036         if (err) {
1037                 pci_disable_msix(pdev);
1038                 goto msi;
1039         }
1040         goto done;
1041
1042 msi:
1043         err = pci_enable_msi(pdev);
1044         if (err)
1045                 goto intx;
1046
1047         err = devm_request_irq(dev, pdev->irq, ioat_dma_do_interrupt, 0,
1048                                "ioat-msi", device);
1049         if (err) {
1050                 pci_disable_msi(pdev);
1051                 goto intx;
1052         }
1053         goto done;
1054
1055 intx:
1056         err = devm_request_irq(dev, pdev->irq, ioat_dma_do_interrupt,
1057                                IRQF_SHARED, "ioat-intx", device);
1058         if (err)
1059                 goto err_no_irq;
1060
1061 done:
1062         if (device->intr_quirk)
1063                 device->intr_quirk(device);
1064         intrctrl |= IOAT_INTRCTRL_MASTER_INT_EN;
1065         writeb(intrctrl, device->reg_base + IOAT_INTRCTRL_OFFSET);
1066         return 0;
1067
1068 err_no_irq:
1069         /* Disable all interrupt generation */
1070         writeb(0, device->reg_base + IOAT_INTRCTRL_OFFSET);
1071         dev_err(dev, "no usable interrupts\n");
1072         return err;
1073 }
1074
1075 static void ioat_disable_interrupts(struct ioatdma_device *device)
1076 {
1077         /* Disable all interrupt generation */
1078         writeb(0, device->reg_base + IOAT_INTRCTRL_OFFSET);
1079 }
1080
1081 int ioat_probe(struct ioatdma_device *device)
1082 {
1083         int err = -ENODEV;
1084         struct dma_device *dma = &device->common;
1085         struct pci_dev *pdev = device->pdev;
1086         struct device *dev = &pdev->dev;
1087
1088         /* DMA coherent memory pool for DMA descriptor allocations */
1089         device->dma_pool = pci_pool_create("dma_desc_pool", pdev,
1090                                            sizeof(struct ioat_dma_descriptor),
1091                                            64, 0);
1092         if (!device->dma_pool) {
1093                 err = -ENOMEM;
1094                 goto err_dma_pool;
1095         }
1096
1097         device->completion_pool = pci_pool_create("completion_pool", pdev,
1098                                                   sizeof(u64), SMP_CACHE_BYTES,
1099                                                   SMP_CACHE_BYTES);
1100
1101         if (!device->completion_pool) {
1102                 err = -ENOMEM;
1103                 goto err_completion_pool;
1104         }
1105
1106         device->enumerate_channels(device);
1107
1108         dma_cap_set(DMA_MEMCPY, dma->cap_mask);
1109         dma->dev = &pdev->dev;
1110
1111         dev_err(dev, "Intel(R) I/OAT DMA Engine found,"
1112                 " %d channels, device version 0x%02x, driver version %s\n",
1113                 dma->chancnt, device->version, IOAT_DMA_VERSION);
1114
1115         if (!dma->chancnt) {
1116                 dev_err(dev, "Intel(R) I/OAT DMA Engine problem found: "
1117                         "zero channels detected\n");
1118                 goto err_setup_interrupts;
1119         }
1120
1121         err = ioat_dma_setup_interrupts(device);
1122         if (err)
1123                 goto err_setup_interrupts;
1124
1125         err = ioat_dma_self_test(device);
1126         if (err)
1127                 goto err_self_test;
1128
1129         return 0;
1130
1131 err_self_test:
1132         ioat_disable_interrupts(device);
1133 err_setup_interrupts:
1134         pci_pool_destroy(device->completion_pool);
1135 err_completion_pool:
1136         pci_pool_destroy(device->dma_pool);
1137 err_dma_pool:
1138         return err;
1139 }
1140
1141 int ioat_register(struct ioatdma_device *device)
1142 {
1143         int err = dma_async_device_register(&device->common);
1144
1145         if (err) {
1146                 ioat_disable_interrupts(device);
1147                 pci_pool_destroy(device->completion_pool);
1148                 pci_pool_destroy(device->dma_pool);
1149         }
1150
1151         return err;
1152 }
1153
1154 /* ioat1_intr_quirk - fix up dma ctrl register to enable / disable msi */
1155 static void ioat1_intr_quirk(struct ioatdma_device *device)
1156 {
1157         struct pci_dev *pdev = device->pdev;
1158         u32 dmactrl;
1159
1160         pci_read_config_dword(pdev, IOAT_PCI_DMACTRL_OFFSET, &dmactrl);
1161         if (pdev->msi_enabled)
1162                 dmactrl |= IOAT_PCI_DMACTRL_MSI_EN;
1163         else
1164                 dmactrl &= ~IOAT_PCI_DMACTRL_MSI_EN;
1165         pci_write_config_dword(pdev, IOAT_PCI_DMACTRL_OFFSET, dmactrl);
1166 }
1167
1168 int ioat1_dma_probe(struct ioatdma_device *device, int dca)
1169 {
1170         struct pci_dev *pdev = device->pdev;
1171         struct dma_device *dma;
1172         int err;
1173
1174         device->intr_quirk = ioat1_intr_quirk;
1175         device->enumerate_channels = ioat1_enumerate_channels;
1176         dma = &device->common;
1177         dma->device_prep_dma_memcpy = ioat1_dma_prep_memcpy;
1178         dma->device_issue_pending = ioat1_dma_memcpy_issue_pending;
1179         dma->device_alloc_chan_resources = ioat1_dma_alloc_chan_resources;
1180         dma->device_free_chan_resources = ioat1_dma_free_chan_resources;
1181         dma->device_is_tx_complete = ioat1_dma_is_complete;
1182
1183         err = ioat_probe(device);
1184         if (err)
1185                 return err;
1186         ioat_set_tcp_copy_break(4096);
1187         err = ioat_register(device);
1188         if (err)
1189                 return err;
1190         if (dca)
1191                 device->dca = ioat_dca_init(pdev, device->reg_base);
1192
1193         INIT_DELAYED_WORK(&device->work, ioat1_chan_watchdog);
1194         schedule_delayed_work(&device->work, WATCHDOG_DELAY);
1195
1196         return err;
1197 }
1198
1199 void ioat_dma_remove(struct ioatdma_device *device)
1200 {
1201         struct dma_device *dma = &device->common;
1202
1203         if (device->version != IOAT_VER_3_0)
1204                 cancel_delayed_work(&device->work);
1205
1206         ioat_disable_interrupts(device);
1207
1208         dma_async_device_unregister(dma);
1209
1210         pci_pool_destroy(device->dma_pool);
1211         pci_pool_destroy(device->completion_pool);
1212
1213         INIT_LIST_HEAD(&dma->channels);
1214 }